Do not sleep in poll() when the need_wakeup flag is set. When this
flag is set, the application needs to explicitly wake up the driver
with a syscall (poll, recvmsg, sendmsg, etc.) to guarantee that Rx
and/or Tx processing will be processed promptly. But the current code
in poll(), sleeps first then wakes up the driver. This means that no
driver processing will occur (baring any interrupts) until the timeout
has expired.
Fix this by checking the need_wakeup flag first and if set, wake the
driver and return to the application. Only if need_wakeup is not set
should the process sleep if there is a timeout set in the poll() call.
Fixes: 77cd0d7b3f ("xsk: add support for need_wakeup flag in AF_XDP rings")
Reported-by: Keith Wiles <keith.wiles@intel.com>
Signed-off-by: Magnus Karlsson <magnus.karlsson@intel.com>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Acked-by: Maciej Fijalkowski <maciej.fijalkowski@intel.com>
Link: https://lore.kernel.org/bpf/20211214102607.7677-1-magnus.karlsson@gmail.com
This is distracting really, let's make this simpler,
because many callers had to take care of this
by themselves, even if on x86 this adds more
code than really needed.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Fix a crash in the buffer pool allocator when a buffer is double
freed. It is possible to trigger this behavior not only from a faulty
driver, but also from user space like this: Create a zero-copy AF_XDP
socket. Load an XDP program that will issue XDP_DROP for all
packets. Put the same umem buffer into the fill ring multiple times,
then bind the socket and send some traffic. This will crash the kernel
as the XDP_DROP action triggers one call to xsk_buff_free()/xp_free()
for every packet dropped. Each call will add the corresponding buffer
entry to the free_list and increase the free_list_cnt. Some entries
will have been added multiple times due to the same buffer being
freed. The buffer allocation code will then traverse this broken list
and since the same buffer is in the list multiple times, it will try
to delete the same buffer twice from the list leading to a crash.
The fix for this is just to test that the buffer has not been added
before in xp_free(). If it has been, just return from the function and
do not put it in the free_list a second time.
Note that this bug was not present in the code before the commit
referenced in the Fixes tag. That code used one list entry per
allocated buffer, so multiple frees did not have any side effects. But
the commit below optimized the usage of the pool and only uses a
single entry per buffer in the umem, meaning that multiple
allocations/frees of the same buffer will also only use one entry,
thus leading to the problem.
Fixes: 47e4075df3 ("xsk: Batched buffer allocation for the pool")
Signed-off-by: Magnus Karlsson <magnus.karlsson@intel.com>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Acked-by: Björn Töpel <bjorn@kernel.org>
Link: https://lore.kernel.org/bpf/20211111075707.21922-1-magnus.karlsson@gmail.com
Fix a build error with clang in __xp_alloc():
[...]
net/xdp/xsk_buff_pool.c:465:15: error: variable 'xskb' is uninitialized
when used here [-Werror,-Wuninitialized]
xp_release(xskb);
^~~~
This is correctly detected by clang, but not gcc. In fact, the xp_release()
statement should not be there at all in the refactored code, just remove it.
Fixes: 94033cd8e7 ("xsk: Optimize for aligned case")
Reported-by: Nathan Chancellor <nathan@kernel.org>
Signed-off-by: Magnus Karlsson <magnus.karlsson@intel.com>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Acked-by: Maciej Fijalkowski <maciej.fijalkowski@intel.com>
Link: https://lore.kernel.org/bpf/20210929061403.8587-1-magnus.karlsson@gmail.com
Optimize for the aligned case by precomputing the parameter values of
the xdp_buff_xsk and xdp_buff structures in the heads array. We can do
this as the heads array size is equal to the number of chunks in the
umem for the aligned case. Then every entry in this array will reflect
a certain chunk/frame and can therefore be prepopulated with the
correct values and we can drop the use of the free_heads stack. Note
that it is not possible to allocate more buffers than what has been
allocated in the aligned case since each chunk can only contain a
single buffer.
We can unfortunately not do this in the unaligned case as one chunk
might contain multiple buffers. In this case, we keep the old scheme
of populating a heads entry every time it is used and using
the free_heads stack.
Also move xp_release() and xp_get_handle() to xsk_buff_pool.h. They
were for some reason in xsk.c even though they are buffer pool
operations.
Signed-off-by: Magnus Karlsson <magnus.karlsson@intel.com>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Link: https://lore.kernel.org/bpf/20210922075613.12186-7-magnus.karlsson@gmail.com
Add a new driver interface xsk_buff_alloc_batch() offering batched
buffer allocations to improve performance. The new interface takes
three arguments: the buffer pool to allocated from, a pointer to an
array of struct xdp_buff pointers which will contain pointers to the
allocated xdp_buffs, and an unsigned integer specifying the max number
of buffers to allocate. The return value is the actual number of
buffers that the allocator managed to allocate and it will be in the
range 0 <= N <= max, where max is the third parameter to the function.
u32 xsk_buff_alloc_batch(struct xsk_buff_pool *pool, struct xdp_buff **xdp,
u32 max);
A second driver interface is also introduced that need to be used in
conjunction with xsk_buff_alloc_batch(). It is a helper that sets the
size of struct xdp_buff and is used by the NIC Rx irq routine when
receiving a packet. This helper sets the three struct members data,
data_meta, and data_end. The two first ones is in the xsk_buff_alloc()
case set in the allocation routine and data_end is set when a packet
is received in the receive irq function. This unfortunately leads to
worse performance since the xdp_buff is touched twice with a long time
period in between leading to an extra cache miss. Instead, we fill out
the xdp_buff with all 3 fields at one single point in time in the
driver, when the size of the packet is known. Hence this helper. Note
that the driver has to use this helper (or set all three fields
itself) when using xsk_buff_alloc_batch(). xsk_buff_alloc() works as
before and does not require this.
void xsk_buff_set_size(struct xdp_buff *xdp, u32 size);
Signed-off-by: Magnus Karlsson <magnus.karlsson@intel.com>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Link: https://lore.kernel.org/bpf/20210922075613.12186-3-magnus.karlsson@gmail.com
Trivial conflict in net/netfilter/nf_tables_api.c.
Duplicate fix in tools/testing/selftests/net/devlink_port_split.py
- take the net-next version.
skmsg, and L4 bpf - keep the bpf code but remove the flags
and err params.
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
This patch introduces a function wrapper to call the sk_error_report
callback. That will prepare to add additional handling whenever
sk_error_report is called, for example to trace socket errors.
Signed-off-by: Alexander Aring <aahringo@redhat.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
XDP_REDIRECT works by a three-step process: the bpf_redirect() and
bpf_redirect_map() helpers will lookup the target of the redirect and store
it (along with some other metadata) in a per-CPU struct bpf_redirect_info.
Next, when the program returns the XDP_REDIRECT return code, the driver
will call xdp_do_redirect() which will use the information thus stored to
actually enqueue the frame into a bulk queue structure (that differs
slightly by map type, but shares the same principle). Finally, before
exiting its NAPI poll loop, the driver will call xdp_do_flush(), which will
flush all the different bulk queues, thus completing the redirect.
Pointers to the map entries will be kept around for this whole sequence of
steps, protected by RCU. However, there is no top-level rcu_read_lock() in
the core code; instead drivers add their own rcu_read_lock() around the XDP
portions of the code, but somewhat inconsistently as Martin discovered[0].
However, things still work because everything happens inside a single NAPI
poll sequence, which means it's between a pair of calls to
local_bh_disable()/local_bh_enable(). So Paul suggested[1] that we could
document this intention by using rcu_dereference_check() with
rcu_read_lock_bh_held() as a second parameter, thus allowing sparse and
lockdep to verify that everything is done correctly.
This patch does just that: we add an __rcu annotation to the map entry
pointers and remove the various comments explaining the NAPI poll assurance
strewn through devmap.c in favour of a longer explanation in filter.c. The
goal is to have one coherent documentation of the entire flow, and rely on
the RCU annotations as a "standard" way of communicating the flow in the
map code (which can additionally be understood by sparse and lockdep).
The RCU annotation replacements result in a fairly straight-forward
replacement where READ_ONCE() becomes rcu_dereference_check(), WRITE_ONCE()
becomes rcu_assign_pointer() and xchg() and cmpxchg() gets wrapped in the
proper constructs to cast the pointer back and forth between __rcu and
__kernel address space (for the benefit of sparse). The one complication is
that xskmap has a few constructions where double-pointers are passed back
and forth; these simply all gain __rcu annotations, and only the final
reference/dereference to the inner-most pointer gets changed.
With this, everything can be run through sparse without eliciting
complaints, and lockdep can verify correctness even without the use of
rcu_read_lock() in the drivers. Subsequent patches will clean these up from
the drivers.
[0] https://lore.kernel.org/bpf/20210415173551.7ma4slcbqeyiba2r@kafai-mbp.dhcp.thefacebook.com/
[1] https://lore.kernel.org/bpf/20210419165837.GA975577@paulmck-ThinkPad-P17-Gen-1/
Signed-off-by: Toke Høiland-Jørgensen <toke@redhat.com>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Link: https://lore.kernel.org/bpf/20210624160609.292325-6-toke@redhat.com
Fix broken Tx ring validation for AF_XDP. The commit under the Fixes
tag, fixed an off-by-one error in the validation but introduced
another error. Descriptors are now let through even if they straddle a
chunk boundary which they are not allowed to do in aligned mode. Worse
is that they are let through even if they straddle the end of the umem
itself, tricking the kernel to read data outside the allowed umem
region which might or might not be mapped at all.
Fix this by reintroducing the old code, but subtract the length by one
to fix the off-by-one error that the original patch was
addressing. The test chunk != chunk_end makes sure packets do not
straddle chunk boundraries. Note that packets of zero length are
allowed in the interface, therefore the test if the length is
non-zero.
Fixes: ac31565c21 ("xsk: Fix for xp_aligned_validate_desc() when len == chunk_size")
Signed-off-by: Magnus Karlsson <magnus.karlsson@intel.com>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Reviewed-by: Xuan Zhuo <xuanzhuo@linux.alibaba.com>
Acked-by: Björn Töpel <bjorn@kernel.org>
Link: https://lore.kernel.org/bpf/20210618075805.14412-1-magnus.karlsson@gmail.com
This patch adds two flags BPF_F_BROADCAST and BPF_F_EXCLUDE_INGRESS to
extend xdp_redirect_map for broadcast support.
With BPF_F_BROADCAST the packet will be broadcasted to all the interfaces
in the map. with BPF_F_EXCLUDE_INGRESS the ingress interface will be
excluded when do broadcasting.
When getting the devices in dev hash map via dev_map_hash_get_next_key(),
there is a possibility that we fall back to the first key when a device
was removed. This will duplicate packets on some interfaces. So just walk
the whole buckets to avoid this issue. For dev array map, we also walk the
whole map to find valid interfaces.
Function bpf_clear_redirect_map() was removed in
commit ee75aef23a ("bpf, xdp: Restructure redirect actions").
Add it back as we need to use ri->map again.
With test topology:
+-------------------+ +-------------------+
| Host A (i40e 10G) | ---------- | eno1(i40e 10G) |
+-------------------+ | |
| Host B |
+-------------------+ | |
| Host C (i40e 10G) | ---------- | eno2(i40e 10G) |
+-------------------+ | |
| +------+ |
| veth0 -- | Peer | |
| veth1 -- | | |
| veth2 -- | NS | |
| +------+ |
+-------------------+
On Host A:
# pktgen/pktgen_sample03_burst_single_flow.sh -i eno1 -d $dst_ip -m $dst_mac -s 64
On Host B(Intel(R) Xeon(R) CPU E5-2690 v3 @ 2.60GHz, 128G Memory):
Use xdp_redirect_map and xdp_redirect_map_multi in samples/bpf for testing.
All the veth peers in the NS have a XDP_DROP program loaded. The
forward_map max_entries in xdp_redirect_map_multi is modify to 4.
Testing the performance impact on the regular xdp_redirect path with and
without patch (to check impact of additional check for broadcast mode):
5.12 rc4 | redirect_map i40e->i40e | 2.0M | 9.7M
5.12 rc4 | redirect_map i40e->veth | 1.7M | 11.8M
5.12 rc4 + patch | redirect_map i40e->i40e | 2.0M | 9.6M
5.12 rc4 + patch | redirect_map i40e->veth | 1.7M | 11.7M
Testing the performance when cloning packets with the redirect_map_multi
test, using a redirect map size of 4, filled with 1-3 devices:
5.12 rc4 + patch | redirect_map multi i40e->veth (x1) | 1.7M | 11.4M
5.12 rc4 + patch | redirect_map multi i40e->veth (x2) | 1.1M | 4.3M
5.12 rc4 + patch | redirect_map multi i40e->veth (x3) | 0.8M | 2.6M
Signed-off-by: Hangbin Liu <liuhangbin@gmail.com>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Acked-by: Toke Høiland-Jørgensen <toke@redhat.com>
Acked-by: Martin KaFai Lau <kafai@fb.com>
Acked-by: John Fastabend <john.fastabend@gmail.com>
Acked-by: Jesper Dangaard Brouer <brouer@redhat.com>
Link: https://lore.kernel.org/bpf/20210519090747.1655268-3-liuhangbin@gmail.com
Use kvcalloc() instead of kcalloc() to support large umems with, on my
server, one million pages or more in the umem.
Reported-by: Dan Siemon <dan@coverfire.com>
Signed-off-by: Magnus Karlsson <magnus.karlsson@intel.com>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Acked-by: Björn Töpel <bjorn@kernel.org>
Acked-by: John Fastabend <john.fastabend@gmail.com>
Link: https://lore.kernel.org/bpf/20210521083301.26921-1-magnus.karlsson@gmail.com
When desc->len is equal to chunk_size, it is legal. But when the
xp_aligned_validate_desc() got chunk_end from desc->addr + desc->len
pointing to the next chunk during the check, it caused the check to
fail.
This problem was first introduced in bbff2f321a ("xsk: new descriptor
addressing scheme"). Later in 2b43470add ("xsk: Introduce AF_XDP buffer
allocation API") this piece of code was moved into the new function called
xp_aligned_validate_desc(). This function was then moved into xsk_queue.h
via 26062b185e ("xsk: Explicitly inline functions and move definitions").
Fixes: bbff2f321a ("xsk: new descriptor addressing scheme")
Signed-off-by: Xuan Zhuo <xuanzhuo@linux.alibaba.com>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Acked-by: Magnus Karlsson <magnus.karlsson@intel.com>
Link: https://lore.kernel.org/bpf/20210428094424.54435-1-xuanzhuo@linux.alibaba.com
DPDK default burst size is 32, however, kernel xsk sendto
syscall can not handle all 32 at one time, and return with
error.
So make kernel XDP socket batch size larger to avoid
unnecessary syscall fail and context switch which will help
to increase performance.
Signed-off-by: Li RongQing <lirongqing@baidu.com>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Acked-by: Magnus Karlsson <magnus.karlsson@intel.com>
Link: https://lore.kernel.org/bpf/1618378752-4191-1-git-send-email-lirongqing@baidu.com
The XDP_REDIRECT implementations for maps and non-maps are fairly
similar, but obviously need to take different code paths depending on
if the target is using a map or not. Today, the redirect targets for
XDP either uses a map, or is based on ifindex.
Here, the map type and id are added to bpf_redirect_info, instead of
the actual map. Map type, map item/ifindex, and the map_id (if any) is
passed to xdp_do_redirect().
For ifindex-based redirect, used by the bpf_redirect() XDP BFP helper,
a special map type/id are used. Map type of UNSPEC together with map id
equal to INT_MAX has the special meaning of an ifindex based
redirect. Note that valid map ids are 1 inclusive, INT_MAX exclusive
([1,INT_MAX[).
In addition to making the code easier to follow, using explicit type
and id in bpf_redirect_info has a slight positive performance impact
by avoiding a pointer indirection for the map type lookup, and instead
use the cacheline for bpf_redirect_info.
Since the actual map is not passed via bpf_redirect_info anymore, the
map lookup is only done in the BPF helper. This means that the
bpf_clear_redirect_map() function can be removed. The actual map item
is RCU protected.
The bpf_redirect_info flags member is not used by XDP, and not
read/written any more. The map member is only written to when
required/used, and not unconditionally.
Signed-off-by: Björn Töpel <bjorn.topel@intel.com>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Reviewed-by: Maciej Fijalkowski <maciej.fijalkowski@intel.com>
Acked-by: Jesper Dangaard Brouer <brouer@redhat.com>
Acked-by: Toke Høiland-Jørgensen <toke@redhat.com>
Link: https://lore.kernel.org/bpf/20210308112907.559576-3-bjorn.topel@gmail.com
Currently the bpf_redirect_map() implementation dispatches to the
correct map-lookup function via a switch-statement. To avoid the
dispatching, this change adds bpf_redirect_map() as a map
operation. Each map provides its bpf_redirect_map() version, and
correct function is automatically selected by the BPF verifier.
A nice side-effect of the code movement is that the map lookup
functions are now local to the map implementation files, which removes
one additional function call.
Signed-off-by: Björn Töpel <bjorn.topel@intel.com>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Acked-by: Jesper Dangaard Brouer <brouer@redhat.com>
Acked-by: Toke Høiland-Jørgensen <toke@redhat.com>
Link: https://lore.kernel.org/bpf/20210308112907.559576-2-bjorn.topel@gmail.com
Currently, the AF_XDP rings uses general smp_{r,w,}mb() barriers on
the kernel-side. On most modern architectures
load-acquire/store-release barriers perform better, and results in
simpler code for circular ring buffers.
This change updates the XDP socket rings to use
load-acquire/store-release barriers.
It is important to note that changing from the old smp_{r,w,}mb()
barriers, to load-acquire/store-release barriers does not break
compatibility. The old semantics work with the new one, and vice
versa.
As pointed out by "Documentation/memory-barriers.txt" in the "SMP
BARRIER PAIRING" section:
"General barriers pair with each other, though they also pair with
most other types of barriers, albeit without multicopy atomicity.
An acquire barrier pairs with a release barrier, but both may also
pair with other barriers, including of course general barriers."
How different barriers behaves and pairs is outlined in
"tools/memory-model/Documentation/cheatsheet.txt".
In order to make sure that compatibility is not broken, LKMM herd7
based litmus tests can be constructed and verified.
We generalize the XDP socket ring to a one entry ring, and create two
scenarios; One where the ring is full, where only the consumer can
proceed, followed by the producer. One where the ring is empty, where
only the producer can proceed, followed by the consumer. Each scenario
is then expanded to four different tests: general producer/general
consumer, general producer/acqrel consumer, acqrel producer/general
consumer, acqrel producer/acqrel consumer. In total eight tests.
The empty ring test:
C spsc-rb+empty
// Simple one entry ring:
// prod cons allowed action prod cons
// 0 0 => prod => 1 0
// 0 1 => cons => 0 0
// 1 0 => cons => 1 1
// 1 1 => prod => 0 1
{}
// We start at prod==0, cons==0, data==0, i.e. nothing has been
// written to the ring. From here only the producer can start, and
// should write 1. Afterwards, consumer can continue and read 1 to
// data. Can we enter state prod==1, cons==1, but consumer observed
// the incorrect value of 0?
P0(int *prod, int *cons, int *data)
{
... producer
}
P1(int *prod, int *cons, int *data)
{
... consumer
}
exists( 1:d=0 /\ prod=1 /\ cons=1 );
The full ring test:
C spsc-rb+full
// Simple one entry ring:
// prod cons allowed action prod cons
// 0 0 => prod => 1 0
// 0 1 => cons => 0 0
// 1 0 => cons => 1 1
// 1 1 => prod => 0 1
{ prod = 1; }
// We start at prod==1, cons==0, data==1, i.e. producer has
// written 0, so from here only the consumer can start, and should
// consume 0. Afterwards, producer can continue and write 1 to
// data. Can we enter state prod==0, cons==1, but consumer observed
// the write of 1?
P0(int *prod, int *cons, int *data)
{
... producer
}
P1(int *prod, int *cons, int *data)
{
... consumer
}
exists( 1:d=1 /\ prod=0 /\ cons=1 );
where P0 and P1 are:
P0(int *prod, int *cons, int *data)
{
int p;
p = READ_ONCE(*prod);
if (READ_ONCE(*cons) == p) {
WRITE_ONCE(*data, 1);
smp_wmb();
WRITE_ONCE(*prod, p ^ 1);
}
}
P0(int *prod, int *cons, int *data)
{
int p;
p = READ_ONCE(*prod);
if (READ_ONCE(*cons) == p) {
WRITE_ONCE(*data, 1);
smp_store_release(prod, p ^ 1);
}
}
P1(int *prod, int *cons, int *data)
{
int c;
int d = -1;
c = READ_ONCE(*cons);
if (READ_ONCE(*prod) != c) {
smp_rmb();
d = READ_ONCE(*data);
smp_mb();
WRITE_ONCE(*cons, c ^ 1);
}
}
P1(int *prod, int *cons, int *data)
{
int c;
int d = -1;
c = READ_ONCE(*cons);
if (smp_load_acquire(prod) != c) {
d = READ_ONCE(*data);
smp_store_release(cons, c ^ 1);
}
}
The full LKMM litmus tests are found at [1].
On x86-64 systems the l2fwd AF_XDP xdpsock sample performance
increases by 1%. This is mostly due to that the smp_mb() is removed,
which is a relatively expensive operation on these
platforms. Weakly-ordered platforms, such as ARM64 might benefit even
more.
[1] https://github.com/bjoto/litmus-xsk
Signed-off-by: Björn Töpel <bjorn.topel@intel.com>
Signed-off-by: Andrii Nakryiko <andrii@kernel.org>
Acked-by: Toke Høiland-Jørgensen <toke@redhat.com>
Link: https://lore.kernel.org/bpf/20210305094113.413544-2-bjorn.topel@gmail.com
This patch is used to construct skb based on page to save memory copy
overhead.
This function is implemented based on IFF_TX_SKB_NO_LINEAR. Only the
network card priv_flags supports IFF_TX_SKB_NO_LINEAR will use page to
directly construct skb. If this feature is not supported, it is still
necessary to copy data to construct skb.
---------------- Performance Testing ------------
The test environment is Aliyun ECS server.
Test cmd:
```
xdpsock -i eth0 -t -S -s <msg size>
```
Test result data:
size 64 512 1024 1500
copy 1916747 1775988 1600203 1440054
page 1974058 1953655 1945463 1904478
percent 3.0% 10.0% 21.58% 32.3%
Signed-off-by: Xuan Zhuo <xuanzhuo@linux.alibaba.com>
Signed-off-by: Alexander Lobakin <alobakin@pm.me>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Reviewed-by: Dust Li <dust.li@linux.alibaba.com>
Acked-by: Magnus Karlsson <magnus.karlsson@intel.com>
Acked-by: John Fastabend <john.fastabend@gmail.com>
Link: https://lore.kernel.org/bpf/20210218204908.5455-6-alobakin@pm.me
xsk_generic_xmit() allocates a new skb and then queues it for
xmitting. The size of new skb's headroom is desc->len, so it comes
to the driver/device with no reserved headroom and/or tailroom.
Lots of drivers need some headroom (and sometimes tailroom) to
prepend (and/or append) some headers or data, e.g. CPU tags,
device-specific headers/descriptors (LSO, TLS etc.), and if case
of no available space skb_cow_head() will reallocate the skb.
Reallocations are unwanted on fast-path, especially when it comes
to XDP, so generic XSK xmit should reserve the spaces declared in
dev->needed_headroom and dev->needed tailroom to avoid them.
Note on max(NET_SKB_PAD, L1_CACHE_ALIGN(dev->needed_headroom)):
Usually, output functions reserve LL_RESERVED_SPACE(dev), which
consists of dev->hard_header_len + dev->needed_headroom, aligned
by 16.
However, on XSK xmit hard header is already here in the chunk, so
hard_header_len is not needed. But it'd still be better to align
data up to cacheline, while reserving no less than driver requests
for headroom. NET_SKB_PAD here is to double-insure there will be
no reallocations even when the driver advertises no needed_headroom,
but in fact need it (not so rare case).
Fixes: 35fcde7f8d ("xsk: support for Tx")
Signed-off-by: Alexander Lobakin <alobakin@pm.me>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Acked-by: Magnus Karlsson <magnus.karlsson@intel.com>
Acked-by: John Fastabend <john.fastabend@gmail.com>
Link: https://lore.kernel.org/bpf/20210218204908.5455-5-alobakin@pm.me
Fold xp_assign_dev and __xp_assign_dev. The former directly calls the
latter.
Signed-off-by: Björn Töpel <bjorn.topel@intel.com>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Reviewed-by: Maciej Fijalkowski <maciej.fijalkowski@intel.com>
Link: https://lore.kernel.org/bpf/20210122105351.11751-3-bjorn.topel@gmail.com
The explicit_free parameter of the __xsk_rcv() function was used to
mark whether the call was via the generic XDP or the native XDP
path. Instead of clutter the code with if-statements and "true/false"
parameters which are hard to understand, simply move the explicit free
to the __xsk_map_redirect() which is always called from the native XDP
path.
Signed-off-by: Björn Töpel <bjorn.topel@intel.com>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Reviewed-by: Maciej Fijalkowski <maciej.fijalkowski@intel.com>
Link: https://lore.kernel.org/bpf/20210122105351.11751-2-bjorn.topel@gmail.com
The number of queues can change by other means, rather than ethtool. For
example, attaching an mqprio qdisc with num_tc > 1 leads to creating
multiple sets of TX queues, which may be then destroyed when mqprio is
deleted. If an AF_XDP socket is created while mqprio is active,
dev->_tx[queue_id].pool will be filled, but then real_num_tx_queues may
decrease with deletion of mqprio, which will mean that the pool won't be
NULLed, and a further increase of the number of TX queues may expose a
dangling pointer.
To avoid any potential misbehavior, this commit clears pool for RX and
TX queues, regardless of real_num_*_queues, still taking into
consideration num_*_queues to avoid overflows.
Fixes: 1c1efc2af1 ("xsk: Create and free buffer pool independently from umem")
Fixes: a41b4f3c58 ("xsk: simplify xdp_clear_umem_at_qid implementation")
Signed-off-by: Maxim Mikityanskiy <maximmi@mellanox.com>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Acked-by: Björn Töpel <bjorn.topel@intel.com>
Link: https://lore.kernel.org/bpf/20210118160333.333439-1-maximmi@mellanox.com
Rollback the reservation in the completion ring when we get a
NETDEV_TX_BUSY. When this error is received from the driver, we are
supposed to let the user application retry the transmit again. And in
order to do this, we need to roll back the failed send so it can be
retried. Unfortunately, we did not cancel the reservation we had made
in the completion ring. By not doing this, we actually make the
completion ring one entry smaller per NETDEV_TX_BUSY error we get, and
after enough of these errors the completion ring will be of size zero
and transmit will stop working.
Fix this by cancelling the reservation when we get a NETDEV_TX_BUSY
error.
Fixes: 642e450b6b ("xsk: Do not discard packet when NETDEV_TX_BUSY")
Reported-by: Xuan Zhuo <xuanzhuo@linux.alibaba.com>
Signed-off-by: Magnus Karlsson <magnus.karlsson@intel.com>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Acked-by: Björn Töpel <bjorn.topel@intel.com>
Link: https://lore.kernel.org/bpf/20201218134525.13119-3-magnus.karlsson@gmail.com
Fix a race when multiple sockets are simultaneously calling sendto()
when the completion ring is shared in the SKB case. This is the case
when you share the same netdev and queue id through the
XDP_SHARED_UMEM bind flag. The problem is that multiple processes can
be in xsk_generic_xmit() and call the backpressure mechanism in
xskq_prod_reserve(xs->pool->cq). As this is a shared resource in this
specific scenario, a race might occur since the rings are
single-producer single-consumer.
Fix this by moving the tx_completion_lock from the socket to the pool
as the pool is shared between the sockets that share the completion
ring. (The pool is not shared when this is not the case.) And then
protect the accesses to xskq_prod_reserve() with this lock. The
tx_completion_lock is renamed cq_lock to better reflect that it
protects accesses to the potentially shared completion ring.
Fixes: 35fcde7f8d ("xsk: support for Tx")
Reported-by: Xuan Zhuo <xuanzhuo@linux.alibaba.com>
Signed-off-by: Magnus Karlsson <magnus.karlsson@intel.com>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Acked-by: Björn Töpel <bjorn.topel@intel.com>
Link: https://lore.kernel.org/bpf/20201218134525.13119-2-magnus.karlsson@gmail.com
Fix a possible memory leak when a bind of an AF_XDP socket fails. When
the fill and completion rings are created, they are tied to the
socket. But when the buffer pool is later created at bind time, the
ownership of these two rings are transferred to the buffer pool as
they might be shared between sockets (and the buffer pool cannot be
created until we know what we are binding to). So, before the buffer
pool is created, these two rings are cleaned up with the socket, and
after they have been transferred they are cleaned up together with
the buffer pool.
The problem is that ownership was transferred before it was absolutely
certain that the buffer pool could be created and initialized
correctly and when one of these errors occurred, the fill and
completion rings did neither belong to the socket nor the pool and
where therefore leaked. Solve this by moving the ownership transfer
to the point where the buffer pool has been completely set up and
there is no way it can fail.
Fixes: 7361f9c3d7 ("xsk: Move fill and completion rings to buffer pool")
Reported-by: syzbot+cfa88ddd0655afa88763@syzkaller.appspotmail.com
Signed-off-by: Magnus Karlsson <magnus.karlsson@intel.com>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Acked-by: Björn Töpel <bjorn.topel@intel.com>
Link: https://lore.kernel.org/bpf/20201214085127.3960-1-magnus.karlsson@gmail.com
Daniel Borkmann says:
====================
pull-request: bpf-next 2020-12-14
1) Expose bpf_sk_storage_*() helpers to iterator programs, from Florent Revest.
2) Add AF_XDP selftests based on veth devs to BPF selftests, from Weqaar Janjua.
3) Support for finding BTF based kernel attach targets through libbpf's
bpf_program__set_attach_target() API, from Andrii Nakryiko.
4) Permit pointers on stack for helper calls in the verifier, from Yonghong Song.
5) Fix overflows in hash map elem size after rlimit removal, from Eric Dumazet.
6) Get rid of direct invocation of llc in BPF selftests, from Andrew Delgadillo.
7) Fix xsk_recvmsg() to reorder socket state check before access, from Björn Töpel.
8) Add new libbpf API helper to retrieve ring buffer epoll fd, from Brendan Jackman.
9) Batch of minor BPF selftest improvements all over the place, from Florian Lehner,
KP Singh, Jiri Olsa and various others.
* https://git.kernel.org/pub/scm/linux/kernel/git/bpf/bpf-next: (31 commits)
selftests/bpf: Add a test for ptr_to_map_value on stack for helper access
bpf: Permits pointers on stack for helper calls
libbpf: Expose libbpf ring_buffer epoll_fd
selftests/bpf: Add set_attach_target() API selftest for module target
libbpf: Support modules in bpf_program__set_attach_target() API
selftests/bpf: Silence ima_setup.sh when not running in verbose mode.
selftests/bpf: Drop the need for LLVM's llc
selftests/bpf: fix bpf_testmod.ko recompilation logic
samples/bpf: Fix possible hang in xdpsock with multiple threads
selftests/bpf: Make selftest compilation work on clang 11
selftests/bpf: Xsk selftests - adding xdpxceiver to .gitignore
selftests/bpf: Drop tcp-{client,server}.py from Makefile
selftests/bpf: Xsk selftests - Bi-directional Sockets - SKB, DRV
selftests/bpf: Xsk selftests - Socket Teardown - SKB, DRV
selftests/bpf: Xsk selftests - DRV POLL, NOPOLL
selftests/bpf: Xsk selftests - SKB POLL, NOPOLL
selftests/bpf: Xsk selftests framework
bpf: Only provide bpf_sock_from_file with CONFIG_NET
bpf: Return -ENOTSUPP when attaching to non-kernel BTF
xsk: Validate socket state in xsk_recvmsg, prior touching socket members
...
====================
Link: https://lore.kernel.org/r/20201214214316.20642-1-daniel@iogearbox.net
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
xdp_return_frame_bulk() needs to pass a xdp_buff
to __xdp_return().
strlcpy got converted to strscpy but here it makes no
functional difference, so just keep the right code.
Conflicts:
net/netfilter/nf_tables_api.c
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
In AF_XDP the socket state needs to be checked, prior touching the
members of the socket. This was not the case for the recvmsg
implementation. Fix that by moving the xsk_is_bound() call.
Fixes: 45a8668184 ("xsk: Add support for recvmsg()")
Reported-by: kernel test robot <oliver.sang@intel.com>
Signed-off-by: Björn Töpel <bjorn.topel@intel.com>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Acked-by: Magnus Karlsson <magnus.karlsson@intel.com>
Link: https://lore.kernel.org/bpf/20201207082008.132263-1-bjorn.topel@gmail.com
If force_zc is set, we should exit out with an error, not fall back to
copy mode.
Fixes: 921b68692a ("xsk: Enable sharing of dma mappings")
Reported-by: Hulk Robot <hulkci@huawei.com>
Signed-off-by: Zhang Changzhong <zhangchangzhong@huawei.com>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Acked-by: Magnus Karlsson <magnus.karlsson@intel.com>
Link: https://lore.kernel.org/bpf/1607077277-41995-1-git-send-email-zhangchangzhong@huawei.com
Alexei Starovoitov says:
====================
pull-request: bpf-next 2020-12-03
The main changes are:
1) Support BTF in kernel modules, from Andrii.
2) Introduce preferred busy-polling, from Björn.
3) bpf_ima_inode_hash() and bpf_bprm_opts_set() helpers, from KP Singh.
4) Memcg-based memory accounting for bpf objects, from Roman.
5) Allow bpf_{s,g}etsockopt from cgroup bind{4,6} hooks, from Stanislav.
* https://git.kernel.org/pub/scm/linux/kernel/git/bpf/bpf-next: (118 commits)
selftests/bpf: Fix invalid use of strncat in test_sockmap
libbpf: Use memcpy instead of strncpy to please GCC
selftests/bpf: Add fentry/fexit/fmod_ret selftest for kernel module
selftests/bpf: Add tp_btf CO-RE reloc test for modules
libbpf: Support attachment of BPF tracing programs to kernel modules
libbpf: Factor out low-level BPF program loading helper
bpf: Allow to specify kernel module BTFs when attaching BPF programs
bpf: Remove hard-coded btf_vmlinux assumption from BPF verifier
selftests/bpf: Add CO-RE relocs selftest relying on kernel module BTF
selftests/bpf: Add support for marking sub-tests as skipped
selftests/bpf: Add bpf_testmod kernel module for testing
libbpf: Add kernel module BTF support for CO-RE relocations
libbpf: Refactor CO-RE relocs to not assume a single BTF object
libbpf: Add internal helper to load BTF data by FD
bpf: Keep module's btf_data_size intact after load
bpf: Fix bpf_put_raw_tracepoint()'s use of __module_address()
selftests/bpf: Add Userspace tests for TCP_WINDOW_CLAMP
bpf: Adds support for setting window clamp
samples/bpf: Fix spelling mistake "recieving" -> "receiving"
bpf: Fix cold build of test_progs-no_alu32
...
====================
Link: https://lore.kernel.org/r/20201204021936.85653-1-alexei.starovoitov@gmail.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Do not use rlimit-based memory accounting for xskmap maps.
It has been replaced with the memcg-based memory accounting.
Signed-off-by: Roman Gushchin <guro@fb.com>
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Acked-by: Song Liu <songliubraving@fb.com>
Link: https://lore.kernel.org/bpf/20201201215900.3569844-31-guro@fb.com
Modify the tx writeable condition from the queue is not full to the
number of present tx queues is less than the half of the total number
of queues. Because the tx queue not full is a very short time, this will
cause a large number of EPOLLOUT events, and cause a large number of
process wake up.
Fixes: 35fcde7f8d ("xsk: support for Tx")
Signed-off-by: Xuan Zhuo <xuanzhuo@linux.alibaba.com>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Acked-by: Magnus Karlsson <magnus.karlsson@intel.com>
Link: https://lore.kernel.org/bpf/508fef55188d4e1160747ead64c6dcda36735880.1606555939.git.xuanzhuo@linux.alibaba.com
datagram_poll will judge the current socket status (EPOLLIN, EPOLLOUT)
based on the traditional socket information (eg: sk_wmem_alloc), but
this does not apply to xsk. So this patch uses sock_poll_wait instead of
datagram_poll, and the mask is calculated by xsk_poll.
Fixes: c497176cb2 ("xsk: add Rx receive functions and poll support")
Signed-off-by: Xuan Zhuo <xuanzhuo@linux.alibaba.com>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Acked-by: Magnus Karlsson <magnus.karlsson@intel.com>
Link: https://lore.kernel.org/bpf/e82f4697438cd63edbf271ebe1918db8261b7c09.1606555939.git.xuanzhuo@linux.alibaba.com
Add napi_id to the xdp_rxq_info structure, and make sure the XDP
socket pick up the napi_id in the Rx path. The napi_id is used to find
the corresponding NAPI structure for socket busy polling.
Signed-off-by: Björn Töpel <bjorn.topel@intel.com>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Acked-by: Ilias Apalodimas <ilias.apalodimas@linaro.org>
Acked-by: Michael S. Tsirkin <mst@redhat.com>
Acked-by: Tariq Toukan <tariqt@nvidia.com>
Link: https://lore.kernel.org/bpf/20201130185205.196029-7-bjorn.topel@gmail.com
Wire-up XDP socket busy-poll support for recvmsg() and sendmsg(). If
the XDP socket prefers busy-polling, make sure that no wakeup/IPI is
performed.
Signed-off-by: Björn Töpel <bjorn.topel@intel.com>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Acked-by: Magnus Karlsson <magnus.karlsson@intel.com>
Link: https://lore.kernel.org/bpf/20201130185205.196029-6-bjorn.topel@gmail.com
Add a check for need wake up in sendmsg(), so that if a user calls
sendmsg() when no wakeup is needed, do not trigger a wakeup.
To simplify the need wakeup check in the syscall, unconditionally
enable the need wakeup flag for Tx. This has a side-effect for poll();
If poll() is called for a socket without enabled need wakeup, a Tx
wakeup is unconditionally performed.
The wakeup matrix for AF_XDP now looks like:
need wakeup | poll() | sendmsg() | recvmsg()
------------+--------------+-------------+------------
disabled | wake Tx | wake Tx | nop
enabled | check flag; | check flag; | check flag;
| wake Tx/Rx | wake Tx | wake Rx
Signed-off-by: Björn Töpel <bjorn.topel@intel.com>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Acked-by: Magnus Karlsson <magnus.karlsson@intel.com>
Link: https://lore.kernel.org/bpf/20201130185205.196029-5-bjorn.topel@gmail.com
Add support for non-blocking recvmsg() to XDP sockets. Previously,
only sendmsg() was supported by XDP socket. Now, for symmetry and the
upcoming busy-polling support, recvmsg() is added.
Signed-off-by: Björn Töpel <bjorn.topel@intel.com>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Acked-by: Magnus Karlsson <magnus.karlsson@intel.com>
Link: https://lore.kernel.org/bpf/20201130185205.196029-4-bjorn.topel@gmail.com
The functions xsk_map_put() and xsk_map_inc() are simple wrappers and
as such, replace these functions with the functions bpf_map_inc() and
bpf_map_put() and remove some error testing code.
Signed-off-by: Zhu Yanjun <zyjzyj2000@gmail.com>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Acked-by: Magnus Karlsson <magnus.karlsson@intel.com>
Link: https://lore.kernel.org/bpf/1606402998-12562-1-git-send-email-yanjunz@nvidia.com
Commit 642e450b6b ("xsk: Do not discard packet when NETDEV_TX_BUSY")
addressed the problem that packets were discarded from the Tx AF_XDP
ring, when the driver returned NETDEV_TX_BUSY. Part of the fix was
bumping the skbuff reference count, so that the buffer would not be
freed by dev_direct_xmit(). A reference count larger than one means
that the skbuff is "shared", which is not the case.
If the "shared" skbuff is sent to the generic XDP receive path,
netif_receive_generic_xdp(), and pskb_expand_head() is entered the
BUG_ON(skb_shared(skb)) will trigger.
This patch adds a variant to dev_direct_xmit(), __dev_direct_xmit(),
where a user can select the skbuff free policy. This allows AF_XDP to
avoid bumping the reference count, but still keep the NETDEV_TX_BUSY
behavior.
Fixes: 642e450b6b ("xsk: Do not discard packet when NETDEV_TX_BUSY")
Reported-by: Yonghong Song <yhs@fb.com>
Signed-off-by: Björn Töpel <bjorn.topel@intel.com>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Link: https://lore.kernel.org/bpf/20201123175600.146255-1-bjorn.topel@gmail.com
Fix incorrect netdev reference count in xsk_bind operation. Incorrect
reference count of the device appears when a user calls bind with the
XDP_ZEROCOPY flag on an interface which does not support zero-copy.
In such a case, an error is returned but the reference count is not
decreased. This change fixes the fault, by decreasing the reference count
in case of such an error.
The problem being corrected appeared in '162c820ed896' for the first time,
and the code was moved to new file location over the time with commit
'c2d3d6a47462'. This specific patch applies to all version starting
from 'c2d3d6a47462'. The same solution should be applied but on different
file (net/xdp/xdp_umem.c) and function (xdp_umem_assign_dev) for versions
from '162c820ed896' to 'c2d3d6a47462' excluded.
Fixes: 162c820ed8 ("xdp: hold device for umem regardless of zero-copy mode")
Signed-off-by: Marek Majtyka <marekx.majtyka@intel.com>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Acked-by: Magnus Karlsson <magnus.karlsson@intel.com>
Link: https://lore.kernel.org/bpf/20201120151443.105903-1-marekx.majtyka@intel.com
Fix a bug that is triggered when a partially setup socket is
destroyed. For a fully setup socket, a socket that has been bound to a
device, the cleanup of the umem is performed at the end of the buffer
pool's cleanup work queue item. This has to be performed in a work
queue, and not in RCU cleanup, as it is doing a vunmap that cannot
execute in interrupt context. However, when a socket has only been
partially set up so that a umem has been created but the buffer pool
has not, the code erroneously directly calls the umem cleanup function
instead of using a work queue, and this leads to a BUG_ON() in
vunmap().
As there in this case is no buffer pool, we cannot use its work queue,
so we need to introduce a work queue for the umem and schedule this for
the cleanup. So in the case there is no pool, we are going to use the
umem's own work queue to schedule the cleanup. But if there is a
pool, the cleanup of the umem is still being performed by the pool's
work queue, as it is important that the umem is cleaned up after the
pool.
Fixes: e5e1a4bc91 ("xsk: Fix possible memory leak at socket close")
Reported-by: Marek Majtyka <marekx.majtyka@intel.com>
Signed-off-by: Magnus Karlsson <magnus.karlsson@intel.com>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Tested-by: Marek Majtyka <marekx.majtyka@intel.com>
Link: https://lore.kernel.org/bpf/1605873219-21629-1-git-send-email-magnus.karlsson@gmail.com
Introduce batched descriptor interfaces in the xsk core code for the
Tx path to be used in the driver to write a code path with higher
performance. This interface will be used by the i40e driver in the
next patch. Though other drivers would likely benefit from this new
interface too.
Note that batching is only implemented for the common case when
there is only one socket bound to the same device and queue id. When
this is not the case, we fall back to the old non-batched version of
the function.
Signed-off-by: Magnus Karlsson <magnus.karlsson@intel.com>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Acked-by: John Fastabend <john.fastabend@gmail.com>
Link: https://lore.kernel.org/bpf/1605525167-14450-5-git-send-email-magnus.karlsson@gmail.com
Introduce one cache line worth of padding between the consumer pointer
and the flags field as well as between the flags field and the start
of the descriptors in all the lockless rings. This so that the x86 HW
adjacency prefetcher will not prefetch the adjacent pointer/field when
only one pointer/field is going to be used. This improves throughput
performance for the l2fwd sample app with 1% on my machine with HW
prefetching turned on in the BIOS.
Signed-off-by: Magnus Karlsson <magnus.karlsson@intel.com>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Acked-by: John Fastabend <john.fastabend@gmail.com>
Link: https://lore.kernel.org/bpf/1605525167-14450-4-git-send-email-magnus.karlsson@gmail.com
Fix a possible memory leak at xsk socket close that is caused by the
refcounting of the umem object being wrong. The reference count of the
umem was decremented only after the pool had been freed. Note that if
the buffer pool is destroyed, it is important that the umem is
destroyed after the pool, otherwise the umem would disappear while the
driver is still running. And as the buffer pool needs to be destroyed
in a work queue, the umem is also (if its refcount reaches zero)
destroyed after the buffer pool in that same work queue.
What was missing is that the refcount also needs to be decremented
when the pool is not freed and when the pool has not even been
created. The first case happens when the refcount of the pool is
higher than 1, i.e. it is still being used by some other socket using
the same device and queue id. In this case, it is safe to decrement
the refcount of the umem outside of the work queue as the umem will
never be freed because the refcount of the umem is always greater than
or equal to the refcount of the buffer pool. The second case is if the
buffer pool has not been created yet, i.e. the socket was closed
before it was bound but after the umem was created. In this case, it
is safe to destroy the umem outside of the work queue, since there is
no pool that can use it by definition.
Fixes: 1c1efc2af1 ("xsk: Create and free buffer pool independently from umem")
Reported-by: syzbot+eb71df123dc2be2c1456@syzkaller.appspotmail.com
Signed-off-by: Magnus Karlsson <magnus.karlsson@intel.com>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Acked-by: Björn Töpel <bjorn.topel@intel.com>
Link: https://lore.kernel.org/bpf/1603801921-2712-1-git-send-email-magnus.karlsson@gmail.com
Alexei Starovoitov says:
====================
pull-request: bpf-next 2020-10-12
The main changes are:
1) The BPF verifier improvements to track register allocation pattern, from Alexei and Yonghong.
2) libbpf relocation support for different size load/store, from Andrii.
3) bpf_redirect_peer() helper and support for inner map array with different max_entries, from Daniel.
4) BPF support for per-cpu variables, form Hao.
5) sockmap improvements, from John.
====================
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Recent work in f4d0525921 ("bpf: Add map_meta_equal map ops") and 134fede4ee
("bpf: Relax max_entries check for most of the inner map types") added support
for dynamic inner max elements for most map-in-map types. Exceptions were maps
like array or prog array where the map_gen_lookup() callback uses the maps'
max_entries field as a constant when emitting instructions.
We recently implemented Maglev consistent hashing into Cilium's load balancer
which uses map-in-map with an outer map being hash and inner being array holding
the Maglev backend table for each service. This has been designed this way in
order to reduce overall memory consumption given the outer hash map allows to
avoid preallocating a large, flat memory area for all services. Also, the
number of service mappings is not always known a-priori.
The use case for dynamic inner array map entries is to further reduce memory
overhead, for example, some services might just have a small number of back
ends while others could have a large number. Right now the Maglev backend table
for small and large number of backends would need to have the same inner array
map entries which adds a lot of unneeded overhead.
Dynamic inner array map entries can be realized by avoiding the inlined code
generation for their lookup. The lookup will still be efficient since it will
be calling into array_map_lookup_elem() directly and thus avoiding retpoline.
The patch adds a BPF_F_INNER_MAP flag to map creation which therefore skips
inline code generation and relaxes array_map_meta_equal() check to ignore both
maps' max_entries. This also still allows to have faster lookups for map-in-map
when BPF_F_INNER_MAP is not specified and hence dynamic max_entries not needed.
Example code generation where inner map is dynamic sized array:
# bpftool p d x i 125
int handle__sys_enter(void * ctx):
; int handle__sys_enter(void *ctx)
0: (b4) w1 = 0
; int key = 0;
1: (63) *(u32 *)(r10 -4) = r1
2: (bf) r2 = r10
;
3: (07) r2 += -4
; inner_map = bpf_map_lookup_elem(&outer_arr_dyn, &key);
4: (18) r1 = map[id:468]
6: (07) r1 += 272
7: (61) r0 = *(u32 *)(r2 +0)
8: (35) if r0 >= 0x3 goto pc+5
9: (67) r0 <<= 3
10: (0f) r0 += r1
11: (79) r0 = *(u64 *)(r0 +0)
12: (15) if r0 == 0x0 goto pc+1
13: (05) goto pc+1
14: (b7) r0 = 0
15: (b4) w6 = -1
; if (!inner_map)
16: (15) if r0 == 0x0 goto pc+6
17: (bf) r2 = r10
;
18: (07) r2 += -4
; val = bpf_map_lookup_elem(inner_map, &key);
19: (bf) r1 = r0 | No inlining but instead
20: (85) call array_map_lookup_elem#149280 | call to array_map_lookup_elem()
; return val ? *val : -1; | for inner array lookup.
21: (15) if r0 == 0x0 goto pc+1
; return val ? *val : -1;
22: (61) r6 = *(u32 *)(r0 +0)
; }
23: (bc) w0 = w6
24: (95) exit
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Acked-by: Andrii Nakryiko <andrii@kernel.org>
Link: https://lore.kernel.org/bpf/20201010234006.7075-4-daniel@iogearbox.net
Introduce one cache line worth of padding between the producer and
consumer pointers in all the lockless rings. This so that the HW
adjacency prefetcher will not prefetch the consumer pointer when the
producer pointer is used and vice versa. This improves throughput
performance for the l2fwd sample app with 2% on my machine with HW
prefetching turned on.
Signed-off-by: Magnus Karlsson <magnus.karlsson@intel.com>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Link: https://lore.kernel.org/bpf/1602166338-21378-1-git-send-email-magnus.karlsson@gmail.com
Jakub Kicinski