Re: [PATCH V1] mmc: core: fix still flush cache when eMMC cache off

[Ulf Hansson]

On 23 March 2017 at 11:45, Bean Huo (beanhuo) <beanhuo@xxxxxxxxxx> wrote:
> Hi,
>
>>On 19 March 2017 at 01:45, Bean Huo (beanhuo) <beanhuo@xxxxxxxxxx> wrote:
>>> This patch fixes the issue that mmc_blk_issue_rq still flushes cache
>>> when eMMC cache has already been off through user space tool, such as
>>> mmc-utils.
>>> The reason is that card->ext_csd.cache_ctrl isn't reset.
>>
>>First, why do you want to turn of the cache ctrl? Is the eMMC device having
>>issues with it? Then we should invent a card quirk instead.
>
>
> Why I turn off it? because I did power loss testing and validation, I should switch it off and on.
> When I do some performance and power loss case validation on several Linux release versions,
> I need to switch off or on cache through user space tool.
> I can't confirm every user that likes me, But I think at least it is not reasonable to
> flush eMMC cache, when internal eMMC cache is off.

Ah, I see. Your use-case seems reasonable while validating robustness
of the eMMC!

>
>>Second, what errors do you encounter when the mmc core tries to flush the
>>cache when it has been turned off? Can you please elaborate on this?
>
>
> No error found, but firstly, please think about overhead introduced by useless flush cache, Unless you
> Don't care this tiny time. second, under the condition of cache off, additional flush cache request still has impact on
> Internal eMMC logic. I don't know What and how impact, but at least it is really exist.

Got it!

However I still don't like the mmc ioctls API to compensate and deal
with all "crazy-ness" that user-space may cause. Cache-ctrl is only
one case out of many.

I see two viable options to solve your problem.
1) Extend mmc_test with a new test(s) for cache ctrl and perhaps
suspend/resume. Isn't this actually exactly what you want?
2) Extend debugfs to be able to turn cache ctrl on/off.

[...]

Kind regards
Uffe