Re: [RFC PATCH] perf: fix panic by mark recursion inside perf_log_throttle

[Peter Zijlstra]

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.