Re: [Intel-gfx] [PATCH 2/3] drm/i915/gt: create workqueue dedicated to wake references

[Tvrtko Ursulin <tvrtko.ursulin@linux.intel.com>]

On 17/05/2023 12:18, Coelho, Luciano wrote:
> On Fri, 2023-05-12 at 13:16 +0100, Tvrtko Ursulin wrote:
>> On 12/05/2023 10:54, Coelho, Luciano wrote:
>>> On Fri, 2023-05-12 at 10:32 +0100, Tvrtko Ursulin wrote:
>>>> On 12/05/2023 10:10, Coelho, Luciano wrote:
>>>>> On Fri, 2023-05-12 at 10:04 +0100, Tvrtko Ursulin wrote:
>>>>>> On 11/05/2023 09:20, Luca Coelho wrote:
>>>>>>> Add a work queue in the intel_wakeref structure to be used exclusively
>>>>>>> by the wake reference mechanism.  This is needed in order to avoid
>>>>>>> using the system workqueue and relying on flush_scheduled_work().
>>>>>>>
>>>>>>> Cc: Tetsuo Handa <penguin-kernel@I-love.SAKURA.ne.jp>
>>>>>>> Cc: Tvrtko Ursulin <tvrtko.ursulin@intel.com>
>>>>>>> Cc: Jani Nikula <jani.nikula@intel.com>
>>>>>>> Cc: Ville Syrjälä <ville.syrjala@linux.intel.com>
>>>>>>> Signed-off-by: Luca Coelho <luciano.coelho@intel.com>
>>>>>>> ---
>>>>>>>      drivers/gpu/drm/i915/gt/intel_engine_cs.c |  7 ++++++-
>>>>>>>      drivers/gpu/drm/i915/gt/intel_engine_pm.c | 15 ++++++++++++--
>>>>>>>      drivers/gpu/drm/i915/gt/intel_engine_pm.h |  3 ++-
>>>>>>>      drivers/gpu/drm/i915/gt/mock_engine.c     |  8 +++++++-
>>>>>>>      drivers/gpu/drm/i915/intel_wakeref.c      | 21 ++++++++++++++-----
>>>>>>>      drivers/gpu/drm/i915/intel_wakeref.h      | 25 +++++++++++++++--------
>>>>>>>      6 files changed, 60 insertions(+), 19 deletions(-)
>>>>>>>
>>>>>>> diff --git a/drivers/gpu/drm/i915/gt/intel_engine_cs.c b/drivers/gpu/drm/i915/gt/intel_engine_cs.c
>>>>>>> index 0aff5bb13c53..6505bfa70cd0 100644
>>>>>>> --- a/drivers/gpu/drm/i915/gt/intel_engine_cs.c
>>>>>>> +++ b/drivers/gpu/drm/i915/gt/intel_engine_cs.c
>>>>>>> @@ -1290,7 +1290,11 @@ static int engine_setup_common(struct intel_engine_cs *engine)
>>>>>>>      		goto err_cmd_parser;
>>>>>>>      
>>>>>>>      	intel_engine_init_execlists(engine);
>>>>>>> -	intel_engine_init__pm(engine);
>>>>>>> +
>>>>>>> +	err = intel_engine_init__pm(engine);
>>>>>>> +	if (err)
>>>>>>> +		goto err_cmd_parser;
>>>>>>> +
>>>>>>>      	intel_engine_init_retire(engine);
>>>>>>>      
>>>>>>>      	/* Use the whole device by default */
>>>>>>> @@ -1525,6 +1529,7 @@ void intel_engine_cleanup_common(struct intel_engine_cs *engine)
>>>>>>>      {
>>>>>>>      	GEM_BUG_ON(!list_empty(&engine->sched_engine->requests));
>>>>>>>      
>>>>>>> +	intel_engine_destroy__pm(engine);
>>>>>>>      	i915_sched_engine_put(engine->sched_engine);
>>>>>>>      	intel_breadcrumbs_put(engine->breadcrumbs);
>>>>>>>      
>>>>>>> diff --git a/drivers/gpu/drm/i915/gt/intel_engine_pm.c b/drivers/gpu/drm/i915/gt/intel_engine_pm.c
>>>>>>> index ee531a5c142c..859b62cf660f 100644
>>>>>>> --- a/drivers/gpu/drm/i915/gt/intel_engine_pm.c
>>>>>>> +++ b/drivers/gpu/drm/i915/gt/intel_engine_pm.c
>>>>>>> @@ -294,14 +294,25 @@ static const struct intel_wakeref_ops wf_ops = {
>>>>>>>      	.put = __engine_park,
>>>>>>>      };
>>>>>>>      
>>>>>>> -void intel_engine_init__pm(struct intel_engine_cs *engine)
>>>>>>> +int intel_engine_init__pm(struct intel_engine_cs *engine)
>>>>>>>      {
>>>>>>>      	struct intel_runtime_pm *rpm = engine->uncore->rpm;
>>>>>>> +	int err;
>>>>>>> +
>>>>>>> +	err = intel_wakeref_init(&engine->wakeref, rpm, &wf_ops);
>>>>>>> +	if (err)
>>>>>>> +		return err;
>>>>>>>      
>>>>>>> -	intel_wakeref_init(&engine->wakeref, rpm, &wf_ops);
>>>>>>>      	intel_engine_init_heartbeat(engine);
>>>>>>>      
>>>>>>>      	intel_gsc_idle_msg_enable(engine);
>>>>>>> +
>>>>>>> +	return 0;
>>>>>>> +}
>>>>>>> +
>>>>>>> +void intel_engine_destroy__pm(struct intel_engine_cs *engine)
>>>>>>> +{
>>>>>>> +	intel_wakeref_destroy(&engine->wakeref);
>>>>>>>      }
>>>>>>>      
>>>>>>>      /**
>>>>>>> diff --git a/drivers/gpu/drm/i915/gt/intel_engine_pm.h b/drivers/gpu/drm/i915/gt/intel_engine_pm.h
>>>>>>> index d68675925b79..e8568f7d10c6 100644
>>>>>>> --- a/drivers/gpu/drm/i915/gt/intel_engine_pm.h
>>>>>>> +++ b/drivers/gpu/drm/i915/gt/intel_engine_pm.h
>>>>>>> @@ -104,7 +104,8 @@ intel_engine_create_kernel_request(struct intel_engine_cs *engine)
>>>>>>>      	return rq;
>>>>>>>      }
>>>>>>>      
>>>>>>> -void intel_engine_init__pm(struct intel_engine_cs *engine);
>>>>>>> +int intel_engine_init__pm(struct intel_engine_cs *engine);
>>>>>>> +void intel_engine_destroy__pm(struct intel_engine_cs *engine);
>>>>>>>      
>>>>>>>      void intel_engine_reset_pinned_contexts(struct intel_engine_cs *engine);
>>>>>>>      
>>>>>>> diff --git a/drivers/gpu/drm/i915/gt/mock_engine.c b/drivers/gpu/drm/i915/gt/mock_engine.c
>>>>>>> index c0637bf799a3..0a3c702c21e2 100644
>>>>>>> --- a/drivers/gpu/drm/i915/gt/mock_engine.c
>>>>>>> +++ b/drivers/gpu/drm/i915/gt/mock_engine.c
>>>>>>> @@ -336,6 +336,7 @@ static void mock_engine_release(struct intel_engine_cs *engine)
>>>>>>>      	intel_context_put(engine->kernel_context);
>>>>>>>      
>>>>>>>      	intel_engine_fini_retire(engine);
>>>>>>> +	intel_engine_destroy__pm(engine);
>>>>>>>      }
>>>>>>>      
>>>>>>>      struct intel_engine_cs *mock_engine(struct drm_i915_private *i915,
>>>>>>> @@ -393,6 +394,7 @@ struct intel_engine_cs *mock_engine(struct drm_i915_private *i915,
>>>>>>>      int mock_engine_init(struct intel_engine_cs *engine)
>>>>>>>      {
>>>>>>>      	struct intel_context *ce;
>>>>>>> +	int err;
>>>>>>>      
>>>>>>>      	INIT_LIST_HEAD(&engine->pinned_contexts_list);
>>>>>>>      
>>>>>>> @@ -402,7 +404,11 @@ int mock_engine_init(struct intel_engine_cs *engine)
>>>>>>>      	engine->sched_engine->private_data = engine;
>>>>>>>      
>>>>>>>      	intel_engine_init_execlists(engine);
>>>>>>> -	intel_engine_init__pm(engine);
>>>>>>> +
>>>>>>> +	err = intel_engine_init__pm(engine);
>>>>>>> +	if (err)
>>>>>>> +		return err;
>>>>>>> +
>>>>>>>      	intel_engine_init_retire(engine);
>>>>>>>      
>>>>>>>      	engine->breadcrumbs = intel_breadcrumbs_create(NULL);
>>>>>>> diff --git a/drivers/gpu/drm/i915/intel_wakeref.c b/drivers/gpu/drm/i915/intel_wakeref.c
>>>>>>> index dfd87d082218..6bae609e1312 100644
>>>>>>> --- a/drivers/gpu/drm/i915/intel_wakeref.c
>>>>>>> +++ b/drivers/gpu/drm/i915/intel_wakeref.c
>>>>>>> @@ -74,7 +74,7 @@ void __intel_wakeref_put_last(struct intel_wakeref *wf, unsigned long flags)
>>>>>>>      
>>>>>>>      	/* Assume we are not in process context and so cannot sleep. */
>>>>>>>      	if (flags & INTEL_WAKEREF_PUT_ASYNC || !mutex_trylock(&wf->mutex)) {
>>>>>>> -		mod_delayed_work(system_wq, &wf->work,
>>>>>>> +		mod_delayed_work(wf->wq, &wf->work,
>>>>>>>      				 FIELD_GET(INTEL_WAKEREF_PUT_DELAY, flags));
>>>>>>>      		return;
>>>>>>>      	}
>>>>>>> @@ -93,10 +93,10 @@ static void __intel_wakeref_put_work(struct work_struct *wrk)
>>>>>>>      	____intel_wakeref_put_last(wf);
>>>>>>>      }
>>>>>>>      
>>>>>>> -void __intel_wakeref_init(struct intel_wakeref *wf,
>>>>>>> -			  struct intel_runtime_pm *rpm,
>>>>>>> -			  const struct intel_wakeref_ops *ops,
>>>>>>> -			  struct intel_wakeref_lockclass *key)
>>>>>>> +int __intel_wakeref_init(struct intel_wakeref *wf,
>>>>>>> +			 struct intel_runtime_pm *rpm,
>>>>>>> +			 const struct intel_wakeref_ops *ops,
>>>>>>> +			 struct intel_wakeref_lockclass *key)
>>>>>>>      {
>>>>>>>      	wf->rpm = rpm;
>>>>>>>      	wf->ops = ops;
>>>>>>> @@ -105,9 +105,20 @@ void __intel_wakeref_init(struct intel_wakeref *wf,
>>>>>>>      	atomic_set(&wf->count, 0);
>>>>>>>      	wf->wakeref = 0;
>>>>>>>      
>>>>>>> +	wf->wq = alloc_workqueue("i1915-wakeref", 0, 0);
>>>>>>
>>>>>> Typo here -
>>>>>
>>>>> Oh, good catch! This is one of my "favorite" typos, for some reason.
>>>>
>>>> Yes, I had the same one. :) Patch 3/3 too.
>>>>
>>>>>>     I wanted to ask however - why does this particular wq
>>>>>> "deserves" to be dedicated and can't just use one of the
>>>>>> drm_i915_private ones?
>>>>>
>>>>> It's because there's no easy way to get access to the drm_i915_private
>>>>> structure from here.  And I don't think this work needs to be in sync
>>>>> with the rest of the works in i915.
>>>>
>>>> Yeah I don't think it needs to be synchronised either. Was just thinking
>>>> if we really need to be creating a bunch of separate workqueues (one per
>>>> engine) for not much use, or instead could just add a backpointer to
>>>> either intel_wakeref or intel_runtime_pm. Latter already has rpm->kdev
>>>> so could plausably be replaced with rpm->i915.
>>>>
>>>> Actually, looking at intel_runtime_pm_init_early, you could get to i915
>>>> via wf->rpm and container_of.
>>>
>>> Yeah, I considered that, but using container_of() can be problematic
>>> when we're not sure where exactly the element is coming from.  My worry
>>> was this:
>>>
>>> int intel_engine_init__pm(struct intel_engine_cs *engine)
>>> {
>>> 	struct intel_runtime_pm *rpm = engine->uncore->rpm;
>>> 	int err;
>>>
>>> 	err = intel_wakeref_init(&engine->wakeref, rpm, &wf_ops);
>>> [...]
>>> }
>>>
>>> In this case, we're getting to __intel_wakeref_init() with an *rpm that
>>> is coming from an intel_uncore structure and not from
>>> drm_i915_private...
>>
>> Right. Yes I agree that would be a flaky/questionable design, even if it
>> worked in practice. I'd just replace rpm->dev with rpm->i915 then. Not
>> feeling *that* strongly about it, but it just feels a waste to create a
>> bunch of workqueues for this.
> 
> How many engines do we typically have at runtime?

Currently not that many I think. There were plans for like 18 per tile, 
but I don't think it will be happening.

> I was looking into getting the i915 struct via rpm->kdev and
> container_of(), but it's not trivial either.  This is how we initialize
> rpm->kdev:
> 
> void intel_runtime_pm_init_early(struct intel_runtime_pm *rpm)
> {
> 	struct drm_i915_private *i915 =
> 			container_of(rpm, struct drm_i915_private, runtime_pm);
> 	struct pci_dev *pdev = to_pci_dev(i915->drm.dev);
> 	struct device *kdev = &pdev->dev;
> 
> 	rpm->kdev = kdev;
> [...]
> }
> 
> So, rpm->kdev actually points to the device structure from inside a
> pci_dev.  i915->drm.dev points to the same device structure, but we
> lose the pointer to the element inside i915, so there's no way to trace
> it back.  We keep only the pointer to the element inside pci_dev.
> 
> Or did you mean that we should change the kdev pointer to an i915
> pointer in intel_runtime_pm and derive kdev when needed? This would
> cause some code shuffle.  And, currently, the wakeref code doesn't have
> any references to drm_i915_private at all (or even to drm for that
> matter)...

Yes, that's what I meant. Struct intel_runtime_pm is only intimately 
tied to i915 via lmem_userfault_list so I think conceptually it is fine. 
But I haven't look at the level of churn it would require.

> I don't know.  We could go either way.  I think it depends mostly on
> how many engines instances we typically have, so we can weigh runtime
> resources vs code complexity...

Or we just go with Tetsuo's approach of a global i915_wq. For me that 
one is completely fine.

Regards,

Tvrtko