On 2021/9/13 下午10:49, Dave Hansen wrote:
> On 9/12/21 8:30 PM, 王贇 wrote:
>> According to the trace we know the story is like this, the NMI
>> triggered perf IRQ throttling and call perf_log_throttle(),
>> which triggered the swevent overflow, and the overflow process
>> do perf_callchain_user() which triggered a user PF, and the PF
>> process triggered perf ftrace which finally lead into a suspected
>> stack overflow.
>>
>> This patch disable ftrace on fault.c, which help to avoid the panic.
> ...
>> +# Disable ftrace to avoid stack overflow.
>> +CFLAGS_REMOVE_fault.o = $(CC_FLAGS_FTRACE)
>
> Was this observed on a mainline kernel?
Yes, it is trigger on linux-next.
>
> How reproducible is this?
>
> I suspect we're going into do_user_addr_fault(), then falling in here:
>
>> if (unlikely(faulthandler_disabled() || !mm)) {
>> bad_area_nosemaphore(regs, error_code, address);
>> return;
>> }
>
Correct, perf_callchain_user() disabled PF which lead into here.
> Then something double faults in perf_swevent_get_recursion_context().
> But, you snipped all of the register dump out so I can't quite see
> what's going on and what might have caused *that* fault. But, in my
> kernel perf_swevent_get_recursion_context+0x0/0x70 is:
>
> mov $0x27d00,%rdx
>
> which is rather unlikely to fault.
Would you like to check the full trace I just sent see if we can get any
clue?
>
> Either way, we don't want to keep ftrace out of fault.c. This patch is
> just a hack, and doesn't really try to fix the underlying problem. This
> situation *should* be handled today. There's code there to handle it.
>
> Something else really funky is going on.
Do you think stack overflow is possible in this case? To be mentioned the NMI
arrive in very high frequency, and reduce perf_event_max_sample_rate to a low
value can also avoid the panic.
Regards,
Michael Wang
>
