Patch series "mm/damon: introduce data attributes monitoring".
TL; DR
======
Extend DAMON for monitoring general data attributes other than accesses.
The short term motivation is lightweight page type (e.g., belonging
cgroup) aware monitoring. In long term, this will help extending DAMON
for multiple access events capture primitives (e.g., page faults and PMU)
and eventually pivotting DAMON to a "Data Attributes Monitoring and
Operations eNgine" in long term.
Background: High Cost of Page Level Properties Monitoring
=========================================================
DAMON is initially introduced as a Data Access MONitor. It has been
extended for not only access monitoring but also data access-aware system
operations (DAMOS). But still the monitoring part is only for data
accesses.
Data access patterns is good information, but some users need more
holistic views. Particularly, users want to show the access pattern
information together with the types of the memory. For example, users who
work for making huge pages efficiently want to know how much of
DAMON-found hot/cold regions are backed by huge pages. Users who run
multiple workloads with different cgroups want to know how much of
DAMON-found hot/cold regions belong to specific cgroups.
For the user demand, we developed a DAMOS extension for page level
properties based monitoring [1], which has landed on 6.14. Using the
feature, users can inform the page level data properties that they are
interested in, in a flexible format that uses DAMOS filters. Then, DAMON
applies the filters to each folio of the entire DAMON region and lets
users know how many bytes of memory in each DAMON region passed the given
filters.
This gives page level detailed and deterministic information to users.
But, because the operation is done at page level, the overhead is
proportional to the memory size. It was useful for test or debugging
purposes on a small number of machines. But it was obviously too heavy to
be enabled always on all machines running the real user workloads. For
real world workloads, it was recommended to use the feature with
user-space controlled sampling approaches. For example, users could do
the page level monitoring only once per hour, on randomly selected one
percent of machines of their fleet. If the runtime and the size of the
fleet is long and big enough, it should provide statistically meaningful
data.
But users are too busy to implement such controls on their own.
Data Attributes Monitoring
==========================
Extend DAMON to monitor not only data accesses, but also general data
attributes. Do the extension while keeping the main promise of DAMON, the
bounded and best-effort minimum overhead.
Allow users to specify what data attributes in addition to the data access
they want to monitor. Users can install one 'data probe' per data
attribute of their interest for this purpose. The 'data probe' should be
able to be applied to any memory, and determine if the given memory has
the appropriate data attribute. E.g., if memory of physical address 42
belongs to cgroup A. Each 'data probe' is configured with filters that
are very similar to the DAMOS filters.
When DAMON checks if each sampling address memory of each region is
accessed since the last check, it applies data probes if registered. Same
to the number of access check-positive samples accounting (nr_accesses),
it accounts the number of each data probe-positive samples in another
per-region counters array, namely 'probe_hits'. When DAMON resets
nr_accesses every aggregation interval, it resets 'probe_hits' together.
Users can read 'probe_hits' just before the values are reset. In this
way, users can know how many hot/cold memory regions have data attributes
of their interest. E.g., 30 percent of this system's hot memory is
belonging to cgroup A, and 80 percent of the cgroup A-belonging hot memory
is backed by huge pages.
Patches Sequence
================
First eight patches implement the core feature, interface and the working
support. Patch 1 introduces data probe data structure, namely
damon_probe. Patch 2 extends damon_ctx for installing data probes. Patch
3 introduces another data structure for filters of each data probe, namely
damon_filter. Patch 4 updates damon_ctx commit function to handle the
probes. Patch 5 extends damon_region for the per-region per-probe
positive samples counter, namely probe_hits. Patch 6 extends
damon_operations for applying probes on the underlying DAMON operations
implementation. Patch 7 updates kdamond_fn() to invoke the probes
applying callback. Patch 8 finally implements the probes support on paddr
ops.
Ten changes for user interface (patches 9-18) come next. Patches 9-13
implements sysfs directories and files for setting data probes, namely
probes directory, probe directory, filters directory, filter directory and
filter directory internal files, respectively. Patch 14 connects the user
inputs that are made via the sysfs files to DAMON core. Following three
patches (patches 15-17) implement sysfs directories and files for showing
the probe_hits to users, namely probes directory, probe directory and hits
files, respectively. Patch 18 introduces a new tracepoint for showing the
probe_hits via tracefs.
Patch 19 adds a selftest for the sysfs files.
Patches 20 and 21 documents the design and usage of the new feature,
respectively.
Seven additional patches (patches 22-28) for monitoring belonging memory
cgroup follow. Depending on the feedback, this part might be separated to
another series in future. Patch 22 defines the DAMON filter type for the
new attribute, namely DAMON_FILTER_TYPE_MEMCG. Patch 23 add the support
on paddr ops. Patch 24 updates the sysfs interface for setup of the
target memcg. Patch 25 move code for easy reuse of the filter target
memcg setup. Patch 26 connects the user input to the core layer.
Finally, patches 27 and 28 update the design and usage documents for the
memcg attribute monitoring support.
Discussion
==========
This allows the page properties monitoring with overhead that is low
enough to be enabled always on real world workloads. Because the sampling
time for access check is reused for data attributes check, the
upper-bounded and best-effort minimum overhead of DAMON is kept. Because
the sampling memory for access check is reused for data attributes check,
additional overhead is minimum.
Still DAMOS-based page level properties monitoring should be useful,
because it provides a deterministic page level information. When in doubt
of the sampling based information, running DAMOS-based one together and
comparing the results would be useful, for debugging and tuning.
Future Works: Mid Term
========================
This version of implementation is limiting the maximum number of data
probes to four. I will try to find a way to remove the limit in future.
I personally think it should be enough for common use cases, though, and
therefore not giving high priority at the moment.
Future Works: Long Term
=======================
There are user requests for extending DAMON with detailed access
information, for example, per-CPUs/threads/read/writes monitoring. For
that, I was working [2] on extending DAMON to use page fault events as
another access check primitives, and making the infrastructure flexible
for future use of yet another access check primitive. Actually there is
another ongoing work [3] for extending DAMON with PMU events. The
motivation of the work is reducing the overhead, though.
In my work [2], I was introducing a new interface for access sampling
primitives control. Now I think this data probe interface can be used for
that, too. That is, data access becomes just one type of data attribute.
Also, pg_idle-confirmed access, page fault-confirmed access, and PMU
event-confirmed access will be different types of data attributes.
The regions adjustment mechanism is currently working based on the access
information. That's because DAMON is designed for data access monitoring.
That is, data access information is the primary interest, and therefore
DAMON adjusts regions in a way that can best-present the information.
Once data access becomes just one of data attributes, there is no reason
to think data access that special. There might be some users not
interested in access at all but want to know the location of memory of
specific type. Data probes interface will allow doing that. Further, we
could extend the interface to let users set any data attribute as the
'primary' attribute. Then, DAMON will split and merge regions in a way
that can best-present the 'primary' attributes.
DAMOS will also be extended, to specify targets based on not only the data
access pattern, but all user-registered data attributes. From this stage,
we may be able to call DAMON as a "Data Attributes Monitoring and
Operations eNgine".
This patch (of 28):
Introduce a data structure for data attribute probe. It is just a linked
list header at this step. It will be extended in a way that it can
determine if a given memory has a specific data attribute.
Link: https://lore.kernel.org/20260518234119.97569-1-sj@kernel.org
Link: https://lore.kernel.org/20260518234119.97569-2-sj@kernel.org
Link: https://lore.kernel.org/20250106193401.109161-1-sj@kernel.org [1]
Link: https://lore.kernel.org/20251208062943.68824-1-sj@kernel.org/ [2]
Link: https://lore.kernel.org/20260423004211.7037-1-akinobu.mita@gmail.com [3]
Signed-off-by: SeongJae 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: "Masami Hiramatsu (Google)" <mhiramat@kernel.org>
Cc: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Shuah Khan <shuah@kernel.org>
Cc: Steven Rostedt <rostedt@goodmis.org>
Cc: Suren Baghdasaryan <surenb@google.com>
Cc: Vlastimil Babka <vbabka@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
By allocating one additional bit in the swap table entry's flags field
alongside the count, we can store the zeromap inline
For 64 bit systems, zeromap will store in the swap table, avoiding zeromap
allocation. It reduces the allocated memory. That is the happy path.
For certain 32-bit archs, there might not be enough bits in the swap table
to contain both PFN and flags. Therefore, conditionally let each cluster
have a zeromap field at build time, and use that instead. If the swapfile
cluster is not fully used, it will still save memory for zeromap. The
empty cluster does not allocate a zeromap. In the worst case, all cluster
are fully populated. We will use memory similar to the previous zeromap
implementation.
A few macros were moved to different headers for build time struct
definition.
[akpm@linux-foundation.org: swap_cluster_alloc_table(): remove unused local `ret]
[akpm@linux-foundation.org: fix unused label `err_free']
Link: https://lore.kernel.org/20260517-swap-table-p4-v5-12-88ae43e064c7@tencent.com
Signed-off-by: Kairui Song <kasong@tencent.com>
Acked-by: Chris Li <chrisl@kernel.org>
Reviewed-by: Youngjun Park <youngjun.park@lge.com>
Cc: Baolin Wang <baolin.wang@linux.alibaba.com>
Cc: Baoquan He <bhe@redhat.com>
Cc: Barry Song <baohua@kernel.org>
Cc: Chengming Zhou <chengming.zhou@linux.dev>
Cc: David Hildenbrand <david@kernel.org>
Cc: Hugh Dickins <hughd@google.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Kemeng Shi <shikemeng@huaweicloud.com>
Cc: Lorenzo Stoakes <ljs@kernel.org>
Cc: Muchun Song <muchun.song@linux.dev>
Cc: Nhat Pham <nphamcs@gmail.com>
Cc: Roman Gushchin <roman.gushchin@linux.dev>
Cc: Shakeel Butt <shakeel.butt@linux.dev>
Cc: Zi Yan <ziy@nvidia.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
To make it possible to allocate large folios directly in swap cache,
provide a new infrastructure helper to handle the swap cache status check,
allocation, and order fallback in the swap cache layer
The new helper replaces the existing swap_cache_alloc_folio. Based on
this, all the separate swap folio allocation that is being done by anon /
shmem before is converted to use this helper directly, unifying folio
allocation for anon, shmem, and readahead.
This slightly consolidates how allocation is synchronized, making it more
stable and less prone to errors. The slot-count and cache-conflict check
is now always performed with the cluster lock held before allocation, and
repeated under the same lock right before cache insertion. This double
check produces a stable result compared to the previous anon and shmem
mTHP allocation implementation, avoids the false-negative conflict checks
that the lockless path can return — large allocations no longer have to
be unwound because the range turned out to be occupied — and aborts
early for already-freed slots, which helps ordinary swapin and especially
readahead, with only a marginal increase in cluster-lock contention (the
lock is very lightly contended and stays local in the first place).
Hence, callers of swap_cache_alloc_folio() no longer need to check the
swap slot count or swap cache status themselves.
And now whoever first successfully allocates a folio in the swap cache
will be the one who charges it and performs the swap-in. The race window
of swapping is also reduced since the loop is much more compact.
Link: https://lore.kernel.org/20260517-swap-table-p4-v5-4-88ae43e064c7@tencent.com
Signed-off-by: Kairui Song <kasong@tencent.com>
Cc: Baolin Wang <baolin.wang@linux.alibaba.com>
Cc: Baoquan He <bhe@redhat.com>
Cc: Barry Song <baohua@kernel.org>
Cc: Chengming Zhou <chengming.zhou@linux.dev>
Cc: Chris Li <chrisl@kernel.org>
Cc: David Hildenbrand <david@kernel.org>
Cc: Hugh Dickins <hughd@google.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Kemeng Shi <shikemeng@huaweicloud.com>
Cc: Lorenzo Stoakes <ljs@kernel.org>
Cc: Muchun Song <muchun.song@linux.dev>
Cc: Nhat Pham <nphamcs@gmail.com>
Cc: Roman Gushchin <roman.gushchin@linux.dev>
Cc: Shakeel Butt <shakeel.butt@linux.dev>
Cc: Youngjun Park <youngjun.park@lge.com>
Cc: Zi Yan <ziy@nvidia.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Patch series "mm, swap: swap table phase IV: unify allocation", v5.
This series unifies the allocation and charging of anon and shmem swap in
folios, provides better synchronization, consolidates the metadata
management, hence dropping the static array and map, and improves the
performance. The static metadata overhead is now close to zero, and
workload performance is slightly improved.
For example, mounting a 1TB swap device saves about 512MB of memory:
Before:
free -m
total used free shared buff/cache available
Mem: 1464 805 346 1 382 658
Swap: 1048575 0 1048575
After:
free -m
total used free shared buff/cache available
Mem: 1464 277 899 1 356 1187
Swap: 1048575 0 1048575
Memory usage is ~512M lower, and we now have a close to 0 static overhead.
It was about 2 bytes per slot before, now roughly 0.09375 bytes per slot
(48 bytes ci info per cluster, which is 512 slots).
Performance test is also looking good, testing Redis in a 2G VM using 6G
ZRAM as swap:
valkey-server --maxmemory 2560M
redis-benchmark -r 3000000 -n 3000000 -d 1024 -c 12 -P 32 -t get
Before: 3385017.283654 RPS
After: 3433309.307292 RPS (1.42% better)
Testing with build kernel under global pressure on a 48c96t system,
limiting the total memory to 8G, using 12G ZRAM, 24 test runs, enabling
THP:
make -j96, using defconfig
Before: user time 2904.59s system time 4773.99s
After: user time 2909.38s system time 4641.55s (2.77% better)
Testing with usemem on a 32c machine using 48G brd ramdisk and 16G RAM, 12
test run:
usemem --init-time -O -y -x -n 48 1G
Before: Throughput (Sum): 6482.58 MB/s Free Latency: 371371.67us
After: Throughput (Sum): 6539.28 MB/s Free Latency: 363059.88us
Seems similar, or slightly better.
This series also reduces memory thrashing, I no longer see any: "Huh
VM_FAULT_OOM leaked out to the #PF handler. Retrying PF", it was shown
several times during stress testing before this series when under great
pressure:
Before: grep -Ri VM_FAULT_OOM <test logs> | wc -l => 18
After: grep -Ri VM_FAULT_OOM <test logs> | wc -l => 0
This patch (of 12):
Instead of trying to return the existing folio if the entry is already
cached in swap_cache_alloc_folio, simply return an error pointer if the
allocation failed, and drop the output argument that indicates what kind
of folio is actually returned.
And a proper wrapper swap_cache_read_folio that decouples and handles the
actual requirement - read in the folio, or return the already read folio
in cache. This is what async swapin and readahead actually required.
As for zswap swap out, the caller just needs to abort if the allocation
fails because the entry is gone or already cached, so removing simplifies
the return argument, making it cleaner.
No feature change.
Link: https://lore.kernel.org/20260517-swap-table-p4-v5-0-88ae43e064c7@tencent.com
Link: https://lore.kernel.org/20260517-swap-table-p4-v5-1-88ae43e064c7@tencent.com
Signed-off-by: Kairui Song <kasong@tencent.com>
Acked-by: Chris Li <chrisl@kernel.org>
Cc: Baolin Wang <baolin.wang@linux.alibaba.com>
Cc: Baoquan He <bhe@redhat.com>
Cc: Barry Song <baohua@kernel.org>
Cc: Chengming Zhou <chengming.zhou@linux.dev>
Cc: David Hildenbrand <david@kernel.org>
Cc: Hugh Dickins <hughd@google.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Kemeng Shi <shikemeng@huaweicloud.com>
Cc: Lorenzo Stoakes <ljs@kernel.org>
Cc: Muchun Song <muchun.song@linux.dev>
Cc: Nhat Pham <nphamcs@gmail.com>
Cc: Roman Gushchin <roman.gushchin@linux.dev>
Cc: Shakeel Butt <shakeel.butt@linux.dev>
Cc: Zi Yan <ziy@nvidia.com>
Cc: Youngjun Park <youngjun.park@lge.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
lookup_swap_cgroup_id() passes swap_cgroup_ctrl[type].map to
__swap_cgroup_id_lookup() without checking that the type was ever
registered via swap_cgroup_swapon(). On a swapless host every ctrl->map
is NULL, so __swap_cgroup_id_lookup() dereferences NULL + a scaled
swp_offset().
Since commit bea67dcc5e ("mm: attempt to batch free swap entries for
zap_pte_range()"), zap_pte_range() -> swap_pte_batch() calls
lookup_swap_cgroup_id() on any non-present, non-none PTE that decodes as a
real swap entry, without first validating it against swap_info[]. A
single PTE corrupted into a type-0 swap entry takes the host down at
process exit.
We hit this in production on a swapless 6.12.58 host: ~1s of
"get_swap_device: Bad swap file entry 3f800204222bb" (do_swap_page() being
correctly defensive about the same entry) followed by
BUG: unable to handle page fault for address: 000003f800204220
RIP: 0010:lookup_swap_cgroup_id+0x2b/0x60
Call Trace:
swap_pte_batch+0xbf/0x230
zap_pte_range+0x4c8/0x780
unmap_page_range+0x190/0x3e0
exit_mmap+0xd9/0x3c0
do_exit+0x20c/0x4b0
syzbot has reported the identical stack.
The source of the PTE corruption is a separate bug; this change makes the
teardown path as robust as the fault path already is. Every other caller
of lookup_swap_cgroup_id() is downstream of a get_swap_device() that has
already validated the entry, so the new branch is cold.
Link: https://lore.kernel.org/20260504-swap-cgroup-fix-7-0-v1-1-f53ff41ee553@linux.dev
Fixes: bea67dcc5e ("mm: attempt to batch free swap entries for zap_pte_range()")
Signed-off-by: Jose Fernandez (Anthropic) <jose.fernandez@linux.dev>
Reported-by: syzbot+e12bd9ca48157add237a@syzkaller.appspotmail.com
Link: https://lore.kernel.org/r/69859728.050a0220.3b3015.0033.GAE@google.com
Assisted-by: Claude:unspecified
Cc: Barry Song <baohua@kernel.org>
Cc: David Hildenbrand <david@kernel.org>
Cc: Hugh Dickins <hughd@google.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Kairui Song <ryncsn@gmail.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Muchun Song <muchun.song@linux.dev>
Cc: Roman Gushchin <roman.gushchin@linux.dev>
Cc: Shakeel Butt <shakeel.butt@linux.dev>
Cc: <stable@vger.kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
The allocator interacts with cgroups which rely on RCU. RCU does not work
everywhere, so the "any context" claim is slightly overstated here.
This should already be enforced by objtool, since this function is not
marked noinstr the x86 build should fail if you call it from a place where
RCU is not watching. But, expecting readers to make that connection for
themselves seems a bit cruel (I don't think there is even any
documentation of what noinstr means at all, let alone the connection with
RCU).
Note this is not claiming that any cgroup code called from the allocator
would actually break if this restriction was violated, it could very well
be that there's no real way for the allocator to act on a cgroup that can
disappear concurrently. But, since it's likely nobody has verified this
one way or another, better to just be safe and declare that RCU is
required. Allocating from an RCU-unsafe context seems a bit crazy anyway.
Link: https://lore.kernel.org/20260519-nolock-rcu-comment-v1-1-4a630c8794e5@google.com
Signed-off-by: Brendan Jackman <jackmanb@google.com>
Suggested-by: Junaid Shahid <junaids@google.com>
Acked-by: Harry Yoo (Oracle) <harry@kernel.org>
Acked-by: Vlastimil Babka (SUSE) <vbabka@kernel.org>
Cc: Alexei Starovoitov <ast@kernel.org>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Suren Baghdasaryan <surenb@google.com>
Cc: Zi Yan <ziy@nvidia.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
page_cache_prev_miss() is documented to return a value outside the
searched range when no gap is found. However, the no-gap-found path
returns xas.xa_index, which after a successful loop is the first index in
the range. As such, that index is misreported as a gap.
The sole caller, page_cache_sync_ra(), uses the return value to estimate
the cached run preceding a sequential read. In some cases, the buggy
return value can undercount the contiguous range by one, shrinking the
readahead window or pushing borderline requests into the small-random-read
branch.
Fix this by returning the start of the range - 1 when no hole is found.
Update page_cache_next_miss() for clarity as well.
Both helpers were previously fixed together in commit 9425c591e0 ("page
cache: fix page_cache_next/prev_miss off by one"), but the fix was
reverted because it caused a hugetlb performance regression. hugetlb no
longer uses these functions and next_miss was subsequently refixed in
commit 901a269ff3 ("filemap: fix page_cache_next_miss() when no hole
found") and commit bbcaee20e0 ("readahead: fix return value of
page_cache_next_miss() when no hole is found"), but prev_miss was not
addressed.
This was found by pointing Claude Opus 4.7 at mm/filemap.c.
Link: https://lore.kernel.org/20260512-prev_miss_fix-v2-1-4af8e5c1ae62@columbia.edu
Fixes: 0d3f929666 ("page cache: Convert hole search to XArray")
Assisted-by: Claude:claude-opus-4-7
Signed-off-by: Tal Zussman <tz2294@columbia.edu>
Reviewed-by: Jan Kara <jack@suse.cz>
Reviewed-by: Vishal Moola <vishal.moola@gmail.com>
Cc: Matthew Wilcox (Oracle) <willy@infradead.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
SeongJae Park