Re: [BUG perf] perf_fetch_caller_regs / rewind_frame_pointer can panic

[Frederic Weisbecker <fweisbec@gmail.com>]

On Thu, Apr 08, 2010 at 11:57:20AM +0200, Eric Dumazet wrote:
> Hello
> 
> Current linux-2.6 tree panics on my dev machine
> 
> 64 bit kernel, 32bit user land
> CONFIG_FRAME_POINTER=y
> 
> perf timechart record &
> 
> Instant crash
> 
> Call Trace:
>  perf_trace_sched_switch+0xd5/0x120
>  schedule+0x6b5/0x860
>  retint_careful+0xd/0x21
>  
> RIP ffffffff81010955 perf_arch_fetch_caller_regs+0x15/0x40
> CR2: 00000000d21f1422
> 
> 
> rewind_frame_pointer() is probably wrong.
> 
> No test performed to check frame is in current stack, or
> that (!user_mode_vm(regs))



user_mode_vm() can not work here as we are actually filling
regs from scratch.

But we indeed need to have a safe dereference to avoid such
crashes. A simple probe_kernel_address() should do the trick.

This API is going to change for the next cycle as it won't need
to rewind further than the first caller. So I'm going to do a
rough probe_kernel_address() fix for the current version. The next
one won't have this problem.



> 
> 
> static inline unsigned long rewind_frame_pointer(int n)
> {
> 	struct stack_frame *frame;
> 
> 	get_bp(frame);
> 
> #ifdef CONFIG_FRAME_POINTER
> 	while (n--)
> 		frame = frame->next_frame;
> #endif
> 
> 	return (unsigned long)frame;
> }
> 
> 
>