Re: [PATCH v4 5/8] mtd: ubi: attach MTD partition from device-tree

[Richard Weinberger <richard@nod.at>]

----- Ursprüngliche Mail -----
> Von: "richard" <richard@nod.at>
> ----- Ursprüngliche Mail -----
>> diff --git a/drivers/mtd/ubi/block.c b/drivers/mtd/ubi/block.c
>> index e0618bbde3613..99b5f502c9dbc 100644
>> --- a/drivers/mtd/ubi/block.c
>> +++ b/drivers/mtd/ubi/block.c
>> @@ -470,7 +470,7 @@ int ubiblock_remove(struct ubi_volume_info *vi, bool force)
>> 	}
>> 
>> 	/* Found a device, let's lock it so we can check if it's busy */
>> -	mutex_lock(&dev->dev_mutex);
>> +	mutex_lock_nested(&dev->dev_mutex, SINGLE_DEPTH_NESTING);
> 
> The usage of mutex_lock_nested() in this patch looks fishy.
> Can you please elaborate a bit more why all these mutexes can be taken twice?
> (Any why not more often).

I think I figured myself.
ubiblock_ops->open() and ->release() are both called with disk->open_mutex held.
ubiblock_open() and ubiblock_release() take dev->dev_mutex.
So, the locking order is open_mutex, followed by dev_mutex.

On the other hand, ubiblock_remove() is called via UBI notify.
It takes first dev_mutex and then calls del_gendisk() which will trigger ubiblock_ops->release()
under disk->open_mutex but takes dev_mutex again.
So, we this not only takes a lock twice but also in reverse order.
mutex_lock_nested() might silence lockdep but I'm not sure whether this is safe at all.

Thanks,
//richard