Re: [PATCH v2 1/2] sched/isolation: Exclude dynamically isolated CPUs from housekeeping masks

[Waiman Long <llong@redhat.com>]

On 11/15/24 2:32 PM, Waiman Long wrote:
> On 11/15/24 10:45 AM, Michal Koutný wrote:
>> Hello.
>>
>> I recently liked the idea of considering isolated CPUs a static (boot
>> time) resource and only use cpusets to place (or remove) sensitive
>> workload from those selected CPUs depending on current needs. (Yes, this
>> may not efficiently utilize the isolated CPUs when reserve them based on
>> maximum needs of a node.)
>
> Thanks for taking a look at this.
>
> Yes, I am moving in this direction too. Boot time statically isolated 
> CPUs have better isolation than is currently possible if we do it 
> dynamically at run time, though we are trying to close the gap.
>
>>
>>
>> On Wed, Aug 21, 2024 at 10:23:11AM GMT, Waiman Long 
>> <longman@redhat.com> wrote:
>>> This patch is a step in that direction by making the housekeeping CPU
>>> mask APIs exclude the dynamically isolated CPUs when they are called
>>> at run time. The housekeeping CPU masks will fall back to the bootup
>>> default when all the dynamically isolated CPUs are released.
>> But when I look at it with the dynamism in mind, I would expect that
>> some API like housekeeping_setup_type(), i.e. modify the set of isolated
>> CPUs are requested and leave it up to the isolation implementation to
>> propagate any changes to respective subsystems. And return an error of
>> type contains a flag for which dynamism isn't implemented yet or not
>> possible.
>
> There are currently 9 different hk_type's defined in 
> include/linux/sched/isolation.h. We are now trying to reduce their 
> number as some of them cannot be set independently. See [1]. I am 
> thinking about doing dynamism in the best effort basis. Of course, we 
> could expose some information about what aspect of dynamic isolation 
> can be enabled at the moment, if necessary.

Forgot to put the link.

[1] https://lore.kernel.org/lkml/20240921190720.106195-1-longman@redhat.com/

>
> Cheers,
> Longman
>