[PATCH] net: leave O_NONBLOCK to the sendmsg and recvmsg syscalls

[PATCH] net: leave O_NONBLOCK to the sendmsg and recvmsg syscalls

[Jens Axboe]

____sys_sendmsg() and ____sys_recvmsg() turn O_NONBLOCK on the socket
into MSG_DONTWAIT. That is right for the system calls, but io_uring uses
__sys_sendmsg_sock() and __sys_recvmsg_sock() for a blocking attempt
from io-wq as well, after the non-blocking attempts with MSG_DONTWAIT
have failed. On an O_NONBLOCK socket that attempt can't block either and
the request completes with -EAGAIN, unlike IORING_OP_SEND and RECV,
which call sock_sendmsg() and sock_recvmsg() directly and don't look at
O_NONBLOCK at all.

Move the O_NONBLOCK handling to ___sys_sendmsg() and ___sys_recvmsg(),
which the sendmsg, sendmmsg, recvmsg and recvmmsg system calls go
through. The io_uring helpers then pass the flags the caller asked for.

Reported-by: Luk?? Lalinsk? <lukas@lalinsky.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>

---

diff --git a/net/socket.c b/net/socket.c
index 2e3b5737c481..08658a9b8065 100644
--- a/net/socket.c
+++ b/n	
@@ -2697,8 +2697,6 @@ static int ____sys_sendmsg(struct socket *sock, struct msghdr *msg_sys,
 	flags &= ~MSG_INTERNAL_SENDMSG_FLAGS;
 	msg_sys->msg_flags = flags;
 
-	if (sock->file->f_flags & O_NONBLOCK)
-		msg_sys->msg_flags |= MSG_DONTWAIT;
 	/*
 	 * If this is sendmmsg() and current destination address is same as
 	 * previously succeeded address, omit asking LSM's decision.
@@ -2762,6 +2760,9 @@ static int ___sys_sendmsg(struct socket *sock, struct user_msghdr __user *msg,
 
 	msg_sys->msg_name = &address;
 
+	if (sock->file->f_flags & O_NONBLOCK)
+		flags |= MSG_DONTWAIT;
+
 	err = sendmsg_copy_msghdr(msg_sys, msg, flags, &iov);
 	if (err < 0)
 		return err;
@@ -2926,9 +2927,6 @@ static int ____sys_recvmsg(struct socket *sock, struct msghdr *msg_sys,
 	/* We assume all kernel code knows the size of sockaddr_storage */
 	msg_sys->msg_namelen = 0;
 
-	if (sock->file->f_flags & O_NONBLOCK)
-		flags |= MSG_DONTWAIT;
-
 	if (unlikely(nosec))
 		err = sock_recvmsg_nosec(sock, msg_sys, flags);
 	else
@@ -2970,6 +2968,9 @@ static int ___sys_recvmsg(struct socket *sock, struct user_msghdr __user *msg,
 	struct sockaddr __user *uaddr;
 	ssize_t err;
 
+	if (sock->file->f_flags & O_NONBLOCK)
+		flags |= MSG_DONTWAIT;
+
 	err = recvmsg_copy_msghdr(msg_sys, msg, flags, &uaddr, &iov);
 	if (err < 0)
 		return err;


-- 
Jens Axboe

Re: [PATCH] net: leave O_NONBLOCK to the sendmsg and recvmsg syscalls

[Eric Dumazet]

Le sam. 7 nov. 2026 à 00:19, Jens Axboe <axboe@kernel.dk> a écrit :
>
> ____sys_sendmsg() and ____sys_recvmsg() turn O_NONBLOCK on the socket
> into MSG_DONTWAIT. That is right for the system calls, but io_uring uses
> __sys_sendmsg_sock() and __sys_recvmsg_sock() for a blocking attempt
> from io-wq as well, after the non-blocking attempts with MSG_DONTWAIT
> have failed. On an O_NONBLOCK socket that attempt can't block either and
> the request completes with -EAGAIN, unlike IORING_OP_SEND and RECV,
> which call sock_sendmsg() and sock_recvmsg() directly and don't look at
> O_NONBLOCK at all.

Makes sense. The socket.c side looks fine to me: all syscall paths
(sendmsg, sendmmsg, recvmsg, recvmmsg, compat) go through ___sys_sendmsg()
or ___sys_recvmsg(), and the copy_msghdr helpers only care about
MSG_CMSG_COMPAT, so setting MSG_DONTWAIT earlier is harmless.

Fixes: tags? And please tag the subject [PATCH net] if this is
meant as a fix.

Fixes: 0fa03c624d8f ("io_uring: add support for sendmsg()")
Fixes: aa1fa28fc73e ("io_uring: add support for recvmsg()")

>
> Move the O_NONBLOCK handling to ___sys_sendmsg() and ___sys_recvmsg(),
> which the sendmsg, sendmmsg, recvmsg and recvmmsg system calls go
> through. The io_uring helpers then pass the flags the caller asked for.
>
> Reported-by: Luk?? Lalinsk? <lukas@lalinsky.com>

The name lost its accents (Lukáš Lalinský). A Closes: or Link: to the
report would be nice too.

> Signed-off-by: Jens Axboe <axboe@kernel.dk>
>
> ---
>
> diff --git a/net/socket.c b/net/socket.c
> index 2e3b5737c481..08658a9b8065 100644
> --- a/net/socket.c
> +++ b/n

This header got truncated, so the patch does not apply as posted.

> @@ -2697,8 +2697,6 @@ static int ____sys_sendmsg(struct socket *sock, struct msghdr *msg_sys,
>         flags &= ~MSG_INTERNAL_SENDMSG_FLAGS;
>         msg_sys->msg_flags = flags;
>

pw-bot: cr

Re: [PATCH] net: leave O_NONBLOCK to the sendmsg and recvmsg syscalls

[Jens Axboe]

On 11/6/26 4:36 PM, Eric Dumazet wrote:
> Le sam. 7 nov. 2026 ? 00:19, Jens Axboe <axboe@kernel.dk> a ?crit :
>>
>> ____sys_sendmsg() and ____sys_recvmsg() turn O_NONBLOCK on the socket
>> into MSG_DONTWAIT. That is right for the system calls, but io_uring uses
>> __sys_sendmsg_sock() and __sys_recvmsg_sock() for a blocking attempt
>> from io-wq as well, after the non-blocking attempts with MSG_DONTWAIT
>> have failed. On an O_NONBLOCK socket that attempt can't block either and
>> the request completes with -EAGAIN, unlike IORING_OP_SEND and RECV,
>> which call sock_sendmsg() and sock_recvmsg() directly and don't look at
>> O_NONBLOCK at all.
> 
> Makes sense. The socket.c side looks fine to me: all syscall paths
> (sendmsg, sendmmsg, recvmsg, recvmmsg, compat) go through ___sys_sendmsg()
> or ___sys_recvmsg(), and the copy_msghdr helpers only care about
> MSG_CMSG_COMPAT, so setting MSG_DONTWAIT earlier is harmless.

Thanks for checking!

> Fixes: tags? And please tag the subject [PATCH net] if this is
> meant as a fix.
> 
> Fixes: 0fa03c624d8f ("io_uring: add support for sendmsg()")
> Fixes: aa1fa28fc73e ("io_uring: add support for recvmsg()")

Good point, I'll add the fixes tags and the subject.

>> Move the O_NONBLOCK handling to ___sys_sendmsg() and ___sys_recvmsg(),
>> which the sendmsg, sendmmsg, recvmsg and recvmmsg system calls go
>> through. The io_uring helpers then pass the flags the caller asked for.
>>
>> Reported-by: Luk?? Lalinsk? <lukas@lalinsky.com>
> 
> The name lost its accents (Luk?? Lalinsk?). A Closes: or Link: to the
> report would be nice too.

Man, not a lot went right in this posting... But yes, it lost it. I'll
add a link too.

>> diff --git a/net/socket.c b/net/socket.c
>> index 2e3b5737c481..08658a9b8065 100644
>> --- a/net/socket.c
>> +++ b/n
> 
> This header got truncated, so the patch does not apply as posted.

I'll send a v2 :)

-- 
Jens Axboe