Re: [PATCH] blk-ioprio: Introduce promote-to-rt policy

[Hou Tao <houtao@huaweicloud.com>]

Hi,

On 2/2/2023 1:33 AM, Bart Van Assche wrote:
> On 1/31/23 20:52, Hou Tao wrote:
>>   /**
>>    * enum prio_policy - I/O priority class policy.
>>    * @POLICY_NO_CHANGE: (default) do not modify the I/O priority class.
>> @@ -27,21 +34,30 @@
>>    * @POLICY_RESTRICT_TO_BE: modify IOPRIO_CLASS_NONE and IOPRIO_CLASS_RT into
>>    *        IOPRIO_CLASS_BE.
>>    * @POLICY_ALL_TO_IDLE: change the I/O priority class into IOPRIO_CLASS_IDLE.
>> - *
>> + * @POLICY_PROMOTE_TO_RT: modify IOPRIO_CLASS_NONE and IOPRIO_CLASS_BE into
>> + *         IOPRIO_CLASS_RT.
>>    * See also <linux/ioprio.h>.
>>    */
>>   enum prio_policy {
>> -    POLICY_NO_CHANGE    = 0,
>> -    POLICY_NONE_TO_RT    = 1,
>> -    POLICY_RESTRICT_TO_BE    = 2,
>> -    POLICY_ALL_TO_IDLE    = 3,
>> +    POLICY_NO_CHANGE    = IOPRIO_CLASS_NONE,
>> +    POLICY_NONE_TO_RT    = IOPRIO_CLASS_RT,
>> +    POLICY_RESTRICT_TO_BE    = IOPRIO_CLASS_BE,
>> +    POLICY_ALL_TO_IDLE    = IOPRIO_CLASS_IDLE,
>> +    POLICY_PROMOTE_TO_RT    = IOPRIO_CLASS_RT | IOPRIO_POL_PROMOTION,
>> +};
>
> The above change complicates the ioprio code. Additionally, I'm concerned that
> it makes the ioprio code slower. Has it been considered to keep the numerical
> values for the existing policies, to assign the number 4 to
> POLICY_PROMOTE_TO_RT and to use a lookup-array in blkcg_set_ioprio() to
> convert the policy number into an IOPRIO_CLASS value?
For the slowness, do you meaning the extra dereference of blkcg->ioprio->policy
when policy is no-change or the handle of IOPRIO_POL_PROMOTION in
blkcg_set_ioprio()? It seems other functions (e.g., ioprio_show_prio_policy()
and ioprio_set_prio_policy()) are not on the hot path. Using a lookup array in
blkcg_set_ioprio() to do the conversion will also be OK, although it will
introduce an extra lookup each time when policy is not no-change. I don't have
strong preference. If you are OK with lookup array in blkcg_set_ioprio(), will
do it in v2.
>
> Thanks,
>
> Bart.
>
>
> .