Hello, Justus Winter, le Mon 05 May 2014 14:10:48 +0200, a écrit : > I believe I have found two problems in the glibc. > > 1. hurd_check_cancel takes 'lock', and then asserts that > 'critical_section_lock' is not taken. However, hurd_thread_cancel > first takes 'critical_section_lock' and then 'lock'. This program > demonstrates this by spinning on both functions.
Indeed, it's another instance of the issue worked around in t/thread-cancel. Roland, this is the reproducer you were asking for in http://lists.gnu.org/archive/html/bug-hurd/2006-08/msg00019.html > 2. The assertion failure actually deadlocks trying to write the > message. Yes, that is not surprising since the assertion is inside a ss->lock section, which _hurd_ctty_output will want to take. Thanks for Samuel > #include <mach.h> > #include <stdio.h> > #include <pthread.h> > > thread_t thread; > > static void * > thread_start(void *arg) > { > thread = mach_thread_self (); > while (1) hurd_check_cancel (); > return NULL; > } > > int > main(int argc, char *argv[]) > { > pthread_t pthread; > int s; > s = pthread_create(&pthread, NULL, &thread_start, NULL); > if (s != 0) return 1; > > while (thread == 0) /* didum */; > printf ("spinning on hurd_thread_cancel (%d)\n", thread); > while (1) hurd_thread_cancel (thread); > return 0; > }
