[PATCH 00/29] use refcount+RCU method to implement lockless slab shrink

From: Qi Zheng
Date: Thu Jun 22 2023 - 04:55:38 EST

Next message: Qi Zheng: "[PATCH 02/29] mm: vmscan: introduce some helpers for dynamically allocating shrinker"
Previous message: kah . jing . lee: "[PATCH 2/2] arch: arm64: configs: Enable UBI and UBIFS"
In reply to: Steven Price: "Re: [PATCH 05/29] drm/panfrost: dynamically allocate the drm-panfrost shrinker"
Next in thread: Qi Zheng: "[PATCH 01/29] mm: shrinker: add shrinker::private_data field"
Messages sorted by: [ date ] [ thread ] [ subject ] [ author ]

Hi all,

1. Background
=============

We used to implement the lockless slab shrink with SRCU [1], but then kernel
test robot reported -88.8% regression in stress-ng.ramfs.ops_per_sec test
case [2], so we reverted it [3].

This patch series aims to re-implement the lockless slab shrink using the
refcount+RCU method proposed by Dave Chinner [4].

[1]. https://lore.kernel.org/lkml/20230313112819.38938-1-zhengqi.arch@xxxxxxxxxxxxx/
[2]. https://lore.kernel.org/lkml/202305230837.db2c233f-yujie.liu@xxxxxxxxx/
[3]. https://lore.kernel.org/all/20230609081518.3039120-1-qi.zheng@xxxxxxxxx/
[4]. https://lore.kernel.org/lkml/ZIJhou1d55d4H1s0@xxxxxxxxxxxxxxxxxxx/

2. Implementation
=================

Currently, the shrinker instances can be divided into the following three types:

a) global shrinker instance statically defined in the kernel, such as
workingset_shadow_shrinker.

b) global shrinker instance statically defined in the kernel modules, such as
mmu_shrinker in x86.

c) shrinker instance embedded in other structures.

For *case a*, the memory of shrinker instance is never freed. For *case b*, the
memory of shrinker instance will be freed after the module is unloaded. But we
will call synchronize_rcu() in free_module() to wait for RCU read-side critical
section to exit. For *case c*, we need to dynamically allocate these shrinker
instances, then the memory of shrinker instance can be dynamically freed alone
by calling kfree_rcu(). Then we can use rcu_read_{lock,unlock}() to ensure that
the shrinker instance is valid.

The shrinker::refcount mechanism ensures that the shrinker instance will not be
run again after unregistration. So the structure that records the pointer of
shrinker instance can be safely freed without waiting for the RCU read-side
critical section.

In this way, while we implement the lockless slab shrink, we don't need to be
blocked in unregister_shrinker() to wait RCU read-side critical section.

PATCH 1 ~ 2: infrastructure for dynamically allocating shrinker instances
PATCH 3 ~ 21: dynamically allocate the shrinker instances in case c
PATCH 22: introduce pool_shrink_rwsem to implement private synchronize_shrinkers()
PATCH 23 ~ 28: implement the lockless slab shrink
PATCH 29: move shrinker-related code into a separate file

3. Testing
==========

3.1 slab shrink stress test
---------------------------

We can reproduce the down_read_trylock() hotspot through the following script:

```

DIR="/root/shrinker/memcg/mnt"

do_create()
{
mkdir -p /sys/fs/cgroup/memory/test
mkdir -p /sys/fs/cgroup/perf_event/test
echo 4G > /sys/fs/cgroup/memory/test/memory.limit_in_bytes
for i in `seq 0 $1`;
do
mkdir -p /sys/fs/cgroup/memory/test/$i;
echo $$ > /sys/fs/cgroup/memory/test/$i/cgroup.procs;
echo $$ > /sys/fs/cgroup/perf_event/test/cgroup.procs;
mkdir -p $DIR/$i;
done
}

do_mount()
{
for i in `seq $1 $2`;
do
mount -t tmpfs $i $DIR/$i;
done
}

do_touch()
{
for i in `seq $1 $2`;
do
echo $$ > /sys/fs/cgroup/memory/test/$i/cgroup.procs;
echo $$ > /sys/fs/cgroup/perf_event/test/cgroup.procs;
dd if=/dev/zero of=$DIR/$i/file$i bs=1M count=1 &
done
}

case "$1" in
touch)
do_touch $2 $3
;;
test)
do_create 4000
do_mount 0 4000
do_touch 0 3000
;;
*)
exit 1
;;
esac
```

Save the above script, then run test and touch commands. Then we can use the
following perf command to view hotspots:

perf top -U -F 999 [-g]

1) Before applying this patchset:

35.34% [kernel] [k] down_read_trylock
18.44% [kernel] [k] shrink_slab
15.98% [kernel] [k] pv_native_safe_halt
15.08% [kernel] [k] up_read
5.33% [kernel] [k] idr_find
2.71% [kernel] [k] _find_next_bit
2.21% [kernel] [k] shrink_node
1.29% [kernel] [k] shrink_lruvec
0.66% [kernel] [k] do_shrink_slab
0.33% [kernel] [k] list_lru_count_one
0.33% [kernel] [k] __radix_tree_lookup
0.25% [kernel] [k] mem_cgroup_iter

- 82.19% 19.49% [kernel] [k] shrink_slab
- 62.00% shrink_slab
36.37% down_read_trylock
15.52% up_read
5.48% idr_find
3.38% _find_next_bit
+ 0.98% do_shrink_slab

2) After applying this patchset:

46.83% [kernel] [k] shrink_slab
20.52% [kernel] [k] pv_native_safe_halt
8.85% [kernel] [k] do_shrink_slab
7.71% [kernel] [k] _find_next_bit
1.72% [kernel] [k] xas_descend
1.70% [kernel] [k] shrink_node
1.44% [kernel] [k] shrink_lruvec
1.43% [kernel] [k] mem_cgroup_iter
1.28% [kernel] [k] xas_load
0.89% [kernel] [k] super_cache_count
0.84% [kernel] [k] xas_start
0.66% [kernel] [k] list_lru_count_one

- 65.50% 40.44% [kernel] [k] shrink_slab
- 22.96% shrink_slab
13.11% _find_next_bit
- 9.91% do_shrink_slab
- 1.59% super_cache_count
0.92% list_lru_count_one

We can see that the first perf hotspot becomes shrink_slab, which is what we
expect.

3.2 registeration and unregisteration stress test
-------------------------------------------------

Run the command below to test:

stress-ng --timeout 60 --times --verify --metrics-brief --ramfs 9 &

1) Before applying this patchset:

setting to a 60 second run per stressor
dispatching hogs: 9 ramfs
stressor bogo ops real time usr time sys time bogo ops/s bogo ops/s
(secs) (secs) (secs) (real time) (usr+sys time)
ramfs 880623 60.02 7.71 226.93 14671.45 3753.09
ramfs:
1 System Management Interrupt
for a 60.03s run time:
5762.40s available CPU time
7.71s user time ( 0.13%)
226.93s system time ( 3.94%)
234.64s total time ( 4.07%)
load average: 8.54 3.06 2.11
passed: 9: ramfs (9)
failed: 0
skipped: 0
successful run completed in 60.03s (1 min, 0.03 secs)

2) After applying this patchset:

setting to a 60 second run per stressor
dispatching hogs: 9 ramfs
stressor bogo ops real time usr time sys time bogo ops/s bogo ops/s
(secs) (secs) (secs) (real time) (usr+sys time)
ramfs 847562 60.02 7.44 230.22 14120.66 3566.23
ramfs:
4 System Management Interrupts
for a 60.12s run time:
5771.95s available CPU time
7.44s user time ( 0.13%)
230.22s system time ( 3.99%)
237.66s total time ( 4.12%)
load average: 8.18 2.43 0.84
passed: 9: ramfs (9)
failed: 0
skipped: 0
successful run completed in 60.12s (1 min, 0.12 secs)

We can see that the ops/s has hardly changed.

This series is based on next-20230613.

Comments and suggestions are welcome.

Thanks,
Qi.

Qi Zheng (29):
mm: shrinker: add shrinker::private_data field
mm: vmscan: introduce some helpers for dynamically allocating shrinker
drm/i915: dynamically allocate the i915_gem_mm shrinker
drm/msm: dynamically allocate the drm-msm_gem shrinker
drm/panfrost: dynamically allocate the drm-panfrost shrinker
dm: dynamically allocate the dm-bufio shrinker
dm zoned: dynamically allocate the dm-zoned-meta shrinker
md/raid5: dynamically allocate the md-raid5 shrinker
bcache: dynamically allocate the md-bcache shrinker
vmw_balloon: dynamically allocate the vmw-balloon shrinker
virtio_balloon: dynamically allocate the virtio-balloon shrinker
mbcache: dynamically allocate the mbcache shrinker
ext4: dynamically allocate the ext4-es shrinker
jbd2,ext4: dynamically allocate the jbd2-journal shrinker
NFSD: dynamically allocate the nfsd-client shrinker
NFSD: dynamically allocate the nfsd-reply shrinker
xfs: dynamically allocate the xfs-buf shrinker
xfs: dynamically allocate the xfs-inodegc shrinker
xfs: dynamically allocate the xfs-qm shrinker
zsmalloc: dynamically allocate the mm-zspool shrinker
fs: super: dynamically allocate the s_shrink
drm/ttm: introduce pool_shrink_rwsem
mm: shrinker: add refcount and completion_wait fields
mm: vmscan: make global slab shrink lockless
mm: vmscan: make memcg slab shrink lockless
mm: shrinker: make count and scan in shrinker debugfs lockless
mm: vmscan: hold write lock to reparent shrinker nr_deferred
mm: shrinkers: convert shrinker_rwsem to mutex
mm: shrinker: move shrinker-related code into a separate file

drivers/gpu/drm/i915/gem/i915_gem_shrinker.c | 27 +-
drivers/gpu/drm/i915/i915_drv.h | 3 +-
drivers/gpu/drm/msm/msm_drv.h | 2 +-
drivers/gpu/drm/msm/msm_gem_shrinker.c | 25 +-
drivers/gpu/drm/panfrost/panfrost_device.h | 2 +-
.../gpu/drm/panfrost/panfrost_gem_shrinker.c | 24 +-
drivers/gpu/drm/ttm/ttm_pool.c | 15 +
drivers/md/bcache/bcache.h | 2 +-
drivers/md/bcache/btree.c | 23 +-
drivers/md/bcache/sysfs.c | 2 +-
drivers/md/dm-bufio.c | 23 +-
drivers/md/dm-cache-metadata.c | 2 +-
drivers/md/dm-thin-metadata.c | 2 +-
drivers/md/dm-zoned-metadata.c | 25 +-
drivers/md/raid5.c | 28 +-
drivers/md/raid5.h | 2 +-
drivers/misc/vmw_balloon.c | 16 +-
drivers/virtio/virtio_balloon.c | 26 +-
fs/btrfs/super.c | 2 +-
fs/ext4/ext4.h | 2 +-
fs/ext4/extents_status.c | 21 +-
fs/jbd2/journal.c | 32 +-
fs/kernfs/mount.c | 2 +-
fs/mbcache.c | 39 +-
fs/nfsd/netns.h | 4 +-
fs/nfsd/nfs4state.c | 20 +-
fs/nfsd/nfscache.c | 33 +-
fs/proc/root.c | 2 +-
fs/super.c | 40 +-
fs/xfs/xfs_buf.c | 25 +-
fs/xfs/xfs_buf.h | 2 +-
fs/xfs/xfs_icache.c | 27 +-
fs/xfs/xfs_mount.c | 4 +-
fs/xfs/xfs_mount.h | 2 +-
fs/xfs/xfs_qm.c | 24 +-
fs/xfs/xfs_qm.h | 2 +-
include/linux/fs.h | 2 +-
include/linux/jbd2.h | 2 +-
include/linux/shrinker.h | 35 +-
mm/Makefile | 4 +-
mm/shrinker.c | 750 ++++++++++++++++++
mm/shrinker_debug.c | 26 +-
mm/vmscan.c | 702 ----------------
mm/zsmalloc.c | 28 +-
44 files changed, 1128 insertions(+), 953 deletions(-)
create mode 100644 mm/shrinker.c

--
2.30.2

Next message: Qi Zheng: "[PATCH 02/29] mm: vmscan: introduce some helpers for dynamically allocating shrinker"
Previous message: kah . jing . lee: "[PATCH 2/2] arch: arm64: configs: Enable UBI and UBIFS"
In reply to: Steven Price: "Re: [PATCH 05/29] drm/panfrost: dynamically allocate the drm-panfrost shrinker"
Next in thread: Qi Zheng: "[PATCH 01/29] mm: shrinker: add shrinker::private_data field"
Messages sorted by: [ date ] [ thread ] [ subject ] [ author ]