[PATCH v3 0/3] Fix races between blkcg code and request queue initialization and cleanup

[PATCH v3 0/3] Fix races between blkcg code and request queue initialization and cleanup

[Bart Van Assche]

Hello Jens,

Recently Joseph Qi identified races between the block cgroup code and request
queue initialization and cleanup. This patch series address these races. Please
consider these patches for kernel v4.17.

Thanks,

Bart.

Changes between v2 and v3:
- Added a third patch that fixes a race between the blkcg code and queue
  cleanup.

Changes between v1 and v2:
- Split a single patch into two patches.
- Dropped blk_alloc_queue_node2() and modified all block drivers that call
  blk_alloc_queue_node().

Bart Van Assche (3):
  block: Add a third argument to blk_alloc_queue_node()
  block: Fix a race between the cgroup code and request queue
    initialization
  block: Fix a race between request queue removal and the block cgroup
    controller

 block/blk-core.c               | 60 +++++++++++++++++++++++++++++++++++-------
 block/blk-mq.c                 |  2 +-
 block/blk-sysfs.c              |  7 -----
 drivers/block/drbd/drbd_main.c |  3 +--
 drivers/block/null_blk.c       |  3 ++-
 drivers/block/umem.c           |  7 +++--
 drivers/ide/ide-probe.c        |  2 +-
 drivers/lightnvm/core.c        |  2 +-
 drivers/md/dm.c                |  2 +-
 drivers/nvdimm/pmem.c          |  2 +-
 drivers/nvme/host/multipath.c  |  2 +-
 drivers/scsi/scsi_lib.c        |  2 +-
 include/linux/blkdev.h         |  3 ++-
 13 files changed, 65 insertions(+), 32 deletions(-)

-- 
2.16.1

[PATCH v3 1/3] block: Add a third argument to blk_alloc_queue_node()

[Bart Van Assche]

This patch does not change any functionality.

Signed-off-by: Bart Van Assche <bart.vanassche@wdc.com>
Cc: Christoph Hellwig <hch@lst.de>
Cc: Joseph Qi <joseph.qi@linux.alibaba.com>
Cc: Philipp Reisner <philipp.reisner@linbit.com>
Cc: Ulf Hansson <ulf.hansson@linaro.org>
Cc: Kees Cook <keescook@chromium.org>
---
 block/blk-core.c              | 7 ++++---
 block/blk-mq.c                | 2 +-
 drivers/block/null_blk.c      | 3 ++-
 drivers/ide/ide-probe.c       | 2 +-
 drivers/lightnvm/core.c       | 2 +-
 drivers/md/dm.c               | 2 +-
 drivers/nvdimm/pmem.c         | 2 +-
 drivers/nvme/host/multipath.c | 2 +-
 drivers/scsi/scsi_lib.c       | 2 +-
 include/linux/blkdev.h        | 3 ++-
 10 files changed, 15 insertions(+), 12 deletions(-)

diff --git a/block/blk-core.c b/block/blk-core.c
index 2d1a7bbe0634..e873a24bf82d 100644
--- a/block/blk-core.c
+++ b/block/blk-core.c
@@ -810,7 +810,7 @@ void blk_exit_rl(struct request_queue *q, struct request_list *rl)
 
 struct request_queue *blk_alloc_queue(gfp_t gfp_mask)
 {
-	return blk_alloc_queue_node(gfp_mask, NUMA_NO_NODE);
+	return blk_alloc_queue_node(gfp_mask, NUMA_NO_NODE, NULL);
 }
 EXPORT_SYMBOL(blk_alloc_queue);
 
@@ -888,7 +888,8 @@ static void blk_rq_timed_out_timer(struct timer_list *t)
 	kblockd_schedule_work(&q->timeout_work);
 }
 
-struct request_queue *blk_alloc_queue_node(gfp_t gfp_mask, int node_id)
+struct request_queue *blk_alloc_queue_node(gfp_t gfp_mask, int node_id,
+					   spinlock_t *lock)
 {
 	struct request_queue *q;
 
@@ -1030,7 +1031,7 @@ blk_init_queue_node(request_fn_proc *rfn, spinlock_t *lock, int node_id)
 {
 	struct request_queue *q;
 
-	q = blk_alloc_queue_node(GFP_KERNEL, node_id);
+	q = blk_alloc_queue_node(GFP_KERNEL, node_id, NULL);
 	if (!q)
 		return NULL;
 
diff --git a/block/blk-mq.c b/block/blk-mq.c
index df93102e2149..ae4e5096f425 100644
--- a/block/blk-mq.c
+++ b/block/blk-mq.c
@@ -2554,7 +2554,7 @@ struct request_queue *blk_mq_init_queue(struct blk_mq_tag_set *set)
 {
 	struct request_queue *uninit_q, *q;
 
-	uninit_q = blk_alloc_queue_node(GFP_KERNEL, set->numa_node);
+	uninit_q = blk_alloc_queue_node(GFP_KERNEL, set->numa_node, NULL);
 	if (!uninit_q)
 		return ERR_PTR(-ENOMEM);
 
diff --git a/drivers/block/null_blk.c b/drivers/block/null_blk.c
index 287a09611c0f..f3d17ea88965 100644
--- a/drivers/block/null_blk.c
+++ b/drivers/block/null_blk.c
@@ -1717,7 +1717,8 @@ static int null_add_dev(struct nullb_device *dev)
 		}
 		null_init_queues(nullb);
 	} else if (dev->queue_mode == NULL_Q_BIO) {
-		nullb->q = blk_alloc_queue_node(GFP_KERNEL, dev->home_node);
+		nullb->q = blk_alloc_queue_node(GFP_KERNEL, dev->home_node,
+						NULL);
 		if (!nullb->q) {
 			rv = -ENOMEM;
 			goto out_cleanup_queues;
diff --git a/drivers/ide/ide-probe.c b/drivers/ide/ide-probe.c
index 17fd55af4d92..2e80a866073c 100644
--- a/drivers/ide/ide-probe.c
+++ b/drivers/ide/ide-probe.c
@@ -766,7 +766,7 @@ static int ide_init_queue(ide_drive_t *drive)
 	 *	limits and LBA48 we could raise it but as yet
 	 *	do not.
 	 */
-	q = blk_alloc_queue_node(GFP_KERNEL, hwif_to_node(hwif));
+	q = blk_alloc_queue_node(GFP_KERNEL, hwif_to_node(hwif), NULL);
 	if (!q)
 		return 1;
 
diff --git a/drivers/lightnvm/core.c b/drivers/lightnvm/core.c
index dcc9e621e651..5f1988df1593 100644
--- a/drivers/lightnvm/core.c
+++ b/drivers/lightnvm/core.c
@@ -384,7 +384,7 @@ static int nvm_create_tgt(struct nvm_dev *dev, struct nvm_ioctl_create *create)
 		goto err_dev;
 	}
 
-	tqueue = blk_alloc_queue_node(GFP_KERNEL, dev->q->node);
+	tqueue = blk_alloc_queue_node(GFP_KERNEL, dev->q->node, NULL);
 	if (!tqueue) {
 		ret = -ENOMEM;
 		goto err_disk;
diff --git a/drivers/md/dm.c b/drivers/md/dm.c
index d6de00f367ef..3c55564f6367 100644
--- a/drivers/md/dm.c
+++ b/drivers/md/dm.c
@@ -1840,7 +1840,7 @@ static struct mapped_device *alloc_dev(int minor)
 	INIT_LIST_HEAD(&md->table_devices);
 	spin_lock_init(&md->uevent_lock);
 
-	md->queue = blk_alloc_queue_node(GFP_KERNEL, numa_node_id);
+	md->queue = blk_alloc_queue_node(GFP_KERNEL, numa_node_id, NULL);
 	if (!md->queue)
 		goto bad;
 	md->queue->queuedata = md;
diff --git a/drivers/nvdimm/pmem.c b/drivers/nvdimm/pmem.c
index 10041ac4032c..cfb15ac50925 100644
--- a/drivers/nvdimm/pmem.c
+++ b/drivers/nvdimm/pmem.c
@@ -344,7 +344,7 @@ static int pmem_attach_disk(struct device *dev,
 		return -EBUSY;
 	}
 
-	q = blk_alloc_queue_node(GFP_KERNEL, dev_to_node(dev));
+	q = blk_alloc_queue_node(GFP_KERNEL, dev_to_node(dev), NULL);
 	if (!q)
 		return -ENOMEM;
 
diff --git a/drivers/nvme/host/multipath.c b/drivers/nvme/host/multipath.c
index 3b211d9e58b8..cf013784fa85 100644
--- a/drivers/nvme/host/multipath.c
+++ b/drivers/nvme/host/multipath.c
@@ -162,7 +162,7 @@ int nvme_mpath_alloc_disk(struct nvme_ctrl *ctrl, struct nvme_ns_head *head)
 	if (!(ctrl->subsys->cmic & (1 << 1)) || !multipath)
 		return 0;
 
-	q = blk_alloc_queue_node(GFP_KERNEL, NUMA_NO_NODE);
+	q = blk_alloc_queue_node(GFP_KERNEL, NUMA_NO_NODE, NULL);
 	if (!q)
 		goto out;
 	q->queuedata = head;
diff --git a/drivers/scsi/scsi_lib.c b/drivers/scsi/scsi_lib.c
index a86df9ca7d1c..71d1135f94d0 100644
--- a/drivers/scsi/scsi_lib.c
+++ b/drivers/scsi/scsi_lib.c
@@ -2223,7 +2223,7 @@ struct request_queue *scsi_old_alloc_queue(struct scsi_device *sdev)
 	struct Scsi_Host *shost = sdev->host;
 	struct request_queue *q;
 
-	q = blk_alloc_queue_node(GFP_KERNEL, NUMA_NO_NODE);
+	q = blk_alloc_queue_node(GFP_KERNEL, NUMA_NO_NODE, NULL);
 	if (!q)
 		return NULL;
 	q->cmd_size = sizeof(struct scsi_cmnd) + shost->hostt->cmd_size;
diff --git a/include/linux/blkdev.h b/include/linux/blkdev.h
index 4f3df807cf8f..d78dd92a9b6a 100644
--- a/include/linux/blkdev.h
+++ b/include/linux/blkdev.h
@@ -1321,7 +1321,8 @@ extern long nr_blockdev_pages(void);
 
 bool __must_check blk_get_queue(struct request_queue *);
 struct request_queue *blk_alloc_queue(gfp_t);
-struct request_queue *blk_alloc_queue_node(gfp_t, int);
+struct request_queue *blk_alloc_queue_node(gfp_t gfp_mask, int node_id,
+					   spinlock_t *lock);
 extern void blk_put_queue(struct request_queue *);
 extern void blk_set_queue_dying(struct request_queue *);
 
-- 
2.16.1

[PATCH v3 2/3] block: Fix a race between the cgroup code and request queue initialization

[Bart Van Assche]

Initialize the request queue lock earlier such that the following
race can no longer occur:

blk_init_queue_node()             blkcg_print_blkgs()
  blk_alloc_queue_node (1)
    q->queue_lock = &q->__queue_lock (2)
    blkcg_init_queue(q) (3)
                                    spin_lock_irq(blkg->q->queue_lock) (4)
  q->queue_lock = lock (5)
                                    spin_unlock_irq(blkg->q->queue_lock) (6)

(1) allocate an uninitialized queue;
(2) initialize queue_lock to its default internal lock;
(3) initialize blkcg part of request queue, which will create blkg and
    then insert it to blkg_list;
(4) traverse blkg_list and find the created blkg, and then take its
    queue lock, here it is the default *internal lock*;
(5) *race window*, now queue_lock is overridden with *driver specified
    lock*;
(6) now unlock *driver specified lock*, not the locked *internal lock*,
    unlock balance breaks.

The changes in this patch are as follows:
- Move the .queue_lock initialization from blk_init_queue_node() into
  blk_alloc_queue_node().
- Only override the .queue_lock pointer for legacy queues because it
  is not useful for blk-mq queues to override this pointer.
- For all all block drivers that initialize .queue_lock explicitly,
  change the blk_alloc_queue() call in the driver into a
  blk_alloc_queue_node() call and remove the explicit .queue_lock
  initialization. Additionally, initialize the spin lock that will
  be used as queue lock earlier if necessary.

Reported-by: Joseph Qi <joseph.qi@linux.alibaba.com>
Signed-off-by: Bart Van Assche <bart.vanassche@wdc.com>
Cc: Christoph Hellwig <hch@lst.de>
Cc: Joseph Qi <joseph.qi@linux.alibaba.com>
Cc: Philipp Reisner <philipp.reisner@linbit.com>
Cc: Ulf Hansson <ulf.hansson@linaro.org>
Cc: Kees Cook <keescook@chromium.org>
---
 block/blk-core.c               | 24 ++++++++++++++++--------
 drivers/block/drbd/drbd_main.c |  3 +--
 drivers/block/umem.c           |  7 +++----
 3 files changed, 20 insertions(+), 14 deletions(-)

diff --git a/block/blk-core.c b/block/blk-core.c
index e873a24bf82d..41c74b37be85 100644
--- a/block/blk-core.c
+++ b/block/blk-core.c
@@ -888,6 +888,19 @@ static void blk_rq_timed_out_timer(struct timer_list *t)
 	kblockd_schedule_work(&q->timeout_work);
 }
 
+/**
+ * blk_alloc_queue_node - allocate a request queue
+ * @gfp_mask: memory allocation flags
+ * @node_id: NUMA node to allocate memory from
+ * @lock: For legacy queues, pointer to a spinlock that will be used to e.g.
+ *        serialize calls to the legacy .request_fn() callback. Ignored for
+ *	  blk-mq request queues.
+ *
+ * Note: pass the queue lock as the third argument to this function instead of
+ * setting the queue lock pointer explicitly to avoid triggering a sporadic
+ * crash in the blkcg code. This function namely calls blkcg_init_queue() and
+ * the queue lock pointer must be set before blkcg_init_queue() is called.
+ */
 struct request_queue *blk_alloc_queue_node(gfp_t gfp_mask, int node_id,
 					   spinlock_t *lock)
 {
@@ -940,11 +953,8 @@ struct request_queue *blk_alloc_queue_node(gfp_t gfp_mask, int node_id,
 	mutex_init(&q->sysfs_lock);
 	spin_lock_init(&q->__queue_lock);
 
-	/*
-	 * By default initialize queue_lock to internal lock and driver can
-	 * override it later if need be.
-	 */
-	q->queue_lock = &q->__queue_lock;
+	if (!q->mq_ops)
+		q->queue_lock = lock ? : &q->__queue_lock;
 
 	/*
 	 * A queue starts its life with bypass turned on to avoid
@@ -1031,13 +1041,11 @@ blk_init_queue_node(request_fn_proc *rfn, spinlock_t *lock, int node_id)
 {
 	struct request_queue *q;
 
-	q = blk_alloc_queue_node(GFP_KERNEL, node_id, NULL);
+	q = blk_alloc_queue_node(GFP_KERNEL, node_id, lock);
 	if (!q)
 		return NULL;
 
 	q->request_fn = rfn;
-	if (lock)
-		q->queue_lock = lock;
 	if (blk_init_allocated_queue(q) < 0) {
 		blk_cleanup_queue(q);
 		return NULL;
diff --git a/drivers/block/drbd/drbd_main.c b/drivers/block/drbd/drbd_main.c
index 0a0394aa1b9c..185f1ef00a7c 100644
--- a/drivers/block/drbd/drbd_main.c
+++ b/drivers/block/drbd/drbd_main.c
@@ -2816,7 +2816,7 @@ enum drbd_ret_code drbd_create_device(struct drbd_config_context *adm_ctx, unsig
 
 	drbd_init_set_defaults(device);
 
-	q = blk_alloc_queue(GFP_KERNEL);
+	q = blk_alloc_queue_node(GFP_KERNEL, NUMA_NO_NODE, &resource->req_lock);
 	if (!q)
 		goto out_no_q;
 	device->rq_queue = q;
@@ -2848,7 +2848,6 @@ enum drbd_ret_code drbd_create_device(struct drbd_config_context *adm_ctx, unsig
 	/* Setting the max_hw_sectors to an odd value of 8kibyte here
 	   This triggers a max_bio_size message upon first attach or connect */
 	blk_queue_max_hw_sectors(q, DRBD_MAX_BIO_SIZE_SAFE >> 8);
-	q->queue_lock = &resource->req_lock;
 
 	device->md_io.page = alloc_page(GFP_KERNEL);
 	if (!device->md_io.page)
diff --git a/drivers/block/umem.c b/drivers/block/umem.c
index 8077123678ad..5c7fb8cc4149 100644
--- a/drivers/block/umem.c
+++ b/drivers/block/umem.c
@@ -888,13 +888,14 @@ static int mm_pci_probe(struct pci_dev *dev, const struct pci_device_id *id)
 	card->Active = -1;	/* no page is active */
 	card->bio = NULL;
 	card->biotail = &card->bio;
+	spin_lock_init(&card->lock);
 
-	card->queue = blk_alloc_queue(GFP_KERNEL);
+	card->queue = blk_alloc_queue_node(GFP_KERNEL, NUMA_NO_NODE,
+					   &card->lock);
 	if (!card->queue)
 		goto failed_alloc;
 
 	blk_queue_make_request(card->queue, mm_make_request);
-	card->queue->queue_lock = &card->lock;
 	card->queue->queuedata = card;
 
 	tasklet_init(&card->tasklet, process_page, (unsigned long)card);
@@ -968,8 +969,6 @@ static int mm_pci_probe(struct pci_dev *dev, const struct pci_device_id *id)
 	dev_printk(KERN_INFO, &card->dev->dev,
 		"Window size %d bytes, IRQ %d\n", data, dev->irq);
 
-	spin_lock_init(&card->lock);
-
 	pci_set_drvdata(dev, card);
 
 	if (pci_write_cmd != 0x0F) 	/* If not Memory Write & Invalidate */
-- 
2.16.1

[PATCH v3 3/3] block: Fix a race between request queue removal and the block cgroup controller

[Bart Van Assche]

Avoid that the following race can occur:

blk_cleanup_queue()               blkcg_print_blkgs()
  spin_lock_irq(lock) (1)           spin_lock_irq(blkg->q->queue_lock) (2,5)
    q->queue_lock = &q->__queue_lock (3)
  spin_unlock_irq(lock) (4)
                                    spin_unlock_irq(blkg->q->queue_lock) (6)

(1) take driver lock;
(2) busy loop for driver lock;
(3) override driver lock with internal lock;
(4) unlock driver lock;
(5) can take driver lock now;
(6) but unlock internal lock.

This change is safe because only the SCSI core and the NVME core keep
a reference on a request queue after having called blk_cleanup_queue().
Neither driver accesses any of the removed data structures between its
blk_cleanup_queue() and blk_put_queue() calls.

Reported-by: Joseph Qi <joseph.qi@linux.alibaba.com>
Signed-off-by: Bart Van Assche <bart.vanassche@wdc.com>
Cc: Jan Kara <jack@suse.com>
---
 block/blk-core.c  | 31 +++++++++++++++++++++++++++++++
 block/blk-sysfs.c |  7 -------
 2 files changed, 31 insertions(+), 7 deletions(-)

diff --git a/block/blk-core.c b/block/blk-core.c
index 41c74b37be85..6febc69a58aa 100644
--- a/block/blk-core.c
+++ b/block/blk-core.c
@@ -719,6 +719,37 @@ void blk_cleanup_queue(struct request_queue *q)
 	del_timer_sync(&q->backing_dev_info->laptop_mode_wb_timer);
 	blk_sync_queue(q);
 
+	/*
+	 * I/O scheduler exit is only safe after the sysfs scheduler attribute
+	 * has been removed.
+	 */
+	WARN_ON_ONCE(q->kobj.state_in_sysfs);
+
+	/*
+	 * Since the I/O scheduler exit code may access cgroup information,
+	 * perform I/O scheduler exit before disassociating from the block
+	 * cgroup controller.
+	 */
+	if (q->elevator) {
+		ioc_clear_queue(q);
+		elevator_exit(q, q->elevator);
+		q->elevator = NULL;
+	}
+
+	/*
+	 * Remove all references to @q from the block cgroup controller before
+	 * restoring @q->queue_lock to avoid that restoring this pointer causes
+	 * e.g. blkcg_print_blkgs() to crash.
+	 */
+	blkcg_exit_queue(q);
+
+	/*
+	 * Since the cgroup code may dereference the @q->backing_dev_info
+	 * pointer, only decrease its reference count after having removed the
+	 * association with the block cgroup controller.
+	 */
+	bdi_put(q->backing_dev_info);
+
 	if (q->mq_ops)
 		blk_mq_free_queue(q);
 	percpu_ref_exit(&q->q_usage_counter);
diff --git a/block/blk-sysfs.c b/block/blk-sysfs.c
index cbea895a5547..fd71a00c9462 100644
--- a/block/blk-sysfs.c
+++ b/block/blk-sysfs.c
@@ -798,13 +798,6 @@ static void __blk_release_queue(struct work_struct *work)
 	if (test_bit(QUEUE_FLAG_POLL_STATS, &q->queue_flags))
 		blk_stat_remove_callback(q, q->poll_cb);
 	blk_stat_free_callback(q->poll_cb);
-	bdi_put(q->backing_dev_info);
-	blkcg_exit_queue(q);
-
-	if (q->elevator) {
-		ioc_clear_queue(q);
-		elevator_exit(q, q->elevator);
-	}
 
 	blk_free_queue_stats(q->stats);
 
-- 
2.16.1

Re: [PATCH v3 0/3] Fix races between blkcg code and request queue initialization and cleanup

[Bart Van Assche]

On 02/09/18 10:44, Bart Van Assche wrote:
> Recently Joseph Qi identified races between the block cgroup code and request
> queue initialization and cleanup. This patch series address these races. Please
> consider these patches for kernel v4.17.

Hello Joseph,

Can you add your Tested-by or Reviewed-by to the patches in this series?

Thanks,

Bart.

Re: [PATCH v3 3/3] block: Fix a race between request queue removal and the block cgroup controller

[Joseph Qi]

Hi Bart,

Sorry for the delayed response since I was on holiday.

On 18/2/10 02:44, Bart Van Assche wrote:
> Avoid that the following race can occur:
> 
> blk_cleanup_queue()               blkcg_print_blkgs()
>   spin_lock_irq(lock) (1)           spin_lock_irq(blkg->q->queue_lock) (2,5)
>     q->queue_lock = &q->__queue_lock (3)
>   spin_unlock_irq(lock) (4)
>                                     spin_unlock_irq(blkg->q->queue_lock) (6)
> 
> (1) take driver lock;
> (2) busy loop for driver lock;
> (3) override driver lock with internal lock;
> (4) unlock driver lock;
> (5) can take driver lock now;
> (6) but unlock internal lock.
> 
> This change is safe because only the SCSI core and the NVME core keep
> a reference on a request queue after having called blk_cleanup_queue().
> Neither driver accesses any of the removed data structures between its
> blk_cleanup_queue() and blk_put_queue() calls.
> 
> Reported-by: Joseph Qi <joseph.qi@linux.alibaba.com>
> Signed-off-by: Bart Van Assche <bart.vanassche@wdc.com>
> Cc: Jan Kara <jack@suse.com>
> ---
>  block/blk-core.c  | 31 +++++++++++++++++++++++++++++++
>  block/blk-sysfs.c |  7 -------
>  2 files changed, 31 insertions(+), 7 deletions(-)
> 
> diff --git a/block/blk-core.c b/block/blk-core.c
> index 41c74b37be85..6febc69a58aa 100644
> --- a/block/blk-core.c
> +++ b/block/blk-core.c
> @@ -719,6 +719,37 @@ void blk_cleanup_queue(struct request_queue *q)
>  	del_timer_sync(&q->backing_dev_info->laptop_mode_wb_timer);
>  	blk_sync_queue(q);
>  
> +	/*
> +	 * I/O scheduler exit is only safe after the sysfs scheduler attribute
> +	 * has been removed.
> +	 */
> +	WARN_ON_ONCE(q->kobj.state_in_sysfs);
> +
I notice that several devices such as loop and zram will call
blk_cleanup_queue before del_gendisk, so it will hit this warning. Is
this normal?

Thanks,
Joseph

> +	/*
> +	 * Since the I/O scheduler exit code may access cgroup information,
> +	 * perform I/O scheduler exit before disassociating from the block
> +	 * cgroup controller.
> +	 */
> +	if (q->elevator) {
> +		ioc_clear_queue(q);
> +		elevator_exit(q, q->elevator);
> +		q->elevator = NULL;
> +	}
> +
> +	/*
> +	 * Remove all references to @q from the block cgroup controller before
> +	 * restoring @q->queue_lock to avoid that restoring this pointer causes
> +	 * e.g. blkcg_print_blkgs() to crash.
> +	 */
> +	blkcg_exit_queue(q);
> +
> +	/*
> +	 * Since the cgroup code may dereference the @q->backing_dev_info
> +	 * pointer, only decrease its reference count after having removed the
> +	 * association with the block cgroup controller.
> +	 */
> +	bdi_put(q->backing_dev_info);
> +
>  	if (q->mq_ops)
>  		blk_mq_free_queue(q);
>  	percpu_ref_exit(&q->q_usage_counter);
> diff --git a/block/blk-sysfs.c b/block/blk-sysfs.c
> index cbea895a5547..fd71a00c9462 100644
> --- a/block/blk-sysfs.c
> +++ b/block/blk-sysfs.c
> @@ -798,13 +798,6 @@ static void __blk_release_queue(struct work_struct *work)
>  	if (test_bit(QUEUE_FLAG_POLL_STATS, &q->queue_flags))
>  		blk_stat_remove_callback(q, q->poll_cb);
>  	blk_stat_free_callback(q->poll_cb);
> -	bdi_put(q->backing_dev_info);
> -	blkcg_exit_queue(q);
> -
> -	if (q->elevator) {
> -		ioc_clear_queue(q);
> -		elevator_exit(q, q->elevator);
> -	}
>  
>  	blk_free_queue_stats(q->stats);
>  
> 

Re: [PATCH v3 3/3] block: Fix a race between request queue removal and the block cgroup controller

[Ming Lei]

On Thu, Feb 22, 2018 at 10:25:28AM +0800, Joseph Qi wrote:
> Hi Bart,
> 
> Sorry for the delayed response since I was on holiday.
> 
> On 18/2/10 02:44, Bart Van Assche wrote:
> > Avoid that the following race can occur:
> > 
> > blk_cleanup_queue()               blkcg_print_blkgs()
> >   spin_lock_irq(lock) (1)           spin_lock_irq(blkg->q->queue_lock) (2,5)
> >     q->queue_lock = &q->__queue_lock (3)
> >   spin_unlock_irq(lock) (4)
> >                                     spin_unlock_irq(blkg->q->queue_lock) (6)
> > 
> > (1) take driver lock;
> > (2) busy loop for driver lock;
> > (3) override driver lock with internal lock;
> > (4) unlock driver lock;
> > (5) can take driver lock now;
> > (6) but unlock internal lock.
> > 
> > This change is safe because only the SCSI core and the NVME core keep
> > a reference on a request queue after having called blk_cleanup_queue().
> > Neither driver accesses any of the removed data structures between its
> > blk_cleanup_queue() and blk_put_queue() calls.
> > 
> > Reported-by: Joseph Qi <joseph.qi@linux.alibaba.com>
> > Signed-off-by: Bart Van Assche <bart.vanassche@wdc.com>
> > Cc: Jan Kara <jack@suse.com>
> > ---
> >  block/blk-core.c  | 31 +++++++++++++++++++++++++++++++
> >  block/blk-sysfs.c |  7 -------
> >  2 files changed, 31 insertions(+), 7 deletions(-)
> > 
> > diff --git a/block/blk-core.c b/block/blk-core.c
> > index 41c74b37be85..6febc69a58aa 100644
> > --- a/block/blk-core.c
> > +++ b/block/blk-core.c
> > @@ -719,6 +719,37 @@ void blk_cleanup_queue(struct request_queue *q)
> >  	del_timer_sync(&q->backing_dev_info->laptop_mode_wb_timer);
> >  	blk_sync_queue(q);
> >  
> > +	/*
> > +	 * I/O scheduler exit is only safe after the sysfs scheduler attribute
> > +	 * has been removed.
> > +	 */
> > +	WARN_ON_ONCE(q->kobj.state_in_sysfs);
> > +
> I notice that several devices such as loop and zram will call
> blk_cleanup_queue before del_gendisk, so it will hit this warning. Is
> this normal?

In theory, it should be a bug to call del_gendisk() after blk_cleanup_queue()
since dirty pages can't be flushed to device any more after queue is
cleaned up.

Thanks,
Ming

Re: [PATCH v3 3/3] block: Fix a race between request queue removal and the block cgroup controller

[Bart Van Assche]
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