NPTL not properly cleaning up threads on SMP systems?

[Thibaut Girka <thib at sitedethib dot com>]

Hi,

I've recently tried to build eglibc from Debian's source package,
but it aborted because of a failed test[1]: nptl/tst-eintr1.
This test failed because one of the pthread_create calls returned EAGAIN.
Indeed, after some time, instead of 12~23 threads,
âgrep Threads /proc/$PID/statusâ reveals much more threads (up to several thousands).

I've attached a simpler testcase that triggers the same issue,
which I have been able to reproduce using a freshly-built glibc from GIT.

I haven't observed this issue on single-processor systems
(ran the test on a dedicated user with a RLIMIT_NPROC of NB_THREADS + 3 for hours without getting EAGAIN).

Regards,
Thibaut Girka.

#include <stdlib.h>
#include <stdio.h>
#include <pthread.h>
#include <sys/types.h>
#include <unistd.h>

#define NB_THREADS	20

void *thread_func(void *arg)
{
	return arg;
}


int main(void)
{
	unsigned int i = 0, j;
	void *result;
	pthread_t threads[NB_THREADS];

	printf("Please watch /proc/%d/task as I'm running!\n", getpid());

	/* Pre-spawn NB_THREAD threads */
	for (j = 0; j < NB_THREADS; j++)
	{
		if (pthread_create(&threads[j], NULL, thread_func, (void *) j) != 0)
		{
			perror("pthread_create");
			return 1;
		}
	}

	/* Now, close one thread, spawn another. FIFO style. */
	while(1)
	{
		if (pthread_join(threads[i % NB_THREADS], &result) != 0)
		{
			perror("pthread_join");
			break;
		}
		if ((unsigned int) result != (i % NB_THREADS))
		{
			printf("Invalid reslut: %d!\n", (unsigned int) result);
			break;
		}
		if (pthread_create(&threads[i % NB_THREADS], NULL, thread_func, (void *) (i % NB_THREADS)) != 0)
		{
			perror("pthread_create");
			break;
		}
		i += 1;
		if (i % 50000 == 0)
			printf("%d threads created\n", i);
	}

	printf("%d threads created\n", i);

	return 0;
}

Attachment: signature.asc
Description: Digital signature