Re: [Libunwind-devel] [PATCH 0/8] libunwind accuracy improvements for x86-64

[Andrew Cagney]

On 20 April 2010 18:52, Lassi Tuura <address@hidden> wrote:
> Hi,
>
> I forgot to mention two things about this patch set:
>
> 1. The new signal frame detection code has a potential issue. The dwarf frame 
> information is only available after calling unw_step(), so if you chance to 
> set things up so you start unwinding at the kernel signal return frame, it 
> will get confused. I don't know how to fix this without major change how 
> unw_init_* work relative to unw_is_signal_frame(). I don't think it's 
> significant, as chances of constructing such a stack frame are slim; I don't 
> think it's possible with local unwind at all, and I don't even know how to 
> create the situation with remote (ptrace) unwinding.

As in back-trace from the signal trampoline?   Single-stepping a
program in a signal-handler will eventually get it back to the signal
trampoline.  Since code trying to single-step out of a handler relies
on correct back-traces, it also turns out it is significant (and users
do notice if the back-trace is wrong and complain :-).

>
> 2. As previously discussed the changes are partially based on earlier work in 
> frysk. The final code isn't really the same - libunwind has changed, I did 
> many things differently, and code in frysk wasn't a complete solution in my 
> tests - but it was useful to look at the frysk version for ideas! Sorry I 
> forgot to add the credit in the original mail; if someone from frysk cares 
> they might want to ask to be added to copyright list.

Mark Wielaard did the frysk work so perhaps a nod to him in the news
or somewhere?