[Intel-gfx] [PATCH v2] drm/i915/execlists: Refactor common engine setup

[Tvrtko Ursulin]

On 29/04/16 11:00, Chris Wilson wrote:
> On Fri, Apr 29, 2016 at 10:50:02AM +0100, Tvrtko Ursulin wrote:
>>
>> On 29/04/16 10:39, Chris Wilson wrote:
>>> On Fri, Apr 29, 2016 at 10:25:41AM +0100, Tvrtko Ursulin wrote:
>>>> On 29/04/16 10:15, Chris Wilson wrote:
>>>>> diff --git a/drivers/gpu/drm/i915/intel_lrc.c b/drivers/gpu/drm/i915/intel_lrc.c
>>>>> index 2e0eaa9fa240..2c94072ab085 100644
>>>>> --- a/drivers/gpu/drm/i915/intel_lrc.c
>>>>> +++ b/drivers/gpu/drm/i915/intel_lrc.c
>>>>> @@ -2016,14 +2016,17 @@ logical_ring_setup(struct drm_device *dev, enum intel_engine_id id)
>>>>>          struct intel_engine_cs *engine = &dev_priv->engine[id];
>>>>>          enum forcewake_domains fw_domains;
>>>>>
>>>>> -       engine->dev = dev;
>>>>> -
>>>>>          engine->id = id;
>>>>>          engine->name = info->name;
>>>>>          engine->exec_id = info->exec_id;
>>>>>          engine->guc_id = info->guc_id;
>>>>>          engine->mmio_base = info->mmio_base;
>>>>>
>>>>> +       /* disable interrupts to this engine before we install ourselves*/
>>>>> +       I915_WRITE_IMR(engine, ~0);
>>>>> +
>>>>> +       engine->dev = dev;
>>>>> +
>>>>>          /* Intentionally left blank. */
>>>>>          engine->buffer = NULL;
>>>>>
>>>>> Make sense?
>>>>
>>>> Not the most elegant because all the hw access we have so far is in
>>>> engine->init_hw. Why can't we just make intel_engine_initialized
>>>> return false until the very last thing in engine constructors?
>>>
>>> In my defence sanitizing the hw before we are ready is common practice
>>> across the driver. The unfun part is that irq install is before gem_init
>>> (because modeset init wants irq enabled for GMBUS/dp-aux). gem init
>>> itself could be split up and moved around so that the setup and init_hw
>>> phases are separate (which would be next on the init ordering hitlist I
>>> hope).
>>>
>>> I want engine->dev/engine->i915 set early so we can use it during setup
>>> and init-hw and so that for_each_engine() works as expected in that
>>> time.
>>
>> Why wouldn't an explicit engine->initialized flag solve that? You
>> could keep setting engine->dev early (as it should be) and then set
>> engine->initialized at the end of per-engine constructors.
>
> Because it becomes irrelevant very shortly. The only interesting
> question remaining is whether or not we should be sanitizing the IMR
> first. It has been suggested elsewhere (in Ville's review of the GT
> interrupt handling) that doing the sanitization would be useful.

How come it becomes irrelevant? Will there not be 
intel_engine_initialized? Because as long as there is, imho it makes 
sense not to use engine->dev for it.

Regards,

Tvrtko