2015-05-15 11:02:55

by Wang Nan

Subject: [PATCH] bpf: fix a race between perf_event_fd and kprobe freeing

According to Alexei Starovoitov (http://lkml.org/lkml/2015/5/15/29),
there is racing between perf_event_fd and 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.

And he suggest to call perf_event_free_bpf_prog() from __free_event()
instead of free_event_rcu() will fix the race,

Signed-off-by: Wang Nan <[email protected]>
---

Hi Alexei Starovoitov,
I tried this patch with identical operations and unable to reproduce
the problem anymore. I think your analysis is right. However, I
heavn't carefully check the base principle. Could you please check
it in your environment?

Thank you.

---
kernel/events/core.c | 3 ++-
1 file changed, 2 insertions(+), 1 deletion(-)

diff --git a/kernel/events/core.c b/kernel/events/core.c
index 81aa3a4..e1f2d5c 100644
--- a/kernel/events/core.c
+++ b/kernel/events/core.c
@@ -3422,7 +3422,6 @@ static void free_event_rcu(struct rcu_head *head)
if (event->ns)
put_pid_ns(event->ns);
perf_event_free_filter(event);
- perf_event_free_bpf_prog(event);
kfree(event);
}

@@ -3564,6 +3563,8 @@ static void __free_event(struct perf_event *event)
module_put(event->pmu->module);
}

+ perf_event_free_bpf_prog(event);
+
call_rcu(&event->rcu_head, free_event_rcu);
}

--
1.8.3.4

2015-05-15 15:59:12

by Alexei Starovoitov

Subject: Re: [PATCH] bpf: fix a race between perf_event_fd and kprobe freeing

On 5/15/15 3:52 AM, Wang Nan wrote:
> According to Alexei Starovoitov (http://lkml.org/lkml/2015/5/15/29),
> there is racing between perf_event_fd and kprobe freeing:

...

> And he suggest to call perf_event_free_bpf_prog() from __free_event()
> instead of free_event_rcu() will fix the race,
...

> @@ -3564,6 +3563,8 @@ static void __free_event(struct perf_event *event)
> module_put(event->pmu->module);
> }
>
> + perf_event_free_bpf_prog(event);
> +
> call_rcu(&event->rcu_head, free_event_rcu);
> }

I don't think that's the right place. It needs to be before destroy().
I will send a patch soon.