Re: [BUG] kernel panic after bpf program removed.

[Alexei Starovoitov <ast@plumgrid.com>]

On 5/14/15 8:54 PM, Wangnan (F) wrote:
> Hi Alexei Starovoitov and other,
>
> I triggered a kernel panic when developing my 'perf bpf' facility. The
> call stack is listed at the bottom of
> this mail.
>
> I attached two bpf programs on 'kmem_cache_free%return' and
> '__alloc_pages_nodemask'. The programs is very simple.
> The panic is raised after closing the bpf program and the perf event
> file. Looks like the panic is caused
> by racing between closing perf event fd and bpf program fd. I'm unable
> to reproduce this problem with similar
> operations.
>
> Following is the exact instruction cause the panic.

thanks for the report.
Looks like pointer 'prog == 0x6c0' is passed into bpf_prog_put,
which means that event->tp_event was freed and memory reused before
free_event_rcu() was called.

I think it's not perf_event_fd racing with prog_fd, but rather
with kprobe freeing:
__free_event()
   event->destroy(event)
     perf_trace_destroy
       perf_trace_event_unreg
which is dropping event->tp_event->perf_refcount
that allows kprobe freeing to proceed in:
unregister_kprobe_event
   trace_remove_event_call
     probe_remove_event_call
and eventually tp_event to get freed.

I think calling perf_event_free_bpf_prog()
from __free_event() instead of free_event_rcu() will fix the race,
but please double check my analysis.
Also please send me a reproducer script. I'd like to see it crashing
first before the fix and not crashing afterwards.