The current implementation of the memcg accounting of the percpu
memory is based on the idea of having two separate sets of chunks for
accounted and non-accounted memory. This approach has an advantage
of not wasting any extra memory for memcg data for non-accounted
chunks, however it complicates the code and leads to a higher chunks
number due to a lower chunk utilization.
Instead of having two chunk types it's possible to declare all* chunks
memcg-aware unless the kernel memory accounting is disabled globally
by a boot option. The size of objcg_array is usually small in
comparison to chunks themselves (it obviously depends on the number of
CPUs), so even if some chunk will have no accounted allocations, the
memory waste isn't significant and will likely be compensated by
a higher chunk utilization. Also, with time more and more percpu
allocations will likely become accounted.
* The first chunk is initialized before the memory cgroup subsystem,
so we don't know for sure whether we need to allocate obj_cgroups.
Because it's small, let's make it free for use. Then we don't need
to allocate obj_cgroups for it.
Signed-off-by: Roman Gushchin <guro@fb.com>
Signed-off-by: Dennis Zhou <dennis@kernel.org>
From Roman ("percpu: partial chunk depopulation"):
In our [Facebook] production experience the percpu memory allocator is
sometimes struggling with returning the memory to the system. A typical
example is a creation of several thousands memory cgroups (each has
several chunks of the percpu data used for vmstats, vmevents,
ref counters etc). Deletion and complete releasing of these cgroups
doesn't always lead to a shrinkage of the percpu memory, so that
sometimes there are several GB's of memory wasted.
The underlying problem is the fragmentation: to release an underlying
chunk all percpu allocations should be released first. The percpu
allocator tends to top up chunks to improve the utilization. It means
new small-ish allocations (e.g. percpu ref counters) are placed onto
almost filled old-ish chunks, effectively pinning them in memory.
This patchset solves this problem by implementing a partial depopulation
of percpu chunks: chunks with many empty pages are being asynchronously
depopulated and the pages are returned to the system.
To illustrate the problem the following script can be used:
--
cd /sys/fs/cgroup
mkdir percpu_test
echo "+memory" > percpu_test/cgroup.subtree_control
cat /proc/meminfo | grep Percpu
for i in `seq 1 1000`; do
mkdir percpu_test/cg_"${i}"
for j in `seq 1 10`; do
mkdir percpu_test/cg_"${i}"_"${j}"
done
done
cat /proc/meminfo | grep Percpu
for i in `seq 1 1000`; do
for j in `seq 1 10`; do
rmdir percpu_test/cg_"${i}"_"${j}"
done
done
sleep 10
cat /proc/meminfo | grep Percpu
for i in `seq 1 1000`; do
rmdir percpu_test/cg_"${i}"
done
rmdir percpu_test
--
It creates 11000 memory cgroups and removes every 10 out of 11.
It prints the initial size of the percpu memory, the size after
creating all cgroups and the size after deleting most of them.
Results:
vanilla:
./percpu_test.sh
Percpu: 7488 kB
Percpu: 481152 kB
Percpu: 481152 kB
with this patchset applied:
./percpu_test.sh
Percpu: 7488 kB
Percpu: 481408 kB
Percpu: 135552 kB
The total size of the percpu memory was reduced by more than 3.5 times.
This patch:
This patch implements partial depopulation of percpu chunks.
As of now, a chunk can be depopulated only as a part of the final
destruction, if there are no more outstanding allocations. However
to minimize a memory waste it might be useful to depopulate a
partially filed chunk, if a small number of outstanding allocations
prevents the chunk from being fully reclaimed.
This patch implements the following depopulation process: it scans
over the chunk pages, looks for a range of empty and populated pages
and performs the depopulation. To avoid races with new allocations,
the chunk is previously isolated. After the depopulation the chunk is
sidelined to a special list or freed. New allocations prefer using
active chunks to sidelined chunks. If a sidelined chunk is used, it is
reintegrated to the active lists.
The depopulation is scheduled on the free path if the chunk is all of
the following:
1) has more than 1/4 of total pages free and populated
2) the system has enough free percpu pages aside of this chunk
3) isn't the reserved chunk
4) isn't the first chunk
If it's already depopulated but got free populated pages, it's a good
target too. The chunk is moved to a special slot,
pcpu_to_depopulate_slot, chunk->isolated is set, and the balance work
item is scheduled. On isolation, these pages are removed from the
pcpu_nr_empty_pop_pages. It is constantly replaced to the
to_depopulate_slot when it meets these qualifications.
pcpu_reclaim_populated() iterates over the to_depopulate_slot until it
becomes empty. The depopulation is performed in the reverse direction to
keep populated pages close to the beginning. Depopulated chunks are
sidelined to preferentially avoid them for new allocations. When no
active chunk can suffice a new allocation, sidelined chunks are first
checked before creating a new chunk.
Signed-off-by: Roman Gushchin <guro@fb.com>
Co-developed-by: Dennis Zhou <dennis@kernel.org>
Signed-off-by: Dennis Zhou <dennis@kernel.org>
Tested-by: Pratik Sampat <psampat@linux.ibm.com>
Signed-off-by: Dennis Zhou <dennis@kernel.org>
Percpu memory is becoming more and more widely used by various subsystems,
and the total amount of memory controlled by the percpu allocator can make
a good part of the total memory.
As an example, bpf maps can consume a lot of percpu memory, and they are
created by a user. Also, some cgroup internals (e.g. memory controller
statistics) can be quite large. On a machine with many CPUs and big
number of cgroups they can consume hundreds of megabytes.
So the lack of memcg accounting is creating a breach in the memory
isolation. Similar to the slab memory, percpu memory should be accounted
by default.
To implement the perpcu accounting it's possible to take the slab memory
accounting as a model to follow. Let's introduce two types of percpu
chunks: root and memcg. What makes memcg chunks different is an
additional space allocated to store memcg membership information. If
__GFP_ACCOUNT is passed on allocation, a memcg chunk should be be used.
If it's possible to charge the corresponding size to the target memory
cgroup, allocation is performed, and the memcg ownership data is recorded.
System-wide allocations are performed using root chunks, so there is no
additional memory overhead.
To implement a fast reparenting of percpu memory on memcg removal, we
don't store mem_cgroup pointers directly: instead we use obj_cgroup API,
introduced for slab accounting.
[akpm@linux-foundation.org: fix CONFIG_MEMCG_KMEM=n build errors and warning]
[akpm@linux-foundation.org: move unreachable code, per Roman]
[cuibixuan@huawei.com: mm/percpu: fix 'defined but not used' warning]
Link: http://lkml.kernel.org/r/6d41b939-a741-b521-a7a2-e7296ec16219@huawei.com
Signed-off-by: Roman Gushchin <guro@fb.com>
Signed-off-by: Bixuan Cui <cuibixuan@huawei.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Reviewed-by: Shakeel Butt <shakeelb@google.com>
Acked-by: Dennis Zhou <dennis@kernel.org>
Cc: Christoph Lameter <cl@linux.com>
Cc: David Rientjes <rientjes@google.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Pekka Enberg <penberg@kernel.org>
Cc: Tejun Heo <tj@kernel.org>
Cc: Tobin C. Harding <tobin@kernel.org>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Waiman Long <longman@redhat.com>
Cc: Bixuan Cui <cuibixuan@huawei.com>
Cc: Michal Koutný <mkoutny@suse.com>
Cc: Stephen Rothwell <sfr@canb.auug.org.au>
Link: http://lkml.kernel.org/r/20200623184515.4132564-3-guro@fb.com
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Currently there are 3 emails tied to me in the kernel tree, I'd rather
dennis@kernel.org be the only one.
Signed-off-by: Dennis Zhou <dennis@kernel.org>
Based on 1 normalized pattern(s):
this file is released under the gplv2
extracted by the scancode license scanner the SPDX license identifier
GPL-2.0-only
has been chosen to replace the boilerplate/reference in 68 file(s).
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Reviewed-by: Armijn Hemel <armijn@tjaldur.nl>
Reviewed-by: Allison Randal <allison@lohutok.net>
Cc: linux-spdx@vger.kernel.org
Link: https://lkml.kernel.org/r/20190531190114.292346262@linutronix.de
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
As mentioned in the last patch, a chunk's hints are no different than a
block just responsible for more bits. This converts chunk level hints to
use a pcpu_block_md to maintain them. This lets us reuse the same hint
helper functions as a block. The left_free and right_free are unused by
the chunk's pcpu_block_md.
Signed-off-by: Dennis Zhou <dennis@kernel.org>
Reviewed-by: Peng Fan <peng.fan@nxp.com>
This patch fixes the starting offset used when scanning chunks to
compute the chunk statistics. The value start_offset (and end_offset)
are managed in bytes while the traversal occurs over bits. Thus for the
reserved and dynamic chunk, it may incorrectly skip over the initial
allocations.
Signed-off-by: Dennis Zhou <dennisszhou@gmail.com>
Signed-off-by: Tejun Heo <tj@kernel.org>
This patch adds first_bit to keep track of the first free bit in the
bitmap. This hint helps prevent scanning of fully allocated blocks.
Signed-off-by: Dennis Zhou <dennisszhou@gmail.com>
Reviewed-by: Josef Bacik <jbacik@fb.com>
Signed-off-by: Tejun Heo <tj@kernel.org>
The percpu memory allocator is experiencing scalability issues when
allocating and freeing large numbers of counters as in BPF.
Additionally, there is a corner case where iteration is triggered over
all chunks if the contig_hint is the right size, but wrong alignment.
This patch replaces the area map allocator with a basic bitmap allocator
implementation. Each subsequent patch will introduce new features and
replace full scanning functions with faster non-scanning options when
possible.
Implementation:
This patchset removes the area map allocator in favor of a bitmap
allocator backed by metadata blocks. The primary goal is to provide
consistency in performance and memory footprint with a focus on small
allocations (< 64 bytes). The bitmap removes the heavy memmove from the
freeing critical path and provides a consistent memory footprint. The
metadata blocks provide a bound on the amount of scanning required by
maintaining a set of hints.
In an effort to make freeing fast, the metadata is updated on the free
path if the new free area makes a page free, a block free, or spans
across blocks. This causes the chunk's contig hint to potentially be
smaller than what it could allocate by up to the smaller of a page or a
block. If the chunk's contig hint is contained within a block, a check
occurs and the hint is kept accurate. Metadata is always kept accurate
on allocation, so there will not be a situation where a chunk has a
later contig hint than available.
Evaluation:
I have primarily done testing against a simple workload of allocation of
1 million objects (2^20) of varying size. Deallocation was done by in
order, alternating, and in reverse. These numbers were collected after
rebasing ontop of a80099a152. I present the worst-case numbers here:
Area Map Allocator:
Object Size | Alloc Time (ms) | Free Time (ms)
----------------------------------------------
4B | 310 | 4770
16B | 557 | 1325
64B | 436 | 273
256B | 776 | 131
1024B | 3280 | 122
Bitmap Allocator:
Object Size | Alloc Time (ms) | Free Time (ms)
----------------------------------------------
4B | 490 | 70
16B | 515 | 75
64B | 610 | 80
256B | 950 | 100
1024B | 3520 | 200
This data demonstrates the inability for the area map allocator to
handle less than ideal situations. In the best case of reverse
deallocation, the area map allocator was able to perform within range
of the bitmap allocator. In the worst case situation, freeing took
nearly 5 seconds for 1 million 4-byte objects. The bitmap allocator
dramatically improves the consistency of the free path. The small
allocations performed nearly identical regardless of the freeing
pattern.
While it does add to the allocation latency, the allocation scenario
here is optimal for the area map allocator. The area map allocator runs
into trouble when it is allocating in chunks where the latter half is
full. It is difficult to replicate this, so I present a variant where
the pages are second half filled. Freeing was done sequentially. Below
are the numbers for this scenario:
Area Map Allocator:
Object Size | Alloc Time (ms) | Free Time (ms)
----------------------------------------------
4B | 4118 | 4892
16B | 1651 | 1163
64B | 598 | 285
256B | 771 | 158
1024B | 3034 | 160
Bitmap Allocator:
Object Size | Alloc Time (ms) | Free Time (ms)
----------------------------------------------
4B | 481 | 67
16B | 506 | 69
64B | 636 | 75
256B | 892 | 90
1024B | 3262 | 147
The data shows a parabolic curve of performance for the area map
allocator. This is due to the memmove operation being the dominant cost
with the lower object sizes as more objects are packed in a chunk and at
higher object sizes, the traversal of the chunk slots is the dominating
cost. The bitmap allocator suffers this problem as well. The above data
shows the inability to scale for the allocation path with the area map
allocator and that the bitmap allocator demonstrates consistent
performance in general.
The second problem of additional scanning can result in the area map
allocator completing in 52 minutes when trying to allocate 1 million
4-byte objects with 8-byte alignment. The same workload takes
approximately 16 seconds to complete for the bitmap allocator.
V2:
Fixed a bug in pcpu_alloc_first_chunk end_offset was setting the bitmap
using bytes instead of bits.
Added a comment to pcpu_cnt_pop_pages to explain bitmap_weight.
Signed-off-by: Dennis Zhou <dennisszhou@gmail.com>
Reviewed-by: Josef Bacik <jbacik@fb.com>
Signed-off-by: Tejun Heo <tj@kernel.org>
pcpu_nr_empty_pop_pages is used to ensure there are a handful of free
pages around to serve atomic allocations. A new field, nr_empty_pop_pages,
is added to the pcpu_chunk struct to keep track of the number of empty
pages. This field is needed as the number of empty populated pages is
globally tracked and deltas are used to update in the bitmap allocator.
Pages that contain a hidden area are not considered to be empty. This
new field is exposed in percpu_stats.
Signed-off-by: Dennis Zhou <dennisszhou@gmail.com>
Reviewed-by: Josef Bacik <jbacik@fb.com>
Signed-off-by: Tejun Heo <tj@kernel.org>
The area map allocator manages the first chunk area by hiding all but
the region it is responsible for serving in the area map. To align this
with the populated page bitmap, end_offset is introduced to keep track
of the delta to end page aligned. The area map is appended with the
page aligned end when necessary to be in line with how the bitmap
allocator requires the ending to be aligned with the LCM of PAGE_SIZE
and the size of each bitmap block. percpu_stats is updated to ignore
this region when present.
Signed-off-by: Dennis Zhou <dennisszhou@gmail.com>
Reviewed-by: Josef Bacik <jbacik@fb.com>
Signed-off-by: Tejun Heo <tj@kernel.org>
Prior this variable was used to manage statistics when the first chunk
had a reserved region. The previous patch introduced start_offset to
keep track of the offset by value rather than boolean. Therefore,
has_reserved can be removed.
Signed-off-by: Dennis Zhou <dennisszhou@gmail.com>
Reviewed-by: Josef Bacik <jbacik@fb.com>
Signed-off-by: Tejun Heo <tj@kernel.org>
Percpu memory holds a minimum threshold of pages that are populated
in order to serve atomic percpu memory requests. This change makes it
easier to verify that there are a minimum number of populated pages
lying around.
Signed-off-by: Dennis Zhou <dennisszhou@gmail.com>
Signed-off-by: Tejun Heo <tj@kernel.org>
This makes the debugfs output for percpu_stats a little easier
to read by changing the spacing of the output to be consistent.
Signed-off-by: Dennis Zhou <dennisszhou@gmail.com>
Signed-off-by: Tejun Heo <tj@kernel.org>
Changes the use of a void buffer to an int buffer for clarity.
Signed-off-by: Dennis Zhou <dennisszhou@gmail.com>
Signed-off-by: Tejun Heo <tj@kernel.org>
There is limited visibility into the use of percpu memory leaving us
unable to reason about correctness of parameters and overall use of
percpu memory. These counters and statistics aim to help understand
basic statistics about percpu memory such as number of allocations over
the lifetime, allocation sizes, and fragmentation.
New Config: PERCPU_STATS
Signed-off-by: Dennis Zhou <dennisz@fb.com>
Signed-off-by: Tejun Heo <tj@kernel.org>
Roman Gushchin
Dennis Zhou
Thomas Gleixner
Kees Cook
Andy Shevchenko
Dennis Zhou
Dennis Zhou