Re: [PATCH v4 5/6] drm: Add fdinfo memory stats

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

On 21/04/2023 12:45, Emil Velikov wrote:
> On Fri, 21 Apr 2023 at 12:23, Tvrtko Ursulin
> <tvrtko.ursulin@linux.intel.com> wrote:
> 
>> On 21/04/2023 11:26, Emil Velikov wrote:
>>> On Wed, 12 Apr 2023 at 23:43, Rob Clark <robdclark@gmail.com> wrote:
>>>
>>>> +/**
>>>> + * enum drm_gem_object_status - bitmask of object state for fdinfo reporting
>>>> + * @DRM_GEM_OBJECT_RESIDENT: object is resident in memory (ie. not unpinned)
>>>> + * @DRM_GEM_OBJECT_PURGEABLE: object marked as purgeable by userspace
>>>> + *
>>>> + * Bitmask of status used for fdinfo memory stats, see &drm_gem_object_funcs.status
>>>> + * and drm_show_fdinfo().  Note that an object can DRM_GEM_OBJECT_PURGEABLE if
>>>> + * it still active or not resident, in which case drm_show_fdinfo() will not
>>>
>>> nit: s/can/can be/;s/if it still/if it is still/
>>>
>>>> + * account for it as purgeable.  So drivers do not need to check if the buffer
>>>> + * is idle and resident to return this bit.  (Ie. userspace can mark a buffer
>>>> + * as purgeable even while it is still busy on the GPU.. it does not _actually_
>>>> + * become puregeable until it becomes idle.  The status gem object func does
>>>
>>> nit: s/puregeable/purgeable/
>>>
>>>
>>> I think we want a similar note in the drm-usage-stats.rst file.
>>>
>>> With the above the whole series is:
>>> Reviewed-by: Emil Velikov <emil.l.velikov@gmail.com>
>>
>> Have you maybe noticed my slightly alternative proposal? (*) I am not a
>> fan of putting this flavour of accounting into the core with no way to
>> opt out. I think it leaves no option but to add more keys in the future
>> for any driver which will not be happy with the core accounting.
>>
>> *) https://patchwork.freedesktop.org/series/116581/
>>
> 
> Indeed I saw it. Not a fan of it, I'm afraid.

Hard to guess the reasons. :)

Anyway, at a minimum I suggest that if the no opt out version has to go 
in, it is clearly documented drm-*-memory-* is *not* about the full 
memory use of the client, but about memory belonging to user visible 
buffer objects *only*. Possibly going as far as naming the keys as 
drm-user-bo-memory-... That way there is a way to implement proper 
drm-*-memory- in the future.

Regards,

Tvrtko

>>> Fwiw: Keeping the i915 patch as part of this series would be great.
>>> Sure i915_drm_client->id becomes dead code, but it's a piece one can
>>> live with for a release or two. Then again it's not my call to make.
>>
>> Rob can take the i915 patch from my RFC too.
>>
> 
> Indeed.
> 
> -Emil