Re: [PATCH 0/1] pipe: only enable the extra wake_up(rd_wait) when epoll is actually used

Re: [PATCH 0/1] pipe: only enable the extra wake_up(rd_wait) when epoll is actually used

[Oleg Nesterov]

On 07/23, Oleg Nesterov wrote:
>
> OK, sashiko has some concerns
>
> 	https://sashiko.dev/#/patchset/amIqmbbZx3NlzUsX%40redhat.com

Let me quote:

	Does skipping this wakeup for non-epoll consumers break io_uring?

	Applications polling pipes via io_uring do not attach an eventpoll context,
	so pipe->epoll_usage will be false. If a writer writes to an empty pipe,
	io_uring receives the wakeup.

	If the writer then writes a second chunk before the first is drained,
	anon_pipe_write() observes was_empty == false and
	pipe_get_epoll_usage() == false, skipping the waitqueue wakeup.

	Could this cause io_uring to miss events and hang permanently, waiting
	for a CQE that will never be emitted for the new data?

and I am starting to think sashiko is right (damn as always ;) and this
patch does affect/break io_uring.

Jens, could you confirm? If yes, we need to update the comments in pipe.c
(I've attached 1/1 at the end, so that you can see what this patch does)

I know nothing about io_uring and io_uring/poll.c is not trivial to say at
least ;) please check my understanding.

It seems that IORING_OP_POLL_ADD / IORING_POLL_ADD_MULTI is edge-triggered
by default! Like EPOLL_CTL_ADD / EPOLLET.

This means that io_uring depends on the "nasty semantics" too, io_poll_wake()
path should add the task work which calls io_req_post_cqe() every time the new
data arrives, even if the pipe was not empty.

Strange...

And. io_poll_parse_events() doesn't set EPOLLET if IORING_POLL_ADD_LEVEL, but
how is it possible to use IORING_POLL_ADD_LEVEL? io_poll_add_prep() only allows
IORING_POLL_ADD_MULTI in flags? OK, I don't understand this code anyway...

Oleg.
---

Subject: [PATCH] pipe: only enable the extra wake_up(rd_wait) when epoll is actually used

pipe_poll() unconditionally sets poll_usage on the first call, forcing
anon_pipe_write() to wake up readers on every write even if the pipe was
not empty. But this is only needed for epoll's "nasty semantics"; poll()
and select() users pay for it for no reason.

Rename it to epoll_usage, and only set it when the caller is actually
using epoll on the read side of the pipe.

Signed-off-by: Oleg Nesterov <oleg@redhat.com>
---
 fs/pipe.c                 | 23 +++++++++++++++++++----
 include/linux/pipe_fs_i.h |  6 ++++--
 2 files changed, 23 insertions(+), 6 deletions(-)

diff --git a/fs/pipe.c b/fs/pipe.c
index 32140cb00d7e..b45b4b8d0b0b 100644
--- a/fs/pipe.c
+++ b/fs/pipe.c
@@ -357,6 +357,23 @@ static inline unsigned int pipe_update_tail(struct pipe_inode_info *pipe,
 	return tail;
 }
 
+static void pipe_set_epoll_usage(struct file *filp, struct pipe_inode_info *pipe)
+{
+#ifdef CONFIG_EPOLL
+	if ((filp->f_mode & FMODE_READ) && filp->f_ep &&
+	    unlikely(!READ_ONCE(pipe->epoll_usage)))
+		WRITE_ONCE(pipe->epoll_usage, true);
+#endif
+}
+
+static bool pipe_get_epoll_usage(struct pipe_inode_info *pipe)
+{
+#ifdef CONFIG_EPOLL
+	return pipe->epoll_usage;
+#endif
+	return false;
+}
+
 static ssize_t
 anon_pipe_read(struct kiocb *iocb, struct iov_iter *to)
 {
@@ -689,7 +706,7 @@ anon_pipe_write(struct kiocb *iocb, struct iov_iter *from)
 	 * Epoll nonsensically wants a wakeup whether the pipe
 	 * was already empty or not.
 	 */
-	if (was_empty || pipe->poll_usage)
+	if (was_empty || pipe_get_epoll_usage(pipe))
 		wake_up_interruptible_sync_poll(&pipe->rd_wait, EPOLLIN | EPOLLRDNORM);
 	kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
 	if (wake_next_writer)
@@ -761,9 +778,7 @@ pipe_poll(struct file *filp, poll_table *wait)
 	union pipe_index idx;
 
 	/* Epoll has some historical nasty semantics, this enables them */
-	if (unlikely(!READ_ONCE(pipe->poll_usage)))
-		WRITE_ONCE(pipe->poll_usage, true);
-
+	pipe_set_epoll_usage(filp, pipe);
 	/*
 	 * Reading pipe state only -- no need for acquiring the semaphore.
 	 *
diff --git a/include/linux/pipe_fs_i.h b/include/linux/pipe_fs_i.h
index 7f6a92ac9704..d06da9bad3a8 100644
--- a/include/linux/pipe_fs_i.h
+++ b/include/linux/pipe_fs_i.h
@@ -74,7 +74,7 @@ union pipe_index {
  *	@files: number of struct file referring this pipe (protected by ->i_lock)
  *	@r_counter: reader counter
  *	@w_counter: writer counter
- *	@poll_usage: is this pipe used for epoll, which has crazy wakeups?
+ *	@epoll_usage: is this pipe used for epoll, which has crazy wakeups?
  *	@fasync_readers: reader side fasync
  *	@fasync_writers: writer side fasync
  *	@bufs: the circular array of pipe buffers
@@ -95,7 +95,9 @@ struct pipe_inode_info {
 	unsigned int files;
 	unsigned int r_counter;
 	unsigned int w_counter;
-	bool poll_usage;
+#ifdef CONFIG_EPOLL
+	bool epoll_usage;
+#endif
 #ifdef CONFIG_WATCH_QUEUE
 	bool note_loss;
 #endif
-- 
2.52.0

Re: [PATCH 0/1] pipe: only enable the extra wake_up(rd_wait) when epoll is actually used

[Mateusz Guzik]

On Fri, Jul 24, 2026 at 3:58 PM Oleg Nesterov <oleg@redhat.com> wrote:
>
> On 07/23, Oleg Nesterov wrote:
> >
> > OK, sashiko has some concerns
> >
> >       https://sashiko.dev/#/patchset/amIqmbbZx3NlzUsX%40redhat.com
>
> Let me quote:
>
>         Does skipping this wakeup for non-epoll consumers break io_uring?
>
>         Applications polling pipes via io_uring do not attach an eventpoll context,
>         so pipe->epoll_usage will be false. If a writer writes to an empty pipe,
>         io_uring receives the wakeup.
>
>         If the writer then writes a second chunk before the first is drained,
>         anon_pipe_write() observes was_empty == false and
>         pipe_get_epoll_usage() == false, skipping the waitqueue wakeup.
>
>         Could this cause io_uring to miss events and hang permanently, waiting
>         for a CQE that will never be emitted for the new data?
>
> and I am starting to think sashiko is right (damn as always ;) and this
> patch does affect/break io_uring.
>
> Jens, could you confirm? If yes, we need to update the comments in pipe.c
> (I've attached 1/1 at the end, so that you can see what this patch does)
>
> I know nothing about io_uring and io_uring/poll.c is not trivial to say at
> least ;) please check my understanding.
>
> It seems that IORING_OP_POLL_ADD / IORING_POLL_ADD_MULTI is edge-triggered
> by default! Like EPOLL_CTL_ADD / EPOLLET.
>
> This means that io_uring depends on the "nasty semantics" too, io_poll_wake()
> path should add the task work which calls io_req_post_cqe() every time the new
> data arrives, even if the pipe was not empty.
>

well io_uring could set the flag as well in that case? it does not
inherently mean random select and poll consumers definitely need this.

Re: [PATCH 0/1] pipe: only enable the extra wake_up(rd_wait) when epoll is actually used

[Oleg Nesterov]

On 07/24, Mateusz Guzik wrote:
>
> On Fri, Jul 24, 2026 at 3:58 PM Oleg Nesterov <oleg@redhat.com> wrote:
> >
> > This means that io_uring depends on the "nasty semantics" too, io_poll_wake()
> > path should add the task work which calls io_req_post_cqe() every time the new
> > data arrives, even if the pipe was not empty.
> >
>
> well io_uring could set the flag as well in that case? it does not
> inherently mean random select and poll consumers definitely need this.

How? I don't think io_uring can do this. It doesn't even know if the file
is pipe or not. Only file_operations->poll() does know...

__Perhaps__ pipe_poll() could check something like wait->_key & EPOLLET to
detect both epoll/io_uring... Not sure, probably not, I'll try to check later.

Oleg.

Re: [PATCH 0/1] pipe: only enable the extra wake_up(rd_wait) when epoll is actually used

[Breno Leitao]

On Fri, Jul 24, 2026 at 04:54:30PM +0200, Oleg Nesterov wrote:
> On 07/24, Mateusz Guzik wrote:
> >
> > On Fri, Jul 24, 2026 at 3:58 PM Oleg Nesterov <oleg@redhat.com> wrote:
> > >
> > > This means that io_uring depends on the "nasty semantics" too, io_poll_wake()
> > > path should add the task work which calls io_req_post_cqe() every time the new
> > > data arrives, even if the pipe was not empty.
> > >
> >
> > well io_uring could set the flag as well in that case? it does not
> > inherently mean random select and poll consumers definitely need this.
> 
> How? I don't think io_uring can do this. It doesn't even know if the file
> is pipe or not. Only file_operations->poll() does know...
> 
> __Perhaps__ pipe_poll() could check something like wait->_key & EPOLLET to
> detect both epoll/io_uring... Not sure, probably not, I'll try to check later.

I got the impression that pipe_poll() sees EPOLLET in wait->_key for
io_uring the same as for EPOLLET epoll. Since d59bd748db0a
("io_uring/poll: disable level triggered poll").

io_poll_add_prep(), which is called before teh proper iouring command,
rejects IORING_POLL_ADD_LEVEL, so io_poll_parse_events() always ORs in
EPOLLET for IORING_OP_POLL_ADD.

The internal apoll path sets EPOLLET unconditionally too
(io_arm_apoll()), so every io_uring poll arming carries it.

Maybe this would work on the set side?

        if ((filp->f_mode & FMODE_READ) && wait &&
            (wait->_key & EPOLLET) &&
            unlikely(!READ_ONCE(pipe->poll_usage)))
                WRITE_ONCE(pipe->poll_usage, true);

--breno

Re: [PATCH 0/1] pipe: only enable the extra wake_up(rd_wait) when epoll is actually used

[Oleg Nesterov]

On 07/27, Breno Leitao wrote:
>
> On Fri, Jul 24, 2026 at 04:54:30PM +0200, Oleg Nesterov wrote:
> >
> > __Perhaps__ pipe_poll() could check something like wait->_key & EPOLLET to
> > detect both epoll/io_uring... Not sure, probably not, I'll try to check later.
>
> I got the impression that pipe_poll() sees EPOLLET in wait->_key for
> io_uring the same as for EPOLLET epoll. Since d59bd748db0a
> ("io_uring/poll: disable level triggered poll").
>
> io_poll_add_prep(), which is called before teh proper iouring command,
> rejects IORING_POLL_ADD_LEVEL, so io_poll_parse_events() always ORs in
> EPOLLET for IORING_OP_POLL_ADD.
>
> The internal apoll path sets EPOLLET unconditionally too
> (io_arm_apoll()), so every io_uring poll arming carries it.
>
> Maybe this would work on the set side?
>
>         if ((filp->f_mode & FMODE_READ) && wait &&
>             (wait->_key & EPOLLET) &&

This is exactly what I meant and what I did. I am writing the changelog,
will send the patch soon today. Both test-cases I wrote pass, same for
tools/testing/selftests/filesystems/epoll/epoll_wakeup_test.c

Oleg.

Re: [PATCH 0/1] pipe: only enable the extra wake_up(rd_wait) when epoll is actually used

[Oleg Nesterov]

On 07/24, Oleg Nesterov wrote:
>
> On 07/23, Oleg Nesterov wrote:
> >
> > OK, sashiko has some concerns
> >
> > 	https://sashiko.dev/#/patchset/amIqmbbZx3NlzUsX%40redhat.com
>
> Let me quote:
>
> 	Does skipping this wakeup for non-epoll consumers break io_uring?
>
> 	Applications polling pipes via io_uring do not attach an eventpoll context,
> 	so pipe->epoll_usage will be false. If a writer writes to an empty pipe,
> 	io_uring receives the wakeup.
>
> 	If the writer then writes a second chunk before the first is drained,
> 	anon_pipe_write() observes was_empty == false and
> 	pipe_get_epoll_usage() == false, skipping the waitqueue wakeup.
>
> 	Could this cause io_uring to miss events and hang permanently, waiting
> 	for a CQE that will never be emitted for the new data?
>
> and I am starting to think sashiko is right (damn as always ;) and this
> patch does affect/break io_uring.
>
> Jens, could you confirm? If yes, we need to update the comments in pipe.c
> (I've attached 1/1 at the end, so that you can see what this patch does)
...
> It seems that IORING_OP_POLL_ADD / IORING_POLL_ADD_MULTI is edge-triggered
> by default! Like EPOLL_CTL_ADD / EPOLLET.

Yes, sashiko is right. With some help from AI I wrote the simple test-case

	#include <unistd.h>
	#include <sys/mman.h>
	#include <sys/epoll.h>
	#include <sys/syscall.h>
	#include <linux/io_uring.h>
	#include <assert.h>

	int main(void)
	{
		struct io_uring_params p = {};
		int fd, pfd[2];

		pipe(pfd);

		fd = syscall(SYS_io_uring_setup, 2, &p);
		assert(fd >= 0);

		void *ring = mmap(0, p.cq_off.cqes + p.cq_entries * sizeof(struct io_uring_cqe),
				  PROT_READ | PROT_WRITE, MAP_SHARED, fd, IORING_OFF_SQ_RING);
		assert(ring != MAP_FAILED);
		*(unsigned *)(ring + p.sq_off.array) = 0;
		*(unsigned *)(ring + p.sq_off.tail) = 1;

		struct io_uring_sqe *sqes = mmap(0, p.sq_entries * sizeof(*sqes),
				  PROT_READ | PROT_WRITE, MAP_SHARED, fd, IORING_OFF_SQES);
		assert(sqes != MAP_FAILED);
		sqes[0].opcode = IORING_OP_POLL_ADD;
		sqes[0].fd = pfd[0];
		sqes[0].len = IORING_POLL_ADD_MULTI;
		sqes[0].poll32_events = EPOLLIN;

		syscall(SYS_io_uring_enter, fd, 1, 0, 0, 0, 0);

		unsigned *cq_head = ring + p.cq_off.head;
		unsigned *cq_tail = ring + p.cq_off.tail;
		for (int i = 0; i < 2; ++i) {
			write(pfd[1], "", 1);
			syscall(SYS_io_uring_enter, fd, 0, 0, IORING_ENTER_GETEVENTS, 0, 0);
			assert(*cq_tail == ++*cq_head);
		}

		return 0;
	}

with this patch the 2nd assert(*cq_tail == ++*cq_head) fails.

And just for the record, another one for epoll

	#include <unistd.h>
	#include <sys/epoll.h>
	#include <assert.h>

	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);

		for (int i = 0; i < 2; ++i) {
			write(pfd[1], "", 1);
			assert(epoll_wait(efd, &evt, 1, 0) == 1);
		}

		return 0;
	}

I'll try to make V2 later.

> And. io_poll_parse_events() doesn't set EPOLLET if IORING_POLL_ADD_LEVEL, but
> how is it possible to use IORING_POLL_ADD_LEVEL? io_poll_add_prep() only allows
> IORING_POLL_ADD_MULTI in flags? OK, I don't understand this code anyway...

OK, I see: d59bd748db0a9 ("io_uring/poll: disable level triggered poll")

Oleg.