Re: [RFC 0/2] DRM standardized memory stats

[Tvrtko Ursulin <tvrtko.ursulin@igalia.com>]

On 21/09/2026 10:19, Christian König wrote:
> On 9/17/26 14:04, Tvrtko Ursulin wrote:
>> Re-floating a simple idea (this time with more Cc) born from an user interest to
>> be able to correctly size the swap space during OS installation, based on the
>> size of the VRAM a discrete GPU might have. And to be able to do it in a vendor
>> agnostic way.
>>
>> Idea is to expose a standardized scheme in sysfs, under the DRM class card, and
>> under a new 'memstat' directory. Such as, example from amdgpu:
>>
>> /sys/class/drm/card1/memstat/
>> ├── gtt
>> │   ├── total_mb
>> │   └── used_mb
>> └── vram
>>      ├── total_mb
>>      └── used_mb
>>
>> Or with concrete numbers:
>>
>> $ grep -Hr . /sys/class/drm/card1/memstat/
>> /sys/class/drm/card1/memstat/vram/total_mb:1024
>> /sys/class/drm/card1/memstat/vram/used_mb:445
>> /sys/class/drm/card1/memstat/gtt/total_mb:7394
>> /sys/class/drm/card1/memstat/gtt/used_mb:71
>>
>> Drivers need to implement a simple DRM driver level callback which needs to
>> report a stable list of interesting memory regions and their respective stats.
>> The region names then become sub-directory names under the new 'memstat'
>> directory, with each region exposing the total size and the current usage.
>>
>> Similar data can already be queried if the dmem cgroup controller is enabled,
>> also only for the participating drivers, by querying the root cgroup. But
>> perhaps sysfs is easier, or perhaps it is too much code for too little benefit.
>> I am curious to hear any opinions.
> 
> The implementation can be polished a bit but I seriously like the idea to standardize that.

Indeed. I think I've fixed all the kobject related flaring bugs locally.

> amdgpu already exposes the information as non-standard sysfs files and adds a bit more, e.g. CPU visible VRAM size, VRAM vendor etc...
> 
> I'm wondering if those shouldn't be added as well and the existing sysfs files then implemented as symlinks.

CPU visible I guess could be added with a small addition to the 2nd 
(amdgpu) patch. So far the implementation is driven trivially from TTM 
placements / range managers, but I could easily add a different stat 
since drivers define what they export.

Existing files as symlinks I do not know who I would implement it. File 
to file is not possible AFAIU, right?

Allowing drivers to put custom files in there might be complicated and 
could create a standardization problem. If vendor is strongly desired I 
could add it as DRM owned attribute perhaps. Like 
/sys/class/drm/card1/memstat/vram/hw_vendor or something.

Regards,

Tvrtko

>> Cc: Maíra Canal <mcanal@igalia.com>
>> Cc: Ludovico de Nittis <ludovico.denittis@collabora.com>
>> Cc: Alex Deucher <alexander.deucher@amd.com>
>> Cc: Christian König <christian.koenig@amd.com>
>> Cc: Maarten Lankhorst <maarten.lankhorst@linux.intel.com>
>> Cc: Maxime Ripard <mripard@kernel.org>
>> Cc: Thomas Zimmermann <tzimmermann@suse.de>
>> Cc: David Airlie <airlied@gmail.com>
>> Cc: Simona Vetter <simona@ffwll.ch>
>>
>> Tvrtko Ursulin (2):
>>    drm: Allow drivers to report standardized memory stats
>>    drm/amdgpu: Wire up DRM memory stats reporting
>>
>>   drivers/gpu/drm/amd/amdgpu/amdgpu.h        |   6 +
>>   drivers/gpu/drm/amd/amdgpu/amdgpu_drv.c    |   2 +-
>>   drivers/gpu/drm/amd/amdgpu/amdgpu_fdinfo.c |  40 +++++--
>>   drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c    |  37 +++++++
>>   drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.h    |   3 +
>>   drivers/gpu/drm/drm_drv.c                  |  10 ++
>>   drivers/gpu/drm/drm_sysfs.c                | 123 +++++++++++++++++++++
>>   include/drm/drm_device.h                   |  19 ++++
>>   include/drm/drm_drv.h                      |   8 ++
>>   include/drm/drm_file.h                     |   9 ++
>>   include/drm/drm_sysfs.h                    |   4 +
>>   11 files changed, 248 insertions(+), 13 deletions(-)
>>
>