Re: [PATCH v2 1/2] cpuhp: Optimize SMT switch operation by batching lock acquisition

[Shrikanth Hegde <sshegde@linux.ibm.com>]

On 2/16/26 6:27 PM, Shrikanth Hegde wrote:
> 

>>            */
>>           if (ctrlval == CPU_SMT_ENABLED && cpu_smt_thread_allowed(cpu))
>>               continue;
>> -        ret = cpu_down_maps_locked(cpu, CPUHP_OFFLINE);
> 
> If you see comment in cpu_down_maps_locked, It says.
> 
>           * Ensure that the control task does not run on the to be offlined
>           * CPU to prevent a deadlock against cfs_b->period_timer.
>           * Also keep at least one housekeeping cpu onlined to avoid 
> generating
>           * an empty sched_domain span
> 
> Now, instead of running on the online_CPU you are now doing the _cpu_down
> functionality in the cpu which is going offline.
> 
> How are you preventing that deadlock?

Ok. I should have looked at updated code. My bad.

This is no longer true after below commit.
de715325cc47 cpu: Revert "cpu/hotplug: Prevent self deadlock on CPU hot-unplug"

> 
>> +        ret = cpu_down_locked(cpu, 0, CPUHP_OFFLINE);
>>           if (ret)
>>               break;
>>           /*
>> @@ -2688,6 +2714,9 @@ int cpuhp_smt_disable(enum cpuhp_smt_control 
>> ctrlval)
>>       }
>>       if (!ret)
>>           cpu_smt_control = ctrlval;
>> +    cpus_write_unlock();
>> +    arch_smt_update();
>> +out:
>>       cpu_maps_update_done();
>>       return ret;
>>   }
>> @@ -2705,6 +2734,8 @@ int cpuhp_smt_enable(void)
>>       int cpu, ret = 0;
>>       cpu_maps_update_begin();
>> +    /* Hold cpus_write_lock() for entire batch operation. */
>> +    cpus_write_lock();
>>       cpu_smt_control = CPU_SMT_ENABLED;
>>       for_each_present_cpu(cpu) {
>>           /* Skip online CPUs and CPUs on offline nodes */
>> @@ -2712,12 +2743,14 @@ int cpuhp_smt_enable(void)
>>               continue;
>>           if (!cpu_smt_thread_allowed(cpu) || ! 
>> topology_is_core_online(cpu))
>>               continue;
>> -        ret = _cpu_up(cpu, 0, CPUHP_ONLINE);
>> +        ret = cpu_up_locked(cpu, 0, CPUHP_ONLINE);
>>           if (ret)
>>               break;
>>           /* See comment in cpuhp_smt_disable() */
>>           cpuhp_online_cpu_device(cpu);
>>       }
>> +    cpus_write_unlock();
>> +    arch_smt_update();
>>       cpu_maps_update_done();
>>       return ret;
>>   }
>