[PATCH] splice, net: Fix splice_to_socket() for O_NONBLOCK socket

[PATCH] splice, net: Fix splice_to_socket() for O_NONBLOCK socket

[Jan Stancek]

LTP sendfile07 [1], which expects sendfile() to return EAGAIN when
transferring data from regular file to a "full" O_NONBLOCK socket,
started failing after commit 2dc334f1a63a ("splice, net: Use
sendmsg(MSG_SPLICE_PAGES) rather than ->sendpage()").
sendfile() no longer immediately returns, but now blocks.

Removed sock_sendpage() handled this case by setting a MSG_DONTWAIT
flag, fix new splice_to_socket() to do the same for O_NONBLOCK sockets.

[1] https://github.com/linux-test-project/ltp/blob/master/testcases/kernel/syscalls/sendfile/sendfile07.c

Fixes: 2dc334f1a63a ("splice, net: Use sendmsg(MSG_SPLICE_PAGES) rather than ->sendpage()")
Signed-off-by: Jan Stancek <jstancek@redhat.com>
---
 fs/splice.c | 2 ++
 1 file changed, 2 insertions(+)

diff --git a/fs/splice.c b/fs/splice.c
index 004eb1c4ce31..3e2a31e1ce6a 100644
--- a/fs/splice.c
+++ b/fs/splice.c
@@ -876,6 +876,8 @@ ssize_t splice_to_socket(struct pipe_inode_info *pipe, struct file *out,
 			msg.msg_flags |= MSG_MORE;
 		if (remain && pipe_occupancy(pipe->head, tail) > 0)
 			msg.msg_flags |= MSG_MORE;
+		if (out->f_flags & O_NONBLOCK)
+			msg.msg_flags |= MSG_DONTWAIT;
 
 		iov_iter_bvec(&msg.msg_iter, ITER_SOURCE, bvec, bc,
 			      len - remain);
-- 
2.31.1

Re: [PATCH] splice, net: Fix splice_to_socket() for O_NONBLOCK socket

[David Howells]

Jan Stancek <jstancek@redhat.com> wrote:

> LTP sendfile07 [1], which expects sendfile() to return EAGAIN when
> transferring data from regular file to a "full" O_NONBLOCK socket,
> started failing after commit 2dc334f1a63a ("splice, net: Use
> sendmsg(MSG_SPLICE_PAGES) rather than ->sendpage()").
> sendfile() no longer immediately returns, but now blocks.
> 
> Removed sock_sendpage() handled this case by setting a MSG_DONTWAIT
> flag, fix new splice_to_socket() to do the same for O_NONBLOCK sockets.

Does this actually work correctly in all circumstances?

The problem might come if you have a splice from a non-rewindable source
through a temporary pipe (eg. sendfile() using splice_direct_to_actor()).

David

Re: [PATCH] splice, net: Fix splice_to_socket() for O_NONBLOCK socket

[Jan Stancek]

On Mon, Jul 24, 2023 at 12:12 PM David Howells <dhowells@redhat.com> wrote:
>
> Jan Stancek <jstancek@redhat.com> wrote:
>
> > LTP sendfile07 [1], which expects sendfile() to return EAGAIN when
> > transferring data from regular file to a "full" O_NONBLOCK socket,
> > started failing after commit 2dc334f1a63a ("splice, net: Use
> > sendmsg(MSG_SPLICE_PAGES) rather than ->sendpage()").
> > sendfile() no longer immediately returns, but now blocks.
> >
> > Removed sock_sendpage() handled this case by setting a MSG_DONTWAIT
> > flag, fix new splice_to_socket() to do the same for O_NONBLOCK sockets.
>
> Does this actually work correctly in all circumstances?
>
> The problem might come if you have a splice from a non-rewindable source
> through a temporary pipe (eg. sendfile() using splice_direct_to_actor()).

I assumed this was safe, since sendfile / splice_direct_to_actor()
requires input to be seekable.

Re: [PATCH] splice, net: Fix splice_to_socket() for O_NONBLOCK socket

[David Howells]

Jan Stancek <jstancek@redhat.com> wrote:

> On Mon, Jul 24, 2023 at 12:12 PM David Howells <dhowells@redhat.com> wrote:
> >
> > Jan Stancek <jstancek@redhat.com> wrote:
> >
> > > LTP sendfile07 [1], which expects sendfile() to return EAGAIN when
> > > transferring data from regular file to a "full" O_NONBLOCK socket,
> > > started failing after commit 2dc334f1a63a ("splice, net: Use
> > > sendmsg(MSG_SPLICE_PAGES) rather than ->sendpage()").
> > > sendfile() no longer immediately returns, but now blocks.
> > >
> > > Removed sock_sendpage() handled this case by setting a MSG_DONTWAIT
> > > flag, fix new splice_to_socket() to do the same for O_NONBLOCK sockets.
> >
> > Does this actually work correctly in all circumstances?
> >
> > The problem might come if you have a splice from a non-rewindable source
> > through a temporary pipe (eg. sendfile() using splice_direct_to_actor()).
> 
> I assumed this was safe, since sendfile / splice_direct_to_actor()
> requires input to be seekable.

Ah!  The test isn't where I was looking for it (in sendfile()) - it's in
splice_direct_to_actor().

I wonder if it's worth making that explicit in do_sendfile() as the
requirement doesn't hold if the output is a pipe (though in such a case,
there's an explicit buffer, so it's not actually a problem).

Anyway, did you want to post this to netdev too so that the networking tree
picks it up?  Feel free to add:

Acked-by: David Howells <dhowells@redhat.com>

Re: [PATCH] splice, net: Fix splice_to_socket() for O_NONBLOCK socket

[Jakub Kicinski]

On Mon, 24 Jul 2023 16:44:07 +0100 David Howells wrote:
> Anyway, did you want to post this to netdev too so that the networking tree
> picks it up?  Feel free to add:

+1, no preference which tree this goes thru, but if no one else claims
it please repost CCing netdev@vger.kernel.org

Re: [PATCH] splice, net: Fix splice_to_socket() for O_NONBLOCK socket

[Jan Stancek]

On Mon, Jul 24, 2023 at 7:09 PM Jakub Kicinski <kuba@kernel.org> wrote:
>
> On Mon, 24 Jul 2023 16:44:07 +0100 David Howells wrote:
> > Anyway, did you want to post this to netdev too so that the networking tree
> > picks it up?  Feel free to add:
>
> +1, no preference which tree this goes thru, but if no one else claims
> it please repost CCing netdev@vger.kernel.org

I'll repost, thanks.