Motivation
Currently a static AtomicLong is used to allocate a unique id whenever a
task is scheduled to any event loop. This could be a source of
contention if delayed tasks are scheduled at a high frequency and can be
easily avoided by having a non-volatile id counter per queue.
Modifications
- Replace static AtomicLong ScheduledFutureTask#nextTaskId with a long
field in AbstractScheduledExecutorService
- Set ScheduledFutureTask#id based on this when adding the task to the
queue (in event loop) instead of at construction time
- Add simple benchmark
Result
Less contention / cache-miss possibility when scheduling future tasks
Before:
Benchmark (num) Mode Cnt Score Error Units
scheduleLots 100000 thrpt 20 346.008 ± 21.931 ops/s
Benchmark (num) Mode Cnt Score Error Units
scheduleLots 100000 thrpt 20 654.824 ± 22.064 ops/s
Motivation:
peek() is implemented in a similar way to poll() for the mpsc queue, thus it is more like a consumer call.
It is possible that we could have multiple thread call peek() and possibly one thread calls poll() at at the same time.
This lead to multiple consumer scenario, which violates the multiple producer single consumer condition and could lead to spin in an infinite loop in peek()
Modification:
Use isEmpty() instead of peek() to check if task queue is empty
Result:
Dont violate the mpsc semantics.
Motivation:
SystemPropertyUtil already uses the AccessController internally so not need to wrap its usage with AccessController as well.
Modifications:
Remove explicit AccessController usage when SystemPropertyUtil is used.
Result:
Code cleanup
Motivation:
At the current moment HttpContentEncoder handle only first value of multiple accept-encoding headers.
Modification:
Join multiple accept-encoding headers to one separated by comma.
Result:
Fixes#9553
Motivation
Currently every call to get() on a promise results in two reads of the
volatile result field when one would suffice. Maybe this is optimized
away but it seems sensible not to rely on that.
Modification
Reimplement get() and get(...) in DefaultPromise to reduce volatile access.
Result
Fewer volatile reads.
Motivation
#9152 reverted some static exception reuse optimizations due to the
problem with Throwable#addSuppressed() raised in #9151. This introduced
a performance issue when promises are cancelled at a high frequency due
to the construction cost of CancellationException at the time that
DefaultPromise#cancel() is called.
Modifications
- Reinstate the prior static CANCELLATION_CAUSE_HOLDER but use it just
as a sentinel to indicate cancellation, constructing a new
CancellationException only if/when one needs to be explicitly
returned/thrown
- Subclass CancellationException, overriding fillInStackTrace() to
minimize the construction cost in these cases
Result
Promises are much cheaper to cancel. Fixes#9522.
Motivation:
There are some extra log level checks (logger.isWarnEnabled()).
Modification:
Remove log level checks (logger.isWarnEnabled()) from io.netty.channel.epoll.AbstractEpollStreamChannel, io.netty.channel.DefaultFileRegion, io.netty.channel.nio.AbstractNioChannel, io.netty.util.HashedWheelTimer, io.netty.handler.stream.ChunkedWriteHandler and io.netty.channel.udt.nio.NioUdtMessageConnectorChannel
Result:
Fixes#9456
Motivation:
The Netty classes are initialized at build time by default for GraalVM Native Image compilation. This is configured via the `--initialize-at-build-time=io.netty` option. While this reduces start-up time it can lead to some problems:
- The class initializer of `io.netty.buffer.PooledByteBufAllocator` looks at the maximum memory size to compute the size of internal buffers. If the class initializer runs during image generation, then the buffers are sized according to the very large heap size that the image generator uses, and Netty allocates several arrays that are 16 MByte. The fix is to initialize the following 3 classes at run time: `io.netty.buffer.PooledByteBufAllocator,io.netty.buffer.ByteBufAllocator,io.netty.buffer.ByteBufUtil`. This fix was dependent on a GraalVM Native Image fix that was included in 19.2.0.
- The class initializer of `io.netty.handler.ssl.util.ThreadLocalInsecureRandom` needs to be initialized at runtime to ensure that the generated values are trully random and not fixed for each generated image.
- The class initializers of `io.netty.buffer.AbstractReferenceCountedByteBuf` and `io.netty.util.AbstractReferenceCounted` compute field offsets. While the field offset recomputation is necessary for correct execution as a native image these initializers also have logic that depends on the presence/absence of `sun.misc.Unsafe`, e.g., via the `-Dio.netty.noUnsafe=true` flag. The fix is to push these initializers to runtime so that the field offset lookups (and the logic depending on them) run at run time. This way no manual substitutions are necessary either.
Modifications:
Add `META-INF/native-image` configuration files that correctly trigger the inialization of the above classes at run time via `--initialize-at-run-time=...` flags.
Result:
Fixes the initialisation issues described above for Netty executables built with GraalVM.
Motivation
The epoll transport was updated in #7834 to decouple setting of the
timerFd from the event loop, so that scheduling delayed tasks does not
require waking up epoll_wait. To achieve this, new overridable hooks
were added in the AbstractScheduledEventExecutor and
SingleThreadEventExecutor superclasses.
However, the minimumDelayScheduledTaskRemoved hook has no current
purpose and I can't envisage a _practical_ need for it. Removing
it would reduce complexity and avoid supporting this specific
API indefinitely. We can add something similar later if needed
but the opposite is not true.
There also isn't a _nice_ way to use the abstractions for
wakeup-avoidance optimizations in other EventLoops that don't have a
decoupled timer.
This PR replaces executeScheduledRunnable and
wakesUpForScheduledRunnable
with two new methods before/afterFutureTaskScheduled that have slightly
different semantics:
- They only apply to additions; given the current internals there's no
practical use for removals
- They allow per-submission wakeup decisions via a boolean return val,
which makes them easier to exploit from other existing EL impls (e.g.
NIO/KQueue)
- They are subjectively "cleaner", taking just the deadline parameter
and not exposing Runnables
- For current EL/queue impls, only the "after" hook is really needed,
but specialized blocking queue impls can conditionally wake on task
submission (I have one lined up)
Also included are further optimization/simplification/fixes to the
timerFd manipulation logic.
Modifications
- Remove AbstractScheduledEventExecutor#minimumDelayScheduledTaskRemoved()
and supporting methods
- Uplift NonWakeupRunnable and corresponding default wakesUpForTask()
impl from SingleThreadEventLoop to SingleThreadEventExecutor
- Change executeScheduledRunnable() to be package-private, and have a
final impl in SingleThreadEventExecutor which triggers new overridable
hooks before/afterFutureTaskScheduled()
- Remove unnecessary use of bookend tasks while draining the task queue
- Use new hooks to add simpler wake-up avoidance optimization to
NioEventLoop (primarily to demonstrate utility/simplicity)
- Reinstate removed EpollTest class
In EpollEventLoop:
- Refactor to use only the new afterFutureTaskScheduled() hook for
updating timerFd
- Fix setTimerFd race condition using a monitor
- Set nextDeadlineNanos to a negative value while the EL is awake and
use this to block timer changes from outside the EL. Restore the
known-set value prior to sleeping, updating timerFd first if necessary
- Don't read from timerFd when processing expiry event
Result
- Cleaner API for integrating with different EL/queue timing impls
- Fixed race condition to avoid missing scheduled wakeups
- Eliminate unnecessary timerFd updates while EL is awake, and
unnecessary expired timerFd reads
- Avoid unnecessary scheduled-task wakeups when using NIO transport
I did not yet further explore the suggestion of using
TFD_TIMER_ABSTIME for the timerFd.
Motivation:
AsciiString.contentEqualsIgnoreCase may return true for non-matching strings of equal length when offset is non zero.
Modifications:
- Correctly take offset into account
- Add unit test
Result:
Fixes#9475
Motivation:
EPOLL supports decoupling the timed wakeup mechanism from the selector call. The EPOLL transport takes advantage of this in order to offer more fine grained timer resolution. However we are current calling timerfd_settime on each call to epoll_wait and this is expensive. We don't have to re-arm the timer on every call to epoll_wait and instead only have to arm the timer when a task is scheduled with an earlier expiration than any other existing scheduled task.
Modifications:
- Before scheduled tasks are added to the task queue, we determine if the new
duration is the soonest to expire, and if so update with timerfd_settime. We
also drain all the tasks at the end of the event loop to make sure we service
any expired tasks and get an accurate next time delay.
- EpollEventLoop maintains a volatile variable which represents the next deadline to expire. This variable is modified inside the event loop thread (before calling epoll_wait) and out side the event loop thread (immediately to ensure proper wakeup time).
- Execute the task queue before the schedule task priority queue. This means we
may delay the processing of scheduled tasks but it ensures we transfer all
pending tasks from the task queue to the scheduled priority queue to run the
soonest to expire scheduled task first.
- Deprecate IORatio on EpollEventLoop, and drain the executor and scheduled queue on each event loop wakeup. Coupling the amount of time we are allowed to drain the executor queue to a proportion of time we process inbound IO may lead to unbounded queue sizes and unpredictable latency.
Result:
Fixes https://github.com/netty/netty/issues/7829
- In most cases this results in less calls to timerfd_settime
- Less event loop wakeups just to check for scheduled tasks executed outside the event loop
- More predictable executor queue and scheduled task queue draining
- More accurate and responsive scheduled task execution
Motivation:
Users' runtime systems may have incompatible dynamic libraries to the ones our
tcnative wrappers link to. Unfortunately, we cannot determine and catch these
scenarios (in which the JVM crashes) but we can make a more educated guess on
what library to load and try to find one that works better before crashing.
Modifications:
1) Build dynamically linked openSSL builds for more OSs (netty-tcnative)
2) Load native linux libraries with matching classifier (first)
Result:
More developers / users can use the dynamically-linked native libraries.
Motivation:
We did miss to call reclaimSpace(...) in one case which can lead to the situation of having the Recycler to not correctly reclaim space and so just create new objects when not needed.
Modifications:
Correctly call reclaimSpace(...)
Result:
Recycler correctly reclaims space in all situations.
Motivation:
Netty homepage(netty.io) serves both "http" and "https".
It's recommended to use https than http.
Modification:
I changed from "http://netty.io" to "https://netty.io"
Result:
No effects.
Motivation:
There are is some unnecessary code (like toString() calls) which can be cleaned up.
Modifications:
- Remove not needed toString() calls
- Simplify subString(...) calls
- Remove some explicit casts when not needed.
Result:
Cleaner code
Motivation:
There is some manual coping of elements of Collections which can be replaced by Collections.addAll(...) and also some unnecessary semicolons.
Modifications:
- Simplify branches
- Use Collections.addAll
- Code cleanup
Result:
Code cleanup
Motivation:
Some methods that either override others or are implemented as part of implementation an interface did miss the `@Override` annotation
Modifications:
Add missing `@Override`s
Result:
Code cleanup
Motivation:
asList should only be used if there are multiple elements.
Modification:
Call to asList with only one argument could be replaced with singletonList
Result:
Cleaner code and a bit of memory savings
Motivation:
Sometimes it is desirable to be able to use a different Queue implementation for the EventLoop of a Channel. This is currently not possible without resort to reflection.
Modifications:
- Add a new constructor to Nio|Epoll|KQueueEventLoopGroup which allows to specify a factory which is used to create the task queue. This was the user can override the default implementation.
- Add test
Result:
Be able to change Queue that is used for the EventLoop.
Motivation
A Semaphore is currently dedicated to this purpose but a simple
CountDownLatch will do.
Modification
Remove private threadLock Semaphore from SingleThreadEventExecutor and just use a CountDownLatch.
Also eliminate use of PlatformDependent.throwException() in startThread
method, and combine some nested if clauses.
Result
Cleaner EventLoop termination notification.
Motivation:
OOME is occurred by increasing suppressedExceptions because other libraries call Throwable#addSuppressed. As we have no control over what other libraries do we need to ensure this can not lead to OOME.
Modifications:
Only use static instances of the Exceptions if we can either dissable addSuppressed or we run on java6.
Result:
Not possible to OOME because of addSuppressed. Fixes https://github.com/netty/netty/issues/9151.
Motivation
@xiaoheng1 reported incorrect behaviour of AsciiString.lastIndexOf in
#9099. Upon closer inspection it appears that it was never implemented
correctly and searches between the provided index and the end of the
string similar to indexOf(...), rather than between the provided index
and the beginning of the string as the javadoc states (and in line with
java.lang.String).
Modifications
Fix AsciiString.lastIndexOf implementation and corresponding unit tests
to behave the same as the equivalent String methods.
Result
Fixes#9099
Motivation:
An instance is always equal to itself. It makes sense to skip processing for this case, which isn't uncommon since `AsciiString` is often memoized within an application when used as HTTP header names.
Modification:
`contentEquals` methods first check for instance equality before doing processing.
Result:
`contentEquals` will be faster when comparing an instance with itself.
I couldn't find any unit tests for these methods, only the static version. Let me know if I should add something to `AsciiStringCharacterTest`.
Came up here:
https://github.com/line/armeria/pull/1731#discussion_r280396280
Motivation:
GraalVM native images are a new way to deliver java applications. Netty is one of the most popular libraries however there are a few limitations that make it impossible to use with native images out of the box. Adding a few metadata (in specific modules will allow the compilation to success and produce working binaries)
Modification:
Added properties files in `META-INF` and substitutions classes (under `internal.svm`) will solve the compilation issues. The substitutions classes are not visible and do not have a public constructor so they are not visible to end users.
Result:
Fixes#8959
This fix is very conservative as it applies the minimum config required to build:
* pure netty servers
* vert.x applications
* grpc applications
The build is having trouble due to checkstyle which does not seem to be able to find the copyright notice on property files.
Motivation:
In GlobalEventExecutorTest we used Thread.sleep(...) which can produce flaky results (as seen on the CI). We should use another alternative during tests.
Modifications:
Replace Thread.sleep(...) with join()
Result:
No more flaky GlobalEventExecutor tests.
Motivation:
IKVM.NET seems to ship a bug sun.misc.Unsafe class, for this reason we should better disable our sun.misc.Unsafe usage when we detect IKVM.NET is used.
Modifications:
Check if IKVM.NET is used and if so do not use sun.misc.Unsafe by default.
Result:
Fixes https://github.com/netty/netty/issues/9035 and https://github.com/netty/netty/issues/8916.
Motivation
AbstractReferenceCounted and AbstractReferenceCountedByteBuf contain
duplicate logic for managing the volatile refcount in an optimized and
consistent manner, which increased in complexity in #8583. It's possible
to extract this into a common helper class now that all access is via an
AtomicIntegerFieldUpdater.
Modifications
- Move duplicate logic into a shared ReferenceCountUpdater class
- Incorporate some additional simplification for the most common single
increment/decrement cases (fewer checks/operations)
Result
Less code duplication, better encapsulation of the "non-trivial"
internal volatile refcount manipulation
Motivation:
PooledByteBufAllocator uses a PoolThreadCache per Thread that allocates / deallocates to minimize the performance overhead. This PoolThreadCache is trimmed after X allocations to free up buffers that are not allocated for a long time. This works out quite well when the app continues to allocate but fails if the app stops to allocate frequently (for whatever reason) and so a lot of memory is wasted and not given back to the arena / freed.
Modifications:
- Add a ThreadExecutorMap that offers multiple methods that wrap Runnable / ThreadFactory / Executor and allow to call ThreadExecutorMap.currentEventExecutor() to get the current executing EventExecutor for the calling Thread.
- Use these methods in the constructors of our EventExecutor implementations (which also covers the EventLoop implementations)
- Add io.netty.allocator.cacheTrimIntervalMillis system property which can be used to specify a fixed rate / interval on which we should try to trim the PoolThreadCache for a EventExecutor that allocates.
- Add PooledByteBufAllocator.trimCurrentThreadCache() to allow the user to trim the cache of the calling thread manually.
- Add testcases
- Introduce FastThreadLocal.getIfExists()
Result:
Allow to better / more frequently trim PoolThreadCache and so give back memory to the area / system.
Motivation:
We dont use ObjectCleaner in our FastThreadLocal anymore so we also dont need to take special care to store it there anymore.
Modifications:
Remove code that is not needed anymore.
Result:
Code cleanup.
Motivation:
GlobalEventExecutor does already provide all guarantees of OrderedEventExecutor so it should implement it.
Modifications:
Let GlobalEventExecutor implement OrderedEventExecutor.
Result:
Make it more clear how execution order is handled in GlobalEventExecutor.
Motivation:
This counter is very useful in order to monitor Netty without having every ByteBufAllocator in the JVM
Modification:
Expose the value of DIRECT_MEMORY_COUNTER as we are already doing for DIRECT_MEMORY_LIMIT.
We are returning -1 in case that DIRECT_MEMORY_COUNTER is not available.
Result:
Be able to get the amount of direct memory used.
Motivation:
We had a typo in NativeLibraryLoader debug log message which could misslead the user.
Modifications:
Fix typo to correctly state java.library.path
Result:
Correct and less confusing log message
