[PATCH] blkcg: handle dying request_queue when associating a blkg

Between v3 [1] and v4 [2] of the blkg association series, the
association point moved from generic_make_request_checks(), which is
called after the request enters the queue, to bio_set_dev(), which is when
the bio is formed before submit_bio(). When the request_queue goes away,
the blkgs supporting the request_queue are destroyed and then the
q->root_blkg is set to %NULL.

This patch adds a %NULL check to blkg_tryget_closest() to prevent the
NPE caused by the above. It also adds a guard to see if the
request_queue is dying when creating a blkg to prevent creating a blkg
for a dead request_queue.

[1] https://lore.kernel.org/lkml/20180911184137.35897-1-dennisszhou@gmail.com/
[2] https://lore.kernel.org/lkml/20181126211946.77067-1-dennis@kernel.org/

Fixes: 5cdf2e3fea5e ("blkcg: associate blkg when associating a device")
Reported-and-tested-by: Ming Lei <redacted>
Signed-off-by: Dennis Zhou <dennis@kernel.org>
---
 block/blk-cgroup.c         | 6 ++++++
 include/linux/blk-cgroup.h | 2 +-
 2 files changed, 7 insertions(+), 1 deletion(-)

diff --git a/block/blk-cgroup.c b/block/blk-cgroup.c
index 6bd0619a7d6e..c30661ddc873 100644
--- a/block/blk-cgroup.c
+++ b/block/blk-cgroup.c

@@ -202,6 +202,12 @@ static struct blkcg_gq *blkg_create(struct blkcg *blkcg,
 	WARN_ON_ONCE(!rcu_read_lock_held());
 	lockdep_assert_held(&q->queue_lock);
 
+	/* request_queue is dying, do not create/recreate a blkg */
+	if (blk_queue_dying(q)) {
+		ret = -ENODEV;
+		goto err_free_blkg;
+	}
+
 	/* blkg holds a reference to blkcg */
 	if (!css_tryget_online(&blkcg->css)) {
 		ret = -ENODEV;

diff --git a/include/linux/blk-cgroup.h b/include/linux/blk-cgroup.h
index bf13ecb0fe4f..f025fd1e22e6 100644
--- a/include/linux/blk-cgroup.h
+++ b/include/linux/blk-cgroup.h

@@ -511,7 +511,7 @@ static inline bool blkg_tryget(struct blkcg_gq *blkg)
  */
 static inline struct blkcg_gq *blkg_tryget_closest(struct blkcg_gq *blkg)
 {
-	while (!percpu_ref_tryget(&blkg->refcnt))
+	while (blkg && !percpu_ref_tryget(&blkg->refcnt))
 		blkg = blkg->parent;
 
 	return blkg;

-- 
2.17.1

On Tue, 2018-12-11 at 18:03 -0500, Dennis Zhou wrote:
+AD4 diff --git a/block/blk-cgroup.c b/block/blk-cgroup.c
+AD4 index 6bd0619a7d6e..c30661ddc873 100644
+AD4 --- a/block/blk-cgroup.c
+AD4 +-+-+- b/block/blk-cgroup.c
+AD4 +AEAAQA -202,6 +-202,12 +AEAAQA static struct blkcg+AF8-gq +ACo-blkg+AF8-create(struct blkcg +ACo-blkcg,
+AD4  	WARN+AF8-ON+AF8-ONCE(+ACE-rcu+AF8-read+AF8-lock+AF8-held())+ADs
+AD4  	lockdep+AF8-assert+AF8-held(+ACY-q-+AD4-queue+AF8-lock)+ADs
+AD4  
+AD4 +-	/+ACo request+AF8-queue is dying, do not create/recreate a blkg +ACo-/
+AD4 +-	if (blk+AF8-queue+AF8-dying(q)) +AHs
+AD4 +-		ret +AD0 -ENODEV+ADs
+AD4 +-		goto err+AF8-free+AF8-blkg+ADs
+AD4 +-	+AH0
+AD4 +-
+AD4  	/+ACo blkg holds a reference to blkcg +ACo-/
+AD4  	if (+ACE-css+AF8-tryget+AF8-online(+ACY-blkcg-+AD4-css)) +AHs
+AD4  		ret +AD0 -ENODEV+ADs

What prevents that the queue state changes after blk+AF8-queue+AF8-dying() has returned
and before blkg+AF8-create() returns? Are you sure you don't need to protect this
code with a blk+AF8-queue+AF8-enter() / blk+AF8-queue+AF8-exit() pair?

Bart.

Hi Bart,

On Tue, Dec 11, 2018 at 03:16:13PM -0800, Bart Van Assche wrote:

On Tue, 2018-12-11 at 18:03 -0500, Dennis Zhou wrote:

quoted

diff --git a/block/blk-cgroup.c b/block/blk-cgroup.c
index 6bd0619a7d6e..c30661ddc873 100644
--- a/block/blk-cgroup.c
+++ b/block/blk-cgroup.c

@@ -202,6 +202,12 @@ static struct blkcg_gq *blkg_create(struct blkcg *blkcg,
 	WARN_ON_ONCE(!rcu_read_lock_held());
 	lockdep_assert_held(&q->queue_lock);
 
+	/* request_queue is dying, do not create/recreate a blkg */
+	if (blk_queue_dying(q)) {
+		ret = -ENODEV;
+		goto err_free_blkg;
+	}
+
 	/* blkg holds a reference to blkcg */
 	if (!css_tryget_online(&blkcg->css)) {
 		ret = -ENODEV;

What prevents that the queue state changes after blk_queue_dying() has returned
and before blkg_create() returns? Are you sure you don't need to protect this
code with a blk_queue_enter() / blk_queue_exit() pair?

Hmmm. So I think the idea is that we rely on normal shutdown as I don't
think there is anything wrong with creating a blkg on a dying
request_queue. When we are doing association, the request_queue should
be pinned by the open call. What we are racing against is when the
request_queue is shutting down, it goes around and destroys the blkgs.
For clarity, QUEUE_FLAG_DYING is set in blk_cleanup_queue() before
calling blk_exit_queue() which eventually calls blkcg_exit_queue().

The use of blk_queue_dying() is to determine whether blkg shutdown has
already started as if we create one after it has started, we may
incorrectly orphan a blkg and leak it. Both blkg creation and
destruction require holding the queue_lock, so if the QUEUE_FLAG_DYING
flag is set after we've checked it, it means blkg destruction hasn't
started because it has to wait on the queue_lock. If QUEUE_FLAG_DYING is
set, then we have no guarantee of knowing what phase blkg destruction is
in leading to a potential leak.

Thanks,
Dennis

On Tue, 2018-12-11 at 23:06 -0500, Dennis Zhou wrote:
+AD4 Hi Bart,
+AD4 
+AD4 On Tue, Dec 11, 2018 at 03:16:13PM -0800, Bart Van Assche wrote:
+AD4 +AD4 On Tue, 2018-12-11 at 18:03 -0500, Dennis Zhou wrote:
+AD4 +AD4 +AD4 diff --git a/block/blk-cgroup.c b/block/blk-cgroup.c
+AD4 +AD4 +AD4 index 6bd0619a7d6e..c30661ddc873 100644
+AD4 +AD4 +AD4 --- a/block/blk-cgroup.c
+AD4 +AD4 +AD4 +-+-+- b/block/blk-cgroup.c
+AD4 +AD4 +AD4 +AEAAQA -202,6 +-202,12 +AEAAQA static struct blkcg+AF8-gq +ACo-blkg+AF8-create(struct blkcg +ACo-blkcg,
+AD4 +AD4 +AD4  	WARN+AF8-ON+AF8-ONCE(+ACE-rcu+AF8-read+AF8-lock+AF8-held())+ADs
+AD4 +AD4 +AD4  	lockdep+AF8-assert+AF8-held(+ACY-q-+AD4-queue+AF8-lock)+ADs
+AD4 +AD4 +AD4  
+AD4 +AD4 +AD4 +-	/+ACo request+AF8-queue is dying, do not create/recreate a blkg +ACo-/
+AD4 +AD4 +AD4 +-	if (blk+AF8-queue+AF8-dying(q)) +AHs
+AD4 +AD4 +AD4 +-		ret +AD0 -ENODEV+ADs
+AD4 +AD4 +AD4 +-		goto err+AF8-free+AF8-blkg+ADs
+AD4 +AD4 +AD4 +-	+AH0
+AD4 +AD4 +AD4 +-
+AD4 +AD4 +AD4  	/+ACo blkg holds a reference to blkcg +ACo-/
+AD4 +AD4 +AD4  	if (+ACE-css+AF8-tryget+AF8-online(+ACY-blkcg-+AD4-css)) +AHs
+AD4 +AD4 +AD4  		ret +AD0 -ENODEV+ADs
+AD4 +AD4 
+AD4 +AD4 What prevents that the queue state changes after blk+AF8-queue+AF8-dying() has returned
+AD4 +AD4 and before blkg+AF8-create() returns? Are you sure you don't need to protect this
+AD4 +AD4 code with a blk+AF8-queue+AF8-enter() / blk+AF8-queue+AF8-exit() pair?
+AD4 +AD4 
+AD4 
+AD4 Hmmm. So I think the idea is that we rely on normal shutdown as I don't
+AD4 think there is anything wrong with creating a blkg on a dying
+AD4 request+AF8-queue. When we are doing association, the request+AF8-queue should
+AD4 be pinned by the open call. What we are racing against is when the
+AD4 request+AF8-queue is shutting down, it goes around and destroys the blkgs.
+AD4 For clarity, QUEUE+AF8-FLAG+AF8-DYING is set in blk+AF8-cleanup+AF8-queue() before
+AD4 calling blk+AF8-exit+AF8-queue() which eventually calls blkcg+AF8-exit+AF8-queue().
+AD4 
+AD4 The use of blk+AF8-queue+AF8-dying() is to determine whether blkg shutdown has
+AD4 already started as if we create one after it has started, we may
+AD4 incorrectly orphan a blkg and leak it. Both blkg creation and
+AD4 destruction require holding the queue+AF8-lock, so if the QUEUE+AF8-FLAG+AF8-DYING
+AD4 flag is set after we've checked it, it means blkg destruction hasn't
+AD4 started because it has to wait on the queue+AF8-lock. If QUEUE+AF8-FLAG+AF8-DYING is
+AD4 set, then we have no guarantee of knowing what phase blkg destruction is
+AD4 in leading to a potential leak.

Hi Dennis,

To answer my own question: since all queue flag manipulations are protected
by the queue lock and since blkg+AF8-create() is called with the queue lock held
the above code does not need any further protection. Hence feel free to add
the following:

Reviewed-by: Bart Van Assche +ADw-bvanassche+AEA-acm.org+AD4

On 12/11/18 4:03 PM, Dennis Zhou wrote:

Between v3 [1] and v4 [2] of the blkg association series, the
association point moved from generic_make_request_checks(), which is
called after the request enters the queue, to bio_set_dev(), which is when
the bio is formed before submit_bio(). When the request_queue goes away,
the blkgs supporting the request_queue are destroyed and then the
q->root_blkg is set to %NULL.

This patch adds a %NULL check to blkg_tryget_closest() to prevent the
NPE caused by the above. It also adds a guard to see if the
request_queue is dying when creating a blkg to prevent creating a blkg
for a dead request_queue.

Applied, thanks Dennis.

-- 
Jens Axboe

On Wed, Dec 12, 2018 at 03:54:52PM -0800, Bart Van Assche wrote:

On Tue, 2018-12-11 at 23:06 -0500, Dennis Zhou wrote:

quoted

Hi Bart,

On Tue, Dec 11, 2018 at 03:16:13PM -0800, Bart Van Assche wrote:

quoted

On Tue, 2018-12-11 at 18:03 -0500, Dennis Zhou wrote:

quoted

diff --git a/block/blk-cgroup.c b/block/blk-cgroup.c
index 6bd0619a7d6e..c30661ddc873 100644
--- a/block/blk-cgroup.c
+++ b/block/blk-cgroup.c

@@ -202,6 +202,12 @@ static struct blkcg_gq *blkg_create(struct blkcg *blkcg,
 	WARN_ON_ONCE(!rcu_read_lock_held());
 	lockdep_assert_held(&q->queue_lock);
 
+	/* request_queue is dying, do not create/recreate a blkg */
+	if (blk_queue_dying(q)) {
+		ret = -ENODEV;
+		goto err_free_blkg;
+	}
+
 	/* blkg holds a reference to blkcg */
 	if (!css_tryget_online(&blkcg->css)) {
 		ret = -ENODEV;

What prevents that the queue state changes after blk_queue_dying() has returned
and before blkg_create() returns? Are you sure you don't need to protect this
code with a blk_queue_enter() / blk_queue_exit() pair?

Hmmm. So I think the idea is that we rely on normal shutdown as I don't
think there is anything wrong with creating a blkg on a dying
request_queue. When we are doing association, the request_queue should
be pinned by the open call. What we are racing against is when the
request_queue is shutting down, it goes around and destroys the blkgs.
For clarity, QUEUE_FLAG_DYING is set in blk_cleanup_queue() before
calling blk_exit_queue() which eventually calls blkcg_exit_queue().

The use of blk_queue_dying() is to determine whether blkg shutdown has
already started as if we create one after it has started, we may
incorrectly orphan a blkg and leak it. Both blkg creation and
destruction require holding the queue_lock, so if the QUEUE_FLAG_DYING
flag is set after we've checked it, it means blkg destruction hasn't
started because it has to wait on the queue_lock. If QUEUE_FLAG_DYING is
set, then we have no guarantee of knowing what phase blkg destruction is
in leading to a potential leak.

Hi Dennis,

To answer my own question: since all queue flag manipulations are protected
by the queue lock and since blkg_create() is called with the queue lock held
the above code does not need any further protection. Hence feel free to add
the following:

Reviewed-by: Bart Van Assche <bvanassche@acm.org>

It seems that Christoph in 57d74df90783 ("block: use atomic bitops for
->queue_flags") changed it so that flag manipulations no longer are
protected by the queue_lock in for-4.21/block. But I think my
explanation above suffices that we will always be able to clean up a
blkg created as long as the QUEUE_FLAG_DYING is not set.

Thanks for reviewing this!
Dennis