[PATCH v2] ubi: block: Fix cleanup handling

[Zhihao Cheng]

在 2023/6/7 17:25, Vincent Whitchurch 写道:
> ubiblock's remove handling has a couple of problems:
> 
>   - It uses the gendisk after put_disk(), resulting in a use-after-free.
> 
>   - There is a circular locking dependency between disk->open_mutex (used
>     from del_gendisk() and blkdev_open()) and dev->dev_mutex (used from
>     ubiblock_open() and ubiblock_remove()).
> 
> Fix these by implementing ->free_disk() and moving the final cleanup
> there.
> 
> Signed-off-by: Vincent Whitchurch <vincent.whitchurch at axis.com>
> ---
> Changes in v2:
> - Combine and rework patches to implement and use ->free_disk().
> - Link to v1: https://lore.kernel.org/r/20230523-ubiblock-remove-v1-0-240bed75849b@axis.com
> ---
>   drivers/mtd/ubi/block.c | 27 ++++++++++++++++++++-------
>   1 file changed, 20 insertions(+), 7 deletions(-)
> 
> diff --git a/drivers/mtd/ubi/block.c b/drivers/mtd/ubi/block.c
> index 3711d7f74600..570e660673ad 100644
> --- a/drivers/mtd/ubi/block.c
> +++ b/drivers/mtd/ubi/block.c
> @@ -293,11 +293,23 @@ static int ubiblock_getgeo(struct block_device *bdev, struct hd_geometry *geo)
>   	return 0;
>   }
>   
> +static void ubiblock_free_disk(struct gendisk *disk)
> +{
> +	struct ubiblock *dev = disk->private_data;
> +
> +	mutex_lock(&devices_mutex);
> +	idr_remove(&ubiblock_minor_idr, disk->first_minor);
> +	mutex_unlock(&devices_mutex);
> +
> +	kfree(dev);
> +}
> +
>   static const struct block_device_operations ubiblock_ops = {
>   	.owner = THIS_MODULE,
>   	.open = ubiblock_open,
>   	.release = ubiblock_release,
>   	.getgeo	= ubiblock_getgeo,
> +	.free_disk = ubiblock_free_disk,
>   };
>   
>   static blk_status_t ubiblock_queue_rq(struct blk_mq_hw_ctx *hctx,
> @@ -452,9 +464,8 @@ static void ubiblock_cleanup(struct ubiblock *dev)
>   	del_gendisk(dev->gd);
>   	/* Finally destroy the blk queue */
>   	dev_info(disk_to_dev(dev->gd), "released");
> -	put_disk(dev->gd);
>   	blk_mq_free_tag_set(&dev->tag_set);
> -	idr_remove(&ubiblock_minor_idr, dev->gd->first_minor);
> +	put_disk(dev->gd);

I thought it's better to do put_disk() first then do 
blk_mq_free_tag_set(), likes nbd, loop does. Will 
put_disk->disk_release->blk_mq_exit_queue->blk_mq_exit_hw_queues access 
tag_set which has been freed by blk_mq_free_tag_set()?

>   }
>   
>   int ubiblock_remove(struct ubi_volume_info *vi)
> @@ -478,11 +489,11 @@ int ubiblock_remove(struct ubi_volume_info *vi)
>   
>   	/* Remove from device list */
>   	list_del(&dev->list);
> -	ubiblock_cleanup(dev);
>   	mutex_unlock(&dev->dev_mutex);
>   	mutex_unlock(&devices_mutex);
>   
> -	kfree(dev);
> +	ubiblock_cleanup(dev);
> +
>   	return 0;
>   
>   out_unlock_dev:
> @@ -623,17 +634,19 @@ static void ubiblock_remove_all(void)
>   {
>   	struct ubiblock *next;
>   	struct ubiblock *dev;
> +	LIST_HEAD(list);
>   
>   	mutex_lock(&devices_mutex);
> -	list_for_each_entry_safe(dev, next, &ubiblock_devices, list) {
> +	list_splice_init(&ubiblock_devices, &list);
> +	mutex_unlock(&devices_mutex);
> +
> +	list_for_each_entry_safe(dev, next, &list, list) {
>   		/* The module is being forcefully removed */
>   		WARN_ON(dev->desc);
>   		/* Remove from device list */
>   		list_del(&dev->list);
>   		ubiblock_cleanup(dev);
> -		kfree(dev);
>   	}
> -	mutex_unlock(&devices_mutex);
>   }
>   
>   int __init ubiblock_init(void)
> 
> ---
> base-commit: 44c026a73be8038f03dbdeef028b642880cf1511
> change-id: 20230523-ubiblock-remove-eab61cf683f0
> 
> Best regards,
>