Re: [PATCH v4 11/15] drm/shmem-helper: Add generic memory shrinker

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

On 5/19/22 17:13, Daniel Vetter wrote:
> On Thu, May 12, 2022 at 10:04:53PM +0300, Dmitry Osipenko wrote:
>> On 5/12/22 20:04, Daniel Vetter wrote:
>>> On Thu, 12 May 2022 at 13:36, Dmitry Osipenko
>>> <dmitry.osipenko@collabora.com> wrote:
>>>>
>>>> On 5/11/22 22:09, Daniel Vetter wrote:
>>>>> On Wed, May 11, 2022 at 07:06:18PM +0300, Dmitry Osipenko wrote:
>>>>>> On 5/11/22 16:09, Daniel Vetter wrote:
>>>>>>>>>>> I'd like to ask you to reduce the scope of the patchset and build the
>>>>>>>>>>> shrinker only for virtio-gpu. I know that I first suggested to build
>>>>>>>>>>> upon shmem helpers, but it seems that it's easier to do that in a later
>>>>>>>>>>> patchset.
>>>>>>>>>> The first version of the VirtIO shrinker didn't support memory eviction.
>>>>>>>>>> Memory eviction support requires page fault handler to be aware of the
>>>>>>>>>> evicted pages, what should we do about it? The page fault handling is a
>>>>>>>>>> part of memory management, hence to me drm-shmem is already kinda a MM.
>>>>>>>>> Hm I still don't get that part, why does that also not go through the
>>>>>>>>> shmem helpers?
>>>>>>>> The drm_gem_shmem_vm_ops includes the page faults handling, it's a
>>>>>>>> helper by itself that is used by DRM drivers.
>>>>>>>>
>>>>>>>> I could try to move all the shrinker logic to the VirtIO and re-invent
>>>>>>>> virtio_gem_shmem_vm_ops, but what is the point of doing this for each
>>>>>>>> driver if we could have it once and for all in the common drm-shmem code?
>>>>>>>>
>>>>>>>> Maybe I should try to factor out all the shrinker logic from drm-shmem
>>>>>>>> into a new drm-shmem-shrinker that could be shared by drivers? Will you
>>>>>>>> be okay with this option?
>>>>>>> I think we're talking past each another a bit. I'm only bringing up the
>>>>>>> purge vs eviction topic we discussed in the other subthread again.
>>>>>>
>>>>>> Thomas asked to move the whole shrinker code to the VirtIO driver and
>>>>>> I's saying that this is not a great idea to me, or am I misunderstanding
>>>>>> the Thomas' suggestion? Thomas?
>>>>>
>>>>> I think it was just me creating a confusion here.
>>>>>
>>>>> fwiw I do also think that shrinker in shmem helpers makes sense, just in
>>>>> case that was also lost in confusion.
>>>>
>>>> Okay, good that we're on the same page now.
>>>>
>>>>>>>>> I'm still confused why drivers need to know the difference
>>>>>>>>> between evition and purging. Or maybe I'm confused again.
>>>>>>>> Example:
>>>>>>>>
>>>>>>>> If userspace uses IOV addresses, then these addresses must be kept
>>>>>>>> reserved while buffer is evicted.
>>>>>>>>
>>>>>>>> If BO is purged, then we don't need to retain the IOV space allocated
>>>>>>>> for the purged BO.
>>>>>>> Yeah but is that actually needed by anyone? If userspace fails to allocate
>>>>>>> another bo because of lack of gpu address space then it's very easy to
>>>>>>> handle that:
>>>>>>>
>>>>>>> 1. Make a rule that "out of gpu address space" gives you a special errno
>>>>>>> code like ENOSPC
>>>>>>>
>>>>>>> 2. If userspace gets that it walks the list of all buffers it marked as
>>>>>>> purgeable and nukes them (whether they have been evicted or not). Then it
>>>>>>> retries the bo allocation.
>>>>>>>
>>>>>>> Alternatively you can do step 2 also directly from the bo alloc ioctl in
>>>>>>> step 1. Either way you clean up va space, and actually a lot more (you
>>>>>>> potentially nuke all buffers marked as purgeable, not just the ones that
>>>>>>> have been purged already) and only when va cleanup is actually needed
>>>>>>>
>>>>>>> Trying to solve this problem at eviction time otoh means:
>>>>>>> - we have this difference between eviction and purging
>>>>>>> - it's still not complete, you still need to glue step 2 above into your
>>>>>>>   driver somehow, and once step 2 above is glued in doing additional
>>>>>>>   cleanup in the purge function is just duplicated logic
>>>>>>>
>>>>>>> So at least in my opinion this isn't the justification we need. And we
>>>>>>> should definitely not just add that complication "in case, for the
>>>>>>> future", if we don't have a real need right now. Adding it later on is
>>>>>>> easy, removing it later on because it just gets in the way and confuses is
>>>>>>> much harder.
>>>>>>
>>>>>> The IOVA space is only one example.
>>>>>>
>>>>>> In case of the VirtIO driver, we may have two memory allocation for a
>>>>>> BO. One is the shmem allcation in guest and the other is in host's vram.
>>>>>> If we will only release the guest's memory on purge, then the vram will
>>>>>> remain allocated until BO is destroyed, which unnecessarily sub-optimal.
>>>>>
>>>>> Hm but why don't you just nuke the memory on the host side too when you
>>>>> evict? Allowing the guest memory to be swapped out while keeping the host
>>>>> memory allocation alive also doesn't make a lot of sense for me. Both can
>>>>> be recreated (I guess at least?) on swap-in.
>>>>
>>>> Shouldn't be very doable or at least worth the efforts. It's userspace
>>>> that manages data uploading, kernel only provides transport for the
>>>> virtio-gpu commands.
>>>>
>>>> Drivers are free to use the same function for both purge() and evict()
>>>> callbacks if they want. Getting rid of the purge() callback creates more
>>>> problems than solves, IMO.
>>>
>>> Hm this still sounds pretty funny and defeats the point of
>>> purgeable/evictable buffers a bit I think. But also I guess we'd
>>> pushed this bikeshed to the max, so I think if you make ->purge
>>> optional and just call ->evict if that's not present, and document it
>>> all in the kerneldoc, then I think that's good.
>>
>> This is a good enough compromise to me.
>>
>>> I just don't think that encouraging drivers to distinguish between
>>> evict/purge is a good idea for almost all of them.
>>
>> Intel's shrinker checks the "madvise" status of BOs and then decides
>> what to do based on it. Perhaps we could move the decision-making about
>> purging to drivers and then it will be single evict() callback, but will
>> drivers really ever need to be responsible for this decision-making or
>> this will be an unnecessary boilerplate code in the drivers? I'll think
>> more about this.
> 
> tbh I wouldn't worry about details, you've convinced me that some
> differentiation between evict and purged makes sense. And yeah maybe
> drivers should have a helper to check that instead of explicit argument,
> but that's a bikeshed color choice which should be fairly easy to adjust
> later on still.

I already reworked patches like you suggested to use single evict() cb
and etc. But those are minor things, the lockings are more important.
I'm now having more h/w on my hands and yesterday found that Lima driver
was getting deadlock using the new dma-buf locking convention that I'm
working on, so I'm now re-testing all thoroughly and will send out v6
once will be confident in it.

-- 
Best regards,
Dmitry