block: don't release queue's sysfs lock during switching elevator
PATCH 1/1 block: don't release queue's sysfs lock during switching elevator
cecf5d87ff20 ("block: split .sysfs_lock into two locks") starts to release & acquire sysfs_lock before registering/un-registering elevator queue during switching elevator for avoiding potential deadlock from showing & storing 'queue/iosched' attributes and removing elevator's kobject.
Turns out there isn't such deadlock because 'q->sysfs_lock' isn't required in .show & .store of queue/iosched's attributes, and just elevator's sysfs lock is acquired in elv_iosched_store() and elv_iosched_show(). So it is safe to hold queue's sysfs lock when registering/un-registering elevator queue.
The biggest issue is that commit cecf5d87ff20 assumes that concurrent write on 'queue/scheduler' can't happen. However, this assumption isn't true, because kernfs_fop_write() only guarantees that concurrent write aren't called on the same open file, but the write could be from different open on the file. So we can't release & re-acquire queue's sysfs lock during switching elevator, otherwise use-after-free on elevator could be triggered.
Fixes the issue by not releasing queue's sysfs lock during switching elevator.
--- block/blk-sysfs.c | 13 ++++--------- block/elevator.c | 31 +------------------------------ 2 files changed, 5 insertions(+), 39 deletions(-)
block/blk-sysfs.c +4 -9
@@ -989,13 +989,11 @@ int blk_register_queue(struct gendisk *disk)blk_mq_debugfs_register (q); } -/* -* The flag of QUEUE_FLAG_REGISTERED isn't set yet, so elevator -* switch won't happen at all. -*/ +mutex_lock (&q->sysfs_lock);if (q->elevator) { ret =elv_register_queue (q, false);if (ret) { +mutex_unlock (&q->sysfs_lock);mutex_unlock (&q->sysfs_dir_lock);kobject_del (&q->kobj);blk_trace_remove_sysfs (dev);
@@ -1005,7 +1003,6 @@ int blk_register_queue(struct gendisk *disk) has_elevator = true; } -mutex_lock (&q->sysfs_lock);blk_queue_flag_set (QUEUE_FLAG_REGISTERED , q);wbt_enable_default (q);blk_throtl_register_queue (q);
@@ -1062,12 +1059,10 @@ void blk_unregister_queue(struct gendisk *disk)kobject_del (&q->kobj);blk_trace_remove_sysfs (disk_to_dev (disk)); -/* -* q->kobj has been removed, so it is safe to check if elevator -* exists without holding q->sysfs_lock. -*/ +mutex_lock (&q->sysfs_lock);if (q->elevator)elv_unregister_queue (q); +mutex_unlock (&q->sysfs_lock);mutex_unlock (&q->sysfs_dir_lock);kobject_put (&disk_to_dev (disk)->kobj);
block/elevator.c +1 -30
@@ -503,9 +503,7 @@ int elv_register_queue(struct request_queue *q, bool uevent)if (uevent)kobject_uevent (&e->kobj,KOBJ_ADD ); -mutex_lock (&q->sysfs_lock); e->registered =1 ; -mutex_unlock (&q->sysfs_lock); }return error; }
@@ -523,11 +521,9 @@ void elv_unregister_queue(struct request_queue *q)kobject_uevent (&e->kobj,KOBJ_REMOVE );kobject_del (&e->kobj); -mutex_lock (&q->sysfs_lock); e->registered =0 ;/* Re-enable throttling in case elevator disabled it */ wbt_enable_default (q); -mutex_unlock (&q->sysfs_lock); } }
@@ -590,32 +586,11 @@ int elevator_switch_mq(struct request_queue *q,lockdep_assert_held (&q->sysfs_lock);if (q->elevator) { -if (q->elevator->registered) { -mutex_unlock (&q->sysfs_lock); - -/* -* Concurrent elevator switch can't happen becasue -* sysfs write is always exclusively on same file. -* -* Also the elevator queue won't be freed after -* sysfs_lock is released becasue kobject_del() in -* blk_unregister_queue() waits for completion of -* .store & .show on its attributes. -*/ +if (q->elevator->registered)elv_unregister_queue (q); -mutex_lock (&q->sysfs_lock); - }ioc_clear_queue (q);elevator_exit (q, q->elevator); - -/* -* sysfs_lock may be dropped, so re-check if queue is -* unregistered. If yes, don't switch to new elevator -* any more -*/ -if (!blk_queue_registered (q)) -return 0 ; } ret =blk_mq_init_sched (q, new_e);
@@ -623,11 +598,7 @@ int elevator_switch_mq(struct request_queue *q, goto out;if (new_e) { -mutex_unlock (&q->sysfs_lock); - ret =elv_register_queue (q, true); - -mutex_lock (&q->sysfs_lock);if (ret) {elevator_exit (q, q->elevator); goto out;
Signature · 1 line
2.20.1
block/elevator.c void elv_unregister_queue(struct request_queue *q) · line 528
Quoted 11 lines · Ming Lei
block/elevator.c +0
-2@@ -523,11 +521,9 @@ void elv_unregister_queue(struct request_queue *q)kobject_uevent (&e->kobj,KOBJ_REMOVE );kobject_del (&e->kobj); -mutex_lock (&q->sysfs_lock); e->registered =0 ;/* Re-enable throttling in case elevator disabled it */ wbt_enable_default (q); -mutex_unlock (&q->sysfs_lock); } }
Does this patch cause sysfs_lock to be held around kobject_del(&e->kobj)? Since sysfs_lock is locked from inside elv_attr_show() and elv_attr_store(), does this mean that this patch reintroduces the lock inversion problem that was fixed recently?
Thanks,
Bart.
Quoted 13 lines · Bart Van Assche
block/elevator.c void elv_unregister_queue(struct request_queue *q) · line 527
Quoted 10 lines · Ming Lei
block/elevator.c +0
-2@@ -523,11 +521,9 @@ void elv_unregister_queue(struct request_queue *q)kobject_uevent (&e->kobj,KOBJ_REMOVE );kobject_del (&e->kobj); -mutex_lock (&q->sysfs_lock); e->registered =0 ;/* Re-enable throttling in case elevator disabled it */ wbt_enable_default (q); -mutex_unlock (&q->sysfs_lock); } }Does this patch cause
sysfs_lockto be held aroundkobject_del(&e->kobj)?
Yes.
Quoted 1 line
Since
sysfs_lockis locked from insideelv_attr_show()andelv_attr_store(),
The request queue's sysfs_lock isn't required in elv_attr_show() and elv_attr_store(), and only elevator's sysfs_lock is needed in the two functions.
Quoted 2 lines
does this mean that this patch reintroduces the lock inversion problem that was fixed recently?
No.
The lock inversion issue only existed on kobjects of q->kobj & q->mq_obj, which was fixed already given the queue's sysfs_lock is required in .show/.store callback of these two kobjects' attributes.
thanks,
Ming
Bart Van Assche
Reviewed-byQuoted 23 lines · Ming Lei
Quoted 13 lines · Bart Van Assche
block/elevator.c void elv_unregister_queue(struct request_queue *q) · line 527
Quoted 10 lines · Ming Lei
block/elevator.c +0
-2@@ -523,11 +521,9 @@ void elv_unregister_queue(struct request_queue *q)kobject_uevent (&e->kobj,KOBJ_REMOVE );kobject_del (&e->kobj); -mutex_lock (&q->sysfs_lock); e->registered =0 ;/* Re-enable throttling in case elevator disabled it */ wbt_enable_default (q); -mutex_unlock (&q->sysfs_lock); } }Does this patch cause
sysfs_lockto be held aroundkobject_del(&e->kobj)?Yes.
Quoted 1 line
Since
sysfs_lockis locked from insideelv_attr_show()andelv_attr_store(),The request queue's
sysfs_lockisn't required inelv_attr_show()andelv_attr_store(), and only elevator'ssysfs_lockis needed in the two functions.
Quoted 8 lines
Quoted 2 lines
does this mean that this patch reintroduces the lock inversion problem that was fixed recently?
No.
The lock inversion issue only existed on kobjects of
q->kobj&q->mq_obj, which was fixed already given the queue'ssysfs_lockis required in .show/.store callback of these two kobjects' attributes.
I had confused e->kobj and eq->kobj, hence my comments. After having taken another look at your patch:
Quoted 20 lines · Ming Lei
cecf5d87ff20("block: split .sysfs_lock into two locks") starts to release & acquiresysfs_lockbefore registering/un-registering elevator queue during switching elevator for avoiding potential deadlock from showing & storing 'queue/iosched' attributes and removing elevator's kobject.Turns out there isn't such deadlock because '
q->sysfs_lock' isn't required in .show & .store of queue/iosched's attributes, and just elevator's sysfs lock is acquired inelv_iosched_store()andelv_iosched_show(). So it is safe to hold queue's sysfs lock when registering/un-registering elevator queue.The biggest issue is that commit
cecf5d87ff20assumes that concurrent write on 'queue/scheduler' can't happen. However, this assumption isn't true, becausekernfs_fop_write()only guarantees that concurrent write aren't called on the same open file, but the write could be from different open on the file. So we can't release & re-acquire queue's sysfs lock during switching elevator, otherwise use-after-free on elevator could be triggered.
Quoted 2 lines
Fixes the issue by not releasing queue's sysfs lock during switching elevator.
Applied, thanks Ming.
Signature · 1 line
Jens Axboe