Re: Fw: Re: 2.6.18-rc2-mm1

[Andi Kleen <ak@muc.de>]

On Fri, Jul 28, 2006 at 09:55:51AM +0100, Jan Beulich wrote:
> >>> Andrew Morton <akpm@osdl.org> 28.07.06 07:42 >>>
> >
> >fyi, Michael's dwarf unwinder seems to have broken.
> >(please follow up on lkml).
> 
> Hmm, not being able to unwind through sysenter_entry is no surprise
> (this simply cannot be properly annotated, as the return address is not
> explicit), but it'd end up in user mode anyway (and the inexact backtrace
> doesn't go past it either). The fallback message is a little mis-leading as
> what is shown is not the left-over backtrace, but the full one (Andi
> probably knows better if/when/why this is supposed to be that way).

Hmm, normally it should dump only the left over entries. On my testing
it did that.

> 
> Likewise for the more puzzling case of not being able to unwind through
> error_code - the left-over trace is again more like a full one. I'm not clear
> why it can't unwind through error_code here; a sufficiently large piece
> of the raw stack dump would be needed to check what's going on here,
> and I just again (don't know how many times I already did this) verified
> that in a similar scenario I get a proper unwind through that point.

Yes I've also seen valid stack traces through error_code

> 
> The third one, getting stuck at __down_failed, is due to the still
> unresolved issue of improper (from the perspective of stack unwinding)
> instruction ordering include/asm-i386/semaphore.h.

I'll fix that. Guess we'll just drop the lock sections.

-Andi