[PATCHv5,6/7] block: protect wbt_lat_usec using q->elevator_lock

Commit Message

Nilay Shroff Feb. 26, 2025, 12:39 p.m. UTC

The wbt latency and state could be updated while initializing the
elevator or exiting the elevator. It could be also updated while
configuring IO latency QoS parameters using cgroup. The elevator
code path is now protected with q->elevator_lock. So we should
protect the access to sysfs attribute wbt_lat_usec using q->elevator
_lock instead of q->sysfs_lock. White we're at it, also protect
ioc_qos_write(), which configures wbt parameters via cgroup, using
q->elevator_lock.

Reviewed-by: Hannes Reinecke <[e-mail]>
Signed-off-by: Nilay Shroff <[e-mail]>
---
 block/blk-iocost.c     |  2 ++
 block/blk-sysfs.c      | 14 ++++++++------
 include/linux/blkdev.h |  4 ++--
 3 files changed, 12 insertions(+), 8 deletions(-)

Patch

diff --git a/block/blk-iocost.c b/block/blk-iocost.c
index 65a1d4427ccf..c68373361301 100644
--- a/block/blk-iocost.c
+++ b/block/blk-iocost.c
@@ -3249,6 +3249,7 @@  static ssize_t ioc_qos_write(struct kernfs_open_file *of, char *input,
 	}
 
 	memflags = blk_mq_freeze_queue(disk->queue);
+	mutex_lock(&disk->queue->elevator_lock);
 	blk_mq_quiesce_queue(disk->queue);
 
 	spin_lock_irq(&ioc->lock);
@@ -3356,6 +3357,7 @@  static ssize_t ioc_qos_write(struct kernfs_open_file *of, char *input,
 	spin_unlock_irq(&ioc->lock);
 
 	blk_mq_unquiesce_queue(disk->queue);
+	mutex_unlock(&disk->queue->elevator_lock);
 	blk_mq_unfreeze_queue(disk->queue, memflags);
 
 	ret = -EINVAL;
diff --git a/block/blk-sysfs.c b/block/blk-sysfs.c
index f1fa57de29ed..46033d640e30 100644
--- a/block/blk-sysfs.c
+++ b/block/blk-sysfs.c
@@ -557,7 +557,7 @@  static ssize_t queue_wb_lat_show(struct gendisk *disk, char *page)
 	ssize_t ret;
 	struct request_queue *q = disk->queue;
 
-	mutex_lock(&q->sysfs_lock);
+	mutex_lock(&q->elevator_lock);
 	if (!wbt_rq_qos(q)) {
 		ret = -EINVAL;
 		goto out;
@@ -570,7 +570,7 @@  static ssize_t queue_wb_lat_show(struct gendisk *disk, char *page)
 
 	ret = sysfs_emit(page, "%llu\n", div_u64(wbt_get_min_lat(q), 1000));
 out:
-	mutex_unlock(&q->sysfs_lock);
+	mutex_unlock(&q->elevator_lock);
 	return ret;
 }
 
@@ -589,8 +589,8 @@  static ssize_t queue_wb_lat_store(struct gendisk *disk, const char *page,
 	if (val < -1)
 		return -EINVAL;
 
-	mutex_lock(&q->sysfs_lock);
 	memflags = blk_mq_freeze_queue(q);
+	mutex_lock(&q->elevator_lock);
 
 	rqos = wbt_rq_qos(q);
 	if (!rqos) {
@@ -619,8 +619,8 @@  static ssize_t queue_wb_lat_store(struct gendisk *disk, const char *page,
 
 	blk_mq_unquiesce_queue(q);
 out:
+	mutex_unlock(&q->elevator_lock);
 	blk_mq_unfreeze_queue(q, memflags);
-	mutex_unlock(&q->sysfs_lock);
 
 	return ret;
 }
@@ -701,7 +701,9 @@  static struct attribute *blk_mq_queue_attrs[] = {
 	 */
 	&elv_iosched_entry.attr,
 	&queue_requests_entry.attr,
-
+#ifdef CONFIG_BLK_WBT
+	&queue_wb_lat_entry.attr,
+#endif
 	/*
 	 * Attributes which don't require locking.
 	 */
@@ -882,10 +884,10 @@  int blk_register_queue(struct gendisk *disk)
 			goto out_crypto_sysfs_unregister;
 		}
 	}
+	wbt_enable_default(disk);
 	mutex_unlock(&q->elevator_lock);
 
 	blk_queue_flag_set(QUEUE_FLAG_REGISTERED, q);
-	wbt_enable_default(disk);
 
 	/* Now everything is ready and send out KOBJ_ADD uevent */
 	kobject_uevent(&disk->queue_kobj, KOBJ_ADD);
diff --git a/include/linux/blkdev.h b/include/linux/blkdev.h
index 3e66ad016a23..0ee3b5c9388e 100644
--- a/include/linux/blkdev.h
+++ b/include/linux/blkdev.h
@@ -563,8 +563,8 @@  struct request_queue {
 	/*
 	 * Protects against I/O scheduler switching, particularly when
 	 * updating q->elevator. Since the elevator update code path may
-	 * also modify q->nr_requests, this lock also protects the sysfs
-	 * attribute nr_requests.
+	 * also modify q->nr_requests and wbt latency, this lock also
+	 * protects the sysfs attributes nr_requests and wbt_lat_usec.
 	 * To ensure proper locking order during an elevator update, first
 	 * freeze the queue, then acquire ->elevator_lock.
 	 */