OK, I think I have it figured out....

The problem is, that when the first signal handler, SIGUSR1, is run the SIGUSR2 signal is blocked. I think this is what Daniel was trying to say, or I didn't provide enough information for him to catch it. So, I've fixed the sigaction call to unblock SIGUSR2 while SIGUSR1 is going and rearranged a few things:

me->stop_info.signal = 0; me->stop_info.last_stop_count = my_stop_count;

/* Tell the thread that wants to stop the world that this */ /* thread has been stopped. Note that sem_post() is */ /* the only async-signal-safe primitive in LinuxThreads. */ sem_post(&GC_suspend_ack_sem);

#if DEBUG_THREADS GC_printf2("Waiting for restart #%d of 0x%lx\n", my_stop_count, my_thread); #endif

/* Wait until that thread tells us to restart by sending */ /* this thread a SIG_THR_RESTART signal. */ if (sigfillset(&mask) != 0) ABORT("sigfillset() failed"); if (sigdelset(&mask, SIG_THR_RESTART) != 0) ABORT("sigdelset() failed"); # ifdef NO_SIGNALS if (sigdelset(&mask, SIGINT) != 0) ABORT("sigdelset() failed"); if (sigdelset(&mask, SIGQUIT) != 0) ABORT("sigdelset() failed"); if (sigdelset(&mask, SIGTERM) != 0) ABORT("sigdelset() failed"); if (sigdelset(&mask, SIGABRT) != 0) ABORT("sigdelset() failed"); # endif

while (me->stop_info.signal != SIG_THR_RESTART) { sigsuspend(&mask); /* Wait for signal */ }

There might still be a bug that I'm just hiding. I think that the problem is while in the handler for SIGUSR1 and someone calls pthread_kill with SIGUSR2, the signal isn't marked as pending because it is masked off. It appears to rely on the following behavior:

thread 1 calls pthread_kill for SIGUSR1. thread 2 enters SIGUSR1 handler and SIGUSR2 is masked. thread 1 calls pthread_kill for SIGUSR2. signal is set pending as it is currently masked off. thread 2 calls sigsuspend thus unblocking SIGUSR2. Signal handler for SIGUSR2 is called and then sigsuspend in SIGUSR1 handler returns.

Is this correct behavior? Should a pthread_kill of a blocked signal pend or should it be dropped?

Right now, I've worked around this.