Hi,

I'm hitting an issue with the assembler implementation of pthread_cond_wait() that is used on x86_64 with nptl.

If a thread is canceled while waiting in pthread_cond_wait(), the condvar seems be left in an inconsistent state. A different thread first calling pthread_cond_signal()t then gets blocked in pthread_cond_destroy(). It seems that the function falsely assumes that there are other waiters in the condvar.

The test program below reproduces the issue, blocking in pthread_cond_destroy(). Switching to the C implementation on x86_64 fixes the problem here.

A trivial fix would just switch back to the C implementation, basically revert e928e223fd. That commit, however, states that the generic implementations are broken on x86_64. I couldn't reproduce that, though.

Has anybody else seen pthread_cond_wait() issues on x86_64?

Kind regards Martin

--

#include <pthread.h>

static pthread_mutex_t m; static pthread_cond_t c; static pthread_t t; static volatile int ready;

static void cancelcb(void *arg) {     pthread_mutex_unlock(&m); }

static void* threadcb(void *arg) {     pthread_mutex_lock(&m);     pthread_cleanup_push(cancelcb, NULL);

ready = 1;     while (1)         pthread_cond_wait(&c, &m);     pthread_cleanup_pop(1); }

int main(int argc, const char *argv[]) {     pthread_mutex_init(&m, NULL);     pthread_cond_init(&c, NULL);

pthread_create(&t, NULL, threadcb, NULL);

while (!ready);

pthread_cancel(t);     pthread_join(t, NULL);

pthread_cond_signal(&c);     pthread_cond_destroy(&c);     pthread_mutex_destroy(&m);

return 0; }