There'd be no work for memcg-aware shrinkers when kernel memory is not
accounted per cgroup, so we can skip allocating per memcg shrinker data.
This saves some memory, avoids holding shrinker_mutex with O(nr_memcgs)
and saves work in shrink_slab_memcg().
Then there are SHRINKER_NONSLAB shrinkers which handle non-kernel memory
so nokmem should not disable their per-memcg behavior. Such shrinkers
(e.g. deferred_split_shrinker) still need access to per-memcg data (see
also commit 0a432dcbeb ("mm: shrinker: make shrinker not depend on
memcg kmem")).
The savings with this patch come on container hosts that create many
superblocks (each with own shrinker) but tracking and processing per-memcg
data is pointless with nokmem (shrink_slab_memcg() is partially guarded
with !memcg_kmem_online already).
The patch uses "boottime" predicate mem_cgroup_kmem_disabled() (not
memcg_kmem_online()) to avoid mistakenly un-MEMCG_AWARE-ing shrinkers
registered before first non-root memcg is mkdir'd.
[mkoutny@suse.com: update comment, per Qi Zheng]
Link: https://lkml.kernel.org/r/20260309-cgroup-ml-nokmem-shrinker-v2-1-3e7a7eefb6c9@suse.com
Link: https://lkml.kernel.org/r/20260225-cgroup-ml-nokmem-shrinker-v1-1-d703899bdda4@suse.com
Signed-off-by: Michal Koutný <mkoutny@suse.com>
Reviewed-by: Roman Gushchin <roman.gushchin@linux.dev>
Acked-by: Qi Zheng <zhengqi.arch@bytedance.com>
Reviewed-by: Muchun Song <muchun.song@linux.dev>
Cc: Dave Chinner <david@fromorbit.com>
Cc: Jan Kara <jack@suse.cz>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Recently, I have been actively contributing to the swap subsystem through
works such as swap-tier patches and flash friendly swap proposal. During
this process, I have consistently reviewed swap table code, some other
patches and fixed several bugs.
As I am already CC'd on many patches and maintaining active interest in
ongoing developments, I would like to officially add myself as a reviewer.
I am committed to contributing to the kernel community with greater
responsibility.
Link: https://lkml.kernel.org/r/20260226010739.3773838-1-youngjun.park@lge.com
Signed-off-by: Youngjun Park <youngjun.park@lge.com>
Reviewed-by: Barry Song <baohua@kernel.org>
Acked-by: Kairui Song <kasong@tencent.com>
Acked-by: Chris Li <chrisl@kernel.org>
Acked-by: Baoquan He <bhe@redhat.com>
Cc: Kemeng Shi <shikemeng@huaweicloud.com>
Cc: Nhat Pham <nphamcs@gmail.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Patch series "Make KHO Stateless", v9.
This series transitions KHO from an xarray-based metadata tracking system
with serialization to a radix tree data structure that can be passed
directly to the next kernel.
The key motivations for this change are to:
- Eliminate the need for data serialization before kexec.
- Remove the KHO finalize state.
- Pass preservation metadata more directly to the next kernel via the FDT.
The new approach uses a radix tree to mark preserved pages. A page's
physical address and its order are encoded into a single value. The tree
is composed of multiple levels of page-sized tables, with leaf nodes being
bitmaps where each set bit represents a preserved page. The physical
address of the radix tree's root is passed in the FDT, allowing the next
kernel to reconstruct the preserved memory map.
This series is broken down into the following patches:
1. kho: Adopt radix tree for preserved memory tracking:
Replaces the xarray-based tracker with the new radix tree
implementation and increments the ABI version.
2. kho: Remove finalize state and clients:
Removes the now-obsolete kho_finalize() function and its usage
from client code and debugfs.
This patch (of 2):
Introduce a radix tree implementation for tracking preserved memory pages
and switch the KHO memory tracking mechanism to use it. This lays the
groundwork for a stateless KHO implementation that eliminates the need for
serialization and the associated "finalize" state.
This patch introduces the core radix tree data structures and constants to
the KHO ABI. It adds the radix tree node and leaf structures, along with
documentation for the radix tree key encoding scheme that combines a
page's physical address and order.
To support broader use by other kernel subsystems, such as hugetlb
preservation, the core radix tree manipulation functions are exported as a
public API.
The xarray-based memory tracking is replaced with this new radix tree
implementation. The core KHO preservation and unpreservation functions
are wired up to use the radix tree helpers. On boot, the second kernel
restores the preserved memory map by walking the radix tree whose root
physical address is passed via the FDT.
The ABI `compatible` version is bumped to "kho-v2" to reflect the
structural changes in the preserved memory map and sub-FDT property names.
This includes renaming "fdt" to "preserved-data" to better reflect that
preserved state may use formats other than FDT.
[ran.xiaokai@zte.com.cn: fix child node parsing for debugfs in/sub_fdts]
Link: https://lkml.kernel.org/r/20260309033530.244508-1-ranxiaokai627@163.com
Link: https://lkml.kernel.org/r/20260206021428.3386442-1-jasonmiu@google.com
Link: https://lkml.kernel.org/r/20260206021428.3386442-2-jasonmiu@google.com
Signed-off-by: Jason Miu <jasonmiu@google.com>
Signed-off-by: Ran Xiaokai <ran.xiaokai@zte.com.cn>
Reviewed-by: Pasha Tatashin <pasha.tatashin@soleen.com>
Reviewed-by: Mike Rapoport (Microsoft) <rppt@kernel.org>
Cc: Alexander Graf <graf@amazon.com>
Cc: Baoquan He <bhe@redhat.com>
Cc: Changyuan Lyu <changyuanl@google.com>
Cc: David Matlack <dmatlack@google.com>
Cc: David Rientjes <rientjes@google.com>
Cc: Jason Gunthorpe <jgg@nvidia.com>
Cc: Pratyush Yadav <pratyush@kernel.org>
Cc: Ran Xiaokai <ran.xiaokai@zte.com.cn>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Commit 8f1081892d ("kho: simplify page initialization in
kho_restore_page()") cleaned up the page initialization logic by moving
the folio and 0-order-page paths into separate functions. It missed
moving the alloc tag initialization.
Do it now to keep the two paths cleanly separated. While at it, touch up
the comments to be a tiny bit shorter (mainly so it doesn't end up
splitting into a multiline comment). This is purely a cosmetic change and
there should be no change in behaviour.
Link: https://lkml.kernel.org/r/20260213085914.2778107-1-pratyush@kernel.org
Signed-off-by: Pratyush Yadav (Google) <pratyush@kernel.org>
Cc: Alexander Graf <graf@amazon.com>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Pasha Tatashin <pasha.tatashin@soleen.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
For example, create three task: hot1 -> cold -> hot2. After all three
task are created, each allocate memory 128MB. the hot1/hot2 task
continuously access 128 MB memory, while the cold task only accesses its
memory briefly and then call madvise(MADV_FREE). However, khugepaged
still prioritizes scanning the cold task and only scans the hot2 task
after completing the scan of the cold task.
All folios in VM_DROPPABLE are lazyfree, Collapsing maintains that
property, so we can just collapse and memory pressure in the future will
free it up. In contrast, collapsing in !VM_DROPPABLE does not maintain
that property, the collapsed folio will not be lazyfree and memory
pressure in the future will not be able to free it up.
So if the user has explicitly informed us via MADV_FREE that this memory
will be freed, and this vma does not have VM_DROPPABLE flags, it is
appropriate for khugepaged to skip it only, thereby avoiding unnecessary
scan and collapse operations to reducing CPU wastage.
Here are the performance test results:
(Throughput bigger is better, other smaller is better)
Testing on x86_64 machine:
| task hot2 | without patch | with patch | delta |
|---------------------|---------------|---------------|---------|
| total accesses time | 3.14 sec | 2.93 sec | -6.69% |
| cycles per access | 4.96 | 2.21 | -55.44% |
| Throughput | 104.38 M/sec | 111.89 M/sec | +7.19% |
| dTLB-load-misses | 284814532 | 69597236 | -75.56% |
Testing on qemu-system-x86_64 -enable-kvm:
| task hot2 | without patch | with patch | delta |
|---------------------|---------------|---------------|---------|
| total accesses time | 3.35 sec | 2.96 sec | -11.64% |
| cycles per access | 7.29 | 2.07 | -71.60% |
| Throughput | 97.67 M/sec | 110.77 M/sec | +13.41% |
| dTLB-load-misses | 241600871 | 3216108 | -98.67% |
[vernon2gm@gmail.com: add comment about VM_DROPPABLE in code, make it clearer]
Link: https://lkml.kernel.org/r/i4uowkt4h2ev47obm5h2vtd4zbk6fyw5g364up7kkjn2vmcikq@auepvqethj5r
Link: https://lkml.kernel.org/r/20260221093918.1456187-5-vernon2gm@gmail.com
Signed-off-by: Vernon Yang <yanglincheng@kylinos.cn>
Acked-by: David Hildenbrand (arm) <david@kernel.org>
Reviewed-by: Lance Yang <lance.yang@linux.dev>
Reviewed-by: Barry Song <baohua@kernel.org>
Cc: Baolin Wang <baolin.wang@linux.alibaba.com>
Cc: Dev Jain <dev.jain@arm.com>
Cc: Liam Howlett <Liam.Howlett@oracle.com>
Cc: Lorenzo Stoakes <lorenzo.stoakes@oracle.com>
Cc: Masami Hiramatsu <mhiramat@kernel.org>
Cc: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
Cc: Nico Pache <npache@redhat.com>
Cc: Ryan Roberts <ryan.roberts@arm.com>
Cc: Steven Rostedt <rostedt@goodmis.org>
Cc: Zi Yan <ziy@nvidia.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Currently, each scan always increases "progress" by HPAGE_PMD_NR,
even if only scanning a single PTE/PMD entry.
- When only scanning a sigle PTE entry, let me provide a detailed
example:
static int hpage_collapse_scan_pmd()
{
for (addr = start_addr, _pte = pte; _pte < pte + HPAGE_PMD_NR;
_pte++, addr += PAGE_SIZE) {
pte_t pteval = ptep_get(_pte);
...
if (pte_uffd_wp(pteval)) { <-- first scan hit
result = SCAN_PTE_UFFD_WP;
goto out_unmap;
}
}
}
During the first scan, if pte_uffd_wp(pteval) is true, the loop exits
directly. In practice, only one PTE is scanned before termination. Here,
"progress += 1" reflects the actual number of PTEs scanned, but previously
"progress += HPAGE_PMD_NR" always.
- When the memory has been collapsed to PMD, let me provide a detailed
example:
The following data is traced by bpftrace on a desktop system. After the
system has been left idle for 10 minutes upon booting, a lot of
SCAN_PMD_MAPPED or SCAN_NO_PTE_TABLE are observed during a full scan by
khugepaged.
From trace_mm_khugepaged_scan_pmd and trace_mm_khugepaged_scan_file, the
following statuses were observed, with frequency mentioned next to them:
SCAN_SUCCEED : 1
SCAN_EXCEED_SHARED_PTE: 2
SCAN_PMD_MAPPED : 142
SCAN_NO_PTE_TABLE : 178
total progress size : 674 MB
Total time : 419 seconds, include khugepaged_scan_sleep_millisecs
The khugepaged_scan list save all task that support collapse into
hugepage, as long as the task is not destroyed, khugepaged will not remove
it from the khugepaged_scan list. This exist a phenomenon where task has
already collapsed all memory regions into hugepage, but khugepaged
continues to scan it, which wastes CPU time and invalid, and due to
khugepaged_scan_sleep_millisecs (default 10s) causes a long wait for
scanning a large number of invalid task, so scanning really valid task is
later.
After applying this patch, when the memory is either SCAN_PMD_MAPPED or
SCAN_NO_PTE_TABLE, just skip it, as follow:
SCAN_EXCEED_SHARED_PTE: 2
SCAN_PMD_MAPPED : 147
SCAN_NO_PTE_TABLE : 173
total progress size : 45 MB
Total time : 20 seconds
SCAN_PTE_MAPPED_HUGEPAGE is the same, for detailed data, refer to
https://lore.kernel.org/linux-mm/4qdu7owpmxfh3ugsue775fxarw5g2gcggbxdf5psj75nnu7z2u@cv2uu2yocaxq
Link: https://lkml.kernel.org/r/20260221093918.1456187-3-vernon2gm@gmail.com
Signed-off-by: Vernon Yang <yanglincheng@kylinos.cn>
Reviewed-by: Dev Jain <dev.jain@arm.com>
Cc: Baolin Wang <baolin.wang@linux.alibaba.com>
Cc: Barry Song <baohua@kernel.org>
Cc: David Hildenbrand (arm) <david@kernel.org>
Cc: Lance Yang <lance.yang@linux.dev>
Cc: Liam Howlett <Liam.Howlett@oracle.com>
Cc: Lorenzo Stoakes <lorenzo.stoakes@oracle.com>
Cc: Masami Hiramatsu <mhiramat@kernel.org>
Cc: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
Cc: Nico Pache <npache@redhat.com>
Cc: Ryan Roberts <ryan.roberts@arm.com>
Cc: Steven Rostedt <rostedt@goodmis.org>
Cc: Zi Yan <ziy@nvidia.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Patch series "Improve khugepaged scan logic", v8.
This series improves the khugepaged scan logic and reduces CPU consumption
by prioritizing scanning tasks that access memory frequently.
The following data is traced by bpftrace[1] on a desktop system. After
the system has been left idle for 10 minutes upon booting, a lot of
SCAN_PMD_MAPPED or SCAN_NO_PTE_TABLE are observed during a full scan by
khugepaged.
@scan_pmd_status[1]: 1 ## SCAN_SUCCEED
@scan_pmd_status[6]: 2 ## SCAN_EXCEED_SHARED_PTE
@scan_pmd_status[3]: 142 ## SCAN_PMD_MAPPED
@scan_pmd_status[2]: 178 ## SCAN_NO_PTE_TABLE
total progress size: 674 MB
Total time : 419 seconds ## include khugepaged_scan_sleep_millisecs
The khugepaged has below phenomenon: the khugepaged list is scanned in a
FIFO manner, as long as the task is not destroyed,
1. the task no longer has memory that can be collapsed into hugepage,
continues scan it always.
2. the task at the front of the khugepaged scan list is cold, they are
still scanned first.
3. everyone scan at intervals of khugepaged_scan_sleep_millisecs
(default 10s). If we always scan the above two cases first, the valid
scan will have to wait for a long time.
For the first case, when the memory is either SCAN_PMD_MAPPED or
SCAN_NO_PTE_TABLE or SCAN_PTE_MAPPED_HUGEPAGE [5], just skip it.
For the second case, if the user has explicitly informed us via
MADV_FREE that these folios will be freed, just skip it only.
The below is some performance test results.
kernbench results (testing on x86_64 machine):
baseline w/o patches test w/ patches
Amean user-32 18522.51 ( 0.00%) 18333.64 * 1.02%*
Amean syst-32 1137.96 ( 0.00%) 1113.79 * 2.12%*
Amean elsp-32 666.04 ( 0.00%) 659.44 * 0.99%*
BAmean-95 user-32 18520.01 ( 0.00%) 18323.57 ( 1.06%)
BAmean-95 syst-32 1137.68 ( 0.00%) 1110.50 ( 2.39%)
BAmean-95 elsp-32 665.92 ( 0.00%) 659.06 ( 1.03%)
BAmean-99 user-32 18520.01 ( 0.00%) 18323.57 ( 1.06%)
BAmean-99 syst-32 1137.68 ( 0.00%) 1110.50 ( 2.39%)
BAmean-99 elsp-32 665.92 ( 0.00%) 659.06 ( 1.03%)
Create three task[2]: hot1 -> cold -> hot2. After all three task are
created, each allocate memory 128MB. the hot1/hot2 task continuously
access 128 MB memory, while the cold task only accesses its memory
briefly andthen call madvise(MADV_FREE). Here are the performance test
results:
(Throughput bigger is better, other smaller is better)
Testing on x86_64 machine:
| task hot2 | without patch | with patch | delta |
|---------------------|---------------|---------------|---------|
| total accesses time | 3.14 sec | 2.93 sec | -6.69% |
| cycles per access | 4.96 | 2.21 | -55.44% |
| Throughput | 104.38 M/sec | 111.89 M/sec | +7.19% |
| dTLB-load-misses | 284814532 | 69597236 | -75.56% |
Testing on qemu-system-x86_64 -enable-kvm:
| task hot2 | without patch | with patch | delta |
|---------------------|---------------|---------------|---------|
| total accesses time | 3.35 sec | 2.96 sec | -11.64% |
| cycles per access | 7.29 | 2.07 | -71.60% |
| Throughput | 97.67 M/sec | 110.77 M/sec | +13.41% |
| dTLB-load-misses | 241600871 | 3216108 | -98.67% |
This patch (of 4):
Add mm_khugepaged_scan event to track the total time for full scan and the
total number of pages scanned of khugepaged.
Link: https://lkml.kernel.org/r/20260221093918.1456187-2-vernon2gm@gmail.com
Signed-off-by: Vernon Yang <yanglincheng@kylinos.cn>
Acked-by: David Hildenbrand (Red Hat) <david@kernel.org>
Reviewed-by: Barry Song <baohua@kernel.org>
Reviewed-by: Lance Yang <lance.yang@linux.dev>
Reviewed-by: Dev Jain <dev.jain@arm.com>
Cc: Baolin Wang <baolin.wang@linux.alibaba.com>
Cc: Dev Jain <dev.jain@arm.com>
Cc: Liam Howlett <Liam.Howlett@oracle.com>
Cc: Lorenzo Stoakes <lorenzo.stoakes@oracle.com>
Cc: Masami Hiramatsu <mhiramat@kernel.org>
Cc: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
Cc: Nico Pache <npache@redhat.com>
Cc: Ryan Roberts <ryan.roberts@arm.com>
Cc: Steven Rostedt <rostedt@goodmis.org>
Cc: Zi Yan <ziy@nvidia.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Reduce 22 declarations of empty_zero_page to 3 and 23 declarations of
ZERO_PAGE() to 4.
Every architecture defines empty_zero_page that way or another, but for the
most of them it is always a page aligned page in BSS and most definitions
of ZERO_PAGE do virt_to_page(empty_zero_page).
Move Linus vetted x86 definition of empty_zero_page and ZERO_PAGE() to the
core MM and drop these definitions in architectures that do not implement
colored zero page (MIPS and s390).
ZERO_PAGE() remains a macro because turning it to a wrapper for a static
inline causes severe pain in header dependencies.
For the most part the change is mechanical, with these being noteworthy:
* alpha: aliased empty_zero_page with ZERO_PGE that was also used for boot
parameters. Switching to a generic empty_zero_page removes the aliasing
and keeps ZERO_PGE for boot parameters only
* arm64: uses __pa_symbol() in ZERO_PAGE() so that definition of
ZERO_PAGE() is kept intact.
* m68k/parisc/um: allocated empty_zero_page from memblock,
although they do not support zero page coloring and having it in BSS
will work fine.
* sparc64 can have empty_zero_page in BSS rather allocate it, but it
can't use virt_to_page() for BSS. Keep it's definition of ZERO_PAGE()
but instead of allocating it, make mem_map_zero point to
empty_zero_page.
* sh: used empty_zero_page for boot parameters at the very early boot.
Rename the parameters page to boot_params_page and let sh use the generic
empty_zero_page.
* hexagon: had an amusing comment about empty_zero_page
/* A handy thing to have if one has the RAM. Declared in head.S */
that unfortunately had to go :)
Link: https://lkml.kernel.org/r/20260211103141.3215197-4-rppt@kernel.org
Signed-off-by: Mike Rapoport (Microsoft) <rppt@kernel.org>
Acked-by: Helge Deller <deller@gmx.de> [parisc]
Tested-by: Helge Deller <deller@gmx.de> [parisc]
Reviewed-by: Christophe Leroy (CS GROUP) <chleroy@kernel.org>
Acked-by: Dave Hansen <dave.hansen@linux.intel.com>
Acked-by: Catalin Marinas <catalin.marinas@arm.com>
Acked-by: Magnus Lindholm <linmag7@gmail.com> [alpha]
Acked-by: Dinh Nguyen <dinguyen@kernel.org> [nios2]
Acked-by: Andreas Larsson <andreas@gaisler.com> [sparc]
Acked-by: David Hildenbrand (Arm) <david@kernel.org>
Acked-by: Liam R. Howlett <Liam.Howlett@oracle.com>
Cc: "Borislav Petkov (AMD)" <bp@alien8.de>
Cc: David S. Miller <davem@davemloft.net>
Cc: Geert Uytterhoeven <geert@linux-m68k.org>
Cc: Guo Ren <guoren@kernel.org>
Cc: Huacai Chen <chenhuacai@kernel.org>
Cc: Ingo Molnar <mingo@redhat.com>
Cc: Johannes Berg <johannes@sipsolutions.net>
Cc: John Paul Adrian Glaubitz <glaubitz@physik.fu-berlin.de>
Cc: Lorenzo Stoakes <lorenzo.stoakes@oracle.com>
Cc: Madhavan Srinivasan <maddy@linux.ibm.com>
Cc: Matt Turner <mattst88@gmail.com>
Cc: Max Filippov <jcmvbkbc@gmail.com>
Cc: Michael Ellerman <mpe@ellerman.id.au>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Michal Simek <monstr@monstr.eu>
Cc: Palmer Dabbelt <palmer@dabbelt.com>
Cc: Richard Weinberger <richard@nod.at>
Cc: Russell King <linux@armlinux.org.uk>
Cc: Stafford Horne <shorne@gmail.com>
Cc: Suren Baghdasaryan <surenb@google.com>
Cc: Vineet Gupta <vgupta@kernel.org>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Will Deacon <will@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Add test_zswap_incompressible() to verify that the zswap_incomp memcg stat
correctly tracks incompressible pages.
The test allocates memory filled with random data from /dev/urandom, which
cannot be effectively compressed by zswap. When this data is swapped out
to zswap, it should be stored as-is and tracked by the zswap_incomp
counter.
The test verifies that:
1. Pages are swapped out to zswap (zswpout increases)
2. Incompressible pages are tracked (zswap_incomp increases)
test:
dd if=/dev/zero of=/swapfile bs=1M count=2048
chmod 600 /swapfile
mkswap /swapfile
swapon /swapfile
echo Y > /sys/module/zswap/parameters/enabled
./test_zswap
TAP version 13
1..8
ok 1 test_zswap_usage
ok 2 test_swapin_nozswap
ok 3 test_zswapin
ok 4 test_zswap_writeback_enabled
ok 5 test_zswap_writeback_disabled
ok 6 test_no_kmem_bypass
ok 7 test_no_invasive_cgroup_shrink
ok 8 test_zswap_incompressible
Totals: pass:8 fail:0 xfail:0 xpass:0 skip:0 error:0
Link: https://lkml.kernel.org/r/20260213071827.5688-3-jiayuan.chen@linux.dev
Signed-off-by: Jiayuan Chen <jiayuan.chen@shopee.com>
Acked-by: Shakeel Butt <shakeel.butt@linux.dev>
Acked-by: Nhat Pham <nphamcs@gmail.com>
Reviewed-by: SeongJae Park <sj@kernel.org>
Cc: Chengming Zhou <chengming.zhou@linux.dev>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Michal Koutný <mkoutny@suse.com>
Cc: Muchun Song <muchun.song@linux.dev>
Cc: Roman Gushchin <roman.gushchin@linux.dev>
Cc: Shuah Khan <shuah@kernel.org>
Cc: Tejun Heo <tj@kernel.org>
Cc: Yosry Ahmed <yosry.ahmed@linux.dev>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Patch series "mm: zswap: add per-memcg stat for incompressible pages", v3.
In containerized environments, knowing which cgroup is contributing
incompressible pages to zswap is essential for effective resource
management. This series adds a new per-memcg stat 'zswap_incomp' to track
incompressible pages, along with a selftest.
This patch (of 2):
The global zswap_stored_incompressible_pages counter was added in commit
dca4437a58 ("mm/zswap: store <PAGE_SIZE compression failed page as-is")
to track how many pages are stored in raw (uncompressed) form in zswap.
However, in containerized environments, knowing which cgroup is
contributing incompressible pages is essential for effective resource
management [1].
Add a new memcg stat 'zswap_incomp' to track incompressible pages per
cgroup. This helps administrators and orchestrators to:
1. Identify workloads that produce incompressible data (e.g., encrypted
data, already-compressed media, random data) and may not benefit from
zswap.
2. Make informed decisions about workload placement - moving
incompressible workloads to nodes with larger swap backing devices
rather than relying on zswap.
3. Debug zswap efficiency issues at the cgroup level without needing to
correlate global stats with individual cgroups.
While the compression ratio can be estimated from existing stats (zswap /
zswapped * PAGE_SIZE), this doesn't distinguish between "uniformly poor
compression" and "a few completely incompressible pages mixed with highly
compressible ones". The zswap_incomp stat provides direct visibility into
the latter case.
Link: https://lkml.kernel.org/r/20260213071827.5688-1-jiayuan.chen@linux.dev
Link: https://lkml.kernel.org/r/20260213071827.5688-2-jiayuan.chen@linux.dev
Link: https://lore.kernel.org/linux-mm/CAF8kJuONDFj4NAksaR4j_WyDbNwNGYLmTe-o76rqU17La=nkOw@mail.gmail.com/ [1]
Signed-off-by: Jiayuan Chen <jiayuan.chen@shopee.com>
Acked-by: Nhat Pham <nphamcs@gmail.com>
Acked-by: Shakeel Butt <shakeel.butt@linux.dev>
Reviewed-by: Yosry Ahmed <yosry.ahmed@linux.dev>
Reviewed-by: SeongJae Park <sj@kernel.org>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Chengming Zhou <chengming.zhou@linux.dev>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Michal Koutný <mkoutny@suse.com>
Cc: Muchun Song <muchun.song@linux.dev>
Cc: Roman Gushchin <roman.gushchin@linux.dev>
Cc: Shuah Khan <shuah@kernel.org>
Cc: Tejun Heo <tj@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
We currently have two different sets of helpers for getting or putting the
private IDs' refcount for order 0 and large folios. This is redundant.
Just use one and always acquire the refcount of the swapout folio size
unless it's zero, and put the refcount using the folio size if the charge
failed, since the folio size can't change. Then there is no need to
update the refcount for tail pages.
Same for freeing, then only one pair of get/put helper is needed now.
The performance might be slightly better, too: both "inc unless zero" and
"add unless zero" use the same cmpxchg implementation. For large folios,
we saved an atomic operation. And for both order 0 and large folios, we
saved a branch.
Link: https://lkml.kernel.org/r/20260213-memcg-privid-v1-1-d8cb7afcf831@tencent.com
Signed-off-by: Kairui Song <kasong@tencent.com>
Acked-by: Johannes Weiner <hannes@cmpxchg.org>
Reviewed-by: Chen Ridong <chenridong@huaweicloud.com>
Acked-by: Shakeel Butt <shakeel.butt@gmail.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Muchun Song <muchun.song@linux.dev>
Cc: Roman Gushchin <roman.gushchin@linux.dev>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
File seals are used on memfd for making shared memory communication with
untrusted peers safer and simpler. Seals provide a guarantee that certain
operations won't be allowed on the file such as writes or truncations.
Maintaining these guarantees across a live update will help keeping such
use cases secure.
These guarantees will also be needed for IOMMUFD preservation with LUO.
Normally when IOMMUFD maps a memfd, it pins all its pages to make sure any
truncation operations on the memfd don't lead to IOMMUFD using freed
memory. This doesn't work with LUO since the preserved memfd might have
completely different pages after a live update, and mapping them back to
the IOMMUFD will cause all sorts of problems. Using and preserving the
seals allows IOMMUFD preservation logic to trust the memfd.
Since the uABI defines seals as an int, preserve them by introducing a new
u32 field. There are currently only 6 possible seals, so the extra bits
are unused and provide room for future expansion. Since the seals are
uABI, it is safe to use them directly in the ABI. While at it, also add a
u32 flags field. It makes sure the struct is nicely aligned, and can be
used later to support things like MFD_CLOEXEC.
Since the serialization structure is changed, bump the version number to
"memfd-v2".
It is important to note that the memfd-v2 version only supports seals that
existed when this version was defined. This set is defined by
MEMFD_LUO_ALL_SEALS. Any new seal might bring a completely different
semantic with it and the parser for memfd-v2 cannot be expected to deal
with that. If there are any future seals added, they will need another
version bump.
Link: https://lkml.kernel.org/r/20260216185946.1215770-3-pratyush@kernel.org
Signed-off-by: Pratyush Yadav (Google) <pratyush@kernel.org>
Tested-by: Samiullah Khawaja <skhawaja@google.com>
Cc: Alexander Graf <graf@amazon.com>
Cc: Baolin Wang <baolin.wang@linux.alibaba.com>
Cc: Hugh Dickins <hughd@google.com>
Cc: Jason Gunthorpe <jgg@nvidia.com>
Cc: Mike Rapoport (Microsoft) <rppt@kernel.org>
Cc: Pasha Tatashin <pasha.tatashin@soleen.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Clean up and simplify how we check if a folio is swapped. The helper
already requires the folio to be in swap cache and locked. That's enough
to pin the swap cluster from being freed, so there is no need to lock
anything else to avoid UAF.
And besides, we have cleaned up and defined the swap operation to be
mostly folio based, and now the only place a folio will have any of its
swap slots' count increased from 0 to 1 is folio_dup_swap, which also
requires the folio lock. So as we are holding the folio lock here, a
folio can't change its swap status from not swapped (all swap slots have a
count of 0) to swapped (any slot has a swap count larger than 0).
So there won't be any false negatives of this helper if we simply depend
on the folio lock to stabilize the cluster.
We are only using this helper to determine if we can and should release
the swap cache. So false positives are completely harmless, and also
already exist before. Depending on the timing, previously, it's also
possible that a racing thread releases the swap count right after
releasing the ci lock and before this helper returns. In any case, the
worst that could happen is we leave a clean swap cache. It will still be
reclaimed when under pressure just fine.
So, in conclusion, we can simplify and make the check much simpler and
lockless. Also, rename it to folio_maybe_swapped to reflect the design.
Link: https://lkml.kernel.org/r/20260218-swap-table-p3-v3-11-f4e34be021a7@tencent.com
Signed-off-by: Kairui Song <kasong@tencent.com>
Acked-by: Chris Li <chrisl@kernel.org>
Cc: Baoquan He <bhe@redhat.com>
Cc: Barry Song <baohua@kernel.org>
Cc: David Hildenbrand <david@kernel.org>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Kairui Song <ryncsn@gmail.com>
Cc: Kemeng Shi <shikemeng@huaweicloud.com>
Cc: kernel test robot <lkp@intel.com>
Cc: Lorenzo Stoakes <lorenzo.stoakes@oracle.com>
Cc: Nhat Pham <nphamcs@gmail.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Now all the infrastructures are ready, switch to using the swap table
only. This is unfortunately a large patch because the whole old counting
mechanism, especially SWP_CONTINUED, has to be gone and switch to the new
mechanism together, with no intermediate steps available.
The swap table is capable of holding up to SWP_TB_COUNT_MAX - 1 counts in
the higher bits of each table entry, so using that, the swap_map can be
completely dropped.
swap_map also had a limit of SWAP_CONT_MAX. Any value beyond that limit
will require a COUNT_CONTINUED page. COUNT_CONTINUED is a bit complex to
maintain, so for the swap table, a simpler approach is used: when the
count goes beyond SWP_TB_COUNT_MAX - 1, the cluster will have an
extend_table allocated, which is a swap cluster-sized array of unsigned
int. The counting is basically offloaded there until the count drops
below SWP_TB_COUNT_MAX again.
Both the swap table and the extend table are cluster-based, so they
exhibit good performance and sparsity.
To make the switch from swap_map to swap table clean, this commit cleans
up and introduces a new set of functions based on the swap table design,
for manipulating swap counts:
- __swap_cluster_dup_entry, __swap_cluster_put_entry,
__swap_cluster_alloc_entry, __swap_cluster_free_entry:
Increase/decrease the count of a swap slot, or alloc / free a swap
slot. This is the internal routine that does the counting work based
on the swap table and handles all the complexities. The caller will
need to lock the cluster before calling them.
All swap count-related update operations are wrapped by these four
helpers.
- swap_dup_entries_cluster, swap_put_entries_cluster:
Increase/decrease the swap count of one or a set of swap slots in the
same cluster range. These two helpers serve as the common routines for
folio_dup_swap & swap_dup_entry_direct, or
folio_put_swap & swap_put_entries_direct.
And use these helpers to replace all existing callers. This helps to
simplify the count tracking by a lot, and the swap_map is gone.
[ryncsn@gmail.com: fix build]
Link: https://lkml.kernel.org/r/aZWuLZi-vYi3vAWe@KASONG-MC4
Link: https://lkml.kernel.org/r/20260218-swap-table-p3-v3-9-f4e34be021a7@tencent.com
Signed-off-by: Kairui Song <kasong@tencent.com>
Suggested-by: Chris Li <chrisl@kernel.org>
Acked-by: Chris Li <chrisl@kernel.org>
Cc: Baoquan He <bhe@redhat.com>
Cc: Barry Song <baohua@kernel.org>
Cc: David Hildenbrand <david@kernel.org>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Kairui Song <ryncsn@gmail.com>
Cc: Kemeng Shi <shikemeng@huaweicloud.com>
Cc: kernel test robot <lkp@intel.com>
Cc: Lorenzo Stoakes <lorenzo.stoakes@oracle.com>
Cc: Nhat Pham <nphamcs@gmail.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Swap table entry will need 4 bits reserved for swap count in the shadow,
so the anon shadow should have its leading 4 bits remain 0.
This should be OK for the foreseeable future. Take 52 bits of physical
address space as an example: for 4K pages, there would be at most 40 bits
for addressable pages. Currently, we have 36 bits available (64 - 1 - 16
- 10 - 1, where XA_VALUE takes 1 bit for marker, MEM_CGROUP_ID_SHIFT takes
16 bits, NODES_SHIFT takes <=10 bits, WORKINGSET flags takes 1 bit).
So in the worst case, we previously need to pack the 40 bits of address in
36 bits fields using a 64K bucket (bucket_order = 4). After this, the
bucket will be increased to 1M. Which should be fine, as on such large
machines, the working set size will be way larger than the bucket size.
And for MGLRU's gen number tracking, it should be even more than enough,
MGLRU's gen number (max_seq) increment is much slower compared to the
eviction counter (nonresident_age).
And after all, either the refault distance or the gen distance is only a
hint that can tolerate inaccuracy just fine.
And the 4 bits can be shrunk to 3, or extended to a higher value if needed
later.
Link: https://lkml.kernel.org/r/20260218-swap-table-p3-v3-5-f4e34be021a7@tencent.com
Signed-off-by: Kairui Song <kasong@tencent.com>
Acked-by: Chris Li <chrisl@kernel.org>
Cc: Baoquan He <bhe@redhat.com>
Cc: Barry Song <baohua@kernel.org>
Cc: David Hildenbrand <david@kernel.org>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Kairui Song <ryncsn@gmail.com>
Cc: Kemeng Shi <shikemeng@huaweicloud.com>
Cc: kernel test robot <lkp@intel.com>
Cc: Lorenzo Stoakes <lorenzo.stoakes@oracle.com>
Cc: Nhat Pham <nphamcs@gmail.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Slightly clean up the swapon process. Add comments about what swap_lock
protects, introduce and rename helpers that wrap swap_map and cluster_info
setup, and do it outside of the swap_lock lock.
This lock protection is not needed for swap_map and cluster_info setup
because all swap users must either hold the percpu ref or hold a stable
allocated swap entry (e.g., locking a folio in the swap cache) before
accessing. So before the swap device is exposed by enable_swap_info,
nothing would use the swap device's map or cluster.
So we are safe to allocate and set up swap data freely first, then expose
the swap device and set the SWP_WRITEOK flag.
Link: https://lkml.kernel.org/r/20260218-swap-table-p3-v3-2-f4e34be021a7@tencent.com
Signed-off-by: Kairui Song <kasong@tencent.com>
Acked-by: Chris Li <chrisl@kernel.org>
Cc: Baoquan He <bhe@redhat.com>
Cc: Barry Song <baohua@kernel.org>
Cc: David Hildenbrand <david@kernel.org>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Kairui Song <ryncsn@gmail.com>
Cc: Kemeng Shi <shikemeng@huaweicloud.com>
Cc: kernel test robot <lkp@intel.com>
Cc: Lorenzo Stoakes <lorenzo.stoakes@oracle.com>
Cc: Nhat Pham <nphamcs@gmail.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
