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

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

On 3/9/22 01:24, Rob Clark wrote:
> On Tue, Mar 8, 2022 at 11:28 AM Dmitry Osipenko
> <dmitry.osipenko@collabora.com> wrote:
>>
>> On 3/8/22 19:29, Rob Clark wrote:
>>> On Tue, Mar 8, 2022 at 5:17 AM Dmitry Osipenko
>>> <dmitry.osipenko@collabora.com> wrote:
>>>>
>>>> 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.
>>>
>>> Will host memory pressure already trigger shrinker in guest?
>>
>> The host memory pressure won't trigger shrinker in guest here. This
>> series will help only with the memory pressure within the guest using a
>> usual "virgl context".
>>
>> Having a host shrinker in a case of "virgl contexts" should be a
>> difficult problem to solve.
> 
> Hmm, I think we just need the balloon driver to trigger the shrinker
> in the guest kernel?  I suppose a driver like drm/virtio might want to
> differentiate between host and guest pressure (ie. consider only
> objects that have host vs guest storage), but even without that,
> freeing up memory in the guest when host is under memory pressure
> seems worthwhile.  Maybe I'm over-simplifying?

Might be the opposite, i.e. me over-complicating :) The variant with
memory ballooning actually could be good and will work for all kinds of
virtio contexts universally. There will be some back-n-forth between
host and guest, but perhaps it will work okay. Thank you for the suggestion.

>>> This is
>>> something I'm quite interested in for "virtgpu native contexts" (ie.
>>> native guest driver with new context type sitting on top of virtgpu),
>>
>> In a case of "native contexts" it should be doable, at least I can't see
>> any obvious problems. The madvise invocations could be passed to the
>> host using a new virtio-gpu command by the guest's madvise IOCTL
>> handler, instead-of/in-addition-to handling madvise in the guest's
>> kernel, and that's it.
> 
> I think we don't want to do that, because MADV:WILLNEED would be by
> far the most frequent guest<->host synchronous round trip.  So from
> that perspective tracking madvise state in guest kernel seems quite
> attractive.

This is a valid concern. I'd assume that the overhead should be
tolerable, but I don't have any actual perf numbers.

> If we really can't track madvise state in the guest for dealing with
> host memory pressure, I think the better option is to introduce
> MADV:WILLNEED_REPLACE, ie. something to tell the host kernel that the
> buffer is needed but the previous contents are not (as long as the GPU
> VA remains the same).  With this the host could allocate new pages if
> needed, and the guest would not need to wait for a reply from host.

If variant with the memory ballooning will work, then it will be
possible to track the state within guest-only. Let's consider the
simplest variant for now.

I'll try to implement the balloon driver support in the v2 and will get
back to you.

>>> since that isn't using host storage
>>
>> s/host/guest ?
> 
> Yes, sorry, I meant that it is not using guest storage.

Thank you for the clarification.