[PATCH v2 0/5] blk-mq: Kyber multiqueue I/O scheduler

[PATCH v2 0/5] blk-mq: Kyber multiqueue I/O scheduler

[Omar Sandoval]

From: Omar Sandoval <osandov@fb.com>

This is v2 of Kyber, an I/O scheduler for multiqueue devices combining
several techniques: the scalable bitmap library, the new blk-stats API,
and queue depth throttling similar to blk-wbt. v1 was here [1].

v2 adds a tunable target synchronous write latency. The heuristics are
more fleshed out and balance read and synchronous write latencies based
on the latency targets. The queueing and dispatch mechanism is the same
as before.

This series is based on block/for-next + the initialization fix series I
sent out yesterday [2]. Patches 1 and 2 implement a new sbitmap
operation. Patch 3 moves a scheduler callback to somewhere more useful.
Patch 4 exports a required function. Patch 5 implements the new
scheduler.

Thanks!

1: http://marc.info/?l=linux-block&m=148978871820916&w=2
2: http://marc.info/?l=linux-block&m=149125578724683&w=2

Omar Sandoval (5):
  sbitmap: add sbitmap_get_shallow() operation
  blk-mq: add shallow depth option for blk_mq_get_tag()
  blk-mq-sched: make completed_request() callback more useful
  blk-mq: export blk_mq_finish_request()
  blk-mq: introduce Kyber multiqueue I/O scheduler

 Documentation/block/kyber-iosched.txt |  14 +
 block/Kconfig.iosched                 |   9 +
 block/Makefile                        |   1 +
 block/blk-mq-sched.h                  |  11 +-
 block/blk-mq-tag.c                    |   5 +-
 block/blk-mq.c                        |   6 +-
 block/blk-mq.h                        |   1 +
 block/kyber-iosched.c                 | 706 ++++++++++++++++++++++++++++++++++
 include/linux/elevator.h              |   2 +-
 include/linux/sbitmap.h               |  55 +++
 lib/sbitmap.c                         |  75 +++-
 11 files changed, 867 insertions(+), 18 deletions(-)
 create mode 100644 Documentation/block/kyber-iosched.txt
 create mode 100644 block/kyber-iosched.c

-- 
2.12.2

[PATCH v2 1/5] sbitmap: add sbitmap_get_shallow() operation

[Omar Sandoval]

From: Omar Sandoval <osandov@fb.com>

This operation supports the use case of limiting the number of bits that
can be allocated for a given operation. Rather than setting aside some
bits at the end of the bitmap, we can set aside bits in each word of the
bitmap. This means we can keep the allocation hints spread out and
support sbitmap_resize() nicely at the cost of lower granularity for the
allowed depth.

Signed-off-by: Omar Sandoval <osandov@fb.com>
---
 include/linux/sbitmap.h | 55 ++++++++++++++++++++++++++++++++++++
 lib/sbitmap.c           | 75 ++++++++++++++++++++++++++++++++++++++++++++-----
 2 files changed, 123 insertions(+), 7 deletions(-)

diff --git a/include/linux/sbitmap.h b/include/linux/sbitmap.h
index d4e0a204c118..a1904aadbc45 100644
--- a/include/linux/sbitmap.h
+++ b/include/linux/sbitmap.h
@@ -176,6 +176,25 @@ void sbitmap_resize(struct sbitmap *sb, unsigned int depth);
 int sbitmap_get(struct sbitmap *sb, unsigned int alloc_hint, bool round_robin);
 
 /**
+ * sbitmap_get_shallow() - Try to allocate a free bit from a &struct sbitmap,
+ * limiting the depth used from each word.
+ * @sb: Bitmap to allocate from.
+ * @alloc_hint: Hint for where to start searching for a free bit.
+ * @shallow_depth: The maximum number of bits to allocate from a single word.
+ *
+ * This rather specific operation allows for having multiple users with
+ * different allocation limits. E.g., there can be a high-priority class that
+ * uses sbitmap_get() and a low-priority class that uses sbitmap_get_shallow()
+ * with a @shallow_depth of (1 << (@sb->shift - 1)). Then, the low-priority
+ * class can only allocate half of the total bits in the bitmap, preventing it
+ * from starving out the high-priority class.
+ *
+ * Return: Non-negative allocated bit number if successful, -1 otherwise.
+ */
+int sbitmap_get_shallow(struct sbitmap *sb, unsigned int alloc_hint,
+			unsigned long shallow_depth);
+
+/**
  * sbitmap_any_bit_set() - Check for a set bit in a &struct sbitmap.
  * @sb: Bitmap to check.
  *
@@ -326,6 +345,19 @@ void sbitmap_queue_resize(struct sbitmap_queue *sbq, unsigned int depth);
 int __sbitmap_queue_get(struct sbitmap_queue *sbq);
 
 /**
+ * __sbitmap_queue_get_shallow() - Try to allocate a free bit from a &struct
+ * sbitmap_queue, limiting the depth used from each word, with preemption
+ * already disabled.
+ * @sbq: Bitmap queue to allocate from.
+ * @shallow_depth: The maximum number of bits to allocate from a single word.
+ * See sbitmap_get_shallow().
+ *
+ * Return: Non-negative allocated bit number if successful, -1 otherwise.
+ */
+int __sbitmap_queue_get_shallow(struct sbitmap_queue *sbq,
+				unsigned int shallow_depth);
+
+/**
  * sbitmap_queue_get() - Try to allocate a free bit from a &struct
  * sbitmap_queue.
  * @sbq: Bitmap queue to allocate from.
@@ -346,6 +378,29 @@ static inline int sbitmap_queue_get(struct sbitmap_queue *sbq,
 }
 
 /**
+ * sbitmap_queue_get_shallow() - Try to allocate a free bit from a &struct
+ * sbitmap_queue, limiting the depth used from each word.
+ * @sbq: Bitmap queue to allocate from.
+ * @cpu: Output parameter; will contain the CPU we ran on (e.g., to be passed to
+ *       sbitmap_queue_clear()).
+ * @shallow_depth: The maximum number of bits to allocate from a single word.
+ * See sbitmap_get_shallow().
+ *
+ * Return: Non-negative allocated bit number if successful, -1 otherwise.
+ */
+static inline int sbitmap_queue_get_shallow(struct sbitmap_queue *sbq,
+					    unsigned int *cpu,
+					    unsigned int shallow_depth)
+{
+	int nr;
+
+	*cpu = get_cpu();
+	nr = __sbitmap_queue_get_shallow(sbq, shallow_depth);
+	put_cpu();
+	return nr;
+}
+
+/**
  * sbitmap_queue_clear() - Free an allocated bit and wake up waiters on a
  * &struct sbitmap_queue.
  * @sbq: Bitmap to free from.
diff --git a/lib/sbitmap.c b/lib/sbitmap.c
index 60e800e0b5a0..80aa8d5463fa 100644
--- a/lib/sbitmap.c
+++ b/lib/sbitmap.c
@@ -79,15 +79,15 @@ void sbitmap_resize(struct sbitmap *sb, unsigned int depth)
 }
 EXPORT_SYMBOL_GPL(sbitmap_resize);
 
-static int __sbitmap_get_word(struct sbitmap_word *word, unsigned int hint,
-			      bool wrap)
+static int __sbitmap_get_word(unsigned long *word, unsigned long depth,
+			      unsigned int hint, bool wrap)
 {
 	unsigned int orig_hint = hint;
 	int nr;
 
 	while (1) {
-		nr = find_next_zero_bit(&word->word, word->depth, hint);
-		if (unlikely(nr >= word->depth)) {
+		nr = find_next_zero_bit(word, depth, hint);
+		if (unlikely(nr >= depth)) {
 			/*
 			 * We started with an offset, and we didn't reset the
 			 * offset to 0 in a failure case, so start from 0 to
@@ -100,11 +100,11 @@ static int __sbitmap_get_word(struct sbitmap_word *word, unsigned int hint,
 			return -1;
 		}
 
-		if (!test_and_set_bit(nr, &word->word))
+		if (!test_and_set_bit(nr, word))
 			break;
 
 		hint = nr + 1;
-		if (hint >= word->depth - 1)
+		if (hint >= depth - 1)
 			hint = 0;
 	}
 
@@ -119,7 +119,8 @@ int sbitmap_get(struct sbitmap *sb, unsigned int alloc_hint, bool round_robin)
 	index = SB_NR_TO_INDEX(sb, alloc_hint);
 
 	for (i = 0; i < sb->map_nr; i++) {
-		nr = __sbitmap_get_word(&sb->map[index],
+		nr = __sbitmap_get_word(&sb->map[index].word,
+					sb->map[index].depth,
 					SB_NR_TO_BIT(sb, alloc_hint),
 					!round_robin);
 		if (nr != -1) {
@@ -141,6 +142,37 @@ int sbitmap_get(struct sbitmap *sb, unsigned int alloc_hint, bool round_robin)
 }
 EXPORT_SYMBOL_GPL(sbitmap_get);
 
+int sbitmap_get_shallow(struct sbitmap *sb, unsigned int alloc_hint,
+			unsigned long shallow_depth)
+{
+	unsigned int i, index;
+	int nr = -1;
+
+	index = SB_NR_TO_INDEX(sb, alloc_hint);
+
+	for (i = 0; i < sb->map_nr; i++) {
+		nr = __sbitmap_get_word(&sb->map[index].word,
+					min(sb->map[index].depth, shallow_depth),
+					SB_NR_TO_BIT(sb, alloc_hint), true);
+		if (nr != -1) {
+			nr += index << sb->shift;
+			break;
+		}
+
+		/* Jump to next index. */
+		index++;
+		alloc_hint = index << sb->shift;
+
+		if (index >= sb->map_nr) {
+			index = 0;
+			alloc_hint = 0;
+		}
+	}
+
+	return nr;
+}
+EXPORT_SYMBOL_GPL(sbitmap_get_shallow);
+
 bool sbitmap_any_bit_set(const struct sbitmap *sb)
 {
 	unsigned int i;
@@ -342,6 +374,35 @@ int __sbitmap_queue_get(struct sbitmap_queue *sbq)
 }
 EXPORT_SYMBOL_GPL(__sbitmap_queue_get);
 
+int __sbitmap_queue_get_shallow(struct sbitmap_queue *sbq,
+				unsigned int shallow_depth)
+{
+	unsigned int hint, depth;
+	int nr;
+
+	hint = this_cpu_read(*sbq->alloc_hint);
+	depth = READ_ONCE(sbq->sb.depth);
+	if (unlikely(hint >= depth)) {
+		hint = depth ? prandom_u32() % depth : 0;
+		this_cpu_write(*sbq->alloc_hint, hint);
+	}
+	nr = sbitmap_get_shallow(&sbq->sb, hint, shallow_depth);
+
+	if (nr == -1) {
+		/* If the map is full, a hint won't do us much good. */
+		this_cpu_write(*sbq->alloc_hint, 0);
+	} else if (nr == hint || unlikely(sbq->round_robin)) {
+		/* Only update the hint if we used it. */
+		hint = nr + 1;
+		if (hint >= depth - 1)
+			hint = 0;
+		this_cpu_write(*sbq->alloc_hint, hint);
+	}
+
+	return nr;
+}
+EXPORT_SYMBOL_GPL(__sbitmap_queue_get_shallow);
+
 static struct sbq_wait_state *sbq_wake_ptr(struct sbitmap_queue *sbq)
 {
 	int i, wake_index;
-- 
2.12.2

[PATCH v2 2/5] blk-mq: add shallow depth option for blk_mq_get_tag()

[Omar Sandoval]

From: Omar Sandoval <osandov@fb.com>

Wire up the sbitmap_get_shallow() operation to the tag code so that a
caller can limit the number of tags available to it.

Signed-off-by: Omar Sandoval <osandov@fb.com>
---
 block/blk-mq-tag.c | 5 ++++-
 block/blk-mq.h     | 1 +
 2 files changed, 5 insertions(+), 1 deletion(-)

diff --git a/block/blk-mq-tag.c b/block/blk-mq-tag.c
index 9d97bfc4d465..d0be72ccb091 100644
--- a/block/blk-mq-tag.c
+++ b/block/blk-mq-tag.c
@@ -96,7 +96,10 @@ static int __blk_mq_get_tag(struct blk_mq_alloc_data *data,
 	if (!(data->flags & BLK_MQ_REQ_INTERNAL) &&
 	    !hctx_may_queue(data->hctx, bt))
 		return -1;
-	return __sbitmap_queue_get(bt);
+	if (data->shallow_depth)
+		return __sbitmap_queue_get_shallow(bt, data->shallow_depth);
+	else
+		return __sbitmap_queue_get(bt);
 }
 
 unsigned int blk_mq_get_tag(struct blk_mq_alloc_data *data)
diff --git a/block/blk-mq.h b/block/blk-mq.h
index 8d49c06fc520..77ec66369f21 100644
--- a/block/blk-mq.h
+++ b/block/blk-mq.h
@@ -141,6 +141,7 @@ struct blk_mq_alloc_data {
 	/* input parameter */
 	struct request_queue *q;
 	unsigned int flags;
+	unsigned int shallow_depth;
 
 	/* input & output parameter */
 	struct blk_mq_ctx *ctx;
-- 
2.12.2

[PATCH v2 3/5] blk-mq-sched: make completed_request() callback more useful

[Omar Sandoval]

From: Omar Sandoval <osandov@fb.com>

Currently, this callback is called right after put_request() and has no
distinguishable purpose. Instead, let's call it before put_request() as
soon as I/O has completed on the request, before we account it in
blk-stat. With this, Kyber can enable stats when it sees a latency
outlier and make sure the outlier gets accounted.

Signed-off-by: Omar Sandoval <osandov@fb.com>
---
 block/blk-mq-sched.h     | 11 +++--------
 block/blk-mq.c           |  5 ++++-
 include/linux/elevator.h |  2 +-
 3 files changed, 8 insertions(+), 10 deletions(-)

diff --git a/block/blk-mq-sched.h b/block/blk-mq-sched.h
index c6e760df0fb4..d846e07bb564 100644
--- a/block/blk-mq-sched.h
+++ b/block/blk-mq-sched.h
@@ -82,17 +82,12 @@ blk_mq_sched_allow_merge(struct request_queue *q, struct request *rq,
 	return true;
 }
 
-static inline void
-blk_mq_sched_completed_request(struct blk_mq_hw_ctx *hctx, struct request *rq)
+static inline void blk_mq_sched_completed_request(struct request *rq)
 {
-	struct elevator_queue *e = hctx->queue->elevator;
+	struct elevator_queue *e = rq->q->elevator;
 
 	if (e && e->type->ops.mq.completed_request)
-		e->type->ops.mq.completed_request(hctx, rq);
-
-	BUG_ON(rq->internal_tag == -1);
-
-	blk_mq_put_tag(hctx, hctx->sched_tags, rq->mq_ctx, rq->internal_tag);
+		e->type->ops.mq.completed_request(rq);
 }
 
 static inline void blk_mq_sched_started_request(struct request *rq)
diff --git a/block/blk-mq.c b/block/blk-mq.c
index b49fdfde7e06..3dbfed6d0e5b 100644
--- a/block/blk-mq.c
+++ b/block/blk-mq.c
@@ -350,7 +350,7 @@ void __blk_mq_finish_request(struct blk_mq_hw_ctx *hctx, struct blk_mq_ctx *ctx,
 	if (rq->tag != -1)
 		blk_mq_put_tag(hctx, hctx->tags, ctx, rq->tag);
 	if (sched_tag != -1)
-		blk_mq_sched_completed_request(hctx, rq);
+		blk_mq_put_tag(hctx, hctx->sched_tags, ctx, sched_tag);
 	blk_mq_sched_restart_queues(hctx);
 	blk_queue_exit(q);
 }
@@ -443,6 +443,9 @@ static void __blk_mq_complete_request(struct request *rq)
 {
 	struct request_queue *q = rq->q;
 
+	if (rq->internal_tag != -1)
+		blk_mq_sched_completed_request(rq);
+
 	blk_mq_stat_add(rq);
 
 	if (!q->softirq_done_fn)
diff --git a/include/linux/elevator.h b/include/linux/elevator.h
index b7ec315ee7e7..3a216318ae73 100644
--- a/include/linux/elevator.h
+++ b/include/linux/elevator.h
@@ -106,7 +106,7 @@ struct elevator_mq_ops {
 	void (*insert_requests)(struct blk_mq_hw_ctx *, struct list_head *, bool);
 	struct request *(*dispatch_request)(struct blk_mq_hw_ctx *);
 	bool (*has_work)(struct blk_mq_hw_ctx *);
-	void (*completed_request)(struct blk_mq_hw_ctx *, struct request *);
+	void (*completed_request)(struct request *);
 	void (*started_request)(struct request *);
 	void (*requeue_request)(struct request *);
 	struct request *(*former_request)(struct request_queue *, struct request *);
-- 
2.12.2

[PATCH v2 4/5] blk-mq: export blk_mq_finish_request()

[Omar Sandoval]

From: Omar Sandoval <osandov@fb.com>

This is required for schedulers that define their own put_request().

Signed-off-by: Omar Sandoval <osandov@fb.com>
---
 block/blk-mq.c | 1 +
 1 file changed, 1 insertion(+)

diff --git a/block/blk-mq.c b/block/blk-mq.c
index 3dbfed6d0e5b..e75fa7a1a653 100644
--- a/block/blk-mq.c
+++ b/block/blk-mq.c
@@ -368,6 +368,7 @@ void blk_mq_finish_request(struct request *rq)
 {
 	blk_mq_finish_hctx_request(blk_mq_map_queue(rq->q, rq->mq_ctx->cpu), rq);
 }
+EXPORT_SYMBOL_GPL(blk_mq_finish_request);
 
 void blk_mq_free_request(struct request *rq)
 {
-- 
2.12.2

[PATCH v2 5/5] blk-mq: introduce Kyber multiqueue I/O scheduler

[Omar Sandoval]

From: Omar Sandoval <osandov@fb.com>

The Kyber I/O scheduler is an I/O scheduler for fast devices designed to
scale to multiple queues. Users configure only two knobs, the target
read and synchronous write latencies, and the scheduler tunes itself to
achieve that latency goal.

The implementation is based on "tokens", built on top of the scalable
bitmap library. Tokens serve as a mechanism for limiting requests. There
are two tiers of tokens: queueing tokens and dispatch tokens.

A queueing token is required to allocate a request. In fact, these
tokens are actually the blk-mq internal scheduler tags, but the
scheduler manages the allocation directly in order to implement its
policy.

Dispatch tokens are device-wide and split up into two scheduling
domains: reads vs. writes. Each hardware queue dispatches batches
round-robin between the scheduling domains as long as tokens are
available for that domain.

These tokens can be used as the mechanism to enable various policies.
The policy Kyber uses is inspired by active queue management techniques
for network routing, similar to blk-wbt. The scheduler monitors
latencies and scales the number of dispatch tokens accordingly. Queueing
tokens are used to prevent starvation of synchronous requests by
asynchronous requests.

Various extensions are possible, including better heuristics and ionice
support. The new scheduler isn't set as the default yet.

Signed-off-by: Omar Sandoval <osandov@fb.com>
---
 Documentation/block/kyber-iosched.txt |  14 +
 block/Kconfig.iosched                 |   9 +
 block/Makefile                        |   1 +
 block/kyber-iosched.c                 | 706 ++++++++++++++++++++++++++++++++++
 4 files changed, 730 insertions(+)
 create mode 100644 Documentation/block/kyber-iosched.txt
 create mode 100644 block/kyber-iosched.c

diff --git a/Documentation/block/kyber-iosched.txt b/Documentation/block/kyber-iosched.txt
new file mode 100644
index 000000000000..e94feacd7edc
--- /dev/null
+++ b/Documentation/block/kyber-iosched.txt
@@ -0,0 +1,14 @@
+Kyber I/O scheduler tunables
+===========================
+
+The only two tunables for the Kyber scheduler are the target latencies for
+reads and synchronous writes. Kyber will throttle requests in order to meet
+these target latencies.
+
+read_lat_nsec
+-------------
+Target latency for reads (in nanoseconds).
+
+write_lat_nsec
+--------------
+Target latency for synchronous writes (in nanoseconds).
diff --git a/block/Kconfig.iosched b/block/Kconfig.iosched
index 58fc8684788d..916e69c68fa4 100644
--- a/block/Kconfig.iosched
+++ b/block/Kconfig.iosched
@@ -69,6 +69,15 @@ config MQ_IOSCHED_DEADLINE
 	---help---
 	  MQ version of the deadline IO scheduler.
 
+config MQ_IOSCHED_KYBER
+	tristate "Kyber I/O scheduler"
+	default y
+	---help---
+	  The Kyber I/O scheduler is a low-overhead scheduler suitable for
+	  multiqueue and other fast devices. Given target latencies for reads and
+	  synchronous writes, it will self-tune queue depths to achieve that
+	  goal.
+
 endmenu
 
 endif
diff --git a/block/Makefile b/block/Makefile
index 081bb680789b..6146d2eaaeaa 100644
--- a/block/Makefile
+++ b/block/Makefile
@@ -20,6 +20,7 @@ obj-$(CONFIG_IOSCHED_NOOP)	+= noop-iosched.o
 obj-$(CONFIG_IOSCHED_DEADLINE)	+= deadline-iosched.o
 obj-$(CONFIG_IOSCHED_CFQ)	+= cfq-iosched.o
 obj-$(CONFIG_MQ_IOSCHED_DEADLINE)	+= mq-deadline.o
+obj-$(CONFIG_MQ_IOSCHED_KYBER)	+= kyber-iosched.o
 
 obj-$(CONFIG_BLOCK_COMPAT)	+= compat_ioctl.o
 obj-$(CONFIG_BLK_CMDLINE_PARSER)	+= cmdline-parser.o
diff --git a/block/kyber-iosched.c b/block/kyber-iosched.c
new file mode 100644
index 000000000000..30253d0758aa
--- /dev/null
+++ b/block/kyber-iosched.c
@@ -0,0 +1,706 @@
+/*
+ * The Kyber I/O scheduler. Controls latency by throttling queue depths using
+ * scalable techniques.
+ *
+ * Copyright (C) 2017 Facebook
+ *
+ * This program is free software; you can redistribute it and/or
+ * modify it under the terms of the GNU General Public
+ * License v2 as published by the Free Software Foundation.
+ *
+ * This program is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
+ * General Public License for more details.
+ *
+ * You should have received a copy of the GNU General Public License
+ * along with this program.  If not, see <https://www.gnu.org/licenses/>.
+ */
+
+#include <linux/kernel.h>
+#include <linux/blkdev.h>
+#include <linux/blk-mq.h>
+#include <linux/elevator.h>
+#include <linux/module.h>
+#include <linux/sbitmap.h>
+
+#include "blk.h"
+#include "blk-mq.h"
+#include "blk-mq-sched.h"
+#include "blk-mq-tag.h"
+#include "blk-stat.h"
+
+/* Scheduling domains. */
+enum {
+	KYBER_READ,
+	KYBER_SYNC_WRITE,
+	KYBER_OTHER, /* Async writes, discard, etc. */
+	KYBER_NUM_DOMAINS,
+};
+
+enum {
+	KYBER_MIN_DEPTH = 256,
+
+	/*
+	 * In order to prevent starvation of synchronous requests by a flood of
+	 * asynchronous requests, we reserve 25% of requests for synchronous
+	 * operations.
+	 */
+	KYBER_ASYNC_PERCENT = 75,
+};
+
+/*
+ * Initial device-wide depths for each scheduling domain.
+ *
+ * Even for fast devices with lots of tags like NVMe, you can saturate
+ * the device with only a fraction of the maximum possible queue depth.
+ * So, we cap these to a reasonable value.
+ */
+static const unsigned int kyber_depth[] = {
+	[KYBER_READ] = 256,
+	[KYBER_SYNC_WRITE] = 128,
+	[KYBER_OTHER] = 64,
+};
+
+/*
+ * Scheduling domain batch sizes. We favor reads.
+ */
+static const unsigned int kyber_batch_size[] = {
+	[KYBER_READ] = 16,
+	[KYBER_SYNC_WRITE] = 8,
+	[KYBER_OTHER] = 8,
+};
+
+struct kyber_queue_data {
+	struct request_queue *q;
+
+	struct blk_stat_callback *cb;
+
+	/*
+	 * The device is divided into multiple scheduling domains based on the
+	 * request type. Each domain has a fixed number of in-flight requests of
+	 * that type device-wide, limited by these tokens.
+	 */
+	struct sbitmap_queue domain_tokens[KYBER_NUM_DOMAINS];
+
+	/*
+	 * Async request percentage, converted to per-word depth for
+	 * sbitmap_get_shallow().
+	 */
+	unsigned int async_depth;
+
+	/* Target latencies in nanoseconds. */
+	u64 read_lat_nsec, write_lat_nsec;
+};
+
+struct kyber_hctx_data {
+	spinlock_t lock;
+	struct list_head rqs[KYBER_NUM_DOMAINS];
+	int cur_domain;
+	unsigned int batching;
+};
+
+static unsigned int rq_sched_domain(const struct request *rq)
+{
+	unsigned int op = rq->cmd_flags;
+
+	if ((op & REQ_OP_MASK) == REQ_OP_READ)
+		return KYBER_READ;
+	else if ((op & REQ_OP_MASK) == REQ_OP_WRITE && op_is_sync(op))
+		return KYBER_SYNC_WRITE;
+	else
+		return KYBER_OTHER;
+}
+
+enum {
+	NONE = 0,
+	GOOD = 1,
+	GREAT = 2,
+	BAD = -1,
+	AWFUL = -2,
+};
+
+#define IS_GOOD(status) ((status) > 0)
+#define IS_BAD(status) ((status) < 0)
+
+static int kyber_lat_status(struct blk_stat_callback *cb,
+			    unsigned int sched_domain, u64 target)
+{
+	u64 latency;
+
+	if (!cb->stat[sched_domain].nr_samples)
+		return NONE;
+
+	latency = cb->stat[sched_domain].mean;
+	if (latency >= 2 * target)
+		return AWFUL;
+	else if (latency > target)
+		return BAD;
+	else if (latency <= target / 2)
+		return GREAT;
+	else /* (latency <= target) */
+		return GOOD;
+}
+
+/*
+ * Adjust the read or synchronous write depth given the status of reads and
+ * writes. The goal is that the latencies of the two domains are fair (i.e., if
+ * one is good, then the other is good).
+ */
+static void kyber_adjust_rw_depth(struct kyber_queue_data *kqd,
+				  unsigned int sched_domain, int this_status,
+				  int other_status)
+{
+	unsigned int orig_depth, depth;
+
+	/*
+	 * If this domain had no samples, or reads and writes are both good or
+	 * both bad, don't adjust the depth.
+	 */
+	if (this_status == NONE ||
+	    (IS_GOOD(this_status) && IS_GOOD(other_status)) ||
+	    (IS_BAD(this_status) && IS_BAD(other_status)))
+		return;
+
+	orig_depth = depth = kqd->domain_tokens[sched_domain].sb.depth;
+
+	if (other_status == NONE) {
+		depth++;
+	} else {
+		switch (this_status) {
+		case GOOD:
+			if (other_status == AWFUL)
+				depth -= max(depth / 4, 1U);
+			else
+				depth -= max(depth / 8, 1U);
+			break;
+		case GREAT:
+			if (other_status == AWFUL)
+				depth /= 2;
+			else
+				depth -= max(depth / 4, 1U);
+			break;
+		case BAD:
+			depth++;
+			break;
+		case AWFUL:
+			if (other_status == GREAT)
+				depth += 2;
+			else
+				depth++;
+			break;
+		}
+	}
+
+	depth = clamp(depth, 1U, kyber_depth[sched_domain]);
+	if (depth != orig_depth)
+		sbitmap_queue_resize(&kqd->domain_tokens[sched_domain], depth);
+}
+
+/*
+ * Adjust the depth of other requests given the status of reads and synchronous
+ * writes. As long as either domain is doing fine, we don't throttle, but if
+ * both domains are doing badly, we throttle heavily.
+ */
+static void kyber_adjust_other_depth(struct kyber_queue_data *kqd,
+				     int read_status, int write_status,
+				     bool have_samples)
+{
+	unsigned int orig_depth, depth;
+	int status;
+
+	orig_depth = depth = kqd->domain_tokens[KYBER_OTHER].sb.depth;
+
+	if (read_status == NONE && write_status == NONE) {
+		depth += 2;
+	} else if (have_samples) {
+		if (read_status == NONE)
+			status = write_status;
+		else if (write_status == NONE)
+			status = read_status;
+		else
+			status = max(read_status, write_status);
+		switch (status) {
+		case GREAT:
+			depth += 2;
+			break;
+		case GOOD:
+			depth++;
+			break;
+		case BAD:
+			depth -= max(depth / 4, 1U);
+			break;
+		case AWFUL:
+			depth /= 2;
+			break;
+		}
+	}
+
+	depth = clamp(depth, 1U, kyber_depth[KYBER_OTHER]);
+	if (depth != orig_depth)
+		sbitmap_queue_resize(&kqd->domain_tokens[KYBER_OTHER], depth);
+}
+
+/*
+ * Apply heuristics for limiting queue depths based on gathered latency
+ * statistics.
+ */
+static void kyber_stat_timer_fn(struct blk_stat_callback *cb)
+{
+	struct kyber_queue_data *kqd = cb->data;
+	int read_status, write_status;
+
+	read_status = kyber_lat_status(cb, KYBER_READ, kqd->read_lat_nsec);
+	write_status = kyber_lat_status(cb, KYBER_SYNC_WRITE, kqd->write_lat_nsec);
+
+	kyber_adjust_rw_depth(kqd, KYBER_READ, read_status, write_status);
+	kyber_adjust_rw_depth(kqd, KYBER_SYNC_WRITE, write_status, read_status);
+	kyber_adjust_other_depth(kqd, read_status, write_status,
+				 cb->stat[KYBER_OTHER].nr_samples != 0);
+
+	/*
+	 * Continue monitoring latencies if we aren't hitting the targets or
+	 * we're still throttling other requests.
+	 */
+	if (!blk_stat_is_active(kqd->cb) &&
+	    ((IS_BAD(read_status) || IS_BAD(write_status) ||
+	      kqd->domain_tokens[KYBER_OTHER].sb.depth < kyber_depth[KYBER_OTHER])))
+		blk_stat_activate_msecs(kqd->cb, 100);
+}
+
+static unsigned int kyber_sched_tags_shift(struct kyber_queue_data *kqd)
+{
+	/*
+	 * All of the hardware queues have the same depth, so we can just grab
+	 * the shift of the first one.
+	 */
+	return kqd->q->queue_hw_ctx[0]->sched_tags->bitmap_tags.sb.shift;
+}
+
+static struct kyber_queue_data *kyber_queue_data_alloc(struct request_queue *q)
+{
+	struct kyber_queue_data *kqd;
+	unsigned int max_tokens;
+	unsigned int shift;
+	int ret = -ENOMEM;
+	int i;
+
+	kqd = kmalloc_node(sizeof(*kqd), GFP_KERNEL, q->node);
+	if (!kqd)
+		goto err;
+	kqd->q = q;
+
+	kqd->cb = blk_stat_alloc_callback(kyber_stat_timer_fn, rq_sched_domain,
+					  KYBER_NUM_DOMAINS, kqd);
+	if (!kqd->cb)
+		goto err_kqd;
+
+	/*
+	 * The maximum number of tokens for any scheduling domain is at least
+	 * the queue depth of a single hardware queue. If the hardware doesn't
+	 * have many tags, still provide a reasonable number.
+	 */
+	max_tokens = max_t(unsigned int, q->tag_set->queue_depth,
+			   KYBER_MIN_DEPTH);
+	for (i = 0; i < KYBER_NUM_DOMAINS; i++) {
+		WARN_ON(!kyber_depth[i]);
+		WARN_ON(!kyber_batch_size[i]);
+		ret = sbitmap_queue_init_node(&kqd->domain_tokens[i],
+					      max_tokens, -1, false, GFP_KERNEL,
+					      q->node);
+		if (ret) {
+			while (--i >= 0)
+				sbitmap_queue_free(&kqd->domain_tokens[i]);
+			goto err_cb;
+		}
+		sbitmap_queue_resize(&kqd->domain_tokens[i], kyber_depth[i]);
+	}
+
+	shift = kyber_sched_tags_shift(kqd);
+	kqd->async_depth = (1U << shift) * KYBER_ASYNC_PERCENT / 100U;
+
+	kqd->read_lat_nsec = 2000000ULL;
+	kqd->write_lat_nsec = 10000000ULL;
+
+	return kqd;
+
+err_cb:
+	blk_stat_free_callback(kqd->cb);
+err_kqd:
+	kfree(kqd);
+err:
+	return ERR_PTR(ret);
+}
+
+static int kyber_init_sched(struct request_queue *q, struct elevator_type *e)
+{
+	struct kyber_queue_data *kqd;
+	struct elevator_queue *eq;
+
+	eq = elevator_alloc(q, e);
+	if (!eq)
+		return -ENOMEM;
+
+	kqd = kyber_queue_data_alloc(q);
+	if (IS_ERR(kqd)) {
+		kobject_put(&eq->kobj);
+		return PTR_ERR(kqd);
+	}
+
+	eq->elevator_data = kqd;
+	q->elevator = eq;
+
+	blk_stat_add_callback(q, kqd->cb);
+
+	return 0;
+}
+
+static void kyber_exit_sched(struct elevator_queue *e)
+{
+	struct kyber_queue_data *kqd = e->elevator_data;
+	struct request_queue *q = kqd->q;
+	int i;
+
+	blk_stat_remove_callback(q, kqd->cb);
+
+	for (i = 0; i < KYBER_NUM_DOMAINS; i++)
+		sbitmap_queue_free(&kqd->domain_tokens[i]);
+	blk_stat_free_callback(kqd->cb);
+	kfree(kqd);
+}
+
+static int kyber_init_hctx(struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx)
+{
+	struct kyber_hctx_data *khd;
+	int i;
+
+	khd = kmalloc_node(sizeof(*khd), GFP_KERNEL, hctx->numa_node);
+	if (!khd)
+		return -ENOMEM;
+
+	spin_lock_init(&khd->lock);
+
+	for (i = 0; i < KYBER_NUM_DOMAINS; i++)
+		INIT_LIST_HEAD(&khd->rqs[i]);
+
+	khd->cur_domain = 0;
+	khd->batching = 0;
+
+	hctx->sched_data = khd;
+
+	return 0;
+}
+
+static void kyber_exit_hctx(struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx)
+{
+	kfree(hctx->sched_data);
+}
+
+static int kyber_get_domain_token(struct kyber_queue_data *kqd,
+				  unsigned int sched_domain)
+{
+	struct sbitmap_queue *domain_tokens;
+
+	domain_tokens = &kqd->domain_tokens[sched_domain];
+	return __sbitmap_queue_get(domain_tokens);
+}
+
+static int rq_get_domain_token(struct request *rq)
+{
+	return (long)rq->elv.priv[0];
+}
+
+static void rq_set_domain_token(struct request *rq, int token)
+{
+	rq->elv.priv[0] = (void *)(long)token;
+}
+
+static void rq_clear_domain_token(struct kyber_queue_data *kqd,
+				  struct request *rq)
+{
+	unsigned int sched_domain;
+	int nr;
+
+	nr = rq_get_domain_token(rq);
+	if (nr != -1) {
+		sched_domain = rq_sched_domain(rq);
+		sbitmap_queue_clear(&kqd->domain_tokens[sched_domain], nr,
+				    rq->mq_ctx->cpu);
+	}
+}
+
+static struct request *kyber_get_request(struct request_queue *q,
+					 unsigned int op,
+					 struct blk_mq_alloc_data *data)
+{
+	struct kyber_queue_data *kqd = q->elevator->elevator_data;
+	struct request *rq;
+
+	/*
+	 * We use the scheduler tags as per-hardware queue queueing tokens.
+	 * Async requests can be limited at this stage.
+	 */
+	if (!op_is_sync(op))
+		data->shallow_depth = kqd->async_depth;
+
+	rq = __blk_mq_alloc_request(data, op);
+	if (rq)
+		rq_set_domain_token(rq, -1);
+	return rq;
+}
+
+static void kyber_put_request(struct request *rq)
+{
+	struct request_queue *q = rq->q;
+	struct kyber_queue_data *kqd = q->elevator->elevator_data;
+
+	rq_clear_domain_token(kqd, rq);
+	blk_mq_finish_request(rq);
+}
+
+static void kyber_completed_request(struct request *rq)
+{
+	struct request_queue *q = rq->q;
+	struct kyber_queue_data *kqd = q->elevator->elevator_data;
+	unsigned int sched_domain;
+	u64 now, latency, target;
+
+	/*
+	 * Check if this request met our latency goal. If not, quickly gather
+	 * some statistics and start throttling.
+	 */
+	sched_domain = rq_sched_domain(rq);
+	switch (sched_domain) {
+	case KYBER_READ:
+		target = kqd->read_lat_nsec;
+		break;
+	case KYBER_SYNC_WRITE:
+		target = kqd->write_lat_nsec;
+		break;
+	default:
+		return;
+	}
+
+	/* If we are already monitoring latencies, don't check again. */
+	if (blk_stat_is_active(kqd->cb))
+		return;
+
+	now = __blk_stat_time(ktime_to_ns(ktime_get()));
+	if (now < blk_stat_time(&rq->issue_stat))
+		return;
+
+	latency = now - blk_stat_time(&rq->issue_stat);
+
+	if (latency > target)
+		blk_stat_activate_msecs(kqd->cb, 10);
+}
+
+static void kyber_flush_busy_ctxs(struct blk_mq_hw_ctx *hctx,
+				  struct kyber_hctx_data *khd)
+{
+	LIST_HEAD(rq_list);
+	struct request *rq, *next;
+
+	blk_mq_flush_busy_ctxs(hctx, &rq_list);
+	list_for_each_entry_safe(rq, next, &rq_list, queuelist) {
+		unsigned int sched_domain;
+
+		sched_domain = rq_sched_domain(rq);
+		list_move_tail(&rq->queuelist, &khd->rqs[sched_domain]);
+	}
+}
+
+static struct request *
+kyber_dispatch_cur_domain(struct blk_mq_hw_ctx *hctx,
+			  struct kyber_queue_data *kqd,
+			  struct kyber_hctx_data *khd,
+			  bool *flushed, bool *no_tokens)
+{
+	struct list_head *rqs;
+	struct request *rq;
+	int nr;
+
+	rqs = &khd->rqs[khd->cur_domain];
+	rq = list_first_entry_or_null(rqs, struct request, queuelist);
+
+	/*
+	 * If there wasn't already a pending request and we haven't flushed the
+	 * software queues yet, flush the software queues and check again.
+	 */
+	if (!rq && !*flushed) {
+		kyber_flush_busy_ctxs(hctx, khd);
+		*flushed = true;
+		rq = list_first_entry_or_null(rqs, struct request, queuelist);
+	}
+
+	if (rq) {
+		nr = kyber_get_domain_token(kqd, khd->cur_domain);
+		if (nr == -1) {
+			*no_tokens = true;
+		} else {
+			khd->batching++;
+			rq_set_domain_token(rq, nr);
+			list_del_init(&rq->queuelist);
+			return rq;
+		}
+	}
+
+	/* There were either no pending requests or no tokens. */
+	return NULL;
+}
+
+/*
+ * Returns a request on success, NULL if there were no requests to dispatch, and
+ * ERR_PTR(-EBUSY) if there were requests to dispatch but no domain tokens for
+ * them.
+ */
+static struct request *__kyber_dispatch_request(struct kyber_queue_data *kqd,
+						struct kyber_hctx_data *khd,
+						struct blk_mq_hw_ctx *hctx)
+{
+	bool flushed = false, no_tokens = false;
+	struct request *rq;
+	int i;
+
+	/*
+	 * First, if we are still entitled to batch, try to dispatch a request
+	 * from the batch.
+	 */
+	if (khd->batching < kyber_batch_size[khd->cur_domain]) {
+		rq = kyber_dispatch_cur_domain(hctx, kqd, khd, &flushed,
+					       &no_tokens);
+		if (rq)
+			return rq;
+	}
+
+	/*
+	 * Either,
+	 * 1. We were no longer entitled to a batch.
+	 * 2. The domain we were batching didn't have any requests.
+	 * 3. The domain we were batching was out of tokens.
+	 *
+	 * Start another batch. Note that this wraps back around to the original
+	 * domain if no other domains have requests or tokens.
+	 */
+	khd->batching = 0;
+	for (i = 0; i < KYBER_NUM_DOMAINS; i++) {
+		if (++khd->cur_domain >= KYBER_NUM_DOMAINS)
+			khd->cur_domain = 0;
+
+		rq = kyber_dispatch_cur_domain(hctx, kqd, khd, &flushed,
+					       &no_tokens);
+		if (rq)
+			return rq;
+	}
+
+	return no_tokens ? ERR_PTR(-EBUSY) : NULL;
+}
+
+static struct request *kyber_dispatch_request(struct blk_mq_hw_ctx *hctx)
+{
+	struct kyber_queue_data *kqd = hctx->queue->elevator->elevator_data;
+	struct kyber_hctx_data *khd = hctx->sched_data;
+	struct request *rq;
+
+	spin_lock(&khd->lock);
+
+	rq = __kyber_dispatch_request(kqd, khd, hctx);
+	if (IS_ERR(rq)) {
+		/*
+		 * We failed to get a domain token. Mark the queue as needing a
+		 * restart and try again in case a token was freed before we set
+		 * the restart bit.
+		 */
+		blk_mq_sched_mark_restart_queue(hctx);
+		rq = __kyber_dispatch_request(kqd, khd, hctx);
+		if (IS_ERR(rq))
+			rq = NULL;
+	}
+
+	spin_unlock(&khd->lock);
+
+	return rq;
+}
+
+static bool kyber_has_work(struct blk_mq_hw_ctx *hctx)
+{
+	struct kyber_hctx_data *khd = hctx->sched_data;
+	int i;
+
+	for (i = 0; i < KYBER_NUM_DOMAINS; i++) {
+		if (!list_empty_careful(&khd->rqs[i]))
+			return true;
+	}
+	return false;
+}
+
+#define KYBER_LAT_SHOW_STORE(op)					\
+static ssize_t kyber_##op##_lat_show(struct elevator_queue *e,		\
+				     char *page)			\
+{									\
+	struct kyber_queue_data *kqd = e->elevator_data;		\
+									\
+	return sprintf(page, "%llu\n", kqd->op##_lat_nsec);		\
+}									\
+									\
+static ssize_t kyber_##op##_lat_store(struct elevator_queue *e,		\
+				      const char *page, size_t count)	\
+{									\
+	struct kyber_queue_data *kqd = e->elevator_data;		\
+	unsigned long long nsec;					\
+	int ret;							\
+									\
+	ret = kstrtoull(page, 10, &nsec);				\
+	if (ret)							\
+		return ret;						\
+									\
+	kqd->op##_lat_nsec = nsec;					\
+									\
+	return count;							\
+}
+KYBER_LAT_SHOW_STORE(read);
+KYBER_LAT_SHOW_STORE(write);
+#undef KYBER_LAT_SHOW_STORE
+
+#define KYBER_LAT_ATTR(op) __ATTR(op##_lat_nsec, 0644, kyber_##op##_lat_show, kyber_##op##_lat_store)
+static struct elv_fs_entry kyber_sched_attrs[] = {
+	KYBER_LAT_ATTR(read),
+	KYBER_LAT_ATTR(write),
+	__ATTR_NULL
+};
+#undef KYBER_LAT_ATTR
+
+static struct elevator_type kyber_sched = {
+	.ops.mq = {
+		.init_sched = kyber_init_sched,
+		.exit_sched = kyber_exit_sched,
+		.init_hctx = kyber_init_hctx,
+		.exit_hctx = kyber_exit_hctx,
+		.get_request = kyber_get_request,
+		.put_request = kyber_put_request,
+		.completed_request = kyber_completed_request,
+		.dispatch_request = kyber_dispatch_request,
+		.has_work = kyber_has_work,
+	},
+	.uses_mq = true,
+	.elevator_attrs = kyber_sched_attrs,
+	.elevator_name = "kyber",
+	.elevator_owner = THIS_MODULE,
+};
+
+static int __init kyber_init(void)
+{
+	return elv_register(&kyber_sched);
+}
+
+static void __exit kyber_exit(void)
+{
+	elv_unregister(&kyber_sched);
+}
+
+module_init(kyber_init);
+module_exit(kyber_exit);
+
+MODULE_AUTHOR("Omar Sandoval");
+MODULE_LICENSE("GPL");
+MODULE_DESCRIPTION("Kyber I/O scheduler");
-- 
2.12.2