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

[Jens Axboe <axboe@kernel.dk>]

On 11/6/26 5:04 AM, Lukáš Lalinský wrote:
> On Tue, Nov 3, 2026, at 5:00 PM, Jens Axboe wrote:
>> Can you try with this on top?
> 
> Hi Jens,
> 
> I tested the three patches together on current master. Everything I
> know of is fixed:
> 
> - After the peer's half-close, a waiting SEND, SENDMSG or WRITEV keeps
>   waiting and completes normally once the peer reads, with or without
>   O_NONBLOCK.
> - The same now holds when the peer has sent urgent data that the
>   application doesn't read.
> - A RECVMSG that is woken repeatedly without data no longer completes
>   with -EAGAIN.
> - The cancel now finds the send, both after a half-close and with
>   unread urgent data, and the send completes with -ECANCELED.
> 
> Our runtime's test suite also passes on the patched kernel.

Great! Patch 1+3 are in Linus's tree now, and I sent out the net
patch as well to netdev just now. Hopefully all will land in
7.4, and I will ensure they all go to stable as well.

-- 
Jens Axboe