Re: [PATCH v1 0/5] Add memory shrinker to VirtIO-GPU DRM driver

[Dmitry Osipenko <dmitry.osipenko@collabora.com>]

On 3/9/22 22:28, Thomas Zimmermann wrote:
> Hi
> 
> Am 09.03.22 um 12:55 schrieb Dmitry Osipenko:
>> Hello,
>>
>> On 3/9/22 11:59, Thomas Zimmermann wrote:
>>> Hi
>>>
>>> Am 08.03.22 um 14:17 schrieb Dmitry Osipenko:
>>>> Hello,
>>>>
>>>> This patchset introduces memory shrinker for the VirtIO-GPU DRM driver.
>>>> During OOM, the shrinker will release BOs that are marked as "not
>>>> needed"
>>>> by userspace using the new madvise IOCTL. The userspace in this case is
>>>> the Mesa VirGL driver, it will mark the cached BOs as "not needed",
>>>> allowing kernel driver to release memory of the cached shmem BOs on
>>>> lowmem
>>>> situations, preventing OOM kills.
>>>
>>> Virtio-gpu is build on top of GEM shmem helpers. I have a prototype
>>> patchset that adds a shrinker to these helpers. If you want to go
>>> further, you could implement something like that instead. Panfrost and
>>> lima also have their own shrinker and could certainly be converted to
>>> the gem-shmem shrinker.
>>
>> I had a thought that it could be possible to unify shrinkers into a
>> common DRM framework. Could you please give me a link to yours prototype
>> patchset?
> 
> I uploaded the patches to
> 
> 
> https://gitlab.freedesktop.org/tzimmermann/linux/-/commits/gem-shmem-cached-mappings
> 
> 
> it's incomplete and un-debugged, but it shows what needs to be done. It
> has the infrastructure, but lacks the changes to the GEM shmem code.
> 
> The reason for this work is to keep GEM shmem pages mapped and allocated
> even while the BO is neither mapped nor pinned.  As it is now, GEM SHMEM
> creates and releases pages on each pin and unpin, and maps and unmaps
> memory ranges on each vmap and vunmap.  It's all wasteful. Only the
> first pin and vmap calls should establish pages and mappings and only
> the purge and free functions should release them.

Hm, aren't maps and pins already refcounted?

> The patchset adds new helpers for BO purging to struct
> drm_gem_object_funcs. With this, I think it might be possible to have
> one global DRM shrinker and let it handle all BOs; independent of each
> BO's memory manager.

Thank you, I'll give it a try.