Re: [RFC PATCH 6/6] cpuset: Add cpuset.isolation_mask file

[Valentin Schneider <valentin.schneider-5wv7dgnIgG8@public.gmane.org>]

On 15/07/21 01:13, Frederic Weisbecker wrote:
> On Wed, Jul 14, 2021 at 06:52:43PM +0200, Peter Zijlstra wrote:
>>
>> cpusets already has means to create paritions; why are you creating
>> something else?
>
> I was about to answer that the semantics of isolcpus, which reference
> a NULL domain, are different from SD_LOAD_BALANCE implied by
> cpuset.sched_load_balance. But then I realize that SD_LOAD_BALANCE has
> been removed.
>
> How cpuset.sched_load_balance is implemented then? Commit
> e669ac8ab952df2f07dee1e1efbf40647d6de332 ("sched: Remove checks against
> SD_LOAD_BALANCE") advertize that setting cpuset.sched_load_balance to 0
> ends up creating NULL domain but that's not what I get. For example if I
> mount a single cpuset root (no other cpuset mountpoints):
>
> $ mount -t cgroup none ./cpuset -o cpuset
> $ cd cpuset
> $ cat cpuset.cpus
> 0-7
> $ cat cpuset.sched_load_balance
> 1
> $ echo 0 > cpuset.sched_load_balance
> $ ls /sys/kernel/debug/domains/cpu1/
> domain0  domain1
>
> I still get the domains on all CPUs...

Huh. That's on v5.14-rc1 with an automounted cpuset:

$ cat /sys/fs/cgroup/cpuset/cpuset.cpus
0-5
$ cat /sys/fs/cgroup/cpuset/cpuset.sched_load_balance
1

$ ls /sys/kernel/debug/sched/domains/cpu*
/sys/kernel/debug/sched/domains/cpu0:
domain0  domain1

/sys/kernel/debug/sched/domains/cpu1:
domain0  domain1

/sys/kernel/debug/sched/domains/cpu2:
domain0  domain1

/sys/kernel/debug/sched/domains/cpu3:
domain0  domain1

/sys/kernel/debug/sched/domains/cpu4:
domain0  domain1

/sys/kernel/debug/sched/domains/cpu5:
domain0  domain1

$ echo 0 > /sys/fs/cgroup/cpuset/cpuset.sched_load_balance
$ ls /sys/kernel/debug/sched/domains/cpu*
/sys/kernel/debug/sched/domains/cpu0:

/sys/kernel/debug/sched/domains/cpu1:

/sys/kernel/debug/sched/domains/cpu2:

/sys/kernel/debug/sched/domains/cpu3:

/sys/kernel/debug/sched/domains/cpu4:

/sys/kernel/debug/sched/domains/cpu5:


I also checked that you can keep cpuset.sched_load_balance=0 at the root
and create exclusive child cpusets with different values of
sched_load_balance, giving you some CPUs attached to the NULL domain and
some others with a sched_domain hierarchy that stays within the cpuset span.