[PATCH v2] mtd: break circular locks in register_mtd_blktrans

[Desmond Cheong Zhi Xi]

On 18/6/21 12:09 am, Desmond Cheong Zhi Xi wrote:
> Syzbot reported a circular locking dependency:
> https://syzkaller.appspot.com/bug?id=7bd106c28e846d1023d4ca915718b1a0905444cb
> 
> This happens because of the following lock dependencies:
> 
> 1. loop_ctl_mutex -> bdev->bd_mutex (when loop_control_ioctl calls
> loop_remove, which then calls del_gendisk; this also happens in
> loop_exit which eventually calls loop_remove)
> 
> 2. bdev->bd_mutex -> mtd_table_mutex (when blkdev_get_by_dev calls
> __blkdev_get, which then calls blktrans_open)
> 
> 3. mtd_table_mutex -> major_names_lock (when register_mtd_blktrans
> calls __register_blkdev)
> 
> 4. major_names_lock -> loop_ctl_mutex (when blk_request_module calls
> loop_probe)
> 
> Hence there's an overall dependency of:
> 
> loop_ctl_mutex   ----------> bdev->bd_mutex
>        ^                            |
>        |                            |
>        |                            v
> major_names_lock <---------  mtd_table_mutex
> 
> We can break this circular dependency by holding mtd_table_mutex only
> for the required critical section in register_mtd_blktrans. This
> avoids the mtd_table_mutex -> major_names_lock dependency.
> 
> Reported-and-tested-by: syzbot+6a8a0d93c91e8fbf2e80 at syzkaller.appspotmail.com
> Co-developed-by: Christoph Hellwig <hch at lst.de>
> Signed-off-by: Christoph Hellwig <hch at lst.de>
> Signed-off-by: Desmond Cheong Zhi Xi <desmondcheongzx at gmail.com>
> ---
> 
> Changes in v1 -> v2:
> 
> Break the circular dependency in register_mtd_blktrans instead of blk_request_module, as suggested by Christoph Hellwig.
> 
>   drivers/mtd/mtd_blkdevs.c | 8 ++------
>   1 file changed, 2 insertions(+), 6 deletions(-)
> 
> diff --git a/drivers/mtd/mtd_blkdevs.c b/drivers/mtd/mtd_blkdevs.c
> index fb8e12d590a1..7d26cfe24d05 100644
> --- a/drivers/mtd/mtd_blkdevs.c
> +++ b/drivers/mtd/mtd_blkdevs.c
> @@ -528,14 +528,10 @@ int register_mtd_blktrans(struct mtd_blktrans_ops *tr)
>   	if (!blktrans_notifier.list.next)
>   		register_mtd_user(&blktrans_notifier);
>   
> -
> -	mutex_lock(&mtd_table_mutex);
> -
>   	ret = register_blkdev(tr->major, tr->name);
>   	if (ret < 0) {
>   		printk(KERN_WARNING "Unable to register %s block device on major %d: %d\n",
>   		       tr->name, tr->major, ret);
> -		mutex_unlock(&mtd_table_mutex);
>   		return ret;
>   	}
>   
> @@ -545,12 +541,12 @@ int register_mtd_blktrans(struct mtd_blktrans_ops *tr)
>   	tr->blkshift = ffs(tr->blksize) - 1;
>   
>   	INIT_LIST_HEAD(&tr->devs);
> -	list_add(&tr->list, &blktrans_majors);
>   
> +	mutex_lock(&mtd_table_mutex);
> +	list_add(&tr->list, &blktrans_majors);
>   	mtd_for_each_device(mtd)
>   		if (mtd->type != MTD_ABSENT)
>   			tr->add_mtd(tr, mtd);
> -
>   	mutex_unlock(&mtd_table_mutex);
>   	return 0;
>   }
> 

Hi maintainers,

Any chance to review this patch?

For additional reference, the mtd_table_mutex --> major_names_lock 
hierarchy that can be removed by this patch also appears in a different 
lock chain:
https://syzkaller.appspot.com/bug?id=cbf5fe846f14a90f05e10df200b08c57941dc750

Best wishes,
Desmond