[RFC,2/2] io_uring: Optimise cq waiting with wait_threshold

Commit Message

Pavel Begunkov Sept. 13, 2019, 10:28 p.m. UTC

From: Pavel Begunkov <[e-mail]>

While waiting for completion events in io_cqring_wait(), the process
will be waken up inside wait_threshold_interruptible() on any request
completion, check num of events in completion queue and potentially go
to sleep again.

Apparently, there could be a lot of such spurious wakeups with lots of
overhead. It especially manifests itself, when min_events is large, and
completions are arriving one by one or in small batches (that usually
is true).

E.g. if device completes requests one by one and io_uring_enter is
waiting for 100 events, then there will be ~99 spurious wakeups.

Use new wait_threshold_*() instead, which won't wake it up until
necessary number of events is collected.

Performance test:
The first thread generates requests (QD=512) one by one, so they will
be completed in the similar pattern. The second thread waiting for
128 events to complete.

Tested with null_blk with 5us delay
and 3.8GHz Intel CPU.

throughput before: 270 KIOPS
throughput after:  370 KIOPS
So, ~40% throughput boost on this exaggerate test.

Signed-off-by: Pavel Begunkov <[e-mail]>
---
 fs/io_uring.c | 21 ++++++++++++---------
 1 file changed, 12 insertions(+), 9 deletions(-)

Patch

diff --git a/fs/io_uring.c b/fs/io_uring.c
index 37395208a729..17d2d30b763a 100644
--- a/fs/io_uring.c
+++ b/fs/io_uring.c
@@ -70,6 +70,7 @@ 
 #include <linux/nospec.h>
 #include <linux/sizes.h>
 #include <linux/hugetlb.h>
+#include <linux/wait_threshold.h>
 
 #include <uapi/linux/io_uring.h>
 
@@ -403,6 +404,13 @@  static struct io_ring_ctx *io_ring_ctx_alloc(struct io_uring_params *p)
 	return ctx;
 }
 
+static unsigned int io_cqring_events(struct io_rings *rings)
+{
+	/* See comment at the top of this file */
+	smp_rmb();
+	return READ_ONCE(rings->cq.tail) - READ_ONCE(rings->cq.head);
+}
+
 static inline bool io_sequence_defer(struct io_ring_ctx *ctx,
 				     struct io_kiocb *req)
 {
@@ -521,7 +529,7 @@  static void io_cqring_fill_event(struct io_ring_ctx *ctx, u64 ki_user_data,
 static void io_cqring_ev_posted(struct io_ring_ctx *ctx)
 {
 	if (waitqueue_active(&ctx->wait))
-		wake_up(&ctx->wait);
+		wake_up_threshold(&ctx->wait, io_cqring_events(ctx->rings));
 	if (waitqueue_active(&ctx->sqo_wait))
 		wake_up(&ctx->sqo_wait);
 	if (ctx->cq_ev_fd)
@@ -546,7 +554,7 @@  static void io_ring_drop_ctx_refs(struct io_ring_ctx *ctx, unsigned refs)
 	percpu_ref_put_many(&ctx->refs, refs);
 
 	if (waitqueue_active(&ctx->wait))
-		wake_up(&ctx->wait);
+		wake_up_threshold(&ctx->wait, io_cqring_events(ctx->rings));
 }
 
 static struct io_kiocb *io_get_req(struct io_ring_ctx *ctx,
@@ -681,12 +689,6 @@  static void io_put_req(struct io_kiocb *req)
 		io_free_req(req);
 }
 
-static unsigned io_cqring_events(struct io_rings *rings)
-{
-	/* See comment at the top of this file */
-	smp_rmb();
-	return READ_ONCE(rings->cq.tail) - READ_ONCE(rings->cq.head);
-}
 
 /*
  * Find and free completed poll iocbs
@@ -2591,7 +2593,8 @@  static int io_cqring_wait(struct io_ring_ctx *ctx, int min_events,
 			return ret;
 	}
 
-	ret = wait_event_interruptible(ctx->wait, io_cqring_events(rings) >= min_events);
+	ret = wait_threshold_interruptible(ctx->wait, min_events,
+					   io_cqring_events(rings));
 	restore_saved_sigmask_unless(ret == -ERESTARTSYS);
 	if (ret == -ERESTARTSYS)
 		ret = -EINTR;