block, bfq: merge bursts of newly-created queues

Many throughput-sensitive workloads are made of several parallel I/O
flows, with all flows generated by the same application, or more
generically by the same task (e.g., system boot). The most
counterproductive action with these workloads is plugging I/O dispatch
when one of the bfq_queues associated with these flows remains
temporarily empty.

To avoid this plugging, BFQ has been using a burst-handling mechanism
for years now. This mechanism has proven effective for throughput, and
not detrimental for service guarantees. This commit pushes this
mechanism a little bit further, basing on the following two facts.

First, all the I/O flows of a the same application or task contribute
to the execution/completion of that common application or task. So the
performance figures that matter are total throughput of the flows and
task-wide I/O latency.  In particular, these flows do not need to be
protected from each other, in terms of individual bandwidth or
latency.

Second, the above fact holds regardless of the number of flows.

Putting these two facts together, this commits merges stably the
bfq_queues associated with these I/O flows, i.e., with the processes
that generate these IO/ flows, regardless of how many the involved
processes are.

To decide whether a set of bfq_queues is actually associated with the
I/O flows of a common application or task, and to merge these queues
stably, this commit operates as follows: given a bfq_queue, say Q2,
currently being created, and the last bfq_queue, say Q1, created
before Q2, Q2 is merged stably with Q1 if
- very little time has elapsed since when Q1 was created
- Q2 has the same ioprio as Q1
- Q2 belongs to the same group as Q1

Merging bfq_queues also reduces scheduling overhead. A fio test with
ten random readers on /dev/nullb shows a throughput boost of 40%, with
a quadcore. Since BFQ's execution time amounts to ~50% of the total
per-request processing time, the above throughput boost implies that
BFQ's overhead is reduced by more than 50%.

Tested-by: Jan Kara <jack@suse.cz>
Signed-off-by: Paolo Valente <paolo.valente@linaro.org>
Tested-by: Oleksandr Natalenko <oleksandr@natalenko.name>
Link: https://lore.kernel.org/r/20210304174627.161-7-paolo.valente@linaro.org
Signed-off-by: Jens Axboe <axboe@kernel.dk>

1 files changed, 249 insertions, 10 deletions

diff --git a/block/bfq-iosched.c b/block/bfq-iosched.c
index a3f59d3065b3..9b7678ad5830 100644
--- a/block/bfq-iosched.c
+++ b/block/bfq-iosched.c
@@ -1075,7 +1075,7 @@ bfq_bfqq_resume_state(struct bfq_queue *bfqq, struct bfq_data *bfqd,
 static int bfqq_process_refs(struct bfq_queue *bfqq)
 {
 	return bfqq->ref - bfqq->allocated - bfqq->entity.on_st_or_in_serv -
-		(bfqq->weight_counter != NULL);
+		(bfqq->weight_counter != NULL) - bfqq->stable_ref;
 }
 
 /* Empty burst list and add just bfqq (see comments on bfq_handle_burst) */
@@ -2628,6 +2628,11 @@ static bool bfq_may_be_close_cooperator(struct bfq_queue *bfqq,
 	return true;
 }
 
+static bool idling_boosts_thr_without_issues(struct bfq_data *bfqd,
+					     struct bfq_queue *bfqq);
+
+static void bfq_put_stable_ref(struct bfq_queue *bfqq);
+
 /*
  * Attempt to schedule a merge of bfqq with the currently in-service
  * queue or with a close queue among the scheduled queues.  Return
@@ -2650,11 +2655,50 @@ static bool bfq_may_be_close_cooperator(struct bfq_queue *bfqq,
  */
 static struct bfq_queue *
 bfq_setup_cooperator(struct bfq_data *bfqd, struct bfq_queue *bfqq,
-		     void *io_struct, bool request)
+		     void *io_struct, bool request, struct bfq_io_cq *bic)
 {
 	struct bfq_queue *in_service_bfqq, *new_bfqq;
 
 	/*
+	 * Check delayed stable merge for rotational or non-queueing
+	 * devs. For this branch to be executed, bfqq must not be
+	 * currently merged with some other queue (i.e., bfqq->bic
+	 * must be non null). If we considered also merged queues,
+	 * then we should also check whether bfqq has already been
+	 * merged with bic->stable_merge_bfqq. But this would be
+	 * costly and complicated.
+	 */
+	if (unlikely(!bfqd->nonrot_with_queueing)) {
+		if (bic->stable_merge_bfqq &&
+		    !bfq_bfqq_just_created(bfqq) &&
+		    time_is_after_jiffies(bfqq->split_time +
+					  msecs_to_jiffies(200))) {
+			struct bfq_queue *stable_merge_bfqq =
+				bic->stable_merge_bfqq;
+			int proc_ref = min(bfqq_process_refs(bfqq),
+					   bfqq_process_refs(stable_merge_bfqq));
+
+			/* deschedule stable merge, because done or aborted here */
+			bfq_put_stable_ref(stable_merge_bfqq);
+
+			bic->stable_merge_bfqq = NULL;
+
+			if (!idling_boosts_thr_without_issues(bfqd, bfqq) &&
+			    proc_ref > 0) {
+				/* next function will take at least one ref */
+				struct bfq_queue *new_bfqq =
+					bfq_setup_merge(bfqq, stable_merge_bfqq);
+
+				bic->stably_merged = true;
+				if (new_bfqq && new_bfqq->bic)
+					new_bfqq->bic->stably_merged = true;
+				return new_bfqq;
+			} else
+				return NULL;
+		}
+	}
+
+	/*
 	 * Do not perform queue merging if the device is non
 	 * rotational and performs internal queueing. In fact, such a
 	 * device reaches a high speed through internal parallelism
@@ -2795,6 +2839,17 @@ static void bfq_bfqq_save_state(struct bfq_queue *bfqq)
 	}
 }
 
+
+static void
+bfq_reassign_last_bfqq(struct bfq_queue *cur_bfqq, struct bfq_queue *new_bfqq)
+{
+	if (cur_bfqq->entity.parent &&
+	    cur_bfqq->entity.parent->last_bfqq_created == cur_bfqq)
+		cur_bfqq->entity.parent->last_bfqq_created = new_bfqq;
+	else if (cur_bfqq->bfqd && cur_bfqq->bfqd->last_bfqq_created == cur_bfqq)
+		cur_bfqq->bfqd->last_bfqq_created = new_bfqq;
+}
+
 void bfq_release_process_ref(struct bfq_data *bfqd, struct bfq_queue *bfqq)
 {
 	/*
@@ -2812,6 +2867,8 @@ void bfq_release_process_ref(struct bfq_data *bfqd, struct bfq_queue *bfqq)
 	    bfqq != bfqd->in_service_queue)
 		bfq_del_bfqq_busy(bfqd, bfqq, false);
 
+	bfq_reassign_last_bfqq(bfqq, NULL);
+
 	bfq_put_queue(bfqq);
 }
 
@@ -2908,6 +2965,9 @@ bfq_merge_bfqqs(struct bfq_data *bfqd, struct bfq_io_cq *bic,
 	 */
 	new_bfqq->pid = -1;
 	bfqq->bic = NULL;
+
+	bfq_reassign_last_bfqq(bfqq, new_bfqq);
+
 	bfq_release_process_ref(bfqd, bfqq);
 }
 
@@ -2935,7 +2995,7 @@ static bool bfq_allow_bio_merge(struct request_queue *q, struct request *rq,
 	 * We take advantage of this function to perform an early merge
 	 * of the queues of possible cooperating processes.
 	 */
-	new_bfqq = bfq_setup_cooperator(bfqd, bfqq, bio, false);
+	new_bfqq = bfq_setup_cooperator(bfqd, bfqq, bio, false, bfqd->bio_bic);
 	if (new_bfqq) {
 		/*
 		 * bic still points to bfqq, then it has not yet been
@@ -5034,6 +5094,12 @@ void bfq_put_queue(struct bfq_queue *bfqq)
 	bfqg_and_blkg_put(bfqg);
 }
 
+static void bfq_put_stable_ref(struct bfq_queue *bfqq)
+{
+	bfqq->stable_ref--;
+	bfq_put_queue(bfqq);
+}
+
 static void bfq_put_cooperator(struct bfq_queue *bfqq)
 {
 	struct bfq_queue *__bfqq, *next;
@@ -5090,6 +5156,24 @@ static void bfq_exit_icq(struct io_cq *icq)
 {
 	struct bfq_io_cq *bic = icq_to_bic(icq);
 
+	if (bic->stable_merge_bfqq) {
+		struct bfq_data *bfqd = bic->stable_merge_bfqq->bfqd;
+
+		/*
+		 * bfqd is NULL if scheduler already exited, and in
+		 * that case this is the last time bfqq is accessed.
+		 */
+		if (bfqd) {
+			unsigned long flags;
+
+			spin_lock_irqsave(&bfqd->lock, flags);
+			bfq_put_stable_ref(bic->stable_merge_bfqq);
+			spin_unlock_irqrestore(&bfqd->lock, flags);
+		} else {
+			bfq_put_stable_ref(bic->stable_merge_bfqq);
+		}
+	}
+
 	bfq_exit_icq_bfqq(bic, true);
 	bfq_exit_icq_bfqq(bic, false);
 }
@@ -5150,7 +5234,8 @@ bfq_set_next_ioprio_data(struct bfq_queue *bfqq, struct bfq_io_cq *bic)
 
 static struct bfq_queue *bfq_get_queue(struct bfq_data *bfqd,
 				       struct bio *bio, bool is_sync,
-				       struct bfq_io_cq *bic);
+				       struct bfq_io_cq *bic,
+				       bool respawn);
 
 static void bfq_check_ioprio_change(struct bfq_io_cq *bic, struct bio *bio)
 {
@@ -5170,7 +5255,7 @@ static void bfq_check_ioprio_change(struct bfq_io_cq *bic, struct bio *bio)
 	bfqq = bic_to_bfqq(bic, false);
 	if (bfqq) {
 		bfq_release_process_ref(bfqd, bfqq);
-		bfqq = bfq_get_queue(bfqd, bio, BLK_RW_ASYNC, bic);
+		bfqq = bfq_get_queue(bfqd, bio, BLK_RW_ASYNC, bic, true);
 		bic_set_bfqq(bic, bfqq, false);
 	}
 
@@ -5213,6 +5298,8 @@ static void bfq_init_bfqq(struct bfq_data *bfqd, struct bfq_queue *bfqq,
 	/* set end request to minus infinity from now */
 	bfqq->ttime.last_end_request = now_ns + 1;
 
+	bfqq->creation_time = jiffies;
+
 	bfqq->io_start_time = now_ns;
 
 	bfq_mark_bfqq_IO_bound(bfqq);
@@ -5262,9 +5349,156 @@ static struct bfq_queue **bfq_async_queue_prio(struct bfq_data *bfqd,
 	}
 }
 
+static struct bfq_queue *
+bfq_do_early_stable_merge(struct bfq_data *bfqd, struct bfq_queue *bfqq,
+			  struct bfq_io_cq *bic,
+			  struct bfq_queue *last_bfqq_created)
+{
+	struct bfq_queue *new_bfqq =
+		bfq_setup_merge(bfqq, last_bfqq_created);
+
+	if (!new_bfqq)
+		return bfqq;
+
+	if (new_bfqq->bic)
+		new_bfqq->bic->stably_merged = true;
+	bic->stably_merged = true;
+
+	/*
+	 * Reusing merge functions. This implies that
+	 * bfqq->bic must be set too, for
+	 * bfq_merge_bfqqs to correctly save bfqq's
+	 * state before killing it.
+	 */
+	bfqq->bic = bic;
+	bfq_merge_bfqqs(bfqd, bic, bfqq, new_bfqq);
+
+	return new_bfqq;
+}
+
+/*
+ * Many throughput-sensitive workloads are made of several parallel
+ * I/O flows, with all flows generated by the same application, or
+ * more generically by the same task (e.g., system boot). The most
+ * counterproductive action with these workloads is plugging I/O
+ * dispatch when one of the bfq_queues associated with these flows
+ * remains temporarily empty.
+ *
+ * To avoid this plugging, BFQ has been using a burst-handling
+ * mechanism for years now. This mechanism has proven effective for
+ * throughput, and not detrimental for service guarantees. The
+ * following function pushes this mechanism a little bit further,
+ * basing on the following two facts.
+ *
+ * First, all the I/O flows of a the same application or task
+ * contribute to the execution/completion of that common application
+ * or task. So the performance figures that matter are total
+ * throughput of the flows and task-wide I/O latency.  In particular,
+ * these flows do not need to be protected from each other, in terms
+ * of individual bandwidth or latency.
+ *
+ * Second, the above fact holds regardless of the number of flows.
+ *
+ * Putting these two facts together, this commits merges stably the
+ * bfq_queues associated with these I/O flows, i.e., with the
+ * processes that generate these IO/ flows, regardless of how many the
+ * involved processes are.
+ *
+ * To decide whether a set of bfq_queues is actually associated with
+ * the I/O flows of a common application or task, and to merge these
+ * queues stably, this function operates as follows: given a bfq_queue,
+ * say Q2, currently being created, and the last bfq_queue, say Q1,
+ * created before Q2, Q2 is merged stably with Q1 if
+ * - very little time has elapsed since when Q1 was created
+ * - Q2 has the same ioprio as Q1
+ * - Q2 belongs to the same group as Q1
+ *
+ * Merging bfq_queues also reduces scheduling overhead. A fio test
+ * with ten random readers on /dev/nullb shows a throughput boost of
+ * 40%, with a quadcore. Since BFQ's execution time amounts to ~50% of
+ * the total per-request processing time, the above throughput boost
+ * implies that BFQ's overhead is reduced by more than 50%.
+ *
+ * This new mechanism most certainly obsoletes the current
+ * burst-handling heuristics. We keep those heuristics for the moment.
+ */
+static struct bfq_queue *bfq_do_or_sched_stable_merge(struct bfq_data *bfqd,
+						      struct bfq_queue *bfqq,
+						      struct bfq_io_cq *bic)
+{
+	struct bfq_queue **source_bfqq = bfqq->entity.parent ?
+		&bfqq->entity.parent->last_bfqq_created :
+		&bfqd->last_bfqq_created;
+
+	struct bfq_queue *last_bfqq_created = *source_bfqq;
+
+	/*
+	 * If last_bfqq_created has not been set yet, then init it. If
+	 * it has been set already, but too long ago, then move it
+	 * forward to bfqq. Finally, move also if bfqq belongs to a
+	 * different group than last_bfqq_created, or if bfqq has a
+	 * different ioprio or ioprio_class. If none of these
+	 * conditions holds true, then try an early stable merge or
+	 * schedule a delayed stable merge.
+	 *
+	 * A delayed merge is scheduled (instead of performing an
+	 * early merge), in case bfqq might soon prove to be more
+	 * throughput-beneficial if not merged. Currently this is
+	 * possible only if bfqd is rotational with no queueing. For
+	 * such a drive, not merging bfqq is better for throughput if
+	 * bfqq happens to contain sequential I/O. So, we wait a
+	 * little bit for enough I/O to flow through bfqq. After that,
+	 * if such an I/O is sequential, then the merge is
+	 * canceled. Otherwise the merge is finally performed.
+	 */
+	if (!last_bfqq_created ||
+	    time_before(last_bfqq_created->creation_time +
+			bfqd->bfq_burst_interval,
+			bfqq->creation_time) ||
+		bfqq->entity.parent != last_bfqq_created->entity.parent ||
+		bfqq->ioprio != last_bfqq_created->ioprio ||
+		bfqq->ioprio_class != last_bfqq_created->ioprio_class)
+		*source_bfqq = bfqq;
+	else if (time_after_eq(last_bfqq_created->creation_time +
+				 bfqd->bfq_burst_interval,
+				 bfqq->creation_time)) {
+		if (likely(bfqd->nonrot_with_queueing))
+			/*
+			 * With this type of drive, leaving
+			 * bfqq alone may provide no
+			 * throughput benefits compared with
+			 * merging bfqq. So merge bfqq now.
+			 */
+			bfqq = bfq_do_early_stable_merge(bfqd, bfqq,
+							 bic,
+							 last_bfqq_created);
+		else { /* schedule tentative stable merge */
+			/*
+			 * get reference on last_bfqq_created,
+			 * to prevent it from being freed,
+			 * until we decide whether to merge
+			 */
+			last_bfqq_created->ref++;
+			/*
+			 * need to keep track of stable refs, to
+			 * compute process refs correctly
+			 */
+			last_bfqq_created->stable_ref++;
+			/*
+			 * Record the bfqq to merge to.
+			 */
+			bic->stable_merge_bfqq = last_bfqq_created;
+		}
+	}
+
+	return bfqq;
+}
+
+
 static struct bfq_queue *bfq_get_queue(struct bfq_data *bfqd,
 				       struct bio *bio, bool is_sync,
-				       struct bfq_io_cq *bic)
+				       struct bfq_io_cq *bic,
+				       bool respawn)
 {
 	const int ioprio = IOPRIO_PRIO_DATA(bic->ioprio);
 	const int ioprio_class = IOPRIO_PRIO_CLASS(bic->ioprio);
@@ -5322,7 +5556,10 @@ static struct bfq_queue *bfq_get_queue(struct bfq_data *bfqd,
 
 out:
 	bfqq->ref++; /* get a process reference to this queue */
-	bfq_log_bfqq(bfqd, bfqq, "get_queue, at end: %p, %d", bfqq, bfqq->ref);
+
+	if (bfqq != &bfqd->oom_bfqq && is_sync && !respawn)
+		bfqq = bfq_do_or_sched_stable_merge(bfqd, bfqq, bic);
+
 	rcu_read_unlock();
 	return bfqq;
 }
@@ -5572,7 +5809,8 @@ static void bfq_rq_enqueued(struct bfq_data *bfqd, struct bfq_queue *bfqq,
 static bool __bfq_insert_request(struct bfq_data *bfqd, struct request *rq)
 {
 	struct bfq_queue *bfqq = RQ_BFQQ(rq),
-		*new_bfqq = bfq_setup_cooperator(bfqd, bfqq, rq, true);
+		*new_bfqq = bfq_setup_cooperator(bfqd, bfqq, rq, true,
+						 RQ_BIC(rq));
 	bool waiting, idle_timer_disabled = false;
 
 	if (new_bfqq) {
@@ -6227,7 +6465,7 @@ static struct bfq_queue *bfq_get_bfqq_handle_split(struct bfq_data *bfqd,
 
 	if (bfqq)
 		bfq_put_queue(bfqq);
-	bfqq = bfq_get_queue(bfqd, bio, is_sync, bic);
+	bfqq = bfq_get_queue(bfqd, bio, is_sync, bic, split);
 
 	bic_set_bfqq(bic, bfqq, is_sync);
 	if (split && is_sync) {
@@ -6348,7 +6586,8 @@ static struct bfq_queue *bfq_init_rq(struct request *rq)
 
 	if (likely(!new_queue)) {
 		/* If the queue was seeky for too long, break it apart. */
-		if (bfq_bfqq_coop(bfqq) && bfq_bfqq_split_coop(bfqq)) {
+		if (bfq_bfqq_coop(bfqq) && bfq_bfqq_split_coop(bfqq) &&
+			!bic->stably_merged) {
 			struct bfq_queue *old_bfqq = bfqq;
 
 			/* Update bic before losing reference to bfqq */