Re: [PATCH v2 1/2] mm/huge_memory.c: skip zero-page remapping when underused THP shrinker is disabled

From: Nico Pache

Date: Mon Nov 16 2026 - 18:09:03 EST




On 11/16/26 3:30 PM, David Hildenbrand (Arm) wrote:
On 10/25/26 18:39, Nico Pache (Red Hat) wrote:
The zero-page remapping in __folio_split() was added to support the
underused THP shrinker. When the shrinker is disabled via
shrink_underused, gate TTU_USE_SHARED_ZEROPAGE on split_underused_thp
to avoid scanning subpages for zero content unnecessarily.

Suggested-by: David Hildenbrand <david@xxxxxxxxxx>
Fixes: b1f202060afe ("mm: remap unused subpages to shared zeropage when splitting isolated thp")
Signed-off-by: Nico Pache <nico.pache@xxxxxxxxx>
---
mm/huge_memory.c | 3 ++-
1 file changed, 2 insertions(+), 1 deletion(-)

diff --git a/mm/huge_memory.c b/mm/huge_memory.c
index bd44d58e8684..2d56d1bff0ea 100644
--- a/mm/huge_memory.c
+++ b/mm/huge_memory.c
@@ -4104,7 +4104,8 @@ static int __folio_freeze_split_anon(struct folio *folio,
out_no_split:
local_irq_enable();
if (anon_vma) {
- if (!ret && !folio_is_device_private(folio))
+ if (!ret && !folio_is_device_private(folio) &&
+ split_underused_thp)
ttu_flags = TTU_USE_SHARED_ZEROPAGE;
remap_anon_folio(folio, 1 << old_order, ttu_flags);
}


I guess the change is not that relevant in practice, because I suspect most
people always run with split_underused_thp=true

Yeah these two patches are only tangentially related. Technically with this patch (1), disabling the underutilized shrinker would indeed fix it, as we no longer are doing zero-page remapping. But the root cause is the split initiated from KSM.

Most people always run with split_underused_thp=true, but they also run with 511 max_pte_nones.

Th question is whether "split_underused_thp" only determines "that we trigger a
split if underused", or also "what to do when splitting a folio".

Yes, it was added for the deferred shrinker, but the semantics of the toggle are
not quite clear. I don't think there is any documentation for it ...

Yeah Im not really sure if the intentions were to make all the splits do this zero-page remapping. And im not fully aware either of the potential impact to the system (whether positive or negative) this could have system wide. You could argue it keeps better system utilization, but you could also argue the cmp is expensive.


Disabling the "split to zeropage" entirely with the toggle off (behavior before
the toggle was introduced) makes sense to me.

But the behavior now already like that in the kernel for a while.


I know that I suggested this, but I wonder, is this patch relevant for your use
case and patch #2?

Technically patch #2 solves the issue im actually trying to tackle. But during this case we also found this functional change which we werent sure if it was intended or not. As you say its poorly defined.

Im ok with or without this patch. Perhaps we can take patch #2 (bar'ing review), and I can try to do some perf testing regarding the effect of this zeropage remapping system wide, and base our judgement off that.


IOW, woudl RHEL customers disable this toggle here when running with KSM on
affected systems?

That would be a potential solution if we only took patch #1. patch #2 is also viable (and a better) solution to solve the KSM regression while keep the shrinker available.

-- Nico