Re: [PATCH] block: make sure last_lookup set as NULL after part deleted

[Hou Tao <houtao1@huawei.com>]

Hi,

On 2020/1/3 12:18, Ming Lei wrote:
> On Fri, Jan 03, 2020 at 11:06:25AM +0800, Hou Tao wrote:
>> Hi,
>>
>> On 2020/1/2 9:23, Ming Lei wrote:
>>> On Tue, Dec 31, 2019 at 10:55:47PM +0800, Hou Tao wrote:
>>>> Hi,
>>>>
snip

>> We have got a seemingly better solution: caching the index of last_lookup in tbl->part[]
>> instead of caching the pointer itself, so we can ensure the validity of returned pointer
>> by ensuring it's not NULL in tbl->part[] as does when last_lookup is NULL or 0.
> 
> Thinking of the problem further, looks we don't need to hold ref for
> .last_lookup.
> 
> What we need is to make sure the partition's ref is increased just
> before assigning .last_lookup, so how about something like the following?
> 
The approach will work for the above case, but it will not work for the following case:

when blk_account_io_done() releases the last ref-counter of last_lookup and calls call_rcu(),
and then a RCU read gets the to-be-freed hd-struct.

blk_account_io_done
  rcu_read_lock()
  // the last ref of last_lookup
  hd_struct_put()
    call_rcu

                              rcu_read_lock
                              read last_lookup

    free()
                              // use-after-free ?
                              hd_struct_try_get

Regards,
Tao

> diff --git a/block/blk-core.c b/block/blk-core.c
> index 089e890ab208..79599f5fd5b7 100644
> --- a/block/blk-core.c
> +++ b/block/blk-core.c
> @@ -1365,18 +1365,6 @@ void blk_account_io_start(struct request *rq, bool new_io)
>  		part_stat_inc(part, merges[rw]);
>  	} else {
>  		part = disk_map_sector_rcu(rq->rq_disk, blk_rq_pos(rq));
> -		if (!hd_struct_try_get(part)) {
> -			/*
> -			 * The partition is already being removed,
> -			 * the request will be accounted on the disk only
> -			 *
> -			 * We take a reference on disk->part0 although that
> -			 * partition will never be deleted, so we can treat
> -			 * it as any other partition.
> -			 */
> -			part = &rq->rq_disk->part0;
> -			hd_struct_get(part);
> -		}
>  		part_inc_in_flight(rq->q, part, rw);
>  		rq->part = part;
>  	}
> diff --git a/block/genhd.c b/block/genhd.c
> index ff6268970ddc..21f4a9b8d24d 100644
> --- a/block/genhd.c
> +++ b/block/genhd.c
> @@ -286,17 +286,24 @@ struct hd_struct *disk_map_sector_rcu(struct gendisk *disk, sector_t sector)
>  	ptbl = rcu_dereference(disk->part_tbl);
>  
>  	part = rcu_dereference(ptbl->last_lookup);
> -	if (part && sector_in_part(part, sector))
> +	if (part && sector_in_part(part, sector)) {
> +		if (!hd_struct_try_get(part))
> +			goto exit;
>  		return part;
> +	}
>  
>  	for (i = 1; i < ptbl->len; i++) {
>  		part = rcu_dereference(ptbl->part[i]);
>  
>  		if (part && sector_in_part(part, sector)) {
> +                       if (!hd_struct_try_get(part))
> +                               goto exit;
>  			rcu_assign_pointer(ptbl->last_lookup, part);
>  			return part;
>  		}
>  	}
> + exit:
> +	hd_struct_get(&disk->part0);
>  	return &disk->part0;
>  }
>  EXPORT_SYMBOL_GPL(disk_map_sector_rcu);
> 
> 
>>
>>> Given partition is actually protected by percpu-refcount now, I guess the
>>> RCU annotation for referencing ->part[partno] and ->last_lookup may not
>>> be necessary, together with the part->rcu_work.
>>>
>> So we will depends on the invocation of of call_rcu() on __percpu_ref_switch_mode() to
>> ensure the RCU readers will find part[i] is NULL before trying to increasing
>> the atomic ref-counter of part[i], right ?
> 
> Yeah.
> 
> Thanks,
> Ming
> 
> 
> .
>