io_uring/net: don't clear msg_inq before io_recv_buf_select() needs it

For bundle receives to function properly, the previous iteration msg_inq
value is needed to make a judgement call on how much data there is to
receive. A previous fix ended up clearing it earlier as an error case
would potentially errantly set IORING_CQE_F_SOCK_NONEMPTY if the request
got failed.

Move the assignment to post assigning buffers for the receive, but
ensure that it's cleared for the buffer selection error case. With that,
buffer selection has the right msg_inq value and can correctly bundle
receives as designed.

Noticed while testing where it was apparent than more than 1 buffer was
never received. After fix was in place, multiple buffers are correctly
picked for receive. This provides a 10x speedup for the test case, as
the buffer size used was 64b.

Fixes: 18414a4a2eab ("io_uring/net: assign kmsg inq/flags before buffer selection")
Signed-off-by: Jens Axboe <axboe@kernel.dk>

diff --git a/io_uring/net.c b/io_uring/net.c
index 7c98c4d5094633cb948bb761d2345e65cfb380f2..cf742bdd2a93e5f2f87e5c141bb0e135fcb04d3c 100644 (file)
--- a/io_uring/net.c
+++ b/io_uring/net.c
@@ -1127,16 +1127,18 @@ int io_recv(struct io_kiocb *req, unsigned int issue_flags)
                flags |= MSG_DONTWAIT;
 
 retry_multishot:
-       kmsg->msg.msg_inq = -1;
-       kmsg->msg.msg_flags = 0;
-
        if (io_do_buffer_select(req)) {
                ret = io_recv_buf_select(req, kmsg, &len, issue_flags);
-               if (unlikely(ret))
+               if (unlikely(ret)) {
+                       kmsg->msg.msg_inq = -1;
                        goto out_free;
+               }
                sr->buf = NULL;
        }
 
+       kmsg->msg.msg_flags = 0;
+       kmsg->msg.msg_inq = -1;
+
        if (flags & MSG_WAITALL)
                min_ret = iov_iter_count(&kmsg->msg.msg_iter);