Re: [PATCH] io_uring/poll: use IOU_F_TWQ_LAZY_WAKE for wakeups

[Jens Axboe <axboe@kernel.dk>]

In the spirit of getting some closure/progress on this, how about this
for starters? Disables lazy wake for poll_exclusive, and provides a flag
that can otherwise be set to disable it as well.
 
diff --git a/include/linux/io_uring_types.h b/include/linux/io_uring_types.h
index d3009d56af0b..03401c6ce5bb 100644
--- a/include/linux/io_uring_types.h
+++ b/include/linux/io_uring_types.h
@@ -431,6 +431,7 @@ enum {
 	/* keep async read/write and isreg together and in order */
 	REQ_F_SUPPORT_NOWAIT_BIT,
 	REQ_F_ISREG_BIT,
+	REQ_F_POLL_NO_LAZY_BIT,
 
 	/* not a real bit, just to check we're not overflowing the space */
 	__REQ_F_LAST_BIT,
@@ -498,6 +499,8 @@ enum {
 	REQ_F_CLEAR_POLLIN	= BIT(REQ_F_CLEAR_POLLIN_BIT),
 	/* hashed into ->cancel_hash_locked, protected by ->uring_lock */
 	REQ_F_HASH_LOCKED	= BIT(REQ_F_HASH_LOCKED_BIT),
+	/* don't use lazy poll wake for this request */
+	REQ_F_POLL_NO_LAZY	= BIT(REQ_F_POLL_NO_LAZY_BIT),
 };
 
 typedef void (*io_req_tw_func_t)(struct io_kiocb *req, struct io_tw_state *ts);
diff --git a/io_uring/poll.c b/io_uring/poll.c
index d38d05edb4fa..4fed5514c379 100644
--- a/io_uring/poll.c
+++ b/io_uring/poll.c
@@ -366,11 +366,16 @@ void io_poll_task_func(struct io_kiocb *req, struct io_tw_state *ts)
 
 static void __io_poll_execute(struct io_kiocb *req, int mask)
 {
+	unsigned flags = 0;
+
 	io_req_set_res(req, mask, 0);
 	req->io_task_work.func = io_poll_task_func;
 
 	trace_io_uring_task_add(req, mask);
-	__io_req_task_work_add(req, IOU_F_TWQ_LAZY_WAKE);
+
+	if (!(req->flags & REQ_F_POLL_NO_LAZY))
+		flags = IOU_F_TWQ_LAZY_WAKE;
+	__io_req_task_work_add(req, flags);
 }
 
 static inline void io_poll_execute(struct io_kiocb *req, int res)
@@ -526,10 +531,17 @@ static void __io_queue_proc(struct io_poll *poll, struct io_poll_table *pt,
 	poll->head = head;
 	poll->wait.private = (void *) wqe_private;
 
-	if (poll->events & EPOLLEXCLUSIVE)
+	if (poll->events & EPOLLEXCLUSIVE) {
+		/*
+		 * Exclusive waits may only wake a limited amount of entries
+		 * rather than all of them, this may interfere with lazy
+		 * wake if someone does wait(events > 1).
+		 */
+		req->flags |= REQ_F_POLL_NO_LAZY;
 		add_wait_queue_exclusive(head, &poll->wait);
-	else
+	} else {
 		add_wait_queue(head, &poll->wait);
+	}
 }
 
 static void io_poll_queue_proc(struct file *file, struct wait_queue_head *head,

-- 
Jens Axboe