Patch series "zsmalloc: small compaction improvements", v2.
A tiny series that can reduce the number of find_alloced_obj() invocations
(which perform a linear scan of sub-page) during compaction. Inspired by
Alexey Romanov's findings.
This patch (of 3):
zspage migration can terminate as soon as it moves the last allocated
object from the source zspage. Add a simple helper zspage_empty() that
tests zspage ->inuse on each migration iteration.
Link: https://lkml.kernel.org/r/20230624053120.643409-2-senozhatsky@chromium.org
Signed-off-by: Sergey Senozhatsky <senozhatsky@chromium.org>
Suggested-by: Alexey Romanov <AVRomanov@sberdevices.ru>
Reviewed-by: Alexey Romanov <avromanov@sberdevices.ru>
Acked-by: Minchan Kim <minchan@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Under memory pressure, we sometimes observe the following crash:
[ 5694.832838] ------------[ cut here ]------------
[ 5694.842093] list_del corruption, ffff888014b6a448->next is LIST_POISON1 (dead000000000100)
[ 5694.858677] WARNING: CPU: 33 PID: 418824 at lib/list_debug.c:47 __list_del_entry_valid+0x42/0x80
[ 5694.961820] CPU: 33 PID: 418824 Comm: fuse_counters.s Kdump: loaded Tainted: G S 5.19.0-0_fbk3_rc3_hoangnhatpzsdynshrv41_10870_g85a9558a25de #1
[ 5694.990194] Hardware name: Wiwynn Twin Lakes MP/Twin Lakes Passive MP, BIOS YMM16 05/24/2021
[ 5695.007072] RIP: 0010:__list_del_entry_valid+0x42/0x80
[ 5695.017351] Code: 08 48 83 c2 22 48 39 d0 74 24 48 8b 10 48 39 f2 75 2c 48 8b 51 08 b0 01 48 39 f2 75 34 c3 48 c7 c7 55 d7 78 82 e8 4e 45 3b 00 <0f> 0b eb 31 48 c7 c7 27 a8 70 82 e8 3e 45 3b 00 0f 0b eb 21 48 c7
[ 5695.054919] RSP: 0018:ffffc90027aef4f0 EFLAGS: 00010246
[ 5695.065366] RAX: 41fe484987275300 RBX: ffff888008988180 RCX: 0000000000000000
[ 5695.079636] RDX: ffff88886006c280 RSI: ffff888860060480 RDI: ffff888860060480
[ 5695.093904] RBP: 0000000000000002 R08: 0000000000000000 R09: ffffc90027aef370
[ 5695.108175] R10: 0000000000000000 R11: ffffffff82fdf1c0 R12: 0000000010000002
[ 5695.122447] R13: ffff888014b6a448 R14: ffff888014b6a420 R15: 00000000138dc240
[ 5695.136717] FS: 00007f23a7d3f740(0000) GS:ffff888860040000(0000) knlGS:0000000000000000
[ 5695.152899] CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033
[ 5695.164388] CR2: 0000560ceaab6ac0 CR3: 000000001c06c001 CR4: 00000000007706e0
[ 5695.178659] DR0: 0000000000000000 DR1: 0000000000000000 DR2: 0000000000000000
[ 5695.192927] DR3: 0000000000000000 DR6: 00000000fffe0ff0 DR7: 0000000000000400
[ 5695.207197] PKRU: 55555554
[ 5695.212602] Call Trace:
[ 5695.217486] <TASK>
[ 5695.221674] zs_map_object+0x91/0x270
[ 5695.229000] zswap_frontswap_store+0x33d/0x870
[ 5695.237885] ? do_raw_spin_lock+0x5d/0xa0
[ 5695.245899] __frontswap_store+0x51/0xb0
[ 5695.253742] swap_writepage+0x3c/0x60
[ 5695.261063] shrink_page_list+0x738/0x1230
[ 5695.269255] shrink_lruvec+0x5ec/0xcd0
[ 5695.276749] ? shrink_slab+0x187/0x5f0
[ 5695.284240] ? mem_cgroup_iter+0x6e/0x120
[ 5695.292255] shrink_node+0x293/0x7b0
[ 5695.299402] do_try_to_free_pages+0xea/0x550
[ 5695.307940] try_to_free_pages+0x19a/0x490
[ 5695.316126] __folio_alloc+0x19ff/0x3e40
[ 5695.323971] ? __filemap_get_folio+0x8a/0x4e0
[ 5695.332681] ? walk_component+0x2a8/0xb50
[ 5695.340697] ? generic_permission+0xda/0x2a0
[ 5695.349231] ? __filemap_get_folio+0x8a/0x4e0
[ 5695.357940] ? walk_component+0x2a8/0xb50
[ 5695.365955] vma_alloc_folio+0x10e/0x570
[ 5695.373796] ? walk_component+0x52/0xb50
[ 5695.381634] wp_page_copy+0x38c/0xc10
[ 5695.388953] ? filename_lookup+0x378/0xbc0
[ 5695.397140] handle_mm_fault+0x87f/0x1800
[ 5695.405157] do_user_addr_fault+0x1bd/0x570
[ 5695.413520] exc_page_fault+0x5d/0x110
[ 5695.421017] asm_exc_page_fault+0x22/0x30
After some investigation, I have found the following issue: unlike other
zswap backends, zsmalloc performs the LRU list update at the object
mapping time, rather than when the slot for the object is allocated.
This deviation was discussed and agreed upon during the review process
of the zsmalloc writeback patch series:
https://lore.kernel.org/lkml/Y3flcAXNxxrvy3ZH@cmpxchg.org/
Unfortunately, this introduces a subtle bug that occurs when there is a
concurrent store and reclaim, which interleave as follows:
zswap_frontswap_store() shrink_worker()
zs_malloc() zs_zpool_shrink()
spin_lock(&pool->lock) zs_reclaim_page()
zspage = find_get_zspage()
spin_unlock(&pool->lock)
spin_lock(&pool->lock)
zspage = list_first_entry(&pool->lru)
list_del(&zspage->lru)
zspage->lru.next = LIST_POISON1
zspage->lru.prev = LIST_POISON2
spin_unlock(&pool->lock)
zs_map_object()
spin_lock(&pool->lock)
if (!list_empty(&zspage->lru))
list_del(&zspage->lru)
CHECK_DATA_CORRUPTION(next == LIST_POISON1) /* BOOM */
With the current upstream code, this issue rarely happens. zswap only
triggers writeback when the pool is already full, at which point all
further store attempts are short-circuited. This creates an implicit
pseudo-serialization between reclaim and store. I am working on a new
zswap shrinking mechanism, which makes interleaving reclaim and store
more likely, exposing this bug.
zbud and z3fold do not have this problem, because they perform the LRU
list update in the alloc function, while still holding the pool's lock.
This patch fixes the aforementioned bug by moving the LRU update back to
zs_malloc(), analogous to zbud and z3fold.
Link: https://lkml.kernel.org/r/20230505185054.2417128-1-nphamcs@gmail.com
Fixes: 64f768c6b3 ("zsmalloc: add a LRU to zs_pool to keep track of zspages in LRU order")
Signed-off-by: Nhat Pham <nphamcs@gmail.com>
Suggested-by: Johannes Weiner <hannes@cmpxchg.org>
Acked-by: Johannes Weiner <hannes@cmpxchg.org>
Reviewed-by: Sergey Senozhatsky <senozhatsky@chromium.org>
Acked-by: Minchan Kim <minchan@kernel.org>
Cc: Dan Streetman <ddstreet@ieee.org>
Cc: Nitin Gupta <ngupta@vflare.org>
Cc: Seth Jennings <sjenning@redhat.com>
Cc: Vitaly Wool <vitaly.wool@konsulko.com>
Cc: <stable@vger.kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
zsmalloc pool can be compacted concurrently by many contexts,
e.g.
cc1 handle_mm_fault()
do_anonymous_page()
__alloc_pages_slowpath()
try_to_free_pages()
do_try_to_free_pages(
lru_gen_shrink_node()
shrink_slab()
do_shrink_slab()
zs_shrinker_scan()
zs_compact()
Pool compaction is currently (basically) single-threaded as
it is performed under pool->lock. Having multiple compaction
threads results in unnecessary contention, as each thread
competes for pool->lock. This, in turn, affects all zsmalloc
operations such as zs_malloc(), zs_map_object(), zs_free(), etc.
Introduce the pool->compaction_in_progress atomic variable,
which ensures that only one compaction context can run at a
time. This reduces overall pool->lock contention in (corner)
cases when many contexts attempt to shrink zspool simultaneously.
Link: https://lkml.kernel.org/r/20230418074639.1903197-1-senozhatsky@chromium.org
Fixes: c0547d0b6a ("zsmalloc: consolidate zs_pool's migrate_lock and size_class's locks")
Signed-off-by: Sergey Senozhatsky <senozhatsky@chromium.org>
Reviewed-by: Yosry Ahmed <yosryahmed@google.com>
Cc: Minchan Kim <minchan@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
The current implementation of the compaction loop fails to set the source
zspage pointer to NULL in all cases, leading to a potential issue where
__zs_compact() could use a stale zspage pointer. This pointer could even
point to a previously freed zspage, causing unexpected behavior in the
putback_zspage() and migrate_write_unlock() functions after returning from
the compaction loop.
Address the issue by ensuring that the source zspage pointer is always set
to NULL when it should be.
Link: https://lkml.kernel.org/r/20230417130850.1784777-1-senozhatsky@chromium.org
Fixes: 5a845e9f2d ("zsmalloc: rework compaction algorithm")
Signed-off-by: Sergey Senozhatsky <senozhatsky@chromium.org>
Reported-by: Yu Zhao <yuzhao@google.com>
Tested-by: Yu Zhao <yuzhao@google.com>
Reviewed-by: Yosry Ahmed <yosryahmed@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
The zsmalloc compaction algorithm has the potential to waste some CPU
cycles, particularly when compacting pages within the same fullness group.
This is due to the way it selects the head page of the fullness list for
source and destination pages, and how it reinserts those pages during each
iteration. The algorithm may first use a page as a migration destination
and then as a migration source, leading to an unnecessary back-and-forth
movement of objects.
Consider the following fullness list:
PageA PageB PageC PageD PageE
During the first iteration, the compaction algorithm will select PageA as
the source and PageB as the destination. All of PageA's objects will be
moved to PageB, and then PageA will be released while PageB is reinserted
into the fullness list.
PageB PageC PageD PageE
During the next iteration, the compaction algorithm will again select the
head of the list as the source and destination, meaning that PageB will
now serve as the source and PageC as the destination. This will result in
the objects being moved away from PageB, the same objects that were just
moved to PageB in the previous iteration.
To prevent this avalanche effect, the compaction algorithm should not
reinsert the destination page between iterations. By doing so, the most
optimal page will continue to be used and its usage ratio will increase,
reducing internal fragmentation. The destination page should only be
reinserted into the fullness list if:
- It becomes full
- No source page is available.
TEST
====
It's very challenging to reliably test this series. I ended up developing
my own synthetic test that has 100% reproducibility. The test generates
significan fragmentation (for each size class) and then performs
compaction for each class individually and tracks the number of memcpy()
in zs_object_copy(), so that we can compare the amount work compaction
does on per-class basis.
Total amount of work (zram mm_stat objs_moved)
----------------------------------------------
Old fullness grouping, old compaction algorithm:
323977 memcpy() in zs_object_copy().
Old fullness grouping, new compaction algorithm:
262944 memcpy() in zs_object_copy().
New fullness grouping, new compaction algorithm:
213978 memcpy() in zs_object_copy().
Per-class compaction memcpy() comparison (T-test)
-------------------------------------------------
x Old fullness grouping, old compaction algorithm
+ Old fullness grouping, new compaction algorithm
N Min Max Median Avg Stddev
x 140 349 3513 2461 2314.1214 806.03271
+ 140 289 2778 2006 1878.1714 641.02073
Difference at 95.0% confidence
-435.95 +/- 170.595
-18.8387% +/- 7.37193%
(Student's t, pooled s = 728.216)
x Old fullness grouping, old compaction algorithm
+ New fullness grouping, new compaction algorithm
N Min Max Median Avg Stddev
x 140 349 3513 2461 2314.1214 806.03271
+ 140 226 2279 1644 1528.4143 524.85268
Difference at 95.0% confidence
-785.707 +/- 159.331
-33.9527% +/- 6.88516%
(Student's t, pooled s = 680.132)
Link: https://lkml.kernel.org/r/20230304034835.2082479-4-senozhatsky@chromium.org
Signed-off-by: Sergey Senozhatsky <senozhatsky@chromium.org>
Acked-by: Minchan Kim <minchan@kernel.org>
Cc: Yosry Ahmed <yosryahmed@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Each zspage maintains ->inuse counter which keeps track of the number of
objects stored in the zspage. The ->inuse counter also determines the
zspage's "fullness group" which is calculated as the ratio of the "inuse"
objects to the total number of objects the zspage can hold
(objs_per_zspage). The closer the ->inuse counter is to objs_per_zspage,
the better.
Each size class maintains several fullness lists, that keep track of
zspages of particular "fullness". Pages within each fullness list are
stored in random order with regard to the ->inuse counter. This is
because sorting the zspages by ->inuse counter each time obj_malloc() or
obj_free() is called would be too expensive. However, the ->inuse counter
is still a crucial factor in many situations.
For the two major zsmalloc operations, zs_malloc() and zs_compact(), we
typically select the head zspage from the corresponding fullness list as
the best candidate zspage. However, this assumption is not always
accurate.
For the zs_malloc() operation, the optimal candidate zspage should have
the highest ->inuse counter. This is because the goal is to maximize the
number of ZS_FULL zspages and make full use of all allocated memory.
For the zs_compact() operation, the optimal source zspage should have the
lowest ->inuse counter. This is because compaction needs to move objects
in use to another page before it can release the zspage and return its
physical pages to the buddy allocator. The fewer objects in use, the
quicker compaction can release the zspage. Additionally, compaction is
measured by the number of pages it releases.
This patch reworks the fullness grouping mechanism. Instead of having two
groups - ZS_ALMOST_EMPTY (usage ratio below 3/4) and ZS_ALMOST_FULL (usage
ration above 3/4) - that result in too many zspages being included in the
ALMOST_EMPTY group for specific classes, size classes maintain a larger
number of fullness lists that give strict guarantees on the minimum and
maximum ->inuse values within each group. Each group represents a 10%
change in the ->inuse ratio compared to neighboring groups. In essence,
there are groups for zspages with 0%, 10%, 20% usage ratios, and so on, up
to 100%.
This enhances the selection of candidate zspages for both zs_malloc() and
zs_compact(). A printout of the ->inuse counters of the first 7 zspages
per (random) class fullness group:
class-768 objs_per_zspage 16:
fullness 100%: empty
fullness 99%: empty
fullness 90%: empty
fullness 80%: empty
fullness 70%: empty
fullness 60%: 8 8 9 9 8 8 8
fullness 50%: empty
fullness 40%: 5 5 6 5 5 5 5
fullness 30%: 4 4 4 4 4 4 4
fullness 20%: 2 3 2 3 3 2 2
fullness 10%: 1 1 1 1 1 1 1
fullness 0%: empty
The zs_malloc() function searches through the groups of pages starting
with the one having the highest usage ratio. This means that it always
selects a zspage from the group with the least internal fragmentation
(highest usage ratio) and makes it even less fragmented by increasing its
usage ratio.
The zs_compact() function, on the other hand, begins by scanning the group
with the highest fragmentation (lowest usage ratio) to locate the source
page. The first available zspage is selected, and then the function moves
downward to find a destination zspage in the group with the lowest
internal fragmentation (highest usage ratio).
Link: https://lkml.kernel.org/r/20230304034835.2082479-3-senozhatsky@chromium.org
Signed-off-by: Sergey Senozhatsky <senozhatsky@chromium.org>
Acked-by: Minchan Kim <minchan@kernel.org>
Cc: Yosry Ahmed <yosryahmed@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Patch series "zsmalloc: fine-grained fullness and new compaction
algorithm", v4.
Existing zsmalloc page fullness grouping leads to suboptimal page
selection for both zs_malloc() and zs_compact(). This patchset reworks
zsmalloc fullness grouping/classification.
Additinally it also implements new compaction algorithm that is expected
to use less CPU-cycles (as it potentially does fewer memcpy-s in
zs_object_copy()).
Test (synthetic) results can be seen in patch 0003.
This patch (of 4):
This optimization has no effect. It only ensures that when a zspage was
added to its corresponding fullness list, its "inuse" counter was higher
or lower than the "inuse" counter of the zspage at the head of the list.
The intention was to keep busy zspages at the head, so they could be
filled up and moved to the ZS_FULL fullness group more quickly. However,
this doesn't work as the "inuse" counter of a zspage can be modified by
obj_free() but the zspage may still belong to the same fullness list. So,
fix_fullness_group() won't change the zspage's position in relation to the
head's "inuse" counter, leading to a largely random order of zspages
within the fullness list.
For instance, consider a printout of the "inuse" counters of the first 10
zspages in a class that holds 93 objects per zspage:
ZS_ALMOST_EMPTY: 36 67 68 64 35 54 63 52
As we can see the zspage with the lowest "inuse" counter
is actually the head of the fullness list.
Remove this pointless "optimisation".
Link: https://lkml.kernel.org/r/20230304034835.2082479-1-senozhatsky@chromium.org
Link: https://lkml.kernel.org/r/20230304034835.2082479-2-senozhatsky@chromium.org
Signed-off-by: Sergey Senozhatsky <senozhatsky@chromium.org>
Acked-by: Minchan Kim <minchan@kernel.org>
Cc: Yosry Ahmed <yosryahmed@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Patch series "zsmalloc: make zspage chain size configurable".
Computers are bad at division. We currently decide the best zspage chain
size (max number of physical pages per-zspage) by looking at a `used
percentage` value. This is not enough as we lose precision during usage
percentage calculations For example, let's look at size class 208:
pages per zspage wasted bytes used%
1 144 96
2 80 99
3 16 99
4 160 99
Current algorithm will select 2 page per zspage configuration, as it's the
first one to reach 99%. However, 3 pages per zspage waste less memory.
Change algorithm and select zspage configuration that has lowest wasted
value.
Link: https://lkml.kernel.org/r/20230118005210.2814763-1-senozhatsky@chromium.org
Link: https://lkml.kernel.org/r/20230118005210.2814763-2-senozhatsky@chromium.org
Signed-off-by: Sergey Senozhatsky <senozhatsky@chromium.org>
Acked-by: Minchan Kim <minchan@kernel.org>
Cc: Mike Kravetz <mike.kravetz@oracle.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Currently, there is a race between zs_free() and zs_reclaim_page():
zs_reclaim_page() finds a handle to an allocated object, but before the
eviction happens, an independent zs_free() call to the same handle could
come in and overwrite the object value stored at the handle with the last
deferred handle. When zs_reclaim_page() finally gets to call the eviction
handler, it will see an invalid object value (i.e the previous deferred
handle instead of the original object value).
This race happens quite infrequently. We only managed to produce it with
out-of-tree developmental code that triggers zsmalloc writeback with a
much higher frequency than usual.
This patch fixes this race by storing the deferred handle in the object
header instead. We differentiate the deferred handle from the other two
cases (handle for allocated object, and linkage for free object) with a
new tag. If zspage reclamation succeeds, we will free these deferred
handles by walking through the zspage objects. On the other hand, if
zspage reclamation fails, we reconstruct the zspage freelist (with the
deferred handle tag and allocated tag) before trying again with the
reclamation.
[arnd@arndb.de: avoid unused-function warning]
Link: https://lkml.kernel.org/r/20230117170507.2651972-1-arnd@kernel.org
Link: https://lkml.kernel.org/r/20230110231701.326724-1-nphamcs@gmail.com
Fixes: 9997bc0175 ("zsmalloc: implement writeback mechanism for zsmalloc")
Signed-off-by: Nhat Pham <nphamcs@gmail.com>
Signed-off-by: Arnd Bergmann <arnd@arndb.de>
Suggested-by: Johannes Weiner <hannes@cmpxchg.org>
Cc: Dan Streetman <ddstreet@ieee.org>
Cc: Minchan Kim <minchan@kernel.org>
Cc: Nitin Gupta <ngupta@vflare.org>
Cc: Sergey Senozhatsky <senozhatsky@chromium.org>
Cc: Seth Jennings <sjenning@redhat.com>
Cc: Vitaly Wool <vitaly.wool@konsulko.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
