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
Clean up some leftovers since there is no way for EROFS to be called
again from a reclaim context.
Signed-off-by: Jingbo Xu <jefflexu@linux.alibaba.com>
Reviewed-by: Gao Xiang <hsiangkao@linux.alibaba.com>
Link: https://lore.kernel.org/r/20240124031945.130782-1-jefflexu@linux.alibaba.com
Signed-off-by: Gao Xiang <hsiangkao@linux.alibaba.com>
In order to reduce memory footprints even further, let's allow
partially filled compressed bvecs for readahead to bail out later.
Signed-off-by: Yue Hu <huyue2@coolpad.com>
Reviewed-by: Gao Xiang <hsiangkao@linux.alibaba.com>
Link: https://lore.kernel.org/r/20231221062341.23901-1-zbestahu@gmail.com
Signed-off-by: Gao Xiang <hsiangkao@linux.alibaba.com>
Let's just disable cached decompression and inplace I/Os for partial
pages as the first step in order to enable sub-page block initial
support. In other words, currently it works primarily based on
temporary short-lived pages. Don't expect too much in terms of
performance.
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/20231206091057.87027-6-hsiangkao@linux.alibaba.com
Currently, compressed sizes are recorded in pages using `pclusterpages`,
However, for tailpacking pclusters, `tailpacking_size` is used instead.
This approach doesn't work when dealing with sub-page blocks. To address
this, let's switch them to the unified `pclustersize` in bytes.
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/20231206091057.87027-3-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
Both MicroLZMA and DEFLATE algorithms can use short-lived pages on
demand for the overlapped inplace I/O decompression.
However, those short-lived pages are actually added to
`be->compressed_pages`. Thus, it should be checked instead of
`pcl->compressed_bvecs`.
The LZ4 algorithm doesn't work like this, so it won't be impacted.
Fixes: 67139e36d9 ("erofs: introduce `z_erofs_parse_in_bvecs'")
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/20231128180431.4116991-1-hsiangkao@linux.alibaba.com
As Linus pointed out [1], lockref_put_return() is fundamentally
designed to be something that can fail. It behaves as a fastpath-only
thing, and the failure case needs to be handled anyway.
Actually, since the new pcluster was just allocated without being
populated, it won't be accessed by others until it is inserted into
XArray, so lockref helpers are actually unneeded here.
Let's just set the proper reference count on initializing.
[1] https://lore.kernel.org/r/CAHk-=whCga8BeQnJ3ZBh_Hfm9ctba_wpF444LpwRybVNMzO6Dw@mail.gmail.com
Fixes: 7674a42f35 ("erofs: use struct lockref to replace handcrafted approach")
Reviewed-by: Chao Yu <chao@kernel.org>
Link: https://lore.kernel.org/r/20231031060524.1103921-1-hsiangkao@linux.alibaba.com
Signed-off-by: Gao Xiang <hsiangkao@linux.alibaba.com>
Currently ztailpacking pclusters are chained with FOLLOWED_NOINPLACE and
not recorded into the managed_pslots XArray.
After commit 7674a42f35 ("erofs: use struct lockref to replace
handcrafted approach"), ztailpacking pclusters won't be freed with
erofs_workgroup_put() anymore, which will cause the following issue:
BUG erofs_pcluster-1 (Tainted: G OE ): Objects remaining in erofs_pcluster-1 on __kmem_cache_shutdown()
Use z_erofs_free_pcluster() directly to free ztailpacking pclusters.
Fixes: 7674a42f35 ("erofs: use struct lockref to replace handcrafted approach")
Signed-off-by: Jingbo Xu <jefflexu@linux.alibaba.com>
Reviewed-by: Gao Xiang <hsiangkao@linux.alibaba.com>
Reviewed-by: Chao Yu <chao@kernel.org>
Link: https://lore.kernel.org/r/20230822110530.96831-1-jefflexu@linux.alibaba.com
Signed-off-by: Gao Xiang <hsiangkao@linux.alibaba.com>
It's a straight-forward conversion except that readahead_folio()
will do folio_put() in advance but it doesn't matter since folios
are still locked.
As before, since file-backed folios (pages for now) are locked, so
we could temporarily use folio->private as an internal counter to
indicate split parts of each folio for the corresponding pclusters
to decompress.
When such counter becomes zero, the folio will be finally unlocked
(see compress.h and z_erofs_onlinepage_endio()).
Reviewed-by: Chao Yu <chao@kernel.org>
Signed-off-by: Gao Xiang <hsiangkao@linux.alibaba.com>
Link: https://lore.kernel.org/r/20230817082813.81180-7-hsiangkao@linux.alibaba.com
EROFS_MAP_FULL_MAPPED is more accurate to decide if caching the last
incomplete pcluster for later read or not.
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/20230817082813.81180-6-hsiangkao@linux.alibaba.com
Some preparation logic should be part of z_erofs_pcluster_begin()
instead of z_erofs_do_read_page(). Let's move now.
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/20230817082813.81180-3-hsiangkao@linux.alibaba.com
{collector,collection} were once reserved in order to indicate different
runtime logical extent instance of multi-reference pclusters.
However, de-duplicated decompression has been landed in a more flexable
way, thus `struct z_erofs_collection` was formally removed in commit
87ca34a706 ("erofs: get rid of `struct z_erofs_collection'").
Let's handle the remaining leftovers, for example:
`z_erofs_collector_begin` => `z_erofs_pcluster_begin`
`z_erofs_collector_end` => `z_erofs_pcluster_end`
as well as some comments. No logic changes.
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/20230817082813.81180-2-hsiangkao@linux.alibaba.com
Don't warn users since -EINTR can be returned due to user interruption.
Also suppress warning messages of readmore.
Signed-off-by: Ferry Meng <mengferry@linux.alibaba.com>
Reviewed-by: Gao Xiang <hsiangkao@linux.alibaba.com>
Reviewed-by: Yue Hu <huyue2@coolpad.com>
Reviewed-by: Chao Yu <chao@kernel.org>
Link: https://lore.kernel.org/r/20230809060637.21311-1-mengferry@linux.alibaba.com
Signed-off-by: Gao Xiang <hsiangkao@linux.alibaba.com>
When handling deduplicated compressed data, there can be multiple
decompressed extents pointing to the same compressed data in one shot.
In such cases, the bvecs which belong to the longest extent will be
selected as the primary bvecs for real decompressors to decode and the
other duplicated bvecs will be directly copied from the primary bvecs.
Previously, only relative offsets of the longest extent were checked to
decompress the primary bvecs. On rare occasions, it can be incorrect
if there are several extents with the same start relative offset.
As a result, some short bvecs could be selected for decompression and
then cause data corruption.
For example, as Shijie Sun reported off-list, considering the following
extents of a file:
117: 903345.. 915250 | 11905 : 385024.. 389120 | 4096
...
119: 919729.. 930323 | 10594 : 385024.. 389120 | 4096
...
124: 968881.. 980786 | 11905 : 385024.. 389120 | 4096
The start relative offset is the same: 2225, but extent 119 (919729..
930323) is shorter than the others.
Let's restrict the bvec length in addition to the start offset if bvecs
are not full.
Reported-by: Shijie Sun <sunshijie@xiaomi.com>
Fixes: 5c2a64252c ("erofs: introduce partial-referenced pclusters")
Tested-by Shijie Sun <sunshijie@xiaomi.com>
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/20230719065459.60083-1-hsiangkao@linux.alibaba.com
z_erofs_do_read_page() may loop infinitely due to the inappropriate
truncation in the below statement. Since the offset is 64 bits and min_t()
truncates the result to 32 bits. The solution is to replace unsigned int
with a 64-bit type, such as erofs_off_t.
cur = end - min_t(unsigned int, offset + end - map->m_la, end);
- For example:
- offset = 0x400160000
- end = 0x370
- map->m_la = 0x160370
- offset + end - map->m_la = 0x400000000
- offset + end - map->m_la = 0x00000000 (truncated as unsigned int)
- Expected result:
- cur = 0
- Actual result:
- cur = 0x370
Signed-off-by: Chunhai Guo <guochunhai@vivo.com>
Fixes: 3883a79abd ("staging: erofs: introduce VLE decompression support")
Reviewed-by: Gao Xiang <hsiangkao@linux.alibaba.com>
Reviewed-by: Chao Yu <chao@kernel.org>
Link: https://lore.kernel.org/r/20230710093410.44071-1-guochunhai@vivo.com
Signed-off-by: Gao Xiang <hsiangkao@linux.alibaba.com>
z_erofs_pcluster_readmore() may take a long time to loop when the page
offset is large enough, which is unnecessary should be prevented.
For example, when the following case is encountered, it will loop 4691368
times, taking about 27 seconds:
- offset = 19217289215
- inode_size = 1442672
Signed-off-by: Chunhai Guo <guochunhai@vivo.com>
Fixes: 386292919c ("erofs: introduce readmore decompression strategy")
Reviewed-by: Gao Xiang <hsiangkao@linux.alibaba.com>
Reviewed-by: Yue Hu <huyue2@coolpad.com>
Reviewed-by: Chao Yu <chao@kernel.org>
Link: https://lore.kernel.org/r/20230710042531.28761-1-guochunhai@vivo.com
Signed-off-by: Gao Xiang <hsiangkao@linux.alibaba.com>
Current check for atomic context is not sufficient as
z_erofs_decompressqueue_endio can be called under rcu lock
from blk_mq_flush_plug_list(). See the stacktrace [1]
In such case we should hand off the decompression work for async
processing rather than trying to do sync decompression in current
context. Patch fixes the detection by checking for
rcu_read_lock_any_held() and while at it use more appropriate
!in_task() check than in_atomic().
Background: Historically erofs would always schedule a kworker for
decompression which would incur the scheduling cost regardless of
the context. But z_erofs_decompressqueue_endio() may not always
be in atomic context and we could actually benefit from doing the
decompression in z_erofs_decompressqueue_endio() if we are in
thread context, for example when running with dm-verity.
This optimization was later added in patch [2] which has shown
improvement in performance benchmarks.
==============================================
[1] Problem stacktrace
[name:core&]BUG: sleeping function called from invalid context at kernel/locking/mutex.c:291
[name:core&]in_atomic(): 0, irqs_disabled(): 0, non_block: 0, pid: 1615, name: CpuMonitorServi
[name:core&]preempt_count: 0, expected: 0
[name:core&]RCU nest depth: 1, expected: 0
CPU: 7 PID: 1615 Comm: CpuMonitorServi Tainted: G S W OE 6.1.25-android14-5-maybe-dirty-mainline #1
Hardware name: MT6897 (DT)
Call trace:
dump_backtrace+0x108/0x15c
show_stack+0x20/0x30
dump_stack_lvl+0x6c/0x8c
dump_stack+0x20/0x48
__might_resched+0x1fc/0x308
__might_sleep+0x50/0x88
mutex_lock+0x2c/0x110
z_erofs_decompress_queue+0x11c/0xc10
z_erofs_decompress_kickoff+0x110/0x1a4
z_erofs_decompressqueue_endio+0x154/0x180
bio_endio+0x1b0/0x1d8
__dm_io_complete+0x22c/0x280
clone_endio+0xe4/0x280
bio_endio+0x1b0/0x1d8
blk_update_request+0x138/0x3a4
blk_mq_plug_issue_direct+0xd4/0x19c
blk_mq_flush_plug_list+0x2b0/0x354
__blk_flush_plug+0x110/0x160
blk_finish_plug+0x30/0x4c
read_pages+0x2fc/0x370
page_cache_ra_unbounded+0xa4/0x23c
page_cache_ra_order+0x290/0x320
do_sync_mmap_readahead+0x108/0x2c0
filemap_fault+0x19c/0x52c
__do_fault+0xc4/0x114
handle_mm_fault+0x5b4/0x1168
do_page_fault+0x338/0x4b4
do_translation_fault+0x40/0x60
do_mem_abort+0x60/0xc8
el0_da+0x4c/0xe0
el0t_64_sync_handler+0xd4/0xfc
el0t_64_sync+0x1a0/0x1a4
[2] Link: https://lore.kernel.org/all/20210317035448.13921-1-huangjianan@oppo.com/
Reported-by: Will Shiu <Will.Shiu@mediatek.com>
Suggested-by: Gao Xiang <xiang@kernel.org>
Signed-off-by: Sandeep Dhavale <dhavale@google.com>
Reviewed-by: Gao Xiang <hsiangkao@linux.alibaba.com>
Reviewed-by: Alexandre Mergnat <amergnat@baylibre.com>
Link: https://lore.kernel.org/r/20230621220848.3379029-1-dhavale@google.com
Signed-off-by: Gao Xiang <hsiangkao@linux.alibaba.com>
Let's avoid the current handcrafted lockref although `struct lockref`
inclusion usually increases extra 4 bytes with an explicit spinlock if
CONFIG_DEBUG_SPINLOCK is off.
Apart from the size difference, note that the meaning of refcount is
also changed to active users. IOWs, it doesn't take an extra refcount
for XArray tree insertion.
I don't observe any significant performance difference at least on
our cloud compute server but the new one indeed simplifies the
overall codebase a bit.
Signed-off-by: Gao Xiang <hsiangkao@linux.alibaba.com>
Reviewed-by: Yue Hu <huyue2@coolpad.com>
Link: https://lore.kernel.org/r/20230529123727.79943-1-hsiangkao@linux.alibaba.com
This patch gets rid of erofs_try_to_free_cached_page() and fold it
into .release_folio().
It also moves managed inode operations into zdata.c, which simplifies
the code a bit. No logic changes.
Signed-off-by: Gao Xiang <hsiangkao@linux.alibaba.com>
Reviewed-by: Yue Hu <huyue2@coolpad.com>
Link: https://lore.kernel.org/r/20230526201459.128169-5-hsiangkao@linux.alibaba.com
After heavily stressing EROFS with several images which include a
hand-crafted image of repeated patterns for more than 46 days, I found
two chains could be linked with each other almost simultaneously and
form a loop so that the entire loop won't be submitted. As a
consequence, the corresponding file pages will remain locked forever.
It can be _only_ observed on data-deduplicated compressed images.
For example, consider two chains with five pclusters in total:
Chain 1: 2->3->4->5 -- The tail pcluster is 5;
Chain 2: 5->1->2 -- The tail pcluster is 2.
Chain 2 could link to Chain 1 with pcluster 5; and Chain 1 could link
to Chain 2 at the same time with pcluster 2.
Since hooked chains are all linked locklessly now, I have no idea how
to simply avoid the race. Instead, let's avoid hooked chains completely
until I could work out a proper way to fix this and end users finally
tell us that it's needed to add it back.
Actually, this optimization can be found with multi-threaded workloads
(especially even more often on deduplicated compressed images), yet I'm
not sure about the overall system impacts of not having this compared
with implementation complexity.
Fixes: 267f2492c8 ("erofs: introduce multi-reference pclusters (fully-referenced)")
Signed-off-by: Gao Xiang <hsiangkao@linux.alibaba.com>
Reviewed-by: Yue Hu <huyue2@coolpad.com>
Link: https://lore.kernel.org/r/20230526201459.128169-4-hsiangkao@linux.alibaba.com
On-stack pagepool is used so that short-lived temporary pages could be
shared within a single I/O request (e.g. among multiple pclusters).
Moving the remaining frontend-related uses into
z_erofs_decompress_frontend to avoid too many arguments.
Signed-off-by: Gao Xiang <hsiangkao@linux.alibaba.com>
Reviewed-by: Yue Hu <huyue2@coolpad.com>
Link: https://lore.kernel.org/r/20230526201459.128169-3-hsiangkao@linux.alibaba.com
If non-bootstrap bvecs cannot be kept in place (very rarely), an extra
short-lived page is allocated.
Let's just allocate it immediately rather than do unnecessary -EAGAIN
return first and retry as a cleanup. Also it's unnecessary to use
__GFP_NOFAIL here since we could gracefully fail out this case instead.
Signed-off-by: Gao Xiang <hsiangkao@linux.alibaba.com>
Reviewed-by: Yue Hu <huyue2@coolpad.com>
Link: https://lore.kernel.org/r/20230526201459.128169-2-hsiangkao@linux.alibaba.com
`end` parameter is no needed since it's pointless for !backmost, we can
handle it with backmost internally. And we only expand the trailing
edge, so the newstart can be replaced with ->headoffset.
Also, remove linux/prefetch.h inclusion since that is not used anymore
after commit 386292919c ("erofs: introduce readmore decompression
strategy").
Signed-off-by: Yue Hu <huyue2@coolpad.com>
Reviewed-by: Gao Xiang <hsiangkao@linux.alibaba.com>
Link: https://lore.kernel.org/r/20230525072605.17857-1-zbestahu@gmail.com
[ Gao Xiang: update commit description. ]
Signed-off-by: Gao Xiang <hsiangkao@linux.alibaba.com>
The struct member is only used to add REQ_RAHEAD during I/O submission.
So it is cleaner to pass it as a parameter than keep it in the struct.
Also, rename function z_erofs_get_sync_decompress_policy() to
z_erofs_is_sync_decompress() for better clarity and conciseness.
Signed-off-by: Yue Hu <huyue2@coolpad.com>
Reviewed-by: Gao Xiang <hsiangkao@linux.alibaba.com>
Link: https://lore.kernel.org/r/20230524063944.1655-1-zbestahu@gmail.com
Signed-off-by: Gao Xiang <hsiangkao@linux.alibaba.com>
No need this helper since it's just a simple wrapper for decompress
method and only one caller. So, let's fold in directly instead.
Signed-off-by: Yue Hu <huyue2@coolpad.com>
Reviewed-by: Gao Xiang <hsiangkao@linux.alibaba.com>
Link: https://lore.kernel.org/r/20230426084449.12781-1-zbestahu@gmail.com
Signed-off-by: Gao Xiang <hsiangkao@linux.alibaba.com>
As Sandeep shown [1], high priority RT per-cpu kthreads are
typically helpful for Android scenarios to minimize the scheduling
latencies.
Switch EROFS_FS_PCPU_KTHREAD_HIPRI on by default if
EROFS_FS_PCPU_KTHREAD is on since it's the typical use cases for
EROFS_FS_PCPU_KTHREAD.
Also clean up unneeded sched_set_normal().
[1] https://lore.kernel.org/r/CAB=BE-SBtO6vcoyLNA9F-9VaN5R0t3o_Zn+FW8GbO6wyUqFneQ@mail.gmail.com
Reviewed-by: Yue Hu <huyue2@coolpad.com>
Reviewed-by: Sandeep Dhavale <dhavale@google.com>
Reviewed-by: Chao Yu <chao@kernel.org>
Signed-off-by: Gao Xiang <hsiangkao@linux.alibaba.com>
Link: https://lore.kernel.org/r/20230522092141.124290-1-hsiangkao@linux.alibaba.com
So that erofs_read_metadata() can read metadata from other inodes
(e.g. packed inode) as well.
Signed-off-by: Jingbo Xu <jefflexu@linux.alibaba.com>
Acked-by: Chao Yu <chao@kernel.org>
Link: https://lore.kernel.org/r/20230407141710.113882-2-jefflexu@linux.alibaba.com
Signed-off-by: Gao Xiang <hsiangkao@linux.alibaba.com>
As the first step of converting hardcoded blocksize to that specified in
on-disk superblock, convert all call sites of hardcoded blocksize to
sb->s_blocksize except for:
1) use sbi->blkszbits instead of sb->s_blocksize in
erofs_superblock_csum_verify() since sb->s_blocksize has not been
updated with the on-disk blocksize yet when the function is called.
2) use inode->i_blkbits instead of sb->s_blocksize in erofs_bread(),
since the inode operated on may be an anonymous inode in fscache mode.
Currently the anonymous inode is allocated from an anonymous mount
maintained in erofs, while in the near future we may allocate anonymous
inodes from a generic API directly and thus have no access to the
anonymous inode's i_sb. Thus we keep the block size in i_blkbits for
anonymous inodes in fscache mode.
Be noted that this patch only gets rid of the hardcoded blocksize, in
preparation for actually setting the on-disk block size in the following
patch. The hard limit of constraining the block size to PAGE_SIZE still
exists until the next patch.
Signed-off-by: Jingbo Xu <jefflexu@linux.alibaba.com>
Reviewed-by: Gao Xiang <hsiangkao@linux.alibaba.com>
Reviewed-by: Yue Hu <huyue2@coolpad.com>
Reviewed-by: Chao Yu <chao@kernel.org>
Link: https://lore.kernel.org/r/20230313135309.75269-2-jefflexu@linux.alibaba.com
[ Gao Xiang: fold a patch to fix incorrect truncated offsets. ]
Link: https://lore.kernel.org/r/20230413035734.15457-1-zhujia.zj@bytedance.com
Signed-off-by: Gao Xiang <hsiangkao@linux.alibaba.com>
Using per-cpu thread pool we can reduce the scheduling latency compared
to workqueue implementation. With this patch scheduling latency and
variation is reduced as per-cpu threads are high priority kthread_workers.
The results were evaluated on arm64 Android devices running 5.10 kernel.
The table below shows resulting improvements of total scheduling latency
for the same app launch benchmark runs with 50 iterations. Scheduling
latency is the latency between when the task (workqueue kworker vs
kthread_worker) became eligible to run to when it actually started
running.
+-------------------------+-----------+----------------+---------+
| | workqueue | kthread_worker | diff |
+-------------------------+-----------+----------------+---------+
| Average (us) | 15253 | 2914 | -80.89% |
| Median (us) | 14001 | 2912 | -79.20% |
| Minimum (us) | 3117 | 1027 | -67.05% |
| Maximum (us) | 30170 | 3805 | -87.39% |
| Standard deviation (us) | 7166 | 359 | |
+-------------------------+-----------+----------------+---------+
Background: Boot times and cold app launch benchmarks are very
important to the Android ecosystem as they directly translate to
responsiveness from user point of view. While EROFS provides
a lot of important features like space savings, we saw some
performance penalty in cold app launch benchmarks in few scenarios.
Analysis showed that the significant variance was coming from the
scheduling cost while decompression cost was more or less the same.
Having per-cpu thread pool we can see from the above table that this
variation is reduced by ~80% on average. This problem was discussed
at LPC 2022. Link to LPC 2022 slides and talk at [1]
[1] https://lpc.events/event/16/contributions/1338/
[ Gao Xiang: At least, we have to add this until WQ_UNBOUND workqueue
issue [2] on many arm64 devices is resolved. ]
[2] https://lore.kernel.org/r/CAJkfWY490-m6wNubkxiTPsW59sfsQs37Wey279LmiRxKt7aQYg@mail.gmail.com
Signed-off-by: Sandeep Dhavale <dhavale@google.com>
Signed-off-by: Gao Xiang <hsiangkao@linux.alibaba.com>
Link: https://lore.kernel.org/r/20230208093322.75816-1-hsiangkao@linux.alibaba.com
Definitions in zdata.h are only used in zdata.c and for internal
use only. No logic changes.
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/20230204093040.97967-4-hsiangkao@linux.alibaba.com
We could just use a boolean in z_erofs_decompressqueue for sync
decompression to simplify the code.
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/20230204093040.97967-2-hsiangkao@linux.alibaba.com
syzkaller reported a memleak:
https://syzkaller.appspot.com/bug?id=62f37ff612f0021641eda5b17f056f1668aa9aed
unreferenced object 0xffff88811009c7f8 (size 136):
...
backtrace:
[<ffffffff821db19b>] z_erofs_do_read_page+0x99b/0x1740
[<ffffffff821dee9e>] z_erofs_readahead+0x24e/0x580
[<ffffffff814bc0d6>] read_pages+0x86/0x3d0
...
syzkaller constructed a case: in z_erofs_register_pcluster(),
ztailpacking = false and map->m_pa = zero. This makes pcl->obj.index be
zero although pcl is not a inline pcluster.
Then following path adds refcount for grp, but the refcount won't be put
because pcl is inline.
z_erofs_readahead()
z_erofs_do_read_page() # for another page
z_erofs_collector_begin()
erofs_find_workgroup()
erofs_workgroup_get()
Since it's illegal for the block address of a non-inlined pcluster to
be zero, add check here to avoid registering the pcluster which would
be leaked.
Fixes: cecf864d3d ("erofs: support inline data decompression")
Reported-by: syzbot+6f8cd9a0155b366d227f@syzkaller.appspotmail.com
Signed-off-by: Chen Zhongjin <chenzhongjin@huawei.com>
Reviewed-by: Yue Hu <huyue2@coolpad.com>
Reviewed-by: Gao Xiang <hsiangkao@linux.alibaba.com>
Reviewed-by: Chao Yu <chao@kernel.org>
Link: https://lore.kernel.org/r/Y42Kz6sVkf+XqJRB@debian
Signed-off-by: Gao Xiang <hsiangkao@linux.alibaba.com>
After commit 4c7e42552b ("erofs: remove useless cache strategy of
DELAYEDALLOC"), only one cached I/O allocation strategy is supported:
When cached I/O is preferred, page allocation is applied without
direct reclaim. If allocation fails, fall back to inplace I/O.
Let's get rid of z_erofs_cache_alloctype. No logical changes.
Reviewed-by: Yue Hu <huyue2@coolpad.com>
Reviewed-by: Chao Yu <chao@kernel.org>
Signed-off-by: Yue Hu <huyue2@coolpad.com>
Signed-off-by: Gao Xiang <hsiangkao@linux.alibaba.com>
Link: https://lore.kernel.org/r/20221206060352.152830-1-xiang@kernel.org
- Fix packed_inode invalid access when reading fragments on crafted
images;
- Add a missing erofs_put_metabuf() in an error path in fscache mode;
- Fix incorrect `count' for unmapped extents in fscache mode;
- Fix use-after-free of fsid and domain_id string when remounting;
- Fix missing xas_retry() in fscache mode.
-----BEGIN PGP SIGNATURE-----
iIcEABYIAC8WIQThPAmQN9sSA0DVxtI5NzHcH7XmBAUCY3OcchEceGlhbmdAa2Vy
bmVsLm9yZwAKCRA5NzHcH7XmBCzgAP92t7Lfu7gBuyhXfCJwJVFK0Iku8j9mhOiT
+C/RVB+9zQEAk/2vy3ULcGN5k6k2q7OgEzNxQ/jM3hVQnoK+sQzwVwA=
=XHmr
-----END PGP SIGNATURE-----
Merge tag 'erofs-for-6.1-rc6-fixes' of git://git.kernel.org/pub/scm/linux/kernel/git/xiang/erofs
Pull erofs fixes from Gao Xiang:
"Most patches randomly fix error paths or corner cases in fscache mode
reported recently. One fixes an invalid access relating to fragments
on crafted images.
Summary:
- Fix packed_inode invalid access when reading fragments on crafted
images
- Add a missing erofs_put_metabuf() in an error path in fscache mode
- Fix incorrect `count' for unmapped extents in fscache mode
- Fix use-after-free of fsid and domain_id string when remounting
- Fix missing xas_retry() in fscache mode"
* tag 'erofs-for-6.1-rc6-fixes' of git://git.kernel.org/pub/scm/linux/kernel/git/xiang/erofs:
erofs: fix missing xas_retry() in fscache mode
erofs: fix use-after-free of fsid and domain_id string
erofs: get correct count for unmapped range in fscache mode
erofs: put metabuf in error path in fscache mode
erofs: fix general protection fault when reading fragment
When psi annotations were added to to btrfs compression reads, the psi
state tracking over add_ra_bio_pages and btrfs_submit_compressed_read was
faulty. A pressure state, once entered, is never left. This results in
incorrectly elevated pressure, which triggers OOM kills.
pflags record the *previous* memstall state when we enter a new one. The
code tried to initialize pflags to 1, and then optimize the leave call
when we either didn't enter a memstall, or were already inside a nested
stall. However, there can be multiple PageWorkingset pages in the bio, at
which point it's that path itself that enters repeatedly and overwrites
pflags. This causes us to miss the exit.
Enter the stall only once if needed, then unwind correctly.
erofs has the same problem, fix that up too. And move the memstall exit
past submit_bio() to restore submit accounting originally added by
b8e24a9300 ("block: annotate refault stalls from IO submission").
Link: https://lkml.kernel.org/r/Y2UHRqthNUwuIQGS@cmpxchg.org
Fixes: 4088a47e78 ("btrfs: add manual PSI accounting for compressed reads")
Fixes: 99486c511f ("erofs: add manual PSI accounting for the compressed address space")
Fixes: 118f3663fb ("block: remove PSI accounting from the bio layer")
Link: https://lore.kernel.org/r/d20a0a85-e415-cf78-27f9-77dd7a94bc8d@leemhuis.info/
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Reported-by: Thorsten Leemhuis <linux@leemhuis.info>
Tested-by: Thorsten Leemhuis <linux@leemhuis.info>
Cc: Chao Yu <chao@kernel.org>
Cc: Chris Mason <clm@fb.com>
Cc: Christoph Hellwig <hch@lst.de>
Cc: David Sterba <dsterba@suse.com>
Cc: Gao Xiang <xiang@kernel.org>
Cc: Jens Axboe <axboe@kernel.dk>
Cc: Josef Bacik <josef@toxicpanda.com>
Cc: Suren Baghdasaryan <surenb@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Partial decompression should be checked after updating length.
It's a new regression when introducing multi-reference pclusters.
Fixes: 2bfab9c0ed ("erofs: record the longest decompressed size in this round")
Reviewed-by: Chao Yu <chao@kernel.org>
Signed-off-by: Gao Xiang <hsiangkao@linux.alibaba.com>
Link: https://lore.kernel.org/r/20221014064915.8103-1-hsiangkao@linux.alibaba.com
