linux-kernel - Re: [PATCH v3] x86/alternatives: remove false sharing in poke_int3

lists.openwall.net		lists / announce owl-users owl-dev john-users john-dev passwdqc-users yescrypt popa3d-users / oss-security kernel-hardening musl sabotage tlsify passwords / crypt-dev xvendor / Bugtraq Full-Disclosure linux-kernel linux-netdev linux-ext4 linux-hardening linux-cve-announce PHC
Open Source and information security mailing list archives
Hash Suite: Windows password security audit tool. GUI, reports in PDF.
[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]
Message-ID: <Z-JX5ImltdTFoFgr@gmail.com>
Date: Tue, 25 Mar 2025 08:14:44 +0100
From: Ingo Molnar <mingo@...nel.org>
To: Eric Dumazet <edumazet@...gle.com>
Cc: Thomas Gleixner <tglx@...utronix.de>, Ingo Molnar <mingo@...hat.com>,
	Borislav Petkov <bp@...en8.de>,
	Dave Hansen <dave.hansen@...ux.intel.com>,
	"H . Peter Anvin" <hpa@...or.com>,
	Peter Zijlstra <peterz@...radead.org>,
	Steven Rostedt <rostedt@...dmis.org>,
	linux-kernel <linux-kernel@...r.kernel.org>,
	Alexei Starovoitov <ast@...nel.org>,
	Daniel Borkmann <daniel@...earbox.net>,
	Masami Hiramatsu <mhiramat@...nel.org>, x86@...nel.org,
	bpf@...r.kernel.org, Eric Dumazet <eric.dumazet@...il.com>,
	Greg Thelen <gthelen@...gle.com>,
	Stephane Eranian <eranian@...gle.com>
Subject: Re: [PATCH v3] x86/alternatives: remove false sharing in
 poke_int3_handler()


* Eric Dumazet <edumazet@...gle.com> wrote:

> 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
> 
> Signed-off-by: Eric Dumazet <edumazet@...gle.com>
> ---
>  arch/x86/kernel/alternative.c | 30 ++++++++++++++++++------------
>  1 file changed, 18 insertions(+), 12 deletions(-)

Thanks for the updates. I've further improved the changelog (see 
attached below), and have tentatively applied it to 
tip:x86/alternatives.

Thanks,

	Ingo

==============================>
From: Eric Dumazet <edumazet@...gle.com>
Date: Tue, 25 Mar 2025 04:33:16 +0000
Subject: [PATCH] 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@...gle.com>
Signed-off-by: Ingo Molnar <mingo@...nel.org>
Cc: Brian Gerst <brgerst@...il.com>
Cc: Juergen Gross <jgross@...e.com>
Cc: H. Peter Anvin <hpa@...or.com>
Cc: Linus Torvalds <torvalds@...ux-foundation.org>
Cc: Kees Cook <keescook@...omium.org>
Cc: Peter Zijlstra <peterz@...radead.org>
Cc: Josh Poimboeuf <jpoimboe@...hat.com>
Link: https://lore.kernel.org/r/20250325043316.874518-1-edumazet@google.com