Skip to contentSkip to conversation

CQ vs SQ timeout accounting

PATCH In review Pavel Begunkov 2 participants 9 messages Apr 18, 2020 – Apr 22, 2020

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

cover CQ vs SQ timeout accounting

The first patch do the job, and triggers after __any__ CQEs, including ones from other timeouts. And it obviously fails to pass the tests. The second patch makes it to count only non-timeout CQEs, and passes the tests. Feel free to squash them.

Fixes Hrvoje's issue. Also, there is one more bug to be fixed. It's related to bulk completion, and manifests itself with both implementations.

Pavel Begunkov (2):
io_uring: trigger timeout after any sqe->off CQEs
io_uring: don't trigger timeout with another t-out

 fs/io_uring.c | 107 ++++++++++++++++----------------------------------
 1 file changed, 33 insertions(+), 74 deletions(-)
Signature · 1 line
2.24.0
#1 lore raw

PATCH 1/2 io_uring: trigger timeout after any sqe->off CQEs

sequence mode timeouts wait not for sqe->off CQEs, but rather sqe->off + number of prior inflight requests with a quirk ignoring other timeouts completions. Wait exactly for sqe->off using completion count (tail) for accounting.

---
 fs/io_uring.c | 120 +++++++++++++++++++-------------------------------
 1 file changed, 46 insertions(+), 74 deletions(-)
fs/io_uring.c +46 -74
index 8ee7b4f72b8f..c34c5dcc90fc 100644
@@ -384,7 +384,8 @@ struct io_timeout {
 	struct file			*file;
 	u64				addr;
 	int				flags;
-	u32				count;
+	u32				off;
+	u32				target_cq;
 };
 
 struct io_rw {
@@ -982,23 +983,6 @@ static struct io_kiocb *io_get_deferred_req(struct io_ring_ctx *ctx)
 	return NULL;
 }
 
-static struct io_kiocb *io_get_timeout_req(struct io_ring_ctx *ctx)
-{
-	struct io_kiocb *req;
-
-	req = list_first_entry_or_null(&ctx->timeout_list, struct io_kiocb, list);
-	if (req) {
-		if (req->flags & REQ_F_TIMEOUT_NOSEQ)
-			return NULL;
-		if (!__req_need_defer(req)) {
-			list_del_init(&req->list);
-			return req;
-		}
-	}
-
-	return NULL;
-}
-
 static void __io_commit_cqring(struct io_ring_ctx *ctx)
 {
 	struct io_rings *rings = ctx->rings;
@@ -1114,12 +1098,42 @@ static void io_kill_timeouts(struct io_ring_ctx *ctx)
 	spin_unlock_irq(&ctx->completion_lock);
 }
 
+static inline bool io_check_in_range(u32 pos, u32 start, u32 end)
+{
+	/* if @end < @start, check for [end, MAX_UINT] + [MAX_UINT, start] */
+	return (pos - start) <= (end - start);
+}
+
+static void __io_flush_timeouts(struct io_ring_ctx *ctx)
+{
+	u32 end, start;
+
+	start = end = ctx->cached_cq_tail;
+	do {
+		struct io_kiocb *req = list_first_entry(&ctx->timeout_list,
+							struct io_kiocb, list);
+
+		if (req->flags & REQ_F_TIMEOUT_NOSEQ)
+			break;
+		/*
+		 * multiple timeouts may have the same target,
+		 * check that @req is in [first_tail, cur_tail]
+		 */
+		if (!io_check_in_range(req->timeout.target_cq, start, end))
+			break;
+
+		list_del_init(&req->list);
+		io_kill_timeout(req);
+		end = ctx->cached_cq_tail;
+	} while (!list_empty(&ctx->timeout_list));
+}
+
 static void io_commit_cqring(struct io_ring_ctx *ctx)
 {
 	struct io_kiocb *req;
 
-	while ((req = io_get_timeout_req(ctx)) != NULL)
-		io_kill_timeout(req);
+	if (!list_empty(&ctx->timeout_list))
+		__io_flush_timeouts(ctx);
 
 	__io_commit_cqring(ctx);
 
@@ -4540,20 +4554,8 @@ static enum hrtimer_restart io_timeout_fn(struct hrtimer *timer)
 	 * We could be racing with timeout deletion. If the list is empty,
 	 * then timeout lookup already found it and will be handling it.
 	 */
-	if (!list_empty(&req->list)) {
-		struct io_kiocb *prev;
-
-		/*
-		 * Adjust the reqs sequence before the current one because it
-		 * will consume a slot in the cq_ring and the cq_tail
-		 * pointer will be increased, otherwise other timeout reqs may
-		 * return in advance without waiting for enough wait_nr.
-		 */
-		prev = req;
-		list_for_each_entry_continue_reverse(prev, &ctx->timeout_list, list)
-			prev->sequence++;
+	if (!list_empty(&req->list))
 		list_del_init(&req->list);
-	}
 
 	io_cqring_fill_event(req, -ETIME);
 	io_commit_cqring(ctx);
@@ -4633,18 +4635,19 @@ static int io_timeout_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe,
 {
 	struct io_timeout_data *data;
 	unsigned flags;
+	u32 off = READ_ONCE(sqe->off);
 
 	if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
 		return -EINVAL;
 	if (sqe->ioprio || sqe->buf_index || sqe->len != 1)
 		return -EINVAL;
-	if (sqe->off && is_timeout_link)
+	if (off && is_timeout_link)
 		return -EINVAL;
 	flags = READ_ONCE(sqe->timeout_flags);
 	if (flags & ~IORING_TIMEOUT_ABS)
 		return -EINVAL;
 
-	req->timeout.count = READ_ONCE(sqe->off);
+	req->timeout.off = off;
 
 	if (!req->io && io_alloc_async_ctx(req))
 		return -ENOMEM;
@@ -4668,68 +4671,37 @@ static int io_timeout_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe,
 static int io_timeout(struct io_kiocb *req)
 {
 	struct io_ring_ctx *ctx = req->ctx;
-	struct io_timeout_data *data;
+	struct io_timeout_data *data = &req->io->timeout;
 	struct list_head *entry;
-	unsigned span = 0;
-	u32 count = req->timeout.count;
-	u32 seq = req->sequence;
+	u32 tail, off = req->timeout.off;
 
-	data = &req->io->timeout;
+	spin_lock_irq(&ctx->completion_lock);
 
 	/*
 	 * sqe->off holds how many events that need to occur for this
 	 * timeout event to be satisfied. If it isn't set, then this is
 	 * a pure timeout request, sequence isn't used.
 	 */
-	if (!count) {
+	if (!off) {
 		req->flags |= REQ_F_TIMEOUT_NOSEQ;
-		spin_lock_irq(&ctx->completion_lock);
 		entry = ctx->timeout_list.prev;
 		goto add;
 	}
 
-	req->sequence = seq + count;
+	tail = ctx->cached_cq_tail;
+	req->timeout.target_cq = tail + off;
 
 	/*
 	 * Insertion sort, ensuring the first entry in the list is always
 	 * the one we need first.
 	 */
-	spin_lock_irq(&ctx->completion_lock);
 	list_for_each_prev(entry, &ctx->timeout_list) {
 		struct io_kiocb *nxt = list_entry(entry, struct io_kiocb, list);
-		unsigned nxt_seq;
-		long long tmp, tmp_nxt;
-		u32 nxt_offset = nxt->timeout.count;
-
-		if (nxt->flags & REQ_F_TIMEOUT_NOSEQ)
-			continue;
-
-		/*
-		 * Since seq + count can overflow, use type long
-		 * long to store it.
-		 */
-		tmp = (long long)seq + count;
-		nxt_seq = nxt->sequence - nxt_offset;
-		tmp_nxt = (long long)nxt_seq + nxt_offset;
+		u32 nxt_off = nxt->timeout.target_cq - tail;
 
-		/*
-		 * cached_sq_head may overflow, and it will never overflow twice
-		 * once there is some timeout req still be valid.
-		 */
-		if (seq < nxt_seq)
-			tmp += UINT_MAX;
-
-		if (tmp > tmp_nxt)
+		if (!(nxt->flags & REQ_F_TIMEOUT_NOSEQ) && (off >= nxt_off))
 			break;
-
-		/*
-		 * Sequence of reqs after the insert one and itself should
-		 * be adjusted because each timeout req consumes a slot.
-		 */
-		span++;
-		nxt->sequence++;
 	}
-	req->sequence -= span;
 add:
 	list_add(&req->list, entry);
 	data->timer.function = io_timeout_fn;
Signature · 1 line
2.24.0
#8 lore raw

PATCH 2/2 io_uring: don't trigger timeout with another t-out

In reply to Pavel Begunkov

When deciding whether to fire a timeout basing on number of completions, ignore CQEs emitted by other timeouts.

---
 fs/io_uring.c | 19 +++----------------
 1 file changed, 3 insertions(+), 16 deletions(-)
fs/io_uring.c +3 -16
index c34c5dcc90fc..d72f82c09ce3 100644
@@ -1098,33 +1098,20 @@ static void io_kill_timeouts(struct io_ring_ctx *ctx)
 	spin_unlock_irq(&ctx->completion_lock);
 }
 
-static inline bool io_check_in_range(u32 pos, u32 start, u32 end)
-{
-	/* if @end < @start, check for [end, MAX_UINT] + [MAX_UINT, start] */
-	return (pos - start) <= (end - start);
-}
-
 static void __io_flush_timeouts(struct io_ring_ctx *ctx)
 {
-	u32 end, start;
-
-	start = end = ctx->cached_cq_tail;
 	do {
 		struct io_kiocb *req = list_first_entry(&ctx->timeout_list,
 							struct io_kiocb, list);
 
 		if (req->flags & REQ_F_TIMEOUT_NOSEQ)
 			break;
-		/*
-		 * multiple timeouts may have the same target,
-		 * check that @req is in [first_tail, cur_tail]
-		 */
-		if (!io_check_in_range(req->timeout.target_cq, start, end))
+		if (req->timeout.target_cq != ctx->cached_cq_tail
+					- atomic_read(&ctx->cq_timeouts))
 			break;
 
 		list_del_init(&req->list);
 		io_kill_timeout(req);
-		end = ctx->cached_cq_tail;
 	} while (!list_empty(&ctx->timeout_list));
 }
 
@@ -4688,7 +4675,7 @@ static int io_timeout(struct io_kiocb *req)
 		goto add;
 	}
 
-	tail = ctx->cached_cq_tail;
+	tail = ctx->cached_cq_tail - atomic_read(&ctx->cq_timeouts);
 	req->timeout.target_cq = tail + off;
 
 	/*
Signature · 1 line
2.24.0
maintainer
#2 lore raw

Re: [PATCH 1/2] io_uring: trigger timeout after any sqe->off CQEs

In reply to Pavel Begunkov
Quoted 25 lines · Pavel Begunkov
fs/io_uring.c +24 -0
+static void __io_flush_timeouts(struct io_ring_ctx *ctx)
+{
+	u32 end, start;
+
+	start = end = ctx->cached_cq_tail;
+	do {
+		struct io_kiocb *req = list_first_entry(&ctx->timeout_list,
+							struct io_kiocb, list);
+
+		if (req->flags & REQ_F_TIMEOUT_NOSEQ)
+			break;
+		/*
+		 * multiple timeouts may have the same target,
+		 * check that @req is in [first_tail, cur_tail]
+		 */
+		if (!io_check_in_range(req->timeout.target_cq, start, end))
+			break;
+
+		list_del_init(&req->list);
+		io_kill_timeout(req);
+		end = ctx->cached_cq_tail;
+	} while (!list_empty(&ctx->timeout_list));
+}
+

fs/io_uring.c static void io_kill_timeouts(struct io_ring_ctx *ctx) · line 1138

Quoted 10 lines
fs/io_uring.c +2 -2
 static void io_commit_cqring(struct io_ring_ctx *ctx)
 {
 	struct io_kiocb *req;
 
-	while ((req = io_get_timeout_req(ctx)) != NULL)
-		io_kill_timeout(req);
+	if (!list_empty(&ctx->timeout_list))
+		__io_flush_timeouts(ctx);
 
 	__io_commit_cqring(ctx);

Any chance we can do this without having to iterate timeouts on the completion path?

Signature · 1 line
Jens Axboe
#3 lore raw

Re: [PATCH 1/2] io_uring: trigger timeout after any sqe->off CQEs

Quoted 34 lines · Jens Axboe
Quoted 33 lines · Pavel Begunkov
fs/io_uring.c +26 -2
+static void __io_flush_timeouts(struct io_ring_ctx *ctx)
+{
+	u32 end, start;
+
+	start = end = ctx->cached_cq_tail;
+	do {
+		struct io_kiocb *req = list_first_entry(&ctx->timeout_list,
+							struct io_kiocb, list);
+
+		if (req->flags & REQ_F_TIMEOUT_NOSEQ)
+			break;
+		/*
+		 * multiple timeouts may have the same target,
+		 * check that @req is in [first_tail, cur_tail]
+		 */
+		if (!io_check_in_range(req->timeout.target_cq, start, end))
+			break;
+
+		list_del_init(&req->list);
+		io_kill_timeout(req);
+		end = ctx->cached_cq_tail;
+	} while (!list_empty(&ctx->timeout_list));
+}
+
 static void io_commit_cqring(struct io_ring_ctx *ctx)
 {
 	struct io_kiocb *req;
 
-	while ((req = io_get_timeout_req(ctx)) != NULL)
-		io_kill_timeout(req);
+	if (!list_empty(&ctx->timeout_list))
+		__io_flush_timeouts(ctx);
Quoted 6 lines
Quoted 2 lines

__io_commit_cqring(ctx);

Any chance we can do this without having to iterate timeouts on the completion path?

If you mean the one in __io_flush_timeouts(), then no, unless we forbid timeouts with identical target sequences + some extra constraints. The loop there is not new, it iterates only over timeouts, that need to be completed, and removes them. That's amortised O(1).

On the other hand, there was a loop in io_timeout_fn() doing in total O(n^2), and it was killed by this patch.

Signature · 1 line
Pavel Begunkov
#4 lore raw

Re: [PATCH 1/2] io_uring: trigger timeout after any sqe->off CQEs

Quoted 42 lines · Pavel Begunkov
Quoted 34 lines · Jens Axboe
Quoted 33 lines · Pavel Begunkov
fs/io_uring.c +26 -2
+static void __io_flush_timeouts(struct io_ring_ctx *ctx)
+{
+	u32 end, start;
+
+	start = end = ctx->cached_cq_tail;
+	do {
+		struct io_kiocb *req = list_first_entry(&ctx->timeout_list,
+							struct io_kiocb, list);
+
+		if (req->flags & REQ_F_TIMEOUT_NOSEQ)
+			break;
+		/*
+		 * multiple timeouts may have the same target,
+		 * check that @req is in [first_tail, cur_tail]
+		 */
+		if (!io_check_in_range(req->timeout.target_cq, start, end))
+			break;
+
+		list_del_init(&req->list);
+		io_kill_timeout(req);
+		end = ctx->cached_cq_tail;
+	} while (!list_empty(&ctx->timeout_list));
+}
+
 static void io_commit_cqring(struct io_ring_ctx *ctx)
 {
 	struct io_kiocb *req;
 
-	while ((req = io_get_timeout_req(ctx)) != NULL)
-		io_kill_timeout(req);
+	if (!list_empty(&ctx->timeout_list))
+		__io_flush_timeouts(ctx);
Quoted 6 lines
Quoted 2 lines

__io_commit_cqring(ctx);

Any chance we can do this without having to iterate timeouts on the completion path?

Quoted 4 lines

If you mean the one in __io_flush_timeouts(), then no, unless we forbid timeouts with identical target sequences + some extra constraints. The loop there is not new, it iterates only over timeouts, that need to be completed, and removes them. That's amortised O(1).

We can think about adding unlock/lock, if that's what you are thinking about.

Quoted 2 lines

On the other hand, there was a loop in io_timeout_fn() doing in total O(n^2), and it was killed by this patch.

Signature · 1 line
Pavel Begunkov
#5 lore raw

Re: [PATCH 1/2] io_uring: trigger timeout after any sqe->off CQEs

Quoted 43 lines · Pavel Begunkov
Quoted 42 lines · Pavel Begunkov
Quoted 40 lines · Jens Axboe
Quoted 25 lines · Pavel Begunkov
fs/io_uring.c +24 -0
+static void __io_flush_timeouts(struct io_ring_ctx *ctx)
+{
+	u32 end, start;
+
+	start = end = ctx->cached_cq_tail;
+	do {
+		struct io_kiocb *req = list_first_entry(&ctx->timeout_list,
+							struct io_kiocb, list);
+
+		if (req->flags & REQ_F_TIMEOUT_NOSEQ)
+			break;
+		/*
+		 * multiple timeouts may have the same target,
+		 * check that @req is in [first_tail, cur_tail]
+		 */
+		if (!io_check_in_range(req->timeout.target_cq, start, end))
+			break;
+
+		list_del_init(&req->list);
+		io_kill_timeout(req);
+		end = ctx->cached_cq_tail;
+	} while (!list_empty(&ctx->timeout_list));
+}
+

fs/io_uring.c static void io_kill_timeouts(struct io_ring_ctx *ctx) · line 1138

Quoted 10 lines
fs/io_uring.c +2 -2
 static void io_commit_cqring(struct io_ring_ctx *ctx)
 {
 	struct io_kiocb *req;
 
-	while ((req = io_get_timeout_req(ctx)) != NULL)
-		io_kill_timeout(req);
+	if (!list_empty(&ctx->timeout_list))
+		__io_flush_timeouts(ctx);
 
 	__io_commit_cqring(ctx);

Any chance we can do this without having to iterate timeouts on the completion path?

Quoted 11 lines
Quoted 4 lines

If you mean the one in __io_flush_timeouts(), then no, unless we forbid timeouts with identical target sequences + some extra constraints. The loop there is not new, it iterates only over timeouts, that need to be completed, and removes them. That's amortised O(1).

We can think about adding unlock/lock, if that's what you are thinking about.

Quoted 2 lines

On the other hand, there was a loop in io_timeout_fn() doing in total O(n^2), and it was killed by this patch.

Any thoughts on this?

I'll return fixing the last timeout bug I saw, but I'd prefer to know on top of what to do that.

Signature · 1 line
Pavel Begunkov
maintainer
#6 lore raw

Re: [PATCH 1/2] io_uring: trigger timeout after any sqe->off CQEs

Quoted 52 lines · Pavel Begunkov
Quoted 51 lines · Pavel Begunkov
Quoted 42 lines · Pavel Begunkov
Quoted 34 lines · Jens Axboe
Quoted 33 lines · Pavel Begunkov
fs/io_uring.c +26 -2
+static void __io_flush_timeouts(struct io_ring_ctx *ctx)
+{
+	u32 end, start;
+
+	start = end = ctx->cached_cq_tail;
+	do {
+		struct io_kiocb *req = list_first_entry(&ctx->timeout_list,
+							struct io_kiocb, list);
+
+		if (req->flags & REQ_F_TIMEOUT_NOSEQ)
+			break;
+		/*
+		 * multiple timeouts may have the same target,
+		 * check that @req is in [first_tail, cur_tail]
+		 */
+		if (!io_check_in_range(req->timeout.target_cq, start, end))
+			break;
+
+		list_del_init(&req->list);
+		io_kill_timeout(req);
+		end = ctx->cached_cq_tail;
+	} while (!list_empty(&ctx->timeout_list));
+}
+
 static void io_commit_cqring(struct io_ring_ctx *ctx)
 {
 	struct io_kiocb *req;
 
-	while ((req = io_get_timeout_req(ctx)) != NULL)
-		io_kill_timeout(req);
+	if (!list_empty(&ctx->timeout_list))
+		__io_flush_timeouts(ctx);
Quoted 6 lines
Quoted 2 lines

__io_commit_cqring(ctx);

Any chance we can do this without having to iterate timeouts on the completion path?

Quoted 4 lines

If you mean the one in __io_flush_timeouts(), then no, unless we forbid timeouts with identical target sequences + some extra constraints. The loop there is not new, it iterates only over timeouts, that need to be completed, and removes them. That's amortised O(1).

We can think about adding unlock/lock, if that's what you are thinking about.

Quoted 8 lines
Quoted 3 lines
Quoted 2 lines

On the other hand, there was a loop in io_timeout_fn() doing in total O(n^2), and it was killed by this patch.

Any thoughts on this?

I'll return fixing the last timeout bug I saw, but I'd prefer to know on top of what to do that.

I think it's fine, but also likely something that we should defer to 5.8. So if there are minor fixes to be done for 5.7, it should be arranged as such.

Signature · 1 line
Jens Axboe
#7 lore raw

Re: [PATCH 1/2] io_uring: trigger timeout after any sqe->off CQEs

Quoted 57 lines · Jens Axboe
Quoted 56 lines · Pavel Begunkov
Quoted 54 lines · Pavel Begunkov
Quoted 42 lines · Pavel Begunkov
Quoted 34 lines · Jens Axboe
Quoted 33 lines · Pavel Begunkov
fs/io_uring.c +26 -2
+static void __io_flush_timeouts(struct io_ring_ctx *ctx)
+{
+	u32 end, start;
+
+	start = end = ctx->cached_cq_tail;
+	do {
+		struct io_kiocb *req = list_first_entry(&ctx->timeout_list,
+							struct io_kiocb, list);
+
+		if (req->flags & REQ_F_TIMEOUT_NOSEQ)
+			break;
+		/*
+		 * multiple timeouts may have the same target,
+		 * check that @req is in [first_tail, cur_tail]
+		 */
+		if (!io_check_in_range(req->timeout.target_cq, start, end))
+			break;
+
+		list_del_init(&req->list);
+		io_kill_timeout(req);
+		end = ctx->cached_cq_tail;
+	} while (!list_empty(&ctx->timeout_list));
+}
+
 static void io_commit_cqring(struct io_ring_ctx *ctx)
 {
 	struct io_kiocb *req;
 
-	while ((req = io_get_timeout_req(ctx)) != NULL)
-		io_kill_timeout(req);
+	if (!list_empty(&ctx->timeout_list))
+		__io_flush_timeouts(ctx);
Quoted 6 lines
Quoted 2 lines

__io_commit_cqring(ctx);

Any chance we can do this without having to iterate timeouts on the completion path?

Quoted 4 lines

If you mean the one in __io_flush_timeouts(), then no, unless we forbid timeouts with identical target sequences + some extra constraints. The loop there is not new, it iterates only over timeouts, that need to be completed, and removes them. That's amortised O(1).

We can think about adding unlock/lock, if that's what you are thinking about.

Quoted 2 lines

On the other hand, there was a loop in io_timeout_fn() doing in total O(n^2), and it was killed by this patch.

Quoted 8 lines
Quoted 4 lines

Any thoughts on this?

I'll return fixing the last timeout bug I saw, but I'd prefer to know on top of what to do that.

I think it's fine, but also likely something that we should defer to 5.8. So if there are minor fixes to be done for 5.7, it should be arranged as such.

Right, totally agree

Signature · 1 line
Pavel Begunkov