On Tue, Sep 06, 2022 at 08:02:39PM -0700, Alexei Starovoitov wrote:
SNIP
> > > __mutex_lock_common kernel/locking/mutex.c:605 [inline]
> > > __mutex_lock+0x13c/0x1350 kernel/locking/mutex.c:747
> > > __pipe_lock fs/pipe.c:103 [inline]
> > > pipe_write+0x132/0x1be0 fs/pipe.c:431
> > > call_write_iter include/linux/fs.h:2188 [inline]
> > > new_sync_write fs/read_write.c:491 [inline]
> > > vfs_write+0x9e9/0xdd0 fs/read_write.c:578
> > > ksys_write+0x1e8/0x250 fs/read_write.c:631
> > > do_syscall_x64 arch/x86/entry/common.c:50 [inline]
> > > do_syscall_64+0x35/0xb0 arch/x86/entry/common.c:80
> > > entry_SYSCALL_64_after_hwframe+0x63/0xcd
> >
> > looks like __bpf_trace_contention_begin needs bpf_prog_active check
> > (like below untested), which would prevent the recursion and bail
> > out after 2nd invocation
> >
> > should be easy to reproduce, will check
> >
> > jirka
> >
> >
> > ---
> > diff --git a/include/trace/bpf_probe.h b/include/trace/bpf_probe.h
> > index 6a13220d2d27..481b057cc8d9 100644
> > --- a/include/trace/bpf_probe.h
> > +++ b/include/trace/bpf_probe.h
> > @@ -4,6 +4,8 @@
> >
> > #ifdef CONFIG_BPF_EVENTS
> >
> > +DECLARE_PER_CPU(int, bpf_prog_active);
> > +
> > #undef __entry
> > #define __entry entry
> >
> > @@ -82,7 +84,11 @@ static notrace void \
> > __bpf_trace_##call(void *__data, proto) \
> > { \
> > struct bpf_prog *prog = __data; \
> > + if (unlikely(__this_cpu_inc_return(bpf_prog_active) != 1)) \
> > + goto out; \
> > CONCATENATE(bpf_trace_run, COUNT_ARGS(args))(prog, CAST_TO_U64(args)); \
> > +out: \
> > + __this_cpu_dec(bpf_prog_active);
>
> I don't think we can use this big hammer here.
> raw_tp progs attached to different hooks need to
> run on the same cpu otherwise we will lose events.
might be good place to use prog->active
I managed to reproduce it localy, will try that
jirka
