Re: [syzbot] [block?] possible deadlock in elv_iosched_store

[Nilay Shroff <nilay@linux.ibm.com>]

On 4/1/25 5:46 PM, Ming Lei wrote:
> On Tue, Apr 01, 2025 at 05:23:56PM +0530, Nilay Shroff wrote:
>>
>>
>> On 3/29/25 7:29 AM, Ming Lei wrote:
>>> On Fri, Mar 28, 2025 at 07:37:25AM -0700, syzbot wrote:
>>>> Hello,
>>>>
>>>> syzbot found the following issue on:
>>>>
>>>> HEAD commit:    1a9239bb4253 Merge tag 'net-next-6.15' of git://git.kernel..
>>>> git tree:       upstream
>>>> console output: https://syzkaller.appspot.com/x/log.txt?x=1384b43f980000
>>>> kernel config:  https://syzkaller.appspot.com/x/.config?x=c7163a109ac459a8
>>>> dashboard link: https://syzkaller.appspot.com/bug?extid=4c7e0f9b94ad65811efb
>>>> compiler:       gcc (Debian 12.2.0-14) 12.2.0, GNU ld (GNU Binutils for Debian) 2.40
>>>> syz repro:      https://syzkaller.appspot.com/x/repro.syz?x=178cfa4c580000
>>>> C reproducer:   https://syzkaller.appspot.com/x/repro.c?x=11a8ca4c580000
>>>>
>>>> Downloadable assets:
>>>> disk image: https://storage.googleapis.com/syzbot-assets/fc7dc9f0d9a7/disk-1a9239bb.raw.xz
>>>> vmlinux: https://storage.googleapis.com/syzbot-assets/f555a3ae03d3/vmlinux-1a9239bb.xz
>>>> kernel image: https://storage.googleapis.com/syzbot-assets/55f6ea74eaf2/bzImage-1a9239bb.xz
>>>>
>>>> IMPORTANT: if you fix the issue, please add the following tag to the commit:
>>>> Reported-by: syzbot+4c7e0f9b94ad65811efb@syzkaller.appspotmail.com
>>>>
>>>
>>> ...
>>>
>>>>
>>>> If you want syzbot to run the reproducer, reply with:
>>>> #syz test: git://repo/address.git branch-or-commit-hash
>>>> If you attach or paste a git patch, syzbot will apply it before testing.
>>>
>>>
>>> diff --git a/block/blk-mq.c b/block/blk-mq.c
>>> index ae8494d88897..d7a103dc258b 100644
>>> --- a/block/blk-mq.c
>>> +++ b/block/blk-mq.c
>>> @@ -4465,14 +4465,12 @@ static struct blk_mq_hw_ctx *blk_mq_alloc_and_init_hctx(
>>>  	return NULL;
>>>  }
>>>  
>>> -static void blk_mq_realloc_hw_ctxs(struct blk_mq_tag_set *set,
>>> -						struct request_queue *q)
>>> +static void __blk_mq_realloc_hw_ctxs(struct blk_mq_tag_set *set,
>>> +				     struct request_queue *q)
>>>  {
>>>  	struct blk_mq_hw_ctx *hctx;
>>>  	unsigned long i, j;
>>>  
>>> -	/* protect against switching io scheduler  */
>>> -	mutex_lock(&q->elevator_lock);
>>>  	for (i = 0; i < set->nr_hw_queues; i++) {
>>>  		int old_node;
>>>  		int node = blk_mq_get_hctx_node(set, i);
>>> @@ -4505,7 +4503,19 @@ static void blk_mq_realloc_hw_ctxs(struct blk_mq_tag_set *set,
>>>  
>>>  	xa_for_each_start(&q->hctx_table, j, hctx, j)
>>>  		blk_mq_exit_hctx(q, set, hctx, j);
>>> -	mutex_unlock(&q->elevator_lock);
>>> +}
>>> +
>>> +static void blk_mq_realloc_hw_ctxs(struct blk_mq_tag_set *set,
>>> +				   struct request_queue *q, bool lock)
>>> +{
>>> +	if (lock) {
>>> +		/* protect against switching io scheduler  */
>>> +		mutex_lock(&q->elevator_lock);
>>> +		__blk_mq_realloc_hw_ctxs(set, q);
>>> +		mutex_unlock(&q->elevator_lock);
>>> +	} else {
>>> +		__blk_mq_realloc_hw_ctxs(set, q);
>>> +	}
>>>  
>>>  	/* unregister cpuhp callbacks for exited hctxs */
>>>  	blk_mq_remove_hw_queues_cpuhp(q);
>>> @@ -4537,7 +4547,7 @@ int blk_mq_init_allocated_queue(struct blk_mq_tag_set *set,
>>>  
>>>  	xa_init(&q->hctx_table);
>>>  
>>> -	blk_mq_realloc_hw_ctxs(set, q);
>>> +	blk_mq_realloc_hw_ctxs(set, q, false);
>>>  	if (!q->nr_hw_queues)
>>>  		goto err_hctxs;
>>>  
>>> @@ -5033,7 +5043,7 @@ static void __blk_mq_update_nr_hw_queues(struct blk_mq_tag_set *set,
>>>  fallback:
>>>  	blk_mq_update_queue_map(set);
>>>  	list_for_each_entry(q, &set->tag_list, tag_set_list) {
>>> -		blk_mq_realloc_hw_ctxs(set, q);
>>> +		blk_mq_realloc_hw_ctxs(set, q, true);
>>>  
>>>  		if (q->nr_hw_queues != set->nr_hw_queues) {
>>>  			int i = prev_nr_hw_queues;
>>>
>>
>> This patch looks good to me, however after we fix this one, I found another splat. 
>> I see that these new splats are side effect of commit ffa1e7ada456 ("block: Make 
>> request_queue lockdep splats show up earlier").
>>
>> IMO in the block layer code (unless it's in an IO submission path or a path where we 
>> have already frozen queue) we may still want to allow memory allocation with GFP_KERNEL. 
>> So in that sense, for example, we may acquire ->elevator_lock followed by fs_reclaim. 
> 
> If any memory GFP_KERNEL allocation grabs ->elevator_lock, it is one real
> deadlock risk.
> 
>> Or in another words, shouldn't it be legitimate to acquire blk layer specific lock and 
>> then allocate memory using GFP_KERNEL assuming we haven't freezed queue or we're not in 
>> IO submission path. But this commit ffa1e7ada456 ("block: Make request_queue lockdep
>> splats show up earlier") now showing up some false-positive splat as well, please see
>> below:
> 
> It depends if we may run GFP_KERNEL allocation with ->elevator_lock.
> 
Okay, so do you think we shall use GFP_NOIO for memory allocation if it's done after we
acquire ->elevator_lock?

Thanks,
--Nilay