On 8/21/20 8:04 PM, Luke Hsiao wrote:
> From: Luke Hsiao <[email protected]>
>
> Currently, io_uring's recvmsg subscribes to both POLLERR and POLLIN. In
> the context of TCP tx zero-copy, this is inefficient since we are only
> reading the error queue and not using recvmsg to read POLLIN responses.
>
> This patch was tested by using a simple sending program to call recvmsg
> using io_uring with MSG_ERRQUEUE set and verifying with printks that the
> POLLIN is correctly unset when the msg flags are MSG_ERRQUEUE.
Sorry, one more minor thing to fix up:
> @@ -4932,6 +4934,11 @@ static bool io_arm_poll_handler(struct io_kiocb *req)
> mask |= POLLIN | POLLRDNORM;
> if (def->pollout)
> mask |= POLLOUT | POLLWRNORM;
> +
> + /* If reading from MSG_ERRQUEUE using recvmsg, ignore POLLIN */
> + if (req->opcode == IORING_OP_RECVMSG && (sqe->msg_flags & MSG_ERRQUEUE))
> + mask &= ~POLLIN;
> +
Don't pass in the sqe here, but use req->sr_msg.msg_flags for this check. This
is actually really important, as you don't want to re-read anything from the
sqe.
I'm actually surprised this one got past Jann :-)
> @@ -6146,7 +6153,7 @@ static void __io_queue_sqe(struct io_kiocb *req, const
> struct io_uring_sqe *sqe,
> * doesn't support non-blocking read/write attempts
> */
> if (ret == -EAGAIN && !(req->flags & REQ_F_NOWAIT)) {
> - if (!io_arm_poll_handler(req)) {
> + if (!io_arm_poll_handler(req, sqe)) {
Also means you can drop this part.
--
Jens Axboe
