2019-06-01 08:08:55 +00:00
|
|
|
// SPDX-License-Identifier: GPL-2.0-only
|
x86/bpf: Clean up non-standard comments, to make the code more readable
So by chance I looked into x86 assembly in arch/x86/net/bpf_jit_comp.c and
noticed the weird and inconsistent comment style it mistakenly learned from
the networking code:
/* Multi-line comment ...
* ... looks like this.
*/
Fix this to use the standard comment style specified in Documentation/CodingStyle
and used in arch/x86/ as well:
/*
* Multi-line comment ...
* ... looks like this.
*/
Also, to quote Linus's ... more explicit views about this:
http://article.gmane.org/gmane.linux.kernel.cryptoapi/21066
> But no, the networking code picked *none* of the above sane formats.
> Instead, it picked these two models that are just half-arsed
> shit-for-brains:
>
> (no)
> /* This is disgusting drug-induced
> * crap, and should die
> */
>
> (no-no-no)
> /* This is also very nasty
> * and visually unbalanced */
>
> Please. The networking code actually has the *worst* possible comment
> style. You can literally find that (no-no-no) style, which is just
> really horribly disgusting and worse than the otherwise fairly similar
> (d) in pretty much every way.
Also improve the comments and some other details while at it:
- Don't mix same-line and previous-line comment style on otherwise
identical code patterns within the same function,
- capitalize 'BPF' and x86 register names consistently,
- capitalize sentences consistently,
- instead of 'x64' use 'x86-64': x64 is a Microsoft specific term,
- use more consistent punctuation,
- use standard coding style in macros as well,
- fix typos and a few other minor details.
Consistent coding style is not optional, at least in arch/x86/.
No change in functionality.
( In case this commit causes conflicts with pending development code
I'll be glad to help resolve any conflicts! )
Acked-by: Thomas Gleixner <tglx@linutronix.de>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Borislav Petkov <bp@alien8.de>
Cc: H. Peter Anvin <hpa@zytor.com>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: David S. Miller <davem@davemloft.net>
Cc: Eric Dumazet <edumazet@google.com>
Cc: Daniel Borkmann <daniel@iogearbox.net>
Cc: Alexei Starovoitov <ast@fb.com>
Cc: Hideaki YOSHIFUJI <yoshfuji@linux-ipv6.org>
Cc: netdev@vger.kernel.org
Cc: linux-kernel@vger.kernel.org
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
2018-04-27 09:54:40 +00:00
|
|
|
/*
|
2021-11-19 16:32:11 +00:00
|
|
|
* BPF JIT compiler
|
2011-04-20 09:27:32 +00:00
|
|
|
*
|
2013-01-31 01:51:44 +00:00
|
|
|
* Copyright (C) 2011-2013 Eric Dumazet (eric.dumazet@gmail.com)
|
2021-11-19 16:32:11 +00:00
|
|
|
* Copyright (c) 2011-2014 PLUMgrid, http://plumgrid.com
|
2011-04-20 09:27:32 +00:00
|
|
|
*/
|
|
|
|
#include <linux/netdevice.h>
|
|
|
|
#include <linux/filter.h>
|
2012-10-27 02:26:22 +00:00
|
|
|
#include <linux/if_vlan.h>
|
2018-02-26 21:13:52 +00:00
|
|
|
#include <linux/bpf.h>
|
2019-11-14 18:57:03 +00:00
|
|
|
#include <linux/memory.h>
|
2019-12-13 17:51:08 +00:00
|
|
|
#include <linux/sort.h>
|
2019-10-16 03:25:03 +00:00
|
|
|
#include <asm/extable.h>
|
2022-09-15 11:11:37 +00:00
|
|
|
#include <asm/ftrace.h>
|
2017-05-08 22:58:11 +00:00
|
|
|
#include <asm/set_memory.h>
|
bpf, x64: implement retpoline for tail call
Implement a retpoline [0] for the BPF tail call JIT'ing that converts
the indirect jump via jmp %rax that is used to make the long jump into
another JITed BPF image. Since this is subject to speculative execution,
we need to control the transient instruction sequence here as well
when CONFIG_RETPOLINE is set, and direct it into a pause + lfence loop.
The latter aligns also with what gcc / clang emits (e.g. [1]).
JIT dump after patch:
# bpftool p d x i 1
0: (18) r2 = map[id:1]
2: (b7) r3 = 0
3: (85) call bpf_tail_call#12
4: (b7) r0 = 2
5: (95) exit
With CONFIG_RETPOLINE:
# bpftool p d j i 1
[...]
33: cmp %edx,0x24(%rsi)
36: jbe 0x0000000000000072 |*
38: mov 0x24(%rbp),%eax
3e: cmp $0x20,%eax
41: ja 0x0000000000000072 |
43: add $0x1,%eax
46: mov %eax,0x24(%rbp)
4c: mov 0x90(%rsi,%rdx,8),%rax
54: test %rax,%rax
57: je 0x0000000000000072 |
59: mov 0x28(%rax),%rax
5d: add $0x25,%rax
61: callq 0x000000000000006d |+
66: pause |
68: lfence |
6b: jmp 0x0000000000000066 |
6d: mov %rax,(%rsp) |
71: retq |
72: mov $0x2,%eax
[...]
* relative fall-through jumps in error case
+ retpoline for indirect jump
Without CONFIG_RETPOLINE:
# bpftool p d j i 1
[...]
33: cmp %edx,0x24(%rsi)
36: jbe 0x0000000000000063 |*
38: mov 0x24(%rbp),%eax
3e: cmp $0x20,%eax
41: ja 0x0000000000000063 |
43: add $0x1,%eax
46: mov %eax,0x24(%rbp)
4c: mov 0x90(%rsi,%rdx,8),%rax
54: test %rax,%rax
57: je 0x0000000000000063 |
59: mov 0x28(%rax),%rax
5d: add $0x25,%rax
61: jmpq *%rax |-
63: mov $0x2,%eax
[...]
* relative fall-through jumps in error case
- plain indirect jump as before
[0] https://support.google.com/faqs/answer/7625886
[1] https://github.com/gcc-mirror/gcc/commit/a31e654fa107be968b802786d747e962c2fcdb2b
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
2018-02-22 14:12:53 +00:00
|
|
|
#include <asm/nospec-branch.h>
|
2019-11-14 18:57:03 +00:00
|
|
|
#include <asm/text-patching.h>
|
2023-09-12 23:31:59 +00:00
|
|
|
#include <asm/unwind.h>
|
2023-12-15 09:12:18 +00:00
|
|
|
#include <asm/cfi.h>
|
2011-04-20 09:27:32 +00:00
|
|
|
|
bpf: Implement BPF exceptions
This patch implements BPF exceptions, and introduces a bpf_throw kfunc
to allow programs to throw exceptions during their execution at runtime.
A bpf_throw invocation is treated as an immediate termination of the
program, returning back to its caller within the kernel, unwinding all
stack frames.
This allows the program to simplify its implementation, by testing for
runtime conditions which the verifier has no visibility into, and assert
that they are true. In case they are not, the program can simply throw
an exception from the other branch.
BPF exceptions are explicitly *NOT* an unlikely slowpath error handling
primitive, and this objective has guided design choices of the
implementation of the them within the kernel (with the bulk of the cost
for unwinding the stack offloaded to the bpf_throw kfunc).
The implementation of this mechanism requires use of add_hidden_subprog
mechanism introduced in the previous patch, which generates a couple of
instructions to move R1 to R0 and exit. The JIT then rewrites the
prologue of this subprog to take the stack pointer and frame pointer as
inputs and reset the stack frame, popping all callee-saved registers
saved by the main subprog. The bpf_throw function then walks the stack
at runtime, and invokes this exception subprog with the stack and frame
pointers as parameters.
Reviewers must take note that currently the main program is made to save
all callee-saved registers on x86_64 during entry into the program. This
is because we must do an equivalent of a lightweight context switch when
unwinding the stack, therefore we need the callee-saved registers of the
caller of the BPF program to be able to return with a sane state.
Note that we have to additionally handle r12, even though it is not used
by the program, because when throwing the exception the program makes an
entry into the kernel which could clobber r12 after saving it on the
stack. To be able to preserve the value we received on program entry, we
push r12 and restore it from the generated subprogram when unwinding the
stack.
For now, bpf_throw invocation fails when lingering resources or locks
exist in that path of the program. In a future followup, bpf_throw will
be extended to perform frame-by-frame unwinding to release lingering
resources for each stack frame, removing this limitation.
Signed-off-by: Kumar Kartikeya Dwivedi <memxor@gmail.com>
Link: https://lore.kernel.org/r/20230912233214.1518551-5-memxor@gmail.com
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
2023-09-12 23:32:01 +00:00
|
|
|
static bool all_callee_regs_used[4] = {true, true, true, true};
|
|
|
|
|
2014-12-05 01:01:24 +00:00
|
|
|
static u8 *emit_code(u8 *ptr, u32 bytes, unsigned int len)
|
2011-04-20 09:27:32 +00:00
|
|
|
{
|
|
|
|
if (len == 1)
|
|
|
|
*ptr = bytes;
|
|
|
|
else if (len == 2)
|
|
|
|
*(u16 *)ptr = bytes;
|
|
|
|
else {
|
|
|
|
*(u32 *)ptr = bytes;
|
|
|
|
barrier();
|
|
|
|
}
|
|
|
|
return ptr + len;
|
|
|
|
}
|
|
|
|
|
2015-05-19 23:59:04 +00:00
|
|
|
#define EMIT(bytes, len) \
|
2021-06-23 11:25:04 +00:00
|
|
|
do { prog = emit_code(prog, bytes, len); } while (0)
|
2011-04-20 09:27:32 +00:00
|
|
|
|
|
|
|
#define EMIT1(b1) EMIT(b1, 1)
|
|
|
|
#define EMIT2(b1, b2) EMIT((b1) + ((b2) << 8), 2)
|
|
|
|
#define EMIT3(b1, b2, b3) EMIT((b1) + ((b2) << 8) + ((b3) << 16), 3)
|
|
|
|
#define EMIT4(b1, b2, b3, b4) EMIT((b1) + ((b2) << 8) + ((b3) << 16) + ((b4) << 24), 4)
|
x86/bpf: Clean up non-standard comments, to make the code more readable
So by chance I looked into x86 assembly in arch/x86/net/bpf_jit_comp.c and
noticed the weird and inconsistent comment style it mistakenly learned from
the networking code:
/* Multi-line comment ...
* ... looks like this.
*/
Fix this to use the standard comment style specified in Documentation/CodingStyle
and used in arch/x86/ as well:
/*
* Multi-line comment ...
* ... looks like this.
*/
Also, to quote Linus's ... more explicit views about this:
http://article.gmane.org/gmane.linux.kernel.cryptoapi/21066
> But no, the networking code picked *none* of the above sane formats.
> Instead, it picked these two models that are just half-arsed
> shit-for-brains:
>
> (no)
> /* This is disgusting drug-induced
> * crap, and should die
> */
>
> (no-no-no)
> /* This is also very nasty
> * and visually unbalanced */
>
> Please. The networking code actually has the *worst* possible comment
> style. You can literally find that (no-no-no) style, which is just
> really horribly disgusting and worse than the otherwise fairly similar
> (d) in pretty much every way.
Also improve the comments and some other details while at it:
- Don't mix same-line and previous-line comment style on otherwise
identical code patterns within the same function,
- capitalize 'BPF' and x86 register names consistently,
- capitalize sentences consistently,
- instead of 'x64' use 'x86-64': x64 is a Microsoft specific term,
- use more consistent punctuation,
- use standard coding style in macros as well,
- fix typos and a few other minor details.
Consistent coding style is not optional, at least in arch/x86/.
No change in functionality.
( In case this commit causes conflicts with pending development code
I'll be glad to help resolve any conflicts! )
Acked-by: Thomas Gleixner <tglx@linutronix.de>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Borislav Petkov <bp@alien8.de>
Cc: H. Peter Anvin <hpa@zytor.com>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: David S. Miller <davem@davemloft.net>
Cc: Eric Dumazet <edumazet@google.com>
Cc: Daniel Borkmann <daniel@iogearbox.net>
Cc: Alexei Starovoitov <ast@fb.com>
Cc: Hideaki YOSHIFUJI <yoshfuji@linux-ipv6.org>
Cc: netdev@vger.kernel.org
Cc: linux-kernel@vger.kernel.org
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
2018-04-27 09:54:40 +00:00
|
|
|
|
net: filter: x86: internal BPF JIT
Maps all internal BPF instructions into x86_64 instructions.
This patch replaces original BPF x64 JIT with internal BPF x64 JIT.
sysctl net.core.bpf_jit_enable is reused as on/off switch.
Performance:
1. old BPF JIT and internal BPF JIT generate equivalent x86_64 code.
No performance difference is observed for filters that were JIT-able before
Example assembler code for BPF filter "tcpdump port 22"
original BPF -> old JIT: original BPF -> internal BPF -> new JIT:
0: push %rbp 0: push %rbp
1: mov %rsp,%rbp 1: mov %rsp,%rbp
4: sub $0x60,%rsp 4: sub $0x228,%rsp
8: mov %rbx,-0x8(%rbp) b: mov %rbx,-0x228(%rbp) // prologue
12: mov %r13,-0x220(%rbp)
19: mov %r14,-0x218(%rbp)
20: mov %r15,-0x210(%rbp)
27: xor %eax,%eax // clear A
c: xor %ebx,%ebx 29: xor %r13,%r13 // clear X
e: mov 0x68(%rdi),%r9d 2c: mov 0x68(%rdi),%r9d
12: sub 0x6c(%rdi),%r9d 30: sub 0x6c(%rdi),%r9d
16: mov 0xd8(%rdi),%r8 34: mov 0xd8(%rdi),%r10
3b: mov %rdi,%rbx
1d: mov $0xc,%esi 3e: mov $0xc,%esi
22: callq 0xffffffffe1021e15 43: callq 0xffffffffe102bd75
27: cmp $0x86dd,%eax 48: cmp $0x86dd,%rax
2c: jne 0x0000000000000069 4f: jne 0x000000000000009a
2e: mov $0x14,%esi 51: mov $0x14,%esi
33: callq 0xffffffffe1021e31 56: callq 0xffffffffe102bd91
38: cmp $0x84,%eax 5b: cmp $0x84,%rax
3d: je 0x0000000000000049 62: je 0x0000000000000074
3f: cmp $0x6,%eax 64: cmp $0x6,%rax
42: je 0x0000000000000049 68: je 0x0000000000000074
44: cmp $0x11,%eax 6a: cmp $0x11,%rax
47: jne 0x00000000000000c6 6e: jne 0x0000000000000117
49: mov $0x36,%esi 74: mov $0x36,%esi
4e: callq 0xffffffffe1021e15 79: callq 0xffffffffe102bd75
53: cmp $0x16,%eax 7e: cmp $0x16,%rax
56: je 0x00000000000000bf 82: je 0x0000000000000110
58: mov $0x38,%esi 88: mov $0x38,%esi
5d: callq 0xffffffffe1021e15 8d: callq 0xffffffffe102bd75
62: cmp $0x16,%eax 92: cmp $0x16,%rax
65: je 0x00000000000000bf 96: je 0x0000000000000110
67: jmp 0x00000000000000c6 98: jmp 0x0000000000000117
69: cmp $0x800,%eax 9a: cmp $0x800,%rax
6e: jne 0x00000000000000c6 a1: jne 0x0000000000000117
70: mov $0x17,%esi a3: mov $0x17,%esi
75: callq 0xffffffffe1021e31 a8: callq 0xffffffffe102bd91
7a: cmp $0x84,%eax ad: cmp $0x84,%rax
7f: je 0x000000000000008b b4: je 0x00000000000000c2
81: cmp $0x6,%eax b6: cmp $0x6,%rax
84: je 0x000000000000008b ba: je 0x00000000000000c2
86: cmp $0x11,%eax bc: cmp $0x11,%rax
89: jne 0x00000000000000c6 c0: jne 0x0000000000000117
8b: mov $0x14,%esi c2: mov $0x14,%esi
90: callq 0xffffffffe1021e15 c7: callq 0xffffffffe102bd75
95: test $0x1fff,%ax cc: test $0x1fff,%rax
99: jne 0x00000000000000c6 d3: jne 0x0000000000000117
d5: mov %rax,%r14
9b: mov $0xe,%esi d8: mov $0xe,%esi
a0: callq 0xffffffffe1021e44 dd: callq 0xffffffffe102bd91 // MSH
e2: and $0xf,%eax
e5: shl $0x2,%eax
e8: mov %rax,%r13
eb: mov %r14,%rax
ee: mov %r13,%rsi
a5: lea 0xe(%rbx),%esi f1: add $0xe,%esi
a8: callq 0xffffffffe1021e0d f4: callq 0xffffffffe102bd6d
ad: cmp $0x16,%eax f9: cmp $0x16,%rax
b0: je 0x00000000000000bf fd: je 0x0000000000000110
ff: mov %r13,%rsi
b2: lea 0x10(%rbx),%esi 102: add $0x10,%esi
b5: callq 0xffffffffe1021e0d 105: callq 0xffffffffe102bd6d
ba: cmp $0x16,%eax 10a: cmp $0x16,%rax
bd: jne 0x00000000000000c6 10e: jne 0x0000000000000117
bf: mov $0xffff,%eax 110: mov $0xffff,%eax
c4: jmp 0x00000000000000c8 115: jmp 0x000000000000011c
c6: xor %eax,%eax 117: mov $0x0,%eax
c8: mov -0x8(%rbp),%rbx 11c: mov -0x228(%rbp),%rbx // epilogue
cc: leaveq 123: mov -0x220(%rbp),%r13
cd: retq 12a: mov -0x218(%rbp),%r14
131: mov -0x210(%rbp),%r15
138: leaveq
139: retq
On fully cached SKBs both JITed functions take 12 nsec to execute.
BPF interpreter executes the program in 30 nsec.
The difference in generated assembler is due to the following:
Old BPF imlements LDX_MSH instruction via sk_load_byte_msh() helper function
inside bpf_jit.S.
New JIT removes the helper and does it explicitly, so ldx_msh cost
is the same for both JITs, but generated code looks longer.
New JIT has 4 registers to save, so prologue/epilogue are larger,
but the cost is within noise on x64.
Old JIT checks whether first insn clears A and if not emits 'xor %eax,%eax'.
New JIT clears %rax unconditionally.
2. old BPF JIT doesn't support ANC_NLATTR, ANC_PAY_OFFSET, ANC_RANDOM
extensions. New JIT supports all BPF extensions.
Performance of such filters improves 2-4 times depending on a filter.
The longer the filter the higher performance gain.
Synthetic benchmarks with many ancillary loads see 20x speedup
which seems to be the maximum gain from JIT
Notes:
. net.core.bpf_jit_enable=2 + tools/net/bpf_jit_disasm is still functional
and can be used to see generated assembler
. there are two jit_compile() functions and code flow for classic filters is:
sk_attach_filter() - load classic BPF
bpf_jit_compile() - try to JIT from classic BPF
sk_convert_filter() - convert classic to internal
bpf_int_jit_compile() - JIT from internal BPF
seccomp and tracing filters will just call bpf_int_jit_compile()
Signed-off-by: Alexei Starovoitov <ast@plumgrid.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
2014-05-14 02:50:46 +00:00
|
|
|
#define EMIT1_off32(b1, off) \
|
x86/bpf: Clean up non-standard comments, to make the code more readable
So by chance I looked into x86 assembly in arch/x86/net/bpf_jit_comp.c and
noticed the weird and inconsistent comment style it mistakenly learned from
the networking code:
/* Multi-line comment ...
* ... looks like this.
*/
Fix this to use the standard comment style specified in Documentation/CodingStyle
and used in arch/x86/ as well:
/*
* Multi-line comment ...
* ... looks like this.
*/
Also, to quote Linus's ... more explicit views about this:
http://article.gmane.org/gmane.linux.kernel.cryptoapi/21066
> But no, the networking code picked *none* of the above sane formats.
> Instead, it picked these two models that are just half-arsed
> shit-for-brains:
>
> (no)
> /* This is disgusting drug-induced
> * crap, and should die
> */
>
> (no-no-no)
> /* This is also very nasty
> * and visually unbalanced */
>
> Please. The networking code actually has the *worst* possible comment
> style. You can literally find that (no-no-no) style, which is just
> really horribly disgusting and worse than the otherwise fairly similar
> (d) in pretty much every way.
Also improve the comments and some other details while at it:
- Don't mix same-line and previous-line comment style on otherwise
identical code patterns within the same function,
- capitalize 'BPF' and x86 register names consistently,
- capitalize sentences consistently,
- instead of 'x64' use 'x86-64': x64 is a Microsoft specific term,
- use more consistent punctuation,
- use standard coding style in macros as well,
- fix typos and a few other minor details.
Consistent coding style is not optional, at least in arch/x86/.
No change in functionality.
( In case this commit causes conflicts with pending development code
I'll be glad to help resolve any conflicts! )
Acked-by: Thomas Gleixner <tglx@linutronix.de>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Borislav Petkov <bp@alien8.de>
Cc: H. Peter Anvin <hpa@zytor.com>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: David S. Miller <davem@davemloft.net>
Cc: Eric Dumazet <edumazet@google.com>
Cc: Daniel Borkmann <daniel@iogearbox.net>
Cc: Alexei Starovoitov <ast@fb.com>
Cc: Hideaki YOSHIFUJI <yoshfuji@linux-ipv6.org>
Cc: netdev@vger.kernel.org
Cc: linux-kernel@vger.kernel.org
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
2018-04-27 09:54:40 +00:00
|
|
|
do { EMIT1(b1); EMIT(off, 4); } while (0)
|
net: filter: x86: internal BPF JIT
Maps all internal BPF instructions into x86_64 instructions.
This patch replaces original BPF x64 JIT with internal BPF x64 JIT.
sysctl net.core.bpf_jit_enable is reused as on/off switch.
Performance:
1. old BPF JIT and internal BPF JIT generate equivalent x86_64 code.
No performance difference is observed for filters that were JIT-able before
Example assembler code for BPF filter "tcpdump port 22"
original BPF -> old JIT: original BPF -> internal BPF -> new JIT:
0: push %rbp 0: push %rbp
1: mov %rsp,%rbp 1: mov %rsp,%rbp
4: sub $0x60,%rsp 4: sub $0x228,%rsp
8: mov %rbx,-0x8(%rbp) b: mov %rbx,-0x228(%rbp) // prologue
12: mov %r13,-0x220(%rbp)
19: mov %r14,-0x218(%rbp)
20: mov %r15,-0x210(%rbp)
27: xor %eax,%eax // clear A
c: xor %ebx,%ebx 29: xor %r13,%r13 // clear X
e: mov 0x68(%rdi),%r9d 2c: mov 0x68(%rdi),%r9d
12: sub 0x6c(%rdi),%r9d 30: sub 0x6c(%rdi),%r9d
16: mov 0xd8(%rdi),%r8 34: mov 0xd8(%rdi),%r10
3b: mov %rdi,%rbx
1d: mov $0xc,%esi 3e: mov $0xc,%esi
22: callq 0xffffffffe1021e15 43: callq 0xffffffffe102bd75
27: cmp $0x86dd,%eax 48: cmp $0x86dd,%rax
2c: jne 0x0000000000000069 4f: jne 0x000000000000009a
2e: mov $0x14,%esi 51: mov $0x14,%esi
33: callq 0xffffffffe1021e31 56: callq 0xffffffffe102bd91
38: cmp $0x84,%eax 5b: cmp $0x84,%rax
3d: je 0x0000000000000049 62: je 0x0000000000000074
3f: cmp $0x6,%eax 64: cmp $0x6,%rax
42: je 0x0000000000000049 68: je 0x0000000000000074
44: cmp $0x11,%eax 6a: cmp $0x11,%rax
47: jne 0x00000000000000c6 6e: jne 0x0000000000000117
49: mov $0x36,%esi 74: mov $0x36,%esi
4e: callq 0xffffffffe1021e15 79: callq 0xffffffffe102bd75
53: cmp $0x16,%eax 7e: cmp $0x16,%rax
56: je 0x00000000000000bf 82: je 0x0000000000000110
58: mov $0x38,%esi 88: mov $0x38,%esi
5d: callq 0xffffffffe1021e15 8d: callq 0xffffffffe102bd75
62: cmp $0x16,%eax 92: cmp $0x16,%rax
65: je 0x00000000000000bf 96: je 0x0000000000000110
67: jmp 0x00000000000000c6 98: jmp 0x0000000000000117
69: cmp $0x800,%eax 9a: cmp $0x800,%rax
6e: jne 0x00000000000000c6 a1: jne 0x0000000000000117
70: mov $0x17,%esi a3: mov $0x17,%esi
75: callq 0xffffffffe1021e31 a8: callq 0xffffffffe102bd91
7a: cmp $0x84,%eax ad: cmp $0x84,%rax
7f: je 0x000000000000008b b4: je 0x00000000000000c2
81: cmp $0x6,%eax b6: cmp $0x6,%rax
84: je 0x000000000000008b ba: je 0x00000000000000c2
86: cmp $0x11,%eax bc: cmp $0x11,%rax
89: jne 0x00000000000000c6 c0: jne 0x0000000000000117
8b: mov $0x14,%esi c2: mov $0x14,%esi
90: callq 0xffffffffe1021e15 c7: callq 0xffffffffe102bd75
95: test $0x1fff,%ax cc: test $0x1fff,%rax
99: jne 0x00000000000000c6 d3: jne 0x0000000000000117
d5: mov %rax,%r14
9b: mov $0xe,%esi d8: mov $0xe,%esi
a0: callq 0xffffffffe1021e44 dd: callq 0xffffffffe102bd91 // MSH
e2: and $0xf,%eax
e5: shl $0x2,%eax
e8: mov %rax,%r13
eb: mov %r14,%rax
ee: mov %r13,%rsi
a5: lea 0xe(%rbx),%esi f1: add $0xe,%esi
a8: callq 0xffffffffe1021e0d f4: callq 0xffffffffe102bd6d
ad: cmp $0x16,%eax f9: cmp $0x16,%rax
b0: je 0x00000000000000bf fd: je 0x0000000000000110
ff: mov %r13,%rsi
b2: lea 0x10(%rbx),%esi 102: add $0x10,%esi
b5: callq 0xffffffffe1021e0d 105: callq 0xffffffffe102bd6d
ba: cmp $0x16,%eax 10a: cmp $0x16,%rax
bd: jne 0x00000000000000c6 10e: jne 0x0000000000000117
bf: mov $0xffff,%eax 110: mov $0xffff,%eax
c4: jmp 0x00000000000000c8 115: jmp 0x000000000000011c
c6: xor %eax,%eax 117: mov $0x0,%eax
c8: mov -0x8(%rbp),%rbx 11c: mov -0x228(%rbp),%rbx // epilogue
cc: leaveq 123: mov -0x220(%rbp),%r13
cd: retq 12a: mov -0x218(%rbp),%r14
131: mov -0x210(%rbp),%r15
138: leaveq
139: retq
On fully cached SKBs both JITed functions take 12 nsec to execute.
BPF interpreter executes the program in 30 nsec.
The difference in generated assembler is due to the following:
Old BPF imlements LDX_MSH instruction via sk_load_byte_msh() helper function
inside bpf_jit.S.
New JIT removes the helper and does it explicitly, so ldx_msh cost
is the same for both JITs, but generated code looks longer.
New JIT has 4 registers to save, so prologue/epilogue are larger,
but the cost is within noise on x64.
Old JIT checks whether first insn clears A and if not emits 'xor %eax,%eax'.
New JIT clears %rax unconditionally.
2. old BPF JIT doesn't support ANC_NLATTR, ANC_PAY_OFFSET, ANC_RANDOM
extensions. New JIT supports all BPF extensions.
Performance of such filters improves 2-4 times depending on a filter.
The longer the filter the higher performance gain.
Synthetic benchmarks with many ancillary loads see 20x speedup
which seems to be the maximum gain from JIT
Notes:
. net.core.bpf_jit_enable=2 + tools/net/bpf_jit_disasm is still functional
and can be used to see generated assembler
. there are two jit_compile() functions and code flow for classic filters is:
sk_attach_filter() - load classic BPF
bpf_jit_compile() - try to JIT from classic BPF
sk_convert_filter() - convert classic to internal
bpf_int_jit_compile() - JIT from internal BPF
seccomp and tracing filters will just call bpf_int_jit_compile()
Signed-off-by: Alexei Starovoitov <ast@plumgrid.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
2014-05-14 02:50:46 +00:00
|
|
|
#define EMIT2_off32(b1, b2, off) \
|
x86/bpf: Clean up non-standard comments, to make the code more readable
So by chance I looked into x86 assembly in arch/x86/net/bpf_jit_comp.c and
noticed the weird and inconsistent comment style it mistakenly learned from
the networking code:
/* Multi-line comment ...
* ... looks like this.
*/
Fix this to use the standard comment style specified in Documentation/CodingStyle
and used in arch/x86/ as well:
/*
* Multi-line comment ...
* ... looks like this.
*/
Also, to quote Linus's ... more explicit views about this:
http://article.gmane.org/gmane.linux.kernel.cryptoapi/21066
> But no, the networking code picked *none* of the above sane formats.
> Instead, it picked these two models that are just half-arsed
> shit-for-brains:
>
> (no)
> /* This is disgusting drug-induced
> * crap, and should die
> */
>
> (no-no-no)
> /* This is also very nasty
> * and visually unbalanced */
>
> Please. The networking code actually has the *worst* possible comment
> style. You can literally find that (no-no-no) style, which is just
> really horribly disgusting and worse than the otherwise fairly similar
> (d) in pretty much every way.
Also improve the comments and some other details while at it:
- Don't mix same-line and previous-line comment style on otherwise
identical code patterns within the same function,
- capitalize 'BPF' and x86 register names consistently,
- capitalize sentences consistently,
- instead of 'x64' use 'x86-64': x64 is a Microsoft specific term,
- use more consistent punctuation,
- use standard coding style in macros as well,
- fix typos and a few other minor details.
Consistent coding style is not optional, at least in arch/x86/.
No change in functionality.
( In case this commit causes conflicts with pending development code
I'll be glad to help resolve any conflicts! )
Acked-by: Thomas Gleixner <tglx@linutronix.de>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Borislav Petkov <bp@alien8.de>
Cc: H. Peter Anvin <hpa@zytor.com>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: David S. Miller <davem@davemloft.net>
Cc: Eric Dumazet <edumazet@google.com>
Cc: Daniel Borkmann <daniel@iogearbox.net>
Cc: Alexei Starovoitov <ast@fb.com>
Cc: Hideaki YOSHIFUJI <yoshfuji@linux-ipv6.org>
Cc: netdev@vger.kernel.org
Cc: linux-kernel@vger.kernel.org
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
2018-04-27 09:54:40 +00:00
|
|
|
do { EMIT2(b1, b2); EMIT(off, 4); } while (0)
|
net: filter: x86: internal BPF JIT
Maps all internal BPF instructions into x86_64 instructions.
This patch replaces original BPF x64 JIT with internal BPF x64 JIT.
sysctl net.core.bpf_jit_enable is reused as on/off switch.
Performance:
1. old BPF JIT and internal BPF JIT generate equivalent x86_64 code.
No performance difference is observed for filters that were JIT-able before
Example assembler code for BPF filter "tcpdump port 22"
original BPF -> old JIT: original BPF -> internal BPF -> new JIT:
0: push %rbp 0: push %rbp
1: mov %rsp,%rbp 1: mov %rsp,%rbp
4: sub $0x60,%rsp 4: sub $0x228,%rsp
8: mov %rbx,-0x8(%rbp) b: mov %rbx,-0x228(%rbp) // prologue
12: mov %r13,-0x220(%rbp)
19: mov %r14,-0x218(%rbp)
20: mov %r15,-0x210(%rbp)
27: xor %eax,%eax // clear A
c: xor %ebx,%ebx 29: xor %r13,%r13 // clear X
e: mov 0x68(%rdi),%r9d 2c: mov 0x68(%rdi),%r9d
12: sub 0x6c(%rdi),%r9d 30: sub 0x6c(%rdi),%r9d
16: mov 0xd8(%rdi),%r8 34: mov 0xd8(%rdi),%r10
3b: mov %rdi,%rbx
1d: mov $0xc,%esi 3e: mov $0xc,%esi
22: callq 0xffffffffe1021e15 43: callq 0xffffffffe102bd75
27: cmp $0x86dd,%eax 48: cmp $0x86dd,%rax
2c: jne 0x0000000000000069 4f: jne 0x000000000000009a
2e: mov $0x14,%esi 51: mov $0x14,%esi
33: callq 0xffffffffe1021e31 56: callq 0xffffffffe102bd91
38: cmp $0x84,%eax 5b: cmp $0x84,%rax
3d: je 0x0000000000000049 62: je 0x0000000000000074
3f: cmp $0x6,%eax 64: cmp $0x6,%rax
42: je 0x0000000000000049 68: je 0x0000000000000074
44: cmp $0x11,%eax 6a: cmp $0x11,%rax
47: jne 0x00000000000000c6 6e: jne 0x0000000000000117
49: mov $0x36,%esi 74: mov $0x36,%esi
4e: callq 0xffffffffe1021e15 79: callq 0xffffffffe102bd75
53: cmp $0x16,%eax 7e: cmp $0x16,%rax
56: je 0x00000000000000bf 82: je 0x0000000000000110
58: mov $0x38,%esi 88: mov $0x38,%esi
5d: callq 0xffffffffe1021e15 8d: callq 0xffffffffe102bd75
62: cmp $0x16,%eax 92: cmp $0x16,%rax
65: je 0x00000000000000bf 96: je 0x0000000000000110
67: jmp 0x00000000000000c6 98: jmp 0x0000000000000117
69: cmp $0x800,%eax 9a: cmp $0x800,%rax
6e: jne 0x00000000000000c6 a1: jne 0x0000000000000117
70: mov $0x17,%esi a3: mov $0x17,%esi
75: callq 0xffffffffe1021e31 a8: callq 0xffffffffe102bd91
7a: cmp $0x84,%eax ad: cmp $0x84,%rax
7f: je 0x000000000000008b b4: je 0x00000000000000c2
81: cmp $0x6,%eax b6: cmp $0x6,%rax
84: je 0x000000000000008b ba: je 0x00000000000000c2
86: cmp $0x11,%eax bc: cmp $0x11,%rax
89: jne 0x00000000000000c6 c0: jne 0x0000000000000117
8b: mov $0x14,%esi c2: mov $0x14,%esi
90: callq 0xffffffffe1021e15 c7: callq 0xffffffffe102bd75
95: test $0x1fff,%ax cc: test $0x1fff,%rax
99: jne 0x00000000000000c6 d3: jne 0x0000000000000117
d5: mov %rax,%r14
9b: mov $0xe,%esi d8: mov $0xe,%esi
a0: callq 0xffffffffe1021e44 dd: callq 0xffffffffe102bd91 // MSH
e2: and $0xf,%eax
e5: shl $0x2,%eax
e8: mov %rax,%r13
eb: mov %r14,%rax
ee: mov %r13,%rsi
a5: lea 0xe(%rbx),%esi f1: add $0xe,%esi
a8: callq 0xffffffffe1021e0d f4: callq 0xffffffffe102bd6d
ad: cmp $0x16,%eax f9: cmp $0x16,%rax
b0: je 0x00000000000000bf fd: je 0x0000000000000110
ff: mov %r13,%rsi
b2: lea 0x10(%rbx),%esi 102: add $0x10,%esi
b5: callq 0xffffffffe1021e0d 105: callq 0xffffffffe102bd6d
ba: cmp $0x16,%eax 10a: cmp $0x16,%rax
bd: jne 0x00000000000000c6 10e: jne 0x0000000000000117
bf: mov $0xffff,%eax 110: mov $0xffff,%eax
c4: jmp 0x00000000000000c8 115: jmp 0x000000000000011c
c6: xor %eax,%eax 117: mov $0x0,%eax
c8: mov -0x8(%rbp),%rbx 11c: mov -0x228(%rbp),%rbx // epilogue
cc: leaveq 123: mov -0x220(%rbp),%r13
cd: retq 12a: mov -0x218(%rbp),%r14
131: mov -0x210(%rbp),%r15
138: leaveq
139: retq
On fully cached SKBs both JITed functions take 12 nsec to execute.
BPF interpreter executes the program in 30 nsec.
The difference in generated assembler is due to the following:
Old BPF imlements LDX_MSH instruction via sk_load_byte_msh() helper function
inside bpf_jit.S.
New JIT removes the helper and does it explicitly, so ldx_msh cost
is the same for both JITs, but generated code looks longer.
New JIT has 4 registers to save, so prologue/epilogue are larger,
but the cost is within noise on x64.
Old JIT checks whether first insn clears A and if not emits 'xor %eax,%eax'.
New JIT clears %rax unconditionally.
2. old BPF JIT doesn't support ANC_NLATTR, ANC_PAY_OFFSET, ANC_RANDOM
extensions. New JIT supports all BPF extensions.
Performance of such filters improves 2-4 times depending on a filter.
The longer the filter the higher performance gain.
Synthetic benchmarks with many ancillary loads see 20x speedup
which seems to be the maximum gain from JIT
Notes:
. net.core.bpf_jit_enable=2 + tools/net/bpf_jit_disasm is still functional
and can be used to see generated assembler
. there are two jit_compile() functions and code flow for classic filters is:
sk_attach_filter() - load classic BPF
bpf_jit_compile() - try to JIT from classic BPF
sk_convert_filter() - convert classic to internal
bpf_int_jit_compile() - JIT from internal BPF
seccomp and tracing filters will just call bpf_int_jit_compile()
Signed-off-by: Alexei Starovoitov <ast@plumgrid.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
2014-05-14 02:50:46 +00:00
|
|
|
#define EMIT3_off32(b1, b2, b3, off) \
|
x86/bpf: Clean up non-standard comments, to make the code more readable
So by chance I looked into x86 assembly in arch/x86/net/bpf_jit_comp.c and
noticed the weird and inconsistent comment style it mistakenly learned from
the networking code:
/* Multi-line comment ...
* ... looks like this.
*/
Fix this to use the standard comment style specified in Documentation/CodingStyle
and used in arch/x86/ as well:
/*
* Multi-line comment ...
* ... looks like this.
*/
Also, to quote Linus's ... more explicit views about this:
http://article.gmane.org/gmane.linux.kernel.cryptoapi/21066
> But no, the networking code picked *none* of the above sane formats.
> Instead, it picked these two models that are just half-arsed
> shit-for-brains:
>
> (no)
> /* This is disgusting drug-induced
> * crap, and should die
> */
>
> (no-no-no)
> /* This is also very nasty
> * and visually unbalanced */
>
> Please. The networking code actually has the *worst* possible comment
> style. You can literally find that (no-no-no) style, which is just
> really horribly disgusting and worse than the otherwise fairly similar
> (d) in pretty much every way.
Also improve the comments and some other details while at it:
- Don't mix same-line and previous-line comment style on otherwise
identical code patterns within the same function,
- capitalize 'BPF' and x86 register names consistently,
- capitalize sentences consistently,
- instead of 'x64' use 'x86-64': x64 is a Microsoft specific term,
- use more consistent punctuation,
- use standard coding style in macros as well,
- fix typos and a few other minor details.
Consistent coding style is not optional, at least in arch/x86/.
No change in functionality.
( In case this commit causes conflicts with pending development code
I'll be glad to help resolve any conflicts! )
Acked-by: Thomas Gleixner <tglx@linutronix.de>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Borislav Petkov <bp@alien8.de>
Cc: H. Peter Anvin <hpa@zytor.com>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: David S. Miller <davem@davemloft.net>
Cc: Eric Dumazet <edumazet@google.com>
Cc: Daniel Borkmann <daniel@iogearbox.net>
Cc: Alexei Starovoitov <ast@fb.com>
Cc: Hideaki YOSHIFUJI <yoshfuji@linux-ipv6.org>
Cc: netdev@vger.kernel.org
Cc: linux-kernel@vger.kernel.org
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
2018-04-27 09:54:40 +00:00
|
|
|
do { EMIT3(b1, b2, b3); EMIT(off, 4); } while (0)
|
net: filter: x86: internal BPF JIT
Maps all internal BPF instructions into x86_64 instructions.
This patch replaces original BPF x64 JIT with internal BPF x64 JIT.
sysctl net.core.bpf_jit_enable is reused as on/off switch.
Performance:
1. old BPF JIT and internal BPF JIT generate equivalent x86_64 code.
No performance difference is observed for filters that were JIT-able before
Example assembler code for BPF filter "tcpdump port 22"
original BPF -> old JIT: original BPF -> internal BPF -> new JIT:
0: push %rbp 0: push %rbp
1: mov %rsp,%rbp 1: mov %rsp,%rbp
4: sub $0x60,%rsp 4: sub $0x228,%rsp
8: mov %rbx,-0x8(%rbp) b: mov %rbx,-0x228(%rbp) // prologue
12: mov %r13,-0x220(%rbp)
19: mov %r14,-0x218(%rbp)
20: mov %r15,-0x210(%rbp)
27: xor %eax,%eax // clear A
c: xor %ebx,%ebx 29: xor %r13,%r13 // clear X
e: mov 0x68(%rdi),%r9d 2c: mov 0x68(%rdi),%r9d
12: sub 0x6c(%rdi),%r9d 30: sub 0x6c(%rdi),%r9d
16: mov 0xd8(%rdi),%r8 34: mov 0xd8(%rdi),%r10
3b: mov %rdi,%rbx
1d: mov $0xc,%esi 3e: mov $0xc,%esi
22: callq 0xffffffffe1021e15 43: callq 0xffffffffe102bd75
27: cmp $0x86dd,%eax 48: cmp $0x86dd,%rax
2c: jne 0x0000000000000069 4f: jne 0x000000000000009a
2e: mov $0x14,%esi 51: mov $0x14,%esi
33: callq 0xffffffffe1021e31 56: callq 0xffffffffe102bd91
38: cmp $0x84,%eax 5b: cmp $0x84,%rax
3d: je 0x0000000000000049 62: je 0x0000000000000074
3f: cmp $0x6,%eax 64: cmp $0x6,%rax
42: je 0x0000000000000049 68: je 0x0000000000000074
44: cmp $0x11,%eax 6a: cmp $0x11,%rax
47: jne 0x00000000000000c6 6e: jne 0x0000000000000117
49: mov $0x36,%esi 74: mov $0x36,%esi
4e: callq 0xffffffffe1021e15 79: callq 0xffffffffe102bd75
53: cmp $0x16,%eax 7e: cmp $0x16,%rax
56: je 0x00000000000000bf 82: je 0x0000000000000110
58: mov $0x38,%esi 88: mov $0x38,%esi
5d: callq 0xffffffffe1021e15 8d: callq 0xffffffffe102bd75
62: cmp $0x16,%eax 92: cmp $0x16,%rax
65: je 0x00000000000000bf 96: je 0x0000000000000110
67: jmp 0x00000000000000c6 98: jmp 0x0000000000000117
69: cmp $0x800,%eax 9a: cmp $0x800,%rax
6e: jne 0x00000000000000c6 a1: jne 0x0000000000000117
70: mov $0x17,%esi a3: mov $0x17,%esi
75: callq 0xffffffffe1021e31 a8: callq 0xffffffffe102bd91
7a: cmp $0x84,%eax ad: cmp $0x84,%rax
7f: je 0x000000000000008b b4: je 0x00000000000000c2
81: cmp $0x6,%eax b6: cmp $0x6,%rax
84: je 0x000000000000008b ba: je 0x00000000000000c2
86: cmp $0x11,%eax bc: cmp $0x11,%rax
89: jne 0x00000000000000c6 c0: jne 0x0000000000000117
8b: mov $0x14,%esi c2: mov $0x14,%esi
90: callq 0xffffffffe1021e15 c7: callq 0xffffffffe102bd75
95: test $0x1fff,%ax cc: test $0x1fff,%rax
99: jne 0x00000000000000c6 d3: jne 0x0000000000000117
d5: mov %rax,%r14
9b: mov $0xe,%esi d8: mov $0xe,%esi
a0: callq 0xffffffffe1021e44 dd: callq 0xffffffffe102bd91 // MSH
e2: and $0xf,%eax
e5: shl $0x2,%eax
e8: mov %rax,%r13
eb: mov %r14,%rax
ee: mov %r13,%rsi
a5: lea 0xe(%rbx),%esi f1: add $0xe,%esi
a8: callq 0xffffffffe1021e0d f4: callq 0xffffffffe102bd6d
ad: cmp $0x16,%eax f9: cmp $0x16,%rax
b0: je 0x00000000000000bf fd: je 0x0000000000000110
ff: mov %r13,%rsi
b2: lea 0x10(%rbx),%esi 102: add $0x10,%esi
b5: callq 0xffffffffe1021e0d 105: callq 0xffffffffe102bd6d
ba: cmp $0x16,%eax 10a: cmp $0x16,%rax
bd: jne 0x00000000000000c6 10e: jne 0x0000000000000117
bf: mov $0xffff,%eax 110: mov $0xffff,%eax
c4: jmp 0x00000000000000c8 115: jmp 0x000000000000011c
c6: xor %eax,%eax 117: mov $0x0,%eax
c8: mov -0x8(%rbp),%rbx 11c: mov -0x228(%rbp),%rbx // epilogue
cc: leaveq 123: mov -0x220(%rbp),%r13
cd: retq 12a: mov -0x218(%rbp),%r14
131: mov -0x210(%rbp),%r15
138: leaveq
139: retq
On fully cached SKBs both JITed functions take 12 nsec to execute.
BPF interpreter executes the program in 30 nsec.
The difference in generated assembler is due to the following:
Old BPF imlements LDX_MSH instruction via sk_load_byte_msh() helper function
inside bpf_jit.S.
New JIT removes the helper and does it explicitly, so ldx_msh cost
is the same for both JITs, but generated code looks longer.
New JIT has 4 registers to save, so prologue/epilogue are larger,
but the cost is within noise on x64.
Old JIT checks whether first insn clears A and if not emits 'xor %eax,%eax'.
New JIT clears %rax unconditionally.
2. old BPF JIT doesn't support ANC_NLATTR, ANC_PAY_OFFSET, ANC_RANDOM
extensions. New JIT supports all BPF extensions.
Performance of such filters improves 2-4 times depending on a filter.
The longer the filter the higher performance gain.
Synthetic benchmarks with many ancillary loads see 20x speedup
which seems to be the maximum gain from JIT
Notes:
. net.core.bpf_jit_enable=2 + tools/net/bpf_jit_disasm is still functional
and can be used to see generated assembler
. there are two jit_compile() functions and code flow for classic filters is:
sk_attach_filter() - load classic BPF
bpf_jit_compile() - try to JIT from classic BPF
sk_convert_filter() - convert classic to internal
bpf_int_jit_compile() - JIT from internal BPF
seccomp and tracing filters will just call bpf_int_jit_compile()
Signed-off-by: Alexei Starovoitov <ast@plumgrid.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
2014-05-14 02:50:46 +00:00
|
|
|
#define EMIT4_off32(b1, b2, b3, b4, off) \
|
x86/bpf: Clean up non-standard comments, to make the code more readable
So by chance I looked into x86 assembly in arch/x86/net/bpf_jit_comp.c and
noticed the weird and inconsistent comment style it mistakenly learned from
the networking code:
/* Multi-line comment ...
* ... looks like this.
*/
Fix this to use the standard comment style specified in Documentation/CodingStyle
and used in arch/x86/ as well:
/*
* Multi-line comment ...
* ... looks like this.
*/
Also, to quote Linus's ... more explicit views about this:
http://article.gmane.org/gmane.linux.kernel.cryptoapi/21066
> But no, the networking code picked *none* of the above sane formats.
> Instead, it picked these two models that are just half-arsed
> shit-for-brains:
>
> (no)
> /* This is disgusting drug-induced
> * crap, and should die
> */
>
> (no-no-no)
> /* This is also very nasty
> * and visually unbalanced */
>
> Please. The networking code actually has the *worst* possible comment
> style. You can literally find that (no-no-no) style, which is just
> really horribly disgusting and worse than the otherwise fairly similar
> (d) in pretty much every way.
Also improve the comments and some other details while at it:
- Don't mix same-line and previous-line comment style on otherwise
identical code patterns within the same function,
- capitalize 'BPF' and x86 register names consistently,
- capitalize sentences consistently,
- instead of 'x64' use 'x86-64': x64 is a Microsoft specific term,
- use more consistent punctuation,
- use standard coding style in macros as well,
- fix typos and a few other minor details.
Consistent coding style is not optional, at least in arch/x86/.
No change in functionality.
( In case this commit causes conflicts with pending development code
I'll be glad to help resolve any conflicts! )
Acked-by: Thomas Gleixner <tglx@linutronix.de>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Borislav Petkov <bp@alien8.de>
Cc: H. Peter Anvin <hpa@zytor.com>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: David S. Miller <davem@davemloft.net>
Cc: Eric Dumazet <edumazet@google.com>
Cc: Daniel Borkmann <daniel@iogearbox.net>
Cc: Alexei Starovoitov <ast@fb.com>
Cc: Hideaki YOSHIFUJI <yoshfuji@linux-ipv6.org>
Cc: netdev@vger.kernel.org
Cc: linux-kernel@vger.kernel.org
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
2018-04-27 09:54:40 +00:00
|
|
|
do { EMIT4(b1, b2, b3, b4); EMIT(off, 4); } while (0)
|
2011-04-20 09:27:32 +00:00
|
|
|
|
2022-03-08 15:30:33 +00:00
|
|
|
#ifdef CONFIG_X86_KERNEL_IBT
|
2023-12-15 09:12:18 +00:00
|
|
|
#define EMIT_ENDBR() EMIT(gen_endbr(), 4)
|
|
|
|
#define EMIT_ENDBR_POISON() EMIT(gen_endbr_poison(), 4)
|
2022-03-08 15:30:33 +00:00
|
|
|
#else
|
|
|
|
#define EMIT_ENDBR()
|
2023-12-15 09:12:18 +00:00
|
|
|
#define EMIT_ENDBR_POISON()
|
2022-03-08 15:30:33 +00:00
|
|
|
#endif
|
|
|
|
|
2014-12-05 01:01:24 +00:00
|
|
|
static bool is_imm8(int value)
|
2011-04-20 09:27:32 +00:00
|
|
|
{
|
|
|
|
return value <= 127 && value >= -128;
|
|
|
|
}
|
|
|
|
|
2014-12-05 01:01:24 +00:00
|
|
|
static bool is_simm32(s64 value)
|
2011-04-20 09:27:32 +00:00
|
|
|
{
|
2018-02-24 00:07:59 +00:00
|
|
|
return value == (s64)(s32)value;
|
|
|
|
}
|
|
|
|
|
|
|
|
static bool is_uimm32(u64 value)
|
|
|
|
{
|
|
|
|
return value == (u64)(u32)value;
|
2011-04-20 09:27:32 +00:00
|
|
|
}
|
|
|
|
|
2014-06-06 21:46:06 +00:00
|
|
|
/* mov dst, src */
|
x86/bpf: Clean up non-standard comments, to make the code more readable
So by chance I looked into x86 assembly in arch/x86/net/bpf_jit_comp.c and
noticed the weird and inconsistent comment style it mistakenly learned from
the networking code:
/* Multi-line comment ...
* ... looks like this.
*/
Fix this to use the standard comment style specified in Documentation/CodingStyle
and used in arch/x86/ as well:
/*
* Multi-line comment ...
* ... looks like this.
*/
Also, to quote Linus's ... more explicit views about this:
http://article.gmane.org/gmane.linux.kernel.cryptoapi/21066
> But no, the networking code picked *none* of the above sane formats.
> Instead, it picked these two models that are just half-arsed
> shit-for-brains:
>
> (no)
> /* This is disgusting drug-induced
> * crap, and should die
> */
>
> (no-no-no)
> /* This is also very nasty
> * and visually unbalanced */
>
> Please. The networking code actually has the *worst* possible comment
> style. You can literally find that (no-no-no) style, which is just
> really horribly disgusting and worse than the otherwise fairly similar
> (d) in pretty much every way.
Also improve the comments and some other details while at it:
- Don't mix same-line and previous-line comment style on otherwise
identical code patterns within the same function,
- capitalize 'BPF' and x86 register names consistently,
- capitalize sentences consistently,
- instead of 'x64' use 'x86-64': x64 is a Microsoft specific term,
- use more consistent punctuation,
- use standard coding style in macros as well,
- fix typos and a few other minor details.
Consistent coding style is not optional, at least in arch/x86/.
No change in functionality.
( In case this commit causes conflicts with pending development code
I'll be glad to help resolve any conflicts! )
Acked-by: Thomas Gleixner <tglx@linutronix.de>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Borislav Petkov <bp@alien8.de>
Cc: H. Peter Anvin <hpa@zytor.com>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: David S. Miller <davem@davemloft.net>
Cc: Eric Dumazet <edumazet@google.com>
Cc: Daniel Borkmann <daniel@iogearbox.net>
Cc: Alexei Starovoitov <ast@fb.com>
Cc: Hideaki YOSHIFUJI <yoshfuji@linux-ipv6.org>
Cc: netdev@vger.kernel.org
Cc: linux-kernel@vger.kernel.org
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
2018-04-27 09:54:40 +00:00
|
|
|
#define EMIT_mov(DST, SRC) \
|
|
|
|
do { \
|
|
|
|
if (DST != SRC) \
|
|
|
|
EMIT3(add_2mod(0x48, DST, SRC), 0x89, add_2reg(0xC0, DST, SRC)); \
|
net: filter: x86: internal BPF JIT
Maps all internal BPF instructions into x86_64 instructions.
This patch replaces original BPF x64 JIT with internal BPF x64 JIT.
sysctl net.core.bpf_jit_enable is reused as on/off switch.
Performance:
1. old BPF JIT and internal BPF JIT generate equivalent x86_64 code.
No performance difference is observed for filters that were JIT-able before
Example assembler code for BPF filter "tcpdump port 22"
original BPF -> old JIT: original BPF -> internal BPF -> new JIT:
0: push %rbp 0: push %rbp
1: mov %rsp,%rbp 1: mov %rsp,%rbp
4: sub $0x60,%rsp 4: sub $0x228,%rsp
8: mov %rbx,-0x8(%rbp) b: mov %rbx,-0x228(%rbp) // prologue
12: mov %r13,-0x220(%rbp)
19: mov %r14,-0x218(%rbp)
20: mov %r15,-0x210(%rbp)
27: xor %eax,%eax // clear A
c: xor %ebx,%ebx 29: xor %r13,%r13 // clear X
e: mov 0x68(%rdi),%r9d 2c: mov 0x68(%rdi),%r9d
12: sub 0x6c(%rdi),%r9d 30: sub 0x6c(%rdi),%r9d
16: mov 0xd8(%rdi),%r8 34: mov 0xd8(%rdi),%r10
3b: mov %rdi,%rbx
1d: mov $0xc,%esi 3e: mov $0xc,%esi
22: callq 0xffffffffe1021e15 43: callq 0xffffffffe102bd75
27: cmp $0x86dd,%eax 48: cmp $0x86dd,%rax
2c: jne 0x0000000000000069 4f: jne 0x000000000000009a
2e: mov $0x14,%esi 51: mov $0x14,%esi
33: callq 0xffffffffe1021e31 56: callq 0xffffffffe102bd91
38: cmp $0x84,%eax 5b: cmp $0x84,%rax
3d: je 0x0000000000000049 62: je 0x0000000000000074
3f: cmp $0x6,%eax 64: cmp $0x6,%rax
42: je 0x0000000000000049 68: je 0x0000000000000074
44: cmp $0x11,%eax 6a: cmp $0x11,%rax
47: jne 0x00000000000000c6 6e: jne 0x0000000000000117
49: mov $0x36,%esi 74: mov $0x36,%esi
4e: callq 0xffffffffe1021e15 79: callq 0xffffffffe102bd75
53: cmp $0x16,%eax 7e: cmp $0x16,%rax
56: je 0x00000000000000bf 82: je 0x0000000000000110
58: mov $0x38,%esi 88: mov $0x38,%esi
5d: callq 0xffffffffe1021e15 8d: callq 0xffffffffe102bd75
62: cmp $0x16,%eax 92: cmp $0x16,%rax
65: je 0x00000000000000bf 96: je 0x0000000000000110
67: jmp 0x00000000000000c6 98: jmp 0x0000000000000117
69: cmp $0x800,%eax 9a: cmp $0x800,%rax
6e: jne 0x00000000000000c6 a1: jne 0x0000000000000117
70: mov $0x17,%esi a3: mov $0x17,%esi
75: callq 0xffffffffe1021e31 a8: callq 0xffffffffe102bd91
7a: cmp $0x84,%eax ad: cmp $0x84,%rax
7f: je 0x000000000000008b b4: je 0x00000000000000c2
81: cmp $0x6,%eax b6: cmp $0x6,%rax
84: je 0x000000000000008b ba: je 0x00000000000000c2
86: cmp $0x11,%eax bc: cmp $0x11,%rax
89: jne 0x00000000000000c6 c0: jne 0x0000000000000117
8b: mov $0x14,%esi c2: mov $0x14,%esi
90: callq 0xffffffffe1021e15 c7: callq 0xffffffffe102bd75
95: test $0x1fff,%ax cc: test $0x1fff,%rax
99: jne 0x00000000000000c6 d3: jne 0x0000000000000117
d5: mov %rax,%r14
9b: mov $0xe,%esi d8: mov $0xe,%esi
a0: callq 0xffffffffe1021e44 dd: callq 0xffffffffe102bd91 // MSH
e2: and $0xf,%eax
e5: shl $0x2,%eax
e8: mov %rax,%r13
eb: mov %r14,%rax
ee: mov %r13,%rsi
a5: lea 0xe(%rbx),%esi f1: add $0xe,%esi
a8: callq 0xffffffffe1021e0d f4: callq 0xffffffffe102bd6d
ad: cmp $0x16,%eax f9: cmp $0x16,%rax
b0: je 0x00000000000000bf fd: je 0x0000000000000110
ff: mov %r13,%rsi
b2: lea 0x10(%rbx),%esi 102: add $0x10,%esi
b5: callq 0xffffffffe1021e0d 105: callq 0xffffffffe102bd6d
ba: cmp $0x16,%eax 10a: cmp $0x16,%rax
bd: jne 0x00000000000000c6 10e: jne 0x0000000000000117
bf: mov $0xffff,%eax 110: mov $0xffff,%eax
c4: jmp 0x00000000000000c8 115: jmp 0x000000000000011c
c6: xor %eax,%eax 117: mov $0x0,%eax
c8: mov -0x8(%rbp),%rbx 11c: mov -0x228(%rbp),%rbx // epilogue
cc: leaveq 123: mov -0x220(%rbp),%r13
cd: retq 12a: mov -0x218(%rbp),%r14
131: mov -0x210(%rbp),%r15
138: leaveq
139: retq
On fully cached SKBs both JITed functions take 12 nsec to execute.
BPF interpreter executes the program in 30 nsec.
The difference in generated assembler is due to the following:
Old BPF imlements LDX_MSH instruction via sk_load_byte_msh() helper function
inside bpf_jit.S.
New JIT removes the helper and does it explicitly, so ldx_msh cost
is the same for both JITs, but generated code looks longer.
New JIT has 4 registers to save, so prologue/epilogue are larger,
but the cost is within noise on x64.
Old JIT checks whether first insn clears A and if not emits 'xor %eax,%eax'.
New JIT clears %rax unconditionally.
2. old BPF JIT doesn't support ANC_NLATTR, ANC_PAY_OFFSET, ANC_RANDOM
extensions. New JIT supports all BPF extensions.
Performance of such filters improves 2-4 times depending on a filter.
The longer the filter the higher performance gain.
Synthetic benchmarks with many ancillary loads see 20x speedup
which seems to be the maximum gain from JIT
Notes:
. net.core.bpf_jit_enable=2 + tools/net/bpf_jit_disasm is still functional
and can be used to see generated assembler
. there are two jit_compile() functions and code flow for classic filters is:
sk_attach_filter() - load classic BPF
bpf_jit_compile() - try to JIT from classic BPF
sk_convert_filter() - convert classic to internal
bpf_int_jit_compile() - JIT from internal BPF
seccomp and tracing filters will just call bpf_int_jit_compile()
Signed-off-by: Alexei Starovoitov <ast@plumgrid.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
2014-05-14 02:50:46 +00:00
|
|
|
} while (0)
|
|
|
|
|
|
|
|
static int bpf_size_to_x86_bytes(int bpf_size)
|
|
|
|
{
|
|
|
|
if (bpf_size == BPF_W)
|
|
|
|
return 4;
|
|
|
|
else if (bpf_size == BPF_H)
|
|
|
|
return 2;
|
|
|
|
else if (bpf_size == BPF_B)
|
|
|
|
return 1;
|
|
|
|
else if (bpf_size == BPF_DW)
|
|
|
|
return 4; /* imm32 */
|
|
|
|
else
|
|
|
|
return 0;
|
|
|
|
}
|
2011-04-20 09:27:32 +00:00
|
|
|
|
x86/bpf: Clean up non-standard comments, to make the code more readable
So by chance I looked into x86 assembly in arch/x86/net/bpf_jit_comp.c and
noticed the weird and inconsistent comment style it mistakenly learned from
the networking code:
/* Multi-line comment ...
* ... looks like this.
*/
Fix this to use the standard comment style specified in Documentation/CodingStyle
and used in arch/x86/ as well:
/*
* Multi-line comment ...
* ... looks like this.
*/
Also, to quote Linus's ... more explicit views about this:
http://article.gmane.org/gmane.linux.kernel.cryptoapi/21066
> But no, the networking code picked *none* of the above sane formats.
> Instead, it picked these two models that are just half-arsed
> shit-for-brains:
>
> (no)
> /* This is disgusting drug-induced
> * crap, and should die
> */
>
> (no-no-no)
> /* This is also very nasty
> * and visually unbalanced */
>
> Please. The networking code actually has the *worst* possible comment
> style. You can literally find that (no-no-no) style, which is just
> really horribly disgusting and worse than the otherwise fairly similar
> (d) in pretty much every way.
Also improve the comments and some other details while at it:
- Don't mix same-line and previous-line comment style on otherwise
identical code patterns within the same function,
- capitalize 'BPF' and x86 register names consistently,
- capitalize sentences consistently,
- instead of 'x64' use 'x86-64': x64 is a Microsoft specific term,
- use more consistent punctuation,
- use standard coding style in macros as well,
- fix typos and a few other minor details.
Consistent coding style is not optional, at least in arch/x86/.
No change in functionality.
( In case this commit causes conflicts with pending development code
I'll be glad to help resolve any conflicts! )
Acked-by: Thomas Gleixner <tglx@linutronix.de>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Borislav Petkov <bp@alien8.de>
Cc: H. Peter Anvin <hpa@zytor.com>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: David S. Miller <davem@davemloft.net>
Cc: Eric Dumazet <edumazet@google.com>
Cc: Daniel Borkmann <daniel@iogearbox.net>
Cc: Alexei Starovoitov <ast@fb.com>
Cc: Hideaki YOSHIFUJI <yoshfuji@linux-ipv6.org>
Cc: netdev@vger.kernel.org
Cc: linux-kernel@vger.kernel.org
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
2018-04-27 09:54:40 +00:00
|
|
|
/*
|
|
|
|
* List of x86 cond jumps opcodes (. + s8)
|
2011-04-20 09:27:32 +00:00
|
|
|
* Add 0x10 (and an extra 0x0f) to generate far jumps (. + s32)
|
|
|
|
*/
|
|
|
|
#define X86_JB 0x72
|
|
|
|
#define X86_JAE 0x73
|
|
|
|
#define X86_JE 0x74
|
|
|
|
#define X86_JNE 0x75
|
|
|
|
#define X86_JBE 0x76
|
|
|
|
#define X86_JA 0x77
|
2017-08-09 23:39:56 +00:00
|
|
|
#define X86_JL 0x7C
|
net: filter: x86: internal BPF JIT
Maps all internal BPF instructions into x86_64 instructions.
This patch replaces original BPF x64 JIT with internal BPF x64 JIT.
sysctl net.core.bpf_jit_enable is reused as on/off switch.
Performance:
1. old BPF JIT and internal BPF JIT generate equivalent x86_64 code.
No performance difference is observed for filters that were JIT-able before
Example assembler code for BPF filter "tcpdump port 22"
original BPF -> old JIT: original BPF -> internal BPF -> new JIT:
0: push %rbp 0: push %rbp
1: mov %rsp,%rbp 1: mov %rsp,%rbp
4: sub $0x60,%rsp 4: sub $0x228,%rsp
8: mov %rbx,-0x8(%rbp) b: mov %rbx,-0x228(%rbp) // prologue
12: mov %r13,-0x220(%rbp)
19: mov %r14,-0x218(%rbp)
20: mov %r15,-0x210(%rbp)
27: xor %eax,%eax // clear A
c: xor %ebx,%ebx 29: xor %r13,%r13 // clear X
e: mov 0x68(%rdi),%r9d 2c: mov 0x68(%rdi),%r9d
12: sub 0x6c(%rdi),%r9d 30: sub 0x6c(%rdi),%r9d
16: mov 0xd8(%rdi),%r8 34: mov 0xd8(%rdi),%r10
3b: mov %rdi,%rbx
1d: mov $0xc,%esi 3e: mov $0xc,%esi
22: callq 0xffffffffe1021e15 43: callq 0xffffffffe102bd75
27: cmp $0x86dd,%eax 48: cmp $0x86dd,%rax
2c: jne 0x0000000000000069 4f: jne 0x000000000000009a
2e: mov $0x14,%esi 51: mov $0x14,%esi
33: callq 0xffffffffe1021e31 56: callq 0xffffffffe102bd91
38: cmp $0x84,%eax 5b: cmp $0x84,%rax
3d: je 0x0000000000000049 62: je 0x0000000000000074
3f: cmp $0x6,%eax 64: cmp $0x6,%rax
42: je 0x0000000000000049 68: je 0x0000000000000074
44: cmp $0x11,%eax 6a: cmp $0x11,%rax
47: jne 0x00000000000000c6 6e: jne 0x0000000000000117
49: mov $0x36,%esi 74: mov $0x36,%esi
4e: callq 0xffffffffe1021e15 79: callq 0xffffffffe102bd75
53: cmp $0x16,%eax 7e: cmp $0x16,%rax
56: je 0x00000000000000bf 82: je 0x0000000000000110
58: mov $0x38,%esi 88: mov $0x38,%esi
5d: callq 0xffffffffe1021e15 8d: callq 0xffffffffe102bd75
62: cmp $0x16,%eax 92: cmp $0x16,%rax
65: je 0x00000000000000bf 96: je 0x0000000000000110
67: jmp 0x00000000000000c6 98: jmp 0x0000000000000117
69: cmp $0x800,%eax 9a: cmp $0x800,%rax
6e: jne 0x00000000000000c6 a1: jne 0x0000000000000117
70: mov $0x17,%esi a3: mov $0x17,%esi
75: callq 0xffffffffe1021e31 a8: callq 0xffffffffe102bd91
7a: cmp $0x84,%eax ad: cmp $0x84,%rax
7f: je 0x000000000000008b b4: je 0x00000000000000c2
81: cmp $0x6,%eax b6: cmp $0x6,%rax
84: je 0x000000000000008b ba: je 0x00000000000000c2
86: cmp $0x11,%eax bc: cmp $0x11,%rax
89: jne 0x00000000000000c6 c0: jne 0x0000000000000117
8b: mov $0x14,%esi c2: mov $0x14,%esi
90: callq 0xffffffffe1021e15 c7: callq 0xffffffffe102bd75
95: test $0x1fff,%ax cc: test $0x1fff,%rax
99: jne 0x00000000000000c6 d3: jne 0x0000000000000117
d5: mov %rax,%r14
9b: mov $0xe,%esi d8: mov $0xe,%esi
a0: callq 0xffffffffe1021e44 dd: callq 0xffffffffe102bd91 // MSH
e2: and $0xf,%eax
e5: shl $0x2,%eax
e8: mov %rax,%r13
eb: mov %r14,%rax
ee: mov %r13,%rsi
a5: lea 0xe(%rbx),%esi f1: add $0xe,%esi
a8: callq 0xffffffffe1021e0d f4: callq 0xffffffffe102bd6d
ad: cmp $0x16,%eax f9: cmp $0x16,%rax
b0: je 0x00000000000000bf fd: je 0x0000000000000110
ff: mov %r13,%rsi
b2: lea 0x10(%rbx),%esi 102: add $0x10,%esi
b5: callq 0xffffffffe1021e0d 105: callq 0xffffffffe102bd6d
ba: cmp $0x16,%eax 10a: cmp $0x16,%rax
bd: jne 0x00000000000000c6 10e: jne 0x0000000000000117
bf: mov $0xffff,%eax 110: mov $0xffff,%eax
c4: jmp 0x00000000000000c8 115: jmp 0x000000000000011c
c6: xor %eax,%eax 117: mov $0x0,%eax
c8: mov -0x8(%rbp),%rbx 11c: mov -0x228(%rbp),%rbx // epilogue
cc: leaveq 123: mov -0x220(%rbp),%r13
cd: retq 12a: mov -0x218(%rbp),%r14
131: mov -0x210(%rbp),%r15
138: leaveq
139: retq
On fully cached SKBs both JITed functions take 12 nsec to execute.
BPF interpreter executes the program in 30 nsec.
The difference in generated assembler is due to the following:
Old BPF imlements LDX_MSH instruction via sk_load_byte_msh() helper function
inside bpf_jit.S.
New JIT removes the helper and does it explicitly, so ldx_msh cost
is the same for both JITs, but generated code looks longer.
New JIT has 4 registers to save, so prologue/epilogue are larger,
but the cost is within noise on x64.
Old JIT checks whether first insn clears A and if not emits 'xor %eax,%eax'.
New JIT clears %rax unconditionally.
2. old BPF JIT doesn't support ANC_NLATTR, ANC_PAY_OFFSET, ANC_RANDOM
extensions. New JIT supports all BPF extensions.
Performance of such filters improves 2-4 times depending on a filter.
The longer the filter the higher performance gain.
Synthetic benchmarks with many ancillary loads see 20x speedup
which seems to be the maximum gain from JIT
Notes:
. net.core.bpf_jit_enable=2 + tools/net/bpf_jit_disasm is still functional
and can be used to see generated assembler
. there are two jit_compile() functions and code flow for classic filters is:
sk_attach_filter() - load classic BPF
bpf_jit_compile() - try to JIT from classic BPF
sk_convert_filter() - convert classic to internal
bpf_int_jit_compile() - JIT from internal BPF
seccomp and tracing filters will just call bpf_int_jit_compile()
Signed-off-by: Alexei Starovoitov <ast@plumgrid.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
2014-05-14 02:50:46 +00:00
|
|
|
#define X86_JGE 0x7D
|
2017-08-09 23:39:56 +00:00
|
|
|
#define X86_JLE 0x7E
|
net: filter: x86: internal BPF JIT
Maps all internal BPF instructions into x86_64 instructions.
This patch replaces original BPF x64 JIT with internal BPF x64 JIT.
sysctl net.core.bpf_jit_enable is reused as on/off switch.
Performance:
1. old BPF JIT and internal BPF JIT generate equivalent x86_64 code.
No performance difference is observed for filters that were JIT-able before
Example assembler code for BPF filter "tcpdump port 22"
original BPF -> old JIT: original BPF -> internal BPF -> new JIT:
0: push %rbp 0: push %rbp
1: mov %rsp,%rbp 1: mov %rsp,%rbp
4: sub $0x60,%rsp 4: sub $0x228,%rsp
8: mov %rbx,-0x8(%rbp) b: mov %rbx,-0x228(%rbp) // prologue
12: mov %r13,-0x220(%rbp)
19: mov %r14,-0x218(%rbp)
20: mov %r15,-0x210(%rbp)
27: xor %eax,%eax // clear A
c: xor %ebx,%ebx 29: xor %r13,%r13 // clear X
e: mov 0x68(%rdi),%r9d 2c: mov 0x68(%rdi),%r9d
12: sub 0x6c(%rdi),%r9d 30: sub 0x6c(%rdi),%r9d
16: mov 0xd8(%rdi),%r8 34: mov 0xd8(%rdi),%r10
3b: mov %rdi,%rbx
1d: mov $0xc,%esi 3e: mov $0xc,%esi
22: callq 0xffffffffe1021e15 43: callq 0xffffffffe102bd75
27: cmp $0x86dd,%eax 48: cmp $0x86dd,%rax
2c: jne 0x0000000000000069 4f: jne 0x000000000000009a
2e: mov $0x14,%esi 51: mov $0x14,%esi
33: callq 0xffffffffe1021e31 56: callq 0xffffffffe102bd91
38: cmp $0x84,%eax 5b: cmp $0x84,%rax
3d: je 0x0000000000000049 62: je 0x0000000000000074
3f: cmp $0x6,%eax 64: cmp $0x6,%rax
42: je 0x0000000000000049 68: je 0x0000000000000074
44: cmp $0x11,%eax 6a: cmp $0x11,%rax
47: jne 0x00000000000000c6 6e: jne 0x0000000000000117
49: mov $0x36,%esi 74: mov $0x36,%esi
4e: callq 0xffffffffe1021e15 79: callq 0xffffffffe102bd75
53: cmp $0x16,%eax 7e: cmp $0x16,%rax
56: je 0x00000000000000bf 82: je 0x0000000000000110
58: mov $0x38,%esi 88: mov $0x38,%esi
5d: callq 0xffffffffe1021e15 8d: callq 0xffffffffe102bd75
62: cmp $0x16,%eax 92: cmp $0x16,%rax
65: je 0x00000000000000bf 96: je 0x0000000000000110
67: jmp 0x00000000000000c6 98: jmp 0x0000000000000117
69: cmp $0x800,%eax 9a: cmp $0x800,%rax
6e: jne 0x00000000000000c6 a1: jne 0x0000000000000117
70: mov $0x17,%esi a3: mov $0x17,%esi
75: callq 0xffffffffe1021e31 a8: callq 0xffffffffe102bd91
7a: cmp $0x84,%eax ad: cmp $0x84,%rax
7f: je 0x000000000000008b b4: je 0x00000000000000c2
81: cmp $0x6,%eax b6: cmp $0x6,%rax
84: je 0x000000000000008b ba: je 0x00000000000000c2
86: cmp $0x11,%eax bc: cmp $0x11,%rax
89: jne 0x00000000000000c6 c0: jne 0x0000000000000117
8b: mov $0x14,%esi c2: mov $0x14,%esi
90: callq 0xffffffffe1021e15 c7: callq 0xffffffffe102bd75
95: test $0x1fff,%ax cc: test $0x1fff,%rax
99: jne 0x00000000000000c6 d3: jne 0x0000000000000117
d5: mov %rax,%r14
9b: mov $0xe,%esi d8: mov $0xe,%esi
a0: callq 0xffffffffe1021e44 dd: callq 0xffffffffe102bd91 // MSH
e2: and $0xf,%eax
e5: shl $0x2,%eax
e8: mov %rax,%r13
eb: mov %r14,%rax
ee: mov %r13,%rsi
a5: lea 0xe(%rbx),%esi f1: add $0xe,%esi
a8: callq 0xffffffffe1021e0d f4: callq 0xffffffffe102bd6d
ad: cmp $0x16,%eax f9: cmp $0x16,%rax
b0: je 0x00000000000000bf fd: je 0x0000000000000110
ff: mov %r13,%rsi
b2: lea 0x10(%rbx),%esi 102: add $0x10,%esi
b5: callq 0xffffffffe1021e0d 105: callq 0xffffffffe102bd6d
ba: cmp $0x16,%eax 10a: cmp $0x16,%rax
bd: jne 0x00000000000000c6 10e: jne 0x0000000000000117
bf: mov $0xffff,%eax 110: mov $0xffff,%eax
c4: jmp 0x00000000000000c8 115: jmp 0x000000000000011c
c6: xor %eax,%eax 117: mov $0x0,%eax
c8: mov -0x8(%rbp),%rbx 11c: mov -0x228(%rbp),%rbx // epilogue
cc: leaveq 123: mov -0x220(%rbp),%r13
cd: retq 12a: mov -0x218(%rbp),%r14
131: mov -0x210(%rbp),%r15
138: leaveq
139: retq
On fully cached SKBs both JITed functions take 12 nsec to execute.
BPF interpreter executes the program in 30 nsec.
The difference in generated assembler is due to the following:
Old BPF imlements LDX_MSH instruction via sk_load_byte_msh() helper function
inside bpf_jit.S.
New JIT removes the helper and does it explicitly, so ldx_msh cost
is the same for both JITs, but generated code looks longer.
New JIT has 4 registers to save, so prologue/epilogue are larger,
but the cost is within noise on x64.
Old JIT checks whether first insn clears A and if not emits 'xor %eax,%eax'.
New JIT clears %rax unconditionally.
2. old BPF JIT doesn't support ANC_NLATTR, ANC_PAY_OFFSET, ANC_RANDOM
extensions. New JIT supports all BPF extensions.
Performance of such filters improves 2-4 times depending on a filter.
The longer the filter the higher performance gain.
Synthetic benchmarks with many ancillary loads see 20x speedup
which seems to be the maximum gain from JIT
Notes:
. net.core.bpf_jit_enable=2 + tools/net/bpf_jit_disasm is still functional
and can be used to see generated assembler
. there are two jit_compile() functions and code flow for classic filters is:
sk_attach_filter() - load classic BPF
bpf_jit_compile() - try to JIT from classic BPF
sk_convert_filter() - convert classic to internal
bpf_int_jit_compile() - JIT from internal BPF
seccomp and tracing filters will just call bpf_int_jit_compile()
Signed-off-by: Alexei Starovoitov <ast@plumgrid.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
2014-05-14 02:50:46 +00:00
|
|
|
#define X86_JG 0x7F
|
2011-04-20 09:27:32 +00:00
|
|
|
|
x86/bpf: Clean up non-standard comments, to make the code more readable
So by chance I looked into x86 assembly in arch/x86/net/bpf_jit_comp.c and
noticed the weird and inconsistent comment style it mistakenly learned from
the networking code:
/* Multi-line comment ...
* ... looks like this.
*/
Fix this to use the standard comment style specified in Documentation/CodingStyle
and used in arch/x86/ as well:
/*
* Multi-line comment ...
* ... looks like this.
*/
Also, to quote Linus's ... more explicit views about this:
http://article.gmane.org/gmane.linux.kernel.cryptoapi/21066
> But no, the networking code picked *none* of the above sane formats.
> Instead, it picked these two models that are just half-arsed
> shit-for-brains:
>
> (no)
> /* This is disgusting drug-induced
> * crap, and should die
> */
>
> (no-no-no)
> /* This is also very nasty
> * and visually unbalanced */
>
> Please. The networking code actually has the *worst* possible comment
> style. You can literally find that (no-no-no) style, which is just
> really horribly disgusting and worse than the otherwise fairly similar
> (d) in pretty much every way.
Also improve the comments and some other details while at it:
- Don't mix same-line and previous-line comment style on otherwise
identical code patterns within the same function,
- capitalize 'BPF' and x86 register names consistently,
- capitalize sentences consistently,
- instead of 'x64' use 'x86-64': x64 is a Microsoft specific term,
- use more consistent punctuation,
- use standard coding style in macros as well,
- fix typos and a few other minor details.
Consistent coding style is not optional, at least in arch/x86/.
No change in functionality.
( In case this commit causes conflicts with pending development code
I'll be glad to help resolve any conflicts! )
Acked-by: Thomas Gleixner <tglx@linutronix.de>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Borislav Petkov <bp@alien8.de>
Cc: H. Peter Anvin <hpa@zytor.com>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: David S. Miller <davem@davemloft.net>
Cc: Eric Dumazet <edumazet@google.com>
Cc: Daniel Borkmann <daniel@iogearbox.net>
Cc: Alexei Starovoitov <ast@fb.com>
Cc: Hideaki YOSHIFUJI <yoshfuji@linux-ipv6.org>
Cc: netdev@vger.kernel.org
Cc: linux-kernel@vger.kernel.org
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
2018-04-27 09:54:40 +00:00
|
|
|
/* Pick a register outside of BPF range for JIT internal work */
|
2016-05-13 17:08:33 +00:00
|
|
|
#define AUX_REG (MAX_BPF_JIT_REG + 1)
|
2019-11-14 18:57:04 +00:00
|
|
|
#define X86_REG_R9 (MAX_BPF_JIT_REG + 2)
|
2024-03-08 01:08:01 +00:00
|
|
|
#define X86_REG_R12 (MAX_BPF_JIT_REG + 3)
|
net: filter: x86: internal BPF JIT
Maps all internal BPF instructions into x86_64 instructions.
This patch replaces original BPF x64 JIT with internal BPF x64 JIT.
sysctl net.core.bpf_jit_enable is reused as on/off switch.
Performance:
1. old BPF JIT and internal BPF JIT generate equivalent x86_64 code.
No performance difference is observed for filters that were JIT-able before
Example assembler code for BPF filter "tcpdump port 22"
original BPF -> old JIT: original BPF -> internal BPF -> new JIT:
0: push %rbp 0: push %rbp
1: mov %rsp,%rbp 1: mov %rsp,%rbp
4: sub $0x60,%rsp 4: sub $0x228,%rsp
8: mov %rbx,-0x8(%rbp) b: mov %rbx,-0x228(%rbp) // prologue
12: mov %r13,-0x220(%rbp)
19: mov %r14,-0x218(%rbp)
20: mov %r15,-0x210(%rbp)
27: xor %eax,%eax // clear A
c: xor %ebx,%ebx 29: xor %r13,%r13 // clear X
e: mov 0x68(%rdi),%r9d 2c: mov 0x68(%rdi),%r9d
12: sub 0x6c(%rdi),%r9d 30: sub 0x6c(%rdi),%r9d
16: mov 0xd8(%rdi),%r8 34: mov 0xd8(%rdi),%r10
3b: mov %rdi,%rbx
1d: mov $0xc,%esi 3e: mov $0xc,%esi
22: callq 0xffffffffe1021e15 43: callq 0xffffffffe102bd75
27: cmp $0x86dd,%eax 48: cmp $0x86dd,%rax
2c: jne 0x0000000000000069 4f: jne 0x000000000000009a
2e: mov $0x14,%esi 51: mov $0x14,%esi
33: callq 0xffffffffe1021e31 56: callq 0xffffffffe102bd91
38: cmp $0x84,%eax 5b: cmp $0x84,%rax
3d: je 0x0000000000000049 62: je 0x0000000000000074
3f: cmp $0x6,%eax 64: cmp $0x6,%rax
42: je 0x0000000000000049 68: je 0x0000000000000074
44: cmp $0x11,%eax 6a: cmp $0x11,%rax
47: jne 0x00000000000000c6 6e: jne 0x0000000000000117
49: mov $0x36,%esi 74: mov $0x36,%esi
4e: callq 0xffffffffe1021e15 79: callq 0xffffffffe102bd75
53: cmp $0x16,%eax 7e: cmp $0x16,%rax
56: je 0x00000000000000bf 82: je 0x0000000000000110
58: mov $0x38,%esi 88: mov $0x38,%esi
5d: callq 0xffffffffe1021e15 8d: callq 0xffffffffe102bd75
62: cmp $0x16,%eax 92: cmp $0x16,%rax
65: je 0x00000000000000bf 96: je 0x0000000000000110
67: jmp 0x00000000000000c6 98: jmp 0x0000000000000117
69: cmp $0x800,%eax 9a: cmp $0x800,%rax
6e: jne 0x00000000000000c6 a1: jne 0x0000000000000117
70: mov $0x17,%esi a3: mov $0x17,%esi
75: callq 0xffffffffe1021e31 a8: callq 0xffffffffe102bd91
7a: cmp $0x84,%eax ad: cmp $0x84,%rax
7f: je 0x000000000000008b b4: je 0x00000000000000c2
81: cmp $0x6,%eax b6: cmp $0x6,%rax
84: je 0x000000000000008b ba: je 0x00000000000000c2
86: cmp $0x11,%eax bc: cmp $0x11,%rax
89: jne 0x00000000000000c6 c0: jne 0x0000000000000117
8b: mov $0x14,%esi c2: mov $0x14,%esi
90: callq 0xffffffffe1021e15 c7: callq 0xffffffffe102bd75
95: test $0x1fff,%ax cc: test $0x1fff,%rax
99: jne 0x00000000000000c6 d3: jne 0x0000000000000117
d5: mov %rax,%r14
9b: mov $0xe,%esi d8: mov $0xe,%esi
a0: callq 0xffffffffe1021e44 dd: callq 0xffffffffe102bd91 // MSH
e2: and $0xf,%eax
e5: shl $0x2,%eax
e8: mov %rax,%r13
eb: mov %r14,%rax
ee: mov %r13,%rsi
a5: lea 0xe(%rbx),%esi f1: add $0xe,%esi
a8: callq 0xffffffffe1021e0d f4: callq 0xffffffffe102bd6d
ad: cmp $0x16,%eax f9: cmp $0x16,%rax
b0: je 0x00000000000000bf fd: je 0x0000000000000110
ff: mov %r13,%rsi
b2: lea 0x10(%rbx),%esi 102: add $0x10,%esi
b5: callq 0xffffffffe1021e0d 105: callq 0xffffffffe102bd6d
ba: cmp $0x16,%eax 10a: cmp $0x16,%rax
bd: jne 0x00000000000000c6 10e: jne 0x0000000000000117
bf: mov $0xffff,%eax 110: mov $0xffff,%eax
c4: jmp 0x00000000000000c8 115: jmp 0x000000000000011c
c6: xor %eax,%eax 117: mov $0x0,%eax
c8: mov -0x8(%rbp),%rbx 11c: mov -0x228(%rbp),%rbx // epilogue
cc: leaveq 123: mov -0x220(%rbp),%r13
cd: retq 12a: mov -0x218(%rbp),%r14
131: mov -0x210(%rbp),%r15
138: leaveq
139: retq
On fully cached SKBs both JITed functions take 12 nsec to execute.
BPF interpreter executes the program in 30 nsec.
The difference in generated assembler is due to the following:
Old BPF imlements LDX_MSH instruction via sk_load_byte_msh() helper function
inside bpf_jit.S.
New JIT removes the helper and does it explicitly, so ldx_msh cost
is the same for both JITs, but generated code looks longer.
New JIT has 4 registers to save, so prologue/epilogue are larger,
but the cost is within noise on x64.
Old JIT checks whether first insn clears A and if not emits 'xor %eax,%eax'.
New JIT clears %rax unconditionally.
2. old BPF JIT doesn't support ANC_NLATTR, ANC_PAY_OFFSET, ANC_RANDOM
extensions. New JIT supports all BPF extensions.
Performance of such filters improves 2-4 times depending on a filter.
The longer the filter the higher performance gain.
Synthetic benchmarks with many ancillary loads see 20x speedup
which seems to be the maximum gain from JIT
Notes:
. net.core.bpf_jit_enable=2 + tools/net/bpf_jit_disasm is still functional
and can be used to see generated assembler
. there are two jit_compile() functions and code flow for classic filters is:
sk_attach_filter() - load classic BPF
bpf_jit_compile() - try to JIT from classic BPF
sk_convert_filter() - convert classic to internal
bpf_int_jit_compile() - JIT from internal BPF
seccomp and tracing filters will just call bpf_int_jit_compile()
Signed-off-by: Alexei Starovoitov <ast@plumgrid.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
2014-05-14 02:50:46 +00:00
|
|
|
|
x86/bpf: Clean up non-standard comments, to make the code more readable
So by chance I looked into x86 assembly in arch/x86/net/bpf_jit_comp.c and
noticed the weird and inconsistent comment style it mistakenly learned from
the networking code:
/* Multi-line comment ...
* ... looks like this.
*/
Fix this to use the standard comment style specified in Documentation/CodingStyle
and used in arch/x86/ as well:
/*
* Multi-line comment ...
* ... looks like this.
*/
Also, to quote Linus's ... more explicit views about this:
http://article.gmane.org/gmane.linux.kernel.cryptoapi/21066
> But no, the networking code picked *none* of the above sane formats.
> Instead, it picked these two models that are just half-arsed
> shit-for-brains:
>
> (no)
> /* This is disgusting drug-induced
> * crap, and should die
> */
>
> (no-no-no)
> /* This is also very nasty
> * and visually unbalanced */
>
> Please. The networking code actually has the *worst* possible comment
> style. You can literally find that (no-no-no) style, which is just
> really horribly disgusting and worse than the otherwise fairly similar
> (d) in pretty much every way.
Also improve the comments and some other details while at it:
- Don't mix same-line and previous-line comment style on otherwise
identical code patterns within the same function,
- capitalize 'BPF' and x86 register names consistently,
- capitalize sentences consistently,
- instead of 'x64' use 'x86-64': x64 is a Microsoft specific term,
- use more consistent punctuation,
- use standard coding style in macros as well,
- fix typos and a few other minor details.
Consistent coding style is not optional, at least in arch/x86/.
No change in functionality.
( In case this commit causes conflicts with pending development code
I'll be glad to help resolve any conflicts! )
Acked-by: Thomas Gleixner <tglx@linutronix.de>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Borislav Petkov <bp@alien8.de>
Cc: H. Peter Anvin <hpa@zytor.com>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: David S. Miller <davem@davemloft.net>
Cc: Eric Dumazet <edumazet@google.com>
Cc: Daniel Borkmann <daniel@iogearbox.net>
Cc: Alexei Starovoitov <ast@fb.com>
Cc: Hideaki YOSHIFUJI <yoshfuji@linux-ipv6.org>
Cc: netdev@vger.kernel.org
Cc: linux-kernel@vger.kernel.org
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
2018-04-27 09:54:40 +00:00
|
|
|
/*
|
|
|
|
* The following table maps BPF registers to x86-64 registers.
|
2016-05-13 17:08:33 +00:00
|
|
|
*
|
x86/bpf: Clean up non-standard comments, to make the code more readable
So by chance I looked into x86 assembly in arch/x86/net/bpf_jit_comp.c and
noticed the weird and inconsistent comment style it mistakenly learned from
the networking code:
/* Multi-line comment ...
* ... looks like this.
*/
Fix this to use the standard comment style specified in Documentation/CodingStyle
and used in arch/x86/ as well:
/*
* Multi-line comment ...
* ... looks like this.
*/
Also, to quote Linus's ... more explicit views about this:
http://article.gmane.org/gmane.linux.kernel.cryptoapi/21066
> But no, the networking code picked *none* of the above sane formats.
> Instead, it picked these two models that are just half-arsed
> shit-for-brains:
>
> (no)
> /* This is disgusting drug-induced
> * crap, and should die
> */
>
> (no-no-no)
> /* This is also very nasty
> * and visually unbalanced */
>
> Please. The networking code actually has the *worst* possible comment
> style. You can literally find that (no-no-no) style, which is just
> really horribly disgusting and worse than the otherwise fairly similar
> (d) in pretty much every way.
Also improve the comments and some other details while at it:
- Don't mix same-line and previous-line comment style on otherwise
identical code patterns within the same function,
- capitalize 'BPF' and x86 register names consistently,
- capitalize sentences consistently,
- instead of 'x64' use 'x86-64': x64 is a Microsoft specific term,
- use more consistent punctuation,
- use standard coding style in macros as well,
- fix typos and a few other minor details.
Consistent coding style is not optional, at least in arch/x86/.
No change in functionality.
( In case this commit causes conflicts with pending development code
I'll be glad to help resolve any conflicts! )
Acked-by: Thomas Gleixner <tglx@linutronix.de>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Borislav Petkov <bp@alien8.de>
Cc: H. Peter Anvin <hpa@zytor.com>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: David S. Miller <davem@davemloft.net>
Cc: Eric Dumazet <edumazet@google.com>
Cc: Daniel Borkmann <daniel@iogearbox.net>
Cc: Alexei Starovoitov <ast@fb.com>
Cc: Hideaki YOSHIFUJI <yoshfuji@linux-ipv6.org>
Cc: netdev@vger.kernel.org
Cc: linux-kernel@vger.kernel.org
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
2018-04-27 09:54:40 +00:00
|
|
|
* x86-64 register R12 is unused, since if used as base address
|
2016-05-13 17:08:33 +00:00
|
|
|
* register in load/store instructions, it always needs an
|
|
|
|
* extra byte of encoding and is callee saved.
|
|
|
|
*
|
2019-11-14 18:57:04 +00:00
|
|
|
* x86-64 register R9 is not used by BPF programs, but can be used by BPF
|
|
|
|
* trampoline. x86-64 register R10 is used for blinding (if enabled).
|
net: filter: x86: internal BPF JIT
Maps all internal BPF instructions into x86_64 instructions.
This patch replaces original BPF x64 JIT with internal BPF x64 JIT.
sysctl net.core.bpf_jit_enable is reused as on/off switch.
Performance:
1. old BPF JIT and internal BPF JIT generate equivalent x86_64 code.
No performance difference is observed for filters that were JIT-able before
Example assembler code for BPF filter "tcpdump port 22"
original BPF -> old JIT: original BPF -> internal BPF -> new JIT:
0: push %rbp 0: push %rbp
1: mov %rsp,%rbp 1: mov %rsp,%rbp
4: sub $0x60,%rsp 4: sub $0x228,%rsp
8: mov %rbx,-0x8(%rbp) b: mov %rbx,-0x228(%rbp) // prologue
12: mov %r13,-0x220(%rbp)
19: mov %r14,-0x218(%rbp)
20: mov %r15,-0x210(%rbp)
27: xor %eax,%eax // clear A
c: xor %ebx,%ebx 29: xor %r13,%r13 // clear X
e: mov 0x68(%rdi),%r9d 2c: mov 0x68(%rdi),%r9d
12: sub 0x6c(%rdi),%r9d 30: sub 0x6c(%rdi),%r9d
16: mov 0xd8(%rdi),%r8 34: mov 0xd8(%rdi),%r10
3b: mov %rdi,%rbx
1d: mov $0xc,%esi 3e: mov $0xc,%esi
22: callq 0xffffffffe1021e15 43: callq 0xffffffffe102bd75
27: cmp $0x86dd,%eax 48: cmp $0x86dd,%rax
2c: jne 0x0000000000000069 4f: jne 0x000000000000009a
2e: mov $0x14,%esi 51: mov $0x14,%esi
33: callq 0xffffffffe1021e31 56: callq 0xffffffffe102bd91
38: cmp $0x84,%eax 5b: cmp $0x84,%rax
3d: je 0x0000000000000049 62: je 0x0000000000000074
3f: cmp $0x6,%eax 64: cmp $0x6,%rax
42: je 0x0000000000000049 68: je 0x0000000000000074
44: cmp $0x11,%eax 6a: cmp $0x11,%rax
47: jne 0x00000000000000c6 6e: jne 0x0000000000000117
49: mov $0x36,%esi 74: mov $0x36,%esi
4e: callq 0xffffffffe1021e15 79: callq 0xffffffffe102bd75
53: cmp $0x16,%eax 7e: cmp $0x16,%rax
56: je 0x00000000000000bf 82: je 0x0000000000000110
58: mov $0x38,%esi 88: mov $0x38,%esi
5d: callq 0xffffffffe1021e15 8d: callq 0xffffffffe102bd75
62: cmp $0x16,%eax 92: cmp $0x16,%rax
65: je 0x00000000000000bf 96: je 0x0000000000000110
67: jmp 0x00000000000000c6 98: jmp 0x0000000000000117
69: cmp $0x800,%eax 9a: cmp $0x800,%rax
6e: jne 0x00000000000000c6 a1: jne 0x0000000000000117
70: mov $0x17,%esi a3: mov $0x17,%esi
75: callq 0xffffffffe1021e31 a8: callq 0xffffffffe102bd91
7a: cmp $0x84,%eax ad: cmp $0x84,%rax
7f: je 0x000000000000008b b4: je 0x00000000000000c2
81: cmp $0x6,%eax b6: cmp $0x6,%rax
84: je 0x000000000000008b ba: je 0x00000000000000c2
86: cmp $0x11,%eax bc: cmp $0x11,%rax
89: jne 0x00000000000000c6 c0: jne 0x0000000000000117
8b: mov $0x14,%esi c2: mov $0x14,%esi
90: callq 0xffffffffe1021e15 c7: callq 0xffffffffe102bd75
95: test $0x1fff,%ax cc: test $0x1fff,%rax
99: jne 0x00000000000000c6 d3: jne 0x0000000000000117
d5: mov %rax,%r14
9b: mov $0xe,%esi d8: mov $0xe,%esi
a0: callq 0xffffffffe1021e44 dd: callq 0xffffffffe102bd91 // MSH
e2: and $0xf,%eax
e5: shl $0x2,%eax
e8: mov %rax,%r13
eb: mov %r14,%rax
ee: mov %r13,%rsi
a5: lea 0xe(%rbx),%esi f1: add $0xe,%esi
a8: callq 0xffffffffe1021e0d f4: callq 0xffffffffe102bd6d
ad: cmp $0x16,%eax f9: cmp $0x16,%rax
b0: je 0x00000000000000bf fd: je 0x0000000000000110
ff: mov %r13,%rsi
b2: lea 0x10(%rbx),%esi 102: add $0x10,%esi
b5: callq 0xffffffffe1021e0d 105: callq 0xffffffffe102bd6d
ba: cmp $0x16,%eax 10a: cmp $0x16,%rax
bd: jne 0x00000000000000c6 10e: jne 0x0000000000000117
bf: mov $0xffff,%eax 110: mov $0xffff,%eax
c4: jmp 0x00000000000000c8 115: jmp 0x000000000000011c
c6: xor %eax,%eax 117: mov $0x0,%eax
c8: mov -0x8(%rbp),%rbx 11c: mov -0x228(%rbp),%rbx // epilogue
cc: leaveq 123: mov -0x220(%rbp),%r13
cd: retq 12a: mov -0x218(%rbp),%r14
131: mov -0x210(%rbp),%r15
138: leaveq
139: retq
On fully cached SKBs both JITed functions take 12 nsec to execute.
BPF interpreter executes the program in 30 nsec.
The difference in generated assembler is due to the following:
Old BPF imlements LDX_MSH instruction via sk_load_byte_msh() helper function
inside bpf_jit.S.
New JIT removes the helper and does it explicitly, so ldx_msh cost
is the same for both JITs, but generated code looks longer.
New JIT has 4 registers to save, so prologue/epilogue are larger,
but the cost is within noise on x64.
Old JIT checks whether first insn clears A and if not emits 'xor %eax,%eax'.
New JIT clears %rax unconditionally.
2. old BPF JIT doesn't support ANC_NLATTR, ANC_PAY_OFFSET, ANC_RANDOM
extensions. New JIT supports all BPF extensions.
Performance of such filters improves 2-4 times depending on a filter.
The longer the filter the higher performance gain.
Synthetic benchmarks with many ancillary loads see 20x speedup
which seems to be the maximum gain from JIT
Notes:
. net.core.bpf_jit_enable=2 + tools/net/bpf_jit_disasm is still functional
and can be used to see generated assembler
. there are two jit_compile() functions and code flow for classic filters is:
sk_attach_filter() - load classic BPF
bpf_jit_compile() - try to JIT from classic BPF
sk_convert_filter() - convert classic to internal
bpf_int_jit_compile() - JIT from internal BPF
seccomp and tracing filters will just call bpf_int_jit_compile()
Signed-off-by: Alexei Starovoitov <ast@plumgrid.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
2014-05-14 02:50:46 +00:00
|
|
|
*/
|
|
|
|
static const int reg2hex[] = {
|
x86/bpf: Clean up non-standard comments, to make the code more readable
So by chance I looked into x86 assembly in arch/x86/net/bpf_jit_comp.c and
noticed the weird and inconsistent comment style it mistakenly learned from
the networking code:
/* Multi-line comment ...
* ... looks like this.
*/
Fix this to use the standard comment style specified in Documentation/CodingStyle
and used in arch/x86/ as well:
/*
* Multi-line comment ...
* ... looks like this.
*/
Also, to quote Linus's ... more explicit views about this:
http://article.gmane.org/gmane.linux.kernel.cryptoapi/21066
> But no, the networking code picked *none* of the above sane formats.
> Instead, it picked these two models that are just half-arsed
> shit-for-brains:
>
> (no)
> /* This is disgusting drug-induced
> * crap, and should die
> */
>
> (no-no-no)
> /* This is also very nasty
> * and visually unbalanced */
>
> Please. The networking code actually has the *worst* possible comment
> style. You can literally find that (no-no-no) style, which is just
> really horribly disgusting and worse than the otherwise fairly similar
> (d) in pretty much every way.
Also improve the comments and some other details while at it:
- Don't mix same-line and previous-line comment style on otherwise
identical code patterns within the same function,
- capitalize 'BPF' and x86 register names consistently,
- capitalize sentences consistently,
- instead of 'x64' use 'x86-64': x64 is a Microsoft specific term,
- use more consistent punctuation,
- use standard coding style in macros as well,
- fix typos and a few other minor details.
Consistent coding style is not optional, at least in arch/x86/.
No change in functionality.
( In case this commit causes conflicts with pending development code
I'll be glad to help resolve any conflicts! )
Acked-by: Thomas Gleixner <tglx@linutronix.de>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Borislav Petkov <bp@alien8.de>
Cc: H. Peter Anvin <hpa@zytor.com>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: David S. Miller <davem@davemloft.net>
Cc: Eric Dumazet <edumazet@google.com>
Cc: Daniel Borkmann <daniel@iogearbox.net>
Cc: Alexei Starovoitov <ast@fb.com>
Cc: Hideaki YOSHIFUJI <yoshfuji@linux-ipv6.org>
Cc: netdev@vger.kernel.org
Cc: linux-kernel@vger.kernel.org
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
2018-04-27 09:54:40 +00:00
|
|
|
[BPF_REG_0] = 0, /* RAX */
|
|
|
|
[BPF_REG_1] = 7, /* RDI */
|
|
|
|
[BPF_REG_2] = 6, /* RSI */
|
|
|
|
[BPF_REG_3] = 2, /* RDX */
|
|
|
|
[BPF_REG_4] = 1, /* RCX */
|
|
|
|
[BPF_REG_5] = 0, /* R8 */
|
|
|
|
[BPF_REG_6] = 3, /* RBX callee saved */
|
|
|
|
[BPF_REG_7] = 5, /* R13 callee saved */
|
|
|
|
[BPF_REG_8] = 6, /* R14 callee saved */
|
|
|
|
[BPF_REG_9] = 7, /* R15 callee saved */
|
|
|
|
[BPF_REG_FP] = 5, /* RBP readonly */
|
|
|
|
[BPF_REG_AX] = 2, /* R10 temp register */
|
|
|
|
[AUX_REG] = 3, /* R11 temp register */
|
2019-11-14 18:57:04 +00:00
|
|
|
[X86_REG_R9] = 1, /* R9 register, 6th function argument */
|
2024-03-08 01:08:01 +00:00
|
|
|
[X86_REG_R12] = 4, /* R12 callee saved */
|
net: filter: x86: internal BPF JIT
Maps all internal BPF instructions into x86_64 instructions.
This patch replaces original BPF x64 JIT with internal BPF x64 JIT.
sysctl net.core.bpf_jit_enable is reused as on/off switch.
Performance:
1. old BPF JIT and internal BPF JIT generate equivalent x86_64 code.
No performance difference is observed for filters that were JIT-able before
Example assembler code for BPF filter "tcpdump port 22"
original BPF -> old JIT: original BPF -> internal BPF -> new JIT:
0: push %rbp 0: push %rbp
1: mov %rsp,%rbp 1: mov %rsp,%rbp
4: sub $0x60,%rsp 4: sub $0x228,%rsp
8: mov %rbx,-0x8(%rbp) b: mov %rbx,-0x228(%rbp) // prologue
12: mov %r13,-0x220(%rbp)
19: mov %r14,-0x218(%rbp)
20: mov %r15,-0x210(%rbp)
27: xor %eax,%eax // clear A
c: xor %ebx,%ebx 29: xor %r13,%r13 // clear X
e: mov 0x68(%rdi),%r9d 2c: mov 0x68(%rdi),%r9d
12: sub 0x6c(%rdi),%r9d 30: sub 0x6c(%rdi),%r9d
16: mov 0xd8(%rdi),%r8 34: mov 0xd8(%rdi),%r10
3b: mov %rdi,%rbx
1d: mov $0xc,%esi 3e: mov $0xc,%esi
22: callq 0xffffffffe1021e15 43: callq 0xffffffffe102bd75
27: cmp $0x86dd,%eax 48: cmp $0x86dd,%rax
2c: jne 0x0000000000000069 4f: jne 0x000000000000009a
2e: mov $0x14,%esi 51: mov $0x14,%esi
33: callq 0xffffffffe1021e31 56: callq 0xffffffffe102bd91
38: cmp $0x84,%eax 5b: cmp $0x84,%rax
3d: je 0x0000000000000049 62: je 0x0000000000000074
3f: cmp $0x6,%eax 64: cmp $0x6,%rax
42: je 0x0000000000000049 68: je 0x0000000000000074
44: cmp $0x11,%eax 6a: cmp $0x11,%rax
47: jne 0x00000000000000c6 6e: jne 0x0000000000000117
49: mov $0x36,%esi 74: mov $0x36,%esi
4e: callq 0xffffffffe1021e15 79: callq 0xffffffffe102bd75
53: cmp $0x16,%eax 7e: cmp $0x16,%rax
56: je 0x00000000000000bf 82: je 0x0000000000000110
58: mov $0x38,%esi 88: mov $0x38,%esi
5d: callq 0xffffffffe1021e15 8d: callq 0xffffffffe102bd75
62: cmp $0x16,%eax 92: cmp $0x16,%rax
65: je 0x00000000000000bf 96: je 0x0000000000000110
67: jmp 0x00000000000000c6 98: jmp 0x0000000000000117
69: cmp $0x800,%eax 9a: cmp $0x800,%rax
6e: jne 0x00000000000000c6 a1: jne 0x0000000000000117
70: mov $0x17,%esi a3: mov $0x17,%esi
75: callq 0xffffffffe1021e31 a8: callq 0xffffffffe102bd91
7a: cmp $0x84,%eax ad: cmp $0x84,%rax
7f: je 0x000000000000008b b4: je 0x00000000000000c2
81: cmp $0x6,%eax b6: cmp $0x6,%rax
84: je 0x000000000000008b ba: je 0x00000000000000c2
86: cmp $0x11,%eax bc: cmp $0x11,%rax
89: jne 0x00000000000000c6 c0: jne 0x0000000000000117
8b: mov $0x14,%esi c2: mov $0x14,%esi
90: callq 0xffffffffe1021e15 c7: callq 0xffffffffe102bd75
95: test $0x1fff,%ax cc: test $0x1fff,%rax
99: jne 0x00000000000000c6 d3: jne 0x0000000000000117
d5: mov %rax,%r14
9b: mov $0xe,%esi d8: mov $0xe,%esi
a0: callq 0xffffffffe1021e44 dd: callq 0xffffffffe102bd91 // MSH
e2: and $0xf,%eax
e5: shl $0x2,%eax
e8: mov %rax,%r13
eb: mov %r14,%rax
ee: mov %r13,%rsi
a5: lea 0xe(%rbx),%esi f1: add $0xe,%esi
a8: callq 0xffffffffe1021e0d f4: callq 0xffffffffe102bd6d
ad: cmp $0x16,%eax f9: cmp $0x16,%rax
b0: je 0x00000000000000bf fd: je 0x0000000000000110
ff: mov %r13,%rsi
b2: lea 0x10(%rbx),%esi 102: add $0x10,%esi
b5: callq 0xffffffffe1021e0d 105: callq 0xffffffffe102bd6d
ba: cmp $0x16,%eax 10a: cmp $0x16,%rax
bd: jne 0x00000000000000c6 10e: jne 0x0000000000000117
bf: mov $0xffff,%eax 110: mov $0xffff,%eax
c4: jmp 0x00000000000000c8 115: jmp 0x000000000000011c
c6: xor %eax,%eax 117: mov $0x0,%eax
c8: mov -0x8(%rbp),%rbx 11c: mov -0x228(%rbp),%rbx // epilogue
cc: leaveq 123: mov -0x220(%rbp),%r13
cd: retq 12a: mov -0x218(%rbp),%r14
131: mov -0x210(%rbp),%r15
138: leaveq
139: retq
On fully cached SKBs both JITed functions take 12 nsec to execute.
BPF interpreter executes the program in 30 nsec.
The difference in generated assembler is due to the following:
Old BPF imlements LDX_MSH instruction via sk_load_byte_msh() helper function
inside bpf_jit.S.
New JIT removes the helper and does it explicitly, so ldx_msh cost
is the same for both JITs, but generated code looks longer.
New JIT has 4 registers to save, so prologue/epilogue are larger,
but the cost is within noise on x64.
Old JIT checks whether first insn clears A and if not emits 'xor %eax,%eax'.
New JIT clears %rax unconditionally.
2. old BPF JIT doesn't support ANC_NLATTR, ANC_PAY_OFFSET, ANC_RANDOM
extensions. New JIT supports all BPF extensions.
Performance of such filters improves 2-4 times depending on a filter.
The longer the filter the higher performance gain.
Synthetic benchmarks with many ancillary loads see 20x speedup
which seems to be the maximum gain from JIT
Notes:
. net.core.bpf_jit_enable=2 + tools/net/bpf_jit_disasm is still functional
and can be used to see generated assembler
. there are two jit_compile() functions and code flow for classic filters is:
sk_attach_filter() - load classic BPF
bpf_jit_compile() - try to JIT from classic BPF
sk_convert_filter() - convert classic to internal
bpf_int_jit_compile() - JIT from internal BPF
seccomp and tracing filters will just call bpf_int_jit_compile()
Signed-off-by: Alexei Starovoitov <ast@plumgrid.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
2014-05-14 02:50:46 +00:00
|
|
|
};
|
|
|
|
|
2019-10-16 03:25:03 +00:00
|
|
|
static const int reg2pt_regs[] = {
|
|
|
|
[BPF_REG_0] = offsetof(struct pt_regs, ax),
|
|
|
|
[BPF_REG_1] = offsetof(struct pt_regs, di),
|
|
|
|
[BPF_REG_2] = offsetof(struct pt_regs, si),
|
|
|
|
[BPF_REG_3] = offsetof(struct pt_regs, dx),
|
|
|
|
[BPF_REG_4] = offsetof(struct pt_regs, cx),
|
|
|
|
[BPF_REG_5] = offsetof(struct pt_regs, r8),
|
|
|
|
[BPF_REG_6] = offsetof(struct pt_regs, bx),
|
|
|
|
[BPF_REG_7] = offsetof(struct pt_regs, r13),
|
|
|
|
[BPF_REG_8] = offsetof(struct pt_regs, r14),
|
|
|
|
[BPF_REG_9] = offsetof(struct pt_regs, r15),
|
|
|
|
};
|
|
|
|
|
x86/bpf: Clean up non-standard comments, to make the code more readable
So by chance I looked into x86 assembly in arch/x86/net/bpf_jit_comp.c and
noticed the weird and inconsistent comment style it mistakenly learned from
the networking code:
/* Multi-line comment ...
* ... looks like this.
*/
Fix this to use the standard comment style specified in Documentation/CodingStyle
and used in arch/x86/ as well:
/*
* Multi-line comment ...
* ... looks like this.
*/
Also, to quote Linus's ... more explicit views about this:
http://article.gmane.org/gmane.linux.kernel.cryptoapi/21066
> But no, the networking code picked *none* of the above sane formats.
> Instead, it picked these two models that are just half-arsed
> shit-for-brains:
>
> (no)
> /* This is disgusting drug-induced
> * crap, and should die
> */
>
> (no-no-no)
> /* This is also very nasty
> * and visually unbalanced */
>
> Please. The networking code actually has the *worst* possible comment
> style. You can literally find that (no-no-no) style, which is just
> really horribly disgusting and worse than the otherwise fairly similar
> (d) in pretty much every way.
Also improve the comments and some other details while at it:
- Don't mix same-line and previous-line comment style on otherwise
identical code patterns within the same function,
- capitalize 'BPF' and x86 register names consistently,
- capitalize sentences consistently,
- instead of 'x64' use 'x86-64': x64 is a Microsoft specific term,
- use more consistent punctuation,
- use standard coding style in macros as well,
- fix typos and a few other minor details.
Consistent coding style is not optional, at least in arch/x86/.
No change in functionality.
( In case this commit causes conflicts with pending development code
I'll be glad to help resolve any conflicts! )
Acked-by: Thomas Gleixner <tglx@linutronix.de>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Borislav Petkov <bp@alien8.de>
Cc: H. Peter Anvin <hpa@zytor.com>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: David S. Miller <davem@davemloft.net>
Cc: Eric Dumazet <edumazet@google.com>
Cc: Daniel Borkmann <daniel@iogearbox.net>
Cc: Alexei Starovoitov <ast@fb.com>
Cc: Hideaki YOSHIFUJI <yoshfuji@linux-ipv6.org>
Cc: netdev@vger.kernel.org
Cc: linux-kernel@vger.kernel.org
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
2018-04-27 09:54:40 +00:00
|
|
|
/*
|
|
|
|
* is_ereg() == true if BPF register 'reg' maps to x86-64 r8..r15
|
net: filter: x86: internal BPF JIT
Maps all internal BPF instructions into x86_64 instructions.
This patch replaces original BPF x64 JIT with internal BPF x64 JIT.
sysctl net.core.bpf_jit_enable is reused as on/off switch.
Performance:
1. old BPF JIT and internal BPF JIT generate equivalent x86_64 code.
No performance difference is observed for filters that were JIT-able before
Example assembler code for BPF filter "tcpdump port 22"
original BPF -> old JIT: original BPF -> internal BPF -> new JIT:
0: push %rbp 0: push %rbp
1: mov %rsp,%rbp 1: mov %rsp,%rbp
4: sub $0x60,%rsp 4: sub $0x228,%rsp
8: mov %rbx,-0x8(%rbp) b: mov %rbx,-0x228(%rbp) // prologue
12: mov %r13,-0x220(%rbp)
19: mov %r14,-0x218(%rbp)
20: mov %r15,-0x210(%rbp)
27: xor %eax,%eax // clear A
c: xor %ebx,%ebx 29: xor %r13,%r13 // clear X
e: mov 0x68(%rdi),%r9d 2c: mov 0x68(%rdi),%r9d
12: sub 0x6c(%rdi),%r9d 30: sub 0x6c(%rdi),%r9d
16: mov 0xd8(%rdi),%r8 34: mov 0xd8(%rdi),%r10
3b: mov %rdi,%rbx
1d: mov $0xc,%esi 3e: mov $0xc,%esi
22: callq 0xffffffffe1021e15 43: callq 0xffffffffe102bd75
27: cmp $0x86dd,%eax 48: cmp $0x86dd,%rax
2c: jne 0x0000000000000069 4f: jne 0x000000000000009a
2e: mov $0x14,%esi 51: mov $0x14,%esi
33: callq 0xffffffffe1021e31 56: callq 0xffffffffe102bd91
38: cmp $0x84,%eax 5b: cmp $0x84,%rax
3d: je 0x0000000000000049 62: je 0x0000000000000074
3f: cmp $0x6,%eax 64: cmp $0x6,%rax
42: je 0x0000000000000049 68: je 0x0000000000000074
44: cmp $0x11,%eax 6a: cmp $0x11,%rax
47: jne 0x00000000000000c6 6e: jne 0x0000000000000117
49: mov $0x36,%esi 74: mov $0x36,%esi
4e: callq 0xffffffffe1021e15 79: callq 0xffffffffe102bd75
53: cmp $0x16,%eax 7e: cmp $0x16,%rax
56: je 0x00000000000000bf 82: je 0x0000000000000110
58: mov $0x38,%esi 88: mov $0x38,%esi
5d: callq 0xffffffffe1021e15 8d: callq 0xffffffffe102bd75
62: cmp $0x16,%eax 92: cmp $0x16,%rax
65: je 0x00000000000000bf 96: je 0x0000000000000110
67: jmp 0x00000000000000c6 98: jmp 0x0000000000000117
69: cmp $0x800,%eax 9a: cmp $0x800,%rax
6e: jne 0x00000000000000c6 a1: jne 0x0000000000000117
70: mov $0x17,%esi a3: mov $0x17,%esi
75: callq 0xffffffffe1021e31 a8: callq 0xffffffffe102bd91
7a: cmp $0x84,%eax ad: cmp $0x84,%rax
7f: je 0x000000000000008b b4: je 0x00000000000000c2
81: cmp $0x6,%eax b6: cmp $0x6,%rax
84: je 0x000000000000008b ba: je 0x00000000000000c2
86: cmp $0x11,%eax bc: cmp $0x11,%rax
89: jne 0x00000000000000c6 c0: jne 0x0000000000000117
8b: mov $0x14,%esi c2: mov $0x14,%esi
90: callq 0xffffffffe1021e15 c7: callq 0xffffffffe102bd75
95: test $0x1fff,%ax cc: test $0x1fff,%rax
99: jne 0x00000000000000c6 d3: jne 0x0000000000000117
d5: mov %rax,%r14
9b: mov $0xe,%esi d8: mov $0xe,%esi
a0: callq 0xffffffffe1021e44 dd: callq 0xffffffffe102bd91 // MSH
e2: and $0xf,%eax
e5: shl $0x2,%eax
e8: mov %rax,%r13
eb: mov %r14,%rax
ee: mov %r13,%rsi
a5: lea 0xe(%rbx),%esi f1: add $0xe,%esi
a8: callq 0xffffffffe1021e0d f4: callq 0xffffffffe102bd6d
ad: cmp $0x16,%eax f9: cmp $0x16,%rax
b0: je 0x00000000000000bf fd: je 0x0000000000000110
ff: mov %r13,%rsi
b2: lea 0x10(%rbx),%esi 102: add $0x10,%esi
b5: callq 0xffffffffe1021e0d 105: callq 0xffffffffe102bd6d
ba: cmp $0x16,%eax 10a: cmp $0x16,%rax
bd: jne 0x00000000000000c6 10e: jne 0x0000000000000117
bf: mov $0xffff,%eax 110: mov $0xffff,%eax
c4: jmp 0x00000000000000c8 115: jmp 0x000000000000011c
c6: xor %eax,%eax 117: mov $0x0,%eax
c8: mov -0x8(%rbp),%rbx 11c: mov -0x228(%rbp),%rbx // epilogue
cc: leaveq 123: mov -0x220(%rbp),%r13
cd: retq 12a: mov -0x218(%rbp),%r14
131: mov -0x210(%rbp),%r15
138: leaveq
139: retq
On fully cached SKBs both JITed functions take 12 nsec to execute.
BPF interpreter executes the program in 30 nsec.
The difference in generated assembler is due to the following:
Old BPF imlements LDX_MSH instruction via sk_load_byte_msh() helper function
inside bpf_jit.S.
New JIT removes the helper and does it explicitly, so ldx_msh cost
is the same for both JITs, but generated code looks longer.
New JIT has 4 registers to save, so prologue/epilogue are larger,
but the cost is within noise on x64.
Old JIT checks whether first insn clears A and if not emits 'xor %eax,%eax'.
New JIT clears %rax unconditionally.
2. old BPF JIT doesn't support ANC_NLATTR, ANC_PAY_OFFSET, ANC_RANDOM
extensions. New JIT supports all BPF extensions.
Performance of such filters improves 2-4 times depending on a filter.
The longer the filter the higher performance gain.
Synthetic benchmarks with many ancillary loads see 20x speedup
which seems to be the maximum gain from JIT
Notes:
. net.core.bpf_jit_enable=2 + tools/net/bpf_jit_disasm is still functional
and can be used to see generated assembler
. there are two jit_compile() functions and code flow for classic filters is:
sk_attach_filter() - load classic BPF
bpf_jit_compile() - try to JIT from classic BPF
sk_convert_filter() - convert classic to internal
bpf_int_jit_compile() - JIT from internal BPF
seccomp and tracing filters will just call bpf_int_jit_compile()
Signed-off-by: Alexei Starovoitov <ast@plumgrid.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
2014-05-14 02:50:46 +00:00
|
|
|
* which need extra byte of encoding.
|
|
|
|
* rax,rcx,...,rbp have simpler encoding
|
|
|
|
*/
|
2014-12-05 01:01:24 +00:00
|
|
|
static bool is_ereg(u32 reg)
|
net: filter: x86: internal BPF JIT
Maps all internal BPF instructions into x86_64 instructions.
This patch replaces original BPF x64 JIT with internal BPF x64 JIT.
sysctl net.core.bpf_jit_enable is reused as on/off switch.
Performance:
1. old BPF JIT and internal BPF JIT generate equivalent x86_64 code.
No performance difference is observed for filters that were JIT-able before
Example assembler code for BPF filter "tcpdump port 22"
original BPF -> old JIT: original BPF -> internal BPF -> new JIT:
0: push %rbp 0: push %rbp
1: mov %rsp,%rbp 1: mov %rsp,%rbp
4: sub $0x60,%rsp 4: sub $0x228,%rsp
8: mov %rbx,-0x8(%rbp) b: mov %rbx,-0x228(%rbp) // prologue
12: mov %r13,-0x220(%rbp)
19: mov %r14,-0x218(%rbp)
20: mov %r15,-0x210(%rbp)
27: xor %eax,%eax // clear A
c: xor %ebx,%ebx 29: xor %r13,%r13 // clear X
e: mov 0x68(%rdi),%r9d 2c: mov 0x68(%rdi),%r9d
12: sub 0x6c(%rdi),%r9d 30: sub 0x6c(%rdi),%r9d
16: mov 0xd8(%rdi),%r8 34: mov 0xd8(%rdi),%r10
3b: mov %rdi,%rbx
1d: mov $0xc,%esi 3e: mov $0xc,%esi
22: callq 0xffffffffe1021e15 43: callq 0xffffffffe102bd75
27: cmp $0x86dd,%eax 48: cmp $0x86dd,%rax
2c: jne 0x0000000000000069 4f: jne 0x000000000000009a
2e: mov $0x14,%esi 51: mov $0x14,%esi
33: callq 0xffffffffe1021e31 56: callq 0xffffffffe102bd91
38: cmp $0x84,%eax 5b: cmp $0x84,%rax
3d: je 0x0000000000000049 62: je 0x0000000000000074
3f: cmp $0x6,%eax 64: cmp $0x6,%rax
42: je 0x0000000000000049 68: je 0x0000000000000074
44: cmp $0x11,%eax 6a: cmp $0x11,%rax
47: jne 0x00000000000000c6 6e: jne 0x0000000000000117
49: mov $0x36,%esi 74: mov $0x36,%esi
4e: callq 0xffffffffe1021e15 79: callq 0xffffffffe102bd75
53: cmp $0x16,%eax 7e: cmp $0x16,%rax
56: je 0x00000000000000bf 82: je 0x0000000000000110
58: mov $0x38,%esi 88: mov $0x38,%esi
5d: callq 0xffffffffe1021e15 8d: callq 0xffffffffe102bd75
62: cmp $0x16,%eax 92: cmp $0x16,%rax
65: je 0x00000000000000bf 96: je 0x0000000000000110
67: jmp 0x00000000000000c6 98: jmp 0x0000000000000117
69: cmp $0x800,%eax 9a: cmp $0x800,%rax
6e: jne 0x00000000000000c6 a1: jne 0x0000000000000117
70: mov $0x17,%esi a3: mov $0x17,%esi
75: callq 0xffffffffe1021e31 a8: callq 0xffffffffe102bd91
7a: cmp $0x84,%eax ad: cmp $0x84,%rax
7f: je 0x000000000000008b b4: je 0x00000000000000c2
81: cmp $0x6,%eax b6: cmp $0x6,%rax
84: je 0x000000000000008b ba: je 0x00000000000000c2
86: cmp $0x11,%eax bc: cmp $0x11,%rax
89: jne 0x00000000000000c6 c0: jne 0x0000000000000117
8b: mov $0x14,%esi c2: mov $0x14,%esi
90: callq 0xffffffffe1021e15 c7: callq 0xffffffffe102bd75
95: test $0x1fff,%ax cc: test $0x1fff,%rax
99: jne 0x00000000000000c6 d3: jne 0x0000000000000117
d5: mov %rax,%r14
9b: mov $0xe,%esi d8: mov $0xe,%esi
a0: callq 0xffffffffe1021e44 dd: callq 0xffffffffe102bd91 // MSH
e2: and $0xf,%eax
e5: shl $0x2,%eax
e8: mov %rax,%r13
eb: mov %r14,%rax
ee: mov %r13,%rsi
a5: lea 0xe(%rbx),%esi f1: add $0xe,%esi
a8: callq 0xffffffffe1021e0d f4: callq 0xffffffffe102bd6d
ad: cmp $0x16,%eax f9: cmp $0x16,%rax
b0: je 0x00000000000000bf fd: je 0x0000000000000110
ff: mov %r13,%rsi
b2: lea 0x10(%rbx),%esi 102: add $0x10,%esi
b5: callq 0xffffffffe1021e0d 105: callq 0xffffffffe102bd6d
ba: cmp $0x16,%eax 10a: cmp $0x16,%rax
bd: jne 0x00000000000000c6 10e: jne 0x0000000000000117
bf: mov $0xffff,%eax 110: mov $0xffff,%eax
c4: jmp 0x00000000000000c8 115: jmp 0x000000000000011c
c6: xor %eax,%eax 117: mov $0x0,%eax
c8: mov -0x8(%rbp),%rbx 11c: mov -0x228(%rbp),%rbx // epilogue
cc: leaveq 123: mov -0x220(%rbp),%r13
cd: retq 12a: mov -0x218(%rbp),%r14
131: mov -0x210(%rbp),%r15
138: leaveq
139: retq
On fully cached SKBs both JITed functions take 12 nsec to execute.
BPF interpreter executes the program in 30 nsec.
The difference in generated assembler is due to the following:
Old BPF imlements LDX_MSH instruction via sk_load_byte_msh() helper function
inside bpf_jit.S.
New JIT removes the helper and does it explicitly, so ldx_msh cost
is the same for both JITs, but generated code looks longer.
New JIT has 4 registers to save, so prologue/epilogue are larger,
but the cost is within noise on x64.
Old JIT checks whether first insn clears A and if not emits 'xor %eax,%eax'.
New JIT clears %rax unconditionally.
2. old BPF JIT doesn't support ANC_NLATTR, ANC_PAY_OFFSET, ANC_RANDOM
extensions. New JIT supports all BPF extensions.
Performance of such filters improves 2-4 times depending on a filter.
The longer the filter the higher performance gain.
Synthetic benchmarks with many ancillary loads see 20x speedup
which seems to be the maximum gain from JIT
Notes:
. net.core.bpf_jit_enable=2 + tools/net/bpf_jit_disasm is still functional
and can be used to see generated assembler
. there are two jit_compile() functions and code flow for classic filters is:
sk_attach_filter() - load classic BPF
bpf_jit_compile() - try to JIT from classic BPF
sk_convert_filter() - convert classic to internal
bpf_int_jit_compile() - JIT from internal BPF
seccomp and tracing filters will just call bpf_int_jit_compile()
Signed-off-by: Alexei Starovoitov <ast@plumgrid.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
2014-05-14 02:50:46 +00:00
|
|
|
{
|
2014-12-04 23:00:48 +00:00
|
|
|
return (1 << reg) & (BIT(BPF_REG_5) |
|
|
|
|
BIT(AUX_REG) |
|
|
|
|
BIT(BPF_REG_7) |
|
|
|
|
BIT(BPF_REG_8) |
|
2016-05-13 17:08:33 +00:00
|
|
|
BIT(BPF_REG_9) |
|
2019-11-14 18:57:04 +00:00
|
|
|
BIT(X86_REG_R9) |
|
2024-03-08 01:08:01 +00:00
|
|
|
BIT(X86_REG_R12) |
|
2016-05-13 17:08:33 +00:00
|
|
|
BIT(BPF_REG_AX));
|
net: filter: x86: internal BPF JIT
Maps all internal BPF instructions into x86_64 instructions.
This patch replaces original BPF x64 JIT with internal BPF x64 JIT.
sysctl net.core.bpf_jit_enable is reused as on/off switch.
Performance:
1. old BPF JIT and internal BPF JIT generate equivalent x86_64 code.
No performance difference is observed for filters that were JIT-able before
Example assembler code for BPF filter "tcpdump port 22"
original BPF -> old JIT: original BPF -> internal BPF -> new JIT:
0: push %rbp 0: push %rbp
1: mov %rsp,%rbp 1: mov %rsp,%rbp
4: sub $0x60,%rsp 4: sub $0x228,%rsp
8: mov %rbx,-0x8(%rbp) b: mov %rbx,-0x228(%rbp) // prologue
12: mov %r13,-0x220(%rbp)
19: mov %r14,-0x218(%rbp)
20: mov %r15,-0x210(%rbp)
27: xor %eax,%eax // clear A
c: xor %ebx,%ebx 29: xor %r13,%r13 // clear X
e: mov 0x68(%rdi),%r9d 2c: mov 0x68(%rdi),%r9d
12: sub 0x6c(%rdi),%r9d 30: sub 0x6c(%rdi),%r9d
16: mov 0xd8(%rdi),%r8 34: mov 0xd8(%rdi),%r10
3b: mov %rdi,%rbx
1d: mov $0xc,%esi 3e: mov $0xc,%esi
22: callq 0xffffffffe1021e15 43: callq 0xffffffffe102bd75
27: cmp $0x86dd,%eax 48: cmp $0x86dd,%rax
2c: jne 0x0000000000000069 4f: jne 0x000000000000009a
2e: mov $0x14,%esi 51: mov $0x14,%esi
33: callq 0xffffffffe1021e31 56: callq 0xffffffffe102bd91
38: cmp $0x84,%eax 5b: cmp $0x84,%rax
3d: je 0x0000000000000049 62: je 0x0000000000000074
3f: cmp $0x6,%eax 64: cmp $0x6,%rax
42: je 0x0000000000000049 68: je 0x0000000000000074
44: cmp $0x11,%eax 6a: cmp $0x11,%rax
47: jne 0x00000000000000c6 6e: jne 0x0000000000000117
49: mov $0x36,%esi 74: mov $0x36,%esi
4e: callq 0xffffffffe1021e15 79: callq 0xffffffffe102bd75
53: cmp $0x16,%eax 7e: cmp $0x16,%rax
56: je 0x00000000000000bf 82: je 0x0000000000000110
58: mov $0x38,%esi 88: mov $0x38,%esi
5d: callq 0xffffffffe1021e15 8d: callq 0xffffffffe102bd75
62: cmp $0x16,%eax 92: cmp $0x16,%rax
65: je 0x00000000000000bf 96: je 0x0000000000000110
67: jmp 0x00000000000000c6 98: jmp 0x0000000000000117
69: cmp $0x800,%eax 9a: cmp $0x800,%rax
6e: jne 0x00000000000000c6 a1: jne 0x0000000000000117
70: mov $0x17,%esi a3: mov $0x17,%esi
75: callq 0xffffffffe1021e31 a8: callq 0xffffffffe102bd91
7a: cmp $0x84,%eax ad: cmp $0x84,%rax
7f: je 0x000000000000008b b4: je 0x00000000000000c2
81: cmp $0x6,%eax b6: cmp $0x6,%rax
84: je 0x000000000000008b ba: je 0x00000000000000c2
86: cmp $0x11,%eax bc: cmp $0x11,%rax
89: jne 0x00000000000000c6 c0: jne 0x0000000000000117
8b: mov $0x14,%esi c2: mov $0x14,%esi
90: callq 0xffffffffe1021e15 c7: callq 0xffffffffe102bd75
95: test $0x1fff,%ax cc: test $0x1fff,%rax
99: jne 0x00000000000000c6 d3: jne 0x0000000000000117
d5: mov %rax,%r14
9b: mov $0xe,%esi d8: mov $0xe,%esi
a0: callq 0xffffffffe1021e44 dd: callq 0xffffffffe102bd91 // MSH
e2: and $0xf,%eax
e5: shl $0x2,%eax
e8: mov %rax,%r13
eb: mov %r14,%rax
ee: mov %r13,%rsi
a5: lea 0xe(%rbx),%esi f1: add $0xe,%esi
a8: callq 0xffffffffe1021e0d f4: callq 0xffffffffe102bd6d
ad: cmp $0x16,%eax f9: cmp $0x16,%rax
b0: je 0x00000000000000bf fd: je 0x0000000000000110
ff: mov %r13,%rsi
b2: lea 0x10(%rbx),%esi 102: add $0x10,%esi
b5: callq 0xffffffffe1021e0d 105: callq 0xffffffffe102bd6d
ba: cmp $0x16,%eax 10a: cmp $0x16,%rax
bd: jne 0x00000000000000c6 10e: jne 0x0000000000000117
bf: mov $0xffff,%eax 110: mov $0xffff,%eax
c4: jmp 0x00000000000000c8 115: jmp 0x000000000000011c
c6: xor %eax,%eax 117: mov $0x0,%eax
c8: mov -0x8(%rbp),%rbx 11c: mov -0x228(%rbp),%rbx // epilogue
cc: leaveq 123: mov -0x220(%rbp),%r13
cd: retq 12a: mov -0x218(%rbp),%r14
131: mov -0x210(%rbp),%r15
138: leaveq
139: retq
On fully cached SKBs both JITed functions take 12 nsec to execute.
BPF interpreter executes the program in 30 nsec.
The difference in generated assembler is due to the following:
Old BPF imlements LDX_MSH instruction via sk_load_byte_msh() helper function
inside bpf_jit.S.
New JIT removes the helper and does it explicitly, so ldx_msh cost
is the same for both JITs, but generated code looks longer.
New JIT has 4 registers to save, so prologue/epilogue are larger,
but the cost is within noise on x64.
Old JIT checks whether first insn clears A and if not emits 'xor %eax,%eax'.
New JIT clears %rax unconditionally.
2. old BPF JIT doesn't support ANC_NLATTR, ANC_PAY_OFFSET, ANC_RANDOM
extensions. New JIT supports all BPF extensions.
Performance of such filters improves 2-4 times depending on a filter.
The longer the filter the higher performance gain.
Synthetic benchmarks with many ancillary loads see 20x speedup
which seems to be the maximum gain from JIT
Notes:
. net.core.bpf_jit_enable=2 + tools/net/bpf_jit_disasm is still functional
and can be used to see generated assembler
. there are two jit_compile() functions and code flow for classic filters is:
sk_attach_filter() - load classic BPF
bpf_jit_compile() - try to JIT from classic BPF
sk_convert_filter() - convert classic to internal
bpf_int_jit_compile() - JIT from internal BPF
seccomp and tracing filters will just call bpf_int_jit_compile()
Signed-off-by: Alexei Starovoitov <ast@plumgrid.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
2014-05-14 02:50:46 +00:00
|
|
|
}
|
|
|
|
|
bpf, x86: Fix encoding for lower 8-bit registers in BPF_STX BPF_B
This patch fixes an encoding bug in emit_stx for BPF_B when the source
register is BPF_REG_FP.
The current implementation for BPF_STX BPF_B in emit_stx saves one REX
byte when the operands can be encoded using Mod-R/M alone. The lower 8
bits of registers %rax, %rbx, %rcx, and %rdx can be accessed without using
a REX prefix via %al, %bl, %cl, and %dl, respectively. Other registers,
(e.g., %rsi, %rdi, %rbp, %rsp) require a REX prefix to use their 8-bit
equivalents (%sil, %dil, %bpl, %spl).
The current code checks if the source for BPF_STX BPF_B is BPF_REG_1
or BPF_REG_2 (which map to %rdi and %rsi), in which case it emits the
required REX prefix. However, it misses the case when the source is
BPF_REG_FP (mapped to %rbp).
The result is that BPF_STX BPF_B with BPF_REG_FP as the source operand
will read from register %ch instead of the correct %bpl. This patch fixes
the problem by fixing and refactoring the check on which registers need
the extra REX byte. Since no BPF registers map to %rsp, there is no need
to handle %spl.
Fixes: 622582786c9e0 ("net: filter: x86: internal BPF JIT")
Signed-off-by: Xi Wang <xi.wang@gmail.com>
Signed-off-by: Luke Nelson <luke.r.nels@gmail.com>
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Link: https://lore.kernel.org/bpf/20200418232655.23870-1-luke.r.nels@gmail.com
2020-04-18 23:26:53 +00:00
|
|
|
/*
|
|
|
|
* is_ereg_8l() == true if BPF register 'reg' is mapped to access x86-64
|
|
|
|
* lower 8-bit registers dil,sil,bpl,spl,r8b..r15b, which need extra byte
|
|
|
|
* of encoding. al,cl,dl,bl have simpler encoding.
|
|
|
|
*/
|
|
|
|
static bool is_ereg_8l(u32 reg)
|
|
|
|
{
|
|
|
|
return is_ereg(reg) ||
|
|
|
|
(1 << reg) & (BIT(BPF_REG_1) |
|
|
|
|
BIT(BPF_REG_2) |
|
|
|
|
BIT(BPF_REG_FP));
|
|
|
|
}
|
|
|
|
|
2018-01-20 00:24:35 +00:00
|
|
|
static bool is_axreg(u32 reg)
|
|
|
|
{
|
|
|
|
return reg == BPF_REG_0;
|
|
|
|
}
|
|
|
|
|
x86/bpf: Clean up non-standard comments, to make the code more readable
So by chance I looked into x86 assembly in arch/x86/net/bpf_jit_comp.c and
noticed the weird and inconsistent comment style it mistakenly learned from
the networking code:
/* Multi-line comment ...
* ... looks like this.
*/
Fix this to use the standard comment style specified in Documentation/CodingStyle
and used in arch/x86/ as well:
/*
* Multi-line comment ...
* ... looks like this.
*/
Also, to quote Linus's ... more explicit views about this:
http://article.gmane.org/gmane.linux.kernel.cryptoapi/21066
> But no, the networking code picked *none* of the above sane formats.
> Instead, it picked these two models that are just half-arsed
> shit-for-brains:
>
> (no)
> /* This is disgusting drug-induced
> * crap, and should die
> */
>
> (no-no-no)
> /* This is also very nasty
> * and visually unbalanced */
>
> Please. The networking code actually has the *worst* possible comment
> style. You can literally find that (no-no-no) style, which is just
> really horribly disgusting and worse than the otherwise fairly similar
> (d) in pretty much every way.
Also improve the comments and some other details while at it:
- Don't mix same-line and previous-line comment style on otherwise
identical code patterns within the same function,
- capitalize 'BPF' and x86 register names consistently,
- capitalize sentences consistently,
- instead of 'x64' use 'x86-64': x64 is a Microsoft specific term,
- use more consistent punctuation,
- use standard coding style in macros as well,
- fix typos and a few other minor details.
Consistent coding style is not optional, at least in arch/x86/.
No change in functionality.
( In case this commit causes conflicts with pending development code
I'll be glad to help resolve any conflicts! )
Acked-by: Thomas Gleixner <tglx@linutronix.de>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Borislav Petkov <bp@alien8.de>
Cc: H. Peter Anvin <hpa@zytor.com>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: David S. Miller <davem@davemloft.net>
Cc: Eric Dumazet <edumazet@google.com>
Cc: Daniel Borkmann <daniel@iogearbox.net>
Cc: Alexei Starovoitov <ast@fb.com>
Cc: Hideaki YOSHIFUJI <yoshfuji@linux-ipv6.org>
Cc: netdev@vger.kernel.org
Cc: linux-kernel@vger.kernel.org
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
2018-04-27 09:54:40 +00:00
|
|
|
/* Add modifiers if 'reg' maps to x86-64 registers R8..R15 */
|
2014-12-05 01:01:24 +00:00
|
|
|
static u8 add_1mod(u8 byte, u32 reg)
|
net: filter: x86: internal BPF JIT
Maps all internal BPF instructions into x86_64 instructions.
This patch replaces original BPF x64 JIT with internal BPF x64 JIT.
sysctl net.core.bpf_jit_enable is reused as on/off switch.
Performance:
1. old BPF JIT and internal BPF JIT generate equivalent x86_64 code.
No performance difference is observed for filters that were JIT-able before
Example assembler code for BPF filter "tcpdump port 22"
original BPF -> old JIT: original BPF -> internal BPF -> new JIT:
0: push %rbp 0: push %rbp
1: mov %rsp,%rbp 1: mov %rsp,%rbp
4: sub $0x60,%rsp 4: sub $0x228,%rsp
8: mov %rbx,-0x8(%rbp) b: mov %rbx,-0x228(%rbp) // prologue
12: mov %r13,-0x220(%rbp)
19: mov %r14,-0x218(%rbp)
20: mov %r15,-0x210(%rbp)
27: xor %eax,%eax // clear A
c: xor %ebx,%ebx 29: xor %r13,%r13 // clear X
e: mov 0x68(%rdi),%r9d 2c: mov 0x68(%rdi),%r9d
12: sub 0x6c(%rdi),%r9d 30: sub 0x6c(%rdi),%r9d
16: mov 0xd8(%rdi),%r8 34: mov 0xd8(%rdi),%r10
3b: mov %rdi,%rbx
1d: mov $0xc,%esi 3e: mov $0xc,%esi
22: callq 0xffffffffe1021e15 43: callq 0xffffffffe102bd75
27: cmp $0x86dd,%eax 48: cmp $0x86dd,%rax
2c: jne 0x0000000000000069 4f: jne 0x000000000000009a
2e: mov $0x14,%esi 51: mov $0x14,%esi
33: callq 0xffffffffe1021e31 56: callq 0xffffffffe102bd91
38: cmp $0x84,%eax 5b: cmp $0x84,%rax
3d: je 0x0000000000000049 62: je 0x0000000000000074
3f: cmp $0x6,%eax 64: cmp $0x6,%rax
42: je 0x0000000000000049 68: je 0x0000000000000074
44: cmp $0x11,%eax 6a: cmp $0x11,%rax
47: jne 0x00000000000000c6 6e: jne 0x0000000000000117
49: mov $0x36,%esi 74: mov $0x36,%esi
4e: callq 0xffffffffe1021e15 79: callq 0xffffffffe102bd75
53: cmp $0x16,%eax 7e: cmp $0x16,%rax
56: je 0x00000000000000bf 82: je 0x0000000000000110
58: mov $0x38,%esi 88: mov $0x38,%esi
5d: callq 0xffffffffe1021e15 8d: callq 0xffffffffe102bd75
62: cmp $0x16,%eax 92: cmp $0x16,%rax
65: je 0x00000000000000bf 96: je 0x0000000000000110
67: jmp 0x00000000000000c6 98: jmp 0x0000000000000117
69: cmp $0x800,%eax 9a: cmp $0x800,%rax
6e: jne 0x00000000000000c6 a1: jne 0x0000000000000117
70: mov $0x17,%esi a3: mov $0x17,%esi
75: callq 0xffffffffe1021e31 a8: callq 0xffffffffe102bd91
7a: cmp $0x84,%eax ad: cmp $0x84,%rax
7f: je 0x000000000000008b b4: je 0x00000000000000c2
81: cmp $0x6,%eax b6: cmp $0x6,%rax
84: je 0x000000000000008b ba: je 0x00000000000000c2
86: cmp $0x11,%eax bc: cmp $0x11,%rax
89: jne 0x00000000000000c6 c0: jne 0x0000000000000117
8b: mov $0x14,%esi c2: mov $0x14,%esi
90: callq 0xffffffffe1021e15 c7: callq 0xffffffffe102bd75
95: test $0x1fff,%ax cc: test $0x1fff,%rax
99: jne 0x00000000000000c6 d3: jne 0x0000000000000117
d5: mov %rax,%r14
9b: mov $0xe,%esi d8: mov $0xe,%esi
a0: callq 0xffffffffe1021e44 dd: callq 0xffffffffe102bd91 // MSH
e2: and $0xf,%eax
e5: shl $0x2,%eax
e8: mov %rax,%r13
eb: mov %r14,%rax
ee: mov %r13,%rsi
a5: lea 0xe(%rbx),%esi f1: add $0xe,%esi
a8: callq 0xffffffffe1021e0d f4: callq 0xffffffffe102bd6d
ad: cmp $0x16,%eax f9: cmp $0x16,%rax
b0: je 0x00000000000000bf fd: je 0x0000000000000110
ff: mov %r13,%rsi
b2: lea 0x10(%rbx),%esi 102: add $0x10,%esi
b5: callq 0xffffffffe1021e0d 105: callq 0xffffffffe102bd6d
ba: cmp $0x16,%eax 10a: cmp $0x16,%rax
bd: jne 0x00000000000000c6 10e: jne 0x0000000000000117
bf: mov $0xffff,%eax 110: mov $0xffff,%eax
c4: jmp 0x00000000000000c8 115: jmp 0x000000000000011c
c6: xor %eax,%eax 117: mov $0x0,%eax
c8: mov -0x8(%rbp),%rbx 11c: mov -0x228(%rbp),%rbx // epilogue
cc: leaveq 123: mov -0x220(%rbp),%r13
cd: retq 12a: mov -0x218(%rbp),%r14
131: mov -0x210(%rbp),%r15
138: leaveq
139: retq
On fully cached SKBs both JITed functions take 12 nsec to execute.
BPF interpreter executes the program in 30 nsec.
The difference in generated assembler is due to the following:
Old BPF imlements LDX_MSH instruction via sk_load_byte_msh() helper function
inside bpf_jit.S.
New JIT removes the helper and does it explicitly, so ldx_msh cost
is the same for both JITs, but generated code looks longer.
New JIT has 4 registers to save, so prologue/epilogue are larger,
but the cost is within noise on x64.
Old JIT checks whether first insn clears A and if not emits 'xor %eax,%eax'.
New JIT clears %rax unconditionally.
2. old BPF JIT doesn't support ANC_NLATTR, ANC_PAY_OFFSET, ANC_RANDOM
extensions. New JIT supports all BPF extensions.
Performance of such filters improves 2-4 times depending on a filter.
The longer the filter the higher performance gain.
Synthetic benchmarks with many ancillary loads see 20x speedup
which seems to be the maximum gain from JIT
Notes:
. net.core.bpf_jit_enable=2 + tools/net/bpf_jit_disasm is still functional
and can be used to see generated assembler
. there are two jit_compile() functions and code flow for classic filters is:
sk_attach_filter() - load classic BPF
bpf_jit_compile() - try to JIT from classic BPF
sk_convert_filter() - convert classic to internal
bpf_int_jit_compile() - JIT from internal BPF
seccomp and tracing filters will just call bpf_int_jit_compile()
Signed-off-by: Alexei Starovoitov <ast@plumgrid.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
2014-05-14 02:50:46 +00:00
|
|
|
{
|
|
|
|
if (is_ereg(reg))
|
|
|
|
byte |= 1;
|
|
|
|
return byte;
|
|
|
|
}
|
|
|
|
|
2014-12-05 01:01:24 +00:00
|
|
|
static u8 add_2mod(u8 byte, u32 r1, u32 r2)
|
net: filter: x86: internal BPF JIT
Maps all internal BPF instructions into x86_64 instructions.
This patch replaces original BPF x64 JIT with internal BPF x64 JIT.
sysctl net.core.bpf_jit_enable is reused as on/off switch.
Performance:
1. old BPF JIT and internal BPF JIT generate equivalent x86_64 code.
No performance difference is observed for filters that were JIT-able before
Example assembler code for BPF filter "tcpdump port 22"
original BPF -> old JIT: original BPF -> internal BPF -> new JIT:
0: push %rbp 0: push %rbp
1: mov %rsp,%rbp 1: mov %rsp,%rbp
4: sub $0x60,%rsp 4: sub $0x228,%rsp
8: mov %rbx,-0x8(%rbp) b: mov %rbx,-0x228(%rbp) // prologue
12: mov %r13,-0x220(%rbp)
19: mov %r14,-0x218(%rbp)
20: mov %r15,-0x210(%rbp)
27: xor %eax,%eax // clear A
c: xor %ebx,%ebx 29: xor %r13,%r13 // clear X
e: mov 0x68(%rdi),%r9d 2c: mov 0x68(%rdi),%r9d
12: sub 0x6c(%rdi),%r9d 30: sub 0x6c(%rdi),%r9d
16: mov 0xd8(%rdi),%r8 34: mov 0xd8(%rdi),%r10
3b: mov %rdi,%rbx
1d: mov $0xc,%esi 3e: mov $0xc,%esi
22: callq 0xffffffffe1021e15 43: callq 0xffffffffe102bd75
27: cmp $0x86dd,%eax 48: cmp $0x86dd,%rax
2c: jne 0x0000000000000069 4f: jne 0x000000000000009a
2e: mov $0x14,%esi 51: mov $0x14,%esi
33: callq 0xffffffffe1021e31 56: callq 0xffffffffe102bd91
38: cmp $0x84,%eax 5b: cmp $0x84,%rax
3d: je 0x0000000000000049 62: je 0x0000000000000074
3f: cmp $0x6,%eax 64: cmp $0x6,%rax
42: je 0x0000000000000049 68: je 0x0000000000000074
44: cmp $0x11,%eax 6a: cmp $0x11,%rax
47: jne 0x00000000000000c6 6e: jne 0x0000000000000117
49: mov $0x36,%esi 74: mov $0x36,%esi
4e: callq 0xffffffffe1021e15 79: callq 0xffffffffe102bd75
53: cmp $0x16,%eax 7e: cmp $0x16,%rax
56: je 0x00000000000000bf 82: je 0x0000000000000110
58: mov $0x38,%esi 88: mov $0x38,%esi
5d: callq 0xffffffffe1021e15 8d: callq 0xffffffffe102bd75
62: cmp $0x16,%eax 92: cmp $0x16,%rax
65: je 0x00000000000000bf 96: je 0x0000000000000110
67: jmp 0x00000000000000c6 98: jmp 0x0000000000000117
69: cmp $0x800,%eax 9a: cmp $0x800,%rax
6e: jne 0x00000000000000c6 a1: jne 0x0000000000000117
70: mov $0x17,%esi a3: mov $0x17,%esi
75: callq 0xffffffffe1021e31 a8: callq 0xffffffffe102bd91
7a: cmp $0x84,%eax ad: cmp $0x84,%rax
7f: je 0x000000000000008b b4: je 0x00000000000000c2
81: cmp $0x6,%eax b6: cmp $0x6,%rax
84: je 0x000000000000008b ba: je 0x00000000000000c2
86: cmp $0x11,%eax bc: cmp $0x11,%rax
89: jne 0x00000000000000c6 c0: jne 0x0000000000000117
8b: mov $0x14,%esi c2: mov $0x14,%esi
90: callq 0xffffffffe1021e15 c7: callq 0xffffffffe102bd75
95: test $0x1fff,%ax cc: test $0x1fff,%rax
99: jne 0x00000000000000c6 d3: jne 0x0000000000000117
d5: mov %rax,%r14
9b: mov $0xe,%esi d8: mov $0xe,%esi
a0: callq 0xffffffffe1021e44 dd: callq 0xffffffffe102bd91 // MSH
e2: and $0xf,%eax
e5: shl $0x2,%eax
e8: mov %rax,%r13
eb: mov %r14,%rax
ee: mov %r13,%rsi
a5: lea 0xe(%rbx),%esi f1: add $0xe,%esi
a8: callq 0xffffffffe1021e0d f4: callq 0xffffffffe102bd6d
ad: cmp $0x16,%eax f9: cmp $0x16,%rax
b0: je 0x00000000000000bf fd: je 0x0000000000000110
ff: mov %r13,%rsi
b2: lea 0x10(%rbx),%esi 102: add $0x10,%esi
b5: callq 0xffffffffe1021e0d 105: callq 0xffffffffe102bd6d
ba: cmp $0x16,%eax 10a: cmp $0x16,%rax
bd: jne 0x00000000000000c6 10e: jne 0x0000000000000117
bf: mov $0xffff,%eax 110: mov $0xffff,%eax
c4: jmp 0x00000000000000c8 115: jmp 0x000000000000011c
c6: xor %eax,%eax 117: mov $0x0,%eax
c8: mov -0x8(%rbp),%rbx 11c: mov -0x228(%rbp),%rbx // epilogue
cc: leaveq 123: mov -0x220(%rbp),%r13
cd: retq 12a: mov -0x218(%rbp),%r14
131: mov -0x210(%rbp),%r15
138: leaveq
139: retq
On fully cached SKBs both JITed functions take 12 nsec to execute.
BPF interpreter executes the program in 30 nsec.
The difference in generated assembler is due to the following:
Old BPF imlements LDX_MSH instruction via sk_load_byte_msh() helper function
inside bpf_jit.S.
New JIT removes the helper and does it explicitly, so ldx_msh cost
is the same for both JITs, but generated code looks longer.
New JIT has 4 registers to save, so prologue/epilogue are larger,
but the cost is within noise on x64.
Old JIT checks whether first insn clears A and if not emits 'xor %eax,%eax'.
New JIT clears %rax unconditionally.
2. old BPF JIT doesn't support ANC_NLATTR, ANC_PAY_OFFSET, ANC_RANDOM
extensions. New JIT supports all BPF extensions.
Performance of such filters improves 2-4 times depending on a filter.
The longer the filter the higher performance gain.
Synthetic benchmarks with many ancillary loads see 20x speedup
which seems to be the maximum gain from JIT
Notes:
. net.core.bpf_jit_enable=2 + tools/net/bpf_jit_disasm is still functional
and can be used to see generated assembler
. there are two jit_compile() functions and code flow for classic filters is:
sk_attach_filter() - load classic BPF
bpf_jit_compile() - try to JIT from classic BPF
sk_convert_filter() - convert classic to internal
bpf_int_jit_compile() - JIT from internal BPF
seccomp and tracing filters will just call bpf_int_jit_compile()
Signed-off-by: Alexei Starovoitov <ast@plumgrid.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
2014-05-14 02:50:46 +00:00
|
|
|
{
|
|
|
|
if (is_ereg(r1))
|
|
|
|
byte |= 1;
|
|
|
|
if (is_ereg(r2))
|
|
|
|
byte |= 4;
|
|
|
|
return byte;
|
|
|
|
}
|
|
|
|
|
2024-03-08 01:08:01 +00:00
|
|
|
static u8 add_3mod(u8 byte, u32 r1, u32 r2, u32 index)
|
|
|
|
{
|
|
|
|
if (is_ereg(r1))
|
|
|
|
byte |= 1;
|
|
|
|
if (is_ereg(index))
|
|
|
|
byte |= 2;
|
|
|
|
if (is_ereg(r2))
|
|
|
|
byte |= 4;
|
|
|
|
return byte;
|
|
|
|
}
|
|
|
|
|
x86/bpf: Clean up non-standard comments, to make the code more readable
So by chance I looked into x86 assembly in arch/x86/net/bpf_jit_comp.c and
noticed the weird and inconsistent comment style it mistakenly learned from
the networking code:
/* Multi-line comment ...
* ... looks like this.
*/
Fix this to use the standard comment style specified in Documentation/CodingStyle
and used in arch/x86/ as well:
/*
* Multi-line comment ...
* ... looks like this.
*/
Also, to quote Linus's ... more explicit views about this:
http://article.gmane.org/gmane.linux.kernel.cryptoapi/21066
> But no, the networking code picked *none* of the above sane formats.
> Instead, it picked these two models that are just half-arsed
> shit-for-brains:
>
> (no)
> /* This is disgusting drug-induced
> * crap, and should die
> */
>
> (no-no-no)
> /* This is also very nasty
> * and visually unbalanced */
>
> Please. The networking code actually has the *worst* possible comment
> style. You can literally find that (no-no-no) style, which is just
> really horribly disgusting and worse than the otherwise fairly similar
> (d) in pretty much every way.
Also improve the comments and some other details while at it:
- Don't mix same-line and previous-line comment style on otherwise
identical code patterns within the same function,
- capitalize 'BPF' and x86 register names consistently,
- capitalize sentences consistently,
- instead of 'x64' use 'x86-64': x64 is a Microsoft specific term,
- use more consistent punctuation,
- use standard coding style in macros as well,
- fix typos and a few other minor details.
Consistent coding style is not optional, at least in arch/x86/.
No change in functionality.
( In case this commit causes conflicts with pending development code
I'll be glad to help resolve any conflicts! )
Acked-by: Thomas Gleixner <tglx@linutronix.de>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Borislav Petkov <bp@alien8.de>
Cc: H. Peter Anvin <hpa@zytor.com>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: David S. Miller <davem@davemloft.net>
Cc: Eric Dumazet <edumazet@google.com>
Cc: Daniel Borkmann <daniel@iogearbox.net>
Cc: Alexei Starovoitov <ast@fb.com>
Cc: Hideaki YOSHIFUJI <yoshfuji@linux-ipv6.org>
Cc: netdev@vger.kernel.org
Cc: linux-kernel@vger.kernel.org
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
2018-04-27 09:54:40 +00:00
|
|
|
/* Encode 'dst_reg' register into x86-64 opcode 'byte' */
|
2014-12-05 01:01:24 +00:00
|
|
|
static u8 add_1reg(u8 byte, u32 dst_reg)
|
net: filter: x86: internal BPF JIT
Maps all internal BPF instructions into x86_64 instructions.
This patch replaces original BPF x64 JIT with internal BPF x64 JIT.
sysctl net.core.bpf_jit_enable is reused as on/off switch.
Performance:
1. old BPF JIT and internal BPF JIT generate equivalent x86_64 code.
No performance difference is observed for filters that were JIT-able before
Example assembler code for BPF filter "tcpdump port 22"
original BPF -> old JIT: original BPF -> internal BPF -> new JIT:
0: push %rbp 0: push %rbp
1: mov %rsp,%rbp 1: mov %rsp,%rbp
4: sub $0x60,%rsp 4: sub $0x228,%rsp
8: mov %rbx,-0x8(%rbp) b: mov %rbx,-0x228(%rbp) // prologue
12: mov %r13,-0x220(%rbp)
19: mov %r14,-0x218(%rbp)
20: mov %r15,-0x210(%rbp)
27: xor %eax,%eax // clear A
c: xor %ebx,%ebx 29: xor %r13,%r13 // clear X
e: mov 0x68(%rdi),%r9d 2c: mov 0x68(%rdi),%r9d
12: sub 0x6c(%rdi),%r9d 30: sub 0x6c(%rdi),%r9d
16: mov 0xd8(%rdi),%r8 34: mov 0xd8(%rdi),%r10
3b: mov %rdi,%rbx
1d: mov $0xc,%esi 3e: mov $0xc,%esi
22: callq 0xffffffffe1021e15 43: callq 0xffffffffe102bd75
27: cmp $0x86dd,%eax 48: cmp $0x86dd,%rax
2c: jne 0x0000000000000069 4f: jne 0x000000000000009a
2e: mov $0x14,%esi 51: mov $0x14,%esi
33: callq 0xffffffffe1021e31 56: callq 0xffffffffe102bd91
38: cmp $0x84,%eax 5b: cmp $0x84,%rax
3d: je 0x0000000000000049 62: je 0x0000000000000074
3f: cmp $0x6,%eax 64: cmp $0x6,%rax
42: je 0x0000000000000049 68: je 0x0000000000000074
44: cmp $0x11,%eax 6a: cmp $0x11,%rax
47: jne 0x00000000000000c6 6e: jne 0x0000000000000117
49: mov $0x36,%esi 74: mov $0x36,%esi
4e: callq 0xffffffffe1021e15 79: callq 0xffffffffe102bd75
53: cmp $0x16,%eax 7e: cmp $0x16,%rax
56: je 0x00000000000000bf 82: je 0x0000000000000110
58: mov $0x38,%esi 88: mov $0x38,%esi
5d: callq 0xffffffffe1021e15 8d: callq 0xffffffffe102bd75
62: cmp $0x16,%eax 92: cmp $0x16,%rax
65: je 0x00000000000000bf 96: je 0x0000000000000110
67: jmp 0x00000000000000c6 98: jmp 0x0000000000000117
69: cmp $0x800,%eax 9a: cmp $0x800,%rax
6e: jne 0x00000000000000c6 a1: jne 0x0000000000000117
70: mov $0x17,%esi a3: mov $0x17,%esi
75: callq 0xffffffffe1021e31 a8: callq 0xffffffffe102bd91
7a: cmp $0x84,%eax ad: cmp $0x84,%rax
7f: je 0x000000000000008b b4: je 0x00000000000000c2
81: cmp $0x6,%eax b6: cmp $0x6,%rax
84: je 0x000000000000008b ba: je 0x00000000000000c2
86: cmp $0x11,%eax bc: cmp $0x11,%rax
89: jne 0x00000000000000c6 c0: jne 0x0000000000000117
8b: mov $0x14,%esi c2: mov $0x14,%esi
90: callq 0xffffffffe1021e15 c7: callq 0xffffffffe102bd75
95: test $0x1fff,%ax cc: test $0x1fff,%rax
99: jne 0x00000000000000c6 d3: jne 0x0000000000000117
d5: mov %rax,%r14
9b: mov $0xe,%esi d8: mov $0xe,%esi
a0: callq 0xffffffffe1021e44 dd: callq 0xffffffffe102bd91 // MSH
e2: and $0xf,%eax
e5: shl $0x2,%eax
e8: mov %rax,%r13
eb: mov %r14,%rax
ee: mov %r13,%rsi
a5: lea 0xe(%rbx),%esi f1: add $0xe,%esi
a8: callq 0xffffffffe1021e0d f4: callq 0xffffffffe102bd6d
ad: cmp $0x16,%eax f9: cmp $0x16,%rax
b0: je 0x00000000000000bf fd: je 0x0000000000000110
ff: mov %r13,%rsi
b2: lea 0x10(%rbx),%esi 102: add $0x10,%esi
b5: callq 0xffffffffe1021e0d 105: callq 0xffffffffe102bd6d
ba: cmp $0x16,%eax 10a: cmp $0x16,%rax
bd: jne 0x00000000000000c6 10e: jne 0x0000000000000117
bf: mov $0xffff,%eax 110: mov $0xffff,%eax
c4: jmp 0x00000000000000c8 115: jmp 0x000000000000011c
c6: xor %eax,%eax 117: mov $0x0,%eax
c8: mov -0x8(%rbp),%rbx 11c: mov -0x228(%rbp),%rbx // epilogue
cc: leaveq 123: mov -0x220(%rbp),%r13
cd: retq 12a: mov -0x218(%rbp),%r14
131: mov -0x210(%rbp),%r15
138: leaveq
139: retq
On fully cached SKBs both JITed functions take 12 nsec to execute.
BPF interpreter executes the program in 30 nsec.
The difference in generated assembler is due to the following:
Old BPF imlements LDX_MSH instruction via sk_load_byte_msh() helper function
inside bpf_jit.S.
New JIT removes the helper and does it explicitly, so ldx_msh cost
is the same for both JITs, but generated code looks longer.
New JIT has 4 registers to save, so prologue/epilogue are larger,
but the cost is within noise on x64.
Old JIT checks whether first insn clears A and if not emits 'xor %eax,%eax'.
New JIT clears %rax unconditionally.
2. old BPF JIT doesn't support ANC_NLATTR, ANC_PAY_OFFSET, ANC_RANDOM
extensions. New JIT supports all BPF extensions.
Performance of such filters improves 2-4 times depending on a filter.
The longer the filter the higher performance gain.
Synthetic benchmarks with many ancillary loads see 20x speedup
which seems to be the maximum gain from JIT
Notes:
. net.core.bpf_jit_enable=2 + tools/net/bpf_jit_disasm is still functional
and can be used to see generated assembler
. there are two jit_compile() functions and code flow for classic filters is:
sk_attach_filter() - load classic BPF
bpf_jit_compile() - try to JIT from classic BPF
sk_convert_filter() - convert classic to internal
bpf_int_jit_compile() - JIT from internal BPF
seccomp and tracing filters will just call bpf_int_jit_compile()
Signed-off-by: Alexei Starovoitov <ast@plumgrid.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
2014-05-14 02:50:46 +00:00
|
|
|
{
|
2014-06-06 21:46:06 +00:00
|
|
|
return byte + reg2hex[dst_reg];
|
net: filter: x86: internal BPF JIT
Maps all internal BPF instructions into x86_64 instructions.
This patch replaces original BPF x64 JIT with internal BPF x64 JIT.
sysctl net.core.bpf_jit_enable is reused as on/off switch.
Performance:
1. old BPF JIT and internal BPF JIT generate equivalent x86_64 code.
No performance difference is observed for filters that were JIT-able before
Example assembler code for BPF filter "tcpdump port 22"
original BPF -> old JIT: original BPF -> internal BPF -> new JIT:
0: push %rbp 0: push %rbp
1: mov %rsp,%rbp 1: mov %rsp,%rbp
4: sub $0x60,%rsp 4: sub $0x228,%rsp
8: mov %rbx,-0x8(%rbp) b: mov %rbx,-0x228(%rbp) // prologue
12: mov %r13,-0x220(%rbp)
19: mov %r14,-0x218(%rbp)
20: mov %r15,-0x210(%rbp)
27: xor %eax,%eax // clear A
c: xor %ebx,%ebx 29: xor %r13,%r13 // clear X
e: mov 0x68(%rdi),%r9d 2c: mov 0x68(%rdi),%r9d
12: sub 0x6c(%rdi),%r9d 30: sub 0x6c(%rdi),%r9d
16: mov 0xd8(%rdi),%r8 34: mov 0xd8(%rdi),%r10
3b: mov %rdi,%rbx
1d: mov $0xc,%esi 3e: mov $0xc,%esi
22: callq 0xffffffffe1021e15 43: callq 0xffffffffe102bd75
27: cmp $0x86dd,%eax 48: cmp $0x86dd,%rax
2c: jne 0x0000000000000069 4f: jne 0x000000000000009a
2e: mov $0x14,%esi 51: mov $0x14,%esi
33: callq 0xffffffffe1021e31 56: callq 0xffffffffe102bd91
38: cmp $0x84,%eax 5b: cmp $0x84,%rax
3d: je 0x0000000000000049 62: je 0x0000000000000074
3f: cmp $0x6,%eax 64: cmp $0x6,%rax
42: je 0x0000000000000049 68: je 0x0000000000000074
44: cmp $0x11,%eax 6a: cmp $0x11,%rax
47: jne 0x00000000000000c6 6e: jne 0x0000000000000117
49: mov $0x36,%esi 74: mov $0x36,%esi
4e: callq 0xffffffffe1021e15 79: callq 0xffffffffe102bd75
53: cmp $0x16,%eax 7e: cmp $0x16,%rax
56: je 0x00000000000000bf 82: je 0x0000000000000110
58: mov $0x38,%esi 88: mov $0x38,%esi
5d: callq 0xffffffffe1021e15 8d: callq 0xffffffffe102bd75
62: cmp $0x16,%eax 92: cmp $0x16,%rax
65: je 0x00000000000000bf 96: je 0x0000000000000110
67: jmp 0x00000000000000c6 98: jmp 0x0000000000000117
69: cmp $0x800,%eax 9a: cmp $0x800,%rax
6e: jne 0x00000000000000c6 a1: jne 0x0000000000000117
70: mov $0x17,%esi a3: mov $0x17,%esi
75: callq 0xffffffffe1021e31 a8: callq 0xffffffffe102bd91
7a: cmp $0x84,%eax ad: cmp $0x84,%rax
7f: je 0x000000000000008b b4: je 0x00000000000000c2
81: cmp $0x6,%eax b6: cmp $0x6,%rax
84: je 0x000000000000008b ba: je 0x00000000000000c2
86: cmp $0x11,%eax bc: cmp $0x11,%rax
89: jne 0x00000000000000c6 c0: jne 0x0000000000000117
8b: mov $0x14,%esi c2: mov $0x14,%esi
90: callq 0xffffffffe1021e15 c7: callq 0xffffffffe102bd75
95: test $0x1fff,%ax cc: test $0x1fff,%rax
99: jne 0x00000000000000c6 d3: jne 0x0000000000000117
d5: mov %rax,%r14
9b: mov $0xe,%esi d8: mov $0xe,%esi
a0: callq 0xffffffffe1021e44 dd: callq 0xffffffffe102bd91 // MSH
e2: and $0xf,%eax
e5: shl $0x2,%eax
e8: mov %rax,%r13
eb: mov %r14,%rax
ee: mov %r13,%rsi
a5: lea 0xe(%rbx),%esi f1: add $0xe,%esi
a8: callq 0xffffffffe1021e0d f4: callq 0xffffffffe102bd6d
ad: cmp $0x16,%eax f9: cmp $0x16,%rax
b0: je 0x00000000000000bf fd: je 0x0000000000000110
ff: mov %r13,%rsi
b2: lea 0x10(%rbx),%esi 102: add $0x10,%esi
b5: callq 0xffffffffe1021e0d 105: callq 0xffffffffe102bd6d
ba: cmp $0x16,%eax 10a: cmp $0x16,%rax
bd: jne 0x00000000000000c6 10e: jne 0x0000000000000117
bf: mov $0xffff,%eax 110: mov $0xffff,%eax
c4: jmp 0x00000000000000c8 115: jmp 0x000000000000011c
c6: xor %eax,%eax 117: mov $0x0,%eax
c8: mov -0x8(%rbp),%rbx 11c: mov -0x228(%rbp),%rbx // epilogue
cc: leaveq 123: mov -0x220(%rbp),%r13
cd: retq 12a: mov -0x218(%rbp),%r14
131: mov -0x210(%rbp),%r15
138: leaveq
139: retq
On fully cached SKBs both JITed functions take 12 nsec to execute.
BPF interpreter executes the program in 30 nsec.
The difference in generated assembler is due to the following:
Old BPF imlements LDX_MSH instruction via sk_load_byte_msh() helper function
inside bpf_jit.S.
New JIT removes the helper and does it explicitly, so ldx_msh cost
is the same for both JITs, but generated code looks longer.
New JIT has 4 registers to save, so prologue/epilogue are larger,
but the cost is within noise on x64.
Old JIT checks whether first insn clears A and if not emits 'xor %eax,%eax'.
New JIT clears %rax unconditionally.
2. old BPF JIT doesn't support ANC_NLATTR, ANC_PAY_OFFSET, ANC_RANDOM
extensions. New JIT supports all BPF extensions.
Performance of such filters improves 2-4 times depending on a filter.
The longer the filter the higher performance gain.
Synthetic benchmarks with many ancillary loads see 20x speedup
which seems to be the maximum gain from JIT
Notes:
. net.core.bpf_jit_enable=2 + tools/net/bpf_jit_disasm is still functional
and can be used to see generated assembler
. there are two jit_compile() functions and code flow for classic filters is:
sk_attach_filter() - load classic BPF
bpf_jit_compile() - try to JIT from classic BPF
sk_convert_filter() - convert classic to internal
bpf_int_jit_compile() - JIT from internal BPF
seccomp and tracing filters will just call bpf_int_jit_compile()
Signed-off-by: Alexei Starovoitov <ast@plumgrid.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
2014-05-14 02:50:46 +00:00
|
|
|
}
|
|
|
|
|
x86/bpf: Clean up non-standard comments, to make the code more readable
So by chance I looked into x86 assembly in arch/x86/net/bpf_jit_comp.c and
noticed the weird and inconsistent comment style it mistakenly learned from
the networking code:
/* Multi-line comment ...
* ... looks like this.
*/
Fix this to use the standard comment style specified in Documentation/CodingStyle
and used in arch/x86/ as well:
/*
* Multi-line comment ...
* ... looks like this.
*/
Also, to quote Linus's ... more explicit views about this:
http://article.gmane.org/gmane.linux.kernel.cryptoapi/21066
> But no, the networking code picked *none* of the above sane formats.
> Instead, it picked these two models that are just half-arsed
> shit-for-brains:
>
> (no)
> /* This is disgusting drug-induced
> * crap, and should die
> */
>
> (no-no-no)
> /* This is also very nasty
> * and visually unbalanced */
>
> Please. The networking code actually has the *worst* possible comment
> style. You can literally find that (no-no-no) style, which is just
> really horribly disgusting and worse than the otherwise fairly similar
> (d) in pretty much every way.
Also improve the comments and some other details while at it:
- Don't mix same-line and previous-line comment style on otherwise
identical code patterns within the same function,
- capitalize 'BPF' and x86 register names consistently,
- capitalize sentences consistently,
- instead of 'x64' use 'x86-64': x64 is a Microsoft specific term,
- use more consistent punctuation,
- use standard coding style in macros as well,
- fix typos and a few other minor details.
Consistent coding style is not optional, at least in arch/x86/.
No change in functionality.
( In case this commit causes conflicts with pending development code
I'll be glad to help resolve any conflicts! )
Acked-by: Thomas Gleixner <tglx@linutronix.de>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Borislav Petkov <bp@alien8.de>
Cc: H. Peter Anvin <hpa@zytor.com>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: David S. Miller <davem@davemloft.net>
Cc: Eric Dumazet <edumazet@google.com>
Cc: Daniel Borkmann <daniel@iogearbox.net>
Cc: Alexei Starovoitov <ast@fb.com>
Cc: Hideaki YOSHIFUJI <yoshfuji@linux-ipv6.org>
Cc: netdev@vger.kernel.org
Cc: linux-kernel@vger.kernel.org
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
2018-04-27 09:54:40 +00:00
|
|
|
/* Encode 'dst_reg' and 'src_reg' registers into x86-64 opcode 'byte' */
|
2014-12-05 01:01:24 +00:00
|
|
|
static u8 add_2reg(u8 byte, u32 dst_reg, u32 src_reg)
|
net: filter: x86: internal BPF JIT
Maps all internal BPF instructions into x86_64 instructions.
This patch replaces original BPF x64 JIT with internal BPF x64 JIT.
sysctl net.core.bpf_jit_enable is reused as on/off switch.
Performance:
1. old BPF JIT and internal BPF JIT generate equivalent x86_64 code.
No performance difference is observed for filters that were JIT-able before
Example assembler code for BPF filter "tcpdump port 22"
original BPF -> old JIT: original BPF -> internal BPF -> new JIT:
0: push %rbp 0: push %rbp
1: mov %rsp,%rbp 1: mov %rsp,%rbp
4: sub $0x60,%rsp 4: sub $0x228,%rsp
8: mov %rbx,-0x8(%rbp) b: mov %rbx,-0x228(%rbp) // prologue
12: mov %r13,-0x220(%rbp)
19: mov %r14,-0x218(%rbp)
20: mov %r15,-0x210(%rbp)
27: xor %eax,%eax // clear A
c: xor %ebx,%ebx 29: xor %r13,%r13 // clear X
e: mov 0x68(%rdi),%r9d 2c: mov 0x68(%rdi),%r9d
12: sub 0x6c(%rdi),%r9d 30: sub 0x6c(%rdi),%r9d
16: mov 0xd8(%rdi),%r8 34: mov 0xd8(%rdi),%r10
3b: mov %rdi,%rbx
1d: mov $0xc,%esi 3e: mov $0xc,%esi
22: callq 0xffffffffe1021e15 43: callq 0xffffffffe102bd75
27: cmp $0x86dd,%eax 48: cmp $0x86dd,%rax
2c: jne 0x0000000000000069 4f: jne 0x000000000000009a
2e: mov $0x14,%esi 51: mov $0x14,%esi
33: callq 0xffffffffe1021e31 56: callq 0xffffffffe102bd91
38: cmp $0x84,%eax 5b: cmp $0x84,%rax
3d: je 0x0000000000000049 62: je 0x0000000000000074
3f: cmp $0x6,%eax 64: cmp $0x6,%rax
42: je 0x0000000000000049 68: je 0x0000000000000074
44: cmp $0x11,%eax 6a: cmp $0x11,%rax
47: jne 0x00000000000000c6 6e: jne 0x0000000000000117
49: mov $0x36,%esi 74: mov $0x36,%esi
4e: callq 0xffffffffe1021e15 79: callq 0xffffffffe102bd75
53: cmp $0x16,%eax 7e: cmp $0x16,%rax
56: je 0x00000000000000bf 82: je 0x0000000000000110
58: mov $0x38,%esi 88: mov $0x38,%esi
5d: callq 0xffffffffe1021e15 8d: callq 0xffffffffe102bd75
62: cmp $0x16,%eax 92: cmp $0x16,%rax
65: je 0x00000000000000bf 96: je 0x0000000000000110
67: jmp 0x00000000000000c6 98: jmp 0x0000000000000117
69: cmp $0x800,%eax 9a: cmp $0x800,%rax
6e: jne 0x00000000000000c6 a1: jne 0x0000000000000117
70: mov $0x17,%esi a3: mov $0x17,%esi
75: callq 0xffffffffe1021e31 a8: callq 0xffffffffe102bd91
7a: cmp $0x84,%eax ad: cmp $0x84,%rax
7f: je 0x000000000000008b b4: je 0x00000000000000c2
81: cmp $0x6,%eax b6: cmp $0x6,%rax
84: je 0x000000000000008b ba: je 0x00000000000000c2
86: cmp $0x11,%eax bc: cmp $0x11,%rax
89: jne 0x00000000000000c6 c0: jne 0x0000000000000117
8b: mov $0x14,%esi c2: mov $0x14,%esi
90: callq 0xffffffffe1021e15 c7: callq 0xffffffffe102bd75
95: test $0x1fff,%ax cc: test $0x1fff,%rax
99: jne 0x00000000000000c6 d3: jne 0x0000000000000117
d5: mov %rax,%r14
9b: mov $0xe,%esi d8: mov $0xe,%esi
a0: callq 0xffffffffe1021e44 dd: callq 0xffffffffe102bd91 // MSH
e2: and $0xf,%eax
e5: shl $0x2,%eax
e8: mov %rax,%r13
eb: mov %r14,%rax
ee: mov %r13,%rsi
a5: lea 0xe(%rbx),%esi f1: add $0xe,%esi
a8: callq 0xffffffffe1021e0d f4: callq 0xffffffffe102bd6d
ad: cmp $0x16,%eax f9: cmp $0x16,%rax
b0: je 0x00000000000000bf fd: je 0x0000000000000110
ff: mov %r13,%rsi
b2: lea 0x10(%rbx),%esi 102: add $0x10,%esi
b5: callq 0xffffffffe1021e0d 105: callq 0xffffffffe102bd6d
ba: cmp $0x16,%eax 10a: cmp $0x16,%rax
bd: jne 0x00000000000000c6 10e: jne 0x0000000000000117
bf: mov $0xffff,%eax 110: mov $0xffff,%eax
c4: jmp 0x00000000000000c8 115: jmp 0x000000000000011c
c6: xor %eax,%eax 117: mov $0x0,%eax
c8: mov -0x8(%rbp),%rbx 11c: mov -0x228(%rbp),%rbx // epilogue
cc: leaveq 123: mov -0x220(%rbp),%r13
cd: retq 12a: mov -0x218(%rbp),%r14
131: mov -0x210(%rbp),%r15
138: leaveq
139: retq
On fully cached SKBs both JITed functions take 12 nsec to execute.
BPF interpreter executes the program in 30 nsec.
The difference in generated assembler is due to the following:
Old BPF imlements LDX_MSH instruction via sk_load_byte_msh() helper function
inside bpf_jit.S.
New JIT removes the helper and does it explicitly, so ldx_msh cost
is the same for both JITs, but generated code looks longer.
New JIT has 4 registers to save, so prologue/epilogue are larger,
but the cost is within noise on x64.
Old JIT checks whether first insn clears A and if not emits 'xor %eax,%eax'.
New JIT clears %rax unconditionally.
2. old BPF JIT doesn't support ANC_NLATTR, ANC_PAY_OFFSET, ANC_RANDOM
extensions. New JIT supports all BPF extensions.
Performance of such filters improves 2-4 times depending on a filter.
The longer the filter the higher performance gain.
Synthetic benchmarks with many ancillary loads see 20x speedup
which seems to be the maximum gain from JIT
Notes:
. net.core.bpf_jit_enable=2 + tools/net/bpf_jit_disasm is still functional
and can be used to see generated assembler
. there are two jit_compile() functions and code flow for classic filters is:
sk_attach_filter() - load classic BPF
bpf_jit_compile() - try to JIT from classic BPF
sk_convert_filter() - convert classic to internal
bpf_int_jit_compile() - JIT from internal BPF
seccomp and tracing filters will just call bpf_int_jit_compile()
Signed-off-by: Alexei Starovoitov <ast@plumgrid.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
2014-05-14 02:50:46 +00:00
|
|
|
{
|
2014-06-06 21:46:06 +00:00
|
|
|
return byte + reg2hex[dst_reg] + (reg2hex[src_reg] << 3);
|
net: filter: x86: internal BPF JIT
Maps all internal BPF instructions into x86_64 instructions.
This patch replaces original BPF x64 JIT with internal BPF x64 JIT.
sysctl net.core.bpf_jit_enable is reused as on/off switch.
Performance:
1. old BPF JIT and internal BPF JIT generate equivalent x86_64 code.
No performance difference is observed for filters that were JIT-able before
Example assembler code for BPF filter "tcpdump port 22"
original BPF -> old JIT: original BPF -> internal BPF -> new JIT:
0: push %rbp 0: push %rbp
1: mov %rsp,%rbp 1: mov %rsp,%rbp
4: sub $0x60,%rsp 4: sub $0x228,%rsp
8: mov %rbx,-0x8(%rbp) b: mov %rbx,-0x228(%rbp) // prologue
12: mov %r13,-0x220(%rbp)
19: mov %r14,-0x218(%rbp)
20: mov %r15,-0x210(%rbp)
27: xor %eax,%eax // clear A
c: xor %ebx,%ebx 29: xor %r13,%r13 // clear X
e: mov 0x68(%rdi),%r9d 2c: mov 0x68(%rdi),%r9d
12: sub 0x6c(%rdi),%r9d 30: sub 0x6c(%rdi),%r9d
16: mov 0xd8(%rdi),%r8 34: mov 0xd8(%rdi),%r10
3b: mov %rdi,%rbx
1d: mov $0xc,%esi 3e: mov $0xc,%esi
22: callq 0xffffffffe1021e15 43: callq 0xffffffffe102bd75
27: cmp $0x86dd,%eax 48: cmp $0x86dd,%rax
2c: jne 0x0000000000000069 4f: jne 0x000000000000009a
2e: mov $0x14,%esi 51: mov $0x14,%esi
33: callq 0xffffffffe1021e31 56: callq 0xffffffffe102bd91
38: cmp $0x84,%eax 5b: cmp $0x84,%rax
3d: je 0x0000000000000049 62: je 0x0000000000000074
3f: cmp $0x6,%eax 64: cmp $0x6,%rax
42: je 0x0000000000000049 68: je 0x0000000000000074
44: cmp $0x11,%eax 6a: cmp $0x11,%rax
47: jne 0x00000000000000c6 6e: jne 0x0000000000000117
49: mov $0x36,%esi 74: mov $0x36,%esi
4e: callq 0xffffffffe1021e15 79: callq 0xffffffffe102bd75
53: cmp $0x16,%eax 7e: cmp $0x16,%rax
56: je 0x00000000000000bf 82: je 0x0000000000000110
58: mov $0x38,%esi 88: mov $0x38,%esi
5d: callq 0xffffffffe1021e15 8d: callq 0xffffffffe102bd75
62: cmp $0x16,%eax 92: cmp $0x16,%rax
65: je 0x00000000000000bf 96: je 0x0000000000000110
67: jmp 0x00000000000000c6 98: jmp 0x0000000000000117
69: cmp $0x800,%eax 9a: cmp $0x800,%rax
6e: jne 0x00000000000000c6 a1: jne 0x0000000000000117
70: mov $0x17,%esi a3: mov $0x17,%esi
75: callq 0xffffffffe1021e31 a8: callq 0xffffffffe102bd91
7a: cmp $0x84,%eax ad: cmp $0x84,%rax
7f: je 0x000000000000008b b4: je 0x00000000000000c2
81: cmp $0x6,%eax b6: cmp $0x6,%rax
84: je 0x000000000000008b ba: je 0x00000000000000c2
86: cmp $0x11,%eax bc: cmp $0x11,%rax
89: jne 0x00000000000000c6 c0: jne 0x0000000000000117
8b: mov $0x14,%esi c2: mov $0x14,%esi
90: callq 0xffffffffe1021e15 c7: callq 0xffffffffe102bd75
95: test $0x1fff,%ax cc: test $0x1fff,%rax
99: jne 0x00000000000000c6 d3: jne 0x0000000000000117
d5: mov %rax,%r14
9b: mov $0xe,%esi d8: mov $0xe,%esi
a0: callq 0xffffffffe1021e44 dd: callq 0xffffffffe102bd91 // MSH
e2: and $0xf,%eax
e5: shl $0x2,%eax
e8: mov %rax,%r13
eb: mov %r14,%rax
ee: mov %r13,%rsi
a5: lea 0xe(%rbx),%esi f1: add $0xe,%esi
a8: callq 0xffffffffe1021e0d f4: callq 0xffffffffe102bd6d
ad: cmp $0x16,%eax f9: cmp $0x16,%rax
b0: je 0x00000000000000bf fd: je 0x0000000000000110
ff: mov %r13,%rsi
b2: lea 0x10(%rbx),%esi 102: add $0x10,%esi
b5: callq 0xffffffffe1021e0d 105: callq 0xffffffffe102bd6d
ba: cmp $0x16,%eax 10a: cmp $0x16,%rax
bd: jne 0x00000000000000c6 10e: jne 0x0000000000000117
bf: mov $0xffff,%eax 110: mov $0xffff,%eax
c4: jmp 0x00000000000000c8 115: jmp 0x000000000000011c
c6: xor %eax,%eax 117: mov $0x0,%eax
c8: mov -0x8(%rbp),%rbx 11c: mov -0x228(%rbp),%rbx // epilogue
cc: leaveq 123: mov -0x220(%rbp),%r13
cd: retq 12a: mov -0x218(%rbp),%r14
131: mov -0x210(%rbp),%r15
138: leaveq
139: retq
On fully cached SKBs both JITed functions take 12 nsec to execute.
BPF interpreter executes the program in 30 nsec.
The difference in generated assembler is due to the following:
Old BPF imlements LDX_MSH instruction via sk_load_byte_msh() helper function
inside bpf_jit.S.
New JIT removes the helper and does it explicitly, so ldx_msh cost
is the same for both JITs, but generated code looks longer.
New JIT has 4 registers to save, so prologue/epilogue are larger,
but the cost is within noise on x64.
Old JIT checks whether first insn clears A and if not emits 'xor %eax,%eax'.
New JIT clears %rax unconditionally.
2. old BPF JIT doesn't support ANC_NLATTR, ANC_PAY_OFFSET, ANC_RANDOM
extensions. New JIT supports all BPF extensions.
Performance of such filters improves 2-4 times depending on a filter.
The longer the filter the higher performance gain.
Synthetic benchmarks with many ancillary loads see 20x speedup
which seems to be the maximum gain from JIT
Notes:
. net.core.bpf_jit_enable=2 + tools/net/bpf_jit_disasm is still functional
and can be used to see generated assembler
. there are two jit_compile() functions and code flow for classic filters is:
sk_attach_filter() - load classic BPF
bpf_jit_compile() - try to JIT from classic BPF
sk_convert_filter() - convert classic to internal
bpf_int_jit_compile() - JIT from internal BPF
seccomp and tracing filters will just call bpf_int_jit_compile()
Signed-off-by: Alexei Starovoitov <ast@plumgrid.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
2014-05-14 02:50:46 +00:00
|
|
|
}
|
|
|
|
|
2021-01-14 18:17:43 +00:00
|
|
|
/* Some 1-byte opcodes for binary ALU operations */
|
|
|
|
static u8 simple_alu_opcodes[] = {
|
|
|
|
[BPF_ADD] = 0x01,
|
|
|
|
[BPF_SUB] = 0x29,
|
|
|
|
[BPF_AND] = 0x21,
|
|
|
|
[BPF_OR] = 0x09,
|
|
|
|
[BPF_XOR] = 0x31,
|
|
|
|
[BPF_LSH] = 0xE0,
|
|
|
|
[BPF_RSH] = 0xE8,
|
|
|
|
[BPF_ARSH] = 0xF8,
|
|
|
|
};
|
|
|
|
|
2014-09-08 06:04:47 +00:00
|
|
|
static void jit_fill_hole(void *area, unsigned int size)
|
|
|
|
{
|
x86/bpf: Clean up non-standard comments, to make the code more readable
So by chance I looked into x86 assembly in arch/x86/net/bpf_jit_comp.c and
noticed the weird and inconsistent comment style it mistakenly learned from
the networking code:
/* Multi-line comment ...
* ... looks like this.
*/
Fix this to use the standard comment style specified in Documentation/CodingStyle
and used in arch/x86/ as well:
/*
* Multi-line comment ...
* ... looks like this.
*/
Also, to quote Linus's ... more explicit views about this:
http://article.gmane.org/gmane.linux.kernel.cryptoapi/21066
> But no, the networking code picked *none* of the above sane formats.
> Instead, it picked these two models that are just half-arsed
> shit-for-brains:
>
> (no)
> /* This is disgusting drug-induced
> * crap, and should die
> */
>
> (no-no-no)
> /* This is also very nasty
> * and visually unbalanced */
>
> Please. The networking code actually has the *worst* possible comment
> style. You can literally find that (no-no-no) style, which is just
> really horribly disgusting and worse than the otherwise fairly similar
> (d) in pretty much every way.
Also improve the comments and some other details while at it:
- Don't mix same-line and previous-line comment style on otherwise
identical code patterns within the same function,
- capitalize 'BPF' and x86 register names consistently,
- capitalize sentences consistently,
- instead of 'x64' use 'x86-64': x64 is a Microsoft specific term,
- use more consistent punctuation,
- use standard coding style in macros as well,
- fix typos and a few other minor details.
Consistent coding style is not optional, at least in arch/x86/.
No change in functionality.
( In case this commit causes conflicts with pending development code
I'll be glad to help resolve any conflicts! )
Acked-by: Thomas Gleixner <tglx@linutronix.de>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Borislav Petkov <bp@alien8.de>
Cc: H. Peter Anvin <hpa@zytor.com>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: David S. Miller <davem@davemloft.net>
Cc: Eric Dumazet <edumazet@google.com>
Cc: Daniel Borkmann <daniel@iogearbox.net>
Cc: Alexei Starovoitov <ast@fb.com>
Cc: Hideaki YOSHIFUJI <yoshfuji@linux-ipv6.org>
Cc: netdev@vger.kernel.org
Cc: linux-kernel@vger.kernel.org
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
2018-04-27 09:54:40 +00:00
|
|
|
/* Fill whole space with INT3 instructions */
|
2014-09-08 06:04:47 +00:00
|
|
|
memset(area, 0xcc, size);
|
|
|
|
}
|
|
|
|
|
2022-05-20 23:57:53 +00:00
|
|
|
int bpf_arch_text_invalidate(void *dst, size_t len)
|
|
|
|
{
|
|
|
|
return IS_ERR_OR_NULL(text_poke_set(dst, 0xcc, len));
|
|
|
|
}
|
|
|
|
|
2014-05-14 02:50:45 +00:00
|
|
|
struct jit_context {
|
x86/bpf: Clean up non-standard comments, to make the code more readable
So by chance I looked into x86 assembly in arch/x86/net/bpf_jit_comp.c and
noticed the weird and inconsistent comment style it mistakenly learned from
the networking code:
/* Multi-line comment ...
* ... looks like this.
*/
Fix this to use the standard comment style specified in Documentation/CodingStyle
and used in arch/x86/ as well:
/*
* Multi-line comment ...
* ... looks like this.
*/
Also, to quote Linus's ... more explicit views about this:
http://article.gmane.org/gmane.linux.kernel.cryptoapi/21066
> But no, the networking code picked *none* of the above sane formats.
> Instead, it picked these two models that are just half-arsed
> shit-for-brains:
>
> (no)
> /* This is disgusting drug-induced
> * crap, and should die
> */
>
> (no-no-no)
> /* This is also very nasty
> * and visually unbalanced */
>
> Please. The networking code actually has the *worst* possible comment
> style. You can literally find that (no-no-no) style, which is just
> really horribly disgusting and worse than the otherwise fairly similar
> (d) in pretty much every way.
Also improve the comments and some other details while at it:
- Don't mix same-line and previous-line comment style on otherwise
identical code patterns within the same function,
- capitalize 'BPF' and x86 register names consistently,
- capitalize sentences consistently,
- instead of 'x64' use 'x86-64': x64 is a Microsoft specific term,
- use more consistent punctuation,
- use standard coding style in macros as well,
- fix typos and a few other minor details.
Consistent coding style is not optional, at least in arch/x86/.
No change in functionality.
( In case this commit causes conflicts with pending development code
I'll be glad to help resolve any conflicts! )
Acked-by: Thomas Gleixner <tglx@linutronix.de>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Borislav Petkov <bp@alien8.de>
Cc: H. Peter Anvin <hpa@zytor.com>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: David S. Miller <davem@davemloft.net>
Cc: Eric Dumazet <edumazet@google.com>
Cc: Daniel Borkmann <daniel@iogearbox.net>
Cc: Alexei Starovoitov <ast@fb.com>
Cc: Hideaki YOSHIFUJI <yoshfuji@linux-ipv6.org>
Cc: netdev@vger.kernel.org
Cc: linux-kernel@vger.kernel.org
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
2018-04-27 09:54:40 +00:00
|
|
|
int cleanup_addr; /* Epilogue code offset */
|
2021-10-26 12:01:47 +00:00
|
|
|
|
|
|
|
/*
|
|
|
|
* Program specific offsets of labels in the code; these rely on the
|
|
|
|
* JIT doing at least 2 passes, recording the position on the first
|
|
|
|
* pass, only to generate the correct offset on the second pass.
|
|
|
|
*/
|
|
|
|
int tail_call_direct_label;
|
|
|
|
int tail_call_indirect_label;
|
2014-05-14 02:50:45 +00:00
|
|
|
};
|
|
|
|
|
x86/bpf: Clean up non-standard comments, to make the code more readable
So by chance I looked into x86 assembly in arch/x86/net/bpf_jit_comp.c and
noticed the weird and inconsistent comment style it mistakenly learned from
the networking code:
/* Multi-line comment ...
* ... looks like this.
*/
Fix this to use the standard comment style specified in Documentation/CodingStyle
and used in arch/x86/ as well:
/*
* Multi-line comment ...
* ... looks like this.
*/
Also, to quote Linus's ... more explicit views about this:
http://article.gmane.org/gmane.linux.kernel.cryptoapi/21066
> But no, the networking code picked *none* of the above sane formats.
> Instead, it picked these two models that are just half-arsed
> shit-for-brains:
>
> (no)
> /* This is disgusting drug-induced
> * crap, and should die
> */
>
> (no-no-no)
> /* This is also very nasty
> * and visually unbalanced */
>
> Please. The networking code actually has the *worst* possible comment
> style. You can literally find that (no-no-no) style, which is just
> really horribly disgusting and worse than the otherwise fairly similar
> (d) in pretty much every way.
Also improve the comments and some other details while at it:
- Don't mix same-line and previous-line comment style on otherwise
identical code patterns within the same function,
- capitalize 'BPF' and x86 register names consistently,
- capitalize sentences consistently,
- instead of 'x64' use 'x86-64': x64 is a Microsoft specific term,
- use more consistent punctuation,
- use standard coding style in macros as well,
- fix typos and a few other minor details.
Consistent coding style is not optional, at least in arch/x86/.
No change in functionality.
( In case this commit causes conflicts with pending development code
I'll be glad to help resolve any conflicts! )
Acked-by: Thomas Gleixner <tglx@linutronix.de>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Borislav Petkov <bp@alien8.de>
Cc: H. Peter Anvin <hpa@zytor.com>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: David S. Miller <davem@davemloft.net>
Cc: Eric Dumazet <edumazet@google.com>
Cc: Daniel Borkmann <daniel@iogearbox.net>
Cc: Alexei Starovoitov <ast@fb.com>
Cc: Hideaki YOSHIFUJI <yoshfuji@linux-ipv6.org>
Cc: netdev@vger.kernel.org
Cc: linux-kernel@vger.kernel.org
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
2018-04-27 09:54:40 +00:00
|
|
|
/* Maximum number of bytes emitted while JITing one eBPF insn */
|
2014-10-11 03:30:23 +00:00
|
|
|
#define BPF_MAX_INSN_SIZE 128
|
|
|
|
#define BPF_INSN_SAFETY 64
|
2019-11-22 20:07:54 +00:00
|
|
|
|
|
|
|
/* Number of bytes emit_patch() needs to generate instructions */
|
|
|
|
#define X86_PATCH_SIZE 5
|
bpf, x64: rework pro/epilogue and tailcall handling in JIT
This commit serves two things:
1) it optimizes BPF prologue/epilogue generation
2) it makes possible to have tailcalls within BPF subprogram
Both points are related to each other since without 1), 2) could not be
achieved.
In [1], Alexei says:
"The prologue will look like:
nop5
xor eax,eax // two new bytes if bpf_tail_call() is used in this
// function
push rbp
mov rbp, rsp
sub rsp, rounded_stack_depth
push rax // zero init tail_call counter
variable number of push rbx,r13,r14,r15
Then bpf_tail_call will pop variable number rbx,..
and final 'pop rax'
Then 'add rsp, size_of_current_stack_frame'
jmp to next function and skip over 'nop5; xor eax,eax; push rpb; mov
rbp, rsp'
This way new function will set its own stack size and will init tail
call
counter with whatever value the parent had.
If next function doesn't use bpf_tail_call it won't have 'xor eax,eax'.
Instead it would need to have 'nop2' in there."
Implement that suggestion.
Since the layout of stack is changed, tail call counter handling can not
rely anymore on popping it to rbx just like it have been handled for
constant prologue case and later overwrite of rbx with actual value of
rbx pushed to stack. Therefore, let's use one of the register (%rcx) that
is considered to be volatile/caller-saved and pop the value of tail call
counter in there in the epilogue.
Drop the BUILD_BUG_ON in emit_prologue and in
emit_bpf_tail_call_indirect where instruction layout is not constant
anymore.
Introduce new poke target, 'tailcall_bypass' to poke descriptor that is
dedicated for skipping the register pops and stack unwind that are
generated right before the actual jump to target program.
For case when the target program is not present, BPF program will skip
the pop instructions and nop5 dedicated for jmpq $target. An example of
such state when only R6 of callee saved registers is used by program:
ffffffffc0513aa1: e9 0e 00 00 00 jmpq 0xffffffffc0513ab4
ffffffffc0513aa6: 5b pop %rbx
ffffffffc0513aa7: 58 pop %rax
ffffffffc0513aa8: 48 81 c4 00 00 00 00 add $0x0,%rsp
ffffffffc0513aaf: 0f 1f 44 00 00 nopl 0x0(%rax,%rax,1)
ffffffffc0513ab4: 48 89 df mov %rbx,%rdi
When target program is inserted, the jump that was there to skip
pops/nop5 will become the nop5, so CPU will go over pops and do the
actual tailcall.
One might ask why there simply can not be pushes after the nop5?
In the following example snippet:
ffffffffc037030c: 48 89 fb mov %rdi,%rbx
(...)
ffffffffc0370332: 5b pop %rbx
ffffffffc0370333: 58 pop %rax
ffffffffc0370334: 48 81 c4 00 00 00 00 add $0x0,%rsp
ffffffffc037033b: 0f 1f 44 00 00 nopl 0x0(%rax,%rax,1)
ffffffffc0370340: 48 81 ec 00 00 00 00 sub $0x0,%rsp
ffffffffc0370347: 50 push %rax
ffffffffc0370348: 53 push %rbx
ffffffffc0370349: 48 89 df mov %rbx,%rdi
ffffffffc037034c: e8 f7 21 00 00 callq 0xffffffffc0372548
There is the bpf2bpf call (at ffffffffc037034c) right after the tailcall
and jump target is not present. ctx is in %rbx register and BPF
subprogram that we will call into on ffffffffc037034c is relying on it,
e.g. it will pick ctx from there. Such code layout is therefore broken
as we would overwrite the content of %rbx with the value that was pushed
on the prologue. That is the reason for the 'bypass' approach.
Special care needs to be taken during the install/update/remove of
tailcall target. In case when target program is not present, the CPU
must not execute the pop instructions that precede the tailcall.
To address that, the following states can be defined:
A nop, unwind, nop
B nop, unwind, tail
C skip, unwind, nop
D skip, unwind, tail
A is forbidden (lead to incorrectness). The state transitions between
tailcall install/update/remove will work as follows:
First install tail call f: C->D->B(f)
* poke the tailcall, after that get rid of the skip
Update tail call f to f': B(f)->B(f')
* poke the tailcall (poke->tailcall_target) and do NOT touch the
poke->tailcall_bypass
Remove tail call: B(f')->C(f')
* poke->tailcall_bypass is poked back to jump, then we wait the RCU
grace period so that other programs will finish its execution and
after that we are safe to remove the poke->tailcall_target
Install new tail call (f''): C(f')->D(f'')->B(f'').
* same as first step
This way CPU can never be exposed to "unwind, tail" state.
Last but not least, when tailcalls get mixed with bpf2bpf calls, it
would be possible to encounter the endless loop due to clearing the
tailcall counter if for example we would use the tailcall3-like from BPF
selftests program that would be subprogram-based, meaning the tailcall
would be present within the BPF subprogram.
This test, broken down to particular steps, would do:
entry -> set tailcall counter to 0, bump it by 1, tailcall to func0
func0 -> call subprog_tail
(we are NOT skipping the first 11 bytes of prologue and this subprogram
has a tailcall, therefore we clear the counter...)
subprog -> do the same thing as entry
and then loop forever.
To address this, the idea is to go through the call chain of bpf2bpf progs
and look for a tailcall presence throughout whole chain. If we saw a single
tail call then each node in this call chain needs to be marked as a subprog
that can reach the tailcall. We would later feed the JIT with this info
and:
- set eax to 0 only when tailcall is reachable and this is the entry prog
- if tailcall is reachable but there's no tailcall in insns of currently
JITed prog then push rax anyway, so that it will be possible to
propagate further down the call chain
- finally if tailcall is reachable, then we need to precede the 'call'
insn with mov rax, [rbp - (stack_depth + 8)]
Tail call related cases from test_verifier kselftest are also working
fine. Sample BPF programs that utilize tail calls (sockex3, tracex5)
work properly as well.
[1]: https://lore.kernel.org/bpf/20200517043227.2gpq22ifoq37ogst@ast-mbp.dhcp.thefacebook.com/
Suggested-by: Alexei Starovoitov <ast@kernel.org>
Signed-off-by: Maciej Fijalkowski <maciej.fijalkowski@intel.com>
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
2020-09-16 21:10:08 +00:00
|
|
|
/* Number of bytes that will be skipped on tailcall */
|
2022-03-08 15:30:33 +00:00
|
|
|
#define X86_TAIL_CALL_OFFSET (11 + ENDBR_INSN_SIZE)
|
2014-10-11 03:30:23 +00:00
|
|
|
|
bpf: Implement BPF exceptions
This patch implements BPF exceptions, and introduces a bpf_throw kfunc
to allow programs to throw exceptions during their execution at runtime.
A bpf_throw invocation is treated as an immediate termination of the
program, returning back to its caller within the kernel, unwinding all
stack frames.
This allows the program to simplify its implementation, by testing for
runtime conditions which the verifier has no visibility into, and assert
that they are true. In case they are not, the program can simply throw
an exception from the other branch.
BPF exceptions are explicitly *NOT* an unlikely slowpath error handling
primitive, and this objective has guided design choices of the
implementation of the them within the kernel (with the bulk of the cost
for unwinding the stack offloaded to the bpf_throw kfunc).
The implementation of this mechanism requires use of add_hidden_subprog
mechanism introduced in the previous patch, which generates a couple of
instructions to move R1 to R0 and exit. The JIT then rewrites the
prologue of this subprog to take the stack pointer and frame pointer as
inputs and reset the stack frame, popping all callee-saved registers
saved by the main subprog. The bpf_throw function then walks the stack
at runtime, and invokes this exception subprog with the stack and frame
pointers as parameters.
Reviewers must take note that currently the main program is made to save
all callee-saved registers on x86_64 during entry into the program. This
is because we must do an equivalent of a lightweight context switch when
unwinding the stack, therefore we need the callee-saved registers of the
caller of the BPF program to be able to return with a sane state.
Note that we have to additionally handle r12, even though it is not used
by the program, because when throwing the exception the program makes an
entry into the kernel which could clobber r12 after saving it on the
stack. To be able to preserve the value we received on program entry, we
push r12 and restore it from the generated subprogram when unwinding the
stack.
For now, bpf_throw invocation fails when lingering resources or locks
exist in that path of the program. In a future followup, bpf_throw will
be extended to perform frame-by-frame unwinding to release lingering
resources for each stack frame, removing this limitation.
Signed-off-by: Kumar Kartikeya Dwivedi <memxor@gmail.com>
Link: https://lore.kernel.org/r/20230912233214.1518551-5-memxor@gmail.com
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
2023-09-12 23:32:01 +00:00
|
|
|
static void push_r12(u8 **pprog)
|
|
|
|
{
|
|
|
|
u8 *prog = *pprog;
|
|
|
|
|
|
|
|
EMIT2(0x41, 0x54); /* push r12 */
|
|
|
|
*pprog = prog;
|
|
|
|
}
|
|
|
|
|
bpf, x64: rework pro/epilogue and tailcall handling in JIT
This commit serves two things:
1) it optimizes BPF prologue/epilogue generation
2) it makes possible to have tailcalls within BPF subprogram
Both points are related to each other since without 1), 2) could not be
achieved.
In [1], Alexei says:
"The prologue will look like:
nop5
xor eax,eax // two new bytes if bpf_tail_call() is used in this
// function
push rbp
mov rbp, rsp
sub rsp, rounded_stack_depth
push rax // zero init tail_call counter
variable number of push rbx,r13,r14,r15
Then bpf_tail_call will pop variable number rbx,..
and final 'pop rax'
Then 'add rsp, size_of_current_stack_frame'
jmp to next function and skip over 'nop5; xor eax,eax; push rpb; mov
rbp, rsp'
This way new function will set its own stack size and will init tail
call
counter with whatever value the parent had.
If next function doesn't use bpf_tail_call it won't have 'xor eax,eax'.
Instead it would need to have 'nop2' in there."
Implement that suggestion.
Since the layout of stack is changed, tail call counter handling can not
rely anymore on popping it to rbx just like it have been handled for
constant prologue case and later overwrite of rbx with actual value of
rbx pushed to stack. Therefore, let's use one of the register (%rcx) that
is considered to be volatile/caller-saved and pop the value of tail call
counter in there in the epilogue.
Drop the BUILD_BUG_ON in emit_prologue and in
emit_bpf_tail_call_indirect where instruction layout is not constant
anymore.
Introduce new poke target, 'tailcall_bypass' to poke descriptor that is
dedicated for skipping the register pops and stack unwind that are
generated right before the actual jump to target program.
For case when the target program is not present, BPF program will skip
the pop instructions and nop5 dedicated for jmpq $target. An example of
such state when only R6 of callee saved registers is used by program:
ffffffffc0513aa1: e9 0e 00 00 00 jmpq 0xffffffffc0513ab4
ffffffffc0513aa6: 5b pop %rbx
ffffffffc0513aa7: 58 pop %rax
ffffffffc0513aa8: 48 81 c4 00 00 00 00 add $0x0,%rsp
ffffffffc0513aaf: 0f 1f 44 00 00 nopl 0x0(%rax,%rax,1)
ffffffffc0513ab4: 48 89 df mov %rbx,%rdi
When target program is inserted, the jump that was there to skip
pops/nop5 will become the nop5, so CPU will go over pops and do the
actual tailcall.
One might ask why there simply can not be pushes after the nop5?
In the following example snippet:
ffffffffc037030c: 48 89 fb mov %rdi,%rbx
(...)
ffffffffc0370332: 5b pop %rbx
ffffffffc0370333: 58 pop %rax
ffffffffc0370334: 48 81 c4 00 00 00 00 add $0x0,%rsp
ffffffffc037033b: 0f 1f 44 00 00 nopl 0x0(%rax,%rax,1)
ffffffffc0370340: 48 81 ec 00 00 00 00 sub $0x0,%rsp
ffffffffc0370347: 50 push %rax
ffffffffc0370348: 53 push %rbx
ffffffffc0370349: 48 89 df mov %rbx,%rdi
ffffffffc037034c: e8 f7 21 00 00 callq 0xffffffffc0372548
There is the bpf2bpf call (at ffffffffc037034c) right after the tailcall
and jump target is not present. ctx is in %rbx register and BPF
subprogram that we will call into on ffffffffc037034c is relying on it,
e.g. it will pick ctx from there. Such code layout is therefore broken
as we would overwrite the content of %rbx with the value that was pushed
on the prologue. That is the reason for the 'bypass' approach.
Special care needs to be taken during the install/update/remove of
tailcall target. In case when target program is not present, the CPU
must not execute the pop instructions that precede the tailcall.
To address that, the following states can be defined:
A nop, unwind, nop
B nop, unwind, tail
C skip, unwind, nop
D skip, unwind, tail
A is forbidden (lead to incorrectness). The state transitions between
tailcall install/update/remove will work as follows:
First install tail call f: C->D->B(f)
* poke the tailcall, after that get rid of the skip
Update tail call f to f': B(f)->B(f')
* poke the tailcall (poke->tailcall_target) and do NOT touch the
poke->tailcall_bypass
Remove tail call: B(f')->C(f')
* poke->tailcall_bypass is poked back to jump, then we wait the RCU
grace period so that other programs will finish its execution and
after that we are safe to remove the poke->tailcall_target
Install new tail call (f''): C(f')->D(f'')->B(f'').
* same as first step
This way CPU can never be exposed to "unwind, tail" state.
Last but not least, when tailcalls get mixed with bpf2bpf calls, it
would be possible to encounter the endless loop due to clearing the
tailcall counter if for example we would use the tailcall3-like from BPF
selftests program that would be subprogram-based, meaning the tailcall
would be present within the BPF subprogram.
This test, broken down to particular steps, would do:
entry -> set tailcall counter to 0, bump it by 1, tailcall to func0
func0 -> call subprog_tail
(we are NOT skipping the first 11 bytes of prologue and this subprogram
has a tailcall, therefore we clear the counter...)
subprog -> do the same thing as entry
and then loop forever.
To address this, the idea is to go through the call chain of bpf2bpf progs
and look for a tailcall presence throughout whole chain. If we saw a single
tail call then each node in this call chain needs to be marked as a subprog
that can reach the tailcall. We would later feed the JIT with this info
and:
- set eax to 0 only when tailcall is reachable and this is the entry prog
- if tailcall is reachable but there's no tailcall in insns of currently
JITed prog then push rax anyway, so that it will be possible to
propagate further down the call chain
- finally if tailcall is reachable, then we need to precede the 'call'
insn with mov rax, [rbp - (stack_depth + 8)]
Tail call related cases from test_verifier kselftest are also working
fine. Sample BPF programs that utilize tail calls (sockex3, tracex5)
work properly as well.
[1]: https://lore.kernel.org/bpf/20200517043227.2gpq22ifoq37ogst@ast-mbp.dhcp.thefacebook.com/
Suggested-by: Alexei Starovoitov <ast@kernel.org>
Signed-off-by: Maciej Fijalkowski <maciej.fijalkowski@intel.com>
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
2020-09-16 21:10:08 +00:00
|
|
|
static void push_callee_regs(u8 **pprog, bool *callee_regs_used)
|
|
|
|
{
|
|
|
|
u8 *prog = *pprog;
|
|
|
|
|
|
|
|
if (callee_regs_used[0])
|
|
|
|
EMIT1(0x53); /* push rbx */
|
|
|
|
if (callee_regs_used[1])
|
|
|
|
EMIT2(0x41, 0x55); /* push r13 */
|
|
|
|
if (callee_regs_used[2])
|
|
|
|
EMIT2(0x41, 0x56); /* push r14 */
|
|
|
|
if (callee_regs_used[3])
|
|
|
|
EMIT2(0x41, 0x57); /* push r15 */
|
|
|
|
*pprog = prog;
|
|
|
|
}
|
|
|
|
|
bpf: Implement BPF exceptions
This patch implements BPF exceptions, and introduces a bpf_throw kfunc
to allow programs to throw exceptions during their execution at runtime.
A bpf_throw invocation is treated as an immediate termination of the
program, returning back to its caller within the kernel, unwinding all
stack frames.
This allows the program to simplify its implementation, by testing for
runtime conditions which the verifier has no visibility into, and assert
that they are true. In case they are not, the program can simply throw
an exception from the other branch.
BPF exceptions are explicitly *NOT* an unlikely slowpath error handling
primitive, and this objective has guided design choices of the
implementation of the them within the kernel (with the bulk of the cost
for unwinding the stack offloaded to the bpf_throw kfunc).
The implementation of this mechanism requires use of add_hidden_subprog
mechanism introduced in the previous patch, which generates a couple of
instructions to move R1 to R0 and exit. The JIT then rewrites the
prologue of this subprog to take the stack pointer and frame pointer as
inputs and reset the stack frame, popping all callee-saved registers
saved by the main subprog. The bpf_throw function then walks the stack
at runtime, and invokes this exception subprog with the stack and frame
pointers as parameters.
Reviewers must take note that currently the main program is made to save
all callee-saved registers on x86_64 during entry into the program. This
is because we must do an equivalent of a lightweight context switch when
unwinding the stack, therefore we need the callee-saved registers of the
caller of the BPF program to be able to return with a sane state.
Note that we have to additionally handle r12, even though it is not used
by the program, because when throwing the exception the program makes an
entry into the kernel which could clobber r12 after saving it on the
stack. To be able to preserve the value we received on program entry, we
push r12 and restore it from the generated subprogram when unwinding the
stack.
For now, bpf_throw invocation fails when lingering resources or locks
exist in that path of the program. In a future followup, bpf_throw will
be extended to perform frame-by-frame unwinding to release lingering
resources for each stack frame, removing this limitation.
Signed-off-by: Kumar Kartikeya Dwivedi <memxor@gmail.com>
Link: https://lore.kernel.org/r/20230912233214.1518551-5-memxor@gmail.com
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
2023-09-12 23:32:01 +00:00
|
|
|
static void pop_r12(u8 **pprog)
|
|
|
|
{
|
|
|
|
u8 *prog = *pprog;
|
|
|
|
|
|
|
|
EMIT2(0x41, 0x5C); /* pop r12 */
|
|
|
|
*pprog = prog;
|
|
|
|
}
|
|
|
|
|
bpf, x64: rework pro/epilogue and tailcall handling in JIT
This commit serves two things:
1) it optimizes BPF prologue/epilogue generation
2) it makes possible to have tailcalls within BPF subprogram
Both points are related to each other since without 1), 2) could not be
achieved.
In [1], Alexei says:
"The prologue will look like:
nop5
xor eax,eax // two new bytes if bpf_tail_call() is used in this
// function
push rbp
mov rbp, rsp
sub rsp, rounded_stack_depth
push rax // zero init tail_call counter
variable number of push rbx,r13,r14,r15
Then bpf_tail_call will pop variable number rbx,..
and final 'pop rax'
Then 'add rsp, size_of_current_stack_frame'
jmp to next function and skip over 'nop5; xor eax,eax; push rpb; mov
rbp, rsp'
This way new function will set its own stack size and will init tail
call
counter with whatever value the parent had.
If next function doesn't use bpf_tail_call it won't have 'xor eax,eax'.
Instead it would need to have 'nop2' in there."
Implement that suggestion.
Since the layout of stack is changed, tail call counter handling can not
rely anymore on popping it to rbx just like it have been handled for
constant prologue case and later overwrite of rbx with actual value of
rbx pushed to stack. Therefore, let's use one of the register (%rcx) that
is considered to be volatile/caller-saved and pop the value of tail call
counter in there in the epilogue.
Drop the BUILD_BUG_ON in emit_prologue and in
emit_bpf_tail_call_indirect where instruction layout is not constant
anymore.
Introduce new poke target, 'tailcall_bypass' to poke descriptor that is
dedicated for skipping the register pops and stack unwind that are
generated right before the actual jump to target program.
For case when the target program is not present, BPF program will skip
the pop instructions and nop5 dedicated for jmpq $target. An example of
such state when only R6 of callee saved registers is used by program:
ffffffffc0513aa1: e9 0e 00 00 00 jmpq 0xffffffffc0513ab4
ffffffffc0513aa6: 5b pop %rbx
ffffffffc0513aa7: 58 pop %rax
ffffffffc0513aa8: 48 81 c4 00 00 00 00 add $0x0,%rsp
ffffffffc0513aaf: 0f 1f 44 00 00 nopl 0x0(%rax,%rax,1)
ffffffffc0513ab4: 48 89 df mov %rbx,%rdi
When target program is inserted, the jump that was there to skip
pops/nop5 will become the nop5, so CPU will go over pops and do the
actual tailcall.
One might ask why there simply can not be pushes after the nop5?
In the following example snippet:
ffffffffc037030c: 48 89 fb mov %rdi,%rbx
(...)
ffffffffc0370332: 5b pop %rbx
ffffffffc0370333: 58 pop %rax
ffffffffc0370334: 48 81 c4 00 00 00 00 add $0x0,%rsp
ffffffffc037033b: 0f 1f 44 00 00 nopl 0x0(%rax,%rax,1)
ffffffffc0370340: 48 81 ec 00 00 00 00 sub $0x0,%rsp
ffffffffc0370347: 50 push %rax
ffffffffc0370348: 53 push %rbx
ffffffffc0370349: 48 89 df mov %rbx,%rdi
ffffffffc037034c: e8 f7 21 00 00 callq 0xffffffffc0372548
There is the bpf2bpf call (at ffffffffc037034c) right after the tailcall
and jump target is not present. ctx is in %rbx register and BPF
subprogram that we will call into on ffffffffc037034c is relying on it,
e.g. it will pick ctx from there. Such code layout is therefore broken
as we would overwrite the content of %rbx with the value that was pushed
on the prologue. That is the reason for the 'bypass' approach.
Special care needs to be taken during the install/update/remove of
tailcall target. In case when target program is not present, the CPU
must not execute the pop instructions that precede the tailcall.
To address that, the following states can be defined:
A nop, unwind, nop
B nop, unwind, tail
C skip, unwind, nop
D skip, unwind, tail
A is forbidden (lead to incorrectness). The state transitions between
tailcall install/update/remove will work as follows:
First install tail call f: C->D->B(f)
* poke the tailcall, after that get rid of the skip
Update tail call f to f': B(f)->B(f')
* poke the tailcall (poke->tailcall_target) and do NOT touch the
poke->tailcall_bypass
Remove tail call: B(f')->C(f')
* poke->tailcall_bypass is poked back to jump, then we wait the RCU
grace period so that other programs will finish its execution and
after that we are safe to remove the poke->tailcall_target
Install new tail call (f''): C(f')->D(f'')->B(f'').
* same as first step
This way CPU can never be exposed to "unwind, tail" state.
Last but not least, when tailcalls get mixed with bpf2bpf calls, it
would be possible to encounter the endless loop due to clearing the
tailcall counter if for example we would use the tailcall3-like from BPF
selftests program that would be subprogram-based, meaning the tailcall
would be present within the BPF subprogram.
This test, broken down to particular steps, would do:
entry -> set tailcall counter to 0, bump it by 1, tailcall to func0
func0 -> call subprog_tail
(we are NOT skipping the first 11 bytes of prologue and this subprogram
has a tailcall, therefore we clear the counter...)
subprog -> do the same thing as entry
and then loop forever.
To address this, the idea is to go through the call chain of bpf2bpf progs
and look for a tailcall presence throughout whole chain. If we saw a single
tail call then each node in this call chain needs to be marked as a subprog
that can reach the tailcall. We would later feed the JIT with this info
and:
- set eax to 0 only when tailcall is reachable and this is the entry prog
- if tailcall is reachable but there's no tailcall in insns of currently
JITed prog then push rax anyway, so that it will be possible to
propagate further down the call chain
- finally if tailcall is reachable, then we need to precede the 'call'
insn with mov rax, [rbp - (stack_depth + 8)]
Tail call related cases from test_verifier kselftest are also working
fine. Sample BPF programs that utilize tail calls (sockex3, tracex5)
work properly as well.
[1]: https://lore.kernel.org/bpf/20200517043227.2gpq22ifoq37ogst@ast-mbp.dhcp.thefacebook.com/
Suggested-by: Alexei Starovoitov <ast@kernel.org>
Signed-off-by: Maciej Fijalkowski <maciej.fijalkowski@intel.com>
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
2020-09-16 21:10:08 +00:00
|
|
|
static void pop_callee_regs(u8 **pprog, bool *callee_regs_used)
|
|
|
|
{
|
|
|
|
u8 *prog = *pprog;
|
|
|
|
|
|
|
|
if (callee_regs_used[3])
|
|
|
|
EMIT2(0x41, 0x5F); /* pop r15 */
|
|
|
|
if (callee_regs_used[2])
|
|
|
|
EMIT2(0x41, 0x5E); /* pop r14 */
|
|
|
|
if (callee_regs_used[1])
|
|
|
|
EMIT2(0x41, 0x5D); /* pop r13 */
|
|
|
|
if (callee_regs_used[0])
|
|
|
|
EMIT1(0x5B); /* pop rbx */
|
|
|
|
*pprog = prog;
|
|
|
|
}
|
2015-05-19 23:59:04 +00:00
|
|
|
|
2024-01-04 14:22:23 +00:00
|
|
|
static void emit_nops(u8 **pprog, int len)
|
|
|
|
{
|
|
|
|
u8 *prog = *pprog;
|
|
|
|
int i, noplen;
|
|
|
|
|
|
|
|
while (len > 0) {
|
|
|
|
noplen = len;
|
|
|
|
|
|
|
|
if (noplen > ASM_NOP_MAX)
|
|
|
|
noplen = ASM_NOP_MAX;
|
|
|
|
|
|
|
|
for (i = 0; i < noplen; i++)
|
|
|
|
EMIT1(x86_nops[noplen][i]);
|
|
|
|
len -= noplen;
|
|
|
|
}
|
|
|
|
|
|
|
|
*pprog = prog;
|
|
|
|
}
|
|
|
|
|
2023-12-15 09:12:18 +00:00
|
|
|
/*
|
|
|
|
* Emit the various CFI preambles, see asm/cfi.h and the comments about FineIBT
|
|
|
|
* in arch/x86/kernel/alternative.c
|
|
|
|
*/
|
|
|
|
|
2023-12-15 09:12:20 +00:00
|
|
|
static void emit_fineibt(u8 **pprog, u32 hash)
|
2023-12-15 09:12:18 +00:00
|
|
|
{
|
|
|
|
u8 *prog = *pprog;
|
|
|
|
|
|
|
|
EMIT_ENDBR();
|
2023-12-15 09:12:19 +00:00
|
|
|
EMIT3_off32(0x41, 0x81, 0xea, hash); /* subl $hash, %r10d */
|
2023-12-15 09:12:18 +00:00
|
|
|
EMIT2(0x74, 0x07); /* jz.d8 +7 */
|
|
|
|
EMIT2(0x0f, 0x0b); /* ud2 */
|
|
|
|
EMIT1(0x90); /* nop */
|
|
|
|
EMIT_ENDBR_POISON();
|
|
|
|
|
|
|
|
*pprog = prog;
|
|
|
|
}
|
|
|
|
|
2023-12-15 09:12:20 +00:00
|
|
|
static void emit_kcfi(u8 **pprog, u32 hash)
|
2023-12-15 09:12:18 +00:00
|
|
|
{
|
|
|
|
u8 *prog = *pprog;
|
|
|
|
|
2023-12-15 09:12:19 +00:00
|
|
|
EMIT1_off32(0xb8, hash); /* movl $hash, %eax */
|
2023-12-15 09:12:18 +00:00
|
|
|
#ifdef CONFIG_CALL_PADDING
|
|
|
|
EMIT1(0x90);
|
|
|
|
EMIT1(0x90);
|
|
|
|
EMIT1(0x90);
|
|
|
|
EMIT1(0x90);
|
|
|
|
EMIT1(0x90);
|
|
|
|
EMIT1(0x90);
|
|
|
|
EMIT1(0x90);
|
|
|
|
EMIT1(0x90);
|
|
|
|
EMIT1(0x90);
|
|
|
|
EMIT1(0x90);
|
|
|
|
EMIT1(0x90);
|
|
|
|
#endif
|
|
|
|
EMIT_ENDBR();
|
|
|
|
|
|
|
|
*pprog = prog;
|
|
|
|
}
|
|
|
|
|
2023-12-15 09:12:20 +00:00
|
|
|
static void emit_cfi(u8 **pprog, u32 hash)
|
2023-12-15 09:12:18 +00:00
|
|
|
{
|
|
|
|
u8 *prog = *pprog;
|
|
|
|
|
|
|
|
switch (cfi_mode) {
|
|
|
|
case CFI_FINEIBT:
|
2023-12-15 09:12:20 +00:00
|
|
|
emit_fineibt(&prog, hash);
|
2023-12-15 09:12:18 +00:00
|
|
|
break;
|
|
|
|
|
|
|
|
case CFI_KCFI:
|
2023-12-15 09:12:20 +00:00
|
|
|
emit_kcfi(&prog, hash);
|
2023-12-15 09:12:18 +00:00
|
|
|
break;
|
|
|
|
|
|
|
|
default:
|
|
|
|
EMIT_ENDBR();
|
|
|
|
break;
|
|
|
|
}
|
|
|
|
|
|
|
|
*pprog = prog;
|
|
|
|
}
|
|
|
|
|
x86/bpf: Clean up non-standard comments, to make the code more readable
So by chance I looked into x86 assembly in arch/x86/net/bpf_jit_comp.c and
noticed the weird and inconsistent comment style it mistakenly learned from
the networking code:
/* Multi-line comment ...
* ... looks like this.
*/
Fix this to use the standard comment style specified in Documentation/CodingStyle
and used in arch/x86/ as well:
/*
* Multi-line comment ...
* ... looks like this.
*/
Also, to quote Linus's ... more explicit views about this:
http://article.gmane.org/gmane.linux.kernel.cryptoapi/21066
> But no, the networking code picked *none* of the above sane formats.
> Instead, it picked these two models that are just half-arsed
> shit-for-brains:
>
> (no)
> /* This is disgusting drug-induced
> * crap, and should die
> */
>
> (no-no-no)
> /* This is also very nasty
> * and visually unbalanced */
>
> Please. The networking code actually has the *worst* possible comment
> style. You can literally find that (no-no-no) style, which is just
> really horribly disgusting and worse than the otherwise fairly similar
> (d) in pretty much every way.
Also improve the comments and some other details while at it:
- Don't mix same-line and previous-line comment style on otherwise
identical code patterns within the same function,
- capitalize 'BPF' and x86 register names consistently,
- capitalize sentences consistently,
- instead of 'x64' use 'x86-64': x64 is a Microsoft specific term,
- use more consistent punctuation,
- use standard coding style in macros as well,
- fix typos and a few other minor details.
Consistent coding style is not optional, at least in arch/x86/.
No change in functionality.
( In case this commit causes conflicts with pending development code
I'll be glad to help resolve any conflicts! )
Acked-by: Thomas Gleixner <tglx@linutronix.de>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Borislav Petkov <bp@alien8.de>
Cc: H. Peter Anvin <hpa@zytor.com>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: David S. Miller <davem@davemloft.net>
Cc: Eric Dumazet <edumazet@google.com>
Cc: Daniel Borkmann <daniel@iogearbox.net>
Cc: Alexei Starovoitov <ast@fb.com>
Cc: Hideaki YOSHIFUJI <yoshfuji@linux-ipv6.org>
Cc: netdev@vger.kernel.org
Cc: linux-kernel@vger.kernel.org
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
2018-04-27 09:54:40 +00:00
|
|
|
/*
|
bpf, x64: rework pro/epilogue and tailcall handling in JIT
This commit serves two things:
1) it optimizes BPF prologue/epilogue generation
2) it makes possible to have tailcalls within BPF subprogram
Both points are related to each other since without 1), 2) could not be
achieved.
In [1], Alexei says:
"The prologue will look like:
nop5
xor eax,eax // two new bytes if bpf_tail_call() is used in this
// function
push rbp
mov rbp, rsp
sub rsp, rounded_stack_depth
push rax // zero init tail_call counter
variable number of push rbx,r13,r14,r15
Then bpf_tail_call will pop variable number rbx,..
and final 'pop rax'
Then 'add rsp, size_of_current_stack_frame'
jmp to next function and skip over 'nop5; xor eax,eax; push rpb; mov
rbp, rsp'
This way new function will set its own stack size and will init tail
call
counter with whatever value the parent had.
If next function doesn't use bpf_tail_call it won't have 'xor eax,eax'.
Instead it would need to have 'nop2' in there."
Implement that suggestion.
Since the layout of stack is changed, tail call counter handling can not
rely anymore on popping it to rbx just like it have been handled for
constant prologue case and later overwrite of rbx with actual value of
rbx pushed to stack. Therefore, let's use one of the register (%rcx) that
is considered to be volatile/caller-saved and pop the value of tail call
counter in there in the epilogue.
Drop the BUILD_BUG_ON in emit_prologue and in
emit_bpf_tail_call_indirect where instruction layout is not constant
anymore.
Introduce new poke target, 'tailcall_bypass' to poke descriptor that is
dedicated for skipping the register pops and stack unwind that are
generated right before the actual jump to target program.
For case when the target program is not present, BPF program will skip
the pop instructions and nop5 dedicated for jmpq $target. An example of
such state when only R6 of callee saved registers is used by program:
ffffffffc0513aa1: e9 0e 00 00 00 jmpq 0xffffffffc0513ab4
ffffffffc0513aa6: 5b pop %rbx
ffffffffc0513aa7: 58 pop %rax
ffffffffc0513aa8: 48 81 c4 00 00 00 00 add $0x0,%rsp
ffffffffc0513aaf: 0f 1f 44 00 00 nopl 0x0(%rax,%rax,1)
ffffffffc0513ab4: 48 89 df mov %rbx,%rdi
When target program is inserted, the jump that was there to skip
pops/nop5 will become the nop5, so CPU will go over pops and do the
actual tailcall.
One might ask why there simply can not be pushes after the nop5?
In the following example snippet:
ffffffffc037030c: 48 89 fb mov %rdi,%rbx
(...)
ffffffffc0370332: 5b pop %rbx
ffffffffc0370333: 58 pop %rax
ffffffffc0370334: 48 81 c4 00 00 00 00 add $0x0,%rsp
ffffffffc037033b: 0f 1f 44 00 00 nopl 0x0(%rax,%rax,1)
ffffffffc0370340: 48 81 ec 00 00 00 00 sub $0x0,%rsp
ffffffffc0370347: 50 push %rax
ffffffffc0370348: 53 push %rbx
ffffffffc0370349: 48 89 df mov %rbx,%rdi
ffffffffc037034c: e8 f7 21 00 00 callq 0xffffffffc0372548
There is the bpf2bpf call (at ffffffffc037034c) right after the tailcall
and jump target is not present. ctx is in %rbx register and BPF
subprogram that we will call into on ffffffffc037034c is relying on it,
e.g. it will pick ctx from there. Such code layout is therefore broken
as we would overwrite the content of %rbx with the value that was pushed
on the prologue. That is the reason for the 'bypass' approach.
Special care needs to be taken during the install/update/remove of
tailcall target. In case when target program is not present, the CPU
must not execute the pop instructions that precede the tailcall.
To address that, the following states can be defined:
A nop, unwind, nop
B nop, unwind, tail
C skip, unwind, nop
D skip, unwind, tail
A is forbidden (lead to incorrectness). The state transitions between
tailcall install/update/remove will work as follows:
First install tail call f: C->D->B(f)
* poke the tailcall, after that get rid of the skip
Update tail call f to f': B(f)->B(f')
* poke the tailcall (poke->tailcall_target) and do NOT touch the
poke->tailcall_bypass
Remove tail call: B(f')->C(f')
* poke->tailcall_bypass is poked back to jump, then we wait the RCU
grace period so that other programs will finish its execution and
after that we are safe to remove the poke->tailcall_target
Install new tail call (f''): C(f')->D(f'')->B(f'').
* same as first step
This way CPU can never be exposed to "unwind, tail" state.
Last but not least, when tailcalls get mixed with bpf2bpf calls, it
would be possible to encounter the endless loop due to clearing the
tailcall counter if for example we would use the tailcall3-like from BPF
selftests program that would be subprogram-based, meaning the tailcall
would be present within the BPF subprogram.
This test, broken down to particular steps, would do:
entry -> set tailcall counter to 0, bump it by 1, tailcall to func0
func0 -> call subprog_tail
(we are NOT skipping the first 11 bytes of prologue and this subprogram
has a tailcall, therefore we clear the counter...)
subprog -> do the same thing as entry
and then loop forever.
To address this, the idea is to go through the call chain of bpf2bpf progs
and look for a tailcall presence throughout whole chain. If we saw a single
tail call then each node in this call chain needs to be marked as a subprog
that can reach the tailcall. We would later feed the JIT with this info
and:
- set eax to 0 only when tailcall is reachable and this is the entry prog
- if tailcall is reachable but there's no tailcall in insns of currently
JITed prog then push rax anyway, so that it will be possible to
propagate further down the call chain
- finally if tailcall is reachable, then we need to precede the 'call'
insn with mov rax, [rbp - (stack_depth + 8)]
Tail call related cases from test_verifier kselftest are also working
fine. Sample BPF programs that utilize tail calls (sockex3, tracex5)
work properly as well.
[1]: https://lore.kernel.org/bpf/20200517043227.2gpq22ifoq37ogst@ast-mbp.dhcp.thefacebook.com/
Suggested-by: Alexei Starovoitov <ast@kernel.org>
Signed-off-by: Maciej Fijalkowski <maciej.fijalkowski@intel.com>
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
2020-09-16 21:10:08 +00:00
|
|
|
* Emit x86-64 prologue code for BPF program.
|
|
|
|
* bpf_tail_call helper will skip the first X86_TAIL_CALL_OFFSET bytes
|
|
|
|
* while jumping to another program
|
2015-05-19 23:59:04 +00:00
|
|
|
*/
|
bpf, x64: rework pro/epilogue and tailcall handling in JIT
This commit serves two things:
1) it optimizes BPF prologue/epilogue generation
2) it makes possible to have tailcalls within BPF subprogram
Both points are related to each other since without 1), 2) could not be
achieved.
In [1], Alexei says:
"The prologue will look like:
nop5
xor eax,eax // two new bytes if bpf_tail_call() is used in this
// function
push rbp
mov rbp, rsp
sub rsp, rounded_stack_depth
push rax // zero init tail_call counter
variable number of push rbx,r13,r14,r15
Then bpf_tail_call will pop variable number rbx,..
and final 'pop rax'
Then 'add rsp, size_of_current_stack_frame'
jmp to next function and skip over 'nop5; xor eax,eax; push rpb; mov
rbp, rsp'
This way new function will set its own stack size and will init tail
call
counter with whatever value the parent had.
If next function doesn't use bpf_tail_call it won't have 'xor eax,eax'.
Instead it would need to have 'nop2' in there."
Implement that suggestion.
Since the layout of stack is changed, tail call counter handling can not
rely anymore on popping it to rbx just like it have been handled for
constant prologue case and later overwrite of rbx with actual value of
rbx pushed to stack. Therefore, let's use one of the register (%rcx) that
is considered to be volatile/caller-saved and pop the value of tail call
counter in there in the epilogue.
Drop the BUILD_BUG_ON in emit_prologue and in
emit_bpf_tail_call_indirect where instruction layout is not constant
anymore.
Introduce new poke target, 'tailcall_bypass' to poke descriptor that is
dedicated for skipping the register pops and stack unwind that are
generated right before the actual jump to target program.
For case when the target program is not present, BPF program will skip
the pop instructions and nop5 dedicated for jmpq $target. An example of
such state when only R6 of callee saved registers is used by program:
ffffffffc0513aa1: e9 0e 00 00 00 jmpq 0xffffffffc0513ab4
ffffffffc0513aa6: 5b pop %rbx
ffffffffc0513aa7: 58 pop %rax
ffffffffc0513aa8: 48 81 c4 00 00 00 00 add $0x0,%rsp
ffffffffc0513aaf: 0f 1f 44 00 00 nopl 0x0(%rax,%rax,1)
ffffffffc0513ab4: 48 89 df mov %rbx,%rdi
When target program is inserted, the jump that was there to skip
pops/nop5 will become the nop5, so CPU will go over pops and do the
actual tailcall.
One might ask why there simply can not be pushes after the nop5?
In the following example snippet:
ffffffffc037030c: 48 89 fb mov %rdi,%rbx
(...)
ffffffffc0370332: 5b pop %rbx
ffffffffc0370333: 58 pop %rax
ffffffffc0370334: 48 81 c4 00 00 00 00 add $0x0,%rsp
ffffffffc037033b: 0f 1f 44 00 00 nopl 0x0(%rax,%rax,1)
ffffffffc0370340: 48 81 ec 00 00 00 00 sub $0x0,%rsp
ffffffffc0370347: 50 push %rax
ffffffffc0370348: 53 push %rbx
ffffffffc0370349: 48 89 df mov %rbx,%rdi
ffffffffc037034c: e8 f7 21 00 00 callq 0xffffffffc0372548
There is the bpf2bpf call (at ffffffffc037034c) right after the tailcall
and jump target is not present. ctx is in %rbx register and BPF
subprogram that we will call into on ffffffffc037034c is relying on it,
e.g. it will pick ctx from there. Such code layout is therefore broken
as we would overwrite the content of %rbx with the value that was pushed
on the prologue. That is the reason for the 'bypass' approach.
Special care needs to be taken during the install/update/remove of
tailcall target. In case when target program is not present, the CPU
must not execute the pop instructions that precede the tailcall.
To address that, the following states can be defined:
A nop, unwind, nop
B nop, unwind, tail
C skip, unwind, nop
D skip, unwind, tail
A is forbidden (lead to incorrectness). The state transitions between
tailcall install/update/remove will work as follows:
First install tail call f: C->D->B(f)
* poke the tailcall, after that get rid of the skip
Update tail call f to f': B(f)->B(f')
* poke the tailcall (poke->tailcall_target) and do NOT touch the
poke->tailcall_bypass
Remove tail call: B(f')->C(f')
* poke->tailcall_bypass is poked back to jump, then we wait the RCU
grace period so that other programs will finish its execution and
after that we are safe to remove the poke->tailcall_target
Install new tail call (f''): C(f')->D(f'')->B(f'').
* same as first step
This way CPU can never be exposed to "unwind, tail" state.
Last but not least, when tailcalls get mixed with bpf2bpf calls, it
would be possible to encounter the endless loop due to clearing the
tailcall counter if for example we would use the tailcall3-like from BPF
selftests program that would be subprogram-based, meaning the tailcall
would be present within the BPF subprogram.
This test, broken down to particular steps, would do:
entry -> set tailcall counter to 0, bump it by 1, tailcall to func0
func0 -> call subprog_tail
(we are NOT skipping the first 11 bytes of prologue and this subprogram
has a tailcall, therefore we clear the counter...)
subprog -> do the same thing as entry
and then loop forever.
To address this, the idea is to go through the call chain of bpf2bpf progs
and look for a tailcall presence throughout whole chain. If we saw a single
tail call then each node in this call chain needs to be marked as a subprog
that can reach the tailcall. We would later feed the JIT with this info
and:
- set eax to 0 only when tailcall is reachable and this is the entry prog
- if tailcall is reachable but there's no tailcall in insns of currently
JITed prog then push rax anyway, so that it will be possible to
propagate further down the call chain
- finally if tailcall is reachable, then we need to precede the 'call'
insn with mov rax, [rbp - (stack_depth + 8)]
Tail call related cases from test_verifier kselftest are also working
fine. Sample BPF programs that utilize tail calls (sockex3, tracex5)
work properly as well.
[1]: https://lore.kernel.org/bpf/20200517043227.2gpq22ifoq37ogst@ast-mbp.dhcp.thefacebook.com/
Suggested-by: Alexei Starovoitov <ast@kernel.org>
Signed-off-by: Maciej Fijalkowski <maciej.fijalkowski@intel.com>
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
2020-09-16 21:10:08 +00:00
|
|
|
static void emit_prologue(u8 **pprog, u32 stack_depth, bool ebpf_from_cbpf,
|
bpf: Implement BPF exceptions
This patch implements BPF exceptions, and introduces a bpf_throw kfunc
to allow programs to throw exceptions during their execution at runtime.
A bpf_throw invocation is treated as an immediate termination of the
program, returning back to its caller within the kernel, unwinding all
stack frames.
This allows the program to simplify its implementation, by testing for
runtime conditions which the verifier has no visibility into, and assert
that they are true. In case they are not, the program can simply throw
an exception from the other branch.
BPF exceptions are explicitly *NOT* an unlikely slowpath error handling
primitive, and this objective has guided design choices of the
implementation of the them within the kernel (with the bulk of the cost
for unwinding the stack offloaded to the bpf_throw kfunc).
The implementation of this mechanism requires use of add_hidden_subprog
mechanism introduced in the previous patch, which generates a couple of
instructions to move R1 to R0 and exit. The JIT then rewrites the
prologue of this subprog to take the stack pointer and frame pointer as
inputs and reset the stack frame, popping all callee-saved registers
saved by the main subprog. The bpf_throw function then walks the stack
at runtime, and invokes this exception subprog with the stack and frame
pointers as parameters.
Reviewers must take note that currently the main program is made to save
all callee-saved registers on x86_64 during entry into the program. This
is because we must do an equivalent of a lightweight context switch when
unwinding the stack, therefore we need the callee-saved registers of the
caller of the BPF program to be able to return with a sane state.
Note that we have to additionally handle r12, even though it is not used
by the program, because when throwing the exception the program makes an
entry into the kernel which could clobber r12 after saving it on the
stack. To be able to preserve the value we received on program entry, we
push r12 and restore it from the generated subprogram when unwinding the
stack.
For now, bpf_throw invocation fails when lingering resources or locks
exist in that path of the program. In a future followup, bpf_throw will
be extended to perform frame-by-frame unwinding to release lingering
resources for each stack frame, removing this limitation.
Signed-off-by: Kumar Kartikeya Dwivedi <memxor@gmail.com>
Link: https://lore.kernel.org/r/20230912233214.1518551-5-memxor@gmail.com
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
2023-09-12 23:32:01 +00:00
|
|
|
bool tail_call_reachable, bool is_subprog,
|
|
|
|
bool is_exception_cb)
|
2011-04-20 09:27:32 +00:00
|
|
|
{
|
2015-05-19 23:59:04 +00:00
|
|
|
u8 *prog = *pprog;
|
2011-04-20 09:27:32 +00:00
|
|
|
|
2023-12-15 09:12:20 +00:00
|
|
|
emit_cfi(&prog, is_subprog ? cfi_bpf_subprog_hash : cfi_bpf_hash);
|
2019-11-14 18:57:13 +00:00
|
|
|
/* BPF trampoline can be made to work without these nops,
|
|
|
|
* but let's waste 5 bytes for now and optimize later
|
|
|
|
*/
|
2024-01-04 14:22:23 +00:00
|
|
|
emit_nops(&prog, X86_PATCH_SIZE);
|
bpf, x64: rework pro/epilogue and tailcall handling in JIT
This commit serves two things:
1) it optimizes BPF prologue/epilogue generation
2) it makes possible to have tailcalls within BPF subprogram
Both points are related to each other since without 1), 2) could not be
achieved.
In [1], Alexei says:
"The prologue will look like:
nop5
xor eax,eax // two new bytes if bpf_tail_call() is used in this
// function
push rbp
mov rbp, rsp
sub rsp, rounded_stack_depth
push rax // zero init tail_call counter
variable number of push rbx,r13,r14,r15
Then bpf_tail_call will pop variable number rbx,..
and final 'pop rax'
Then 'add rsp, size_of_current_stack_frame'
jmp to next function and skip over 'nop5; xor eax,eax; push rpb; mov
rbp, rsp'
This way new function will set its own stack size and will init tail
call
counter with whatever value the parent had.
If next function doesn't use bpf_tail_call it won't have 'xor eax,eax'.
Instead it would need to have 'nop2' in there."
Implement that suggestion.
Since the layout of stack is changed, tail call counter handling can not
rely anymore on popping it to rbx just like it have been handled for
constant prologue case and later overwrite of rbx with actual value of
rbx pushed to stack. Therefore, let's use one of the register (%rcx) that
is considered to be volatile/caller-saved and pop the value of tail call
counter in there in the epilogue.
Drop the BUILD_BUG_ON in emit_prologue and in
emit_bpf_tail_call_indirect where instruction layout is not constant
anymore.
Introduce new poke target, 'tailcall_bypass' to poke descriptor that is
dedicated for skipping the register pops and stack unwind that are
generated right before the actual jump to target program.
For case when the target program is not present, BPF program will skip
the pop instructions and nop5 dedicated for jmpq $target. An example of
such state when only R6 of callee saved registers is used by program:
ffffffffc0513aa1: e9 0e 00 00 00 jmpq 0xffffffffc0513ab4
ffffffffc0513aa6: 5b pop %rbx
ffffffffc0513aa7: 58 pop %rax
ffffffffc0513aa8: 48 81 c4 00 00 00 00 add $0x0,%rsp
ffffffffc0513aaf: 0f 1f 44 00 00 nopl 0x0(%rax,%rax,1)
ffffffffc0513ab4: 48 89 df mov %rbx,%rdi
When target program is inserted, the jump that was there to skip
pops/nop5 will become the nop5, so CPU will go over pops and do the
actual tailcall.
One might ask why there simply can not be pushes after the nop5?
In the following example snippet:
ffffffffc037030c: 48 89 fb mov %rdi,%rbx
(...)
ffffffffc0370332: 5b pop %rbx
ffffffffc0370333: 58 pop %rax
ffffffffc0370334: 48 81 c4 00 00 00 00 add $0x0,%rsp
ffffffffc037033b: 0f 1f 44 00 00 nopl 0x0(%rax,%rax,1)
ffffffffc0370340: 48 81 ec 00 00 00 00 sub $0x0,%rsp
ffffffffc0370347: 50 push %rax
ffffffffc0370348: 53 push %rbx
ffffffffc0370349: 48 89 df mov %rbx,%rdi
ffffffffc037034c: e8 f7 21 00 00 callq 0xffffffffc0372548
There is the bpf2bpf call (at ffffffffc037034c) right after the tailcall
and jump target is not present. ctx is in %rbx register and BPF
subprogram that we will call into on ffffffffc037034c is relying on it,
e.g. it will pick ctx from there. Such code layout is therefore broken
as we would overwrite the content of %rbx with the value that was pushed
on the prologue. That is the reason for the 'bypass' approach.
Special care needs to be taken during the install/update/remove of
tailcall target. In case when target program is not present, the CPU
must not execute the pop instructions that precede the tailcall.
To address that, the following states can be defined:
A nop, unwind, nop
B nop, unwind, tail
C skip, unwind, nop
D skip, unwind, tail
A is forbidden (lead to incorrectness). The state transitions between
tailcall install/update/remove will work as follows:
First install tail call f: C->D->B(f)
* poke the tailcall, after that get rid of the skip
Update tail call f to f': B(f)->B(f')
* poke the tailcall (poke->tailcall_target) and do NOT touch the
poke->tailcall_bypass
Remove tail call: B(f')->C(f')
* poke->tailcall_bypass is poked back to jump, then we wait the RCU
grace period so that other programs will finish its execution and
after that we are safe to remove the poke->tailcall_target
Install new tail call (f''): C(f')->D(f'')->B(f'').
* same as first step
This way CPU can never be exposed to "unwind, tail" state.
Last but not least, when tailcalls get mixed with bpf2bpf calls, it
would be possible to encounter the endless loop due to clearing the
tailcall counter if for example we would use the tailcall3-like from BPF
selftests program that would be subprogram-based, meaning the tailcall
would be present within the BPF subprogram.
This test, broken down to particular steps, would do:
entry -> set tailcall counter to 0, bump it by 1, tailcall to func0
func0 -> call subprog_tail
(we are NOT skipping the first 11 bytes of prologue and this subprogram
has a tailcall, therefore we clear the counter...)
subprog -> do the same thing as entry
and then loop forever.
To address this, the idea is to go through the call chain of bpf2bpf progs
and look for a tailcall presence throughout whole chain. If we saw a single
tail call then each node in this call chain needs to be marked as a subprog
that can reach the tailcall. We would later feed the JIT with this info
and:
- set eax to 0 only when tailcall is reachable and this is the entry prog
- if tailcall is reachable but there's no tailcall in insns of currently
JITed prog then push rax anyway, so that it will be possible to
propagate further down the call chain
- finally if tailcall is reachable, then we need to precede the 'call'
insn with mov rax, [rbp - (stack_depth + 8)]
Tail call related cases from test_verifier kselftest are also working
fine. Sample BPF programs that utilize tail calls (sockex3, tracex5)
work properly as well.
[1]: https://lore.kernel.org/bpf/20200517043227.2gpq22ifoq37ogst@ast-mbp.dhcp.thefacebook.com/
Suggested-by: Alexei Starovoitov <ast@kernel.org>
Signed-off-by: Maciej Fijalkowski <maciej.fijalkowski@intel.com>
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
2020-09-16 21:10:08 +00:00
|
|
|
if (!ebpf_from_cbpf) {
|
|
|
|
if (tail_call_reachable && !is_subprog)
|
2023-09-12 15:04:40 +00:00
|
|
|
/* When it's the entry of the whole tailcall context,
|
|
|
|
* zeroing rax means initialising tail_call_cnt.
|
|
|
|
*/
|
bpf, x64: rework pro/epilogue and tailcall handling in JIT
This commit serves two things:
1) it optimizes BPF prologue/epilogue generation
2) it makes possible to have tailcalls within BPF subprogram
Both points are related to each other since without 1), 2) could not be
achieved.
In [1], Alexei says:
"The prologue will look like:
nop5
xor eax,eax // two new bytes if bpf_tail_call() is used in this
// function
push rbp
mov rbp, rsp
sub rsp, rounded_stack_depth
push rax // zero init tail_call counter
variable number of push rbx,r13,r14,r15
Then bpf_tail_call will pop variable number rbx,..
and final 'pop rax'
Then 'add rsp, size_of_current_stack_frame'
jmp to next function and skip over 'nop5; xor eax,eax; push rpb; mov
rbp, rsp'
This way new function will set its own stack size and will init tail
call
counter with whatever value the parent had.
If next function doesn't use bpf_tail_call it won't have 'xor eax,eax'.
Instead it would need to have 'nop2' in there."
Implement that suggestion.
Since the layout of stack is changed, tail call counter handling can not
rely anymore on popping it to rbx just like it have been handled for
constant prologue case and later overwrite of rbx with actual value of
rbx pushed to stack. Therefore, let's use one of the register (%rcx) that
is considered to be volatile/caller-saved and pop the value of tail call
counter in there in the epilogue.
Drop the BUILD_BUG_ON in emit_prologue and in
emit_bpf_tail_call_indirect where instruction layout is not constant
anymore.
Introduce new poke target, 'tailcall_bypass' to poke descriptor that is
dedicated for skipping the register pops and stack unwind that are
generated right before the actual jump to target program.
For case when the target program is not present, BPF program will skip
the pop instructions and nop5 dedicated for jmpq $target. An example of
such state when only R6 of callee saved registers is used by program:
ffffffffc0513aa1: e9 0e 00 00 00 jmpq 0xffffffffc0513ab4
ffffffffc0513aa6: 5b pop %rbx
ffffffffc0513aa7: 58 pop %rax
ffffffffc0513aa8: 48 81 c4 00 00 00 00 add $0x0,%rsp
ffffffffc0513aaf: 0f 1f 44 00 00 nopl 0x0(%rax,%rax,1)
ffffffffc0513ab4: 48 89 df mov %rbx,%rdi
When target program is inserted, the jump that was there to skip
pops/nop5 will become the nop5, so CPU will go over pops and do the
actual tailcall.
One might ask why there simply can not be pushes after the nop5?
In the following example snippet:
ffffffffc037030c: 48 89 fb mov %rdi,%rbx
(...)
ffffffffc0370332: 5b pop %rbx
ffffffffc0370333: 58 pop %rax
ffffffffc0370334: 48 81 c4 00 00 00 00 add $0x0,%rsp
ffffffffc037033b: 0f 1f 44 00 00 nopl 0x0(%rax,%rax,1)
ffffffffc0370340: 48 81 ec 00 00 00 00 sub $0x0,%rsp
ffffffffc0370347: 50 push %rax
ffffffffc0370348: 53 push %rbx
ffffffffc0370349: 48 89 df mov %rbx,%rdi
ffffffffc037034c: e8 f7 21 00 00 callq 0xffffffffc0372548
There is the bpf2bpf call (at ffffffffc037034c) right after the tailcall
and jump target is not present. ctx is in %rbx register and BPF
subprogram that we will call into on ffffffffc037034c is relying on it,
e.g. it will pick ctx from there. Such code layout is therefore broken
as we would overwrite the content of %rbx with the value that was pushed
on the prologue. That is the reason for the 'bypass' approach.
Special care needs to be taken during the install/update/remove of
tailcall target. In case when target program is not present, the CPU
must not execute the pop instructions that precede the tailcall.
To address that, the following states can be defined:
A nop, unwind, nop
B nop, unwind, tail
C skip, unwind, nop
D skip, unwind, tail
A is forbidden (lead to incorrectness). The state transitions between
tailcall install/update/remove will work as follows:
First install tail call f: C->D->B(f)
* poke the tailcall, after that get rid of the skip
Update tail call f to f': B(f)->B(f')
* poke the tailcall (poke->tailcall_target) and do NOT touch the
poke->tailcall_bypass
Remove tail call: B(f')->C(f')
* poke->tailcall_bypass is poked back to jump, then we wait the RCU
grace period so that other programs will finish its execution and
after that we are safe to remove the poke->tailcall_target
Install new tail call (f''): C(f')->D(f'')->B(f'').
* same as first step
This way CPU can never be exposed to "unwind, tail" state.
Last but not least, when tailcalls get mixed with bpf2bpf calls, it
would be possible to encounter the endless loop due to clearing the
tailcall counter if for example we would use the tailcall3-like from BPF
selftests program that would be subprogram-based, meaning the tailcall
would be present within the BPF subprogram.
This test, broken down to particular steps, would do:
entry -> set tailcall counter to 0, bump it by 1, tailcall to func0
func0 -> call subprog_tail
(we are NOT skipping the first 11 bytes of prologue and this subprogram
has a tailcall, therefore we clear the counter...)
subprog -> do the same thing as entry
and then loop forever.
To address this, the idea is to go through the call chain of bpf2bpf progs
and look for a tailcall presence throughout whole chain. If we saw a single
tail call then each node in this call chain needs to be marked as a subprog
that can reach the tailcall. We would later feed the JIT with this info
and:
- set eax to 0 only when tailcall is reachable and this is the entry prog
- if tailcall is reachable but there's no tailcall in insns of currently
JITed prog then push rax anyway, so that it will be possible to
propagate further down the call chain
- finally if tailcall is reachable, then we need to precede the 'call'
insn with mov rax, [rbp - (stack_depth + 8)]
Tail call related cases from test_verifier kselftest are also working
fine. Sample BPF programs that utilize tail calls (sockex3, tracex5)
work properly as well.
[1]: https://lore.kernel.org/bpf/20200517043227.2gpq22ifoq37ogst@ast-mbp.dhcp.thefacebook.com/
Suggested-by: Alexei Starovoitov <ast@kernel.org>
Signed-off-by: Maciej Fijalkowski <maciej.fijalkowski@intel.com>
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
2020-09-16 21:10:08 +00:00
|
|
|
EMIT2(0x31, 0xC0); /* xor eax, eax */
|
|
|
|
else
|
2023-09-12 15:04:40 +00:00
|
|
|
/* Keep the same instruction layout. */
|
bpf, x64: rework pro/epilogue and tailcall handling in JIT
This commit serves two things:
1) it optimizes BPF prologue/epilogue generation
2) it makes possible to have tailcalls within BPF subprogram
Both points are related to each other since without 1), 2) could not be
achieved.
In [1], Alexei says:
"The prologue will look like:
nop5
xor eax,eax // two new bytes if bpf_tail_call() is used in this
// function
push rbp
mov rbp, rsp
sub rsp, rounded_stack_depth
push rax // zero init tail_call counter
variable number of push rbx,r13,r14,r15
Then bpf_tail_call will pop variable number rbx,..
and final 'pop rax'
Then 'add rsp, size_of_current_stack_frame'
jmp to next function and skip over 'nop5; xor eax,eax; push rpb; mov
rbp, rsp'
This way new function will set its own stack size and will init tail
call
counter with whatever value the parent had.
If next function doesn't use bpf_tail_call it won't have 'xor eax,eax'.
Instead it would need to have 'nop2' in there."
Implement that suggestion.
Since the layout of stack is changed, tail call counter handling can not
rely anymore on popping it to rbx just like it have been handled for
constant prologue case and later overwrite of rbx with actual value of
rbx pushed to stack. Therefore, let's use one of the register (%rcx) that
is considered to be volatile/caller-saved and pop the value of tail call
counter in there in the epilogue.
Drop the BUILD_BUG_ON in emit_prologue and in
emit_bpf_tail_call_indirect where instruction layout is not constant
anymore.
Introduce new poke target, 'tailcall_bypass' to poke descriptor that is
dedicated for skipping the register pops and stack unwind that are
generated right before the actual jump to target program.
For case when the target program is not present, BPF program will skip
the pop instructions and nop5 dedicated for jmpq $target. An example of
such state when only R6 of callee saved registers is used by program:
ffffffffc0513aa1: e9 0e 00 00 00 jmpq 0xffffffffc0513ab4
ffffffffc0513aa6: 5b pop %rbx
ffffffffc0513aa7: 58 pop %rax
ffffffffc0513aa8: 48 81 c4 00 00 00 00 add $0x0,%rsp
ffffffffc0513aaf: 0f 1f 44 00 00 nopl 0x0(%rax,%rax,1)
ffffffffc0513ab4: 48 89 df mov %rbx,%rdi
When target program is inserted, the jump that was there to skip
pops/nop5 will become the nop5, so CPU will go over pops and do the
actual tailcall.
One might ask why there simply can not be pushes after the nop5?
In the following example snippet:
ffffffffc037030c: 48 89 fb mov %rdi,%rbx
(...)
ffffffffc0370332: 5b pop %rbx
ffffffffc0370333: 58 pop %rax
ffffffffc0370334: 48 81 c4 00 00 00 00 add $0x0,%rsp
ffffffffc037033b: 0f 1f 44 00 00 nopl 0x0(%rax,%rax,1)
ffffffffc0370340: 48 81 ec 00 00 00 00 sub $0x0,%rsp
ffffffffc0370347: 50 push %rax
ffffffffc0370348: 53 push %rbx
ffffffffc0370349: 48 89 df mov %rbx,%rdi
ffffffffc037034c: e8 f7 21 00 00 callq 0xffffffffc0372548
There is the bpf2bpf call (at ffffffffc037034c) right after the tailcall
and jump target is not present. ctx is in %rbx register and BPF
subprogram that we will call into on ffffffffc037034c is relying on it,
e.g. it will pick ctx from there. Such code layout is therefore broken
as we would overwrite the content of %rbx with the value that was pushed
on the prologue. That is the reason for the 'bypass' approach.
Special care needs to be taken during the install/update/remove of
tailcall target. In case when target program is not present, the CPU
must not execute the pop instructions that precede the tailcall.
To address that, the following states can be defined:
A nop, unwind, nop
B nop, unwind, tail
C skip, unwind, nop
D skip, unwind, tail
A is forbidden (lead to incorrectness). The state transitions between
tailcall install/update/remove will work as follows:
First install tail call f: C->D->B(f)
* poke the tailcall, after that get rid of the skip
Update tail call f to f': B(f)->B(f')
* poke the tailcall (poke->tailcall_target) and do NOT touch the
poke->tailcall_bypass
Remove tail call: B(f')->C(f')
* poke->tailcall_bypass is poked back to jump, then we wait the RCU
grace period so that other programs will finish its execution and
after that we are safe to remove the poke->tailcall_target
Install new tail call (f''): C(f')->D(f'')->B(f'').
* same as first step
This way CPU can never be exposed to "unwind, tail" state.
Last but not least, when tailcalls get mixed with bpf2bpf calls, it
would be possible to encounter the endless loop due to clearing the
tailcall counter if for example we would use the tailcall3-like from BPF
selftests program that would be subprogram-based, meaning the tailcall
would be present within the BPF subprogram.
This test, broken down to particular steps, would do:
entry -> set tailcall counter to 0, bump it by 1, tailcall to func0
func0 -> call subprog_tail
(we are NOT skipping the first 11 bytes of prologue and this subprogram
has a tailcall, therefore we clear the counter...)
subprog -> do the same thing as entry
and then loop forever.
To address this, the idea is to go through the call chain of bpf2bpf progs
and look for a tailcall presence throughout whole chain. If we saw a single
tail call then each node in this call chain needs to be marked as a subprog
that can reach the tailcall. We would later feed the JIT with this info
and:
- set eax to 0 only when tailcall is reachable and this is the entry prog
- if tailcall is reachable but there's no tailcall in insns of currently
JITed prog then push rax anyway, so that it will be possible to
propagate further down the call chain
- finally if tailcall is reachable, then we need to precede the 'call'
insn with mov rax, [rbp - (stack_depth + 8)]
Tail call related cases from test_verifier kselftest are also working
fine. Sample BPF programs that utilize tail calls (sockex3, tracex5)
work properly as well.
[1]: https://lore.kernel.org/bpf/20200517043227.2gpq22ifoq37ogst@ast-mbp.dhcp.thefacebook.com/
Suggested-by: Alexei Starovoitov <ast@kernel.org>
Signed-off-by: Maciej Fijalkowski <maciej.fijalkowski@intel.com>
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
2020-09-16 21:10:08 +00:00
|
|
|
EMIT2(0x66, 0x90); /* nop2 */
|
|
|
|
}
|
bpf: Implement BPF exceptions
This patch implements BPF exceptions, and introduces a bpf_throw kfunc
to allow programs to throw exceptions during their execution at runtime.
A bpf_throw invocation is treated as an immediate termination of the
program, returning back to its caller within the kernel, unwinding all
stack frames.
This allows the program to simplify its implementation, by testing for
runtime conditions which the verifier has no visibility into, and assert
that they are true. In case they are not, the program can simply throw
an exception from the other branch.
BPF exceptions are explicitly *NOT* an unlikely slowpath error handling
primitive, and this objective has guided design choices of the
implementation of the them within the kernel (with the bulk of the cost
for unwinding the stack offloaded to the bpf_throw kfunc).
The implementation of this mechanism requires use of add_hidden_subprog
mechanism introduced in the previous patch, which generates a couple of
instructions to move R1 to R0 and exit. The JIT then rewrites the
prologue of this subprog to take the stack pointer and frame pointer as
inputs and reset the stack frame, popping all callee-saved registers
saved by the main subprog. The bpf_throw function then walks the stack
at runtime, and invokes this exception subprog with the stack and frame
pointers as parameters.
Reviewers must take note that currently the main program is made to save
all callee-saved registers on x86_64 during entry into the program. This
is because we must do an equivalent of a lightweight context switch when
unwinding the stack, therefore we need the callee-saved registers of the
caller of the BPF program to be able to return with a sane state.
Note that we have to additionally handle r12, even though it is not used
by the program, because when throwing the exception the program makes an
entry into the kernel which could clobber r12 after saving it on the
stack. To be able to preserve the value we received on program entry, we
push r12 and restore it from the generated subprogram when unwinding the
stack.
For now, bpf_throw invocation fails when lingering resources or locks
exist in that path of the program. In a future followup, bpf_throw will
be extended to perform frame-by-frame unwinding to release lingering
resources for each stack frame, removing this limitation.
Signed-off-by: Kumar Kartikeya Dwivedi <memxor@gmail.com>
Link: https://lore.kernel.org/r/20230912233214.1518551-5-memxor@gmail.com
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
2023-09-12 23:32:01 +00:00
|
|
|
/* Exception callback receives FP as third parameter */
|
|
|
|
if (is_exception_cb) {
|
|
|
|
EMIT3(0x48, 0x89, 0xF4); /* mov rsp, rsi */
|
|
|
|
EMIT3(0x48, 0x89, 0xD5); /* mov rbp, rdx */
|
|
|
|
/* The main frame must have exception_boundary as true, so we
|
|
|
|
* first restore those callee-saved regs from stack, before
|
|
|
|
* reusing the stack frame.
|
|
|
|
*/
|
|
|
|
pop_callee_regs(&prog, all_callee_regs_used);
|
|
|
|
pop_r12(&prog);
|
|
|
|
/* Reset the stack frame. */
|
|
|
|
EMIT3(0x48, 0x89, 0xEC); /* mov rsp, rbp */
|
|
|
|
} else {
|
|
|
|
EMIT1(0x55); /* push rbp */
|
|
|
|
EMIT3(0x48, 0x89, 0xE5); /* mov rbp, rsp */
|
|
|
|
}
|
2022-03-08 15:30:33 +00:00
|
|
|
|
|
|
|
/* X86_TAIL_CALL_OFFSET is here */
|
|
|
|
EMIT_ENDBR();
|
|
|
|
|
2019-06-14 22:43:28 +00:00
|
|
|
/* sub rsp, rounded_stack_depth */
|
2020-09-29 20:46:53 +00:00
|
|
|
if (stack_depth)
|
|
|
|
EMIT3_off32(0x48, 0x81, 0xEC, round_up(stack_depth, 8));
|
bpf, x64: rework pro/epilogue and tailcall handling in JIT
This commit serves two things:
1) it optimizes BPF prologue/epilogue generation
2) it makes possible to have tailcalls within BPF subprogram
Both points are related to each other since without 1), 2) could not be
achieved.
In [1], Alexei says:
"The prologue will look like:
nop5
xor eax,eax // two new bytes if bpf_tail_call() is used in this
// function
push rbp
mov rbp, rsp
sub rsp, rounded_stack_depth
push rax // zero init tail_call counter
variable number of push rbx,r13,r14,r15
Then bpf_tail_call will pop variable number rbx,..
and final 'pop rax'
Then 'add rsp, size_of_current_stack_frame'
jmp to next function and skip over 'nop5; xor eax,eax; push rpb; mov
rbp, rsp'
This way new function will set its own stack size and will init tail
call
counter with whatever value the parent had.
If next function doesn't use bpf_tail_call it won't have 'xor eax,eax'.
Instead it would need to have 'nop2' in there."
Implement that suggestion.
Since the layout of stack is changed, tail call counter handling can not
rely anymore on popping it to rbx just like it have been handled for
constant prologue case and later overwrite of rbx with actual value of
rbx pushed to stack. Therefore, let's use one of the register (%rcx) that
is considered to be volatile/caller-saved and pop the value of tail call
counter in there in the epilogue.
Drop the BUILD_BUG_ON in emit_prologue and in
emit_bpf_tail_call_indirect where instruction layout is not constant
anymore.
Introduce new poke target, 'tailcall_bypass' to poke descriptor that is
dedicated for skipping the register pops and stack unwind that are
generated right before the actual jump to target program.
For case when the target program is not present, BPF program will skip
the pop instructions and nop5 dedicated for jmpq $target. An example of
such state when only R6 of callee saved registers is used by program:
ffffffffc0513aa1: e9 0e 00 00 00 jmpq 0xffffffffc0513ab4
ffffffffc0513aa6: 5b pop %rbx
ffffffffc0513aa7: 58 pop %rax
ffffffffc0513aa8: 48 81 c4 00 00 00 00 add $0x0,%rsp
ffffffffc0513aaf: 0f 1f 44 00 00 nopl 0x0(%rax,%rax,1)
ffffffffc0513ab4: 48 89 df mov %rbx,%rdi
When target program is inserted, the jump that was there to skip
pops/nop5 will become the nop5, so CPU will go over pops and do the
actual tailcall.
One might ask why there simply can not be pushes after the nop5?
In the following example snippet:
ffffffffc037030c: 48 89 fb mov %rdi,%rbx
(...)
ffffffffc0370332: 5b pop %rbx
ffffffffc0370333: 58 pop %rax
ffffffffc0370334: 48 81 c4 00 00 00 00 add $0x0,%rsp
ffffffffc037033b: 0f 1f 44 00 00 nopl 0x0(%rax,%rax,1)
ffffffffc0370340: 48 81 ec 00 00 00 00 sub $0x0,%rsp
ffffffffc0370347: 50 push %rax
ffffffffc0370348: 53 push %rbx
ffffffffc0370349: 48 89 df mov %rbx,%rdi
ffffffffc037034c: e8 f7 21 00 00 callq 0xffffffffc0372548
There is the bpf2bpf call (at ffffffffc037034c) right after the tailcall
and jump target is not present. ctx is in %rbx register and BPF
subprogram that we will call into on ffffffffc037034c is relying on it,
e.g. it will pick ctx from there. Such code layout is therefore broken
as we would overwrite the content of %rbx with the value that was pushed
on the prologue. That is the reason for the 'bypass' approach.
Special care needs to be taken during the install/update/remove of
tailcall target. In case when target program is not present, the CPU
must not execute the pop instructions that precede the tailcall.
To address that, the following states can be defined:
A nop, unwind, nop
B nop, unwind, tail
C skip, unwind, nop
D skip, unwind, tail
A is forbidden (lead to incorrectness). The state transitions between
tailcall install/update/remove will work as follows:
First install tail call f: C->D->B(f)
* poke the tailcall, after that get rid of the skip
Update tail call f to f': B(f)->B(f')
* poke the tailcall (poke->tailcall_target) and do NOT touch the
poke->tailcall_bypass
Remove tail call: B(f')->C(f')
* poke->tailcall_bypass is poked back to jump, then we wait the RCU
grace period so that other programs will finish its execution and
after that we are safe to remove the poke->tailcall_target
Install new tail call (f''): C(f')->D(f'')->B(f'').
* same as first step
This way CPU can never be exposed to "unwind, tail" state.
Last but not least, when tailcalls get mixed with bpf2bpf calls, it
would be possible to encounter the endless loop due to clearing the
tailcall counter if for example we would use the tailcall3-like from BPF
selftests program that would be subprogram-based, meaning the tailcall
would be present within the BPF subprogram.
This test, broken down to particular steps, would do:
entry -> set tailcall counter to 0, bump it by 1, tailcall to func0
func0 -> call subprog_tail
(we are NOT skipping the first 11 bytes of prologue and this subprogram
has a tailcall, therefore we clear the counter...)
subprog -> do the same thing as entry
and then loop forever.
To address this, the idea is to go through the call chain of bpf2bpf progs
and look for a tailcall presence throughout whole chain. If we saw a single
tail call then each node in this call chain needs to be marked as a subprog
that can reach the tailcall. We would later feed the JIT with this info
and:
- set eax to 0 only when tailcall is reachable and this is the entry prog
- if tailcall is reachable but there's no tailcall in insns of currently
JITed prog then push rax anyway, so that it will be possible to
propagate further down the call chain
- finally if tailcall is reachable, then we need to precede the 'call'
insn with mov rax, [rbp - (stack_depth + 8)]
Tail call related cases from test_verifier kselftest are also working
fine. Sample BPF programs that utilize tail calls (sockex3, tracex5)
work properly as well.
[1]: https://lore.kernel.org/bpf/20200517043227.2gpq22ifoq37ogst@ast-mbp.dhcp.thefacebook.com/
Suggested-by: Alexei Starovoitov <ast@kernel.org>
Signed-off-by: Maciej Fijalkowski <maciej.fijalkowski@intel.com>
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
2020-09-16 21:10:08 +00:00
|
|
|
if (tail_call_reachable)
|
|
|
|
EMIT1(0x50); /* push rax */
|
2015-05-19 23:59:04 +00:00
|
|
|
*pprog = prog;
|
|
|
|
}
|
|
|
|
|
bpf, x86: Emit patchable direct jump as tail call
Add initial code emission for *direct* jumps for tail call maps in
order to avoid the retpoline overhead from a493a87f38cf ("bpf, x64:
implement retpoline for tail call") for situations that allow for
it, meaning, for known constant keys at verification time which are
used as index into the tail call map. In case of Cilium which makes
heavy use of tail calls, constant keys are used in the vast majority,
only for a single occurrence we use a dynamic key.
High level outline is that if the target prog is NULL in the map, we
emit a 5-byte nop for the fall-through case and if not, we emit a
5-byte direct relative jmp to the target bpf_func + skipped prologue
offset. Later during runtime, we patch these 5-byte nop/jmps upon
tail call map update or deletions dynamically. Note that on x86-64
the direct jmp works as we reuse the same stack frame and skip
prologue (as opposed to some other JIT implementations).
One of the issues is that the tail call map slots can change at any
given time even during JITing. Therefore, we have two passes: i) emit
nops for all patchable locations during main JITing phase until we
declare prog->jited = 1 eventually. At this point the image is stable,
not public yet and with all jmps disabled. While JITing, we collect
additional info like poke->ip in order to remember the patch location
for later modifications. In ii) bpf_tail_call_direct_fixup() walks
over the progs poke_tab, locks the tail call maps poke_mutex to
prevent from parallel updates and patches in the right locations via
__bpf_arch_text_poke(). Note, the main bpf_arch_text_poke() cannot
be used at this point since we're not yet exposed to kallsyms. For
the update we use plain memcpy() since the image is not public and
still in read-write mode. After patching, we activate that poke entry
through poke->ip_stable. Meaning, at this point any tail call map
updates/deletions are not going to ignore that poke entry anymore.
Then, bpf_arch_text_poke() might still occur on the read-write image
until we finally locked it as read-only. Both modifications on the
given image are under text_mutex to avoid interference with each
other when update requests come in in parallel for different tail
call maps (current one we have locked in JIT and different one where
poke->ip_stable was already set).
Example prog:
# ./bpftool p d x i 1655
0: (b7) r3 = 0
1: (18) r2 = map[id:526]
3: (85) call bpf_tail_call#12
4: (b7) r0 = 1
5: (95) exit
Before:
# ./bpftool p d j i 1655
0xffffffffc076e55c:
0: nopl 0x0(%rax,%rax,1)
5: push %rbp
6: mov %rsp,%rbp
9: sub $0x200,%rsp
10: push %rbx
11: push %r13
13: push %r14
15: push %r15
17: pushq $0x0 _
19: xor %edx,%edx |_ index (arg 3)
1b: movabs $0xffff88d95cc82600,%rsi |_ map (arg 2)
25: mov %edx,%edx | index >= array->map.max_entries
27: cmp %edx,0x24(%rsi) |
2a: jbe 0x0000000000000066 |_
2c: mov -0x224(%rbp),%eax | tail call limit check
32: cmp $0x20,%eax |
35: ja 0x0000000000000066 |
37: add $0x1,%eax |
3a: mov %eax,-0x224(%rbp) |_
40: mov 0xd0(%rsi,%rdx,8),%rax |_ prog = array->ptrs[index]
48: test %rax,%rax | prog == NULL check
4b: je 0x0000000000000066 |_
4d: mov 0x30(%rax),%rax | goto *(prog->bpf_func + prologue_size)
51: add $0x19,%rax |
55: callq 0x0000000000000061 | retpoline for indirect jump
5a: pause |
5c: lfence |
5f: jmp 0x000000000000005a |
61: mov %rax,(%rsp) |
65: retq |_
66: mov $0x1,%eax
6b: pop %rbx
6c: pop %r15
6e: pop %r14
70: pop %r13
72: pop %rbx
73: leaveq
74: retq
After; state after JIT:
# ./bpftool p d j i 1655
0xffffffffc08e8930:
0: nopl 0x0(%rax,%rax,1)
5: push %rbp
6: mov %rsp,%rbp
9: sub $0x200,%rsp
10: push %rbx
11: push %r13
13: push %r14
15: push %r15
17: pushq $0x0 _
19: xor %edx,%edx |_ index (arg 3)
1b: movabs $0xffff9d8afd74c000,%rsi |_ map (arg 2)
25: mov -0x224(%rbp),%eax | tail call limit check
2b: cmp $0x20,%eax |
2e: ja 0x000000000000003e |
30: add $0x1,%eax |
33: mov %eax,-0x224(%rbp) |_
39: jmpq 0xfffffffffffd1785 |_ [direct] goto *(prog->bpf_func + prologue_size)
3e: mov $0x1,%eax
43: pop %rbx
44: pop %r15
46: pop %r14
48: pop %r13
4a: pop %rbx
4b: leaveq
4c: retq
After; state after map update (target prog):
# ./bpftool p d j i 1655
0xffffffffc08e8930:
0: nopl 0x0(%rax,%rax,1)
5: push %rbp
6: mov %rsp,%rbp
9: sub $0x200,%rsp
10: push %rbx
11: push %r13
13: push %r14
15: push %r15
17: pushq $0x0
19: xor %edx,%edx
1b: movabs $0xffff9d8afd74c000,%rsi
25: mov -0x224(%rbp),%eax
2b: cmp $0x20,%eax .
2e: ja 0x000000000000003e .
30: add $0x1,%eax .
33: mov %eax,-0x224(%rbp) |_
39: jmpq 0xffffffffffb09f55 |_ goto *(prog->bpf_func + prologue_size)
3e: mov $0x1,%eax
43: pop %rbx
44: pop %r15
46: pop %r14
48: pop %r13
4a: pop %rbx
4b: leaveq
4c: retq
After; state after map update (no prog):
# ./bpftool p d j i 1655
0xffffffffc08e8930:
0: nopl 0x0(%rax,%rax,1)
5: push %rbp
6: mov %rsp,%rbp
9: sub $0x200,%rsp
10: push %rbx
11: push %r13
13: push %r14
15: push %r15
17: pushq $0x0
19: xor %edx,%edx
1b: movabs $0xffff9d8afd74c000,%rsi
25: mov -0x224(%rbp),%eax
2b: cmp $0x20,%eax .
2e: ja 0x000000000000003e .
30: add $0x1,%eax .
33: mov %eax,-0x224(%rbp) |_
39: nopl 0x0(%rax,%rax,1) |_ fall-through nop
3e: mov $0x1,%eax
43: pop %rbx
44: pop %r15
46: pop %r14
48: pop %r13
4a: pop %rbx
4b: leaveq
4c: retq
Nice bonus is that this also shrinks the code emission quite a bit
for every tail call invocation.
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Link: https://lore.kernel.org/bpf/6ada4c1c9d35eeb5f4ecfab94593dafa6b5c4b09.1574452833.git.daniel@iogearbox.net
2019-11-22 20:08:00 +00:00
|
|
|
static int emit_patch(u8 **pprog, void *func, void *ip, u8 opcode)
|
|
|
|
{
|
|
|
|
u8 *prog = *pprog;
|
|
|
|
s64 offset;
|
|
|
|
|
|
|
|
offset = func - (ip + X86_PATCH_SIZE);
|
|
|
|
if (!is_simm32(offset)) {
|
|
|
|
pr_err("Target call %p is out of range\n", func);
|
|
|
|
return -ERANGE;
|
|
|
|
}
|
|
|
|
EMIT1_off32(opcode, offset);
|
|
|
|
*pprog = prog;
|
|
|
|
return 0;
|
|
|
|
}
|
|
|
|
|
|
|
|
static int emit_call(u8 **pprog, void *func, void *ip)
|
|
|
|
{
|
|
|
|
return emit_patch(pprog, func, ip, 0xE8);
|
|
|
|
}
|
|
|
|
|
2022-09-15 11:11:34 +00:00
|
|
|
static int emit_rsb_call(u8 **pprog, void *func, void *ip)
|
|
|
|
{
|
|
|
|
OPTIMIZER_HIDE_VAR(func);
|
2024-04-01 18:55:30 +00:00
|
|
|
ip += x86_call_depth_emit_accounting(pprog, func, ip);
|
2022-09-15 11:11:34 +00:00
|
|
|
return emit_patch(pprog, func, ip, 0xE8);
|
|
|
|
}
|
|
|
|
|
bpf, x86: Emit patchable direct jump as tail call
Add initial code emission for *direct* jumps for tail call maps in
order to avoid the retpoline overhead from a493a87f38cf ("bpf, x64:
implement retpoline for tail call") for situations that allow for
it, meaning, for known constant keys at verification time which are
used as index into the tail call map. In case of Cilium which makes
heavy use of tail calls, constant keys are used in the vast majority,
only for a single occurrence we use a dynamic key.
High level outline is that if the target prog is NULL in the map, we
emit a 5-byte nop for the fall-through case and if not, we emit a
5-byte direct relative jmp to the target bpf_func + skipped prologue
offset. Later during runtime, we patch these 5-byte nop/jmps upon
tail call map update or deletions dynamically. Note that on x86-64
the direct jmp works as we reuse the same stack frame and skip
prologue (as opposed to some other JIT implementations).
One of the issues is that the tail call map slots can change at any
given time even during JITing. Therefore, we have two passes: i) emit
nops for all patchable locations during main JITing phase until we
declare prog->jited = 1 eventually. At this point the image is stable,
not public yet and with all jmps disabled. While JITing, we collect
additional info like poke->ip in order to remember the patch location
for later modifications. In ii) bpf_tail_call_direct_fixup() walks
over the progs poke_tab, locks the tail call maps poke_mutex to
prevent from parallel updates and patches in the right locations via
__bpf_arch_text_poke(). Note, the main bpf_arch_text_poke() cannot
be used at this point since we're not yet exposed to kallsyms. For
the update we use plain memcpy() since the image is not public and
still in read-write mode. After patching, we activate that poke entry
through poke->ip_stable. Meaning, at this point any tail call map
updates/deletions are not going to ignore that poke entry anymore.
Then, bpf_arch_text_poke() might still occur on the read-write image
until we finally locked it as read-only. Both modifications on the
given image are under text_mutex to avoid interference with each
other when update requests come in in parallel for different tail
call maps (current one we have locked in JIT and different one where
poke->ip_stable was already set).
Example prog:
# ./bpftool p d x i 1655
0: (b7) r3 = 0
1: (18) r2 = map[id:526]
3: (85) call bpf_tail_call#12
4: (b7) r0 = 1
5: (95) exit
Before:
# ./bpftool p d j i 1655
0xffffffffc076e55c:
0: nopl 0x0(%rax,%rax,1)
5: push %rbp
6: mov %rsp,%rbp
9: sub $0x200,%rsp
10: push %rbx
11: push %r13
13: push %r14
15: push %r15
17: pushq $0x0 _
19: xor %edx,%edx |_ index (arg 3)
1b: movabs $0xffff88d95cc82600,%rsi |_ map (arg 2)
25: mov %edx,%edx | index >= array->map.max_entries
27: cmp %edx,0x24(%rsi) |
2a: jbe 0x0000000000000066 |_
2c: mov -0x224(%rbp),%eax | tail call limit check
32: cmp $0x20,%eax |
35: ja 0x0000000000000066 |
37: add $0x1,%eax |
3a: mov %eax,-0x224(%rbp) |_
40: mov 0xd0(%rsi,%rdx,8),%rax |_ prog = array->ptrs[index]
48: test %rax,%rax | prog == NULL check
4b: je 0x0000000000000066 |_
4d: mov 0x30(%rax),%rax | goto *(prog->bpf_func + prologue_size)
51: add $0x19,%rax |
55: callq 0x0000000000000061 | retpoline for indirect jump
5a: pause |
5c: lfence |
5f: jmp 0x000000000000005a |
61: mov %rax,(%rsp) |
65: retq |_
66: mov $0x1,%eax
6b: pop %rbx
6c: pop %r15
6e: pop %r14
70: pop %r13
72: pop %rbx
73: leaveq
74: retq
After; state after JIT:
# ./bpftool p d j i 1655
0xffffffffc08e8930:
0: nopl 0x0(%rax,%rax,1)
5: push %rbp
6: mov %rsp,%rbp
9: sub $0x200,%rsp
10: push %rbx
11: push %r13
13: push %r14
15: push %r15
17: pushq $0x0 _
19: xor %edx,%edx |_ index (arg 3)
1b: movabs $0xffff9d8afd74c000,%rsi |_ map (arg 2)
25: mov -0x224(%rbp),%eax | tail call limit check
2b: cmp $0x20,%eax |
2e: ja 0x000000000000003e |
30: add $0x1,%eax |
33: mov %eax,-0x224(%rbp) |_
39: jmpq 0xfffffffffffd1785 |_ [direct] goto *(prog->bpf_func + prologue_size)
3e: mov $0x1,%eax
43: pop %rbx
44: pop %r15
46: pop %r14
48: pop %r13
4a: pop %rbx
4b: leaveq
4c: retq
After; state after map update (target prog):
# ./bpftool p d j i 1655
0xffffffffc08e8930:
0: nopl 0x0(%rax,%rax,1)
5: push %rbp
6: mov %rsp,%rbp
9: sub $0x200,%rsp
10: push %rbx
11: push %r13
13: push %r14
15: push %r15
17: pushq $0x0
19: xor %edx,%edx
1b: movabs $0xffff9d8afd74c000,%rsi
25: mov -0x224(%rbp),%eax
2b: cmp $0x20,%eax .
2e: ja 0x000000000000003e .
30: add $0x1,%eax .
33: mov %eax,-0x224(%rbp) |_
39: jmpq 0xffffffffffb09f55 |_ goto *(prog->bpf_func + prologue_size)
3e: mov $0x1,%eax
43: pop %rbx
44: pop %r15
46: pop %r14
48: pop %r13
4a: pop %rbx
4b: leaveq
4c: retq
After; state after map update (no prog):
# ./bpftool p d j i 1655
0xffffffffc08e8930:
0: nopl 0x0(%rax,%rax,1)
5: push %rbp
6: mov %rsp,%rbp
9: sub $0x200,%rsp
10: push %rbx
11: push %r13
13: push %r14
15: push %r15
17: pushq $0x0
19: xor %edx,%edx
1b: movabs $0xffff9d8afd74c000,%rsi
25: mov -0x224(%rbp),%eax
2b: cmp $0x20,%eax .
2e: ja 0x000000000000003e .
30: add $0x1,%eax .
33: mov %eax,-0x224(%rbp) |_
39: nopl 0x0(%rax,%rax,1) |_ fall-through nop
3e: mov $0x1,%eax
43: pop %rbx
44: pop %r15
46: pop %r14
48: pop %r13
4a: pop %rbx
4b: leaveq
4c: retq
Nice bonus is that this also shrinks the code emission quite a bit
for every tail call invocation.
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Link: https://lore.kernel.org/bpf/6ada4c1c9d35eeb5f4ecfab94593dafa6b5c4b09.1574452833.git.daniel@iogearbox.net
2019-11-22 20:08:00 +00:00
|
|
|
static int emit_jump(u8 **pprog, void *func, void *ip)
|
|
|
|
{
|
|
|
|
return emit_patch(pprog, func, ip, 0xE9);
|
|
|
|
}
|
|
|
|
|
|
|
|
static int __bpf_arch_text_poke(void *ip, enum bpf_text_poke_type t,
|
2022-02-04 18:57:42 +00:00
|
|
|
void *old_addr, void *new_addr)
|
bpf, x86: Emit patchable direct jump as tail call
Add initial code emission for *direct* jumps for tail call maps in
order to avoid the retpoline overhead from a493a87f38cf ("bpf, x64:
implement retpoline for tail call") for situations that allow for
it, meaning, for known constant keys at verification time which are
used as index into the tail call map. In case of Cilium which makes
heavy use of tail calls, constant keys are used in the vast majority,
only for a single occurrence we use a dynamic key.
High level outline is that if the target prog is NULL in the map, we
emit a 5-byte nop for the fall-through case and if not, we emit a
5-byte direct relative jmp to the target bpf_func + skipped prologue
offset. Later during runtime, we patch these 5-byte nop/jmps upon
tail call map update or deletions dynamically. Note that on x86-64
the direct jmp works as we reuse the same stack frame and skip
prologue (as opposed to some other JIT implementations).
One of the issues is that the tail call map slots can change at any
given time even during JITing. Therefore, we have two passes: i) emit
nops for all patchable locations during main JITing phase until we
declare prog->jited = 1 eventually. At this point the image is stable,
not public yet and with all jmps disabled. While JITing, we collect
additional info like poke->ip in order to remember the patch location
for later modifications. In ii) bpf_tail_call_direct_fixup() walks
over the progs poke_tab, locks the tail call maps poke_mutex to
prevent from parallel updates and patches in the right locations via
__bpf_arch_text_poke(). Note, the main bpf_arch_text_poke() cannot
be used at this point since we're not yet exposed to kallsyms. For
the update we use plain memcpy() since the image is not public and
still in read-write mode. After patching, we activate that poke entry
through poke->ip_stable. Meaning, at this point any tail call map
updates/deletions are not going to ignore that poke entry anymore.
Then, bpf_arch_text_poke() might still occur on the read-write image
until we finally locked it as read-only. Both modifications on the
given image are under text_mutex to avoid interference with each
other when update requests come in in parallel for different tail
call maps (current one we have locked in JIT and different one where
poke->ip_stable was already set).
Example prog:
# ./bpftool p d x i 1655
0: (b7) r3 = 0
1: (18) r2 = map[id:526]
3: (85) call bpf_tail_call#12
4: (b7) r0 = 1
5: (95) exit
Before:
# ./bpftool p d j i 1655
0xffffffffc076e55c:
0: nopl 0x0(%rax,%rax,1)
5: push %rbp
6: mov %rsp,%rbp
9: sub $0x200,%rsp
10: push %rbx
11: push %r13
13: push %r14
15: push %r15
17: pushq $0x0 _
19: xor %edx,%edx |_ index (arg 3)
1b: movabs $0xffff88d95cc82600,%rsi |_ map (arg 2)
25: mov %edx,%edx | index >= array->map.max_entries
27: cmp %edx,0x24(%rsi) |
2a: jbe 0x0000000000000066 |_
2c: mov -0x224(%rbp),%eax | tail call limit check
32: cmp $0x20,%eax |
35: ja 0x0000000000000066 |
37: add $0x1,%eax |
3a: mov %eax,-0x224(%rbp) |_
40: mov 0xd0(%rsi,%rdx,8),%rax |_ prog = array->ptrs[index]
48: test %rax,%rax | prog == NULL check
4b: je 0x0000000000000066 |_
4d: mov 0x30(%rax),%rax | goto *(prog->bpf_func + prologue_size)
51: add $0x19,%rax |
55: callq 0x0000000000000061 | retpoline for indirect jump
5a: pause |
5c: lfence |
5f: jmp 0x000000000000005a |
61: mov %rax,(%rsp) |
65: retq |_
66: mov $0x1,%eax
6b: pop %rbx
6c: pop %r15
6e: pop %r14
70: pop %r13
72: pop %rbx
73: leaveq
74: retq
After; state after JIT:
# ./bpftool p d j i 1655
0xffffffffc08e8930:
0: nopl 0x0(%rax,%rax,1)
5: push %rbp
6: mov %rsp,%rbp
9: sub $0x200,%rsp
10: push %rbx
11: push %r13
13: push %r14
15: push %r15
17: pushq $0x0 _
19: xor %edx,%edx |_ index (arg 3)
1b: movabs $0xffff9d8afd74c000,%rsi |_ map (arg 2)
25: mov -0x224(%rbp),%eax | tail call limit check
2b: cmp $0x20,%eax |
2e: ja 0x000000000000003e |
30: add $0x1,%eax |
33: mov %eax,-0x224(%rbp) |_
39: jmpq 0xfffffffffffd1785 |_ [direct] goto *(prog->bpf_func + prologue_size)
3e: mov $0x1,%eax
43: pop %rbx
44: pop %r15
46: pop %r14
48: pop %r13
4a: pop %rbx
4b: leaveq
4c: retq
After; state after map update (target prog):
# ./bpftool p d j i 1655
0xffffffffc08e8930:
0: nopl 0x0(%rax,%rax,1)
5: push %rbp
6: mov %rsp,%rbp
9: sub $0x200,%rsp
10: push %rbx
11: push %r13
13: push %r14
15: push %r15
17: pushq $0x0
19: xor %edx,%edx
1b: movabs $0xffff9d8afd74c000,%rsi
25: mov -0x224(%rbp),%eax
2b: cmp $0x20,%eax .
2e: ja 0x000000000000003e .
30: add $0x1,%eax .
33: mov %eax,-0x224(%rbp) |_
39: jmpq 0xffffffffffb09f55 |_ goto *(prog->bpf_func + prologue_size)
3e: mov $0x1,%eax
43: pop %rbx
44: pop %r15
46: pop %r14
48: pop %r13
4a: pop %rbx
4b: leaveq
4c: retq
After; state after map update (no prog):
# ./bpftool p d j i 1655
0xffffffffc08e8930:
0: nopl 0x0(%rax,%rax,1)
5: push %rbp
6: mov %rsp,%rbp
9: sub $0x200,%rsp
10: push %rbx
11: push %r13
13: push %r14
15: push %r15
17: pushq $0x0
19: xor %edx,%edx
1b: movabs $0xffff9d8afd74c000,%rsi
25: mov -0x224(%rbp),%eax
2b: cmp $0x20,%eax .
2e: ja 0x000000000000003e .
30: add $0x1,%eax .
33: mov %eax,-0x224(%rbp) |_
39: nopl 0x0(%rax,%rax,1) |_ fall-through nop
3e: mov $0x1,%eax
43: pop %rbx
44: pop %r15
46: pop %r14
48: pop %r13
4a: pop %rbx
4b: leaveq
4c: retq
Nice bonus is that this also shrinks the code emission quite a bit
for every tail call invocation.
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Link: https://lore.kernel.org/bpf/6ada4c1c9d35eeb5f4ecfab94593dafa6b5c4b09.1574452833.git.daniel@iogearbox.net
2019-11-22 20:08:00 +00:00
|
|
|
{
|
x86: Remove dynamic NOP selection
This ensures that a NOP is a NOP and not a random other instruction that
is also a NOP. It allows simplification of dynamic code patching that
wants to verify existing code before writing new instructions (ftrace,
jump_label, static_call, etc..).
Differentiating on NOPs is not a feature.
This pessimises 32bit (DONTCARE) and 32bit on 64bit CPUs (CARELESS).
32bit is not a performance target.
Everything x86_64 since AMD K10 (2007) and Intel IvyBridge (2012) is
fine with using NOPL (as opposed to prefix NOP). And per FEATURE_NOPL
being required for x86_64, all x86_64 CPUs can use NOPL. So stop
caring about NOPs, simplify things and get on with life.
[ The problem seems to be that some uarchs can only decode NOPL on a
single front-end port while others have severe decode penalties for
excessive prefixes. All modern uarchs can handle both, except Atom,
which has prefix penalties. ]
[ Also, much doubt you can actually measure any of this on normal
workloads. ]
After this, FEATURE_NOPL is unused except for required-features for
x86_64. FEATURE_K8 is only used for PTI.
[ bp: Kernel build measurements showed ~0.3s slowdown on Sandybridge
which is hardly a slowdown. Get rid of X86_FEATURE_K7, while at it. ]
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Signed-off-by: Borislav Petkov <bp@suse.de>
Acked-by: Alexei Starovoitov <alexei.starovoitov@gmail.com> # bpf
Acked-by: Linus Torvalds <torvalds@linuxfoundation.org>
Link: https://lkml.kernel.org/r/20210312115749.065275711@infradead.org
2021-03-12 11:32:54 +00:00
|
|
|
const u8 *nop_insn = x86_nops[5];
|
2019-11-24 00:39:42 +00:00
|
|
|
u8 old_insn[X86_PATCH_SIZE];
|
|
|
|
u8 new_insn[X86_PATCH_SIZE];
|
bpf, x86: Emit patchable direct jump as tail call
Add initial code emission for *direct* jumps for tail call maps in
order to avoid the retpoline overhead from a493a87f38cf ("bpf, x64:
implement retpoline for tail call") for situations that allow for
it, meaning, for known constant keys at verification time which are
used as index into the tail call map. In case of Cilium which makes
heavy use of tail calls, constant keys are used in the vast majority,
only for a single occurrence we use a dynamic key.
High level outline is that if the target prog is NULL in the map, we
emit a 5-byte nop for the fall-through case and if not, we emit a
5-byte direct relative jmp to the target bpf_func + skipped prologue
offset. Later during runtime, we patch these 5-byte nop/jmps upon
tail call map update or deletions dynamically. Note that on x86-64
the direct jmp works as we reuse the same stack frame and skip
prologue (as opposed to some other JIT implementations).
One of the issues is that the tail call map slots can change at any
given time even during JITing. Therefore, we have two passes: i) emit
nops for all patchable locations during main JITing phase until we
declare prog->jited = 1 eventually. At this point the image is stable,
not public yet and with all jmps disabled. While JITing, we collect
additional info like poke->ip in order to remember the patch location
for later modifications. In ii) bpf_tail_call_direct_fixup() walks
over the progs poke_tab, locks the tail call maps poke_mutex to
prevent from parallel updates and patches in the right locations via
__bpf_arch_text_poke(). Note, the main bpf_arch_text_poke() cannot
be used at this point since we're not yet exposed to kallsyms. For
the update we use plain memcpy() since the image is not public and
still in read-write mode. After patching, we activate that poke entry
through poke->ip_stable. Meaning, at this point any tail call map
updates/deletions are not going to ignore that poke entry anymore.
Then, bpf_arch_text_poke() might still occur on the read-write image
until we finally locked it as read-only. Both modifications on the
given image are under text_mutex to avoid interference with each
other when update requests come in in parallel for different tail
call maps (current one we have locked in JIT and different one where
poke->ip_stable was already set).
Example prog:
# ./bpftool p d x i 1655
0: (b7) r3 = 0
1: (18) r2 = map[id:526]
3: (85) call bpf_tail_call#12
4: (b7) r0 = 1
5: (95) exit
Before:
# ./bpftool p d j i 1655
0xffffffffc076e55c:
0: nopl 0x0(%rax,%rax,1)
5: push %rbp
6: mov %rsp,%rbp
9: sub $0x200,%rsp
10: push %rbx
11: push %r13
13: push %r14
15: push %r15
17: pushq $0x0 _
19: xor %edx,%edx |_ index (arg 3)
1b: movabs $0xffff88d95cc82600,%rsi |_ map (arg 2)
25: mov %edx,%edx | index >= array->map.max_entries
27: cmp %edx,0x24(%rsi) |
2a: jbe 0x0000000000000066 |_
2c: mov -0x224(%rbp),%eax | tail call limit check
32: cmp $0x20,%eax |
35: ja 0x0000000000000066 |
37: add $0x1,%eax |
3a: mov %eax,-0x224(%rbp) |_
40: mov 0xd0(%rsi,%rdx,8),%rax |_ prog = array->ptrs[index]
48: test %rax,%rax | prog == NULL check
4b: je 0x0000000000000066 |_
4d: mov 0x30(%rax),%rax | goto *(prog->bpf_func + prologue_size)
51: add $0x19,%rax |
55: callq 0x0000000000000061 | retpoline for indirect jump
5a: pause |
5c: lfence |
5f: jmp 0x000000000000005a |
61: mov %rax,(%rsp) |
65: retq |_
66: mov $0x1,%eax
6b: pop %rbx
6c: pop %r15
6e: pop %r14
70: pop %r13
72: pop %rbx
73: leaveq
74: retq
After; state after JIT:
# ./bpftool p d j i 1655
0xffffffffc08e8930:
0: nopl 0x0(%rax,%rax,1)
5: push %rbp
6: mov %rsp,%rbp
9: sub $0x200,%rsp
10: push %rbx
11: push %r13
13: push %r14
15: push %r15
17: pushq $0x0 _
19: xor %edx,%edx |_ index (arg 3)
1b: movabs $0xffff9d8afd74c000,%rsi |_ map (arg 2)
25: mov -0x224(%rbp),%eax | tail call limit check
2b: cmp $0x20,%eax |
2e: ja 0x000000000000003e |
30: add $0x1,%eax |
33: mov %eax,-0x224(%rbp) |_
39: jmpq 0xfffffffffffd1785 |_ [direct] goto *(prog->bpf_func + prologue_size)
3e: mov $0x1,%eax
43: pop %rbx
44: pop %r15
46: pop %r14
48: pop %r13
4a: pop %rbx
4b: leaveq
4c: retq
After; state after map update (target prog):
# ./bpftool p d j i 1655
0xffffffffc08e8930:
0: nopl 0x0(%rax,%rax,1)
5: push %rbp
6: mov %rsp,%rbp
9: sub $0x200,%rsp
10: push %rbx
11: push %r13
13: push %r14
15: push %r15
17: pushq $0x0
19: xor %edx,%edx
1b: movabs $0xffff9d8afd74c000,%rsi
25: mov -0x224(%rbp),%eax
2b: cmp $0x20,%eax .
2e: ja 0x000000000000003e .
30: add $0x1,%eax .
33: mov %eax,-0x224(%rbp) |_
39: jmpq 0xffffffffffb09f55 |_ goto *(prog->bpf_func + prologue_size)
3e: mov $0x1,%eax
43: pop %rbx
44: pop %r15
46: pop %r14
48: pop %r13
4a: pop %rbx
4b: leaveq
4c: retq
After; state after map update (no prog):
# ./bpftool p d j i 1655
0xffffffffc08e8930:
0: nopl 0x0(%rax,%rax,1)
5: push %rbp
6: mov %rsp,%rbp
9: sub $0x200,%rsp
10: push %rbx
11: push %r13
13: push %r14
15: push %r15
17: pushq $0x0
19: xor %edx,%edx
1b: movabs $0xffff9d8afd74c000,%rsi
25: mov -0x224(%rbp),%eax
2b: cmp $0x20,%eax .
2e: ja 0x000000000000003e .
30: add $0x1,%eax .
33: mov %eax,-0x224(%rbp) |_
39: nopl 0x0(%rax,%rax,1) |_ fall-through nop
3e: mov $0x1,%eax
43: pop %rbx
44: pop %r15
46: pop %r14
48: pop %r13
4a: pop %rbx
4b: leaveq
4c: retq
Nice bonus is that this also shrinks the code emission quite a bit
for every tail call invocation.
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Link: https://lore.kernel.org/bpf/6ada4c1c9d35eeb5f4ecfab94593dafa6b5c4b09.1574452833.git.daniel@iogearbox.net
2019-11-22 20:08:00 +00:00
|
|
|
u8 *prog;
|
|
|
|
int ret;
|
|
|
|
|
2019-11-24 00:39:42 +00:00
|
|
|
memcpy(old_insn, nop_insn, X86_PATCH_SIZE);
|
|
|
|
if (old_addr) {
|
|
|
|
prog = old_insn;
|
|
|
|
ret = t == BPF_MOD_CALL ?
|
|
|
|
emit_call(&prog, old_addr, ip) :
|
|
|
|
emit_jump(&prog, old_addr, ip);
|
|
|
|
if (ret)
|
|
|
|
return ret;
|
bpf, x86: Emit patchable direct jump as tail call
Add initial code emission for *direct* jumps for tail call maps in
order to avoid the retpoline overhead from a493a87f38cf ("bpf, x64:
implement retpoline for tail call") for situations that allow for
it, meaning, for known constant keys at verification time which are
used as index into the tail call map. In case of Cilium which makes
heavy use of tail calls, constant keys are used in the vast majority,
only for a single occurrence we use a dynamic key.
High level outline is that if the target prog is NULL in the map, we
emit a 5-byte nop for the fall-through case and if not, we emit a
5-byte direct relative jmp to the target bpf_func + skipped prologue
offset. Later during runtime, we patch these 5-byte nop/jmps upon
tail call map update or deletions dynamically. Note that on x86-64
the direct jmp works as we reuse the same stack frame and skip
prologue (as opposed to some other JIT implementations).
One of the issues is that the tail call map slots can change at any
given time even during JITing. Therefore, we have two passes: i) emit
nops for all patchable locations during main JITing phase until we
declare prog->jited = 1 eventually. At this point the image is stable,
not public yet and with all jmps disabled. While JITing, we collect
additional info like poke->ip in order to remember the patch location
for later modifications. In ii) bpf_tail_call_direct_fixup() walks
over the progs poke_tab, locks the tail call maps poke_mutex to
prevent from parallel updates and patches in the right locations via
__bpf_arch_text_poke(). Note, the main bpf_arch_text_poke() cannot
be used at this point since we're not yet exposed to kallsyms. For
the update we use plain memcpy() since the image is not public and
still in read-write mode. After patching, we activate that poke entry
through poke->ip_stable. Meaning, at this point any tail call map
updates/deletions are not going to ignore that poke entry anymore.
Then, bpf_arch_text_poke() might still occur on the read-write image
until we finally locked it as read-only. Both modifications on the
given image are under text_mutex to avoid interference with each
other when update requests come in in parallel for different tail
call maps (current one we have locked in JIT and different one where
poke->ip_stable was already set).
Example prog:
# ./bpftool p d x i 1655
0: (b7) r3 = 0
1: (18) r2 = map[id:526]
3: (85) call bpf_tail_call#12
4: (b7) r0 = 1
5: (95) exit
Before:
# ./bpftool p d j i 1655
0xffffffffc076e55c:
0: nopl 0x0(%rax,%rax,1)
5: push %rbp
6: mov %rsp,%rbp
9: sub $0x200,%rsp
10: push %rbx
11: push %r13
13: push %r14
15: push %r15
17: pushq $0x0 _
19: xor %edx,%edx |_ index (arg 3)
1b: movabs $0xffff88d95cc82600,%rsi |_ map (arg 2)
25: mov %edx,%edx | index >= array->map.max_entries
27: cmp %edx,0x24(%rsi) |
2a: jbe 0x0000000000000066 |_
2c: mov -0x224(%rbp),%eax | tail call limit check
32: cmp $0x20,%eax |
35: ja 0x0000000000000066 |
37: add $0x1,%eax |
3a: mov %eax,-0x224(%rbp) |_
40: mov 0xd0(%rsi,%rdx,8),%rax |_ prog = array->ptrs[index]
48: test %rax,%rax | prog == NULL check
4b: je 0x0000000000000066 |_
4d: mov 0x30(%rax),%rax | goto *(prog->bpf_func + prologue_size)
51: add $0x19,%rax |
55: callq 0x0000000000000061 | retpoline for indirect jump
5a: pause |
5c: lfence |
5f: jmp 0x000000000000005a |
61: mov %rax,(%rsp) |
65: retq |_
66: mov $0x1,%eax
6b: pop %rbx
6c: pop %r15
6e: pop %r14
70: pop %r13
72: pop %rbx
73: leaveq
74: retq
After; state after JIT:
# ./bpftool p d j i 1655
0xffffffffc08e8930:
0: nopl 0x0(%rax,%rax,1)
5: push %rbp
6: mov %rsp,%rbp
9: sub $0x200,%rsp
10: push %rbx
11: push %r13
13: push %r14
15: push %r15
17: pushq $0x0 _
19: xor %edx,%edx |_ index (arg 3)
1b: movabs $0xffff9d8afd74c000,%rsi |_ map (arg 2)
25: mov -0x224(%rbp),%eax | tail call limit check
2b: cmp $0x20,%eax |
2e: ja 0x000000000000003e |
30: add $0x1,%eax |
33: mov %eax,-0x224(%rbp) |_
39: jmpq 0xfffffffffffd1785 |_ [direct] goto *(prog->bpf_func + prologue_size)
3e: mov $0x1,%eax
43: pop %rbx
44: pop %r15
46: pop %r14
48: pop %r13
4a: pop %rbx
4b: leaveq
4c: retq
After; state after map update (target prog):
# ./bpftool p d j i 1655
0xffffffffc08e8930:
0: nopl 0x0(%rax,%rax,1)
5: push %rbp
6: mov %rsp,%rbp
9: sub $0x200,%rsp
10: push %rbx
11: push %r13
13: push %r14
15: push %r15
17: pushq $0x0
19: xor %edx,%edx
1b: movabs $0xffff9d8afd74c000,%rsi
25: mov -0x224(%rbp),%eax
2b: cmp $0x20,%eax .
2e: ja 0x000000000000003e .
30: add $0x1,%eax .
33: mov %eax,-0x224(%rbp) |_
39: jmpq 0xffffffffffb09f55 |_ goto *(prog->bpf_func + prologue_size)
3e: mov $0x1,%eax
43: pop %rbx
44: pop %r15
46: pop %r14
48: pop %r13
4a: pop %rbx
4b: leaveq
4c: retq
After; state after map update (no prog):
# ./bpftool p d j i 1655
0xffffffffc08e8930:
0: nopl 0x0(%rax,%rax,1)
5: push %rbp
6: mov %rsp,%rbp
9: sub $0x200,%rsp
10: push %rbx
11: push %r13
13: push %r14
15: push %r15
17: pushq $0x0
19: xor %edx,%edx
1b: movabs $0xffff9d8afd74c000,%rsi
25: mov -0x224(%rbp),%eax
2b: cmp $0x20,%eax .
2e: ja 0x000000000000003e .
30: add $0x1,%eax .
33: mov %eax,-0x224(%rbp) |_
39: nopl 0x0(%rax,%rax,1) |_ fall-through nop
3e: mov $0x1,%eax
43: pop %rbx
44: pop %r15
46: pop %r14
48: pop %r13
4a: pop %rbx
4b: leaveq
4c: retq
Nice bonus is that this also shrinks the code emission quite a bit
for every tail call invocation.
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Link: https://lore.kernel.org/bpf/6ada4c1c9d35eeb5f4ecfab94593dafa6b5c4b09.1574452833.git.daniel@iogearbox.net
2019-11-22 20:08:00 +00:00
|
|
|
}
|
|
|
|
|
2019-11-24 00:39:42 +00:00
|
|
|
memcpy(new_insn, nop_insn, X86_PATCH_SIZE);
|
|
|
|
if (new_addr) {
|
|
|
|
prog = new_insn;
|
|
|
|
ret = t == BPF_MOD_CALL ?
|
|
|
|
emit_call(&prog, new_addr, ip) :
|
|
|
|
emit_jump(&prog, new_addr, ip);
|
|
|
|
if (ret)
|
|
|
|
return ret;
|
bpf, x86: Emit patchable direct jump as tail call
Add initial code emission for *direct* jumps for tail call maps in
order to avoid the retpoline overhead from a493a87f38cf ("bpf, x64:
implement retpoline for tail call") for situations that allow for
it, meaning, for known constant keys at verification time which are
used as index into the tail call map. In case of Cilium which makes
heavy use of tail calls, constant keys are used in the vast majority,
only for a single occurrence we use a dynamic key.
High level outline is that if the target prog is NULL in the map, we
emit a 5-byte nop for the fall-through case and if not, we emit a
5-byte direct relative jmp to the target bpf_func + skipped prologue
offset. Later during runtime, we patch these 5-byte nop/jmps upon
tail call map update or deletions dynamically. Note that on x86-64
the direct jmp works as we reuse the same stack frame and skip
prologue (as opposed to some other JIT implementations).
One of the issues is that the tail call map slots can change at any
given time even during JITing. Therefore, we have two passes: i) emit
nops for all patchable locations during main JITing phase until we
declare prog->jited = 1 eventually. At this point the image is stable,
not public yet and with all jmps disabled. While JITing, we collect
additional info like poke->ip in order to remember the patch location
for later modifications. In ii) bpf_tail_call_direct_fixup() walks
over the progs poke_tab, locks the tail call maps poke_mutex to
prevent from parallel updates and patches in the right locations via
__bpf_arch_text_poke(). Note, the main bpf_arch_text_poke() cannot
be used at this point since we're not yet exposed to kallsyms. For
the update we use plain memcpy() since the image is not public and
still in read-write mode. After patching, we activate that poke entry
through poke->ip_stable. Meaning, at this point any tail call map
updates/deletions are not going to ignore that poke entry anymore.
Then, bpf_arch_text_poke() might still occur on the read-write image
until we finally locked it as read-only. Both modifications on the
given image are under text_mutex to avoid interference with each
other when update requests come in in parallel for different tail
call maps (current one we have locked in JIT and different one where
poke->ip_stable was already set).
Example prog:
# ./bpftool p d x i 1655
0: (b7) r3 = 0
1: (18) r2 = map[id:526]
3: (85) call bpf_tail_call#12
4: (b7) r0 = 1
5: (95) exit
Before:
# ./bpftool p d j i 1655
0xffffffffc076e55c:
0: nopl 0x0(%rax,%rax,1)
5: push %rbp
6: mov %rsp,%rbp
9: sub $0x200,%rsp
10: push %rbx
11: push %r13
13: push %r14
15: push %r15
17: pushq $0x0 _
19: xor %edx,%edx |_ index (arg 3)
1b: movabs $0xffff88d95cc82600,%rsi |_ map (arg 2)
25: mov %edx,%edx | index >= array->map.max_entries
27: cmp %edx,0x24(%rsi) |
2a: jbe 0x0000000000000066 |_
2c: mov -0x224(%rbp),%eax | tail call limit check
32: cmp $0x20,%eax |
35: ja 0x0000000000000066 |
37: add $0x1,%eax |
3a: mov %eax,-0x224(%rbp) |_
40: mov 0xd0(%rsi,%rdx,8),%rax |_ prog = array->ptrs[index]
48: test %rax,%rax | prog == NULL check
4b: je 0x0000000000000066 |_
4d: mov 0x30(%rax),%rax | goto *(prog->bpf_func + prologue_size)
51: add $0x19,%rax |
55: callq 0x0000000000000061 | retpoline for indirect jump
5a: pause |
5c: lfence |
5f: jmp 0x000000000000005a |
61: mov %rax,(%rsp) |
65: retq |_
66: mov $0x1,%eax
6b: pop %rbx
6c: pop %r15
6e: pop %r14
70: pop %r13
72: pop %rbx
73: leaveq
74: retq
After; state after JIT:
# ./bpftool p d j i 1655
0xffffffffc08e8930:
0: nopl 0x0(%rax,%rax,1)
5: push %rbp
6: mov %rsp,%rbp
9: sub $0x200,%rsp
10: push %rbx
11: push %r13
13: push %r14
15: push %r15
17: pushq $0x0 _
19: xor %edx,%edx |_ index (arg 3)
1b: movabs $0xffff9d8afd74c000,%rsi |_ map (arg 2)
25: mov -0x224(%rbp),%eax | tail call limit check
2b: cmp $0x20,%eax |
2e: ja 0x000000000000003e |
30: add $0x1,%eax |
33: mov %eax,-0x224(%rbp) |_
39: jmpq 0xfffffffffffd1785 |_ [direct] goto *(prog->bpf_func + prologue_size)
3e: mov $0x1,%eax
43: pop %rbx
44: pop %r15
46: pop %r14
48: pop %r13
4a: pop %rbx
4b: leaveq
4c: retq
After; state after map update (target prog):
# ./bpftool p d j i 1655
0xffffffffc08e8930:
0: nopl 0x0(%rax,%rax,1)
5: push %rbp
6: mov %rsp,%rbp
9: sub $0x200,%rsp
10: push %rbx
11: push %r13
13: push %r14
15: push %r15
17: pushq $0x0
19: xor %edx,%edx
1b: movabs $0xffff9d8afd74c000,%rsi
25: mov -0x224(%rbp),%eax
2b: cmp $0x20,%eax .
2e: ja 0x000000000000003e .
30: add $0x1,%eax .
33: mov %eax,-0x224(%rbp) |_
39: jmpq 0xffffffffffb09f55 |_ goto *(prog->bpf_func + prologue_size)
3e: mov $0x1,%eax
43: pop %rbx
44: pop %r15
46: pop %r14
48: pop %r13
4a: pop %rbx
4b: leaveq
4c: retq
After; state after map update (no prog):
# ./bpftool p d j i 1655
0xffffffffc08e8930:
0: nopl 0x0(%rax,%rax,1)
5: push %rbp
6: mov %rsp,%rbp
9: sub $0x200,%rsp
10: push %rbx
11: push %r13
13: push %r14
15: push %r15
17: pushq $0x0
19: xor %edx,%edx
1b: movabs $0xffff9d8afd74c000,%rsi
25: mov -0x224(%rbp),%eax
2b: cmp $0x20,%eax .
2e: ja 0x000000000000003e .
30: add $0x1,%eax .
33: mov %eax,-0x224(%rbp) |_
39: nopl 0x0(%rax,%rax,1) |_ fall-through nop
3e: mov $0x1,%eax
43: pop %rbx
44: pop %r15
46: pop %r14
48: pop %r13
4a: pop %rbx
4b: leaveq
4c: retq
Nice bonus is that this also shrinks the code emission quite a bit
for every tail call invocation.
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Link: https://lore.kernel.org/bpf/6ada4c1c9d35eeb5f4ecfab94593dafa6b5c4b09.1574452833.git.daniel@iogearbox.net
2019-11-22 20:08:00 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
ret = -EBUSY;
|
|
|
|
mutex_lock(&text_mutex);
|
|
|
|
if (memcmp(ip, old_insn, X86_PATCH_SIZE))
|
|
|
|
goto out;
|
bpf, x64: rework pro/epilogue and tailcall handling in JIT
This commit serves two things:
1) it optimizes BPF prologue/epilogue generation
2) it makes possible to have tailcalls within BPF subprogram
Both points are related to each other since without 1), 2) could not be
achieved.
In [1], Alexei says:
"The prologue will look like:
nop5
xor eax,eax // two new bytes if bpf_tail_call() is used in this
// function
push rbp
mov rbp, rsp
sub rsp, rounded_stack_depth
push rax // zero init tail_call counter
variable number of push rbx,r13,r14,r15
Then bpf_tail_call will pop variable number rbx,..
and final 'pop rax'
Then 'add rsp, size_of_current_stack_frame'
jmp to next function and skip over 'nop5; xor eax,eax; push rpb; mov
rbp, rsp'
This way new function will set its own stack size and will init tail
call
counter with whatever value the parent had.
If next function doesn't use bpf_tail_call it won't have 'xor eax,eax'.
Instead it would need to have 'nop2' in there."
Implement that suggestion.
Since the layout of stack is changed, tail call counter handling can not
rely anymore on popping it to rbx just like it have been handled for
constant prologue case and later overwrite of rbx with actual value of
rbx pushed to stack. Therefore, let's use one of the register (%rcx) that
is considered to be volatile/caller-saved and pop the value of tail call
counter in there in the epilogue.
Drop the BUILD_BUG_ON in emit_prologue and in
emit_bpf_tail_call_indirect where instruction layout is not constant
anymore.
Introduce new poke target, 'tailcall_bypass' to poke descriptor that is
dedicated for skipping the register pops and stack unwind that are
generated right before the actual jump to target program.
For case when the target program is not present, BPF program will skip
the pop instructions and nop5 dedicated for jmpq $target. An example of
such state when only R6 of callee saved registers is used by program:
ffffffffc0513aa1: e9 0e 00 00 00 jmpq 0xffffffffc0513ab4
ffffffffc0513aa6: 5b pop %rbx
ffffffffc0513aa7: 58 pop %rax
ffffffffc0513aa8: 48 81 c4 00 00 00 00 add $0x0,%rsp
ffffffffc0513aaf: 0f 1f 44 00 00 nopl 0x0(%rax,%rax,1)
ffffffffc0513ab4: 48 89 df mov %rbx,%rdi
When target program is inserted, the jump that was there to skip
pops/nop5 will become the nop5, so CPU will go over pops and do the
actual tailcall.
One might ask why there simply can not be pushes after the nop5?
In the following example snippet:
ffffffffc037030c: 48 89 fb mov %rdi,%rbx
(...)
ffffffffc0370332: 5b pop %rbx
ffffffffc0370333: 58 pop %rax
ffffffffc0370334: 48 81 c4 00 00 00 00 add $0x0,%rsp
ffffffffc037033b: 0f 1f 44 00 00 nopl 0x0(%rax,%rax,1)
ffffffffc0370340: 48 81 ec 00 00 00 00 sub $0x0,%rsp
ffffffffc0370347: 50 push %rax
ffffffffc0370348: 53 push %rbx
ffffffffc0370349: 48 89 df mov %rbx,%rdi
ffffffffc037034c: e8 f7 21 00 00 callq 0xffffffffc0372548
There is the bpf2bpf call (at ffffffffc037034c) right after the tailcall
and jump target is not present. ctx is in %rbx register and BPF
subprogram that we will call into on ffffffffc037034c is relying on it,
e.g. it will pick ctx from there. Such code layout is therefore broken
as we would overwrite the content of %rbx with the value that was pushed
on the prologue. That is the reason for the 'bypass' approach.
Special care needs to be taken during the install/update/remove of
tailcall target. In case when target program is not present, the CPU
must not execute the pop instructions that precede the tailcall.
To address that, the following states can be defined:
A nop, unwind, nop
B nop, unwind, tail
C skip, unwind, nop
D skip, unwind, tail
A is forbidden (lead to incorrectness). The state transitions between
tailcall install/update/remove will work as follows:
First install tail call f: C->D->B(f)
* poke the tailcall, after that get rid of the skip
Update tail call f to f': B(f)->B(f')
* poke the tailcall (poke->tailcall_target) and do NOT touch the
poke->tailcall_bypass
Remove tail call: B(f')->C(f')
* poke->tailcall_bypass is poked back to jump, then we wait the RCU
grace period so that other programs will finish its execution and
after that we are safe to remove the poke->tailcall_target
Install new tail call (f''): C(f')->D(f'')->B(f'').
* same as first step
This way CPU can never be exposed to "unwind, tail" state.
Last but not least, when tailcalls get mixed with bpf2bpf calls, it
would be possible to encounter the endless loop due to clearing the
tailcall counter if for example we would use the tailcall3-like from BPF
selftests program that would be subprogram-based, meaning the tailcall
would be present within the BPF subprogram.
This test, broken down to particular steps, would do:
entry -> set tailcall counter to 0, bump it by 1, tailcall to func0
func0 -> call subprog_tail
(we are NOT skipping the first 11 bytes of prologue and this subprogram
has a tailcall, therefore we clear the counter...)
subprog -> do the same thing as entry
and then loop forever.
To address this, the idea is to go through the call chain of bpf2bpf progs
and look for a tailcall presence throughout whole chain. If we saw a single
tail call then each node in this call chain needs to be marked as a subprog
that can reach the tailcall. We would later feed the JIT with this info
and:
- set eax to 0 only when tailcall is reachable and this is the entry prog
- if tailcall is reachable but there's no tailcall in insns of currently
JITed prog then push rax anyway, so that it will be possible to
propagate further down the call chain
- finally if tailcall is reachable, then we need to precede the 'call'
insn with mov rax, [rbp - (stack_depth + 8)]
Tail call related cases from test_verifier kselftest are also working
fine. Sample BPF programs that utilize tail calls (sockex3, tracex5)
work properly as well.
[1]: https://lore.kernel.org/bpf/20200517043227.2gpq22ifoq37ogst@ast-mbp.dhcp.thefacebook.com/
Suggested-by: Alexei Starovoitov <ast@kernel.org>
Signed-off-by: Maciej Fijalkowski <maciej.fijalkowski@intel.com>
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
2020-09-16 21:10:08 +00:00
|
|
|
ret = 1;
|
2019-11-24 00:39:42 +00:00
|
|
|
if (memcmp(ip, new_insn, X86_PATCH_SIZE)) {
|
2022-02-04 18:57:42 +00:00
|
|
|
text_poke_bp(ip, new_insn, X86_PATCH_SIZE, NULL);
|
bpf, x64: rework pro/epilogue and tailcall handling in JIT
This commit serves two things:
1) it optimizes BPF prologue/epilogue generation
2) it makes possible to have tailcalls within BPF subprogram
Both points are related to each other since without 1), 2) could not be
achieved.
In [1], Alexei says:
"The prologue will look like:
nop5
xor eax,eax // two new bytes if bpf_tail_call() is used in this
// function
push rbp
mov rbp, rsp
sub rsp, rounded_stack_depth
push rax // zero init tail_call counter
variable number of push rbx,r13,r14,r15
Then bpf_tail_call will pop variable number rbx,..
and final 'pop rax'
Then 'add rsp, size_of_current_stack_frame'
jmp to next function and skip over 'nop5; xor eax,eax; push rpb; mov
rbp, rsp'
This way new function will set its own stack size and will init tail
call
counter with whatever value the parent had.
If next function doesn't use bpf_tail_call it won't have 'xor eax,eax'.
Instead it would need to have 'nop2' in there."
Implement that suggestion.
Since the layout of stack is changed, tail call counter handling can not
rely anymore on popping it to rbx just like it have been handled for
constant prologue case and later overwrite of rbx with actual value of
rbx pushed to stack. Therefore, let's use one of the register (%rcx) that
is considered to be volatile/caller-saved and pop the value of tail call
counter in there in the epilogue.
Drop the BUILD_BUG_ON in emit_prologue and in
emit_bpf_tail_call_indirect where instruction layout is not constant
anymore.
Introduce new poke target, 'tailcall_bypass' to poke descriptor that is
dedicated for skipping the register pops and stack unwind that are
generated right before the actual jump to target program.
For case when the target program is not present, BPF program will skip
the pop instructions and nop5 dedicated for jmpq $target. An example of
such state when only R6 of callee saved registers is used by program:
ffffffffc0513aa1: e9 0e 00 00 00 jmpq 0xffffffffc0513ab4
ffffffffc0513aa6: 5b pop %rbx
ffffffffc0513aa7: 58 pop %rax
ffffffffc0513aa8: 48 81 c4 00 00 00 00 add $0x0,%rsp
ffffffffc0513aaf: 0f 1f 44 00 00 nopl 0x0(%rax,%rax,1)
ffffffffc0513ab4: 48 89 df mov %rbx,%rdi
When target program is inserted, the jump that was there to skip
pops/nop5 will become the nop5, so CPU will go over pops and do the
actual tailcall.
One might ask why there simply can not be pushes after the nop5?
In the following example snippet:
ffffffffc037030c: 48 89 fb mov %rdi,%rbx
(...)
ffffffffc0370332: 5b pop %rbx
ffffffffc0370333: 58 pop %rax
ffffffffc0370334: 48 81 c4 00 00 00 00 add $0x0,%rsp
ffffffffc037033b: 0f 1f 44 00 00 nopl 0x0(%rax,%rax,1)
ffffffffc0370340: 48 81 ec 00 00 00 00 sub $0x0,%rsp
ffffffffc0370347: 50 push %rax
ffffffffc0370348: 53 push %rbx
ffffffffc0370349: 48 89 df mov %rbx,%rdi
ffffffffc037034c: e8 f7 21 00 00 callq 0xffffffffc0372548
There is the bpf2bpf call (at ffffffffc037034c) right after the tailcall
and jump target is not present. ctx is in %rbx register and BPF
subprogram that we will call into on ffffffffc037034c is relying on it,
e.g. it will pick ctx from there. Such code layout is therefore broken
as we would overwrite the content of %rbx with the value that was pushed
on the prologue. That is the reason for the 'bypass' approach.
Special care needs to be taken during the install/update/remove of
tailcall target. In case when target program is not present, the CPU
must not execute the pop instructions that precede the tailcall.
To address that, the following states can be defined:
A nop, unwind, nop
B nop, unwind, tail
C skip, unwind, nop
D skip, unwind, tail
A is forbidden (lead to incorrectness). The state transitions between
tailcall install/update/remove will work as follows:
First install tail call f: C->D->B(f)
* poke the tailcall, after that get rid of the skip
Update tail call f to f': B(f)->B(f')
* poke the tailcall (poke->tailcall_target) and do NOT touch the
poke->tailcall_bypass
Remove tail call: B(f')->C(f')
* poke->tailcall_bypass is poked back to jump, then we wait the RCU
grace period so that other programs will finish its execution and
after that we are safe to remove the poke->tailcall_target
Install new tail call (f''): C(f')->D(f'')->B(f'').
* same as first step
This way CPU can never be exposed to "unwind, tail" state.
Last but not least, when tailcalls get mixed with bpf2bpf calls, it
would be possible to encounter the endless loop due to clearing the
tailcall counter if for example we would use the tailcall3-like from BPF
selftests program that would be subprogram-based, meaning the tailcall
would be present within the BPF subprogram.
This test, broken down to particular steps, would do:
entry -> set tailcall counter to 0, bump it by 1, tailcall to func0
func0 -> call subprog_tail
(we are NOT skipping the first 11 bytes of prologue and this subprogram
has a tailcall, therefore we clear the counter...)
subprog -> do the same thing as entry
and then loop forever.
To address this, the idea is to go through the call chain of bpf2bpf progs
and look for a tailcall presence throughout whole chain. If we saw a single
tail call then each node in this call chain needs to be marked as a subprog
that can reach the tailcall. We would later feed the JIT with this info
and:
- set eax to 0 only when tailcall is reachable and this is the entry prog
- if tailcall is reachable but there's no tailcall in insns of currently
JITed prog then push rax anyway, so that it will be possible to
propagate further down the call chain
- finally if tailcall is reachable, then we need to precede the 'call'
insn with mov rax, [rbp - (stack_depth + 8)]
Tail call related cases from test_verifier kselftest are also working
fine. Sample BPF programs that utilize tail calls (sockex3, tracex5)
work properly as well.
[1]: https://lore.kernel.org/bpf/20200517043227.2gpq22ifoq37ogst@ast-mbp.dhcp.thefacebook.com/
Suggested-by: Alexei Starovoitov <ast@kernel.org>
Signed-off-by: Maciej Fijalkowski <maciej.fijalkowski@intel.com>
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
2020-09-16 21:10:08 +00:00
|
|
|
ret = 0;
|
2019-11-24 00:39:42 +00:00
|
|
|
}
|
bpf, x86: Emit patchable direct jump as tail call
Add initial code emission for *direct* jumps for tail call maps in
order to avoid the retpoline overhead from a493a87f38cf ("bpf, x64:
implement retpoline for tail call") for situations that allow for
it, meaning, for known constant keys at verification time which are
used as index into the tail call map. In case of Cilium which makes
heavy use of tail calls, constant keys are used in the vast majority,
only for a single occurrence we use a dynamic key.
High level outline is that if the target prog is NULL in the map, we
emit a 5-byte nop for the fall-through case and if not, we emit a
5-byte direct relative jmp to the target bpf_func + skipped prologue
offset. Later during runtime, we patch these 5-byte nop/jmps upon
tail call map update or deletions dynamically. Note that on x86-64
the direct jmp works as we reuse the same stack frame and skip
prologue (as opposed to some other JIT implementations).
One of the issues is that the tail call map slots can change at any
given time even during JITing. Therefore, we have two passes: i) emit
nops for all patchable locations during main JITing phase until we
declare prog->jited = 1 eventually. At this point the image is stable,
not public yet and with all jmps disabled. While JITing, we collect
additional info like poke->ip in order to remember the patch location
for later modifications. In ii) bpf_tail_call_direct_fixup() walks
over the progs poke_tab, locks the tail call maps poke_mutex to
prevent from parallel updates and patches in the right locations via
__bpf_arch_text_poke(). Note, the main bpf_arch_text_poke() cannot
be used at this point since we're not yet exposed to kallsyms. For
the update we use plain memcpy() since the image is not public and
still in read-write mode. After patching, we activate that poke entry
through poke->ip_stable. Meaning, at this point any tail call map
updates/deletions are not going to ignore that poke entry anymore.
Then, bpf_arch_text_poke() might still occur on the read-write image
until we finally locked it as read-only. Both modifications on the
given image are under text_mutex to avoid interference with each
other when update requests come in in parallel for different tail
call maps (current one we have locked in JIT and different one where
poke->ip_stable was already set).
Example prog:
# ./bpftool p d x i 1655
0: (b7) r3 = 0
1: (18) r2 = map[id:526]
3: (85) call bpf_tail_call#12
4: (b7) r0 = 1
5: (95) exit
Before:
# ./bpftool p d j i 1655
0xffffffffc076e55c:
0: nopl 0x0(%rax,%rax,1)
5: push %rbp
6: mov %rsp,%rbp
9: sub $0x200,%rsp
10: push %rbx
11: push %r13
13: push %r14
15: push %r15
17: pushq $0x0 _
19: xor %edx,%edx |_ index (arg 3)
1b: movabs $0xffff88d95cc82600,%rsi |_ map (arg 2)
25: mov %edx,%edx | index >= array->map.max_entries
27: cmp %edx,0x24(%rsi) |
2a: jbe 0x0000000000000066 |_
2c: mov -0x224(%rbp),%eax | tail call limit check
32: cmp $0x20,%eax |
35: ja 0x0000000000000066 |
37: add $0x1,%eax |
3a: mov %eax,-0x224(%rbp) |_
40: mov 0xd0(%rsi,%rdx,8),%rax |_ prog = array->ptrs[index]
48: test %rax,%rax | prog == NULL check
4b: je 0x0000000000000066 |_
4d: mov 0x30(%rax),%rax | goto *(prog->bpf_func + prologue_size)
51: add $0x19,%rax |
55: callq 0x0000000000000061 | retpoline for indirect jump
5a: pause |
5c: lfence |
5f: jmp 0x000000000000005a |
61: mov %rax,(%rsp) |
65: retq |_
66: mov $0x1,%eax
6b: pop %rbx
6c: pop %r15
6e: pop %r14
70: pop %r13
72: pop %rbx
73: leaveq
74: retq
After; state after JIT:
# ./bpftool p d j i 1655
0xffffffffc08e8930:
0: nopl 0x0(%rax,%rax,1)
5: push %rbp
6: mov %rsp,%rbp
9: sub $0x200,%rsp
10: push %rbx
11: push %r13
13: push %r14
15: push %r15
17: pushq $0x0 _
19: xor %edx,%edx |_ index (arg 3)
1b: movabs $0xffff9d8afd74c000,%rsi |_ map (arg 2)
25: mov -0x224(%rbp),%eax | tail call limit check
2b: cmp $0x20,%eax |
2e: ja 0x000000000000003e |
30: add $0x1,%eax |
33: mov %eax,-0x224(%rbp) |_
39: jmpq 0xfffffffffffd1785 |_ [direct] goto *(prog->bpf_func + prologue_size)
3e: mov $0x1,%eax
43: pop %rbx
44: pop %r15
46: pop %r14
48: pop %r13
4a: pop %rbx
4b: leaveq
4c: retq
After; state after map update (target prog):
# ./bpftool p d j i 1655
0xffffffffc08e8930:
0: nopl 0x0(%rax,%rax,1)
5: push %rbp
6: mov %rsp,%rbp
9: sub $0x200,%rsp
10: push %rbx
11: push %r13
13: push %r14
15: push %r15
17: pushq $0x0
19: xor %edx,%edx
1b: movabs $0xffff9d8afd74c000,%rsi
25: mov -0x224(%rbp),%eax
2b: cmp $0x20,%eax .
2e: ja 0x000000000000003e .
30: add $0x1,%eax .
33: mov %eax,-0x224(%rbp) |_
39: jmpq 0xffffffffffb09f55 |_ goto *(prog->bpf_func + prologue_size)
3e: mov $0x1,%eax
43: pop %rbx
44: pop %r15
46: pop %r14
48: pop %r13
4a: pop %rbx
4b: leaveq
4c: retq
After; state after map update (no prog):
# ./bpftool p d j i 1655
0xffffffffc08e8930:
0: nopl 0x0(%rax,%rax,1)
5: push %rbp
6: mov %rsp,%rbp
9: sub $0x200,%rsp
10: push %rbx
11: push %r13
13: push %r14
15: push %r15
17: pushq $0x0
19: xor %edx,%edx
1b: movabs $0xffff9d8afd74c000,%rsi
25: mov -0x224(%rbp),%eax
2b: cmp $0x20,%eax .
2e: ja 0x000000000000003e .
30: add $0x1,%eax .
33: mov %eax,-0x224(%rbp) |_
39: nopl 0x0(%rax,%rax,1) |_ fall-through nop
3e: mov $0x1,%eax
43: pop %rbx
44: pop %r15
46: pop %r14
48: pop %r13
4a: pop %rbx
4b: leaveq
4c: retq
Nice bonus is that this also shrinks the code emission quite a bit
for every tail call invocation.
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Link: https://lore.kernel.org/bpf/6ada4c1c9d35eeb5f4ecfab94593dafa6b5c4b09.1574452833.git.daniel@iogearbox.net
2019-11-22 20:08:00 +00:00
|
|
|
out:
|
|
|
|
mutex_unlock(&text_mutex);
|
|
|
|
return ret;
|
|
|
|
}
|
|
|
|
|
|
|
|
int bpf_arch_text_poke(void *ip, enum bpf_text_poke_type t,
|
|
|
|
void *old_addr, void *new_addr)
|
|
|
|
{
|
|
|
|
if (!is_kernel_text((long)ip) &&
|
|
|
|
!is_bpf_text_address((long)ip))
|
|
|
|
/* BPF poking in modules is not supported */
|
|
|
|
return -EINVAL;
|
|
|
|
|
2022-03-08 15:30:33 +00:00
|
|
|
/*
|
|
|
|
* See emit_prologue(), for IBT builds the trampoline hook is preceded
|
|
|
|
* with an ENDBR instruction.
|
|
|
|
*/
|
|
|
|
if (is_endbr(*(u32 *)ip))
|
|
|
|
ip += ENDBR_INSN_SIZE;
|
|
|
|
|
2022-02-04 18:57:42 +00:00
|
|
|
return __bpf_arch_text_poke(ip, t, old_addr, new_addr);
|
bpf, x86: Emit patchable direct jump as tail call
Add initial code emission for *direct* jumps for tail call maps in
order to avoid the retpoline overhead from a493a87f38cf ("bpf, x64:
implement retpoline for tail call") for situations that allow for
it, meaning, for known constant keys at verification time which are
used as index into the tail call map. In case of Cilium which makes
heavy use of tail calls, constant keys are used in the vast majority,
only for a single occurrence we use a dynamic key.
High level outline is that if the target prog is NULL in the map, we
emit a 5-byte nop for the fall-through case and if not, we emit a
5-byte direct relative jmp to the target bpf_func + skipped prologue
offset. Later during runtime, we patch these 5-byte nop/jmps upon
tail call map update or deletions dynamically. Note that on x86-64
the direct jmp works as we reuse the same stack frame and skip
prologue (as opposed to some other JIT implementations).
One of the issues is that the tail call map slots can change at any
given time even during JITing. Therefore, we have two passes: i) emit
nops for all patchable locations during main JITing phase until we
declare prog->jited = 1 eventually. At this point the image is stable,
not public yet and with all jmps disabled. While JITing, we collect
additional info like poke->ip in order to remember the patch location
for later modifications. In ii) bpf_tail_call_direct_fixup() walks
over the progs poke_tab, locks the tail call maps poke_mutex to
prevent from parallel updates and patches in the right locations via
__bpf_arch_text_poke(). Note, the main bpf_arch_text_poke() cannot
be used at this point since we're not yet exposed to kallsyms. For
the update we use plain memcpy() since the image is not public and
still in read-write mode. After patching, we activate that poke entry
through poke->ip_stable. Meaning, at this point any tail call map
updates/deletions are not going to ignore that poke entry anymore.
Then, bpf_arch_text_poke() might still occur on the read-write image
until we finally locked it as read-only. Both modifications on the
given image are under text_mutex to avoid interference with each
other when update requests come in in parallel for different tail
call maps (current one we have locked in JIT and different one where
poke->ip_stable was already set).
Example prog:
# ./bpftool p d x i 1655
0: (b7) r3 = 0
1: (18) r2 = map[id:526]
3: (85) call bpf_tail_call#12
4: (b7) r0 = 1
5: (95) exit
Before:
# ./bpftool p d j i 1655
0xffffffffc076e55c:
0: nopl 0x0(%rax,%rax,1)
5: push %rbp
6: mov %rsp,%rbp
9: sub $0x200,%rsp
10: push %rbx
11: push %r13
13: push %r14
15: push %r15
17: pushq $0x0 _
19: xor %edx,%edx |_ index (arg 3)
1b: movabs $0xffff88d95cc82600,%rsi |_ map (arg 2)
25: mov %edx,%edx | index >= array->map.max_entries
27: cmp %edx,0x24(%rsi) |
2a: jbe 0x0000000000000066 |_
2c: mov -0x224(%rbp),%eax | tail call limit check
32: cmp $0x20,%eax |
35: ja 0x0000000000000066 |
37: add $0x1,%eax |
3a: mov %eax,-0x224(%rbp) |_
40: mov 0xd0(%rsi,%rdx,8),%rax |_ prog = array->ptrs[index]
48: test %rax,%rax | prog == NULL check
4b: je 0x0000000000000066 |_
4d: mov 0x30(%rax),%rax | goto *(prog->bpf_func + prologue_size)
51: add $0x19,%rax |
55: callq 0x0000000000000061 | retpoline for indirect jump
5a: pause |
5c: lfence |
5f: jmp 0x000000000000005a |
61: mov %rax,(%rsp) |
65: retq |_
66: mov $0x1,%eax
6b: pop %rbx
6c: pop %r15
6e: pop %r14
70: pop %r13
72: pop %rbx
73: leaveq
74: retq
After; state after JIT:
# ./bpftool p d j i 1655
0xffffffffc08e8930:
0: nopl 0x0(%rax,%rax,1)
5: push %rbp
6: mov %rsp,%rbp
9: sub $0x200,%rsp
10: push %rbx
11: push %r13
13: push %r14
15: push %r15
17: pushq $0x0 _
19: xor %edx,%edx |_ index (arg 3)
1b: movabs $0xffff9d8afd74c000,%rsi |_ map (arg 2)
25: mov -0x224(%rbp),%eax | tail call limit check
2b: cmp $0x20,%eax |
2e: ja 0x000000000000003e |
30: add $0x1,%eax |
33: mov %eax,-0x224(%rbp) |_
39: jmpq 0xfffffffffffd1785 |_ [direct] goto *(prog->bpf_func + prologue_size)
3e: mov $0x1,%eax
43: pop %rbx
44: pop %r15
46: pop %r14
48: pop %r13
4a: pop %rbx
4b: leaveq
4c: retq
After; state after map update (target prog):
# ./bpftool p d j i 1655
0xffffffffc08e8930:
0: nopl 0x0(%rax,%rax,1)
5: push %rbp
6: mov %rsp,%rbp
9: sub $0x200,%rsp
10: push %rbx
11: push %r13
13: push %r14
15: push %r15
17: pushq $0x0
19: xor %edx,%edx
1b: movabs $0xffff9d8afd74c000,%rsi
25: mov -0x224(%rbp),%eax
2b: cmp $0x20,%eax .
2e: ja 0x000000000000003e .
30: add $0x1,%eax .
33: mov %eax,-0x224(%rbp) |_
39: jmpq 0xffffffffffb09f55 |_ goto *(prog->bpf_func + prologue_size)
3e: mov $0x1,%eax
43: pop %rbx
44: pop %r15
46: pop %r14
48: pop %r13
4a: pop %rbx
4b: leaveq
4c: retq
After; state after map update (no prog):
# ./bpftool p d j i 1655
0xffffffffc08e8930:
0: nopl 0x0(%rax,%rax,1)
5: push %rbp
6: mov %rsp,%rbp
9: sub $0x200,%rsp
10: push %rbx
11: push %r13
13: push %r14
15: push %r15
17: pushq $0x0
19: xor %edx,%edx
1b: movabs $0xffff9d8afd74c000,%rsi
25: mov -0x224(%rbp),%eax
2b: cmp $0x20,%eax .
2e: ja 0x000000000000003e .
30: add $0x1,%eax .
33: mov %eax,-0x224(%rbp) |_
39: nopl 0x0(%rax,%rax,1) |_ fall-through nop
3e: mov $0x1,%eax
43: pop %rbx
44: pop %r15
46: pop %r14
48: pop %r13
4a: pop %rbx
4b: leaveq
4c: retq
Nice bonus is that this also shrinks the code emission quite a bit
for every tail call invocation.
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Link: https://lore.kernel.org/bpf/6ada4c1c9d35eeb5f4ecfab94593dafa6b5c4b09.1574452833.git.daniel@iogearbox.net
2019-11-22 20:08:00 +00:00
|
|
|
}
|
|
|
|
|
2021-10-26 12:01:48 +00:00
|
|
|
#define EMIT_LFENCE() EMIT3(0x0F, 0xAE, 0xE8)
|
|
|
|
|
|
|
|
static void emit_indirect_jump(u8 **pprog, int reg, u8 *ip)
|
|
|
|
{
|
|
|
|
u8 *prog = *pprog;
|
|
|
|
|
2022-02-16 19:57:00 +00:00
|
|
|
if (cpu_feature_enabled(X86_FEATURE_RETPOLINE_LFENCE)) {
|
2021-10-26 12:01:48 +00:00
|
|
|
EMIT_LFENCE();
|
|
|
|
EMIT2(0xFF, 0xE0 + reg);
|
|
|
|
} else if (cpu_feature_enabled(X86_FEATURE_RETPOLINE)) {
|
2022-03-28 11:13:41 +00:00
|
|
|
OPTIMIZER_HIDE_VAR(reg);
|
2022-09-15 11:11:28 +00:00
|
|
|
if (cpu_feature_enabled(X86_FEATURE_CALL_DEPTH))
|
|
|
|
emit_jump(&prog, &__x86_indirect_jump_thunk_array[reg], ip);
|
|
|
|
else
|
|
|
|
emit_jump(&prog, &__x86_indirect_thunk_array[reg], ip);
|
2022-06-14 21:15:34 +00:00
|
|
|
} else {
|
2022-09-08 10:04:50 +00:00
|
|
|
EMIT2(0xFF, 0xE0 + reg); /* jmp *%\reg */
|
2023-11-21 16:07:33 +00:00
|
|
|
if (IS_ENABLED(CONFIG_MITIGATION_RETPOLINE) || IS_ENABLED(CONFIG_MITIGATION_SLS))
|
2022-09-08 10:04:50 +00:00
|
|
|
EMIT1(0xCC); /* int3 */
|
2022-06-14 21:15:34 +00:00
|
|
|
}
|
2021-10-26 12:01:48 +00:00
|
|
|
|
|
|
|
*pprog = prog;
|
|
|
|
}
|
|
|
|
|
2022-06-14 21:15:41 +00:00
|
|
|
static void emit_return(u8 **pprog, u8 *ip)
|
|
|
|
{
|
|
|
|
u8 *prog = *pprog;
|
|
|
|
|
|
|
|
if (cpu_feature_enabled(X86_FEATURE_RETHUNK)) {
|
2022-09-15 11:11:25 +00:00
|
|
|
emit_jump(&prog, x86_return_thunk, ip);
|
2022-06-14 21:15:41 +00:00
|
|
|
} else {
|
|
|
|
EMIT1(0xC3); /* ret */
|
2023-11-21 16:07:33 +00:00
|
|
|
if (IS_ENABLED(CONFIG_MITIGATION_SLS))
|
2022-06-14 21:15:41 +00:00
|
|
|
EMIT1(0xCC); /* int3 */
|
|
|
|
}
|
2021-10-26 12:01:48 +00:00
|
|
|
|
|
|
|
*pprog = prog;
|
|
|
|
}
|
|
|
|
|
x86/bpf: Clean up non-standard comments, to make the code more readable
So by chance I looked into x86 assembly in arch/x86/net/bpf_jit_comp.c and
noticed the weird and inconsistent comment style it mistakenly learned from
the networking code:
/* Multi-line comment ...
* ... looks like this.
*/
Fix this to use the standard comment style specified in Documentation/CodingStyle
and used in arch/x86/ as well:
/*
* Multi-line comment ...
* ... looks like this.
*/
Also, to quote Linus's ... more explicit views about this:
http://article.gmane.org/gmane.linux.kernel.cryptoapi/21066
> But no, the networking code picked *none* of the above sane formats.
> Instead, it picked these two models that are just half-arsed
> shit-for-brains:
>
> (no)
> /* This is disgusting drug-induced
> * crap, and should die
> */
>
> (no-no-no)
> /* This is also very nasty
> * and visually unbalanced */
>
> Please. The networking code actually has the *worst* possible comment
> style. You can literally find that (no-no-no) style, which is just
> really horribly disgusting and worse than the otherwise fairly similar
> (d) in pretty much every way.
Also improve the comments and some other details while at it:
- Don't mix same-line and previous-line comment style on otherwise
identical code patterns within the same function,
- capitalize 'BPF' and x86 register names consistently,
- capitalize sentences consistently,
- instead of 'x64' use 'x86-64': x64 is a Microsoft specific term,
- use more consistent punctuation,
- use standard coding style in macros as well,
- fix typos and a few other minor details.
Consistent coding style is not optional, at least in arch/x86/.
No change in functionality.
( In case this commit causes conflicts with pending development code
I'll be glad to help resolve any conflicts! )
Acked-by: Thomas Gleixner <tglx@linutronix.de>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Borislav Petkov <bp@alien8.de>
Cc: H. Peter Anvin <hpa@zytor.com>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: David S. Miller <davem@davemloft.net>
Cc: Eric Dumazet <edumazet@google.com>
Cc: Daniel Borkmann <daniel@iogearbox.net>
Cc: Alexei Starovoitov <ast@fb.com>
Cc: Hideaki YOSHIFUJI <yoshfuji@linux-ipv6.org>
Cc: netdev@vger.kernel.org
Cc: linux-kernel@vger.kernel.org
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
2018-04-27 09:54:40 +00:00
|
|
|
/*
|
|
|
|
* Generate the following code:
|
|
|
|
*
|
2015-05-19 23:59:04 +00:00
|
|
|
* ... bpf_tail_call(void *ctx, struct bpf_array *array, u64 index) ...
|
|
|
|
* if (index >= array->map.max_entries)
|
|
|
|
* goto out;
|
bpf: Change value of MAX_TAIL_CALL_CNT from 32 to 33
In the current code, the actual max tail call count is 33 which is greater
than MAX_TAIL_CALL_CNT (defined as 32). The actual limit is not consistent
with the meaning of MAX_TAIL_CALL_CNT and thus confusing at first glance.
We can see the historical evolution from commit 04fd61ab36ec ("bpf: allow
bpf programs to tail-call other bpf programs") and commit f9dabe016b63
("bpf: Undo off-by-one in interpreter tail call count limit"). In order
to avoid changing existing behavior, the actual limit is 33 now, this is
reasonable.
After commit 874be05f525e ("bpf, tests: Add tail call test suite"), we can
see there exists failed testcase.
On all archs when CONFIG_BPF_JIT_ALWAYS_ON is not set:
# echo 0 > /proc/sys/net/core/bpf_jit_enable
# modprobe test_bpf
# dmesg | grep -w FAIL
Tail call error path, max count reached jited:0 ret 34 != 33 FAIL
On some archs:
# echo 1 > /proc/sys/net/core/bpf_jit_enable
# modprobe test_bpf
# dmesg | grep -w FAIL
Tail call error path, max count reached jited:1 ret 34 != 33 FAIL
Although the above failed testcase has been fixed in commit 18935a72eb25
("bpf/tests: Fix error in tail call limit tests"), it would still be good
to change the value of MAX_TAIL_CALL_CNT from 32 to 33 to make the code
more readable.
The 32-bit x86 JIT was using a limit of 32, just fix the wrong comments and
limit to 33 tail calls as the constant MAX_TAIL_CALL_CNT updated. For the
mips64 JIT, use "ori" instead of "addiu" as suggested by Johan Almbladh.
For the riscv JIT, use RV_REG_TCC directly to save one register move as
suggested by Björn Töpel. For the other implementations, no function changes,
it does not change the current limit 33, the new value of MAX_TAIL_CALL_CNT
can reflect the actual max tail call count, the related tail call testcases
in test_bpf module and selftests can work well for the interpreter and the
JIT.
Here are the test results on x86_64:
# uname -m
x86_64
# echo 0 > /proc/sys/net/core/bpf_jit_enable
# modprobe test_bpf test_suite=test_tail_calls
# dmesg | tail -1
test_bpf: test_tail_calls: Summary: 8 PASSED, 0 FAILED, [0/8 JIT'ed]
# rmmod test_bpf
# echo 1 > /proc/sys/net/core/bpf_jit_enable
# modprobe test_bpf test_suite=test_tail_calls
# dmesg | tail -1
test_bpf: test_tail_calls: Summary: 8 PASSED, 0 FAILED, [8/8 JIT'ed]
# rmmod test_bpf
# ./test_progs -t tailcalls
#142 tailcalls:OK
Summary: 1/11 PASSED, 0 SKIPPED, 0 FAILED
Signed-off-by: Tiezhu Yang <yangtiezhu@loongson.cn>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Tested-by: Johan Almbladh <johan.almbladh@anyfinetworks.com>
Tested-by: Ilya Leoshkevich <iii@linux.ibm.com>
Acked-by: Björn Töpel <bjorn@kernel.org>
Acked-by: Johan Almbladh <johan.almbladh@anyfinetworks.com>
Acked-by: Ilya Leoshkevich <iii@linux.ibm.com>
Link: https://lore.kernel.org/bpf/1636075800-3264-1-git-send-email-yangtiezhu@loongson.cn
2021-11-05 01:30:00 +00:00
|
|
|
* if (tail_call_cnt++ >= MAX_TAIL_CALL_CNT)
|
2015-05-19 23:59:04 +00:00
|
|
|
* goto out;
|
2015-08-06 07:02:33 +00:00
|
|
|
* prog = array->ptrs[index];
|
2015-05-19 23:59:04 +00:00
|
|
|
* if (prog == NULL)
|
|
|
|
* goto out;
|
|
|
|
* goto *(prog->bpf_func + prologue_size);
|
|
|
|
* out:
|
|
|
|
*/
|
bpf: Implement BPF exceptions
This patch implements BPF exceptions, and introduces a bpf_throw kfunc
to allow programs to throw exceptions during their execution at runtime.
A bpf_throw invocation is treated as an immediate termination of the
program, returning back to its caller within the kernel, unwinding all
stack frames.
This allows the program to simplify its implementation, by testing for
runtime conditions which the verifier has no visibility into, and assert
that they are true. In case they are not, the program can simply throw
an exception from the other branch.
BPF exceptions are explicitly *NOT* an unlikely slowpath error handling
primitive, and this objective has guided design choices of the
implementation of the them within the kernel (with the bulk of the cost
for unwinding the stack offloaded to the bpf_throw kfunc).
The implementation of this mechanism requires use of add_hidden_subprog
mechanism introduced in the previous patch, which generates a couple of
instructions to move R1 to R0 and exit. The JIT then rewrites the
prologue of this subprog to take the stack pointer and frame pointer as
inputs and reset the stack frame, popping all callee-saved registers
saved by the main subprog. The bpf_throw function then walks the stack
at runtime, and invokes this exception subprog with the stack and frame
pointers as parameters.
Reviewers must take note that currently the main program is made to save
all callee-saved registers on x86_64 during entry into the program. This
is because we must do an equivalent of a lightweight context switch when
unwinding the stack, therefore we need the callee-saved registers of the
caller of the BPF program to be able to return with a sane state.
Note that we have to additionally handle r12, even though it is not used
by the program, because when throwing the exception the program makes an
entry into the kernel which could clobber r12 after saving it on the
stack. To be able to preserve the value we received on program entry, we
push r12 and restore it from the generated subprogram when unwinding the
stack.
For now, bpf_throw invocation fails when lingering resources or locks
exist in that path of the program. In a future followup, bpf_throw will
be extended to perform frame-by-frame unwinding to release lingering
resources for each stack frame, removing this limitation.
Signed-off-by: Kumar Kartikeya Dwivedi <memxor@gmail.com>
Link: https://lore.kernel.org/r/20230912233214.1518551-5-memxor@gmail.com
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
2023-09-12 23:32:01 +00:00
|
|
|
static void emit_bpf_tail_call_indirect(struct bpf_prog *bpf_prog,
|
|
|
|
u8 **pprog, bool *callee_regs_used,
|
2021-10-26 12:01:47 +00:00
|
|
|
u32 stack_depth, u8 *ip,
|
|
|
|
struct jit_context *ctx)
|
2015-05-19 23:59:04 +00:00
|
|
|
{
|
bpf, x64: rework pro/epilogue and tailcall handling in JIT
This commit serves two things:
1) it optimizes BPF prologue/epilogue generation
2) it makes possible to have tailcalls within BPF subprogram
Both points are related to each other since without 1), 2) could not be
achieved.
In [1], Alexei says:
"The prologue will look like:
nop5
xor eax,eax // two new bytes if bpf_tail_call() is used in this
// function
push rbp
mov rbp, rsp
sub rsp, rounded_stack_depth
push rax // zero init tail_call counter
variable number of push rbx,r13,r14,r15
Then bpf_tail_call will pop variable number rbx,..
and final 'pop rax'
Then 'add rsp, size_of_current_stack_frame'
jmp to next function and skip over 'nop5; xor eax,eax; push rpb; mov
rbp, rsp'
This way new function will set its own stack size and will init tail
call
counter with whatever value the parent had.
If next function doesn't use bpf_tail_call it won't have 'xor eax,eax'.
Instead it would need to have 'nop2' in there."
Implement that suggestion.
Since the layout of stack is changed, tail call counter handling can not
rely anymore on popping it to rbx just like it have been handled for
constant prologue case and later overwrite of rbx with actual value of
rbx pushed to stack. Therefore, let's use one of the register (%rcx) that
is considered to be volatile/caller-saved and pop the value of tail call
counter in there in the epilogue.
Drop the BUILD_BUG_ON in emit_prologue and in
emit_bpf_tail_call_indirect where instruction layout is not constant
anymore.
Introduce new poke target, 'tailcall_bypass' to poke descriptor that is
dedicated for skipping the register pops and stack unwind that are
generated right before the actual jump to target program.
For case when the target program is not present, BPF program will skip
the pop instructions and nop5 dedicated for jmpq $target. An example of
such state when only R6 of callee saved registers is used by program:
ffffffffc0513aa1: e9 0e 00 00 00 jmpq 0xffffffffc0513ab4
ffffffffc0513aa6: 5b pop %rbx
ffffffffc0513aa7: 58 pop %rax
ffffffffc0513aa8: 48 81 c4 00 00 00 00 add $0x0,%rsp
ffffffffc0513aaf: 0f 1f 44 00 00 nopl 0x0(%rax,%rax,1)
ffffffffc0513ab4: 48 89 df mov %rbx,%rdi
When target program is inserted, the jump that was there to skip
pops/nop5 will become the nop5, so CPU will go over pops and do the
actual tailcall.
One might ask why there simply can not be pushes after the nop5?
In the following example snippet:
ffffffffc037030c: 48 89 fb mov %rdi,%rbx
(...)
ffffffffc0370332: 5b pop %rbx
ffffffffc0370333: 58 pop %rax
ffffffffc0370334: 48 81 c4 00 00 00 00 add $0x0,%rsp
ffffffffc037033b: 0f 1f 44 00 00 nopl 0x0(%rax,%rax,1)
ffffffffc0370340: 48 81 ec 00 00 00 00 sub $0x0,%rsp
ffffffffc0370347: 50 push %rax
ffffffffc0370348: 53 push %rbx
ffffffffc0370349: 48 89 df mov %rbx,%rdi
ffffffffc037034c: e8 f7 21 00 00 callq 0xffffffffc0372548
There is the bpf2bpf call (at ffffffffc037034c) right after the tailcall
and jump target is not present. ctx is in %rbx register and BPF
subprogram that we will call into on ffffffffc037034c is relying on it,
e.g. it will pick ctx from there. Such code layout is therefore broken
as we would overwrite the content of %rbx with the value that was pushed
on the prologue. That is the reason for the 'bypass' approach.
Special care needs to be taken during the install/update/remove of
tailcall target. In case when target program is not present, the CPU
must not execute the pop instructions that precede the tailcall.
To address that, the following states can be defined:
A nop, unwind, nop
B nop, unwind, tail
C skip, unwind, nop
D skip, unwind, tail
A is forbidden (lead to incorrectness). The state transitions between
tailcall install/update/remove will work as follows:
First install tail call f: C->D->B(f)
* poke the tailcall, after that get rid of the skip
Update tail call f to f': B(f)->B(f')
* poke the tailcall (poke->tailcall_target) and do NOT touch the
poke->tailcall_bypass
Remove tail call: B(f')->C(f')
* poke->tailcall_bypass is poked back to jump, then we wait the RCU
grace period so that other programs will finish its execution and
after that we are safe to remove the poke->tailcall_target
Install new tail call (f''): C(f')->D(f'')->B(f'').
* same as first step
This way CPU can never be exposed to "unwind, tail" state.
Last but not least, when tailcalls get mixed with bpf2bpf calls, it
would be possible to encounter the endless loop due to clearing the
tailcall counter if for example we would use the tailcall3-like from BPF
selftests program that would be subprogram-based, meaning the tailcall
would be present within the BPF subprogram.
This test, broken down to particular steps, would do:
entry -> set tailcall counter to 0, bump it by 1, tailcall to func0
func0 -> call subprog_tail
(we are NOT skipping the first 11 bytes of prologue and this subprogram
has a tailcall, therefore we clear the counter...)
subprog -> do the same thing as entry
and then loop forever.
To address this, the idea is to go through the call chain of bpf2bpf progs
and look for a tailcall presence throughout whole chain. If we saw a single
tail call then each node in this call chain needs to be marked as a subprog
that can reach the tailcall. We would later feed the JIT with this info
and:
- set eax to 0 only when tailcall is reachable and this is the entry prog
- if tailcall is reachable but there's no tailcall in insns of currently
JITed prog then push rax anyway, so that it will be possible to
propagate further down the call chain
- finally if tailcall is reachable, then we need to precede the 'call'
insn with mov rax, [rbp - (stack_depth + 8)]
Tail call related cases from test_verifier kselftest are also working
fine. Sample BPF programs that utilize tail calls (sockex3, tracex5)
work properly as well.
[1]: https://lore.kernel.org/bpf/20200517043227.2gpq22ifoq37ogst@ast-mbp.dhcp.thefacebook.com/
Suggested-by: Alexei Starovoitov <ast@kernel.org>
Signed-off-by: Maciej Fijalkowski <maciej.fijalkowski@intel.com>
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
2020-09-16 21:10:08 +00:00
|
|
|
int tcc_off = -4 - round_up(stack_depth, 8);
|
2021-10-26 12:01:47 +00:00
|
|
|
u8 *prog = *pprog, *start = *pprog;
|
|
|
|
int offset;
|
2020-09-29 20:46:53 +00:00
|
|
|
|
x86/bpf: Clean up non-standard comments, to make the code more readable
So by chance I looked into x86 assembly in arch/x86/net/bpf_jit_comp.c and
noticed the weird and inconsistent comment style it mistakenly learned from
the networking code:
/* Multi-line comment ...
* ... looks like this.
*/
Fix this to use the standard comment style specified in Documentation/CodingStyle
and used in arch/x86/ as well:
/*
* Multi-line comment ...
* ... looks like this.
*/
Also, to quote Linus's ... more explicit views about this:
http://article.gmane.org/gmane.linux.kernel.cryptoapi/21066
> But no, the networking code picked *none* of the above sane formats.
> Instead, it picked these two models that are just half-arsed
> shit-for-brains:
>
> (no)
> /* This is disgusting drug-induced
> * crap, and should die
> */
>
> (no-no-no)
> /* This is also very nasty
> * and visually unbalanced */
>
> Please. The networking code actually has the *worst* possible comment
> style. You can literally find that (no-no-no) style, which is just
> really horribly disgusting and worse than the otherwise fairly similar
> (d) in pretty much every way.
Also improve the comments and some other details while at it:
- Don't mix same-line and previous-line comment style on otherwise
identical code patterns within the same function,
- capitalize 'BPF' and x86 register names consistently,
- capitalize sentences consistently,
- instead of 'x64' use 'x86-64': x64 is a Microsoft specific term,
- use more consistent punctuation,
- use standard coding style in macros as well,
- fix typos and a few other minor details.
Consistent coding style is not optional, at least in arch/x86/.
No change in functionality.
( In case this commit causes conflicts with pending development code
I'll be glad to help resolve any conflicts! )
Acked-by: Thomas Gleixner <tglx@linutronix.de>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Borislav Petkov <bp@alien8.de>
Cc: H. Peter Anvin <hpa@zytor.com>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: David S. Miller <davem@davemloft.net>
Cc: Eric Dumazet <edumazet@google.com>
Cc: Daniel Borkmann <daniel@iogearbox.net>
Cc: Alexei Starovoitov <ast@fb.com>
Cc: Hideaki YOSHIFUJI <yoshfuji@linux-ipv6.org>
Cc: netdev@vger.kernel.org
Cc: linux-kernel@vger.kernel.org
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
2018-04-27 09:54:40 +00:00
|
|
|
/*
|
|
|
|
* rdi - pointer to ctx
|
2015-05-19 23:59:04 +00:00
|
|
|
* rsi - pointer to bpf_array
|
|
|
|
* rdx - index in bpf_array
|
|
|
|
*/
|
|
|
|
|
x86/bpf: Clean up non-standard comments, to make the code more readable
So by chance I looked into x86 assembly in arch/x86/net/bpf_jit_comp.c and
noticed the weird and inconsistent comment style it mistakenly learned from
the networking code:
/* Multi-line comment ...
* ... looks like this.
*/
Fix this to use the standard comment style specified in Documentation/CodingStyle
and used in arch/x86/ as well:
/*
* Multi-line comment ...
* ... looks like this.
*/
Also, to quote Linus's ... more explicit views about this:
http://article.gmane.org/gmane.linux.kernel.cryptoapi/21066
> But no, the networking code picked *none* of the above sane formats.
> Instead, it picked these two models that are just half-arsed
> shit-for-brains:
>
> (no)
> /* This is disgusting drug-induced
> * crap, and should die
> */
>
> (no-no-no)
> /* This is also very nasty
> * and visually unbalanced */
>
> Please. The networking code actually has the *worst* possible comment
> style. You can literally find that (no-no-no) style, which is just
> really horribly disgusting and worse than the otherwise fairly similar
> (d) in pretty much every way.
Also improve the comments and some other details while at it:
- Don't mix same-line and previous-line comment style on otherwise
identical code patterns within the same function,
- capitalize 'BPF' and x86 register names consistently,
- capitalize sentences consistently,
- instead of 'x64' use 'x86-64': x64 is a Microsoft specific term,
- use more consistent punctuation,
- use standard coding style in macros as well,
- fix typos and a few other minor details.
Consistent coding style is not optional, at least in arch/x86/.
No change in functionality.
( In case this commit causes conflicts with pending development code
I'll be glad to help resolve any conflicts! )
Acked-by: Thomas Gleixner <tglx@linutronix.de>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Borislav Petkov <bp@alien8.de>
Cc: H. Peter Anvin <hpa@zytor.com>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: David S. Miller <davem@davemloft.net>
Cc: Eric Dumazet <edumazet@google.com>
Cc: Daniel Borkmann <daniel@iogearbox.net>
Cc: Alexei Starovoitov <ast@fb.com>
Cc: Hideaki YOSHIFUJI <yoshfuji@linux-ipv6.org>
Cc: netdev@vger.kernel.org
Cc: linux-kernel@vger.kernel.org
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
2018-04-27 09:54:40 +00:00
|
|
|
/*
|
|
|
|
* if (index >= array->map.max_entries)
|
|
|
|
* goto out;
|
2015-05-19 23:59:04 +00:00
|
|
|
*/
|
2017-10-03 22:37:20 +00:00
|
|
|
EMIT2(0x89, 0xD2); /* mov edx, edx */
|
|
|
|
EMIT3(0x39, 0x56, /* cmp dword ptr [rsi + 16], edx */
|
2015-05-19 23:59:04 +00:00
|
|
|
offsetof(struct bpf_array, map.max_entries));
|
2021-10-26 12:01:47 +00:00
|
|
|
|
|
|
|
offset = ctx->tail_call_indirect_label - (prog + 2 - start);
|
|
|
|
EMIT2(X86_JBE, offset); /* jbe out */
|
2015-05-19 23:59:04 +00:00
|
|
|
|
x86/bpf: Clean up non-standard comments, to make the code more readable
So by chance I looked into x86 assembly in arch/x86/net/bpf_jit_comp.c and
noticed the weird and inconsistent comment style it mistakenly learned from
the networking code:
/* Multi-line comment ...
* ... looks like this.
*/
Fix this to use the standard comment style specified in Documentation/CodingStyle
and used in arch/x86/ as well:
/*
* Multi-line comment ...
* ... looks like this.
*/
Also, to quote Linus's ... more explicit views about this:
http://article.gmane.org/gmane.linux.kernel.cryptoapi/21066
> But no, the networking code picked *none* of the above sane formats.
> Instead, it picked these two models that are just half-arsed
> shit-for-brains:
>
> (no)
> /* This is disgusting drug-induced
> * crap, and should die
> */
>
> (no-no-no)
> /* This is also very nasty
> * and visually unbalanced */
>
> Please. The networking code actually has the *worst* possible comment
> style. You can literally find that (no-no-no) style, which is just
> really horribly disgusting and worse than the otherwise fairly similar
> (d) in pretty much every way.
Also improve the comments and some other details while at it:
- Don't mix same-line and previous-line comment style on otherwise
identical code patterns within the same function,
- capitalize 'BPF' and x86 register names consistently,
- capitalize sentences consistently,
- instead of 'x64' use 'x86-64': x64 is a Microsoft specific term,
- use more consistent punctuation,
- use standard coding style in macros as well,
- fix typos and a few other minor details.
Consistent coding style is not optional, at least in arch/x86/.
No change in functionality.
( In case this commit causes conflicts with pending development code
I'll be glad to help resolve any conflicts! )
Acked-by: Thomas Gleixner <tglx@linutronix.de>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Borislav Petkov <bp@alien8.de>
Cc: H. Peter Anvin <hpa@zytor.com>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: David S. Miller <davem@davemloft.net>
Cc: Eric Dumazet <edumazet@google.com>
Cc: Daniel Borkmann <daniel@iogearbox.net>
Cc: Alexei Starovoitov <ast@fb.com>
Cc: Hideaki YOSHIFUJI <yoshfuji@linux-ipv6.org>
Cc: netdev@vger.kernel.org
Cc: linux-kernel@vger.kernel.org
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
2018-04-27 09:54:40 +00:00
|
|
|
/*
|
bpf: Change value of MAX_TAIL_CALL_CNT from 32 to 33
In the current code, the actual max tail call count is 33 which is greater
than MAX_TAIL_CALL_CNT (defined as 32). The actual limit is not consistent
with the meaning of MAX_TAIL_CALL_CNT and thus confusing at first glance.
We can see the historical evolution from commit 04fd61ab36ec ("bpf: allow
bpf programs to tail-call other bpf programs") and commit f9dabe016b63
("bpf: Undo off-by-one in interpreter tail call count limit"). In order
to avoid changing existing behavior, the actual limit is 33 now, this is
reasonable.
After commit 874be05f525e ("bpf, tests: Add tail call test suite"), we can
see there exists failed testcase.
On all archs when CONFIG_BPF_JIT_ALWAYS_ON is not set:
# echo 0 > /proc/sys/net/core/bpf_jit_enable
# modprobe test_bpf
# dmesg | grep -w FAIL
Tail call error path, max count reached jited:0 ret 34 != 33 FAIL
On some archs:
# echo 1 > /proc/sys/net/core/bpf_jit_enable
# modprobe test_bpf
# dmesg | grep -w FAIL
Tail call error path, max count reached jited:1 ret 34 != 33 FAIL
Although the above failed testcase has been fixed in commit 18935a72eb25
("bpf/tests: Fix error in tail call limit tests"), it would still be good
to change the value of MAX_TAIL_CALL_CNT from 32 to 33 to make the code
more readable.
The 32-bit x86 JIT was using a limit of 32, just fix the wrong comments and
limit to 33 tail calls as the constant MAX_TAIL_CALL_CNT updated. For the
mips64 JIT, use "ori" instead of "addiu" as suggested by Johan Almbladh.
For the riscv JIT, use RV_REG_TCC directly to save one register move as
suggested by Björn Töpel. For the other implementations, no function changes,
it does not change the current limit 33, the new value of MAX_TAIL_CALL_CNT
can reflect the actual max tail call count, the related tail call testcases
in test_bpf module and selftests can work well for the interpreter and the
JIT.
Here are the test results on x86_64:
# uname -m
x86_64
# echo 0 > /proc/sys/net/core/bpf_jit_enable
# modprobe test_bpf test_suite=test_tail_calls
# dmesg | tail -1
test_bpf: test_tail_calls: Summary: 8 PASSED, 0 FAILED, [0/8 JIT'ed]
# rmmod test_bpf
# echo 1 > /proc/sys/net/core/bpf_jit_enable
# modprobe test_bpf test_suite=test_tail_calls
# dmesg | tail -1
test_bpf: test_tail_calls: Summary: 8 PASSED, 0 FAILED, [8/8 JIT'ed]
# rmmod test_bpf
# ./test_progs -t tailcalls
#142 tailcalls:OK
Summary: 1/11 PASSED, 0 SKIPPED, 0 FAILED
Signed-off-by: Tiezhu Yang <yangtiezhu@loongson.cn>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Tested-by: Johan Almbladh <johan.almbladh@anyfinetworks.com>
Tested-by: Ilya Leoshkevich <iii@linux.ibm.com>
Acked-by: Björn Töpel <bjorn@kernel.org>
Acked-by: Johan Almbladh <johan.almbladh@anyfinetworks.com>
Acked-by: Ilya Leoshkevich <iii@linux.ibm.com>
Link: https://lore.kernel.org/bpf/1636075800-3264-1-git-send-email-yangtiezhu@loongson.cn
2021-11-05 01:30:00 +00:00
|
|
|
* if (tail_call_cnt++ >= MAX_TAIL_CALL_CNT)
|
x86/bpf: Clean up non-standard comments, to make the code more readable
So by chance I looked into x86 assembly in arch/x86/net/bpf_jit_comp.c and
noticed the weird and inconsistent comment style it mistakenly learned from
the networking code:
/* Multi-line comment ...
* ... looks like this.
*/
Fix this to use the standard comment style specified in Documentation/CodingStyle
and used in arch/x86/ as well:
/*
* Multi-line comment ...
* ... looks like this.
*/
Also, to quote Linus's ... more explicit views about this:
http://article.gmane.org/gmane.linux.kernel.cryptoapi/21066
> But no, the networking code picked *none* of the above sane formats.
> Instead, it picked these two models that are just half-arsed
> shit-for-brains:
>
> (no)
> /* This is disgusting drug-induced
> * crap, and should die
> */
>
> (no-no-no)
> /* This is also very nasty
> * and visually unbalanced */
>
> Please. The networking code actually has the *worst* possible comment
> style. You can literally find that (no-no-no) style, which is just
> really horribly disgusting and worse than the otherwise fairly similar
> (d) in pretty much every way.
Also improve the comments and some other details while at it:
- Don't mix same-line and previous-line comment style on otherwise
identical code patterns within the same function,
- capitalize 'BPF' and x86 register names consistently,
- capitalize sentences consistently,
- instead of 'x64' use 'x86-64': x64 is a Microsoft specific term,
- use more consistent punctuation,
- use standard coding style in macros as well,
- fix typos and a few other minor details.
Consistent coding style is not optional, at least in arch/x86/.
No change in functionality.
( In case this commit causes conflicts with pending development code
I'll be glad to help resolve any conflicts! )
Acked-by: Thomas Gleixner <tglx@linutronix.de>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Borislav Petkov <bp@alien8.de>
Cc: H. Peter Anvin <hpa@zytor.com>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: David S. Miller <davem@davemloft.net>
Cc: Eric Dumazet <edumazet@google.com>
Cc: Daniel Borkmann <daniel@iogearbox.net>
Cc: Alexei Starovoitov <ast@fb.com>
Cc: Hideaki YOSHIFUJI <yoshfuji@linux-ipv6.org>
Cc: netdev@vger.kernel.org
Cc: linux-kernel@vger.kernel.org
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
2018-04-27 09:54:40 +00:00
|
|
|
* goto out;
|
2015-05-19 23:59:04 +00:00
|
|
|
*/
|
bpf, x64: rework pro/epilogue and tailcall handling in JIT
This commit serves two things:
1) it optimizes BPF prologue/epilogue generation
2) it makes possible to have tailcalls within BPF subprogram
Both points are related to each other since without 1), 2) could not be
achieved.
In [1], Alexei says:
"The prologue will look like:
nop5
xor eax,eax // two new bytes if bpf_tail_call() is used in this
// function
push rbp
mov rbp, rsp
sub rsp, rounded_stack_depth
push rax // zero init tail_call counter
variable number of push rbx,r13,r14,r15
Then bpf_tail_call will pop variable number rbx,..
and final 'pop rax'
Then 'add rsp, size_of_current_stack_frame'
jmp to next function and skip over 'nop5; xor eax,eax; push rpb; mov
rbp, rsp'
This way new function will set its own stack size and will init tail
call
counter with whatever value the parent had.
If next function doesn't use bpf_tail_call it won't have 'xor eax,eax'.
Instead it would need to have 'nop2' in there."
Implement that suggestion.
Since the layout of stack is changed, tail call counter handling can not
rely anymore on popping it to rbx just like it have been handled for
constant prologue case and later overwrite of rbx with actual value of
rbx pushed to stack. Therefore, let's use one of the register (%rcx) that
is considered to be volatile/caller-saved and pop the value of tail call
counter in there in the epilogue.
Drop the BUILD_BUG_ON in emit_prologue and in
emit_bpf_tail_call_indirect where instruction layout is not constant
anymore.
Introduce new poke target, 'tailcall_bypass' to poke descriptor that is
dedicated for skipping the register pops and stack unwind that are
generated right before the actual jump to target program.
For case when the target program is not present, BPF program will skip
the pop instructions and nop5 dedicated for jmpq $target. An example of
such state when only R6 of callee saved registers is used by program:
ffffffffc0513aa1: e9 0e 00 00 00 jmpq 0xffffffffc0513ab4
ffffffffc0513aa6: 5b pop %rbx
ffffffffc0513aa7: 58 pop %rax
ffffffffc0513aa8: 48 81 c4 00 00 00 00 add $0x0,%rsp
ffffffffc0513aaf: 0f 1f 44 00 00 nopl 0x0(%rax,%rax,1)
ffffffffc0513ab4: 48 89 df mov %rbx,%rdi
When target program is inserted, the jump that was there to skip
pops/nop5 will become the nop5, so CPU will go over pops and do the
actual tailcall.
One might ask why there simply can not be pushes after the nop5?
In the following example snippet:
ffffffffc037030c: 48 89 fb mov %rdi,%rbx
(...)
ffffffffc0370332: 5b pop %rbx
ffffffffc0370333: 58 pop %rax
ffffffffc0370334: 48 81 c4 00 00 00 00 add $0x0,%rsp
ffffffffc037033b: 0f 1f 44 00 00 nopl 0x0(%rax,%rax,1)
ffffffffc0370340: 48 81 ec 00 00 00 00 sub $0x0,%rsp
ffffffffc0370347: 50 push %rax
ffffffffc0370348: 53 push %rbx
ffffffffc0370349: 48 89 df mov %rbx,%rdi
ffffffffc037034c: e8 f7 21 00 00 callq 0xffffffffc0372548
There is the bpf2bpf call (at ffffffffc037034c) right after the tailcall
and jump target is not present. ctx is in %rbx register and BPF
subprogram that we will call into on ffffffffc037034c is relying on it,
e.g. it will pick ctx from there. Such code layout is therefore broken
as we would overwrite the content of %rbx with the value that was pushed
on the prologue. That is the reason for the 'bypass' approach.
Special care needs to be taken during the install/update/remove of
tailcall target. In case when target program is not present, the CPU
must not execute the pop instructions that precede the tailcall.
To address that, the following states can be defined:
A nop, unwind, nop
B nop, unwind, tail
C skip, unwind, nop
D skip, unwind, tail
A is forbidden (lead to incorrectness). The state transitions between
tailcall install/update/remove will work as follows:
First install tail call f: C->D->B(f)
* poke the tailcall, after that get rid of the skip
Update tail call f to f': B(f)->B(f')
* poke the tailcall (poke->tailcall_target) and do NOT touch the
poke->tailcall_bypass
Remove tail call: B(f')->C(f')
* poke->tailcall_bypass is poked back to jump, then we wait the RCU
grace period so that other programs will finish its execution and
after that we are safe to remove the poke->tailcall_target
Install new tail call (f''): C(f')->D(f'')->B(f'').
* same as first step
This way CPU can never be exposed to "unwind, tail" state.
Last but not least, when tailcalls get mixed with bpf2bpf calls, it
would be possible to encounter the endless loop due to clearing the
tailcall counter if for example we would use the tailcall3-like from BPF
selftests program that would be subprogram-based, meaning the tailcall
would be present within the BPF subprogram.
This test, broken down to particular steps, would do:
entry -> set tailcall counter to 0, bump it by 1, tailcall to func0
func0 -> call subprog_tail
(we are NOT skipping the first 11 bytes of prologue and this subprogram
has a tailcall, therefore we clear the counter...)
subprog -> do the same thing as entry
and then loop forever.
To address this, the idea is to go through the call chain of bpf2bpf progs
and look for a tailcall presence throughout whole chain. If we saw a single
tail call then each node in this call chain needs to be marked as a subprog
that can reach the tailcall. We would later feed the JIT with this info
and:
- set eax to 0 only when tailcall is reachable and this is the entry prog
- if tailcall is reachable but there's no tailcall in insns of currently
JITed prog then push rax anyway, so that it will be possible to
propagate further down the call chain
- finally if tailcall is reachable, then we need to precede the 'call'
insn with mov rax, [rbp - (stack_depth + 8)]
Tail call related cases from test_verifier kselftest are also working
fine. Sample BPF programs that utilize tail calls (sockex3, tracex5)
work properly as well.
[1]: https://lore.kernel.org/bpf/20200517043227.2gpq22ifoq37ogst@ast-mbp.dhcp.thefacebook.com/
Suggested-by: Alexei Starovoitov <ast@kernel.org>
Signed-off-by: Maciej Fijalkowski <maciej.fijalkowski@intel.com>
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
2020-09-16 21:10:08 +00:00
|
|
|
EMIT2_off32(0x8B, 0x85, tcc_off); /* mov eax, dword ptr [rbp - tcc_off] */
|
2015-05-19 23:59:04 +00:00
|
|
|
EMIT3(0x83, 0xF8, MAX_TAIL_CALL_CNT); /* cmp eax, MAX_TAIL_CALL_CNT */
|
2021-10-26 12:01:47 +00:00
|
|
|
|
|
|
|
offset = ctx->tail_call_indirect_label - (prog + 2 - start);
|
bpf: Change value of MAX_TAIL_CALL_CNT from 32 to 33
In the current code, the actual max tail call count is 33 which is greater
than MAX_TAIL_CALL_CNT (defined as 32). The actual limit is not consistent
with the meaning of MAX_TAIL_CALL_CNT and thus confusing at first glance.
We can see the historical evolution from commit 04fd61ab36ec ("bpf: allow
bpf programs to tail-call other bpf programs") and commit f9dabe016b63
("bpf: Undo off-by-one in interpreter tail call count limit"). In order
to avoid changing existing behavior, the actual limit is 33 now, this is
reasonable.
After commit 874be05f525e ("bpf, tests: Add tail call test suite"), we can
see there exists failed testcase.
On all archs when CONFIG_BPF_JIT_ALWAYS_ON is not set:
# echo 0 > /proc/sys/net/core/bpf_jit_enable
# modprobe test_bpf
# dmesg | grep -w FAIL
Tail call error path, max count reached jited:0 ret 34 != 33 FAIL
On some archs:
# echo 1 > /proc/sys/net/core/bpf_jit_enable
# modprobe test_bpf
# dmesg | grep -w FAIL
Tail call error path, max count reached jited:1 ret 34 != 33 FAIL
Although the above failed testcase has been fixed in commit 18935a72eb25
("bpf/tests: Fix error in tail call limit tests"), it would still be good
to change the value of MAX_TAIL_CALL_CNT from 32 to 33 to make the code
more readable.
The 32-bit x86 JIT was using a limit of 32, just fix the wrong comments and
limit to 33 tail calls as the constant MAX_TAIL_CALL_CNT updated. For the
mips64 JIT, use "ori" instead of "addiu" as suggested by Johan Almbladh.
For the riscv JIT, use RV_REG_TCC directly to save one register move as
suggested by Björn Töpel. For the other implementations, no function changes,
it does not change the current limit 33, the new value of MAX_TAIL_CALL_CNT
can reflect the actual max tail call count, the related tail call testcases
in test_bpf module and selftests can work well for the interpreter and the
JIT.
Here are the test results on x86_64:
# uname -m
x86_64
# echo 0 > /proc/sys/net/core/bpf_jit_enable
# modprobe test_bpf test_suite=test_tail_calls
# dmesg | tail -1
test_bpf: test_tail_calls: Summary: 8 PASSED, 0 FAILED, [0/8 JIT'ed]
# rmmod test_bpf
# echo 1 > /proc/sys/net/core/bpf_jit_enable
# modprobe test_bpf test_suite=test_tail_calls
# dmesg | tail -1
test_bpf: test_tail_calls: Summary: 8 PASSED, 0 FAILED, [8/8 JIT'ed]
# rmmod test_bpf
# ./test_progs -t tailcalls
#142 tailcalls:OK
Summary: 1/11 PASSED, 0 SKIPPED, 0 FAILED
Signed-off-by: Tiezhu Yang <yangtiezhu@loongson.cn>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Tested-by: Johan Almbladh <johan.almbladh@anyfinetworks.com>
Tested-by: Ilya Leoshkevich <iii@linux.ibm.com>
Acked-by: Björn Töpel <bjorn@kernel.org>
Acked-by: Johan Almbladh <johan.almbladh@anyfinetworks.com>
Acked-by: Ilya Leoshkevich <iii@linux.ibm.com>
Link: https://lore.kernel.org/bpf/1636075800-3264-1-git-send-email-yangtiezhu@loongson.cn
2021-11-05 01:30:00 +00:00
|
|
|
EMIT2(X86_JAE, offset); /* jae out */
|
2015-05-19 23:59:04 +00:00
|
|
|
EMIT3(0x83, 0xC0, 0x01); /* add eax, 1 */
|
bpf, x64: rework pro/epilogue and tailcall handling in JIT
This commit serves two things:
1) it optimizes BPF prologue/epilogue generation
2) it makes possible to have tailcalls within BPF subprogram
Both points are related to each other since without 1), 2) could not be
achieved.
In [1], Alexei says:
"The prologue will look like:
nop5
xor eax,eax // two new bytes if bpf_tail_call() is used in this
// function
push rbp
mov rbp, rsp
sub rsp, rounded_stack_depth
push rax // zero init tail_call counter
variable number of push rbx,r13,r14,r15
Then bpf_tail_call will pop variable number rbx,..
and final 'pop rax'
Then 'add rsp, size_of_current_stack_frame'
jmp to next function and skip over 'nop5; xor eax,eax; push rpb; mov
rbp, rsp'
This way new function will set its own stack size and will init tail
call
counter with whatever value the parent had.
If next function doesn't use bpf_tail_call it won't have 'xor eax,eax'.
Instead it would need to have 'nop2' in there."
Implement that suggestion.
Since the layout of stack is changed, tail call counter handling can not
rely anymore on popping it to rbx just like it have been handled for
constant prologue case and later overwrite of rbx with actual value of
rbx pushed to stack. Therefore, let's use one of the register (%rcx) that
is considered to be volatile/caller-saved and pop the value of tail call
counter in there in the epilogue.
Drop the BUILD_BUG_ON in emit_prologue and in
emit_bpf_tail_call_indirect where instruction layout is not constant
anymore.
Introduce new poke target, 'tailcall_bypass' to poke descriptor that is
dedicated for skipping the register pops and stack unwind that are
generated right before the actual jump to target program.
For case when the target program is not present, BPF program will skip
the pop instructions and nop5 dedicated for jmpq $target. An example of
such state when only R6 of callee saved registers is used by program:
ffffffffc0513aa1: e9 0e 00 00 00 jmpq 0xffffffffc0513ab4
ffffffffc0513aa6: 5b pop %rbx
ffffffffc0513aa7: 58 pop %rax
ffffffffc0513aa8: 48 81 c4 00 00 00 00 add $0x0,%rsp
ffffffffc0513aaf: 0f 1f 44 00 00 nopl 0x0(%rax,%rax,1)
ffffffffc0513ab4: 48 89 df mov %rbx,%rdi
When target program is inserted, the jump that was there to skip
pops/nop5 will become the nop5, so CPU will go over pops and do the
actual tailcall.
One might ask why there simply can not be pushes after the nop5?
In the following example snippet:
ffffffffc037030c: 48 89 fb mov %rdi,%rbx
(...)
ffffffffc0370332: 5b pop %rbx
ffffffffc0370333: 58 pop %rax
ffffffffc0370334: 48 81 c4 00 00 00 00 add $0x0,%rsp
ffffffffc037033b: 0f 1f 44 00 00 nopl 0x0(%rax,%rax,1)
ffffffffc0370340: 48 81 ec 00 00 00 00 sub $0x0,%rsp
ffffffffc0370347: 50 push %rax
ffffffffc0370348: 53 push %rbx
ffffffffc0370349: 48 89 df mov %rbx,%rdi
ffffffffc037034c: e8 f7 21 00 00 callq 0xffffffffc0372548
There is the bpf2bpf call (at ffffffffc037034c) right after the tailcall
and jump target is not present. ctx is in %rbx register and BPF
subprogram that we will call into on ffffffffc037034c is relying on it,
e.g. it will pick ctx from there. Such code layout is therefore broken
as we would overwrite the content of %rbx with the value that was pushed
on the prologue. That is the reason for the 'bypass' approach.
Special care needs to be taken during the install/update/remove of
tailcall target. In case when target program is not present, the CPU
must not execute the pop instructions that precede the tailcall.
To address that, the following states can be defined:
A nop, unwind, nop
B nop, unwind, tail
C skip, unwind, nop
D skip, unwind, tail
A is forbidden (lead to incorrectness). The state transitions between
tailcall install/update/remove will work as follows:
First install tail call f: C->D->B(f)
* poke the tailcall, after that get rid of the skip
Update tail call f to f': B(f)->B(f')
* poke the tailcall (poke->tailcall_target) and do NOT touch the
poke->tailcall_bypass
Remove tail call: B(f')->C(f')
* poke->tailcall_bypass is poked back to jump, then we wait the RCU
grace period so that other programs will finish its execution and
after that we are safe to remove the poke->tailcall_target
Install new tail call (f''): C(f')->D(f'')->B(f'').
* same as first step
This way CPU can never be exposed to "unwind, tail" state.
Last but not least, when tailcalls get mixed with bpf2bpf calls, it
would be possible to encounter the endless loop due to clearing the
tailcall counter if for example we would use the tailcall3-like from BPF
selftests program that would be subprogram-based, meaning the tailcall
would be present within the BPF subprogram.
This test, broken down to particular steps, would do:
entry -> set tailcall counter to 0, bump it by 1, tailcall to func0
func0 -> call subprog_tail
(we are NOT skipping the first 11 bytes of prologue and this subprogram
has a tailcall, therefore we clear the counter...)
subprog -> do the same thing as entry
and then loop forever.
To address this, the idea is to go through the call chain of bpf2bpf progs
and look for a tailcall presence throughout whole chain. If we saw a single
tail call then each node in this call chain needs to be marked as a subprog
that can reach the tailcall. We would later feed the JIT with this info
and:
- set eax to 0 only when tailcall is reachable and this is the entry prog
- if tailcall is reachable but there's no tailcall in insns of currently
JITed prog then push rax anyway, so that it will be possible to
propagate further down the call chain
- finally if tailcall is reachable, then we need to precede the 'call'
insn with mov rax, [rbp - (stack_depth + 8)]
Tail call related cases from test_verifier kselftest are also working
fine. Sample BPF programs that utilize tail calls (sockex3, tracex5)
work properly as well.
[1]: https://lore.kernel.org/bpf/20200517043227.2gpq22ifoq37ogst@ast-mbp.dhcp.thefacebook.com/
Suggested-by: Alexei Starovoitov <ast@kernel.org>
Signed-off-by: Maciej Fijalkowski <maciej.fijalkowski@intel.com>
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
2020-09-16 21:10:08 +00:00
|
|
|
EMIT2_off32(0x89, 0x85, tcc_off); /* mov dword ptr [rbp - tcc_off], eax */
|
2015-05-19 23:59:04 +00:00
|
|
|
|
2015-08-06 07:02:33 +00:00
|
|
|
/* prog = array->ptrs[index]; */
|
2020-09-16 21:10:04 +00:00
|
|
|
EMIT4_off32(0x48, 0x8B, 0x8C, 0xD6, /* mov rcx, [rsi + rdx * 8 + offsetof(...)] */
|
2015-08-06 07:02:33 +00:00
|
|
|
offsetof(struct bpf_array, ptrs));
|
2015-05-19 23:59:04 +00:00
|
|
|
|
x86/bpf: Clean up non-standard comments, to make the code more readable
So by chance I looked into x86 assembly in arch/x86/net/bpf_jit_comp.c and
noticed the weird and inconsistent comment style it mistakenly learned from
the networking code:
/* Multi-line comment ...
* ... looks like this.
*/
Fix this to use the standard comment style specified in Documentation/CodingStyle
and used in arch/x86/ as well:
/*
* Multi-line comment ...
* ... looks like this.
*/
Also, to quote Linus's ... more explicit views about this:
http://article.gmane.org/gmane.linux.kernel.cryptoapi/21066
> But no, the networking code picked *none* of the above sane formats.
> Instead, it picked these two models that are just half-arsed
> shit-for-brains:
>
> (no)
> /* This is disgusting drug-induced
> * crap, and should die
> */
>
> (no-no-no)
> /* This is also very nasty
> * and visually unbalanced */
>
> Please. The networking code actually has the *worst* possible comment
> style. You can literally find that (no-no-no) style, which is just
> really horribly disgusting and worse than the otherwise fairly similar
> (d) in pretty much every way.
Also improve the comments and some other details while at it:
- Don't mix same-line and previous-line comment style on otherwise
identical code patterns within the same function,
- capitalize 'BPF' and x86 register names consistently,
- capitalize sentences consistently,
- instead of 'x64' use 'x86-64': x64 is a Microsoft specific term,
- use more consistent punctuation,
- use standard coding style in macros as well,
- fix typos and a few other minor details.
Consistent coding style is not optional, at least in arch/x86/.
No change in functionality.
( In case this commit causes conflicts with pending development code
I'll be glad to help resolve any conflicts! )
Acked-by: Thomas Gleixner <tglx@linutronix.de>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Borislav Petkov <bp@alien8.de>
Cc: H. Peter Anvin <hpa@zytor.com>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: David S. Miller <davem@davemloft.net>
Cc: Eric Dumazet <edumazet@google.com>
Cc: Daniel Borkmann <daniel@iogearbox.net>
Cc: Alexei Starovoitov <ast@fb.com>
Cc: Hideaki YOSHIFUJI <yoshfuji@linux-ipv6.org>
Cc: netdev@vger.kernel.org
Cc: linux-kernel@vger.kernel.org
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
2018-04-27 09:54:40 +00:00
|
|
|
/*
|
|
|
|
* if (prog == NULL)
|
|
|
|
* goto out;
|
2015-05-19 23:59:04 +00:00
|
|
|
*/
|
bpf, x64: rework pro/epilogue and tailcall handling in JIT
This commit serves two things:
1) it optimizes BPF prologue/epilogue generation
2) it makes possible to have tailcalls within BPF subprogram
Both points are related to each other since without 1), 2) could not be
achieved.
In [1], Alexei says:
"The prologue will look like:
nop5
xor eax,eax // two new bytes if bpf_tail_call() is used in this
// function
push rbp
mov rbp, rsp
sub rsp, rounded_stack_depth
push rax // zero init tail_call counter
variable number of push rbx,r13,r14,r15
Then bpf_tail_call will pop variable number rbx,..
and final 'pop rax'
Then 'add rsp, size_of_current_stack_frame'
jmp to next function and skip over 'nop5; xor eax,eax; push rpb; mov
rbp, rsp'
This way new function will set its own stack size and will init tail
call
counter with whatever value the parent had.
If next function doesn't use bpf_tail_call it won't have 'xor eax,eax'.
Instead it would need to have 'nop2' in there."
Implement that suggestion.
Since the layout of stack is changed, tail call counter handling can not
rely anymore on popping it to rbx just like it have been handled for
constant prologue case and later overwrite of rbx with actual value of
rbx pushed to stack. Therefore, let's use one of the register (%rcx) that
is considered to be volatile/caller-saved and pop the value of tail call
counter in there in the epilogue.
Drop the BUILD_BUG_ON in emit_prologue and in
emit_bpf_tail_call_indirect where instruction layout is not constant
anymore.
Introduce new poke target, 'tailcall_bypass' to poke descriptor that is
dedicated for skipping the register pops and stack unwind that are
generated right before the actual jump to target program.
For case when the target program is not present, BPF program will skip
the pop instructions and nop5 dedicated for jmpq $target. An example of
such state when only R6 of callee saved registers is used by program:
ffffffffc0513aa1: e9 0e 00 00 00 jmpq 0xffffffffc0513ab4
ffffffffc0513aa6: 5b pop %rbx
ffffffffc0513aa7: 58 pop %rax
ffffffffc0513aa8: 48 81 c4 00 00 00 00 add $0x0,%rsp
ffffffffc0513aaf: 0f 1f 44 00 00 nopl 0x0(%rax,%rax,1)
ffffffffc0513ab4: 48 89 df mov %rbx,%rdi
When target program is inserted, the jump that was there to skip
pops/nop5 will become the nop5, so CPU will go over pops and do the
actual tailcall.
One might ask why there simply can not be pushes after the nop5?
In the following example snippet:
ffffffffc037030c: 48 89 fb mov %rdi,%rbx
(...)
ffffffffc0370332: 5b pop %rbx
ffffffffc0370333: 58 pop %rax
ffffffffc0370334: 48 81 c4 00 00 00 00 add $0x0,%rsp
ffffffffc037033b: 0f 1f 44 00 00 nopl 0x0(%rax,%rax,1)
ffffffffc0370340: 48 81 ec 00 00 00 00 sub $0x0,%rsp
ffffffffc0370347: 50 push %rax
ffffffffc0370348: 53 push %rbx
ffffffffc0370349: 48 89 df mov %rbx,%rdi
ffffffffc037034c: e8 f7 21 00 00 callq 0xffffffffc0372548
There is the bpf2bpf call (at ffffffffc037034c) right after the tailcall
and jump target is not present. ctx is in %rbx register and BPF
subprogram that we will call into on ffffffffc037034c is relying on it,
e.g. it will pick ctx from there. Such code layout is therefore broken
as we would overwrite the content of %rbx with the value that was pushed
on the prologue. That is the reason for the 'bypass' approach.
Special care needs to be taken during the install/update/remove of
tailcall target. In case when target program is not present, the CPU
must not execute the pop instructions that precede the tailcall.
To address that, the following states can be defined:
A nop, unwind, nop
B nop, unwind, tail
C skip, unwind, nop
D skip, unwind, tail
A is forbidden (lead to incorrectness). The state transitions between
tailcall install/update/remove will work as follows:
First install tail call f: C->D->B(f)
* poke the tailcall, after that get rid of the skip
Update tail call f to f': B(f)->B(f')
* poke the tailcall (poke->tailcall_target) and do NOT touch the
poke->tailcall_bypass
Remove tail call: B(f')->C(f')
* poke->tailcall_bypass is poked back to jump, then we wait the RCU
grace period so that other programs will finish its execution and
after that we are safe to remove the poke->tailcall_target
Install new tail call (f''): C(f')->D(f'')->B(f'').
* same as first step
This way CPU can never be exposed to "unwind, tail" state.
Last but not least, when tailcalls get mixed with bpf2bpf calls, it
would be possible to encounter the endless loop due to clearing the
tailcall counter if for example we would use the tailcall3-like from BPF
selftests program that would be subprogram-based, meaning the tailcall
would be present within the BPF subprogram.
This test, broken down to particular steps, would do:
entry -> set tailcall counter to 0, bump it by 1, tailcall to func0
func0 -> call subprog_tail
(we are NOT skipping the first 11 bytes of prologue and this subprogram
has a tailcall, therefore we clear the counter...)
subprog -> do the same thing as entry
and then loop forever.
To address this, the idea is to go through the call chain of bpf2bpf progs
and look for a tailcall presence throughout whole chain. If we saw a single
tail call then each node in this call chain needs to be marked as a subprog
that can reach the tailcall. We would later feed the JIT with this info
and:
- set eax to 0 only when tailcall is reachable and this is the entry prog
- if tailcall is reachable but there's no tailcall in insns of currently
JITed prog then push rax anyway, so that it will be possible to
propagate further down the call chain
- finally if tailcall is reachable, then we need to precede the 'call'
insn with mov rax, [rbp - (stack_depth + 8)]
Tail call related cases from test_verifier kselftest are also working
fine. Sample BPF programs that utilize tail calls (sockex3, tracex5)
work properly as well.
[1]: https://lore.kernel.org/bpf/20200517043227.2gpq22ifoq37ogst@ast-mbp.dhcp.thefacebook.com/
Suggested-by: Alexei Starovoitov <ast@kernel.org>
Signed-off-by: Maciej Fijalkowski <maciej.fijalkowski@intel.com>
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
2020-09-16 21:10:08 +00:00
|
|
|
EMIT3(0x48, 0x85, 0xC9); /* test rcx,rcx */
|
2015-05-19 23:59:04 +00:00
|
|
|
|
2021-10-26 12:01:47 +00:00
|
|
|
offset = ctx->tail_call_indirect_label - (prog + 2 - start);
|
|
|
|
EMIT2(X86_JE, offset); /* je out */
|
|
|
|
|
bpf: Implement BPF exceptions
This patch implements BPF exceptions, and introduces a bpf_throw kfunc
to allow programs to throw exceptions during their execution at runtime.
A bpf_throw invocation is treated as an immediate termination of the
program, returning back to its caller within the kernel, unwinding all
stack frames.
This allows the program to simplify its implementation, by testing for
runtime conditions which the verifier has no visibility into, and assert
that they are true. In case they are not, the program can simply throw
an exception from the other branch.
BPF exceptions are explicitly *NOT* an unlikely slowpath error handling
primitive, and this objective has guided design choices of the
implementation of the them within the kernel (with the bulk of the cost
for unwinding the stack offloaded to the bpf_throw kfunc).
The implementation of this mechanism requires use of add_hidden_subprog
mechanism introduced in the previous patch, which generates a couple of
instructions to move R1 to R0 and exit. The JIT then rewrites the
prologue of this subprog to take the stack pointer and frame pointer as
inputs and reset the stack frame, popping all callee-saved registers
saved by the main subprog. The bpf_throw function then walks the stack
at runtime, and invokes this exception subprog with the stack and frame
pointers as parameters.
Reviewers must take note that currently the main program is made to save
all callee-saved registers on x86_64 during entry into the program. This
is because we must do an equivalent of a lightweight context switch when
unwinding the stack, therefore we need the callee-saved registers of the
caller of the BPF program to be able to return with a sane state.
Note that we have to additionally handle r12, even though it is not used
by the program, because when throwing the exception the program makes an
entry into the kernel which could clobber r12 after saving it on the
stack. To be able to preserve the value we received on program entry, we
push r12 and restore it from the generated subprogram when unwinding the
stack.
For now, bpf_throw invocation fails when lingering resources or locks
exist in that path of the program. In a future followup, bpf_throw will
be extended to perform frame-by-frame unwinding to release lingering
resources for each stack frame, removing this limitation.
Signed-off-by: Kumar Kartikeya Dwivedi <memxor@gmail.com>
Link: https://lore.kernel.org/r/20230912233214.1518551-5-memxor@gmail.com
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
2023-09-12 23:32:01 +00:00
|
|
|
if (bpf_prog->aux->exception_boundary) {
|
|
|
|
pop_callee_regs(&prog, all_callee_regs_used);
|
|
|
|
pop_r12(&prog);
|
|
|
|
} else {
|
|
|
|
pop_callee_regs(&prog, callee_regs_used);
|
2024-03-08 01:08:01 +00:00
|
|
|
if (bpf_arena_get_kern_vm_start(bpf_prog->aux->arena))
|
|
|
|
pop_r12(&prog);
|
bpf: Implement BPF exceptions
This patch implements BPF exceptions, and introduces a bpf_throw kfunc
to allow programs to throw exceptions during their execution at runtime.
A bpf_throw invocation is treated as an immediate termination of the
program, returning back to its caller within the kernel, unwinding all
stack frames.
This allows the program to simplify its implementation, by testing for
runtime conditions which the verifier has no visibility into, and assert
that they are true. In case they are not, the program can simply throw
an exception from the other branch.
BPF exceptions are explicitly *NOT* an unlikely slowpath error handling
primitive, and this objective has guided design choices of the
implementation of the them within the kernel (with the bulk of the cost
for unwinding the stack offloaded to the bpf_throw kfunc).
The implementation of this mechanism requires use of add_hidden_subprog
mechanism introduced in the previous patch, which generates a couple of
instructions to move R1 to R0 and exit. The JIT then rewrites the
prologue of this subprog to take the stack pointer and frame pointer as
inputs and reset the stack frame, popping all callee-saved registers
saved by the main subprog. The bpf_throw function then walks the stack
at runtime, and invokes this exception subprog with the stack and frame
pointers as parameters.
Reviewers must take note that currently the main program is made to save
all callee-saved registers on x86_64 during entry into the program. This
is because we must do an equivalent of a lightweight context switch when
unwinding the stack, therefore we need the callee-saved registers of the
caller of the BPF program to be able to return with a sane state.
Note that we have to additionally handle r12, even though it is not used
by the program, because when throwing the exception the program makes an
entry into the kernel which could clobber r12 after saving it on the
stack. To be able to preserve the value we received on program entry, we
push r12 and restore it from the generated subprogram when unwinding the
stack.
For now, bpf_throw invocation fails when lingering resources or locks
exist in that path of the program. In a future followup, bpf_throw will
be extended to perform frame-by-frame unwinding to release lingering
resources for each stack frame, removing this limitation.
Signed-off-by: Kumar Kartikeya Dwivedi <memxor@gmail.com>
Link: https://lore.kernel.org/r/20230912233214.1518551-5-memxor@gmail.com
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
2023-09-12 23:32:01 +00:00
|
|
|
}
|
bpf, x64: rework pro/epilogue and tailcall handling in JIT
This commit serves two things:
1) it optimizes BPF prologue/epilogue generation
2) it makes possible to have tailcalls within BPF subprogram
Both points are related to each other since without 1), 2) could not be
achieved.
In [1], Alexei says:
"The prologue will look like:
nop5
xor eax,eax // two new bytes if bpf_tail_call() is used in this
// function
push rbp
mov rbp, rsp
sub rsp, rounded_stack_depth
push rax // zero init tail_call counter
variable number of push rbx,r13,r14,r15
Then bpf_tail_call will pop variable number rbx,..
and final 'pop rax'
Then 'add rsp, size_of_current_stack_frame'
jmp to next function and skip over 'nop5; xor eax,eax; push rpb; mov
rbp, rsp'
This way new function will set its own stack size and will init tail
call
counter with whatever value the parent had.
If next function doesn't use bpf_tail_call it won't have 'xor eax,eax'.
Instead it would need to have 'nop2' in there."
Implement that suggestion.
Since the layout of stack is changed, tail call counter handling can not
rely anymore on popping it to rbx just like it have been handled for
constant prologue case and later overwrite of rbx with actual value of
rbx pushed to stack. Therefore, let's use one of the register (%rcx) that
is considered to be volatile/caller-saved and pop the value of tail call
counter in there in the epilogue.
Drop the BUILD_BUG_ON in emit_prologue and in
emit_bpf_tail_call_indirect where instruction layout is not constant
anymore.
Introduce new poke target, 'tailcall_bypass' to poke descriptor that is
dedicated for skipping the register pops and stack unwind that are
generated right before the actual jump to target program.
For case when the target program is not present, BPF program will skip
the pop instructions and nop5 dedicated for jmpq $target. An example of
such state when only R6 of callee saved registers is used by program:
ffffffffc0513aa1: e9 0e 00 00 00 jmpq 0xffffffffc0513ab4
ffffffffc0513aa6: 5b pop %rbx
ffffffffc0513aa7: 58 pop %rax
ffffffffc0513aa8: 48 81 c4 00 00 00 00 add $0x0,%rsp
ffffffffc0513aaf: 0f 1f 44 00 00 nopl 0x0(%rax,%rax,1)
ffffffffc0513ab4: 48 89 df mov %rbx,%rdi
When target program is inserted, the jump that was there to skip
pops/nop5 will become the nop5, so CPU will go over pops and do the
actual tailcall.
One might ask why there simply can not be pushes after the nop5?
In the following example snippet:
ffffffffc037030c: 48 89 fb mov %rdi,%rbx
(...)
ffffffffc0370332: 5b pop %rbx
ffffffffc0370333: 58 pop %rax
ffffffffc0370334: 48 81 c4 00 00 00 00 add $0x0,%rsp
ffffffffc037033b: 0f 1f 44 00 00 nopl 0x0(%rax,%rax,1)
ffffffffc0370340: 48 81 ec 00 00 00 00 sub $0x0,%rsp
ffffffffc0370347: 50 push %rax
ffffffffc0370348: 53 push %rbx
ffffffffc0370349: 48 89 df mov %rbx,%rdi
ffffffffc037034c: e8 f7 21 00 00 callq 0xffffffffc0372548
There is the bpf2bpf call (at ffffffffc037034c) right after the tailcall
and jump target is not present. ctx is in %rbx register and BPF
subprogram that we will call into on ffffffffc037034c is relying on it,
e.g. it will pick ctx from there. Such code layout is therefore broken
as we would overwrite the content of %rbx with the value that was pushed
on the prologue. That is the reason for the 'bypass' approach.
Special care needs to be taken during the install/update/remove of
tailcall target. In case when target program is not present, the CPU
must not execute the pop instructions that precede the tailcall.
To address that, the following states can be defined:
A nop, unwind, nop
B nop, unwind, tail
C skip, unwind, nop
D skip, unwind, tail
A is forbidden (lead to incorrectness). The state transitions between
tailcall install/update/remove will work as follows:
First install tail call f: C->D->B(f)
* poke the tailcall, after that get rid of the skip
Update tail call f to f': B(f)->B(f')
* poke the tailcall (poke->tailcall_target) and do NOT touch the
poke->tailcall_bypass
Remove tail call: B(f')->C(f')
* poke->tailcall_bypass is poked back to jump, then we wait the RCU
grace period so that other programs will finish its execution and
after that we are safe to remove the poke->tailcall_target
Install new tail call (f''): C(f')->D(f'')->B(f'').
* same as first step
This way CPU can never be exposed to "unwind, tail" state.
Last but not least, when tailcalls get mixed with bpf2bpf calls, it
would be possible to encounter the endless loop due to clearing the
tailcall counter if for example we would use the tailcall3-like from BPF
selftests program that would be subprogram-based, meaning the tailcall
would be present within the BPF subprogram.
This test, broken down to particular steps, would do:
entry -> set tailcall counter to 0, bump it by 1, tailcall to func0
func0 -> call subprog_tail
(we are NOT skipping the first 11 bytes of prologue and this subprogram
has a tailcall, therefore we clear the counter...)
subprog -> do the same thing as entry
and then loop forever.
To address this, the idea is to go through the call chain of bpf2bpf progs
and look for a tailcall presence throughout whole chain. If we saw a single
tail call then each node in this call chain needs to be marked as a subprog
that can reach the tailcall. We would later feed the JIT with this info
and:
- set eax to 0 only when tailcall is reachable and this is the entry prog
- if tailcall is reachable but there's no tailcall in insns of currently
JITed prog then push rax anyway, so that it will be possible to
propagate further down the call chain
- finally if tailcall is reachable, then we need to precede the 'call'
insn with mov rax, [rbp - (stack_depth + 8)]
Tail call related cases from test_verifier kselftest are also working
fine. Sample BPF programs that utilize tail calls (sockex3, tracex5)
work properly as well.
[1]: https://lore.kernel.org/bpf/20200517043227.2gpq22ifoq37ogst@ast-mbp.dhcp.thefacebook.com/
Suggested-by: Alexei Starovoitov <ast@kernel.org>
Signed-off-by: Maciej Fijalkowski <maciej.fijalkowski@intel.com>
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
2020-09-16 21:10:08 +00:00
|
|
|
|
|
|
|
EMIT1(0x58); /* pop rax */
|
2020-09-29 20:46:53 +00:00
|
|
|
if (stack_depth)
|
|
|
|
EMIT3_off32(0x48, 0x81, 0xC4, /* add rsp, sd */
|
|
|
|
round_up(stack_depth, 8));
|
bpf, x64: rework pro/epilogue and tailcall handling in JIT
This commit serves two things:
1) it optimizes BPF prologue/epilogue generation
2) it makes possible to have tailcalls within BPF subprogram
Both points are related to each other since without 1), 2) could not be
achieved.
In [1], Alexei says:
"The prologue will look like:
nop5
xor eax,eax // two new bytes if bpf_tail_call() is used in this
// function
push rbp
mov rbp, rsp
sub rsp, rounded_stack_depth
push rax // zero init tail_call counter
variable number of push rbx,r13,r14,r15
Then bpf_tail_call will pop variable number rbx,..
and final 'pop rax'
Then 'add rsp, size_of_current_stack_frame'
jmp to next function and skip over 'nop5; xor eax,eax; push rpb; mov
rbp, rsp'
This way new function will set its own stack size and will init tail
call
counter with whatever value the parent had.
If next function doesn't use bpf_tail_call it won't have 'xor eax,eax'.
Instead it would need to have 'nop2' in there."
Implement that suggestion.
Since the layout of stack is changed, tail call counter handling can not
rely anymore on popping it to rbx just like it have been handled for
constant prologue case and later overwrite of rbx with actual value of
rbx pushed to stack. Therefore, let's use one of the register (%rcx) that
is considered to be volatile/caller-saved and pop the value of tail call
counter in there in the epilogue.
Drop the BUILD_BUG_ON in emit_prologue and in
emit_bpf_tail_call_indirect where instruction layout is not constant
anymore.
Introduce new poke target, 'tailcall_bypass' to poke descriptor that is
dedicated for skipping the register pops and stack unwind that are
generated right before the actual jump to target program.
For case when the target program is not present, BPF program will skip
the pop instructions and nop5 dedicated for jmpq $target. An example of
such state when only R6 of callee saved registers is used by program:
ffffffffc0513aa1: e9 0e 00 00 00 jmpq 0xffffffffc0513ab4
ffffffffc0513aa6: 5b pop %rbx
ffffffffc0513aa7: 58 pop %rax
ffffffffc0513aa8: 48 81 c4 00 00 00 00 add $0x0,%rsp
ffffffffc0513aaf: 0f 1f 44 00 00 nopl 0x0(%rax,%rax,1)
ffffffffc0513ab4: 48 89 df mov %rbx,%rdi
When target program is inserted, the jump that was there to skip
pops/nop5 will become the nop5, so CPU will go over pops and do the
actual tailcall.
One might ask why there simply can not be pushes after the nop5?
In the following example snippet:
ffffffffc037030c: 48 89 fb mov %rdi,%rbx
(...)
ffffffffc0370332: 5b pop %rbx
ffffffffc0370333: 58 pop %rax
ffffffffc0370334: 48 81 c4 00 00 00 00 add $0x0,%rsp
ffffffffc037033b: 0f 1f 44 00 00 nopl 0x0(%rax,%rax,1)
ffffffffc0370340: 48 81 ec 00 00 00 00 sub $0x0,%rsp
ffffffffc0370347: 50 push %rax
ffffffffc0370348: 53 push %rbx
ffffffffc0370349: 48 89 df mov %rbx,%rdi
ffffffffc037034c: e8 f7 21 00 00 callq 0xffffffffc0372548
There is the bpf2bpf call (at ffffffffc037034c) right after the tailcall
and jump target is not present. ctx is in %rbx register and BPF
subprogram that we will call into on ffffffffc037034c is relying on it,
e.g. it will pick ctx from there. Such code layout is therefore broken
as we would overwrite the content of %rbx with the value that was pushed
on the prologue. That is the reason for the 'bypass' approach.
Special care needs to be taken during the install/update/remove of
tailcall target. In case when target program is not present, the CPU
must not execute the pop instructions that precede the tailcall.
To address that, the following states can be defined:
A nop, unwind, nop
B nop, unwind, tail
C skip, unwind, nop
D skip, unwind, tail
A is forbidden (lead to incorrectness). The state transitions between
tailcall install/update/remove will work as follows:
First install tail call f: C->D->B(f)
* poke the tailcall, after that get rid of the skip
Update tail call f to f': B(f)->B(f')
* poke the tailcall (poke->tailcall_target) and do NOT touch the
poke->tailcall_bypass
Remove tail call: B(f')->C(f')
* poke->tailcall_bypass is poked back to jump, then we wait the RCU
grace period so that other programs will finish its execution and
after that we are safe to remove the poke->tailcall_target
Install new tail call (f''): C(f')->D(f'')->B(f'').
* same as first step
This way CPU can never be exposed to "unwind, tail" state.
Last but not least, when tailcalls get mixed with bpf2bpf calls, it
would be possible to encounter the endless loop due to clearing the
tailcall counter if for example we would use the tailcall3-like from BPF
selftests program that would be subprogram-based, meaning the tailcall
would be present within the BPF subprogram.
This test, broken down to particular steps, would do:
entry -> set tailcall counter to 0, bump it by 1, tailcall to func0
func0 -> call subprog_tail
(we are NOT skipping the first 11 bytes of prologue and this subprogram
has a tailcall, therefore we clear the counter...)
subprog -> do the same thing as entry
and then loop forever.
To address this, the idea is to go through the call chain of bpf2bpf progs
and look for a tailcall presence throughout whole chain. If we saw a single
tail call then each node in this call chain needs to be marked as a subprog
that can reach the tailcall. We would later feed the JIT with this info
and:
- set eax to 0 only when tailcall is reachable and this is the entry prog
- if tailcall is reachable but there's no tailcall in insns of currently
JITed prog then push rax anyway, so that it will be possible to
propagate further down the call chain
- finally if tailcall is reachable, then we need to precede the 'call'
insn with mov rax, [rbp - (stack_depth + 8)]
Tail call related cases from test_verifier kselftest are also working
fine. Sample BPF programs that utilize tail calls (sockex3, tracex5)
work properly as well.
[1]: https://lore.kernel.org/bpf/20200517043227.2gpq22ifoq37ogst@ast-mbp.dhcp.thefacebook.com/
Suggested-by: Alexei Starovoitov <ast@kernel.org>
Signed-off-by: Maciej Fijalkowski <maciej.fijalkowski@intel.com>
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
2020-09-16 21:10:08 +00:00
|
|
|
|
|
|
|
/* goto *(prog->bpf_func + X86_TAIL_CALL_OFFSET); */
|
2020-09-16 21:10:04 +00:00
|
|
|
EMIT4(0x48, 0x8B, 0x49, /* mov rcx, qword ptr [rcx + 32] */
|
2015-05-19 23:59:04 +00:00
|
|
|
offsetof(struct bpf_prog, bpf_func));
|
bpf, x64: rework pro/epilogue and tailcall handling in JIT
This commit serves two things:
1) it optimizes BPF prologue/epilogue generation
2) it makes possible to have tailcalls within BPF subprogram
Both points are related to each other since without 1), 2) could not be
achieved.
In [1], Alexei says:
"The prologue will look like:
nop5
xor eax,eax // two new bytes if bpf_tail_call() is used in this
// function
push rbp
mov rbp, rsp
sub rsp, rounded_stack_depth
push rax // zero init tail_call counter
variable number of push rbx,r13,r14,r15
Then bpf_tail_call will pop variable number rbx,..
and final 'pop rax'
Then 'add rsp, size_of_current_stack_frame'
jmp to next function and skip over 'nop5; xor eax,eax; push rpb; mov
rbp, rsp'
This way new function will set its own stack size and will init tail
call
counter with whatever value the parent had.
If next function doesn't use bpf_tail_call it won't have 'xor eax,eax'.
Instead it would need to have 'nop2' in there."
Implement that suggestion.
Since the layout of stack is changed, tail call counter handling can not
rely anymore on popping it to rbx just like it have been handled for
constant prologue case and later overwrite of rbx with actual value of
rbx pushed to stack. Therefore, let's use one of the register (%rcx) that
is considered to be volatile/caller-saved and pop the value of tail call
counter in there in the epilogue.
Drop the BUILD_BUG_ON in emit_prologue and in
emit_bpf_tail_call_indirect where instruction layout is not constant
anymore.
Introduce new poke target, 'tailcall_bypass' to poke descriptor that is
dedicated for skipping the register pops and stack unwind that are
generated right before the actual jump to target program.
For case when the target program is not present, BPF program will skip
the pop instructions and nop5 dedicated for jmpq $target. An example of
such state when only R6 of callee saved registers is used by program:
ffffffffc0513aa1: e9 0e 00 00 00 jmpq 0xffffffffc0513ab4
ffffffffc0513aa6: 5b pop %rbx
ffffffffc0513aa7: 58 pop %rax
ffffffffc0513aa8: 48 81 c4 00 00 00 00 add $0x0,%rsp
ffffffffc0513aaf: 0f 1f 44 00 00 nopl 0x0(%rax,%rax,1)
ffffffffc0513ab4: 48 89 df mov %rbx,%rdi
When target program is inserted, the jump that was there to skip
pops/nop5 will become the nop5, so CPU will go over pops and do the
actual tailcall.
One might ask why there simply can not be pushes after the nop5?
In the following example snippet:
ffffffffc037030c: 48 89 fb mov %rdi,%rbx
(...)
ffffffffc0370332: 5b pop %rbx
ffffffffc0370333: 58 pop %rax
ffffffffc0370334: 48 81 c4 00 00 00 00 add $0x0,%rsp
ffffffffc037033b: 0f 1f 44 00 00 nopl 0x0(%rax,%rax,1)
ffffffffc0370340: 48 81 ec 00 00 00 00 sub $0x0,%rsp
ffffffffc0370347: 50 push %rax
ffffffffc0370348: 53 push %rbx
ffffffffc0370349: 48 89 df mov %rbx,%rdi
ffffffffc037034c: e8 f7 21 00 00 callq 0xffffffffc0372548
There is the bpf2bpf call (at ffffffffc037034c) right after the tailcall
and jump target is not present. ctx is in %rbx register and BPF
subprogram that we will call into on ffffffffc037034c is relying on it,
e.g. it will pick ctx from there. Such code layout is therefore broken
as we would overwrite the content of %rbx with the value that was pushed
on the prologue. That is the reason for the 'bypass' approach.
Special care needs to be taken during the install/update/remove of
tailcall target. In case when target program is not present, the CPU
must not execute the pop instructions that precede the tailcall.
To address that, the following states can be defined:
A nop, unwind, nop
B nop, unwind, tail
C skip, unwind, nop
D skip, unwind, tail
A is forbidden (lead to incorrectness). The state transitions between
tailcall install/update/remove will work as follows:
First install tail call f: C->D->B(f)
* poke the tailcall, after that get rid of the skip
Update tail call f to f': B(f)->B(f')
* poke the tailcall (poke->tailcall_target) and do NOT touch the
poke->tailcall_bypass
Remove tail call: B(f')->C(f')
* poke->tailcall_bypass is poked back to jump, then we wait the RCU
grace period so that other programs will finish its execution and
after that we are safe to remove the poke->tailcall_target
Install new tail call (f''): C(f')->D(f'')->B(f'').
* same as first step
This way CPU can never be exposed to "unwind, tail" state.
Last but not least, when tailcalls get mixed with bpf2bpf calls, it
would be possible to encounter the endless loop due to clearing the
tailcall counter if for example we would use the tailcall3-like from BPF
selftests program that would be subprogram-based, meaning the tailcall
would be present within the BPF subprogram.
This test, broken down to particular steps, would do:
entry -> set tailcall counter to 0, bump it by 1, tailcall to func0
func0 -> call subprog_tail
(we are NOT skipping the first 11 bytes of prologue and this subprogram
has a tailcall, therefore we clear the counter...)
subprog -> do the same thing as entry
and then loop forever.
To address this, the idea is to go through the call chain of bpf2bpf progs
and look for a tailcall presence throughout whole chain. If we saw a single
tail call then each node in this call chain needs to be marked as a subprog
that can reach the tailcall. We would later feed the JIT with this info
and:
- set eax to 0 only when tailcall is reachable and this is the entry prog
- if tailcall is reachable but there's no tailcall in insns of currently
JITed prog then push rax anyway, so that it will be possible to
propagate further down the call chain
- finally if tailcall is reachable, then we need to precede the 'call'
insn with mov rax, [rbp - (stack_depth + 8)]
Tail call related cases from test_verifier kselftest are also working
fine. Sample BPF programs that utilize tail calls (sockex3, tracex5)
work properly as well.
[1]: https://lore.kernel.org/bpf/20200517043227.2gpq22ifoq37ogst@ast-mbp.dhcp.thefacebook.com/
Suggested-by: Alexei Starovoitov <ast@kernel.org>
Signed-off-by: Maciej Fijalkowski <maciej.fijalkowski@intel.com>
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
2020-09-16 21:10:08 +00:00
|
|
|
EMIT4(0x48, 0x83, 0xC1, /* add rcx, X86_TAIL_CALL_OFFSET */
|
|
|
|
X86_TAIL_CALL_OFFSET);
|
x86/bpf: Clean up non-standard comments, to make the code more readable
So by chance I looked into x86 assembly in arch/x86/net/bpf_jit_comp.c and
noticed the weird and inconsistent comment style it mistakenly learned from
the networking code:
/* Multi-line comment ...
* ... looks like this.
*/
Fix this to use the standard comment style specified in Documentation/CodingStyle
and used in arch/x86/ as well:
/*
* Multi-line comment ...
* ... looks like this.
*/
Also, to quote Linus's ... more explicit views about this:
http://article.gmane.org/gmane.linux.kernel.cryptoapi/21066
> But no, the networking code picked *none* of the above sane formats.
> Instead, it picked these two models that are just half-arsed
> shit-for-brains:
>
> (no)
> /* This is disgusting drug-induced
> * crap, and should die
> */
>
> (no-no-no)
> /* This is also very nasty
> * and visually unbalanced */
>
> Please. The networking code actually has the *worst* possible comment
> style. You can literally find that (no-no-no) style, which is just
> really horribly disgusting and worse than the otherwise fairly similar
> (d) in pretty much every way.
Also improve the comments and some other details while at it:
- Don't mix same-line and previous-line comment style on otherwise
identical code patterns within the same function,
- capitalize 'BPF' and x86 register names consistently,
- capitalize sentences consistently,
- instead of 'x64' use 'x86-64': x64 is a Microsoft specific term,
- use more consistent punctuation,
- use standard coding style in macros as well,
- fix typos and a few other minor details.
Consistent coding style is not optional, at least in arch/x86/.
No change in functionality.
( In case this commit causes conflicts with pending development code
I'll be glad to help resolve any conflicts! )
Acked-by: Thomas Gleixner <tglx@linutronix.de>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Borislav Petkov <bp@alien8.de>
Cc: H. Peter Anvin <hpa@zytor.com>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: David S. Miller <davem@davemloft.net>
Cc: Eric Dumazet <edumazet@google.com>
Cc: Daniel Borkmann <daniel@iogearbox.net>
Cc: Alexei Starovoitov <ast@fb.com>
Cc: Hideaki YOSHIFUJI <yoshfuji@linux-ipv6.org>
Cc: netdev@vger.kernel.org
Cc: linux-kernel@vger.kernel.org
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
2018-04-27 09:54:40 +00:00
|
|
|
/*
|
2020-09-16 21:10:04 +00:00
|
|
|
* Now we're ready to jump into next BPF program
|
2015-05-19 23:59:04 +00:00
|
|
|
* rdi == ctx (1st arg)
|
bpf, x64: rework pro/epilogue and tailcall handling in JIT
This commit serves two things:
1) it optimizes BPF prologue/epilogue generation
2) it makes possible to have tailcalls within BPF subprogram
Both points are related to each other since without 1), 2) could not be
achieved.
In [1], Alexei says:
"The prologue will look like:
nop5
xor eax,eax // two new bytes if bpf_tail_call() is used in this
// function
push rbp
mov rbp, rsp
sub rsp, rounded_stack_depth
push rax // zero init tail_call counter
variable number of push rbx,r13,r14,r15
Then bpf_tail_call will pop variable number rbx,..
and final 'pop rax'
Then 'add rsp, size_of_current_stack_frame'
jmp to next function and skip over 'nop5; xor eax,eax; push rpb; mov
rbp, rsp'
This way new function will set its own stack size and will init tail
call
counter with whatever value the parent had.
If next function doesn't use bpf_tail_call it won't have 'xor eax,eax'.
Instead it would need to have 'nop2' in there."
Implement that suggestion.
Since the layout of stack is changed, tail call counter handling can not
rely anymore on popping it to rbx just like it have been handled for
constant prologue case and later overwrite of rbx with actual value of
rbx pushed to stack. Therefore, let's use one of the register (%rcx) that
is considered to be volatile/caller-saved and pop the value of tail call
counter in there in the epilogue.
Drop the BUILD_BUG_ON in emit_prologue and in
emit_bpf_tail_call_indirect where instruction layout is not constant
anymore.
Introduce new poke target, 'tailcall_bypass' to poke descriptor that is
dedicated for skipping the register pops and stack unwind that are
generated right before the actual jump to target program.
For case when the target program is not present, BPF program will skip
the pop instructions and nop5 dedicated for jmpq $target. An example of
such state when only R6 of callee saved registers is used by program:
ffffffffc0513aa1: e9 0e 00 00 00 jmpq 0xffffffffc0513ab4
ffffffffc0513aa6: 5b pop %rbx
ffffffffc0513aa7: 58 pop %rax
ffffffffc0513aa8: 48 81 c4 00 00 00 00 add $0x0,%rsp
ffffffffc0513aaf: 0f 1f 44 00 00 nopl 0x0(%rax,%rax,1)
ffffffffc0513ab4: 48 89 df mov %rbx,%rdi
When target program is inserted, the jump that was there to skip
pops/nop5 will become the nop5, so CPU will go over pops and do the
actual tailcall.
One might ask why there simply can not be pushes after the nop5?
In the following example snippet:
ffffffffc037030c: 48 89 fb mov %rdi,%rbx
(...)
ffffffffc0370332: 5b pop %rbx
ffffffffc0370333: 58 pop %rax
ffffffffc0370334: 48 81 c4 00 00 00 00 add $0x0,%rsp
ffffffffc037033b: 0f 1f 44 00 00 nopl 0x0(%rax,%rax,1)
ffffffffc0370340: 48 81 ec 00 00 00 00 sub $0x0,%rsp
ffffffffc0370347: 50 push %rax
ffffffffc0370348: 53 push %rbx
ffffffffc0370349: 48 89 df mov %rbx,%rdi
ffffffffc037034c: e8 f7 21 00 00 callq 0xffffffffc0372548
There is the bpf2bpf call (at ffffffffc037034c) right after the tailcall
and jump target is not present. ctx is in %rbx register and BPF
subprogram that we will call into on ffffffffc037034c is relying on it,
e.g. it will pick ctx from there. Such code layout is therefore broken
as we would overwrite the content of %rbx with the value that was pushed
on the prologue. That is the reason for the 'bypass' approach.
Special care needs to be taken during the install/update/remove of
tailcall target. In case when target program is not present, the CPU
must not execute the pop instructions that precede the tailcall.
To address that, the following states can be defined:
A nop, unwind, nop
B nop, unwind, tail
C skip, unwind, nop
D skip, unwind, tail
A is forbidden (lead to incorrectness). The state transitions between
tailcall install/update/remove will work as follows:
First install tail call f: C->D->B(f)
* poke the tailcall, after that get rid of the skip
Update tail call f to f': B(f)->B(f')
* poke the tailcall (poke->tailcall_target) and do NOT touch the
poke->tailcall_bypass
Remove tail call: B(f')->C(f')
* poke->tailcall_bypass is poked back to jump, then we wait the RCU
grace period so that other programs will finish its execution and
after that we are safe to remove the poke->tailcall_target
Install new tail call (f''): C(f')->D(f'')->B(f'').
* same as first step
This way CPU can never be exposed to "unwind, tail" state.
Last but not least, when tailcalls get mixed with bpf2bpf calls, it
would be possible to encounter the endless loop due to clearing the
tailcall counter if for example we would use the tailcall3-like from BPF
selftests program that would be subprogram-based, meaning the tailcall
would be present within the BPF subprogram.
This test, broken down to particular steps, would do:
entry -> set tailcall counter to 0, bump it by 1, tailcall to func0
func0 -> call subprog_tail
(we are NOT skipping the first 11 bytes of prologue and this subprogram
has a tailcall, therefore we clear the counter...)
subprog -> do the same thing as entry
and then loop forever.
To address this, the idea is to go through the call chain of bpf2bpf progs
and look for a tailcall presence throughout whole chain. If we saw a single
tail call then each node in this call chain needs to be marked as a subprog
that can reach the tailcall. We would later feed the JIT with this info
and:
- set eax to 0 only when tailcall is reachable and this is the entry prog
- if tailcall is reachable but there's no tailcall in insns of currently
JITed prog then push rax anyway, so that it will be possible to
propagate further down the call chain
- finally if tailcall is reachable, then we need to precede the 'call'
insn with mov rax, [rbp - (stack_depth + 8)]
Tail call related cases from test_verifier kselftest are also working
fine. Sample BPF programs that utilize tail calls (sockex3, tracex5)
work properly as well.
[1]: https://lore.kernel.org/bpf/20200517043227.2gpq22ifoq37ogst@ast-mbp.dhcp.thefacebook.com/
Suggested-by: Alexei Starovoitov <ast@kernel.org>
Signed-off-by: Maciej Fijalkowski <maciej.fijalkowski@intel.com>
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
2020-09-16 21:10:08 +00:00
|
|
|
* rcx == prog->bpf_func + X86_TAIL_CALL_OFFSET
|
2015-05-19 23:59:04 +00:00
|
|
|
*/
|
2021-10-26 12:01:48 +00:00
|
|
|
emit_indirect_jump(&prog, 1 /* rcx */, ip + (prog - start));
|
2015-05-19 23:59:04 +00:00
|
|
|
|
|
|
|
/* out: */
|
2021-10-26 12:01:47 +00:00
|
|
|
ctx->tail_call_indirect_label = prog - start;
|
2015-05-19 23:59:04 +00:00
|
|
|
*pprog = prog;
|
|
|
|
}
|
|
|
|
|
bpf: Implement BPF exceptions
This patch implements BPF exceptions, and introduces a bpf_throw kfunc
to allow programs to throw exceptions during their execution at runtime.
A bpf_throw invocation is treated as an immediate termination of the
program, returning back to its caller within the kernel, unwinding all
stack frames.
This allows the program to simplify its implementation, by testing for
runtime conditions which the verifier has no visibility into, and assert
that they are true. In case they are not, the program can simply throw
an exception from the other branch.
BPF exceptions are explicitly *NOT* an unlikely slowpath error handling
primitive, and this objective has guided design choices of the
implementation of the them within the kernel (with the bulk of the cost
for unwinding the stack offloaded to the bpf_throw kfunc).
The implementation of this mechanism requires use of add_hidden_subprog
mechanism introduced in the previous patch, which generates a couple of
instructions to move R1 to R0 and exit. The JIT then rewrites the
prologue of this subprog to take the stack pointer and frame pointer as
inputs and reset the stack frame, popping all callee-saved registers
saved by the main subprog. The bpf_throw function then walks the stack
at runtime, and invokes this exception subprog with the stack and frame
pointers as parameters.
Reviewers must take note that currently the main program is made to save
all callee-saved registers on x86_64 during entry into the program. This
is because we must do an equivalent of a lightweight context switch when
unwinding the stack, therefore we need the callee-saved registers of the
caller of the BPF program to be able to return with a sane state.
Note that we have to additionally handle r12, even though it is not used
by the program, because when throwing the exception the program makes an
entry into the kernel which could clobber r12 after saving it on the
stack. To be able to preserve the value we received on program entry, we
push r12 and restore it from the generated subprogram when unwinding the
stack.
For now, bpf_throw invocation fails when lingering resources or locks
exist in that path of the program. In a future followup, bpf_throw will
be extended to perform frame-by-frame unwinding to release lingering
resources for each stack frame, removing this limitation.
Signed-off-by: Kumar Kartikeya Dwivedi <memxor@gmail.com>
Link: https://lore.kernel.org/r/20230912233214.1518551-5-memxor@gmail.com
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
2023-09-12 23:32:01 +00:00
|
|
|
static void emit_bpf_tail_call_direct(struct bpf_prog *bpf_prog,
|
|
|
|
struct bpf_jit_poke_descriptor *poke,
|
2021-10-26 12:01:47 +00:00
|
|
|
u8 **pprog, u8 *ip,
|
|
|
|
bool *callee_regs_used, u32 stack_depth,
|
|
|
|
struct jit_context *ctx)
|
bpf, x86: Emit patchable direct jump as tail call
Add initial code emission for *direct* jumps for tail call maps in
order to avoid the retpoline overhead from a493a87f38cf ("bpf, x64:
implement retpoline for tail call") for situations that allow for
it, meaning, for known constant keys at verification time which are
used as index into the tail call map. In case of Cilium which makes
heavy use of tail calls, constant keys are used in the vast majority,
only for a single occurrence we use a dynamic key.
High level outline is that if the target prog is NULL in the map, we
emit a 5-byte nop for the fall-through case and if not, we emit a
5-byte direct relative jmp to the target bpf_func + skipped prologue
offset. Later during runtime, we patch these 5-byte nop/jmps upon
tail call map update or deletions dynamically. Note that on x86-64
the direct jmp works as we reuse the same stack frame and skip
prologue (as opposed to some other JIT implementations).
One of the issues is that the tail call map slots can change at any
given time even during JITing. Therefore, we have two passes: i) emit
nops for all patchable locations during main JITing phase until we
declare prog->jited = 1 eventually. At this point the image is stable,
not public yet and with all jmps disabled. While JITing, we collect
additional info like poke->ip in order to remember the patch location
for later modifications. In ii) bpf_tail_call_direct_fixup() walks
over the progs poke_tab, locks the tail call maps poke_mutex to
prevent from parallel updates and patches in the right locations via
__bpf_arch_text_poke(). Note, the main bpf_arch_text_poke() cannot
be used at this point since we're not yet exposed to kallsyms. For
the update we use plain memcpy() since the image is not public and
still in read-write mode. After patching, we activate that poke entry
through poke->ip_stable. Meaning, at this point any tail call map
updates/deletions are not going to ignore that poke entry anymore.
Then, bpf_arch_text_poke() might still occur on the read-write image
until we finally locked it as read-only. Both modifications on the
given image are under text_mutex to avoid interference with each
other when update requests come in in parallel for different tail
call maps (current one we have locked in JIT and different one where
poke->ip_stable was already set).
Example prog:
# ./bpftool p d x i 1655
0: (b7) r3 = 0
1: (18) r2 = map[id:526]
3: (85) call bpf_tail_call#12
4: (b7) r0 = 1
5: (95) exit
Before:
# ./bpftool p d j i 1655
0xffffffffc076e55c:
0: nopl 0x0(%rax,%rax,1)
5: push %rbp
6: mov %rsp,%rbp
9: sub $0x200,%rsp
10: push %rbx
11: push %r13
13: push %r14
15: push %r15
17: pushq $0x0 _
19: xor %edx,%edx |_ index (arg 3)
1b: movabs $0xffff88d95cc82600,%rsi |_ map (arg 2)
25: mov %edx,%edx | index >= array->map.max_entries
27: cmp %edx,0x24(%rsi) |
2a: jbe 0x0000000000000066 |_
2c: mov -0x224(%rbp),%eax | tail call limit check
32: cmp $0x20,%eax |
35: ja 0x0000000000000066 |
37: add $0x1,%eax |
3a: mov %eax,-0x224(%rbp) |_
40: mov 0xd0(%rsi,%rdx,8),%rax |_ prog = array->ptrs[index]
48: test %rax,%rax | prog == NULL check
4b: je 0x0000000000000066 |_
4d: mov 0x30(%rax),%rax | goto *(prog->bpf_func + prologue_size)
51: add $0x19,%rax |
55: callq 0x0000000000000061 | retpoline for indirect jump
5a: pause |
5c: lfence |
5f: jmp 0x000000000000005a |
61: mov %rax,(%rsp) |
65: retq |_
66: mov $0x1,%eax
6b: pop %rbx
6c: pop %r15
6e: pop %r14
70: pop %r13
72: pop %rbx
73: leaveq
74: retq
After; state after JIT:
# ./bpftool p d j i 1655
0xffffffffc08e8930:
0: nopl 0x0(%rax,%rax,1)
5: push %rbp
6: mov %rsp,%rbp
9: sub $0x200,%rsp
10: push %rbx
11: push %r13
13: push %r14
15: push %r15
17: pushq $0x0 _
19: xor %edx,%edx |_ index (arg 3)
1b: movabs $0xffff9d8afd74c000,%rsi |_ map (arg 2)
25: mov -0x224(%rbp),%eax | tail call limit check
2b: cmp $0x20,%eax |
2e: ja 0x000000000000003e |
30: add $0x1,%eax |
33: mov %eax,-0x224(%rbp) |_
39: jmpq 0xfffffffffffd1785 |_ [direct] goto *(prog->bpf_func + prologue_size)
3e: mov $0x1,%eax
43: pop %rbx
44: pop %r15
46: pop %r14
48: pop %r13
4a: pop %rbx
4b: leaveq
4c: retq
After; state after map update (target prog):
# ./bpftool p d j i 1655
0xffffffffc08e8930:
0: nopl 0x0(%rax,%rax,1)
5: push %rbp
6: mov %rsp,%rbp
9: sub $0x200,%rsp
10: push %rbx
11: push %r13
13: push %r14
15: push %r15
17: pushq $0x0
19: xor %edx,%edx
1b: movabs $0xffff9d8afd74c000,%rsi
25: mov -0x224(%rbp),%eax
2b: cmp $0x20,%eax .
2e: ja 0x000000000000003e .
30: add $0x1,%eax .
33: mov %eax,-0x224(%rbp) |_
39: jmpq 0xffffffffffb09f55 |_ goto *(prog->bpf_func + prologue_size)
3e: mov $0x1,%eax
43: pop %rbx
44: pop %r15
46: pop %r14
48: pop %r13
4a: pop %rbx
4b: leaveq
4c: retq
After; state after map update (no prog):
# ./bpftool p d j i 1655
0xffffffffc08e8930:
0: nopl 0x0(%rax,%rax,1)
5: push %rbp
6: mov %rsp,%rbp
9: sub $0x200,%rsp
10: push %rbx
11: push %r13
13: push %r14
15: push %r15
17: pushq $0x0
19: xor %edx,%edx
1b: movabs $0xffff9d8afd74c000,%rsi
25: mov -0x224(%rbp),%eax
2b: cmp $0x20,%eax .
2e: ja 0x000000000000003e .
30: add $0x1,%eax .
33: mov %eax,-0x224(%rbp) |_
39: nopl 0x0(%rax,%rax,1) |_ fall-through nop
3e: mov $0x1,%eax
43: pop %rbx
44: pop %r15
46: pop %r14
48: pop %r13
4a: pop %rbx
4b: leaveq
4c: retq
Nice bonus is that this also shrinks the code emission quite a bit
for every tail call invocation.
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Link: https://lore.kernel.org/bpf/6ada4c1c9d35eeb5f4ecfab94593dafa6b5c4b09.1574452833.git.daniel@iogearbox.net
2019-11-22 20:08:00 +00:00
|
|
|
{
|
bpf, x64: rework pro/epilogue and tailcall handling in JIT
This commit serves two things:
1) it optimizes BPF prologue/epilogue generation
2) it makes possible to have tailcalls within BPF subprogram
Both points are related to each other since without 1), 2) could not be
achieved.
In [1], Alexei says:
"The prologue will look like:
nop5
xor eax,eax // two new bytes if bpf_tail_call() is used in this
// function
push rbp
mov rbp, rsp
sub rsp, rounded_stack_depth
push rax // zero init tail_call counter
variable number of push rbx,r13,r14,r15
Then bpf_tail_call will pop variable number rbx,..
and final 'pop rax'
Then 'add rsp, size_of_current_stack_frame'
jmp to next function and skip over 'nop5; xor eax,eax; push rpb; mov
rbp, rsp'
This way new function will set its own stack size and will init tail
call
counter with whatever value the parent had.
If next function doesn't use bpf_tail_call it won't have 'xor eax,eax'.
Instead it would need to have 'nop2' in there."
Implement that suggestion.
Since the layout of stack is changed, tail call counter handling can not
rely anymore on popping it to rbx just like it have been handled for
constant prologue case and later overwrite of rbx with actual value of
rbx pushed to stack. Therefore, let's use one of the register (%rcx) that
is considered to be volatile/caller-saved and pop the value of tail call
counter in there in the epilogue.
Drop the BUILD_BUG_ON in emit_prologue and in
emit_bpf_tail_call_indirect where instruction layout is not constant
anymore.
Introduce new poke target, 'tailcall_bypass' to poke descriptor that is
dedicated for skipping the register pops and stack unwind that are
generated right before the actual jump to target program.
For case when the target program is not present, BPF program will skip
the pop instructions and nop5 dedicated for jmpq $target. An example of
such state when only R6 of callee saved registers is used by program:
ffffffffc0513aa1: e9 0e 00 00 00 jmpq 0xffffffffc0513ab4
ffffffffc0513aa6: 5b pop %rbx
ffffffffc0513aa7: 58 pop %rax
ffffffffc0513aa8: 48 81 c4 00 00 00 00 add $0x0,%rsp
ffffffffc0513aaf: 0f 1f 44 00 00 nopl 0x0(%rax,%rax,1)
ffffffffc0513ab4: 48 89 df mov %rbx,%rdi
When target program is inserted, the jump that was there to skip
pops/nop5 will become the nop5, so CPU will go over pops and do the
actual tailcall.
One might ask why there simply can not be pushes after the nop5?
In the following example snippet:
ffffffffc037030c: 48 89 fb mov %rdi,%rbx
(...)
ffffffffc0370332: 5b pop %rbx
ffffffffc0370333: 58 pop %rax
ffffffffc0370334: 48 81 c4 00 00 00 00 add $0x0,%rsp
ffffffffc037033b: 0f 1f 44 00 00 nopl 0x0(%rax,%rax,1)
ffffffffc0370340: 48 81 ec 00 00 00 00 sub $0x0,%rsp
ffffffffc0370347: 50 push %rax
ffffffffc0370348: 53 push %rbx
ffffffffc0370349: 48 89 df mov %rbx,%rdi
ffffffffc037034c: e8 f7 21 00 00 callq 0xffffffffc0372548
There is the bpf2bpf call (at ffffffffc037034c) right after the tailcall
and jump target is not present. ctx is in %rbx register and BPF
subprogram that we will call into on ffffffffc037034c is relying on it,
e.g. it will pick ctx from there. Such code layout is therefore broken
as we would overwrite the content of %rbx with the value that was pushed
on the prologue. That is the reason for the 'bypass' approach.
Special care needs to be taken during the install/update/remove of
tailcall target. In case when target program is not present, the CPU
must not execute the pop instructions that precede the tailcall.
To address that, the following states can be defined:
A nop, unwind, nop
B nop, unwind, tail
C skip, unwind, nop
D skip, unwind, tail
A is forbidden (lead to incorrectness). The state transitions between
tailcall install/update/remove will work as follows:
First install tail call f: C->D->B(f)
* poke the tailcall, after that get rid of the skip
Update tail call f to f': B(f)->B(f')
* poke the tailcall (poke->tailcall_target) and do NOT touch the
poke->tailcall_bypass
Remove tail call: B(f')->C(f')
* poke->tailcall_bypass is poked back to jump, then we wait the RCU
grace period so that other programs will finish its execution and
after that we are safe to remove the poke->tailcall_target
Install new tail call (f''): C(f')->D(f'')->B(f'').
* same as first step
This way CPU can never be exposed to "unwind, tail" state.
Last but not least, when tailcalls get mixed with bpf2bpf calls, it
would be possible to encounter the endless loop due to clearing the
tailcall counter if for example we would use the tailcall3-like from BPF
selftests program that would be subprogram-based, meaning the tailcall
would be present within the BPF subprogram.
This test, broken down to particular steps, would do:
entry -> set tailcall counter to 0, bump it by 1, tailcall to func0
func0 -> call subprog_tail
(we are NOT skipping the first 11 bytes of prologue and this subprogram
has a tailcall, therefore we clear the counter...)
subprog -> do the same thing as entry
and then loop forever.
To address this, the idea is to go through the call chain of bpf2bpf progs
and look for a tailcall presence throughout whole chain. If we saw a single
tail call then each node in this call chain needs to be marked as a subprog
that can reach the tailcall. We would later feed the JIT with this info
and:
- set eax to 0 only when tailcall is reachable and this is the entry prog
- if tailcall is reachable but there's no tailcall in insns of currently
JITed prog then push rax anyway, so that it will be possible to
propagate further down the call chain
- finally if tailcall is reachable, then we need to precede the 'call'
insn with mov rax, [rbp - (stack_depth + 8)]
Tail call related cases from test_verifier kselftest are also working
fine. Sample BPF programs that utilize tail calls (sockex3, tracex5)
work properly as well.
[1]: https://lore.kernel.org/bpf/20200517043227.2gpq22ifoq37ogst@ast-mbp.dhcp.thefacebook.com/
Suggested-by: Alexei Starovoitov <ast@kernel.org>
Signed-off-by: Maciej Fijalkowski <maciej.fijalkowski@intel.com>
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
2020-09-16 21:10:08 +00:00
|
|
|
int tcc_off = -4 - round_up(stack_depth, 8);
|
2021-10-26 12:01:47 +00:00
|
|
|
u8 *prog = *pprog, *start = *pprog;
|
|
|
|
int offset;
|
bpf, x64: rework pro/epilogue and tailcall handling in JIT
This commit serves two things:
1) it optimizes BPF prologue/epilogue generation
2) it makes possible to have tailcalls within BPF subprogram
Both points are related to each other since without 1), 2) could not be
achieved.
In [1], Alexei says:
"The prologue will look like:
nop5
xor eax,eax // two new bytes if bpf_tail_call() is used in this
// function
push rbp
mov rbp, rsp
sub rsp, rounded_stack_depth
push rax // zero init tail_call counter
variable number of push rbx,r13,r14,r15
Then bpf_tail_call will pop variable number rbx,..
and final 'pop rax'
Then 'add rsp, size_of_current_stack_frame'
jmp to next function and skip over 'nop5; xor eax,eax; push rpb; mov
rbp, rsp'
This way new function will set its own stack size and will init tail
call
counter with whatever value the parent had.
If next function doesn't use bpf_tail_call it won't have 'xor eax,eax'.
Instead it would need to have 'nop2' in there."
Implement that suggestion.
Since the layout of stack is changed, tail call counter handling can not
rely anymore on popping it to rbx just like it have been handled for
constant prologue case and later overwrite of rbx with actual value of
rbx pushed to stack. Therefore, let's use one of the register (%rcx) that
is considered to be volatile/caller-saved and pop the value of tail call
counter in there in the epilogue.
Drop the BUILD_BUG_ON in emit_prologue and in
emit_bpf_tail_call_indirect where instruction layout is not constant
anymore.
Introduce new poke target, 'tailcall_bypass' to poke descriptor that is
dedicated for skipping the register pops and stack unwind that are
generated right before the actual jump to target program.
For case when the target program is not present, BPF program will skip
the pop instructions and nop5 dedicated for jmpq $target. An example of
such state when only R6 of callee saved registers is used by program:
ffffffffc0513aa1: e9 0e 00 00 00 jmpq 0xffffffffc0513ab4
ffffffffc0513aa6: 5b pop %rbx
ffffffffc0513aa7: 58 pop %rax
ffffffffc0513aa8: 48 81 c4 00 00 00 00 add $0x0,%rsp
ffffffffc0513aaf: 0f 1f 44 00 00 nopl 0x0(%rax,%rax,1)
ffffffffc0513ab4: 48 89 df mov %rbx,%rdi
When target program is inserted, the jump that was there to skip
pops/nop5 will become the nop5, so CPU will go over pops and do the
actual tailcall.
One might ask why there simply can not be pushes after the nop5?
In the following example snippet:
ffffffffc037030c: 48 89 fb mov %rdi,%rbx
(...)
ffffffffc0370332: 5b pop %rbx
ffffffffc0370333: 58 pop %rax
ffffffffc0370334: 48 81 c4 00 00 00 00 add $0x0,%rsp
ffffffffc037033b: 0f 1f 44 00 00 nopl 0x0(%rax,%rax,1)
ffffffffc0370340: 48 81 ec 00 00 00 00 sub $0x0,%rsp
ffffffffc0370347: 50 push %rax
ffffffffc0370348: 53 push %rbx
ffffffffc0370349: 48 89 df mov %rbx,%rdi
ffffffffc037034c: e8 f7 21 00 00 callq 0xffffffffc0372548
There is the bpf2bpf call (at ffffffffc037034c) right after the tailcall
and jump target is not present. ctx is in %rbx register and BPF
subprogram that we will call into on ffffffffc037034c is relying on it,
e.g. it will pick ctx from there. Such code layout is therefore broken
as we would overwrite the content of %rbx with the value that was pushed
on the prologue. That is the reason for the 'bypass' approach.
Special care needs to be taken during the install/update/remove of
tailcall target. In case when target program is not present, the CPU
must not execute the pop instructions that precede the tailcall.
To address that, the following states can be defined:
A nop, unwind, nop
B nop, unwind, tail
C skip, unwind, nop
D skip, unwind, tail
A is forbidden (lead to incorrectness). The state transitions between
tailcall install/update/remove will work as follows:
First install tail call f: C->D->B(f)
* poke the tailcall, after that get rid of the skip
Update tail call f to f': B(f)->B(f')
* poke the tailcall (poke->tailcall_target) and do NOT touch the
poke->tailcall_bypass
Remove tail call: B(f')->C(f')
* poke->tailcall_bypass is poked back to jump, then we wait the RCU
grace period so that other programs will finish its execution and
after that we are safe to remove the poke->tailcall_target
Install new tail call (f''): C(f')->D(f'')->B(f'').
* same as first step
This way CPU can never be exposed to "unwind, tail" state.
Last but not least, when tailcalls get mixed with bpf2bpf calls, it
would be possible to encounter the endless loop due to clearing the
tailcall counter if for example we would use the tailcall3-like from BPF
selftests program that would be subprogram-based, meaning the tailcall
would be present within the BPF subprogram.
This test, broken down to particular steps, would do:
entry -> set tailcall counter to 0, bump it by 1, tailcall to func0
func0 -> call subprog_tail
(we are NOT skipping the first 11 bytes of prologue and this subprogram
has a tailcall, therefore we clear the counter...)
subprog -> do the same thing as entry
and then loop forever.
To address this, the idea is to go through the call chain of bpf2bpf progs
and look for a tailcall presence throughout whole chain. If we saw a single
tail call then each node in this call chain needs to be marked as a subprog
that can reach the tailcall. We would later feed the JIT with this info
and:
- set eax to 0 only when tailcall is reachable and this is the entry prog
- if tailcall is reachable but there's no tailcall in insns of currently
JITed prog then push rax anyway, so that it will be possible to
propagate further down the call chain
- finally if tailcall is reachable, then we need to precede the 'call'
insn with mov rax, [rbp - (stack_depth + 8)]
Tail call related cases from test_verifier kselftest are also working
fine. Sample BPF programs that utilize tail calls (sockex3, tracex5)
work properly as well.
[1]: https://lore.kernel.org/bpf/20200517043227.2gpq22ifoq37ogst@ast-mbp.dhcp.thefacebook.com/
Suggested-by: Alexei Starovoitov <ast@kernel.org>
Signed-off-by: Maciej Fijalkowski <maciej.fijalkowski@intel.com>
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
2020-09-16 21:10:08 +00:00
|
|
|
|
bpf, x86: Emit patchable direct jump as tail call
Add initial code emission for *direct* jumps for tail call maps in
order to avoid the retpoline overhead from a493a87f38cf ("bpf, x64:
implement retpoline for tail call") for situations that allow for
it, meaning, for known constant keys at verification time which are
used as index into the tail call map. In case of Cilium which makes
heavy use of tail calls, constant keys are used in the vast majority,
only for a single occurrence we use a dynamic key.
High level outline is that if the target prog is NULL in the map, we
emit a 5-byte nop for the fall-through case and if not, we emit a
5-byte direct relative jmp to the target bpf_func + skipped prologue
offset. Later during runtime, we patch these 5-byte nop/jmps upon
tail call map update or deletions dynamically. Note that on x86-64
the direct jmp works as we reuse the same stack frame and skip
prologue (as opposed to some other JIT implementations).
One of the issues is that the tail call map slots can change at any
given time even during JITing. Therefore, we have two passes: i) emit
nops for all patchable locations during main JITing phase until we
declare prog->jited = 1 eventually. At this point the image is stable,
not public yet and with all jmps disabled. While JITing, we collect
additional info like poke->ip in order to remember the patch location
for later modifications. In ii) bpf_tail_call_direct_fixup() walks
over the progs poke_tab, locks the tail call maps poke_mutex to
prevent from parallel updates and patches in the right locations via
__bpf_arch_text_poke(). Note, the main bpf_arch_text_poke() cannot
be used at this point since we're not yet exposed to kallsyms. For
the update we use plain memcpy() since the image is not public and
still in read-write mode. After patching, we activate that poke entry
through poke->ip_stable. Meaning, at this point any tail call map
updates/deletions are not going to ignore that poke entry anymore.
Then, bpf_arch_text_poke() might still occur on the read-write image
until we finally locked it as read-only. Both modifications on the
given image are under text_mutex to avoid interference with each
other when update requests come in in parallel for different tail
call maps (current one we have locked in JIT and different one where
poke->ip_stable was already set).
Example prog:
# ./bpftool p d x i 1655
0: (b7) r3 = 0
1: (18) r2 = map[id:526]
3: (85) call bpf_tail_call#12
4: (b7) r0 = 1
5: (95) exit
Before:
# ./bpftool p d j i 1655
0xffffffffc076e55c:
0: nopl 0x0(%rax,%rax,1)
5: push %rbp
6: mov %rsp,%rbp
9: sub $0x200,%rsp
10: push %rbx
11: push %r13
13: push %r14
15: push %r15
17: pushq $0x0 _
19: xor %edx,%edx |_ index (arg 3)
1b: movabs $0xffff88d95cc82600,%rsi |_ map (arg 2)
25: mov %edx,%edx | index >= array->map.max_entries
27: cmp %edx,0x24(%rsi) |
2a: jbe 0x0000000000000066 |_
2c: mov -0x224(%rbp),%eax | tail call limit check
32: cmp $0x20,%eax |
35: ja 0x0000000000000066 |
37: add $0x1,%eax |
3a: mov %eax,-0x224(%rbp) |_
40: mov 0xd0(%rsi,%rdx,8),%rax |_ prog = array->ptrs[index]
48: test %rax,%rax | prog == NULL check
4b: je 0x0000000000000066 |_
4d: mov 0x30(%rax),%rax | goto *(prog->bpf_func + prologue_size)
51: add $0x19,%rax |
55: callq 0x0000000000000061 | retpoline for indirect jump
5a: pause |
5c: lfence |
5f: jmp 0x000000000000005a |
61: mov %rax,(%rsp) |
65: retq |_
66: mov $0x1,%eax
6b: pop %rbx
6c: pop %r15
6e: pop %r14
70: pop %r13
72: pop %rbx
73: leaveq
74: retq
After; state after JIT:
# ./bpftool p d j i 1655
0xffffffffc08e8930:
0: nopl 0x0(%rax,%rax,1)
5: push %rbp
6: mov %rsp,%rbp
9: sub $0x200,%rsp
10: push %rbx
11: push %r13
13: push %r14
15: push %r15
17: pushq $0x0 _
19: xor %edx,%edx |_ index (arg 3)
1b: movabs $0xffff9d8afd74c000,%rsi |_ map (arg 2)
25: mov -0x224(%rbp),%eax | tail call limit check
2b: cmp $0x20,%eax |
2e: ja 0x000000000000003e |
30: add $0x1,%eax |
33: mov %eax,-0x224(%rbp) |_
39: jmpq 0xfffffffffffd1785 |_ [direct] goto *(prog->bpf_func + prologue_size)
3e: mov $0x1,%eax
43: pop %rbx
44: pop %r15
46: pop %r14
48: pop %r13
4a: pop %rbx
4b: leaveq
4c: retq
After; state after map update (target prog):
# ./bpftool p d j i 1655
0xffffffffc08e8930:
0: nopl 0x0(%rax,%rax,1)
5: push %rbp
6: mov %rsp,%rbp
9: sub $0x200,%rsp
10: push %rbx
11: push %r13
13: push %r14
15: push %r15
17: pushq $0x0
19: xor %edx,%edx
1b: movabs $0xffff9d8afd74c000,%rsi
25: mov -0x224(%rbp),%eax
2b: cmp $0x20,%eax .
2e: ja 0x000000000000003e .
30: add $0x1,%eax .
33: mov %eax,-0x224(%rbp) |_
39: jmpq 0xffffffffffb09f55 |_ goto *(prog->bpf_func + prologue_size)
3e: mov $0x1,%eax
43: pop %rbx
44: pop %r15
46: pop %r14
48: pop %r13
4a: pop %rbx
4b: leaveq
4c: retq
After; state after map update (no prog):
# ./bpftool p d j i 1655
0xffffffffc08e8930:
0: nopl 0x0(%rax,%rax,1)
5: push %rbp
6: mov %rsp,%rbp
9: sub $0x200,%rsp
10: push %rbx
11: push %r13
13: push %r14
15: push %r15
17: pushq $0x0
19: xor %edx,%edx
1b: movabs $0xffff9d8afd74c000,%rsi
25: mov -0x224(%rbp),%eax
2b: cmp $0x20,%eax .
2e: ja 0x000000000000003e .
30: add $0x1,%eax .
33: mov %eax,-0x224(%rbp) |_
39: nopl 0x0(%rax,%rax,1) |_ fall-through nop
3e: mov $0x1,%eax
43: pop %rbx
44: pop %r15
46: pop %r14
48: pop %r13
4a: pop %rbx
4b: leaveq
4c: retq
Nice bonus is that this also shrinks the code emission quite a bit
for every tail call invocation.
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Link: https://lore.kernel.org/bpf/6ada4c1c9d35eeb5f4ecfab94593dafa6b5c4b09.1574452833.git.daniel@iogearbox.net
2019-11-22 20:08:00 +00:00
|
|
|
/*
|
bpf: Change value of MAX_TAIL_CALL_CNT from 32 to 33
In the current code, the actual max tail call count is 33 which is greater
than MAX_TAIL_CALL_CNT (defined as 32). The actual limit is not consistent
with the meaning of MAX_TAIL_CALL_CNT and thus confusing at first glance.
We can see the historical evolution from commit 04fd61ab36ec ("bpf: allow
bpf programs to tail-call other bpf programs") and commit f9dabe016b63
("bpf: Undo off-by-one in interpreter tail call count limit"). In order
to avoid changing existing behavior, the actual limit is 33 now, this is
reasonable.
After commit 874be05f525e ("bpf, tests: Add tail call test suite"), we can
see there exists failed testcase.
On all archs when CONFIG_BPF_JIT_ALWAYS_ON is not set:
# echo 0 > /proc/sys/net/core/bpf_jit_enable
# modprobe test_bpf
# dmesg | grep -w FAIL
Tail call error path, max count reached jited:0 ret 34 != 33 FAIL
On some archs:
# echo 1 > /proc/sys/net/core/bpf_jit_enable
# modprobe test_bpf
# dmesg | grep -w FAIL
Tail call error path, max count reached jited:1 ret 34 != 33 FAIL
Although the above failed testcase has been fixed in commit 18935a72eb25
("bpf/tests: Fix error in tail call limit tests"), it would still be good
to change the value of MAX_TAIL_CALL_CNT from 32 to 33 to make the code
more readable.
The 32-bit x86 JIT was using a limit of 32, just fix the wrong comments and
limit to 33 tail calls as the constant MAX_TAIL_CALL_CNT updated. For the
mips64 JIT, use "ori" instead of "addiu" as suggested by Johan Almbladh.
For the riscv JIT, use RV_REG_TCC directly to save one register move as
suggested by Björn Töpel. For the other implementations, no function changes,
it does not change the current limit 33, the new value of MAX_TAIL_CALL_CNT
can reflect the actual max tail call count, the related tail call testcases
in test_bpf module and selftests can work well for the interpreter and the
JIT.
Here are the test results on x86_64:
# uname -m
x86_64
# echo 0 > /proc/sys/net/core/bpf_jit_enable
# modprobe test_bpf test_suite=test_tail_calls
# dmesg | tail -1
test_bpf: test_tail_calls: Summary: 8 PASSED, 0 FAILED, [0/8 JIT'ed]
# rmmod test_bpf
# echo 1 > /proc/sys/net/core/bpf_jit_enable
# modprobe test_bpf test_suite=test_tail_calls
# dmesg | tail -1
test_bpf: test_tail_calls: Summary: 8 PASSED, 0 FAILED, [8/8 JIT'ed]
# rmmod test_bpf
# ./test_progs -t tailcalls
#142 tailcalls:OK
Summary: 1/11 PASSED, 0 SKIPPED, 0 FAILED
Signed-off-by: Tiezhu Yang <yangtiezhu@loongson.cn>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Tested-by: Johan Almbladh <johan.almbladh@anyfinetworks.com>
Tested-by: Ilya Leoshkevich <iii@linux.ibm.com>
Acked-by: Björn Töpel <bjorn@kernel.org>
Acked-by: Johan Almbladh <johan.almbladh@anyfinetworks.com>
Acked-by: Ilya Leoshkevich <iii@linux.ibm.com>
Link: https://lore.kernel.org/bpf/1636075800-3264-1-git-send-email-yangtiezhu@loongson.cn
2021-11-05 01:30:00 +00:00
|
|
|
* if (tail_call_cnt++ >= MAX_TAIL_CALL_CNT)
|
bpf, x86: Emit patchable direct jump as tail call
Add initial code emission for *direct* jumps for tail call maps in
order to avoid the retpoline overhead from a493a87f38cf ("bpf, x64:
implement retpoline for tail call") for situations that allow for
it, meaning, for known constant keys at verification time which are
used as index into the tail call map. In case of Cilium which makes
heavy use of tail calls, constant keys are used in the vast majority,
only for a single occurrence we use a dynamic key.
High level outline is that if the target prog is NULL in the map, we
emit a 5-byte nop for the fall-through case and if not, we emit a
5-byte direct relative jmp to the target bpf_func + skipped prologue
offset. Later during runtime, we patch these 5-byte nop/jmps upon
tail call map update or deletions dynamically. Note that on x86-64
the direct jmp works as we reuse the same stack frame and skip
prologue (as opposed to some other JIT implementations).
One of the issues is that the tail call map slots can change at any
given time even during JITing. Therefore, we have two passes: i) emit
nops for all patchable locations during main JITing phase until we
declare prog->jited = 1 eventually. At this point the image is stable,
not public yet and with all jmps disabled. While JITing, we collect
additional info like poke->ip in order to remember the patch location
for later modifications. In ii) bpf_tail_call_direct_fixup() walks
over the progs poke_tab, locks the tail call maps poke_mutex to
prevent from parallel updates and patches in the right locations via
__bpf_arch_text_poke(). Note, the main bpf_arch_text_poke() cannot
be used at this point since we're not yet exposed to kallsyms. For
the update we use plain memcpy() since the image is not public and
still in read-write mode. After patching, we activate that poke entry
through poke->ip_stable. Meaning, at this point any tail call map
updates/deletions are not going to ignore that poke entry anymore.
Then, bpf_arch_text_poke() might still occur on the read-write image
until we finally locked it as read-only. Both modifications on the
given image are under text_mutex to avoid interference with each
other when update requests come in in parallel for different tail
call maps (current one we have locked in JIT and different one where
poke->ip_stable was already set).
Example prog:
# ./bpftool p d x i 1655
0: (b7) r3 = 0
1: (18) r2 = map[id:526]
3: (85) call bpf_tail_call#12
4: (b7) r0 = 1
5: (95) exit
Before:
# ./bpftool p d j i 1655
0xffffffffc076e55c:
0: nopl 0x0(%rax,%rax,1)
5: push %rbp
6: mov %rsp,%rbp
9: sub $0x200,%rsp
10: push %rbx
11: push %r13
13: push %r14
15: push %r15
17: pushq $0x0 _
19: xor %edx,%edx |_ index (arg 3)
1b: movabs $0xffff88d95cc82600,%rsi |_ map (arg 2)
25: mov %edx,%edx | index >= array->map.max_entries
27: cmp %edx,0x24(%rsi) |
2a: jbe 0x0000000000000066 |_
2c: mov -0x224(%rbp),%eax | tail call limit check
32: cmp $0x20,%eax |
35: ja 0x0000000000000066 |
37: add $0x1,%eax |
3a: mov %eax,-0x224(%rbp) |_
40: mov 0xd0(%rsi,%rdx,8),%rax |_ prog = array->ptrs[index]
48: test %rax,%rax | prog == NULL check
4b: je 0x0000000000000066 |_
4d: mov 0x30(%rax),%rax | goto *(prog->bpf_func + prologue_size)
51: add $0x19,%rax |
55: callq 0x0000000000000061 | retpoline for indirect jump
5a: pause |
5c: lfence |
5f: jmp 0x000000000000005a |
61: mov %rax,(%rsp) |
65: retq |_
66: mov $0x1,%eax
6b: pop %rbx
6c: pop %r15
6e: pop %r14
70: pop %r13
72: pop %rbx
73: leaveq
74: retq
After; state after JIT:
# ./bpftool p d j i 1655
0xffffffffc08e8930:
0: nopl 0x0(%rax,%rax,1)
5: push %rbp
6: mov %rsp,%rbp
9: sub $0x200,%rsp
10: push %rbx
11: push %r13
13: push %r14
15: push %r15
17: pushq $0x0 _
19: xor %edx,%edx |_ index (arg 3)
1b: movabs $0xffff9d8afd74c000,%rsi |_ map (arg 2)
25: mov -0x224(%rbp),%eax | tail call limit check
2b: cmp $0x20,%eax |
2e: ja 0x000000000000003e |
30: add $0x1,%eax |
33: mov %eax,-0x224(%rbp) |_
39: jmpq 0xfffffffffffd1785 |_ [direct] goto *(prog->bpf_func + prologue_size)
3e: mov $0x1,%eax
43: pop %rbx
44: pop %r15
46: pop %r14
48: pop %r13
4a: pop %rbx
4b: leaveq
4c: retq
After; state after map update (target prog):
# ./bpftool p d j i 1655
0xffffffffc08e8930:
0: nopl 0x0(%rax,%rax,1)
5: push %rbp
6: mov %rsp,%rbp
9: sub $0x200,%rsp
10: push %rbx
11: push %r13
13: push %r14
15: push %r15
17: pushq $0x0
19: xor %edx,%edx
1b: movabs $0xffff9d8afd74c000,%rsi
25: mov -0x224(%rbp),%eax
2b: cmp $0x20,%eax .
2e: ja 0x000000000000003e .
30: add $0x1,%eax .
33: mov %eax,-0x224(%rbp) |_
39: jmpq 0xffffffffffb09f55 |_ goto *(prog->bpf_func + prologue_size)
3e: mov $0x1,%eax
43: pop %rbx
44: pop %r15
46: pop %r14
48: pop %r13
4a: pop %rbx
4b: leaveq
4c: retq
After; state after map update (no prog):
# ./bpftool p d j i 1655
0xffffffffc08e8930:
0: nopl 0x0(%rax,%rax,1)
5: push %rbp
6: mov %rsp,%rbp
9: sub $0x200,%rsp
10: push %rbx
11: push %r13
13: push %r14
15: push %r15
17: pushq $0x0
19: xor %edx,%edx
1b: movabs $0xffff9d8afd74c000,%rsi
25: mov -0x224(%rbp),%eax
2b: cmp $0x20,%eax .
2e: ja 0x000000000000003e .
30: add $0x1,%eax .
33: mov %eax,-0x224(%rbp) |_
39: nopl 0x0(%rax,%rax,1) |_ fall-through nop
3e: mov $0x1,%eax
43: pop %rbx
44: pop %r15
46: pop %r14
48: pop %r13
4a: pop %rbx
4b: leaveq
4c: retq
Nice bonus is that this also shrinks the code emission quite a bit
for every tail call invocation.
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Link: https://lore.kernel.org/bpf/6ada4c1c9d35eeb5f4ecfab94593dafa6b5c4b09.1574452833.git.daniel@iogearbox.net
2019-11-22 20:08:00 +00:00
|
|
|
* goto out;
|
|
|
|
*/
|
bpf, x64: rework pro/epilogue and tailcall handling in JIT
This commit serves two things:
1) it optimizes BPF prologue/epilogue generation
2) it makes possible to have tailcalls within BPF subprogram
Both points are related to each other since without 1), 2) could not be
achieved.
In [1], Alexei says:
"The prologue will look like:
nop5
xor eax,eax // two new bytes if bpf_tail_call() is used in this
// function
push rbp
mov rbp, rsp
sub rsp, rounded_stack_depth
push rax // zero init tail_call counter
variable number of push rbx,r13,r14,r15
Then bpf_tail_call will pop variable number rbx,..
and final 'pop rax'
Then 'add rsp, size_of_current_stack_frame'
jmp to next function and skip over 'nop5; xor eax,eax; push rpb; mov
rbp, rsp'
This way new function will set its own stack size and will init tail
call
counter with whatever value the parent had.
If next function doesn't use bpf_tail_call it won't have 'xor eax,eax'.
Instead it would need to have 'nop2' in there."
Implement that suggestion.
Since the layout of stack is changed, tail call counter handling can not
rely anymore on popping it to rbx just like it have been handled for
constant prologue case and later overwrite of rbx with actual value of
rbx pushed to stack. Therefore, let's use one of the register (%rcx) that
is considered to be volatile/caller-saved and pop the value of tail call
counter in there in the epilogue.
Drop the BUILD_BUG_ON in emit_prologue and in
emit_bpf_tail_call_indirect where instruction layout is not constant
anymore.
Introduce new poke target, 'tailcall_bypass' to poke descriptor that is
dedicated for skipping the register pops and stack unwind that are
generated right before the actual jump to target program.
For case when the target program is not present, BPF program will skip
the pop instructions and nop5 dedicated for jmpq $target. An example of
such state when only R6 of callee saved registers is used by program:
ffffffffc0513aa1: e9 0e 00 00 00 jmpq 0xffffffffc0513ab4
ffffffffc0513aa6: 5b pop %rbx
ffffffffc0513aa7: 58 pop %rax
ffffffffc0513aa8: 48 81 c4 00 00 00 00 add $0x0,%rsp
ffffffffc0513aaf: 0f 1f 44 00 00 nopl 0x0(%rax,%rax,1)
ffffffffc0513ab4: 48 89 df mov %rbx,%rdi
When target program is inserted, the jump that was there to skip
pops/nop5 will become the nop5, so CPU will go over pops and do the
actual tailcall.
One might ask why there simply can not be pushes after the nop5?
In the following example snippet:
ffffffffc037030c: 48 89 fb mov %rdi,%rbx
(...)
ffffffffc0370332: 5b pop %rbx
ffffffffc0370333: 58 pop %rax
ffffffffc0370334: 48 81 c4 00 00 00 00 add $0x0,%rsp
ffffffffc037033b: 0f 1f 44 00 00 nopl 0x0(%rax,%rax,1)
ffffffffc0370340: 48 81 ec 00 00 00 00 sub $0x0,%rsp
ffffffffc0370347: 50 push %rax
ffffffffc0370348: 53 push %rbx
ffffffffc0370349: 48 89 df mov %rbx,%rdi
ffffffffc037034c: e8 f7 21 00 00 callq 0xffffffffc0372548
There is the bpf2bpf call (at ffffffffc037034c) right after the tailcall
and jump target is not present. ctx is in %rbx register and BPF
subprogram that we will call into on ffffffffc037034c is relying on it,
e.g. it will pick ctx from there. Such code layout is therefore broken
as we would overwrite the content of %rbx with the value that was pushed
on the prologue. That is the reason for the 'bypass' approach.
Special care needs to be taken during the install/update/remove of
tailcall target. In case when target program is not present, the CPU
must not execute the pop instructions that precede the tailcall.
To address that, the following states can be defined:
A nop, unwind, nop
B nop, unwind, tail
C skip, unwind, nop
D skip, unwind, tail
A is forbidden (lead to incorrectness). The state transitions between
tailcall install/update/remove will work as follows:
First install tail call f: C->D->B(f)
* poke the tailcall, after that get rid of the skip
Update tail call f to f': B(f)->B(f')
* poke the tailcall (poke->tailcall_target) and do NOT touch the
poke->tailcall_bypass
Remove tail call: B(f')->C(f')
* poke->tailcall_bypass is poked back to jump, then we wait the RCU
grace period so that other programs will finish its execution and
after that we are safe to remove the poke->tailcall_target
Install new tail call (f''): C(f')->D(f'')->B(f'').
* same as first step
This way CPU can never be exposed to "unwind, tail" state.
Last but not least, when tailcalls get mixed with bpf2bpf calls, it
would be possible to encounter the endless loop due to clearing the
tailcall counter if for example we would use the tailcall3-like from BPF
selftests program that would be subprogram-based, meaning the tailcall
would be present within the BPF subprogram.
This test, broken down to particular steps, would do:
entry -> set tailcall counter to 0, bump it by 1, tailcall to func0
func0 -> call subprog_tail
(we are NOT skipping the first 11 bytes of prologue and this subprogram
has a tailcall, therefore we clear the counter...)
subprog -> do the same thing as entry
and then loop forever.
To address this, the idea is to go through the call chain of bpf2bpf progs
and look for a tailcall presence throughout whole chain. If we saw a single
tail call then each node in this call chain needs to be marked as a subprog
that can reach the tailcall. We would later feed the JIT with this info
and:
- set eax to 0 only when tailcall is reachable and this is the entry prog
- if tailcall is reachable but there's no tailcall in insns of currently
JITed prog then push rax anyway, so that it will be possible to
propagate further down the call chain
- finally if tailcall is reachable, then we need to precede the 'call'
insn with mov rax, [rbp - (stack_depth + 8)]
Tail call related cases from test_verifier kselftest are also working
fine. Sample BPF programs that utilize tail calls (sockex3, tracex5)
work properly as well.
[1]: https://lore.kernel.org/bpf/20200517043227.2gpq22ifoq37ogst@ast-mbp.dhcp.thefacebook.com/
Suggested-by: Alexei Starovoitov <ast@kernel.org>
Signed-off-by: Maciej Fijalkowski <maciej.fijalkowski@intel.com>
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
2020-09-16 21:10:08 +00:00
|
|
|
EMIT2_off32(0x8B, 0x85, tcc_off); /* mov eax, dword ptr [rbp - tcc_off] */
|
bpf, x86: Emit patchable direct jump as tail call
Add initial code emission for *direct* jumps for tail call maps in
order to avoid the retpoline overhead from a493a87f38cf ("bpf, x64:
implement retpoline for tail call") for situations that allow for
it, meaning, for known constant keys at verification time which are
used as index into the tail call map. In case of Cilium which makes
heavy use of tail calls, constant keys are used in the vast majority,
only for a single occurrence we use a dynamic key.
High level outline is that if the target prog is NULL in the map, we
emit a 5-byte nop for the fall-through case and if not, we emit a
5-byte direct relative jmp to the target bpf_func + skipped prologue
offset. Later during runtime, we patch these 5-byte nop/jmps upon
tail call map update or deletions dynamically. Note that on x86-64
the direct jmp works as we reuse the same stack frame and skip
prologue (as opposed to some other JIT implementations).
One of the issues is that the tail call map slots can change at any
given time even during JITing. Therefore, we have two passes: i) emit
nops for all patchable locations during main JITing phase until we
declare prog->jited = 1 eventually. At this point the image is stable,
not public yet and with all jmps disabled. While JITing, we collect
additional info like poke->ip in order to remember the patch location
for later modifications. In ii) bpf_tail_call_direct_fixup() walks
over the progs poke_tab, locks the tail call maps poke_mutex to
prevent from parallel updates and patches in the right locations via
__bpf_arch_text_poke(). Note, the main bpf_arch_text_poke() cannot
be used at this point since we're not yet exposed to kallsyms. For
the update we use plain memcpy() since the image is not public and
still in read-write mode. After patching, we activate that poke entry
through poke->ip_stable. Meaning, at this point any tail call map
updates/deletions are not going to ignore that poke entry anymore.
Then, bpf_arch_text_poke() might still occur on the read-write image
until we finally locked it as read-only. Both modifications on the
given image are under text_mutex to avoid interference with each
other when update requests come in in parallel for different tail
call maps (current one we have locked in JIT and different one where
poke->ip_stable was already set).
Example prog:
# ./bpftool p d x i 1655
0: (b7) r3 = 0
1: (18) r2 = map[id:526]
3: (85) call bpf_tail_call#12
4: (b7) r0 = 1
5: (95) exit
Before:
# ./bpftool p d j i 1655
0xffffffffc076e55c:
0: nopl 0x0(%rax,%rax,1)
5: push %rbp
6: mov %rsp,%rbp
9: sub $0x200,%rsp
10: push %rbx
11: push %r13
13: push %r14
15: push %r15
17: pushq $0x0 _
19: xor %edx,%edx |_ index (arg 3)
1b: movabs $0xffff88d95cc82600,%rsi |_ map (arg 2)
25: mov %edx,%edx | index >= array->map.max_entries
27: cmp %edx,0x24(%rsi) |
2a: jbe 0x0000000000000066 |_
2c: mov -0x224(%rbp),%eax | tail call limit check
32: cmp $0x20,%eax |
35: ja 0x0000000000000066 |
37: add $0x1,%eax |
3a: mov %eax,-0x224(%rbp) |_
40: mov 0xd0(%rsi,%rdx,8),%rax |_ prog = array->ptrs[index]
48: test %rax,%rax | prog == NULL check
4b: je 0x0000000000000066 |_
4d: mov 0x30(%rax),%rax | goto *(prog->bpf_func + prologue_size)
51: add $0x19,%rax |
55: callq 0x0000000000000061 | retpoline for indirect jump
5a: pause |
5c: lfence |
5f: jmp 0x000000000000005a |
61: mov %rax,(%rsp) |
65: retq |_
66: mov $0x1,%eax
6b: pop %rbx
6c: pop %r15
6e: pop %r14
70: pop %r13
72: pop %rbx
73: leaveq
74: retq
After; state after JIT:
# ./bpftool p d j i 1655
0xffffffffc08e8930:
0: nopl 0x0(%rax,%rax,1)
5: push %rbp
6: mov %rsp,%rbp
9: sub $0x200,%rsp
10: push %rbx
11: push %r13
13: push %r14
15: push %r15
17: pushq $0x0 _
19: xor %edx,%edx |_ index (arg 3)
1b: movabs $0xffff9d8afd74c000,%rsi |_ map (arg 2)
25: mov -0x224(%rbp),%eax | tail call limit check
2b: cmp $0x20,%eax |
2e: ja 0x000000000000003e |
30: add $0x1,%eax |
33: mov %eax,-0x224(%rbp) |_
39: jmpq 0xfffffffffffd1785 |_ [direct] goto *(prog->bpf_func + prologue_size)
3e: mov $0x1,%eax
43: pop %rbx
44: pop %r15
46: pop %r14
48: pop %r13
4a: pop %rbx
4b: leaveq
4c: retq
After; state after map update (target prog):
# ./bpftool p d j i 1655
0xffffffffc08e8930:
0: nopl 0x0(%rax,%rax,1)
5: push %rbp
6: mov %rsp,%rbp
9: sub $0x200,%rsp
10: push %rbx
11: push %r13
13: push %r14
15: push %r15
17: pushq $0x0
19: xor %edx,%edx
1b: movabs $0xffff9d8afd74c000,%rsi
25: mov -0x224(%rbp),%eax
2b: cmp $0x20,%eax .
2e: ja 0x000000000000003e .
30: add $0x1,%eax .
33: mov %eax,-0x224(%rbp) |_
39: jmpq 0xffffffffffb09f55 |_ goto *(prog->bpf_func + prologue_size)
3e: mov $0x1,%eax
43: pop %rbx
44: pop %r15
46: pop %r14
48: pop %r13
4a: pop %rbx
4b: leaveq
4c: retq
After; state after map update (no prog):
# ./bpftool p d j i 1655
0xffffffffc08e8930:
0: nopl 0x0(%rax,%rax,1)
5: push %rbp
6: mov %rsp,%rbp
9: sub $0x200,%rsp
10: push %rbx
11: push %r13
13: push %r14
15: push %r15
17: pushq $0x0
19: xor %edx,%edx
1b: movabs $0xffff9d8afd74c000,%rsi
25: mov -0x224(%rbp),%eax
2b: cmp $0x20,%eax .
2e: ja 0x000000000000003e .
30: add $0x1,%eax .
33: mov %eax,-0x224(%rbp) |_
39: nopl 0x0(%rax,%rax,1) |_ fall-through nop
3e: mov $0x1,%eax
43: pop %rbx
44: pop %r15
46: pop %r14
48: pop %r13
4a: pop %rbx
4b: leaveq
4c: retq
Nice bonus is that this also shrinks the code emission quite a bit
for every tail call invocation.
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Link: https://lore.kernel.org/bpf/6ada4c1c9d35eeb5f4ecfab94593dafa6b5c4b09.1574452833.git.daniel@iogearbox.net
2019-11-22 20:08:00 +00:00
|
|
|
EMIT3(0x83, 0xF8, MAX_TAIL_CALL_CNT); /* cmp eax, MAX_TAIL_CALL_CNT */
|
2021-10-26 12:01:47 +00:00
|
|
|
|
|
|
|
offset = ctx->tail_call_direct_label - (prog + 2 - start);
|
bpf: Change value of MAX_TAIL_CALL_CNT from 32 to 33
In the current code, the actual max tail call count is 33 which is greater
than MAX_TAIL_CALL_CNT (defined as 32). The actual limit is not consistent
with the meaning of MAX_TAIL_CALL_CNT and thus confusing at first glance.
We can see the historical evolution from commit 04fd61ab36ec ("bpf: allow
bpf programs to tail-call other bpf programs") and commit f9dabe016b63
("bpf: Undo off-by-one in interpreter tail call count limit"). In order
to avoid changing existing behavior, the actual limit is 33 now, this is
reasonable.
After commit 874be05f525e ("bpf, tests: Add tail call test suite"), we can
see there exists failed testcase.
On all archs when CONFIG_BPF_JIT_ALWAYS_ON is not set:
# echo 0 > /proc/sys/net/core/bpf_jit_enable
# modprobe test_bpf
# dmesg | grep -w FAIL
Tail call error path, max count reached jited:0 ret 34 != 33 FAIL
On some archs:
# echo 1 > /proc/sys/net/core/bpf_jit_enable
# modprobe test_bpf
# dmesg | grep -w FAIL
Tail call error path, max count reached jited:1 ret 34 != 33 FAIL
Although the above failed testcase has been fixed in commit 18935a72eb25
("bpf/tests: Fix error in tail call limit tests"), it would still be good
to change the value of MAX_TAIL_CALL_CNT from 32 to 33 to make the code
more readable.
The 32-bit x86 JIT was using a limit of 32, just fix the wrong comments and
limit to 33 tail calls as the constant MAX_TAIL_CALL_CNT updated. For the
mips64 JIT, use "ori" instead of "addiu" as suggested by Johan Almbladh.
For the riscv JIT, use RV_REG_TCC directly to save one register move as
suggested by Björn Töpel. For the other implementations, no function changes,
it does not change the current limit 33, the new value of MAX_TAIL_CALL_CNT
can reflect the actual max tail call count, the related tail call testcases
in test_bpf module and selftests can work well for the interpreter and the
JIT.
Here are the test results on x86_64:
# uname -m
x86_64
# echo 0 > /proc/sys/net/core/bpf_jit_enable
# modprobe test_bpf test_suite=test_tail_calls
# dmesg | tail -1
test_bpf: test_tail_calls: Summary: 8 PASSED, 0 FAILED, [0/8 JIT'ed]
# rmmod test_bpf
# echo 1 > /proc/sys/net/core/bpf_jit_enable
# modprobe test_bpf test_suite=test_tail_calls
# dmesg | tail -1
test_bpf: test_tail_calls: Summary: 8 PASSED, 0 FAILED, [8/8 JIT'ed]
# rmmod test_bpf
# ./test_progs -t tailcalls
#142 tailcalls:OK
Summary: 1/11 PASSED, 0 SKIPPED, 0 FAILED
Signed-off-by: Tiezhu Yang <yangtiezhu@loongson.cn>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Tested-by: Johan Almbladh <johan.almbladh@anyfinetworks.com>
Tested-by: Ilya Leoshkevich <iii@linux.ibm.com>
Acked-by: Björn Töpel <bjorn@kernel.org>
Acked-by: Johan Almbladh <johan.almbladh@anyfinetworks.com>
Acked-by: Ilya Leoshkevich <iii@linux.ibm.com>
Link: https://lore.kernel.org/bpf/1636075800-3264-1-git-send-email-yangtiezhu@loongson.cn
2021-11-05 01:30:00 +00:00
|
|
|
EMIT2(X86_JAE, offset); /* jae out */
|
bpf, x86: Emit patchable direct jump as tail call
Add initial code emission for *direct* jumps for tail call maps in
order to avoid the retpoline overhead from a493a87f38cf ("bpf, x64:
implement retpoline for tail call") for situations that allow for
it, meaning, for known constant keys at verification time which are
used as index into the tail call map. In case of Cilium which makes
heavy use of tail calls, constant keys are used in the vast majority,
only for a single occurrence we use a dynamic key.
High level outline is that if the target prog is NULL in the map, we
emit a 5-byte nop for the fall-through case and if not, we emit a
5-byte direct relative jmp to the target bpf_func + skipped prologue
offset. Later during runtime, we patch these 5-byte nop/jmps upon
tail call map update or deletions dynamically. Note that on x86-64
the direct jmp works as we reuse the same stack frame and skip
prologue (as opposed to some other JIT implementations).
One of the issues is that the tail call map slots can change at any
given time even during JITing. Therefore, we have two passes: i) emit
nops for all patchable locations during main JITing phase until we
declare prog->jited = 1 eventually. At this point the image is stable,
not public yet and with all jmps disabled. While JITing, we collect
additional info like poke->ip in order to remember the patch location
for later modifications. In ii) bpf_tail_call_direct_fixup() walks
over the progs poke_tab, locks the tail call maps poke_mutex to
prevent from parallel updates and patches in the right locations via
__bpf_arch_text_poke(). Note, the main bpf_arch_text_poke() cannot
be used at this point since we're not yet exposed to kallsyms. For
the update we use plain memcpy() since the image is not public and
still in read-write mode. After patching, we activate that poke entry
through poke->ip_stable. Meaning, at this point any tail call map
updates/deletions are not going to ignore that poke entry anymore.
Then, bpf_arch_text_poke() might still occur on the read-write image
until we finally locked it as read-only. Both modifications on the
given image are under text_mutex to avoid interference with each
other when update requests come in in parallel for different tail
call maps (current one we have locked in JIT and different one where
poke->ip_stable was already set).
Example prog:
# ./bpftool p d x i 1655
0: (b7) r3 = 0
1: (18) r2 = map[id:526]
3: (85) call bpf_tail_call#12
4: (b7) r0 = 1
5: (95) exit
Before:
# ./bpftool p d j i 1655
0xffffffffc076e55c:
0: nopl 0x0(%rax,%rax,1)
5: push %rbp
6: mov %rsp,%rbp
9: sub $0x200,%rsp
10: push %rbx
11: push %r13
13: push %r14
15: push %r15
17: pushq $0x0 _
19: xor %edx,%edx |_ index (arg 3)
1b: movabs $0xffff88d95cc82600,%rsi |_ map (arg 2)
25: mov %edx,%edx | index >= array->map.max_entries
27: cmp %edx,0x24(%rsi) |
2a: jbe 0x0000000000000066 |_
2c: mov -0x224(%rbp),%eax | tail call limit check
32: cmp $0x20,%eax |
35: ja 0x0000000000000066 |
37: add $0x1,%eax |
3a: mov %eax,-0x224(%rbp) |_
40: mov 0xd0(%rsi,%rdx,8),%rax |_ prog = array->ptrs[index]
48: test %rax,%rax | prog == NULL check
4b: je 0x0000000000000066 |_
4d: mov 0x30(%rax),%rax | goto *(prog->bpf_func + prologue_size)
51: add $0x19,%rax |
55: callq 0x0000000000000061 | retpoline for indirect jump
5a: pause |
5c: lfence |
5f: jmp 0x000000000000005a |
61: mov %rax,(%rsp) |
65: retq |_
66: mov $0x1,%eax
6b: pop %rbx
6c: pop %r15
6e: pop %r14
70: pop %r13
72: pop %rbx
73: leaveq
74: retq
After; state after JIT:
# ./bpftool p d j i 1655
0xffffffffc08e8930:
0: nopl 0x0(%rax,%rax,1)
5: push %rbp
6: mov %rsp,%rbp
9: sub $0x200,%rsp
10: push %rbx
11: push %r13
13: push %r14
15: push %r15
17: pushq $0x0 _
19: xor %edx,%edx |_ index (arg 3)
1b: movabs $0xffff9d8afd74c000,%rsi |_ map (arg 2)
25: mov -0x224(%rbp),%eax | tail call limit check
2b: cmp $0x20,%eax |
2e: ja 0x000000000000003e |
30: add $0x1,%eax |
33: mov %eax,-0x224(%rbp) |_
39: jmpq 0xfffffffffffd1785 |_ [direct] goto *(prog->bpf_func + prologue_size)
3e: mov $0x1,%eax
43: pop %rbx
44: pop %r15
46: pop %r14
48: pop %r13
4a: pop %rbx
4b: leaveq
4c: retq
After; state after map update (target prog):
# ./bpftool p d j i 1655
0xffffffffc08e8930:
0: nopl 0x0(%rax,%rax,1)
5: push %rbp
6: mov %rsp,%rbp
9: sub $0x200,%rsp
10: push %rbx
11: push %r13
13: push %r14
15: push %r15
17: pushq $0x0
19: xor %edx,%edx
1b: movabs $0xffff9d8afd74c000,%rsi
25: mov -0x224(%rbp),%eax
2b: cmp $0x20,%eax .
2e: ja 0x000000000000003e .
30: add $0x1,%eax .
33: mov %eax,-0x224(%rbp) |_
39: jmpq 0xffffffffffb09f55 |_ goto *(prog->bpf_func + prologue_size)
3e: mov $0x1,%eax
43: pop %rbx
44: pop %r15
46: pop %r14
48: pop %r13
4a: pop %rbx
4b: leaveq
4c: retq
After; state after map update (no prog):
# ./bpftool p d j i 1655
0xffffffffc08e8930:
0: nopl 0x0(%rax,%rax,1)
5: push %rbp
6: mov %rsp,%rbp
9: sub $0x200,%rsp
10: push %rbx
11: push %r13
13: push %r14
15: push %r15
17: pushq $0x0
19: xor %edx,%edx
1b: movabs $0xffff9d8afd74c000,%rsi
25: mov -0x224(%rbp),%eax
2b: cmp $0x20,%eax .
2e: ja 0x000000000000003e .
30: add $0x1,%eax .
33: mov %eax,-0x224(%rbp) |_
39: nopl 0x0(%rax,%rax,1) |_ fall-through nop
3e: mov $0x1,%eax
43: pop %rbx
44: pop %r15
46: pop %r14
48: pop %r13
4a: pop %rbx
4b: leaveq
4c: retq
Nice bonus is that this also shrinks the code emission quite a bit
for every tail call invocation.
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Link: https://lore.kernel.org/bpf/6ada4c1c9d35eeb5f4ecfab94593dafa6b5c4b09.1574452833.git.daniel@iogearbox.net
2019-11-22 20:08:00 +00:00
|
|
|
EMIT3(0x83, 0xC0, 0x01); /* add eax, 1 */
|
bpf, x64: rework pro/epilogue and tailcall handling in JIT
This commit serves two things:
1) it optimizes BPF prologue/epilogue generation
2) it makes possible to have tailcalls within BPF subprogram
Both points are related to each other since without 1), 2) could not be
achieved.
In [1], Alexei says:
"The prologue will look like:
nop5
xor eax,eax // two new bytes if bpf_tail_call() is used in this
// function
push rbp
mov rbp, rsp
sub rsp, rounded_stack_depth
push rax // zero init tail_call counter
variable number of push rbx,r13,r14,r15
Then bpf_tail_call will pop variable number rbx,..
and final 'pop rax'
Then 'add rsp, size_of_current_stack_frame'
jmp to next function and skip over 'nop5; xor eax,eax; push rpb; mov
rbp, rsp'
This way new function will set its own stack size and will init tail
call
counter with whatever value the parent had.
If next function doesn't use bpf_tail_call it won't have 'xor eax,eax'.
Instead it would need to have 'nop2' in there."
Implement that suggestion.
Since the layout of stack is changed, tail call counter handling can not
rely anymore on popping it to rbx just like it have been handled for
constant prologue case and later overwrite of rbx with actual value of
rbx pushed to stack. Therefore, let's use one of the register (%rcx) that
is considered to be volatile/caller-saved and pop the value of tail call
counter in there in the epilogue.
Drop the BUILD_BUG_ON in emit_prologue and in
emit_bpf_tail_call_indirect where instruction layout is not constant
anymore.
Introduce new poke target, 'tailcall_bypass' to poke descriptor that is
dedicated for skipping the register pops and stack unwind that are
generated right before the actual jump to target program.
For case when the target program is not present, BPF program will skip
the pop instructions and nop5 dedicated for jmpq $target. An example of
such state when only R6 of callee saved registers is used by program:
ffffffffc0513aa1: e9 0e 00 00 00 jmpq 0xffffffffc0513ab4
ffffffffc0513aa6: 5b pop %rbx
ffffffffc0513aa7: 58 pop %rax
ffffffffc0513aa8: 48 81 c4 00 00 00 00 add $0x0,%rsp
ffffffffc0513aaf: 0f 1f 44 00 00 nopl 0x0(%rax,%rax,1)
ffffffffc0513ab4: 48 89 df mov %rbx,%rdi
When target program is inserted, the jump that was there to skip
pops/nop5 will become the nop5, so CPU will go over pops and do the
actual tailcall.
One might ask why there simply can not be pushes after the nop5?
In the following example snippet:
ffffffffc037030c: 48 89 fb mov %rdi,%rbx
(...)
ffffffffc0370332: 5b pop %rbx
ffffffffc0370333: 58 pop %rax
ffffffffc0370334: 48 81 c4 00 00 00 00 add $0x0,%rsp
ffffffffc037033b: 0f 1f 44 00 00 nopl 0x0(%rax,%rax,1)
ffffffffc0370340: 48 81 ec 00 00 00 00 sub $0x0,%rsp
ffffffffc0370347: 50 push %rax
ffffffffc0370348: 53 push %rbx
ffffffffc0370349: 48 89 df mov %rbx,%rdi
ffffffffc037034c: e8 f7 21 00 00 callq 0xffffffffc0372548
There is the bpf2bpf call (at ffffffffc037034c) right after the tailcall
and jump target is not present. ctx is in %rbx register and BPF
subprogram that we will call into on ffffffffc037034c is relying on it,
e.g. it will pick ctx from there. Such code layout is therefore broken
as we would overwrite the content of %rbx with the value that was pushed
on the prologue. That is the reason for the 'bypass' approach.
Special care needs to be taken during the install/update/remove of
tailcall target. In case when target program is not present, the CPU
must not execute the pop instructions that precede the tailcall.
To address that, the following states can be defined:
A nop, unwind, nop
B nop, unwind, tail
C skip, unwind, nop
D skip, unwind, tail
A is forbidden (lead to incorrectness). The state transitions between
tailcall install/update/remove will work as follows:
First install tail call f: C->D->B(f)
* poke the tailcall, after that get rid of the skip
Update tail call f to f': B(f)->B(f')
* poke the tailcall (poke->tailcall_target) and do NOT touch the
poke->tailcall_bypass
Remove tail call: B(f')->C(f')
* poke->tailcall_bypass is poked back to jump, then we wait the RCU
grace period so that other programs will finish its execution and
after that we are safe to remove the poke->tailcall_target
Install new tail call (f''): C(f')->D(f'')->B(f'').
* same as first step
This way CPU can never be exposed to "unwind, tail" state.
Last but not least, when tailcalls get mixed with bpf2bpf calls, it
would be possible to encounter the endless loop due to clearing the
tailcall counter if for example we would use the tailcall3-like from BPF
selftests program that would be subprogram-based, meaning the tailcall
would be present within the BPF subprogram.
This test, broken down to particular steps, would do:
entry -> set tailcall counter to 0, bump it by 1, tailcall to func0
func0 -> call subprog_tail
(we are NOT skipping the first 11 bytes of prologue and this subprogram
has a tailcall, therefore we clear the counter...)
subprog -> do the same thing as entry
and then loop forever.
To address this, the idea is to go through the call chain of bpf2bpf progs
and look for a tailcall presence throughout whole chain. If we saw a single
tail call then each node in this call chain needs to be marked as a subprog
that can reach the tailcall. We would later feed the JIT with this info
and:
- set eax to 0 only when tailcall is reachable and this is the entry prog
- if tailcall is reachable but there's no tailcall in insns of currently
JITed prog then push rax anyway, so that it will be possible to
propagate further down the call chain
- finally if tailcall is reachable, then we need to precede the 'call'
insn with mov rax, [rbp - (stack_depth + 8)]
Tail call related cases from test_verifier kselftest are also working
fine. Sample BPF programs that utilize tail calls (sockex3, tracex5)
work properly as well.
[1]: https://lore.kernel.org/bpf/20200517043227.2gpq22ifoq37ogst@ast-mbp.dhcp.thefacebook.com/
Suggested-by: Alexei Starovoitov <ast@kernel.org>
Signed-off-by: Maciej Fijalkowski <maciej.fijalkowski@intel.com>
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
2020-09-16 21:10:08 +00:00
|
|
|
EMIT2_off32(0x89, 0x85, tcc_off); /* mov dword ptr [rbp - tcc_off], eax */
|
bpf, x86: Emit patchable direct jump as tail call
Add initial code emission for *direct* jumps for tail call maps in
order to avoid the retpoline overhead from a493a87f38cf ("bpf, x64:
implement retpoline for tail call") for situations that allow for
it, meaning, for known constant keys at verification time which are
used as index into the tail call map. In case of Cilium which makes
heavy use of tail calls, constant keys are used in the vast majority,
only for a single occurrence we use a dynamic key.
High level outline is that if the target prog is NULL in the map, we
emit a 5-byte nop for the fall-through case and if not, we emit a
5-byte direct relative jmp to the target bpf_func + skipped prologue
offset. Later during runtime, we patch these 5-byte nop/jmps upon
tail call map update or deletions dynamically. Note that on x86-64
the direct jmp works as we reuse the same stack frame and skip
prologue (as opposed to some other JIT implementations).
One of the issues is that the tail call map slots can change at any
given time even during JITing. Therefore, we have two passes: i) emit
nops for all patchable locations during main JITing phase until we
declare prog->jited = 1 eventually. At this point the image is stable,
not public yet and with all jmps disabled. While JITing, we collect
additional info like poke->ip in order to remember the patch location
for later modifications. In ii) bpf_tail_call_direct_fixup() walks
over the progs poke_tab, locks the tail call maps poke_mutex to
prevent from parallel updates and patches in the right locations via
__bpf_arch_text_poke(). Note, the main bpf_arch_text_poke() cannot
be used at this point since we're not yet exposed to kallsyms. For
the update we use plain memcpy() since the image is not public and
still in read-write mode. After patching, we activate that poke entry
through poke->ip_stable. Meaning, at this point any tail call map
updates/deletions are not going to ignore that poke entry anymore.
Then, bpf_arch_text_poke() might still occur on the read-write image
until we finally locked it as read-only. Both modifications on the
given image are under text_mutex to avoid interference with each
other when update requests come in in parallel for different tail
call maps (current one we have locked in JIT and different one where
poke->ip_stable was already set).
Example prog:
# ./bpftool p d x i 1655
0: (b7) r3 = 0
1: (18) r2 = map[id:526]
3: (85) call bpf_tail_call#12
4: (b7) r0 = 1
5: (95) exit
Before:
# ./bpftool p d j i 1655
0xffffffffc076e55c:
0: nopl 0x0(%rax,%rax,1)
5: push %rbp
6: mov %rsp,%rbp
9: sub $0x200,%rsp
10: push %rbx
11: push %r13
13: push %r14
15: push %r15
17: pushq $0x0 _
19: xor %edx,%edx |_ index (arg 3)
1b: movabs $0xffff88d95cc82600,%rsi |_ map (arg 2)
25: mov %edx,%edx | index >= array->map.max_entries
27: cmp %edx,0x24(%rsi) |
2a: jbe 0x0000000000000066 |_
2c: mov -0x224(%rbp),%eax | tail call limit check
32: cmp $0x20,%eax |
35: ja 0x0000000000000066 |
37: add $0x1,%eax |
3a: mov %eax,-0x224(%rbp) |_
40: mov 0xd0(%rsi,%rdx,8),%rax |_ prog = array->ptrs[index]
48: test %rax,%rax | prog == NULL check
4b: je 0x0000000000000066 |_
4d: mov 0x30(%rax),%rax | goto *(prog->bpf_func + prologue_size)
51: add $0x19,%rax |
55: callq 0x0000000000000061 | retpoline for indirect jump
5a: pause |
5c: lfence |
5f: jmp 0x000000000000005a |
61: mov %rax,(%rsp) |
65: retq |_
66: mov $0x1,%eax
6b: pop %rbx
6c: pop %r15
6e: pop %r14
70: pop %r13
72: pop %rbx
73: leaveq
74: retq
After; state after JIT:
# ./bpftool p d j i 1655
0xffffffffc08e8930:
0: nopl 0x0(%rax,%rax,1)
5: push %rbp
6: mov %rsp,%rbp
9: sub $0x200,%rsp
10: push %rbx
11: push %r13
13: push %r14
15: push %r15
17: pushq $0x0 _
19: xor %edx,%edx |_ index (arg 3)
1b: movabs $0xffff9d8afd74c000,%rsi |_ map (arg 2)
25: mov -0x224(%rbp),%eax | tail call limit check
2b: cmp $0x20,%eax |
2e: ja 0x000000000000003e |
30: add $0x1,%eax |
33: mov %eax,-0x224(%rbp) |_
39: jmpq 0xfffffffffffd1785 |_ [direct] goto *(prog->bpf_func + prologue_size)
3e: mov $0x1,%eax
43: pop %rbx
44: pop %r15
46: pop %r14
48: pop %r13
4a: pop %rbx
4b: leaveq
4c: retq
After; state after map update (target prog):
# ./bpftool p d j i 1655
0xffffffffc08e8930:
0: nopl 0x0(%rax,%rax,1)
5: push %rbp
6: mov %rsp,%rbp
9: sub $0x200,%rsp
10: push %rbx
11: push %r13
13: push %r14
15: push %r15
17: pushq $0x0
19: xor %edx,%edx
1b: movabs $0xffff9d8afd74c000,%rsi
25: mov -0x224(%rbp),%eax
2b: cmp $0x20,%eax .
2e: ja 0x000000000000003e .
30: add $0x1,%eax .
33: mov %eax,-0x224(%rbp) |_
39: jmpq 0xffffffffffb09f55 |_ goto *(prog->bpf_func + prologue_size)
3e: mov $0x1,%eax
43: pop %rbx
44: pop %r15
46: pop %r14
48: pop %r13
4a: pop %rbx
4b: leaveq
4c: retq
After; state after map update (no prog):
# ./bpftool p d j i 1655
0xffffffffc08e8930:
0: nopl 0x0(%rax,%rax,1)
5: push %rbp
6: mov %rsp,%rbp
9: sub $0x200,%rsp
10: push %rbx
11: push %r13
13: push %r14
15: push %r15
17: pushq $0x0
19: xor %edx,%edx
1b: movabs $0xffff9d8afd74c000,%rsi
25: mov -0x224(%rbp),%eax
2b: cmp $0x20,%eax .
2e: ja 0x000000000000003e .
30: add $0x1,%eax .
33: mov %eax,-0x224(%rbp) |_
39: nopl 0x0(%rax,%rax,1) |_ fall-through nop
3e: mov $0x1,%eax
43: pop %rbx
44: pop %r15
46: pop %r14
48: pop %r13
4a: pop %rbx
4b: leaveq
4c: retq
Nice bonus is that this also shrinks the code emission quite a bit
for every tail call invocation.
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Link: https://lore.kernel.org/bpf/6ada4c1c9d35eeb5f4ecfab94593dafa6b5c4b09.1574452833.git.daniel@iogearbox.net
2019-11-22 20:08:00 +00:00
|
|
|
|
2021-10-26 12:01:47 +00:00
|
|
|
poke->tailcall_bypass = ip + (prog - start);
|
bpf, x64: rework pro/epilogue and tailcall handling in JIT
This commit serves two things:
1) it optimizes BPF prologue/epilogue generation
2) it makes possible to have tailcalls within BPF subprogram
Both points are related to each other since without 1), 2) could not be
achieved.
In [1], Alexei says:
"The prologue will look like:
nop5
xor eax,eax // two new bytes if bpf_tail_call() is used in this
// function
push rbp
mov rbp, rsp
sub rsp, rounded_stack_depth
push rax // zero init tail_call counter
variable number of push rbx,r13,r14,r15
Then bpf_tail_call will pop variable number rbx,..
and final 'pop rax'
Then 'add rsp, size_of_current_stack_frame'
jmp to next function and skip over 'nop5; xor eax,eax; push rpb; mov
rbp, rsp'
This way new function will set its own stack size and will init tail
call
counter with whatever value the parent had.
If next function doesn't use bpf_tail_call it won't have 'xor eax,eax'.
Instead it would need to have 'nop2' in there."
Implement that suggestion.
Since the layout of stack is changed, tail call counter handling can not
rely anymore on popping it to rbx just like it have been handled for
constant prologue case and later overwrite of rbx with actual value of
rbx pushed to stack. Therefore, let's use one of the register (%rcx) that
is considered to be volatile/caller-saved and pop the value of tail call
counter in there in the epilogue.
Drop the BUILD_BUG_ON in emit_prologue and in
emit_bpf_tail_call_indirect where instruction layout is not constant
anymore.
Introduce new poke target, 'tailcall_bypass' to poke descriptor that is
dedicated for skipping the register pops and stack unwind that are
generated right before the actual jump to target program.
For case when the target program is not present, BPF program will skip
the pop instructions and nop5 dedicated for jmpq $target. An example of
such state when only R6 of callee saved registers is used by program:
ffffffffc0513aa1: e9 0e 00 00 00 jmpq 0xffffffffc0513ab4
ffffffffc0513aa6: 5b pop %rbx
ffffffffc0513aa7: 58 pop %rax
ffffffffc0513aa8: 48 81 c4 00 00 00 00 add $0x0,%rsp
ffffffffc0513aaf: 0f 1f 44 00 00 nopl 0x0(%rax,%rax,1)
ffffffffc0513ab4: 48 89 df mov %rbx,%rdi
When target program is inserted, the jump that was there to skip
pops/nop5 will become the nop5, so CPU will go over pops and do the
actual tailcall.
One might ask why there simply can not be pushes after the nop5?
In the following example snippet:
ffffffffc037030c: 48 89 fb mov %rdi,%rbx
(...)
ffffffffc0370332: 5b pop %rbx
ffffffffc0370333: 58 pop %rax
ffffffffc0370334: 48 81 c4 00 00 00 00 add $0x0,%rsp
ffffffffc037033b: 0f 1f 44 00 00 nopl 0x0(%rax,%rax,1)
ffffffffc0370340: 48 81 ec 00 00 00 00 sub $0x0,%rsp
ffffffffc0370347: 50 push %rax
ffffffffc0370348: 53 push %rbx
ffffffffc0370349: 48 89 df mov %rbx,%rdi
ffffffffc037034c: e8 f7 21 00 00 callq 0xffffffffc0372548
There is the bpf2bpf call (at ffffffffc037034c) right after the tailcall
and jump target is not present. ctx is in %rbx register and BPF
subprogram that we will call into on ffffffffc037034c is relying on it,
e.g. it will pick ctx from there. Such code layout is therefore broken
as we would overwrite the content of %rbx with the value that was pushed
on the prologue. That is the reason for the 'bypass' approach.
Special care needs to be taken during the install/update/remove of
tailcall target. In case when target program is not present, the CPU
must not execute the pop instructions that precede the tailcall.
To address that, the following states can be defined:
A nop, unwind, nop
B nop, unwind, tail
C skip, unwind, nop
D skip, unwind, tail
A is forbidden (lead to incorrectness). The state transitions between
tailcall install/update/remove will work as follows:
First install tail call f: C->D->B(f)
* poke the tailcall, after that get rid of the skip
Update tail call f to f': B(f)->B(f')
* poke the tailcall (poke->tailcall_target) and do NOT touch the
poke->tailcall_bypass
Remove tail call: B(f')->C(f')
* poke->tailcall_bypass is poked back to jump, then we wait the RCU
grace period so that other programs will finish its execution and
after that we are safe to remove the poke->tailcall_target
Install new tail call (f''): C(f')->D(f'')->B(f'').
* same as first step
This way CPU can never be exposed to "unwind, tail" state.
Last but not least, when tailcalls get mixed with bpf2bpf calls, it
would be possible to encounter the endless loop due to clearing the
tailcall counter if for example we would use the tailcall3-like from BPF
selftests program that would be subprogram-based, meaning the tailcall
would be present within the BPF subprogram.
This test, broken down to particular steps, would do:
entry -> set tailcall counter to 0, bump it by 1, tailcall to func0
func0 -> call subprog_tail
(we are NOT skipping the first 11 bytes of prologue and this subprogram
has a tailcall, therefore we clear the counter...)
subprog -> do the same thing as entry
and then loop forever.
To address this, the idea is to go through the call chain of bpf2bpf progs
and look for a tailcall presence throughout whole chain. If we saw a single
tail call then each node in this call chain needs to be marked as a subprog
that can reach the tailcall. We would later feed the JIT with this info
and:
- set eax to 0 only when tailcall is reachable and this is the entry prog
- if tailcall is reachable but there's no tailcall in insns of currently
JITed prog then push rax anyway, so that it will be possible to
propagate further down the call chain
- finally if tailcall is reachable, then we need to precede the 'call'
insn with mov rax, [rbp - (stack_depth + 8)]
Tail call related cases from test_verifier kselftest are also working
fine. Sample BPF programs that utilize tail calls (sockex3, tracex5)
work properly as well.
[1]: https://lore.kernel.org/bpf/20200517043227.2gpq22ifoq37ogst@ast-mbp.dhcp.thefacebook.com/
Suggested-by: Alexei Starovoitov <ast@kernel.org>
Signed-off-by: Maciej Fijalkowski <maciej.fijalkowski@intel.com>
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
2020-09-16 21:10:08 +00:00
|
|
|
poke->adj_off = X86_TAIL_CALL_OFFSET;
|
2021-10-26 12:01:47 +00:00
|
|
|
poke->tailcall_target = ip + ctx->tail_call_direct_label - X86_PATCH_SIZE;
|
bpf, x64: rework pro/epilogue and tailcall handling in JIT
This commit serves two things:
1) it optimizes BPF prologue/epilogue generation
2) it makes possible to have tailcalls within BPF subprogram
Both points are related to each other since without 1), 2) could not be
achieved.
In [1], Alexei says:
"The prologue will look like:
nop5
xor eax,eax // two new bytes if bpf_tail_call() is used in this
// function
push rbp
mov rbp, rsp
sub rsp, rounded_stack_depth
push rax // zero init tail_call counter
variable number of push rbx,r13,r14,r15
Then bpf_tail_call will pop variable number rbx,..
and final 'pop rax'
Then 'add rsp, size_of_current_stack_frame'
jmp to next function and skip over 'nop5; xor eax,eax; push rpb; mov
rbp, rsp'
This way new function will set its own stack size and will init tail
call
counter with whatever value the parent had.
If next function doesn't use bpf_tail_call it won't have 'xor eax,eax'.
Instead it would need to have 'nop2' in there."
Implement that suggestion.
Since the layout of stack is changed, tail call counter handling can not
rely anymore on popping it to rbx just like it have been handled for
constant prologue case and later overwrite of rbx with actual value of
rbx pushed to stack. Therefore, let's use one of the register (%rcx) that
is considered to be volatile/caller-saved and pop the value of tail call
counter in there in the epilogue.
Drop the BUILD_BUG_ON in emit_prologue and in
emit_bpf_tail_call_indirect where instruction layout is not constant
anymore.
Introduce new poke target, 'tailcall_bypass' to poke descriptor that is
dedicated for skipping the register pops and stack unwind that are
generated right before the actual jump to target program.
For case when the target program is not present, BPF program will skip
the pop instructions and nop5 dedicated for jmpq $target. An example of
such state when only R6 of callee saved registers is used by program:
ffffffffc0513aa1: e9 0e 00 00 00 jmpq 0xffffffffc0513ab4
ffffffffc0513aa6: 5b pop %rbx
ffffffffc0513aa7: 58 pop %rax
ffffffffc0513aa8: 48 81 c4 00 00 00 00 add $0x0,%rsp
ffffffffc0513aaf: 0f 1f 44 00 00 nopl 0x0(%rax,%rax,1)
ffffffffc0513ab4: 48 89 df mov %rbx,%rdi
When target program is inserted, the jump that was there to skip
pops/nop5 will become the nop5, so CPU will go over pops and do the
actual tailcall.
One might ask why there simply can not be pushes after the nop5?
In the following example snippet:
ffffffffc037030c: 48 89 fb mov %rdi,%rbx
(...)
ffffffffc0370332: 5b pop %rbx
ffffffffc0370333: 58 pop %rax
ffffffffc0370334: 48 81 c4 00 00 00 00 add $0x0,%rsp
ffffffffc037033b: 0f 1f 44 00 00 nopl 0x0(%rax,%rax,1)
ffffffffc0370340: 48 81 ec 00 00 00 00 sub $0x0,%rsp
ffffffffc0370347: 50 push %rax
ffffffffc0370348: 53 push %rbx
ffffffffc0370349: 48 89 df mov %rbx,%rdi
ffffffffc037034c: e8 f7 21 00 00 callq 0xffffffffc0372548
There is the bpf2bpf call (at ffffffffc037034c) right after the tailcall
and jump target is not present. ctx is in %rbx register and BPF
subprogram that we will call into on ffffffffc037034c is relying on it,
e.g. it will pick ctx from there. Such code layout is therefore broken
as we would overwrite the content of %rbx with the value that was pushed
on the prologue. That is the reason for the 'bypass' approach.
Special care needs to be taken during the install/update/remove of
tailcall target. In case when target program is not present, the CPU
must not execute the pop instructions that precede the tailcall.
To address that, the following states can be defined:
A nop, unwind, nop
B nop, unwind, tail
C skip, unwind, nop
D skip, unwind, tail
A is forbidden (lead to incorrectness). The state transitions between
tailcall install/update/remove will work as follows:
First install tail call f: C->D->B(f)
* poke the tailcall, after that get rid of the skip
Update tail call f to f': B(f)->B(f')
* poke the tailcall (poke->tailcall_target) and do NOT touch the
poke->tailcall_bypass
Remove tail call: B(f')->C(f')
* poke->tailcall_bypass is poked back to jump, then we wait the RCU
grace period so that other programs will finish its execution and
after that we are safe to remove the poke->tailcall_target
Install new tail call (f''): C(f')->D(f'')->B(f'').
* same as first step
This way CPU can never be exposed to "unwind, tail" state.
Last but not least, when tailcalls get mixed with bpf2bpf calls, it
would be possible to encounter the endless loop due to clearing the
tailcall counter if for example we would use the tailcall3-like from BPF
selftests program that would be subprogram-based, meaning the tailcall
would be present within the BPF subprogram.
This test, broken down to particular steps, would do:
entry -> set tailcall counter to 0, bump it by 1, tailcall to func0
func0 -> call subprog_tail
(we are NOT skipping the first 11 bytes of prologue and this subprogram
has a tailcall, therefore we clear the counter...)
subprog -> do the same thing as entry
and then loop forever.
To address this, the idea is to go through the call chain of bpf2bpf progs
and look for a tailcall presence throughout whole chain. If we saw a single
tail call then each node in this call chain needs to be marked as a subprog
that can reach the tailcall. We would later feed the JIT with this info
and:
- set eax to 0 only when tailcall is reachable and this is the entry prog
- if tailcall is reachable but there's no tailcall in insns of currently
JITed prog then push rax anyway, so that it will be possible to
propagate further down the call chain
- finally if tailcall is reachable, then we need to precede the 'call'
insn with mov rax, [rbp - (stack_depth + 8)]
Tail call related cases from test_verifier kselftest are also working
fine. Sample BPF programs that utilize tail calls (sockex3, tracex5)
work properly as well.
[1]: https://lore.kernel.org/bpf/20200517043227.2gpq22ifoq37ogst@ast-mbp.dhcp.thefacebook.com/
Suggested-by: Alexei Starovoitov <ast@kernel.org>
Signed-off-by: Maciej Fijalkowski <maciej.fijalkowski@intel.com>
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
2020-09-16 21:10:08 +00:00
|
|
|
poke->bypass_addr = (u8 *)poke->tailcall_target + X86_PATCH_SIZE;
|
|
|
|
|
|
|
|
emit_jump(&prog, (u8 *)poke->tailcall_target + X86_PATCH_SIZE,
|
|
|
|
poke->tailcall_bypass);
|
|
|
|
|
bpf: Implement BPF exceptions
This patch implements BPF exceptions, and introduces a bpf_throw kfunc
to allow programs to throw exceptions during their execution at runtime.
A bpf_throw invocation is treated as an immediate termination of the
program, returning back to its caller within the kernel, unwinding all
stack frames.
This allows the program to simplify its implementation, by testing for
runtime conditions which the verifier has no visibility into, and assert
that they are true. In case they are not, the program can simply throw
an exception from the other branch.
BPF exceptions are explicitly *NOT* an unlikely slowpath error handling
primitive, and this objective has guided design choices of the
implementation of the them within the kernel (with the bulk of the cost
for unwinding the stack offloaded to the bpf_throw kfunc).
The implementation of this mechanism requires use of add_hidden_subprog
mechanism introduced in the previous patch, which generates a couple of
instructions to move R1 to R0 and exit. The JIT then rewrites the
prologue of this subprog to take the stack pointer and frame pointer as
inputs and reset the stack frame, popping all callee-saved registers
saved by the main subprog. The bpf_throw function then walks the stack
at runtime, and invokes this exception subprog with the stack and frame
pointers as parameters.
Reviewers must take note that currently the main program is made to save
all callee-saved registers on x86_64 during entry into the program. This
is because we must do an equivalent of a lightweight context switch when
unwinding the stack, therefore we need the callee-saved registers of the
caller of the BPF program to be able to return with a sane state.
Note that we have to additionally handle r12, even though it is not used
by the program, because when throwing the exception the program makes an
entry into the kernel which could clobber r12 after saving it on the
stack. To be able to preserve the value we received on program entry, we
push r12 and restore it from the generated subprogram when unwinding the
stack.
For now, bpf_throw invocation fails when lingering resources or locks
exist in that path of the program. In a future followup, bpf_throw will
be extended to perform frame-by-frame unwinding to release lingering
resources for each stack frame, removing this limitation.
Signed-off-by: Kumar Kartikeya Dwivedi <memxor@gmail.com>
Link: https://lore.kernel.org/r/20230912233214.1518551-5-memxor@gmail.com
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
2023-09-12 23:32:01 +00:00
|
|
|
if (bpf_prog->aux->exception_boundary) {
|
|
|
|
pop_callee_regs(&prog, all_callee_regs_used);
|
|
|
|
pop_r12(&prog);
|
|
|
|
} else {
|
|
|
|
pop_callee_regs(&prog, callee_regs_used);
|
2024-03-08 01:08:01 +00:00
|
|
|
if (bpf_arena_get_kern_vm_start(bpf_prog->aux->arena))
|
|
|
|
pop_r12(&prog);
|
bpf: Implement BPF exceptions
This patch implements BPF exceptions, and introduces a bpf_throw kfunc
to allow programs to throw exceptions during their execution at runtime.
A bpf_throw invocation is treated as an immediate termination of the
program, returning back to its caller within the kernel, unwinding all
stack frames.
This allows the program to simplify its implementation, by testing for
runtime conditions which the verifier has no visibility into, and assert
that they are true. In case they are not, the program can simply throw
an exception from the other branch.
BPF exceptions are explicitly *NOT* an unlikely slowpath error handling
primitive, and this objective has guided design choices of the
implementation of the them within the kernel (with the bulk of the cost
for unwinding the stack offloaded to the bpf_throw kfunc).
The implementation of this mechanism requires use of add_hidden_subprog
mechanism introduced in the previous patch, which generates a couple of
instructions to move R1 to R0 and exit. The JIT then rewrites the
prologue of this subprog to take the stack pointer and frame pointer as
inputs and reset the stack frame, popping all callee-saved registers
saved by the main subprog. The bpf_throw function then walks the stack
at runtime, and invokes this exception subprog with the stack and frame
pointers as parameters.
Reviewers must take note that currently the main program is made to save
all callee-saved registers on x86_64 during entry into the program. This
is because we must do an equivalent of a lightweight context switch when
unwinding the stack, therefore we need the callee-saved registers of the
caller of the BPF program to be able to return with a sane state.
Note that we have to additionally handle r12, even though it is not used
by the program, because when throwing the exception the program makes an
entry into the kernel which could clobber r12 after saving it on the
stack. To be able to preserve the value we received on program entry, we
push r12 and restore it from the generated subprogram when unwinding the
stack.
For now, bpf_throw invocation fails when lingering resources or locks
exist in that path of the program. In a future followup, bpf_throw will
be extended to perform frame-by-frame unwinding to release lingering
resources for each stack frame, removing this limitation.
Signed-off-by: Kumar Kartikeya Dwivedi <memxor@gmail.com>
Link: https://lore.kernel.org/r/20230912233214.1518551-5-memxor@gmail.com
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
2023-09-12 23:32:01 +00:00
|
|
|
}
|
|
|
|
|
bpf, x64: rework pro/epilogue and tailcall handling in JIT
This commit serves two things:
1) it optimizes BPF prologue/epilogue generation
2) it makes possible to have tailcalls within BPF subprogram
Both points are related to each other since without 1), 2) could not be
achieved.
In [1], Alexei says:
"The prologue will look like:
nop5
xor eax,eax // two new bytes if bpf_tail_call() is used in this
// function
push rbp
mov rbp, rsp
sub rsp, rounded_stack_depth
push rax // zero init tail_call counter
variable number of push rbx,r13,r14,r15
Then bpf_tail_call will pop variable number rbx,..
and final 'pop rax'
Then 'add rsp, size_of_current_stack_frame'
jmp to next function and skip over 'nop5; xor eax,eax; push rpb; mov
rbp, rsp'
This way new function will set its own stack size and will init tail
call
counter with whatever value the parent had.
If next function doesn't use bpf_tail_call it won't have 'xor eax,eax'.
Instead it would need to have 'nop2' in there."
Implement that suggestion.
Since the layout of stack is changed, tail call counter handling can not
rely anymore on popping it to rbx just like it have been handled for
constant prologue case and later overwrite of rbx with actual value of
rbx pushed to stack. Therefore, let's use one of the register (%rcx) that
is considered to be volatile/caller-saved and pop the value of tail call
counter in there in the epilogue.
Drop the BUILD_BUG_ON in emit_prologue and in
emit_bpf_tail_call_indirect where instruction layout is not constant
anymore.
Introduce new poke target, 'tailcall_bypass' to poke descriptor that is
dedicated for skipping the register pops and stack unwind that are
generated right before the actual jump to target program.
For case when the target program is not present, BPF program will skip
the pop instructions and nop5 dedicated for jmpq $target. An example of
such state when only R6 of callee saved registers is used by program:
ffffffffc0513aa1: e9 0e 00 00 00 jmpq 0xffffffffc0513ab4
ffffffffc0513aa6: 5b pop %rbx
ffffffffc0513aa7: 58 pop %rax
ffffffffc0513aa8: 48 81 c4 00 00 00 00 add $0x0,%rsp
ffffffffc0513aaf: 0f 1f 44 00 00 nopl 0x0(%rax,%rax,1)
ffffffffc0513ab4: 48 89 df mov %rbx,%rdi
When target program is inserted, the jump that was there to skip
pops/nop5 will become the nop5, so CPU will go over pops and do the
actual tailcall.
One might ask why there simply can not be pushes after the nop5?
In the following example snippet:
ffffffffc037030c: 48 89 fb mov %rdi,%rbx
(...)
ffffffffc0370332: 5b pop %rbx
ffffffffc0370333: 58 pop %rax
ffffffffc0370334: 48 81 c4 00 00 00 00 add $0x0,%rsp
ffffffffc037033b: 0f 1f 44 00 00 nopl 0x0(%rax,%rax,1)
ffffffffc0370340: 48 81 ec 00 00 00 00 sub $0x0,%rsp
ffffffffc0370347: 50 push %rax
ffffffffc0370348: 53 push %rbx
ffffffffc0370349: 48 89 df mov %rbx,%rdi
ffffffffc037034c: e8 f7 21 00 00 callq 0xffffffffc0372548
There is the bpf2bpf call (at ffffffffc037034c) right after the tailcall
and jump target is not present. ctx is in %rbx register and BPF
subprogram that we will call into on ffffffffc037034c is relying on it,
e.g. it will pick ctx from there. Such code layout is therefore broken
as we would overwrite the content of %rbx with the value that was pushed
on the prologue. That is the reason for the 'bypass' approach.
Special care needs to be taken during the install/update/remove of
tailcall target. In case when target program is not present, the CPU
must not execute the pop instructions that precede the tailcall.
To address that, the following states can be defined:
A nop, unwind, nop
B nop, unwind, tail
C skip, unwind, nop
D skip, unwind, tail
A is forbidden (lead to incorrectness). The state transitions between
tailcall install/update/remove will work as follows:
First install tail call f: C->D->B(f)
* poke the tailcall, after that get rid of the skip
Update tail call f to f': B(f)->B(f')
* poke the tailcall (poke->tailcall_target) and do NOT touch the
poke->tailcall_bypass
Remove tail call: B(f')->C(f')
* poke->tailcall_bypass is poked back to jump, then we wait the RCU
grace period so that other programs will finish its execution and
after that we are safe to remove the poke->tailcall_target
Install new tail call (f''): C(f')->D(f'')->B(f'').
* same as first step
This way CPU can never be exposed to "unwind, tail" state.
Last but not least, when tailcalls get mixed with bpf2bpf calls, it
would be possible to encounter the endless loop due to clearing the
tailcall counter if for example we would use the tailcall3-like from BPF
selftests program that would be subprogram-based, meaning the tailcall
would be present within the BPF subprogram.
This test, broken down to particular steps, would do:
entry -> set tailcall counter to 0, bump it by 1, tailcall to func0
func0 -> call subprog_tail
(we are NOT skipping the first 11 bytes of prologue and this subprogram
has a tailcall, therefore we clear the counter...)
subprog -> do the same thing as entry
and then loop forever.
To address this, the idea is to go through the call chain of bpf2bpf progs
and look for a tailcall presence throughout whole chain. If we saw a single
tail call then each node in this call chain needs to be marked as a subprog
that can reach the tailcall. We would later feed the JIT with this info
and:
- set eax to 0 only when tailcall is reachable and this is the entry prog
- if tailcall is reachable but there's no tailcall in insns of currently
JITed prog then push rax anyway, so that it will be possible to
propagate further down the call chain
- finally if tailcall is reachable, then we need to precede the 'call'
insn with mov rax, [rbp - (stack_depth + 8)]
Tail call related cases from test_verifier kselftest are also working
fine. Sample BPF programs that utilize tail calls (sockex3, tracex5)
work properly as well.
[1]: https://lore.kernel.org/bpf/20200517043227.2gpq22ifoq37ogst@ast-mbp.dhcp.thefacebook.com/
Suggested-by: Alexei Starovoitov <ast@kernel.org>
Signed-off-by: Maciej Fijalkowski <maciej.fijalkowski@intel.com>
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
2020-09-16 21:10:08 +00:00
|
|
|
EMIT1(0x58); /* pop rax */
|
2020-09-29 20:46:53 +00:00
|
|
|
if (stack_depth)
|
|
|
|
EMIT3_off32(0x48, 0x81, 0xC4, round_up(stack_depth, 8));
|
bpf, x86: Emit patchable direct jump as tail call
Add initial code emission for *direct* jumps for tail call maps in
order to avoid the retpoline overhead from a493a87f38cf ("bpf, x64:
implement retpoline for tail call") for situations that allow for
it, meaning, for known constant keys at verification time which are
used as index into the tail call map. In case of Cilium which makes
heavy use of tail calls, constant keys are used in the vast majority,
only for a single occurrence we use a dynamic key.
High level outline is that if the target prog is NULL in the map, we
emit a 5-byte nop for the fall-through case and if not, we emit a
5-byte direct relative jmp to the target bpf_func + skipped prologue
offset. Later during runtime, we patch these 5-byte nop/jmps upon
tail call map update or deletions dynamically. Note that on x86-64
the direct jmp works as we reuse the same stack frame and skip
prologue (as opposed to some other JIT implementations).
One of the issues is that the tail call map slots can change at any
given time even during JITing. Therefore, we have two passes: i) emit
nops for all patchable locations during main JITing phase until we
declare prog->jited = 1 eventually. At this point the image is stable,
not public yet and with all jmps disabled. While JITing, we collect
additional info like poke->ip in order to remember the patch location
for later modifications. In ii) bpf_tail_call_direct_fixup() walks
over the progs poke_tab, locks the tail call maps poke_mutex to
prevent from parallel updates and patches in the right locations via
__bpf_arch_text_poke(). Note, the main bpf_arch_text_poke() cannot
be used at this point since we're not yet exposed to kallsyms. For
the update we use plain memcpy() since the image is not public and
still in read-write mode. After patching, we activate that poke entry
through poke->ip_stable. Meaning, at this point any tail call map
updates/deletions are not going to ignore that poke entry anymore.
Then, bpf_arch_text_poke() might still occur on the read-write image
until we finally locked it as read-only. Both modifications on the
given image are under text_mutex to avoid interference with each
other when update requests come in in parallel for different tail
call maps (current one we have locked in JIT and different one where
poke->ip_stable was already set).
Example prog:
# ./bpftool p d x i 1655
0: (b7) r3 = 0
1: (18) r2 = map[id:526]
3: (85) call bpf_tail_call#12
4: (b7) r0 = 1
5: (95) exit
Before:
# ./bpftool p d j i 1655
0xffffffffc076e55c:
0: nopl 0x0(%rax,%rax,1)
5: push %rbp
6: mov %rsp,%rbp
9: sub $0x200,%rsp
10: push %rbx
11: push %r13
13: push %r14
15: push %r15
17: pushq $0x0 _
19: xor %edx,%edx |_ index (arg 3)
1b: movabs $0xffff88d95cc82600,%rsi |_ map (arg 2)
25: mov %edx,%edx | index >= array->map.max_entries
27: cmp %edx,0x24(%rsi) |
2a: jbe 0x0000000000000066 |_
2c: mov -0x224(%rbp),%eax | tail call limit check
32: cmp $0x20,%eax |
35: ja 0x0000000000000066 |
37: add $0x1,%eax |
3a: mov %eax,-0x224(%rbp) |_
40: mov 0xd0(%rsi,%rdx,8),%rax |_ prog = array->ptrs[index]
48: test %rax,%rax | prog == NULL check
4b: je 0x0000000000000066 |_
4d: mov 0x30(%rax),%rax | goto *(prog->bpf_func + prologue_size)
51: add $0x19,%rax |
55: callq 0x0000000000000061 | retpoline for indirect jump
5a: pause |
5c: lfence |
5f: jmp 0x000000000000005a |
61: mov %rax,(%rsp) |
65: retq |_
66: mov $0x1,%eax
6b: pop %rbx
6c: pop %r15
6e: pop %r14
70: pop %r13
72: pop %rbx
73: leaveq
74: retq
After; state after JIT:
# ./bpftool p d j i 1655
0xffffffffc08e8930:
0: nopl 0x0(%rax,%rax,1)
5: push %rbp
6: mov %rsp,%rbp
9: sub $0x200,%rsp
10: push %rbx
11: push %r13
13: push %r14
15: push %r15
17: pushq $0x0 _
19: xor %edx,%edx |_ index (arg 3)
1b: movabs $0xffff9d8afd74c000,%rsi |_ map (arg 2)
25: mov -0x224(%rbp),%eax | tail call limit check
2b: cmp $0x20,%eax |
2e: ja 0x000000000000003e |
30: add $0x1,%eax |
33: mov %eax,-0x224(%rbp) |_
39: jmpq 0xfffffffffffd1785 |_ [direct] goto *(prog->bpf_func + prologue_size)
3e: mov $0x1,%eax
43: pop %rbx
44: pop %r15
46: pop %r14
48: pop %r13
4a: pop %rbx
4b: leaveq
4c: retq
After; state after map update (target prog):
# ./bpftool p d j i 1655
0xffffffffc08e8930:
0: nopl 0x0(%rax,%rax,1)
5: push %rbp
6: mov %rsp,%rbp
9: sub $0x200,%rsp
10: push %rbx
11: push %r13
13: push %r14
15: push %r15
17: pushq $0x0
19: xor %edx,%edx
1b: movabs $0xffff9d8afd74c000,%rsi
25: mov -0x224(%rbp),%eax
2b: cmp $0x20,%eax .
2e: ja 0x000000000000003e .
30: add $0x1,%eax .
33: mov %eax,-0x224(%rbp) |_
39: jmpq 0xffffffffffb09f55 |_ goto *(prog->bpf_func + prologue_size)
3e: mov $0x1,%eax
43: pop %rbx
44: pop %r15
46: pop %r14
48: pop %r13
4a: pop %rbx
4b: leaveq
4c: retq
After; state after map update (no prog):
# ./bpftool p d j i 1655
0xffffffffc08e8930:
0: nopl 0x0(%rax,%rax,1)
5: push %rbp
6: mov %rsp,%rbp
9: sub $0x200,%rsp
10: push %rbx
11: push %r13
13: push %r14
15: push %r15
17: pushq $0x0
19: xor %edx,%edx
1b: movabs $0xffff9d8afd74c000,%rsi
25: mov -0x224(%rbp),%eax
2b: cmp $0x20,%eax .
2e: ja 0x000000000000003e .
30: add $0x1,%eax .
33: mov %eax,-0x224(%rbp) |_
39: nopl 0x0(%rax,%rax,1) |_ fall-through nop
3e: mov $0x1,%eax
43: pop %rbx
44: pop %r15
46: pop %r14
48: pop %r13
4a: pop %rbx
4b: leaveq
4c: retq
Nice bonus is that this also shrinks the code emission quite a bit
for every tail call invocation.
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Link: https://lore.kernel.org/bpf/6ada4c1c9d35eeb5f4ecfab94593dafa6b5c4b09.1574452833.git.daniel@iogearbox.net
2019-11-22 20:08:00 +00:00
|
|
|
|
2024-01-04 14:22:23 +00:00
|
|
|
emit_nops(&prog, X86_PATCH_SIZE);
|
2021-10-26 12:01:47 +00:00
|
|
|
|
bpf, x86: Emit patchable direct jump as tail call
Add initial code emission for *direct* jumps for tail call maps in
order to avoid the retpoline overhead from a493a87f38cf ("bpf, x64:
implement retpoline for tail call") for situations that allow for
it, meaning, for known constant keys at verification time which are
used as index into the tail call map. In case of Cilium which makes
heavy use of tail calls, constant keys are used in the vast majority,
only for a single occurrence we use a dynamic key.
High level outline is that if the target prog is NULL in the map, we
emit a 5-byte nop for the fall-through case and if not, we emit a
5-byte direct relative jmp to the target bpf_func + skipped prologue
offset. Later during runtime, we patch these 5-byte nop/jmps upon
tail call map update or deletions dynamically. Note that on x86-64
the direct jmp works as we reuse the same stack frame and skip
prologue (as opposed to some other JIT implementations).
One of the issues is that the tail call map slots can change at any
given time even during JITing. Therefore, we have two passes: i) emit
nops for all patchable locations during main JITing phase until we
declare prog->jited = 1 eventually. At this point the image is stable,
not public yet and with all jmps disabled. While JITing, we collect
additional info like poke->ip in order to remember the patch location
for later modifications. In ii) bpf_tail_call_direct_fixup() walks
over the progs poke_tab, locks the tail call maps poke_mutex to
prevent from parallel updates and patches in the right locations via
__bpf_arch_text_poke(). Note, the main bpf_arch_text_poke() cannot
be used at this point since we're not yet exposed to kallsyms. For
the update we use plain memcpy() since the image is not public and
still in read-write mode. After patching, we activate that poke entry
through poke->ip_stable. Meaning, at this point any tail call map
updates/deletions are not going to ignore that poke entry anymore.
Then, bpf_arch_text_poke() might still occur on the read-write image
until we finally locked it as read-only. Both modifications on the
given image are under text_mutex to avoid interference with each
other when update requests come in in parallel for different tail
call maps (current one we have locked in JIT and different one where
poke->ip_stable was already set).
Example prog:
# ./bpftool p d x i 1655
0: (b7) r3 = 0
1: (18) r2 = map[id:526]
3: (85) call bpf_tail_call#12
4: (b7) r0 = 1
5: (95) exit
Before:
# ./bpftool p d j i 1655
0xffffffffc076e55c:
0: nopl 0x0(%rax,%rax,1)
5: push %rbp
6: mov %rsp,%rbp
9: sub $0x200,%rsp
10: push %rbx
11: push %r13
13: push %r14
15: push %r15
17: pushq $0x0 _
19: xor %edx,%edx |_ index (arg 3)
1b: movabs $0xffff88d95cc82600,%rsi |_ map (arg 2)
25: mov %edx,%edx | index >= array->map.max_entries
27: cmp %edx,0x24(%rsi) |
2a: jbe 0x0000000000000066 |_
2c: mov -0x224(%rbp),%eax | tail call limit check
32: cmp $0x20,%eax |
35: ja 0x0000000000000066 |
37: add $0x1,%eax |
3a: mov %eax,-0x224(%rbp) |_
40: mov 0xd0(%rsi,%rdx,8),%rax |_ prog = array->ptrs[index]
48: test %rax,%rax | prog == NULL check
4b: je 0x0000000000000066 |_
4d: mov 0x30(%rax),%rax | goto *(prog->bpf_func + prologue_size)
51: add $0x19,%rax |
55: callq 0x0000000000000061 | retpoline for indirect jump
5a: pause |
5c: lfence |
5f: jmp 0x000000000000005a |
61: mov %rax,(%rsp) |
65: retq |_
66: mov $0x1,%eax
6b: pop %rbx
6c: pop %r15
6e: pop %r14
70: pop %r13
72: pop %rbx
73: leaveq
74: retq
After; state after JIT:
# ./bpftool p d j i 1655
0xffffffffc08e8930:
0: nopl 0x0(%rax,%rax,1)
5: push %rbp
6: mov %rsp,%rbp
9: sub $0x200,%rsp
10: push %rbx
11: push %r13
13: push %r14
15: push %r15
17: pushq $0x0 _
19: xor %edx,%edx |_ index (arg 3)
1b: movabs $0xffff9d8afd74c000,%rsi |_ map (arg 2)
25: mov -0x224(%rbp),%eax | tail call limit check
2b: cmp $0x20,%eax |
2e: ja 0x000000000000003e |
30: add $0x1,%eax |
33: mov %eax,-0x224(%rbp) |_
39: jmpq 0xfffffffffffd1785 |_ [direct] goto *(prog->bpf_func + prologue_size)
3e: mov $0x1,%eax
43: pop %rbx
44: pop %r15
46: pop %r14
48: pop %r13
4a: pop %rbx
4b: leaveq
4c: retq
After; state after map update (target prog):
# ./bpftool p d j i 1655
0xffffffffc08e8930:
0: nopl 0x0(%rax,%rax,1)
5: push %rbp
6: mov %rsp,%rbp
9: sub $0x200,%rsp
10: push %rbx
11: push %r13
13: push %r14
15: push %r15
17: pushq $0x0
19: xor %edx,%edx
1b: movabs $0xffff9d8afd74c000,%rsi
25: mov -0x224(%rbp),%eax
2b: cmp $0x20,%eax .
2e: ja 0x000000000000003e .
30: add $0x1,%eax .
33: mov %eax,-0x224(%rbp) |_
39: jmpq 0xffffffffffb09f55 |_ goto *(prog->bpf_func + prologue_size)
3e: mov $0x1,%eax
43: pop %rbx
44: pop %r15
46: pop %r14
48: pop %r13
4a: pop %rbx
4b: leaveq
4c: retq
After; state after map update (no prog):
# ./bpftool p d j i 1655
0xffffffffc08e8930:
0: nopl 0x0(%rax,%rax,1)
5: push %rbp
6: mov %rsp,%rbp
9: sub $0x200,%rsp
10: push %rbx
11: push %r13
13: push %r14
15: push %r15
17: pushq $0x0
19: xor %edx,%edx
1b: movabs $0xffff9d8afd74c000,%rsi
25: mov -0x224(%rbp),%eax
2b: cmp $0x20,%eax .
2e: ja 0x000000000000003e .
30: add $0x1,%eax .
33: mov %eax,-0x224(%rbp) |_
39: nopl 0x0(%rax,%rax,1) |_ fall-through nop
3e: mov $0x1,%eax
43: pop %rbx
44: pop %r15
46: pop %r14
48: pop %r13
4a: pop %rbx
4b: leaveq
4c: retq
Nice bonus is that this also shrinks the code emission quite a bit
for every tail call invocation.
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Link: https://lore.kernel.org/bpf/6ada4c1c9d35eeb5f4ecfab94593dafa6b5c4b09.1574452833.git.daniel@iogearbox.net
2019-11-22 20:08:00 +00:00
|
|
|
/* out: */
|
2021-10-26 12:01:47 +00:00
|
|
|
ctx->tail_call_direct_label = prog - start;
|
bpf, x86: Emit patchable direct jump as tail call
Add initial code emission for *direct* jumps for tail call maps in
order to avoid the retpoline overhead from a493a87f38cf ("bpf, x64:
implement retpoline for tail call") for situations that allow for
it, meaning, for known constant keys at verification time which are
used as index into the tail call map. In case of Cilium which makes
heavy use of tail calls, constant keys are used in the vast majority,
only for a single occurrence we use a dynamic key.
High level outline is that if the target prog is NULL in the map, we
emit a 5-byte nop for the fall-through case and if not, we emit a
5-byte direct relative jmp to the target bpf_func + skipped prologue
offset. Later during runtime, we patch these 5-byte nop/jmps upon
tail call map update or deletions dynamically. Note that on x86-64
the direct jmp works as we reuse the same stack frame and skip
prologue (as opposed to some other JIT implementations).
One of the issues is that the tail call map slots can change at any
given time even during JITing. Therefore, we have two passes: i) emit
nops for all patchable locations during main JITing phase until we
declare prog->jited = 1 eventually. At this point the image is stable,
not public yet and with all jmps disabled. While JITing, we collect
additional info like poke->ip in order to remember the patch location
for later modifications. In ii) bpf_tail_call_direct_fixup() walks
over the progs poke_tab, locks the tail call maps poke_mutex to
prevent from parallel updates and patches in the right locations via
__bpf_arch_text_poke(). Note, the main bpf_arch_text_poke() cannot
be used at this point since we're not yet exposed to kallsyms. For
the update we use plain memcpy() since the image is not public and
still in read-write mode. After patching, we activate that poke entry
through poke->ip_stable. Meaning, at this point any tail call map
updates/deletions are not going to ignore that poke entry anymore.
Then, bpf_arch_text_poke() might still occur on the read-write image
until we finally locked it as read-only. Both modifications on the
given image are under text_mutex to avoid interference with each
other when update requests come in in parallel for different tail
call maps (current one we have locked in JIT and different one where
poke->ip_stable was already set).
Example prog:
# ./bpftool p d x i 1655
0: (b7) r3 = 0
1: (18) r2 = map[id:526]
3: (85) call bpf_tail_call#12
4: (b7) r0 = 1
5: (95) exit
Before:
# ./bpftool p d j i 1655
0xffffffffc076e55c:
0: nopl 0x0(%rax,%rax,1)
5: push %rbp
6: mov %rsp,%rbp
9: sub $0x200,%rsp
10: push %rbx
11: push %r13
13: push %r14
15: push %r15
17: pushq $0x0 _
19: xor %edx,%edx |_ index (arg 3)
1b: movabs $0xffff88d95cc82600,%rsi |_ map (arg 2)
25: mov %edx,%edx | index >= array->map.max_entries
27: cmp %edx,0x24(%rsi) |
2a: jbe 0x0000000000000066 |_
2c: mov -0x224(%rbp),%eax | tail call limit check
32: cmp $0x20,%eax |
35: ja 0x0000000000000066 |
37: add $0x1,%eax |
3a: mov %eax,-0x224(%rbp) |_
40: mov 0xd0(%rsi,%rdx,8),%rax |_ prog = array->ptrs[index]
48: test %rax,%rax | prog == NULL check
4b: je 0x0000000000000066 |_
4d: mov 0x30(%rax),%rax | goto *(prog->bpf_func + prologue_size)
51: add $0x19,%rax |
55: callq 0x0000000000000061 | retpoline for indirect jump
5a: pause |
5c: lfence |
5f: jmp 0x000000000000005a |
61: mov %rax,(%rsp) |
65: retq |_
66: mov $0x1,%eax
6b: pop %rbx
6c: pop %r15
6e: pop %r14
70: pop %r13
72: pop %rbx
73: leaveq
74: retq
After; state after JIT:
# ./bpftool p d j i 1655
0xffffffffc08e8930:
0: nopl 0x0(%rax,%rax,1)
5: push %rbp
6: mov %rsp,%rbp
9: sub $0x200,%rsp
10: push %rbx
11: push %r13
13: push %r14
15: push %r15
17: pushq $0x0 _
19: xor %edx,%edx |_ index (arg 3)
1b: movabs $0xffff9d8afd74c000,%rsi |_ map (arg 2)
25: mov -0x224(%rbp),%eax | tail call limit check
2b: cmp $0x20,%eax |
2e: ja 0x000000000000003e |
30: add $0x1,%eax |
33: mov %eax,-0x224(%rbp) |_
39: jmpq 0xfffffffffffd1785 |_ [direct] goto *(prog->bpf_func + prologue_size)
3e: mov $0x1,%eax
43: pop %rbx
44: pop %r15
46: pop %r14
48: pop %r13
4a: pop %rbx
4b: leaveq
4c: retq
After; state after map update (target prog):
# ./bpftool p d j i 1655
0xffffffffc08e8930:
0: nopl 0x0(%rax,%rax,1)
5: push %rbp
6: mov %rsp,%rbp
9: sub $0x200,%rsp
10: push %rbx
11: push %r13
13: push %r14
15: push %r15
17: pushq $0x0
19: xor %edx,%edx
1b: movabs $0xffff9d8afd74c000,%rsi
25: mov -0x224(%rbp),%eax
2b: cmp $0x20,%eax .
2e: ja 0x000000000000003e .
30: add $0x1,%eax .
33: mov %eax,-0x224(%rbp) |_
39: jmpq 0xffffffffffb09f55 |_ goto *(prog->bpf_func + prologue_size)
3e: mov $0x1,%eax
43: pop %rbx
44: pop %r15
46: pop %r14
48: pop %r13
4a: pop %rbx
4b: leaveq
4c: retq
After; state after map update (no prog):
# ./bpftool p d j i 1655
0xffffffffc08e8930:
0: nopl 0x0(%rax,%rax,1)
5: push %rbp
6: mov %rsp,%rbp
9: sub $0x200,%rsp
10: push %rbx
11: push %r13
13: push %r14
15: push %r15
17: pushq $0x0
19: xor %edx,%edx
1b: movabs $0xffff9d8afd74c000,%rsi
25: mov -0x224(%rbp),%eax
2b: cmp $0x20,%eax .
2e: ja 0x000000000000003e .
30: add $0x1,%eax .
33: mov %eax,-0x224(%rbp) |_
39: nopl 0x0(%rax,%rax,1) |_ fall-through nop
3e: mov $0x1,%eax
43: pop %rbx
44: pop %r15
46: pop %r14
48: pop %r13
4a: pop %rbx
4b: leaveq
4c: retq
Nice bonus is that this also shrinks the code emission quite a bit
for every tail call invocation.
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Link: https://lore.kernel.org/bpf/6ada4c1c9d35eeb5f4ecfab94593dafa6b5c4b09.1574452833.git.daniel@iogearbox.net
2019-11-22 20:08:00 +00:00
|
|
|
|
|
|
|
*pprog = prog;
|
|
|
|
}
|
|
|
|
|
|
|
|
static void bpf_tail_call_direct_fixup(struct bpf_prog *prog)
|
|
|
|
{
|
|
|
|
struct bpf_jit_poke_descriptor *poke;
|
|
|
|
struct bpf_array *array;
|
|
|
|
struct bpf_prog *target;
|
|
|
|
int i, ret;
|
|
|
|
|
|
|
|
for (i = 0; i < prog->aux->size_poke_tab; i++) {
|
|
|
|
poke = &prog->aux->poke_tab[i];
|
bpf: Track subprog poke descriptors correctly and fix use-after-free
Subprograms are calling map_poke_track(), but on program release there is no
hook to call map_poke_untrack(). However, on program release, the aux memory
(and poke descriptor table) is freed even though we still have a reference to
it in the element list of the map aux data. When we run map_poke_run(), we then
end up accessing free'd memory, triggering KASAN in prog_array_map_poke_run():
[...]
[ 402.824689] BUG: KASAN: use-after-free in prog_array_map_poke_run+0xc2/0x34e
[ 402.824698] Read of size 4 at addr ffff8881905a7940 by task hubble-fgs/4337
[ 402.824705] CPU: 1 PID: 4337 Comm: hubble-fgs Tainted: G I 5.12.0+ #399
[ 402.824715] Call Trace:
[ 402.824719] dump_stack+0x93/0xc2
[ 402.824727] print_address_description.constprop.0+0x1a/0x140
[ 402.824736] ? prog_array_map_poke_run+0xc2/0x34e
[ 402.824740] ? prog_array_map_poke_run+0xc2/0x34e
[ 402.824744] kasan_report.cold+0x7c/0xd8
[ 402.824752] ? prog_array_map_poke_run+0xc2/0x34e
[ 402.824757] prog_array_map_poke_run+0xc2/0x34e
[ 402.824765] bpf_fd_array_map_update_elem+0x124/0x1a0
[...]
The elements concerned are walked as follows:
for (i = 0; i < elem->aux->size_poke_tab; i++) {
poke = &elem->aux->poke_tab[i];
[...]
The access to size_poke_tab is a 4 byte read, verified by checking offsets
in the KASAN dump:
[ 402.825004] The buggy address belongs to the object at ffff8881905a7800
which belongs to the cache kmalloc-1k of size 1024
[ 402.825008] The buggy address is located 320 bytes inside of
1024-byte region [ffff8881905a7800, ffff8881905a7c00)
The pahole output of bpf_prog_aux:
struct bpf_prog_aux {
[...]
/* --- cacheline 5 boundary (320 bytes) --- */
u32 size_poke_tab; /* 320 4 */
[...]
In general, subprograms do not necessarily manage their own data structures.
For example, BTF func_info and linfo are just pointers to the main program
structure. This allows reference counting and cleanup to be done on the latter
which simplifies their management a bit. The aux->poke_tab struct, however,
did not follow this logic. The initial proposed fix for this use-after-free
bug further embedded poke data tracking into the subprogram with proper
reference counting. However, Daniel and Alexei questioned why we were treating
these objects special; I agree, its unnecessary. The fix here removes the per
subprogram poke table allocation and map tracking and instead simply points
the aux->poke_tab pointer at the main programs poke table. This way, map
tracking is simplified to the main program and we do not need to manage them
per subprogram.
This also means, bpf_prog_free_deferred(), which unwinds the program reference
counting and kfrees objects, needs to ensure that we don't try to double free
the poke_tab when free'ing the subprog structures. This is easily solved by
NULL'ing the poke_tab pointer. The second detail is to ensure that per
subprogram JIT logic only does fixups on poke_tab[] entries it owns. To do
this, we add a pointer in the poke structure to point at the subprogram value
so JITs can easily check while walking the poke_tab structure if the current
entry belongs to the current program. The aux pointer is stable and therefore
suitable for such comparison. On the jit_subprogs() error path, we omit
cleaning up the poke->aux field because these are only ever referenced from
the JIT side, but on error we will never make it to the JIT, so its fine to
leave them dangling. Removing these pointers would complicate the error path
for no reason. However, we do need to untrack all poke descriptors from the
main program as otherwise they could race with the freeing of JIT memory from
the subprograms. Lastly, a748c6975dea3 ("bpf: propagate poke descriptors to
subprograms") had an off-by-one on the subprogram instruction index range
check as it was testing 'insn_idx >= subprog_start && insn_idx <= subprog_end'.
However, subprog_end is the next subprogram's start instruction.
Fixes: a748c6975dea3 ("bpf: propagate poke descriptors to subprograms")
Signed-off-by: John Fastabend <john.fastabend@gmail.com>
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Co-developed-by: Daniel Borkmann <daniel@iogearbox.net>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Link: https://lore.kernel.org/bpf/20210707223848.14580-2-john.fastabend@gmail.com
2021-07-07 22:38:47 +00:00
|
|
|
if (poke->aux && poke->aux != prog->aux)
|
|
|
|
continue;
|
|
|
|
|
2020-09-16 21:10:06 +00:00
|
|
|
WARN_ON_ONCE(READ_ONCE(poke->tailcall_target_stable));
|
bpf, x86: Emit patchable direct jump as tail call
Add initial code emission for *direct* jumps for tail call maps in
order to avoid the retpoline overhead from a493a87f38cf ("bpf, x64:
implement retpoline for tail call") for situations that allow for
it, meaning, for known constant keys at verification time which are
used as index into the tail call map. In case of Cilium which makes
heavy use of tail calls, constant keys are used in the vast majority,
only for a single occurrence we use a dynamic key.
High level outline is that if the target prog is NULL in the map, we
emit a 5-byte nop for the fall-through case and if not, we emit a
5-byte direct relative jmp to the target bpf_func + skipped prologue
offset. Later during runtime, we patch these 5-byte nop/jmps upon
tail call map update or deletions dynamically. Note that on x86-64
the direct jmp works as we reuse the same stack frame and skip
prologue (as opposed to some other JIT implementations).
One of the issues is that the tail call map slots can change at any
given time even during JITing. Therefore, we have two passes: i) emit
nops for all patchable locations during main JITing phase until we
declare prog->jited = 1 eventually. At this point the image is stable,
not public yet and with all jmps disabled. While JITing, we collect
additional info like poke->ip in order to remember the patch location
for later modifications. In ii) bpf_tail_call_direct_fixup() walks
over the progs poke_tab, locks the tail call maps poke_mutex to
prevent from parallel updates and patches in the right locations via
__bpf_arch_text_poke(). Note, the main bpf_arch_text_poke() cannot
be used at this point since we're not yet exposed to kallsyms. For
the update we use plain memcpy() since the image is not public and
still in read-write mode. After patching, we activate that poke entry
through poke->ip_stable. Meaning, at this point any tail call map
updates/deletions are not going to ignore that poke entry anymore.
Then, bpf_arch_text_poke() might still occur on the read-write image
until we finally locked it as read-only. Both modifications on the
given image are under text_mutex to avoid interference with each
other when update requests come in in parallel for different tail
call maps (current one we have locked in JIT and different one where
poke->ip_stable was already set).
Example prog:
# ./bpftool p d x i 1655
0: (b7) r3 = 0
1: (18) r2 = map[id:526]
3: (85) call bpf_tail_call#12
4: (b7) r0 = 1
5: (95) exit
Before:
# ./bpftool p d j i 1655
0xffffffffc076e55c:
0: nopl 0x0(%rax,%rax,1)
5: push %rbp
6: mov %rsp,%rbp
9: sub $0x200,%rsp
10: push %rbx
11: push %r13
13: push %r14
15: push %r15
17: pushq $0x0 _
19: xor %edx,%edx |_ index (arg 3)
1b: movabs $0xffff88d95cc82600,%rsi |_ map (arg 2)
25: mov %edx,%edx | index >= array->map.max_entries
27: cmp %edx,0x24(%rsi) |
2a: jbe 0x0000000000000066 |_
2c: mov -0x224(%rbp),%eax | tail call limit check
32: cmp $0x20,%eax |
35: ja 0x0000000000000066 |
37: add $0x1,%eax |
3a: mov %eax,-0x224(%rbp) |_
40: mov 0xd0(%rsi,%rdx,8),%rax |_ prog = array->ptrs[index]
48: test %rax,%rax | prog == NULL check
4b: je 0x0000000000000066 |_
4d: mov 0x30(%rax),%rax | goto *(prog->bpf_func + prologue_size)
51: add $0x19,%rax |
55: callq 0x0000000000000061 | retpoline for indirect jump
5a: pause |
5c: lfence |
5f: jmp 0x000000000000005a |
61: mov %rax,(%rsp) |
65: retq |_
66: mov $0x1,%eax
6b: pop %rbx
6c: pop %r15
6e: pop %r14
70: pop %r13
72: pop %rbx
73: leaveq
74: retq
After; state after JIT:
# ./bpftool p d j i 1655
0xffffffffc08e8930:
0: nopl 0x0(%rax,%rax,1)
5: push %rbp
6: mov %rsp,%rbp
9: sub $0x200,%rsp
10: push %rbx
11: push %r13
13: push %r14
15: push %r15
17: pushq $0x0 _
19: xor %edx,%edx |_ index (arg 3)
1b: movabs $0xffff9d8afd74c000,%rsi |_ map (arg 2)
25: mov -0x224(%rbp),%eax | tail call limit check
2b: cmp $0x20,%eax |
2e: ja 0x000000000000003e |
30: add $0x1,%eax |
33: mov %eax,-0x224(%rbp) |_
39: jmpq 0xfffffffffffd1785 |_ [direct] goto *(prog->bpf_func + prologue_size)
3e: mov $0x1,%eax
43: pop %rbx
44: pop %r15
46: pop %r14
48: pop %r13
4a: pop %rbx
4b: leaveq
4c: retq
After; state after map update (target prog):
# ./bpftool p d j i 1655
0xffffffffc08e8930:
0: nopl 0x0(%rax,%rax,1)
5: push %rbp
6: mov %rsp,%rbp
9: sub $0x200,%rsp
10: push %rbx
11: push %r13
13: push %r14
15: push %r15
17: pushq $0x0
19: xor %edx,%edx
1b: movabs $0xffff9d8afd74c000,%rsi
25: mov -0x224(%rbp),%eax
2b: cmp $0x20,%eax .
2e: ja 0x000000000000003e .
30: add $0x1,%eax .
33: mov %eax,-0x224(%rbp) |_
39: jmpq 0xffffffffffb09f55 |_ goto *(prog->bpf_func + prologue_size)
3e: mov $0x1,%eax
43: pop %rbx
44: pop %r15
46: pop %r14
48: pop %r13
4a: pop %rbx
4b: leaveq
4c: retq
After; state after map update (no prog):
# ./bpftool p d j i 1655
0xffffffffc08e8930:
0: nopl 0x0(%rax,%rax,1)
5: push %rbp
6: mov %rsp,%rbp
9: sub $0x200,%rsp
10: push %rbx
11: push %r13
13: push %r14
15: push %r15
17: pushq $0x0
19: xor %edx,%edx
1b: movabs $0xffff9d8afd74c000,%rsi
25: mov -0x224(%rbp),%eax
2b: cmp $0x20,%eax .
2e: ja 0x000000000000003e .
30: add $0x1,%eax .
33: mov %eax,-0x224(%rbp) |_
39: nopl 0x0(%rax,%rax,1) |_ fall-through nop
3e: mov $0x1,%eax
43: pop %rbx
44: pop %r15
46: pop %r14
48: pop %r13
4a: pop %rbx
4b: leaveq
4c: retq
Nice bonus is that this also shrinks the code emission quite a bit
for every tail call invocation.
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Link: https://lore.kernel.org/bpf/6ada4c1c9d35eeb5f4ecfab94593dafa6b5c4b09.1574452833.git.daniel@iogearbox.net
2019-11-22 20:08:00 +00:00
|
|
|
|
|
|
|
if (poke->reason != BPF_POKE_REASON_TAIL_CALL)
|
|
|
|
continue;
|
|
|
|
|
|
|
|
array = container_of(poke->tail_call.map, struct bpf_array, map);
|
|
|
|
mutex_lock(&array->aux->poke_mutex);
|
|
|
|
target = array->ptrs[poke->tail_call.key];
|
|
|
|
if (target) {
|
2020-09-16 21:10:06 +00:00
|
|
|
ret = __bpf_arch_text_poke(poke->tailcall_target,
|
|
|
|
BPF_MOD_JUMP, NULL,
|
bpf, x86: Emit patchable direct jump as tail call
Add initial code emission for *direct* jumps for tail call maps in
order to avoid the retpoline overhead from a493a87f38cf ("bpf, x64:
implement retpoline for tail call") for situations that allow for
it, meaning, for known constant keys at verification time which are
used as index into the tail call map. In case of Cilium which makes
heavy use of tail calls, constant keys are used in the vast majority,
only for a single occurrence we use a dynamic key.
High level outline is that if the target prog is NULL in the map, we
emit a 5-byte nop for the fall-through case and if not, we emit a
5-byte direct relative jmp to the target bpf_func + skipped prologue
offset. Later during runtime, we patch these 5-byte nop/jmps upon
tail call map update or deletions dynamically. Note that on x86-64
the direct jmp works as we reuse the same stack frame and skip
prologue (as opposed to some other JIT implementations).
One of the issues is that the tail call map slots can change at any
given time even during JITing. Therefore, we have two passes: i) emit
nops for all patchable locations during main JITing phase until we
declare prog->jited = 1 eventually. At this point the image is stable,
not public yet and with all jmps disabled. While JITing, we collect
additional info like poke->ip in order to remember the patch location
for later modifications. In ii) bpf_tail_call_direct_fixup() walks
over the progs poke_tab, locks the tail call maps poke_mutex to
prevent from parallel updates and patches in the right locations via
__bpf_arch_text_poke(). Note, the main bpf_arch_text_poke() cannot
be used at this point since we're not yet exposed to kallsyms. For
the update we use plain memcpy() since the image is not public and
still in read-write mode. After patching, we activate that poke entry
through poke->ip_stable. Meaning, at this point any tail call map
updates/deletions are not going to ignore that poke entry anymore.
Then, bpf_arch_text_poke() might still occur on the read-write image
until we finally locked it as read-only. Both modifications on the
given image are under text_mutex to avoid interference with each
other when update requests come in in parallel for different tail
call maps (current one we have locked in JIT and different one where
poke->ip_stable was already set).
Example prog:
# ./bpftool p d x i 1655
0: (b7) r3 = 0
1: (18) r2 = map[id:526]
3: (85) call bpf_tail_call#12
4: (b7) r0 = 1
5: (95) exit
Before:
# ./bpftool p d j i 1655
0xffffffffc076e55c:
0: nopl 0x0(%rax,%rax,1)
5: push %rbp
6: mov %rsp,%rbp
9: sub $0x200,%rsp
10: push %rbx
11: push %r13
13: push %r14
15: push %r15
17: pushq $0x0 _
19: xor %edx,%edx |_ index (arg 3)
1b: movabs $0xffff88d95cc82600,%rsi |_ map (arg 2)
25: mov %edx,%edx | index >= array->map.max_entries
27: cmp %edx,0x24(%rsi) |
2a: jbe 0x0000000000000066 |_
2c: mov -0x224(%rbp),%eax | tail call limit check
32: cmp $0x20,%eax |
35: ja 0x0000000000000066 |
37: add $0x1,%eax |
3a: mov %eax,-0x224(%rbp) |_
40: mov 0xd0(%rsi,%rdx,8),%rax |_ prog = array->ptrs[index]
48: test %rax,%rax | prog == NULL check
4b: je 0x0000000000000066 |_
4d: mov 0x30(%rax),%rax | goto *(prog->bpf_func + prologue_size)
51: add $0x19,%rax |
55: callq 0x0000000000000061 | retpoline for indirect jump
5a: pause |
5c: lfence |
5f: jmp 0x000000000000005a |
61: mov %rax,(%rsp) |
65: retq |_
66: mov $0x1,%eax
6b: pop %rbx
6c: pop %r15
6e: pop %r14
70: pop %r13
72: pop %rbx
73: leaveq
74: retq
After; state after JIT:
# ./bpftool p d j i 1655
0xffffffffc08e8930:
0: nopl 0x0(%rax,%rax,1)
5: push %rbp
6: mov %rsp,%rbp
9: sub $0x200,%rsp
10: push %rbx
11: push %r13
13: push %r14
15: push %r15
17: pushq $0x0 _
19: xor %edx,%edx |_ index (arg 3)
1b: movabs $0xffff9d8afd74c000,%rsi |_ map (arg 2)
25: mov -0x224(%rbp),%eax | tail call limit check
2b: cmp $0x20,%eax |
2e: ja 0x000000000000003e |
30: add $0x1,%eax |
33: mov %eax,-0x224(%rbp) |_
39: jmpq 0xfffffffffffd1785 |_ [direct] goto *(prog->bpf_func + prologue_size)
3e: mov $0x1,%eax
43: pop %rbx
44: pop %r15
46: pop %r14
48: pop %r13
4a: pop %rbx
4b: leaveq
4c: retq
After; state after map update (target prog):
# ./bpftool p d j i 1655
0xffffffffc08e8930:
0: nopl 0x0(%rax,%rax,1)
5: push %rbp
6: mov %rsp,%rbp
9: sub $0x200,%rsp
10: push %rbx
11: push %r13
13: push %r14
15: push %r15
17: pushq $0x0
19: xor %edx,%edx
1b: movabs $0xffff9d8afd74c000,%rsi
25: mov -0x224(%rbp),%eax
2b: cmp $0x20,%eax .
2e: ja 0x000000000000003e .
30: add $0x1,%eax .
33: mov %eax,-0x224(%rbp) |_
39: jmpq 0xffffffffffb09f55 |_ goto *(prog->bpf_func + prologue_size)
3e: mov $0x1,%eax
43: pop %rbx
44: pop %r15
46: pop %r14
48: pop %r13
4a: pop %rbx
4b: leaveq
4c: retq
After; state after map update (no prog):
# ./bpftool p d j i 1655
0xffffffffc08e8930:
0: nopl 0x0(%rax,%rax,1)
5: push %rbp
6: mov %rsp,%rbp
9: sub $0x200,%rsp
10: push %rbx
11: push %r13
13: push %r14
15: push %r15
17: pushq $0x0
19: xor %edx,%edx
1b: movabs $0xffff9d8afd74c000,%rsi
25: mov -0x224(%rbp),%eax
2b: cmp $0x20,%eax .
2e: ja 0x000000000000003e .
30: add $0x1,%eax .
33: mov %eax,-0x224(%rbp) |_
39: nopl 0x0(%rax,%rax,1) |_ fall-through nop
3e: mov $0x1,%eax
43: pop %rbx
44: pop %r15
46: pop %r14
48: pop %r13
4a: pop %rbx
4b: leaveq
4c: retq
Nice bonus is that this also shrinks the code emission quite a bit
for every tail call invocation.
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Link: https://lore.kernel.org/bpf/6ada4c1c9d35eeb5f4ecfab94593dafa6b5c4b09.1574452833.git.daniel@iogearbox.net
2019-11-22 20:08:00 +00:00
|
|
|
(u8 *)target->bpf_func +
|
2022-02-04 18:57:42 +00:00
|
|
|
poke->adj_off);
|
bpf, x86: Emit patchable direct jump as tail call
Add initial code emission for *direct* jumps for tail call maps in
order to avoid the retpoline overhead from a493a87f38cf ("bpf, x64:
implement retpoline for tail call") for situations that allow for
it, meaning, for known constant keys at verification time which are
used as index into the tail call map. In case of Cilium which makes
heavy use of tail calls, constant keys are used in the vast majority,
only for a single occurrence we use a dynamic key.
High level outline is that if the target prog is NULL in the map, we
emit a 5-byte nop for the fall-through case and if not, we emit a
5-byte direct relative jmp to the target bpf_func + skipped prologue
offset. Later during runtime, we patch these 5-byte nop/jmps upon
tail call map update or deletions dynamically. Note that on x86-64
the direct jmp works as we reuse the same stack frame and skip
prologue (as opposed to some other JIT implementations).
One of the issues is that the tail call map slots can change at any
given time even during JITing. Therefore, we have two passes: i) emit
nops for all patchable locations during main JITing phase until we
declare prog->jited = 1 eventually. At this point the image is stable,
not public yet and with all jmps disabled. While JITing, we collect
additional info like poke->ip in order to remember the patch location
for later modifications. In ii) bpf_tail_call_direct_fixup() walks
over the progs poke_tab, locks the tail call maps poke_mutex to
prevent from parallel updates and patches in the right locations via
__bpf_arch_text_poke(). Note, the main bpf_arch_text_poke() cannot
be used at this point since we're not yet exposed to kallsyms. For
the update we use plain memcpy() since the image is not public and
still in read-write mode. After patching, we activate that poke entry
through poke->ip_stable. Meaning, at this point any tail call map
updates/deletions are not going to ignore that poke entry anymore.
Then, bpf_arch_text_poke() might still occur on the read-write image
until we finally locked it as read-only. Both modifications on the
given image are under text_mutex to avoid interference with each
other when update requests come in in parallel for different tail
call maps (current one we have locked in JIT and different one where
poke->ip_stable was already set).
Example prog:
# ./bpftool p d x i 1655
0: (b7) r3 = 0
1: (18) r2 = map[id:526]
3: (85) call bpf_tail_call#12
4: (b7) r0 = 1
5: (95) exit
Before:
# ./bpftool p d j i 1655
0xffffffffc076e55c:
0: nopl 0x0(%rax,%rax,1)
5: push %rbp
6: mov %rsp,%rbp
9: sub $0x200,%rsp
10: push %rbx
11: push %r13
13: push %r14
15: push %r15
17: pushq $0x0 _
19: xor %edx,%edx |_ index (arg 3)
1b: movabs $0xffff88d95cc82600,%rsi |_ map (arg 2)
25: mov %edx,%edx | index >= array->map.max_entries
27: cmp %edx,0x24(%rsi) |
2a: jbe 0x0000000000000066 |_
2c: mov -0x224(%rbp),%eax | tail call limit check
32: cmp $0x20,%eax |
35: ja 0x0000000000000066 |
37: add $0x1,%eax |
3a: mov %eax,-0x224(%rbp) |_
40: mov 0xd0(%rsi,%rdx,8),%rax |_ prog = array->ptrs[index]
48: test %rax,%rax | prog == NULL check
4b: je 0x0000000000000066 |_
4d: mov 0x30(%rax),%rax | goto *(prog->bpf_func + prologue_size)
51: add $0x19,%rax |
55: callq 0x0000000000000061 | retpoline for indirect jump
5a: pause |
5c: lfence |
5f: jmp 0x000000000000005a |
61: mov %rax,(%rsp) |
65: retq |_
66: mov $0x1,%eax
6b: pop %rbx
6c: pop %r15
6e: pop %r14
70: pop %r13
72: pop %rbx
73: leaveq
74: retq
After; state after JIT:
# ./bpftool p d j i 1655
0xffffffffc08e8930:
0: nopl 0x0(%rax,%rax,1)
5: push %rbp
6: mov %rsp,%rbp
9: sub $0x200,%rsp
10: push %rbx
11: push %r13
13: push %r14
15: push %r15
17: pushq $0x0 _
19: xor %edx,%edx |_ index (arg 3)
1b: movabs $0xffff9d8afd74c000,%rsi |_ map (arg 2)
25: mov -0x224(%rbp),%eax | tail call limit check
2b: cmp $0x20,%eax |
2e: ja 0x000000000000003e |
30: add $0x1,%eax |
33: mov %eax,-0x224(%rbp) |_
39: jmpq 0xfffffffffffd1785 |_ [direct] goto *(prog->bpf_func + prologue_size)
3e: mov $0x1,%eax
43: pop %rbx
44: pop %r15
46: pop %r14
48: pop %r13
4a: pop %rbx
4b: leaveq
4c: retq
After; state after map update (target prog):
# ./bpftool p d j i 1655
0xffffffffc08e8930:
0: nopl 0x0(%rax,%rax,1)
5: push %rbp
6: mov %rsp,%rbp
9: sub $0x200,%rsp
10: push %rbx
11: push %r13
13: push %r14
15: push %r15
17: pushq $0x0
19: xor %edx,%edx
1b: movabs $0xffff9d8afd74c000,%rsi
25: mov -0x224(%rbp),%eax
2b: cmp $0x20,%eax .
2e: ja 0x000000000000003e .
30: add $0x1,%eax .
33: mov %eax,-0x224(%rbp) |_
39: jmpq 0xffffffffffb09f55 |_ goto *(prog->bpf_func + prologue_size)
3e: mov $0x1,%eax
43: pop %rbx
44: pop %r15
46: pop %r14
48: pop %r13
4a: pop %rbx
4b: leaveq
4c: retq
After; state after map update (no prog):
# ./bpftool p d j i 1655
0xffffffffc08e8930:
0: nopl 0x0(%rax,%rax,1)
5: push %rbp
6: mov %rsp,%rbp
9: sub $0x200,%rsp
10: push %rbx
11: push %r13
13: push %r14
15: push %r15
17: pushq $0x0
19: xor %edx,%edx
1b: movabs $0xffff9d8afd74c000,%rsi
25: mov -0x224(%rbp),%eax
2b: cmp $0x20,%eax .
2e: ja 0x000000000000003e .
30: add $0x1,%eax .
33: mov %eax,-0x224(%rbp) |_
39: nopl 0x0(%rax,%rax,1) |_ fall-through nop
3e: mov $0x1,%eax
43: pop %rbx
44: pop %r15
46: pop %r14
48: pop %r13
4a: pop %rbx
4b: leaveq
4c: retq
Nice bonus is that this also shrinks the code emission quite a bit
for every tail call invocation.
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Link: https://lore.kernel.org/bpf/6ada4c1c9d35eeb5f4ecfab94593dafa6b5c4b09.1574452833.git.daniel@iogearbox.net
2019-11-22 20:08:00 +00:00
|
|
|
BUG_ON(ret < 0);
|
bpf, x64: rework pro/epilogue and tailcall handling in JIT
This commit serves two things:
1) it optimizes BPF prologue/epilogue generation
2) it makes possible to have tailcalls within BPF subprogram
Both points are related to each other since without 1), 2) could not be
achieved.
In [1], Alexei says:
"The prologue will look like:
nop5
xor eax,eax // two new bytes if bpf_tail_call() is used in this
// function
push rbp
mov rbp, rsp
sub rsp, rounded_stack_depth
push rax // zero init tail_call counter
variable number of push rbx,r13,r14,r15
Then bpf_tail_call will pop variable number rbx,..
and final 'pop rax'
Then 'add rsp, size_of_current_stack_frame'
jmp to next function and skip over 'nop5; xor eax,eax; push rpb; mov
rbp, rsp'
This way new function will set its own stack size and will init tail
call
counter with whatever value the parent had.
If next function doesn't use bpf_tail_call it won't have 'xor eax,eax'.
Instead it would need to have 'nop2' in there."
Implement that suggestion.
Since the layout of stack is changed, tail call counter handling can not
rely anymore on popping it to rbx just like it have been handled for
constant prologue case and later overwrite of rbx with actual value of
rbx pushed to stack. Therefore, let's use one of the register (%rcx) that
is considered to be volatile/caller-saved and pop the value of tail call
counter in there in the epilogue.
Drop the BUILD_BUG_ON in emit_prologue and in
emit_bpf_tail_call_indirect where instruction layout is not constant
anymore.
Introduce new poke target, 'tailcall_bypass' to poke descriptor that is
dedicated for skipping the register pops and stack unwind that are
generated right before the actual jump to target program.
For case when the target program is not present, BPF program will skip
the pop instructions and nop5 dedicated for jmpq $target. An example of
such state when only R6 of callee saved registers is used by program:
ffffffffc0513aa1: e9 0e 00 00 00 jmpq 0xffffffffc0513ab4
ffffffffc0513aa6: 5b pop %rbx
ffffffffc0513aa7: 58 pop %rax
ffffffffc0513aa8: 48 81 c4 00 00 00 00 add $0x0,%rsp
ffffffffc0513aaf: 0f 1f 44 00 00 nopl 0x0(%rax,%rax,1)
ffffffffc0513ab4: 48 89 df mov %rbx,%rdi
When target program is inserted, the jump that was there to skip
pops/nop5 will become the nop5, so CPU will go over pops and do the
actual tailcall.
One might ask why there simply can not be pushes after the nop5?
In the following example snippet:
ffffffffc037030c: 48 89 fb mov %rdi,%rbx
(...)
ffffffffc0370332: 5b pop %rbx
ffffffffc0370333: 58 pop %rax
ffffffffc0370334: 48 81 c4 00 00 00 00 add $0x0,%rsp
ffffffffc037033b: 0f 1f 44 00 00 nopl 0x0(%rax,%rax,1)
ffffffffc0370340: 48 81 ec 00 00 00 00 sub $0x0,%rsp
ffffffffc0370347: 50 push %rax
ffffffffc0370348: 53 push %rbx
ffffffffc0370349: 48 89 df mov %rbx,%rdi
ffffffffc037034c: e8 f7 21 00 00 callq 0xffffffffc0372548
There is the bpf2bpf call (at ffffffffc037034c) right after the tailcall
and jump target is not present. ctx is in %rbx register and BPF
subprogram that we will call into on ffffffffc037034c is relying on it,
e.g. it will pick ctx from there. Such code layout is therefore broken
as we would overwrite the content of %rbx with the value that was pushed
on the prologue. That is the reason for the 'bypass' approach.
Special care needs to be taken during the install/update/remove of
tailcall target. In case when target program is not present, the CPU
must not execute the pop instructions that precede the tailcall.
To address that, the following states can be defined:
A nop, unwind, nop
B nop, unwind, tail
C skip, unwind, nop
D skip, unwind, tail
A is forbidden (lead to incorrectness). The state transitions between
tailcall install/update/remove will work as follows:
First install tail call f: C->D->B(f)
* poke the tailcall, after that get rid of the skip
Update tail call f to f': B(f)->B(f')
* poke the tailcall (poke->tailcall_target) and do NOT touch the
poke->tailcall_bypass
Remove tail call: B(f')->C(f')
* poke->tailcall_bypass is poked back to jump, then we wait the RCU
grace period so that other programs will finish its execution and
after that we are safe to remove the poke->tailcall_target
Install new tail call (f''): C(f')->D(f'')->B(f'').
* same as first step
This way CPU can never be exposed to "unwind, tail" state.
Last but not least, when tailcalls get mixed with bpf2bpf calls, it
would be possible to encounter the endless loop due to clearing the
tailcall counter if for example we would use the tailcall3-like from BPF
selftests program that would be subprogram-based, meaning the tailcall
would be present within the BPF subprogram.
This test, broken down to particular steps, would do:
entry -> set tailcall counter to 0, bump it by 1, tailcall to func0
func0 -> call subprog_tail
(we are NOT skipping the first 11 bytes of prologue and this subprogram
has a tailcall, therefore we clear the counter...)
subprog -> do the same thing as entry
and then loop forever.
To address this, the idea is to go through the call chain of bpf2bpf progs
and look for a tailcall presence throughout whole chain. If we saw a single
tail call then each node in this call chain needs to be marked as a subprog
that can reach the tailcall. We would later feed the JIT with this info
and:
- set eax to 0 only when tailcall is reachable and this is the entry prog
- if tailcall is reachable but there's no tailcall in insns of currently
JITed prog then push rax anyway, so that it will be possible to
propagate further down the call chain
- finally if tailcall is reachable, then we need to precede the 'call'
insn with mov rax, [rbp - (stack_depth + 8)]
Tail call related cases from test_verifier kselftest are also working
fine. Sample BPF programs that utilize tail calls (sockex3, tracex5)
work properly as well.
[1]: https://lore.kernel.org/bpf/20200517043227.2gpq22ifoq37ogst@ast-mbp.dhcp.thefacebook.com/
Suggested-by: Alexei Starovoitov <ast@kernel.org>
Signed-off-by: Maciej Fijalkowski <maciej.fijalkowski@intel.com>
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
2020-09-16 21:10:08 +00:00
|
|
|
ret = __bpf_arch_text_poke(poke->tailcall_bypass,
|
|
|
|
BPF_MOD_JUMP,
|
|
|
|
(u8 *)poke->tailcall_target +
|
2022-02-04 18:57:42 +00:00
|
|
|
X86_PATCH_SIZE, NULL);
|
bpf, x64: rework pro/epilogue and tailcall handling in JIT
This commit serves two things:
1) it optimizes BPF prologue/epilogue generation
2) it makes possible to have tailcalls within BPF subprogram
Both points are related to each other since without 1), 2) could not be
achieved.
In [1], Alexei says:
"The prologue will look like:
nop5
xor eax,eax // two new bytes if bpf_tail_call() is used in this
// function
push rbp
mov rbp, rsp
sub rsp, rounded_stack_depth
push rax // zero init tail_call counter
variable number of push rbx,r13,r14,r15
Then bpf_tail_call will pop variable number rbx,..
and final 'pop rax'
Then 'add rsp, size_of_current_stack_frame'
jmp to next function and skip over 'nop5; xor eax,eax; push rpb; mov
rbp, rsp'
This way new function will set its own stack size and will init tail
call
counter with whatever value the parent had.
If next function doesn't use bpf_tail_call it won't have 'xor eax,eax'.
Instead it would need to have 'nop2' in there."
Implement that suggestion.
Since the layout of stack is changed, tail call counter handling can not
rely anymore on popping it to rbx just like it have been handled for
constant prologue case and later overwrite of rbx with actual value of
rbx pushed to stack. Therefore, let's use one of the register (%rcx) that
is considered to be volatile/caller-saved and pop the value of tail call
counter in there in the epilogue.
Drop the BUILD_BUG_ON in emit_prologue and in
emit_bpf_tail_call_indirect where instruction layout is not constant
anymore.
Introduce new poke target, 'tailcall_bypass' to poke descriptor that is
dedicated for skipping the register pops and stack unwind that are
generated right before the actual jump to target program.
For case when the target program is not present, BPF program will skip
the pop instructions and nop5 dedicated for jmpq $target. An example of
such state when only R6 of callee saved registers is used by program:
ffffffffc0513aa1: e9 0e 00 00 00 jmpq 0xffffffffc0513ab4
ffffffffc0513aa6: 5b pop %rbx
ffffffffc0513aa7: 58 pop %rax
ffffffffc0513aa8: 48 81 c4 00 00 00 00 add $0x0,%rsp
ffffffffc0513aaf: 0f 1f 44 00 00 nopl 0x0(%rax,%rax,1)
ffffffffc0513ab4: 48 89 df mov %rbx,%rdi
When target program is inserted, the jump that was there to skip
pops/nop5 will become the nop5, so CPU will go over pops and do the
actual tailcall.
One might ask why there simply can not be pushes after the nop5?
In the following example snippet:
ffffffffc037030c: 48 89 fb mov %rdi,%rbx
(...)
ffffffffc0370332: 5b pop %rbx
ffffffffc0370333: 58 pop %rax
ffffffffc0370334: 48 81 c4 00 00 00 00 add $0x0,%rsp
ffffffffc037033b: 0f 1f 44 00 00 nopl 0x0(%rax,%rax,1)
ffffffffc0370340: 48 81 ec 00 00 00 00 sub $0x0,%rsp
ffffffffc0370347: 50 push %rax
ffffffffc0370348: 53 push %rbx
ffffffffc0370349: 48 89 df mov %rbx,%rdi
ffffffffc037034c: e8 f7 21 00 00 callq 0xffffffffc0372548
There is the bpf2bpf call (at ffffffffc037034c) right after the tailcall
and jump target is not present. ctx is in %rbx register and BPF
subprogram that we will call into on ffffffffc037034c is relying on it,
e.g. it will pick ctx from there. Such code layout is therefore broken
as we would overwrite the content of %rbx with the value that was pushed
on the prologue. That is the reason for the 'bypass' approach.
Special care needs to be taken during the install/update/remove of
tailcall target. In case when target program is not present, the CPU
must not execute the pop instructions that precede the tailcall.
To address that, the following states can be defined:
A nop, unwind, nop
B nop, unwind, tail
C skip, unwind, nop
D skip, unwind, tail
A is forbidden (lead to incorrectness). The state transitions between
tailcall install/update/remove will work as follows:
First install tail call f: C->D->B(f)
* poke the tailcall, after that get rid of the skip
Update tail call f to f': B(f)->B(f')
* poke the tailcall (poke->tailcall_target) and do NOT touch the
poke->tailcall_bypass
Remove tail call: B(f')->C(f')
* poke->tailcall_bypass is poked back to jump, then we wait the RCU
grace period so that other programs will finish its execution and
after that we are safe to remove the poke->tailcall_target
Install new tail call (f''): C(f')->D(f'')->B(f'').
* same as first step
This way CPU can never be exposed to "unwind, tail" state.
Last but not least, when tailcalls get mixed with bpf2bpf calls, it
would be possible to encounter the endless loop due to clearing the
tailcall counter if for example we would use the tailcall3-like from BPF
selftests program that would be subprogram-based, meaning the tailcall
would be present within the BPF subprogram.
This test, broken down to particular steps, would do:
entry -> set tailcall counter to 0, bump it by 1, tailcall to func0
func0 -> call subprog_tail
(we are NOT skipping the first 11 bytes of prologue and this subprogram
has a tailcall, therefore we clear the counter...)
subprog -> do the same thing as entry
and then loop forever.
To address this, the idea is to go through the call chain of bpf2bpf progs
and look for a tailcall presence throughout whole chain. If we saw a single
tail call then each node in this call chain needs to be marked as a subprog
that can reach the tailcall. We would later feed the JIT with this info
and:
- set eax to 0 only when tailcall is reachable and this is the entry prog
- if tailcall is reachable but there's no tailcall in insns of currently
JITed prog then push rax anyway, so that it will be possible to
propagate further down the call chain
- finally if tailcall is reachable, then we need to precede the 'call'
insn with mov rax, [rbp - (stack_depth + 8)]
Tail call related cases from test_verifier kselftest are also working
fine. Sample BPF programs that utilize tail calls (sockex3, tracex5)
work properly as well.
[1]: https://lore.kernel.org/bpf/20200517043227.2gpq22ifoq37ogst@ast-mbp.dhcp.thefacebook.com/
Suggested-by: Alexei Starovoitov <ast@kernel.org>
Signed-off-by: Maciej Fijalkowski <maciej.fijalkowski@intel.com>
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
2020-09-16 21:10:08 +00:00
|
|
|
BUG_ON(ret < 0);
|
bpf, x86: Emit patchable direct jump as tail call
Add initial code emission for *direct* jumps for tail call maps in
order to avoid the retpoline overhead from a493a87f38cf ("bpf, x64:
implement retpoline for tail call") for situations that allow for
it, meaning, for known constant keys at verification time which are
used as index into the tail call map. In case of Cilium which makes
heavy use of tail calls, constant keys are used in the vast majority,
only for a single occurrence we use a dynamic key.
High level outline is that if the target prog is NULL in the map, we
emit a 5-byte nop for the fall-through case and if not, we emit a
5-byte direct relative jmp to the target bpf_func + skipped prologue
offset. Later during runtime, we patch these 5-byte nop/jmps upon
tail call map update or deletions dynamically. Note that on x86-64
the direct jmp works as we reuse the same stack frame and skip
prologue (as opposed to some other JIT implementations).
One of the issues is that the tail call map slots can change at any
given time even during JITing. Therefore, we have two passes: i) emit
nops for all patchable locations during main JITing phase until we
declare prog->jited = 1 eventually. At this point the image is stable,
not public yet and with all jmps disabled. While JITing, we collect
additional info like poke->ip in order to remember the patch location
for later modifications. In ii) bpf_tail_call_direct_fixup() walks
over the progs poke_tab, locks the tail call maps poke_mutex to
prevent from parallel updates and patches in the right locations via
__bpf_arch_text_poke(). Note, the main bpf_arch_text_poke() cannot
be used at this point since we're not yet exposed to kallsyms. For
the update we use plain memcpy() since the image is not public and
still in read-write mode. After patching, we activate that poke entry
through poke->ip_stable. Meaning, at this point any tail call map
updates/deletions are not going to ignore that poke entry anymore.
Then, bpf_arch_text_poke() might still occur on the read-write image
until we finally locked it as read-only. Both modifications on the
given image are under text_mutex to avoid interference with each
other when update requests come in in parallel for different tail
call maps (current one we have locked in JIT and different one where
poke->ip_stable was already set).
Example prog:
# ./bpftool p d x i 1655
0: (b7) r3 = 0
1: (18) r2 = map[id:526]
3: (85) call bpf_tail_call#12
4: (b7) r0 = 1
5: (95) exit
Before:
# ./bpftool p d j i 1655
0xffffffffc076e55c:
0: nopl 0x0(%rax,%rax,1)
5: push %rbp
6: mov %rsp,%rbp
9: sub $0x200,%rsp
10: push %rbx
11: push %r13
13: push %r14
15: push %r15
17: pushq $0x0 _
19: xor %edx,%edx |_ index (arg 3)
1b: movabs $0xffff88d95cc82600,%rsi |_ map (arg 2)
25: mov %edx,%edx | index >= array->map.max_entries
27: cmp %edx,0x24(%rsi) |
2a: jbe 0x0000000000000066 |_
2c: mov -0x224(%rbp),%eax | tail call limit check
32: cmp $0x20,%eax |
35: ja 0x0000000000000066 |
37: add $0x1,%eax |
3a: mov %eax,-0x224(%rbp) |_
40: mov 0xd0(%rsi,%rdx,8),%rax |_ prog = array->ptrs[index]
48: test %rax,%rax | prog == NULL check
4b: je 0x0000000000000066 |_
4d: mov 0x30(%rax),%rax | goto *(prog->bpf_func + prologue_size)
51: add $0x19,%rax |
55: callq 0x0000000000000061 | retpoline for indirect jump
5a: pause |
5c: lfence |
5f: jmp 0x000000000000005a |
61: mov %rax,(%rsp) |
65: retq |_
66: mov $0x1,%eax
6b: pop %rbx
6c: pop %r15
6e: pop %r14
70: pop %r13
72: pop %rbx
73: leaveq
74: retq
After; state after JIT:
# ./bpftool p d j i 1655
0xffffffffc08e8930:
0: nopl 0x0(%rax,%rax,1)
5: push %rbp
6: mov %rsp,%rbp
9: sub $0x200,%rsp
10: push %rbx
11: push %r13
13: push %r14
15: push %r15
17: pushq $0x0 _
19: xor %edx,%edx |_ index (arg 3)
1b: movabs $0xffff9d8afd74c000,%rsi |_ map (arg 2)
25: mov -0x224(%rbp),%eax | tail call limit check
2b: cmp $0x20,%eax |
2e: ja 0x000000000000003e |
30: add $0x1,%eax |
33: mov %eax,-0x224(%rbp) |_
39: jmpq 0xfffffffffffd1785 |_ [direct] goto *(prog->bpf_func + prologue_size)
3e: mov $0x1,%eax
43: pop %rbx
44: pop %r15
46: pop %r14
48: pop %r13
4a: pop %rbx
4b: leaveq
4c: retq
After; state after map update (target prog):
# ./bpftool p d j i 1655
0xffffffffc08e8930:
0: nopl 0x0(%rax,%rax,1)
5: push %rbp
6: mov %rsp,%rbp
9: sub $0x200,%rsp
10: push %rbx
11: push %r13
13: push %r14
15: push %r15
17: pushq $0x0
19: xor %edx,%edx
1b: movabs $0xffff9d8afd74c000,%rsi
25: mov -0x224(%rbp),%eax
2b: cmp $0x20,%eax .
2e: ja 0x000000000000003e .
30: add $0x1,%eax .
33: mov %eax,-0x224(%rbp) |_
39: jmpq 0xffffffffffb09f55 |_ goto *(prog->bpf_func + prologue_size)
3e: mov $0x1,%eax
43: pop %rbx
44: pop %r15
46: pop %r14
48: pop %r13
4a: pop %rbx
4b: leaveq
4c: retq
After; state after map update (no prog):
# ./bpftool p d j i 1655
0xffffffffc08e8930:
0: nopl 0x0(%rax,%rax,1)
5: push %rbp
6: mov %rsp,%rbp
9: sub $0x200,%rsp
10: push %rbx
11: push %r13
13: push %r14
15: push %r15
17: pushq $0x0
19: xor %edx,%edx
1b: movabs $0xffff9d8afd74c000,%rsi
25: mov -0x224(%rbp),%eax
2b: cmp $0x20,%eax .
2e: ja 0x000000000000003e .
30: add $0x1,%eax .
33: mov %eax,-0x224(%rbp) |_
39: nopl 0x0(%rax,%rax,1) |_ fall-through nop
3e: mov $0x1,%eax
43: pop %rbx
44: pop %r15
46: pop %r14
48: pop %r13
4a: pop %rbx
4b: leaveq
4c: retq
Nice bonus is that this also shrinks the code emission quite a bit
for every tail call invocation.
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Link: https://lore.kernel.org/bpf/6ada4c1c9d35eeb5f4ecfab94593dafa6b5c4b09.1574452833.git.daniel@iogearbox.net
2019-11-22 20:08:00 +00:00
|
|
|
}
|
2020-09-16 21:10:06 +00:00
|
|
|
WRITE_ONCE(poke->tailcall_target_stable, true);
|
bpf, x86: Emit patchable direct jump as tail call
Add initial code emission for *direct* jumps for tail call maps in
order to avoid the retpoline overhead from a493a87f38cf ("bpf, x64:
implement retpoline for tail call") for situations that allow for
it, meaning, for known constant keys at verification time which are
used as index into the tail call map. In case of Cilium which makes
heavy use of tail calls, constant keys are used in the vast majority,
only for a single occurrence we use a dynamic key.
High level outline is that if the target prog is NULL in the map, we
emit a 5-byte nop for the fall-through case and if not, we emit a
5-byte direct relative jmp to the target bpf_func + skipped prologue
offset. Later during runtime, we patch these 5-byte nop/jmps upon
tail call map update or deletions dynamically. Note that on x86-64
the direct jmp works as we reuse the same stack frame and skip
prologue (as opposed to some other JIT implementations).
One of the issues is that the tail call map slots can change at any
given time even during JITing. Therefore, we have two passes: i) emit
nops for all patchable locations during main JITing phase until we
declare prog->jited = 1 eventually. At this point the image is stable,
not public yet and with all jmps disabled. While JITing, we collect
additional info like poke->ip in order to remember the patch location
for later modifications. In ii) bpf_tail_call_direct_fixup() walks
over the progs poke_tab, locks the tail call maps poke_mutex to
prevent from parallel updates and patches in the right locations via
__bpf_arch_text_poke(). Note, the main bpf_arch_text_poke() cannot
be used at this point since we're not yet exposed to kallsyms. For
the update we use plain memcpy() since the image is not public and
still in read-write mode. After patching, we activate that poke entry
through poke->ip_stable. Meaning, at this point any tail call map
updates/deletions are not going to ignore that poke entry anymore.
Then, bpf_arch_text_poke() might still occur on the read-write image
until we finally locked it as read-only. Both modifications on the
given image are under text_mutex to avoid interference with each
other when update requests come in in parallel for different tail
call maps (current one we have locked in JIT and different one where
poke->ip_stable was already set).
Example prog:
# ./bpftool p d x i 1655
0: (b7) r3 = 0
1: (18) r2 = map[id:526]
3: (85) call bpf_tail_call#12
4: (b7) r0 = 1
5: (95) exit
Before:
# ./bpftool p d j i 1655
0xffffffffc076e55c:
0: nopl 0x0(%rax,%rax,1)
5: push %rbp
6: mov %rsp,%rbp
9: sub $0x200,%rsp
10: push %rbx
11: push %r13
13: push %r14
15: push %r15
17: pushq $0x0 _
19: xor %edx,%edx |_ index (arg 3)
1b: movabs $0xffff88d95cc82600,%rsi |_ map (arg 2)
25: mov %edx,%edx | index >= array->map.max_entries
27: cmp %edx,0x24(%rsi) |
2a: jbe 0x0000000000000066 |_
2c: mov -0x224(%rbp),%eax | tail call limit check
32: cmp $0x20,%eax |
35: ja 0x0000000000000066 |
37: add $0x1,%eax |
3a: mov %eax,-0x224(%rbp) |_
40: mov 0xd0(%rsi,%rdx,8),%rax |_ prog = array->ptrs[index]
48: test %rax,%rax | prog == NULL check
4b: je 0x0000000000000066 |_
4d: mov 0x30(%rax),%rax | goto *(prog->bpf_func + prologue_size)
51: add $0x19,%rax |
55: callq 0x0000000000000061 | retpoline for indirect jump
5a: pause |
5c: lfence |
5f: jmp 0x000000000000005a |
61: mov %rax,(%rsp) |
65: retq |_
66: mov $0x1,%eax
6b: pop %rbx
6c: pop %r15
6e: pop %r14
70: pop %r13
72: pop %rbx
73: leaveq
74: retq
After; state after JIT:
# ./bpftool p d j i 1655
0xffffffffc08e8930:
0: nopl 0x0(%rax,%rax,1)
5: push %rbp
6: mov %rsp,%rbp
9: sub $0x200,%rsp
10: push %rbx
11: push %r13
13: push %r14
15: push %r15
17: pushq $0x0 _
19: xor %edx,%edx |_ index (arg 3)
1b: movabs $0xffff9d8afd74c000,%rsi |_ map (arg 2)
25: mov -0x224(%rbp),%eax | tail call limit check
2b: cmp $0x20,%eax |
2e: ja 0x000000000000003e |
30: add $0x1,%eax |
33: mov %eax,-0x224(%rbp) |_
39: jmpq 0xfffffffffffd1785 |_ [direct] goto *(prog->bpf_func + prologue_size)
3e: mov $0x1,%eax
43: pop %rbx
44: pop %r15
46: pop %r14
48: pop %r13
4a: pop %rbx
4b: leaveq
4c: retq
After; state after map update (target prog):
# ./bpftool p d j i 1655
0xffffffffc08e8930:
0: nopl 0x0(%rax,%rax,1)
5: push %rbp
6: mov %rsp,%rbp
9: sub $0x200,%rsp
10: push %rbx
11: push %r13
13: push %r14
15: push %r15
17: pushq $0x0
19: xor %edx,%edx
1b: movabs $0xffff9d8afd74c000,%rsi
25: mov -0x224(%rbp),%eax
2b: cmp $0x20,%eax .
2e: ja 0x000000000000003e .
30: add $0x1,%eax .
33: mov %eax,-0x224(%rbp) |_
39: jmpq 0xffffffffffb09f55 |_ goto *(prog->bpf_func + prologue_size)
3e: mov $0x1,%eax
43: pop %rbx
44: pop %r15
46: pop %r14
48: pop %r13
4a: pop %rbx
4b: leaveq
4c: retq
After; state after map update (no prog):
# ./bpftool p d j i 1655
0xffffffffc08e8930:
0: nopl 0x0(%rax,%rax,1)
5: push %rbp
6: mov %rsp,%rbp
9: sub $0x200,%rsp
10: push %rbx
11: push %r13
13: push %r14
15: push %r15
17: pushq $0x0
19: xor %edx,%edx
1b: movabs $0xffff9d8afd74c000,%rsi
25: mov -0x224(%rbp),%eax
2b: cmp $0x20,%eax .
2e: ja 0x000000000000003e .
30: add $0x1,%eax .
33: mov %eax,-0x224(%rbp) |_
39: nopl 0x0(%rax,%rax,1) |_ fall-through nop
3e: mov $0x1,%eax
43: pop %rbx
44: pop %r15
46: pop %r14
48: pop %r13
4a: pop %rbx
4b: leaveq
4c: retq
Nice bonus is that this also shrinks the code emission quite a bit
for every tail call invocation.
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Link: https://lore.kernel.org/bpf/6ada4c1c9d35eeb5f4ecfab94593dafa6b5c4b09.1574452833.git.daniel@iogearbox.net
2019-11-22 20:08:00 +00:00
|
|
|
mutex_unlock(&array->aux->poke_mutex);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2018-02-24 00:07:59 +00:00
|
|
|
static void emit_mov_imm32(u8 **pprog, bool sign_propagate,
|
|
|
|
u32 dst_reg, const u32 imm32)
|
|
|
|
{
|
|
|
|
u8 *prog = *pprog;
|
|
|
|
u8 b1, b2, b3;
|
|
|
|
|
x86/bpf: Clean up non-standard comments, to make the code more readable
So by chance I looked into x86 assembly in arch/x86/net/bpf_jit_comp.c and
noticed the weird and inconsistent comment style it mistakenly learned from
the networking code:
/* Multi-line comment ...
* ... looks like this.
*/
Fix this to use the standard comment style specified in Documentation/CodingStyle
and used in arch/x86/ as well:
/*
* Multi-line comment ...
* ... looks like this.
*/
Also, to quote Linus's ... more explicit views about this:
http://article.gmane.org/gmane.linux.kernel.cryptoapi/21066
> But no, the networking code picked *none* of the above sane formats.
> Instead, it picked these two models that are just half-arsed
> shit-for-brains:
>
> (no)
> /* This is disgusting drug-induced
> * crap, and should die
> */
>
> (no-no-no)
> /* This is also very nasty
> * and visually unbalanced */
>
> Please. The networking code actually has the *worst* possible comment
> style. You can literally find that (no-no-no) style, which is just
> really horribly disgusting and worse than the otherwise fairly similar
> (d) in pretty much every way.
Also improve the comments and some other details while at it:
- Don't mix same-line and previous-line comment style on otherwise
identical code patterns within the same function,
- capitalize 'BPF' and x86 register names consistently,
- capitalize sentences consistently,
- instead of 'x64' use 'x86-64': x64 is a Microsoft specific term,
- use more consistent punctuation,
- use standard coding style in macros as well,
- fix typos and a few other minor details.
Consistent coding style is not optional, at least in arch/x86/.
No change in functionality.
( In case this commit causes conflicts with pending development code
I'll be glad to help resolve any conflicts! )
Acked-by: Thomas Gleixner <tglx@linutronix.de>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Borislav Petkov <bp@alien8.de>
Cc: H. Peter Anvin <hpa@zytor.com>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: David S. Miller <davem@davemloft.net>
Cc: Eric Dumazet <edumazet@google.com>
Cc: Daniel Borkmann <daniel@iogearbox.net>
Cc: Alexei Starovoitov <ast@fb.com>
Cc: Hideaki YOSHIFUJI <yoshfuji@linux-ipv6.org>
Cc: netdev@vger.kernel.org
Cc: linux-kernel@vger.kernel.org
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
2018-04-27 09:54:40 +00:00
|
|
|
/*
|
|
|
|
* Optimization: if imm32 is positive, use 'mov %eax, imm32'
|
2018-02-24 00:07:59 +00:00
|
|
|
* (which zero-extends imm32) to save 2 bytes.
|
|
|
|
*/
|
|
|
|
if (sign_propagate && (s32)imm32 < 0) {
|
|
|
|
/* 'mov %rax, imm32' sign extends imm32 */
|
|
|
|
b1 = add_1mod(0x48, dst_reg);
|
|
|
|
b2 = 0xC7;
|
|
|
|
b3 = 0xC0;
|
|
|
|
EMIT3_off32(b1, b2, add_1reg(b3, dst_reg), imm32);
|
|
|
|
goto done;
|
|
|
|
}
|
|
|
|
|
x86/bpf: Clean up non-standard comments, to make the code more readable
So by chance I looked into x86 assembly in arch/x86/net/bpf_jit_comp.c and
noticed the weird and inconsistent comment style it mistakenly learned from
the networking code:
/* Multi-line comment ...
* ... looks like this.
*/
Fix this to use the standard comment style specified in Documentation/CodingStyle
and used in arch/x86/ as well:
/*
* Multi-line comment ...
* ... looks like this.
*/
Also, to quote Linus's ... more explicit views about this:
http://article.gmane.org/gmane.linux.kernel.cryptoapi/21066
> But no, the networking code picked *none* of the above sane formats.
> Instead, it picked these two models that are just half-arsed
> shit-for-brains:
>
> (no)
> /* This is disgusting drug-induced
> * crap, and should die
> */
>
> (no-no-no)
> /* This is also very nasty
> * and visually unbalanced */
>
> Please. The networking code actually has the *worst* possible comment
> style. You can literally find that (no-no-no) style, which is just
> really horribly disgusting and worse than the otherwise fairly similar
> (d) in pretty much every way.
Also improve the comments and some other details while at it:
- Don't mix same-line and previous-line comment style on otherwise
identical code patterns within the same function,
- capitalize 'BPF' and x86 register names consistently,
- capitalize sentences consistently,
- instead of 'x64' use 'x86-64': x64 is a Microsoft specific term,
- use more consistent punctuation,
- use standard coding style in macros as well,
- fix typos and a few other minor details.
Consistent coding style is not optional, at least in arch/x86/.
No change in functionality.
( In case this commit causes conflicts with pending development code
I'll be glad to help resolve any conflicts! )
Acked-by: Thomas Gleixner <tglx@linutronix.de>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Borislav Petkov <bp@alien8.de>
Cc: H. Peter Anvin <hpa@zytor.com>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: David S. Miller <davem@davemloft.net>
Cc: Eric Dumazet <edumazet@google.com>
Cc: Daniel Borkmann <daniel@iogearbox.net>
Cc: Alexei Starovoitov <ast@fb.com>
Cc: Hideaki YOSHIFUJI <yoshfuji@linux-ipv6.org>
Cc: netdev@vger.kernel.org
Cc: linux-kernel@vger.kernel.org
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
2018-04-27 09:54:40 +00:00
|
|
|
/*
|
|
|
|
* Optimization: if imm32 is zero, use 'xor %eax, %eax'
|
2018-02-24 00:07:59 +00:00
|
|
|
* to save 3 bytes.
|
|
|
|
*/
|
|
|
|
if (imm32 == 0) {
|
|
|
|
if (is_ereg(dst_reg))
|
|
|
|
EMIT1(add_2mod(0x40, dst_reg, dst_reg));
|
|
|
|
b2 = 0x31; /* xor */
|
|
|
|
b3 = 0xC0;
|
|
|
|
EMIT2(b2, add_2reg(b3, dst_reg, dst_reg));
|
|
|
|
goto done;
|
|
|
|
}
|
|
|
|
|
|
|
|
/* mov %eax, imm32 */
|
|
|
|
if (is_ereg(dst_reg))
|
|
|
|
EMIT1(add_1mod(0x40, dst_reg));
|
|
|
|
EMIT1_off32(add_1reg(0xB8, dst_reg), imm32);
|
|
|
|
done:
|
|
|
|
*pprog = prog;
|
|
|
|
}
|
|
|
|
|
|
|
|
static void emit_mov_imm64(u8 **pprog, u32 dst_reg,
|
|
|
|
const u32 imm32_hi, const u32 imm32_lo)
|
|
|
|
{
|
|
|
|
u8 *prog = *pprog;
|
|
|
|
|
|
|
|
if (is_uimm32(((u64)imm32_hi << 32) | (u32)imm32_lo)) {
|
x86/bpf: Clean up non-standard comments, to make the code more readable
So by chance I looked into x86 assembly in arch/x86/net/bpf_jit_comp.c and
noticed the weird and inconsistent comment style it mistakenly learned from
the networking code:
/* Multi-line comment ...
* ... looks like this.
*/
Fix this to use the standard comment style specified in Documentation/CodingStyle
and used in arch/x86/ as well:
/*
* Multi-line comment ...
* ... looks like this.
*/
Also, to quote Linus's ... more explicit views about this:
http://article.gmane.org/gmane.linux.kernel.cryptoapi/21066
> But no, the networking code picked *none* of the above sane formats.
> Instead, it picked these two models that are just half-arsed
> shit-for-brains:
>
> (no)
> /* This is disgusting drug-induced
> * crap, and should die
> */
>
> (no-no-no)
> /* This is also very nasty
> * and visually unbalanced */
>
> Please. The networking code actually has the *worst* possible comment
> style. You can literally find that (no-no-no) style, which is just
> really horribly disgusting and worse than the otherwise fairly similar
> (d) in pretty much every way.
Also improve the comments and some other details while at it:
- Don't mix same-line and previous-line comment style on otherwise
identical code patterns within the same function,
- capitalize 'BPF' and x86 register names consistently,
- capitalize sentences consistently,
- instead of 'x64' use 'x86-64': x64 is a Microsoft specific term,
- use more consistent punctuation,
- use standard coding style in macros as well,
- fix typos and a few other minor details.
Consistent coding style is not optional, at least in arch/x86/.
No change in functionality.
( In case this commit causes conflicts with pending development code
I'll be glad to help resolve any conflicts! )
Acked-by: Thomas Gleixner <tglx@linutronix.de>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Borislav Petkov <bp@alien8.de>
Cc: H. Peter Anvin <hpa@zytor.com>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: David S. Miller <davem@davemloft.net>
Cc: Eric Dumazet <edumazet@google.com>
Cc: Daniel Borkmann <daniel@iogearbox.net>
Cc: Alexei Starovoitov <ast@fb.com>
Cc: Hideaki YOSHIFUJI <yoshfuji@linux-ipv6.org>
Cc: netdev@vger.kernel.org
Cc: linux-kernel@vger.kernel.org
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
2018-04-27 09:54:40 +00:00
|
|
|
/*
|
|
|
|
* For emitting plain u32, where sign bit must not be
|
2018-02-24 00:07:59 +00:00
|
|
|
* propagated LLVM tends to load imm64 over mov32
|
|
|
|
* directly, so save couple of bytes by just doing
|
|
|
|
* 'mov %eax, imm32' instead.
|
|
|
|
*/
|
|
|
|
emit_mov_imm32(&prog, false, dst_reg, imm32_lo);
|
|
|
|
} else {
|
2022-09-26 15:33:37 +00:00
|
|
|
/* movabsq rax, imm64 */
|
2018-02-24 00:07:59 +00:00
|
|
|
EMIT2(add_1mod(0x48, dst_reg), add_1reg(0xB8, dst_reg));
|
|
|
|
EMIT(imm32_lo, 4);
|
|
|
|
EMIT(imm32_hi, 4);
|
|
|
|
}
|
|
|
|
|
|
|
|
*pprog = prog;
|
|
|
|
}
|
|
|
|
|
2018-02-24 00:08:01 +00:00
|
|
|
static void emit_mov_reg(u8 **pprog, bool is64, u32 dst_reg, u32 src_reg)
|
|
|
|
{
|
|
|
|
u8 *prog = *pprog;
|
|
|
|
|
|
|
|
if (is64) {
|
|
|
|
/* mov dst, src */
|
|
|
|
EMIT_mov(dst_reg, src_reg);
|
|
|
|
} else {
|
|
|
|
/* mov32 dst, src */
|
|
|
|
if (is_ereg(dst_reg) || is_ereg(src_reg))
|
|
|
|
EMIT1(add_2mod(0x40, dst_reg, src_reg));
|
|
|
|
EMIT2(0x89, add_2reg(0xC0, dst_reg, src_reg));
|
|
|
|
}
|
|
|
|
|
|
|
|
*pprog = prog;
|
|
|
|
}
|
|
|
|
|
2023-07-28 01:12:02 +00:00
|
|
|
static void emit_movsx_reg(u8 **pprog, int num_bits, bool is64, u32 dst_reg,
|
|
|
|
u32 src_reg)
|
|
|
|
{
|
|
|
|
u8 *prog = *pprog;
|
|
|
|
|
|
|
|
if (is64) {
|
|
|
|
/* movs[b,w,l]q dst, src */
|
|
|
|
if (num_bits == 8)
|
|
|
|
EMIT4(add_2mod(0x48, src_reg, dst_reg), 0x0f, 0xbe,
|
|
|
|
add_2reg(0xC0, src_reg, dst_reg));
|
|
|
|
else if (num_bits == 16)
|
|
|
|
EMIT4(add_2mod(0x48, src_reg, dst_reg), 0x0f, 0xbf,
|
|
|
|
add_2reg(0xC0, src_reg, dst_reg));
|
|
|
|
else if (num_bits == 32)
|
|
|
|
EMIT3(add_2mod(0x48, src_reg, dst_reg), 0x63,
|
|
|
|
add_2reg(0xC0, src_reg, dst_reg));
|
|
|
|
} else {
|
|
|
|
/* movs[b,w]l dst, src */
|
|
|
|
if (num_bits == 8) {
|
|
|
|
EMIT4(add_2mod(0x40, src_reg, dst_reg), 0x0f, 0xbe,
|
|
|
|
add_2reg(0xC0, src_reg, dst_reg));
|
|
|
|
} else if (num_bits == 16) {
|
|
|
|
if (is_ereg(dst_reg) || is_ereg(src_reg))
|
|
|
|
EMIT1(add_2mod(0x40, src_reg, dst_reg));
|
|
|
|
EMIT3(add_2mod(0x0f, src_reg, dst_reg), 0xbf,
|
|
|
|
add_2reg(0xC0, src_reg, dst_reg));
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
*pprog = prog;
|
|
|
|
}
|
|
|
|
|
2021-01-14 18:17:41 +00:00
|
|
|
/* Emit the suffix (ModR/M etc) for addressing *(ptr_reg + off) and val_reg */
|
|
|
|
static void emit_insn_suffix(u8 **pprog, u32 ptr_reg, u32 val_reg, int off)
|
|
|
|
{
|
|
|
|
u8 *prog = *pprog;
|
|
|
|
|
|
|
|
if (is_imm8(off)) {
|
|
|
|
/* 1-byte signed displacement.
|
|
|
|
*
|
|
|
|
* If off == 0 we could skip this and save one extra byte, but
|
|
|
|
* special case of x86 R13 which always needs an offset is not
|
|
|
|
* worth the hassle
|
|
|
|
*/
|
|
|
|
EMIT2(add_2reg(0x40, ptr_reg, val_reg), off);
|
|
|
|
} else {
|
|
|
|
/* 4-byte signed displacement */
|
|
|
|
EMIT1_off32(add_2reg(0x80, ptr_reg, val_reg), off);
|
|
|
|
}
|
|
|
|
*pprog = prog;
|
|
|
|
}
|
|
|
|
|
2024-03-08 01:08:01 +00:00
|
|
|
static void emit_insn_suffix_SIB(u8 **pprog, u32 ptr_reg, u32 val_reg, u32 index_reg, int off)
|
|
|
|
{
|
|
|
|
u8 *prog = *pprog;
|
|
|
|
|
|
|
|
if (is_imm8(off)) {
|
|
|
|
EMIT3(add_2reg(0x44, BPF_REG_0, val_reg), add_2reg(0, ptr_reg, index_reg) /* SIB */, off);
|
|
|
|
} else {
|
|
|
|
EMIT2_off32(add_2reg(0x84, BPF_REG_0, val_reg), add_2reg(0, ptr_reg, index_reg) /* SIB */, off);
|
|
|
|
}
|
|
|
|
*pprog = prog;
|
|
|
|
}
|
|
|
|
|
2021-01-14 18:17:42 +00:00
|
|
|
/*
|
|
|
|
* Emit a REX byte if it will be necessary to address these registers
|
|
|
|
*/
|
|
|
|
static void maybe_emit_mod(u8 **pprog, u32 dst_reg, u32 src_reg, bool is64)
|
|
|
|
{
|
|
|
|
u8 *prog = *pprog;
|
|
|
|
|
|
|
|
if (is64)
|
|
|
|
EMIT1(add_2mod(0x48, dst_reg, src_reg));
|
|
|
|
else if (is_ereg(dst_reg) || is_ereg(src_reg))
|
|
|
|
EMIT1(add_2mod(0x40, dst_reg, src_reg));
|
|
|
|
*pprog = prog;
|
|
|
|
}
|
|
|
|
|
2021-10-06 19:41:35 +00:00
|
|
|
/*
|
|
|
|
* Similar version of maybe_emit_mod() for a single register
|
|
|
|
*/
|
|
|
|
static void maybe_emit_1mod(u8 **pprog, u32 reg, bool is64)
|
|
|
|
{
|
|
|
|
u8 *prog = *pprog;
|
|
|
|
|
|
|
|
if (is64)
|
|
|
|
EMIT1(add_1mod(0x48, reg));
|
|
|
|
else if (is_ereg(reg))
|
|
|
|
EMIT1(add_1mod(0x40, reg));
|
|
|
|
*pprog = prog;
|
|
|
|
}
|
|
|
|
|
2019-11-14 18:57:02 +00:00
|
|
|
/* LDX: dst_reg = *(u8*)(src_reg + off) */
|
|
|
|
static void emit_ldx(u8 **pprog, u32 size, u32 dst_reg, u32 src_reg, int off)
|
|
|
|
{
|
|
|
|
u8 *prog = *pprog;
|
|
|
|
|
|
|
|
switch (size) {
|
|
|
|
case BPF_B:
|
|
|
|
/* Emit 'movzx rax, byte ptr [rax + off]' */
|
|
|
|
EMIT3(add_2mod(0x48, src_reg, dst_reg), 0x0F, 0xB6);
|
|
|
|
break;
|
|
|
|
case BPF_H:
|
|
|
|
/* Emit 'movzx rax, word ptr [rax + off]' */
|
|
|
|
EMIT3(add_2mod(0x48, src_reg, dst_reg), 0x0F, 0xB7);
|
|
|
|
break;
|
|
|
|
case BPF_W:
|
|
|
|
/* Emit 'mov eax, dword ptr [rax+0x14]' */
|
|
|
|
if (is_ereg(dst_reg) || is_ereg(src_reg))
|
|
|
|
EMIT2(add_2mod(0x40, src_reg, dst_reg), 0x8B);
|
|
|
|
else
|
|
|
|
EMIT1(0x8B);
|
|
|
|
break;
|
|
|
|
case BPF_DW:
|
|
|
|
/* Emit 'mov rax, qword ptr [rax+0x14]' */
|
|
|
|
EMIT2(add_2mod(0x48, src_reg, dst_reg), 0x8B);
|
|
|
|
break;
|
|
|
|
}
|
2021-01-14 18:17:41 +00:00
|
|
|
emit_insn_suffix(&prog, src_reg, dst_reg, off);
|
2019-11-14 18:57:02 +00:00
|
|
|
*pprog = prog;
|
|
|
|
}
|
|
|
|
|
2023-07-28 01:11:56 +00:00
|
|
|
/* LDSX: dst_reg = *(s8*)(src_reg + off) */
|
|
|
|
static void emit_ldsx(u8 **pprog, u32 size, u32 dst_reg, u32 src_reg, int off)
|
|
|
|
{
|
|
|
|
u8 *prog = *pprog;
|
|
|
|
|
|
|
|
switch (size) {
|
|
|
|
case BPF_B:
|
|
|
|
/* Emit 'movsx rax, byte ptr [rax + off]' */
|
|
|
|
EMIT3(add_2mod(0x48, src_reg, dst_reg), 0x0F, 0xBE);
|
|
|
|
break;
|
|
|
|
case BPF_H:
|
|
|
|
/* Emit 'movsx rax, word ptr [rax + off]' */
|
|
|
|
EMIT3(add_2mod(0x48, src_reg, dst_reg), 0x0F, 0xBF);
|
|
|
|
break;
|
|
|
|
case BPF_W:
|
|
|
|
/* Emit 'movsx rax, dword ptr [rax+0x14]' */
|
|
|
|
EMIT2(add_2mod(0x48, src_reg, dst_reg), 0x63);
|
|
|
|
break;
|
|
|
|
}
|
|
|
|
emit_insn_suffix(&prog, src_reg, dst_reg, off);
|
|
|
|
*pprog = prog;
|
|
|
|
}
|
|
|
|
|
2024-03-08 01:08:01 +00:00
|
|
|
static void emit_ldx_index(u8 **pprog, u32 size, u32 dst_reg, u32 src_reg, u32 index_reg, int off)
|
|
|
|
{
|
|
|
|
u8 *prog = *pprog;
|
|
|
|
|
|
|
|
switch (size) {
|
|
|
|
case BPF_B:
|
|
|
|
/* movzx rax, byte ptr [rax + r12 + off] */
|
|
|
|
EMIT3(add_3mod(0x40, src_reg, dst_reg, index_reg), 0x0F, 0xB6);
|
|
|
|
break;
|
|
|
|
case BPF_H:
|
|
|
|
/* movzx rax, word ptr [rax + r12 + off] */
|
|
|
|
EMIT3(add_3mod(0x40, src_reg, dst_reg, index_reg), 0x0F, 0xB7);
|
|
|
|
break;
|
|
|
|
case BPF_W:
|
|
|
|
/* mov eax, dword ptr [rax + r12 + off] */
|
|
|
|
EMIT2(add_3mod(0x40, src_reg, dst_reg, index_reg), 0x8B);
|
|
|
|
break;
|
|
|
|
case BPF_DW:
|
|
|
|
/* mov rax, qword ptr [rax + r12 + off] */
|
|
|
|
EMIT2(add_3mod(0x48, src_reg, dst_reg, index_reg), 0x8B);
|
|
|
|
break;
|
|
|
|
}
|
|
|
|
emit_insn_suffix_SIB(&prog, src_reg, dst_reg, index_reg, off);
|
|
|
|
*pprog = prog;
|
|
|
|
}
|
|
|
|
|
|
|
|
static void emit_ldx_r12(u8 **pprog, u32 size, u32 dst_reg, u32 src_reg, int off)
|
|
|
|
{
|
|
|
|
emit_ldx_index(pprog, size, dst_reg, src_reg, X86_REG_R12, off);
|
|
|
|
}
|
|
|
|
|
2019-11-14 18:57:02 +00:00
|
|
|
/* STX: *(u8*)(dst_reg + off) = src_reg */
|
|
|
|
static void emit_stx(u8 **pprog, u32 size, u32 dst_reg, u32 src_reg, int off)
|
|
|
|
{
|
|
|
|
u8 *prog = *pprog;
|
|
|
|
|
|
|
|
switch (size) {
|
|
|
|
case BPF_B:
|
|
|
|
/* Emit 'mov byte ptr [rax + off], al' */
|
bpf, x86: Fix encoding for lower 8-bit registers in BPF_STX BPF_B
This patch fixes an encoding bug in emit_stx for BPF_B when the source
register is BPF_REG_FP.
The current implementation for BPF_STX BPF_B in emit_stx saves one REX
byte when the operands can be encoded using Mod-R/M alone. The lower 8
bits of registers %rax, %rbx, %rcx, and %rdx can be accessed without using
a REX prefix via %al, %bl, %cl, and %dl, respectively. Other registers,
(e.g., %rsi, %rdi, %rbp, %rsp) require a REX prefix to use their 8-bit
equivalents (%sil, %dil, %bpl, %spl).
The current code checks if the source for BPF_STX BPF_B is BPF_REG_1
or BPF_REG_2 (which map to %rdi and %rsi), in which case it emits the
required REX prefix. However, it misses the case when the source is
BPF_REG_FP (mapped to %rbp).
The result is that BPF_STX BPF_B with BPF_REG_FP as the source operand
will read from register %ch instead of the correct %bpl. This patch fixes
the problem by fixing and refactoring the check on which registers need
the extra REX byte. Since no BPF registers map to %rsp, there is no need
to handle %spl.
Fixes: 622582786c9e0 ("net: filter: x86: internal BPF JIT")
Signed-off-by: Xi Wang <xi.wang@gmail.com>
Signed-off-by: Luke Nelson <luke.r.nels@gmail.com>
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Link: https://lore.kernel.org/bpf/20200418232655.23870-1-luke.r.nels@gmail.com
2020-04-18 23:26:53 +00:00
|
|
|
if (is_ereg(dst_reg) || is_ereg_8l(src_reg))
|
|
|
|
/* Add extra byte for eregs or SIL,DIL,BPL in src_reg */
|
2019-11-14 18:57:02 +00:00
|
|
|
EMIT2(add_2mod(0x40, dst_reg, src_reg), 0x88);
|
|
|
|
else
|
|
|
|
EMIT1(0x88);
|
|
|
|
break;
|
|
|
|
case BPF_H:
|
|
|
|
if (is_ereg(dst_reg) || is_ereg(src_reg))
|
|
|
|
EMIT3(0x66, add_2mod(0x40, dst_reg, src_reg), 0x89);
|
|
|
|
else
|
|
|
|
EMIT2(0x66, 0x89);
|
|
|
|
break;
|
|
|
|
case BPF_W:
|
|
|
|
if (is_ereg(dst_reg) || is_ereg(src_reg))
|
|
|
|
EMIT2(add_2mod(0x40, dst_reg, src_reg), 0x89);
|
|
|
|
else
|
|
|
|
EMIT1(0x89);
|
|
|
|
break;
|
|
|
|
case BPF_DW:
|
|
|
|
EMIT2(add_2mod(0x48, dst_reg, src_reg), 0x89);
|
|
|
|
break;
|
|
|
|
}
|
2021-01-14 18:17:41 +00:00
|
|
|
emit_insn_suffix(&prog, dst_reg, src_reg, off);
|
2019-11-14 18:57:02 +00:00
|
|
|
*pprog = prog;
|
|
|
|
}
|
|
|
|
|
2024-03-08 01:08:01 +00:00
|
|
|
/* STX: *(u8*)(dst_reg + index_reg + off) = src_reg */
|
|
|
|
static void emit_stx_index(u8 **pprog, u32 size, u32 dst_reg, u32 src_reg, u32 index_reg, int off)
|
|
|
|
{
|
|
|
|
u8 *prog = *pprog;
|
|
|
|
|
|
|
|
switch (size) {
|
|
|
|
case BPF_B:
|
|
|
|
/* mov byte ptr [rax + r12 + off], al */
|
|
|
|
EMIT2(add_3mod(0x40, dst_reg, src_reg, index_reg), 0x88);
|
|
|
|
break;
|
|
|
|
case BPF_H:
|
|
|
|
/* mov word ptr [rax + r12 + off], ax */
|
|
|
|
EMIT3(0x66, add_3mod(0x40, dst_reg, src_reg, index_reg), 0x89);
|
|
|
|
break;
|
|
|
|
case BPF_W:
|
|
|
|
/* mov dword ptr [rax + r12 + 1], eax */
|
|
|
|
EMIT2(add_3mod(0x40, dst_reg, src_reg, index_reg), 0x89);
|
|
|
|
break;
|
|
|
|
case BPF_DW:
|
|
|
|
/* mov qword ptr [rax + r12 + 1], rax */
|
|
|
|
EMIT2(add_3mod(0x48, dst_reg, src_reg, index_reg), 0x89);
|
|
|
|
break;
|
|
|
|
}
|
|
|
|
emit_insn_suffix_SIB(&prog, dst_reg, src_reg, index_reg, off);
|
|
|
|
*pprog = prog;
|
|
|
|
}
|
|
|
|
|
|
|
|
static void emit_stx_r12(u8 **pprog, u32 size, u32 dst_reg, u32 src_reg, int off)
|
|
|
|
{
|
|
|
|
emit_stx_index(pprog, size, dst_reg, src_reg, X86_REG_R12, off);
|
|
|
|
}
|
|
|
|
|
|
|
|
/* ST: *(u8*)(dst_reg + index_reg + off) = imm32 */
|
|
|
|
static void emit_st_index(u8 **pprog, u32 size, u32 dst_reg, u32 index_reg, int off, int imm)
|
|
|
|
{
|
|
|
|
u8 *prog = *pprog;
|
|
|
|
|
|
|
|
switch (size) {
|
|
|
|
case BPF_B:
|
|
|
|
/* mov byte ptr [rax + r12 + off], imm8 */
|
|
|
|
EMIT2(add_3mod(0x40, dst_reg, 0, index_reg), 0xC6);
|
|
|
|
break;
|
|
|
|
case BPF_H:
|
|
|
|
/* mov word ptr [rax + r12 + off], imm16 */
|
|
|
|
EMIT3(0x66, add_3mod(0x40, dst_reg, 0, index_reg), 0xC7);
|
|
|
|
break;
|
|
|
|
case BPF_W:
|
|
|
|
/* mov dword ptr [rax + r12 + 1], imm32 */
|
|
|
|
EMIT2(add_3mod(0x40, dst_reg, 0, index_reg), 0xC7);
|
|
|
|
break;
|
|
|
|
case BPF_DW:
|
|
|
|
/* mov qword ptr [rax + r12 + 1], imm32 */
|
|
|
|
EMIT2(add_3mod(0x48, dst_reg, 0, index_reg), 0xC7);
|
|
|
|
break;
|
|
|
|
}
|
|
|
|
emit_insn_suffix_SIB(&prog, dst_reg, 0, index_reg, off);
|
|
|
|
EMIT(imm, bpf_size_to_x86_bytes(size));
|
|
|
|
*pprog = prog;
|
|
|
|
}
|
|
|
|
|
|
|
|
static void emit_st_r12(u8 **pprog, u32 size, u32 dst_reg, int off, int imm)
|
|
|
|
{
|
|
|
|
emit_st_index(pprog, size, dst_reg, X86_REG_R12, off, imm);
|
|
|
|
}
|
|
|
|
|
2021-01-14 18:17:44 +00:00
|
|
|
static int emit_atomic(u8 **pprog, u8 atomic_op,
|
|
|
|
u32 dst_reg, u32 src_reg, s16 off, u8 bpf_size)
|
|
|
|
{
|
|
|
|
u8 *prog = *pprog;
|
|
|
|
|
|
|
|
EMIT1(0xF0); /* lock prefix */
|
|
|
|
|
|
|
|
maybe_emit_mod(&prog, dst_reg, src_reg, bpf_size == BPF_DW);
|
|
|
|
|
|
|
|
/* emit opcode */
|
|
|
|
switch (atomic_op) {
|
|
|
|
case BPF_ADD:
|
2021-01-14 18:17:49 +00:00
|
|
|
case BPF_AND:
|
|
|
|
case BPF_OR:
|
|
|
|
case BPF_XOR:
|
2021-01-14 18:17:44 +00:00
|
|
|
/* lock *(u32/u64*)(dst_reg + off) <op>= src_reg */
|
|
|
|
EMIT1(simple_alu_opcodes[atomic_op]);
|
|
|
|
break;
|
2021-01-14 18:17:46 +00:00
|
|
|
case BPF_ADD | BPF_FETCH:
|
|
|
|
/* src_reg = atomic_fetch_add(dst_reg + off, src_reg); */
|
|
|
|
EMIT2(0x0F, 0xC1);
|
|
|
|
break;
|
2021-01-14 18:17:47 +00:00
|
|
|
case BPF_XCHG:
|
|
|
|
/* src_reg = atomic_xchg(dst_reg + off, src_reg); */
|
|
|
|
EMIT1(0x87);
|
|
|
|
break;
|
|
|
|
case BPF_CMPXCHG:
|
|
|
|
/* r0 = atomic_cmpxchg(dst_reg + off, r0, src_reg); */
|
|
|
|
EMIT2(0x0F, 0xB1);
|
|
|
|
break;
|
2021-01-14 18:17:44 +00:00
|
|
|
default:
|
|
|
|
pr_err("bpf_jit: unknown atomic opcode %02x\n", atomic_op);
|
|
|
|
return -EFAULT;
|
|
|
|
}
|
|
|
|
|
|
|
|
emit_insn_suffix(&prog, dst_reg, src_reg, off);
|
|
|
|
|
|
|
|
*pprog = prog;
|
|
|
|
return 0;
|
|
|
|
}
|
|
|
|
|
2024-03-08 01:08:01 +00:00
|
|
|
#define DONT_CLEAR 1
|
|
|
|
|
2021-09-08 13:29:18 +00:00
|
|
|
bool ex_handler_bpf(const struct exception_table_entry *x, struct pt_regs *regs)
|
2019-10-16 03:25:03 +00:00
|
|
|
{
|
|
|
|
u32 reg = x->fixup >> 8;
|
|
|
|
|
|
|
|
/* jump over faulting load and clear dest register */
|
2024-03-08 01:08:01 +00:00
|
|
|
if (reg != DONT_CLEAR)
|
|
|
|
*(unsigned long *)((void *)regs + reg) = 0;
|
2019-10-16 03:25:03 +00:00
|
|
|
regs->ip += x->fixup & 0xff;
|
|
|
|
return true;
|
|
|
|
}
|
|
|
|
|
bpf, x64: rework pro/epilogue and tailcall handling in JIT
This commit serves two things:
1) it optimizes BPF prologue/epilogue generation
2) it makes possible to have tailcalls within BPF subprogram
Both points are related to each other since without 1), 2) could not be
achieved.
In [1], Alexei says:
"The prologue will look like:
nop5
xor eax,eax // two new bytes if bpf_tail_call() is used in this
// function
push rbp
mov rbp, rsp
sub rsp, rounded_stack_depth
push rax // zero init tail_call counter
variable number of push rbx,r13,r14,r15
Then bpf_tail_call will pop variable number rbx,..
and final 'pop rax'
Then 'add rsp, size_of_current_stack_frame'
jmp to next function and skip over 'nop5; xor eax,eax; push rpb; mov
rbp, rsp'
This way new function will set its own stack size and will init tail
call
counter with whatever value the parent had.
If next function doesn't use bpf_tail_call it won't have 'xor eax,eax'.
Instead it would need to have 'nop2' in there."
Implement that suggestion.
Since the layout of stack is changed, tail call counter handling can not
rely anymore on popping it to rbx just like it have been handled for
constant prologue case and later overwrite of rbx with actual value of
rbx pushed to stack. Therefore, let's use one of the register (%rcx) that
is considered to be volatile/caller-saved and pop the value of tail call
counter in there in the epilogue.
Drop the BUILD_BUG_ON in emit_prologue and in
emit_bpf_tail_call_indirect where instruction layout is not constant
anymore.
Introduce new poke target, 'tailcall_bypass' to poke descriptor that is
dedicated for skipping the register pops and stack unwind that are
generated right before the actual jump to target program.
For case when the target program is not present, BPF program will skip
the pop instructions and nop5 dedicated for jmpq $target. An example of
such state when only R6 of callee saved registers is used by program:
ffffffffc0513aa1: e9 0e 00 00 00 jmpq 0xffffffffc0513ab4
ffffffffc0513aa6: 5b pop %rbx
ffffffffc0513aa7: 58 pop %rax
ffffffffc0513aa8: 48 81 c4 00 00 00 00 add $0x0,%rsp
ffffffffc0513aaf: 0f 1f 44 00 00 nopl 0x0(%rax,%rax,1)
ffffffffc0513ab4: 48 89 df mov %rbx,%rdi
When target program is inserted, the jump that was there to skip
pops/nop5 will become the nop5, so CPU will go over pops and do the
actual tailcall.
One might ask why there simply can not be pushes after the nop5?
In the following example snippet:
ffffffffc037030c: 48 89 fb mov %rdi,%rbx
(...)
ffffffffc0370332: 5b pop %rbx
ffffffffc0370333: 58 pop %rax
ffffffffc0370334: 48 81 c4 00 00 00 00 add $0x0,%rsp
ffffffffc037033b: 0f 1f 44 00 00 nopl 0x0(%rax,%rax,1)
ffffffffc0370340: 48 81 ec 00 00 00 00 sub $0x0,%rsp
ffffffffc0370347: 50 push %rax
ffffffffc0370348: 53 push %rbx
ffffffffc0370349: 48 89 df mov %rbx,%rdi
ffffffffc037034c: e8 f7 21 00 00 callq 0xffffffffc0372548
There is the bpf2bpf call (at ffffffffc037034c) right after the tailcall
and jump target is not present. ctx is in %rbx register and BPF
subprogram that we will call into on ffffffffc037034c is relying on it,
e.g. it will pick ctx from there. Such code layout is therefore broken
as we would overwrite the content of %rbx with the value that was pushed
on the prologue. That is the reason for the 'bypass' approach.
Special care needs to be taken during the install/update/remove of
tailcall target. In case when target program is not present, the CPU
must not execute the pop instructions that precede the tailcall.
To address that, the following states can be defined:
A nop, unwind, nop
B nop, unwind, tail
C skip, unwind, nop
D skip, unwind, tail
A is forbidden (lead to incorrectness). The state transitions between
tailcall install/update/remove will work as follows:
First install tail call f: C->D->B(f)
* poke the tailcall, after that get rid of the skip
Update tail call f to f': B(f)->B(f')
* poke the tailcall (poke->tailcall_target) and do NOT touch the
poke->tailcall_bypass
Remove tail call: B(f')->C(f')
* poke->tailcall_bypass is poked back to jump, then we wait the RCU
grace period so that other programs will finish its execution and
after that we are safe to remove the poke->tailcall_target
Install new tail call (f''): C(f')->D(f'')->B(f'').
* same as first step
This way CPU can never be exposed to "unwind, tail" state.
Last but not least, when tailcalls get mixed with bpf2bpf calls, it
would be possible to encounter the endless loop due to clearing the
tailcall counter if for example we would use the tailcall3-like from BPF
selftests program that would be subprogram-based, meaning the tailcall
would be present within the BPF subprogram.
This test, broken down to particular steps, would do:
entry -> set tailcall counter to 0, bump it by 1, tailcall to func0
func0 -> call subprog_tail
(we are NOT skipping the first 11 bytes of prologue and this subprogram
has a tailcall, therefore we clear the counter...)
subprog -> do the same thing as entry
and then loop forever.
To address this, the idea is to go through the call chain of bpf2bpf progs
and look for a tailcall presence throughout whole chain. If we saw a single
tail call then each node in this call chain needs to be marked as a subprog
that can reach the tailcall. We would later feed the JIT with this info
and:
- set eax to 0 only when tailcall is reachable and this is the entry prog
- if tailcall is reachable but there's no tailcall in insns of currently
JITed prog then push rax anyway, so that it will be possible to
propagate further down the call chain
- finally if tailcall is reachable, then we need to precede the 'call'
insn with mov rax, [rbp - (stack_depth + 8)]
Tail call related cases from test_verifier kselftest are also working
fine. Sample BPF programs that utilize tail calls (sockex3, tracex5)
work properly as well.
[1]: https://lore.kernel.org/bpf/20200517043227.2gpq22ifoq37ogst@ast-mbp.dhcp.thefacebook.com/
Suggested-by: Alexei Starovoitov <ast@kernel.org>
Signed-off-by: Maciej Fijalkowski <maciej.fijalkowski@intel.com>
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
2020-09-16 21:10:08 +00:00
|
|
|
static void detect_reg_usage(struct bpf_insn *insn, int insn_cnt,
|
|
|
|
bool *regs_used, bool *tail_call_seen)
|
|
|
|
{
|
|
|
|
int i;
|
|
|
|
|
|
|
|
for (i = 1; i <= insn_cnt; i++, insn++) {
|
|
|
|
if (insn->code == (BPF_JMP | BPF_TAIL_CALL))
|
|
|
|
*tail_call_seen = true;
|
|
|
|
if (insn->dst_reg == BPF_REG_6 || insn->src_reg == BPF_REG_6)
|
|
|
|
regs_used[0] = true;
|
|
|
|
if (insn->dst_reg == BPF_REG_7 || insn->src_reg == BPF_REG_7)
|
|
|
|
regs_used[1] = true;
|
|
|
|
if (insn->dst_reg == BPF_REG_8 || insn->src_reg == BPF_REG_8)
|
|
|
|
regs_used[2] = true;
|
|
|
|
if (insn->dst_reg == BPF_REG_9 || insn->src_reg == BPF_REG_9)
|
|
|
|
regs_used[3] = true;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
bpf,x64: use shrx/sarx/shlx when available
BMI2 provides 3 shift instructions (shrx, sarx and shlx) that use VEX
encoding but target general purpose registers [1]. They allow the shift
count in any general purpose register and have the same performance as
non BMI2 shift instructions [2].
Instead of shr/sar/shl that implicitly use %cl (lowest 8 bit of %rcx),
emit their more flexible alternatives provided in BMI2 when advantageous;
keep using the non BMI2 instructions when shift count is already in
BPF_REG_4/%rcx as non BMI2 instructions are shorter.
To summarize, when BMI2 is available:
-------------------------------------------------
| arbitrary dst
=================================================
src == ecx | shl dst, cl
-------------------------------------------------
src != ecx | shlx dst, dst, src
-------------------------------------------------
And no additional register shuffling is needed.
A concrete example between non BMI2 and BMI2 codegen. To shift %rsi by
%rdi:
Without BMI2:
ef3: push %rcx
51
ef4: mov %rdi,%rcx
48 89 f9
ef7: shl %cl,%rsi
48 d3 e6
efa: pop %rcx
59
With BMI2:
f0b: shlx %rdi,%rsi,%rsi
c4 e2 c1 f7 f6
[1] https://en.wikipedia.org/wiki/X86_Bit_manipulation_instruction_set
[2] https://www.agner.org/optimize/instruction_tables.pdf
Signed-off-by: Jie Meng <jmeng@fb.com>
Link: https://lore.kernel.org/r/20221007202348.1118830-3-jmeng@fb.com
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
2022-10-07 20:23:48 +00:00
|
|
|
/* emit the 3-byte VEX prefix
|
|
|
|
*
|
|
|
|
* r: same as rex.r, extra bit for ModRM reg field
|
|
|
|
* x: same as rex.x, extra bit for SIB index field
|
|
|
|
* b: same as rex.b, extra bit for ModRM r/m, or SIB base
|
|
|
|
* m: opcode map select, encoding escape bytes e.g. 0x0f38
|
|
|
|
* w: same as rex.w (32 bit or 64 bit) or opcode specific
|
|
|
|
* src_reg2: additional source reg (encoded as BPF reg)
|
|
|
|
* l: vector length (128 bit or 256 bit) or reserved
|
|
|
|
* pp: opcode prefix (none, 0x66, 0xf2 or 0xf3)
|
|
|
|
*/
|
|
|
|
static void emit_3vex(u8 **pprog, bool r, bool x, bool b, u8 m,
|
|
|
|
bool w, u8 src_reg2, bool l, u8 pp)
|
|
|
|
{
|
|
|
|
u8 *prog = *pprog;
|
|
|
|
const u8 b0 = 0xc4; /* first byte of 3-byte VEX prefix */
|
|
|
|
u8 b1, b2;
|
|
|
|
u8 vvvv = reg2hex[src_reg2];
|
|
|
|
|
|
|
|
/* reg2hex gives only the lower 3 bit of vvvv */
|
|
|
|
if (is_ereg(src_reg2))
|
|
|
|
vvvv |= 1 << 3;
|
|
|
|
|
|
|
|
/*
|
|
|
|
* 2nd byte of 3-byte VEX prefix
|
|
|
|
* ~ means bit inverted encoding
|
|
|
|
*
|
|
|
|
* 7 0
|
|
|
|
* +---+---+---+---+---+---+---+---+
|
|
|
|
* |~R |~X |~B | m |
|
|
|
|
* +---+---+---+---+---+---+---+---+
|
|
|
|
*/
|
|
|
|
b1 = (!r << 7) | (!x << 6) | (!b << 5) | (m & 0x1f);
|
|
|
|
/*
|
|
|
|
* 3rd byte of 3-byte VEX prefix
|
|
|
|
*
|
|
|
|
* 7 0
|
|
|
|
* +---+---+---+---+---+---+---+---+
|
|
|
|
* | W | ~vvvv | L | pp |
|
|
|
|
* +---+---+---+---+---+---+---+---+
|
|
|
|
*/
|
|
|
|
b2 = (w << 7) | ((~vvvv & 0xf) << 3) | (l << 2) | (pp & 3);
|
|
|
|
|
|
|
|
EMIT3(b0, b1, b2);
|
|
|
|
*pprog = prog;
|
|
|
|
}
|
|
|
|
|
|
|
|
/* emit BMI2 shift instruction */
|
|
|
|
static void emit_shiftx(u8 **pprog, u32 dst_reg, u8 src_reg, bool is64, u8 op)
|
|
|
|
{
|
|
|
|
u8 *prog = *pprog;
|
|
|
|
bool r = is_ereg(dst_reg);
|
|
|
|
u8 m = 2; /* escape code 0f38 */
|
|
|
|
|
|
|
|
emit_3vex(&prog, r, false, r, m, is64, src_reg, false, op);
|
|
|
|
EMIT2(0xf7, add_2reg(0xC0, dst_reg, dst_reg));
|
|
|
|
*pprog = prog;
|
|
|
|
}
|
|
|
|
|
bpf,x64: Pad NOPs to make images converge more easily
The x64 bpf jit expects bpf images converge within the given passes, but
it could fail to do so with some corner cases. For example:
l0: ja 40
l1: ja 40
[... repeated ja 40 ]
l39: ja 40
l40: ret #0
This bpf program contains 40 "ja 40" instructions which are effectively
NOPs and designed to be replaced with valid code dynamically. Ideally,
bpf jit should optimize those "ja 40" instructions out when translating
the bpf instructions into x64 machine code. However, do_jit() can only
remove one "ja 40" for offset==0 on each pass, so it requires at least
40 runs to eliminate those JMPs and exceeds the current limit of
passes(20). In the end, the program got rejected when BPF_JIT_ALWAYS_ON
is set even though it's legit as a classic socket filter.
To make bpf images more likely converge within 20 passes, this commit
pads some instructions with NOPs in the last 5 passes:
1. conditional jumps
A possible size variance comes from the adoption of imm8 JMP. If the
offset is imm8, we calculate the size difference of this BPF instruction
between the previous and the current pass and fill the gap with NOPs.
To avoid the recalculation of jump offset, those NOPs are inserted before
the JMP code, so we have to subtract the 2 bytes of imm8 JMP when
calculating the NOP number.
2. BPF_JA
There are two conditions for BPF_JA.
a.) nop jumps
If this instruction is not optimized out in the previous pass,
instead of removing it, we insert the equivalent size of NOPs.
b.) label jumps
Similar to condition jumps, we prepend NOPs right before the JMP
code.
To make the code concise, emit_nops() is modified to use the signed len and
return the number of inserted NOPs.
For bpf-to-bpf, we always enable padding for the extra pass since there
is only one extra run and the jump padding doesn't affected the images
that converge without padding.
After applying this patch, the corner case was loaded with the following
jit code:
flen=45 proglen=77 pass=17 image=ffffffffc03367d4 from=jump pid=10097
JIT code: 00000000: 0f 1f 44 00 00 55 48 89 e5 53 41 55 31 c0 45 31
JIT code: 00000010: ed 48 89 fb eb 30 eb 2e eb 2c eb 2a eb 28 eb 26
JIT code: 00000020: eb 24 eb 22 eb 20 eb 1e eb 1c eb 1a eb 18 eb 16
JIT code: 00000030: eb 14 eb 12 eb 10 eb 0e eb 0c eb 0a eb 08 eb 06
JIT code: 00000040: eb 04 eb 02 66 90 31 c0 41 5d 5b c9 c3
0: 0f 1f 44 00 00 nop DWORD PTR [rax+rax*1+0x0]
5: 55 push rbp
6: 48 89 e5 mov rbp,rsp
9: 53 push rbx
a: 41 55 push r13
c: 31 c0 xor eax,eax
e: 45 31 ed xor r13d,r13d
11: 48 89 fb mov rbx,rdi
14: eb 30 jmp 0x46
16: eb 2e jmp 0x46
...
3e: eb 06 jmp 0x46
40: eb 04 jmp 0x46
42: eb 02 jmp 0x46
44: 66 90 xchg ax,ax
46: 31 c0 xor eax,eax
48: 41 5d pop r13
4a: 5b pop rbx
4b: c9 leave
4c: c3 ret
At the 16th pass, 15 jumps were already optimized out, and one jump was
replaced with NOPs at 44 and the image converged at the 17th pass.
v4:
- Add the detailed comments about the possible padding bytes
v3:
- Copy the instructions of prologue separately or the size calculation
of the first BPF instruction would include the prologue.
- Replace WARN_ONCE() with pr_err() and EFAULT
- Use MAX_PASSES in the for loop condition check
- Remove the "padded" flag from x64_jit_data. For the extra pass of
subprogs, padding is always enabled since it won't hurt the images
that converge without padding.
v2:
- Simplify the sample code in the description and provide the jit code
- Check the expected padding bytes with WARN_ONCE
- Move the 'padded' flag to 'struct x64_jit_data'
Signed-off-by: Gary Lin <glin@suse.com>
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Link: https://lore.kernel.org/bpf/20210119102501.511-2-glin@suse.com
2021-01-19 10:24:59 +00:00
|
|
|
#define INSN_SZ_DIFF (((addrs[i] - addrs[i - 1]) - (prog - temp)))
|
|
|
|
|
2023-09-12 15:04:41 +00:00
|
|
|
/* mov rax, qword ptr [rbp - rounded_stack_depth - 8] */
|
|
|
|
#define RESTORE_TAIL_CALL_CNT(stack) \
|
|
|
|
EMIT3_off32(0x48, 0x8B, 0x85, -round_up(stack, 8) - 8)
|
|
|
|
|
2022-02-04 18:57:42 +00:00
|
|
|
static int do_jit(struct bpf_prog *bpf_prog, int *addrs, u8 *image, u8 *rw_image,
|
bpf,x64: Pad NOPs to make images converge more easily
The x64 bpf jit expects bpf images converge within the given passes, but
it could fail to do so with some corner cases. For example:
l0: ja 40
l1: ja 40
[... repeated ja 40 ]
l39: ja 40
l40: ret #0
This bpf program contains 40 "ja 40" instructions which are effectively
NOPs and designed to be replaced with valid code dynamically. Ideally,
bpf jit should optimize those "ja 40" instructions out when translating
the bpf instructions into x64 machine code. However, do_jit() can only
remove one "ja 40" for offset==0 on each pass, so it requires at least
40 runs to eliminate those JMPs and exceeds the current limit of
passes(20). In the end, the program got rejected when BPF_JIT_ALWAYS_ON
is set even though it's legit as a classic socket filter.
To make bpf images more likely converge within 20 passes, this commit
pads some instructions with NOPs in the last 5 passes:
1. conditional jumps
A possible size variance comes from the adoption of imm8 JMP. If the
offset is imm8, we calculate the size difference of this BPF instruction
between the previous and the current pass and fill the gap with NOPs.
To avoid the recalculation of jump offset, those NOPs are inserted before
the JMP code, so we have to subtract the 2 bytes of imm8 JMP when
calculating the NOP number.
2. BPF_JA
There are two conditions for BPF_JA.
a.) nop jumps
If this instruction is not optimized out in the previous pass,
instead of removing it, we insert the equivalent size of NOPs.
b.) label jumps
Similar to condition jumps, we prepend NOPs right before the JMP
code.
To make the code concise, emit_nops() is modified to use the signed len and
return the number of inserted NOPs.
For bpf-to-bpf, we always enable padding for the extra pass since there
is only one extra run and the jump padding doesn't affected the images
that converge without padding.
After applying this patch, the corner case was loaded with the following
jit code:
flen=45 proglen=77 pass=17 image=ffffffffc03367d4 from=jump pid=10097
JIT code: 00000000: 0f 1f 44 00 00 55 48 89 e5 53 41 55 31 c0 45 31
JIT code: 00000010: ed 48 89 fb eb 30 eb 2e eb 2c eb 2a eb 28 eb 26
JIT code: 00000020: eb 24 eb 22 eb 20 eb 1e eb 1c eb 1a eb 18 eb 16
JIT code: 00000030: eb 14 eb 12 eb 10 eb 0e eb 0c eb 0a eb 08 eb 06
JIT code: 00000040: eb 04 eb 02 66 90 31 c0 41 5d 5b c9 c3
0: 0f 1f 44 00 00 nop DWORD PTR [rax+rax*1+0x0]
5: 55 push rbp
6: 48 89 e5 mov rbp,rsp
9: 53 push rbx
a: 41 55 push r13
c: 31 c0 xor eax,eax
e: 45 31 ed xor r13d,r13d
11: 48 89 fb mov rbx,rdi
14: eb 30 jmp 0x46
16: eb 2e jmp 0x46
...
3e: eb 06 jmp 0x46
40: eb 04 jmp 0x46
42: eb 02 jmp 0x46
44: 66 90 xchg ax,ax
46: 31 c0 xor eax,eax
48: 41 5d pop r13
4a: 5b pop rbx
4b: c9 leave
4c: c3 ret
At the 16th pass, 15 jumps were already optimized out, and one jump was
replaced with NOPs at 44 and the image converged at the 17th pass.
v4:
- Add the detailed comments about the possible padding bytes
v3:
- Copy the instructions of prologue separately or the size calculation
of the first BPF instruction would include the prologue.
- Replace WARN_ONCE() with pr_err() and EFAULT
- Use MAX_PASSES in the for loop condition check
- Remove the "padded" flag from x64_jit_data. For the extra pass of
subprogs, padding is always enabled since it won't hurt the images
that converge without padding.
v2:
- Simplify the sample code in the description and provide the jit code
- Check the expected padding bytes with WARN_ONCE
- Move the 'padded' flag to 'struct x64_jit_data'
Signed-off-by: Gary Lin <glin@suse.com>
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Link: https://lore.kernel.org/bpf/20210119102501.511-2-glin@suse.com
2021-01-19 10:24:59 +00:00
|
|
|
int oldproglen, struct jit_context *ctx, bool jmp_padding)
|
2015-05-19 23:59:04 +00:00
|
|
|
{
|
bpf, x64: rework pro/epilogue and tailcall handling in JIT
This commit serves two things:
1) it optimizes BPF prologue/epilogue generation
2) it makes possible to have tailcalls within BPF subprogram
Both points are related to each other since without 1), 2) could not be
achieved.
In [1], Alexei says:
"The prologue will look like:
nop5
xor eax,eax // two new bytes if bpf_tail_call() is used in this
// function
push rbp
mov rbp, rsp
sub rsp, rounded_stack_depth
push rax // zero init tail_call counter
variable number of push rbx,r13,r14,r15
Then bpf_tail_call will pop variable number rbx,..
and final 'pop rax'
Then 'add rsp, size_of_current_stack_frame'
jmp to next function and skip over 'nop5; xor eax,eax; push rpb; mov
rbp, rsp'
This way new function will set its own stack size and will init tail
call
counter with whatever value the parent had.
If next function doesn't use bpf_tail_call it won't have 'xor eax,eax'.
Instead it would need to have 'nop2' in there."
Implement that suggestion.
Since the layout of stack is changed, tail call counter handling can not
rely anymore on popping it to rbx just like it have been handled for
constant prologue case and later overwrite of rbx with actual value of
rbx pushed to stack. Therefore, let's use one of the register (%rcx) that
is considered to be volatile/caller-saved and pop the value of tail call
counter in there in the epilogue.
Drop the BUILD_BUG_ON in emit_prologue and in
emit_bpf_tail_call_indirect where instruction layout is not constant
anymore.
Introduce new poke target, 'tailcall_bypass' to poke descriptor that is
dedicated for skipping the register pops and stack unwind that are
generated right before the actual jump to target program.
For case when the target program is not present, BPF program will skip
the pop instructions and nop5 dedicated for jmpq $target. An example of
such state when only R6 of callee saved registers is used by program:
ffffffffc0513aa1: e9 0e 00 00 00 jmpq 0xffffffffc0513ab4
ffffffffc0513aa6: 5b pop %rbx
ffffffffc0513aa7: 58 pop %rax
ffffffffc0513aa8: 48 81 c4 00 00 00 00 add $0x0,%rsp
ffffffffc0513aaf: 0f 1f 44 00 00 nopl 0x0(%rax,%rax,1)
ffffffffc0513ab4: 48 89 df mov %rbx,%rdi
When target program is inserted, the jump that was there to skip
pops/nop5 will become the nop5, so CPU will go over pops and do the
actual tailcall.
One might ask why there simply can not be pushes after the nop5?
In the following example snippet:
ffffffffc037030c: 48 89 fb mov %rdi,%rbx
(...)
ffffffffc0370332: 5b pop %rbx
ffffffffc0370333: 58 pop %rax
ffffffffc0370334: 48 81 c4 00 00 00 00 add $0x0,%rsp
ffffffffc037033b: 0f 1f 44 00 00 nopl 0x0(%rax,%rax,1)
ffffffffc0370340: 48 81 ec 00 00 00 00 sub $0x0,%rsp
ffffffffc0370347: 50 push %rax
ffffffffc0370348: 53 push %rbx
ffffffffc0370349: 48 89 df mov %rbx,%rdi
ffffffffc037034c: e8 f7 21 00 00 callq 0xffffffffc0372548
There is the bpf2bpf call (at ffffffffc037034c) right after the tailcall
and jump target is not present. ctx is in %rbx register and BPF
subprogram that we will call into on ffffffffc037034c is relying on it,
e.g. it will pick ctx from there. Such code layout is therefore broken
as we would overwrite the content of %rbx with the value that was pushed
on the prologue. That is the reason for the 'bypass' approach.
Special care needs to be taken during the install/update/remove of
tailcall target. In case when target program is not present, the CPU
must not execute the pop instructions that precede the tailcall.
To address that, the following states can be defined:
A nop, unwind, nop
B nop, unwind, tail
C skip, unwind, nop
D skip, unwind, tail
A is forbidden (lead to incorrectness). The state transitions between
tailcall install/update/remove will work as follows:
First install tail call f: C->D->B(f)
* poke the tailcall, after that get rid of the skip
Update tail call f to f': B(f)->B(f')
* poke the tailcall (poke->tailcall_target) and do NOT touch the
poke->tailcall_bypass
Remove tail call: B(f')->C(f')
* poke->tailcall_bypass is poked back to jump, then we wait the RCU
grace period so that other programs will finish its execution and
after that we are safe to remove the poke->tailcall_target
Install new tail call (f''): C(f')->D(f'')->B(f'').
* same as first step
This way CPU can never be exposed to "unwind, tail" state.
Last but not least, when tailcalls get mixed with bpf2bpf calls, it
would be possible to encounter the endless loop due to clearing the
tailcall counter if for example we would use the tailcall3-like from BPF
selftests program that would be subprogram-based, meaning the tailcall
would be present within the BPF subprogram.
This test, broken down to particular steps, would do:
entry -> set tailcall counter to 0, bump it by 1, tailcall to func0
func0 -> call subprog_tail
(we are NOT skipping the first 11 bytes of prologue and this subprogram
has a tailcall, therefore we clear the counter...)
subprog -> do the same thing as entry
and then loop forever.
To address this, the idea is to go through the call chain of bpf2bpf progs
and look for a tailcall presence throughout whole chain. If we saw a single
tail call then each node in this call chain needs to be marked as a subprog
that can reach the tailcall. We would later feed the JIT with this info
and:
- set eax to 0 only when tailcall is reachable and this is the entry prog
- if tailcall is reachable but there's no tailcall in insns of currently
JITed prog then push rax anyway, so that it will be possible to
propagate further down the call chain
- finally if tailcall is reachable, then we need to precede the 'call'
insn with mov rax, [rbp - (stack_depth + 8)]
Tail call related cases from test_verifier kselftest are also working
fine. Sample BPF programs that utilize tail calls (sockex3, tracex5)
work properly as well.
[1]: https://lore.kernel.org/bpf/20200517043227.2gpq22ifoq37ogst@ast-mbp.dhcp.thefacebook.com/
Suggested-by: Alexei Starovoitov <ast@kernel.org>
Signed-off-by: Maciej Fijalkowski <maciej.fijalkowski@intel.com>
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
2020-09-16 21:10:08 +00:00
|
|
|
bool tail_call_reachable = bpf_prog->aux->tail_call_reachable;
|
2015-05-19 23:59:04 +00:00
|
|
|
struct bpf_insn *insn = bpf_prog->insnsi;
|
bpf, x64: rework pro/epilogue and tailcall handling in JIT
This commit serves two things:
1) it optimizes BPF prologue/epilogue generation
2) it makes possible to have tailcalls within BPF subprogram
Both points are related to each other since without 1), 2) could not be
achieved.
In [1], Alexei says:
"The prologue will look like:
nop5
xor eax,eax // two new bytes if bpf_tail_call() is used in this
// function
push rbp
mov rbp, rsp
sub rsp, rounded_stack_depth
push rax // zero init tail_call counter
variable number of push rbx,r13,r14,r15
Then bpf_tail_call will pop variable number rbx,..
and final 'pop rax'
Then 'add rsp, size_of_current_stack_frame'
jmp to next function and skip over 'nop5; xor eax,eax; push rpb; mov
rbp, rsp'
This way new function will set its own stack size and will init tail
call
counter with whatever value the parent had.
If next function doesn't use bpf_tail_call it won't have 'xor eax,eax'.
Instead it would need to have 'nop2' in there."
Implement that suggestion.
Since the layout of stack is changed, tail call counter handling can not
rely anymore on popping it to rbx just like it have been handled for
constant prologue case and later overwrite of rbx with actual value of
rbx pushed to stack. Therefore, let's use one of the register (%rcx) that
is considered to be volatile/caller-saved and pop the value of tail call
counter in there in the epilogue.
Drop the BUILD_BUG_ON in emit_prologue and in
emit_bpf_tail_call_indirect where instruction layout is not constant
anymore.
Introduce new poke target, 'tailcall_bypass' to poke descriptor that is
dedicated for skipping the register pops and stack unwind that are
generated right before the actual jump to target program.
For case when the target program is not present, BPF program will skip
the pop instructions and nop5 dedicated for jmpq $target. An example of
such state when only R6 of callee saved registers is used by program:
ffffffffc0513aa1: e9 0e 00 00 00 jmpq 0xffffffffc0513ab4
ffffffffc0513aa6: 5b pop %rbx
ffffffffc0513aa7: 58 pop %rax
ffffffffc0513aa8: 48 81 c4 00 00 00 00 add $0x0,%rsp
ffffffffc0513aaf: 0f 1f 44 00 00 nopl 0x0(%rax,%rax,1)
ffffffffc0513ab4: 48 89 df mov %rbx,%rdi
When target program is inserted, the jump that was there to skip
pops/nop5 will become the nop5, so CPU will go over pops and do the
actual tailcall.
One might ask why there simply can not be pushes after the nop5?
In the following example snippet:
ffffffffc037030c: 48 89 fb mov %rdi,%rbx
(...)
ffffffffc0370332: 5b pop %rbx
ffffffffc0370333: 58 pop %rax
ffffffffc0370334: 48 81 c4 00 00 00 00 add $0x0,%rsp
ffffffffc037033b: 0f 1f 44 00 00 nopl 0x0(%rax,%rax,1)
ffffffffc0370340: 48 81 ec 00 00 00 00 sub $0x0,%rsp
ffffffffc0370347: 50 push %rax
ffffffffc0370348: 53 push %rbx
ffffffffc0370349: 48 89 df mov %rbx,%rdi
ffffffffc037034c: e8 f7 21 00 00 callq 0xffffffffc0372548
There is the bpf2bpf call (at ffffffffc037034c) right after the tailcall
and jump target is not present. ctx is in %rbx register and BPF
subprogram that we will call into on ffffffffc037034c is relying on it,
e.g. it will pick ctx from there. Such code layout is therefore broken
as we would overwrite the content of %rbx with the value that was pushed
on the prologue. That is the reason for the 'bypass' approach.
Special care needs to be taken during the install/update/remove of
tailcall target. In case when target program is not present, the CPU
must not execute the pop instructions that precede the tailcall.
To address that, the following states can be defined:
A nop, unwind, nop
B nop, unwind, tail
C skip, unwind, nop
D skip, unwind, tail
A is forbidden (lead to incorrectness). The state transitions between
tailcall install/update/remove will work as follows:
First install tail call f: C->D->B(f)
* poke the tailcall, after that get rid of the skip
Update tail call f to f': B(f)->B(f')
* poke the tailcall (poke->tailcall_target) and do NOT touch the
poke->tailcall_bypass
Remove tail call: B(f')->C(f')
* poke->tailcall_bypass is poked back to jump, then we wait the RCU
grace period so that other programs will finish its execution and
after that we are safe to remove the poke->tailcall_target
Install new tail call (f''): C(f')->D(f'')->B(f'').
* same as first step
This way CPU can never be exposed to "unwind, tail" state.
Last but not least, when tailcalls get mixed with bpf2bpf calls, it
would be possible to encounter the endless loop due to clearing the
tailcall counter if for example we would use the tailcall3-like from BPF
selftests program that would be subprogram-based, meaning the tailcall
would be present within the BPF subprogram.
This test, broken down to particular steps, would do:
entry -> set tailcall counter to 0, bump it by 1, tailcall to func0
func0 -> call subprog_tail
(we are NOT skipping the first 11 bytes of prologue and this subprogram
has a tailcall, therefore we clear the counter...)
subprog -> do the same thing as entry
and then loop forever.
To address this, the idea is to go through the call chain of bpf2bpf progs
and look for a tailcall presence throughout whole chain. If we saw a single
tail call then each node in this call chain needs to be marked as a subprog
that can reach the tailcall. We would later feed the JIT with this info
and:
- set eax to 0 only when tailcall is reachable and this is the entry prog
- if tailcall is reachable but there's no tailcall in insns of currently
JITed prog then push rax anyway, so that it will be possible to
propagate further down the call chain
- finally if tailcall is reachable, then we need to precede the 'call'
insn with mov rax, [rbp - (stack_depth + 8)]
Tail call related cases from test_verifier kselftest are also working
fine. Sample BPF programs that utilize tail calls (sockex3, tracex5)
work properly as well.
[1]: https://lore.kernel.org/bpf/20200517043227.2gpq22ifoq37ogst@ast-mbp.dhcp.thefacebook.com/
Suggested-by: Alexei Starovoitov <ast@kernel.org>
Signed-off-by: Maciej Fijalkowski <maciej.fijalkowski@intel.com>
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
2020-09-16 21:10:08 +00:00
|
|
|
bool callee_regs_used[4] = {};
|
2015-05-19 23:59:04 +00:00
|
|
|
int insn_cnt = bpf_prog->len;
|
bpf, x64: rework pro/epilogue and tailcall handling in JIT
This commit serves two things:
1) it optimizes BPF prologue/epilogue generation
2) it makes possible to have tailcalls within BPF subprogram
Both points are related to each other since without 1), 2) could not be
achieved.
In [1], Alexei says:
"The prologue will look like:
nop5
xor eax,eax // two new bytes if bpf_tail_call() is used in this
// function
push rbp
mov rbp, rsp
sub rsp, rounded_stack_depth
push rax // zero init tail_call counter
variable number of push rbx,r13,r14,r15
Then bpf_tail_call will pop variable number rbx,..
and final 'pop rax'
Then 'add rsp, size_of_current_stack_frame'
jmp to next function and skip over 'nop5; xor eax,eax; push rpb; mov
rbp, rsp'
This way new function will set its own stack size and will init tail
call
counter with whatever value the parent had.
If next function doesn't use bpf_tail_call it won't have 'xor eax,eax'.
Instead it would need to have 'nop2' in there."
Implement that suggestion.
Since the layout of stack is changed, tail call counter handling can not
rely anymore on popping it to rbx just like it have been handled for
constant prologue case and later overwrite of rbx with actual value of
rbx pushed to stack. Therefore, let's use one of the register (%rcx) that
is considered to be volatile/caller-saved and pop the value of tail call
counter in there in the epilogue.
Drop the BUILD_BUG_ON in emit_prologue and in
emit_bpf_tail_call_indirect where instruction layout is not constant
anymore.
Introduce new poke target, 'tailcall_bypass' to poke descriptor that is
dedicated for skipping the register pops and stack unwind that are
generated right before the actual jump to target program.
For case when the target program is not present, BPF program will skip
the pop instructions and nop5 dedicated for jmpq $target. An example of
such state when only R6 of callee saved registers is used by program:
ffffffffc0513aa1: e9 0e 00 00 00 jmpq 0xffffffffc0513ab4
ffffffffc0513aa6: 5b pop %rbx
ffffffffc0513aa7: 58 pop %rax
ffffffffc0513aa8: 48 81 c4 00 00 00 00 add $0x0,%rsp
ffffffffc0513aaf: 0f 1f 44 00 00 nopl 0x0(%rax,%rax,1)
ffffffffc0513ab4: 48 89 df mov %rbx,%rdi
When target program is inserted, the jump that was there to skip
pops/nop5 will become the nop5, so CPU will go over pops and do the
actual tailcall.
One might ask why there simply can not be pushes after the nop5?
In the following example snippet:
ffffffffc037030c: 48 89 fb mov %rdi,%rbx
(...)
ffffffffc0370332: 5b pop %rbx
ffffffffc0370333: 58 pop %rax
ffffffffc0370334: 48 81 c4 00 00 00 00 add $0x0,%rsp
ffffffffc037033b: 0f 1f 44 00 00 nopl 0x0(%rax,%rax,1)
ffffffffc0370340: 48 81 ec 00 00 00 00 sub $0x0,%rsp
ffffffffc0370347: 50 push %rax
ffffffffc0370348: 53 push %rbx
ffffffffc0370349: 48 89 df mov %rbx,%rdi
ffffffffc037034c: e8 f7 21 00 00 callq 0xffffffffc0372548
There is the bpf2bpf call (at ffffffffc037034c) right after the tailcall
and jump target is not present. ctx is in %rbx register and BPF
subprogram that we will call into on ffffffffc037034c is relying on it,
e.g. it will pick ctx from there. Such code layout is therefore broken
as we would overwrite the content of %rbx with the value that was pushed
on the prologue. That is the reason for the 'bypass' approach.
Special care needs to be taken during the install/update/remove of
tailcall target. In case when target program is not present, the CPU
must not execute the pop instructions that precede the tailcall.
To address that, the following states can be defined:
A nop, unwind, nop
B nop, unwind, tail
C skip, unwind, nop
D skip, unwind, tail
A is forbidden (lead to incorrectness). The state transitions between
tailcall install/update/remove will work as follows:
First install tail call f: C->D->B(f)
* poke the tailcall, after that get rid of the skip
Update tail call f to f': B(f)->B(f')
* poke the tailcall (poke->tailcall_target) and do NOT touch the
poke->tailcall_bypass
Remove tail call: B(f')->C(f')
* poke->tailcall_bypass is poked back to jump, then we wait the RCU
grace period so that other programs will finish its execution and
after that we are safe to remove the poke->tailcall_target
Install new tail call (f''): C(f')->D(f'')->B(f'').
* same as first step
This way CPU can never be exposed to "unwind, tail" state.
Last but not least, when tailcalls get mixed with bpf2bpf calls, it
would be possible to encounter the endless loop due to clearing the
tailcall counter if for example we would use the tailcall3-like from BPF
selftests program that would be subprogram-based, meaning the tailcall
would be present within the BPF subprogram.
This test, broken down to particular steps, would do:
entry -> set tailcall counter to 0, bump it by 1, tailcall to func0
func0 -> call subprog_tail
(we are NOT skipping the first 11 bytes of prologue and this subprogram
has a tailcall, therefore we clear the counter...)
subprog -> do the same thing as entry
and then loop forever.
To address this, the idea is to go through the call chain of bpf2bpf progs
and look for a tailcall presence throughout whole chain. If we saw a single
tail call then each node in this call chain needs to be marked as a subprog
that can reach the tailcall. We would later feed the JIT with this info
and:
- set eax to 0 only when tailcall is reachable and this is the entry prog
- if tailcall is reachable but there's no tailcall in insns of currently
JITed prog then push rax anyway, so that it will be possible to
propagate further down the call chain
- finally if tailcall is reachable, then we need to precede the 'call'
insn with mov rax, [rbp - (stack_depth + 8)]
Tail call related cases from test_verifier kselftest are also working
fine. Sample BPF programs that utilize tail calls (sockex3, tracex5)
work properly as well.
[1]: https://lore.kernel.org/bpf/20200517043227.2gpq22ifoq37ogst@ast-mbp.dhcp.thefacebook.com/
Suggested-by: Alexei Starovoitov <ast@kernel.org>
Signed-off-by: Maciej Fijalkowski <maciej.fijalkowski@intel.com>
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
2020-09-16 21:10:08 +00:00
|
|
|
bool tail_call_seen = false;
|
2015-05-19 23:59:04 +00:00
|
|
|
bool seen_exit = false;
|
|
|
|
u8 temp[BPF_MAX_INSN_SIZE + BPF_INSN_SAFETY];
|
bpf: Add x86-64 JIT support for bpf_addr_space_cast instruction.
LLVM generates bpf_addr_space_cast instruction while translating
pointers between native (zero) address space and
__attribute__((address_space(N))).
The addr_space=1 is reserved as bpf_arena address space.
rY = addr_space_cast(rX, 0, 1) is processed by the verifier and
converted to normal 32-bit move: wX = wY
rY = addr_space_cast(rX, 1, 0) has to be converted by JIT:
aux_reg = upper_32_bits of arena->user_vm_start
aux_reg <<= 32
wX = wY // clear upper 32 bits of dst register
if (wX) // if not zero add upper bits of user_vm_start
wX |= aux_reg
JIT can do it more efficiently:
mov dst_reg32, src_reg32 // 32-bit move
shl dst_reg, 32
or dst_reg, user_vm_start
rol dst_reg, 32
xor r11, r11
test dst_reg32, dst_reg32 // check if lower 32-bit are zero
cmove r11, dst_reg // if so, set dst_reg to zero
// Intel swapped src/dst register encoding in CMOVcc
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Signed-off-by: Andrii Nakryiko <andrii@kernel.org>
Acked-by: Eduard Zingerman <eddyz87@gmail.com>
Link: https://lore.kernel.org/bpf/20240308010812.89848-5-alexei.starovoitov@gmail.com
2024-03-08 01:08:02 +00:00
|
|
|
u64 arena_vm_start, user_vm_start;
|
2021-06-23 11:25:04 +00:00
|
|
|
int i, excnt = 0;
|
bpf,x64: Pad NOPs to make images converge more easily
The x64 bpf jit expects bpf images converge within the given passes, but
it could fail to do so with some corner cases. For example:
l0: ja 40
l1: ja 40
[... repeated ja 40 ]
l39: ja 40
l40: ret #0
This bpf program contains 40 "ja 40" instructions which are effectively
NOPs and designed to be replaced with valid code dynamically. Ideally,
bpf jit should optimize those "ja 40" instructions out when translating
the bpf instructions into x64 machine code. However, do_jit() can only
remove one "ja 40" for offset==0 on each pass, so it requires at least
40 runs to eliminate those JMPs and exceeds the current limit of
passes(20). In the end, the program got rejected when BPF_JIT_ALWAYS_ON
is set even though it's legit as a classic socket filter.
To make bpf images more likely converge within 20 passes, this commit
pads some instructions with NOPs in the last 5 passes:
1. conditional jumps
A possible size variance comes from the adoption of imm8 JMP. If the
offset is imm8, we calculate the size difference of this BPF instruction
between the previous and the current pass and fill the gap with NOPs.
To avoid the recalculation of jump offset, those NOPs are inserted before
the JMP code, so we have to subtract the 2 bytes of imm8 JMP when
calculating the NOP number.
2. BPF_JA
There are two conditions for BPF_JA.
a.) nop jumps
If this instruction is not optimized out in the previous pass,
instead of removing it, we insert the equivalent size of NOPs.
b.) label jumps
Similar to condition jumps, we prepend NOPs right before the JMP
code.
To make the code concise, emit_nops() is modified to use the signed len and
return the number of inserted NOPs.
For bpf-to-bpf, we always enable padding for the extra pass since there
is only one extra run and the jump padding doesn't affected the images
that converge without padding.
After applying this patch, the corner case was loaded with the following
jit code:
flen=45 proglen=77 pass=17 image=ffffffffc03367d4 from=jump pid=10097
JIT code: 00000000: 0f 1f 44 00 00 55 48 89 e5 53 41 55 31 c0 45 31
JIT code: 00000010: ed 48 89 fb eb 30 eb 2e eb 2c eb 2a eb 28 eb 26
JIT code: 00000020: eb 24 eb 22 eb 20 eb 1e eb 1c eb 1a eb 18 eb 16
JIT code: 00000030: eb 14 eb 12 eb 10 eb 0e eb 0c eb 0a eb 08 eb 06
JIT code: 00000040: eb 04 eb 02 66 90 31 c0 41 5d 5b c9 c3
0: 0f 1f 44 00 00 nop DWORD PTR [rax+rax*1+0x0]
5: 55 push rbp
6: 48 89 e5 mov rbp,rsp
9: 53 push rbx
a: 41 55 push r13
c: 31 c0 xor eax,eax
e: 45 31 ed xor r13d,r13d
11: 48 89 fb mov rbx,rdi
14: eb 30 jmp 0x46
16: eb 2e jmp 0x46
...
3e: eb 06 jmp 0x46
40: eb 04 jmp 0x46
42: eb 02 jmp 0x46
44: 66 90 xchg ax,ax
46: 31 c0 xor eax,eax
48: 41 5d pop r13
4a: 5b pop rbx
4b: c9 leave
4c: c3 ret
At the 16th pass, 15 jumps were already optimized out, and one jump was
replaced with NOPs at 44 and the image converged at the 17th pass.
v4:
- Add the detailed comments about the possible padding bytes
v3:
- Copy the instructions of prologue separately or the size calculation
of the first BPF instruction would include the prologue.
- Replace WARN_ONCE() with pr_err() and EFAULT
- Use MAX_PASSES in the for loop condition check
- Remove the "padded" flag from x64_jit_data. For the extra pass of
subprogs, padding is always enabled since it won't hurt the images
that converge without padding.
v2:
- Simplify the sample code in the description and provide the jit code
- Check the expected padding bytes with WARN_ONCE
- Move the 'padded' flag to 'struct x64_jit_data'
Signed-off-by: Gary Lin <glin@suse.com>
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Link: https://lore.kernel.org/bpf/20210119102501.511-2-glin@suse.com
2021-01-19 10:24:59 +00:00
|
|
|
int ilen, proglen = 0;
|
2015-05-19 23:59:04 +00:00
|
|
|
u8 *prog = temp;
|
2021-01-14 18:17:44 +00:00
|
|
|
int err;
|
2015-05-19 23:59:04 +00:00
|
|
|
|
2024-03-08 01:08:01 +00:00
|
|
|
arena_vm_start = bpf_arena_get_kern_vm_start(bpf_prog->aux->arena);
|
bpf: Add x86-64 JIT support for bpf_addr_space_cast instruction.
LLVM generates bpf_addr_space_cast instruction while translating
pointers between native (zero) address space and
__attribute__((address_space(N))).
The addr_space=1 is reserved as bpf_arena address space.
rY = addr_space_cast(rX, 0, 1) is processed by the verifier and
converted to normal 32-bit move: wX = wY
rY = addr_space_cast(rX, 1, 0) has to be converted by JIT:
aux_reg = upper_32_bits of arena->user_vm_start
aux_reg <<= 32
wX = wY // clear upper 32 bits of dst register
if (wX) // if not zero add upper bits of user_vm_start
wX |= aux_reg
JIT can do it more efficiently:
mov dst_reg32, src_reg32 // 32-bit move
shl dst_reg, 32
or dst_reg, user_vm_start
rol dst_reg, 32
xor r11, r11
test dst_reg32, dst_reg32 // check if lower 32-bit are zero
cmove r11, dst_reg // if so, set dst_reg to zero
// Intel swapped src/dst register encoding in CMOVcc
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Signed-off-by: Andrii Nakryiko <andrii@kernel.org>
Acked-by: Eduard Zingerman <eddyz87@gmail.com>
Link: https://lore.kernel.org/bpf/20240308010812.89848-5-alexei.starovoitov@gmail.com
2024-03-08 01:08:02 +00:00
|
|
|
user_vm_start = bpf_arena_get_user_vm_start(bpf_prog->aux->arena);
|
2024-03-08 01:08:01 +00:00
|
|
|
|
bpf, x64: rework pro/epilogue and tailcall handling in JIT
This commit serves two things:
1) it optimizes BPF prologue/epilogue generation
2) it makes possible to have tailcalls within BPF subprogram
Both points are related to each other since without 1), 2) could not be
achieved.
In [1], Alexei says:
"The prologue will look like:
nop5
xor eax,eax // two new bytes if bpf_tail_call() is used in this
// function
push rbp
mov rbp, rsp
sub rsp, rounded_stack_depth
push rax // zero init tail_call counter
variable number of push rbx,r13,r14,r15
Then bpf_tail_call will pop variable number rbx,..
and final 'pop rax'
Then 'add rsp, size_of_current_stack_frame'
jmp to next function and skip over 'nop5; xor eax,eax; push rpb; mov
rbp, rsp'
This way new function will set its own stack size and will init tail
call
counter with whatever value the parent had.
If next function doesn't use bpf_tail_call it won't have 'xor eax,eax'.
Instead it would need to have 'nop2' in there."
Implement that suggestion.
Since the layout of stack is changed, tail call counter handling can not
rely anymore on popping it to rbx just like it have been handled for
constant prologue case and later overwrite of rbx with actual value of
rbx pushed to stack. Therefore, let's use one of the register (%rcx) that
is considered to be volatile/caller-saved and pop the value of tail call
counter in there in the epilogue.
Drop the BUILD_BUG_ON in emit_prologue and in
emit_bpf_tail_call_indirect where instruction layout is not constant
anymore.
Introduce new poke target, 'tailcall_bypass' to poke descriptor that is
dedicated for skipping the register pops and stack unwind that are
generated right before the actual jump to target program.
For case when the target program is not present, BPF program will skip
the pop instructions and nop5 dedicated for jmpq $target. An example of
such state when only R6 of callee saved registers is used by program:
ffffffffc0513aa1: e9 0e 00 00 00 jmpq 0xffffffffc0513ab4
ffffffffc0513aa6: 5b pop %rbx
ffffffffc0513aa7: 58 pop %rax
ffffffffc0513aa8: 48 81 c4 00 00 00 00 add $0x0,%rsp
ffffffffc0513aaf: 0f 1f 44 00 00 nopl 0x0(%rax,%rax,1)
ffffffffc0513ab4: 48 89 df mov %rbx,%rdi
When target program is inserted, the jump that was there to skip
pops/nop5 will become the nop5, so CPU will go over pops and do the
actual tailcall.
One might ask why there simply can not be pushes after the nop5?
In the following example snippet:
ffffffffc037030c: 48 89 fb mov %rdi,%rbx
(...)
ffffffffc0370332: 5b pop %rbx
ffffffffc0370333: 58 pop %rax
ffffffffc0370334: 48 81 c4 00 00 00 00 add $0x0,%rsp
ffffffffc037033b: 0f 1f 44 00 00 nopl 0x0(%rax,%rax,1)
ffffffffc0370340: 48 81 ec 00 00 00 00 sub $0x0,%rsp
ffffffffc0370347: 50 push %rax
ffffffffc0370348: 53 push %rbx
ffffffffc0370349: 48 89 df mov %rbx,%rdi
ffffffffc037034c: e8 f7 21 00 00 callq 0xffffffffc0372548
There is the bpf2bpf call (at ffffffffc037034c) right after the tailcall
and jump target is not present. ctx is in %rbx register and BPF
subprogram that we will call into on ffffffffc037034c is relying on it,
e.g. it will pick ctx from there. Such code layout is therefore broken
as we would overwrite the content of %rbx with the value that was pushed
on the prologue. That is the reason for the 'bypass' approach.
Special care needs to be taken during the install/update/remove of
tailcall target. In case when target program is not present, the CPU
must not execute the pop instructions that precede the tailcall.
To address that, the following states can be defined:
A nop, unwind, nop
B nop, unwind, tail
C skip, unwind, nop
D skip, unwind, tail
A is forbidden (lead to incorrectness). The state transitions between
tailcall install/update/remove will work as follows:
First install tail call f: C->D->B(f)
* poke the tailcall, after that get rid of the skip
Update tail call f to f': B(f)->B(f')
* poke the tailcall (poke->tailcall_target) and do NOT touch the
poke->tailcall_bypass
Remove tail call: B(f')->C(f')
* poke->tailcall_bypass is poked back to jump, then we wait the RCU
grace period so that other programs will finish its execution and
after that we are safe to remove the poke->tailcall_target
Install new tail call (f''): C(f')->D(f'')->B(f'').
* same as first step
This way CPU can never be exposed to "unwind, tail" state.
Last but not least, when tailcalls get mixed with bpf2bpf calls, it
would be possible to encounter the endless loop due to clearing the
tailcall counter if for example we would use the tailcall3-like from BPF
selftests program that would be subprogram-based, meaning the tailcall
would be present within the BPF subprogram.
This test, broken down to particular steps, would do:
entry -> set tailcall counter to 0, bump it by 1, tailcall to func0
func0 -> call subprog_tail
(we are NOT skipping the first 11 bytes of prologue and this subprogram
has a tailcall, therefore we clear the counter...)
subprog -> do the same thing as entry
and then loop forever.
To address this, the idea is to go through the call chain of bpf2bpf progs
and look for a tailcall presence throughout whole chain. If we saw a single
tail call then each node in this call chain needs to be marked as a subprog
that can reach the tailcall. We would later feed the JIT with this info
and:
- set eax to 0 only when tailcall is reachable and this is the entry prog
- if tailcall is reachable but there's no tailcall in insns of currently
JITed prog then push rax anyway, so that it will be possible to
propagate further down the call chain
- finally if tailcall is reachable, then we need to precede the 'call'
insn with mov rax, [rbp - (stack_depth + 8)]
Tail call related cases from test_verifier kselftest are also working
fine. Sample BPF programs that utilize tail calls (sockex3, tracex5)
work properly as well.
[1]: https://lore.kernel.org/bpf/20200517043227.2gpq22ifoq37ogst@ast-mbp.dhcp.thefacebook.com/
Suggested-by: Alexei Starovoitov <ast@kernel.org>
Signed-off-by: Maciej Fijalkowski <maciej.fijalkowski@intel.com>
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
2020-09-16 21:10:08 +00:00
|
|
|
detect_reg_usage(insn, insn_cnt, callee_regs_used,
|
|
|
|
&tail_call_seen);
|
|
|
|
|
|
|
|
/* tail call's presence in current prog implies it is reachable */
|
|
|
|
tail_call_reachable |= tail_call_seen;
|
|
|
|
|
2018-02-24 00:08:02 +00:00
|
|
|
emit_prologue(&prog, bpf_prog->aux->stack_depth,
|
bpf, x64: rework pro/epilogue and tailcall handling in JIT
This commit serves two things:
1) it optimizes BPF prologue/epilogue generation
2) it makes possible to have tailcalls within BPF subprogram
Both points are related to each other since without 1), 2) could not be
achieved.
In [1], Alexei says:
"The prologue will look like:
nop5
xor eax,eax // two new bytes if bpf_tail_call() is used in this
// function
push rbp
mov rbp, rsp
sub rsp, rounded_stack_depth
push rax // zero init tail_call counter
variable number of push rbx,r13,r14,r15
Then bpf_tail_call will pop variable number rbx,..
and final 'pop rax'
Then 'add rsp, size_of_current_stack_frame'
jmp to next function and skip over 'nop5; xor eax,eax; push rpb; mov
rbp, rsp'
This way new function will set its own stack size and will init tail
call
counter with whatever value the parent had.
If next function doesn't use bpf_tail_call it won't have 'xor eax,eax'.
Instead it would need to have 'nop2' in there."
Implement that suggestion.
Since the layout of stack is changed, tail call counter handling can not
rely anymore on popping it to rbx just like it have been handled for
constant prologue case and later overwrite of rbx with actual value of
rbx pushed to stack. Therefore, let's use one of the register (%rcx) that
is considered to be volatile/caller-saved and pop the value of tail call
counter in there in the epilogue.
Drop the BUILD_BUG_ON in emit_prologue and in
emit_bpf_tail_call_indirect where instruction layout is not constant
anymore.
Introduce new poke target, 'tailcall_bypass' to poke descriptor that is
dedicated for skipping the register pops and stack unwind that are
generated right before the actual jump to target program.
For case when the target program is not present, BPF program will skip
the pop instructions and nop5 dedicated for jmpq $target. An example of
such state when only R6 of callee saved registers is used by program:
ffffffffc0513aa1: e9 0e 00 00 00 jmpq 0xffffffffc0513ab4
ffffffffc0513aa6: 5b pop %rbx
ffffffffc0513aa7: 58 pop %rax
ffffffffc0513aa8: 48 81 c4 00 00 00 00 add $0x0,%rsp
ffffffffc0513aaf: 0f 1f 44 00 00 nopl 0x0(%rax,%rax,1)
ffffffffc0513ab4: 48 89 df mov %rbx,%rdi
When target program is inserted, the jump that was there to skip
pops/nop5 will become the nop5, so CPU will go over pops and do the
actual tailcall.
One might ask why there simply can not be pushes after the nop5?
In the following example snippet:
ffffffffc037030c: 48 89 fb mov %rdi,%rbx
(...)
ffffffffc0370332: 5b pop %rbx
ffffffffc0370333: 58 pop %rax
ffffffffc0370334: 48 81 c4 00 00 00 00 add $0x0,%rsp
ffffffffc037033b: 0f 1f 44 00 00 nopl 0x0(%rax,%rax,1)
ffffffffc0370340: 48 81 ec 00 00 00 00 sub $0x0,%rsp
ffffffffc0370347: 50 push %rax
ffffffffc0370348: 53 push %rbx
ffffffffc0370349: 48 89 df mov %rbx,%rdi
ffffffffc037034c: e8 f7 21 00 00 callq 0xffffffffc0372548
There is the bpf2bpf call (at ffffffffc037034c) right after the tailcall
and jump target is not present. ctx is in %rbx register and BPF
subprogram that we will call into on ffffffffc037034c is relying on it,
e.g. it will pick ctx from there. Such code layout is therefore broken
as we would overwrite the content of %rbx with the value that was pushed
on the prologue. That is the reason for the 'bypass' approach.
Special care needs to be taken during the install/update/remove of
tailcall target. In case when target program is not present, the CPU
must not execute the pop instructions that precede the tailcall.
To address that, the following states can be defined:
A nop, unwind, nop
B nop, unwind, tail
C skip, unwind, nop
D skip, unwind, tail
A is forbidden (lead to incorrectness). The state transitions between
tailcall install/update/remove will work as follows:
First install tail call f: C->D->B(f)
* poke the tailcall, after that get rid of the skip
Update tail call f to f': B(f)->B(f')
* poke the tailcall (poke->tailcall_target) and do NOT touch the
poke->tailcall_bypass
Remove tail call: B(f')->C(f')
* poke->tailcall_bypass is poked back to jump, then we wait the RCU
grace period so that other programs will finish its execution and
after that we are safe to remove the poke->tailcall_target
Install new tail call (f''): C(f')->D(f'')->B(f'').
* same as first step
This way CPU can never be exposed to "unwind, tail" state.
Last but not least, when tailcalls get mixed with bpf2bpf calls, it
would be possible to encounter the endless loop due to clearing the
tailcall counter if for example we would use the tailcall3-like from BPF
selftests program that would be subprogram-based, meaning the tailcall
would be present within the BPF subprogram.
This test, broken down to particular steps, would do:
entry -> set tailcall counter to 0, bump it by 1, tailcall to func0
func0 -> call subprog_tail
(we are NOT skipping the first 11 bytes of prologue and this subprogram
has a tailcall, therefore we clear the counter...)
subprog -> do the same thing as entry
and then loop forever.
To address this, the idea is to go through the call chain of bpf2bpf progs
and look for a tailcall presence throughout whole chain. If we saw a single
tail call then each node in this call chain needs to be marked as a subprog
that can reach the tailcall. We would later feed the JIT with this info
and:
- set eax to 0 only when tailcall is reachable and this is the entry prog
- if tailcall is reachable but there's no tailcall in insns of currently
JITed prog then push rax anyway, so that it will be possible to
propagate further down the call chain
- finally if tailcall is reachable, then we need to precede the 'call'
insn with mov rax, [rbp - (stack_depth + 8)]
Tail call related cases from test_verifier kselftest are also working
fine. Sample BPF programs that utilize tail calls (sockex3, tracex5)
work properly as well.
[1]: https://lore.kernel.org/bpf/20200517043227.2gpq22ifoq37ogst@ast-mbp.dhcp.thefacebook.com/
Suggested-by: Alexei Starovoitov <ast@kernel.org>
Signed-off-by: Maciej Fijalkowski <maciej.fijalkowski@intel.com>
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
2020-09-16 21:10:08 +00:00
|
|
|
bpf_prog_was_classic(bpf_prog), tail_call_reachable,
|
bpf: Implement BPF exceptions
This patch implements BPF exceptions, and introduces a bpf_throw kfunc
to allow programs to throw exceptions during their execution at runtime.
A bpf_throw invocation is treated as an immediate termination of the
program, returning back to its caller within the kernel, unwinding all
stack frames.
This allows the program to simplify its implementation, by testing for
runtime conditions which the verifier has no visibility into, and assert
that they are true. In case they are not, the program can simply throw
an exception from the other branch.
BPF exceptions are explicitly *NOT* an unlikely slowpath error handling
primitive, and this objective has guided design choices of the
implementation of the them within the kernel (with the bulk of the cost
for unwinding the stack offloaded to the bpf_throw kfunc).
The implementation of this mechanism requires use of add_hidden_subprog
mechanism introduced in the previous patch, which generates a couple of
instructions to move R1 to R0 and exit. The JIT then rewrites the
prologue of this subprog to take the stack pointer and frame pointer as
inputs and reset the stack frame, popping all callee-saved registers
saved by the main subprog. The bpf_throw function then walks the stack
at runtime, and invokes this exception subprog with the stack and frame
pointers as parameters.
Reviewers must take note that currently the main program is made to save
all callee-saved registers on x86_64 during entry into the program. This
is because we must do an equivalent of a lightweight context switch when
unwinding the stack, therefore we need the callee-saved registers of the
caller of the BPF program to be able to return with a sane state.
Note that we have to additionally handle r12, even though it is not used
by the program, because when throwing the exception the program makes an
entry into the kernel which could clobber r12 after saving it on the
stack. To be able to preserve the value we received on program entry, we
push r12 and restore it from the generated subprogram when unwinding the
stack.
For now, bpf_throw invocation fails when lingering resources or locks
exist in that path of the program. In a future followup, bpf_throw will
be extended to perform frame-by-frame unwinding to release lingering
resources for each stack frame, removing this limitation.
Signed-off-by: Kumar Kartikeya Dwivedi <memxor@gmail.com>
Link: https://lore.kernel.org/r/20230912233214.1518551-5-memxor@gmail.com
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
2023-09-12 23:32:01 +00:00
|
|
|
bpf_is_subprog(bpf_prog), bpf_prog->aux->exception_cb);
|
|
|
|
/* Exception callback will clobber callee regs for its own use, and
|
|
|
|
* restore the original callee regs from main prog's stack frame.
|
|
|
|
*/
|
|
|
|
if (bpf_prog->aux->exception_boundary) {
|
|
|
|
/* We also need to save r12, which is not mapped to any BPF
|
|
|
|
* register, as we throw after entry into the kernel, which may
|
|
|
|
* overwrite r12.
|
|
|
|
*/
|
|
|
|
push_r12(&prog);
|
|
|
|
push_callee_regs(&prog, all_callee_regs_used);
|
|
|
|
} else {
|
2024-03-08 01:08:01 +00:00
|
|
|
if (arena_vm_start)
|
|
|
|
push_r12(&prog);
|
bpf: Implement BPF exceptions
This patch implements BPF exceptions, and introduces a bpf_throw kfunc
to allow programs to throw exceptions during their execution at runtime.
A bpf_throw invocation is treated as an immediate termination of the
program, returning back to its caller within the kernel, unwinding all
stack frames.
This allows the program to simplify its implementation, by testing for
runtime conditions which the verifier has no visibility into, and assert
that they are true. In case they are not, the program can simply throw
an exception from the other branch.
BPF exceptions are explicitly *NOT* an unlikely slowpath error handling
primitive, and this objective has guided design choices of the
implementation of the them within the kernel (with the bulk of the cost
for unwinding the stack offloaded to the bpf_throw kfunc).
The implementation of this mechanism requires use of add_hidden_subprog
mechanism introduced in the previous patch, which generates a couple of
instructions to move R1 to R0 and exit. The JIT then rewrites the
prologue of this subprog to take the stack pointer and frame pointer as
inputs and reset the stack frame, popping all callee-saved registers
saved by the main subprog. The bpf_throw function then walks the stack
at runtime, and invokes this exception subprog with the stack and frame
pointers as parameters.
Reviewers must take note that currently the main program is made to save
all callee-saved registers on x86_64 during entry into the program. This
is because we must do an equivalent of a lightweight context switch when
unwinding the stack, therefore we need the callee-saved registers of the
caller of the BPF program to be able to return with a sane state.
Note that we have to additionally handle r12, even though it is not used
by the program, because when throwing the exception the program makes an
entry into the kernel which could clobber r12 after saving it on the
stack. To be able to preserve the value we received on program entry, we
push r12 and restore it from the generated subprogram when unwinding the
stack.
For now, bpf_throw invocation fails when lingering resources or locks
exist in that path of the program. In a future followup, bpf_throw will
be extended to perform frame-by-frame unwinding to release lingering
resources for each stack frame, removing this limitation.
Signed-off-by: Kumar Kartikeya Dwivedi <memxor@gmail.com>
Link: https://lore.kernel.org/r/20230912233214.1518551-5-memxor@gmail.com
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
2023-09-12 23:32:01 +00:00
|
|
|
push_callee_regs(&prog, callee_regs_used);
|
|
|
|
}
|
2024-03-08 01:08:01 +00:00
|
|
|
if (arena_vm_start)
|
|
|
|
emit_mov_imm64(&prog, X86_REG_R12,
|
|
|
|
arena_vm_start >> 32, (u32) arena_vm_start);
|
bpf,x64: Pad NOPs to make images converge more easily
The x64 bpf jit expects bpf images converge within the given passes, but
it could fail to do so with some corner cases. For example:
l0: ja 40
l1: ja 40
[... repeated ja 40 ]
l39: ja 40
l40: ret #0
This bpf program contains 40 "ja 40" instructions which are effectively
NOPs and designed to be replaced with valid code dynamically. Ideally,
bpf jit should optimize those "ja 40" instructions out when translating
the bpf instructions into x64 machine code. However, do_jit() can only
remove one "ja 40" for offset==0 on each pass, so it requires at least
40 runs to eliminate those JMPs and exceeds the current limit of
passes(20). In the end, the program got rejected when BPF_JIT_ALWAYS_ON
is set even though it's legit as a classic socket filter.
To make bpf images more likely converge within 20 passes, this commit
pads some instructions with NOPs in the last 5 passes:
1. conditional jumps
A possible size variance comes from the adoption of imm8 JMP. If the
offset is imm8, we calculate the size difference of this BPF instruction
between the previous and the current pass and fill the gap with NOPs.
To avoid the recalculation of jump offset, those NOPs are inserted before
the JMP code, so we have to subtract the 2 bytes of imm8 JMP when
calculating the NOP number.
2. BPF_JA
There are two conditions for BPF_JA.
a.) nop jumps
If this instruction is not optimized out in the previous pass,
instead of removing it, we insert the equivalent size of NOPs.
b.) label jumps
Similar to condition jumps, we prepend NOPs right before the JMP
code.
To make the code concise, emit_nops() is modified to use the signed len and
return the number of inserted NOPs.
For bpf-to-bpf, we always enable padding for the extra pass since there
is only one extra run and the jump padding doesn't affected the images
that converge without padding.
After applying this patch, the corner case was loaded with the following
jit code:
flen=45 proglen=77 pass=17 image=ffffffffc03367d4 from=jump pid=10097
JIT code: 00000000: 0f 1f 44 00 00 55 48 89 e5 53 41 55 31 c0 45 31
JIT code: 00000010: ed 48 89 fb eb 30 eb 2e eb 2c eb 2a eb 28 eb 26
JIT code: 00000020: eb 24 eb 22 eb 20 eb 1e eb 1c eb 1a eb 18 eb 16
JIT code: 00000030: eb 14 eb 12 eb 10 eb 0e eb 0c eb 0a eb 08 eb 06
JIT code: 00000040: eb 04 eb 02 66 90 31 c0 41 5d 5b c9 c3
0: 0f 1f 44 00 00 nop DWORD PTR [rax+rax*1+0x0]
5: 55 push rbp
6: 48 89 e5 mov rbp,rsp
9: 53 push rbx
a: 41 55 push r13
c: 31 c0 xor eax,eax
e: 45 31 ed xor r13d,r13d
11: 48 89 fb mov rbx,rdi
14: eb 30 jmp 0x46
16: eb 2e jmp 0x46
...
3e: eb 06 jmp 0x46
40: eb 04 jmp 0x46
42: eb 02 jmp 0x46
44: 66 90 xchg ax,ax
46: 31 c0 xor eax,eax
48: 41 5d pop r13
4a: 5b pop rbx
4b: c9 leave
4c: c3 ret
At the 16th pass, 15 jumps were already optimized out, and one jump was
replaced with NOPs at 44 and the image converged at the 17th pass.
v4:
- Add the detailed comments about the possible padding bytes
v3:
- Copy the instructions of prologue separately or the size calculation
of the first BPF instruction would include the prologue.
- Replace WARN_ONCE() with pr_err() and EFAULT
- Use MAX_PASSES in the for loop condition check
- Remove the "padded" flag from x64_jit_data. For the extra pass of
subprogs, padding is always enabled since it won't hurt the images
that converge without padding.
v2:
- Simplify the sample code in the description and provide the jit code
- Check the expected padding bytes with WARN_ONCE
- Move the 'padded' flag to 'struct x64_jit_data'
Signed-off-by: Gary Lin <glin@suse.com>
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Link: https://lore.kernel.org/bpf/20210119102501.511-2-glin@suse.com
2021-01-19 10:24:59 +00:00
|
|
|
|
|
|
|
ilen = prog - temp;
|
2022-02-04 18:57:42 +00:00
|
|
|
if (rw_image)
|
|
|
|
memcpy(rw_image + proglen, temp, ilen);
|
bpf,x64: Pad NOPs to make images converge more easily
The x64 bpf jit expects bpf images converge within the given passes, but
it could fail to do so with some corner cases. For example:
l0: ja 40
l1: ja 40
[... repeated ja 40 ]
l39: ja 40
l40: ret #0
This bpf program contains 40 "ja 40" instructions which are effectively
NOPs and designed to be replaced with valid code dynamically. Ideally,
bpf jit should optimize those "ja 40" instructions out when translating
the bpf instructions into x64 machine code. However, do_jit() can only
remove one "ja 40" for offset==0 on each pass, so it requires at least
40 runs to eliminate those JMPs and exceeds the current limit of
passes(20). In the end, the program got rejected when BPF_JIT_ALWAYS_ON
is set even though it's legit as a classic socket filter.
To make bpf images more likely converge within 20 passes, this commit
pads some instructions with NOPs in the last 5 passes:
1. conditional jumps
A possible size variance comes from the adoption of imm8 JMP. If the
offset is imm8, we calculate the size difference of this BPF instruction
between the previous and the current pass and fill the gap with NOPs.
To avoid the recalculation of jump offset, those NOPs are inserted before
the JMP code, so we have to subtract the 2 bytes of imm8 JMP when
calculating the NOP number.
2. BPF_JA
There are two conditions for BPF_JA.
a.) nop jumps
If this instruction is not optimized out in the previous pass,
instead of removing it, we insert the equivalent size of NOPs.
b.) label jumps
Similar to condition jumps, we prepend NOPs right before the JMP
code.
To make the code concise, emit_nops() is modified to use the signed len and
return the number of inserted NOPs.
For bpf-to-bpf, we always enable padding for the extra pass since there
is only one extra run and the jump padding doesn't affected the images
that converge without padding.
After applying this patch, the corner case was loaded with the following
jit code:
flen=45 proglen=77 pass=17 image=ffffffffc03367d4 from=jump pid=10097
JIT code: 00000000: 0f 1f 44 00 00 55 48 89 e5 53 41 55 31 c0 45 31
JIT code: 00000010: ed 48 89 fb eb 30 eb 2e eb 2c eb 2a eb 28 eb 26
JIT code: 00000020: eb 24 eb 22 eb 20 eb 1e eb 1c eb 1a eb 18 eb 16
JIT code: 00000030: eb 14 eb 12 eb 10 eb 0e eb 0c eb 0a eb 08 eb 06
JIT code: 00000040: eb 04 eb 02 66 90 31 c0 41 5d 5b c9 c3
0: 0f 1f 44 00 00 nop DWORD PTR [rax+rax*1+0x0]
5: 55 push rbp
6: 48 89 e5 mov rbp,rsp
9: 53 push rbx
a: 41 55 push r13
c: 31 c0 xor eax,eax
e: 45 31 ed xor r13d,r13d
11: 48 89 fb mov rbx,rdi
14: eb 30 jmp 0x46
16: eb 2e jmp 0x46
...
3e: eb 06 jmp 0x46
40: eb 04 jmp 0x46
42: eb 02 jmp 0x46
44: 66 90 xchg ax,ax
46: 31 c0 xor eax,eax
48: 41 5d pop r13
4a: 5b pop rbx
4b: c9 leave
4c: c3 ret
At the 16th pass, 15 jumps were already optimized out, and one jump was
replaced with NOPs at 44 and the image converged at the 17th pass.
v4:
- Add the detailed comments about the possible padding bytes
v3:
- Copy the instructions of prologue separately or the size calculation
of the first BPF instruction would include the prologue.
- Replace WARN_ONCE() with pr_err() and EFAULT
- Use MAX_PASSES in the for loop condition check
- Remove the "padded" flag from x64_jit_data. For the extra pass of
subprogs, padding is always enabled since it won't hurt the images
that converge without padding.
v2:
- Simplify the sample code in the description and provide the jit code
- Check the expected padding bytes with WARN_ONCE
- Move the 'padded' flag to 'struct x64_jit_data'
Signed-off-by: Gary Lin <glin@suse.com>
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Link: https://lore.kernel.org/bpf/20210119102501.511-2-glin@suse.com
2021-01-19 10:24:59 +00:00
|
|
|
proglen += ilen;
|
|
|
|
addrs[0] = proglen;
|
|
|
|
prog = temp;
|
2015-05-19 23:59:04 +00:00
|
|
|
|
2019-07-31 01:38:26 +00:00
|
|
|
for (i = 1; i <= insn_cnt; i++, insn++) {
|
2014-06-06 21:46:06 +00:00
|
|
|
const s32 imm32 = insn->imm;
|
|
|
|
u32 dst_reg = insn->dst_reg;
|
|
|
|
u32 src_reg = insn->src_reg;
|
2018-02-24 00:07:59 +00:00
|
|
|
u8 b2 = 0, b3 = 0;
|
2021-02-02 05:38:37 +00:00
|
|
|
u8 *start_of_ldx;
|
net: filter: x86: internal BPF JIT
Maps all internal BPF instructions into x86_64 instructions.
This patch replaces original BPF x64 JIT with internal BPF x64 JIT.
sysctl net.core.bpf_jit_enable is reused as on/off switch.
Performance:
1. old BPF JIT and internal BPF JIT generate equivalent x86_64 code.
No performance difference is observed for filters that were JIT-able before
Example assembler code for BPF filter "tcpdump port 22"
original BPF -> old JIT: original BPF -> internal BPF -> new JIT:
0: push %rbp 0: push %rbp
1: mov %rsp,%rbp 1: mov %rsp,%rbp
4: sub $0x60,%rsp 4: sub $0x228,%rsp
8: mov %rbx,-0x8(%rbp) b: mov %rbx,-0x228(%rbp) // prologue
12: mov %r13,-0x220(%rbp)
19: mov %r14,-0x218(%rbp)
20: mov %r15,-0x210(%rbp)
27: xor %eax,%eax // clear A
c: xor %ebx,%ebx 29: xor %r13,%r13 // clear X
e: mov 0x68(%rdi),%r9d 2c: mov 0x68(%rdi),%r9d
12: sub 0x6c(%rdi),%r9d 30: sub 0x6c(%rdi),%r9d
16: mov 0xd8(%rdi),%r8 34: mov 0xd8(%rdi),%r10
3b: mov %rdi,%rbx
1d: mov $0xc,%esi 3e: mov $0xc,%esi
22: callq 0xffffffffe1021e15 43: callq 0xffffffffe102bd75
27: cmp $0x86dd,%eax 48: cmp $0x86dd,%rax
2c: jne 0x0000000000000069 4f: jne 0x000000000000009a
2e: mov $0x14,%esi 51: mov $0x14,%esi
33: callq 0xffffffffe1021e31 56: callq 0xffffffffe102bd91
38: cmp $0x84,%eax 5b: cmp $0x84,%rax
3d: je 0x0000000000000049 62: je 0x0000000000000074
3f: cmp $0x6,%eax 64: cmp $0x6,%rax
42: je 0x0000000000000049 68: je 0x0000000000000074
44: cmp $0x11,%eax 6a: cmp $0x11,%rax
47: jne 0x00000000000000c6 6e: jne 0x0000000000000117
49: mov $0x36,%esi 74: mov $0x36,%esi
4e: callq 0xffffffffe1021e15 79: callq 0xffffffffe102bd75
53: cmp $0x16,%eax 7e: cmp $0x16,%rax
56: je 0x00000000000000bf 82: je 0x0000000000000110
58: mov $0x38,%esi 88: mov $0x38,%esi
5d: callq 0xffffffffe1021e15 8d: callq 0xffffffffe102bd75
62: cmp $0x16,%eax 92: cmp $0x16,%rax
65: je 0x00000000000000bf 96: je 0x0000000000000110
67: jmp 0x00000000000000c6 98: jmp 0x0000000000000117
69: cmp $0x800,%eax 9a: cmp $0x800,%rax
6e: jne 0x00000000000000c6 a1: jne 0x0000000000000117
70: mov $0x17,%esi a3: mov $0x17,%esi
75: callq 0xffffffffe1021e31 a8: callq 0xffffffffe102bd91
7a: cmp $0x84,%eax ad: cmp $0x84,%rax
7f: je 0x000000000000008b b4: je 0x00000000000000c2
81: cmp $0x6,%eax b6: cmp $0x6,%rax
84: je 0x000000000000008b ba: je 0x00000000000000c2
86: cmp $0x11,%eax bc: cmp $0x11,%rax
89: jne 0x00000000000000c6 c0: jne 0x0000000000000117
8b: mov $0x14,%esi c2: mov $0x14,%esi
90: callq 0xffffffffe1021e15 c7: callq 0xffffffffe102bd75
95: test $0x1fff,%ax cc: test $0x1fff,%rax
99: jne 0x00000000000000c6 d3: jne 0x0000000000000117
d5: mov %rax,%r14
9b: mov $0xe,%esi d8: mov $0xe,%esi
a0: callq 0xffffffffe1021e44 dd: callq 0xffffffffe102bd91 // MSH
e2: and $0xf,%eax
e5: shl $0x2,%eax
e8: mov %rax,%r13
eb: mov %r14,%rax
ee: mov %r13,%rsi
a5: lea 0xe(%rbx),%esi f1: add $0xe,%esi
a8: callq 0xffffffffe1021e0d f4: callq 0xffffffffe102bd6d
ad: cmp $0x16,%eax f9: cmp $0x16,%rax
b0: je 0x00000000000000bf fd: je 0x0000000000000110
ff: mov %r13,%rsi
b2: lea 0x10(%rbx),%esi 102: add $0x10,%esi
b5: callq 0xffffffffe1021e0d 105: callq 0xffffffffe102bd6d
ba: cmp $0x16,%eax 10a: cmp $0x16,%rax
bd: jne 0x00000000000000c6 10e: jne 0x0000000000000117
bf: mov $0xffff,%eax 110: mov $0xffff,%eax
c4: jmp 0x00000000000000c8 115: jmp 0x000000000000011c
c6: xor %eax,%eax 117: mov $0x0,%eax
c8: mov -0x8(%rbp),%rbx 11c: mov -0x228(%rbp),%rbx // epilogue
cc: leaveq 123: mov -0x220(%rbp),%r13
cd: retq 12a: mov -0x218(%rbp),%r14
131: mov -0x210(%rbp),%r15
138: leaveq
139: retq
On fully cached SKBs both JITed functions take 12 nsec to execute.
BPF interpreter executes the program in 30 nsec.
The difference in generated assembler is due to the following:
Old BPF imlements LDX_MSH instruction via sk_load_byte_msh() helper function
inside bpf_jit.S.
New JIT removes the helper and does it explicitly, so ldx_msh cost
is the same for both JITs, but generated code looks longer.
New JIT has 4 registers to save, so prologue/epilogue are larger,
but the cost is within noise on x64.
Old JIT checks whether first insn clears A and if not emits 'xor %eax,%eax'.
New JIT clears %rax unconditionally.
2. old BPF JIT doesn't support ANC_NLATTR, ANC_PAY_OFFSET, ANC_RANDOM
extensions. New JIT supports all BPF extensions.
Performance of such filters improves 2-4 times depending on a filter.
The longer the filter the higher performance gain.
Synthetic benchmarks with many ancillary loads see 20x speedup
which seems to be the maximum gain from JIT
Notes:
. net.core.bpf_jit_enable=2 + tools/net/bpf_jit_disasm is still functional
and can be used to see generated assembler
. there are two jit_compile() functions and code flow for classic filters is:
sk_attach_filter() - load classic BPF
bpf_jit_compile() - try to JIT from classic BPF
sk_convert_filter() - convert classic to internal
bpf_int_jit_compile() - JIT from internal BPF
seccomp and tracing filters will just call bpf_int_jit_compile()
Signed-off-by: Alexei Starovoitov <ast@plumgrid.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
2014-05-14 02:50:46 +00:00
|
|
|
s64 jmp_offset;
|
bpf, x86: Improve PROBE_MEM runtime load check
This patch rewrites the runtime PROBE_MEM check insns emitted by the BPF
JIT in order to ensure load safety. The changes in the patch fix two
issues with the previous logic and more generally improve size of
emitted code. Paragraphs between this one and "FIX 1" below explain the
purpose of the runtime check and examine the current implementation.
When a load is marked PROBE_MEM - e.g. due to PTR_UNTRUSTED access - the
address being loaded from is not necessarily valid. The BPF jit sets up
exception handlers for each such load which catch page faults and 0 out
the destination register.
Arbitrary register-relative loads can escape this exception handling
mechanism. Specifically, a load like dst_reg = *(src_reg + off) will not
trigger BPF exception handling if (src_reg + off) is outside of kernel
address space, resulting in an uncaught page fault. A concrete example
of such behavior is a program like:
struct result {
char space[40];
long a;
};
/* if err, returns ERR_PTR(-EINVAL) */
struct result *ptr = get_ptr_maybe_err();
long x = ptr->a;
If get_ptr_maybe_err returns ERR_PTR(-EINVAL) and the result isn't
checked for err, 'result' will be (u64)-EINVAL, a number close to
U64_MAX. The ptr->a load will be > U64_MAX and will wrap over to a small
positive u64, which will be in userspace and thus not covered by BPF
exception handling mechanism.
In order to prevent such loads from occurring, the BPF jit emits some
instructions which do runtime checking of (src_reg + off) and skip the
actual load if it's out of range. As an example, here are instructions
emitted for a %rdi = *(%rdi + 0x10) PROBE_MEM load:
72: movabs $0x800000000010,%r11 --|
7c: cmp %r11,%rdi |- 72 - 7f: Check 1
7f: jb 0x000000000000008d --|
81: mov %rdi,%r11 -----|
84: add $0x0000000000000010,%r11 |- 81-8b: Check 2
8b: jnc 0x0000000000000091 -----|
8d: xor %edi,%edi ---- 0 out dest
8f: jmp 0x0000000000000095
91: mov 0x10(%rdi),%rdi ---- Actual load
95:
The JIT considers kernel address space to start at MAX_TASK_SIZE +
PAGE_SIZE. Determining whether a load will be outside of kernel address
space should be a simple check:
(src_reg + off) >= MAX_TASK_SIZE + PAGE_SIZE
But because there is only one spare register when the checking logic is
emitted, this logic is split into two checks:
Check 1: src_reg >= (MAX_TASK_SIZE + PAGE_SIZE - off)
Check 2: src_reg + off doesn't wrap over U64_MAX and result in small pos u64
Emitted insns implementing Checks 1 and 2 are annotated in the above
example. Check 1 can be done with a single spare register since the
source reg by definition is the left-hand-side of the inequality.
Since adding 'off' to both sides of Check 1's inequality results in the
original inequality we want, it's equivalent to testing that inequality.
Except in the case where src_reg + off wraps past U64_MAX, which is why
Check 2 needs to actually add src_reg + off if Check 1 passes - again
using the single spare reg.
FIX 1: The Check 1 inequality listed above is not what current code is
doing. Current code is a bit more pessimistic, instead checking:
src_reg >= (MAX_TASK_SIZE + PAGE_SIZE + abs(off))
The 0x800000000010 in above example is from this current check. If Check
1 was corrected to use the correct right-hand-side, the value would be
0x7ffffffffff0. This patch changes the checking logic more broadly (FIX
2 below will elaborate), fixing this issue as a side-effect of the
rewrite. Regardless, it's important to understand why Check 1 should've
been doing MAX_TASK_SIZE + PAGE_SIZE - off before proceeding.
FIX 2: Current code relies on a 'jnc' to determine whether src_reg + off
addition wrapped over. For negative offsets this logic is incorrect.
Consider Check 2 insns emitted when off = -0x10:
81: mov %rdi,%r11
84: add 0xfffffffffffffff0,%r11
8b: jnc 0x0000000000000091
2's complement representation of -0x10 is a large positive u64. Any
value of src_reg that passes Check 1 will result in carry flag being set
after (src_reg + off) addition. So a load with any negative offset will
always fail Check 2 at runtime and never do the actual load. This patch
fixes the negative offset issue by rewriting both checks in order to not
rely on carry flag.
The rewrite takes advantage of the fact that, while we only have one
scratch reg to hold arbitrary values, we know the offset at JIT time.
This we can use src_reg as a temporary scratch reg to hold src_reg +
offset since we can return it to its original value by later subtracting
offset. As a result we can directly check the original inequality we
care about:
(src_reg + off) >= MAX_TASK_SIZE + PAGE_SIZE
For a load like %rdi = *(%rsi + -0x10), this results in emitted code:
43: movabs $0x800000000000,%r11
4d: add $0xfffffffffffffff0,%rsi --- src_reg += off
54: cmp %r11,%rsi --- Check original inequality
57: jae 0x000000000000005d
59: xor %edi,%edi
5b: jmp 0x0000000000000061
5d: mov 0x0(%rdi),%rsi --- Actual Load
61: sub $0xfffffffffffffff0,%rsi --- src_reg -= off
Note that the actual load is always done with offset 0, since previous
insns have already done src_reg += off. Regardless of whether the new
check succeeds or fails, insn 61 is always executed, returning src_reg
to its original value.
Because the goal of these checks is to ensure that loaded-from address
will be protected by BPF exception handler, the new check can safely
ignore any wrapover from insn 4d. If such wrapped-over address passes
insn 54 + 57's cmp-and-jmp it will have such protection so the load can
proceed.
IMPROVEMENTS: The above improved logic is 8 insns vs original logic's 9,
and has 1 fewer jmp. The number of checking insns can be further
improved in common scenarios:
If src_reg == dst_reg, the actual load insn will clobber src_reg, so
there's no original src_reg state for the sub insn immediately following
the load to restore, so it can be omitted. In fact, it must be omitted
since it would incorrectly subtract from the result of the load if it
wasn't. So for src_reg == dst_reg, JIT emits these insns:
3c: movabs $0x800000000000,%r11
46: add $0xfffffffffffffff0,%rdi
4d: cmp %r11,%rdi
50: jae 0x0000000000000056
52: xor %edi,%edi
54: jmp 0x000000000000005a
56: mov 0x0(%rdi),%rdi
5a:
The only difference from larger example being the omitted sub, which
would've been insn 5a in this example.
If offset == 0, we can similarly omit the sub as in previous case, since
there's nothing added to subtract. For the same reason we can omit the
addition as well, resulting in JIT emitting these insns:
46: movabs $0x800000000000,%r11
4d: cmp %r11,%rdi
50: jae 0x0000000000000056
52: xor %edi,%edi
54: jmp 0x000000000000005a
56: mov 0x0(%rdi),%rdi
5a:
Although the above example also has src_reg == dst_reg, the same
offset == 0 optimization is valid to apply if src_reg != dst_reg.
To summarize the improvements in emitted insn count for the
check-and-load:
BEFORE: 8 check insns, 3 jmps
AFTER (general case): 7 check insns, 2 jmps (12.5% fewer insn, 33% jmp)
AFTER (src == dst): 6 check insns, 2 jmps (25% fewer insn)
AFTER (offset == 0): 5 check insns, 2 jmps (37.5% fewer insn)
(Above counts don't include the 1 load insn, just checking around it)
Based on BPF bytecode + JITted x86 insn I saw while experimenting with
these improvements, I expect the src_reg == dst_reg case to occur most
often, followed by offset == 0, then the general case.
Signed-off-by: Dave Marchevsky <davemarchevsky@fb.com>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Acked-by: Yonghong Song <yhs@fb.com>
Link: https://lore.kernel.org/bpf/20221216214319.3408356-1-davemarchevsky@fb.com
2022-12-16 21:43:18 +00:00
|
|
|
s16 insn_off;
|
net: filter: x86: internal BPF JIT
Maps all internal BPF instructions into x86_64 instructions.
This patch replaces original BPF x64 JIT with internal BPF x64 JIT.
sysctl net.core.bpf_jit_enable is reused as on/off switch.
Performance:
1. old BPF JIT and internal BPF JIT generate equivalent x86_64 code.
No performance difference is observed for filters that were JIT-able before
Example assembler code for BPF filter "tcpdump port 22"
original BPF -> old JIT: original BPF -> internal BPF -> new JIT:
0: push %rbp 0: push %rbp
1: mov %rsp,%rbp 1: mov %rsp,%rbp
4: sub $0x60,%rsp 4: sub $0x228,%rsp
8: mov %rbx,-0x8(%rbp) b: mov %rbx,-0x228(%rbp) // prologue
12: mov %r13,-0x220(%rbp)
19: mov %r14,-0x218(%rbp)
20: mov %r15,-0x210(%rbp)
27: xor %eax,%eax // clear A
c: xor %ebx,%ebx 29: xor %r13,%r13 // clear X
e: mov 0x68(%rdi),%r9d 2c: mov 0x68(%rdi),%r9d
12: sub 0x6c(%rdi),%r9d 30: sub 0x6c(%rdi),%r9d
16: mov 0xd8(%rdi),%r8 34: mov 0xd8(%rdi),%r10
3b: mov %rdi,%rbx
1d: mov $0xc,%esi 3e: mov $0xc,%esi
22: callq 0xffffffffe1021e15 43: callq 0xffffffffe102bd75
27: cmp $0x86dd,%eax 48: cmp $0x86dd,%rax
2c: jne 0x0000000000000069 4f: jne 0x000000000000009a
2e: mov $0x14,%esi 51: mov $0x14,%esi
33: callq 0xffffffffe1021e31 56: callq 0xffffffffe102bd91
38: cmp $0x84,%eax 5b: cmp $0x84,%rax
3d: je 0x0000000000000049 62: je 0x0000000000000074
3f: cmp $0x6,%eax 64: cmp $0x6,%rax
42: je 0x0000000000000049 68: je 0x0000000000000074
44: cmp $0x11,%eax 6a: cmp $0x11,%rax
47: jne 0x00000000000000c6 6e: jne 0x0000000000000117
49: mov $0x36,%esi 74: mov $0x36,%esi
4e: callq 0xffffffffe1021e15 79: callq 0xffffffffe102bd75
53: cmp $0x16,%eax 7e: cmp $0x16,%rax
56: je 0x00000000000000bf 82: je 0x0000000000000110
58: mov $0x38,%esi 88: mov $0x38,%esi
5d: callq 0xffffffffe1021e15 8d: callq 0xffffffffe102bd75
62: cmp $0x16,%eax 92: cmp $0x16,%rax
65: je 0x00000000000000bf 96: je 0x0000000000000110
67: jmp 0x00000000000000c6 98: jmp 0x0000000000000117
69: cmp $0x800,%eax 9a: cmp $0x800,%rax
6e: jne 0x00000000000000c6 a1: jne 0x0000000000000117
70: mov $0x17,%esi a3: mov $0x17,%esi
75: callq 0xffffffffe1021e31 a8: callq 0xffffffffe102bd91
7a: cmp $0x84,%eax ad: cmp $0x84,%rax
7f: je 0x000000000000008b b4: je 0x00000000000000c2
81: cmp $0x6,%eax b6: cmp $0x6,%rax
84: je 0x000000000000008b ba: je 0x00000000000000c2
86: cmp $0x11,%eax bc: cmp $0x11,%rax
89: jne 0x00000000000000c6 c0: jne 0x0000000000000117
8b: mov $0x14,%esi c2: mov $0x14,%esi
90: callq 0xffffffffe1021e15 c7: callq 0xffffffffe102bd75
95: test $0x1fff,%ax cc: test $0x1fff,%rax
99: jne 0x00000000000000c6 d3: jne 0x0000000000000117
d5: mov %rax,%r14
9b: mov $0xe,%esi d8: mov $0xe,%esi
a0: callq 0xffffffffe1021e44 dd: callq 0xffffffffe102bd91 // MSH
e2: and $0xf,%eax
e5: shl $0x2,%eax
e8: mov %rax,%r13
eb: mov %r14,%rax
ee: mov %r13,%rsi
a5: lea 0xe(%rbx),%esi f1: add $0xe,%esi
a8: callq 0xffffffffe1021e0d f4: callq 0xffffffffe102bd6d
ad: cmp $0x16,%eax f9: cmp $0x16,%rax
b0: je 0x00000000000000bf fd: je 0x0000000000000110
ff: mov %r13,%rsi
b2: lea 0x10(%rbx),%esi 102: add $0x10,%esi
b5: callq 0xffffffffe1021e0d 105: callq 0xffffffffe102bd6d
ba: cmp $0x16,%eax 10a: cmp $0x16,%rax
bd: jne 0x00000000000000c6 10e: jne 0x0000000000000117
bf: mov $0xffff,%eax 110: mov $0xffff,%eax
c4: jmp 0x00000000000000c8 115: jmp 0x000000000000011c
c6: xor %eax,%eax 117: mov $0x0,%eax
c8: mov -0x8(%rbp),%rbx 11c: mov -0x228(%rbp),%rbx // epilogue
cc: leaveq 123: mov -0x220(%rbp),%r13
cd: retq 12a: mov -0x218(%rbp),%r14
131: mov -0x210(%rbp),%r15
138: leaveq
139: retq
On fully cached SKBs both JITed functions take 12 nsec to execute.
BPF interpreter executes the program in 30 nsec.
The difference in generated assembler is due to the following:
Old BPF imlements LDX_MSH instruction via sk_load_byte_msh() helper function
inside bpf_jit.S.
New JIT removes the helper and does it explicitly, so ldx_msh cost
is the same for both JITs, but generated code looks longer.
New JIT has 4 registers to save, so prologue/epilogue are larger,
but the cost is within noise on x64.
Old JIT checks whether first insn clears A and if not emits 'xor %eax,%eax'.
New JIT clears %rax unconditionally.
2. old BPF JIT doesn't support ANC_NLATTR, ANC_PAY_OFFSET, ANC_RANDOM
extensions. New JIT supports all BPF extensions.
Performance of such filters improves 2-4 times depending on a filter.
The longer the filter the higher performance gain.
Synthetic benchmarks with many ancillary loads see 20x speedup
which seems to be the maximum gain from JIT
Notes:
. net.core.bpf_jit_enable=2 + tools/net/bpf_jit_disasm is still functional
and can be used to see generated assembler
. there are two jit_compile() functions and code flow for classic filters is:
sk_attach_filter() - load classic BPF
bpf_jit_compile() - try to JIT from classic BPF
sk_convert_filter() - convert classic to internal
bpf_int_jit_compile() - JIT from internal BPF
seccomp and tracing filters will just call bpf_int_jit_compile()
Signed-off-by: Alexei Starovoitov <ast@plumgrid.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
2014-05-14 02:50:46 +00:00
|
|
|
u8 jmp_cond;
|
|
|
|
u8 *func;
|
bpf,x64: Pad NOPs to make images converge more easily
The x64 bpf jit expects bpf images converge within the given passes, but
it could fail to do so with some corner cases. For example:
l0: ja 40
l1: ja 40
[... repeated ja 40 ]
l39: ja 40
l40: ret #0
This bpf program contains 40 "ja 40" instructions which are effectively
NOPs and designed to be replaced with valid code dynamically. Ideally,
bpf jit should optimize those "ja 40" instructions out when translating
the bpf instructions into x64 machine code. However, do_jit() can only
remove one "ja 40" for offset==0 on each pass, so it requires at least
40 runs to eliminate those JMPs and exceeds the current limit of
passes(20). In the end, the program got rejected when BPF_JIT_ALWAYS_ON
is set even though it's legit as a classic socket filter.
To make bpf images more likely converge within 20 passes, this commit
pads some instructions with NOPs in the last 5 passes:
1. conditional jumps
A possible size variance comes from the adoption of imm8 JMP. If the
offset is imm8, we calculate the size difference of this BPF instruction
between the previous and the current pass and fill the gap with NOPs.
To avoid the recalculation of jump offset, those NOPs are inserted before
the JMP code, so we have to subtract the 2 bytes of imm8 JMP when
calculating the NOP number.
2. BPF_JA
There are two conditions for BPF_JA.
a.) nop jumps
If this instruction is not optimized out in the previous pass,
instead of removing it, we insert the equivalent size of NOPs.
b.) label jumps
Similar to condition jumps, we prepend NOPs right before the JMP
code.
To make the code concise, emit_nops() is modified to use the signed len and
return the number of inserted NOPs.
For bpf-to-bpf, we always enable padding for the extra pass since there
is only one extra run and the jump padding doesn't affected the images
that converge without padding.
After applying this patch, the corner case was loaded with the following
jit code:
flen=45 proglen=77 pass=17 image=ffffffffc03367d4 from=jump pid=10097
JIT code: 00000000: 0f 1f 44 00 00 55 48 89 e5 53 41 55 31 c0 45 31
JIT code: 00000010: ed 48 89 fb eb 30 eb 2e eb 2c eb 2a eb 28 eb 26
JIT code: 00000020: eb 24 eb 22 eb 20 eb 1e eb 1c eb 1a eb 18 eb 16
JIT code: 00000030: eb 14 eb 12 eb 10 eb 0e eb 0c eb 0a eb 08 eb 06
JIT code: 00000040: eb 04 eb 02 66 90 31 c0 41 5d 5b c9 c3
0: 0f 1f 44 00 00 nop DWORD PTR [rax+rax*1+0x0]
5: 55 push rbp
6: 48 89 e5 mov rbp,rsp
9: 53 push rbx
a: 41 55 push r13
c: 31 c0 xor eax,eax
e: 45 31 ed xor r13d,r13d
11: 48 89 fb mov rbx,rdi
14: eb 30 jmp 0x46
16: eb 2e jmp 0x46
...
3e: eb 06 jmp 0x46
40: eb 04 jmp 0x46
42: eb 02 jmp 0x46
44: 66 90 xchg ax,ax
46: 31 c0 xor eax,eax
48: 41 5d pop r13
4a: 5b pop rbx
4b: c9 leave
4c: c3 ret
At the 16th pass, 15 jumps were already optimized out, and one jump was
replaced with NOPs at 44 and the image converged at the 17th pass.
v4:
- Add the detailed comments about the possible padding bytes
v3:
- Copy the instructions of prologue separately or the size calculation
of the first BPF instruction would include the prologue.
- Replace WARN_ONCE() with pr_err() and EFAULT
- Use MAX_PASSES in the for loop condition check
- Remove the "padded" flag from x64_jit_data. For the extra pass of
subprogs, padding is always enabled since it won't hurt the images
that converge without padding.
v2:
- Simplify the sample code in the description and provide the jit code
- Check the expected padding bytes with WARN_ONCE
- Move the 'padded' flag to 'struct x64_jit_data'
Signed-off-by: Gary Lin <glin@suse.com>
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Link: https://lore.kernel.org/bpf/20210119102501.511-2-glin@suse.com
2021-01-19 10:24:59 +00:00
|
|
|
int nops;
|
net: filter: x86: internal BPF JIT
Maps all internal BPF instructions into x86_64 instructions.
This patch replaces original BPF x64 JIT with internal BPF x64 JIT.
sysctl net.core.bpf_jit_enable is reused as on/off switch.
Performance:
1. old BPF JIT and internal BPF JIT generate equivalent x86_64 code.
No performance difference is observed for filters that were JIT-able before
Example assembler code for BPF filter "tcpdump port 22"
original BPF -> old JIT: original BPF -> internal BPF -> new JIT:
0: push %rbp 0: push %rbp
1: mov %rsp,%rbp 1: mov %rsp,%rbp
4: sub $0x60,%rsp 4: sub $0x228,%rsp
8: mov %rbx,-0x8(%rbp) b: mov %rbx,-0x228(%rbp) // prologue
12: mov %r13,-0x220(%rbp)
19: mov %r14,-0x218(%rbp)
20: mov %r15,-0x210(%rbp)
27: xor %eax,%eax // clear A
c: xor %ebx,%ebx 29: xor %r13,%r13 // clear X
e: mov 0x68(%rdi),%r9d 2c: mov 0x68(%rdi),%r9d
12: sub 0x6c(%rdi),%r9d 30: sub 0x6c(%rdi),%r9d
16: mov 0xd8(%rdi),%r8 34: mov 0xd8(%rdi),%r10
3b: mov %rdi,%rbx
1d: mov $0xc,%esi 3e: mov $0xc,%esi
22: callq 0xffffffffe1021e15 43: callq 0xffffffffe102bd75
27: cmp $0x86dd,%eax 48: cmp $0x86dd,%rax
2c: jne 0x0000000000000069 4f: jne 0x000000000000009a
2e: mov $0x14,%esi 51: mov $0x14,%esi
33: callq 0xffffffffe1021e31 56: callq 0xffffffffe102bd91
38: cmp $0x84,%eax 5b: cmp $0x84,%rax
3d: je 0x0000000000000049 62: je 0x0000000000000074
3f: cmp $0x6,%eax 64: cmp $0x6,%rax
42: je 0x0000000000000049 68: je 0x0000000000000074
44: cmp $0x11,%eax 6a: cmp $0x11,%rax
47: jne 0x00000000000000c6 6e: jne 0x0000000000000117
49: mov $0x36,%esi 74: mov $0x36,%esi
4e: callq 0xffffffffe1021e15 79: callq 0xffffffffe102bd75
53: cmp $0x16,%eax 7e: cmp $0x16,%rax
56: je 0x00000000000000bf 82: je 0x0000000000000110
58: mov $0x38,%esi 88: mov $0x38,%esi
5d: callq 0xffffffffe1021e15 8d: callq 0xffffffffe102bd75
62: cmp $0x16,%eax 92: cmp $0x16,%rax
65: je 0x00000000000000bf 96: je 0x0000000000000110
67: jmp 0x00000000000000c6 98: jmp 0x0000000000000117
69: cmp $0x800,%eax 9a: cmp $0x800,%rax
6e: jne 0x00000000000000c6 a1: jne 0x0000000000000117
70: mov $0x17,%esi a3: mov $0x17,%esi
75: callq 0xffffffffe1021e31 a8: callq 0xffffffffe102bd91
7a: cmp $0x84,%eax ad: cmp $0x84,%rax
7f: je 0x000000000000008b b4: je 0x00000000000000c2
81: cmp $0x6,%eax b6: cmp $0x6,%rax
84: je 0x000000000000008b ba: je 0x00000000000000c2
86: cmp $0x11,%eax bc: cmp $0x11,%rax
89: jne 0x00000000000000c6 c0: jne 0x0000000000000117
8b: mov $0x14,%esi c2: mov $0x14,%esi
90: callq 0xffffffffe1021e15 c7: callq 0xffffffffe102bd75
95: test $0x1fff,%ax cc: test $0x1fff,%rax
99: jne 0x00000000000000c6 d3: jne 0x0000000000000117
d5: mov %rax,%r14
9b: mov $0xe,%esi d8: mov $0xe,%esi
a0: callq 0xffffffffe1021e44 dd: callq 0xffffffffe102bd91 // MSH
e2: and $0xf,%eax
e5: shl $0x2,%eax
e8: mov %rax,%r13
eb: mov %r14,%rax
ee: mov %r13,%rsi
a5: lea 0xe(%rbx),%esi f1: add $0xe,%esi
a8: callq 0xffffffffe1021e0d f4: callq 0xffffffffe102bd6d
ad: cmp $0x16,%eax f9: cmp $0x16,%rax
b0: je 0x00000000000000bf fd: je 0x0000000000000110
ff: mov %r13,%rsi
b2: lea 0x10(%rbx),%esi 102: add $0x10,%esi
b5: callq 0xffffffffe1021e0d 105: callq 0xffffffffe102bd6d
ba: cmp $0x16,%eax 10a: cmp $0x16,%rax
bd: jne 0x00000000000000c6 10e: jne 0x0000000000000117
bf: mov $0xffff,%eax 110: mov $0xffff,%eax
c4: jmp 0x00000000000000c8 115: jmp 0x000000000000011c
c6: xor %eax,%eax 117: mov $0x0,%eax
c8: mov -0x8(%rbp),%rbx 11c: mov -0x228(%rbp),%rbx // epilogue
cc: leaveq 123: mov -0x220(%rbp),%r13
cd: retq 12a: mov -0x218(%rbp),%r14
131: mov -0x210(%rbp),%r15
138: leaveq
139: retq
On fully cached SKBs both JITed functions take 12 nsec to execute.
BPF interpreter executes the program in 30 nsec.
The difference in generated assembler is due to the following:
Old BPF imlements LDX_MSH instruction via sk_load_byte_msh() helper function
inside bpf_jit.S.
New JIT removes the helper and does it explicitly, so ldx_msh cost
is the same for both JITs, but generated code looks longer.
New JIT has 4 registers to save, so prologue/epilogue are larger,
but the cost is within noise on x64.
Old JIT checks whether first insn clears A and if not emits 'xor %eax,%eax'.
New JIT clears %rax unconditionally.
2. old BPF JIT doesn't support ANC_NLATTR, ANC_PAY_OFFSET, ANC_RANDOM
extensions. New JIT supports all BPF extensions.
Performance of such filters improves 2-4 times depending on a filter.
The longer the filter the higher performance gain.
Synthetic benchmarks with many ancillary loads see 20x speedup
which seems to be the maximum gain from JIT
Notes:
. net.core.bpf_jit_enable=2 + tools/net/bpf_jit_disasm is still functional
and can be used to see generated assembler
. there are two jit_compile() functions and code flow for classic filters is:
sk_attach_filter() - load classic BPF
bpf_jit_compile() - try to JIT from classic BPF
sk_convert_filter() - convert classic to internal
bpf_int_jit_compile() - JIT from internal BPF
seccomp and tracing filters will just call bpf_int_jit_compile()
Signed-off-by: Alexei Starovoitov <ast@plumgrid.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
2014-05-14 02:50:46 +00:00
|
|
|
|
|
|
|
switch (insn->code) {
|
|
|
|
/* ALU */
|
|
|
|
case BPF_ALU | BPF_ADD | BPF_X:
|
|
|
|
case BPF_ALU | BPF_SUB | BPF_X:
|
|
|
|
case BPF_ALU | BPF_AND | BPF_X:
|
|
|
|
case BPF_ALU | BPF_OR | BPF_X:
|
|
|
|
case BPF_ALU | BPF_XOR | BPF_X:
|
|
|
|
case BPF_ALU64 | BPF_ADD | BPF_X:
|
|
|
|
case BPF_ALU64 | BPF_SUB | BPF_X:
|
|
|
|
case BPF_ALU64 | BPF_AND | BPF_X:
|
|
|
|
case BPF_ALU64 | BPF_OR | BPF_X:
|
|
|
|
case BPF_ALU64 | BPF_XOR | BPF_X:
|
2021-01-14 18:17:42 +00:00
|
|
|
maybe_emit_mod(&prog, dst_reg, src_reg,
|
|
|
|
BPF_CLASS(insn->code) == BPF_ALU64);
|
2021-01-14 18:17:43 +00:00
|
|
|
b2 = simple_alu_opcodes[BPF_OP(insn->code)];
|
2014-06-06 21:46:06 +00:00
|
|
|
EMIT2(b2, add_2reg(0xC0, dst_reg, src_reg));
|
net: filter: x86: internal BPF JIT
Maps all internal BPF instructions into x86_64 instructions.
This patch replaces original BPF x64 JIT with internal BPF x64 JIT.
sysctl net.core.bpf_jit_enable is reused as on/off switch.
Performance:
1. old BPF JIT and internal BPF JIT generate equivalent x86_64 code.
No performance difference is observed for filters that were JIT-able before
Example assembler code for BPF filter "tcpdump port 22"
original BPF -> old JIT: original BPF -> internal BPF -> new JIT:
0: push %rbp 0: push %rbp
1: mov %rsp,%rbp 1: mov %rsp,%rbp
4: sub $0x60,%rsp 4: sub $0x228,%rsp
8: mov %rbx,-0x8(%rbp) b: mov %rbx,-0x228(%rbp) // prologue
12: mov %r13,-0x220(%rbp)
19: mov %r14,-0x218(%rbp)
20: mov %r15,-0x210(%rbp)
27: xor %eax,%eax // clear A
c: xor %ebx,%ebx 29: xor %r13,%r13 // clear X
e: mov 0x68(%rdi),%r9d 2c: mov 0x68(%rdi),%r9d
12: sub 0x6c(%rdi),%r9d 30: sub 0x6c(%rdi),%r9d
16: mov 0xd8(%rdi),%r8 34: mov 0xd8(%rdi),%r10
3b: mov %rdi,%rbx
1d: mov $0xc,%esi 3e: mov $0xc,%esi
22: callq 0xffffffffe1021e15 43: callq 0xffffffffe102bd75
27: cmp $0x86dd,%eax 48: cmp $0x86dd,%rax
2c: jne 0x0000000000000069 4f: jne 0x000000000000009a
2e: mov $0x14,%esi 51: mov $0x14,%esi
33: callq 0xffffffffe1021e31 56: callq 0xffffffffe102bd91
38: cmp $0x84,%eax 5b: cmp $0x84,%rax
3d: je 0x0000000000000049 62: je 0x0000000000000074
3f: cmp $0x6,%eax 64: cmp $0x6,%rax
42: je 0x0000000000000049 68: je 0x0000000000000074
44: cmp $0x11,%eax 6a: cmp $0x11,%rax
47: jne 0x00000000000000c6 6e: jne 0x0000000000000117
49: mov $0x36,%esi 74: mov $0x36,%esi
4e: callq 0xffffffffe1021e15 79: callq 0xffffffffe102bd75
53: cmp $0x16,%eax 7e: cmp $0x16,%rax
56: je 0x00000000000000bf 82: je 0x0000000000000110
58: mov $0x38,%esi 88: mov $0x38,%esi
5d: callq 0xffffffffe1021e15 8d: callq 0xffffffffe102bd75
62: cmp $0x16,%eax 92: cmp $0x16,%rax
65: je 0x00000000000000bf 96: je 0x0000000000000110
67: jmp 0x00000000000000c6 98: jmp 0x0000000000000117
69: cmp $0x800,%eax 9a: cmp $0x800,%rax
6e: jne 0x00000000000000c6 a1: jne 0x0000000000000117
70: mov $0x17,%esi a3: mov $0x17,%esi
75: callq 0xffffffffe1021e31 a8: callq 0xffffffffe102bd91
7a: cmp $0x84,%eax ad: cmp $0x84,%rax
7f: je 0x000000000000008b b4: je 0x00000000000000c2
81: cmp $0x6,%eax b6: cmp $0x6,%rax
84: je 0x000000000000008b ba: je 0x00000000000000c2
86: cmp $0x11,%eax bc: cmp $0x11,%rax
89: jne 0x00000000000000c6 c0: jne 0x0000000000000117
8b: mov $0x14,%esi c2: mov $0x14,%esi
90: callq 0xffffffffe1021e15 c7: callq 0xffffffffe102bd75
95: test $0x1fff,%ax cc: test $0x1fff,%rax
99: jne 0x00000000000000c6 d3: jne 0x0000000000000117
d5: mov %rax,%r14
9b: mov $0xe,%esi d8: mov $0xe,%esi
a0: callq 0xffffffffe1021e44 dd: callq 0xffffffffe102bd91 // MSH
e2: and $0xf,%eax
e5: shl $0x2,%eax
e8: mov %rax,%r13
eb: mov %r14,%rax
ee: mov %r13,%rsi
a5: lea 0xe(%rbx),%esi f1: add $0xe,%esi
a8: callq 0xffffffffe1021e0d f4: callq 0xffffffffe102bd6d
ad: cmp $0x16,%eax f9: cmp $0x16,%rax
b0: je 0x00000000000000bf fd: je 0x0000000000000110
ff: mov %r13,%rsi
b2: lea 0x10(%rbx),%esi 102: add $0x10,%esi
b5: callq 0xffffffffe1021e0d 105: callq 0xffffffffe102bd6d
ba: cmp $0x16,%eax 10a: cmp $0x16,%rax
bd: jne 0x00000000000000c6 10e: jne 0x0000000000000117
bf: mov $0xffff,%eax 110: mov $0xffff,%eax
c4: jmp 0x00000000000000c8 115: jmp 0x000000000000011c
c6: xor %eax,%eax 117: mov $0x0,%eax
c8: mov -0x8(%rbp),%rbx 11c: mov -0x228(%rbp),%rbx // epilogue
cc: leaveq 123: mov -0x220(%rbp),%r13
cd: retq 12a: mov -0x218(%rbp),%r14
131: mov -0x210(%rbp),%r15
138: leaveq
139: retq
On fully cached SKBs both JITed functions take 12 nsec to execute.
BPF interpreter executes the program in 30 nsec.
The difference in generated assembler is due to the following:
Old BPF imlements LDX_MSH instruction via sk_load_byte_msh() helper function
inside bpf_jit.S.
New JIT removes the helper and does it explicitly, so ldx_msh cost
is the same for both JITs, but generated code looks longer.
New JIT has 4 registers to save, so prologue/epilogue are larger,
but the cost is within noise on x64.
Old JIT checks whether first insn clears A and if not emits 'xor %eax,%eax'.
New JIT clears %rax unconditionally.
2. old BPF JIT doesn't support ANC_NLATTR, ANC_PAY_OFFSET, ANC_RANDOM
extensions. New JIT supports all BPF extensions.
Performance of such filters improves 2-4 times depending on a filter.
The longer the filter the higher performance gain.
Synthetic benchmarks with many ancillary loads see 20x speedup
which seems to be the maximum gain from JIT
Notes:
. net.core.bpf_jit_enable=2 + tools/net/bpf_jit_disasm is still functional
and can be used to see generated assembler
. there are two jit_compile() functions and code flow for classic filters is:
sk_attach_filter() - load classic BPF
bpf_jit_compile() - try to JIT from classic BPF
sk_convert_filter() - convert classic to internal
bpf_int_jit_compile() - JIT from internal BPF
seccomp and tracing filters will just call bpf_int_jit_compile()
Signed-off-by: Alexei Starovoitov <ast@plumgrid.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
2014-05-14 02:50:46 +00:00
|
|
|
break;
|
2011-04-20 09:27:32 +00:00
|
|
|
|
net: filter: x86: internal BPF JIT
Maps all internal BPF instructions into x86_64 instructions.
This patch replaces original BPF x64 JIT with internal BPF x64 JIT.
sysctl net.core.bpf_jit_enable is reused as on/off switch.
Performance:
1. old BPF JIT and internal BPF JIT generate equivalent x86_64 code.
No performance difference is observed for filters that were JIT-able before
Example assembler code for BPF filter "tcpdump port 22"
original BPF -> old JIT: original BPF -> internal BPF -> new JIT:
0: push %rbp 0: push %rbp
1: mov %rsp,%rbp 1: mov %rsp,%rbp
4: sub $0x60,%rsp 4: sub $0x228,%rsp
8: mov %rbx,-0x8(%rbp) b: mov %rbx,-0x228(%rbp) // prologue
12: mov %r13,-0x220(%rbp)
19: mov %r14,-0x218(%rbp)
20: mov %r15,-0x210(%rbp)
27: xor %eax,%eax // clear A
c: xor %ebx,%ebx 29: xor %r13,%r13 // clear X
e: mov 0x68(%rdi),%r9d 2c: mov 0x68(%rdi),%r9d
12: sub 0x6c(%rdi),%r9d 30: sub 0x6c(%rdi),%r9d
16: mov 0xd8(%rdi),%r8 34: mov 0xd8(%rdi),%r10
3b: mov %rdi,%rbx
1d: mov $0xc,%esi 3e: mov $0xc,%esi
22: callq 0xffffffffe1021e15 43: callq 0xffffffffe102bd75
27: cmp $0x86dd,%eax 48: cmp $0x86dd,%rax
2c: jne 0x0000000000000069 4f: jne 0x000000000000009a
2e: mov $0x14,%esi 51: mov $0x14,%esi
33: callq 0xffffffffe1021e31 56: callq 0xffffffffe102bd91
38: cmp $0x84,%eax 5b: cmp $0x84,%rax
3d: je 0x0000000000000049 62: je 0x0000000000000074
3f: cmp $0x6,%eax 64: cmp $0x6,%rax
42: je 0x0000000000000049 68: je 0x0000000000000074
44: cmp $0x11,%eax 6a: cmp $0x11,%rax
47: jne 0x00000000000000c6 6e: jne 0x0000000000000117
49: mov $0x36,%esi 74: mov $0x36,%esi
4e: callq 0xffffffffe1021e15 79: callq 0xffffffffe102bd75
53: cmp $0x16,%eax 7e: cmp $0x16,%rax
56: je 0x00000000000000bf 82: je 0x0000000000000110
58: mov $0x38,%esi 88: mov $0x38,%esi
5d: callq 0xffffffffe1021e15 8d: callq 0xffffffffe102bd75
62: cmp $0x16,%eax 92: cmp $0x16,%rax
65: je 0x00000000000000bf 96: je 0x0000000000000110
67: jmp 0x00000000000000c6 98: jmp 0x0000000000000117
69: cmp $0x800,%eax 9a: cmp $0x800,%rax
6e: jne 0x00000000000000c6 a1: jne 0x0000000000000117
70: mov $0x17,%esi a3: mov $0x17,%esi
75: callq 0xffffffffe1021e31 a8: callq 0xffffffffe102bd91
7a: cmp $0x84,%eax ad: cmp $0x84,%rax
7f: je 0x000000000000008b b4: je 0x00000000000000c2
81: cmp $0x6,%eax b6: cmp $0x6,%rax
84: je 0x000000000000008b ba: je 0x00000000000000c2
86: cmp $0x11,%eax bc: cmp $0x11,%rax
89: jne 0x00000000000000c6 c0: jne 0x0000000000000117
8b: mov $0x14,%esi c2: mov $0x14,%esi
90: callq 0xffffffffe1021e15 c7: callq 0xffffffffe102bd75
95: test $0x1fff,%ax cc: test $0x1fff,%rax
99: jne 0x00000000000000c6 d3: jne 0x0000000000000117
d5: mov %rax,%r14
9b: mov $0xe,%esi d8: mov $0xe,%esi
a0: callq 0xffffffffe1021e44 dd: callq 0xffffffffe102bd91 // MSH
e2: and $0xf,%eax
e5: shl $0x2,%eax
e8: mov %rax,%r13
eb: mov %r14,%rax
ee: mov %r13,%rsi
a5: lea 0xe(%rbx),%esi f1: add $0xe,%esi
a8: callq 0xffffffffe1021e0d f4: callq 0xffffffffe102bd6d
ad: cmp $0x16,%eax f9: cmp $0x16,%rax
b0: je 0x00000000000000bf fd: je 0x0000000000000110
ff: mov %r13,%rsi
b2: lea 0x10(%rbx),%esi 102: add $0x10,%esi
b5: callq 0xffffffffe1021e0d 105: callq 0xffffffffe102bd6d
ba: cmp $0x16,%eax 10a: cmp $0x16,%rax
bd: jne 0x00000000000000c6 10e: jne 0x0000000000000117
bf: mov $0xffff,%eax 110: mov $0xffff,%eax
c4: jmp 0x00000000000000c8 115: jmp 0x000000000000011c
c6: xor %eax,%eax 117: mov $0x0,%eax
c8: mov -0x8(%rbp),%rbx 11c: mov -0x228(%rbp),%rbx // epilogue
cc: leaveq 123: mov -0x220(%rbp),%r13
cd: retq 12a: mov -0x218(%rbp),%r14
131: mov -0x210(%rbp),%r15
138: leaveq
139: retq
On fully cached SKBs both JITed functions take 12 nsec to execute.
BPF interpreter executes the program in 30 nsec.
The difference in generated assembler is due to the following:
Old BPF imlements LDX_MSH instruction via sk_load_byte_msh() helper function
inside bpf_jit.S.
New JIT removes the helper and does it explicitly, so ldx_msh cost
is the same for both JITs, but generated code looks longer.
New JIT has 4 registers to save, so prologue/epilogue are larger,
but the cost is within noise on x64.
Old JIT checks whether first insn clears A and if not emits 'xor %eax,%eax'.
New JIT clears %rax unconditionally.
2. old BPF JIT doesn't support ANC_NLATTR, ANC_PAY_OFFSET, ANC_RANDOM
extensions. New JIT supports all BPF extensions.
Performance of such filters improves 2-4 times depending on a filter.
The longer the filter the higher performance gain.
Synthetic benchmarks with many ancillary loads see 20x speedup
which seems to be the maximum gain from JIT
Notes:
. net.core.bpf_jit_enable=2 + tools/net/bpf_jit_disasm is still functional
and can be used to see generated assembler
. there are two jit_compile() functions and code flow for classic filters is:
sk_attach_filter() - load classic BPF
bpf_jit_compile() - try to JIT from classic BPF
sk_convert_filter() - convert classic to internal
bpf_int_jit_compile() - JIT from internal BPF
seccomp and tracing filters will just call bpf_int_jit_compile()
Signed-off-by: Alexei Starovoitov <ast@plumgrid.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
2014-05-14 02:50:46 +00:00
|
|
|
case BPF_ALU64 | BPF_MOV | BPF_X:
|
bpf: Add x86-64 JIT support for bpf_addr_space_cast instruction.
LLVM generates bpf_addr_space_cast instruction while translating
pointers between native (zero) address space and
__attribute__((address_space(N))).
The addr_space=1 is reserved as bpf_arena address space.
rY = addr_space_cast(rX, 0, 1) is processed by the verifier and
converted to normal 32-bit move: wX = wY
rY = addr_space_cast(rX, 1, 0) has to be converted by JIT:
aux_reg = upper_32_bits of arena->user_vm_start
aux_reg <<= 32
wX = wY // clear upper 32 bits of dst register
if (wX) // if not zero add upper bits of user_vm_start
wX |= aux_reg
JIT can do it more efficiently:
mov dst_reg32, src_reg32 // 32-bit move
shl dst_reg, 32
or dst_reg, user_vm_start
rol dst_reg, 32
xor r11, r11
test dst_reg32, dst_reg32 // check if lower 32-bit are zero
cmove r11, dst_reg // if so, set dst_reg to zero
// Intel swapped src/dst register encoding in CMOVcc
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Signed-off-by: Andrii Nakryiko <andrii@kernel.org>
Acked-by: Eduard Zingerman <eddyz87@gmail.com>
Link: https://lore.kernel.org/bpf/20240308010812.89848-5-alexei.starovoitov@gmail.com
2024-03-08 01:08:02 +00:00
|
|
|
if (insn->off == BPF_ADDR_SPACE_CAST &&
|
|
|
|
insn->imm == 1U << 16) {
|
|
|
|
if (dst_reg != src_reg)
|
|
|
|
/* 32-bit mov */
|
|
|
|
emit_mov_reg(&prog, false, dst_reg, src_reg);
|
|
|
|
/* shl dst_reg, 32 */
|
|
|
|
maybe_emit_1mod(&prog, dst_reg, true);
|
|
|
|
EMIT3(0xC1, add_1reg(0xE0, dst_reg), 32);
|
|
|
|
|
|
|
|
/* or dst_reg, user_vm_start */
|
|
|
|
maybe_emit_1mod(&prog, dst_reg, true);
|
|
|
|
if (is_axreg(dst_reg))
|
|
|
|
EMIT1_off32(0x0D, user_vm_start >> 32);
|
|
|
|
else
|
|
|
|
EMIT2_off32(0x81, add_1reg(0xC8, dst_reg), user_vm_start >> 32);
|
|
|
|
|
|
|
|
/* rol dst_reg, 32 */
|
|
|
|
maybe_emit_1mod(&prog, dst_reg, true);
|
|
|
|
EMIT3(0xC1, add_1reg(0xC0, dst_reg), 32);
|
|
|
|
|
|
|
|
/* xor r11, r11 */
|
|
|
|
EMIT3(0x4D, 0x31, 0xDB);
|
|
|
|
|
|
|
|
/* test dst_reg32, dst_reg32; check if lower 32-bit are zero */
|
|
|
|
maybe_emit_mod(&prog, dst_reg, dst_reg, false);
|
|
|
|
EMIT2(0x85, add_2reg(0xC0, dst_reg, dst_reg));
|
|
|
|
|
|
|
|
/* cmove r11, dst_reg; if so, set dst_reg to zero */
|
|
|
|
/* WARNING: Intel swapped src/dst register encoding in CMOVcc !!! */
|
|
|
|
maybe_emit_mod(&prog, AUX_REG, dst_reg, true);
|
|
|
|
EMIT3(0x0F, 0x44, add_2reg(0xC0, AUX_REG, dst_reg));
|
|
|
|
break;
|
|
|
|
}
|
|
|
|
fallthrough;
|
net: filter: x86: internal BPF JIT
Maps all internal BPF instructions into x86_64 instructions.
This patch replaces original BPF x64 JIT with internal BPF x64 JIT.
sysctl net.core.bpf_jit_enable is reused as on/off switch.
Performance:
1. old BPF JIT and internal BPF JIT generate equivalent x86_64 code.
No performance difference is observed for filters that were JIT-able before
Example assembler code for BPF filter "tcpdump port 22"
original BPF -> old JIT: original BPF -> internal BPF -> new JIT:
0: push %rbp 0: push %rbp
1: mov %rsp,%rbp 1: mov %rsp,%rbp
4: sub $0x60,%rsp 4: sub $0x228,%rsp
8: mov %rbx,-0x8(%rbp) b: mov %rbx,-0x228(%rbp) // prologue
12: mov %r13,-0x220(%rbp)
19: mov %r14,-0x218(%rbp)
20: mov %r15,-0x210(%rbp)
27: xor %eax,%eax // clear A
c: xor %ebx,%ebx 29: xor %r13,%r13 // clear X
e: mov 0x68(%rdi),%r9d 2c: mov 0x68(%rdi),%r9d
12: sub 0x6c(%rdi),%r9d 30: sub 0x6c(%rdi),%r9d
16: mov 0xd8(%rdi),%r8 34: mov 0xd8(%rdi),%r10
3b: mov %rdi,%rbx
1d: mov $0xc,%esi 3e: mov $0xc,%esi
22: callq 0xffffffffe1021e15 43: callq 0xffffffffe102bd75
27: cmp $0x86dd,%eax 48: cmp $0x86dd,%rax
2c: jne 0x0000000000000069 4f: jne 0x000000000000009a
2e: mov $0x14,%esi 51: mov $0x14,%esi
33: callq 0xffffffffe1021e31 56: callq 0xffffffffe102bd91
38: cmp $0x84,%eax 5b: cmp $0x84,%rax
3d: je 0x0000000000000049 62: je 0x0000000000000074
3f: cmp $0x6,%eax 64: cmp $0x6,%rax
42: je 0x0000000000000049 68: je 0x0000000000000074
44: cmp $0x11,%eax 6a: cmp $0x11,%rax
47: jne 0x00000000000000c6 6e: jne 0x0000000000000117
49: mov $0x36,%esi 74: mov $0x36,%esi
4e: callq 0xffffffffe1021e15 79: callq 0xffffffffe102bd75
53: cmp $0x16,%eax 7e: cmp $0x16,%rax
56: je 0x00000000000000bf 82: je 0x0000000000000110
58: mov $0x38,%esi 88: mov $0x38,%esi
5d: callq 0xffffffffe1021e15 8d: callq 0xffffffffe102bd75
62: cmp $0x16,%eax 92: cmp $0x16,%rax
65: je 0x00000000000000bf 96: je 0x0000000000000110
67: jmp 0x00000000000000c6 98: jmp 0x0000000000000117
69: cmp $0x800,%eax 9a: cmp $0x800,%rax
6e: jne 0x00000000000000c6 a1: jne 0x0000000000000117
70: mov $0x17,%esi a3: mov $0x17,%esi
75: callq 0xffffffffe1021e31 a8: callq 0xffffffffe102bd91
7a: cmp $0x84,%eax ad: cmp $0x84,%rax
7f: je 0x000000000000008b b4: je 0x00000000000000c2
81: cmp $0x6,%eax b6: cmp $0x6,%rax
84: je 0x000000000000008b ba: je 0x00000000000000c2
86: cmp $0x11,%eax bc: cmp $0x11,%rax
89: jne 0x00000000000000c6 c0: jne 0x0000000000000117
8b: mov $0x14,%esi c2: mov $0x14,%esi
90: callq 0xffffffffe1021e15 c7: callq 0xffffffffe102bd75
95: test $0x1fff,%ax cc: test $0x1fff,%rax
99: jne 0x00000000000000c6 d3: jne 0x0000000000000117
d5: mov %rax,%r14
9b: mov $0xe,%esi d8: mov $0xe,%esi
a0: callq 0xffffffffe1021e44 dd: callq 0xffffffffe102bd91 // MSH
e2: and $0xf,%eax
e5: shl $0x2,%eax
e8: mov %rax,%r13
eb: mov %r14,%rax
ee: mov %r13,%rsi
a5: lea 0xe(%rbx),%esi f1: add $0xe,%esi
a8: callq 0xffffffffe1021e0d f4: callq 0xffffffffe102bd6d
ad: cmp $0x16,%eax f9: cmp $0x16,%rax
b0: je 0x00000000000000bf fd: je 0x0000000000000110
ff: mov %r13,%rsi
b2: lea 0x10(%rbx),%esi 102: add $0x10,%esi
b5: callq 0xffffffffe1021e0d 105: callq 0xffffffffe102bd6d
ba: cmp $0x16,%eax 10a: cmp $0x16,%rax
bd: jne 0x00000000000000c6 10e: jne 0x0000000000000117
bf: mov $0xffff,%eax 110: mov $0xffff,%eax
c4: jmp 0x00000000000000c8 115: jmp 0x000000000000011c
c6: xor %eax,%eax 117: mov $0x0,%eax
c8: mov -0x8(%rbp),%rbx 11c: mov -0x228(%rbp),%rbx // epilogue
cc: leaveq 123: mov -0x220(%rbp),%r13
cd: retq 12a: mov -0x218(%rbp),%r14
131: mov -0x210(%rbp),%r15
138: leaveq
139: retq
On fully cached SKBs both JITed functions take 12 nsec to execute.
BPF interpreter executes the program in 30 nsec.
The difference in generated assembler is due to the following:
Old BPF imlements LDX_MSH instruction via sk_load_byte_msh() helper function
inside bpf_jit.S.
New JIT removes the helper and does it explicitly, so ldx_msh cost
is the same for both JITs, but generated code looks longer.
New JIT has 4 registers to save, so prologue/epilogue are larger,
but the cost is within noise on x64.
Old JIT checks whether first insn clears A and if not emits 'xor %eax,%eax'.
New JIT clears %rax unconditionally.
2. old BPF JIT doesn't support ANC_NLATTR, ANC_PAY_OFFSET, ANC_RANDOM
extensions. New JIT supports all BPF extensions.
Performance of such filters improves 2-4 times depending on a filter.
The longer the filter the higher performance gain.
Synthetic benchmarks with many ancillary loads see 20x speedup
which seems to be the maximum gain from JIT
Notes:
. net.core.bpf_jit_enable=2 + tools/net/bpf_jit_disasm is still functional
and can be used to see generated assembler
. there are two jit_compile() functions and code flow for classic filters is:
sk_attach_filter() - load classic BPF
bpf_jit_compile() - try to JIT from classic BPF
sk_convert_filter() - convert classic to internal
bpf_int_jit_compile() - JIT from internal BPF
seccomp and tracing filters will just call bpf_int_jit_compile()
Signed-off-by: Alexei Starovoitov <ast@plumgrid.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
2014-05-14 02:50:46 +00:00
|
|
|
case BPF_ALU | BPF_MOV | BPF_X:
|
2023-07-28 01:12:02 +00:00
|
|
|
if (insn->off == 0)
|
|
|
|
emit_mov_reg(&prog,
|
|
|
|
BPF_CLASS(insn->code) == BPF_ALU64,
|
|
|
|
dst_reg, src_reg);
|
|
|
|
else
|
|
|
|
emit_movsx_reg(&prog, insn->off,
|
|
|
|
BPF_CLASS(insn->code) == BPF_ALU64,
|
|
|
|
dst_reg, src_reg);
|
net: filter: x86: internal BPF JIT
Maps all internal BPF instructions into x86_64 instructions.
This patch replaces original BPF x64 JIT with internal BPF x64 JIT.
sysctl net.core.bpf_jit_enable is reused as on/off switch.
Performance:
1. old BPF JIT and internal BPF JIT generate equivalent x86_64 code.
No performance difference is observed for filters that were JIT-able before
Example assembler code for BPF filter "tcpdump port 22"
original BPF -> old JIT: original BPF -> internal BPF -> new JIT:
0: push %rbp 0: push %rbp
1: mov %rsp,%rbp 1: mov %rsp,%rbp
4: sub $0x60,%rsp 4: sub $0x228,%rsp
8: mov %rbx,-0x8(%rbp) b: mov %rbx,-0x228(%rbp) // prologue
12: mov %r13,-0x220(%rbp)
19: mov %r14,-0x218(%rbp)
20: mov %r15,-0x210(%rbp)
27: xor %eax,%eax // clear A
c: xor %ebx,%ebx 29: xor %r13,%r13 // clear X
e: mov 0x68(%rdi),%r9d 2c: mov 0x68(%rdi),%r9d
12: sub 0x6c(%rdi),%r9d 30: sub 0x6c(%rdi),%r9d
16: mov 0xd8(%rdi),%r8 34: mov 0xd8(%rdi),%r10
3b: mov %rdi,%rbx
1d: mov $0xc,%esi 3e: mov $0xc,%esi
22: callq 0xffffffffe1021e15 43: callq 0xffffffffe102bd75
27: cmp $0x86dd,%eax 48: cmp $0x86dd,%rax
2c: jne 0x0000000000000069 4f: jne 0x000000000000009a
2e: mov $0x14,%esi 51: mov $0x14,%esi
33: callq 0xffffffffe1021e31 56: callq 0xffffffffe102bd91
38: cmp $0x84,%eax 5b: cmp $0x84,%rax
3d: je 0x0000000000000049 62: je 0x0000000000000074
3f: cmp $0x6,%eax 64: cmp $0x6,%rax
42: je 0x0000000000000049 68: je 0x0000000000000074
44: cmp $0x11,%eax 6a: cmp $0x11,%rax
47: jne 0x00000000000000c6 6e: jne 0x0000000000000117
49: mov $0x36,%esi 74: mov $0x36,%esi
4e: callq 0xffffffffe1021e15 79: callq 0xffffffffe102bd75
53: cmp $0x16,%eax 7e: cmp $0x16,%rax
56: je 0x00000000000000bf 82: je 0x0000000000000110
58: mov $0x38,%esi 88: mov $0x38,%esi
5d: callq 0xffffffffe1021e15 8d: callq 0xffffffffe102bd75
62: cmp $0x16,%eax 92: cmp $0x16,%rax
65: je 0x00000000000000bf 96: je 0x0000000000000110
67: jmp 0x00000000000000c6 98: jmp 0x0000000000000117
69: cmp $0x800,%eax 9a: cmp $0x800,%rax
6e: jne 0x00000000000000c6 a1: jne 0x0000000000000117
70: mov $0x17,%esi a3: mov $0x17,%esi
75: callq 0xffffffffe1021e31 a8: callq 0xffffffffe102bd91
7a: cmp $0x84,%eax ad: cmp $0x84,%rax
7f: je 0x000000000000008b b4: je 0x00000000000000c2
81: cmp $0x6,%eax b6: cmp $0x6,%rax
84: je 0x000000000000008b ba: je 0x00000000000000c2
86: cmp $0x11,%eax bc: cmp $0x11,%rax
89: jne 0x00000000000000c6 c0: jne 0x0000000000000117
8b: mov $0x14,%esi c2: mov $0x14,%esi
90: callq 0xffffffffe1021e15 c7: callq 0xffffffffe102bd75
95: test $0x1fff,%ax cc: test $0x1fff,%rax
99: jne 0x00000000000000c6 d3: jne 0x0000000000000117
d5: mov %rax,%r14
9b: mov $0xe,%esi d8: mov $0xe,%esi
a0: callq 0xffffffffe1021e44 dd: callq 0xffffffffe102bd91 // MSH
e2: and $0xf,%eax
e5: shl $0x2,%eax
e8: mov %rax,%r13
eb: mov %r14,%rax
ee: mov %r13,%rsi
a5: lea 0xe(%rbx),%esi f1: add $0xe,%esi
a8: callq 0xffffffffe1021e0d f4: callq 0xffffffffe102bd6d
ad: cmp $0x16,%eax f9: cmp $0x16,%rax
b0: je 0x00000000000000bf fd: je 0x0000000000000110
ff: mov %r13,%rsi
b2: lea 0x10(%rbx),%esi 102: add $0x10,%esi
b5: callq 0xffffffffe1021e0d 105: callq 0xffffffffe102bd6d
ba: cmp $0x16,%eax 10a: cmp $0x16,%rax
bd: jne 0x00000000000000c6 10e: jne 0x0000000000000117
bf: mov $0xffff,%eax 110: mov $0xffff,%eax
c4: jmp 0x00000000000000c8 115: jmp 0x000000000000011c
c6: xor %eax,%eax 117: mov $0x0,%eax
c8: mov -0x8(%rbp),%rbx 11c: mov -0x228(%rbp),%rbx // epilogue
cc: leaveq 123: mov -0x220(%rbp),%r13
cd: retq 12a: mov -0x218(%rbp),%r14
131: mov -0x210(%rbp),%r15
138: leaveq
139: retq
On fully cached SKBs both JITed functions take 12 nsec to execute.
BPF interpreter executes the program in 30 nsec.
The difference in generated assembler is due to the following:
Old BPF imlements LDX_MSH instruction via sk_load_byte_msh() helper function
inside bpf_jit.S.
New JIT removes the helper and does it explicitly, so ldx_msh cost
is the same for both JITs, but generated code looks longer.
New JIT has 4 registers to save, so prologue/epilogue are larger,
but the cost is within noise on x64.
Old JIT checks whether first insn clears A and if not emits 'xor %eax,%eax'.
New JIT clears %rax unconditionally.
2. old BPF JIT doesn't support ANC_NLATTR, ANC_PAY_OFFSET, ANC_RANDOM
extensions. New JIT supports all BPF extensions.
Performance of such filters improves 2-4 times depending on a filter.
The longer the filter the higher performance gain.
Synthetic benchmarks with many ancillary loads see 20x speedup
which seems to be the maximum gain from JIT
Notes:
. net.core.bpf_jit_enable=2 + tools/net/bpf_jit_disasm is still functional
and can be used to see generated assembler
. there are two jit_compile() functions and code flow for classic filters is:
sk_attach_filter() - load classic BPF
bpf_jit_compile() - try to JIT from classic BPF
sk_convert_filter() - convert classic to internal
bpf_int_jit_compile() - JIT from internal BPF
seccomp and tracing filters will just call bpf_int_jit_compile()
Signed-off-by: Alexei Starovoitov <ast@plumgrid.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
2014-05-14 02:50:46 +00:00
|
|
|
break;
|
2011-04-20 09:27:32 +00:00
|
|
|
|
2014-06-06 21:46:06 +00:00
|
|
|
/* neg dst */
|
net: filter: x86: internal BPF JIT
Maps all internal BPF instructions into x86_64 instructions.
This patch replaces original BPF x64 JIT with internal BPF x64 JIT.
sysctl net.core.bpf_jit_enable is reused as on/off switch.
Performance:
1. old BPF JIT and internal BPF JIT generate equivalent x86_64 code.
No performance difference is observed for filters that were JIT-able before
Example assembler code for BPF filter "tcpdump port 22"
original BPF -> old JIT: original BPF -> internal BPF -> new JIT:
0: push %rbp 0: push %rbp
1: mov %rsp,%rbp 1: mov %rsp,%rbp
4: sub $0x60,%rsp 4: sub $0x228,%rsp
8: mov %rbx,-0x8(%rbp) b: mov %rbx,-0x228(%rbp) // prologue
12: mov %r13,-0x220(%rbp)
19: mov %r14,-0x218(%rbp)
20: mov %r15,-0x210(%rbp)
27: xor %eax,%eax // clear A
c: xor %ebx,%ebx 29: xor %r13,%r13 // clear X
e: mov 0x68(%rdi),%r9d 2c: mov 0x68(%rdi),%r9d
12: sub 0x6c(%rdi),%r9d 30: sub 0x6c(%rdi),%r9d
16: mov 0xd8(%rdi),%r8 34: mov 0xd8(%rdi),%r10
3b: mov %rdi,%rbx
1d: mov $0xc,%esi 3e: mov $0xc,%esi
22: callq 0xffffffffe1021e15 43: callq 0xffffffffe102bd75
27: cmp $0x86dd,%eax 48: cmp $0x86dd,%rax
2c: jne 0x0000000000000069 4f: jne 0x000000000000009a
2e: mov $0x14,%esi 51: mov $0x14,%esi
33: callq 0xffffffffe1021e31 56: callq 0xffffffffe102bd91
38: cmp $0x84,%eax 5b: cmp $0x84,%rax
3d: je 0x0000000000000049 62: je 0x0000000000000074
3f: cmp $0x6,%eax 64: cmp $0x6,%rax
42: je 0x0000000000000049 68: je 0x0000000000000074
44: cmp $0x11,%eax 6a: cmp $0x11,%rax
47: jne 0x00000000000000c6 6e: jne 0x0000000000000117
49: mov $0x36,%esi 74: mov $0x36,%esi
4e: callq 0xffffffffe1021e15 79: callq 0xffffffffe102bd75
53: cmp $0x16,%eax 7e: cmp $0x16,%rax
56: je 0x00000000000000bf 82: je 0x0000000000000110
58: mov $0x38,%esi 88: mov $0x38,%esi
5d: callq 0xffffffffe1021e15 8d: callq 0xffffffffe102bd75
62: cmp $0x16,%eax 92: cmp $0x16,%rax
65: je 0x00000000000000bf 96: je 0x0000000000000110
67: jmp 0x00000000000000c6 98: jmp 0x0000000000000117
69: cmp $0x800,%eax 9a: cmp $0x800,%rax
6e: jne 0x00000000000000c6 a1: jne 0x0000000000000117
70: mov $0x17,%esi a3: mov $0x17,%esi
75: callq 0xffffffffe1021e31 a8: callq 0xffffffffe102bd91
7a: cmp $0x84,%eax ad: cmp $0x84,%rax
7f: je 0x000000000000008b b4: je 0x00000000000000c2
81: cmp $0x6,%eax b6: cmp $0x6,%rax
84: je 0x000000000000008b ba: je 0x00000000000000c2
86: cmp $0x11,%eax bc: cmp $0x11,%rax
89: jne 0x00000000000000c6 c0: jne 0x0000000000000117
8b: mov $0x14,%esi c2: mov $0x14,%esi
90: callq 0xffffffffe1021e15 c7: callq 0xffffffffe102bd75
95: test $0x1fff,%ax cc: test $0x1fff,%rax
99: jne 0x00000000000000c6 d3: jne 0x0000000000000117
d5: mov %rax,%r14
9b: mov $0xe,%esi d8: mov $0xe,%esi
a0: callq 0xffffffffe1021e44 dd: callq 0xffffffffe102bd91 // MSH
e2: and $0xf,%eax
e5: shl $0x2,%eax
e8: mov %rax,%r13
eb: mov %r14,%rax
ee: mov %r13,%rsi
a5: lea 0xe(%rbx),%esi f1: add $0xe,%esi
a8: callq 0xffffffffe1021e0d f4: callq 0xffffffffe102bd6d
ad: cmp $0x16,%eax f9: cmp $0x16,%rax
b0: je 0x00000000000000bf fd: je 0x0000000000000110
ff: mov %r13,%rsi
b2: lea 0x10(%rbx),%esi 102: add $0x10,%esi
b5: callq 0xffffffffe1021e0d 105: callq 0xffffffffe102bd6d
ba: cmp $0x16,%eax 10a: cmp $0x16,%rax
bd: jne 0x00000000000000c6 10e: jne 0x0000000000000117
bf: mov $0xffff,%eax 110: mov $0xffff,%eax
c4: jmp 0x00000000000000c8 115: jmp 0x000000000000011c
c6: xor %eax,%eax 117: mov $0x0,%eax
c8: mov -0x8(%rbp),%rbx 11c: mov -0x228(%rbp),%rbx // epilogue
cc: leaveq 123: mov -0x220(%rbp),%r13
cd: retq 12a: mov -0x218(%rbp),%r14
131: mov -0x210(%rbp),%r15
138: leaveq
139: retq
On fully cached SKBs both JITed functions take 12 nsec to execute.
BPF interpreter executes the program in 30 nsec.
The difference in generated assembler is due to the following:
Old BPF imlements LDX_MSH instruction via sk_load_byte_msh() helper function
inside bpf_jit.S.
New JIT removes the helper and does it explicitly, so ldx_msh cost
is the same for both JITs, but generated code looks longer.
New JIT has 4 registers to save, so prologue/epilogue are larger,
but the cost is within noise on x64.
Old JIT checks whether first insn clears A and if not emits 'xor %eax,%eax'.
New JIT clears %rax unconditionally.
2. old BPF JIT doesn't support ANC_NLATTR, ANC_PAY_OFFSET, ANC_RANDOM
extensions. New JIT supports all BPF extensions.
Performance of such filters improves 2-4 times depending on a filter.
The longer the filter the higher performance gain.
Synthetic benchmarks with many ancillary loads see 20x speedup
which seems to be the maximum gain from JIT
Notes:
. net.core.bpf_jit_enable=2 + tools/net/bpf_jit_disasm is still functional
and can be used to see generated assembler
. there are two jit_compile() functions and code flow for classic filters is:
sk_attach_filter() - load classic BPF
bpf_jit_compile() - try to JIT from classic BPF
sk_convert_filter() - convert classic to internal
bpf_int_jit_compile() - JIT from internal BPF
seccomp and tracing filters will just call bpf_int_jit_compile()
Signed-off-by: Alexei Starovoitov <ast@plumgrid.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
2014-05-14 02:50:46 +00:00
|
|
|
case BPF_ALU | BPF_NEG:
|
|
|
|
case BPF_ALU64 | BPF_NEG:
|
2021-10-06 19:41:35 +00:00
|
|
|
maybe_emit_1mod(&prog, dst_reg,
|
|
|
|
BPF_CLASS(insn->code) == BPF_ALU64);
|
2014-06-06 21:46:06 +00:00
|
|
|
EMIT2(0xF7, add_1reg(0xD8, dst_reg));
|
net: filter: x86: internal BPF JIT
Maps all internal BPF instructions into x86_64 instructions.
This patch replaces original BPF x64 JIT with internal BPF x64 JIT.
sysctl net.core.bpf_jit_enable is reused as on/off switch.
Performance:
1. old BPF JIT and internal BPF JIT generate equivalent x86_64 code.
No performance difference is observed for filters that were JIT-able before
Example assembler code for BPF filter "tcpdump port 22"
original BPF -> old JIT: original BPF -> internal BPF -> new JIT:
0: push %rbp 0: push %rbp
1: mov %rsp,%rbp 1: mov %rsp,%rbp
4: sub $0x60,%rsp 4: sub $0x228,%rsp
8: mov %rbx,-0x8(%rbp) b: mov %rbx,-0x228(%rbp) // prologue
12: mov %r13,-0x220(%rbp)
19: mov %r14,-0x218(%rbp)
20: mov %r15,-0x210(%rbp)
27: xor %eax,%eax // clear A
c: xor %ebx,%ebx 29: xor %r13,%r13 // clear X
e: mov 0x68(%rdi),%r9d 2c: mov 0x68(%rdi),%r9d
12: sub 0x6c(%rdi),%r9d 30: sub 0x6c(%rdi),%r9d
16: mov 0xd8(%rdi),%r8 34: mov 0xd8(%rdi),%r10
3b: mov %rdi,%rbx
1d: mov $0xc,%esi 3e: mov $0xc,%esi
22: callq 0xffffffffe1021e15 43: callq 0xffffffffe102bd75
27: cmp $0x86dd,%eax 48: cmp $0x86dd,%rax
2c: jne 0x0000000000000069 4f: jne 0x000000000000009a
2e: mov $0x14,%esi 51: mov $0x14,%esi
33: callq 0xffffffffe1021e31 56: callq 0xffffffffe102bd91
38: cmp $0x84,%eax 5b: cmp $0x84,%rax
3d: je 0x0000000000000049 62: je 0x0000000000000074
3f: cmp $0x6,%eax 64: cmp $0x6,%rax
42: je 0x0000000000000049 68: je 0x0000000000000074
44: cmp $0x11,%eax 6a: cmp $0x11,%rax
47: jne 0x00000000000000c6 6e: jne 0x0000000000000117
49: mov $0x36,%esi 74: mov $0x36,%esi
4e: callq 0xffffffffe1021e15 79: callq 0xffffffffe102bd75
53: cmp $0x16,%eax 7e: cmp $0x16,%rax
56: je 0x00000000000000bf 82: je 0x0000000000000110
58: mov $0x38,%esi 88: mov $0x38,%esi
5d: callq 0xffffffffe1021e15 8d: callq 0xffffffffe102bd75
62: cmp $0x16,%eax 92: cmp $0x16,%rax
65: je 0x00000000000000bf 96: je 0x0000000000000110
67: jmp 0x00000000000000c6 98: jmp 0x0000000000000117
69: cmp $0x800,%eax 9a: cmp $0x800,%rax
6e: jne 0x00000000000000c6 a1: jne 0x0000000000000117
70: mov $0x17,%esi a3: mov $0x17,%esi
75: callq 0xffffffffe1021e31 a8: callq 0xffffffffe102bd91
7a: cmp $0x84,%eax ad: cmp $0x84,%rax
7f: je 0x000000000000008b b4: je 0x00000000000000c2
81: cmp $0x6,%eax b6: cmp $0x6,%rax
84: je 0x000000000000008b ba: je 0x00000000000000c2
86: cmp $0x11,%eax bc: cmp $0x11,%rax
89: jne 0x00000000000000c6 c0: jne 0x0000000000000117
8b: mov $0x14,%esi c2: mov $0x14,%esi
90: callq 0xffffffffe1021e15 c7: callq 0xffffffffe102bd75
95: test $0x1fff,%ax cc: test $0x1fff,%rax
99: jne 0x00000000000000c6 d3: jne 0x0000000000000117
d5: mov %rax,%r14
9b: mov $0xe,%esi d8: mov $0xe,%esi
a0: callq 0xffffffffe1021e44 dd: callq 0xffffffffe102bd91 // MSH
e2: and $0xf,%eax
e5: shl $0x2,%eax
e8: mov %rax,%r13
eb: mov %r14,%rax
ee: mov %r13,%rsi
a5: lea 0xe(%rbx),%esi f1: add $0xe,%esi
a8: callq 0xffffffffe1021e0d f4: callq 0xffffffffe102bd6d
ad: cmp $0x16,%eax f9: cmp $0x16,%rax
b0: je 0x00000000000000bf fd: je 0x0000000000000110
ff: mov %r13,%rsi
b2: lea 0x10(%rbx),%esi 102: add $0x10,%esi
b5: callq 0xffffffffe1021e0d 105: callq 0xffffffffe102bd6d
ba: cmp $0x16,%eax 10a: cmp $0x16,%rax
bd: jne 0x00000000000000c6 10e: jne 0x0000000000000117
bf: mov $0xffff,%eax 110: mov $0xffff,%eax
c4: jmp 0x00000000000000c8 115: jmp 0x000000000000011c
c6: xor %eax,%eax 117: mov $0x0,%eax
c8: mov -0x8(%rbp),%rbx 11c: mov -0x228(%rbp),%rbx // epilogue
cc: leaveq 123: mov -0x220(%rbp),%r13
cd: retq 12a: mov -0x218(%rbp),%r14
131: mov -0x210(%rbp),%r15
138: leaveq
139: retq
On fully cached SKBs both JITed functions take 12 nsec to execute.
BPF interpreter executes the program in 30 nsec.
The difference in generated assembler is due to the following:
Old BPF imlements LDX_MSH instruction via sk_load_byte_msh() helper function
inside bpf_jit.S.
New JIT removes the helper and does it explicitly, so ldx_msh cost
is the same for both JITs, but generated code looks longer.
New JIT has 4 registers to save, so prologue/epilogue are larger,
but the cost is within noise on x64.
Old JIT checks whether first insn clears A and if not emits 'xor %eax,%eax'.
New JIT clears %rax unconditionally.
2. old BPF JIT doesn't support ANC_NLATTR, ANC_PAY_OFFSET, ANC_RANDOM
extensions. New JIT supports all BPF extensions.
Performance of such filters improves 2-4 times depending on a filter.
The longer the filter the higher performance gain.
Synthetic benchmarks with many ancillary loads see 20x speedup
which seems to be the maximum gain from JIT
Notes:
. net.core.bpf_jit_enable=2 + tools/net/bpf_jit_disasm is still functional
and can be used to see generated assembler
. there are two jit_compile() functions and code flow for classic filters is:
sk_attach_filter() - load classic BPF
bpf_jit_compile() - try to JIT from classic BPF
sk_convert_filter() - convert classic to internal
bpf_int_jit_compile() - JIT from internal BPF
seccomp and tracing filters will just call bpf_int_jit_compile()
Signed-off-by: Alexei Starovoitov <ast@plumgrid.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
2014-05-14 02:50:46 +00:00
|
|
|
break;
|
|
|
|
|
|
|
|
case BPF_ALU | BPF_ADD | BPF_K:
|
|
|
|
case BPF_ALU | BPF_SUB | BPF_K:
|
|
|
|
case BPF_ALU | BPF_AND | BPF_K:
|
|
|
|
case BPF_ALU | BPF_OR | BPF_K:
|
|
|
|
case BPF_ALU | BPF_XOR | BPF_K:
|
|
|
|
case BPF_ALU64 | BPF_ADD | BPF_K:
|
|
|
|
case BPF_ALU64 | BPF_SUB | BPF_K:
|
|
|
|
case BPF_ALU64 | BPF_AND | BPF_K:
|
|
|
|
case BPF_ALU64 | BPF_OR | BPF_K:
|
|
|
|
case BPF_ALU64 | BPF_XOR | BPF_K:
|
2021-10-06 19:41:35 +00:00
|
|
|
maybe_emit_1mod(&prog, dst_reg,
|
|
|
|
BPF_CLASS(insn->code) == BPF_ALU64);
|
net: filter: x86: internal BPF JIT
Maps all internal BPF instructions into x86_64 instructions.
This patch replaces original BPF x64 JIT with internal BPF x64 JIT.
sysctl net.core.bpf_jit_enable is reused as on/off switch.
Performance:
1. old BPF JIT and internal BPF JIT generate equivalent x86_64 code.
No performance difference is observed for filters that were JIT-able before
Example assembler code for BPF filter "tcpdump port 22"
original BPF -> old JIT: original BPF -> internal BPF -> new JIT:
0: push %rbp 0: push %rbp
1: mov %rsp,%rbp 1: mov %rsp,%rbp
4: sub $0x60,%rsp 4: sub $0x228,%rsp
8: mov %rbx,-0x8(%rbp) b: mov %rbx,-0x228(%rbp) // prologue
12: mov %r13,-0x220(%rbp)
19: mov %r14,-0x218(%rbp)
20: mov %r15,-0x210(%rbp)
27: xor %eax,%eax // clear A
c: xor %ebx,%ebx 29: xor %r13,%r13 // clear X
e: mov 0x68(%rdi),%r9d 2c: mov 0x68(%rdi),%r9d
12: sub 0x6c(%rdi),%r9d 30: sub 0x6c(%rdi),%r9d
16: mov 0xd8(%rdi),%r8 34: mov 0xd8(%rdi),%r10
3b: mov %rdi,%rbx
1d: mov $0xc,%esi 3e: mov $0xc,%esi
22: callq 0xffffffffe1021e15 43: callq 0xffffffffe102bd75
27: cmp $0x86dd,%eax 48: cmp $0x86dd,%rax
2c: jne 0x0000000000000069 4f: jne 0x000000000000009a
2e: mov $0x14,%esi 51: mov $0x14,%esi
33: callq 0xffffffffe1021e31 56: callq 0xffffffffe102bd91
38: cmp $0x84,%eax 5b: cmp $0x84,%rax
3d: je 0x0000000000000049 62: je 0x0000000000000074
3f: cmp $0x6,%eax 64: cmp $0x6,%rax
42: je 0x0000000000000049 68: je 0x0000000000000074
44: cmp $0x11,%eax 6a: cmp $0x11,%rax
47: jne 0x00000000000000c6 6e: jne 0x0000000000000117
49: mov $0x36,%esi 74: mov $0x36,%esi
4e: callq 0xffffffffe1021e15 79: callq 0xffffffffe102bd75
53: cmp $0x16,%eax 7e: cmp $0x16,%rax
56: je 0x00000000000000bf 82: je 0x0000000000000110
58: mov $0x38,%esi 88: mov $0x38,%esi
5d: callq 0xffffffffe1021e15 8d: callq 0xffffffffe102bd75
62: cmp $0x16,%eax 92: cmp $0x16,%rax
65: je 0x00000000000000bf 96: je 0x0000000000000110
67: jmp 0x00000000000000c6 98: jmp 0x0000000000000117
69: cmp $0x800,%eax 9a: cmp $0x800,%rax
6e: jne 0x00000000000000c6 a1: jne 0x0000000000000117
70: mov $0x17,%esi a3: mov $0x17,%esi
75: callq 0xffffffffe1021e31 a8: callq 0xffffffffe102bd91
7a: cmp $0x84,%eax ad: cmp $0x84,%rax
7f: je 0x000000000000008b b4: je 0x00000000000000c2
81: cmp $0x6,%eax b6: cmp $0x6,%rax
84: je 0x000000000000008b ba: je 0x00000000000000c2
86: cmp $0x11,%eax bc: cmp $0x11,%rax
89: jne 0x00000000000000c6 c0: jne 0x0000000000000117
8b: mov $0x14,%esi c2: mov $0x14,%esi
90: callq 0xffffffffe1021e15 c7: callq 0xffffffffe102bd75
95: test $0x1fff,%ax cc: test $0x1fff,%rax
99: jne 0x00000000000000c6 d3: jne 0x0000000000000117
d5: mov %rax,%r14
9b: mov $0xe,%esi d8: mov $0xe,%esi
a0: callq 0xffffffffe1021e44 dd: callq 0xffffffffe102bd91 // MSH
e2: and $0xf,%eax
e5: shl $0x2,%eax
e8: mov %rax,%r13
eb: mov %r14,%rax
ee: mov %r13,%rsi
a5: lea 0xe(%rbx),%esi f1: add $0xe,%esi
a8: callq 0xffffffffe1021e0d f4: callq 0xffffffffe102bd6d
ad: cmp $0x16,%eax f9: cmp $0x16,%rax
b0: je 0x00000000000000bf fd: je 0x0000000000000110
ff: mov %r13,%rsi
b2: lea 0x10(%rbx),%esi 102: add $0x10,%esi
b5: callq 0xffffffffe1021e0d 105: callq 0xffffffffe102bd6d
ba: cmp $0x16,%eax 10a: cmp $0x16,%rax
bd: jne 0x00000000000000c6 10e: jne 0x0000000000000117
bf: mov $0xffff,%eax 110: mov $0xffff,%eax
c4: jmp 0x00000000000000c8 115: jmp 0x000000000000011c
c6: xor %eax,%eax 117: mov $0x0,%eax
c8: mov -0x8(%rbp),%rbx 11c: mov -0x228(%rbp),%rbx // epilogue
cc: leaveq 123: mov -0x220(%rbp),%r13
cd: retq 12a: mov -0x218(%rbp),%r14
131: mov -0x210(%rbp),%r15
138: leaveq
139: retq
On fully cached SKBs both JITed functions take 12 nsec to execute.
BPF interpreter executes the program in 30 nsec.
The difference in generated assembler is due to the following:
Old BPF imlements LDX_MSH instruction via sk_load_byte_msh() helper function
inside bpf_jit.S.
New JIT removes the helper and does it explicitly, so ldx_msh cost
is the same for both JITs, but generated code looks longer.
New JIT has 4 registers to save, so prologue/epilogue are larger,
but the cost is within noise on x64.
Old JIT checks whether first insn clears A and if not emits 'xor %eax,%eax'.
New JIT clears %rax unconditionally.
2. old BPF JIT doesn't support ANC_NLATTR, ANC_PAY_OFFSET, ANC_RANDOM
extensions. New JIT supports all BPF extensions.
Performance of such filters improves 2-4 times depending on a filter.
The longer the filter the higher performance gain.
Synthetic benchmarks with many ancillary loads see 20x speedup
which seems to be the maximum gain from JIT
Notes:
. net.core.bpf_jit_enable=2 + tools/net/bpf_jit_disasm is still functional
and can be used to see generated assembler
. there are two jit_compile() functions and code flow for classic filters is:
sk_attach_filter() - load classic BPF
bpf_jit_compile() - try to JIT from classic BPF
sk_convert_filter() - convert classic to internal
bpf_int_jit_compile() - JIT from internal BPF
seccomp and tracing filters will just call bpf_int_jit_compile()
Signed-off-by: Alexei Starovoitov <ast@plumgrid.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
2014-05-14 02:50:46 +00:00
|
|
|
|
x86/bpf: Clean up non-standard comments, to make the code more readable
So by chance I looked into x86 assembly in arch/x86/net/bpf_jit_comp.c and
noticed the weird and inconsistent comment style it mistakenly learned from
the networking code:
/* Multi-line comment ...
* ... looks like this.
*/
Fix this to use the standard comment style specified in Documentation/CodingStyle
and used in arch/x86/ as well:
/*
* Multi-line comment ...
* ... looks like this.
*/
Also, to quote Linus's ... more explicit views about this:
http://article.gmane.org/gmane.linux.kernel.cryptoapi/21066
> But no, the networking code picked *none* of the above sane formats.
> Instead, it picked these two models that are just half-arsed
> shit-for-brains:
>
> (no)
> /* This is disgusting drug-induced
> * crap, and should die
> */
>
> (no-no-no)
> /* This is also very nasty
> * and visually unbalanced */
>
> Please. The networking code actually has the *worst* possible comment
> style. You can literally find that (no-no-no) style, which is just
> really horribly disgusting and worse than the otherwise fairly similar
> (d) in pretty much every way.
Also improve the comments and some other details while at it:
- Don't mix same-line and previous-line comment style on otherwise
identical code patterns within the same function,
- capitalize 'BPF' and x86 register names consistently,
- capitalize sentences consistently,
- instead of 'x64' use 'x86-64': x64 is a Microsoft specific term,
- use more consistent punctuation,
- use standard coding style in macros as well,
- fix typos and a few other minor details.
Consistent coding style is not optional, at least in arch/x86/.
No change in functionality.
( In case this commit causes conflicts with pending development code
I'll be glad to help resolve any conflicts! )
Acked-by: Thomas Gleixner <tglx@linutronix.de>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Borislav Petkov <bp@alien8.de>
Cc: H. Peter Anvin <hpa@zytor.com>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: David S. Miller <davem@davemloft.net>
Cc: Eric Dumazet <edumazet@google.com>
Cc: Daniel Borkmann <daniel@iogearbox.net>
Cc: Alexei Starovoitov <ast@fb.com>
Cc: Hideaki YOSHIFUJI <yoshfuji@linux-ipv6.org>
Cc: netdev@vger.kernel.org
Cc: linux-kernel@vger.kernel.org
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
2018-04-27 09:54:40 +00:00
|
|
|
/*
|
|
|
|
* b3 holds 'normal' opcode, b2 short form only valid
|
2018-01-20 00:24:35 +00:00
|
|
|
* in case dst is eax/rax.
|
|
|
|
*/
|
net: filter: x86: internal BPF JIT
Maps all internal BPF instructions into x86_64 instructions.
This patch replaces original BPF x64 JIT with internal BPF x64 JIT.
sysctl net.core.bpf_jit_enable is reused as on/off switch.
Performance:
1. old BPF JIT and internal BPF JIT generate equivalent x86_64 code.
No performance difference is observed for filters that were JIT-able before
Example assembler code for BPF filter "tcpdump port 22"
original BPF -> old JIT: original BPF -> internal BPF -> new JIT:
0: push %rbp 0: push %rbp
1: mov %rsp,%rbp 1: mov %rsp,%rbp
4: sub $0x60,%rsp 4: sub $0x228,%rsp
8: mov %rbx,-0x8(%rbp) b: mov %rbx,-0x228(%rbp) // prologue
12: mov %r13,-0x220(%rbp)
19: mov %r14,-0x218(%rbp)
20: mov %r15,-0x210(%rbp)
27: xor %eax,%eax // clear A
c: xor %ebx,%ebx 29: xor %r13,%r13 // clear X
e: mov 0x68(%rdi),%r9d 2c: mov 0x68(%rdi),%r9d
12: sub 0x6c(%rdi),%r9d 30: sub 0x6c(%rdi),%r9d
16: mov 0xd8(%rdi),%r8 34: mov 0xd8(%rdi),%r10
3b: mov %rdi,%rbx
1d: mov $0xc,%esi 3e: mov $0xc,%esi
22: callq 0xffffffffe1021e15 43: callq 0xffffffffe102bd75
27: cmp $0x86dd,%eax 48: cmp $0x86dd,%rax
2c: jne 0x0000000000000069 4f: jne 0x000000000000009a
2e: mov $0x14,%esi 51: mov $0x14,%esi
33: callq 0xffffffffe1021e31 56: callq 0xffffffffe102bd91
38: cmp $0x84,%eax 5b: cmp $0x84,%rax
3d: je 0x0000000000000049 62: je 0x0000000000000074
3f: cmp $0x6,%eax 64: cmp $0x6,%rax
42: je 0x0000000000000049 68: je 0x0000000000000074
44: cmp $0x11,%eax 6a: cmp $0x11,%rax
47: jne 0x00000000000000c6 6e: jne 0x0000000000000117
49: mov $0x36,%esi 74: mov $0x36,%esi
4e: callq 0xffffffffe1021e15 79: callq 0xffffffffe102bd75
53: cmp $0x16,%eax 7e: cmp $0x16,%rax
56: je 0x00000000000000bf 82: je 0x0000000000000110
58: mov $0x38,%esi 88: mov $0x38,%esi
5d: callq 0xffffffffe1021e15 8d: callq 0xffffffffe102bd75
62: cmp $0x16,%eax 92: cmp $0x16,%rax
65: je 0x00000000000000bf 96: je 0x0000000000000110
67: jmp 0x00000000000000c6 98: jmp 0x0000000000000117
69: cmp $0x800,%eax 9a: cmp $0x800,%rax
6e: jne 0x00000000000000c6 a1: jne 0x0000000000000117
70: mov $0x17,%esi a3: mov $0x17,%esi
75: callq 0xffffffffe1021e31 a8: callq 0xffffffffe102bd91
7a: cmp $0x84,%eax ad: cmp $0x84,%rax
7f: je 0x000000000000008b b4: je 0x00000000000000c2
81: cmp $0x6,%eax b6: cmp $0x6,%rax
84: je 0x000000000000008b ba: je 0x00000000000000c2
86: cmp $0x11,%eax bc: cmp $0x11,%rax
89: jne 0x00000000000000c6 c0: jne 0x0000000000000117
8b: mov $0x14,%esi c2: mov $0x14,%esi
90: callq 0xffffffffe1021e15 c7: callq 0xffffffffe102bd75
95: test $0x1fff,%ax cc: test $0x1fff,%rax
99: jne 0x00000000000000c6 d3: jne 0x0000000000000117
d5: mov %rax,%r14
9b: mov $0xe,%esi d8: mov $0xe,%esi
a0: callq 0xffffffffe1021e44 dd: callq 0xffffffffe102bd91 // MSH
e2: and $0xf,%eax
e5: shl $0x2,%eax
e8: mov %rax,%r13
eb: mov %r14,%rax
ee: mov %r13,%rsi
a5: lea 0xe(%rbx),%esi f1: add $0xe,%esi
a8: callq 0xffffffffe1021e0d f4: callq 0xffffffffe102bd6d
ad: cmp $0x16,%eax f9: cmp $0x16,%rax
b0: je 0x00000000000000bf fd: je 0x0000000000000110
ff: mov %r13,%rsi
b2: lea 0x10(%rbx),%esi 102: add $0x10,%esi
b5: callq 0xffffffffe1021e0d 105: callq 0xffffffffe102bd6d
ba: cmp $0x16,%eax 10a: cmp $0x16,%rax
bd: jne 0x00000000000000c6 10e: jne 0x0000000000000117
bf: mov $0xffff,%eax 110: mov $0xffff,%eax
c4: jmp 0x00000000000000c8 115: jmp 0x000000000000011c
c6: xor %eax,%eax 117: mov $0x0,%eax
c8: mov -0x8(%rbp),%rbx 11c: mov -0x228(%rbp),%rbx // epilogue
cc: leaveq 123: mov -0x220(%rbp),%r13
cd: retq 12a: mov -0x218(%rbp),%r14
131: mov -0x210(%rbp),%r15
138: leaveq
139: retq
On fully cached SKBs both JITed functions take 12 nsec to execute.
BPF interpreter executes the program in 30 nsec.
The difference in generated assembler is due to the following:
Old BPF imlements LDX_MSH instruction via sk_load_byte_msh() helper function
inside bpf_jit.S.
New JIT removes the helper and does it explicitly, so ldx_msh cost
is the same for both JITs, but generated code looks longer.
New JIT has 4 registers to save, so prologue/epilogue are larger,
but the cost is within noise on x64.
Old JIT checks whether first insn clears A and if not emits 'xor %eax,%eax'.
New JIT clears %rax unconditionally.
2. old BPF JIT doesn't support ANC_NLATTR, ANC_PAY_OFFSET, ANC_RANDOM
extensions. New JIT supports all BPF extensions.
Performance of such filters improves 2-4 times depending on a filter.
The longer the filter the higher performance gain.
Synthetic benchmarks with many ancillary loads see 20x speedup
which seems to be the maximum gain from JIT
Notes:
. net.core.bpf_jit_enable=2 + tools/net/bpf_jit_disasm is still functional
and can be used to see generated assembler
. there are two jit_compile() functions and code flow for classic filters is:
sk_attach_filter() - load classic BPF
bpf_jit_compile() - try to JIT from classic BPF
sk_convert_filter() - convert classic to internal
bpf_int_jit_compile() - JIT from internal BPF
seccomp and tracing filters will just call bpf_int_jit_compile()
Signed-off-by: Alexei Starovoitov <ast@plumgrid.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
2014-05-14 02:50:46 +00:00
|
|
|
switch (BPF_OP(insn->code)) {
|
2018-01-20 00:24:35 +00:00
|
|
|
case BPF_ADD:
|
|
|
|
b3 = 0xC0;
|
|
|
|
b2 = 0x05;
|
|
|
|
break;
|
|
|
|
case BPF_SUB:
|
|
|
|
b3 = 0xE8;
|
|
|
|
b2 = 0x2D;
|
|
|
|
break;
|
|
|
|
case BPF_AND:
|
|
|
|
b3 = 0xE0;
|
|
|
|
b2 = 0x25;
|
|
|
|
break;
|
|
|
|
case BPF_OR:
|
|
|
|
b3 = 0xC8;
|
|
|
|
b2 = 0x0D;
|
|
|
|
break;
|
|
|
|
case BPF_XOR:
|
|
|
|
b3 = 0xF0;
|
|
|
|
b2 = 0x35;
|
|
|
|
break;
|
net: filter: x86: internal BPF JIT
Maps all internal BPF instructions into x86_64 instructions.
This patch replaces original BPF x64 JIT with internal BPF x64 JIT.
sysctl net.core.bpf_jit_enable is reused as on/off switch.
Performance:
1. old BPF JIT and internal BPF JIT generate equivalent x86_64 code.
No performance difference is observed for filters that were JIT-able before
Example assembler code for BPF filter "tcpdump port 22"
original BPF -> old JIT: original BPF -> internal BPF -> new JIT:
0: push %rbp 0: push %rbp
1: mov %rsp,%rbp 1: mov %rsp,%rbp
4: sub $0x60,%rsp 4: sub $0x228,%rsp
8: mov %rbx,-0x8(%rbp) b: mov %rbx,-0x228(%rbp) // prologue
12: mov %r13,-0x220(%rbp)
19: mov %r14,-0x218(%rbp)
20: mov %r15,-0x210(%rbp)
27: xor %eax,%eax // clear A
c: xor %ebx,%ebx 29: xor %r13,%r13 // clear X
e: mov 0x68(%rdi),%r9d 2c: mov 0x68(%rdi),%r9d
12: sub 0x6c(%rdi),%r9d 30: sub 0x6c(%rdi),%r9d
16: mov 0xd8(%rdi),%r8 34: mov 0xd8(%rdi),%r10
3b: mov %rdi,%rbx
1d: mov $0xc,%esi 3e: mov $0xc,%esi
22: callq 0xffffffffe1021e15 43: callq 0xffffffffe102bd75
27: cmp $0x86dd,%eax 48: cmp $0x86dd,%rax
2c: jne 0x0000000000000069 4f: jne 0x000000000000009a
2e: mov $0x14,%esi 51: mov $0x14,%esi
33: callq 0xffffffffe1021e31 56: callq 0xffffffffe102bd91
38: cmp $0x84,%eax 5b: cmp $0x84,%rax
3d: je 0x0000000000000049 62: je 0x0000000000000074
3f: cmp $0x6,%eax 64: cmp $0x6,%rax
42: je 0x0000000000000049 68: je 0x0000000000000074
44: cmp $0x11,%eax 6a: cmp $0x11,%rax
47: jne 0x00000000000000c6 6e: jne 0x0000000000000117
49: mov $0x36,%esi 74: mov $0x36,%esi
4e: callq 0xffffffffe1021e15 79: callq 0xffffffffe102bd75
53: cmp $0x16,%eax 7e: cmp $0x16,%rax
56: je 0x00000000000000bf 82: je 0x0000000000000110
58: mov $0x38,%esi 88: mov $0x38,%esi
5d: callq 0xffffffffe1021e15 8d: callq 0xffffffffe102bd75
62: cmp $0x16,%eax 92: cmp $0x16,%rax
65: je 0x00000000000000bf 96: je 0x0000000000000110
67: jmp 0x00000000000000c6 98: jmp 0x0000000000000117
69: cmp $0x800,%eax 9a: cmp $0x800,%rax
6e: jne 0x00000000000000c6 a1: jne 0x0000000000000117
70: mov $0x17,%esi a3: mov $0x17,%esi
75: callq 0xffffffffe1021e31 a8: callq 0xffffffffe102bd91
7a: cmp $0x84,%eax ad: cmp $0x84,%rax
7f: je 0x000000000000008b b4: je 0x00000000000000c2
81: cmp $0x6,%eax b6: cmp $0x6,%rax
84: je 0x000000000000008b ba: je 0x00000000000000c2
86: cmp $0x11,%eax bc: cmp $0x11,%rax
89: jne 0x00000000000000c6 c0: jne 0x0000000000000117
8b: mov $0x14,%esi c2: mov $0x14,%esi
90: callq 0xffffffffe1021e15 c7: callq 0xffffffffe102bd75
95: test $0x1fff,%ax cc: test $0x1fff,%rax
99: jne 0x00000000000000c6 d3: jne 0x0000000000000117
d5: mov %rax,%r14
9b: mov $0xe,%esi d8: mov $0xe,%esi
a0: callq 0xffffffffe1021e44 dd: callq 0xffffffffe102bd91 // MSH
e2: and $0xf,%eax
e5: shl $0x2,%eax
e8: mov %rax,%r13
eb: mov %r14,%rax
ee: mov %r13,%rsi
a5: lea 0xe(%rbx),%esi f1: add $0xe,%esi
a8: callq 0xffffffffe1021e0d f4: callq 0xffffffffe102bd6d
ad: cmp $0x16,%eax f9: cmp $0x16,%rax
b0: je 0x00000000000000bf fd: je 0x0000000000000110
ff: mov %r13,%rsi
b2: lea 0x10(%rbx),%esi 102: add $0x10,%esi
b5: callq 0xffffffffe1021e0d 105: callq 0xffffffffe102bd6d
ba: cmp $0x16,%eax 10a: cmp $0x16,%rax
bd: jne 0x00000000000000c6 10e: jne 0x0000000000000117
bf: mov $0xffff,%eax 110: mov $0xffff,%eax
c4: jmp 0x00000000000000c8 115: jmp 0x000000000000011c
c6: xor %eax,%eax 117: mov $0x0,%eax
c8: mov -0x8(%rbp),%rbx 11c: mov -0x228(%rbp),%rbx // epilogue
cc: leaveq 123: mov -0x220(%rbp),%r13
cd: retq 12a: mov -0x218(%rbp),%r14
131: mov -0x210(%rbp),%r15
138: leaveq
139: retq
On fully cached SKBs both JITed functions take 12 nsec to execute.
BPF interpreter executes the program in 30 nsec.
The difference in generated assembler is due to the following:
Old BPF imlements LDX_MSH instruction via sk_load_byte_msh() helper function
inside bpf_jit.S.
New JIT removes the helper and does it explicitly, so ldx_msh cost
is the same for both JITs, but generated code looks longer.
New JIT has 4 registers to save, so prologue/epilogue are larger,
but the cost is within noise on x64.
Old JIT checks whether first insn clears A and if not emits 'xor %eax,%eax'.
New JIT clears %rax unconditionally.
2. old BPF JIT doesn't support ANC_NLATTR, ANC_PAY_OFFSET, ANC_RANDOM
extensions. New JIT supports all BPF extensions.
Performance of such filters improves 2-4 times depending on a filter.
The longer the filter the higher performance gain.
Synthetic benchmarks with many ancillary loads see 20x speedup
which seems to be the maximum gain from JIT
Notes:
. net.core.bpf_jit_enable=2 + tools/net/bpf_jit_disasm is still functional
and can be used to see generated assembler
. there are two jit_compile() functions and code flow for classic filters is:
sk_attach_filter() - load classic BPF
bpf_jit_compile() - try to JIT from classic BPF
sk_convert_filter() - convert classic to internal
bpf_int_jit_compile() - JIT from internal BPF
seccomp and tracing filters will just call bpf_int_jit_compile()
Signed-off-by: Alexei Starovoitov <ast@plumgrid.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
2014-05-14 02:50:46 +00:00
|
|
|
}
|
|
|
|
|
2014-06-06 21:46:06 +00:00
|
|
|
if (is_imm8(imm32))
|
|
|
|
EMIT3(0x83, add_1reg(b3, dst_reg), imm32);
|
2018-01-20 00:24:35 +00:00
|
|
|
else if (is_axreg(dst_reg))
|
|
|
|
EMIT1_off32(b2, imm32);
|
net: filter: x86: internal BPF JIT
Maps all internal BPF instructions into x86_64 instructions.
This patch replaces original BPF x64 JIT with internal BPF x64 JIT.
sysctl net.core.bpf_jit_enable is reused as on/off switch.
Performance:
1. old BPF JIT and internal BPF JIT generate equivalent x86_64 code.
No performance difference is observed for filters that were JIT-able before
Example assembler code for BPF filter "tcpdump port 22"
original BPF -> old JIT: original BPF -> internal BPF -> new JIT:
0: push %rbp 0: push %rbp
1: mov %rsp,%rbp 1: mov %rsp,%rbp
4: sub $0x60,%rsp 4: sub $0x228,%rsp
8: mov %rbx,-0x8(%rbp) b: mov %rbx,-0x228(%rbp) // prologue
12: mov %r13,-0x220(%rbp)
19: mov %r14,-0x218(%rbp)
20: mov %r15,-0x210(%rbp)
27: xor %eax,%eax // clear A
c: xor %ebx,%ebx 29: xor %r13,%r13 // clear X
e: mov 0x68(%rdi),%r9d 2c: mov 0x68(%rdi),%r9d
12: sub 0x6c(%rdi),%r9d 30: sub 0x6c(%rdi),%r9d
16: mov 0xd8(%rdi),%r8 34: mov 0xd8(%rdi),%r10
3b: mov %rdi,%rbx
1d: mov $0xc,%esi 3e: mov $0xc,%esi
22: callq 0xffffffffe1021e15 43: callq 0xffffffffe102bd75
27: cmp $0x86dd,%eax 48: cmp $0x86dd,%rax
2c: jne 0x0000000000000069 4f: jne 0x000000000000009a
2e: mov $0x14,%esi 51: mov $0x14,%esi
33: callq 0xffffffffe1021e31 56: callq 0xffffffffe102bd91
38: cmp $0x84,%eax 5b: cmp $0x84,%rax
3d: je 0x0000000000000049 62: je 0x0000000000000074
3f: cmp $0x6,%eax 64: cmp $0x6,%rax
42: je 0x0000000000000049 68: je 0x0000000000000074
44: cmp $0x11,%eax 6a: cmp $0x11,%rax
47: jne 0x00000000000000c6 6e: jne 0x0000000000000117
49: mov $0x36,%esi 74: mov $0x36,%esi
4e: callq 0xffffffffe1021e15 79: callq 0xffffffffe102bd75
53: cmp $0x16,%eax 7e: cmp $0x16,%rax
56: je 0x00000000000000bf 82: je 0x0000000000000110
58: mov $0x38,%esi 88: mov $0x38,%esi
5d: callq 0xffffffffe1021e15 8d: callq 0xffffffffe102bd75
62: cmp $0x16,%eax 92: cmp $0x16,%rax
65: je 0x00000000000000bf 96: je 0x0000000000000110
67: jmp 0x00000000000000c6 98: jmp 0x0000000000000117
69: cmp $0x800,%eax 9a: cmp $0x800,%rax
6e: jne 0x00000000000000c6 a1: jne 0x0000000000000117
70: mov $0x17,%esi a3: mov $0x17,%esi
75: callq 0xffffffffe1021e31 a8: callq 0xffffffffe102bd91
7a: cmp $0x84,%eax ad: cmp $0x84,%rax
7f: je 0x000000000000008b b4: je 0x00000000000000c2
81: cmp $0x6,%eax b6: cmp $0x6,%rax
84: je 0x000000000000008b ba: je 0x00000000000000c2
86: cmp $0x11,%eax bc: cmp $0x11,%rax
89: jne 0x00000000000000c6 c0: jne 0x0000000000000117
8b: mov $0x14,%esi c2: mov $0x14,%esi
90: callq 0xffffffffe1021e15 c7: callq 0xffffffffe102bd75
95: test $0x1fff,%ax cc: test $0x1fff,%rax
99: jne 0x00000000000000c6 d3: jne 0x0000000000000117
d5: mov %rax,%r14
9b: mov $0xe,%esi d8: mov $0xe,%esi
a0: callq 0xffffffffe1021e44 dd: callq 0xffffffffe102bd91 // MSH
e2: and $0xf,%eax
e5: shl $0x2,%eax
e8: mov %rax,%r13
eb: mov %r14,%rax
ee: mov %r13,%rsi
a5: lea 0xe(%rbx),%esi f1: add $0xe,%esi
a8: callq 0xffffffffe1021e0d f4: callq 0xffffffffe102bd6d
ad: cmp $0x16,%eax f9: cmp $0x16,%rax
b0: je 0x00000000000000bf fd: je 0x0000000000000110
ff: mov %r13,%rsi
b2: lea 0x10(%rbx),%esi 102: add $0x10,%esi
b5: callq 0xffffffffe1021e0d 105: callq 0xffffffffe102bd6d
ba: cmp $0x16,%eax 10a: cmp $0x16,%rax
bd: jne 0x00000000000000c6 10e: jne 0x0000000000000117
bf: mov $0xffff,%eax 110: mov $0xffff,%eax
c4: jmp 0x00000000000000c8 115: jmp 0x000000000000011c
c6: xor %eax,%eax 117: mov $0x0,%eax
c8: mov -0x8(%rbp),%rbx 11c: mov -0x228(%rbp),%rbx // epilogue
cc: leaveq 123: mov -0x220(%rbp),%r13
cd: retq 12a: mov -0x218(%rbp),%r14
131: mov -0x210(%rbp),%r15
138: leaveq
139: retq
On fully cached SKBs both JITed functions take 12 nsec to execute.
BPF interpreter executes the program in 30 nsec.
The difference in generated assembler is due to the following:
Old BPF imlements LDX_MSH instruction via sk_load_byte_msh() helper function
inside bpf_jit.S.
New JIT removes the helper and does it explicitly, so ldx_msh cost
is the same for both JITs, but generated code looks longer.
New JIT has 4 registers to save, so prologue/epilogue are larger,
but the cost is within noise on x64.
Old JIT checks whether first insn clears A and if not emits 'xor %eax,%eax'.
New JIT clears %rax unconditionally.
2. old BPF JIT doesn't support ANC_NLATTR, ANC_PAY_OFFSET, ANC_RANDOM
extensions. New JIT supports all BPF extensions.
Performance of such filters improves 2-4 times depending on a filter.
The longer the filter the higher performance gain.
Synthetic benchmarks with many ancillary loads see 20x speedup
which seems to be the maximum gain from JIT
Notes:
. net.core.bpf_jit_enable=2 + tools/net/bpf_jit_disasm is still functional
and can be used to see generated assembler
. there are two jit_compile() functions and code flow for classic filters is:
sk_attach_filter() - load classic BPF
bpf_jit_compile() - try to JIT from classic BPF
sk_convert_filter() - convert classic to internal
bpf_int_jit_compile() - JIT from internal BPF
seccomp and tracing filters will just call bpf_int_jit_compile()
Signed-off-by: Alexei Starovoitov <ast@plumgrid.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
2014-05-14 02:50:46 +00:00
|
|
|
else
|
2014-06-06 21:46:06 +00:00
|
|
|
EMIT2_off32(0x81, add_1reg(b3, dst_reg), imm32);
|
net: filter: x86: internal BPF JIT
Maps all internal BPF instructions into x86_64 instructions.
This patch replaces original BPF x64 JIT with internal BPF x64 JIT.
sysctl net.core.bpf_jit_enable is reused as on/off switch.
Performance:
1. old BPF JIT and internal BPF JIT generate equivalent x86_64 code.
No performance difference is observed for filters that were JIT-able before
Example assembler code for BPF filter "tcpdump port 22"
original BPF -> old JIT: original BPF -> internal BPF -> new JIT:
0: push %rbp 0: push %rbp
1: mov %rsp,%rbp 1: mov %rsp,%rbp
4: sub $0x60,%rsp 4: sub $0x228,%rsp
8: mov %rbx,-0x8(%rbp) b: mov %rbx,-0x228(%rbp) // prologue
12: mov %r13,-0x220(%rbp)
19: mov %r14,-0x218(%rbp)
20: mov %r15,-0x210(%rbp)
27: xor %eax,%eax // clear A
c: xor %ebx,%ebx 29: xor %r13,%r13 // clear X
e: mov 0x68(%rdi),%r9d 2c: mov 0x68(%rdi),%r9d
12: sub 0x6c(%rdi),%r9d 30: sub 0x6c(%rdi),%r9d
16: mov 0xd8(%rdi),%r8 34: mov 0xd8(%rdi),%r10
3b: mov %rdi,%rbx
1d: mov $0xc,%esi 3e: mov $0xc,%esi
22: callq 0xffffffffe1021e15 43: callq 0xffffffffe102bd75
27: cmp $0x86dd,%eax 48: cmp $0x86dd,%rax
2c: jne 0x0000000000000069 4f: jne 0x000000000000009a
2e: mov $0x14,%esi 51: mov $0x14,%esi
33: callq 0xffffffffe1021e31 56: callq 0xffffffffe102bd91
38: cmp $0x84,%eax 5b: cmp $0x84,%rax
3d: je 0x0000000000000049 62: je 0x0000000000000074
3f: cmp $0x6,%eax 64: cmp $0x6,%rax
42: je 0x0000000000000049 68: je 0x0000000000000074
44: cmp $0x11,%eax 6a: cmp $0x11,%rax
47: jne 0x00000000000000c6 6e: jne 0x0000000000000117
49: mov $0x36,%esi 74: mov $0x36,%esi
4e: callq 0xffffffffe1021e15 79: callq 0xffffffffe102bd75
53: cmp $0x16,%eax 7e: cmp $0x16,%rax
56: je 0x00000000000000bf 82: je 0x0000000000000110
58: mov $0x38,%esi 88: mov $0x38,%esi
5d: callq 0xffffffffe1021e15 8d: callq 0xffffffffe102bd75
62: cmp $0x16,%eax 92: cmp $0x16,%rax
65: je 0x00000000000000bf 96: je 0x0000000000000110
67: jmp 0x00000000000000c6 98: jmp 0x0000000000000117
69: cmp $0x800,%eax 9a: cmp $0x800,%rax
6e: jne 0x00000000000000c6 a1: jne 0x0000000000000117
70: mov $0x17,%esi a3: mov $0x17,%esi
75: callq 0xffffffffe1021e31 a8: callq 0xffffffffe102bd91
7a: cmp $0x84,%eax ad: cmp $0x84,%rax
7f: je 0x000000000000008b b4: je 0x00000000000000c2
81: cmp $0x6,%eax b6: cmp $0x6,%rax
84: je 0x000000000000008b ba: je 0x00000000000000c2
86: cmp $0x11,%eax bc: cmp $0x11,%rax
89: jne 0x00000000000000c6 c0: jne 0x0000000000000117
8b: mov $0x14,%esi c2: mov $0x14,%esi
90: callq 0xffffffffe1021e15 c7: callq 0xffffffffe102bd75
95: test $0x1fff,%ax cc: test $0x1fff,%rax
99: jne 0x00000000000000c6 d3: jne 0x0000000000000117
d5: mov %rax,%r14
9b: mov $0xe,%esi d8: mov $0xe,%esi
a0: callq 0xffffffffe1021e44 dd: callq 0xffffffffe102bd91 // MSH
e2: and $0xf,%eax
e5: shl $0x2,%eax
e8: mov %rax,%r13
eb: mov %r14,%rax
ee: mov %r13,%rsi
a5: lea 0xe(%rbx),%esi f1: add $0xe,%esi
a8: callq 0xffffffffe1021e0d f4: callq 0xffffffffe102bd6d
ad: cmp $0x16,%eax f9: cmp $0x16,%rax
b0: je 0x00000000000000bf fd: je 0x0000000000000110
ff: mov %r13,%rsi
b2: lea 0x10(%rbx),%esi 102: add $0x10,%esi
b5: callq 0xffffffffe1021e0d 105: callq 0xffffffffe102bd6d
ba: cmp $0x16,%eax 10a: cmp $0x16,%rax
bd: jne 0x00000000000000c6 10e: jne 0x0000000000000117
bf: mov $0xffff,%eax 110: mov $0xffff,%eax
c4: jmp 0x00000000000000c8 115: jmp 0x000000000000011c
c6: xor %eax,%eax 117: mov $0x0,%eax
c8: mov -0x8(%rbp),%rbx 11c: mov -0x228(%rbp),%rbx // epilogue
cc: leaveq 123: mov -0x220(%rbp),%r13
cd: retq 12a: mov -0x218(%rbp),%r14
131: mov -0x210(%rbp),%r15
138: leaveq
139: retq
On fully cached SKBs both JITed functions take 12 nsec to execute.
BPF interpreter executes the program in 30 nsec.
The difference in generated assembler is due to the following:
Old BPF imlements LDX_MSH instruction via sk_load_byte_msh() helper function
inside bpf_jit.S.
New JIT removes the helper and does it explicitly, so ldx_msh cost
is the same for both JITs, but generated code looks longer.
New JIT has 4 registers to save, so prologue/epilogue are larger,
but the cost is within noise on x64.
Old JIT checks whether first insn clears A and if not emits 'xor %eax,%eax'.
New JIT clears %rax unconditionally.
2. old BPF JIT doesn't support ANC_NLATTR, ANC_PAY_OFFSET, ANC_RANDOM
extensions. New JIT supports all BPF extensions.
Performance of such filters improves 2-4 times depending on a filter.
The longer the filter the higher performance gain.
Synthetic benchmarks with many ancillary loads see 20x speedup
which seems to be the maximum gain from JIT
Notes:
. net.core.bpf_jit_enable=2 + tools/net/bpf_jit_disasm is still functional
and can be used to see generated assembler
. there are two jit_compile() functions and code flow for classic filters is:
sk_attach_filter() - load classic BPF
bpf_jit_compile() - try to JIT from classic BPF
sk_convert_filter() - convert classic to internal
bpf_int_jit_compile() - JIT from internal BPF
seccomp and tracing filters will just call bpf_int_jit_compile()
Signed-off-by: Alexei Starovoitov <ast@plumgrid.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
2014-05-14 02:50:46 +00:00
|
|
|
break;
|
|
|
|
|
|
|
|
case BPF_ALU64 | BPF_MOV | BPF_K:
|
|
|
|
case BPF_ALU | BPF_MOV | BPF_K:
|
2018-02-24 00:07:59 +00:00
|
|
|
emit_mov_imm32(&prog, BPF_CLASS(insn->code) == BPF_ALU64,
|
|
|
|
dst_reg, imm32);
|
net: filter: x86: internal BPF JIT
Maps all internal BPF instructions into x86_64 instructions.
This patch replaces original BPF x64 JIT with internal BPF x64 JIT.
sysctl net.core.bpf_jit_enable is reused as on/off switch.
Performance:
1. old BPF JIT and internal BPF JIT generate equivalent x86_64 code.
No performance difference is observed for filters that were JIT-able before
Example assembler code for BPF filter "tcpdump port 22"
original BPF -> old JIT: original BPF -> internal BPF -> new JIT:
0: push %rbp 0: push %rbp
1: mov %rsp,%rbp 1: mov %rsp,%rbp
4: sub $0x60,%rsp 4: sub $0x228,%rsp
8: mov %rbx,-0x8(%rbp) b: mov %rbx,-0x228(%rbp) // prologue
12: mov %r13,-0x220(%rbp)
19: mov %r14,-0x218(%rbp)
20: mov %r15,-0x210(%rbp)
27: xor %eax,%eax // clear A
c: xor %ebx,%ebx 29: xor %r13,%r13 // clear X
e: mov 0x68(%rdi),%r9d 2c: mov 0x68(%rdi),%r9d
12: sub 0x6c(%rdi),%r9d 30: sub 0x6c(%rdi),%r9d
16: mov 0xd8(%rdi),%r8 34: mov 0xd8(%rdi),%r10
3b: mov %rdi,%rbx
1d: mov $0xc,%esi 3e: mov $0xc,%esi
22: callq 0xffffffffe1021e15 43: callq 0xffffffffe102bd75
27: cmp $0x86dd,%eax 48: cmp $0x86dd,%rax
2c: jne 0x0000000000000069 4f: jne 0x000000000000009a
2e: mov $0x14,%esi 51: mov $0x14,%esi
33: callq 0xffffffffe1021e31 56: callq 0xffffffffe102bd91
38: cmp $0x84,%eax 5b: cmp $0x84,%rax
3d: je 0x0000000000000049 62: je 0x0000000000000074
3f: cmp $0x6,%eax 64: cmp $0x6,%rax
42: je 0x0000000000000049 68: je 0x0000000000000074
44: cmp $0x11,%eax 6a: cmp $0x11,%rax
47: jne 0x00000000000000c6 6e: jne 0x0000000000000117
49: mov $0x36,%esi 74: mov $0x36,%esi
4e: callq 0xffffffffe1021e15 79: callq 0xffffffffe102bd75
53: cmp $0x16,%eax 7e: cmp $0x16,%rax
56: je 0x00000000000000bf 82: je 0x0000000000000110
58: mov $0x38,%esi 88: mov $0x38,%esi
5d: callq 0xffffffffe1021e15 8d: callq 0xffffffffe102bd75
62: cmp $0x16,%eax 92: cmp $0x16,%rax
65: je 0x00000000000000bf 96: je 0x0000000000000110
67: jmp 0x00000000000000c6 98: jmp 0x0000000000000117
69: cmp $0x800,%eax 9a: cmp $0x800,%rax
6e: jne 0x00000000000000c6 a1: jne 0x0000000000000117
70: mov $0x17,%esi a3: mov $0x17,%esi
75: callq 0xffffffffe1021e31 a8: callq 0xffffffffe102bd91
7a: cmp $0x84,%eax ad: cmp $0x84,%rax
7f: je 0x000000000000008b b4: je 0x00000000000000c2
81: cmp $0x6,%eax b6: cmp $0x6,%rax
84: je 0x000000000000008b ba: je 0x00000000000000c2
86: cmp $0x11,%eax bc: cmp $0x11,%rax
89: jne 0x00000000000000c6 c0: jne 0x0000000000000117
8b: mov $0x14,%esi c2: mov $0x14,%esi
90: callq 0xffffffffe1021e15 c7: callq 0xffffffffe102bd75
95: test $0x1fff,%ax cc: test $0x1fff,%rax
99: jne 0x00000000000000c6 d3: jne 0x0000000000000117
d5: mov %rax,%r14
9b: mov $0xe,%esi d8: mov $0xe,%esi
a0: callq 0xffffffffe1021e44 dd: callq 0xffffffffe102bd91 // MSH
e2: and $0xf,%eax
e5: shl $0x2,%eax
e8: mov %rax,%r13
eb: mov %r14,%rax
ee: mov %r13,%rsi
a5: lea 0xe(%rbx),%esi f1: add $0xe,%esi
a8: callq 0xffffffffe1021e0d f4: callq 0xffffffffe102bd6d
ad: cmp $0x16,%eax f9: cmp $0x16,%rax
b0: je 0x00000000000000bf fd: je 0x0000000000000110
ff: mov %r13,%rsi
b2: lea 0x10(%rbx),%esi 102: add $0x10,%esi
b5: callq 0xffffffffe1021e0d 105: callq 0xffffffffe102bd6d
ba: cmp $0x16,%eax 10a: cmp $0x16,%rax
bd: jne 0x00000000000000c6 10e: jne 0x0000000000000117
bf: mov $0xffff,%eax 110: mov $0xffff,%eax
c4: jmp 0x00000000000000c8 115: jmp 0x000000000000011c
c6: xor %eax,%eax 117: mov $0x0,%eax
c8: mov -0x8(%rbp),%rbx 11c: mov -0x228(%rbp),%rbx // epilogue
cc: leaveq 123: mov -0x220(%rbp),%r13
cd: retq 12a: mov -0x218(%rbp),%r14
131: mov -0x210(%rbp),%r15
138: leaveq
139: retq
On fully cached SKBs both JITed functions take 12 nsec to execute.
BPF interpreter executes the program in 30 nsec.
The difference in generated assembler is due to the following:
Old BPF imlements LDX_MSH instruction via sk_load_byte_msh() helper function
inside bpf_jit.S.
New JIT removes the helper and does it explicitly, so ldx_msh cost
is the same for both JITs, but generated code looks longer.
New JIT has 4 registers to save, so prologue/epilogue are larger,
but the cost is within noise on x64.
Old JIT checks whether first insn clears A and if not emits 'xor %eax,%eax'.
New JIT clears %rax unconditionally.
2. old BPF JIT doesn't support ANC_NLATTR, ANC_PAY_OFFSET, ANC_RANDOM
extensions. New JIT supports all BPF extensions.
Performance of such filters improves 2-4 times depending on a filter.
The longer the filter the higher performance gain.
Synthetic benchmarks with many ancillary loads see 20x speedup
which seems to be the maximum gain from JIT
Notes:
. net.core.bpf_jit_enable=2 + tools/net/bpf_jit_disasm is still functional
and can be used to see generated assembler
. there are two jit_compile() functions and code flow for classic filters is:
sk_attach_filter() - load classic BPF
bpf_jit_compile() - try to JIT from classic BPF
sk_convert_filter() - convert classic to internal
bpf_int_jit_compile() - JIT from internal BPF
seccomp and tracing filters will just call bpf_int_jit_compile()
Signed-off-by: Alexei Starovoitov <ast@plumgrid.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
2014-05-14 02:50:46 +00:00
|
|
|
break;
|
|
|
|
|
net: filter: add "load 64-bit immediate" eBPF instruction
add BPF_LD_IMM64 instruction to load 64-bit immediate value into a register.
All previous instructions were 8-byte. This is first 16-byte instruction.
Two consecutive 'struct bpf_insn' blocks are interpreted as single instruction:
insn[0].code = BPF_LD | BPF_DW | BPF_IMM
insn[0].dst_reg = destination register
insn[0].imm = lower 32-bit
insn[1].code = 0
insn[1].imm = upper 32-bit
All unused fields must be zero.
Classic BPF has similar instruction: BPF_LD | BPF_W | BPF_IMM
which loads 32-bit immediate value into a register.
x64 JITs it as single 'movabsq %rax, imm64'
arm64 may JIT as sequence of four 'movk x0, #imm16, lsl #shift' insn
Note that old eBPF programs are binary compatible with new interpreter.
It helps eBPF programs load 64-bit constant into a register with one
instruction instead of using two registers and 4 instructions:
BPF_MOV32_IMM(R1, imm32)
BPF_ALU64_IMM(BPF_LSH, R1, 32)
BPF_MOV32_IMM(R2, imm32)
BPF_ALU64_REG(BPF_OR, R1, R2)
User space generated programs will use this instruction to load constants only.
To tell kernel that user space needs a pointer the _pseudo_ variant of
this instruction may be added later, which will use extra bits of encoding
to indicate what type of pointer user space is asking kernel to provide.
For example 'off' or 'src_reg' fields can be used for such purpose.
src_reg = 1 could mean that user space is asking kernel to validate and
load in-kernel map pointer.
src_reg = 2 could mean that user space needs readonly data section pointer
src_reg = 3 could mean that user space needs a pointer to per-cpu local data
All such future pseudo instructions will not be carrying the actual pointer
as part of the instruction, but rather will be treated as a request to kernel
to provide one. The kernel will verify the request_for_a_pointer, then
will drop _pseudo_ marking and will store actual internal pointer inside
the instruction, so the end result is the interpreter and JITs never
see pseudo BPF_LD_IMM64 insns and only operate on generic BPF_LD_IMM64 that
loads 64-bit immediate into a register. User space never operates on direct
pointers and verifier can easily recognize request_for_pointer vs other
instructions.
Signed-off-by: Alexei Starovoitov <ast@plumgrid.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
2014-09-05 05:17:17 +00:00
|
|
|
case BPF_LD | BPF_IMM | BPF_DW:
|
2018-02-24 00:07:59 +00:00
|
|
|
emit_mov_imm64(&prog, dst_reg, insn[1].imm, insn[0].imm);
|
net: filter: add "load 64-bit immediate" eBPF instruction
add BPF_LD_IMM64 instruction to load 64-bit immediate value into a register.
All previous instructions were 8-byte. This is first 16-byte instruction.
Two consecutive 'struct bpf_insn' blocks are interpreted as single instruction:
insn[0].code = BPF_LD | BPF_DW | BPF_IMM
insn[0].dst_reg = destination register
insn[0].imm = lower 32-bit
insn[1].code = 0
insn[1].imm = upper 32-bit
All unused fields must be zero.
Classic BPF has similar instruction: BPF_LD | BPF_W | BPF_IMM
which loads 32-bit immediate value into a register.
x64 JITs it as single 'movabsq %rax, imm64'
arm64 may JIT as sequence of four 'movk x0, #imm16, lsl #shift' insn
Note that old eBPF programs are binary compatible with new interpreter.
It helps eBPF programs load 64-bit constant into a register with one
instruction instead of using two registers and 4 instructions:
BPF_MOV32_IMM(R1, imm32)
BPF_ALU64_IMM(BPF_LSH, R1, 32)
BPF_MOV32_IMM(R2, imm32)
BPF_ALU64_REG(BPF_OR, R1, R2)
User space generated programs will use this instruction to load constants only.
To tell kernel that user space needs a pointer the _pseudo_ variant of
this instruction may be added later, which will use extra bits of encoding
to indicate what type of pointer user space is asking kernel to provide.
For example 'off' or 'src_reg' fields can be used for such purpose.
src_reg = 1 could mean that user space is asking kernel to validate and
load in-kernel map pointer.
src_reg = 2 could mean that user space needs readonly data section pointer
src_reg = 3 could mean that user space needs a pointer to per-cpu local data
All such future pseudo instructions will not be carrying the actual pointer
as part of the instruction, but rather will be treated as a request to kernel
to provide one. The kernel will verify the request_for_a_pointer, then
will drop _pseudo_ marking and will store actual internal pointer inside
the instruction, so the end result is the interpreter and JITs never
see pseudo BPF_LD_IMM64 insns and only operate on generic BPF_LD_IMM64 that
loads 64-bit immediate into a register. User space never operates on direct
pointers and verifier can easily recognize request_for_pointer vs other
instructions.
Signed-off-by: Alexei Starovoitov <ast@plumgrid.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
2014-09-05 05:17:17 +00:00
|
|
|
insn++;
|
|
|
|
i++;
|
|
|
|
break;
|
|
|
|
|
2014-06-06 21:46:06 +00:00
|
|
|
/* dst %= src, dst /= src, dst %= imm32, dst /= imm32 */
|
net: filter: x86: internal BPF JIT
Maps all internal BPF instructions into x86_64 instructions.
This patch replaces original BPF x64 JIT with internal BPF x64 JIT.
sysctl net.core.bpf_jit_enable is reused as on/off switch.
Performance:
1. old BPF JIT and internal BPF JIT generate equivalent x86_64 code.
No performance difference is observed for filters that were JIT-able before
Example assembler code for BPF filter "tcpdump port 22"
original BPF -> old JIT: original BPF -> internal BPF -> new JIT:
0: push %rbp 0: push %rbp
1: mov %rsp,%rbp 1: mov %rsp,%rbp
4: sub $0x60,%rsp 4: sub $0x228,%rsp
8: mov %rbx,-0x8(%rbp) b: mov %rbx,-0x228(%rbp) // prologue
12: mov %r13,-0x220(%rbp)
19: mov %r14,-0x218(%rbp)
20: mov %r15,-0x210(%rbp)
27: xor %eax,%eax // clear A
c: xor %ebx,%ebx 29: xor %r13,%r13 // clear X
e: mov 0x68(%rdi),%r9d 2c: mov 0x68(%rdi),%r9d
12: sub 0x6c(%rdi),%r9d 30: sub 0x6c(%rdi),%r9d
16: mov 0xd8(%rdi),%r8 34: mov 0xd8(%rdi),%r10
3b: mov %rdi,%rbx
1d: mov $0xc,%esi 3e: mov $0xc,%esi
22: callq 0xffffffffe1021e15 43: callq 0xffffffffe102bd75
27: cmp $0x86dd,%eax 48: cmp $0x86dd,%rax
2c: jne 0x0000000000000069 4f: jne 0x000000000000009a
2e: mov $0x14,%esi 51: mov $0x14,%esi
33: callq 0xffffffffe1021e31 56: callq 0xffffffffe102bd91
38: cmp $0x84,%eax 5b: cmp $0x84,%rax
3d: je 0x0000000000000049 62: je 0x0000000000000074
3f: cmp $0x6,%eax 64: cmp $0x6,%rax
42: je 0x0000000000000049 68: je 0x0000000000000074
44: cmp $0x11,%eax 6a: cmp $0x11,%rax
47: jne 0x00000000000000c6 6e: jne 0x0000000000000117
49: mov $0x36,%esi 74: mov $0x36,%esi
4e: callq 0xffffffffe1021e15 79: callq 0xffffffffe102bd75
53: cmp $0x16,%eax 7e: cmp $0x16,%rax
56: je 0x00000000000000bf 82: je 0x0000000000000110
58: mov $0x38,%esi 88: mov $0x38,%esi
5d: callq 0xffffffffe1021e15 8d: callq 0xffffffffe102bd75
62: cmp $0x16,%eax 92: cmp $0x16,%rax
65: je 0x00000000000000bf 96: je 0x0000000000000110
67: jmp 0x00000000000000c6 98: jmp 0x0000000000000117
69: cmp $0x800,%eax 9a: cmp $0x800,%rax
6e: jne 0x00000000000000c6 a1: jne 0x0000000000000117
70: mov $0x17,%esi a3: mov $0x17,%esi
75: callq 0xffffffffe1021e31 a8: callq 0xffffffffe102bd91
7a: cmp $0x84,%eax ad: cmp $0x84,%rax
7f: je 0x000000000000008b b4: je 0x00000000000000c2
81: cmp $0x6,%eax b6: cmp $0x6,%rax
84: je 0x000000000000008b ba: je 0x00000000000000c2
86: cmp $0x11,%eax bc: cmp $0x11,%rax
89: jne 0x00000000000000c6 c0: jne 0x0000000000000117
8b: mov $0x14,%esi c2: mov $0x14,%esi
90: callq 0xffffffffe1021e15 c7: callq 0xffffffffe102bd75
95: test $0x1fff,%ax cc: test $0x1fff,%rax
99: jne 0x00000000000000c6 d3: jne 0x0000000000000117
d5: mov %rax,%r14
9b: mov $0xe,%esi d8: mov $0xe,%esi
a0: callq 0xffffffffe1021e44 dd: callq 0xffffffffe102bd91 // MSH
e2: and $0xf,%eax
e5: shl $0x2,%eax
e8: mov %rax,%r13
eb: mov %r14,%rax
ee: mov %r13,%rsi
a5: lea 0xe(%rbx),%esi f1: add $0xe,%esi
a8: callq 0xffffffffe1021e0d f4: callq 0xffffffffe102bd6d
ad: cmp $0x16,%eax f9: cmp $0x16,%rax
b0: je 0x00000000000000bf fd: je 0x0000000000000110
ff: mov %r13,%rsi
b2: lea 0x10(%rbx),%esi 102: add $0x10,%esi
b5: callq 0xffffffffe1021e0d 105: callq 0xffffffffe102bd6d
ba: cmp $0x16,%eax 10a: cmp $0x16,%rax
bd: jne 0x00000000000000c6 10e: jne 0x0000000000000117
bf: mov $0xffff,%eax 110: mov $0xffff,%eax
c4: jmp 0x00000000000000c8 115: jmp 0x000000000000011c
c6: xor %eax,%eax 117: mov $0x0,%eax
c8: mov -0x8(%rbp),%rbx 11c: mov -0x228(%rbp),%rbx // epilogue
cc: leaveq 123: mov -0x220(%rbp),%r13
cd: retq 12a: mov -0x218(%rbp),%r14
131: mov -0x210(%rbp),%r15
138: leaveq
139: retq
On fully cached SKBs both JITed functions take 12 nsec to execute.
BPF interpreter executes the program in 30 nsec.
The difference in generated assembler is due to the following:
Old BPF imlements LDX_MSH instruction via sk_load_byte_msh() helper function
inside bpf_jit.S.
New JIT removes the helper and does it explicitly, so ldx_msh cost
is the same for both JITs, but generated code looks longer.
New JIT has 4 registers to save, so prologue/epilogue are larger,
but the cost is within noise on x64.
Old JIT checks whether first insn clears A and if not emits 'xor %eax,%eax'.
New JIT clears %rax unconditionally.
2. old BPF JIT doesn't support ANC_NLATTR, ANC_PAY_OFFSET, ANC_RANDOM
extensions. New JIT supports all BPF extensions.
Performance of such filters improves 2-4 times depending on a filter.
The longer the filter the higher performance gain.
Synthetic benchmarks with many ancillary loads see 20x speedup
which seems to be the maximum gain from JIT
Notes:
. net.core.bpf_jit_enable=2 + tools/net/bpf_jit_disasm is still functional
and can be used to see generated assembler
. there are two jit_compile() functions and code flow for classic filters is:
sk_attach_filter() - load classic BPF
bpf_jit_compile() - try to JIT from classic BPF
sk_convert_filter() - convert classic to internal
bpf_int_jit_compile() - JIT from internal BPF
seccomp and tracing filters will just call bpf_int_jit_compile()
Signed-off-by: Alexei Starovoitov <ast@plumgrid.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
2014-05-14 02:50:46 +00:00
|
|
|
case BPF_ALU | BPF_MOD | BPF_X:
|
|
|
|
case BPF_ALU | BPF_DIV | BPF_X:
|
|
|
|
case BPF_ALU | BPF_MOD | BPF_K:
|
|
|
|
case BPF_ALU | BPF_DIV | BPF_K:
|
|
|
|
case BPF_ALU64 | BPF_MOD | BPF_X:
|
|
|
|
case BPF_ALU64 | BPF_DIV | BPF_X:
|
|
|
|
case BPF_ALU64 | BPF_MOD | BPF_K:
|
bpf, x64: Save bytes for DIV by reducing reg copies
Instead of unconditionally performing push/pop on %rax/%rdx in case of
division/modulo, we can save a few bytes in case of destination register
being either BPF r0 (%rax) or r3 (%rdx) since the result is written in
there anyway.
Also, we do not need to copy the source to %r11 unless the source is either
%rax, %rdx or an immediate.
For example, before the patch:
22: push %rax
23: push %rdx
24: mov %rsi,%r11
27: xor %edx,%edx
29: div %r11
2c: mov %rax,%r11
2f: pop %rdx
30: pop %rax
31: mov %r11,%rax
After:
22: push %rdx
23: xor %edx,%edx
25: div %rsi
28: pop %rdx
Signed-off-by: Jie Meng <jmeng@fb.com>
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Tested-by: Daniel Borkmann <daniel@iogearbox.net>
Acked-by: Daniel Borkmann <daniel@iogearbox.net>
Link: https://lore.kernel.org/bpf/20211002035626.2041910-1-jmeng@fb.com
2021-10-02 03:56:26 +00:00
|
|
|
case BPF_ALU64 | BPF_DIV | BPF_K: {
|
|
|
|
bool is64 = BPF_CLASS(insn->code) == BPF_ALU64;
|
net: filter: x86: internal BPF JIT
Maps all internal BPF instructions into x86_64 instructions.
This patch replaces original BPF x64 JIT with internal BPF x64 JIT.
sysctl net.core.bpf_jit_enable is reused as on/off switch.
Performance:
1. old BPF JIT and internal BPF JIT generate equivalent x86_64 code.
No performance difference is observed for filters that were JIT-able before
Example assembler code for BPF filter "tcpdump port 22"
original BPF -> old JIT: original BPF -> internal BPF -> new JIT:
0: push %rbp 0: push %rbp
1: mov %rsp,%rbp 1: mov %rsp,%rbp
4: sub $0x60,%rsp 4: sub $0x228,%rsp
8: mov %rbx,-0x8(%rbp) b: mov %rbx,-0x228(%rbp) // prologue
12: mov %r13,-0x220(%rbp)
19: mov %r14,-0x218(%rbp)
20: mov %r15,-0x210(%rbp)
27: xor %eax,%eax // clear A
c: xor %ebx,%ebx 29: xor %r13,%r13 // clear X
e: mov 0x68(%rdi),%r9d 2c: mov 0x68(%rdi),%r9d
12: sub 0x6c(%rdi),%r9d 30: sub 0x6c(%rdi),%r9d
16: mov 0xd8(%rdi),%r8 34: mov 0xd8(%rdi),%r10
3b: mov %rdi,%rbx
1d: mov $0xc,%esi 3e: mov $0xc,%esi
22: callq 0xffffffffe1021e15 43: callq 0xffffffffe102bd75
27: cmp $0x86dd,%eax 48: cmp $0x86dd,%rax
2c: jne 0x0000000000000069 4f: jne 0x000000000000009a
2e: mov $0x14,%esi 51: mov $0x14,%esi
33: callq 0xffffffffe1021e31 56: callq 0xffffffffe102bd91
38: cmp $0x84,%eax 5b: cmp $0x84,%rax
3d: je 0x0000000000000049 62: je 0x0000000000000074
3f: cmp $0x6,%eax 64: cmp $0x6,%rax
42: je 0x0000000000000049 68: je 0x0000000000000074
44: cmp $0x11,%eax 6a: cmp $0x11,%rax
47: jne 0x00000000000000c6 6e: jne 0x0000000000000117
49: mov $0x36,%esi 74: mov $0x36,%esi
4e: callq 0xffffffffe1021e15 79: callq 0xffffffffe102bd75
53: cmp $0x16,%eax 7e: cmp $0x16,%rax
56: je 0x00000000000000bf 82: je 0x0000000000000110
58: mov $0x38,%esi 88: mov $0x38,%esi
5d: callq 0xffffffffe1021e15 8d: callq 0xffffffffe102bd75
62: cmp $0x16,%eax 92: cmp $0x16,%rax
65: je 0x00000000000000bf 96: je 0x0000000000000110
67: jmp 0x00000000000000c6 98: jmp 0x0000000000000117
69: cmp $0x800,%eax 9a: cmp $0x800,%rax
6e: jne 0x00000000000000c6 a1: jne 0x0000000000000117
70: mov $0x17,%esi a3: mov $0x17,%esi
75: callq 0xffffffffe1021e31 a8: callq 0xffffffffe102bd91
7a: cmp $0x84,%eax ad: cmp $0x84,%rax
7f: je 0x000000000000008b b4: je 0x00000000000000c2
81: cmp $0x6,%eax b6: cmp $0x6,%rax
84: je 0x000000000000008b ba: je 0x00000000000000c2
86: cmp $0x11,%eax bc: cmp $0x11,%rax
89: jne 0x00000000000000c6 c0: jne 0x0000000000000117
8b: mov $0x14,%esi c2: mov $0x14,%esi
90: callq 0xffffffffe1021e15 c7: callq 0xffffffffe102bd75
95: test $0x1fff,%ax cc: test $0x1fff,%rax
99: jne 0x00000000000000c6 d3: jne 0x0000000000000117
d5: mov %rax,%r14
9b: mov $0xe,%esi d8: mov $0xe,%esi
a0: callq 0xffffffffe1021e44 dd: callq 0xffffffffe102bd91 // MSH
e2: and $0xf,%eax
e5: shl $0x2,%eax
e8: mov %rax,%r13
eb: mov %r14,%rax
ee: mov %r13,%rsi
a5: lea 0xe(%rbx),%esi f1: add $0xe,%esi
a8: callq 0xffffffffe1021e0d f4: callq 0xffffffffe102bd6d
ad: cmp $0x16,%eax f9: cmp $0x16,%rax
b0: je 0x00000000000000bf fd: je 0x0000000000000110
ff: mov %r13,%rsi
b2: lea 0x10(%rbx),%esi 102: add $0x10,%esi
b5: callq 0xffffffffe1021e0d 105: callq 0xffffffffe102bd6d
ba: cmp $0x16,%eax 10a: cmp $0x16,%rax
bd: jne 0x00000000000000c6 10e: jne 0x0000000000000117
bf: mov $0xffff,%eax 110: mov $0xffff,%eax
c4: jmp 0x00000000000000c8 115: jmp 0x000000000000011c
c6: xor %eax,%eax 117: mov $0x0,%eax
c8: mov -0x8(%rbp),%rbx 11c: mov -0x228(%rbp),%rbx // epilogue
cc: leaveq 123: mov -0x220(%rbp),%r13
cd: retq 12a: mov -0x218(%rbp),%r14
131: mov -0x210(%rbp),%r15
138: leaveq
139: retq
On fully cached SKBs both JITed functions take 12 nsec to execute.
BPF interpreter executes the program in 30 nsec.
The difference in generated assembler is due to the following:
Old BPF imlements LDX_MSH instruction via sk_load_byte_msh() helper function
inside bpf_jit.S.
New JIT removes the helper and does it explicitly, so ldx_msh cost
is the same for both JITs, but generated code looks longer.
New JIT has 4 registers to save, so prologue/epilogue are larger,
but the cost is within noise on x64.
Old JIT checks whether first insn clears A and if not emits 'xor %eax,%eax'.
New JIT clears %rax unconditionally.
2. old BPF JIT doesn't support ANC_NLATTR, ANC_PAY_OFFSET, ANC_RANDOM
extensions. New JIT supports all BPF extensions.
Performance of such filters improves 2-4 times depending on a filter.
The longer the filter the higher performance gain.
Synthetic benchmarks with many ancillary loads see 20x speedup
which seems to be the maximum gain from JIT
Notes:
. net.core.bpf_jit_enable=2 + tools/net/bpf_jit_disasm is still functional
and can be used to see generated assembler
. there are two jit_compile() functions and code flow for classic filters is:
sk_attach_filter() - load classic BPF
bpf_jit_compile() - try to JIT from classic BPF
sk_convert_filter() - convert classic to internal
bpf_int_jit_compile() - JIT from internal BPF
seccomp and tracing filters will just call bpf_int_jit_compile()
Signed-off-by: Alexei Starovoitov <ast@plumgrid.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
2014-05-14 02:50:46 +00:00
|
|
|
|
bpf, x64: Save bytes for DIV by reducing reg copies
Instead of unconditionally performing push/pop on %rax/%rdx in case of
division/modulo, we can save a few bytes in case of destination register
being either BPF r0 (%rax) or r3 (%rdx) since the result is written in
there anyway.
Also, we do not need to copy the source to %r11 unless the source is either
%rax, %rdx or an immediate.
For example, before the patch:
22: push %rax
23: push %rdx
24: mov %rsi,%r11
27: xor %edx,%edx
29: div %r11
2c: mov %rax,%r11
2f: pop %rdx
30: pop %rax
31: mov %r11,%rax
After:
22: push %rdx
23: xor %edx,%edx
25: div %rsi
28: pop %rdx
Signed-off-by: Jie Meng <jmeng@fb.com>
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Tested-by: Daniel Borkmann <daniel@iogearbox.net>
Acked-by: Daniel Borkmann <daniel@iogearbox.net>
Link: https://lore.kernel.org/bpf/20211002035626.2041910-1-jmeng@fb.com
2021-10-02 03:56:26 +00:00
|
|
|
if (dst_reg != BPF_REG_0)
|
|
|
|
EMIT1(0x50); /* push rax */
|
|
|
|
if (dst_reg != BPF_REG_3)
|
|
|
|
EMIT1(0x52); /* push rdx */
|
|
|
|
|
|
|
|
if (BPF_SRC(insn->code) == BPF_X) {
|
|
|
|
if (src_reg == BPF_REG_0 ||
|
|
|
|
src_reg == BPF_REG_3) {
|
|
|
|
/* mov r11, src_reg */
|
|
|
|
EMIT_mov(AUX_REG, src_reg);
|
|
|
|
src_reg = AUX_REG;
|
|
|
|
}
|
|
|
|
} else {
|
2014-06-06 21:46:06 +00:00
|
|
|
/* mov r11, imm32 */
|
|
|
|
EMIT3_off32(0x49, 0xC7, 0xC3, imm32);
|
bpf, x64: Save bytes for DIV by reducing reg copies
Instead of unconditionally performing push/pop on %rax/%rdx in case of
division/modulo, we can save a few bytes in case of destination register
being either BPF r0 (%rax) or r3 (%rdx) since the result is written in
there anyway.
Also, we do not need to copy the source to %r11 unless the source is either
%rax, %rdx or an immediate.
For example, before the patch:
22: push %rax
23: push %rdx
24: mov %rsi,%r11
27: xor %edx,%edx
29: div %r11
2c: mov %rax,%r11
2f: pop %rdx
30: pop %rax
31: mov %r11,%rax
After:
22: push %rdx
23: xor %edx,%edx
25: div %rsi
28: pop %rdx
Signed-off-by: Jie Meng <jmeng@fb.com>
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Tested-by: Daniel Borkmann <daniel@iogearbox.net>
Acked-by: Daniel Borkmann <daniel@iogearbox.net>
Link: https://lore.kernel.org/bpf/20211002035626.2041910-1-jmeng@fb.com
2021-10-02 03:56:26 +00:00
|
|
|
src_reg = AUX_REG;
|
|
|
|
}
|
net: filter: x86: internal BPF JIT
Maps all internal BPF instructions into x86_64 instructions.
This patch replaces original BPF x64 JIT with internal BPF x64 JIT.
sysctl net.core.bpf_jit_enable is reused as on/off switch.
Performance:
1. old BPF JIT and internal BPF JIT generate equivalent x86_64 code.
No performance difference is observed for filters that were JIT-able before
Example assembler code for BPF filter "tcpdump port 22"
original BPF -> old JIT: original BPF -> internal BPF -> new JIT:
0: push %rbp 0: push %rbp
1: mov %rsp,%rbp 1: mov %rsp,%rbp
4: sub $0x60,%rsp 4: sub $0x228,%rsp
8: mov %rbx,-0x8(%rbp) b: mov %rbx,-0x228(%rbp) // prologue
12: mov %r13,-0x220(%rbp)
19: mov %r14,-0x218(%rbp)
20: mov %r15,-0x210(%rbp)
27: xor %eax,%eax // clear A
c: xor %ebx,%ebx 29: xor %r13,%r13 // clear X
e: mov 0x68(%rdi),%r9d 2c: mov 0x68(%rdi),%r9d
12: sub 0x6c(%rdi),%r9d 30: sub 0x6c(%rdi),%r9d
16: mov 0xd8(%rdi),%r8 34: mov 0xd8(%rdi),%r10
3b: mov %rdi,%rbx
1d: mov $0xc,%esi 3e: mov $0xc,%esi
22: callq 0xffffffffe1021e15 43: callq 0xffffffffe102bd75
27: cmp $0x86dd,%eax 48: cmp $0x86dd,%rax
2c: jne 0x0000000000000069 4f: jne 0x000000000000009a
2e: mov $0x14,%esi 51: mov $0x14,%esi
33: callq 0xffffffffe1021e31 56: callq 0xffffffffe102bd91
38: cmp $0x84,%eax 5b: cmp $0x84,%rax
3d: je 0x0000000000000049 62: je 0x0000000000000074
3f: cmp $0x6,%eax 64: cmp $0x6,%rax
42: je 0x0000000000000049 68: je 0x0000000000000074
44: cmp $0x11,%eax 6a: cmp $0x11,%rax
47: jne 0x00000000000000c6 6e: jne 0x0000000000000117
49: mov $0x36,%esi 74: mov $0x36,%esi
4e: callq 0xffffffffe1021e15 79: callq 0xffffffffe102bd75
53: cmp $0x16,%eax 7e: cmp $0x16,%rax
56: je 0x00000000000000bf 82: je 0x0000000000000110
58: mov $0x38,%esi 88: mov $0x38,%esi
5d: callq 0xffffffffe1021e15 8d: callq 0xffffffffe102bd75
62: cmp $0x16,%eax 92: cmp $0x16,%rax
65: je 0x00000000000000bf 96: je 0x0000000000000110
67: jmp 0x00000000000000c6 98: jmp 0x0000000000000117
69: cmp $0x800,%eax 9a: cmp $0x800,%rax
6e: jne 0x00000000000000c6 a1: jne 0x0000000000000117
70: mov $0x17,%esi a3: mov $0x17,%esi
75: callq 0xffffffffe1021e31 a8: callq 0xffffffffe102bd91
7a: cmp $0x84,%eax ad: cmp $0x84,%rax
7f: je 0x000000000000008b b4: je 0x00000000000000c2
81: cmp $0x6,%eax b6: cmp $0x6,%rax
84: je 0x000000000000008b ba: je 0x00000000000000c2
86: cmp $0x11,%eax bc: cmp $0x11,%rax
89: jne 0x00000000000000c6 c0: jne 0x0000000000000117
8b: mov $0x14,%esi c2: mov $0x14,%esi
90: callq 0xffffffffe1021e15 c7: callq 0xffffffffe102bd75
95: test $0x1fff,%ax cc: test $0x1fff,%rax
99: jne 0x00000000000000c6 d3: jne 0x0000000000000117
d5: mov %rax,%r14
9b: mov $0xe,%esi d8: mov $0xe,%esi
a0: callq 0xffffffffe1021e44 dd: callq 0xffffffffe102bd91 // MSH
e2: and $0xf,%eax
e5: shl $0x2,%eax
e8: mov %rax,%r13
eb: mov %r14,%rax
ee: mov %r13,%rsi
a5: lea 0xe(%rbx),%esi f1: add $0xe,%esi
a8: callq 0xffffffffe1021e0d f4: callq 0xffffffffe102bd6d
ad: cmp $0x16,%eax f9: cmp $0x16,%rax
b0: je 0x00000000000000bf fd: je 0x0000000000000110
ff: mov %r13,%rsi
b2: lea 0x10(%rbx),%esi 102: add $0x10,%esi
b5: callq 0xffffffffe1021e0d 105: callq 0xffffffffe102bd6d
ba: cmp $0x16,%eax 10a: cmp $0x16,%rax
bd: jne 0x00000000000000c6 10e: jne 0x0000000000000117
bf: mov $0xffff,%eax 110: mov $0xffff,%eax
c4: jmp 0x00000000000000c8 115: jmp 0x000000000000011c
c6: xor %eax,%eax 117: mov $0x0,%eax
c8: mov -0x8(%rbp),%rbx 11c: mov -0x228(%rbp),%rbx // epilogue
cc: leaveq 123: mov -0x220(%rbp),%r13
cd: retq 12a: mov -0x218(%rbp),%r14
131: mov -0x210(%rbp),%r15
138: leaveq
139: retq
On fully cached SKBs both JITed functions take 12 nsec to execute.
BPF interpreter executes the program in 30 nsec.
The difference in generated assembler is due to the following:
Old BPF imlements LDX_MSH instruction via sk_load_byte_msh() helper function
inside bpf_jit.S.
New JIT removes the helper and does it explicitly, so ldx_msh cost
is the same for both JITs, but generated code looks longer.
New JIT has 4 registers to save, so prologue/epilogue are larger,
but the cost is within noise on x64.
Old JIT checks whether first insn clears A and if not emits 'xor %eax,%eax'.
New JIT clears %rax unconditionally.
2. old BPF JIT doesn't support ANC_NLATTR, ANC_PAY_OFFSET, ANC_RANDOM
extensions. New JIT supports all BPF extensions.
Performance of such filters improves 2-4 times depending on a filter.
The longer the filter the higher performance gain.
Synthetic benchmarks with many ancillary loads see 20x speedup
which seems to be the maximum gain from JIT
Notes:
. net.core.bpf_jit_enable=2 + tools/net/bpf_jit_disasm is still functional
and can be used to see generated assembler
. there are two jit_compile() functions and code flow for classic filters is:
sk_attach_filter() - load classic BPF
bpf_jit_compile() - try to JIT from classic BPF
sk_convert_filter() - convert classic to internal
bpf_int_jit_compile() - JIT from internal BPF
seccomp and tracing filters will just call bpf_int_jit_compile()
Signed-off-by: Alexei Starovoitov <ast@plumgrid.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
2014-05-14 02:50:46 +00:00
|
|
|
|
bpf, x64: Save bytes for DIV by reducing reg copies
Instead of unconditionally performing push/pop on %rax/%rdx in case of
division/modulo, we can save a few bytes in case of destination register
being either BPF r0 (%rax) or r3 (%rdx) since the result is written in
there anyway.
Also, we do not need to copy the source to %r11 unless the source is either
%rax, %rdx or an immediate.
For example, before the patch:
22: push %rax
23: push %rdx
24: mov %rsi,%r11
27: xor %edx,%edx
29: div %r11
2c: mov %rax,%r11
2f: pop %rdx
30: pop %rax
31: mov %r11,%rax
After:
22: push %rdx
23: xor %edx,%edx
25: div %rsi
28: pop %rdx
Signed-off-by: Jie Meng <jmeng@fb.com>
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Tested-by: Daniel Borkmann <daniel@iogearbox.net>
Acked-by: Daniel Borkmann <daniel@iogearbox.net>
Link: https://lore.kernel.org/bpf/20211002035626.2041910-1-jmeng@fb.com
2021-10-02 03:56:26 +00:00
|
|
|
if (dst_reg != BPF_REG_0)
|
|
|
|
/* mov rax, dst_reg */
|
|
|
|
emit_mov_reg(&prog, is64, BPF_REG_0, dst_reg);
|
net: filter: x86: internal BPF JIT
Maps all internal BPF instructions into x86_64 instructions.
This patch replaces original BPF x64 JIT with internal BPF x64 JIT.
sysctl net.core.bpf_jit_enable is reused as on/off switch.
Performance:
1. old BPF JIT and internal BPF JIT generate equivalent x86_64 code.
No performance difference is observed for filters that were JIT-able before
Example assembler code for BPF filter "tcpdump port 22"
original BPF -> old JIT: original BPF -> internal BPF -> new JIT:
0: push %rbp 0: push %rbp
1: mov %rsp,%rbp 1: mov %rsp,%rbp
4: sub $0x60,%rsp 4: sub $0x228,%rsp
8: mov %rbx,-0x8(%rbp) b: mov %rbx,-0x228(%rbp) // prologue
12: mov %r13,-0x220(%rbp)
19: mov %r14,-0x218(%rbp)
20: mov %r15,-0x210(%rbp)
27: xor %eax,%eax // clear A
c: xor %ebx,%ebx 29: xor %r13,%r13 // clear X
e: mov 0x68(%rdi),%r9d 2c: mov 0x68(%rdi),%r9d
12: sub 0x6c(%rdi),%r9d 30: sub 0x6c(%rdi),%r9d
16: mov 0xd8(%rdi),%r8 34: mov 0xd8(%rdi),%r10
3b: mov %rdi,%rbx
1d: mov $0xc,%esi 3e: mov $0xc,%esi
22: callq 0xffffffffe1021e15 43: callq 0xffffffffe102bd75
27: cmp $0x86dd,%eax 48: cmp $0x86dd,%rax
2c: jne 0x0000000000000069 4f: jne 0x000000000000009a
2e: mov $0x14,%esi 51: mov $0x14,%esi
33: callq 0xffffffffe1021e31 56: callq 0xffffffffe102bd91
38: cmp $0x84,%eax 5b: cmp $0x84,%rax
3d: je 0x0000000000000049 62: je 0x0000000000000074
3f: cmp $0x6,%eax 64: cmp $0x6,%rax
42: je 0x0000000000000049 68: je 0x0000000000000074
44: cmp $0x11,%eax 6a: cmp $0x11,%rax
47: jne 0x00000000000000c6 6e: jne 0x0000000000000117
49: mov $0x36,%esi 74: mov $0x36,%esi
4e: callq 0xffffffffe1021e15 79: callq 0xffffffffe102bd75
53: cmp $0x16,%eax 7e: cmp $0x16,%rax
56: je 0x00000000000000bf 82: je 0x0000000000000110
58: mov $0x38,%esi 88: mov $0x38,%esi
5d: callq 0xffffffffe1021e15 8d: callq 0xffffffffe102bd75
62: cmp $0x16,%eax 92: cmp $0x16,%rax
65: je 0x00000000000000bf 96: je 0x0000000000000110
67: jmp 0x00000000000000c6 98: jmp 0x0000000000000117
69: cmp $0x800,%eax 9a: cmp $0x800,%rax
6e: jne 0x00000000000000c6 a1: jne 0x0000000000000117
70: mov $0x17,%esi a3: mov $0x17,%esi
75: callq 0xffffffffe1021e31 a8: callq 0xffffffffe102bd91
7a: cmp $0x84,%eax ad: cmp $0x84,%rax
7f: je 0x000000000000008b b4: je 0x00000000000000c2
81: cmp $0x6,%eax b6: cmp $0x6,%rax
84: je 0x000000000000008b ba: je 0x00000000000000c2
86: cmp $0x11,%eax bc: cmp $0x11,%rax
89: jne 0x00000000000000c6 c0: jne 0x0000000000000117
8b: mov $0x14,%esi c2: mov $0x14,%esi
90: callq 0xffffffffe1021e15 c7: callq 0xffffffffe102bd75
95: test $0x1fff,%ax cc: test $0x1fff,%rax
99: jne 0x00000000000000c6 d3: jne 0x0000000000000117
d5: mov %rax,%r14
9b: mov $0xe,%esi d8: mov $0xe,%esi
a0: callq 0xffffffffe1021e44 dd: callq 0xffffffffe102bd91 // MSH
e2: and $0xf,%eax
e5: shl $0x2,%eax
e8: mov %rax,%r13
eb: mov %r14,%rax
ee: mov %r13,%rsi
a5: lea 0xe(%rbx),%esi f1: add $0xe,%esi
a8: callq 0xffffffffe1021e0d f4: callq 0xffffffffe102bd6d
ad: cmp $0x16,%eax f9: cmp $0x16,%rax
b0: je 0x00000000000000bf fd: je 0x0000000000000110
ff: mov %r13,%rsi
b2: lea 0x10(%rbx),%esi 102: add $0x10,%esi
b5: callq 0xffffffffe1021e0d 105: callq 0xffffffffe102bd6d
ba: cmp $0x16,%eax 10a: cmp $0x16,%rax
bd: jne 0x00000000000000c6 10e: jne 0x0000000000000117
bf: mov $0xffff,%eax 110: mov $0xffff,%eax
c4: jmp 0x00000000000000c8 115: jmp 0x000000000000011c
c6: xor %eax,%eax 117: mov $0x0,%eax
c8: mov -0x8(%rbp),%rbx 11c: mov -0x228(%rbp),%rbx // epilogue
cc: leaveq 123: mov -0x220(%rbp),%r13
cd: retq 12a: mov -0x218(%rbp),%r14
131: mov -0x210(%rbp),%r15
138: leaveq
139: retq
On fully cached SKBs both JITed functions take 12 nsec to execute.
BPF interpreter executes the program in 30 nsec.
The difference in generated assembler is due to the following:
Old BPF imlements LDX_MSH instruction via sk_load_byte_msh() helper function
inside bpf_jit.S.
New JIT removes the helper and does it explicitly, so ldx_msh cost
is the same for both JITs, but generated code looks longer.
New JIT has 4 registers to save, so prologue/epilogue are larger,
but the cost is within noise on x64.
Old JIT checks whether first insn clears A and if not emits 'xor %eax,%eax'.
New JIT clears %rax unconditionally.
2. old BPF JIT doesn't support ANC_NLATTR, ANC_PAY_OFFSET, ANC_RANDOM
extensions. New JIT supports all BPF extensions.
Performance of such filters improves 2-4 times depending on a filter.
The longer the filter the higher performance gain.
Synthetic benchmarks with many ancillary loads see 20x speedup
which seems to be the maximum gain from JIT
Notes:
. net.core.bpf_jit_enable=2 + tools/net/bpf_jit_disasm is still functional
and can be used to see generated assembler
. there are two jit_compile() functions and code flow for classic filters is:
sk_attach_filter() - load classic BPF
bpf_jit_compile() - try to JIT from classic BPF
sk_convert_filter() - convert classic to internal
bpf_int_jit_compile() - JIT from internal BPF
seccomp and tracing filters will just call bpf_int_jit_compile()
Signed-off-by: Alexei Starovoitov <ast@plumgrid.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
2014-05-14 02:50:46 +00:00
|
|
|
|
2023-07-28 01:12:19 +00:00
|
|
|
if (insn->off == 0) {
|
|
|
|
/*
|
|
|
|
* xor edx, edx
|
|
|
|
* equivalent to 'xor rdx, rdx', but one byte less
|
|
|
|
*/
|
|
|
|
EMIT2(0x31, 0xd2);
|
|
|
|
|
|
|
|
/* div src_reg */
|
|
|
|
maybe_emit_1mod(&prog, src_reg, is64);
|
|
|
|
EMIT2(0xF7, add_1reg(0xF0, src_reg));
|
|
|
|
} else {
|
|
|
|
if (BPF_CLASS(insn->code) == BPF_ALU)
|
|
|
|
EMIT1(0x99); /* cdq */
|
|
|
|
else
|
|
|
|
EMIT2(0x48, 0x99); /* cqo */
|
|
|
|
|
|
|
|
/* idiv src_reg */
|
|
|
|
maybe_emit_1mod(&prog, src_reg, is64);
|
|
|
|
EMIT2(0xF7, add_1reg(0xF8, src_reg));
|
|
|
|
}
|
net: filter: x86: internal BPF JIT
Maps all internal BPF instructions into x86_64 instructions.
This patch replaces original BPF x64 JIT with internal BPF x64 JIT.
sysctl net.core.bpf_jit_enable is reused as on/off switch.
Performance:
1. old BPF JIT and internal BPF JIT generate equivalent x86_64 code.
No performance difference is observed for filters that were JIT-able before
Example assembler code for BPF filter "tcpdump port 22"
original BPF -> old JIT: original BPF -> internal BPF -> new JIT:
0: push %rbp 0: push %rbp
1: mov %rsp,%rbp 1: mov %rsp,%rbp
4: sub $0x60,%rsp 4: sub $0x228,%rsp
8: mov %rbx,-0x8(%rbp) b: mov %rbx,-0x228(%rbp) // prologue
12: mov %r13,-0x220(%rbp)
19: mov %r14,-0x218(%rbp)
20: mov %r15,-0x210(%rbp)
27: xor %eax,%eax // clear A
c: xor %ebx,%ebx 29: xor %r13,%r13 // clear X
e: mov 0x68(%rdi),%r9d 2c: mov 0x68(%rdi),%r9d
12: sub 0x6c(%rdi),%r9d 30: sub 0x6c(%rdi),%r9d
16: mov 0xd8(%rdi),%r8 34: mov 0xd8(%rdi),%r10
3b: mov %rdi,%rbx
1d: mov $0xc,%esi 3e: mov $0xc,%esi
22: callq 0xffffffffe1021e15 43: callq 0xffffffffe102bd75
27: cmp $0x86dd,%eax 48: cmp $0x86dd,%rax
2c: jne 0x0000000000000069 4f: jne 0x000000000000009a
2e: mov $0x14,%esi 51: mov $0x14,%esi
33: callq 0xffffffffe1021e31 56: callq 0xffffffffe102bd91
38: cmp $0x84,%eax 5b: cmp $0x84,%rax
3d: je 0x0000000000000049 62: je 0x0000000000000074
3f: cmp $0x6,%eax 64: cmp $0x6,%rax
42: je 0x0000000000000049 68: je 0x0000000000000074
44: cmp $0x11,%eax 6a: cmp $0x11,%rax
47: jne 0x00000000000000c6 6e: jne 0x0000000000000117
49: mov $0x36,%esi 74: mov $0x36,%esi
4e: callq 0xffffffffe1021e15 79: callq 0xffffffffe102bd75
53: cmp $0x16,%eax 7e: cmp $0x16,%rax
56: je 0x00000000000000bf 82: je 0x0000000000000110
58: mov $0x38,%esi 88: mov $0x38,%esi
5d: callq 0xffffffffe1021e15 8d: callq 0xffffffffe102bd75
62: cmp $0x16,%eax 92: cmp $0x16,%rax
65: je 0x00000000000000bf 96: je 0x0000000000000110
67: jmp 0x00000000000000c6 98: jmp 0x0000000000000117
69: cmp $0x800,%eax 9a: cmp $0x800,%rax
6e: jne 0x00000000000000c6 a1: jne 0x0000000000000117
70: mov $0x17,%esi a3: mov $0x17,%esi
75: callq 0xffffffffe1021e31 a8: callq 0xffffffffe102bd91
7a: cmp $0x84,%eax ad: cmp $0x84,%rax
7f: je 0x000000000000008b b4: je 0x00000000000000c2
81: cmp $0x6,%eax b6: cmp $0x6,%rax
84: je 0x000000000000008b ba: je 0x00000000000000c2
86: cmp $0x11,%eax bc: cmp $0x11,%rax
89: jne 0x00000000000000c6 c0: jne 0x0000000000000117
8b: mov $0x14,%esi c2: mov $0x14,%esi
90: callq 0xffffffffe1021e15 c7: callq 0xffffffffe102bd75
95: test $0x1fff,%ax cc: test $0x1fff,%rax
99: jne 0x00000000000000c6 d3: jne 0x0000000000000117
d5: mov %rax,%r14
9b: mov $0xe,%esi d8: mov $0xe,%esi
a0: callq 0xffffffffe1021e44 dd: callq 0xffffffffe102bd91 // MSH
e2: and $0xf,%eax
e5: shl $0x2,%eax
e8: mov %rax,%r13
eb: mov %r14,%rax
ee: mov %r13,%rsi
a5: lea 0xe(%rbx),%esi f1: add $0xe,%esi
a8: callq 0xffffffffe1021e0d f4: callq 0xffffffffe102bd6d
ad: cmp $0x16,%eax f9: cmp $0x16,%rax
b0: je 0x00000000000000bf fd: je 0x0000000000000110
ff: mov %r13,%rsi
b2: lea 0x10(%rbx),%esi 102: add $0x10,%esi
b5: callq 0xffffffffe1021e0d 105: callq 0xffffffffe102bd6d
ba: cmp $0x16,%eax 10a: cmp $0x16,%rax
bd: jne 0x00000000000000c6 10e: jne 0x0000000000000117
bf: mov $0xffff,%eax 110: mov $0xffff,%eax
c4: jmp 0x00000000000000c8 115: jmp 0x000000000000011c
c6: xor %eax,%eax 117: mov $0x0,%eax
c8: mov -0x8(%rbp),%rbx 11c: mov -0x228(%rbp),%rbx // epilogue
cc: leaveq 123: mov -0x220(%rbp),%r13
cd: retq 12a: mov -0x218(%rbp),%r14
131: mov -0x210(%rbp),%r15
138: leaveq
139: retq
On fully cached SKBs both JITed functions take 12 nsec to execute.
BPF interpreter executes the program in 30 nsec.
The difference in generated assembler is due to the following:
Old BPF imlements LDX_MSH instruction via sk_load_byte_msh() helper function
inside bpf_jit.S.
New JIT removes the helper and does it explicitly, so ldx_msh cost
is the same for both JITs, but generated code looks longer.
New JIT has 4 registers to save, so prologue/epilogue are larger,
but the cost is within noise on x64.
Old JIT checks whether first insn clears A and if not emits 'xor %eax,%eax'.
New JIT clears %rax unconditionally.
2. old BPF JIT doesn't support ANC_NLATTR, ANC_PAY_OFFSET, ANC_RANDOM
extensions. New JIT supports all BPF extensions.
Performance of such filters improves 2-4 times depending on a filter.
The longer the filter the higher performance gain.
Synthetic benchmarks with many ancillary loads see 20x speedup
which seems to be the maximum gain from JIT
Notes:
. net.core.bpf_jit_enable=2 + tools/net/bpf_jit_disasm is still functional
and can be used to see generated assembler
. there are two jit_compile() functions and code flow for classic filters is:
sk_attach_filter() - load classic BPF
bpf_jit_compile() - try to JIT from classic BPF
sk_convert_filter() - convert classic to internal
bpf_int_jit_compile() - JIT from internal BPF
seccomp and tracing filters will just call bpf_int_jit_compile()
Signed-off-by: Alexei Starovoitov <ast@plumgrid.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
2014-05-14 02:50:46 +00:00
|
|
|
|
bpf, x64: Save bytes for DIV by reducing reg copies
Instead of unconditionally performing push/pop on %rax/%rdx in case of
division/modulo, we can save a few bytes in case of destination register
being either BPF r0 (%rax) or r3 (%rdx) since the result is written in
there anyway.
Also, we do not need to copy the source to %r11 unless the source is either
%rax, %rdx or an immediate.
For example, before the patch:
22: push %rax
23: push %rdx
24: mov %rsi,%r11
27: xor %edx,%edx
29: div %r11
2c: mov %rax,%r11
2f: pop %rdx
30: pop %rax
31: mov %r11,%rax
After:
22: push %rdx
23: xor %edx,%edx
25: div %rsi
28: pop %rdx
Signed-off-by: Jie Meng <jmeng@fb.com>
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Tested-by: Daniel Borkmann <daniel@iogearbox.net>
Acked-by: Daniel Borkmann <daniel@iogearbox.net>
Link: https://lore.kernel.org/bpf/20211002035626.2041910-1-jmeng@fb.com
2021-10-02 03:56:26 +00:00
|
|
|
if (BPF_OP(insn->code) == BPF_MOD &&
|
|
|
|
dst_reg != BPF_REG_3)
|
|
|
|
/* mov dst_reg, rdx */
|
|
|
|
emit_mov_reg(&prog, is64, dst_reg, BPF_REG_3);
|
|
|
|
else if (BPF_OP(insn->code) == BPF_DIV &&
|
|
|
|
dst_reg != BPF_REG_0)
|
|
|
|
/* mov dst_reg, rax */
|
|
|
|
emit_mov_reg(&prog, is64, dst_reg, BPF_REG_0);
|
net: filter: x86: internal BPF JIT
Maps all internal BPF instructions into x86_64 instructions.
This patch replaces original BPF x64 JIT with internal BPF x64 JIT.
sysctl net.core.bpf_jit_enable is reused as on/off switch.
Performance:
1. old BPF JIT and internal BPF JIT generate equivalent x86_64 code.
No performance difference is observed for filters that were JIT-able before
Example assembler code for BPF filter "tcpdump port 22"
original BPF -> old JIT: original BPF -> internal BPF -> new JIT:
0: push %rbp 0: push %rbp
1: mov %rsp,%rbp 1: mov %rsp,%rbp
4: sub $0x60,%rsp 4: sub $0x228,%rsp
8: mov %rbx,-0x8(%rbp) b: mov %rbx,-0x228(%rbp) // prologue
12: mov %r13,-0x220(%rbp)
19: mov %r14,-0x218(%rbp)
20: mov %r15,-0x210(%rbp)
27: xor %eax,%eax // clear A
c: xor %ebx,%ebx 29: xor %r13,%r13 // clear X
e: mov 0x68(%rdi),%r9d 2c: mov 0x68(%rdi),%r9d
12: sub 0x6c(%rdi),%r9d 30: sub 0x6c(%rdi),%r9d
16: mov 0xd8(%rdi),%r8 34: mov 0xd8(%rdi),%r10
3b: mov %rdi,%rbx
1d: mov $0xc,%esi 3e: mov $0xc,%esi
22: callq 0xffffffffe1021e15 43: callq 0xffffffffe102bd75
27: cmp $0x86dd,%eax 48: cmp $0x86dd,%rax
2c: jne 0x0000000000000069 4f: jne 0x000000000000009a
2e: mov $0x14,%esi 51: mov $0x14,%esi
33: callq 0xffffffffe1021e31 56: callq 0xffffffffe102bd91
38: cmp $0x84,%eax 5b: cmp $0x84,%rax
3d: je 0x0000000000000049 62: je 0x0000000000000074
3f: cmp $0x6,%eax 64: cmp $0x6,%rax
42: je 0x0000000000000049 68: je 0x0000000000000074
44: cmp $0x11,%eax 6a: cmp $0x11,%rax
47: jne 0x00000000000000c6 6e: jne 0x0000000000000117
49: mov $0x36,%esi 74: mov $0x36,%esi
4e: callq 0xffffffffe1021e15 79: callq 0xffffffffe102bd75
53: cmp $0x16,%eax 7e: cmp $0x16,%rax
56: je 0x00000000000000bf 82: je 0x0000000000000110
58: mov $0x38,%esi 88: mov $0x38,%esi
5d: callq 0xffffffffe1021e15 8d: callq 0xffffffffe102bd75
62: cmp $0x16,%eax 92: cmp $0x16,%rax
65: je 0x00000000000000bf 96: je 0x0000000000000110
67: jmp 0x00000000000000c6 98: jmp 0x0000000000000117
69: cmp $0x800,%eax 9a: cmp $0x800,%rax
6e: jne 0x00000000000000c6 a1: jne 0x0000000000000117
70: mov $0x17,%esi a3: mov $0x17,%esi
75: callq 0xffffffffe1021e31 a8: callq 0xffffffffe102bd91
7a: cmp $0x84,%eax ad: cmp $0x84,%rax
7f: je 0x000000000000008b b4: je 0x00000000000000c2
81: cmp $0x6,%eax b6: cmp $0x6,%rax
84: je 0x000000000000008b ba: je 0x00000000000000c2
86: cmp $0x11,%eax bc: cmp $0x11,%rax
89: jne 0x00000000000000c6 c0: jne 0x0000000000000117
8b: mov $0x14,%esi c2: mov $0x14,%esi
90: callq 0xffffffffe1021e15 c7: callq 0xffffffffe102bd75
95: test $0x1fff,%ax cc: test $0x1fff,%rax
99: jne 0x00000000000000c6 d3: jne 0x0000000000000117
d5: mov %rax,%r14
9b: mov $0xe,%esi d8: mov $0xe,%esi
a0: callq 0xffffffffe1021e44 dd: callq 0xffffffffe102bd91 // MSH
e2: and $0xf,%eax
e5: shl $0x2,%eax
e8: mov %rax,%r13
eb: mov %r14,%rax
ee: mov %r13,%rsi
a5: lea 0xe(%rbx),%esi f1: add $0xe,%esi
a8: callq 0xffffffffe1021e0d f4: callq 0xffffffffe102bd6d
ad: cmp $0x16,%eax f9: cmp $0x16,%rax
b0: je 0x00000000000000bf fd: je 0x0000000000000110
ff: mov %r13,%rsi
b2: lea 0x10(%rbx),%esi 102: add $0x10,%esi
b5: callq 0xffffffffe1021e0d 105: callq 0xffffffffe102bd6d
ba: cmp $0x16,%eax 10a: cmp $0x16,%rax
bd: jne 0x00000000000000c6 10e: jne 0x0000000000000117
bf: mov $0xffff,%eax 110: mov $0xffff,%eax
c4: jmp 0x00000000000000c8 115: jmp 0x000000000000011c
c6: xor %eax,%eax 117: mov $0x0,%eax
c8: mov -0x8(%rbp),%rbx 11c: mov -0x228(%rbp),%rbx // epilogue
cc: leaveq 123: mov -0x220(%rbp),%r13
cd: retq 12a: mov -0x218(%rbp),%r14
131: mov -0x210(%rbp),%r15
138: leaveq
139: retq
On fully cached SKBs both JITed functions take 12 nsec to execute.
BPF interpreter executes the program in 30 nsec.
The difference in generated assembler is due to the following:
Old BPF imlements LDX_MSH instruction via sk_load_byte_msh() helper function
inside bpf_jit.S.
New JIT removes the helper and does it explicitly, so ldx_msh cost
is the same for both JITs, but generated code looks longer.
New JIT has 4 registers to save, so prologue/epilogue are larger,
but the cost is within noise on x64.
Old JIT checks whether first insn clears A and if not emits 'xor %eax,%eax'.
New JIT clears %rax unconditionally.
2. old BPF JIT doesn't support ANC_NLATTR, ANC_PAY_OFFSET, ANC_RANDOM
extensions. New JIT supports all BPF extensions.
Performance of such filters improves 2-4 times depending on a filter.
The longer the filter the higher performance gain.
Synthetic benchmarks with many ancillary loads see 20x speedup
which seems to be the maximum gain from JIT
Notes:
. net.core.bpf_jit_enable=2 + tools/net/bpf_jit_disasm is still functional
and can be used to see generated assembler
. there are two jit_compile() functions and code flow for classic filters is:
sk_attach_filter() - load classic BPF
bpf_jit_compile() - try to JIT from classic BPF
sk_convert_filter() - convert classic to internal
bpf_int_jit_compile() - JIT from internal BPF
seccomp and tracing filters will just call bpf_int_jit_compile()
Signed-off-by: Alexei Starovoitov <ast@plumgrid.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
2014-05-14 02:50:46 +00:00
|
|
|
|
bpf, x64: Save bytes for DIV by reducing reg copies
Instead of unconditionally performing push/pop on %rax/%rdx in case of
division/modulo, we can save a few bytes in case of destination register
being either BPF r0 (%rax) or r3 (%rdx) since the result is written in
there anyway.
Also, we do not need to copy the source to %r11 unless the source is either
%rax, %rdx or an immediate.
For example, before the patch:
22: push %rax
23: push %rdx
24: mov %rsi,%r11
27: xor %edx,%edx
29: div %r11
2c: mov %rax,%r11
2f: pop %rdx
30: pop %rax
31: mov %r11,%rax
After:
22: push %rdx
23: xor %edx,%edx
25: div %rsi
28: pop %rdx
Signed-off-by: Jie Meng <jmeng@fb.com>
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Tested-by: Daniel Borkmann <daniel@iogearbox.net>
Acked-by: Daniel Borkmann <daniel@iogearbox.net>
Link: https://lore.kernel.org/bpf/20211002035626.2041910-1-jmeng@fb.com
2021-10-02 03:56:26 +00:00
|
|
|
if (dst_reg != BPF_REG_3)
|
|
|
|
EMIT1(0x5A); /* pop rdx */
|
|
|
|
if (dst_reg != BPF_REG_0)
|
|
|
|
EMIT1(0x58); /* pop rax */
|
net: filter: x86: internal BPF JIT
Maps all internal BPF instructions into x86_64 instructions.
This patch replaces original BPF x64 JIT with internal BPF x64 JIT.
sysctl net.core.bpf_jit_enable is reused as on/off switch.
Performance:
1. old BPF JIT and internal BPF JIT generate equivalent x86_64 code.
No performance difference is observed for filters that were JIT-able before
Example assembler code for BPF filter "tcpdump port 22"
original BPF -> old JIT: original BPF -> internal BPF -> new JIT:
0: push %rbp 0: push %rbp
1: mov %rsp,%rbp 1: mov %rsp,%rbp
4: sub $0x60,%rsp 4: sub $0x228,%rsp
8: mov %rbx,-0x8(%rbp) b: mov %rbx,-0x228(%rbp) // prologue
12: mov %r13,-0x220(%rbp)
19: mov %r14,-0x218(%rbp)
20: mov %r15,-0x210(%rbp)
27: xor %eax,%eax // clear A
c: xor %ebx,%ebx 29: xor %r13,%r13 // clear X
e: mov 0x68(%rdi),%r9d 2c: mov 0x68(%rdi),%r9d
12: sub 0x6c(%rdi),%r9d 30: sub 0x6c(%rdi),%r9d
16: mov 0xd8(%rdi),%r8 34: mov 0xd8(%rdi),%r10
3b: mov %rdi,%rbx
1d: mov $0xc,%esi 3e: mov $0xc,%esi
22: callq 0xffffffffe1021e15 43: callq 0xffffffffe102bd75
27: cmp $0x86dd,%eax 48: cmp $0x86dd,%rax
2c: jne 0x0000000000000069 4f: jne 0x000000000000009a
2e: mov $0x14,%esi 51: mov $0x14,%esi
33: callq 0xffffffffe1021e31 56: callq 0xffffffffe102bd91
38: cmp $0x84,%eax 5b: cmp $0x84,%rax
3d: je 0x0000000000000049 62: je 0x0000000000000074
3f: cmp $0x6,%eax 64: cmp $0x6,%rax
42: je 0x0000000000000049 68: je 0x0000000000000074
44: cmp $0x11,%eax 6a: cmp $0x11,%rax
47: jne 0x00000000000000c6 6e: jne 0x0000000000000117
49: mov $0x36,%esi 74: mov $0x36,%esi
4e: callq 0xffffffffe1021e15 79: callq 0xffffffffe102bd75
53: cmp $0x16,%eax 7e: cmp $0x16,%rax
56: je 0x00000000000000bf 82: je 0x0000000000000110
58: mov $0x38,%esi 88: mov $0x38,%esi
5d: callq 0xffffffffe1021e15 8d: callq 0xffffffffe102bd75
62: cmp $0x16,%eax 92: cmp $0x16,%rax
65: je 0x00000000000000bf 96: je 0x0000000000000110
67: jmp 0x00000000000000c6 98: jmp 0x0000000000000117
69: cmp $0x800,%eax 9a: cmp $0x800,%rax
6e: jne 0x00000000000000c6 a1: jne 0x0000000000000117
70: mov $0x17,%esi a3: mov $0x17,%esi
75: callq 0xffffffffe1021e31 a8: callq 0xffffffffe102bd91
7a: cmp $0x84,%eax ad: cmp $0x84,%rax
7f: je 0x000000000000008b b4: je 0x00000000000000c2
81: cmp $0x6,%eax b6: cmp $0x6,%rax
84: je 0x000000000000008b ba: je 0x00000000000000c2
86: cmp $0x11,%eax bc: cmp $0x11,%rax
89: jne 0x00000000000000c6 c0: jne 0x0000000000000117
8b: mov $0x14,%esi c2: mov $0x14,%esi
90: callq 0xffffffffe1021e15 c7: callq 0xffffffffe102bd75
95: test $0x1fff,%ax cc: test $0x1fff,%rax
99: jne 0x00000000000000c6 d3: jne 0x0000000000000117
d5: mov %rax,%r14
9b: mov $0xe,%esi d8: mov $0xe,%esi
a0: callq 0xffffffffe1021e44 dd: callq 0xffffffffe102bd91 // MSH
e2: and $0xf,%eax
e5: shl $0x2,%eax
e8: mov %rax,%r13
eb: mov %r14,%rax
ee: mov %r13,%rsi
a5: lea 0xe(%rbx),%esi f1: add $0xe,%esi
a8: callq 0xffffffffe1021e0d f4: callq 0xffffffffe102bd6d
ad: cmp $0x16,%eax f9: cmp $0x16,%rax
b0: je 0x00000000000000bf fd: je 0x0000000000000110
ff: mov %r13,%rsi
b2: lea 0x10(%rbx),%esi 102: add $0x10,%esi
b5: callq 0xffffffffe1021e0d 105: callq 0xffffffffe102bd6d
ba: cmp $0x16,%eax 10a: cmp $0x16,%rax
bd: jne 0x00000000000000c6 10e: jne 0x0000000000000117
bf: mov $0xffff,%eax 110: mov $0xffff,%eax
c4: jmp 0x00000000000000c8 115: jmp 0x000000000000011c
c6: xor %eax,%eax 117: mov $0x0,%eax
c8: mov -0x8(%rbp),%rbx 11c: mov -0x228(%rbp),%rbx // epilogue
cc: leaveq 123: mov -0x220(%rbp),%r13
cd: retq 12a: mov -0x218(%rbp),%r14
131: mov -0x210(%rbp),%r15
138: leaveq
139: retq
On fully cached SKBs both JITed functions take 12 nsec to execute.
BPF interpreter executes the program in 30 nsec.
The difference in generated assembler is due to the following:
Old BPF imlements LDX_MSH instruction via sk_load_byte_msh() helper function
inside bpf_jit.S.
New JIT removes the helper and does it explicitly, so ldx_msh cost
is the same for both JITs, but generated code looks longer.
New JIT has 4 registers to save, so prologue/epilogue are larger,
but the cost is within noise on x64.
Old JIT checks whether first insn clears A and if not emits 'xor %eax,%eax'.
New JIT clears %rax unconditionally.
2. old BPF JIT doesn't support ANC_NLATTR, ANC_PAY_OFFSET, ANC_RANDOM
extensions. New JIT supports all BPF extensions.
Performance of such filters improves 2-4 times depending on a filter.
The longer the filter the higher performance gain.
Synthetic benchmarks with many ancillary loads see 20x speedup
which seems to be the maximum gain from JIT
Notes:
. net.core.bpf_jit_enable=2 + tools/net/bpf_jit_disasm is still functional
and can be used to see generated assembler
. there are two jit_compile() functions and code flow for classic filters is:
sk_attach_filter() - load classic BPF
bpf_jit_compile() - try to JIT from classic BPF
sk_convert_filter() - convert classic to internal
bpf_int_jit_compile() - JIT from internal BPF
seccomp and tracing filters will just call bpf_int_jit_compile()
Signed-off-by: Alexei Starovoitov <ast@plumgrid.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
2014-05-14 02:50:46 +00:00
|
|
|
break;
|
bpf, x64: Save bytes for DIV by reducing reg copies
Instead of unconditionally performing push/pop on %rax/%rdx in case of
division/modulo, we can save a few bytes in case of destination register
being either BPF r0 (%rax) or r3 (%rdx) since the result is written in
there anyway.
Also, we do not need to copy the source to %r11 unless the source is either
%rax, %rdx or an immediate.
For example, before the patch:
22: push %rax
23: push %rdx
24: mov %rsi,%r11
27: xor %edx,%edx
29: div %r11
2c: mov %rax,%r11
2f: pop %rdx
30: pop %rax
31: mov %r11,%rax
After:
22: push %rdx
23: xor %edx,%edx
25: div %rsi
28: pop %rdx
Signed-off-by: Jie Meng <jmeng@fb.com>
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Tested-by: Daniel Borkmann <daniel@iogearbox.net>
Acked-by: Daniel Borkmann <daniel@iogearbox.net>
Link: https://lore.kernel.org/bpf/20211002035626.2041910-1-jmeng@fb.com
2021-10-02 03:56:26 +00:00
|
|
|
}
|
net: filter: x86: internal BPF JIT
Maps all internal BPF instructions into x86_64 instructions.
This patch replaces original BPF x64 JIT with internal BPF x64 JIT.
sysctl net.core.bpf_jit_enable is reused as on/off switch.
Performance:
1. old BPF JIT and internal BPF JIT generate equivalent x86_64 code.
No performance difference is observed for filters that were JIT-able before
Example assembler code for BPF filter "tcpdump port 22"
original BPF -> old JIT: original BPF -> internal BPF -> new JIT:
0: push %rbp 0: push %rbp
1: mov %rsp,%rbp 1: mov %rsp,%rbp
4: sub $0x60,%rsp 4: sub $0x228,%rsp
8: mov %rbx,-0x8(%rbp) b: mov %rbx,-0x228(%rbp) // prologue
12: mov %r13,-0x220(%rbp)
19: mov %r14,-0x218(%rbp)
20: mov %r15,-0x210(%rbp)
27: xor %eax,%eax // clear A
c: xor %ebx,%ebx 29: xor %r13,%r13 // clear X
e: mov 0x68(%rdi),%r9d 2c: mov 0x68(%rdi),%r9d
12: sub 0x6c(%rdi),%r9d 30: sub 0x6c(%rdi),%r9d
16: mov 0xd8(%rdi),%r8 34: mov 0xd8(%rdi),%r10
3b: mov %rdi,%rbx
1d: mov $0xc,%esi 3e: mov $0xc,%esi
22: callq 0xffffffffe1021e15 43: callq 0xffffffffe102bd75
27: cmp $0x86dd,%eax 48: cmp $0x86dd,%rax
2c: jne 0x0000000000000069 4f: jne 0x000000000000009a
2e: mov $0x14,%esi 51: mov $0x14,%esi
33: callq 0xffffffffe1021e31 56: callq 0xffffffffe102bd91
38: cmp $0x84,%eax 5b: cmp $0x84,%rax
3d: je 0x0000000000000049 62: je 0x0000000000000074
3f: cmp $0x6,%eax 64: cmp $0x6,%rax
42: je 0x0000000000000049 68: je 0x0000000000000074
44: cmp $0x11,%eax 6a: cmp $0x11,%rax
47: jne 0x00000000000000c6 6e: jne 0x0000000000000117
49: mov $0x36,%esi 74: mov $0x36,%esi
4e: callq 0xffffffffe1021e15 79: callq 0xffffffffe102bd75
53: cmp $0x16,%eax 7e: cmp $0x16,%rax
56: je 0x00000000000000bf 82: je 0x0000000000000110
58: mov $0x38,%esi 88: mov $0x38,%esi
5d: callq 0xffffffffe1021e15 8d: callq 0xffffffffe102bd75
62: cmp $0x16,%eax 92: cmp $0x16,%rax
65: je 0x00000000000000bf 96: je 0x0000000000000110
67: jmp 0x00000000000000c6 98: jmp 0x0000000000000117
69: cmp $0x800,%eax 9a: cmp $0x800,%rax
6e: jne 0x00000000000000c6 a1: jne 0x0000000000000117
70: mov $0x17,%esi a3: mov $0x17,%esi
75: callq 0xffffffffe1021e31 a8: callq 0xffffffffe102bd91
7a: cmp $0x84,%eax ad: cmp $0x84,%rax
7f: je 0x000000000000008b b4: je 0x00000000000000c2
81: cmp $0x6,%eax b6: cmp $0x6,%rax
84: je 0x000000000000008b ba: je 0x00000000000000c2
86: cmp $0x11,%eax bc: cmp $0x11,%rax
89: jne 0x00000000000000c6 c0: jne 0x0000000000000117
8b: mov $0x14,%esi c2: mov $0x14,%esi
90: callq 0xffffffffe1021e15 c7: callq 0xffffffffe102bd75
95: test $0x1fff,%ax cc: test $0x1fff,%rax
99: jne 0x00000000000000c6 d3: jne 0x0000000000000117
d5: mov %rax,%r14
9b: mov $0xe,%esi d8: mov $0xe,%esi
a0: callq 0xffffffffe1021e44 dd: callq 0xffffffffe102bd91 // MSH
e2: and $0xf,%eax
e5: shl $0x2,%eax
e8: mov %rax,%r13
eb: mov %r14,%rax
ee: mov %r13,%rsi
a5: lea 0xe(%rbx),%esi f1: add $0xe,%esi
a8: callq 0xffffffffe1021e0d f4: callq 0xffffffffe102bd6d
ad: cmp $0x16,%eax f9: cmp $0x16,%rax
b0: je 0x00000000000000bf fd: je 0x0000000000000110
ff: mov %r13,%rsi
b2: lea 0x10(%rbx),%esi 102: add $0x10,%esi
b5: callq 0xffffffffe1021e0d 105: callq 0xffffffffe102bd6d
ba: cmp $0x16,%eax 10a: cmp $0x16,%rax
bd: jne 0x00000000000000c6 10e: jne 0x0000000000000117
bf: mov $0xffff,%eax 110: mov $0xffff,%eax
c4: jmp 0x00000000000000c8 115: jmp 0x000000000000011c
c6: xor %eax,%eax 117: mov $0x0,%eax
c8: mov -0x8(%rbp),%rbx 11c: mov -0x228(%rbp),%rbx // epilogue
cc: leaveq 123: mov -0x220(%rbp),%r13
cd: retq 12a: mov -0x218(%rbp),%r14
131: mov -0x210(%rbp),%r15
138: leaveq
139: retq
On fully cached SKBs both JITed functions take 12 nsec to execute.
BPF interpreter executes the program in 30 nsec.
The difference in generated assembler is due to the following:
Old BPF imlements LDX_MSH instruction via sk_load_byte_msh() helper function
inside bpf_jit.S.
New JIT removes the helper and does it explicitly, so ldx_msh cost
is the same for both JITs, but generated code looks longer.
New JIT has 4 registers to save, so prologue/epilogue are larger,
but the cost is within noise on x64.
Old JIT checks whether first insn clears A and if not emits 'xor %eax,%eax'.
New JIT clears %rax unconditionally.
2. old BPF JIT doesn't support ANC_NLATTR, ANC_PAY_OFFSET, ANC_RANDOM
extensions. New JIT supports all BPF extensions.
Performance of such filters improves 2-4 times depending on a filter.
The longer the filter the higher performance gain.
Synthetic benchmarks with many ancillary loads see 20x speedup
which seems to be the maximum gain from JIT
Notes:
. net.core.bpf_jit_enable=2 + tools/net/bpf_jit_disasm is still functional
and can be used to see generated assembler
. there are two jit_compile() functions and code flow for classic filters is:
sk_attach_filter() - load classic BPF
bpf_jit_compile() - try to JIT from classic BPF
sk_convert_filter() - convert classic to internal
bpf_int_jit_compile() - JIT from internal BPF
seccomp and tracing filters will just call bpf_int_jit_compile()
Signed-off-by: Alexei Starovoitov <ast@plumgrid.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
2014-05-14 02:50:46 +00:00
|
|
|
|
|
|
|
case BPF_ALU | BPF_MUL | BPF_K:
|
|
|
|
case BPF_ALU64 | BPF_MUL | BPF_K:
|
2021-10-06 19:41:35 +00:00
|
|
|
maybe_emit_mod(&prog, dst_reg, dst_reg,
|
|
|
|
BPF_CLASS(insn->code) == BPF_ALU64);
|
net: filter: x86: internal BPF JIT
Maps all internal BPF instructions into x86_64 instructions.
This patch replaces original BPF x64 JIT with internal BPF x64 JIT.
sysctl net.core.bpf_jit_enable is reused as on/off switch.
Performance:
1. old BPF JIT and internal BPF JIT generate equivalent x86_64 code.
No performance difference is observed for filters that were JIT-able before
Example assembler code for BPF filter "tcpdump port 22"
original BPF -> old JIT: original BPF -> internal BPF -> new JIT:
0: push %rbp 0: push %rbp
1: mov %rsp,%rbp 1: mov %rsp,%rbp
4: sub $0x60,%rsp 4: sub $0x228,%rsp
8: mov %rbx,-0x8(%rbp) b: mov %rbx,-0x228(%rbp) // prologue
12: mov %r13,-0x220(%rbp)
19: mov %r14,-0x218(%rbp)
20: mov %r15,-0x210(%rbp)
27: xor %eax,%eax // clear A
c: xor %ebx,%ebx 29: xor %r13,%r13 // clear X
e: mov 0x68(%rdi),%r9d 2c: mov 0x68(%rdi),%r9d
12: sub 0x6c(%rdi),%r9d 30: sub 0x6c(%rdi),%r9d
16: mov 0xd8(%rdi),%r8 34: mov 0xd8(%rdi),%r10
3b: mov %rdi,%rbx
1d: mov $0xc,%esi 3e: mov $0xc,%esi
22: callq 0xffffffffe1021e15 43: callq 0xffffffffe102bd75
27: cmp $0x86dd,%eax 48: cmp $0x86dd,%rax
2c: jne 0x0000000000000069 4f: jne 0x000000000000009a
2e: mov $0x14,%esi 51: mov $0x14,%esi
33: callq 0xffffffffe1021e31 56: callq 0xffffffffe102bd91
38: cmp $0x84,%eax 5b: cmp $0x84,%rax
3d: je 0x0000000000000049 62: je 0x0000000000000074
3f: cmp $0x6,%eax 64: cmp $0x6,%rax
42: je 0x0000000000000049 68: je 0x0000000000000074
44: cmp $0x11,%eax 6a: cmp $0x11,%rax
47: jne 0x00000000000000c6 6e: jne 0x0000000000000117
49: mov $0x36,%esi 74: mov $0x36,%esi
4e: callq 0xffffffffe1021e15 79: callq 0xffffffffe102bd75
53: cmp $0x16,%eax 7e: cmp $0x16,%rax
56: je 0x00000000000000bf 82: je 0x0000000000000110
58: mov $0x38,%esi 88: mov $0x38,%esi
5d: callq 0xffffffffe1021e15 8d: callq 0xffffffffe102bd75
62: cmp $0x16,%eax 92: cmp $0x16,%rax
65: je 0x00000000000000bf 96: je 0x0000000000000110
67: jmp 0x00000000000000c6 98: jmp 0x0000000000000117
69: cmp $0x800,%eax 9a: cmp $0x800,%rax
6e: jne 0x00000000000000c6 a1: jne 0x0000000000000117
70: mov $0x17,%esi a3: mov $0x17,%esi
75: callq 0xffffffffe1021e31 a8: callq 0xffffffffe102bd91
7a: cmp $0x84,%eax ad: cmp $0x84,%rax
7f: je 0x000000000000008b b4: je 0x00000000000000c2
81: cmp $0x6,%eax b6: cmp $0x6,%rax
84: je 0x000000000000008b ba: je 0x00000000000000c2
86: cmp $0x11,%eax bc: cmp $0x11,%rax
89: jne 0x00000000000000c6 c0: jne 0x0000000000000117
8b: mov $0x14,%esi c2: mov $0x14,%esi
90: callq 0xffffffffe1021e15 c7: callq 0xffffffffe102bd75
95: test $0x1fff,%ax cc: test $0x1fff,%rax
99: jne 0x00000000000000c6 d3: jne 0x0000000000000117
d5: mov %rax,%r14
9b: mov $0xe,%esi d8: mov $0xe,%esi
a0: callq 0xffffffffe1021e44 dd: callq 0xffffffffe102bd91 // MSH
e2: and $0xf,%eax
e5: shl $0x2,%eax
e8: mov %rax,%r13
eb: mov %r14,%rax
ee: mov %r13,%rsi
a5: lea 0xe(%rbx),%esi f1: add $0xe,%esi
a8: callq 0xffffffffe1021e0d f4: callq 0xffffffffe102bd6d
ad: cmp $0x16,%eax f9: cmp $0x16,%rax
b0: je 0x00000000000000bf fd: je 0x0000000000000110
ff: mov %r13,%rsi
b2: lea 0x10(%rbx),%esi 102: add $0x10,%esi
b5: callq 0xffffffffe1021e0d 105: callq 0xffffffffe102bd6d
ba: cmp $0x16,%eax 10a: cmp $0x16,%rax
bd: jne 0x00000000000000c6 10e: jne 0x0000000000000117
bf: mov $0xffff,%eax 110: mov $0xffff,%eax
c4: jmp 0x00000000000000c8 115: jmp 0x000000000000011c
c6: xor %eax,%eax 117: mov $0x0,%eax
c8: mov -0x8(%rbp),%rbx 11c: mov -0x228(%rbp),%rbx // epilogue
cc: leaveq 123: mov -0x220(%rbp),%r13
cd: retq 12a: mov -0x218(%rbp),%r14
131: mov -0x210(%rbp),%r15
138: leaveq
139: retq
On fully cached SKBs both JITed functions take 12 nsec to execute.
BPF interpreter executes the program in 30 nsec.
The difference in generated assembler is due to the following:
Old BPF imlements LDX_MSH instruction via sk_load_byte_msh() helper function
inside bpf_jit.S.
New JIT removes the helper and does it explicitly, so ldx_msh cost
is the same for both JITs, but generated code looks longer.
New JIT has 4 registers to save, so prologue/epilogue are larger,
but the cost is within noise on x64.
Old JIT checks whether first insn clears A and if not emits 'xor %eax,%eax'.
New JIT clears %rax unconditionally.
2. old BPF JIT doesn't support ANC_NLATTR, ANC_PAY_OFFSET, ANC_RANDOM
extensions. New JIT supports all BPF extensions.
Performance of such filters improves 2-4 times depending on a filter.
The longer the filter the higher performance gain.
Synthetic benchmarks with many ancillary loads see 20x speedup
which seems to be the maximum gain from JIT
Notes:
. net.core.bpf_jit_enable=2 + tools/net/bpf_jit_disasm is still functional
and can be used to see generated assembler
. there are two jit_compile() functions and code flow for classic filters is:
sk_attach_filter() - load classic BPF
bpf_jit_compile() - try to JIT from classic BPF
sk_convert_filter() - convert classic to internal
bpf_int_jit_compile() - JIT from internal BPF
seccomp and tracing filters will just call bpf_int_jit_compile()
Signed-off-by: Alexei Starovoitov <ast@plumgrid.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
2014-05-14 02:50:46 +00:00
|
|
|
|
2021-09-13 21:13:37 +00:00
|
|
|
if (is_imm8(imm32))
|
|
|
|
/* imul dst_reg, dst_reg, imm8 */
|
|
|
|
EMIT3(0x6B, add_2reg(0xC0, dst_reg, dst_reg),
|
|
|
|
imm32);
|
net: filter: x86: internal BPF JIT
Maps all internal BPF instructions into x86_64 instructions.
This patch replaces original BPF x64 JIT with internal BPF x64 JIT.
sysctl net.core.bpf_jit_enable is reused as on/off switch.
Performance:
1. old BPF JIT and internal BPF JIT generate equivalent x86_64 code.
No performance difference is observed for filters that were JIT-able before
Example assembler code for BPF filter "tcpdump port 22"
original BPF -> old JIT: original BPF -> internal BPF -> new JIT:
0: push %rbp 0: push %rbp
1: mov %rsp,%rbp 1: mov %rsp,%rbp
4: sub $0x60,%rsp 4: sub $0x228,%rsp
8: mov %rbx,-0x8(%rbp) b: mov %rbx,-0x228(%rbp) // prologue
12: mov %r13,-0x220(%rbp)
19: mov %r14,-0x218(%rbp)
20: mov %r15,-0x210(%rbp)
27: xor %eax,%eax // clear A
c: xor %ebx,%ebx 29: xor %r13,%r13 // clear X
e: mov 0x68(%rdi),%r9d 2c: mov 0x68(%rdi),%r9d
12: sub 0x6c(%rdi),%r9d 30: sub 0x6c(%rdi),%r9d
16: mov 0xd8(%rdi),%r8 34: mov 0xd8(%rdi),%r10
3b: mov %rdi,%rbx
1d: mov $0xc,%esi 3e: mov $0xc,%esi
22: callq 0xffffffffe1021e15 43: callq 0xffffffffe102bd75
27: cmp $0x86dd,%eax 48: cmp $0x86dd,%rax
2c: jne 0x0000000000000069 4f: jne 0x000000000000009a
2e: mov $0x14,%esi 51: mov $0x14,%esi
33: callq 0xffffffffe1021e31 56: callq 0xffffffffe102bd91
38: cmp $0x84,%eax 5b: cmp $0x84,%rax
3d: je 0x0000000000000049 62: je 0x0000000000000074
3f: cmp $0x6,%eax 64: cmp $0x6,%rax
42: je 0x0000000000000049 68: je 0x0000000000000074
44: cmp $0x11,%eax 6a: cmp $0x11,%rax
47: jne 0x00000000000000c6 6e: jne 0x0000000000000117
49: mov $0x36,%esi 74: mov $0x36,%esi
4e: callq 0xffffffffe1021e15 79: callq 0xffffffffe102bd75
53: cmp $0x16,%eax 7e: cmp $0x16,%rax
56: je 0x00000000000000bf 82: je 0x0000000000000110
58: mov $0x38,%esi 88: mov $0x38,%esi
5d: callq 0xffffffffe1021e15 8d: callq 0xffffffffe102bd75
62: cmp $0x16,%eax 92: cmp $0x16,%rax
65: je 0x00000000000000bf 96: je 0x0000000000000110
67: jmp 0x00000000000000c6 98: jmp 0x0000000000000117
69: cmp $0x800,%eax 9a: cmp $0x800,%rax
6e: jne 0x00000000000000c6 a1: jne 0x0000000000000117
70: mov $0x17,%esi a3: mov $0x17,%esi
75: callq 0xffffffffe1021e31 a8: callq 0xffffffffe102bd91
7a: cmp $0x84,%eax ad: cmp $0x84,%rax
7f: je 0x000000000000008b b4: je 0x00000000000000c2
81: cmp $0x6,%eax b6: cmp $0x6,%rax
84: je 0x000000000000008b ba: je 0x00000000000000c2
86: cmp $0x11,%eax bc: cmp $0x11,%rax
89: jne 0x00000000000000c6 c0: jne 0x0000000000000117
8b: mov $0x14,%esi c2: mov $0x14,%esi
90: callq 0xffffffffe1021e15 c7: callq 0xffffffffe102bd75
95: test $0x1fff,%ax cc: test $0x1fff,%rax
99: jne 0x00000000000000c6 d3: jne 0x0000000000000117
d5: mov %rax,%r14
9b: mov $0xe,%esi d8: mov $0xe,%esi
a0: callq 0xffffffffe1021e44 dd: callq 0xffffffffe102bd91 // MSH
e2: and $0xf,%eax
e5: shl $0x2,%eax
e8: mov %rax,%r13
eb: mov %r14,%rax
ee: mov %r13,%rsi
a5: lea 0xe(%rbx),%esi f1: add $0xe,%esi
a8: callq 0xffffffffe1021e0d f4: callq 0xffffffffe102bd6d
ad: cmp $0x16,%eax f9: cmp $0x16,%rax
b0: je 0x00000000000000bf fd: je 0x0000000000000110
ff: mov %r13,%rsi
b2: lea 0x10(%rbx),%esi 102: add $0x10,%esi
b5: callq 0xffffffffe1021e0d 105: callq 0xffffffffe102bd6d
ba: cmp $0x16,%eax 10a: cmp $0x16,%rax
bd: jne 0x00000000000000c6 10e: jne 0x0000000000000117
bf: mov $0xffff,%eax 110: mov $0xffff,%eax
c4: jmp 0x00000000000000c8 115: jmp 0x000000000000011c
c6: xor %eax,%eax 117: mov $0x0,%eax
c8: mov -0x8(%rbp),%rbx 11c: mov -0x228(%rbp),%rbx // epilogue
cc: leaveq 123: mov -0x220(%rbp),%r13
cd: retq 12a: mov -0x218(%rbp),%r14
131: mov -0x210(%rbp),%r15
138: leaveq
139: retq
On fully cached SKBs both JITed functions take 12 nsec to execute.
BPF interpreter executes the program in 30 nsec.
The difference in generated assembler is due to the following:
Old BPF imlements LDX_MSH instruction via sk_load_byte_msh() helper function
inside bpf_jit.S.
New JIT removes the helper and does it explicitly, so ldx_msh cost
is the same for both JITs, but generated code looks longer.
New JIT has 4 registers to save, so prologue/epilogue are larger,
but the cost is within noise on x64.
Old JIT checks whether first insn clears A and if not emits 'xor %eax,%eax'.
New JIT clears %rax unconditionally.
2. old BPF JIT doesn't support ANC_NLATTR, ANC_PAY_OFFSET, ANC_RANDOM
extensions. New JIT supports all BPF extensions.
Performance of such filters improves 2-4 times depending on a filter.
The longer the filter the higher performance gain.
Synthetic benchmarks with many ancillary loads see 20x speedup
which seems to be the maximum gain from JIT
Notes:
. net.core.bpf_jit_enable=2 + tools/net/bpf_jit_disasm is still functional
and can be used to see generated assembler
. there are two jit_compile() functions and code flow for classic filters is:
sk_attach_filter() - load classic BPF
bpf_jit_compile() - try to JIT from classic BPF
sk_convert_filter() - convert classic to internal
bpf_int_jit_compile() - JIT from internal BPF
seccomp and tracing filters will just call bpf_int_jit_compile()
Signed-off-by: Alexei Starovoitov <ast@plumgrid.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
2014-05-14 02:50:46 +00:00
|
|
|
else
|
2021-09-13 21:13:37 +00:00
|
|
|
/* imul dst_reg, dst_reg, imm32 */
|
|
|
|
EMIT2_off32(0x69,
|
|
|
|
add_2reg(0xC0, dst_reg, dst_reg),
|
|
|
|
imm32);
|
|
|
|
break;
|
net: filter: x86: internal BPF JIT
Maps all internal BPF instructions into x86_64 instructions.
This patch replaces original BPF x64 JIT with internal BPF x64 JIT.
sysctl net.core.bpf_jit_enable is reused as on/off switch.
Performance:
1. old BPF JIT and internal BPF JIT generate equivalent x86_64 code.
No performance difference is observed for filters that were JIT-able before
Example assembler code for BPF filter "tcpdump port 22"
original BPF -> old JIT: original BPF -> internal BPF -> new JIT:
0: push %rbp 0: push %rbp
1: mov %rsp,%rbp 1: mov %rsp,%rbp
4: sub $0x60,%rsp 4: sub $0x228,%rsp
8: mov %rbx,-0x8(%rbp) b: mov %rbx,-0x228(%rbp) // prologue
12: mov %r13,-0x220(%rbp)
19: mov %r14,-0x218(%rbp)
20: mov %r15,-0x210(%rbp)
27: xor %eax,%eax // clear A
c: xor %ebx,%ebx 29: xor %r13,%r13 // clear X
e: mov 0x68(%rdi),%r9d 2c: mov 0x68(%rdi),%r9d
12: sub 0x6c(%rdi),%r9d 30: sub 0x6c(%rdi),%r9d
16: mov 0xd8(%rdi),%r8 34: mov 0xd8(%rdi),%r10
3b: mov %rdi,%rbx
1d: mov $0xc,%esi 3e: mov $0xc,%esi
22: callq 0xffffffffe1021e15 43: callq 0xffffffffe102bd75
27: cmp $0x86dd,%eax 48: cmp $0x86dd,%rax
2c: jne 0x0000000000000069 4f: jne 0x000000000000009a
2e: mov $0x14,%esi 51: mov $0x14,%esi
33: callq 0xffffffffe1021e31 56: callq 0xffffffffe102bd91
38: cmp $0x84,%eax 5b: cmp $0x84,%rax
3d: je 0x0000000000000049 62: je 0x0000000000000074
3f: cmp $0x6,%eax 64: cmp $0x6,%rax
42: je 0x0000000000000049 68: je 0x0000000000000074
44: cmp $0x11,%eax 6a: cmp $0x11,%rax
47: jne 0x00000000000000c6 6e: jne 0x0000000000000117
49: mov $0x36,%esi 74: mov $0x36,%esi
4e: callq 0xffffffffe1021e15 79: callq 0xffffffffe102bd75
53: cmp $0x16,%eax 7e: cmp $0x16,%rax
56: je 0x00000000000000bf 82: je 0x0000000000000110
58: mov $0x38,%esi 88: mov $0x38,%esi
5d: callq 0xffffffffe1021e15 8d: callq 0xffffffffe102bd75
62: cmp $0x16,%eax 92: cmp $0x16,%rax
65: je 0x00000000000000bf 96: je 0x0000000000000110
67: jmp 0x00000000000000c6 98: jmp 0x0000000000000117
69: cmp $0x800,%eax 9a: cmp $0x800,%rax
6e: jne 0x00000000000000c6 a1: jne 0x0000000000000117
70: mov $0x17,%esi a3: mov $0x17,%esi
75: callq 0xffffffffe1021e31 a8: callq 0xffffffffe102bd91
7a: cmp $0x84,%eax ad: cmp $0x84,%rax
7f: je 0x000000000000008b b4: je 0x00000000000000c2
81: cmp $0x6,%eax b6: cmp $0x6,%rax
84: je 0x000000000000008b ba: je 0x00000000000000c2
86: cmp $0x11,%eax bc: cmp $0x11,%rax
89: jne 0x00000000000000c6 c0: jne 0x0000000000000117
8b: mov $0x14,%esi c2: mov $0x14,%esi
90: callq 0xffffffffe1021e15 c7: callq 0xffffffffe102bd75
95: test $0x1fff,%ax cc: test $0x1fff,%rax
99: jne 0x00000000000000c6 d3: jne 0x0000000000000117
d5: mov %rax,%r14
9b: mov $0xe,%esi d8: mov $0xe,%esi
a0: callq 0xffffffffe1021e44 dd: callq 0xffffffffe102bd91 // MSH
e2: and $0xf,%eax
e5: shl $0x2,%eax
e8: mov %rax,%r13
eb: mov %r14,%rax
ee: mov %r13,%rsi
a5: lea 0xe(%rbx),%esi f1: add $0xe,%esi
a8: callq 0xffffffffe1021e0d f4: callq 0xffffffffe102bd6d
ad: cmp $0x16,%eax f9: cmp $0x16,%rax
b0: je 0x00000000000000bf fd: je 0x0000000000000110
ff: mov %r13,%rsi
b2: lea 0x10(%rbx),%esi 102: add $0x10,%esi
b5: callq 0xffffffffe1021e0d 105: callq 0xffffffffe102bd6d
ba: cmp $0x16,%eax 10a: cmp $0x16,%rax
bd: jne 0x00000000000000c6 10e: jne 0x0000000000000117
bf: mov $0xffff,%eax 110: mov $0xffff,%eax
c4: jmp 0x00000000000000c8 115: jmp 0x000000000000011c
c6: xor %eax,%eax 117: mov $0x0,%eax
c8: mov -0x8(%rbp),%rbx 11c: mov -0x228(%rbp),%rbx // epilogue
cc: leaveq 123: mov -0x220(%rbp),%r13
cd: retq 12a: mov -0x218(%rbp),%r14
131: mov -0x210(%rbp),%r15
138: leaveq
139: retq
On fully cached SKBs both JITed functions take 12 nsec to execute.
BPF interpreter executes the program in 30 nsec.
The difference in generated assembler is due to the following:
Old BPF imlements LDX_MSH instruction via sk_load_byte_msh() helper function
inside bpf_jit.S.
New JIT removes the helper and does it explicitly, so ldx_msh cost
is the same for both JITs, but generated code looks longer.
New JIT has 4 registers to save, so prologue/epilogue are larger,
but the cost is within noise on x64.
Old JIT checks whether first insn clears A and if not emits 'xor %eax,%eax'.
New JIT clears %rax unconditionally.
2. old BPF JIT doesn't support ANC_NLATTR, ANC_PAY_OFFSET, ANC_RANDOM
extensions. New JIT supports all BPF extensions.
Performance of such filters improves 2-4 times depending on a filter.
The longer the filter the higher performance gain.
Synthetic benchmarks with many ancillary loads see 20x speedup
which seems to be the maximum gain from JIT
Notes:
. net.core.bpf_jit_enable=2 + tools/net/bpf_jit_disasm is still functional
and can be used to see generated assembler
. there are two jit_compile() functions and code flow for classic filters is:
sk_attach_filter() - load classic BPF
bpf_jit_compile() - try to JIT from classic BPF
sk_convert_filter() - convert classic to internal
bpf_int_jit_compile() - JIT from internal BPF
seccomp and tracing filters will just call bpf_int_jit_compile()
Signed-off-by: Alexei Starovoitov <ast@plumgrid.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
2014-05-14 02:50:46 +00:00
|
|
|
|
2021-09-13 21:13:37 +00:00
|
|
|
case BPF_ALU | BPF_MUL | BPF_X:
|
|
|
|
case BPF_ALU64 | BPF_MUL | BPF_X:
|
2021-10-06 19:41:35 +00:00
|
|
|
maybe_emit_mod(&prog, src_reg, dst_reg,
|
|
|
|
BPF_CLASS(insn->code) == BPF_ALU64);
|
net: filter: x86: internal BPF JIT
Maps all internal BPF instructions into x86_64 instructions.
This patch replaces original BPF x64 JIT with internal BPF x64 JIT.
sysctl net.core.bpf_jit_enable is reused as on/off switch.
Performance:
1. old BPF JIT and internal BPF JIT generate equivalent x86_64 code.
No performance difference is observed for filters that were JIT-able before
Example assembler code for BPF filter "tcpdump port 22"
original BPF -> old JIT: original BPF -> internal BPF -> new JIT:
0: push %rbp 0: push %rbp
1: mov %rsp,%rbp 1: mov %rsp,%rbp
4: sub $0x60,%rsp 4: sub $0x228,%rsp
8: mov %rbx,-0x8(%rbp) b: mov %rbx,-0x228(%rbp) // prologue
12: mov %r13,-0x220(%rbp)
19: mov %r14,-0x218(%rbp)
20: mov %r15,-0x210(%rbp)
27: xor %eax,%eax // clear A
c: xor %ebx,%ebx 29: xor %r13,%r13 // clear X
e: mov 0x68(%rdi),%r9d 2c: mov 0x68(%rdi),%r9d
12: sub 0x6c(%rdi),%r9d 30: sub 0x6c(%rdi),%r9d
16: mov 0xd8(%rdi),%r8 34: mov 0xd8(%rdi),%r10
3b: mov %rdi,%rbx
1d: mov $0xc,%esi 3e: mov $0xc,%esi
22: callq 0xffffffffe1021e15 43: callq 0xffffffffe102bd75
27: cmp $0x86dd,%eax 48: cmp $0x86dd,%rax
2c: jne 0x0000000000000069 4f: jne 0x000000000000009a
2e: mov $0x14,%esi 51: mov $0x14,%esi
33: callq 0xffffffffe1021e31 56: callq 0xffffffffe102bd91
38: cmp $0x84,%eax 5b: cmp $0x84,%rax
3d: je 0x0000000000000049 62: je 0x0000000000000074
3f: cmp $0x6,%eax 64: cmp $0x6,%rax
42: je 0x0000000000000049 68: je 0x0000000000000074
44: cmp $0x11,%eax 6a: cmp $0x11,%rax
47: jne 0x00000000000000c6 6e: jne 0x0000000000000117
49: mov $0x36,%esi 74: mov $0x36,%esi
4e: callq 0xffffffffe1021e15 79: callq 0xffffffffe102bd75
53: cmp $0x16,%eax 7e: cmp $0x16,%rax
56: je 0x00000000000000bf 82: je 0x0000000000000110
58: mov $0x38,%esi 88: mov $0x38,%esi
5d: callq 0xffffffffe1021e15 8d: callq 0xffffffffe102bd75
62: cmp $0x16,%eax 92: cmp $0x16,%rax
65: je 0x00000000000000bf 96: je 0x0000000000000110
67: jmp 0x00000000000000c6 98: jmp 0x0000000000000117
69: cmp $0x800,%eax 9a: cmp $0x800,%rax
6e: jne 0x00000000000000c6 a1: jne 0x0000000000000117
70: mov $0x17,%esi a3: mov $0x17,%esi
75: callq 0xffffffffe1021e31 a8: callq 0xffffffffe102bd91
7a: cmp $0x84,%eax ad: cmp $0x84,%rax
7f: je 0x000000000000008b b4: je 0x00000000000000c2
81: cmp $0x6,%eax b6: cmp $0x6,%rax
84: je 0x000000000000008b ba: je 0x00000000000000c2
86: cmp $0x11,%eax bc: cmp $0x11,%rax
89: jne 0x00000000000000c6 c0: jne 0x0000000000000117
8b: mov $0x14,%esi c2: mov $0x14,%esi
90: callq 0xffffffffe1021e15 c7: callq 0xffffffffe102bd75
95: test $0x1fff,%ax cc: test $0x1fff,%rax
99: jne 0x00000000000000c6 d3: jne 0x0000000000000117
d5: mov %rax,%r14
9b: mov $0xe,%esi d8: mov $0xe,%esi
a0: callq 0xffffffffe1021e44 dd: callq 0xffffffffe102bd91 // MSH
e2: and $0xf,%eax
e5: shl $0x2,%eax
e8: mov %rax,%r13
eb: mov %r14,%rax
ee: mov %r13,%rsi
a5: lea 0xe(%rbx),%esi f1: add $0xe,%esi
a8: callq 0xffffffffe1021e0d f4: callq 0xffffffffe102bd6d
ad: cmp $0x16,%eax f9: cmp $0x16,%rax
b0: je 0x00000000000000bf fd: je 0x0000000000000110
ff: mov %r13,%rsi
b2: lea 0x10(%rbx),%esi 102: add $0x10,%esi
b5: callq 0xffffffffe1021e0d 105: callq 0xffffffffe102bd6d
ba: cmp $0x16,%eax 10a: cmp $0x16,%rax
bd: jne 0x00000000000000c6 10e: jne 0x0000000000000117
bf: mov $0xffff,%eax 110: mov $0xffff,%eax
c4: jmp 0x00000000000000c8 115: jmp 0x000000000000011c
c6: xor %eax,%eax 117: mov $0x0,%eax
c8: mov -0x8(%rbp),%rbx 11c: mov -0x228(%rbp),%rbx // epilogue
cc: leaveq 123: mov -0x220(%rbp),%r13
cd: retq 12a: mov -0x218(%rbp),%r14
131: mov -0x210(%rbp),%r15
138: leaveq
139: retq
On fully cached SKBs both JITed functions take 12 nsec to execute.
BPF interpreter executes the program in 30 nsec.
The difference in generated assembler is due to the following:
Old BPF imlements LDX_MSH instruction via sk_load_byte_msh() helper function
inside bpf_jit.S.
New JIT removes the helper and does it explicitly, so ldx_msh cost
is the same for both JITs, but generated code looks longer.
New JIT has 4 registers to save, so prologue/epilogue are larger,
but the cost is within noise on x64.
Old JIT checks whether first insn clears A and if not emits 'xor %eax,%eax'.
New JIT clears %rax unconditionally.
2. old BPF JIT doesn't support ANC_NLATTR, ANC_PAY_OFFSET, ANC_RANDOM
extensions. New JIT supports all BPF extensions.
Performance of such filters improves 2-4 times depending on a filter.
The longer the filter the higher performance gain.
Synthetic benchmarks with many ancillary loads see 20x speedup
which seems to be the maximum gain from JIT
Notes:
. net.core.bpf_jit_enable=2 + tools/net/bpf_jit_disasm is still functional
and can be used to see generated assembler
. there are two jit_compile() functions and code flow for classic filters is:
sk_attach_filter() - load classic BPF
bpf_jit_compile() - try to JIT from classic BPF
sk_convert_filter() - convert classic to internal
bpf_int_jit_compile() - JIT from internal BPF
seccomp and tracing filters will just call bpf_int_jit_compile()
Signed-off-by: Alexei Starovoitov <ast@plumgrid.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
2014-05-14 02:50:46 +00:00
|
|
|
|
2021-09-13 21:13:37 +00:00
|
|
|
/* imul dst_reg, src_reg */
|
|
|
|
EMIT3(0x0F, 0xAF, add_2reg(0xC0, src_reg, dst_reg));
|
net: filter: x86: internal BPF JIT
Maps all internal BPF instructions into x86_64 instructions.
This patch replaces original BPF x64 JIT with internal BPF x64 JIT.
sysctl net.core.bpf_jit_enable is reused as on/off switch.
Performance:
1. old BPF JIT and internal BPF JIT generate equivalent x86_64 code.
No performance difference is observed for filters that were JIT-able before
Example assembler code for BPF filter "tcpdump port 22"
original BPF -> old JIT: original BPF -> internal BPF -> new JIT:
0: push %rbp 0: push %rbp
1: mov %rsp,%rbp 1: mov %rsp,%rbp
4: sub $0x60,%rsp 4: sub $0x228,%rsp
8: mov %rbx,-0x8(%rbp) b: mov %rbx,-0x228(%rbp) // prologue
12: mov %r13,-0x220(%rbp)
19: mov %r14,-0x218(%rbp)
20: mov %r15,-0x210(%rbp)
27: xor %eax,%eax // clear A
c: xor %ebx,%ebx 29: xor %r13,%r13 // clear X
e: mov 0x68(%rdi),%r9d 2c: mov 0x68(%rdi),%r9d
12: sub 0x6c(%rdi),%r9d 30: sub 0x6c(%rdi),%r9d
16: mov 0xd8(%rdi),%r8 34: mov 0xd8(%rdi),%r10
3b: mov %rdi,%rbx
1d: mov $0xc,%esi 3e: mov $0xc,%esi
22: callq 0xffffffffe1021e15 43: callq 0xffffffffe102bd75
27: cmp $0x86dd,%eax 48: cmp $0x86dd,%rax
2c: jne 0x0000000000000069 4f: jne 0x000000000000009a
2e: mov $0x14,%esi 51: mov $0x14,%esi
33: callq 0xffffffffe1021e31 56: callq 0xffffffffe102bd91
38: cmp $0x84,%eax 5b: cmp $0x84,%rax
3d: je 0x0000000000000049 62: je 0x0000000000000074
3f: cmp $0x6,%eax 64: cmp $0x6,%rax
42: je 0x0000000000000049 68: je 0x0000000000000074
44: cmp $0x11,%eax 6a: cmp $0x11,%rax
47: jne 0x00000000000000c6 6e: jne 0x0000000000000117
49: mov $0x36,%esi 74: mov $0x36,%esi
4e: callq 0xffffffffe1021e15 79: callq 0xffffffffe102bd75
53: cmp $0x16,%eax 7e: cmp $0x16,%rax
56: je 0x00000000000000bf 82: je 0x0000000000000110
58: mov $0x38,%esi 88: mov $0x38,%esi
5d: callq 0xffffffffe1021e15 8d: callq 0xffffffffe102bd75
62: cmp $0x16,%eax 92: cmp $0x16,%rax
65: je 0x00000000000000bf 96: je 0x0000000000000110
67: jmp 0x00000000000000c6 98: jmp 0x0000000000000117
69: cmp $0x800,%eax 9a: cmp $0x800,%rax
6e: jne 0x00000000000000c6 a1: jne 0x0000000000000117
70: mov $0x17,%esi a3: mov $0x17,%esi
75: callq 0xffffffffe1021e31 a8: callq 0xffffffffe102bd91
7a: cmp $0x84,%eax ad: cmp $0x84,%rax
7f: je 0x000000000000008b b4: je 0x00000000000000c2
81: cmp $0x6,%eax b6: cmp $0x6,%rax
84: je 0x000000000000008b ba: je 0x00000000000000c2
86: cmp $0x11,%eax bc: cmp $0x11,%rax
89: jne 0x00000000000000c6 c0: jne 0x0000000000000117
8b: mov $0x14,%esi c2: mov $0x14,%esi
90: callq 0xffffffffe1021e15 c7: callq 0xffffffffe102bd75
95: test $0x1fff,%ax cc: test $0x1fff,%rax
99: jne 0x00000000000000c6 d3: jne 0x0000000000000117
d5: mov %rax,%r14
9b: mov $0xe,%esi d8: mov $0xe,%esi
a0: callq 0xffffffffe1021e44 dd: callq 0xffffffffe102bd91 // MSH
e2: and $0xf,%eax
e5: shl $0x2,%eax
e8: mov %rax,%r13
eb: mov %r14,%rax
ee: mov %r13,%rsi
a5: lea 0xe(%rbx),%esi f1: add $0xe,%esi
a8: callq 0xffffffffe1021e0d f4: callq 0xffffffffe102bd6d
ad: cmp $0x16,%eax f9: cmp $0x16,%rax
b0: je 0x00000000000000bf fd: je 0x0000000000000110
ff: mov %r13,%rsi
b2: lea 0x10(%rbx),%esi 102: add $0x10,%esi
b5: callq 0xffffffffe1021e0d 105: callq 0xffffffffe102bd6d
ba: cmp $0x16,%eax 10a: cmp $0x16,%rax
bd: jne 0x00000000000000c6 10e: jne 0x0000000000000117
bf: mov $0xffff,%eax 110: mov $0xffff,%eax
c4: jmp 0x00000000000000c8 115: jmp 0x000000000000011c
c6: xor %eax,%eax 117: mov $0x0,%eax
c8: mov -0x8(%rbp),%rbx 11c: mov -0x228(%rbp),%rbx // epilogue
cc: leaveq 123: mov -0x220(%rbp),%r13
cd: retq 12a: mov -0x218(%rbp),%r14
131: mov -0x210(%rbp),%r15
138: leaveq
139: retq
On fully cached SKBs both JITed functions take 12 nsec to execute.
BPF interpreter executes the program in 30 nsec.
The difference in generated assembler is due to the following:
Old BPF imlements LDX_MSH instruction via sk_load_byte_msh() helper function
inside bpf_jit.S.
New JIT removes the helper and does it explicitly, so ldx_msh cost
is the same for both JITs, but generated code looks longer.
New JIT has 4 registers to save, so prologue/epilogue are larger,
but the cost is within noise on x64.
Old JIT checks whether first insn clears A and if not emits 'xor %eax,%eax'.
New JIT clears %rax unconditionally.
2. old BPF JIT doesn't support ANC_NLATTR, ANC_PAY_OFFSET, ANC_RANDOM
extensions. New JIT supports all BPF extensions.
Performance of such filters improves 2-4 times depending on a filter.
The longer the filter the higher performance gain.
Synthetic benchmarks with many ancillary loads see 20x speedup
which seems to be the maximum gain from JIT
Notes:
. net.core.bpf_jit_enable=2 + tools/net/bpf_jit_disasm is still functional
and can be used to see generated assembler
. there are two jit_compile() functions and code flow for classic filters is:
sk_attach_filter() - load classic BPF
bpf_jit_compile() - try to JIT from classic BPF
sk_convert_filter() - convert classic to internal
bpf_int_jit_compile() - JIT from internal BPF
seccomp and tracing filters will just call bpf_int_jit_compile()
Signed-off-by: Alexei Starovoitov <ast@plumgrid.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
2014-05-14 02:50:46 +00:00
|
|
|
break;
|
2021-09-13 21:13:37 +00:00
|
|
|
|
x86/bpf: Clean up non-standard comments, to make the code more readable
So by chance I looked into x86 assembly in arch/x86/net/bpf_jit_comp.c and
noticed the weird and inconsistent comment style it mistakenly learned from
the networking code:
/* Multi-line comment ...
* ... looks like this.
*/
Fix this to use the standard comment style specified in Documentation/CodingStyle
and used in arch/x86/ as well:
/*
* Multi-line comment ...
* ... looks like this.
*/
Also, to quote Linus's ... more explicit views about this:
http://article.gmane.org/gmane.linux.kernel.cryptoapi/21066
> But no, the networking code picked *none* of the above sane formats.
> Instead, it picked these two models that are just half-arsed
> shit-for-brains:
>
> (no)
> /* This is disgusting drug-induced
> * crap, and should die
> */
>
> (no-no-no)
> /* This is also very nasty
> * and visually unbalanced */
>
> Please. The networking code actually has the *worst* possible comment
> style. You can literally find that (no-no-no) style, which is just
> really horribly disgusting and worse than the otherwise fairly similar
> (d) in pretty much every way.
Also improve the comments and some other details while at it:
- Don't mix same-line and previous-line comment style on otherwise
identical code patterns within the same function,
- capitalize 'BPF' and x86 register names consistently,
- capitalize sentences consistently,
- instead of 'x64' use 'x86-64': x64 is a Microsoft specific term,
- use more consistent punctuation,
- use standard coding style in macros as well,
- fix typos and a few other minor details.
Consistent coding style is not optional, at least in arch/x86/.
No change in functionality.
( In case this commit causes conflicts with pending development code
I'll be glad to help resolve any conflicts! )
Acked-by: Thomas Gleixner <tglx@linutronix.de>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Borislav Petkov <bp@alien8.de>
Cc: H. Peter Anvin <hpa@zytor.com>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: David S. Miller <davem@davemloft.net>
Cc: Eric Dumazet <edumazet@google.com>
Cc: Daniel Borkmann <daniel@iogearbox.net>
Cc: Alexei Starovoitov <ast@fb.com>
Cc: Hideaki YOSHIFUJI <yoshfuji@linux-ipv6.org>
Cc: netdev@vger.kernel.org
Cc: linux-kernel@vger.kernel.org
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
2018-04-27 09:54:40 +00:00
|
|
|
/* Shifts */
|
net: filter: x86: internal BPF JIT
Maps all internal BPF instructions into x86_64 instructions.
This patch replaces original BPF x64 JIT with internal BPF x64 JIT.
sysctl net.core.bpf_jit_enable is reused as on/off switch.
Performance:
1. old BPF JIT and internal BPF JIT generate equivalent x86_64 code.
No performance difference is observed for filters that were JIT-able before
Example assembler code for BPF filter "tcpdump port 22"
original BPF -> old JIT: original BPF -> internal BPF -> new JIT:
0: push %rbp 0: push %rbp
1: mov %rsp,%rbp 1: mov %rsp,%rbp
4: sub $0x60,%rsp 4: sub $0x228,%rsp
8: mov %rbx,-0x8(%rbp) b: mov %rbx,-0x228(%rbp) // prologue
12: mov %r13,-0x220(%rbp)
19: mov %r14,-0x218(%rbp)
20: mov %r15,-0x210(%rbp)
27: xor %eax,%eax // clear A
c: xor %ebx,%ebx 29: xor %r13,%r13 // clear X
e: mov 0x68(%rdi),%r9d 2c: mov 0x68(%rdi),%r9d
12: sub 0x6c(%rdi),%r9d 30: sub 0x6c(%rdi),%r9d
16: mov 0xd8(%rdi),%r8 34: mov 0xd8(%rdi),%r10
3b: mov %rdi,%rbx
1d: mov $0xc,%esi 3e: mov $0xc,%esi
22: callq 0xffffffffe1021e15 43: callq 0xffffffffe102bd75
27: cmp $0x86dd,%eax 48: cmp $0x86dd,%rax
2c: jne 0x0000000000000069 4f: jne 0x000000000000009a
2e: mov $0x14,%esi 51: mov $0x14,%esi
33: callq 0xffffffffe1021e31 56: callq 0xffffffffe102bd91
38: cmp $0x84,%eax 5b: cmp $0x84,%rax
3d: je 0x0000000000000049 62: je 0x0000000000000074
3f: cmp $0x6,%eax 64: cmp $0x6,%rax
42: je 0x0000000000000049 68: je 0x0000000000000074
44: cmp $0x11,%eax 6a: cmp $0x11,%rax
47: jne 0x00000000000000c6 6e: jne 0x0000000000000117
49: mov $0x36,%esi 74: mov $0x36,%esi
4e: callq 0xffffffffe1021e15 79: callq 0xffffffffe102bd75
53: cmp $0x16,%eax 7e: cmp $0x16,%rax
56: je 0x00000000000000bf 82: je 0x0000000000000110
58: mov $0x38,%esi 88: mov $0x38,%esi
5d: callq 0xffffffffe1021e15 8d: callq 0xffffffffe102bd75
62: cmp $0x16,%eax 92: cmp $0x16,%rax
65: je 0x00000000000000bf 96: je 0x0000000000000110
67: jmp 0x00000000000000c6 98: jmp 0x0000000000000117
69: cmp $0x800,%eax 9a: cmp $0x800,%rax
6e: jne 0x00000000000000c6 a1: jne 0x0000000000000117
70: mov $0x17,%esi a3: mov $0x17,%esi
75: callq 0xffffffffe1021e31 a8: callq 0xffffffffe102bd91
7a: cmp $0x84,%eax ad: cmp $0x84,%rax
7f: je 0x000000000000008b b4: je 0x00000000000000c2
81: cmp $0x6,%eax b6: cmp $0x6,%rax
84: je 0x000000000000008b ba: je 0x00000000000000c2
86: cmp $0x11,%eax bc: cmp $0x11,%rax
89: jne 0x00000000000000c6 c0: jne 0x0000000000000117
8b: mov $0x14,%esi c2: mov $0x14,%esi
90: callq 0xffffffffe1021e15 c7: callq 0xffffffffe102bd75
95: test $0x1fff,%ax cc: test $0x1fff,%rax
99: jne 0x00000000000000c6 d3: jne 0x0000000000000117
d5: mov %rax,%r14
9b: mov $0xe,%esi d8: mov $0xe,%esi
a0: callq 0xffffffffe1021e44 dd: callq 0xffffffffe102bd91 // MSH
e2: and $0xf,%eax
e5: shl $0x2,%eax
e8: mov %rax,%r13
eb: mov %r14,%rax
ee: mov %r13,%rsi
a5: lea 0xe(%rbx),%esi f1: add $0xe,%esi
a8: callq 0xffffffffe1021e0d f4: callq 0xffffffffe102bd6d
ad: cmp $0x16,%eax f9: cmp $0x16,%rax
b0: je 0x00000000000000bf fd: je 0x0000000000000110
ff: mov %r13,%rsi
b2: lea 0x10(%rbx),%esi 102: add $0x10,%esi
b5: callq 0xffffffffe1021e0d 105: callq 0xffffffffe102bd6d
ba: cmp $0x16,%eax 10a: cmp $0x16,%rax
bd: jne 0x00000000000000c6 10e: jne 0x0000000000000117
bf: mov $0xffff,%eax 110: mov $0xffff,%eax
c4: jmp 0x00000000000000c8 115: jmp 0x000000000000011c
c6: xor %eax,%eax 117: mov $0x0,%eax
c8: mov -0x8(%rbp),%rbx 11c: mov -0x228(%rbp),%rbx // epilogue
cc: leaveq 123: mov -0x220(%rbp),%r13
cd: retq 12a: mov -0x218(%rbp),%r14
131: mov -0x210(%rbp),%r15
138: leaveq
139: retq
On fully cached SKBs both JITed functions take 12 nsec to execute.
BPF interpreter executes the program in 30 nsec.
The difference in generated assembler is due to the following:
Old BPF imlements LDX_MSH instruction via sk_load_byte_msh() helper function
inside bpf_jit.S.
New JIT removes the helper and does it explicitly, so ldx_msh cost
is the same for both JITs, but generated code looks longer.
New JIT has 4 registers to save, so prologue/epilogue are larger,
but the cost is within noise on x64.
Old JIT checks whether first insn clears A and if not emits 'xor %eax,%eax'.
New JIT clears %rax unconditionally.
2. old BPF JIT doesn't support ANC_NLATTR, ANC_PAY_OFFSET, ANC_RANDOM
extensions. New JIT supports all BPF extensions.
Performance of such filters improves 2-4 times depending on a filter.
The longer the filter the higher performance gain.
Synthetic benchmarks with many ancillary loads see 20x speedup
which seems to be the maximum gain from JIT
Notes:
. net.core.bpf_jit_enable=2 + tools/net/bpf_jit_disasm is still functional
and can be used to see generated assembler
. there are two jit_compile() functions and code flow for classic filters is:
sk_attach_filter() - load classic BPF
bpf_jit_compile() - try to JIT from classic BPF
sk_convert_filter() - convert classic to internal
bpf_int_jit_compile() - JIT from internal BPF
seccomp and tracing filters will just call bpf_int_jit_compile()
Signed-off-by: Alexei Starovoitov <ast@plumgrid.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
2014-05-14 02:50:46 +00:00
|
|
|
case BPF_ALU | BPF_LSH | BPF_K:
|
|
|
|
case BPF_ALU | BPF_RSH | BPF_K:
|
|
|
|
case BPF_ALU | BPF_ARSH | BPF_K:
|
|
|
|
case BPF_ALU64 | BPF_LSH | BPF_K:
|
|
|
|
case BPF_ALU64 | BPF_RSH | BPF_K:
|
|
|
|
case BPF_ALU64 | BPF_ARSH | BPF_K:
|
2021-10-06 19:41:35 +00:00
|
|
|
maybe_emit_1mod(&prog, dst_reg,
|
|
|
|
BPF_CLASS(insn->code) == BPF_ALU64);
|
net: filter: x86: internal BPF JIT
Maps all internal BPF instructions into x86_64 instructions.
This patch replaces original BPF x64 JIT with internal BPF x64 JIT.
sysctl net.core.bpf_jit_enable is reused as on/off switch.
Performance:
1. old BPF JIT and internal BPF JIT generate equivalent x86_64 code.
No performance difference is observed for filters that were JIT-able before
Example assembler code for BPF filter "tcpdump port 22"
original BPF -> old JIT: original BPF -> internal BPF -> new JIT:
0: push %rbp 0: push %rbp
1: mov %rsp,%rbp 1: mov %rsp,%rbp
4: sub $0x60,%rsp 4: sub $0x228,%rsp
8: mov %rbx,-0x8(%rbp) b: mov %rbx,-0x228(%rbp) // prologue
12: mov %r13,-0x220(%rbp)
19: mov %r14,-0x218(%rbp)
20: mov %r15,-0x210(%rbp)
27: xor %eax,%eax // clear A
c: xor %ebx,%ebx 29: xor %r13,%r13 // clear X
e: mov 0x68(%rdi),%r9d 2c: mov 0x68(%rdi),%r9d
12: sub 0x6c(%rdi),%r9d 30: sub 0x6c(%rdi),%r9d
16: mov 0xd8(%rdi),%r8 34: mov 0xd8(%rdi),%r10
3b: mov %rdi,%rbx
1d: mov $0xc,%esi 3e: mov $0xc,%esi
22: callq 0xffffffffe1021e15 43: callq 0xffffffffe102bd75
27: cmp $0x86dd,%eax 48: cmp $0x86dd,%rax
2c: jne 0x0000000000000069 4f: jne 0x000000000000009a
2e: mov $0x14,%esi 51: mov $0x14,%esi
33: callq 0xffffffffe1021e31 56: callq 0xffffffffe102bd91
38: cmp $0x84,%eax 5b: cmp $0x84,%rax
3d: je 0x0000000000000049 62: je 0x0000000000000074
3f: cmp $0x6,%eax 64: cmp $0x6,%rax
42: je 0x0000000000000049 68: je 0x0000000000000074
44: cmp $0x11,%eax 6a: cmp $0x11,%rax
47: jne 0x00000000000000c6 6e: jne 0x0000000000000117
49: mov $0x36,%esi 74: mov $0x36,%esi
4e: callq 0xffffffffe1021e15 79: callq 0xffffffffe102bd75
53: cmp $0x16,%eax 7e: cmp $0x16,%rax
56: je 0x00000000000000bf 82: je 0x0000000000000110
58: mov $0x38,%esi 88: mov $0x38,%esi
5d: callq 0xffffffffe1021e15 8d: callq 0xffffffffe102bd75
62: cmp $0x16,%eax 92: cmp $0x16,%rax
65: je 0x00000000000000bf 96: je 0x0000000000000110
67: jmp 0x00000000000000c6 98: jmp 0x0000000000000117
69: cmp $0x800,%eax 9a: cmp $0x800,%rax
6e: jne 0x00000000000000c6 a1: jne 0x0000000000000117
70: mov $0x17,%esi a3: mov $0x17,%esi
75: callq 0xffffffffe1021e31 a8: callq 0xffffffffe102bd91
7a: cmp $0x84,%eax ad: cmp $0x84,%rax
7f: je 0x000000000000008b b4: je 0x00000000000000c2
81: cmp $0x6,%eax b6: cmp $0x6,%rax
84: je 0x000000000000008b ba: je 0x00000000000000c2
86: cmp $0x11,%eax bc: cmp $0x11,%rax
89: jne 0x00000000000000c6 c0: jne 0x0000000000000117
8b: mov $0x14,%esi c2: mov $0x14,%esi
90: callq 0xffffffffe1021e15 c7: callq 0xffffffffe102bd75
95: test $0x1fff,%ax cc: test $0x1fff,%rax
99: jne 0x00000000000000c6 d3: jne 0x0000000000000117
d5: mov %rax,%r14
9b: mov $0xe,%esi d8: mov $0xe,%esi
a0: callq 0xffffffffe1021e44 dd: callq 0xffffffffe102bd91 // MSH
e2: and $0xf,%eax
e5: shl $0x2,%eax
e8: mov %rax,%r13
eb: mov %r14,%rax
ee: mov %r13,%rsi
a5: lea 0xe(%rbx),%esi f1: add $0xe,%esi
a8: callq 0xffffffffe1021e0d f4: callq 0xffffffffe102bd6d
ad: cmp $0x16,%eax f9: cmp $0x16,%rax
b0: je 0x00000000000000bf fd: je 0x0000000000000110
ff: mov %r13,%rsi
b2: lea 0x10(%rbx),%esi 102: add $0x10,%esi
b5: callq 0xffffffffe1021e0d 105: callq 0xffffffffe102bd6d
ba: cmp $0x16,%eax 10a: cmp $0x16,%rax
bd: jne 0x00000000000000c6 10e: jne 0x0000000000000117
bf: mov $0xffff,%eax 110: mov $0xffff,%eax
c4: jmp 0x00000000000000c8 115: jmp 0x000000000000011c
c6: xor %eax,%eax 117: mov $0x0,%eax
c8: mov -0x8(%rbp),%rbx 11c: mov -0x228(%rbp),%rbx // epilogue
cc: leaveq 123: mov -0x220(%rbp),%r13
cd: retq 12a: mov -0x218(%rbp),%r14
131: mov -0x210(%rbp),%r15
138: leaveq
139: retq
On fully cached SKBs both JITed functions take 12 nsec to execute.
BPF interpreter executes the program in 30 nsec.
The difference in generated assembler is due to the following:
Old BPF imlements LDX_MSH instruction via sk_load_byte_msh() helper function
inside bpf_jit.S.
New JIT removes the helper and does it explicitly, so ldx_msh cost
is the same for both JITs, but generated code looks longer.
New JIT has 4 registers to save, so prologue/epilogue are larger,
but the cost is within noise on x64.
Old JIT checks whether first insn clears A and if not emits 'xor %eax,%eax'.
New JIT clears %rax unconditionally.
2. old BPF JIT doesn't support ANC_NLATTR, ANC_PAY_OFFSET, ANC_RANDOM
extensions. New JIT supports all BPF extensions.
Performance of such filters improves 2-4 times depending on a filter.
The longer the filter the higher performance gain.
Synthetic benchmarks with many ancillary loads see 20x speedup
which seems to be the maximum gain from JIT
Notes:
. net.core.bpf_jit_enable=2 + tools/net/bpf_jit_disasm is still functional
and can be used to see generated assembler
. there are two jit_compile() functions and code flow for classic filters is:
sk_attach_filter() - load classic BPF
bpf_jit_compile() - try to JIT from classic BPF
sk_convert_filter() - convert classic to internal
bpf_int_jit_compile() - JIT from internal BPF
seccomp and tracing filters will just call bpf_int_jit_compile()
Signed-off-by: Alexei Starovoitov <ast@plumgrid.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
2014-05-14 02:50:46 +00:00
|
|
|
|
2021-01-14 18:17:43 +00:00
|
|
|
b3 = simple_alu_opcodes[BPF_OP(insn->code)];
|
2018-02-24 00:07:58 +00:00
|
|
|
if (imm32 == 1)
|
|
|
|
EMIT2(0xD1, add_1reg(b3, dst_reg));
|
|
|
|
else
|
|
|
|
EMIT3(0xC1, add_1reg(b3, dst_reg), imm32);
|
net: filter: x86: internal BPF JIT
Maps all internal BPF instructions into x86_64 instructions.
This patch replaces original BPF x64 JIT with internal BPF x64 JIT.
sysctl net.core.bpf_jit_enable is reused as on/off switch.
Performance:
1. old BPF JIT and internal BPF JIT generate equivalent x86_64 code.
No performance difference is observed for filters that were JIT-able before
Example assembler code for BPF filter "tcpdump port 22"
original BPF -> old JIT: original BPF -> internal BPF -> new JIT:
0: push %rbp 0: push %rbp
1: mov %rsp,%rbp 1: mov %rsp,%rbp
4: sub $0x60,%rsp 4: sub $0x228,%rsp
8: mov %rbx,-0x8(%rbp) b: mov %rbx,-0x228(%rbp) // prologue
12: mov %r13,-0x220(%rbp)
19: mov %r14,-0x218(%rbp)
20: mov %r15,-0x210(%rbp)
27: xor %eax,%eax // clear A
c: xor %ebx,%ebx 29: xor %r13,%r13 // clear X
e: mov 0x68(%rdi),%r9d 2c: mov 0x68(%rdi),%r9d
12: sub 0x6c(%rdi),%r9d 30: sub 0x6c(%rdi),%r9d
16: mov 0xd8(%rdi),%r8 34: mov 0xd8(%rdi),%r10
3b: mov %rdi,%rbx
1d: mov $0xc,%esi 3e: mov $0xc,%esi
22: callq 0xffffffffe1021e15 43: callq 0xffffffffe102bd75
27: cmp $0x86dd,%eax 48: cmp $0x86dd,%rax
2c: jne 0x0000000000000069 4f: jne 0x000000000000009a
2e: mov $0x14,%esi 51: mov $0x14,%esi
33: callq 0xffffffffe1021e31 56: callq 0xffffffffe102bd91
38: cmp $0x84,%eax 5b: cmp $0x84,%rax
3d: je 0x0000000000000049 62: je 0x0000000000000074
3f: cmp $0x6,%eax 64: cmp $0x6,%rax
42: je 0x0000000000000049 68: je 0x0000000000000074
44: cmp $0x11,%eax 6a: cmp $0x11,%rax
47: jne 0x00000000000000c6 6e: jne 0x0000000000000117
49: mov $0x36,%esi 74: mov $0x36,%esi
4e: callq 0xffffffffe1021e15 79: callq 0xffffffffe102bd75
53: cmp $0x16,%eax 7e: cmp $0x16,%rax
56: je 0x00000000000000bf 82: je 0x0000000000000110
58: mov $0x38,%esi 88: mov $0x38,%esi
5d: callq 0xffffffffe1021e15 8d: callq 0xffffffffe102bd75
62: cmp $0x16,%eax 92: cmp $0x16,%rax
65: je 0x00000000000000bf 96: je 0x0000000000000110
67: jmp 0x00000000000000c6 98: jmp 0x0000000000000117
69: cmp $0x800,%eax 9a: cmp $0x800,%rax
6e: jne 0x00000000000000c6 a1: jne 0x0000000000000117
70: mov $0x17,%esi a3: mov $0x17,%esi
75: callq 0xffffffffe1021e31 a8: callq 0xffffffffe102bd91
7a: cmp $0x84,%eax ad: cmp $0x84,%rax
7f: je 0x000000000000008b b4: je 0x00000000000000c2
81: cmp $0x6,%eax b6: cmp $0x6,%rax
84: je 0x000000000000008b ba: je 0x00000000000000c2
86: cmp $0x11,%eax bc: cmp $0x11,%rax
89: jne 0x00000000000000c6 c0: jne 0x0000000000000117
8b: mov $0x14,%esi c2: mov $0x14,%esi
90: callq 0xffffffffe1021e15 c7: callq 0xffffffffe102bd75
95: test $0x1fff,%ax cc: test $0x1fff,%rax
99: jne 0x00000000000000c6 d3: jne 0x0000000000000117
d5: mov %rax,%r14
9b: mov $0xe,%esi d8: mov $0xe,%esi
a0: callq 0xffffffffe1021e44 dd: callq 0xffffffffe102bd91 // MSH
e2: and $0xf,%eax
e5: shl $0x2,%eax
e8: mov %rax,%r13
eb: mov %r14,%rax
ee: mov %r13,%rsi
a5: lea 0xe(%rbx),%esi f1: add $0xe,%esi
a8: callq 0xffffffffe1021e0d f4: callq 0xffffffffe102bd6d
ad: cmp $0x16,%eax f9: cmp $0x16,%rax
b0: je 0x00000000000000bf fd: je 0x0000000000000110
ff: mov %r13,%rsi
b2: lea 0x10(%rbx),%esi 102: add $0x10,%esi
b5: callq 0xffffffffe1021e0d 105: callq 0xffffffffe102bd6d
ba: cmp $0x16,%eax 10a: cmp $0x16,%rax
bd: jne 0x00000000000000c6 10e: jne 0x0000000000000117
bf: mov $0xffff,%eax 110: mov $0xffff,%eax
c4: jmp 0x00000000000000c8 115: jmp 0x000000000000011c
c6: xor %eax,%eax 117: mov $0x0,%eax
c8: mov -0x8(%rbp),%rbx 11c: mov -0x228(%rbp),%rbx // epilogue
cc: leaveq 123: mov -0x220(%rbp),%r13
cd: retq 12a: mov -0x218(%rbp),%r14
131: mov -0x210(%rbp),%r15
138: leaveq
139: retq
On fully cached SKBs both JITed functions take 12 nsec to execute.
BPF interpreter executes the program in 30 nsec.
The difference in generated assembler is due to the following:
Old BPF imlements LDX_MSH instruction via sk_load_byte_msh() helper function
inside bpf_jit.S.
New JIT removes the helper and does it explicitly, so ldx_msh cost
is the same for both JITs, but generated code looks longer.
New JIT has 4 registers to save, so prologue/epilogue are larger,
but the cost is within noise on x64.
Old JIT checks whether first insn clears A and if not emits 'xor %eax,%eax'.
New JIT clears %rax unconditionally.
2. old BPF JIT doesn't support ANC_NLATTR, ANC_PAY_OFFSET, ANC_RANDOM
extensions. New JIT supports all BPF extensions.
Performance of such filters improves 2-4 times depending on a filter.
The longer the filter the higher performance gain.
Synthetic benchmarks with many ancillary loads see 20x speedup
which seems to be the maximum gain from JIT
Notes:
. net.core.bpf_jit_enable=2 + tools/net/bpf_jit_disasm is still functional
and can be used to see generated assembler
. there are two jit_compile() functions and code flow for classic filters is:
sk_attach_filter() - load classic BPF
bpf_jit_compile() - try to JIT from classic BPF
sk_convert_filter() - convert classic to internal
bpf_int_jit_compile() - JIT from internal BPF
seccomp and tracing filters will just call bpf_int_jit_compile()
Signed-off-by: Alexei Starovoitov <ast@plumgrid.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
2014-05-14 02:50:46 +00:00
|
|
|
break;
|
|
|
|
|
2014-08-25 19:27:02 +00:00
|
|
|
case BPF_ALU | BPF_LSH | BPF_X:
|
|
|
|
case BPF_ALU | BPF_RSH | BPF_X:
|
|
|
|
case BPF_ALU | BPF_ARSH | BPF_X:
|
|
|
|
case BPF_ALU64 | BPF_LSH | BPF_X:
|
|
|
|
case BPF_ALU64 | BPF_RSH | BPF_X:
|
|
|
|
case BPF_ALU64 | BPF_ARSH | BPF_X:
|
bpf,x64: use shrx/sarx/shlx when available
BMI2 provides 3 shift instructions (shrx, sarx and shlx) that use VEX
encoding but target general purpose registers [1]. They allow the shift
count in any general purpose register and have the same performance as
non BMI2 shift instructions [2].
Instead of shr/sar/shl that implicitly use %cl (lowest 8 bit of %rcx),
emit their more flexible alternatives provided in BMI2 when advantageous;
keep using the non BMI2 instructions when shift count is already in
BPF_REG_4/%rcx as non BMI2 instructions are shorter.
To summarize, when BMI2 is available:
-------------------------------------------------
| arbitrary dst
=================================================
src == ecx | shl dst, cl
-------------------------------------------------
src != ecx | shlx dst, dst, src
-------------------------------------------------
And no additional register shuffling is needed.
A concrete example between non BMI2 and BMI2 codegen. To shift %rsi by
%rdi:
Without BMI2:
ef3: push %rcx
51
ef4: mov %rdi,%rcx
48 89 f9
ef7: shl %cl,%rsi
48 d3 e6
efa: pop %rcx
59
With BMI2:
f0b: shlx %rdi,%rsi,%rsi
c4 e2 c1 f7 f6
[1] https://en.wikipedia.org/wiki/X86_Bit_manipulation_instruction_set
[2] https://www.agner.org/optimize/instruction_tables.pdf
Signed-off-by: Jie Meng <jmeng@fb.com>
Link: https://lore.kernel.org/r/20221007202348.1118830-3-jmeng@fb.com
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
2022-10-07 20:23:48 +00:00
|
|
|
/* BMI2 shifts aren't better when shift count is already in rcx */
|
|
|
|
if (boot_cpu_has(X86_FEATURE_BMI2) && src_reg != BPF_REG_4) {
|
|
|
|
/* shrx/sarx/shlx dst_reg, dst_reg, src_reg */
|
|
|
|
bool w = (BPF_CLASS(insn->code) == BPF_ALU64);
|
|
|
|
u8 op;
|
|
|
|
|
|
|
|
switch (BPF_OP(insn->code)) {
|
|
|
|
case BPF_LSH:
|
|
|
|
op = 1; /* prefix 0x66 */
|
|
|
|
break;
|
|
|
|
case BPF_RSH:
|
|
|
|
op = 3; /* prefix 0xf2 */
|
|
|
|
break;
|
|
|
|
case BPF_ARSH:
|
|
|
|
op = 2; /* prefix 0xf3 */
|
|
|
|
break;
|
|
|
|
}
|
2014-08-25 19:27:02 +00:00
|
|
|
|
bpf,x64: use shrx/sarx/shlx when available
BMI2 provides 3 shift instructions (shrx, sarx and shlx) that use VEX
encoding but target general purpose registers [1]. They allow the shift
count in any general purpose register and have the same performance as
non BMI2 shift instructions [2].
Instead of shr/sar/shl that implicitly use %cl (lowest 8 bit of %rcx),
emit their more flexible alternatives provided in BMI2 when advantageous;
keep using the non BMI2 instructions when shift count is already in
BPF_REG_4/%rcx as non BMI2 instructions are shorter.
To summarize, when BMI2 is available:
-------------------------------------------------
| arbitrary dst
=================================================
src == ecx | shl dst, cl
-------------------------------------------------
src != ecx | shlx dst, dst, src
-------------------------------------------------
And no additional register shuffling is needed.
A concrete example between non BMI2 and BMI2 codegen. To shift %rsi by
%rdi:
Without BMI2:
ef3: push %rcx
51
ef4: mov %rdi,%rcx
48 89 f9
ef7: shl %cl,%rsi
48 d3 e6
efa: pop %rcx
59
With BMI2:
f0b: shlx %rdi,%rsi,%rsi
c4 e2 c1 f7 f6
[1] https://en.wikipedia.org/wiki/X86_Bit_manipulation_instruction_set
[2] https://www.agner.org/optimize/instruction_tables.pdf
Signed-off-by: Jie Meng <jmeng@fb.com>
Link: https://lore.kernel.org/r/20221007202348.1118830-3-jmeng@fb.com
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
2022-10-07 20:23:48 +00:00
|
|
|
emit_shiftx(&prog, dst_reg, src_reg, w, op);
|
|
|
|
|
|
|
|
break;
|
2014-08-25 19:27:02 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
if (src_reg != BPF_REG_4) { /* common case */
|
bpf,x64: avoid unnecessary instructions when shift dest is ecx
x64 JIT produces redundant instructions when a shift operation's
destination register is BPF_REG_4/ecx and this patch removes them.
Specifically, when dest reg is BPF_REG_4 but the src isn't, we
needn't push and pop ecx around shift only to get it overwritten
by r11 immediately afterwards.
In the rare case when both dest and src registers are BPF_REG_4,
a single shift instruction is sufficient and we don't need the
two MOV instructions around the shift.
To summarize using shift left as an example, without patch:
-------------------------------------------------
| dst == ecx | dst != ecx
=================================================
src == ecx | mov r11, ecx | shl dst, cl
| shl r11, ecx |
| mov ecx, r11 |
-------------------------------------------------
src != ecx | mov r11, ecx | push ecx
| push ecx | mov ecx, src
| mov ecx, src | shl dst, cl
| shl r11, cl | pop ecx
| pop ecx |
| mov ecx, r11 |
-------------------------------------------------
With patch:
-------------------------------------------------
| dst == ecx | dst != ecx
=================================================
src == ecx | shl ecx, cl | shl dst, cl
-------------------------------------------------
src != ecx | mov r11, ecx | push ecx
| mov ecx, src | mov ecx, src
| shl r11, cl | shl dst, cl
| mov ecx, r11 | pop ecx
-------------------------------------------------
Signed-off-by: Jie Meng <jmeng@fb.com>
Link: https://lore.kernel.org/r/20221007202348.1118830-2-jmeng@fb.com
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
2022-10-07 20:23:47 +00:00
|
|
|
/* Check for bad case when dst_reg == rcx */
|
|
|
|
if (dst_reg == BPF_REG_4) {
|
|
|
|
/* mov r11, dst_reg */
|
|
|
|
EMIT_mov(AUX_REG, dst_reg);
|
|
|
|
dst_reg = AUX_REG;
|
|
|
|
} else {
|
|
|
|
EMIT1(0x51); /* push rcx */
|
|
|
|
}
|
2014-08-25 19:27:02 +00:00
|
|
|
/* mov rcx, src_reg */
|
|
|
|
EMIT_mov(BPF_REG_4, src_reg);
|
|
|
|
}
|
|
|
|
|
|
|
|
/* shl %rax, %cl | shr %rax, %cl | sar %rax, %cl */
|
2021-10-06 19:41:35 +00:00
|
|
|
maybe_emit_1mod(&prog, dst_reg,
|
|
|
|
BPF_CLASS(insn->code) == BPF_ALU64);
|
2014-08-25 19:27:02 +00:00
|
|
|
|
2021-01-14 18:17:43 +00:00
|
|
|
b3 = simple_alu_opcodes[BPF_OP(insn->code)];
|
2014-08-25 19:27:02 +00:00
|
|
|
EMIT2(0xD3, add_1reg(b3, dst_reg));
|
|
|
|
|
bpf,x64: avoid unnecessary instructions when shift dest is ecx
x64 JIT produces redundant instructions when a shift operation's
destination register is BPF_REG_4/ecx and this patch removes them.
Specifically, when dest reg is BPF_REG_4 but the src isn't, we
needn't push and pop ecx around shift only to get it overwritten
by r11 immediately afterwards.
In the rare case when both dest and src registers are BPF_REG_4,
a single shift instruction is sufficient and we don't need the
two MOV instructions around the shift.
To summarize using shift left as an example, without patch:
-------------------------------------------------
| dst == ecx | dst != ecx
=================================================
src == ecx | mov r11, ecx | shl dst, cl
| shl r11, ecx |
| mov ecx, r11 |
-------------------------------------------------
src != ecx | mov r11, ecx | push ecx
| push ecx | mov ecx, src
| mov ecx, src | shl dst, cl
| shl r11, cl | pop ecx
| pop ecx |
| mov ecx, r11 |
-------------------------------------------------
With patch:
-------------------------------------------------
| dst == ecx | dst != ecx
=================================================
src == ecx | shl ecx, cl | shl dst, cl
-------------------------------------------------
src != ecx | mov r11, ecx | push ecx
| mov ecx, src | mov ecx, src
| shl r11, cl | shl dst, cl
| mov ecx, r11 | pop ecx
-------------------------------------------------
Signed-off-by: Jie Meng <jmeng@fb.com>
Link: https://lore.kernel.org/r/20221007202348.1118830-2-jmeng@fb.com
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
2022-10-07 20:23:47 +00:00
|
|
|
if (src_reg != BPF_REG_4) {
|
|
|
|
if (insn->dst_reg == BPF_REG_4)
|
|
|
|
/* mov dst_reg, r11 */
|
|
|
|
EMIT_mov(insn->dst_reg, AUX_REG);
|
|
|
|
else
|
|
|
|
EMIT1(0x59); /* pop rcx */
|
|
|
|
}
|
2014-08-25 19:27:02 +00:00
|
|
|
|
|
|
|
break;
|
|
|
|
|
net: filter: x86: internal BPF JIT
Maps all internal BPF instructions into x86_64 instructions.
This patch replaces original BPF x64 JIT with internal BPF x64 JIT.
sysctl net.core.bpf_jit_enable is reused as on/off switch.
Performance:
1. old BPF JIT and internal BPF JIT generate equivalent x86_64 code.
No performance difference is observed for filters that were JIT-able before
Example assembler code for BPF filter "tcpdump port 22"
original BPF -> old JIT: original BPF -> internal BPF -> new JIT:
0: push %rbp 0: push %rbp
1: mov %rsp,%rbp 1: mov %rsp,%rbp
4: sub $0x60,%rsp 4: sub $0x228,%rsp
8: mov %rbx,-0x8(%rbp) b: mov %rbx,-0x228(%rbp) // prologue
12: mov %r13,-0x220(%rbp)
19: mov %r14,-0x218(%rbp)
20: mov %r15,-0x210(%rbp)
27: xor %eax,%eax // clear A
c: xor %ebx,%ebx 29: xor %r13,%r13 // clear X
e: mov 0x68(%rdi),%r9d 2c: mov 0x68(%rdi),%r9d
12: sub 0x6c(%rdi),%r9d 30: sub 0x6c(%rdi),%r9d
16: mov 0xd8(%rdi),%r8 34: mov 0xd8(%rdi),%r10
3b: mov %rdi,%rbx
1d: mov $0xc,%esi 3e: mov $0xc,%esi
22: callq 0xffffffffe1021e15 43: callq 0xffffffffe102bd75
27: cmp $0x86dd,%eax 48: cmp $0x86dd,%rax
2c: jne 0x0000000000000069 4f: jne 0x000000000000009a
2e: mov $0x14,%esi 51: mov $0x14,%esi
33: callq 0xffffffffe1021e31 56: callq 0xffffffffe102bd91
38: cmp $0x84,%eax 5b: cmp $0x84,%rax
3d: je 0x0000000000000049 62: je 0x0000000000000074
3f: cmp $0x6,%eax 64: cmp $0x6,%rax
42: je 0x0000000000000049 68: je 0x0000000000000074
44: cmp $0x11,%eax 6a: cmp $0x11,%rax
47: jne 0x00000000000000c6 6e: jne 0x0000000000000117
49: mov $0x36,%esi 74: mov $0x36,%esi
4e: callq 0xffffffffe1021e15 79: callq 0xffffffffe102bd75
53: cmp $0x16,%eax 7e: cmp $0x16,%rax
56: je 0x00000000000000bf 82: je 0x0000000000000110
58: mov $0x38,%esi 88: mov $0x38,%esi
5d: callq 0xffffffffe1021e15 8d: callq 0xffffffffe102bd75
62: cmp $0x16,%eax 92: cmp $0x16,%rax
65: je 0x00000000000000bf 96: je 0x0000000000000110
67: jmp 0x00000000000000c6 98: jmp 0x0000000000000117
69: cmp $0x800,%eax 9a: cmp $0x800,%rax
6e: jne 0x00000000000000c6 a1: jne 0x0000000000000117
70: mov $0x17,%esi a3: mov $0x17,%esi
75: callq 0xffffffffe1021e31 a8: callq 0xffffffffe102bd91
7a: cmp $0x84,%eax ad: cmp $0x84,%rax
7f: je 0x000000000000008b b4: je 0x00000000000000c2
81: cmp $0x6,%eax b6: cmp $0x6,%rax
84: je 0x000000000000008b ba: je 0x00000000000000c2
86: cmp $0x11,%eax bc: cmp $0x11,%rax
89: jne 0x00000000000000c6 c0: jne 0x0000000000000117
8b: mov $0x14,%esi c2: mov $0x14,%esi
90: callq 0xffffffffe1021e15 c7: callq 0xffffffffe102bd75
95: test $0x1fff,%ax cc: test $0x1fff,%rax
99: jne 0x00000000000000c6 d3: jne 0x0000000000000117
d5: mov %rax,%r14
9b: mov $0xe,%esi d8: mov $0xe,%esi
a0: callq 0xffffffffe1021e44 dd: callq 0xffffffffe102bd91 // MSH
e2: and $0xf,%eax
e5: shl $0x2,%eax
e8: mov %rax,%r13
eb: mov %r14,%rax
ee: mov %r13,%rsi
a5: lea 0xe(%rbx),%esi f1: add $0xe,%esi
a8: callq 0xffffffffe1021e0d f4: callq 0xffffffffe102bd6d
ad: cmp $0x16,%eax f9: cmp $0x16,%rax
b0: je 0x00000000000000bf fd: je 0x0000000000000110
ff: mov %r13,%rsi
b2: lea 0x10(%rbx),%esi 102: add $0x10,%esi
b5: callq 0xffffffffe1021e0d 105: callq 0xffffffffe102bd6d
ba: cmp $0x16,%eax 10a: cmp $0x16,%rax
bd: jne 0x00000000000000c6 10e: jne 0x0000000000000117
bf: mov $0xffff,%eax 110: mov $0xffff,%eax
c4: jmp 0x00000000000000c8 115: jmp 0x000000000000011c
c6: xor %eax,%eax 117: mov $0x0,%eax
c8: mov -0x8(%rbp),%rbx 11c: mov -0x228(%rbp),%rbx // epilogue
cc: leaveq 123: mov -0x220(%rbp),%r13
cd: retq 12a: mov -0x218(%rbp),%r14
131: mov -0x210(%rbp),%r15
138: leaveq
139: retq
On fully cached SKBs both JITed functions take 12 nsec to execute.
BPF interpreter executes the program in 30 nsec.
The difference in generated assembler is due to the following:
Old BPF imlements LDX_MSH instruction via sk_load_byte_msh() helper function
inside bpf_jit.S.
New JIT removes the helper and does it explicitly, so ldx_msh cost
is the same for both JITs, but generated code looks longer.
New JIT has 4 registers to save, so prologue/epilogue are larger,
but the cost is within noise on x64.
Old JIT checks whether first insn clears A and if not emits 'xor %eax,%eax'.
New JIT clears %rax unconditionally.
2. old BPF JIT doesn't support ANC_NLATTR, ANC_PAY_OFFSET, ANC_RANDOM
extensions. New JIT supports all BPF extensions.
Performance of such filters improves 2-4 times depending on a filter.
The longer the filter the higher performance gain.
Synthetic benchmarks with many ancillary loads see 20x speedup
which seems to be the maximum gain from JIT
Notes:
. net.core.bpf_jit_enable=2 + tools/net/bpf_jit_disasm is still functional
and can be used to see generated assembler
. there are two jit_compile() functions and code flow for classic filters is:
sk_attach_filter() - load classic BPF
bpf_jit_compile() - try to JIT from classic BPF
sk_convert_filter() - convert classic to internal
bpf_int_jit_compile() - JIT from internal BPF
seccomp and tracing filters will just call bpf_int_jit_compile()
Signed-off-by: Alexei Starovoitov <ast@plumgrid.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
2014-05-14 02:50:46 +00:00
|
|
|
case BPF_ALU | BPF_END | BPF_FROM_BE:
|
2023-07-28 01:12:13 +00:00
|
|
|
case BPF_ALU64 | BPF_END | BPF_FROM_LE:
|
2014-06-06 21:46:06 +00:00
|
|
|
switch (imm32) {
|
net: filter: x86: internal BPF JIT
Maps all internal BPF instructions into x86_64 instructions.
This patch replaces original BPF x64 JIT with internal BPF x64 JIT.
sysctl net.core.bpf_jit_enable is reused as on/off switch.
Performance:
1. old BPF JIT and internal BPF JIT generate equivalent x86_64 code.
No performance difference is observed for filters that were JIT-able before
Example assembler code for BPF filter "tcpdump port 22"
original BPF -> old JIT: original BPF -> internal BPF -> new JIT:
0: push %rbp 0: push %rbp
1: mov %rsp,%rbp 1: mov %rsp,%rbp
4: sub $0x60,%rsp 4: sub $0x228,%rsp
8: mov %rbx,-0x8(%rbp) b: mov %rbx,-0x228(%rbp) // prologue
12: mov %r13,-0x220(%rbp)
19: mov %r14,-0x218(%rbp)
20: mov %r15,-0x210(%rbp)
27: xor %eax,%eax // clear A
c: xor %ebx,%ebx 29: xor %r13,%r13 // clear X
e: mov 0x68(%rdi),%r9d 2c: mov 0x68(%rdi),%r9d
12: sub 0x6c(%rdi),%r9d 30: sub 0x6c(%rdi),%r9d
16: mov 0xd8(%rdi),%r8 34: mov 0xd8(%rdi),%r10
3b: mov %rdi,%rbx
1d: mov $0xc,%esi 3e: mov $0xc,%esi
22: callq 0xffffffffe1021e15 43: callq 0xffffffffe102bd75
27: cmp $0x86dd,%eax 48: cmp $0x86dd,%rax
2c: jne 0x0000000000000069 4f: jne 0x000000000000009a
2e: mov $0x14,%esi 51: mov $0x14,%esi
33: callq 0xffffffffe1021e31 56: callq 0xffffffffe102bd91
38: cmp $0x84,%eax 5b: cmp $0x84,%rax
3d: je 0x0000000000000049 62: je 0x0000000000000074
3f: cmp $0x6,%eax 64: cmp $0x6,%rax
42: je 0x0000000000000049 68: je 0x0000000000000074
44: cmp $0x11,%eax 6a: cmp $0x11,%rax
47: jne 0x00000000000000c6 6e: jne 0x0000000000000117
49: mov $0x36,%esi 74: mov $0x36,%esi
4e: callq 0xffffffffe1021e15 79: callq 0xffffffffe102bd75
53: cmp $0x16,%eax 7e: cmp $0x16,%rax
56: je 0x00000000000000bf 82: je 0x0000000000000110
58: mov $0x38,%esi 88: mov $0x38,%esi
5d: callq 0xffffffffe1021e15 8d: callq 0xffffffffe102bd75
62: cmp $0x16,%eax 92: cmp $0x16,%rax
65: je 0x00000000000000bf 96: je 0x0000000000000110
67: jmp 0x00000000000000c6 98: jmp 0x0000000000000117
69: cmp $0x800,%eax 9a: cmp $0x800,%rax
6e: jne 0x00000000000000c6 a1: jne 0x0000000000000117
70: mov $0x17,%esi a3: mov $0x17,%esi
75: callq 0xffffffffe1021e31 a8: callq 0xffffffffe102bd91
7a: cmp $0x84,%eax ad: cmp $0x84,%rax
7f: je 0x000000000000008b b4: je 0x00000000000000c2
81: cmp $0x6,%eax b6: cmp $0x6,%rax
84: je 0x000000000000008b ba: je 0x00000000000000c2
86: cmp $0x11,%eax bc: cmp $0x11,%rax
89: jne 0x00000000000000c6 c0: jne 0x0000000000000117
8b: mov $0x14,%esi c2: mov $0x14,%esi
90: callq 0xffffffffe1021e15 c7: callq 0xffffffffe102bd75
95: test $0x1fff,%ax cc: test $0x1fff,%rax
99: jne 0x00000000000000c6 d3: jne 0x0000000000000117
d5: mov %rax,%r14
9b: mov $0xe,%esi d8: mov $0xe,%esi
a0: callq 0xffffffffe1021e44 dd: callq 0xffffffffe102bd91 // MSH
e2: and $0xf,%eax
e5: shl $0x2,%eax
e8: mov %rax,%r13
eb: mov %r14,%rax
ee: mov %r13,%rsi
a5: lea 0xe(%rbx),%esi f1: add $0xe,%esi
a8: callq 0xffffffffe1021e0d f4: callq 0xffffffffe102bd6d
ad: cmp $0x16,%eax f9: cmp $0x16,%rax
b0: je 0x00000000000000bf fd: je 0x0000000000000110
ff: mov %r13,%rsi
b2: lea 0x10(%rbx),%esi 102: add $0x10,%esi
b5: callq 0xffffffffe1021e0d 105: callq 0xffffffffe102bd6d
ba: cmp $0x16,%eax 10a: cmp $0x16,%rax
bd: jne 0x00000000000000c6 10e: jne 0x0000000000000117
bf: mov $0xffff,%eax 110: mov $0xffff,%eax
c4: jmp 0x00000000000000c8 115: jmp 0x000000000000011c
c6: xor %eax,%eax 117: mov $0x0,%eax
c8: mov -0x8(%rbp),%rbx 11c: mov -0x228(%rbp),%rbx // epilogue
cc: leaveq 123: mov -0x220(%rbp),%r13
cd: retq 12a: mov -0x218(%rbp),%r14
131: mov -0x210(%rbp),%r15
138: leaveq
139: retq
On fully cached SKBs both JITed functions take 12 nsec to execute.
BPF interpreter executes the program in 30 nsec.
The difference in generated assembler is due to the following:
Old BPF imlements LDX_MSH instruction via sk_load_byte_msh() helper function
inside bpf_jit.S.
New JIT removes the helper and does it explicitly, so ldx_msh cost
is the same for both JITs, but generated code looks longer.
New JIT has 4 registers to save, so prologue/epilogue are larger,
but the cost is within noise on x64.
Old JIT checks whether first insn clears A and if not emits 'xor %eax,%eax'.
New JIT clears %rax unconditionally.
2. old BPF JIT doesn't support ANC_NLATTR, ANC_PAY_OFFSET, ANC_RANDOM
extensions. New JIT supports all BPF extensions.
Performance of such filters improves 2-4 times depending on a filter.
The longer the filter the higher performance gain.
Synthetic benchmarks with many ancillary loads see 20x speedup
which seems to be the maximum gain from JIT
Notes:
. net.core.bpf_jit_enable=2 + tools/net/bpf_jit_disasm is still functional
and can be used to see generated assembler
. there are two jit_compile() functions and code flow for classic filters is:
sk_attach_filter() - load classic BPF
bpf_jit_compile() - try to JIT from classic BPF
sk_convert_filter() - convert classic to internal
bpf_int_jit_compile() - JIT from internal BPF
seccomp and tracing filters will just call bpf_int_jit_compile()
Signed-off-by: Alexei Starovoitov <ast@plumgrid.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
2014-05-14 02:50:46 +00:00
|
|
|
case 16:
|
x86/bpf: Clean up non-standard comments, to make the code more readable
So by chance I looked into x86 assembly in arch/x86/net/bpf_jit_comp.c and
noticed the weird and inconsistent comment style it mistakenly learned from
the networking code:
/* Multi-line comment ...
* ... looks like this.
*/
Fix this to use the standard comment style specified in Documentation/CodingStyle
and used in arch/x86/ as well:
/*
* Multi-line comment ...
* ... looks like this.
*/
Also, to quote Linus's ... more explicit views about this:
http://article.gmane.org/gmane.linux.kernel.cryptoapi/21066
> But no, the networking code picked *none* of the above sane formats.
> Instead, it picked these two models that are just half-arsed
> shit-for-brains:
>
> (no)
> /* This is disgusting drug-induced
> * crap, and should die
> */
>
> (no-no-no)
> /* This is also very nasty
> * and visually unbalanced */
>
> Please. The networking code actually has the *worst* possible comment
> style. You can literally find that (no-no-no) style, which is just
> really horribly disgusting and worse than the otherwise fairly similar
> (d) in pretty much every way.
Also improve the comments and some other details while at it:
- Don't mix same-line and previous-line comment style on otherwise
identical code patterns within the same function,
- capitalize 'BPF' and x86 register names consistently,
- capitalize sentences consistently,
- instead of 'x64' use 'x86-64': x64 is a Microsoft specific term,
- use more consistent punctuation,
- use standard coding style in macros as well,
- fix typos and a few other minor details.
Consistent coding style is not optional, at least in arch/x86/.
No change in functionality.
( In case this commit causes conflicts with pending development code
I'll be glad to help resolve any conflicts! )
Acked-by: Thomas Gleixner <tglx@linutronix.de>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Borislav Petkov <bp@alien8.de>
Cc: H. Peter Anvin <hpa@zytor.com>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: David S. Miller <davem@davemloft.net>
Cc: Eric Dumazet <edumazet@google.com>
Cc: Daniel Borkmann <daniel@iogearbox.net>
Cc: Alexei Starovoitov <ast@fb.com>
Cc: Hideaki YOSHIFUJI <yoshfuji@linux-ipv6.org>
Cc: netdev@vger.kernel.org
Cc: linux-kernel@vger.kernel.org
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
2018-04-27 09:54:40 +00:00
|
|
|
/* Emit 'ror %ax, 8' to swap lower 2 bytes */
|
net: filter: x86: internal BPF JIT
Maps all internal BPF instructions into x86_64 instructions.
This patch replaces original BPF x64 JIT with internal BPF x64 JIT.
sysctl net.core.bpf_jit_enable is reused as on/off switch.
Performance:
1. old BPF JIT and internal BPF JIT generate equivalent x86_64 code.
No performance difference is observed for filters that were JIT-able before
Example assembler code for BPF filter "tcpdump port 22"
original BPF -> old JIT: original BPF -> internal BPF -> new JIT:
0: push %rbp 0: push %rbp
1: mov %rsp,%rbp 1: mov %rsp,%rbp
4: sub $0x60,%rsp 4: sub $0x228,%rsp
8: mov %rbx,-0x8(%rbp) b: mov %rbx,-0x228(%rbp) // prologue
12: mov %r13,-0x220(%rbp)
19: mov %r14,-0x218(%rbp)
20: mov %r15,-0x210(%rbp)
27: xor %eax,%eax // clear A
c: xor %ebx,%ebx 29: xor %r13,%r13 // clear X
e: mov 0x68(%rdi),%r9d 2c: mov 0x68(%rdi),%r9d
12: sub 0x6c(%rdi),%r9d 30: sub 0x6c(%rdi),%r9d
16: mov 0xd8(%rdi),%r8 34: mov 0xd8(%rdi),%r10
3b: mov %rdi,%rbx
1d: mov $0xc,%esi 3e: mov $0xc,%esi
22: callq 0xffffffffe1021e15 43: callq 0xffffffffe102bd75
27: cmp $0x86dd,%eax 48: cmp $0x86dd,%rax
2c: jne 0x0000000000000069 4f: jne 0x000000000000009a
2e: mov $0x14,%esi 51: mov $0x14,%esi
33: callq 0xffffffffe1021e31 56: callq 0xffffffffe102bd91
38: cmp $0x84,%eax 5b: cmp $0x84,%rax
3d: je 0x0000000000000049 62: je 0x0000000000000074
3f: cmp $0x6,%eax 64: cmp $0x6,%rax
42: je 0x0000000000000049 68: je 0x0000000000000074
44: cmp $0x11,%eax 6a: cmp $0x11,%rax
47: jne 0x00000000000000c6 6e: jne 0x0000000000000117
49: mov $0x36,%esi 74: mov $0x36,%esi
4e: callq 0xffffffffe1021e15 79: callq 0xffffffffe102bd75
53: cmp $0x16,%eax 7e: cmp $0x16,%rax
56: je 0x00000000000000bf 82: je 0x0000000000000110
58: mov $0x38,%esi 88: mov $0x38,%esi
5d: callq 0xffffffffe1021e15 8d: callq 0xffffffffe102bd75
62: cmp $0x16,%eax 92: cmp $0x16,%rax
65: je 0x00000000000000bf 96: je 0x0000000000000110
67: jmp 0x00000000000000c6 98: jmp 0x0000000000000117
69: cmp $0x800,%eax 9a: cmp $0x800,%rax
6e: jne 0x00000000000000c6 a1: jne 0x0000000000000117
70: mov $0x17,%esi a3: mov $0x17,%esi
75: callq 0xffffffffe1021e31 a8: callq 0xffffffffe102bd91
7a: cmp $0x84,%eax ad: cmp $0x84,%rax
7f: je 0x000000000000008b b4: je 0x00000000000000c2
81: cmp $0x6,%eax b6: cmp $0x6,%rax
84: je 0x000000000000008b ba: je 0x00000000000000c2
86: cmp $0x11,%eax bc: cmp $0x11,%rax
89: jne 0x00000000000000c6 c0: jne 0x0000000000000117
8b: mov $0x14,%esi c2: mov $0x14,%esi
90: callq 0xffffffffe1021e15 c7: callq 0xffffffffe102bd75
95: test $0x1fff,%ax cc: test $0x1fff,%rax
99: jne 0x00000000000000c6 d3: jne 0x0000000000000117
d5: mov %rax,%r14
9b: mov $0xe,%esi d8: mov $0xe,%esi
a0: callq 0xffffffffe1021e44 dd: callq 0xffffffffe102bd91 // MSH
e2: and $0xf,%eax
e5: shl $0x2,%eax
e8: mov %rax,%r13
eb: mov %r14,%rax
ee: mov %r13,%rsi
a5: lea 0xe(%rbx),%esi f1: add $0xe,%esi
a8: callq 0xffffffffe1021e0d f4: callq 0xffffffffe102bd6d
ad: cmp $0x16,%eax f9: cmp $0x16,%rax
b0: je 0x00000000000000bf fd: je 0x0000000000000110
ff: mov %r13,%rsi
b2: lea 0x10(%rbx),%esi 102: add $0x10,%esi
b5: callq 0xffffffffe1021e0d 105: callq 0xffffffffe102bd6d
ba: cmp $0x16,%eax 10a: cmp $0x16,%rax
bd: jne 0x00000000000000c6 10e: jne 0x0000000000000117
bf: mov $0xffff,%eax 110: mov $0xffff,%eax
c4: jmp 0x00000000000000c8 115: jmp 0x000000000000011c
c6: xor %eax,%eax 117: mov $0x0,%eax
c8: mov -0x8(%rbp),%rbx 11c: mov -0x228(%rbp),%rbx // epilogue
cc: leaveq 123: mov -0x220(%rbp),%r13
cd: retq 12a: mov -0x218(%rbp),%r14
131: mov -0x210(%rbp),%r15
138: leaveq
139: retq
On fully cached SKBs both JITed functions take 12 nsec to execute.
BPF interpreter executes the program in 30 nsec.
The difference in generated assembler is due to the following:
Old BPF imlements LDX_MSH instruction via sk_load_byte_msh() helper function
inside bpf_jit.S.
New JIT removes the helper and does it explicitly, so ldx_msh cost
is the same for both JITs, but generated code looks longer.
New JIT has 4 registers to save, so prologue/epilogue are larger,
but the cost is within noise on x64.
Old JIT checks whether first insn clears A and if not emits 'xor %eax,%eax'.
New JIT clears %rax unconditionally.
2. old BPF JIT doesn't support ANC_NLATTR, ANC_PAY_OFFSET, ANC_RANDOM
extensions. New JIT supports all BPF extensions.
Performance of such filters improves 2-4 times depending on a filter.
The longer the filter the higher performance gain.
Synthetic benchmarks with many ancillary loads see 20x speedup
which seems to be the maximum gain from JIT
Notes:
. net.core.bpf_jit_enable=2 + tools/net/bpf_jit_disasm is still functional
and can be used to see generated assembler
. there are two jit_compile() functions and code flow for classic filters is:
sk_attach_filter() - load classic BPF
bpf_jit_compile() - try to JIT from classic BPF
sk_convert_filter() - convert classic to internal
bpf_int_jit_compile() - JIT from internal BPF
seccomp and tracing filters will just call bpf_int_jit_compile()
Signed-off-by: Alexei Starovoitov <ast@plumgrid.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
2014-05-14 02:50:46 +00:00
|
|
|
EMIT1(0x66);
|
2014-06-06 21:46:06 +00:00
|
|
|
if (is_ereg(dst_reg))
|
net: filter: x86: internal BPF JIT
Maps all internal BPF instructions into x86_64 instructions.
This patch replaces original BPF x64 JIT with internal BPF x64 JIT.
sysctl net.core.bpf_jit_enable is reused as on/off switch.
Performance:
1. old BPF JIT and internal BPF JIT generate equivalent x86_64 code.
No performance difference is observed for filters that were JIT-able before
Example assembler code for BPF filter "tcpdump port 22"
original BPF -> old JIT: original BPF -> internal BPF -> new JIT:
0: push %rbp 0: push %rbp
1: mov %rsp,%rbp 1: mov %rsp,%rbp
4: sub $0x60,%rsp 4: sub $0x228,%rsp
8: mov %rbx,-0x8(%rbp) b: mov %rbx,-0x228(%rbp) // prologue
12: mov %r13,-0x220(%rbp)
19: mov %r14,-0x218(%rbp)
20: mov %r15,-0x210(%rbp)
27: xor %eax,%eax // clear A
c: xor %ebx,%ebx 29: xor %r13,%r13 // clear X
e: mov 0x68(%rdi),%r9d 2c: mov 0x68(%rdi),%r9d
12: sub 0x6c(%rdi),%r9d 30: sub 0x6c(%rdi),%r9d
16: mov 0xd8(%rdi),%r8 34: mov 0xd8(%rdi),%r10
3b: mov %rdi,%rbx
1d: mov $0xc,%esi 3e: mov $0xc,%esi
22: callq 0xffffffffe1021e15 43: callq 0xffffffffe102bd75
27: cmp $0x86dd,%eax 48: cmp $0x86dd,%rax
2c: jne 0x0000000000000069 4f: jne 0x000000000000009a
2e: mov $0x14,%esi 51: mov $0x14,%esi
33: callq 0xffffffffe1021e31 56: callq 0xffffffffe102bd91
38: cmp $0x84,%eax 5b: cmp $0x84,%rax
3d: je 0x0000000000000049 62: je 0x0000000000000074
3f: cmp $0x6,%eax 64: cmp $0x6,%rax
42: je 0x0000000000000049 68: je 0x0000000000000074
44: cmp $0x11,%eax 6a: cmp $0x11,%rax
47: jne 0x00000000000000c6 6e: jne 0x0000000000000117
49: mov $0x36,%esi 74: mov $0x36,%esi
4e: callq 0xffffffffe1021e15 79: callq 0xffffffffe102bd75
53: cmp $0x16,%eax 7e: cmp $0x16,%rax
56: je 0x00000000000000bf 82: je 0x0000000000000110
58: mov $0x38,%esi 88: mov $0x38,%esi
5d: callq 0xffffffffe1021e15 8d: callq 0xffffffffe102bd75
62: cmp $0x16,%eax 92: cmp $0x16,%rax
65: je 0x00000000000000bf 96: je 0x0000000000000110
67: jmp 0x00000000000000c6 98: jmp 0x0000000000000117
69: cmp $0x800,%eax 9a: cmp $0x800,%rax
6e: jne 0x00000000000000c6 a1: jne 0x0000000000000117
70: mov $0x17,%esi a3: mov $0x17,%esi
75: callq 0xffffffffe1021e31 a8: callq 0xffffffffe102bd91
7a: cmp $0x84,%eax ad: cmp $0x84,%rax
7f: je 0x000000000000008b b4: je 0x00000000000000c2
81: cmp $0x6,%eax b6: cmp $0x6,%rax
84: je 0x000000000000008b ba: je 0x00000000000000c2
86: cmp $0x11,%eax bc: cmp $0x11,%rax
89: jne 0x00000000000000c6 c0: jne 0x0000000000000117
8b: mov $0x14,%esi c2: mov $0x14,%esi
90: callq 0xffffffffe1021e15 c7: callq 0xffffffffe102bd75
95: test $0x1fff,%ax cc: test $0x1fff,%rax
99: jne 0x00000000000000c6 d3: jne 0x0000000000000117
d5: mov %rax,%r14
9b: mov $0xe,%esi d8: mov $0xe,%esi
a0: callq 0xffffffffe1021e44 dd: callq 0xffffffffe102bd91 // MSH
e2: and $0xf,%eax
e5: shl $0x2,%eax
e8: mov %rax,%r13
eb: mov %r14,%rax
ee: mov %r13,%rsi
a5: lea 0xe(%rbx),%esi f1: add $0xe,%esi
a8: callq 0xffffffffe1021e0d f4: callq 0xffffffffe102bd6d
ad: cmp $0x16,%eax f9: cmp $0x16,%rax
b0: je 0x00000000000000bf fd: je 0x0000000000000110
ff: mov %r13,%rsi
b2: lea 0x10(%rbx),%esi 102: add $0x10,%esi
b5: callq 0xffffffffe1021e0d 105: callq 0xffffffffe102bd6d
ba: cmp $0x16,%eax 10a: cmp $0x16,%rax
bd: jne 0x00000000000000c6 10e: jne 0x0000000000000117
bf: mov $0xffff,%eax 110: mov $0xffff,%eax
c4: jmp 0x00000000000000c8 115: jmp 0x000000000000011c
c6: xor %eax,%eax 117: mov $0x0,%eax
c8: mov -0x8(%rbp),%rbx 11c: mov -0x228(%rbp),%rbx // epilogue
cc: leaveq 123: mov -0x220(%rbp),%r13
cd: retq 12a: mov -0x218(%rbp),%r14
131: mov -0x210(%rbp),%r15
138: leaveq
139: retq
On fully cached SKBs both JITed functions take 12 nsec to execute.
BPF interpreter executes the program in 30 nsec.
The difference in generated assembler is due to the following:
Old BPF imlements LDX_MSH instruction via sk_load_byte_msh() helper function
inside bpf_jit.S.
New JIT removes the helper and does it explicitly, so ldx_msh cost
is the same for both JITs, but generated code looks longer.
New JIT has 4 registers to save, so prologue/epilogue are larger,
but the cost is within noise on x64.
Old JIT checks whether first insn clears A and if not emits 'xor %eax,%eax'.
New JIT clears %rax unconditionally.
2. old BPF JIT doesn't support ANC_NLATTR, ANC_PAY_OFFSET, ANC_RANDOM
extensions. New JIT supports all BPF extensions.
Performance of such filters improves 2-4 times depending on a filter.
The longer the filter the higher performance gain.
Synthetic benchmarks with many ancillary loads see 20x speedup
which seems to be the maximum gain from JIT
Notes:
. net.core.bpf_jit_enable=2 + tools/net/bpf_jit_disasm is still functional
and can be used to see generated assembler
. there are two jit_compile() functions and code flow for classic filters is:
sk_attach_filter() - load classic BPF
bpf_jit_compile() - try to JIT from classic BPF
sk_convert_filter() - convert classic to internal
bpf_int_jit_compile() - JIT from internal BPF
seccomp and tracing filters will just call bpf_int_jit_compile()
Signed-off-by: Alexei Starovoitov <ast@plumgrid.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
2014-05-14 02:50:46 +00:00
|
|
|
EMIT1(0x41);
|
2014-06-06 21:46:06 +00:00
|
|
|
EMIT3(0xC1, add_1reg(0xC8, dst_reg), 8);
|
2015-05-12 06:25:16 +00:00
|
|
|
|
x86/bpf: Clean up non-standard comments, to make the code more readable
So by chance I looked into x86 assembly in arch/x86/net/bpf_jit_comp.c and
noticed the weird and inconsistent comment style it mistakenly learned from
the networking code:
/* Multi-line comment ...
* ... looks like this.
*/
Fix this to use the standard comment style specified in Documentation/CodingStyle
and used in arch/x86/ as well:
/*
* Multi-line comment ...
* ... looks like this.
*/
Also, to quote Linus's ... more explicit views about this:
http://article.gmane.org/gmane.linux.kernel.cryptoapi/21066
> But no, the networking code picked *none* of the above sane formats.
> Instead, it picked these two models that are just half-arsed
> shit-for-brains:
>
> (no)
> /* This is disgusting drug-induced
> * crap, and should die
> */
>
> (no-no-no)
> /* This is also very nasty
> * and visually unbalanced */
>
> Please. The networking code actually has the *worst* possible comment
> style. You can literally find that (no-no-no) style, which is just
> really horribly disgusting and worse than the otherwise fairly similar
> (d) in pretty much every way.
Also improve the comments and some other details while at it:
- Don't mix same-line and previous-line comment style on otherwise
identical code patterns within the same function,
- capitalize 'BPF' and x86 register names consistently,
- capitalize sentences consistently,
- instead of 'x64' use 'x86-64': x64 is a Microsoft specific term,
- use more consistent punctuation,
- use standard coding style in macros as well,
- fix typos and a few other minor details.
Consistent coding style is not optional, at least in arch/x86/.
No change in functionality.
( In case this commit causes conflicts with pending development code
I'll be glad to help resolve any conflicts! )
Acked-by: Thomas Gleixner <tglx@linutronix.de>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Borislav Petkov <bp@alien8.de>
Cc: H. Peter Anvin <hpa@zytor.com>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: David S. Miller <davem@davemloft.net>
Cc: Eric Dumazet <edumazet@google.com>
Cc: Daniel Borkmann <daniel@iogearbox.net>
Cc: Alexei Starovoitov <ast@fb.com>
Cc: Hideaki YOSHIFUJI <yoshfuji@linux-ipv6.org>
Cc: netdev@vger.kernel.org
Cc: linux-kernel@vger.kernel.org
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
2018-04-27 09:54:40 +00:00
|
|
|
/* Emit 'movzwl eax, ax' */
|
2015-05-12 06:25:16 +00:00
|
|
|
if (is_ereg(dst_reg))
|
|
|
|
EMIT3(0x45, 0x0F, 0xB7);
|
|
|
|
else
|
|
|
|
EMIT2(0x0F, 0xB7);
|
|
|
|
EMIT1(add_2reg(0xC0, dst_reg, dst_reg));
|
net: filter: x86: internal BPF JIT
Maps all internal BPF instructions into x86_64 instructions.
This patch replaces original BPF x64 JIT with internal BPF x64 JIT.
sysctl net.core.bpf_jit_enable is reused as on/off switch.
Performance:
1. old BPF JIT and internal BPF JIT generate equivalent x86_64 code.
No performance difference is observed for filters that were JIT-able before
Example assembler code for BPF filter "tcpdump port 22"
original BPF -> old JIT: original BPF -> internal BPF -> new JIT:
0: push %rbp 0: push %rbp
1: mov %rsp,%rbp 1: mov %rsp,%rbp
4: sub $0x60,%rsp 4: sub $0x228,%rsp
8: mov %rbx,-0x8(%rbp) b: mov %rbx,-0x228(%rbp) // prologue
12: mov %r13,-0x220(%rbp)
19: mov %r14,-0x218(%rbp)
20: mov %r15,-0x210(%rbp)
27: xor %eax,%eax // clear A
c: xor %ebx,%ebx 29: xor %r13,%r13 // clear X
e: mov 0x68(%rdi),%r9d 2c: mov 0x68(%rdi),%r9d
12: sub 0x6c(%rdi),%r9d 30: sub 0x6c(%rdi),%r9d
16: mov 0xd8(%rdi),%r8 34: mov 0xd8(%rdi),%r10
3b: mov %rdi,%rbx
1d: mov $0xc,%esi 3e: mov $0xc,%esi
22: callq 0xffffffffe1021e15 43: callq 0xffffffffe102bd75
27: cmp $0x86dd,%eax 48: cmp $0x86dd,%rax
2c: jne 0x0000000000000069 4f: jne 0x000000000000009a
2e: mov $0x14,%esi 51: mov $0x14,%esi
33: callq 0xffffffffe1021e31 56: callq 0xffffffffe102bd91
38: cmp $0x84,%eax 5b: cmp $0x84,%rax
3d: je 0x0000000000000049 62: je 0x0000000000000074
3f: cmp $0x6,%eax 64: cmp $0x6,%rax
42: je 0x0000000000000049 68: je 0x0000000000000074
44: cmp $0x11,%eax 6a: cmp $0x11,%rax
47: jne 0x00000000000000c6 6e: jne 0x0000000000000117
49: mov $0x36,%esi 74: mov $0x36,%esi
4e: callq 0xffffffffe1021e15 79: callq 0xffffffffe102bd75
53: cmp $0x16,%eax 7e: cmp $0x16,%rax
56: je 0x00000000000000bf 82: je 0x0000000000000110
58: mov $0x38,%esi 88: mov $0x38,%esi
5d: callq 0xffffffffe1021e15 8d: callq 0xffffffffe102bd75
62: cmp $0x16,%eax 92: cmp $0x16,%rax
65: je 0x00000000000000bf 96: je 0x0000000000000110
67: jmp 0x00000000000000c6 98: jmp 0x0000000000000117
69: cmp $0x800,%eax 9a: cmp $0x800,%rax
6e: jne 0x00000000000000c6 a1: jne 0x0000000000000117
70: mov $0x17,%esi a3: mov $0x17,%esi
75: callq 0xffffffffe1021e31 a8: callq 0xffffffffe102bd91
7a: cmp $0x84,%eax ad: cmp $0x84,%rax
7f: je 0x000000000000008b b4: je 0x00000000000000c2
81: cmp $0x6,%eax b6: cmp $0x6,%rax
84: je 0x000000000000008b ba: je 0x00000000000000c2
86: cmp $0x11,%eax bc: cmp $0x11,%rax
89: jne 0x00000000000000c6 c0: jne 0x0000000000000117
8b: mov $0x14,%esi c2: mov $0x14,%esi
90: callq 0xffffffffe1021e15 c7: callq 0xffffffffe102bd75
95: test $0x1fff,%ax cc: test $0x1fff,%rax
99: jne 0x00000000000000c6 d3: jne 0x0000000000000117
d5: mov %rax,%r14
9b: mov $0xe,%esi d8: mov $0xe,%esi
a0: callq 0xffffffffe1021e44 dd: callq 0xffffffffe102bd91 // MSH
e2: and $0xf,%eax
e5: shl $0x2,%eax
e8: mov %rax,%r13
eb: mov %r14,%rax
ee: mov %r13,%rsi
a5: lea 0xe(%rbx),%esi f1: add $0xe,%esi
a8: callq 0xffffffffe1021e0d f4: callq 0xffffffffe102bd6d
ad: cmp $0x16,%eax f9: cmp $0x16,%rax
b0: je 0x00000000000000bf fd: je 0x0000000000000110
ff: mov %r13,%rsi
b2: lea 0x10(%rbx),%esi 102: add $0x10,%esi
b5: callq 0xffffffffe1021e0d 105: callq 0xffffffffe102bd6d
ba: cmp $0x16,%eax 10a: cmp $0x16,%rax
bd: jne 0x00000000000000c6 10e: jne 0x0000000000000117
bf: mov $0xffff,%eax 110: mov $0xffff,%eax
c4: jmp 0x00000000000000c8 115: jmp 0x000000000000011c
c6: xor %eax,%eax 117: mov $0x0,%eax
c8: mov -0x8(%rbp),%rbx 11c: mov -0x228(%rbp),%rbx // epilogue
cc: leaveq 123: mov -0x220(%rbp),%r13
cd: retq 12a: mov -0x218(%rbp),%r14
131: mov -0x210(%rbp),%r15
138: leaveq
139: retq
On fully cached SKBs both JITed functions take 12 nsec to execute.
BPF interpreter executes the program in 30 nsec.
The difference in generated assembler is due to the following:
Old BPF imlements LDX_MSH instruction via sk_load_byte_msh() helper function
inside bpf_jit.S.
New JIT removes the helper and does it explicitly, so ldx_msh cost
is the same for both JITs, but generated code looks longer.
New JIT has 4 registers to save, so prologue/epilogue are larger,
but the cost is within noise on x64.
Old JIT checks whether first insn clears A and if not emits 'xor %eax,%eax'.
New JIT clears %rax unconditionally.
2. old BPF JIT doesn't support ANC_NLATTR, ANC_PAY_OFFSET, ANC_RANDOM
extensions. New JIT supports all BPF extensions.
Performance of such filters improves 2-4 times depending on a filter.
The longer the filter the higher performance gain.
Synthetic benchmarks with many ancillary loads see 20x speedup
which seems to be the maximum gain from JIT
Notes:
. net.core.bpf_jit_enable=2 + tools/net/bpf_jit_disasm is still functional
and can be used to see generated assembler
. there are two jit_compile() functions and code flow for classic filters is:
sk_attach_filter() - load classic BPF
bpf_jit_compile() - try to JIT from classic BPF
sk_convert_filter() - convert classic to internal
bpf_int_jit_compile() - JIT from internal BPF
seccomp and tracing filters will just call bpf_int_jit_compile()
Signed-off-by: Alexei Starovoitov <ast@plumgrid.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
2014-05-14 02:50:46 +00:00
|
|
|
break;
|
|
|
|
case 32:
|
x86/bpf: Clean up non-standard comments, to make the code more readable
So by chance I looked into x86 assembly in arch/x86/net/bpf_jit_comp.c and
noticed the weird and inconsistent comment style it mistakenly learned from
the networking code:
/* Multi-line comment ...
* ... looks like this.
*/
Fix this to use the standard comment style specified in Documentation/CodingStyle
and used in arch/x86/ as well:
/*
* Multi-line comment ...
* ... looks like this.
*/
Also, to quote Linus's ... more explicit views about this:
http://article.gmane.org/gmane.linux.kernel.cryptoapi/21066
> But no, the networking code picked *none* of the above sane formats.
> Instead, it picked these two models that are just half-arsed
> shit-for-brains:
>
> (no)
> /* This is disgusting drug-induced
> * crap, and should die
> */
>
> (no-no-no)
> /* This is also very nasty
> * and visually unbalanced */
>
> Please. The networking code actually has the *worst* possible comment
> style. You can literally find that (no-no-no) style, which is just
> really horribly disgusting and worse than the otherwise fairly similar
> (d) in pretty much every way.
Also improve the comments and some other details while at it:
- Don't mix same-line and previous-line comment style on otherwise
identical code patterns within the same function,
- capitalize 'BPF' and x86 register names consistently,
- capitalize sentences consistently,
- instead of 'x64' use 'x86-64': x64 is a Microsoft specific term,
- use more consistent punctuation,
- use standard coding style in macros as well,
- fix typos and a few other minor details.
Consistent coding style is not optional, at least in arch/x86/.
No change in functionality.
( In case this commit causes conflicts with pending development code
I'll be glad to help resolve any conflicts! )
Acked-by: Thomas Gleixner <tglx@linutronix.de>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Borislav Petkov <bp@alien8.de>
Cc: H. Peter Anvin <hpa@zytor.com>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: David S. Miller <davem@davemloft.net>
Cc: Eric Dumazet <edumazet@google.com>
Cc: Daniel Borkmann <daniel@iogearbox.net>
Cc: Alexei Starovoitov <ast@fb.com>
Cc: Hideaki YOSHIFUJI <yoshfuji@linux-ipv6.org>
Cc: netdev@vger.kernel.org
Cc: linux-kernel@vger.kernel.org
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
2018-04-27 09:54:40 +00:00
|
|
|
/* Emit 'bswap eax' to swap lower 4 bytes */
|
2014-06-06 21:46:06 +00:00
|
|
|
if (is_ereg(dst_reg))
|
net: filter: x86: internal BPF JIT
Maps all internal BPF instructions into x86_64 instructions.
This patch replaces original BPF x64 JIT with internal BPF x64 JIT.
sysctl net.core.bpf_jit_enable is reused as on/off switch.
Performance:
1. old BPF JIT and internal BPF JIT generate equivalent x86_64 code.
No performance difference is observed for filters that were JIT-able before
Example assembler code for BPF filter "tcpdump port 22"
original BPF -> old JIT: original BPF -> internal BPF -> new JIT:
0: push %rbp 0: push %rbp
1: mov %rsp,%rbp 1: mov %rsp,%rbp
4: sub $0x60,%rsp 4: sub $0x228,%rsp
8: mov %rbx,-0x8(%rbp) b: mov %rbx,-0x228(%rbp) // prologue
12: mov %r13,-0x220(%rbp)
19: mov %r14,-0x218(%rbp)
20: mov %r15,-0x210(%rbp)
27: xor %eax,%eax // clear A
c: xor %ebx,%ebx 29: xor %r13,%r13 // clear X
e: mov 0x68(%rdi),%r9d 2c: mov 0x68(%rdi),%r9d
12: sub 0x6c(%rdi),%r9d 30: sub 0x6c(%rdi),%r9d
16: mov 0xd8(%rdi),%r8 34: mov 0xd8(%rdi),%r10
3b: mov %rdi,%rbx
1d: mov $0xc,%esi 3e: mov $0xc,%esi
22: callq 0xffffffffe1021e15 43: callq 0xffffffffe102bd75
27: cmp $0x86dd,%eax 48: cmp $0x86dd,%rax
2c: jne 0x0000000000000069 4f: jne 0x000000000000009a
2e: mov $0x14,%esi 51: mov $0x14,%esi
33: callq 0xffffffffe1021e31 56: callq 0xffffffffe102bd91
38: cmp $0x84,%eax 5b: cmp $0x84,%rax
3d: je 0x0000000000000049 62: je 0x0000000000000074
3f: cmp $0x6,%eax 64: cmp $0x6,%rax
42: je 0x0000000000000049 68: je 0x0000000000000074
44: cmp $0x11,%eax 6a: cmp $0x11,%rax
47: jne 0x00000000000000c6 6e: jne 0x0000000000000117
49: mov $0x36,%esi 74: mov $0x36,%esi
4e: callq 0xffffffffe1021e15 79: callq 0xffffffffe102bd75
53: cmp $0x16,%eax 7e: cmp $0x16,%rax
56: je 0x00000000000000bf 82: je 0x0000000000000110
58: mov $0x38,%esi 88: mov $0x38,%esi
5d: callq 0xffffffffe1021e15 8d: callq 0xffffffffe102bd75
62: cmp $0x16,%eax 92: cmp $0x16,%rax
65: je 0x00000000000000bf 96: je 0x0000000000000110
67: jmp 0x00000000000000c6 98: jmp 0x0000000000000117
69: cmp $0x800,%eax 9a: cmp $0x800,%rax
6e: jne 0x00000000000000c6 a1: jne 0x0000000000000117
70: mov $0x17,%esi a3: mov $0x17,%esi
75: callq 0xffffffffe1021e31 a8: callq 0xffffffffe102bd91
7a: cmp $0x84,%eax ad: cmp $0x84,%rax
7f: je 0x000000000000008b b4: je 0x00000000000000c2
81: cmp $0x6,%eax b6: cmp $0x6,%rax
84: je 0x000000000000008b ba: je 0x00000000000000c2
86: cmp $0x11,%eax bc: cmp $0x11,%rax
89: jne 0x00000000000000c6 c0: jne 0x0000000000000117
8b: mov $0x14,%esi c2: mov $0x14,%esi
90: callq 0xffffffffe1021e15 c7: callq 0xffffffffe102bd75
95: test $0x1fff,%ax cc: test $0x1fff,%rax
99: jne 0x00000000000000c6 d3: jne 0x0000000000000117
d5: mov %rax,%r14
9b: mov $0xe,%esi d8: mov $0xe,%esi
a0: callq 0xffffffffe1021e44 dd: callq 0xffffffffe102bd91 // MSH
e2: and $0xf,%eax
e5: shl $0x2,%eax
e8: mov %rax,%r13
eb: mov %r14,%rax
ee: mov %r13,%rsi
a5: lea 0xe(%rbx),%esi f1: add $0xe,%esi
a8: callq 0xffffffffe1021e0d f4: callq 0xffffffffe102bd6d
ad: cmp $0x16,%eax f9: cmp $0x16,%rax
b0: je 0x00000000000000bf fd: je 0x0000000000000110
ff: mov %r13,%rsi
b2: lea 0x10(%rbx),%esi 102: add $0x10,%esi
b5: callq 0xffffffffe1021e0d 105: callq 0xffffffffe102bd6d
ba: cmp $0x16,%eax 10a: cmp $0x16,%rax
bd: jne 0x00000000000000c6 10e: jne 0x0000000000000117
bf: mov $0xffff,%eax 110: mov $0xffff,%eax
c4: jmp 0x00000000000000c8 115: jmp 0x000000000000011c
c6: xor %eax,%eax 117: mov $0x0,%eax
c8: mov -0x8(%rbp),%rbx 11c: mov -0x228(%rbp),%rbx // epilogue
cc: leaveq 123: mov -0x220(%rbp),%r13
cd: retq 12a: mov -0x218(%rbp),%r14
131: mov -0x210(%rbp),%r15
138: leaveq
139: retq
On fully cached SKBs both JITed functions take 12 nsec to execute.
BPF interpreter executes the program in 30 nsec.
The difference in generated assembler is due to the following:
Old BPF imlements LDX_MSH instruction via sk_load_byte_msh() helper function
inside bpf_jit.S.
New JIT removes the helper and does it explicitly, so ldx_msh cost
is the same for both JITs, but generated code looks longer.
New JIT has 4 registers to save, so prologue/epilogue are larger,
but the cost is within noise on x64.
Old JIT checks whether first insn clears A and if not emits 'xor %eax,%eax'.
New JIT clears %rax unconditionally.
2. old BPF JIT doesn't support ANC_NLATTR, ANC_PAY_OFFSET, ANC_RANDOM
extensions. New JIT supports all BPF extensions.
Performance of such filters improves 2-4 times depending on a filter.
The longer the filter the higher performance gain.
Synthetic benchmarks with many ancillary loads see 20x speedup
which seems to be the maximum gain from JIT
Notes:
. net.core.bpf_jit_enable=2 + tools/net/bpf_jit_disasm is still functional
and can be used to see generated assembler
. there are two jit_compile() functions and code flow for classic filters is:
sk_attach_filter() - load classic BPF
bpf_jit_compile() - try to JIT from classic BPF
sk_convert_filter() - convert classic to internal
bpf_int_jit_compile() - JIT from internal BPF
seccomp and tracing filters will just call bpf_int_jit_compile()
Signed-off-by: Alexei Starovoitov <ast@plumgrid.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
2014-05-14 02:50:46 +00:00
|
|
|
EMIT2(0x41, 0x0F);
|
2011-04-20 09:27:32 +00:00
|
|
|
else
|
net: filter: x86: internal BPF JIT
Maps all internal BPF instructions into x86_64 instructions.
This patch replaces original BPF x64 JIT with internal BPF x64 JIT.
sysctl net.core.bpf_jit_enable is reused as on/off switch.
Performance:
1. old BPF JIT and internal BPF JIT generate equivalent x86_64 code.
No performance difference is observed for filters that were JIT-able before
Example assembler code for BPF filter "tcpdump port 22"
original BPF -> old JIT: original BPF -> internal BPF -> new JIT:
0: push %rbp 0: push %rbp
1: mov %rsp,%rbp 1: mov %rsp,%rbp
4: sub $0x60,%rsp 4: sub $0x228,%rsp
8: mov %rbx,-0x8(%rbp) b: mov %rbx,-0x228(%rbp) // prologue
12: mov %r13,-0x220(%rbp)
19: mov %r14,-0x218(%rbp)
20: mov %r15,-0x210(%rbp)
27: xor %eax,%eax // clear A
c: xor %ebx,%ebx 29: xor %r13,%r13 // clear X
e: mov 0x68(%rdi),%r9d 2c: mov 0x68(%rdi),%r9d
12: sub 0x6c(%rdi),%r9d 30: sub 0x6c(%rdi),%r9d
16: mov 0xd8(%rdi),%r8 34: mov 0xd8(%rdi),%r10
3b: mov %rdi,%rbx
1d: mov $0xc,%esi 3e: mov $0xc,%esi
22: callq 0xffffffffe1021e15 43: callq 0xffffffffe102bd75
27: cmp $0x86dd,%eax 48: cmp $0x86dd,%rax
2c: jne 0x0000000000000069 4f: jne 0x000000000000009a
2e: mov $0x14,%esi 51: mov $0x14,%esi
33: callq 0xffffffffe1021e31 56: callq 0xffffffffe102bd91
38: cmp $0x84,%eax 5b: cmp $0x84,%rax
3d: je 0x0000000000000049 62: je 0x0000000000000074
3f: cmp $0x6,%eax 64: cmp $0x6,%rax
42: je 0x0000000000000049 68: je 0x0000000000000074
44: cmp $0x11,%eax 6a: cmp $0x11,%rax
47: jne 0x00000000000000c6 6e: jne 0x0000000000000117
49: mov $0x36,%esi 74: mov $0x36,%esi
4e: callq 0xffffffffe1021e15 79: callq 0xffffffffe102bd75
53: cmp $0x16,%eax 7e: cmp $0x16,%rax
56: je 0x00000000000000bf 82: je 0x0000000000000110
58: mov $0x38,%esi 88: mov $0x38,%esi
5d: callq 0xffffffffe1021e15 8d: callq 0xffffffffe102bd75
62: cmp $0x16,%eax 92: cmp $0x16,%rax
65: je 0x00000000000000bf 96: je 0x0000000000000110
67: jmp 0x00000000000000c6 98: jmp 0x0000000000000117
69: cmp $0x800,%eax 9a: cmp $0x800,%rax
6e: jne 0x00000000000000c6 a1: jne 0x0000000000000117
70: mov $0x17,%esi a3: mov $0x17,%esi
75: callq 0xffffffffe1021e31 a8: callq 0xffffffffe102bd91
7a: cmp $0x84,%eax ad: cmp $0x84,%rax
7f: je 0x000000000000008b b4: je 0x00000000000000c2
81: cmp $0x6,%eax b6: cmp $0x6,%rax
84: je 0x000000000000008b ba: je 0x00000000000000c2
86: cmp $0x11,%eax bc: cmp $0x11,%rax
89: jne 0x00000000000000c6 c0: jne 0x0000000000000117
8b: mov $0x14,%esi c2: mov $0x14,%esi
90: callq 0xffffffffe1021e15 c7: callq 0xffffffffe102bd75
95: test $0x1fff,%ax cc: test $0x1fff,%rax
99: jne 0x00000000000000c6 d3: jne 0x0000000000000117
d5: mov %rax,%r14
9b: mov $0xe,%esi d8: mov $0xe,%esi
a0: callq 0xffffffffe1021e44 dd: callq 0xffffffffe102bd91 // MSH
e2: and $0xf,%eax
e5: shl $0x2,%eax
e8: mov %rax,%r13
eb: mov %r14,%rax
ee: mov %r13,%rsi
a5: lea 0xe(%rbx),%esi f1: add $0xe,%esi
a8: callq 0xffffffffe1021e0d f4: callq 0xffffffffe102bd6d
ad: cmp $0x16,%eax f9: cmp $0x16,%rax
b0: je 0x00000000000000bf fd: je 0x0000000000000110
ff: mov %r13,%rsi
b2: lea 0x10(%rbx),%esi 102: add $0x10,%esi
b5: callq 0xffffffffe1021e0d 105: callq 0xffffffffe102bd6d
ba: cmp $0x16,%eax 10a: cmp $0x16,%rax
bd: jne 0x00000000000000c6 10e: jne 0x0000000000000117
bf: mov $0xffff,%eax 110: mov $0xffff,%eax
c4: jmp 0x00000000000000c8 115: jmp 0x000000000000011c
c6: xor %eax,%eax 117: mov $0x0,%eax
c8: mov -0x8(%rbp),%rbx 11c: mov -0x228(%rbp),%rbx // epilogue
cc: leaveq 123: mov -0x220(%rbp),%r13
cd: retq 12a: mov -0x218(%rbp),%r14
131: mov -0x210(%rbp),%r15
138: leaveq
139: retq
On fully cached SKBs both JITed functions take 12 nsec to execute.
BPF interpreter executes the program in 30 nsec.
The difference in generated assembler is due to the following:
Old BPF imlements LDX_MSH instruction via sk_load_byte_msh() helper function
inside bpf_jit.S.
New JIT removes the helper and does it explicitly, so ldx_msh cost
is the same for both JITs, but generated code looks longer.
New JIT has 4 registers to save, so prologue/epilogue are larger,
but the cost is within noise on x64.
Old JIT checks whether first insn clears A and if not emits 'xor %eax,%eax'.
New JIT clears %rax unconditionally.
2. old BPF JIT doesn't support ANC_NLATTR, ANC_PAY_OFFSET, ANC_RANDOM
extensions. New JIT supports all BPF extensions.
Performance of such filters improves 2-4 times depending on a filter.
The longer the filter the higher performance gain.
Synthetic benchmarks with many ancillary loads see 20x speedup
which seems to be the maximum gain from JIT
Notes:
. net.core.bpf_jit_enable=2 + tools/net/bpf_jit_disasm is still functional
and can be used to see generated assembler
. there are two jit_compile() functions and code flow for classic filters is:
sk_attach_filter() - load classic BPF
bpf_jit_compile() - try to JIT from classic BPF
sk_convert_filter() - convert classic to internal
bpf_int_jit_compile() - JIT from internal BPF
seccomp and tracing filters will just call bpf_int_jit_compile()
Signed-off-by: Alexei Starovoitov <ast@plumgrid.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
2014-05-14 02:50:46 +00:00
|
|
|
EMIT1(0x0F);
|
2014-06-06 21:46:06 +00:00
|
|
|
EMIT1(add_1reg(0xC8, dst_reg));
|
2011-04-20 09:27:32 +00:00
|
|
|
break;
|
net: filter: x86: internal BPF JIT
Maps all internal BPF instructions into x86_64 instructions.
This patch replaces original BPF x64 JIT with internal BPF x64 JIT.
sysctl net.core.bpf_jit_enable is reused as on/off switch.
Performance:
1. old BPF JIT and internal BPF JIT generate equivalent x86_64 code.
No performance difference is observed for filters that were JIT-able before
Example assembler code for BPF filter "tcpdump port 22"
original BPF -> old JIT: original BPF -> internal BPF -> new JIT:
0: push %rbp 0: push %rbp
1: mov %rsp,%rbp 1: mov %rsp,%rbp
4: sub $0x60,%rsp 4: sub $0x228,%rsp
8: mov %rbx,-0x8(%rbp) b: mov %rbx,-0x228(%rbp) // prologue
12: mov %r13,-0x220(%rbp)
19: mov %r14,-0x218(%rbp)
20: mov %r15,-0x210(%rbp)
27: xor %eax,%eax // clear A
c: xor %ebx,%ebx 29: xor %r13,%r13 // clear X
e: mov 0x68(%rdi),%r9d 2c: mov 0x68(%rdi),%r9d
12: sub 0x6c(%rdi),%r9d 30: sub 0x6c(%rdi),%r9d
16: mov 0xd8(%rdi),%r8 34: mov 0xd8(%rdi),%r10
3b: mov %rdi,%rbx
1d: mov $0xc,%esi 3e: mov $0xc,%esi
22: callq 0xffffffffe1021e15 43: callq 0xffffffffe102bd75
27: cmp $0x86dd,%eax 48: cmp $0x86dd,%rax
2c: jne 0x0000000000000069 4f: jne 0x000000000000009a
2e: mov $0x14,%esi 51: mov $0x14,%esi
33: callq 0xffffffffe1021e31 56: callq 0xffffffffe102bd91
38: cmp $0x84,%eax 5b: cmp $0x84,%rax
3d: je 0x0000000000000049 62: je 0x0000000000000074
3f: cmp $0x6,%eax 64: cmp $0x6,%rax
42: je 0x0000000000000049 68: je 0x0000000000000074
44: cmp $0x11,%eax 6a: cmp $0x11,%rax
47: jne 0x00000000000000c6 6e: jne 0x0000000000000117
49: mov $0x36,%esi 74: mov $0x36,%esi
4e: callq 0xffffffffe1021e15 79: callq 0xffffffffe102bd75
53: cmp $0x16,%eax 7e: cmp $0x16,%rax
56: je 0x00000000000000bf 82: je 0x0000000000000110
58: mov $0x38,%esi 88: mov $0x38,%esi
5d: callq 0xffffffffe1021e15 8d: callq 0xffffffffe102bd75
62: cmp $0x16,%eax 92: cmp $0x16,%rax
65: je 0x00000000000000bf 96: je 0x0000000000000110
67: jmp 0x00000000000000c6 98: jmp 0x0000000000000117
69: cmp $0x800,%eax 9a: cmp $0x800,%rax
6e: jne 0x00000000000000c6 a1: jne 0x0000000000000117
70: mov $0x17,%esi a3: mov $0x17,%esi
75: callq 0xffffffffe1021e31 a8: callq 0xffffffffe102bd91
7a: cmp $0x84,%eax ad: cmp $0x84,%rax
7f: je 0x000000000000008b b4: je 0x00000000000000c2
81: cmp $0x6,%eax b6: cmp $0x6,%rax
84: je 0x000000000000008b ba: je 0x00000000000000c2
86: cmp $0x11,%eax bc: cmp $0x11,%rax
89: jne 0x00000000000000c6 c0: jne 0x0000000000000117
8b: mov $0x14,%esi c2: mov $0x14,%esi
90: callq 0xffffffffe1021e15 c7: callq 0xffffffffe102bd75
95: test $0x1fff,%ax cc: test $0x1fff,%rax
99: jne 0x00000000000000c6 d3: jne 0x0000000000000117
d5: mov %rax,%r14
9b: mov $0xe,%esi d8: mov $0xe,%esi
a0: callq 0xffffffffe1021e44 dd: callq 0xffffffffe102bd91 // MSH
e2: and $0xf,%eax
e5: shl $0x2,%eax
e8: mov %rax,%r13
eb: mov %r14,%rax
ee: mov %r13,%rsi
a5: lea 0xe(%rbx),%esi f1: add $0xe,%esi
a8: callq 0xffffffffe1021e0d f4: callq 0xffffffffe102bd6d
ad: cmp $0x16,%eax f9: cmp $0x16,%rax
b0: je 0x00000000000000bf fd: je 0x0000000000000110
ff: mov %r13,%rsi
b2: lea 0x10(%rbx),%esi 102: add $0x10,%esi
b5: callq 0xffffffffe1021e0d 105: callq 0xffffffffe102bd6d
ba: cmp $0x16,%eax 10a: cmp $0x16,%rax
bd: jne 0x00000000000000c6 10e: jne 0x0000000000000117
bf: mov $0xffff,%eax 110: mov $0xffff,%eax
c4: jmp 0x00000000000000c8 115: jmp 0x000000000000011c
c6: xor %eax,%eax 117: mov $0x0,%eax
c8: mov -0x8(%rbp),%rbx 11c: mov -0x228(%rbp),%rbx // epilogue
cc: leaveq 123: mov -0x220(%rbp),%r13
cd: retq 12a: mov -0x218(%rbp),%r14
131: mov -0x210(%rbp),%r15
138: leaveq
139: retq
On fully cached SKBs both JITed functions take 12 nsec to execute.
BPF interpreter executes the program in 30 nsec.
The difference in generated assembler is due to the following:
Old BPF imlements LDX_MSH instruction via sk_load_byte_msh() helper function
inside bpf_jit.S.
New JIT removes the helper and does it explicitly, so ldx_msh cost
is the same for both JITs, but generated code looks longer.
New JIT has 4 registers to save, so prologue/epilogue are larger,
but the cost is within noise on x64.
Old JIT checks whether first insn clears A and if not emits 'xor %eax,%eax'.
New JIT clears %rax unconditionally.
2. old BPF JIT doesn't support ANC_NLATTR, ANC_PAY_OFFSET, ANC_RANDOM
extensions. New JIT supports all BPF extensions.
Performance of such filters improves 2-4 times depending on a filter.
The longer the filter the higher performance gain.
Synthetic benchmarks with many ancillary loads see 20x speedup
which seems to be the maximum gain from JIT
Notes:
. net.core.bpf_jit_enable=2 + tools/net/bpf_jit_disasm is still functional
and can be used to see generated assembler
. there are two jit_compile() functions and code flow for classic filters is:
sk_attach_filter() - load classic BPF
bpf_jit_compile() - try to JIT from classic BPF
sk_convert_filter() - convert classic to internal
bpf_int_jit_compile() - JIT from internal BPF
seccomp and tracing filters will just call bpf_int_jit_compile()
Signed-off-by: Alexei Starovoitov <ast@plumgrid.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
2014-05-14 02:50:46 +00:00
|
|
|
case 64:
|
x86/bpf: Clean up non-standard comments, to make the code more readable
So by chance I looked into x86 assembly in arch/x86/net/bpf_jit_comp.c and
noticed the weird and inconsistent comment style it mistakenly learned from
the networking code:
/* Multi-line comment ...
* ... looks like this.
*/
Fix this to use the standard comment style specified in Documentation/CodingStyle
and used in arch/x86/ as well:
/*
* Multi-line comment ...
* ... looks like this.
*/
Also, to quote Linus's ... more explicit views about this:
http://article.gmane.org/gmane.linux.kernel.cryptoapi/21066
> But no, the networking code picked *none* of the above sane formats.
> Instead, it picked these two models that are just half-arsed
> shit-for-brains:
>
> (no)
> /* This is disgusting drug-induced
> * crap, and should die
> */
>
> (no-no-no)
> /* This is also very nasty
> * and visually unbalanced */
>
> Please. The networking code actually has the *worst* possible comment
> style. You can literally find that (no-no-no) style, which is just
> really horribly disgusting and worse than the otherwise fairly similar
> (d) in pretty much every way.
Also improve the comments and some other details while at it:
- Don't mix same-line and previous-line comment style on otherwise
identical code patterns within the same function,
- capitalize 'BPF' and x86 register names consistently,
- capitalize sentences consistently,
- instead of 'x64' use 'x86-64': x64 is a Microsoft specific term,
- use more consistent punctuation,
- use standard coding style in macros as well,
- fix typos and a few other minor details.
Consistent coding style is not optional, at least in arch/x86/.
No change in functionality.
( In case this commit causes conflicts with pending development code
I'll be glad to help resolve any conflicts! )
Acked-by: Thomas Gleixner <tglx@linutronix.de>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Borislav Petkov <bp@alien8.de>
Cc: H. Peter Anvin <hpa@zytor.com>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: David S. Miller <davem@davemloft.net>
Cc: Eric Dumazet <edumazet@google.com>
Cc: Daniel Borkmann <daniel@iogearbox.net>
Cc: Alexei Starovoitov <ast@fb.com>
Cc: Hideaki YOSHIFUJI <yoshfuji@linux-ipv6.org>
Cc: netdev@vger.kernel.org
Cc: linux-kernel@vger.kernel.org
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
2018-04-27 09:54:40 +00:00
|
|
|
/* Emit 'bswap rax' to swap 8 bytes */
|
2014-06-06 21:46:06 +00:00
|
|
|
EMIT3(add_1mod(0x48, dst_reg), 0x0F,
|
|
|
|
add_1reg(0xC8, dst_reg));
|
2013-01-31 01:51:44 +00:00
|
|
|
break;
|
|
|
|
}
|
net: filter: x86: internal BPF JIT
Maps all internal BPF instructions into x86_64 instructions.
This patch replaces original BPF x64 JIT with internal BPF x64 JIT.
sysctl net.core.bpf_jit_enable is reused as on/off switch.
Performance:
1. old BPF JIT and internal BPF JIT generate equivalent x86_64 code.
No performance difference is observed for filters that were JIT-able before
Example assembler code for BPF filter "tcpdump port 22"
original BPF -> old JIT: original BPF -> internal BPF -> new JIT:
0: push %rbp 0: push %rbp
1: mov %rsp,%rbp 1: mov %rsp,%rbp
4: sub $0x60,%rsp 4: sub $0x228,%rsp
8: mov %rbx,-0x8(%rbp) b: mov %rbx,-0x228(%rbp) // prologue
12: mov %r13,-0x220(%rbp)
19: mov %r14,-0x218(%rbp)
20: mov %r15,-0x210(%rbp)
27: xor %eax,%eax // clear A
c: xor %ebx,%ebx 29: xor %r13,%r13 // clear X
e: mov 0x68(%rdi),%r9d 2c: mov 0x68(%rdi),%r9d
12: sub 0x6c(%rdi),%r9d 30: sub 0x6c(%rdi),%r9d
16: mov 0xd8(%rdi),%r8 34: mov 0xd8(%rdi),%r10
3b: mov %rdi,%rbx
1d: mov $0xc,%esi 3e: mov $0xc,%esi
22: callq 0xffffffffe1021e15 43: callq 0xffffffffe102bd75
27: cmp $0x86dd,%eax 48: cmp $0x86dd,%rax
2c: jne 0x0000000000000069 4f: jne 0x000000000000009a
2e: mov $0x14,%esi 51: mov $0x14,%esi
33: callq 0xffffffffe1021e31 56: callq 0xffffffffe102bd91
38: cmp $0x84,%eax 5b: cmp $0x84,%rax
3d: je 0x0000000000000049 62: je 0x0000000000000074
3f: cmp $0x6,%eax 64: cmp $0x6,%rax
42: je 0x0000000000000049 68: je 0x0000000000000074
44: cmp $0x11,%eax 6a: cmp $0x11,%rax
47: jne 0x00000000000000c6 6e: jne 0x0000000000000117
49: mov $0x36,%esi 74: mov $0x36,%esi
4e: callq 0xffffffffe1021e15 79: callq 0xffffffffe102bd75
53: cmp $0x16,%eax 7e: cmp $0x16,%rax
56: je 0x00000000000000bf 82: je 0x0000000000000110
58: mov $0x38,%esi 88: mov $0x38,%esi
5d: callq 0xffffffffe1021e15 8d: callq 0xffffffffe102bd75
62: cmp $0x16,%eax 92: cmp $0x16,%rax
65: je 0x00000000000000bf 96: je 0x0000000000000110
67: jmp 0x00000000000000c6 98: jmp 0x0000000000000117
69: cmp $0x800,%eax 9a: cmp $0x800,%rax
6e: jne 0x00000000000000c6 a1: jne 0x0000000000000117
70: mov $0x17,%esi a3: mov $0x17,%esi
75: callq 0xffffffffe1021e31 a8: callq 0xffffffffe102bd91
7a: cmp $0x84,%eax ad: cmp $0x84,%rax
7f: je 0x000000000000008b b4: je 0x00000000000000c2
81: cmp $0x6,%eax b6: cmp $0x6,%rax
84: je 0x000000000000008b ba: je 0x00000000000000c2
86: cmp $0x11,%eax bc: cmp $0x11,%rax
89: jne 0x00000000000000c6 c0: jne 0x0000000000000117
8b: mov $0x14,%esi c2: mov $0x14,%esi
90: callq 0xffffffffe1021e15 c7: callq 0xffffffffe102bd75
95: test $0x1fff,%ax cc: test $0x1fff,%rax
99: jne 0x00000000000000c6 d3: jne 0x0000000000000117
d5: mov %rax,%r14
9b: mov $0xe,%esi d8: mov $0xe,%esi
a0: callq 0xffffffffe1021e44 dd: callq 0xffffffffe102bd91 // MSH
e2: and $0xf,%eax
e5: shl $0x2,%eax
e8: mov %rax,%r13
eb: mov %r14,%rax
ee: mov %r13,%rsi
a5: lea 0xe(%rbx),%esi f1: add $0xe,%esi
a8: callq 0xffffffffe1021e0d f4: callq 0xffffffffe102bd6d
ad: cmp $0x16,%eax f9: cmp $0x16,%rax
b0: je 0x00000000000000bf fd: je 0x0000000000000110
ff: mov %r13,%rsi
b2: lea 0x10(%rbx),%esi 102: add $0x10,%esi
b5: callq 0xffffffffe1021e0d 105: callq 0xffffffffe102bd6d
ba: cmp $0x16,%eax 10a: cmp $0x16,%rax
bd: jne 0x00000000000000c6 10e: jne 0x0000000000000117
bf: mov $0xffff,%eax 110: mov $0xffff,%eax
c4: jmp 0x00000000000000c8 115: jmp 0x000000000000011c
c6: xor %eax,%eax 117: mov $0x0,%eax
c8: mov -0x8(%rbp),%rbx 11c: mov -0x228(%rbp),%rbx // epilogue
cc: leaveq 123: mov -0x220(%rbp),%r13
cd: retq 12a: mov -0x218(%rbp),%r14
131: mov -0x210(%rbp),%r15
138: leaveq
139: retq
On fully cached SKBs both JITed functions take 12 nsec to execute.
BPF interpreter executes the program in 30 nsec.
The difference in generated assembler is due to the following:
Old BPF imlements LDX_MSH instruction via sk_load_byte_msh() helper function
inside bpf_jit.S.
New JIT removes the helper and does it explicitly, so ldx_msh cost
is the same for both JITs, but generated code looks longer.
New JIT has 4 registers to save, so prologue/epilogue are larger,
but the cost is within noise on x64.
Old JIT checks whether first insn clears A and if not emits 'xor %eax,%eax'.
New JIT clears %rax unconditionally.
2. old BPF JIT doesn't support ANC_NLATTR, ANC_PAY_OFFSET, ANC_RANDOM
extensions. New JIT supports all BPF extensions.
Performance of such filters improves 2-4 times depending on a filter.
The longer the filter the higher performance gain.
Synthetic benchmarks with many ancillary loads see 20x speedup
which seems to be the maximum gain from JIT
Notes:
. net.core.bpf_jit_enable=2 + tools/net/bpf_jit_disasm is still functional
and can be used to see generated assembler
. there are two jit_compile() functions and code flow for classic filters is:
sk_attach_filter() - load classic BPF
bpf_jit_compile() - try to JIT from classic BPF
sk_convert_filter() - convert classic to internal
bpf_int_jit_compile() - JIT from internal BPF
seccomp and tracing filters will just call bpf_int_jit_compile()
Signed-off-by: Alexei Starovoitov <ast@plumgrid.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
2014-05-14 02:50:46 +00:00
|
|
|
break;
|
|
|
|
|
|
|
|
case BPF_ALU | BPF_END | BPF_FROM_LE:
|
2015-05-12 06:25:16 +00:00
|
|
|
switch (imm32) {
|
|
|
|
case 16:
|
x86/bpf: Clean up non-standard comments, to make the code more readable
So by chance I looked into x86 assembly in arch/x86/net/bpf_jit_comp.c and
noticed the weird and inconsistent comment style it mistakenly learned from
the networking code:
/* Multi-line comment ...
* ... looks like this.
*/
Fix this to use the standard comment style specified in Documentation/CodingStyle
and used in arch/x86/ as well:
/*
* Multi-line comment ...
* ... looks like this.
*/
Also, to quote Linus's ... more explicit views about this:
http://article.gmane.org/gmane.linux.kernel.cryptoapi/21066
> But no, the networking code picked *none* of the above sane formats.
> Instead, it picked these two models that are just half-arsed
> shit-for-brains:
>
> (no)
> /* This is disgusting drug-induced
> * crap, and should die
> */
>
> (no-no-no)
> /* This is also very nasty
> * and visually unbalanced */
>
> Please. The networking code actually has the *worst* possible comment
> style. You can literally find that (no-no-no) style, which is just
> really horribly disgusting and worse than the otherwise fairly similar
> (d) in pretty much every way.
Also improve the comments and some other details while at it:
- Don't mix same-line and previous-line comment style on otherwise
identical code patterns within the same function,
- capitalize 'BPF' and x86 register names consistently,
- capitalize sentences consistently,
- instead of 'x64' use 'x86-64': x64 is a Microsoft specific term,
- use more consistent punctuation,
- use standard coding style in macros as well,
- fix typos and a few other minor details.
Consistent coding style is not optional, at least in arch/x86/.
No change in functionality.
( In case this commit causes conflicts with pending development code
I'll be glad to help resolve any conflicts! )
Acked-by: Thomas Gleixner <tglx@linutronix.de>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Borislav Petkov <bp@alien8.de>
Cc: H. Peter Anvin <hpa@zytor.com>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: David S. Miller <davem@davemloft.net>
Cc: Eric Dumazet <edumazet@google.com>
Cc: Daniel Borkmann <daniel@iogearbox.net>
Cc: Alexei Starovoitov <ast@fb.com>
Cc: Hideaki YOSHIFUJI <yoshfuji@linux-ipv6.org>
Cc: netdev@vger.kernel.org
Cc: linux-kernel@vger.kernel.org
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
2018-04-27 09:54:40 +00:00
|
|
|
/*
|
|
|
|
* Emit 'movzwl eax, ax' to zero extend 16-bit
|
2015-05-12 06:25:16 +00:00
|
|
|
* into 64 bit
|
|
|
|
*/
|
|
|
|
if (is_ereg(dst_reg))
|
|
|
|
EMIT3(0x45, 0x0F, 0xB7);
|
|
|
|
else
|
|
|
|
EMIT2(0x0F, 0xB7);
|
|
|
|
EMIT1(add_2reg(0xC0, dst_reg, dst_reg));
|
|
|
|
break;
|
|
|
|
case 32:
|
x86/bpf: Clean up non-standard comments, to make the code more readable
So by chance I looked into x86 assembly in arch/x86/net/bpf_jit_comp.c and
noticed the weird and inconsistent comment style it mistakenly learned from
the networking code:
/* Multi-line comment ...
* ... looks like this.
*/
Fix this to use the standard comment style specified in Documentation/CodingStyle
and used in arch/x86/ as well:
/*
* Multi-line comment ...
* ... looks like this.
*/
Also, to quote Linus's ... more explicit views about this:
http://article.gmane.org/gmane.linux.kernel.cryptoapi/21066
> But no, the networking code picked *none* of the above sane formats.
> Instead, it picked these two models that are just half-arsed
> shit-for-brains:
>
> (no)
> /* This is disgusting drug-induced
> * crap, and should die
> */
>
> (no-no-no)
> /* This is also very nasty
> * and visually unbalanced */
>
> Please. The networking code actually has the *worst* possible comment
> style. You can literally find that (no-no-no) style, which is just
> really horribly disgusting and worse than the otherwise fairly similar
> (d) in pretty much every way.
Also improve the comments and some other details while at it:
- Don't mix same-line and previous-line comment style on otherwise
identical code patterns within the same function,
- capitalize 'BPF' and x86 register names consistently,
- capitalize sentences consistently,
- instead of 'x64' use 'x86-64': x64 is a Microsoft specific term,
- use more consistent punctuation,
- use standard coding style in macros as well,
- fix typos and a few other minor details.
Consistent coding style is not optional, at least in arch/x86/.
No change in functionality.
( In case this commit causes conflicts with pending development code
I'll be glad to help resolve any conflicts! )
Acked-by: Thomas Gleixner <tglx@linutronix.de>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Borislav Petkov <bp@alien8.de>
Cc: H. Peter Anvin <hpa@zytor.com>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: David S. Miller <davem@davemloft.net>
Cc: Eric Dumazet <edumazet@google.com>
Cc: Daniel Borkmann <daniel@iogearbox.net>
Cc: Alexei Starovoitov <ast@fb.com>
Cc: Hideaki YOSHIFUJI <yoshfuji@linux-ipv6.org>
Cc: netdev@vger.kernel.org
Cc: linux-kernel@vger.kernel.org
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
2018-04-27 09:54:40 +00:00
|
|
|
/* Emit 'mov eax, eax' to clear upper 32-bits */
|
2015-05-12 06:25:16 +00:00
|
|
|
if (is_ereg(dst_reg))
|
|
|
|
EMIT1(0x45);
|
|
|
|
EMIT2(0x89, add_2reg(0xC0, dst_reg, dst_reg));
|
|
|
|
break;
|
|
|
|
case 64:
|
|
|
|
/* nop */
|
|
|
|
break;
|
|
|
|
}
|
net: filter: x86: internal BPF JIT
Maps all internal BPF instructions into x86_64 instructions.
This patch replaces original BPF x64 JIT with internal BPF x64 JIT.
sysctl net.core.bpf_jit_enable is reused as on/off switch.
Performance:
1. old BPF JIT and internal BPF JIT generate equivalent x86_64 code.
No performance difference is observed for filters that were JIT-able before
Example assembler code for BPF filter "tcpdump port 22"
original BPF -> old JIT: original BPF -> internal BPF -> new JIT:
0: push %rbp 0: push %rbp
1: mov %rsp,%rbp 1: mov %rsp,%rbp
4: sub $0x60,%rsp 4: sub $0x228,%rsp
8: mov %rbx,-0x8(%rbp) b: mov %rbx,-0x228(%rbp) // prologue
12: mov %r13,-0x220(%rbp)
19: mov %r14,-0x218(%rbp)
20: mov %r15,-0x210(%rbp)
27: xor %eax,%eax // clear A
c: xor %ebx,%ebx 29: xor %r13,%r13 // clear X
e: mov 0x68(%rdi),%r9d 2c: mov 0x68(%rdi),%r9d
12: sub 0x6c(%rdi),%r9d 30: sub 0x6c(%rdi),%r9d
16: mov 0xd8(%rdi),%r8 34: mov 0xd8(%rdi),%r10
3b: mov %rdi,%rbx
1d: mov $0xc,%esi 3e: mov $0xc,%esi
22: callq 0xffffffffe1021e15 43: callq 0xffffffffe102bd75
27: cmp $0x86dd,%eax 48: cmp $0x86dd,%rax
2c: jne 0x0000000000000069 4f: jne 0x000000000000009a
2e: mov $0x14,%esi 51: mov $0x14,%esi
33: callq 0xffffffffe1021e31 56: callq 0xffffffffe102bd91
38: cmp $0x84,%eax 5b: cmp $0x84,%rax
3d: je 0x0000000000000049 62: je 0x0000000000000074
3f: cmp $0x6,%eax 64: cmp $0x6,%rax
42: je 0x0000000000000049 68: je 0x0000000000000074
44: cmp $0x11,%eax 6a: cmp $0x11,%rax
47: jne 0x00000000000000c6 6e: jne 0x0000000000000117
49: mov $0x36,%esi 74: mov $0x36,%esi
4e: callq 0xffffffffe1021e15 79: callq 0xffffffffe102bd75
53: cmp $0x16,%eax 7e: cmp $0x16,%rax
56: je 0x00000000000000bf 82: je 0x0000000000000110
58: mov $0x38,%esi 88: mov $0x38,%esi
5d: callq 0xffffffffe1021e15 8d: callq 0xffffffffe102bd75
62: cmp $0x16,%eax 92: cmp $0x16,%rax
65: je 0x00000000000000bf 96: je 0x0000000000000110
67: jmp 0x00000000000000c6 98: jmp 0x0000000000000117
69: cmp $0x800,%eax 9a: cmp $0x800,%rax
6e: jne 0x00000000000000c6 a1: jne 0x0000000000000117
70: mov $0x17,%esi a3: mov $0x17,%esi
75: callq 0xffffffffe1021e31 a8: callq 0xffffffffe102bd91
7a: cmp $0x84,%eax ad: cmp $0x84,%rax
7f: je 0x000000000000008b b4: je 0x00000000000000c2
81: cmp $0x6,%eax b6: cmp $0x6,%rax
84: je 0x000000000000008b ba: je 0x00000000000000c2
86: cmp $0x11,%eax bc: cmp $0x11,%rax
89: jne 0x00000000000000c6 c0: jne 0x0000000000000117
8b: mov $0x14,%esi c2: mov $0x14,%esi
90: callq 0xffffffffe1021e15 c7: callq 0xffffffffe102bd75
95: test $0x1fff,%ax cc: test $0x1fff,%rax
99: jne 0x00000000000000c6 d3: jne 0x0000000000000117
d5: mov %rax,%r14
9b: mov $0xe,%esi d8: mov $0xe,%esi
a0: callq 0xffffffffe1021e44 dd: callq 0xffffffffe102bd91 // MSH
e2: and $0xf,%eax
e5: shl $0x2,%eax
e8: mov %rax,%r13
eb: mov %r14,%rax
ee: mov %r13,%rsi
a5: lea 0xe(%rbx),%esi f1: add $0xe,%esi
a8: callq 0xffffffffe1021e0d f4: callq 0xffffffffe102bd6d
ad: cmp $0x16,%eax f9: cmp $0x16,%rax
b0: je 0x00000000000000bf fd: je 0x0000000000000110
ff: mov %r13,%rsi
b2: lea 0x10(%rbx),%esi 102: add $0x10,%esi
b5: callq 0xffffffffe1021e0d 105: callq 0xffffffffe102bd6d
ba: cmp $0x16,%eax 10a: cmp $0x16,%rax
bd: jne 0x00000000000000c6 10e: jne 0x0000000000000117
bf: mov $0xffff,%eax 110: mov $0xffff,%eax
c4: jmp 0x00000000000000c8 115: jmp 0x000000000000011c
c6: xor %eax,%eax 117: mov $0x0,%eax
c8: mov -0x8(%rbp),%rbx 11c: mov -0x228(%rbp),%rbx // epilogue
cc: leaveq 123: mov -0x220(%rbp),%r13
cd: retq 12a: mov -0x218(%rbp),%r14
131: mov -0x210(%rbp),%r15
138: leaveq
139: retq
On fully cached SKBs both JITed functions take 12 nsec to execute.
BPF interpreter executes the program in 30 nsec.
The difference in generated assembler is due to the following:
Old BPF imlements LDX_MSH instruction via sk_load_byte_msh() helper function
inside bpf_jit.S.
New JIT removes the helper and does it explicitly, so ldx_msh cost
is the same for both JITs, but generated code looks longer.
New JIT has 4 registers to save, so prologue/epilogue are larger,
but the cost is within noise on x64.
Old JIT checks whether first insn clears A and if not emits 'xor %eax,%eax'.
New JIT clears %rax unconditionally.
2. old BPF JIT doesn't support ANC_NLATTR, ANC_PAY_OFFSET, ANC_RANDOM
extensions. New JIT supports all BPF extensions.
Performance of such filters improves 2-4 times depending on a filter.
The longer the filter the higher performance gain.
Synthetic benchmarks with many ancillary loads see 20x speedup
which seems to be the maximum gain from JIT
Notes:
. net.core.bpf_jit_enable=2 + tools/net/bpf_jit_disasm is still functional
and can be used to see generated assembler
. there are two jit_compile() functions and code flow for classic filters is:
sk_attach_filter() - load classic BPF
bpf_jit_compile() - try to JIT from classic BPF
sk_convert_filter() - convert classic to internal
bpf_int_jit_compile() - JIT from internal BPF
seccomp and tracing filters will just call bpf_int_jit_compile()
Signed-off-by: Alexei Starovoitov <ast@plumgrid.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
2014-05-14 02:50:46 +00:00
|
|
|
break;
|
|
|
|
|
2021-07-13 08:18:31 +00:00
|
|
|
/* speculation barrier */
|
|
|
|
case BPF_ST | BPF_NOSPEC:
|
2022-10-05 17:00:39 +00:00
|
|
|
EMIT_LFENCE();
|
2021-07-13 08:18:31 +00:00
|
|
|
break;
|
|
|
|
|
2014-06-06 21:46:06 +00:00
|
|
|
/* ST: *(u8*)(dst_reg + off) = imm */
|
net: filter: x86: internal BPF JIT
Maps all internal BPF instructions into x86_64 instructions.
This patch replaces original BPF x64 JIT with internal BPF x64 JIT.
sysctl net.core.bpf_jit_enable is reused as on/off switch.
Performance:
1. old BPF JIT and internal BPF JIT generate equivalent x86_64 code.
No performance difference is observed for filters that were JIT-able before
Example assembler code for BPF filter "tcpdump port 22"
original BPF -> old JIT: original BPF -> internal BPF -> new JIT:
0: push %rbp 0: push %rbp
1: mov %rsp,%rbp 1: mov %rsp,%rbp
4: sub $0x60,%rsp 4: sub $0x228,%rsp
8: mov %rbx,-0x8(%rbp) b: mov %rbx,-0x228(%rbp) // prologue
12: mov %r13,-0x220(%rbp)
19: mov %r14,-0x218(%rbp)
20: mov %r15,-0x210(%rbp)
27: xor %eax,%eax // clear A
c: xor %ebx,%ebx 29: xor %r13,%r13 // clear X
e: mov 0x68(%rdi),%r9d 2c: mov 0x68(%rdi),%r9d
12: sub 0x6c(%rdi),%r9d 30: sub 0x6c(%rdi),%r9d
16: mov 0xd8(%rdi),%r8 34: mov 0xd8(%rdi),%r10
3b: mov %rdi,%rbx
1d: mov $0xc,%esi 3e: mov $0xc,%esi
22: callq 0xffffffffe1021e15 43: callq 0xffffffffe102bd75
27: cmp $0x86dd,%eax 48: cmp $0x86dd,%rax
2c: jne 0x0000000000000069 4f: jne 0x000000000000009a
2e: mov $0x14,%esi 51: mov $0x14,%esi
33: callq 0xffffffffe1021e31 56: callq 0xffffffffe102bd91
38: cmp $0x84,%eax 5b: cmp $0x84,%rax
3d: je 0x0000000000000049 62: je 0x0000000000000074
3f: cmp $0x6,%eax 64: cmp $0x6,%rax
42: je 0x0000000000000049 68: je 0x0000000000000074
44: cmp $0x11,%eax 6a: cmp $0x11,%rax
47: jne 0x00000000000000c6 6e: jne 0x0000000000000117
49: mov $0x36,%esi 74: mov $0x36,%esi
4e: callq 0xffffffffe1021e15 79: callq 0xffffffffe102bd75
53: cmp $0x16,%eax 7e: cmp $0x16,%rax
56: je 0x00000000000000bf 82: je 0x0000000000000110
58: mov $0x38,%esi 88: mov $0x38,%esi
5d: callq 0xffffffffe1021e15 8d: callq 0xffffffffe102bd75
62: cmp $0x16,%eax 92: cmp $0x16,%rax
65: je 0x00000000000000bf 96: je 0x0000000000000110
67: jmp 0x00000000000000c6 98: jmp 0x0000000000000117
69: cmp $0x800,%eax 9a: cmp $0x800,%rax
6e: jne 0x00000000000000c6 a1: jne 0x0000000000000117
70: mov $0x17,%esi a3: mov $0x17,%esi
75: callq 0xffffffffe1021e31 a8: callq 0xffffffffe102bd91
7a: cmp $0x84,%eax ad: cmp $0x84,%rax
7f: je 0x000000000000008b b4: je 0x00000000000000c2
81: cmp $0x6,%eax b6: cmp $0x6,%rax
84: je 0x000000000000008b ba: je 0x00000000000000c2
86: cmp $0x11,%eax bc: cmp $0x11,%rax
89: jne 0x00000000000000c6 c0: jne 0x0000000000000117
8b: mov $0x14,%esi c2: mov $0x14,%esi
90: callq 0xffffffffe1021e15 c7: callq 0xffffffffe102bd75
95: test $0x1fff,%ax cc: test $0x1fff,%rax
99: jne 0x00000000000000c6 d3: jne 0x0000000000000117
d5: mov %rax,%r14
9b: mov $0xe,%esi d8: mov $0xe,%esi
a0: callq 0xffffffffe1021e44 dd: callq 0xffffffffe102bd91 // MSH
e2: and $0xf,%eax
e5: shl $0x2,%eax
e8: mov %rax,%r13
eb: mov %r14,%rax
ee: mov %r13,%rsi
a5: lea 0xe(%rbx),%esi f1: add $0xe,%esi
a8: callq 0xffffffffe1021e0d f4: callq 0xffffffffe102bd6d
ad: cmp $0x16,%eax f9: cmp $0x16,%rax
b0: je 0x00000000000000bf fd: je 0x0000000000000110
ff: mov %r13,%rsi
b2: lea 0x10(%rbx),%esi 102: add $0x10,%esi
b5: callq 0xffffffffe1021e0d 105: callq 0xffffffffe102bd6d
ba: cmp $0x16,%eax 10a: cmp $0x16,%rax
bd: jne 0x00000000000000c6 10e: jne 0x0000000000000117
bf: mov $0xffff,%eax 110: mov $0xffff,%eax
c4: jmp 0x00000000000000c8 115: jmp 0x000000000000011c
c6: xor %eax,%eax 117: mov $0x0,%eax
c8: mov -0x8(%rbp),%rbx 11c: mov -0x228(%rbp),%rbx // epilogue
cc: leaveq 123: mov -0x220(%rbp),%r13
cd: retq 12a: mov -0x218(%rbp),%r14
131: mov -0x210(%rbp),%r15
138: leaveq
139: retq
On fully cached SKBs both JITed functions take 12 nsec to execute.
BPF interpreter executes the program in 30 nsec.
The difference in generated assembler is due to the following:
Old BPF imlements LDX_MSH instruction via sk_load_byte_msh() helper function
inside bpf_jit.S.
New JIT removes the helper and does it explicitly, so ldx_msh cost
is the same for both JITs, but generated code looks longer.
New JIT has 4 registers to save, so prologue/epilogue are larger,
but the cost is within noise on x64.
Old JIT checks whether first insn clears A and if not emits 'xor %eax,%eax'.
New JIT clears %rax unconditionally.
2. old BPF JIT doesn't support ANC_NLATTR, ANC_PAY_OFFSET, ANC_RANDOM
extensions. New JIT supports all BPF extensions.
Performance of such filters improves 2-4 times depending on a filter.
The longer the filter the higher performance gain.
Synthetic benchmarks with many ancillary loads see 20x speedup
which seems to be the maximum gain from JIT
Notes:
. net.core.bpf_jit_enable=2 + tools/net/bpf_jit_disasm is still functional
and can be used to see generated assembler
. there are two jit_compile() functions and code flow for classic filters is:
sk_attach_filter() - load classic BPF
bpf_jit_compile() - try to JIT from classic BPF
sk_convert_filter() - convert classic to internal
bpf_int_jit_compile() - JIT from internal BPF
seccomp and tracing filters will just call bpf_int_jit_compile()
Signed-off-by: Alexei Starovoitov <ast@plumgrid.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
2014-05-14 02:50:46 +00:00
|
|
|
case BPF_ST | BPF_MEM | BPF_B:
|
2014-06-06 21:46:06 +00:00
|
|
|
if (is_ereg(dst_reg))
|
net: filter: x86: internal BPF JIT
Maps all internal BPF instructions into x86_64 instructions.
This patch replaces original BPF x64 JIT with internal BPF x64 JIT.
sysctl net.core.bpf_jit_enable is reused as on/off switch.
Performance:
1. old BPF JIT and internal BPF JIT generate equivalent x86_64 code.
No performance difference is observed for filters that were JIT-able before
Example assembler code for BPF filter "tcpdump port 22"
original BPF -> old JIT: original BPF -> internal BPF -> new JIT:
0: push %rbp 0: push %rbp
1: mov %rsp,%rbp 1: mov %rsp,%rbp
4: sub $0x60,%rsp 4: sub $0x228,%rsp
8: mov %rbx,-0x8(%rbp) b: mov %rbx,-0x228(%rbp) // prologue
12: mov %r13,-0x220(%rbp)
19: mov %r14,-0x218(%rbp)
20: mov %r15,-0x210(%rbp)
27: xor %eax,%eax // clear A
c: xor %ebx,%ebx 29: xor %r13,%r13 // clear X
e: mov 0x68(%rdi),%r9d 2c: mov 0x68(%rdi),%r9d
12: sub 0x6c(%rdi),%r9d 30: sub 0x6c(%rdi),%r9d
16: mov 0xd8(%rdi),%r8 34: mov 0xd8(%rdi),%r10
3b: mov %rdi,%rbx
1d: mov $0xc,%esi 3e: mov $0xc,%esi
22: callq 0xffffffffe1021e15 43: callq 0xffffffffe102bd75
27: cmp $0x86dd,%eax 48: cmp $0x86dd,%rax
2c: jne 0x0000000000000069 4f: jne 0x000000000000009a
2e: mov $0x14,%esi 51: mov $0x14,%esi
33: callq 0xffffffffe1021e31 56: callq 0xffffffffe102bd91
38: cmp $0x84,%eax 5b: cmp $0x84,%rax
3d: je 0x0000000000000049 62: je 0x0000000000000074
3f: cmp $0x6,%eax 64: cmp $0x6,%rax
42: je 0x0000000000000049 68: je 0x0000000000000074
44: cmp $0x11,%eax 6a: cmp $0x11,%rax
47: jne 0x00000000000000c6 6e: jne 0x0000000000000117
49: mov $0x36,%esi 74: mov $0x36,%esi
4e: callq 0xffffffffe1021e15 79: callq 0xffffffffe102bd75
53: cmp $0x16,%eax 7e: cmp $0x16,%rax
56: je 0x00000000000000bf 82: je 0x0000000000000110
58: mov $0x38,%esi 88: mov $0x38,%esi
5d: callq 0xffffffffe1021e15 8d: callq 0xffffffffe102bd75
62: cmp $0x16,%eax 92: cmp $0x16,%rax
65: je 0x00000000000000bf 96: je 0x0000000000000110
67: jmp 0x00000000000000c6 98: jmp 0x0000000000000117
69: cmp $0x800,%eax 9a: cmp $0x800,%rax
6e: jne 0x00000000000000c6 a1: jne 0x0000000000000117
70: mov $0x17,%esi a3: mov $0x17,%esi
75: callq 0xffffffffe1021e31 a8: callq 0xffffffffe102bd91
7a: cmp $0x84,%eax ad: cmp $0x84,%rax
7f: je 0x000000000000008b b4: je 0x00000000000000c2
81: cmp $0x6,%eax b6: cmp $0x6,%rax
84: je 0x000000000000008b ba: je 0x00000000000000c2
86: cmp $0x11,%eax bc: cmp $0x11,%rax
89: jne 0x00000000000000c6 c0: jne 0x0000000000000117
8b: mov $0x14,%esi c2: mov $0x14,%esi
90: callq 0xffffffffe1021e15 c7: callq 0xffffffffe102bd75
95: test $0x1fff,%ax cc: test $0x1fff,%rax
99: jne 0x00000000000000c6 d3: jne 0x0000000000000117
d5: mov %rax,%r14
9b: mov $0xe,%esi d8: mov $0xe,%esi
a0: callq 0xffffffffe1021e44 dd: callq 0xffffffffe102bd91 // MSH
e2: and $0xf,%eax
e5: shl $0x2,%eax
e8: mov %rax,%r13
eb: mov %r14,%rax
ee: mov %r13,%rsi
a5: lea 0xe(%rbx),%esi f1: add $0xe,%esi
a8: callq 0xffffffffe1021e0d f4: callq 0xffffffffe102bd6d
ad: cmp $0x16,%eax f9: cmp $0x16,%rax
b0: je 0x00000000000000bf fd: je 0x0000000000000110
ff: mov %r13,%rsi
b2: lea 0x10(%rbx),%esi 102: add $0x10,%esi
b5: callq 0xffffffffe1021e0d 105: callq 0xffffffffe102bd6d
ba: cmp $0x16,%eax 10a: cmp $0x16,%rax
bd: jne 0x00000000000000c6 10e: jne 0x0000000000000117
bf: mov $0xffff,%eax 110: mov $0xffff,%eax
c4: jmp 0x00000000000000c8 115: jmp 0x000000000000011c
c6: xor %eax,%eax 117: mov $0x0,%eax
c8: mov -0x8(%rbp),%rbx 11c: mov -0x228(%rbp),%rbx // epilogue
cc: leaveq 123: mov -0x220(%rbp),%r13
cd: retq 12a: mov -0x218(%rbp),%r14
131: mov -0x210(%rbp),%r15
138: leaveq
139: retq
On fully cached SKBs both JITed functions take 12 nsec to execute.
BPF interpreter executes the program in 30 nsec.
The difference in generated assembler is due to the following:
Old BPF imlements LDX_MSH instruction via sk_load_byte_msh() helper function
inside bpf_jit.S.
New JIT removes the helper and does it explicitly, so ldx_msh cost
is the same for both JITs, but generated code looks longer.
New JIT has 4 registers to save, so prologue/epilogue are larger,
but the cost is within noise on x64.
Old JIT checks whether first insn clears A and if not emits 'xor %eax,%eax'.
New JIT clears %rax unconditionally.
2. old BPF JIT doesn't support ANC_NLATTR, ANC_PAY_OFFSET, ANC_RANDOM
extensions. New JIT supports all BPF extensions.
Performance of such filters improves 2-4 times depending on a filter.
The longer the filter the higher performance gain.
Synthetic benchmarks with many ancillary loads see 20x speedup
which seems to be the maximum gain from JIT
Notes:
. net.core.bpf_jit_enable=2 + tools/net/bpf_jit_disasm is still functional
and can be used to see generated assembler
. there are two jit_compile() functions and code flow for classic filters is:
sk_attach_filter() - load classic BPF
bpf_jit_compile() - try to JIT from classic BPF
sk_convert_filter() - convert classic to internal
bpf_int_jit_compile() - JIT from internal BPF
seccomp and tracing filters will just call bpf_int_jit_compile()
Signed-off-by: Alexei Starovoitov <ast@plumgrid.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
2014-05-14 02:50:46 +00:00
|
|
|
EMIT2(0x41, 0xC6);
|
|
|
|
else
|
|
|
|
EMIT1(0xC6);
|
|
|
|
goto st;
|
|
|
|
case BPF_ST | BPF_MEM | BPF_H:
|
2014-06-06 21:46:06 +00:00
|
|
|
if (is_ereg(dst_reg))
|
net: filter: x86: internal BPF JIT
Maps all internal BPF instructions into x86_64 instructions.
This patch replaces original BPF x64 JIT with internal BPF x64 JIT.
sysctl net.core.bpf_jit_enable is reused as on/off switch.
Performance:
1. old BPF JIT and internal BPF JIT generate equivalent x86_64 code.
No performance difference is observed for filters that were JIT-able before
Example assembler code for BPF filter "tcpdump port 22"
original BPF -> old JIT: original BPF -> internal BPF -> new JIT:
0: push %rbp 0: push %rbp
1: mov %rsp,%rbp 1: mov %rsp,%rbp
4: sub $0x60,%rsp 4: sub $0x228,%rsp
8: mov %rbx,-0x8(%rbp) b: mov %rbx,-0x228(%rbp) // prologue
12: mov %r13,-0x220(%rbp)
19: mov %r14,-0x218(%rbp)
20: mov %r15,-0x210(%rbp)
27: xor %eax,%eax // clear A
c: xor %ebx,%ebx 29: xor %r13,%r13 // clear X
e: mov 0x68(%rdi),%r9d 2c: mov 0x68(%rdi),%r9d
12: sub 0x6c(%rdi),%r9d 30: sub 0x6c(%rdi),%r9d
16: mov 0xd8(%rdi),%r8 34: mov 0xd8(%rdi),%r10
3b: mov %rdi,%rbx
1d: mov $0xc,%esi 3e: mov $0xc,%esi
22: callq 0xffffffffe1021e15 43: callq 0xffffffffe102bd75
27: cmp $0x86dd,%eax 48: cmp $0x86dd,%rax
2c: jne 0x0000000000000069 4f: jne 0x000000000000009a
2e: mov $0x14,%esi 51: mov $0x14,%esi
33: callq 0xffffffffe1021e31 56: callq 0xffffffffe102bd91
38: cmp $0x84,%eax 5b: cmp $0x84,%rax
3d: je 0x0000000000000049 62: je 0x0000000000000074
3f: cmp $0x6,%eax 64: cmp $0x6,%rax
42: je 0x0000000000000049 68: je 0x0000000000000074
44: cmp $0x11,%eax 6a: cmp $0x11,%rax
47: jne 0x00000000000000c6 6e: jne 0x0000000000000117
49: mov $0x36,%esi 74: mov $0x36,%esi
4e: callq 0xffffffffe1021e15 79: callq 0xffffffffe102bd75
53: cmp $0x16,%eax 7e: cmp $0x16,%rax
56: je 0x00000000000000bf 82: je 0x0000000000000110
58: mov $0x38,%esi 88: mov $0x38,%esi
5d: callq 0xffffffffe1021e15 8d: callq 0xffffffffe102bd75
62: cmp $0x16,%eax 92: cmp $0x16,%rax
65: je 0x00000000000000bf 96: je 0x0000000000000110
67: jmp 0x00000000000000c6 98: jmp 0x0000000000000117
69: cmp $0x800,%eax 9a: cmp $0x800,%rax
6e: jne 0x00000000000000c6 a1: jne 0x0000000000000117
70: mov $0x17,%esi a3: mov $0x17,%esi
75: callq 0xffffffffe1021e31 a8: callq 0xffffffffe102bd91
7a: cmp $0x84,%eax ad: cmp $0x84,%rax
7f: je 0x000000000000008b b4: je 0x00000000000000c2
81: cmp $0x6,%eax b6: cmp $0x6,%rax
84: je 0x000000000000008b ba: je 0x00000000000000c2
86: cmp $0x11,%eax bc: cmp $0x11,%rax
89: jne 0x00000000000000c6 c0: jne 0x0000000000000117
8b: mov $0x14,%esi c2: mov $0x14,%esi
90: callq 0xffffffffe1021e15 c7: callq 0xffffffffe102bd75
95: test $0x1fff,%ax cc: test $0x1fff,%rax
99: jne 0x00000000000000c6 d3: jne 0x0000000000000117
d5: mov %rax,%r14
9b: mov $0xe,%esi d8: mov $0xe,%esi
a0: callq 0xffffffffe1021e44 dd: callq 0xffffffffe102bd91 // MSH
e2: and $0xf,%eax
e5: shl $0x2,%eax
e8: mov %rax,%r13
eb: mov %r14,%rax
ee: mov %r13,%rsi
a5: lea 0xe(%rbx),%esi f1: add $0xe,%esi
a8: callq 0xffffffffe1021e0d f4: callq 0xffffffffe102bd6d
ad: cmp $0x16,%eax f9: cmp $0x16,%rax
b0: je 0x00000000000000bf fd: je 0x0000000000000110
ff: mov %r13,%rsi
b2: lea 0x10(%rbx),%esi 102: add $0x10,%esi
b5: callq 0xffffffffe1021e0d 105: callq 0xffffffffe102bd6d
ba: cmp $0x16,%eax 10a: cmp $0x16,%rax
bd: jne 0x00000000000000c6 10e: jne 0x0000000000000117
bf: mov $0xffff,%eax 110: mov $0xffff,%eax
c4: jmp 0x00000000000000c8 115: jmp 0x000000000000011c
c6: xor %eax,%eax 117: mov $0x0,%eax
c8: mov -0x8(%rbp),%rbx 11c: mov -0x228(%rbp),%rbx // epilogue
cc: leaveq 123: mov -0x220(%rbp),%r13
cd: retq 12a: mov -0x218(%rbp),%r14
131: mov -0x210(%rbp),%r15
138: leaveq
139: retq
On fully cached SKBs both JITed functions take 12 nsec to execute.
BPF interpreter executes the program in 30 nsec.
The difference in generated assembler is due to the following:
Old BPF imlements LDX_MSH instruction via sk_load_byte_msh() helper function
inside bpf_jit.S.
New JIT removes the helper and does it explicitly, so ldx_msh cost
is the same for both JITs, but generated code looks longer.
New JIT has 4 registers to save, so prologue/epilogue are larger,
but the cost is within noise on x64.
Old JIT checks whether first insn clears A and if not emits 'xor %eax,%eax'.
New JIT clears %rax unconditionally.
2. old BPF JIT doesn't support ANC_NLATTR, ANC_PAY_OFFSET, ANC_RANDOM
extensions. New JIT supports all BPF extensions.
Performance of such filters improves 2-4 times depending on a filter.
The longer the filter the higher performance gain.
Synthetic benchmarks with many ancillary loads see 20x speedup
which seems to be the maximum gain from JIT
Notes:
. net.core.bpf_jit_enable=2 + tools/net/bpf_jit_disasm is still functional
and can be used to see generated assembler
. there are two jit_compile() functions and code flow for classic filters is:
sk_attach_filter() - load classic BPF
bpf_jit_compile() - try to JIT from classic BPF
sk_convert_filter() - convert classic to internal
bpf_int_jit_compile() - JIT from internal BPF
seccomp and tracing filters will just call bpf_int_jit_compile()
Signed-off-by: Alexei Starovoitov <ast@plumgrid.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
2014-05-14 02:50:46 +00:00
|
|
|
EMIT3(0x66, 0x41, 0xC7);
|
|
|
|
else
|
|
|
|
EMIT2(0x66, 0xC7);
|
|
|
|
goto st;
|
|
|
|
case BPF_ST | BPF_MEM | BPF_W:
|
2014-06-06 21:46:06 +00:00
|
|
|
if (is_ereg(dst_reg))
|
net: filter: x86: internal BPF JIT
Maps all internal BPF instructions into x86_64 instructions.
This patch replaces original BPF x64 JIT with internal BPF x64 JIT.
sysctl net.core.bpf_jit_enable is reused as on/off switch.
Performance:
1. old BPF JIT and internal BPF JIT generate equivalent x86_64 code.
No performance difference is observed for filters that were JIT-able before
Example assembler code for BPF filter "tcpdump port 22"
original BPF -> old JIT: original BPF -> internal BPF -> new JIT:
0: push %rbp 0: push %rbp
1: mov %rsp,%rbp 1: mov %rsp,%rbp
4: sub $0x60,%rsp 4: sub $0x228,%rsp
8: mov %rbx,-0x8(%rbp) b: mov %rbx,-0x228(%rbp) // prologue
12: mov %r13,-0x220(%rbp)
19: mov %r14,-0x218(%rbp)
20: mov %r15,-0x210(%rbp)
27: xor %eax,%eax // clear A
c: xor %ebx,%ebx 29: xor %r13,%r13 // clear X
e: mov 0x68(%rdi),%r9d 2c: mov 0x68(%rdi),%r9d
12: sub 0x6c(%rdi),%r9d 30: sub 0x6c(%rdi),%r9d
16: mov 0xd8(%rdi),%r8 34: mov 0xd8(%rdi),%r10
3b: mov %rdi,%rbx
1d: mov $0xc,%esi 3e: mov $0xc,%esi
22: callq 0xffffffffe1021e15 43: callq 0xffffffffe102bd75
27: cmp $0x86dd,%eax 48: cmp $0x86dd,%rax
2c: jne 0x0000000000000069 4f: jne 0x000000000000009a
2e: mov $0x14,%esi 51: mov $0x14,%esi
33: callq 0xffffffffe1021e31 56: callq 0xffffffffe102bd91
38: cmp $0x84,%eax 5b: cmp $0x84,%rax
3d: je 0x0000000000000049 62: je 0x0000000000000074
3f: cmp $0x6,%eax 64: cmp $0x6,%rax
42: je 0x0000000000000049 68: je 0x0000000000000074
44: cmp $0x11,%eax 6a: cmp $0x11,%rax
47: jne 0x00000000000000c6 6e: jne 0x0000000000000117
49: mov $0x36,%esi 74: mov $0x36,%esi
4e: callq 0xffffffffe1021e15 79: callq 0xffffffffe102bd75
53: cmp $0x16,%eax 7e: cmp $0x16,%rax
56: je 0x00000000000000bf 82: je 0x0000000000000110
58: mov $0x38,%esi 88: mov $0x38,%esi
5d: callq 0xffffffffe1021e15 8d: callq 0xffffffffe102bd75
62: cmp $0x16,%eax 92: cmp $0x16,%rax
65: je 0x00000000000000bf 96: je 0x0000000000000110
67: jmp 0x00000000000000c6 98: jmp 0x0000000000000117
69: cmp $0x800,%eax 9a: cmp $0x800,%rax
6e: jne 0x00000000000000c6 a1: jne 0x0000000000000117
70: mov $0x17,%esi a3: mov $0x17,%esi
75: callq 0xffffffffe1021e31 a8: callq 0xffffffffe102bd91
7a: cmp $0x84,%eax ad: cmp $0x84,%rax
7f: je 0x000000000000008b b4: je 0x00000000000000c2
81: cmp $0x6,%eax b6: cmp $0x6,%rax
84: je 0x000000000000008b ba: je 0x00000000000000c2
86: cmp $0x11,%eax bc: cmp $0x11,%rax
89: jne 0x00000000000000c6 c0: jne 0x0000000000000117
8b: mov $0x14,%esi c2: mov $0x14,%esi
90: callq 0xffffffffe1021e15 c7: callq 0xffffffffe102bd75
95: test $0x1fff,%ax cc: test $0x1fff,%rax
99: jne 0x00000000000000c6 d3: jne 0x0000000000000117
d5: mov %rax,%r14
9b: mov $0xe,%esi d8: mov $0xe,%esi
a0: callq 0xffffffffe1021e44 dd: callq 0xffffffffe102bd91 // MSH
e2: and $0xf,%eax
e5: shl $0x2,%eax
e8: mov %rax,%r13
eb: mov %r14,%rax
ee: mov %r13,%rsi
a5: lea 0xe(%rbx),%esi f1: add $0xe,%esi
a8: callq 0xffffffffe1021e0d f4: callq 0xffffffffe102bd6d
ad: cmp $0x16,%eax f9: cmp $0x16,%rax
b0: je 0x00000000000000bf fd: je 0x0000000000000110
ff: mov %r13,%rsi
b2: lea 0x10(%rbx),%esi 102: add $0x10,%esi
b5: callq 0xffffffffe1021e0d 105: callq 0xffffffffe102bd6d
ba: cmp $0x16,%eax 10a: cmp $0x16,%rax
bd: jne 0x00000000000000c6 10e: jne 0x0000000000000117
bf: mov $0xffff,%eax 110: mov $0xffff,%eax
c4: jmp 0x00000000000000c8 115: jmp 0x000000000000011c
c6: xor %eax,%eax 117: mov $0x0,%eax
c8: mov -0x8(%rbp),%rbx 11c: mov -0x228(%rbp),%rbx // epilogue
cc: leaveq 123: mov -0x220(%rbp),%r13
cd: retq 12a: mov -0x218(%rbp),%r14
131: mov -0x210(%rbp),%r15
138: leaveq
139: retq
On fully cached SKBs both JITed functions take 12 nsec to execute.
BPF interpreter executes the program in 30 nsec.
The difference in generated assembler is due to the following:
Old BPF imlements LDX_MSH instruction via sk_load_byte_msh() helper function
inside bpf_jit.S.
New JIT removes the helper and does it explicitly, so ldx_msh cost
is the same for both JITs, but generated code looks longer.
New JIT has 4 registers to save, so prologue/epilogue are larger,
but the cost is within noise on x64.
Old JIT checks whether first insn clears A and if not emits 'xor %eax,%eax'.
New JIT clears %rax unconditionally.
2. old BPF JIT doesn't support ANC_NLATTR, ANC_PAY_OFFSET, ANC_RANDOM
extensions. New JIT supports all BPF extensions.
Performance of such filters improves 2-4 times depending on a filter.
The longer the filter the higher performance gain.
Synthetic benchmarks with many ancillary loads see 20x speedup
which seems to be the maximum gain from JIT
Notes:
. net.core.bpf_jit_enable=2 + tools/net/bpf_jit_disasm is still functional
and can be used to see generated assembler
. there are two jit_compile() functions and code flow for classic filters is:
sk_attach_filter() - load classic BPF
bpf_jit_compile() - try to JIT from classic BPF
sk_convert_filter() - convert classic to internal
bpf_int_jit_compile() - JIT from internal BPF
seccomp and tracing filters will just call bpf_int_jit_compile()
Signed-off-by: Alexei Starovoitov <ast@plumgrid.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
2014-05-14 02:50:46 +00:00
|
|
|
EMIT2(0x41, 0xC7);
|
|
|
|
else
|
|
|
|
EMIT1(0xC7);
|
|
|
|
goto st;
|
|
|
|
case BPF_ST | BPF_MEM | BPF_DW:
|
2014-06-06 21:46:06 +00:00
|
|
|
EMIT2(add_1mod(0x48, dst_reg), 0xC7);
|
net: filter: x86: internal BPF JIT
Maps all internal BPF instructions into x86_64 instructions.
This patch replaces original BPF x64 JIT with internal BPF x64 JIT.
sysctl net.core.bpf_jit_enable is reused as on/off switch.
Performance:
1. old BPF JIT and internal BPF JIT generate equivalent x86_64 code.
No performance difference is observed for filters that were JIT-able before
Example assembler code for BPF filter "tcpdump port 22"
original BPF -> old JIT: original BPF -> internal BPF -> new JIT:
0: push %rbp 0: push %rbp
1: mov %rsp,%rbp 1: mov %rsp,%rbp
4: sub $0x60,%rsp 4: sub $0x228,%rsp
8: mov %rbx,-0x8(%rbp) b: mov %rbx,-0x228(%rbp) // prologue
12: mov %r13,-0x220(%rbp)
19: mov %r14,-0x218(%rbp)
20: mov %r15,-0x210(%rbp)
27: xor %eax,%eax // clear A
c: xor %ebx,%ebx 29: xor %r13,%r13 // clear X
e: mov 0x68(%rdi),%r9d 2c: mov 0x68(%rdi),%r9d
12: sub 0x6c(%rdi),%r9d 30: sub 0x6c(%rdi),%r9d
16: mov 0xd8(%rdi),%r8 34: mov 0xd8(%rdi),%r10
3b: mov %rdi,%rbx
1d: mov $0xc,%esi 3e: mov $0xc,%esi
22: callq 0xffffffffe1021e15 43: callq 0xffffffffe102bd75
27: cmp $0x86dd,%eax 48: cmp $0x86dd,%rax
2c: jne 0x0000000000000069 4f: jne 0x000000000000009a
2e: mov $0x14,%esi 51: mov $0x14,%esi
33: callq 0xffffffffe1021e31 56: callq 0xffffffffe102bd91
38: cmp $0x84,%eax 5b: cmp $0x84,%rax
3d: je 0x0000000000000049 62: je 0x0000000000000074
3f: cmp $0x6,%eax 64: cmp $0x6,%rax
42: je 0x0000000000000049 68: je 0x0000000000000074
44: cmp $0x11,%eax 6a: cmp $0x11,%rax
47: jne 0x00000000000000c6 6e: jne 0x0000000000000117
49: mov $0x36,%esi 74: mov $0x36,%esi
4e: callq 0xffffffffe1021e15 79: callq 0xffffffffe102bd75
53: cmp $0x16,%eax 7e: cmp $0x16,%rax
56: je 0x00000000000000bf 82: je 0x0000000000000110
58: mov $0x38,%esi 88: mov $0x38,%esi
5d: callq 0xffffffffe1021e15 8d: callq 0xffffffffe102bd75
62: cmp $0x16,%eax 92: cmp $0x16,%rax
65: je 0x00000000000000bf 96: je 0x0000000000000110
67: jmp 0x00000000000000c6 98: jmp 0x0000000000000117
69: cmp $0x800,%eax 9a: cmp $0x800,%rax
6e: jne 0x00000000000000c6 a1: jne 0x0000000000000117
70: mov $0x17,%esi a3: mov $0x17,%esi
75: callq 0xffffffffe1021e31 a8: callq 0xffffffffe102bd91
7a: cmp $0x84,%eax ad: cmp $0x84,%rax
7f: je 0x000000000000008b b4: je 0x00000000000000c2
81: cmp $0x6,%eax b6: cmp $0x6,%rax
84: je 0x000000000000008b ba: je 0x00000000000000c2
86: cmp $0x11,%eax bc: cmp $0x11,%rax
89: jne 0x00000000000000c6 c0: jne 0x0000000000000117
8b: mov $0x14,%esi c2: mov $0x14,%esi
90: callq 0xffffffffe1021e15 c7: callq 0xffffffffe102bd75
95: test $0x1fff,%ax cc: test $0x1fff,%rax
99: jne 0x00000000000000c6 d3: jne 0x0000000000000117
d5: mov %rax,%r14
9b: mov $0xe,%esi d8: mov $0xe,%esi
a0: callq 0xffffffffe1021e44 dd: callq 0xffffffffe102bd91 // MSH
e2: and $0xf,%eax
e5: shl $0x2,%eax
e8: mov %rax,%r13
eb: mov %r14,%rax
ee: mov %r13,%rsi
a5: lea 0xe(%rbx),%esi f1: add $0xe,%esi
a8: callq 0xffffffffe1021e0d f4: callq 0xffffffffe102bd6d
ad: cmp $0x16,%eax f9: cmp $0x16,%rax
b0: je 0x00000000000000bf fd: je 0x0000000000000110
ff: mov %r13,%rsi
b2: lea 0x10(%rbx),%esi 102: add $0x10,%esi
b5: callq 0xffffffffe1021e0d 105: callq 0xffffffffe102bd6d
ba: cmp $0x16,%eax 10a: cmp $0x16,%rax
bd: jne 0x00000000000000c6 10e: jne 0x0000000000000117
bf: mov $0xffff,%eax 110: mov $0xffff,%eax
c4: jmp 0x00000000000000c8 115: jmp 0x000000000000011c
c6: xor %eax,%eax 117: mov $0x0,%eax
c8: mov -0x8(%rbp),%rbx 11c: mov -0x228(%rbp),%rbx // epilogue
cc: leaveq 123: mov -0x220(%rbp),%r13
cd: retq 12a: mov -0x218(%rbp),%r14
131: mov -0x210(%rbp),%r15
138: leaveq
139: retq
On fully cached SKBs both JITed functions take 12 nsec to execute.
BPF interpreter executes the program in 30 nsec.
The difference in generated assembler is due to the following:
Old BPF imlements LDX_MSH instruction via sk_load_byte_msh() helper function
inside bpf_jit.S.
New JIT removes the helper and does it explicitly, so ldx_msh cost
is the same for both JITs, but generated code looks longer.
New JIT has 4 registers to save, so prologue/epilogue are larger,
but the cost is within noise on x64.
Old JIT checks whether first insn clears A and if not emits 'xor %eax,%eax'.
New JIT clears %rax unconditionally.
2. old BPF JIT doesn't support ANC_NLATTR, ANC_PAY_OFFSET, ANC_RANDOM
extensions. New JIT supports all BPF extensions.
Performance of such filters improves 2-4 times depending on a filter.
The longer the filter the higher performance gain.
Synthetic benchmarks with many ancillary loads see 20x speedup
which seems to be the maximum gain from JIT
Notes:
. net.core.bpf_jit_enable=2 + tools/net/bpf_jit_disasm is still functional
and can be used to see generated assembler
. there are two jit_compile() functions and code flow for classic filters is:
sk_attach_filter() - load classic BPF
bpf_jit_compile() - try to JIT from classic BPF
sk_convert_filter() - convert classic to internal
bpf_int_jit_compile() - JIT from internal BPF
seccomp and tracing filters will just call bpf_int_jit_compile()
Signed-off-by: Alexei Starovoitov <ast@plumgrid.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
2014-05-14 02:50:46 +00:00
|
|
|
|
|
|
|
st: if (is_imm8(insn->off))
|
2014-06-06 21:46:06 +00:00
|
|
|
EMIT2(add_1reg(0x40, dst_reg), insn->off);
|
net: filter: x86: internal BPF JIT
Maps all internal BPF instructions into x86_64 instructions.
This patch replaces original BPF x64 JIT with internal BPF x64 JIT.
sysctl net.core.bpf_jit_enable is reused as on/off switch.
Performance:
1. old BPF JIT and internal BPF JIT generate equivalent x86_64 code.
No performance difference is observed for filters that were JIT-able before
Example assembler code for BPF filter "tcpdump port 22"
original BPF -> old JIT: original BPF -> internal BPF -> new JIT:
0: push %rbp 0: push %rbp
1: mov %rsp,%rbp 1: mov %rsp,%rbp
4: sub $0x60,%rsp 4: sub $0x228,%rsp
8: mov %rbx,-0x8(%rbp) b: mov %rbx,-0x228(%rbp) // prologue
12: mov %r13,-0x220(%rbp)
19: mov %r14,-0x218(%rbp)
20: mov %r15,-0x210(%rbp)
27: xor %eax,%eax // clear A
c: xor %ebx,%ebx 29: xor %r13,%r13 // clear X
e: mov 0x68(%rdi),%r9d 2c: mov 0x68(%rdi),%r9d
12: sub 0x6c(%rdi),%r9d 30: sub 0x6c(%rdi),%r9d
16: mov 0xd8(%rdi),%r8 34: mov 0xd8(%rdi),%r10
3b: mov %rdi,%rbx
1d: mov $0xc,%esi 3e: mov $0xc,%esi
22: callq 0xffffffffe1021e15 43: callq 0xffffffffe102bd75
27: cmp $0x86dd,%eax 48: cmp $0x86dd,%rax
2c: jne 0x0000000000000069 4f: jne 0x000000000000009a
2e: mov $0x14,%esi 51: mov $0x14,%esi
33: callq 0xffffffffe1021e31 56: callq 0xffffffffe102bd91
38: cmp $0x84,%eax 5b: cmp $0x84,%rax
3d: je 0x0000000000000049 62: je 0x0000000000000074
3f: cmp $0x6,%eax 64: cmp $0x6,%rax
42: je 0x0000000000000049 68: je 0x0000000000000074
44: cmp $0x11,%eax 6a: cmp $0x11,%rax
47: jne 0x00000000000000c6 6e: jne 0x0000000000000117
49: mov $0x36,%esi 74: mov $0x36,%esi
4e: callq 0xffffffffe1021e15 79: callq 0xffffffffe102bd75
53: cmp $0x16,%eax 7e: cmp $0x16,%rax
56: je 0x00000000000000bf 82: je 0x0000000000000110
58: mov $0x38,%esi 88: mov $0x38,%esi
5d: callq 0xffffffffe1021e15 8d: callq 0xffffffffe102bd75
62: cmp $0x16,%eax 92: cmp $0x16,%rax
65: je 0x00000000000000bf 96: je 0x0000000000000110
67: jmp 0x00000000000000c6 98: jmp 0x0000000000000117
69: cmp $0x800,%eax 9a: cmp $0x800,%rax
6e: jne 0x00000000000000c6 a1: jne 0x0000000000000117
70: mov $0x17,%esi a3: mov $0x17,%esi
75: callq 0xffffffffe1021e31 a8: callq 0xffffffffe102bd91
7a: cmp $0x84,%eax ad: cmp $0x84,%rax
7f: je 0x000000000000008b b4: je 0x00000000000000c2
81: cmp $0x6,%eax b6: cmp $0x6,%rax
84: je 0x000000000000008b ba: je 0x00000000000000c2
86: cmp $0x11,%eax bc: cmp $0x11,%rax
89: jne 0x00000000000000c6 c0: jne 0x0000000000000117
8b: mov $0x14,%esi c2: mov $0x14,%esi
90: callq 0xffffffffe1021e15 c7: callq 0xffffffffe102bd75
95: test $0x1fff,%ax cc: test $0x1fff,%rax
99: jne 0x00000000000000c6 d3: jne 0x0000000000000117
d5: mov %rax,%r14
9b: mov $0xe,%esi d8: mov $0xe,%esi
a0: callq 0xffffffffe1021e44 dd: callq 0xffffffffe102bd91 // MSH
e2: and $0xf,%eax
e5: shl $0x2,%eax
e8: mov %rax,%r13
eb: mov %r14,%rax
ee: mov %r13,%rsi
a5: lea 0xe(%rbx),%esi f1: add $0xe,%esi
a8: callq 0xffffffffe1021e0d f4: callq 0xffffffffe102bd6d
ad: cmp $0x16,%eax f9: cmp $0x16,%rax
b0: je 0x00000000000000bf fd: je 0x0000000000000110
ff: mov %r13,%rsi
b2: lea 0x10(%rbx),%esi 102: add $0x10,%esi
b5: callq 0xffffffffe1021e0d 105: callq 0xffffffffe102bd6d
ba: cmp $0x16,%eax 10a: cmp $0x16,%rax
bd: jne 0x00000000000000c6 10e: jne 0x0000000000000117
bf: mov $0xffff,%eax 110: mov $0xffff,%eax
c4: jmp 0x00000000000000c8 115: jmp 0x000000000000011c
c6: xor %eax,%eax 117: mov $0x0,%eax
c8: mov -0x8(%rbp),%rbx 11c: mov -0x228(%rbp),%rbx // epilogue
cc: leaveq 123: mov -0x220(%rbp),%r13
cd: retq 12a: mov -0x218(%rbp),%r14
131: mov -0x210(%rbp),%r15
138: leaveq
139: retq
On fully cached SKBs both JITed functions take 12 nsec to execute.
BPF interpreter executes the program in 30 nsec.
The difference in generated assembler is due to the following:
Old BPF imlements LDX_MSH instruction via sk_load_byte_msh() helper function
inside bpf_jit.S.
New JIT removes the helper and does it explicitly, so ldx_msh cost
is the same for both JITs, but generated code looks longer.
New JIT has 4 registers to save, so prologue/epilogue are larger,
but the cost is within noise on x64.
Old JIT checks whether first insn clears A and if not emits 'xor %eax,%eax'.
New JIT clears %rax unconditionally.
2. old BPF JIT doesn't support ANC_NLATTR, ANC_PAY_OFFSET, ANC_RANDOM
extensions. New JIT supports all BPF extensions.
Performance of such filters improves 2-4 times depending on a filter.
The longer the filter the higher performance gain.
Synthetic benchmarks with many ancillary loads see 20x speedup
which seems to be the maximum gain from JIT
Notes:
. net.core.bpf_jit_enable=2 + tools/net/bpf_jit_disasm is still functional
and can be used to see generated assembler
. there are two jit_compile() functions and code flow for classic filters is:
sk_attach_filter() - load classic BPF
bpf_jit_compile() - try to JIT from classic BPF
sk_convert_filter() - convert classic to internal
bpf_int_jit_compile() - JIT from internal BPF
seccomp and tracing filters will just call bpf_int_jit_compile()
Signed-off-by: Alexei Starovoitov <ast@plumgrid.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
2014-05-14 02:50:46 +00:00
|
|
|
else
|
2014-06-06 21:46:06 +00:00
|
|
|
EMIT1_off32(add_1reg(0x80, dst_reg), insn->off);
|
net: filter: x86: internal BPF JIT
Maps all internal BPF instructions into x86_64 instructions.
This patch replaces original BPF x64 JIT with internal BPF x64 JIT.
sysctl net.core.bpf_jit_enable is reused as on/off switch.
Performance:
1. old BPF JIT and internal BPF JIT generate equivalent x86_64 code.
No performance difference is observed for filters that were JIT-able before
Example assembler code for BPF filter "tcpdump port 22"
original BPF -> old JIT: original BPF -> internal BPF -> new JIT:
0: push %rbp 0: push %rbp
1: mov %rsp,%rbp 1: mov %rsp,%rbp
4: sub $0x60,%rsp 4: sub $0x228,%rsp
8: mov %rbx,-0x8(%rbp) b: mov %rbx,-0x228(%rbp) // prologue
12: mov %r13,-0x220(%rbp)
19: mov %r14,-0x218(%rbp)
20: mov %r15,-0x210(%rbp)
27: xor %eax,%eax // clear A
c: xor %ebx,%ebx 29: xor %r13,%r13 // clear X
e: mov 0x68(%rdi),%r9d 2c: mov 0x68(%rdi),%r9d
12: sub 0x6c(%rdi),%r9d 30: sub 0x6c(%rdi),%r9d
16: mov 0xd8(%rdi),%r8 34: mov 0xd8(%rdi),%r10
3b: mov %rdi,%rbx
1d: mov $0xc,%esi 3e: mov $0xc,%esi
22: callq 0xffffffffe1021e15 43: callq 0xffffffffe102bd75
27: cmp $0x86dd,%eax 48: cmp $0x86dd,%rax
2c: jne 0x0000000000000069 4f: jne 0x000000000000009a
2e: mov $0x14,%esi 51: mov $0x14,%esi
33: callq 0xffffffffe1021e31 56: callq 0xffffffffe102bd91
38: cmp $0x84,%eax 5b: cmp $0x84,%rax
3d: je 0x0000000000000049 62: je 0x0000000000000074
3f: cmp $0x6,%eax 64: cmp $0x6,%rax
42: je 0x0000000000000049 68: je 0x0000000000000074
44: cmp $0x11,%eax 6a: cmp $0x11,%rax
47: jne 0x00000000000000c6 6e: jne 0x0000000000000117
49: mov $0x36,%esi 74: mov $0x36,%esi
4e: callq 0xffffffffe1021e15 79: callq 0xffffffffe102bd75
53: cmp $0x16,%eax 7e: cmp $0x16,%rax
56: je 0x00000000000000bf 82: je 0x0000000000000110
58: mov $0x38,%esi 88: mov $0x38,%esi
5d: callq 0xffffffffe1021e15 8d: callq 0xffffffffe102bd75
62: cmp $0x16,%eax 92: cmp $0x16,%rax
65: je 0x00000000000000bf 96: je 0x0000000000000110
67: jmp 0x00000000000000c6 98: jmp 0x0000000000000117
69: cmp $0x800,%eax 9a: cmp $0x800,%rax
6e: jne 0x00000000000000c6 a1: jne 0x0000000000000117
70: mov $0x17,%esi a3: mov $0x17,%esi
75: callq 0xffffffffe1021e31 a8: callq 0xffffffffe102bd91
7a: cmp $0x84,%eax ad: cmp $0x84,%rax
7f: je 0x000000000000008b b4: je 0x00000000000000c2
81: cmp $0x6,%eax b6: cmp $0x6,%rax
84: je 0x000000000000008b ba: je 0x00000000000000c2
86: cmp $0x11,%eax bc: cmp $0x11,%rax
89: jne 0x00000000000000c6 c0: jne 0x0000000000000117
8b: mov $0x14,%esi c2: mov $0x14,%esi
90: callq 0xffffffffe1021e15 c7: callq 0xffffffffe102bd75
95: test $0x1fff,%ax cc: test $0x1fff,%rax
99: jne 0x00000000000000c6 d3: jne 0x0000000000000117
d5: mov %rax,%r14
9b: mov $0xe,%esi d8: mov $0xe,%esi
a0: callq 0xffffffffe1021e44 dd: callq 0xffffffffe102bd91 // MSH
e2: and $0xf,%eax
e5: shl $0x2,%eax
e8: mov %rax,%r13
eb: mov %r14,%rax
ee: mov %r13,%rsi
a5: lea 0xe(%rbx),%esi f1: add $0xe,%esi
a8: callq 0xffffffffe1021e0d f4: callq 0xffffffffe102bd6d
ad: cmp $0x16,%eax f9: cmp $0x16,%rax
b0: je 0x00000000000000bf fd: je 0x0000000000000110
ff: mov %r13,%rsi
b2: lea 0x10(%rbx),%esi 102: add $0x10,%esi
b5: callq 0xffffffffe1021e0d 105: callq 0xffffffffe102bd6d
ba: cmp $0x16,%eax 10a: cmp $0x16,%rax
bd: jne 0x00000000000000c6 10e: jne 0x0000000000000117
bf: mov $0xffff,%eax 110: mov $0xffff,%eax
c4: jmp 0x00000000000000c8 115: jmp 0x000000000000011c
c6: xor %eax,%eax 117: mov $0x0,%eax
c8: mov -0x8(%rbp),%rbx 11c: mov -0x228(%rbp),%rbx // epilogue
cc: leaveq 123: mov -0x220(%rbp),%r13
cd: retq 12a: mov -0x218(%rbp),%r14
131: mov -0x210(%rbp),%r15
138: leaveq
139: retq
On fully cached SKBs both JITed functions take 12 nsec to execute.
BPF interpreter executes the program in 30 nsec.
The difference in generated assembler is due to the following:
Old BPF imlements LDX_MSH instruction via sk_load_byte_msh() helper function
inside bpf_jit.S.
New JIT removes the helper and does it explicitly, so ldx_msh cost
is the same for both JITs, but generated code looks longer.
New JIT has 4 registers to save, so prologue/epilogue are larger,
but the cost is within noise on x64.
Old JIT checks whether first insn clears A and if not emits 'xor %eax,%eax'.
New JIT clears %rax unconditionally.
2. old BPF JIT doesn't support ANC_NLATTR, ANC_PAY_OFFSET, ANC_RANDOM
extensions. New JIT supports all BPF extensions.
Performance of such filters improves 2-4 times depending on a filter.
The longer the filter the higher performance gain.
Synthetic benchmarks with many ancillary loads see 20x speedup
which seems to be the maximum gain from JIT
Notes:
. net.core.bpf_jit_enable=2 + tools/net/bpf_jit_disasm is still functional
and can be used to see generated assembler
. there are two jit_compile() functions and code flow for classic filters is:
sk_attach_filter() - load classic BPF
bpf_jit_compile() - try to JIT from classic BPF
sk_convert_filter() - convert classic to internal
bpf_int_jit_compile() - JIT from internal BPF
seccomp and tracing filters will just call bpf_int_jit_compile()
Signed-off-by: Alexei Starovoitov <ast@plumgrid.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
2014-05-14 02:50:46 +00:00
|
|
|
|
2014-06-06 21:46:06 +00:00
|
|
|
EMIT(imm32, bpf_size_to_x86_bytes(BPF_SIZE(insn->code)));
|
net: filter: x86: internal BPF JIT
Maps all internal BPF instructions into x86_64 instructions.
This patch replaces original BPF x64 JIT with internal BPF x64 JIT.
sysctl net.core.bpf_jit_enable is reused as on/off switch.
Performance:
1. old BPF JIT and internal BPF JIT generate equivalent x86_64 code.
No performance difference is observed for filters that were JIT-able before
Example assembler code for BPF filter "tcpdump port 22"
original BPF -> old JIT: original BPF -> internal BPF -> new JIT:
0: push %rbp 0: push %rbp
1: mov %rsp,%rbp 1: mov %rsp,%rbp
4: sub $0x60,%rsp 4: sub $0x228,%rsp
8: mov %rbx,-0x8(%rbp) b: mov %rbx,-0x228(%rbp) // prologue
12: mov %r13,-0x220(%rbp)
19: mov %r14,-0x218(%rbp)
20: mov %r15,-0x210(%rbp)
27: xor %eax,%eax // clear A
c: xor %ebx,%ebx 29: xor %r13,%r13 // clear X
e: mov 0x68(%rdi),%r9d 2c: mov 0x68(%rdi),%r9d
12: sub 0x6c(%rdi),%r9d 30: sub 0x6c(%rdi),%r9d
16: mov 0xd8(%rdi),%r8 34: mov 0xd8(%rdi),%r10
3b: mov %rdi,%rbx
1d: mov $0xc,%esi 3e: mov $0xc,%esi
22: callq 0xffffffffe1021e15 43: callq 0xffffffffe102bd75
27: cmp $0x86dd,%eax 48: cmp $0x86dd,%rax
2c: jne 0x0000000000000069 4f: jne 0x000000000000009a
2e: mov $0x14,%esi 51: mov $0x14,%esi
33: callq 0xffffffffe1021e31 56: callq 0xffffffffe102bd91
38: cmp $0x84,%eax 5b: cmp $0x84,%rax
3d: je 0x0000000000000049 62: je 0x0000000000000074
3f: cmp $0x6,%eax 64: cmp $0x6,%rax
42: je 0x0000000000000049 68: je 0x0000000000000074
44: cmp $0x11,%eax 6a: cmp $0x11,%rax
47: jne 0x00000000000000c6 6e: jne 0x0000000000000117
49: mov $0x36,%esi 74: mov $0x36,%esi
4e: callq 0xffffffffe1021e15 79: callq 0xffffffffe102bd75
53: cmp $0x16,%eax 7e: cmp $0x16,%rax
56: je 0x00000000000000bf 82: je 0x0000000000000110
58: mov $0x38,%esi 88: mov $0x38,%esi
5d: callq 0xffffffffe1021e15 8d: callq 0xffffffffe102bd75
62: cmp $0x16,%eax 92: cmp $0x16,%rax
65: je 0x00000000000000bf 96: je 0x0000000000000110
67: jmp 0x00000000000000c6 98: jmp 0x0000000000000117
69: cmp $0x800,%eax 9a: cmp $0x800,%rax
6e: jne 0x00000000000000c6 a1: jne 0x0000000000000117
70: mov $0x17,%esi a3: mov $0x17,%esi
75: callq 0xffffffffe1021e31 a8: callq 0xffffffffe102bd91
7a: cmp $0x84,%eax ad: cmp $0x84,%rax
7f: je 0x000000000000008b b4: je 0x00000000000000c2
81: cmp $0x6,%eax b6: cmp $0x6,%rax
84: je 0x000000000000008b ba: je 0x00000000000000c2
86: cmp $0x11,%eax bc: cmp $0x11,%rax
89: jne 0x00000000000000c6 c0: jne 0x0000000000000117
8b: mov $0x14,%esi c2: mov $0x14,%esi
90: callq 0xffffffffe1021e15 c7: callq 0xffffffffe102bd75
95: test $0x1fff,%ax cc: test $0x1fff,%rax
99: jne 0x00000000000000c6 d3: jne 0x0000000000000117
d5: mov %rax,%r14
9b: mov $0xe,%esi d8: mov $0xe,%esi
a0: callq 0xffffffffe1021e44 dd: callq 0xffffffffe102bd91 // MSH
e2: and $0xf,%eax
e5: shl $0x2,%eax
e8: mov %rax,%r13
eb: mov %r14,%rax
ee: mov %r13,%rsi
a5: lea 0xe(%rbx),%esi f1: add $0xe,%esi
a8: callq 0xffffffffe1021e0d f4: callq 0xffffffffe102bd6d
ad: cmp $0x16,%eax f9: cmp $0x16,%rax
b0: je 0x00000000000000bf fd: je 0x0000000000000110
ff: mov %r13,%rsi
b2: lea 0x10(%rbx),%esi 102: add $0x10,%esi
b5: callq 0xffffffffe1021e0d 105: callq 0xffffffffe102bd6d
ba: cmp $0x16,%eax 10a: cmp $0x16,%rax
bd: jne 0x00000000000000c6 10e: jne 0x0000000000000117
bf: mov $0xffff,%eax 110: mov $0xffff,%eax
c4: jmp 0x00000000000000c8 115: jmp 0x000000000000011c
c6: xor %eax,%eax 117: mov $0x0,%eax
c8: mov -0x8(%rbp),%rbx 11c: mov -0x228(%rbp),%rbx // epilogue
cc: leaveq 123: mov -0x220(%rbp),%r13
cd: retq 12a: mov -0x218(%rbp),%r14
131: mov -0x210(%rbp),%r15
138: leaveq
139: retq
On fully cached SKBs both JITed functions take 12 nsec to execute.
BPF interpreter executes the program in 30 nsec.
The difference in generated assembler is due to the following:
Old BPF imlements LDX_MSH instruction via sk_load_byte_msh() helper function
inside bpf_jit.S.
New JIT removes the helper and does it explicitly, so ldx_msh cost
is the same for both JITs, but generated code looks longer.
New JIT has 4 registers to save, so prologue/epilogue are larger,
but the cost is within noise on x64.
Old JIT checks whether first insn clears A and if not emits 'xor %eax,%eax'.
New JIT clears %rax unconditionally.
2. old BPF JIT doesn't support ANC_NLATTR, ANC_PAY_OFFSET, ANC_RANDOM
extensions. New JIT supports all BPF extensions.
Performance of such filters improves 2-4 times depending on a filter.
The longer the filter the higher performance gain.
Synthetic benchmarks with many ancillary loads see 20x speedup
which seems to be the maximum gain from JIT
Notes:
. net.core.bpf_jit_enable=2 + tools/net/bpf_jit_disasm is still functional
and can be used to see generated assembler
. there are two jit_compile() functions and code flow for classic filters is:
sk_attach_filter() - load classic BPF
bpf_jit_compile() - try to JIT from classic BPF
sk_convert_filter() - convert classic to internal
bpf_int_jit_compile() - JIT from internal BPF
seccomp and tracing filters will just call bpf_int_jit_compile()
Signed-off-by: Alexei Starovoitov <ast@plumgrid.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
2014-05-14 02:50:46 +00:00
|
|
|
break;
|
|
|
|
|
2014-06-06 21:46:06 +00:00
|
|
|
/* STX: *(u8*)(dst_reg + off) = src_reg */
|
net: filter: x86: internal BPF JIT
Maps all internal BPF instructions into x86_64 instructions.
This patch replaces original BPF x64 JIT with internal BPF x64 JIT.
sysctl net.core.bpf_jit_enable is reused as on/off switch.
Performance:
1. old BPF JIT and internal BPF JIT generate equivalent x86_64 code.
No performance difference is observed for filters that were JIT-able before
Example assembler code for BPF filter "tcpdump port 22"
original BPF -> old JIT: original BPF -> internal BPF -> new JIT:
0: push %rbp 0: push %rbp
1: mov %rsp,%rbp 1: mov %rsp,%rbp
4: sub $0x60,%rsp 4: sub $0x228,%rsp
8: mov %rbx,-0x8(%rbp) b: mov %rbx,-0x228(%rbp) // prologue
12: mov %r13,-0x220(%rbp)
19: mov %r14,-0x218(%rbp)
20: mov %r15,-0x210(%rbp)
27: xor %eax,%eax // clear A
c: xor %ebx,%ebx 29: xor %r13,%r13 // clear X
e: mov 0x68(%rdi),%r9d 2c: mov 0x68(%rdi),%r9d
12: sub 0x6c(%rdi),%r9d 30: sub 0x6c(%rdi),%r9d
16: mov 0xd8(%rdi),%r8 34: mov 0xd8(%rdi),%r10
3b: mov %rdi,%rbx
1d: mov $0xc,%esi 3e: mov $0xc,%esi
22: callq 0xffffffffe1021e15 43: callq 0xffffffffe102bd75
27: cmp $0x86dd,%eax 48: cmp $0x86dd,%rax
2c: jne 0x0000000000000069 4f: jne 0x000000000000009a
2e: mov $0x14,%esi 51: mov $0x14,%esi
33: callq 0xffffffffe1021e31 56: callq 0xffffffffe102bd91
38: cmp $0x84,%eax 5b: cmp $0x84,%rax
3d: je 0x0000000000000049 62: je 0x0000000000000074
3f: cmp $0x6,%eax 64: cmp $0x6,%rax
42: je 0x0000000000000049 68: je 0x0000000000000074
44: cmp $0x11,%eax 6a: cmp $0x11,%rax
47: jne 0x00000000000000c6 6e: jne 0x0000000000000117
49: mov $0x36,%esi 74: mov $0x36,%esi
4e: callq 0xffffffffe1021e15 79: callq 0xffffffffe102bd75
53: cmp $0x16,%eax 7e: cmp $0x16,%rax
56: je 0x00000000000000bf 82: je 0x0000000000000110
58: mov $0x38,%esi 88: mov $0x38,%esi
5d: callq 0xffffffffe1021e15 8d: callq 0xffffffffe102bd75
62: cmp $0x16,%eax 92: cmp $0x16,%rax
65: je 0x00000000000000bf 96: je 0x0000000000000110
67: jmp 0x00000000000000c6 98: jmp 0x0000000000000117
69: cmp $0x800,%eax 9a: cmp $0x800,%rax
6e: jne 0x00000000000000c6 a1: jne 0x0000000000000117
70: mov $0x17,%esi a3: mov $0x17,%esi
75: callq 0xffffffffe1021e31 a8: callq 0xffffffffe102bd91
7a: cmp $0x84,%eax ad: cmp $0x84,%rax
7f: je 0x000000000000008b b4: je 0x00000000000000c2
81: cmp $0x6,%eax b6: cmp $0x6,%rax
84: je 0x000000000000008b ba: je 0x00000000000000c2
86: cmp $0x11,%eax bc: cmp $0x11,%rax
89: jne 0x00000000000000c6 c0: jne 0x0000000000000117
8b: mov $0x14,%esi c2: mov $0x14,%esi
90: callq 0xffffffffe1021e15 c7: callq 0xffffffffe102bd75
95: test $0x1fff,%ax cc: test $0x1fff,%rax
99: jne 0x00000000000000c6 d3: jne 0x0000000000000117
d5: mov %rax,%r14
9b: mov $0xe,%esi d8: mov $0xe,%esi
a0: callq 0xffffffffe1021e44 dd: callq 0xffffffffe102bd91 // MSH
e2: and $0xf,%eax
e5: shl $0x2,%eax
e8: mov %rax,%r13
eb: mov %r14,%rax
ee: mov %r13,%rsi
a5: lea 0xe(%rbx),%esi f1: add $0xe,%esi
a8: callq 0xffffffffe1021e0d f4: callq 0xffffffffe102bd6d
ad: cmp $0x16,%eax f9: cmp $0x16,%rax
b0: je 0x00000000000000bf fd: je 0x0000000000000110
ff: mov %r13,%rsi
b2: lea 0x10(%rbx),%esi 102: add $0x10,%esi
b5: callq 0xffffffffe1021e0d 105: callq 0xffffffffe102bd6d
ba: cmp $0x16,%eax 10a: cmp $0x16,%rax
bd: jne 0x00000000000000c6 10e: jne 0x0000000000000117
bf: mov $0xffff,%eax 110: mov $0xffff,%eax
c4: jmp 0x00000000000000c8 115: jmp 0x000000000000011c
c6: xor %eax,%eax 117: mov $0x0,%eax
c8: mov -0x8(%rbp),%rbx 11c: mov -0x228(%rbp),%rbx // epilogue
cc: leaveq 123: mov -0x220(%rbp),%r13
cd: retq 12a: mov -0x218(%rbp),%r14
131: mov -0x210(%rbp),%r15
138: leaveq
139: retq
On fully cached SKBs both JITed functions take 12 nsec to execute.
BPF interpreter executes the program in 30 nsec.
The difference in generated assembler is due to the following:
Old BPF imlements LDX_MSH instruction via sk_load_byte_msh() helper function
inside bpf_jit.S.
New JIT removes the helper and does it explicitly, so ldx_msh cost
is the same for both JITs, but generated code looks longer.
New JIT has 4 registers to save, so prologue/epilogue are larger,
but the cost is within noise on x64.
Old JIT checks whether first insn clears A and if not emits 'xor %eax,%eax'.
New JIT clears %rax unconditionally.
2. old BPF JIT doesn't support ANC_NLATTR, ANC_PAY_OFFSET, ANC_RANDOM
extensions. New JIT supports all BPF extensions.
Performance of such filters improves 2-4 times depending on a filter.
The longer the filter the higher performance gain.
Synthetic benchmarks with many ancillary loads see 20x speedup
which seems to be the maximum gain from JIT
Notes:
. net.core.bpf_jit_enable=2 + tools/net/bpf_jit_disasm is still functional
and can be used to see generated assembler
. there are two jit_compile() functions and code flow for classic filters is:
sk_attach_filter() - load classic BPF
bpf_jit_compile() - try to JIT from classic BPF
sk_convert_filter() - convert classic to internal
bpf_int_jit_compile() - JIT from internal BPF
seccomp and tracing filters will just call bpf_int_jit_compile()
Signed-off-by: Alexei Starovoitov <ast@plumgrid.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
2014-05-14 02:50:46 +00:00
|
|
|
case BPF_STX | BPF_MEM | BPF_B:
|
|
|
|
case BPF_STX | BPF_MEM | BPF_H:
|
|
|
|
case BPF_STX | BPF_MEM | BPF_W:
|
|
|
|
case BPF_STX | BPF_MEM | BPF_DW:
|
2019-11-14 18:57:02 +00:00
|
|
|
emit_stx(&prog, BPF_SIZE(insn->code), dst_reg, src_reg, insn->off);
|
net: filter: x86: internal BPF JIT
Maps all internal BPF instructions into x86_64 instructions.
This patch replaces original BPF x64 JIT with internal BPF x64 JIT.
sysctl net.core.bpf_jit_enable is reused as on/off switch.
Performance:
1. old BPF JIT and internal BPF JIT generate equivalent x86_64 code.
No performance difference is observed for filters that were JIT-able before
Example assembler code for BPF filter "tcpdump port 22"
original BPF -> old JIT: original BPF -> internal BPF -> new JIT:
0: push %rbp 0: push %rbp
1: mov %rsp,%rbp 1: mov %rsp,%rbp
4: sub $0x60,%rsp 4: sub $0x228,%rsp
8: mov %rbx,-0x8(%rbp) b: mov %rbx,-0x228(%rbp) // prologue
12: mov %r13,-0x220(%rbp)
19: mov %r14,-0x218(%rbp)
20: mov %r15,-0x210(%rbp)
27: xor %eax,%eax // clear A
c: xor %ebx,%ebx 29: xor %r13,%r13 // clear X
e: mov 0x68(%rdi),%r9d 2c: mov 0x68(%rdi),%r9d
12: sub 0x6c(%rdi),%r9d 30: sub 0x6c(%rdi),%r9d
16: mov 0xd8(%rdi),%r8 34: mov 0xd8(%rdi),%r10
3b: mov %rdi,%rbx
1d: mov $0xc,%esi 3e: mov $0xc,%esi
22: callq 0xffffffffe1021e15 43: callq 0xffffffffe102bd75
27: cmp $0x86dd,%eax 48: cmp $0x86dd,%rax
2c: jne 0x0000000000000069 4f: jne 0x000000000000009a
2e: mov $0x14,%esi 51: mov $0x14,%esi
33: callq 0xffffffffe1021e31 56: callq 0xffffffffe102bd91
38: cmp $0x84,%eax 5b: cmp $0x84,%rax
3d: je 0x0000000000000049 62: je 0x0000000000000074
3f: cmp $0x6,%eax 64: cmp $0x6,%rax
42: je 0x0000000000000049 68: je 0x0000000000000074
44: cmp $0x11,%eax 6a: cmp $0x11,%rax
47: jne 0x00000000000000c6 6e: jne 0x0000000000000117
49: mov $0x36,%esi 74: mov $0x36,%esi
4e: callq 0xffffffffe1021e15 79: callq 0xffffffffe102bd75
53: cmp $0x16,%eax 7e: cmp $0x16,%rax
56: je 0x00000000000000bf 82: je 0x0000000000000110
58: mov $0x38,%esi 88: mov $0x38,%esi
5d: callq 0xffffffffe1021e15 8d: callq 0xffffffffe102bd75
62: cmp $0x16,%eax 92: cmp $0x16,%rax
65: je 0x00000000000000bf 96: je 0x0000000000000110
67: jmp 0x00000000000000c6 98: jmp 0x0000000000000117
69: cmp $0x800,%eax 9a: cmp $0x800,%rax
6e: jne 0x00000000000000c6 a1: jne 0x0000000000000117
70: mov $0x17,%esi a3: mov $0x17,%esi
75: callq 0xffffffffe1021e31 a8: callq 0xffffffffe102bd91
7a: cmp $0x84,%eax ad: cmp $0x84,%rax
7f: je 0x000000000000008b b4: je 0x00000000000000c2
81: cmp $0x6,%eax b6: cmp $0x6,%rax
84: je 0x000000000000008b ba: je 0x00000000000000c2
86: cmp $0x11,%eax bc: cmp $0x11,%rax
89: jne 0x00000000000000c6 c0: jne 0x0000000000000117
8b: mov $0x14,%esi c2: mov $0x14,%esi
90: callq 0xffffffffe1021e15 c7: callq 0xffffffffe102bd75
95: test $0x1fff,%ax cc: test $0x1fff,%rax
99: jne 0x00000000000000c6 d3: jne 0x0000000000000117
d5: mov %rax,%r14
9b: mov $0xe,%esi d8: mov $0xe,%esi
a0: callq 0xffffffffe1021e44 dd: callq 0xffffffffe102bd91 // MSH
e2: and $0xf,%eax
e5: shl $0x2,%eax
e8: mov %rax,%r13
eb: mov %r14,%rax
ee: mov %r13,%rsi
a5: lea 0xe(%rbx),%esi f1: add $0xe,%esi
a8: callq 0xffffffffe1021e0d f4: callq 0xffffffffe102bd6d
ad: cmp $0x16,%eax f9: cmp $0x16,%rax
b0: je 0x00000000000000bf fd: je 0x0000000000000110
ff: mov %r13,%rsi
b2: lea 0x10(%rbx),%esi 102: add $0x10,%esi
b5: callq 0xffffffffe1021e0d 105: callq 0xffffffffe102bd6d
ba: cmp $0x16,%eax 10a: cmp $0x16,%rax
bd: jne 0x00000000000000c6 10e: jne 0x0000000000000117
bf: mov $0xffff,%eax 110: mov $0xffff,%eax
c4: jmp 0x00000000000000c8 115: jmp 0x000000000000011c
c6: xor %eax,%eax 117: mov $0x0,%eax
c8: mov -0x8(%rbp),%rbx 11c: mov -0x228(%rbp),%rbx // epilogue
cc: leaveq 123: mov -0x220(%rbp),%r13
cd: retq 12a: mov -0x218(%rbp),%r14
131: mov -0x210(%rbp),%r15
138: leaveq
139: retq
On fully cached SKBs both JITed functions take 12 nsec to execute.
BPF interpreter executes the program in 30 nsec.
The difference in generated assembler is due to the following:
Old BPF imlements LDX_MSH instruction via sk_load_byte_msh() helper function
inside bpf_jit.S.
New JIT removes the helper and does it explicitly, so ldx_msh cost
is the same for both JITs, but generated code looks longer.
New JIT has 4 registers to save, so prologue/epilogue are larger,
but the cost is within noise on x64.
Old JIT checks whether first insn clears A and if not emits 'xor %eax,%eax'.
New JIT clears %rax unconditionally.
2. old BPF JIT doesn't support ANC_NLATTR, ANC_PAY_OFFSET, ANC_RANDOM
extensions. New JIT supports all BPF extensions.
Performance of such filters improves 2-4 times depending on a filter.
The longer the filter the higher performance gain.
Synthetic benchmarks with many ancillary loads see 20x speedup
which seems to be the maximum gain from JIT
Notes:
. net.core.bpf_jit_enable=2 + tools/net/bpf_jit_disasm is still functional
and can be used to see generated assembler
. there are two jit_compile() functions and code flow for classic filters is:
sk_attach_filter() - load classic BPF
bpf_jit_compile() - try to JIT from classic BPF
sk_convert_filter() - convert classic to internal
bpf_int_jit_compile() - JIT from internal BPF
seccomp and tracing filters will just call bpf_int_jit_compile()
Signed-off-by: Alexei Starovoitov <ast@plumgrid.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
2014-05-14 02:50:46 +00:00
|
|
|
break;
|
|
|
|
|
2024-03-08 01:08:01 +00:00
|
|
|
case BPF_ST | BPF_PROBE_MEM32 | BPF_B:
|
|
|
|
case BPF_ST | BPF_PROBE_MEM32 | BPF_H:
|
|
|
|
case BPF_ST | BPF_PROBE_MEM32 | BPF_W:
|
|
|
|
case BPF_ST | BPF_PROBE_MEM32 | BPF_DW:
|
|
|
|
start_of_ldx = prog;
|
|
|
|
emit_st_r12(&prog, BPF_SIZE(insn->code), dst_reg, insn->off, insn->imm);
|
|
|
|
goto populate_extable;
|
|
|
|
|
|
|
|
/* LDX: dst_reg = *(u8*)(src_reg + r12 + off) */
|
|
|
|
case BPF_LDX | BPF_PROBE_MEM32 | BPF_B:
|
|
|
|
case BPF_LDX | BPF_PROBE_MEM32 | BPF_H:
|
|
|
|
case BPF_LDX | BPF_PROBE_MEM32 | BPF_W:
|
|
|
|
case BPF_LDX | BPF_PROBE_MEM32 | BPF_DW:
|
|
|
|
case BPF_STX | BPF_PROBE_MEM32 | BPF_B:
|
|
|
|
case BPF_STX | BPF_PROBE_MEM32 | BPF_H:
|
|
|
|
case BPF_STX | BPF_PROBE_MEM32 | BPF_W:
|
|
|
|
case BPF_STX | BPF_PROBE_MEM32 | BPF_DW:
|
|
|
|
start_of_ldx = prog;
|
|
|
|
if (BPF_CLASS(insn->code) == BPF_LDX)
|
|
|
|
emit_ldx_r12(&prog, BPF_SIZE(insn->code), dst_reg, src_reg, insn->off);
|
|
|
|
else
|
|
|
|
emit_stx_r12(&prog, BPF_SIZE(insn->code), dst_reg, src_reg, insn->off);
|
|
|
|
populate_extable:
|
|
|
|
{
|
|
|
|
struct exception_table_entry *ex;
|
|
|
|
u8 *_insn = image + proglen + (start_of_ldx - temp);
|
|
|
|
s64 delta;
|
|
|
|
|
|
|
|
if (!bpf_prog->aux->extable)
|
|
|
|
break;
|
|
|
|
|
|
|
|
if (excnt >= bpf_prog->aux->num_exentries) {
|
|
|
|
pr_err("mem32 extable bug\n");
|
|
|
|
return -EFAULT;
|
|
|
|
}
|
|
|
|
ex = &bpf_prog->aux->extable[excnt++];
|
|
|
|
|
|
|
|
delta = _insn - (u8 *)&ex->insn;
|
|
|
|
/* switch ex to rw buffer for writes */
|
|
|
|
ex = (void *)rw_image + ((void *)ex - (void *)image);
|
|
|
|
|
|
|
|
ex->insn = delta;
|
|
|
|
|
|
|
|
ex->data = EX_TYPE_BPF;
|
|
|
|
|
|
|
|
ex->fixup = (prog - start_of_ldx) |
|
|
|
|
((BPF_CLASS(insn->code) == BPF_LDX ? reg2pt_regs[dst_reg] : DONT_CLEAR) << 8);
|
|
|
|
}
|
|
|
|
break;
|
|
|
|
|
2014-06-06 21:46:06 +00:00
|
|
|
/* LDX: dst_reg = *(u8*)(src_reg + off) */
|
net: filter: x86: internal BPF JIT
Maps all internal BPF instructions into x86_64 instructions.
This patch replaces original BPF x64 JIT with internal BPF x64 JIT.
sysctl net.core.bpf_jit_enable is reused as on/off switch.
Performance:
1. old BPF JIT and internal BPF JIT generate equivalent x86_64 code.
No performance difference is observed for filters that were JIT-able before
Example assembler code for BPF filter "tcpdump port 22"
original BPF -> old JIT: original BPF -> internal BPF -> new JIT:
0: push %rbp 0: push %rbp
1: mov %rsp,%rbp 1: mov %rsp,%rbp
4: sub $0x60,%rsp 4: sub $0x228,%rsp
8: mov %rbx,-0x8(%rbp) b: mov %rbx,-0x228(%rbp) // prologue
12: mov %r13,-0x220(%rbp)
19: mov %r14,-0x218(%rbp)
20: mov %r15,-0x210(%rbp)
27: xor %eax,%eax // clear A
c: xor %ebx,%ebx 29: xor %r13,%r13 // clear X
e: mov 0x68(%rdi),%r9d 2c: mov 0x68(%rdi),%r9d
12: sub 0x6c(%rdi),%r9d 30: sub 0x6c(%rdi),%r9d
16: mov 0xd8(%rdi),%r8 34: mov 0xd8(%rdi),%r10
3b: mov %rdi,%rbx
1d: mov $0xc,%esi 3e: mov $0xc,%esi
22: callq 0xffffffffe1021e15 43: callq 0xffffffffe102bd75
27: cmp $0x86dd,%eax 48: cmp $0x86dd,%rax
2c: jne 0x0000000000000069 4f: jne 0x000000000000009a
2e: mov $0x14,%esi 51: mov $0x14,%esi
33: callq 0xffffffffe1021e31 56: callq 0xffffffffe102bd91
38: cmp $0x84,%eax 5b: cmp $0x84,%rax
3d: je 0x0000000000000049 62: je 0x0000000000000074
3f: cmp $0x6,%eax 64: cmp $0x6,%rax
42: je 0x0000000000000049 68: je 0x0000000000000074
44: cmp $0x11,%eax 6a: cmp $0x11,%rax
47: jne 0x00000000000000c6 6e: jne 0x0000000000000117
49: mov $0x36,%esi 74: mov $0x36,%esi
4e: callq 0xffffffffe1021e15 79: callq 0xffffffffe102bd75
53: cmp $0x16,%eax 7e: cmp $0x16,%rax
56: je 0x00000000000000bf 82: je 0x0000000000000110
58: mov $0x38,%esi 88: mov $0x38,%esi
5d: callq 0xffffffffe1021e15 8d: callq 0xffffffffe102bd75
62: cmp $0x16,%eax 92: cmp $0x16,%rax
65: je 0x00000000000000bf 96: je 0x0000000000000110
67: jmp 0x00000000000000c6 98: jmp 0x0000000000000117
69: cmp $0x800,%eax 9a: cmp $0x800,%rax
6e: jne 0x00000000000000c6 a1: jne 0x0000000000000117
70: mov $0x17,%esi a3: mov $0x17,%esi
75: callq 0xffffffffe1021e31 a8: callq 0xffffffffe102bd91
7a: cmp $0x84,%eax ad: cmp $0x84,%rax
7f: je 0x000000000000008b b4: je 0x00000000000000c2
81: cmp $0x6,%eax b6: cmp $0x6,%rax
84: je 0x000000000000008b ba: je 0x00000000000000c2
86: cmp $0x11,%eax bc: cmp $0x11,%rax
89: jne 0x00000000000000c6 c0: jne 0x0000000000000117
8b: mov $0x14,%esi c2: mov $0x14,%esi
90: callq 0xffffffffe1021e15 c7: callq 0xffffffffe102bd75
95: test $0x1fff,%ax cc: test $0x1fff,%rax
99: jne 0x00000000000000c6 d3: jne 0x0000000000000117
d5: mov %rax,%r14
9b: mov $0xe,%esi d8: mov $0xe,%esi
a0: callq 0xffffffffe1021e44 dd: callq 0xffffffffe102bd91 // MSH
e2: and $0xf,%eax
e5: shl $0x2,%eax
e8: mov %rax,%r13
eb: mov %r14,%rax
ee: mov %r13,%rsi
a5: lea 0xe(%rbx),%esi f1: add $0xe,%esi
a8: callq 0xffffffffe1021e0d f4: callq 0xffffffffe102bd6d
ad: cmp $0x16,%eax f9: cmp $0x16,%rax
b0: je 0x00000000000000bf fd: je 0x0000000000000110
ff: mov %r13,%rsi
b2: lea 0x10(%rbx),%esi 102: add $0x10,%esi
b5: callq 0xffffffffe1021e0d 105: callq 0xffffffffe102bd6d
ba: cmp $0x16,%eax 10a: cmp $0x16,%rax
bd: jne 0x00000000000000c6 10e: jne 0x0000000000000117
bf: mov $0xffff,%eax 110: mov $0xffff,%eax
c4: jmp 0x00000000000000c8 115: jmp 0x000000000000011c
c6: xor %eax,%eax 117: mov $0x0,%eax
c8: mov -0x8(%rbp),%rbx 11c: mov -0x228(%rbp),%rbx // epilogue
cc: leaveq 123: mov -0x220(%rbp),%r13
cd: retq 12a: mov -0x218(%rbp),%r14
131: mov -0x210(%rbp),%r15
138: leaveq
139: retq
On fully cached SKBs both JITed functions take 12 nsec to execute.
BPF interpreter executes the program in 30 nsec.
The difference in generated assembler is due to the following:
Old BPF imlements LDX_MSH instruction via sk_load_byte_msh() helper function
inside bpf_jit.S.
New JIT removes the helper and does it explicitly, so ldx_msh cost
is the same for both JITs, but generated code looks longer.
New JIT has 4 registers to save, so prologue/epilogue are larger,
but the cost is within noise on x64.
Old JIT checks whether first insn clears A and if not emits 'xor %eax,%eax'.
New JIT clears %rax unconditionally.
2. old BPF JIT doesn't support ANC_NLATTR, ANC_PAY_OFFSET, ANC_RANDOM
extensions. New JIT supports all BPF extensions.
Performance of such filters improves 2-4 times depending on a filter.
The longer the filter the higher performance gain.
Synthetic benchmarks with many ancillary loads see 20x speedup
which seems to be the maximum gain from JIT
Notes:
. net.core.bpf_jit_enable=2 + tools/net/bpf_jit_disasm is still functional
and can be used to see generated assembler
. there are two jit_compile() functions and code flow for classic filters is:
sk_attach_filter() - load classic BPF
bpf_jit_compile() - try to JIT from classic BPF
sk_convert_filter() - convert classic to internal
bpf_int_jit_compile() - JIT from internal BPF
seccomp and tracing filters will just call bpf_int_jit_compile()
Signed-off-by: Alexei Starovoitov <ast@plumgrid.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
2014-05-14 02:50:46 +00:00
|
|
|
case BPF_LDX | BPF_MEM | BPF_B:
|
2019-10-16 03:25:03 +00:00
|
|
|
case BPF_LDX | BPF_PROBE_MEM | BPF_B:
|
net: filter: x86: internal BPF JIT
Maps all internal BPF instructions into x86_64 instructions.
This patch replaces original BPF x64 JIT with internal BPF x64 JIT.
sysctl net.core.bpf_jit_enable is reused as on/off switch.
Performance:
1. old BPF JIT and internal BPF JIT generate equivalent x86_64 code.
No performance difference is observed for filters that were JIT-able before
Example assembler code for BPF filter "tcpdump port 22"
original BPF -> old JIT: original BPF -> internal BPF -> new JIT:
0: push %rbp 0: push %rbp
1: mov %rsp,%rbp 1: mov %rsp,%rbp
4: sub $0x60,%rsp 4: sub $0x228,%rsp
8: mov %rbx,-0x8(%rbp) b: mov %rbx,-0x228(%rbp) // prologue
12: mov %r13,-0x220(%rbp)
19: mov %r14,-0x218(%rbp)
20: mov %r15,-0x210(%rbp)
27: xor %eax,%eax // clear A
c: xor %ebx,%ebx 29: xor %r13,%r13 // clear X
e: mov 0x68(%rdi),%r9d 2c: mov 0x68(%rdi),%r9d
12: sub 0x6c(%rdi),%r9d 30: sub 0x6c(%rdi),%r9d
16: mov 0xd8(%rdi),%r8 34: mov 0xd8(%rdi),%r10
3b: mov %rdi,%rbx
1d: mov $0xc,%esi 3e: mov $0xc,%esi
22: callq 0xffffffffe1021e15 43: callq 0xffffffffe102bd75
27: cmp $0x86dd,%eax 48: cmp $0x86dd,%rax
2c: jne 0x0000000000000069 4f: jne 0x000000000000009a
2e: mov $0x14,%esi 51: mov $0x14,%esi
33: callq 0xffffffffe1021e31 56: callq 0xffffffffe102bd91
38: cmp $0x84,%eax 5b: cmp $0x84,%rax
3d: je 0x0000000000000049 62: je 0x0000000000000074
3f: cmp $0x6,%eax 64: cmp $0x6,%rax
42: je 0x0000000000000049 68: je 0x0000000000000074
44: cmp $0x11,%eax 6a: cmp $0x11,%rax
47: jne 0x00000000000000c6 6e: jne 0x0000000000000117
49: mov $0x36,%esi 74: mov $0x36,%esi
4e: callq 0xffffffffe1021e15 79: callq 0xffffffffe102bd75
53: cmp $0x16,%eax 7e: cmp $0x16,%rax
56: je 0x00000000000000bf 82: je 0x0000000000000110
58: mov $0x38,%esi 88: mov $0x38,%esi
5d: callq 0xffffffffe1021e15 8d: callq 0xffffffffe102bd75
62: cmp $0x16,%eax 92: cmp $0x16,%rax
65: je 0x00000000000000bf 96: je 0x0000000000000110
67: jmp 0x00000000000000c6 98: jmp 0x0000000000000117
69: cmp $0x800,%eax 9a: cmp $0x800,%rax
6e: jne 0x00000000000000c6 a1: jne 0x0000000000000117
70: mov $0x17,%esi a3: mov $0x17,%esi
75: callq 0xffffffffe1021e31 a8: callq 0xffffffffe102bd91
7a: cmp $0x84,%eax ad: cmp $0x84,%rax
7f: je 0x000000000000008b b4: je 0x00000000000000c2
81: cmp $0x6,%eax b6: cmp $0x6,%rax
84: je 0x000000000000008b ba: je 0x00000000000000c2
86: cmp $0x11,%eax bc: cmp $0x11,%rax
89: jne 0x00000000000000c6 c0: jne 0x0000000000000117
8b: mov $0x14,%esi c2: mov $0x14,%esi
90: callq 0xffffffffe1021e15 c7: callq 0xffffffffe102bd75
95: test $0x1fff,%ax cc: test $0x1fff,%rax
99: jne 0x00000000000000c6 d3: jne 0x0000000000000117
d5: mov %rax,%r14
9b: mov $0xe,%esi d8: mov $0xe,%esi
a0: callq 0xffffffffe1021e44 dd: callq 0xffffffffe102bd91 // MSH
e2: and $0xf,%eax
e5: shl $0x2,%eax
e8: mov %rax,%r13
eb: mov %r14,%rax
ee: mov %r13,%rsi
a5: lea 0xe(%rbx),%esi f1: add $0xe,%esi
a8: callq 0xffffffffe1021e0d f4: callq 0xffffffffe102bd6d
ad: cmp $0x16,%eax f9: cmp $0x16,%rax
b0: je 0x00000000000000bf fd: je 0x0000000000000110
ff: mov %r13,%rsi
b2: lea 0x10(%rbx),%esi 102: add $0x10,%esi
b5: callq 0xffffffffe1021e0d 105: callq 0xffffffffe102bd6d
ba: cmp $0x16,%eax 10a: cmp $0x16,%rax
bd: jne 0x00000000000000c6 10e: jne 0x0000000000000117
bf: mov $0xffff,%eax 110: mov $0xffff,%eax
c4: jmp 0x00000000000000c8 115: jmp 0x000000000000011c
c6: xor %eax,%eax 117: mov $0x0,%eax
c8: mov -0x8(%rbp),%rbx 11c: mov -0x228(%rbp),%rbx // epilogue
cc: leaveq 123: mov -0x220(%rbp),%r13
cd: retq 12a: mov -0x218(%rbp),%r14
131: mov -0x210(%rbp),%r15
138: leaveq
139: retq
On fully cached SKBs both JITed functions take 12 nsec to execute.
BPF interpreter executes the program in 30 nsec.
The difference in generated assembler is due to the following:
Old BPF imlements LDX_MSH instruction via sk_load_byte_msh() helper function
inside bpf_jit.S.
New JIT removes the helper and does it explicitly, so ldx_msh cost
is the same for both JITs, but generated code looks longer.
New JIT has 4 registers to save, so prologue/epilogue are larger,
but the cost is within noise on x64.
Old JIT checks whether first insn clears A and if not emits 'xor %eax,%eax'.
New JIT clears %rax unconditionally.
2. old BPF JIT doesn't support ANC_NLATTR, ANC_PAY_OFFSET, ANC_RANDOM
extensions. New JIT supports all BPF extensions.
Performance of such filters improves 2-4 times depending on a filter.
The longer the filter the higher performance gain.
Synthetic benchmarks with many ancillary loads see 20x speedup
which seems to be the maximum gain from JIT
Notes:
. net.core.bpf_jit_enable=2 + tools/net/bpf_jit_disasm is still functional
and can be used to see generated assembler
. there are two jit_compile() functions and code flow for classic filters is:
sk_attach_filter() - load classic BPF
bpf_jit_compile() - try to JIT from classic BPF
sk_convert_filter() - convert classic to internal
bpf_int_jit_compile() - JIT from internal BPF
seccomp and tracing filters will just call bpf_int_jit_compile()
Signed-off-by: Alexei Starovoitov <ast@plumgrid.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
2014-05-14 02:50:46 +00:00
|
|
|
case BPF_LDX | BPF_MEM | BPF_H:
|
2019-10-16 03:25:03 +00:00
|
|
|
case BPF_LDX | BPF_PROBE_MEM | BPF_H:
|
net: filter: x86: internal BPF JIT
Maps all internal BPF instructions into x86_64 instructions.
This patch replaces original BPF x64 JIT with internal BPF x64 JIT.
sysctl net.core.bpf_jit_enable is reused as on/off switch.
Performance:
1. old BPF JIT and internal BPF JIT generate equivalent x86_64 code.
No performance difference is observed for filters that were JIT-able before
Example assembler code for BPF filter "tcpdump port 22"
original BPF -> old JIT: original BPF -> internal BPF -> new JIT:
0: push %rbp 0: push %rbp
1: mov %rsp,%rbp 1: mov %rsp,%rbp
4: sub $0x60,%rsp 4: sub $0x228,%rsp
8: mov %rbx,-0x8(%rbp) b: mov %rbx,-0x228(%rbp) // prologue
12: mov %r13,-0x220(%rbp)
19: mov %r14,-0x218(%rbp)
20: mov %r15,-0x210(%rbp)
27: xor %eax,%eax // clear A
c: xor %ebx,%ebx 29: xor %r13,%r13 // clear X
e: mov 0x68(%rdi),%r9d 2c: mov 0x68(%rdi),%r9d
12: sub 0x6c(%rdi),%r9d 30: sub 0x6c(%rdi),%r9d
16: mov 0xd8(%rdi),%r8 34: mov 0xd8(%rdi),%r10
3b: mov %rdi,%rbx
1d: mov $0xc,%esi 3e: mov $0xc,%esi
22: callq 0xffffffffe1021e15 43: callq 0xffffffffe102bd75
27: cmp $0x86dd,%eax 48: cmp $0x86dd,%rax
2c: jne 0x0000000000000069 4f: jne 0x000000000000009a
2e: mov $0x14,%esi 51: mov $0x14,%esi
33: callq 0xffffffffe1021e31 56: callq 0xffffffffe102bd91
38: cmp $0x84,%eax 5b: cmp $0x84,%rax
3d: je 0x0000000000000049 62: je 0x0000000000000074
3f: cmp $0x6,%eax 64: cmp $0x6,%rax
42: je 0x0000000000000049 68: je 0x0000000000000074
44: cmp $0x11,%eax 6a: cmp $0x11,%rax
47: jne 0x00000000000000c6 6e: jne 0x0000000000000117
49: mov $0x36,%esi 74: mov $0x36,%esi
4e: callq 0xffffffffe1021e15 79: callq 0xffffffffe102bd75
53: cmp $0x16,%eax 7e: cmp $0x16,%rax
56: je 0x00000000000000bf 82: je 0x0000000000000110
58: mov $0x38,%esi 88: mov $0x38,%esi
5d: callq 0xffffffffe1021e15 8d: callq 0xffffffffe102bd75
62: cmp $0x16,%eax 92: cmp $0x16,%rax
65: je 0x00000000000000bf 96: je 0x0000000000000110
67: jmp 0x00000000000000c6 98: jmp 0x0000000000000117
69: cmp $0x800,%eax 9a: cmp $0x800,%rax
6e: jne 0x00000000000000c6 a1: jne 0x0000000000000117
70: mov $0x17,%esi a3: mov $0x17,%esi
75: callq 0xffffffffe1021e31 a8: callq 0xffffffffe102bd91
7a: cmp $0x84,%eax ad: cmp $0x84,%rax
7f: je 0x000000000000008b b4: je 0x00000000000000c2
81: cmp $0x6,%eax b6: cmp $0x6,%rax
84: je 0x000000000000008b ba: je 0x00000000000000c2
86: cmp $0x11,%eax bc: cmp $0x11,%rax
89: jne 0x00000000000000c6 c0: jne 0x0000000000000117
8b: mov $0x14,%esi c2: mov $0x14,%esi
90: callq 0xffffffffe1021e15 c7: callq 0xffffffffe102bd75
95: test $0x1fff,%ax cc: test $0x1fff,%rax
99: jne 0x00000000000000c6 d3: jne 0x0000000000000117
d5: mov %rax,%r14
9b: mov $0xe,%esi d8: mov $0xe,%esi
a0: callq 0xffffffffe1021e44 dd: callq 0xffffffffe102bd91 // MSH
e2: and $0xf,%eax
e5: shl $0x2,%eax
e8: mov %rax,%r13
eb: mov %r14,%rax
ee: mov %r13,%rsi
a5: lea 0xe(%rbx),%esi f1: add $0xe,%esi
a8: callq 0xffffffffe1021e0d f4: callq 0xffffffffe102bd6d
ad: cmp $0x16,%eax f9: cmp $0x16,%rax
b0: je 0x00000000000000bf fd: je 0x0000000000000110
ff: mov %r13,%rsi
b2: lea 0x10(%rbx),%esi 102: add $0x10,%esi
b5: callq 0xffffffffe1021e0d 105: callq 0xffffffffe102bd6d
ba: cmp $0x16,%eax 10a: cmp $0x16,%rax
bd: jne 0x00000000000000c6 10e: jne 0x0000000000000117
bf: mov $0xffff,%eax 110: mov $0xffff,%eax
c4: jmp 0x00000000000000c8 115: jmp 0x000000000000011c
c6: xor %eax,%eax 117: mov $0x0,%eax
c8: mov -0x8(%rbp),%rbx 11c: mov -0x228(%rbp),%rbx // epilogue
cc: leaveq 123: mov -0x220(%rbp),%r13
cd: retq 12a: mov -0x218(%rbp),%r14
131: mov -0x210(%rbp),%r15
138: leaveq
139: retq
On fully cached SKBs both JITed functions take 12 nsec to execute.
BPF interpreter executes the program in 30 nsec.
The difference in generated assembler is due to the following:
Old BPF imlements LDX_MSH instruction via sk_load_byte_msh() helper function
inside bpf_jit.S.
New JIT removes the helper and does it explicitly, so ldx_msh cost
is the same for both JITs, but generated code looks longer.
New JIT has 4 registers to save, so prologue/epilogue are larger,
but the cost is within noise on x64.
Old JIT checks whether first insn clears A and if not emits 'xor %eax,%eax'.
New JIT clears %rax unconditionally.
2. old BPF JIT doesn't support ANC_NLATTR, ANC_PAY_OFFSET, ANC_RANDOM
extensions. New JIT supports all BPF extensions.
Performance of such filters improves 2-4 times depending on a filter.
The longer the filter the higher performance gain.
Synthetic benchmarks with many ancillary loads see 20x speedup
which seems to be the maximum gain from JIT
Notes:
. net.core.bpf_jit_enable=2 + tools/net/bpf_jit_disasm is still functional
and can be used to see generated assembler
. there are two jit_compile() functions and code flow for classic filters is:
sk_attach_filter() - load classic BPF
bpf_jit_compile() - try to JIT from classic BPF
sk_convert_filter() - convert classic to internal
bpf_int_jit_compile() - JIT from internal BPF
seccomp and tracing filters will just call bpf_int_jit_compile()
Signed-off-by: Alexei Starovoitov <ast@plumgrid.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
2014-05-14 02:50:46 +00:00
|
|
|
case BPF_LDX | BPF_MEM | BPF_W:
|
2019-10-16 03:25:03 +00:00
|
|
|
case BPF_LDX | BPF_PROBE_MEM | BPF_W:
|
net: filter: x86: internal BPF JIT
Maps all internal BPF instructions into x86_64 instructions.
This patch replaces original BPF x64 JIT with internal BPF x64 JIT.
sysctl net.core.bpf_jit_enable is reused as on/off switch.
Performance:
1. old BPF JIT and internal BPF JIT generate equivalent x86_64 code.
No performance difference is observed for filters that were JIT-able before
Example assembler code for BPF filter "tcpdump port 22"
original BPF -> old JIT: original BPF -> internal BPF -> new JIT:
0: push %rbp 0: push %rbp
1: mov %rsp,%rbp 1: mov %rsp,%rbp
4: sub $0x60,%rsp 4: sub $0x228,%rsp
8: mov %rbx,-0x8(%rbp) b: mov %rbx,-0x228(%rbp) // prologue
12: mov %r13,-0x220(%rbp)
19: mov %r14,-0x218(%rbp)
20: mov %r15,-0x210(%rbp)
27: xor %eax,%eax // clear A
c: xor %ebx,%ebx 29: xor %r13,%r13 // clear X
e: mov 0x68(%rdi),%r9d 2c: mov 0x68(%rdi),%r9d
12: sub 0x6c(%rdi),%r9d 30: sub 0x6c(%rdi),%r9d
16: mov 0xd8(%rdi),%r8 34: mov 0xd8(%rdi),%r10
3b: mov %rdi,%rbx
1d: mov $0xc,%esi 3e: mov $0xc,%esi
22: callq 0xffffffffe1021e15 43: callq 0xffffffffe102bd75
27: cmp $0x86dd,%eax 48: cmp $0x86dd,%rax
2c: jne 0x0000000000000069 4f: jne 0x000000000000009a
2e: mov $0x14,%esi 51: mov $0x14,%esi
33: callq 0xffffffffe1021e31 56: callq 0xffffffffe102bd91
38: cmp $0x84,%eax 5b: cmp $0x84,%rax
3d: je 0x0000000000000049 62: je 0x0000000000000074
3f: cmp $0x6,%eax 64: cmp $0x6,%rax
42: je 0x0000000000000049 68: je 0x0000000000000074
44: cmp $0x11,%eax 6a: cmp $0x11,%rax
47: jne 0x00000000000000c6 6e: jne 0x0000000000000117
49: mov $0x36,%esi 74: mov $0x36,%esi
4e: callq 0xffffffffe1021e15 79: callq 0xffffffffe102bd75
53: cmp $0x16,%eax 7e: cmp $0x16,%rax
56: je 0x00000000000000bf 82: je 0x0000000000000110
58: mov $0x38,%esi 88: mov $0x38,%esi
5d: callq 0xffffffffe1021e15 8d: callq 0xffffffffe102bd75
62: cmp $0x16,%eax 92: cmp $0x16,%rax
65: je 0x00000000000000bf 96: je 0x0000000000000110
67: jmp 0x00000000000000c6 98: jmp 0x0000000000000117
69: cmp $0x800,%eax 9a: cmp $0x800,%rax
6e: jne 0x00000000000000c6 a1: jne 0x0000000000000117
70: mov $0x17,%esi a3: mov $0x17,%esi
75: callq 0xffffffffe1021e31 a8: callq 0xffffffffe102bd91
7a: cmp $0x84,%eax ad: cmp $0x84,%rax
7f: je 0x000000000000008b b4: je 0x00000000000000c2
81: cmp $0x6,%eax b6: cmp $0x6,%rax
84: je 0x000000000000008b ba: je 0x00000000000000c2
86: cmp $0x11,%eax bc: cmp $0x11,%rax
89: jne 0x00000000000000c6 c0: jne 0x0000000000000117
8b: mov $0x14,%esi c2: mov $0x14,%esi
90: callq 0xffffffffe1021e15 c7: callq 0xffffffffe102bd75
95: test $0x1fff,%ax cc: test $0x1fff,%rax
99: jne 0x00000000000000c6 d3: jne 0x0000000000000117
d5: mov %rax,%r14
9b: mov $0xe,%esi d8: mov $0xe,%esi
a0: callq 0xffffffffe1021e44 dd: callq 0xffffffffe102bd91 // MSH
e2: and $0xf,%eax
e5: shl $0x2,%eax
e8: mov %rax,%r13
eb: mov %r14,%rax
ee: mov %r13,%rsi
a5: lea 0xe(%rbx),%esi f1: add $0xe,%esi
a8: callq 0xffffffffe1021e0d f4: callq 0xffffffffe102bd6d
ad: cmp $0x16,%eax f9: cmp $0x16,%rax
b0: je 0x00000000000000bf fd: je 0x0000000000000110
ff: mov %r13,%rsi
b2: lea 0x10(%rbx),%esi 102: add $0x10,%esi
b5: callq 0xffffffffe1021e0d 105: callq 0xffffffffe102bd6d
ba: cmp $0x16,%eax 10a: cmp $0x16,%rax
bd: jne 0x00000000000000c6 10e: jne 0x0000000000000117
bf: mov $0xffff,%eax 110: mov $0xffff,%eax
c4: jmp 0x00000000000000c8 115: jmp 0x000000000000011c
c6: xor %eax,%eax 117: mov $0x0,%eax
c8: mov -0x8(%rbp),%rbx 11c: mov -0x228(%rbp),%rbx // epilogue
cc: leaveq 123: mov -0x220(%rbp),%r13
cd: retq 12a: mov -0x218(%rbp),%r14
131: mov -0x210(%rbp),%r15
138: leaveq
139: retq
On fully cached SKBs both JITed functions take 12 nsec to execute.
BPF interpreter executes the program in 30 nsec.
The difference in generated assembler is due to the following:
Old BPF imlements LDX_MSH instruction via sk_load_byte_msh() helper function
inside bpf_jit.S.
New JIT removes the helper and does it explicitly, so ldx_msh cost
is the same for both JITs, but generated code looks longer.
New JIT has 4 registers to save, so prologue/epilogue are larger,
but the cost is within noise on x64.
Old JIT checks whether first insn clears A and if not emits 'xor %eax,%eax'.
New JIT clears %rax unconditionally.
2. old BPF JIT doesn't support ANC_NLATTR, ANC_PAY_OFFSET, ANC_RANDOM
extensions. New JIT supports all BPF extensions.
Performance of such filters improves 2-4 times depending on a filter.
The longer the filter the higher performance gain.
Synthetic benchmarks with many ancillary loads see 20x speedup
which seems to be the maximum gain from JIT
Notes:
. net.core.bpf_jit_enable=2 + tools/net/bpf_jit_disasm is still functional
and can be used to see generated assembler
. there are two jit_compile() functions and code flow for classic filters is:
sk_attach_filter() - load classic BPF
bpf_jit_compile() - try to JIT from classic BPF
sk_convert_filter() - convert classic to internal
bpf_int_jit_compile() - JIT from internal BPF
seccomp and tracing filters will just call bpf_int_jit_compile()
Signed-off-by: Alexei Starovoitov <ast@plumgrid.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
2014-05-14 02:50:46 +00:00
|
|
|
case BPF_LDX | BPF_MEM | BPF_DW:
|
2019-10-16 03:25:03 +00:00
|
|
|
case BPF_LDX | BPF_PROBE_MEM | BPF_DW:
|
2023-07-28 01:11:56 +00:00
|
|
|
/* LDXS: dst_reg = *(s8*)(src_reg + off) */
|
|
|
|
case BPF_LDX | BPF_MEMSX | BPF_B:
|
|
|
|
case BPF_LDX | BPF_MEMSX | BPF_H:
|
|
|
|
case BPF_LDX | BPF_MEMSX | BPF_W:
|
|
|
|
case BPF_LDX | BPF_PROBE_MEMSX | BPF_B:
|
|
|
|
case BPF_LDX | BPF_PROBE_MEMSX | BPF_H:
|
|
|
|
case BPF_LDX | BPF_PROBE_MEMSX | BPF_W:
|
bpf, x86: Improve PROBE_MEM runtime load check
This patch rewrites the runtime PROBE_MEM check insns emitted by the BPF
JIT in order to ensure load safety. The changes in the patch fix two
issues with the previous logic and more generally improve size of
emitted code. Paragraphs between this one and "FIX 1" below explain the
purpose of the runtime check and examine the current implementation.
When a load is marked PROBE_MEM - e.g. due to PTR_UNTRUSTED access - the
address being loaded from is not necessarily valid. The BPF jit sets up
exception handlers for each such load which catch page faults and 0 out
the destination register.
Arbitrary register-relative loads can escape this exception handling
mechanism. Specifically, a load like dst_reg = *(src_reg + off) will not
trigger BPF exception handling if (src_reg + off) is outside of kernel
address space, resulting in an uncaught page fault. A concrete example
of such behavior is a program like:
struct result {
char space[40];
long a;
};
/* if err, returns ERR_PTR(-EINVAL) */
struct result *ptr = get_ptr_maybe_err();
long x = ptr->a;
If get_ptr_maybe_err returns ERR_PTR(-EINVAL) and the result isn't
checked for err, 'result' will be (u64)-EINVAL, a number close to
U64_MAX. The ptr->a load will be > U64_MAX and will wrap over to a small
positive u64, which will be in userspace and thus not covered by BPF
exception handling mechanism.
In order to prevent such loads from occurring, the BPF jit emits some
instructions which do runtime checking of (src_reg + off) and skip the
actual load if it's out of range. As an example, here are instructions
emitted for a %rdi = *(%rdi + 0x10) PROBE_MEM load:
72: movabs $0x800000000010,%r11 --|
7c: cmp %r11,%rdi |- 72 - 7f: Check 1
7f: jb 0x000000000000008d --|
81: mov %rdi,%r11 -----|
84: add $0x0000000000000010,%r11 |- 81-8b: Check 2
8b: jnc 0x0000000000000091 -----|
8d: xor %edi,%edi ---- 0 out dest
8f: jmp 0x0000000000000095
91: mov 0x10(%rdi),%rdi ---- Actual load
95:
The JIT considers kernel address space to start at MAX_TASK_SIZE +
PAGE_SIZE. Determining whether a load will be outside of kernel address
space should be a simple check:
(src_reg + off) >= MAX_TASK_SIZE + PAGE_SIZE
But because there is only one spare register when the checking logic is
emitted, this logic is split into two checks:
Check 1: src_reg >= (MAX_TASK_SIZE + PAGE_SIZE - off)
Check 2: src_reg + off doesn't wrap over U64_MAX and result in small pos u64
Emitted insns implementing Checks 1 and 2 are annotated in the above
example. Check 1 can be done with a single spare register since the
source reg by definition is the left-hand-side of the inequality.
Since adding 'off' to both sides of Check 1's inequality results in the
original inequality we want, it's equivalent to testing that inequality.
Except in the case where src_reg + off wraps past U64_MAX, which is why
Check 2 needs to actually add src_reg + off if Check 1 passes - again
using the single spare reg.
FIX 1: The Check 1 inequality listed above is not what current code is
doing. Current code is a bit more pessimistic, instead checking:
src_reg >= (MAX_TASK_SIZE + PAGE_SIZE + abs(off))
The 0x800000000010 in above example is from this current check. If Check
1 was corrected to use the correct right-hand-side, the value would be
0x7ffffffffff0. This patch changes the checking logic more broadly (FIX
2 below will elaborate), fixing this issue as a side-effect of the
rewrite. Regardless, it's important to understand why Check 1 should've
been doing MAX_TASK_SIZE + PAGE_SIZE - off before proceeding.
FIX 2: Current code relies on a 'jnc' to determine whether src_reg + off
addition wrapped over. For negative offsets this logic is incorrect.
Consider Check 2 insns emitted when off = -0x10:
81: mov %rdi,%r11
84: add 0xfffffffffffffff0,%r11
8b: jnc 0x0000000000000091
2's complement representation of -0x10 is a large positive u64. Any
value of src_reg that passes Check 1 will result in carry flag being set
after (src_reg + off) addition. So a load with any negative offset will
always fail Check 2 at runtime and never do the actual load. This patch
fixes the negative offset issue by rewriting both checks in order to not
rely on carry flag.
The rewrite takes advantage of the fact that, while we only have one
scratch reg to hold arbitrary values, we know the offset at JIT time.
This we can use src_reg as a temporary scratch reg to hold src_reg +
offset since we can return it to its original value by later subtracting
offset. As a result we can directly check the original inequality we
care about:
(src_reg + off) >= MAX_TASK_SIZE + PAGE_SIZE
For a load like %rdi = *(%rsi + -0x10), this results in emitted code:
43: movabs $0x800000000000,%r11
4d: add $0xfffffffffffffff0,%rsi --- src_reg += off
54: cmp %r11,%rsi --- Check original inequality
57: jae 0x000000000000005d
59: xor %edi,%edi
5b: jmp 0x0000000000000061
5d: mov 0x0(%rdi),%rsi --- Actual Load
61: sub $0xfffffffffffffff0,%rsi --- src_reg -= off
Note that the actual load is always done with offset 0, since previous
insns have already done src_reg += off. Regardless of whether the new
check succeeds or fails, insn 61 is always executed, returning src_reg
to its original value.
Because the goal of these checks is to ensure that loaded-from address
will be protected by BPF exception handler, the new check can safely
ignore any wrapover from insn 4d. If such wrapped-over address passes
insn 54 + 57's cmp-and-jmp it will have such protection so the load can
proceed.
IMPROVEMENTS: The above improved logic is 8 insns vs original logic's 9,
and has 1 fewer jmp. The number of checking insns can be further
improved in common scenarios:
If src_reg == dst_reg, the actual load insn will clobber src_reg, so
there's no original src_reg state for the sub insn immediately following
the load to restore, so it can be omitted. In fact, it must be omitted
since it would incorrectly subtract from the result of the load if it
wasn't. So for src_reg == dst_reg, JIT emits these insns:
3c: movabs $0x800000000000,%r11
46: add $0xfffffffffffffff0,%rdi
4d: cmp %r11,%rdi
50: jae 0x0000000000000056
52: xor %edi,%edi
54: jmp 0x000000000000005a
56: mov 0x0(%rdi),%rdi
5a:
The only difference from larger example being the omitted sub, which
would've been insn 5a in this example.
If offset == 0, we can similarly omit the sub as in previous case, since
there's nothing added to subtract. For the same reason we can omit the
addition as well, resulting in JIT emitting these insns:
46: movabs $0x800000000000,%r11
4d: cmp %r11,%rdi
50: jae 0x0000000000000056
52: xor %edi,%edi
54: jmp 0x000000000000005a
56: mov 0x0(%rdi),%rdi
5a:
Although the above example also has src_reg == dst_reg, the same
offset == 0 optimization is valid to apply if src_reg != dst_reg.
To summarize the improvements in emitted insn count for the
check-and-load:
BEFORE: 8 check insns, 3 jmps
AFTER (general case): 7 check insns, 2 jmps (12.5% fewer insn, 33% jmp)
AFTER (src == dst): 6 check insns, 2 jmps (25% fewer insn)
AFTER (offset == 0): 5 check insns, 2 jmps (37.5% fewer insn)
(Above counts don't include the 1 load insn, just checking around it)
Based on BPF bytecode + JITted x86 insn I saw while experimenting with
these improvements, I expect the src_reg == dst_reg case to occur most
often, followed by offset == 0, then the general case.
Signed-off-by: Dave Marchevsky <davemarchevsky@fb.com>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Acked-by: Yonghong Song <yhs@fb.com>
Link: https://lore.kernel.org/bpf/20221216214319.3408356-1-davemarchevsky@fb.com
2022-12-16 21:43:18 +00:00
|
|
|
insn_off = insn->off;
|
|
|
|
|
2023-07-28 01:11:56 +00:00
|
|
|
if (BPF_MODE(insn->code) == BPF_PROBE_MEM ||
|
|
|
|
BPF_MODE(insn->code) == BPF_PROBE_MEMSX) {
|
bpf, x86: Improve PROBE_MEM runtime load check
This patch rewrites the runtime PROBE_MEM check insns emitted by the BPF
JIT in order to ensure load safety. The changes in the patch fix two
issues with the previous logic and more generally improve size of
emitted code. Paragraphs between this one and "FIX 1" below explain the
purpose of the runtime check and examine the current implementation.
When a load is marked PROBE_MEM - e.g. due to PTR_UNTRUSTED access - the
address being loaded from is not necessarily valid. The BPF jit sets up
exception handlers for each such load which catch page faults and 0 out
the destination register.
Arbitrary register-relative loads can escape this exception handling
mechanism. Specifically, a load like dst_reg = *(src_reg + off) will not
trigger BPF exception handling if (src_reg + off) is outside of kernel
address space, resulting in an uncaught page fault. A concrete example
of such behavior is a program like:
struct result {
char space[40];
long a;
};
/* if err, returns ERR_PTR(-EINVAL) */
struct result *ptr = get_ptr_maybe_err();
long x = ptr->a;
If get_ptr_maybe_err returns ERR_PTR(-EINVAL) and the result isn't
checked for err, 'result' will be (u64)-EINVAL, a number close to
U64_MAX. The ptr->a load will be > U64_MAX and will wrap over to a small
positive u64, which will be in userspace and thus not covered by BPF
exception handling mechanism.
In order to prevent such loads from occurring, the BPF jit emits some
instructions which do runtime checking of (src_reg + off) and skip the
actual load if it's out of range. As an example, here are instructions
emitted for a %rdi = *(%rdi + 0x10) PROBE_MEM load:
72: movabs $0x800000000010,%r11 --|
7c: cmp %r11,%rdi |- 72 - 7f: Check 1
7f: jb 0x000000000000008d --|
81: mov %rdi,%r11 -----|
84: add $0x0000000000000010,%r11 |- 81-8b: Check 2
8b: jnc 0x0000000000000091 -----|
8d: xor %edi,%edi ---- 0 out dest
8f: jmp 0x0000000000000095
91: mov 0x10(%rdi),%rdi ---- Actual load
95:
The JIT considers kernel address space to start at MAX_TASK_SIZE +
PAGE_SIZE. Determining whether a load will be outside of kernel address
space should be a simple check:
(src_reg + off) >= MAX_TASK_SIZE + PAGE_SIZE
But because there is only one spare register when the checking logic is
emitted, this logic is split into two checks:
Check 1: src_reg >= (MAX_TASK_SIZE + PAGE_SIZE - off)
Check 2: src_reg + off doesn't wrap over U64_MAX and result in small pos u64
Emitted insns implementing Checks 1 and 2 are annotated in the above
example. Check 1 can be done with a single spare register since the
source reg by definition is the left-hand-side of the inequality.
Since adding 'off' to both sides of Check 1's inequality results in the
original inequality we want, it's equivalent to testing that inequality.
Except in the case where src_reg + off wraps past U64_MAX, which is why
Check 2 needs to actually add src_reg + off if Check 1 passes - again
using the single spare reg.
FIX 1: The Check 1 inequality listed above is not what current code is
doing. Current code is a bit more pessimistic, instead checking:
src_reg >= (MAX_TASK_SIZE + PAGE_SIZE + abs(off))
The 0x800000000010 in above example is from this current check. If Check
1 was corrected to use the correct right-hand-side, the value would be
0x7ffffffffff0. This patch changes the checking logic more broadly (FIX
2 below will elaborate), fixing this issue as a side-effect of the
rewrite. Regardless, it's important to understand why Check 1 should've
been doing MAX_TASK_SIZE + PAGE_SIZE - off before proceeding.
FIX 2: Current code relies on a 'jnc' to determine whether src_reg + off
addition wrapped over. For negative offsets this logic is incorrect.
Consider Check 2 insns emitted when off = -0x10:
81: mov %rdi,%r11
84: add 0xfffffffffffffff0,%r11
8b: jnc 0x0000000000000091
2's complement representation of -0x10 is a large positive u64. Any
value of src_reg that passes Check 1 will result in carry flag being set
after (src_reg + off) addition. So a load with any negative offset will
always fail Check 2 at runtime and never do the actual load. This patch
fixes the negative offset issue by rewriting both checks in order to not
rely on carry flag.
The rewrite takes advantage of the fact that, while we only have one
scratch reg to hold arbitrary values, we know the offset at JIT time.
This we can use src_reg as a temporary scratch reg to hold src_reg +
offset since we can return it to its original value by later subtracting
offset. As a result we can directly check the original inequality we
care about:
(src_reg + off) >= MAX_TASK_SIZE + PAGE_SIZE
For a load like %rdi = *(%rsi + -0x10), this results in emitted code:
43: movabs $0x800000000000,%r11
4d: add $0xfffffffffffffff0,%rsi --- src_reg += off
54: cmp %r11,%rsi --- Check original inequality
57: jae 0x000000000000005d
59: xor %edi,%edi
5b: jmp 0x0000000000000061
5d: mov 0x0(%rdi),%rsi --- Actual Load
61: sub $0xfffffffffffffff0,%rsi --- src_reg -= off
Note that the actual load is always done with offset 0, since previous
insns have already done src_reg += off. Regardless of whether the new
check succeeds or fails, insn 61 is always executed, returning src_reg
to its original value.
Because the goal of these checks is to ensure that loaded-from address
will be protected by BPF exception handler, the new check can safely
ignore any wrapover from insn 4d. If such wrapped-over address passes
insn 54 + 57's cmp-and-jmp it will have such protection so the load can
proceed.
IMPROVEMENTS: The above improved logic is 8 insns vs original logic's 9,
and has 1 fewer jmp. The number of checking insns can be further
improved in common scenarios:
If src_reg == dst_reg, the actual load insn will clobber src_reg, so
there's no original src_reg state for the sub insn immediately following
the load to restore, so it can be omitted. In fact, it must be omitted
since it would incorrectly subtract from the result of the load if it
wasn't. So for src_reg == dst_reg, JIT emits these insns:
3c: movabs $0x800000000000,%r11
46: add $0xfffffffffffffff0,%rdi
4d: cmp %r11,%rdi
50: jae 0x0000000000000056
52: xor %edi,%edi
54: jmp 0x000000000000005a
56: mov 0x0(%rdi),%rdi
5a:
The only difference from larger example being the omitted sub, which
would've been insn 5a in this example.
If offset == 0, we can similarly omit the sub as in previous case, since
there's nothing added to subtract. For the same reason we can omit the
addition as well, resulting in JIT emitting these insns:
46: movabs $0x800000000000,%r11
4d: cmp %r11,%rdi
50: jae 0x0000000000000056
52: xor %edi,%edi
54: jmp 0x000000000000005a
56: mov 0x0(%rdi),%rdi
5a:
Although the above example also has src_reg == dst_reg, the same
offset == 0 optimization is valid to apply if src_reg != dst_reg.
To summarize the improvements in emitted insn count for the
check-and-load:
BEFORE: 8 check insns, 3 jmps
AFTER (general case): 7 check insns, 2 jmps (12.5% fewer insn, 33% jmp)
AFTER (src == dst): 6 check insns, 2 jmps (25% fewer insn)
AFTER (offset == 0): 5 check insns, 2 jmps (37.5% fewer insn)
(Above counts don't include the 1 load insn, just checking around it)
Based on BPF bytecode + JITted x86 insn I saw while experimenting with
these improvements, I expect the src_reg == dst_reg case to occur most
often, followed by offset == 0, then the general case.
Signed-off-by: Dave Marchevsky <davemarchevsky@fb.com>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Acked-by: Yonghong Song <yhs@fb.com>
Link: https://lore.kernel.org/bpf/20221216214319.3408356-1-davemarchevsky@fb.com
2022-12-16 21:43:18 +00:00
|
|
|
/* Conservatively check that src_reg + insn->off is a kernel address:
|
|
|
|
* src_reg + insn->off >= TASK_SIZE_MAX + PAGE_SIZE
|
|
|
|
* src_reg is used as scratch for src_reg += insn->off and restored
|
|
|
|
* after emit_ldx if necessary
|
2021-12-15 03:25:13 +00:00
|
|
|
*/
|
|
|
|
|
bpf, x86: Improve PROBE_MEM runtime load check
This patch rewrites the runtime PROBE_MEM check insns emitted by the BPF
JIT in order to ensure load safety. The changes in the patch fix two
issues with the previous logic and more generally improve size of
emitted code. Paragraphs between this one and "FIX 1" below explain the
purpose of the runtime check and examine the current implementation.
When a load is marked PROBE_MEM - e.g. due to PTR_UNTRUSTED access - the
address being loaded from is not necessarily valid. The BPF jit sets up
exception handlers for each such load which catch page faults and 0 out
the destination register.
Arbitrary register-relative loads can escape this exception handling
mechanism. Specifically, a load like dst_reg = *(src_reg + off) will not
trigger BPF exception handling if (src_reg + off) is outside of kernel
address space, resulting in an uncaught page fault. A concrete example
of such behavior is a program like:
struct result {
char space[40];
long a;
};
/* if err, returns ERR_PTR(-EINVAL) */
struct result *ptr = get_ptr_maybe_err();
long x = ptr->a;
If get_ptr_maybe_err returns ERR_PTR(-EINVAL) and the result isn't
checked for err, 'result' will be (u64)-EINVAL, a number close to
U64_MAX. The ptr->a load will be > U64_MAX and will wrap over to a small
positive u64, which will be in userspace and thus not covered by BPF
exception handling mechanism.
In order to prevent such loads from occurring, the BPF jit emits some
instructions which do runtime checking of (src_reg + off) and skip the
actual load if it's out of range. As an example, here are instructions
emitted for a %rdi = *(%rdi + 0x10) PROBE_MEM load:
72: movabs $0x800000000010,%r11 --|
7c: cmp %r11,%rdi |- 72 - 7f: Check 1
7f: jb 0x000000000000008d --|
81: mov %rdi,%r11 -----|
84: add $0x0000000000000010,%r11 |- 81-8b: Check 2
8b: jnc 0x0000000000000091 -----|
8d: xor %edi,%edi ---- 0 out dest
8f: jmp 0x0000000000000095
91: mov 0x10(%rdi),%rdi ---- Actual load
95:
The JIT considers kernel address space to start at MAX_TASK_SIZE +
PAGE_SIZE. Determining whether a load will be outside of kernel address
space should be a simple check:
(src_reg + off) >= MAX_TASK_SIZE + PAGE_SIZE
But because there is only one spare register when the checking logic is
emitted, this logic is split into two checks:
Check 1: src_reg >= (MAX_TASK_SIZE + PAGE_SIZE - off)
Check 2: src_reg + off doesn't wrap over U64_MAX and result in small pos u64
Emitted insns implementing Checks 1 and 2 are annotated in the above
example. Check 1 can be done with a single spare register since the
source reg by definition is the left-hand-side of the inequality.
Since adding 'off' to both sides of Check 1's inequality results in the
original inequality we want, it's equivalent to testing that inequality.
Except in the case where src_reg + off wraps past U64_MAX, which is why
Check 2 needs to actually add src_reg + off if Check 1 passes - again
using the single spare reg.
FIX 1: The Check 1 inequality listed above is not what current code is
doing. Current code is a bit more pessimistic, instead checking:
src_reg >= (MAX_TASK_SIZE + PAGE_SIZE + abs(off))
The 0x800000000010 in above example is from this current check. If Check
1 was corrected to use the correct right-hand-side, the value would be
0x7ffffffffff0. This patch changes the checking logic more broadly (FIX
2 below will elaborate), fixing this issue as a side-effect of the
rewrite. Regardless, it's important to understand why Check 1 should've
been doing MAX_TASK_SIZE + PAGE_SIZE - off before proceeding.
FIX 2: Current code relies on a 'jnc' to determine whether src_reg + off
addition wrapped over. For negative offsets this logic is incorrect.
Consider Check 2 insns emitted when off = -0x10:
81: mov %rdi,%r11
84: add 0xfffffffffffffff0,%r11
8b: jnc 0x0000000000000091
2's complement representation of -0x10 is a large positive u64. Any
value of src_reg that passes Check 1 will result in carry flag being set
after (src_reg + off) addition. So a load with any negative offset will
always fail Check 2 at runtime and never do the actual load. This patch
fixes the negative offset issue by rewriting both checks in order to not
rely on carry flag.
The rewrite takes advantage of the fact that, while we only have one
scratch reg to hold arbitrary values, we know the offset at JIT time.
This we can use src_reg as a temporary scratch reg to hold src_reg +
offset since we can return it to its original value by later subtracting
offset. As a result we can directly check the original inequality we
care about:
(src_reg + off) >= MAX_TASK_SIZE + PAGE_SIZE
For a load like %rdi = *(%rsi + -0x10), this results in emitted code:
43: movabs $0x800000000000,%r11
4d: add $0xfffffffffffffff0,%rsi --- src_reg += off
54: cmp %r11,%rsi --- Check original inequality
57: jae 0x000000000000005d
59: xor %edi,%edi
5b: jmp 0x0000000000000061
5d: mov 0x0(%rdi),%rsi --- Actual Load
61: sub $0xfffffffffffffff0,%rsi --- src_reg -= off
Note that the actual load is always done with offset 0, since previous
insns have already done src_reg += off. Regardless of whether the new
check succeeds or fails, insn 61 is always executed, returning src_reg
to its original value.
Because the goal of these checks is to ensure that loaded-from address
will be protected by BPF exception handler, the new check can safely
ignore any wrapover from insn 4d. If such wrapped-over address passes
insn 54 + 57's cmp-and-jmp it will have such protection so the load can
proceed.
IMPROVEMENTS: The above improved logic is 8 insns vs original logic's 9,
and has 1 fewer jmp. The number of checking insns can be further
improved in common scenarios:
If src_reg == dst_reg, the actual load insn will clobber src_reg, so
there's no original src_reg state for the sub insn immediately following
the load to restore, so it can be omitted. In fact, it must be omitted
since it would incorrectly subtract from the result of the load if it
wasn't. So for src_reg == dst_reg, JIT emits these insns:
3c: movabs $0x800000000000,%r11
46: add $0xfffffffffffffff0,%rdi
4d: cmp %r11,%rdi
50: jae 0x0000000000000056
52: xor %edi,%edi
54: jmp 0x000000000000005a
56: mov 0x0(%rdi),%rdi
5a:
The only difference from larger example being the omitted sub, which
would've been insn 5a in this example.
If offset == 0, we can similarly omit the sub as in previous case, since
there's nothing added to subtract. For the same reason we can omit the
addition as well, resulting in JIT emitting these insns:
46: movabs $0x800000000000,%r11
4d: cmp %r11,%rdi
50: jae 0x0000000000000056
52: xor %edi,%edi
54: jmp 0x000000000000005a
56: mov 0x0(%rdi),%rdi
5a:
Although the above example also has src_reg == dst_reg, the same
offset == 0 optimization is valid to apply if src_reg != dst_reg.
To summarize the improvements in emitted insn count for the
check-and-load:
BEFORE: 8 check insns, 3 jmps
AFTER (general case): 7 check insns, 2 jmps (12.5% fewer insn, 33% jmp)
AFTER (src == dst): 6 check insns, 2 jmps (25% fewer insn)
AFTER (offset == 0): 5 check insns, 2 jmps (37.5% fewer insn)
(Above counts don't include the 1 load insn, just checking around it)
Based on BPF bytecode + JITted x86 insn I saw while experimenting with
these improvements, I expect the src_reg == dst_reg case to occur most
often, followed by offset == 0, then the general case.
Signed-off-by: Dave Marchevsky <davemarchevsky@fb.com>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Acked-by: Yonghong Song <yhs@fb.com>
Link: https://lore.kernel.org/bpf/20221216214319.3408356-1-davemarchevsky@fb.com
2022-12-16 21:43:18 +00:00
|
|
|
u64 limit = TASK_SIZE_MAX + PAGE_SIZE;
|
|
|
|
u8 *end_of_jmp;
|
|
|
|
|
|
|
|
/* At end of these emitted checks, insn->off will have been added
|
|
|
|
* to src_reg, so no need to do relative load with insn->off offset
|
2021-12-15 03:25:13 +00:00
|
|
|
*/
|
bpf, x86: Improve PROBE_MEM runtime load check
This patch rewrites the runtime PROBE_MEM check insns emitted by the BPF
JIT in order to ensure load safety. The changes in the patch fix two
issues with the previous logic and more generally improve size of
emitted code. Paragraphs between this one and "FIX 1" below explain the
purpose of the runtime check and examine the current implementation.
When a load is marked PROBE_MEM - e.g. due to PTR_UNTRUSTED access - the
address being loaded from is not necessarily valid. The BPF jit sets up
exception handlers for each such load which catch page faults and 0 out
the destination register.
Arbitrary register-relative loads can escape this exception handling
mechanism. Specifically, a load like dst_reg = *(src_reg + off) will not
trigger BPF exception handling if (src_reg + off) is outside of kernel
address space, resulting in an uncaught page fault. A concrete example
of such behavior is a program like:
struct result {
char space[40];
long a;
};
/* if err, returns ERR_PTR(-EINVAL) */
struct result *ptr = get_ptr_maybe_err();
long x = ptr->a;
If get_ptr_maybe_err returns ERR_PTR(-EINVAL) and the result isn't
checked for err, 'result' will be (u64)-EINVAL, a number close to
U64_MAX. The ptr->a load will be > U64_MAX and will wrap over to a small
positive u64, which will be in userspace and thus not covered by BPF
exception handling mechanism.
In order to prevent such loads from occurring, the BPF jit emits some
instructions which do runtime checking of (src_reg + off) and skip the
actual load if it's out of range. As an example, here are instructions
emitted for a %rdi = *(%rdi + 0x10) PROBE_MEM load:
72: movabs $0x800000000010,%r11 --|
7c: cmp %r11,%rdi |- 72 - 7f: Check 1
7f: jb 0x000000000000008d --|
81: mov %rdi,%r11 -----|
84: add $0x0000000000000010,%r11 |- 81-8b: Check 2
8b: jnc 0x0000000000000091 -----|
8d: xor %edi,%edi ---- 0 out dest
8f: jmp 0x0000000000000095
91: mov 0x10(%rdi),%rdi ---- Actual load
95:
The JIT considers kernel address space to start at MAX_TASK_SIZE +
PAGE_SIZE. Determining whether a load will be outside of kernel address
space should be a simple check:
(src_reg + off) >= MAX_TASK_SIZE + PAGE_SIZE
But because there is only one spare register when the checking logic is
emitted, this logic is split into two checks:
Check 1: src_reg >= (MAX_TASK_SIZE + PAGE_SIZE - off)
Check 2: src_reg + off doesn't wrap over U64_MAX and result in small pos u64
Emitted insns implementing Checks 1 and 2 are annotated in the above
example. Check 1 can be done with a single spare register since the
source reg by definition is the left-hand-side of the inequality.
Since adding 'off' to both sides of Check 1's inequality results in the
original inequality we want, it's equivalent to testing that inequality.
Except in the case where src_reg + off wraps past U64_MAX, which is why
Check 2 needs to actually add src_reg + off if Check 1 passes - again
using the single spare reg.
FIX 1: The Check 1 inequality listed above is not what current code is
doing. Current code is a bit more pessimistic, instead checking:
src_reg >= (MAX_TASK_SIZE + PAGE_SIZE + abs(off))
The 0x800000000010 in above example is from this current check. If Check
1 was corrected to use the correct right-hand-side, the value would be
0x7ffffffffff0. This patch changes the checking logic more broadly (FIX
2 below will elaborate), fixing this issue as a side-effect of the
rewrite. Regardless, it's important to understand why Check 1 should've
been doing MAX_TASK_SIZE + PAGE_SIZE - off before proceeding.
FIX 2: Current code relies on a 'jnc' to determine whether src_reg + off
addition wrapped over. For negative offsets this logic is incorrect.
Consider Check 2 insns emitted when off = -0x10:
81: mov %rdi,%r11
84: add 0xfffffffffffffff0,%r11
8b: jnc 0x0000000000000091
2's complement representation of -0x10 is a large positive u64. Any
value of src_reg that passes Check 1 will result in carry flag being set
after (src_reg + off) addition. So a load with any negative offset will
always fail Check 2 at runtime and never do the actual load. This patch
fixes the negative offset issue by rewriting both checks in order to not
rely on carry flag.
The rewrite takes advantage of the fact that, while we only have one
scratch reg to hold arbitrary values, we know the offset at JIT time.
This we can use src_reg as a temporary scratch reg to hold src_reg +
offset since we can return it to its original value by later subtracting
offset. As a result we can directly check the original inequality we
care about:
(src_reg + off) >= MAX_TASK_SIZE + PAGE_SIZE
For a load like %rdi = *(%rsi + -0x10), this results in emitted code:
43: movabs $0x800000000000,%r11
4d: add $0xfffffffffffffff0,%rsi --- src_reg += off
54: cmp %r11,%rsi --- Check original inequality
57: jae 0x000000000000005d
59: xor %edi,%edi
5b: jmp 0x0000000000000061
5d: mov 0x0(%rdi),%rsi --- Actual Load
61: sub $0xfffffffffffffff0,%rsi --- src_reg -= off
Note that the actual load is always done with offset 0, since previous
insns have already done src_reg += off. Regardless of whether the new
check succeeds or fails, insn 61 is always executed, returning src_reg
to its original value.
Because the goal of these checks is to ensure that loaded-from address
will be protected by BPF exception handler, the new check can safely
ignore any wrapover from insn 4d. If such wrapped-over address passes
insn 54 + 57's cmp-and-jmp it will have such protection so the load can
proceed.
IMPROVEMENTS: The above improved logic is 8 insns vs original logic's 9,
and has 1 fewer jmp. The number of checking insns can be further
improved in common scenarios:
If src_reg == dst_reg, the actual load insn will clobber src_reg, so
there's no original src_reg state for the sub insn immediately following
the load to restore, so it can be omitted. In fact, it must be omitted
since it would incorrectly subtract from the result of the load if it
wasn't. So for src_reg == dst_reg, JIT emits these insns:
3c: movabs $0x800000000000,%r11
46: add $0xfffffffffffffff0,%rdi
4d: cmp %r11,%rdi
50: jae 0x0000000000000056
52: xor %edi,%edi
54: jmp 0x000000000000005a
56: mov 0x0(%rdi),%rdi
5a:
The only difference from larger example being the omitted sub, which
would've been insn 5a in this example.
If offset == 0, we can similarly omit the sub as in previous case, since
there's nothing added to subtract. For the same reason we can omit the
addition as well, resulting in JIT emitting these insns:
46: movabs $0x800000000000,%r11
4d: cmp %r11,%rdi
50: jae 0x0000000000000056
52: xor %edi,%edi
54: jmp 0x000000000000005a
56: mov 0x0(%rdi),%rdi
5a:
Although the above example also has src_reg == dst_reg, the same
offset == 0 optimization is valid to apply if src_reg != dst_reg.
To summarize the improvements in emitted insn count for the
check-and-load:
BEFORE: 8 check insns, 3 jmps
AFTER (general case): 7 check insns, 2 jmps (12.5% fewer insn, 33% jmp)
AFTER (src == dst): 6 check insns, 2 jmps (25% fewer insn)
AFTER (offset == 0): 5 check insns, 2 jmps (37.5% fewer insn)
(Above counts don't include the 1 load insn, just checking around it)
Based on BPF bytecode + JITted x86 insn I saw while experimenting with
these improvements, I expect the src_reg == dst_reg case to occur most
often, followed by offset == 0, then the general case.
Signed-off-by: Dave Marchevsky <davemarchevsky@fb.com>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Acked-by: Yonghong Song <yhs@fb.com>
Link: https://lore.kernel.org/bpf/20221216214319.3408356-1-davemarchevsky@fb.com
2022-12-16 21:43:18 +00:00
|
|
|
insn_off = 0;
|
2021-12-15 03:25:13 +00:00
|
|
|
|
|
|
|
/* movabsq r11, limit */
|
|
|
|
EMIT2(add_1mod(0x48, AUX_REG), add_1reg(0xB8, AUX_REG));
|
|
|
|
EMIT((u32)limit, 4);
|
|
|
|
EMIT(limit >> 32, 4);
|
bpf, x86: Improve PROBE_MEM runtime load check
This patch rewrites the runtime PROBE_MEM check insns emitted by the BPF
JIT in order to ensure load safety. The changes in the patch fix two
issues with the previous logic and more generally improve size of
emitted code. Paragraphs between this one and "FIX 1" below explain the
purpose of the runtime check and examine the current implementation.
When a load is marked PROBE_MEM - e.g. due to PTR_UNTRUSTED access - the
address being loaded from is not necessarily valid. The BPF jit sets up
exception handlers for each such load which catch page faults and 0 out
the destination register.
Arbitrary register-relative loads can escape this exception handling
mechanism. Specifically, a load like dst_reg = *(src_reg + off) will not
trigger BPF exception handling if (src_reg + off) is outside of kernel
address space, resulting in an uncaught page fault. A concrete example
of such behavior is a program like:
struct result {
char space[40];
long a;
};
/* if err, returns ERR_PTR(-EINVAL) */
struct result *ptr = get_ptr_maybe_err();
long x = ptr->a;
If get_ptr_maybe_err returns ERR_PTR(-EINVAL) and the result isn't
checked for err, 'result' will be (u64)-EINVAL, a number close to
U64_MAX. The ptr->a load will be > U64_MAX and will wrap over to a small
positive u64, which will be in userspace and thus not covered by BPF
exception handling mechanism.
In order to prevent such loads from occurring, the BPF jit emits some
instructions which do runtime checking of (src_reg + off) and skip the
actual load if it's out of range. As an example, here are instructions
emitted for a %rdi = *(%rdi + 0x10) PROBE_MEM load:
72: movabs $0x800000000010,%r11 --|
7c: cmp %r11,%rdi |- 72 - 7f: Check 1
7f: jb 0x000000000000008d --|
81: mov %rdi,%r11 -----|
84: add $0x0000000000000010,%r11 |- 81-8b: Check 2
8b: jnc 0x0000000000000091 -----|
8d: xor %edi,%edi ---- 0 out dest
8f: jmp 0x0000000000000095
91: mov 0x10(%rdi),%rdi ---- Actual load
95:
The JIT considers kernel address space to start at MAX_TASK_SIZE +
PAGE_SIZE. Determining whether a load will be outside of kernel address
space should be a simple check:
(src_reg + off) >= MAX_TASK_SIZE + PAGE_SIZE
But because there is only one spare register when the checking logic is
emitted, this logic is split into two checks:
Check 1: src_reg >= (MAX_TASK_SIZE + PAGE_SIZE - off)
Check 2: src_reg + off doesn't wrap over U64_MAX and result in small pos u64
Emitted insns implementing Checks 1 and 2 are annotated in the above
example. Check 1 can be done with a single spare register since the
source reg by definition is the left-hand-side of the inequality.
Since adding 'off' to both sides of Check 1's inequality results in the
original inequality we want, it's equivalent to testing that inequality.
Except in the case where src_reg + off wraps past U64_MAX, which is why
Check 2 needs to actually add src_reg + off if Check 1 passes - again
using the single spare reg.
FIX 1: The Check 1 inequality listed above is not what current code is
doing. Current code is a bit more pessimistic, instead checking:
src_reg >= (MAX_TASK_SIZE + PAGE_SIZE + abs(off))
The 0x800000000010 in above example is from this current check. If Check
1 was corrected to use the correct right-hand-side, the value would be
0x7ffffffffff0. This patch changes the checking logic more broadly (FIX
2 below will elaborate), fixing this issue as a side-effect of the
rewrite. Regardless, it's important to understand why Check 1 should've
been doing MAX_TASK_SIZE + PAGE_SIZE - off before proceeding.
FIX 2: Current code relies on a 'jnc' to determine whether src_reg + off
addition wrapped over. For negative offsets this logic is incorrect.
Consider Check 2 insns emitted when off = -0x10:
81: mov %rdi,%r11
84: add 0xfffffffffffffff0,%r11
8b: jnc 0x0000000000000091
2's complement representation of -0x10 is a large positive u64. Any
value of src_reg that passes Check 1 will result in carry flag being set
after (src_reg + off) addition. So a load with any negative offset will
always fail Check 2 at runtime and never do the actual load. This patch
fixes the negative offset issue by rewriting both checks in order to not
rely on carry flag.
The rewrite takes advantage of the fact that, while we only have one
scratch reg to hold arbitrary values, we know the offset at JIT time.
This we can use src_reg as a temporary scratch reg to hold src_reg +
offset since we can return it to its original value by later subtracting
offset. As a result we can directly check the original inequality we
care about:
(src_reg + off) >= MAX_TASK_SIZE + PAGE_SIZE
For a load like %rdi = *(%rsi + -0x10), this results in emitted code:
43: movabs $0x800000000000,%r11
4d: add $0xfffffffffffffff0,%rsi --- src_reg += off
54: cmp %r11,%rsi --- Check original inequality
57: jae 0x000000000000005d
59: xor %edi,%edi
5b: jmp 0x0000000000000061
5d: mov 0x0(%rdi),%rsi --- Actual Load
61: sub $0xfffffffffffffff0,%rsi --- src_reg -= off
Note that the actual load is always done with offset 0, since previous
insns have already done src_reg += off. Regardless of whether the new
check succeeds or fails, insn 61 is always executed, returning src_reg
to its original value.
Because the goal of these checks is to ensure that loaded-from address
will be protected by BPF exception handler, the new check can safely
ignore any wrapover from insn 4d. If such wrapped-over address passes
insn 54 + 57's cmp-and-jmp it will have such protection so the load can
proceed.
IMPROVEMENTS: The above improved logic is 8 insns vs original logic's 9,
and has 1 fewer jmp. The number of checking insns can be further
improved in common scenarios:
If src_reg == dst_reg, the actual load insn will clobber src_reg, so
there's no original src_reg state for the sub insn immediately following
the load to restore, so it can be omitted. In fact, it must be omitted
since it would incorrectly subtract from the result of the load if it
wasn't. So for src_reg == dst_reg, JIT emits these insns:
3c: movabs $0x800000000000,%r11
46: add $0xfffffffffffffff0,%rdi
4d: cmp %r11,%rdi
50: jae 0x0000000000000056
52: xor %edi,%edi
54: jmp 0x000000000000005a
56: mov 0x0(%rdi),%rdi
5a:
The only difference from larger example being the omitted sub, which
would've been insn 5a in this example.
If offset == 0, we can similarly omit the sub as in previous case, since
there's nothing added to subtract. For the same reason we can omit the
addition as well, resulting in JIT emitting these insns:
46: movabs $0x800000000000,%r11
4d: cmp %r11,%rdi
50: jae 0x0000000000000056
52: xor %edi,%edi
54: jmp 0x000000000000005a
56: mov 0x0(%rdi),%rdi
5a:
Although the above example also has src_reg == dst_reg, the same
offset == 0 optimization is valid to apply if src_reg != dst_reg.
To summarize the improvements in emitted insn count for the
check-and-load:
BEFORE: 8 check insns, 3 jmps
AFTER (general case): 7 check insns, 2 jmps (12.5% fewer insn, 33% jmp)
AFTER (src == dst): 6 check insns, 2 jmps (25% fewer insn)
AFTER (offset == 0): 5 check insns, 2 jmps (37.5% fewer insn)
(Above counts don't include the 1 load insn, just checking around it)
Based on BPF bytecode + JITted x86 insn I saw while experimenting with
these improvements, I expect the src_reg == dst_reg case to occur most
often, followed by offset == 0, then the general case.
Signed-off-by: Dave Marchevsky <davemarchevsky@fb.com>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Acked-by: Yonghong Song <yhs@fb.com>
Link: https://lore.kernel.org/bpf/20221216214319.3408356-1-davemarchevsky@fb.com
2022-12-16 21:43:18 +00:00
|
|
|
|
|
|
|
if (insn->off) {
|
|
|
|
/* add src_reg, insn->off */
|
|
|
|
maybe_emit_1mod(&prog, src_reg, true);
|
|
|
|
EMIT2_off32(0x81, add_1reg(0xC0, src_reg), insn->off);
|
|
|
|
}
|
|
|
|
|
2021-12-15 03:25:13 +00:00
|
|
|
/* cmp src_reg, r11 */
|
|
|
|
maybe_emit_mod(&prog, src_reg, AUX_REG, true);
|
|
|
|
EMIT2(0x39, add_2reg(0xC0, src_reg, AUX_REG));
|
bpf, x86: Improve PROBE_MEM runtime load check
This patch rewrites the runtime PROBE_MEM check insns emitted by the BPF
JIT in order to ensure load safety. The changes in the patch fix two
issues with the previous logic and more generally improve size of
emitted code. Paragraphs between this one and "FIX 1" below explain the
purpose of the runtime check and examine the current implementation.
When a load is marked PROBE_MEM - e.g. due to PTR_UNTRUSTED access - the
address being loaded from is not necessarily valid. The BPF jit sets up
exception handlers for each such load which catch page faults and 0 out
the destination register.
Arbitrary register-relative loads can escape this exception handling
mechanism. Specifically, a load like dst_reg = *(src_reg + off) will not
trigger BPF exception handling if (src_reg + off) is outside of kernel
address space, resulting in an uncaught page fault. A concrete example
of such behavior is a program like:
struct result {
char space[40];
long a;
};
/* if err, returns ERR_PTR(-EINVAL) */
struct result *ptr = get_ptr_maybe_err();
long x = ptr->a;
If get_ptr_maybe_err returns ERR_PTR(-EINVAL) and the result isn't
checked for err, 'result' will be (u64)-EINVAL, a number close to
U64_MAX. The ptr->a load will be > U64_MAX and will wrap over to a small
positive u64, which will be in userspace and thus not covered by BPF
exception handling mechanism.
In order to prevent such loads from occurring, the BPF jit emits some
instructions which do runtime checking of (src_reg + off) and skip the
actual load if it's out of range. As an example, here are instructions
emitted for a %rdi = *(%rdi + 0x10) PROBE_MEM load:
72: movabs $0x800000000010,%r11 --|
7c: cmp %r11,%rdi |- 72 - 7f: Check 1
7f: jb 0x000000000000008d --|
81: mov %rdi,%r11 -----|
84: add $0x0000000000000010,%r11 |- 81-8b: Check 2
8b: jnc 0x0000000000000091 -----|
8d: xor %edi,%edi ---- 0 out dest
8f: jmp 0x0000000000000095
91: mov 0x10(%rdi),%rdi ---- Actual load
95:
The JIT considers kernel address space to start at MAX_TASK_SIZE +
PAGE_SIZE. Determining whether a load will be outside of kernel address
space should be a simple check:
(src_reg + off) >= MAX_TASK_SIZE + PAGE_SIZE
But because there is only one spare register when the checking logic is
emitted, this logic is split into two checks:
Check 1: src_reg >= (MAX_TASK_SIZE + PAGE_SIZE - off)
Check 2: src_reg + off doesn't wrap over U64_MAX and result in small pos u64
Emitted insns implementing Checks 1 and 2 are annotated in the above
example. Check 1 can be done with a single spare register since the
source reg by definition is the left-hand-side of the inequality.
Since adding 'off' to both sides of Check 1's inequality results in the
original inequality we want, it's equivalent to testing that inequality.
Except in the case where src_reg + off wraps past U64_MAX, which is why
Check 2 needs to actually add src_reg + off if Check 1 passes - again
using the single spare reg.
FIX 1: The Check 1 inequality listed above is not what current code is
doing. Current code is a bit more pessimistic, instead checking:
src_reg >= (MAX_TASK_SIZE + PAGE_SIZE + abs(off))
The 0x800000000010 in above example is from this current check. If Check
1 was corrected to use the correct right-hand-side, the value would be
0x7ffffffffff0. This patch changes the checking logic more broadly (FIX
2 below will elaborate), fixing this issue as a side-effect of the
rewrite. Regardless, it's important to understand why Check 1 should've
been doing MAX_TASK_SIZE + PAGE_SIZE - off before proceeding.
FIX 2: Current code relies on a 'jnc' to determine whether src_reg + off
addition wrapped over. For negative offsets this logic is incorrect.
Consider Check 2 insns emitted when off = -0x10:
81: mov %rdi,%r11
84: add 0xfffffffffffffff0,%r11
8b: jnc 0x0000000000000091
2's complement representation of -0x10 is a large positive u64. Any
value of src_reg that passes Check 1 will result in carry flag being set
after (src_reg + off) addition. So a load with any negative offset will
always fail Check 2 at runtime and never do the actual load. This patch
fixes the negative offset issue by rewriting both checks in order to not
rely on carry flag.
The rewrite takes advantage of the fact that, while we only have one
scratch reg to hold arbitrary values, we know the offset at JIT time.
This we can use src_reg as a temporary scratch reg to hold src_reg +
offset since we can return it to its original value by later subtracting
offset. As a result we can directly check the original inequality we
care about:
(src_reg + off) >= MAX_TASK_SIZE + PAGE_SIZE
For a load like %rdi = *(%rsi + -0x10), this results in emitted code:
43: movabs $0x800000000000,%r11
4d: add $0xfffffffffffffff0,%rsi --- src_reg += off
54: cmp %r11,%rsi --- Check original inequality
57: jae 0x000000000000005d
59: xor %edi,%edi
5b: jmp 0x0000000000000061
5d: mov 0x0(%rdi),%rsi --- Actual Load
61: sub $0xfffffffffffffff0,%rsi --- src_reg -= off
Note that the actual load is always done with offset 0, since previous
insns have already done src_reg += off. Regardless of whether the new
check succeeds or fails, insn 61 is always executed, returning src_reg
to its original value.
Because the goal of these checks is to ensure that loaded-from address
will be protected by BPF exception handler, the new check can safely
ignore any wrapover from insn 4d. If such wrapped-over address passes
insn 54 + 57's cmp-and-jmp it will have such protection so the load can
proceed.
IMPROVEMENTS: The above improved logic is 8 insns vs original logic's 9,
and has 1 fewer jmp. The number of checking insns can be further
improved in common scenarios:
If src_reg == dst_reg, the actual load insn will clobber src_reg, so
there's no original src_reg state for the sub insn immediately following
the load to restore, so it can be omitted. In fact, it must be omitted
since it would incorrectly subtract from the result of the load if it
wasn't. So for src_reg == dst_reg, JIT emits these insns:
3c: movabs $0x800000000000,%r11
46: add $0xfffffffffffffff0,%rdi
4d: cmp %r11,%rdi
50: jae 0x0000000000000056
52: xor %edi,%edi
54: jmp 0x000000000000005a
56: mov 0x0(%rdi),%rdi
5a:
The only difference from larger example being the omitted sub, which
would've been insn 5a in this example.
If offset == 0, we can similarly omit the sub as in previous case, since
there's nothing added to subtract. For the same reason we can omit the
addition as well, resulting in JIT emitting these insns:
46: movabs $0x800000000000,%r11
4d: cmp %r11,%rdi
50: jae 0x0000000000000056
52: xor %edi,%edi
54: jmp 0x000000000000005a
56: mov 0x0(%rdi),%rdi
5a:
Although the above example also has src_reg == dst_reg, the same
offset == 0 optimization is valid to apply if src_reg != dst_reg.
To summarize the improvements in emitted insn count for the
check-and-load:
BEFORE: 8 check insns, 3 jmps
AFTER (general case): 7 check insns, 2 jmps (12.5% fewer insn, 33% jmp)
AFTER (src == dst): 6 check insns, 2 jmps (25% fewer insn)
AFTER (offset == 0): 5 check insns, 2 jmps (37.5% fewer insn)
(Above counts don't include the 1 load insn, just checking around it)
Based on BPF bytecode + JITted x86 insn I saw while experimenting with
these improvements, I expect the src_reg == dst_reg case to occur most
often, followed by offset == 0, then the general case.
Signed-off-by: Dave Marchevsky <davemarchevsky@fb.com>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Acked-by: Yonghong Song <yhs@fb.com>
Link: https://lore.kernel.org/bpf/20221216214319.3408356-1-davemarchevsky@fb.com
2022-12-16 21:43:18 +00:00
|
|
|
|
|
|
|
/* if unsigned '>=', goto load */
|
|
|
|
EMIT2(X86_JAE, 0);
|
|
|
|
end_of_jmp = prog;
|
2021-12-15 03:25:13 +00:00
|
|
|
|
2021-02-02 05:38:37 +00:00
|
|
|
/* xor dst_reg, dst_reg */
|
|
|
|
emit_mov_imm32(&prog, false, dst_reg, 0);
|
|
|
|
/* jmp byte_after_ldx */
|
|
|
|
EMIT2(0xEB, 0);
|
|
|
|
|
bpf, x86: Improve PROBE_MEM runtime load check
This patch rewrites the runtime PROBE_MEM check insns emitted by the BPF
JIT in order to ensure load safety. The changes in the patch fix two
issues with the previous logic and more generally improve size of
emitted code. Paragraphs between this one and "FIX 1" below explain the
purpose of the runtime check and examine the current implementation.
When a load is marked PROBE_MEM - e.g. due to PTR_UNTRUSTED access - the
address being loaded from is not necessarily valid. The BPF jit sets up
exception handlers for each such load which catch page faults and 0 out
the destination register.
Arbitrary register-relative loads can escape this exception handling
mechanism. Specifically, a load like dst_reg = *(src_reg + off) will not
trigger BPF exception handling if (src_reg + off) is outside of kernel
address space, resulting in an uncaught page fault. A concrete example
of such behavior is a program like:
struct result {
char space[40];
long a;
};
/* if err, returns ERR_PTR(-EINVAL) */
struct result *ptr = get_ptr_maybe_err();
long x = ptr->a;
If get_ptr_maybe_err returns ERR_PTR(-EINVAL) and the result isn't
checked for err, 'result' will be (u64)-EINVAL, a number close to
U64_MAX. The ptr->a load will be > U64_MAX and will wrap over to a small
positive u64, which will be in userspace and thus not covered by BPF
exception handling mechanism.
In order to prevent such loads from occurring, the BPF jit emits some
instructions which do runtime checking of (src_reg + off) and skip the
actual load if it's out of range. As an example, here are instructions
emitted for a %rdi = *(%rdi + 0x10) PROBE_MEM load:
72: movabs $0x800000000010,%r11 --|
7c: cmp %r11,%rdi |- 72 - 7f: Check 1
7f: jb 0x000000000000008d --|
81: mov %rdi,%r11 -----|
84: add $0x0000000000000010,%r11 |- 81-8b: Check 2
8b: jnc 0x0000000000000091 -----|
8d: xor %edi,%edi ---- 0 out dest
8f: jmp 0x0000000000000095
91: mov 0x10(%rdi),%rdi ---- Actual load
95:
The JIT considers kernel address space to start at MAX_TASK_SIZE +
PAGE_SIZE. Determining whether a load will be outside of kernel address
space should be a simple check:
(src_reg + off) >= MAX_TASK_SIZE + PAGE_SIZE
But because there is only one spare register when the checking logic is
emitted, this logic is split into two checks:
Check 1: src_reg >= (MAX_TASK_SIZE + PAGE_SIZE - off)
Check 2: src_reg + off doesn't wrap over U64_MAX and result in small pos u64
Emitted insns implementing Checks 1 and 2 are annotated in the above
example. Check 1 can be done with a single spare register since the
source reg by definition is the left-hand-side of the inequality.
Since adding 'off' to both sides of Check 1's inequality results in the
original inequality we want, it's equivalent to testing that inequality.
Except in the case where src_reg + off wraps past U64_MAX, which is why
Check 2 needs to actually add src_reg + off if Check 1 passes - again
using the single spare reg.
FIX 1: The Check 1 inequality listed above is not what current code is
doing. Current code is a bit more pessimistic, instead checking:
src_reg >= (MAX_TASK_SIZE + PAGE_SIZE + abs(off))
The 0x800000000010 in above example is from this current check. If Check
1 was corrected to use the correct right-hand-side, the value would be
0x7ffffffffff0. This patch changes the checking logic more broadly (FIX
2 below will elaborate), fixing this issue as a side-effect of the
rewrite. Regardless, it's important to understand why Check 1 should've
been doing MAX_TASK_SIZE + PAGE_SIZE - off before proceeding.
FIX 2: Current code relies on a 'jnc' to determine whether src_reg + off
addition wrapped over. For negative offsets this logic is incorrect.
Consider Check 2 insns emitted when off = -0x10:
81: mov %rdi,%r11
84: add 0xfffffffffffffff0,%r11
8b: jnc 0x0000000000000091
2's complement representation of -0x10 is a large positive u64. Any
value of src_reg that passes Check 1 will result in carry flag being set
after (src_reg + off) addition. So a load with any negative offset will
always fail Check 2 at runtime and never do the actual load. This patch
fixes the negative offset issue by rewriting both checks in order to not
rely on carry flag.
The rewrite takes advantage of the fact that, while we only have one
scratch reg to hold arbitrary values, we know the offset at JIT time.
This we can use src_reg as a temporary scratch reg to hold src_reg +
offset since we can return it to its original value by later subtracting
offset. As a result we can directly check the original inequality we
care about:
(src_reg + off) >= MAX_TASK_SIZE + PAGE_SIZE
For a load like %rdi = *(%rsi + -0x10), this results in emitted code:
43: movabs $0x800000000000,%r11
4d: add $0xfffffffffffffff0,%rsi --- src_reg += off
54: cmp %r11,%rsi --- Check original inequality
57: jae 0x000000000000005d
59: xor %edi,%edi
5b: jmp 0x0000000000000061
5d: mov 0x0(%rdi),%rsi --- Actual Load
61: sub $0xfffffffffffffff0,%rsi --- src_reg -= off
Note that the actual load is always done with offset 0, since previous
insns have already done src_reg += off. Regardless of whether the new
check succeeds or fails, insn 61 is always executed, returning src_reg
to its original value.
Because the goal of these checks is to ensure that loaded-from address
will be protected by BPF exception handler, the new check can safely
ignore any wrapover from insn 4d. If such wrapped-over address passes
insn 54 + 57's cmp-and-jmp it will have such protection so the load can
proceed.
IMPROVEMENTS: The above improved logic is 8 insns vs original logic's 9,
and has 1 fewer jmp. The number of checking insns can be further
improved in common scenarios:
If src_reg == dst_reg, the actual load insn will clobber src_reg, so
there's no original src_reg state for the sub insn immediately following
the load to restore, so it can be omitted. In fact, it must be omitted
since it would incorrectly subtract from the result of the load if it
wasn't. So for src_reg == dst_reg, JIT emits these insns:
3c: movabs $0x800000000000,%r11
46: add $0xfffffffffffffff0,%rdi
4d: cmp %r11,%rdi
50: jae 0x0000000000000056
52: xor %edi,%edi
54: jmp 0x000000000000005a
56: mov 0x0(%rdi),%rdi
5a:
The only difference from larger example being the omitted sub, which
would've been insn 5a in this example.
If offset == 0, we can similarly omit the sub as in previous case, since
there's nothing added to subtract. For the same reason we can omit the
addition as well, resulting in JIT emitting these insns:
46: movabs $0x800000000000,%r11
4d: cmp %r11,%rdi
50: jae 0x0000000000000056
52: xor %edi,%edi
54: jmp 0x000000000000005a
56: mov 0x0(%rdi),%rdi
5a:
Although the above example also has src_reg == dst_reg, the same
offset == 0 optimization is valid to apply if src_reg != dst_reg.
To summarize the improvements in emitted insn count for the
check-and-load:
BEFORE: 8 check insns, 3 jmps
AFTER (general case): 7 check insns, 2 jmps (12.5% fewer insn, 33% jmp)
AFTER (src == dst): 6 check insns, 2 jmps (25% fewer insn)
AFTER (offset == 0): 5 check insns, 2 jmps (37.5% fewer insn)
(Above counts don't include the 1 load insn, just checking around it)
Based on BPF bytecode + JITted x86 insn I saw while experimenting with
these improvements, I expect the src_reg == dst_reg case to occur most
often, followed by offset == 0, then the general case.
Signed-off-by: Dave Marchevsky <davemarchevsky@fb.com>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Acked-by: Yonghong Song <yhs@fb.com>
Link: https://lore.kernel.org/bpf/20221216214319.3408356-1-davemarchevsky@fb.com
2022-12-16 21:43:18 +00:00
|
|
|
/* populate jmp_offset for JAE above to jump to start_of_ldx */
|
2021-02-02 05:38:37 +00:00
|
|
|
start_of_ldx = prog;
|
bpf, x86: Improve PROBE_MEM runtime load check
This patch rewrites the runtime PROBE_MEM check insns emitted by the BPF
JIT in order to ensure load safety. The changes in the patch fix two
issues with the previous logic and more generally improve size of
emitted code. Paragraphs between this one and "FIX 1" below explain the
purpose of the runtime check and examine the current implementation.
When a load is marked PROBE_MEM - e.g. due to PTR_UNTRUSTED access - the
address being loaded from is not necessarily valid. The BPF jit sets up
exception handlers for each such load which catch page faults and 0 out
the destination register.
Arbitrary register-relative loads can escape this exception handling
mechanism. Specifically, a load like dst_reg = *(src_reg + off) will not
trigger BPF exception handling if (src_reg + off) is outside of kernel
address space, resulting in an uncaught page fault. A concrete example
of such behavior is a program like:
struct result {
char space[40];
long a;
};
/* if err, returns ERR_PTR(-EINVAL) */
struct result *ptr = get_ptr_maybe_err();
long x = ptr->a;
If get_ptr_maybe_err returns ERR_PTR(-EINVAL) and the result isn't
checked for err, 'result' will be (u64)-EINVAL, a number close to
U64_MAX. The ptr->a load will be > U64_MAX and will wrap over to a small
positive u64, which will be in userspace and thus not covered by BPF
exception handling mechanism.
In order to prevent such loads from occurring, the BPF jit emits some
instructions which do runtime checking of (src_reg + off) and skip the
actual load if it's out of range. As an example, here are instructions
emitted for a %rdi = *(%rdi + 0x10) PROBE_MEM load:
72: movabs $0x800000000010,%r11 --|
7c: cmp %r11,%rdi |- 72 - 7f: Check 1
7f: jb 0x000000000000008d --|
81: mov %rdi,%r11 -----|
84: add $0x0000000000000010,%r11 |- 81-8b: Check 2
8b: jnc 0x0000000000000091 -----|
8d: xor %edi,%edi ---- 0 out dest
8f: jmp 0x0000000000000095
91: mov 0x10(%rdi),%rdi ---- Actual load
95:
The JIT considers kernel address space to start at MAX_TASK_SIZE +
PAGE_SIZE. Determining whether a load will be outside of kernel address
space should be a simple check:
(src_reg + off) >= MAX_TASK_SIZE + PAGE_SIZE
But because there is only one spare register when the checking logic is
emitted, this logic is split into two checks:
Check 1: src_reg >= (MAX_TASK_SIZE + PAGE_SIZE - off)
Check 2: src_reg + off doesn't wrap over U64_MAX and result in small pos u64
Emitted insns implementing Checks 1 and 2 are annotated in the above
example. Check 1 can be done with a single spare register since the
source reg by definition is the left-hand-side of the inequality.
Since adding 'off' to both sides of Check 1's inequality results in the
original inequality we want, it's equivalent to testing that inequality.
Except in the case where src_reg + off wraps past U64_MAX, which is why
Check 2 needs to actually add src_reg + off if Check 1 passes - again
using the single spare reg.
FIX 1: The Check 1 inequality listed above is not what current code is
doing. Current code is a bit more pessimistic, instead checking:
src_reg >= (MAX_TASK_SIZE + PAGE_SIZE + abs(off))
The 0x800000000010 in above example is from this current check. If Check
1 was corrected to use the correct right-hand-side, the value would be
0x7ffffffffff0. This patch changes the checking logic more broadly (FIX
2 below will elaborate), fixing this issue as a side-effect of the
rewrite. Regardless, it's important to understand why Check 1 should've
been doing MAX_TASK_SIZE + PAGE_SIZE - off before proceeding.
FIX 2: Current code relies on a 'jnc' to determine whether src_reg + off
addition wrapped over. For negative offsets this logic is incorrect.
Consider Check 2 insns emitted when off = -0x10:
81: mov %rdi,%r11
84: add 0xfffffffffffffff0,%r11
8b: jnc 0x0000000000000091
2's complement representation of -0x10 is a large positive u64. Any
value of src_reg that passes Check 1 will result in carry flag being set
after (src_reg + off) addition. So a load with any negative offset will
always fail Check 2 at runtime and never do the actual load. This patch
fixes the negative offset issue by rewriting both checks in order to not
rely on carry flag.
The rewrite takes advantage of the fact that, while we only have one
scratch reg to hold arbitrary values, we know the offset at JIT time.
This we can use src_reg as a temporary scratch reg to hold src_reg +
offset since we can return it to its original value by later subtracting
offset. As a result we can directly check the original inequality we
care about:
(src_reg + off) >= MAX_TASK_SIZE + PAGE_SIZE
For a load like %rdi = *(%rsi + -0x10), this results in emitted code:
43: movabs $0x800000000000,%r11
4d: add $0xfffffffffffffff0,%rsi --- src_reg += off
54: cmp %r11,%rsi --- Check original inequality
57: jae 0x000000000000005d
59: xor %edi,%edi
5b: jmp 0x0000000000000061
5d: mov 0x0(%rdi),%rsi --- Actual Load
61: sub $0xfffffffffffffff0,%rsi --- src_reg -= off
Note that the actual load is always done with offset 0, since previous
insns have already done src_reg += off. Regardless of whether the new
check succeeds or fails, insn 61 is always executed, returning src_reg
to its original value.
Because the goal of these checks is to ensure that loaded-from address
will be protected by BPF exception handler, the new check can safely
ignore any wrapover from insn 4d. If such wrapped-over address passes
insn 54 + 57's cmp-and-jmp it will have such protection so the load can
proceed.
IMPROVEMENTS: The above improved logic is 8 insns vs original logic's 9,
and has 1 fewer jmp. The number of checking insns can be further
improved in common scenarios:
If src_reg == dst_reg, the actual load insn will clobber src_reg, so
there's no original src_reg state for the sub insn immediately following
the load to restore, so it can be omitted. In fact, it must be omitted
since it would incorrectly subtract from the result of the load if it
wasn't. So for src_reg == dst_reg, JIT emits these insns:
3c: movabs $0x800000000000,%r11
46: add $0xfffffffffffffff0,%rdi
4d: cmp %r11,%rdi
50: jae 0x0000000000000056
52: xor %edi,%edi
54: jmp 0x000000000000005a
56: mov 0x0(%rdi),%rdi
5a:
The only difference from larger example being the omitted sub, which
would've been insn 5a in this example.
If offset == 0, we can similarly omit the sub as in previous case, since
there's nothing added to subtract. For the same reason we can omit the
addition as well, resulting in JIT emitting these insns:
46: movabs $0x800000000000,%r11
4d: cmp %r11,%rdi
50: jae 0x0000000000000056
52: xor %edi,%edi
54: jmp 0x000000000000005a
56: mov 0x0(%rdi),%rdi
5a:
Although the above example also has src_reg == dst_reg, the same
offset == 0 optimization is valid to apply if src_reg != dst_reg.
To summarize the improvements in emitted insn count for the
check-and-load:
BEFORE: 8 check insns, 3 jmps
AFTER (general case): 7 check insns, 2 jmps (12.5% fewer insn, 33% jmp)
AFTER (src == dst): 6 check insns, 2 jmps (25% fewer insn)
AFTER (offset == 0): 5 check insns, 2 jmps (37.5% fewer insn)
(Above counts don't include the 1 load insn, just checking around it)
Based on BPF bytecode + JITted x86 insn I saw while experimenting with
these improvements, I expect the src_reg == dst_reg case to occur most
often, followed by offset == 0, then the general case.
Signed-off-by: Dave Marchevsky <davemarchevsky@fb.com>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Acked-by: Yonghong Song <yhs@fb.com>
Link: https://lore.kernel.org/bpf/20221216214319.3408356-1-davemarchevsky@fb.com
2022-12-16 21:43:18 +00:00
|
|
|
end_of_jmp[-1] = start_of_ldx - end_of_jmp;
|
2021-02-02 05:38:37 +00:00
|
|
|
}
|
2023-07-28 01:11:56 +00:00
|
|
|
if (BPF_MODE(insn->code) == BPF_PROBE_MEMSX ||
|
|
|
|
BPF_MODE(insn->code) == BPF_MEMSX)
|
|
|
|
emit_ldsx(&prog, BPF_SIZE(insn->code), dst_reg, src_reg, insn_off);
|
|
|
|
else
|
|
|
|
emit_ldx(&prog, BPF_SIZE(insn->code), dst_reg, src_reg, insn_off);
|
|
|
|
if (BPF_MODE(insn->code) == BPF_PROBE_MEM ||
|
|
|
|
BPF_MODE(insn->code) == BPF_PROBE_MEMSX) {
|
2019-10-16 03:25:03 +00:00
|
|
|
struct exception_table_entry *ex;
|
2021-06-22 11:00:26 +00:00
|
|
|
u8 *_insn = image + proglen + (start_of_ldx - temp);
|
2019-10-16 03:25:03 +00:00
|
|
|
s64 delta;
|
|
|
|
|
2021-02-02 05:38:37 +00:00
|
|
|
/* populate jmp_offset for JMP above */
|
|
|
|
start_of_ldx[-1] = prog - start_of_ldx;
|
|
|
|
|
bpf, x86: Improve PROBE_MEM runtime load check
This patch rewrites the runtime PROBE_MEM check insns emitted by the BPF
JIT in order to ensure load safety. The changes in the patch fix two
issues with the previous logic and more generally improve size of
emitted code. Paragraphs between this one and "FIX 1" below explain the
purpose of the runtime check and examine the current implementation.
When a load is marked PROBE_MEM - e.g. due to PTR_UNTRUSTED access - the
address being loaded from is not necessarily valid. The BPF jit sets up
exception handlers for each such load which catch page faults and 0 out
the destination register.
Arbitrary register-relative loads can escape this exception handling
mechanism. Specifically, a load like dst_reg = *(src_reg + off) will not
trigger BPF exception handling if (src_reg + off) is outside of kernel
address space, resulting in an uncaught page fault. A concrete example
of such behavior is a program like:
struct result {
char space[40];
long a;
};
/* if err, returns ERR_PTR(-EINVAL) */
struct result *ptr = get_ptr_maybe_err();
long x = ptr->a;
If get_ptr_maybe_err returns ERR_PTR(-EINVAL) and the result isn't
checked for err, 'result' will be (u64)-EINVAL, a number close to
U64_MAX. The ptr->a load will be > U64_MAX and will wrap over to a small
positive u64, which will be in userspace and thus not covered by BPF
exception handling mechanism.
In order to prevent such loads from occurring, the BPF jit emits some
instructions which do runtime checking of (src_reg + off) and skip the
actual load if it's out of range. As an example, here are instructions
emitted for a %rdi = *(%rdi + 0x10) PROBE_MEM load:
72: movabs $0x800000000010,%r11 --|
7c: cmp %r11,%rdi |- 72 - 7f: Check 1
7f: jb 0x000000000000008d --|
81: mov %rdi,%r11 -----|
84: add $0x0000000000000010,%r11 |- 81-8b: Check 2
8b: jnc 0x0000000000000091 -----|
8d: xor %edi,%edi ---- 0 out dest
8f: jmp 0x0000000000000095
91: mov 0x10(%rdi),%rdi ---- Actual load
95:
The JIT considers kernel address space to start at MAX_TASK_SIZE +
PAGE_SIZE. Determining whether a load will be outside of kernel address
space should be a simple check:
(src_reg + off) >= MAX_TASK_SIZE + PAGE_SIZE
But because there is only one spare register when the checking logic is
emitted, this logic is split into two checks:
Check 1: src_reg >= (MAX_TASK_SIZE + PAGE_SIZE - off)
Check 2: src_reg + off doesn't wrap over U64_MAX and result in small pos u64
Emitted insns implementing Checks 1 and 2 are annotated in the above
example. Check 1 can be done with a single spare register since the
source reg by definition is the left-hand-side of the inequality.
Since adding 'off' to both sides of Check 1's inequality results in the
original inequality we want, it's equivalent to testing that inequality.
Except in the case where src_reg + off wraps past U64_MAX, which is why
Check 2 needs to actually add src_reg + off if Check 1 passes - again
using the single spare reg.
FIX 1: The Check 1 inequality listed above is not what current code is
doing. Current code is a bit more pessimistic, instead checking:
src_reg >= (MAX_TASK_SIZE + PAGE_SIZE + abs(off))
The 0x800000000010 in above example is from this current check. If Check
1 was corrected to use the correct right-hand-side, the value would be
0x7ffffffffff0. This patch changes the checking logic more broadly (FIX
2 below will elaborate), fixing this issue as a side-effect of the
rewrite. Regardless, it's important to understand why Check 1 should've
been doing MAX_TASK_SIZE + PAGE_SIZE - off before proceeding.
FIX 2: Current code relies on a 'jnc' to determine whether src_reg + off
addition wrapped over. For negative offsets this logic is incorrect.
Consider Check 2 insns emitted when off = -0x10:
81: mov %rdi,%r11
84: add 0xfffffffffffffff0,%r11
8b: jnc 0x0000000000000091
2's complement representation of -0x10 is a large positive u64. Any
value of src_reg that passes Check 1 will result in carry flag being set
after (src_reg + off) addition. So a load with any negative offset will
always fail Check 2 at runtime and never do the actual load. This patch
fixes the negative offset issue by rewriting both checks in order to not
rely on carry flag.
The rewrite takes advantage of the fact that, while we only have one
scratch reg to hold arbitrary values, we know the offset at JIT time.
This we can use src_reg as a temporary scratch reg to hold src_reg +
offset since we can return it to its original value by later subtracting
offset. As a result we can directly check the original inequality we
care about:
(src_reg + off) >= MAX_TASK_SIZE + PAGE_SIZE
For a load like %rdi = *(%rsi + -0x10), this results in emitted code:
43: movabs $0x800000000000,%r11
4d: add $0xfffffffffffffff0,%rsi --- src_reg += off
54: cmp %r11,%rsi --- Check original inequality
57: jae 0x000000000000005d
59: xor %edi,%edi
5b: jmp 0x0000000000000061
5d: mov 0x0(%rdi),%rsi --- Actual Load
61: sub $0xfffffffffffffff0,%rsi --- src_reg -= off
Note that the actual load is always done with offset 0, since previous
insns have already done src_reg += off. Regardless of whether the new
check succeeds or fails, insn 61 is always executed, returning src_reg
to its original value.
Because the goal of these checks is to ensure that loaded-from address
will be protected by BPF exception handler, the new check can safely
ignore any wrapover from insn 4d. If such wrapped-over address passes
insn 54 + 57's cmp-and-jmp it will have such protection so the load can
proceed.
IMPROVEMENTS: The above improved logic is 8 insns vs original logic's 9,
and has 1 fewer jmp. The number of checking insns can be further
improved in common scenarios:
If src_reg == dst_reg, the actual load insn will clobber src_reg, so
there's no original src_reg state for the sub insn immediately following
the load to restore, so it can be omitted. In fact, it must be omitted
since it would incorrectly subtract from the result of the load if it
wasn't. So for src_reg == dst_reg, JIT emits these insns:
3c: movabs $0x800000000000,%r11
46: add $0xfffffffffffffff0,%rdi
4d: cmp %r11,%rdi
50: jae 0x0000000000000056
52: xor %edi,%edi
54: jmp 0x000000000000005a
56: mov 0x0(%rdi),%rdi
5a:
The only difference from larger example being the omitted sub, which
would've been insn 5a in this example.
If offset == 0, we can similarly omit the sub as in previous case, since
there's nothing added to subtract. For the same reason we can omit the
addition as well, resulting in JIT emitting these insns:
46: movabs $0x800000000000,%r11
4d: cmp %r11,%rdi
50: jae 0x0000000000000056
52: xor %edi,%edi
54: jmp 0x000000000000005a
56: mov 0x0(%rdi),%rdi
5a:
Although the above example also has src_reg == dst_reg, the same
offset == 0 optimization is valid to apply if src_reg != dst_reg.
To summarize the improvements in emitted insn count for the
check-and-load:
BEFORE: 8 check insns, 3 jmps
AFTER (general case): 7 check insns, 2 jmps (12.5% fewer insn, 33% jmp)
AFTER (src == dst): 6 check insns, 2 jmps (25% fewer insn)
AFTER (offset == 0): 5 check insns, 2 jmps (37.5% fewer insn)
(Above counts don't include the 1 load insn, just checking around it)
Based on BPF bytecode + JITted x86 insn I saw while experimenting with
these improvements, I expect the src_reg == dst_reg case to occur most
often, followed by offset == 0, then the general case.
Signed-off-by: Dave Marchevsky <davemarchevsky@fb.com>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Acked-by: Yonghong Song <yhs@fb.com>
Link: https://lore.kernel.org/bpf/20221216214319.3408356-1-davemarchevsky@fb.com
2022-12-16 21:43:18 +00:00
|
|
|
if (insn->off && src_reg != dst_reg) {
|
|
|
|
/* sub src_reg, insn->off
|
|
|
|
* Restore src_reg after "add src_reg, insn->off" in prev
|
|
|
|
* if statement. But if src_reg == dst_reg, emit_ldx
|
|
|
|
* above already clobbered src_reg, so no need to restore.
|
|
|
|
* If add src_reg, insn->off was unnecessary, no need to
|
|
|
|
* restore either.
|
|
|
|
*/
|
|
|
|
maybe_emit_1mod(&prog, src_reg, true);
|
|
|
|
EMIT2_off32(0x81, add_1reg(0xE8, src_reg), insn->off);
|
|
|
|
}
|
|
|
|
|
2019-10-16 03:25:03 +00:00
|
|
|
if (!bpf_prog->aux->extable)
|
|
|
|
break;
|
|
|
|
|
|
|
|
if (excnt >= bpf_prog->aux->num_exentries) {
|
|
|
|
pr_err("ex gen bug\n");
|
|
|
|
return -EFAULT;
|
|
|
|
}
|
|
|
|
ex = &bpf_prog->aux->extable[excnt++];
|
|
|
|
|
|
|
|
delta = _insn - (u8 *)&ex->insn;
|
|
|
|
if (!is_simm32(delta)) {
|
|
|
|
pr_err("extable->insn doesn't fit into 32-bit\n");
|
|
|
|
return -EFAULT;
|
|
|
|
}
|
2022-02-04 18:57:42 +00:00
|
|
|
/* switch ex to rw buffer for writes */
|
|
|
|
ex = (void *)rw_image + ((void *)ex - (void *)image);
|
|
|
|
|
2019-10-16 03:25:03 +00:00
|
|
|
ex->insn = delta;
|
|
|
|
|
2021-11-10 10:01:09 +00:00
|
|
|
ex->data = EX_TYPE_BPF;
|
2019-10-16 03:25:03 +00:00
|
|
|
|
|
|
|
if (dst_reg > BPF_REG_9) {
|
|
|
|
pr_err("verifier error\n");
|
|
|
|
return -EFAULT;
|
|
|
|
}
|
|
|
|
/*
|
|
|
|
* Compute size of x86 insn and its target dest x86 register.
|
|
|
|
* ex_handler_bpf() will use lower 8 bits to adjust
|
|
|
|
* pt_regs->ip to jump over this x86 instruction
|
|
|
|
* and upper bits to figure out which pt_regs to zero out.
|
|
|
|
* End result: x86 insn "mov rbx, qword ptr [rax+0x14]"
|
|
|
|
* of 4 bytes will be ignored and rbx will be zero inited.
|
|
|
|
*/
|
2021-12-16 02:38:30 +00:00
|
|
|
ex->fixup = (prog - start_of_ldx) | (reg2pt_regs[dst_reg] << 8);
|
2019-10-16 03:25:03 +00:00
|
|
|
}
|
net: filter: x86: internal BPF JIT
Maps all internal BPF instructions into x86_64 instructions.
This patch replaces original BPF x64 JIT with internal BPF x64 JIT.
sysctl net.core.bpf_jit_enable is reused as on/off switch.
Performance:
1. old BPF JIT and internal BPF JIT generate equivalent x86_64 code.
No performance difference is observed for filters that were JIT-able before
Example assembler code for BPF filter "tcpdump port 22"
original BPF -> old JIT: original BPF -> internal BPF -> new JIT:
0: push %rbp 0: push %rbp
1: mov %rsp,%rbp 1: mov %rsp,%rbp
4: sub $0x60,%rsp 4: sub $0x228,%rsp
8: mov %rbx,-0x8(%rbp) b: mov %rbx,-0x228(%rbp) // prologue
12: mov %r13,-0x220(%rbp)
19: mov %r14,-0x218(%rbp)
20: mov %r15,-0x210(%rbp)
27: xor %eax,%eax // clear A
c: xor %ebx,%ebx 29: xor %r13,%r13 // clear X
e: mov 0x68(%rdi),%r9d 2c: mov 0x68(%rdi),%r9d
12: sub 0x6c(%rdi),%r9d 30: sub 0x6c(%rdi),%r9d
16: mov 0xd8(%rdi),%r8 34: mov 0xd8(%rdi),%r10
3b: mov %rdi,%rbx
1d: mov $0xc,%esi 3e: mov $0xc,%esi
22: callq 0xffffffffe1021e15 43: callq 0xffffffffe102bd75
27: cmp $0x86dd,%eax 48: cmp $0x86dd,%rax
2c: jne 0x0000000000000069 4f: jne 0x000000000000009a
2e: mov $0x14,%esi 51: mov $0x14,%esi
33: callq 0xffffffffe1021e31 56: callq 0xffffffffe102bd91
38: cmp $0x84,%eax 5b: cmp $0x84,%rax
3d: je 0x0000000000000049 62: je 0x0000000000000074
3f: cmp $0x6,%eax 64: cmp $0x6,%rax
42: je 0x0000000000000049 68: je 0x0000000000000074
44: cmp $0x11,%eax 6a: cmp $0x11,%rax
47: jne 0x00000000000000c6 6e: jne 0x0000000000000117
49: mov $0x36,%esi 74: mov $0x36,%esi
4e: callq 0xffffffffe1021e15 79: callq 0xffffffffe102bd75
53: cmp $0x16,%eax 7e: cmp $0x16,%rax
56: je 0x00000000000000bf 82: je 0x0000000000000110
58: mov $0x38,%esi 88: mov $0x38,%esi
5d: callq 0xffffffffe1021e15 8d: callq 0xffffffffe102bd75
62: cmp $0x16,%eax 92: cmp $0x16,%rax
65: je 0x00000000000000bf 96: je 0x0000000000000110
67: jmp 0x00000000000000c6 98: jmp 0x0000000000000117
69: cmp $0x800,%eax 9a: cmp $0x800,%rax
6e: jne 0x00000000000000c6 a1: jne 0x0000000000000117
70: mov $0x17,%esi a3: mov $0x17,%esi
75: callq 0xffffffffe1021e31 a8: callq 0xffffffffe102bd91
7a: cmp $0x84,%eax ad: cmp $0x84,%rax
7f: je 0x000000000000008b b4: je 0x00000000000000c2
81: cmp $0x6,%eax b6: cmp $0x6,%rax
84: je 0x000000000000008b ba: je 0x00000000000000c2
86: cmp $0x11,%eax bc: cmp $0x11,%rax
89: jne 0x00000000000000c6 c0: jne 0x0000000000000117
8b: mov $0x14,%esi c2: mov $0x14,%esi
90: callq 0xffffffffe1021e15 c7: callq 0xffffffffe102bd75
95: test $0x1fff,%ax cc: test $0x1fff,%rax
99: jne 0x00000000000000c6 d3: jne 0x0000000000000117
d5: mov %rax,%r14
9b: mov $0xe,%esi d8: mov $0xe,%esi
a0: callq 0xffffffffe1021e44 dd: callq 0xffffffffe102bd91 // MSH
e2: and $0xf,%eax
e5: shl $0x2,%eax
e8: mov %rax,%r13
eb: mov %r14,%rax
ee: mov %r13,%rsi
a5: lea 0xe(%rbx),%esi f1: add $0xe,%esi
a8: callq 0xffffffffe1021e0d f4: callq 0xffffffffe102bd6d
ad: cmp $0x16,%eax f9: cmp $0x16,%rax
b0: je 0x00000000000000bf fd: je 0x0000000000000110
ff: mov %r13,%rsi
b2: lea 0x10(%rbx),%esi 102: add $0x10,%esi
b5: callq 0xffffffffe1021e0d 105: callq 0xffffffffe102bd6d
ba: cmp $0x16,%eax 10a: cmp $0x16,%rax
bd: jne 0x00000000000000c6 10e: jne 0x0000000000000117
bf: mov $0xffff,%eax 110: mov $0xffff,%eax
c4: jmp 0x00000000000000c8 115: jmp 0x000000000000011c
c6: xor %eax,%eax 117: mov $0x0,%eax
c8: mov -0x8(%rbp),%rbx 11c: mov -0x228(%rbp),%rbx // epilogue
cc: leaveq 123: mov -0x220(%rbp),%r13
cd: retq 12a: mov -0x218(%rbp),%r14
131: mov -0x210(%rbp),%r15
138: leaveq
139: retq
On fully cached SKBs both JITed functions take 12 nsec to execute.
BPF interpreter executes the program in 30 nsec.
The difference in generated assembler is due to the following:
Old BPF imlements LDX_MSH instruction via sk_load_byte_msh() helper function
inside bpf_jit.S.
New JIT removes the helper and does it explicitly, so ldx_msh cost
is the same for both JITs, but generated code looks longer.
New JIT has 4 registers to save, so prologue/epilogue are larger,
but the cost is within noise on x64.
Old JIT checks whether first insn clears A and if not emits 'xor %eax,%eax'.
New JIT clears %rax unconditionally.
2. old BPF JIT doesn't support ANC_NLATTR, ANC_PAY_OFFSET, ANC_RANDOM
extensions. New JIT supports all BPF extensions.
Performance of such filters improves 2-4 times depending on a filter.
The longer the filter the higher performance gain.
Synthetic benchmarks with many ancillary loads see 20x speedup
which seems to be the maximum gain from JIT
Notes:
. net.core.bpf_jit_enable=2 + tools/net/bpf_jit_disasm is still functional
and can be used to see generated assembler
. there are two jit_compile() functions and code flow for classic filters is:
sk_attach_filter() - load classic BPF
bpf_jit_compile() - try to JIT from classic BPF
sk_convert_filter() - convert classic to internal
bpf_int_jit_compile() - JIT from internal BPF
seccomp and tracing filters will just call bpf_int_jit_compile()
Signed-off-by: Alexei Starovoitov <ast@plumgrid.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
2014-05-14 02:50:46 +00:00
|
|
|
break;
|
|
|
|
|
2021-01-14 18:17:44 +00:00
|
|
|
case BPF_STX | BPF_ATOMIC | BPF_W:
|
|
|
|
case BPF_STX | BPF_ATOMIC | BPF_DW:
|
2021-01-14 18:17:49 +00:00
|
|
|
if (insn->imm == (BPF_AND | BPF_FETCH) ||
|
|
|
|
insn->imm == (BPF_OR | BPF_FETCH) ||
|
|
|
|
insn->imm == (BPF_XOR | BPF_FETCH)) {
|
|
|
|
bool is64 = BPF_SIZE(insn->code) == BPF_DW;
|
2021-02-16 12:53:07 +00:00
|
|
|
u32 real_src_reg = src_reg;
|
bpf, x86: Fix bpf mapping of atomic fetch implementation
Fix the case where the dst register maps to %rax as otherwise this produces
an incorrect mapping with the implementation in 981f94c3e921 ("bpf: Add
bitwise atomic instructions") as %rax is clobbered given it's part of the
cmpxchg as operand.
The issue is similar to b29dd96b905f ("bpf, x86: Fix BPF_FETCH atomic and/or/
xor with r0 as src") just that the case of dst register was missed.
Before, dst=r0 (%rax) src=r2 (%rsi):
[...]
c5: mov %rax,%r10
c8: mov 0x0(%rax),%rax <---+ (broken)
cc: mov %rax,%r11 |
cf: and %rsi,%r11 |
d2: lock cmpxchg %r11,0x0(%rax) <---+
d8: jne 0x00000000000000c8 |
da: mov %rax,%rsi |
dd: mov %r10,%rax |
[...] |
|
After, dst=r0 (%rax) src=r2 (%rsi): |
|
[...] |
da: mov %rax,%r10 |
dd: mov 0x0(%r10),%rax <---+ (fixed)
e1: mov %rax,%r11 |
e4: and %rsi,%r11 |
e7: lock cmpxchg %r11,0x0(%r10) <---+
ed: jne 0x00000000000000dd
ef: mov %rax,%rsi
f2: mov %r10,%rax
[...]
The remaining combinations were fine as-is though:
After, dst=r9 (%r15) src=r0 (%rax):
[...]
dc: mov %rax,%r10
df: mov 0x0(%r15),%rax
e3: mov %rax,%r11
e6: and %r10,%r11
e9: lock cmpxchg %r11,0x0(%r15)
ef: jne 0x00000000000000df _
f1: mov %rax,%r10 | (unneeded, but
f4: mov %r10,%rax _| not a problem)
[...]
After, dst=r9 (%r15) src=r4 (%rcx):
[...]
de: mov %rax,%r10
e1: mov 0x0(%r15),%rax
e5: mov %rax,%r11
e8: and %rcx,%r11
eb: lock cmpxchg %r11,0x0(%r15)
f1: jne 0x00000000000000e1
f3: mov %rax,%rcx
f6: mov %r10,%rax
[...]
The case of dst == src register is rejected by the verifier and
therefore not supported, but x86 JIT also handles this case just
fine.
After, dst=r0 (%rax) src=r0 (%rax):
[...]
eb: mov %rax,%r10
ee: mov 0x0(%r10),%rax
f2: mov %rax,%r11
f5: and %r10,%r11
f8: lock cmpxchg %r11,0x0(%r10)
fe: jne 0x00000000000000ee
100: mov %rax,%r10
103: mov %r10,%rax
[...]
Fixes: 981f94c3e921 ("bpf: Add bitwise atomic instructions")
Reported-by: Johan Almbladh <johan.almbladh@anyfinetworks.com>
Signed-off-by: Johan Almbladh <johan.almbladh@anyfinetworks.com>
Co-developed-by: Daniel Borkmann <daniel@iogearbox.net>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Reviewed-by: Brendan Jackman <jackmanb@google.com>
Acked-by: Alexei Starovoitov <ast@kernel.org>
2021-09-27 13:11:57 +00:00
|
|
|
u32 real_dst_reg = dst_reg;
|
|
|
|
u8 *branch_target;
|
2021-01-14 18:17:49 +00:00
|
|
|
|
|
|
|
/*
|
|
|
|
* Can't be implemented with a single x86 insn.
|
|
|
|
* Need to do a CMPXCHG loop.
|
|
|
|
*/
|
|
|
|
|
|
|
|
/* Will need RAX as a CMPXCHG operand so save R0 */
|
|
|
|
emit_mov_reg(&prog, true, BPF_REG_AX, BPF_REG_0);
|
2021-02-16 12:53:07 +00:00
|
|
|
if (src_reg == BPF_REG_0)
|
|
|
|
real_src_reg = BPF_REG_AX;
|
bpf, x86: Fix bpf mapping of atomic fetch implementation
Fix the case where the dst register maps to %rax as otherwise this produces
an incorrect mapping with the implementation in 981f94c3e921 ("bpf: Add
bitwise atomic instructions") as %rax is clobbered given it's part of the
cmpxchg as operand.
The issue is similar to b29dd96b905f ("bpf, x86: Fix BPF_FETCH atomic and/or/
xor with r0 as src") just that the case of dst register was missed.
Before, dst=r0 (%rax) src=r2 (%rsi):
[...]
c5: mov %rax,%r10
c8: mov 0x0(%rax),%rax <---+ (broken)
cc: mov %rax,%r11 |
cf: and %rsi,%r11 |
d2: lock cmpxchg %r11,0x0(%rax) <---+
d8: jne 0x00000000000000c8 |
da: mov %rax,%rsi |
dd: mov %r10,%rax |
[...] |
|
After, dst=r0 (%rax) src=r2 (%rsi): |
|
[...] |
da: mov %rax,%r10 |
dd: mov 0x0(%r10),%rax <---+ (fixed)
e1: mov %rax,%r11 |
e4: and %rsi,%r11 |
e7: lock cmpxchg %r11,0x0(%r10) <---+
ed: jne 0x00000000000000dd
ef: mov %rax,%rsi
f2: mov %r10,%rax
[...]
The remaining combinations were fine as-is though:
After, dst=r9 (%r15) src=r0 (%rax):
[...]
dc: mov %rax,%r10
df: mov 0x0(%r15),%rax
e3: mov %rax,%r11
e6: and %r10,%r11
e9: lock cmpxchg %r11,0x0(%r15)
ef: jne 0x00000000000000df _
f1: mov %rax,%r10 | (unneeded, but
f4: mov %r10,%rax _| not a problem)
[...]
After, dst=r9 (%r15) src=r4 (%rcx):
[...]
de: mov %rax,%r10
e1: mov 0x0(%r15),%rax
e5: mov %rax,%r11
e8: and %rcx,%r11
eb: lock cmpxchg %r11,0x0(%r15)
f1: jne 0x00000000000000e1
f3: mov %rax,%rcx
f6: mov %r10,%rax
[...]
The case of dst == src register is rejected by the verifier and
therefore not supported, but x86 JIT also handles this case just
fine.
After, dst=r0 (%rax) src=r0 (%rax):
[...]
eb: mov %rax,%r10
ee: mov 0x0(%r10),%rax
f2: mov %rax,%r11
f5: and %r10,%r11
f8: lock cmpxchg %r11,0x0(%r10)
fe: jne 0x00000000000000ee
100: mov %rax,%r10
103: mov %r10,%rax
[...]
Fixes: 981f94c3e921 ("bpf: Add bitwise atomic instructions")
Reported-by: Johan Almbladh <johan.almbladh@anyfinetworks.com>
Signed-off-by: Johan Almbladh <johan.almbladh@anyfinetworks.com>
Co-developed-by: Daniel Borkmann <daniel@iogearbox.net>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Reviewed-by: Brendan Jackman <jackmanb@google.com>
Acked-by: Alexei Starovoitov <ast@kernel.org>
2021-09-27 13:11:57 +00:00
|
|
|
if (dst_reg == BPF_REG_0)
|
|
|
|
real_dst_reg = BPF_REG_AX;
|
2021-02-16 12:53:07 +00:00
|
|
|
|
2021-01-14 18:17:49 +00:00
|
|
|
branch_target = prog;
|
|
|
|
/* Load old value */
|
|
|
|
emit_ldx(&prog, BPF_SIZE(insn->code),
|
bpf, x86: Fix bpf mapping of atomic fetch implementation
Fix the case where the dst register maps to %rax as otherwise this produces
an incorrect mapping with the implementation in 981f94c3e921 ("bpf: Add
bitwise atomic instructions") as %rax is clobbered given it's part of the
cmpxchg as operand.
The issue is similar to b29dd96b905f ("bpf, x86: Fix BPF_FETCH atomic and/or/
xor with r0 as src") just that the case of dst register was missed.
Before, dst=r0 (%rax) src=r2 (%rsi):
[...]
c5: mov %rax,%r10
c8: mov 0x0(%rax),%rax <---+ (broken)
cc: mov %rax,%r11 |
cf: and %rsi,%r11 |
d2: lock cmpxchg %r11,0x0(%rax) <---+
d8: jne 0x00000000000000c8 |
da: mov %rax,%rsi |
dd: mov %r10,%rax |
[...] |
|
After, dst=r0 (%rax) src=r2 (%rsi): |
|
[...] |
da: mov %rax,%r10 |
dd: mov 0x0(%r10),%rax <---+ (fixed)
e1: mov %rax,%r11 |
e4: and %rsi,%r11 |
e7: lock cmpxchg %r11,0x0(%r10) <---+
ed: jne 0x00000000000000dd
ef: mov %rax,%rsi
f2: mov %r10,%rax
[...]
The remaining combinations were fine as-is though:
After, dst=r9 (%r15) src=r0 (%rax):
[...]
dc: mov %rax,%r10
df: mov 0x0(%r15),%rax
e3: mov %rax,%r11
e6: and %r10,%r11
e9: lock cmpxchg %r11,0x0(%r15)
ef: jne 0x00000000000000df _
f1: mov %rax,%r10 | (unneeded, but
f4: mov %r10,%rax _| not a problem)
[...]
After, dst=r9 (%r15) src=r4 (%rcx):
[...]
de: mov %rax,%r10
e1: mov 0x0(%r15),%rax
e5: mov %rax,%r11
e8: and %rcx,%r11
eb: lock cmpxchg %r11,0x0(%r15)
f1: jne 0x00000000000000e1
f3: mov %rax,%rcx
f6: mov %r10,%rax
[...]
The case of dst == src register is rejected by the verifier and
therefore not supported, but x86 JIT also handles this case just
fine.
After, dst=r0 (%rax) src=r0 (%rax):
[...]
eb: mov %rax,%r10
ee: mov 0x0(%r10),%rax
f2: mov %rax,%r11
f5: and %r10,%r11
f8: lock cmpxchg %r11,0x0(%r10)
fe: jne 0x00000000000000ee
100: mov %rax,%r10
103: mov %r10,%rax
[...]
Fixes: 981f94c3e921 ("bpf: Add bitwise atomic instructions")
Reported-by: Johan Almbladh <johan.almbladh@anyfinetworks.com>
Signed-off-by: Johan Almbladh <johan.almbladh@anyfinetworks.com>
Co-developed-by: Daniel Borkmann <daniel@iogearbox.net>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Reviewed-by: Brendan Jackman <jackmanb@google.com>
Acked-by: Alexei Starovoitov <ast@kernel.org>
2021-09-27 13:11:57 +00:00
|
|
|
BPF_REG_0, real_dst_reg, insn->off);
|
2021-01-14 18:17:49 +00:00
|
|
|
/*
|
|
|
|
* Perform the (commutative) operation locally,
|
|
|
|
* put the result in the AUX_REG.
|
|
|
|
*/
|
|
|
|
emit_mov_reg(&prog, is64, AUX_REG, BPF_REG_0);
|
2021-02-16 12:53:07 +00:00
|
|
|
maybe_emit_mod(&prog, AUX_REG, real_src_reg, is64);
|
2021-01-14 18:17:49 +00:00
|
|
|
EMIT2(simple_alu_opcodes[BPF_OP(insn->imm)],
|
2021-02-16 12:53:07 +00:00
|
|
|
add_2reg(0xC0, AUX_REG, real_src_reg));
|
2021-01-14 18:17:49 +00:00
|
|
|
/* Attempt to swap in new value */
|
|
|
|
err = emit_atomic(&prog, BPF_CMPXCHG,
|
bpf, x86: Fix bpf mapping of atomic fetch implementation
Fix the case where the dst register maps to %rax as otherwise this produces
an incorrect mapping with the implementation in 981f94c3e921 ("bpf: Add
bitwise atomic instructions") as %rax is clobbered given it's part of the
cmpxchg as operand.
The issue is similar to b29dd96b905f ("bpf, x86: Fix BPF_FETCH atomic and/or/
xor with r0 as src") just that the case of dst register was missed.
Before, dst=r0 (%rax) src=r2 (%rsi):
[...]
c5: mov %rax,%r10
c8: mov 0x0(%rax),%rax <---+ (broken)
cc: mov %rax,%r11 |
cf: and %rsi,%r11 |
d2: lock cmpxchg %r11,0x0(%rax) <---+
d8: jne 0x00000000000000c8 |
da: mov %rax,%rsi |
dd: mov %r10,%rax |
[...] |
|
After, dst=r0 (%rax) src=r2 (%rsi): |
|
[...] |
da: mov %rax,%r10 |
dd: mov 0x0(%r10),%rax <---+ (fixed)
e1: mov %rax,%r11 |
e4: and %rsi,%r11 |
e7: lock cmpxchg %r11,0x0(%r10) <---+
ed: jne 0x00000000000000dd
ef: mov %rax,%rsi
f2: mov %r10,%rax
[...]
The remaining combinations were fine as-is though:
After, dst=r9 (%r15) src=r0 (%rax):
[...]
dc: mov %rax,%r10
df: mov 0x0(%r15),%rax
e3: mov %rax,%r11
e6: and %r10,%r11
e9: lock cmpxchg %r11,0x0(%r15)
ef: jne 0x00000000000000df _
f1: mov %rax,%r10 | (unneeded, but
f4: mov %r10,%rax _| not a problem)
[...]
After, dst=r9 (%r15) src=r4 (%rcx):
[...]
de: mov %rax,%r10
e1: mov 0x0(%r15),%rax
e5: mov %rax,%r11
e8: and %rcx,%r11
eb: lock cmpxchg %r11,0x0(%r15)
f1: jne 0x00000000000000e1
f3: mov %rax,%rcx
f6: mov %r10,%rax
[...]
The case of dst == src register is rejected by the verifier and
therefore not supported, but x86 JIT also handles this case just
fine.
After, dst=r0 (%rax) src=r0 (%rax):
[...]
eb: mov %rax,%r10
ee: mov 0x0(%r10),%rax
f2: mov %rax,%r11
f5: and %r10,%r11
f8: lock cmpxchg %r11,0x0(%r10)
fe: jne 0x00000000000000ee
100: mov %rax,%r10
103: mov %r10,%rax
[...]
Fixes: 981f94c3e921 ("bpf: Add bitwise atomic instructions")
Reported-by: Johan Almbladh <johan.almbladh@anyfinetworks.com>
Signed-off-by: Johan Almbladh <johan.almbladh@anyfinetworks.com>
Co-developed-by: Daniel Borkmann <daniel@iogearbox.net>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Reviewed-by: Brendan Jackman <jackmanb@google.com>
Acked-by: Alexei Starovoitov <ast@kernel.org>
2021-09-27 13:11:57 +00:00
|
|
|
real_dst_reg, AUX_REG,
|
|
|
|
insn->off,
|
2021-01-14 18:17:49 +00:00
|
|
|
BPF_SIZE(insn->code));
|
|
|
|
if (WARN_ON(err))
|
|
|
|
return err;
|
|
|
|
/*
|
|
|
|
* ZF tells us whether we won the race. If it's
|
|
|
|
* cleared we need to try again.
|
|
|
|
*/
|
|
|
|
EMIT2(X86_JNE, -(prog - branch_target) - 2);
|
|
|
|
/* Return the pre-modification value */
|
2021-02-16 12:53:07 +00:00
|
|
|
emit_mov_reg(&prog, is64, real_src_reg, BPF_REG_0);
|
2021-01-14 18:17:49 +00:00
|
|
|
/* Restore R0 after clobbering RAX */
|
|
|
|
emit_mov_reg(&prog, true, BPF_REG_0, BPF_REG_AX);
|
|
|
|
break;
|
|
|
|
}
|
|
|
|
|
2021-01-14 18:17:44 +00:00
|
|
|
err = emit_atomic(&prog, insn->imm, dst_reg, src_reg,
|
bpf, x86: Fix bpf mapping of atomic fetch implementation
Fix the case where the dst register maps to %rax as otherwise this produces
an incorrect mapping with the implementation in 981f94c3e921 ("bpf: Add
bitwise atomic instructions") as %rax is clobbered given it's part of the
cmpxchg as operand.
The issue is similar to b29dd96b905f ("bpf, x86: Fix BPF_FETCH atomic and/or/
xor with r0 as src") just that the case of dst register was missed.
Before, dst=r0 (%rax) src=r2 (%rsi):
[...]
c5: mov %rax,%r10
c8: mov 0x0(%rax),%rax <---+ (broken)
cc: mov %rax,%r11 |
cf: and %rsi,%r11 |
d2: lock cmpxchg %r11,0x0(%rax) <---+
d8: jne 0x00000000000000c8 |
da: mov %rax,%rsi |
dd: mov %r10,%rax |
[...] |
|
After, dst=r0 (%rax) src=r2 (%rsi): |
|
[...] |
da: mov %rax,%r10 |
dd: mov 0x0(%r10),%rax <---+ (fixed)
e1: mov %rax,%r11 |
e4: and %rsi,%r11 |
e7: lock cmpxchg %r11,0x0(%r10) <---+
ed: jne 0x00000000000000dd
ef: mov %rax,%rsi
f2: mov %r10,%rax
[...]
The remaining combinations were fine as-is though:
After, dst=r9 (%r15) src=r0 (%rax):
[...]
dc: mov %rax,%r10
df: mov 0x0(%r15),%rax
e3: mov %rax,%r11
e6: and %r10,%r11
e9: lock cmpxchg %r11,0x0(%r15)
ef: jne 0x00000000000000df _
f1: mov %rax,%r10 | (unneeded, but
f4: mov %r10,%rax _| not a problem)
[...]
After, dst=r9 (%r15) src=r4 (%rcx):
[...]
de: mov %rax,%r10
e1: mov 0x0(%r15),%rax
e5: mov %rax,%r11
e8: and %rcx,%r11
eb: lock cmpxchg %r11,0x0(%r15)
f1: jne 0x00000000000000e1
f3: mov %rax,%rcx
f6: mov %r10,%rax
[...]
The case of dst == src register is rejected by the verifier and
therefore not supported, but x86 JIT also handles this case just
fine.
After, dst=r0 (%rax) src=r0 (%rax):
[...]
eb: mov %rax,%r10
ee: mov 0x0(%r10),%rax
f2: mov %rax,%r11
f5: and %r10,%r11
f8: lock cmpxchg %r11,0x0(%r10)
fe: jne 0x00000000000000ee
100: mov %rax,%r10
103: mov %r10,%rax
[...]
Fixes: 981f94c3e921 ("bpf: Add bitwise atomic instructions")
Reported-by: Johan Almbladh <johan.almbladh@anyfinetworks.com>
Signed-off-by: Johan Almbladh <johan.almbladh@anyfinetworks.com>
Co-developed-by: Daniel Borkmann <daniel@iogearbox.net>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Reviewed-by: Brendan Jackman <jackmanb@google.com>
Acked-by: Alexei Starovoitov <ast@kernel.org>
2021-09-27 13:11:57 +00:00
|
|
|
insn->off, BPF_SIZE(insn->code));
|
2021-01-14 18:17:44 +00:00
|
|
|
if (err)
|
|
|
|
return err;
|
net: filter: x86: internal BPF JIT
Maps all internal BPF instructions into x86_64 instructions.
This patch replaces original BPF x64 JIT with internal BPF x64 JIT.
sysctl net.core.bpf_jit_enable is reused as on/off switch.
Performance:
1. old BPF JIT and internal BPF JIT generate equivalent x86_64 code.
No performance difference is observed for filters that were JIT-able before
Example assembler code for BPF filter "tcpdump port 22"
original BPF -> old JIT: original BPF -> internal BPF -> new JIT:
0: push %rbp 0: push %rbp
1: mov %rsp,%rbp 1: mov %rsp,%rbp
4: sub $0x60,%rsp 4: sub $0x228,%rsp
8: mov %rbx,-0x8(%rbp) b: mov %rbx,-0x228(%rbp) // prologue
12: mov %r13,-0x220(%rbp)
19: mov %r14,-0x218(%rbp)
20: mov %r15,-0x210(%rbp)
27: xor %eax,%eax // clear A
c: xor %ebx,%ebx 29: xor %r13,%r13 // clear X
e: mov 0x68(%rdi),%r9d 2c: mov 0x68(%rdi),%r9d
12: sub 0x6c(%rdi),%r9d 30: sub 0x6c(%rdi),%r9d
16: mov 0xd8(%rdi),%r8 34: mov 0xd8(%rdi),%r10
3b: mov %rdi,%rbx
1d: mov $0xc,%esi 3e: mov $0xc,%esi
22: callq 0xffffffffe1021e15 43: callq 0xffffffffe102bd75
27: cmp $0x86dd,%eax 48: cmp $0x86dd,%rax
2c: jne 0x0000000000000069 4f: jne 0x000000000000009a
2e: mov $0x14,%esi 51: mov $0x14,%esi
33: callq 0xffffffffe1021e31 56: callq 0xffffffffe102bd91
38: cmp $0x84,%eax 5b: cmp $0x84,%rax
3d: je 0x0000000000000049 62: je 0x0000000000000074
3f: cmp $0x6,%eax 64: cmp $0x6,%rax
42: je 0x0000000000000049 68: je 0x0000000000000074
44: cmp $0x11,%eax 6a: cmp $0x11,%rax
47: jne 0x00000000000000c6 6e: jne 0x0000000000000117
49: mov $0x36,%esi 74: mov $0x36,%esi
4e: callq 0xffffffffe1021e15 79: callq 0xffffffffe102bd75
53: cmp $0x16,%eax 7e: cmp $0x16,%rax
56: je 0x00000000000000bf 82: je 0x0000000000000110
58: mov $0x38,%esi 88: mov $0x38,%esi
5d: callq 0xffffffffe1021e15 8d: callq 0xffffffffe102bd75
62: cmp $0x16,%eax 92: cmp $0x16,%rax
65: je 0x00000000000000bf 96: je 0x0000000000000110
67: jmp 0x00000000000000c6 98: jmp 0x0000000000000117
69: cmp $0x800,%eax 9a: cmp $0x800,%rax
6e: jne 0x00000000000000c6 a1: jne 0x0000000000000117
70: mov $0x17,%esi a3: mov $0x17,%esi
75: callq 0xffffffffe1021e31 a8: callq 0xffffffffe102bd91
7a: cmp $0x84,%eax ad: cmp $0x84,%rax
7f: je 0x000000000000008b b4: je 0x00000000000000c2
81: cmp $0x6,%eax b6: cmp $0x6,%rax
84: je 0x000000000000008b ba: je 0x00000000000000c2
86: cmp $0x11,%eax bc: cmp $0x11,%rax
89: jne 0x00000000000000c6 c0: jne 0x0000000000000117
8b: mov $0x14,%esi c2: mov $0x14,%esi
90: callq 0xffffffffe1021e15 c7: callq 0xffffffffe102bd75
95: test $0x1fff,%ax cc: test $0x1fff,%rax
99: jne 0x00000000000000c6 d3: jne 0x0000000000000117
d5: mov %rax,%r14
9b: mov $0xe,%esi d8: mov $0xe,%esi
a0: callq 0xffffffffe1021e44 dd: callq 0xffffffffe102bd91 // MSH
e2: and $0xf,%eax
e5: shl $0x2,%eax
e8: mov %rax,%r13
eb: mov %r14,%rax
ee: mov %r13,%rsi
a5: lea 0xe(%rbx),%esi f1: add $0xe,%esi
a8: callq 0xffffffffe1021e0d f4: callq 0xffffffffe102bd6d
ad: cmp $0x16,%eax f9: cmp $0x16,%rax
b0: je 0x00000000000000bf fd: je 0x0000000000000110
ff: mov %r13,%rsi
b2: lea 0x10(%rbx),%esi 102: add $0x10,%esi
b5: callq 0xffffffffe1021e0d 105: callq 0xffffffffe102bd6d
ba: cmp $0x16,%eax 10a: cmp $0x16,%rax
bd: jne 0x00000000000000c6 10e: jne 0x0000000000000117
bf: mov $0xffff,%eax 110: mov $0xffff,%eax
c4: jmp 0x00000000000000c8 115: jmp 0x000000000000011c
c6: xor %eax,%eax 117: mov $0x0,%eax
c8: mov -0x8(%rbp),%rbx 11c: mov -0x228(%rbp),%rbx // epilogue
cc: leaveq 123: mov -0x220(%rbp),%r13
cd: retq 12a: mov -0x218(%rbp),%r14
131: mov -0x210(%rbp),%r15
138: leaveq
139: retq
On fully cached SKBs both JITed functions take 12 nsec to execute.
BPF interpreter executes the program in 30 nsec.
The difference in generated assembler is due to the following:
Old BPF imlements LDX_MSH instruction via sk_load_byte_msh() helper function
inside bpf_jit.S.
New JIT removes the helper and does it explicitly, so ldx_msh cost
is the same for both JITs, but generated code looks longer.
New JIT has 4 registers to save, so prologue/epilogue are larger,
but the cost is within noise on x64.
Old JIT checks whether first insn clears A and if not emits 'xor %eax,%eax'.
New JIT clears %rax unconditionally.
2. old BPF JIT doesn't support ANC_NLATTR, ANC_PAY_OFFSET, ANC_RANDOM
extensions. New JIT supports all BPF extensions.
Performance of such filters improves 2-4 times depending on a filter.
The longer the filter the higher performance gain.
Synthetic benchmarks with many ancillary loads see 20x speedup
which seems to be the maximum gain from JIT
Notes:
. net.core.bpf_jit_enable=2 + tools/net/bpf_jit_disasm is still functional
and can be used to see generated assembler
. there are two jit_compile() functions and code flow for classic filters is:
sk_attach_filter() - load classic BPF
bpf_jit_compile() - try to JIT from classic BPF
sk_convert_filter() - convert classic to internal
bpf_int_jit_compile() - JIT from internal BPF
seccomp and tracing filters will just call bpf_int_jit_compile()
Signed-off-by: Alexei Starovoitov <ast@plumgrid.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
2014-05-14 02:50:46 +00:00
|
|
|
break;
|
|
|
|
|
|
|
|
/* call */
|
2022-09-15 11:11:34 +00:00
|
|
|
case BPF_JMP | BPF_CALL: {
|
2024-04-01 18:55:30 +00:00
|
|
|
u8 *ip = image + addrs[i - 1];
|
2022-09-15 11:11:34 +00:00
|
|
|
|
2014-06-06 21:46:06 +00:00
|
|
|
func = (u8 *) __bpf_call_base + imm32;
|
bpf, x64: rework pro/epilogue and tailcall handling in JIT
This commit serves two things:
1) it optimizes BPF prologue/epilogue generation
2) it makes possible to have tailcalls within BPF subprogram
Both points are related to each other since without 1), 2) could not be
achieved.
In [1], Alexei says:
"The prologue will look like:
nop5
xor eax,eax // two new bytes if bpf_tail_call() is used in this
// function
push rbp
mov rbp, rsp
sub rsp, rounded_stack_depth
push rax // zero init tail_call counter
variable number of push rbx,r13,r14,r15
Then bpf_tail_call will pop variable number rbx,..
and final 'pop rax'
Then 'add rsp, size_of_current_stack_frame'
jmp to next function and skip over 'nop5; xor eax,eax; push rpb; mov
rbp, rsp'
This way new function will set its own stack size and will init tail
call
counter with whatever value the parent had.
If next function doesn't use bpf_tail_call it won't have 'xor eax,eax'.
Instead it would need to have 'nop2' in there."
Implement that suggestion.
Since the layout of stack is changed, tail call counter handling can not
rely anymore on popping it to rbx just like it have been handled for
constant prologue case and later overwrite of rbx with actual value of
rbx pushed to stack. Therefore, let's use one of the register (%rcx) that
is considered to be volatile/caller-saved and pop the value of tail call
counter in there in the epilogue.
Drop the BUILD_BUG_ON in emit_prologue and in
emit_bpf_tail_call_indirect where instruction layout is not constant
anymore.
Introduce new poke target, 'tailcall_bypass' to poke descriptor that is
dedicated for skipping the register pops and stack unwind that are
generated right before the actual jump to target program.
For case when the target program is not present, BPF program will skip
the pop instructions and nop5 dedicated for jmpq $target. An example of
such state when only R6 of callee saved registers is used by program:
ffffffffc0513aa1: e9 0e 00 00 00 jmpq 0xffffffffc0513ab4
ffffffffc0513aa6: 5b pop %rbx
ffffffffc0513aa7: 58 pop %rax
ffffffffc0513aa8: 48 81 c4 00 00 00 00 add $0x0,%rsp
ffffffffc0513aaf: 0f 1f 44 00 00 nopl 0x0(%rax,%rax,1)
ffffffffc0513ab4: 48 89 df mov %rbx,%rdi
When target program is inserted, the jump that was there to skip
pops/nop5 will become the nop5, so CPU will go over pops and do the
actual tailcall.
One might ask why there simply can not be pushes after the nop5?
In the following example snippet:
ffffffffc037030c: 48 89 fb mov %rdi,%rbx
(...)
ffffffffc0370332: 5b pop %rbx
ffffffffc0370333: 58 pop %rax
ffffffffc0370334: 48 81 c4 00 00 00 00 add $0x0,%rsp
ffffffffc037033b: 0f 1f 44 00 00 nopl 0x0(%rax,%rax,1)
ffffffffc0370340: 48 81 ec 00 00 00 00 sub $0x0,%rsp
ffffffffc0370347: 50 push %rax
ffffffffc0370348: 53 push %rbx
ffffffffc0370349: 48 89 df mov %rbx,%rdi
ffffffffc037034c: e8 f7 21 00 00 callq 0xffffffffc0372548
There is the bpf2bpf call (at ffffffffc037034c) right after the tailcall
and jump target is not present. ctx is in %rbx register and BPF
subprogram that we will call into on ffffffffc037034c is relying on it,
e.g. it will pick ctx from there. Such code layout is therefore broken
as we would overwrite the content of %rbx with the value that was pushed
on the prologue. That is the reason for the 'bypass' approach.
Special care needs to be taken during the install/update/remove of
tailcall target. In case when target program is not present, the CPU
must not execute the pop instructions that precede the tailcall.
To address that, the following states can be defined:
A nop, unwind, nop
B nop, unwind, tail
C skip, unwind, nop
D skip, unwind, tail
A is forbidden (lead to incorrectness). The state transitions between
tailcall install/update/remove will work as follows:
First install tail call f: C->D->B(f)
* poke the tailcall, after that get rid of the skip
Update tail call f to f': B(f)->B(f')
* poke the tailcall (poke->tailcall_target) and do NOT touch the
poke->tailcall_bypass
Remove tail call: B(f')->C(f')
* poke->tailcall_bypass is poked back to jump, then we wait the RCU
grace period so that other programs will finish its execution and
after that we are safe to remove the poke->tailcall_target
Install new tail call (f''): C(f')->D(f'')->B(f'').
* same as first step
This way CPU can never be exposed to "unwind, tail" state.
Last but not least, when tailcalls get mixed with bpf2bpf calls, it
would be possible to encounter the endless loop due to clearing the
tailcall counter if for example we would use the tailcall3-like from BPF
selftests program that would be subprogram-based, meaning the tailcall
would be present within the BPF subprogram.
This test, broken down to particular steps, would do:
entry -> set tailcall counter to 0, bump it by 1, tailcall to func0
func0 -> call subprog_tail
(we are NOT skipping the first 11 bytes of prologue and this subprogram
has a tailcall, therefore we clear the counter...)
subprog -> do the same thing as entry
and then loop forever.
To address this, the idea is to go through the call chain of bpf2bpf progs
and look for a tailcall presence throughout whole chain. If we saw a single
tail call then each node in this call chain needs to be marked as a subprog
that can reach the tailcall. We would later feed the JIT with this info
and:
- set eax to 0 only when tailcall is reachable and this is the entry prog
- if tailcall is reachable but there's no tailcall in insns of currently
JITed prog then push rax anyway, so that it will be possible to
propagate further down the call chain
- finally if tailcall is reachable, then we need to precede the 'call'
insn with mov rax, [rbp - (stack_depth + 8)]
Tail call related cases from test_verifier kselftest are also working
fine. Sample BPF programs that utilize tail calls (sockex3, tracex5)
work properly as well.
[1]: https://lore.kernel.org/bpf/20200517043227.2gpq22ifoq37ogst@ast-mbp.dhcp.thefacebook.com/
Suggested-by: Alexei Starovoitov <ast@kernel.org>
Signed-off-by: Maciej Fijalkowski <maciej.fijalkowski@intel.com>
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
2020-09-16 21:10:08 +00:00
|
|
|
if (tail_call_reachable) {
|
2023-09-12 15:04:41 +00:00
|
|
|
RESTORE_TAIL_CALL_CNT(bpf_prog->aux->stack_depth);
|
2024-04-01 18:55:30 +00:00
|
|
|
ip += 7;
|
bpf, x64: rework pro/epilogue and tailcall handling in JIT
This commit serves two things:
1) it optimizes BPF prologue/epilogue generation
2) it makes possible to have tailcalls within BPF subprogram
Both points are related to each other since without 1), 2) could not be
achieved.
In [1], Alexei says:
"The prologue will look like:
nop5
xor eax,eax // two new bytes if bpf_tail_call() is used in this
// function
push rbp
mov rbp, rsp
sub rsp, rounded_stack_depth
push rax // zero init tail_call counter
variable number of push rbx,r13,r14,r15
Then bpf_tail_call will pop variable number rbx,..
and final 'pop rax'
Then 'add rsp, size_of_current_stack_frame'
jmp to next function and skip over 'nop5; xor eax,eax; push rpb; mov
rbp, rsp'
This way new function will set its own stack size and will init tail
call
counter with whatever value the parent had.
If next function doesn't use bpf_tail_call it won't have 'xor eax,eax'.
Instead it would need to have 'nop2' in there."
Implement that suggestion.
Since the layout of stack is changed, tail call counter handling can not
rely anymore on popping it to rbx just like it have been handled for
constant prologue case and later overwrite of rbx with actual value of
rbx pushed to stack. Therefore, let's use one of the register (%rcx) that
is considered to be volatile/caller-saved and pop the value of tail call
counter in there in the epilogue.
Drop the BUILD_BUG_ON in emit_prologue and in
emit_bpf_tail_call_indirect where instruction layout is not constant
anymore.
Introduce new poke target, 'tailcall_bypass' to poke descriptor that is
dedicated for skipping the register pops and stack unwind that are
generated right before the actual jump to target program.
For case when the target program is not present, BPF program will skip
the pop instructions and nop5 dedicated for jmpq $target. An example of
such state when only R6 of callee saved registers is used by program:
ffffffffc0513aa1: e9 0e 00 00 00 jmpq 0xffffffffc0513ab4
ffffffffc0513aa6: 5b pop %rbx
ffffffffc0513aa7: 58 pop %rax
ffffffffc0513aa8: 48 81 c4 00 00 00 00 add $0x0,%rsp
ffffffffc0513aaf: 0f 1f 44 00 00 nopl 0x0(%rax,%rax,1)
ffffffffc0513ab4: 48 89 df mov %rbx,%rdi
When target program is inserted, the jump that was there to skip
pops/nop5 will become the nop5, so CPU will go over pops and do the
actual tailcall.
One might ask why there simply can not be pushes after the nop5?
In the following example snippet:
ffffffffc037030c: 48 89 fb mov %rdi,%rbx
(...)
ffffffffc0370332: 5b pop %rbx
ffffffffc0370333: 58 pop %rax
ffffffffc0370334: 48 81 c4 00 00 00 00 add $0x0,%rsp
ffffffffc037033b: 0f 1f 44 00 00 nopl 0x0(%rax,%rax,1)
ffffffffc0370340: 48 81 ec 00 00 00 00 sub $0x0,%rsp
ffffffffc0370347: 50 push %rax
ffffffffc0370348: 53 push %rbx
ffffffffc0370349: 48 89 df mov %rbx,%rdi
ffffffffc037034c: e8 f7 21 00 00 callq 0xffffffffc0372548
There is the bpf2bpf call (at ffffffffc037034c) right after the tailcall
and jump target is not present. ctx is in %rbx register and BPF
subprogram that we will call into on ffffffffc037034c is relying on it,
e.g. it will pick ctx from there. Such code layout is therefore broken
as we would overwrite the content of %rbx with the value that was pushed
on the prologue. That is the reason for the 'bypass' approach.
Special care needs to be taken during the install/update/remove of
tailcall target. In case when target program is not present, the CPU
must not execute the pop instructions that precede the tailcall.
To address that, the following states can be defined:
A nop, unwind, nop
B nop, unwind, tail
C skip, unwind, nop
D skip, unwind, tail
A is forbidden (lead to incorrectness). The state transitions between
tailcall install/update/remove will work as follows:
First install tail call f: C->D->B(f)
* poke the tailcall, after that get rid of the skip
Update tail call f to f': B(f)->B(f')
* poke the tailcall (poke->tailcall_target) and do NOT touch the
poke->tailcall_bypass
Remove tail call: B(f')->C(f')
* poke->tailcall_bypass is poked back to jump, then we wait the RCU
grace period so that other programs will finish its execution and
after that we are safe to remove the poke->tailcall_target
Install new tail call (f''): C(f')->D(f'')->B(f'').
* same as first step
This way CPU can never be exposed to "unwind, tail" state.
Last but not least, when tailcalls get mixed with bpf2bpf calls, it
would be possible to encounter the endless loop due to clearing the
tailcall counter if for example we would use the tailcall3-like from BPF
selftests program that would be subprogram-based, meaning the tailcall
would be present within the BPF subprogram.
This test, broken down to particular steps, would do:
entry -> set tailcall counter to 0, bump it by 1, tailcall to func0
func0 -> call subprog_tail
(we are NOT skipping the first 11 bytes of prologue and this subprogram
has a tailcall, therefore we clear the counter...)
subprog -> do the same thing as entry
and then loop forever.
To address this, the idea is to go through the call chain of bpf2bpf progs
and look for a tailcall presence throughout whole chain. If we saw a single
tail call then each node in this call chain needs to be marked as a subprog
that can reach the tailcall. We would later feed the JIT with this info
and:
- set eax to 0 only when tailcall is reachable and this is the entry prog
- if tailcall is reachable but there's no tailcall in insns of currently
JITed prog then push rax anyway, so that it will be possible to
propagate further down the call chain
- finally if tailcall is reachable, then we need to precede the 'call'
insn with mov rax, [rbp - (stack_depth + 8)]
Tail call related cases from test_verifier kselftest are also working
fine. Sample BPF programs that utilize tail calls (sockex3, tracex5)
work properly as well.
[1]: https://lore.kernel.org/bpf/20200517043227.2gpq22ifoq37ogst@ast-mbp.dhcp.thefacebook.com/
Suggested-by: Alexei Starovoitov <ast@kernel.org>
Signed-off-by: Maciej Fijalkowski <maciej.fijalkowski@intel.com>
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
2020-09-16 21:10:08 +00:00
|
|
|
}
|
2024-04-01 18:55:30 +00:00
|
|
|
if (!imm32)
|
|
|
|
return -EINVAL;
|
|
|
|
ip += x86_call_depth_emit_accounting(&prog, func, ip);
|
|
|
|
if (emit_call(&prog, func, ip))
|
2022-09-15 11:11:34 +00:00
|
|
|
return -EINVAL;
|
net: filter: x86: internal BPF JIT
Maps all internal BPF instructions into x86_64 instructions.
This patch replaces original BPF x64 JIT with internal BPF x64 JIT.
sysctl net.core.bpf_jit_enable is reused as on/off switch.
Performance:
1. old BPF JIT and internal BPF JIT generate equivalent x86_64 code.
No performance difference is observed for filters that were JIT-able before
Example assembler code for BPF filter "tcpdump port 22"
original BPF -> old JIT: original BPF -> internal BPF -> new JIT:
0: push %rbp 0: push %rbp
1: mov %rsp,%rbp 1: mov %rsp,%rbp
4: sub $0x60,%rsp 4: sub $0x228,%rsp
8: mov %rbx,-0x8(%rbp) b: mov %rbx,-0x228(%rbp) // prologue
12: mov %r13,-0x220(%rbp)
19: mov %r14,-0x218(%rbp)
20: mov %r15,-0x210(%rbp)
27: xor %eax,%eax // clear A
c: xor %ebx,%ebx 29: xor %r13,%r13 // clear X
e: mov 0x68(%rdi),%r9d 2c: mov 0x68(%rdi),%r9d
12: sub 0x6c(%rdi),%r9d 30: sub 0x6c(%rdi),%r9d
16: mov 0xd8(%rdi),%r8 34: mov 0xd8(%rdi),%r10
3b: mov %rdi,%rbx
1d: mov $0xc,%esi 3e: mov $0xc,%esi
22: callq 0xffffffffe1021e15 43: callq 0xffffffffe102bd75
27: cmp $0x86dd,%eax 48: cmp $0x86dd,%rax
2c: jne 0x0000000000000069 4f: jne 0x000000000000009a
2e: mov $0x14,%esi 51: mov $0x14,%esi
33: callq 0xffffffffe1021e31 56: callq 0xffffffffe102bd91
38: cmp $0x84,%eax 5b: cmp $0x84,%rax
3d: je 0x0000000000000049 62: je 0x0000000000000074
3f: cmp $0x6,%eax 64: cmp $0x6,%rax
42: je 0x0000000000000049 68: je 0x0000000000000074
44: cmp $0x11,%eax 6a: cmp $0x11,%rax
47: jne 0x00000000000000c6 6e: jne 0x0000000000000117
49: mov $0x36,%esi 74: mov $0x36,%esi
4e: callq 0xffffffffe1021e15 79: callq 0xffffffffe102bd75
53: cmp $0x16,%eax 7e: cmp $0x16,%rax
56: je 0x00000000000000bf 82: je 0x0000000000000110
58: mov $0x38,%esi 88: mov $0x38,%esi
5d: callq 0xffffffffe1021e15 8d: callq 0xffffffffe102bd75
62: cmp $0x16,%eax 92: cmp $0x16,%rax
65: je 0x00000000000000bf 96: je 0x0000000000000110
67: jmp 0x00000000000000c6 98: jmp 0x0000000000000117
69: cmp $0x800,%eax 9a: cmp $0x800,%rax
6e: jne 0x00000000000000c6 a1: jne 0x0000000000000117
70: mov $0x17,%esi a3: mov $0x17,%esi
75: callq 0xffffffffe1021e31 a8: callq 0xffffffffe102bd91
7a: cmp $0x84,%eax ad: cmp $0x84,%rax
7f: je 0x000000000000008b b4: je 0x00000000000000c2
81: cmp $0x6,%eax b6: cmp $0x6,%rax
84: je 0x000000000000008b ba: je 0x00000000000000c2
86: cmp $0x11,%eax bc: cmp $0x11,%rax
89: jne 0x00000000000000c6 c0: jne 0x0000000000000117
8b: mov $0x14,%esi c2: mov $0x14,%esi
90: callq 0xffffffffe1021e15 c7: callq 0xffffffffe102bd75
95: test $0x1fff,%ax cc: test $0x1fff,%rax
99: jne 0x00000000000000c6 d3: jne 0x0000000000000117
d5: mov %rax,%r14
9b: mov $0xe,%esi d8: mov $0xe,%esi
a0: callq 0xffffffffe1021e44 dd: callq 0xffffffffe102bd91 // MSH
e2: and $0xf,%eax
e5: shl $0x2,%eax
e8: mov %rax,%r13
eb: mov %r14,%rax
ee: mov %r13,%rsi
a5: lea 0xe(%rbx),%esi f1: add $0xe,%esi
a8: callq 0xffffffffe1021e0d f4: callq 0xffffffffe102bd6d
ad: cmp $0x16,%eax f9: cmp $0x16,%rax
b0: je 0x00000000000000bf fd: je 0x0000000000000110
ff: mov %r13,%rsi
b2: lea 0x10(%rbx),%esi 102: add $0x10,%esi
b5: callq 0xffffffffe1021e0d 105: callq 0xffffffffe102bd6d
ba: cmp $0x16,%eax 10a: cmp $0x16,%rax
bd: jne 0x00000000000000c6 10e: jne 0x0000000000000117
bf: mov $0xffff,%eax 110: mov $0xffff,%eax
c4: jmp 0x00000000000000c8 115: jmp 0x000000000000011c
c6: xor %eax,%eax 117: mov $0x0,%eax
c8: mov -0x8(%rbp),%rbx 11c: mov -0x228(%rbp),%rbx // epilogue
cc: leaveq 123: mov -0x220(%rbp),%r13
cd: retq 12a: mov -0x218(%rbp),%r14
131: mov -0x210(%rbp),%r15
138: leaveq
139: retq
On fully cached SKBs both JITed functions take 12 nsec to execute.
BPF interpreter executes the program in 30 nsec.
The difference in generated assembler is due to the following:
Old BPF imlements LDX_MSH instruction via sk_load_byte_msh() helper function
inside bpf_jit.S.
New JIT removes the helper and does it explicitly, so ldx_msh cost
is the same for both JITs, but generated code looks longer.
New JIT has 4 registers to save, so prologue/epilogue are larger,
but the cost is within noise on x64.
Old JIT checks whether first insn clears A and if not emits 'xor %eax,%eax'.
New JIT clears %rax unconditionally.
2. old BPF JIT doesn't support ANC_NLATTR, ANC_PAY_OFFSET, ANC_RANDOM
extensions. New JIT supports all BPF extensions.
Performance of such filters improves 2-4 times depending on a filter.
The longer the filter the higher performance gain.
Synthetic benchmarks with many ancillary loads see 20x speedup
which seems to be the maximum gain from JIT
Notes:
. net.core.bpf_jit_enable=2 + tools/net/bpf_jit_disasm is still functional
and can be used to see generated assembler
. there are two jit_compile() functions and code flow for classic filters is:
sk_attach_filter() - load classic BPF
bpf_jit_compile() - try to JIT from classic BPF
sk_convert_filter() - convert classic to internal
bpf_int_jit_compile() - JIT from internal BPF
seccomp and tracing filters will just call bpf_int_jit_compile()
Signed-off-by: Alexei Starovoitov <ast@plumgrid.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
2014-05-14 02:50:46 +00:00
|
|
|
break;
|
2022-09-15 11:11:34 +00:00
|
|
|
}
|
net: filter: x86: internal BPF JIT
Maps all internal BPF instructions into x86_64 instructions.
This patch replaces original BPF x64 JIT with internal BPF x64 JIT.
sysctl net.core.bpf_jit_enable is reused as on/off switch.
Performance:
1. old BPF JIT and internal BPF JIT generate equivalent x86_64 code.
No performance difference is observed for filters that were JIT-able before
Example assembler code for BPF filter "tcpdump port 22"
original BPF -> old JIT: original BPF -> internal BPF -> new JIT:
0: push %rbp 0: push %rbp
1: mov %rsp,%rbp 1: mov %rsp,%rbp
4: sub $0x60,%rsp 4: sub $0x228,%rsp
8: mov %rbx,-0x8(%rbp) b: mov %rbx,-0x228(%rbp) // prologue
12: mov %r13,-0x220(%rbp)
19: mov %r14,-0x218(%rbp)
20: mov %r15,-0x210(%rbp)
27: xor %eax,%eax // clear A
c: xor %ebx,%ebx 29: xor %r13,%r13 // clear X
e: mov 0x68(%rdi),%r9d 2c: mov 0x68(%rdi),%r9d
12: sub 0x6c(%rdi),%r9d 30: sub 0x6c(%rdi),%r9d
16: mov 0xd8(%rdi),%r8 34: mov 0xd8(%rdi),%r10
3b: mov %rdi,%rbx
1d: mov $0xc,%esi 3e: mov $0xc,%esi
22: callq 0xffffffffe1021e15 43: callq 0xffffffffe102bd75
27: cmp $0x86dd,%eax 48: cmp $0x86dd,%rax
2c: jne 0x0000000000000069 4f: jne 0x000000000000009a
2e: mov $0x14,%esi 51: mov $0x14,%esi
33: callq 0xffffffffe1021e31 56: callq 0xffffffffe102bd91
38: cmp $0x84,%eax 5b: cmp $0x84,%rax
3d: je 0x0000000000000049 62: je 0x0000000000000074
3f: cmp $0x6,%eax 64: cmp $0x6,%rax
42: je 0x0000000000000049 68: je 0x0000000000000074
44: cmp $0x11,%eax 6a: cmp $0x11,%rax
47: jne 0x00000000000000c6 6e: jne 0x0000000000000117
49: mov $0x36,%esi 74: mov $0x36,%esi
4e: callq 0xffffffffe1021e15 79: callq 0xffffffffe102bd75
53: cmp $0x16,%eax 7e: cmp $0x16,%rax
56: je 0x00000000000000bf 82: je 0x0000000000000110
58: mov $0x38,%esi 88: mov $0x38,%esi
5d: callq 0xffffffffe1021e15 8d: callq 0xffffffffe102bd75
62: cmp $0x16,%eax 92: cmp $0x16,%rax
65: je 0x00000000000000bf 96: je 0x0000000000000110
67: jmp 0x00000000000000c6 98: jmp 0x0000000000000117
69: cmp $0x800,%eax 9a: cmp $0x800,%rax
6e: jne 0x00000000000000c6 a1: jne 0x0000000000000117
70: mov $0x17,%esi a3: mov $0x17,%esi
75: callq 0xffffffffe1021e31 a8: callq 0xffffffffe102bd91
7a: cmp $0x84,%eax ad: cmp $0x84,%rax
7f: je 0x000000000000008b b4: je 0x00000000000000c2
81: cmp $0x6,%eax b6: cmp $0x6,%rax
84: je 0x000000000000008b ba: je 0x00000000000000c2
86: cmp $0x11,%eax bc: cmp $0x11,%rax
89: jne 0x00000000000000c6 c0: jne 0x0000000000000117
8b: mov $0x14,%esi c2: mov $0x14,%esi
90: callq 0xffffffffe1021e15 c7: callq 0xffffffffe102bd75
95: test $0x1fff,%ax cc: test $0x1fff,%rax
99: jne 0x00000000000000c6 d3: jne 0x0000000000000117
d5: mov %rax,%r14
9b: mov $0xe,%esi d8: mov $0xe,%esi
a0: callq 0xffffffffe1021e44 dd: callq 0xffffffffe102bd91 // MSH
e2: and $0xf,%eax
e5: shl $0x2,%eax
e8: mov %rax,%r13
eb: mov %r14,%rax
ee: mov %r13,%rsi
a5: lea 0xe(%rbx),%esi f1: add $0xe,%esi
a8: callq 0xffffffffe1021e0d f4: callq 0xffffffffe102bd6d
ad: cmp $0x16,%eax f9: cmp $0x16,%rax
b0: je 0x00000000000000bf fd: je 0x0000000000000110
ff: mov %r13,%rsi
b2: lea 0x10(%rbx),%esi 102: add $0x10,%esi
b5: callq 0xffffffffe1021e0d 105: callq 0xffffffffe102bd6d
ba: cmp $0x16,%eax 10a: cmp $0x16,%rax
bd: jne 0x00000000000000c6 10e: jne 0x0000000000000117
bf: mov $0xffff,%eax 110: mov $0xffff,%eax
c4: jmp 0x00000000000000c8 115: jmp 0x000000000000011c
c6: xor %eax,%eax 117: mov $0x0,%eax
c8: mov -0x8(%rbp),%rbx 11c: mov -0x228(%rbp),%rbx // epilogue
cc: leaveq 123: mov -0x220(%rbp),%r13
cd: retq 12a: mov -0x218(%rbp),%r14
131: mov -0x210(%rbp),%r15
138: leaveq
139: retq
On fully cached SKBs both JITed functions take 12 nsec to execute.
BPF interpreter executes the program in 30 nsec.
The difference in generated assembler is due to the following:
Old BPF imlements LDX_MSH instruction via sk_load_byte_msh() helper function
inside bpf_jit.S.
New JIT removes the helper and does it explicitly, so ldx_msh cost
is the same for both JITs, but generated code looks longer.
New JIT has 4 registers to save, so prologue/epilogue are larger,
but the cost is within noise on x64.
Old JIT checks whether first insn clears A and if not emits 'xor %eax,%eax'.
New JIT clears %rax unconditionally.
2. old BPF JIT doesn't support ANC_NLATTR, ANC_PAY_OFFSET, ANC_RANDOM
extensions. New JIT supports all BPF extensions.
Performance of such filters improves 2-4 times depending on a filter.
The longer the filter the higher performance gain.
Synthetic benchmarks with many ancillary loads see 20x speedup
which seems to be the maximum gain from JIT
Notes:
. net.core.bpf_jit_enable=2 + tools/net/bpf_jit_disasm is still functional
and can be used to see generated assembler
. there are two jit_compile() functions and code flow for classic filters is:
sk_attach_filter() - load classic BPF
bpf_jit_compile() - try to JIT from classic BPF
sk_convert_filter() - convert classic to internal
bpf_int_jit_compile() - JIT from internal BPF
seccomp and tracing filters will just call bpf_int_jit_compile()
Signed-off-by: Alexei Starovoitov <ast@plumgrid.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
2014-05-14 02:50:46 +00:00
|
|
|
|
2017-05-30 20:31:27 +00:00
|
|
|
case BPF_JMP | BPF_TAIL_CALL:
|
bpf, x86: Emit patchable direct jump as tail call
Add initial code emission for *direct* jumps for tail call maps in
order to avoid the retpoline overhead from a493a87f38cf ("bpf, x64:
implement retpoline for tail call") for situations that allow for
it, meaning, for known constant keys at verification time which are
used as index into the tail call map. In case of Cilium which makes
heavy use of tail calls, constant keys are used in the vast majority,
only for a single occurrence we use a dynamic key.
High level outline is that if the target prog is NULL in the map, we
emit a 5-byte nop for the fall-through case and if not, we emit a
5-byte direct relative jmp to the target bpf_func + skipped prologue
offset. Later during runtime, we patch these 5-byte nop/jmps upon
tail call map update or deletions dynamically. Note that on x86-64
the direct jmp works as we reuse the same stack frame and skip
prologue (as opposed to some other JIT implementations).
One of the issues is that the tail call map slots can change at any
given time even during JITing. Therefore, we have two passes: i) emit
nops for all patchable locations during main JITing phase until we
declare prog->jited = 1 eventually. At this point the image is stable,
not public yet and with all jmps disabled. While JITing, we collect
additional info like poke->ip in order to remember the patch location
for later modifications. In ii) bpf_tail_call_direct_fixup() walks
over the progs poke_tab, locks the tail call maps poke_mutex to
prevent from parallel updates and patches in the right locations via
__bpf_arch_text_poke(). Note, the main bpf_arch_text_poke() cannot
be used at this point since we're not yet exposed to kallsyms. For
the update we use plain memcpy() since the image is not public and
still in read-write mode. After patching, we activate that poke entry
through poke->ip_stable. Meaning, at this point any tail call map
updates/deletions are not going to ignore that poke entry anymore.
Then, bpf_arch_text_poke() might still occur on the read-write image
until we finally locked it as read-only. Both modifications on the
given image are under text_mutex to avoid interference with each
other when update requests come in in parallel for different tail
call maps (current one we have locked in JIT and different one where
poke->ip_stable was already set).
Example prog:
# ./bpftool p d x i 1655
0: (b7) r3 = 0
1: (18) r2 = map[id:526]
3: (85) call bpf_tail_call#12
4: (b7) r0 = 1
5: (95) exit
Before:
# ./bpftool p d j i 1655
0xffffffffc076e55c:
0: nopl 0x0(%rax,%rax,1)
5: push %rbp
6: mov %rsp,%rbp
9: sub $0x200,%rsp
10: push %rbx
11: push %r13
13: push %r14
15: push %r15
17: pushq $0x0 _
19: xor %edx,%edx |_ index (arg 3)
1b: movabs $0xffff88d95cc82600,%rsi |_ map (arg 2)
25: mov %edx,%edx | index >= array->map.max_entries
27: cmp %edx,0x24(%rsi) |
2a: jbe 0x0000000000000066 |_
2c: mov -0x224(%rbp),%eax | tail call limit check
32: cmp $0x20,%eax |
35: ja 0x0000000000000066 |
37: add $0x1,%eax |
3a: mov %eax,-0x224(%rbp) |_
40: mov 0xd0(%rsi,%rdx,8),%rax |_ prog = array->ptrs[index]
48: test %rax,%rax | prog == NULL check
4b: je 0x0000000000000066 |_
4d: mov 0x30(%rax),%rax | goto *(prog->bpf_func + prologue_size)
51: add $0x19,%rax |
55: callq 0x0000000000000061 | retpoline for indirect jump
5a: pause |
5c: lfence |
5f: jmp 0x000000000000005a |
61: mov %rax,(%rsp) |
65: retq |_
66: mov $0x1,%eax
6b: pop %rbx
6c: pop %r15
6e: pop %r14
70: pop %r13
72: pop %rbx
73: leaveq
74: retq
After; state after JIT:
# ./bpftool p d j i 1655
0xffffffffc08e8930:
0: nopl 0x0(%rax,%rax,1)
5: push %rbp
6: mov %rsp,%rbp
9: sub $0x200,%rsp
10: push %rbx
11: push %r13
13: push %r14
15: push %r15
17: pushq $0x0 _
19: xor %edx,%edx |_ index (arg 3)
1b: movabs $0xffff9d8afd74c000,%rsi |_ map (arg 2)
25: mov -0x224(%rbp),%eax | tail call limit check
2b: cmp $0x20,%eax |
2e: ja 0x000000000000003e |
30: add $0x1,%eax |
33: mov %eax,-0x224(%rbp) |_
39: jmpq 0xfffffffffffd1785 |_ [direct] goto *(prog->bpf_func + prologue_size)
3e: mov $0x1,%eax
43: pop %rbx
44: pop %r15
46: pop %r14
48: pop %r13
4a: pop %rbx
4b: leaveq
4c: retq
After; state after map update (target prog):
# ./bpftool p d j i 1655
0xffffffffc08e8930:
0: nopl 0x0(%rax,%rax,1)
5: push %rbp
6: mov %rsp,%rbp
9: sub $0x200,%rsp
10: push %rbx
11: push %r13
13: push %r14
15: push %r15
17: pushq $0x0
19: xor %edx,%edx
1b: movabs $0xffff9d8afd74c000,%rsi
25: mov -0x224(%rbp),%eax
2b: cmp $0x20,%eax .
2e: ja 0x000000000000003e .
30: add $0x1,%eax .
33: mov %eax,-0x224(%rbp) |_
39: jmpq 0xffffffffffb09f55 |_ goto *(prog->bpf_func + prologue_size)
3e: mov $0x1,%eax
43: pop %rbx
44: pop %r15
46: pop %r14
48: pop %r13
4a: pop %rbx
4b: leaveq
4c: retq
After; state after map update (no prog):
# ./bpftool p d j i 1655
0xffffffffc08e8930:
0: nopl 0x0(%rax,%rax,1)
5: push %rbp
6: mov %rsp,%rbp
9: sub $0x200,%rsp
10: push %rbx
11: push %r13
13: push %r14
15: push %r15
17: pushq $0x0
19: xor %edx,%edx
1b: movabs $0xffff9d8afd74c000,%rsi
25: mov -0x224(%rbp),%eax
2b: cmp $0x20,%eax .
2e: ja 0x000000000000003e .
30: add $0x1,%eax .
33: mov %eax,-0x224(%rbp) |_
39: nopl 0x0(%rax,%rax,1) |_ fall-through nop
3e: mov $0x1,%eax
43: pop %rbx
44: pop %r15
46: pop %r14
48: pop %r13
4a: pop %rbx
4b: leaveq
4c: retq
Nice bonus is that this also shrinks the code emission quite a bit
for every tail call invocation.
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Link: https://lore.kernel.org/bpf/6ada4c1c9d35eeb5f4ecfab94593dafa6b5c4b09.1574452833.git.daniel@iogearbox.net
2019-11-22 20:08:00 +00:00
|
|
|
if (imm32)
|
bpf: Implement BPF exceptions
This patch implements BPF exceptions, and introduces a bpf_throw kfunc
to allow programs to throw exceptions during their execution at runtime.
A bpf_throw invocation is treated as an immediate termination of the
program, returning back to its caller within the kernel, unwinding all
stack frames.
This allows the program to simplify its implementation, by testing for
runtime conditions which the verifier has no visibility into, and assert
that they are true. In case they are not, the program can simply throw
an exception from the other branch.
BPF exceptions are explicitly *NOT* an unlikely slowpath error handling
primitive, and this objective has guided design choices of the
implementation of the them within the kernel (with the bulk of the cost
for unwinding the stack offloaded to the bpf_throw kfunc).
The implementation of this mechanism requires use of add_hidden_subprog
mechanism introduced in the previous patch, which generates a couple of
instructions to move R1 to R0 and exit. The JIT then rewrites the
prologue of this subprog to take the stack pointer and frame pointer as
inputs and reset the stack frame, popping all callee-saved registers
saved by the main subprog. The bpf_throw function then walks the stack
at runtime, and invokes this exception subprog with the stack and frame
pointers as parameters.
Reviewers must take note that currently the main program is made to save
all callee-saved registers on x86_64 during entry into the program. This
is because we must do an equivalent of a lightweight context switch when
unwinding the stack, therefore we need the callee-saved registers of the
caller of the BPF program to be able to return with a sane state.
Note that we have to additionally handle r12, even though it is not used
by the program, because when throwing the exception the program makes an
entry into the kernel which could clobber r12 after saving it on the
stack. To be able to preserve the value we received on program entry, we
push r12 and restore it from the generated subprogram when unwinding the
stack.
For now, bpf_throw invocation fails when lingering resources or locks
exist in that path of the program. In a future followup, bpf_throw will
be extended to perform frame-by-frame unwinding to release lingering
resources for each stack frame, removing this limitation.
Signed-off-by: Kumar Kartikeya Dwivedi <memxor@gmail.com>
Link: https://lore.kernel.org/r/20230912233214.1518551-5-memxor@gmail.com
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
2023-09-12 23:32:01 +00:00
|
|
|
emit_bpf_tail_call_direct(bpf_prog,
|
|
|
|
&bpf_prog->aux->poke_tab[imm32 - 1],
|
2021-10-26 12:01:47 +00:00
|
|
|
&prog, image + addrs[i - 1],
|
bpf, x64: rework pro/epilogue and tailcall handling in JIT
This commit serves two things:
1) it optimizes BPF prologue/epilogue generation
2) it makes possible to have tailcalls within BPF subprogram
Both points are related to each other since without 1), 2) could not be
achieved.
In [1], Alexei says:
"The prologue will look like:
nop5
xor eax,eax // two new bytes if bpf_tail_call() is used in this
// function
push rbp
mov rbp, rsp
sub rsp, rounded_stack_depth
push rax // zero init tail_call counter
variable number of push rbx,r13,r14,r15
Then bpf_tail_call will pop variable number rbx,..
and final 'pop rax'
Then 'add rsp, size_of_current_stack_frame'
jmp to next function and skip over 'nop5; xor eax,eax; push rpb; mov
rbp, rsp'
This way new function will set its own stack size and will init tail
call
counter with whatever value the parent had.
If next function doesn't use bpf_tail_call it won't have 'xor eax,eax'.
Instead it would need to have 'nop2' in there."
Implement that suggestion.
Since the layout of stack is changed, tail call counter handling can not
rely anymore on popping it to rbx just like it have been handled for
constant prologue case and later overwrite of rbx with actual value of
rbx pushed to stack. Therefore, let's use one of the register (%rcx) that
is considered to be volatile/caller-saved and pop the value of tail call
counter in there in the epilogue.
Drop the BUILD_BUG_ON in emit_prologue and in
emit_bpf_tail_call_indirect where instruction layout is not constant
anymore.
Introduce new poke target, 'tailcall_bypass' to poke descriptor that is
dedicated for skipping the register pops and stack unwind that are
generated right before the actual jump to target program.
For case when the target program is not present, BPF program will skip
the pop instructions and nop5 dedicated for jmpq $target. An example of
such state when only R6 of callee saved registers is used by program:
ffffffffc0513aa1: e9 0e 00 00 00 jmpq 0xffffffffc0513ab4
ffffffffc0513aa6: 5b pop %rbx
ffffffffc0513aa7: 58 pop %rax
ffffffffc0513aa8: 48 81 c4 00 00 00 00 add $0x0,%rsp
ffffffffc0513aaf: 0f 1f 44 00 00 nopl 0x0(%rax,%rax,1)
ffffffffc0513ab4: 48 89 df mov %rbx,%rdi
When target program is inserted, the jump that was there to skip
pops/nop5 will become the nop5, so CPU will go over pops and do the
actual tailcall.
One might ask why there simply can not be pushes after the nop5?
In the following example snippet:
ffffffffc037030c: 48 89 fb mov %rdi,%rbx
(...)
ffffffffc0370332: 5b pop %rbx
ffffffffc0370333: 58 pop %rax
ffffffffc0370334: 48 81 c4 00 00 00 00 add $0x0,%rsp
ffffffffc037033b: 0f 1f 44 00 00 nopl 0x0(%rax,%rax,1)
ffffffffc0370340: 48 81 ec 00 00 00 00 sub $0x0,%rsp
ffffffffc0370347: 50 push %rax
ffffffffc0370348: 53 push %rbx
ffffffffc0370349: 48 89 df mov %rbx,%rdi
ffffffffc037034c: e8 f7 21 00 00 callq 0xffffffffc0372548
There is the bpf2bpf call (at ffffffffc037034c) right after the tailcall
and jump target is not present. ctx is in %rbx register and BPF
subprogram that we will call into on ffffffffc037034c is relying on it,
e.g. it will pick ctx from there. Such code layout is therefore broken
as we would overwrite the content of %rbx with the value that was pushed
on the prologue. That is the reason for the 'bypass' approach.
Special care needs to be taken during the install/update/remove of
tailcall target. In case when target program is not present, the CPU
must not execute the pop instructions that precede the tailcall.
To address that, the following states can be defined:
A nop, unwind, nop
B nop, unwind, tail
C skip, unwind, nop
D skip, unwind, tail
A is forbidden (lead to incorrectness). The state transitions between
tailcall install/update/remove will work as follows:
First install tail call f: C->D->B(f)
* poke the tailcall, after that get rid of the skip
Update tail call f to f': B(f)->B(f')
* poke the tailcall (poke->tailcall_target) and do NOT touch the
poke->tailcall_bypass
Remove tail call: B(f')->C(f')
* poke->tailcall_bypass is poked back to jump, then we wait the RCU
grace period so that other programs will finish its execution and
after that we are safe to remove the poke->tailcall_target
Install new tail call (f''): C(f')->D(f'')->B(f'').
* same as first step
This way CPU can never be exposed to "unwind, tail" state.
Last but not least, when tailcalls get mixed with bpf2bpf calls, it
would be possible to encounter the endless loop due to clearing the
tailcall counter if for example we would use the tailcall3-like from BPF
selftests program that would be subprogram-based, meaning the tailcall
would be present within the BPF subprogram.
This test, broken down to particular steps, would do:
entry -> set tailcall counter to 0, bump it by 1, tailcall to func0
func0 -> call subprog_tail
(we are NOT skipping the first 11 bytes of prologue and this subprogram
has a tailcall, therefore we clear the counter...)
subprog -> do the same thing as entry
and then loop forever.
To address this, the idea is to go through the call chain of bpf2bpf progs
and look for a tailcall presence throughout whole chain. If we saw a single
tail call then each node in this call chain needs to be marked as a subprog
that can reach the tailcall. We would later feed the JIT with this info
and:
- set eax to 0 only when tailcall is reachable and this is the entry prog
- if tailcall is reachable but there's no tailcall in insns of currently
JITed prog then push rax anyway, so that it will be possible to
propagate further down the call chain
- finally if tailcall is reachable, then we need to precede the 'call'
insn with mov rax, [rbp - (stack_depth + 8)]
Tail call related cases from test_verifier kselftest are also working
fine. Sample BPF programs that utilize tail calls (sockex3, tracex5)
work properly as well.
[1]: https://lore.kernel.org/bpf/20200517043227.2gpq22ifoq37ogst@ast-mbp.dhcp.thefacebook.com/
Suggested-by: Alexei Starovoitov <ast@kernel.org>
Signed-off-by: Maciej Fijalkowski <maciej.fijalkowski@intel.com>
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
2020-09-16 21:10:08 +00:00
|
|
|
callee_regs_used,
|
2021-10-26 12:01:47 +00:00
|
|
|
bpf_prog->aux->stack_depth,
|
|
|
|
ctx);
|
bpf, x86: Emit patchable direct jump as tail call
Add initial code emission for *direct* jumps for tail call maps in
order to avoid the retpoline overhead from a493a87f38cf ("bpf, x64:
implement retpoline for tail call") for situations that allow for
it, meaning, for known constant keys at verification time which are
used as index into the tail call map. In case of Cilium which makes
heavy use of tail calls, constant keys are used in the vast majority,
only for a single occurrence we use a dynamic key.
High level outline is that if the target prog is NULL in the map, we
emit a 5-byte nop for the fall-through case and if not, we emit a
5-byte direct relative jmp to the target bpf_func + skipped prologue
offset. Later during runtime, we patch these 5-byte nop/jmps upon
tail call map update or deletions dynamically. Note that on x86-64
the direct jmp works as we reuse the same stack frame and skip
prologue (as opposed to some other JIT implementations).
One of the issues is that the tail call map slots can change at any
given time even during JITing. Therefore, we have two passes: i) emit
nops for all patchable locations during main JITing phase until we
declare prog->jited = 1 eventually. At this point the image is stable,
not public yet and with all jmps disabled. While JITing, we collect
additional info like poke->ip in order to remember the patch location
for later modifications. In ii) bpf_tail_call_direct_fixup() walks
over the progs poke_tab, locks the tail call maps poke_mutex to
prevent from parallel updates and patches in the right locations via
__bpf_arch_text_poke(). Note, the main bpf_arch_text_poke() cannot
be used at this point since we're not yet exposed to kallsyms. For
the update we use plain memcpy() since the image is not public and
still in read-write mode. After patching, we activate that poke entry
through poke->ip_stable. Meaning, at this point any tail call map
updates/deletions are not going to ignore that poke entry anymore.
Then, bpf_arch_text_poke() might still occur on the read-write image
until we finally locked it as read-only. Both modifications on the
given image are under text_mutex to avoid interference with each
other when update requests come in in parallel for different tail
call maps (current one we have locked in JIT and different one where
poke->ip_stable was already set).
Example prog:
# ./bpftool p d x i 1655
0: (b7) r3 = 0
1: (18) r2 = map[id:526]
3: (85) call bpf_tail_call#12
4: (b7) r0 = 1
5: (95) exit
Before:
# ./bpftool p d j i 1655
0xffffffffc076e55c:
0: nopl 0x0(%rax,%rax,1)
5: push %rbp
6: mov %rsp,%rbp
9: sub $0x200,%rsp
10: push %rbx
11: push %r13
13: push %r14
15: push %r15
17: pushq $0x0 _
19: xor %edx,%edx |_ index (arg 3)
1b: movabs $0xffff88d95cc82600,%rsi |_ map (arg 2)
25: mov %edx,%edx | index >= array->map.max_entries
27: cmp %edx,0x24(%rsi) |
2a: jbe 0x0000000000000066 |_
2c: mov -0x224(%rbp),%eax | tail call limit check
32: cmp $0x20,%eax |
35: ja 0x0000000000000066 |
37: add $0x1,%eax |
3a: mov %eax,-0x224(%rbp) |_
40: mov 0xd0(%rsi,%rdx,8),%rax |_ prog = array->ptrs[index]
48: test %rax,%rax | prog == NULL check
4b: je 0x0000000000000066 |_
4d: mov 0x30(%rax),%rax | goto *(prog->bpf_func + prologue_size)
51: add $0x19,%rax |
55: callq 0x0000000000000061 | retpoline for indirect jump
5a: pause |
5c: lfence |
5f: jmp 0x000000000000005a |
61: mov %rax,(%rsp) |
65: retq |_
66: mov $0x1,%eax
6b: pop %rbx
6c: pop %r15
6e: pop %r14
70: pop %r13
72: pop %rbx
73: leaveq
74: retq
After; state after JIT:
# ./bpftool p d j i 1655
0xffffffffc08e8930:
0: nopl 0x0(%rax,%rax,1)
5: push %rbp
6: mov %rsp,%rbp
9: sub $0x200,%rsp
10: push %rbx
11: push %r13
13: push %r14
15: push %r15
17: pushq $0x0 _
19: xor %edx,%edx |_ index (arg 3)
1b: movabs $0xffff9d8afd74c000,%rsi |_ map (arg 2)
25: mov -0x224(%rbp),%eax | tail call limit check
2b: cmp $0x20,%eax |
2e: ja 0x000000000000003e |
30: add $0x1,%eax |
33: mov %eax,-0x224(%rbp) |_
39: jmpq 0xfffffffffffd1785 |_ [direct] goto *(prog->bpf_func + prologue_size)
3e: mov $0x1,%eax
43: pop %rbx
44: pop %r15
46: pop %r14
48: pop %r13
4a: pop %rbx
4b: leaveq
4c: retq
After; state after map update (target prog):
# ./bpftool p d j i 1655
0xffffffffc08e8930:
0: nopl 0x0(%rax,%rax,1)
5: push %rbp
6: mov %rsp,%rbp
9: sub $0x200,%rsp
10: push %rbx
11: push %r13
13: push %r14
15: push %r15
17: pushq $0x0
19: xor %edx,%edx
1b: movabs $0xffff9d8afd74c000,%rsi
25: mov -0x224(%rbp),%eax
2b: cmp $0x20,%eax .
2e: ja 0x000000000000003e .
30: add $0x1,%eax .
33: mov %eax,-0x224(%rbp) |_
39: jmpq 0xffffffffffb09f55 |_ goto *(prog->bpf_func + prologue_size)
3e: mov $0x1,%eax
43: pop %rbx
44: pop %r15
46: pop %r14
48: pop %r13
4a: pop %rbx
4b: leaveq
4c: retq
After; state after map update (no prog):
# ./bpftool p d j i 1655
0xffffffffc08e8930:
0: nopl 0x0(%rax,%rax,1)
5: push %rbp
6: mov %rsp,%rbp
9: sub $0x200,%rsp
10: push %rbx
11: push %r13
13: push %r14
15: push %r15
17: pushq $0x0
19: xor %edx,%edx
1b: movabs $0xffff9d8afd74c000,%rsi
25: mov -0x224(%rbp),%eax
2b: cmp $0x20,%eax .
2e: ja 0x000000000000003e .
30: add $0x1,%eax .
33: mov %eax,-0x224(%rbp) |_
39: nopl 0x0(%rax,%rax,1) |_ fall-through nop
3e: mov $0x1,%eax
43: pop %rbx
44: pop %r15
46: pop %r14
48: pop %r13
4a: pop %rbx
4b: leaveq
4c: retq
Nice bonus is that this also shrinks the code emission quite a bit
for every tail call invocation.
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Link: https://lore.kernel.org/bpf/6ada4c1c9d35eeb5f4ecfab94593dafa6b5c4b09.1574452833.git.daniel@iogearbox.net
2019-11-22 20:08:00 +00:00
|
|
|
else
|
bpf: Implement BPF exceptions
This patch implements BPF exceptions, and introduces a bpf_throw kfunc
to allow programs to throw exceptions during their execution at runtime.
A bpf_throw invocation is treated as an immediate termination of the
program, returning back to its caller within the kernel, unwinding all
stack frames.
This allows the program to simplify its implementation, by testing for
runtime conditions which the verifier has no visibility into, and assert
that they are true. In case they are not, the program can simply throw
an exception from the other branch.
BPF exceptions are explicitly *NOT* an unlikely slowpath error handling
primitive, and this objective has guided design choices of the
implementation of the them within the kernel (with the bulk of the cost
for unwinding the stack offloaded to the bpf_throw kfunc).
The implementation of this mechanism requires use of add_hidden_subprog
mechanism introduced in the previous patch, which generates a couple of
instructions to move R1 to R0 and exit. The JIT then rewrites the
prologue of this subprog to take the stack pointer and frame pointer as
inputs and reset the stack frame, popping all callee-saved registers
saved by the main subprog. The bpf_throw function then walks the stack
at runtime, and invokes this exception subprog with the stack and frame
pointers as parameters.
Reviewers must take note that currently the main program is made to save
all callee-saved registers on x86_64 during entry into the program. This
is because we must do an equivalent of a lightweight context switch when
unwinding the stack, therefore we need the callee-saved registers of the
caller of the BPF program to be able to return with a sane state.
Note that we have to additionally handle r12, even though it is not used
by the program, because when throwing the exception the program makes an
entry into the kernel which could clobber r12 after saving it on the
stack. To be able to preserve the value we received on program entry, we
push r12 and restore it from the generated subprogram when unwinding the
stack.
For now, bpf_throw invocation fails when lingering resources or locks
exist in that path of the program. In a future followup, bpf_throw will
be extended to perform frame-by-frame unwinding to release lingering
resources for each stack frame, removing this limitation.
Signed-off-by: Kumar Kartikeya Dwivedi <memxor@gmail.com>
Link: https://lore.kernel.org/r/20230912233214.1518551-5-memxor@gmail.com
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
2023-09-12 23:32:01 +00:00
|
|
|
emit_bpf_tail_call_indirect(bpf_prog,
|
|
|
|
&prog,
|
bpf, x64: rework pro/epilogue and tailcall handling in JIT
This commit serves two things:
1) it optimizes BPF prologue/epilogue generation
2) it makes possible to have tailcalls within BPF subprogram
Both points are related to each other since without 1), 2) could not be
achieved.
In [1], Alexei says:
"The prologue will look like:
nop5
xor eax,eax // two new bytes if bpf_tail_call() is used in this
// function
push rbp
mov rbp, rsp
sub rsp, rounded_stack_depth
push rax // zero init tail_call counter
variable number of push rbx,r13,r14,r15
Then bpf_tail_call will pop variable number rbx,..
and final 'pop rax'
Then 'add rsp, size_of_current_stack_frame'
jmp to next function and skip over 'nop5; xor eax,eax; push rpb; mov
rbp, rsp'
This way new function will set its own stack size and will init tail
call
counter with whatever value the parent had.
If next function doesn't use bpf_tail_call it won't have 'xor eax,eax'.
Instead it would need to have 'nop2' in there."
Implement that suggestion.
Since the layout of stack is changed, tail call counter handling can not
rely anymore on popping it to rbx just like it have been handled for
constant prologue case and later overwrite of rbx with actual value of
rbx pushed to stack. Therefore, let's use one of the register (%rcx) that
is considered to be volatile/caller-saved and pop the value of tail call
counter in there in the epilogue.
Drop the BUILD_BUG_ON in emit_prologue and in
emit_bpf_tail_call_indirect where instruction layout is not constant
anymore.
Introduce new poke target, 'tailcall_bypass' to poke descriptor that is
dedicated for skipping the register pops and stack unwind that are
generated right before the actual jump to target program.
For case when the target program is not present, BPF program will skip
the pop instructions and nop5 dedicated for jmpq $target. An example of
such state when only R6 of callee saved registers is used by program:
ffffffffc0513aa1: e9 0e 00 00 00 jmpq 0xffffffffc0513ab4
ffffffffc0513aa6: 5b pop %rbx
ffffffffc0513aa7: 58 pop %rax
ffffffffc0513aa8: 48 81 c4 00 00 00 00 add $0x0,%rsp
ffffffffc0513aaf: 0f 1f 44 00 00 nopl 0x0(%rax,%rax,1)
ffffffffc0513ab4: 48 89 df mov %rbx,%rdi
When target program is inserted, the jump that was there to skip
pops/nop5 will become the nop5, so CPU will go over pops and do the
actual tailcall.
One might ask why there simply can not be pushes after the nop5?
In the following example snippet:
ffffffffc037030c: 48 89 fb mov %rdi,%rbx
(...)
ffffffffc0370332: 5b pop %rbx
ffffffffc0370333: 58 pop %rax
ffffffffc0370334: 48 81 c4 00 00 00 00 add $0x0,%rsp
ffffffffc037033b: 0f 1f 44 00 00 nopl 0x0(%rax,%rax,1)
ffffffffc0370340: 48 81 ec 00 00 00 00 sub $0x0,%rsp
ffffffffc0370347: 50 push %rax
ffffffffc0370348: 53 push %rbx
ffffffffc0370349: 48 89 df mov %rbx,%rdi
ffffffffc037034c: e8 f7 21 00 00 callq 0xffffffffc0372548
There is the bpf2bpf call (at ffffffffc037034c) right after the tailcall
and jump target is not present. ctx is in %rbx register and BPF
subprogram that we will call into on ffffffffc037034c is relying on it,
e.g. it will pick ctx from there. Such code layout is therefore broken
as we would overwrite the content of %rbx with the value that was pushed
on the prologue. That is the reason for the 'bypass' approach.
Special care needs to be taken during the install/update/remove of
tailcall target. In case when target program is not present, the CPU
must not execute the pop instructions that precede the tailcall.
To address that, the following states can be defined:
A nop, unwind, nop
B nop, unwind, tail
C skip, unwind, nop
D skip, unwind, tail
A is forbidden (lead to incorrectness). The state transitions between
tailcall install/update/remove will work as follows:
First install tail call f: C->D->B(f)
* poke the tailcall, after that get rid of the skip
Update tail call f to f': B(f)->B(f')
* poke the tailcall (poke->tailcall_target) and do NOT touch the
poke->tailcall_bypass
Remove tail call: B(f')->C(f')
* poke->tailcall_bypass is poked back to jump, then we wait the RCU
grace period so that other programs will finish its execution and
after that we are safe to remove the poke->tailcall_target
Install new tail call (f''): C(f')->D(f'')->B(f'').
* same as first step
This way CPU can never be exposed to "unwind, tail" state.
Last but not least, when tailcalls get mixed with bpf2bpf calls, it
would be possible to encounter the endless loop due to clearing the
tailcall counter if for example we would use the tailcall3-like from BPF
selftests program that would be subprogram-based, meaning the tailcall
would be present within the BPF subprogram.
This test, broken down to particular steps, would do:
entry -> set tailcall counter to 0, bump it by 1, tailcall to func0
func0 -> call subprog_tail
(we are NOT skipping the first 11 bytes of prologue and this subprogram
has a tailcall, therefore we clear the counter...)
subprog -> do the same thing as entry
and then loop forever.
To address this, the idea is to go through the call chain of bpf2bpf progs
and look for a tailcall presence throughout whole chain. If we saw a single
tail call then each node in this call chain needs to be marked as a subprog
that can reach the tailcall. We would later feed the JIT with this info
and:
- set eax to 0 only when tailcall is reachable and this is the entry prog
- if tailcall is reachable but there's no tailcall in insns of currently
JITed prog then push rax anyway, so that it will be possible to
propagate further down the call chain
- finally if tailcall is reachable, then we need to precede the 'call'
insn with mov rax, [rbp - (stack_depth + 8)]
Tail call related cases from test_verifier kselftest are also working
fine. Sample BPF programs that utilize tail calls (sockex3, tracex5)
work properly as well.
[1]: https://lore.kernel.org/bpf/20200517043227.2gpq22ifoq37ogst@ast-mbp.dhcp.thefacebook.com/
Suggested-by: Alexei Starovoitov <ast@kernel.org>
Signed-off-by: Maciej Fijalkowski <maciej.fijalkowski@intel.com>
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
2020-09-16 21:10:08 +00:00
|
|
|
callee_regs_used,
|
2021-10-26 12:01:47 +00:00
|
|
|
bpf_prog->aux->stack_depth,
|
|
|
|
image + addrs[i - 1],
|
|
|
|
ctx);
|
2015-05-19 23:59:04 +00:00
|
|
|
break;
|
|
|
|
|
net: filter: x86: internal BPF JIT
Maps all internal BPF instructions into x86_64 instructions.
This patch replaces original BPF x64 JIT with internal BPF x64 JIT.
sysctl net.core.bpf_jit_enable is reused as on/off switch.
Performance:
1. old BPF JIT and internal BPF JIT generate equivalent x86_64 code.
No performance difference is observed for filters that were JIT-able before
Example assembler code for BPF filter "tcpdump port 22"
original BPF -> old JIT: original BPF -> internal BPF -> new JIT:
0: push %rbp 0: push %rbp
1: mov %rsp,%rbp 1: mov %rsp,%rbp
4: sub $0x60,%rsp 4: sub $0x228,%rsp
8: mov %rbx,-0x8(%rbp) b: mov %rbx,-0x228(%rbp) // prologue
12: mov %r13,-0x220(%rbp)
19: mov %r14,-0x218(%rbp)
20: mov %r15,-0x210(%rbp)
27: xor %eax,%eax // clear A
c: xor %ebx,%ebx 29: xor %r13,%r13 // clear X
e: mov 0x68(%rdi),%r9d 2c: mov 0x68(%rdi),%r9d
12: sub 0x6c(%rdi),%r9d 30: sub 0x6c(%rdi),%r9d
16: mov 0xd8(%rdi),%r8 34: mov 0xd8(%rdi),%r10
3b: mov %rdi,%rbx
1d: mov $0xc,%esi 3e: mov $0xc,%esi
22: callq 0xffffffffe1021e15 43: callq 0xffffffffe102bd75
27: cmp $0x86dd,%eax 48: cmp $0x86dd,%rax
2c: jne 0x0000000000000069 4f: jne 0x000000000000009a
2e: mov $0x14,%esi 51: mov $0x14,%esi
33: callq 0xffffffffe1021e31 56: callq 0xffffffffe102bd91
38: cmp $0x84,%eax 5b: cmp $0x84,%rax
3d: je 0x0000000000000049 62: je 0x0000000000000074
3f: cmp $0x6,%eax 64: cmp $0x6,%rax
42: je 0x0000000000000049 68: je 0x0000000000000074
44: cmp $0x11,%eax 6a: cmp $0x11,%rax
47: jne 0x00000000000000c6 6e: jne 0x0000000000000117
49: mov $0x36,%esi 74: mov $0x36,%esi
4e: callq 0xffffffffe1021e15 79: callq 0xffffffffe102bd75
53: cmp $0x16,%eax 7e: cmp $0x16,%rax
56: je 0x00000000000000bf 82: je 0x0000000000000110
58: mov $0x38,%esi 88: mov $0x38,%esi
5d: callq 0xffffffffe1021e15 8d: callq 0xffffffffe102bd75
62: cmp $0x16,%eax 92: cmp $0x16,%rax
65: je 0x00000000000000bf 96: je 0x0000000000000110
67: jmp 0x00000000000000c6 98: jmp 0x0000000000000117
69: cmp $0x800,%eax 9a: cmp $0x800,%rax
6e: jne 0x00000000000000c6 a1: jne 0x0000000000000117
70: mov $0x17,%esi a3: mov $0x17,%esi
75: callq 0xffffffffe1021e31 a8: callq 0xffffffffe102bd91
7a: cmp $0x84,%eax ad: cmp $0x84,%rax
7f: je 0x000000000000008b b4: je 0x00000000000000c2
81: cmp $0x6,%eax b6: cmp $0x6,%rax
84: je 0x000000000000008b ba: je 0x00000000000000c2
86: cmp $0x11,%eax bc: cmp $0x11,%rax
89: jne 0x00000000000000c6 c0: jne 0x0000000000000117
8b: mov $0x14,%esi c2: mov $0x14,%esi
90: callq 0xffffffffe1021e15 c7: callq 0xffffffffe102bd75
95: test $0x1fff,%ax cc: test $0x1fff,%rax
99: jne 0x00000000000000c6 d3: jne 0x0000000000000117
d5: mov %rax,%r14
9b: mov $0xe,%esi d8: mov $0xe,%esi
a0: callq 0xffffffffe1021e44 dd: callq 0xffffffffe102bd91 // MSH
e2: and $0xf,%eax
e5: shl $0x2,%eax
e8: mov %rax,%r13
eb: mov %r14,%rax
ee: mov %r13,%rsi
a5: lea 0xe(%rbx),%esi f1: add $0xe,%esi
a8: callq 0xffffffffe1021e0d f4: callq 0xffffffffe102bd6d
ad: cmp $0x16,%eax f9: cmp $0x16,%rax
b0: je 0x00000000000000bf fd: je 0x0000000000000110
ff: mov %r13,%rsi
b2: lea 0x10(%rbx),%esi 102: add $0x10,%esi
b5: callq 0xffffffffe1021e0d 105: callq 0xffffffffe102bd6d
ba: cmp $0x16,%eax 10a: cmp $0x16,%rax
bd: jne 0x00000000000000c6 10e: jne 0x0000000000000117
bf: mov $0xffff,%eax 110: mov $0xffff,%eax
c4: jmp 0x00000000000000c8 115: jmp 0x000000000000011c
c6: xor %eax,%eax 117: mov $0x0,%eax
c8: mov -0x8(%rbp),%rbx 11c: mov -0x228(%rbp),%rbx // epilogue
cc: leaveq 123: mov -0x220(%rbp),%r13
cd: retq 12a: mov -0x218(%rbp),%r14
131: mov -0x210(%rbp),%r15
138: leaveq
139: retq
On fully cached SKBs both JITed functions take 12 nsec to execute.
BPF interpreter executes the program in 30 nsec.
The difference in generated assembler is due to the following:
Old BPF imlements LDX_MSH instruction via sk_load_byte_msh() helper function
inside bpf_jit.S.
New JIT removes the helper and does it explicitly, so ldx_msh cost
is the same for both JITs, but generated code looks longer.
New JIT has 4 registers to save, so prologue/epilogue are larger,
but the cost is within noise on x64.
Old JIT checks whether first insn clears A and if not emits 'xor %eax,%eax'.
New JIT clears %rax unconditionally.
2. old BPF JIT doesn't support ANC_NLATTR, ANC_PAY_OFFSET, ANC_RANDOM
extensions. New JIT supports all BPF extensions.
Performance of such filters improves 2-4 times depending on a filter.
The longer the filter the higher performance gain.
Synthetic benchmarks with many ancillary loads see 20x speedup
which seems to be the maximum gain from JIT
Notes:
. net.core.bpf_jit_enable=2 + tools/net/bpf_jit_disasm is still functional
and can be used to see generated assembler
. there are two jit_compile() functions and code flow for classic filters is:
sk_attach_filter() - load classic BPF
bpf_jit_compile() - try to JIT from classic BPF
sk_convert_filter() - convert classic to internal
bpf_int_jit_compile() - JIT from internal BPF
seccomp and tracing filters will just call bpf_int_jit_compile()
Signed-off-by: Alexei Starovoitov <ast@plumgrid.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
2014-05-14 02:50:46 +00:00
|
|
|
/* cond jump */
|
|
|
|
case BPF_JMP | BPF_JEQ | BPF_X:
|
|
|
|
case BPF_JMP | BPF_JNE | BPF_X:
|
|
|
|
case BPF_JMP | BPF_JGT | BPF_X:
|
2017-08-09 23:39:56 +00:00
|
|
|
case BPF_JMP | BPF_JLT | BPF_X:
|
net: filter: x86: internal BPF JIT
Maps all internal BPF instructions into x86_64 instructions.
This patch replaces original BPF x64 JIT with internal BPF x64 JIT.
sysctl net.core.bpf_jit_enable is reused as on/off switch.
Performance:
1. old BPF JIT and internal BPF JIT generate equivalent x86_64 code.
No performance difference is observed for filters that were JIT-able before
Example assembler code for BPF filter "tcpdump port 22"
original BPF -> old JIT: original BPF -> internal BPF -> new JIT:
0: push %rbp 0: push %rbp
1: mov %rsp,%rbp 1: mov %rsp,%rbp
4: sub $0x60,%rsp 4: sub $0x228,%rsp
8: mov %rbx,-0x8(%rbp) b: mov %rbx,-0x228(%rbp) // prologue
12: mov %r13,-0x220(%rbp)
19: mov %r14,-0x218(%rbp)
20: mov %r15,-0x210(%rbp)
27: xor %eax,%eax // clear A
c: xor %ebx,%ebx 29: xor %r13,%r13 // clear X
e: mov 0x68(%rdi),%r9d 2c: mov 0x68(%rdi),%r9d
12: sub 0x6c(%rdi),%r9d 30: sub 0x6c(%rdi),%r9d
16: mov 0xd8(%rdi),%r8 34: mov 0xd8(%rdi),%r10
3b: mov %rdi,%rbx
1d: mov $0xc,%esi 3e: mov $0xc,%esi
22: callq 0xffffffffe1021e15 43: callq 0xffffffffe102bd75
27: cmp $0x86dd,%eax 48: cmp $0x86dd,%rax
2c: jne 0x0000000000000069 4f: jne 0x000000000000009a
2e: mov $0x14,%esi 51: mov $0x14,%esi
33: callq 0xffffffffe1021e31 56: callq 0xffffffffe102bd91
38: cmp $0x84,%eax 5b: cmp $0x84,%rax
3d: je 0x0000000000000049 62: je 0x0000000000000074
3f: cmp $0x6,%eax 64: cmp $0x6,%rax
42: je 0x0000000000000049 68: je 0x0000000000000074
44: cmp $0x11,%eax 6a: cmp $0x11,%rax
47: jne 0x00000000000000c6 6e: jne 0x0000000000000117
49: mov $0x36,%esi 74: mov $0x36,%esi
4e: callq 0xffffffffe1021e15 79: callq 0xffffffffe102bd75
53: cmp $0x16,%eax 7e: cmp $0x16,%rax
56: je 0x00000000000000bf 82: je 0x0000000000000110
58: mov $0x38,%esi 88: mov $0x38,%esi
5d: callq 0xffffffffe1021e15 8d: callq 0xffffffffe102bd75
62: cmp $0x16,%eax 92: cmp $0x16,%rax
65: je 0x00000000000000bf 96: je 0x0000000000000110
67: jmp 0x00000000000000c6 98: jmp 0x0000000000000117
69: cmp $0x800,%eax 9a: cmp $0x800,%rax
6e: jne 0x00000000000000c6 a1: jne 0x0000000000000117
70: mov $0x17,%esi a3: mov $0x17,%esi
75: callq 0xffffffffe1021e31 a8: callq 0xffffffffe102bd91
7a: cmp $0x84,%eax ad: cmp $0x84,%rax
7f: je 0x000000000000008b b4: je 0x00000000000000c2
81: cmp $0x6,%eax b6: cmp $0x6,%rax
84: je 0x000000000000008b ba: je 0x00000000000000c2
86: cmp $0x11,%eax bc: cmp $0x11,%rax
89: jne 0x00000000000000c6 c0: jne 0x0000000000000117
8b: mov $0x14,%esi c2: mov $0x14,%esi
90: callq 0xffffffffe1021e15 c7: callq 0xffffffffe102bd75
95: test $0x1fff,%ax cc: test $0x1fff,%rax
99: jne 0x00000000000000c6 d3: jne 0x0000000000000117
d5: mov %rax,%r14
9b: mov $0xe,%esi d8: mov $0xe,%esi
a0: callq 0xffffffffe1021e44 dd: callq 0xffffffffe102bd91 // MSH
e2: and $0xf,%eax
e5: shl $0x2,%eax
e8: mov %rax,%r13
eb: mov %r14,%rax
ee: mov %r13,%rsi
a5: lea 0xe(%rbx),%esi f1: add $0xe,%esi
a8: callq 0xffffffffe1021e0d f4: callq 0xffffffffe102bd6d
ad: cmp $0x16,%eax f9: cmp $0x16,%rax
b0: je 0x00000000000000bf fd: je 0x0000000000000110
ff: mov %r13,%rsi
b2: lea 0x10(%rbx),%esi 102: add $0x10,%esi
b5: callq 0xffffffffe1021e0d 105: callq 0xffffffffe102bd6d
ba: cmp $0x16,%eax 10a: cmp $0x16,%rax
bd: jne 0x00000000000000c6 10e: jne 0x0000000000000117
bf: mov $0xffff,%eax 110: mov $0xffff,%eax
c4: jmp 0x00000000000000c8 115: jmp 0x000000000000011c
c6: xor %eax,%eax 117: mov $0x0,%eax
c8: mov -0x8(%rbp),%rbx 11c: mov -0x228(%rbp),%rbx // epilogue
cc: leaveq 123: mov -0x220(%rbp),%r13
cd: retq 12a: mov -0x218(%rbp),%r14
131: mov -0x210(%rbp),%r15
138: leaveq
139: retq
On fully cached SKBs both JITed functions take 12 nsec to execute.
BPF interpreter executes the program in 30 nsec.
The difference in generated assembler is due to the following:
Old BPF imlements LDX_MSH instruction via sk_load_byte_msh() helper function
inside bpf_jit.S.
New JIT removes the helper and does it explicitly, so ldx_msh cost
is the same for both JITs, but generated code looks longer.
New JIT has 4 registers to save, so prologue/epilogue are larger,
but the cost is within noise on x64.
Old JIT checks whether first insn clears A and if not emits 'xor %eax,%eax'.
New JIT clears %rax unconditionally.
2. old BPF JIT doesn't support ANC_NLATTR, ANC_PAY_OFFSET, ANC_RANDOM
extensions. New JIT supports all BPF extensions.
Performance of such filters improves 2-4 times depending on a filter.
The longer the filter the higher performance gain.
Synthetic benchmarks with many ancillary loads see 20x speedup
which seems to be the maximum gain from JIT
Notes:
. net.core.bpf_jit_enable=2 + tools/net/bpf_jit_disasm is still functional
and can be used to see generated assembler
. there are two jit_compile() functions and code flow for classic filters is:
sk_attach_filter() - load classic BPF
bpf_jit_compile() - try to JIT from classic BPF
sk_convert_filter() - convert classic to internal
bpf_int_jit_compile() - JIT from internal BPF
seccomp and tracing filters will just call bpf_int_jit_compile()
Signed-off-by: Alexei Starovoitov <ast@plumgrid.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
2014-05-14 02:50:46 +00:00
|
|
|
case BPF_JMP | BPF_JGE | BPF_X:
|
2017-08-09 23:39:56 +00:00
|
|
|
case BPF_JMP | BPF_JLE | BPF_X:
|
net: filter: x86: internal BPF JIT
Maps all internal BPF instructions into x86_64 instructions.
This patch replaces original BPF x64 JIT with internal BPF x64 JIT.
sysctl net.core.bpf_jit_enable is reused as on/off switch.
Performance:
1. old BPF JIT and internal BPF JIT generate equivalent x86_64 code.
No performance difference is observed for filters that were JIT-able before
Example assembler code for BPF filter "tcpdump port 22"
original BPF -> old JIT: original BPF -> internal BPF -> new JIT:
0: push %rbp 0: push %rbp
1: mov %rsp,%rbp 1: mov %rsp,%rbp
4: sub $0x60,%rsp 4: sub $0x228,%rsp
8: mov %rbx,-0x8(%rbp) b: mov %rbx,-0x228(%rbp) // prologue
12: mov %r13,-0x220(%rbp)
19: mov %r14,-0x218(%rbp)
20: mov %r15,-0x210(%rbp)
27: xor %eax,%eax // clear A
c: xor %ebx,%ebx 29: xor %r13,%r13 // clear X
e: mov 0x68(%rdi),%r9d 2c: mov 0x68(%rdi),%r9d
12: sub 0x6c(%rdi),%r9d 30: sub 0x6c(%rdi),%r9d
16: mov 0xd8(%rdi),%r8 34: mov 0xd8(%rdi),%r10
3b: mov %rdi,%rbx
1d: mov $0xc,%esi 3e: mov $0xc,%esi
22: callq 0xffffffffe1021e15 43: callq 0xffffffffe102bd75
27: cmp $0x86dd,%eax 48: cmp $0x86dd,%rax
2c: jne 0x0000000000000069 4f: jne 0x000000000000009a
2e: mov $0x14,%esi 51: mov $0x14,%esi
33: callq 0xffffffffe1021e31 56: callq 0xffffffffe102bd91
38: cmp $0x84,%eax 5b: cmp $0x84,%rax
3d: je 0x0000000000000049 62: je 0x0000000000000074
3f: cmp $0x6,%eax 64: cmp $0x6,%rax
42: je 0x0000000000000049 68: je 0x0000000000000074
44: cmp $0x11,%eax 6a: cmp $0x11,%rax
47: jne 0x00000000000000c6 6e: jne 0x0000000000000117
49: mov $0x36,%esi 74: mov $0x36,%esi
4e: callq 0xffffffffe1021e15 79: callq 0xffffffffe102bd75
53: cmp $0x16,%eax 7e: cmp $0x16,%rax
56: je 0x00000000000000bf 82: je 0x0000000000000110
58: mov $0x38,%esi 88: mov $0x38,%esi
5d: callq 0xffffffffe1021e15 8d: callq 0xffffffffe102bd75
62: cmp $0x16,%eax 92: cmp $0x16,%rax
65: je 0x00000000000000bf 96: je 0x0000000000000110
67: jmp 0x00000000000000c6 98: jmp 0x0000000000000117
69: cmp $0x800,%eax 9a: cmp $0x800,%rax
6e: jne 0x00000000000000c6 a1: jne 0x0000000000000117
70: mov $0x17,%esi a3: mov $0x17,%esi
75: callq 0xffffffffe1021e31 a8: callq 0xffffffffe102bd91
7a: cmp $0x84,%eax ad: cmp $0x84,%rax
7f: je 0x000000000000008b b4: je 0x00000000000000c2
81: cmp $0x6,%eax b6: cmp $0x6,%rax
84: je 0x000000000000008b ba: je 0x00000000000000c2
86: cmp $0x11,%eax bc: cmp $0x11,%rax
89: jne 0x00000000000000c6 c0: jne 0x0000000000000117
8b: mov $0x14,%esi c2: mov $0x14,%esi
90: callq 0xffffffffe1021e15 c7: callq 0xffffffffe102bd75
95: test $0x1fff,%ax cc: test $0x1fff,%rax
99: jne 0x00000000000000c6 d3: jne 0x0000000000000117
d5: mov %rax,%r14
9b: mov $0xe,%esi d8: mov $0xe,%esi
a0: callq 0xffffffffe1021e44 dd: callq 0xffffffffe102bd91 // MSH
e2: and $0xf,%eax
e5: shl $0x2,%eax
e8: mov %rax,%r13
eb: mov %r14,%rax
ee: mov %r13,%rsi
a5: lea 0xe(%rbx),%esi f1: add $0xe,%esi
a8: callq 0xffffffffe1021e0d f4: callq 0xffffffffe102bd6d
ad: cmp $0x16,%eax f9: cmp $0x16,%rax
b0: je 0x00000000000000bf fd: je 0x0000000000000110
ff: mov %r13,%rsi
b2: lea 0x10(%rbx),%esi 102: add $0x10,%esi
b5: callq 0xffffffffe1021e0d 105: callq 0xffffffffe102bd6d
ba: cmp $0x16,%eax 10a: cmp $0x16,%rax
bd: jne 0x00000000000000c6 10e: jne 0x0000000000000117
bf: mov $0xffff,%eax 110: mov $0xffff,%eax
c4: jmp 0x00000000000000c8 115: jmp 0x000000000000011c
c6: xor %eax,%eax 117: mov $0x0,%eax
c8: mov -0x8(%rbp),%rbx 11c: mov -0x228(%rbp),%rbx // epilogue
cc: leaveq 123: mov -0x220(%rbp),%r13
cd: retq 12a: mov -0x218(%rbp),%r14
131: mov -0x210(%rbp),%r15
138: leaveq
139: retq
On fully cached SKBs both JITed functions take 12 nsec to execute.
BPF interpreter executes the program in 30 nsec.
The difference in generated assembler is due to the following:
Old BPF imlements LDX_MSH instruction via sk_load_byte_msh() helper function
inside bpf_jit.S.
New JIT removes the helper and does it explicitly, so ldx_msh cost
is the same for both JITs, but generated code looks longer.
New JIT has 4 registers to save, so prologue/epilogue are larger,
but the cost is within noise on x64.
Old JIT checks whether first insn clears A and if not emits 'xor %eax,%eax'.
New JIT clears %rax unconditionally.
2. old BPF JIT doesn't support ANC_NLATTR, ANC_PAY_OFFSET, ANC_RANDOM
extensions. New JIT supports all BPF extensions.
Performance of such filters improves 2-4 times depending on a filter.
The longer the filter the higher performance gain.
Synthetic benchmarks with many ancillary loads see 20x speedup
which seems to be the maximum gain from JIT
Notes:
. net.core.bpf_jit_enable=2 + tools/net/bpf_jit_disasm is still functional
and can be used to see generated assembler
. there are two jit_compile() functions and code flow for classic filters is:
sk_attach_filter() - load classic BPF
bpf_jit_compile() - try to JIT from classic BPF
sk_convert_filter() - convert classic to internal
bpf_int_jit_compile() - JIT from internal BPF
seccomp and tracing filters will just call bpf_int_jit_compile()
Signed-off-by: Alexei Starovoitov <ast@plumgrid.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
2014-05-14 02:50:46 +00:00
|
|
|
case BPF_JMP | BPF_JSGT | BPF_X:
|
2017-08-09 23:39:56 +00:00
|
|
|
case BPF_JMP | BPF_JSLT | BPF_X:
|
net: filter: x86: internal BPF JIT
Maps all internal BPF instructions into x86_64 instructions.
This patch replaces original BPF x64 JIT with internal BPF x64 JIT.
sysctl net.core.bpf_jit_enable is reused as on/off switch.
Performance:
1. old BPF JIT and internal BPF JIT generate equivalent x86_64 code.
No performance difference is observed for filters that were JIT-able before
Example assembler code for BPF filter "tcpdump port 22"
original BPF -> old JIT: original BPF -> internal BPF -> new JIT:
0: push %rbp 0: push %rbp
1: mov %rsp,%rbp 1: mov %rsp,%rbp
4: sub $0x60,%rsp 4: sub $0x228,%rsp
8: mov %rbx,-0x8(%rbp) b: mov %rbx,-0x228(%rbp) // prologue
12: mov %r13,-0x220(%rbp)
19: mov %r14,-0x218(%rbp)
20: mov %r15,-0x210(%rbp)
27: xor %eax,%eax // clear A
c: xor %ebx,%ebx 29: xor %r13,%r13 // clear X
e: mov 0x68(%rdi),%r9d 2c: mov 0x68(%rdi),%r9d
12: sub 0x6c(%rdi),%r9d 30: sub 0x6c(%rdi),%r9d
16: mov 0xd8(%rdi),%r8 34: mov 0xd8(%rdi),%r10
3b: mov %rdi,%rbx
1d: mov $0xc,%esi 3e: mov $0xc,%esi
22: callq 0xffffffffe1021e15 43: callq 0xffffffffe102bd75
27: cmp $0x86dd,%eax 48: cmp $0x86dd,%rax
2c: jne 0x0000000000000069 4f: jne 0x000000000000009a
2e: mov $0x14,%esi 51: mov $0x14,%esi
33: callq 0xffffffffe1021e31 56: callq 0xffffffffe102bd91
38: cmp $0x84,%eax 5b: cmp $0x84,%rax
3d: je 0x0000000000000049 62: je 0x0000000000000074
3f: cmp $0x6,%eax 64: cmp $0x6,%rax
42: je 0x0000000000000049 68: je 0x0000000000000074
44: cmp $0x11,%eax 6a: cmp $0x11,%rax
47: jne 0x00000000000000c6 6e: jne 0x0000000000000117
49: mov $0x36,%esi 74: mov $0x36,%esi
4e: callq 0xffffffffe1021e15 79: callq 0xffffffffe102bd75
53: cmp $0x16,%eax 7e: cmp $0x16,%rax
56: je 0x00000000000000bf 82: je 0x0000000000000110
58: mov $0x38,%esi 88: mov $0x38,%esi
5d: callq 0xffffffffe1021e15 8d: callq 0xffffffffe102bd75
62: cmp $0x16,%eax 92: cmp $0x16,%rax
65: je 0x00000000000000bf 96: je 0x0000000000000110
67: jmp 0x00000000000000c6 98: jmp 0x0000000000000117
69: cmp $0x800,%eax 9a: cmp $0x800,%rax
6e: jne 0x00000000000000c6 a1: jne 0x0000000000000117
70: mov $0x17,%esi a3: mov $0x17,%esi
75: callq 0xffffffffe1021e31 a8: callq 0xffffffffe102bd91
7a: cmp $0x84,%eax ad: cmp $0x84,%rax
7f: je 0x000000000000008b b4: je 0x00000000000000c2
81: cmp $0x6,%eax b6: cmp $0x6,%rax
84: je 0x000000000000008b ba: je 0x00000000000000c2
86: cmp $0x11,%eax bc: cmp $0x11,%rax
89: jne 0x00000000000000c6 c0: jne 0x0000000000000117
8b: mov $0x14,%esi c2: mov $0x14,%esi
90: callq 0xffffffffe1021e15 c7: callq 0xffffffffe102bd75
95: test $0x1fff,%ax cc: test $0x1fff,%rax
99: jne 0x00000000000000c6 d3: jne 0x0000000000000117
d5: mov %rax,%r14
9b: mov $0xe,%esi d8: mov $0xe,%esi
a0: callq 0xffffffffe1021e44 dd: callq 0xffffffffe102bd91 // MSH
e2: and $0xf,%eax
e5: shl $0x2,%eax
e8: mov %rax,%r13
eb: mov %r14,%rax
ee: mov %r13,%rsi
a5: lea 0xe(%rbx),%esi f1: add $0xe,%esi
a8: callq 0xffffffffe1021e0d f4: callq 0xffffffffe102bd6d
ad: cmp $0x16,%eax f9: cmp $0x16,%rax
b0: je 0x00000000000000bf fd: je 0x0000000000000110
ff: mov %r13,%rsi
b2: lea 0x10(%rbx),%esi 102: add $0x10,%esi
b5: callq 0xffffffffe1021e0d 105: callq 0xffffffffe102bd6d
ba: cmp $0x16,%eax 10a: cmp $0x16,%rax
bd: jne 0x00000000000000c6 10e: jne 0x0000000000000117
bf: mov $0xffff,%eax 110: mov $0xffff,%eax
c4: jmp 0x00000000000000c8 115: jmp 0x000000000000011c
c6: xor %eax,%eax 117: mov $0x0,%eax
c8: mov -0x8(%rbp),%rbx 11c: mov -0x228(%rbp),%rbx // epilogue
cc: leaveq 123: mov -0x220(%rbp),%r13
cd: retq 12a: mov -0x218(%rbp),%r14
131: mov -0x210(%rbp),%r15
138: leaveq
139: retq
On fully cached SKBs both JITed functions take 12 nsec to execute.
BPF interpreter executes the program in 30 nsec.
The difference in generated assembler is due to the following:
Old BPF imlements LDX_MSH instruction via sk_load_byte_msh() helper function
inside bpf_jit.S.
New JIT removes the helper and does it explicitly, so ldx_msh cost
is the same for both JITs, but generated code looks longer.
New JIT has 4 registers to save, so prologue/epilogue are larger,
but the cost is within noise on x64.
Old JIT checks whether first insn clears A and if not emits 'xor %eax,%eax'.
New JIT clears %rax unconditionally.
2. old BPF JIT doesn't support ANC_NLATTR, ANC_PAY_OFFSET, ANC_RANDOM
extensions. New JIT supports all BPF extensions.
Performance of such filters improves 2-4 times depending on a filter.
The longer the filter the higher performance gain.
Synthetic benchmarks with many ancillary loads see 20x speedup
which seems to be the maximum gain from JIT
Notes:
. net.core.bpf_jit_enable=2 + tools/net/bpf_jit_disasm is still functional
and can be used to see generated assembler
. there are two jit_compile() functions and code flow for classic filters is:
sk_attach_filter() - load classic BPF
bpf_jit_compile() - try to JIT from classic BPF
sk_convert_filter() - convert classic to internal
bpf_int_jit_compile() - JIT from internal BPF
seccomp and tracing filters will just call bpf_int_jit_compile()
Signed-off-by: Alexei Starovoitov <ast@plumgrid.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
2014-05-14 02:50:46 +00:00
|
|
|
case BPF_JMP | BPF_JSGE | BPF_X:
|
2017-08-09 23:39:56 +00:00
|
|
|
case BPF_JMP | BPF_JSLE | BPF_X:
|
2019-01-26 17:26:06 +00:00
|
|
|
case BPF_JMP32 | BPF_JEQ | BPF_X:
|
|
|
|
case BPF_JMP32 | BPF_JNE | BPF_X:
|
|
|
|
case BPF_JMP32 | BPF_JGT | BPF_X:
|
|
|
|
case BPF_JMP32 | BPF_JLT | BPF_X:
|
|
|
|
case BPF_JMP32 | BPF_JGE | BPF_X:
|
|
|
|
case BPF_JMP32 | BPF_JLE | BPF_X:
|
|
|
|
case BPF_JMP32 | BPF_JSGT | BPF_X:
|
|
|
|
case BPF_JMP32 | BPF_JSLT | BPF_X:
|
|
|
|
case BPF_JMP32 | BPF_JSGE | BPF_X:
|
|
|
|
case BPF_JMP32 | BPF_JSLE | BPF_X:
|
2014-06-06 21:46:06 +00:00
|
|
|
/* cmp dst_reg, src_reg */
|
2021-01-14 18:17:42 +00:00
|
|
|
maybe_emit_mod(&prog, dst_reg, src_reg,
|
|
|
|
BPF_CLASS(insn->code) == BPF_JMP);
|
2019-01-26 17:26:06 +00:00
|
|
|
EMIT2(0x39, add_2reg(0xC0, dst_reg, src_reg));
|
net: filter: x86: internal BPF JIT
Maps all internal BPF instructions into x86_64 instructions.
This patch replaces original BPF x64 JIT with internal BPF x64 JIT.
sysctl net.core.bpf_jit_enable is reused as on/off switch.
Performance:
1. old BPF JIT and internal BPF JIT generate equivalent x86_64 code.
No performance difference is observed for filters that were JIT-able before
Example assembler code for BPF filter "tcpdump port 22"
original BPF -> old JIT: original BPF -> internal BPF -> new JIT:
0: push %rbp 0: push %rbp
1: mov %rsp,%rbp 1: mov %rsp,%rbp
4: sub $0x60,%rsp 4: sub $0x228,%rsp
8: mov %rbx,-0x8(%rbp) b: mov %rbx,-0x228(%rbp) // prologue
12: mov %r13,-0x220(%rbp)
19: mov %r14,-0x218(%rbp)
20: mov %r15,-0x210(%rbp)
27: xor %eax,%eax // clear A
c: xor %ebx,%ebx 29: xor %r13,%r13 // clear X
e: mov 0x68(%rdi),%r9d 2c: mov 0x68(%rdi),%r9d
12: sub 0x6c(%rdi),%r9d 30: sub 0x6c(%rdi),%r9d
16: mov 0xd8(%rdi),%r8 34: mov 0xd8(%rdi),%r10
3b: mov %rdi,%rbx
1d: mov $0xc,%esi 3e: mov $0xc,%esi
22: callq 0xffffffffe1021e15 43: callq 0xffffffffe102bd75
27: cmp $0x86dd,%eax 48: cmp $0x86dd,%rax
2c: jne 0x0000000000000069 4f: jne 0x000000000000009a
2e: mov $0x14,%esi 51: mov $0x14,%esi
33: callq 0xffffffffe1021e31 56: callq 0xffffffffe102bd91
38: cmp $0x84,%eax 5b: cmp $0x84,%rax
3d: je 0x0000000000000049 62: je 0x0000000000000074
3f: cmp $0x6,%eax 64: cmp $0x6,%rax
42: je 0x0000000000000049 68: je 0x0000000000000074
44: cmp $0x11,%eax 6a: cmp $0x11,%rax
47: jne 0x00000000000000c6 6e: jne 0x0000000000000117
49: mov $0x36,%esi 74: mov $0x36,%esi
4e: callq 0xffffffffe1021e15 79: callq 0xffffffffe102bd75
53: cmp $0x16,%eax 7e: cmp $0x16,%rax
56: je 0x00000000000000bf 82: je 0x0000000000000110
58: mov $0x38,%esi 88: mov $0x38,%esi
5d: callq 0xffffffffe1021e15 8d: callq 0xffffffffe102bd75
62: cmp $0x16,%eax 92: cmp $0x16,%rax
65: je 0x00000000000000bf 96: je 0x0000000000000110
67: jmp 0x00000000000000c6 98: jmp 0x0000000000000117
69: cmp $0x800,%eax 9a: cmp $0x800,%rax
6e: jne 0x00000000000000c6 a1: jne 0x0000000000000117
70: mov $0x17,%esi a3: mov $0x17,%esi
75: callq 0xffffffffe1021e31 a8: callq 0xffffffffe102bd91
7a: cmp $0x84,%eax ad: cmp $0x84,%rax
7f: je 0x000000000000008b b4: je 0x00000000000000c2
81: cmp $0x6,%eax b6: cmp $0x6,%rax
84: je 0x000000000000008b ba: je 0x00000000000000c2
86: cmp $0x11,%eax bc: cmp $0x11,%rax
89: jne 0x00000000000000c6 c0: jne 0x0000000000000117
8b: mov $0x14,%esi c2: mov $0x14,%esi
90: callq 0xffffffffe1021e15 c7: callq 0xffffffffe102bd75
95: test $0x1fff,%ax cc: test $0x1fff,%rax
99: jne 0x00000000000000c6 d3: jne 0x0000000000000117
d5: mov %rax,%r14
9b: mov $0xe,%esi d8: mov $0xe,%esi
a0: callq 0xffffffffe1021e44 dd: callq 0xffffffffe102bd91 // MSH
e2: and $0xf,%eax
e5: shl $0x2,%eax
e8: mov %rax,%r13
eb: mov %r14,%rax
ee: mov %r13,%rsi
a5: lea 0xe(%rbx),%esi f1: add $0xe,%esi
a8: callq 0xffffffffe1021e0d f4: callq 0xffffffffe102bd6d
ad: cmp $0x16,%eax f9: cmp $0x16,%rax
b0: je 0x00000000000000bf fd: je 0x0000000000000110
ff: mov %r13,%rsi
b2: lea 0x10(%rbx),%esi 102: add $0x10,%esi
b5: callq 0xffffffffe1021e0d 105: callq 0xffffffffe102bd6d
ba: cmp $0x16,%eax 10a: cmp $0x16,%rax
bd: jne 0x00000000000000c6 10e: jne 0x0000000000000117
bf: mov $0xffff,%eax 110: mov $0xffff,%eax
c4: jmp 0x00000000000000c8 115: jmp 0x000000000000011c
c6: xor %eax,%eax 117: mov $0x0,%eax
c8: mov -0x8(%rbp),%rbx 11c: mov -0x228(%rbp),%rbx // epilogue
cc: leaveq 123: mov -0x220(%rbp),%r13
cd: retq 12a: mov -0x218(%rbp),%r14
131: mov -0x210(%rbp),%r15
138: leaveq
139: retq
On fully cached SKBs both JITed functions take 12 nsec to execute.
BPF interpreter executes the program in 30 nsec.
The difference in generated assembler is due to the following:
Old BPF imlements LDX_MSH instruction via sk_load_byte_msh() helper function
inside bpf_jit.S.
New JIT removes the helper and does it explicitly, so ldx_msh cost
is the same for both JITs, but generated code looks longer.
New JIT has 4 registers to save, so prologue/epilogue are larger,
but the cost is within noise on x64.
Old JIT checks whether first insn clears A and if not emits 'xor %eax,%eax'.
New JIT clears %rax unconditionally.
2. old BPF JIT doesn't support ANC_NLATTR, ANC_PAY_OFFSET, ANC_RANDOM
extensions. New JIT supports all BPF extensions.
Performance of such filters improves 2-4 times depending on a filter.
The longer the filter the higher performance gain.
Synthetic benchmarks with many ancillary loads see 20x speedup
which seems to be the maximum gain from JIT
Notes:
. net.core.bpf_jit_enable=2 + tools/net/bpf_jit_disasm is still functional
and can be used to see generated assembler
. there are two jit_compile() functions and code flow for classic filters is:
sk_attach_filter() - load classic BPF
bpf_jit_compile() - try to JIT from classic BPF
sk_convert_filter() - convert classic to internal
bpf_int_jit_compile() - JIT from internal BPF
seccomp and tracing filters will just call bpf_int_jit_compile()
Signed-off-by: Alexei Starovoitov <ast@plumgrid.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
2014-05-14 02:50:46 +00:00
|
|
|
goto emit_cond_jmp;
|
|
|
|
|
|
|
|
case BPF_JMP | BPF_JSET | BPF_X:
|
2019-01-26 17:26:06 +00:00
|
|
|
case BPF_JMP32 | BPF_JSET | BPF_X:
|
2014-06-06 21:46:06 +00:00
|
|
|
/* test dst_reg, src_reg */
|
2021-01-14 18:17:42 +00:00
|
|
|
maybe_emit_mod(&prog, dst_reg, src_reg,
|
|
|
|
BPF_CLASS(insn->code) == BPF_JMP);
|
2019-01-26 17:26:06 +00:00
|
|
|
EMIT2(0x85, add_2reg(0xC0, dst_reg, src_reg));
|
net: filter: x86: internal BPF JIT
Maps all internal BPF instructions into x86_64 instructions.
This patch replaces original BPF x64 JIT with internal BPF x64 JIT.
sysctl net.core.bpf_jit_enable is reused as on/off switch.
Performance:
1. old BPF JIT and internal BPF JIT generate equivalent x86_64 code.
No performance difference is observed for filters that were JIT-able before
Example assembler code for BPF filter "tcpdump port 22"
original BPF -> old JIT: original BPF -> internal BPF -> new JIT:
0: push %rbp 0: push %rbp
1: mov %rsp,%rbp 1: mov %rsp,%rbp
4: sub $0x60,%rsp 4: sub $0x228,%rsp
8: mov %rbx,-0x8(%rbp) b: mov %rbx,-0x228(%rbp) // prologue
12: mov %r13,-0x220(%rbp)
19: mov %r14,-0x218(%rbp)
20: mov %r15,-0x210(%rbp)
27: xor %eax,%eax // clear A
c: xor %ebx,%ebx 29: xor %r13,%r13 // clear X
e: mov 0x68(%rdi),%r9d 2c: mov 0x68(%rdi),%r9d
12: sub 0x6c(%rdi),%r9d 30: sub 0x6c(%rdi),%r9d
16: mov 0xd8(%rdi),%r8 34: mov 0xd8(%rdi),%r10
3b: mov %rdi,%rbx
1d: mov $0xc,%esi 3e: mov $0xc,%esi
22: callq 0xffffffffe1021e15 43: callq 0xffffffffe102bd75
27: cmp $0x86dd,%eax 48: cmp $0x86dd,%rax
2c: jne 0x0000000000000069 4f: jne 0x000000000000009a
2e: mov $0x14,%esi 51: mov $0x14,%esi
33: callq 0xffffffffe1021e31 56: callq 0xffffffffe102bd91
38: cmp $0x84,%eax 5b: cmp $0x84,%rax
3d: je 0x0000000000000049 62: je 0x0000000000000074
3f: cmp $0x6,%eax 64: cmp $0x6,%rax
42: je 0x0000000000000049 68: je 0x0000000000000074
44: cmp $0x11,%eax 6a: cmp $0x11,%rax
47: jne 0x00000000000000c6 6e: jne 0x0000000000000117
49: mov $0x36,%esi 74: mov $0x36,%esi
4e: callq 0xffffffffe1021e15 79: callq 0xffffffffe102bd75
53: cmp $0x16,%eax 7e: cmp $0x16,%rax
56: je 0x00000000000000bf 82: je 0x0000000000000110
58: mov $0x38,%esi 88: mov $0x38,%esi
5d: callq 0xffffffffe1021e15 8d: callq 0xffffffffe102bd75
62: cmp $0x16,%eax 92: cmp $0x16,%rax
65: je 0x00000000000000bf 96: je 0x0000000000000110
67: jmp 0x00000000000000c6 98: jmp 0x0000000000000117
69: cmp $0x800,%eax 9a: cmp $0x800,%rax
6e: jne 0x00000000000000c6 a1: jne 0x0000000000000117
70: mov $0x17,%esi a3: mov $0x17,%esi
75: callq 0xffffffffe1021e31 a8: callq 0xffffffffe102bd91
7a: cmp $0x84,%eax ad: cmp $0x84,%rax
7f: je 0x000000000000008b b4: je 0x00000000000000c2
81: cmp $0x6,%eax b6: cmp $0x6,%rax
84: je 0x000000000000008b ba: je 0x00000000000000c2
86: cmp $0x11,%eax bc: cmp $0x11,%rax
89: jne 0x00000000000000c6 c0: jne 0x0000000000000117
8b: mov $0x14,%esi c2: mov $0x14,%esi
90: callq 0xffffffffe1021e15 c7: callq 0xffffffffe102bd75
95: test $0x1fff,%ax cc: test $0x1fff,%rax
99: jne 0x00000000000000c6 d3: jne 0x0000000000000117
d5: mov %rax,%r14
9b: mov $0xe,%esi d8: mov $0xe,%esi
a0: callq 0xffffffffe1021e44 dd: callq 0xffffffffe102bd91 // MSH
e2: and $0xf,%eax
e5: shl $0x2,%eax
e8: mov %rax,%r13
eb: mov %r14,%rax
ee: mov %r13,%rsi
a5: lea 0xe(%rbx),%esi f1: add $0xe,%esi
a8: callq 0xffffffffe1021e0d f4: callq 0xffffffffe102bd6d
ad: cmp $0x16,%eax f9: cmp $0x16,%rax
b0: je 0x00000000000000bf fd: je 0x0000000000000110
ff: mov %r13,%rsi
b2: lea 0x10(%rbx),%esi 102: add $0x10,%esi
b5: callq 0xffffffffe1021e0d 105: callq 0xffffffffe102bd6d
ba: cmp $0x16,%eax 10a: cmp $0x16,%rax
bd: jne 0x00000000000000c6 10e: jne 0x0000000000000117
bf: mov $0xffff,%eax 110: mov $0xffff,%eax
c4: jmp 0x00000000000000c8 115: jmp 0x000000000000011c
c6: xor %eax,%eax 117: mov $0x0,%eax
c8: mov -0x8(%rbp),%rbx 11c: mov -0x228(%rbp),%rbx // epilogue
cc: leaveq 123: mov -0x220(%rbp),%r13
cd: retq 12a: mov -0x218(%rbp),%r14
131: mov -0x210(%rbp),%r15
138: leaveq
139: retq
On fully cached SKBs both JITed functions take 12 nsec to execute.
BPF interpreter executes the program in 30 nsec.
The difference in generated assembler is due to the following:
Old BPF imlements LDX_MSH instruction via sk_load_byte_msh() helper function
inside bpf_jit.S.
New JIT removes the helper and does it explicitly, so ldx_msh cost
is the same for both JITs, but generated code looks longer.
New JIT has 4 registers to save, so prologue/epilogue are larger,
but the cost is within noise on x64.
Old JIT checks whether first insn clears A and if not emits 'xor %eax,%eax'.
New JIT clears %rax unconditionally.
2. old BPF JIT doesn't support ANC_NLATTR, ANC_PAY_OFFSET, ANC_RANDOM
extensions. New JIT supports all BPF extensions.
Performance of such filters improves 2-4 times depending on a filter.
The longer the filter the higher performance gain.
Synthetic benchmarks with many ancillary loads see 20x speedup
which seems to be the maximum gain from JIT
Notes:
. net.core.bpf_jit_enable=2 + tools/net/bpf_jit_disasm is still functional
and can be used to see generated assembler
. there are two jit_compile() functions and code flow for classic filters is:
sk_attach_filter() - load classic BPF
bpf_jit_compile() - try to JIT from classic BPF
sk_convert_filter() - convert classic to internal
bpf_int_jit_compile() - JIT from internal BPF
seccomp and tracing filters will just call bpf_int_jit_compile()
Signed-off-by: Alexei Starovoitov <ast@plumgrid.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
2014-05-14 02:50:46 +00:00
|
|
|
goto emit_cond_jmp;
|
|
|
|
|
|
|
|
case BPF_JMP | BPF_JSET | BPF_K:
|
2019-01-26 17:26:06 +00:00
|
|
|
case BPF_JMP32 | BPF_JSET | BPF_K:
|
2014-06-06 21:46:06 +00:00
|
|
|
/* test dst_reg, imm32 */
|
2021-10-06 19:41:35 +00:00
|
|
|
maybe_emit_1mod(&prog, dst_reg,
|
|
|
|
BPF_CLASS(insn->code) == BPF_JMP);
|
2014-06-06 21:46:06 +00:00
|
|
|
EMIT2_off32(0xF7, add_1reg(0xC0, dst_reg), imm32);
|
net: filter: x86: internal BPF JIT
Maps all internal BPF instructions into x86_64 instructions.
This patch replaces original BPF x64 JIT with internal BPF x64 JIT.
sysctl net.core.bpf_jit_enable is reused as on/off switch.
Performance:
1. old BPF JIT and internal BPF JIT generate equivalent x86_64 code.
No performance difference is observed for filters that were JIT-able before
Example assembler code for BPF filter "tcpdump port 22"
original BPF -> old JIT: original BPF -> internal BPF -> new JIT:
0: push %rbp 0: push %rbp
1: mov %rsp,%rbp 1: mov %rsp,%rbp
4: sub $0x60,%rsp 4: sub $0x228,%rsp
8: mov %rbx,-0x8(%rbp) b: mov %rbx,-0x228(%rbp) // prologue
12: mov %r13,-0x220(%rbp)
19: mov %r14,-0x218(%rbp)
20: mov %r15,-0x210(%rbp)
27: xor %eax,%eax // clear A
c: xor %ebx,%ebx 29: xor %r13,%r13 // clear X
e: mov 0x68(%rdi),%r9d 2c: mov 0x68(%rdi),%r9d
12: sub 0x6c(%rdi),%r9d 30: sub 0x6c(%rdi),%r9d
16: mov 0xd8(%rdi),%r8 34: mov 0xd8(%rdi),%r10
3b: mov %rdi,%rbx
1d: mov $0xc,%esi 3e: mov $0xc,%esi
22: callq 0xffffffffe1021e15 43: callq 0xffffffffe102bd75
27: cmp $0x86dd,%eax 48: cmp $0x86dd,%rax
2c: jne 0x0000000000000069 4f: jne 0x000000000000009a
2e: mov $0x14,%esi 51: mov $0x14,%esi
33: callq 0xffffffffe1021e31 56: callq 0xffffffffe102bd91
38: cmp $0x84,%eax 5b: cmp $0x84,%rax
3d: je 0x0000000000000049 62: je 0x0000000000000074
3f: cmp $0x6,%eax 64: cmp $0x6,%rax
42: je 0x0000000000000049 68: je 0x0000000000000074
44: cmp $0x11,%eax 6a: cmp $0x11,%rax
47: jne 0x00000000000000c6 6e: jne 0x0000000000000117
49: mov $0x36,%esi 74: mov $0x36,%esi
4e: callq 0xffffffffe1021e15 79: callq 0xffffffffe102bd75
53: cmp $0x16,%eax 7e: cmp $0x16,%rax
56: je 0x00000000000000bf 82: je 0x0000000000000110
58: mov $0x38,%esi 88: mov $0x38,%esi
5d: callq 0xffffffffe1021e15 8d: callq 0xffffffffe102bd75
62: cmp $0x16,%eax 92: cmp $0x16,%rax
65: je 0x00000000000000bf 96: je 0x0000000000000110
67: jmp 0x00000000000000c6 98: jmp 0x0000000000000117
69: cmp $0x800,%eax 9a: cmp $0x800,%rax
6e: jne 0x00000000000000c6 a1: jne 0x0000000000000117
70: mov $0x17,%esi a3: mov $0x17,%esi
75: callq 0xffffffffe1021e31 a8: callq 0xffffffffe102bd91
7a: cmp $0x84,%eax ad: cmp $0x84,%rax
7f: je 0x000000000000008b b4: je 0x00000000000000c2
81: cmp $0x6,%eax b6: cmp $0x6,%rax
84: je 0x000000000000008b ba: je 0x00000000000000c2
86: cmp $0x11,%eax bc: cmp $0x11,%rax
89: jne 0x00000000000000c6 c0: jne 0x0000000000000117
8b: mov $0x14,%esi c2: mov $0x14,%esi
90: callq 0xffffffffe1021e15 c7: callq 0xffffffffe102bd75
95: test $0x1fff,%ax cc: test $0x1fff,%rax
99: jne 0x00000000000000c6 d3: jne 0x0000000000000117
d5: mov %rax,%r14
9b: mov $0xe,%esi d8: mov $0xe,%esi
a0: callq 0xffffffffe1021e44 dd: callq 0xffffffffe102bd91 // MSH
e2: and $0xf,%eax
e5: shl $0x2,%eax
e8: mov %rax,%r13
eb: mov %r14,%rax
ee: mov %r13,%rsi
a5: lea 0xe(%rbx),%esi f1: add $0xe,%esi
a8: callq 0xffffffffe1021e0d f4: callq 0xffffffffe102bd6d
ad: cmp $0x16,%eax f9: cmp $0x16,%rax
b0: je 0x00000000000000bf fd: je 0x0000000000000110
ff: mov %r13,%rsi
b2: lea 0x10(%rbx),%esi 102: add $0x10,%esi
b5: callq 0xffffffffe1021e0d 105: callq 0xffffffffe102bd6d
ba: cmp $0x16,%eax 10a: cmp $0x16,%rax
bd: jne 0x00000000000000c6 10e: jne 0x0000000000000117
bf: mov $0xffff,%eax 110: mov $0xffff,%eax
c4: jmp 0x00000000000000c8 115: jmp 0x000000000000011c
c6: xor %eax,%eax 117: mov $0x0,%eax
c8: mov -0x8(%rbp),%rbx 11c: mov -0x228(%rbp),%rbx // epilogue
cc: leaveq 123: mov -0x220(%rbp),%r13
cd: retq 12a: mov -0x218(%rbp),%r14
131: mov -0x210(%rbp),%r15
138: leaveq
139: retq
On fully cached SKBs both JITed functions take 12 nsec to execute.
BPF interpreter executes the program in 30 nsec.
The difference in generated assembler is due to the following:
Old BPF imlements LDX_MSH instruction via sk_load_byte_msh() helper function
inside bpf_jit.S.
New JIT removes the helper and does it explicitly, so ldx_msh cost
is the same for both JITs, but generated code looks longer.
New JIT has 4 registers to save, so prologue/epilogue are larger,
but the cost is within noise on x64.
Old JIT checks whether first insn clears A and if not emits 'xor %eax,%eax'.
New JIT clears %rax unconditionally.
2. old BPF JIT doesn't support ANC_NLATTR, ANC_PAY_OFFSET, ANC_RANDOM
extensions. New JIT supports all BPF extensions.
Performance of such filters improves 2-4 times depending on a filter.
The longer the filter the higher performance gain.
Synthetic benchmarks with many ancillary loads see 20x speedup
which seems to be the maximum gain from JIT
Notes:
. net.core.bpf_jit_enable=2 + tools/net/bpf_jit_disasm is still functional
and can be used to see generated assembler
. there are two jit_compile() functions and code flow for classic filters is:
sk_attach_filter() - load classic BPF
bpf_jit_compile() - try to JIT from classic BPF
sk_convert_filter() - convert classic to internal
bpf_int_jit_compile() - JIT from internal BPF
seccomp and tracing filters will just call bpf_int_jit_compile()
Signed-off-by: Alexei Starovoitov <ast@plumgrid.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
2014-05-14 02:50:46 +00:00
|
|
|
goto emit_cond_jmp;
|
|
|
|
|
|
|
|
case BPF_JMP | BPF_JEQ | BPF_K:
|
|
|
|
case BPF_JMP | BPF_JNE | BPF_K:
|
|
|
|
case BPF_JMP | BPF_JGT | BPF_K:
|
2017-08-09 23:39:56 +00:00
|
|
|
case BPF_JMP | BPF_JLT | BPF_K:
|
net: filter: x86: internal BPF JIT
Maps all internal BPF instructions into x86_64 instructions.
This patch replaces original BPF x64 JIT with internal BPF x64 JIT.
sysctl net.core.bpf_jit_enable is reused as on/off switch.
Performance:
1. old BPF JIT and internal BPF JIT generate equivalent x86_64 code.
No performance difference is observed for filters that were JIT-able before
Example assembler code for BPF filter "tcpdump port 22"
original BPF -> old JIT: original BPF -> internal BPF -> new JIT:
0: push %rbp 0: push %rbp
1: mov %rsp,%rbp 1: mov %rsp,%rbp
4: sub $0x60,%rsp 4: sub $0x228,%rsp
8: mov %rbx,-0x8(%rbp) b: mov %rbx,-0x228(%rbp) // prologue
12: mov %r13,-0x220(%rbp)
19: mov %r14,-0x218(%rbp)
20: mov %r15,-0x210(%rbp)
27: xor %eax,%eax // clear A
c: xor %ebx,%ebx 29: xor %r13,%r13 // clear X
e: mov 0x68(%rdi),%r9d 2c: mov 0x68(%rdi),%r9d
12: sub 0x6c(%rdi),%r9d 30: sub 0x6c(%rdi),%r9d
16: mov 0xd8(%rdi),%r8 34: mov 0xd8(%rdi),%r10
3b: mov %rdi,%rbx
1d: mov $0xc,%esi 3e: mov $0xc,%esi
22: callq 0xffffffffe1021e15 43: callq 0xffffffffe102bd75
27: cmp $0x86dd,%eax 48: cmp $0x86dd,%rax
2c: jne 0x0000000000000069 4f: jne 0x000000000000009a
2e: mov $0x14,%esi 51: mov $0x14,%esi
33: callq 0xffffffffe1021e31 56: callq 0xffffffffe102bd91
38: cmp $0x84,%eax 5b: cmp $0x84,%rax
3d: je 0x0000000000000049 62: je 0x0000000000000074
3f: cmp $0x6,%eax 64: cmp $0x6,%rax
42: je 0x0000000000000049 68: je 0x0000000000000074
44: cmp $0x11,%eax 6a: cmp $0x11,%rax
47: jne 0x00000000000000c6 6e: jne 0x0000000000000117
49: mov $0x36,%esi 74: mov $0x36,%esi
4e: callq 0xffffffffe1021e15 79: callq 0xffffffffe102bd75
53: cmp $0x16,%eax 7e: cmp $0x16,%rax
56: je 0x00000000000000bf 82: je 0x0000000000000110
58: mov $0x38,%esi 88: mov $0x38,%esi
5d: callq 0xffffffffe1021e15 8d: callq 0xffffffffe102bd75
62: cmp $0x16,%eax 92: cmp $0x16,%rax
65: je 0x00000000000000bf 96: je 0x0000000000000110
67: jmp 0x00000000000000c6 98: jmp 0x0000000000000117
69: cmp $0x800,%eax 9a: cmp $0x800,%rax
6e: jne 0x00000000000000c6 a1: jne 0x0000000000000117
70: mov $0x17,%esi a3: mov $0x17,%esi
75: callq 0xffffffffe1021e31 a8: callq 0xffffffffe102bd91
7a: cmp $0x84,%eax ad: cmp $0x84,%rax
7f: je 0x000000000000008b b4: je 0x00000000000000c2
81: cmp $0x6,%eax b6: cmp $0x6,%rax
84: je 0x000000000000008b ba: je 0x00000000000000c2
86: cmp $0x11,%eax bc: cmp $0x11,%rax
89: jne 0x00000000000000c6 c0: jne 0x0000000000000117
8b: mov $0x14,%esi c2: mov $0x14,%esi
90: callq 0xffffffffe1021e15 c7: callq 0xffffffffe102bd75
95: test $0x1fff,%ax cc: test $0x1fff,%rax
99: jne 0x00000000000000c6 d3: jne 0x0000000000000117
d5: mov %rax,%r14
9b: mov $0xe,%esi d8: mov $0xe,%esi
a0: callq 0xffffffffe1021e44 dd: callq 0xffffffffe102bd91 // MSH
e2: and $0xf,%eax
e5: shl $0x2,%eax
e8: mov %rax,%r13
eb: mov %r14,%rax
ee: mov %r13,%rsi
a5: lea 0xe(%rbx),%esi f1: add $0xe,%esi
a8: callq 0xffffffffe1021e0d f4: callq 0xffffffffe102bd6d
ad: cmp $0x16,%eax f9: cmp $0x16,%rax
b0: je 0x00000000000000bf fd: je 0x0000000000000110
ff: mov %r13,%rsi
b2: lea 0x10(%rbx),%esi 102: add $0x10,%esi
b5: callq 0xffffffffe1021e0d 105: callq 0xffffffffe102bd6d
ba: cmp $0x16,%eax 10a: cmp $0x16,%rax
bd: jne 0x00000000000000c6 10e: jne 0x0000000000000117
bf: mov $0xffff,%eax 110: mov $0xffff,%eax
c4: jmp 0x00000000000000c8 115: jmp 0x000000000000011c
c6: xor %eax,%eax 117: mov $0x0,%eax
c8: mov -0x8(%rbp),%rbx 11c: mov -0x228(%rbp),%rbx // epilogue
cc: leaveq 123: mov -0x220(%rbp),%r13
cd: retq 12a: mov -0x218(%rbp),%r14
131: mov -0x210(%rbp),%r15
138: leaveq
139: retq
On fully cached SKBs both JITed functions take 12 nsec to execute.
BPF interpreter executes the program in 30 nsec.
The difference in generated assembler is due to the following:
Old BPF imlements LDX_MSH instruction via sk_load_byte_msh() helper function
inside bpf_jit.S.
New JIT removes the helper and does it explicitly, so ldx_msh cost
is the same for both JITs, but generated code looks longer.
New JIT has 4 registers to save, so prologue/epilogue are larger,
but the cost is within noise on x64.
Old JIT checks whether first insn clears A and if not emits 'xor %eax,%eax'.
New JIT clears %rax unconditionally.
2. old BPF JIT doesn't support ANC_NLATTR, ANC_PAY_OFFSET, ANC_RANDOM
extensions. New JIT supports all BPF extensions.
Performance of such filters improves 2-4 times depending on a filter.
The longer the filter the higher performance gain.
Synthetic benchmarks with many ancillary loads see 20x speedup
which seems to be the maximum gain from JIT
Notes:
. net.core.bpf_jit_enable=2 + tools/net/bpf_jit_disasm is still functional
and can be used to see generated assembler
. there are two jit_compile() functions and code flow for classic filters is:
sk_attach_filter() - load classic BPF
bpf_jit_compile() - try to JIT from classic BPF
sk_convert_filter() - convert classic to internal
bpf_int_jit_compile() - JIT from internal BPF
seccomp and tracing filters will just call bpf_int_jit_compile()
Signed-off-by: Alexei Starovoitov <ast@plumgrid.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
2014-05-14 02:50:46 +00:00
|
|
|
case BPF_JMP | BPF_JGE | BPF_K:
|
2017-08-09 23:39:56 +00:00
|
|
|
case BPF_JMP | BPF_JLE | BPF_K:
|
net: filter: x86: internal BPF JIT
Maps all internal BPF instructions into x86_64 instructions.
This patch replaces original BPF x64 JIT with internal BPF x64 JIT.
sysctl net.core.bpf_jit_enable is reused as on/off switch.
Performance:
1. old BPF JIT and internal BPF JIT generate equivalent x86_64 code.
No performance difference is observed for filters that were JIT-able before
Example assembler code for BPF filter "tcpdump port 22"
original BPF -> old JIT: original BPF -> internal BPF -> new JIT:
0: push %rbp 0: push %rbp
1: mov %rsp,%rbp 1: mov %rsp,%rbp
4: sub $0x60,%rsp 4: sub $0x228,%rsp
8: mov %rbx,-0x8(%rbp) b: mov %rbx,-0x228(%rbp) // prologue
12: mov %r13,-0x220(%rbp)
19: mov %r14,-0x218(%rbp)
20: mov %r15,-0x210(%rbp)
27: xor %eax,%eax // clear A
c: xor %ebx,%ebx 29: xor %r13,%r13 // clear X
e: mov 0x68(%rdi),%r9d 2c: mov 0x68(%rdi),%r9d
12: sub 0x6c(%rdi),%r9d 30: sub 0x6c(%rdi),%r9d
16: mov 0xd8(%rdi),%r8 34: mov 0xd8(%rdi),%r10
3b: mov %rdi,%rbx
1d: mov $0xc,%esi 3e: mov $0xc,%esi
22: callq 0xffffffffe1021e15 43: callq 0xffffffffe102bd75
27: cmp $0x86dd,%eax 48: cmp $0x86dd,%rax
2c: jne 0x0000000000000069 4f: jne 0x000000000000009a
2e: mov $0x14,%esi 51: mov $0x14,%esi
33: callq 0xffffffffe1021e31 56: callq 0xffffffffe102bd91
38: cmp $0x84,%eax 5b: cmp $0x84,%rax
3d: je 0x0000000000000049 62: je 0x0000000000000074
3f: cmp $0x6,%eax 64: cmp $0x6,%rax
42: je 0x0000000000000049 68: je 0x0000000000000074
44: cmp $0x11,%eax 6a: cmp $0x11,%rax
47: jne 0x00000000000000c6 6e: jne 0x0000000000000117
49: mov $0x36,%esi 74: mov $0x36,%esi
4e: callq 0xffffffffe1021e15 79: callq 0xffffffffe102bd75
53: cmp $0x16,%eax 7e: cmp $0x16,%rax
56: je 0x00000000000000bf 82: je 0x0000000000000110
58: mov $0x38,%esi 88: mov $0x38,%esi
5d: callq 0xffffffffe1021e15 8d: callq 0xffffffffe102bd75
62: cmp $0x16,%eax 92: cmp $0x16,%rax
65: je 0x00000000000000bf 96: je 0x0000000000000110
67: jmp 0x00000000000000c6 98: jmp 0x0000000000000117
69: cmp $0x800,%eax 9a: cmp $0x800,%rax
6e: jne 0x00000000000000c6 a1: jne 0x0000000000000117
70: mov $0x17,%esi a3: mov $0x17,%esi
75: callq 0xffffffffe1021e31 a8: callq 0xffffffffe102bd91
7a: cmp $0x84,%eax ad: cmp $0x84,%rax
7f: je 0x000000000000008b b4: je 0x00000000000000c2
81: cmp $0x6,%eax b6: cmp $0x6,%rax
84: je 0x000000000000008b ba: je 0x00000000000000c2
86: cmp $0x11,%eax bc: cmp $0x11,%rax
89: jne 0x00000000000000c6 c0: jne 0x0000000000000117
8b: mov $0x14,%esi c2: mov $0x14,%esi
90: callq 0xffffffffe1021e15 c7: callq 0xffffffffe102bd75
95: test $0x1fff,%ax cc: test $0x1fff,%rax
99: jne 0x00000000000000c6 d3: jne 0x0000000000000117
d5: mov %rax,%r14
9b: mov $0xe,%esi d8: mov $0xe,%esi
a0: callq 0xffffffffe1021e44 dd: callq 0xffffffffe102bd91 // MSH
e2: and $0xf,%eax
e5: shl $0x2,%eax
e8: mov %rax,%r13
eb: mov %r14,%rax
ee: mov %r13,%rsi
a5: lea 0xe(%rbx),%esi f1: add $0xe,%esi
a8: callq 0xffffffffe1021e0d f4: callq 0xffffffffe102bd6d
ad: cmp $0x16,%eax f9: cmp $0x16,%rax
b0: je 0x00000000000000bf fd: je 0x0000000000000110
ff: mov %r13,%rsi
b2: lea 0x10(%rbx),%esi 102: add $0x10,%esi
b5: callq 0xffffffffe1021e0d 105: callq 0xffffffffe102bd6d
ba: cmp $0x16,%eax 10a: cmp $0x16,%rax
bd: jne 0x00000000000000c6 10e: jne 0x0000000000000117
bf: mov $0xffff,%eax 110: mov $0xffff,%eax
c4: jmp 0x00000000000000c8 115: jmp 0x000000000000011c
c6: xor %eax,%eax 117: mov $0x0,%eax
c8: mov -0x8(%rbp),%rbx 11c: mov -0x228(%rbp),%rbx // epilogue
cc: leaveq 123: mov -0x220(%rbp),%r13
cd: retq 12a: mov -0x218(%rbp),%r14
131: mov -0x210(%rbp),%r15
138: leaveq
139: retq
On fully cached SKBs both JITed functions take 12 nsec to execute.
BPF interpreter executes the program in 30 nsec.
The difference in generated assembler is due to the following:
Old BPF imlements LDX_MSH instruction via sk_load_byte_msh() helper function
inside bpf_jit.S.
New JIT removes the helper and does it explicitly, so ldx_msh cost
is the same for both JITs, but generated code looks longer.
New JIT has 4 registers to save, so prologue/epilogue are larger,
but the cost is within noise on x64.
Old JIT checks whether first insn clears A and if not emits 'xor %eax,%eax'.
New JIT clears %rax unconditionally.
2. old BPF JIT doesn't support ANC_NLATTR, ANC_PAY_OFFSET, ANC_RANDOM
extensions. New JIT supports all BPF extensions.
Performance of such filters improves 2-4 times depending on a filter.
The longer the filter the higher performance gain.
Synthetic benchmarks with many ancillary loads see 20x speedup
which seems to be the maximum gain from JIT
Notes:
. net.core.bpf_jit_enable=2 + tools/net/bpf_jit_disasm is still functional
and can be used to see generated assembler
. there are two jit_compile() functions and code flow for classic filters is:
sk_attach_filter() - load classic BPF
bpf_jit_compile() - try to JIT from classic BPF
sk_convert_filter() - convert classic to internal
bpf_int_jit_compile() - JIT from internal BPF
seccomp and tracing filters will just call bpf_int_jit_compile()
Signed-off-by: Alexei Starovoitov <ast@plumgrid.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
2014-05-14 02:50:46 +00:00
|
|
|
case BPF_JMP | BPF_JSGT | BPF_K:
|
2017-08-09 23:39:56 +00:00
|
|
|
case BPF_JMP | BPF_JSLT | BPF_K:
|
net: filter: x86: internal BPF JIT
Maps all internal BPF instructions into x86_64 instructions.
This patch replaces original BPF x64 JIT with internal BPF x64 JIT.
sysctl net.core.bpf_jit_enable is reused as on/off switch.
Performance:
1. old BPF JIT and internal BPF JIT generate equivalent x86_64 code.
No performance difference is observed for filters that were JIT-able before
Example assembler code for BPF filter "tcpdump port 22"
original BPF -> old JIT: original BPF -> internal BPF -> new JIT:
0: push %rbp 0: push %rbp
1: mov %rsp,%rbp 1: mov %rsp,%rbp
4: sub $0x60,%rsp 4: sub $0x228,%rsp
8: mov %rbx,-0x8(%rbp) b: mov %rbx,-0x228(%rbp) // prologue
12: mov %r13,-0x220(%rbp)
19: mov %r14,-0x218(%rbp)
20: mov %r15,-0x210(%rbp)
27: xor %eax,%eax // clear A
c: xor %ebx,%ebx 29: xor %r13,%r13 // clear X
e: mov 0x68(%rdi),%r9d 2c: mov 0x68(%rdi),%r9d
12: sub 0x6c(%rdi),%r9d 30: sub 0x6c(%rdi),%r9d
16: mov 0xd8(%rdi),%r8 34: mov 0xd8(%rdi),%r10
3b: mov %rdi,%rbx
1d: mov $0xc,%esi 3e: mov $0xc,%esi
22: callq 0xffffffffe1021e15 43: callq 0xffffffffe102bd75
27: cmp $0x86dd,%eax 48: cmp $0x86dd,%rax
2c: jne 0x0000000000000069 4f: jne 0x000000000000009a
2e: mov $0x14,%esi 51: mov $0x14,%esi
33: callq 0xffffffffe1021e31 56: callq 0xffffffffe102bd91
38: cmp $0x84,%eax 5b: cmp $0x84,%rax
3d: je 0x0000000000000049 62: je 0x0000000000000074
3f: cmp $0x6,%eax 64: cmp $0x6,%rax
42: je 0x0000000000000049 68: je 0x0000000000000074
44: cmp $0x11,%eax 6a: cmp $0x11,%rax
47: jne 0x00000000000000c6 6e: jne 0x0000000000000117
49: mov $0x36,%esi 74: mov $0x36,%esi
4e: callq 0xffffffffe1021e15 79: callq 0xffffffffe102bd75
53: cmp $0x16,%eax 7e: cmp $0x16,%rax
56: je 0x00000000000000bf 82: je 0x0000000000000110
58: mov $0x38,%esi 88: mov $0x38,%esi
5d: callq 0xffffffffe1021e15 8d: callq 0xffffffffe102bd75
62: cmp $0x16,%eax 92: cmp $0x16,%rax
65: je 0x00000000000000bf 96: je 0x0000000000000110
67: jmp 0x00000000000000c6 98: jmp 0x0000000000000117
69: cmp $0x800,%eax 9a: cmp $0x800,%rax
6e: jne 0x00000000000000c6 a1: jne 0x0000000000000117
70: mov $0x17,%esi a3: mov $0x17,%esi
75: callq 0xffffffffe1021e31 a8: callq 0xffffffffe102bd91
7a: cmp $0x84,%eax ad: cmp $0x84,%rax
7f: je 0x000000000000008b b4: je 0x00000000000000c2
81: cmp $0x6,%eax b6: cmp $0x6,%rax
84: je 0x000000000000008b ba: je 0x00000000000000c2
86: cmp $0x11,%eax bc: cmp $0x11,%rax
89: jne 0x00000000000000c6 c0: jne 0x0000000000000117
8b: mov $0x14,%esi c2: mov $0x14,%esi
90: callq 0xffffffffe1021e15 c7: callq 0xffffffffe102bd75
95: test $0x1fff,%ax cc: test $0x1fff,%rax
99: jne 0x00000000000000c6 d3: jne 0x0000000000000117
d5: mov %rax,%r14
9b: mov $0xe,%esi d8: mov $0xe,%esi
a0: callq 0xffffffffe1021e44 dd: callq 0xffffffffe102bd91 // MSH
e2: and $0xf,%eax
e5: shl $0x2,%eax
e8: mov %rax,%r13
eb: mov %r14,%rax
ee: mov %r13,%rsi
a5: lea 0xe(%rbx),%esi f1: add $0xe,%esi
a8: callq 0xffffffffe1021e0d f4: callq 0xffffffffe102bd6d
ad: cmp $0x16,%eax f9: cmp $0x16,%rax
b0: je 0x00000000000000bf fd: je 0x0000000000000110
ff: mov %r13,%rsi
b2: lea 0x10(%rbx),%esi 102: add $0x10,%esi
b5: callq 0xffffffffe1021e0d 105: callq 0xffffffffe102bd6d
ba: cmp $0x16,%eax 10a: cmp $0x16,%rax
bd: jne 0x00000000000000c6 10e: jne 0x0000000000000117
bf: mov $0xffff,%eax 110: mov $0xffff,%eax
c4: jmp 0x00000000000000c8 115: jmp 0x000000000000011c
c6: xor %eax,%eax 117: mov $0x0,%eax
c8: mov -0x8(%rbp),%rbx 11c: mov -0x228(%rbp),%rbx // epilogue
cc: leaveq 123: mov -0x220(%rbp),%r13
cd: retq 12a: mov -0x218(%rbp),%r14
131: mov -0x210(%rbp),%r15
138: leaveq
139: retq
On fully cached SKBs both JITed functions take 12 nsec to execute.
BPF interpreter executes the program in 30 nsec.
The difference in generated assembler is due to the following:
Old BPF imlements LDX_MSH instruction via sk_load_byte_msh() helper function
inside bpf_jit.S.
New JIT removes the helper and does it explicitly, so ldx_msh cost
is the same for both JITs, but generated code looks longer.
New JIT has 4 registers to save, so prologue/epilogue are larger,
but the cost is within noise on x64.
Old JIT checks whether first insn clears A and if not emits 'xor %eax,%eax'.
New JIT clears %rax unconditionally.
2. old BPF JIT doesn't support ANC_NLATTR, ANC_PAY_OFFSET, ANC_RANDOM
extensions. New JIT supports all BPF extensions.
Performance of such filters improves 2-4 times depending on a filter.
The longer the filter the higher performance gain.
Synthetic benchmarks with many ancillary loads see 20x speedup
which seems to be the maximum gain from JIT
Notes:
. net.core.bpf_jit_enable=2 + tools/net/bpf_jit_disasm is still functional
and can be used to see generated assembler
. there are two jit_compile() functions and code flow for classic filters is:
sk_attach_filter() - load classic BPF
bpf_jit_compile() - try to JIT from classic BPF
sk_convert_filter() - convert classic to internal
bpf_int_jit_compile() - JIT from internal BPF
seccomp and tracing filters will just call bpf_int_jit_compile()
Signed-off-by: Alexei Starovoitov <ast@plumgrid.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
2014-05-14 02:50:46 +00:00
|
|
|
case BPF_JMP | BPF_JSGE | BPF_K:
|
2017-08-09 23:39:56 +00:00
|
|
|
case BPF_JMP | BPF_JSLE | BPF_K:
|
2019-01-26 17:26:06 +00:00
|
|
|
case BPF_JMP32 | BPF_JEQ | BPF_K:
|
|
|
|
case BPF_JMP32 | BPF_JNE | BPF_K:
|
|
|
|
case BPF_JMP32 | BPF_JGT | BPF_K:
|
|
|
|
case BPF_JMP32 | BPF_JLT | BPF_K:
|
|
|
|
case BPF_JMP32 | BPF_JGE | BPF_K:
|
|
|
|
case BPF_JMP32 | BPF_JLE | BPF_K:
|
|
|
|
case BPF_JMP32 | BPF_JSGT | BPF_K:
|
|
|
|
case BPF_JMP32 | BPF_JSLT | BPF_K:
|
|
|
|
case BPF_JMP32 | BPF_JSGE | BPF_K:
|
|
|
|
case BPF_JMP32 | BPF_JSLE | BPF_K:
|
2019-10-02 23:45:11 +00:00
|
|
|
/* test dst_reg, dst_reg to save one extra byte */
|
|
|
|
if (imm32 == 0) {
|
2021-01-14 18:17:42 +00:00
|
|
|
maybe_emit_mod(&prog, dst_reg, dst_reg,
|
|
|
|
BPF_CLASS(insn->code) == BPF_JMP);
|
2019-10-02 23:45:11 +00:00
|
|
|
EMIT2(0x85, add_2reg(0xC0, dst_reg, dst_reg));
|
|
|
|
goto emit_cond_jmp;
|
|
|
|
}
|
|
|
|
|
2014-06-06 21:46:06 +00:00
|
|
|
/* cmp dst_reg, imm8/32 */
|
2021-10-06 19:41:35 +00:00
|
|
|
maybe_emit_1mod(&prog, dst_reg,
|
|
|
|
BPF_CLASS(insn->code) == BPF_JMP);
|
net: filter: x86: internal BPF JIT
Maps all internal BPF instructions into x86_64 instructions.
This patch replaces original BPF x64 JIT with internal BPF x64 JIT.
sysctl net.core.bpf_jit_enable is reused as on/off switch.
Performance:
1. old BPF JIT and internal BPF JIT generate equivalent x86_64 code.
No performance difference is observed for filters that were JIT-able before
Example assembler code for BPF filter "tcpdump port 22"
original BPF -> old JIT: original BPF -> internal BPF -> new JIT:
0: push %rbp 0: push %rbp
1: mov %rsp,%rbp 1: mov %rsp,%rbp
4: sub $0x60,%rsp 4: sub $0x228,%rsp
8: mov %rbx,-0x8(%rbp) b: mov %rbx,-0x228(%rbp) // prologue
12: mov %r13,-0x220(%rbp)
19: mov %r14,-0x218(%rbp)
20: mov %r15,-0x210(%rbp)
27: xor %eax,%eax // clear A
c: xor %ebx,%ebx 29: xor %r13,%r13 // clear X
e: mov 0x68(%rdi),%r9d 2c: mov 0x68(%rdi),%r9d
12: sub 0x6c(%rdi),%r9d 30: sub 0x6c(%rdi),%r9d
16: mov 0xd8(%rdi),%r8 34: mov 0xd8(%rdi),%r10
3b: mov %rdi,%rbx
1d: mov $0xc,%esi 3e: mov $0xc,%esi
22: callq 0xffffffffe1021e15 43: callq 0xffffffffe102bd75
27: cmp $0x86dd,%eax 48: cmp $0x86dd,%rax
2c: jne 0x0000000000000069 4f: jne 0x000000000000009a
2e: mov $0x14,%esi 51: mov $0x14,%esi
33: callq 0xffffffffe1021e31 56: callq 0xffffffffe102bd91
38: cmp $0x84,%eax 5b: cmp $0x84,%rax
3d: je 0x0000000000000049 62: je 0x0000000000000074
3f: cmp $0x6,%eax 64: cmp $0x6,%rax
42: je 0x0000000000000049 68: je 0x0000000000000074
44: cmp $0x11,%eax 6a: cmp $0x11,%rax
47: jne 0x00000000000000c6 6e: jne 0x0000000000000117
49: mov $0x36,%esi 74: mov $0x36,%esi
4e: callq 0xffffffffe1021e15 79: callq 0xffffffffe102bd75
53: cmp $0x16,%eax 7e: cmp $0x16,%rax
56: je 0x00000000000000bf 82: je 0x0000000000000110
58: mov $0x38,%esi 88: mov $0x38,%esi
5d: callq 0xffffffffe1021e15 8d: callq 0xffffffffe102bd75
62: cmp $0x16,%eax 92: cmp $0x16,%rax
65: je 0x00000000000000bf 96: je 0x0000000000000110
67: jmp 0x00000000000000c6 98: jmp 0x0000000000000117
69: cmp $0x800,%eax 9a: cmp $0x800,%rax
6e: jne 0x00000000000000c6 a1: jne 0x0000000000000117
70: mov $0x17,%esi a3: mov $0x17,%esi
75: callq 0xffffffffe1021e31 a8: callq 0xffffffffe102bd91
7a: cmp $0x84,%eax ad: cmp $0x84,%rax
7f: je 0x000000000000008b b4: je 0x00000000000000c2
81: cmp $0x6,%eax b6: cmp $0x6,%rax
84: je 0x000000000000008b ba: je 0x00000000000000c2
86: cmp $0x11,%eax bc: cmp $0x11,%rax
89: jne 0x00000000000000c6 c0: jne 0x0000000000000117
8b: mov $0x14,%esi c2: mov $0x14,%esi
90: callq 0xffffffffe1021e15 c7: callq 0xffffffffe102bd75
95: test $0x1fff,%ax cc: test $0x1fff,%rax
99: jne 0x00000000000000c6 d3: jne 0x0000000000000117
d5: mov %rax,%r14
9b: mov $0xe,%esi d8: mov $0xe,%esi
a0: callq 0xffffffffe1021e44 dd: callq 0xffffffffe102bd91 // MSH
e2: and $0xf,%eax
e5: shl $0x2,%eax
e8: mov %rax,%r13
eb: mov %r14,%rax
ee: mov %r13,%rsi
a5: lea 0xe(%rbx),%esi f1: add $0xe,%esi
a8: callq 0xffffffffe1021e0d f4: callq 0xffffffffe102bd6d
ad: cmp $0x16,%eax f9: cmp $0x16,%rax
b0: je 0x00000000000000bf fd: je 0x0000000000000110
ff: mov %r13,%rsi
b2: lea 0x10(%rbx),%esi 102: add $0x10,%esi
b5: callq 0xffffffffe1021e0d 105: callq 0xffffffffe102bd6d
ba: cmp $0x16,%eax 10a: cmp $0x16,%rax
bd: jne 0x00000000000000c6 10e: jne 0x0000000000000117
bf: mov $0xffff,%eax 110: mov $0xffff,%eax
c4: jmp 0x00000000000000c8 115: jmp 0x000000000000011c
c6: xor %eax,%eax 117: mov $0x0,%eax
c8: mov -0x8(%rbp),%rbx 11c: mov -0x228(%rbp),%rbx // epilogue
cc: leaveq 123: mov -0x220(%rbp),%r13
cd: retq 12a: mov -0x218(%rbp),%r14
131: mov -0x210(%rbp),%r15
138: leaveq
139: retq
On fully cached SKBs both JITed functions take 12 nsec to execute.
BPF interpreter executes the program in 30 nsec.
The difference in generated assembler is due to the following:
Old BPF imlements LDX_MSH instruction via sk_load_byte_msh() helper function
inside bpf_jit.S.
New JIT removes the helper and does it explicitly, so ldx_msh cost
is the same for both JITs, but generated code looks longer.
New JIT has 4 registers to save, so prologue/epilogue are larger,
but the cost is within noise on x64.
Old JIT checks whether first insn clears A and if not emits 'xor %eax,%eax'.
New JIT clears %rax unconditionally.
2. old BPF JIT doesn't support ANC_NLATTR, ANC_PAY_OFFSET, ANC_RANDOM
extensions. New JIT supports all BPF extensions.
Performance of such filters improves 2-4 times depending on a filter.
The longer the filter the higher performance gain.
Synthetic benchmarks with many ancillary loads see 20x speedup
which seems to be the maximum gain from JIT
Notes:
. net.core.bpf_jit_enable=2 + tools/net/bpf_jit_disasm is still functional
and can be used to see generated assembler
. there are two jit_compile() functions and code flow for classic filters is:
sk_attach_filter() - load classic BPF
bpf_jit_compile() - try to JIT from classic BPF
sk_convert_filter() - convert classic to internal
bpf_int_jit_compile() - JIT from internal BPF
seccomp and tracing filters will just call bpf_int_jit_compile()
Signed-off-by: Alexei Starovoitov <ast@plumgrid.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
2014-05-14 02:50:46 +00:00
|
|
|
|
2014-06-06 21:46:06 +00:00
|
|
|
if (is_imm8(imm32))
|
|
|
|
EMIT3(0x83, add_1reg(0xF8, dst_reg), imm32);
|
net: filter: x86: internal BPF JIT
Maps all internal BPF instructions into x86_64 instructions.
This patch replaces original BPF x64 JIT with internal BPF x64 JIT.
sysctl net.core.bpf_jit_enable is reused as on/off switch.
Performance:
1. old BPF JIT and internal BPF JIT generate equivalent x86_64 code.
No performance difference is observed for filters that were JIT-able before
Example assembler code for BPF filter "tcpdump port 22"
original BPF -> old JIT: original BPF -> internal BPF -> new JIT:
0: push %rbp 0: push %rbp
1: mov %rsp,%rbp 1: mov %rsp,%rbp
4: sub $0x60,%rsp 4: sub $0x228,%rsp
8: mov %rbx,-0x8(%rbp) b: mov %rbx,-0x228(%rbp) // prologue
12: mov %r13,-0x220(%rbp)
19: mov %r14,-0x218(%rbp)
20: mov %r15,-0x210(%rbp)
27: xor %eax,%eax // clear A
c: xor %ebx,%ebx 29: xor %r13,%r13 // clear X
e: mov 0x68(%rdi),%r9d 2c: mov 0x68(%rdi),%r9d
12: sub 0x6c(%rdi),%r9d 30: sub 0x6c(%rdi),%r9d
16: mov 0xd8(%rdi),%r8 34: mov 0xd8(%rdi),%r10
3b: mov %rdi,%rbx
1d: mov $0xc,%esi 3e: mov $0xc,%esi
22: callq 0xffffffffe1021e15 43: callq 0xffffffffe102bd75
27: cmp $0x86dd,%eax 48: cmp $0x86dd,%rax
2c: jne 0x0000000000000069 4f: jne 0x000000000000009a
2e: mov $0x14,%esi 51: mov $0x14,%esi
33: callq 0xffffffffe1021e31 56: callq 0xffffffffe102bd91
38: cmp $0x84,%eax 5b: cmp $0x84,%rax
3d: je 0x0000000000000049 62: je 0x0000000000000074
3f: cmp $0x6,%eax 64: cmp $0x6,%rax
42: je 0x0000000000000049 68: je 0x0000000000000074
44: cmp $0x11,%eax 6a: cmp $0x11,%rax
47: jne 0x00000000000000c6 6e: jne 0x0000000000000117
49: mov $0x36,%esi 74: mov $0x36,%esi
4e: callq 0xffffffffe1021e15 79: callq 0xffffffffe102bd75
53: cmp $0x16,%eax 7e: cmp $0x16,%rax
56: je 0x00000000000000bf 82: je 0x0000000000000110
58: mov $0x38,%esi 88: mov $0x38,%esi
5d: callq 0xffffffffe1021e15 8d: callq 0xffffffffe102bd75
62: cmp $0x16,%eax 92: cmp $0x16,%rax
65: je 0x00000000000000bf 96: je 0x0000000000000110
67: jmp 0x00000000000000c6 98: jmp 0x0000000000000117
69: cmp $0x800,%eax 9a: cmp $0x800,%rax
6e: jne 0x00000000000000c6 a1: jne 0x0000000000000117
70: mov $0x17,%esi a3: mov $0x17,%esi
75: callq 0xffffffffe1021e31 a8: callq 0xffffffffe102bd91
7a: cmp $0x84,%eax ad: cmp $0x84,%rax
7f: je 0x000000000000008b b4: je 0x00000000000000c2
81: cmp $0x6,%eax b6: cmp $0x6,%rax
84: je 0x000000000000008b ba: je 0x00000000000000c2
86: cmp $0x11,%eax bc: cmp $0x11,%rax
89: jne 0x00000000000000c6 c0: jne 0x0000000000000117
8b: mov $0x14,%esi c2: mov $0x14,%esi
90: callq 0xffffffffe1021e15 c7: callq 0xffffffffe102bd75
95: test $0x1fff,%ax cc: test $0x1fff,%rax
99: jne 0x00000000000000c6 d3: jne 0x0000000000000117
d5: mov %rax,%r14
9b: mov $0xe,%esi d8: mov $0xe,%esi
a0: callq 0xffffffffe1021e44 dd: callq 0xffffffffe102bd91 // MSH
e2: and $0xf,%eax
e5: shl $0x2,%eax
e8: mov %rax,%r13
eb: mov %r14,%rax
ee: mov %r13,%rsi
a5: lea 0xe(%rbx),%esi f1: add $0xe,%esi
a8: callq 0xffffffffe1021e0d f4: callq 0xffffffffe102bd6d
ad: cmp $0x16,%eax f9: cmp $0x16,%rax
b0: je 0x00000000000000bf fd: je 0x0000000000000110
ff: mov %r13,%rsi
b2: lea 0x10(%rbx),%esi 102: add $0x10,%esi
b5: callq 0xffffffffe1021e0d 105: callq 0xffffffffe102bd6d
ba: cmp $0x16,%eax 10a: cmp $0x16,%rax
bd: jne 0x00000000000000c6 10e: jne 0x0000000000000117
bf: mov $0xffff,%eax 110: mov $0xffff,%eax
c4: jmp 0x00000000000000c8 115: jmp 0x000000000000011c
c6: xor %eax,%eax 117: mov $0x0,%eax
c8: mov -0x8(%rbp),%rbx 11c: mov -0x228(%rbp),%rbx // epilogue
cc: leaveq 123: mov -0x220(%rbp),%r13
cd: retq 12a: mov -0x218(%rbp),%r14
131: mov -0x210(%rbp),%r15
138: leaveq
139: retq
On fully cached SKBs both JITed functions take 12 nsec to execute.
BPF interpreter executes the program in 30 nsec.
The difference in generated assembler is due to the following:
Old BPF imlements LDX_MSH instruction via sk_load_byte_msh() helper function
inside bpf_jit.S.
New JIT removes the helper and does it explicitly, so ldx_msh cost
is the same for both JITs, but generated code looks longer.
New JIT has 4 registers to save, so prologue/epilogue are larger,
but the cost is within noise on x64.
Old JIT checks whether first insn clears A and if not emits 'xor %eax,%eax'.
New JIT clears %rax unconditionally.
2. old BPF JIT doesn't support ANC_NLATTR, ANC_PAY_OFFSET, ANC_RANDOM
extensions. New JIT supports all BPF extensions.
Performance of such filters improves 2-4 times depending on a filter.
The longer the filter the higher performance gain.
Synthetic benchmarks with many ancillary loads see 20x speedup
which seems to be the maximum gain from JIT
Notes:
. net.core.bpf_jit_enable=2 + tools/net/bpf_jit_disasm is still functional
and can be used to see generated assembler
. there are two jit_compile() functions and code flow for classic filters is:
sk_attach_filter() - load classic BPF
bpf_jit_compile() - try to JIT from classic BPF
sk_convert_filter() - convert classic to internal
bpf_int_jit_compile() - JIT from internal BPF
seccomp and tracing filters will just call bpf_int_jit_compile()
Signed-off-by: Alexei Starovoitov <ast@plumgrid.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
2014-05-14 02:50:46 +00:00
|
|
|
else
|
2014-06-06 21:46:06 +00:00
|
|
|
EMIT2_off32(0x81, add_1reg(0xF8, dst_reg), imm32);
|
net: filter: x86: internal BPF JIT
Maps all internal BPF instructions into x86_64 instructions.
This patch replaces original BPF x64 JIT with internal BPF x64 JIT.
sysctl net.core.bpf_jit_enable is reused as on/off switch.
Performance:
1. old BPF JIT and internal BPF JIT generate equivalent x86_64 code.
No performance difference is observed for filters that were JIT-able before
Example assembler code for BPF filter "tcpdump port 22"
original BPF -> old JIT: original BPF -> internal BPF -> new JIT:
0: push %rbp 0: push %rbp
1: mov %rsp,%rbp 1: mov %rsp,%rbp
4: sub $0x60,%rsp 4: sub $0x228,%rsp
8: mov %rbx,-0x8(%rbp) b: mov %rbx,-0x228(%rbp) // prologue
12: mov %r13,-0x220(%rbp)
19: mov %r14,-0x218(%rbp)
20: mov %r15,-0x210(%rbp)
27: xor %eax,%eax // clear A
c: xor %ebx,%ebx 29: xor %r13,%r13 // clear X
e: mov 0x68(%rdi),%r9d 2c: mov 0x68(%rdi),%r9d
12: sub 0x6c(%rdi),%r9d 30: sub 0x6c(%rdi),%r9d
16: mov 0xd8(%rdi),%r8 34: mov 0xd8(%rdi),%r10
3b: mov %rdi,%rbx
1d: mov $0xc,%esi 3e: mov $0xc,%esi
22: callq 0xffffffffe1021e15 43: callq 0xffffffffe102bd75
27: cmp $0x86dd,%eax 48: cmp $0x86dd,%rax
2c: jne 0x0000000000000069 4f: jne 0x000000000000009a
2e: mov $0x14,%esi 51: mov $0x14,%esi
33: callq 0xffffffffe1021e31 56: callq 0xffffffffe102bd91
38: cmp $0x84,%eax 5b: cmp $0x84,%rax
3d: je 0x0000000000000049 62: je 0x0000000000000074
3f: cmp $0x6,%eax 64: cmp $0x6,%rax
42: je 0x0000000000000049 68: je 0x0000000000000074
44: cmp $0x11,%eax 6a: cmp $0x11,%rax
47: jne 0x00000000000000c6 6e: jne 0x0000000000000117
49: mov $0x36,%esi 74: mov $0x36,%esi
4e: callq 0xffffffffe1021e15 79: callq 0xffffffffe102bd75
53: cmp $0x16,%eax 7e: cmp $0x16,%rax
56: je 0x00000000000000bf 82: je 0x0000000000000110
58: mov $0x38,%esi 88: mov $0x38,%esi
5d: callq 0xffffffffe1021e15 8d: callq 0xffffffffe102bd75
62: cmp $0x16,%eax 92: cmp $0x16,%rax
65: je 0x00000000000000bf 96: je 0x0000000000000110
67: jmp 0x00000000000000c6 98: jmp 0x0000000000000117
69: cmp $0x800,%eax 9a: cmp $0x800,%rax
6e: jne 0x00000000000000c6 a1: jne 0x0000000000000117
70: mov $0x17,%esi a3: mov $0x17,%esi
75: callq 0xffffffffe1021e31 a8: callq 0xffffffffe102bd91
7a: cmp $0x84,%eax ad: cmp $0x84,%rax
7f: je 0x000000000000008b b4: je 0x00000000000000c2
81: cmp $0x6,%eax b6: cmp $0x6,%rax
84: je 0x000000000000008b ba: je 0x00000000000000c2
86: cmp $0x11,%eax bc: cmp $0x11,%rax
89: jne 0x00000000000000c6 c0: jne 0x0000000000000117
8b: mov $0x14,%esi c2: mov $0x14,%esi
90: callq 0xffffffffe1021e15 c7: callq 0xffffffffe102bd75
95: test $0x1fff,%ax cc: test $0x1fff,%rax
99: jne 0x00000000000000c6 d3: jne 0x0000000000000117
d5: mov %rax,%r14
9b: mov $0xe,%esi d8: mov $0xe,%esi
a0: callq 0xffffffffe1021e44 dd: callq 0xffffffffe102bd91 // MSH
e2: and $0xf,%eax
e5: shl $0x2,%eax
e8: mov %rax,%r13
eb: mov %r14,%rax
ee: mov %r13,%rsi
a5: lea 0xe(%rbx),%esi f1: add $0xe,%esi
a8: callq 0xffffffffe1021e0d f4: callq 0xffffffffe102bd6d
ad: cmp $0x16,%eax f9: cmp $0x16,%rax
b0: je 0x00000000000000bf fd: je 0x0000000000000110
ff: mov %r13,%rsi
b2: lea 0x10(%rbx),%esi 102: add $0x10,%esi
b5: callq 0xffffffffe1021e0d 105: callq 0xffffffffe102bd6d
ba: cmp $0x16,%eax 10a: cmp $0x16,%rax
bd: jne 0x00000000000000c6 10e: jne 0x0000000000000117
bf: mov $0xffff,%eax 110: mov $0xffff,%eax
c4: jmp 0x00000000000000c8 115: jmp 0x000000000000011c
c6: xor %eax,%eax 117: mov $0x0,%eax
c8: mov -0x8(%rbp),%rbx 11c: mov -0x228(%rbp),%rbx // epilogue
cc: leaveq 123: mov -0x220(%rbp),%r13
cd: retq 12a: mov -0x218(%rbp),%r14
131: mov -0x210(%rbp),%r15
138: leaveq
139: retq
On fully cached SKBs both JITed functions take 12 nsec to execute.
BPF interpreter executes the program in 30 nsec.
The difference in generated assembler is due to the following:
Old BPF imlements LDX_MSH instruction via sk_load_byte_msh() helper function
inside bpf_jit.S.
New JIT removes the helper and does it explicitly, so ldx_msh cost
is the same for both JITs, but generated code looks longer.
New JIT has 4 registers to save, so prologue/epilogue are larger,
but the cost is within noise on x64.
Old JIT checks whether first insn clears A and if not emits 'xor %eax,%eax'.
New JIT clears %rax unconditionally.
2. old BPF JIT doesn't support ANC_NLATTR, ANC_PAY_OFFSET, ANC_RANDOM
extensions. New JIT supports all BPF extensions.
Performance of such filters improves 2-4 times depending on a filter.
The longer the filter the higher performance gain.
Synthetic benchmarks with many ancillary loads see 20x speedup
which seems to be the maximum gain from JIT
Notes:
. net.core.bpf_jit_enable=2 + tools/net/bpf_jit_disasm is still functional
and can be used to see generated assembler
. there are two jit_compile() functions and code flow for classic filters is:
sk_attach_filter() - load classic BPF
bpf_jit_compile() - try to JIT from classic BPF
sk_convert_filter() - convert classic to internal
bpf_int_jit_compile() - JIT from internal BPF
seccomp and tracing filters will just call bpf_int_jit_compile()
Signed-off-by: Alexei Starovoitov <ast@plumgrid.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
2014-05-14 02:50:46 +00:00
|
|
|
|
x86/bpf: Clean up non-standard comments, to make the code more readable
So by chance I looked into x86 assembly in arch/x86/net/bpf_jit_comp.c and
noticed the weird and inconsistent comment style it mistakenly learned from
the networking code:
/* Multi-line comment ...
* ... looks like this.
*/
Fix this to use the standard comment style specified in Documentation/CodingStyle
and used in arch/x86/ as well:
/*
* Multi-line comment ...
* ... looks like this.
*/
Also, to quote Linus's ... more explicit views about this:
http://article.gmane.org/gmane.linux.kernel.cryptoapi/21066
> But no, the networking code picked *none* of the above sane formats.
> Instead, it picked these two models that are just half-arsed
> shit-for-brains:
>
> (no)
> /* This is disgusting drug-induced
> * crap, and should die
> */
>
> (no-no-no)
> /* This is also very nasty
> * and visually unbalanced */
>
> Please. The networking code actually has the *worst* possible comment
> style. You can literally find that (no-no-no) style, which is just
> really horribly disgusting and worse than the otherwise fairly similar
> (d) in pretty much every way.
Also improve the comments and some other details while at it:
- Don't mix same-line and previous-line comment style on otherwise
identical code patterns within the same function,
- capitalize 'BPF' and x86 register names consistently,
- capitalize sentences consistently,
- instead of 'x64' use 'x86-64': x64 is a Microsoft specific term,
- use more consistent punctuation,
- use standard coding style in macros as well,
- fix typos and a few other minor details.
Consistent coding style is not optional, at least in arch/x86/.
No change in functionality.
( In case this commit causes conflicts with pending development code
I'll be glad to help resolve any conflicts! )
Acked-by: Thomas Gleixner <tglx@linutronix.de>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Borislav Petkov <bp@alien8.de>
Cc: H. Peter Anvin <hpa@zytor.com>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: David S. Miller <davem@davemloft.net>
Cc: Eric Dumazet <edumazet@google.com>
Cc: Daniel Borkmann <daniel@iogearbox.net>
Cc: Alexei Starovoitov <ast@fb.com>
Cc: Hideaki YOSHIFUJI <yoshfuji@linux-ipv6.org>
Cc: netdev@vger.kernel.org
Cc: linux-kernel@vger.kernel.org
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
2018-04-27 09:54:40 +00:00
|
|
|
emit_cond_jmp: /* Convert BPF opcode to x86 */
|
net: filter: x86: internal BPF JIT
Maps all internal BPF instructions into x86_64 instructions.
This patch replaces original BPF x64 JIT with internal BPF x64 JIT.
sysctl net.core.bpf_jit_enable is reused as on/off switch.
Performance:
1. old BPF JIT and internal BPF JIT generate equivalent x86_64 code.
No performance difference is observed for filters that were JIT-able before
Example assembler code for BPF filter "tcpdump port 22"
original BPF -> old JIT: original BPF -> internal BPF -> new JIT:
0: push %rbp 0: push %rbp
1: mov %rsp,%rbp 1: mov %rsp,%rbp
4: sub $0x60,%rsp 4: sub $0x228,%rsp
8: mov %rbx,-0x8(%rbp) b: mov %rbx,-0x228(%rbp) // prologue
12: mov %r13,-0x220(%rbp)
19: mov %r14,-0x218(%rbp)
20: mov %r15,-0x210(%rbp)
27: xor %eax,%eax // clear A
c: xor %ebx,%ebx 29: xor %r13,%r13 // clear X
e: mov 0x68(%rdi),%r9d 2c: mov 0x68(%rdi),%r9d
12: sub 0x6c(%rdi),%r9d 30: sub 0x6c(%rdi),%r9d
16: mov 0xd8(%rdi),%r8 34: mov 0xd8(%rdi),%r10
3b: mov %rdi,%rbx
1d: mov $0xc,%esi 3e: mov $0xc,%esi
22: callq 0xffffffffe1021e15 43: callq 0xffffffffe102bd75
27: cmp $0x86dd,%eax 48: cmp $0x86dd,%rax
2c: jne 0x0000000000000069 4f: jne 0x000000000000009a
2e: mov $0x14,%esi 51: mov $0x14,%esi
33: callq 0xffffffffe1021e31 56: callq 0xffffffffe102bd91
38: cmp $0x84,%eax 5b: cmp $0x84,%rax
3d: je 0x0000000000000049 62: je 0x0000000000000074
3f: cmp $0x6,%eax 64: cmp $0x6,%rax
42: je 0x0000000000000049 68: je 0x0000000000000074
44: cmp $0x11,%eax 6a: cmp $0x11,%rax
47: jne 0x00000000000000c6 6e: jne 0x0000000000000117
49: mov $0x36,%esi 74: mov $0x36,%esi
4e: callq 0xffffffffe1021e15 79: callq 0xffffffffe102bd75
53: cmp $0x16,%eax 7e: cmp $0x16,%rax
56: je 0x00000000000000bf 82: je 0x0000000000000110
58: mov $0x38,%esi 88: mov $0x38,%esi
5d: callq 0xffffffffe1021e15 8d: callq 0xffffffffe102bd75
62: cmp $0x16,%eax 92: cmp $0x16,%rax
65: je 0x00000000000000bf 96: je 0x0000000000000110
67: jmp 0x00000000000000c6 98: jmp 0x0000000000000117
69: cmp $0x800,%eax 9a: cmp $0x800,%rax
6e: jne 0x00000000000000c6 a1: jne 0x0000000000000117
70: mov $0x17,%esi a3: mov $0x17,%esi
75: callq 0xffffffffe1021e31 a8: callq 0xffffffffe102bd91
7a: cmp $0x84,%eax ad: cmp $0x84,%rax
7f: je 0x000000000000008b b4: je 0x00000000000000c2
81: cmp $0x6,%eax b6: cmp $0x6,%rax
84: je 0x000000000000008b ba: je 0x00000000000000c2
86: cmp $0x11,%eax bc: cmp $0x11,%rax
89: jne 0x00000000000000c6 c0: jne 0x0000000000000117
8b: mov $0x14,%esi c2: mov $0x14,%esi
90: callq 0xffffffffe1021e15 c7: callq 0xffffffffe102bd75
95: test $0x1fff,%ax cc: test $0x1fff,%rax
99: jne 0x00000000000000c6 d3: jne 0x0000000000000117
d5: mov %rax,%r14
9b: mov $0xe,%esi d8: mov $0xe,%esi
a0: callq 0xffffffffe1021e44 dd: callq 0xffffffffe102bd91 // MSH
e2: and $0xf,%eax
e5: shl $0x2,%eax
e8: mov %rax,%r13
eb: mov %r14,%rax
ee: mov %r13,%rsi
a5: lea 0xe(%rbx),%esi f1: add $0xe,%esi
a8: callq 0xffffffffe1021e0d f4: callq 0xffffffffe102bd6d
ad: cmp $0x16,%eax f9: cmp $0x16,%rax
b0: je 0x00000000000000bf fd: je 0x0000000000000110
ff: mov %r13,%rsi
b2: lea 0x10(%rbx),%esi 102: add $0x10,%esi
b5: callq 0xffffffffe1021e0d 105: callq 0xffffffffe102bd6d
ba: cmp $0x16,%eax 10a: cmp $0x16,%rax
bd: jne 0x00000000000000c6 10e: jne 0x0000000000000117
bf: mov $0xffff,%eax 110: mov $0xffff,%eax
c4: jmp 0x00000000000000c8 115: jmp 0x000000000000011c
c6: xor %eax,%eax 117: mov $0x0,%eax
c8: mov -0x8(%rbp),%rbx 11c: mov -0x228(%rbp),%rbx // epilogue
cc: leaveq 123: mov -0x220(%rbp),%r13
cd: retq 12a: mov -0x218(%rbp),%r14
131: mov -0x210(%rbp),%r15
138: leaveq
139: retq
On fully cached SKBs both JITed functions take 12 nsec to execute.
BPF interpreter executes the program in 30 nsec.
The difference in generated assembler is due to the following:
Old BPF imlements LDX_MSH instruction via sk_load_byte_msh() helper function
inside bpf_jit.S.
New JIT removes the helper and does it explicitly, so ldx_msh cost
is the same for both JITs, but generated code looks longer.
New JIT has 4 registers to save, so prologue/epilogue are larger,
but the cost is within noise on x64.
Old JIT checks whether first insn clears A and if not emits 'xor %eax,%eax'.
New JIT clears %rax unconditionally.
2. old BPF JIT doesn't support ANC_NLATTR, ANC_PAY_OFFSET, ANC_RANDOM
extensions. New JIT supports all BPF extensions.
Performance of such filters improves 2-4 times depending on a filter.
The longer the filter the higher performance gain.
Synthetic benchmarks with many ancillary loads see 20x speedup
which seems to be the maximum gain from JIT
Notes:
. net.core.bpf_jit_enable=2 + tools/net/bpf_jit_disasm is still functional
and can be used to see generated assembler
. there are two jit_compile() functions and code flow for classic filters is:
sk_attach_filter() - load classic BPF
bpf_jit_compile() - try to JIT from classic BPF
sk_convert_filter() - convert classic to internal
bpf_int_jit_compile() - JIT from internal BPF
seccomp and tracing filters will just call bpf_int_jit_compile()
Signed-off-by: Alexei Starovoitov <ast@plumgrid.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
2014-05-14 02:50:46 +00:00
|
|
|
switch (BPF_OP(insn->code)) {
|
|
|
|
case BPF_JEQ:
|
|
|
|
jmp_cond = X86_JE;
|
|
|
|
break;
|
|
|
|
case BPF_JSET:
|
|
|
|
case BPF_JNE:
|
|
|
|
jmp_cond = X86_JNE;
|
|
|
|
break;
|
|
|
|
case BPF_JGT:
|
|
|
|
/* GT is unsigned '>', JA in x86 */
|
|
|
|
jmp_cond = X86_JA;
|
|
|
|
break;
|
2017-08-09 23:39:56 +00:00
|
|
|
case BPF_JLT:
|
|
|
|
/* LT is unsigned '<', JB in x86 */
|
|
|
|
jmp_cond = X86_JB;
|
|
|
|
break;
|
net: filter: x86: internal BPF JIT
Maps all internal BPF instructions into x86_64 instructions.
This patch replaces original BPF x64 JIT with internal BPF x64 JIT.
sysctl net.core.bpf_jit_enable is reused as on/off switch.
Performance:
1. old BPF JIT and internal BPF JIT generate equivalent x86_64 code.
No performance difference is observed for filters that were JIT-able before
Example assembler code for BPF filter "tcpdump port 22"
original BPF -> old JIT: original BPF -> internal BPF -> new JIT:
0: push %rbp 0: push %rbp
1: mov %rsp,%rbp 1: mov %rsp,%rbp
4: sub $0x60,%rsp 4: sub $0x228,%rsp
8: mov %rbx,-0x8(%rbp) b: mov %rbx,-0x228(%rbp) // prologue
12: mov %r13,-0x220(%rbp)
19: mov %r14,-0x218(%rbp)
20: mov %r15,-0x210(%rbp)
27: xor %eax,%eax // clear A
c: xor %ebx,%ebx 29: xor %r13,%r13 // clear X
e: mov 0x68(%rdi),%r9d 2c: mov 0x68(%rdi),%r9d
12: sub 0x6c(%rdi),%r9d 30: sub 0x6c(%rdi),%r9d
16: mov 0xd8(%rdi),%r8 34: mov 0xd8(%rdi),%r10
3b: mov %rdi,%rbx
1d: mov $0xc,%esi 3e: mov $0xc,%esi
22: callq 0xffffffffe1021e15 43: callq 0xffffffffe102bd75
27: cmp $0x86dd,%eax 48: cmp $0x86dd,%rax
2c: jne 0x0000000000000069 4f: jne 0x000000000000009a
2e: mov $0x14,%esi 51: mov $0x14,%esi
33: callq 0xffffffffe1021e31 56: callq 0xffffffffe102bd91
38: cmp $0x84,%eax 5b: cmp $0x84,%rax
3d: je 0x0000000000000049 62: je 0x0000000000000074
3f: cmp $0x6,%eax 64: cmp $0x6,%rax
42: je 0x0000000000000049 68: je 0x0000000000000074
44: cmp $0x11,%eax 6a: cmp $0x11,%rax
47: jne 0x00000000000000c6 6e: jne 0x0000000000000117
49: mov $0x36,%esi 74: mov $0x36,%esi
4e: callq 0xffffffffe1021e15 79: callq 0xffffffffe102bd75
53: cmp $0x16,%eax 7e: cmp $0x16,%rax
56: je 0x00000000000000bf 82: je 0x0000000000000110
58: mov $0x38,%esi 88: mov $0x38,%esi
5d: callq 0xffffffffe1021e15 8d: callq 0xffffffffe102bd75
62: cmp $0x16,%eax 92: cmp $0x16,%rax
65: je 0x00000000000000bf 96: je 0x0000000000000110
67: jmp 0x00000000000000c6 98: jmp 0x0000000000000117
69: cmp $0x800,%eax 9a: cmp $0x800,%rax
6e: jne 0x00000000000000c6 a1: jne 0x0000000000000117
70: mov $0x17,%esi a3: mov $0x17,%esi
75: callq 0xffffffffe1021e31 a8: callq 0xffffffffe102bd91
7a: cmp $0x84,%eax ad: cmp $0x84,%rax
7f: je 0x000000000000008b b4: je 0x00000000000000c2
81: cmp $0x6,%eax b6: cmp $0x6,%rax
84: je 0x000000000000008b ba: je 0x00000000000000c2
86: cmp $0x11,%eax bc: cmp $0x11,%rax
89: jne 0x00000000000000c6 c0: jne 0x0000000000000117
8b: mov $0x14,%esi c2: mov $0x14,%esi
90: callq 0xffffffffe1021e15 c7: callq 0xffffffffe102bd75
95: test $0x1fff,%ax cc: test $0x1fff,%rax
99: jne 0x00000000000000c6 d3: jne 0x0000000000000117
d5: mov %rax,%r14
9b: mov $0xe,%esi d8: mov $0xe,%esi
a0: callq 0xffffffffe1021e44 dd: callq 0xffffffffe102bd91 // MSH
e2: and $0xf,%eax
e5: shl $0x2,%eax
e8: mov %rax,%r13
eb: mov %r14,%rax
ee: mov %r13,%rsi
a5: lea 0xe(%rbx),%esi f1: add $0xe,%esi
a8: callq 0xffffffffe1021e0d f4: callq 0xffffffffe102bd6d
ad: cmp $0x16,%eax f9: cmp $0x16,%rax
b0: je 0x00000000000000bf fd: je 0x0000000000000110
ff: mov %r13,%rsi
b2: lea 0x10(%rbx),%esi 102: add $0x10,%esi
b5: callq 0xffffffffe1021e0d 105: callq 0xffffffffe102bd6d
ba: cmp $0x16,%eax 10a: cmp $0x16,%rax
bd: jne 0x00000000000000c6 10e: jne 0x0000000000000117
bf: mov $0xffff,%eax 110: mov $0xffff,%eax
c4: jmp 0x00000000000000c8 115: jmp 0x000000000000011c
c6: xor %eax,%eax 117: mov $0x0,%eax
c8: mov -0x8(%rbp),%rbx 11c: mov -0x228(%rbp),%rbx // epilogue
cc: leaveq 123: mov -0x220(%rbp),%r13
cd: retq 12a: mov -0x218(%rbp),%r14
131: mov -0x210(%rbp),%r15
138: leaveq
139: retq
On fully cached SKBs both JITed functions take 12 nsec to execute.
BPF interpreter executes the program in 30 nsec.
The difference in generated assembler is due to the following:
Old BPF imlements LDX_MSH instruction via sk_load_byte_msh() helper function
inside bpf_jit.S.
New JIT removes the helper and does it explicitly, so ldx_msh cost
is the same for both JITs, but generated code looks longer.
New JIT has 4 registers to save, so prologue/epilogue are larger,
but the cost is within noise on x64.
Old JIT checks whether first insn clears A and if not emits 'xor %eax,%eax'.
New JIT clears %rax unconditionally.
2. old BPF JIT doesn't support ANC_NLATTR, ANC_PAY_OFFSET, ANC_RANDOM
extensions. New JIT supports all BPF extensions.
Performance of such filters improves 2-4 times depending on a filter.
The longer the filter the higher performance gain.
Synthetic benchmarks with many ancillary loads see 20x speedup
which seems to be the maximum gain from JIT
Notes:
. net.core.bpf_jit_enable=2 + tools/net/bpf_jit_disasm is still functional
and can be used to see generated assembler
. there are two jit_compile() functions and code flow for classic filters is:
sk_attach_filter() - load classic BPF
bpf_jit_compile() - try to JIT from classic BPF
sk_convert_filter() - convert classic to internal
bpf_int_jit_compile() - JIT from internal BPF
seccomp and tracing filters will just call bpf_int_jit_compile()
Signed-off-by: Alexei Starovoitov <ast@plumgrid.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
2014-05-14 02:50:46 +00:00
|
|
|
case BPF_JGE:
|
|
|
|
/* GE is unsigned '>=', JAE in x86 */
|
|
|
|
jmp_cond = X86_JAE;
|
|
|
|
break;
|
2017-08-09 23:39:56 +00:00
|
|
|
case BPF_JLE:
|
|
|
|
/* LE is unsigned '<=', JBE in x86 */
|
|
|
|
jmp_cond = X86_JBE;
|
|
|
|
break;
|
net: filter: x86: internal BPF JIT
Maps all internal BPF instructions into x86_64 instructions.
This patch replaces original BPF x64 JIT with internal BPF x64 JIT.
sysctl net.core.bpf_jit_enable is reused as on/off switch.
Performance:
1. old BPF JIT and internal BPF JIT generate equivalent x86_64 code.
No performance difference is observed for filters that were JIT-able before
Example assembler code for BPF filter "tcpdump port 22"
original BPF -> old JIT: original BPF -> internal BPF -> new JIT:
0: push %rbp 0: push %rbp
1: mov %rsp,%rbp 1: mov %rsp,%rbp
4: sub $0x60,%rsp 4: sub $0x228,%rsp
8: mov %rbx,-0x8(%rbp) b: mov %rbx,-0x228(%rbp) // prologue
12: mov %r13,-0x220(%rbp)
19: mov %r14,-0x218(%rbp)
20: mov %r15,-0x210(%rbp)
27: xor %eax,%eax // clear A
c: xor %ebx,%ebx 29: xor %r13,%r13 // clear X
e: mov 0x68(%rdi),%r9d 2c: mov 0x68(%rdi),%r9d
12: sub 0x6c(%rdi),%r9d 30: sub 0x6c(%rdi),%r9d
16: mov 0xd8(%rdi),%r8 34: mov 0xd8(%rdi),%r10
3b: mov %rdi,%rbx
1d: mov $0xc,%esi 3e: mov $0xc,%esi
22: callq 0xffffffffe1021e15 43: callq 0xffffffffe102bd75
27: cmp $0x86dd,%eax 48: cmp $0x86dd,%rax
2c: jne 0x0000000000000069 4f: jne 0x000000000000009a
2e: mov $0x14,%esi 51: mov $0x14,%esi
33: callq 0xffffffffe1021e31 56: callq 0xffffffffe102bd91
38: cmp $0x84,%eax 5b: cmp $0x84,%rax
3d: je 0x0000000000000049 62: je 0x0000000000000074
3f: cmp $0x6,%eax 64: cmp $0x6,%rax
42: je 0x0000000000000049 68: je 0x0000000000000074
44: cmp $0x11,%eax 6a: cmp $0x11,%rax
47: jne 0x00000000000000c6 6e: jne 0x0000000000000117
49: mov $0x36,%esi 74: mov $0x36,%esi
4e: callq 0xffffffffe1021e15 79: callq 0xffffffffe102bd75
53: cmp $0x16,%eax 7e: cmp $0x16,%rax
56: je 0x00000000000000bf 82: je 0x0000000000000110
58: mov $0x38,%esi 88: mov $0x38,%esi
5d: callq 0xffffffffe1021e15 8d: callq 0xffffffffe102bd75
62: cmp $0x16,%eax 92: cmp $0x16,%rax
65: je 0x00000000000000bf 96: je 0x0000000000000110
67: jmp 0x00000000000000c6 98: jmp 0x0000000000000117
69: cmp $0x800,%eax 9a: cmp $0x800,%rax
6e: jne 0x00000000000000c6 a1: jne 0x0000000000000117
70: mov $0x17,%esi a3: mov $0x17,%esi
75: callq 0xffffffffe1021e31 a8: callq 0xffffffffe102bd91
7a: cmp $0x84,%eax ad: cmp $0x84,%rax
7f: je 0x000000000000008b b4: je 0x00000000000000c2
81: cmp $0x6,%eax b6: cmp $0x6,%rax
84: je 0x000000000000008b ba: je 0x00000000000000c2
86: cmp $0x11,%eax bc: cmp $0x11,%rax
89: jne 0x00000000000000c6 c0: jne 0x0000000000000117
8b: mov $0x14,%esi c2: mov $0x14,%esi
90: callq 0xffffffffe1021e15 c7: callq 0xffffffffe102bd75
95: test $0x1fff,%ax cc: test $0x1fff,%rax
99: jne 0x00000000000000c6 d3: jne 0x0000000000000117
d5: mov %rax,%r14
9b: mov $0xe,%esi d8: mov $0xe,%esi
a0: callq 0xffffffffe1021e44 dd: callq 0xffffffffe102bd91 // MSH
e2: and $0xf,%eax
e5: shl $0x2,%eax
e8: mov %rax,%r13
eb: mov %r14,%rax
ee: mov %r13,%rsi
a5: lea 0xe(%rbx),%esi f1: add $0xe,%esi
a8: callq 0xffffffffe1021e0d f4: callq 0xffffffffe102bd6d
ad: cmp $0x16,%eax f9: cmp $0x16,%rax
b0: je 0x00000000000000bf fd: je 0x0000000000000110
ff: mov %r13,%rsi
b2: lea 0x10(%rbx),%esi 102: add $0x10,%esi
b5: callq 0xffffffffe1021e0d 105: callq 0xffffffffe102bd6d
ba: cmp $0x16,%eax 10a: cmp $0x16,%rax
bd: jne 0x00000000000000c6 10e: jne 0x0000000000000117
bf: mov $0xffff,%eax 110: mov $0xffff,%eax
c4: jmp 0x00000000000000c8 115: jmp 0x000000000000011c
c6: xor %eax,%eax 117: mov $0x0,%eax
c8: mov -0x8(%rbp),%rbx 11c: mov -0x228(%rbp),%rbx // epilogue
cc: leaveq 123: mov -0x220(%rbp),%r13
cd: retq 12a: mov -0x218(%rbp),%r14
131: mov -0x210(%rbp),%r15
138: leaveq
139: retq
On fully cached SKBs both JITed functions take 12 nsec to execute.
BPF interpreter executes the program in 30 nsec.
The difference in generated assembler is due to the following:
Old BPF imlements LDX_MSH instruction via sk_load_byte_msh() helper function
inside bpf_jit.S.
New JIT removes the helper and does it explicitly, so ldx_msh cost
is the same for both JITs, but generated code looks longer.
New JIT has 4 registers to save, so prologue/epilogue are larger,
but the cost is within noise on x64.
Old JIT checks whether first insn clears A and if not emits 'xor %eax,%eax'.
New JIT clears %rax unconditionally.
2. old BPF JIT doesn't support ANC_NLATTR, ANC_PAY_OFFSET, ANC_RANDOM
extensions. New JIT supports all BPF extensions.
Performance of such filters improves 2-4 times depending on a filter.
The longer the filter the higher performance gain.
Synthetic benchmarks with many ancillary loads see 20x speedup
which seems to be the maximum gain from JIT
Notes:
. net.core.bpf_jit_enable=2 + tools/net/bpf_jit_disasm is still functional
and can be used to see generated assembler
. there are two jit_compile() functions and code flow for classic filters is:
sk_attach_filter() - load classic BPF
bpf_jit_compile() - try to JIT from classic BPF
sk_convert_filter() - convert classic to internal
bpf_int_jit_compile() - JIT from internal BPF
seccomp and tracing filters will just call bpf_int_jit_compile()
Signed-off-by: Alexei Starovoitov <ast@plumgrid.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
2014-05-14 02:50:46 +00:00
|
|
|
case BPF_JSGT:
|
x86/bpf: Clean up non-standard comments, to make the code more readable
So by chance I looked into x86 assembly in arch/x86/net/bpf_jit_comp.c and
noticed the weird and inconsistent comment style it mistakenly learned from
the networking code:
/* Multi-line comment ...
* ... looks like this.
*/
Fix this to use the standard comment style specified in Documentation/CodingStyle
and used in arch/x86/ as well:
/*
* Multi-line comment ...
* ... looks like this.
*/
Also, to quote Linus's ... more explicit views about this:
http://article.gmane.org/gmane.linux.kernel.cryptoapi/21066
> But no, the networking code picked *none* of the above sane formats.
> Instead, it picked these two models that are just half-arsed
> shit-for-brains:
>
> (no)
> /* This is disgusting drug-induced
> * crap, and should die
> */
>
> (no-no-no)
> /* This is also very nasty
> * and visually unbalanced */
>
> Please. The networking code actually has the *worst* possible comment
> style. You can literally find that (no-no-no) style, which is just
> really horribly disgusting and worse than the otherwise fairly similar
> (d) in pretty much every way.
Also improve the comments and some other details while at it:
- Don't mix same-line and previous-line comment style on otherwise
identical code patterns within the same function,
- capitalize 'BPF' and x86 register names consistently,
- capitalize sentences consistently,
- instead of 'x64' use 'x86-64': x64 is a Microsoft specific term,
- use more consistent punctuation,
- use standard coding style in macros as well,
- fix typos and a few other minor details.
Consistent coding style is not optional, at least in arch/x86/.
No change in functionality.
( In case this commit causes conflicts with pending development code
I'll be glad to help resolve any conflicts! )
Acked-by: Thomas Gleixner <tglx@linutronix.de>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Borislav Petkov <bp@alien8.de>
Cc: H. Peter Anvin <hpa@zytor.com>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: David S. Miller <davem@davemloft.net>
Cc: Eric Dumazet <edumazet@google.com>
Cc: Daniel Borkmann <daniel@iogearbox.net>
Cc: Alexei Starovoitov <ast@fb.com>
Cc: Hideaki YOSHIFUJI <yoshfuji@linux-ipv6.org>
Cc: netdev@vger.kernel.org
Cc: linux-kernel@vger.kernel.org
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
2018-04-27 09:54:40 +00:00
|
|
|
/* Signed '>', GT in x86 */
|
net: filter: x86: internal BPF JIT
Maps all internal BPF instructions into x86_64 instructions.
This patch replaces original BPF x64 JIT with internal BPF x64 JIT.
sysctl net.core.bpf_jit_enable is reused as on/off switch.
Performance:
1. old BPF JIT and internal BPF JIT generate equivalent x86_64 code.
No performance difference is observed for filters that were JIT-able before
Example assembler code for BPF filter "tcpdump port 22"
original BPF -> old JIT: original BPF -> internal BPF -> new JIT:
0: push %rbp 0: push %rbp
1: mov %rsp,%rbp 1: mov %rsp,%rbp
4: sub $0x60,%rsp 4: sub $0x228,%rsp
8: mov %rbx,-0x8(%rbp) b: mov %rbx,-0x228(%rbp) // prologue
12: mov %r13,-0x220(%rbp)
19: mov %r14,-0x218(%rbp)
20: mov %r15,-0x210(%rbp)
27: xor %eax,%eax // clear A
c: xor %ebx,%ebx 29: xor %r13,%r13 // clear X
e: mov 0x68(%rdi),%r9d 2c: mov 0x68(%rdi),%r9d
12: sub 0x6c(%rdi),%r9d 30: sub 0x6c(%rdi),%r9d
16: mov 0xd8(%rdi),%r8 34: mov 0xd8(%rdi),%r10
3b: mov %rdi,%rbx
1d: mov $0xc,%esi 3e: mov $0xc,%esi
22: callq 0xffffffffe1021e15 43: callq 0xffffffffe102bd75
27: cmp $0x86dd,%eax 48: cmp $0x86dd,%rax
2c: jne 0x0000000000000069 4f: jne 0x000000000000009a
2e: mov $0x14,%esi 51: mov $0x14,%esi
33: callq 0xffffffffe1021e31 56: callq 0xffffffffe102bd91
38: cmp $0x84,%eax 5b: cmp $0x84,%rax
3d: je 0x0000000000000049 62: je 0x0000000000000074
3f: cmp $0x6,%eax 64: cmp $0x6,%rax
42: je 0x0000000000000049 68: je 0x0000000000000074
44: cmp $0x11,%eax 6a: cmp $0x11,%rax
47: jne 0x00000000000000c6 6e: jne 0x0000000000000117
49: mov $0x36,%esi 74: mov $0x36,%esi
4e: callq 0xffffffffe1021e15 79: callq 0xffffffffe102bd75
53: cmp $0x16,%eax 7e: cmp $0x16,%rax
56: je 0x00000000000000bf 82: je 0x0000000000000110
58: mov $0x38,%esi 88: mov $0x38,%esi
5d: callq 0xffffffffe1021e15 8d: callq 0xffffffffe102bd75
62: cmp $0x16,%eax 92: cmp $0x16,%rax
65: je 0x00000000000000bf 96: je 0x0000000000000110
67: jmp 0x00000000000000c6 98: jmp 0x0000000000000117
69: cmp $0x800,%eax 9a: cmp $0x800,%rax
6e: jne 0x00000000000000c6 a1: jne 0x0000000000000117
70: mov $0x17,%esi a3: mov $0x17,%esi
75: callq 0xffffffffe1021e31 a8: callq 0xffffffffe102bd91
7a: cmp $0x84,%eax ad: cmp $0x84,%rax
7f: je 0x000000000000008b b4: je 0x00000000000000c2
81: cmp $0x6,%eax b6: cmp $0x6,%rax
84: je 0x000000000000008b ba: je 0x00000000000000c2
86: cmp $0x11,%eax bc: cmp $0x11,%rax
89: jne 0x00000000000000c6 c0: jne 0x0000000000000117
8b: mov $0x14,%esi c2: mov $0x14,%esi
90: callq 0xffffffffe1021e15 c7: callq 0xffffffffe102bd75
95: test $0x1fff,%ax cc: test $0x1fff,%rax
99: jne 0x00000000000000c6 d3: jne 0x0000000000000117
d5: mov %rax,%r14
9b: mov $0xe,%esi d8: mov $0xe,%esi
a0: callq 0xffffffffe1021e44 dd: callq 0xffffffffe102bd91 // MSH
e2: and $0xf,%eax
e5: shl $0x2,%eax
e8: mov %rax,%r13
eb: mov %r14,%rax
ee: mov %r13,%rsi
a5: lea 0xe(%rbx),%esi f1: add $0xe,%esi
a8: callq 0xffffffffe1021e0d f4: callq 0xffffffffe102bd6d
ad: cmp $0x16,%eax f9: cmp $0x16,%rax
b0: je 0x00000000000000bf fd: je 0x0000000000000110
ff: mov %r13,%rsi
b2: lea 0x10(%rbx),%esi 102: add $0x10,%esi
b5: callq 0xffffffffe1021e0d 105: callq 0xffffffffe102bd6d
ba: cmp $0x16,%eax 10a: cmp $0x16,%rax
bd: jne 0x00000000000000c6 10e: jne 0x0000000000000117
bf: mov $0xffff,%eax 110: mov $0xffff,%eax
c4: jmp 0x00000000000000c8 115: jmp 0x000000000000011c
c6: xor %eax,%eax 117: mov $0x0,%eax
c8: mov -0x8(%rbp),%rbx 11c: mov -0x228(%rbp),%rbx // epilogue
cc: leaveq 123: mov -0x220(%rbp),%r13
cd: retq 12a: mov -0x218(%rbp),%r14
131: mov -0x210(%rbp),%r15
138: leaveq
139: retq
On fully cached SKBs both JITed functions take 12 nsec to execute.
BPF interpreter executes the program in 30 nsec.
The difference in generated assembler is due to the following:
Old BPF imlements LDX_MSH instruction via sk_load_byte_msh() helper function
inside bpf_jit.S.
New JIT removes the helper and does it explicitly, so ldx_msh cost
is the same for both JITs, but generated code looks longer.
New JIT has 4 registers to save, so prologue/epilogue are larger,
but the cost is within noise on x64.
Old JIT checks whether first insn clears A and if not emits 'xor %eax,%eax'.
New JIT clears %rax unconditionally.
2. old BPF JIT doesn't support ANC_NLATTR, ANC_PAY_OFFSET, ANC_RANDOM
extensions. New JIT supports all BPF extensions.
Performance of such filters improves 2-4 times depending on a filter.
The longer the filter the higher performance gain.
Synthetic benchmarks with many ancillary loads see 20x speedup
which seems to be the maximum gain from JIT
Notes:
. net.core.bpf_jit_enable=2 + tools/net/bpf_jit_disasm is still functional
and can be used to see generated assembler
. there are two jit_compile() functions and code flow for classic filters is:
sk_attach_filter() - load classic BPF
bpf_jit_compile() - try to JIT from classic BPF
sk_convert_filter() - convert classic to internal
bpf_int_jit_compile() - JIT from internal BPF
seccomp and tracing filters will just call bpf_int_jit_compile()
Signed-off-by: Alexei Starovoitov <ast@plumgrid.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
2014-05-14 02:50:46 +00:00
|
|
|
jmp_cond = X86_JG;
|
|
|
|
break;
|
2017-08-09 23:39:56 +00:00
|
|
|
case BPF_JSLT:
|
x86/bpf: Clean up non-standard comments, to make the code more readable
So by chance I looked into x86 assembly in arch/x86/net/bpf_jit_comp.c and
noticed the weird and inconsistent comment style it mistakenly learned from
the networking code:
/* Multi-line comment ...
* ... looks like this.
*/
Fix this to use the standard comment style specified in Documentation/CodingStyle
and used in arch/x86/ as well:
/*
* Multi-line comment ...
* ... looks like this.
*/
Also, to quote Linus's ... more explicit views about this:
http://article.gmane.org/gmane.linux.kernel.cryptoapi/21066
> But no, the networking code picked *none* of the above sane formats.
> Instead, it picked these two models that are just half-arsed
> shit-for-brains:
>
> (no)
> /* This is disgusting drug-induced
> * crap, and should die
> */
>
> (no-no-no)
> /* This is also very nasty
> * and visually unbalanced */
>
> Please. The networking code actually has the *worst* possible comment
> style. You can literally find that (no-no-no) style, which is just
> really horribly disgusting and worse than the otherwise fairly similar
> (d) in pretty much every way.
Also improve the comments and some other details while at it:
- Don't mix same-line and previous-line comment style on otherwise
identical code patterns within the same function,
- capitalize 'BPF' and x86 register names consistently,
- capitalize sentences consistently,
- instead of 'x64' use 'x86-64': x64 is a Microsoft specific term,
- use more consistent punctuation,
- use standard coding style in macros as well,
- fix typos and a few other minor details.
Consistent coding style is not optional, at least in arch/x86/.
No change in functionality.
( In case this commit causes conflicts with pending development code
I'll be glad to help resolve any conflicts! )
Acked-by: Thomas Gleixner <tglx@linutronix.de>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Borislav Petkov <bp@alien8.de>
Cc: H. Peter Anvin <hpa@zytor.com>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: David S. Miller <davem@davemloft.net>
Cc: Eric Dumazet <edumazet@google.com>
Cc: Daniel Borkmann <daniel@iogearbox.net>
Cc: Alexei Starovoitov <ast@fb.com>
Cc: Hideaki YOSHIFUJI <yoshfuji@linux-ipv6.org>
Cc: netdev@vger.kernel.org
Cc: linux-kernel@vger.kernel.org
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
2018-04-27 09:54:40 +00:00
|
|
|
/* Signed '<', LT in x86 */
|
2017-08-09 23:39:56 +00:00
|
|
|
jmp_cond = X86_JL;
|
|
|
|
break;
|
net: filter: x86: internal BPF JIT
Maps all internal BPF instructions into x86_64 instructions.
This patch replaces original BPF x64 JIT with internal BPF x64 JIT.
sysctl net.core.bpf_jit_enable is reused as on/off switch.
Performance:
1. old BPF JIT and internal BPF JIT generate equivalent x86_64 code.
No performance difference is observed for filters that were JIT-able before
Example assembler code for BPF filter "tcpdump port 22"
original BPF -> old JIT: original BPF -> internal BPF -> new JIT:
0: push %rbp 0: push %rbp
1: mov %rsp,%rbp 1: mov %rsp,%rbp
4: sub $0x60,%rsp 4: sub $0x228,%rsp
8: mov %rbx,-0x8(%rbp) b: mov %rbx,-0x228(%rbp) // prologue
12: mov %r13,-0x220(%rbp)
19: mov %r14,-0x218(%rbp)
20: mov %r15,-0x210(%rbp)
27: xor %eax,%eax // clear A
c: xor %ebx,%ebx 29: xor %r13,%r13 // clear X
e: mov 0x68(%rdi),%r9d 2c: mov 0x68(%rdi),%r9d
12: sub 0x6c(%rdi),%r9d 30: sub 0x6c(%rdi),%r9d
16: mov 0xd8(%rdi),%r8 34: mov 0xd8(%rdi),%r10
3b: mov %rdi,%rbx
1d: mov $0xc,%esi 3e: mov $0xc,%esi
22: callq 0xffffffffe1021e15 43: callq 0xffffffffe102bd75
27: cmp $0x86dd,%eax 48: cmp $0x86dd,%rax
2c: jne 0x0000000000000069 4f: jne 0x000000000000009a
2e: mov $0x14,%esi 51: mov $0x14,%esi
33: callq 0xffffffffe1021e31 56: callq 0xffffffffe102bd91
38: cmp $0x84,%eax 5b: cmp $0x84,%rax
3d: je 0x0000000000000049 62: je 0x0000000000000074
3f: cmp $0x6,%eax 64: cmp $0x6,%rax
42: je 0x0000000000000049 68: je 0x0000000000000074
44: cmp $0x11,%eax 6a: cmp $0x11,%rax
47: jne 0x00000000000000c6 6e: jne 0x0000000000000117
49: mov $0x36,%esi 74: mov $0x36,%esi
4e: callq 0xffffffffe1021e15 79: callq 0xffffffffe102bd75
53: cmp $0x16,%eax 7e: cmp $0x16,%rax
56: je 0x00000000000000bf 82: je 0x0000000000000110
58: mov $0x38,%esi 88: mov $0x38,%esi
5d: callq 0xffffffffe1021e15 8d: callq 0xffffffffe102bd75
62: cmp $0x16,%eax 92: cmp $0x16,%rax
65: je 0x00000000000000bf 96: je 0x0000000000000110
67: jmp 0x00000000000000c6 98: jmp 0x0000000000000117
69: cmp $0x800,%eax 9a: cmp $0x800,%rax
6e: jne 0x00000000000000c6 a1: jne 0x0000000000000117
70: mov $0x17,%esi a3: mov $0x17,%esi
75: callq 0xffffffffe1021e31 a8: callq 0xffffffffe102bd91
7a: cmp $0x84,%eax ad: cmp $0x84,%rax
7f: je 0x000000000000008b b4: je 0x00000000000000c2
81: cmp $0x6,%eax b6: cmp $0x6,%rax
84: je 0x000000000000008b ba: je 0x00000000000000c2
86: cmp $0x11,%eax bc: cmp $0x11,%rax
89: jne 0x00000000000000c6 c0: jne 0x0000000000000117
8b: mov $0x14,%esi c2: mov $0x14,%esi
90: callq 0xffffffffe1021e15 c7: callq 0xffffffffe102bd75
95: test $0x1fff,%ax cc: test $0x1fff,%rax
99: jne 0x00000000000000c6 d3: jne 0x0000000000000117
d5: mov %rax,%r14
9b: mov $0xe,%esi d8: mov $0xe,%esi
a0: callq 0xffffffffe1021e44 dd: callq 0xffffffffe102bd91 // MSH
e2: and $0xf,%eax
e5: shl $0x2,%eax
e8: mov %rax,%r13
eb: mov %r14,%rax
ee: mov %r13,%rsi
a5: lea 0xe(%rbx),%esi f1: add $0xe,%esi
a8: callq 0xffffffffe1021e0d f4: callq 0xffffffffe102bd6d
ad: cmp $0x16,%eax f9: cmp $0x16,%rax
b0: je 0x00000000000000bf fd: je 0x0000000000000110
ff: mov %r13,%rsi
b2: lea 0x10(%rbx),%esi 102: add $0x10,%esi
b5: callq 0xffffffffe1021e0d 105: callq 0xffffffffe102bd6d
ba: cmp $0x16,%eax 10a: cmp $0x16,%rax
bd: jne 0x00000000000000c6 10e: jne 0x0000000000000117
bf: mov $0xffff,%eax 110: mov $0xffff,%eax
c4: jmp 0x00000000000000c8 115: jmp 0x000000000000011c
c6: xor %eax,%eax 117: mov $0x0,%eax
c8: mov -0x8(%rbp),%rbx 11c: mov -0x228(%rbp),%rbx // epilogue
cc: leaveq 123: mov -0x220(%rbp),%r13
cd: retq 12a: mov -0x218(%rbp),%r14
131: mov -0x210(%rbp),%r15
138: leaveq
139: retq
On fully cached SKBs both JITed functions take 12 nsec to execute.
BPF interpreter executes the program in 30 nsec.
The difference in generated assembler is due to the following:
Old BPF imlements LDX_MSH instruction via sk_load_byte_msh() helper function
inside bpf_jit.S.
New JIT removes the helper and does it explicitly, so ldx_msh cost
is the same for both JITs, but generated code looks longer.
New JIT has 4 registers to save, so prologue/epilogue are larger,
but the cost is within noise on x64.
Old JIT checks whether first insn clears A and if not emits 'xor %eax,%eax'.
New JIT clears %rax unconditionally.
2. old BPF JIT doesn't support ANC_NLATTR, ANC_PAY_OFFSET, ANC_RANDOM
extensions. New JIT supports all BPF extensions.
Performance of such filters improves 2-4 times depending on a filter.
The longer the filter the higher performance gain.
Synthetic benchmarks with many ancillary loads see 20x speedup
which seems to be the maximum gain from JIT
Notes:
. net.core.bpf_jit_enable=2 + tools/net/bpf_jit_disasm is still functional
and can be used to see generated assembler
. there are two jit_compile() functions and code flow for classic filters is:
sk_attach_filter() - load classic BPF
bpf_jit_compile() - try to JIT from classic BPF
sk_convert_filter() - convert classic to internal
bpf_int_jit_compile() - JIT from internal BPF
seccomp and tracing filters will just call bpf_int_jit_compile()
Signed-off-by: Alexei Starovoitov <ast@plumgrid.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
2014-05-14 02:50:46 +00:00
|
|
|
case BPF_JSGE:
|
x86/bpf: Clean up non-standard comments, to make the code more readable
So by chance I looked into x86 assembly in arch/x86/net/bpf_jit_comp.c and
noticed the weird and inconsistent comment style it mistakenly learned from
the networking code:
/* Multi-line comment ...
* ... looks like this.
*/
Fix this to use the standard comment style specified in Documentation/CodingStyle
and used in arch/x86/ as well:
/*
* Multi-line comment ...
* ... looks like this.
*/
Also, to quote Linus's ... more explicit views about this:
http://article.gmane.org/gmane.linux.kernel.cryptoapi/21066
> But no, the networking code picked *none* of the above sane formats.
> Instead, it picked these two models that are just half-arsed
> shit-for-brains:
>
> (no)
> /* This is disgusting drug-induced
> * crap, and should die
> */
>
> (no-no-no)
> /* This is also very nasty
> * and visually unbalanced */
>
> Please. The networking code actually has the *worst* possible comment
> style. You can literally find that (no-no-no) style, which is just
> really horribly disgusting and worse than the otherwise fairly similar
> (d) in pretty much every way.
Also improve the comments and some other details while at it:
- Don't mix same-line and previous-line comment style on otherwise
identical code patterns within the same function,
- capitalize 'BPF' and x86 register names consistently,
- capitalize sentences consistently,
- instead of 'x64' use 'x86-64': x64 is a Microsoft specific term,
- use more consistent punctuation,
- use standard coding style in macros as well,
- fix typos and a few other minor details.
Consistent coding style is not optional, at least in arch/x86/.
No change in functionality.
( In case this commit causes conflicts with pending development code
I'll be glad to help resolve any conflicts! )
Acked-by: Thomas Gleixner <tglx@linutronix.de>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Borislav Petkov <bp@alien8.de>
Cc: H. Peter Anvin <hpa@zytor.com>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: David S. Miller <davem@davemloft.net>
Cc: Eric Dumazet <edumazet@google.com>
Cc: Daniel Borkmann <daniel@iogearbox.net>
Cc: Alexei Starovoitov <ast@fb.com>
Cc: Hideaki YOSHIFUJI <yoshfuji@linux-ipv6.org>
Cc: netdev@vger.kernel.org
Cc: linux-kernel@vger.kernel.org
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
2018-04-27 09:54:40 +00:00
|
|
|
/* Signed '>=', GE in x86 */
|
net: filter: x86: internal BPF JIT
Maps all internal BPF instructions into x86_64 instructions.
This patch replaces original BPF x64 JIT with internal BPF x64 JIT.
sysctl net.core.bpf_jit_enable is reused as on/off switch.
Performance:
1. old BPF JIT and internal BPF JIT generate equivalent x86_64 code.
No performance difference is observed for filters that were JIT-able before
Example assembler code for BPF filter "tcpdump port 22"
original BPF -> old JIT: original BPF -> internal BPF -> new JIT:
0: push %rbp 0: push %rbp
1: mov %rsp,%rbp 1: mov %rsp,%rbp
4: sub $0x60,%rsp 4: sub $0x228,%rsp
8: mov %rbx,-0x8(%rbp) b: mov %rbx,-0x228(%rbp) // prologue
12: mov %r13,-0x220(%rbp)
19: mov %r14,-0x218(%rbp)
20: mov %r15,-0x210(%rbp)
27: xor %eax,%eax // clear A
c: xor %ebx,%ebx 29: xor %r13,%r13 // clear X
e: mov 0x68(%rdi),%r9d 2c: mov 0x68(%rdi),%r9d
12: sub 0x6c(%rdi),%r9d 30: sub 0x6c(%rdi),%r9d
16: mov 0xd8(%rdi),%r8 34: mov 0xd8(%rdi),%r10
3b: mov %rdi,%rbx
1d: mov $0xc,%esi 3e: mov $0xc,%esi
22: callq 0xffffffffe1021e15 43: callq 0xffffffffe102bd75
27: cmp $0x86dd,%eax 48: cmp $0x86dd,%rax
2c: jne 0x0000000000000069 4f: jne 0x000000000000009a
2e: mov $0x14,%esi 51: mov $0x14,%esi
33: callq 0xffffffffe1021e31 56: callq 0xffffffffe102bd91
38: cmp $0x84,%eax 5b: cmp $0x84,%rax
3d: je 0x0000000000000049 62: je 0x0000000000000074
3f: cmp $0x6,%eax 64: cmp $0x6,%rax
42: je 0x0000000000000049 68: je 0x0000000000000074
44: cmp $0x11,%eax 6a: cmp $0x11,%rax
47: jne 0x00000000000000c6 6e: jne 0x0000000000000117
49: mov $0x36,%esi 74: mov $0x36,%esi
4e: callq 0xffffffffe1021e15 79: callq 0xffffffffe102bd75
53: cmp $0x16,%eax 7e: cmp $0x16,%rax
56: je 0x00000000000000bf 82: je 0x0000000000000110
58: mov $0x38,%esi 88: mov $0x38,%esi
5d: callq 0xffffffffe1021e15 8d: callq 0xffffffffe102bd75
62: cmp $0x16,%eax 92: cmp $0x16,%rax
65: je 0x00000000000000bf 96: je 0x0000000000000110
67: jmp 0x00000000000000c6 98: jmp 0x0000000000000117
69: cmp $0x800,%eax 9a: cmp $0x800,%rax
6e: jne 0x00000000000000c6 a1: jne 0x0000000000000117
70: mov $0x17,%esi a3: mov $0x17,%esi
75: callq 0xffffffffe1021e31 a8: callq 0xffffffffe102bd91
7a: cmp $0x84,%eax ad: cmp $0x84,%rax
7f: je 0x000000000000008b b4: je 0x00000000000000c2
81: cmp $0x6,%eax b6: cmp $0x6,%rax
84: je 0x000000000000008b ba: je 0x00000000000000c2
86: cmp $0x11,%eax bc: cmp $0x11,%rax
89: jne 0x00000000000000c6 c0: jne 0x0000000000000117
8b: mov $0x14,%esi c2: mov $0x14,%esi
90: callq 0xffffffffe1021e15 c7: callq 0xffffffffe102bd75
95: test $0x1fff,%ax cc: test $0x1fff,%rax
99: jne 0x00000000000000c6 d3: jne 0x0000000000000117
d5: mov %rax,%r14
9b: mov $0xe,%esi d8: mov $0xe,%esi
a0: callq 0xffffffffe1021e44 dd: callq 0xffffffffe102bd91 // MSH
e2: and $0xf,%eax
e5: shl $0x2,%eax
e8: mov %rax,%r13
eb: mov %r14,%rax
ee: mov %r13,%rsi
a5: lea 0xe(%rbx),%esi f1: add $0xe,%esi
a8: callq 0xffffffffe1021e0d f4: callq 0xffffffffe102bd6d
ad: cmp $0x16,%eax f9: cmp $0x16,%rax
b0: je 0x00000000000000bf fd: je 0x0000000000000110
ff: mov %r13,%rsi
b2: lea 0x10(%rbx),%esi 102: add $0x10,%esi
b5: callq 0xffffffffe1021e0d 105: callq 0xffffffffe102bd6d
ba: cmp $0x16,%eax 10a: cmp $0x16,%rax
bd: jne 0x00000000000000c6 10e: jne 0x0000000000000117
bf: mov $0xffff,%eax 110: mov $0xffff,%eax
c4: jmp 0x00000000000000c8 115: jmp 0x000000000000011c
c6: xor %eax,%eax 117: mov $0x0,%eax
c8: mov -0x8(%rbp),%rbx 11c: mov -0x228(%rbp),%rbx // epilogue
cc: leaveq 123: mov -0x220(%rbp),%r13
cd: retq 12a: mov -0x218(%rbp),%r14
131: mov -0x210(%rbp),%r15
138: leaveq
139: retq
On fully cached SKBs both JITed functions take 12 nsec to execute.
BPF interpreter executes the program in 30 nsec.
The difference in generated assembler is due to the following:
Old BPF imlements LDX_MSH instruction via sk_load_byte_msh() helper function
inside bpf_jit.S.
New JIT removes the helper and does it explicitly, so ldx_msh cost
is the same for both JITs, but generated code looks longer.
New JIT has 4 registers to save, so prologue/epilogue are larger,
but the cost is within noise on x64.
Old JIT checks whether first insn clears A and if not emits 'xor %eax,%eax'.
New JIT clears %rax unconditionally.
2. old BPF JIT doesn't support ANC_NLATTR, ANC_PAY_OFFSET, ANC_RANDOM
extensions. New JIT supports all BPF extensions.
Performance of such filters improves 2-4 times depending on a filter.
The longer the filter the higher performance gain.
Synthetic benchmarks with many ancillary loads see 20x speedup
which seems to be the maximum gain from JIT
Notes:
. net.core.bpf_jit_enable=2 + tools/net/bpf_jit_disasm is still functional
and can be used to see generated assembler
. there are two jit_compile() functions and code flow for classic filters is:
sk_attach_filter() - load classic BPF
bpf_jit_compile() - try to JIT from classic BPF
sk_convert_filter() - convert classic to internal
bpf_int_jit_compile() - JIT from internal BPF
seccomp and tracing filters will just call bpf_int_jit_compile()
Signed-off-by: Alexei Starovoitov <ast@plumgrid.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
2014-05-14 02:50:46 +00:00
|
|
|
jmp_cond = X86_JGE;
|
|
|
|
break;
|
2017-08-09 23:39:56 +00:00
|
|
|
case BPF_JSLE:
|
x86/bpf: Clean up non-standard comments, to make the code more readable
So by chance I looked into x86 assembly in arch/x86/net/bpf_jit_comp.c and
noticed the weird and inconsistent comment style it mistakenly learned from
the networking code:
/* Multi-line comment ...
* ... looks like this.
*/
Fix this to use the standard comment style specified in Documentation/CodingStyle
and used in arch/x86/ as well:
/*
* Multi-line comment ...
* ... looks like this.
*/
Also, to quote Linus's ... more explicit views about this:
http://article.gmane.org/gmane.linux.kernel.cryptoapi/21066
> But no, the networking code picked *none* of the above sane formats.
> Instead, it picked these two models that are just half-arsed
> shit-for-brains:
>
> (no)
> /* This is disgusting drug-induced
> * crap, and should die
> */
>
> (no-no-no)
> /* This is also very nasty
> * and visually unbalanced */
>
> Please. The networking code actually has the *worst* possible comment
> style. You can literally find that (no-no-no) style, which is just
> really horribly disgusting and worse than the otherwise fairly similar
> (d) in pretty much every way.
Also improve the comments and some other details while at it:
- Don't mix same-line and previous-line comment style on otherwise
identical code patterns within the same function,
- capitalize 'BPF' and x86 register names consistently,
- capitalize sentences consistently,
- instead of 'x64' use 'x86-64': x64 is a Microsoft specific term,
- use more consistent punctuation,
- use standard coding style in macros as well,
- fix typos and a few other minor details.
Consistent coding style is not optional, at least in arch/x86/.
No change in functionality.
( In case this commit causes conflicts with pending development code
I'll be glad to help resolve any conflicts! )
Acked-by: Thomas Gleixner <tglx@linutronix.de>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Borislav Petkov <bp@alien8.de>
Cc: H. Peter Anvin <hpa@zytor.com>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: David S. Miller <davem@davemloft.net>
Cc: Eric Dumazet <edumazet@google.com>
Cc: Daniel Borkmann <daniel@iogearbox.net>
Cc: Alexei Starovoitov <ast@fb.com>
Cc: Hideaki YOSHIFUJI <yoshfuji@linux-ipv6.org>
Cc: netdev@vger.kernel.org
Cc: linux-kernel@vger.kernel.org
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
2018-04-27 09:54:40 +00:00
|
|
|
/* Signed '<=', LE in x86 */
|
2017-08-09 23:39:56 +00:00
|
|
|
jmp_cond = X86_JLE;
|
|
|
|
break;
|
x86/bpf: Clean up non-standard comments, to make the code more readable
So by chance I looked into x86 assembly in arch/x86/net/bpf_jit_comp.c and
noticed the weird and inconsistent comment style it mistakenly learned from
the networking code:
/* Multi-line comment ...
* ... looks like this.
*/
Fix this to use the standard comment style specified in Documentation/CodingStyle
and used in arch/x86/ as well:
/*
* Multi-line comment ...
* ... looks like this.
*/
Also, to quote Linus's ... more explicit views about this:
http://article.gmane.org/gmane.linux.kernel.cryptoapi/21066
> But no, the networking code picked *none* of the above sane formats.
> Instead, it picked these two models that are just half-arsed
> shit-for-brains:
>
> (no)
> /* This is disgusting drug-induced
> * crap, and should die
> */
>
> (no-no-no)
> /* This is also very nasty
> * and visually unbalanced */
>
> Please. The networking code actually has the *worst* possible comment
> style. You can literally find that (no-no-no) style, which is just
> really horribly disgusting and worse than the otherwise fairly similar
> (d) in pretty much every way.
Also improve the comments and some other details while at it:
- Don't mix same-line and previous-line comment style on otherwise
identical code patterns within the same function,
- capitalize 'BPF' and x86 register names consistently,
- capitalize sentences consistently,
- instead of 'x64' use 'x86-64': x64 is a Microsoft specific term,
- use more consistent punctuation,
- use standard coding style in macros as well,
- fix typos and a few other minor details.
Consistent coding style is not optional, at least in arch/x86/.
No change in functionality.
( In case this commit causes conflicts with pending development code
I'll be glad to help resolve any conflicts! )
Acked-by: Thomas Gleixner <tglx@linutronix.de>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Borislav Petkov <bp@alien8.de>
Cc: H. Peter Anvin <hpa@zytor.com>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: David S. Miller <davem@davemloft.net>
Cc: Eric Dumazet <edumazet@google.com>
Cc: Daniel Borkmann <daniel@iogearbox.net>
Cc: Alexei Starovoitov <ast@fb.com>
Cc: Hideaki YOSHIFUJI <yoshfuji@linux-ipv6.org>
Cc: netdev@vger.kernel.org
Cc: linux-kernel@vger.kernel.org
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
2018-04-27 09:54:40 +00:00
|
|
|
default: /* to silence GCC warning */
|
net: filter: x86: internal BPF JIT
Maps all internal BPF instructions into x86_64 instructions.
This patch replaces original BPF x64 JIT with internal BPF x64 JIT.
sysctl net.core.bpf_jit_enable is reused as on/off switch.
Performance:
1. old BPF JIT and internal BPF JIT generate equivalent x86_64 code.
No performance difference is observed for filters that were JIT-able before
Example assembler code for BPF filter "tcpdump port 22"
original BPF -> old JIT: original BPF -> internal BPF -> new JIT:
0: push %rbp 0: push %rbp
1: mov %rsp,%rbp 1: mov %rsp,%rbp
4: sub $0x60,%rsp 4: sub $0x228,%rsp
8: mov %rbx,-0x8(%rbp) b: mov %rbx,-0x228(%rbp) // prologue
12: mov %r13,-0x220(%rbp)
19: mov %r14,-0x218(%rbp)
20: mov %r15,-0x210(%rbp)
27: xor %eax,%eax // clear A
c: xor %ebx,%ebx 29: xor %r13,%r13 // clear X
e: mov 0x68(%rdi),%r9d 2c: mov 0x68(%rdi),%r9d
12: sub 0x6c(%rdi),%r9d 30: sub 0x6c(%rdi),%r9d
16: mov 0xd8(%rdi),%r8 34: mov 0xd8(%rdi),%r10
3b: mov %rdi,%rbx
1d: mov $0xc,%esi 3e: mov $0xc,%esi
22: callq 0xffffffffe1021e15 43: callq 0xffffffffe102bd75
27: cmp $0x86dd,%eax 48: cmp $0x86dd,%rax
2c: jne 0x0000000000000069 4f: jne 0x000000000000009a
2e: mov $0x14,%esi 51: mov $0x14,%esi
33: callq 0xffffffffe1021e31 56: callq 0xffffffffe102bd91
38: cmp $0x84,%eax 5b: cmp $0x84,%rax
3d: je 0x0000000000000049 62: je 0x0000000000000074
3f: cmp $0x6,%eax 64: cmp $0x6,%rax
42: je 0x0000000000000049 68: je 0x0000000000000074
44: cmp $0x11,%eax 6a: cmp $0x11,%rax
47: jne 0x00000000000000c6 6e: jne 0x0000000000000117
49: mov $0x36,%esi 74: mov $0x36,%esi
4e: callq 0xffffffffe1021e15 79: callq 0xffffffffe102bd75
53: cmp $0x16,%eax 7e: cmp $0x16,%rax
56: je 0x00000000000000bf 82: je 0x0000000000000110
58: mov $0x38,%esi 88: mov $0x38,%esi
5d: callq 0xffffffffe1021e15 8d: callq 0xffffffffe102bd75
62: cmp $0x16,%eax 92: cmp $0x16,%rax
65: je 0x00000000000000bf 96: je 0x0000000000000110
67: jmp 0x00000000000000c6 98: jmp 0x0000000000000117
69: cmp $0x800,%eax 9a: cmp $0x800,%rax
6e: jne 0x00000000000000c6 a1: jne 0x0000000000000117
70: mov $0x17,%esi a3: mov $0x17,%esi
75: callq 0xffffffffe1021e31 a8: callq 0xffffffffe102bd91
7a: cmp $0x84,%eax ad: cmp $0x84,%rax
7f: je 0x000000000000008b b4: je 0x00000000000000c2
81: cmp $0x6,%eax b6: cmp $0x6,%rax
84: je 0x000000000000008b ba: je 0x00000000000000c2
86: cmp $0x11,%eax bc: cmp $0x11,%rax
89: jne 0x00000000000000c6 c0: jne 0x0000000000000117
8b: mov $0x14,%esi c2: mov $0x14,%esi
90: callq 0xffffffffe1021e15 c7: callq 0xffffffffe102bd75
95: test $0x1fff,%ax cc: test $0x1fff,%rax
99: jne 0x00000000000000c6 d3: jne 0x0000000000000117
d5: mov %rax,%r14
9b: mov $0xe,%esi d8: mov $0xe,%esi
a0: callq 0xffffffffe1021e44 dd: callq 0xffffffffe102bd91 // MSH
e2: and $0xf,%eax
e5: shl $0x2,%eax
e8: mov %rax,%r13
eb: mov %r14,%rax
ee: mov %r13,%rsi
a5: lea 0xe(%rbx),%esi f1: add $0xe,%esi
a8: callq 0xffffffffe1021e0d f4: callq 0xffffffffe102bd6d
ad: cmp $0x16,%eax f9: cmp $0x16,%rax
b0: je 0x00000000000000bf fd: je 0x0000000000000110
ff: mov %r13,%rsi
b2: lea 0x10(%rbx),%esi 102: add $0x10,%esi
b5: callq 0xffffffffe1021e0d 105: callq 0xffffffffe102bd6d
ba: cmp $0x16,%eax 10a: cmp $0x16,%rax
bd: jne 0x00000000000000c6 10e: jne 0x0000000000000117
bf: mov $0xffff,%eax 110: mov $0xffff,%eax
c4: jmp 0x00000000000000c8 115: jmp 0x000000000000011c
c6: xor %eax,%eax 117: mov $0x0,%eax
c8: mov -0x8(%rbp),%rbx 11c: mov -0x228(%rbp),%rbx // epilogue
cc: leaveq 123: mov -0x220(%rbp),%r13
cd: retq 12a: mov -0x218(%rbp),%r14
131: mov -0x210(%rbp),%r15
138: leaveq
139: retq
On fully cached SKBs both JITed functions take 12 nsec to execute.
BPF interpreter executes the program in 30 nsec.
The difference in generated assembler is due to the following:
Old BPF imlements LDX_MSH instruction via sk_load_byte_msh() helper function
inside bpf_jit.S.
New JIT removes the helper and does it explicitly, so ldx_msh cost
is the same for both JITs, but generated code looks longer.
New JIT has 4 registers to save, so prologue/epilogue are larger,
but the cost is within noise on x64.
Old JIT checks whether first insn clears A and if not emits 'xor %eax,%eax'.
New JIT clears %rax unconditionally.
2. old BPF JIT doesn't support ANC_NLATTR, ANC_PAY_OFFSET, ANC_RANDOM
extensions. New JIT supports all BPF extensions.
Performance of such filters improves 2-4 times depending on a filter.
The longer the filter the higher performance gain.
Synthetic benchmarks with many ancillary loads see 20x speedup
which seems to be the maximum gain from JIT
Notes:
. net.core.bpf_jit_enable=2 + tools/net/bpf_jit_disasm is still functional
and can be used to see generated assembler
. there are two jit_compile() functions and code flow for classic filters is:
sk_attach_filter() - load classic BPF
bpf_jit_compile() - try to JIT from classic BPF
sk_convert_filter() - convert classic to internal
bpf_int_jit_compile() - JIT from internal BPF
seccomp and tracing filters will just call bpf_int_jit_compile()
Signed-off-by: Alexei Starovoitov <ast@plumgrid.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
2014-05-14 02:50:46 +00:00
|
|
|
return -EFAULT;
|
|
|
|
}
|
|
|
|
jmp_offset = addrs[i + insn->off] - addrs[i];
|
|
|
|
if (is_imm8(jmp_offset)) {
|
bpf,x64: Pad NOPs to make images converge more easily
The x64 bpf jit expects bpf images converge within the given passes, but
it could fail to do so with some corner cases. For example:
l0: ja 40
l1: ja 40
[... repeated ja 40 ]
l39: ja 40
l40: ret #0
This bpf program contains 40 "ja 40" instructions which are effectively
NOPs and designed to be replaced with valid code dynamically. Ideally,
bpf jit should optimize those "ja 40" instructions out when translating
the bpf instructions into x64 machine code. However, do_jit() can only
remove one "ja 40" for offset==0 on each pass, so it requires at least
40 runs to eliminate those JMPs and exceeds the current limit of
passes(20). In the end, the program got rejected when BPF_JIT_ALWAYS_ON
is set even though it's legit as a classic socket filter.
To make bpf images more likely converge within 20 passes, this commit
pads some instructions with NOPs in the last 5 passes:
1. conditional jumps
A possible size variance comes from the adoption of imm8 JMP. If the
offset is imm8, we calculate the size difference of this BPF instruction
between the previous and the current pass and fill the gap with NOPs.
To avoid the recalculation of jump offset, those NOPs are inserted before
the JMP code, so we have to subtract the 2 bytes of imm8 JMP when
calculating the NOP number.
2. BPF_JA
There are two conditions for BPF_JA.
a.) nop jumps
If this instruction is not optimized out in the previous pass,
instead of removing it, we insert the equivalent size of NOPs.
b.) label jumps
Similar to condition jumps, we prepend NOPs right before the JMP
code.
To make the code concise, emit_nops() is modified to use the signed len and
return the number of inserted NOPs.
For bpf-to-bpf, we always enable padding for the extra pass since there
is only one extra run and the jump padding doesn't affected the images
that converge without padding.
After applying this patch, the corner case was loaded with the following
jit code:
flen=45 proglen=77 pass=17 image=ffffffffc03367d4 from=jump pid=10097
JIT code: 00000000: 0f 1f 44 00 00 55 48 89 e5 53 41 55 31 c0 45 31
JIT code: 00000010: ed 48 89 fb eb 30 eb 2e eb 2c eb 2a eb 28 eb 26
JIT code: 00000020: eb 24 eb 22 eb 20 eb 1e eb 1c eb 1a eb 18 eb 16
JIT code: 00000030: eb 14 eb 12 eb 10 eb 0e eb 0c eb 0a eb 08 eb 06
JIT code: 00000040: eb 04 eb 02 66 90 31 c0 41 5d 5b c9 c3
0: 0f 1f 44 00 00 nop DWORD PTR [rax+rax*1+0x0]
5: 55 push rbp
6: 48 89 e5 mov rbp,rsp
9: 53 push rbx
a: 41 55 push r13
c: 31 c0 xor eax,eax
e: 45 31 ed xor r13d,r13d
11: 48 89 fb mov rbx,rdi
14: eb 30 jmp 0x46
16: eb 2e jmp 0x46
...
3e: eb 06 jmp 0x46
40: eb 04 jmp 0x46
42: eb 02 jmp 0x46
44: 66 90 xchg ax,ax
46: 31 c0 xor eax,eax
48: 41 5d pop r13
4a: 5b pop rbx
4b: c9 leave
4c: c3 ret
At the 16th pass, 15 jumps were already optimized out, and one jump was
replaced with NOPs at 44 and the image converged at the 17th pass.
v4:
- Add the detailed comments about the possible padding bytes
v3:
- Copy the instructions of prologue separately or the size calculation
of the first BPF instruction would include the prologue.
- Replace WARN_ONCE() with pr_err() and EFAULT
- Use MAX_PASSES in the for loop condition check
- Remove the "padded" flag from x64_jit_data. For the extra pass of
subprogs, padding is always enabled since it won't hurt the images
that converge without padding.
v2:
- Simplify the sample code in the description and provide the jit code
- Check the expected padding bytes with WARN_ONCE
- Move the 'padded' flag to 'struct x64_jit_data'
Signed-off-by: Gary Lin <glin@suse.com>
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Link: https://lore.kernel.org/bpf/20210119102501.511-2-glin@suse.com
2021-01-19 10:24:59 +00:00
|
|
|
if (jmp_padding) {
|
|
|
|
/* To keep the jmp_offset valid, the extra bytes are
|
2021-03-18 14:28:01 +00:00
|
|
|
* padded before the jump insn, so we subtract the
|
bpf,x64: Pad NOPs to make images converge more easily
The x64 bpf jit expects bpf images converge within the given passes, but
it could fail to do so with some corner cases. For example:
l0: ja 40
l1: ja 40
[... repeated ja 40 ]
l39: ja 40
l40: ret #0
This bpf program contains 40 "ja 40" instructions which are effectively
NOPs and designed to be replaced with valid code dynamically. Ideally,
bpf jit should optimize those "ja 40" instructions out when translating
the bpf instructions into x64 machine code. However, do_jit() can only
remove one "ja 40" for offset==0 on each pass, so it requires at least
40 runs to eliminate those JMPs and exceeds the current limit of
passes(20). In the end, the program got rejected when BPF_JIT_ALWAYS_ON
is set even though it's legit as a classic socket filter.
To make bpf images more likely converge within 20 passes, this commit
pads some instructions with NOPs in the last 5 passes:
1. conditional jumps
A possible size variance comes from the adoption of imm8 JMP. If the
offset is imm8, we calculate the size difference of this BPF instruction
between the previous and the current pass and fill the gap with NOPs.
To avoid the recalculation of jump offset, those NOPs are inserted before
the JMP code, so we have to subtract the 2 bytes of imm8 JMP when
calculating the NOP number.
2. BPF_JA
There are two conditions for BPF_JA.
a.) nop jumps
If this instruction is not optimized out in the previous pass,
instead of removing it, we insert the equivalent size of NOPs.
b.) label jumps
Similar to condition jumps, we prepend NOPs right before the JMP
code.
To make the code concise, emit_nops() is modified to use the signed len and
return the number of inserted NOPs.
For bpf-to-bpf, we always enable padding for the extra pass since there
is only one extra run and the jump padding doesn't affected the images
that converge without padding.
After applying this patch, the corner case was loaded with the following
jit code:
flen=45 proglen=77 pass=17 image=ffffffffc03367d4 from=jump pid=10097
JIT code: 00000000: 0f 1f 44 00 00 55 48 89 e5 53 41 55 31 c0 45 31
JIT code: 00000010: ed 48 89 fb eb 30 eb 2e eb 2c eb 2a eb 28 eb 26
JIT code: 00000020: eb 24 eb 22 eb 20 eb 1e eb 1c eb 1a eb 18 eb 16
JIT code: 00000030: eb 14 eb 12 eb 10 eb 0e eb 0c eb 0a eb 08 eb 06
JIT code: 00000040: eb 04 eb 02 66 90 31 c0 41 5d 5b c9 c3
0: 0f 1f 44 00 00 nop DWORD PTR [rax+rax*1+0x0]
5: 55 push rbp
6: 48 89 e5 mov rbp,rsp
9: 53 push rbx
a: 41 55 push r13
c: 31 c0 xor eax,eax
e: 45 31 ed xor r13d,r13d
11: 48 89 fb mov rbx,rdi
14: eb 30 jmp 0x46
16: eb 2e jmp 0x46
...
3e: eb 06 jmp 0x46
40: eb 04 jmp 0x46
42: eb 02 jmp 0x46
44: 66 90 xchg ax,ax
46: 31 c0 xor eax,eax
48: 41 5d pop r13
4a: 5b pop rbx
4b: c9 leave
4c: c3 ret
At the 16th pass, 15 jumps were already optimized out, and one jump was
replaced with NOPs at 44 and the image converged at the 17th pass.
v4:
- Add the detailed comments about the possible padding bytes
v3:
- Copy the instructions of prologue separately or the size calculation
of the first BPF instruction would include the prologue.
- Replace WARN_ONCE() with pr_err() and EFAULT
- Use MAX_PASSES in the for loop condition check
- Remove the "padded" flag from x64_jit_data. For the extra pass of
subprogs, padding is always enabled since it won't hurt the images
that converge without padding.
v2:
- Simplify the sample code in the description and provide the jit code
- Check the expected padding bytes with WARN_ONCE
- Move the 'padded' flag to 'struct x64_jit_data'
Signed-off-by: Gary Lin <glin@suse.com>
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Link: https://lore.kernel.org/bpf/20210119102501.511-2-glin@suse.com
2021-01-19 10:24:59 +00:00
|
|
|
* 2 bytes of jmp_cond insn from INSN_SZ_DIFF.
|
|
|
|
*
|
|
|
|
* If the previous pass already emits an imm8
|
|
|
|
* jmp_cond, then this BPF insn won't shrink, so
|
|
|
|
* "nops" is 0.
|
|
|
|
*
|
|
|
|
* On the other hand, if the previous pass emits an
|
|
|
|
* imm32 jmp_cond, the extra 4 bytes(*) is padded to
|
|
|
|
* keep the image from shrinking further.
|
|
|
|
*
|
|
|
|
* (*) imm32 jmp_cond is 6 bytes, and imm8 jmp_cond
|
|
|
|
* is 2 bytes, so the size difference is 4 bytes.
|
|
|
|
*/
|
|
|
|
nops = INSN_SZ_DIFF - 2;
|
|
|
|
if (nops != 0 && nops != 4) {
|
|
|
|
pr_err("unexpected jmp_cond padding: %d bytes\n",
|
|
|
|
nops);
|
|
|
|
return -EFAULT;
|
|
|
|
}
|
2021-06-23 11:25:04 +00:00
|
|
|
emit_nops(&prog, nops);
|
bpf,x64: Pad NOPs to make images converge more easily
The x64 bpf jit expects bpf images converge within the given passes, but
it could fail to do so with some corner cases. For example:
l0: ja 40
l1: ja 40
[... repeated ja 40 ]
l39: ja 40
l40: ret #0
This bpf program contains 40 "ja 40" instructions which are effectively
NOPs and designed to be replaced with valid code dynamically. Ideally,
bpf jit should optimize those "ja 40" instructions out when translating
the bpf instructions into x64 machine code. However, do_jit() can only
remove one "ja 40" for offset==0 on each pass, so it requires at least
40 runs to eliminate those JMPs and exceeds the current limit of
passes(20). In the end, the program got rejected when BPF_JIT_ALWAYS_ON
is set even though it's legit as a classic socket filter.
To make bpf images more likely converge within 20 passes, this commit
pads some instructions with NOPs in the last 5 passes:
1. conditional jumps
A possible size variance comes from the adoption of imm8 JMP. If the
offset is imm8, we calculate the size difference of this BPF instruction
between the previous and the current pass and fill the gap with NOPs.
To avoid the recalculation of jump offset, those NOPs are inserted before
the JMP code, so we have to subtract the 2 bytes of imm8 JMP when
calculating the NOP number.
2. BPF_JA
There are two conditions for BPF_JA.
a.) nop jumps
If this instruction is not optimized out in the previous pass,
instead of removing it, we insert the equivalent size of NOPs.
b.) label jumps
Similar to condition jumps, we prepend NOPs right before the JMP
code.
To make the code concise, emit_nops() is modified to use the signed len and
return the number of inserted NOPs.
For bpf-to-bpf, we always enable padding for the extra pass since there
is only one extra run and the jump padding doesn't affected the images
that converge without padding.
After applying this patch, the corner case was loaded with the following
jit code:
flen=45 proglen=77 pass=17 image=ffffffffc03367d4 from=jump pid=10097
JIT code: 00000000: 0f 1f 44 00 00 55 48 89 e5 53 41 55 31 c0 45 31
JIT code: 00000010: ed 48 89 fb eb 30 eb 2e eb 2c eb 2a eb 28 eb 26
JIT code: 00000020: eb 24 eb 22 eb 20 eb 1e eb 1c eb 1a eb 18 eb 16
JIT code: 00000030: eb 14 eb 12 eb 10 eb 0e eb 0c eb 0a eb 08 eb 06
JIT code: 00000040: eb 04 eb 02 66 90 31 c0 41 5d 5b c9 c3
0: 0f 1f 44 00 00 nop DWORD PTR [rax+rax*1+0x0]
5: 55 push rbp
6: 48 89 e5 mov rbp,rsp
9: 53 push rbx
a: 41 55 push r13
c: 31 c0 xor eax,eax
e: 45 31 ed xor r13d,r13d
11: 48 89 fb mov rbx,rdi
14: eb 30 jmp 0x46
16: eb 2e jmp 0x46
...
3e: eb 06 jmp 0x46
40: eb 04 jmp 0x46
42: eb 02 jmp 0x46
44: 66 90 xchg ax,ax
46: 31 c0 xor eax,eax
48: 41 5d pop r13
4a: 5b pop rbx
4b: c9 leave
4c: c3 ret
At the 16th pass, 15 jumps were already optimized out, and one jump was
replaced with NOPs at 44 and the image converged at the 17th pass.
v4:
- Add the detailed comments about the possible padding bytes
v3:
- Copy the instructions of prologue separately or the size calculation
of the first BPF instruction would include the prologue.
- Replace WARN_ONCE() with pr_err() and EFAULT
- Use MAX_PASSES in the for loop condition check
- Remove the "padded" flag from x64_jit_data. For the extra pass of
subprogs, padding is always enabled since it won't hurt the images
that converge without padding.
v2:
- Simplify the sample code in the description and provide the jit code
- Check the expected padding bytes with WARN_ONCE
- Move the 'padded' flag to 'struct x64_jit_data'
Signed-off-by: Gary Lin <glin@suse.com>
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Link: https://lore.kernel.org/bpf/20210119102501.511-2-glin@suse.com
2021-01-19 10:24:59 +00:00
|
|
|
}
|
net: filter: x86: internal BPF JIT
Maps all internal BPF instructions into x86_64 instructions.
This patch replaces original BPF x64 JIT with internal BPF x64 JIT.
sysctl net.core.bpf_jit_enable is reused as on/off switch.
Performance:
1. old BPF JIT and internal BPF JIT generate equivalent x86_64 code.
No performance difference is observed for filters that were JIT-able before
Example assembler code for BPF filter "tcpdump port 22"
original BPF -> old JIT: original BPF -> internal BPF -> new JIT:
0: push %rbp 0: push %rbp
1: mov %rsp,%rbp 1: mov %rsp,%rbp
4: sub $0x60,%rsp 4: sub $0x228,%rsp
8: mov %rbx,-0x8(%rbp) b: mov %rbx,-0x228(%rbp) // prologue
12: mov %r13,-0x220(%rbp)
19: mov %r14,-0x218(%rbp)
20: mov %r15,-0x210(%rbp)
27: xor %eax,%eax // clear A
c: xor %ebx,%ebx 29: xor %r13,%r13 // clear X
e: mov 0x68(%rdi),%r9d 2c: mov 0x68(%rdi),%r9d
12: sub 0x6c(%rdi),%r9d 30: sub 0x6c(%rdi),%r9d
16: mov 0xd8(%rdi),%r8 34: mov 0xd8(%rdi),%r10
3b: mov %rdi,%rbx
1d: mov $0xc,%esi 3e: mov $0xc,%esi
22: callq 0xffffffffe1021e15 43: callq 0xffffffffe102bd75
27: cmp $0x86dd,%eax 48: cmp $0x86dd,%rax
2c: jne 0x0000000000000069 4f: jne 0x000000000000009a
2e: mov $0x14,%esi 51: mov $0x14,%esi
33: callq 0xffffffffe1021e31 56: callq 0xffffffffe102bd91
38: cmp $0x84,%eax 5b: cmp $0x84,%rax
3d: je 0x0000000000000049 62: je 0x0000000000000074
3f: cmp $0x6,%eax 64: cmp $0x6,%rax
42: je 0x0000000000000049 68: je 0x0000000000000074
44: cmp $0x11,%eax 6a: cmp $0x11,%rax
47: jne 0x00000000000000c6 6e: jne 0x0000000000000117
49: mov $0x36,%esi 74: mov $0x36,%esi
4e: callq 0xffffffffe1021e15 79: callq 0xffffffffe102bd75
53: cmp $0x16,%eax 7e: cmp $0x16,%rax
56: je 0x00000000000000bf 82: je 0x0000000000000110
58: mov $0x38,%esi 88: mov $0x38,%esi
5d: callq 0xffffffffe1021e15 8d: callq 0xffffffffe102bd75
62: cmp $0x16,%eax 92: cmp $0x16,%rax
65: je 0x00000000000000bf 96: je 0x0000000000000110
67: jmp 0x00000000000000c6 98: jmp 0x0000000000000117
69: cmp $0x800,%eax 9a: cmp $0x800,%rax
6e: jne 0x00000000000000c6 a1: jne 0x0000000000000117
70: mov $0x17,%esi a3: mov $0x17,%esi
75: callq 0xffffffffe1021e31 a8: callq 0xffffffffe102bd91
7a: cmp $0x84,%eax ad: cmp $0x84,%rax
7f: je 0x000000000000008b b4: je 0x00000000000000c2
81: cmp $0x6,%eax b6: cmp $0x6,%rax
84: je 0x000000000000008b ba: je 0x00000000000000c2
86: cmp $0x11,%eax bc: cmp $0x11,%rax
89: jne 0x00000000000000c6 c0: jne 0x0000000000000117
8b: mov $0x14,%esi c2: mov $0x14,%esi
90: callq 0xffffffffe1021e15 c7: callq 0xffffffffe102bd75
95: test $0x1fff,%ax cc: test $0x1fff,%rax
99: jne 0x00000000000000c6 d3: jne 0x0000000000000117
d5: mov %rax,%r14
9b: mov $0xe,%esi d8: mov $0xe,%esi
a0: callq 0xffffffffe1021e44 dd: callq 0xffffffffe102bd91 // MSH
e2: and $0xf,%eax
e5: shl $0x2,%eax
e8: mov %rax,%r13
eb: mov %r14,%rax
ee: mov %r13,%rsi
a5: lea 0xe(%rbx),%esi f1: add $0xe,%esi
a8: callq 0xffffffffe1021e0d f4: callq 0xffffffffe102bd6d
ad: cmp $0x16,%eax f9: cmp $0x16,%rax
b0: je 0x00000000000000bf fd: je 0x0000000000000110
ff: mov %r13,%rsi
b2: lea 0x10(%rbx),%esi 102: add $0x10,%esi
b5: callq 0xffffffffe1021e0d 105: callq 0xffffffffe102bd6d
ba: cmp $0x16,%eax 10a: cmp $0x16,%rax
bd: jne 0x00000000000000c6 10e: jne 0x0000000000000117
bf: mov $0xffff,%eax 110: mov $0xffff,%eax
c4: jmp 0x00000000000000c8 115: jmp 0x000000000000011c
c6: xor %eax,%eax 117: mov $0x0,%eax
c8: mov -0x8(%rbp),%rbx 11c: mov -0x228(%rbp),%rbx // epilogue
cc: leaveq 123: mov -0x220(%rbp),%r13
cd: retq 12a: mov -0x218(%rbp),%r14
131: mov -0x210(%rbp),%r15
138: leaveq
139: retq
On fully cached SKBs both JITed functions take 12 nsec to execute.
BPF interpreter executes the program in 30 nsec.
The difference in generated assembler is due to the following:
Old BPF imlements LDX_MSH instruction via sk_load_byte_msh() helper function
inside bpf_jit.S.
New JIT removes the helper and does it explicitly, so ldx_msh cost
is the same for both JITs, but generated code looks longer.
New JIT has 4 registers to save, so prologue/epilogue are larger,
but the cost is within noise on x64.
Old JIT checks whether first insn clears A and if not emits 'xor %eax,%eax'.
New JIT clears %rax unconditionally.
2. old BPF JIT doesn't support ANC_NLATTR, ANC_PAY_OFFSET, ANC_RANDOM
extensions. New JIT supports all BPF extensions.
Performance of such filters improves 2-4 times depending on a filter.
The longer the filter the higher performance gain.
Synthetic benchmarks with many ancillary loads see 20x speedup
which seems to be the maximum gain from JIT
Notes:
. net.core.bpf_jit_enable=2 + tools/net/bpf_jit_disasm is still functional
and can be used to see generated assembler
. there are two jit_compile() functions and code flow for classic filters is:
sk_attach_filter() - load classic BPF
bpf_jit_compile() - try to JIT from classic BPF
sk_convert_filter() - convert classic to internal
bpf_int_jit_compile() - JIT from internal BPF
seccomp and tracing filters will just call bpf_int_jit_compile()
Signed-off-by: Alexei Starovoitov <ast@plumgrid.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
2014-05-14 02:50:46 +00:00
|
|
|
EMIT2(jmp_cond, jmp_offset);
|
|
|
|
} else if (is_simm32(jmp_offset)) {
|
|
|
|
EMIT2_off32(0x0F, jmp_cond + 0x10, jmp_offset);
|
|
|
|
} else {
|
|
|
|
pr_err("cond_jmp gen bug %llx\n", jmp_offset);
|
|
|
|
return -EFAULT;
|
|
|
|
}
|
|
|
|
|
|
|
|
break;
|
2011-04-20 09:27:32 +00:00
|
|
|
|
net: filter: x86: internal BPF JIT
Maps all internal BPF instructions into x86_64 instructions.
This patch replaces original BPF x64 JIT with internal BPF x64 JIT.
sysctl net.core.bpf_jit_enable is reused as on/off switch.
Performance:
1. old BPF JIT and internal BPF JIT generate equivalent x86_64 code.
No performance difference is observed for filters that were JIT-able before
Example assembler code for BPF filter "tcpdump port 22"
original BPF -> old JIT: original BPF -> internal BPF -> new JIT:
0: push %rbp 0: push %rbp
1: mov %rsp,%rbp 1: mov %rsp,%rbp
4: sub $0x60,%rsp 4: sub $0x228,%rsp
8: mov %rbx,-0x8(%rbp) b: mov %rbx,-0x228(%rbp) // prologue
12: mov %r13,-0x220(%rbp)
19: mov %r14,-0x218(%rbp)
20: mov %r15,-0x210(%rbp)
27: xor %eax,%eax // clear A
c: xor %ebx,%ebx 29: xor %r13,%r13 // clear X
e: mov 0x68(%rdi),%r9d 2c: mov 0x68(%rdi),%r9d
12: sub 0x6c(%rdi),%r9d 30: sub 0x6c(%rdi),%r9d
16: mov 0xd8(%rdi),%r8 34: mov 0xd8(%rdi),%r10
3b: mov %rdi,%rbx
1d: mov $0xc,%esi 3e: mov $0xc,%esi
22: callq 0xffffffffe1021e15 43: callq 0xffffffffe102bd75
27: cmp $0x86dd,%eax 48: cmp $0x86dd,%rax
2c: jne 0x0000000000000069 4f: jne 0x000000000000009a
2e: mov $0x14,%esi 51: mov $0x14,%esi
33: callq 0xffffffffe1021e31 56: callq 0xffffffffe102bd91
38: cmp $0x84,%eax 5b: cmp $0x84,%rax
3d: je 0x0000000000000049 62: je 0x0000000000000074
3f: cmp $0x6,%eax 64: cmp $0x6,%rax
42: je 0x0000000000000049 68: je 0x0000000000000074
44: cmp $0x11,%eax 6a: cmp $0x11,%rax
47: jne 0x00000000000000c6 6e: jne 0x0000000000000117
49: mov $0x36,%esi 74: mov $0x36,%esi
4e: callq 0xffffffffe1021e15 79: callq 0xffffffffe102bd75
53: cmp $0x16,%eax 7e: cmp $0x16,%rax
56: je 0x00000000000000bf 82: je 0x0000000000000110
58: mov $0x38,%esi 88: mov $0x38,%esi
5d: callq 0xffffffffe1021e15 8d: callq 0xffffffffe102bd75
62: cmp $0x16,%eax 92: cmp $0x16,%rax
65: je 0x00000000000000bf 96: je 0x0000000000000110
67: jmp 0x00000000000000c6 98: jmp 0x0000000000000117
69: cmp $0x800,%eax 9a: cmp $0x800,%rax
6e: jne 0x00000000000000c6 a1: jne 0x0000000000000117
70: mov $0x17,%esi a3: mov $0x17,%esi
75: callq 0xffffffffe1021e31 a8: callq 0xffffffffe102bd91
7a: cmp $0x84,%eax ad: cmp $0x84,%rax
7f: je 0x000000000000008b b4: je 0x00000000000000c2
81: cmp $0x6,%eax b6: cmp $0x6,%rax
84: je 0x000000000000008b ba: je 0x00000000000000c2
86: cmp $0x11,%eax bc: cmp $0x11,%rax
89: jne 0x00000000000000c6 c0: jne 0x0000000000000117
8b: mov $0x14,%esi c2: mov $0x14,%esi
90: callq 0xffffffffe1021e15 c7: callq 0xffffffffe102bd75
95: test $0x1fff,%ax cc: test $0x1fff,%rax
99: jne 0x00000000000000c6 d3: jne 0x0000000000000117
d5: mov %rax,%r14
9b: mov $0xe,%esi d8: mov $0xe,%esi
a0: callq 0xffffffffe1021e44 dd: callq 0xffffffffe102bd91 // MSH
e2: and $0xf,%eax
e5: shl $0x2,%eax
e8: mov %rax,%r13
eb: mov %r14,%rax
ee: mov %r13,%rsi
a5: lea 0xe(%rbx),%esi f1: add $0xe,%esi
a8: callq 0xffffffffe1021e0d f4: callq 0xffffffffe102bd6d
ad: cmp $0x16,%eax f9: cmp $0x16,%rax
b0: je 0x00000000000000bf fd: je 0x0000000000000110
ff: mov %r13,%rsi
b2: lea 0x10(%rbx),%esi 102: add $0x10,%esi
b5: callq 0xffffffffe1021e0d 105: callq 0xffffffffe102bd6d
ba: cmp $0x16,%eax 10a: cmp $0x16,%rax
bd: jne 0x00000000000000c6 10e: jne 0x0000000000000117
bf: mov $0xffff,%eax 110: mov $0xffff,%eax
c4: jmp 0x00000000000000c8 115: jmp 0x000000000000011c
c6: xor %eax,%eax 117: mov $0x0,%eax
c8: mov -0x8(%rbp),%rbx 11c: mov -0x228(%rbp),%rbx // epilogue
cc: leaveq 123: mov -0x220(%rbp),%r13
cd: retq 12a: mov -0x218(%rbp),%r14
131: mov -0x210(%rbp),%r15
138: leaveq
139: retq
On fully cached SKBs both JITed functions take 12 nsec to execute.
BPF interpreter executes the program in 30 nsec.
The difference in generated assembler is due to the following:
Old BPF imlements LDX_MSH instruction via sk_load_byte_msh() helper function
inside bpf_jit.S.
New JIT removes the helper and does it explicitly, so ldx_msh cost
is the same for both JITs, but generated code looks longer.
New JIT has 4 registers to save, so prologue/epilogue are larger,
but the cost is within noise on x64.
Old JIT checks whether first insn clears A and if not emits 'xor %eax,%eax'.
New JIT clears %rax unconditionally.
2. old BPF JIT doesn't support ANC_NLATTR, ANC_PAY_OFFSET, ANC_RANDOM
extensions. New JIT supports all BPF extensions.
Performance of such filters improves 2-4 times depending on a filter.
The longer the filter the higher performance gain.
Synthetic benchmarks with many ancillary loads see 20x speedup
which seems to be the maximum gain from JIT
Notes:
. net.core.bpf_jit_enable=2 + tools/net/bpf_jit_disasm is still functional
and can be used to see generated assembler
. there are two jit_compile() functions and code flow for classic filters is:
sk_attach_filter() - load classic BPF
bpf_jit_compile() - try to JIT from classic BPF
sk_convert_filter() - convert classic to internal
bpf_int_jit_compile() - JIT from internal BPF
seccomp and tracing filters will just call bpf_int_jit_compile()
Signed-off-by: Alexei Starovoitov <ast@plumgrid.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
2014-05-14 02:50:46 +00:00
|
|
|
case BPF_JMP | BPF_JA:
|
2023-07-28 01:12:31 +00:00
|
|
|
case BPF_JMP32 | BPF_JA:
|
|
|
|
if (BPF_CLASS(insn->code) == BPF_JMP) {
|
|
|
|
if (insn->off == -1)
|
|
|
|
/* -1 jmp instructions will always jump
|
|
|
|
* backwards two bytes. Explicitly handling
|
|
|
|
* this case avoids wasting too many passes
|
|
|
|
* when there are long sequences of replaced
|
|
|
|
* dead code.
|
|
|
|
*/
|
|
|
|
jmp_offset = -2;
|
|
|
|
else
|
|
|
|
jmp_offset = addrs[i + insn->off] - addrs[i];
|
|
|
|
} else {
|
|
|
|
if (insn->imm == -1)
|
|
|
|
jmp_offset = -2;
|
|
|
|
else
|
|
|
|
jmp_offset = addrs[i + insn->imm] - addrs[i];
|
|
|
|
}
|
bpf, x64: fix JIT emission for dead code
Commit 2a5418a13fcf ("bpf: improve dead code sanitizing") replaced dead
code with a series of ja-1 instructions, for safety. That made JIT
compilation much more complex for some BPF programs. One instance of such
programs is, for example:
bool flag = false
...
/* A bunch of other code */
...
if (flag)
do_something()
In some cases llvm is not able to remove at compile time the code for
do_something(), so the generated BPF program ends up with a large amount
of dead instructions. In one specific real life example, there are two
series of ~500 and ~1000 dead instructions in the program. When the
verifier replaces them with a series of ja-1 instructions, it causes an
interesting behavior at JIT time.
During the first pass, since all the instructions are estimated at 64
bytes, the ja-1 instructions end up being translated as 5 bytes JMP
instructions (0xE9), since the jump offsets become increasingly large (>
127) as each instruction gets discovered to be 5 bytes instead of the
estimated 64.
Starting from the second pass, the first N instructions of the ja-1
sequence get translated into 2 bytes JMPs (0xEB) because the jump offsets
become <= 127 this time. In particular, N is defined as roughly 127 / (5
- 2) ~= 42. So, each further pass will make the subsequent N JMP
instructions shrink from 5 to 2 bytes, making the image shrink every time.
This means that in order to have the entire program converge, there need
to be, in the real example above, at least ~1000 / 42 ~= 24 passes just
for translating the dead code. If we add this number to the passes needed
to translate the other non dead code, it brings such program to 40+
passes, and JIT doesn't complete. Ultimately the userspace loader fails
because such BPF program was supposed to be part of a prog array owner
being JITed.
While it is certainly possible to try to refactor such programs to help
the compiler remove dead code, the behavior is not really intuitive and it
puts further burden on the BPF developer who is not expecting such
behavior. To make things worse, such programs are working just fine in all
the kernel releases prior to the ja-1 fix.
A possible approach to mitigate this behavior consists into noticing that
for ja-1 instructions we don't really need to rely on the estimated size
of the previous and current instructions, we know that a -1 BPF jump
offset can be safely translated into a 0xEB instruction with a jump offset
of -2.
Such fix brings the BPF program in the previous example to complete again
in ~9 passes.
Fixes: 2a5418a13fcf ("bpf: improve dead code sanitizing")
Signed-off-by: Gianluca Borello <g.borello@gmail.com>
Acked-by: Alexei Starovoitov <ast@kernel.org>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
2018-04-25 05:42:16 +00:00
|
|
|
|
bpf,x64: Pad NOPs to make images converge more easily
The x64 bpf jit expects bpf images converge within the given passes, but
it could fail to do so with some corner cases. For example:
l0: ja 40
l1: ja 40
[... repeated ja 40 ]
l39: ja 40
l40: ret #0
This bpf program contains 40 "ja 40" instructions which are effectively
NOPs and designed to be replaced with valid code dynamically. Ideally,
bpf jit should optimize those "ja 40" instructions out when translating
the bpf instructions into x64 machine code. However, do_jit() can only
remove one "ja 40" for offset==0 on each pass, so it requires at least
40 runs to eliminate those JMPs and exceeds the current limit of
passes(20). In the end, the program got rejected when BPF_JIT_ALWAYS_ON
is set even though it's legit as a classic socket filter.
To make bpf images more likely converge within 20 passes, this commit
pads some instructions with NOPs in the last 5 passes:
1. conditional jumps
A possible size variance comes from the adoption of imm8 JMP. If the
offset is imm8, we calculate the size difference of this BPF instruction
between the previous and the current pass and fill the gap with NOPs.
To avoid the recalculation of jump offset, those NOPs are inserted before
the JMP code, so we have to subtract the 2 bytes of imm8 JMP when
calculating the NOP number.
2. BPF_JA
There are two conditions for BPF_JA.
a.) nop jumps
If this instruction is not optimized out in the previous pass,
instead of removing it, we insert the equivalent size of NOPs.
b.) label jumps
Similar to condition jumps, we prepend NOPs right before the JMP
code.
To make the code concise, emit_nops() is modified to use the signed len and
return the number of inserted NOPs.
For bpf-to-bpf, we always enable padding for the extra pass since there
is only one extra run and the jump padding doesn't affected the images
that converge without padding.
After applying this patch, the corner case was loaded with the following
jit code:
flen=45 proglen=77 pass=17 image=ffffffffc03367d4 from=jump pid=10097
JIT code: 00000000: 0f 1f 44 00 00 55 48 89 e5 53 41 55 31 c0 45 31
JIT code: 00000010: ed 48 89 fb eb 30 eb 2e eb 2c eb 2a eb 28 eb 26
JIT code: 00000020: eb 24 eb 22 eb 20 eb 1e eb 1c eb 1a eb 18 eb 16
JIT code: 00000030: eb 14 eb 12 eb 10 eb 0e eb 0c eb 0a eb 08 eb 06
JIT code: 00000040: eb 04 eb 02 66 90 31 c0 41 5d 5b c9 c3
0: 0f 1f 44 00 00 nop DWORD PTR [rax+rax*1+0x0]
5: 55 push rbp
6: 48 89 e5 mov rbp,rsp
9: 53 push rbx
a: 41 55 push r13
c: 31 c0 xor eax,eax
e: 45 31 ed xor r13d,r13d
11: 48 89 fb mov rbx,rdi
14: eb 30 jmp 0x46
16: eb 2e jmp 0x46
...
3e: eb 06 jmp 0x46
40: eb 04 jmp 0x46
42: eb 02 jmp 0x46
44: 66 90 xchg ax,ax
46: 31 c0 xor eax,eax
48: 41 5d pop r13
4a: 5b pop rbx
4b: c9 leave
4c: c3 ret
At the 16th pass, 15 jumps were already optimized out, and one jump was
replaced with NOPs at 44 and the image converged at the 17th pass.
v4:
- Add the detailed comments about the possible padding bytes
v3:
- Copy the instructions of prologue separately or the size calculation
of the first BPF instruction would include the prologue.
- Replace WARN_ONCE() with pr_err() and EFAULT
- Use MAX_PASSES in the for loop condition check
- Remove the "padded" flag from x64_jit_data. For the extra pass of
subprogs, padding is always enabled since it won't hurt the images
that converge without padding.
v2:
- Simplify the sample code in the description and provide the jit code
- Check the expected padding bytes with WARN_ONCE
- Move the 'padded' flag to 'struct x64_jit_data'
Signed-off-by: Gary Lin <glin@suse.com>
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Link: https://lore.kernel.org/bpf/20210119102501.511-2-glin@suse.com
2021-01-19 10:24:59 +00:00
|
|
|
if (!jmp_offset) {
|
|
|
|
/*
|
|
|
|
* If jmp_padding is enabled, the extra nops will
|
|
|
|
* be inserted. Otherwise, optimize out nop jumps.
|
|
|
|
*/
|
|
|
|
if (jmp_padding) {
|
|
|
|
/* There are 3 possible conditions.
|
|
|
|
* (1) This BPF_JA is already optimized out in
|
|
|
|
* the previous run, so there is no need
|
|
|
|
* to pad any extra byte (0 byte).
|
|
|
|
* (2) The previous pass emits an imm8 jmp,
|
|
|
|
* so we pad 2 bytes to match the previous
|
|
|
|
* insn size.
|
|
|
|
* (3) Similarly, the previous pass emits an
|
|
|
|
* imm32 jmp, and 5 bytes is padded.
|
|
|
|
*/
|
|
|
|
nops = INSN_SZ_DIFF;
|
|
|
|
if (nops != 0 && nops != 2 && nops != 5) {
|
|
|
|
pr_err("unexpected nop jump padding: %d bytes\n",
|
|
|
|
nops);
|
|
|
|
return -EFAULT;
|
|
|
|
}
|
2021-06-23 11:25:04 +00:00
|
|
|
emit_nops(&prog, nops);
|
bpf,x64: Pad NOPs to make images converge more easily
The x64 bpf jit expects bpf images converge within the given passes, but
it could fail to do so with some corner cases. For example:
l0: ja 40
l1: ja 40
[... repeated ja 40 ]
l39: ja 40
l40: ret #0
This bpf program contains 40 "ja 40" instructions which are effectively
NOPs and designed to be replaced with valid code dynamically. Ideally,
bpf jit should optimize those "ja 40" instructions out when translating
the bpf instructions into x64 machine code. However, do_jit() can only
remove one "ja 40" for offset==0 on each pass, so it requires at least
40 runs to eliminate those JMPs and exceeds the current limit of
passes(20). In the end, the program got rejected when BPF_JIT_ALWAYS_ON
is set even though it's legit as a classic socket filter.
To make bpf images more likely converge within 20 passes, this commit
pads some instructions with NOPs in the last 5 passes:
1. conditional jumps
A possible size variance comes from the adoption of imm8 JMP. If the
offset is imm8, we calculate the size difference of this BPF instruction
between the previous and the current pass and fill the gap with NOPs.
To avoid the recalculation of jump offset, those NOPs are inserted before
the JMP code, so we have to subtract the 2 bytes of imm8 JMP when
calculating the NOP number.
2. BPF_JA
There are two conditions for BPF_JA.
a.) nop jumps
If this instruction is not optimized out in the previous pass,
instead of removing it, we insert the equivalent size of NOPs.
b.) label jumps
Similar to condition jumps, we prepend NOPs right before the JMP
code.
To make the code concise, emit_nops() is modified to use the signed len and
return the number of inserted NOPs.
For bpf-to-bpf, we always enable padding for the extra pass since there
is only one extra run and the jump padding doesn't affected the images
that converge without padding.
After applying this patch, the corner case was loaded with the following
jit code:
flen=45 proglen=77 pass=17 image=ffffffffc03367d4 from=jump pid=10097
JIT code: 00000000: 0f 1f 44 00 00 55 48 89 e5 53 41 55 31 c0 45 31
JIT code: 00000010: ed 48 89 fb eb 30 eb 2e eb 2c eb 2a eb 28 eb 26
JIT code: 00000020: eb 24 eb 22 eb 20 eb 1e eb 1c eb 1a eb 18 eb 16
JIT code: 00000030: eb 14 eb 12 eb 10 eb 0e eb 0c eb 0a eb 08 eb 06
JIT code: 00000040: eb 04 eb 02 66 90 31 c0 41 5d 5b c9 c3
0: 0f 1f 44 00 00 nop DWORD PTR [rax+rax*1+0x0]
5: 55 push rbp
6: 48 89 e5 mov rbp,rsp
9: 53 push rbx
a: 41 55 push r13
c: 31 c0 xor eax,eax
e: 45 31 ed xor r13d,r13d
11: 48 89 fb mov rbx,rdi
14: eb 30 jmp 0x46
16: eb 2e jmp 0x46
...
3e: eb 06 jmp 0x46
40: eb 04 jmp 0x46
42: eb 02 jmp 0x46
44: 66 90 xchg ax,ax
46: 31 c0 xor eax,eax
48: 41 5d pop r13
4a: 5b pop rbx
4b: c9 leave
4c: c3 ret
At the 16th pass, 15 jumps were already optimized out, and one jump was
replaced with NOPs at 44 and the image converged at the 17th pass.
v4:
- Add the detailed comments about the possible padding bytes
v3:
- Copy the instructions of prologue separately or the size calculation
of the first BPF instruction would include the prologue.
- Replace WARN_ONCE() with pr_err() and EFAULT
- Use MAX_PASSES in the for loop condition check
- Remove the "padded" flag from x64_jit_data. For the extra pass of
subprogs, padding is always enabled since it won't hurt the images
that converge without padding.
v2:
- Simplify the sample code in the description and provide the jit code
- Check the expected padding bytes with WARN_ONCE
- Move the 'padded' flag to 'struct x64_jit_data'
Signed-off-by: Gary Lin <glin@suse.com>
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Link: https://lore.kernel.org/bpf/20210119102501.511-2-glin@suse.com
2021-01-19 10:24:59 +00:00
|
|
|
}
|
net: filter: x86: internal BPF JIT
Maps all internal BPF instructions into x86_64 instructions.
This patch replaces original BPF x64 JIT with internal BPF x64 JIT.
sysctl net.core.bpf_jit_enable is reused as on/off switch.
Performance:
1. old BPF JIT and internal BPF JIT generate equivalent x86_64 code.
No performance difference is observed for filters that were JIT-able before
Example assembler code for BPF filter "tcpdump port 22"
original BPF -> old JIT: original BPF -> internal BPF -> new JIT:
0: push %rbp 0: push %rbp
1: mov %rsp,%rbp 1: mov %rsp,%rbp
4: sub $0x60,%rsp 4: sub $0x228,%rsp
8: mov %rbx,-0x8(%rbp) b: mov %rbx,-0x228(%rbp) // prologue
12: mov %r13,-0x220(%rbp)
19: mov %r14,-0x218(%rbp)
20: mov %r15,-0x210(%rbp)
27: xor %eax,%eax // clear A
c: xor %ebx,%ebx 29: xor %r13,%r13 // clear X
e: mov 0x68(%rdi),%r9d 2c: mov 0x68(%rdi),%r9d
12: sub 0x6c(%rdi),%r9d 30: sub 0x6c(%rdi),%r9d
16: mov 0xd8(%rdi),%r8 34: mov 0xd8(%rdi),%r10
3b: mov %rdi,%rbx
1d: mov $0xc,%esi 3e: mov $0xc,%esi
22: callq 0xffffffffe1021e15 43: callq 0xffffffffe102bd75
27: cmp $0x86dd,%eax 48: cmp $0x86dd,%rax
2c: jne 0x0000000000000069 4f: jne 0x000000000000009a
2e: mov $0x14,%esi 51: mov $0x14,%esi
33: callq 0xffffffffe1021e31 56: callq 0xffffffffe102bd91
38: cmp $0x84,%eax 5b: cmp $0x84,%rax
3d: je 0x0000000000000049 62: je 0x0000000000000074
3f: cmp $0x6,%eax 64: cmp $0x6,%rax
42: je 0x0000000000000049 68: je 0x0000000000000074
44: cmp $0x11,%eax 6a: cmp $0x11,%rax
47: jne 0x00000000000000c6 6e: jne 0x0000000000000117
49: mov $0x36,%esi 74: mov $0x36,%esi
4e: callq 0xffffffffe1021e15 79: callq 0xffffffffe102bd75
53: cmp $0x16,%eax 7e: cmp $0x16,%rax
56: je 0x00000000000000bf 82: je 0x0000000000000110
58: mov $0x38,%esi 88: mov $0x38,%esi
5d: callq 0xffffffffe1021e15 8d: callq 0xffffffffe102bd75
62: cmp $0x16,%eax 92: cmp $0x16,%rax
65: je 0x00000000000000bf 96: je 0x0000000000000110
67: jmp 0x00000000000000c6 98: jmp 0x0000000000000117
69: cmp $0x800,%eax 9a: cmp $0x800,%rax
6e: jne 0x00000000000000c6 a1: jne 0x0000000000000117
70: mov $0x17,%esi a3: mov $0x17,%esi
75: callq 0xffffffffe1021e31 a8: callq 0xffffffffe102bd91
7a: cmp $0x84,%eax ad: cmp $0x84,%rax
7f: je 0x000000000000008b b4: je 0x00000000000000c2
81: cmp $0x6,%eax b6: cmp $0x6,%rax
84: je 0x000000000000008b ba: je 0x00000000000000c2
86: cmp $0x11,%eax bc: cmp $0x11,%rax
89: jne 0x00000000000000c6 c0: jne 0x0000000000000117
8b: mov $0x14,%esi c2: mov $0x14,%esi
90: callq 0xffffffffe1021e15 c7: callq 0xffffffffe102bd75
95: test $0x1fff,%ax cc: test $0x1fff,%rax
99: jne 0x00000000000000c6 d3: jne 0x0000000000000117
d5: mov %rax,%r14
9b: mov $0xe,%esi d8: mov $0xe,%esi
a0: callq 0xffffffffe1021e44 dd: callq 0xffffffffe102bd91 // MSH
e2: and $0xf,%eax
e5: shl $0x2,%eax
e8: mov %rax,%r13
eb: mov %r14,%rax
ee: mov %r13,%rsi
a5: lea 0xe(%rbx),%esi f1: add $0xe,%esi
a8: callq 0xffffffffe1021e0d f4: callq 0xffffffffe102bd6d
ad: cmp $0x16,%eax f9: cmp $0x16,%rax
b0: je 0x00000000000000bf fd: je 0x0000000000000110
ff: mov %r13,%rsi
b2: lea 0x10(%rbx),%esi 102: add $0x10,%esi
b5: callq 0xffffffffe1021e0d 105: callq 0xffffffffe102bd6d
ba: cmp $0x16,%eax 10a: cmp $0x16,%rax
bd: jne 0x00000000000000c6 10e: jne 0x0000000000000117
bf: mov $0xffff,%eax 110: mov $0xffff,%eax
c4: jmp 0x00000000000000c8 115: jmp 0x000000000000011c
c6: xor %eax,%eax 117: mov $0x0,%eax
c8: mov -0x8(%rbp),%rbx 11c: mov -0x228(%rbp),%rbx // epilogue
cc: leaveq 123: mov -0x220(%rbp),%r13
cd: retq 12a: mov -0x218(%rbp),%r14
131: mov -0x210(%rbp),%r15
138: leaveq
139: retq
On fully cached SKBs both JITed functions take 12 nsec to execute.
BPF interpreter executes the program in 30 nsec.
The difference in generated assembler is due to the following:
Old BPF imlements LDX_MSH instruction via sk_load_byte_msh() helper function
inside bpf_jit.S.
New JIT removes the helper and does it explicitly, so ldx_msh cost
is the same for both JITs, but generated code looks longer.
New JIT has 4 registers to save, so prologue/epilogue are larger,
but the cost is within noise on x64.
Old JIT checks whether first insn clears A and if not emits 'xor %eax,%eax'.
New JIT clears %rax unconditionally.
2. old BPF JIT doesn't support ANC_NLATTR, ANC_PAY_OFFSET, ANC_RANDOM
extensions. New JIT supports all BPF extensions.
Performance of such filters improves 2-4 times depending on a filter.
The longer the filter the higher performance gain.
Synthetic benchmarks with many ancillary loads see 20x speedup
which seems to be the maximum gain from JIT
Notes:
. net.core.bpf_jit_enable=2 + tools/net/bpf_jit_disasm is still functional
and can be used to see generated assembler
. there are two jit_compile() functions and code flow for classic filters is:
sk_attach_filter() - load classic BPF
bpf_jit_compile() - try to JIT from classic BPF
sk_convert_filter() - convert classic to internal
bpf_int_jit_compile() - JIT from internal BPF
seccomp and tracing filters will just call bpf_int_jit_compile()
Signed-off-by: Alexei Starovoitov <ast@plumgrid.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
2014-05-14 02:50:46 +00:00
|
|
|
break;
|
bpf,x64: Pad NOPs to make images converge more easily
The x64 bpf jit expects bpf images converge within the given passes, but
it could fail to do so with some corner cases. For example:
l0: ja 40
l1: ja 40
[... repeated ja 40 ]
l39: ja 40
l40: ret #0
This bpf program contains 40 "ja 40" instructions which are effectively
NOPs and designed to be replaced with valid code dynamically. Ideally,
bpf jit should optimize those "ja 40" instructions out when translating
the bpf instructions into x64 machine code. However, do_jit() can only
remove one "ja 40" for offset==0 on each pass, so it requires at least
40 runs to eliminate those JMPs and exceeds the current limit of
passes(20). In the end, the program got rejected when BPF_JIT_ALWAYS_ON
is set even though it's legit as a classic socket filter.
To make bpf images more likely converge within 20 passes, this commit
pads some instructions with NOPs in the last 5 passes:
1. conditional jumps
A possible size variance comes from the adoption of imm8 JMP. If the
offset is imm8, we calculate the size difference of this BPF instruction
between the previous and the current pass and fill the gap with NOPs.
To avoid the recalculation of jump offset, those NOPs are inserted before
the JMP code, so we have to subtract the 2 bytes of imm8 JMP when
calculating the NOP number.
2. BPF_JA
There are two conditions for BPF_JA.
a.) nop jumps
If this instruction is not optimized out in the previous pass,
instead of removing it, we insert the equivalent size of NOPs.
b.) label jumps
Similar to condition jumps, we prepend NOPs right before the JMP
code.
To make the code concise, emit_nops() is modified to use the signed len and
return the number of inserted NOPs.
For bpf-to-bpf, we always enable padding for the extra pass since there
is only one extra run and the jump padding doesn't affected the images
that converge without padding.
After applying this patch, the corner case was loaded with the following
jit code:
flen=45 proglen=77 pass=17 image=ffffffffc03367d4 from=jump pid=10097
JIT code: 00000000: 0f 1f 44 00 00 55 48 89 e5 53 41 55 31 c0 45 31
JIT code: 00000010: ed 48 89 fb eb 30 eb 2e eb 2c eb 2a eb 28 eb 26
JIT code: 00000020: eb 24 eb 22 eb 20 eb 1e eb 1c eb 1a eb 18 eb 16
JIT code: 00000030: eb 14 eb 12 eb 10 eb 0e eb 0c eb 0a eb 08 eb 06
JIT code: 00000040: eb 04 eb 02 66 90 31 c0 41 5d 5b c9 c3
0: 0f 1f 44 00 00 nop DWORD PTR [rax+rax*1+0x0]
5: 55 push rbp
6: 48 89 e5 mov rbp,rsp
9: 53 push rbx
a: 41 55 push r13
c: 31 c0 xor eax,eax
e: 45 31 ed xor r13d,r13d
11: 48 89 fb mov rbx,rdi
14: eb 30 jmp 0x46
16: eb 2e jmp 0x46
...
3e: eb 06 jmp 0x46
40: eb 04 jmp 0x46
42: eb 02 jmp 0x46
44: 66 90 xchg ax,ax
46: 31 c0 xor eax,eax
48: 41 5d pop r13
4a: 5b pop rbx
4b: c9 leave
4c: c3 ret
At the 16th pass, 15 jumps were already optimized out, and one jump was
replaced with NOPs at 44 and the image converged at the 17th pass.
v4:
- Add the detailed comments about the possible padding bytes
v3:
- Copy the instructions of prologue separately or the size calculation
of the first BPF instruction would include the prologue.
- Replace WARN_ONCE() with pr_err() and EFAULT
- Use MAX_PASSES in the for loop condition check
- Remove the "padded" flag from x64_jit_data. For the extra pass of
subprogs, padding is always enabled since it won't hurt the images
that converge without padding.
v2:
- Simplify the sample code in the description and provide the jit code
- Check the expected padding bytes with WARN_ONCE
- Move the 'padded' flag to 'struct x64_jit_data'
Signed-off-by: Gary Lin <glin@suse.com>
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Link: https://lore.kernel.org/bpf/20210119102501.511-2-glin@suse.com
2021-01-19 10:24:59 +00:00
|
|
|
}
|
net: filter: x86: internal BPF JIT
Maps all internal BPF instructions into x86_64 instructions.
This patch replaces original BPF x64 JIT with internal BPF x64 JIT.
sysctl net.core.bpf_jit_enable is reused as on/off switch.
Performance:
1. old BPF JIT and internal BPF JIT generate equivalent x86_64 code.
No performance difference is observed for filters that were JIT-able before
Example assembler code for BPF filter "tcpdump port 22"
original BPF -> old JIT: original BPF -> internal BPF -> new JIT:
0: push %rbp 0: push %rbp
1: mov %rsp,%rbp 1: mov %rsp,%rbp
4: sub $0x60,%rsp 4: sub $0x228,%rsp
8: mov %rbx,-0x8(%rbp) b: mov %rbx,-0x228(%rbp) // prologue
12: mov %r13,-0x220(%rbp)
19: mov %r14,-0x218(%rbp)
20: mov %r15,-0x210(%rbp)
27: xor %eax,%eax // clear A
c: xor %ebx,%ebx 29: xor %r13,%r13 // clear X
e: mov 0x68(%rdi),%r9d 2c: mov 0x68(%rdi),%r9d
12: sub 0x6c(%rdi),%r9d 30: sub 0x6c(%rdi),%r9d
16: mov 0xd8(%rdi),%r8 34: mov 0xd8(%rdi),%r10
3b: mov %rdi,%rbx
1d: mov $0xc,%esi 3e: mov $0xc,%esi
22: callq 0xffffffffe1021e15 43: callq 0xffffffffe102bd75
27: cmp $0x86dd,%eax 48: cmp $0x86dd,%rax
2c: jne 0x0000000000000069 4f: jne 0x000000000000009a
2e: mov $0x14,%esi 51: mov $0x14,%esi
33: callq 0xffffffffe1021e31 56: callq 0xffffffffe102bd91
38: cmp $0x84,%eax 5b: cmp $0x84,%rax
3d: je 0x0000000000000049 62: je 0x0000000000000074
3f: cmp $0x6,%eax 64: cmp $0x6,%rax
42: je 0x0000000000000049 68: je 0x0000000000000074
44: cmp $0x11,%eax 6a: cmp $0x11,%rax
47: jne 0x00000000000000c6 6e: jne 0x0000000000000117
49: mov $0x36,%esi 74: mov $0x36,%esi
4e: callq 0xffffffffe1021e15 79: callq 0xffffffffe102bd75
53: cmp $0x16,%eax 7e: cmp $0x16,%rax
56: je 0x00000000000000bf 82: je 0x0000000000000110
58: mov $0x38,%esi 88: mov $0x38,%esi
5d: callq 0xffffffffe1021e15 8d: callq 0xffffffffe102bd75
62: cmp $0x16,%eax 92: cmp $0x16,%rax
65: je 0x00000000000000bf 96: je 0x0000000000000110
67: jmp 0x00000000000000c6 98: jmp 0x0000000000000117
69: cmp $0x800,%eax 9a: cmp $0x800,%rax
6e: jne 0x00000000000000c6 a1: jne 0x0000000000000117
70: mov $0x17,%esi a3: mov $0x17,%esi
75: callq 0xffffffffe1021e31 a8: callq 0xffffffffe102bd91
7a: cmp $0x84,%eax ad: cmp $0x84,%rax
7f: je 0x000000000000008b b4: je 0x00000000000000c2
81: cmp $0x6,%eax b6: cmp $0x6,%rax
84: je 0x000000000000008b ba: je 0x00000000000000c2
86: cmp $0x11,%eax bc: cmp $0x11,%rax
89: jne 0x00000000000000c6 c0: jne 0x0000000000000117
8b: mov $0x14,%esi c2: mov $0x14,%esi
90: callq 0xffffffffe1021e15 c7: callq 0xffffffffe102bd75
95: test $0x1fff,%ax cc: test $0x1fff,%rax
99: jne 0x00000000000000c6 d3: jne 0x0000000000000117
d5: mov %rax,%r14
9b: mov $0xe,%esi d8: mov $0xe,%esi
a0: callq 0xffffffffe1021e44 dd: callq 0xffffffffe102bd91 // MSH
e2: and $0xf,%eax
e5: shl $0x2,%eax
e8: mov %rax,%r13
eb: mov %r14,%rax
ee: mov %r13,%rsi
a5: lea 0xe(%rbx),%esi f1: add $0xe,%esi
a8: callq 0xffffffffe1021e0d f4: callq 0xffffffffe102bd6d
ad: cmp $0x16,%eax f9: cmp $0x16,%rax
b0: je 0x00000000000000bf fd: je 0x0000000000000110
ff: mov %r13,%rsi
b2: lea 0x10(%rbx),%esi 102: add $0x10,%esi
b5: callq 0xffffffffe1021e0d 105: callq 0xffffffffe102bd6d
ba: cmp $0x16,%eax 10a: cmp $0x16,%rax
bd: jne 0x00000000000000c6 10e: jne 0x0000000000000117
bf: mov $0xffff,%eax 110: mov $0xffff,%eax
c4: jmp 0x00000000000000c8 115: jmp 0x000000000000011c
c6: xor %eax,%eax 117: mov $0x0,%eax
c8: mov -0x8(%rbp),%rbx 11c: mov -0x228(%rbp),%rbx // epilogue
cc: leaveq 123: mov -0x220(%rbp),%r13
cd: retq 12a: mov -0x218(%rbp),%r14
131: mov -0x210(%rbp),%r15
138: leaveq
139: retq
On fully cached SKBs both JITed functions take 12 nsec to execute.
BPF interpreter executes the program in 30 nsec.
The difference in generated assembler is due to the following:
Old BPF imlements LDX_MSH instruction via sk_load_byte_msh() helper function
inside bpf_jit.S.
New JIT removes the helper and does it explicitly, so ldx_msh cost
is the same for both JITs, but generated code looks longer.
New JIT has 4 registers to save, so prologue/epilogue are larger,
but the cost is within noise on x64.
Old JIT checks whether first insn clears A and if not emits 'xor %eax,%eax'.
New JIT clears %rax unconditionally.
2. old BPF JIT doesn't support ANC_NLATTR, ANC_PAY_OFFSET, ANC_RANDOM
extensions. New JIT supports all BPF extensions.
Performance of such filters improves 2-4 times depending on a filter.
The longer the filter the higher performance gain.
Synthetic benchmarks with many ancillary loads see 20x speedup
which seems to be the maximum gain from JIT
Notes:
. net.core.bpf_jit_enable=2 + tools/net/bpf_jit_disasm is still functional
and can be used to see generated assembler
. there are two jit_compile() functions and code flow for classic filters is:
sk_attach_filter() - load classic BPF
bpf_jit_compile() - try to JIT from classic BPF
sk_convert_filter() - convert classic to internal
bpf_int_jit_compile() - JIT from internal BPF
seccomp and tracing filters will just call bpf_int_jit_compile()
Signed-off-by: Alexei Starovoitov <ast@plumgrid.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
2014-05-14 02:50:46 +00:00
|
|
|
emit_jmp:
|
|
|
|
if (is_imm8(jmp_offset)) {
|
bpf,x64: Pad NOPs to make images converge more easily
The x64 bpf jit expects bpf images converge within the given passes, but
it could fail to do so with some corner cases. For example:
l0: ja 40
l1: ja 40
[... repeated ja 40 ]
l39: ja 40
l40: ret #0
This bpf program contains 40 "ja 40" instructions which are effectively
NOPs and designed to be replaced with valid code dynamically. Ideally,
bpf jit should optimize those "ja 40" instructions out when translating
the bpf instructions into x64 machine code. However, do_jit() can only
remove one "ja 40" for offset==0 on each pass, so it requires at least
40 runs to eliminate those JMPs and exceeds the current limit of
passes(20). In the end, the program got rejected when BPF_JIT_ALWAYS_ON
is set even though it's legit as a classic socket filter.
To make bpf images more likely converge within 20 passes, this commit
pads some instructions with NOPs in the last 5 passes:
1. conditional jumps
A possible size variance comes from the adoption of imm8 JMP. If the
offset is imm8, we calculate the size difference of this BPF instruction
between the previous and the current pass and fill the gap with NOPs.
To avoid the recalculation of jump offset, those NOPs are inserted before
the JMP code, so we have to subtract the 2 bytes of imm8 JMP when
calculating the NOP number.
2. BPF_JA
There are two conditions for BPF_JA.
a.) nop jumps
If this instruction is not optimized out in the previous pass,
instead of removing it, we insert the equivalent size of NOPs.
b.) label jumps
Similar to condition jumps, we prepend NOPs right before the JMP
code.
To make the code concise, emit_nops() is modified to use the signed len and
return the number of inserted NOPs.
For bpf-to-bpf, we always enable padding for the extra pass since there
is only one extra run and the jump padding doesn't affected the images
that converge without padding.
After applying this patch, the corner case was loaded with the following
jit code:
flen=45 proglen=77 pass=17 image=ffffffffc03367d4 from=jump pid=10097
JIT code: 00000000: 0f 1f 44 00 00 55 48 89 e5 53 41 55 31 c0 45 31
JIT code: 00000010: ed 48 89 fb eb 30 eb 2e eb 2c eb 2a eb 28 eb 26
JIT code: 00000020: eb 24 eb 22 eb 20 eb 1e eb 1c eb 1a eb 18 eb 16
JIT code: 00000030: eb 14 eb 12 eb 10 eb 0e eb 0c eb 0a eb 08 eb 06
JIT code: 00000040: eb 04 eb 02 66 90 31 c0 41 5d 5b c9 c3
0: 0f 1f 44 00 00 nop DWORD PTR [rax+rax*1+0x0]
5: 55 push rbp
6: 48 89 e5 mov rbp,rsp
9: 53 push rbx
a: 41 55 push r13
c: 31 c0 xor eax,eax
e: 45 31 ed xor r13d,r13d
11: 48 89 fb mov rbx,rdi
14: eb 30 jmp 0x46
16: eb 2e jmp 0x46
...
3e: eb 06 jmp 0x46
40: eb 04 jmp 0x46
42: eb 02 jmp 0x46
44: 66 90 xchg ax,ax
46: 31 c0 xor eax,eax
48: 41 5d pop r13
4a: 5b pop rbx
4b: c9 leave
4c: c3 ret
At the 16th pass, 15 jumps were already optimized out, and one jump was
replaced with NOPs at 44 and the image converged at the 17th pass.
v4:
- Add the detailed comments about the possible padding bytes
v3:
- Copy the instructions of prologue separately or the size calculation
of the first BPF instruction would include the prologue.
- Replace WARN_ONCE() with pr_err() and EFAULT
- Use MAX_PASSES in the for loop condition check
- Remove the "padded" flag from x64_jit_data. For the extra pass of
subprogs, padding is always enabled since it won't hurt the images
that converge without padding.
v2:
- Simplify the sample code in the description and provide the jit code
- Check the expected padding bytes with WARN_ONCE
- Move the 'padded' flag to 'struct x64_jit_data'
Signed-off-by: Gary Lin <glin@suse.com>
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Link: https://lore.kernel.org/bpf/20210119102501.511-2-glin@suse.com
2021-01-19 10:24:59 +00:00
|
|
|
if (jmp_padding) {
|
|
|
|
/* To avoid breaking jmp_offset, the extra bytes
|
|
|
|
* are padded before the actual jmp insn, so
|
2021-03-18 14:28:01 +00:00
|
|
|
* 2 bytes is subtracted from INSN_SZ_DIFF.
|
bpf,x64: Pad NOPs to make images converge more easily
The x64 bpf jit expects bpf images converge within the given passes, but
it could fail to do so with some corner cases. For example:
l0: ja 40
l1: ja 40
[... repeated ja 40 ]
l39: ja 40
l40: ret #0
This bpf program contains 40 "ja 40" instructions which are effectively
NOPs and designed to be replaced with valid code dynamically. Ideally,
bpf jit should optimize those "ja 40" instructions out when translating
the bpf instructions into x64 machine code. However, do_jit() can only
remove one "ja 40" for offset==0 on each pass, so it requires at least
40 runs to eliminate those JMPs and exceeds the current limit of
passes(20). In the end, the program got rejected when BPF_JIT_ALWAYS_ON
is set even though it's legit as a classic socket filter.
To make bpf images more likely converge within 20 passes, this commit
pads some instructions with NOPs in the last 5 passes:
1. conditional jumps
A possible size variance comes from the adoption of imm8 JMP. If the
offset is imm8, we calculate the size difference of this BPF instruction
between the previous and the current pass and fill the gap with NOPs.
To avoid the recalculation of jump offset, those NOPs are inserted before
the JMP code, so we have to subtract the 2 bytes of imm8 JMP when
calculating the NOP number.
2. BPF_JA
There are two conditions for BPF_JA.
a.) nop jumps
If this instruction is not optimized out in the previous pass,
instead of removing it, we insert the equivalent size of NOPs.
b.) label jumps
Similar to condition jumps, we prepend NOPs right before the JMP
code.
To make the code concise, emit_nops() is modified to use the signed len and
return the number of inserted NOPs.
For bpf-to-bpf, we always enable padding for the extra pass since there
is only one extra run and the jump padding doesn't affected the images
that converge without padding.
After applying this patch, the corner case was loaded with the following
jit code:
flen=45 proglen=77 pass=17 image=ffffffffc03367d4 from=jump pid=10097
JIT code: 00000000: 0f 1f 44 00 00 55 48 89 e5 53 41 55 31 c0 45 31
JIT code: 00000010: ed 48 89 fb eb 30 eb 2e eb 2c eb 2a eb 28 eb 26
JIT code: 00000020: eb 24 eb 22 eb 20 eb 1e eb 1c eb 1a eb 18 eb 16
JIT code: 00000030: eb 14 eb 12 eb 10 eb 0e eb 0c eb 0a eb 08 eb 06
JIT code: 00000040: eb 04 eb 02 66 90 31 c0 41 5d 5b c9 c3
0: 0f 1f 44 00 00 nop DWORD PTR [rax+rax*1+0x0]
5: 55 push rbp
6: 48 89 e5 mov rbp,rsp
9: 53 push rbx
a: 41 55 push r13
c: 31 c0 xor eax,eax
e: 45 31 ed xor r13d,r13d
11: 48 89 fb mov rbx,rdi
14: eb 30 jmp 0x46
16: eb 2e jmp 0x46
...
3e: eb 06 jmp 0x46
40: eb 04 jmp 0x46
42: eb 02 jmp 0x46
44: 66 90 xchg ax,ax
46: 31 c0 xor eax,eax
48: 41 5d pop r13
4a: 5b pop rbx
4b: c9 leave
4c: c3 ret
At the 16th pass, 15 jumps were already optimized out, and one jump was
replaced with NOPs at 44 and the image converged at the 17th pass.
v4:
- Add the detailed comments about the possible padding bytes
v3:
- Copy the instructions of prologue separately or the size calculation
of the first BPF instruction would include the prologue.
- Replace WARN_ONCE() with pr_err() and EFAULT
- Use MAX_PASSES in the for loop condition check
- Remove the "padded" flag from x64_jit_data. For the extra pass of
subprogs, padding is always enabled since it won't hurt the images
that converge without padding.
v2:
- Simplify the sample code in the description and provide the jit code
- Check the expected padding bytes with WARN_ONCE
- Move the 'padded' flag to 'struct x64_jit_data'
Signed-off-by: Gary Lin <glin@suse.com>
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Link: https://lore.kernel.org/bpf/20210119102501.511-2-glin@suse.com
2021-01-19 10:24:59 +00:00
|
|
|
*
|
|
|
|
* If the previous pass already emits an imm8
|
|
|
|
* jmp, there is nothing to pad (0 byte).
|
|
|
|
*
|
|
|
|
* If it emits an imm32 jmp (5 bytes) previously
|
|
|
|
* and now an imm8 jmp (2 bytes), then we pad
|
|
|
|
* (5 - 2 = 3) bytes to stop the image from
|
|
|
|
* shrinking further.
|
|
|
|
*/
|
|
|
|
nops = INSN_SZ_DIFF - 2;
|
|
|
|
if (nops != 0 && nops != 3) {
|
|
|
|
pr_err("unexpected jump padding: %d bytes\n",
|
|
|
|
nops);
|
|
|
|
return -EFAULT;
|
|
|
|
}
|
2021-06-23 11:25:04 +00:00
|
|
|
emit_nops(&prog, INSN_SZ_DIFF - 2);
|
bpf,x64: Pad NOPs to make images converge more easily
The x64 bpf jit expects bpf images converge within the given passes, but
it could fail to do so with some corner cases. For example:
l0: ja 40
l1: ja 40
[... repeated ja 40 ]
l39: ja 40
l40: ret #0
This bpf program contains 40 "ja 40" instructions which are effectively
NOPs and designed to be replaced with valid code dynamically. Ideally,
bpf jit should optimize those "ja 40" instructions out when translating
the bpf instructions into x64 machine code. However, do_jit() can only
remove one "ja 40" for offset==0 on each pass, so it requires at least
40 runs to eliminate those JMPs and exceeds the current limit of
passes(20). In the end, the program got rejected when BPF_JIT_ALWAYS_ON
is set even though it's legit as a classic socket filter.
To make bpf images more likely converge within 20 passes, this commit
pads some instructions with NOPs in the last 5 passes:
1. conditional jumps
A possible size variance comes from the adoption of imm8 JMP. If the
offset is imm8, we calculate the size difference of this BPF instruction
between the previous and the current pass and fill the gap with NOPs.
To avoid the recalculation of jump offset, those NOPs are inserted before
the JMP code, so we have to subtract the 2 bytes of imm8 JMP when
calculating the NOP number.
2. BPF_JA
There are two conditions for BPF_JA.
a.) nop jumps
If this instruction is not optimized out in the previous pass,
instead of removing it, we insert the equivalent size of NOPs.
b.) label jumps
Similar to condition jumps, we prepend NOPs right before the JMP
code.
To make the code concise, emit_nops() is modified to use the signed len and
return the number of inserted NOPs.
For bpf-to-bpf, we always enable padding for the extra pass since there
is only one extra run and the jump padding doesn't affected the images
that converge without padding.
After applying this patch, the corner case was loaded with the following
jit code:
flen=45 proglen=77 pass=17 image=ffffffffc03367d4 from=jump pid=10097
JIT code: 00000000: 0f 1f 44 00 00 55 48 89 e5 53 41 55 31 c0 45 31
JIT code: 00000010: ed 48 89 fb eb 30 eb 2e eb 2c eb 2a eb 28 eb 26
JIT code: 00000020: eb 24 eb 22 eb 20 eb 1e eb 1c eb 1a eb 18 eb 16
JIT code: 00000030: eb 14 eb 12 eb 10 eb 0e eb 0c eb 0a eb 08 eb 06
JIT code: 00000040: eb 04 eb 02 66 90 31 c0 41 5d 5b c9 c3
0: 0f 1f 44 00 00 nop DWORD PTR [rax+rax*1+0x0]
5: 55 push rbp
6: 48 89 e5 mov rbp,rsp
9: 53 push rbx
a: 41 55 push r13
c: 31 c0 xor eax,eax
e: 45 31 ed xor r13d,r13d
11: 48 89 fb mov rbx,rdi
14: eb 30 jmp 0x46
16: eb 2e jmp 0x46
...
3e: eb 06 jmp 0x46
40: eb 04 jmp 0x46
42: eb 02 jmp 0x46
44: 66 90 xchg ax,ax
46: 31 c0 xor eax,eax
48: 41 5d pop r13
4a: 5b pop rbx
4b: c9 leave
4c: c3 ret
At the 16th pass, 15 jumps were already optimized out, and one jump was
replaced with NOPs at 44 and the image converged at the 17th pass.
v4:
- Add the detailed comments about the possible padding bytes
v3:
- Copy the instructions of prologue separately or the size calculation
of the first BPF instruction would include the prologue.
- Replace WARN_ONCE() with pr_err() and EFAULT
- Use MAX_PASSES in the for loop condition check
- Remove the "padded" flag from x64_jit_data. For the extra pass of
subprogs, padding is always enabled since it won't hurt the images
that converge without padding.
v2:
- Simplify the sample code in the description and provide the jit code
- Check the expected padding bytes with WARN_ONCE
- Move the 'padded' flag to 'struct x64_jit_data'
Signed-off-by: Gary Lin <glin@suse.com>
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Link: https://lore.kernel.org/bpf/20210119102501.511-2-glin@suse.com
2021-01-19 10:24:59 +00:00
|
|
|
}
|
net: filter: x86: internal BPF JIT
Maps all internal BPF instructions into x86_64 instructions.
This patch replaces original BPF x64 JIT with internal BPF x64 JIT.
sysctl net.core.bpf_jit_enable is reused as on/off switch.
Performance:
1. old BPF JIT and internal BPF JIT generate equivalent x86_64 code.
No performance difference is observed for filters that were JIT-able before
Example assembler code for BPF filter "tcpdump port 22"
original BPF -> old JIT: original BPF -> internal BPF -> new JIT:
0: push %rbp 0: push %rbp
1: mov %rsp,%rbp 1: mov %rsp,%rbp
4: sub $0x60,%rsp 4: sub $0x228,%rsp
8: mov %rbx,-0x8(%rbp) b: mov %rbx,-0x228(%rbp) // prologue
12: mov %r13,-0x220(%rbp)
19: mov %r14,-0x218(%rbp)
20: mov %r15,-0x210(%rbp)
27: xor %eax,%eax // clear A
c: xor %ebx,%ebx 29: xor %r13,%r13 // clear X
e: mov 0x68(%rdi),%r9d 2c: mov 0x68(%rdi),%r9d
12: sub 0x6c(%rdi),%r9d 30: sub 0x6c(%rdi),%r9d
16: mov 0xd8(%rdi),%r8 34: mov 0xd8(%rdi),%r10
3b: mov %rdi,%rbx
1d: mov $0xc,%esi 3e: mov $0xc,%esi
22: callq 0xffffffffe1021e15 43: callq 0xffffffffe102bd75
27: cmp $0x86dd,%eax 48: cmp $0x86dd,%rax
2c: jne 0x0000000000000069 4f: jne 0x000000000000009a
2e: mov $0x14,%esi 51: mov $0x14,%esi
33: callq 0xffffffffe1021e31 56: callq 0xffffffffe102bd91
38: cmp $0x84,%eax 5b: cmp $0x84,%rax
3d: je 0x0000000000000049 62: je 0x0000000000000074
3f: cmp $0x6,%eax 64: cmp $0x6,%rax
42: je 0x0000000000000049 68: je 0x0000000000000074
44: cmp $0x11,%eax 6a: cmp $0x11,%rax
47: jne 0x00000000000000c6 6e: jne 0x0000000000000117
49: mov $0x36,%esi 74: mov $0x36,%esi
4e: callq 0xffffffffe1021e15 79: callq 0xffffffffe102bd75
53: cmp $0x16,%eax 7e: cmp $0x16,%rax
56: je 0x00000000000000bf 82: je 0x0000000000000110
58: mov $0x38,%esi 88: mov $0x38,%esi
5d: callq 0xffffffffe1021e15 8d: callq 0xffffffffe102bd75
62: cmp $0x16,%eax 92: cmp $0x16,%rax
65: je 0x00000000000000bf 96: je 0x0000000000000110
67: jmp 0x00000000000000c6 98: jmp 0x0000000000000117
69: cmp $0x800,%eax 9a: cmp $0x800,%rax
6e: jne 0x00000000000000c6 a1: jne 0x0000000000000117
70: mov $0x17,%esi a3: mov $0x17,%esi
75: callq 0xffffffffe1021e31 a8: callq 0xffffffffe102bd91
7a: cmp $0x84,%eax ad: cmp $0x84,%rax
7f: je 0x000000000000008b b4: je 0x00000000000000c2
81: cmp $0x6,%eax b6: cmp $0x6,%rax
84: je 0x000000000000008b ba: je 0x00000000000000c2
86: cmp $0x11,%eax bc: cmp $0x11,%rax
89: jne 0x00000000000000c6 c0: jne 0x0000000000000117
8b: mov $0x14,%esi c2: mov $0x14,%esi
90: callq 0xffffffffe1021e15 c7: callq 0xffffffffe102bd75
95: test $0x1fff,%ax cc: test $0x1fff,%rax
99: jne 0x00000000000000c6 d3: jne 0x0000000000000117
d5: mov %rax,%r14
9b: mov $0xe,%esi d8: mov $0xe,%esi
a0: callq 0xffffffffe1021e44 dd: callq 0xffffffffe102bd91 // MSH
e2: and $0xf,%eax
e5: shl $0x2,%eax
e8: mov %rax,%r13
eb: mov %r14,%rax
ee: mov %r13,%rsi
a5: lea 0xe(%rbx),%esi f1: add $0xe,%esi
a8: callq 0xffffffffe1021e0d f4: callq 0xffffffffe102bd6d
ad: cmp $0x16,%eax f9: cmp $0x16,%rax
b0: je 0x00000000000000bf fd: je 0x0000000000000110
ff: mov %r13,%rsi
b2: lea 0x10(%rbx),%esi 102: add $0x10,%esi
b5: callq 0xffffffffe1021e0d 105: callq 0xffffffffe102bd6d
ba: cmp $0x16,%eax 10a: cmp $0x16,%rax
bd: jne 0x00000000000000c6 10e: jne 0x0000000000000117
bf: mov $0xffff,%eax 110: mov $0xffff,%eax
c4: jmp 0x00000000000000c8 115: jmp 0x000000000000011c
c6: xor %eax,%eax 117: mov $0x0,%eax
c8: mov -0x8(%rbp),%rbx 11c: mov -0x228(%rbp),%rbx // epilogue
cc: leaveq 123: mov -0x220(%rbp),%r13
cd: retq 12a: mov -0x218(%rbp),%r14
131: mov -0x210(%rbp),%r15
138: leaveq
139: retq
On fully cached SKBs both JITed functions take 12 nsec to execute.
BPF interpreter executes the program in 30 nsec.
The difference in generated assembler is due to the following:
Old BPF imlements LDX_MSH instruction via sk_load_byte_msh() helper function
inside bpf_jit.S.
New JIT removes the helper and does it explicitly, so ldx_msh cost
is the same for both JITs, but generated code looks longer.
New JIT has 4 registers to save, so prologue/epilogue are larger,
but the cost is within noise on x64.
Old JIT checks whether first insn clears A and if not emits 'xor %eax,%eax'.
New JIT clears %rax unconditionally.
2. old BPF JIT doesn't support ANC_NLATTR, ANC_PAY_OFFSET, ANC_RANDOM
extensions. New JIT supports all BPF extensions.
Performance of such filters improves 2-4 times depending on a filter.
The longer the filter the higher performance gain.
Synthetic benchmarks with many ancillary loads see 20x speedup
which seems to be the maximum gain from JIT
Notes:
. net.core.bpf_jit_enable=2 + tools/net/bpf_jit_disasm is still functional
and can be used to see generated assembler
. there are two jit_compile() functions and code flow for classic filters is:
sk_attach_filter() - load classic BPF
bpf_jit_compile() - try to JIT from classic BPF
sk_convert_filter() - convert classic to internal
bpf_int_jit_compile() - JIT from internal BPF
seccomp and tracing filters will just call bpf_int_jit_compile()
Signed-off-by: Alexei Starovoitov <ast@plumgrid.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
2014-05-14 02:50:46 +00:00
|
|
|
EMIT2(0xEB, jmp_offset);
|
|
|
|
} else if (is_simm32(jmp_offset)) {
|
|
|
|
EMIT1_off32(0xE9, jmp_offset);
|
|
|
|
} else {
|
|
|
|
pr_err("jmp gen bug %llx\n", jmp_offset);
|
|
|
|
return -EFAULT;
|
|
|
|
}
|
|
|
|
break;
|
|
|
|
|
|
|
|
case BPF_JMP | BPF_EXIT:
|
2014-11-29 22:46:13 +00:00
|
|
|
if (seen_exit) {
|
net: filter: x86: internal BPF JIT
Maps all internal BPF instructions into x86_64 instructions.
This patch replaces original BPF x64 JIT with internal BPF x64 JIT.
sysctl net.core.bpf_jit_enable is reused as on/off switch.
Performance:
1. old BPF JIT and internal BPF JIT generate equivalent x86_64 code.
No performance difference is observed for filters that were JIT-able before
Example assembler code for BPF filter "tcpdump port 22"
original BPF -> old JIT: original BPF -> internal BPF -> new JIT:
0: push %rbp 0: push %rbp
1: mov %rsp,%rbp 1: mov %rsp,%rbp
4: sub $0x60,%rsp 4: sub $0x228,%rsp
8: mov %rbx,-0x8(%rbp) b: mov %rbx,-0x228(%rbp) // prologue
12: mov %r13,-0x220(%rbp)
19: mov %r14,-0x218(%rbp)
20: mov %r15,-0x210(%rbp)
27: xor %eax,%eax // clear A
c: xor %ebx,%ebx 29: xor %r13,%r13 // clear X
e: mov 0x68(%rdi),%r9d 2c: mov 0x68(%rdi),%r9d
12: sub 0x6c(%rdi),%r9d 30: sub 0x6c(%rdi),%r9d
16: mov 0xd8(%rdi),%r8 34: mov 0xd8(%rdi),%r10
3b: mov %rdi,%rbx
1d: mov $0xc,%esi 3e: mov $0xc,%esi
22: callq 0xffffffffe1021e15 43: callq 0xffffffffe102bd75
27: cmp $0x86dd,%eax 48: cmp $0x86dd,%rax
2c: jne 0x0000000000000069 4f: jne 0x000000000000009a
2e: mov $0x14,%esi 51: mov $0x14,%esi
33: callq 0xffffffffe1021e31 56: callq 0xffffffffe102bd91
38: cmp $0x84,%eax 5b: cmp $0x84,%rax
3d: je 0x0000000000000049 62: je 0x0000000000000074
3f: cmp $0x6,%eax 64: cmp $0x6,%rax
42: je 0x0000000000000049 68: je 0x0000000000000074
44: cmp $0x11,%eax 6a: cmp $0x11,%rax
47: jne 0x00000000000000c6 6e: jne 0x0000000000000117
49: mov $0x36,%esi 74: mov $0x36,%esi
4e: callq 0xffffffffe1021e15 79: callq 0xffffffffe102bd75
53: cmp $0x16,%eax 7e: cmp $0x16,%rax
56: je 0x00000000000000bf 82: je 0x0000000000000110
58: mov $0x38,%esi 88: mov $0x38,%esi
5d: callq 0xffffffffe1021e15 8d: callq 0xffffffffe102bd75
62: cmp $0x16,%eax 92: cmp $0x16,%rax
65: je 0x00000000000000bf 96: je 0x0000000000000110
67: jmp 0x00000000000000c6 98: jmp 0x0000000000000117
69: cmp $0x800,%eax 9a: cmp $0x800,%rax
6e: jne 0x00000000000000c6 a1: jne 0x0000000000000117
70: mov $0x17,%esi a3: mov $0x17,%esi
75: callq 0xffffffffe1021e31 a8: callq 0xffffffffe102bd91
7a: cmp $0x84,%eax ad: cmp $0x84,%rax
7f: je 0x000000000000008b b4: je 0x00000000000000c2
81: cmp $0x6,%eax b6: cmp $0x6,%rax
84: je 0x000000000000008b ba: je 0x00000000000000c2
86: cmp $0x11,%eax bc: cmp $0x11,%rax
89: jne 0x00000000000000c6 c0: jne 0x0000000000000117
8b: mov $0x14,%esi c2: mov $0x14,%esi
90: callq 0xffffffffe1021e15 c7: callq 0xffffffffe102bd75
95: test $0x1fff,%ax cc: test $0x1fff,%rax
99: jne 0x00000000000000c6 d3: jne 0x0000000000000117
d5: mov %rax,%r14
9b: mov $0xe,%esi d8: mov $0xe,%esi
a0: callq 0xffffffffe1021e44 dd: callq 0xffffffffe102bd91 // MSH
e2: and $0xf,%eax
e5: shl $0x2,%eax
e8: mov %rax,%r13
eb: mov %r14,%rax
ee: mov %r13,%rsi
a5: lea 0xe(%rbx),%esi f1: add $0xe,%esi
a8: callq 0xffffffffe1021e0d f4: callq 0xffffffffe102bd6d
ad: cmp $0x16,%eax f9: cmp $0x16,%rax
b0: je 0x00000000000000bf fd: je 0x0000000000000110
ff: mov %r13,%rsi
b2: lea 0x10(%rbx),%esi 102: add $0x10,%esi
b5: callq 0xffffffffe1021e0d 105: callq 0xffffffffe102bd6d
ba: cmp $0x16,%eax 10a: cmp $0x16,%rax
bd: jne 0x00000000000000c6 10e: jne 0x0000000000000117
bf: mov $0xffff,%eax 110: mov $0xffff,%eax
c4: jmp 0x00000000000000c8 115: jmp 0x000000000000011c
c6: xor %eax,%eax 117: mov $0x0,%eax
c8: mov -0x8(%rbp),%rbx 11c: mov -0x228(%rbp),%rbx // epilogue
cc: leaveq 123: mov -0x220(%rbp),%r13
cd: retq 12a: mov -0x218(%rbp),%r14
131: mov -0x210(%rbp),%r15
138: leaveq
139: retq
On fully cached SKBs both JITed functions take 12 nsec to execute.
BPF interpreter executes the program in 30 nsec.
The difference in generated assembler is due to the following:
Old BPF imlements LDX_MSH instruction via sk_load_byte_msh() helper function
inside bpf_jit.S.
New JIT removes the helper and does it explicitly, so ldx_msh cost
is the same for both JITs, but generated code looks longer.
New JIT has 4 registers to save, so prologue/epilogue are larger,
but the cost is within noise on x64.
Old JIT checks whether first insn clears A and if not emits 'xor %eax,%eax'.
New JIT clears %rax unconditionally.
2. old BPF JIT doesn't support ANC_NLATTR, ANC_PAY_OFFSET, ANC_RANDOM
extensions. New JIT supports all BPF extensions.
Performance of such filters improves 2-4 times depending on a filter.
The longer the filter the higher performance gain.
Synthetic benchmarks with many ancillary loads see 20x speedup
which seems to be the maximum gain from JIT
Notes:
. net.core.bpf_jit_enable=2 + tools/net/bpf_jit_disasm is still functional
and can be used to see generated assembler
. there are two jit_compile() functions and code flow for classic filters is:
sk_attach_filter() - load classic BPF
bpf_jit_compile() - try to JIT from classic BPF
sk_convert_filter() - convert classic to internal
bpf_int_jit_compile() - JIT from internal BPF
seccomp and tracing filters will just call bpf_int_jit_compile()
Signed-off-by: Alexei Starovoitov <ast@plumgrid.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
2014-05-14 02:50:46 +00:00
|
|
|
jmp_offset = ctx->cleanup_addr - addrs[i];
|
|
|
|
goto emit_jmp;
|
|
|
|
}
|
2014-11-29 22:46:13 +00:00
|
|
|
seen_exit = true;
|
x86/bpf: Clean up non-standard comments, to make the code more readable
So by chance I looked into x86 assembly in arch/x86/net/bpf_jit_comp.c and
noticed the weird and inconsistent comment style it mistakenly learned from
the networking code:
/* Multi-line comment ...
* ... looks like this.
*/
Fix this to use the standard comment style specified in Documentation/CodingStyle
and used in arch/x86/ as well:
/*
* Multi-line comment ...
* ... looks like this.
*/
Also, to quote Linus's ... more explicit views about this:
http://article.gmane.org/gmane.linux.kernel.cryptoapi/21066
> But no, the networking code picked *none* of the above sane formats.
> Instead, it picked these two models that are just half-arsed
> shit-for-brains:
>
> (no)
> /* This is disgusting drug-induced
> * crap, and should die
> */
>
> (no-no-no)
> /* This is also very nasty
> * and visually unbalanced */
>
> Please. The networking code actually has the *worst* possible comment
> style. You can literally find that (no-no-no) style, which is just
> really horribly disgusting and worse than the otherwise fairly similar
> (d) in pretty much every way.
Also improve the comments and some other details while at it:
- Don't mix same-line and previous-line comment style on otherwise
identical code patterns within the same function,
- capitalize 'BPF' and x86 register names consistently,
- capitalize sentences consistently,
- instead of 'x64' use 'x86-64': x64 is a Microsoft specific term,
- use more consistent punctuation,
- use standard coding style in macros as well,
- fix typos and a few other minor details.
Consistent coding style is not optional, at least in arch/x86/.
No change in functionality.
( In case this commit causes conflicts with pending development code
I'll be glad to help resolve any conflicts! )
Acked-by: Thomas Gleixner <tglx@linutronix.de>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Borislav Petkov <bp@alien8.de>
Cc: H. Peter Anvin <hpa@zytor.com>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: David S. Miller <davem@davemloft.net>
Cc: Eric Dumazet <edumazet@google.com>
Cc: Daniel Borkmann <daniel@iogearbox.net>
Cc: Alexei Starovoitov <ast@fb.com>
Cc: Hideaki YOSHIFUJI <yoshfuji@linux-ipv6.org>
Cc: netdev@vger.kernel.org
Cc: linux-kernel@vger.kernel.org
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
2018-04-27 09:54:40 +00:00
|
|
|
/* Update cleanup_addr */
|
net: filter: x86: internal BPF JIT
Maps all internal BPF instructions into x86_64 instructions.
This patch replaces original BPF x64 JIT with internal BPF x64 JIT.
sysctl net.core.bpf_jit_enable is reused as on/off switch.
Performance:
1. old BPF JIT and internal BPF JIT generate equivalent x86_64 code.
No performance difference is observed for filters that were JIT-able before
Example assembler code for BPF filter "tcpdump port 22"
original BPF -> old JIT: original BPF -> internal BPF -> new JIT:
0: push %rbp 0: push %rbp
1: mov %rsp,%rbp 1: mov %rsp,%rbp
4: sub $0x60,%rsp 4: sub $0x228,%rsp
8: mov %rbx,-0x8(%rbp) b: mov %rbx,-0x228(%rbp) // prologue
12: mov %r13,-0x220(%rbp)
19: mov %r14,-0x218(%rbp)
20: mov %r15,-0x210(%rbp)
27: xor %eax,%eax // clear A
c: xor %ebx,%ebx 29: xor %r13,%r13 // clear X
e: mov 0x68(%rdi),%r9d 2c: mov 0x68(%rdi),%r9d
12: sub 0x6c(%rdi),%r9d 30: sub 0x6c(%rdi),%r9d
16: mov 0xd8(%rdi),%r8 34: mov 0xd8(%rdi),%r10
3b: mov %rdi,%rbx
1d: mov $0xc,%esi 3e: mov $0xc,%esi
22: callq 0xffffffffe1021e15 43: callq 0xffffffffe102bd75
27: cmp $0x86dd,%eax 48: cmp $0x86dd,%rax
2c: jne 0x0000000000000069 4f: jne 0x000000000000009a
2e: mov $0x14,%esi 51: mov $0x14,%esi
33: callq 0xffffffffe1021e31 56: callq 0xffffffffe102bd91
38: cmp $0x84,%eax 5b: cmp $0x84,%rax
3d: je 0x0000000000000049 62: je 0x0000000000000074
3f: cmp $0x6,%eax 64: cmp $0x6,%rax
42: je 0x0000000000000049 68: je 0x0000000000000074
44: cmp $0x11,%eax 6a: cmp $0x11,%rax
47: jne 0x00000000000000c6 6e: jne 0x0000000000000117
49: mov $0x36,%esi 74: mov $0x36,%esi
4e: callq 0xffffffffe1021e15 79: callq 0xffffffffe102bd75
53: cmp $0x16,%eax 7e: cmp $0x16,%rax
56: je 0x00000000000000bf 82: je 0x0000000000000110
58: mov $0x38,%esi 88: mov $0x38,%esi
5d: callq 0xffffffffe1021e15 8d: callq 0xffffffffe102bd75
62: cmp $0x16,%eax 92: cmp $0x16,%rax
65: je 0x00000000000000bf 96: je 0x0000000000000110
67: jmp 0x00000000000000c6 98: jmp 0x0000000000000117
69: cmp $0x800,%eax 9a: cmp $0x800,%rax
6e: jne 0x00000000000000c6 a1: jne 0x0000000000000117
70: mov $0x17,%esi a3: mov $0x17,%esi
75: callq 0xffffffffe1021e31 a8: callq 0xffffffffe102bd91
7a: cmp $0x84,%eax ad: cmp $0x84,%rax
7f: je 0x000000000000008b b4: je 0x00000000000000c2
81: cmp $0x6,%eax b6: cmp $0x6,%rax
84: je 0x000000000000008b ba: je 0x00000000000000c2
86: cmp $0x11,%eax bc: cmp $0x11,%rax
89: jne 0x00000000000000c6 c0: jne 0x0000000000000117
8b: mov $0x14,%esi c2: mov $0x14,%esi
90: callq 0xffffffffe1021e15 c7: callq 0xffffffffe102bd75
95: test $0x1fff,%ax cc: test $0x1fff,%rax
99: jne 0x00000000000000c6 d3: jne 0x0000000000000117
d5: mov %rax,%r14
9b: mov $0xe,%esi d8: mov $0xe,%esi
a0: callq 0xffffffffe1021e44 dd: callq 0xffffffffe102bd91 // MSH
e2: and $0xf,%eax
e5: shl $0x2,%eax
e8: mov %rax,%r13
eb: mov %r14,%rax
ee: mov %r13,%rsi
a5: lea 0xe(%rbx),%esi f1: add $0xe,%esi
a8: callq 0xffffffffe1021e0d f4: callq 0xffffffffe102bd6d
ad: cmp $0x16,%eax f9: cmp $0x16,%rax
b0: je 0x00000000000000bf fd: je 0x0000000000000110
ff: mov %r13,%rsi
b2: lea 0x10(%rbx),%esi 102: add $0x10,%esi
b5: callq 0xffffffffe1021e0d 105: callq 0xffffffffe102bd6d
ba: cmp $0x16,%eax 10a: cmp $0x16,%rax
bd: jne 0x00000000000000c6 10e: jne 0x0000000000000117
bf: mov $0xffff,%eax 110: mov $0xffff,%eax
c4: jmp 0x00000000000000c8 115: jmp 0x000000000000011c
c6: xor %eax,%eax 117: mov $0x0,%eax
c8: mov -0x8(%rbp),%rbx 11c: mov -0x228(%rbp),%rbx // epilogue
cc: leaveq 123: mov -0x220(%rbp),%r13
cd: retq 12a: mov -0x218(%rbp),%r14
131: mov -0x210(%rbp),%r15
138: leaveq
139: retq
On fully cached SKBs both JITed functions take 12 nsec to execute.
BPF interpreter executes the program in 30 nsec.
The difference in generated assembler is due to the following:
Old BPF imlements LDX_MSH instruction via sk_load_byte_msh() helper function
inside bpf_jit.S.
New JIT removes the helper and does it explicitly, so ldx_msh cost
is the same for both JITs, but generated code looks longer.
New JIT has 4 registers to save, so prologue/epilogue are larger,
but the cost is within noise on x64.
Old JIT checks whether first insn clears A and if not emits 'xor %eax,%eax'.
New JIT clears %rax unconditionally.
2. old BPF JIT doesn't support ANC_NLATTR, ANC_PAY_OFFSET, ANC_RANDOM
extensions. New JIT supports all BPF extensions.
Performance of such filters improves 2-4 times depending on a filter.
The longer the filter the higher performance gain.
Synthetic benchmarks with many ancillary loads see 20x speedup
which seems to be the maximum gain from JIT
Notes:
. net.core.bpf_jit_enable=2 + tools/net/bpf_jit_disasm is still functional
and can be used to see generated assembler
. there are two jit_compile() functions and code flow for classic filters is:
sk_attach_filter() - load classic BPF
bpf_jit_compile() - try to JIT from classic BPF
sk_convert_filter() - convert classic to internal
bpf_int_jit_compile() - JIT from internal BPF
seccomp and tracing filters will just call bpf_int_jit_compile()
Signed-off-by: Alexei Starovoitov <ast@plumgrid.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
2014-05-14 02:50:46 +00:00
|
|
|
ctx->cleanup_addr = proglen;
|
bpf: Implement BPF exceptions
This patch implements BPF exceptions, and introduces a bpf_throw kfunc
to allow programs to throw exceptions during their execution at runtime.
A bpf_throw invocation is treated as an immediate termination of the
program, returning back to its caller within the kernel, unwinding all
stack frames.
This allows the program to simplify its implementation, by testing for
runtime conditions which the verifier has no visibility into, and assert
that they are true. In case they are not, the program can simply throw
an exception from the other branch.
BPF exceptions are explicitly *NOT* an unlikely slowpath error handling
primitive, and this objective has guided design choices of the
implementation of the them within the kernel (with the bulk of the cost
for unwinding the stack offloaded to the bpf_throw kfunc).
The implementation of this mechanism requires use of add_hidden_subprog
mechanism introduced in the previous patch, which generates a couple of
instructions to move R1 to R0 and exit. The JIT then rewrites the
prologue of this subprog to take the stack pointer and frame pointer as
inputs and reset the stack frame, popping all callee-saved registers
saved by the main subprog. The bpf_throw function then walks the stack
at runtime, and invokes this exception subprog with the stack and frame
pointers as parameters.
Reviewers must take note that currently the main program is made to save
all callee-saved registers on x86_64 during entry into the program. This
is because we must do an equivalent of a lightweight context switch when
unwinding the stack, therefore we need the callee-saved registers of the
caller of the BPF program to be able to return with a sane state.
Note that we have to additionally handle r12, even though it is not used
by the program, because when throwing the exception the program makes an
entry into the kernel which could clobber r12 after saving it on the
stack. To be able to preserve the value we received on program entry, we
push r12 and restore it from the generated subprogram when unwinding the
stack.
For now, bpf_throw invocation fails when lingering resources or locks
exist in that path of the program. In a future followup, bpf_throw will
be extended to perform frame-by-frame unwinding to release lingering
resources for each stack frame, removing this limitation.
Signed-off-by: Kumar Kartikeya Dwivedi <memxor@gmail.com>
Link: https://lore.kernel.org/r/20230912233214.1518551-5-memxor@gmail.com
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
2023-09-12 23:32:01 +00:00
|
|
|
if (bpf_prog->aux->exception_boundary) {
|
|
|
|
pop_callee_regs(&prog, all_callee_regs_used);
|
|
|
|
pop_r12(&prog);
|
|
|
|
} else {
|
|
|
|
pop_callee_regs(&prog, callee_regs_used);
|
2024-03-08 01:08:01 +00:00
|
|
|
if (arena_vm_start)
|
|
|
|
pop_r12(&prog);
|
bpf: Implement BPF exceptions
This patch implements BPF exceptions, and introduces a bpf_throw kfunc
to allow programs to throw exceptions during their execution at runtime.
A bpf_throw invocation is treated as an immediate termination of the
program, returning back to its caller within the kernel, unwinding all
stack frames.
This allows the program to simplify its implementation, by testing for
runtime conditions which the verifier has no visibility into, and assert
that they are true. In case they are not, the program can simply throw
an exception from the other branch.
BPF exceptions are explicitly *NOT* an unlikely slowpath error handling
primitive, and this objective has guided design choices of the
implementation of the them within the kernel (with the bulk of the cost
for unwinding the stack offloaded to the bpf_throw kfunc).
The implementation of this mechanism requires use of add_hidden_subprog
mechanism introduced in the previous patch, which generates a couple of
instructions to move R1 to R0 and exit. The JIT then rewrites the
prologue of this subprog to take the stack pointer and frame pointer as
inputs and reset the stack frame, popping all callee-saved registers
saved by the main subprog. The bpf_throw function then walks the stack
at runtime, and invokes this exception subprog with the stack and frame
pointers as parameters.
Reviewers must take note that currently the main program is made to save
all callee-saved registers on x86_64 during entry into the program. This
is because we must do an equivalent of a lightweight context switch when
unwinding the stack, therefore we need the callee-saved registers of the
caller of the BPF program to be able to return with a sane state.
Note that we have to additionally handle r12, even though it is not used
by the program, because when throwing the exception the program makes an
entry into the kernel which could clobber r12 after saving it on the
stack. To be able to preserve the value we received on program entry, we
push r12 and restore it from the generated subprogram when unwinding the
stack.
For now, bpf_throw invocation fails when lingering resources or locks
exist in that path of the program. In a future followup, bpf_throw will
be extended to perform frame-by-frame unwinding to release lingering
resources for each stack frame, removing this limitation.
Signed-off-by: Kumar Kartikeya Dwivedi <memxor@gmail.com>
Link: https://lore.kernel.org/r/20230912233214.1518551-5-memxor@gmail.com
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
2023-09-12 23:32:01 +00:00
|
|
|
}
|
2019-06-14 22:43:28 +00:00
|
|
|
EMIT1(0xC9); /* leave */
|
2022-06-14 21:15:41 +00:00
|
|
|
emit_return(&prog, image + addrs[i - 1] + (prog - temp));
|
net: filter: x86: internal BPF JIT
Maps all internal BPF instructions into x86_64 instructions.
This patch replaces original BPF x64 JIT with internal BPF x64 JIT.
sysctl net.core.bpf_jit_enable is reused as on/off switch.
Performance:
1. old BPF JIT and internal BPF JIT generate equivalent x86_64 code.
No performance difference is observed for filters that were JIT-able before
Example assembler code for BPF filter "tcpdump port 22"
original BPF -> old JIT: original BPF -> internal BPF -> new JIT:
0: push %rbp 0: push %rbp
1: mov %rsp,%rbp 1: mov %rsp,%rbp
4: sub $0x60,%rsp 4: sub $0x228,%rsp
8: mov %rbx,-0x8(%rbp) b: mov %rbx,-0x228(%rbp) // prologue
12: mov %r13,-0x220(%rbp)
19: mov %r14,-0x218(%rbp)
20: mov %r15,-0x210(%rbp)
27: xor %eax,%eax // clear A
c: xor %ebx,%ebx 29: xor %r13,%r13 // clear X
e: mov 0x68(%rdi),%r9d 2c: mov 0x68(%rdi),%r9d
12: sub 0x6c(%rdi),%r9d 30: sub 0x6c(%rdi),%r9d
16: mov 0xd8(%rdi),%r8 34: mov 0xd8(%rdi),%r10
3b: mov %rdi,%rbx
1d: mov $0xc,%esi 3e: mov $0xc,%esi
22: callq 0xffffffffe1021e15 43: callq 0xffffffffe102bd75
27: cmp $0x86dd,%eax 48: cmp $0x86dd,%rax
2c: jne 0x0000000000000069 4f: jne 0x000000000000009a
2e: mov $0x14,%esi 51: mov $0x14,%esi
33: callq 0xffffffffe1021e31 56: callq 0xffffffffe102bd91
38: cmp $0x84,%eax 5b: cmp $0x84,%rax
3d: je 0x0000000000000049 62: je 0x0000000000000074
3f: cmp $0x6,%eax 64: cmp $0x6,%rax
42: je 0x0000000000000049 68: je 0x0000000000000074
44: cmp $0x11,%eax 6a: cmp $0x11,%rax
47: jne 0x00000000000000c6 6e: jne 0x0000000000000117
49: mov $0x36,%esi 74: mov $0x36,%esi
4e: callq 0xffffffffe1021e15 79: callq 0xffffffffe102bd75
53: cmp $0x16,%eax 7e: cmp $0x16,%rax
56: je 0x00000000000000bf 82: je 0x0000000000000110
58: mov $0x38,%esi 88: mov $0x38,%esi
5d: callq 0xffffffffe1021e15 8d: callq 0xffffffffe102bd75
62: cmp $0x16,%eax 92: cmp $0x16,%rax
65: je 0x00000000000000bf 96: je 0x0000000000000110
67: jmp 0x00000000000000c6 98: jmp 0x0000000000000117
69: cmp $0x800,%eax 9a: cmp $0x800,%rax
6e: jne 0x00000000000000c6 a1: jne 0x0000000000000117
70: mov $0x17,%esi a3: mov $0x17,%esi
75: callq 0xffffffffe1021e31 a8: callq 0xffffffffe102bd91
7a: cmp $0x84,%eax ad: cmp $0x84,%rax
7f: je 0x000000000000008b b4: je 0x00000000000000c2
81: cmp $0x6,%eax b6: cmp $0x6,%rax
84: je 0x000000000000008b ba: je 0x00000000000000c2
86: cmp $0x11,%eax bc: cmp $0x11,%rax
89: jne 0x00000000000000c6 c0: jne 0x0000000000000117
8b: mov $0x14,%esi c2: mov $0x14,%esi
90: callq 0xffffffffe1021e15 c7: callq 0xffffffffe102bd75
95: test $0x1fff,%ax cc: test $0x1fff,%rax
99: jne 0x00000000000000c6 d3: jne 0x0000000000000117
d5: mov %rax,%r14
9b: mov $0xe,%esi d8: mov $0xe,%esi
a0: callq 0xffffffffe1021e44 dd: callq 0xffffffffe102bd91 // MSH
e2: and $0xf,%eax
e5: shl $0x2,%eax
e8: mov %rax,%r13
eb: mov %r14,%rax
ee: mov %r13,%rsi
a5: lea 0xe(%rbx),%esi f1: add $0xe,%esi
a8: callq 0xffffffffe1021e0d f4: callq 0xffffffffe102bd6d
ad: cmp $0x16,%eax f9: cmp $0x16,%rax
b0: je 0x00000000000000bf fd: je 0x0000000000000110
ff: mov %r13,%rsi
b2: lea 0x10(%rbx),%esi 102: add $0x10,%esi
b5: callq 0xffffffffe1021e0d 105: callq 0xffffffffe102bd6d
ba: cmp $0x16,%eax 10a: cmp $0x16,%rax
bd: jne 0x00000000000000c6 10e: jne 0x0000000000000117
bf: mov $0xffff,%eax 110: mov $0xffff,%eax
c4: jmp 0x00000000000000c8 115: jmp 0x000000000000011c
c6: xor %eax,%eax 117: mov $0x0,%eax
c8: mov -0x8(%rbp),%rbx 11c: mov -0x228(%rbp),%rbx // epilogue
cc: leaveq 123: mov -0x220(%rbp),%r13
cd: retq 12a: mov -0x218(%rbp),%r14
131: mov -0x210(%rbp),%r15
138: leaveq
139: retq
On fully cached SKBs both JITed functions take 12 nsec to execute.
BPF interpreter executes the program in 30 nsec.
The difference in generated assembler is due to the following:
Old BPF imlements LDX_MSH instruction via sk_load_byte_msh() helper function
inside bpf_jit.S.
New JIT removes the helper and does it explicitly, so ldx_msh cost
is the same for both JITs, but generated code looks longer.
New JIT has 4 registers to save, so prologue/epilogue are larger,
but the cost is within noise on x64.
Old JIT checks whether first insn clears A and if not emits 'xor %eax,%eax'.
New JIT clears %rax unconditionally.
2. old BPF JIT doesn't support ANC_NLATTR, ANC_PAY_OFFSET, ANC_RANDOM
extensions. New JIT supports all BPF extensions.
Performance of such filters improves 2-4 times depending on a filter.
The longer the filter the higher performance gain.
Synthetic benchmarks with many ancillary loads see 20x speedup
which seems to be the maximum gain from JIT
Notes:
. net.core.bpf_jit_enable=2 + tools/net/bpf_jit_disasm is still functional
and can be used to see generated assembler
. there are two jit_compile() functions and code flow for classic filters is:
sk_attach_filter() - load classic BPF
bpf_jit_compile() - try to JIT from classic BPF
sk_convert_filter() - convert classic to internal
bpf_int_jit_compile() - JIT from internal BPF
seccomp and tracing filters will just call bpf_int_jit_compile()
Signed-off-by: Alexei Starovoitov <ast@plumgrid.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
2014-05-14 02:50:46 +00:00
|
|
|
break;
|
|
|
|
|
2014-05-14 02:50:45 +00:00
|
|
|
default:
|
x86/bpf: Clean up non-standard comments, to make the code more readable
So by chance I looked into x86 assembly in arch/x86/net/bpf_jit_comp.c and
noticed the weird and inconsistent comment style it mistakenly learned from
the networking code:
/* Multi-line comment ...
* ... looks like this.
*/
Fix this to use the standard comment style specified in Documentation/CodingStyle
and used in arch/x86/ as well:
/*
* Multi-line comment ...
* ... looks like this.
*/
Also, to quote Linus's ... more explicit views about this:
http://article.gmane.org/gmane.linux.kernel.cryptoapi/21066
> But no, the networking code picked *none* of the above sane formats.
> Instead, it picked these two models that are just half-arsed
> shit-for-brains:
>
> (no)
> /* This is disgusting drug-induced
> * crap, and should die
> */
>
> (no-no-no)
> /* This is also very nasty
> * and visually unbalanced */
>
> Please. The networking code actually has the *worst* possible comment
> style. You can literally find that (no-no-no) style, which is just
> really horribly disgusting and worse than the otherwise fairly similar
> (d) in pretty much every way.
Also improve the comments and some other details while at it:
- Don't mix same-line and previous-line comment style on otherwise
identical code patterns within the same function,
- capitalize 'BPF' and x86 register names consistently,
- capitalize sentences consistently,
- instead of 'x64' use 'x86-64': x64 is a Microsoft specific term,
- use more consistent punctuation,
- use standard coding style in macros as well,
- fix typos and a few other minor details.
Consistent coding style is not optional, at least in arch/x86/.
No change in functionality.
( In case this commit causes conflicts with pending development code
I'll be glad to help resolve any conflicts! )
Acked-by: Thomas Gleixner <tglx@linutronix.de>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Borislav Petkov <bp@alien8.de>
Cc: H. Peter Anvin <hpa@zytor.com>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: David S. Miller <davem@davemloft.net>
Cc: Eric Dumazet <edumazet@google.com>
Cc: Daniel Borkmann <daniel@iogearbox.net>
Cc: Alexei Starovoitov <ast@fb.com>
Cc: Hideaki YOSHIFUJI <yoshfuji@linux-ipv6.org>
Cc: netdev@vger.kernel.org
Cc: linux-kernel@vger.kernel.org
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
2018-04-27 09:54:40 +00:00
|
|
|
/*
|
|
|
|
* By design x86-64 JIT should support all BPF instructions.
|
net: filter: x86: internal BPF JIT
Maps all internal BPF instructions into x86_64 instructions.
This patch replaces original BPF x64 JIT with internal BPF x64 JIT.
sysctl net.core.bpf_jit_enable is reused as on/off switch.
Performance:
1. old BPF JIT and internal BPF JIT generate equivalent x86_64 code.
No performance difference is observed for filters that were JIT-able before
Example assembler code for BPF filter "tcpdump port 22"
original BPF -> old JIT: original BPF -> internal BPF -> new JIT:
0: push %rbp 0: push %rbp
1: mov %rsp,%rbp 1: mov %rsp,%rbp
4: sub $0x60,%rsp 4: sub $0x228,%rsp
8: mov %rbx,-0x8(%rbp) b: mov %rbx,-0x228(%rbp) // prologue
12: mov %r13,-0x220(%rbp)
19: mov %r14,-0x218(%rbp)
20: mov %r15,-0x210(%rbp)
27: xor %eax,%eax // clear A
c: xor %ebx,%ebx 29: xor %r13,%r13 // clear X
e: mov 0x68(%rdi),%r9d 2c: mov 0x68(%rdi),%r9d
12: sub 0x6c(%rdi),%r9d 30: sub 0x6c(%rdi),%r9d
16: mov 0xd8(%rdi),%r8 34: mov 0xd8(%rdi),%r10
3b: mov %rdi,%rbx
1d: mov $0xc,%esi 3e: mov $0xc,%esi
22: callq 0xffffffffe1021e15 43: callq 0xffffffffe102bd75
27: cmp $0x86dd,%eax 48: cmp $0x86dd,%rax
2c: jne 0x0000000000000069 4f: jne 0x000000000000009a
2e: mov $0x14,%esi 51: mov $0x14,%esi
33: callq 0xffffffffe1021e31 56: callq 0xffffffffe102bd91
38: cmp $0x84,%eax 5b: cmp $0x84,%rax
3d: je 0x0000000000000049 62: je 0x0000000000000074
3f: cmp $0x6,%eax 64: cmp $0x6,%rax
42: je 0x0000000000000049 68: je 0x0000000000000074
44: cmp $0x11,%eax 6a: cmp $0x11,%rax
47: jne 0x00000000000000c6 6e: jne 0x0000000000000117
49: mov $0x36,%esi 74: mov $0x36,%esi
4e: callq 0xffffffffe1021e15 79: callq 0xffffffffe102bd75
53: cmp $0x16,%eax 7e: cmp $0x16,%rax
56: je 0x00000000000000bf 82: je 0x0000000000000110
58: mov $0x38,%esi 88: mov $0x38,%esi
5d: callq 0xffffffffe1021e15 8d: callq 0xffffffffe102bd75
62: cmp $0x16,%eax 92: cmp $0x16,%rax
65: je 0x00000000000000bf 96: je 0x0000000000000110
67: jmp 0x00000000000000c6 98: jmp 0x0000000000000117
69: cmp $0x800,%eax 9a: cmp $0x800,%rax
6e: jne 0x00000000000000c6 a1: jne 0x0000000000000117
70: mov $0x17,%esi a3: mov $0x17,%esi
75: callq 0xffffffffe1021e31 a8: callq 0xffffffffe102bd91
7a: cmp $0x84,%eax ad: cmp $0x84,%rax
7f: je 0x000000000000008b b4: je 0x00000000000000c2
81: cmp $0x6,%eax b6: cmp $0x6,%rax
84: je 0x000000000000008b ba: je 0x00000000000000c2
86: cmp $0x11,%eax bc: cmp $0x11,%rax
89: jne 0x00000000000000c6 c0: jne 0x0000000000000117
8b: mov $0x14,%esi c2: mov $0x14,%esi
90: callq 0xffffffffe1021e15 c7: callq 0xffffffffe102bd75
95: test $0x1fff,%ax cc: test $0x1fff,%rax
99: jne 0x00000000000000c6 d3: jne 0x0000000000000117
d5: mov %rax,%r14
9b: mov $0xe,%esi d8: mov $0xe,%esi
a0: callq 0xffffffffe1021e44 dd: callq 0xffffffffe102bd91 // MSH
e2: and $0xf,%eax
e5: shl $0x2,%eax
e8: mov %rax,%r13
eb: mov %r14,%rax
ee: mov %r13,%rsi
a5: lea 0xe(%rbx),%esi f1: add $0xe,%esi
a8: callq 0xffffffffe1021e0d f4: callq 0xffffffffe102bd6d
ad: cmp $0x16,%eax f9: cmp $0x16,%rax
b0: je 0x00000000000000bf fd: je 0x0000000000000110
ff: mov %r13,%rsi
b2: lea 0x10(%rbx),%esi 102: add $0x10,%esi
b5: callq 0xffffffffe1021e0d 105: callq 0xffffffffe102bd6d
ba: cmp $0x16,%eax 10a: cmp $0x16,%rax
bd: jne 0x00000000000000c6 10e: jne 0x0000000000000117
bf: mov $0xffff,%eax 110: mov $0xffff,%eax
c4: jmp 0x00000000000000c8 115: jmp 0x000000000000011c
c6: xor %eax,%eax 117: mov $0x0,%eax
c8: mov -0x8(%rbp),%rbx 11c: mov -0x228(%rbp),%rbx // epilogue
cc: leaveq 123: mov -0x220(%rbp),%r13
cd: retq 12a: mov -0x218(%rbp),%r14
131: mov -0x210(%rbp),%r15
138: leaveq
139: retq
On fully cached SKBs both JITed functions take 12 nsec to execute.
BPF interpreter executes the program in 30 nsec.
The difference in generated assembler is due to the following:
Old BPF imlements LDX_MSH instruction via sk_load_byte_msh() helper function
inside bpf_jit.S.
New JIT removes the helper and does it explicitly, so ldx_msh cost
is the same for both JITs, but generated code looks longer.
New JIT has 4 registers to save, so prologue/epilogue are larger,
but the cost is within noise on x64.
Old JIT checks whether first insn clears A and if not emits 'xor %eax,%eax'.
New JIT clears %rax unconditionally.
2. old BPF JIT doesn't support ANC_NLATTR, ANC_PAY_OFFSET, ANC_RANDOM
extensions. New JIT supports all BPF extensions.
Performance of such filters improves 2-4 times depending on a filter.
The longer the filter the higher performance gain.
Synthetic benchmarks with many ancillary loads see 20x speedup
which seems to be the maximum gain from JIT
Notes:
. net.core.bpf_jit_enable=2 + tools/net/bpf_jit_disasm is still functional
and can be used to see generated assembler
. there are two jit_compile() functions and code flow for classic filters is:
sk_attach_filter() - load classic BPF
bpf_jit_compile() - try to JIT from classic BPF
sk_convert_filter() - convert classic to internal
bpf_int_jit_compile() - JIT from internal BPF
seccomp and tracing filters will just call bpf_int_jit_compile()
Signed-off-by: Alexei Starovoitov <ast@plumgrid.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
2014-05-14 02:50:46 +00:00
|
|
|
* This error will be seen if new instruction was added
|
x86/bpf: Clean up non-standard comments, to make the code more readable
So by chance I looked into x86 assembly in arch/x86/net/bpf_jit_comp.c and
noticed the weird and inconsistent comment style it mistakenly learned from
the networking code:
/* Multi-line comment ...
* ... looks like this.
*/
Fix this to use the standard comment style specified in Documentation/CodingStyle
and used in arch/x86/ as well:
/*
* Multi-line comment ...
* ... looks like this.
*/
Also, to quote Linus's ... more explicit views about this:
http://article.gmane.org/gmane.linux.kernel.cryptoapi/21066
> But no, the networking code picked *none* of the above sane formats.
> Instead, it picked these two models that are just half-arsed
> shit-for-brains:
>
> (no)
> /* This is disgusting drug-induced
> * crap, and should die
> */
>
> (no-no-no)
> /* This is also very nasty
> * and visually unbalanced */
>
> Please. The networking code actually has the *worst* possible comment
> style. You can literally find that (no-no-no) style, which is just
> really horribly disgusting and worse than the otherwise fairly similar
> (d) in pretty much every way.
Also improve the comments and some other details while at it:
- Don't mix same-line and previous-line comment style on otherwise
identical code patterns within the same function,
- capitalize 'BPF' and x86 register names consistently,
- capitalize sentences consistently,
- instead of 'x64' use 'x86-64': x64 is a Microsoft specific term,
- use more consistent punctuation,
- use standard coding style in macros as well,
- fix typos and a few other minor details.
Consistent coding style is not optional, at least in arch/x86/.
No change in functionality.
( In case this commit causes conflicts with pending development code
I'll be glad to help resolve any conflicts! )
Acked-by: Thomas Gleixner <tglx@linutronix.de>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Borislav Petkov <bp@alien8.de>
Cc: H. Peter Anvin <hpa@zytor.com>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: David S. Miller <davem@davemloft.net>
Cc: Eric Dumazet <edumazet@google.com>
Cc: Daniel Borkmann <daniel@iogearbox.net>
Cc: Alexei Starovoitov <ast@fb.com>
Cc: Hideaki YOSHIFUJI <yoshfuji@linux-ipv6.org>
Cc: netdev@vger.kernel.org
Cc: linux-kernel@vger.kernel.org
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
2018-04-27 09:54:40 +00:00
|
|
|
* to the interpreter, but not to the JIT, or if there is
|
|
|
|
* junk in bpf_prog.
|
net: filter: x86: internal BPF JIT
Maps all internal BPF instructions into x86_64 instructions.
This patch replaces original BPF x64 JIT with internal BPF x64 JIT.
sysctl net.core.bpf_jit_enable is reused as on/off switch.
Performance:
1. old BPF JIT and internal BPF JIT generate equivalent x86_64 code.
No performance difference is observed for filters that were JIT-able before
Example assembler code for BPF filter "tcpdump port 22"
original BPF -> old JIT: original BPF -> internal BPF -> new JIT:
0: push %rbp 0: push %rbp
1: mov %rsp,%rbp 1: mov %rsp,%rbp
4: sub $0x60,%rsp 4: sub $0x228,%rsp
8: mov %rbx,-0x8(%rbp) b: mov %rbx,-0x228(%rbp) // prologue
12: mov %r13,-0x220(%rbp)
19: mov %r14,-0x218(%rbp)
20: mov %r15,-0x210(%rbp)
27: xor %eax,%eax // clear A
c: xor %ebx,%ebx 29: xor %r13,%r13 // clear X
e: mov 0x68(%rdi),%r9d 2c: mov 0x68(%rdi),%r9d
12: sub 0x6c(%rdi),%r9d 30: sub 0x6c(%rdi),%r9d
16: mov 0xd8(%rdi),%r8 34: mov 0xd8(%rdi),%r10
3b: mov %rdi,%rbx
1d: mov $0xc,%esi 3e: mov $0xc,%esi
22: callq 0xffffffffe1021e15 43: callq 0xffffffffe102bd75
27: cmp $0x86dd,%eax 48: cmp $0x86dd,%rax
2c: jne 0x0000000000000069 4f: jne 0x000000000000009a
2e: mov $0x14,%esi 51: mov $0x14,%esi
33: callq 0xffffffffe1021e31 56: callq 0xffffffffe102bd91
38: cmp $0x84,%eax 5b: cmp $0x84,%rax
3d: je 0x0000000000000049 62: je 0x0000000000000074
3f: cmp $0x6,%eax 64: cmp $0x6,%rax
42: je 0x0000000000000049 68: je 0x0000000000000074
44: cmp $0x11,%eax 6a: cmp $0x11,%rax
47: jne 0x00000000000000c6 6e: jne 0x0000000000000117
49: mov $0x36,%esi 74: mov $0x36,%esi
4e: callq 0xffffffffe1021e15 79: callq 0xffffffffe102bd75
53: cmp $0x16,%eax 7e: cmp $0x16,%rax
56: je 0x00000000000000bf 82: je 0x0000000000000110
58: mov $0x38,%esi 88: mov $0x38,%esi
5d: callq 0xffffffffe1021e15 8d: callq 0xffffffffe102bd75
62: cmp $0x16,%eax 92: cmp $0x16,%rax
65: je 0x00000000000000bf 96: je 0x0000000000000110
67: jmp 0x00000000000000c6 98: jmp 0x0000000000000117
69: cmp $0x800,%eax 9a: cmp $0x800,%rax
6e: jne 0x00000000000000c6 a1: jne 0x0000000000000117
70: mov $0x17,%esi a3: mov $0x17,%esi
75: callq 0xffffffffe1021e31 a8: callq 0xffffffffe102bd91
7a: cmp $0x84,%eax ad: cmp $0x84,%rax
7f: je 0x000000000000008b b4: je 0x00000000000000c2
81: cmp $0x6,%eax b6: cmp $0x6,%rax
84: je 0x000000000000008b ba: je 0x00000000000000c2
86: cmp $0x11,%eax bc: cmp $0x11,%rax
89: jne 0x00000000000000c6 c0: jne 0x0000000000000117
8b: mov $0x14,%esi c2: mov $0x14,%esi
90: callq 0xffffffffe1021e15 c7: callq 0xffffffffe102bd75
95: test $0x1fff,%ax cc: test $0x1fff,%rax
99: jne 0x00000000000000c6 d3: jne 0x0000000000000117
d5: mov %rax,%r14
9b: mov $0xe,%esi d8: mov $0xe,%esi
a0: callq 0xffffffffe1021e44 dd: callq 0xffffffffe102bd91 // MSH
e2: and $0xf,%eax
e5: shl $0x2,%eax
e8: mov %rax,%r13
eb: mov %r14,%rax
ee: mov %r13,%rsi
a5: lea 0xe(%rbx),%esi f1: add $0xe,%esi
a8: callq 0xffffffffe1021e0d f4: callq 0xffffffffe102bd6d
ad: cmp $0x16,%eax f9: cmp $0x16,%rax
b0: je 0x00000000000000bf fd: je 0x0000000000000110
ff: mov %r13,%rsi
b2: lea 0x10(%rbx),%esi 102: add $0x10,%esi
b5: callq 0xffffffffe1021e0d 105: callq 0xffffffffe102bd6d
ba: cmp $0x16,%eax 10a: cmp $0x16,%rax
bd: jne 0x00000000000000c6 10e: jne 0x0000000000000117
bf: mov $0xffff,%eax 110: mov $0xffff,%eax
c4: jmp 0x00000000000000c8 115: jmp 0x000000000000011c
c6: xor %eax,%eax 117: mov $0x0,%eax
c8: mov -0x8(%rbp),%rbx 11c: mov -0x228(%rbp),%rbx // epilogue
cc: leaveq 123: mov -0x220(%rbp),%r13
cd: retq 12a: mov -0x218(%rbp),%r14
131: mov -0x210(%rbp),%r15
138: leaveq
139: retq
On fully cached SKBs both JITed functions take 12 nsec to execute.
BPF interpreter executes the program in 30 nsec.
The difference in generated assembler is due to the following:
Old BPF imlements LDX_MSH instruction via sk_load_byte_msh() helper function
inside bpf_jit.S.
New JIT removes the helper and does it explicitly, so ldx_msh cost
is the same for both JITs, but generated code looks longer.
New JIT has 4 registers to save, so prologue/epilogue are larger,
but the cost is within noise on x64.
Old JIT checks whether first insn clears A and if not emits 'xor %eax,%eax'.
New JIT clears %rax unconditionally.
2. old BPF JIT doesn't support ANC_NLATTR, ANC_PAY_OFFSET, ANC_RANDOM
extensions. New JIT supports all BPF extensions.
Performance of such filters improves 2-4 times depending on a filter.
The longer the filter the higher performance gain.
Synthetic benchmarks with many ancillary loads see 20x speedup
which seems to be the maximum gain from JIT
Notes:
. net.core.bpf_jit_enable=2 + tools/net/bpf_jit_disasm is still functional
and can be used to see generated assembler
. there are two jit_compile() functions and code flow for classic filters is:
sk_attach_filter() - load classic BPF
bpf_jit_compile() - try to JIT from classic BPF
sk_convert_filter() - convert classic to internal
bpf_int_jit_compile() - JIT from internal BPF
seccomp and tracing filters will just call bpf_int_jit_compile()
Signed-off-by: Alexei Starovoitov <ast@plumgrid.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
2014-05-14 02:50:46 +00:00
|
|
|
*/
|
|
|
|
pr_err("bpf_jit: unknown opcode %02x\n", insn->code);
|
2014-05-14 02:50:45 +00:00
|
|
|
return -EINVAL;
|
|
|
|
}
|
net: filter: x86: internal BPF JIT
Maps all internal BPF instructions into x86_64 instructions.
This patch replaces original BPF x64 JIT with internal BPF x64 JIT.
sysctl net.core.bpf_jit_enable is reused as on/off switch.
Performance:
1. old BPF JIT and internal BPF JIT generate equivalent x86_64 code.
No performance difference is observed for filters that were JIT-able before
Example assembler code for BPF filter "tcpdump port 22"
original BPF -> old JIT: original BPF -> internal BPF -> new JIT:
0: push %rbp 0: push %rbp
1: mov %rsp,%rbp 1: mov %rsp,%rbp
4: sub $0x60,%rsp 4: sub $0x228,%rsp
8: mov %rbx,-0x8(%rbp) b: mov %rbx,-0x228(%rbp) // prologue
12: mov %r13,-0x220(%rbp)
19: mov %r14,-0x218(%rbp)
20: mov %r15,-0x210(%rbp)
27: xor %eax,%eax // clear A
c: xor %ebx,%ebx 29: xor %r13,%r13 // clear X
e: mov 0x68(%rdi),%r9d 2c: mov 0x68(%rdi),%r9d
12: sub 0x6c(%rdi),%r9d 30: sub 0x6c(%rdi),%r9d
16: mov 0xd8(%rdi),%r8 34: mov 0xd8(%rdi),%r10
3b: mov %rdi,%rbx
1d: mov $0xc,%esi 3e: mov $0xc,%esi
22: callq 0xffffffffe1021e15 43: callq 0xffffffffe102bd75
27: cmp $0x86dd,%eax 48: cmp $0x86dd,%rax
2c: jne 0x0000000000000069 4f: jne 0x000000000000009a
2e: mov $0x14,%esi 51: mov $0x14,%esi
33: callq 0xffffffffe1021e31 56: callq 0xffffffffe102bd91
38: cmp $0x84,%eax 5b: cmp $0x84,%rax
3d: je 0x0000000000000049 62: je 0x0000000000000074
3f: cmp $0x6,%eax 64: cmp $0x6,%rax
42: je 0x0000000000000049 68: je 0x0000000000000074
44: cmp $0x11,%eax 6a: cmp $0x11,%rax
47: jne 0x00000000000000c6 6e: jne 0x0000000000000117
49: mov $0x36,%esi 74: mov $0x36,%esi
4e: callq 0xffffffffe1021e15 79: callq 0xffffffffe102bd75
53: cmp $0x16,%eax 7e: cmp $0x16,%rax
56: je 0x00000000000000bf 82: je 0x0000000000000110
58: mov $0x38,%esi 88: mov $0x38,%esi
5d: callq 0xffffffffe1021e15 8d: callq 0xffffffffe102bd75
62: cmp $0x16,%eax 92: cmp $0x16,%rax
65: je 0x00000000000000bf 96: je 0x0000000000000110
67: jmp 0x00000000000000c6 98: jmp 0x0000000000000117
69: cmp $0x800,%eax 9a: cmp $0x800,%rax
6e: jne 0x00000000000000c6 a1: jne 0x0000000000000117
70: mov $0x17,%esi a3: mov $0x17,%esi
75: callq 0xffffffffe1021e31 a8: callq 0xffffffffe102bd91
7a: cmp $0x84,%eax ad: cmp $0x84,%rax
7f: je 0x000000000000008b b4: je 0x00000000000000c2
81: cmp $0x6,%eax b6: cmp $0x6,%rax
84: je 0x000000000000008b ba: je 0x00000000000000c2
86: cmp $0x11,%eax bc: cmp $0x11,%rax
89: jne 0x00000000000000c6 c0: jne 0x0000000000000117
8b: mov $0x14,%esi c2: mov $0x14,%esi
90: callq 0xffffffffe1021e15 c7: callq 0xffffffffe102bd75
95: test $0x1fff,%ax cc: test $0x1fff,%rax
99: jne 0x00000000000000c6 d3: jne 0x0000000000000117
d5: mov %rax,%r14
9b: mov $0xe,%esi d8: mov $0xe,%esi
a0: callq 0xffffffffe1021e44 dd: callq 0xffffffffe102bd91 // MSH
e2: and $0xf,%eax
e5: shl $0x2,%eax
e8: mov %rax,%r13
eb: mov %r14,%rax
ee: mov %r13,%rsi
a5: lea 0xe(%rbx),%esi f1: add $0xe,%esi
a8: callq 0xffffffffe1021e0d f4: callq 0xffffffffe102bd6d
ad: cmp $0x16,%eax f9: cmp $0x16,%rax
b0: je 0x00000000000000bf fd: je 0x0000000000000110
ff: mov %r13,%rsi
b2: lea 0x10(%rbx),%esi 102: add $0x10,%esi
b5: callq 0xffffffffe1021e0d 105: callq 0xffffffffe102bd6d
ba: cmp $0x16,%eax 10a: cmp $0x16,%rax
bd: jne 0x00000000000000c6 10e: jne 0x0000000000000117
bf: mov $0xffff,%eax 110: mov $0xffff,%eax
c4: jmp 0x00000000000000c8 115: jmp 0x000000000000011c
c6: xor %eax,%eax 117: mov $0x0,%eax
c8: mov -0x8(%rbp),%rbx 11c: mov -0x228(%rbp),%rbx // epilogue
cc: leaveq 123: mov -0x220(%rbp),%r13
cd: retq 12a: mov -0x218(%rbp),%r14
131: mov -0x210(%rbp),%r15
138: leaveq
139: retq
On fully cached SKBs both JITed functions take 12 nsec to execute.
BPF interpreter executes the program in 30 nsec.
The difference in generated assembler is due to the following:
Old BPF imlements LDX_MSH instruction via sk_load_byte_msh() helper function
inside bpf_jit.S.
New JIT removes the helper and does it explicitly, so ldx_msh cost
is the same for both JITs, but generated code looks longer.
New JIT has 4 registers to save, so prologue/epilogue are larger,
but the cost is within noise on x64.
Old JIT checks whether first insn clears A and if not emits 'xor %eax,%eax'.
New JIT clears %rax unconditionally.
2. old BPF JIT doesn't support ANC_NLATTR, ANC_PAY_OFFSET, ANC_RANDOM
extensions. New JIT supports all BPF extensions.
Performance of such filters improves 2-4 times depending on a filter.
The longer the filter the higher performance gain.
Synthetic benchmarks with many ancillary loads see 20x speedup
which seems to be the maximum gain from JIT
Notes:
. net.core.bpf_jit_enable=2 + tools/net/bpf_jit_disasm is still functional
and can be used to see generated assembler
. there are two jit_compile() functions and code flow for classic filters is:
sk_attach_filter() - load classic BPF
bpf_jit_compile() - try to JIT from classic BPF
sk_convert_filter() - convert classic to internal
bpf_int_jit_compile() - JIT from internal BPF
seccomp and tracing filters will just call bpf_int_jit_compile()
Signed-off-by: Alexei Starovoitov <ast@plumgrid.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
2014-05-14 02:50:46 +00:00
|
|
|
|
2014-05-14 02:50:45 +00:00
|
|
|
ilen = prog - temp;
|
2014-10-11 03:30:23 +00:00
|
|
|
if (ilen > BPF_MAX_INSN_SIZE) {
|
2017-02-16 21:24:49 +00:00
|
|
|
pr_err("bpf_jit: fatal insn size error\n");
|
2014-10-11 03:30:23 +00:00
|
|
|
return -EFAULT;
|
|
|
|
}
|
|
|
|
|
2014-05-14 02:50:45 +00:00
|
|
|
if (image) {
|
2021-04-05 21:52:15 +00:00
|
|
|
/*
|
|
|
|
* When populating the image, assert that:
|
|
|
|
*
|
|
|
|
* i) We do not write beyond the allocated space, and
|
|
|
|
* ii) addrs[i] did not change from the prior run, in order
|
|
|
|
* to validate assumptions made for computing branch
|
|
|
|
* displacements.
|
|
|
|
*/
|
|
|
|
if (unlikely(proglen + ilen > oldproglen ||
|
|
|
|
proglen + ilen != addrs[i])) {
|
2017-02-16 21:24:49 +00:00
|
|
|
pr_err("bpf_jit: fatal error\n");
|
2014-05-14 02:50:45 +00:00
|
|
|
return -EFAULT;
|
2011-04-20 09:27:32 +00:00
|
|
|
}
|
2022-02-04 18:57:42 +00:00
|
|
|
memcpy(rw_image + proglen, temp, ilen);
|
2011-04-20 09:27:32 +00:00
|
|
|
}
|
2014-05-14 02:50:45 +00:00
|
|
|
proglen += ilen;
|
|
|
|
addrs[i] = proglen;
|
|
|
|
prog = temp;
|
|
|
|
}
|
2019-10-16 03:25:03 +00:00
|
|
|
|
|
|
|
if (image && excnt != bpf_prog->aux->num_exentries) {
|
|
|
|
pr_err("extable is not populated\n");
|
|
|
|
return -EFAULT;
|
|
|
|
}
|
2014-05-14 02:50:45 +00:00
|
|
|
return proglen;
|
|
|
|
}
|
|
|
|
|
bpf, x86: allow function arguments up to 12 for TRACING
For now, the BPF program of type BPF_PROG_TYPE_TRACING can only be used
on the kernel functions whose arguments count less than or equal to 6, if
not considering '> 8 bytes' struct argument. This is not friendly at all,
as too many functions have arguments count more than 6.
According to the current kernel version, below is a statistics of the
function arguments count:
argument count | function count
7 | 704
8 | 270
9 | 84
10 | 47
11 | 47
12 | 27
13 | 22
14 | 5
15 | 0
16 | 1
Therefore, let's enhance it by increasing the function arguments count
allowed in arch_prepare_bpf_trampoline(), for now, only x86_64.
For the case that we don't need to call origin function, which means
without BPF_TRAMP_F_CALL_ORIG, we need only copy the function arguments
that stored in the frame of the caller to current frame. The 7th and later
arguments are stored in "$rbp + 0x18", and they will be copied to the
stack area following where register values are saved.
For the case with BPF_TRAMP_F_CALL_ORIG, we need prepare the arguments
in stack before call origin function, which means we need alloc extra
"8 * (arg_count - 6)" memory in the top of the stack. Note, there should
not be any data be pushed to the stack before calling the origin function.
So 'rbx' value will be stored on a stack position higher than where stack
arguments are stored for BPF_TRAMP_F_CALL_ORIG.
According to the research of Yonghong, struct members should be all in
register or all on the stack. Meanwhile, the compiler will pass the
argument on regs if the remaining regs can hold the argument. Therefore,
we need save the arguments in order. Otherwise, disorder of the args can
happen. For example:
struct foo_struct {
long a;
int b;
};
int foo(char, char, char, char, char, struct foo_struct,
char);
the arg1-5,arg7 will be passed by regs, and arg6 will by stack. Therefore,
we should save/restore the arguments in the same order with the
declaration of foo(). And the args used as ctx in stack will be like this:
reg_arg6 -- copy from regs
stack_arg2 -- copy from stack
stack_arg1
reg_arg5 -- copy from regs
reg_arg4
reg_arg3
reg_arg2
reg_arg1
We use EMIT3_off32() or EMIT4() for "lea" and "sub". The range of the
imm in "lea" and "sub" is [-128, 127] if EMIT4() is used. Therefore,
we use EMIT3_off32() instead if the imm out of the range.
It works well for the FENTRY/FEXIT/MODIFY_RETURN.
Signed-off-by: Menglong Dong <imagedong@tencent.com>
Acked-by: Yonghong Song <yhs@fb.com>
Link: https://lore.kernel.org/r/20230713040738.1789742-3-imagedong@tencent.com
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
2023-07-13 04:07:37 +00:00
|
|
|
static void clean_stack_garbage(const struct btf_func_model *m,
|
|
|
|
u8 **pprog, int nr_stack_slots,
|
|
|
|
int stack_size)
|
2019-11-14 18:57:04 +00:00
|
|
|
{
|
bpf, x86: allow function arguments up to 12 for TRACING
For now, the BPF program of type BPF_PROG_TYPE_TRACING can only be used
on the kernel functions whose arguments count less than or equal to 6, if
not considering '> 8 bytes' struct argument. This is not friendly at all,
as too many functions have arguments count more than 6.
According to the current kernel version, below is a statistics of the
function arguments count:
argument count | function count
7 | 704
8 | 270
9 | 84
10 | 47
11 | 47
12 | 27
13 | 22
14 | 5
15 | 0
16 | 1
Therefore, let's enhance it by increasing the function arguments count
allowed in arch_prepare_bpf_trampoline(), for now, only x86_64.
For the case that we don't need to call origin function, which means
without BPF_TRAMP_F_CALL_ORIG, we need only copy the function arguments
that stored in the frame of the caller to current frame. The 7th and later
arguments are stored in "$rbp + 0x18", and they will be copied to the
stack area following where register values are saved.
For the case with BPF_TRAMP_F_CALL_ORIG, we need prepare the arguments
in stack before call origin function, which means we need alloc extra
"8 * (arg_count - 6)" memory in the top of the stack. Note, there should
not be any data be pushed to the stack before calling the origin function.
So 'rbx' value will be stored on a stack position higher than where stack
arguments are stored for BPF_TRAMP_F_CALL_ORIG.
According to the research of Yonghong, struct members should be all in
register or all on the stack. Meanwhile, the compiler will pass the
argument on regs if the remaining regs can hold the argument. Therefore,
we need save the arguments in order. Otherwise, disorder of the args can
happen. For example:
struct foo_struct {
long a;
int b;
};
int foo(char, char, char, char, char, struct foo_struct,
char);
the arg1-5,arg7 will be passed by regs, and arg6 will by stack. Therefore,
we should save/restore the arguments in the same order with the
declaration of foo(). And the args used as ctx in stack will be like this:
reg_arg6 -- copy from regs
stack_arg2 -- copy from stack
stack_arg1
reg_arg5 -- copy from regs
reg_arg4
reg_arg3
reg_arg2
reg_arg1
We use EMIT3_off32() or EMIT4() for "lea" and "sub". The range of the
imm in "lea" and "sub" is [-128, 127] if EMIT4() is used. Therefore,
we use EMIT3_off32() instead if the imm out of the range.
It works well for the FENTRY/FEXIT/MODIFY_RETURN.
Signed-off-by: Menglong Dong <imagedong@tencent.com>
Acked-by: Yonghong Song <yhs@fb.com>
Link: https://lore.kernel.org/r/20230713040738.1789742-3-imagedong@tencent.com
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
2023-07-13 04:07:37 +00:00
|
|
|
int arg_size, off;
|
|
|
|
u8 *prog;
|
|
|
|
|
|
|
|
/* Generally speaking, the compiler will pass the arguments
|
|
|
|
* on-stack with "push" instruction, which will take 8-byte
|
|
|
|
* on the stack. In this case, there won't be garbage values
|
|
|
|
* while we copy the arguments from origin stack frame to current
|
|
|
|
* in BPF_DW.
|
|
|
|
*
|
|
|
|
* However, sometimes the compiler will only allocate 4-byte on
|
|
|
|
* the stack for the arguments. For now, this case will only
|
|
|
|
* happen if there is only one argument on-stack and its size
|
|
|
|
* not more than 4 byte. In this case, there will be garbage
|
|
|
|
* values on the upper 4-byte where we store the argument on
|
|
|
|
* current stack frame.
|
|
|
|
*
|
|
|
|
* arguments on origin stack:
|
|
|
|
*
|
|
|
|
* stack_arg_1(4-byte) xxx(4-byte)
|
|
|
|
*
|
|
|
|
* what we copy:
|
|
|
|
*
|
|
|
|
* stack_arg_1(8-byte): stack_arg_1(origin) xxx
|
|
|
|
*
|
|
|
|
* and the xxx is the garbage values which we should clean here.
|
|
|
|
*/
|
|
|
|
if (nr_stack_slots != 1)
|
|
|
|
return;
|
|
|
|
|
|
|
|
/* the size of the last argument */
|
|
|
|
arg_size = m->arg_size[m->nr_args - 1];
|
|
|
|
if (arg_size <= 4) {
|
|
|
|
off = -(stack_size - 4);
|
|
|
|
prog = *pprog;
|
|
|
|
/* mov DWORD PTR [rbp + off], 0 */
|
|
|
|
if (!is_imm8(off))
|
|
|
|
EMIT2_off32(0xC7, 0x85, off);
|
|
|
|
else
|
|
|
|
EMIT3(0xC7, 0x45, off);
|
|
|
|
EMIT(0, 4);
|
|
|
|
*pprog = prog;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
/* get the count of the regs that are used to pass arguments */
|
|
|
|
static int get_nr_used_regs(const struct btf_func_model *m)
|
|
|
|
{
|
|
|
|
int i, arg_regs, nr_used_regs = 0;
|
|
|
|
|
|
|
|
for (i = 0; i < min_t(int, m->nr_args, MAX_BPF_FUNC_ARGS); i++) {
|
|
|
|
arg_regs = (m->arg_size[i] + 7) / 8;
|
|
|
|
if (nr_used_regs + arg_regs <= 6)
|
|
|
|
nr_used_regs += arg_regs;
|
|
|
|
|
|
|
|
if (nr_used_regs >= 6)
|
|
|
|
break;
|
|
|
|
}
|
|
|
|
|
|
|
|
return nr_used_regs;
|
|
|
|
}
|
|
|
|
|
|
|
|
static void save_args(const struct btf_func_model *m, u8 **prog,
|
|
|
|
int stack_size, bool for_call_origin)
|
|
|
|
{
|
2023-07-19 11:03:30 +00:00
|
|
|
int arg_regs, first_off = 0, nr_regs = 0, nr_stack_slots = 0;
|
bpf, x86: allow function arguments up to 12 for TRACING
For now, the BPF program of type BPF_PROG_TYPE_TRACING can only be used
on the kernel functions whose arguments count less than or equal to 6, if
not considering '> 8 bytes' struct argument. This is not friendly at all,
as too many functions have arguments count more than 6.
According to the current kernel version, below is a statistics of the
function arguments count:
argument count | function count
7 | 704
8 | 270
9 | 84
10 | 47
11 | 47
12 | 27
13 | 22
14 | 5
15 | 0
16 | 1
Therefore, let's enhance it by increasing the function arguments count
allowed in arch_prepare_bpf_trampoline(), for now, only x86_64.
For the case that we don't need to call origin function, which means
without BPF_TRAMP_F_CALL_ORIG, we need only copy the function arguments
that stored in the frame of the caller to current frame. The 7th and later
arguments are stored in "$rbp + 0x18", and they will be copied to the
stack area following where register values are saved.
For the case with BPF_TRAMP_F_CALL_ORIG, we need prepare the arguments
in stack before call origin function, which means we need alloc extra
"8 * (arg_count - 6)" memory in the top of the stack. Note, there should
not be any data be pushed to the stack before calling the origin function.
So 'rbx' value will be stored on a stack position higher than where stack
arguments are stored for BPF_TRAMP_F_CALL_ORIG.
According to the research of Yonghong, struct members should be all in
register or all on the stack. Meanwhile, the compiler will pass the
argument on regs if the remaining regs can hold the argument. Therefore,
we need save the arguments in order. Otherwise, disorder of the args can
happen. For example:
struct foo_struct {
long a;
int b;
};
int foo(char, char, char, char, char, struct foo_struct,
char);
the arg1-5,arg7 will be passed by regs, and arg6 will by stack. Therefore,
we should save/restore the arguments in the same order with the
declaration of foo(). And the args used as ctx in stack will be like this:
reg_arg6 -- copy from regs
stack_arg2 -- copy from stack
stack_arg1
reg_arg5 -- copy from regs
reg_arg4
reg_arg3
reg_arg2
reg_arg1
We use EMIT3_off32() or EMIT4() for "lea" and "sub". The range of the
imm in "lea" and "sub" is [-128, 127] if EMIT4() is used. Therefore,
we use EMIT3_off32() instead if the imm out of the range.
It works well for the FENTRY/FEXIT/MODIFY_RETURN.
Signed-off-by: Menglong Dong <imagedong@tencent.com>
Acked-by: Yonghong Song <yhs@fb.com>
Link: https://lore.kernel.org/r/20230713040738.1789742-3-imagedong@tencent.com
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
2023-07-13 04:07:37 +00:00
|
|
|
int i, j;
|
2023-01-05 03:50:26 +00:00
|
|
|
|
2019-11-14 18:57:04 +00:00
|
|
|
/* Store function arguments to stack.
|
|
|
|
* For a function that accepts two pointers the sequence will be:
|
|
|
|
* mov QWORD PTR [rbp-0x10],rdi
|
|
|
|
* mov QWORD PTR [rbp-0x8],rsi
|
|
|
|
*/
|
bpf, x86: allow function arguments up to 12 for TRACING
For now, the BPF program of type BPF_PROG_TYPE_TRACING can only be used
on the kernel functions whose arguments count less than or equal to 6, if
not considering '> 8 bytes' struct argument. This is not friendly at all,
as too many functions have arguments count more than 6.
According to the current kernel version, below is a statistics of the
function arguments count:
argument count | function count
7 | 704
8 | 270
9 | 84
10 | 47
11 | 47
12 | 27
13 | 22
14 | 5
15 | 0
16 | 1
Therefore, let's enhance it by increasing the function arguments count
allowed in arch_prepare_bpf_trampoline(), for now, only x86_64.
For the case that we don't need to call origin function, which means
without BPF_TRAMP_F_CALL_ORIG, we need only copy the function arguments
that stored in the frame of the caller to current frame. The 7th and later
arguments are stored in "$rbp + 0x18", and they will be copied to the
stack area following where register values are saved.
For the case with BPF_TRAMP_F_CALL_ORIG, we need prepare the arguments
in stack before call origin function, which means we need alloc extra
"8 * (arg_count - 6)" memory in the top of the stack. Note, there should
not be any data be pushed to the stack before calling the origin function.
So 'rbx' value will be stored on a stack position higher than where stack
arguments are stored for BPF_TRAMP_F_CALL_ORIG.
According to the research of Yonghong, struct members should be all in
register or all on the stack. Meanwhile, the compiler will pass the
argument on regs if the remaining regs can hold the argument. Therefore,
we need save the arguments in order. Otherwise, disorder of the args can
happen. For example:
struct foo_struct {
long a;
int b;
};
int foo(char, char, char, char, char, struct foo_struct,
char);
the arg1-5,arg7 will be passed by regs, and arg6 will by stack. Therefore,
we should save/restore the arguments in the same order with the
declaration of foo(). And the args used as ctx in stack will be like this:
reg_arg6 -- copy from regs
stack_arg2 -- copy from stack
stack_arg1
reg_arg5 -- copy from regs
reg_arg4
reg_arg3
reg_arg2
reg_arg1
We use EMIT3_off32() or EMIT4() for "lea" and "sub". The range of the
imm in "lea" and "sub" is [-128, 127] if EMIT4() is used. Therefore,
we use EMIT3_off32() instead if the imm out of the range.
It works well for the FENTRY/FEXIT/MODIFY_RETURN.
Signed-off-by: Menglong Dong <imagedong@tencent.com>
Acked-by: Yonghong Song <yhs@fb.com>
Link: https://lore.kernel.org/r/20230713040738.1789742-3-imagedong@tencent.com
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
2023-07-13 04:07:37 +00:00
|
|
|
for (i = 0; i < min_t(int, m->nr_args, MAX_BPF_FUNC_ARGS); i++) {
|
|
|
|
arg_regs = (m->arg_size[i] + 7) / 8;
|
|
|
|
|
|
|
|
/* According to the research of Yonghong, struct members
|
|
|
|
* should be all in register or all on the stack.
|
|
|
|
* Meanwhile, the compiler will pass the argument on regs
|
|
|
|
* if the remaining regs can hold the argument.
|
|
|
|
*
|
|
|
|
* Disorder of the args can happen. For example:
|
|
|
|
*
|
|
|
|
* struct foo_struct {
|
|
|
|
* long a;
|
|
|
|
* int b;
|
|
|
|
* };
|
|
|
|
* int foo(char, char, char, char, char, struct foo_struct,
|
|
|
|
* char);
|
|
|
|
*
|
|
|
|
* the arg1-5,arg7 will be passed by regs, and arg6 will
|
|
|
|
* by stack.
|
|
|
|
*/
|
|
|
|
if (nr_regs + arg_regs > 6) {
|
|
|
|
/* copy function arguments from origin stack frame
|
|
|
|
* into current stack frame.
|
|
|
|
*
|
|
|
|
* The starting address of the arguments on-stack
|
|
|
|
* is:
|
|
|
|
* rbp + 8(push rbp) +
|
|
|
|
* 8(return addr of origin call) +
|
|
|
|
* 8(return addr of the caller)
|
|
|
|
* which means: rbp + 24
|
|
|
|
*/
|
|
|
|
for (j = 0; j < arg_regs; j++) {
|
|
|
|
emit_ldx(prog, BPF_DW, BPF_REG_0, BPF_REG_FP,
|
|
|
|
nr_stack_slots * 8 + 0x18);
|
|
|
|
emit_stx(prog, BPF_DW, BPF_REG_FP, BPF_REG_0,
|
|
|
|
-stack_size);
|
|
|
|
|
|
|
|
if (!nr_stack_slots)
|
|
|
|
first_off = stack_size;
|
|
|
|
stack_size -= 8;
|
|
|
|
nr_stack_slots++;
|
|
|
|
}
|
|
|
|
} else {
|
|
|
|
/* Only copy the arguments on-stack to current
|
|
|
|
* 'stack_size' and ignore the regs, used to
|
2024-01-03 00:40:11 +00:00
|
|
|
* prepare the arguments on-stack for origin call.
|
bpf, x86: allow function arguments up to 12 for TRACING
For now, the BPF program of type BPF_PROG_TYPE_TRACING can only be used
on the kernel functions whose arguments count less than or equal to 6, if
not considering '> 8 bytes' struct argument. This is not friendly at all,
as too many functions have arguments count more than 6.
According to the current kernel version, below is a statistics of the
function arguments count:
argument count | function count
7 | 704
8 | 270
9 | 84
10 | 47
11 | 47
12 | 27
13 | 22
14 | 5
15 | 0
16 | 1
Therefore, let's enhance it by increasing the function arguments count
allowed in arch_prepare_bpf_trampoline(), for now, only x86_64.
For the case that we don't need to call origin function, which means
without BPF_TRAMP_F_CALL_ORIG, we need only copy the function arguments
that stored in the frame of the caller to current frame. The 7th and later
arguments are stored in "$rbp + 0x18", and they will be copied to the
stack area following where register values are saved.
For the case with BPF_TRAMP_F_CALL_ORIG, we need prepare the arguments
in stack before call origin function, which means we need alloc extra
"8 * (arg_count - 6)" memory in the top of the stack. Note, there should
not be any data be pushed to the stack before calling the origin function.
So 'rbx' value will be stored on a stack position higher than where stack
arguments are stored for BPF_TRAMP_F_CALL_ORIG.
According to the research of Yonghong, struct members should be all in
register or all on the stack. Meanwhile, the compiler will pass the
argument on regs if the remaining regs can hold the argument. Therefore,
we need save the arguments in order. Otherwise, disorder of the args can
happen. For example:
struct foo_struct {
long a;
int b;
};
int foo(char, char, char, char, char, struct foo_struct,
char);
the arg1-5,arg7 will be passed by regs, and arg6 will by stack. Therefore,
we should save/restore the arguments in the same order with the
declaration of foo(). And the args used as ctx in stack will be like this:
reg_arg6 -- copy from regs
stack_arg2 -- copy from stack
stack_arg1
reg_arg5 -- copy from regs
reg_arg4
reg_arg3
reg_arg2
reg_arg1
We use EMIT3_off32() or EMIT4() for "lea" and "sub". The range of the
imm in "lea" and "sub" is [-128, 127] if EMIT4() is used. Therefore,
we use EMIT3_off32() instead if the imm out of the range.
It works well for the FENTRY/FEXIT/MODIFY_RETURN.
Signed-off-by: Menglong Dong <imagedong@tencent.com>
Acked-by: Yonghong Song <yhs@fb.com>
Link: https://lore.kernel.org/r/20230713040738.1789742-3-imagedong@tencent.com
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
2023-07-13 04:07:37 +00:00
|
|
|
*/
|
|
|
|
if (for_call_origin) {
|
|
|
|
nr_regs += arg_regs;
|
|
|
|
continue;
|
|
|
|
}
|
|
|
|
|
|
|
|
/* copy the arguments from regs into stack */
|
|
|
|
for (j = 0; j < arg_regs; j++) {
|
|
|
|
emit_stx(prog, BPF_DW, BPF_REG_FP,
|
|
|
|
nr_regs == 5 ? X86_REG_R9 : BPF_REG_1 + nr_regs,
|
|
|
|
-stack_size);
|
|
|
|
stack_size -= 8;
|
|
|
|
nr_regs++;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
clean_stack_garbage(m, prog, nr_stack_slots, first_off);
|
2019-11-14 18:57:04 +00:00
|
|
|
}
|
|
|
|
|
bpf, x86: allow function arguments up to 12 for TRACING
For now, the BPF program of type BPF_PROG_TYPE_TRACING can only be used
on the kernel functions whose arguments count less than or equal to 6, if
not considering '> 8 bytes' struct argument. This is not friendly at all,
as too many functions have arguments count more than 6.
According to the current kernel version, below is a statistics of the
function arguments count:
argument count | function count
7 | 704
8 | 270
9 | 84
10 | 47
11 | 47
12 | 27
13 | 22
14 | 5
15 | 0
16 | 1
Therefore, let's enhance it by increasing the function arguments count
allowed in arch_prepare_bpf_trampoline(), for now, only x86_64.
For the case that we don't need to call origin function, which means
without BPF_TRAMP_F_CALL_ORIG, we need only copy the function arguments
that stored in the frame of the caller to current frame. The 7th and later
arguments are stored in "$rbp + 0x18", and they will be copied to the
stack area following where register values are saved.
For the case with BPF_TRAMP_F_CALL_ORIG, we need prepare the arguments
in stack before call origin function, which means we need alloc extra
"8 * (arg_count - 6)" memory in the top of the stack. Note, there should
not be any data be pushed to the stack before calling the origin function.
So 'rbx' value will be stored on a stack position higher than where stack
arguments are stored for BPF_TRAMP_F_CALL_ORIG.
According to the research of Yonghong, struct members should be all in
register or all on the stack. Meanwhile, the compiler will pass the
argument on regs if the remaining regs can hold the argument. Therefore,
we need save the arguments in order. Otherwise, disorder of the args can
happen. For example:
struct foo_struct {
long a;
int b;
};
int foo(char, char, char, char, char, struct foo_struct,
char);
the arg1-5,arg7 will be passed by regs, and arg6 will by stack. Therefore,
we should save/restore the arguments in the same order with the
declaration of foo(). And the args used as ctx in stack will be like this:
reg_arg6 -- copy from regs
stack_arg2 -- copy from stack
stack_arg1
reg_arg5 -- copy from regs
reg_arg4
reg_arg3
reg_arg2
reg_arg1
We use EMIT3_off32() or EMIT4() for "lea" and "sub". The range of the
imm in "lea" and "sub" is [-128, 127] if EMIT4() is used. Therefore,
we use EMIT3_off32() instead if the imm out of the range.
It works well for the FENTRY/FEXIT/MODIFY_RETURN.
Signed-off-by: Menglong Dong <imagedong@tencent.com>
Acked-by: Yonghong Song <yhs@fb.com>
Link: https://lore.kernel.org/r/20230713040738.1789742-3-imagedong@tencent.com
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
2023-07-13 04:07:37 +00:00
|
|
|
static void restore_regs(const struct btf_func_model *m, u8 **prog,
|
2019-11-14 18:57:04 +00:00
|
|
|
int stack_size)
|
|
|
|
{
|
bpf, x86: allow function arguments up to 12 for TRACING
For now, the BPF program of type BPF_PROG_TYPE_TRACING can only be used
on the kernel functions whose arguments count less than or equal to 6, if
not considering '> 8 bytes' struct argument. This is not friendly at all,
as too many functions have arguments count more than 6.
According to the current kernel version, below is a statistics of the
function arguments count:
argument count | function count
7 | 704
8 | 270
9 | 84
10 | 47
11 | 47
12 | 27
13 | 22
14 | 5
15 | 0
16 | 1
Therefore, let's enhance it by increasing the function arguments count
allowed in arch_prepare_bpf_trampoline(), for now, only x86_64.
For the case that we don't need to call origin function, which means
without BPF_TRAMP_F_CALL_ORIG, we need only copy the function arguments
that stored in the frame of the caller to current frame. The 7th and later
arguments are stored in "$rbp + 0x18", and they will be copied to the
stack area following where register values are saved.
For the case with BPF_TRAMP_F_CALL_ORIG, we need prepare the arguments
in stack before call origin function, which means we need alloc extra
"8 * (arg_count - 6)" memory in the top of the stack. Note, there should
not be any data be pushed to the stack before calling the origin function.
So 'rbx' value will be stored on a stack position higher than where stack
arguments are stored for BPF_TRAMP_F_CALL_ORIG.
According to the research of Yonghong, struct members should be all in
register or all on the stack. Meanwhile, the compiler will pass the
argument on regs if the remaining regs can hold the argument. Therefore,
we need save the arguments in order. Otherwise, disorder of the args can
happen. For example:
struct foo_struct {
long a;
int b;
};
int foo(char, char, char, char, char, struct foo_struct,
char);
the arg1-5,arg7 will be passed by regs, and arg6 will by stack. Therefore,
we should save/restore the arguments in the same order with the
declaration of foo(). And the args used as ctx in stack will be like this:
reg_arg6 -- copy from regs
stack_arg2 -- copy from stack
stack_arg1
reg_arg5 -- copy from regs
reg_arg4
reg_arg3
reg_arg2
reg_arg1
We use EMIT3_off32() or EMIT4() for "lea" and "sub". The range of the
imm in "lea" and "sub" is [-128, 127] if EMIT4() is used. Therefore,
we use EMIT3_off32() instead if the imm out of the range.
It works well for the FENTRY/FEXIT/MODIFY_RETURN.
Signed-off-by: Menglong Dong <imagedong@tencent.com>
Acked-by: Yonghong Song <yhs@fb.com>
Link: https://lore.kernel.org/r/20230713040738.1789742-3-imagedong@tencent.com
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
2023-07-13 04:07:37 +00:00
|
|
|
int i, j, arg_regs, nr_regs = 0;
|
2019-11-14 18:57:04 +00:00
|
|
|
|
|
|
|
/* Restore function arguments from stack.
|
|
|
|
* For a function that accepts two pointers the sequence will be:
|
|
|
|
* EMIT4(0x48, 0x8B, 0x7D, 0xF0); mov rdi,QWORD PTR [rbp-0x10]
|
|
|
|
* EMIT4(0x48, 0x8B, 0x75, 0xF8); mov rsi,QWORD PTR [rbp-0x8]
|
bpf, x86: allow function arguments up to 12 for TRACING
For now, the BPF program of type BPF_PROG_TYPE_TRACING can only be used
on the kernel functions whose arguments count less than or equal to 6, if
not considering '> 8 bytes' struct argument. This is not friendly at all,
as too many functions have arguments count more than 6.
According to the current kernel version, below is a statistics of the
function arguments count:
argument count | function count
7 | 704
8 | 270
9 | 84
10 | 47
11 | 47
12 | 27
13 | 22
14 | 5
15 | 0
16 | 1
Therefore, let's enhance it by increasing the function arguments count
allowed in arch_prepare_bpf_trampoline(), for now, only x86_64.
For the case that we don't need to call origin function, which means
without BPF_TRAMP_F_CALL_ORIG, we need only copy the function arguments
that stored in the frame of the caller to current frame. The 7th and later
arguments are stored in "$rbp + 0x18", and they will be copied to the
stack area following where register values are saved.
For the case with BPF_TRAMP_F_CALL_ORIG, we need prepare the arguments
in stack before call origin function, which means we need alloc extra
"8 * (arg_count - 6)" memory in the top of the stack. Note, there should
not be any data be pushed to the stack before calling the origin function.
So 'rbx' value will be stored on a stack position higher than where stack
arguments are stored for BPF_TRAMP_F_CALL_ORIG.
According to the research of Yonghong, struct members should be all in
register or all on the stack. Meanwhile, the compiler will pass the
argument on regs if the remaining regs can hold the argument. Therefore,
we need save the arguments in order. Otherwise, disorder of the args can
happen. For example:
struct foo_struct {
long a;
int b;
};
int foo(char, char, char, char, char, struct foo_struct,
char);
the arg1-5,arg7 will be passed by regs, and arg6 will by stack. Therefore,
we should save/restore the arguments in the same order with the
declaration of foo(). And the args used as ctx in stack will be like this:
reg_arg6 -- copy from regs
stack_arg2 -- copy from stack
stack_arg1
reg_arg5 -- copy from regs
reg_arg4
reg_arg3
reg_arg2
reg_arg1
We use EMIT3_off32() or EMIT4() for "lea" and "sub". The range of the
imm in "lea" and "sub" is [-128, 127] if EMIT4() is used. Therefore,
we use EMIT3_off32() instead if the imm out of the range.
It works well for the FENTRY/FEXIT/MODIFY_RETURN.
Signed-off-by: Menglong Dong <imagedong@tencent.com>
Acked-by: Yonghong Song <yhs@fb.com>
Link: https://lore.kernel.org/r/20230713040738.1789742-3-imagedong@tencent.com
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
2023-07-13 04:07:37 +00:00
|
|
|
*
|
|
|
|
* The logic here is similar to what we do in save_args()
|
2019-11-14 18:57:04 +00:00
|
|
|
*/
|
bpf, x86: allow function arguments up to 12 for TRACING
For now, the BPF program of type BPF_PROG_TYPE_TRACING can only be used
on the kernel functions whose arguments count less than or equal to 6, if
not considering '> 8 bytes' struct argument. This is not friendly at all,
as too many functions have arguments count more than 6.
According to the current kernel version, below is a statistics of the
function arguments count:
argument count | function count
7 | 704
8 | 270
9 | 84
10 | 47
11 | 47
12 | 27
13 | 22
14 | 5
15 | 0
16 | 1
Therefore, let's enhance it by increasing the function arguments count
allowed in arch_prepare_bpf_trampoline(), for now, only x86_64.
For the case that we don't need to call origin function, which means
without BPF_TRAMP_F_CALL_ORIG, we need only copy the function arguments
that stored in the frame of the caller to current frame. The 7th and later
arguments are stored in "$rbp + 0x18", and they will be copied to the
stack area following where register values are saved.
For the case with BPF_TRAMP_F_CALL_ORIG, we need prepare the arguments
in stack before call origin function, which means we need alloc extra
"8 * (arg_count - 6)" memory in the top of the stack. Note, there should
not be any data be pushed to the stack before calling the origin function.
So 'rbx' value will be stored on a stack position higher than where stack
arguments are stored for BPF_TRAMP_F_CALL_ORIG.
According to the research of Yonghong, struct members should be all in
register or all on the stack. Meanwhile, the compiler will pass the
argument on regs if the remaining regs can hold the argument. Therefore,
we need save the arguments in order. Otherwise, disorder of the args can
happen. For example:
struct foo_struct {
long a;
int b;
};
int foo(char, char, char, char, char, struct foo_struct,
char);
the arg1-5,arg7 will be passed by regs, and arg6 will by stack. Therefore,
we should save/restore the arguments in the same order with the
declaration of foo(). And the args used as ctx in stack will be like this:
reg_arg6 -- copy from regs
stack_arg2 -- copy from stack
stack_arg1
reg_arg5 -- copy from regs
reg_arg4
reg_arg3
reg_arg2
reg_arg1
We use EMIT3_off32() or EMIT4() for "lea" and "sub". The range of the
imm in "lea" and "sub" is [-128, 127] if EMIT4() is used. Therefore,
we use EMIT3_off32() instead if the imm out of the range.
It works well for the FENTRY/FEXIT/MODIFY_RETURN.
Signed-off-by: Menglong Dong <imagedong@tencent.com>
Acked-by: Yonghong Song <yhs@fb.com>
Link: https://lore.kernel.org/r/20230713040738.1789742-3-imagedong@tencent.com
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
2023-07-13 04:07:37 +00:00
|
|
|
for (i = 0; i < min_t(int, m->nr_args, MAX_BPF_FUNC_ARGS); i++) {
|
|
|
|
arg_regs = (m->arg_size[i] + 7) / 8;
|
|
|
|
if (nr_regs + arg_regs <= 6) {
|
|
|
|
for (j = 0; j < arg_regs; j++) {
|
|
|
|
emit_ldx(prog, BPF_DW,
|
|
|
|
nr_regs == 5 ? X86_REG_R9 : BPF_REG_1 + nr_regs,
|
|
|
|
BPF_REG_FP,
|
|
|
|
-stack_size);
|
|
|
|
stack_size -= 8;
|
|
|
|
nr_regs++;
|
|
|
|
}
|
|
|
|
} else {
|
|
|
|
stack_size -= 8 * arg_regs;
|
|
|
|
}
|
|
|
|
|
|
|
|
if (nr_regs >= 6)
|
|
|
|
break;
|
|
|
|
}
|
2019-11-14 18:57:04 +00:00
|
|
|
}
|
|
|
|
|
2020-03-04 19:18:48 +00:00
|
|
|
static int invoke_bpf_prog(const struct btf_func_model *m, u8 **pprog,
|
2022-05-10 20:59:19 +00:00
|
|
|
struct bpf_tramp_link *l, int stack_size,
|
2023-12-06 22:40:54 +00:00
|
|
|
int run_ctx_off, bool save_ret,
|
|
|
|
void *image, void *rw_image)
|
2020-03-04 19:18:48 +00:00
|
|
|
{
|
|
|
|
u8 *prog = *pprog;
|
2021-02-10 03:36:29 +00:00
|
|
|
u8 *jmp_insn;
|
2022-05-10 20:59:20 +00:00
|
|
|
int ctx_cookie_off = offsetof(struct bpf_tramp_run_ctx, bpf_cookie);
|
2022-05-10 20:59:19 +00:00
|
|
|
struct bpf_prog *p = l->link.prog;
|
2022-05-10 20:59:21 +00:00
|
|
|
u64 cookie = l->cookie;
|
2020-03-04 19:18:48 +00:00
|
|
|
|
2022-05-10 20:59:21 +00:00
|
|
|
/* mov rdi, cookie */
|
|
|
|
emit_mov_imm64(&prog, BPF_REG_1, (long) cookie >> 32, (u32) (long) cookie);
|
2022-05-10 20:59:20 +00:00
|
|
|
|
|
|
|
/* Prepare struct bpf_tramp_run_ctx.
|
|
|
|
*
|
|
|
|
* bpf_tramp_run_ctx is already preserved by
|
|
|
|
* arch_prepare_bpf_trampoline().
|
|
|
|
*
|
|
|
|
* mov QWORD PTR [rbp - run_ctx_off + ctx_cookie_off], rdi
|
|
|
|
*/
|
|
|
|
emit_stx(&prog, BPF_DW, BPF_REG_FP, BPF_REG_1, -run_ctx_off + ctx_cookie_off);
|
|
|
|
|
2021-02-10 03:36:29 +00:00
|
|
|
/* arg1: mov rdi, progs[i] */
|
|
|
|
emit_mov_imm64(&prog, BPF_REG_1, (long) p >> 32, (u32) (long) p);
|
2022-05-10 20:59:20 +00:00
|
|
|
/* arg2: lea rsi, [rbp - ctx_cookie_off] */
|
bpf, x86: allow function arguments up to 12 for TRACING
For now, the BPF program of type BPF_PROG_TYPE_TRACING can only be used
on the kernel functions whose arguments count less than or equal to 6, if
not considering '> 8 bytes' struct argument. This is not friendly at all,
as too many functions have arguments count more than 6.
According to the current kernel version, below is a statistics of the
function arguments count:
argument count | function count
7 | 704
8 | 270
9 | 84
10 | 47
11 | 47
12 | 27
13 | 22
14 | 5
15 | 0
16 | 1
Therefore, let's enhance it by increasing the function arguments count
allowed in arch_prepare_bpf_trampoline(), for now, only x86_64.
For the case that we don't need to call origin function, which means
without BPF_TRAMP_F_CALL_ORIG, we need only copy the function arguments
that stored in the frame of the caller to current frame. The 7th and later
arguments are stored in "$rbp + 0x18", and they will be copied to the
stack area following where register values are saved.
For the case with BPF_TRAMP_F_CALL_ORIG, we need prepare the arguments
in stack before call origin function, which means we need alloc extra
"8 * (arg_count - 6)" memory in the top of the stack. Note, there should
not be any data be pushed to the stack before calling the origin function.
So 'rbx' value will be stored on a stack position higher than where stack
arguments are stored for BPF_TRAMP_F_CALL_ORIG.
According to the research of Yonghong, struct members should be all in
register or all on the stack. Meanwhile, the compiler will pass the
argument on regs if the remaining regs can hold the argument. Therefore,
we need save the arguments in order. Otherwise, disorder of the args can
happen. For example:
struct foo_struct {
long a;
int b;
};
int foo(char, char, char, char, char, struct foo_struct,
char);
the arg1-5,arg7 will be passed by regs, and arg6 will by stack. Therefore,
we should save/restore the arguments in the same order with the
declaration of foo(). And the args used as ctx in stack will be like this:
reg_arg6 -- copy from regs
stack_arg2 -- copy from stack
stack_arg1
reg_arg5 -- copy from regs
reg_arg4
reg_arg3
reg_arg2
reg_arg1
We use EMIT3_off32() or EMIT4() for "lea" and "sub". The range of the
imm in "lea" and "sub" is [-128, 127] if EMIT4() is used. Therefore,
we use EMIT3_off32() instead if the imm out of the range.
It works well for the FENTRY/FEXIT/MODIFY_RETURN.
Signed-off-by: Menglong Dong <imagedong@tencent.com>
Acked-by: Yonghong Song <yhs@fb.com>
Link: https://lore.kernel.org/r/20230713040738.1789742-3-imagedong@tencent.com
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
2023-07-13 04:07:37 +00:00
|
|
|
if (!is_imm8(-run_ctx_off))
|
|
|
|
EMIT3_off32(0x48, 0x8D, 0xB5, -run_ctx_off);
|
|
|
|
else
|
|
|
|
EMIT4(0x48, 0x8D, 0x75, -run_ctx_off);
|
2022-05-10 20:59:20 +00:00
|
|
|
|
2023-12-06 22:40:54 +00:00
|
|
|
if (emit_rsb_call(&prog, bpf_trampoline_enter(p), image + (prog - (u8 *)rw_image)))
|
2022-06-28 17:43:06 +00:00
|
|
|
return -EINVAL;
|
2021-02-10 03:36:28 +00:00
|
|
|
/* remember prog start time returned by __bpf_prog_enter */
|
|
|
|
emit_mov_reg(&prog, true, BPF_REG_6, BPF_REG_0);
|
2020-03-04 19:18:48 +00:00
|
|
|
|
2021-02-10 03:36:29 +00:00
|
|
|
/* if (__bpf_prog_enter*(prog) == 0)
|
|
|
|
* goto skip_exec_of_prog;
|
|
|
|
*/
|
|
|
|
EMIT3(0x48, 0x85, 0xC0); /* test rax,rax */
|
|
|
|
/* emit 2 nops that will be replaced with JE insn */
|
|
|
|
jmp_insn = prog;
|
|
|
|
emit_nops(&prog, 2);
|
|
|
|
|
2020-03-04 19:18:48 +00:00
|
|
|
/* arg1: lea rdi, [rbp - stack_size] */
|
bpf, x86: allow function arguments up to 12 for TRACING
For now, the BPF program of type BPF_PROG_TYPE_TRACING can only be used
on the kernel functions whose arguments count less than or equal to 6, if
not considering '> 8 bytes' struct argument. This is not friendly at all,
as too many functions have arguments count more than 6.
According to the current kernel version, below is a statistics of the
function arguments count:
argument count | function count
7 | 704
8 | 270
9 | 84
10 | 47
11 | 47
12 | 27
13 | 22
14 | 5
15 | 0
16 | 1
Therefore, let's enhance it by increasing the function arguments count
allowed in arch_prepare_bpf_trampoline(), for now, only x86_64.
For the case that we don't need to call origin function, which means
without BPF_TRAMP_F_CALL_ORIG, we need only copy the function arguments
that stored in the frame of the caller to current frame. The 7th and later
arguments are stored in "$rbp + 0x18", and they will be copied to the
stack area following where register values are saved.
For the case with BPF_TRAMP_F_CALL_ORIG, we need prepare the arguments
in stack before call origin function, which means we need alloc extra
"8 * (arg_count - 6)" memory in the top of the stack. Note, there should
not be any data be pushed to the stack before calling the origin function.
So 'rbx' value will be stored on a stack position higher than where stack
arguments are stored for BPF_TRAMP_F_CALL_ORIG.
According to the research of Yonghong, struct members should be all in
register or all on the stack. Meanwhile, the compiler will pass the
argument on regs if the remaining regs can hold the argument. Therefore,
we need save the arguments in order. Otherwise, disorder of the args can
happen. For example:
struct foo_struct {
long a;
int b;
};
int foo(char, char, char, char, char, struct foo_struct,
char);
the arg1-5,arg7 will be passed by regs, and arg6 will by stack. Therefore,
we should save/restore the arguments in the same order with the
declaration of foo(). And the args used as ctx in stack will be like this:
reg_arg6 -- copy from regs
stack_arg2 -- copy from stack
stack_arg1
reg_arg5 -- copy from regs
reg_arg4
reg_arg3
reg_arg2
reg_arg1
We use EMIT3_off32() or EMIT4() for "lea" and "sub". The range of the
imm in "lea" and "sub" is [-128, 127] if EMIT4() is used. Therefore,
we use EMIT3_off32() instead if the imm out of the range.
It works well for the FENTRY/FEXIT/MODIFY_RETURN.
Signed-off-by: Menglong Dong <imagedong@tencent.com>
Acked-by: Yonghong Song <yhs@fb.com>
Link: https://lore.kernel.org/r/20230713040738.1789742-3-imagedong@tencent.com
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
2023-07-13 04:07:37 +00:00
|
|
|
if (!is_imm8(-stack_size))
|
|
|
|
EMIT3_off32(0x48, 0x8D, 0xBD, -stack_size);
|
|
|
|
else
|
|
|
|
EMIT4(0x48, 0x8D, 0x7D, -stack_size);
|
2020-03-04 19:18:48 +00:00
|
|
|
/* arg2: progs[i]->insnsi for interpreter */
|
|
|
|
if (!p->jited)
|
|
|
|
emit_mov_imm64(&prog, BPF_REG_2,
|
|
|
|
(long) p->insnsi >> 32,
|
|
|
|
(u32) (long) p->insnsi);
|
|
|
|
/* call JITed bpf program or interpreter */
|
2023-12-06 22:40:54 +00:00
|
|
|
if (emit_rsb_call(&prog, p->bpf_func, image + (prog - (u8 *)rw_image)))
|
2020-03-04 19:18:48 +00:00
|
|
|
return -EINVAL;
|
|
|
|
|
2021-09-14 02:33:51 +00:00
|
|
|
/*
|
|
|
|
* BPF_TRAMP_MODIFY_RETURN trampolines can modify the return
|
2020-03-04 19:18:49 +00:00
|
|
|
* of the previous call which is then passed on the stack to
|
|
|
|
* the next BPF program.
|
2021-09-14 02:33:51 +00:00
|
|
|
*
|
|
|
|
* BPF_TRAMP_FENTRY trampoline may need to return the return
|
|
|
|
* value of BPF_PROG_TYPE_STRUCT_OPS prog.
|
2020-03-04 19:18:49 +00:00
|
|
|
*/
|
2021-09-14 02:33:51 +00:00
|
|
|
if (save_ret)
|
2020-03-04 19:18:49 +00:00
|
|
|
emit_stx(&prog, BPF_DW, BPF_REG_FP, BPF_REG_0, -8);
|
|
|
|
|
2021-02-10 03:36:29 +00:00
|
|
|
/* replace 2 nops with JE insn, since jmp target is known */
|
|
|
|
jmp_insn[0] = X86_JE;
|
|
|
|
jmp_insn[1] = prog - jmp_insn - 2;
|
|
|
|
|
2021-02-10 03:36:28 +00:00
|
|
|
/* arg1: mov rdi, progs[i] */
|
|
|
|
emit_mov_imm64(&prog, BPF_REG_1, (long) p >> 32, (u32) (long) p);
|
|
|
|
/* arg2: mov rsi, rbx <- start time in nsec */
|
|
|
|
emit_mov_reg(&prog, true, BPF_REG_2, BPF_REG_6);
|
2022-05-10 20:59:20 +00:00
|
|
|
/* arg3: lea rdx, [rbp - run_ctx_off] */
|
bpf, x86: allow function arguments up to 12 for TRACING
For now, the BPF program of type BPF_PROG_TYPE_TRACING can only be used
on the kernel functions whose arguments count less than or equal to 6, if
not considering '> 8 bytes' struct argument. This is not friendly at all,
as too many functions have arguments count more than 6.
According to the current kernel version, below is a statistics of the
function arguments count:
argument count | function count
7 | 704
8 | 270
9 | 84
10 | 47
11 | 47
12 | 27
13 | 22
14 | 5
15 | 0
16 | 1
Therefore, let's enhance it by increasing the function arguments count
allowed in arch_prepare_bpf_trampoline(), for now, only x86_64.
For the case that we don't need to call origin function, which means
without BPF_TRAMP_F_CALL_ORIG, we need only copy the function arguments
that stored in the frame of the caller to current frame. The 7th and later
arguments are stored in "$rbp + 0x18", and they will be copied to the
stack area following where register values are saved.
For the case with BPF_TRAMP_F_CALL_ORIG, we need prepare the arguments
in stack before call origin function, which means we need alloc extra
"8 * (arg_count - 6)" memory in the top of the stack. Note, there should
not be any data be pushed to the stack before calling the origin function.
So 'rbx' value will be stored on a stack position higher than where stack
arguments are stored for BPF_TRAMP_F_CALL_ORIG.
According to the research of Yonghong, struct members should be all in
register or all on the stack. Meanwhile, the compiler will pass the
argument on regs if the remaining regs can hold the argument. Therefore,
we need save the arguments in order. Otherwise, disorder of the args can
happen. For example:
struct foo_struct {
long a;
int b;
};
int foo(char, char, char, char, char, struct foo_struct,
char);
the arg1-5,arg7 will be passed by regs, and arg6 will by stack. Therefore,
we should save/restore the arguments in the same order with the
declaration of foo(). And the args used as ctx in stack will be like this:
reg_arg6 -- copy from regs
stack_arg2 -- copy from stack
stack_arg1
reg_arg5 -- copy from regs
reg_arg4
reg_arg3
reg_arg2
reg_arg1
We use EMIT3_off32() or EMIT4() for "lea" and "sub". The range of the
imm in "lea" and "sub" is [-128, 127] if EMIT4() is used. Therefore,
we use EMIT3_off32() instead if the imm out of the range.
It works well for the FENTRY/FEXIT/MODIFY_RETURN.
Signed-off-by: Menglong Dong <imagedong@tencent.com>
Acked-by: Yonghong Song <yhs@fb.com>
Link: https://lore.kernel.org/r/20230713040738.1789742-3-imagedong@tencent.com
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
2023-07-13 04:07:37 +00:00
|
|
|
if (!is_imm8(-run_ctx_off))
|
|
|
|
EMIT3_off32(0x48, 0x8D, 0x95, -run_ctx_off);
|
|
|
|
else
|
|
|
|
EMIT4(0x48, 0x8D, 0x55, -run_ctx_off);
|
2023-12-06 22:40:54 +00:00
|
|
|
if (emit_rsb_call(&prog, bpf_trampoline_exit(p), image + (prog - (u8 *)rw_image)))
|
2022-06-28 17:43:06 +00:00
|
|
|
return -EINVAL;
|
2020-03-04 19:18:48 +00:00
|
|
|
|
|
|
|
*pprog = prog;
|
|
|
|
return 0;
|
|
|
|
}
|
|
|
|
|
|
|
|
static void emit_align(u8 **pprog, u32 align)
|
|
|
|
{
|
|
|
|
u8 *target, *prog = *pprog;
|
|
|
|
|
|
|
|
target = PTR_ALIGN(prog, align);
|
|
|
|
if (target != prog)
|
|
|
|
emit_nops(&prog, target - prog);
|
|
|
|
|
|
|
|
*pprog = prog;
|
|
|
|
}
|
|
|
|
|
|
|
|
static int emit_cond_near_jump(u8 **pprog, void *func, void *ip, u8 jmp_cond)
|
|
|
|
{
|
|
|
|
u8 *prog = *pprog;
|
|
|
|
s64 offset;
|
|
|
|
|
|
|
|
offset = func - (ip + 2 + 4);
|
|
|
|
if (!is_simm32(offset)) {
|
|
|
|
pr_err("Target %p is out of range\n", func);
|
|
|
|
return -EINVAL;
|
|
|
|
}
|
|
|
|
EMIT2_off32(0x0F, jmp_cond + 0x10, offset);
|
|
|
|
*pprog = prog;
|
|
|
|
return 0;
|
|
|
|
}
|
|
|
|
|
bpf: Introduce BPF_MAP_TYPE_STRUCT_OPS
The patch introduces BPF_MAP_TYPE_STRUCT_OPS. The map value
is a kernel struct with its func ptr implemented in bpf prog.
This new map is the interface to register/unregister/introspect
a bpf implemented kernel struct.
The kernel struct is actually embedded inside another new struct
(or called the "value" struct in the code). For example,
"struct tcp_congestion_ops" is embbeded in:
struct bpf_struct_ops_tcp_congestion_ops {
refcount_t refcnt;
enum bpf_struct_ops_state state;
struct tcp_congestion_ops data; /* <-- kernel subsystem struct here */
}
The map value is "struct bpf_struct_ops_tcp_congestion_ops".
The "bpftool map dump" will then be able to show the
state ("inuse"/"tobefree") and the number of subsystem's refcnt (e.g.
number of tcp_sock in the tcp_congestion_ops case). This "value" struct
is created automatically by a macro. Having a separate "value" struct
will also make extending "struct bpf_struct_ops_XYZ" easier (e.g. adding
"void (*init)(void)" to "struct bpf_struct_ops_XYZ" to do some
initialization works before registering the struct_ops to the kernel
subsystem). The libbpf will take care of finding and populating the
"struct bpf_struct_ops_XYZ" from "struct XYZ".
Register a struct_ops to a kernel subsystem:
1. Load all needed BPF_PROG_TYPE_STRUCT_OPS prog(s)
2. Create a BPF_MAP_TYPE_STRUCT_OPS with attr->btf_vmlinux_value_type_id
set to the btf id "struct bpf_struct_ops_tcp_congestion_ops" of the
running kernel.
Instead of reusing the attr->btf_value_type_id,
btf_vmlinux_value_type_id s added such that attr->btf_fd can still be
used as the "user" btf which could store other useful sysadmin/debug
info that may be introduced in the furture,
e.g. creation-date/compiler-details/map-creator...etc.
3. Create a "struct bpf_struct_ops_tcp_congestion_ops" object as described
in the running kernel btf. Populate the value of this object.
The function ptr should be populated with the prog fds.
4. Call BPF_MAP_UPDATE with the object created in (3) as
the map value. The key is always "0".
During BPF_MAP_UPDATE, the code that saves the kernel-func-ptr's
args as an array of u64 is generated. BPF_MAP_UPDATE also allows
the specific struct_ops to do some final checks in "st_ops->init_member()"
(e.g. ensure all mandatory func ptrs are implemented).
If everything looks good, it will register this kernel struct
to the kernel subsystem. The map will not allow further update
from this point.
Unregister a struct_ops from the kernel subsystem:
BPF_MAP_DELETE with key "0".
Introspect a struct_ops:
BPF_MAP_LOOKUP_ELEM with key "0". The map value returned will
have the prog _id_ populated as the func ptr.
The map value state (enum bpf_struct_ops_state) will transit from:
INIT (map created) =>
INUSE (map updated, i.e. reg) =>
TOBEFREE (map value deleted, i.e. unreg)
The kernel subsystem needs to call bpf_struct_ops_get() and
bpf_struct_ops_put() to manage the "refcnt" in the
"struct bpf_struct_ops_XYZ". This patch uses a separate refcnt
for the purose of tracking the subsystem usage. Another approach
is to reuse the map->refcnt and then "show" (i.e. during map_lookup)
the subsystem's usage by doing map->refcnt - map->usercnt to filter out
the map-fd/pinned-map usage. However, that will also tie down the
future semantics of map->refcnt and map->usercnt.
The very first subsystem's refcnt (during reg()) holds one
count to map->refcnt. When the very last subsystem's refcnt
is gone, it will also release the map->refcnt. All bpf_prog will be
freed when the map->refcnt reaches 0 (i.e. during map_free()).
Here is how the bpftool map command will look like:
[root@arch-fb-vm1 bpf]# bpftool map show
6: struct_ops name dctcp flags 0x0
key 4B value 256B max_entries 1 memlock 4096B
btf_id 6
[root@arch-fb-vm1 bpf]# bpftool map dump id 6
[{
"value": {
"refcnt": {
"refs": {
"counter": 1
}
},
"state": 1,
"data": {
"list": {
"next": 0,
"prev": 0
},
"key": 0,
"flags": 2,
"init": 24,
"release": 0,
"ssthresh": 25,
"cong_avoid": 30,
"set_state": 27,
"cwnd_event": 28,
"in_ack_event": 26,
"undo_cwnd": 29,
"pkts_acked": 0,
"min_tso_segs": 0,
"sndbuf_expand": 0,
"cong_control": 0,
"get_info": 0,
"name": [98,112,102,95,100,99,116,99,112,0,0,0,0,0,0,0
],
"owner": 0
}
}
}
]
Misc Notes:
* bpf_struct_ops_map_sys_lookup_elem() is added for syscall lookup.
It does an inplace update on "*value" instead returning a pointer
to syscall.c. Otherwise, it needs a separate copy of "zero" value
for the BPF_STRUCT_OPS_STATE_INIT to avoid races.
* The bpf_struct_ops_map_delete_elem() is also called without
preempt_disable() from map_delete_elem(). It is because
the "->unreg()" may requires sleepable context, e.g.
the "tcp_unregister_congestion_control()".
* "const" is added to some of the existing "struct btf_func_model *"
function arg to avoid a compiler warning caused by this patch.
Signed-off-by: Martin KaFai Lau <kafai@fb.com>
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Acked-by: Andrii Nakryiko <andriin@fb.com>
Acked-by: Yonghong Song <yhs@fb.com>
Link: https://lore.kernel.org/bpf/20200109003505.3855919-1-kafai@fb.com
2020-01-09 00:35:05 +00:00
|
|
|
static int invoke_bpf(const struct btf_func_model *m, u8 **pprog,
|
2022-05-10 20:59:19 +00:00
|
|
|
struct bpf_tramp_links *tl, int stack_size,
|
2023-12-06 22:40:54 +00:00
|
|
|
int run_ctx_off, bool save_ret,
|
|
|
|
void *image, void *rw_image)
|
2019-11-14 18:57:04 +00:00
|
|
|
{
|
2020-03-04 19:18:48 +00:00
|
|
|
int i;
|
2019-11-14 18:57:04 +00:00
|
|
|
u8 *prog = *pprog;
|
|
|
|
|
2022-05-10 20:59:19 +00:00
|
|
|
for (i = 0; i < tl->nr_links; i++) {
|
|
|
|
if (invoke_bpf_prog(m, &prog, tl->links[i], stack_size,
|
2023-12-06 22:40:54 +00:00
|
|
|
run_ctx_off, save_ret, image, rw_image))
|
2020-03-04 19:18:49 +00:00
|
|
|
return -EINVAL;
|
|
|
|
}
|
|
|
|
*pprog = prog;
|
|
|
|
return 0;
|
|
|
|
}
|
|
|
|
|
|
|
|
static int invoke_bpf_mod_ret(const struct btf_func_model *m, u8 **pprog,
|
2022-05-10 20:59:19 +00:00
|
|
|
struct bpf_tramp_links *tl, int stack_size,
|
2023-12-06 22:40:54 +00:00
|
|
|
int run_ctx_off, u8 **branches,
|
|
|
|
void *image, void *rw_image)
|
2020-03-04 19:18:49 +00:00
|
|
|
{
|
|
|
|
u8 *prog = *pprog;
|
2021-06-23 11:25:04 +00:00
|
|
|
int i;
|
2020-03-04 19:18:49 +00:00
|
|
|
|
|
|
|
/* The first fmod_ret program will receive a garbage return value.
|
|
|
|
* Set this to 0 to avoid confusing the program.
|
|
|
|
*/
|
|
|
|
emit_mov_imm32(&prog, false, BPF_REG_0, 0);
|
|
|
|
emit_stx(&prog, BPF_DW, BPF_REG_FP, BPF_REG_0, -8);
|
2022-05-10 20:59:19 +00:00
|
|
|
for (i = 0; i < tl->nr_links; i++) {
|
2023-12-06 22:40:54 +00:00
|
|
|
if (invoke_bpf_prog(m, &prog, tl->links[i], stack_size, run_ctx_off, true,
|
|
|
|
image, rw_image))
|
2019-11-14 18:57:04 +00:00
|
|
|
return -EINVAL;
|
2020-03-04 19:18:49 +00:00
|
|
|
|
2020-03-11 00:39:06 +00:00
|
|
|
/* mod_ret prog stored return value into [rbp - 8]. Emit:
|
|
|
|
* if (*(u64 *)(rbp - 8) != 0)
|
2020-03-04 19:18:49 +00:00
|
|
|
* goto do_fexit;
|
|
|
|
*/
|
2020-03-11 00:39:06 +00:00
|
|
|
/* cmp QWORD PTR [rbp - 0x8], 0x0 */
|
|
|
|
EMIT4(0x48, 0x83, 0x7d, 0xf8); EMIT1(0x00);
|
2020-03-04 19:18:49 +00:00
|
|
|
|
|
|
|
/* Save the location of the branch and Generate 6 nops
|
|
|
|
* (4 bytes for an offset and 2 bytes for the jump) These nops
|
|
|
|
* are replaced with a conditional jump once do_fexit (i.e. the
|
|
|
|
* start of the fexit invocation) is finalized.
|
|
|
|
*/
|
|
|
|
branches[i] = prog;
|
|
|
|
emit_nops(&prog, 4 + 2);
|
2019-11-14 18:57:04 +00:00
|
|
|
}
|
2020-03-04 19:18:49 +00:00
|
|
|
|
2019-11-14 18:57:04 +00:00
|
|
|
*pprog = prog;
|
|
|
|
return 0;
|
|
|
|
}
|
|
|
|
|
|
|
|
/* Example:
|
|
|
|
* __be16 eth_type_trans(struct sk_buff *skb, struct net_device *dev);
|
|
|
|
* its 'struct btf_func_model' will be nr_args=2
|
|
|
|
* The assembly code when eth_type_trans is executing after trampoline:
|
|
|
|
*
|
|
|
|
* push rbp
|
|
|
|
* mov rbp, rsp
|
|
|
|
* sub rsp, 16 // space for skb and dev
|
|
|
|
* push rbx // temp regs to pass start time
|
|
|
|
* mov qword ptr [rbp - 16], rdi // save skb pointer to stack
|
|
|
|
* mov qword ptr [rbp - 8], rsi // save dev pointer to stack
|
|
|
|
* call __bpf_prog_enter // rcu_read_lock and preempt_disable
|
|
|
|
* mov rbx, rax // remember start time in bpf stats are enabled
|
|
|
|
* lea rdi, [rbp - 16] // R1==ctx of bpf prog
|
|
|
|
* call addr_of_jited_FENTRY_prog
|
|
|
|
* movabsq rdi, 64bit_addr_of_struct_bpf_prog // unused if bpf stats are off
|
|
|
|
* mov rsi, rbx // prog start time
|
|
|
|
* call __bpf_prog_exit // rcu_read_unlock, preempt_enable and stats math
|
|
|
|
* mov rdi, qword ptr [rbp - 16] // restore skb pointer from stack
|
|
|
|
* mov rsi, qword ptr [rbp - 8] // restore dev pointer from stack
|
|
|
|
* pop rbx
|
|
|
|
* leave
|
|
|
|
* ret
|
|
|
|
*
|
|
|
|
* eth_type_trans has 5 byte nop at the beginning. These 5 bytes will be
|
|
|
|
* replaced with 'call generated_bpf_trampoline'. When it returns
|
|
|
|
* eth_type_trans will continue executing with original skb and dev pointers.
|
|
|
|
*
|
|
|
|
* The assembly code when eth_type_trans is called from trampoline:
|
|
|
|
*
|
|
|
|
* push rbp
|
|
|
|
* mov rbp, rsp
|
|
|
|
* sub rsp, 24 // space for skb, dev, return value
|
|
|
|
* push rbx // temp regs to pass start time
|
|
|
|
* mov qword ptr [rbp - 24], rdi // save skb pointer to stack
|
|
|
|
* mov qword ptr [rbp - 16], rsi // save dev pointer to stack
|
|
|
|
* call __bpf_prog_enter // rcu_read_lock and preempt_disable
|
|
|
|
* mov rbx, rax // remember start time if bpf stats are enabled
|
|
|
|
* lea rdi, [rbp - 24] // R1==ctx of bpf prog
|
|
|
|
* call addr_of_jited_FENTRY_prog // bpf prog can access skb and dev
|
|
|
|
* movabsq rdi, 64bit_addr_of_struct_bpf_prog // unused if bpf stats are off
|
|
|
|
* mov rsi, rbx // prog start time
|
|
|
|
* call __bpf_prog_exit // rcu_read_unlock, preempt_enable and stats math
|
|
|
|
* mov rdi, qword ptr [rbp - 24] // restore skb pointer from stack
|
|
|
|
* mov rsi, qword ptr [rbp - 16] // restore dev pointer from stack
|
|
|
|
* call eth_type_trans+5 // execute body of eth_type_trans
|
|
|
|
* mov qword ptr [rbp - 8], rax // save return value
|
|
|
|
* call __bpf_prog_enter // rcu_read_lock and preempt_disable
|
|
|
|
* mov rbx, rax // remember start time in bpf stats are enabled
|
|
|
|
* lea rdi, [rbp - 24] // R1==ctx of bpf prog
|
|
|
|
* call addr_of_jited_FEXIT_prog // bpf prog can access skb, dev, return value
|
|
|
|
* movabsq rdi, 64bit_addr_of_struct_bpf_prog // unused if bpf stats are off
|
|
|
|
* mov rsi, rbx // prog start time
|
|
|
|
* call __bpf_prog_exit // rcu_read_unlock, preempt_enable and stats math
|
|
|
|
* mov rax, qword ptr [rbp - 8] // restore eth_type_trans's return value
|
|
|
|
* pop rbx
|
|
|
|
* leave
|
|
|
|
* add rsp, 8 // skip eth_type_trans's frame
|
|
|
|
* ret // return to its caller
|
|
|
|
*/
|
2023-12-06 22:40:54 +00:00
|
|
|
static int __arch_prepare_bpf_trampoline(struct bpf_tramp_image *im, void *rw_image,
|
|
|
|
void *rw_image_end, void *image,
|
2023-12-06 22:40:52 +00:00
|
|
|
const struct btf_func_model *m, u32 flags,
|
|
|
|
struct bpf_tramp_links *tlinks,
|
|
|
|
void *func_addr)
|
2019-11-14 18:57:04 +00:00
|
|
|
{
|
2023-01-05 03:50:26 +00:00
|
|
|
int i, ret, nr_regs = m->nr_args, stack_size = 0;
|
bpf, x86: allow function arguments up to 12 for TRACING
For now, the BPF program of type BPF_PROG_TYPE_TRACING can only be used
on the kernel functions whose arguments count less than or equal to 6, if
not considering '> 8 bytes' struct argument. This is not friendly at all,
as too many functions have arguments count more than 6.
According to the current kernel version, below is a statistics of the
function arguments count:
argument count | function count
7 | 704
8 | 270
9 | 84
10 | 47
11 | 47
12 | 27
13 | 22
14 | 5
15 | 0
16 | 1
Therefore, let's enhance it by increasing the function arguments count
allowed in arch_prepare_bpf_trampoline(), for now, only x86_64.
For the case that we don't need to call origin function, which means
without BPF_TRAMP_F_CALL_ORIG, we need only copy the function arguments
that stored in the frame of the caller to current frame. The 7th and later
arguments are stored in "$rbp + 0x18", and they will be copied to the
stack area following where register values are saved.
For the case with BPF_TRAMP_F_CALL_ORIG, we need prepare the arguments
in stack before call origin function, which means we need alloc extra
"8 * (arg_count - 6)" memory in the top of the stack. Note, there should
not be any data be pushed to the stack before calling the origin function.
So 'rbx' value will be stored on a stack position higher than where stack
arguments are stored for BPF_TRAMP_F_CALL_ORIG.
According to the research of Yonghong, struct members should be all in
register or all on the stack. Meanwhile, the compiler will pass the
argument on regs if the remaining regs can hold the argument. Therefore,
we need save the arguments in order. Otherwise, disorder of the args can
happen. For example:
struct foo_struct {
long a;
int b;
};
int foo(char, char, char, char, char, struct foo_struct,
char);
the arg1-5,arg7 will be passed by regs, and arg6 will by stack. Therefore,
we should save/restore the arguments in the same order with the
declaration of foo(). And the args used as ctx in stack will be like this:
reg_arg6 -- copy from regs
stack_arg2 -- copy from stack
stack_arg1
reg_arg5 -- copy from regs
reg_arg4
reg_arg3
reg_arg2
reg_arg1
We use EMIT3_off32() or EMIT4() for "lea" and "sub". The range of the
imm in "lea" and "sub" is [-128, 127] if EMIT4() is used. Therefore,
we use EMIT3_off32() instead if the imm out of the range.
It works well for the FENTRY/FEXIT/MODIFY_RETURN.
Signed-off-by: Menglong Dong <imagedong@tencent.com>
Acked-by: Yonghong Song <yhs@fb.com>
Link: https://lore.kernel.org/r/20230713040738.1789742-3-imagedong@tencent.com
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
2023-07-13 04:07:37 +00:00
|
|
|
int regs_off, nregs_off, ip_off, run_ctx_off, arg_stack_off, rbx_off;
|
2022-05-10 20:59:19 +00:00
|
|
|
struct bpf_tramp_links *fentry = &tlinks[BPF_TRAMP_FENTRY];
|
|
|
|
struct bpf_tramp_links *fexit = &tlinks[BPF_TRAMP_FEXIT];
|
|
|
|
struct bpf_tramp_links *fmod_ret = &tlinks[BPF_TRAMP_MODIFY_RETURN];
|
2022-09-26 15:33:37 +00:00
|
|
|
void *orig_call = func_addr;
|
2020-03-04 19:18:49 +00:00
|
|
|
u8 **branches = NULL;
|
2019-11-14 18:57:04 +00:00
|
|
|
u8 *prog;
|
2021-09-14 02:33:51 +00:00
|
|
|
bool save_ret;
|
2019-11-14 18:57:04 +00:00
|
|
|
|
2023-12-15 09:12:20 +00:00
|
|
|
/*
|
|
|
|
* F_INDIRECT is only compatible with F_RET_FENTRY_RET, it is
|
|
|
|
* explicitly incompatible with F_CALL_ORIG | F_SKIP_FRAME | F_IP_ARG
|
|
|
|
* because @func_addr.
|
|
|
|
*/
|
|
|
|
WARN_ON_ONCE((flags & BPF_TRAMP_F_INDIRECT) &&
|
|
|
|
(flags & ~(BPF_TRAMP_F_INDIRECT | BPF_TRAMP_F_RET_FENTRY_RET)));
|
|
|
|
|
2023-01-05 03:50:26 +00:00
|
|
|
/* extra registers for struct arguments */
|
2023-12-15 09:12:20 +00:00
|
|
|
for (i = 0; i < m->nr_args; i++) {
|
2022-08-31 15:26:52 +00:00
|
|
|
if (m->arg_flags[i] & BTF_FMODEL_STRUCT_ARG)
|
2023-01-05 03:50:26 +00:00
|
|
|
nr_regs += (m->arg_size[i] + 7) / 8 - 1;
|
2023-12-15 09:12:20 +00:00
|
|
|
}
|
2023-01-05 03:50:26 +00:00
|
|
|
|
bpf, x86: allow function arguments up to 12 for TRACING
For now, the BPF program of type BPF_PROG_TYPE_TRACING can only be used
on the kernel functions whose arguments count less than or equal to 6, if
not considering '> 8 bytes' struct argument. This is not friendly at all,
as too many functions have arguments count more than 6.
According to the current kernel version, below is a statistics of the
function arguments count:
argument count | function count
7 | 704
8 | 270
9 | 84
10 | 47
11 | 47
12 | 27
13 | 22
14 | 5
15 | 0
16 | 1
Therefore, let's enhance it by increasing the function arguments count
allowed in arch_prepare_bpf_trampoline(), for now, only x86_64.
For the case that we don't need to call origin function, which means
without BPF_TRAMP_F_CALL_ORIG, we need only copy the function arguments
that stored in the frame of the caller to current frame. The 7th and later
arguments are stored in "$rbp + 0x18", and they will be copied to the
stack area following where register values are saved.
For the case with BPF_TRAMP_F_CALL_ORIG, we need prepare the arguments
in stack before call origin function, which means we need alloc extra
"8 * (arg_count - 6)" memory in the top of the stack. Note, there should
not be any data be pushed to the stack before calling the origin function.
So 'rbx' value will be stored on a stack position higher than where stack
arguments are stored for BPF_TRAMP_F_CALL_ORIG.
According to the research of Yonghong, struct members should be all in
register or all on the stack. Meanwhile, the compiler will pass the
argument on regs if the remaining regs can hold the argument. Therefore,
we need save the arguments in order. Otherwise, disorder of the args can
happen. For example:
struct foo_struct {
long a;
int b;
};
int foo(char, char, char, char, char, struct foo_struct,
char);
the arg1-5,arg7 will be passed by regs, and arg6 will by stack. Therefore,
we should save/restore the arguments in the same order with the
declaration of foo(). And the args used as ctx in stack will be like this:
reg_arg6 -- copy from regs
stack_arg2 -- copy from stack
stack_arg1
reg_arg5 -- copy from regs
reg_arg4
reg_arg3
reg_arg2
reg_arg1
We use EMIT3_off32() or EMIT4() for "lea" and "sub". The range of the
imm in "lea" and "sub" is [-128, 127] if EMIT4() is used. Therefore,
we use EMIT3_off32() instead if the imm out of the range.
It works well for the FENTRY/FEXIT/MODIFY_RETURN.
Signed-off-by: Menglong Dong <imagedong@tencent.com>
Acked-by: Yonghong Song <yhs@fb.com>
Link: https://lore.kernel.org/r/20230713040738.1789742-3-imagedong@tencent.com
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
2023-07-13 04:07:37 +00:00
|
|
|
/* x86-64 supports up to MAX_BPF_FUNC_ARGS arguments. 1-6
|
|
|
|
* are passed through regs, the remains are through stack.
|
|
|
|
*/
|
|
|
|
if (nr_regs > MAX_BPF_FUNC_ARGS)
|
2022-08-31 15:26:52 +00:00
|
|
|
return -ENOTSUPP;
|
|
|
|
|
2021-12-08 19:32:43 +00:00
|
|
|
/* Generated trampoline stack layout:
|
|
|
|
*
|
|
|
|
* RBP + 8 [ return address ]
|
|
|
|
* RBP + 0 [ RBP ]
|
|
|
|
*
|
|
|
|
* RBP - 8 [ return value ] BPF_TRAMP_F_CALL_ORIG or
|
|
|
|
* BPF_TRAMP_F_RET_FENTRY_RET flags
|
|
|
|
*
|
|
|
|
* [ reg_argN ] always
|
|
|
|
* [ ... ]
|
|
|
|
* RBP - regs_off [ reg_arg1 ] program's ctx pointer
|
|
|
|
*
|
2023-01-05 03:50:26 +00:00
|
|
|
* RBP - nregs_off [ regs count ] always
|
2021-12-08 19:32:44 +00:00
|
|
|
*
|
2021-12-08 19:32:43 +00:00
|
|
|
* RBP - ip_off [ traced function ] BPF_TRAMP_F_IP_ARG flag
|
2022-05-10 20:59:20 +00:00
|
|
|
*
|
bpf, x86: allow function arguments up to 12 for TRACING
For now, the BPF program of type BPF_PROG_TYPE_TRACING can only be used
on the kernel functions whose arguments count less than or equal to 6, if
not considering '> 8 bytes' struct argument. This is not friendly at all,
as too many functions have arguments count more than 6.
According to the current kernel version, below is a statistics of the
function arguments count:
argument count | function count
7 | 704
8 | 270
9 | 84
10 | 47
11 | 47
12 | 27
13 | 22
14 | 5
15 | 0
16 | 1
Therefore, let's enhance it by increasing the function arguments count
allowed in arch_prepare_bpf_trampoline(), for now, only x86_64.
For the case that we don't need to call origin function, which means
without BPF_TRAMP_F_CALL_ORIG, we need only copy the function arguments
that stored in the frame of the caller to current frame. The 7th and later
arguments are stored in "$rbp + 0x18", and they will be copied to the
stack area following where register values are saved.
For the case with BPF_TRAMP_F_CALL_ORIG, we need prepare the arguments
in stack before call origin function, which means we need alloc extra
"8 * (arg_count - 6)" memory in the top of the stack. Note, there should
not be any data be pushed to the stack before calling the origin function.
So 'rbx' value will be stored on a stack position higher than where stack
arguments are stored for BPF_TRAMP_F_CALL_ORIG.
According to the research of Yonghong, struct members should be all in
register or all on the stack. Meanwhile, the compiler will pass the
argument on regs if the remaining regs can hold the argument. Therefore,
we need save the arguments in order. Otherwise, disorder of the args can
happen. For example:
struct foo_struct {
long a;
int b;
};
int foo(char, char, char, char, char, struct foo_struct,
char);
the arg1-5,arg7 will be passed by regs, and arg6 will by stack. Therefore,
we should save/restore the arguments in the same order with the
declaration of foo(). And the args used as ctx in stack will be like this:
reg_arg6 -- copy from regs
stack_arg2 -- copy from stack
stack_arg1
reg_arg5 -- copy from regs
reg_arg4
reg_arg3
reg_arg2
reg_arg1
We use EMIT3_off32() or EMIT4() for "lea" and "sub". The range of the
imm in "lea" and "sub" is [-128, 127] if EMIT4() is used. Therefore,
we use EMIT3_off32() instead if the imm out of the range.
It works well for the FENTRY/FEXIT/MODIFY_RETURN.
Signed-off-by: Menglong Dong <imagedong@tencent.com>
Acked-by: Yonghong Song <yhs@fb.com>
Link: https://lore.kernel.org/r/20230713040738.1789742-3-imagedong@tencent.com
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
2023-07-13 04:07:37 +00:00
|
|
|
* RBP - rbx_off [ rbx value ] always
|
|
|
|
*
|
2022-05-10 20:59:20 +00:00
|
|
|
* RBP - run_ctx_off [ bpf_tramp_run_ctx ]
|
bpf, x86: allow function arguments up to 12 for TRACING
For now, the BPF program of type BPF_PROG_TYPE_TRACING can only be used
on the kernel functions whose arguments count less than or equal to 6, if
not considering '> 8 bytes' struct argument. This is not friendly at all,
as too many functions have arguments count more than 6.
According to the current kernel version, below is a statistics of the
function arguments count:
argument count | function count
7 | 704
8 | 270
9 | 84
10 | 47
11 | 47
12 | 27
13 | 22
14 | 5
15 | 0
16 | 1
Therefore, let's enhance it by increasing the function arguments count
allowed in arch_prepare_bpf_trampoline(), for now, only x86_64.
For the case that we don't need to call origin function, which means
without BPF_TRAMP_F_CALL_ORIG, we need only copy the function arguments
that stored in the frame of the caller to current frame. The 7th and later
arguments are stored in "$rbp + 0x18", and they will be copied to the
stack area following where register values are saved.
For the case with BPF_TRAMP_F_CALL_ORIG, we need prepare the arguments
in stack before call origin function, which means we need alloc extra
"8 * (arg_count - 6)" memory in the top of the stack. Note, there should
not be any data be pushed to the stack before calling the origin function.
So 'rbx' value will be stored on a stack position higher than where stack
arguments are stored for BPF_TRAMP_F_CALL_ORIG.
According to the research of Yonghong, struct members should be all in
register or all on the stack. Meanwhile, the compiler will pass the
argument on regs if the remaining regs can hold the argument. Therefore,
we need save the arguments in order. Otherwise, disorder of the args can
happen. For example:
struct foo_struct {
long a;
int b;
};
int foo(char, char, char, char, char, struct foo_struct,
char);
the arg1-5,arg7 will be passed by regs, and arg6 will by stack. Therefore,
we should save/restore the arguments in the same order with the
declaration of foo(). And the args used as ctx in stack will be like this:
reg_arg6 -- copy from regs
stack_arg2 -- copy from stack
stack_arg1
reg_arg5 -- copy from regs
reg_arg4
reg_arg3
reg_arg2
reg_arg1
We use EMIT3_off32() or EMIT4() for "lea" and "sub". The range of the
imm in "lea" and "sub" is [-128, 127] if EMIT4() is used. Therefore,
we use EMIT3_off32() instead if the imm out of the range.
It works well for the FENTRY/FEXIT/MODIFY_RETURN.
Signed-off-by: Menglong Dong <imagedong@tencent.com>
Acked-by: Yonghong Song <yhs@fb.com>
Link: https://lore.kernel.org/r/20230713040738.1789742-3-imagedong@tencent.com
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
2023-07-13 04:07:37 +00:00
|
|
|
*
|
|
|
|
* [ stack_argN ] BPF_TRAMP_F_CALL_ORIG
|
|
|
|
* [ ... ]
|
|
|
|
* [ stack_arg2 ]
|
|
|
|
* RBP - arg_stack_off [ stack_arg1 ]
|
2023-09-12 15:04:41 +00:00
|
|
|
* RSP [ tail_call_cnt ] BPF_TRAMP_F_TAIL_CALL_CTX
|
2021-12-08 19:32:43 +00:00
|
|
|
*/
|
|
|
|
|
2021-09-14 02:33:51 +00:00
|
|
|
/* room for return value of orig_call or fentry prog */
|
|
|
|
save_ret = flags & (BPF_TRAMP_F_CALL_ORIG | BPF_TRAMP_F_RET_FENTRY_RET);
|
|
|
|
if (save_ret)
|
|
|
|
stack_size += 8;
|
2019-11-14 18:57:04 +00:00
|
|
|
|
2023-01-05 03:50:26 +00:00
|
|
|
stack_size += nr_regs * 8;
|
2021-12-08 19:32:43 +00:00
|
|
|
regs_off = stack_size;
|
|
|
|
|
2023-01-05 03:50:26 +00:00
|
|
|
/* regs count */
|
2021-12-08 19:32:44 +00:00
|
|
|
stack_size += 8;
|
2023-01-05 03:50:26 +00:00
|
|
|
nregs_off = stack_size;
|
2021-12-08 19:32:44 +00:00
|
|
|
|
2021-07-14 09:43:53 +00:00
|
|
|
if (flags & BPF_TRAMP_F_IP_ARG)
|
|
|
|
stack_size += 8; /* room for IP address argument */
|
|
|
|
|
2021-12-08 19:32:43 +00:00
|
|
|
ip_off = stack_size;
|
|
|
|
|
bpf, x86: allow function arguments up to 12 for TRACING
For now, the BPF program of type BPF_PROG_TYPE_TRACING can only be used
on the kernel functions whose arguments count less than or equal to 6, if
not considering '> 8 bytes' struct argument. This is not friendly at all,
as too many functions have arguments count more than 6.
According to the current kernel version, below is a statistics of the
function arguments count:
argument count | function count
7 | 704
8 | 270
9 | 84
10 | 47
11 | 47
12 | 27
13 | 22
14 | 5
15 | 0
16 | 1
Therefore, let's enhance it by increasing the function arguments count
allowed in arch_prepare_bpf_trampoline(), for now, only x86_64.
For the case that we don't need to call origin function, which means
without BPF_TRAMP_F_CALL_ORIG, we need only copy the function arguments
that stored in the frame of the caller to current frame. The 7th and later
arguments are stored in "$rbp + 0x18", and they will be copied to the
stack area following where register values are saved.
For the case with BPF_TRAMP_F_CALL_ORIG, we need prepare the arguments
in stack before call origin function, which means we need alloc extra
"8 * (arg_count - 6)" memory in the top of the stack. Note, there should
not be any data be pushed to the stack before calling the origin function.
So 'rbx' value will be stored on a stack position higher than where stack
arguments are stored for BPF_TRAMP_F_CALL_ORIG.
According to the research of Yonghong, struct members should be all in
register or all on the stack. Meanwhile, the compiler will pass the
argument on regs if the remaining regs can hold the argument. Therefore,
we need save the arguments in order. Otherwise, disorder of the args can
happen. For example:
struct foo_struct {
long a;
int b;
};
int foo(char, char, char, char, char, struct foo_struct,
char);
the arg1-5,arg7 will be passed by regs, and arg6 will by stack. Therefore,
we should save/restore the arguments in the same order with the
declaration of foo(). And the args used as ctx in stack will be like this:
reg_arg6 -- copy from regs
stack_arg2 -- copy from stack
stack_arg1
reg_arg5 -- copy from regs
reg_arg4
reg_arg3
reg_arg2
reg_arg1
We use EMIT3_off32() or EMIT4() for "lea" and "sub". The range of the
imm in "lea" and "sub" is [-128, 127] if EMIT4() is used. Therefore,
we use EMIT3_off32() instead if the imm out of the range.
It works well for the FENTRY/FEXIT/MODIFY_RETURN.
Signed-off-by: Menglong Dong <imagedong@tencent.com>
Acked-by: Yonghong Song <yhs@fb.com>
Link: https://lore.kernel.org/r/20230713040738.1789742-3-imagedong@tencent.com
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
2023-07-13 04:07:37 +00:00
|
|
|
stack_size += 8;
|
|
|
|
rbx_off = stack_size;
|
|
|
|
|
2022-05-10 20:59:20 +00:00
|
|
|
stack_size += (sizeof(struct bpf_tramp_run_ctx) + 7) & ~0x7;
|
|
|
|
run_ctx_off = stack_size;
|
|
|
|
|
bpf, x86: allow function arguments up to 12 for TRACING
For now, the BPF program of type BPF_PROG_TYPE_TRACING can only be used
on the kernel functions whose arguments count less than or equal to 6, if
not considering '> 8 bytes' struct argument. This is not friendly at all,
as too many functions have arguments count more than 6.
According to the current kernel version, below is a statistics of the
function arguments count:
argument count | function count
7 | 704
8 | 270
9 | 84
10 | 47
11 | 47
12 | 27
13 | 22
14 | 5
15 | 0
16 | 1
Therefore, let's enhance it by increasing the function arguments count
allowed in arch_prepare_bpf_trampoline(), for now, only x86_64.
For the case that we don't need to call origin function, which means
without BPF_TRAMP_F_CALL_ORIG, we need only copy the function arguments
that stored in the frame of the caller to current frame. The 7th and later
arguments are stored in "$rbp + 0x18", and they will be copied to the
stack area following where register values are saved.
For the case with BPF_TRAMP_F_CALL_ORIG, we need prepare the arguments
in stack before call origin function, which means we need alloc extra
"8 * (arg_count - 6)" memory in the top of the stack. Note, there should
not be any data be pushed to the stack before calling the origin function.
So 'rbx' value will be stored on a stack position higher than where stack
arguments are stored for BPF_TRAMP_F_CALL_ORIG.
According to the research of Yonghong, struct members should be all in
register or all on the stack. Meanwhile, the compiler will pass the
argument on regs if the remaining regs can hold the argument. Therefore,
we need save the arguments in order. Otherwise, disorder of the args can
happen. For example:
struct foo_struct {
long a;
int b;
};
int foo(char, char, char, char, char, struct foo_struct,
char);
the arg1-5,arg7 will be passed by regs, and arg6 will by stack. Therefore,
we should save/restore the arguments in the same order with the
declaration of foo(). And the args used as ctx in stack will be like this:
reg_arg6 -- copy from regs
stack_arg2 -- copy from stack
stack_arg1
reg_arg5 -- copy from regs
reg_arg4
reg_arg3
reg_arg2
reg_arg1
We use EMIT3_off32() or EMIT4() for "lea" and "sub". The range of the
imm in "lea" and "sub" is [-128, 127] if EMIT4() is used. Therefore,
we use EMIT3_off32() instead if the imm out of the range.
It works well for the FENTRY/FEXIT/MODIFY_RETURN.
Signed-off-by: Menglong Dong <imagedong@tencent.com>
Acked-by: Yonghong Song <yhs@fb.com>
Link: https://lore.kernel.org/r/20230713040738.1789742-3-imagedong@tencent.com
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
2023-07-13 04:07:37 +00:00
|
|
|
if (nr_regs > 6 && (flags & BPF_TRAMP_F_CALL_ORIG)) {
|
|
|
|
/* the space that used to pass arguments on-stack */
|
|
|
|
stack_size += (nr_regs - get_nr_used_regs(m)) * 8;
|
|
|
|
/* make sure the stack pointer is 16-byte aligned if we
|
|
|
|
* need pass arguments on stack, which means
|
|
|
|
* [stack_size + 8(rbp) + 8(rip) + 8(origin rip)]
|
|
|
|
* should be 16-byte aligned. Following code depend on
|
|
|
|
* that stack_size is already 8-byte aligned.
|
|
|
|
*/
|
|
|
|
stack_size += (stack_size % 16) ? 0 : 8;
|
|
|
|
}
|
|
|
|
|
|
|
|
arg_stack_off = stack_size;
|
|
|
|
|
2022-03-08 15:30:33 +00:00
|
|
|
if (flags & BPF_TRAMP_F_SKIP_FRAME) {
|
2019-11-14 18:57:04 +00:00
|
|
|
/* skip patched call instruction and point orig_call to actual
|
|
|
|
* body of the kernel function.
|
|
|
|
*/
|
2022-03-08 15:30:33 +00:00
|
|
|
if (is_endbr(*(u32 *)orig_call))
|
|
|
|
orig_call += ENDBR_INSN_SIZE;
|
2019-11-22 20:07:54 +00:00
|
|
|
orig_call += X86_PATCH_SIZE;
|
2022-03-08 15:30:33 +00:00
|
|
|
}
|
2019-11-14 18:57:04 +00:00
|
|
|
|
2023-12-06 22:40:54 +00:00
|
|
|
prog = rw_image;
|
2019-11-14 18:57:04 +00:00
|
|
|
|
2023-12-15 09:12:20 +00:00
|
|
|
if (flags & BPF_TRAMP_F_INDIRECT) {
|
|
|
|
/*
|
|
|
|
* Indirect call for bpf_struct_ops
|
|
|
|
*/
|
|
|
|
emit_cfi(&prog, cfi_get_func_hash(func_addr));
|
|
|
|
} else {
|
|
|
|
/*
|
|
|
|
* Direct-call fentry stub, as such it needs accounting for the
|
|
|
|
* __fentry__ call.
|
|
|
|
*/
|
2024-04-01 18:55:30 +00:00
|
|
|
x86_call_depth_emit_accounting(&prog, NULL, image);
|
2023-12-15 09:12:20 +00:00
|
|
|
}
|
2019-11-14 18:57:04 +00:00
|
|
|
EMIT1(0x55); /* push rbp */
|
|
|
|
EMIT3(0x48, 0x89, 0xE5); /* mov rbp, rsp */
|
2023-12-15 09:12:20 +00:00
|
|
|
if (!is_imm8(stack_size)) {
|
bpf, x86: allow function arguments up to 12 for TRACING
For now, the BPF program of type BPF_PROG_TYPE_TRACING can only be used
on the kernel functions whose arguments count less than or equal to 6, if
not considering '> 8 bytes' struct argument. This is not friendly at all,
as too many functions have arguments count more than 6.
According to the current kernel version, below is a statistics of the
function arguments count:
argument count | function count
7 | 704
8 | 270
9 | 84
10 | 47
11 | 47
12 | 27
13 | 22
14 | 5
15 | 0
16 | 1
Therefore, let's enhance it by increasing the function arguments count
allowed in arch_prepare_bpf_trampoline(), for now, only x86_64.
For the case that we don't need to call origin function, which means
without BPF_TRAMP_F_CALL_ORIG, we need only copy the function arguments
that stored in the frame of the caller to current frame. The 7th and later
arguments are stored in "$rbp + 0x18", and they will be copied to the
stack area following where register values are saved.
For the case with BPF_TRAMP_F_CALL_ORIG, we need prepare the arguments
in stack before call origin function, which means we need alloc extra
"8 * (arg_count - 6)" memory in the top of the stack. Note, there should
not be any data be pushed to the stack before calling the origin function.
So 'rbx' value will be stored on a stack position higher than where stack
arguments are stored for BPF_TRAMP_F_CALL_ORIG.
According to the research of Yonghong, struct members should be all in
register or all on the stack. Meanwhile, the compiler will pass the
argument on regs if the remaining regs can hold the argument. Therefore,
we need save the arguments in order. Otherwise, disorder of the args can
happen. For example:
struct foo_struct {
long a;
int b;
};
int foo(char, char, char, char, char, struct foo_struct,
char);
the arg1-5,arg7 will be passed by regs, and arg6 will by stack. Therefore,
we should save/restore the arguments in the same order with the
declaration of foo(). And the args used as ctx in stack will be like this:
reg_arg6 -- copy from regs
stack_arg2 -- copy from stack
stack_arg1
reg_arg5 -- copy from regs
reg_arg4
reg_arg3
reg_arg2
reg_arg1
We use EMIT3_off32() or EMIT4() for "lea" and "sub". The range of the
imm in "lea" and "sub" is [-128, 127] if EMIT4() is used. Therefore,
we use EMIT3_off32() instead if the imm out of the range.
It works well for the FENTRY/FEXIT/MODIFY_RETURN.
Signed-off-by: Menglong Dong <imagedong@tencent.com>
Acked-by: Yonghong Song <yhs@fb.com>
Link: https://lore.kernel.org/r/20230713040738.1789742-3-imagedong@tencent.com
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
2023-07-13 04:07:37 +00:00
|
|
|
/* sub rsp, stack_size */
|
|
|
|
EMIT3_off32(0x48, 0x81, 0xEC, stack_size);
|
2023-12-15 09:12:20 +00:00
|
|
|
} else {
|
bpf, x86: allow function arguments up to 12 for TRACING
For now, the BPF program of type BPF_PROG_TYPE_TRACING can only be used
on the kernel functions whose arguments count less than or equal to 6, if
not considering '> 8 bytes' struct argument. This is not friendly at all,
as too many functions have arguments count more than 6.
According to the current kernel version, below is a statistics of the
function arguments count:
argument count | function count
7 | 704
8 | 270
9 | 84
10 | 47
11 | 47
12 | 27
13 | 22
14 | 5
15 | 0
16 | 1
Therefore, let's enhance it by increasing the function arguments count
allowed in arch_prepare_bpf_trampoline(), for now, only x86_64.
For the case that we don't need to call origin function, which means
without BPF_TRAMP_F_CALL_ORIG, we need only copy the function arguments
that stored in the frame of the caller to current frame. The 7th and later
arguments are stored in "$rbp + 0x18", and they will be copied to the
stack area following where register values are saved.
For the case with BPF_TRAMP_F_CALL_ORIG, we need prepare the arguments
in stack before call origin function, which means we need alloc extra
"8 * (arg_count - 6)" memory in the top of the stack. Note, there should
not be any data be pushed to the stack before calling the origin function.
So 'rbx' value will be stored on a stack position higher than where stack
arguments are stored for BPF_TRAMP_F_CALL_ORIG.
According to the research of Yonghong, struct members should be all in
register or all on the stack. Meanwhile, the compiler will pass the
argument on regs if the remaining regs can hold the argument. Therefore,
we need save the arguments in order. Otherwise, disorder of the args can
happen. For example:
struct foo_struct {
long a;
int b;
};
int foo(char, char, char, char, char, struct foo_struct,
char);
the arg1-5,arg7 will be passed by regs, and arg6 will by stack. Therefore,
we should save/restore the arguments in the same order with the
declaration of foo(). And the args used as ctx in stack will be like this:
reg_arg6 -- copy from regs
stack_arg2 -- copy from stack
stack_arg1
reg_arg5 -- copy from regs
reg_arg4
reg_arg3
reg_arg2
reg_arg1
We use EMIT3_off32() or EMIT4() for "lea" and "sub". The range of the
imm in "lea" and "sub" is [-128, 127] if EMIT4() is used. Therefore,
we use EMIT3_off32() instead if the imm out of the range.
It works well for the FENTRY/FEXIT/MODIFY_RETURN.
Signed-off-by: Menglong Dong <imagedong@tencent.com>
Acked-by: Yonghong Song <yhs@fb.com>
Link: https://lore.kernel.org/r/20230713040738.1789742-3-imagedong@tencent.com
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
2023-07-13 04:07:37 +00:00
|
|
|
/* sub rsp, stack_size */
|
|
|
|
EMIT4(0x48, 0x83, 0xEC, stack_size);
|
2023-12-15 09:12:20 +00:00
|
|
|
}
|
2023-09-12 15:04:41 +00:00
|
|
|
if (flags & BPF_TRAMP_F_TAIL_CALL_CTX)
|
|
|
|
EMIT1(0x50); /* push rax */
|
bpf, x86: allow function arguments up to 12 for TRACING
For now, the BPF program of type BPF_PROG_TYPE_TRACING can only be used
on the kernel functions whose arguments count less than or equal to 6, if
not considering '> 8 bytes' struct argument. This is not friendly at all,
as too many functions have arguments count more than 6.
According to the current kernel version, below is a statistics of the
function arguments count:
argument count | function count
7 | 704
8 | 270
9 | 84
10 | 47
11 | 47
12 | 27
13 | 22
14 | 5
15 | 0
16 | 1
Therefore, let's enhance it by increasing the function arguments count
allowed in arch_prepare_bpf_trampoline(), for now, only x86_64.
For the case that we don't need to call origin function, which means
without BPF_TRAMP_F_CALL_ORIG, we need only copy the function arguments
that stored in the frame of the caller to current frame. The 7th and later
arguments are stored in "$rbp + 0x18", and they will be copied to the
stack area following where register values are saved.
For the case with BPF_TRAMP_F_CALL_ORIG, we need prepare the arguments
in stack before call origin function, which means we need alloc extra
"8 * (arg_count - 6)" memory in the top of the stack. Note, there should
not be any data be pushed to the stack before calling the origin function.
So 'rbx' value will be stored on a stack position higher than where stack
arguments are stored for BPF_TRAMP_F_CALL_ORIG.
According to the research of Yonghong, struct members should be all in
register or all on the stack. Meanwhile, the compiler will pass the
argument on regs if the remaining regs can hold the argument. Therefore,
we need save the arguments in order. Otherwise, disorder of the args can
happen. For example:
struct foo_struct {
long a;
int b;
};
int foo(char, char, char, char, char, struct foo_struct,
char);
the arg1-5,arg7 will be passed by regs, and arg6 will by stack. Therefore,
we should save/restore the arguments in the same order with the
declaration of foo(). And the args used as ctx in stack will be like this:
reg_arg6 -- copy from regs
stack_arg2 -- copy from stack
stack_arg1
reg_arg5 -- copy from regs
reg_arg4
reg_arg3
reg_arg2
reg_arg1
We use EMIT3_off32() or EMIT4() for "lea" and "sub". The range of the
imm in "lea" and "sub" is [-128, 127] if EMIT4() is used. Therefore,
we use EMIT3_off32() instead if the imm out of the range.
It works well for the FENTRY/FEXIT/MODIFY_RETURN.
Signed-off-by: Menglong Dong <imagedong@tencent.com>
Acked-by: Yonghong Song <yhs@fb.com>
Link: https://lore.kernel.org/r/20230713040738.1789742-3-imagedong@tencent.com
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
2023-07-13 04:07:37 +00:00
|
|
|
/* mov QWORD PTR [rbp - rbx_off], rbx */
|
|
|
|
emit_stx(&prog, BPF_DW, BPF_REG_FP, BPF_REG_6, -rbx_off);
|
2019-11-14 18:57:04 +00:00
|
|
|
|
2022-08-31 15:26:52 +00:00
|
|
|
/* Store number of argument registers of the traced function:
|
2023-01-05 03:50:26 +00:00
|
|
|
* mov rax, nr_regs
|
|
|
|
* mov QWORD PTR [rbp - nregs_off], rax
|
2021-12-08 19:32:44 +00:00
|
|
|
*/
|
2023-01-05 03:50:26 +00:00
|
|
|
emit_mov_imm64(&prog, BPF_REG_0, 0, (u32) nr_regs);
|
|
|
|
emit_stx(&prog, BPF_DW, BPF_REG_FP, BPF_REG_0, -nregs_off);
|
2021-12-08 19:32:44 +00:00
|
|
|
|
2021-07-14 09:43:53 +00:00
|
|
|
if (flags & BPF_TRAMP_F_IP_ARG) {
|
|
|
|
/* Store IP address of the traced function:
|
2022-09-26 15:33:37 +00:00
|
|
|
* movabsq rax, func_addr
|
2021-12-08 19:32:43 +00:00
|
|
|
* mov QWORD PTR [rbp - ip_off], rax
|
2021-07-14 09:43:53 +00:00
|
|
|
*/
|
2022-09-26 15:33:37 +00:00
|
|
|
emit_mov_imm64(&prog, BPF_REG_0, (long) func_addr >> 32, (u32) (long) func_addr);
|
2021-12-08 19:32:43 +00:00
|
|
|
emit_stx(&prog, BPF_DW, BPF_REG_FP, BPF_REG_0, -ip_off);
|
2021-07-14 09:43:53 +00:00
|
|
|
}
|
|
|
|
|
bpf, x86: allow function arguments up to 12 for TRACING
For now, the BPF program of type BPF_PROG_TYPE_TRACING can only be used
on the kernel functions whose arguments count less than or equal to 6, if
not considering '> 8 bytes' struct argument. This is not friendly at all,
as too many functions have arguments count more than 6.
According to the current kernel version, below is a statistics of the
function arguments count:
argument count | function count
7 | 704
8 | 270
9 | 84
10 | 47
11 | 47
12 | 27
13 | 22
14 | 5
15 | 0
16 | 1
Therefore, let's enhance it by increasing the function arguments count
allowed in arch_prepare_bpf_trampoline(), for now, only x86_64.
For the case that we don't need to call origin function, which means
without BPF_TRAMP_F_CALL_ORIG, we need only copy the function arguments
that stored in the frame of the caller to current frame. The 7th and later
arguments are stored in "$rbp + 0x18", and they will be copied to the
stack area following where register values are saved.
For the case with BPF_TRAMP_F_CALL_ORIG, we need prepare the arguments
in stack before call origin function, which means we need alloc extra
"8 * (arg_count - 6)" memory in the top of the stack. Note, there should
not be any data be pushed to the stack before calling the origin function.
So 'rbx' value will be stored on a stack position higher than where stack
arguments are stored for BPF_TRAMP_F_CALL_ORIG.
According to the research of Yonghong, struct members should be all in
register or all on the stack. Meanwhile, the compiler will pass the
argument on regs if the remaining regs can hold the argument. Therefore,
we need save the arguments in order. Otherwise, disorder of the args can
happen. For example:
struct foo_struct {
long a;
int b;
};
int foo(char, char, char, char, char, struct foo_struct,
char);
the arg1-5,arg7 will be passed by regs, and arg6 will by stack. Therefore,
we should save/restore the arguments in the same order with the
declaration of foo(). And the args used as ctx in stack will be like this:
reg_arg6 -- copy from regs
stack_arg2 -- copy from stack
stack_arg1
reg_arg5 -- copy from regs
reg_arg4
reg_arg3
reg_arg2
reg_arg1
We use EMIT3_off32() or EMIT4() for "lea" and "sub". The range of the
imm in "lea" and "sub" is [-128, 127] if EMIT4() is used. Therefore,
we use EMIT3_off32() instead if the imm out of the range.
It works well for the FENTRY/FEXIT/MODIFY_RETURN.
Signed-off-by: Menglong Dong <imagedong@tencent.com>
Acked-by: Yonghong Song <yhs@fb.com>
Link: https://lore.kernel.org/r/20230713040738.1789742-3-imagedong@tencent.com
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
2023-07-13 04:07:37 +00:00
|
|
|
save_args(m, &prog, regs_off, false);
|
2019-11-14 18:57:04 +00:00
|
|
|
|
2021-03-16 21:00:07 +00:00
|
|
|
if (flags & BPF_TRAMP_F_CALL_ORIG) {
|
|
|
|
/* arg1: mov rdi, im */
|
|
|
|
emit_mov_imm64(&prog, BPF_REG_1, (long) im >> 32, (u32) (long) im);
|
2023-12-06 22:40:54 +00:00
|
|
|
if (emit_rsb_call(&prog, __bpf_tramp_enter,
|
|
|
|
image + (prog - (u8 *)rw_image))) {
|
2021-03-16 21:00:07 +00:00
|
|
|
ret = -EINVAL;
|
|
|
|
goto cleanup;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2023-12-15 09:12:20 +00:00
|
|
|
if (fentry->nr_links) {
|
2022-05-10 20:59:20 +00:00
|
|
|
if (invoke_bpf(m, &prog, fentry, regs_off, run_ctx_off,
|
2023-12-06 22:40:54 +00:00
|
|
|
flags & BPF_TRAMP_F_RET_FENTRY_RET, image, rw_image))
|
2019-11-14 18:57:04 +00:00
|
|
|
return -EINVAL;
|
2023-12-15 09:12:20 +00:00
|
|
|
}
|
2019-11-14 18:57:04 +00:00
|
|
|
|
2022-05-10 20:59:19 +00:00
|
|
|
if (fmod_ret->nr_links) {
|
|
|
|
branches = kcalloc(fmod_ret->nr_links, sizeof(u8 *),
|
2020-03-04 19:18:49 +00:00
|
|
|
GFP_KERNEL);
|
|
|
|
if (!branches)
|
|
|
|
return -ENOMEM;
|
|
|
|
|
2021-12-08 19:32:43 +00:00
|
|
|
if (invoke_bpf_mod_ret(m, &prog, fmod_ret, regs_off,
|
2023-12-06 22:40:54 +00:00
|
|
|
run_ctx_off, branches, image, rw_image)) {
|
2020-03-04 19:18:49 +00:00
|
|
|
ret = -EINVAL;
|
|
|
|
goto cleanup;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2019-11-14 18:57:04 +00:00
|
|
|
if (flags & BPF_TRAMP_F_CALL_ORIG) {
|
bpf, x86: allow function arguments up to 12 for TRACING
For now, the BPF program of type BPF_PROG_TYPE_TRACING can only be used
on the kernel functions whose arguments count less than or equal to 6, if
not considering '> 8 bytes' struct argument. This is not friendly at all,
as too many functions have arguments count more than 6.
According to the current kernel version, below is a statistics of the
function arguments count:
argument count | function count
7 | 704
8 | 270
9 | 84
10 | 47
11 | 47
12 | 27
13 | 22
14 | 5
15 | 0
16 | 1
Therefore, let's enhance it by increasing the function arguments count
allowed in arch_prepare_bpf_trampoline(), for now, only x86_64.
For the case that we don't need to call origin function, which means
without BPF_TRAMP_F_CALL_ORIG, we need only copy the function arguments
that stored in the frame of the caller to current frame. The 7th and later
arguments are stored in "$rbp + 0x18", and they will be copied to the
stack area following where register values are saved.
For the case with BPF_TRAMP_F_CALL_ORIG, we need prepare the arguments
in stack before call origin function, which means we need alloc extra
"8 * (arg_count - 6)" memory in the top of the stack. Note, there should
not be any data be pushed to the stack before calling the origin function.
So 'rbx' value will be stored on a stack position higher than where stack
arguments are stored for BPF_TRAMP_F_CALL_ORIG.
According to the research of Yonghong, struct members should be all in
register or all on the stack. Meanwhile, the compiler will pass the
argument on regs if the remaining regs can hold the argument. Therefore,
we need save the arguments in order. Otherwise, disorder of the args can
happen. For example:
struct foo_struct {
long a;
int b;
};
int foo(char, char, char, char, char, struct foo_struct,
char);
the arg1-5,arg7 will be passed by regs, and arg6 will by stack. Therefore,
we should save/restore the arguments in the same order with the
declaration of foo(). And the args used as ctx in stack will be like this:
reg_arg6 -- copy from regs
stack_arg2 -- copy from stack
stack_arg1
reg_arg5 -- copy from regs
reg_arg4
reg_arg3
reg_arg2
reg_arg1
We use EMIT3_off32() or EMIT4() for "lea" and "sub". The range of the
imm in "lea" and "sub" is [-128, 127] if EMIT4() is used. Therefore,
we use EMIT3_off32() instead if the imm out of the range.
It works well for the FENTRY/FEXIT/MODIFY_RETURN.
Signed-off-by: Menglong Dong <imagedong@tencent.com>
Acked-by: Yonghong Song <yhs@fb.com>
Link: https://lore.kernel.org/r/20230713040738.1789742-3-imagedong@tencent.com
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
2023-07-13 04:07:37 +00:00
|
|
|
restore_regs(m, &prog, regs_off);
|
|
|
|
save_args(m, &prog, arg_stack_off, true);
|
2019-11-14 18:57:04 +00:00
|
|
|
|
2023-12-15 09:12:20 +00:00
|
|
|
if (flags & BPF_TRAMP_F_TAIL_CALL_CTX) {
|
2023-09-12 15:04:41 +00:00
|
|
|
/* Before calling the original function, restore the
|
|
|
|
* tail_call_cnt from stack to rax.
|
|
|
|
*/
|
|
|
|
RESTORE_TAIL_CALL_CNT(stack_size);
|
2023-12-15 09:12:20 +00:00
|
|
|
}
|
2023-09-12 15:04:41 +00:00
|
|
|
|
2022-07-20 00:21:25 +00:00
|
|
|
if (flags & BPF_TRAMP_F_ORIG_STACK) {
|
2023-09-12 15:04:41 +00:00
|
|
|
emit_ldx(&prog, BPF_DW, BPF_REG_6, BPF_REG_FP, 8);
|
|
|
|
EMIT2(0xff, 0xd3); /* call *rbx */
|
2022-07-20 00:21:25 +00:00
|
|
|
} else {
|
|
|
|
/* call original function */
|
2023-12-06 22:40:54 +00:00
|
|
|
if (emit_rsb_call(&prog, orig_call, image + (prog - (u8 *)rw_image))) {
|
2022-07-20 00:21:25 +00:00
|
|
|
ret = -EINVAL;
|
|
|
|
goto cleanup;
|
|
|
|
}
|
2020-03-04 19:18:49 +00:00
|
|
|
}
|
2019-11-14 18:57:04 +00:00
|
|
|
/* remember return value in a stack for bpf prog to access */
|
|
|
|
emit_stx(&prog, BPF_DW, BPF_REG_FP, BPF_REG_0, -8);
|
2023-12-06 22:40:54 +00:00
|
|
|
im->ip_after_call = image + (prog - (u8 *)rw_image);
|
2024-01-04 14:22:23 +00:00
|
|
|
emit_nops(&prog, X86_PATCH_SIZE);
|
2019-11-14 18:57:04 +00:00
|
|
|
}
|
|
|
|
|
2022-05-10 20:59:19 +00:00
|
|
|
if (fmod_ret->nr_links) {
|
2020-03-04 19:18:49 +00:00
|
|
|
/* From Intel 64 and IA-32 Architectures Optimization
|
|
|
|
* Reference Manual, 3.4.1.4 Code Alignment, Assembly/Compiler
|
|
|
|
* Coding Rule 11: All branch targets should be 16-byte
|
|
|
|
* aligned.
|
|
|
|
*/
|
|
|
|
emit_align(&prog, 16);
|
|
|
|
/* Update the branches saved in invoke_bpf_mod_ret with the
|
|
|
|
* aligned address of do_fexit.
|
|
|
|
*/
|
2023-12-15 09:12:20 +00:00
|
|
|
for (i = 0; i < fmod_ret->nr_links; i++) {
|
2023-12-06 22:40:54 +00:00
|
|
|
emit_cond_near_jump(&branches[i], image + (prog - (u8 *)rw_image),
|
|
|
|
image + (branches[i] - (u8 *)rw_image), X86_JNE);
|
2023-12-15 09:12:20 +00:00
|
|
|
}
|
2020-03-04 19:18:49 +00:00
|
|
|
}
|
|
|
|
|
2023-12-15 09:12:20 +00:00
|
|
|
if (fexit->nr_links) {
|
2023-12-06 22:40:54 +00:00
|
|
|
if (invoke_bpf(m, &prog, fexit, regs_off, run_ctx_off,
|
|
|
|
false, image, rw_image)) {
|
2020-03-04 19:18:49 +00:00
|
|
|
ret = -EINVAL;
|
|
|
|
goto cleanup;
|
|
|
|
}
|
2023-12-15 09:12:20 +00:00
|
|
|
}
|
2019-11-14 18:57:04 +00:00
|
|
|
|
|
|
|
if (flags & BPF_TRAMP_F_RESTORE_REGS)
|
bpf, x86: allow function arguments up to 12 for TRACING
For now, the BPF program of type BPF_PROG_TYPE_TRACING can only be used
on the kernel functions whose arguments count less than or equal to 6, if
not considering '> 8 bytes' struct argument. This is not friendly at all,
as too many functions have arguments count more than 6.
According to the current kernel version, below is a statistics of the
function arguments count:
argument count | function count
7 | 704
8 | 270
9 | 84
10 | 47
11 | 47
12 | 27
13 | 22
14 | 5
15 | 0
16 | 1
Therefore, let's enhance it by increasing the function arguments count
allowed in arch_prepare_bpf_trampoline(), for now, only x86_64.
For the case that we don't need to call origin function, which means
without BPF_TRAMP_F_CALL_ORIG, we need only copy the function arguments
that stored in the frame of the caller to current frame. The 7th and later
arguments are stored in "$rbp + 0x18", and they will be copied to the
stack area following where register values are saved.
For the case with BPF_TRAMP_F_CALL_ORIG, we need prepare the arguments
in stack before call origin function, which means we need alloc extra
"8 * (arg_count - 6)" memory in the top of the stack. Note, there should
not be any data be pushed to the stack before calling the origin function.
So 'rbx' value will be stored on a stack position higher than where stack
arguments are stored for BPF_TRAMP_F_CALL_ORIG.
According to the research of Yonghong, struct members should be all in
register or all on the stack. Meanwhile, the compiler will pass the
argument on regs if the remaining regs can hold the argument. Therefore,
we need save the arguments in order. Otherwise, disorder of the args can
happen. For example:
struct foo_struct {
long a;
int b;
};
int foo(char, char, char, char, char, struct foo_struct,
char);
the arg1-5,arg7 will be passed by regs, and arg6 will by stack. Therefore,
we should save/restore the arguments in the same order with the
declaration of foo(). And the args used as ctx in stack will be like this:
reg_arg6 -- copy from regs
stack_arg2 -- copy from stack
stack_arg1
reg_arg5 -- copy from regs
reg_arg4
reg_arg3
reg_arg2
reg_arg1
We use EMIT3_off32() or EMIT4() for "lea" and "sub". The range of the
imm in "lea" and "sub" is [-128, 127] if EMIT4() is used. Therefore,
we use EMIT3_off32() instead if the imm out of the range.
It works well for the FENTRY/FEXIT/MODIFY_RETURN.
Signed-off-by: Menglong Dong <imagedong@tencent.com>
Acked-by: Yonghong Song <yhs@fb.com>
Link: https://lore.kernel.org/r/20230713040738.1789742-3-imagedong@tencent.com
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
2023-07-13 04:07:37 +00:00
|
|
|
restore_regs(m, &prog, regs_off);
|
2019-11-14 18:57:04 +00:00
|
|
|
|
2020-03-04 19:18:49 +00:00
|
|
|
/* This needs to be done regardless. If there were fmod_ret programs,
|
|
|
|
* the return value is only updated on the stack and still needs to be
|
|
|
|
* restored to R0.
|
|
|
|
*/
|
2021-03-16 21:00:07 +00:00
|
|
|
if (flags & BPF_TRAMP_F_CALL_ORIG) {
|
2023-12-06 22:40:54 +00:00
|
|
|
im->ip_epilogue = image + (prog - (u8 *)rw_image);
|
2021-03-16 21:00:07 +00:00
|
|
|
/* arg1: mov rdi, im */
|
|
|
|
emit_mov_imm64(&prog, BPF_REG_1, (long) im >> 32, (u32) (long) im);
|
2023-12-06 22:40:54 +00:00
|
|
|
if (emit_rsb_call(&prog, __bpf_tramp_exit, image + (prog - (u8 *)rw_image))) {
|
2021-03-16 21:00:07 +00:00
|
|
|
ret = -EINVAL;
|
|
|
|
goto cleanup;
|
|
|
|
}
|
2023-12-15 09:12:20 +00:00
|
|
|
} else if (flags & BPF_TRAMP_F_TAIL_CALL_CTX) {
|
2023-09-12 15:04:41 +00:00
|
|
|
/* Before running the original function, restore the
|
|
|
|
* tail_call_cnt from stack to rax.
|
|
|
|
*/
|
|
|
|
RESTORE_TAIL_CALL_CNT(stack_size);
|
2023-12-15 09:12:20 +00:00
|
|
|
}
|
2023-09-12 15:04:41 +00:00
|
|
|
|
2021-09-14 02:33:51 +00:00
|
|
|
/* restore return value of orig_call or fentry prog back into RAX */
|
|
|
|
if (save_ret)
|
|
|
|
emit_ldx(&prog, BPF_DW, BPF_REG_0, BPF_REG_FP, -8);
|
2019-11-14 18:57:04 +00:00
|
|
|
|
bpf, x86: allow function arguments up to 12 for TRACING
For now, the BPF program of type BPF_PROG_TYPE_TRACING can only be used
on the kernel functions whose arguments count less than or equal to 6, if
not considering '> 8 bytes' struct argument. This is not friendly at all,
as too many functions have arguments count more than 6.
According to the current kernel version, below is a statistics of the
function arguments count:
argument count | function count
7 | 704
8 | 270
9 | 84
10 | 47
11 | 47
12 | 27
13 | 22
14 | 5
15 | 0
16 | 1
Therefore, let's enhance it by increasing the function arguments count
allowed in arch_prepare_bpf_trampoline(), for now, only x86_64.
For the case that we don't need to call origin function, which means
without BPF_TRAMP_F_CALL_ORIG, we need only copy the function arguments
that stored in the frame of the caller to current frame. The 7th and later
arguments are stored in "$rbp + 0x18", and they will be copied to the
stack area following where register values are saved.
For the case with BPF_TRAMP_F_CALL_ORIG, we need prepare the arguments
in stack before call origin function, which means we need alloc extra
"8 * (arg_count - 6)" memory in the top of the stack. Note, there should
not be any data be pushed to the stack before calling the origin function.
So 'rbx' value will be stored on a stack position higher than where stack
arguments are stored for BPF_TRAMP_F_CALL_ORIG.
According to the research of Yonghong, struct members should be all in
register or all on the stack. Meanwhile, the compiler will pass the
argument on regs if the remaining regs can hold the argument. Therefore,
we need save the arguments in order. Otherwise, disorder of the args can
happen. For example:
struct foo_struct {
long a;
int b;
};
int foo(char, char, char, char, char, struct foo_struct,
char);
the arg1-5,arg7 will be passed by regs, and arg6 will by stack. Therefore,
we should save/restore the arguments in the same order with the
declaration of foo(). And the args used as ctx in stack will be like this:
reg_arg6 -- copy from regs
stack_arg2 -- copy from stack
stack_arg1
reg_arg5 -- copy from regs
reg_arg4
reg_arg3
reg_arg2
reg_arg1
We use EMIT3_off32() or EMIT4() for "lea" and "sub". The range of the
imm in "lea" and "sub" is [-128, 127] if EMIT4() is used. Therefore,
we use EMIT3_off32() instead if the imm out of the range.
It works well for the FENTRY/FEXIT/MODIFY_RETURN.
Signed-off-by: Menglong Dong <imagedong@tencent.com>
Acked-by: Yonghong Song <yhs@fb.com>
Link: https://lore.kernel.org/r/20230713040738.1789742-3-imagedong@tencent.com
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
2023-07-13 04:07:37 +00:00
|
|
|
emit_ldx(&prog, BPF_DW, BPF_REG_6, BPF_REG_FP, -rbx_off);
|
2019-11-14 18:57:04 +00:00
|
|
|
EMIT1(0xC9); /* leave */
|
2023-12-15 09:12:20 +00:00
|
|
|
if (flags & BPF_TRAMP_F_SKIP_FRAME) {
|
2019-11-14 18:57:04 +00:00
|
|
|
/* skip our return address and return to parent */
|
|
|
|
EMIT4(0x48, 0x83, 0xC4, 8); /* add rsp, 8 */
|
2023-12-15 09:12:20 +00:00
|
|
|
}
|
2023-12-06 22:40:54 +00:00
|
|
|
emit_return(&prog, image + (prog - (u8 *)rw_image));
|
bpf: Introduce BPF_MAP_TYPE_STRUCT_OPS
The patch introduces BPF_MAP_TYPE_STRUCT_OPS. The map value
is a kernel struct with its func ptr implemented in bpf prog.
This new map is the interface to register/unregister/introspect
a bpf implemented kernel struct.
The kernel struct is actually embedded inside another new struct
(or called the "value" struct in the code). For example,
"struct tcp_congestion_ops" is embbeded in:
struct bpf_struct_ops_tcp_congestion_ops {
refcount_t refcnt;
enum bpf_struct_ops_state state;
struct tcp_congestion_ops data; /* <-- kernel subsystem struct here */
}
The map value is "struct bpf_struct_ops_tcp_congestion_ops".
The "bpftool map dump" will then be able to show the
state ("inuse"/"tobefree") and the number of subsystem's refcnt (e.g.
number of tcp_sock in the tcp_congestion_ops case). This "value" struct
is created automatically by a macro. Having a separate "value" struct
will also make extending "struct bpf_struct_ops_XYZ" easier (e.g. adding
"void (*init)(void)" to "struct bpf_struct_ops_XYZ" to do some
initialization works before registering the struct_ops to the kernel
subsystem). The libbpf will take care of finding and populating the
"struct bpf_struct_ops_XYZ" from "struct XYZ".
Register a struct_ops to a kernel subsystem:
1. Load all needed BPF_PROG_TYPE_STRUCT_OPS prog(s)
2. Create a BPF_MAP_TYPE_STRUCT_OPS with attr->btf_vmlinux_value_type_id
set to the btf id "struct bpf_struct_ops_tcp_congestion_ops" of the
running kernel.
Instead of reusing the attr->btf_value_type_id,
btf_vmlinux_value_type_id s added such that attr->btf_fd can still be
used as the "user" btf which could store other useful sysadmin/debug
info that may be introduced in the furture,
e.g. creation-date/compiler-details/map-creator...etc.
3. Create a "struct bpf_struct_ops_tcp_congestion_ops" object as described
in the running kernel btf. Populate the value of this object.
The function ptr should be populated with the prog fds.
4. Call BPF_MAP_UPDATE with the object created in (3) as
the map value. The key is always "0".
During BPF_MAP_UPDATE, the code that saves the kernel-func-ptr's
args as an array of u64 is generated. BPF_MAP_UPDATE also allows
the specific struct_ops to do some final checks in "st_ops->init_member()"
(e.g. ensure all mandatory func ptrs are implemented).
If everything looks good, it will register this kernel struct
to the kernel subsystem. The map will not allow further update
from this point.
Unregister a struct_ops from the kernel subsystem:
BPF_MAP_DELETE with key "0".
Introspect a struct_ops:
BPF_MAP_LOOKUP_ELEM with key "0". The map value returned will
have the prog _id_ populated as the func ptr.
The map value state (enum bpf_struct_ops_state) will transit from:
INIT (map created) =>
INUSE (map updated, i.e. reg) =>
TOBEFREE (map value deleted, i.e. unreg)
The kernel subsystem needs to call bpf_struct_ops_get() and
bpf_struct_ops_put() to manage the "refcnt" in the
"struct bpf_struct_ops_XYZ". This patch uses a separate refcnt
for the purose of tracking the subsystem usage. Another approach
is to reuse the map->refcnt and then "show" (i.e. during map_lookup)
the subsystem's usage by doing map->refcnt - map->usercnt to filter out
the map-fd/pinned-map usage. However, that will also tie down the
future semantics of map->refcnt and map->usercnt.
The very first subsystem's refcnt (during reg()) holds one
count to map->refcnt. When the very last subsystem's refcnt
is gone, it will also release the map->refcnt. All bpf_prog will be
freed when the map->refcnt reaches 0 (i.e. during map_free()).
Here is how the bpftool map command will look like:
[root@arch-fb-vm1 bpf]# bpftool map show
6: struct_ops name dctcp flags 0x0
key 4B value 256B max_entries 1 memlock 4096B
btf_id 6
[root@arch-fb-vm1 bpf]# bpftool map dump id 6
[{
"value": {
"refcnt": {
"refs": {
"counter": 1
}
},
"state": 1,
"data": {
"list": {
"next": 0,
"prev": 0
},
"key": 0,
"flags": 2,
"init": 24,
"release": 0,
"ssthresh": 25,
"cong_avoid": 30,
"set_state": 27,
"cwnd_event": 28,
"in_ack_event": 26,
"undo_cwnd": 29,
"pkts_acked": 0,
"min_tso_segs": 0,
"sndbuf_expand": 0,
"cong_control": 0,
"get_info": 0,
"name": [98,112,102,95,100,99,116,99,112,0,0,0,0,0,0,0
],
"owner": 0
}
}
}
]
Misc Notes:
* bpf_struct_ops_map_sys_lookup_elem() is added for syscall lookup.
It does an inplace update on "*value" instead returning a pointer
to syscall.c. Otherwise, it needs a separate copy of "zero" value
for the BPF_STRUCT_OPS_STATE_INIT to avoid races.
* The bpf_struct_ops_map_delete_elem() is also called without
preempt_disable() from map_delete_elem(). It is because
the "->unreg()" may requires sleepable context, e.g.
the "tcp_unregister_congestion_control()".
* "const" is added to some of the existing "struct btf_func_model *"
function arg to avoid a compiler warning caused by this patch.
Signed-off-by: Martin KaFai Lau <kafai@fb.com>
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Acked-by: Andrii Nakryiko <andriin@fb.com>
Acked-by: Yonghong Song <yhs@fb.com>
Link: https://lore.kernel.org/bpf/20200109003505.3855919-1-kafai@fb.com
2020-01-09 00:35:05 +00:00
|
|
|
/* Make sure the trampoline generation logic doesn't overflow */
|
2023-12-06 22:40:54 +00:00
|
|
|
if (WARN_ON_ONCE(prog > (u8 *)rw_image_end - BPF_INSN_SAFETY)) {
|
2020-03-04 19:18:49 +00:00
|
|
|
ret = -EFAULT;
|
|
|
|
goto cleanup;
|
|
|
|
}
|
2023-12-06 22:40:54 +00:00
|
|
|
ret = prog - (u8 *)rw_image + BPF_INSN_SAFETY;
|
2020-03-04 19:18:49 +00:00
|
|
|
|
|
|
|
cleanup:
|
|
|
|
kfree(branches);
|
|
|
|
return ret;
|
2019-11-14 18:57:04 +00:00
|
|
|
}
|
|
|
|
|
2023-12-06 22:40:54 +00:00
|
|
|
void *arch_alloc_bpf_trampoline(unsigned int size)
|
|
|
|
{
|
|
|
|
return bpf_prog_pack_alloc(size, jit_fill_hole);
|
|
|
|
}
|
|
|
|
|
|
|
|
void arch_free_bpf_trampoline(void *image, unsigned int size)
|
|
|
|
{
|
|
|
|
bpf_prog_pack_free(image, size);
|
|
|
|
}
|
|
|
|
|
|
|
|
void arch_protect_bpf_trampoline(void *image, unsigned int size)
|
|
|
|
{
|
|
|
|
}
|
|
|
|
|
|
|
|
void arch_unprotect_bpf_trampoline(void *image, unsigned int size)
|
|
|
|
{
|
|
|
|
}
|
|
|
|
|
2023-12-06 22:40:52 +00:00
|
|
|
int arch_prepare_bpf_trampoline(struct bpf_tramp_image *im, void *image, void *image_end,
|
|
|
|
const struct btf_func_model *m, u32 flags,
|
|
|
|
struct bpf_tramp_links *tlinks,
|
|
|
|
void *func_addr)
|
|
|
|
{
|
2023-12-06 22:40:54 +00:00
|
|
|
void *rw_image, *tmp;
|
|
|
|
int ret;
|
|
|
|
u32 size = image_end - image;
|
|
|
|
|
|
|
|
/* rw_image doesn't need to be in module memory range, so we can
|
|
|
|
* use kvmalloc.
|
|
|
|
*/
|
|
|
|
rw_image = kvmalloc(size, GFP_KERNEL);
|
|
|
|
if (!rw_image)
|
|
|
|
return -ENOMEM;
|
|
|
|
|
|
|
|
ret = __arch_prepare_bpf_trampoline(im, rw_image, rw_image + size, image, m,
|
|
|
|
flags, tlinks, func_addr);
|
|
|
|
if (ret < 0)
|
|
|
|
goto out;
|
|
|
|
|
|
|
|
tmp = bpf_arch_text_copy(image, rw_image, size);
|
|
|
|
if (IS_ERR(tmp))
|
|
|
|
ret = PTR_ERR(tmp);
|
|
|
|
out:
|
|
|
|
kvfree(rw_image);
|
|
|
|
return ret;
|
2023-12-06 22:40:52 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
int arch_bpf_trampoline_size(const struct btf_func_model *m, u32 flags,
|
|
|
|
struct bpf_tramp_links *tlinks, void *func_addr)
|
|
|
|
{
|
|
|
|
struct bpf_tramp_image im;
|
|
|
|
void *image;
|
|
|
|
int ret;
|
|
|
|
|
|
|
|
/* Allocate a temporary buffer for __arch_prepare_bpf_trampoline().
|
|
|
|
* This will NOT cause fragmentation in direct map, as we do not
|
|
|
|
* call set_memory_*() on this buffer.
|
|
|
|
*
|
|
|
|
* We cannot use kvmalloc here, because we need image to be in
|
|
|
|
* module memory range.
|
|
|
|
*/
|
|
|
|
image = bpf_jit_alloc_exec(PAGE_SIZE);
|
|
|
|
if (!image)
|
|
|
|
return -ENOMEM;
|
|
|
|
|
2023-12-06 22:40:54 +00:00
|
|
|
ret = __arch_prepare_bpf_trampoline(&im, image, image + PAGE_SIZE, image,
|
|
|
|
m, flags, tlinks, func_addr);
|
2023-12-06 22:40:52 +00:00
|
|
|
bpf_jit_free_exec(image);
|
|
|
|
return ret;
|
|
|
|
}
|
|
|
|
|
2022-09-26 18:47:38 +00:00
|
|
|
static int emit_bpf_dispatcher(u8 **pprog, int a, int b, s64 *progs, u8 *image, u8 *buf)
|
2019-12-13 17:51:08 +00:00
|
|
|
{
|
2020-03-04 19:18:48 +00:00
|
|
|
u8 *jg_reloc, *prog = *pprog;
|
2021-06-23 11:25:04 +00:00
|
|
|
int pivot, err, jg_bytes = 1;
|
2019-12-13 17:51:08 +00:00
|
|
|
s64 jg_offset;
|
|
|
|
|
|
|
|
if (a == b) {
|
|
|
|
/* Leaf node of recursion, i.e. not a range of indices
|
|
|
|
* anymore.
|
|
|
|
*/
|
|
|
|
EMIT1(add_1mod(0x48, BPF_REG_3)); /* cmp rdx,func */
|
|
|
|
if (!is_simm32(progs[a]))
|
|
|
|
return -1;
|
|
|
|
EMIT2_off32(0x81, add_1reg(0xF8, BPF_REG_3),
|
|
|
|
progs[a]);
|
|
|
|
err = emit_cond_near_jump(&prog, /* je func */
|
2022-09-26 18:47:38 +00:00
|
|
|
(void *)progs[a], image + (prog - buf),
|
2019-12-13 17:51:08 +00:00
|
|
|
X86_JE);
|
|
|
|
if (err)
|
|
|
|
return err;
|
|
|
|
|
2022-09-26 18:47:38 +00:00
|
|
|
emit_indirect_jump(&prog, 2 /* rdx */, image + (prog - buf));
|
2019-12-13 17:51:08 +00:00
|
|
|
|
|
|
|
*pprog = prog;
|
|
|
|
return 0;
|
|
|
|
}
|
|
|
|
|
|
|
|
/* Not a leaf node, so we pivot, and recursively descend into
|
|
|
|
* the lower and upper ranges.
|
|
|
|
*/
|
|
|
|
pivot = (b - a) / 2;
|
|
|
|
EMIT1(add_1mod(0x48, BPF_REG_3)); /* cmp rdx,func */
|
|
|
|
if (!is_simm32(progs[a + pivot]))
|
|
|
|
return -1;
|
|
|
|
EMIT2_off32(0x81, add_1reg(0xF8, BPF_REG_3), progs[a + pivot]);
|
|
|
|
|
|
|
|
if (pivot > 2) { /* jg upper_part */
|
|
|
|
/* Require near jump. */
|
|
|
|
jg_bytes = 4;
|
|
|
|
EMIT2_off32(0x0F, X86_JG + 0x10, 0);
|
|
|
|
} else {
|
|
|
|
EMIT2(X86_JG, 0);
|
|
|
|
}
|
|
|
|
jg_reloc = prog;
|
|
|
|
|
|
|
|
err = emit_bpf_dispatcher(&prog, a, a + pivot, /* emit lower_part */
|
2022-09-26 18:47:38 +00:00
|
|
|
progs, image, buf);
|
2019-12-13 17:51:08 +00:00
|
|
|
if (err)
|
|
|
|
return err;
|
|
|
|
|
2019-12-13 17:51:12 +00:00
|
|
|
/* From Intel 64 and IA-32 Architectures Optimization
|
|
|
|
* Reference Manual, 3.4.1.4 Code Alignment, Assembly/Compiler
|
|
|
|
* Coding Rule 11: All branch targets should be 16-byte
|
|
|
|
* aligned.
|
|
|
|
*/
|
2020-03-04 19:18:48 +00:00
|
|
|
emit_align(&prog, 16);
|
2019-12-13 17:51:08 +00:00
|
|
|
jg_offset = prog - jg_reloc;
|
|
|
|
emit_code(jg_reloc - jg_bytes, jg_offset, jg_bytes);
|
|
|
|
|
|
|
|
err = emit_bpf_dispatcher(&prog, a + pivot + 1, /* emit upper_part */
|
2022-09-26 18:47:38 +00:00
|
|
|
b, progs, image, buf);
|
2019-12-13 17:51:08 +00:00
|
|
|
if (err)
|
|
|
|
return err;
|
|
|
|
|
|
|
|
*pprog = prog;
|
|
|
|
return 0;
|
|
|
|
}
|
|
|
|
|
|
|
|
static int cmp_ips(const void *a, const void *b)
|
|
|
|
{
|
|
|
|
const s64 *ipa = a;
|
|
|
|
const s64 *ipb = b;
|
|
|
|
|
|
|
|
if (*ipa > *ipb)
|
|
|
|
return 1;
|
|
|
|
if (*ipa < *ipb)
|
|
|
|
return -1;
|
|
|
|
return 0;
|
|
|
|
}
|
|
|
|
|
2022-09-26 18:47:38 +00:00
|
|
|
int arch_prepare_bpf_dispatcher(void *image, void *buf, s64 *funcs, int num_funcs)
|
2019-12-13 17:51:08 +00:00
|
|
|
{
|
2022-09-26 18:47:38 +00:00
|
|
|
u8 *prog = buf;
|
2019-12-13 17:51:08 +00:00
|
|
|
|
|
|
|
sort(funcs, num_funcs, sizeof(funcs[0]), cmp_ips, NULL);
|
2022-09-26 18:47:38 +00:00
|
|
|
return emit_bpf_dispatcher(&prog, 0, num_funcs - 1, funcs, image, buf);
|
2019-12-13 17:51:08 +00:00
|
|
|
}
|
|
|
|
|
2017-12-15 01:55:15 +00:00
|
|
|
struct x64_jit_data {
|
2022-02-04 18:57:42 +00:00
|
|
|
struct bpf_binary_header *rw_header;
|
2017-12-15 01:55:15 +00:00
|
|
|
struct bpf_binary_header *header;
|
|
|
|
int *addrs;
|
|
|
|
u8 *image;
|
|
|
|
int proglen;
|
|
|
|
struct jit_context ctx;
|
|
|
|
};
|
|
|
|
|
bpf,x64: Pad NOPs to make images converge more easily
The x64 bpf jit expects bpf images converge within the given passes, but
it could fail to do so with some corner cases. For example:
l0: ja 40
l1: ja 40
[... repeated ja 40 ]
l39: ja 40
l40: ret #0
This bpf program contains 40 "ja 40" instructions which are effectively
NOPs and designed to be replaced with valid code dynamically. Ideally,
bpf jit should optimize those "ja 40" instructions out when translating
the bpf instructions into x64 machine code. However, do_jit() can only
remove one "ja 40" for offset==0 on each pass, so it requires at least
40 runs to eliminate those JMPs and exceeds the current limit of
passes(20). In the end, the program got rejected when BPF_JIT_ALWAYS_ON
is set even though it's legit as a classic socket filter.
To make bpf images more likely converge within 20 passes, this commit
pads some instructions with NOPs in the last 5 passes:
1. conditional jumps
A possible size variance comes from the adoption of imm8 JMP. If the
offset is imm8, we calculate the size difference of this BPF instruction
between the previous and the current pass and fill the gap with NOPs.
To avoid the recalculation of jump offset, those NOPs are inserted before
the JMP code, so we have to subtract the 2 bytes of imm8 JMP when
calculating the NOP number.
2. BPF_JA
There are two conditions for BPF_JA.
a.) nop jumps
If this instruction is not optimized out in the previous pass,
instead of removing it, we insert the equivalent size of NOPs.
b.) label jumps
Similar to condition jumps, we prepend NOPs right before the JMP
code.
To make the code concise, emit_nops() is modified to use the signed len and
return the number of inserted NOPs.
For bpf-to-bpf, we always enable padding for the extra pass since there
is only one extra run and the jump padding doesn't affected the images
that converge without padding.
After applying this patch, the corner case was loaded with the following
jit code:
flen=45 proglen=77 pass=17 image=ffffffffc03367d4 from=jump pid=10097
JIT code: 00000000: 0f 1f 44 00 00 55 48 89 e5 53 41 55 31 c0 45 31
JIT code: 00000010: ed 48 89 fb eb 30 eb 2e eb 2c eb 2a eb 28 eb 26
JIT code: 00000020: eb 24 eb 22 eb 20 eb 1e eb 1c eb 1a eb 18 eb 16
JIT code: 00000030: eb 14 eb 12 eb 10 eb 0e eb 0c eb 0a eb 08 eb 06
JIT code: 00000040: eb 04 eb 02 66 90 31 c0 41 5d 5b c9 c3
0: 0f 1f 44 00 00 nop DWORD PTR [rax+rax*1+0x0]
5: 55 push rbp
6: 48 89 e5 mov rbp,rsp
9: 53 push rbx
a: 41 55 push r13
c: 31 c0 xor eax,eax
e: 45 31 ed xor r13d,r13d
11: 48 89 fb mov rbx,rdi
14: eb 30 jmp 0x46
16: eb 2e jmp 0x46
...
3e: eb 06 jmp 0x46
40: eb 04 jmp 0x46
42: eb 02 jmp 0x46
44: 66 90 xchg ax,ax
46: 31 c0 xor eax,eax
48: 41 5d pop r13
4a: 5b pop rbx
4b: c9 leave
4c: c3 ret
At the 16th pass, 15 jumps were already optimized out, and one jump was
replaced with NOPs at 44 and the image converged at the 17th pass.
v4:
- Add the detailed comments about the possible padding bytes
v3:
- Copy the instructions of prologue separately or the size calculation
of the first BPF instruction would include the prologue.
- Replace WARN_ONCE() with pr_err() and EFAULT
- Use MAX_PASSES in the for loop condition check
- Remove the "padded" flag from x64_jit_data. For the extra pass of
subprogs, padding is always enabled since it won't hurt the images
that converge without padding.
v2:
- Simplify the sample code in the description and provide the jit code
- Check the expected padding bytes with WARN_ONCE
- Move the 'padded' flag to 'struct x64_jit_data'
Signed-off-by: Gary Lin <glin@suse.com>
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Link: https://lore.kernel.org/bpf/20210119102501.511-2-glin@suse.com
2021-01-19 10:24:59 +00:00
|
|
|
#define MAX_PASSES 20
|
|
|
|
#define PADDING_PASSES (MAX_PASSES - 5)
|
|
|
|
|
2016-05-13 17:08:31 +00:00
|
|
|
struct bpf_prog *bpf_int_jit_compile(struct bpf_prog *prog)
|
2014-05-14 02:50:45 +00:00
|
|
|
{
|
2022-02-04 18:57:42 +00:00
|
|
|
struct bpf_binary_header *rw_header = NULL;
|
2014-05-14 02:50:45 +00:00
|
|
|
struct bpf_binary_header *header = NULL;
|
2016-05-13 17:08:33 +00:00
|
|
|
struct bpf_prog *tmp, *orig_prog = prog;
|
2017-12-15 01:55:15 +00:00
|
|
|
struct x64_jit_data *jit_data;
|
2014-05-14 02:50:45 +00:00
|
|
|
int proglen, oldproglen = 0;
|
|
|
|
struct jit_context ctx = {};
|
2016-05-13 17:08:33 +00:00
|
|
|
bool tmp_blinded = false;
|
2017-12-15 01:55:15 +00:00
|
|
|
bool extra_pass = false;
|
bpf,x64: Pad NOPs to make images converge more easily
The x64 bpf jit expects bpf images converge within the given passes, but
it could fail to do so with some corner cases. For example:
l0: ja 40
l1: ja 40
[... repeated ja 40 ]
l39: ja 40
l40: ret #0
This bpf program contains 40 "ja 40" instructions which are effectively
NOPs and designed to be replaced with valid code dynamically. Ideally,
bpf jit should optimize those "ja 40" instructions out when translating
the bpf instructions into x64 machine code. However, do_jit() can only
remove one "ja 40" for offset==0 on each pass, so it requires at least
40 runs to eliminate those JMPs and exceeds the current limit of
passes(20). In the end, the program got rejected when BPF_JIT_ALWAYS_ON
is set even though it's legit as a classic socket filter.
To make bpf images more likely converge within 20 passes, this commit
pads some instructions with NOPs in the last 5 passes:
1. conditional jumps
A possible size variance comes from the adoption of imm8 JMP. If the
offset is imm8, we calculate the size difference of this BPF instruction
between the previous and the current pass and fill the gap with NOPs.
To avoid the recalculation of jump offset, those NOPs are inserted before
the JMP code, so we have to subtract the 2 bytes of imm8 JMP when
calculating the NOP number.
2. BPF_JA
There are two conditions for BPF_JA.
a.) nop jumps
If this instruction is not optimized out in the previous pass,
instead of removing it, we insert the equivalent size of NOPs.
b.) label jumps
Similar to condition jumps, we prepend NOPs right before the JMP
code.
To make the code concise, emit_nops() is modified to use the signed len and
return the number of inserted NOPs.
For bpf-to-bpf, we always enable padding for the extra pass since there
is only one extra run and the jump padding doesn't affected the images
that converge without padding.
After applying this patch, the corner case was loaded with the following
jit code:
flen=45 proglen=77 pass=17 image=ffffffffc03367d4 from=jump pid=10097
JIT code: 00000000: 0f 1f 44 00 00 55 48 89 e5 53 41 55 31 c0 45 31
JIT code: 00000010: ed 48 89 fb eb 30 eb 2e eb 2c eb 2a eb 28 eb 26
JIT code: 00000020: eb 24 eb 22 eb 20 eb 1e eb 1c eb 1a eb 18 eb 16
JIT code: 00000030: eb 14 eb 12 eb 10 eb 0e eb 0c eb 0a eb 08 eb 06
JIT code: 00000040: eb 04 eb 02 66 90 31 c0 41 5d 5b c9 c3
0: 0f 1f 44 00 00 nop DWORD PTR [rax+rax*1+0x0]
5: 55 push rbp
6: 48 89 e5 mov rbp,rsp
9: 53 push rbx
a: 41 55 push r13
c: 31 c0 xor eax,eax
e: 45 31 ed xor r13d,r13d
11: 48 89 fb mov rbx,rdi
14: eb 30 jmp 0x46
16: eb 2e jmp 0x46
...
3e: eb 06 jmp 0x46
40: eb 04 jmp 0x46
42: eb 02 jmp 0x46
44: 66 90 xchg ax,ax
46: 31 c0 xor eax,eax
48: 41 5d pop r13
4a: 5b pop rbx
4b: c9 leave
4c: c3 ret
At the 16th pass, 15 jumps were already optimized out, and one jump was
replaced with NOPs at 44 and the image converged at the 17th pass.
v4:
- Add the detailed comments about the possible padding bytes
v3:
- Copy the instructions of prologue separately or the size calculation
of the first BPF instruction would include the prologue.
- Replace WARN_ONCE() with pr_err() and EFAULT
- Use MAX_PASSES in the for loop condition check
- Remove the "padded" flag from x64_jit_data. For the extra pass of
subprogs, padding is always enabled since it won't hurt the images
that converge without padding.
v2:
- Simplify the sample code in the description and provide the jit code
- Check the expected padding bytes with WARN_ONCE
- Move the 'padded' flag to 'struct x64_jit_data'
Signed-off-by: Gary Lin <glin@suse.com>
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Link: https://lore.kernel.org/bpf/20210119102501.511-2-glin@suse.com
2021-01-19 10:24:59 +00:00
|
|
|
bool padding = false;
|
2022-02-04 18:57:42 +00:00
|
|
|
u8 *rw_image = NULL;
|
2014-05-14 02:50:45 +00:00
|
|
|
u8 *image = NULL;
|
|
|
|
int *addrs;
|
|
|
|
int pass;
|
|
|
|
int i;
|
|
|
|
|
2017-12-15 01:55:14 +00:00
|
|
|
if (!prog->jit_requested)
|
2016-05-13 17:08:33 +00:00
|
|
|
return orig_prog;
|
|
|
|
|
|
|
|
tmp = bpf_jit_blind_constants(prog);
|
x86/bpf: Clean up non-standard comments, to make the code more readable
So by chance I looked into x86 assembly in arch/x86/net/bpf_jit_comp.c and
noticed the weird and inconsistent comment style it mistakenly learned from
the networking code:
/* Multi-line comment ...
* ... looks like this.
*/
Fix this to use the standard comment style specified in Documentation/CodingStyle
and used in arch/x86/ as well:
/*
* Multi-line comment ...
* ... looks like this.
*/
Also, to quote Linus's ... more explicit views about this:
http://article.gmane.org/gmane.linux.kernel.cryptoapi/21066
> But no, the networking code picked *none* of the above sane formats.
> Instead, it picked these two models that are just half-arsed
> shit-for-brains:
>
> (no)
> /* This is disgusting drug-induced
> * crap, and should die
> */
>
> (no-no-no)
> /* This is also very nasty
> * and visually unbalanced */
>
> Please. The networking code actually has the *worst* possible comment
> style. You can literally find that (no-no-no) style, which is just
> really horribly disgusting and worse than the otherwise fairly similar
> (d) in pretty much every way.
Also improve the comments and some other details while at it:
- Don't mix same-line and previous-line comment style on otherwise
identical code patterns within the same function,
- capitalize 'BPF' and x86 register names consistently,
- capitalize sentences consistently,
- instead of 'x64' use 'x86-64': x64 is a Microsoft specific term,
- use more consistent punctuation,
- use standard coding style in macros as well,
- fix typos and a few other minor details.
Consistent coding style is not optional, at least in arch/x86/.
No change in functionality.
( In case this commit causes conflicts with pending development code
I'll be glad to help resolve any conflicts! )
Acked-by: Thomas Gleixner <tglx@linutronix.de>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Borislav Petkov <bp@alien8.de>
Cc: H. Peter Anvin <hpa@zytor.com>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: David S. Miller <davem@davemloft.net>
Cc: Eric Dumazet <edumazet@google.com>
Cc: Daniel Borkmann <daniel@iogearbox.net>
Cc: Alexei Starovoitov <ast@fb.com>
Cc: Hideaki YOSHIFUJI <yoshfuji@linux-ipv6.org>
Cc: netdev@vger.kernel.org
Cc: linux-kernel@vger.kernel.org
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
2018-04-27 09:54:40 +00:00
|
|
|
/*
|
|
|
|
* If blinding was requested and we failed during blinding,
|
2016-05-13 17:08:33 +00:00
|
|
|
* we must fall back to the interpreter.
|
|
|
|
*/
|
|
|
|
if (IS_ERR(tmp))
|
|
|
|
return orig_prog;
|
|
|
|
if (tmp != prog) {
|
|
|
|
tmp_blinded = true;
|
|
|
|
prog = tmp;
|
|
|
|
}
|
2011-04-20 09:27:32 +00:00
|
|
|
|
2017-12-15 01:55:15 +00:00
|
|
|
jit_data = prog->aux->jit_data;
|
|
|
|
if (!jit_data) {
|
|
|
|
jit_data = kzalloc(sizeof(*jit_data), GFP_KERNEL);
|
|
|
|
if (!jit_data) {
|
|
|
|
prog = orig_prog;
|
|
|
|
goto out;
|
|
|
|
}
|
|
|
|
prog->aux->jit_data = jit_data;
|
|
|
|
}
|
|
|
|
addrs = jit_data->addrs;
|
|
|
|
if (addrs) {
|
|
|
|
ctx = jit_data->ctx;
|
|
|
|
oldproglen = jit_data->proglen;
|
|
|
|
image = jit_data->image;
|
|
|
|
header = jit_data->header;
|
2022-02-04 18:57:42 +00:00
|
|
|
rw_header = jit_data->rw_header;
|
|
|
|
rw_image = (void *)rw_header + ((void *)image - (void *)header);
|
2017-12-15 01:55:15 +00:00
|
|
|
extra_pass = true;
|
bpf,x64: Pad NOPs to make images converge more easily
The x64 bpf jit expects bpf images converge within the given passes, but
it could fail to do so with some corner cases. For example:
l0: ja 40
l1: ja 40
[... repeated ja 40 ]
l39: ja 40
l40: ret #0
This bpf program contains 40 "ja 40" instructions which are effectively
NOPs and designed to be replaced with valid code dynamically. Ideally,
bpf jit should optimize those "ja 40" instructions out when translating
the bpf instructions into x64 machine code. However, do_jit() can only
remove one "ja 40" for offset==0 on each pass, so it requires at least
40 runs to eliminate those JMPs and exceeds the current limit of
passes(20). In the end, the program got rejected when BPF_JIT_ALWAYS_ON
is set even though it's legit as a classic socket filter.
To make bpf images more likely converge within 20 passes, this commit
pads some instructions with NOPs in the last 5 passes:
1. conditional jumps
A possible size variance comes from the adoption of imm8 JMP. If the
offset is imm8, we calculate the size difference of this BPF instruction
between the previous and the current pass and fill the gap with NOPs.
To avoid the recalculation of jump offset, those NOPs are inserted before
the JMP code, so we have to subtract the 2 bytes of imm8 JMP when
calculating the NOP number.
2. BPF_JA
There are two conditions for BPF_JA.
a.) nop jumps
If this instruction is not optimized out in the previous pass,
instead of removing it, we insert the equivalent size of NOPs.
b.) label jumps
Similar to condition jumps, we prepend NOPs right before the JMP
code.
To make the code concise, emit_nops() is modified to use the signed len and
return the number of inserted NOPs.
For bpf-to-bpf, we always enable padding for the extra pass since there
is only one extra run and the jump padding doesn't affected the images
that converge without padding.
After applying this patch, the corner case was loaded with the following
jit code:
flen=45 proglen=77 pass=17 image=ffffffffc03367d4 from=jump pid=10097
JIT code: 00000000: 0f 1f 44 00 00 55 48 89 e5 53 41 55 31 c0 45 31
JIT code: 00000010: ed 48 89 fb eb 30 eb 2e eb 2c eb 2a eb 28 eb 26
JIT code: 00000020: eb 24 eb 22 eb 20 eb 1e eb 1c eb 1a eb 18 eb 16
JIT code: 00000030: eb 14 eb 12 eb 10 eb 0e eb 0c eb 0a eb 08 eb 06
JIT code: 00000040: eb 04 eb 02 66 90 31 c0 41 5d 5b c9 c3
0: 0f 1f 44 00 00 nop DWORD PTR [rax+rax*1+0x0]
5: 55 push rbp
6: 48 89 e5 mov rbp,rsp
9: 53 push rbx
a: 41 55 push r13
c: 31 c0 xor eax,eax
e: 45 31 ed xor r13d,r13d
11: 48 89 fb mov rbx,rdi
14: eb 30 jmp 0x46
16: eb 2e jmp 0x46
...
3e: eb 06 jmp 0x46
40: eb 04 jmp 0x46
42: eb 02 jmp 0x46
44: 66 90 xchg ax,ax
46: 31 c0 xor eax,eax
48: 41 5d pop r13
4a: 5b pop rbx
4b: c9 leave
4c: c3 ret
At the 16th pass, 15 jumps were already optimized out, and one jump was
replaced with NOPs at 44 and the image converged at the 17th pass.
v4:
- Add the detailed comments about the possible padding bytes
v3:
- Copy the instructions of prologue separately or the size calculation
of the first BPF instruction would include the prologue.
- Replace WARN_ONCE() with pr_err() and EFAULT
- Use MAX_PASSES in the for loop condition check
- Remove the "padded" flag from x64_jit_data. For the extra pass of
subprogs, padding is always enabled since it won't hurt the images
that converge without padding.
v2:
- Simplify the sample code in the description and provide the jit code
- Check the expected padding bytes with WARN_ONCE
- Move the 'padded' flag to 'struct x64_jit_data'
Signed-off-by: Gary Lin <glin@suse.com>
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Link: https://lore.kernel.org/bpf/20210119102501.511-2-glin@suse.com
2021-01-19 10:24:59 +00:00
|
|
|
padding = true;
|
2017-12-15 01:55:15 +00:00
|
|
|
goto skip_init_addrs;
|
|
|
|
}
|
bpf, x86: Use kvmalloc_array instead kmalloc_array in bpf_jit_comp
x86 bpf_jit_comp.c used kmalloc_array to store jited addresses
for each bpf insn. With a large bpf program, we have see the
following allocation failures in our production server:
page allocation failure: order:5, mode:0x40cc0(GFP_KERNEL|__GFP_COMP),
nodemask=(null),cpuset=/,mems_allowed=0"
Call Trace:
dump_stack+0x50/0x70
warn_alloc.cold.120+0x72/0xd2
? __alloc_pages_direct_compact+0x157/0x160
__alloc_pages_slowpath+0xcdb/0xd00
? get_page_from_freelist+0xe44/0x1600
? vunmap_page_range+0x1ba/0x340
__alloc_pages_nodemask+0x2c9/0x320
kmalloc_order+0x18/0x80
kmalloc_order_trace+0x1d/0xa0
bpf_int_jit_compile+0x1e2/0x484
? kmalloc_order_trace+0x1d/0xa0
bpf_prog_select_runtime+0xc3/0x150
bpf_prog_load+0x480/0x720
? __mod_memcg_lruvec_state+0x21/0x100
__do_sys_bpf+0xc31/0x2040
? close_pdeo+0x86/0xe0
do_syscall_64+0x42/0x110
entry_SYSCALL_64_after_hwframe+0x44/0xa9
RIP: 0033:0x7f2f300f7fa9
Code: Bad RIP value.
Dumped assembly:
ffffffff810b6d70 <bpf_int_jit_compile>:
; {
ffffffff810b6d70: e8 eb a5 b4 00 callq 0xffffffff81c01360 <__fentry__>
ffffffff810b6d75: 41 57 pushq %r15
...
ffffffff810b6f39: e9 72 fe ff ff jmp 0xffffffff810b6db0 <bpf_int_jit_compile+0x40>
; addrs = kmalloc_array(prog->len + 1, sizeof(*addrs), GFP_KERNEL);
ffffffff810b6f3e: 8b 45 0c movl 12(%rbp), %eax
; return __kmalloc(bytes, flags);
ffffffff810b6f41: be c0 0c 00 00 movl $3264, %esi
; addrs = kmalloc_array(prog->len + 1, sizeof(*addrs), GFP_KERNEL);
ffffffff810b6f46: 8d 78 01 leal 1(%rax), %edi
; if (unlikely(check_mul_overflow(n, size, &bytes)))
ffffffff810b6f49: 48 c1 e7 02 shlq $2, %rdi
; return __kmalloc(bytes, flags);
ffffffff810b6f4d: e8 8e 0c 1d 00 callq 0xffffffff81287be0 <__kmalloc>
; if (!addrs) {
ffffffff810b6f52: 48 85 c0 testq %rax, %rax
Change kmalloc_array() to kvmalloc_array() to avoid potential
allocation error for big bpf programs.
Signed-off-by: Yonghong Song <yhs@fb.com>
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Link: https://lore.kernel.org/bpf/20210309015647.3657852-1-yhs@fb.com
2021-03-09 01:56:47 +00:00
|
|
|
addrs = kvmalloc_array(prog->len + 1, sizeof(*addrs), GFP_KERNEL);
|
2016-05-13 17:08:33 +00:00
|
|
|
if (!addrs) {
|
|
|
|
prog = orig_prog;
|
2017-12-15 01:55:15 +00:00
|
|
|
goto out_addrs;
|
2016-05-13 17:08:33 +00:00
|
|
|
}
|
2014-05-14 02:50:45 +00:00
|
|
|
|
x86/bpf: Clean up non-standard comments, to make the code more readable
So by chance I looked into x86 assembly in arch/x86/net/bpf_jit_comp.c and
noticed the weird and inconsistent comment style it mistakenly learned from
the networking code:
/* Multi-line comment ...
* ... looks like this.
*/
Fix this to use the standard comment style specified in Documentation/CodingStyle
and used in arch/x86/ as well:
/*
* Multi-line comment ...
* ... looks like this.
*/
Also, to quote Linus's ... more explicit views about this:
http://article.gmane.org/gmane.linux.kernel.cryptoapi/21066
> But no, the networking code picked *none* of the above sane formats.
> Instead, it picked these two models that are just half-arsed
> shit-for-brains:
>
> (no)
> /* This is disgusting drug-induced
> * crap, and should die
> */
>
> (no-no-no)
> /* This is also very nasty
> * and visually unbalanced */
>
> Please. The networking code actually has the *worst* possible comment
> style. You can literally find that (no-no-no) style, which is just
> really horribly disgusting and worse than the otherwise fairly similar
> (d) in pretty much every way.
Also improve the comments and some other details while at it:
- Don't mix same-line and previous-line comment style on otherwise
identical code patterns within the same function,
- capitalize 'BPF' and x86 register names consistently,
- capitalize sentences consistently,
- instead of 'x64' use 'x86-64': x64 is a Microsoft specific term,
- use more consistent punctuation,
- use standard coding style in macros as well,
- fix typos and a few other minor details.
Consistent coding style is not optional, at least in arch/x86/.
No change in functionality.
( In case this commit causes conflicts with pending development code
I'll be glad to help resolve any conflicts! )
Acked-by: Thomas Gleixner <tglx@linutronix.de>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Borislav Petkov <bp@alien8.de>
Cc: H. Peter Anvin <hpa@zytor.com>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: David S. Miller <davem@davemloft.net>
Cc: Eric Dumazet <edumazet@google.com>
Cc: Daniel Borkmann <daniel@iogearbox.net>
Cc: Alexei Starovoitov <ast@fb.com>
Cc: Hideaki YOSHIFUJI <yoshfuji@linux-ipv6.org>
Cc: netdev@vger.kernel.org
Cc: linux-kernel@vger.kernel.org
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
2018-04-27 09:54:40 +00:00
|
|
|
/*
|
|
|
|
* Before first pass, make a rough estimation of addrs[]
|
|
|
|
* each BPF instruction is translated to less than 64 bytes
|
2014-05-14 02:50:45 +00:00
|
|
|
*/
|
2019-07-31 01:38:26 +00:00
|
|
|
for (proglen = 0, i = 0; i <= prog->len; i++) {
|
2014-05-14 02:50:45 +00:00
|
|
|
proglen += 64;
|
|
|
|
addrs[i] = proglen;
|
|
|
|
}
|
|
|
|
ctx.cleanup_addr = proglen;
|
2017-12-15 01:55:15 +00:00
|
|
|
skip_init_addrs:
|
2014-05-14 02:50:45 +00:00
|
|
|
|
x86/bpf: Clean up non-standard comments, to make the code more readable
So by chance I looked into x86 assembly in arch/x86/net/bpf_jit_comp.c and
noticed the weird and inconsistent comment style it mistakenly learned from
the networking code:
/* Multi-line comment ...
* ... looks like this.
*/
Fix this to use the standard comment style specified in Documentation/CodingStyle
and used in arch/x86/ as well:
/*
* Multi-line comment ...
* ... looks like this.
*/
Also, to quote Linus's ... more explicit views about this:
http://article.gmane.org/gmane.linux.kernel.cryptoapi/21066
> But no, the networking code picked *none* of the above sane formats.
> Instead, it picked these two models that are just half-arsed
> shit-for-brains:
>
> (no)
> /* This is disgusting drug-induced
> * crap, and should die
> */
>
> (no-no-no)
> /* This is also very nasty
> * and visually unbalanced */
>
> Please. The networking code actually has the *worst* possible comment
> style. You can literally find that (no-no-no) style, which is just
> really horribly disgusting and worse than the otherwise fairly similar
> (d) in pretty much every way.
Also improve the comments and some other details while at it:
- Don't mix same-line and previous-line comment style on otherwise
identical code patterns within the same function,
- capitalize 'BPF' and x86 register names consistently,
- capitalize sentences consistently,
- instead of 'x64' use 'x86-64': x64 is a Microsoft specific term,
- use more consistent punctuation,
- use standard coding style in macros as well,
- fix typos and a few other minor details.
Consistent coding style is not optional, at least in arch/x86/.
No change in functionality.
( In case this commit causes conflicts with pending development code
I'll be glad to help resolve any conflicts! )
Acked-by: Thomas Gleixner <tglx@linutronix.de>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Borislav Petkov <bp@alien8.de>
Cc: H. Peter Anvin <hpa@zytor.com>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: David S. Miller <davem@davemloft.net>
Cc: Eric Dumazet <edumazet@google.com>
Cc: Daniel Borkmann <daniel@iogearbox.net>
Cc: Alexei Starovoitov <ast@fb.com>
Cc: Hideaki YOSHIFUJI <yoshfuji@linux-ipv6.org>
Cc: netdev@vger.kernel.org
Cc: linux-kernel@vger.kernel.org
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
2018-04-27 09:54:40 +00:00
|
|
|
/*
|
|
|
|
* JITed image shrinks with every pass and the loop iterates
|
|
|
|
* until the image stops shrinking. Very large BPF programs
|
2015-05-22 22:42:55 +00:00
|
|
|
* may converge on the last pass. In such case do one more
|
x86/bpf: Clean up non-standard comments, to make the code more readable
So by chance I looked into x86 assembly in arch/x86/net/bpf_jit_comp.c and
noticed the weird and inconsistent comment style it mistakenly learned from
the networking code:
/* Multi-line comment ...
* ... looks like this.
*/
Fix this to use the standard comment style specified in Documentation/CodingStyle
and used in arch/x86/ as well:
/*
* Multi-line comment ...
* ... looks like this.
*/
Also, to quote Linus's ... more explicit views about this:
http://article.gmane.org/gmane.linux.kernel.cryptoapi/21066
> But no, the networking code picked *none* of the above sane formats.
> Instead, it picked these two models that are just half-arsed
> shit-for-brains:
>
> (no)
> /* This is disgusting drug-induced
> * crap, and should die
> */
>
> (no-no-no)
> /* This is also very nasty
> * and visually unbalanced */
>
> Please. The networking code actually has the *worst* possible comment
> style. You can literally find that (no-no-no) style, which is just
> really horribly disgusting and worse than the otherwise fairly similar
> (d) in pretty much every way.
Also improve the comments and some other details while at it:
- Don't mix same-line and previous-line comment style on otherwise
identical code patterns within the same function,
- capitalize 'BPF' and x86 register names consistently,
- capitalize sentences consistently,
- instead of 'x64' use 'x86-64': x64 is a Microsoft specific term,
- use more consistent punctuation,
- use standard coding style in macros as well,
- fix typos and a few other minor details.
Consistent coding style is not optional, at least in arch/x86/.
No change in functionality.
( In case this commit causes conflicts with pending development code
I'll be glad to help resolve any conflicts! )
Acked-by: Thomas Gleixner <tglx@linutronix.de>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Borislav Petkov <bp@alien8.de>
Cc: H. Peter Anvin <hpa@zytor.com>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: David S. Miller <davem@davemloft.net>
Cc: Eric Dumazet <edumazet@google.com>
Cc: Daniel Borkmann <daniel@iogearbox.net>
Cc: Alexei Starovoitov <ast@fb.com>
Cc: Hideaki YOSHIFUJI <yoshfuji@linux-ipv6.org>
Cc: netdev@vger.kernel.org
Cc: linux-kernel@vger.kernel.org
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
2018-04-27 09:54:40 +00:00
|
|
|
* pass to emit the final image.
|
2015-05-22 22:42:55 +00:00
|
|
|
*/
|
bpf,x64: Pad NOPs to make images converge more easily
The x64 bpf jit expects bpf images converge within the given passes, but
it could fail to do so with some corner cases. For example:
l0: ja 40
l1: ja 40
[... repeated ja 40 ]
l39: ja 40
l40: ret #0
This bpf program contains 40 "ja 40" instructions which are effectively
NOPs and designed to be replaced with valid code dynamically. Ideally,
bpf jit should optimize those "ja 40" instructions out when translating
the bpf instructions into x64 machine code. However, do_jit() can only
remove one "ja 40" for offset==0 on each pass, so it requires at least
40 runs to eliminate those JMPs and exceeds the current limit of
passes(20). In the end, the program got rejected when BPF_JIT_ALWAYS_ON
is set even though it's legit as a classic socket filter.
To make bpf images more likely converge within 20 passes, this commit
pads some instructions with NOPs in the last 5 passes:
1. conditional jumps
A possible size variance comes from the adoption of imm8 JMP. If the
offset is imm8, we calculate the size difference of this BPF instruction
between the previous and the current pass and fill the gap with NOPs.
To avoid the recalculation of jump offset, those NOPs are inserted before
the JMP code, so we have to subtract the 2 bytes of imm8 JMP when
calculating the NOP number.
2. BPF_JA
There are two conditions for BPF_JA.
a.) nop jumps
If this instruction is not optimized out in the previous pass,
instead of removing it, we insert the equivalent size of NOPs.
b.) label jumps
Similar to condition jumps, we prepend NOPs right before the JMP
code.
To make the code concise, emit_nops() is modified to use the signed len and
return the number of inserted NOPs.
For bpf-to-bpf, we always enable padding for the extra pass since there
is only one extra run and the jump padding doesn't affected the images
that converge without padding.
After applying this patch, the corner case was loaded with the following
jit code:
flen=45 proglen=77 pass=17 image=ffffffffc03367d4 from=jump pid=10097
JIT code: 00000000: 0f 1f 44 00 00 55 48 89 e5 53 41 55 31 c0 45 31
JIT code: 00000010: ed 48 89 fb eb 30 eb 2e eb 2c eb 2a eb 28 eb 26
JIT code: 00000020: eb 24 eb 22 eb 20 eb 1e eb 1c eb 1a eb 18 eb 16
JIT code: 00000030: eb 14 eb 12 eb 10 eb 0e eb 0c eb 0a eb 08 eb 06
JIT code: 00000040: eb 04 eb 02 66 90 31 c0 41 5d 5b c9 c3
0: 0f 1f 44 00 00 nop DWORD PTR [rax+rax*1+0x0]
5: 55 push rbp
6: 48 89 e5 mov rbp,rsp
9: 53 push rbx
a: 41 55 push r13
c: 31 c0 xor eax,eax
e: 45 31 ed xor r13d,r13d
11: 48 89 fb mov rbx,rdi
14: eb 30 jmp 0x46
16: eb 2e jmp 0x46
...
3e: eb 06 jmp 0x46
40: eb 04 jmp 0x46
42: eb 02 jmp 0x46
44: 66 90 xchg ax,ax
46: 31 c0 xor eax,eax
48: 41 5d pop r13
4a: 5b pop rbx
4b: c9 leave
4c: c3 ret
At the 16th pass, 15 jumps were already optimized out, and one jump was
replaced with NOPs at 44 and the image converged at the 17th pass.
v4:
- Add the detailed comments about the possible padding bytes
v3:
- Copy the instructions of prologue separately or the size calculation
of the first BPF instruction would include the prologue.
- Replace WARN_ONCE() with pr_err() and EFAULT
- Use MAX_PASSES in the for loop condition check
- Remove the "padded" flag from x64_jit_data. For the extra pass of
subprogs, padding is always enabled since it won't hurt the images
that converge without padding.
v2:
- Simplify the sample code in the description and provide the jit code
- Check the expected padding bytes with WARN_ONCE
- Move the 'padded' flag to 'struct x64_jit_data'
Signed-off-by: Gary Lin <glin@suse.com>
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Link: https://lore.kernel.org/bpf/20210119102501.511-2-glin@suse.com
2021-01-19 10:24:59 +00:00
|
|
|
for (pass = 0; pass < MAX_PASSES || image; pass++) {
|
|
|
|
if (!padding && pass >= PADDING_PASSES)
|
|
|
|
padding = true;
|
2022-02-04 18:57:42 +00:00
|
|
|
proglen = do_jit(prog, addrs, image, rw_image, oldproglen, &ctx, padding);
|
2014-05-14 02:50:45 +00:00
|
|
|
if (proglen <= 0) {
|
2018-05-02 18:12:22 +00:00
|
|
|
out_image:
|
2014-05-14 02:50:45 +00:00
|
|
|
image = NULL;
|
2022-03-02 17:51:26 +00:00
|
|
|
if (header) {
|
|
|
|
bpf_arch_text_copy(&header->size, &rw_header->size,
|
|
|
|
sizeof(rw_header->size));
|
2022-02-04 18:57:42 +00:00
|
|
|
bpf_jit_binary_pack_free(header, rw_header);
|
2022-03-02 17:51:26 +00:00
|
|
|
}
|
2022-03-09 12:33:18 +00:00
|
|
|
/* Fall back to interpreter mode */
|
2016-05-13 17:08:33 +00:00
|
|
|
prog = orig_prog;
|
2022-03-09 12:33:18 +00:00
|
|
|
if (extra_pass) {
|
|
|
|
prog->bpf_func = NULL;
|
|
|
|
prog->jited = 0;
|
|
|
|
prog->jited_len = 0;
|
|
|
|
}
|
2016-05-13 17:08:33 +00:00
|
|
|
goto out_addrs;
|
2014-05-14 02:50:45 +00:00
|
|
|
}
|
2011-04-20 09:27:32 +00:00
|
|
|
if (image) {
|
2014-10-11 03:30:23 +00:00
|
|
|
if (proglen != oldproglen) {
|
2014-05-14 02:50:45 +00:00
|
|
|
pr_err("bpf_jit: proglen=%d != oldproglen=%d\n",
|
|
|
|
proglen, oldproglen);
|
2018-05-02 18:12:22 +00:00
|
|
|
goto out_image;
|
2014-10-11 03:30:23 +00:00
|
|
|
}
|
2011-04-20 09:27:32 +00:00
|
|
|
break;
|
|
|
|
}
|
|
|
|
if (proglen == oldproglen) {
|
2019-10-16 03:25:03 +00:00
|
|
|
/*
|
|
|
|
* The number of entries in extable is the number of BPF_LDX
|
|
|
|
* insns that access kernel memory via "pointer to BTF type".
|
|
|
|
* The verifier changed their opcode from LDX|MEM|size
|
|
|
|
* to LDX|PROBE_MEM|size to make JITing easier.
|
|
|
|
*/
|
|
|
|
u32 align = __alignof__(struct exception_table_entry);
|
|
|
|
u32 extable_size = prog->aux->num_exentries *
|
|
|
|
sizeof(struct exception_table_entry);
|
|
|
|
|
|
|
|
/* allocate module memory for x86 insns and extable */
|
2022-02-04 18:57:42 +00:00
|
|
|
header = bpf_jit_binary_pack_alloc(roundup(proglen, align) + extable_size,
|
|
|
|
&image, align, &rw_header, &rw_image,
|
|
|
|
jit_fill_hole);
|
2016-05-13 17:08:33 +00:00
|
|
|
if (!header) {
|
|
|
|
prog = orig_prog;
|
|
|
|
goto out_addrs;
|
|
|
|
}
|
2019-10-16 03:25:03 +00:00
|
|
|
prog->aux->extable = (void *) image + roundup(proglen, align);
|
2011-04-20 09:27:32 +00:00
|
|
|
}
|
|
|
|
oldproglen = proglen;
|
2018-03-07 21:10:01 +00:00
|
|
|
cond_resched();
|
2011-04-20 09:27:32 +00:00
|
|
|
}
|
2013-03-21 21:22:03 +00:00
|
|
|
|
2011-04-20 09:27:32 +00:00
|
|
|
if (bpf_jit_enable > 1)
|
2023-06-09 00:54:39 +00:00
|
|
|
bpf_jit_dump(prog->len, proglen, pass + 1, rw_image);
|
2011-04-20 09:27:32 +00:00
|
|
|
|
|
|
|
if (image) {
|
2017-12-15 01:55:15 +00:00
|
|
|
if (!prog->is_func || extra_pass) {
|
2022-02-04 18:57:42 +00:00
|
|
|
/*
|
|
|
|
* bpf_jit_binary_pack_finalize fails in two scenarios:
|
|
|
|
* 1) header is not pointing to proper module memory;
|
|
|
|
* 2) the arch doesn't support bpf_arch_text_copy().
|
|
|
|
*
|
2022-02-08 06:25:33 +00:00
|
|
|
* Both cases are serious bugs and justify WARN_ON.
|
2022-02-04 18:57:42 +00:00
|
|
|
*/
|
2022-02-08 06:25:33 +00:00
|
|
|
if (WARN_ON(bpf_jit_binary_pack_finalize(prog, header, rw_header))) {
|
2022-03-09 12:33:18 +00:00
|
|
|
/* header has been freed */
|
|
|
|
header = NULL;
|
|
|
|
goto out_image;
|
2022-02-08 06:25:33 +00:00
|
|
|
}
|
|
|
|
|
bpf, x86: Emit patchable direct jump as tail call
Add initial code emission for *direct* jumps for tail call maps in
order to avoid the retpoline overhead from a493a87f38cf ("bpf, x64:
implement retpoline for tail call") for situations that allow for
it, meaning, for known constant keys at verification time which are
used as index into the tail call map. In case of Cilium which makes
heavy use of tail calls, constant keys are used in the vast majority,
only for a single occurrence we use a dynamic key.
High level outline is that if the target prog is NULL in the map, we
emit a 5-byte nop for the fall-through case and if not, we emit a
5-byte direct relative jmp to the target bpf_func + skipped prologue
offset. Later during runtime, we patch these 5-byte nop/jmps upon
tail call map update or deletions dynamically. Note that on x86-64
the direct jmp works as we reuse the same stack frame and skip
prologue (as opposed to some other JIT implementations).
One of the issues is that the tail call map slots can change at any
given time even during JITing. Therefore, we have two passes: i) emit
nops for all patchable locations during main JITing phase until we
declare prog->jited = 1 eventually. At this point the image is stable,
not public yet and with all jmps disabled. While JITing, we collect
additional info like poke->ip in order to remember the patch location
for later modifications. In ii) bpf_tail_call_direct_fixup() walks
over the progs poke_tab, locks the tail call maps poke_mutex to
prevent from parallel updates and patches in the right locations via
__bpf_arch_text_poke(). Note, the main bpf_arch_text_poke() cannot
be used at this point since we're not yet exposed to kallsyms. For
the update we use plain memcpy() since the image is not public and
still in read-write mode. After patching, we activate that poke entry
through poke->ip_stable. Meaning, at this point any tail call map
updates/deletions are not going to ignore that poke entry anymore.
Then, bpf_arch_text_poke() might still occur on the read-write image
until we finally locked it as read-only. Both modifications on the
given image are under text_mutex to avoid interference with each
other when update requests come in in parallel for different tail
call maps (current one we have locked in JIT and different one where
poke->ip_stable was already set).
Example prog:
# ./bpftool p d x i 1655
0: (b7) r3 = 0
1: (18) r2 = map[id:526]
3: (85) call bpf_tail_call#12
4: (b7) r0 = 1
5: (95) exit
Before:
# ./bpftool p d j i 1655
0xffffffffc076e55c:
0: nopl 0x0(%rax,%rax,1)
5: push %rbp
6: mov %rsp,%rbp
9: sub $0x200,%rsp
10: push %rbx
11: push %r13
13: push %r14
15: push %r15
17: pushq $0x0 _
19: xor %edx,%edx |_ index (arg 3)
1b: movabs $0xffff88d95cc82600,%rsi |_ map (arg 2)
25: mov %edx,%edx | index >= array->map.max_entries
27: cmp %edx,0x24(%rsi) |
2a: jbe 0x0000000000000066 |_
2c: mov -0x224(%rbp),%eax | tail call limit check
32: cmp $0x20,%eax |
35: ja 0x0000000000000066 |
37: add $0x1,%eax |
3a: mov %eax,-0x224(%rbp) |_
40: mov 0xd0(%rsi,%rdx,8),%rax |_ prog = array->ptrs[index]
48: test %rax,%rax | prog == NULL check
4b: je 0x0000000000000066 |_
4d: mov 0x30(%rax),%rax | goto *(prog->bpf_func + prologue_size)
51: add $0x19,%rax |
55: callq 0x0000000000000061 | retpoline for indirect jump
5a: pause |
5c: lfence |
5f: jmp 0x000000000000005a |
61: mov %rax,(%rsp) |
65: retq |_
66: mov $0x1,%eax
6b: pop %rbx
6c: pop %r15
6e: pop %r14
70: pop %r13
72: pop %rbx
73: leaveq
74: retq
After; state after JIT:
# ./bpftool p d j i 1655
0xffffffffc08e8930:
0: nopl 0x0(%rax,%rax,1)
5: push %rbp
6: mov %rsp,%rbp
9: sub $0x200,%rsp
10: push %rbx
11: push %r13
13: push %r14
15: push %r15
17: pushq $0x0 _
19: xor %edx,%edx |_ index (arg 3)
1b: movabs $0xffff9d8afd74c000,%rsi |_ map (arg 2)
25: mov -0x224(%rbp),%eax | tail call limit check
2b: cmp $0x20,%eax |
2e: ja 0x000000000000003e |
30: add $0x1,%eax |
33: mov %eax,-0x224(%rbp) |_
39: jmpq 0xfffffffffffd1785 |_ [direct] goto *(prog->bpf_func + prologue_size)
3e: mov $0x1,%eax
43: pop %rbx
44: pop %r15
46: pop %r14
48: pop %r13
4a: pop %rbx
4b: leaveq
4c: retq
After; state after map update (target prog):
# ./bpftool p d j i 1655
0xffffffffc08e8930:
0: nopl 0x0(%rax,%rax,1)
5: push %rbp
6: mov %rsp,%rbp
9: sub $0x200,%rsp
10: push %rbx
11: push %r13
13: push %r14
15: push %r15
17: pushq $0x0
19: xor %edx,%edx
1b: movabs $0xffff9d8afd74c000,%rsi
25: mov -0x224(%rbp),%eax
2b: cmp $0x20,%eax .
2e: ja 0x000000000000003e .
30: add $0x1,%eax .
33: mov %eax,-0x224(%rbp) |_
39: jmpq 0xffffffffffb09f55 |_ goto *(prog->bpf_func + prologue_size)
3e: mov $0x1,%eax
43: pop %rbx
44: pop %r15
46: pop %r14
48: pop %r13
4a: pop %rbx
4b: leaveq
4c: retq
After; state after map update (no prog):
# ./bpftool p d j i 1655
0xffffffffc08e8930:
0: nopl 0x0(%rax,%rax,1)
5: push %rbp
6: mov %rsp,%rbp
9: sub $0x200,%rsp
10: push %rbx
11: push %r13
13: push %r14
15: push %r15
17: pushq $0x0
19: xor %edx,%edx
1b: movabs $0xffff9d8afd74c000,%rsi
25: mov -0x224(%rbp),%eax
2b: cmp $0x20,%eax .
2e: ja 0x000000000000003e .
30: add $0x1,%eax .
33: mov %eax,-0x224(%rbp) |_
39: nopl 0x0(%rax,%rax,1) |_ fall-through nop
3e: mov $0x1,%eax
43: pop %rbx
44: pop %r15
46: pop %r14
48: pop %r13
4a: pop %rbx
4b: leaveq
4c: retq
Nice bonus is that this also shrinks the code emission quite a bit
for every tail call invocation.
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Link: https://lore.kernel.org/bpf/6ada4c1c9d35eeb5f4ecfab94593dafa6b5c4b09.1574452833.git.daniel@iogearbox.net
2019-11-22 20:08:00 +00:00
|
|
|
bpf_tail_call_direct_fixup(prog);
|
2017-12-15 01:55:15 +00:00
|
|
|
} else {
|
|
|
|
jit_data->addrs = addrs;
|
|
|
|
jit_data->ctx = ctx;
|
|
|
|
jit_data->proglen = proglen;
|
|
|
|
jit_data->image = image;
|
|
|
|
jit_data->header = header;
|
2022-02-04 18:57:42 +00:00
|
|
|
jit_data->rw_header = rw_header;
|
2017-12-15 01:55:15 +00:00
|
|
|
}
|
2023-12-15 09:12:18 +00:00
|
|
|
/*
|
|
|
|
* ctx.prog_offset is used when CFI preambles put code *before*
|
|
|
|
* the function. See emit_cfi(). For FineIBT specifically this code
|
|
|
|
* can also be executed and bpf_prog_kallsyms_add() will
|
|
|
|
* generate an additional symbol to cover this, hence also
|
|
|
|
* decrement proglen.
|
|
|
|
*/
|
|
|
|
prog->bpf_func = (void *)image + cfi_get_offset();
|
2015-09-29 23:41:50 +00:00
|
|
|
prog->jited = 1;
|
2023-12-15 09:12:18 +00:00
|
|
|
prog->jited_len = proglen - cfi_get_offset();
|
2017-01-06 23:26:33 +00:00
|
|
|
} else {
|
|
|
|
prog = orig_prog;
|
2011-04-20 09:27:32 +00:00
|
|
|
}
|
2016-05-13 17:08:33 +00:00
|
|
|
|
bpf, x64: fix memleak when not converging on calls
The JIT logic in jit_subprogs() is as follows: for all subprogs we
allocate a bpf_prog_alloc(), populate it (prog->is_func = 1 here),
and pass it to bpf_int_jit_compile(). If a failure occurred during
JIT and prog->jited is not set, then we bail out from attempting to
JIT the whole program, and punt to the interpreter instead. In case
JITing went successful, we fixup BPF call offsets and do another
pass to bpf_int_jit_compile() (extra_pass is true at that point) to
complete JITing calls. Given that requires to pass JIT context around
addrs and jit_data from x86 JIT are freed in the extra_pass in
bpf_int_jit_compile() when calls are involved (if not, they can
be freed immediately). However, if in the original pass, the JIT
image didn't converge then we leak addrs and jit_data since image
itself is NULL, the prog->is_func is set and extra_pass is false
in that case, meaning both will become unreachable and are never
cleaned up, therefore we need to free as well on !image. Only x64
JIT is affected.
Fixes: 1c2a088a6626 ("bpf: x64: add JIT support for multi-function programs")
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Acked-by: Alexei Starovoitov <ast@kernel.org>
Acked-by: David S. Miller <davem@davemloft.net>
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
2018-05-02 18:12:23 +00:00
|
|
|
if (!image || !prog->is_func || extra_pass) {
|
2018-12-08 00:42:25 +00:00
|
|
|
if (image)
|
2019-07-31 01:38:26 +00:00
|
|
|
bpf_prog_fill_jited_linfo(prog, addrs + 1);
|
2016-05-13 17:08:33 +00:00
|
|
|
out_addrs:
|
bpf, x86: Use kvmalloc_array instead kmalloc_array in bpf_jit_comp
x86 bpf_jit_comp.c used kmalloc_array to store jited addresses
for each bpf insn. With a large bpf program, we have see the
following allocation failures in our production server:
page allocation failure: order:5, mode:0x40cc0(GFP_KERNEL|__GFP_COMP),
nodemask=(null),cpuset=/,mems_allowed=0"
Call Trace:
dump_stack+0x50/0x70
warn_alloc.cold.120+0x72/0xd2
? __alloc_pages_direct_compact+0x157/0x160
__alloc_pages_slowpath+0xcdb/0xd00
? get_page_from_freelist+0xe44/0x1600
? vunmap_page_range+0x1ba/0x340
__alloc_pages_nodemask+0x2c9/0x320
kmalloc_order+0x18/0x80
kmalloc_order_trace+0x1d/0xa0
bpf_int_jit_compile+0x1e2/0x484
? kmalloc_order_trace+0x1d/0xa0
bpf_prog_select_runtime+0xc3/0x150
bpf_prog_load+0x480/0x720
? __mod_memcg_lruvec_state+0x21/0x100
__do_sys_bpf+0xc31/0x2040
? close_pdeo+0x86/0xe0
do_syscall_64+0x42/0x110
entry_SYSCALL_64_after_hwframe+0x44/0xa9
RIP: 0033:0x7f2f300f7fa9
Code: Bad RIP value.
Dumped assembly:
ffffffff810b6d70 <bpf_int_jit_compile>:
; {
ffffffff810b6d70: e8 eb a5 b4 00 callq 0xffffffff81c01360 <__fentry__>
ffffffff810b6d75: 41 57 pushq %r15
...
ffffffff810b6f39: e9 72 fe ff ff jmp 0xffffffff810b6db0 <bpf_int_jit_compile+0x40>
; addrs = kmalloc_array(prog->len + 1, sizeof(*addrs), GFP_KERNEL);
ffffffff810b6f3e: 8b 45 0c movl 12(%rbp), %eax
; return __kmalloc(bytes, flags);
ffffffff810b6f41: be c0 0c 00 00 movl $3264, %esi
; addrs = kmalloc_array(prog->len + 1, sizeof(*addrs), GFP_KERNEL);
ffffffff810b6f46: 8d 78 01 leal 1(%rax), %edi
; if (unlikely(check_mul_overflow(n, size, &bytes)))
ffffffff810b6f49: 48 c1 e7 02 shlq $2, %rdi
; return __kmalloc(bytes, flags);
ffffffff810b6f4d: e8 8e 0c 1d 00 callq 0xffffffff81287be0 <__kmalloc>
; if (!addrs) {
ffffffff810b6f52: 48 85 c0 testq %rax, %rax
Change kmalloc_array() to kvmalloc_array() to avoid potential
allocation error for big bpf programs.
Signed-off-by: Yonghong Song <yhs@fb.com>
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Link: https://lore.kernel.org/bpf/20210309015647.3657852-1-yhs@fb.com
2021-03-09 01:56:47 +00:00
|
|
|
kvfree(addrs);
|
2017-12-15 01:55:15 +00:00
|
|
|
kfree(jit_data);
|
|
|
|
prog->aux->jit_data = NULL;
|
|
|
|
}
|
2016-05-13 17:08:33 +00:00
|
|
|
out:
|
|
|
|
if (tmp_blinded)
|
|
|
|
bpf_jit_prog_release_other(prog, prog == orig_prog ?
|
|
|
|
tmp : orig_prog);
|
2016-05-13 17:08:31 +00:00
|
|
|
return prog;
|
2011-04-20 09:27:32 +00:00
|
|
|
}
|
bpf: Support bpf program calling kernel function
This patch adds support to BPF verifier to allow bpf program calling
kernel function directly.
The use case included in this set is to allow bpf-tcp-cc to directly
call some tcp-cc helper functions (e.g. "tcp_cong_avoid_ai()"). Those
functions have already been used by some kernel tcp-cc implementations.
This set will also allow the bpf-tcp-cc program to directly call the
kernel tcp-cc implementation, For example, a bpf_dctcp may only want to
implement its own dctcp_cwnd_event() and reuse other dctcp_*() directly
from the kernel tcp_dctcp.c instead of reimplementing (or
copy-and-pasting) them.
The tcp-cc kernel functions mentioned above will be white listed
for the struct_ops bpf-tcp-cc programs to use in a later patch.
The white listed functions are not bounded to a fixed ABI contract.
Those functions have already been used by the existing kernel tcp-cc.
If any of them has changed, both in-tree and out-of-tree kernel tcp-cc
implementations have to be changed. The same goes for the struct_ops
bpf-tcp-cc programs which have to be adjusted accordingly.
This patch is to make the required changes in the bpf verifier.
First change is in btf.c, it adds a case in "btf_check_func_arg_match()".
When the passed in "btf->kernel_btf == true", it means matching the
verifier regs' states with a kernel function. This will handle the
PTR_TO_BTF_ID reg. It also maps PTR_TO_SOCK_COMMON, PTR_TO_SOCKET,
and PTR_TO_TCP_SOCK to its kernel's btf_id.
In the later libbpf patch, the insn calling a kernel function will
look like:
insn->code == (BPF_JMP | BPF_CALL)
insn->src_reg == BPF_PSEUDO_KFUNC_CALL /* <- new in this patch */
insn->imm == func_btf_id /* btf_id of the running kernel */
[ For the future calling function-in-kernel-module support, an array
of module btf_fds can be passed at the load time and insn->off
can be used to index into this array. ]
At the early stage of verifier, the verifier will collect all kernel
function calls into "struct bpf_kfunc_desc". Those
descriptors are stored in "prog->aux->kfunc_tab" and will
be available to the JIT. Since this "add" operation is similar
to the current "add_subprog()" and looking for the same insn->code,
they are done together in the new "add_subprog_and_kfunc()".
In the "do_check()" stage, the new "check_kfunc_call()" is added
to verify the kernel function call instruction:
1. Ensure the kernel function can be used by a particular BPF_PROG_TYPE.
A new bpf_verifier_ops "check_kfunc_call" is added to do that.
The bpf-tcp-cc struct_ops program will implement this function in
a later patch.
2. Call "btf_check_kfunc_args_match()" to ensure the regs can be
used as the args of a kernel function.
3. Mark the regs' type, subreg_def, and zext_dst.
At the later do_misc_fixups() stage, the new fixup_kfunc_call()
will replace the insn->imm with the function address (relative
to __bpf_call_base). If needed, the jit can find the btf_func_model
by calling the new bpf_jit_find_kfunc_model(prog, insn).
With the imm set to the function address, "bpftool prog dump xlated"
will be able to display the kernel function calls the same way as
it displays other bpf helper calls.
gpl_compatible program is required to call kernel function.
This feature currently requires JIT.
The verifier selftests are adjusted because of the changes in
the verbose log in add_subprog_and_kfunc().
Signed-off-by: Martin KaFai Lau <kafai@fb.com>
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Link: https://lore.kernel.org/bpf/20210325015142.1544736-1-kafai@fb.com
2021-03-25 01:51:42 +00:00
|
|
|
|
|
|
|
bool bpf_jit_supports_kfunc_call(void)
|
|
|
|
{
|
|
|
|
return true;
|
|
|
|
}
|
2022-02-04 18:57:39 +00:00
|
|
|
|
|
|
|
void *bpf_arch_text_copy(void *dst, void *src, size_t len)
|
|
|
|
{
|
|
|
|
if (text_poke_copy(dst, src, len) == NULL)
|
|
|
|
return ERR_PTR(-EINVAL);
|
|
|
|
return dst;
|
|
|
|
}
|
2022-06-17 10:57:34 +00:00
|
|
|
|
|
|
|
/* Indicate the JIT backend supports mixing bpf2bpf and tailcalls. */
|
|
|
|
bool bpf_jit_supports_subprog_tailcalls(void)
|
|
|
|
{
|
|
|
|
return true;
|
|
|
|
}
|
bpf, x86: fix freeing of not-finalized bpf_prog_pack
syzbot reported a few issues with bpf_prog_pack [1], [2]. This only happens
with multiple subprogs. In jit_subprogs(), we first call bpf_int_jit_compile()
on each sub program. And then, we call it on each sub program again. jit_data
is not freed in the first call of bpf_int_jit_compile(). Similarly we don't
call bpf_jit_binary_pack_finalize() in the first call of bpf_int_jit_compile().
If bpf_int_jit_compile() failed for one sub program, we will call
bpf_jit_binary_pack_finalize() for this sub program. However, we don't have a
chance to call it for other sub programs. Then we will hit "goto out_free" in
jit_subprogs(), and call bpf_jit_free on some subprograms that haven't got
bpf_jit_binary_pack_finalize() yet.
At this point, bpf_jit_binary_pack_free() is called and the whole 2MB page is
freed erroneously.
Fix this with a custom bpf_jit_free() for x86_64, which calls
bpf_jit_binary_pack_finalize() if necessary. Also, with custom
bpf_jit_free(), bpf_prog_aux->use_bpf_prog_pack is not needed any more,
remove it.
Fixes: 1022a5498f6f ("bpf, x86_64: Use bpf_jit_binary_pack_alloc")
[1] https://syzkaller.appspot.com/bug?extid=2f649ec6d2eea1495a8f
[2] https://syzkaller.appspot.com/bug?extid=87f65c75f4a72db05445
Reported-by: syzbot+2f649ec6d2eea1495a8f@syzkaller.appspotmail.com
Reported-by: syzbot+87f65c75f4a72db05445@syzkaller.appspotmail.com
Signed-off-by: Song Liu <song@kernel.org>
Link: https://lore.kernel.org/r/20220706002612.4013790-1-song@kernel.org
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
2022-07-06 00:26:12 +00:00
|
|
|
|
|
|
|
void bpf_jit_free(struct bpf_prog *prog)
|
|
|
|
{
|
|
|
|
if (prog->jited) {
|
|
|
|
struct x64_jit_data *jit_data = prog->aux->jit_data;
|
|
|
|
struct bpf_binary_header *hdr;
|
|
|
|
|
|
|
|
/*
|
|
|
|
* If we fail the final pass of JIT (from jit_subprogs),
|
|
|
|
* the program may not be finalized yet. Call finalize here
|
|
|
|
* before freeing it.
|
|
|
|
*/
|
|
|
|
if (jit_data) {
|
|
|
|
bpf_jit_binary_pack_finalize(prog, jit_data->header,
|
|
|
|
jit_data->rw_header);
|
|
|
|
kvfree(jit_data->addrs);
|
|
|
|
kfree(jit_data);
|
|
|
|
}
|
2023-12-15 09:12:18 +00:00
|
|
|
prog->bpf_func = (void *)prog->bpf_func - cfi_get_offset();
|
bpf, x86: fix freeing of not-finalized bpf_prog_pack
syzbot reported a few issues with bpf_prog_pack [1], [2]. This only happens
with multiple subprogs. In jit_subprogs(), we first call bpf_int_jit_compile()
on each sub program. And then, we call it on each sub program again. jit_data
is not freed in the first call of bpf_int_jit_compile(). Similarly we don't
call bpf_jit_binary_pack_finalize() in the first call of bpf_int_jit_compile().
If bpf_int_jit_compile() failed for one sub program, we will call
bpf_jit_binary_pack_finalize() for this sub program. However, we don't have a
chance to call it for other sub programs. Then we will hit "goto out_free" in
jit_subprogs(), and call bpf_jit_free on some subprograms that haven't got
bpf_jit_binary_pack_finalize() yet.
At this point, bpf_jit_binary_pack_free() is called and the whole 2MB page is
freed erroneously.
Fix this with a custom bpf_jit_free() for x86_64, which calls
bpf_jit_binary_pack_finalize() if necessary. Also, with custom
bpf_jit_free(), bpf_prog_aux->use_bpf_prog_pack is not needed any more,
remove it.
Fixes: 1022a5498f6f ("bpf, x86_64: Use bpf_jit_binary_pack_alloc")
[1] https://syzkaller.appspot.com/bug?extid=2f649ec6d2eea1495a8f
[2] https://syzkaller.appspot.com/bug?extid=87f65c75f4a72db05445
Reported-by: syzbot+2f649ec6d2eea1495a8f@syzkaller.appspotmail.com
Reported-by: syzbot+87f65c75f4a72db05445@syzkaller.appspotmail.com
Signed-off-by: Song Liu <song@kernel.org>
Link: https://lore.kernel.org/r/20220706002612.4013790-1-song@kernel.org
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
2022-07-06 00:26:12 +00:00
|
|
|
hdr = bpf_jit_binary_pack_hdr(prog);
|
|
|
|
bpf_jit_binary_pack_free(hdr, NULL);
|
|
|
|
WARN_ON_ONCE(!bpf_prog_kallsyms_verify_off(prog));
|
|
|
|
}
|
|
|
|
|
|
|
|
bpf_prog_unlock_free(prog);
|
|
|
|
}
|
2023-09-12 23:31:59 +00:00
|
|
|
|
|
|
|
bool bpf_jit_supports_exceptions(void)
|
|
|
|
{
|
|
|
|
/* We unwind through both kernel frames (starting from within bpf_throw
|
2023-09-18 15:52:33 +00:00
|
|
|
* call) and BPF frames. Therefore we require ORC unwinder to be enabled
|
|
|
|
* to walk kernel frames and reach BPF frames in the stack trace.
|
2023-09-12 23:31:59 +00:00
|
|
|
*/
|
2023-09-18 15:52:33 +00:00
|
|
|
return IS_ENABLED(CONFIG_UNWINDER_ORC);
|
2023-09-12 23:31:59 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
void arch_bpf_stack_walk(bool (*consume_fn)(void *cookie, u64 ip, u64 sp, u64 bp), void *cookie)
|
|
|
|
{
|
2023-09-18 15:52:33 +00:00
|
|
|
#if defined(CONFIG_UNWINDER_ORC)
|
2023-09-12 23:31:59 +00:00
|
|
|
struct unwind_state state;
|
|
|
|
unsigned long addr;
|
|
|
|
|
|
|
|
for (unwind_start(&state, current, NULL, NULL); !unwind_done(&state);
|
|
|
|
unwind_next_frame(&state)) {
|
|
|
|
addr = unwind_get_return_address(&state);
|
|
|
|
if (!addr || !consume_fn(cookie, (u64)addr, (u64)state.sp, (u64)state.bp))
|
|
|
|
break;
|
|
|
|
}
|
|
|
|
return;
|
|
|
|
#endif
|
|
|
|
WARN(1, "verification of programs using bpf_throw should have failed\n");
|
|
|
|
}
|
2023-12-06 08:30:40 +00:00
|
|
|
|
|
|
|
void bpf_arch_poke_desc_update(struct bpf_jit_poke_descriptor *poke,
|
|
|
|
struct bpf_prog *new, struct bpf_prog *old)
|
|
|
|
{
|
|
|
|
u8 *old_addr, *new_addr, *old_bypass_addr;
|
|
|
|
int ret;
|
|
|
|
|
|
|
|
old_bypass_addr = old ? NULL : poke->bypass_addr;
|
|
|
|
old_addr = old ? (u8 *)old->bpf_func + poke->adj_off : NULL;
|
|
|
|
new_addr = new ? (u8 *)new->bpf_func + poke->adj_off : NULL;
|
|
|
|
|
|
|
|
/*
|
|
|
|
* On program loading or teardown, the program's kallsym entry
|
|
|
|
* might not be in place, so we use __bpf_arch_text_poke to skip
|
|
|
|
* the kallsyms check.
|
|
|
|
*/
|
|
|
|
if (new) {
|
|
|
|
ret = __bpf_arch_text_poke(poke->tailcall_target,
|
|
|
|
BPF_MOD_JUMP,
|
|
|
|
old_addr, new_addr);
|
|
|
|
BUG_ON(ret < 0);
|
|
|
|
if (!old) {
|
|
|
|
ret = __bpf_arch_text_poke(poke->tailcall_bypass,
|
|
|
|
BPF_MOD_JUMP,
|
|
|
|
poke->bypass_addr,
|
|
|
|
NULL);
|
|
|
|
BUG_ON(ret < 0);
|
|
|
|
}
|
|
|
|
} else {
|
|
|
|
ret = __bpf_arch_text_poke(poke->tailcall_bypass,
|
|
|
|
BPF_MOD_JUMP,
|
|
|
|
old_bypass_addr,
|
|
|
|
poke->bypass_addr);
|
|
|
|
BUG_ON(ret < 0);
|
|
|
|
/* let other CPUs finish the execution of program
|
|
|
|
* so that it will not possible to expose them
|
|
|
|
* to invalid nop, stack unwind, nop state
|
|
|
|
*/
|
|
|
|
if (!ret)
|
|
|
|
synchronize_rcu();
|
|
|
|
ret = __bpf_arch_text_poke(poke->tailcall_target,
|
|
|
|
BPF_MOD_JUMP,
|
|
|
|
old_addr, NULL);
|
|
|
|
BUG_ON(ret < 0);
|
|
|
|
}
|
|
|
|
}
|
2024-01-05 10:48:17 +00:00
|
|
|
|
bpf: Add x86-64 JIT support for bpf_addr_space_cast instruction.
LLVM generates bpf_addr_space_cast instruction while translating
pointers between native (zero) address space and
__attribute__((address_space(N))).
The addr_space=1 is reserved as bpf_arena address space.
rY = addr_space_cast(rX, 0, 1) is processed by the verifier and
converted to normal 32-bit move: wX = wY
rY = addr_space_cast(rX, 1, 0) has to be converted by JIT:
aux_reg = upper_32_bits of arena->user_vm_start
aux_reg <<= 32
wX = wY // clear upper 32 bits of dst register
if (wX) // if not zero add upper bits of user_vm_start
wX |= aux_reg
JIT can do it more efficiently:
mov dst_reg32, src_reg32 // 32-bit move
shl dst_reg, 32
or dst_reg, user_vm_start
rol dst_reg, 32
xor r11, r11
test dst_reg32, dst_reg32 // check if lower 32-bit are zero
cmove r11, dst_reg // if so, set dst_reg to zero
// Intel swapped src/dst register encoding in CMOVcc
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Signed-off-by: Andrii Nakryiko <andrii@kernel.org>
Acked-by: Eduard Zingerman <eddyz87@gmail.com>
Link: https://lore.kernel.org/bpf/20240308010812.89848-5-alexei.starovoitov@gmail.com
2024-03-08 01:08:02 +00:00
|
|
|
bool bpf_jit_supports_arena(void)
|
|
|
|
{
|
|
|
|
return true;
|
|
|
|
}
|
|
|
|
|
2024-01-05 10:48:17 +00:00
|
|
|
bool bpf_jit_supports_ptr_xchg(void)
|
|
|
|
{
|
|
|
|
return true;
|
|
|
|
}
|