[PATCH] io_uring: make POLL_ADD/POLL_REMOVE scale better

[PATCH] io_uring: make POLL_ADD/POLL_REMOVE scale better

[Jens Axboe]

One of the obvious use cases for these commands is networking, where
it's not uncommon to have tons of sockets open and polled for. The
current implementation uses a list for insertion and lookup, which works
fine for file based use cases where the count is usually low, it breaks
down somewhat for higher number of files / sockets. A test case with
30k sockets being polled for and cancelled takes:

real    0m6.968s
user    0m0.002s
sys     0m6.936s

with the patch it takes:

real    0m0.233s
user    0m0.010s
sys     0m0.176s

If you go to 50k sockets, it gets even more abysmal with the current
code:

real    0m40.602s
user    0m0.010s
sys     0m40.555s

with the patch it takes:

real    0m0.398s
user    0m0.000s
sys     0m0.341s

Change is pretty straight forward, just replace the cancel_list with
a red/black tree instead.

Signed-off-by: Jens Axboe <axboe@kernel.dk>

---

diff --git a/fs/io_uring.c b/fs/io_uring.c
index 55f8b1d378df..5ad652fa24b8 100644
--- a/fs/io_uring.c
+++ b/fs/io_uring.c
@@ -271,7 +271,7 @@ struct io_ring_ctx {
 		 * manipulate the list, hence no extra locking is needed there.
 		 */
 		struct list_head	poll_list;
-		struct list_head	cancel_list;
+		struct rb_root		cancel_tree;
 
 		spinlock_t		inflight_lock;
 		struct list_head	inflight_list;
@@ -323,7 +323,10 @@ struct io_kiocb {
 	struct sqe_submit	submit;
 
 	struct io_ring_ctx	*ctx;
-	struct list_head	list;
+	union {
+		struct list_head	list;
+		struct rb_node		rb_node;
+	};
 	struct list_head	link_list;
 	unsigned int		flags;
 	refcount_t		refs;
@@ -433,7 +436,7 @@ static struct io_ring_ctx *io_ring_ctx_alloc(struct io_uring_params *p)
 	init_waitqueue_head(&ctx->wait);
 	spin_lock_init(&ctx->completion_lock);
 	INIT_LIST_HEAD(&ctx->poll_list);
-	INIT_LIST_HEAD(&ctx->cancel_list);
+	ctx->cancel_tree = RB_ROOT;
 	INIT_LIST_HEAD(&ctx->defer_list);
 	INIT_LIST_HEAD(&ctx->timeout_list);
 	init_waitqueue_head(&ctx->inflight_wait);
@@ -1934,6 +1937,14 @@ static int io_accept(struct io_kiocb *req, const struct io_uring_sqe *sqe,
 #endif
 }
 
+static inline void io_poll_remove_req(struct io_kiocb *req)
+{
+	if (!RB_EMPTY_NODE(&req->rb_node)) {
+		rb_erase(&req->rb_node, &req->ctx->cancel_tree);
+		RB_CLEAR_NODE(&req->rb_node);
+	}
+}
+
 static void io_poll_remove_one(struct io_kiocb *req)
 {
 	struct io_poll_iocb *poll = &req->poll;
@@ -1945,17 +1956,17 @@ static void io_poll_remove_one(struct io_kiocb *req)
 		io_queue_async_work(req);
 	}
 	spin_unlock(&poll->head->lock);
-
-	list_del_init(&req->list);
+	io_poll_remove_req(req);
 }
 
 static void io_poll_remove_all(struct io_ring_ctx *ctx)
 {
+	struct rb_node *node;
 	struct io_kiocb *req;
 
 	spin_lock_irq(&ctx->completion_lock);
-	while (!list_empty(&ctx->cancel_list)) {
-		req = list_first_entry(&ctx->cancel_list, struct io_kiocb,list);
+	while ((node = rb_first(&ctx->cancel_tree)) != NULL) {
+		req = rb_entry(node, struct io_kiocb, rb_node);
 		io_poll_remove_one(req);
 	}
 	spin_unlock_irq(&ctx->completion_lock);
@@ -1963,13 +1974,21 @@ static void io_poll_remove_all(struct io_ring_ctx *ctx)
 
 static int io_poll_cancel(struct io_ring_ctx *ctx, __u64 sqe_addr)
 {
+	struct rb_node *p, *parent = NULL;
 	struct io_kiocb *req;
 
-	list_for_each_entry(req, &ctx->cancel_list, list) {
-		if (req->user_data != sqe_addr)
-			continue;
-		io_poll_remove_one(req);
-		return 0;
+	p = ctx->cancel_tree.rb_node;
+	while (p) {
+		parent = p;
+		req = rb_entry(parent, struct io_kiocb, rb_node);
+		if (sqe_addr < req->user_data) {
+			p = p->rb_left;
+		} else if (sqe_addr > req->user_data) {
+			p = p->rb_right;
+		} else {
+			io_poll_remove_one(req);
+			return 0;
+		}
 	}
 
 	return -ENOENT;
@@ -2039,7 +2058,7 @@ static void io_poll_complete_work(struct io_wq_work **workptr)
 		spin_unlock_irq(&ctx->completion_lock);
 		return;
 	}
-	list_del_init(&req->list);
+	io_poll_remove_req(req);
 	io_poll_complete(req, mask);
 	spin_unlock_irq(&ctx->completion_lock);
 
@@ -2073,7 +2092,7 @@ static int io_poll_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
 	 * for finalizing the request, mark us as having grabbed that already.
 	 */
 	if (mask && spin_trylock_irqsave(&ctx->completion_lock, flags)) {
-		list_del(&req->list);
+		io_poll_remove_req(req);
 		io_poll_complete(req, mask);
 		req->flags |= REQ_F_COMP_LOCKED;
 		io_put_req(req);
@@ -2108,6 +2127,25 @@ static void io_poll_queue_proc(struct file *file, struct wait_queue_head *head,
 	add_wait_queue(head, &pt->req->poll.wait);
 }
 
+static void io_poll_req_insert(struct io_kiocb *req)
+{
+	struct io_ring_ctx *ctx = req->ctx;
+	struct rb_node **p = &ctx->cancel_tree.rb_node;
+	struct rb_node *parent = NULL;
+	struct io_kiocb *tmp;
+
+	while (*p) {
+		parent = *p;
+		tmp = rb_entry(parent, struct io_kiocb, rb_node);
+		if (req->user_data < tmp->user_data)
+			p = &(*p)->rb_left;
+		else
+			p = &(*p)->rb_right;
+	}
+	rb_link_node(&req->rb_node, parent, p);
+	rb_insert_color(&req->rb_node, &ctx->cancel_tree);
+}
+
 static int io_poll_add(struct io_kiocb *req, const struct io_uring_sqe *sqe,
 		       struct io_kiocb **nxt)
 {
@@ -2129,6 +2167,7 @@ static int io_poll_add(struct io_kiocb *req, const struct io_uring_sqe *sqe,
 	INIT_IO_WORK(&req->work, io_poll_complete_work);
 	events = READ_ONCE(sqe->poll_events);
 	poll->events = demangle_poll(events) | EPOLLERR | EPOLLHUP;
+	RB_CLEAR_NODE(&req->rb_node);
 
 	poll->head = NULL;
 	poll->done = false;
@@ -2161,7 +2200,7 @@ static int io_poll_add(struct io_kiocb *req, const struct io_uring_sqe *sqe,
 		else if (cancel)
 			WRITE_ONCE(poll->canceled, true);
 		else if (!poll->done) /* actually waiting for an event */
-			list_add_tail(&req->list, &ctx->cancel_list);
+			io_poll_req_insert(req);
 		spin_unlock(&poll->head->lock);
 	}
 	if (mask) { /* no async, we'd stolen it */

-- 
Jens Axboe