Skip to contentSkip to conversation

block: move sched_tags allocation/de-allocation outside of locking context

PATCH Superseded Nilay Shroff 4 participants 15 messages Jun 30, 2025 – Jul 1, 2025

Subsystems (2)

v6 of 8 Superseded Latest is v8 →

lore
Share & tools
lore
#0 lore raw

v6 cover block: move sched_tags allocation/de-allocation outside of locking context

Hi,

There have been a few reports[1] indicating potential lockdep warnings due to a lock dependency from the percpu allocator to the elevator lock. This patch series aims to eliminate that dependency.

The series consists of three patches:
The first patch is preparatory patch and just move elevator queue allocation logic from ->init_sched into blk_mq_init_sched.

The second patch in the series restructures sched_tags allocation and deallocation during elevator update/switch operations to ensure these actions are performed entirely outside the ->freeze_lock and ->elevator_ lock. This eliminates the percpu allocator’s lock dependency on the elevator and freeze lock during scheduler transitions.

The third patch introduces batch allocation and deallocation helpers for sched_tags. These helpers are used during __blk_mq_update_nr_hw_queues() to decouple sched_tags memory management from both the elevator and freeze locks, addressing the lockdep concerns in the nr_hw_queues update path.

[1] https://lore.kernel.org/all/[email protected]/

Changes since v5:
- Fixed smatch warning reported by kernel test robot here:
https://lore.kernel.org/all/[email protected]/
Link to v5: https://lore.kernel.org/all/[email protected]/

Changes since v4:
- Define a local Xarray variable in __blk_mq_update_nr_hw_queues to store sched_tags, instead of storing it in an Xarray defined in 'struct elevator_tags' (Ming Lei)
Link to v4: https://lore.kernel.org/all/[email protected]/

Changes since v3:
- Further split the patchset into three patch series so that we can have a separate patch for sched_tags batch allocation/deallocation (Ming Lei)
- Use Xarray to store and load the sched_tags (Ming Lei)
- Unexport elevator_alloc() as we no longer need to use it outside of block layer core (hch)
- unwind the sched_tags allocation and free tags when we it fails in the middle of allocation (hch)
- Move struct elevator_queue header from commin header to elevator.c as there's no user of it outside elevator.c (Ming Lei, hch)
Link to v3: https://lore.kernel.org/all/[email protected]/

Change since v2:
- Split the patch into a two patch series. The first patch updates ->init_sched elevator API change and second patch handles the sched tags allocation/de-allocation logic (Ming Lei)
- Address sched tags allocation/deallocation logic while running in the context of nr_hw_queue update so that we can handle all possible scenarios in a single patchest (Ming Lei)
Link to v2: https://lore.kernel.org/all/[email protected]/

Changes since v1:
- As the lifetime of elevator queue and sched tags are same, allocate and move sched tags under struct elevator_queue (Ming Lei)
Link to v1: https://lore.kernel.org/all/[email protected]/

Nilay Shroff (3):
block: move elevator queue allocation logic into blk_mq_init_sched
block: fix lockdep warning caused by lock dependency in
elv_iosched_store
block: fix potential deadlock while running nr_hw_queue update

 block/bfq-iosched.c   |  13 +--
 block/blk-mq-sched.c  | 220 ++++++++++++++++++++++++++++--------------
 block/blk-mq-sched.h  |  12 ++-
 block/blk-mq.c        |  11 ++-
 block/blk.h           |   2 +-
 block/elevator.c      |  50 ++++++++--
 block/elevator.h      |  16 ++-
 block/kyber-iosched.c |  11 +--
 block/mq-deadline.c   |  14 +--
 9 files changed, 234 insertions(+), 115 deletions(-)
Signature · 1 line
2.49.0
#1 lore raw

PATCH v6 1/3 block: move elevator queue allocation logic into blk_mq_init_sched

In preparation for allocating sched_tags before freezing the request queue and acquiring ->elevator_lock, move the elevator queue allocation logic from the elevator ops ->init_sched callback into blk_mq_init_sched. As elevator_alloc is now only invoked from block layer core, we don't need to export it, so unexport elevator_alloc function.

This refactoring provides a centralized location for elevator queue initialization, which makes it easier to store pre-allocated sched_tags in the struct elevator_queue during later changes.

---
 block/bfq-iosched.c   | 13 +++----------
 block/blk-mq-sched.c  | 11 ++++++++---
 block/elevator.c      |  1 -
 block/elevator.h      |  2 +-
 block/kyber-iosched.c | 11 ++---------
 block/mq-deadline.c   | 14 ++------------
 6 files changed, 16 insertions(+), 36 deletions(-)
block/bfq-iosched.c +3 -10
index 0cb1e9873aab..fd26dc1901b0 100644
@@ -7232,22 +7232,16 @@ static void bfq_init_root_group(struct bfq_group *root_group,
 	root_group->sched_data.bfq_class_idle_last_service = jiffies;
 }
 
-static int bfq_init_queue(struct request_queue *q, struct elevator_type *e)
+static int bfq_init_queue(struct request_queue *q, struct elevator_queue *eq)
 {
 	struct bfq_data *bfqd;
-	struct elevator_queue *eq;
 	unsigned int i;
 	struct blk_independent_access_ranges *ia_ranges = q->disk->ia_ranges;
 
-	eq = elevator_alloc(q, e);
-	if (!eq)
-		return -ENOMEM;
-
 	bfqd = kzalloc_node(sizeof(*bfqd), GFP_KERNEL, q->node);
-	if (!bfqd) {
-		kobject_put(&eq->kobj);
+	if (!bfqd)
 		return -ENOMEM;
-	}
+
 	eq->elevator_data = bfqd;
 
 	spin_lock_irq(&q->queue_lock);
@@ -7405,7 +7399,6 @@ static int bfq_init_queue(struct request_queue *q, struct elevator_type *e)
 
 out_free:
 	kfree(bfqd);
-	kobject_put(&eq->kobj);
 	return -ENOMEM;
 }
 
block/blk-mq-sched.c +8 -3
index 55a0fd105147..359e0704e09b 100644
@@ -475,10 +475,14 @@ int blk_mq_init_sched(struct request_queue *q, struct elevator_type *e)
 	q->nr_requests = 2 * min_t(unsigned int, q->tag_set->queue_depth,
 				   BLKDEV_DEFAULT_RQ);
 
+	eq = elevator_alloc(q, e);
+	if (!eq)
+		return -ENOMEM;
+
 	if (blk_mq_is_shared_tags(flags)) {
 		ret = blk_mq_init_sched_shared_tags(q);
 		if (ret)
-			return ret;
+			goto err_put_elevator;
 	}
 
 	queue_for_each_hw_ctx(q, hctx, i) {
@@ -487,7 +491,7 @@ int blk_mq_init_sched(struct request_queue *q, struct elevator_type *e)
 			goto err_free_map_and_rqs;
 	}
 
-	ret = e->ops.init_sched(q, e);
+	ret = e->ops.init_sched(q, eq);
 	if (ret)
 		goto err_free_map_and_rqs;
 
@@ -508,7 +512,8 @@ int blk_mq_init_sched(struct request_queue *q, struct elevator_type *e)
 err_free_map_and_rqs:
 	blk_mq_sched_free_rqs(q);
 	blk_mq_sched_tags_teardown(q, flags);
-
+err_put_elevator:
+	kobject_put(&eq->kobj);
 	q->elevator = NULL;
 	return ret;
 }
block/elevator.c +0 -1
index ab22542e6cf0..770874040f79 100644
@@ -148,7 +148,6 @@ struct elevator_queue *elevator_alloc(struct request_queue *q,
 
 	return eq;
 }
-EXPORT_SYMBOL(elevator_alloc);
 
 static void elevator_release(struct kobject *kobj)
 {
block/elevator.h +1 -1
index a07ce773a38f..a4de5f9ad790 100644
@@ -24,7 +24,7 @@ struct blk_mq_alloc_data;
 struct blk_mq_hw_ctx;
 
 struct elevator_mq_ops {
-	int (*init_sched)(struct request_queue *, struct elevator_type *);
+	int (*init_sched)(struct request_queue *, struct elevator_queue *);
 	void (*exit_sched)(struct elevator_queue *);
 	int (*init_hctx)(struct blk_mq_hw_ctx *, unsigned int);
 	void (*exit_hctx)(struct blk_mq_hw_ctx *, unsigned int);
block/kyber-iosched.c +2 -9
index 4dba8405bd01..7b6832cb3a8d 100644
@@ -402,20 +402,13 @@ static struct kyber_queue_data *kyber_queue_data_alloc(struct request_queue *q)
 	return ERR_PTR(ret);
 }
 
-static int kyber_init_sched(struct request_queue *q, struct elevator_type *e)
+static int kyber_init_sched(struct request_queue *q, struct elevator_queue *eq)
 {
 	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);
+	if (IS_ERR(kqd))
 		return PTR_ERR(kqd);
-	}
 
 	blk_stat_enable_accounting(q);
 
block/mq-deadline.c +2 -12
index 2edf1cac06d5..7b6caf30e00a 100644
@@ -568,20 +568,14 @@ static void dd_exit_sched(struct elevator_queue *e)
 /*
  * initialize elevator private data (deadline_data).
  */
-static int dd_init_sched(struct request_queue *q, struct elevator_type *e)
+static int dd_init_sched(struct request_queue *q, struct elevator_queue *eq)
 {
 	struct deadline_data *dd;
-	struct elevator_queue *eq;
 	enum dd_prio prio;
-	int ret = -ENOMEM;
-
-	eq = elevator_alloc(q, e);
-	if (!eq)
-		return ret;
 
 	dd = kzalloc_node(sizeof(*dd), GFP_KERNEL, q->node);
 	if (!dd)
-		goto put_eq;
+		return -ENOMEM;
 
 	eq->elevator_data = dd;
 
@@ -608,10 +602,6 @@ static int dd_init_sched(struct request_queue *q, struct elevator_type *e)
 
 	q->elevator = eq;
 	return 0;
-
-put_eq:
-	kobject_put(&eq->kobj);
-	return ret;
 }
 
 /*
Signature · 1 line
2.50.0
#4 lore raw

PATCH v6 2/3 block: fix lockdep warning caused by lock dependency in elv_iosched_store

In reply to Nilay Shroff

Recent lockdep reports [1] have revealed a potential deadlock caused by a lock dependency between the percpu allocator lock and the elevator lock. This issue can be avoided by ensuring that the allocation and release of scheduler tags (sched_tags) are performed outside the elevator lock. Furthermore, the queue does not need to be remain frozen during these operations.

To address this, move all sched_tags allocations and deallocations outside of both the ->elevator_lock and the ->freeze_lock. Since the lifetime of the elevator queue and its associated sched_tags is closely tied, the allocated sched_tags are now stored in the elevator queue structure. Then, during the actual elevator switch (which runs under ->freeze_lock and ->elevator_lock), the pre-allocated sched_tags are assigned to the appropriate q->hctx. Once the elevator switch is complete and the locks are released, the old elevator queue and its associated sched_tags are freed.

This commit specifically addresses the allocation/deallocation of sched_ tags during elevator switching. Note that sched_tags may also be allocated in other contexts, such as during nr_hw_queues updates. Supporting that use case will require batch allocation/deallocation, which will be handled in a follow-up patch.

This restructuring ensures that sched_tags memory management occurs entirely outside of the ->elevator_lock and ->freeze_lock context, eliminating the lock dependency problem seen during scheduler updates.

[1] https://lore.kernel.org/all/[email protected]/

---
 block/blk-mq-sched.c | 155 +++++++++++++++++++++++--------------------
 block/blk-mq-sched.h |   8 ++-
 block/elevator.c     |  47 +++++++++++--
 block/elevator.h     |  14 +++-
 4 files changed, 144 insertions(+), 80 deletions(-)
block/blk-mq-sched.c +83 -72
index 359e0704e09b..2d6d1ebdd8fb 100644
@@ -374,64 +374,17 @@ bool blk_mq_sched_try_insert_merge(struct request_queue *q, struct request *rq,
 }
 EXPORT_SYMBOL_GPL(blk_mq_sched_try_insert_merge);
 
-static int blk_mq_sched_alloc_map_and_rqs(struct request_queue *q,
-					  struct blk_mq_hw_ctx *hctx,
-					  unsigned int hctx_idx)
-{
-	if (blk_mq_is_shared_tags(q->tag_set->flags)) {
-		hctx->sched_tags = q->sched_shared_tags;
-		return 0;
-	}
-
-	hctx->sched_tags = blk_mq_alloc_map_and_rqs(q->tag_set, hctx_idx,
-						    q->nr_requests);
-
-	if (!hctx->sched_tags)
-		return -ENOMEM;
-	return 0;
-}
-
-static void blk_mq_exit_sched_shared_tags(struct request_queue *queue)
-{
-	blk_mq_free_rq_map(queue->sched_shared_tags);
-	queue->sched_shared_tags = NULL;
-}
-
 /* called in queue's release handler, tagset has gone away */
 static void blk_mq_sched_tags_teardown(struct request_queue *q, unsigned int flags)
 {
 	struct blk_mq_hw_ctx *hctx;
 	unsigned long i;
 
-	queue_for_each_hw_ctx(q, hctx, i) {
-		if (hctx->sched_tags) {
-			if (!blk_mq_is_shared_tags(flags))
-				blk_mq_free_rq_map(hctx->sched_tags);
-			hctx->sched_tags = NULL;
-		}
-	}
+	queue_for_each_hw_ctx(q, hctx, i)
+		hctx->sched_tags = NULL;
 
 	if (blk_mq_is_shared_tags(flags))
-		blk_mq_exit_sched_shared_tags(q);
-}
-
-static int blk_mq_init_sched_shared_tags(struct request_queue *queue)
-{
-	struct blk_mq_tag_set *set = queue->tag_set;
-
-	/*
-	 * Set initial depth at max so that we don't need to reallocate for
-	 * updating nr_requests.
-	 */
-	queue->sched_shared_tags = blk_mq_alloc_map_and_rqs(set,
-						BLK_MQ_NO_HCTX_IDX,
-						MAX_SCHED_RQ);
-	if (!queue->sched_shared_tags)
-		return -ENOMEM;
-
-	blk_mq_tag_update_sched_shared_tags(queue);
-
-	return 0;
+		q->sched_shared_tags = NULL;
 }
 
 void blk_mq_sched_reg_debugfs(struct request_queue *q)
@@ -458,8 +411,75 @@ void blk_mq_sched_unreg_debugfs(struct request_queue *q)
 	mutex_unlock(&q->debugfs_mutex);
 }
 
+void blk_mq_free_sched_tags(struct elevator_tags *et,
+		struct blk_mq_tag_set *set)
+{
+	unsigned long i;
+
+	/* Shared tags are stored at index 0 in @tags. */
+	if (blk_mq_is_shared_tags(set->flags))
+		blk_mq_free_map_and_rqs(set, et->tags[0], BLK_MQ_NO_HCTX_IDX);
+	else {
+		for (i = 0; i < et->nr_hw_queues; i++)
+			blk_mq_free_map_and_rqs(set, et->tags[i], i);
+	}
+
+	kfree(et);
+}
+
+struct elevator_tags *blk_mq_alloc_sched_tags(struct blk_mq_tag_set *set,
+		unsigned int nr_hw_queues)
+{
+	unsigned int nr_tags;
+	int i;
+	struct elevator_tags *et;
+	gfp_t gfp = GFP_NOIO | __GFP_ZERO | __GFP_NOWARN | __GFP_NORETRY;
+
+	if (blk_mq_is_shared_tags(set->flags))
+		nr_tags = 1;
+	else
+		nr_tags = nr_hw_queues;
+
+	et = kmalloc(sizeof(struct elevator_tags) +
+			nr_tags * sizeof(struct blk_mq_tags *), gfp);
+	if (!et)
+		return NULL;
+	/*
+	 * Default to double of smaller one between hw queue_depth and
+	 * 128, since we don't split into sync/async like the old code
+	 * did. Additionally, this is a per-hw queue depth.
+	 */
+	et->nr_requests = 2 * min_t(unsigned int, set->queue_depth,
+			BLKDEV_DEFAULT_RQ);
+	et->nr_hw_queues = nr_hw_queues;
+
+	if (blk_mq_is_shared_tags(set->flags)) {
+		/* Shared tags are stored at index 0 in @tags. */
+		et->tags[0] = blk_mq_alloc_map_and_rqs(set, BLK_MQ_NO_HCTX_IDX,
+					MAX_SCHED_RQ);
+		if (!et->tags[0])
+			goto out;
+	} else {
+		for (i = 0; i < et->nr_hw_queues; i++) {
+			et->tags[i] = blk_mq_alloc_map_and_rqs(set, i,
+					et->nr_requests);
+			if (!et->tags[i])
+				goto out_unwind;
+		}
+	}
+
+	return et;
+out_unwind:
+	while (--i >= 0)
+		blk_mq_free_map_and_rqs(set, et->tags[i], i);
+out:
+	kfree(et);
+	return NULL;
+}
+
 /* caller must have a reference to @e, will grab another one if successful */
-int blk_mq_init_sched(struct request_queue *q, struct elevator_type *e)
+int blk_mq_init_sched(struct request_queue *q, struct elevator_type *e,
+		struct elevator_tags *et)
 {
 	unsigned int flags = q->tag_set->flags;
 	struct blk_mq_hw_ctx *hctx;
@@ -467,40 +487,33 @@ int blk_mq_init_sched(struct request_queue *q, struct elevator_type *e)
 	unsigned long i;
 	int ret;
 
-	/*
-	 * Default to double of smaller one between hw queue_depth and 128,
-	 * since we don't split into sync/async like the old code did.
-	 * Additionally, this is a per-hw queue depth.
-	 */
-	q->nr_requests = 2 * min_t(unsigned int, q->tag_set->queue_depth,
-				   BLKDEV_DEFAULT_RQ);
-
-	eq = elevator_alloc(q, e);
+	eq = elevator_alloc(q, e, et);
 	if (!eq)
 		return -ENOMEM;
 
+	q->nr_requests = et->nr_requests;
+
 	if (blk_mq_is_shared_tags(flags)) {
-		ret = blk_mq_init_sched_shared_tags(q);
-		if (ret)
-			goto err_put_elevator;
+		/* Shared tags are stored at index 0 in @et->tags. */
+		q->sched_shared_tags = et->tags[0];
+		blk_mq_tag_update_sched_shared_tags(q);
 	}
 
 	queue_for_each_hw_ctx(q, hctx, i) {
-		ret = blk_mq_sched_alloc_map_and_rqs(q, hctx, i);
-		if (ret)
-			goto err_free_map_and_rqs;
+		if (blk_mq_is_shared_tags(flags))
+			hctx->sched_tags = q->sched_shared_tags;
+		else
+			hctx->sched_tags = et->tags[i];
 	}
 
 	ret = e->ops.init_sched(q, eq);
 	if (ret)
-		goto err_free_map_and_rqs;
+		goto out;
 
 	queue_for_each_hw_ctx(q, hctx, i) {
 		if (e->ops.init_hctx) {
 			ret = e->ops.init_hctx(hctx, i);
 			if (ret) {
-				eq = q->elevator;
-				blk_mq_sched_free_rqs(q);
 				blk_mq_exit_sched(q, eq);
 				kobject_put(&eq->kobj);
 				return ret;
@@ -509,10 +522,8 @@ int blk_mq_init_sched(struct request_queue *q, struct elevator_type *e)
 	}
 	return 0;
 
-err_free_map_and_rqs:
-	blk_mq_sched_free_rqs(q);
+out:
 	blk_mq_sched_tags_teardown(q, flags);
-err_put_elevator:
 	kobject_put(&eq->kobj);
 	q->elevator = NULL;
 	return ret;
block/blk-mq-sched.h +7 -1
index 1326526bb733..0cde00cd1c47 100644
@@ -18,10 +18,16 @@ void __blk_mq_sched_restart(struct blk_mq_hw_ctx *hctx);
 
 void blk_mq_sched_dispatch_requests(struct blk_mq_hw_ctx *hctx);
 
-int blk_mq_init_sched(struct request_queue *q, struct elevator_type *e);
+int blk_mq_init_sched(struct request_queue *q, struct elevator_type *e,
+		struct elevator_tags *et);
 void blk_mq_exit_sched(struct request_queue *q, struct elevator_queue *e);
 void blk_mq_sched_free_rqs(struct request_queue *q);
 
+struct elevator_tags *blk_mq_alloc_sched_tags(struct blk_mq_tag_set *set,
+		unsigned int nr_hw_queues);
+void blk_mq_free_sched_tags(struct elevator_tags *et,
+		struct blk_mq_tag_set *set);
+
 static inline void blk_mq_sched_restart(struct blk_mq_hw_ctx *hctx)
 {
 	if (test_bit(BLK_MQ_S_SCHED_RESTART, &hctx->state))
block/elevator.c +42 -5

121 lines not on this page · raw message

block/elevator.h +12 -2
index a4de5f9ad790..adc5c157e17e 100644
@@ -23,6 +23,15 @@ enum elv_merge {
 struct blk_mq_alloc_data;
 struct blk_mq_hw_ctx;
 
+struct elevator_tags {
+	/* num. of hardware queues for which tags are allocated */
+	unsigned int nr_hw_queues;
+	/* depth used while allocating tags */
+	unsigned int nr_requests;
+	/* shared tag is stored at index 0 */
+	struct blk_mq_tags *tags[];
+};
+
 struct elevator_mq_ops {
 	int (*init_sched)(struct request_queue *, struct elevator_queue *);
 	void (*exit_sched)(struct elevator_queue *);
@@ -113,6 +122,7 @@ struct request *elv_rqhash_find(struct request_queue *q, sector_t offset);
 struct elevator_queue
 {
 	struct elevator_type *type;
+	struct elevator_tags *et;
 	void *elevator_data;
 	struct kobject kobj;
 	struct mutex sysfs_lock;
@@ -152,8 +162,8 @@ ssize_t elv_iosched_show(struct gendisk *disk, char *page);
 ssize_t elv_iosched_store(struct gendisk *disk, const char *page, size_t count);
 
 extern bool elv_bio_merge_ok(struct request *, struct bio *);
-extern struct elevator_queue *elevator_alloc(struct request_queue *,
-					struct elevator_type *);
+struct elevator_queue *elevator_alloc(struct request_queue *,
+		struct elevator_type *, struct elevator_tags *);
 
 /*
  * Helper functions.
Signature · 1 line
2.50.0
#9 lore raw

PATCH v6 3/3 block: fix potential deadlock while running nr_hw_queue update

In reply to Nilay Shroff

Move scheduler tags (sched_tags) allocation and deallocation outside both the ->elevator_lock and ->freeze_lock when updating nr_hw_queues. This change breaks the dependency chain from the percpu allocator lock to the elevator lock, helping to prevent potential deadlocks, as observed in the reported lockdep splat[1].

This commit introduces batch allocation and deallocation helpers for sched_tags, which are now used from within __blk_mq_update_nr_hw_queues routine while iterating through the tagset.

With this change, all sched_tags memory management is handled entirely outside the ->elevator_lock and the ->freeze_lock context, thereby eliminating the lock dependency that could otherwise manifest during nr_hw_queues updates.

[1] https://lore.kernel.org/all/[email protected]/

---
 block/blk-mq-sched.c | 63 ++++++++++++++++++++++++++++++++++++++++++++
 block/blk-mq-sched.h |  4 +++
 block/blk-mq.c       | 11 +++++++-
 block/blk.h          |  2 +-
 block/elevator.c     |  4 +--
 5 files changed, 80 insertions(+), 4 deletions(-)
block/blk-mq-sched.c +63 -0
index 2d6d1ebdd8fb..da802df34a8c 100644
@@ -427,6 +427,30 @@ void blk_mq_free_sched_tags(struct elevator_tags *et,
 	kfree(et);
 }
 
+void blk_mq_free_sched_tags_batch(struct xarray *et_table,
+		struct blk_mq_tag_set *set)
+{
+	struct request_queue *q;
+	struct elevator_tags *et;
+
+	lockdep_assert_held_write(&set->update_nr_hwq_lock);
+
+	list_for_each_entry(q, &set->tag_list, tag_set_list) {
+		/*
+		 * Accessing q->elevator without holding q->elevator_lock is
+		 * safe because we're holding here set->update_nr_hwq_lock in
+		 * the writer context. So, scheduler update/switch code (which
+		 * acquires the same lock but in the reader context) can't run
+		 * concurrently.
+		 */
+		if (q->elevator) {
+			et = xa_load(et_table, q->id);
+			if (et)
+				blk_mq_free_sched_tags(et, set);
+		}
+	}
+}
+
 struct elevator_tags *blk_mq_alloc_sched_tags(struct blk_mq_tag_set *set,
 		unsigned int nr_hw_queues)
 {
@@ -477,6 +501,45 @@ struct elevator_tags *blk_mq_alloc_sched_tags(struct blk_mq_tag_set *set,
 	return NULL;
 }
 
+int blk_mq_alloc_sched_tags_batch(struct xarray *et_table,
+		struct blk_mq_tag_set *set, unsigned int nr_hw_queues)
+{
+	struct request_queue *q;
+	struct elevator_tags *et;
+	gfp_t gfp = GFP_NOIO | __GFP_ZERO | __GFP_NOWARN | __GFP_NORETRY;
+
+	lockdep_assert_held_write(&set->update_nr_hwq_lock);
+
+	list_for_each_entry(q, &set->tag_list, tag_set_list) {
+		/*
+		 * Accessing q->elevator without holding q->elevator_lock is
+		 * safe because we're holding here set->update_nr_hwq_lock in
+		 * the writer context. So, scheduler update/switch code (which
+		 * acquires the same lock but in the reader context) can't run
+		 * concurrently.
+		 */
+		if (q->elevator) {
+			et = blk_mq_alloc_sched_tags(set, nr_hw_queues);
+			if (!et)
+				goto out_unwind;
+			if (xa_insert(et_table, q->id, et, gfp))
+				goto out_free_tags;
+		}
+	}
+	return 0;
+out_free_tags:
+	blk_mq_free_sched_tags(et, set);
+out_unwind:
+	list_for_each_entry_continue_reverse(q, &set->tag_list, tag_set_list) {
+		if (q->elevator) {
+			et = xa_load(et_table, q->id);
+			if (et)
+				blk_mq_free_sched_tags(et, set);
+		}
+	}
+	return -ENOMEM;
+}
+
 /* caller must have a reference to @e, will grab another one if successful */
 int blk_mq_init_sched(struct request_queue *q, struct elevator_type *e,
 		struct elevator_tags *et)
block/blk-mq-sched.h +4 -0
index 0cde00cd1c47..b554e1d55950 100644
@@ -25,8 +25,12 @@ void blk_mq_sched_free_rqs(struct request_queue *q);
 
 struct elevator_tags *blk_mq_alloc_sched_tags(struct blk_mq_tag_set *set,
 		unsigned int nr_hw_queues);
+int blk_mq_alloc_sched_tags_batch(struct xarray *et_table,
+		struct blk_mq_tag_set *set, unsigned int nr_hw_queues);
 void blk_mq_free_sched_tags(struct elevator_tags *et,
 		struct blk_mq_tag_set *set);
+void blk_mq_free_sched_tags_batch(struct xarray *et_table,
+		struct blk_mq_tag_set *set);
 
 static inline void blk_mq_sched_restart(struct blk_mq_hw_ctx *hctx)
 {
block/blk-mq.c +10 -1
index 4806b867e37d..a68b658ce07b 100644
@@ -4972,6 +4972,7 @@ static void __blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set,
 	struct request_queue *q;
 	int prev_nr_hw_queues = set->nr_hw_queues;
 	unsigned int memflags;
+	struct xarray et_table;
 	int i;
 
 	lockdep_assert_held(&set->tag_list_lock);
@@ -4984,6 +4985,11 @@ static void __blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set,
 		return;
 
 	memflags = memalloc_noio_save();
+
+	xa_init(&et_table);
+	if (blk_mq_alloc_sched_tags_batch(&et_table, set, nr_hw_queues) < 0)
+		goto out_memalloc_restore;
+
 	list_for_each_entry(q, &set->tag_list, tag_set_list) {
 		blk_mq_debugfs_unregister_hctxs(q);
 		blk_mq_sysfs_unregister_hctxs(q);
@@ -4995,6 +5001,7 @@ static void __blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set,
 	if (blk_mq_realloc_tag_set_tags(set, nr_hw_queues) < 0) {
 		list_for_each_entry(q, &set->tag_list, tag_set_list)
 			blk_mq_unfreeze_queue_nomemrestore(q);
+		blk_mq_free_sched_tags_batch(&et_table, set);
 		goto reregister;
 	}
 
@@ -5019,7 +5026,7 @@ static void __blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set,
 
 	/* elv_update_nr_hw_queues() unfreeze queue for us */
 	list_for_each_entry(q, &set->tag_list, tag_set_list)
-		elv_update_nr_hw_queues(q);
+		elv_update_nr_hw_queues(q, &et_table);
 
 reregister:
 	list_for_each_entry(q, &set->tag_list, tag_set_list) {
@@ -5029,7 +5036,9 @@ static void __blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set,
 		blk_mq_remove_hw_queues_cpuhp(q);
 		blk_mq_add_hw_queues_cpuhp(q);
 	}
+out_memalloc_restore:
 	memalloc_noio_restore(memflags);
+	xa_destroy(&et_table);
 
 	/* Free the excess tags when nr_hw_queues shrink. */
 	for (i = set->nr_hw_queues; i < prev_nr_hw_queues; i++)
block/blk.h +1 -1
index 37ec459fe656..c6d1d1458388 100644
@@ -321,7 +321,7 @@ bool blk_bio_list_merge(struct request_queue *q, struct list_head *list,
 
 bool blk_insert_flush(struct request *rq);
 
-void elv_update_nr_hw_queues(struct request_queue *q);
+void elv_update_nr_hw_queues(struct request_queue *q, struct xarray *et_table);
 void elevator_set_default(struct request_queue *q);
 void elevator_set_none(struct request_queue *q);
 
block/elevator.c +2 -2
index 50f4b78efe66..8ba8b869d5a4 100644
@@ -705,7 +705,7 @@ static int elevator_change(struct request_queue *q, struct elv_change_ctx *ctx)
  * The I/O scheduler depends on the number of hardware queues, this forces a
  * reattachment when nr_hw_queues changes.
  */
-void elv_update_nr_hw_queues(struct request_queue *q)
+void elv_update_nr_hw_queues(struct request_queue *q, struct xarray *et_table)
 {
 	struct blk_mq_tag_set *set = q->tag_set;
 	struct elv_change_ctx ctx = {};
@@ -720,7 +720,7 @@ void elv_update_nr_hw_queues(struct request_queue *q)
 	 * acquires same lock in the reader context) can't run concurrently.
 	 */
 	if (q->elevator) {
-		ctx.et = blk_mq_alloc_sched_tags(set, set->nr_hw_queues);
+		ctx.et = xa_load(et_table, q->id);
 		if (!ctx.et) {
 			WARN_ON_ONCE(1);
 			return;
Signature · 1 line
2.50.0
Reviewed-by on v6 1/3
#3 lore raw

Re: [PATCHv6 1/3] block: move elevator queue allocation logic into blk_mq_init_sched

In reply to Nilay Shroff
Quoted 12 lines · Nilay Shroff

In preparation for allocating sched_tags before freezing the request queue and acquiring ->elevator_lock, move the elevator queue allocation logic from the elevator ops ->init_sched callback into blk_mq_init_sched. As elevator_alloc is now only invoked from block layer core, we don't need to export it, so unexport elevator_alloc function.

This refactoring provides a centralized location for elevator queue initialization, which makes it easier to store pre-allocated sched_tags in the struct elevator_queue during later changes.

Quoted 9 lines
---
  block/bfq-iosched.c   | 13 +++----------
  block/blk-mq-sched.c  | 11 ++++++++---
  block/elevator.c      |  1 -
  block/elevator.h      |  2 +-
  block/kyber-iosched.c | 11 ++---------
  block/mq-deadline.c   | 14 ++------------
  6 files changed, 16 insertions(+), 36 deletions(-)

Cheers,

Hannes

Signature · 4 lines
Dr. Hannes Reinecke                  Kernel Storage Architect
[email protected]                                +49 911 74053 688
SUSE Software Solutions GmbH, Frankenstr. 146, 90461 Nürnberg
HRB 36809 (AG Nürnberg), GF: I. Totev, A. McDonald, W. Knoblich
#6 lore raw

Re: [PATCHv6 2/3] block: fix lockdep warning caused by lock dependency in elv_iosched_store

In reply to Nilay Shroff
Quoted 171 lines · Nilay Shroff

Recent lockdep reports [1] have revealed a potential deadlock caused by a lock dependency between the percpu allocator lock and the elevator lock. This issue can be avoided by ensuring that the allocation and release of scheduler tags (sched_tags) are performed outside the elevator lock. Furthermore, the queue does not need to be remain frozen during these operations.

To address this, move all sched_tags allocations and deallocations outside of both the ->elevator_lock and the ->freeze_lock. Since the lifetime of the elevator queue and its associated sched_tags is closely tied, the allocated sched_tags are now stored in the elevator queue structure. Then, during the actual elevator switch (which runs under ->freeze_lock and ->elevator_lock), the pre-allocated sched_tags are assigned to the appropriate q->hctx. Once the elevator switch is complete and the locks are released, the old elevator queue and its associated sched_tags are freed.

This commit specifically addresses the allocation/deallocation of sched_ tags during elevator switching. Note that sched_tags may also be allocated in other contexts, such as during nr_hw_queues updates. Supporting that use case will require batch allocation/deallocation, which will be handled in a follow-up patch.

This restructuring ensures that sched_tags memory management occurs entirely outside of the ->elevator_lock and ->freeze_lock context, eliminating the lock dependency problem seen during scheduler updates.

[1] https://lore.kernel.org/all/[email protected]/

---
  block/blk-mq-sched.c | 155 +++++++++++++++++++++++--------------------
  block/blk-mq-sched.h |   8 ++-
  block/elevator.c     |  47 +++++++++++--
  block/elevator.h     |  14 +++-
  4 files changed, 144 insertions(+), 80 deletions(-)
block/blk-mq-sched.c +59 -50
index 359e0704e09b..2d6d1ebdd8fb 100644
@@ -374,64 +374,17 @@ bool blk_mq_sched_try_insert_merge(struct request_queue *q, struct request *rq,
  }
  EXPORT_SYMBOL_GPL(blk_mq_sched_try_insert_merge);
 
-static int blk_mq_sched_alloc_map_and_rqs(struct request_queue *q,
-					  struct blk_mq_hw_ctx *hctx,
-					  unsigned int hctx_idx)
-{
-	if (blk_mq_is_shared_tags(q->tag_set->flags)) {
-		hctx->sched_tags = q->sched_shared_tags;
-		return 0;
-	}
-
-	hctx->sched_tags = blk_mq_alloc_map_and_rqs(q->tag_set, hctx_idx,
-						    q->nr_requests);
-
-	if (!hctx->sched_tags)
-		return -ENOMEM;
-	return 0;
-}
-
-static void blk_mq_exit_sched_shared_tags(struct request_queue *queue)
-{
-	blk_mq_free_rq_map(queue->sched_shared_tags);
-	queue->sched_shared_tags = NULL;
-}
-
  /* called in queue's release handler, tagset has gone away */
  static void blk_mq_sched_tags_teardown(struct request_queue *q, unsigned int flags)
  {
  	struct blk_mq_hw_ctx *hctx;
  	unsigned long i;
 
-	queue_for_each_hw_ctx(q, hctx, i) {
-		if (hctx->sched_tags) {
-			if (!blk_mq_is_shared_tags(flags))
-				blk_mq_free_rq_map(hctx->sched_tags);
-			hctx->sched_tags = NULL;
-		}
-	}
+	queue_for_each_hw_ctx(q, hctx, i)
+		hctx->sched_tags = NULL;
 
  	if (blk_mq_is_shared_tags(flags))
-		blk_mq_exit_sched_shared_tags(q);
-}
-
-static int blk_mq_init_sched_shared_tags(struct request_queue *queue)
-{
-	struct blk_mq_tag_set *set = queue->tag_set;
-
-	/*
-	 * Set initial depth at max so that we don't need to reallocate for
-	 * updating nr_requests.
-	 */
-	queue->sched_shared_tags = blk_mq_alloc_map_and_rqs(set,
-						BLK_MQ_NO_HCTX_IDX,
-						MAX_SCHED_RQ);
-	if (!queue->sched_shared_tags)
-		return -ENOMEM;
-
-	blk_mq_tag_update_sched_shared_tags(queue);
-
-	return 0;
+		q->sched_shared_tags = NULL;
  }
 
  void blk_mq_sched_reg_debugfs(struct request_queue *q)
@@ -458,8 +411,75 @@ void blk_mq_sched_unreg_debugfs(struct request_queue *q)
  	mutex_unlock(&q->debugfs_mutex);
  }
 
+void blk_mq_free_sched_tags(struct elevator_tags *et,
+		struct blk_mq_tag_set *set)
+{
+	unsigned long i;
+
+	/* Shared tags are stored at index 0 in @tags. */
+	if (blk_mq_is_shared_tags(set->flags))
+		blk_mq_free_map_and_rqs(set, et->tags[0], BLK_MQ_NO_HCTX_IDX);
+	else {
+		for (i = 0; i < et->nr_hw_queues; i++)
+			blk_mq_free_map_and_rqs(set, et->tags[i], i);
+	}
+
+	kfree(et);
+}
+
+struct elevator_tags *blk_mq_alloc_sched_tags(struct blk_mq_tag_set *set,
+		unsigned int nr_hw_queues)
+{
+	unsigned int nr_tags;
+	int i;
+	struct elevator_tags *et;
+	gfp_t gfp = GFP_NOIO | __GFP_ZERO | __GFP_NOWARN | __GFP_NORETRY;
+
+	if (blk_mq_is_shared_tags(set->flags))
+		nr_tags = 1;
+	else
+		nr_tags = nr_hw_queues;
+
+	et = kmalloc(sizeof(struct elevator_tags) +
+			nr_tags * sizeof(struct blk_mq_tags *), gfp);
+	if (!et)
+		return NULL;
+	/*
+	 * Default to double of smaller one between hw queue_depth and
+	 * 128, since we don't split into sync/async like the old code
+	 * did. Additionally, this is a per-hw queue depth.
+	 */
+	et->nr_requests = 2 * min_t(unsigned int, set->queue_depth,
+			BLKDEV_DEFAULT_RQ);
+	et->nr_hw_queues = nr_hw_queues;
+
+	if (blk_mq_is_shared_tags(set->flags)) {
+		/* Shared tags are stored at index 0 in @tags. */
+		et->tags[0] = blk_mq_alloc_map_and_rqs(set, BLK_MQ_NO_HCTX_IDX,
+					MAX_SCHED_RQ);
+		if (!et->tags[0])
+			goto out;
+	} else {
+		for (i = 0; i < et->nr_hw_queues; i++) {
+			et->tags[i] = blk_mq_alloc_map_and_rqs(set, i,
+					et->nr_requests);
+			if (!et->tags[i])
+				goto out_unwind;
+		}
+	}

block/blk-mq-sched.c void blk_mq_sched_unreg_debugfs(struct request_queue *q) · line 479

Quoted 10 lines
block/blk-mq-sched.c +10 -0
+
+	return et;
+out_unwind:
+	while (--i >= 0)
+		blk_mq_free_map_and_rqs(set, et->tags[i], i);
+out:
+	kfree(et);
+	return NULL;
+}
+

As smatch stated, the unwind pattern is a bit odd. Maybe move the unwind into the 'else' branch, and us a conditional to invoke it:

if (i < et->nr_hw_queues)
while (--i >= 0)
blk_mq_free_map_and_request()

Cheers,

Hannes

Signature · 4 lines
Dr. Hannes Reinecke                  Kernel Storage Architect
[email protected]                                +49 911 74053 688
SUSE Software Solutions GmbH, Frankenstr. 146, 90461 Nürnberg
HRB 36809 (AG Nürnberg), GF: I. Totev, A. McDonald, W. Knoblich
#11 lore raw

Re: [PATCHv6 3/3] block: fix potential deadlock while running nr_hw_queue update

In reply to Nilay Shroff
Quoted 50 lines · Nilay Shroff

Move scheduler tags (sched_tags) allocation and deallocation outside both the ->elevator_lock and ->freeze_lock when updating nr_hw_queues. This change breaks the dependency chain from the percpu allocator lock to the elevator lock, helping to prevent potential deadlocks, as observed in the reported lockdep splat[1].

This commit introduces batch allocation and deallocation helpers for sched_tags, which are now used from within __blk_mq_update_nr_hw_queues routine while iterating through the tagset.

With this change, all sched_tags memory management is handled entirely outside the ->elevator_lock and the ->freeze_lock context, thereby eliminating the lock dependency that could otherwise manifest during nr_hw_queues updates.

[1] https://lore.kernel.org/all/[email protected]/

---
  block/blk-mq-sched.c | 63 ++++++++++++++++++++++++++++++++++++++++++++
  block/blk-mq-sched.h |  4 +++
  block/blk-mq.c       | 11 +++++++-
  block/blk.h          |  2 +-
  block/elevator.c     |  4 +--
  5 files changed, 80 insertions(+), 4 deletions(-)
block/blk-mq-sched.c +14 -0
index 2d6d1ebdd8fb..da802df34a8c 100644
@@ -427,6 +427,30 @@ void blk_mq_free_sched_tags(struct elevator_tags *et,
  	kfree(et);
  }
 
+void blk_mq_free_sched_tags_batch(struct xarray *et_table,
+		struct blk_mq_tag_set *set)
+{
+	struct request_queue *q;
+	struct elevator_tags *et;
+
+	lockdep_assert_held_write(&set->update_nr_hwq_lock);
+
+	list_for_each_entry(q, &set->tag_list, tag_set_list) {
+		/*
+		 * Accessing q->elevator without holding q->elevator_lock is
+		 * safe because we're holding here set->update_nr_hwq_lock in
+		 * the writer context. So, scheduler update/switch code (which
+		 * acquires the same lock but in the reader context) can't run

block/blk-mq-sched.c void blk_mq_free_sched_tags(struct elevator_tags *et, · line 453

Quoted 10 lines
block/blk-mq-sched.c +10 -0
+		 * concurrently.
+		 */
+		if (q->elevator) {
+			et = xa_load(et_table, q->id);
+			if (et)
+				blk_mq_free_sched_tags(et, set);
+		}
+	}
+}
+

Odd. Do we allow empty sched_tags?
Why isn't this an error (or at least a warning)?

Cheers,

Hannes

Signature · 4 lines
Dr. Hannes Reinecke                  Kernel Storage Architect
[email protected]                                +49 911 74053 688
SUSE Software Solutions GmbH, Frankenstr. 146, 90461 Nürnberg
HRB 36809 (AG Nürnberg), GF: I. Totev, A. McDonald, W. Knoblich
#7 lore raw

Re: [PATCHv6 2/3] block: fix lockdep warning caused by lock dependency in elv_iosched_store

In reply to Hannes Reinecke
Quoted 55 lines · Hannes Reinecke
Quoted 40 lines
block/blk-mq-sched.c +40 -0
+struct elevator_tags *blk_mq_alloc_sched_tags(struct blk_mq_tag_set *set,
+        unsigned int nr_hw_queues)
+{
+    unsigned int nr_tags;
+    int i;
+    struct elevator_tags *et;
+    gfp_t gfp = GFP_NOIO | __GFP_ZERO | __GFP_NOWARN | __GFP_NORETRY;
+
+    if (blk_mq_is_shared_tags(set->flags))
+        nr_tags = 1;
+    else
+        nr_tags = nr_hw_queues;
+
+    et = kmalloc(sizeof(struct elevator_tags) +
+            nr_tags * sizeof(struct blk_mq_tags *), gfp);
+    if (!et)
+        return NULL;
+    /*
+     * Default to double of smaller one between hw queue_depth and
+     * 128, since we don't split into sync/async like the old code
+     * did. Additionally, this is a per-hw queue depth.
+     */
+    et->nr_requests = 2 * min_t(unsigned int, set->queue_depth,
+            BLKDEV_DEFAULT_RQ);
+    et->nr_hw_queues = nr_hw_queues;
+
+    if (blk_mq_is_shared_tags(set->flags)) {
+        /* Shared tags are stored at index 0 in @tags. */
+        et->tags[0] = blk_mq_alloc_map_and_rqs(set, BLK_MQ_NO_HCTX_IDX,
+                    MAX_SCHED_RQ);
+        if (!et->tags[0])
+            goto out;
+    } else {
+        for (i = 0; i < et->nr_hw_queues; i++) {
+            et->tags[i] = blk_mq_alloc_map_and_rqs(set, i,
+                    et->nr_requests);
+            if (!et->tags[i])
+                goto out_unwind;
+        }
+    }

block/blk-mq-sched.c void blk_mq_sched_unreg_debugfs(struct request_queue *q) · line 479

Quoted 10 lines
block/blk-mq-sched.c +10 -0
+
+    return et;
+out_unwind:
+    while (--i >= 0)
+        blk_mq_free_map_and_rqs(set, et->tags[i], i);
+out:
+    kfree(et);
+    return NULL;
+}
+

As smatch stated, the unwind pattern is a bit odd. Maybe move the unwind into the 'else' branch, and us a conditional to invoke it:

Quoted 4 lines

if (i < et->nr_hw_queues)
  while (--i >= 0)
    blk_mq_free_map_and_request()

I believe the 'if (i < et->nr_hw_queues)' check is unnecessary here. When we jump to the @out_unwind label, @i is always less than @et->nr_hw_queues because the for loop exits early (on allocation failure) before reaching the upper bound. If @i had reached @et->nr_hw_queues, the loop would have completed and we wouldn't jump to @out_unwind at all — we’d simply return @et instead.

The Smatch flagged the unwind loop due to the use of an unsigned @i in the previous patch. In that case, if the first allocation (i == 0) fails, then '--i' underflows to UINT_MAX, and the condition '--i >= 0' is always true — hence the warning.

This patch corrects that by declaring @i as a signed int, so that '--i >= 0' behaves as expected and avoids the Smatch warning.

So, I don't think an extra condition like 'if (i < et->nr_hw_queues)' is needed around the unwind loop. Agreed?

Thnaks,
--Nilay

#12 lore raw

Re: [PATCHv6 3/3] block: fix potential deadlock while running nr_hw_queue update

In reply to Hannes Reinecke
Quoted 54 lines · Hannes Reinecke
Quoted 53 lines · Nilay Shroff

Move scheduler tags (sched_tags) allocation and deallocation outside both the ->elevator_lock and ->freeze_lock when updating nr_hw_queues. This change breaks the dependency chain from the percpu allocator lock to the elevator lock, helping to prevent potential deadlocks, as observed in the reported lockdep splat[1].

This commit introduces batch allocation and deallocation helpers for sched_tags, which are now used from within __blk_mq_update_nr_hw_queues routine while iterating through the tagset.

With this change, all sched_tags memory management is handled entirely outside the ->elevator_lock and the ->freeze_lock context, thereby eliminating the lock dependency that could otherwise manifest during nr_hw_queues updates.

[1] https://lore.kernel.org/all/[email protected]/

---
  block/blk-mq-sched.c | 63 ++++++++++++++++++++++++++++++++++++++++++++ block/blk-mq-sched.h |  4 +++
  block/blk-mq.c       | 11 +++++++-
  block/blk.h          |  2 +-
  block/elevator.c     |  4 +--

  5 files changed, 80 insertions(+), 4 deletions(-)
block/blk-mq-sched.c +17 -0
index 2d6d1ebdd8fb..da802df34a8c 100644
@@ -427,6 +427,30 @@ void blk_mq_free_sched_tags(struct elevator_tags *et,
       kfree(et);
   }
   +void blk_mq_free_sched_tags_batch(struct xarray *et_table,
+        struct blk_mq_tag_set *set)
+{
+    struct request_queue *q;
+    struct elevator_tags *et;
+
+    lockdep_assert_held_write(&set->update_nr_hwq_lock);
+
+    list_for_each_entry(q, &set->tag_list, tag_set_list) {
+        /*
+         * Accessing q->elevator without holding q->elevator_lock is
+         * safe because we're holding here set->update_nr_hwq_lock in
+         * the writer context. So, scheduler update/switch code (which
+         * acquires the same lock but in the reader context) can't run
+         * concurrently.
+         */
+        if (q->elevator) {
+            et = xa_load(et_table, q->id);
Quoted 9 lines

block/blk-mq-sched.c void blk_mq_free_sched_tags(struct elevator_tags *et, · line 453

Quoted 6 lines
block/blk-mq-sched.c +6 -0
+            if (et)
+                blk_mq_free_sched_tags(et, set);
+        }
+    }
+}
+

Odd. Do we allow empty sched_tags?
Why isn't this an error (or at least a warning)?

We don't allow empty sched_tags and xa_load is not supposed to return an empty sched_tags. It's just a paranoia guard that we evaluate whether @et is non-NULL. As this function returns nothing we don't return any error code to the caller here. But yes, we may add a warning here if that seems reasonable. I'd add a 'WARN_ON(!et)' and update it in the next patch.

Thanks,
--Nilay

Reviewed-by on v6 2/3
#8 lore raw

Re: [PATCHv6 2/3] block: fix lockdep warning caused by lock dependency in elv_iosched_store

In reply to Nilay Shroff
Quoted 27 lines · Nilay Shroff

Recent lockdep reports [1] have revealed a potential deadlock caused by a lock dependency between the percpu allocator lock and the elevator lock. This issue can be avoided by ensuring that the allocation and release of scheduler tags (sched_tags) are performed outside the elevator lock. Furthermore, the queue does not need to be remain frozen during these operations.

To address this, move all sched_tags allocations and deallocations outside of both the ->elevator_lock and the ->freeze_lock. Since the lifetime of the elevator queue and its associated sched_tags is closely tied, the allocated sched_tags are now stored in the elevator queue structure. Then, during the actual elevator switch (which runs under ->freeze_lock and ->elevator_lock), the pre-allocated sched_tags are assigned to the appropriate q->hctx. Once the elevator switch is complete and the locks are released, the old elevator queue and its associated sched_tags are freed.

This commit specifically addresses the allocation/deallocation of sched_ tags during elevator switching. Note that sched_tags may also be allocated in other contexts, such as during nr_hw_queues updates. Supporting that use case will require batch allocation/deallocation, which will be handled in a follow-up patch.

This restructuring ensures that sched_tags memory management occurs entirely outside of the ->elevator_lock and ->freeze_lock context, eliminating the lock dependency problem seen during scheduler updates.

Quoted 5 lines

[1] https://lore.kernel.org/all/[email protected]/

Thanks,
Ming

Reviewed-by on v6 3/3
#13 lore raw

Re: [PATCHv6 3/3] block: fix potential deadlock while running nr_hw_queue update

In reply to Nilay Shroff
Quoted 195 lines · Nilay Shroff

Move scheduler tags (sched_tags) allocation and deallocation outside both the ->elevator_lock and ->freeze_lock when updating nr_hw_queues. This change breaks the dependency chain from the percpu allocator lock to the elevator lock, helping to prevent potential deadlocks, as observed in the reported lockdep splat[1].

This commit introduces batch allocation and deallocation helpers for sched_tags, which are now used from within __blk_mq_update_nr_hw_queues routine while iterating through the tagset.

With this change, all sched_tags memory management is handled entirely outside the ->elevator_lock and the ->freeze_lock context, thereby eliminating the lock dependency that could otherwise manifest during nr_hw_queues updates.

[1] https://lore.kernel.org/all/[email protected]/

---
 block/blk-mq-sched.c | 63 ++++++++++++++++++++++++++++++++++++++++++++
 block/blk-mq-sched.h |  4 +++
 block/blk-mq.c       | 11 +++++++-
 block/blk.h          |  2 +-
 block/elevator.c     |  4 +--
 5 files changed, 80 insertions(+), 4 deletions(-)
block/blk-mq-sched.c +63 -0
index 2d6d1ebdd8fb..da802df34a8c 100644
@@ -427,6 +427,30 @@ void blk_mq_free_sched_tags(struct elevator_tags *et,
 	kfree(et);
 }
 
+void blk_mq_free_sched_tags_batch(struct xarray *et_table,
+		struct blk_mq_tag_set *set)
+{
+	struct request_queue *q;
+	struct elevator_tags *et;
+
+	lockdep_assert_held_write(&set->update_nr_hwq_lock);
+
+	list_for_each_entry(q, &set->tag_list, tag_set_list) {
+		/*
+		 * Accessing q->elevator without holding q->elevator_lock is
+		 * safe because we're holding here set->update_nr_hwq_lock in
+		 * the writer context. So, scheduler update/switch code (which
+		 * acquires the same lock but in the reader context) can't run
+		 * concurrently.
+		 */
+		if (q->elevator) {
+			et = xa_load(et_table, q->id);
+			if (et)
+				blk_mq_free_sched_tags(et, set);
+		}
+	}
+}
+
 struct elevator_tags *blk_mq_alloc_sched_tags(struct blk_mq_tag_set *set,
 		unsigned int nr_hw_queues)
 {
@@ -477,6 +501,45 @@ struct elevator_tags *blk_mq_alloc_sched_tags(struct blk_mq_tag_set *set,
 	return NULL;
 }
 
+int blk_mq_alloc_sched_tags_batch(struct xarray *et_table,
+		struct blk_mq_tag_set *set, unsigned int nr_hw_queues)
+{
+	struct request_queue *q;
+	struct elevator_tags *et;
+	gfp_t gfp = GFP_NOIO | __GFP_ZERO | __GFP_NOWARN | __GFP_NORETRY;
+
+	lockdep_assert_held_write(&set->update_nr_hwq_lock);
+
+	list_for_each_entry(q, &set->tag_list, tag_set_list) {
+		/*
+		 * Accessing q->elevator without holding q->elevator_lock is
+		 * safe because we're holding here set->update_nr_hwq_lock in
+		 * the writer context. So, scheduler update/switch code (which
+		 * acquires the same lock but in the reader context) can't run
+		 * concurrently.
+		 */
+		if (q->elevator) {
+			et = blk_mq_alloc_sched_tags(set, nr_hw_queues);
+			if (!et)
+				goto out_unwind;
+			if (xa_insert(et_table, q->id, et, gfp))
+				goto out_free_tags;
+		}
+	}
+	return 0;
+out_free_tags:
+	blk_mq_free_sched_tags(et, set);
+out_unwind:
+	list_for_each_entry_continue_reverse(q, &set->tag_list, tag_set_list) {
+		if (q->elevator) {
+			et = xa_load(et_table, q->id);
+			if (et)
+				blk_mq_free_sched_tags(et, set);
+		}
+	}
+	return -ENOMEM;
+}
+
 /* caller must have a reference to @e, will grab another one if successful */
 int blk_mq_init_sched(struct request_queue *q, struct elevator_type *e,
 		struct elevator_tags *et)
block/blk-mq-sched.h +4 -0
index 0cde00cd1c47..b554e1d55950 100644
@@ -25,8 +25,12 @@ void blk_mq_sched_free_rqs(struct request_queue *q);
 
 struct elevator_tags *blk_mq_alloc_sched_tags(struct blk_mq_tag_set *set,
 		unsigned int nr_hw_queues);
+int blk_mq_alloc_sched_tags_batch(struct xarray *et_table,
+		struct blk_mq_tag_set *set, unsigned int nr_hw_queues);
 void blk_mq_free_sched_tags(struct elevator_tags *et,
 		struct blk_mq_tag_set *set);
+void blk_mq_free_sched_tags_batch(struct xarray *et_table,
+		struct blk_mq_tag_set *set);
 
 static inline void blk_mq_sched_restart(struct blk_mq_hw_ctx *hctx)
 {
block/blk-mq.c +10 -1
index 4806b867e37d..a68b658ce07b 100644
@@ -4972,6 +4972,7 @@ static void __blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set,
 	struct request_queue *q;
 	int prev_nr_hw_queues = set->nr_hw_queues;
 	unsigned int memflags;
+	struct xarray et_table;
 	int i;
 
 	lockdep_assert_held(&set->tag_list_lock);
@@ -4984,6 +4985,11 @@ static void __blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set,
 		return;
 
 	memflags = memalloc_noio_save();
+
+	xa_init(&et_table);
+	if (blk_mq_alloc_sched_tags_batch(&et_table, set, nr_hw_queues) < 0)
+		goto out_memalloc_restore;
+
 	list_for_each_entry(q, &set->tag_list, tag_set_list) {
 		blk_mq_debugfs_unregister_hctxs(q);
 		blk_mq_sysfs_unregister_hctxs(q);
@@ -4995,6 +5001,7 @@ static void __blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set,
 	if (blk_mq_realloc_tag_set_tags(set, nr_hw_queues) < 0) {
 		list_for_each_entry(q, &set->tag_list, tag_set_list)
 			blk_mq_unfreeze_queue_nomemrestore(q);
+		blk_mq_free_sched_tags_batch(&et_table, set);
 		goto reregister;
 	}
 
@@ -5019,7 +5026,7 @@ static void __blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set,
 
 	/* elv_update_nr_hw_queues() unfreeze queue for us */
 	list_for_each_entry(q, &set->tag_list, tag_set_list)
-		elv_update_nr_hw_queues(q);
+		elv_update_nr_hw_queues(q, &et_table);
 
 reregister:
 	list_for_each_entry(q, &set->tag_list, tag_set_list) {
@@ -5029,7 +5036,9 @@ static void __blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set,
 		blk_mq_remove_hw_queues_cpuhp(q);
 		blk_mq_add_hw_queues_cpuhp(q);
 	}
+out_memalloc_restore:
 	memalloc_noio_restore(memflags);
+	xa_destroy(&et_table);
 
 	/* Free the excess tags when nr_hw_queues shrink. */
 	for (i = set->nr_hw_queues; i < prev_nr_hw_queues; i++)
block/blk.h +1 -1
index 37ec459fe656..c6d1d1458388 100644
@@ -321,7 +321,7 @@ bool blk_bio_list_merge(struct request_queue *q, struct list_head *list,
 
 bool blk_insert_flush(struct request *rq);
 
-void elv_update_nr_hw_queues(struct request_queue *q);
+void elv_update_nr_hw_queues(struct request_queue *q, struct xarray *et_table);
 void elevator_set_default(struct request_queue *q);
 void elevator_set_none(struct request_queue *q);
 
block/elevator.c +0 -0
index 50f4b78efe66..8ba8b869d5a4 100644

block/elevator.c static int elevator_change(struct request_queue *q, struct elv_change_ctx *ctx) · line 708

Quoted 5 lines
block/elevator.c +1 -1
@@ -705,7 +705,7 @@ static int elevator_change(struct request_queue *q, struct elv_change_ctx *ctx)
  * The I/O scheduler depends on the number of hardware queues, this forces a
  * reattachment when nr_hw_queues changes.
  */
-void elv_update_nr_hw_queues(struct request_queue *q)
+void elv_update_nr_hw_queues(struct request_queue *q, struct xarray *et_table)

et_table isn't necessary to expose to elv_update_nr_hw_queues(), and it is less readable than passing 'struct elevator_tags *' directly, but it can be one followup improvement.

Anyway:

Thanks,
Ming

#14 lore raw

Re: [PATCHv6 3/3] block: fix potential deadlock while running nr_hw_queue update

Quoted 6 lines · Ming Lei
Quoted 6 lines
block/elevator.c +0 -0
index 50f4b78efe66..8ba8b869d5a4 100644
@@ -705,7 +705,7 @@ static int elevator_change(struct request_queue *q, struct elv_change_ctx *ctx)
  * The I/O scheduler depends on the number of hardware queues, this forces a
Quoted 9 lines

block/elevator.c static int elevator_change(struct request_queue *q, struct elv_change_ctx *ctx) · line 708

Quoted 4 lines
block/elevator.c +1 -1
  * reattachment when nr_hw_queues changes.
  */
-void elv_update_nr_hw_queues(struct request_queue *q)
+void elv_update_nr_hw_queues(struct request_queue *q, struct xarray *et_table)

et_table isn't necessary to expose to elv_update_nr_hw_queues(), and it is less readable than passing 'struct elevator_tags *' directly, but it can be one followup improvement.

Yeah, makes sense... I will update this in the subsequent patchset.

Thanks,
--Nilay