Deadlock with cpu profiling on 64 bit about gperftools HOT 3 CLOSED

libunwind doesn't call malloc or sbrk by design. But it does call mmap for 
memory
allocation:

libunwind.git/include/libunwind_i.h

#define GET_MEMORY(mem, size_in_bytes)                              \
do {                                                                        \
  /* Hopefully, mmap() goes straight through to a system call stub...  */   \
  mem = mmap (0, size, PROT_READ | PROT_WRITE, MAP_PRIVATE | MAP_ANONYMOUS, \
              -1, 0);

But the dynamic linker may call some mmap hook before issuing the actual system 
call.
Further such hooks may make recursive calls to GetStackTrace().

in_get_stack_trace is a mechanism intended to resolve this situation. But as you
found out, it has problems in the presence of async signals. 

The spin lock is guarding concurrent accesses to in_get_stack_trace. 
We should probably use cmpxchg based code there.

We've just finished writing a bugfix, that should be present in the next 
release.

This should be fixed in perftools 0.95 (just released).