Patch series "Buddy allocator like (or non-uniform) folio split", v10.
This patchset adds a new buddy allocator like (or non-uniform) large folio
split from a order-n folio to order-m with m < n. It reduces
1. the total number of after-split folios from 2^(n-m) to n-m+1;
2. the amount of memory needed for multi-index xarray split from 2^(n/6-m/6) to
n/6-m/6, assuming XA_CHUNK_SHIFT=6;
3. keep more large folios after a split from all order-m folios to
order-(n-1) to order-m folios.
For example, to split an order-9 to order-0, folio split generates 10 (or
11 for anonymous memory) folios instead of 512, allocates 1 xa_node
instead of 8, and leaves 1 order-8, 1 order-7, ..., 1 order-1 and 2
order-0 folios (or 4 order-0 for anonymous memory) instead of 512 order-0
folios.
Instead of duplicating existing split_huge_page*() code, __folio_split()
is introduced as the shared backend code for both
split_huge_page_to_list_to_order() and folio_split(). __folio_split() can
support both uniform split and buddy allocator like (or non-uniform)
split. All existing split_huge_page*() users can be gradually converted
to use folio_split() if possible. In this patchset, I converted
truncate_inode_partial_folio() to use folio_split().
xfstests quick group passed for both tmpfs and xfs. I also
semi-replicated Hugh's test[12] and ran it without any issue for almost 24
hours.
This patch (of 8):
A preparation patch for non-uniform folio split, which always split a
folio into half iteratively, and minimal xarray entry split.
Currently, xas_split_alloc() and xas_split() always split all slots from a
multi-index entry. They cost the same number of xa_node as the
to-be-split slots. For example, to split an order-9 entry, which takes
2^(9-6)=8 slots, assuming XA_CHUNK_SHIFT is 6 (!CONFIG_BASE_SMALL), 8
xa_node are needed. Instead xas_try_split() is intended to be used
iteratively to split the order-9 entry into 2 order-8 entries, then split
one order-8 entry, based on the given index, to 2 order-7 entries, ...,
and split one order-1 entry to 2 order-0 entries. When splitting the
order-6 entry and a new xa_node is needed, xas_try_split() will try to
allocate one if possible. As a result, xas_try_split() would only need 1
xa_node instead of 8.
When a new xa_node is needed during the split, xas_try_split() can try to
allocate one but no more. -ENOMEM will be return if a node cannot be
allocated. -EINVAL will be return if a sibling node is split or cascade
split happens, where two or more new nodes are needed, and these are not
supported by xas_try_split().
xas_split_alloc() and xas_split() split an order-9 to order-0:
---------------------------------
| | | | | | | | |
| 0 | 1 | 2 | 3 | 4 | 5 | 6 | 7 |
| | | | | | | | |
---------------------------------
| | | |
------- --- --- -------
| | ... | |
V V V V
----------- ----------- ----------- -----------
| xa_node | | xa_node | ... | xa_node | | xa_node |
----------- ----------- ----------- -----------
xas_try_split() splits an order-9 to order-0:
---------------------------------
| | | | | | | | |
| 0 | 1 | 2 | 3 | 4 | 5 | 6 | 7 |
| | | | | | | | |
---------------------------------
|
|
V
-----------
| xa_node |
-----------
Link: https://lkml.kernel.org/r/20250307174001.242794-1-ziy@nvidia.com
Link: https://lkml.kernel.org/r/20250307174001.242794-2-ziy@nvidia.com
Signed-off-by: Zi Yan <ziy@nvidia.com>
Cc: Baolin Wang <baolin.wang@linux.alibaba.com>
Cc: David Hildenbrand <david@redhat.com>
Cc: Hugh Dickins <hughd@google.com>
Cc: John Hubbard <jhubbard@nvidia.com>
Cc: Kefeng Wang <wangkefeng.wang@huawei.com>
Cc: Kirill A. Shuemov <kirill.shutemov@linux.intel.com>
Cc: Miaohe Lin <linmiaohe@huawei.com>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Ryan Roberts <ryan.roberts@arm.com>
Cc: Yang Shi <yang@os.amperecomputing.com>
Cc: Yu Zhao <yuzhao@google.com>
Cc: Zi Yan <ziy@nvidia.com>
Cc: Kairui Song <kasong@tencent.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
pte_page() is more expensive than pte_pfn() (often it's defined as
pfn_to_page(pte_pfn())), so it makes sense to do the conversion to pfn
once (by calling folio_pfn()) rather than convert the pfn to a page each
time.
While this is a very small advantage, the main motivation is removing a
reference to folio->page.
Link: https://lkml.kernel.org/r/20250226163131.3795869-1-willy@infradead.org
Signed-off-by: Matthew Wilcox (Oracle) <willy@infradead.org>
Cc: Muchun Song <muchun.song@linux.dev>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
According to the code logic, the first parameter of the sub-function
__get_vm_area_node() should be size instead of real_size.
Then in __get_vm_area_node(), the size will be aligned, so the redundant
alignment operation is deleted.
The use of the real_size variable causes code redundancy, so it is removed
to simplify the code.
The real prefix is generally used to indicate the adjusted value of a
parameter, but according to the code logic, it should indicate the
original value, so it is recommended to rename it to original_align.
Link: https://lkml.kernel.org/r/20250306072131.800499-1-liuye@kylinos.cn
Signed-off-by: Liu Ye <liuye@kylinos.cn>
Reviewed-by: "Uladzislau Rezki (Sony)" <urezki@gmail.com>
Cc: Christop Hellwig <hch@infradead.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Patch series "mm: page_ext: Introduce new iteration API", v3.
Introduction
============
[ Thanks to David Hildenbrand for identifying the root cause of this
issue and proving guidance on how to fix it. The new API idea, bugs
and misconceptions are all mine though ]
Currently, trying to reserve 1G pages with page_owner=on and sparsemem
causes a crash. The reproducer is very simple:
1. Build the kernel with CONFIG_SPARSEMEM=y and the table extensions
2. Pass 'default_hugepagesz=1 page_owner=on' in the kernel command-line
3. Reserve one 1G page at run-time, this should crash (see patch 1 for
backtrace)
[ A crash with page_table_check is also possible, but harder to trigger ]
Apparently, starting with commit cf54f310d0 ("mm/hugetlb: use __GFP_COMP
for gigantic folios") we now pass the full allocation order to page
extension clients and the page extension implementation assumes that all
PFNs of an allocation range will be stored in the same memory section (which
is not true for 1G pages).
To fix this, this series introduces a new iteration API for page extension
objects. The API checks if the next page extension object can be retrieved
from the current section or if it needs to look up for it in another
section.
Please, find all details in patch 1.
I tested this series on arm64 and x86 by reserving 1G pages at run-time
and doing kernel builds (always with page_owner=on and page_table_check=on).
This patch (of 3):
The page extension implementation assumes that all page extensions of a
given page order are stored in the same memory section. The function
page_ext_next() relies on this assumption by adding an offset to the
current object to return the next adjacent page extension.
This behavior works as expected for flatmem but fails for sparsemem when
using 1G pages. The commit cf54f310d0 ("mm/hugetlb: use __GFP_COMP for
gigantic folios") exposes this issue, making it possible for a crash when
using page_owner or page_table_check page extensions.
The problem is that for 1G pages, the page extensions may span memory
section boundaries and be stored in different memory sections. This issue
was not visible before commit cf54f310d0 ("mm/hugetlb: use __GFP_COMP
for gigantic folios") because alloc_contig_pages() never passed more than
MAX_PAGE_ORDER to post_alloc_hook(). However, the series introducing
mentioned commit changed this behavior allowing the full 1G page order to
be passed.
Reproducer:
1. Build the kernel with CONFIG_SPARSEMEM=y and table extensions
support
2. Pass 'default_hugepagesz=1 page_owner=on' in the kernel command-line
3. Reserve one 1G page at run-time, this should crash (backtrace below)
To address this issue, this commit introduces a new API for iterating
through page extensions. The main iteration macro is for_each_page_ext()
and it must be called with the RCU read lock taken. Here's an usage
example:
"""
struct page_ext_iter iter;
struct page_ext *page_ext;
...
rcu_read_lock();
for_each_page_ext(page, 1 << order, page_ext, iter) {
struct my_page_ext *obj = get_my_page_ext_obj(page_ext);
...
}
rcu_read_unlock();
"""
The loop construct uses page_ext_iter_next() which checks to see if we
have crossed sections in the iteration. In this case,
page_ext_iter_next() retrieves the next page_ext object from another
section.
Thanks to David Hildenbrand for helping identify the root cause and
providing suggestions on how to fix and optmize the solution (final
implementation and bugs are all mine through).
Lastly, here's the backtrace, without kasan you can get random crashes:
[ 76.052526] BUG: KASAN: slab-out-of-bounds in __update_page_owner_handle+0x238/0x298
[ 76.060283] Write of size 4 at addr ffff07ff96240038 by task tee/3598
[ 76.066714]
[ 76.068203] CPU: 88 UID: 0 PID: 3598 Comm: tee Kdump: loaded Not tainted 6.13.0-rep1 #3
[ 76.076202] Hardware name: WIWYNN Mt.Jade Server System B81.030Z1.0007/Mt.Jade Motherboard, BIOS 2.10.20220810 (SCP: 2.10.20220810) 2022/08/10
[ 76.088972] Call trace:
[ 76.091411] show_stack+0x20/0x38 (C)
[ 76.095073] dump_stack_lvl+0x80/0xf8
[ 76.098733] print_address_description.constprop.0+0x88/0x398
[ 76.104476] print_report+0xa8/0x278
[ 76.108041] kasan_report+0xa8/0xf8
[ 76.111520] __asan_report_store4_noabort+0x20/0x30
[ 76.116391] __update_page_owner_handle+0x238/0x298
[ 76.121259] __set_page_owner+0xdc/0x140
[ 76.125173] post_alloc_hook+0x190/0x1d8
[ 76.129090] alloc_contig_range_noprof+0x54c/0x890
[ 76.133874] alloc_contig_pages_noprof+0x35c/0x4a8
[ 76.138656] alloc_gigantic_folio.isra.0+0x2c0/0x368
[ 76.143616] only_alloc_fresh_hugetlb_folio.isra.0+0x24/0x150
[ 76.149353] alloc_pool_huge_folio+0x11c/0x1f8
[ 76.153787] set_max_huge_pages+0x364/0xca8
[ 76.157961] __nr_hugepages_store_common+0xb0/0x1a0
[ 76.162829] nr_hugepages_store+0x108/0x118
[ 76.167003] kobj_attr_store+0x3c/0x70
[ 76.170745] sysfs_kf_write+0xfc/0x188
[ 76.174492] kernfs_fop_write_iter+0x274/0x3e0
[ 76.178927] vfs_write+0x64c/0x8e0
[ 76.182323] ksys_write+0xf8/0x1f0
[ 76.185716] __arm64_sys_write+0x74/0xb0
[ 76.189630] invoke_syscall.constprop.0+0xd8/0x1e0
[ 76.194412] do_el0_svc+0x164/0x1e0
[ 76.197891] el0_svc+0x40/0xe0
[ 76.200939] el0t_64_sync_handler+0x144/0x168
[ 76.205287] el0t_64_sync+0x1ac/0x1b0
Link: https://lkml.kernel.org/r/cover.1741301089.git.luizcap@redhat.com
Link: https://lkml.kernel.org/r/a45893880b7e1601082d39d2c5c8b50bcc096305.1741301089.git.luizcap@redhat.com
Fixes: cf54f310d0 ("mm/hugetlb: use __GFP_COMP for gigantic folios")
Signed-off-by: Luiz Capitulino <luizcap@redhat.com>
Acked-by: David Hildenbrand <david@redhat.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Luiz Capitulino <luizcap@redhat.com>
Cc: Muchun Song <muchun.song@linux.dev>
Cc: Pasha Tatashin <pasha.tatashin@soleen.com>
Cc: Yu Zhao <yuzhao@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
damon_sysfs_cmd_request is DAMON sysfs interface's own synchronization
mechanism for accessing DAMON internal data via damon_callback hooks. All
the users are now migrated to damon_call() and damos_walk(), so nobody
really uses it. No one writes to the data structure but reading code is
still remained. Remove the reading code and the entire data structure.
Link: https://lkml.kernel.org/r/20250306175908.66300-8-sj@kernel.org
Signed-off-by: SeongJae Park <sj@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
damon_sysfs_cmd_request_callback() is the damon_callback hook functions
that were used to handle user requests that need to read and/or write
DAMON internal data. All the usages are now updated to use damon_call()
or damos_walk(), though. Remove it and its callers.
Link: https://lkml.kernel.org/r/20250306175908.66300-7-sj@kernel.org
Signed-off-by: SeongJae Park <sj@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
damon_sysfs_handle_cmd() handles user requests that it can directly handle
on its own. For requests that need to be handled from damon_callback
hooks, it uses DAMON sysfs interface's own synchronous damon_callback
hooks management mechanism, namely damon_sysfs_cmd_request. Now all user
requests are handled without damon_callback hooks, so
damon_sysfs_cmd_request client code in damon_sysfs_andle_cmd() does
nothing in real. Remove the unnecessary code.
Link: https://lkml.kernel.org/r/20250306175908.66300-6-sj@kernel.org
Signed-off-by: SeongJae Park <sj@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
DAMON sysfs interface is using damon_callback->after_aggregation hook with
its self-implemented synchronization mechanism for the hook. It is
inefficient, complicated, and take up to one aggregation interval to
complete, which can be long on some configs.
Use damon_call() instead. It provides a synchronization mechanism that
built inside DAMON's core layer, so more efficient than DAMON sysfs
interface's own one. Also it isolates the implementation inside the core
layer, and hence it makes the code easier to maintain. Finally, it takes
up to one sampling interval, which is much shorter than the aggregation
interval in common setups.
Link: https://lkml.kernel.org/r/20250306175908.66300-5-sj@kernel.org
Signed-off-by: SeongJae Park <sj@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Currently all DAMON kernel API callers do online DAMON parameters commit
from damon_callback->after_aggregation because only those are safe place
to call the DAMON monitoring attributes update function, namely
damon_set_attrs().
Because damon_callback hooks provide no synchronization, the callers work
in asynchronous ways or implement their own inefficient and complicated
synchronization mechanisms. It also means online DAMON parameters commit
can take up to one aggregation interval. On large systems having long
aggregation intervals, that can be too slow. The synchronization can be
done in more efficient and simple way while removing the latency
constraint if it can be done using damon_call().
The fact that damon_call() can be executed in the middle of the
aggregation makes damon_set_attrs() unsafe to be called from it, though.
Two real problems can occur in the case. First, converting the not yet
completely aggregated nr_accesses for new user-set intervals can arguably
degrade the accuracy or at least make the logic complicated. Second,
kdamond_reset_aggregated() will not be called after the monitoring results
update, so next aggregation starts from unclean state. This can result in
inconsistent and unexpected nr_accesses_bp.
Make it safe as follows. Catch the middle-of-the-aggregation case from
damon_set_attrs() by checking the passed_sample_intervals and
next_aggregationsis of the context. And pass the information to
nr_accesses conversion logic. The logic works as before if it is not the
case (called after the current aggregation is completed). If it is the
case (committing parameters in the middle of the aggregation), it drops
the nr_accesses information that so far aggregated, and make the status
same to the beginning of this aggregation, but as if the last aggregation
was started with the updated sampling/aggregation intervals.
The middle-of-aggregastion check introduce yet another edge case, though.
This happens because kdamond_tune_intervals() can also call
damon_set_attrs() with the middle-of-aggregation check. Consider
damon_call() for parameters commit and kdamond_tune_intervals() are called
in same iteration of kdamond main loop. Because kdamond_tune_interval()
is called for aggregation intervals, it should be the end of the
aggregation. The first damon_set_attrs() call from kdamond_call()
understands it is the end of the aggregation and correctly handle it.
But, because the damon_set_attrs() updated next_aggregation_sis of the
context. Hence, the second damon_set_attrs() invocation from
kdamond_tune_interval() believes it is called in the middle of the
aggregation. It therefore resets aggregated information so far. After
that, kdamond_reset_interval() is called and double-reset the aggregated
information. Avoid this case, too, by setting the next_aggregation_sis
before kdamond_tune_intervals() is invoked.
Link: https://lkml.kernel.org/r/20250306175908.66300-4-sj@kernel.org
Signed-off-by: SeongJae Park <sj@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
kdamond_call() callers may iterate the regions, so better to call it when
the number of regions is as small as possible. It is when
kdamond_merge_regions() is finished. Invoke it on the point.
This change is also aimed to make future changes for carrying online
parameters commit with damon_call() easier. The commit operation should
be able to make sequence between other aggregation interval based
operations including regioins merging and aggregation reset. Placing
damon_call() invocation after the regions merging makes the sequence
handling simpler.
Link: https://lkml.kernel.org/r/20250306175908.66300-3-sj@kernel.org
Signed-off-by: SeongJae Park <sj@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Patch series "mm/damon/sysfs: commit parameters online via damon_call()".
Due to the lack of ways to synchronously access DAMON internal data, DAMON
sysfs interface is using damon_callback hooks with its own synchronization
mechanism. The mechanism is built on top of damon_callback hooks in an
ineifficient and complicated way.
Patch series "mm/damon: replace most damon_callback usages in sysfs with
new core functions", which starts with commit e035320fd3
("mm/damon/sysfs-schemes: remove unnecessary schemes existence check in
damon_sysfs_schemes_clear_regions()") introduced two new DAMON kernel API
functions that providing the synchronous access, replaced most
damon_callback hooks usage in DAMON sysfs interface, and cleaned up
unnecessary code.
Continue the replacement and cleanup works. Update the last DAMON sysfs'
usage of its own synchronization mechanism, namely online DAMON parameters
commit, to use damon_call() instead of the damon_callback hooks and the
hard-to-maintain core-external synchronization mechanism. Then remove the
no more be used code due to the change, and more unused code that just not
yet cleaned up.
The first four patches (patches 1-4) of this series makes DAMON sysfs
interface's online parameters commit to use damon_call(). Then, following
three patches (patches 5-7) remove the DAMON sysfs interface's own
synchronization mechanism and its usages, which is no more be used by
anyone due to the first four patches. Finally, six patches (8-13) do more
cleanup of outdated comment and unused code.
This patch (of 13):
Online DAMON parameters commit via DAMON sysfs interface can make kdamond
stop. This behavior was made because it can make the implementation
simpler. The implementation tries committing the parameter without
validation. If it finds something wrong in the middle of the parameters
update, it returns error without reverting the partially committed
parameters back. It is safe though, since it immediately breaks kdamond
main loop in the case of the error return.
Users can make the wrong parameters by mistake, though. Stopping kdamond
in the case is not very useful behavior. Also this makes it difficult to
utilize damon_call() instead of damon_callback hook for online parameters
update, since damon_call() cannot immediately break kdamond main loop in
the middle.
Validate the input parameters and return error when it fails before
starting parameters updates. In case of mistakenly wrong parameters,
kdamond can continue running with the old and valid parameters.
Link: https://lkml.kernel.org/r/20250306175908.66300-1-sj@kernel.org
Link: https://lkml.kernel.org/r/20250306175908.66300-2-sj@kernel.org
Signed-off-by: SeongJae Park <sj@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Currently, we crash the kernel when a decompression failure occurs in
zswap (either because of memory corruption, or a bug in the compression
algorithm). This is overkill. We should only SIGBUS the unfortunate
process asking for the zswap entry on zswap load, and skip the corrupted
entry in zswap writeback.
See [1] for a recent upstream discussion about this.
The zswap writeback case is relatively straightforward to fix. For the
zswap_load() case, we change the return behavior:
* Return 0 on success.
* Return -ENOENT (with the folio locked) if zswap does not own the
swapped out content.
* Return -EIO if zswap owns the swapped out content, but encounters a
decompression failure for some reasons. The folio will be unlocked,
but not be marked up-to-date, which will eventually cause the process
requesting the page to SIGBUS (see the handling of not-up-to-date
folio in do_swap_page() in mm/memory.c), without crashing the kernel.
* Return -EINVAL if we encounter a large folio, as large folio should
not be swapped in while zswap is being used. Similar to the -EIO case,
we also unlock the folio but do not mark it as up-to-date to SIGBUS
the faulting process.
As a side effect, we require one extra zswap tree traversal in the load
and writeback paths. Quick benchmarking on a kernel build test shows no
performance difference:
With the new scheme:
real: mean: 125.1s, stdev: 0.12s
user: mean: 3265.23s, stdev: 9.62s
sys: mean: 2156.41s, stdev: 13.98s
The old scheme:
real: mean: 125.78s, stdev: 0.45s
user: mean: 3287.18s, stdev: 5.95s
sys: mean: 2177.08s, stdev: 26.52s
[nphamcs@gmail.com: fix documentation of zswap_load()]
Link: https://lkml.kernel.org/r/20250306222453.1269456-1-nphamcs@gmail.com
Link: https://lore.kernel.org/all/ZsiLElTykamcYZ6J@casper.infradead.org/ [1]
Link: https://lkml.kernel.org/r/20250306205011.784787-1-nphamcs@gmail.com
Signed-off-by: Nhat Pham <nphamcs@gmail.com>
Suggested-by: Matthew Wilcox <willy@infradead.org>
Suggested-by: Yosry Ahmed <yosry.ahmed@linux.dev>
Suggested-by: Johannes Weiner <hannes@cmpxchg.org>
Reviewed-by: Chengming Zhou <chengming.zhou@linux.dev>
Acked-by: Johannes Weiner <hannes@cmpxchg.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
In alloc_surplus_hugetlb_folio(), we increase nr_huge_pages and
surplus_huge_pages separately. In the middle window, if we set
nr_hugepages to smaller and satisfy count < persistent_huge_pages(h), the
surplus_huge_pages will be increased by adjust_pool_surplus().
After adding delay in the middle window, we can reproduce the problem
easily by following step:
1. echo 3 > /proc/sys/vm/nr_overcommit_hugepages
2. mmap two hugepages. When nr_huge_pages=2 and surplus_huge_pages=1,
goto step 3.
3. echo 0 > /proc/sys/vm/nr_huge_pages
Finally, nr_huge_pages is less than surplus_huge_pages.
To fix the problem, call only_alloc_fresh_hugetlb_folio() instead and
move down __prep_account_new_huge_page() into the hugetlb_lock.
Link: https://lkml.kernel.org/r/20250305035409.2391344-1-liushixin2@huawei.com
Fixes: 0c397daea1 ("mm, hugetlb: further simplify hugetlb allocation API")
Signed-off-by: Liu Shixin <liushixin2@huawei.com>
Acked-by: Peter Xu <peterx@redhat.com>
Acked-by: Oscar Salvador <osalvador@suse.de>
Cc: David Hildenbrand <david@redhat.com>
Cc: Kefeng Wang <wangkefeng.wang@huawei.com>
Cc: Liu Shixin <liushixin2@huawei.com>
Cc: Muchun Song <muchun.song@linux.dev>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Unlike their name and assumed purposes, {core,ops}_filters DAMOS sysfs
directories are allowing installing any type of filters. As a first step
for preventing such wrong installments, add information about filters that
handled by what layer should the installed to the given filters directory
in the DAMOS sysfs internal data structures.
Link: https://lkml.kernel.org/r/20250305222733.59089-6-sj@kernel.org
Signed-off-by: SeongJae Park <sj@kernel.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Connect user inputs for files under core_filters and ops_filters with
DAMON, so that the files can really function. Becasuse {core,ops}_filters
are easier to be managed in terms of expecting filters evaluation order,
add filters in {core,ops}_filters before 'filters' directory.
Link: https://lkml.kernel.org/r/20250305222733.59089-4-sj@kernel.org
Signed-off-by: SeongJae Park <sj@kernel.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Implement two DAMOS sysfs directories for managing core and operations
layer handled filters separately. Those are named as 'core_filters' and
'ops_filters', and have files hierarchy same to 'filters'. This commit is
only populating and cleaning up the directories, not really connecting the
files with DAMON. Following changes will make the connections.
Link: https://lkml.kernel.org/r/20250305222733.59089-3-sj@kernel.org
Signed-off-by: SeongJae Park <sj@kernel.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Patch series "mm/damon: add sysfs dirs for managing DAMOS filters based on
handling layers".
DAMOS filters are categorized into two groups based on their handling
layers, namely core and operations layers. The categorization affects
when each filter is evaluated. Core layer handled filters are evaluated
first. The order meant nothing before, but introduction of allow filters
changed that.
DAMOS sysfs interface provides single directory for filters, namely
'filters'. Users can install any filters in any order there. DAMON will
internally categorize those into core and operations layer handled ones,
and apply the evaluation order rule. The ordering rule is clearly
documented. But the interface could still confuse users since it is
allowed to install filters on the directory in mixed ways.
Add two sysfs directories for managing filters by handling layers, namely
'core_filters' and 'ops_filters' for filters that handled by core and
operations layer, respectively. Those are avoided to be used for
installing filters that not handled by the assumed layers.
For backward compatibility, keep 'filters' directory with its curernt
behavior. Filters installed in the directory will be added to DAMON after
those of 'core_filters' and 'ops_filters' directories, with the automatic
categorizations. Also recommend users to use the new directories while
noticing 'filters' directory could be deprecated in future on the usage
documents.
Note that new directories provide all features that were provided with
'filters', but just in a more clear way. Deprecating 'filters' in future
will hence not make an irreversal feature loss.
This patch (of 8):
damon_sysfs_scheme_set_filters() is using a hard-coded directory name,
"filters". Refactor for general named directories of same files
hierarchy, to use from upcoming changes for adding sibling directories
having files same to those of "filters", and named as "core_filters" and
"ops_filters".
[arnd@arndb.deL avoid Wformat-security warning]
Link: https://lkml.kernel.org/r/20250310135142.4176976-1-arnd@kernel.org
Link: https://lkml.kernel.org/r/20250305222733.59089-1-sj@kernel.org
Link: https://lkml.kernel.org/r/20250305222733.59089-2-sj@kernel.org
Signed-off-by: SeongJae Park <sj@kernel.org>
Signed-off-by: Arnd Bergmann <arnd@arndb.de>
Cc: Jonathan Corbet <corbet@lwn.net>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Everything is in place to stop using the per-page mapcounts in large
folios: the mapcount of tail pages will always be logically 0 (-1 value),
just like it currently is for hugetlb folios already, and the page
mapcount of the head page is either 0 (-1 value) or contains a page type
(e.g., hugetlb).
Maintaining _nr_pages_mapped without per-page mapcounts is impossible, so
that one also has to go with CONFIG_NO_PAGE_MAPCOUNT.
There are two remaining implications:
(1) Per-node, per-cgroup and per-lruvec stats of "NR_ANON_MAPPED"
("mapped anonymous memory") and "NR_FILE_MAPPED"
("mapped file memory"):
As soon as any page of the folio is mapped -- folio_mapped() -- we
now account the complete folio as mapped. Once the last page is
unmapped -- !folio_mapped() -- we account the complete folio as
unmapped.
This implies that ...
* "AnonPages" and "Mapped" in /proc/meminfo and
/sys/devices/system/node/*/meminfo
* cgroup v2: "anon" and "file_mapped" in "memory.stat" and
"memory.numa_stat"
* cgroup v1: "rss" and "mapped_file" in "memory.stat" and
"memory.numa_stat
... can now appear higher than before. But note that these folios do
consume that memory, simply not all pages are actually currently
mapped.
It's worth nothing that other accounting in the kernel (esp. cgroup
charging on allocation) is not affected by this change.
[why oh why is "anon" called "rss" in cgroup v1]
(2) Detecting partial mappings
Detecting whether anon THPs are partially mapped gets a bit more
unreliable. As long as a single MM maps such a large folio
("exclusively mapped"), we can reliably detect it. Especially before
fork() / after a short-lived child process quit, we will detect
partial mappings reliably, which is the common case.
In essence, if the average per-page mapcount in an anon THP is < 1,
we know for sure that we have a partial mapping.
However, as soon as multiple MMs are involved, we might miss detecting
partial mappings: this might be relevant with long-lived child
processes. If we have a fully-mapped anon folio before fork(), once
our child processes and our parent all unmap (zap/COW) the same pages
(but not the complete folio), we might not detect the partial mapping.
However, once the child processes quit we would detect the partial
mapping.
How relevant this case is in practice remains to be seen.
Swapout/migration will likely mitigate this.
In the future, RMAP walkers could check for that for that case
(e.g., when collecting access bits during reclaim) and simply flag
them for deferred-splitting.
Link: https://lkml.kernel.org/r/20250303163014.1128035-21-david@redhat.com
Signed-off-by: David Hildenbrand <david@redhat.com>
Cc: Andy Lutomirks^H^Hski <luto@kernel.org>
Cc: Borislav Betkov <bp@alien8.de>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Ingo Molnar <mingo@redhat.com>
Cc: Jann Horn <jannh@google.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>
Cc: Lance Yang <ioworker0@gmail.com>
Cc: Liam Howlett <liam.howlett@oracle.com>
Cc: Lorenzo Stoakes <lorenzo.stoakes@oracle.com>
Cc: Matthew Wilcow (Oracle) <willy@infradead.org>
Cc: Michal Koutn <mkoutny@suse.com>
Cc: Muchun Song <muchun.song@linux.dev>
Cc: tejun heo <tj@kernel.org>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Zefan Li <lizefan.x@bytedance.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Let's implement an alternative when per-page mapcounts in large folios are
no longer maintained -- soon with CONFIG_NO_PAGE_MAPCOUNT.
When computing the output for smaps / smaps_rollups, in particular when
calculating the USS (Unique Set Size) and the PSS (Proportional Set Size),
we still rely on per-page mapcounts.
To determine private vs. shared, we'll use folio_likely_mapped_shared(),
similar to how we handle PM_MMAP_EXCLUSIVE. Similarly, we might now
under-estimate the USS and count pages towards "shared" that are actually
"private" ("exclusively mapped").
When calculating the PSS, we'll now also use the average per-page mapcount
for large folios: this can result in both, an over-estimation and an
under-estimation of the PSS. The difference is not expected to matter
much in practice, but we'll have to learn as we go.
We can now provide folio_precise_page_mapcount() only with
CONFIG_PAGE_MAPCOUNT, and remove one of the last users of per-page
mapcounts when CONFIG_NO_PAGE_MAPCOUNT is enabled.
Document the new behavior.
Link: https://lkml.kernel.org/r/20250303163014.1128035-20-david@redhat.com
Signed-off-by: David Hildenbrand <david@redhat.com>
Cc: Andy Lutomirks^H^Hski <luto@kernel.org>
Cc: Borislav Betkov <bp@alien8.de>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Ingo Molnar <mingo@redhat.com>
Cc: Jann Horn <jannh@google.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>
Cc: Lance Yang <ioworker0@gmail.com>
Cc: Liam Howlett <liam.howlett@oracle.com>
Cc: Lorenzo Stoakes <lorenzo.stoakes@oracle.com>
Cc: Matthew Wilcow (Oracle) <willy@infradead.org>
Cc: Michal Koutn <mkoutny@suse.com>
Cc: Muchun Song <muchun.song@linux.dev>
Cc: tejun heo <tj@kernel.org>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Zefan Li <lizefan.x@bytedance.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
