Re: [PATCH 13/15] sched/isolation: Implement sysfs interface for dynamic housekeeping

Re: [PATCH 13/15] sched/isolation: Implement sysfs interface for dynamic housekeeping

[Qiliang Yuan]

On Wed, Mar 25, 2026 at 03:04:32PM +0100, Peter Zijlstra wrote:
> Why? What was wrong with cpusets?

This is the central point of the architecture. The distinction I was 
trying to address is:

1. Task Isolation (Current CPUSets):
   The `cpuset` subsystem (especially `cpuset.cpus.partition = isolated`) 
   is excellent at managing task placement and load balancing. It 
   ensures no user tasks are pushed to isolated CPUs.

2. Kernel Overhead Isolation (Housekeeping):
   Currently, `cpusets` do not manage kernel-internal overhead like RCU 
   callbacks, timers, or unbound workqueues. These are managed by the 
   global `housekeeping_cpumask`, which is settled at boot via 
   `isolcpus`/`nohz_full` and is static.

DHEI fills this second gap by making the housekeeping mask dynamic. 
However, I agree that a parallel sysfs interface is redundant.

In V13, I will move the control interface to `cpuset`. The root cpuset 
will serve as the primary interface, allowing changes in the cpuset 
partition state to automatically trigger the migration of kernel 
housekeeping overhead. This achieves "Full Dynamic Isolation" (both tasks 
and kernel overhead) through a single, unified interface.

Best regards,
Qiliang

Re: [PATCH 13/15] sched/isolation: Implement sysfs interface for dynamic housekeeping

[Tejun Heo]

cc'ing Waiman.

On Mon, Mar 30, 2026 at 07:46:20PM +0800, Qiliang Yuan wrote:
> On Wed, Mar 25, 2026 at 03:04:32PM +0100, Peter Zijlstra wrote:
> > Why? What was wrong with cpusets?
> 
> This is the central point of the architecture. The distinction I was 
> trying to address is:
> 
> 1. Task Isolation (Current CPUSets):
>    The `cpuset` subsystem (especially `cpuset.cpus.partition = isolated`) 
>    is excellent at managing task placement and load balancing. It 
>    ensures no user tasks are pushed to isolated CPUs.
> 
> 2. Kernel Overhead Isolation (Housekeeping):
>    Currently, `cpusets` do not manage kernel-internal overhead like RCU 
>    callbacks, timers, or unbound workqueues. These are managed by the 
>    global `housekeeping_cpumask`, which is settled at boot via 
>    `isolcpus`/`nohz_full` and is static.
> 
> DHEI fills this second gap by making the housekeeping mask dynamic. 
> However, I agree that a parallel sysfs interface is redundant.
> 
> In V13, I will move the control interface to `cpuset`. The root cpuset 
> will serve as the primary interface, allowing changes in the cpuset 
> partition state to automatically trigger the migration of kernel 
> housekeeping overhead. This achieves "Full Dynamic Isolation" (both tasks 
> and kernel overhead) through a single, unified interface.

Please discuss with Frederic and Waiman first because they have been working
towards making cpuset to cover what cpuisol does. I don't think we want two
separate mechanisms for the same thing and don't see why this would need to
be its own thing when it has to be coupled with task isolation anyway.

Thanks.

-- 
tejun

Re: [PATCH 13/15] sched/isolation: Implement sysfs interface for dynamic housekeeping

[Waiman Long]

On 3/30/26 7:46 AM, Qiliang Yuan wrote:
> On Wed, Mar 25, 2026 at 03:04:32PM +0100, Peter Zijlstra wrote:
>> Why? What was wrong with cpusets?
> This is the central point of the architecture. The distinction I was
> trying to address is:
>
> 1. Task Isolation (Current CPUSets):
>     The `cpuset` subsystem (especially `cpuset.cpus.partition = isolated`)
>     is excellent at managing task placement and load balancing. It
>     ensures no user tasks are pushed to isolated CPUs.
>
> 2. Kernel Overhead Isolation (Housekeeping):
>     Currently, `cpusets` do not manage kernel-internal overhead like RCU
>     callbacks, timers, or unbound workqueues. These are managed by the
>     global `housekeeping_cpumask`, which is settled at boot via
>     `isolcpus`/`nohz_full` and is static.

My plan is to extend the cpuset isolated partition mechanism to isolate 
other kernel noise currently covered by the nohz_full and manged_irq 
boot command line. That will makes the HK_TYPE_KERNEL_NOISE cpumask 
modifiable at run time. It is not a direct modification of the HK 
cpumasks as advocated by this patch series but an indirect one via the 
creation of the appropriate isolated cpuset partitions.

Cheers,
Longman