[RFC PATCH 3/5] mm/vma: add support for peer to peer to device vma

[Jason Gunthorpe]

On Tue, Jan 29, 2019 at 07:08:06PM -0500, Jerome Glisse wrote:
> On Tue, Jan 29, 2019 at 11:02:25PM +0000, Jason Gunthorpe wrote:
> > On Tue, Jan 29, 2019 at 03:44:00PM -0500, Jerome Glisse wrote:
> > 
> > > > But this API doesn't seem to offer any control - I thought that
> > > > control was all coming from the mm/hmm notifiers triggering p2p_unmaps?
> > > 
> > > The control is within the driver implementation of those callbacks. 
> > 
> > Seems like what you mean by control is 'the exporter gets to choose
> > the physical address at the instant of map' - which seems reasonable
> > for GPU.
> > 
> > 
> > > will only allow p2p map to succeed for objects that have been tagged by the
> > > userspace in some way ie the userspace application is in control of what
> > > can be map to peer device.
> > 
> > I would have thought this means the VMA for the object is created
> > without the map/unmap ops? Or are GPU objects and VMAs unrelated?
> 
> GPU object and VMA are unrelated in all open source GPU driver i am
> somewhat familiar with (AMD, Intel, NVidia). You can create a GPU
> object and never map it (and thus never have it associated with a
> vma) and in fact this is very common. For graphic you usualy only
> have hand full of the hundreds of GPU object your application have
> mapped.

I mean the other way does every VMA with a p2p_map/unmap point to
exactly one GPU object?

ie I'm surprised you say that p2p_map needs to have policy, I would
have though the policy is applied when the VMA is created (ie objects
that are not for p2p do not have p2p_map set), and even for GPU
p2p_map should really only have to do with window allocation and pure
'can I even do p2p' type functionality.

> Idea is that we can only ask exporter to be predictable and still allow
> them to fail if things are really going bad.

I think hot unplug / PCI error recovery is one of the 'really going
bad' cases..

> I think i put it in the comment above the ops but in any cases i should
> write something in documentation with example and thorough guideline.
> Note that there won't be any mmu notifier to mmap of a device file
> unless the device driver calls for it or there is a syscall like munmap
> or mremap or mprotect well any syscall that work on vma.

This is something we might need to explore, does calling
zap_vma_ptes() invoke enough notifiers that a MMU notifiers or HMM
mirror consumer will release any p2p maps on that VMA?

> If we ever want to support full pin then we might have to add a
> flag so that GPU driver can refuse an importer that wants things
> pin forever.

This would become interesting for VFIO and RDMA at least - I don't
think VFIO has anything like SVA so it would want to import a p2p_map
and indicate that it will not respond to MMU notifiers.

GPU can refuse, but maybe RDMA would allow it...

Jason