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

[Jerome Glisse <jglisse@redhat.com>]

On Tue, Jan 29, 2019 at 08:24:36PM +0000, Jason Gunthorpe wrote:
> On Tue, Jan 29, 2019 at 02:50:55PM -0500, Jerome Glisse wrote:
> 
> > GPU driver do want more control :) GPU driver are moving things around
> > all the time and they have more memory than bar space (on newer platform
> > AMD GPU do resize the bar but it is not the rule for all GPUs). So
> > GPU driver do actualy manage their BAR address space and they map and
> > unmap thing there. They can not allow someone to just pin stuff there
> > randomly or this would disrupt their regular work flow. Hence they need
> > control and they might implement threshold for instance if they have
> > more than N pages of bar space map for peer to peer then they can decide
> > to fall back to main memory for any new peer mapping.
> 
> 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. So
driver implementation can refuse to map by returning an error on p2p_map
or it can decide to use main memory by migrating its object to main memory
and populating the dma address array with dma_page_map() of the main memory
pages. Driver like GPU can have policy on top of that for instance they
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. This is needed for GPU driver as we do want
userspace involvement on what object are allowed to have p2p access and
also so that we can report to userspace when we are running out of BAR
addresses for this to work as intended (ie not falling back to main memory)
so that application can take appropriate actions (like decide what to
prioritize).

For moving things around after a successful p2p_map yes the exporting
device have to call for instance zap_vma_ptes() or something similar.
This will trigger notifier call and the importing device will invalidate
its mapping. Once it is invalidated then the exporting device can
point new call of p2p_map (for the same range) to new memory (obviously
the exporting device have to synchronize any concurrent call to p2p_map
with the invalidation).

> 
> I would think that the importing driver can assume the BAR page is
> kept alive until it calls unmap (presumably triggered by notifiers)?
> 
> ie the exporting driver sees the BAR page as pinned until unmap.

The intention with this patchset is that it is not pin ie the importer
device _must_ abide by all mmu notifier invalidations and they can
happen at anytime. The importing device can however re-p2p_map the
same range after an invalidation.

I would like to restrict this to importer that can invalidate for
now because i believe all the first device to use can support the
invalidation.

Also when using HMM private device memory we _can not_ pin virtual
address to device memory as otherwise CPU access would have to SIGBUS
or SEGFAULT and we do not want that. So this was also a motivation to
keep thing consistent for the importer for both cases.

Cheers,
Jérôme