On 1/27/23 16:17, Christian König wrote:
Am 27.01.23 um 15:44 schrieb Danilo Krummrich:
[SNIP]
What you want is one component for tracking the VA allocations
(drm_mm based) and a different component/interface for tracking the
VA mappings (probably rb tree based).
That's what the GPUVA manager is doing. There are gpuva_regions
which correspond to VA allocations and gpuvas which represent the
mappings. Both are tracked separately (currently both with a
separate drm_mm, though). However, the GPUVA manager needs to take
regions into account when dealing with mappings to make sure the
GPUVA manager doesn't propose drivers to merge over region
boundaries. Speaking from userspace PoV, the kernel wouldn't merge
mappings from different VKBuffer objects even if they're virtually
and physically contiguous.
That are two completely different things and shouldn't be handled in
a single component.
They are different things, but they're related in a way that for
handling the mappings (in particular merging and sparse) the GPUVA
manager needs to know the VA allocation (or region) boundaries.
I have the feeling there might be a misunderstanding. Userspace is in
charge to actually allocate a portion of VA space and manage it. The
GPUVA manager just needs to know about those VA space allocations and
hence keeps track of them.
The GPUVA manager is not meant to be an allocator in the sense of
finding and providing a hole for a given request.
Maybe the non-ideal choice of using drm_mm was implying something else.
Uff, well long story short that doesn't even remotely match the
requirements. This way the GPUVA manager won't be usable for a whole
bunch of use cases.
What we have are mappings which say X needs to point to Y with this and
hw dependent flags.
The whole idea of having ranges is not going to fly. Neither with AMD
GPUs and I strongly think not with Intels XA either.
A range in the sense of the GPUVA manager simply represents a VA space
allocation (which in case of Nouveau is taken in userspace). Userspace
allocates the portion of VA space and lets the kernel know about it. The
current implementation needs that for the named reasons. So, I think
there is no reason why this would work with one GPU, but not with
another. It's just part of the design choice of the manager.
And I'm absolutely happy to discuss the details of the manager
implementation though.
We should probably talk about the design of the GPUVA manager once
more when this should be applicable to all GPU drivers.
That's what I try to figure out with this RFC, how to make it
appicable for all GPU drivers, so I'm happy to discuss this. :-)
Yeah, that was really good idea :) That proposal here is really far away
from the actual requirements.
And those are the ones I'm looking for. Do you mind sharing the
requirements for amdgpu in particular?
For sparse residency the kernel also needs to know the region
boundaries to make sure that it keeps sparse mappings around.
What?
When userspace creates a new VKBuffer with the
VK_BUFFER_CREATE_SPARSE_BINDING_BIT the kernel may need to create
sparse mappings in order to ensure that using this buffer without any
memory backed mappings doesn't fault the GPU.
Currently, the implementation does this the following way:
1. Userspace creates a new VKBuffer and hence allocates a portion of
the VA space for it. It calls into the kernel indicating the new VA
space region and the fact that the region is sparse.
2. The kernel picks up the region and stores it in the GPUVA manager,
the driver creates the corresponding sparse mappings / page table
entries.
3. Userspace might ask the driver to create a couple of memory backed
mappings for this particular VA region. The GPUVA manager stores the
mapping parameters, the driver creates the corresponding page table
entries.
4. Userspace might ask to unmap all the memory backed mappings from
this particular VA region. The GPUVA manager removes the mapping
parameters, the driver cleans up the corresponding page table entries.
However, the driver also needs to re-create the sparse mappings, since
it's a sparse buffer, hence it needs to know the boundaries of the
region it needs to create the sparse mappings in.
Again, this is not how things are working. First of all the kernel
absolutely should *NOT* know about those regions.
What we have inside the kernel is the information what happens if an
address X is accessed. On AMD HW this can be:
1. Route to the PCIe bus because the mapped BO is stored in system memory.
2. Route to the internal MC because the mapped BO is stored in local
memory.
3. Route to other GPUs in the same hive.
4. Route to some doorbell to kick of other work.
...
x. Ignore write, return 0 on reads (this is what is used for sparse
mappings).
x+1. Trigger a recoverable page fault. This is used for things like SVA.
x+2. Trigger a non-recoverable page fault. This is used for things like
unmapped regions where access is illegal.
All this is plus some hw specific caching flags.
When Vulkan allocates a sparse VKBuffer what should happen is the
following:
1. The Vulkan driver somehow figures out a VA region A..B for the
buffer. This can be in userspace (libdrm_amdgpu) or kernel (drm_mm), but
essentially is currently driver specific.
Right, for Nouveau we have this in userspace as well.
2. The kernel gets a request to map the VA range A..B as sparse, meaning
that it updates the page tables from A..B with the sparse setting.
3. User space asks kernel to map a couple of memory backings at location
A+1, A+10, A+15 etc....
4. The VKBuffer is de-allocated, userspace asks kernel to update region
A..B to not map anything (usually triggers a non-recoverable fault).
Until here this seems to be identical to what I'm doing.
It'd be interesting to know how amdgpu handles everything that
potentially happens between your 3) and 4). More specifically, how are
the page tables changed when memory backed mappings are mapped on a
sparse range? What happens when the memory backed mappings are unmapped,
but the VKBuffer isn't de-allocated, and hence sparse mappings need to
be re-deployed?
Let's assume the sparse VKBuffer (and hence the VA space allocation) is
pretty large. In Nouveau the corresponding PTEs would have a rather huge
page size to cover this. Now, if small memory backed mappings are mapped
to this huge sparse buffer, in Nouveau we'd allocate a new PT with a
corresponding smaller page size overlaying the sparse mappings PTEs.
How would this look like in amdgpu?
When you want to unify this between hw drivers I strongly suggest to
completely start from scratch once more.
First of all don't think about those mappings as VMAs, that won't work
because VMAs are usually something large. Think of this as individual
PTEs controlled by the application. similar how COW mappings and struct
pages are handled inside the kernel.
Why do you consider tracking single PTEs superior to tracking VMAs? All
the properties for a page you mentioned above should be equal for the
entirety of pages of a whole (memory backed) mapping, aren't they?
Then I would start with the VA allocation manager. You could probably
base that on drm_mm. We handle it differently in amdgpu currently, but I
think this is something we could change.
It was not my intention to come up with an actual allocator for the VA
space in the sense of actually finding a free and fitting hole in the VA
space.
For Nouveau (and XE, I think) we have this in userspace and from what
you've written previously I thought the same applies for amdgpu?
Then come up with something close to the amdgpu VM system. I'm pretty
sure that should work for Nouveau and Intel XA as well. In other words
you just have a bunch of very very small structures which represents
mappings and a larger structure which combine all mappings of a specific
type, e.g. all mappings of a BO or all sparse mappings etc...
Considering what you wrote above I assume that small structures /
mappings in this paragraph refer to PTEs.
Immediately, I don't really see how this fine grained resolution of
single PTEs would help implementing this in Nouveau. Actually, I think
it would even complicate the handling of PTs, but I would need to think
about this a bit more.
Merging of regions is actually not mandatory. We don't do it in amdgpu
and can live with the additional mappings pretty well. But I think this
can differ between drivers.
Regards,
Christian.
