damon_probe_hits_wsum() could overflow in weird setups. Users could set
the weight unreasonably high. They could also set the aggregation
interval unreasonably high compared to the sampling interval. Such user
setup is unlikely. Even if such setup is used, damon_has_probe_weights()
always returns false, so the overflow cannot happen. The function may be
completed in future, though. Even if the overflow happens, the
consequence is degraded monitoring results for the unreasonable setup. It
is just a trivial user experience issue.
It is still better to be prevented unless the cost is expensive. Avoid
the overflow by adding the parameter validation in the core layer
parameters validation function.
Link: https://lore.kernel.org/20260710134651.18084-11-sj@kernel.org
Signed-off-by: SJ Park <sj@kernel.org>
Cc: David Hildenbrand <david@kernel.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Liam R. Howlett <liam@infradead.org>
Cc: Lorenzo Stoakes <ljs@kernel.org>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Suren Baghdasaryan <surenb@google.com>
Cc: Vlastimil Babka <vbabka@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
When any damon_probe->weight is set, DAMON will do only probe monitoring.
probe_hits is 'unsigned char'. It could overflow when the aggregation
interval is larger than the sampling interval times 256.
damon_as_probe_weights() always return false, so such overflow cannot
happen. Even if it happens, it only degrades the monitoring results.
That said, the overflow is not intentional. It is better to be prevented
as long as the cost is not expensive. Disallow the overflow by adding a
validation logic on the core layer parameters validation function.
Link: https://lore.kernel.org/20260710134651.18084-10-sj@kernel.org
Signed-off-by: SJ Park <sj@kernel.org>
Cc: David Hildenbrand <david@kernel.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Liam R. Howlett <liam@infradead.org>
Cc: Lorenzo Stoakes <ljs@kernel.org>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Suren Baghdasaryan <surenb@google.com>
Cc: Vlastimil Babka <vbabka@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
When probe weights are set, users may want DAMON monitoring results to be
optimized for the weights. For that, regions adjustment should work for
the weighted sum of probe hits. Extend damon_merge_regions_of() to detect
if the weights are set, and work with probe hits in the case.
The weights setup detection function is incomplete. It always returns
false. It is intentional, so that more changes to completely support
weights can be made in an incremental but safe way. Until the function is
completed, all changes depend on it is no-op, so DAMON works in the
current mode.
Link: https://lore.kernel.org/20260710134651.18084-9-sj@kernel.org
Signed-off-by: SJ Park <sj@kernel.org>
Cc: David Hildenbrand <david@kernel.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Liam R. Howlett <liam@infradead.org>
Cc: Lorenzo Stoakes <ljs@kernel.org>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Suren Baghdasaryan <surenb@google.com>
Cc: Vlastimil Babka <vbabka@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Patch series "mm/damon: introduce data attributes only monitoring".
TL;DR: Introduce a way to get DAMON's best effort accuracy monitoring of
user-demanding non-access data attributes.
Background
==========
DAMON was initially designed for only access monitoring. It turned out
users want to get the information together with more data attributes. For
example, some users want to know how much of a hot memory region belongs
to huge pages or specific cgroups. Page level properties based monitoring
was introduced with commit 626ffabe67 ("mm/damon: clarify trying vs
applying on damos_stat kernel-doc comment") to fill the gap. Because it
works only at snapshot level and snapshot capturing in the mode can induce
high overhead, commit 45c49d9fd6 ("mm/damon/core: introduce struct
damon_probe") introduced data attributes monitoring.
Data attributes monitoring treats the attributes as only additional and
subordinate information. Data access monitoring is always turned on, and
regions are adjusted for best accuracy of the access information. In some
cases, users may be primarily interested in the attributes more than the
access. They might even not care about the access information at all.
Because DAMON treats data accesses as the only primary information, such
users cannot get high quality attributes information.
Design and Implementation
=========================
Introduce another way for treating data attributes as the primary
information. Add 'weight' property to each data attribute probe. When
any of the weights are set, the mode is enabled. Data access monitoring
is completely turned off in the mode. For region adjustment, the weighted
sum of probe hit counters is used instead of the nr_accesses.
Using the weights, users can specify to what attributes they are
interested in to what degree. DAMON will adjust the regions and provide
the best-effort quality monitoring that is optimized for the user demands.
Extend damon_operations for efficient use of probe hits. Update regions
merge and kdamond main logic to support the new mode. Add a new struct
field and a sysfs file for API callers and ABI users, respectively.
Test
====
On ~7 GiB memory idle system, run a simple AI-assisted program. The
program allocates and faults 2 GiB anonymous pages. Then, it does nothing
but wait until the user terminates it. Hence, the system ~2 GiB of
anonymous pages with no active accesses.
Monitor the distribution of the anonymous pages using DAMON attributes
monitoring mode, using DAMON user-space tool, damo [1].
$ sudo ./damo start --probe_filter allow anon
$ sudo ./damo report access --dont_merge_regions
heatmap: 00000000000000000000000000000000000000000000000399999995111111146666666666666666
# min/max temperatures: -2,470,000,000, -1,620,000,000, column size: 99.800 MiB
intervals: sample 5 ms aggr 100 ms (max access hz 200)
# <start> <size> <freq> <age> <probe hits>
0 4.000 KiB 79.840 MiB 0 hz 24.700 s 2
1 79.844 MiB 718.562 MiB 0 hz 24.700 s 8
2 798.406 MiB 793.148 MiB 0 hz 24.700 s 7
3 1.554 GiB 797.828 MiB 0 hz 24.700 s 7
4 2.333 GiB 794.668 MiB 0 hz 24.600 s 8
5 3.109 GiB 791.117 MiB 0 hz 24.500 s 0
6 3.882 GiB 785.312 MiB 0 hz 24 s 2
7 4.649 GiB 787.867 MiB 0 hz 16.200 s 6
8 5.418 GiB 784.477 MiB 0 hz 23.300 s 6
9 6.184 GiB 783.820 MiB 0 hz 18.200 s 9
10 6.950 GiB 797.730 MiB 0 hz 18.900 s 7
11 7.729 GiB 69.625 MiB 0 hz 18.900 s 0
memory bw estimate: 0 B per second
total size: 7.797 GiB
record DAMON intervals: sample 5 ms, aggr 100 ms
Note that the line after the line starting with "intervals:" is not
provided by the current version of 'damo'. I manually added the legends
line for easier understanding of these results.
Each of the 12 lines after the legend line shows the DAMON-found regions.
Each line shows 1) index of the region, 2) start address of the region, 3)
size of the region, 4) access frequency of the region, 5) age (how long
the access frequency on the region was kept) of the region, and finally 6)
the probe hit count.
Because data access is the primary information that adjusts region for,
and there is only nearly zero access on the system, regions are naively
adjusted with the same size. Still <probe hits> show different
distribution of the anonymous pages, but it is obviously very rough
information.
Switch to the attributes only mode and show how it changes the picture:
$ sudo ./damo tune --probe_filter allow anon --probe_weight 100
$ sudo ./damo report access --dont_merge_regions
heatmap: 88888888888888888889888999999889999999000004888888888888889999988888898888888888
# min/max temperatures: -4,430,000,000, 0, column size: 99.800 MiB
intervals: sample 5 ms aggr 100 ms (max access hz 200)
# <start> <size> <freq> <age> <probe hits>
0 4.000 KiB 60.445 MiB 0 hz 700 ms 0
1 60.449 MiB 1.363 MiB 0 hz 600 ms 18
2 61.812 MiB 144.000 KiB 0 hz 0 ns 1
3 61.953 MiB 1.922 MiB 0 hz 2.400 s 19
4 63.875 MiB 12.133 MiB 0 hz 200 ms 0
[...]
500 5.132 GiB 8.000 KiB 0 hz 2 m 15.800 s 20
501 5.132 GiB 8.000 KiB 0 hz 2 m 16.200 s 0
502 5.132 GiB 16.000 KiB 0 hz 2 m 16.900 s 20
503 5.132 GiB 24.000 KiB 0 hz 2 m 14.200 s 0
504 5.132 GiB 8.000 KiB 0 hz 2 m 14.900 s 20
[...]
923 7.534 GiB 126.637 MiB 0 hz 0 ns 6
924 7.658 GiB 252.000 KiB 0 hz 54.800 s 20
925 7.658 GiB 142.242 MiB 0 hz 300 ms 0
memory bw estimate: 0 B per second
total size: 7.797 GiB
record DAMON intervals: sample 5 ms, aggr 100 ms
As expected, regions are adjusted to provide the best accurate picture for
the anonymous pages distribution (<probe hits>). The region 0 (60.445 MiB
memory from the address 4.000 KiB) has nearly zero anonymous pages. The
region 1 (1.363 MiB memory from the address 60.449 MiB) is nearly full
with anonymous pages. Region 500 (8 KiB memory from the address 5.132
GiB) is certainly two anonymous pages.
Future Work
===========
Attributes only monitoring disables access monitoring. We will enable
that in future, by extending the supported attributes to include data
accesses. This patch series, and the future work are parts of the ongoing
project [2] for extending DAMON. The project aims to extend DAMON with
primitives other than page table accessed bits such as AMD IBS, Intel
PEBS, and Arm SPE, to provide more powerful and detailed information like
per-CPUs/threads/reads/writes monitoring.
Patches Sequence
================
Patch 1 introduces damon_probe->weight for specifying the weights of each
attribute. Patches 2-6 extends apply_probe() damon_ops callback to
efficiently support the new mode. Patch 7 fixes wrong use of abs() in the
regions merge code. Patch 8 extends regions merge function to work with
probe hits in the mode. Patch 8 also introduces the function for
detecting the mode enablement but always returns false, for safe and
incremental changes. Patches 9 and 10 adds user parameters validation to
prevent theoretical overflow of probe hits and the weighted sum. Patches
11-14 incrementally update kdamond_fn() to support the mode. Patch 15
completes the mode detection function implementation, so that the new mode
really works. Patch 16 introduces a new sysfs file for ABI users.
Finally, patches 17-19 respectively updates design, usage and ABI
documents for the new feature and interfaces.
[1] https://github.com/damonitor/damo
[2] https://lore.kernel.org/20260525225208.1179-1-sj@kernel.org/
This patch (of 19):
Add a new field, weight to damon_probe struct. The field is used to
specify the degree of the API caller's interest to the data attribute of
the probe.
Link: https://lore.kernel.org/20260710134651.18084-1-sj@kernel.org
Link: https://lore.kernel.org/20260710134651.18084-2-sj@kernel.org
Link: https://github.com/damonitor/damo [1]
Link: https://lore.kernel.org/20260525225208.1179-1-sj@kernel.org/ [2]
Signed-off-by: SJ Park <sj@kernel.org>
Cc: David Hildenbrand <david@kernel.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Liam R. Howlett <liam@infradead.org>
Cc: Lorenzo Stoakes <ljs@kernel.org>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Suren Baghdasaryan <surenb@google.com>
Cc: Vlastimil Babka <vbabka@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Since commit f002882ca3 ("mm: merge folio_is_secretmem() and
folio_fast_pin_allowed() into gup_fast_folio_allowed()"),
gup_fast_folio_allowed() falls back to the slow path for any order-0 folio
with a NULL mapping when CONFIG_SECRETMEM=y. This causes a performance
regression for drivers that allocate pages with alloc_page() and insert
them into VMAs via vm_insert_page(). These pages legitimately have a NULL
folio->mapping, but they cannot be secretmem pages.
Secretmem pages are always added to the secretmem inode's page cache via
filemap_add_folio(), which sets folio->mapping to the inode's i_mapping.
A folio with a NULL mapping can never be a secretmem folio. The
NULL-mapping check was intended to handle truncated file-backed pages (a
reject_file_backed concern), not secretmem detection.
When only check_secretmem is true (and reject_file_backed is false), a
NULL mapping is sufficient to prove the folio is not secretmem, so the
fast path can proceed.
Link: https://lore.kernel.org/20260708005745.164928-1-jhubbard@nvidia.com
Fixes: f002882ca3 ("mm: merge folio_is_secretmem() and folio_fast_pin_allowed() into gup_fast_folio_allowed()")
Signed-off-by: John Hubbard <jhubbard@nvidia.com>
Tested-by: Sourab Gupta <sougupta@nvidia.com>
Acked-by: David Hildenbrand (Arm) <david@kernel.org>
Cc: Alistair Popple <apopple@nvidia.com>
Cc: Balbir Singh <balbirs@nvidia.com>
Cc: Zi Yan <ziy@nvidia.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
The KSM NUMA merge test allocates identical pages on different NUMA nodes
and verifies KSM behavior with merge_across_nodes enabled and disabled.
On systems with memoryless NUMA nodes, for example:
#numactl -H
available: 2 nodes (0,4)
.....
node 0 cpus: 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
node 0 size: 14825 MB
node 0 free: 1382 MB
node 4 cpus:
node 4 size: 0 MB
node 4 free: 0 MB
the test may attempt to allocate memory on a node without memory, causing
numa_alloc_onnode() to fail and resulting in a spurious test failure.
The test currently checks numa_num_configured_nodes() to determine whether
sufficient NUMA nodes are available. However, configured nodes do not
necessarily have memory.
Reuse the existing get_first_mem_node() and get_next_mem_node() helpers to
locate NUMA nodes that actually contain memory, and skip the test when
fewer than two such nodes are available.
Before patch:
---------------------------
running ./ksm_tests -N -m 1
---------------------------
mbind: Invalid argument
ok 1 KSM NUMA merging
Totals: pass:1 fail:0 xfail:0 xpass:0 skip:0 error:0
[PASS]
ok 1 ksm_tests -N -m 1
---------------------------
running ./ksm_tests -N -m 0
---------------------------
mbind: Invalid argument
not ok 1 KSM NUMA merging
Totals: pass:0 fail:1 xfail:0 xpass:0 skip:0 error:0
[FAIL]
not ok 2 ksm_tests -N -m 0 # exit=1
After patch:
---------------------------
running ./ksm_tests -N -m 1
---------------------------
At least 2 NUMA nodes with memory must be available
ok 1
SKIP KSM NUMA merging
Totals: pass:0 fail:0 xfail:0 xpass:0 skip:1 error:0
[PASS]
ok 1 ksm_tests -N -m 1
---------------------------
running ./ksm_tests -N -m 0
---------------------------
At least 2 NUMA nodes with memory must be available
ok 1
SKIP KSM NUMA merging
Totals: pass:0 fail:0 xfail:0 xpass:0 skip:1 error:0
[PASS]
ok 2 ksm_tests -N -m 0
Link: https://lore.kernel.org/78a3b0e3fb94004c0710872c5bab6f7381b7d63c.1783446924.git.sayalip@linux.ibm.com
Fixes: e3820ab252 ("selftest/vm: fix ksm selftest to run with different NUMA topologies")
Co-developed-by: David Hildenbrand (Arm) <david@kernel.org>
Signed-off-by: David Hildenbrand (Arm) <david@kernel.org>
Signed-off-by: Sayali Patil <sayalip@linux.ibm.com>
Cc: Dev Jain <dev.jain@arm.com>
Cc: Liam Howlett <liam@infradead.org>
Cc: Miaohe Lin <linmiaohe@huawei.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Oscar Salvador <osalvador@suse.de>
Cc: "Ritesh Harjani (IBM)" <ritesh.list@gmail.com>
Cc: Shuah Khan <shuah@kernel.org>
Cc: Zi Yan <ziy@nvidia.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Patch series "selftests/mm: avoid false failures in hugetlb and KSM
tests", v3.
This series fixes issues in the hugetlb and KSM MM selftest categories
that can report failures when the prerequisites for the tests are not
satisfied.
Patch 1 updates the hugetlb selftest helpers to handle -EINVAL when
attempting to configure gigantic HugeTLB pages via nr_hugepages. PowerPC
hash MMU pSeries systems expose gigantic hugepage sizes but do not allow
runtime allocation of such pages, causing the sysfs write to fail. Handle
this case gracefully and continue running the test instead of aborting.
Patch 2 fixes the KSM NUMA merge test on systems with memoryless NUMA
nodes. The test currently relies on the number of configured NUMA nodes
and may attempt allocations on nodes that have no memory, resulting in
spurious failures. Use the existing helpers to identify NUMA nodes that
contain memory and skip the test when fewer than two such nodes are
available.
Patch 3 fixes a pre-existing operator precedence issue in ksm_tests, where
a ternary expression combined with logical OR operators could be evaluated
differently than intended. Added parentheses to ensure the correct
evaluation order.
These changes improve handling of unsupported test configurations and
unmet test prerequisites, avoiding spurious failures.
This patch (of 3):
Some MM selftests attempt to configure the amount of HugeTLB pages of
different sizes by writing to nr_hugepages.
PowerPC hash MMU pSeries systems advertise gigantic hugepage sizes but do
not support runtime allocation of such pages, writes to the corresponding
nr_hugepages file fail with -EINVAL. This causes the test to bail out
even though the failure is due to a platform limitation rather than the
functionality being tested.
Ignore -EINVAL when configuring nr_hugepages so that tests continue to run
on systems where gigantic hugepage allocation is unsupported.
Before patch:
-------------------------
running ./hugetlb-madvise
-------------------------
TAP version 13
1..1
[INFO] detected hugetlb page size: 16777216 KiB
[INFO] detected hugetlb page size: 16384 KiB
ok 1 MADV_DONTNEED and MADV_REMOVE on hugetlb
Totals: pass:1 fail:0 xfail:0 xpass:0 skip:0 error:0
Bail out! /sys/kernel/mm/hugepages/hugepages-16777216kB/nr_hugepages
write(0) failed: Invalid argument
Totals: pass:0 fail:0 xfail:0 xpass:0 skip:0 error:0
[FAIL]
After patch:
-------------------------
running ./hugetlb-madvise
-------------------------
TAP version 13
1..1
[INFO] detected hugetlb page size: 16777216 KiB
[INFO] detected hugetlb page size: 16384 KiB
ok 1 MADV_DONTNEED and MADV_REMOVE on hugetlb
Totals: pass:1 fail:0 xfail:0 xpass:0 skip:0 error:0
[PASS]
Link: https://lore.kernel.org/cover.1783446924.git.sayalip@linux.ibm.com
Link: https://lore.kernel.org/2e3b585cbb30b2fc495dcd49d75de6f6da61861c.1783446924.git.sayalip@linux.ibm.com
Fixes: 27477b28b7 ("selftests/mm: hugepage_settings: add APIs to get and set nr_hugepages")
Co-developed-by: David Hildenbrand (Arm) <david@kernel.org>
Signed-off-by: David Hildenbrand (Arm) <david@kernel.org>
Signed-off-by: Sayali Patil <sayalip@linux.ibm.com>
Cc: Dev Jain <dev.jain@arm.com>
Cc: Liam Howlett <liam@infradead.org>
Cc: Miaohe Lin <linmiaohe@huawei.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Oscar Salvador <osalvador@suse.de>
Cc: "Ritesh Harjani (IBM)" <ritesh.list@gmail.com>
Cc: Shuah Khan <shuah@kernel.org>
Cc: Zi Yan <ziy@nvidia.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Patch series "KSM: use linear_page_index in collect_procs_ksm()", v2.
In collect_procs_ksm() which is used to collect processes when the error
hit an ksm page, there is the same issue with rmap_walk_ksm (see the
previous discussion at [1]). So we apply the similar logic changes to the
collect_procs_ksm().
The patch [1/2] move the initializaion of addr from the position inside
loop to the position before the loop, since the variable will not change
in the loop.
The patch [2/2] optimize collect_procs_ksm by passing a suitable page
offset range to the anon_vma_interval_tree_foreach loop to reduce
ineffective checks.
This patch (of 2):
Similar to 318d87b8fa ("ksm: initialize the addr only once in
rmap_walk_ksm"), only initialize the addr once in rmap_walk_ksm because
the addr variable doesn't change across iterations.
Link: https://lore.kernel.org/20260709173212190rZdwynySRyLr9EtPuXBRU@zte.com.cn
Link: https://lore.kernel.org/all/20260703162253688u8Str9eFLR8TGCmo7nIOF@zte.com.cn/ [1]
Signed-off-by: xu xin <xu.xin16@zte.com.cn>
Acked-by: David Hildenbrand (Arm) <david@kernel.org>
Cc: Chengming Zhou <chengming.zhou@linux.dev>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
User impact / Why this matters to Linux users
=============================================
When a system runs with KSM enabled and memory becomes tight, KSM pages
may be swapped out or migrated. The kernel then performs a reverse map
walk by rmap_walk_ksm to locate all page table entries that reference
these pages. If A large number of unrelated VMAs can attach to a single
anon_vma related with this KSM page, then rmap_walk might be severe
performance bottleneck. In our embedded test environment, we observed
~20,000 VMAs sharing one anon_vma without any fork purely from VMA
splits
which cause 200~700ms duration of rmap_walk_ksm.
When one of those VMAs mapped a KSM page, then this KSM page's rmapping
will become bottleneck with hold its anon_vma lock for a long time. The
anon_vma lock is not only used by KSM; it is a core lock protecting the
VMA interval tree and is acquired by many critical memory operations:
' Page faults: do_anonymous_page(), do_wp_page() (during COW)
' Memory reclaim: try_to_unmap()
' Page migration & compaction: migrate_pages(), compact_zone()
' mlock / munlock: mlock_fixup()
' Process exit: exit_mmap() (tearing down VMAs)
' Cgroup memory accounting: mem_cgroup_move_charge()
If one thread holds the anon_vma lock for hundreds of milliseconds
because of an inefficient KSM rmap walk, any other thread that
tries to acquire the same lock (e.g., an application taking a page
fault, kswapd reclaiming pages, or a migration thread) will block.
This leads to stalled application threads, increased latency
spikes, and in extreme cases container timeouts or watchdog
triggers.
This patch reduces the worst-case anon_vma lock hold time during
ksm_rmap_walk from >500 ms to <1 ms, thereby almost eliminating
this source of lock contention and improving system responsiveness
under memory pressure.
Real-world examples:
====================
- JVM / Go runtime: These use mmap for heap regions and later call
mprotect(PROT_NONE) for garbage collection barriers or guard pages,
splitting the original VMA into thousands of small pieces over time.
- Database engines (MySQL, PostgreSQL): Large shared memory buffers
or anonymous mappings are managed with madvise(MADV_DONTNEED) to
release specific pages, which also splits VMAs.
Root Cause
==========
Through local debugging trace analysis, we found that most of the
latency of rmap_walk_ksm occurs within anon_vma_interval_tree_foreach,
leading to an excessively long hold time on the anon_vma lock (even
reaching 500ms or more), which in turn causes upper-layer applications
(waiting for the anon_vma lock) to be blocked for extended periods.
Further investigation revealed that 99.9% of iterations inside the
anon_vma_interval_tree_foreach loop are skipped due to the first check
"if (addr < vma->vm_start || addr >= vma->vm_end)), indicating that a
large number of loop iterations are ineffective. This inefficiency
arises because the start page index and the end page index parameters
passed to anon_vma_interval_tree_foreach span the entire address space
from 0 to ULONG_MAX, resulting in very poor loop efficiency.
Solution
========
We cannot rely solely on anon_vma to locate all PTEs mapping this page
but also need to have the original page's linear_page_index. Since the
implementation of anon_vma_interval_tree_foreach it essentially
iterates to find a suitable VMA such that the provided page index
falls within the candidate's vm_pgoff range.
vm_pgoff <= original linear page offset <= (vm_pgoff + vma_pages(v) - 1)
Fortunately, an earlier commit introduced the linear_page_index to struct
ksm_rmap_item, allowing for optimizing the RMAP walk.
Test results
============
A rmap testbench can be obtained with two Out-Of-Tree patches at [1][2].
After applying the OOT patches and building rmap_benchmark from:
tools/testing/rmap/rmap_benchmark.c, we can start the performance test.
The testing result in QEMU is shown as follows:
KSM rmapping Maximum duration Average duration
Before: 705.12 ms (705119858 ns) 532.04 ms (532041586 ns)
After: 1.67 ms (1665917 ns) 1.44 ms (1443784 ns)
The benchmark numbers are realistic, since we observed ~20,000 VMAs
sharing one anon_vma on a production system running a Java application
with KSM enabled. The lock hold time before the patch was measured at
228ms (max) during rmap walks triggered by memory compaction and page
migration. The benchmark reproduces that VMA count and lockhold
behavior in a controlled environment.
Link: https://lore.kernel.org/20260703162510242nxmjbcLy5ccp1dbZSK3EU@zte.com.cn
Link: https://lore.kernel.org/all/202605301703094695zmVgcSC27BNR0rH0N8_x@zte.com.cn [1]
Link: https://lore.kernel.org/all/20260530170404509QpJmBtpSjn3uQHeVKA2iA@zte.com.cn/ [2]
Co-developed-by: Wang Yaxin <wang.yaxin@zte.com.cn>
Signed-off-by: Wang Yaxin <wang.yaxin@zte.com.cn>
Signed-off-by: xu xin <xu.xin16@zte.com.cn>
Acked-by: David Hildenbrand (Arm) <david@kernel.org>
Cc: Chengming Zhou <chengming.zhou@linux.dev>
Cc: Hugh Dickins <hughd@google.com>
Cc: "Liam R. Howlett" <liam@infradead.org>
Cc: Lorenzo Stoakes <ljs@kernel.org>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Suren Baghdasaryan <surenb@google.com>
Cc: Vlastimil Babka <vbabka@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Patch series "KSM: performance optimizations for rmap_walk_ksm", v11.
This series fixes a severe KSM reverse-mapping performance problem that
can freeze applications for hundreds of milliseconds under memory pressure
especially when a lot of unrelated VMAs sharing a single anon_vma.
Two key highlights:
1. Lock hold time drops from >500ms to <2ms
- In our benchmark (20,000 VMAs sharing an anon_vma), worst-case
anon_vma lock hold time during KSM rmap walk went from 705ms
down to 1.67ms (max) and 1.44ms (avg).
2. Real user impact
- The anon_vma lock is also acquired by page faults, reclaim,
migration, compaction, mlock, exit_mmap, and cgroup accounting.
- A long hold due to inefficient rmap walks stalls application
threads, causing latency spikes, reduced throughput, or even
container timeouts.
- The problem occurs even without fork() – VMA splitting (e.g.,
via mprotect or madvise over time) can create tens of thousands
of VMAs all attached to the same anon_vma.
Real-world examples:
- JVM / Go runtime: These use mmap for heap regions and later call
mprotect(PROT_NONE) for garbage collection barriers or guard pages,
splitting the original VMA into thousands of small pieces over time.
- Database engines (MySQL, PostgreSQL): Large shared memory buffers or
anonymous mappings are managed with madvise(MADV_DONTNEED) to release
specific pages, which also splits VMAs.
Why the benchmark numbers are realistic: We observed ~20,000 VMAs sharing
one anon_vma on a production system running a Java application with KSM
enabled. The lock hold time before the patch was measured at 228 ms
(max) during rmap walks triggered by memory compaction and page migration.
The benchmark reproduces that VMA count and lock‑hold behavior in a
controlled environment.
For systems that do not have thousands of VMAs per anon_vma, the patch
adds negligible overhead (a single pgoff comparison). For systems that do
suffer from this issue, the improvement is dramatic: 1) Worst‑case
anon_vma lock hold time drops from hundreds of milliseconds to under
2 ms.2)This directly reduces blocking of parallel operations that need
the same lock – page faults, reclaim, migration, compaction, mlock, and
exit_mmap.
End‑users will see lower tail latency (fewer application stalls), higher
throughput under memory pressure, and no more spurious lockup warnings or
container timeouts caused by excessive lock hold times.
In short: workloads that do not hit this pathological pattern are
unaffected; those that do will see a 100x to 500x reduction in lock hold
times, which translates directly into a more responsive system.
This patch (of 3):
As preparation for KSM rmap optimizations, let's track the original
linear_page_index() of a de-duplicated page in its ksm_rmap_item, so we
can efficiently search for the page in an address space, avoiding scanning
the entire address space. This was previously discussed in [1, 2].
To avoid growing ksm_rmap_item, let's squeeze it into the existing
structure by overlying some members (oldchecksum, age, remaining_skips)
that are only relevant while on the unstable tree. The new entry will
only be relevant for entries in the stable tree.
However, as the age information is read by should_skip_rmap_item() with
the smart-scanning approach even while we have an entry in the stable
tree, but the page changes (no longer a KSM page, for example due to COW),
we have to change the handling there a bit.
We'll calculate the linear page index in try_to_merge_with_ksm_page(),
when adding it to the stable tree, and reset the index (to reset overlayed
data) when removing an item from the stable tree -- in
remove_rmap_item_from_tree(), remove_node_from_stable_tree() and
break_cow().
To be specially clarified, the reason for resetting the stored index at
break_cow() is:
- When a page successfully becomes a KSM page (i.e., after
stable_tree_append() sets STABLE_FLAG), both anon_vma and the index are
stored and remain valid.
- However, during the merging process, there are several failure paths
where we already prepared an rmap item to be added to the stable tree,
but must revert that as some part of the merge process failed. Examples
include:
1 The second call to try_to_merge_with_ksm_page() fails in
try_to_merge_two_pages().
2 stable_tree_insert() fails in cmp_and_merge_page().
In such cases, break_cow() is invoked to break the COW mapping and
discard the KSM state.
Currently, break_cow() already contains a
put_anon_vma(rmap_item->anon_vma) to release the reference taken during
the aborted merge. Because the index is logically paired with anon_vma
(both are only meaningful when the rmap_item is in a stable state), it
must also be cleared (or reset) in break_cow() to avoid leaving stale
linear_page_index values that could confuse subsequent rmap walks or
scanning logic.
Link: https://lore.kernel.org/20260703162253688u8Str9eFLR8TGCmo7nIOF@zte.com.cn
Link: https://lore.kernel.org/20260703162357853iIa-RP7if9hRlAIuTh5La@zte.com.cn
Link: https://lore.kernel.org/all/adTPQSb-qSSHviJN@lucifer/ [1]
Link: https://lore.kernel.org/all/202604091806051535BJWZ_FTtdIm3Snk24ei_@zte.com.cn/ [2]
Signed-off-by: xu xin <xu.xin16@zte.com.cn>
Acked-by: David Hildenbrand (Arm) <david@kernel.org>
Cc: Chengming Zhou <chengming.zhou@linux.dev>
Cc: Hugh Dickins <hughd@google.com>
Cc: "Liam R. Howlett" <liam@infradead.org>
Cc: Lorenzo Stoakes <ljs@kernel.org>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Suren Baghdasaryan <surenb@google.com>
Cc: Vlastimil Babka <vbabka@kernel.org>
Cc: Wang Yaxin <wang.yaxin@zte.com.cn>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
If damon_update_monitoring_result() is called at the end of the
aggregation interval, probe_hits is not reset. That's because the value
will be exposed to the user via damon_region_aggregated trace event.
Meanwhile, damon_probe_hits_mvsum() can be called in this state. Due to
its logic, it will return a value that is incorrectly high. This could
happen if the user requested DAMOS schemes applied regions sysfs files
update exactly in the time sequence.
The impact is minor, but better to avoid. Check the timing and simply
return the fully aggregated last_probe_hits, like
damon_nr_accesses_mvsum() also does. It is not 100% accurate since it is
the last interval's aggregation. But better than the value that is
completely reset.
Link: https://lore.kernel.org/20260708135359.122587-8-sj@kernel.org
Signed-off-by: SJ Park <sj@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Users can update DAMON parameters at runtime. If the samples and/or
aggregation intervals are updated in this way, monitoring results
depending on the intervals should also be updated for a more accurate
snapshot. The age and nr_accesses are properly updated, while probe_hits
are not updated in the way. Do the update.
Link: https://lore.kernel.org/20260708135359.122587-7-sj@kernel.org
Signed-off-by: SJ Park <sj@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Patch series "mm/damon: update probe hits for runtime parameter commits".
DAMON users can update DAMON parameters such as sampling and aggregation
intervals at runtime. For such changes, monitoring results that depend on
the intervals should be properly updated for better accuracy. For
example, the access frequency counter (nr_accesses) is updated. The data
attribute monitoring counter (probe_hits) is not being updated, though.
Do the updates for new parameters.
Patch 1 removes obsolete comments and test code for a function that this
series will touch. Patches 2-5 rename functions that are being used for
nr_accesses update, to be able to be used for probe_hits without
confusion. Patch 6 does the probe_hits update. Patch 7 update
damon_probe_hits_mvsum() to cover a corner case from the update for better
accuracy.
This patch (of 7):
The comments on damon_nr_accesses_to_accesses_bp() and its unit test warn
it can trigger division-by-zero when the aggregation interval is zero.
Commit 35d4a3cf70 ("mm/damon/ops-common: handle extreme intervals in
damon_hot_score()") modified damon_max_nr_accesses() to always return
non-zero. Hence no division-by-zero of the note can happen. Remove the
obsolete comment on the function. The test code was written to test the
division-by-zero case, which cannot happen anymore. Having it makes no
sense. Entirely remove the test code and its comment.
Link: https://lore.kernel.org/20260708135359.122587-1-sj@kernel.org
Link: https://lore.kernel.org/20260708135359.122587-2-sj@kernel.org
Signed-off-by: SJ Park <sj@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
When vma_iter_prealloc() fails in do_mmap(), it jumps to error_just_free
as a error path of this function, but there are several possible issues.
1) It jumps to error_just_free without updating ret to -ENOMEM, meaning
do_mmap() will return 0 on failure.
2) The error path unconditionally frees the region struct. Since the
region was already added to the global nommu_region_tree via
add_nommu_region(), leaving it makes a potential dangling pointer in
the tree and may cause a use-after-free on the next tree walk.
3) If do_mmap() finds an existing overlapping shared region, it
increments its usage, sets region to this existing pregion, and jumps
to share:
region = pregion;
result = start;
goto share;
When vma_iter_prealloc() fails and jumps to error_just_free, the
error path unconditionally frees the region:
error:
...
if (region->vm_file)
fput(region->vm_file);
kmem_cache_free(vm_region_jar, region);
This potentially leaves a dangling pointer in nommu_region_tree and
causes RB-tree corruption.
4) When establishing a new private mapping, do_mmap_private() allocates
physical pages and assigns them to region->vm_start:
base = alloc_pages_exact(total << PAGE_SHIFT, GFP_KERNEL);
...
region->vm_start = (unsigned long) base;
If we later fail at vma_iter_prealloc() and jump to error_just_free,
the region struct is freed, but the backing physical memory isn't freed
via free_page_series().
5) In the error label of do_mmap(), the vm_area_struct allocated is
freed by vm_area_free(vma) but not called after vma_close(), leaving
potential memory leak which should be handled by a custom .close
handler of vm_ops.
This commit fixes those issues by introducing new jump label,
error_vma_iter_prealloc, to correctly handle the error case of
vma_iter_prealloc(), updating ret value (1), and move the region updates
after the place that the allocation is finished (2).
Additionally, the commit removes the existing goto label, error, and
consolidates to error_just_free as existing `goto error;` code blocks
always release nommu_region_sem.
Moreover, it only frees region allocated in this request to avoid
freeing the shared, existing region shared by other processes (3), and
free physical memory when do_mmap_private() allocates (4). It also add
vma_close() before vm_area_free() to fix the potential leak (5).
Those issues are discovered by Sashiko, linked below.
Link: https://lore.kernel.org/20260708083829.576036-1-thehajime@gmail.com
Link: https://sashiko.dev/#/patchset/20260702012830.667205-1-thehajime%40gmail.com
Link: https://sashiko.dev/#/patchset/c8513ee5aa8444ec9bf6c276043c9f833016a2fa.1783304131.git.thehajime%40gmail.com
Link: https://sashiko.dev/#/patchset/20260707235137.498738-1-thehajime%40gmail.com
Fixes: b5df092264 ("mm: set up vma iterator for vma_iter_prealloc() calls")
Signed-off-by: Hajime Tazaki <thehajime@gmail.com>
Cc: Geert Uytterhoeven <geert@linux-m68k.org>
Cc: Jann Horn <jannh@google.com>
Cc: Liam R. Howlett <liam@infradead.org>
Cc: Lorenzo Stoakes <ljs@kernel.org>
Cc: Pedro Falcato <pfalcato@suse.de>
Cc: Vlastimil Babka <vbabka@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
