This is the mail archive of the
gcc-patches@gcc.gnu.org
mailing list for the GCC project.
Re: PATCH: Turn on -fomit-frame-pointer by default for 32bit Linux/x86
On Wed, Aug 4, 2010 at 2:16 PM, Andi Kleen <andi@firstfloor.org> wrote:
>> The issue here is not that the unwind info is missing; it's that glibc
>> backtrace() is not async-signal safe (which, given that it calls malloc
>> shouldn't be a big surprise :-).
>
> Are you sure? AFAIK it only calls malloc when you register frames,
> but not during unwinding itself.
Yes, I double-checked this morning.
The self-deadlock sequence observed with glibc-2.11.1 when calling
backtrace() from inside TCMalloc malloc() was the same as in
http://gcc.gnu.org/bugzilla/show_bug.cgi?id=24724:
backtrace() (first time around) calls dlopen("libgcc_s.so.1"), which
(recursively) calls malloc(), which self-deadlocks.
I am sure this particular problem could be resolved, but I am also sure
there are great many other ways in which backtrace() is not async-signal
safe.
For example, libgcc_s.so.1 _Unwind_Backtrace (which is what glibc
backtrace on x86 dispatches to) calls back into dl_iterate_phdr, which
locks loader lock => kaboom!
--
Paul Pluzhnikov