Potential race condition

[NolanDeveloper]

According to Execution of signals signal handler may be called whenever there's transition between kernel and user mode. fork() is a syscall, so there should be such transition when it returns here. cleanup() function is installed for the number of termination signals before fork(). So it seems like it might be possible that cleanup() is called between the moment when fork() has finished and monitored_pid is assigned. This would make cleanup() to _exit() here. Then child process would continue running as long as it wants instead of being terminated.

I wasn't able to reproduce this issue.

[pixelb]

Well I repro'd with:

-  monitored_pid = fork ();
+  int tpid = fork ();
+  if (tpid) sleep(10); /* delay monitor */
+  monitored_pid = tpid;

If you run timeout -v -- 30 sleep 60, and then kill the timeout process within 10s, the sleep remains

[pixelb]

Actually I think I see a related race.
If fork() fails, monitored_pid is -1.
If a signal is received before the parent process exits, then a signal is sent to pid -1, which is a all processes for which we have permission to send signals :(

[pixelb]

Fixed with commit ab4ffc8