]> Gentwo Git Trees - linux/.git/commit
x86/alternatives: Improve code-patching scalability by removing false sharing in...
authorEric Dumazet <edumazet@google.com>
Fri, 11 Apr 2025 05:40:13 +0000 (07:40 +0200)
committerIngo Molnar <mingo@kernel.org>
Fri, 11 Apr 2025 09:01:33 +0000 (11:01 +0200)
commit4334336e769bea1351ab82b22b06118c81bd217f
treeb4ac5ac52f3bd7e8f9964062489a6d306360ee57
parent0af2f6be1b4281385b618cb86ad946eded089ac8
x86/alternatives: Improve code-patching scalability by removing false sharing in poke_int3_handler()

eBPF programs can be run 50,000,000 times per second on busy servers.

Whenever /proc/sys/kernel/bpf_stats_enabled is turned off,
hundreds of calls sites are patched from text_poke_bp_batch()
and we see a huge loss of performance due to false sharing
on bp_desc.refs lasting up to three seconds.

   51.30%  server_bin       [kernel.kallsyms]           [k] poke_int3_handler
            |
            |--46.45%--poke_int3_handler
            |          exc_int3
            |          asm_exc_int3
            |          |
            |          |--24.26%--cls_bpf_classify
            |          |          tcf_classify
            |          |          __dev_queue_xmit
            |          |          ip6_finish_output2
            |          |          ip6_output
            |          |          ip6_xmit
            |          |          inet6_csk_xmit
            |          |          __tcp_transmit_skb

Fix this by replacing bp_desc.refs with a per-cpu bp_refs.

Before the patch, on a host with 240 cores (480 threads):

  $ sysctl -wq kernel.bpf_stats_enabled=0

  text_poke_bp_batch(nr_entries=164) : Took 2655300 usec

  $ bpftool prog | grep run_time_ns
  ...
  105: sched_cls  name hn_egress  tag 699fc5eea64144e3  gpl run_time_ns
  3009063719 run_cnt 82757845 : average cost is 36 nsec per call

After this patch:

  $ sysctl -wq kernel.bpf_stats_enabled=0

  text_poke_bp_batch(nr_entries=164) : Took 702 usec

  $ bpftool prog | grep run_time_ns
  ...
  105: sched_cls  name hn_egress  tag 699fc5eea64144e3  gpl run_time_ns
  1928223019 run_cnt 67682728 : average cost is 28 nsec per call

Ie. text-patching performance improved 3700x: from 2.65 seconds
to 0.0007 seconds.

Since the atomic_cond_read_acquire(refs, !VAL) spin-loop was not triggered
even once in my tests, add an unlikely() annotation, because this appears
to be the common case.

[ mingo: Improved the changelog some more. ]

Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Cc: Juergen Gross <jgross@suse.com>
Cc: "H . Peter Anvin" <hpa@zytor.com>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Brian Gerst <brgerst@gmail.com>
Cc: Kees Cook <keescook@chromium.org>
Cc: Josh Poimboeuf <jpoimboe@redhat.com>
Link: https://lore.kernel.org/r/20250411054105.2341982-2-mingo@kernel.org
arch/x86/kernel/alternative.c