Re: [PATCH v5 4/6] cgroup/cpuset: Don't update isolated_cpus from CPU hotplug

[Waiman Long <llong@redhat.com>]

On 2/12/26 10:28 PM, Chen Ridong wrote:
>
> On 2026/2/13 0:46, Waiman Long wrote:
>> As any change to isolated_cpus is going to be propagated to the
>> HK_TYPE_DOMAIN housekeeping cpumask, it can be problematic if
>> housekeeping cpumasks are directly being modified from the CPU hotplug
>> code path. This is especially the case if we are going to enable dynamic
>> update to the nohz_full housekeeping cpumask (HK_TYPE_KERNEL_NOISE)
>> in the near future with the help of CPU hotplug.
>>
>> Avoid these potential problems by changing the cpuset code to not
>> updating isolated_cpus when calling from CPU hotplug. A new special
>> PRS_INVALID_ISOLCPUS is added to indicate the current cpuset is an
>> invalid partition but its effective_xcpus are still in isolated_cpus.
>> This special state will be set if an isolated partition becomes invalid
>> due to the shutdown of the last active CPU in that partition. We also
>> need to keep the effective_xcpus even if exclusive_cpus isn't set.
>>
>> When changes are made to "cpuset.cpus", "cpuset.cpus.exclusive" or
>> "cpuset.cpus.partition" of a PRS_INVALID_ISOLCPUS cpuset, its state
>> will be reset back to PRS_INVALID_ISOLATED and its effective_xcpus will
>> be removed from isolated_cpus before proceeding.
>>
>> As CPU hotplug will no longer update isolated_cpus, some of the test
>> cases in test_cpuset_prs.h will have to be updated to match the new
>> expected results. Some new test cases are also added to confirm that
>> "cpuset.cpus.isolated" and HK_TYPE_DOMAIN housekeeping cpumask will
>> both be updated.
>>
>> Signed-off-by: Waiman Long <longman@redhat.com>
>> ---
>>   kernel/cgroup/cpuset.c                        | 85 ++++++++++++++++---
>>   .../selftests/cgroup/test_cpuset_prs.sh       | 21 +++--
>>   2 files changed, 87 insertions(+), 19 deletions(-)
>>
>> diff --git a/kernel/cgroup/cpuset.c b/kernel/cgroup/cpuset.c
>> index c792380f9b60..48b7f275085b 100644
>> --- a/kernel/cgroup/cpuset.c
>> +++ b/kernel/cgroup/cpuset.c
>> @@ -159,6 +159,8 @@ static bool force_sd_rebuild;			/* RWCS */
>>    *   2 - partition root without load balancing (isolated)
>>    *  -1 - invalid partition root
>>    *  -2 - invalid isolated partition root
>> + *  -3 - invalid isolated partition root but with effective xcpus still
>> + *	 in isolated_cpus (set from CPU hotplug side)
>>    *
>>    *  There are 2 types of partitions - local or remote. Local partitions are
>>    *  those whose parents are partition root themselves. Setting of
>> @@ -187,6 +189,7 @@ static bool force_sd_rebuild;			/* RWCS */
>>   #define PRS_ISOLATED		2
>>   #define PRS_INVALID_ROOT	-1
>>   #define PRS_INVALID_ISOLATED	-2
>> +#define PRS_INVALID_ISOLCPUS	-3 /* Effective xcpus still in isolated_cpus */
>>   
>>   /*
>>    * Temporary cpumasks for working with partitions that are passed among
>> @@ -382,6 +385,30 @@ static inline bool is_in_v2_mode(void)
>>   	      (cpuset_cgrp_subsys.root->flags & CGRP_ROOT_CPUSET_V2_MODE);
>>   }
>>   
>> +/*
>> + * If the given cpuset has a partition state of PRS_INVALID_ISOLCPUS,
>> + * remove its effective_xcpus from isolated_cpus and reset its state to
>> + * PRS_INVALID_ISOLATED. Also clear effective_xcpus if exclusive_cpus is
>> + * empty.
>> + */
>> +static void fix_invalid_isolcpus(struct cpuset *cs, struct cpuset *trialcs)
>> +{
>> +	if (likely(cs->partition_root_state != PRS_INVALID_ISOLCPUS))
>> +		return;
>> +	WARN_ON_ONCE(cpumask_empty(cs->effective_xcpus));
>> +	spin_lock_irq(&callback_lock);
>> +	cpumask_andnot(isolated_cpus, isolated_cpus, cs->effective_xcpus);
>> +	if (cpumask_empty(cs->exclusive_cpus))
>> +		cpumask_clear(cs->effective_xcpus);
>> +	cs->partition_root_state = PRS_INVALID_ISOLATED;
>> +	spin_unlock_irq(&callback_lock);
>> +	isolated_cpus_updating = true;
>> +	if (trialcs) {
>> +		trialcs->partition_root_state = PRS_INVALID_ISOLATED;
>> +		cpumask_copy(trialcs->effective_xcpus, cs->effective_xcpus);
>> +	}
>> +}
> When fix_invalid_isolcpus is called from changing cpus/exclusive cpus, should we
> copy cs->effective_xcpus to trialcs->effective_xcpus?
>
> I tested as follow steps(using the whole series):
>
>   # cd /sys/fs/cgroup/
>   # mkdir test
>   # echo 1 > cpuset.cpus.
>   # cd test/
>   # echo 1 > cpuset.cpus.exclusive
>   # echo $$ > cgroup.procs
>   # echo isolated > cpuset.cpus.partition
>   # cat cpuset.cpus.partition
> isolated
>   # echo 0 > /sys/devices/system/cpu/cpu1/online
>   # cat cpuset.cpus.partition
> isolated invalid
>   # echo 2 > cpuset.cpus.exclusive
>   # cat cpuset.cpus.partition
> isolated invalid (Parent unable to distribute cpu downstream)
>
> After changing cpuset.cpus.exclusive to 2, the test cpuset should
> become valid again, but it remains invalid.

Right, changes to trialcs->effective_xcpus is unnecessary() as 
compute_trialcs_excpus() will be called before fix_invalid_isolcpus() is 
invoked. Will fix that in the next version.

Thanks,
Longman