Re: [PATCH] x86: Get bp from the IRQ regs instead of directly from the CPU

[Soeren Sandmann]

Ingo Molnar <mingo@xxxxxxx> writes:

> >  arch/x86/kernel/cpu/perf_event.c |   10 +++++++++-
> >  1 files changed, 9 insertions(+), 1 deletions(-)
> > 
> > diff --git a/arch/x86/kernel/cpu/perf_event.c b/arch/x86/kernel/cpu/perf_event.c
> > index b5801c3..39b1d0c 100644
> > --- a/arch/x86/kernel/cpu/perf_event.c
> > +++ b/arch/x86/kernel/cpu/perf_event.c
> > @@ -2177,10 +2177,18 @@ static const struct stacktrace_ops backtrace_ops = {
> >  static void
> >  perf_callchain_kernel(struct pt_regs *regs, struct perf_callchain_entry *entry)
> >  {
> > +	unsigned long bp;
> > +    
> >  	callchain_store(entry, PERF_CONTEXT_KERNEL);
> >  	callchain_store(entry, regs->ip);
> >  
> > -	dump_trace(NULL, regs, NULL, 0, &backtrace_ops, entry);
> > +#ifdef CONFIG_FRAME_POINTER
> > +	bp = regs->bp;
> > +#else
> > +	bp = 0;
> > +#endif
> > +	
> > +	dump_trace(NULL, regs, NULL, bp, &backtrace_ops, entry);
> >  }
> 
> Wouldnt it be better to push this logic into dump_trace() itself? That 
> way other ways of backtrace generation would be improved as well, not 
> just perf events call-chains.

Yes, it would, and I wrote the patch for that below, but then
discovered that getting bp from the IRQ registers makes stacktracing
not work at all on 64 bit. As I read entry_64.S, rbp ends up being
stored in regs->bx; maybe that's related.

I'll try and track down what is going on later.


Soren