Re: [PATCH v5 3/3] cgroup/cpuset: Keep user set cpus affinity

[Waiman Long <longman@redhat.com>]

On 8/16/22 16:15, Tejun Heo wrote:
> On Tue, Aug 16, 2022 at 03:27:34PM -0400, Waiman Long wrote:
>> +static int cpuset_set_cpus_allowed_ptr(struct task_struct *p,
>> +				       const struct cpumask *mask)
>> +{
>> +	cpumask_var_t new_mask;
>> +	int ret;
>> +
>> +	if (!READ_ONCE(p->user_cpus_ptr)) {
>> +		ret = set_cpus_allowed_ptr(p, mask);
>> +		/*
>> +		 * If user_cpus_ptr becomes set now, we are racing with
>> +		 * a concurrent sched_setaffinity(). So use the newly
>> +		 * set user_cpus_ptr and retry again.
>> +		 *
>> +		 * TODO: We cannot detect change in the cpumask pointed to
>> +		 * by user_cpus_ptr. We will have to add a sequence number
>> +		 * if such a race needs to be addressed.
>> +		 */
> This is too ugly and obviously broken. Let's please do it properly.

Actually, there is similar construct in __sched_setaffinity():

again:
         retval = __set_cpus_allowed_ptr(p, new_mask, SCA_CHECK);
         if (retval)
                 goto out_free_new_mask;

         cpuset_cpus_allowed(p, cpus_allowed);
         if (!cpumask_subset(new_mask, cpus_allowed)) {
                 /*
                  * We must have raced with a concurrent cpuset update.
                  * Just reset the cpumask to the cpuset's cpus_allowed.
                  */
                 cpumask_copy(new_mask, cpus_allowed);
                 goto again;
         }

It is hard to synchronize different subsystems atomically without 
running into locking issue. Let me think about what can be done in this 
case.

Is using a sequence number to check for race with retry good enough?

Cheers,
Longman