io_uring: optimise submission side poll_refs

The final poll_refs put in __io_arm_poll_handler() takes quite some
cycles. When we're arming from the original task context task_work won't
be run, so in this case we can assume that we won't race with task_works
and so not take the initial ownership ref.

One caveat is that after arming a poll we may race with it, so we have
to add a bunch of io_poll_get_ownership() hidden inside of
io_poll_can_finish_inline() whenever we want to complete arming inline.
For the same reason we can't just set REQ_F_DOUBLE_POLL in
__io_queue_proc() and so need to sync with the first poll entry by
taking its wq head lock.

Signed-off-by: Pavel Begunkov <asml.silence@gmail.com>
Link: https://lore.kernel.org/r/8825315d7f5e182ac1578a031e546f79b1c97d01.1655990418.git.asml.silence@gmail.com
Signed-off-by: Jens Axboe <axboe@kernel.dk>

io_uring/poll.c

@@ -34,6 +34,7 @@ struct io_poll_table {
 	struct io_kiocb *req;
 	int nr_entries;
 	int error;
+	bool owning;
 	/* output value, set only if arm poll returns >0 */
 	__poll_t result_mask;
 };
@@ -374,6 +375,27 @@ static int io_poll_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
 	return 1;
 }
 
+static void io_poll_double_prepare(struct io_kiocb *req)
+{
+	struct wait_queue_head *head;
+	struct io_poll *poll = io_poll_get_single(req);
+
+	/* head is RCU protected, see io_poll_remove_entries() comments */
+	rcu_read_lock();
+	head = smp_load_acquire(&poll->head);
+	if (head) {
+		/*
+		 * poll arm may not hold ownership and so race with
+		 * io_poll_wake() by modifying req->flags. There is only one
+		 * poll entry queued, serialise with it by taking its head lock.
+		 */
+		spin_lock_irq(&head->lock);
+		req->flags |= REQ_F_DOUBLE_POLL;
+		spin_unlock_irq(&head->lock);
+	}
+	rcu_read_unlock();
+}
+
 static void __io_queue_proc(struct io_poll *poll, struct io_poll_table *pt,
 			    struct wait_queue_head *head,
 			    struct io_poll **poll_ptr)
@@ -405,16 +427,19 @@ static void __io_queue_proc(struct io_poll *poll, struct io_poll_table *pt,
 			pt->error = -ENOMEM;
 			return;
 		}
+
+		io_poll_double_prepare(req);
 		/* mark as double wq entry */
 		wqe_private |= IO_WQE_F_DOUBLE;
-		req->flags |= REQ_F_DOUBLE_POLL;
 		io_init_poll_iocb(poll, first->events, first->wait.func);
 		*poll_ptr = poll;
 		if (req->opcode == IORING_OP_POLL_ADD)
 			req->flags |= REQ_F_ASYNC_DATA;
+	} else {
+		/* fine to modify, there is no poll queued to race with us */
+		req->flags |= REQ_F_SINGLE_POLL;
 	}
 
-	req->flags |= REQ_F_SINGLE_POLL;
 	pt->nr_entries++;
 	poll->head = head;
 	poll->wait.private = (void *) wqe_private;
@@ -435,6 +460,12 @@ static void io_poll_queue_proc(struct file *file, struct wait_queue_head *head,
 			(struct io_poll **) &pt->req->async_data);
 }
 
+static bool io_poll_can_finish_inline(struct io_kiocb *req,
+				      struct io_poll_table *pt)
+{
+	return pt->owning || io_poll_get_ownership(req);
+}
+
 /*
  * Returns 0 when it's handed over for polling. The caller owns the requests if
  * it returns non-zero, but otherwise should not touch it. Negative values
@@ -443,7 +474,8 @@ static void io_poll_queue_proc(struct file *file, struct wait_queue_head *head,
  */
 static int __io_arm_poll_handler(struct io_kiocb *req,
 				 struct io_poll *poll,
-				 struct io_poll_table *ipt, __poll_t mask)
+				 struct io_poll_table *ipt, __poll_t mask,
+				 unsigned issue_flags)
 {
 	struct io_ring_ctx *ctx = req->ctx;
 	int v;
@@ -452,34 +484,45 @@ static int __io_arm_poll_handler(struct io_kiocb *req,
 	req->work.cancel_seq = atomic_read(&ctx->cancel_seq);
 	io_init_poll_iocb(poll, mask, io_poll_wake);
 	poll->file = req->file;
-
 	req->apoll_events = poll->events;
 
 	ipt->pt._key = mask;
 	ipt->req = req;
 	ipt->error = 0;
 	ipt->nr_entries = 0;
-
 	/*
-	 * Take the ownership to delay any tw execution up until we're done
+	 * Polling is either completed here or via task_work, so if we're in the
-	 * with poll arming. see io_poll_get_ownership().
+	 * task context we're naturally serialised with tw by merit of running
+	 * the same task. When it's io-wq, take the ownership to prevent tw
+	 * from running. However, when we're in the task context, skip taking
+	 * it as an optimisation.
+	 *
+	 * Note: even though the request won't be completed/freed, without
+	 * ownership we still can race with io_poll_wake().
+	 * io_poll_can_finish_inline() tries to deal with that.
 	 */
-	atomic_set(&req->poll_refs, 1);
+	ipt->owning = issue_flags & IO_URING_F_UNLOCKED;
+
+	atomic_set(&req->poll_refs, (int)ipt->owning);
 	mask = vfs_poll(req->file, &ipt->pt) & poll->events;
 
 	if (unlikely(ipt->error || !ipt->nr_entries)) {
 		io_poll_remove_entries(req);
 
-		if (mask && (poll->events & EPOLLET)) {
+		if (!io_poll_can_finish_inline(req, ipt)) {
+			io_poll_mark_cancelled(req);
+			return 0;
+		} else if (mask && (poll->events & EPOLLET)) {
 			ipt->result_mask = mask;
 			return 1;
-		} else {
-			return ipt->error ?: -EINVAL;
 		}
+		return ipt->error ?: -EINVAL;
 	}
 
 	if (mask &&
 	   ((poll->events & (EPOLLET|EPOLLONESHOT)) == (EPOLLET|EPOLLONESHOT))) {
+		if (!io_poll_can_finish_inline(req, ipt))
+			return 0;
 		io_poll_remove_entries(req);
 		ipt->result_mask = mask;
 		/* no one else has access to the req, forget about the ref */
@@ -491,11 +534,13 @@ static int __io_arm_poll_handler(struct io_kiocb *req,
 	else
 		io_poll_req_insert(req);
 
-	if (mask && (poll->events & EPOLLET)) {
+	if (mask && (poll->events & EPOLLET) &&
+	    io_poll_can_finish_inline(req, ipt)) {
 		__io_poll_execute(req, mask);
 		return 0;
 	}
 
+	if (ipt->owning) {
 		/*
 		 * Release ownership. If someone tried to queue a tw while it was
 		 * locked, kick it off for them.
@@ -503,6 +548,7 @@ static int __io_arm_poll_handler(struct io_kiocb *req,
 		v = atomic_dec_return(&req->poll_refs);
 		if (unlikely(v & IO_POLL_REF_MASK))
 			__io_poll_execute(req, 0);
+	}
 	return 0;
 }
 
@@ -585,7 +631,7 @@ int io_arm_poll_handler(struct io_kiocb *req, unsigned issue_flags)
 
 	io_kbuf_recycle(req, issue_flags);
 
-	ret = __io_arm_poll_handler(req, &apoll->poll, &ipt, mask);
+	ret = __io_arm_poll_handler(req, &apoll->poll, &ipt, mask, issue_flags);
 	if (ret)
 		return ret > 0 ? IO_APOLL_READY : IO_APOLL_ABORTED;
 	trace_io_uring_poll_arm(req, mask, apoll->poll.events);
@@ -817,7 +863,7 @@ int io_poll_add(struct io_kiocb *req, unsigned int issue_flags)
 	else
 		req->flags &= ~REQ_F_HASH_LOCKED;
 
-	ret = __io_arm_poll_handler(req, poll, &ipt, poll->events);
+	ret = __io_arm_poll_handler(req, poll, &ipt, poll->events, issue_flags);
 	if (ret > 0) {
 		io_req_set_res(req, ipt.result_mask, 0);
 		return IOU_OK;