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

[Nilay Shroff <nilay@linux.ibm.com>]

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/0659ea8d-a463-47c8-9180-43c719e106eb@linux.ibm.com/

Reported-by: Stefan Haberland <sth@linux.ibm.com>
Closes: https://lore.kernel.org/all/0659ea8d-a463-47c8-9180-43c719e106eb@linux.ibm.com/
Signed-off-by: Nilay Shroff <nilay@linux.ibm.com>
---
 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(-)

diff --git a/block/blk-mq-sched.c b/block/blk-mq-sched.c
index 2d6d1ebdd8fb..da802df34a8c 100644
--- a/block/blk-mq-sched.c
+++ b/block/blk-mq-sched.c
@@ -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)
diff --git a/block/blk-mq-sched.h b/block/blk-mq-sched.h
index 0cde00cd1c47..b554e1d55950 100644
--- a/block/blk-mq-sched.h
+++ b/block/blk-mq-sched.h
@@ -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)
 {
diff --git a/block/blk-mq.c b/block/blk-mq.c
index 4806b867e37d..a68b658ce07b 100644
--- a/block/blk-mq.c
+++ b/block/blk-mq.c
@@ -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++)
diff --git a/block/blk.h b/block/blk.h
index 37ec459fe656..c6d1d1458388 100644
--- a/block/blk.h
+++ b/block/blk.h
@@ -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);
 
diff --git a/block/elevator.c b/block/elevator.c
index 50f4b78efe66..8ba8b869d5a4 100644
--- a/block/elevator.c
+++ b/block/elevator.c
@@ -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;
-- 
2.50.0