Re: [PATCH v2] blk-iocost: fix shift-out-of-bounds in iocg_hick_delay()

From: Li Jinlin
Date: Mon Nov 28 2022 - 23:53:24 EST




On 2022/11/29 10:59, Yu Kuai wrote:
> Hi,
>
> 在 2022/11/29 10:49, Li Jinlin 写道:
>>
>>
>> On 2022/11/29 9:14, Yu Kuai wrote:
>>> Hi,
>>>
>>> 在 2022/11/29 3:58, Tejun Heo 写道:
>>>> On Mon, Nov 28, 2022 at 11:04:13AM +0800, Li Jinlin wrote:
>>>>>        /* calculate the current delay in effect - 1/2 every second */
>>>>>        tdelta = now->now - iocg->delay_at;
>>>>>        if (iocg->delay)
>>>>> -        delay = iocg->delay >> div64_u64(tdelta, USEC_PER_SEC);
>>>>> +        delay = iocg->delay >>
>>>>> +            min_t(u64, div64_u64(tdelta, USEC_PER_SEC), 63);
>>>>
>>>> I replied earlier but the right thing to do here is setting delay to 0 if
>>>> the shift is >= 64.
>>>
>>> Perhaps following change will make more sense?
>>>
>>> @@ -1322,18 +1323,19 @@ static bool iocg_kick_delay(struct ioc_gq *iocg, struct ioc_now *now)
>>>   {
>>>          struct ioc *ioc = iocg->ioc;
>>>          struct blkcg_gq *blkg = iocg_to_blkg(iocg);
>>> -       u64 tdelta, delay, new_delay;
>>> +       u64 delay = 0;
>>> +       u64 new_delay;
>>>          s64 vover, vover_pct;
>>>          u32 hwa;
>>>
>>>          lockdep_assert_held(&iocg->waitq.lock);
>>>
>>>          /* calculate the current delay in effect - 1/2 every second */
>>> -       tdelta = now->now - iocg->delay_at;
>>> -       if (iocg->delay)
>>> +       if (iocg->delay && now->now > iocg->delay_at) {
>>> +               u64 tdelta = now->now - iocg->delay_at;
>>> +
>>>                  delay = iocg->delay >> div64_u64(tdelta, USEC_PER_SEC);
>>> -       else
>>> -               delay = 0;
>>> +       }
>>>
>> I think "now->now > iocg->delay_at" is unnecessary, it is almost inevitable.
>
> From what I see, following can only happen if now->now < iocg->delay_at:
>
> "shift exponent 18446744073709"
>
You are right.

But I didn't see any ubsan reported at now->now - iocg->delay_at.
Need to confirm this.

Jinlin
Thanks.

> Or something else triggers it?
>
> Thanks,
> Kuai
>