The problem occurs in this scenario:
1. We find a synchronization point.
2. We attmept to create the job.
3. We fail because the job table is full: `mtctx->nextJobID > mtctx->doneJobID + mtctx->jobIDMask`.
4. We call `ZSTDMT_compressStream_generic` again.
5. We forget that we're at a sync point already, and we continue looking
for the next sync point.
This fix is to detect if we're currently paused at a sync point, and if
we are then don't load any more input.
Caught by zstreamtest. I modified it to make the bug occur more often
(~1/100K -> ~1/200) and verified that it is fixed after. I then ran a
few hundred thousand unmodified zstreamtest iterations to verify.
When zstdmt cannot get a buffer and `ZSTD_e_end` is passed an empty
compression job can be created. Additionally, `mtctx->frameEnded` can be
set to 1, which could potentially cause problems like unterminated blocks.
The fix is to adjust to `ZSTD_e_flush` even when we can't get a buffer.
This commit leaves only the functions used by zstd_compress.c. All other
functions have been removed from the API. The ZSTDMT unit tests in
fuzzer.c and zstreamtest.c have been rewritten to use the ZSTD API. And
the --mt zstreamtest tests have been ripped out.
Simplifies the code and removes blocking from zstdmt.
At this point we could completely delete
`ZSTDMT_compress_advanced_internal()`. However I'm leaving it in because
I think we want to do that in the zstd-1.5.0 release, in case anyone is
still using the ZSTDMT API, even though it is not installed by default.
Fixes#2327.
Pass in the `ZSTD_cParamMode_e` to select how we define our cparams.
Based on the mode we either take the `dictSize` into account or we set
it to `0`. See the documentation for `ZSTD_cParamMode_e`.
Some of the modes currently share the same behavior. But they have
distinct modes because they are drastically different cases. E.g.
compression + reprocessing the dictionary and creating a cdict.
Additionally, when downsizing the hashLog and chainLog take the
(adjusted) dictionary size into account, since the size of the
dictionary gets added onto the window size.
Adds a simple test to ensure that we aren't downsizing too far.
The DDS structure can't be copied into the working tables like the DMS.
So it doesn't need to account for the source size when sizing its
parameters, just the dictionary size.
Conditions to trigger:
* CDict is loaded as raw content.
* CDict starts with the zstd dictionary magic number.
* The CDict is reprocessed (not attached or copied).
* The new API is used (streaming or `ZSTD_compress2()`).
Bug: The dictionary is loaded as a zstd dictionary, not a raw content
dictionary, because the dict content type is set to `ZSTD_dct_auto`.
Fix: Pass in the dictionary content type from cdict creation to the call
to `ZSTD_compress_insertDictionary()`.
Test: Added a test case that exposes the bug, and fixed the raw
content tests to not modify the `dictBuffer`, which makes all future
tests with the `dictBuffer` raw content, which doesn't seem intentional.
Rename ADDRESS_SANITIZER -> ZSTD_ADDRESS_SANITIZER and same for
MEMORY_SANITIZER. Also set it to 0/1 instead of checking for defined.
This allows the user to override ASAN/MSAN detection for platforms that
don't support it.
This clarifies operator precedence, and quiets cppcheck in
the Kernel Test Robot. I think this is a slight bonus to
readability, so I am accepting the suggestion.
Even if the discrepancies are at the moment benign, it's probably better to
standardize on using the one true initializer, rather than trying (and failing)
to correctly duplicate its behavior.
Rewrite ZSTD_createCDict_advanced() as a wrapper around
ZSTD_createCDict_advanced2(). Evaluate whether to use DDSS mode *after* fully
resolving cparams. If not, fall back.
Rather than restrict our temp chain table to 2 ** chainLog entries, this
commit uses all available space to reach further back to gather longer
chains to pack into the DDSS chain table.
Rather than interleave all of the chain table entries, tying each entry's
position to the corresponding position in the input, this commit changes the
layout so that all the entries in a single chain are laid out next to each
other. The last entry in the hash table's bucket for this hash is now a packed
pointer of position + length of this chain.
This cannot be merged as written, since it allocates temporary memory inside
ZSTD_dedicatedDictSearch_lazy_loadDictionary().
Previously, if DDSS was enabled on a CCtx and a dictionary was inserted into
the CCtx, the CCtx MatchState would be filled as a DDSS struct, causing
segfaults etc. This changes the check to use whether the MatchState is marked
as using the DDSS (which is only ever set for CDict MatchStates), rather than
looking at the CCtxParams.
Entries in the hashTable chain cache aren't subject to the same aliasing that
the circular chain table is subject to. As such, we don't need to stop when we
cross the chain limit. We can delve deeper. :)
This caused us to double-search the first position and fail to search the
last position in the chain, slowing down search and making it less effective.
This makes it clear that not only is the feature allowed here, we're actually
using it, as opposed to the CCtxParam field, in which it's enabled, but we may
or may not be using it.
The unused function definitions are hidden behind a
`#ifndef ZSTD_NO_UNUSED_FUNCTIONS` check.
Initially hiding all functions which are unused and take up more than
2KB of stack space, because these will show up as warnings in the
Linux Kernel build system.
Unintentional integer overflow (OVERFLOW_BEFORE_WIDEN)
overflow_before_widen: Potentially overflowing expression:
cdict->dictContentSize * 6U
with type unsigned int (32 bits, unsigned) is evaluated using 32-bit
arithmetic, and then used in a context that expects an expression of
type U64 (64 bits, unsigned).
new version easier to vectorize
leads to smaller code and faster execution
notably at the last recombination stage
(basically, fixed cost per block).
Assembly inspected with godbolt
On my laptop, with `clang` and `-mavx2` :
2K block : 1280 MB/s -> 1550 MB/s
8K block : 1750 MB/s -> 1860 MB/s
The bound check condition should always be met because we selected `set_basic` as
our encoding type. But that code is very far away, so assert it is true so if it is
ever false we can catch it, and add a bounds check.
Fixes#2213.
Allow compression to use dictionaries with missing symbols in their
entropy tables. We set the FSE repeat mode to check when there are
missing symbols, and set the FSE repeat mode to valid when all symbols
are present.
Note that when not all symbols are present, the heuristics which favor
dictionary tables for lower compression levels won't activate.
Tested by manually creating a dictionary with missing symbols of every
type, and validing that the compressor rejects it before this change,
and accepts it after this change. Also, I ran the `dictionary_loader`
fuzzer for >1 hour of CPU time without running into cases where
compression succeeds, but decompression fails.
Fixes#2174.
Exposed when loading a dictionary < LDM minMatch bytes in MT mode.
Test Plan:
```
CC=clang make -j zstreamtest MOREFLAGS="-O0 -fsanitize=address"
./zstreamtest -vv -i100000000 -t1 --newapi -s7065 -t3925297
```
TODO: Add an explicit test that loads a small dictionary in MT mode
This commit pulls out the internals of `ZSTD_estimateCCtxSize_usingCCtxParams`
into a helper. It then migrates two other callsites to use that helper,
a small optimization for `ZSTD_estimateCStreamSize_usingCCtxParams`, which
folds the buffer sizing into the helper, and then `ZSTD_resetCCtx_internal`,
which is more invasive.
This attempts to guarantee that the estimates returned to users are always
correct.
`ZSTD_estimateCCtxSize()` provides estimates for one-shot compression, which
is guaranteed not to buffer inputs or outputs. So it ignores the sizes of the
buffers, assuming they'll be zero. However, the actual workspace allocation
logic always allocates those buffers, and when running under ASAN, the
workspace surrounds every allocation with 256 bytes of redzone. So the 0-sized
buffers end up consuming 512 bytes of space, which is accounted for in the
actual allocation path through the use of `ZSTD_cwksp_alloc_size()` but isn't
in the estimation path, since it ignores the buffers entirely.
This commit fixes this.
Fixes:
Enable RLE blocks for superblock mode
Fix the limitation that the literals block must shrink. Instead, when we're within 200 bytes of the next header byte size, we will just use the next one up. That way we should (almost?) always have space for the table.
Remove the limitation that the first sub-block MUST have compressed literals and be compressed. Now one sub-block MUST be compressed (otherwise we fall back to raw block which is okay, since that is streamable). If no block has compressed literals that is okay, we will fix up the next Huffman table.
Handle the case where the last sub-block is uncompressed (maybe it is very small). Before it would skip superblock in this case, now we allow the last sub-block to be uncompressed. To do this we need to regenerate the correct repcodes.
Respect disableLiteralsCompression in superblock mode
Fix superblock mode to handle a block consisting of only compressed literals
Fix a off by 1 error in superblock mode that disabled it whenever there were last literals
Fix superblock mode with long literals/matches (> 0xFFFF)
Allow superblock mode to repeat Huffman tables
Respect ZSTD_minGain().
Tests:
Simple check for the condition in #2096.
When the simple_round_trip fuzzer enables superblock mode, it checks that the compressed size isn't expanded too much.
Remaining limitations:
O(targetCBlockSize^2) because we recompute statistics every sequence
Unable to split literals of length > targetCBlockSize into multiple sequences
Refuses to generate sub-blocks that don't shrink the compressed data, so we could end up with large sub-blocks. We should emit those sections as uncompressed blocks instead.
...
Fixes#2096
* adding long support for patch-from
* adding refPrefix to dictionary_decompress
* adding refPrefix to dictionary_loader
* conversion nit
* triggering log mode on chainLog < fileLog and removing old threshold
* adding refPrefix to dictionary_round_trip
* adding docs
* adding enableldm + forceWindow test for dict
* separate patch-from logic into FIO_adjustParamsForPatchFromMode
* moving memLimit adjustment to outside ifdefs (need for decomp)
* removing refPrefix gate on dictionary_round_trip
* rebase on top of dev refPrefix change
* making sure refPrefx + ldm is < 1% of srcSize
* combining notes for patch-from
* moving memlimit logic inside fileio.c
* adding display for optimal parser and long mode trigger
* conversion nit
* fuzzer found heap-overflow fix
* another conversion nit
* moving FIO_adjustMemLimitForPatchFromMode outside ifndef
* making params immutable
* moving memLimit update before createDictBuffer call
* making maxSrcSize unsigned long long
* making dictSize and maxSrcSize params unsigned long long
* error on files larger than 4gb
* extend refPrefix test to include round trip
* conversion to size_t
* making sure ldm is at least 10x better
* removing break
* including zstd_compress_internal and removing redundant macros
* exposing ZSTD_cycleLog()
* using cycleLog instead of chainLog
* add some more docs about user optimizations
* formatting
`CHECK_F` macro moved to `error_private.h` (shared between `fse_compress.c` and `fse_decompress.c`). `ZSTD_limitCopy()` moved to `zstd_internal.h` (shared between `zstd_compress.c` and `zstd_decompress.c`). Erroneous build artefact `zstd.h` removed from repo.
To complement the single-file decoder a new script was added to create an amalgamated single-file of all of the Zstd source, along with examples and (simple) tests.
* All copyright lines now have -2020 instead of -present
* All copyright lines include "Facebook, Inc"
* All licenses are now standardized
The copyright in `threading.{h,c}` is not changed because it comes from
zstdmt.
The copyright and license of `divsufsort.{h,c}` is not changed.
Super blocks must never violate the zstd block bound of input_size + ZSTD_blockHeaderSize. The individual sub-blocks may, but not the super block. If the superblock violates the block bound we are liable to violate ZSTD_compressBound(), which we must not do. Whenever the super block violates the block bound we instead emit an uncompressed block.
This means we increase the latency because of the single uncompressed block. I fix this by enabling streaming an uncompressed block, so the latency of an uncompressed block is 1 byte. This doesn't reduce the latency of the buffer-less API, but I don't think we really care.
* I added a test case that verifies that the decompression has 1 byte latency.
* I rely on existing zstreamtest / fuzzer / libfuzzer regression tests for correctness. During development I had several correctness bugs, and they easily caught them.
* The added assert that the superblock doesn't violate the block bound will help us discover any missed conditions (though I think I got them all).
Credit to OSS-Fuzz.
* Allow zero sized buffers in `stream_decompress`. Ensure that we never have two
zero sized buffers in a row so we guarantee forwards progress.
* Make case 4 in `stream_round_trip` do a zero sized buffers call followed by
a full call to guarantee forwards progress.
* Fix `limitCopy()` in legacy decoders.
* Fix memcpy in `zstdmt_compress.c`.
Catches the bug fixed in PR #1939
* Adding fail logging for superblock flow
* Dividing by targetCBlockSize instead of blockSize
* Adding new const and using more acurate formula for nbBlocks
* Only do dstCapacity check if using superblock
* Remvoing disabling logic
* Updating test to make it catch more extreme case of previou bug
* Also updating comment
* Only taking compressEnd shortcut on non-superblock
Fixes new fuzz issue
Credit to OSS-Fuzz
* Initializing unsigned value
* Initialilzing to 1 instead of 0 because its more conservative
* Unconditionoally setting to check first and then checking zero
* Moving bool to before block for c90
* Move check set before block
Fixes a fuzz issue where dictionary_round_trip failed because the compressor was generating corrupt files thanks to zero weights in the table.
* Only setting loaded dict huf table to valid on non-zero
* Adding hasNoZeroWeights test to fse tables
* Forbiding nbBits != 0 when weight == 0
* Reverting the last commit
* Setting table log to 0 when weight == 0
* Small (invalid) zero weight dict test
* Small (valid) zero weight dict test
* Initializing repeatMode vars to check before zero check
* Removing FSE changes to seperate pr
* Reverting accidentally changed file
* Negating bool, using unsigned, optimization nit
This parameter is unused in single-threaded compression. We should make it
behave like the other multithread-only parameters, for which we only accept
zero when we are not built with multithreading.
* Silently skip dictionaries less than 8 bytes, unless using `ZSTD_dct_fullDict`.
This changes the compressor, which silently skips dictionaries <= 8 bytes.
* Allow repcodes that are equal to the dictionary content size, since it is in bounds.
Compression ratio of fast strategies (levels 1 & 2)
was seriously reduced, due to accidental disabling of Literals compression.
Credit to @QrczakMK, which perfectly described the issue, and implementation details,
making the fix straightforward.
Example : initCStream with level 1 on synthetic sample P50 :
Before : 5,273,976 bytes
After : 3,154,678 bytes
ZSTD_compress (for comparison) : 3,154,550
Fix#1787.
To follow : refactor the test which was supposed to catch this issue (and failed)
* Bump `WILDCOPY_OVERLENGTH` to 16 to fix the wildcopy overread.
* Optimize `ZSTD_wildcopy()` by removing unnecessary branches and
unrolling the loop.
* Extract `ZSTD_overlapCopy8()` into its own function.
* Add `ZSTD_safecopy()` for `ZSTD_execSequenceEnd()`. It is
optimized for single long sequences, since that is the important
case that can end up in `ZSTD_execSequenceEnd()`. Without this
optimization, decompressing a block with 1 long match goes
from 5.7 GB/s to 800 MB/s.
* Refactor `ZSTD_execSequenceEnd()`.
* Increase the literal copy shortcut to 16.
* Add a shortcut for offset >= 16.
* Simplify `ZSTD_execSequence()` by pushing more cases into
`ZSTD_execSequenceEnd()`.
* Delete `ZSTD_execSequenceLong()` since it is exactly the
same as `ZSTD_execSequence()`.
clang-8 seeds +17.5% on silesia and +21.8% on enwik8.
gcc-9 sees +12% on silesia and +15.5% on enwik8.
TODO: More detailed measurements, and on more datasets.
Crdit to OSS-Fuzz for finding the wildcopy overread.
This led to a nasty edgecase, where index reduction for modes that don't use
the h3 table would have a degenerate table (size 4) allocated and marked clean,
but which would not be re-indexed.
The source matchState is potentially at a lower current index, which means
that any extra table space not overwritten by the copy may now contain
invalid indices. The simple solution is to unconditionally shrink the valid
table area to just the area overwritten.
Summary: The idea behind wildcopy is that it can be cheaper to copy more bytes (say 8) than it is to copy less (say, 3). This change takes that further by exploiting some properties:
1. it's almost always OK to copy 16 bytes instead of 8, which means fewer copy instructions, and fewer branches
2. A 16 byte chunk size means that ~90% of wildcopy invocations will have a trip count of 1, so branch prediction will be improved.
Speedup on Xeon E5-2680v4 is in the range of 3-5%.
Measured wildcopy length distributions on silesia.tar:
level <=8 <=16 <=24 >24
1 78.05% 11.49% 3.52% 6.94%
3 82.14% 8.99% 2.44% 6.43%
6 85.81% 6.51% 2.92% 4.76%
8 83.02% 7.31% 3.64% 6.03%
10 84.13% 6.67% 3.29% 5.91%
15 77.58% 7.55% 5.21% 9.66%
16 80.07% 7.20% 3.98% 8.75%
Test Plan: benchmark silesia, make check
The nbSeq "short" format (1-byte)
is compatible with any value < 128.
However, the code would cautiously only accept values < 127.
This is not an error, because the general 2-bytes format
is compatible with small values < 128.
Hence the inefficiency never triggered any warning.
Spotted by Intel's Smita Kumar.
* [ldm] Fix bug in overflow correction with large job size
* [zstdmt] Respect ZSTDMT_JOBSIZE_MAX (1G in 64-bit mode)
* [test] Add test that exposes the bug
Sadly the test fails on our CI because it uses too much memory, so
I had to comment it out.
When we wrote one byte beyond the end of the buffer for RLE
blocks back in 1.3.7, we would then have `op > oend`. That is
a problem when we use `oend - op` for the size of the destination
buffer, and allows further writes beyond the end of the buffer for
the rest of the function. Lets assert that it doesn't happen.
Also : minor speed optimization :
shortcut to ZSTD_reset_matchState() rather than the full reset process.
It still needs to be completed with ZSTD_continueCCtx() for proper initialization.
Also : changed position of LDM hash tables in the context,
so that the "regular" hash tables can be at a predictable position,
hence allowing the shortcut to ZSTD_reset_matchState() without complex conditions.
* Extract the overflow correction into a helper function.
* Load the dictionary `ZSTD_CHUNKSIZE_MAX = 512 MB` bytes at a time
and overflow correct between each chunk.
Data corruption could happen when all these conditions are true:
* You are using multithreading mode
* Your overlap size is >= 512 MB (implies window size >= 512 MB)
* You are using a strategy >= ZSTD_btlazy
* You are compressing more than 4 GB
The problem is that when loading a large dictionary we don't do
overflow correction. We can only load 512 MB at a time, and may
need to do overflow correction before each chunk.
We would only skip at most 192 bytes at a time before this diff.
This was added to optimize long matches and skip the middle of the
match. However, it doesn't handle the case of repetitive data.
This patch keeps the optimization, but also handles repetitive data
by taking the max of the two return values.
```
> for n in $(seq 9); do echo strategy=$n; dd status=none if=/dev/zero bs=1024k count=1000 | command time -f %U ./zstd --zstd=strategy=$n >/dev/null; done
strategy=1
0.27
strategy=2
0.23
strategy=3
0.27
strategy=4
0.43
strategy=5
0.56
strategy=6
0.43
strategy=7
0.34
strategy=8
0.34
strategy=9
0.35
```
At level 19 with multithreading the compressed size of `silesia.tar` regresses 300 bytes, and `enwik8` regresses 100 bytes.
In single threaded mode `enwik8` is also within 100 bytes, and I didn't test `silesia.tar`.
Fixes Issue #1634.
fast mode does the same thing as before :
it pre-emptively invalidates any index that could lead to offset > maxDistance.
It's supposed to help speed.
But this logic is performed inside zstd_fast,
so that other strategies can select a different behavior.
It's re-synchronized with nextToUpdate at beginning of each block.
It only needs to be tracked from within zstd_opt block parser.
Made the logic clear, so that no code tried to maintain this variable.
An even better solution would be to make nextToUpdate3
an internal variable of ZSTD_compressBlock_opt_generic().
That would make it possible to remove it from ZSTD_matchState_t,
thus restricting its visibility to only where it's actually useful.
This would require deeper changes though,
since the matchState is the natural structure to transport parameters into and inside the parser.
ZSTDMT was broken when compiled without ZSTD_MULTITHREAD defined,
because `ZSTD_CCtx_setParameter(cctx, ZSTD_c_nbWorkers, nbWorkerss)`
failed. It was detected by the MSVC test which runs the fuzzer with
multithreading disabled.
This is a very niche use case of a deprecated API, because the API is
inefficient and synchronous, since `threading.h` will be synchronous.
Users almost certainly don't want this, and anyone who tested their code
should realize that it is broken. Therefore, I think it is safe to
require `ZSTD_MULTITHREAD` to be defined to use ZSTDMT.
`ZSTD_compress2()` wouldn't wait for multithreaded compression to
finish. We didn't find this because ZSTDMT will block when it can
compress all in one go, but it can't do that if it doesn't have enough
output space, or if `ZSTD_c_rsyncable` is enabled.
Since we will already sometimes block when using `ZSTD_e_end`, I've
changed `ZSTD_e_end` and `ZSTD_e_flush` to guarantee maximum forward
progress. This simplifies the API, and helps users avoid the easy bug
that was made in `ZSTD_compress2()`
* Found by the libfuzzer fuzzers.
* Added a test case that catches the problem.
* I will make the fuzzers sometimes allocate less than
`ZSTD_compressBound()` output space.
This PR is based on top of PR #1563.
The optimization is to process two input pointers per loop.
It is based on ideas from [igzip] level 1, and talking to @gbtucker.
| Platform | Silesia | Enwik8 |
|-------------------------|-------------|--------|
| OSX clang-10 | +5.3% | +5.4% |
| i9 5 GHz gcc-8 | +6.6% | +6.6% |
| i9 5 GHz clang-7 | +8.0% | +8.0% |
| Skylake 2.4 GHz gcc-4.8 | +6.3% | +7.9% |
| Skylake 2.4 GHz clang-7 | +6.2% | +7.5% |
Testing on all Silesia files on my Intel i9-9900k with gcc-8
| Silesia File | Ratio Change | Speed Change |
|--------------|--------------|--------------|
| silesia.tar | +0.17% | +6.6% |
| dickens | +0.25% | +7.0% |
| mozilla | +0.02% | +6.8% |
| mr | -0.30% | +10.9% |
| nci | +1.28% | +4.5% |
| ooffice | -0.35% | +10.7% |
| osdb | +0.75% | +9.8% |
| reymont | +0.65% | +4.6% |
| samba | +0.70% | +5.9% |
| sao | -0.01% | +14.0% |
| webster | +0.30% | +5.5% |
| xml | +0.92% | +5.3% |
| x-ray | -0.00% | +1.4% |
Same tests on Calgary. For brevity, I've only included files
where compression ratio regressed or was much better.
| Calgary File | Ratio Change | Speed Change |
|--------------|--------------|--------------|
| calgary.tar | +0.30% | +7.1% |
| geo | -0.14% | +25.0% |
| obj1 | -0.46% | +15.2% |
| obj2 | -0.18% | +6.0% |
| pic | +1.80% | +9.3% |
| trans | -0.35% | +5.5% |
We gain 0.1% of compression ratio on Silesia.
We gain 0.3% of compression ratio on enwik8.
I also tested on the GitHub and hg-commands datasets without a dictionary,
and we gain a small amount of compression ratio on each, as well as speed.
I tested the negative compression levels on Silesia on my
Intel i9-9900k with gcc-8:
| Level | Ratio Change | Speed Change |
|-------|--------------|--------------|
| -1 | +0.13% | +6.4% |
| -2 | +4.6% | -1.5% |
| -3 | +7.5% | -4.8% |
| -4 | +8.5% | -6.9% |
| -5 | +9.1% | -9.1% |
Roughly, the negative levels now scale half as quickly. E.g. the new
level 16 is roughly equivalent to the old level 8, but a bit quicker
and smaller. If you don't think this is the right trade off, we can
change it to multiply the step size by 2, instead of adding 1. I think
this makes sense, because it gives a bit slower ratio decay.
[igzip]: https://github.com/01org/isa-l/tree/master/igzip
