io_uring: reduce scheduling due to tw

Every task_work will try to wake the task to be executed, which causes
excessive scheduling and additional overhead. For some tw it's
justified, but others won't do much but post a single CQE.

When a task waits for multiple cqes, every such task_work will wake it
up. Instead, the task may give a hint about how many cqes it waits for,
io_req_local_work_add() will compare against it and skip wake ups
if #cqes + #tw is not enough to satisfy the waiting condition. Task_work
that uses the optimisation should be simple enough and never post more
than one CQE. It's also ignored for non DEFER_TASKRUN rings.

Signed-off-by: Pavel Begunkov <asml.silence@gmail.com>
Link: https://lore.kernel.org/r/d2b77e99d1e86624d8a69f7037d764b739dcd225.1680782017.git.asml.silence@gmail.com
Signed-off-by: Jens Axboe <axboe@kernel.dk>

include/linux/io_uring_types.h

@@ -296,7 +296,7 @@ struct io_ring_ctx {
 		spinlock_t		completion_lock;
 
 		bool			poll_multi_queue;
-		bool			cq_waiting;
+		atomic_t		cq_wait_nr;
 
 		/*
 		 * ->iopoll_list is protected by the ctx->uring_lock for
@@ -566,6 +566,7 @@ struct io_kiocb {
 	atomic_t			refs;
 	atomic_t			poll_refs;
 	struct io_task_work		io_task_work;
+	unsigned			nr_tw;
 	/* for polled requests, i.e. IORING_OP_POLL_ADD and async armed poll */
 	union {
 		struct hlist_node	hash_node;

io_uring/io_uring.c

@@ -1300,34 +1300,58 @@ static __cold void io_fallback_tw(struct io_uring_task *tctx)
 	}
 }
 
-static void io_req_local_work_add(struct io_kiocb *req)
+static void io_req_local_work_add(struct io_kiocb *req, unsigned flags)
 {
 	struct io_ring_ctx *ctx = req->ctx;
+	unsigned nr_wait, nr_tw, nr_tw_prev;
 	struct llist_node *first;
 
+	if (req->flags & (REQ_F_LINK | REQ_F_HARDLINK))
+		flags &= ~IOU_F_TWQ_LAZY_WAKE;
+
 	first = READ_ONCE(ctx->work_llist.first);
 	do {
+		nr_tw_prev = 0;
+		if (first) {
+			struct io_kiocb *first_req = container_of(first,
+							struct io_kiocb,
+							io_task_work.node);
+			/*
+			 * Might be executed at any moment, rely on
+			 * SLAB_TYPESAFE_BY_RCU to keep it alive.
+			 */
+			nr_tw_prev = READ_ONCE(first_req->nr_tw);
+		}
+		nr_tw = nr_tw_prev + 1;
+		/* Large enough to fail the nr_wait comparison below */
+		if (!(flags & IOU_F_TWQ_LAZY_WAKE))
+			nr_tw = -1U;
+
+		req->nr_tw = nr_tw;
 		req->io_task_work.node.next = first;
 	} while (!try_cmpxchg(&ctx->work_llist.first, &first,
 			      &req->io_task_work.node));
 
-	if (first)
-		return;
-
-	/* needed for the following wake up */
-	smp_mb__after_atomic();
-
+	if (!first) {
 		if (unlikely(atomic_read(&req->task->io_uring->in_cancel))) {
 			io_move_task_work_from_local(ctx);
 			return;
 		}
-
 		if (ctx->flags & IORING_SETUP_TASKRUN_FLAG)
 			atomic_or(IORING_SQ_TASKRUN, &ctx->rings->sq_flags);
 		if (ctx->has_evfd)
 			io_eventfd_signal(ctx);
+	}
 
-	if (READ_ONCE(ctx->cq_waiting))
+	nr_wait = atomic_read(&ctx->cq_wait_nr);
+	/* no one is waiting */
+	if (!nr_wait)
+		return;
+	/* either not enough or the previous add has already woken it up */
+	if (nr_wait > nr_tw || nr_tw_prev >= nr_wait)
+		return;
+	/* pairs with set_current_state() in io_cqring_wait() */
+	smp_mb__after_atomic();
 	wake_up_state(ctx->submitter_task, TASK_INTERRUPTIBLE);
 }
 
@@ -1339,7 +1363,7 @@ void __io_req_task_work_add(struct io_kiocb *req, unsigned flags)
 	if (!(flags & IOU_F_TWQ_FORCE_NORMAL) &&
 	    (ctx->flags & IORING_SETUP_DEFER_TASKRUN)) {
 		rcu_read_lock();
-		io_req_local_work_add(req);
+		io_req_local_work_add(req, flags);
 		rcu_read_unlock();
 		return;
 	}
@@ -2625,7 +2649,9 @@ static int io_cqring_wait(struct io_ring_ctx *ctx, int min_events,
 		unsigned long check_cq;
 
 		if (ctx->flags & IORING_SETUP_DEFER_TASKRUN) {
-			WRITE_ONCE(ctx->cq_waiting, 1);
+			int nr_wait = (int) iowq.cq_tail - READ_ONCE(ctx->rings->cq.tail);
+
+			atomic_set(&ctx->cq_wait_nr, nr_wait);
 			set_current_state(TASK_INTERRUPTIBLE);
 		} else {
 			prepare_to_wait_exclusive(&ctx->cq_wait, &iowq.wq,
@@ -2634,7 +2660,7 @@ static int io_cqring_wait(struct io_ring_ctx *ctx, int min_events,
 
 		ret = io_cqring_wait_schedule(ctx, &iowq);
 		__set_current_state(TASK_RUNNING);
-		WRITE_ONCE(ctx->cq_waiting, 0);
+		atomic_set(&ctx->cq_wait_nr, 0);
 
 		if (ret < 0)
 			break;
@@ -4517,7 +4543,7 @@ static int __init io_uring_init(void)
 	io_uring_optable_init();
 
 	req_cachep = KMEM_CACHE(io_kiocb, SLAB_HWCACHE_ALIGN | SLAB_PANIC |
-				SLAB_ACCOUNT);
+				SLAB_ACCOUNT | SLAB_TYPESAFE_BY_RCU);
 	return 0;
 };
 __initcall(io_uring_init);

io_uring/io_uring.h

@@ -18,6 +18,15 @@
 enum {
 	/* don't use deferred task_work */
 	IOU_F_TWQ_FORCE_NORMAL			= 1,
+
+	/*
+	 * A hint to not wake right away but delay until there are enough of
+	 * tw's queued to match the number of CQEs the task is waiting for.
+	 *
+	 * Must not be used wirh requests generating more than one CQE.
+	 * It's also ignored unless IORING_SETUP_DEFER_TASKRUN is set.
+	 */
+	IOU_F_TWQ_LAZY_WAKE			= 2,
 };
 
 enum {

io_uring/notif.c

@@ -31,7 +31,7 @@ static void io_tx_ubuf_callback(struct sk_buff *skb, struct ubuf_info *uarg,
 	struct io_kiocb *notif = cmd_to_io_kiocb(nd);
 
 	if (refcount_dec_and_test(&uarg->refcnt))
-		io_req_task_work_add(notif);
+		__io_req_task_work_add(notif, IOU_F_TWQ_LAZY_WAKE);
 }
 
 static void io_tx_ubuf_callback_ext(struct sk_buff *skb, struct ubuf_info *uarg,

io_uring/notif.h

@@ -33,7 +33,7 @@ static inline void io_notif_flush(struct io_kiocb *notif)
 
 	/* drop slot's master ref */
 	if (refcount_dec_and_test(&nd->uarg.refcnt))
-		io_req_task_work_add(notif);
+		__io_req_task_work_add(notif, IOU_F_TWQ_LAZY_WAKE);
 }
 
 static inline int io_notif_account_mem(struct io_kiocb *notif, unsigned len)

io_uring/rw.c

@@ -304,7 +304,7 @@ static void io_complete_rw(struct kiocb *kiocb, long res)
 		return;
 	io_req_set_res(req, io_fixup_rw_res(req, res), 0);
 	req->io_task_work.func = io_req_rw_complete;
-	io_req_task_work_add(req);
+	__io_req_task_work_add(req, IOU_F_TWQ_LAZY_WAKE);
 }
 
 static void io_complete_rw_iopoll(struct kiocb *kiocb, long res)