Re: [Libunwind-devel] crashes trying to get a backtrace from a signal handler (ARM)

[Lassi Tuura]

Hi,

> The frame above is always __default_sa_restorer() from /lib/libc.so.0.
> At least that's what my backtrace outputs before it segfaults.

OK so either the unwind info for that frame is incorrect and sends libunwind
into la-la-land, or whatever is above is not something libunwind understands.

If libunwind gets sent to an invalid addresses, it's not always completely
robust. Maybe the ARM bits fall-back handler require better protection to
avoid stomping around in invalid address space regions?

> For now I have backed out the experimental frame-chain unwinding for the
> ARM platform. But I am willing to help you to debug this further.

I know precious little about ARM specifically; maybe someone else on the list
can help. I can try to help with the generic unwind issues. I'd try to figure
out first if the failing address is valid. If not, figure out what's wrong in
__default_sa_restorer() unwind info; if yes, understand why libunwind doesn't
understand it - either why it doesn't find frame description for it, or what
wrong assumptions the fallback handler makes.

You can try reading "readelf -Wwf /lib/libc.so.0" output for the restorer and
see if it matches the disassembly of the function.

It's also useful to compare GDB's back traces with libunwind ones to see where
things go wrong and why.

Regards,
Lassi