Fix race with timer signal in stackprof_start()

[XrXr]

I got a SEGV core dump with the stack trace as follows:

#0  __pthread_kill_implementation (no_tid=0, signo=14, threadid=0) at ./nptl/pthread_kill.c:50
#1  __pthread_kill_internal (signo=14, threadid=0) at ./nptl/pthread_kill.c:78
#2  __GI___pthread_kill (threadid=0, signo=14) at ./nptl/pthread_kill.c:89
#3  <signal handler called> () at /lib/x86_64-linux-gnu/libc.so.6
#4  stackprof_start (argc=<optimized out>, argv=<optimized out>, self=<optimized out>) at stackprof.c:228
#5  vm_call_cfunc_with_frame_
...

Notice that threadid=0 in the top frame -- the SEGV comes from inside
libc as it tries to dereference threadid.

The signal comes from stackprof's signal handler:

if (pthread_self() != _stackprof.target_thread) {
    pthread_kill(_stackprof.target_thread, sig);
    return;
}

During stackprof_start(), _stackprof.target_thread is 0.

You can recreate the stack trace in the crash with a program that does
pthread_kill(0, SIGALRM):

#include <signal.h>

int
main(void)
{
   pthread_kill(0, SIGALRM);
}

Only set running after target_thread is set to avoid this crash in
case the timer expires after settimer() but before setting
target_thread.

Also, since the ordering is important here, make running
volatile sig_atomic_t to prevent the compiler from doing unwanted
reordering.

[XrXr]

Just in case you're worried about the codegen consequences of switching to violatile sig_atomic_t (the standard thing to do for things checked in signal handlers), it gets compiled down to regular memory operations on x64:

(gdb) disassemble/s stackprof_start
228	    // Do this last since timer set above might expires
229	    _stackprof.running = 1;
   0x0000000000003824 <+260>:	mov    DWORD PTR [rip+0x4c92],0x1        # 0x84c0 <_stackprof>

It's mainly to stop the optimizer from going rogue.

[XrXr]

Ah, this is fixing the same problem as #226. #226 seems more comprehensive, and should fix the problem I witnessed, too.

[froydnj]

Thank you for noticing #226! Anytime volatile is avoided, the world gets just a little bit better. 😄

[XrXr]

In some distant future, everyone will be on C11 and everyone will have access to standard atomics and the standard memory model :D