shrink_lruvec() drives reclaim in SWAP_CLUSTER_MAX (32) chunks, but
isolate_lru_folios() may scan far more than that per call on a single
LRU.  The excess is never charged back, so shrink_lruvec() keeps
rescanning the same folios round after round.  When reclaim targets a
lower zone, the same scanner also keeps walking zone-ineligible folios
that can never satisfy the allocation, inflating nr_reclaimed into a
false progress that delays the OOM.

This series fixes both:

  [1/2] Charge the isolate overshoot against the scan quota so the
        next round skips already-scanned folios.
  [2/2] Stop scanning once too many zone-ineligible folios have been
        skipped, instead of force-isolating them.

Wupeng Ma (2):
  mm: vmscan: charge isolate overshoot against scan quota
  mm: vmscan: stop scanning ineligible folios after max_nr_skipped

 mm/vmscan.c | 48 +++++++++++++++++++++++++++++++++---------------
 1 file changed, 33 insertions(+), 15 deletions(-)

-- 
2.43.0

On Tue, 11 Aug 2026 15:08:47 +0800 Wupeng Ma <mawupeng1@huawei.com> wrote:

> shrink_lruvec() drives reclaim in SWAP_CLUSTER_MAX (32) chunks, but
> isolate_lru_folios() may scan far more than that per call on a single
> LRU.  The excess is never charged back, so shrink_lruvec() keeps
> rescanning the same folios round after round.  When reclaim targets a
> lower zone, the same scanner also keeps walking zone-ineligible folios
> that can never satisfy the allocation, inflating nr_reclaimed into a
> false progress that delays the OOM.
> 
> This series fixes both:
> 
>   [1/2] Charge the isolate overshoot against the scan quota so the
>         next round skips already-scanned folios.
>   [2/2] Stop scanning once too many zone-ineligible folios have been
>         skipped, instead of force-isolating them.

This all sounds fairly bad.  Is there some report?  Real-world test
results?  Laboratory test results?  Something to help us understand the
impact of the issues and the situations in which they occur.

Thanks.

On 周三 2026-8-12 09:42, Andrew Morton wrote:
> On Tue, 11 Aug 2026 15:08:47 +0800 Wupeng Ma <mawupeng1@huawei.com> wrote:
> 
>> shrink_lruvec() drives reclaim in SWAP_CLUSTER_MAX (32) chunks, but
>> isolate_lru_folios() may scan far more than that per call on a single
>> LRU.  The excess is never charged back, so shrink_lruvec() keeps
>> rescanning the same folios round after round.  When reclaim targets a
>> lower zone, the same scanner also keeps walking zone-ineligible folios
>> that can never satisfy the allocation, inflating nr_reclaimed into a
>> false progress that delays the OOM.
>>
>> This series fixes both:
>>
>>   [1/2] Charge the isolate overshoot against the scan quota so the
>>         next round skips already-scanned folios.
>>   [2/2] Stop scanning once too many zone-ineligible folios have been
>>         skipped, instead of force-isolating them.
> 
> This all sounds fairly bad.  Is there some report?  Real-world test
> results?  Laboratory test results?  Something to help us understand the
> impact of the issues and the situations in which they occur.
> 

Sorry — this background belongs in the cover letter and I left it
out.  Adding it here.

## Situation

We observed slow, unexpected OOM behavior during extreme stress testing,
and while digging into the reclaim code during that analysis we spotted
these latent risks in isolate_lru_folios() — the overshoot never
being charged back, and the force-isolate path on lower-zone reclaim.
The concerns below are the ones surfaced from reading the code, then
confirmed by constructing the situation deliberately.

The problem only appears when reclaim targets a lower zone while the
LRU holds folios from a higher zone:

  - reclaim_idx points at DMA32/DMA (the triggering allocation asked
    for a lower zone, e.g. __GFP_DMA32), and
  - the LRU holds folios from a higher zone (Normal/Movable).

Normal userspace allocations go to the highest zone, so reclaim_idx
never points below it and the zone-skip branch is never taken.  It
needs a real lower-zone allocator to drain that zone below watermark;
that is also why it went unnoticed upstream — the 1c7b17cf hard-lockup
fix that introduced the force-isolate path was found only on a ~1 TB
box running DMA32 module allocations.

## Impact

Reproduced on x86 QEMU, 7.2-rc6, with a kernel module doing
__GFP_DMA32 allocations to drain DMA32 below watermark (LRU folios
sit in Movable):

  - A single isolate_lru_folios() call scanned 32794 pages and took
    25 (32769 skipped): 99.9% wasted on ineligible folios.
  - Across one run, isolate was called 339 times, 140-165 of which
    isolated nothing (taken=0, pure empty scans).
  - shrink_lruvec() charges only the 32-page quota per round and
    never refunds the overshoot, so the same skipped folios are
    rescanned round after round.  vmstat on a memcg OOM path:
    Δpgscan_direct / Δpgsteal_direct = 2836624 / 112719 = 25.2x
    (isolate -> shrink_folio_list returns the folio -> isolate again).

Once max_nr_skipped hits SWAP_CLUSTER_MAX_SKIPPED, the current code
force-isolates the remaining ineligible folios.  On the inactive LRU
they reach shrink_folio_list() and get reclaimed though they can
never satisfy the allocation, inflating nr_reclaimed and resetting
no_progress_loops in should_reclaim_retry(), delaying the OOM.

## A/B results (same .config, md5-identical)

| metric                  | baseline   | patched | note |
|-------------------------|------------|---------|------|
| empty scans (taken=0)   | 140-165    | 0       | stable across runs |
| isolate calls           | 339        | 7       | 48x fewer |
| total pages scanned     | 343711     | 3140    | 109x fewer |
| single-call max scan    | 32794      | 3104    | |

Empty-scan 0 is the stable evidence (holds every run).  The
Δpgscan/Δpgsteal ratio is volatile (baseline 56-26675x, patched
65-1324x, ranges overlap) and is not relied on alone.

## Caveats and reproduction

  - Triggering needs a lower-zone-pressured box.  To confirm the code
    analysis, the situation was constructed on a small x86 QEMU VM
    (1500M, CONFIG_LRU_GEN=n) with kernel cmdline
    `movable_zone=DMA32 kernelcore=256M` (DMA32 small, Normal empty,
    Movable large), then a kernel module doing
    `alloc_page(GFP_DMA32)` drains DMA32 below watermark.  LRU folios
    sit in Movable and are zone-skipped while reclaiming for DMA32.
    Observed via the `mm_vmscan_lru_isolate` tracepoint and
    /proc/vmstat (pgscan_direct, pgsteal_direct).

> Thanks.

Hi, maintainers
  Kindly ping.

On 周三 2026-8-12 10:29, mawupeng wrote:
> 
> 
> On 周三 2026-8-12 09:42, Andrew Morton wrote:
>> On Tue, 11 Aug 2026 15:08:47 +0800 Wupeng Ma <mawupeng1@huawei.com> wrote:
>>
>>> shrink_lruvec() drives reclaim in SWAP_CLUSTER_MAX (32) chunks, but
>>> isolate_lru_folios() may scan far more than that per call on a single
>>> LRU.  The excess is never charged back, so shrink_lruvec() keeps
>>> rescanning the same folios round after round.  When reclaim targets a
>>> lower zone, the same scanner also keeps walking zone-ineligible folios
>>> that can never satisfy the allocation, inflating nr_reclaimed into a
>>> false progress that delays the OOM.
>>>
>>> This series fixes both:
>>>
>>>   [1/2] Charge the isolate overshoot against the scan quota so the
>>>         next round skips already-scanned folios.
>>>   [2/2] Stop scanning once too many zone-ineligible folios have been
>>>         skipped, instead of force-isolating them.
>>
>> This all sounds fairly bad.  Is there some report?  Real-world test
>> results?  Laboratory test results?  Something to help us understand the
>> impact of the issues and the situations in which they occur.
>>
> 
> Sorry — this background belongs in the cover letter and I left it
> out.  Adding it here.
> 
> ## Situation
> 
> We observed slow, unexpected OOM behavior during extreme stress testing,
> and while digging into the reclaim code during that analysis we spotted
> these latent risks in isolate_lru_folios() — the overshoot never
> being charged back, and the force-isolate path on lower-zone reclaim.
> The concerns below are the ones surfaced from reading the code, then
> confirmed by constructing the situation deliberately.
> 
> The problem only appears when reclaim targets a lower zone while the
> LRU holds folios from a higher zone:
> 
>   - reclaim_idx points at DMA32/DMA (the triggering allocation asked
>     for a lower zone, e.g. __GFP_DMA32), and
>   - the LRU holds folios from a higher zone (Normal/Movable).
> 
> Normal userspace allocations go to the highest zone, so reclaim_idx
> never points below it and the zone-skip branch is never taken.  It
> needs a real lower-zone allocator to drain that zone below watermark;
> that is also why it went unnoticed upstream — the 1c7b17cf hard-lockup
> fix that introduced the force-isolate path was found only on a ~1 TB
> box running DMA32 module allocations.
> 
> ## Impact
> 
> Reproduced on x86 QEMU, 7.2-rc6, with a kernel module doing
> __GFP_DMA32 allocations to drain DMA32 below watermark (LRU folios
> sit in Movable):
> 
>   - A single isolate_lru_folios() call scanned 32794 pages and took
>     25 (32769 skipped): 99.9% wasted on ineligible folios.
>   - Across one run, isolate was called 339 times, 140-165 of which
>     isolated nothing (taken=0, pure empty scans).
>   - shrink_lruvec() charges only the 32-page quota per round and
>     never refunds the overshoot, so the same skipped folios are
>     rescanned round after round.  vmstat on a memcg OOM path:
>     Δpgscan_direct / Δpgsteal_direct = 2836624 / 112719 = 25.2x
>     (isolate -> shrink_folio_list returns the folio -> isolate again).
> 
> Once max_nr_skipped hits SWAP_CLUSTER_MAX_SKIPPED, the current code
> force-isolates the remaining ineligible folios.  On the inactive LRU
> they reach shrink_folio_list() and get reclaimed though they can
> never satisfy the allocation, inflating nr_reclaimed and resetting
> no_progress_loops in should_reclaim_retry(), delaying the OOM.
> 
> ## A/B results (same .config, md5-identical)
> 
> | metric                  | baseline   | patched | note |
> |-------------------------|------------|---------|------|
> | empty scans (taken=0)   | 140-165    | 0       | stable across runs |
> | isolate calls           | 339        | 7       | 48x fewer |
> | total pages scanned     | 343711     | 3140    | 109x fewer |
> | single-call max scan    | 32794      | 3104    | |
> 
> Empty-scan 0 is the stable evidence (holds every run).  The
> Δpgscan/Δpgsteal ratio is volatile (baseline 56-26675x, patched
> 65-1324x, ranges overlap) and is not relied on alone.
> 
> ## Caveats and reproduction
> 
>   - Triggering needs a lower-zone-pressured box.  To confirm the code
>     analysis, the situation was constructed on a small x86 QEMU VM
>     (1500M, CONFIG_LRU_GEN=n) with kernel cmdline
>     `movable_zone=DMA32 kernelcore=256M` (DMA32 small, Normal empty,
>     Movable large), then a kernel module doing
>     `alloc_page(GFP_DMA32)` drains DMA32 below watermark.  LRU folios
>     sit in Movable and are zone-skipped while reclaiming for DMA32.
>     Observed via the `mm_vmscan_lru_isolate` tracepoint and
>     /proc/vmstat (pgscan_direct, pgsteal_direct).
> 
>> Thanks.
> 

On Tue, 25 Aug 2026 08:59:54 +0800 mawupeng <mawupeng1@huawei.com> wrote:

> Hi, maintainers
>   Kindly ping.

It's been quite a while.  It would be helpful to refresh/retest/resend,
please.

While doing this, please update the changelogging to help reviewers
understand the end-user impact of the issue.

Please also give some though to whether we should backport these fixes
into earlier kernels, with a cc:stable.  Given the unusual
circumstances which are required to hit this, I'm thinking "no", but
feel free to disagree!

On 周日 2026-8-30 11:18, Andrew Morton wrote:
> On Tue, 25 Aug 2026 08:59:54 +0800 mawupeng <mawupeng1@huawei.com> wrote:
> 
>> Hi, maintainers
>>   Kindly ping.
> 
> It's been quite a while.  It would be helpful to refresh/retest/resend,
> please.

Thanks, Andrew. I’ll retest it with the background (situation, end-user impact,
reproduction, A/B) folded into the cover letter and re-tested on the current -rc.

> 
> While doing this, please update the changelogging to help reviewers
> understand the end-user impact of the issue.
> 
> Please also give some though to whether we should backport these fixes
> into earlier kernels, with a cc:stable.  Given the unusual
> circumstances which are required to hit this, I'm thinking "no", but
> feel free to disagree!

On cc:stable — I agree it doesn’t seem warranted. The conditions needed
to trigger this are unusual enough that I don’t think a backport is
justified, so I’d lean toward not adding Cc:stable. Happy to revisit if
I’m missing something.

>