[PATCH v18 25/26] drm/virtio: Support shmem shrinking

[Dmitry Osipenko]

On 11/4/23 01:55, Gurchetan Singh wrote:
> On Sun, Oct 29, 2023 at 4:03 PM Dmitry Osipenko <
> dmitry.osipenko at collabora.com> wrote:
> 
>> Support generic drm-shmem memory shrinker and add new madvise IOCTL to
>> the VirtIO-GPU driver. BO cache manager of Mesa driver will mark BOs as
>> "don't need" using the new IOCTL to let shrinker purge the marked BOs on
>> OOM, the shrinker will also evict unpurgeable shmem BOs from memory if
>> guest supports SWAP file or partition.
>>
>> Acked-by: Gerd Hoffmann <kraxel at redhat.com>
>> Signed-off-by: Daniel Almeida <daniel.almeida at collabora.com>
>> Signed-off-by: Dmitry Osipenko <dmitry.osipenko at collabora.com>
>> ---
>>  drivers/gpu/drm/virtio/virtgpu_drv.h    | 13 +++++-
>>  drivers/gpu/drm/virtio/virtgpu_gem.c    | 35 ++++++++++++++
>>  drivers/gpu/drm/virtio/virtgpu_ioctl.c  | 25 ++++++++++
>>  drivers/gpu/drm/virtio/virtgpu_kms.c    |  8 ++++
>>  drivers/gpu/drm/virtio/virtgpu_object.c | 61 +++++++++++++++++++++++++
>>  drivers/gpu/drm/virtio/virtgpu_vq.c     | 40 ++++++++++++++++
>>  include/uapi/drm/virtgpu_drm.h          | 14 ++++++
>>  7 files changed, 195 insertions(+), 1 deletion(-)
...
> 
> Link to open-source userspace?

https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/15278

I'll add it to the commit message

> Also, don't you need a VIRTGPU_PARAM_MADVISE_SUPPORTED or is the plan to
> use a DRM version?

The ioctl() returns error if DRM_VIRTGPU_MADVISE unsupported, extra
flags not needed by userspace

> Overall, the series for a generic shrinker seems useful for a wide variety
> of DRM drivers, such as Panfrost.
> 
> For virtio-gpu, it could be a tad VIRGIL specific: since other context
> types (gfxstream gles, DRM, vk contexts) decrease memory consumption by
> either not allocating shadow buffers for textures/buffers, or using blob
> memory.
> 
> Maybe we need to design with blob in mind, since we expect virgl to be
> deprecated.  On Android, it basically is with ANGLE and native contexts.
> On Linux, Zink looks good too.  Even with memory issues fixed, virgl is
> unattractive compared to those solutions.

We should finish shmem first since started with it, then move to blobs.

My rough idea for the blobs is to use a timer-based approach to avoid
frequent virtio syncing with host that will be bad for performance
otherwise. Virtio-gpu kernel driver will maintain a list of purgeable
blobs and will send the list of idling blobs down to host after a period
of inactivity.

Virgl may be not interesting for CrOS, but Qemu will continue supporting
it. I also expect that today's ARM Chromebooks which use Virgl and only
support GL will continue to use Virgl for the next 4 years.

> Android specific idea: I wonder if we could tie GEM blob buffers usage to
> the lmkd and kill based on that ... ?
> 
> https://source.android.com/docs/core/perf/lmkd
> 
> Is there GEM buffer accounting infrastructure already?

Are you talking about killing a guest APP that uses host blobs when host
is under pressure? I'm not familiar with how arcvm works, but may assume
that it runs a VM per APP. In that case VM is just another process from
the lmkd perspective that will be taken down on OOM, i.e. blob buffers
already should be seen as a part of a VM's memory by lmkd when they
reside in sysmem.

-- 
Best regards,
Dmitry