[Intel-gfx] [PATCH] drm/i915: Infrastructure for supporting different GGTT views per object

[Tvrtko Ursulin]

On 12/01/2014 04:07 PM, Daniel Vetter wrote:
> On Mon, Dec 01, 2014 at 11:32:42AM +0000, Tvrtko Ursulin wrote:
>> On 11/28/2014 05:31 PM, Daniel Vetter wrote:
>>> On Thu, Nov 27, 2014 at 02:52:44PM +0000, Tvrtko Ursulin wrote:
>>>> From: Tvrtko Ursulin <tvrtko.ursulin at intel.com>
>>>> diff --git a/drivers/gpu/drm/i915/i915_gem.c b/drivers/gpu/drm/i915/i915_gem.c
>>>> index 86cf428..6213c07 100644
>>>> --- a/drivers/gpu/drm/i915/i915_gem.c
>>>> +++ b/drivers/gpu/drm/i915/i915_gem.c
>>>> @@ -2090,8 +2090,7 @@ i915_gem_shrink(struct drm_i915_private *dev_priv,
>>>>   			/* For the unbound phase, this should be a no-op! */
>>>>   			list_for_each_entry_safe(vma, v,
>>>>   						 &obj->vma_list, vma_link)
>>>> -				if (i915_vma_unbind(vma))
>>>> -					break;
>>>> +				i915_vma_unbind(vma);
>>>
>>> Why drop the early break if a vma_unbind fails? Looks like a superflous
>>> hunk to me.
>>
>> I wasn't sure about this. (Does it makes sense to try and unbind other VMAs
>> if one couldn't be unbound?)
>>
>> In fact, looking at it now, I am not sure about the unbind flow
>> (i915_vma_unbind). Won't i915_gem_object_retire move all VMAs to inactive
>> list on first VMA unbind? Retire only on last VMA going away?
>
> Yeah only the first vma_unbind might fail with the current code. The
> problem though is that you ignore all failures.

I am not sure what you mean. Why only the first unbind can fail?

The part I was unsure about was this break removal in the shrinker. 
Whether or not it makes sense to go through all VMAs regardless if one 
failed to unbind? Is there any space to be gained by doing that?

Alternatively, I also looked at it as: If it doesn't make sense to go 
through all of then, then what to do if the first unbind succeeds and 
some other fails?  End results sounds the same as trying to unbind as 
much as possible. So I opted for doing that.

My second concern is that object retire on 1st VMA unbind. Should that 
only be done when the last VMA is going away?

As it stands (in my v2 patch) it can move all VMAs onto the inactive 
list when first one is unbound which looks wrong.

> Aside: In general this is also about reducing the size of the diff to only
> have essential changes. I'm happy when people throw in more cleanup, but
> it should be done as follow-up/prep patches. This is because often the
> only evidence for fixing a bug we have is "it bisects to this commit". So
> making commits in complex code like gem minimal is the name of the game.

Sure, as soon as one understands what is essential and what is not. :)

>>>> diff --git a/drivers/gpu/drm/i915/i915_gpu_error.c b/drivers/gpu/drm/i915/i915_gpu_error.c
>>>> index 89a2f3d..77f1bdc 100644
>>>> --- a/drivers/gpu/drm/i915/i915_gpu_error.c
>>>> +++ b/drivers/gpu/drm/i915/i915_gpu_error.c
>>>> @@ -717,10 +717,8 @@ static u32 capture_pinned_bo(struct drm_i915_error_buffer *err,
>>>>   			break;
>>>>
>>>>   		list_for_each_entry(vma, &obj->vma_list, vma_link)
>>>> -			if (vma->vm == vm && vma->pin_count > 0) {
>>>> +			if (vma->vm == vm && vma->pin_count > 0)
>>>>   				capture_bo(err++, vma);
>>>> -				break;
>>>
>>> Not fully sure about this one, but can't hurt I guess.
>>
>> Not sure if it is useful at the moment or at all?
>
> Probably not useful right now. Otoh if we ever wire up the display fault
> registers on modern platforms this migh become useful to cross-check that
> the current display plane register settings match up with the
> corresponding buffer. Won't hurt either though.
>
> If you feel like make it a separate patch perhaps.

I don't know, it sounds like an overkill to do that for this short hunk 
so I prefer to leave it in if you don't mind.

Regards,

Tvrtko