Re: [PATCH v2] scsi: core: Fix hang of freezing queue between blocking and running device

[Bart Van Assche <bvanassche@acm.org>]

On 8/9/21 7:13 AM, lijinlin3@huawei.com wrote:
> From: Li Jinlin <lijinlin3@huawei.com>
> 
> We found a hang issue, the test steps are as follows:
>    1. blocking device via scsi_device_set_state()
>    2. dd if=/dev/sda of=/mnt/t.log bs=1M count=10
>    3. echo none > /sys/block/sda/queue/scheduler
>    4. echo "running" >/sys/block/sda/device/state
> 
> Step 3 and 4 should finish this work after step 4, but they hangs.
> 
>    CPU#0               CPU#1                CPU#2
>    ---------------     ----------------     ----------------
>                                             Step 1: blocking device
> 
>                                             Step 2: dd xxxx
>                                                    ^^^^^^ get request
>                                                           q_usage_counter++
> 
>                        Step 3: switching scheculer
>                        elv_iosched_store
>                          elevator_switch
>                            blk_mq_freeze_queue
>                              blk_freeze_queue
>                                > blk_freeze_queue_start
>                                  ^^^^^^ mq_freeze_depth++
> 
>                                > blk_mq_run_hw_queues
>                                  ^^^^^^ can't run queue when dev blocked
> 
>                                > blk_mq_freeze_queue_wait
>                                  ^^^^^^ Hang here!!!
>                                         wait q_usage_counter==0
> 
>    Step 4: running device
>    store_state_field
>      scsi_rescan_device
>        scsi_attach_vpd
>          scsi_vpd_inquiry
>            __scsi_execute
>              blk_get_request
>                blk_mq_alloc_request
>                  blk_queue_enter
>                  ^^^^^^ Hang here!!!
>                         wait mq_freeze_depth==0
> 
>      blk_mq_run_hw_queues
>      ^^^^^^ dispatch IO, q_usage_counter will reduce to zero
> 
>                              blk_mq_unfreeze_queue
>                              ^^^^^ mq_freeze_depth--
> 
> Step 3 and 4 wait for each other.
> 
> To fix this, we need to run queue before rescanning device when the device
> state changes to SDEV_RUNNING.
> 
> Fixes: f0f82e2476f6 ("scsi: core: Fix capacity set to zero after offlinining device")
> Signed-off-by: Li Jinlin <lijinlin3@huawei.com>
> Signed-off-by: Qiu Laibin <qiulaibin@huawei.com>
> ---
> changes since v1 send with Message-ID:
> 20210805143231.1713299-1-lijinlin3@huawei.com
> 
>   - Modify the subject to make it distinct
>   - Modify the message to fix typo and make it distinct
>   - Reduce the number of SOB
> 
>   drivers/scsi/scsi_sysfs.c | 6 +++---
>   1 file changed, 3 insertions(+), 3 deletions(-)
> 
> diff --git a/drivers/scsi/scsi_sysfs.c b/drivers/scsi/scsi_sysfs.c
> index c3a710bceba0..aa701582c950 100644
> --- a/drivers/scsi/scsi_sysfs.c
> +++ b/drivers/scsi/scsi_sysfs.c
> @@ -809,12 +809,12 @@ store_state_field(struct device *dev, struct device_attribute *attr,
>   	ret = scsi_device_set_state(sdev, state);
>   	/*
>   	 * If the device state changes to SDEV_RUNNING, we need to
> -	 * rescan the device to revalidate it, and run the queue to
> -	 * avoid I/O hang.
> +	 * run the queue to avoid I/O hang, and rescan the device
> +	 * to revalidate it.
>   	 */
>   	if (ret == 0 && state == SDEV_RUNNING) {
> -		scsi_rescan_device(dev);
>   		blk_mq_run_hw_queues(sdev->request_queue, true);
> +		scsi_rescan_device(dev);
>   	}
>   	mutex_unlock(&sdev->state_mutex);

The patch looks fine to me but I think the comment in 
store_state_field() should be expanded. Although the description in the 
commit message makes it clear how I/O may hang, that is not clear from 
the source code comment. Please mention in the comment that running the 
queue first is necessary because another thread may be waiting inside 
blk_mq_freeze_queue_wait() and because that call may be waiting for 
pending I/O to finish.

Thanks,

Bart.