Re: io_uring: SENDMSG/WRITEV return -EAGAIN on O_NONBLOCK socket after peer half-close

[Jens Axboe <axboe@kernel.dk>]

On 11/2/26 7:00 AM, Luk?? Lalinsk? wrote:
> On Thu, Oct 29, 2026, at 9:42 PM, Luk?? Lalinsk? wrote:
>> There is one more inconsistency even with both patches applied:
>> in the urgent data case, WRITEV on an O_NONBLOCK socket still completes
>> with -EAGAIN, while SEND and SENDMSG on the same socket keep waiting.
>> So whether WRITEV waits or fails depends on whether the peer sent
>> urgent data, which the application can't really predict. It's not a
>> problem for me, since I always use SEND/SENDMSG on sockets, but it is
>> an inconsistent behavior.
> 
> Hi Jens,
> 
> I ran into one more issue while testing this. The SEND or SENDMSG
> operations become uncancelable for a brief time after submission,
> either in the half-closed state before your patches, or with the urgent
> data flag after your patches. If ASYNC_CANCEL is submitted right
> after the send on such socket, it completes with -ENOENT, but the
> send stays pending. A second cancel issued later finds it. Without the
> urgent data, the first cancel finds it and the send completes with
> -ECANCELED. Your patches made it work properly for the half-closed
> case.

Can you try with this on top?

diff --git a/io_uring/poll.c b/io_uring/poll.c
index ac2979586b4f..82485e1350dd 100644
--- a/io_uring/poll.c
+++ b/io_uring/poll.c
@@ -739,7 +739,7 @@ int io_arm_apoll(struct io_kiocb *req, unsigned issue_flags, __poll_t mask)
 int io_arm_poll_handler(struct io_kiocb *req, unsigned issue_flags)
 {
 	const struct io_issue_def *def = &io_issue_defs[req->opcode];
-	__poll_t mask = POLLPRI | POLLERR;
+	__poll_t mask = EPOLLERR;
 
 	if (!def->pollin && !def->pollout)
 		return IO_APOLL_ABORTED;
@@ -747,7 +747,8 @@ int io_arm_poll_handler(struct io_kiocb *req, unsigned issue_flags)
 		return IO_APOLL_ABORTED;
 
 	if (def->pollin) {
-		mask |= EPOLLIN | EPOLLRDNORM;
+		/* urgent data is an event for a reader only */
+		mask |= EPOLLIN | EPOLLRDNORM | EPOLLPRI;
 
 		/* If reading from MSG_ERRQUEUE using recvmsg, ignore POLLIN */
 		if (req->flags & REQ_F_CLEAR_POLLIN)

-- 
Jens Axboe