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

[Jerome Glisse]

On Wed, Jan 30, 2019 at 10:33:39AM +0000, Koenig, Christian wrote:
> Am 30.01.19 um 09:02 schrieb Christoph Hellwig:
> > On Tue, Jan 29, 2019 at 08:58:35PM +0000, Jason Gunthorpe wrote:
> >> On Tue, Jan 29, 2019 at 01:39:49PM -0700, Logan Gunthorpe wrote:
> >>
> >>> implement the mapping. And I don't think we should have 'special' vma's
> >>> for this (though we may need something to ensure we don't get mapping
> >>> requests mixed with different types of pages...).
> >> I think Jerome explained the point here is to have a 'special vma'
> >> rather than a 'special struct page' as, really, we don't need a
> >> struct page at all to make this work.
> >>
> >> If I recall your earlier attempts at adding struct page for BAR
> >> memory, it ran aground on issues related to O_DIRECT/sgls, etc, etc.
> > Struct page is what makes O_DIRECT work, using sgls or biovecs, etc on
> > it work.  Without struct page none of the above can work at all.  That
> > is why we use struct page for backing BARs in the existing P2P code.
> > Not that I'm a particular fan of creating struct page for this device
> > memory, but without major invasive surgery to large parts of the kernel
> > it is the only way to make it work.
> 
> The problem seems to be that struct page does two things:
> 
> 1. Memory management for system memory.
> 2. The object to work with in the I/O layer.
> 
> This was done because a good part of that stuff overlaps, like reference 
> counting how often a page is used.  The problem now is that this doesn't 
> work very well for device memory in some cases.
> 
> For example on GPUs you usually have a large amount of memory which is 
> not even accessible by the CPU. In other words you can't easily create a 
> struct page for it because you can't reference it with a physical CPU 
> address.
> 
> Maybe struct page should be split up into smaller structures? I mean 
> it's really overloaded with data.

I think the simpler answer is that we do not want to allow GUP or any-
thing similar to pin BAR or device memory. Doing so can only hurt us
long term by fragmenting the GPU memory and forbidding us to move thing
around. For transparent use of device memory within a process this is
definitly forbidden to pin.

I do not see any good reasons we would like to pin device memory for
the existing GPU GEM objects. Userspace always had a very low expectation
on what it can do with mmap of those object and i believe it is better
to keep expectation low here and says nothing will work with those
pointer. I just do not see a valid and compelling use case to change
that :)

Even outside GPU driver, device driver like RDMA just want to share their
doorbell to other device and they do not want to see those doorbell page
use in direct I/O or anything similar AFAICT.

Cheers,
Jérôme