[PATCH v4 00/10] Add MEMORY_DEVICE_COHERENT for coherent device memory mapping

Jason Gunthorpe jgg at nvidia.com
Wed Feb 2 15:42:32 UTC 2022


On Wed, Feb 02, 2022 at 03:57:50PM +0100, Christoph Hellwig wrote:
> On Thu, Jan 27, 2022 at 02:32:58PM -0800, Andrew Morton wrote:
> > On Wed, 26 Jan 2022 21:09:39 -0600 Alex Sierra <alex.sierra at amd.com> wrote:
> > 
> > > This patch series introduces MEMORY_DEVICE_COHERENT, a type of memory
> > > owned by a device that can be mapped into CPU page tables like
> > > MEMORY_DEVICE_GENERIC and can also be migrated like
> > > MEMORY_DEVICE_PRIVATE.
> > 
> > Some more reviewer input appears to be desirable here.
> > 
> > I was going to tentatively add it to -mm and -next, but problems. 
> > 5.17-rc1's mm/migrate.c:migrate_vma_check_page() is rather different
> > from the tree you patched.  Please redo, refresh and resend?
> 
> I really hate adding more types with the weird one off page refcount.
> We need to clean that mess up first.

Is there anyone who could give an outline of what is needed to make
fsdax use compound pages/folios for its PMD stuff?

I already suggested removing that as a way forward, and was shot down,
but nobody is standing up to maintain this code and fix it :(

We got devdax and the DRM stuff fixed now, so FSDAX is the next
blocker on this work.

The people who want this to advance have no idea about FSs or what to
do, unfortunately.

Jason


More information about the amd-gfx mailing list