[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]
Message-ID: <CAH0uvojGoLX6mpK9wA1cw-EO-y_fUmdndAU8eZ1pa70Lc_rvvw@mail.gmail.com>
Date: Fri, 30 May 2025 17:00:38 -0700
From: Howard Chu <howardchu95@...il.com>
To: Namhyung Kim <namhyung@...nel.org>
Cc: Steven Rostedt <rostedt@...dmis.org>, Mathieu Desnoyers <mathieu.desnoyers@...icios.com>, acme@...nel.org,
mingo@...hat.com, mark.rutland@....com, alexander.shishkin@...ux.intel.com,
jolsa@...nel.org, irogers@...gle.com, adrian.hunter@...el.com,
peterz@...radead.org, kan.liang@...ux.intel.com,
linux-perf-users@...r.kernel.org, linux-kernel@...r.kernel.org,
Song Liu <song@...nel.org>
Subject: Re: [RFC PATCH v1] perf trace: Mitigate failures in parallel perf
trace instances
Hello Namhyung,
On Fri, May 30, 2025 at 4:37 PM Namhyung Kim <namhyung@...nel.org> wrote:
>
> Hello,
>
> (Adding tracing folks)
(That's so convenient wow)
>
> On Wed, May 28, 2025 at 11:55:36PM -0700, Howard Chu wrote:
> > perf trace utilizes the tracepoint utility, the only filter in perf
> > trace is a filter on syscall type. For example, if perf traces only
> > openat, then it filters all the other syscalls, such as readlinkat,
> > readv, etc.
> >
> > This filtering is flawed. Consider this case: two perf trace
> > instances are running at the same time, trace instance A tracing
> > readlinkat, trace instance B tracing openat. When an openat syscall
> > enters, it triggers both BPF programs (sys_enter) in both perf trace
> > instances, these kernel functions will be executed:
> >
> > perf_syscall_enter
> > perf_call_bpf_enter
> > trace_call_bpf
> > bpf_prog_run_array
> >
> > In bpf_prog_run_array:
> > ~~~
> > while ((prog = READ_ONCE(item->prog))) {
> > run_ctx.bpf_cookie = item->bpf_cookie;
> > ret &= run_prog(prog, ctx);
> > item++;
> > }
> > ~~~
> >
> > I'm not a BPF expert, but by tinkering I found that if one of the BPF
> > programs returns 0, there will be no tracepoint sample. That is,
> >
> > (Is there a sample?) = ProgRetA & ProgRetB & ProgRetC
> >
> > Where ProgRetA is the return value of one of the BPF programs in the BPF
> > program array.
> >
> > Go back to the case, when two perf trace instances are tracing two
> > different syscalls, again, A is tracing readlinkat, B is tracing openat,
> > when an openat syscall enters, it triggers the sys_enter program in
> > instance A, call it ProgA, and the sys_enter program in instance B,
> > ProgB, now ProgA will return 0 because ProgA cares about readlinkat only,
> > even though ProgB returns 1; (Is there a sample?) = ProgRetA (0) &
> > ProgRetB (1) = 0. So there won't be a tracepoint sample in B's output,
> > when there really should be one.
>
> Sounds like a bug. I think it should run bpf programs attached to the
> current perf_event only. Isn't it the case for tracepoint + perf + bpf?
I really can't answer that question.
>
> >
> > I also want to point out that openat and readlinkat have augmented
> > output, so my example might not be accurate, but it does explain the
> > current perf-trace-in-parallel dilemma.
> >
> > Now for augmented output, it is different. When it calls
> > bpf_perf_event_output, there is a sample. There won't be no ProgRetA &
> > ProgRetB... thing. So I will send another RFC patch to enable
> > parallelism using this feature. Also, augmented_output creates a sample
> > on it's own, so returning 1 will create a duplicated sample, when
> > augmented, just return 0 instead.
>
> Yes, it's bpf-output and tracepoint respectively. Maybe we should
> always return 1 not to drop syscalls unintentionally and perf can
> discard duplicated samples.
I like this.
>
> Another approach would be return 0 always and use bpf-output for
> unaugmented syscalls too. But I'm afraid it'd affect other perf tools
> using tracepoints.
Yep.
>
> >
> > Is this approach perfect? Absolutely not, there will likely be some
> > performance overhead on the kernel side. It is just a quick dirty fix
> > that makes perf trace run in parallel without failing. This patch is an
> > explanation on the reason of failures and possibly, a link used in a
> > nack comment.
>
> Thanks for your work, but I'm afraid it'd still miss some syscalls as it
> returns 0 sometimes.
My bad... For example this:
if (pid_filter__has(&pids_filtered, getpid()))
return 0;
This patch is practically meaningless, but it passes the parallel tests.
Thanks,
Howard
Powered by blists - more mailing lists