Re: [PATCH] cfq: Take whether cfq group is changed into account when choosing service tree

[Gui Jianfeng]

Corrado Zoccolo wrote:
> Hi Gui,
> On Mon, Dec 14, 2009 at 10:54 AM, Gui Jianfeng
> <guijianfeng@xxxxxxxxxxxxxx> wrote:
>> Corrado Zoccolo wrote:
>>> Hi,
>> Currently, IIUC, only the workload that didn't use up its slice will be saved, and only
>> such workloads are restoring when a group is resumed. So sometimes, we'll still get the
>> previous serving_type and workload_expires. Am i missing something?
> You are right. cfq_choose_cfqg should set the workload as expired if
> !cfqg->saved_workload_slice (just set cfqd->workload_expires = jiffies
> - 1), so the workload will be chosen again as the lowest keyed one.
> Can you send a patch to fix this?

Will do.

Thanks
Gui

>>
>>> I have one more concern, though.
>>> RT priority has now changed meaning. Before, an RT task would always
>>> have priority access to the disk. Now, a BE task in a different group,
>>> with lower weight, can steal the disk from the RT task.
>>> A way to preserve the old meaning is to consider wheter a group has RT
>>> tasks inside when sorting groups tree, and putting those groups at the
>>> front.
>>> Usually, RT tasks will be put in the root group, and this (if
>>> group_isolation=0) will automatically make sure that also the noidle
>>> workload gets serviced quickly after RT tasks release the disk. We
>>> could even enforce that, with group_isolation=0, all RT tasks are put
>>> in the root group.
>>>
>>> The rationale behind this suggestion is that groups are for user
>>> processes, while RT is system wide, since it is only root that can
>>> grant it.
>>  I agree, and one more thing, currently we can't see fairness between different
>>  idle tasks in different groups. Because we only allow idle cfqq dispatch one request
>>  for its dispatch round even if it's the only task in the cgroup, group always loose it
>>  share. So whether we can rely on group_isolation, when group_isolation == 1 we provide
>>  isolation for idle tasks.
> Agreed.
> 
> Thanks,
> Corrado
> 
>> Thanks
>> Gui
>>
>>
>>
> 
> 
> 

-- 
Regards
Gui Jianfeng

--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/