Re: [RFC PATCH] perf: fix panic by mark recursion inside perf_log_throttle
From: 王贇
Date: Mon Sep 13 2021 - 21:58:52 EST
On 2021/9/13 下午6:24, Peter Zijlstra wrote:
> On Mon, Sep 13, 2021 at 11:00:47AM +0800, 王贇 wrote:
>>
>>
>> On 2021/9/10 下午11:38, Peter Zijlstra wrote:
>>> On Thu, Sep 09, 2021 at 11:13:21AM +0800, 王贇 wrote:
>>>> When running with ftrace function enabled, we observed panic
>>>> as below:
>>>>
>>>> traps: PANIC: double fault, error_code: 0x0
>>>> [snip]
>>>> RIP: 0010:perf_swevent_get_recursion_context+0x0/0x70
>>>> [snip]
>>>> Call Trace:
>>>> <NMI>
>>>> perf_trace_buf_alloc+0x26/0xd0
>>>> perf_ftrace_function_call+0x18f/0x2e0
>>>> kernelmode_fixup_or_oops+0x5/0x120
>>>> __bad_area_nosemaphore+0x1b8/0x280
>>>> do_user_addr_fault+0x410/0x920
>>>> exc_page_fault+0x92/0x300
>>>> asm_exc_page_fault+0x1e/0x30
>>>> RIP: 0010:__get_user_nocheck_8+0x6/0x13
>>>> perf_callchain_user+0x266/0x2f0
>>>> get_perf_callchain+0x194/0x210
>>>> perf_callchain+0xa3/0xc0
>>>> perf_prepare_sample+0xa5/0xa60
>>>> perf_event_output_forward+0x7b/0x1b0
>>>> __perf_event_overflow+0x67/0x120
>>>> perf_swevent_overflow+0xcb/0x110
>>>> perf_swevent_event+0xb0/0xf0
>>>> perf_tp_event+0x292/0x410
>>>> perf_trace_run_bpf_submit+0x87/0xc0
>>>> perf_trace_lock_acquire+0x12b/0x170
>>>> lock_acquire+0x1bf/0x2e0
>>>> perf_output_begin+0x70/0x4b0
>>>> perf_log_throttle+0xe2/0x1a0
>>>> perf_event_nmi_handler+0x30/0x50
>>>> nmi_handle+0xba/0x2a0
>>>> default_do_nmi+0x45/0xf0
>>>> exc_nmi+0x155/0x170
>>>> end_repeat_nmi+0x16/0x55
>>>
>>> kernel/events/Makefile has:
>>>
>>> ifdef CONFIG_FUNCTION_TRACER
>>> CFLAGS_REMOVE_core.o = $(CC_FLAGS_FTRACE)
>>> endif
>>>
>>> Which, afaict, should avoid the above, no?
>>
>> I'm afraid it's not working for this case, the
>> start point of tracing is at lock_acquire() which
>> is not from 'kernel/events/core', the following PF
>> related function are also not from 'core', prevent
>> ftrace on 'core' can't prevent this from happen...
>
> I'm confused tho; where does the #DF come from? Because taking a #PF
> from NMI should be perfectly fine.
>
> AFAICT that callchain is something like:
>
> NMI
> perf_event_nmi_handler()
> (part of the chain is missing here)
> perf_log_throttle()
> perf_output_begin() /* events/ring_buffer.c */
> rcu_read_lock()
> rcu_lock_acquire()
> lock_acquire()
> trace_lock_acquire() --> perf_trace_foo
>
> ...
> perf_callchain()
> perf_callchain_user()
> #PF (fully expected during a userspace callchain)
> (some stuff, until the first __fentry)
> perf_trace_function_call
> perf_trace_buf_alloc()
> perf_swevent_get_recursion_context()
> *BOOM*
>
> Now, supposedly we then take another #PF from get_recursion_context() or
> something, but that doesn't make sense. That should just work...
>
> Can you figure out what's going wrong there? going with the RIP, this
> almost looks like 'swhash->recursion' goes splat, but again that makes
> no sense, that's a per-cpu variable.
That's true, I actually have tried several approach to avoid the issue, but
it trigger panic as long as we access 'swhash->recursion', the array should
be accessible but somehow broken, that's why I consider this a suspected
stack overflow, since nmi repeated and trace seems very long, but just a
suspect...
Regards,
Michael Wang
>