Pull erofs updates from Gao Xiang:
"Still no new features for this cycle, as some ongoing improvements
remain premature for now.
This includes a micro-optimization for the superblock checksum, along
with minor bugfixes and code cleanups, as usual:
- Micro-optimize superblock checksum
- Avoid overly large bvecs[] for file-backed mounts
- Some leftover folio conversion in z_erofs_bind_cache()
- Minor bugfixes and cleanups"
* tag 'erofs-for-6.14-rc1' of git://git.kernel.org/pub/scm/linux/kernel/git/xiang/erofs:
erofs: refine z_erofs_get_extent_compressedlen()
erofs: remove dead code in erofs_fc_parse_param
erofs: return SHRINK_EMPTY if no objects to free
erofs: convert z_erofs_bind_cache() to folios
erofs: tidy up zdata.c
erofs: get rid of `z_erofs_next_pcluster_t`
erofs: simplify z_erofs_load_compact_lcluster()
erofs: fix potential return value overflow of z_erofs_shrink_scan()
erofs: shorten bvecs[] for file-backed mounts
erofs: micro-optimize superblock checksum
fs: erofs: xattr.c change kzalloc to kcalloc
Pull kthread updates from Frederic Weisbecker:
"Kthreads affinity follow either of 4 existing different patterns:
1) Per-CPU kthreads must stay affine to a single CPU and never
execute relevant code on any other CPU. This is currently handled
by smpboot code which takes care of CPU-hotplug operations.
Affinity here is a correctness constraint.
2) Some kthreads _have_ to be affine to a specific set of CPUs and
can't run anywhere else. The affinity is set through
kthread_bind_mask() and the subsystem takes care by itself to
handle CPU-hotplug operations. Affinity here is assumed to be a
correctness constraint.
3) Per-node kthreads _prefer_ to be affine to a specific NUMA node.
This is not a correctness constraint but merely a preference in
terms of memory locality. kswapd and kcompactd both fall into this
category. The affinity is set manually like for any other task and
CPU-hotplug is supposed to be handled by the relevant subsystem so
that the task is properly reaffined whenever a given CPU from the
node comes up. Also care should be taken so that the node affinity
doesn't cross isolated (nohz_full) cpumask boundaries.
4) Similar to the previous point except kthreads have a _preferred_
affinity different than a node. Both RCU boost kthreads and RCU
exp kworkers fall into this category as they refer to "RCU nodes"
from a distinctly distributed tree.
Currently the preferred affinity patterns (3 and 4) have at least 4
identified users, with more or less success when it comes to handle
CPU-hotplug operations and CPU isolation. Each of which do it in its
own ad-hoc way.
This is an infrastructure proposal to handle this with the following
API changes:
- kthread_create_on_node() automatically affines the created kthread
to its target node unless it has been set as per-cpu or bound with
kthread_bind[_mask]() before the first wake-up.
- kthread_affine_preferred() is a new function that can be called
right after kthread_create_on_node() to specify a preferred
affinity different than the specified node.
When the preferred affinity can't be applied because the possible
targets are offline or isolated (nohz_full), the kthread is affine to
the housekeeping CPUs (which means to all online CPUs most of the time
or only the non-nohz_full CPUs when nohz_full= is set).
kswapd, kcompactd, RCU boost kthreads and RCU exp kworkers have been
converted, along with a few old drivers.
Summary of the changes:
- Consolidate a bunch of ad-hoc implementations of
kthread_run_on_cpu()
- Introduce task_cpu_fallback_mask() that defines the default last
resort affinity of a task to become nohz_full aware
- Add some correctness check to ensure kthread_bind() is always
called before the first kthread wake up.
- Default affine kthread to its preferred node.
- Convert kswapd / kcompactd and remove their halfway working ad-hoc
affinity implementation
- Implement kthreads preferred affinity
- Unify kthread worker and kthread API's style
- Convert RCU kthreads to the new API and remove the ad-hoc affinity
implementation"
* tag 'kthread-for-6.14-rc1' of git://git.kernel.org/pub/scm/linux/kernel/git/frederic/linux-dynticks:
kthread: modify kernel-doc function name to match code
rcu: Use kthread preferred affinity for RCU exp kworkers
treewide: Introduce kthread_run_worker[_on_cpu]()
kthread: Unify kthread_create_on_cpu() and kthread_create_worker_on_cpu() automatic format
rcu: Use kthread preferred affinity for RCU boost
kthread: Implement preferred affinity
mm: Create/affine kswapd to its preferred node
mm: Create/affine kcompactd to its preferred node
kthread: Default affine kthread to its preferred NUMA node
kthread: Make sure kthread hasn't started while binding it
sched,arm64: Handle CPU isolation on last resort fallback rq selection
arm64: Exclude nohz_full CPUs from 32bits el0 support
lib: test_objpool: Use kthread_run_on_cpu()
kallsyms: Use kthread_run_on_cpu()
soc/qman: test: Use kthread_run_on_cpu()
arm/bL_switcher: Use kthread_run_on_cpu()
Pull misc vfs updates from Christian Brauner:
"Features:
- Support caching symlink lengths in inodes
The size is stored in a new union utilizing the same space as
i_devices, thus avoiding growing the struct or taking up any more
space
When utilized it dodges strlen() in vfs_readlink(), giving about
1.5% speed up when issuing readlink on /initrd.img on ext4
- Add RWF_DONTCACHE iocb and FOP_DONTCACHE file_operations flag
If a file system supports uncached buffered IO, it may set
FOP_DONTCACHE and enable support for RWF_DONTCACHE.
If RWF_DONTCACHE is attempted without the file system supporting
it, it'll get errored with -EOPNOTSUPP
- Enable VBOXGUEST and VBOXSF_FS on ARM64
Now that VirtualBox is able to run as a host on arm64 (e.g. the
Apple M3 processors) we can enable VBOXSF_FS (and in turn
VBOXGUEST) for this architecture.
Tested with various runs of bonnie++ and dbench on an Apple MacBook
Pro with the latest Virtualbox 7.1.4 r165100 installed
Cleanups:
- Delay sysctl_nr_open check in expand_files()
- Use kernel-doc includes in fiemap docbook
- Use page->private instead of page->index in watch_queue
- Use a consume fence in mnt_idmap() as it's heavily used in
link_path_walk()
- Replace magic number 7 with ARRAY_SIZE() in fc_log
- Sort out a stale comment about races between fd alloc and dup2()
- Fix return type of do_mount() from long to int
- Various cosmetic cleanups for the lockref code
Fixes:
- Annotate spinning as unlikely() in __read_seqcount_begin
The annotation already used to be there, but got lost in commit
52ac39e5db ("seqlock: seqcount_t: Implement all read APIs as
statement expressions")
- Fix proc_handler for sysctl_nr_open
- Flush delayed work in delayed fput()
- Fix grammar and spelling in propagate_umount()
- Fix ESP not readable during coredump
In /proc/PID/stat, there is the kstkesp field which is the stack
pointer of a thread. While the thread is active, this field reads
zero. But during a coredump, it should have a valid value
However, at the moment, kstkesp is zero even during coredump
- Don't wake up the writer if the pipe is still full
- Fix unbalanced user_access_end() in select code"
* tag 'vfs-6.14-rc1.misc' of git://git.kernel.org/pub/scm/linux/kernel/git/vfs/vfs: (28 commits)
gfs2: use lockref_init for qd_lockref
erofs: use lockref_init for pcl->lockref
dcache: use lockref_init for d_lockref
lockref: add a lockref_init helper
lockref: drop superfluous externs
lockref: use bool for false/true returns
lockref: improve the lockref_get_not_zero description
lockref: remove lockref_put_not_zero
fs: Fix return type of do_mount() from long to int
select: Fix unbalanced user_access_end()
vbox: Enable VBOXGUEST and VBOXSF_FS on ARM64
pipe_read: don't wake up the writer if the pipe is still full
selftests: coredump: Add stackdump test
fs/proc: do_task_stat: Fix ESP not readable during coredump
fs: add RWF_DONTCACHE iocb and FOP_DONTCACHE file_operations flag
fs: sort out a stale comment about races between fd alloc and dup2
fs: Fix grammar and spelling in propagate_umount()
fs: fc_log replace magic number 7 with ARRAY_SIZE()
fs: use a consume fence in mnt_idmap()
file: flush delayed work in delayed fput()
...
kthread_create() creates a kthread without running it yet. kthread_run()
creates a kthread and runs it.
On the other hand, kthread_create_worker() creates a kthread worker and
runs it.
This difference in behaviours is confusing. Also there is no way to
create a kthread worker and affine it using kthread_bind_mask() or
kthread_affine_preferred() before starting it.
Consolidate the behaviours and introduce kthread_run_worker[_on_cpu]()
that behaves just like kthread_run(). kthread_create_worker[_on_cpu]()
will now only create a kthread worker without starting it.
Signed-off-by: Frederic Weisbecker <frederic@kernel.org>
Signed-off-by: Dan Carpenter <dan.carpenter@linaro.org>
kthread_create_on_cpu() uses the CPU argument as an implicit and unique
printf argument to add to the format whereas
kthread_create_worker_on_cpu() still relies on explicitly passing the
printf arguments. This difference in behaviour is error prone and
doesn't help standardizing per-CPU kthread names.
Unify the behaviours and convert kthread_create_worker_on_cpu() to
use the printf behaviour of kthread_create_on_cpu().
Signed-off-by: Frederic Weisbecker <frederic@kernel.org>
Once a pcluster is fully decompressed and there are no attached cached
folios, its corresponding `struct z_erofs_pcluster` will be freed. This
will significantly reduce the frequency of calls to erofs_shrink_scan()
and the memory allocated for `struct z_erofs_pcluster`.
The tables below show approximately a 96% reduction in the calls to
erofs_shrink_scan() and in the memory allocated for `struct
z_erofs_pcluster` after applying this patch. The results were obtained
by performing a test to copy a 4.1GB partition on ARM64 Android devices
running the 6.6 kernel with an 8-core CPU and 12GB of memory.
1. The reduction in calls to erofs_shrink_scan():
+-----------------+-----------+----------+---------+
| | w/o patch | w/ patch | diff |
+-----------------+-----------+----------+---------+
| Average (times) | 11390 | 390 | -96.57% |
+-----------------+-----------+----------+---------+
2. The reduction in memory released by erofs_shrink_scan():
+-----------------+-----------+----------+---------+
| | w/o patch | w/ patch | diff |
+-----------------+-----------+----------+---------+
| Average (Byte) | 133612656 | 4434552 | -96.68% |
+-----------------+-----------+----------+---------+
Signed-off-by: Chunhai Guo <guochunhai@vivo.com>
Reviewed-by: Gao Xiang <hsiangkao@linux.alibaba.com>
Link: https://lore.kernel.org/r/20241112043235.546164-1-guochunhai@vivo.com
Signed-off-by: Gao Xiang <hsiangkao@linux.alibaba.com>
`struct erofs_workgroup` was introduced to provide a unique header
for all physically indexed objects. However, after big pclusters and
shared pclusters are implemented upstream, it seems that all EROFS
encoded data (which requires transformation) can be represented with
`struct z_erofs_pcluster` directly.
Move all members into `struct z_erofs_pcluster` for simplicity.
Reviewed-by: Chao Yu <chao@kernel.org>
Signed-off-by: Gao Xiang <hsiangkao@linux.alibaba.com>
Link: https://lore.kernel.org/r/20241021035323.3280682-3-hsiangkao@linux.alibaba.com
This patch aims to allocate bvpages and short-lived compressed pages
from the reserved pool first.
After applying this patch, there are three benefits.
1. It reduces the page allocation time.
The bvpages and short-lived compressed pages account for about 4% of
the pages allocated from the system in the multi-app launch benchmarks
[1]. It reduces the page allocation time accordingly and lowers the
likelihood of blockage by page allocation in low memory scenarios.
2. The pages in the reserved pool will be allocated on demand.
Currently, bvpages and short-lived compressed pages are short-lived
pages allocated from the system, and the pages in the reserved pool all
originate from short-lived pages. Consequently, the number of reserved
pool pages will increase to z_erofs_rsv_nrpages over time.
With this patch, all short-lived pages are allocated from the reserved
pool first, so the number of reserved pool pages will only increase when
there are not enough pages. Thus, even if z_erofs_rsv_nrpages is set to
a large number for specific reasons, the actual number of reserved pool
pages may remain low as per demand. In the multi-app launch benchmarks
[1], z_erofs_rsv_nrpages is set at 256, while the number of reserved
pool pages remains below 64.
3. When erofs cache decompression is disabled
(EROFS_ZIP_CACHE_DISABLED), all pages will *only* be allocated from
the reserved pool for erofs. This will significantly reduce the memory
pressure from erofs.
[1] For additional details on the multi-app launch benchmarks, please
refer to commit 0f6273ab46 ("erofs: add a reserved buffer pool for lz4
decompression").
Signed-off-by: Chunhai Guo <guochunhai@vivo.com>
Reviewed-by: Gao Xiang <hsiangkao@linux.alibaba.com>
Reviewed-by: Chao Yu <chao@kernel.org>
Link: https://lore.kernel.org/r/20240906121110.3701889-1-guochunhai@vivo.com
Signed-off-by: Gao Xiang <hsiangkao@linux.alibaba.com>
With iterative development, our codebase can now deal with compressed
buffer misses properly if both in-place I/O and compressed buffer
allocation fail.
Note that if readahead fails (with non-uptodate folios), the original
request will then fall back to synchronous read, and `.read_folio()`
should return appropriate errnos; otherwise -EIO will be passed to
user space, which is unexpected.
To simplify rarely encountered failure paths, a mimic decompression
will be just used. Before that, failure reasons are recorded in
compressed_bvecs[] and they also act as placeholders to avoid in-place
pages. They will be parsed just before decompression and then pass
back to `.read_folio()`.
Reviewed-by: Chao Yu <chao@kernel.org>
Signed-off-by: Gao Xiang <hsiangkao@linux.alibaba.com>
Link: https://lore.kernel.org/r/20240905084732.2684515-1-hsiangkao@linux.alibaba.com
syzbot reported a task hang issue due to a deadlock case where it is
waiting for the folio lock of a cached folio that will be used for
cache I/Os.
After looking into the crafted fuzzed image, I found it's formed with
several overlapped big pclusters as below:
Ext: logical offset | length : physical offset | length
0: 0.. 16384 | 16384 : 151552.. 167936 | 16384
1: 16384.. 32768 | 16384 : 155648.. 172032 | 16384
2: 32768.. 49152 | 16384 : 537223168.. 537239552 | 16384
...
Here, extent 0/1 are physically overlapped although it's entirely
_impossible_ for normal filesystem images generated by mkfs.
First, managed folios containing compressed data will be marked as
up-to-date and then unlocked immediately (unlike in-place folios) when
compressed I/Os are complete. If physical blocks are not submitted in
the incremental order, there should be separate BIOs to avoid dependency
issues. However, the current code mis-arranges z_erofs_fill_bio_vec()
and BIO submission which causes unexpected BIO waits.
Second, managed folios will be connected to their own pclusters for
efficient inter-queries. However, this is somewhat hard to implement
easily if overlapped big pclusters exist. Again, these only appear in
fuzzed images so let's simply fall back to temporary short-lived pages
for correctness.
Additionally, it justifies that referenced managed folios cannot be
truncated for now and reverts part of commit 2080ca1ed3 ("erofs: tidy
up `struct z_erofs_bvec`") for simplicity although it shouldn't be any
difference.
Reported-by: syzbot+4fc98ed414ae63d1ada2@syzkaller.appspotmail.com
Reported-by: syzbot+de04e06b28cfecf2281c@syzkaller.appspotmail.com
Reported-by: syzbot+c8c8238b394be4a1087d@syzkaller.appspotmail.com
Tested-by: syzbot+4fc98ed414ae63d1ada2@syzkaller.appspotmail.com
Closes: https://lore.kernel.org/r/0000000000002fda01061e334873@google.com
Fixes: 8e6c8fa9f2 ("erofs: enable big pcluster feature")
Signed-off-by: Gao Xiang <hsiangkao@linux.alibaba.com>
Link: https://lore.kernel.org/r/20240910070847.3356592-1-hsiangkao@linux.alibaba.com
Smatch complains that:
fs/erofs/zdata.c:1047 z_erofs_scan_folio()
error: uninitialized symbol 'err'.
The issue is if we hit this (!(map->m_flags & EROFS_MAP_MAPPED)) {
condition then "err" isn't set. It's inside a loop so we would have to
hit that condition on every iteration. Initialize "err" to zero to
solve this.
Fixes: 5b9654efb6 ("erofs: teach z_erofs_scan_folios() to handle multi-page folios")
Signed-off-by: Dan Carpenter <dan.carpenter@linaro.org>
Link: https://lore.kernel.org/r/f78ab50e-ed6d-4275-8dd4-a4159fa565a2@stanley.mountain
Signed-off-by: Gao Xiang <hsiangkao@linux.alibaba.com>
LZ4 always reuses the decompressed buffer as its LZ77 sliding window
(dynamic dictionary) for optimal performance. However, in specific
cases, the output buffer may not fully contain valid page cache pages,
resulting in the use of short-lived pages for temporary purposes.
Due to the limited sliding window size, LZ4 shortlived bounce pages can
also be reused in a sliding manner, so each bounce page can be vmapped
multiple times in different relative positions by design. In order to
avoiding double frees, currently, reuse counts are recorded via page
refcount, but it will no longer be used as-is in the future world of
Memdescs.
Just maintain a lookup table to check if a shortlived page is reused.
Signed-off-by: Gao Xiang <hsiangkao@linux.alibaba.com>
Link: https://lore.kernel.org/r/20240711053659.1364989-1-hsiangkao@linux.alibaba.com
After revisiting the design, I believe `struct z_erofs_bvec` should
be page-based instead of folio-based due to the reasons below:
- The minimized memory mapping block is a page;
- Under the certain circumstances, only temporary pages needs to be
used instead of folios since refcount, mapcount for such pages are
unnecessary;
- Decompressors handle all types of pages including temporary pages,
not only folios.
When handling `struct z_erofs_bvec`, all folio-related information
is now accessed using the page_folio() helper.
The final goal of this round adaptation is to eliminate direct
accesses to `struct page` in the EROFS codebase, except for some
exceptions like `z_erofs_is_shortlived_page()` and
`z_erofs_page_is_invalidated()`, which require a new helper to
determine the memdesc type of an arbitrary page.
Actually large folios of compressed files seem to work now, yet I tend
to conduct more tests before officially enabling this for all scenarios.
Signed-off-by: Gao Xiang <hsiangkao@linux.alibaba.com>
Link: https://lore.kernel.org/r/20240703120051.3653452-4-hsiangkao@linux.alibaba.com
Previously, a folio just contains one page. In order to enable large
folios, z_erofs_scan_folios() needs to handle multi-page folios.
First, this patch eliminates all gotos. Instead, the new loop deal
with multiple parts in each folio. It's simple to handle the parts
which belong to unmapped extents or fragment extents; but for encoded
extents, the page boundaries needs to be considered for `tight` and
`split` to keep inplace I/Os work correctly: when a part crosses the
page boundary, they needs to be reseted properly.
Besides, simplify `tight` derivation since Z_EROFS_PCLUSTER_HOOKED
has been removed for quite a while.
Signed-off-by: Gao Xiang <hsiangkao@linux.alibaba.com>
Link: https://lore.kernel.org/r/20240703120051.3653452-3-hsiangkao@linux.alibaba.com
Callers are happier that way, especially since we no longer need to
play with splitting offset into block number and offset within block,
passing the former to erofs_bread(), then adding the latter...
erofs_bread() always reads entire pages, anyway.
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
Online folios are locked file-backed folios which will eventually
keep decoded (e.g. decompressed) data of each inode for end users to
utilize. It may belong to a few pclusters and contain other data (e.g.
compressed data for inplace I/Os) temporarily in a time-sharing manner
to reduce memory footprints for low-ended storage devices with high
latencies under heary I/O pressure.
Apart from folio_end_read() usage, it's a straight-forward conversion.
Reviewed-by: Chao Yu <chao@kernel.org>
Signed-off-by: Gao Xiang <hsiangkao@linux.alibaba.com>
Link: https://lore.kernel.org/r/20240305091448.1384242-1-hsiangkao@linux.alibaba.com
Even with inplace decompression, sometimes very few temporary buffers
may be still needed for a single decompression shot (e.g. 16 pages for
64k sliding window or 4 pages for 16k sliding window). In low-memory
scenarios, it would be better to try to allocate with GFP_NOWAIT on
readahead first. That can help reduce the time spent on page allocation
under durative memory pressure.
Here are detailed performance numbers under multi-app launch benchmark
workload [1] on ARM64 Android devices (8-core CPU and 8GB of memory)
running a 5.15 LTS kernel with EROFS of 4k pclusters:
+----------------------------------------------+
| LZ4 | vanilla | patched | diff |
|----------------+---------+---------+---------|
| Average (ms) | 3364 | 2684 | -20.21% | [64k sliding window]
|----------------+---------+---------+---------|
| Average (ms) | 2079 | 1610 | -22.56% | [16k sliding window]
+----------------------------------------------+
The total size of system images for 4k pclusters is almost unchanged:
(64k sliding window) 9,117,044 KB
(16k sliding window) 9,113,096 KB
Therefore, in addition to switch the sliding window from 64k to 16k,
after applying this patch, it can eventually save 52.14% (3364 -> 1610)
on average with no memory reservation. That is particularly useful for
embedded devices with limited resources.
[1] https://lore.kernel.org/r/20240109074143.4138783-1-guochunhai@vivo.com
Suggested-by: Gao Xiang <xiang@kernel.org>
Signed-off-by: Chunhai Guo <guochunhai@vivo.com>
Signed-off-by: Gao Xiang <hsiangkao@linux.alibaba.com>
Reviewed-by: Yue Hu <huyue2@coolpad.com>
Link: https://lore.kernel.org/r/20240126140142.201718-1-hsiangkao@linux.alibaba.com
I encountered a race issue after lengthy (~594647 secs) stress tests on
a 64k-page arm64 VM with several 4k-block EROFS images. The timing
is like below:
z_erofs_try_inplace_io z_erofs_fill_bio_vec
cmpxchg(&compressed_bvecs[].page,
NULL, ..)
[access bufvec]
compressed_bvecs[] = *bvec;
Previously, z_erofs_submit_queue() just accessed bufvec->page only, so
other fields in bufvec didn't matter. After the subpage block support
is landed, .offset and .end can be used too, but filling bufvec isn't
an atomic operation which can cause inconsistency.
Let's use a spinlock to keep the atomicity of each bufvec. More
specifically, just reuse the existing spinlock `pcl->obj.lockref.lock`
since it's rarely used (also it takes a short time if even used) as long
as the pcluster has a reference.
Fixes: 192351616a ("erofs: support I/O submission for sub-page compressed blocks")
Signed-off-by: Gao Xiang <hsiangkao@linux.alibaba.com>
Reviewed-by: Yue Hu <huyue2@coolpad.com>
Reviewed-by: Sandeep Dhavale <dhavale@google.com>
Link: https://lore.kernel.org/r/20240125120039.3228103-1-hsiangkao@linux.alibaba.com
Add a basic I/O submission path first to support sub-page blocks:
- Temporary short-lived pages will be used entirely;
- In-place I/O pages can be used partially, but compressed pages need
to be able to be mapped in contiguous virtual memory.
As a start, currently cache decompression is explicitly disabled for
sub-page blocks, which will be supported in the future.
Reviewed-by: Yue Hu <huyue2@coolpad.com>
Reviewed-by: Chao Yu <chao@kernel.org>
Signed-off-by: Gao Xiang <hsiangkao@linux.alibaba.com>
Link: https://lore.kernel.org/r/20231206091057.87027-2-hsiangkao@linux.alibaba.com
