On Wed, Apr 12, 2017 at 4:22 AM, Dan Williams [off-list ref] wrote:

On Tue, Apr 11, 2017 at 10:42 AM, Oliver O'Halloran [off-list ref] wrote:

quoted

Hi all,

This series adds support for ZONE_DEVICE and the pmem api on powerpc. Namely,
support for altmaps and the various bits and pieces required for DAX PMD faults.
The first two patches touch generic mm/ code, but otherwise this is fairly well
contained in arch/powerpc.

If the nvdimm folks could sanity check this series I'd appreciate it.

Quick feedback: I'm in the process of cleaning up and resubmitting my
patch set to push the pmem api down into the driver directly.

    https://lwn.net/Articles/713064/

That's been on my radar for a while and I was hoping it would be in
4.12. Moving operations into the driver makes a lot of sense from a
design perspective and it should make supporting some of the
contutto's eccentricities a bit easier.

I'm also reworking memory hotplug to allow sub-section allocations
which has collided with Michal Hocko's hotplug reworks. It will be
good to have some more eyes on that work to understand the cross-arch
implications.

    https://lkml.org/lkml/2017/3/19/146

I'd been putting off looking at this since I figured it would clash
with the hotplug rework and HMM, but I'll see if I can get it working
on ppc.

quoted

Series is based on next-20170411, but it should apply elsewhere with minor
fixups to arch_{add|remove}_memory due to conflicts with HMM.  For those
interested in testing this, there is a driver and matching firmware that carves
out some system memory for use as an emulated Con Tutto memory card.

Driver: https://github.com/oohal/linux/tree/contutto-next
Firmware: https://github.com/oohal/skiboot/tree/fake-contutto

Edit core/init.c:686 to control the amount of memory borrowed for the emulated
device.  I'm keeping the driver out of tree for a until 4.13 since I plan on
reworking the firmware interface anyway and There's at least one showstopper
bug.

Is this memory card I/O-cache coherent? I.e. existing dma mapping api
can hand out mappings to it? Just trying to figure out if this the
existing pmem-definition of ZONE_DEVICE or a new one.

As far as the rest of the system is concerned Con Tutto memory is
identical to normal system memory. All accesses to the card's memory
is mediated by a memory controller which participates in the memory
coherency protocol. That said, the link between the card and the
system is non-coherent so logic on the FPGA can access memory
incoherently. I'm primarily interested in using the card as a memory
platform so I haven't spent a lot of time thinking about the latter
use case, but a different concept of device memory might be required
there.

Oliver