sigprocmask or (better) pthread_sigmask is required to fix #11859 bug.

---

Python has a test for "broken pthread_sigmask". Extract of configure.in:

      AC_MSG_CHECKING(if PTHREAD_SCOPE_SYSTEM is supported)
      AC_CACHE_VAL(ac_cv_pthread_system_supported,
      [AC_RUN_IFELSE([AC_LANG_SOURCE([[#include <pthread.h>
      void *foo(void *parm) {
        return NULL;
      }
      main() {
        pthread_attr_t attr;
        pthread_t id;
        if (pthread_attr_init(&attr)) exit(-1);
        if (pthread_attr_setscope(&attr, PTHREAD_SCOPE_SYSTEM)) exit(-1);
        if (pthread_create(&id, &attr, foo, NULL)) exit(-1);
        exit(0);
      }]])],
      [ac_cv_pthread_system_supported=yes],
      [ac_cv_pthread_system_supported=no],
      [ac_cv_pthread_system_supported=no])
      ])
      AC_MSG_RESULT($ac_cv_pthread_system_supported)
      if test "$ac_cv_pthread_system_supported" = "yes"; then
        AC_DEFINE(PTHREAD_SYSTEM_SCHED_SUPPORTED, 1, [Defined if PTHREAD_SCOPE_SYSTEM supported.])
      fi
      AC_CHECK_FUNCS(pthread_sigmask,
        [case $ac_sys_system in
        CYGWIN*)
          AC_DEFINE(HAVE_BROKEN_PTHREAD_SIGMASK, 1,
            [Define if pthread_sigmask() does not work on your system.])
            ;;
        esac])

Extract of Python/thread_pthread.h:

/* On platforms that don't use standard POSIX threads pthread_sigmask()
 * isn't present.  DEC threads uses sigprocmask() instead as do most
 * other UNIX International compliant systems that don't have the full
 * pthread implementation.
 */
#if defined(HAVE_PTHREAD_SIGMASK) && !defined(HAVE_BROKEN_PTHREAD_SIGMASK)
#  define SET_THREAD_SIGMASK pthread_sigmask
#else
#  define SET_THREAD_SIGMASK sigprocmask
#endif

---

Because today more and more programs rely on threads, it is maybe not a good idea to provide a binding of sigprocmask(). I would prefer to only add pthread_sigmask() which has a determistic behaviour with threads. So only compile signal.pthread_sigmask() if pthread API is present and pthread_sigmask "is not broken".

---

About the patch: the doc should explain that the signal masks are inherited for child processes (fork() and execve()). I don't know if this behaviour is specific to Linux or not.

If we only use pthread_sigmask(), the doc is wrong: "Set the signal mask for the process." It's not for the process but only for the current thread.

How does it work if I change the signal mask of the main thread and then I create a new thread: the signal mask is inherited, or a default mask is used instead?

---

The new faulthandler uses a thread to implement a timeout: the thread uses pthread_sigmask() or sigprocmask() to ignore all signals. If I don't set the signal mask, some tests fail: check that a system call (like reading from a pipe) can be interrupted by signal. The problem is that signal may be send to the faulthandler thread, instead of the main thread. Hum, while I'm writing this, I realize that I should maybe not fallback to sigprocmask() because it may mask signals for the whole process (all threads)!