Re: [RFC PATCH 0/6] mm/damon: Add mTHP-aware collapse/split with ARM SPE feedback

Re: [RFC PATCH 0/6] mm/damon: Add mTHP-aware collapse/split with ARM SPE feedback

[SeongJae Park]

On Thu, 18 Jun 2026 21:13:07 +0800 wang lian <lianux.mm@gmail.com> wrote:

> 
> 
> > On Jun 18, 2026, at 19:03, Gutierrez Asier <gutierrez.asier@huawei-partners.com> wrote:
> > 
> > Hi Wang,
> > 
> > On 6/18/2026 12:48 PM, Wang Lian wrote:
> >> Received an off-list report that DAMON significantly overestimates
> >> hot memory in KVM/QEMU deployments with THP-backed tmpfs guest memory
> >> running Oracle workloads.
> >> 
> >> The root cause is structural: a PMD entry covers 512 4KB subpages with
> >> a single Access Flag (AF) bit. When any one subpage is accessed, the entire
> >> 2MB region appears "hot" to DAMON. On ARM64, this is compounded by the
> >> hardware AF mechanism -- the AF is only set on a TLB miss. Consequently, when the
> >> working set fits entirely within the L2 TLB (e.g., a 16MB working set with 2MB THP
> >> running on a Kunpeng 920's 2048-entry L2 TLB), DAMON becomes completely blind to
> >> subsequent accesses. x86 is not subject to this specific blindness under similar
> >> conditions.
> > 
> > Have you tried setting the minimum region size to 2MB?
> > 
> >> We reproduced this memory inflation on a Kunpeng 920 platform using a synthetic
> >> workload (8GB mmap with a 0.2% sparse hotspot, i.e. 16MB actually hot):
> >> THP=always causes DAMON to report the entire 8GB as hot, while THP=never
> >> reports only a few hundred MB -- a 512x overestimate relative to the actual
> >> 16MB hotspot under THP, and a ~33x gap between the two THP modes. ARM SPE hardware profiling
> >> independently confirms this asymmetry: out of 2,005 THPs sampled system-wide
> >> over 10 seconds, 97% had fewer than 10% of their 4KB subpages actually accessed.
> > 
> > THP always will just collapse the entire PID into huge pages anyway. This
> > is outside DAMON's control.
> > 
> > Have you tried setting THP to never and running DAMON with DAMON_COLLAPSE
> > action?
> > 
> >> To mitigate this, this series extends the existing DAMOS_COLLAPSE action to be
> >> mTHP-aware via a new target_order field, and introduces a new
> >> DAMOS_MTHP_SPLIT action. This enables DAMON to proactively split PMD THPs
> >> into smaller mTHPs when most subpages are probed as cold, and collapse them
> >> back when beneficial. To resolve the sub-PMD monitoring blindness, the split
> >> path can incorporate fine-grained hardware feedback from ARM SPE.
> >> The hardware feedback loop (damon_spe_folio_heatmap) implements a two-pass
> >> signal filter: it first identifies the peak chunk access count, and then marks
> >> sub-chunks with >= 1/10 of the peak count as hot, effectively filtering out
> >> SPE sampling noise. A configurable hot_threshold (default 30%) controls the
> >> split decision: only folios with a hot fraction below this threshold are
> >> eligible for splitting. When no SPE data is available, the infrastructure
> >> gracefully falls back to explicit PTE-level scanning via folio_walk.
> >> 
> >> Currently, SPE data is fed from userspace via debugfs (e.g., perf script piped
> >> through a histogram builder into /sys/kernel/debug/damon/spe_feed).
> >> 
> >> Collapse path (patches 1-3):
> >>  DAMON scheme action=COLLAPSE, target_order=N
> >>  -> damos_va_collapse() -> damon_collapse_folio_range()
> >>  -> collapse_huge_page()
> >> 
> >> Split path (patches 4-5):
> >>  DAMON scheme action=MTHP_SPLIT, target_order=N, hot_threshold=M
> >>  -> damos_va_mthp_split() -> damon_spe_hot_fraction()
> >>  -> split_folio_to_order()
> >> 
> >> SPE feedback infrastructure (patch 6):
> >>  perf script -> spe_hist -> debugfs spe_feed
> >>  -> per-folio rbtree {THP-aligned PFN -> access_count[512]}
> >>  -> damon_spe_folio_heatmap() -> hot_bitmap -> split decision
> >> 
> >> The userspace helper tools (including the spe_hist histogram builder and
> >> validation scripts) are archived at:
> >>  https://github.com/lianux-mm/damon_spe
> >> 
> >> Testing was performed on a Kunpeng 920 system (256 cores, 249GB RAM, base kernel
> >> 7.1.0-rc5+):
> >> 
> >>  T1 ARM64 blind spot: A 16MB THP workload (where 8 PMDs fit entirely within the
> >>     L2 TLB) resulted in DAMON detecting 0 regions. Conversely, using 512MB
> >>     with 4KB base pages, or a 16GB THP layout (exceeding L2 TLB reach), allowed
> >>     DAMON to function normally.
> >> 
> >>  T2 THP inflation: With an 8GB mmap and 16MB actually hot (0.2%),
> >>     THP=always: DAMON reported 8GB hot (512x vs ground truth);
> >>     THP=never: ~245MB (15x vs ground truth).  The THP-induced gap
> >>     between the two modes was ~33x.
> >> 
> >>  T3 RocksDB: Fragmented malloc allocation prevented THP formation, and DAMON
> >>     behaved normally. We could not reproduce THP inflation with RocksDB.
> >>     The workloads fundamentally vulnerable to this structural issue remain KVM
> >>     guests, JVM large heaps, and PostgreSQL shared_buffers.
> >> 
> >>  T4 min=0 deadlock break: A 256MB THP induced the DAMON blind spot.
> >>     Triggering an unconditional mthp_split (via nr_accesses/min=0) successfully
> >>     shattered the space into 16384x16KB folios, allowing DAMON to fully recover.
> >> 
> >>  T5 ARM SPE histogram: Out of 2005 sampled THPs, 97% exhibited <10% hot subpages.
> >>     A typical trace showed PFN 0x820db800 accumulated 39,794 hardware accesses
> >>     concentrated across only 3 out of 512 subpages.
> > The SPE stuff fits SeongJae's goals for DAMON-X, I think. Maybe this is something
> > we should keep in the user space and let the kernel provide only the API to add
> > different metrics, including PMU and SPE.
> 
> Hi Asier,
> 
> Thanks for your prompt and constructive reply. I really appreciate your 
> detailed analysis of the mTHP and SPE interaction.

Indeed, very helpful comments.  Thank you Asier!

> 
> Your point regarding the design boundary—whether this fits better in 
> user space or aligned with DAMON-X—is highly valuable. 

Actually Asier is saying about the perf event-based monitoring extension [1].
DAMON-X [2] is another project.

> 
> Since SeongJae (SJ) will look into this thread tomorrow, let us sync up 
> then. I look forward to cooperating with both of you to refine this 
> design and find the best architectural fit for the subsystem.

As I also replied, I'd also prefer this to be aligned with the perf event-based
extension roadmap.

[1] https://lore.kernel.org/all/20260525225208.1179-1-sj@kernel.org/
[2] https://lwn.net/Articles/1071256/


Thanks,
SJ

[...]

[RFC PATCH 0/6] mm/damon: Add mTHP-aware collapse/split with ARM SPE feedback

[Wang Lian]

Received an off-list report that DAMON significantly overestimates
hot memory in KVM/QEMU deployments with THP-backed tmpfs guest memory
running Oracle workloads.

The root cause is structural: a PMD entry covers 512 4KB subpages with
a single Access Flag (AF) bit. When any one subpage is accessed, the entire
2MB region appears "hot" to DAMON. On ARM64, this is compounded by the
hardware AF mechanism -- the AF is only set on a TLB miss. Consequently, when the
working set fits entirely within the L2 TLB (e.g., a 16MB working set with 2MB THP
running on a Kunpeng 920's 2048-entry L2 TLB), DAMON becomes completely blind to
subsequent accesses. x86 is not subject to this specific blindness under similar
conditions.

We reproduced this memory inflation on a Kunpeng 920 platform using a synthetic
workload (8GB mmap with a 0.2% sparse hotspot, i.e. 16MB actually hot):
THP=always causes DAMON to report the entire 8GB as hot, while THP=never
reports only a few hundred MB -- a 512x overestimate relative to the actual
16MB hotspot under THP, and a ~33x gap between the two THP modes. ARM SPE hardware profiling
independently confirms this asymmetry: out of 2,005 THPs sampled system-wide
over 10 seconds, 97% had fewer than 10% of their 4KB subpages actually accessed.

To mitigate this, this series extends the existing DAMOS_COLLAPSE action to be
mTHP-aware via a new target_order field, and introduces a new
DAMOS_MTHP_SPLIT action. This enables DAMON to proactively split PMD THPs
into smaller mTHPs when most subpages are probed as cold, and collapse them
back when beneficial. To resolve the sub-PMD monitoring blindness, the split
path can incorporate fine-grained hardware feedback from ARM SPE.

The hardware feedback loop (damon_spe_folio_heatmap) implements a two-pass
signal filter: it first identifies the peak chunk access count, and then marks
sub-chunks with >= 1/10 of the peak count as hot, effectively filtering out
SPE sampling noise. A configurable hot_threshold (default 30%) controls the
split decision: only folios with a hot fraction below this threshold are
eligible for splitting. When no SPE data is available, the infrastructure
gracefully falls back to explicit PTE-level scanning via folio_walk.

Currently, SPE data is fed from userspace via debugfs (e.g., perf script piped
through a histogram builder into /sys/kernel/debug/damon/spe_feed).

Collapse path (patches 1-3):
  DAMON scheme action=COLLAPSE, target_order=N
  -> damos_va_collapse() -> damon_collapse_folio_range()
  -> collapse_huge_page()

Split path (patches 4-5):
  DAMON scheme action=MTHP_SPLIT, target_order=N, hot_threshold=M
  -> damos_va_mthp_split() -> damon_spe_hot_fraction()
  -> split_folio_to_order()

SPE feedback infrastructure (patch 6):
  perf script -> spe_hist -> debugfs spe_feed
  -> per-folio rbtree {THP-aligned PFN -> access_count[512]}
  -> damon_spe_folio_heatmap() -> hot_bitmap -> split decision

The userspace helper tools (including the spe_hist histogram builder and
validation scripts) are archived at:
  https://github.com/lianux-mm/damon_spe

Testing was performed on a Kunpeng 920 system (256 cores, 249GB RAM, base kernel
7.1.0-rc5+):

  T1 ARM64 blind spot: A 16MB THP workload (where 8 PMDs fit entirely within the
     L2 TLB) resulted in DAMON detecting 0 regions. Conversely, using 512MB
     with 4KB base pages, or a 16GB THP layout (exceeding L2 TLB reach), allowed
     DAMON to function normally.

  T2 THP inflation: With an 8GB mmap and 16MB actually hot (0.2%),
     THP=always: DAMON reported 8GB hot (512x vs ground truth);
     THP=never: ~245MB (15x vs ground truth).  The THP-induced gap
     between the two modes was ~33x.

  T3 RocksDB: Fragmented malloc allocation prevented THP formation, and DAMON
     behaved normally. We could not reproduce THP inflation with RocksDB.
     The workloads fundamentally vulnerable to this structural issue remain KVM
     guests, JVM large heaps, and PostgreSQL shared_buffers.

  T4 min=0 deadlock break: A 256MB THP induced the DAMON blind spot.
     Triggering an unconditional mthp_split (via nr_accesses/min=0) successfully
     shattered the space into 16384x16KB folios, allowing DAMON to fully recover.

  T5 ARM SPE histogram: Out of 2005 sampled THPs, 97% exhibited <10% hot subpages.
     A typical trace showed PFN 0x820db800 accumulated 39,794 hardware accesses
     concentrated across only 3 out of 512 subpages.

  End-to-end: Verified hot/cold discrimination. The SPE feed preserved a 90%
     hot THP intact, while successfully splitting a 25% cold THP into 128x16KB folios.

Known limitations:
- The full KVM + Oracle production chain has not yet been benchmarked end-to-end.
  While individual component verification is complete, full integration testing
  is planned in collaboration with Sangfor.
- khugepaged may aggressively re-collapse the mTHPs that DAMON splits. A
  coordination/back-off mechanism is required to avoid ping-pong effects.
- SPE data is currently funneled via a userspace daemon and debugfs. Direct
  kernel-side perf_event sampling integration is planned as a follow-up.
- The rbtree entry TTL (30s) and signal threshold (1/10 of peak) are empirical
  defaults subject to further tuning.
- The ARM64 DAMON blind spot (WSS < L2 TLB reach) is a pre-existing hardware-MMU
  characteristic, not introduced by this series. Setting nr_accesses/min=0
  serves as an effective workaround for the split path.

Reported-by: Chaobing Dai <daichaobing@sangfor.com.cn>
Cc: SeongJae Park <sj@kernel.org>
Cc: Andrew Morton <akpm@linux-foundation.org>
Cc: Nico Pache <npache@redhat.com>
Cc: Asier Gutierrez <gutierrez.asier@huawei-partners.com>
Cc: linux-mm@kvack.org
Cc: linux-kernel@vger.kernel.org
Signed-off-by: Wang Lian <lianux.mm@gmail.com>

Wang Lian (6):
  mm/damon: add target_order field for DAMOS_COLLAPSE
  mm/khugepaged: add damon_collapse_folio_range() for external callers
  mm/damon/vaddr: implement mTHP-aware DAMOS_COLLAPSE handler
  mm/damon: introduce DAMOS_MTHP_SPLIT action and hot_threshold
  mm/damon/vaddr: implement DAMOS_MTHP_SPLIT handler
  mm/damon: add SPE feedback for sub-THP split decisions

 include/linux/damon.h      |  18 ++
 include/linux/khugepaged.h |   3 +
 mm/damon/Kconfig           |  12 +
 mm/damon/Makefile          |   1 +
 mm/damon/core.c            |   3 +
 mm/damon/spe.c             | 505 +++++++++++++++++++++++++++++++++++++
 mm/damon/spe.h             |  62 +++++
 mm/damon/sysfs-schemes.c   |  96 +++++++
 mm/damon/vaddr.c           | 118 +++++++++
 mm/khugepaged.c            |  39 +++
 10 files changed, 857 insertions(+)
 create mode 100644 mm/damon/spe.c
 create mode 100644 mm/damon/spe.h

--
2.50.1 (Apple Git-155)

Re: [RFC PATCH 0/6] mm/damon: Add mTHP-aware collapse/split with ARM SPE feedback

[Gutierrez Asier]

Hi Wang,

On 6/18/2026 12:48 PM, Wang Lian wrote:
> Received an off-list report that DAMON significantly overestimates
> hot memory in KVM/QEMU deployments with THP-backed tmpfs guest memory
> running Oracle workloads.
> 
> The root cause is structural: a PMD entry covers 512 4KB subpages with
> a single Access Flag (AF) bit. When any one subpage is accessed, the entire
> 2MB region appears "hot" to DAMON. On ARM64, this is compounded by the
> hardware AF mechanism -- the AF is only set on a TLB miss. Consequently, when the
> working set fits entirely within the L2 TLB (e.g., a 16MB working set with 2MB THP
> running on a Kunpeng 920's 2048-entry L2 TLB), DAMON becomes completely blind to
> subsequent accesses. x86 is not subject to this specific blindness under similar
> conditions.

Have you tried setting the minimum region size to 2MB?

> We reproduced this memory inflation on a Kunpeng 920 platform using a synthetic
> workload (8GB mmap with a 0.2% sparse hotspot, i.e. 16MB actually hot):
> THP=always causes DAMON to report the entire 8GB as hot, while THP=never
> reports only a few hundred MB -- a 512x overestimate relative to the actual
> 16MB hotspot under THP, and a ~33x gap between the two THP modes. ARM SPE hardware profiling
> independently confirms this asymmetry: out of 2,005 THPs sampled system-wide
> over 10 seconds, 97% had fewer than 10% of their 4KB subpages actually accessed.

THP always will just collapse the entire PID into huge pages anyway. This
is outside DAMON's control.

Have you tried setting THP to never and running DAMON with DAMON_COLLAPSE
action?

> To mitigate this, this series extends the existing DAMOS_COLLAPSE action to be
> mTHP-aware via a new target_order field, and introduces a new
> DAMOS_MTHP_SPLIT action. This enables DAMON to proactively split PMD THPs
> into smaller mTHPs when most subpages are probed as cold, and collapse them
> back when beneficial. To resolve the sub-PMD monitoring blindness, the split
> path can incorporate fine-grained hardware feedback from ARM SPE.
> The hardware feedback loop (damon_spe_folio_heatmap) implements a two-pass
> signal filter: it first identifies the peak chunk access count, and then marks
> sub-chunks with >= 1/10 of the peak count as hot, effectively filtering out
> SPE sampling noise. A configurable hot_threshold (default 30%) controls the
> split decision: only folios with a hot fraction below this threshold are
> eligible for splitting. When no SPE data is available, the infrastructure
> gracefully falls back to explicit PTE-level scanning via folio_walk.
> 
> Currently, SPE data is fed from userspace via debugfs (e.g., perf script piped
> through a histogram builder into /sys/kernel/debug/damon/spe_feed).
> 
> Collapse path (patches 1-3):
>   DAMON scheme action=COLLAPSE, target_order=N
>   -> damos_va_collapse() -> damon_collapse_folio_range()
>   -> collapse_huge_page()
> 
> Split path (patches 4-5):
>   DAMON scheme action=MTHP_SPLIT, target_order=N, hot_threshold=M
>   -> damos_va_mthp_split() -> damon_spe_hot_fraction()
>   -> split_folio_to_order()
> 
> SPE feedback infrastructure (patch 6):
>   perf script -> spe_hist -> debugfs spe_feed
>   -> per-folio rbtree {THP-aligned PFN -> access_count[512]}
>   -> damon_spe_folio_heatmap() -> hot_bitmap -> split decision
> 
> The userspace helper tools (including the spe_hist histogram builder and
> validation scripts) are archived at:
>   https://github.com/lianux-mm/damon_spe
> 
> Testing was performed on a Kunpeng 920 system (256 cores, 249GB RAM, base kernel
> 7.1.0-rc5+):
> 
>   T1 ARM64 blind spot: A 16MB THP workload (where 8 PMDs fit entirely within the
>      L2 TLB) resulted in DAMON detecting 0 regions. Conversely, using 512MB
>      with 4KB base pages, or a 16GB THP layout (exceeding L2 TLB reach), allowed
>      DAMON to function normally.
> 
>   T2 THP inflation: With an 8GB mmap and 16MB actually hot (0.2%),
>      THP=always: DAMON reported 8GB hot (512x vs ground truth);
>      THP=never: ~245MB (15x vs ground truth).  The THP-induced gap
>      between the two modes was ~33x.
> 
>   T3 RocksDB: Fragmented malloc allocation prevented THP formation, and DAMON
>      behaved normally. We could not reproduce THP inflation with RocksDB.
>      The workloads fundamentally vulnerable to this structural issue remain KVM
>      guests, JVM large heaps, and PostgreSQL shared_buffers.
> 
>   T4 min=0 deadlock break: A 256MB THP induced the DAMON blind spot.
>      Triggering an unconditional mthp_split (via nr_accesses/min=0) successfully
>      shattered the space into 16384x16KB folios, allowing DAMON to fully recover.
> 
>   T5 ARM SPE histogram: Out of 2005 sampled THPs, 97% exhibited <10% hot subpages.
>      A typical trace showed PFN 0x820db800 accumulated 39,794 hardware accesses
>      concentrated across only 3 out of 512 subpages.
The SPE stuff fits SeongJae's goals for DAMON-X, I think. Maybe this is something
we should keep in the user space and let the kernel provide only the API to add
different metrics, including PMU and SPE.
>   End-to-end: Verified hot/cold discrimination. The SPE feed preserved a 90%
>      hot THP intact, while successfully splitting a 25% cold THP into 128x16KB folios.
> 
> Known limitations:
> - The full KVM + Oracle production chain has not yet been benchmarked end-to-end.
>   While individual component verification is complete, full integration testing
>   is planned in collaboration with Sangfor.
> - khugepaged may aggressively re-collapse the mTHPs that DAMON splits. A
>   coordination/back-off mechanism is required to avoid ping-pong effects.
> - SPE data is currently funneled via a userspace daemon and debugfs. Direct
>   kernel-side perf_event sampling integration is planned as a follow-up.
> - The rbtree entry TTL (30s) and signal threshold (1/10 of peak) are empirical
>   defaults subject to further tuning.
> - The ARM64 DAMON blind spot (WSS < L2 TLB reach) is a pre-existing hardware-MMU
>   characteristic, not introduced by this series. Setting nr_accesses/min=0
>   serves as an effective workaround for the split path.
> 
> Reported-by: Chaobing Dai <daichaobing@sangfor.com.cn>
> Cc: SeongJae Park <sj@kernel.org>
> Cc: Andrew Morton <akpm@linux-foundation.org>
> Cc: Nico Pache <npache@redhat.com>
> Cc: Asier Gutierrez <gutierrez.asier@huawei-partners.com>
> Cc: linux-mm@kvack.org
> Cc: linux-kernel@vger.kernel.org
> Signed-off-by: Wang Lian <lianux.mm@gmail.com>
> 
> Wang Lian (6):
>   mm/damon: add target_order field for DAMOS_COLLAPSE
>   mm/khugepaged: add damon_collapse_folio_range() for external callers
>   mm/damon/vaddr: implement mTHP-aware DAMOS_COLLAPSE handler
>   mm/damon: introduce DAMOS_MTHP_SPLIT action and hot_threshold
>   mm/damon/vaddr: implement DAMOS_MTHP_SPLIT handler
>   mm/damon: add SPE feedback for sub-THP split decisions
> 
>  include/linux/damon.h      |  18 ++
>  include/linux/khugepaged.h |   3 +
>  mm/damon/Kconfig           |  12 +
>  mm/damon/Makefile          |   1 +
>  mm/damon/core.c            |   3 +
>  mm/damon/spe.c             | 505 +++++++++++++++++++++++++++++++++++++
>  mm/damon/spe.h             |  62 +++++
>  mm/damon/sysfs-schemes.c   |  96 +++++++
>  mm/damon/vaddr.c           | 118 +++++++++
>  mm/khugepaged.c            |  39 +++
>  10 files changed, 857 insertions(+)
>  create mode 100644 mm/damon/spe.c
>  create mode 100644 mm/damon/spe.h
> 
> --
> 2.50.1 (Apple Git-155)
> 

-- 
Asier Gutierrez
Huawei

Re: [RFC PATCH 0/6] mm/damon: Add mTHP-aware collapse/split with ARM SPE feedback

[SeongJae Park]

Hello Lian,

On Thu, 18 Jun 2026 17:48:32 +0800 Wang Lian <lianux.mm@gmail.com> wrote:

> Received an off-list report that DAMON significantly overestimates
> hot memory in KVM/QEMU deployments with THP-backed tmpfs guest memory
> running Oracle workloads.
> 
> The root cause is structural: a PMD entry covers 512 4KB subpages with
> a single Access Flag (AF) bit. When any one subpage is accessed, the entire
> 2MB region appears "hot" to DAMON. On ARM64,

This makes sense to me.  I also agree this could caused the reported problem.
And this is a known limitation of DAMON.  My suggestion for straightforward
workaround of this problem is, using 'age' information of DAMON for better
identification of the hot memory.

That is, I don't expect real hot data in real production systems will evenly
scattered.  Even if they are, I don't expect they will all evenly frequently
accessed.  Only a few of those would be accessed frequently for long.  Even if
that is, there would be data that frequently for longer.  You could show the
distriibution of the pattern and find X % of hottest memory as hot.

We invented idle time percentiles [1] for a similar purpose, though it is more
focusing on finding cold memory.

I understand this patch series is trying to make more fundamental and better
solution on hardware that can do better.  Makes sense to me.

> this is compounded by the
> hardware AF mechanism -- the AF is only set on a TLB miss. Consequently, when the
> working set fits entirely within the L2 TLB (e.g., a 16MB working set with 2MB THP
> running on a Kunpeng 920's 2048-entry L2 TLB), DAMON becomes completely blind to
> subsequent accesses.

This makes sense to me.  However, I don't get how this is contributing to the
problem.  Could you please elaborate?

> x86 is not subject to this specific blindness under similar
> conditions.

To my understanding on x86, same issue exists.  If TLB hits, Aceessed bit is
not set, and DAMON shows it as unaccessed.  Am I missing something?

> 
> We reproduced this memory inflation on a Kunpeng 920 platform using a synthetic
> workload (8GB mmap with a 0.2% sparse hotspot, i.e. 16MB actually hot):
> THP=always causes DAMON to report the entire 8GB as hot, while THP=never
> reports only a few hundred MB -- a 512x overestimate relative to the actual
> 16MB hotspot under THP, and a ~33x gap between the two THP modes. ARM SPE hardware profiling
> independently confirms this asymmetry: out of 2,005 THPs sampled system-wide
> over 10 seconds, 97% had fewer than 10% of their 4KB subpages actually accessed.

I don't think the real world production systems to have this very artificial
access pattern.  I believe (or, hope) use of 'age' can work around the issue in
a reasonable level for many cases.  I understand this setup is only for PoC,
and I think this is well designed test for the purpose.  Thank you for sharing
this.

> 
> To mitigate this, this series extends the existing DAMOS_COLLAPSE action to be
> mTHP-aware via a new target_order field,

Makes sensee, and sounds nice.  Definitely no one size fits all!

> and introduces a new
> DAMOS_MTHP_SPLIT action. This enables DAMON to proactively split PMD THPs
> into smaller mTHPs

Nice!  Asier was planning to do similar work in future.  I think you could
collaborate to reduce unnecessary duplicates!

I'd suggest making the name simpler and consistent to DAMOS_COLLAPSE, though.
Say, DAMOS_SPLIT ?

> when most subpages are probed as cold, and collapse them
> back when beneficial. To resolve the sub-PMD monitoring blindness, the split
> path can incorporate fine-grained hardware feedback from ARM SPE.
> 
> The hardware feedback loop (damon_spe_folio_heatmap) implements a two-pass
> signal filter: it first identifies the peak chunk access count, and then marks
> sub-chunks with >= 1/10 of the peak count as hot, effectively filtering out
> SPE sampling noise. A configurable hot_threshold (default 30%) controls the
> split decision: only folios with a hot fraction below this threshold are
> eligible for splitting. When no SPE data is available, the infrastructure
> gracefully falls back to explicit PTE-level scanning via folio_walk.
> 
> Currently, SPE data is fed from userspace via debugfs (e.g., perf script piped
> through a histogram builder into /sys/kernel/debug/damon/spe_feed).

So you implemented a debugfs interface?  That must be a nice approach for PoC.
But it may be difficult to be upstreamed as is.

You could build a control plane that decides the exact address ranges to split,
and directly feed it to DAMOS using DAMOS address filter.  max_nr_snapshots can
also be useful for making such kind of user space controls more deterministic.

For simpler user-space control, utilizing user_input DAMOS quota goal [2]
should also be another option.

We are also planning [3] to extend DAMON for perf events.  On top of it, we
might be able to extend it further to utilize ARM SPE by DAMON itself, and do
all this without the user space help but only DAMOS.

Baseed on below 'limitations' section, I understand this is only for PoC at the
moment, and you plan to explore the perf event based approach.  I'd also
recommend that.

> 
> Collapse path (patches 1-3):
>   DAMON scheme action=COLLAPSE, target_order=N
>   -> damos_va_collapse() -> damon_collapse_folio_range()
>   -> collapse_huge_page()
> 
> Split path (patches 4-5):
>   DAMON scheme action=MTHP_SPLIT, target_order=N, hot_threshold=M
>   -> damos_va_mthp_split() -> damon_spe_hot_fraction()
>   -> split_folio_to_order()
> 
> SPE feedback infrastructure (patch 6):
>   perf script -> spe_hist -> debugfs spe_feed
>   -> per-folio rbtree {THP-aligned PFN -> access_count[512]}
>   -> damon_spe_folio_heatmap() -> hot_bitmap -> split decision
> 
> The userspace helper tools (including the spe_hist histogram builder and
> validation scripts) are archived at:
>   https://github.com/lianux-mm/damon_spe

Thank you for making all the grateful code open!

> 
> Testing was performed on a Kunpeng 920 system (256 cores, 249GB RAM, base kernel
> 7.1.0-rc5+):
> 
>   T1 ARM64 blind spot: A 16MB THP workload (where 8 PMDs fit entirely within the
>      L2 TLB) resulted in DAMON detecting 0 regions. Conversely, using 512MB
>      with 4KB base pages, or a 16GB THP layout (exceeding L2 TLB reach), allowed
>      DAMON to function normally.
> 
>   T2 THP inflation: With an 8GB mmap and 16MB actually hot (0.2%),
>      THP=always: DAMON reported 8GB hot (512x vs ground truth);
>      THP=never: ~245MB (15x vs ground truth).  The THP-induced gap
>      between the two modes was ~33x.
> 
>   T3 RocksDB: Fragmented malloc allocation prevented THP formation, and DAMON
>      behaved normally. We could not reproduce THP inflation with RocksDB.
>      The workloads fundamentally vulnerable to this structural issue remain KVM
>      guests, JVM large heaps, and PostgreSQL shared_buffers.
> 
>   T4 min=0 deadlock break: A 256MB THP induced the DAMON blind spot.
>      Triggering an unconditional mthp_split (via nr_accesses/min=0) successfully
>      shattered the space into 16384x16KB folios, allowing DAMON to fully recover.
> 
>   T5 ARM SPE histogram: Out of 2005 sampled THPs, 97% exhibited <10% hot subpages.
>      A typical trace showed PFN 0x820db800 accumulated 39,794 hardware accesses
>      concentrated across only 3 out of 512 subpages.
> 
>   End-to-end: Verified hot/cold discrimination. The SPE feed preserved a 90%
>      hot THP intact, while successfully splitting a 25% cold THP into 128x16KB folios.
> 
> Known limitations:
> - The full KVM + Oracle production chain has not yet been benchmarked end-to-end.
>   While individual component verification is complete, full integration testing
>   is planned in collaboration with Sangfor.
> - khugepaged may aggressively re-collapse the mTHPs that DAMON splits. A
>   coordination/back-off mechanism is required to avoid ping-pong effects.

Do you really need to khugepaged together, when you already have
DAMOS_COLLAPSE, and anyway you are running DAMON for hugepage splits?

> - SPE data is currently funneled via a userspace daemon and debugfs. Direct
>   kernel-side perf_event sampling integration is planned as a follow-up.

Nice, I think this will make our projects aligned and reduce unnecessary
duplicates.  I'd encourage you to try this path.

> - The rbtree entry TTL (30s) and signal threshold (1/10 of peak) are empirical
>   defaults subject to further tuning.

I don't fully understand this part.  Could you please elaborate?

> - The ARM64 DAMON blind spot (WSS < L2 TLB reach) is a pre-existing hardware-MMU
>   characteristic, not introduced by this series. Setting nr_accesses/min=0
>   serves as an effective workaround for the split path.

I don't fully understand this, too.  Could you please elaborate and enlighten
me?

> 
> Reported-by: Chaobing Dai <daichaobing@sangfor.com.cn>
> Cc: SeongJae Park <sj@kernel.org>
> Cc: Andrew Morton <akpm@linux-foundation.org>
> Cc: Nico Pache <npache@redhat.com>
> Cc: Asier Gutierrez <gutierrez.asier@huawei-partners.com>
> Cc: linux-mm@kvack.org
> Cc: linux-kernel@vger.kernel.org
> Signed-off-by: Wang Lian <lianux.mm@gmail.com>
> 
> Wang Lian (6):
>   mm/damon: add target_order field for DAMOS_COLLAPSE
>   mm/khugepaged: add damon_collapse_folio_range() for external callers
>   mm/damon/vaddr: implement mTHP-aware DAMOS_COLLAPSE handler
>   mm/damon: introduce DAMOS_MTHP_SPLIT action and hot_threshold
>   mm/damon/vaddr: implement DAMOS_MTHP_SPLIT handler
>   mm/damon: add SPE feedback for sub-THP split decisions
> 
>  include/linux/damon.h      |  18 ++
>  include/linux/khugepaged.h |   3 +
>  mm/damon/Kconfig           |  12 +
>  mm/damon/Makefile          |   1 +
>  mm/damon/core.c            |   3 +
>  mm/damon/spe.c             | 505 +++++++++++++++++++++++++++++++++++++
>  mm/damon/spe.h             |  62 +++++
>  mm/damon/sysfs-schemes.c   |  96 +++++++
>  mm/damon/vaddr.c           | 118 +++++++++
>  mm/khugepaged.c            |  39 +++
>  10 files changed, 857 insertions(+)
>  create mode 100644 mm/damon/spe.c
>  create mode 100644 mm/damon/spe.h

Because this is an RFC and we found high level TODO (trying perf event based
appraoch instead of debugfs), I will skip reviewing the details.  If you have
specific parts that want my detailed review, let me know.

Also, the perf event based monitoring is a long term project.  The ETA is the
LSFMMBPF'27.  If you cannot wait until the time, maybe you could try the
alternative approaches (using address filter or user_input quota goal) and
upstreaming dependent parts (DAMOS_COLLAPSE extension for mTHP and DAMOS_SPLIT)
first could also be a nice approach, in my opinion.

[1] https://origin.kernel.org/doc/html/latest/admin-guide/mm/damon/stat.html#memory-idle-ms-percentiles
[2] https://origin.kernel.org/doc/html/latest/mm/damon/design.html#aim-oriented-feedback-driven-auto-tuning
[3] https://lore.kernel.org/20251128193947.80866-1-sj@kernel.org/


Thanks,
SJ

[...]

Re: [RFC PATCH 0/6] mm/damon: Add mTHP-aware collapse/split with ARM SPE feedback

[SeongJae Park]

On Thu, 18 Jun 2026 18:47:16 -0700 SeongJae Park <sj@kernel.org> wrote:

> Hello Lian,
> 
> On Thu, 18 Jun 2026 17:48:32 +0800 Wang Lian <lianux.mm@gmail.com> wrote:
> 
> > Received an off-list report that DAMON significantly overestimates
> > hot memory in KVM/QEMU deployments with THP-backed tmpfs guest memory
> > running Oracle workloads.
> > 
> > The root cause is structural: a PMD entry covers 512 4KB subpages with
> > a single Access Flag (AF) bit. When any one subpage is accessed, the entire
> > 2MB region appears "hot" to DAMON. On ARM64,
> 
> This makes sense to me.  I also agree this could caused the reported problem.
> And this is a known limitation of DAMON.  My suggestion for straightforward
> workaround of this problem is, using 'age' information of DAMON for better
> identification of the hot memory.
> 
> That is, I don't expect real hot data in real production systems will evenly
> scattered.  Even if they are, I don't expect they will all evenly frequently
> accessed.  Only a few of those would be accessed frequently for long.  Even if
> that is, there would be data that frequently for longer.  You could show the
> distriibution of the pattern and find X % of hottest memory as hot.
> 
> We invented idle time percentiles [1] for a similar purpose, though it is more
> focusing on finding cold memory.
> 
> I understand this patch series is trying to make more fundamental and better
> solution on hardware that can do better.  Makes sense to me.
> 
> > this is compounded by the
> > hardware AF mechanism -- the AF is only set on a TLB miss. Consequently, when the
> > working set fits entirely within the L2 TLB (e.g., a 16MB working set with 2MB THP
> > running on a Kunpeng 920's 2048-entry L2 TLB), DAMON becomes completely blind to
> > subsequent accesses.
> 
> This makes sense to me.  However, I don't get how this is contributing to the
> problem.  Could you please elaborate?
> 
> > x86 is not subject to this specific blindness under similar
> > conditions.
> 
> To my understanding on x86, same issue exists.  If TLB hits, Aceessed bit is
> not set, and DAMON shows it as unaccessed.  Am I missing something?
> 
> > 
> > We reproduced this memory inflation on a Kunpeng 920 platform using a synthetic
> > workload (8GB mmap with a 0.2% sparse hotspot, i.e. 16MB actually hot):
> > THP=always causes DAMON to report the entire 8GB as hot, while THP=never
> > reports only a few hundred MB -- a 512x overestimate relative to the actual
> > 16MB hotspot under THP, and a ~33x gap between the two THP modes. ARM SPE hardware profiling
> > independently confirms this asymmetry: out of 2,005 THPs sampled system-wide
> > over 10 seconds, 97% had fewer than 10% of their 4KB subpages actually accessed.
> 
> I don't think the real world production systems to have this very artificial
> access pattern.  I believe (or, hope) use of 'age' can work around the issue in
> a reasonable level for many cases.  I understand this setup is only for PoC,
> and I think this is well designed test for the purpose.  Thank you for sharing
> this.
> 
> > 
> > To mitigate this, this series extends the existing DAMOS_COLLAPSE action to be
> > mTHP-aware via a new target_order field,
> 
> Makes sensee, and sounds nice.  Definitely no one size fits all!
> 
> > and introduces a new
> > DAMOS_MTHP_SPLIT action. This enables DAMON to proactively split PMD THPs
> > into smaller mTHPs
> 
> Nice!  Asier was planning to do similar work in future.  I think you could
> collaborate to reduce unnecessary duplicates!
> 
> I'd suggest making the name simpler and consistent to DAMOS_COLLAPSE, though.
> Say, DAMOS_SPLIT ?
> 
> > when most subpages are probed as cold, and collapse them
> > back when beneficial. To resolve the sub-PMD monitoring blindness, the split
> > path can incorporate fine-grained hardware feedback from ARM SPE.
> > 
> > The hardware feedback loop (damon_spe_folio_heatmap) implements a two-pass
> > signal filter: it first identifies the peak chunk access count, and then marks
> > sub-chunks with >= 1/10 of the peak count as hot, effectively filtering out
> > SPE sampling noise. A configurable hot_threshold (default 30%) controls the
> > split decision: only folios with a hot fraction below this threshold are
> > eligible for splitting. When no SPE data is available, the infrastructure
> > gracefully falls back to explicit PTE-level scanning via folio_walk.
> > 
> > Currently, SPE data is fed from userspace via debugfs (e.g., perf script piped
> > through a histogram builder into /sys/kernel/debug/damon/spe_feed).
> 
> So you implemented a debugfs interface?  That must be a nice approach for PoC.
> But it may be difficult to be upstreamed as is.
> 
> You could build a control plane that decides the exact address ranges to split,
> and directly feed it to DAMOS using DAMOS address filter.  max_nr_snapshots can
> also be useful for making such kind of user space controls more deterministic.
> 
> For simpler user-space control, utilizing user_input DAMOS quota goal [2]
> should also be another option.
> 
> We are also planning [3] to extend DAMON for perf events.  On top of it, we
> might be able to extend it further to utilize ARM SPE by DAMON itself, and do
> all this without the user space help but only DAMOS.
> 
> Baseed on below 'limitations' section, I understand this is only for PoC at the
> moment, and you plan to explore the perf event based approach.  I'd also
> recommend that.
> 
> > 
> > Collapse path (patches 1-3):
> >   DAMON scheme action=COLLAPSE, target_order=N
> >   -> damos_va_collapse() -> damon_collapse_folio_range()
> >   -> collapse_huge_page()
> > 
> > Split path (patches 4-5):
> >   DAMON scheme action=MTHP_SPLIT, target_order=N, hot_threshold=M
> >   -> damos_va_mthp_split() -> damon_spe_hot_fraction()
> >   -> split_folio_to_order()
> > 
> > SPE feedback infrastructure (patch 6):
> >   perf script -> spe_hist -> debugfs spe_feed
> >   -> per-folio rbtree {THP-aligned PFN -> access_count[512]}
> >   -> damon_spe_folio_heatmap() -> hot_bitmap -> split decision
> > 
> > The userspace helper tools (including the spe_hist histogram builder and
> > validation scripts) are archived at:
> >   https://github.com/lianux-mm/damon_spe
> 
> Thank you for making all the grateful code open!
> 
> > 
> > Testing was performed on a Kunpeng 920 system (256 cores, 249GB RAM, base kernel
> > 7.1.0-rc5+):
> > 
> >   T1 ARM64 blind spot: A 16MB THP workload (where 8 PMDs fit entirely within the
> >      L2 TLB) resulted in DAMON detecting 0 regions. Conversely, using 512MB
> >      with 4KB base pages, or a 16GB THP layout (exceeding L2 TLB reach), allowed
> >      DAMON to function normally.
> > 
> >   T2 THP inflation: With an 8GB mmap and 16MB actually hot (0.2%),
> >      THP=always: DAMON reported 8GB hot (512x vs ground truth);
> >      THP=never: ~245MB (15x vs ground truth).  The THP-induced gap
> >      between the two modes was ~33x.
> > 
> >   T3 RocksDB: Fragmented malloc allocation prevented THP formation, and DAMON
> >      behaved normally. We could not reproduce THP inflation with RocksDB.
> >      The workloads fundamentally vulnerable to this structural issue remain KVM
> >      guests, JVM large heaps, and PostgreSQL shared_buffers.
> > 
> >   T4 min=0 deadlock break: A 256MB THP induced the DAMON blind spot.
> >      Triggering an unconditional mthp_split (via nr_accesses/min=0) successfully
> >      shattered the space into 16384x16KB folios, allowing DAMON to fully recover.
> > 
> >   T5 ARM SPE histogram: Out of 2005 sampled THPs, 97% exhibited <10% hot subpages.
> >      A typical trace showed PFN 0x820db800 accumulated 39,794 hardware accesses
> >      concentrated across only 3 out of 512 subpages.
> > 
> >   End-to-end: Verified hot/cold discrimination. The SPE feed preserved a 90%
> >      hot THP intact, while successfully splitting a 25% cold THP into 128x16KB folios.
> > 
> > Known limitations:
> > - The full KVM + Oracle production chain has not yet been benchmarked end-to-end.
> >   While individual component verification is complete, full integration testing
> >   is planned in collaboration with Sangfor.
> > - khugepaged may aggressively re-collapse the mTHPs that DAMON splits. A
> >   coordination/back-off mechanism is required to avoid ping-pong effects.
> 
> Do you really need to khugepaged together, when you already have
> DAMOS_COLLAPSE, and anyway you are running DAMON for hugepage splits?
> 
> > - SPE data is currently funneled via a userspace daemon and debugfs. Direct
> >   kernel-side perf_event sampling integration is planned as a follow-up.
> 
> Nice, I think this will make our projects aligned and reduce unnecessary
> duplicates.  I'd encourage you to try this path.
> 
> > - The rbtree entry TTL (30s) and signal threshold (1/10 of peak) are empirical
> >   defaults subject to further tuning.
> 
> I don't fully understand this part.  Could you please elaborate?
> 
> > - The ARM64 DAMON blind spot (WSS < L2 TLB reach) is a pre-existing hardware-MMU
> >   characteristic, not introduced by this series. Setting nr_accesses/min=0
> >   serves as an effective workaround for the split path.
> 
> I don't fully understand this, too.  Could you please elaborate and enlighten
> me?
> 
> > 
> > Reported-by: Chaobing Dai <daichaobing@sangfor.com.cn>
> > Cc: SeongJae Park <sj@kernel.org>
> > Cc: Andrew Morton <akpm@linux-foundation.org>
> > Cc: Nico Pache <npache@redhat.com>
> > Cc: Asier Gutierrez <gutierrez.asier@huawei-partners.com>
> > Cc: linux-mm@kvack.org
> > Cc: linux-kernel@vger.kernel.org
> > Signed-off-by: Wang Lian <lianux.mm@gmail.com>
> > 
> > Wang Lian (6):
> >   mm/damon: add target_order field for DAMOS_COLLAPSE
> >   mm/khugepaged: add damon_collapse_folio_range() for external callers
> >   mm/damon/vaddr: implement mTHP-aware DAMOS_COLLAPSE handler
> >   mm/damon: introduce DAMOS_MTHP_SPLIT action and hot_threshold
> >   mm/damon/vaddr: implement DAMOS_MTHP_SPLIT handler
> >   mm/damon: add SPE feedback for sub-THP split decisions
> > 
> >  include/linux/damon.h      |  18 ++
> >  include/linux/khugepaged.h |   3 +
> >  mm/damon/Kconfig           |  12 +
> >  mm/damon/Makefile          |   1 +
> >  mm/damon/core.c            |   3 +
> >  mm/damon/spe.c             | 505 +++++++++++++++++++++++++++++++++++++
> >  mm/damon/spe.h             |  62 +++++
> >  mm/damon/sysfs-schemes.c   |  96 +++++++
> >  mm/damon/vaddr.c           | 118 +++++++++
> >  mm/khugepaged.c            |  39 +++
> >  10 files changed, 857 insertions(+)
> >  create mode 100644 mm/damon/spe.c
> >  create mode 100644 mm/damon/spe.h
> 
> Because this is an RFC and we found high level TODO (trying perf event based
> appraoch instead of debugfs), I will skip reviewing the details.  If you have
> specific parts that want my detailed review, let me know.
> 
> Also, the perf event based monitoring is a long term project.  The ETA is the
> LSFMMBPF'27.  If you cannot wait until the time, maybe you could try the
> alternative approaches (using address filter or user_input quota goal) and
> upstreaming dependent parts (DAMOS_COLLAPSE extension for mTHP and DAMOS_SPLIT)
> first could also be a nice approach, in my opinion.
> 
> [1] https://origin.kernel.org/doc/html/latest/admin-guide/mm/damon/stat.html#memory-idle-ms-percentiles
> [2] https://origin.kernel.org/doc/html/latest/mm/damon/design.html#aim-oriented-feedback-driven-auto-tuning
> [3] https://lore.kernel.org/20251128193947.80866-1-sj@kernel.org/

The above link ([3]) is wrong, sorry.  Please use below.

[3] https://lore.kernel.org/20260525225208.1179-1-sj@kernel.org/


Thanks,
SJ

[...]

Re: [RFC PATCH 0/6] mm/damon: Add mTHP-aware collapse/split with ARM SPE feedback

[SeongJae Park]

+ damon@lists.linux.dev

Please Cc damon@lists.linux.dev from the next revision, and all DAMON patches
in future.


Thanks,
SJ

On Thu, 18 Jun 2026 18:54:23 -0700 SeongJae Park <sj@kernel.org> wrote:

> On Thu, 18 Jun 2026 18:47:16 -0700 SeongJae Park <sj@kernel.org> wrote:
> 
> > Hello Lian,
> > 
> > On Thu, 18 Jun 2026 17:48:32 +0800 Wang Lian <lianux.mm@gmail.com> wrote:
> > 
> > > Received an off-list report that DAMON significantly overestimates
> > > hot memory in KVM/QEMU deployments with THP-backed tmpfs guest memory
> > > running Oracle workloads.
> > > 
> > > The root cause is structural: a PMD entry covers 512 4KB subpages with
> > > a single Access Flag (AF) bit. When any one subpage is accessed, the entire
> > > 2MB region appears "hot" to DAMON. On ARM64,
> > 
> > This makes sense to me.  I also agree this could caused the reported problem.
> > And this is a known limitation of DAMON.  My suggestion for straightforward
> > workaround of this problem is, using 'age' information of DAMON for better
> > identification of the hot memory.
> > 
> > That is, I don't expect real hot data in real production systems will evenly
> > scattered.  Even if they are, I don't expect they will all evenly frequently
> > accessed.  Only a few of those would be accessed frequently for long.  Even if
> > that is, there would be data that frequently for longer.  You could show the
> > distriibution of the pattern and find X % of hottest memory as hot.
> > 
> > We invented idle time percentiles [1] for a similar purpose, though it is more
> > focusing on finding cold memory.
> > 
> > I understand this patch series is trying to make more fundamental and better
> > solution on hardware that can do better.  Makes sense to me.
> > 
> > > this is compounded by the
> > > hardware AF mechanism -- the AF is only set on a TLB miss. Consequently, when the
> > > working set fits entirely within the L2 TLB (e.g., a 16MB working set with 2MB THP
> > > running on a Kunpeng 920's 2048-entry L2 TLB), DAMON becomes completely blind to
> > > subsequent accesses.
> > 
> > This makes sense to me.  However, I don't get how this is contributing to the
> > problem.  Could you please elaborate?
> > 
> > > x86 is not subject to this specific blindness under similar
> > > conditions.
> > 
> > To my understanding on x86, same issue exists.  If TLB hits, Aceessed bit is
> > not set, and DAMON shows it as unaccessed.  Am I missing something?
> > 
> > > 
> > > We reproduced this memory inflation on a Kunpeng 920 platform using a synthetic
> > > workload (8GB mmap with a 0.2% sparse hotspot, i.e. 16MB actually hot):
> > > THP=always causes DAMON to report the entire 8GB as hot, while THP=never
> > > reports only a few hundred MB -- a 512x overestimate relative to the actual
> > > 16MB hotspot under THP, and a ~33x gap between the two THP modes. ARM SPE hardware profiling
> > > independently confirms this asymmetry: out of 2,005 THPs sampled system-wide
> > > over 10 seconds, 97% had fewer than 10% of their 4KB subpages actually accessed.
> > 
> > I don't think the real world production systems to have this very artificial
> > access pattern.  I believe (or, hope) use of 'age' can work around the issue in
> > a reasonable level for many cases.  I understand this setup is only for PoC,
> > and I think this is well designed test for the purpose.  Thank you for sharing
> > this.
> > 
> > > 
> > > To mitigate this, this series extends the existing DAMOS_COLLAPSE action to be
> > > mTHP-aware via a new target_order field,
> > 
> > Makes sensee, and sounds nice.  Definitely no one size fits all!
> > 
> > > and introduces a new
> > > DAMOS_MTHP_SPLIT action. This enables DAMON to proactively split PMD THPs
> > > into smaller mTHPs
> > 
> > Nice!  Asier was planning to do similar work in future.  I think you could
> > collaborate to reduce unnecessary duplicates!
> > 
> > I'd suggest making the name simpler and consistent to DAMOS_COLLAPSE, though.
> > Say, DAMOS_SPLIT ?
> > 
> > > when most subpages are probed as cold, and collapse them
> > > back when beneficial. To resolve the sub-PMD monitoring blindness, the split
> > > path can incorporate fine-grained hardware feedback from ARM SPE.
> > > 
> > > The hardware feedback loop (damon_spe_folio_heatmap) implements a two-pass
> > > signal filter: it first identifies the peak chunk access count, and then marks
> > > sub-chunks with >= 1/10 of the peak count as hot, effectively filtering out
> > > SPE sampling noise. A configurable hot_threshold (default 30%) controls the
> > > split decision: only folios with a hot fraction below this threshold are
> > > eligible for splitting. When no SPE data is available, the infrastructure
> > > gracefully falls back to explicit PTE-level scanning via folio_walk.
> > > 
> > > Currently, SPE data is fed from userspace via debugfs (e.g., perf script piped
> > > through a histogram builder into /sys/kernel/debug/damon/spe_feed).
> > 
> > So you implemented a debugfs interface?  That must be a nice approach for PoC.
> > But it may be difficult to be upstreamed as is.
> > 
> > You could build a control plane that decides the exact address ranges to split,
> > and directly feed it to DAMOS using DAMOS address filter.  max_nr_snapshots can
> > also be useful for making such kind of user space controls more deterministic.
> > 
> > For simpler user-space control, utilizing user_input DAMOS quota goal [2]
> > should also be another option.
> > 
> > We are also planning [3] to extend DAMON for perf events.  On top of it, we
> > might be able to extend it further to utilize ARM SPE by DAMON itself, and do
> > all this without the user space help but only DAMOS.
> > 
> > Baseed on below 'limitations' section, I understand this is only for PoC at the
> > moment, and you plan to explore the perf event based approach.  I'd also
> > recommend that.
> > 
> > > 
> > > Collapse path (patches 1-3):
> > >   DAMON scheme action=COLLAPSE, target_order=N
> > >   -> damos_va_collapse() -> damon_collapse_folio_range()
> > >   -> collapse_huge_page()
> > > 
> > > Split path (patches 4-5):
> > >   DAMON scheme action=MTHP_SPLIT, target_order=N, hot_threshold=M
> > >   -> damos_va_mthp_split() -> damon_spe_hot_fraction()
> > >   -> split_folio_to_order()
> > > 
> > > SPE feedback infrastructure (patch 6):
> > >   perf script -> spe_hist -> debugfs spe_feed
> > >   -> per-folio rbtree {THP-aligned PFN -> access_count[512]}
> > >   -> damon_spe_folio_heatmap() -> hot_bitmap -> split decision
> > > 
> > > The userspace helper tools (including the spe_hist histogram builder and
> > > validation scripts) are archived at:
> > >   https://github.com/lianux-mm/damon_spe
> > 
> > Thank you for making all the grateful code open!
> > 
> > > 
> > > Testing was performed on a Kunpeng 920 system (256 cores, 249GB RAM, base kernel
> > > 7.1.0-rc5+):
> > > 
> > >   T1 ARM64 blind spot: A 16MB THP workload (where 8 PMDs fit entirely within the
> > >      L2 TLB) resulted in DAMON detecting 0 regions. Conversely, using 512MB
> > >      with 4KB base pages, or a 16GB THP layout (exceeding L2 TLB reach), allowed
> > >      DAMON to function normally.
> > > 
> > >   T2 THP inflation: With an 8GB mmap and 16MB actually hot (0.2%),
> > >      THP=always: DAMON reported 8GB hot (512x vs ground truth);
> > >      THP=never: ~245MB (15x vs ground truth).  The THP-induced gap
> > >      between the two modes was ~33x.
> > > 
> > >   T3 RocksDB: Fragmented malloc allocation prevented THP formation, and DAMON
> > >      behaved normally. We could not reproduce THP inflation with RocksDB.
> > >      The workloads fundamentally vulnerable to this structural issue remain KVM
> > >      guests, JVM large heaps, and PostgreSQL shared_buffers.
> > > 
> > >   T4 min=0 deadlock break: A 256MB THP induced the DAMON blind spot.
> > >      Triggering an unconditional mthp_split (via nr_accesses/min=0) successfully
> > >      shattered the space into 16384x16KB folios, allowing DAMON to fully recover.
> > > 
> > >   T5 ARM SPE histogram: Out of 2005 sampled THPs, 97% exhibited <10% hot subpages.
> > >      A typical trace showed PFN 0x820db800 accumulated 39,794 hardware accesses
> > >      concentrated across only 3 out of 512 subpages.
> > > 
> > >   End-to-end: Verified hot/cold discrimination. The SPE feed preserved a 90%
> > >      hot THP intact, while successfully splitting a 25% cold THP into 128x16KB folios.
> > > 
> > > Known limitations:
> > > - The full KVM + Oracle production chain has not yet been benchmarked end-to-end.
> > >   While individual component verification is complete, full integration testing
> > >   is planned in collaboration with Sangfor.
> > > - khugepaged may aggressively re-collapse the mTHPs that DAMON splits. A
> > >   coordination/back-off mechanism is required to avoid ping-pong effects.
> > 
> > Do you really need to khugepaged together, when you already have
> > DAMOS_COLLAPSE, and anyway you are running DAMON for hugepage splits?
> > 
> > > - SPE data is currently funneled via a userspace daemon and debugfs. Direct
> > >   kernel-side perf_event sampling integration is planned as a follow-up.
> > 
> > Nice, I think this will make our projects aligned and reduce unnecessary
> > duplicates.  I'd encourage you to try this path.
> > 
> > > - The rbtree entry TTL (30s) and signal threshold (1/10 of peak) are empirical
> > >   defaults subject to further tuning.
> > 
> > I don't fully understand this part.  Could you please elaborate?
> > 
> > > - The ARM64 DAMON blind spot (WSS < L2 TLB reach) is a pre-existing hardware-MMU
> > >   characteristic, not introduced by this series. Setting nr_accesses/min=0
> > >   serves as an effective workaround for the split path.
> > 
> > I don't fully understand this, too.  Could you please elaborate and enlighten
> > me?
> > 
> > > 
> > > Reported-by: Chaobing Dai <daichaobing@sangfor.com.cn>
> > > Cc: SeongJae Park <sj@kernel.org>
> > > Cc: Andrew Morton <akpm@linux-foundation.org>
> > > Cc: Nico Pache <npache@redhat.com>
> > > Cc: Asier Gutierrez <gutierrez.asier@huawei-partners.com>
> > > Cc: linux-mm@kvack.org
> > > Cc: linux-kernel@vger.kernel.org
> > > Signed-off-by: Wang Lian <lianux.mm@gmail.com>
> > > 
> > > Wang Lian (6):
> > >   mm/damon: add target_order field for DAMOS_COLLAPSE
> > >   mm/khugepaged: add damon_collapse_folio_range() for external callers
> > >   mm/damon/vaddr: implement mTHP-aware DAMOS_COLLAPSE handler
> > >   mm/damon: introduce DAMOS_MTHP_SPLIT action and hot_threshold
> > >   mm/damon/vaddr: implement DAMOS_MTHP_SPLIT handler
> > >   mm/damon: add SPE feedback for sub-THP split decisions
> > > 
> > >  include/linux/damon.h      |  18 ++
> > >  include/linux/khugepaged.h |   3 +
> > >  mm/damon/Kconfig           |  12 +
> > >  mm/damon/Makefile          |   1 +
> > >  mm/damon/core.c            |   3 +
> > >  mm/damon/spe.c             | 505 +++++++++++++++++++++++++++++++++++++
> > >  mm/damon/spe.h             |  62 +++++
> > >  mm/damon/sysfs-schemes.c   |  96 +++++++
> > >  mm/damon/vaddr.c           | 118 +++++++++
> > >  mm/khugepaged.c            |  39 +++
> > >  10 files changed, 857 insertions(+)
> > >  create mode 100644 mm/damon/spe.c
> > >  create mode 100644 mm/damon/spe.h
> > 
> > Because this is an RFC and we found high level TODO (trying perf event based
> > appraoch instead of debugfs), I will skip reviewing the details.  If you have
> > specific parts that want my detailed review, let me know.
> > 
> > Also, the perf event based monitoring is a long term project.  The ETA is the
> > LSFMMBPF'27.  If you cannot wait until the time, maybe you could try the
> > alternative approaches (using address filter or user_input quota goal) and
> > upstreaming dependent parts (DAMOS_COLLAPSE extension for mTHP and DAMOS_SPLIT)
> > first could also be a nice approach, in my opinion.
> > 
> > [1] https://origin.kernel.org/doc/html/latest/admin-guide/mm/damon/stat.html#memory-idle-ms-percentiles
> > [2] https://origin.kernel.org/doc/html/latest/mm/damon/design.html#aim-oriented-feedback-driven-auto-tuning
> > [3] https://lore.kernel.org/20251128193947.80866-1-sj@kernel.org/
> 
> The above link ([3]) is wrong, sorry.  Please use below.
> 
> [3] https://lore.kernel.org/20260525225208.1179-1-sj@kernel.org/
> 
> 
> Thanks,
> SJ
> 
> [...]
> 

Sent using hkml (https://github.com/sjp38/hackermail)

Re: [RFC PATCH 0/6] mm/damon: Add mTHP-aware collapse/split with ARM SPE feedback

[Wang Lian]

Hi SeongJae,

Thank you for the thorough and thoughtful review.  Your feedback on the
x86 AF behavior was an important correction -- I'll address that and
your other questions below.

On Thu, 18 Jun 2026 SeongJae Park <sj@kernel.org> wrote:

> This makes sense to me.  I also agree this could caused the reported
> problem.  And this is a known limitation of DAMON.  My suggestion for
> straightforward workaround of this problem is, using 'age' information
> of DAMON for better identification of the hot memory.

Thank you for pointing out idle time percentiles [1].  We agree that 'age'
helps differentiate frequently-accessed from occasionally-accessed regions,
and it is a good workaround for many cases.

However, age operates at region granularity, which is still at or above
PMD level for THP-mapped memory.  When only a few 4KB subpages within a
2MB THP are hot, age tells us the region has been accessed recently, but
not which subpages are hot.  The split decision needs sub-PMD information,
which is what the SPE heatmap provides.

That said, combining age with split could be valuable: split only regions
that have been consistently hot (high age) AND have sparse sub-page access
patterns.  We will explore this.

> > On ARM64, this is compounded by the hardware AF mechanism -- the AF
> > is only set on a TLB miss.
>
> This makes sense to me.  However, I don't get how this is contributing
> to the problem.  Could you please elaborate?

The AF-on-TLB-miss behavior creates a second-order problem that directly
exacerbates the overestimation. 

When DAMON's mkold path clears the PMD AF, it deliberately skips the TLB 
flush to minimize overhead. If the dense working set fits entirely within 
the L2 TLB (e.g., 16MB workload using 8 PMD entries on Kunpeng 920's 2048-entry 
L2 TLB), subsequent hardware accesses hit the valid, stale TLB entries 
directly. The hardware MMU never generates a page table walk, so the 
in-memory PMD AF stays 0. 

Consequently, DAMON sees `nr_accesses = 0` and assumes the region is completely 
cold, making it impossible to naturally track the sub-page usage shifts. When 
sporadic/noise accesses later hit other parts of this "seemingly cold" PMD 
and trigger an isolated TLB refilling, DAMON abruptly sees the whole 2MB 
as hot. This binary oscillation (completely blind vs. fully hot) is what 
drives the massive overestimation under THP.

We confirmed this TLB-reach aspect empirically via our T1 test:
  16MB THP (8 PMDs, 0.4% of L2 TLB reach) -> DAMON tracks 0 accesses (blind)
  16GB THP (8192 PMDs, 400% of L2 TLB reach) -> DAMON tracks normally due to natural eviction

> > x86 is not subject to this specific blindness under similar
> > conditions.
>
> To my understanding on x86, same issue exists.  If TLB hits, Aceessed
> bit is not set, and DAMON shows it as unaccessed.  Am I missing
> something?

You are entirely right, and I was wrong on this point. I re-checked the 
kernel source and verified that x86's ptep_test_and_clear_young() does NOT 
flush the TLB. Even ptep_clear_flush_young() on x86 deliberately skips the 
flush as a performance optimization (arch/x86/mm/pgtable.c:486-502). The 
same optimization architectural behavior exists on PowerPC and RISC-V.

Therefore, both architectures are theoretically vulnerable to this stale-TLB 
blind spot under identical tightly-fit workloads. Our initial assumption 
was biased because T1 was only conducted on ARM64. We will reproduce the 
T1 setup on x86 to verify the exact behavior, and I will correct this 
claim in the v2 cover letter. Thank you for catching this mistake.

> Nice!  Asier was planning to do similar work in future.  I think you
> could collaborate to reduce unnecessary duplicates!

Great to hear! We would be happy to collaborate with Asier. I'll reach
out to him to coordinate our efforts.

> I'd suggest making the name simpler and consistent to DAMOS_COLLAPSE,
> though.  Say, DAMOS_SPLIT ?

Agreed. DAMOS_SPLIT is cleaner and fits the existing naming convention 
perfectly. Will rename in v2.

> So you implemented a debugfs interface?  That must be a nice approach
> for PoC.  But it may be difficult to be upstreamed as is.
>
> You could build a control plane that decides the exact address ranges
> to split, and directly feed it to DAMOS using DAMOS address filter.

The native perf event approach [3] aligns perfectly with our long-term 
Phase 2c plan, and we are highly interested in collaborating on it to 
eliminate the userspace daemon and debugfs bridge entirely.

However, since native kernel-side SPE handling is a long-term item, we 
will follow your pragmatic alternative suggestion for v2: use DAMOS address 
filters or user_input quota goals [2] to feed the split decisions from 
userspace cleanly. This allows us to upstream the core infrastructure 
(mTHP target_order for collapse and the new DAMOS_SPLIT action) first.

> Do you really need to khugepaged together, when you already have
> DAMOS_COLLAPSE, and anyway you are running DAMON for hugepage splits?

Excellent point. Running both concurrently on the same VMA introduces 
redundancy and heavy ping-pong effects. 

Option (b) is definitely cleaner: we will let DAMON handle both split and 
re-collapse decisions using its own access data. To make this robust in 
production environments where khugepaged is globally enabled, we will 
explore having the DAMOS_SPLIT path temporarily mark the target ranges 
(e.g., via a pseudo-VM_NOHUGEPAGE backing off mechanism) to prevent 
khugepaged from immediately undoing DAMON's work.

> > - The rbtree entry TTL (30s) and signal threshold (1/10 of peak) are
> >   empirical defaults subject to further tuning.
>
> I don't fully understand this part.  Could you please elaborate?

Since ARM SPE samples hardware accesses instruction-by-instruction, the raw 
data is highly statistical and noisy. 

The TTL (30s) defines the lifecycle of our per-folio rbtree tracking entries. 
Entries not updated within 30 seconds are pruned to prevent stale tracking data 
from corrupting split decisions after a workload phase change. 30s is selected 
to comfortably outlive DAMON's aggregation intervals while keeping the rbtree 
memory footprint tightly bounded.

The signal threshold (1/10 of peak) filters out the statistical sampling noise. 
Instead of treating any subpage with access > 0 as hot, the algorithm finds the 
peak access count inside the 2MB region and only marks sub-chunks with >= 1/10 
of that peak as genuinely hot. On Kunpeng 920, this specific threshold successfully 
reduced false-hot subpage classifications from ~50% to <5%. We plan to make 
these parameters sysfs-configurable.

> > - The ARM64 DAMON blind spot (WSS < L2 TLB reach) is a pre-existing
> >   hardware-MMU characteristic, not introduced by this series.  Setting
> >   nr_accesses/min=0 serves as an effective workaround for the split path.
>
> I don't fully understand this, too.  Could you please elaborate and
> enlighten me?

The blind spot creates an operational deadlock for the split infrastructure:
  1. WSS < TLB reach -> All THP entries stay cached in TLB.
  2. DAMON's page-table scan yields `nr_accesses = 0` globally.
  3. A scheme requiring `nr_accesses.min = 1` never fires -> DAMOS_SPLIT is never invoked.
  4. THPs remain unsplit -> WSS remains within TLB reach -> Loop returns to step 1.

Setting `nr_accesses.min = 0` and `max = 0` breaks this deadlock. It forces 
DAMON to evaluate these seemingly "dead/cold" regions. Once the split handler 
invokes, it checks the ARM SPE telemetry (which captures data directly from the 
instruction pipeline, completely bypassing the MMU page-table AF limitation). 
If SPE reveals a sparse access heatmap, the split is executed. Once shattered into 
mTHP/base pages, the TLB reach drops, natural TLB misses resume, and DAMON's 
standard page-table tracking fully recovers.


Thanks again for your guidance. The action items for v2 are locked in:
  1. Rename DAMOS_MTHP_SPLIT -> DAMOS_SPLIT.
  2. Drop debugfs in favor of DAMOS address filters / control plane.
  3. Correct x86 AF behavior statements in the cover letter.
  4. Coordinate with Asier on split/collapse unification.
  5. Implement back-off to prevent khugepaged ping-pong under Option (b).

Best regards,
Wang Lian

Re: [RFC PATCH 0/6] mm/damon: Add mTHP-aware collapse/split with ARM SPE feedback

[Gutierrez Asier]

On 6/19/2026 6:40 AM, Wang Lian wrote:
> Hi SeongJae,
> 
> Thank you for the thorough and thoughtful review.  Your feedback on the
> x86 AF behavior was an important correction -- I'll address that and
> your other questions below.
> 
> On Thu, 18 Jun 2026 SeongJae Park <sj@kernel.org> wrote:
> 
>> This makes sense to me.  I also agree this could caused the reported
>> problem.  And this is a known limitation of DAMON.  My suggestion for
>> straightforward workaround of this problem is, using 'age' information
>> of DAMON for better identification of the hot memory.
> 
> Thank you for pointing out idle time percentiles [1].  We agree that 'age'
> helps differentiate frequently-accessed from occasionally-accessed regions,
> and it is a good workaround for many cases.
> 
> However, age operates at region granularity, which is still at or above
> PMD level for THP-mapped memory.  When only a few 4KB subpages within a
> 2MB THP are hot, age tells us the region has been accessed recently, but
> not which subpages are hot.  The split decision needs sub-PMD information,
> which is what the SPE heatmap provides.
> 
> That said, combining age with split could be valuable: split only regions
> that have been consistently hot (high age) AND have sparse sub-page access
> patterns.  We will explore this.
> 
>>> On ARM64, this is compounded by the hardware AF mechanism -- the AF
>>> is only set on a TLB miss.
>>
>> This makes sense to me.  However, I don't get how this is contributing
>> to the problem.  Could you please elaborate?
> 
> The AF-on-TLB-miss behavior creates a second-order problem that directly
> exacerbates the overestimation. 
> 
> When DAMON's mkold path clears the PMD AF, it deliberately skips the TLB 
> flush to minimize overhead. If the dense working set fits entirely within 
> the L2 TLB (e.g., 16MB workload using 8 PMD entries on Kunpeng 920's 2048-entry 
> L2 TLB), subsequent hardware accesses hit the valid, stale TLB entries 
> directly. The hardware MMU never generates a page table walk, so the 
> in-memory PMD AF stays 0. 
> 
> Consequently, DAMON sees `nr_accesses = 0` and assumes the region is completely 
> cold, making it impossible to naturally track the sub-page usage shifts. When 
> sporadic/noise accesses later hit other parts of this "seemingly cold" PMD 
> and trigger an isolated TLB refilling, DAMON abruptly sees the whole 2MB 
> as hot. This binary oscillation (completely blind vs. fully hot) is what 
> drives the massive overestimation under THP.
> 
> We confirmed this TLB-reach aspect empirically via our T1 test:
>   16MB THP (8 PMDs, 0.4% of L2 TLB reach) -> DAMON tracks 0 accesses (blind)
>   16GB THP (8192 PMDs, 400% of L2 TLB reach) -> DAMON tracks normally due to natural eviction
> 
>>> x86 is not subject to this specific blindness under similar
>>> conditions.
>>
>> To my understanding on x86, same issue exists.  If TLB hits, Aceessed
>> bit is not set, and DAMON shows it as unaccessed.  Am I missing
>> something?
> 
> You are entirely right, and I was wrong on this point. I re-checked the 
> kernel source and verified that x86's ptep_test_and_clear_young() does NOT 
> flush the TLB. Even ptep_clear_flush_young() on x86 deliberately skips the 
> flush as a performance optimization (arch/x86/mm/pgtable.c:486-502). The 
> same optimization architectural behavior exists on PowerPC and RISC-V.
> 
> Therefore, both architectures are theoretically vulnerable to this stale-TLB 
> blind spot under identical tightly-fit workloads. Our initial assumption 
> was biased because T1 was only conducted on ARM64. We will reproduce the 
> T1 setup on x86 to verify the exact behavior, and I will correct this 
> claim in the v2 cover letter. Thank you for catching this mistake.
> 
>> Nice!  Asier was planning to do similar work in future.  I think you
>> could collaborate to reduce unnecessary duplicates!
> 
> Great to hear! We would be happy to collaborate with Asier. I'll reach
> out to him to coordinate our efforts.
Sure, I will be happy to cooperate.
>> I'd suggest making the name simpler and consistent to DAMOS_COLLAPSE,
>> though.  Say, DAMOS_SPLIT ?
> 
> Agreed. DAMOS_SPLIT is cleaner and fits the existing naming convention 
> perfectly. Will rename in v2.
> 
>> So you implemented a debugfs interface?  That must be a nice approach
>> for PoC.  But it may be difficult to be upstreamed as is.
>>
>> You could build a control plane that decides the exact address ranges
>> to split, and directly feed it to DAMOS using DAMOS address filter.
> 
> The native perf event approach [3] aligns perfectly with our long-term 
> Phase 2c plan, and we are highly interested in collaborating on it to 
> eliminate the userspace daemon and debugfs bridge entirely.
> 
> However, since native kernel-side SPE handling is a long-term item, we 
> will follow your pragmatic alternative suggestion for v2: use DAMOS address 
> filters or user_input quota goals [2] to feed the split decisions from 
> userspace cleanly. This allows us to upstream the core infrastructure 
> (mTHP target_order for collapse and the new DAMOS_SPLIT action) first.
> 
>> Do you really need to khugepaged together, when you already have
>> DAMOS_COLLAPSE, and anyway you are running DAMON for hugepage splits?
> 
> Excellent point. Running both concurrently on the same VMA introduces 
> redundancy and heavy ping-pong effects. 
> 
> Option (b) is definitely cleaner: we will let DAMON handle both split and 
> re-collapse decisions using its own access data. To make this robust in 
> production environments where khugepaged is globally enabled, we will 
> explore having the DAMOS_SPLIT path temporarily mark the target ranges 
> (e.g., via a pseudo-VM_NOHUGEPAGE backing off mechanism) to prevent 
> khugepaged from immediately undoing DAMON's work.
> 
>>> - The rbtree entry TTL (30s) and signal threshold (1/10 of peak) are
>>>   empirical defaults subject to further tuning.
>>
>> I don't fully understand this part.  Could you please elaborate?
> 
> Since ARM SPE samples hardware accesses instruction-by-instruction, the raw 
> data is highly statistical and noisy. 
> 
> The TTL (30s) defines the lifecycle of our per-folio rbtree tracking entries. 
> Entries not updated within 30 seconds are pruned to prevent stale tracking data 
> from corrupting split decisions after a workload phase change. 30s is selected 
> to comfortably outlive DAMON's aggregation intervals while keeping the rbtree 
> memory footprint tightly bounded.
> 
> The signal threshold (1/10 of peak) filters out the statistical sampling noise. 
> Instead of treating any subpage with access > 0 as hot, the algorithm finds the 
> peak access count inside the 2MB region and only marks sub-chunks with >= 1/10 
> of that peak as genuinely hot. On Kunpeng 920, this specific threshold successfully 
> reduced false-hot subpage classifications from ~50% to <5%. We plan to make 
> these parameters sysfs-configurable.
> 
>>> - The ARM64 DAMON blind spot (WSS < L2 TLB reach) is a pre-existing
>>>   hardware-MMU characteristic, not introduced by this series.  Setting
>>>   nr_accesses/min=0 serves as an effective workaround for the split path.
>>
>> I don't fully understand this, too.  Could you please elaborate and
>> enlighten me?
> 
> The blind spot creates an operational deadlock for the split infrastructure:
>   1. WSS < TLB reach -> All THP entries stay cached in TLB.
>   2. DAMON's page-table scan yields `nr_accesses = 0` globally.
>   3. A scheme requiring `nr_accesses.min = 1` never fires -> DAMOS_SPLIT is never invoked.
>   4. THPs remain unsplit -> WSS remains within TLB reach -> Loop returns to step 1.
> 
> Setting `nr_accesses.min = 0` and `max = 0` breaks this deadlock. It forces 
> DAMON to evaluate these seemingly "dead/cold" regions. Once the split handler 
> invokes, it checks the ARM SPE telemetry (which captures data directly from the 
> instruction pipeline, completely bypassing the MMU page-table AF limitation). 
> If SPE reveals a sparse access heatmap, the split is executed. Once shattered into 
> mTHP/base pages, the TLB reach drops, natural TLB misses resume, and DAMON's 
> standard page-table tracking fully recovers.
> 
> 
> Thanks again for your guidance. The action items for v2 are locked in:
>   1. Rename DAMOS_MTHP_SPLIT -> DAMOS_SPLIT.
>   2. Drop debugfs in favor of DAMOS address filters / control plane.
>   3. Correct x86 AF behavior statements in the cover letter.
>   4. Coordinate with Asier on split/collapse unification.
>   5. Implement back-off to prevent khugepaged ping-pong under Option (b).
> 
> Best regards,
> Wang Lian

-- 
Asier Gutierrez
Huawei

Re: [RFC PATCH 0/6] mm/damon: Add mTHP-aware collapse/split with ARM SPE feedback

[SeongJae Park]

+ damon@lists.linux.dev

On Fri, 19 Jun 2026 17:31:33 +0300 Gutierrez Asier <gutierrez.asier@huawei-partners.com> wrote:

> 
> 
> On 6/19/2026 6:40 AM, Wang Lian wrote:
> > Hi SeongJae,
> > 
> > Thank you for the thorough and thoughtful review.  Your feedback on the
> > x86 AF behavior was an important correction -- I'll address that and
> > your other questions below.
> > 
> > On Thu, 18 Jun 2026 SeongJae Park <sj@kernel.org> wrote:
> > 
> >> This makes sense to me.  I also agree this could caused the reported
> >> problem.  And this is a known limitation of DAMON.  My suggestion for
> >> straightforward workaround of this problem is, using 'age' information
> >> of DAMON for better identification of the hot memory.
> > 
> > Thank you for pointing out idle time percentiles [1].  We agree that 'age'
> > helps differentiate frequently-accessed from occasionally-accessed regions,
> > and it is a good workaround for many cases.
> > 
> > However, age operates at region granularity, which is still at or above
> > PMD level for THP-mapped memory.  When only a few 4KB subpages within a
> > 2MB THP are hot, age tells us the region has been accessed recently, but
> > not which subpages are hot.  The split decision needs sub-PMD information,
> > which is what the SPE heatmap provides.
> > 
> > That said, combining age with split could be valuable: split only regions
> > that have been consistently hot (high age) AND have sparse sub-page access
> > patterns.  We will explore this.

Yes, I agree.  Using features like SPE in addition to 'age' will make it much
better.

> > 
> >>> On ARM64, this is compounded by the hardware AF mechanism -- the AF
> >>> is only set on a TLB miss.
> >>
> >> This makes sense to me.  However, I don't get how this is contributing
> >> to the problem.  Could you please elaborate?
> > 
> > The AF-on-TLB-miss behavior creates a second-order problem that directly
> > exacerbates the overestimation. 
> > 
> > When DAMON's mkold path clears the PMD AF, it deliberately skips the TLB 
> > flush to minimize overhead. If the dense working set fits entirely within 
> > the L2 TLB (e.g., 16MB workload using 8 PMD entries on Kunpeng 920's 2048-entry 
> > L2 TLB), subsequent hardware accesses hit the valid, stale TLB entries 
> > directly. The hardware MMU never generates a page table walk, so the 
> > in-memory PMD AF stays 0. 

Yes, that all makes sense.  But, let's not call it "stale" TLB entry.  TLB is
only for translation and the entry is doing its role.  We do not flush TLB by
purpose.  Nothing is stale here.

> > 
> > Consequently, DAMON sees `nr_accesses = 0` and assumes the region is completely 
> > cold, making it impossible to naturally track the sub-page usage shifts. When 
> > sporadic/noise accesses later hit other parts of this "seemingly cold" PMD 
> > and trigger an isolated TLB refilling, DAMON abruptly sees the whole 2MB 
> > as hot. This binary oscillation (completely blind vs. fully hot) is what 
> > drives the massive overestimation under THP.
> > 
> > We confirmed this TLB-reach aspect empirically via our T1 test:
> >   16MB THP (8 PMDs, 0.4% of L2 TLB reach) -> DAMON tracks 0 accesses (blind)
> >   16GB THP (8192 PMDs, 400% of L2 TLB reach) -> DAMON tracks normally due to natural eviction

So, it is a problem different from the previously mentioned one (showing more
hot memory), correct?  Has it reported from a real production?

DAMON does not flush TLB assuming real production systems would have anyway
large amount of working set that naturally flush TLB.  If there are real
production systems that this assumption doesn't apply, we may neeed to think
this again.

Anyway, the cover letter would be better to make this point clear.

> > 
> >>> x86 is not subject to this specific blindness under similar
> >>> conditions.
> >>
> >> To my understanding on x86, same issue exists.  If TLB hits, Aceessed
> >> bit is not set, and DAMON shows it as unaccessed.  Am I missing
> >> something?
> > 
> > You are entirely right, and I was wrong on this point. I re-checked the 
> > kernel source and verified that x86's ptep_test_and_clear_young() does NOT 
> > flush the TLB. Even ptep_clear_flush_young() on x86 deliberately skips the 
> > flush as a performance optimization (arch/x86/mm/pgtable.c:486-502). The 
> > same optimization architectural behavior exists on PowerPC and RISC-V.
> > 
> > Therefore, both architectures are theoretically vulnerable to this stale-TLB 
> > blind spot under identical tightly-fit workloads. Our initial assumption 
> > was biased because T1 was only conducted on ARM64. We will reproduce the 
> > T1 setup on x86 to verify the exact behavior, and I will correct this 
> > claim in the v2 cover letter. Thank you for catching this mistake.
> > 
> >> Nice!  Asier was planning to do similar work in future.  I think you
> >> could collaborate to reduce unnecessary duplicates!
> > 
> > Great to hear! We would be happy to collaborate with Asier. I'll reach
> > out to him to coordinate our efforts.
> Sure, I will be happy to cooperate.

Great, looking forward to your fantastic coworks!

> >> I'd suggest making the name simpler and consistent to DAMOS_COLLAPSE,
> >> though.  Say, DAMOS_SPLIT ?
> > 
> > Agreed. DAMOS_SPLIT is cleaner and fits the existing naming convention 
> > perfectly. Will rename in v2.
> > 
> >> So you implemented a debugfs interface?  That must be a nice approach
> >> for PoC.  But it may be difficult to be upstreamed as is.
> >>
> >> You could build a control plane that decides the exact address ranges
> >> to split, and directly feed it to DAMOS using DAMOS address filter.
> > 
> > The native perf event approach [3] aligns perfectly with our long-term 
> > Phase 2c plan, and we are highly interested in collaborating on it to 
> > eliminate the userspace daemon and debugfs bridge entirely.
> > 
> > However, since native kernel-side SPE handling is a long-term item, we 
> > will follow your pragmatic alternative suggestion for v2: use DAMOS address 
> > filters or user_input quota goals [2] to feed the split decisions from 
> > userspace cleanly. This allows us to upstream the core infrastructure 
> > (mTHP target_order for collapse and the new DAMOS_SPLIT action) first.
> > 
> >> Do you really need to khugepaged together, when you already have
> >> DAMOS_COLLAPSE, and anyway you are running DAMON for hugepage splits?
> > 
> > Excellent point. Running both concurrently on the same VMA introduces 
> > redundancy and heavy ping-pong effects. 
> > 
> > Option (b) is definitely cleaner: we will let DAMON handle both split and 
> > re-collapse decisions using its own access data. To make this robust in 
> > production environments where khugepaged is globally enabled, we will 
> > explore having the DAMOS_SPLIT path temporarily mark the target ranges 
> > (e.g., via a pseudo-VM_NOHUGEPAGE backing off mechanism) to prevent 
> > khugepaged from immediately undoing DAMON's work.

Can't you simply turn off khugepaged?

> > 
> >>> - The rbtree entry TTL (30s) and signal threshold (1/10 of peak) are
> >>>   empirical defaults subject to further tuning.
> >>
> >> I don't fully understand this part.  Could you please elaborate?
> > 
> > Since ARM SPE samples hardware accesses instruction-by-instruction, the raw 
> > data is highly statistical and noisy. 
> > 
> > The TTL (30s) defines the lifecycle of our per-folio rbtree tracking entries. 
> > Entries not updated within 30 seconds are pruned to prevent stale tracking data 
> > from corrupting split decisions after a workload phase change. 30s is selected 
> > to comfortably outlive DAMON's aggregation intervals while keeping the rbtree 
> > memory footprint tightly bounded.
> > 
> > The signal threshold (1/10 of peak) filters out the statistical sampling noise. 
> > Instead of treating any subpage with access > 0 as hot, the algorithm finds the 
> > peak access count inside the 2MB region and only marks sub-chunks with >= 1/10 
> > of that peak as genuinely hot. On Kunpeng 920, this specific threshold successfully 
> > reduced false-hot subpage classifications from ~50% to <5%. We plan to make 
> > these parameters sysfs-configurable.
> > 
> >>> - The ARM64 DAMON blind spot (WSS < L2 TLB reach) is a pre-existing
> >>>   hardware-MMU characteristic, not introduced by this series.  Setting
> >>>   nr_accesses/min=0 serves as an effective workaround for the split path.
> >>
> >> I don't fully understand this, too.  Could you please elaborate and
> >> enlighten me?
> > 
> > The blind spot creates an operational deadlock for the split infrastructure:
> >   1. WSS < TLB reach -> All THP entries stay cached in TLB.

But, is this really common on real production systems?

> >   2. DAMON's page-table scan yields `nr_accesses = 0` globally.
> >   3. A scheme requiring `nr_accesses.min = 1` never fires -> DAMOS_SPLIT is never invoked.

Why would you run DAMOS_SPLIT action for regions having 1 or higher
nr_accesses?  Having nr_accesses 1 or higher means the region was accessed, and
I asume you want to split THP that cold.  Shouldn't it rather target regions
having only 0 nr_accesses, as it means it is cold?

> >   4. THPs remain unsplit -> WSS remains within TLB reach -> Loop returns to step 1.
> > 
> > Setting `nr_accesses.min = 0` and `max = 0` breaks this deadlock.

Ah, yes, now I got it.  And this seems the right approach to me.  FYI
min_nr_accesses and max_nr_accesses are the terms we  usually use.

> > It forces 
> > DAMON to evaluate these seemingly "dead/cold" regions. Once the split handler 
> > invokes, it checks the ARM SPE telemetry (which captures data directly from the 
> > instruction pipeline, completely bypassing the MMU page-table AF limitation). 
> > If SPE reveals a sparse access heatmap, the split is executed. Once shattered into 
> > mTHP/base pages, the TLB reach drops, natural TLB misses resume, and DAMON's 
> > standard page-table tracking fully recovers.
> > 
> > 
> > Thanks again for your guidance. The action items for v2 are locked in:
> >   1. Rename DAMOS_MTHP_SPLIT -> DAMOS_SPLIT.
> >   2. Drop debugfs in favor of DAMOS address filters / control plane.
> >   3. Correct x86 AF behavior statements in the cover letter.
> >   4. Coordinate with Asier on split/collapse unification.

Sounds good.  Looking forward to the next version!  Also, if the blind spot is
a problem reported from the real production systems, please clarify it.

> >   5. Implement back-off to prevent khugepaged ping-pong under Option (b).

Again, why can't you just turn kdamond off?


Thanks,
SJ

[...]