Hey,
On 2023-05-11 12:14, Tvrtko Ursulin wrote:
>
> On 10/05/2023 19:46, Tejun Heo wrote:
>> Hello,
>>
>> On Wed, May 10, 2023 at 04:59:01PM +0200, Maarten Lankhorst wrote:
>>> The misc controller is not granular enough. A single computer may have any
>>> number of
>>> graphics cards, some of them with multiple regions of vram inside a single
>>> card.
>>
>> Extending the misc controller to support dynamic keys shouldn't be that
>> difficult.
>>
>> ...
>>> In the next version, I will move all the code for handling the resource
>>> limit to
>>> TTM's eviction layer, because otherwise it cannot handle the resource limit
>>> correctly.
>>>
>>> The effect of moving the code to TTM, is that it will make the code even
>>> more generic
>>> for drivers that have vram and use TTM. When using TTM, you only have to
>>> describe your
>>> VRAM, update some fields in the TTM manager and (un)register your device
>>> with the
>>> cgroup handler on (un)load. It's quite trivial to add vram accounting to
>>> amdgpu and
>>> nouveau. [2]
>>>
>>> If you want to add a knob for scheduling weight for a process, it makes
>>> sense to
>>> also add resource usage as a knob, otherwise the effect of that knob is very
>>> limited. So even for Tvrtko's original proposed usecase, it would make
>>> sense.
>>
>> It does make sense but unlike Tvrtko's scheduling weights what's being
>> proposed doesn't seem to encapsulate GPU memory resource in a generic enough
>> manner at least to my untrained eyes. ie. w/ drm.weight, I don't need any
>> specific knoweldge of how a specific GPU operates to say "this guy should
>> get 2x processing power over that guy". This more or less holds for other
>> major resources including CPU, memory and IO. What you're proposing seems a
>> lot more tied to hardware details and users would have to know a lot more
>> about how memory is configured on that particular GPU.
>>
>> Now, if this is inherent to how all, or at least most, GPUs operate, sure,
>> but otherwise let's start small in terms of interface and not take up space
>> which should be for something universal. If this turns out to be the way,
>> expanding to take up the generic interface space isn't difficult.
>>
>> I don't know GPU space so please educate me where I'm wrong.
>
> I was hoping it might be passable in principle (in some form, pending
> discussion on semantics) given how Maarten's proposal starts with only very
> specific per-device-per-memory regions controls, which is applicable to many
> devices. And hard limit at that, which probably has less complicated
> semantics, or at least implementation.
>
> My understanding of the proposal is that the allocation either fits, or it
> evicts from the parent's hierarchy (if possible) and then fits, or it fails.
> Which sounds simple enough.
Yeah, for vram itś that simple. I think for mapped and sysmem regions we may
require the
possiblity to ignore the limits as well, as it's possible to move to those
regions from
eviction. We probably don't want eviction to fail because of too low limits.
> I do however agree that it is a limited use case. So from the negative side
> of the debating camp I have to ask if this use case could be simply satisfied
> by providing a driver/device global over commit yes/no control? In other
> words, is it really important to partition the vram space ahead of time, and
> from the kernel side too? Wouldn't the relevant (highly specialised)
> applications work just fine with global over commit disabled? Even if the
> option to configure their maximum allowed working set from the userspace side
> was needed.
Disabling overcommit? Do you mean pinning the memory workload? This causes a
denial of service if
done without limits, and that's what we're trying to avoid. There is no need
for immunity from
eviction. Overcommit is still useful inside the cgroup itself, we only want
immunity from being
evicted by another process performing another workload.
> Or if we conclude cgroup controller is the way to go, would adding less
> specific limits make it more palatable? I am thinking here some generic
> "device memory resident". Not per device, not per memory region. So
> userspace/admins have some chance of configuring generic limits. That would
> require coming up with more generic use cases though so another thing to
> discuss. Like who would use that and for what.
You would run into ambiguity with that. I think it's fine to assume any number
of vram regions. In most cases, the number is 1.
I don't see that number going much higher, 2 is already on the high side. A
simple controller could just half each region separately.
> Assuming also we can agree that "device memory resident" is a stable/solid
> concept across drm. Should be easier than for integrated GPUs, for which I
> have to admit I currently don't remember if allocations are already
> consistently covered by the memory controller. Even if they are ownership is
> probably
