With CONFIG_ZSTD_COMPRESS=m and CONFIG_ZSTD_DECOMPRESS=y we end up in
a situation when files from lib/zstd/common/ are compiled once to be
linked later for ZSTD_DECOMPRESS (build-in) and ZSTD_COMPRESS (module)
even though CFLAGS are different for builtins and modules.
So far somehow this was not a problem but enabling LLVM LTO exposes
the problem as:
ld.lld: error: linking module flags 'Code Model': IDs have conflicting values in 'lib/built-in.a(zstd_common.o at 5868)' and 'ld-temp.o'
This particular conflict is caused by KBUILD_CFLAGS=-mcmodel=medium vs.
KBUILD_CFLAGS_MODULE=-mcmodel=large , modules use the large model on
POWERPC as explained at
https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/tree/arch/powerpc/Makefile?h=v5.18-rc4#n127
but the current use of common files is wrong anyway.
This works around the issue by introducing a zstd_common module with
shared code.
Signed-off-by: Alexey Kardashevskiy <aik@ozlabs.ru>
Signed-off-by: Masahiro Yamada <masahiroy@kernel.org>
After the update to zstd-1.4.10 passing -O3 is no longer necessary to
get good performance from zstd. Using the default optimization level -O2
is sufficient to get good performance.
I've measured no significant change to compression speed, and a ~1%
decompression speed loss, which is acceptable.
This fixes the reported parisc -Wframe-larger-than=1536 errors [0]. The
gcc-8-hppa-linux-gnu compiler performed very poorly with -O3, generating
stacks that are ~3KB. With -O2 these same functions generate stacks in
the < 100B, completely fixing the problem. Function size deltas are
listed below:
ZSTD_compressBlock_fast_extDict_generic: 3800 -> 68
ZSTD_compressBlock_fast: 2216 -> 40
ZSTD_compressBlock_fast_dictMatchState: 1848 -> 64
ZSTD_compressBlock_doubleFast_extDict_generic: 3744 -> 76
ZSTD_fillDoubleHashTable: 3252 -> 0
ZSTD_compressBlock_doubleFast: 5856 -> 36
ZSTD_compressBlock_doubleFast_dictMatchState: 5380 -> 84
ZSTD_copmressBlock_lazy2: 2420 -> 72
Additionally, this improves the reported code bloat [1]. With gcc-11
bloat-o-meter shows an 80KB code size improvement:
```
> ../scripts/bloat-o-meter vmlinux.old vmlinux
add/remove: 31/8 grow/shrink: 24/155 up/down: 25734/-107924 (-82190)
Total: Before=6418562, After=6336372, chg -1.28%
```
Compared to before the zstd-1.4.10 update we see a total code size
regression of 105KB, down from 374KB at v5.16-rc1:
```
> ../scripts/bloat-o-meter vmlinux.old vmlinux
add/remove: 292/62 grow/shrink: 56/88 up/down: 235009/-127487 (107522)
Total: Before=6228850, After=6336372, chg +1.73%
```
[0] https://lkml.org/lkml/2021/11/15/710
[1] https://lkml.org/lkml/2021/11/14/189
Link: https://lore.kernel.org/r/20211117014949.1169186-4-nickrterrell@gmail.com/
Link: https://lore.kernel.org/r/20211117201459.1194876-4-nickrterrell@gmail.com/
Reported-by: Geert Uytterhoeven <geert@linux-m68k.org>
Tested-by: Geert Uytterhoeven <geert@linux-m68k.org>
Reviewed-by: Geert Uytterhoeven <geert@linux-m68k.org>
Signed-off-by: Nick Terrell <terrelln@fb.com>
Upgrade to the latest upstream zstd version 1.4.10.
This patch is 100% generated from upstream zstd commit 20821a46f412 [0].
This patch is very large because it is transitioning from the custom
kernel zstd to using upstream directly. The new zstd follows upstreams
file structure which is different. Future update patches will be much
smaller because they will only contain the changes from one upstream
zstd release.
As an aid for review I've created a commit [1] that shows the diff
between upstream zstd as-is (which doesn't compile), and the zstd
code imported in this patch. The verion of zstd in this patch is
generated from upstream with changes applied by automation to replace
upstreams libc dependencies, remove unnecessary portability macros,
replace `/**` comments with `/*` comments, and use the kernel's xxhash
instead of bundling it.
The benefits of this patch are as follows:
1. Using upstream directly with automated script to generate kernel
code. This allows us to update the kernel every upstream release, so
the kernel gets the latest bug fixes and performance improvements,
and doesn't get 3 years out of date again. The automation and the
translated code are tested every upstream commit to ensure it
continues to work.
2. Upgrades from a custom zstd based on 1.3.1 to 1.4.10, getting 3 years
of performance improvements and bug fixes. On x86_64 I've measured
15% faster BtrFS and SquashFS decompression+read speeds, 35% faster
kernel decompression, and 30% faster ZRAM decompression+read speeds.
3. Zstd-1.4.10 supports negative compression levels, which allow zstd to
match or subsume lzo's performance.
4. Maintains the same kernel-specific wrapper API, so no callers have to
be modified with zstd version updates.
One concern that was brought up was stack usage. Upstream zstd had
already removed most of its heavy stack usage functions, but I just
removed the last functions that allocate arrays on the stack. I've
measured the high water mark for both compression and decompression
before and after this patch. Decompression is approximately neutral,
using about 1.2KB of stack space. Compression levels up to 3 regressed
from 1.4KB -> 1.6KB, and higher compression levels regressed from 1.5KB
-> 2KB. We've added unit tests upstream to prevent further regression.
I believe that this is a reasonable increase, and if it does end up
causing problems, this commit can be cleanly reverted, because it only
touches zstd.
I chose the bulk update instead of replaying upstream commits because
there have been ~3500 upstream commits since the 1.3.1 release, zstd
wasn't ready to be used in the kernel as-is before a month ago, and not
all upstream zstd commits build. The bulk update preserves bisectablity
because bugs can be bisected to the zstd version update. At that point
the update can be reverted, and we can work with upstream to find and
fix the bug.
Note that upstream zstd release 1.4.10 doesn't exist yet. I have cut a
staging branch at 20821a46f412 [0] and will apply any changes requested
to the staging branch. Once we're ready to merge this update I will cut
a zstd release at the commit we merge, so we have a known zstd release
in the kernel.
The implementation of the kernel API is contained in
zstd_compress_module.c and zstd_decompress_module.c.
[0] 20821a46f4
[1] e0fa481d0e
Signed-off-by: Nick Terrell <terrelln@fb.com>
Tested By: Paul Jones <paul@pauljones.id.au>
Tested-by: Oleksandr Natalenko <oleksandr@natalenko.name>
Tested-by: Sedat Dilek <sedat.dilek@gmail.com> # LLVM/Clang v13.0.0 on x86-64
Tested-by: Jean-Denis Girard <jd.girard@sysnux.pf>
Add SPDX license identifiers to all Make/Kconfig files which:
- Have no license information of any form
These files fall under the project license, GPL v2 only. The resulting SPDX
license identifier is:
GPL-2.0-only
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Now, Kbuild nicely handles composite objects to avoid multiple
definition.
Makefiles can simply add the same objects multiple times across
composite objects.
Signed-off-by: Masahiro Yamada <yamada.masahiro@socionext.com>
Add zstd compression and decompression kernel modules.
zstd offers a wide varity of compression speed and quality trade-offs.
It can compress at speeds approaching lz4, and quality approaching lzma.
zstd decompressions at speeds more than twice as fast as zlib, and
decompression speed remains roughly the same across all compression levels.
The code was ported from the upstream zstd source repository. The
`linux/zstd.h` header was modified to match linux kernel style.
The cross-platform and allocation code was stripped out. Instead zstd
requires the caller to pass a preallocated workspace. The source files
were clang-formatted [1] to match the Linux Kernel style as much as
possible. Otherwise, the code was unmodified. We would like to avoid
as much further manual modification to the source code as possible, so it
will be easier to keep the kernel zstd up to date.
I benchmarked zstd compression as a special character device. I ran zstd
and zlib compression at several levels, as well as performing no
compression, which measure the time spent copying the data to kernel space.
Data is passed to the compresser 4096 B at a time. The benchmark file is
located in the upstream zstd source repository under
`contrib/linux-kernel/zstd_compress_test.c` [2].
I ran the benchmarks on a Ubuntu 14.04 VM with 2 cores and 4 GiB of RAM.
The VM is running on a MacBook Pro with a 3.1 GHz Intel Core i7 processor,
16 GB of RAM, and a SSD. I benchmarked using `silesia.tar` [3], which is
211,988,480 B large. Run the following commands for the benchmark:
sudo modprobe zstd_compress_test
sudo mknod zstd_compress_test c 245 0
sudo cp silesia.tar zstd_compress_test
The time is reported by the time of the userland `cp`.
The MB/s is computed with
1,536,217,008 B / time(buffer size, hash)
which includes the time to copy from userland.
The Adjusted MB/s is computed with
1,536,217,088 B / (time(buffer size, hash) - time(buffer size, none)).
The memory reported is the amount of memory the compressor requests.
| Method | Size (B) | Time (s) | Ratio | MB/s | Adj MB/s | Mem (MB) |
|----------|----------|----------|-------|---------|----------|----------|
| none | 11988480 | 0.100 | 1 | 2119.88 | - | - |
| zstd -1 | 73645762 | 1.044 | 2.878 | 203.05 | 224.56 | 1.23 |
| zstd -3 | 66988878 | 1.761 | 3.165 | 120.38 | 127.63 | 2.47 |
| zstd -5 | 65001259 | 2.563 | 3.261 | 82.71 | 86.07 | 2.86 |
| zstd -10 | 60165346 | 13.242 | 3.523 | 16.01 | 16.13 | 13.22 |
| zstd -15 | 58009756 | 47.601 | 3.654 | 4.45 | 4.46 | 21.61 |
| zstd -19 | 54014593 | 102.835 | 3.925 | 2.06 | 2.06 | 60.15 |
| zlib -1 | 77260026 | 2.895 | 2.744 | 73.23 | 75.85 | 0.27 |
| zlib -3 | 72972206 | 4.116 | 2.905 | 51.50 | 52.79 | 0.27 |
| zlib -6 | 68190360 | 9.633 | 3.109 | 22.01 | 22.24 | 0.27 |
| zlib -9 | 67613382 | 22.554 | 3.135 | 9.40 | 9.44 | 0.27 |
I benchmarked zstd decompression using the same method on the same machine.
The benchmark file is located in the upstream zstd repo under
`contrib/linux-kernel/zstd_decompress_test.c` [4]. The memory reported is
the amount of memory required to decompress data compressed with the given
compression level. If you know the maximum size of your input, you can
reduce the memory usage of decompression irrespective of the compression
level.
| Method | Time (s) | MB/s | Adjusted MB/s | Memory (MB) |
|----------|----------|---------|---------------|-------------|
| none | 0.025 | 8479.54 | - | - |
| zstd -1 | 0.358 | 592.15 | 636.60 | 0.84 |
| zstd -3 | 0.396 | 535.32 | 571.40 | 1.46 |
| zstd -5 | 0.396 | 535.32 | 571.40 | 1.46 |
| zstd -10 | 0.374 | 566.81 | 607.42 | 2.51 |
| zstd -15 | 0.379 | 559.34 | 598.84 | 4.61 |
| zstd -19 | 0.412 | 514.54 | 547.77 | 8.80 |
| zlib -1 | 0.940 | 225.52 | 231.68 | 0.04 |
| zlib -3 | 0.883 | 240.08 | 247.07 | 0.04 |
| zlib -6 | 0.844 | 251.17 | 258.84 | 0.04 |
| zlib -9 | 0.837 | 253.27 | 287.64 | 0.04 |
Tested in userland using the test-suite in the zstd repo under
`contrib/linux-kernel/test/UserlandTest.cpp` [5] by mocking the kernel
functions. Fuzz tested using libfuzzer [6] with the fuzz harnesses under
`contrib/linux-kernel/test/{RoundTripCrash.c,DecompressCrash.c}` [7] [8]
with ASAN, UBSAN, and MSAN. Additionaly, it was tested while testing the
BtrFS and SquashFS patches coming next.
[1] https://clang.llvm.org/docs/ClangFormat.html
[2] https://github.com/facebook/zstd/blob/dev/contrib/linux-kernel/zstd_compress_test.c
[3] http://sun.aei.polsl.pl/~sdeor/index.php?page=silesia
[4] https://github.com/facebook/zstd/blob/dev/contrib/linux-kernel/zstd_decompress_test.c
[5] https://github.com/facebook/zstd/blob/dev/contrib/linux-kernel/test/UserlandTest.cpp
[6] http://llvm.org/docs/LibFuzzer.html
[7] https://github.com/facebook/zstd/blob/dev/contrib/linux-kernel/test/RoundTripCrash.c
[8] https://github.com/facebook/zstd/blob/dev/contrib/linux-kernel/test/DecompressCrash.c
zstd source repository: https://github.com/facebook/zstd
Signed-off-by: Nick Terrell <terrelln@fb.com>
Signed-off-by: Chris Mason <clm@fb.com>
Alexey Kardashevskiy