pipes && EPOLLET, again

[Oleg Nesterov <oleg@redhat.com>]

Linus,

On 03/04, Oleg Nesterov wrote:
>
> and we need to cleanup the poll_usage
> logic first.

We have already discussed this before, I'll probably do this later,
but lets forget it for now.

Don't we need the trivial one-liner below anyway?

I am not saying this is a bug, but please consider

	#include <stdio.h>
	#include <stdbool.h>
	#include <stdlib.h>
	#include <unistd.h>
	#include <sys/epoll.h>
	#include <assert.h>

	static char buf[16 * 4096];

	int main(void)
	{
		int pfd[2], efd;
		struct epoll_event evt = { .events = EPOLLIN | EPOLLET };

		pipe(pfd);
		efd = epoll_create1(0);
		epoll_ctl(efd, EPOLL_CTL_ADD, pfd[0], &evt);

		write(pfd[1], buf, 4096);
		assert(epoll_wait(efd, &evt, 1, 0) == 1);

		if (!fork()) {
			write(pfd[1], buf, sizeof(buf));
			assert(0);
		}

		sleep(1);
		assert(epoll_wait(efd, &evt, 1, 0) == 1);

		return 0;
	}

the 2nd epoll_wait() fails, despite the fact that the child has already
written 15 * PAGE_SIZE bytes. This doesn't look consistent to me...

Oleg.
---

diff --git a/fs/pipe.c b/fs/pipe.c
index b0641f75b1ba..8a32257cc74f 100644
--- a/fs/pipe.c
+++ b/fs/pipe.c
@@ -554,7 +554,7 @@ anon_pipe_write(struct kiocb *iocb, struct iov_iter *from)
 		 * become empty while we dropped the lock.
 		 */
 		mutex_unlock(&pipe->mutex);
-		if (was_empty)
+		if (was_empty || pipe->poll_usage)
 			wake_up_interruptible_sync_poll(&pipe->rd_wait, EPOLLIN | EPOLLRDNORM);
 		kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
 		wait_event_interruptible_exclusive(pipe->wr_wait, pipe_writable(pipe));