[Intel-gfx] [PATCH 2/2] drm/i915/pmu: Fix CPU hotplug with multiple GPUs

[Tvrtko Ursulin]

On 20/10/2020 13:40, Chris Wilson wrote:
> Quoting Tvrtko Ursulin (2020-10-20 13:33:12)
>>
>> On 20/10/2020 13:10, Chris Wilson wrote:
>>> Quoting Chris Wilson (2020-10-20 12:59:57)
>>>> Quoting Tvrtko Ursulin (2020-10-20 11:08:22)
>>>>> From: Tvrtko Ursulin <tvrtko.ursulin at intel.com>
>>>>>
>>>>> Since we keep a driver global mask of online CPUs and base the decision
>>>>> whether PMU needs to be migrated upon it, we need to make sure the
>>>>> migration is done for all registered PMUs (so GPUs).
>>>>>
>>>>> To do this we need to track the current CPU for each PMU and base the
>>>>> decision on whether to migrate on a comparison between global and local
>>>>> state.
>>>>>
>>>>> At the same time, since dynamic CPU hotplug notification slots are a
>>>>> scarce resource and given how we already register the multi instance type
>>>>> state, we can and should add multiple instance of the i915 PMU to this
>>>>> same state and not allocate a new one for every GPU.
>>>>>
>>>>> v2:
>>>>>    * Use pr_notice. (Chris)
>>>>>
>>>>> Signed-off-by: Tvrtko Ursulin <tvrtko.ursulin at intel.com>
>>>>> Suggested-by: Daniel Vetter <daniel.vetter at intel.com> # dynamic slot optimisation
>>>>> Cc: Chris Wilson <chris at chris-wilson.co.uk>
>>>>> ---
>>>>>    drivers/gpu/drm/i915/i915_pci.c |  7 ++++-
>>>>>    drivers/gpu/drm/i915/i915_pmu.c | 50 ++++++++++++++++++++-------------
>>>>>    drivers/gpu/drm/i915/i915_pmu.h |  6 +++-
>>>>>    3 files changed, 41 insertions(+), 22 deletions(-)
>>>>>
>>>>> diff --git a/drivers/gpu/drm/i915/i915_pci.c b/drivers/gpu/drm/i915/i915_pci.c
>>>>> index 27964ac0638a..a384f51c91c1 100644
>>>>> --- a/drivers/gpu/drm/i915/i915_pci.c
>>>>> +++ b/drivers/gpu/drm/i915/i915_pci.c
>>>>> @@ -1150,9 +1150,13 @@ static int __init i915_init(void)
>>>>>                   return 0;
>>>>>           }
>>>>>    
>>>>> +       i915_pmu_init();
>>>>> +
>>>>>           err = pci_register_driver(&i915_pci_driver);
>>>>> -       if (err)
>>>>> +       if (err) {
>>>>> +               i915_pmu_exit();
>>>>>                   return err;
>>>>> +       }
>>>>>    
>>>>>           i915_perf_sysctl_register();
>>>>>           return 0;
>>>>> @@ -1166,6 +1170,7 @@ static void __exit i915_exit(void)
>>>>>           i915_perf_sysctl_unregister();
>>>>>           pci_unregister_driver(&i915_pci_driver);
>>>>>           i915_globals_exit();
>>>>> +       i915_pmu_exit();
>>>>>    }
>>>>>    
>>>>>    module_init(i915_init);
>>>>> diff --git a/drivers/gpu/drm/i915/i915_pmu.c b/drivers/gpu/drm/i915/i915_pmu.c
>>>>> index 51ed7d0efcdc..0d6c0945621e 100644
>>>>> --- a/drivers/gpu/drm/i915/i915_pmu.c
>>>>> +++ b/drivers/gpu/drm/i915/i915_pmu.c
>>>>> @@ -30,6 +30,7 @@
>>>>>    #define ENGINE_SAMPLE_BITS (1 << I915_PMU_SAMPLE_BITS)
>>>>>    
>>>>>    static cpumask_t i915_pmu_cpumask;
>>>>> +static unsigned int i915_pmu_target_cpu = -1;
>>>>>    
>>>>>    static u8 engine_config_sample(u64 config)
>>>>>    {
>>>>> @@ -1049,25 +1050,32 @@ static int i915_pmu_cpu_online(unsigned int cpu, struct hlist_node *node)
>>>>>    static int i915_pmu_cpu_offline(unsigned int cpu, struct hlist_node *node)
>>>>>    {
>>>>>           struct i915_pmu *pmu = hlist_entry_safe(node, typeof(*pmu), cpuhp.node);
>>>>> -       unsigned int target;
>>>>> +       unsigned int target = i915_pmu_target_cpu;
>>>>
>>>> So we still have multiple callbacks, one per pmu. But each callback is
>>>> now stored in a list from the cpuhp_slot instead of each callback having
>>>> its own slot.
>>>>
>>>>>    
>>>>>           GEM_BUG_ON(!pmu->base.event_init);
>>>>>    
>>>>>           if (cpumask_test_and_clear_cpu(cpu, &i915_pmu_cpumask)) {
>>>>
>>>> On first callback...
>>>>
>>>>>                   target = cpumask_any_but(topology_sibling_cpumask(cpu), cpu);
>>>>
>>>> Pick any other cpu.
>>>>
>>>>> +
>>>>>                   /* Migrate events if there is a valid target */
>>>>>                   if (target < nr_cpu_ids) {
>>>>>                           cpumask_set_cpu(target, &i915_pmu_cpumask);
>>>>> -                       perf_pmu_migrate_context(&pmu->base, cpu, target);
>>>>> +                       i915_pmu_target_cpu = target;
>>>>
>>>> Store target for all callbacks.
>>>>
>>>>>                   }
>>>>>           }
>>>>>    
>>>>> +       if (target < nr_cpu_ids && target != pmu->cpuhp.cpu) {
>>>>
>>>> If global [i915_pmu_target_cpu] target has changed, update perf.
>>>>
>>>>> +               perf_pmu_migrate_context(&pmu->base, cpu, target);
>>>>> +               pmu->cpuhp.cpu = target;
>>>>
>>>> It is claimed that cpuhp_state_remove_instance() will call the offline
>>>> callback for all online cpus... Do we need a pmu->base.state != STOPPED
>>>> guard?
>>>
>>> s/claimed/it definitely does :)/
>>>
>>> Or rather pmu->closed.
>>
>> Hm why? You think perf_pmu_migrate_context accesses something in the PMU
>> outside of the already protected entry points?
> 
> If this callback is being called for every online when we unplug one
> device, we then believe that no cpus remain online for all other devices.
> Should a cpu then be offlined, target is -1u so greater than
> nr_cpu_online and we move the perf context to the void, worst case, in
> the best case we fail to migrate the perf context off the dying cpu.

Well spotted nasty interaction, thanks.

Regards,

Tvrtko