[PATCH] um: implement arch_sync_kernel_mappings

[Anton Ivanov]

On 16/03/2021 11:04, Johannes Berg wrote:
> On Tue, 2021-03-16 at 10:49 +0000, Anton Ivanov wrote:
>>>> I get slightly better performance on userspace,
>>>> etc compared to your patch. It is marginal -
>>>> sub 1% close to the experimental error.
>>>
>>> Why does that even affect userspace at all, btw? We're talking about
>>> kernel mappings?
>>
>> My userspace tests do heavy IO.
> 
> OK, but even then, my patch should be an improvement on anything that
> needs updated ranges since it avoids the segfault(s), and not really
> make anything worse that *doesn't*?
> 
> But then again, I'm not sure you said that this patch was an improvement
> or not for your workload?

Your patch showed no difference on heavy IO userspace. Roughly the same numbers I usually get.

Syncing pages at a PMD level change showed some marginal improvement on a heavy IO testcase but very close to the margin of error - in the realm of 0.2-0.3%

I would really need to do 20+ runs to have reliable numbers on that to confirm as it is too close to the margin of error.

> 
> johannes
> 
> 

-- 
Anton R. Ivanov
Cambridgegreys Limited. Registered in England. Company Number 10273661
https://www.cambridgegreys.com/