Re: [RFC PATCH v3 0/8] mm,numa: N_PRIVATE node isolation for device-managed memory

[Gregory Price <gourry@gourry.net>]

On Tue, Jan 13, 2026 at 02:24:40PM +1100, Balbir Singh wrote:
> On 1/13/26 12:30, Gregory Price wrote:
> > On Mon, Jan 12, 2026 at 05:17:53PM -0800, dan.j.williams@intel.com wrote:
> >>
> >> I think what Balbir is saying is that the _PUBLIC is implied and can be
> >> omitted. It is true that N_MEMORY[_PUBLIC] already indicates multi-zone
> >> support. So N_MEMORY_PRIVATE makes sense to me as something that it is
> >> distinct from N_{HIGH,NORMAL}_MEMORY which are subsets of N_MEMORY.
> >> Distinct to prompt "go read the documentation to figure out why this
> >> thing looks not like the others".
> > 
> > Ah, ack.  Will update for v4 once i give some thought to the compression
> > stuff and the cgroups notes.
> > 
> > I would love if the ZONE_DEVICE folks could also chime in on whether the
> > callback structures for pgmap and hmm might be re-usable here, but might
> > take a few more versions to get the attention of everyone.
> > 
> 
> I see ZONE_DEVICE as a parallel construct to N_MEMORY_PRIVATE. ZONE_DEVICE
> is memory managed by devices and already isolated from the allocator. Do you
> see a need for both? I do see the need for migration between the two, but
> I suspect you want to have ZONE_DEVICE as a valid zone inside of N_MEMORY_PRIVATE?
> 

I see N_MEMORY_PRIVATE replacing some ZONE_DEVICE patterns.

N_MEMORY_PRIVATE essentially means some driver controls how allocation
occurs, and some components of mm/ can be enlightened to allow certain
types of N_MEMORY_PRIVATE nodes to be used directly (e.g. DEMOTE_ONLY
nodes could be used by vmscan.c but not by page_alloc.c as a fallback
node).

But you could totally have a driver hotplug an N_PRIVATE node and not
register the NID anywhere.  In that case the driver would allow
allocation either via something like

fd = open("/dev/my_driver_file", ...);
buf = mmap(fd, ...);
buf[0] = 0x0;
/* Fault into driver, driver allocates w/ __GFP flag for private node */

or just some ioctl()

ioctl(fd, ALLOC_SOME_MEMORY, ...);

The driver wouldn't have to reimplement allocator logic, and could
register its own set of callbacks to manage how the memory is allowed to
be mapped into page tables and such (my understanding is hmm.c already
has some of this control, that could be re-used - and pgmap exists for
ZONE_DEVICE, this could be re-used in some way).

~Gregory