Re: [RFC PATCH 0/4] x86/mm/cpa: merge small mappings whenever possible

[Aaron Lu]

Hi Kirill,

Thanks a lot for the feedback.

On 8/9/2022 6:04 PM, Kirill A. Shutemov wrote:
> On Mon, Aug 08, 2022 at 10:56:45PM +0800, Aaron Lu wrote:
>> This is an early RFC. While all reviews are welcome, reviewing this code
>> now will be a waste of time for the x86 subsystem maintainers. I would,
>> however, appreciate a preliminary review from the folks on the to and cc
>> list. I'm posting it to the list in case anyone else is interested in
>> seeing this early version.
> 
> Last time[1] I tried to merge pages back in direct mapping it lead to
> substantial performance regression for some workloads. I cannot find the
> report right now, but I remember it was something graphics related.
>

Do you happen to remember the workload name? I can try running it.

> Have you done any performance evaluation?
> 

Not yet, I was mostly concentrating on correctness. In addition to the
graphics workload, do you have anything else in mind that may be
sensitive to such a change?

I think maybe I can run patch4's mode0 test with and without this merge
functionality and see how performance would change since mode0 is
essentially doing various set_memory_X() calls on different cpus
simultaneously which can trigger a lot of splits and merges. Sounds good?

> My take away was that the merge has to be batched. Like log where changes
> to direct mapping happens and come back to then and merge when the number
> of changes cross some limit.
> 

Appreciate your suggestion.

> Also I don't see you handling set_memory_4k(). Huh?
>

Ah Right, I missed that. Currently set_memory_4k() is not specially
handled and can be mistakenly merged. Will fix this in later versions.

> [1] https://lore.kernel.org/lkml/20200416213229.19174-1-kirill.shutemov@xxxxxxxxxxxxxxx/
> 

Thanks!