Re: [PATCH V14 00/24] mmc: Add Command Queue support

[Ulf Hansson]

Linus, Adrian,

On 28 November 2017 at 10:42, Linus Walleij <linus.walleij@xxxxxxxxxx> wrote:
> On Tue, Nov 21, 2017 at 2:42 PM, Adrian Hunter <adrian.hunter@xxxxxxxxx> wrote:
>
>> Here is V14 of the hardware command queue patches without the software
>> command queue patches, now using blk-mq and now with blk-mq support for
>> non-CQE I/O.
>>
>> V14 includes a number of fixes to existing code, changes to default to
>> blk-mq, and adds patches to remove legacy code.
>
> I have looked over the code, I was unable to find a good mergebase to apply
> it on (I guess it is based on linux-next at some date in the past) so mostly
> I just looked at it overall, and I can solidly say that this patch series:
>
> Acked-by: Linus Walleij <linus.walleij@xxxxxxxxxx>

Great, thanks!

>
> I gave some more explicit review on some initial patches that I think
> should go in as fixes.

Thanks, and already taken care of.

>
> I do not expect it to perform any less than the previous iteration on my
> systems where it was already performing well and Bartlomiej also has
> confirmed that the patch set works for him.
>
> Ulf: I suggest this be applied (+/- some rebasing) early for v4.15.

Yes, I am up for that!

I have now also completed my review of the series and in the end, most
of my comments turned out to be of minor issues, hopefully easily
addressed.

>
> I am positively convinced that we can make things work on top of this.
>
>> HW CMDQ offers 25% - 50% better random multi-threaded I/O.  I see a slight
>> 2% drop in sequential read speed but no change to sequential write.
>
> Fully acceptable I think.
>
>> Non-CQE blk-mq showed a 3% decrease in sequential read performance.  This
>> seemed to be coming from the inferior latency of running work items compared
>> with a dedicated thread.  Hacking blk-mq workqueue to be unbound reduced the
>> performance degradation from 3% to 1%.
>
> Also acceptable I think.
>
>> While we should look at changing blk-mq to give better workqueue performance,
>> a bigger gain is likely to be made by adding a new host API to enable the
>> next already-prepared request to be issued directly from within ->done()
>> callback of the current request.

I assume that is taken care of by adding the new host cap
(MMC_CAP_DIRECT_COMPLETE).

So, then it's just a matter of adopting all host drivers, and when we
are done with that, we can remove that cap. :-)

>
> I agree.
>
> Yours,
> Linus Walleij

I am awaiting a re-based v15 version - and eager to apply it! :-)

Thanks and kind regards
Uffe