[Libunwind-devel] [patch] Fix for fast unwind and tcmalloc not playing well together.

[Paul Pluzhnikov]

Greetings,

I am now trying to use Lassi's fast trace, and the results are exceptionally
good: on some of my test cases the overhead of using libunwind (compared
to frame-based unwinder) went from 9x to 1.2x ;-)

Lassi, thanks!

One problem that cropped up: tcmalloc wants to record stack traces on free,
and this causes a crash when a thread is exiting.

Glibc calls thread-specific dtors in the order in which the keys were added,
so the first dtor is the trace_cache_free() one. Then thread-specific
data for some other key is free()d, which calls into unw_backtrace(),
which uses dangling cache and munmapped cache->frames.

Attached patch
1. delays destruction of trace cache to the last possible moment, and
2. clears tls_cache, so the dangling pointer will not be used.

Tested on Linux/x86_64 (Ubuntu 10.04); no new failures.

Thanks,
-- 
Paul Pluzhnikov

libunwind-delay-dtor-20111019.txt
Description: Text document