Re: [OFFLIST PATCH 2/2] cgroup: use subsystem-specific rstat locks to avoid contention

[JP Kobryn <inwardvessel@gmail.com>]

On 5/21/25 4:50 PM, Shakeel Butt wrote:
> On Wed, May 21, 2025 at 04:47:01PM -0700, JP Kobryn wrote:
>>
>>
>> On 5/21/25 4:33 PM, Shakeel Butt wrote:
>>> On Wed, May 21, 2025 at 04:23:44PM -0700, Shakeel Butt wrote:
>>>> On Thu, May 22, 2025 at 12:23:44AM +0200, Klara Modin wrote:
>>>>> Hi,
>>>>>
>>>>> On 2025-04-28 23:15:58 -0700, Shakeel Butt wrote:
>>>>>> Please ignore this patch as it was sent by mistake.
>>>>>
>>>>> This seems to have made it into next:
>>>>>
>>>>> 748922dcfabd ("cgroup: use subsystem-specific rstat locks to avoid contention")
>>>>>
>>>>> It causes a BUG and eventually a panic on my Raspberry Pi 1:
>>>>>
>>>>> WARNING: CPU: 0 PID: 0 at mm/percpu.c:1766 pcpu_alloc_noprof (mm/percpu.c:1766 (discriminator 2))
>>>>> illegal size (0) or align (4) for percpu allocation
>>>>
>>>> Ok this config is without CONFIG_SMP and on such configs we have:
>>>>
>>>> typedef struct { } arch_spinlock_t;
>>>>
>>>> So, we are doing ss->rstat_ss_cpu_lock = alloc_percpu(0).
>>>>
>>>> Hmm, let me think more on how to fix this.
>>>>
>>>
>>> I think following is the simplest fix:
>>>
>>> diff --git a/kernel/cgroup/rstat.c b/kernel/cgroup/rstat.c
>>> index 7dd396ae3c68..aab09495192e 100644
>>> --- a/kernel/cgroup/rstat.c
>>> +++ b/kernel/cgroup/rstat.c
>>> @@ -511,7 +511,10 @@ int __init ss_rstat_init(struct cgroup_subsys *ss)
>>>    	int cpu;
>>>    	if (ss) {
>>> -		ss->rstat_ss_cpu_lock = alloc_percpu(raw_spinlock_t);
>>> +		size_t size = sizeof(raw_spinlock_t) ?: 1;
>>> +
>>> +		ss->rstat_ss_cpu_lock = __alloc_percpu(size,
>>> +						__alignof__(raw_spinlock_t));
>>
>> Thanks for narrowing this one down so fast. Would this approach be more
>> straightforward?
>>
>> if (ss) {
>> #ifdef CONFIG_SMP
>> 	ss->rstat_ss_cpu_lock = alloc_percpu(raw_spinlock_t);
>> #endif
>>
>> Since on non-smp the lock functions are no-ops, leaving the ss cpu lock
>> can perhaps be left NULL. I could include a comment as well explaining
>> why.
>>
>>>    		if (!ss->rstat_ss_cpu_lock)
>>>    			return -ENOMEM;
> 
> Include this check and return -ENOMEM in the ifdef as well.

Good call. I will get a patch out tonight.

> 
>>>    	}
>>