Re: Another approach of UFSHPB

From: Daejun Park
Date: Wed May 27 2020 - 05:15:27 EST


On 2020-05-26 Bart Van Assche wrote:
>On 2020-05-25 23:15, Avri Altman wrote:
>>> On 2020-05-24 22:40, Daejun Park wrote:
>>>> The HPB driver is close to the UFS core function, but it is not essential
>>>> for operating UFS device. With reference to this article
>>>> (https://lwn.net/Articles/645810/), we implemented extended UFS-feature
>>>> as bus model. Because the HPB driver consumes the user's main memory, it
>>> should
>>>> support bind / unbind functionality as needed. We implemented the HPB
>>> driver
>>>> can be unbind / unload on runtime.
>>>
>>> I do not agree that the bus model is the best choice for freeing cache
>>> memory if it is no longer needed. A shrinker is probably a much better
>>> choice because the callback functions in a shrinker get invoked when a
>>> system is under memory pressure. See also register_shrinker(),
>>> unregister_shrinker() and struct shrinker in include/linux/shrinker.h.
>>
>> Since this discussion is closely related to cache allocation,
>> What is your opinion about allocating the pages dynamically as the regions
>> Are being activated/deactivated, in oppose of how it is done today -
>> Statically on init for the entire max-active-subregions?

> Memory that is statically allocated cannot be used for any other purpose
> (e.g. page cache) without triggering the associated shrinker. As far as
> I know shrinkers are only triggered when (close to) out of memory. So
> dynamically allocating memory as needed is probably a better strategy
> than statically allocating the entire region at initialization time.

To improve UFS device performance using the HPB driver,
the number of active-subregions above a certain threshold is essential.
If the number of active-subregions is lower than the threshold,
the performance improvement by using HPB will be reduced.
Also, due to frequent and active/inactive protocol overhead,
performance may be worse than when the HPB feature is not used.

Therefore, it is better to unbind/unload HPB driver than
to reduce the number of active subregions below the threshold.
We designed the HPB driver to make the UFS device work
even when the module is unloaded.

Thanks,

Daejun