Re: [PATCH RFT] blk-mq: optimize queue tag busy iter for shared_tags

[John Garry]

On 22/12/2021 12:06, Kashyap Desai wrote:
> > But I did not think that my patch would help mpi3mr since it does not use
> > host_tagset.
> Internally we are testing performance check for mpi3mr. We want to make sure
> performance is not impacted due to shared host_tagset before we apply the
> feature.

Hmmm... I thought that you said previously that it was not necessary for 
this HW. But I am not familiar with the driver or HW.

As an aside, I assume that this driver uses none IO sched by default, 
but drivers using host_tagset use mq-deadline – I’m not sure if that is 
what you want.

> > > We can drop
> > > request of this RFT since I tested above series and it serve the same
> > > purpose.
> > ok, fine.
> >
> > And just to confirm, do you now think that we need to fix any older kernel
> > with some backport of my changes? I think that we would just need to
> > consider 5.16 (when it becomes stable), 5.15, and and 5.10
> It need full patch set (below) + associated fixes.
> [1] " blk-mq: Use shared tags for shared sbitmap support" Commit -
> e155b0c238b20f0a866f4334d292656665836c8a
>
> Below commit cannot go to stable without [1].
> https://git.kernel.org/pub/scm/linux/kernel/git/axboe/linux-block.git/commit/?h=for-5.17/block&id=fea9f92f1748083cb82049ed503be30c3d3a9b69
>
> I am not sure if stable requirement fits in this case. I mean large
> patch-set is OK ?

The two other patches in the series wouldn't need to be backported, so 
it should be possible. You would just need a very good reason, though. 
And we would need to know whether 5.10 and 5.15 are required - they use 
shared sbitmap. As I mentioned a few times, contention in 
blk_mq_queue_tag_busy_iter() and callees would not be so high as 
blk_mq_tags.lock and blk_mq_tags.rqs are not shared there.

Thanks,
John