Motivation:
Often a temporary ByteBuffer is used which can be cached to reduce the GC pressure.
Modifications:
Cache the ByteBuffer in the PoolThreadCache as well.
Result:
Less GC.
Motivation
#8563 highlighted race conditions introduced by the prior optimistic
update optimization in 83a19d5650. These
were known at the time but considered acceptable given the perf
benefit in high contention scenarios.
This PR proposes a modified approach which provides roughly half the
gains but stronger concurrency semantics. Race conditions still exist
but their scope is narrowed to much less likely cases (releases
coinciding with retain overflow), and even in those
cases certain guarantees are still assured. Once release() returns true,
all subsequent release/retains are guaranteed to throw, and in
particular deallocate will be called at most once.
Modifications
- Use even numbers internally (including -ve) for live refcounts
- "Final" release changes to odd number (equivalent to refcount 0)
- Retain still uses faster getAndAdd, release uses CAS loop
- First CAS attempt uses non-volatile read
- Thread.yield() after a failed CAS provides a net gain
Result
More (though not completely) robust concurrency semantics for ref
counting; increased latency under high contention, but still roughly
twice as fast as the original logic. Bench results to follow
Motivation:
ByteBuf is used everywhere so we should try hard to be able to make things inlinable. During benchmarks it showed that writeCharSequence(...) fails to inline writeUtf8 because it is too big even if its hots.
Modifications:
Move less common code-path to extra method to allow inlining.
Result:
Be able to inline writeUtf8 in most cases.
Motivation:
Often a temporary ByteBuffer is used which can be cached to reduce the GC pressure.
Modifications:
Add a Deque per PoolChunk which will be used for caching.
Result:
Less GC.
Motivation:
When we create new chunk with memory aligned, the offset of direct memory should be
'alignment - address & (alignment - 1)', not just 'address & (alignment - 1)'.
Modification:
Change offset calculating formula to offset = alignment - address & (alignment - 1) in PoolArena.DirectArena#offsetCacheLine and add a unit test to assert that.
Result:
Correctly calculate offset.
Motivation:
ByteBuf.retainedSlice() and similar methods produce sliced buffers with
an independent refcount to the buffer that they wrap.
One of the optimizations in 10539f4dc7 was
to use the ref to the unwrapped buffer object for added slices, but this
did not take into account the above special case when later releasing.
Thanks to @rkapsi for discovering this via #8495.
Modifications:
Since a reference to the slice is still kept in the Component class,
just changed Component.freeIfNecessary() to release the slice in
preference to the unwrapped buf.
Also added a unit test which reproduces the bug.
Result:
Fixes#8495
Motivation:
Two similar bugs were introduced by myself in separate recent PRs #8393
and #8464, while optimizing the assignment/handling of temporary arrays
in ByteBufUtil and UnsafeByteBufUtil.
The temp arrays allocated for buffering data written to an OutputStream
are incorrectly sized to the full length of the data to copy rather than
being capped at WRITE_CHUNK_SIZE.
Unfortunately one of these is in the 4.1.31.Final release, I'm really
sorry and will be more careful in future.
This kind of thing is tricky to cover in unit tests.
Modifications:
Revert the temp array allocations back to their original sizes.
Avoid making duplicate calls to ByteBuf.capacity() in a couple of places
in ByteBufUtil (unrelated thing I noticed, can remove it from this PR if
desired!)
Result:
Temporary byte arrays will be reverted to their originally intended
sizes.
Motivation:
#8388 introduced a reusable ThreadLocal<byte[]> for use in
decodeString(...). It can be used in more places in the buffer package
to avoid temporary allocations of small arrays.
Modifications:
Encapsulate use of the ThreadLocal in a static package-private
ByteBufUtil.threadLocalTempArray(int) method, and make use of it from a
handful of new places including ByteBufUtil.readBytes(...).
Result:
Fewer short-lived small byte array allocations.
Motivation:
CompositeByteBuf is a powerful and versatile abstraction, allowing for
manipulation of large data without copying bytes. There is still a
non-negligible cost to reading/writing however relative to "singular"
ByteBufs, and this can be mostly eliminated with some rework of the
internals.
My use case is message modification/transformation while zero-copy
proxying. For example replacing a string within a large message with one
of a different length
Modifications:
- No longer slice added buffers and unwrap added slices
- Components store target buf offset relative to position in
composite buf
- Less allocations, object footprint, pointer indirection, offset
arithmetic
- Use Component[] rather than ArrayList<Component>
- Avoid pointer indirection and duplicate bounds check, more
efficient backing array growth
- Facilitates optimization when doing bulk-inserts - inserting n
ByteBufs behind m is now O(m + n) instead of O(mn)
- Avoid unnecessary casting and method call indirection via superclass
- Eliminate some duplicate range/ref checks via non-checking versions of
toComponentIndex and findComponent
- Add simple fast-path for toComponentIndex(0); add racy cache of
last-accessed Component to findComponent(int)
- Override forEachByte0(...) and forEachByteDesc0(...) methods
- Make use of RecyclableArrayList in nioBuffers(int, int) (in line with
FasterCompositeByteBuf impl)
- Modify addComponents0(boolean,int,Iterable) to use the Iterable
directly rather than copy to an array first (and possibly to an
ArrayList before that)
- Optimize addComponents0(boolean,int,ByteBuf[],int) to not perform
repeated array insertions and avoid second loop for offset updates
- Simplify other logic in various places, in particular the general
pattern used where a sub-range is iterated over
- Add benchmarks to demonstrate some improvements
While refactoring I also came across a couple of clear bugs. They are
fixed in these changes but I will open another PR with unit tests and
fixes to the current version.
Result:
Much faster creation, manipulation, and access; many fewer allocations
and smaller footprint. Benchmark results to follow.
Motivation:
Unpooled.wrap(byte[]...) and Unpooled.wrap(ByteBuffer...) currently
allocate/copy an intermediate ByteBuf ArrayList and array, which can be
avoided.
Modifications:
- Define new internal ByteWrapper interface and add a CompositeByteBuf
constructor which takes a ByteWrapper with an array of the type that it
wraps, and modify the appropriate Unpooled.wrap(...) methods to take
advantage of it
- Tidy up other constructors in CompositeByteBuf to remove duplication
and misleading len arg (which is really an end offset into provided
array)
Result:
Less allocation/copying when wrapping byte[] and ByteBuffer arrays,
tidier code.
Motivation:
I came across two bugs:
- Components removed due to capacity reduction aren't released
- Offsets aren't set correctly on empty components that are added
between existing components
Modifications:
Add unit tests which expose these bugs, fix them.
Result:
Bugs are fixed
Motivation:
There are currently many more places where this could be used which were
possibly not considered when the method was added.
If https://github.com/netty/netty/pull/8388 is included in its current
form, a number of these places could additionally make use of the same
BYTE_ARRAYS threadlocal.
There's also a couple of adjacent places where an optimistically-pooled
heap buffer is used for temp byte storage which could use the
threadlocal too in preference to allocating a temp heap bytebuf wrapper.
For example
https://github.com/netty/netty/blob/4.1/buffer/src/main/java/io/netty/buffer/ByteBufUtil.java#L1417.
Modifications:
Replace new byte[] with PlatformDependent.allocateUninitializedArray()
where appropriate; make use of ByteBufUtil.getBytes() in some places
which currently perform the equivalent logic, including avoiding copy of
backing array if possible (although would be rare).
Result:
Further potential speed-up with java9+ and appropriate compile flags.
Many of these places could be on latency-sensitive code paths.
* Optimize AbstractByteBuf.getCharSequence() in US_ASCII case
Motivation:
Inspired by https://github.com/netty/netty/pull/8388, I noticed this
simple optimization to avoid char[] allocation (also suggested in a TODO
here).
Modifications:
Return an AsciiString from AbstractByteBuf.getCharSequence() if
requested charset is US_ASCII or ISO_8859_1 (latter thanks to
@Scottmitch's suggestion). Also tweak unit tests not to require Strings
and include a new benchmark to demonstrate the speedup.
Result:
Speed-up of AbstractByteBuf.getCharSequence() in ascii and iso 8859/1
cases
Motivation:
CompositeByteBuf.decompose(...) did not correctly slice the content and so produced an incorrect representation of the data.
Modifications:
- Rewrote implementation to fix bug and also improved it to reduce GC
- Add unit tests.
Result:
Fixes https://github.com/netty/netty/issues/8400.
Motivation:
While looking at the nice optimization done in
https://github.com/netty/netty/pull/8347 I couldn't help noticing the
logic could be simplified further. Apologies if this is just my OCD and
inappropriate!
Modifications:
Reduce amount of code used for ByteBufInputStream.readLine()
Result:
Slightly smaller and simpler code
Motivation:
Avoid creating any StringBuilder instance if
ByteBufInputStream::readLine isn't used
Modifications:
The StringBuilder instance is lazy allocated on demand and
are added new test case branches to address the increased
complexity of ByteBufInputStream::readLine
Result:
Reduced GC activity if ByteBufInputStream::readLine isn't used
Motivation:
We should just directly init the refCnt to 1 and not use the AtomicIntegerFieldUpdater.
Modifications:
Just assing directly to 1.
Result:
Cleaner code and possible a bit faster as the JVM / JIT may be able to optimize the first store easily.
Motiviation:
At the moment whenever ensureAccessible() is called in our ByteBuf implementations (which is basically on each operation) we will do a volatile read. That per-se is not such a bad thing but the problem here is that it will also reduce the the optimizations that the compiler / jit can do. For example as these are volatile it can not eliminate multiple loads of it when inline the methods of ByteBuf which happens quite frequently because most of them a quite small and very hot. That is especially true for all the methods that act on primitives.
It gets even worse as people often call a lot of these after each other in the same method or even use method chaining here.
The idea of the change is basically just ue a non-volatile read for the ensureAccessible() check as its a best-effort implementation to detect acting on already released buffers anyway as even with a volatile read it could happen that the user will release it in another thread before we actual access the buffer after the reference check.
Modifications:
- Try to do a non-volatile read using sun.misc.Unsafe if we can use it.
- Add a benchmark
Result:
Big performance win when multiple ByteBuf methods are called from a method.
With the change:
UnsafeByteBufBenchmark.setGetLongUnsafeByteBuf thrpt 20 281395842,128 ± 5050792,296 ops/s
Before the change:
UnsafeByteBufBenchmark.setGetLongUnsafeByteBuf thrpt 20 217419832,801 ± 5080579,030 ops/s
Motivation:
The JVM isn't always able to hoist out/reduce bounds checking (due to ref counting operations etc etc) hence making it configurable could improve performances for most CPU intensive use cases.
Modifications:
Each AbstractByteBuf bounds check has been tested against a new static final configuration property similar to checkAccessible ie io.netty.buffer.bytebuf.checkBounds.
Result:
Any user could disable ByteBuf bounds checking in order to get extra performances.
* Allow to use native transports when sun.misc.Unsafe is not present on the system
Motivation:
We should be able to use the native transports (epoll / kqueue) even when sun.misc.Unsafe is not present on the system. This is especially important as Java11 will be released soon and does not allow access to it by default.
Modifications:
- Correctly disable usage of sun.misc.Unsafe when -PnoUnsafe is used while running the build
- Correctly increment metric when UnpooledDirectByteBuf is allocated. This was uncovered once -PnoUnsafe usage was fixed.
- Implement fallbacks in all our native transport code for when sun.misc.Unsafe is not present.
Result:
Fixes https://github.com/netty/netty/issues/8229.
In nioBuffer(int,int) in CompositeByteBuf , we create a sub-array of nioBuffers for the components that are in range, then concatenate all the components in range into a single bigger buffer.
However, if the call to nioBuffers() returned only one sub-buffer, then we are copying it to a newly-allocated buffer "merged" for no reason.
Motivation:
Profiler for Spark shows a lot of time spent in put() method inside nioBuffer(), while usually no copy of data is required.
Modification:
This change skips this last step and just returns a duplicate of the single buffer returned by the call to nioBuffers(), which will in most implementation not copy the data
Result:
No copy when the source is only 1 buffer
Motivation:
We need to add special handling for WrappedCompositeByteBuf as these also extend AbstractByteBuf, otherwise we will not correctly adjust / read the writerIndex during processing.
Modifications:
- Add instanceof checks for WrappedCompositeByteBuf as well.
- Add testcases
Result:
Fixes https://github.com/netty/netty/issues/8152.
Motivation:
5b1fe611a6 introduced the usage of a finalizer as last resort for PoolThreadCache. As we may call free() from the FastThreadLocal.onRemoval(...) and finalize() we need to guard against multiple calls as otherwise we will corrupt internal state (that is used for metrics).
Modifications:
Use AtomicBoolean to guard against multiple calls of PoolThreadCache.free().
Result:
No more corruption of internal state caused by calling PoolThreadCache.free() multuple times.
Motivation:
Recent PR https://github.com/netty/netty/pull/8040 introduced
Unpooled.wrappedUnmodifiableBuffer(ByteBuf...) which has the same
behaviour but wraps the provided array directly. This is preferred for
most uses (including varargs-based use) and if there are any unusual
cases of an explicit array which is re-used before the ByteBuf is
finished with, it can just be copied first.
Modifications:
Added @Deprecated annotation and javadoc to
Unpooled.unmodifiableBuffer(ByteBuf...).
Result:
Unpooled.unmodifiableBuffer(ByteBuf...) will be deprecated.
Motivation:
ObjectCleaner does start a Thread to handle the cleaning of resources which leaks into the users application. We should not use it in netty itself to make things more predictable.
Modifications:
- Remove usage of ObjectCleaner and use finalize as a replacement when possible.
- Clarify javadocs for FastThreadLocal.onRemoval(...) to ensure its clear that remove() is not guaranteed to be called when the Thread completees and so this method is not enough to guarantee cleanup for this case.
Result:
Fixes https://github.com/netty/netty/issues/8017.
Motivation:
Unpooled.unmodifiableBuffer() is currently used to efficiently write
arrays of ByteBufs via FixedCompositeByteBuf, but involves an allocation
and content-copy of the provided ByteBuf array which in many (most?)
cases shouldn't be necessary.
Modifications:
Modify the internal FixedCompositeByteBuf class to support wrapping the
provided ByteBuf array directly. Control this behaviour with a
constructor flag and expose the "unsafe" version via a new
Unpooled.wrappedUnmodifiableBuffer(ByteBuf...) method.
Result:
Less garbage on IO paths. I would guess pretty much all existing usage
of unmodifiableBuffer() could use the copy-free version but assume it's
not safe to change its default behaviour.
Motivation:
Eliminate avoidable backing array reallocations when constructing
composite ByteBufs from existing buffer arrays/Iterables. This also
applies to the Unpooled.wrappedBuffer(...) methods.
Modifications:
Ensure the initial components ComponentList is sized at least as large
as the provided buffer array/Iterable in the CompositeByteBuffer
constructors.
In single-arg Unpooled.wrappedBuffer(...) methods, set maxNumComponents
to the count of provided buffers, rather than a fixed default of 16. It
seems likely that most usage of these involves wrapping a list without
subsequent modification, particularly since they return a ByteBuf rather
than CompositeByteBuf. If a different/larger max is required there are
already the wrappedBuffer(int, ...) variants.
In fact the current behaviour could be considered inconsistent - if you
call Unpooled.wrappedBuffer(int, ByteBuf) with a single buffer, you
might expect to subsequently be able to add buffers to it (since you
specified a max related to consolidation), but it will in fact return
just a slice of the provided ByteBuf.
Result:
Fewer and smaller allocations in some cases when using CompositeByteBufs
or Unpooled.wrappedBuffer(...).
Motivation:
Currently there is not a clear way to provide a byte array to a netty
ByteBuf and be informed when it is released. This is a would be a
valuable addition for projects that integrate with netty but also pool
their own byte arrays.
Modification:
Modified the UnpooledHeapByteBuf class so that the freeArray method is
protected visibility instead of default. This will allow a user to
subclass the UnpooledHeapByteBuf, provide a byte array, and override
freeArray to return the byte array to a pool when it is called.
Additionally this makes this implementation equivalent to
UnpooledDirectByteBuf (freeDirect is protected).
Additionally allocateArray is also made protect to provide another override
option for subclasses.
Result:
Users can override UnpooledHeapByteBuf#freeArray and
UnpooledHeapByteBuf#allocateArray.
Motivation:
When I read the source code, I found that the comment of PoolChunk is out of date, it may confuses readers with the description about memoryMap.
Modifications:
update the last passage of the comment of the PoolChunk class.
Result:
No change to any source code , just update comment.
Motivation:
When a buffer is over-released, the current error message of `IllegalReferenceCountException` is `refCnt: XXX, increment: XXX`, which is confusing. The correct message should be `refCnt: XXX, decrement: XXX`.
Modifications:
Pass `-decrement` to create `IllegalReferenceCountException`.
Result:
The error message will be `refCnt: XXX, decrement: XXX` when a buffer is over-released.
Motivation:
It should be possible to write a ReadOnlyByteBufferBuf to a channel without errors. However, ReadOnlyByteBufferBuf does not override isWritable and ensureWritable, which can cause some handlers to mistakenly assume they can write to the ReadOnlyByteBufferBuf, resulting in ReadOnlyBufferException.
Modification:
Added isWritable and ensureWritable method overrides on ReadOnlyByteBufferBuf to indicate that it is never writable. Added tests for these methods.
Result:
Can successfully write ReadOnlyByteBufferBuf to a channel with an SslHandler (or any other handler which may attempt to write to the ByteBuf it receives).
Motivation:
The `AbstractByteBuf#equals` method doesn't take into account the
class of buffer instance. So the two buffers with different classes
must have the same `hashCode` values if `equals` method returns `true`.
But `EmptyByteBuf#hashCode` is not consistent with `#hashCode`
of the empty `AbstractByteBuf`, that is violates the contract and
can lead to errors.
Modifications:
Return `1` in `EmptyByteBuf#hashCode`.
Result:
Consistent behavior of `EmptyByteBuf#hashCode` and `AbstractByteBuf#hashCode`.
Motivation:
The `ByteBuf#slice` and `ByteBuf#duplicate` methods should check
an accessibility to prevent creation slice or duplicate
of released buffer. At now this works not in the all scenarios.
Modifications:
Add missed checks.
Result:
More correct and consistent behavior of `ByteBuf` methods.
Motivation:
The `#ensureAccessible` method in `UnpooledHeapByteBuf#capacity` used
to prevent NPE if buffer is released and `array` is `null`. In all
other implementations of `ByteBuf` the accessible is not checked by
`capacity` method. We can assign an empty array to `array`
in the `deallocate` and don't worry about NPE in the `#capacity`.
This will help reduce the number of repeated calls of the
`#ensureAccessible` in many operations with `UnpooledHeapByteBuf`.
Modifications:
1. Remove `#ensureAccessible` call from `UnpooledHeapByteBuf#capacity`.
Use the `EmptyArrays#EMPTY_BYTES` instead of `null` in `#deallocate`.
2. Fix access checks in `AbstractUnsafeSwappedByteBuf` and
`AbstractByteBuf#slice` that relied on `#ensureAccessible`
in `UnpooledHeapByteBuf#capacity`. This was found by unit tests.
Result:
Less double calls of `#ensureAccessible` for `UnpooledHeapByteBuf`.
Motivation:
Currently copying a direct ByteBuf copies it fully into the heap before writing it to an output stream.
The can result in huge memory usage on the heap.
Modification:
copy the bytebuf contents via an 8k buffer into the output stream
Result:
Fixes#7804
Motivation:
We should allow to access the memoryAddress / array of the FixedCompositeByteBuf when it only wraps a single ByteBuf. We do the same for CompositeByteBuf.
Modifications:
- Check how many buffers FixedCompositeByteBuf wraps and depending on it delegate the access to the memoryAddress / array
- Add unit tests.
Result:
Fixes [#7752].
Motivation:
If someone invoke writeByte(), markWriterIndex(), readByte() in order first, and then invoke resetWriterIndex() should be throw a IndexOutOfBoundsException to obey the rule that the buffer declared "0 <= readerIndex <= writerIndex <= capacity".
Modification:
Changed the code writerIndex = markedWriterIndex; into writerIndex(markedWriterIndex); to make the check affect
Result:
Throw IndexOutOfBoundsException if any invalid happened in resetWriterIndex.
Motivation:
Read-only heap ByteBuffer doesn't expose array: the existent method to perform copies to direct ByteBuf involves the creation of a (maybe pooled) additional heap ByteBuf instance and copy
Modifications:
To avoid stressing the allocator with additional (and stealth) heap ByteBuf allocations is provided a method to perform copies using the (pooled) internal NIO buffer
Result:
Copies from read-only heap ByteBuffer to direct ByteBuf won't create any intermediate ByteBuf
Motivation:
To avoid eager allocation of the destination and to perform length prefixed encoding of UTF-8 string with forward only access pattern
Modifications:
The original writeUtf8 is modified by allowing customization of the reserved bytes on the destination buffer and is introduced an exact UTF-8 length estimator.
Result:
Is now possible to perform length first encoding with UTF-8 well-formed char sequences following a forward only write access pattern on the destination buffer.
Motivation:
ByteBufUtil by default will cache DirectByteBuffer objects, and the
associated direct memory (up to 64k). In combination with the Recycler which may
cache up to 32k elements per thread may lead to a large amount of direct
memory being retained per EventLoop thread. As traffic spikes come this
may be perceived as a memory leak because the memory in the Recycler
will never be reclaimed.
Modifications:
- By default we shouldn't cache DirectByteBuffer objects.
Result:
Less direct memory consumption due to caching DirectByteBuffer objects.
Motivation:
There is some cleanup that can be done.
Modifications:
- Use intializer list expression where possible
- Remove unused imports.
Result:
Cleaner code.
Motivation:
We need the memoryAddress of a direct buffer when using our native transports. For this reason ReadOnlyUnsafeDirectByteBuf.memoryAddress() should not throw.
Modifications:
- Correctly override ReadOnlyUnsafeDirectByteBuf.memoryAddress() and hasMemoryAddress()
- Add test case
Result:
Fixes [#7672].
Motivation:
We saw some timeouts on the CI when the leak detection is enabled.
Modifications:
- Use smaller number of operations in test
- Increase timeout
Result:
CI not times out.
Motivation:
ByteBufUtil.isText(...) may produce unexpected results if called concurrently on the same ByteBuffer.
Modifications:
- Don't use internalNioBuffer where it is not safe.
- Add unit test.
Result:
ByteBufUtil.isText is thread-safe.
Motivation:
Usages of HttpResponseStatus may result in more object allocation then necessary due to not looking for cached objects and the AsciiString parsing method not being used due to CharSequence method being used instead.
Modifications:
- HttpResponseDecoder should attempt to get the HttpResponseStatus from cache instead of allocating a new object
- HttpResponseStatus#parseLine(CharSequence) should check if the type is AsciiString and redirect to the AsciiString parsing method which may not require an additional toString call
- HttpResponseStatus#parseLine(AsciiString) can be optimized and doesn't require and may not require object allocation
Result:
Less allocations when dealing with HttpResponseStatus.
Motivation:
Depending on the implementation of ByteBuf nioBuffer(...) and nioBuffers(...) may either share the content or return a ByteBuffer that contains a copy of the content.
Modifications:
Fix javadocs.
Result:
Correct docs.
Motivation:
Calling ByteBuf.toString(Charset) on the same buffer from multiple threads at the same time produces unexpected results, such as various exceptions and/or corrupted output. This is because ByteBufUtil.decodeString(...) is taking the source ByteBuffer for CharsetDecoder.decode() from ByteBuf.internalNioBuffer(int, int), which is not thread-safe.
Modification:
Call ByteBuf.nioBuffer() instead of ByteBuf.internalNioBuffer() to get the source buffer to pass to CharsetDecoder.decode().
Result:
Fixes the possible race condition.
Motivation:
We did not correctly take the position into account when wrapping a ByteBuffer via ReadOnlyUnsafeDirectByteBuf as we obtained the memory address from the original ByteBuffer and not the slice we take.
Modifications:
- Correctly use the slice to obtain memory address.
- Add test case.
Result:
Fixes [#7565].
Motivation:
There is no guarantee that FastThreadLocal.onRemoval(...) is called if the FastThreadLocal is used by "non" FastThreacLocalThreads. This can lead to all sort of problems, like for example memory leaks as direct memory is not correctly cleaned up etc.
Beside this we use ThreadDeathWatcher to check if we need to release buffers back to the pool when thread local caches are collected. In the past ThreadDeathWatcher was used which will need to "wakeup" every second to check if the registered Threads are still alive. If we can ensure FastThreadLocal.onRemoval(...) is called we do not need this anymore.
Modifications:
- Introduce ObjectCleaner and use it to ensure FastThreadLocal.onRemoval(...) is always called when a Thread is collected.
- Deprecate ThreadDeathWatcher
- Add unit tests.
Result:
Consistent way of cleanup FastThreadLocals when a Thread is collected.
Motivation:
We used subList in CompositeByteBuf to remove ranges of elements from the internal storage. Beside this we also used an foreach loop in a few cases which will crate an Iterator.
Modifications:
- Use our own sub-class of ArrayList which exposes removeRange(...). This allows to remove a range of elements without an extra allocation.
- Use an old style for loop to iterate over the elements to reduce object allocations.
Result:
Less allocations.
Automatic-Module-Name entry provides a stable JDK9 module name, when Netty is used in a modular JDK9 applications. More info: http://blog.joda.org/2017/05/java-se-9-jpms-automatic-modules.html
When Netty migrates to JDK9 in the future, the entry can be replaced by actual module-info descriptor.
Modification:
The POM-s are configured to put the correct module names to the manifest.
Result:
Fixes#7218.
Motivation:
We dont need to use the ThreadDeathWatcher if we use a FastThreadLocalThread for which we wrap the Runnable and ensure we call FastThreadLocal.removeAll() once the Runnable completes.
Modifications:
- Dont use a ThreadDeathWatcher if we are sure we will call FastThreadLocal.removeAll()
- Add unit test.
Result:
Less overhead / running theads if you only allocate / deallocate from FastThreadLocalThreads.
Motivation:
AbstractByteBuf#readSlice relied upon the bounds checking of the slice operation in order to detect index out of bounds conditions. However the slice bounds checking operation allows for the slice to go beyond the writer index, and this is out of bounds for a read operation.
Modifications:
- AbstractByteBuf#readSlice and AbstractByteBuf#readRetainedSlice should ensure the desired amount of bytes are readable before taking a slice
Result:
No reading of undefined data in AbstractByteBuf#readSlice and AbstractByteBuf#readRetainedSlice.
Motivation:
When calling CompositeBytebuf.copy() and copy(...) we currently use Unpooled to allocate the buffer. This is not really correct and may produce more GC then needed. We should use the allocator that was used when creating the CompositeByteBuf to allocate the new buffer which may be for example the PooledByteBufAllocator.
Modifications:
- Use alloc() to allocate the new buffer.
- Add tests
- Fix tests that depend on the copy to be backed by an byte-array without checking hasArray() first.
Result:
Fixes [#7393].
Motivation:
Even if it's a super micro-optimization (most JVM could optimize such
cases in runtime), in theory (and according to some perf tests) it
may help a bit. It also makes a code more clear and allows you to
access such methods in the test scope directly, without instance of
the class.
Modifications:
Add 'static' modifier for all methods, where it possible. Mostly in
test scope.
Result:
Cleaner code with proper 'static' modifiers.
Motivation:
Javadoc of the `ByteBufUtil#copy(AsciiString, int, ByteBuf, int, int)` is incorrect.
Modifications:
Fix it.
Result:
The description of the `#copy` method is not misleading.
Motivation:
In the `ByteBufOutputStream` we can use an appropriate methods of `ByteBuf`
to reduce calls of virtual methods and do not copying converting logic.
Modifications:
- Use an appropriate methods of `ByteBuf`
- Remove redundant conversions (int -> byte, int -> char).
- Use `ByteBuf#writeCharSequence` in the `writeBytes(String)'.
Result:
Less code duplication. A `writeBytes(String)` method is faster.
No unnecessary conversions. More consistent and cleaner code.
Configuring this is tough because there is split between highly shared (and accessed) objects and lightly accessed objects.
Modification:
There are a number of changes here. In relative order of importance:
API / Functionality changes:
* Max records and max sample records are gone. Only "target" records, the number of records tries to retain is exposed.
* Records are sampled based on the number of already stored records. The likelihood of recording a new sample is `2^(-n)`, where `n` is the number of currently stored elements.
* Records are stored in a concurrent stack structure rather than a list. This avoids a head and tail. Since the stack is only read once, there is no need to maintain head and tail pointers
* The properties of this imply that the very first and very last access are always recorded. When deciding to sample, the top element is replaced rather than pushed.
* Samples that happen between the first and last accesses now have a chance of being recorded. Previously only the final few were kept.
* Sampling is no longer deterministic. Previously, a deterministic access pattern meant that you could conceivably always miss some access points.
* Sampling has a linear ramp for low values and and exponentially backs off roughly equal to 2^n. This means that for 1,000,000 accesses, about 20 will actually be kept. I have an elegant proof for this which is too large to fit in this commit message.
Code changes:
* All locks are gone. Because sampling rarely needs to do a write, there is almost 0 contention. The dropped records counter is slightly contentious, but this could be removed or changed to a LongAdder. This was not done because of memory concerns.
* Stack trace exclusion is done outside of RLD. Classes can opt to remove some of their methods.
* Stack trace exclusion is faster, since it uses String.equals, often getting a pointer compare due to interning. Previously it used contains()
* Leak printing is outputted fairly differently. I tried to preserve as much of the original formatting as possible, but some things didn't make sense to keep.
Result:
More useful leak reporting.
Faster:
```
Before:
Benchmark (recordTimes) Mode Cnt Score Error Units
ResourceLeakDetectorRecordBenchmark.record 8 thrpt 20 136293.404 ± 7669.454 ops/s
ResourceLeakDetectorRecordBenchmark.record 16 thrpt 20 72805.720 ± 3710.864 ops/s
ResourceLeakDetectorRecordBenchmark.recordWithHint 8 thrpt 20 139131.215 ± 4882.751 ops/s
ResourceLeakDetectorRecordBenchmark.recordWithHint 16 thrpt 20 74146.313 ± 4999.246 ops/s
After:
Benchmark (recordTimes) Mode Cnt Score Error Units
ResourceLeakDetectorRecordBenchmark.record 8 thrpt 20 155281.969 ± 5301.399 ops/s
ResourceLeakDetectorRecordBenchmark.record 16 thrpt 20 77866.239 ± 3821.054 ops/s
ResourceLeakDetectorRecordBenchmark.recordWithHint 8 thrpt 20 153360.036 ± 8611.353 ops/s
ResourceLeakDetectorRecordBenchmark.recordWithHint 16 thrpt 20 78670.804 ± 2399.149 ops/s
```
Motivation:
Highly retained and released objects have contention on their ref
count. Currently, the ref count is updated using compareAndSet
with care to make sure the count doesn't overflow, double free, or
revive the object.
Profiling has shown that a non trivial (~1%) of CPU time on gRPC
latency benchmarks is from the ref count updating.
Modification:
Rather than pessimistically assuming the ref count will be invalid,
optimistically update it assuming it will be. If the update was
wrong, then use the slow path to revert the change and throw an
execption. Most of the time, the ref counts are correct.
This changes from using compareAndSet to getAndAdd, which emits a
different CPU instruction on x86 (CMPXCHG to XADD). Because the
CPU knows it will modifiy the memory, it can avoid contention.
On a highly contended machine, this can be about 2x faster.
There is a downside to the new approach. The ref counters can
temporarily enter invalid states if over retained or over released.
The code does handle these overflow and underflow scenarios, but it
is possible that another concurrent access may push the failure to
a different location. For example:
Time 1 Thread 1: obj.retain(INT_MAX - 1)
Time 2 Thread 1: obj.retain(2)
Time 2 Thread 2: obj.retain(1)
Previously Thread 2 would always succeed and Thread 1 would always
fail on the second access. Now, thread 2 could fail while thread 1
is rolling back its change.
====
There are a few reasons why I think this is okay:
1. Buggy code is going to have bugs. An exception _is_ going to be
thrown. This just causes the other threads to notice the state
is messed up and stop early.
2. If high retention counts are a use case, then ref count should
be a long rather than an int.
3. The critical section is greatly reduced compared to the previous
version, so the likelihood of this happening is lower
4. On error, the code always rollsback the change atomically, so
there is no possibility of corruption.
Result:
Faster refcounting
```
BEFORE:
Benchmark (delay) Mode Cnt Score Error Units
AbstractReferenceCountedByteBufBenchmark.retainRelease_contended 1 sample 2901361 804.579 ± 1.835 ns/op
AbstractReferenceCountedByteBufBenchmark.retainRelease_contended 10 sample 3038729 785.376 ± 16.471 ns/op
AbstractReferenceCountedByteBufBenchmark.retainRelease_contended 100 sample 2899401 817.392 ± 6.668 ns/op
AbstractReferenceCountedByteBufBenchmark.retainRelease_contended 1000 sample 3650566 2077.700 ± 0.600 ns/op
AbstractReferenceCountedByteBufBenchmark.retainRelease_contended 10000 sample 3005467 19949.334 ± 4.243 ns/op
AbstractReferenceCountedByteBufBenchmark.retainRelease_uncontended 1 sample 456091 48.610 ± 1.162 ns/op
AbstractReferenceCountedByteBufBenchmark.retainRelease_uncontended 10 sample 732051 62.599 ± 0.815 ns/op
AbstractReferenceCountedByteBufBenchmark.retainRelease_uncontended 100 sample 778925 228.629 ± 1.205 ns/op
AbstractReferenceCountedByteBufBenchmark.retainRelease_uncontended 1000 sample 633682 2002.987 ± 2.856 ns/op
AbstractReferenceCountedByteBufBenchmark.retainRelease_uncontended 10000 sample 506442 19735.345 ± 12.312 ns/op
AFTER:
Benchmark (delay) Mode Cnt Score Error Units
AbstractReferenceCountedByteBufBenchmark.retainRelease_contended 1 sample 3761980 383.436 ± 1.315 ns/op
AbstractReferenceCountedByteBufBenchmark.retainRelease_contended 10 sample 3667304 474.429 ± 1.101 ns/op
AbstractReferenceCountedByteBufBenchmark.retainRelease_contended 100 sample 3039374 479.267 ± 0.435 ns/op
AbstractReferenceCountedByteBufBenchmark.retainRelease_contended 1000 sample 3709210 2044.603 ± 0.989 ns/op
AbstractReferenceCountedByteBufBenchmark.retainRelease_contended 10000 sample 3011591 19904.227 ± 18.025 ns/op
AbstractReferenceCountedByteBufBenchmark.retainRelease_uncontended 1 sample 494975 52.269 ± 8.345 ns/op
AbstractReferenceCountedByteBufBenchmark.retainRelease_uncontended 10 sample 771094 62.290 ± 0.795 ns/op
AbstractReferenceCountedByteBufBenchmark.retainRelease_uncontended 100 sample 763230 235.044 ± 1.552 ns/op
AbstractReferenceCountedByteBufBenchmark.retainRelease_uncontended 1000 sample 634037 2006.578 ± 3.574 ns/op
AbstractReferenceCountedByteBufBenchmark.retainRelease_uncontended 10000 sample 506284 19742.605 ± 13.729 ns/op
```
Motivation:
Most, but not all defaults are statically exposed on
PooledByteBufAllocator. This makes it cumbersome to make a custom
allocator where most of the defaults remain the same.
Modification:
Expose useCacheForAllThreads, and Direct preferred. The latter is
needed because it is under the internal package, and public code
should probably not depend on it.
Result:
More customizeable allocators
Motivation:
The constrcutors a protected atm but the classes are public. We should make the constructors public as well to make it easier to write your own ByteBufAllocator.
Modifications:
Change constructors to be public and add some javadocs.
Result:
Easier to create own ByteBufAllocator.
