Re: [PATCH 1/2] perf/x86/rapl: Move the pmu allocation out of CPU hotplug

[Liang, Kan]

On 2024-10-08 4:36 p.m., Thomas Gleixner wrote:
> On Tue, Oct 08 2024 at 22:23, Thomas Gleixner wrote:
>> On Tue, Oct 08 2024 at 16:10, Kan Liang wrote:
>>> On 2024-10-08 12:33 p.m., Thomas Gleixner wrote:
>>>> On Fri, Sep 13 2024 at 10:10, kan liang wrote:
>>>>> +static void __init init_rapl_pmu(void)
>>>>> +{
>>>>> +	struct rapl_pmu *pmu;
>>>>> +	s32 rapl_pmu_idx;
>>>>> +	int cpu;
>>>>> +
>>>>> +	cpus_read_lock();
>>>>> +
>>>>> +	for_each_cpu(cpu, cpu_online_mask) {
>>>>
>>>> How is that supposed to work, when not all CPUs are online when
>>>> init_rapl_pmus() is invoked?
>>>>
>>>
>>> RAPL is a module. The module_init() is called during do_initcalls(),
>>> which is after the smp_init(). The cpu_online_mask has been setup in the
>>> smp_init().
>>>
>>> I also patched the kernel to double check. The cpu_online_mask indeed
>>> shows all the online CPUs.
>>>
>>> [    7.021212] smp: Brought up 1 node, 48 CPUs
>>> [    7.021212] smpboot: Total of 48 processors activated (211200.00
>>> BogoMIPS)
>>> ... ...
>>> [   16.557323] RAPL PMU: rapl_pmu_init: cpu_online_mask 0xffffffffffff
>>
>>  1) Start your kernel with maxcpus=2 (not recommended, but ...)
>>  2) Load the module
>>  3) Online the rest of the CPUs from userspace
>>
>> If your machine has more than one die you might be surprised...

Thanks. I will find a 2 sockets machine and give it a try.

> 
> You can make this work because the new topology code allows you to
> retrieve the possible number of cores/dies/packages even when they have
> not been onlined yet. 
>

Actually, I think the possible CPU mask should be good enough here. The
init_rapl_pmu() just intends to allocate the space for a pmu in each die.

The worst case of using a possible mask is that some space may be
wasted, when there is no online CPUs on a die. But it should be an
unusual case. It should be harmless.

Thanks,
Kan