Re: [PATCH -next v10 3/4] block, bfq: refactor the counting of 'num_groups_with_pending_reqs'

[Yu Kuai <yukuai1@huaweicloud.com>]

Hi,

在 2022/09/14 17:00, Jan Kara 写道:
> Hi guys!
> 
> On Wed 14-09-22 16:15:26, Yu Kuai wrote:
>> 在 2022/09/14 15:50, Paolo VALENTE 写道:
>>>
>>>
>>>> Il giorno 14 set 2022, alle ore 03:55, Yu Kuai <yukuai1@huaweicloud.com> ha scritto:
>>>>
>>>>
>>>>
>>>> 在 2022/09/07 9:16, Yu Kuai 写道:
>>>>> Hi, Paolo!
>>>>> 在 2022/09/06 17:37, Paolo Valente 写道:
>>>>>>
>>>>>>
>>>>>>> Il giorno 26 ago 2022, alle ore 04:34, Yu Kuai <yukuai1@huaweicloud.com> ha scritto:
>>>>>>>
>>>>>>> Hi, Paolo!
>>>>>>>
>>>>>>> 在 2022/08/25 22:59, Paolo Valente 写道:
>>>>>>>>> Il giorno 11 ago 2022, alle ore 03:19, Yu Kuai <yukuai1@huaweicloud.com <mailto:yukuai1@huaweicloud.com>> ha scritto:
>>>>>>>>>
>>>>>>>>> Hi, Paolo
>>>>>>>>>
>>>>>>>>> 在 2022/08/10 18:49, Paolo Valente 写道:
>>>>>>>>>>> Il giorno 27 lug 2022, alle ore 14:11, Yu Kuai <yukuai1@huaweicloud.com <mailto:yukuai1@huaweicloud.com>> ha scritto:
>>>>>>>>>>>
>>>>>>>>>>> Hi, Paolo
>>>>>>>>>>>
>>>>>>>>>> hi
>>>>>>>>>>> Are you still interested in this patchset?
>>>>>>>>>>>
>>>>>>>>>> Yes. Sorry for replying very late again.
>>>>>>>>>> Probably the last fix that you suggest is enough, but I'm a little bit
>>>>>>>>>> concerned that it may be a little hasty.  In fact, before this fix, we
>>>>>>>>>> exchanged several messages, and I didn't seem to be very good at
>>>>>>>>>> convincing you about the need to keep into account also in-service
>>>>>>>>>> I/O.  So, my question is: are you sure that now you have a
>>>>>>>>>
>>>>>>>>> I'm confused here, I'm pretty aware that in-service I/O(as said pending
>>>>>>>>> requests is the patchset) should be counted, as you suggested in v7, are
>>>>>>>>> you still thinking that the way in this patchset is problematic?
>>>>>>>>>
>>>>>>>>> I'll try to explain again that how to track is bfqq has pending pending
>>>>>>>>> requests, please let me know if you still think there are some problems:
>>>>>>>>>
>>>>>>>>> patch 1 support to track if bfqq has pending requests, it's
>>>>>>>>> done by setting the flag 'entity->in_groups_with_pending_reqs' when the
>>>>>>>>> first request is inserted to bfqq, and it's cleared when the last
>>>>>>>>> request is completed. specifically the flag is set in
>>>>>>>>> bfq_add_bfqq_busy() when 'bfqq->dispatched' if false, and it's cleared
>>>>>>>>> both in bfq_completed_request() and bfq_del_bfqq_busy() when
>>>>>>>>> 'bfqq->diapatched' is false.
>>>>>>>>>
>>>>>>>> This general description seems correct to me. Have you already sent a new version of your patchset?
>>>>>>>
>>>>>>> It's glad that we finially on the same page here.
>>>>>>>
>>>>>>
>>>>>> Yep. Sorry for my chronicle delay.
>>>>> Better late than never 😁
>>>>>>
>>>>>>> Please take a look at patch 1, which already impelement the above
>>>>>>> descriptions, it seems to me there is no need to send a new version
>>>>>>> for now. If you think there are still some other problems, please let
>>>>>>> me know.
>>>>>>>
>>>>>>
>>>>>> Patch 1 seems ok to me. I seem to have only one pending comment on this patch (3/4) instead. Let me paste previous stuff here for your convenience:
>>>>> That sounds good.
>>>>>>
>>>>>>>>
>>>>>>>> -    /*
>>>>>>>> -     * Next function is invoked last, because it causes bfqq to be
>>>>>>>> -     * freed if the following holds: bfqq is not in service and
>>>>>>>> -     * has no dispatched request. DO NOT use bfqq after the next
>>>>>>>> -     * function invocation.
>>>>>>>> -     */
>>>>>>> I would really love it if you leave this comment.  I added it after
>>>>>>> suffering a lot for a nasty UAF.  Of course the first sentence may
>>>>>>> need to be adjusted if the code that precedes it is to be removed.
>>>>>>> Same as above, if this patch is applied, this function will be gone.
>>>>
>>>> Hi, I'm curious while I'm trying to add the comment, before this
>>>> patchset, can bfqq be freed when bfq_weights_tree_remove is called?
>>>>
>>>> bfq_completed_request
>>>> bfqq->dispatched--
>>>> if (!bfqq->dispatched && !bfq_bfqq_busy(bfqq))
>>>>    bfq_weights_tree_remove(bfqd, bfqq);
>>>>
>>>> // continue to use bfqq
>>>>
>>>> It seems to me this is problematic if so, because bfqq is used after
>>>> bfq_weights_tree_remove() is called.
>>>>
>>>
>>> It is.  Yet, IIRC, I verified that bfqq was not used after that free,
>>> and I added that comment as a heads-up.  What is a scenario (before
>>> your pending modifications) where this use-after-free happens?
>>>
>>
>> No, it never happens, I just notice it because it'll be weird if I
>> place the comment where bfq_weights_tree_remove() is called, since bfqq
>> will still be accessed.
>>
>> If the suituation that the comment says is possible, perhaps we should
>> move bfq_weights_tree_remove() to the last of bfq_completed_request().
>> However, it seems that we haven't meet the problem for quite a long
>> time...
> 
> I'm bit confused which comment you are speaking about but
> bfq_completed_request() gets called only from bfq_finish_requeue_request()
> and the request itself still holds a reference to bfqq. Only later in
> bfq_finish_requeue_request() when we do:
> 
> 	bfqq_request_freed(bfqq);
> 	bfq_put_queue(bfqq);
> 
> bfqq can get freed.

Yes, you're right. Then I think the only place that
bfq_weights_tree_remove() can free bfqq is from bfq_del_bfqq_busy().
I'll move the following comment with a little adjustment here, which is
from bfq_weights_tree_remove() before this patchset:

         /*
         ┊* Next function is invoked last, because it causes bfqq to be
         ┊* freed. DO NOT use bfqq after the next function invocation.
         ┊*/

Thanks,
Kuai

> 
> 								Honza
>