Motivation:
writeUtf8 can suffer from inlining issues and/or megamorphic call-sites on the hot path due to ByteBuf hierarchy
Modifications:
Duplicate and specialize the code paths to reduce the need of polymorphic calls
Result:
Performance are more stable in user code
Motivation:
If ByteBufUtil.getBytes() is called with copy=false, it does not
correctly check that the underlying array can be shared in some cases.
In particular:
* It does not check that the arrayOffset() is zero. This causes it to
incorrectly return the underlying array if the other conditions are
met. The returned array will be longer than requested, with additional
unwanted bytes at its start.
* It assumes that the capacity() of the ByteBuf is equal to the backing
array length. This is not true for some types of ByteBuf, such as
PooledHeapByteBuf. This causes it to incorrectly return the underlying
array if the other conditions are met. The returned array will be
longer than requested, with additional unwanted bytes at its end.
Modifications:
This commit fixes the two bugs by:
* Checking that the arrayOffset() is zero before returning the
underlying array.
* Comparing the requested length to the underlying array's length,
rather than the ByteBuf's capacity, before returning the underlying
array.
This commit also adds a series of test cases for ByteBufUtil.getBytes().
Result:
ByteBufUtil.getBytes() now correctly checks whether the underlying array
can be shared or not.
The test cases will ensure the bug is not reintroduced in the future.
Motivation
This is used solely for the DataOutput#writeUTF8() method, which may
often not be used.
Modifications
Lazily construct the contained DataOutputStream in ByteBufOutputStream.
Result
Saves an allocation in some common cases
Motivation
ByteBuf has an isAccessible method which was introduced as part of ref
counting optimizations but there are some places still doing
accessibility checks by accessing the volatile refCnt() directly.
Modifications
- Have PooledNonRetained(Duplicate|Sliced)ByteBuf#isAccessible() use
their refcount delegate's isAccessible() method
- Add static isAccessible(buf) and ensureAccessible(buf) methods to
ByteBufUtil
(since ByteBuf#isAccessible() is package-private)
- Adjust DefaultByteBufHolder and similar classes to use these methods
rather than access refCnt() directly
Result
- More efficient accessibility checks in more places
Motivation:
We shouldn't call incSmallAllocation() in a synchronized block as its backed by a concurrent datastructure
Modifications:
Move call of incSmallAllocation() out of synchronized block
Result:
Minimize scope of synchronized block
Motivation:
For size from 512 bytes to chunkSize, we use a buddy algorithm. The
drawback is that it has a large internal fragmentation.
Modifications:
1. add SizeClassesMetric and SizeClasses
2. remove tiny size, now we have small, normal and huge size
3. rewrite the structure of PoolChunk
4. rewrite pooled allocate algorithm in PoolChunk
5. when allocate subpage, using lowest common multiple of pageSize and
elemSize instead of pageSize.
6. add more tests in PooledByteBufAllocatorTest and PoolArenaTest
Result:
Reduce internal fragmentation and closes#3910
Motivation:
We should include as much details as possible when throwing an IllegalArgumentException because of overflow in CompositeByteBuf
Modifications:
Add more details and factor out check into a static method to share code
Result:
Make it more clear why an operations failed
Motivation
An NPE was reported in #10245, caused by a regression introduced in
#8939. This in particular affects ByteToMessageDecoders that use the
COMPOSITE_CUMULATOR.
Modification
- Unwrap WrappedCompositeByteBufs passed to
CompositeByteBuf#addFlattenedComponents(...) method before accessing
internal components field
- Extend unit test to cover this case and ensure more of the
CompositeByteBuf tests are also run on the wrapped variant
Results
Fixes#10245
Motivation:
We can make use of our optimized implementations for UTF-8 and US-ASCII if the user request a copy of a sequence for these charsets
Modifications:
- Add fastpath implementation
- Add unit tests
Result:
Fixes https://github.com/netty/netty/issues/10205
Motivation:
We need to ensure we not overflow when adding buffers to a CompositeByteBuf
Modifications:
- Correctly validate overflow before adding to the internal storage
- Add testcase
Result:
Fixes https://github.com/netty/netty/issues/10194
Motivation:
We have found out that ByteBufUtil.indexOf can be inefficient for substring search on
ByteBuf, both in terms of algorithm complexity (worst case O(needle.readableBytes *
haystack.readableBytes)), and in constant factor (esp. on Composite buffers).
With implementation of more performant search algorithms we have seen improvements on
the order of magnitude.
Modifications:
This change introduces three search algorithms:
1. Knuth Morris Pratt - classical textbook algorithm, a good default choice.
2. Bit mask based algorithm - stable performance on any input, but limited to maximum
search substring (the needle) length of 64 bytes.
3. Aho–Corasick - worse performance and higher memory consumption than [1] and [2], but
it supports multiple substring (the needles) search simultaneously, by inspecting every
byte of the haystack only once.
Each algorithm processes every byte of underlying buffer only once, they are implemented
as ByteProcessor.
Result:
Efficient search algorithms with linear time complexity available in Netty (I will share
benchmark results in a comment on a PR).
Motivation:
PoolChunk requires a link to a PoolThreadCache to init ByteBuf. Currently the link is retrieved from a thread local: arena.parent.threadCache().
It has some performance cost. At the beginning of the allocation call the PoolThreadCache is already retrieved from the thread local. The reference can be propagated through the calls and used.
Modifications:
Replace second lookup of PoolThreadCache during ByteBuf init by propagation of a reference to PoolThreadCache down in the allocation stack explicitly
Result:
Improve performance of ByteBuf allocation
--Before--
Benchmark (size) (tokens) (useThreadCache) Mode Cnt Score Error Units
SimpleByteBufPooledAllocatorBenchmark.getAndRelease 123 0 true avgt 20 57.112 ± 1.004 ns/op
SimpleByteBufPooledAllocatorBenchmark.getAndRelease 123 100 true avgt 20 222.827 ± 1.307 ns/op
--After--
Benchmark (size) (tokens) (useThreadCache) Mode Cnt Score Error Units
SimpleByteBufPooledAllocatorBenchmark.getAndRelease 123 0 true avgt 20 50.732 ± 1.321 ns/op
SimpleByteBufPooledAllocatorBenchmark.getAndRelease 123 100 true avgt 20 216.892 ± 3.806 ns/op
Motivation:
PoolChunk.usage() method has non-trivial computations. It is used currently in hot path methods invoked when an allocation and de-allocation are happened.
The idea is to replace usage() output comparison against percent thresholds by Chunk.freeBytes plain comparison against absolute thresholds. In such way the majority of computations from the threshold conditions are moved to init logic.
Modifications:
Replace PoolChunk.usage() conditions in PoolChunkList with equivalent conditions for PoolChunk.freeBytes()
Result:
Improve performance of allocation and de-allocation of ByteBuf from normal size cache pool
Motivation:
The current implementation of log2 in PoolThreadCache uses a loop and less efficient than an version based on Integer.numberOfLeadingZeros (intrinsic).
Modifications:
Replace the current log2 implementation in PoolThreadCache with a version based on Integer.numberOfLeadingZeros
Result:
It can improve performance slightly during allocation and de-allocation of ByteBuf using pooled allocator.
Motivation:
FixedCompositeByteBuf.isDirect() should return the same value as EMPTY_BUFFER.isDirect() when constructed via an empty array
Modifications:
- Return correct value when constructed via empty array.
- Add unit test
Result:
FixedCompositeByteBuf.isDirect() returns correct value
Motivation:
We had a typo in the system property name that was used to lookup the cache trime interval. We should ensure we use the correct naming when lookup the property
Modifications:
- Support the old and the new (correct) naming of the property when configure the cache trim interval.
- Log something if someone uses the old (deprecated) name
Result:
Fixes https://github.com/netty/netty/issues/9981
Motivation
Recent optimization #9765 introduced a bug where the native indices of
the internal reused duplicate nio buffer are not properly reset prior to
using it to copy data during a reallocation operation. This can result
in BufferOverflowExceptions thrown during ByteBuf capacity changes.
The code path in question applies only to pooled direct buffers when
Unsafe is disabled or not available.
Modification
Ensure ByteBuffer#clear() is always called on the reused internal nio
buffer prior to returning it from PooledByteBuf#internalNioBuffer()
(protected method); add unit test that exposes the bug.
Result
Fixes#9911
# Motivation:
`DefaultByteBufHolder.equals()` considers another object equal if it's an instance of `ByteBufferHolder` and if the contents of two objects are equal. However, the behavior of `equals` method is not a part of the `ByteBufHolder` contract so `DefaultByteBufHolder`'s version may be causing violation of the symmetric property if other classes have different logic.
There are already a few classes that are affected by this: `DefaultHttp2GoAwayFrame`, `DefaultHttp2UnknownFrame`, and `SctpMessage` are all overriding `equals` method breaking the symmetric property.
Another effect of this behavior is that all instances with empty data are considered equal. That may not be desireable in the situations when instances are created for predefined constants, e.g. `FullBulkStringRedisMessage.NULL_INSTANCE` and `FullBulkStringRedisMessage.EMPTY_INSTANCE` in `codec-redis`.
# Modification:
Make `DefaultByteBufHolder.equals()` implementation only work for the objects of the same class.
# Result:
- The symmetric property of the `equals` method is restored for the classes in question.
- Instances of different classes are not considered equal even if the content of the data they hold are the same.
Motivation
While working on other changes I noticed some opportunities to
streamline a few things in AbstractByteBuf.
Modifications
- Avoid duplicate ensureAccessible() checks in discard(Some)ReadBytes()
and ensureWritable0(int) methods
- Simplify ensureWritable0(int) logic
- Make some conditional checks more concise
Result
Cleaner, possibly faster code
Motivation:
Currently when use of Unsafe is disabled and an internal reallocation is
performed for a direct PooledByteBuf, a one-off temporary duplicate is
made of the source and destination backing nio buffers so that the
copy can be done in a threadsafe manner.
The need for this can be reduced by sharing the temporary duplicate
buffer that is already stored in the corresponding destination
PooledByteBuf instance.
Modifications:
Have PoolArena#memoryCopy(...) take the destination PooledByteBuf
instead of the underlying mem reference and offset, and use
internalNioBuffer() to obtain/initialize a reusable duplicate of the
backing nio buffer.
Result:
Fewer temporary allocations when resizing direct pooled ByteBufs in the
non-Unsafe case
Motivation
There's currently no way to determine whether an arbitrary ByteBuf
behaves internally like a "singluar" buffer or a composite one, and this
can be important to know when making decisions about how to manipulate
it in an efficient way.
An example of this is the ByteBuf#discardReadBytes() method which
increases the writable bytes for a contiguous buffer (by readerIndex)
but does not for a composite one.
Unfortunately !(buf instanceof CompositeByteBuf) is not reliable, since
for example this will be true in the case of a sliced CompositeByteBuf
or some third-party composite implementation.
isContiguous was chosen over isComposite since we want to assume "not
contiguous" in the unknown/default case - the doc will it clear that
false does not imply composite.
Modifications
- Add ByteBuf#isContiguous() which returns true by default
- Override the "concrete" ByteBuf impls to return true and ensure
wrapped/derived impls delegate it appropriately
- Include some basic unit tests
Result
Better assumptions/decisions possible when manipulating arbitrary
ByteBufs, for example when combining/cumulating them.
Motivation:
At the moment we directly extend the Recycler base class in our code which makes it hard to experiment with different Object pool implementation. It would be nice to be able to switch from one to another by using a system property in the future. This would also allow to more easily test things like https://github.com/netty/netty/pull/8052.
Modifications:
- Introduce ObjectPool class with static method that we now use internally to obtain an ObjectPool implementation.
- Wrap the Recycler into an ObjectPool and return it for now
Result:
Preparation for different ObjectPool implementations
Motivation
Currently doc != code and so one needs to change. Though behaviour as
currently documented might be more intuitive, we don't want to break
anyone so will adjust the doc instead. See #9503 for discussion.
Modifications
Correct the javadoc of indexOf(...) method in ByteBuf abstract class.
Results
Correct javadoc
