block/mq-deadline: Add cgroup support

Maintain statistics per cgroup and export these to user space. These
statistics are essential for verifying whether the proper I/O priorities
have been assigned to requests. An example of the statistics data with
this patch applied:

$ cat /sys/fs/cgroup/io.stat
11:2 rbytes=0 wbytes=0 rios=3 wios=0 dbytes=0 dios=0 [NONE] dispatched=0 inserted=0 merged=171 [RT] dispatched=0 inserted=0 merged=0 [BE] dispatched=0 inserted=0 merged=0 [IDLE] dispatched=0 inserted=0 merged=0
8:32 rbytes=2142720 wbytes=0 rios=105 wios=0 dbytes=0 dios=0 [NONE] dispatched=0 inserted=0 merged=171 [RT] dispatched=0 inserted=0 merged=0 [BE] dispatched=0 inserted=0 merged=0 [IDLE] dispatched=0 inserted=0 merged=0

Cc: Damien Le Moal <damien.lemoal@wdc.com>
Cc: Hannes Reinecke <hare@suse.de>
Cc: Christoph Hellwig <hch@lst.de>
Cc: Ming Lei <ming.lei@redhat.com>
Cc: Johannes Thumshirn <johannes.thumshirn@wdc.com>
Cc: Himanshu Madhani <himanshu.madhani@oracle.com>
Signed-off-by: Bart Van Assche <bvanassche@acm.org>
Link: https://lore.kernel.org/r/20210618004456.7280-16-bvanassche@acm.org
Signed-off-by: Jens Axboe <axboe@kernel.dk>

5 files changed, 308 insertions, 14 deletions

diff --git a/block/Kconfig.iosched b/block/Kconfig.iosched
index 2f2158e05a91..64053d67a97b 100644
--- a/block/Kconfig.iosched
+++ b/block/Kconfig.iosched
@@ -9,6 +9,12 @@ config MQ_IOSCHED_DEADLINE
 	help
 	  MQ version of the deadline IO scheduler.
 
+config MQ_IOSCHED_DEADLINE_CGROUP
+       tristate
+       default y
+       depends on MQ_IOSCHED_DEADLINE
+       depends on BLK_CGROUP
+
 config MQ_IOSCHED_KYBER
 	tristate "Kyber I/O scheduler"
 	default y
diff --git a/block/Makefile b/block/Makefile
index af3d044abaf1..b9db5d4edfc8 100644
--- a/block/Makefile
+++ b/block/Makefile
@@ -21,6 +21,8 @@ obj-$(CONFIG_BLK_CGROUP_IOPRIO)	+= blk-ioprio.o
 obj-$(CONFIG_BLK_CGROUP_IOLATENCY)	+= blk-iolatency.o
 obj-$(CONFIG_BLK_CGROUP_IOCOST)	+= blk-iocost.o
 obj-$(CONFIG_MQ_IOSCHED_DEADLINE)	+= mq-deadline.o
+mq-deadline-y += mq-deadline-main.o
+mq-deadline-$(CONFIG_MQ_IOSCHED_DEADLINE_CGROUP)+= mq-deadline-cgroup.o
 obj-$(CONFIG_MQ_IOSCHED_KYBER)	+= kyber-iosched.o
 bfq-y				:= bfq-iosched.o bfq-wf2q.o bfq-cgroup.o
 obj-$(CONFIG_IOSCHED_BFQ)	+= bfq.o
diff --git a/block/mq-deadline-cgroup.c b/block/mq-deadline-cgroup.c
new file mode 100644
index 000000000000..3b4bfddec39f
--- /dev/null
+++ b/block/mq-deadline-cgroup.c
@@ -0,0 +1,126 @@
+// SPDX-License-Identifier: GPL-2.0
+
+#include <linux/blk-cgroup.h>
+#include <linux/ioprio.h>
+
+#include "mq-deadline-cgroup.h"
+
+static struct blkcg_policy dd_blkcg_policy;
+
+static struct blkcg_policy_data *dd_cpd_alloc(gfp_t gfp)
+{
+	struct dd_blkcg *pd;
+
+	pd = kzalloc(sizeof(*pd), gfp);
+	if (!pd)
+		return NULL;
+	pd->stats = alloc_percpu_gfp(typeof(*pd->stats),
+				     GFP_KERNEL | __GFP_ZERO);
+	if (!pd->stats) {
+		kfree(pd);
+		return NULL;
+	}
+	return &pd->cpd;
+}
+
+static void dd_cpd_free(struct blkcg_policy_data *cpd)
+{
+	struct dd_blkcg *dd_blkcg = container_of(cpd, typeof(*dd_blkcg), cpd);
+
+	free_percpu(dd_blkcg->stats);
+	kfree(dd_blkcg);
+}
+
+static struct dd_blkcg *dd_blkcg_from_pd(struct blkg_policy_data *pd)
+{
+	return container_of(blkcg_to_cpd(pd->blkg->blkcg, &dd_blkcg_policy),
+			    struct dd_blkcg, cpd);
+}
+
+/*
+ * Convert an association between a block cgroup and a request queue into a
+ * pointer to the mq-deadline information associated with a (blkcg, queue) pair.
+ */
+struct dd_blkcg *dd_blkcg_from_bio(struct bio *bio)
+{
+	struct blkg_policy_data *pd;
+
+	pd = blkg_to_pd(bio->bi_blkg, &dd_blkcg_policy);
+	if (!pd)
+		return NULL;
+
+	return dd_blkcg_from_pd(pd);
+}
+
+static size_t dd_pd_stat(struct blkg_policy_data *pd, char *buf, size_t size)
+{
+	static const char *const prio_class_name[] = {
+		[IOPRIO_CLASS_NONE]	= "NONE",
+		[IOPRIO_CLASS_RT]	= "RT",
+		[IOPRIO_CLASS_BE]	= "BE",
+		[IOPRIO_CLASS_IDLE]	= "IDLE",
+	};
+	struct dd_blkcg *blkcg = dd_blkcg_from_pd(pd);
+	int res = 0;
+	u8 prio;
+
+	for (prio = 0; prio < ARRAY_SIZE(blkcg->stats->stats); prio++)
+		res += scnprintf(buf + res, size - res,
+			" [%s] dispatched=%u inserted=%u merged=%u",
+			prio_class_name[prio],
+			ddcg_sum(blkcg, dispatched, prio) +
+			ddcg_sum(blkcg, merged, prio) -
+			ddcg_sum(blkcg, completed, prio),
+			ddcg_sum(blkcg, inserted, prio) -
+			ddcg_sum(blkcg, completed, prio),
+			ddcg_sum(blkcg, merged, prio));
+
+	return res;
+}
+
+static struct blkg_policy_data *dd_pd_alloc(gfp_t gfp, struct request_queue *q,
+					    struct blkcg *blkcg)
+{
+	struct dd_blkg *pd;
+
+	pd = kzalloc(sizeof(*pd), gfp);
+	if (!pd)
+		return NULL;
+	return &pd->pd;
+}
+
+static void dd_pd_free(struct blkg_policy_data *pd)
+{
+	struct dd_blkg *dd_blkg = container_of(pd, typeof(*dd_blkg), pd);
+
+	kfree(dd_blkg);
+}
+
+static struct blkcg_policy dd_blkcg_policy = {
+	.cpd_alloc_fn		= dd_cpd_alloc,
+	.cpd_free_fn		= dd_cpd_free,
+
+	.pd_alloc_fn		= dd_pd_alloc,
+	.pd_free_fn		= dd_pd_free,
+	.pd_stat_fn		= dd_pd_stat,
+};
+
+int dd_activate_policy(struct request_queue *q)
+{
+	return blkcg_activate_policy(q, &dd_blkcg_policy);
+}
+
+void dd_deactivate_policy(struct request_queue *q)
+{
+	blkcg_deactivate_policy(q, &dd_blkcg_policy);
+}
+
+int __init dd_blkcg_init(void)
+{
+	return blkcg_policy_register(&dd_blkcg_policy);
+}
+
+void __exit dd_blkcg_exit(void)
+{
+	blkcg_policy_unregister(&dd_blkcg_policy);
+}
diff --git a/block/mq-deadline-cgroup.h b/block/mq-deadline-cgroup.h
new file mode 100644
index 000000000000..0143fd74f3ce
--- /dev/null
+++ b/block/mq-deadline-cgroup.h
@@ -0,0 +1,114 @@
+/* SPDX-License-Identifier: GPL-2.0 */
+
+#if !defined(_MQ_DEADLINE_CGROUP_H_)
+#define _MQ_DEADLINE_CGROUP_H_
+
+#include <linux/blk-cgroup.h>
+
+struct request_queue;
+
+/**
+ * struct io_stats_per_prio - I/O statistics per I/O priority class.
+ * @inserted: Number of inserted requests.
+ * @merged: Number of merged requests.
+ * @dispatched: Number of dispatched requests.
+ * @completed: Number of I/O completions.
+ */
+struct io_stats_per_prio {
+	local_t inserted;
+	local_t merged;
+	local_t dispatched;
+	local_t completed;
+};
+
+/* I/O statistics per I/O cgroup per I/O priority class (IOPRIO_CLASS_*). */
+struct blkcg_io_stats {
+	struct io_stats_per_prio stats[4];
+};
+
+/**
+ * struct dd_blkcg - Per cgroup data.
+ * @cpd: blkcg_policy_data structure.
+ * @stats: I/O statistics.
+ */
+struct dd_blkcg {
+	struct blkcg_policy_data cpd;	/* must be the first member */
+	struct blkcg_io_stats __percpu *stats;
+};
+
+/*
+ * Count one event of type 'event_type' and with I/O priority class
+ * 'prio_class'.
+ */
+#define ddcg_count(ddcg, event_type, prio_class) do {			\
+if (ddcg) {								\
+	struct blkcg_io_stats *io_stats = get_cpu_ptr((ddcg)->stats);	\
+									\
+	BUILD_BUG_ON(!__same_type((ddcg), struct dd_blkcg *));		\
+	BUILD_BUG_ON(!__same_type((prio_class), u8));			\
+	local_inc(&io_stats->stats[(prio_class)].event_type);		\
+	put_cpu_ptr(io_stats);						\
+}									\
+} while (0)
+
+/*
+ * Returns the total number of ddcg_count(ddcg, event_type, prio_class) calls
+ * across all CPUs. No locking or barriers since it is fine if the returned
+ * sum is slightly outdated.
+ */
+#define ddcg_sum(ddcg, event_type, prio) ({				\
+	unsigned int cpu;						\
+	u32 sum = 0;							\
+									\
+	BUILD_BUG_ON(!__same_type((ddcg), struct dd_blkcg *));		\
+	BUILD_BUG_ON(!__same_type((prio), u8));				\
+	for_each_present_cpu(cpu)					\
+		sum += local_read(&per_cpu_ptr((ddcg)->stats, cpu)->	\
+				  stats[(prio)].event_type);		\
+	sum;								\
+})
+
+#ifdef CONFIG_BLK_CGROUP
+
+/**
+ * struct dd_blkg - Per (cgroup, request queue) data.
+ * @pd: blkg_policy_data structure.
+ */
+struct dd_blkg {
+	struct blkg_policy_data pd;	/* must be the first member */
+};
+
+struct dd_blkcg *dd_blkcg_from_bio(struct bio *bio);
+int dd_activate_policy(struct request_queue *q);
+void dd_deactivate_policy(struct request_queue *q);
+int __init dd_blkcg_init(void);
+void __exit dd_blkcg_exit(void);
+
+#else /* CONFIG_BLK_CGROUP */
+
+static inline struct dd_blkcg *dd_blkcg_from_bio(struct bio *bio)
+{
+	return NULL;
+}
+
+static inline int dd_activate_policy(struct request_queue *q)
+{
+	return 0;
+}
+
+static inline void dd_deactivate_policy(struct request_queue *q)
+{
+}
+
+static inline int dd_blkcg_init(void)
+{
+	return 0;
+}
+
+static inline void dd_blkcg_exit(void)
+{
+}
+
+#endif /* CONFIG_BLK_CGROUP */
+
+#endif /* _MQ_DEADLINE_CGROUP_H_ */
diff --git a/block/mq-deadline.c b/block/mq-deadline-main.c
index 04d9d6b3745b..58a401ea8f56 100644
--- a/block/mq-deadline.c
+++ b/block/mq-deadline-main.c
@@ -25,6 +25,7 @@
 #include "blk-mq-debugfs.h"
 #include "blk-mq-tag.h"
 #include "blk-mq-sched.h"
+#include "mq-deadline-cgroup.h"
 
 /*
  * See Documentation/block/deadline-iosched.rst
@@ -51,14 +52,6 @@ enum dd_prio {
 
 enum { DD_PRIO_COUNT = 3 };
 
-/* I/O statistics per I/O priority. */
-struct io_stats_per_prio {
-	local_t inserted;
-	local_t merged;
-	local_t dispatched;
-	local_t completed;
-};
-
 /* I/O statistics for all I/O priorities (enum dd_prio). */
 struct io_stats {
 	struct io_stats_per_prio stats[DD_PRIO_COUNT];
@@ -81,6 +74,9 @@ struct deadline_data {
 	 * run time data
 	 */
 
+	/* Request queue that owns this data structure. */
+	struct request_queue *queue;
+
 	struct dd_per_prio per_prio[DD_PRIO_COUNT];
 
 	/* Data direction of latest dispatched request. */
@@ -232,8 +228,10 @@ static void dd_merged_requests(struct request_queue *q, struct request *req,
 	struct deadline_data *dd = q->elevator->elevator_data;
 	const u8 ioprio_class = dd_rq_ioclass(next);
 	const enum dd_prio prio = ioprio_class_to_prio[ioprio_class];
+	struct dd_blkcg *blkcg = next->elv.priv[0];
 
 	dd_count(dd, merged, prio);
+	ddcg_count(blkcg, merged, ioprio_class);
 
 	/*
 	 * if next expires before rq, assign its expire time to rq
@@ -370,6 +368,7 @@ static struct request *__dd_dispatch_request(struct deadline_data *dd,
 {
 	struct request *rq, *next_rq;
 	enum dd_data_dir data_dir;
+	struct dd_blkcg *blkcg;
 	enum dd_prio prio;
 	u8 ioprio_class;
 
@@ -464,6 +463,8 @@ done:
 	ioprio_class = dd_rq_ioclass(rq);
 	prio = ioprio_class_to_prio[ioprio_class];
 	dd_count(dd, dispatched, prio);
+	blkcg = rq->elv.priv[0];
+	ddcg_count(blkcg, dispatched, ioprio_class);
 	/*
 	 * If the request needs its target zone locked, do it.
 	 */
@@ -540,6 +541,8 @@ static void dd_exit_sched(struct elevator_queue *e)
 	struct deadline_data *dd = e->elevator_data;
 	enum dd_prio prio;
 
+	dd_deactivate_policy(dd->queue);
+
 	for (prio = 0; prio <= DD_PRIO_MAX; prio++) {
 		struct dd_per_prio *per_prio = &dd->per_prio[prio];
 
@@ -553,7 +556,7 @@ static void dd_exit_sched(struct elevator_queue *e)
 }
 
 /*
- * initialize elevator private data (deadline_data).
+ * Initialize elevator private data (deadline_data) and associate with blkcg.
  */
 static int dd_init_sched(struct request_queue *q, struct elevator_type *e)
 {
@@ -562,6 +565,12 @@ static int dd_init_sched(struct request_queue *q, struct elevator_type *e)
 	enum dd_prio prio;
 	int ret = -ENOMEM;
 
+	/*
+	 * Initialization would be very tricky if the queue is not frozen,
+	 * hence the warning statement below.
+	 */
+	WARN_ON_ONCE(!percpu_ref_is_zero(&q->q_usage_counter));
+
 	eq = elevator_alloc(q, e);
 	if (!eq)
 		return ret;
@@ -577,6 +586,8 @@ static int dd_init_sched(struct request_queue *q, struct elevator_type *e)
 	if (!dd->stats)
 		goto free_dd;
 
+	dd->queue = q;
+
 	for (prio = 0; prio <= DD_PRIO_MAX; prio++) {
 		struct dd_per_prio *per_prio = &dd->per_prio[prio];
 
@@ -595,9 +606,17 @@ static int dd_init_sched(struct request_queue *q, struct elevator_type *e)
 	spin_lock_init(&dd->lock);
 	spin_lock_init(&dd->zone_lock);
 
+	ret = dd_activate_policy(q);
+	if (ret)
+		goto free_stats;
+
+	ret = 0;
 	q->elevator = eq;
 	return 0;
 
+free_stats:
+	free_percpu(dd->stats);
+
 free_dd:
 	kfree(dd);
 
@@ -670,6 +689,7 @@ static void dd_insert_request(struct blk_mq_hw_ctx *hctx, struct request *rq,
 	u8 ioprio_class = IOPRIO_PRIO_CLASS(ioprio);
 	struct dd_per_prio *per_prio;
 	enum dd_prio prio;
+	struct dd_blkcg *blkcg;
 
 	lockdep_assert_held(&dd->lock);
 
@@ -679,8 +699,19 @@ static void dd_insert_request(struct blk_mq_hw_ctx *hctx, struct request *rq,
 	 */
 	blk_req_zone_write_unlock(rq);
 
+	/*
+	 * If a block cgroup has been associated with the submitter and if an
+	 * I/O priority has been set in the associated block cgroup, use the
+	 * lowest of the cgroup priority and the request priority for the
+	 * request. If no priority has been set in the request, use the cgroup
+	 * priority.
+	 */
 	prio = ioprio_class_to_prio[ioprio_class];
 	dd_count(dd, inserted, prio);
+	blkcg = dd_blkcg_from_bio(rq->bio);
+	ddcg_count(blkcg, inserted, ioprio_class);
+	WARN_ON_ONCE(rq->elv.priv[0]);
+	rq->elv.priv[0] = blkcg;
 
 	if (blk_mq_sched_try_insert_merge(q, rq))
 		return;
@@ -727,12 +758,10 @@ static void dd_insert_requests(struct blk_mq_hw_ctx *hctx,
 	spin_unlock(&dd->lock);
 }
 
-/*
- * Nothing to do here. This is defined only to ensure that .finish_request
- * method is called upon request completion.
- */
+/* Callback from inside blk_mq_rq_ctx_init(). */
 static void dd_prepare_request(struct request *rq)
 {
+	rq->elv.priv[0] = NULL;
 }
 
 /*
@@ -755,11 +784,13 @@ static void dd_finish_request(struct request *rq)
 {
 	struct request_queue *q = rq->q;
 	struct deadline_data *dd = q->elevator->elevator_data;
+	struct dd_blkcg *blkcg = rq->elv.priv[0];
 	const u8 ioprio_class = dd_rq_ioclass(rq);
 	const enum dd_prio prio = ioprio_class_to_prio[ioprio_class];
 	struct dd_per_prio *per_prio = &dd->per_prio[prio];
 
 	dd_count(dd, completed, prio);
+	ddcg_count(blkcg, completed, ioprio_class);
 
 	if (blk_queue_is_zoned(q)) {
 		unsigned long flags;
@@ -1079,11 +1110,26 @@ MODULE_ALIAS("mq-deadline-iosched");
 
 static int __init deadline_init(void)
 {
-	return elv_register(&mq_deadline);
+	int ret;
+
+	ret = elv_register(&mq_deadline);
+	if (ret)
+		goto out;
+	ret = dd_blkcg_init();
+	if (ret)
+		goto unreg;
+
+out:
+	return ret;
+
+unreg:
+	elv_unregister(&mq_deadline);
+	goto out;
 }
 
 static void __exit deadline_exit(void)
 {
+	dd_blkcg_exit();
 	elv_unregister(&mq_deadline);
 }