[PATCH] block: don't acquire .sysfs_lock before removing mq & iosched kobjects

[PATCH] block: don't acquire .sysfs_lock before removing mq & iosched kobjects

[Ming Lei]

The kernfs built-in lock of 'kn->count' is held in sysfs .show/.store
path. Meantime, inside block's .show/.store callback, q->sysfs_lock is
required.

However, when mq & iosched kobjects are removed via
blk_mq_unregister_dev() & elv_unregister_queue(), q->sysfs_lock is held
too. This way causes AB-BA lock because the kernfs built-in lock of
'kn-count' is required inside kobject_del() too, see the lockdep warning[1].

On the other hand, it isn't necessary to acquire q->sysfs_lock for
both blk_mq_unregister_dev() & elv_unregister_queue() because
clearing REGISTERED flag prevents storing to 'queue/scheduler'
from being happened. Also sysfs write(store) is exclusive, so no
necessary to hold the lock for elv_unregister_queue() when it is
called in switching elevator path.

Fixes the issue by not holding the q->sysfs_lock for blk_mq_unregister_dev() &
elv_unregister_queue().

[1] lockdep warning
======================================================
WARNING: possible circular locking dependency detected
5.3.0-rc3-00044-g73277fc75ea0 #1380 Not tainted
------------------------------------------------------
rmmod/777 is trying to acquire lock:
00000000ac50e981 (kn->count#202){++++}, at: kernfs_remove_by_name_ns+0x59/0x72

but task is already holding lock:
00000000fb16ae21 (&q->sysfs_lock){+.+.}, at: blk_unregister_queue+0x78/0x10b

which lock already depends on the new lock.

the existing dependency chain (in reverse order) is:

-> #1 (&q->sysfs_lock){+.+.}:
       __lock_acquire+0x95f/0xa2f
       lock_acquire+0x1b4/0x1e8
       __mutex_lock+0x14a/0xa9b
       blk_mq_hw_sysfs_show+0x63/0xb6
       sysfs_kf_seq_show+0x11f/0x196
       seq_read+0x2cd/0x5f2
       vfs_read+0xc7/0x18c
       ksys_read+0xc4/0x13e
       do_syscall_64+0xa7/0x295
       entry_SYSCALL_64_after_hwframe+0x49/0xbe

-> #0 (kn->count#202){++++}:
       check_prev_add+0x5d2/0xc45
       validate_chain+0xed3/0xf94
       __lock_acquire+0x95f/0xa2f
       lock_acquire+0x1b4/0x1e8
       __kernfs_remove+0x237/0x40b
       kernfs_remove_by_name_ns+0x59/0x72
       remove_files+0x61/0x96
       sysfs_remove_group+0x81/0xa4
       sysfs_remove_groups+0x3b/0x44
       kobject_del+0x44/0x94
       blk_mq_unregister_dev+0x83/0xdd
       blk_unregister_queue+0xa0/0x10b
       del_gendisk+0x259/0x3fa
       null_del_dev+0x8b/0x1c3 [null_blk]
       null_exit+0x5c/0x95 [null_blk]
       __se_sys_delete_module+0x204/0x337
       do_syscall_64+0xa7/0x295
       entry_SYSCALL_64_after_hwframe+0x49/0xbe

other info that might help us debug this:

 Possible unsafe locking scenario:

       CPU0                    CPU1
       ----                    ----
  lock(&q->sysfs_lock);
                               lock(kn->count#202);
                               lock(&q->sysfs_lock);
  lock(kn->count#202);

 *** DEADLOCK ***

2 locks held by rmmod/777:
 #0: 00000000e69bd9de (&lock){+.+.}, at: null_exit+0x2e/0x95 [null_blk]
 #1: 00000000fb16ae21 (&q->sysfs_lock){+.+.}, at: blk_unregister_queue+0x78/0x10b

stack backtrace:
CPU: 0 PID: 777 Comm: rmmod Not tainted 5.3.0-rc3-00044-g73277fc75ea0 #1380
Hardware name: QEMU Standard PC (Q35 + ICH9, 2009), BIOS ?-20180724_192412-buildhw-07.phx4
Call Trace:
 dump_stack+0x9a/0xe6
 check_noncircular+0x207/0x251
 ? print_circular_bug+0x32a/0x32a
 ? find_usage_backwards+0x84/0xb0
 check_prev_add+0x5d2/0xc45
 validate_chain+0xed3/0xf94
 ? check_prev_add+0xc45/0xc45
 ? mark_lock+0x11b/0x804
 ? check_usage_forwards+0x1ca/0x1ca
 __lock_acquire+0x95f/0xa2f
 lock_acquire+0x1b4/0x1e8
 ? kernfs_remove_by_name_ns+0x59/0x72
 __kernfs_remove+0x237/0x40b
 ? kernfs_remove_by_name_ns+0x59/0x72
 ? kernfs_next_descendant_post+0x7d/0x7d
 ? strlen+0x10/0x23
 ? strcmp+0x22/0x44
 kernfs_remove_by_name_ns+0x59/0x72
 remove_files+0x61/0x96
 sysfs_remove_group+0x81/0xa4
 sysfs_remove_groups+0x3b/0x44
 kobject_del+0x44/0x94
 blk_mq_unregister_dev+0x83/0xdd
 blk_unregister_queue+0xa0/0x10b
 del_gendisk+0x259/0x3fa
 ? disk_events_poll_msecs_store+0x12b/0x12b
 ? check_flags+0x1ea/0x204
 ? mark_held_locks+0x1f/0x7a
 null_del_dev+0x8b/0x1c3 [null_blk]
 null_exit+0x5c/0x95 [null_blk]
 __se_sys_delete_module+0x204/0x337
 ? free_module+0x39f/0x39f
 ? blkcg_maybe_throttle_current+0x8a/0x718
 ? rwlock_bug+0x62/0x62
 ? __blkcg_punt_bio_submit+0xd0/0xd0
 ? trace_hardirqs_on_thunk+0x1a/0x20
 ? mark_held_locks+0x1f/0x7a
 ? do_syscall_64+0x4c/0x295
 do_syscall_64+0xa7/0x295
 entry_SYSCALL_64_after_hwframe+0x49/0xbe
RIP: 0033:0x7fb696cdbe6b
Code: 73 01 c3 48 8b 0d 1d 20 0c 00 f7 d8 64 89 01 48 83 c8 ff c3 66 2e 0f 1f 84 00 00 008
RSP: 002b:00007ffec9588788 EFLAGS: 00000206 ORIG_RAX: 00000000000000b0
RAX: ffffffffffffffda RBX: 0000559e589137c0 RCX: 00007fb696cdbe6b
RDX: 000000000000000a RSI: 0000000000000800 RDI: 0000559e58913828
RBP: 0000000000000000 R08: 00007ffec9587701 R09: 0000000000000000
R10: 00007fb696d4eae0 R11: 0000000000000206 R12: 00007ffec95889b0
R13: 00007ffec95896b3 R14: 0000559e58913260 R15: 0000559e589137c0

Cc: Greg KH <gregkh@linuxfoundation.org>
Cc: Mike Snitzer <snitzer@redhat.com>
Cc: Bart Van Assche <bvanassche@acm.org>
Signed-off-by: Ming Lei <ming.lei@redhat.com>
---
 block/blk-mq-sysfs.c |  4 ++--
 block/blk-sysfs.c    |  9 ++++-----
 block/elevator.c     | 16 +++++++++++++---
 3 files changed, 19 insertions(+), 10 deletions(-)

diff --git a/block/blk-mq-sysfs.c b/block/blk-mq-sysfs.c
index e0b97c22726c..d5e805281130 100644
--- a/block/blk-mq-sysfs.c
+++ b/block/blk-mq-sysfs.c
@@ -247,8 +247,6 @@ void blk_mq_unregister_dev(struct device *dev, struct request_queue *q)
 	struct blk_mq_hw_ctx *hctx;
 	int i;
 
-	lockdep_assert_held(&q->sysfs_lock);
-
 	queue_for_each_hw_ctx(q, hctx, i)
 		blk_mq_unregister_hctx(hctx);
 
@@ -256,7 +254,9 @@ void blk_mq_unregister_dev(struct device *dev, struct request_queue *q)
 	kobject_del(q->mq_kobj);
 	kobject_put(&dev->kobj);
 
+	mutex_lock(&q->sysfs_lock);
 	q->mq_sysfs_init_done = false;
+	mutex_unlock(&q->sysfs_lock);
 }
 
 void blk_mq_hctx_kobj_init(struct blk_mq_hw_ctx *hctx)
diff --git a/block/blk-sysfs.c b/block/blk-sysfs.c
index 977c659dcd18..46f033b48917 100644
--- a/block/blk-sysfs.c
+++ b/block/blk-sysfs.c
@@ -1021,6 +1021,7 @@ EXPORT_SYMBOL_GPL(blk_register_queue);
 void blk_unregister_queue(struct gendisk *disk)
 {
 	struct request_queue *q = disk->queue;
+	bool has_elevator;
 
 	if (WARN_ON(!q))
 		return;
@@ -1035,8 +1036,9 @@ void blk_unregister_queue(struct gendisk *disk)
 	 * concurrent elv_iosched_store() calls.
 	 */
 	mutex_lock(&q->sysfs_lock);
-
 	blk_queue_flag_clear(QUEUE_FLAG_REGISTERED, q);
+	has_elevator = q->elevator;
+	mutex_unlock(&q->sysfs_lock);
 
 	/*
 	 * Remove the sysfs attributes before unregistering the queue data
@@ -1044,16 +1046,13 @@ void blk_unregister_queue(struct gendisk *disk)
 	 */
 	if (queue_is_mq(q))
 		blk_mq_unregister_dev(disk_to_dev(disk), q);
-	mutex_unlock(&q->sysfs_lock);
 
 	kobject_uevent(&q->kobj, KOBJ_REMOVE);
 	kobject_del(&q->kobj);
 	blk_trace_remove_sysfs(disk_to_dev(disk));
 
-	mutex_lock(&q->sysfs_lock);
-	if (q->elevator)
+	if (has_elevator)
 		elv_unregister_queue(q);
-	mutex_unlock(&q->sysfs_lock);
 
 	kobject_put(&disk_to_dev(disk)->kobj);
 }
diff --git a/block/elevator.c b/block/elevator.c
index 2f17d66d0e61..4a12c60ba946 100644
--- a/block/elevator.c
+++ b/block/elevator.c
@@ -495,16 +495,17 @@ int elv_register_queue(struct request_queue *q)
 
 void elv_unregister_queue(struct request_queue *q)
 {
-	lockdep_assert_held(&q->sysfs_lock);
-
 	if (q) {
 		struct elevator_queue *e = q->elevator;
 
 		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);
 	}
 }
 
@@ -567,8 +568,17 @@ int elevator_switch_mq(struct request_queue *q,
 	lockdep_assert_held(&q->sysfs_lock);
 
 	if (q->elevator) {
-		if (q->elevator->registered)
+		if (q->elevator->registered) {
+			/*
+			 * sysfs write is exclusively, release
+			 * sysfs_lock for avoiding deadlock with
+			 * sysfs built-in lock which is required
+			 * in either .show or .store path.
+			 */
+			mutex_unlock(&q->sysfs_lock);
 			elv_unregister_queue(q);
+			mutex_lock(&q->sysfs_lock);
+		}
 		ioc_clear_queue(q);
 		elevator_exit(q, q->elevator);
 	}
-- 
2.20.1

Re: [PATCH] block: don't acquire .sysfs_lock before removing mq & iosched kobjects

[Bart Van Assche]

On 8/16/19 6:55 AM, Ming Lei wrote:
> The kernfs built-in lock of 'kn->count' is held in sysfs .show/.store
> path. Meantime, inside block's .show/.store callback, q->sysfs_lock is
> required.
> 
> However, when mq & iosched kobjects are removed via
> blk_mq_unregister_dev() & elv_unregister_queue(), q->sysfs_lock is held
> too. This way causes AB-BA lock because the kernfs built-in lock of
> 'kn-count' is required inside kobject_del() too, see the lockdep warning[1].
> 
> On the other hand, it isn't necessary to acquire q->sysfs_lock for
> both blk_mq_unregister_dev() & elv_unregister_queue() because
> clearing REGISTERED flag prevents storing to 'queue/scheduler'
> from being happened. Also sysfs write(store) is exclusive, so no
> necessary to hold the lock for elv_unregister_queue() when it is
> called in switching elevator path.
> 
> Fixes the issue by not holding the q->sysfs_lock for blk_mq_unregister_dev() &
> elv_unregister_queue().

Have you considered to split sysfs_lock into multiple mutexes? Today it 
is very hard to verify the correctness of block layer code that uses 
sysfs_lock because it has not been documented anywhere what that mutex 
protects. I think that mutex should be split into at least two mutexes: 
one that protects switching I/O schedulers and another one that protects 
hctx->tags and hctx->sched_tags.

Thanks,

Bart.

Re: [PATCH] block: don't acquire .sysfs_lock before removing mq & iosched kobjects

[Ming Lei]

On Fri, Aug 16, 2019 at 08:14:13AM -0700, Bart Van Assche wrote:
> On 8/16/19 6:55 AM, Ming Lei wrote:
> > The kernfs built-in lock of 'kn->count' is held in sysfs .show/.store
> > path. Meantime, inside block's .show/.store callback, q->sysfs_lock is
> > required.
> > 
> > However, when mq & iosched kobjects are removed via
> > blk_mq_unregister_dev() & elv_unregister_queue(), q->sysfs_lock is held
> > too. This way causes AB-BA lock because the kernfs built-in lock of
> > 'kn-count' is required inside kobject_del() too, see the lockdep warning[1].
> > 
> > On the other hand, it isn't necessary to acquire q->sysfs_lock for
> > both blk_mq_unregister_dev() & elv_unregister_queue() because
> > clearing REGISTERED flag prevents storing to 'queue/scheduler'
> > from being happened. Also sysfs write(store) is exclusive, so no
> > necessary to hold the lock for elv_unregister_queue() when it is
> > called in switching elevator path.
> > 
> > Fixes the issue by not holding the q->sysfs_lock for blk_mq_unregister_dev() &
> > elv_unregister_queue().
> 
> Have you considered to split sysfs_lock into multiple mutexes? Today it is

So far, not consider it. At least now, I just don't see the need to hold
sysfs_lock for both blk_mq_unregister_dev() & elv_unregister_queue().
Then we can fix the deadlock issue which can be triggered quite easily,
and the fix should be for -stable.

Yeah, I agree that sysfs_lock has been used too widely.

> very hard to verify the correctness of block layer code that uses sysfs_lock
> because it has not been documented anywhere what that mutex protects. I
> think that mutex should be split into at least two mutexes: one that
> protects switching I/O schedulers and another one that protects hctx->tags
> and hctx->sched_tags.

sysfs_lock is required in any .show & .store callback, and switching I/O
scheduler is done in .store(), then hctx->sched_tags is protected by sysfs_lock
too.

hctx->tags is tagset wide or host-wide, which is protected by set->tag_list_lock.


Thanks,
Ming

Re: [PATCH] block: don't acquire .sysfs_lock before removing mq & iosched kobjects

[Bart Van Assche]

On 8/19/19 1:15 AM, Ming Lei wrote:
> hctx->tags is tagset wide or host-wide, which is protected by set->tag_list_lock.

Isn't the purpose of set->tag_list_lock to protect set->tag_list?

Bart.

Re: [PATCH] block: don't acquire .sysfs_lock before removing mq & iosched kobjects

[Ming Lei]

On Tue, Aug 20, 2019 at 02:07:43PM -0700, Bart Van Assche wrote:
> On 8/19/19 1:15 AM, Ming Lei wrote:
> > hctx->tags is tagset wide or host-wide, which is protected by set->tag_list_lock.
> 
> Isn't the purpose of set->tag_list_lock to protect set->tag_list?

In theory, .tags is shared among all queues in tagset, so it should be
protected by tagset wide lock, what is why set->tag_list_lock is held
in blk_mq_update_nr_hw_queues().

However, both sysfs and debugfs read/write doesn't hold this lock because
the .tags won't be shrunk in blk_mq_update_nr_requests(), and
blk_mq_update_nr_hw_queues un-registers debugfs/sysfs before changing .tags.


thanks,
Ming

Re: [PATCH] block: don't acquire .sysfs_lock before removing mq & iosched kobjects

[Bart Van Assche]

On 8/16/19 6:55 AM, Ming Lei wrote:
> diff --git a/block/blk-sysfs.c b/block/blk-sysfs.c
> index 977c659dcd18..46f033b48917 100644
> --- a/block/blk-sysfs.c
> +++ b/block/blk-sysfs.c
> @@ -1021,6 +1021,7 @@ EXPORT_SYMBOL_GPL(blk_register_queue);
>   void blk_unregister_queue(struct gendisk *disk)
>   {
>   	struct request_queue *q = disk->queue;
> +	bool has_elevator;
>   
>   	if (WARN_ON(!q))
>   		return;
> @@ -1035,8 +1036,9 @@ void blk_unregister_queue(struct gendisk *disk)
>   	 * concurrent elv_iosched_store() calls.
>   	 */
>   	mutex_lock(&q->sysfs_lock);
> -
>   	blk_queue_flag_clear(QUEUE_FLAG_REGISTERED, q);
> +	has_elevator = q->elevator;
> +	mutex_unlock(&q->sysfs_lock);

blk_queue_flag_clear() modifies queue flags atomically so no need to 
hold sysfs_lock around calls of that function.

> @@ -1044,16 +1046,13 @@ void blk_unregister_queue(struct gendisk *disk)
>   	 */
>   	if (queue_is_mq(q))
>   		blk_mq_unregister_dev(disk_to_dev(disk), q);
> -	mutex_unlock(&q->sysfs_lock);
>   
>   	kobject_uevent(&q->kobj, KOBJ_REMOVE);
>   	kobject_del(&q->kobj);
>   	blk_trace_remove_sysfs(disk_to_dev(disk));
>   
> -	mutex_lock(&q->sysfs_lock);
> -	if (q->elevator)
> +	if (has_elevator)
>   		elv_unregister_queue(q);
> -	mutex_unlock(&q->sysfs_lock);

Have you considered to move the q->elevator check into 
elv_unregister_queue() such that no new 'has_elevator' variable has to 
be introduced in this function?

Thanks,

Bart.

Re: [PATCH] block: don't acquire .sysfs_lock before removing mq & iosched kobjects

[Ming Lei]

On Fri, Aug 16, 2019 at 08:31:06AM -0700, Bart Van Assche wrote:
> On 8/16/19 6:55 AM, Ming Lei wrote:
> > diff --git a/block/blk-sysfs.c b/block/blk-sysfs.c
> > index 977c659dcd18..46f033b48917 100644
> > --- a/block/blk-sysfs.c
> > +++ b/block/blk-sysfs.c
> > @@ -1021,6 +1021,7 @@ EXPORT_SYMBOL_GPL(blk_register_queue);
> >   void blk_unregister_queue(struct gendisk *disk)
> >   {
> >   	struct request_queue *q = disk->queue;
> > +	bool has_elevator;
> >   	if (WARN_ON(!q))
> >   		return;
> > @@ -1035,8 +1036,9 @@ void blk_unregister_queue(struct gendisk *disk)
> >   	 * concurrent elv_iosched_store() calls.
> >   	 */
> >   	mutex_lock(&q->sysfs_lock);
> > -
> >   	blk_queue_flag_clear(QUEUE_FLAG_REGISTERED, q);
> > +	has_elevator = q->elevator;
> > +	mutex_unlock(&q->sysfs_lock);
> 
> blk_queue_flag_clear() modifies queue flags atomically so no need to hold
> sysfs_lock around calls of that function.

If you take a look at the above comment, you will see why the sysfs lock
is needed.

> 
> > @@ -1044,16 +1046,13 @@ void blk_unregister_queue(struct gendisk *disk)
> >   	 */
> >   	if (queue_is_mq(q))
> >   		blk_mq_unregister_dev(disk_to_dev(disk), q);
> > -	mutex_unlock(&q->sysfs_lock);
> >   	kobject_uevent(&q->kobj, KOBJ_REMOVE);
> >   	kobject_del(&q->kobj);
> >   	blk_trace_remove_sysfs(disk_to_dev(disk));
> > -	mutex_lock(&q->sysfs_lock);
> > -	if (q->elevator)
> > +	if (has_elevator)
> >   		elv_unregister_queue(q);
> > -	mutex_unlock(&q->sysfs_lock);
> 
> Have you considered to move the q->elevator check into
> elv_unregister_queue() such that no new 'has_elevator' variable has to be
> introduced in this function?

No, I'd keep to read 'q->elevator' with .sysfs_lock.

Thanks,
Ming

Re: [PATCH] block: don't acquire .sysfs_lock before removing mq & iosched kobjects

[Bart Van Assche]

On 8/16/19 6:55 AM, Ming Lei wrote:
> @@ -567,8 +568,17 @@ int elevator_switch_mq(struct request_queue *q,
>   	lockdep_assert_held(&q->sysfs_lock);
>   
>   	if (q->elevator) {
> -		if (q->elevator->registered)
> +		if (q->elevator->registered) {
> +			/*
> +			 * sysfs write is exclusively, release
> +			 * sysfs_lock for avoiding deadlock with
> +			 * sysfs built-in lock which is required
> +			 * in either .show or .store path.
> +			 */
> +			mutex_unlock(&q->sysfs_lock);
>   			elv_unregister_queue(q);
> +			mutex_lock(&q->sysfs_lock);
> +		}
>   		ioc_clear_queue(q);
>   		elevator_exit(q, q->elevator);
>   	}

Hi Ming,

I don't like this part of the patch. Consider the following call chain:

queue_attr_store() -> elv_iosched_store() -> __elevator_change() ->
elevator_switch() -> elevator_switch_mq().

queue_attr_store() locks sysfs_lock to serialize sysfs attribute show
and store callbacks. So the above changes unlocks sysfs_lock from inside
such a callback function and hence breaks that serialization. Can you
have a look at the alternative patch below?

Thanks,

Bart.


Subject: [PATCH] block: Fix lock inversion triggered during request queue removal

Call blk_mq_unregister_dev() after having deleted q->kobj. Move
the kobject_uevent(q->mq_kobj, KOBJ_REMOVE) call from inside
blk_mq_unregister_dev() to its caller.

---
  block/blk-mq-sysfs.c |  5 ++---
  block/blk-sysfs.c    | 19 +++----------------
  block/elevator.c     |  2 --
  3 files changed, 5 insertions(+), 21 deletions(-)

diff --git a/block/blk-mq-sysfs.c b/block/blk-mq-sysfs.c
index d6e1a9bd7131..0ec968009791 100644
--- a/block/blk-mq-sysfs.c
+++ b/block/blk-mq-sysfs.c
@@ -270,16 +270,15 @@ void blk_mq_unregister_dev(struct device *dev, struct request_queue *q)
  	struct blk_mq_hw_ctx *hctx;
  	int i;

-	lockdep_assert_held(&q->sysfs_lock);
-
  	queue_for_each_hw_ctx(q, hctx, i)
  		blk_mq_unregister_hctx(hctx);

-	kobject_uevent(q->mq_kobj, KOBJ_REMOVE);
  	kobject_del(q->mq_kobj);
  	kobject_put(&dev->kobj);

+	mutex_lock(&q->sysfs_lock);
  	q->mq_sysfs_init_done = false;
+	mutex_unlock(&q->sysfs_lock);
  }

  void blk_mq_hctx_kobj_init(struct blk_mq_hw_ctx *hctx)
diff --git a/block/blk-sysfs.c b/block/blk-sysfs.c
index 977c659dcd18..e6f8cd99aded 100644
--- a/block/blk-sysfs.c
+++ b/block/blk-sysfs.c
@@ -1029,31 +1029,18 @@ void blk_unregister_queue(struct gendisk *disk)
  	if (!test_bit(QUEUE_FLAG_REGISTERED, &q->queue_flags))
  		return;

-	/*
-	 * Since sysfs_remove_dir() prevents adding new directory entries
-	 * before removal of existing entries starts, protect against
-	 * concurrent elv_iosched_store() calls.
-	 */
-	mutex_lock(&q->sysfs_lock);
-
  	blk_queue_flag_clear(QUEUE_FLAG_REGISTERED, q);

-	/*
-	 * Remove the sysfs attributes before unregistering the queue data
-	 * structures that can be modified through sysfs.
-	 */
  	if (queue_is_mq(q))
-		blk_mq_unregister_dev(disk_to_dev(disk), q);
-	mutex_unlock(&q->sysfs_lock);
-
+		kobject_uevent(q->mq_kobj, KOBJ_REMOVE);
  	kobject_uevent(&q->kobj, KOBJ_REMOVE);
  	kobject_del(&q->kobj);
  	blk_trace_remove_sysfs(disk_to_dev(disk));

-	mutex_lock(&q->sysfs_lock);
+	if (queue_is_mq(q))
+		blk_mq_unregister_dev(disk_to_dev(disk), q);
  	if (q->elevator)
  		elv_unregister_queue(q);
-	mutex_unlock(&q->sysfs_lock);

  	kobject_put(&disk_to_dev(disk)->kobj);
  }
diff --git a/block/elevator.c b/block/elevator.c
index 2f17d66d0e61..128e7cf032e1 100644
--- a/block/elevator.c
+++ b/block/elevator.c
@@ -495,8 +495,6 @@ int elv_register_queue(struct request_queue *q)

  void elv_unregister_queue(struct request_queue *q)
  {
-	lockdep_assert_held(&q->sysfs_lock);
-
  	if (q) {
  		struct elevator_queue *e = q->elevator;

Re: [PATCH] block: don't acquire .sysfs_lock before removing mq & iosched kobjects

[Ming Lei]

On Tue, Aug 20, 2019 at 02:21:10PM -0700, Bart Van Assche wrote:
> On 8/16/19 6:55 AM, Ming Lei wrote:
> > @@ -567,8 +568,17 @@ int elevator_switch_mq(struct request_queue *q,
> >   	lockdep_assert_held(&q->sysfs_lock);
> >   	if (q->elevator) {
> > -		if (q->elevator->registered)
> > +		if (q->elevator->registered) {
> > +			/*
> > +			 * sysfs write is exclusively, release
> > +			 * sysfs_lock for avoiding deadlock with
> > +			 * sysfs built-in lock which is required
> > +			 * in either .show or .store path.
> > +			 */
> > +			mutex_unlock(&q->sysfs_lock);
> >   			elv_unregister_queue(q);
> > +			mutex_lock(&q->sysfs_lock);
> > +		}
> >   		ioc_clear_queue(q);
> >   		elevator_exit(q, q->elevator);
> >   	}
> 
> Hi Ming,
> 
> I don't like this part of the patch. Consider the following call chain:
> 
> queue_attr_store() -> elv_iosched_store() -> __elevator_change() ->
> elevator_switch() -> elevator_switch_mq().
> 
> queue_attr_store() locks sysfs_lock to serialize sysfs attribute show
> and store callbacks. So the above changes unlocks sysfs_lock from inside
> such a callback function and hence breaks that serialization.

I guess we have to break the serialization, because we can't hold
q->sysfs_lock when deleting iosched kobjects. And we may re-check
QUEUE_FLAG_REGISTERED after grabbing .sysfs_lock again, and stop
to switch to new scheduler if queue is un-registered.

> Can you
> have a look at the alternative patch below?

Sure.

> 
> Thanks,
> 
> Bart.
> 
> 
> Subject: [PATCH] block: Fix lock inversion triggered during request queue removal
> 
> Call blk_mq_unregister_dev() after having deleted q->kobj. Move
> the kobject_uevent(q->mq_kobj, KOBJ_REMOVE) call from inside
> blk_mq_unregister_dev() to its caller.
> 
> ---
>  block/blk-mq-sysfs.c |  5 ++---
>  block/blk-sysfs.c    | 19 +++----------------
>  block/elevator.c     |  2 --
>  3 files changed, 5 insertions(+), 21 deletions(-)
> 
> diff --git a/block/blk-mq-sysfs.c b/block/blk-mq-sysfs.c
> index d6e1a9bd7131..0ec968009791 100644
> --- a/block/blk-mq-sysfs.c
> +++ b/block/blk-mq-sysfs.c
> @@ -270,16 +270,15 @@ void blk_mq_unregister_dev(struct device *dev, struct request_queue *q)
>  	struct blk_mq_hw_ctx *hctx;
>  	int i;
> 
> -	lockdep_assert_held(&q->sysfs_lock);
> -
>  	queue_for_each_hw_ctx(q, hctx, i)
>  		blk_mq_unregister_hctx(hctx);
> 
> -	kobject_uevent(q->mq_kobj, KOBJ_REMOVE);
>  	kobject_del(q->mq_kobj);
>  	kobject_put(&dev->kobj);
> 
> +	mutex_lock(&q->sysfs_lock);
>  	q->mq_sysfs_init_done = false;
> +	mutex_unlock(&q->sysfs_lock);
>  }
> 
>  void blk_mq_hctx_kobj_init(struct blk_mq_hw_ctx *hctx)
> diff --git a/block/blk-sysfs.c b/block/blk-sysfs.c
> index 977c659dcd18..e6f8cd99aded 100644
> --- a/block/blk-sysfs.c
> +++ b/block/blk-sysfs.c
> @@ -1029,31 +1029,18 @@ void blk_unregister_queue(struct gendisk *disk)
>  	if (!test_bit(QUEUE_FLAG_REGISTERED, &q->queue_flags))
>  		return;
> 
> -	/*
> -	 * Since sysfs_remove_dir() prevents adding new directory entries
> -	 * before removal of existing entries starts, protect against
> -	 * concurrent elv_iosched_store() calls.
> -	 */
> -	mutex_lock(&q->sysfs_lock);
> -
>  	blk_queue_flag_clear(QUEUE_FLAG_REGISTERED, q);

.sysfs_lock has to be held for clearing 'REGISTERED', we have to wait
for concurrent store to 'queue/scheduler' before un-registering queue.

> 
> -	/*
> -	 * Remove the sysfs attributes before unregistering the queue data
> -	 * structures that can be modified through sysfs.
> -	 */
>  	if (queue_is_mq(q))
> -		blk_mq_unregister_dev(disk_to_dev(disk), q);
> -	mutex_unlock(&q->sysfs_lock);
> -
> +		kobject_uevent(q->mq_kobj, KOBJ_REMOVE);

Could you explain why you move the above line here?

>  	kobject_uevent(&q->kobj, KOBJ_REMOVE);
>  	kobject_del(&q->kobj);
>  	blk_trace_remove_sysfs(disk_to_dev(disk));
> 
> -	mutex_lock(&q->sysfs_lock);
> +	if (queue_is_mq(q))
> +		blk_mq_unregister_dev(disk_to_dev(disk), q);
>  	if (q->elevator)
>  		elv_unregister_queue(q);
> -	mutex_unlock(&q->sysfs_lock);
> 
>  	kobject_put(&disk_to_dev(disk)->kobj);
>  }
> diff --git a/block/elevator.c b/block/elevator.c
> index 2f17d66d0e61..128e7cf032e1 100644
> --- a/block/elevator.c
> +++ b/block/elevator.c
> @@ -495,8 +495,6 @@ int elv_register_queue(struct request_queue *q)
> 
>  void elv_unregister_queue(struct request_queue *q)
>  {
> -	lockdep_assert_held(&q->sysfs_lock);
> -
>  	if (q) {
>  		struct elevator_queue *e = q->elevator;

elv_unregister_queue() is still called in elevator_switch_mq() with
.sysfs_lock held, then potential AB-BA lock still exists.

The same issue exists on blk_mq_sysfs_unregister() too.


Thanks,
Ming

Re: [PATCH] block: don't acquire .sysfs_lock before removing mq & iosched kobjects

[Bart Van Assche]

On 8/20/19 8:00 PM, Ming Lei wrote:
> On Tue, Aug 20, 2019 at 02:21:10PM -0700, Bart Van Assche wrote:
>> -	/*
>> -	 * Remove the sysfs attributes before unregistering the queue data
>> -	 * structures that can be modified through sysfs.
>> -	 */
>>   	if (queue_is_mq(q))
>> -		blk_mq_unregister_dev(disk_to_dev(disk), q);
>> -	mutex_unlock(&q->sysfs_lock);
>> -
>> +		kobject_uevent(q->mq_kobj, KOBJ_REMOVE);
> 
> Could you explain why you move the above line here?

I'm not sure whether kobject_del() deletes any objects attached to the 
deleted kobj. This change ensures that kobject_uevent() is called before 
the parent object of q->mq_kobj is deleted.

Bart.

Re: [PATCH] block: don't acquire .sysfs_lock before removing mq & iosched kobjects

[Ming Lei]

On Wed, Aug 21, 2019 at 08:41:32AM -0700, Bart Van Assche wrote:
> On 8/20/19 8:00 PM, Ming Lei wrote:
> > On Tue, Aug 20, 2019 at 02:21:10PM -0700, Bart Van Assche wrote:
> > > -	/*
> > > -	 * Remove the sysfs attributes before unregistering the queue data
> > > -	 * structures that can be modified through sysfs.
> > > -	 */
> > >   	if (queue_is_mq(q))
> > > -		blk_mq_unregister_dev(disk_to_dev(disk), q);
> > > -	mutex_unlock(&q->sysfs_lock);
> > > -
> > > +		kobject_uevent(q->mq_kobj, KOBJ_REMOVE);
> > 
> > Could you explain why you move the above line here?
> 
> I'm not sure whether kobject_del() deletes any objects attached to the
> deleted kobj. This change ensures that kobject_uevent() is called before the
> parent object of q->mq_kobj is deleted.

From comment of kernfs_remove(), all subdirectories and files will be
removed.

kobject_del
	sysfs_remove_dir
		kernfs_remove

	/**
	 * kernfs_remove - remove a kernfs_node recursively
	 * @kn: the kernfs_node to remove
	 *
	 * Remove @kn along with all its subdirectories and files.
	 */
	void kernfs_remove(struct kernfs_node *kn)


Thanks,
Ming