Re: [-next v2 4/4] cpuset: add helpers for cpus read and cpuset_mutex locks

[Waiman Long <llong@redhat.com>]

On 8/13/25 11:58 PM, Chen Ridong wrote:
>
> On 2025/8/14 11:27, Waiman Long wrote:
>> On 8/13/25 11:13 PM, Waiman Long wrote:
>>> On 8/13/25 8:44 PM, Chen Ridong wrote:
>>>> On 2025/8/14 4:09, Waiman Long wrote:
>>>>> On 8/13/25 4:29 AM, Chen Ridong wrote:
>>>>>> From: Chen Ridong <chenridong@huawei.com>
>>>>>>
>>>>>> cpuset: add helpers for cpus_read_lock and cpuset_mutex
>>>>>>
>>>>>> Replace repetitive locking patterns with new helpers:
>>>>>> - cpus_read_cpuset_lock()
>>>>>> - cpus_read_cpuset_unlock()
>>>>>>
>>>>>> This makes the code cleaner and ensures consistent lock ordering.
>>>>>>
>>>>>> Signed-off-by: Chen Ridong <chenridong@huawei.com>
>>>>>> ---
>>>>>>     kernel/cgroup/cpuset-internal.h |  2 ++
>>>>>>     kernel/cgroup/cpuset-v1.c       | 12 +++------
>>>>>>     kernel/cgroup/cpuset.c          | 48 +++++++++++++++------------------
>>>>>>     3 files changed, 28 insertions(+), 34 deletions(-)
>>>>>>
>>>>>> diff --git a/kernel/cgroup/cpuset-internal.h b/kernel/cgroup/cpuset-internal.h
>>>>>> index 75b3aef39231..6fb00c96044d 100644
>>>>>> --- a/kernel/cgroup/cpuset-internal.h
>>>>>> +++ b/kernel/cgroup/cpuset-internal.h
>>>>>> @@ -276,6 +276,8 @@ int cpuset_update_flag(cpuset_flagbits_t bit, struct cpuset *cs, int
>>>>>> turning_on)
>>>>>>     ssize_t cpuset_write_resmask(struct kernfs_open_file *of,
>>>>>>                         char *buf, size_t nbytes, loff_t off);
>>>>>>     int cpuset_common_seq_show(struct seq_file *sf, void *v);
>>>>>> +void cpus_read_cpuset_lock(void);
>>>>>> +void cpus_read_cpuset_unlock(void);
>>>>> The names are not intuitive. I would prefer just extend the cpuset_lock/unlock to include
>>>>> cpus_read_lock/unlock and we use cpuset_lock/unlock consistently in the cpuset code. Also, there is
>>>>> now no external user of cpuset_lock/unlock, we may as well remove them from include/linux/cpuset.h.
>>>>>
>>>>> Cheers,
>>>>> Longman
>>>> I like the idea and have considered it.
>>>> However, I noticed that cpuset_locked is being used in __sched_setscheduler.
>>> Right, I overloooked the cpuset_lock() call in kernel/sched/syscall.c. So we can't remove it from
>>> include/linux/cpuset.h.
>>>
>>> This call is invoked to ensure cpusets information is stable. However, it doesn't hurt if the
>>> cpus_read_lock() is also acquired as a result. Alternatively, we can use a name like
>>> cpuset_full_lock() to include cpus_read_lock().
>> I have a correction. According to commit d74b27d63a8b ("cgroup/cpuset: Change cpuset_rwsem and
>> hotplug lock order") , sched_scheduler() can be called while holding cpus_hotplug_lock. So we should
>> keep cpuset_lock() as it is.
>>
>> Cheers,
>> Longman
> Thank you Longman, this is very helpful.
>
> I had considered whether we can add cpus_read_lock() to the cpuset_lock, but based on your
> explanation, I now understand this approach would not work.
>
> For clarity, would it be acceptable to rename:
> cpus_read_cpuset_lock() -> cpuset_full_lock()
> cpus_read_cpuset_unlock() -> cpuset_full_unlock()

Yes, that is what I want to see. Note that taking both cpus_read_lock() 
and cpuset_mutex are needed to modify cpuset data. Taking just 
cpuset_mutex will prevent other from making changes to the cpuset data, 
but is not enough to make modification.

Cheers,
Longman

>