Re: [discuss] Re: [PATCH 2/3] reliable stack trace support (x86-64)

[Andi Kleen <ak@suse.de>]

On Thursday 18 May 2006 11:37, Jan Beulich wrote:
> >>> Andi Kleen <ak@suse.de> 16.05.06 19:05 >>>
> >
> >On Tuesday 16 May 2006 18:06, Jan Beulich wrote:
> >> >>> Andi Kleen <ak@suse.de> 16.05.06 17:13 >>>
> >>
> >> On Tuesday 16 May 2006 16:21, Jan Beulich wrote:
> >> >> These are the x86_64-specific pieces to enable reliable stack traces.
> >> >> The only restriction with this is that it currently cannot unwind
> >> >> across the interrupt->normal stack boundary, as that transition is
> >> >> lacking proper annotation.
> >> >
> >> >It would be nice if you could submit a patch to fix that.
> >>
> >> But I don't know how to fix it. See my other mail
> >
> >which mail?
>
> Reply to Ingo (with you on cc) regarding patch 1/3. Just saying that I
> don't know much about expressions here.
>

Hmm, maybe we can find somebody who does. But then they would
first need to be implemented in the unwinder anyways, I guess.
Could be nasty agreed.

> >> - I have no experience with expressions, nor have I ever seen them in
> >> use.
> >
> >I remember Jim Houston used a hack of just loading the old stack into a
> > register and defining that as a base register in CFI. I guess i would be
> > willing to trade a few moves for that (should be pretty much free on a
> > OOO CPU anyways) You think that trick would work?
>
> I don't think that would, because without CONFIG_DEBUG_INFO none of the
> preserved registers get saved, hence there's no register to use for this.
> Thus the price would not only be a move, but also a save/push and a  
> reload/pop.

Three instructions?  We might be still able to afford that.

push/pop is probably not needed because the stack frame will be already
set up (ok possibly after a few instructions, but a small window might be 
tolerable)

> >> >> +#define UNW_PC(frame) (frame)->regs.rip
> >> >> +#define UNW_SP(frame) (frame)->regs.rsp
> >> >
> >> >I think we alreay have instruction_pointer(). Better add a
> >> > stack_pointer() in ptrace.h too.
> >>
> >> I could do that, but the macros will have to remain, as they don't
> >> access pt_regs dierectly, so I guess it'd be pointless to change it.
> >
> >UNW_PC() is instruction_pointer(&frame->regs), isn't it?
>
> Yes. But the intention is that the user of UNW_PC doesn't need to know any
> details of what fields frame has (i.e. the parameter of UNW_PC must only be
> frame), so you can't replace it with instruction_pointer().

Maybe I'm dense but I still don't get - frame has a pt_regs so why 
isn't the caller allowed to know about that fact?

-Andi