Slot cache is no longer needed now, removing it and all related code.
- vm-scalability with: `usemem --init-time -O -y -x -R -31 1G`,
12G memory cgroup using simulated pmem as SWAP (32G pmem, 32 CPUs),
16 test runs for each case, measuring the total throughput:
Before (KB/s) (stdev) After (KB/s) (stdev)
Random (4K): 424907.60 (24410.78) 414745.92 (34554.78)
Random (64K): 163308.82 (11635.72) 167314.50 (18434.99)
Sequential (4K, !-R): 6150056.79 (103205.90) 6321469.06 (115878.16)
The performance changes are below noise level.
- Build linux kernel with make -j96, using 4K folio with 1.5G memory
cgroup limit and 64K folio with 2G memory cgroup limit, on top of tmpfs,
12 test runs, measuring the system time:
Before (s) (stdev) After (s) (stdev)
make -j96 (4K): 6445.69 (61.95) 6408.80 (69.46)
make -j96 (64K): 6841.71 (409.04) 6437.99 (435.55)
Similar to above, 64k mTHP case showed a slight improvement.
Link: https://lkml.kernel.org/r/20250313165935.63303-7-ryncsn@gmail.com
Signed-off-by: Kairui Song <kasong@tencent.com>
Reviewed-by: Baoquan He <bhe@redhat.com>
Cc: Baolin Wang <baolin.wang@linux.alibaba.com>
Cc: Barry Song <v-songbaohua@oppo.com>
Cc: Chris Li <chrisl@kernel.org>
Cc: "Huang, Ying" <ying.huang@linux.alibaba.com>
Cc: Hugh Dickins <hughd@google.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Kalesh Singh <kaleshsingh@google.com>
Cc: Matthew Wilcow (Oracle) <willy@infradead.org>
Cc: Nhat Pham <nphamcs@gmail.com>
Cc: Yosry Ahmed <yosryahmed@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Current allocation workflow first traverses the plist with a global lock
held, after choosing a device, it uses the percpu cluster on that swap
device. This commit moves the percpu cluster variable out of being tied
to individual swap devices, making it a global percpu variable, and will
be used directly for allocation as a fast path.
The global percpu cluster variable will never point to a HDD device, and
allocations on a HDD device are still globally serialized.
This improves the allocator performance and prepares for removal of the
slot cache in later commits. There shouldn't be much observable behavior
change, except one thing: this changes how swap device allocation rotation
works.
Currently, each allocation will rotate the plist, and because of the
existence of slot cache (one order 0 allocation usually returns 64
entries), swap devices of the same priority are rotated for every 64 order
0 entries consumed. High order allocations are different, they will
bypass the slot cache, and so swap device is rotated for every 16K, 32K,
or up to 2M allocation.
The rotation rule was never clearly defined or documented, it was changed
several times without mentioning.
After this commit, and once slot cache is gone in later commits, swap
device rotation will happen for every consumed cluster. Ideally non-HDD
devices will be rotated if 2M space has been consumed for each order.
Fragmented clusters will rotate the device faster, which seems OK. HDD
devices is rotated for every allocation regardless of the allocation
order, which should be OK too and trivial.
This commit also slightly changes allocation behaviour for slot cache.
The new added cluster allocation fast path may allocate entries from
different device to the slot cache, this is not observable from user
space, only impact performance very slightly, and slot cache will be just
gone in next commit, so this can be ignored.
Link: https://lkml.kernel.org/r/20250313165935.63303-6-ryncsn@gmail.com
Signed-off-by: Kairui Song <kasong@tencent.com>
Cc: Baolin Wang <baolin.wang@linux.alibaba.com>
Cc: Baoquan He <bhe@redhat.com>
Cc: Barry Song <v-songbaohua@oppo.com>
Cc: Chris Li <chrisl@kernel.org>
Cc: "Huang, Ying" <ying.huang@linux.alibaba.com>
Cc: Hugh Dickins <hughd@google.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Kalesh Singh <kaleshsingh@google.com>
Cc: Matthew Wilcow (Oracle) <willy@infradead.org>
Cc: Nhat Pham <nphamcs@gmail.com>
Cc: Yosry Ahmed <yosryahmed@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Patch series "mm, swap: remove swap slot cache", v3.
Slot cache was initially introduced by commit 67afa38e01 ("mm/swap: add
cache for swap slots allocation") to reduce the lock contention of
si->lock.
Previous series "mm, swap: rework of swap allocator locks" [1] removed
swap slot cache for freeing path as freeing path no longer touches
si->lock in most cased. Allocation path also have slight to none
contention on si->lock since that series, but slot cache still helps to
reduce other overheads, like counters and the plist.
This series removes the slot cache from allocation path too, by using the
cluster as allocation fast path and also reduce other overheads.
Now slot cache is completely gone, the code is much simplified without
obvious feature or performance change, also clean up related workaround.
Also this should avoid other potential issues, e.g. the long pinning of
swap slots: swap slot cache pins swap slots with HAS_CACHE, causing
reclaim or allocation fail to use these slots on scanning.
The only behavior change is the swap device allocation rotation mechanism,
as explained in the patch "mm, swap: use percpu cluster as allocation fast
path".
Test results are looking good after deleting the swap slot cache:
- vm-scalability with: `usemem --init-time -O -y -x -R -31 1G`,
12G memory cgroup using simulated pmem as SWAP (32G pmem, 32 CPUs),
16 test runs for each case, measuring the total throughput:
Before (KB/s) (stdev) After (KB/s) (stdev)
Random (4K): 424907.60 (24410.78) 414745.92 (34554.78)
Random (64K): 163308.82 (11635.72) 167314.50 (18434.99)
Sequential (4K, !-R): 6150056.79 (103205.90) 6321469.06 (115878.16)
- Build linux kernel with make -j96, using 4K folio with 1.5G memory
cgroup limit and 64K folio with 2G memory cgroup limit, on top of tmpfs,
12 test runs, measuring the system time:
Before (s) (stdev) After (s) (stdev)
make -j96 (4K): 6445.69 (61.95) 6408.80 (69.46)
make -j96 (64K): 6841.71 (409.04) 6437.99 (435.55)
The performance is unchanged, slightly better in some cases.
[1] https://lore.kernel.org/linux-mm/20250113175732.48099-1-ryncsn@gmail.com/
This patch (of 7):
Swap allocator will do swap cache reclaim to recycle HAS_CACHE slots for
allocation. It initiates the reclaim from the offset to be reclaimed and
looks up the corresponding folio. The lookup process is lockless, so it's
possible the folio will be removed from the swap cache and given a
different swap entry before the reclaim locks the folio. If it happens,
the reclaim will end up reclaiming an irrelevant folio, and return wrong
return value.
This shouldn't cause any problem with correctness or stability, but it is
indeed confusing and unexpected, and will increase fragmentation, decrease
performance.
Fix this by checking whether the folio is still pointing to the offset the
allocator want to reclaim before reclaiming it.
Link: https://lkml.kernel.org/r/20250313165935.63303-1-ryncsn@gmail.com
Link: https://lkml.kernel.org/r/20250313165935.63303-2-ryncsn@gmail.com
Signed-off-by: Kairui Song <kasong@tencent.com>
Reviewed-by: Baoquan He <bhe@redhat.com>
Cc: Baolin Wang <baolin.wang@linux.alibaba.com>
Cc: Barry Song <v-songbaohua@oppo.com>
Cc: Chris Li <chrisl@kernel.org>
Cc: "Huang, Ying" <ying.huang@linux.alibaba.com>
Cc: Hugh Dickins <hughd@google.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Kairui Song <kasong@tencent.com>
Cc: Kalesh Singh <kaleshsingh@google.com>
Cc: Matthew Wilcow (Oracle) <willy@infradead.org>
Cc: Nhat Pham <nphamcs@gmail.com>
Cc: Yosry Ahmed <yosryahmed@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
When a process consumes a UE in a page, the memory failure handler
attempts to collect information for a potential SIGBUS. If the page is an
anonymous page, page_mapped_in_vma(page, vma) is invoked in order to
1. retrieve the vaddr from the process' address space,
2. verify that the vaddr is indeed mapped to the poisoned page,
where 'page' is the precise small page with UE.
It's been observed that when injecting poison to a non-head subpage of an
anonymous hugetlb page, no SIGBUS shows up, while injecting to the head
page produces a SIGBUS. The cause is that, though hugetlb_walk() returns
a valid pmd entry (on x86), but check_pte() detects mismatch between the
head page per the pmd and the input subpage. Thus the vaddr is considered
not mapped to the subpage and the process is not collected for SIGBUS
purpose. This is the calling stack:
collect_procs_anon
page_mapped_in_vma
page_vma_mapped_walk
hugetlb_walk
huge_pte_lock
check_pte
check_pte() header says that it
"check if [pvmw->pfn, @pvmw->pfn + @pvmw->nr_pages) is mapped at the @pvmw->pte"
but practically works only if pvmw->pfn is the head page pfn at pvmw->pte.
Hindsight acknowledging that some pvmw->pte could point to a hugepage of
some sort such that it makes sense to make check_pte() work for hugepage.
Link: https://lkml.kernel.org/r/20250224211445.2663312-1-jane.chu@oracle.com
Signed-off-by: Jane Chu <jane.chu@oracle.com>
Cc: Hugh Dickins <hughd@google.com>
Cc: Kirill A. Shuemov <kirill.shutemov@linux.intel.com>
Cc: linmiaohe <linmiaohe@huawei.com>
Cc: Matthew Wilcow (Oracle) <willy@infradead.org>
Cc: Peter Xu <peterx@redhat.com>
Cc: <stable@vger.kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
The fallback code searches for the biggest buddy first in an attempt to
steal the whole block and encourage type grouping down the line.
The approach used to be this:
- Non-movable requests will split the largest buddy and steal the
remainder. This splits up contiguity, but it allows subsequent
requests of this type to fall back into adjacent space.
- Movable requests go and look for the smallest buddy instead. The
thinking is that movable requests can be compacted, so grouping is
less important than retaining contiguity.
c0cd6f557b ("mm: page_alloc: fix freelist movement during block
conversion") enforces freelist type hygiene, which restricts stealing to
either claiming the whole block or just taking the requested chunk; no
additional pages or buddy remainders can be stolen any more.
The patch mishandled when to switch to finding the smallest buddy in that
new reality. As a result, it may steal the exact request size, but from
the biggest buddy. This causes fracturing for no good reason.
Fix this by committing to the new behavior: either steal the whole block,
or fall back to the smallest buddy.
Remove single-page stealing from steal_suitable_fallback(). Rename it to
try_to_steal_block() to make the intentions clear. If this fails, always
fall back to the smallest buddy.
The following is from 4 runs of mmtest's thpchallenge. "Pollute" is
single page fallback, "steal" is conversion of a partially used block.
The numbers for free block conversions (omitted) are comparable.
vanilla patched
@pollute[unmovable from reclaimable]: 27 106
@pollute[unmovable from movable]: 82 46
@pollute[reclaimable from unmovable]: 256 83
@pollute[reclaimable from movable]: 46 8
@pollute[movable from unmovable]: 4841 868
@pollute[movable from reclaimable]: 5278 12568
@steal[unmovable from reclaimable]: 11 12
@steal[unmovable from movable]: 113 49
@steal[reclaimable from unmovable]: 19 34
@steal[reclaimable from movable]: 47 21
@steal[movable from unmovable]: 250 183
@steal[movable from reclaimable]: 81 93
The allocator appears to do a better job at keeping stealing and polluting
to the first fallback preference. As a result, the numbers for "from
movable" - the least preferred fallback option, and most detrimental to
compactability - are down across the board.
Link: https://lkml.kernel.org/r/20250225001023.1494422-2-hannes@cmpxchg.org
Fixes: c0cd6f557b ("mm: page_alloc: fix freelist movement during block conversion")
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Suggested-by: Vlastimil Babka <vbabka@suse.cz>
Reviewed-by: Brendan Jackman <jackmanb@google.com>
Reviewed-by: Vlastimil Babka <vbabka@suse.cz>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Patch series "fs/proc/task_mmu: add guard region bit to pagemap".
Currently there is no means of determining whether a given page in a
mapping range is designated a guard region (as installed via madvise()
using the MADV_GUARD_INSTALL flag).
This is generally not an issue, but in some instances users may wish to
determine whether this is the case.
This series adds this ability via /proc/$pid/pagemap, updates the
documentation and adds a self test to assert that this functions
correctly.
This patch (of 2):
Currently there is no means by which users can determine whether a given
page in memory is in fact a guard region, that is having had the
MADV_GUARD_INSTALL madvise() flag applied to it.
This is intentional, as to provide this information in VMA metadata would
contradict the intent of the feature (providing a means to change fault
behaviour at a page table level rather than a VMA level), and would
require VMA metadata operations to scan page tables, which is
unacceptable.
In many cases, users have no need to reflect and determine what regions
have been designated guard regions, as it is the user who has established
them in the first place.
But in some instances, such as monitoring software, or software that
relies upon being able to ascertain the nature of mappings within a remote
process for instance, it becomes useful to be able to determine which
pages have the guard region marker applied.
This patch makes use of an unused pagemap bit (58) to provide this
information.
This patch updates the documentation at the same time as making the change
such that the implementation of the feature and the documentation of it
are tied together.
Link: https://lkml.kernel.org/r/cover.1740139449.git.lorenzo.stoakes@oracle.com
Link: https://lkml.kernel.org/r/521d99c08b975fb06a1e7201e971cc24d68196d1.1740139449.git.lorenzo.stoakes@oracle.com
Signed-off-by: Lorenzo Stoakes <lorenzo.stoakes@oracle.com>
Acked-by: David Hildenbrand <david@redhat.com>
Cc: Jann Horn <jannh@google.com>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Kalesh Singh <kaleshsingh@google.com>
Cc: Liam Howlett <liam.howlett@oracle.com>
Cc: Matthew Wilcow (Oracle) <willy@infradead.org>
Cc: "Paul E . McKenney" <paulmck@kernel.org>
Cc: Shuah Khan <shuah@kernel.org>
Cc: Suren Baghdasaryan <surenb@google.com>
Cc: Vlastimil Babka <vbabka@suse.cz>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
If running from a directory that can't be read by unprivileged users,
executing on-fault-test via the nobody user will fail.
The kselftest build does give the file the correct permissions, but after
being installed it might be in a directory without global execute
permissions.
Since the script can't safely fix that, just skip if it happens. Note
that the stderr of the `ls` command is unfiltered meaning the user sees a
"permission denied" error that can help inform them why the test was
skipped.
Link: https://lkml.kernel.org/r/20250311-mm-selftests-v4-11-dec210a658f5@google.com
Signed-off-by: Brendan Jackman <jackmanb@google.com>
Cc: Dev Jain <dev.jain@arm.com>
Cc: Lorenzo Stoakes <lorenzo.stoakes@oracle.com>
Cc: Mateusz Guzik <mjguzik@gmail.com>
Cc: Shuah Khan <shuah@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Patch series "selftests/mm: Some cleanups from trying to run them", v4.
I never had much luck running mm selftests so I spent a few hours digging
into why.
Looks like most of the reason is missing SKIP checks, so this series is
just adding a bunch of those that I found. I did not do anything like all
of them, just the ones I spotted in gup_longterm, gup_test, mmap,
userfaultfd and memfd_secret.
It's a bit unfortunate to have to skip those tests when ftruncate() fails,
but I don't have time to dig deep enough into it to actually make them
pass. I have observed the issue on 9pfs and heard rumours that NFS has a
similar problem.
I'm now able to run these test groups successfully:
- mmap
- gup_test
- compaction
- migration
- page_frag
- userfaultfd
- mlock
I've never gone past "Waiting for hugetlb memory to get depleted", in the
hugetlb tests. I don't know if they are stuck or if they would eventually
work if I was patient enough (testing on a 1G machine). I have not
investigated further.
I had some issues with mlock tests failing due to -ENOSRCH from mlock2(),
I can no longer reproduce that though, things work OK now.
Of the remaining tests there may be others that work fine, but there's no
convenient way to survey the whole output of run_vmtests.sh so I'm just
going test by test here.
In my spare moments I am slowly chipping away at a setup to run these
tests continuously in a reasonably hermetic QEMU environment via
virtme-ng:
5fad4b9c59/README.md
Hopefully that will eventually offer a way to provide a "canned"
environment where the tests are known to work, which can be fairly easily
reproduced by any developer.
This patch (of 12):
Just reporting failure doesn't tell you what went wrong. This can fail in
different ways so report errno to help the reader get started debugging.
Link: https://lkml.kernel.org/r/20250311-mm-selftests-v4-0-dec210a658f5@google.com
Link: https://lkml.kernel.org/r/20250311-mm-selftests-v4-1-dec210a658f5@google.com
Signed-off-by: Brendan Jackman <jackmanb@google.com>
Reviewed-by: Dev Jain <dev.jain@arm.com>
Cc: Lorenzo Stoakes <lorenzo.stoakes@oracle.com>
Cc: Mateusz Guzik <mjguzik@gmail.com>
Cc: Shuah Khan <shuah@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Current object mapping API is a little cumbersome. First, it's
inconsistent, sometimes it returns with page-faults disabled and sometimes
with page-faults enabled. Second, and most importantly, it enforces
atomicity restrictions on its users. zs_map_object() has to return a
liner object address which is not always possible because some objects
span multiple physical (non-contiguous) pages. For such objects zsmalloc
uses a per-CPU buffer to which object's data is copied before a pointer to
that per-CPU buffer is returned back to the caller. This leads to
another, final, issue - extra memcpy(). Since the caller gets a pointer
to per-CPU buffer it can memcpy() data only to that buffer, and during
zs_unmap_object() zsmalloc will memcpy() from that per-CPU buffer to
physical pages that object in question spans across.
New API splits functions by access mode:
- zs_obj_read_begin(handle, local_copy)
Returns a pointer to handle memory. For objects that span two
physical pages a local_copy buffer is used to store object's
data before the address is returned to the caller. Otherwise
the object's page is kmap_local mapped directly.
- zs_obj_read_end(handle, buf)
Unmaps the page if it was kmap_local mapped by zs_obj_read_begin().
- zs_obj_write(handle, buf, len)
Copies len-bytes from compression buffer to handle memory
(takes care of objects that span two pages). This does not
need any additional (e.g. per-CPU) buffers and writes the data
directly to zsmalloc pool pages.
In terms of performance, on a synthetic and completely reproducible
test that allocates fixed number of objects of fixed sizes and
iterates over those objects, first mapping in RO then in RW mode:
OLD API
=======
3 first results out of 10
369,205,778 instructions # 0.80 insn per cycle
40,467,926 branches # 113.732 M/sec
369,002,122 instructions # 0.62 insn per cycle
40,426,145 branches # 189.361 M/sec
369,036,706 instructions # 0.63 insn per cycle
40,430,860 branches # 204.105 M/sec
[..]
NEW API
=======
3 first results out of 10
265,799,293 instructions # 0.51 insn per cycle
29,834,567 branches # 170.281 M/sec
265,765,970 instructions # 0.55 insn per cycle
29,829,019 branches # 161.602 M/sec
265,764,702 instructions # 0.51 insn per cycle
29,828,015 branches # 189.677 M/sec
[..]
T-test on all 10 runs
=====================
Difference at 95.0% confidence
-1.03219e+08 +/- 55308.7
-27.9705% +/- 0.0149878%
(Student's t, pooled s = 58864.4)
The old API will stay around until the remaining users switch to the new
one. After that we'll also remove zsmalloc per-CPU buffer and CPU hotplug
handling.
The split of map(RO) and map(WO) into read_{begin/end}/write is suggested
by Yosry Ahmed.
Link: https://lkml.kernel.org/r/20250303022425.285971-15-senozhatsky@chromium.org
Signed-off-by: Sergey Senozhatsky <senozhatsky@chromium.org>
Suggested-by: Yosry Ahmed <yosry.ahmed@linux.dev>
Reviewed-by: Yosry Ahmed <yosry.ahmed@linux.dev>
Cc: Hillf Danton <hdanton@sina.com>
Cc: Kairui Song <ryncsn@gmail.com>
Cc: Minchan Kim <minchan@kernel.org>
Cc: Sebastian Andrzej Siewior <bigeasy@linutronix.de>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
In order to implement preemptible object mapping we need a zspage lock
that satisfies several preconditions:
- it should be reader-write type of a lock
- it should be possible to hold it from any context, but also being
preemptible if the context allows it
- we never sleep while acquiring but can sleep while holding in read
mode
An rwsemaphore doesn't suffice, due to atomicity requirements, rwlock
doesn't satisfy due to reader-preemptability requirement. It's also worth
to mention, that per-zspage rwsem is a little too memory heavy (we can
easily have double digits megabytes used only on rwsemaphores).
Switch over from rwlock_t to a atomic_t-based implementation of a
reader-writer semaphore that satisfies all of the preconditions.
The spin-lock based zspage_lock is suggested by Hillf Danton.
Link: https://lkml.kernel.org/r/20250303022425.285971-14-senozhatsky@chromium.org
Signed-off-by: Sergey Senozhatsky <senozhatsky@chromium.org>
Suggested-by: Hillf Danton <hdanton@sina.com>
Cc: Kairui Song <ryncsn@gmail.com>
Cc: Minchan Kim <minchan@kernel.org>
Cc: Sebastian Andrzej Siewior <bigeasy@linutronix.de>
Cc: Yosry Ahmed <yosry.ahmed@linux.dev>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Allocate post-processing target in place_pp_slot(). This simplifies
scan_slots_for_writeback() and scan_slots_for_recompress() loops because
we don't need to track pps pointer state anymore. Previously we have to
explicitly NULL the point if it has been added to a post-processing bucket
or re-use previously allocated pointer otherwise and make sure we don't
leak the memory in the end.
We are also fine doing GFP_NOIO allocation, as post-processing can be
called under memory pressure so we better pick as many slots as we can as
soon as we can and start post-processing them, possibly saving the memory.
Allocation failure there is not fatal, we will post-process whatever we
put into the buckets on previous iterations.
Link: https://lkml.kernel.org/r/20250303022425.285971-12-senozhatsky@chromium.org
Signed-off-by: Sergey Senozhatsky <senozhatsky@chromium.org>
Cc: Hillf Danton <hdanton@sina.com>
Cc: Kairui Song <ryncsn@gmail.com>
Cc: Minchan Kim <minchan@kernel.org>
Cc: Sebastian Andrzej Siewior <bigeasy@linutronix.de>
Cc: Yosry Ahmed <yosry.ahmed@linux.dev>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
This reworks recompression loop handling:
- set a rule that stream-put NULLs the stream pointer If the loop
returns with a non-NULL stream then it's a successful recompression,
otherwise the stream should always be NULL.
- do not count the number of recompressions Mark object as
incompressible as soon as the algorithm with the highest priority failed
to compress that object.
- count compression errors as resource usage Even if compression has
failed, we still need to bump num_recomp_pages counter.
Link: https://lkml.kernel.org/r/20250303022425.285971-11-senozhatsky@chromium.org
Signed-off-by: Sergey Senozhatsky <senozhatsky@chromium.org>
Cc: Hillf Danton <hdanton@sina.com>
Cc: Kairui Song <ryncsn@gmail.com>
Cc: Minchan Kim <minchan@kernel.org>
Cc: Sebastian Andrzej Siewior <bigeasy@linutronix.de>
Cc: Yosry Ahmed <yosry.ahmed@linux.dev>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Previously zram write() was atomic which required us to pass
__GFP_KSWAPD_RECLAIM to zsmalloc handle allocation on a fast path and
attempt a slow path allocation (with recompression) if the fast path
failed.
Since we are not in atomic context anymore we can permit direct reclaim
during handle allocation, and hence can have a single allocation path.
There is no slow path anymore so we don't unlock per-CPU stream (and don't
lose compressed data) which means that there is no need to do
recompression now (which should reduce CPU and battery usage).
Link: https://lkml.kernel.org/r/20250303022425.285971-6-senozhatsky@chromium.org
Signed-off-by: Sergey Senozhatsky <senozhatsky@chromium.org>
Cc: Hillf Danton <hdanton@sina.com>
Cc: Kairui Song <ryncsn@gmail.com>
Cc: Minchan Kim <minchan@kernel.org>
Cc: Sebastian Andrzej Siewior <bigeasy@linutronix.de>
Cc: Yosry Ahmed <yosry.ahmed@linux.dev>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Currently, per-CPU stream access is done from a non-preemptible (atomic)
section, which imposes the same atomicity requirements on compression
backends as entry spin-lock, and makes it impossible to use algorithms
that can schedule/wait/sleep during compression and decompression.
Switch to preemptible per-CPU model, similar to the one used in zswap.
Instead of a per-CPU local lock, each stream carries a mutex which is
locked throughout entire time zram uses it for compression or
decompression, so that cpu-dead event waits for zram to stop using a
particular per-CPU stream and release it.
Link: https://lkml.kernel.org/r/20250303022425.285971-3-senozhatsky@chromium.org
Signed-off-by: Sergey Senozhatsky <senozhatsky@chromium.org>
Suggested-by: Yosry Ahmed <yosry.ahmed@linux.dev>
Reviewed-by: Yosry Ahmed <yosry.ahmed@linux.dev>
Cc: Hillf Danton <hdanton@sina.com>
Cc: Kairui Song <ryncsn@gmail.com>
Cc: Minchan Kim <minchan@kernel.org>
Cc: Sebastian Andrzej Siewior <bigeasy@linutronix.de>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Patch series "zsmalloc/zram: there be preemption", v10.
Currently zram runs compression and decompression in non-preemptible
sections, e.g.
zcomp_stream_get() // grabs CPU local lock
zcomp_compress()
or
zram_slot_lock() // grabs entry spin-lock
zcomp_stream_get() // grabs CPU local lock
zs_map_object() // grabs rwlock and CPU local lock
zcomp_decompress()
Potentially a little troublesome for a number of reasons.
For instance, this makes it impossible to use async compression algorithms
or/and H/W compression algorithms, which can wait for OP completion or
resource availability. This also restricts what compression algorithms
can do internally, for example, zstd can allocate internal state memory
for C/D dictionaries:
do_fsync()
do_writepages()
zram_bio_write()
zram_write_page() // become non-preemptible
zcomp_compress()
zstd_compress()
ZSTD_compress_usingCDict()
ZSTD_compressBegin_usingCDict_internal()
ZSTD_resetCCtx_usingCDict()
ZSTD_resetCCtx_internal()
zstd_custom_alloc() // memory allocation
Not to mention that the system can be configured to maximize compression
ratio at a cost of CPU/HW time (e.g. lz4hc or deflate with very high
compression level) so zram can stay in non-preemptible section (even under
spin-lock or/and rwlock) for an extended period of time. Aside from
compression algorithms, this also restricts what zram can do. One
particular example is zram_write_page() zsmalloc handle allocation, which
has an optimistic allocation (disallowing direct reclaim) and a
pessimistic fallback path, which then forces zram to compress the page one
more time.
This series changes zram to not directly impose atomicity restrictions on
compression algorithms (and on itself), which makes zram write() fully
preemptible; zram read(), sadly, is not always preemptible yet. There are
still indirect atomicity restrictions imposed by zsmalloc(). One notable
example is object mapping API, which returns with: a) local CPU lock held
b) zspage rwlock held
First, zsmalloc's zspage lock is converted from rwlock to a special type
of RW-lookalike look with some extra guarantees/features. Second, a new
handle mapping is introduced which doesn't use per-CPU buffers (and hence
no local CPU lock), does fewer memcpy() calls, but requires users to
provide a pointer to temp buffer for object copy-in (when needed). Third,
zram is converted to the new zsmalloc mapping API and thus zram read()
becomes preemptible.
This patch (of 19):
Concurrent modifications of meta table entries is now handled by per-entry
spin-lock. This has a number of shortcomings.
First, this imposes atomic requirements on compression backends. zram can
call both zcomp_compress() and zcomp_decompress() under entry spin-lock,
which implies that we can use only compression algorithms that don't
schedule/sleep/wait during compression and decompression. This, for
instance, makes it impossible to use some of the ASYNC compression
algorithms (H/W compression, etc.) implementations.
Second, this can potentially trigger watchdogs. For example, entry
re-compression with secondary algorithms is performed under entry
spin-lock. Given that we chain secondary compression algorithms and that
some of them can be configured for best compression ratio (and worst
compression speed) zram can stay under spin-lock for quite some time.
Having a per-entry mutex (or, for instance, a rw-semaphore) significantly
increases sizeof() of each entry and hence the meta table. Therefore
entry locking returns back to bit locking, as before, however, this time
also preempt-rt friendly, because if waits-on-bit instead of
spinning-on-bit. Lock owners are also now permitted to schedule, which is
a first step on the path of making zram non-atomic.
Link: https://lkml.kernel.org/r/20250303022425.285971-1-senozhatsky@chromium.org
Link: https://lkml.kernel.org/r/20250303022425.285971-2-senozhatsky@chromium.org
Signed-off-by: Sergey Senozhatsky <senozhatsky@chromium.org>
Cc: Hillf Danton <hdanton@sina.com>
Cc: Kairui Song <ryncsn@gmail.com>
Cc: Minchan Kim <minchan@kernel.org>
Cc: Sebastian Andrzej Siewior <bigeasy@linutronix.de>
Cc: Yosry Ahmed <yosry.ahmed@linux.dev>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
