This reintroduces a concept removed by: commit d6cb41cc44 ("mm, hugetlb:
remove hugepages_treat_as_movable sysctl")
This sysctl provides flexibility between ZONE_MOVABLE use cases:
1) onlining memory in ZONE_MOVABLE to maintain hotplug compatibility
2) onlining memory in ZONE_MOVABLE to make hugepage allocate reliable
When ZONE_MOVABLE is used to make huge page allocation more reliable,
disallowing gigantic pages memory in this region is pointless. If hotplug
is not a requirement, we can loosen the restrictions to allow 1GB gigantic
pages in ZONE_MOVABLE.
Since 1GB can be difficult to migrate / has impacts on compaction /
defragmentation, we don't enable this by default. Notably, 1GB pages can
only be migrated if another 1GB page is available - so hot-unplug will
fail if such a page cannot be found.
However, since there are scenarios where gigantic pages are migratable, we
should allow use of these on movable regions.
When not valid 1GB is available for migration, hot-unplug will retry
indefinitely (or until interrupted). For example:
echo 0 > node0/hugepages/..-1GB/nr_hugepages # clear node0 1GB pages
echo 1 > node1/hugepages/..-1GB/nr_hugepages # reserve node1 1GB page
./alloc_huge_node1 & # Allocate a 1GB page on node1
./node1_offline & # attempt to offline all node1 memory
echo 1 > node0/hugepages/..-1GB/nr_hugepages # reserve node0 1GB page
In this example, node1_offline will block indefinitely until the final
step, when a node0 1GB page is made available.
Note: Boot-time CMA is not possible for driver-managed hotplug memory, as
CMA requires the memory to be registered as SystemRAM at boot time.
Additionally, 1GB huge pages are not supported by THP.
Link: https://lkml.kernel.org/r/20251221125603.2364174-1-gourry@gourry.net
Signed-off-by: Gregory Price <gourry@gourry.net>
Suggested-by: David Rientjes <rientjes@google.com>
Link: https://lore.kernel.org/all/20180201193132.Hk7vI_xaU%25akpm@linux-foundation.org/
Acked-by: David Hildenbrand (Red Hat) <david@kernel.org>
Acked-by: David Rientjes <rientjes@google.com>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Michal Hocko <mhocko@suse.com>
Cc: "David Hildenbrand (Red Hat)" <david@kernel.org>
Cc: Gregory Price <gourry@gourry.net>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Liam Howlett <liam.howlett@oracle.com>
Cc: Lorenzo Stoakes <lorenzo.stoakes@oracle.com>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Muchun Song <muchun.song@linux.dev>
Cc: Oscar Salvador <osalvador@suse.de>
Cc: Suren Baghdasaryan <surenb@google.com>
Cc: Vlastimil Babka <vbabka@suse.cz>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Since commit cc638f329e ("mm, thp: tweak reclaim/compaction effort of
local-only and all-node allocations"), THP page fault allocations have
settled on the following scheme (from the commit log):
1. local node only THP allocation with no reclaim, just compaction.
2. for madvised VMA's or when synchronous compaction is enabled always - THP
allocation from any node with effort determined by global defrag setting
and VMA madvise
3. fallback to base pages on any node
Recent customer reports however revealed we have a gap in step 1 above.
What we have seen is excessive reclaim due to THP page faults on a NUMA
node that's close to its high watermark, while other nodes have plenty of
free memory.
The problem with step 1 is that it promises no reclaim after the
compaction attempt, however reclaim is only avoided for certain compaction
outcomes (deferred, or skipped due to insufficient free base pages), and
not e.g. when compaction is actually performed but fails (we did see
compact_fail vmstat counter increasing).
THP page faults can therefore exhibit a zone_reclaim_mode-like behavior,
which is not the intention.
Thus add a check for __GFP_THISNODE that corresponds to this exact
situation and prevents continuing with reclaim/compaction once the initial
compaction attempt isn't successful in allocating the page.
Note that commit cc638f329e has not introduced this over-reclaim
possibility; it appears to exist in some form since commit 2f0799a0ff
("mm, thp: restore node-local hugepage allocations"). Followup commits
b39d0ee263 ("mm, page_alloc: avoid expensive reclaim when compaction may
not succeed") and cc638f329e have moved in the right direction, but left
the abovementioned gap.
Link: https://lkml.kernel.org/r/20251219-costly-noretry-thisnode-fix-v1-1-e1085a4a0c34@suse.cz
Fixes: 2f0799a0ff ("mm, thp: restore node-local hugepage allocations")
Signed-off-by: Vlastimil Babka <vbabka@suse.cz>
Acked-by: Michal Hocko <mhocko@suse.com>
Acked-by: Johannes Weiner <hannes@cmpxchg.org>
Acked-by: Pedro Falcato <pfalcato@suse.de>
Acked-by: Zi Yan <ziy@nvidia.com>
Cc: Brendan Jackman <jackmanb@google.com>
Cc: "David Hildenbrand (Red Hat)" <david@kernel.org>
Cc: David Rientjes <rientjes@google.com>
Cc: Joshua Hahn <joshua.hahnjy@gmail.com>
Cc: Liam Howlett <liam.howlett@oracle.com>
Cc: Lorenzo Stoakes <lorenzo.stoakes@oracle.com>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Suren Baghdasaryan <surenb@google.com>
Cc: <stable@vger.kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
The KMSG_COMPONENT macro is a leftover of the s390 specific "kernel
message catalog" from 2008 [1] which never made it upstream.
The macro was added to s390 code to allow for an out-of-tree patch which
used this to generate unique message ids. Also this out-of-tree doesn't
exist anymore.
The pattern of how the KMSG_COMPONENT is used was partially also used for
non s390 specific code, for whatever reasons.
Remove the macro in order to get rid of a pointless indirection.
Link: https://lkml.kernel.org/r/20251126143602.2207435-1-hca@linux.ibm.com
Link: https://lwn.net/Articles/292650/ [1]
Signed-off-by: Heiko Carstens <hca@linux.ibm.com>
Reviewed-by: Sergey Senozhatsky <senozhatsky@chromium.org>
Cc: Jens Axboe <axboe@kernel.dk>
Cc: Minchan Kim <minchan@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Use bool for 'hashdist' and replace simple_strtoul() with kstrtobool() for
parsing the 'hashdist=' boot parameter. Unlike simple_strtoul(), which
returns an unsigned long, kstrtobool() converts the string directly to
bool and avoids implicit casting.
Check the return value of kstrtobool() and reject invalid values. This
adds error handling while preserving behavior for existing values, and
removes use of the deprecated simple_strtoul() helper. The current code
silently sets 'hashdist = 0' if parsing fails, instead of leaving the
default value (HASHDIST_DEFAULT) unchanged.
Additionally, kstrtobool() accepts common boolean strings such as "on" and
"off".
Link: https://lkml.kernel.org/r/20251217110214.50807-1-thorsten.blum@linux.dev
Signed-off-by: Thorsten Blum <thorsten.blum@linux.dev>
Reviewed-by: Matthew Wilcox (Oracle) <willy@infradead.org>
Reviewed-by: Mike Rapoport (Microsoft) <rppt@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Laptop mode was introduced to save battery, by delaying and consolidating
writes and thereby maximize the time rotating hard drives wouldn't have to
spin.
Luckily, rotating hard drives, with their high spin-up times and power
draw, are a thing of the past for battery-powered devices. Reclaim has
also since changed to not write single filesystem pages anymore, and
regular filesystem writeback is lumpy by design.
The juice doesn't appear worth the squeeze anymore. The footprint of the
feature is small, but nevertheless it's a complicating factor in mm,
block, filesystems. Developers don't think about it, and it likely hasn't
been tested with new reclaim and writeback changes in years.
Let's sunset it. Keep the sysctl with a deprecation warning around for a
few more cycles, but remove all functionality behind it.
[akpm@linux-foundation.org: fix Documentation/admin-guide/laptops/index.rst]
Link: https://lkml.kernel.org/r/20251216185201.GH905277@cmpxchg.org
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Suggested-by: Christoph Hellwig <hch@infradead.org>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Acked-by: Jens Axboe <axboe@kernel.dk>
Reviewed-by: Shakeel Butt <shakeel.butt@linux.dev>
Acked-by: Michal Hocko <mhocko@suse.com>
Cc: Deepanshu Kartikey <kartikey406@gmail.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
pp_in_progress makes sure that only one post-processing (writeback or
recomrpession) is active at any given time. Functionality wise it,
basically, shadows zram init_lock, when init_lock is acquired in writer
mode.
Switch recompress_store() and writeback_store() to take zram init_lock in
writer mode, like all store() sysfs handlers should do, so that we can
drop pp_in_progress. Recompression and writeback can be somewhat slow, so
holding init_lock in writer mode can block zram attrs reads, but in
reality the only zram attrs reads that take place are mm_stat reads, and
usually it's the same process that reads mm_stat and does recompression or
writeback.
Link: https://lkml.kernel.org/r/20251216071342.687993-1-senozhatsky@chromium.org
Signed-off-by: Sergey Senozhatsky <senozhatsky@chromium.org>
Suggested-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Cc: Brian Geffon <bgeffon@google.com>
Cc: Minchan Kim <minchan@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
DAMON users can read DAMOS stats via DAMON sysfs interface. It enables
efficient, simple and flexible usages of the stats. Especially for
systems not having advanced tools like perf or bpftrace, that can be
useful. But if the advanced tools are available, exposing the stats via
tracepoint can reduce unnecessary reimplementation of the wheels. Add a
new tracepoint for DAMOS stats, namely damos_stat_after_apply_interval.
The tracepoint is triggered for each scheme's apply interval and exposes
the whole stat values. If the user needs sub-apply interval information
for any chance, damos_before_apply tracepoint could be used.
Link: https://lkml.kernel.org/r/20251216080128.42991-13-sj@kernel.org
Signed-off-by: SeongJae Park <sj@kernel.org>
Reviewed-by: Steven Rostedt (Google) <rostedt@goodmis.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Liam Howlett <liam.howlett@oracle.com>
Cc: Lorenzo Stoakes <lorenzo.stoakes@oracle.com>
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: Suren Baghdasaryan <surenb@google.com>
Cc: Vlastimil Babka <vbabka@suse.cz>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
There are DAMOS use cases that require user-space centric control of its
activation and deactivation. Having the control plane on the user-space,
or using DAMOS as a way for monitoring results collection are such
examples.
DAMON parameters online commit, DAMOS quotas and watermarks can be useful
for this purpose. However, those features work only at the
sub-DAMON-snapshot level. In some use cases, the DAMON-snapshot level
control is required. For example, in DAMOS-based monitoring results
collection use case, the user online-installs a DAMOS scheme with
DAMOS_STAT action, wait it be applied to whole regions of a single
DAMON-snapshot, retrieves the stats and tried regions information, and
online-uninstall the scheme. It is efficient to ensure the lifetime of
the scheme as no more no less one snapshot consumption.
To support such use cases, introduce a new DAMOS core API per-scheme
parameter, namely max_nr_snapshots. As the name implies, it is the upper
limit of nr_snapshots, which is a DAMOS stat that represents the number of
DAMON-snapshots that the scheme has fully applied. If the limit is set
with a non-zero value and nr_snapshots reaches or exceeds the limit, the
scheme is deactivated.
Link: https://lkml.kernel.org/r/20251216080128.42991-8-sj@kernel.org
Signed-off-by: SeongJae Park <sj@kernel.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Liam Howlett <liam.howlett@oracle.com>
Cc: Lorenzo Stoakes <lorenzo.stoakes@oracle.com>
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: Steven Rostedt <rostedt@goodmis.org>
Cc: Suren Baghdasaryan <surenb@google.com>
Cc: Vlastimil Babka <vbabka@suse.cz>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Patch series "mm/damon: introduce {,max_}nr_snapshots and tracepoint for
damos stats".
Introduce three changes for improving DAMOS stat's provided information,
deterministic control, and reading usability.
DAMOS provides stats that are important for understanding its behavior.
It lacks information about how many DAMON-generated monitoring output
snapshots it has worked on. Add a new stat, nr_snapshots, to show the
information.
Users can control DAMOS schemes in multiple ways. Using the online
parameters commit feature, they can install and uninstall DAMOS schemes
whenever they want while keeping DAMON runs. DAMOS quotas and watermarks
can be used for manually or automatically turning on/off or adjusting the
aggressiveness of the scheme. DAMOS filters can be used for applying the
scheme to specific memory entities based on their types and locations.
Some users want their DAMOS scheme to be applied to only specific number
of DAMON snapshots, for more deterministic control. One example use case
is tracepoint based snapshot reading. Add a new knob, max_nr_snapshots,
to support this. If the nr_snapshots parameter becomes same to or greater
than the value of this parameter, the scheme is deactivated.
Users can read DAMOS stats via DAMON's sysfs interface. For deep level
investigations on environments having advanced tools like perf and
bpftrace, exposing the stats via a tracepoint can be useful. Implement a
new tracepoint, namely damon:damos_stat_after_apply_interval.
First five patches (patches 1-5) of this series implement the new stat,
nr_snapshots, on the core layer (patch 1), expose on DAMON sysfs user
interface (patch 2), and update documents (patches 3-5).
Following six patches (patches 6-11) are for the new stat based DAMOS
deactivation (max_nr_snapshots). The first one (patch 6) of this group
updates a kernel-doc comment before making further changes. Then an
implementation of it on the core layer (patch 7), an introduction of a new
DAMON sysfs interface file for users of the feature (patch 8), and three
updates of the documents (patches 9-11) follow.
The final one (patch 12) introduces the new tracepoint that exposes the
DAMOS stat values for each scheme apply interval.
This patch (of 12):
DAMON generates monitoring results snapshots for every sampling interval.
DAMOS applies given schemes on the regions of the snapshots, for every
apply interval of the scheme.
DAMOS stat informs a given scheme has tried to how many memory entities
and applied, in the region and byte level. In some use cases including
user-space oriented tuning and investigations, it is useful to know that
in the DAMON-snapshot level. Introduce a new stat, namely nr_snapshots
for DAMON core API callers.
[sj@kernel.org: fix wrong list_is_last() call in damons_is_last_region()]
Link: https://lkml.kernel.org/r/20260114152049.99727-1-sj@kernel.org
Link: https://lkml.kernel.org/r/20251216080128.42991-1-sj@kernel.org
Link: https://lkml.kernel.org/r/20251216080128.42991-2-sj@kernel.org
Signed-off-by: SeongJae Park <sj@kernel.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Liam Howlett <liam.howlett@oracle.com>
Cc: Lorenzo Stoakes <lorenzo.stoakes@oracle.com>
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: Steven Rostedt <rostedt@goodmis.org>
Cc: Suren Baghdasaryan <surenb@google.com>
Cc: Vlastimil Babka <vbabka@suse.cz>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Use ALIGNMENT_SMALLFOLIO instead of ALIGNMENT_MTHP when allocating small
folios to ensure correct memory alignment for the test case.
Before: test allocates small folios with 64KB alignment
(ALIGNMENT_MTHP) when only 4KB alignment (ALIGNMENT_SMALLFOLIO) is
needed. This wastes address space and may cause allocation failures on
systems with fragmented memory.
Worst-case impact: this only affects thp_swap_allocator_test tool
behavior.
Link: https://lkml.kernel.org/r/20251209031745.2723120-1-kaushlendra.kumar@intel.com
Signed-off-by: Kaushlendra Kumar <kaushlendra.kumar@intel.com>
Reviewed-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Currently, DAMON does not proactively clean up invalid monitoring targets
during its runtime. When some monitored processes exit, DAMON continues
to make the following unnecessary function calls,
--damon_for_each_target--
--damon_for_each_region--
damon_do_apply_schemes
damos_apply_scheme
damon_va_apply_scheme
damos_madvise
damon_get_mm
it is only in the damon_get_mm() function that it may finally discover the
target no longer exists, which wastes CPU resources. A simple idea is to
check for the existence of monitoring targets within the
kdamond_need_stop() function and promptly clean up non-existent targets.
However, SJ pointed out that this approach is problematic because the
online commit logic incorrectly uses list indices to update the monitoring
state. This can lead to data loss if the target list is changed
concurrently. Meanwhile, SJ suggests checking for target existence at the
damon_for_each_target level, and if a target does not exist, simply skip
it and proceed to the next one.
Link: https://lkml.kernel.org/r/20251210052508.264433-1-lienze@kylinos.cn
Signed-off-by: Enze Li <lienze@kylinos.cn>
Suggested-by: SeongJae Park <sj@kernel.org>
Reviewed-by: SeongJae Park <sj@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
The TODO comment in target_nid_store() suggested adding range validation
for target_nid. As discussed in [1], the current behavior of accepting
any integer value is intentional. DAMON sysfs aims to remain flexible,
including supporting users who prepare node IDs before future NUMA hotplug
events.
Because this behavior matches the broader design philosophy of the DAMON
sysfs interface, the TODO comment is now misleading. This patch removes
the comment without introducing any behavioral change.
No functional changes.
Link: https://lkml.kernel.org/r/20251211032722.4928-2-swarajgaikwad1925@gmail.com
Link: https://lore.kernel.org/lkml/20251210150930.57679-1-sj@kernel.org/ [1]
Signed-off-by: Swaraj Gaikwad <swarajgaikwad1925@gmail.com>
Suggested-by: SeongJae Park <sj@kernel.org>
Reviewed-by: SeongJae Park <sj@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
folio_zero_user() does straight zeroing without caring about temporal
locality for caches.
This replaced commit c6ddfb6c58 ("mm, clear_huge_page: move order
algorithm into a separate function") where we cleared a page at a time
converging to the faulting page from the left and the right.
To retain limited temporal locality, split the clearing in three parts:
the faulting page and its immediate neighbourhood, and the regions on its
left and right. We clear the local neighbourhood last to maximize chances
of it sticking around in the cache.
Performance
===
AMD Genoa (EPYC 9J14, cpus=2 sockets * 96 cores * 2 threads,
memory=2.2 TB, L1d=16K/thread, L2=512K/thread, L3=2MB/thread)
vm-scalability/anon-w-seq-hugetlb: this workload runs with 384 processes
(one for each CPU) each zeroing anonymously mapped hugetlb memory which is
then accessed sequentially. stime utime
discontiguous-page 1739.93 ( +- 6.15% ) 1016.61 ( +- 4.75% )
contiguous-page 1853.70 ( +- 2.51% ) 1187.13 ( +- 3.50% )
batched-pages 1756.75 ( +- 2.98% ) 1133.32 ( +- 4.89% )
neighbourhood-last 1725.18 ( +- 4.59% ) 1123.78 ( +- 7.38% )
Both stime and utime largely respond somewhat expectedly. There is a fair
amount of run to run variation but the general trend is that the stime
drops and utime increases. There are a few oddities, like contiguous-page
performing very differently from batched-pages.
As such this is likely an uncommon pattern where we saturate the memory
bandwidth (since all CPUs are running the test) and at the same time are
cache constrained because we access the entire region.
Kernel make (make -j 12 bzImage):
stime utime
discontiguous-page 199.29 ( +- 0.63% ) 1431.67 ( +- .04% )
contiguous-page 193.76 ( +- 0.58% ) 1433.60 ( +- .05% )
batched-pages 193.92 ( +- 0.76% ) 1431.04 ( +- .08% )
neighbourhood-last 194.46 ( +- 0.68% ) 1431.51 ( +- .06% )
For make the utime stays relatively flat with a fairly small (-2.4%)
improvement in the stime.
Link: https://lkml.kernel.org/r/20260107072009.1615991-9-ankur.a.arora@oracle.com
Signed-off-by: Ankur Arora <ankur.a.arora@oracle.com>
Reviewed-by: Raghavendra K T <raghavendra.kt@amd.com>
Tested-by: Raghavendra K T <raghavendra.kt@amd.com>
Acked-by: David Hildenbrand (Red Hat) <david@kernel.org>
Cc: Andy Lutomirski <luto@kernel.org>
Cc: "Borislav Petkov (AMD)" <bp@alien8.de>
Cc: Boris Ostrovsky <boris.ostrovsky@oracle.com>
Cc: "H. Peter Anvin" <hpa@zytor.com>
Cc: Ingo Molnar <mingo@redhat.com>
Cc: Konrad Rzessutek Wilk <konrad.wilk@oracle.com>
Cc: Lance Yang <ioworker0@gmail.com>
Cc: "Liam R. Howlett" <Liam.Howlett@oracle.com>
Cc: Li Zhe <lizhe.67@bytedance.com>
Cc: Lorenzo Stoakes <lorenzo.stoakes@oracle.com>
Cc: Mateusz Guzik <mjguzik@gmail.com>
Cc: Matthew Wilcox (Oracle) <willy@infradead.org>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Suren Baghdasaryan <surenb@google.com>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Vlastimil Babka <vbabka@suse.cz>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Use batch clearing in clear_contig_highpages() instead of clearing a
single page at a time. Exposing larger ranges enables the processor to
optimize based on extent.
To do this we just switch to using clear_user_highpages() which would in
turn use clear_user_pages() or clear_pages().
Batched clearing, when running under non-preemptible models, however, has
latency considerations. In particular, we need periodic invocations of
cond_resched() to keep to reasonable preemption latencies. This is a
problem because the clearing primitives do not, or might not be able to,
call cond_resched() to check if preemption is needed.
So, limit the worst case preemption latency by doing the clearing in units
of no more than PROCESS_PAGES_NON_PREEMPT_BATCH pages. (Preemptible
models already define away most of cond_resched(), so the batch size is
ignored when running under those.)
PROCESS_PAGES_NON_PREEMPT_BATCH: for architectures with "fast" clear-pages
(ones that define clear_pages()), we define it as 32MB worth of pages.
This is meant to be large enough to allow the processor to optimize the
operation and yet small enough that we see reasonable preemption latency
for when this optimization is not possible (ex. slow microarchitectures,
memory bandwidth saturation.)
This specific value also allows for a cacheline allocation elision
optimization (which might help unrelated applications by not evicting
potentially useful cache lines) that kicks in recent generations of AMD
Zen processors at around LLC-size (32MB is a typical size).
At the same time 32MB is small enough that even with poor clearing
bandwidth (say ~10GBps), time to clear 32MB should be well below the
scheduler's default warning threshold
(sysctl_resched_latency_warn_ms=100).
"Slow" architectures (don't have clear_pages()) will continue to use the
base value (single page).
Performance
==
Testing a demand fault workload shows a decent improvement in bandwidth
with pg-sz=1GB. Bandwidth with pg-sz=2MB stays flat.
$ perf bench mem mmap -p $pg-sz -f demand -s 64GB -l 5
contiguous-pages batched-pages
(GBps +- %stdev) (GBps +- %stdev)
pg-sz=2MB 23.58 +- 1.95% 25.34 +- 1.18% + 7.50% preempt=*
pg-sz=1GB 25.09 +- 0.79% 39.22 +- 2.32% + 56.31% preempt=none|voluntary
pg-sz=1GB 25.71 +- 0.03% 52.73 +- 0.20% [#] +110.16% preempt=full|lazy
[#] We perform much better with preempt=full|lazy because, not
needing explicit invocations of cond_resched() we can clear the
full extent (pg-sz=1GB) as a single unit which the processor
can optimize for.
(Unless otherwise noted, all numbers are on AMD Genoa (EPYC 9J13);
region-size=64GB, local node; 2.56 GHz, boost=0.)
Analysis
==
pg-sz=1GB: the improvement we see falls in two buckets depending on the
batch size in use.
For batch-size=32MB the number of cachelines allocated (L1-dcache-loads)
-- which stay relatively flat for smaller batches, start to drop off
because cacheline allocation elision kicks in. And as can be seen below,
at batch-size=1GB, we stop allocating cachelines almost entirely. (Not
visible here but from testing with intermediate sizes, the allocation
change kicks in only at batch-size=32MB and ramps up from there.)
contigous-pages 6,949,417,798 L1-dcache-loads # 883.599 M/sec ( +- 0.01% ) (35.75%)
3,226,709,573 L1-dcache-load-misses # 46.43% of all L1-dcache accesses ( +- 0.05% ) (35.75%)
batched,32MB 2,290,365,772 L1-dcache-loads # 471.171 M/sec ( +- 0.36% ) (35.72%)
1,144,426,272 L1-dcache-load-misses # 49.97% of all L1-dcache accesses ( +- 0.58% ) (35.70%)
batched,1GB 63,914,157 L1-dcache-loads # 17.464 M/sec ( +- 8.08% ) (35.73%)
22,074,367 L1-dcache-load-misses # 34.54% of all L1-dcache accesses ( +- 16.70% ) (35.70%)
The dropoff is also visible in L2 prefetch hits (miss numbers are
on similar lines):
contiguous-pages 3,464,861,312 l2_pf_hit_l2.all # 437.722 M/sec ( +- 0.74% ) (15.69%)
batched,32MB 883,750,087 l2_pf_hit_l2.all # 181.223 M/sec ( +- 1.18% ) (15.71%)
batched,1GB 8,967,943 l2_pf_hit_l2.all # 2.450 M/sec ( +- 17.92% ) (15.77%)
This largely decouples the frontend from the backend since the clearing
operation does not need to wait on loads from memory (we still need
cacheline ownership but that's a shorter path). This is most visible if
we rerun the test above with (boost=1, 3.66 GHz).
$ perf bench mem mmap -p $pg-sz -f demand -s 64GB -l 5
contiguous-pages batched-pages
(GBps +- %stdev) (GBps +- %stdev)
pg-sz=2MB 26.08 +- 1.72% 26.13 +- 0.92% - preempt=*
pg-sz=1GB 26.99 +- 0.62% 48.85 +- 2.19% + 80.99% preempt=none|voluntary
pg-sz=1GB 27.69 +- 0.18% 75.18 +- 0.25% +171.50% preempt=full|lazy
Comparing the batched-pages numbers from the boost=0 ones and these: for a
clock-speed gain of 42% we gain 24.5% for batch-size=32MB and 42.5% for
batch-size=1GB. In comparison the baseline contiguous-pages case and both
the pg-sz=2MB ones are largely backend bound so gain no more than ~10%.
Other platforms tested, Intel Icelakex (Oracle X9) and ARM64 Neoverse-N1
(Ampere Altra) both show an improvement of ~35% for pg-sz=2MB|1GB. The
first goes from around 8GBps to 11GBps and the second from 32GBps to 44
GBPs.
[ankur.a.arora@oracle.com: move the unit computation and make it a const
Link: https://lkml.kernel.org/r/20260108060406.1693853-1-ankur.a.arora@oracle.com
Link: https://lkml.kernel.org/r/20260107072009.1615991-8-ankur.a.arora@oracle.com
Signed-off-by: Ankur Arora <ankur.a.arora@oracle.com>
Acked-by: David Hildenbrand (Red Hat) <david@kernel.org>
Cc: Andy Lutomirski <luto@kernel.org>
Cc: "Borislav Petkov (AMD)" <bp@alien8.de>
Cc: Boris Ostrovsky <boris.ostrovsky@oracle.com>
Cc: "H. Peter Anvin" <hpa@zytor.com>
Cc: Ingo Molnar <mingo@redhat.com>
Cc: Konrad Rzessutek Wilk <konrad.wilk@oracle.com>
Cc: Lance Yang <ioworker0@gmail.com>
Cc: "Liam R. Howlett" <Liam.Howlett@oracle.com>
Cc: Li Zhe <lizhe.67@bytedance.com>
Cc: Lorenzo Stoakes <lorenzo.stoakes@oracle.com>
Cc: Mateusz Guzik <mjguzik@gmail.com>
Cc: Matthew Wilcox (Oracle) <willy@infradead.org>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Raghavendra K T <raghavendra.kt@amd.com>
Cc: Suren Baghdasaryan <surenb@google.com>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Vlastimil Babka <vbabka@suse.cz>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
process_huge_pages(), used to clear hugepages, is optimized for cache
locality. In particular it processes a hugepage in 4KB page units and in
a difficult to predict order: clearing pages in the periphery in a
backwards or forwards direction, then converging inwards to the faulting
page (or page specified via base_addr.)
This helps maximize temporal locality at time of access. However, while
it keeps stores inside a 4KB page sequential, pages are ordered
semi-randomly in a way that is not easy for the processor to predict.
This limits the clearing bandwidth to what's available in a 4KB page.
Consider the baseline bandwidth:
$ perf bench mem mmap -p 2MB -f populate -s 64GB -l 3
# Running 'mem/mmap' benchmark:
# function 'populate' (Eagerly populated mmap())
# Copying 64GB bytes ...
11.791097 GB/sec
(Unless otherwise noted, all numbers are on AMD Genoa (EPYC 9J13);
region-size=64GB, local node; 2.56 GHz, boost=0.)
11.79 GBps amounts to around 323ns/4KB. With memory access latency of
~100ns, that doesn't leave much time to help from, say, hardware
prefetchers.
(Note that since this is a purely write workload, it's reasonable
to assume that the processor does not need to prefetch any cachelines.
However, for a processor to skip the prefetch, it would need to look
at the access pattern, and see that full cachelines were being written.
This might be easily visible if clear_page() was using, say x86 string
instructions; less so if it were using a store loop. In any case, the
existence of these kind predictors or appropriately helpful threshold
values is implementation specific.
Additionally, even when the processor can skip the prefetch, coherence
protocols will still need to establish exclusive ownership
necessitating communication with remote caches.)
With that, the change is quite straight-forward. Instead of clearing
pages discontiguously, clear contiguously: switch to a loop around
clear_user_highpage().
Performance
==
Testing a demand fault workload shows a decent improvement in bandwidth
with pg-sz=2MB. Performance of pg-sz=1GB does not change because it has
always used straight clearing.
$ perf bench mem mmap -p $pg-sz -f demand -s 64GB -l 5
discontiguous-pages contiguous-pages
(baseline)
(GBps +- %stdev) (GBps +- %stdev)
pg-sz=2MB 11.76 +- 1.10% 23.58 +- 1.95% +100.51%
pg-sz=1GB 24.85 +- 2.41% 25.40 +- 1.33% -
Analysis (pg-sz=2MB)
==
At L1 data cache level, nothing changes. The processor continues to
access the same number of cachelines, allocating and missing them as it
writes to them.
discontiguous-pages 7,394,341,051 L1-dcache-loads # 445.172 M/sec ( +- 0.04% ) (35.73%)
3,292,247,227 L1-dcache-load-misses # 44.52% of all L1-dcache accesses ( +- 0.01% ) (35.73%)
contiguous-pages 7,205,105,282 L1-dcache-loads # 861.895 M/sec ( +- 0.02% ) (35.75%)
3,241,584,535 L1-dcache-load-misses # 44.99% of all L1-dcache accesses ( +- 0.00% ) (35.74%)
The L2 prefetcher, however, is now able to prefetch ~22% more cachelines
(L2 prefetch miss rate also goes up significantly showing that we are
backend limited):
discontiguous-pages 2,835,860,245 l2_pf_hit_l2.all # 170.242 M/sec ( +- 0.12% ) (15.65%)
contiguous-pages 3,472,055,269 l2_pf_hit_l2.all # 411.319 M/sec ( +- 0.62% ) (15.67%)
That sill leaves a large gap between the ~22% improvement in prefetch and
the ~100% improvement in bandwidth but better prefetching seems to
streamline the traffic well enough that most of the data starts comes from
the L2 leading to substantially fewer cache-misses at the LLC:
discontiguous-pages 8,493,499,137 cache-references # 511.416 M/sec ( +- 0.15% ) (50.01%)
930,501,344 cache-misses # 10.96% of all cache refs ( +- 0.52% ) (50.01%)
contiguous-pages 9,421,926,416 cache-references # 1.120 G/sec ( +- 0.09% ) (50.02%)
68,787,247 cache-misses # 0.73% of all cache refs ( +- 0.15% ) (50.03%)
In addition, there are a few minor frontend optimizations: clear_pages()
on x86 is now fully inlined, so we don't have a CALL/RET pair (which isn't
free when using RETHUNK speculative execution mitigation as we do on my
test system.) The loop in clear_contig_highpages() is also easier to
predict (especially when handling faults) as compared to that in
process_huge_pages().
discontiguous-pages 980,014,411 branches # 59.005 M/sec (31.26%)
discontiguous-pages 180,897,177 branch-misses # 18.46% of all branches (31.26%)
contiguous-pages 515,630,550 branches # 62.654 M/sec (31.27%)
contiguous-pages 78,039,496 branch-misses # 15.13% of all branches (31.28%)
Note that although clearing contiguously is easier to optimize for the
processor, it does not, sadly, mean that the processor will necessarily
take advantage of it. For instance this change does not result in any
improvement in my tests on Intel Icelakex (Oracle X9), or on ARM64
Neoverse-N1 (Ampere Altra).
Link: https://lkml.kernel.org/r/20260107072009.1615991-7-ankur.a.arora@oracle.com
Signed-off-by: Ankur Arora <ankur.a.arora@oracle.com>
Reviewed-by: Raghavendra K T <raghavendra.kt@amd.com>
Tested-by: Raghavendra K T <raghavendra.kt@amd.com>
Acked-by: David Hildenbrand (Red Hat) <david@kernel.org>
Cc: Andy Lutomirski <luto@kernel.org>
Cc: "Borislav Petkov (AMD)" <bp@alien8.de>
Cc: Boris Ostrovsky <boris.ostrovsky@oracle.com>
Cc: "H. Peter Anvin" <hpa@zytor.com>
Cc: Ingo Molnar <mingo@redhat.com>
Cc: Konrad Rzessutek Wilk <konrad.wilk@oracle.com>
Cc: Lance Yang <ioworker0@gmail.com>
Cc: "Liam R. Howlett" <Liam.Howlett@oracle.com>
Cc: Li Zhe <lizhe.67@bytedance.com>
Cc: Lorenzo Stoakes <lorenzo.stoakes@oracle.com>
Cc: Mateusz Guzik <mjguzik@gmail.com>
Cc: Matthew Wilcox (Oracle) <willy@infradead.org>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Suren Baghdasaryan <surenb@google.com>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Vlastimil Babka <vbabka@suse.cz>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Performance when clearing with string instructions (x86-64-stosq and
similar) can vary significantly based on the chunk-size used.
$ perf bench mem memset -k 4KB -s 4GB -f x86-64-stosq
# Running 'mem/memset' benchmark:
# function 'x86-64-stosq' (movsq-based memset() in arch/x86/lib/memset_64.S)
# Copying 4GB bytes ...
13.748208 GB/sec
$ perf bench mem memset -k 2MB -s 4GB -f x86-64-stosq
# Running 'mem/memset' benchmark:
# function 'x86-64-stosq' (movsq-based memset() in
# arch/x86/lib/memset_64.S)
# Copying 4GB bytes ...
15.067900 GB/sec
$ perf bench mem memset -k 1GB -s 4GB -f x86-64-stosq
# Running 'mem/memset' benchmark:
# function 'x86-64-stosq' (movsq-based memset() in arch/x86/lib/memset_64.S)
# Copying 4GB bytes ...
38.104311 GB/sec
(Both on AMD Milan.)
With a change in chunk-size from 4KB to 1GB, we see the performance go
from 13.7 GB/sec to 38.1 GB/sec. For the chunk-size of 2MB the change
isn't quite as drastic but it is worth adding a clear_page() variant that
can handle contiguous page-extents.
Link: https://lkml.kernel.org/r/20260107072009.1615991-6-ankur.a.arora@oracle.com
Signed-off-by: Ankur Arora <ankur.a.arora@oracle.com>
Tested-by: Raghavendra K T <raghavendra.kt@amd.com>
Reviewed-by: David Hildenbrand (Red Hat) <david@kernel.org>
Cc: Andy Lutomirski <luto@kernel.org>
Cc: "Borislav Petkov (AMD)" <bp@alien8.de>
Cc: Boris Ostrovsky <boris.ostrovsky@oracle.com>
Cc: David Hildenbrand <david@redhat.com>
Cc: "H. Peter Anvin" <hpa@zytor.com>
Cc: Ingo Molnar <mingo@redhat.com>
Cc: Konrad Rzessutek Wilk <konrad.wilk@oracle.com>
Cc: Lance Yang <ioworker0@gmail.com>
Cc: "Liam R. Howlett" <Liam.Howlett@oracle.com>
Cc: Li Zhe <lizhe.67@bytedance.com>
Cc: Lorenzo Stoakes <lorenzo.stoakes@oracle.com>
Cc: Mateusz Guzik <mjguzik@gmail.com>
Cc: Matthew Wilcox (Oracle) <willy@infradead.org>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Suren Baghdasaryan <surenb@google.com>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Vlastimil Babka <vbabka@suse.cz>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Patch series "mm: folio_zero_user: clear page ranges", v11.
This series adds clearing of contiguous page ranges for hugepages.
The series improves on the current discontiguous clearing approach in two
ways:
- clear pages in a contiguous fashion.
- use batched clearing via clear_pages() wherever exposed.
The first is useful because it allows us to make much better use of
hardware prefetchers.
The second, enables advertising the real extent to the processor. Where
specific instructions support it (ex. string instructions on x86; "mops"
on arm64 etc), a processor can optimize based on this because, instead of
seeing a sequence of 8-byte stores, or a sequence of 4KB pages, it sees a
larger unit being operated on.
For instance, AMD Zen uarchs (for extents larger than LLC-size) switch to
a mode where they start eliding cacheline allocation. This is helpful not
just because it results in higher bandwidth, but also because now the
cache is not evicting useful cachelines and replacing them with zeroes.
Demand faulting a 64GB region shows performance improvement:
$ perf bench mem mmap -p $pg-sz -f demand -s 64GB -l 5
baseline +series
(GBps +- %stdev) (GBps +- %stdev)
pg-sz=2MB 11.76 +- 1.10% 25.34 +- 1.18% [*] +115.47% preempt=*
pg-sz=1GB 24.85 +- 2.41% 39.22 +- 2.32% + 57.82% preempt=none|voluntary
pg-sz=1GB (similar) 52.73 +- 0.20% [#] +112.19% preempt=full|lazy
[*] This improvement is because switching to sequential clearing
allows the hardware prefetchers to do a much better job.
[#] For pg-sz=1GB a large part of the improvement is because of the
cacheline elision mentioned above. preempt=full|lazy improves upon
that because, not needing explicit invocations of cond_resched() to
ensure reasonable preemption latency, it can clear the full extent
as a single unit. In comparison the maximum extent used for
preempt=none|voluntary is PROCESS_PAGES_NON_PREEMPT_BATCH (32MB).
When provided the full extent the processor forgoes allocating
cachelines on this path almost entirely.
(The hope is that eventually, in the fullness of time, the lazy
preemption model will be able to do the same job that none or
voluntary models are used for, allowing us to do away with
cond_resched().)
Raghavendra also tested previous version of the series on AMD Genoa and
sees similar improvement [1] with preempt=lazy.
$ perf bench mem map -p $page-size -f populate -s 64GB -l 10
base patched change
pg-sz=2MB 12.731939 GB/sec 26.304263 GB/sec 106.6%
pg-sz=1GB 26.232423 GB/sec 61.174836 GB/sec 133.2%
This patch (of 8):
Let's drop all variants that effectively map to clear_page() and provide
it in a generic variant instead.
We'll use the macro clear_user_page to indicate whether an architecture
provides it's own variant.
Also, clear_user_page() is only called from the generic variant of
clear_user_highpage(), so define it only if the architecture does not
provide a clear_user_highpage(). And, for simplicity define it in
linux/highmem.h.
Note that for parisc, clear_page() and clear_user_page() map to
clear_page_asm(), so we can just get rid of the custom clear_user_page()
implementation. There is a clear_user_page_asm() function on parisc, that
seems to be unused. Not sure what's up with that.
Link: https://lkml.kernel.org/r/20260107072009.1615991-1-ankur.a.arora@oracle.com
Link: https://lkml.kernel.org/r/20260107072009.1615991-2-ankur.a.arora@oracle.com
Signed-off-by: David Hildenbrand <david@redhat.com>
Co-developed-by: Ankur Arora <ankur.a.arora@oracle.com>
Signed-off-by: Ankur Arora <ankur.a.arora@oracle.com>
Cc: Andy Lutomirski <luto@kernel.org>
Cc: Ankur Arora <ankur.a.arora@oracle.com>
Cc: "Borislav Petkov (AMD)" <bp@alien8.de>
Cc: Boris Ostrovsky <boris.ostrovsky@oracle.com>
Cc: David Hildenbrand <david@kernel.org>
Cc: "H. Peter Anvin" <hpa@zytor.com>
Cc: Ingo Molnar <mingo@redhat.com>
Cc: Konrad Rzessutek Wilk <konrad.wilk@oracle.com>
Cc: Lance Yang <ioworker0@gmail.com>
Cc: "Liam R. Howlett" <Liam.Howlett@oracle.com>
Cc: Li Zhe <lizhe.67@bytedance.com>
Cc: Lorenzo Stoakes <lorenzo.stoakes@oracle.com>
Cc: Mateusz Guzik <mjguzik@gmail.com>
Cc: Matthew Wilcox (Oracle) <willy@infradead.org>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Raghavendra K T <raghavendra.kt@amd.com>
Cc: Suren Baghdasaryan <surenb@google.com>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Vlastimil Babka <vbabka@suse.cz>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
