[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]
Message-ID: <CANn89i+fmyJ8p=vBpwBy38yhVMCJv8XjrTkrXSUnSGedboCM_Q@mail.gmail.com>
Date: Mon, 24 Mar 2025 04:59:12 +0100
From: Eric Dumazet <edumazet@...gle.com>
To: Ingo Molnar <mingo@...nel.org>
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] x86/alternatives: remove false sharing in poke_int3_handler()
On Sun, Mar 23, 2025 at 10:38 PM Ingo Molnar <mingo@...nel.org> wrote:
>
>
> * Eric Dumazet <edumazet@...gle.com> wrote:
>
> > eBPF programs can be run 20,000,000+ 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 critical loss of performance due to false sharing
> > on bp_desc.refs lasting up to three seconds.
>
> > @@ -2413,8 +2415,12 @@ static void text_poke_bp_batch(struct text_poke_loc *tp, unsigned int nr_entries
> > /*
> > * Remove and wait for refs to be zero.
> > */
> > - if (!atomic_dec_and_test(&bp_desc.refs))
> > - atomic_cond_read_acquire(&bp_desc.refs, !VAL);
> > + for_each_possible_cpu(i) {
> > + atomic_t *refs = per_cpu_ptr(&bp_refs, i);
> > +
> > + if (!atomic_dec_and_test(refs))
> > + atomic_cond_read_acquire(refs, !VAL);
> > + }
>
> So your patch changes text_poke_bp_batch() to busy-spin-wait for
> bp_refs to go to zero on all 480 CPUs.
>
> Your measurement is using /proc/sys/kernel/bpf_stats_enabled on a
> single CPU, right?
Yes, some daemon enables bpf_stats for a small amount of time (one
second) to gather stats
on eBPF run time costs. (bpftool prog | grep run_time)
One eBPF selftest can also do this.
tools/testing/selftests/bpf/prog_tests/enable_stats.c
>
> What's the adversarial workload here? Spamming bpf_stats_enabled on all
> CPUs in parallel? Or mixing it with some other text_poke_bp_batch()
> user if bpf_stats_enabled serializes access?
The workload is having ~480 cpus running various eBPF programs all
over the places,
In the perf bit I added in the changelog, we see an eBPF program
hooked at the xmit of each packet.
But the fd = bpf_enable_stats(BPF_STATS_RUN_TIME) / .... / close(fd)
only happens from time to time,
because of the supposed extra cost of fetching two extra time stamps.
BTW, before the patch stats on my test host look like
105: sched_cls name hn_egress tag 699fc5eea64144e3 gpl run_time_ns
3009063719 run_cnt 82757845
-> average cost is 36 nsec per call
And after the patch :
105: sched_cls name hn_egress tag 699fc5eea64144e3 gpl run_time_ns
1928223019 run_cnt 67682728
-> average cost is 28 nsec per call
>
> Does anything undesirable happen in that case?
The case of multiple threads trying to flip bpf_stats_enabled is
handled by bpf_stats_enabled_mutex.
Thanks !
>
> Thanks,
>
> Ingo
Powered by blists - more mailing lists