This patch fix bug reported in https://bugzilla.kernel.org/show_bug.cgi?id=73331,
after the patch http://www.spinics.net/lists/kvm/msg105230.html applied, there is
some progress and the L2 can boot up, however, slowly. The original idea of this
fix vid injection patch is from "Zhang, Yang Z" <yang.z.zhang@intel.com>.
Interrupt which delivered by vid should be injected to L1 by L0 if current is in
L1, or should be injected to L2 by L0 through the old injection way if L1 doesn't
have set External-interrupt exiting bit. The current logic doen't consider these
cases. This patch fix it by vid intr to L1 if current is L1 or L2 through old
injection way if L1 doen't have External-interrupt exiting bit set.
Signed-off-by: Wanpeng Li <wanpeng.li@linux.intel.com>
Signed-off-by: "Zhang, Yang Z" <yang.z.zhang@intel.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Certain instructions (e.g., mwait and monitor) cause a #UD exception when they
are executed in user mode. This is in contrast to the regular privileged
instructions which cause #GP. In order not to mess with SVM interception of
mwait and monitor which assumes privilege level assertions take place before
interception, a flag has been added.
Signed-off-by: Nadav Amit <namit@cs.technion.ac.il>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Certain instructions, such as monitor and xsave do not support big real mode
and cause a #GP exception if any of the accessed bytes effective address are
not within [0, 0xffff]. This patch introduces a flag to mark these
instructions, including the necassary checks.
Signed-off-by: Nadav Amit <namit@cs.technion.ac.il>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Emulator accesses are always done a page at a time, either by the emulator
itself (for fetches) or because we need to query the MMU for address
translations. Speed up these accesses by using kvm_read_guest_page
and, in the case of fetches, by inlining kvm_read_guest_virt_helper and
dropping the loop around kvm_read_guest_page.
This final tweak saves 30-100 more clock cycles (4-10%), bringing the
count (as measured by kvm-unit-tests) down to 720-1100 clock cycles on
a Sandy Bridge Xeon host, compared to 2300-3200 before the whole series
and 925-1700 after the first two low-hanging fruit changes.
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
When the CS base is not page-aligned, the linear address of the code could
get close to the page boundary (e.g. 0x...ffe) even if the EIP value is
not. So we need to first linearize the address, and only then compute
the number of valid bytes that can be fetched.
This happens relatively often when executing real mode code.
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
We do not need a memory copying loop anymore in insn_fetch; we
can use a byte-aligned pointer to access instruction fields directly
from the fetch_cache. This eliminates 50-150 cycles (corresponding to
a 5-10% improvement in performance) from each instruction.
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
do_insn_fetch_bytes will only be called once in a given insn_fetch and
insn_fetch_arr, because in fact it will only be called at most twice
for any instruction and the first call is explicit in x86_decode_insn.
This observation lets us hoist the call out of the memory copying loop.
It does not buy performance, because most fetches are one byte long
anyway, but it prepares for the next patch.
The overflow check is tricky, but correct. Because do_insn_fetch_bytes
has already been called once, we know that fc->end is at least 15. So
it is okay to subtract the number of bytes we want to read.
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Hoist the common case up from do_insn_fetch_byte to do_insn_fetch,
and prime the fetch_cache in x86_decode_insn. This helps a bit the
compiler and the branch predictor, but above all it lays the
ground for further changes in the next few patches.
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
rip_relative is only set if decode_modrm runs, and if you have ModRM
you will also have a memopp. We can then access memopp unconditionally.
Note that rip_relative cannot be hoisted up to decode_modrm, or you
break "mov $0, xyz(%rip)".
Also, move typecast on "out of range value" of mem.ea to decode_modrm.
Together, all these optimizations save about 50 cycles on each emulated
instructions (4-6%).
Signed-off-by: Bandan Das <bsd@redhat.com>
[Fix immediate operands with rip-relative addressing. - Paolo]
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
x86_decode_insn already sets a default for seg_override,
so remove it from the zeroed area. Also replace set/get functions
with direct access to the field.
Signed-off-by: Bandan Das <bsd@redhat.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
A lot of initializations are unnecessary as they get set to
appropriate values before actually being used. Optimize
placement of fields in x86_emulate_ctxt
Signed-off-by: Bandan Das <bsd@redhat.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Remove the if conditional - that will help us avoid
an "else initialize to 0" Also, rearrange operators
for slightly better code.
Signed-off-by: Bandan Das <bsd@redhat.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
The same information can be gleaned from ctxt->d and avoids having
to zero/NULL initialize intercept and check_perm
Signed-off-by: Bandan Das <bsd@redhat.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Core emulator functions all belong in emulator.c,
x86 should have no knowledge of emulator internals
Signed-off-by: Bandan Das <bsd@redhat.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
The "if/return" checks are useless, because we return X86EMUL_CONTINUE
anyway if we do not return.
Reviewed-by: Marcelo Tosatti <mtosatti@redhat.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
We can just blindly move all 16 bytes of ctxt->src's value to ctxt->dst.
write_register_operand will take care of writing only the lower bytes.
Avoiding a call to memcpy (the compiler optimizes it out) gains about
200 cycles on kvm-unit-tests for register-to-register moves, and makes
them about as fast as arithmetic instructions.
We could perhaps get a larger speedup by moving all instructions _except_
moves out of x86_emulate_insn, removing opcode_len, and replacing the
switch statement with an inlined em_mov.
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
There are several checks for "peculiar" aspects of instructions in both
x86_decode_insn and x86_emulate_insn. Group them together, and guard
them with a single "if" that lets the processor quickly skip them all.
Make this more effective by adding two more flag bits that say whether the
.intercept and .check_perm fields are valid. We will reuse these
flags later to avoid initializing fields of the emulate_ctxt struct.
This skims about 30 cycles for each emulated instructions, which is
approximately a 3% improvement.
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Despite the provisions to emulate up to 130 consecutive instructions, in
practice KVM will emulate just one before exiting handle_invalid_guest_state,
because x86_emulate_instruction always sets KVM_REQ_EVENT.
However, we only need to do this if an interrupt could be injected,
which happens a) if an interrupt shadow bit (STI or MOV SS) has gone
away; b) if the interrupt flag has just been set (other instructions
than STI can set it without enabling an interrupt shadow).
This cuts another 700-900 cycles from the cost of emulating an
instruction (measured on a Sandy Bridge Xeon: 1650-2600 cycles
before the patch on kvm-unit-tests, 925-1700 afterwards).
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
For the next patch we will need to know the full state of the
interrupt shadow; we will then set KVM_REQ_EVENT when one bit
is cleared.
However, right now get_interrupt_shadow only returns the one
corresponding to the emulated instruction, or an unconditional
0 if the emulated instruction does not have an interrupt shadow.
This is confusing and does not allow us to check for cleared
bits as mentioned above.
Clean the callback up, and modify toggle_interruptibility to
match the comment above the call. As a small result, the
call to set_interrupt_shadow will be skipped in the common
case where int_shadow == 0 && mask == 0.
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
About 25% of the time spent in emulation of invalid guest state
is wasted in checking whether emulation is required for the next
instruction. However, this almost never changes except when a
segment register (or TR or LDTR) changes, or when there is a mode
transition (i.e. CR0 changes).
In fact, vmx_set_segment and vmx_set_cr0 already modify
vmx->emulation_required (except that the former for some reason
uses |= instead of just an assignment). So there is no need to
call guest_state_valid in the emulation loop.
Emulation performance test results indicate 1650-2600 cycles
for common instructions, versus 2300-3200 before this patch on
a Sandy Bridge Xeon.
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Since commit 575203 the MCE subsystem in the Linux kernel for AMD sets bit 18
in MSR_K7_HWCR. Running such a kernel as a guest in KVM on an AMD host results
in a GPE injected into the guest because kvm_set_msr_common returns 1. This
patch fixes this by masking bit 18 from the MSR value desired by the guest.
Signed-off-by: Matthias Lange <matthias.lange@kernkonzept.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
We encountered a scenario in which after an INIT is delivered, a pending
interrupt is delivered, although it was sent before the INIT. As the SDM
states in section 10.4.7.1, the ISR and the IRR should be cleared after INIT as
KVM does. This also means that pending interrupts should be cleared. This
patch clears upon reset (and INIT) the pending interrupts; and at the same
occassion clears the pending exceptions, since they may cause a similar issue.
Signed-off-by: Nadav Amit <namit@cs.technion.ac.il>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
We have noticed that qemu-kvm hangs early in the BIOS when runnning nested
under some versions of VMware ESXi.
The problem we believe is because KVM assumes that the platform preserves
the 'G' but for any segment register. The SVM specification itemizes the
segment attribute bits that are observed by the CPU, but the (G)ranularity bit
is not one of the bits itemized, for any segment. Though current AMD CPUs keep
track of the (G)ranularity bit for all segment registers other than CS, the
specification does not require it. VMware's virtual CPU may not track the
(G)ranularity bit for any segment register.
Since kvm already synthesizes the (G)ranularity bit for the CS segment. It
should do so for all segments. The patch below does that, and helps get rid of
the hangs. Patch applies on top of Linus' tree.
Signed-off-by: Jim Mattson <jmattson@vmware.com>
Signed-off-by: Alok N Kataria <akataria@vmware.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
In two cases lapic.c does not use the apic_debug macro correctly. This patch
fixes them.
Signed-off-by: Nadav Amit <namit@cs.technion.ac.il>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
I've observed kvmclock being marked as unstable on a modern
single-socket system with a stable TSC and qemu-1.6.2 or qemu-2.0.0.
The culprit was failure in TSC matching because of overflow of
kvm_arch::nr_vcpus_matched_tsc in case there were multiple TSC writes
in a single synchronization cycle.
Turns out that qemu does multiple TSC writes during init, below is the
evidence of that (qemu-2.0.0):
The first one:
0xffffffffa08ff2b4 : vmx_write_tsc_offset+0xa4/0xb0 [kvm_intel]
0xffffffffa04c9c05 : kvm_write_tsc+0x1a5/0x360 [kvm]
0xffffffffa04cfd6b : kvm_arch_vcpu_postcreate+0x4b/0x80 [kvm]
0xffffffffa04b8188 : kvm_vm_ioctl+0x418/0x750 [kvm]
The second one:
0xffffffffa08ff2b4 : vmx_write_tsc_offset+0xa4/0xb0 [kvm_intel]
0xffffffffa04c9c05 : kvm_write_tsc+0x1a5/0x360 [kvm]
0xffffffffa090610d : vmx_set_msr+0x29d/0x350 [kvm_intel]
0xffffffffa04be83b : do_set_msr+0x3b/0x60 [kvm]
0xffffffffa04c10a8 : msr_io+0xc8/0x160 [kvm]
0xffffffffa04caeb6 : kvm_arch_vcpu_ioctl+0xc86/0x1060 [kvm]
0xffffffffa04b6797 : kvm_vcpu_ioctl+0xc7/0x5a0 [kvm]
#0 kvm_vcpu_ioctl at /build/buildd/qemu-2.0.0+dfsg/kvm-all.c:1780
#1 kvm_put_msrs at /build/buildd/qemu-2.0.0+dfsg/target-i386/kvm.c:1270
#2 kvm_arch_put_registers at /build/buildd/qemu-2.0.0+dfsg/target-i386/kvm.c:1909
#3 kvm_cpu_synchronize_post_init at /build/buildd/qemu-2.0.0+dfsg/kvm-all.c:1641
#4 cpu_synchronize_post_init at /build/buildd/qemu-2.0.0+dfsg/include/sysemu/kvm.h:330
#5 cpu_synchronize_all_post_init () at /build/buildd/qemu-2.0.0+dfsg/cpus.c:521
#6 main at /build/buildd/qemu-2.0.0+dfsg/vl.c:4390
The third one:
0xffffffffa08ff2b4 : vmx_write_tsc_offset+0xa4/0xb0 [kvm_intel]
0xffffffffa04c9c05 : kvm_write_tsc+0x1a5/0x360 [kvm]
0xffffffffa090610d : vmx_set_msr+0x29d/0x350 [kvm_intel]
0xffffffffa04be83b : do_set_msr+0x3b/0x60 [kvm]
0xffffffffa04c10a8 : msr_io+0xc8/0x160 [kvm]
0xffffffffa04caeb6 : kvm_arch_vcpu_ioctl+0xc86/0x1060 [kvm]
0xffffffffa04b6797 : kvm_vcpu_ioctl+0xc7/0x5a0 [kvm]
#0 kvm_vcpu_ioctl at /build/buildd/qemu-2.0.0+dfsg/kvm-all.c:1780
#1 kvm_put_msrs at /build/buildd/qemu-2.0.0+dfsg/target-i386/kvm.c:1270
#2 kvm_arch_put_registers at /build/buildd/qemu-2.0.0+dfsg/target-i386/kvm.c:1909
#3 kvm_cpu_synchronize_post_reset at /build/buildd/qemu-2.0.0+dfsg/kvm-all.c:1635
#4 cpu_synchronize_post_reset at /build/buildd/qemu-2.0.0+dfsg/include/sysemu/kvm.h:323
#5 cpu_synchronize_all_post_reset () at /build/buildd/qemu-2.0.0+dfsg/cpus.c:512
#6 main at /build/buildd/qemu-2.0.0+dfsg/vl.c:4482
The fix is to count each vCPU only once when matched, so that
nr_vcpus_matched_tsc holds the size of the matched set. This is
achieved by reusing generation counters. Every vCPU with
this_tsc_generation == cur_tsc_generation is in the matched set. The
match set is cleared by setting cur_tsc_generation to a value which no
other vCPU is set to (by incrementing it).
I needed to bump up the counter size form u8 to u64 to ensure it never
overflows. Otherwise in cases TSC is not written the same number of
times on each vCPU the counter could overflow and incorrectly indicate
some vCPUs as being in the matched set. This scenario seems unlikely
but I'm not sure if it can be disregarded.
Signed-off-by: Tomasz Grabiec <tgrabiec@cloudius-systems.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Obtaining the port number from DX is bogus as a) there are immediate
port accesses and b) user space may have changed the register content
while processing the PIO access. Forward the correct value from the
instruction emulator instead.
Signed-off-by: Jan Kiszka <jan.kiszka@siemens.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
The access size of an in/ins is reported in dst_bytes, and that of
out/outs in src_bytes.
Signed-off-by: Jan Kiszka <jan.kiszka@siemens.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
First, kvm_read_guest returns 0 on success. And then we need to take the
access size into account when testing the bitmap: intercept if any of
bits corresponding to the access is set.
Signed-off-by: Jan Kiszka <jan.kiszka@siemens.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
CLTS only changes TS which is not monitored by selected CR0
interception. So skip any attempt to translate WRITE_CR0 to
CR0_SEL_WRITE for this instruction.
Signed-off-by: Jan Kiszka <jan.kiszka@siemens.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
A struct member variable is set to the same value more than once
This was found using a static code analysis program called cppcheck.
Signed-off-by: Rickard Strandqvist <rickard_strandqvist@spectrumdigital.se>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
We import the CPL via SS.DPL since ae9fedc793. However, we fail to
export it this way so far. This caused spurious guest crashes, e.g. of
Linux when accessing the vmport from guest user space which triggered
register saving/restoring to/from host user space.
Signed-off-by: Jan Kiszka <jan.kiszka@siemens.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
VMX instructions use 32-bit operands in 32-bit mode, and 64-bit operands in
64-bit mode. The current implementation is broken since it does not use the
register operands correctly, and always uses 64-bit for reads and writes.
Moreover, write to memory in vmwrite only considers long-mode, so it ignores
cs.l. This patch fixes this behavior.
Signed-off-by: Nadav Amit <namit@cs.technion.ac.il>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
On 32-bit mode only bits [31:0] of the CR should be used for setting the CR
value. Otherwise, the host may incorrectly assume the value is invalid if bits
[63:32] are not zero. Moreover, the CR is currently being read twice when CR8
is used. Last, nested mov-cr exiting is modified to handle the CR value
correctly as well.
Signed-off-by: Nadav Amit <namit@cs.technion.ac.il>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Currently, the hypercall handling routine only considers LME as an indication
to whether the guest uses 32/64-bit mode. This is incosistent with hyperv
hypercalls handling and against the common sense of considering cs.l as well.
This patch uses is_64_bit_mode instead of is_long_mode for that matter. In
addition, the result is masked in respect to the guest execution mode. Last, it
changes kvm_hv_hypercall to use is_64_bit_mode as well to simplify the code.
Signed-off-by: Nadav Amit <namit@cs.technion.ac.il>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
When the guest sets DR6 and DR7, KVM asserts the high 32-bits are clear, and
otherwise injects a #GP exception. This exception should only be injected only
if running in long-mode.
Signed-off-by: Nadav Amit <namit@cs.technion.ac.il>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Many real CPUs get this wrong as well, but ours is totally off: bits 9:1
define the highest index value.
Signed-off-by: Jan Kiszka <jan.kiszka@siemens.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Allow L1 to "leak" its debug controls into L2, i.e. permit cleared
VM_{ENTRY_LOAD,EXIT_SAVE}_DEBUG_CONTROLS. This requires to manually
transfer the state of DR7 and IA32_DEBUGCTLMSR from L1 into L2 as both
run on different VMCS.
Signed-off-by: Jan Kiszka <jan.kiszka@siemens.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
SDM says bits 1, 4-6, 8, 13-16, and 26 have to be set.
Signed-off-by: Jan Kiszka <jan.kiszka@siemens.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
We already have this control enabled by exposing a broken
MSR_IA32_VMX_PROCBASED_CTLS value. This will properly advertise our
capability once the value is fixed by clearing the right bits in
MSR_IA32_VMX_TRUE_PROCBASED_CTLS. We also have to ensure to test the
right value on L2 entry.
Signed-off-by: Jan Kiszka <jan.kiszka@siemens.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
We already implemented them but failed to advertise them. Currently they
all return the identical values to the capability MSRs they are
augmenting. So there is no change in exposed features yet.
Drop related comments at this chance that are partially incorrect and
redundant anyway.
Signed-off-by: Jan Kiszka <jan.kiszka@siemens.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
The spec says those controls are at bit position 2 - makes 4 as value.
The impact of this mistake is effectively zero as we only use them to
ensure that these features are set at position 2 (or, previously, 1) in
MSR_IA32_VMX_{EXIT,ENTRY}_CTLS - which is and will be always true
according to the spec.
Signed-off-by: Jan Kiszka <jan.kiszka@siemens.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
On long-mode the current NOP (0x90) emulation still writes back to RAX. As a
result, EAX is zero-extended and the high 32-bits of RAX are cleared.
Signed-off-by: Nadav Amit <namit@cs.technion.ac.il>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Even if the condition of cmov is not satisfied, bits[63:32] should be cleared.
This is clearly stated in Intel's CMOVcc documentation. The solution is to
reassign the destination onto itself if the condition is unsatisfied. For that
matter the original destination value needs to be read.
Signed-off-by: Nadav Amit <namit@cs.technion.ac.il>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Return unhandlable error on inter-privilege level ret instruction. This is
since the current emulation does not check the privilege level correctly when
loading the CS, and does not pop RSP/SS as needed.
Cc: stable@vger.kernel.org
Signed-off-by: Nadav Amit <namit@cs.technion.ac.il>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
The emulator does not emulate the xadd instruction correctly if the two
operands are the same. In this (unlikely) situation the result should be the
sum of X and X (2X) when it is currently X. The solution is to first perform
writeback to the source, before writing to the destination. The only
instruction which should be affected is xadd, as the other instructions that
perform writeback to the source use the extended accumlator (e.g., RAX:RDX).
Signed-off-by: Nadav Amit <namit@cs.technion.ac.il>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
The current emulation of bit operations ignores the offset from the destination
on 64-bit target memory operands. This patch fixes this behavior.
Signed-off-by: Nadav Amit <namit@cs.technion.ac.il>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
use mm.h definition
Cc: Gleb Natapov <gleb@kernel.org>
Cc: Paolo Bonzini <pbonzini@redhat.com>
Signed-off-by: Fabian Frederick <fabf@skynet.be>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
