Re: [RFC 1/5] mm/hmm: HMM API to enable P2P DMA for device private pages

["Thomas Hellström" <thomas.hellstrom@linux.intel.com>]

On Tue, 2025-01-28 at 11:16 -0400, Jason Gunthorpe wrote:
> On Tue, Jan 28, 2025 at 03:48:54PM +0100, Thomas Hellström wrote:
> > On Tue, 2025-01-28 at 09:20 -0400, Jason Gunthorpe wrote:
> > > On Tue, Jan 28, 2025 at 09:51:52AM +0100, Thomas Hellström wrote:
> > > 
> > > > How would the pgmap device know whether P2P is actually
> > > > possible
> > > > without knowing the client device, (like calling
> > > > pci_p2pdma_distance)
> > > > and also if looking into access control, whether it is allowed?
> > > 
> > > The DMA API will do this, this happens after this patch is put on
> > > top
> > > of Leon's DMA API patches. The mapping operation will fail and it
> > > will
> > > likely be fatal to whatever is going on.
> > >  
> > > get_dma_pfn_for_device() returns a new PFN, but that is not a DMA
> > > mapped address, it is just a PFN that has another struct page
> > > under
> > > it.
> > > 
> > > There is an implicit assumption here that P2P will work and we
> > > don't
> > > need a 3rd case to handle non-working P2P..
> > 
> > OK. We will have the case where we want pfnmaps with driver-private
> > fast interconnects to return "interconnect possible, don't migrate"
> > whereas possibly other gpus and other devices would return
> > "interconnect unsuitable, do migrate", so (as I understand it)
> > something requiring a more flexible interface than this.
> 
> I'm not sure this doesn't handle that case?
> 
> Here we are talking about having DEVICE_PRIVATE struct page
> mappings. On a GPU this should represent GPU local memory that is
> non-coherent with the CPU, and not mapped into the CPU.
> 
> This series supports three case:
> 
>  1) pgmap->owner == range->dev_private_owner
>     This is "driver private fast interconnect" in this case HMM
> should
>     immediately return the page. The calling driver understands the
>     private parts of the pgmap and computes the private interconnect
>     address.
> 
>     This requires organizing your driver so that all private
>     interconnect has the same pgmap->owner.

Yes, although that makes this map static, since pgmap->owner has to be
set at pgmap creation time. and we were during initial discussions
looking at something dynamic here. However I think we can probably do
with a per-driver owner for now and get back if that's not sufficient.

> 
>  2) The page is DEVICE_PRIVATE and get_dma_pfn_for_device() exists.
>     The exporting driver has the option to return a P2P struct page
>     that can be used for PCI P2P without any migration. In a PCI GPU
>     context this means the GPU has mapped its local memory to a PCI
>     address. The assumption is that P2P always works and so this
>     address can be DMA'd from.

So do I understand it correctly, that the driver then needs to set up
one device_private struct page and one pcie_p2p struct page for each
page of device memory participating in this way?

> 
>  3) Migrate back to CPU memory - then eveything works.
> 
> Is that not enough? Where do you want something different?
> 
> > > > but leaves any dma- mapping or pfn mangling to be done after
> > > > the
> > > > call to hmm_range_fault(), since hmm_range_fault() really only
> > > > needs
> > > > to know whether it has to migrate to system or not.
> > > 
> > > See above, this is already the case..
> > 
> > Well what I meant was at hmm_range_fault() time only consider
> > whether
> > to migrate or not. Afterwards at dma-mapping time you'd expose the
> > alternative pfns that could be used for dma-mapping.
> 
> That sounds like you are talking about multipath, we are not really
> ready to tackle general multipath yet at the DMA API level, IMHO.
> 
> If you are just talking about your private multi-path, then that is
> already handled..

No, the issue I'm having with this is really why would
hmm_range_fault() need the new pfn when it could easily be obtained
from the device-private pfn by the hmm_range_fault() caller? The only
thing hmm_range_fault() needs to know is, again, whether to migrate or
not. But I guess if the plan is to have hmm_range_fault() call
pci_p2pdma_distance() on it, and we don't want the exporter to do that,
it makes sense.

> 
> > We were actually looking at a solution where the pagemap implements
> > something along
> > 
> > bool devmem_allowed(pagemap, client); //for hmm_range_fault
> > 
> > plus dma_map() and dma_unmap() methods.
> 
> This sounds like dmabuf philosophy, and I don't think we should go in
> this direction. The hmm caller should always be responsible for dma
> mapping and we need to improve the DMA API to make this work better,
> not build side hacks like this.
> 
> You can read my feelings and reasoning on this topic within this huge
> thread:
> 
> https://lore.kernel.org/dri-devel/20250108132358.GP5556@nvidia.com/
> 
> > In this way you'd don't need to expose special p2p dma pages and
> > the
> 
> Removing the "special p2p dma pages" has to be done by improving the
> DMA API to understand how to map phsyical addresses without struct
> page. We are working toward this, slowly.

> pgmap->ops->dma_map/unmap() ideas just repeat the DMABUF mistake
> of mis-using the DMA API for P2P cases. Today you cannot correctly
> DMA
> map P2P memory without the struct page.

Yeah, I don't want to drag hmm into that discussion, although
admittedly the idea of pgmap->ops->dma_map/unmap mimics the dma-buf
behaviour.

So anyway what we'll do is to try to use an interconnect-common owner
for now and revisit the problem if that's not sufficient so we can come
up with an acceptable solution.


/Thomas