On Wed, Jan 15, 2025 at 11:06:53AM +0100, Christian König wrote:

Am 15.01.25 um 09:55 schrieb Simona Vetter:

quoted

quoted

quoted

If we add something
new, we need clear rules and not just "here's the kvm code that uses it".
That's how we've done dma-buf at first, and it was a terrible mess of
mismatched expecations.

Yes, that would be wrong. It should be self defined within dmabuf and
kvm should adopt to it, move semantics and all.

Ack.

I feel like we have a plan here.

I think I have to object a bit on that.

quoted

  Summary from my side:

- Sort out pin vs revocable vs dynamic/moveable semantics, make sure
   importers have no surprises.

- Adopt whatever new dma-api datastructures pops out of the dma-api
   reworks.

- Add pfn based memory access as yet another optional access method, with
   helpers so that exporters who support this get all the others for free.

I don't see a strict ordering between these, imo should be driven by
actual users of the dma-buf api.

Already done:

- dmem cgroup so that we can resource control device pinnings just landed
   in drm-next for next merge window. So that part is imo sorted and we can
   charge ahead with pinning into device memory without all the concerns
   we've had years ago when discussing that for p2p dma-buf support.

   But there might be some work so that we can p2p pin without requiring
   dynamic attachments only, I haven't checked whether that needs
   adjustment in dma-buf.c code or just in exporters.

Anything missing?

Well as far as I can see this use case is not a good fit for the DMA-buf
interfaces in the first place. DMA-buf deals with devices and buffer
exchange.

What's necessary here instead is to give an importing VM full access on some
memory for their specific use case.

That full access includes CPU and DMA mappings, modifying caching
attributes, potentially setting encryption keys for specific ranges etc....
etc...

In other words we have a lot of things the importer here should be able to
do which we don't want most of the DMA-buf importers to do.

This proposal isn't about forcing existing exporters to allow importers to
do new stuff. That stays as-is, because it would break things.

It's about adding yet another interface to get at the underlying data, and
we have tons of those already. The only difference is that if we don't
butcher the design, we'll be able to implement all the existing dma-buf
interfaces on top of this new pfn interface, for some neat maximal
compatibility.

But fundamentally there's never been an expectation that you can take any
arbitrary dma-buf and pass it any arbitrary importer, and that is must
work. The fundamental promise is that if it _does_ work, then
- it's zero copy
- and fast, or as fast as we can make it

I don't see this any different than all the much more specific prposals
and existing code, where a subset of importers/exporters have special
rules so that e.g. gpu interconnect or vfio uuid based sharing works.
pfn-based sharing is just yet another flavor that exists to get the max
amount of speed out of interconnects.

Cheers, Sima

The semantics for things like pin vs revocable vs dynamic/moveable seems
similar, but that's basically it.

As far as I know the TEE subsystem also represents their allocations as file
descriptors. If I'm not completely mistaken this use case most likely fit's
better there.

quoted

I feel like this is small enough that m-l archives is good enough. For
some of the bigger projects we do in graphics we sometimes create entries
in our kerneldoc with wip design consensus and things like that. But
feels like overkill here.

quoted

My general desire is to move all of RDMA's MR process away from
scatterlist and work using only the new DMA API. This will save *huge*
amounts of memory in common workloads and be the basis for non-struct
page DMA support, including P2P.

Yeah a more memory efficient structure than the scatterlist would be
really nice. That would even benefit the very special dma-buf exporters
where you cannot get a pfn and only the dma_addr_t, altough most of those
(all maybe even?) have contig buffers, so your scatterlist has only one
entry. But it would definitely be nice from a design pov.

Completely agree on that part.

Scatterlist have a some design flaws, especially mixing the input and out
parameters of the DMA API into the same structure.

Additional to that DMA addresses are basically missing which bus they belong
to and details how the access should be made (e.g. snoop vs no-snoop
etc...).

quoted

Aside: A way to more efficiently create compressed scatterlists would be
neat too, because a lot of drivers hand-roll that and it's a bit brittle
and kinda silly to duplicate. With compressed I mean just a single entry
for a contig range, in practice thanks to huge pages/folios and allocators
trying to hand out contig ranges if there's plenty of memory that saves a
lot of memory too. But currently it's a bit a pain to construct these
efficiently, mostly it's just a two-pass approach and then trying to free
surplus memory or krealloc to fit. Also I don't have good ideas here, but
dma-api folks might have some from looking at too many things that create
scatterlists.

I mailed with Christoph about that a while back as well and we both agreed
that it would probably be a good idea to start defining a data structure to
better encapsulate DMA addresses.

It's just that nobody had time for that yet and/or I wasn't looped in in the
final discussion about it.

Regards,
Christian.

quoted

-Sima

-- 
Simona Vetter
Software Engineer, Intel Corporation
http://blog.ffwll.ch