Skip to contentSkip to conversation

singly linked list for chains

PATCH In review Pavel Begunkov 2 participants 8 messages Oct 28, 2020 – Oct 30, 2020

Subsystems (2)
lore
Share & tools
lore
#3 lore raw

PATCH v2 3/4 io_uring: link requests with singly linked list

In reply to Pavel Begunkov

Singly linked list for keeping linked requests is enough, because we almost always operate on the head and traverse forward with the exception of linked timeouts going 1 hop backwards.

Replace ->link_list with a handmade singly linked list. Also kill REQ_F_LINK_HEAD in favour of checking a newly added ->list for NULL directly.

That saves 8B in io_kiocb, is not as heavy as list fixup, makes better use of cache by not touching a previous request (i.e. last request of the link) each time on list modification and optimises cache use further in the following patch, and actually makes travesal easier removing in the end some lines. Also, keeping invariant in ->list instead of having REQ_F_LINK_HEAD is less error-prone.

---
 fs/io_uring.c | 132 ++++++++++++++++++++------------------------------
 1 file changed, 52 insertions(+), 80 deletions(-)
fs/io_uring.c +52 -80
index 4eb1cb8a8bf8..b8b5131dd017 100644
@@ -577,7 +577,6 @@ enum {
 	REQ_F_FORCE_ASYNC_BIT	= IOSQE_ASYNC_BIT,
 	REQ_F_BUFFER_SELECT_BIT	= IOSQE_BUFFER_SELECT_BIT,
 
-	REQ_F_LINK_HEAD_BIT,
 	REQ_F_FAIL_LINK_BIT,
 	REQ_F_INFLIGHT_BIT,
 	REQ_F_CUR_POS_BIT,
@@ -609,8 +608,6 @@ enum {
 	/* IOSQE_BUFFER_SELECT */
 	REQ_F_BUFFER_SELECT	= BIT(REQ_F_BUFFER_SELECT_BIT),
 
-	/* head of a link */
-	REQ_F_LINK_HEAD		= BIT(REQ_F_LINK_HEAD_BIT),
 	/* fail rest of links */
 	REQ_F_FAIL_LINK		= BIT(REQ_F_FAIL_LINK_BIT),
 	/* on inflight list */
@@ -689,7 +686,7 @@ struct io_kiocb {
 	struct task_struct		*task;
 	u64				user_data;
 
-	struct list_head		link_list;
+	struct io_kiocb			*link;
 
 	/*
 	 * 1. used with ctx->iopoll_list with reads/writes
@@ -986,6 +983,9 @@ struct sock *io_uring_get_socket(struct file *file)
 }
 EXPORT_SYMBOL(io_uring_get_socket);
 
+#define io_for_each_link(pos, head) \
+	for (pos = (head); pos; pos = pos->link)
+
 static inline void io_clean_op(struct io_kiocb *req)
 {
 	if (req->flags & (REQ_F_NEED_CLEANUP | REQ_F_BUFFER_SELECTED |
@@ -1404,10 +1404,8 @@ static void io_prep_async_link(struct io_kiocb *req)
 {
 	struct io_kiocb *cur;
 
-	io_prep_async_work(req);
-	if (req->flags & REQ_F_LINK_HEAD)
-		list_for_each_entry(cur, &req->link_list, link_list)
-			io_prep_async_work(cur);
+	io_for_each_link(cur, req)
+		io_prep_async_work(cur);
 }
 
 static struct io_kiocb *__io_queue_async_work(struct io_kiocb *req)
@@ -1854,6 +1852,14 @@ static void __io_free_req(struct io_kiocb *req)
 	percpu_ref_put(&ctx->refs);
 }
 
+static inline void io_remove_next_linked(struct io_kiocb *req)
+{
+	struct io_kiocb *nxt = req->link;
+
+	req->link = nxt->link;
+	nxt->link = NULL;
+}
+
 static void io_kill_linked_timeout(struct io_kiocb *req)
 {
 	struct io_ring_ctx *ctx = req->ctx;
@@ -1862,8 +1868,8 @@ static void io_kill_linked_timeout(struct io_kiocb *req)
 	unsigned long flags;
 
 	spin_lock_irqsave(&ctx->completion_lock, flags);
-	link = list_first_entry_or_null(&req->link_list, struct io_kiocb,
-					link_list);
+	link = req->link;
+
 	/*
 	 * Can happen if a linked timeout fired and link had been like
 	 * req -> link t-out -> link t-out [-> ...]
@@ -1872,7 +1878,7 @@ static void io_kill_linked_timeout(struct io_kiocb *req)
 		struct io_timeout_data *io = link->async_data;
 		int ret;
 
-		list_del_init(&link->link_list);
+		io_remove_next_linked(req);
 		link->timeout.head = NULL;
 		ret = hrtimer_try_to_cancel(&io->timer);
 		if (ret != -1) {
@@ -1890,41 +1896,22 @@ static void io_kill_linked_timeout(struct io_kiocb *req)
 	}
 }
 
-static struct io_kiocb *io_req_link_next(struct io_kiocb *req)
-{
-	struct io_kiocb *nxt;
 
-	/*
-	 * The list should never be empty when we are called here. But could
-	 * potentially happen if the chain is messed up, check to be on the
-	 * safe side.
-	 */
-	if (unlikely(list_empty(&req->link_list)))
-		return NULL;
-
-	nxt = list_first_entry(&req->link_list, struct io_kiocb, link_list);
-	list_del_init(&req->link_list);
-	if (!list_empty(&nxt->link_list))
-		nxt->flags |= REQ_F_LINK_HEAD;
-	return nxt;
-}
-
-/*
- * Called if REQ_F_LINK_HEAD is set, and we fail the head request
- */
 static void io_fail_links(struct io_kiocb *req)
 {
+	struct io_kiocb *link, *nxt;
 	struct io_ring_ctx *ctx = req->ctx;
 	unsigned long flags;
 
 	spin_lock_irqsave(&ctx->completion_lock, flags);
-	while (!list_empty(&req->link_list)) {
-		struct io_kiocb *link = list_first_entry(&req->link_list,
-						struct io_kiocb, link_list);
+	link = req->link;
+	req->link = NULL;
 
-		list_del_init(&link->link_list);
-		trace_io_uring_fail_link(req, link);
+	while (link) {
+		nxt = link->link;
+		link->link = NULL;
 
+		trace_io_uring_fail_link(req, link);
 		io_cqring_fill_event(link, -ECANCELED);
 
 		/*
@@ -1936,8 +1923,8 @@ static void io_fail_links(struct io_kiocb *req)
 			io_put_req_deferred(link, 2);
 		else
 			io_double_put_req(link);
+		link = nxt;
 	}
-
 	io_commit_cqring(ctx);
 	spin_unlock_irqrestore(&ctx->completion_lock, flags);
 
@@ -1946,7 +1933,6 @@ static void io_fail_links(struct io_kiocb *req)
 
 static struct io_kiocb *__io_req_find_next(struct io_kiocb *req)
 {
-	req->flags &= ~REQ_F_LINK_HEAD;
 	if (req->flags & REQ_F_LINK_TIMEOUT)
 		io_kill_linked_timeout(req);
 
@@ -1956,15 +1942,19 @@ static struct io_kiocb *__io_req_find_next(struct io_kiocb *req)
 	 * dependencies to the next request. In case of failure, fail the rest
 	 * of the chain.
 	 */
-	if (likely(!(req->flags & REQ_F_FAIL_LINK)))
-		return io_req_link_next(req);
+	if (likely(!(req->flags & REQ_F_FAIL_LINK))) {
+		struct io_kiocb *nxt = req->link;
+
+		req->link = NULL;
+		return nxt;
+	}
 	io_fail_links(req);
 	return NULL;
 }
 
-static struct io_kiocb *io_req_find_next(struct io_kiocb *req)
+static inline struct io_kiocb *io_req_find_next(struct io_kiocb *req)
 {
-	if (likely(!(req->flags & REQ_F_LINK_HEAD)))
+	if (likely(!(req->link) && !(req->flags & REQ_F_LINK_TIMEOUT)))
 		return NULL;
 	return __io_req_find_next(req);
 }
@@ -2058,7 +2048,7 @@ static void io_req_task_queue(struct io_kiocb *req)
 	}
 }
 
-static void io_queue_next(struct io_kiocb *req)
+static inline void io_queue_next(struct io_kiocb *req)
 {
 	struct io_kiocb *nxt = io_req_find_next(req);
 
@@ -2115,8 +2105,7 @@ static void io_req_free_batch(struct req_batch *rb, struct io_kiocb *req)
 		io_free_req(req);
 		return;
 	}
-	if (req->flags & REQ_F_LINK_HEAD)
-		io_queue_next(req);
+	io_queue_next(req);
 
 	if (req->task != rb->task) {
 		if (rb->task) {
@@ -5751,11 +5740,10 @@ static u32 io_get_sequence(struct io_kiocb *req)
 {
 	struct io_kiocb *pos;
 	struct io_ring_ctx *ctx = req->ctx;
-	u32 total_submitted, nr_reqs = 1;
+	u32 total_submitted, nr_reqs = 0;
 
-	if (req->flags & REQ_F_LINK_HEAD)
-		list_for_each_entry(pos, &req->link_list, link_list)
-			nr_reqs++;
+	io_for_each_link(pos, req)
+		nr_reqs++;
 
 	total_submitted = ctx->cached_sq_head - ctx->cached_sq_dropped;
 	return total_submitted - nr_reqs;
@@ -6100,7 +6088,7 @@ static enum hrtimer_restart io_link_timeout_fn(struct hrtimer *timer)
 	 * race with the completion of the linked work.
 	 */
 	if (prev && refcount_inc_not_zero(&prev->refs))
-		list_del_init(&req->link_list);
+		io_remove_next_linked(prev);
 	else
 		prev = NULL;
 	spin_unlock_irqrestore(&ctx->completion_lock, flags);
@@ -6118,8 +6106,8 @@ static enum hrtimer_restart io_link_timeout_fn(struct hrtimer *timer)
 static void __io_queue_linked_timeout(struct io_kiocb *req)
 {
 	/*
-	 * If the list is now empty, then our linked request finished before
-	 * we got a chance to setup the timer
+	 * If the back reference is NULL, then our linked request finished
+	 * before we got a chance to setup the timer
 	 */
 	if (req->timeout.head) {
 		struct io_timeout_data *data = req->async_data;
@@ -6144,16 +6132,10 @@ static void io_queue_linked_timeout(struct io_kiocb *req)
 
 static struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req)
 {
-	struct io_kiocb *nxt;
+	struct io_kiocb *nxt = req->link;
 
-	if (!(req->flags & REQ_F_LINK_HEAD))
-		return NULL;
-	if (req->flags & REQ_F_LINK_TIMEOUT)
-		return NULL;
-
-	nxt = list_first_entry_or_null(&req->link_list, struct io_kiocb,
-					link_list);
-	if (!nxt || nxt->opcode != IORING_OP_LINK_TIMEOUT)
+	if (!nxt || (req->flags & REQ_F_LINK_TIMEOUT) ||
+	    nxt->opcode != IORING_OP_LINK_TIMEOUT)
 		return NULL;
 
 	nxt->timeout.head = req;
@@ -6301,7 +6283,7 @@ static int io_submit_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe,
 			return ret;
 		}
 		trace_io_uring_link(ctx, req, head);
-		list_add_tail(&req->link_list, &head->link_list);
+		link->last->link = req;
 		link->last = req;
 
 		/* last request of a link, enqueue the link */
@@ -6315,9 +6297,6 @@ static int io_submit_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe,
 			ctx->drain_next = 0;
 		}
 		if (req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) {
-			req->flags |= REQ_F_LINK_HEAD;
-			INIT_LIST_HEAD(&req->link_list);
-
 			ret = io_req_defer_prep(req, sqe);
 			if (unlikely(ret))
 				req->flags |= REQ_F_FAIL_LINK;
@@ -6450,6 +6429,7 @@ static int io_init_req(struct io_ring_ctx *ctx, struct io_kiocb *req,
 	req->file = NULL;
 	req->ctx = ctx;
 	req->flags = 0;
+	req->link = NULL;
 	/* one is dropped after submission, the other at completion */
 	refcount_set(&req->refs, 2);
 	req->task = current;
@@ -8385,29 +8365,21 @@ static bool io_match_link(struct io_kiocb *preq, struct io_kiocb *req)
 {
 	struct io_kiocb *link;
 
-	if (!(preq->flags & REQ_F_LINK_HEAD))
-		return false;
-
-	list_for_each_entry(link, &preq->link_list, link_list) {
+	io_for_each_link(link, preq->link) {
 		if (link == req)
 			return true;
 	}
-
 	return false;
 }
 
-static bool io_match_link_files(struct io_kiocb *req,
+static bool io_match_link_files(struct io_kiocb *head,
 				struct files_struct *files)
 {
-	struct io_kiocb *link;
+	struct io_kiocb *req;
 
-	if (io_match_files(req, files))
-		return true;
-	if (req->flags & REQ_F_LINK_HEAD) {
-		list_for_each_entry(link, &req->link_list, link_list) {
-			if (io_match_files(link, files))
-				return true;
-		}
+	io_for_each_link(req, head) {
+		if (io_match_files(req, files))
+			return true;
 	}
 	return false;
 }
Signature · 1 line
2.24.0
#4 lore raw

PATCH v2 4/4 io_uring: toss io_kiocb fields for better caching

In reply to Pavel Begunkov

We've got extra 8 bytes in the 2nd cacheline, put ->fixed_file_refs there, so inline execution path mostly doesn't touch the 3rd cacheline for fixed_file requests as well.

---
 fs/io_uring.c | 3 +--
 1 file changed, 1 insertion(+), 2 deletions(-)
fs/io_uring.c +1 -2
index b8b5131dd017..4298396028cc 100644
@@ -687,14 +687,13 @@ struct io_kiocb {
 	u64				user_data;
 
 	struct io_kiocb			*link;
+	struct percpu_ref		*fixed_file_refs;
 
 	/*
 	 * 1. used with ctx->iopoll_list with reads/writes
 	 * 2. to track reqs with ->files (see io_op_def::file_table)
 	 */
 	struct list_head		inflight_entry;
-
-	struct percpu_ref		*fixed_file_refs;
 	struct callback_head		task_work;
 	/* for polled requests, i.e. IORING_OP_POLL_ADD and async armed poll */
 	struct hlist_node		hash_node;
Signature · 1 line
2.24.0
#2 lore raw

PATCH v2 2/4 io_uring: track link timeout's master explicitly

In reply to Pavel Begunkov

In preparation for converting singly linked lists for chaining requests, make linked timeouts save requests that they're responsible for and not count on doubly linked list for back referencing.

---
 fs/io_uring.c | 24 ++++++++++++------------
 1 file changed, 12 insertions(+), 12 deletions(-)
fs/io_uring.c +12 -12
index ce092d6cab73..4eb1cb8a8bf8 100644
@@ -445,6 +445,8 @@ struct io_timeout {
 	u32				off;
 	u32				target_seq;
 	struct list_head		list;
+	/* head of the link, used by linked timeouts only */
+	struct io_kiocb			*head;
 };
 
 struct io_timeout_rem {
@@ -1871,6 +1873,7 @@ static void io_kill_linked_timeout(struct io_kiocb *req)
 		int ret;
 
 		list_del_init(&link->link_list);
+		link->timeout.head = NULL;
 		ret = hrtimer_try_to_cancel(&io->timer);
 		if (ret != -1) {
 			io_cqring_fill_event(link, -ECANCELED);
@@ -6084,26 +6087,22 @@ static enum hrtimer_restart io_link_timeout_fn(struct hrtimer *timer)
 {
 	struct io_timeout_data *data = container_of(timer,
 						struct io_timeout_data, timer);
-	struct io_kiocb *req = data->req;
+	struct io_kiocb *prev, *req = data->req;
 	struct io_ring_ctx *ctx = req->ctx;
-	struct io_kiocb *prev = NULL;
 	unsigned long flags;
 
 	spin_lock_irqsave(&ctx->completion_lock, flags);
+	prev = req->timeout.head;
+	req->timeout.head = NULL;
 
 	/*
 	 * We don't expect the list to be empty, that will only happen if we
 	 * race with the completion of the linked work.
 	 */
-	if (!list_empty(&req->link_list)) {
-		prev = list_entry(req->link_list.prev, struct io_kiocb,
-				  link_list);
-		if (refcount_inc_not_zero(&prev->refs))
-			list_del_init(&req->link_list);
-		else
-			prev = NULL;
-	}
-
+	if (prev && refcount_inc_not_zero(&prev->refs))
+		list_del_init(&req->link_list);
+	else
+		prev = NULL;
 	spin_unlock_irqrestore(&ctx->completion_lock, flags);
 
 	if (prev) {
@@ -6122,7 +6121,7 @@ static void __io_queue_linked_timeout(struct io_kiocb *req)
 	 * If the list is now empty, then our linked request finished before
 	 * we got a chance to setup the timer
 	 */
-	if (!list_empty(&req->link_list)) {
+	if (req->timeout.head) {
 		struct io_timeout_data *data = req->async_data;
 
 		data->timer.function = io_link_timeout_fn;
@@ -6157,6 +6156,7 @@ static struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req)
 	if (!nxt || nxt->opcode != IORING_OP_LINK_TIMEOUT)
 		return NULL;
 
+	nxt->timeout.head = req;
 	nxt->flags |= REQ_F_LTIMEOUT_ACTIVE;
 	req->flags |= REQ_F_LINK_TIMEOUT;
 	return nxt;
Signature · 1 line
2.24.0
#0 lore raw

v2 cover singly linked list for chains

v2: split the meat patch.

I did not re-benchmark it, but as stated before: 5030 vs 5160 KIOPS by a naive (consistent) nop benchmark that submits 32 linked nops and then submits them in a loop. Worth trying with some better hardware.

Pavel Begunkov (4):
io_uring: track link's head and tail during submit
io_uring: track link timeout's master explicitly
io_uring: link requests with singly linked list
io_uring: toss io_kiocb fields for better caching

 fs/io_uring.c | 181 ++++++++++++++++++++++----------------------------
 1 file changed, 80 insertions(+), 101 deletions(-)
Signature · 1 line
2.24.0
#1 lore raw

PATCH v2 1/4 io_uring: track link's head and tail during submit

Explicitly save not only a link's head in io_submit_sqe[s]() but the tail as well. That's in preparation for keeping linked requests in a singly linked list.

---
 fs/io_uring.c | 24 ++++++++++++++++--------
 1 file changed, 16 insertions(+), 8 deletions(-)
fs/io_uring.c +16 -8
index 3d244c61a5c3..ce092d6cab73 100644
@@ -6262,8 +6262,13 @@ static inline void io_queue_link_head(struct io_kiocb *req,
 		io_queue_sqe(req, NULL, cs);
 }
 
+struct io_submit_link {
+	struct io_kiocb *head;
+	struct io_kiocb *last;
+};
+
 static int io_submit_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe,
-			 struct io_kiocb **link, struct io_comp_state *cs)
+			 struct io_submit_link *link, struct io_comp_state *cs)
 {
 	struct io_ring_ctx *ctx = req->ctx;
 	int ret;
@@ -6275,8 +6280,8 @@ static int io_submit_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe,
 	 * submitted sync once the chain is complete. If none of those
 	 * conditions are true (normal request), then just queue it.
 	 */
-	if (*link) {
-		struct io_kiocb *head = *link;
+	if (link->head) {
+		struct io_kiocb *head = link->head;
 
 		/*
 		 * Taking sequential execution of a link, draining both sides
@@ -6297,11 +6302,12 @@ static int io_submit_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe,
 		}
 		trace_io_uring_link(ctx, req, head);
 		list_add_tail(&req->link_list, &head->link_list);
+		link->last = req;
 
 		/* last request of a link, enqueue the link */
 		if (!(req->flags & (REQ_F_LINK | REQ_F_HARDLINK))) {
 			io_queue_link_head(head, cs);
-			*link = NULL;
+			link->head = NULL;
 		}
 	} else {
 		if (unlikely(ctx->drain_next)) {
@@ -6315,7 +6321,8 @@ static int io_submit_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe,
 			ret = io_req_defer_prep(req, sqe);
 			if (unlikely(ret))
 				req->flags |= REQ_F_FAIL_LINK;
-			*link = req;
+			link->head = req;
+			link->last = req;
 		} else {
 			io_queue_sqe(req, sqe, cs);
 		}
@@ -6495,7 +6502,7 @@ static int io_init_req(struct io_ring_ctx *ctx, struct io_kiocb *req,
 static int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr)
 {
 	struct io_submit_state state;
-	struct io_kiocb *link = NULL;
+	struct io_submit_link link;
 	int i, submitted = 0;
 
 	/* if we have a backlog and couldn't flush it all, return BUSY */
@@ -6515,6 +6522,7 @@ static int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr)
 	refcount_add(nr, &current->usage);
 
 	io_submit_state_start(&state, ctx, nr);
+	link.head = NULL;
 
 	for (i = 0; i < nr; i++) {
 		const struct io_uring_sqe *sqe;
@@ -6560,8 +6568,8 @@ static int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr)
 		percpu_counter_sub(&tctx->inflight, unused);
 		put_task_struct_many(current, unused);
 	}
-	if (link)
-		io_queue_link_head(link, &state.comp);
+	if (link.head)
+		io_queue_link_head(link.head, &state.comp);
 	io_submit_state_end(&state);
 
 	 /* Commit SQ ring head once we've consumed and submitted all SQEs */
Signature · 1 line
2.24.0
maintainer
#5 lore raw
In reply to Pavel Begunkov
Quoted 5 lines · Pavel Begunkov

v2: split the meat patch.

I did not re-benchmark it, but as stated before: 5030 vs 5160 KIOPS by a naive (consistent) nop benchmark that submits 32 linked nops and then submits them in a loop. Worth trying with some better hardware.

Applied for 5.11 - I'll give it a spin here too and see if I find any improvements. But a smaller io_kiocb is always a win.

Signature · 1 line
Jens Axboe
#6 lore raw
Quoted 9 lines · Jens Axboe
Quoted 5 lines · Pavel Begunkov

v2: split the meat patch.

I did not re-benchmark it, but as stated before: 5030 vs 5160 KIOPS by a naive (consistent) nop benchmark that submits 32 linked nops and then submits them in a loop. Worth trying with some better hardware.

Applied for 5.11 - I'll give it a spin here too and see if I find any improvements. But a smaller io_kiocb is always a win.

Might be a good time for a generic single list. There are at least 3 places I remember: bio, io-wq, io_uring

Signature · 1 line
Pavel Begunkov
maintainer
#7 lore raw
Quoted 8 lines · Pavel Begunkov
Quoted 7 lines · Jens Axboe
Quoted 5 lines · Pavel Begunkov

v2: split the meat patch.

I did not re-benchmark it, but as stated before: 5030 vs 5160 KIOPS by a naive (consistent) nop benchmark that submits 32 linked nops and then submits them in a loop. Worth trying with some better hardware.

Quoted 5 lines
Quoted 2 lines

Applied for 5.11 - I'll give it a spin here too and see if I find any improvements. But a smaller io_kiocb is always a win.

Might be a good time for a generic single list. There are at least 3 places I remember: bio, io-wq, io_uring

Yes, kind of weird that we don't have one. There's the normal doubly linked one, and then there's the single head for hash tables. But I could not find anything good for io-wq, which is why I rolled my own. Would be nice to turn it into a generic one and reuse it everywhere (and I bet it would grow more users).

The lockless lists (llist) is singular, but overkill if you already have locking you're under.

Signature · 1 line
Jens Axboe