Message-ID: <CAP045ApsNDc-wJSSY0-BC+HMvWErUYk=GAt6P+J_8Q6dcdXj4Q@mail.gmail.com>
Date: Fri, 12 Jul 2024 16:05:16 -0700
From: Kyle Huey <me@...ehuey.com>
To: Jiri Olsa <olsajiri@...il.com>
Cc: Joe Damato <jdamato@...tly.com>, linux-kernel@...r.kernel.org, bpf@...r.kernel.org, 
	netdev@...r.kernel.org, acme@...nel.org, andrii.nakryiko@...il.com, 
	elver@...gle.com, khuey@...ehuey.com, mingo@...nel.org, namhyung@...nel.org, 
	peterz@...radead.org, robert@...llahan.org, yonghong.song@...ux.dev, 
	mkarsten@...terloo.ca, kuba@...nel.org
Subject: Re: [bpf?] [net-next ?] [RESEND] possible bpf overflow/output bug
 introduced in 6.10rc1 ?

On Fri, Jul 12, 2024 at 3:49 PM Kyle Huey <me@...ehuey.com> wrote:
>
> On Fri, Jul 12, 2024 at 3:18 PM Jiri Olsa <olsajiri@...il.com> wrote:
> >
> > On Fri, Jul 12, 2024 at 09:53:53AM -0700, Joe Damato wrote:
> > > Greetings:
> > >
> > > (I am reposting this question after 2 days and to a wider audience
> > > as I didn't hear back [1]; my apologies it just seemed like a
> > > possible bug slipped into 6.10-rc1 and I wanted to bring attention
> > > to it before 6.10 is released.)
> > >
> > > While testing some unrelated networking code with Martin Karsten (cc'd on
> > > this email) we discovered what appears to be some sort of overflow bug in
> > > bpf.
> > >
> > > git bisect suggests that commit f11f10bfa1ca ("perf/bpf: Call BPF handler
> > > directly, not through overflow machinery") is the first commit where the
> > > (I assume) buggy behavior appears.
> >
> > heya, nice catch!
> >
> > I can reproduce.. it seems that after f11f10bfa1ca we allow to run tracepoint
> > program as perf event overflow program
> >
> > bpftrace's bpf program returns 1 which means that perf_trace_run_bpf_submit
> > will continue to execute perf_tp_event and then:
> >
> >   perf_tp_event
> >     perf_swevent_event
> >       __perf_event_overflow
> >         bpf_overflow_handler
> >
> > bpf_overflow_handler then executes event->prog on wrong arguments, which
> > results in wrong 'work' data in bpftrace output
> >
> > I can 'fix' that by checking the event type before running the program like
> > in the change below, but I wonder there's probably better fix
> >
> > Kyle, any idea?
>
> Thanks for doing the hard work here Jiri. I did see the original email
> a couple days ago but the cause was far from obvious to me so I was
> waiting until I had more time to dig in.
>
> The issue here is that kernel/trace/bpf_trace.c pokes at event->prog
> directly, so the assumption made in my patch series (based on the
> suggested patch at
> https://lore.kernel.org/lkml/ZXJJa5re536_e7c1@google.com/) that having
> a BPF program in event->prog means we also use the BPF overflow
> handler is wrong.
>
> I'll think about how to fix it.
>
> - Kyle

The good news is that perf_event_attach_bpf_prog() (where we have a
program but no overflow handler) and perf_event_set_bpf_handler()
(where we have a program and an overflow handler) appear to be
mutually exclusive, gated on perf_event_is_tracing(). So I believe we
can fix this with a more generic version of your patch.

- Kyle

>
> > >
> > > Running the following on my machine as of the commit mentioned above:
> > >
> > >   bpftrace -e 'tracepoint:napi:napi_poll { @[args->work] = count(); }'
> > >
> > > while simultaneously transferring data to the target machine (in my case, I
> > > scp'd a 100MiB file of zeros in a loop) results in very strange output
> > > (snipped):
> > >
> > >   @[11]: 5
> > >   @[18]: 5
> > >   @[-30590]: 6
> > >   @[10]: 7
> > >   @[14]: 9
> > >
> > > It does not seem that the driver I am using on my test system (mlx5) would
> > > ever return a negative value from its napi poll function and likewise for
> > > the driver Martin is using (mlx4).
> > >
> > > As such, I don't think it is possible for args->work to ever be a large
> > > negative number, but perhaps I am misunderstanding something?
> > >
> > > I would like to note that commit 14e40a9578b7 ("perf/bpf: Remove #ifdef
> > > CONFIG_BPF_SYSCALL from struct perf_event members") does not exhibit this
> > > behavior and the output seems reasonable on my test system. Martin confirms
> > > the same for both commits on his test system, which uses different hardware
> > > than mine.
> > >
> > > Is this an expected side effect of this change? I would expect it is not
> > > and that the output is a bug of some sort. My apologies in that I am not
> > > particularly familiar with the bpf code and cannot suggest what the root
> > > cause might be.
> > >
> > > If it is not a bug:
> > >   1. Sorry for the noise :(
> >
> > your report is great, thanks a lot!
> >
> > jirka
> >
> >
> > >   2. Can anyone suggest what this output might mean or how the
> > >      script run above should be modified? AFAIK this is a fairly
> > >      common bpftrace that many folks run for profiling/debugging
> > >      purposes.
> > >
> > > Thanks,
> > > Joe
> > >
> > > [1]: https://lore.kernel.org/bpf/Zo64cpho2cFQiOeE@LQ3V64L9R2/T/#u
> >
> > ---
> > diff --git a/kernel/events/core.c b/kernel/events/core.c
> > index c6a6936183d5..0045dc754ef7 100644
> > --- a/kernel/events/core.c
> > +++ b/kernel/events/core.c
> > @@ -9580,7 +9580,7 @@ static int bpf_overflow_handler(struct perf_event *event,
> >                 goto out;
> >         rcu_read_lock();
> >         prog = READ_ONCE(event->prog);
> > -       if (prog) {
> > +       if (prog && prog->type == BPF_PROG_TYPE_PERF_EVENT) {
> >                 perf_prepare_sample(data, event, regs);
> >                 ret = bpf_prog_run(prog, &ctx);
> >         }

Powered by blists - more mailing lists