Re: Strange block/scsi/workqueue issue

[Jens Axboe <jaxboe@fusionio.com>]

On 2011-04-13 19:01, James Bottomley wrote:
> While you still have the problematic system, can you try this patch?  It
> avoids changing anything in block (other than to add a missing state
> guard for the elv_next_request).  If it works, we can defer the sync vs
> async discussion and use it for a -stable fix.
> 
> Thanks,
> 
> James
> 
> ---
> 
> diff --git a/block/blk.h b/block/blk.h
> index c8db371..11d0d25 100644
> --- a/block/blk.h
> +++ b/block/blk.h
> @@ -62,7 +62,8 @@ static inline struct request *__elv_next_request(struct request_queue *q)
>  			return rq;
>  		}
>  
> -		if (!q->elevator->ops->elevator_dispatch_fn(q, 0))
> +		if (test_bit(QUEUE_FLAG_DEAD, &q->queue_flags) ||
> +		    !q->elevator->ops->elevator_dispatch_fn(q, 0))
>  			return NULL;
>  	}
>  }
> diff --git a/drivers/scsi/scsi_sysfs.c b/drivers/scsi/scsi_sysfs.c
> index e44ff64..5aa4246 100644
> --- a/drivers/scsi/scsi_sysfs.c
> +++ b/drivers/scsi/scsi_sysfs.c
> @@ -322,14 +322,9 @@ static void scsi_device_dev_release_usercontext(struct work_struct *work)
>  		kfree(evt);
>  	}
>  
> -	if (sdev->request_queue) {
> -		sdev->request_queue->queuedata = NULL;
> -		/* user context needed to free queue */
> -		scsi_free_queue(sdev->request_queue);
> -		/* temporary expedient, try to catch use of queue lock
> -		 * after free of sdev */
> -		sdev->request_queue = NULL;
> -	}
> +	/* temporary expedient, try to catch use of queue lock after
> +	 * free of sdev */
> +	sdev->request_queue = NULL;
>  
>  	scsi_target_reap(scsi_target(sdev));
>  
> @@ -937,6 +932,11 @@ void __scsi_remove_device(struct scsi_device *sdev)
>  	if (sdev->host->hostt->slave_destroy)
>  		sdev->host->hostt->slave_destroy(sdev);
>  	transport_destroy_device(dev);
> +	/* Setting this to NULL causes the request function to reject
> +	 * any I/O requests */
> +	sdev->request_queue->queuedata = NULL;
> +	/* Freeing the queue signals to block that we're done */
> +	scsi_free_queue(sdev->request_queue);
>  	put_device(dev);
>  }

This patch looks pretty clean. Shouldn't you serialize that ->queuedata
= NULL assignment with the queue lock, though?

-- 
Jens Axboe