Re: [PATCH 00/60] Private NUMA nodes

From: Yury Norov

Date: Sun Nov 22 2026 - 16:12:47 EST


On Sun, Nov 22, 2026 at 02:11:31PM -0500, Gregory Price (Meta) wrote:
> On 2026-11-22 11:51 -0500, Yury Norov wrote:
> > On Fri, Nov 20, 2026 at 01:08:51PM -0500, Gregory Price via B4 Relay wrote:
> > > This is v6 of the Private NUMA node series. The series falls into four
> > > parts, each building on the one before:
> > >
> > > 1-18 mm: introduce N_MEMORY_COMMON
> >
> > I walked through this part, mostly. It looks like you can substantially
> > decrease the patch count if you keep N_MEMORY meaning unchanged, and
> > add another type, like N_MEMORY_NOALLOC.
> >
>
> I don't see how this would reduce the patch count substatially, you
> still need to convert existing N_MEMORY users and need to differentiate
> between general purpose memory nodes and private memory nodes.

Please check this quick attempt with my codex. The patch count is reduced
to 52:

https://github.com/norov/linux/tree/private-nodes-v6-n-memory-folded

The brief summary for dropped/simplified patches is:

02 -> dropped - State helpers: no publication/clearing helpers needed.
06–11 -> dropped - Migration, tiers, DAMON, scheduler, stack locality:
existing N_MEMORY checks already select ordinary memory.
14–15 -> dropped - Hash distribution, kswapd recovery: existing N_MEMORY
counts remain correct.
03 -> 02 - Common-memory state: documentation only; no new state or sysfs
attribute.
05 -> 04 - Interleave/tmpfs policies: only the three-line weighted-interleave
hotplug filter remains.
12 -> 05 - Reclaim: only hotplug kswapd gating remains; iterator conversion
moves to the reclaim-state patch.
13 -> 06 - Compaction: only system-wide/proactive filtering and hotplug
kcompactd gating remain.
16 -> 07 - OOM constraints: only capacity-summing intersections remain.
27 -> 18 - Private-memory state: expands to 16 files; topology/accounting
users must include both memory sets.
31 -> 22 - Cpuset handling: requires ordinary memory while preserving
allowed private nodes in the returned mask.
44 -> 36,
46 -> 38,
50 -> 42 - Service states: direct updates at boot,
online, and offline sites replace shared helpers, increasing
duplication.

The new N_MEMORY_NOALLOC patch is not split intentionally. If you do it, it
would probably turn into 4 individual patches.

The more important thing is that N_MEMORY and the proposed N_MEMORY_NOALLOC
is a disjoint, contrary to N_MEMORY and N_MEMORY_COMMON.

> And N_MEMORY_NOALLOC is not a good name.

Sure, I'm not attached to any particular name. Although,
N_MEMORY_COMMON nothing better to me. It answers the question
'what for this thing' even worse.

Thanks,
Yury