Re: [PATCH 0/1] futex: check PF_KTHREAD rather than !p->mm to filter out kthreads

[Oleg Nesterov <oleg@redhat.com>]

Hi Darren,

On 02/09, Darren Hart wrote:
>
> I will prepare the test that Peter suggested so we have something to
> test with now as well as run for regressions over time.

Just in case, I already wrote the stupid test-case:

	#include <stdio.h>
	#include <unistd.h>
	#include <signal.h>
	#include <sys/syscall.h>
	#include <sys/wait.h>
	#include <sys/mman.h>
	#include <assert.h>

	#define FUTEX_LOCK_PI	6

	struct robust_list {
		struct robust_list *next;
	};

	struct robust_list_head {
		struct robust_list list;
		long futex_offset;
		struct robust_list *list_op_pending;
	};

	int main(void)
	{
		int *mutex = mmap(NULL, 4, PROT_READ|PROT_WRITE,
					MAP_ANONYMOUS|MAP_SHARED, -1,0);
		assert((void *)mutex != MAP_FAILED);

		for (;;) {
			int err, pid = fork();

			if (!pid) {
				struct robust_list_head head;

				head.list.next = &head.list;
				head.futex_offset = 0;
				head.list_op_pending = (void *)mutex + 1;

				assert(syscall(__NR_set_robust_list, &head, sizeof(head)) == 0);
				kill(getpid(), SIGSTOP);
				_exit(0);
			}

			assert(waitpid(-1, NULL, WSTOPPED) == pid);

			*mutex = pid;
			kill(pid, SIGKILL);
			err = syscall(__NR_futex, mutex, FUTEX_LOCK_PI, 0,0,0);
			assert(wait(NULL) == pid);

			if (err) {
				printf("err=%d %m\n", err);
				kill(0, SIGKILL);
			}
		}

		return 0;
	}

it needs ~20 secs to fail on my machine. Probably it can be improved.

> Hrm, if we cleared the TID mask, and the pi chain is empty, can we not
> clear the waiters?

In this case the waiter should take care, I guess.


OK. I'll try to make at least the 1st fix today (EXITING -> EXITPIDONE
livelock).

Oleg.