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
Currently an epoll_ctl syscall is made every time there is a change to
the event interest flags (EPOLLIN, EPOLLOUT, etc) of a channel. These
are only done in the event loop so can be aggregated into 0 or 1 such
calls per channel prior to the next call to epoll_wait.
Modifications
I think further streamlining/simplification is possible but for now I've
tried to minimize structural changes and added the aggregation beneath
the existing flag manipulation logic.
A new AbstractChannel#activeFlags field records the flags last set on
the epoll fd for that channel. Calls to setFlag/clearFlag update the
flags field as before but instead of calling epoll_ctl immediately, just
set or clear a bit for the channel in a new bitset in the associated
EpollEventLoop to reflect whether there's any change to the last set
value.
Prior to calling epoll_wait the event loop makes the appropriate
epoll_ctl(EPOLL_CTL_MOD) call once for each channel who's bit is set.
Result
Fewer syscalls, particularly in some auto-read=false cases. Simplified
error handling from centralization of these calls.
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:
EpollDatagramChannel#localAddress returns wrong information when
EpollDatagramChannel is created with InternetProtocolFamily,
and EpollDatagramChannel#localAddress is invoked BEFORE the actual binding.
This is a regression caused by change
e17ce934da
Modifications:
EpollDatagramChannel() and EpollDatagramChannel(InternetProtocolFamily family)
do not cache local/remote address
Result:
Rebinding on the same address without "reuse port" works
EpollDatagramChannel#localAddress returns correct address
Motivation:
On openSUSE (probably more), 64 bit builds use lib64, e.g. /usr/lib64, and
configure picks this up and builds the native library in
native-build/target/lib64 where maven is not looking.
Modifications:
Explicitly specify --libdir=${project.build.directory}/native-build/target/lib
during configuration.
Result:
Maven uses the correct lib directory.
Motivation:
We recently made a change to use ET for the eventfd and not trigger a read each time. This testcase proves everything works as expected.
Modifications:
Add testcase that verifies thqat the wakeups happen correctly
Result:
More tests
Motivation
The AbstractEpollStreamChannel::spliceTo(FileDescriptor, ...) methods
take an offset parameter but this was effectively ignored due to what
looks like a typo in the corresponding JNI function impl. Instead it
would always use the file's own native offset.
Modification
- Fix typo in netty_epoll_native_splice0() and offset accounting in
AbstractEpollStreamChannel::SpliceFdTask.
- Modify unit test to include an invocation of the public spliceTo
method using non-zero offset.
Result
spliceTo FD methods work as expected when an offset is provided.
Motivation
I noticed this while looking at something else.
AbstractEpollStreamChannel::spliceQueue is an MPSC queue but only
accessed from the event loop. So it could be just changed to e.g. an
ArrayDeque. This PR instead reverts to using is as an MPSC queue to
avoid dispatching a task to the EL, as appears was the original
intention.
Modification
Change AbstractEpollStreamChannel::spliceQueue to be volatile and lazily
initialized via double-checked locking. Add tasks directly to the queue
from the public methods rather than possibly waking the EL just to
enqueue.
An alternative is just to change PlatformDependent.newMpscQueue() to new
ArrayDeque() and be done with it :)
Result
Less disruptive channel/fd-splicing.
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:
Compiling with -Werror,-Wuninitialized complains about the sockaddrs being uninitialized.
I believe this is because the init function netty_unix_socket_initSockaddr is in a
separate compilation unit. Since this code isn't on the criticial path, it's easy
to just memset the variables rather than suppress the warning.
Modification:
Always clear the sockaddrs, even if they will be initialized later.
Result:
Able to compile with warnings turned on
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:
c9aaa93d83 added the ability to specify an EventLoopTaskQueueFactory but did place it under MultithreadEventLoopGroup while not really belongs there.
Modifications:
Make EventLoopTaskQueueFactory a top-level interface
Result:
More logical code layout.
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:
We did not have support for enable / disable loopback mode in our native epoll transport and also missed the implemention to access the configured interface.
Modifications:
Add implementation and adjust test to cover it
Result:
More complete multicast support with native epoll transport
Motivation:
The wakeup logic in EpollEventLoop is overly complex
Modification:
* Simplify the race to wakeup the loop
* Dont let the event loop wake up itself (it's already awake!)
* Make event loop check if there are any more tasks after preparing to
sleep. There is small window where the non-eventloop writers can issue
eventfd writes here, but that is okay.
Result:
Cleaner wakeup logic.
Benchmarks:
```
BEFORE
Benchmark Mode Cnt Score Error Units
EpollSocketChannelBenchmark.executeMulti thrpt 20 408381.411 ± 2857.498 ops/s
EpollSocketChannelBenchmark.executeSingle thrpt 20 157022.360 ± 1240.573 ops/s
EpollSocketChannelBenchmark.pingPong thrpt 20 60571.704 ± 331.125 ops/s
Benchmark Mode Cnt Score Error Units
EpollSocketChannelBenchmark.executeMulti thrpt 20 440546.953 ± 1652.823 ops/s
EpollSocketChannelBenchmark.executeSingle thrpt 20 168114.751 ± 1176.609 ops/s
EpollSocketChannelBenchmark.pingPong thrpt 20 61231.878 ± 520.108 ops/s
```
Motivation:
We do not need to issue a read on timerfd and eventfd when the EventLoop wakes up if we register these as Edge-Triggered. This removes the overhead of 2 syscalls and so helps to reduce latency.
Modifications:
- Ensure we register the timerfd and eventfd with EPOLLET flag
- If eventfd_write fails with EAGAIN, call eventfd_read and try eventfd_write again as we only use it as wake-up mechanism.
Result:
Less syscalls and so reducing overhead.
Co-authored-by: Carl Mastrangelo <carl@carlmastrangelo.com>
Motivation:
When EpollDatagramChannel is created with an existing FileDescriptor we should detect the correct InternetProtocolFamily.
Modifications:
Obtain the InternetProtocolFamily from the given FD
Result:
Use correct InternetProtocolFamily when EpollDatagramChannel is created via existing FileDescriptor
Motivation:
Provide epoll/native multicast to support high load multicast users (we are using it for a high load telecomm app at my day job).
Modification:
Added support for source specific and any source multicast for epoll transport. Some caveats: no support for disabling loop back mode, retrieval of interface and block operation, all of which tend to be less frequently used.
Result:
Provides epoll transport multicast for common use cases.
Co-authored-by: Norman Maurer <norman_maurer@apple.com>
Motivation:
The current KQueueEventLoop implementation does not process concurrent domain socket channel registration/unregistration in the order they actual
happen since unregistration are delated by an event loop task scheduling. When a domain socket is closed, it's file descriptor might be reused
quickly and therefore trigger a new channel registration using the same descriptor.
Consequently the KQueueEventLoop#add(AbstractKQueueChannel) method will overwrite the current inactive channels having the same descriptor
and the delayed KQueueEventLoop#remove(AbstractKQueueChannel) will remove the active channel that replaced the inactive one.
As active channels are registered, events for this file descriptor won't be processed anymore and the channels will never be closed.
The same problem can also happen in EpollEventLoop. Beside this we also may never remove the AbstractEpollChannel from the internal map
when it is unregistered which will prevent it from be GC'ed
Modifications:
- Change logic of native KQueue and Epoll implementations to ensure we correctly handle the case of FD reuse
- Only try to update kevent / epoll if the Channel is still open (as otherwise it will be handled by kqueue / epoll itself)
- Correctly remove AbstractEpollChannel from internal map in all cases
- Make implementation of closeAll() consistent for Epoll and KQueueEventLoop
Result:
KQueue and Epoll native transports correctly handle FD reuse
Co-authored-by: Norman Maurer <norman_maurer@apple.com>
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
These aren't needed, only one field from each class is used. It also showed as an ambiguous identifier compilation error in my IDE even though javac is obviously fine with it.
Modifications
Static-import explicit ChannelOption fields in EpollDomainSocketChannelConfig instead of using .* wildcard.
Result
Cleaner / more consistent code.
Motivation
These implementations delegate most of their methods to an existing Handle and previously extended RecvByteBufAllocator.DelegatingHandle. This was reverted in #6322 with the introduction of ExtendedHandle but it's not clear to me why it needed to be - the code looks a lot cleaner.
Modifications
Have (Epoll|KQueue)RecvByteAllocatorHandle extend DelegatingHandle again, while still implementing ExtendedHandle.
Result
Less code.
Motivation:
At the moment we throw a ChannelException if netty_epoll_linuxsocket_setTcpMd5Sig fails. This is inconsistent with other methods which throw a IOException.
Modifications:
Throw IOException
Result:
More correct and consistent exception usage in epoll transport
Motivation:
We currently only cover ipv4 multicast in the testsuite but we should also have tests for ipv6.
Modifications:
- Add test for ipv6
- Ensure we only try to run multicast test for ipv4 / ipv6 if the loopback interface supports it.
Result:
Better test coverage
Motivation:
When netty_epoll_linuxsocket_setTcpMd5Sig fails to init the sockaddr we should throw an exception and not silently return.
Modifications:
Throw exception if init of sockaddr fails.
Result:
Correctly report back error to user.
…nterface, block or loopback-mode-disabled operations).
Motivation:
Provide epoll/native multicast to support high load multicast users (we are using it for a high load telecomm app at my day job).
Modification:
Added support for (ipv4 only) source specific and any source multicast for epoll transport. Some caveats (beyond no ipv6 support initially - there’s a bit of work to add in join and leave group specifically around SSM, as ipv6 uses different data structures for this): no support for disabling loop back mode, retrieval of interface and block operation, all of which tend to be less frequently used.
Result:
Provides epoll transport multicast for IPv4 for common use cases. Understand if you’d prefer to hold off until ipv6 is included but not sure when I’ll be able to get to that.
Motivation:
Systems depending on Netty may benefit (telemetry, alternative even loop scheduling algorithms) from knowing the number of channels assigned to each EventLoop.
Modification:
Expose the number of channels registered in the EventLoop via SingleThreadEventLoop.registeredChannels.
Result:
Fixes#8276.
Motivation:
com.puppycrawl.tools checkstyle < 8.18 was reported to contain a possible security flaw. We should upgrade.
Modifications:
- Upgrade netty-build and checkstyle.
- Fix checkstyle errors
Result:
Fixes https://github.com/netty/netty/issues/8968.
Motivation:
Since DomainSocketChannel is a DuplexChannel, which be able to shutdown input or output individually on demands, but ALLOW_HALF_CLOSURE channel option has not been supported yet.
I thought this could be a missing feature of Unix domain socket, so here the PR for it.
Modifications:
1. Added allHalfClosure property both in EpollDomainSocketChannelConfig and KQueueDomainSocketChannelConfig,
2. Enabled isAllowHalfClosure method of native channel to support domain channel config,
3. Created EpollDomainSocketShutdownOutputByPeerTest and KQueueDomainSocketShutdownOutputByPeerTest to verify the change.
Result:
ALLOW_HALF_CLOSURE channel option can be set with DomainSocketChannel, and no more warning of Unknown channel option 'ALLOW_HALF_CLOSURE'.
Motivation:
At some point we needed --add-opens=java.base/java.nio=ALL-UNNAMED to run our native tests but this is not true anymore.
Modifications:
Remove --add-opens=java.base/java.nio=ALL-UNNAMED when running native tests.
Result:
Remove obsolate jvm arg.
Motivation:
`DefaultFileRegion.transferTo` will return 0 all the time when we request more data then the actual file size. This may result in a busy spin while processing the fileregion during writes.
Modifications:
- If we wrote 0 bytes check if the underlying file size is smaller then the requested count and if so throw an IOException
- Add DefaultFileRegionTest
- Add a test to the testsuite
Result:
Fixes https://github.com/netty/netty/issues/8868.
Motivation:
The SSLEngine does provide a way to signal to the caller that it may need to execute a blocking / long-running task which then can be offloaded to an Executor to ensure the I/O thread is not blocked. Currently how we handle this in SslHandler is not really optimal as while we offload to the Executor we still block the I/O Thread.
Modifications:
- Correctly support offloading the task to the Executor while suspending processing of SSL in the I/O Thread
- Add new methods to SslContext to specify the Executor when creating a SslHandler
- Remove @deprecated annotations from SslHandler constructor that takes an Executor
- Adjust tests to also run with the Executor to ensure all works as expected.
Result:
Be able to offload long running tasks to an Executor when using SslHandler. Partly fixes https://github.com/netty/netty/issues/7862 and https://github.com/netty/netty/issues/7020.
Motivation:
We missed to extend a few tests from the testsuite and so also run these with our native KQueue and Epoll transport.
Modifications:
Extend tests and so run these for our native transports as well.
Result:
More tests.
Motivation:
When using a linux distribution that supports sendmmsg(...) we allocated enough direct memory per EpollEventLoop to be able to write IOV_MAX number of iovecs per message that can be written per sendmmsg.
The number of messages that can be written per sendmmsg(...) call is limited by UIO_MAX_IOV.
In practice this resulted in an allocation of 16MB direct memory per EpollEventLoop instance that stayed allocated until the EpollEventLoop was shutdown which happens as part of the shutdown of the enclosing EpollEVentLoopGroup.
This resulted in quite some heavy direct memory usage in practice even when in practice we have very slim changes to ever need all of the memory.
Modification:
Adjust NativeDatagramPacketArray to share one IovArray instance across all NativeDatagramPacket instances it holds. This limits the max number of iovecs we can write across all messages to IOV_MAX per sendmmsg(...) call.
This in practice will still be enough to allow us to write multiple messages with one syscall while keep the memory overhead to a minimum.
Result:
Smaller direct memory footprint per EpollEventLoop when using EpollDatagramChannel on distributions that support sendmmsg(...).
Fixes https://github.com/netty/netty/issues/8814
Motivation:
In the native code EpollDatagramChannel / KQueueDatagramChannel creates a DatagramSocketAddress object for each received UDP datagram even when in connected mode as it uses the recvfrom(...) / recvmsg(...) method. Creating these is quite heavy in terms of allocations as internally, char[], String, Inet4Address, InetAddressHolder, InetSocketAddressHolder, InetAddress[], byte[] objects are getting generated when constructing the object. When in connected mode we can just use regular read(...) calls which do not need to allocate all of these.
Modifications:
- When in connected mode use read(...) and NOT recvfrom(..) / readmsg(...) to reduce allocations when possible.
- Adjust tests to ensure read works as expected when in connected mode.
Result:
Less allocations and GC when using native datagram channels in connected mode. Fixes https://github.com/netty/netty/issues/8770.
Motivation:
We have a utility method to check for > 0 and >0 arguments. We should use it.
Modification:
use checkPositive/checkPositiveOrZero instead of if statement.
Result:
Re-use utility method.
Motivation:
We need to update to a new checkstyle plugin to allow the usage of lambdas.
Modifications:
- Update to new plugin version.
- Fix checkstyle problems.
Result:
Be able to use checkstyle plugin which supports new Java syntax.
Motivation:
While using Load Balancers or HA support is needed there are cases when UDP channel need to bind to IP Address which is not available on network interfaces locally.
Modification:
Modified EpollDatagramChannelConfig to allow IP_FREEBIND option
Result:
Fixes ##8727.
Motivation:
Most of the maven modules do not explicitly declare their
dependencies and rely on transitivity, which is not always correct.
Modifications:
For all maven modules, add all of their dependencies to pom.xml
Result:
All of the (essentially non-transitive) depepdencies of the modules are explicitly declared in pom.xml
Motivation:
https://github.com/netty/netty/issues/8444 reports that there is some issue with negative values passed to timerfd_settime. This test verifies that everything is working as expected.
Modifications:
Add testcase.
Result:
Test to verify expected behaviour.
Motivation:
epoll_wait should work in 4.1.30 like it did in 4.1.29.
Modifications:
Revert Integer.MAX_VALUE back to MAX_SCHEDULED_TIMERFD_NS (999,999,999).
Add unit test.
Result:
epoll_wait will no longer throw EINVAL.
Motivation:
Add an option (through a SelectStrategy return code) to have the Netty event loop thread to do busy-wait on the epoll.
The reason for this change is to avoid the context switch cost that comes when the event loop thread is blocked on the epoll_wait() call.
On average, the context switch has a penalty of ~13usec.
This benefits both:
The latency when reading from a socket
Scheduling tasks to be executed on the event loop thread.
The tradeoff, when enabling this feature, is that the event loop thread will be using 100% cpu, even when inactive.
Modification:
Added SelectStrategy option to return BUSY_WAIT
Epoll loop will do a epoll_wait() with no timeout
Use pause instruction to hint to processor that we're in a busy loop
Result:
When enabled, minimizes impact of context switch in the critical path
Motivation
The EpollChannelConfig (same for KQueues) and its subclasses repeatetly declare their own channel field which leads to a 3x repetition for each config instance. Given the fields are protected or package-private it's exposing the code code to "field hiding" bugs.
Modifications
Use the the existing protected channel field from the DefaultChannelConfig class and simply cast it when needed.
Result
Fixes#8331
Motivation:
The Epoll transport checks to see if there are any scheduled tasks
before entering epoll_wait, and resets the timerfd just before.
This causes an extra syscall to timerfd_settime before doing any
actual work. When scheduled tasks aren't added frequently, or
tasks are added with later deadlines, this is unnecessary.
Modification:
Check the *deadline* of the peeked task in EpollEventLoop, rather
than the *delay*. If it hasn't changed since last time, don't
re-arm the timer
Result:
About 2us faster on gRPC RTT 50pct latency benchmarks.
Before (2 runs for 5 minutes, 1 minute of warmup):
```
50.0%ile Latency (in nanos): 64267
90.0%ile Latency (in nanos): 72851
95.0%ile Latency (in nanos): 78903
99.0%ile Latency (in nanos): 92327
99.9%ile Latency (in nanos): 119691
100.0%ile Latency (in nanos): 13347327
QPS: 14933
50.0%ile Latency (in nanos): 63907
90.0%ile Latency (in nanos): 73055
95.0%ile Latency (in nanos): 79443
99.0%ile Latency (in nanos): 93739
99.9%ile Latency (in nanos): 123583
100.0%ile Latency (in nanos): 14028287
QPS: 14936
```
After:
```
50.0%ile Latency (in nanos): 62123
90.0%ile Latency (in nanos): 70795
95.0%ile Latency (in nanos): 76895
99.0%ile Latency (in nanos): 90887
99.9%ile Latency (in nanos): 117819
100.0%ile Latency (in nanos): 14126591
QPS: 15387
50.0%ile Latency (in nanos): 61021
90.0%ile Latency (in nanos): 70311
95.0%ile Latency (in nanos): 76687
99.0%ile Latency (in nanos): 90887
99.9%ile Latency (in nanos): 119527
100.0%ile Latency (in nanos): 6351615
QPS: 15571
```
Motivation:
When using Epoll based transport, allow applications to configure SO_BUSY_POLL socket option:
SO_BUSY_POLL (since Linux 3.11)
Sets the approximate time in microseconds to busy poll on a
blocking receive when there is no data. Increasing this value
requires CAP_NET_ADMIN. The default for this option is con‐
trolled by the /proc/sys/net/core/busy_read file.
The value in the /proc/sys/net/core/busy_poll file determines
how long select(2) and poll(2) will busy poll when they oper‐
ate on sockets with SO_BUSY_POLL set and no events to report
are found.
In both cases, busy polling will only be done when the socket
last received data from a network device that supports this
option.
While busy polling may improve latency of some applications,
care must be taken when using it since this will increase both
CPU utilization and power usage.
Modification:
Added SO_BUSY_POLL socket option
Result:
Able to configure SO_BUSY_POLL from Netty
Motivation:
We should ensure we call *UnLoad when we detect an error during calling *OnLoad and previous *OnLoad calls were succesfull.
Modifications:
Correctly call *UnLoad when needed.
Result:
More correct code and no leaks when an error happens during loading the native lib.
* 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.
Motivation:
We should support to load multiple shaded versions of the same netty artifact as netty is often used in multiple dependencies.
This is related to https://github.com/netty/netty/issues/7272.
Modifications:
- Use -fvisibility=hidden when compiling and use JNIEXPORT for things we really want to have exported
- Ensure fields are declared as static so these are not exported
- Adjust testsuite-shading to use install_name_tool on MacOS to change the id of the lib. Otherwise the wrong may be used.
Result:
Be able to use multiple shaded versions of the same netty artifact.
Motivation:
Avoid unnecessary native memory allocation if UDP / TCP isn't being
used.
Modifications:
Create the reused NativeDatagramPacketArray and IovArray upon first use
instead of EpollEventLoop construction.
Also correct related comment in NativeDatagramPacketArray.
Result:
Reduced native memory use when using epoll in many cases
Motivation:
We can store the NativeDatagramPacketArray directly in the EpollEventLoop. This removes the need of using FastThreadLocal.
Modifications:
- Store NativeDatagramPacketArray directly in the EpollEventLoop (just as we do with IovArray as well).
Result:
Less FastThreadLocal usage and more consistent code.
Motivation:
Epoll and Kqueue channels have internal state which forces
a single read operation after channel construction. This
violates the Channel#read() interface which indicates that
data shouldn't be delivered until this method is called.
The behavior is also inconsistent with the NIO transport.
Modifications:
- Epoll and Kqueue shouldn't unconditionally read upon
initialization, and instead should rely upon Channel#read()
or auto_read.
Result:
Epoll and Kqueue are more consistent with NIO.
Motivation:
We should allow to schedule tasks with a delay up to Long.MAX_VALUE as we did pre 4.1.25.Final.
Modifications:
Just ensure we not overflow and put the correct max limits in place when schedule a timer. At worse we will get a wakeup to early and then schedule a new timeout.
Result:
Fixes https://github.com/netty/netty/issues/7970.
* Read until all data is consumed when EOF is detected even if readPending is false and auto-read is disabled.
Motivation:
We should better always notify the user of EOF even if the user did not request any data as otherwise we may never be notified when the remote peer closes the connection. This should be ok as the amount of extra data we may read and so fire through the pipeline is limited by SO_RECVBUF.
Modifications:
- Always drain the socket when EOF is detected.
- Add testcase
Result:
No risk for the user to be not notified of EOF.
Motivation:
Sometimes it's useful to disable native transports / native ssl to debug a problem. We should allow to do so with a system property so people not need to adjust code for this.
Modifications:
Add system properties which allow to disable native transport and native ssl.
Result:
Easier to disable native code usage without code changes.
Motivation:
DatagramPacket.recipient() doesn't return the actual destination IP, but the IP the app is bound to.
Modification:
- IP_RECVORIGDSTADDR option is enabled for UDP sockets, which allows retrieval of ancillary information containing the original recipient.
- _recvFrom(...) function from transport-native-unix-common/src/main/c/netty_unix_socket.c is modified such that if IP_RECVORIGDSTADDR is set, recvmsg is used instead of recvfrom; enabling the retrieval of the original recipient.
- DatagramSocketAddress also contains a 'local' address, representing the recipient.
- EpollDatagramChannel is updated to return the retrieved recipient address instead of the address the channel is bound to.
Result:
Fixes#4950.
Motivation:
Using a very huge delay when calling schedule(...) may cause an Selector error when calling select(...) later on. We should gaurd against such a big value.
Modifications:
- Add guard against a very huge value.
- Added tests.
Result:
Fixes [#7365]
Motivation:
This allows netty to operate in 'transparent proxy' mode for UDP, intercepting connections
to other addresses by means of Linux firewalling rules, as per
https://www.kernel.org/doc/Documentation/networking/tproxy.txt
Modification:
Add IP_TRANSPARENT option.
Result:
Allows setting and getting of the IP_TRANSPARENT option, which allows retrieval of the ultimate socket address originally requested.
Motivation:
AbstractNioByteChannel will detect that the remote end of the socket has
been closed and propagate a user event through the pipeline. However if
the user has auto read on, or calls read again, we may propagate the
same user events again. If the underlying transport continuously
notifies us that there is read activity this will happen in a spin loop
which consumes unnecessary CPU.
Modifications:
- AbstractNioByteChannel's unsafe read() should check if the input side
of the socket has been shutdown before processing the event. This is
consistent with EPOLL and KQUEUE transports.
- add unit test with @normanmaurer's help, and make transports consistent with respect to user events
Result:
No more read spin loop in NIO when the channel is half closed.
Motivation:
The flush task is currently using flush() which will have the affect of have the flush traverse the whole ChannelPipeline and also flush messages that were written since we gave up flushing. This is not really correct as we should only continue to flush messages that were flushed at the point in time when the flush task was submitted for execution if the user not explicit call flush() by him/herself.
Modification:
Call *Unsafe.flush0() via the flush task which will only continue flushing messages that were marked as flushed before.
Result:
More correct behaviour when the flush task is used.
Motivation:
b215794de3 recently introduced a change in behavior where writeSpinCount provided a limit for how many write operations were attempted per flush operation. However when the write quantum was meet the selector write flag was not cleared, and the channel unsafe flush0 method has an optimization which prematurely exits if the write flag is set. This may lead to no write progress being made under the following scenario:
- flush is called, but the socket can't accept all data, we set the write flag
- the selector wakes us up because the socket is writable, we write data and use the writeSpinCount quantum
- we then schedule a flush() on the EventLoop to execute later, however it the flush0 optimization prematurely exits because the write flag is still set
In this scenario the socket is still writable so the EventLoop may never notify us that the socket is writable, and therefore we may never attempt to flush data to the OS.
Modifications:
- When the writeSpinCount quantum is exceeded we should clear the selector write flag
Result:
Fixes https://github.com/netty/netty/issues/7729
Motivation:
IovArray implements MessageProcessor, and the processMessage method will continue to be called during iteration until it returns true. A recent commit b215794de3 changed the return value to only return true if any component of a CompositeByteBuf was added as a result of the method call. However this results in the iteration continuing, and potentially subsequent smaller buffers maybe added, which will result in out of order writes and generally corrupts data.
Modifications:
- IovArray#add should return false so that the MessageProcessor#processMessage will stop iterating.
Result:
Native transports which use IovArray will not corrupt data during gathering writes of CompositeByteBuf objects.
Motivation:
FileDescriptor#writev calls JNI code, and that JNI code dereferences a NULL pointer which crashes the application. This occurs when writing a single CompositeByteBuf object with more than one component.
Modifications:
- Initialize the iovec iterator properly to avoid the core dump
- Fix the array length calculation if we aren't able to fit all the ByteBuffer objects in the iovec array
Result:
No more core dump.
Motivation:
The writeSpinCount currently loops over the same buffer, gathering
write, file write, or other write operation multiple times but will
continue writing until there is nothing left or the OS doesn't accept
any data for that specific write. However if the OS keeps accepting
writes there is no way to limit how much time we spend on a specific
socket. This can lead to unfair consumption of resources dedicated to a
single socket.
We currently don't limit the amount of bytes we attempt to write per
gathering write. If there are many more bytes pending relative to the
SO_SNDBUF size we will end up building iov arrays with more elements
than can be written, which results in extra iteration, conditionals,
and book keeping.
Modifications:
- writeSpinCount should limit the number of system calls we make to
write data, instead of applying to individual write operations
- IovArray should support a maximum number of bytes
- IovArray should support composite buffers of greater than size 1024
- We should auto-scale the amount of data that we attempt to write per
gathering write operation relative to SO_SNDBUF and how much data is
successfully written
- The non-unsafe path should also support a maximum number of bytes,
and respect the IOV_MAX limit
Result:
Write resource consumption can be bounded and gathering writes have
a limit relative to the amount of data which can actually be accepted
by the socket.
Motivation:
We used NetUtil.isIpV4StackPreferred() when loading JNI code which tries to load NetworkInterface in its static initializer. Unfortunally a lock on the NetworkInterface class init may be already hold somewhere else which may cause a loader deadlock.
Modifications:
Add a new Socket.initialize() method that will be called when init the library and pass everything needed to the JNI level so we not need to call back to java.
Result:
Fixes [#7458].
Motivation:
AbstractChannel attempts to "filter" messages which are written [1]. A goal of this process is to copy from heap to direct if necessary. However implementations of this method [2][3] may translate a buffer with 0 readable bytes to EMPTY_BUFFER. This may mask a user error where an empty buffer is written but already released.
Modifications:
Replace safeRelease(...) with release(...) to ensure we propagate reference count issues.
Result:
Fixes [#7383]
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:
To better isolate OS system calls we should not call getsockopt directly but use our netty_unix_socket_getOption0 function. See is a followup of f115bf5.
Modifications:
Export netty_unix_socket_getOption0 by declaring it in the header file and use it
Result:
Better isolation of system calls.
Motivation:
If a user calls EpollSocketChannelConfig.getOptions() and TCP_FASTOPEN_CONNECT is not supported we throw an exception.
Modifications:
- Just return 0 if ENOPROTOOPT is set.
- Add testcase
Result:
getOptions() works as epxected.
Motivation:
We need to set readPending to false when we detect a EOF while issue a read as otherwise we may not unregister from the Selector / Epoll / KQueue and so keep on receving wakeups.
The important bit is that we may even get a wakeup for a read event but will still will only be able to read 0 bytes from the socket, so we need to be very careful when we clear the readPending. This can happen because we generally using edge-triggered mode for our native transports and because of the nature of edge-triggered we may schedule an read event just to find out there is nothing left to read atm (because we completely drained the socket on the previous read).
Modifications:
Set readPending to false when EOF is detected.
Result:
Fixes [#7255].
This reverts commit 413c7c2cd8 as it introduced an regression when edge-triggered mode is used which is true for our native transports by default. With 413c7c2cd8 included it was possible that we set readPending to false by mistake even if we would be interested in read more.
Motivation:
Linux kernel 4.11 introduced a new socket option,
TCP_FASTOPEN_CONNECT, that greatly simplifies making TCP Fast Open
connections on client side. Usually simply setting the flag before
connect() call is enough, no more changes are required.
Details can be found in kernel commit:
https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/commit/?id=19f6d3f3
Modifications:
TCP_FASTOPEN_CONNECT socket option was added to EpollChannelOption
class.
Result:
Netty clients can easily make TCP Fast Open connections. Simply
calling option(EpollChannelOption.TCP_FASTOPEN_CONNECT, true) in
client bootstrap is enough (given recent enough kernel).
Motivation:
readPending is currently only set to false if data is delivered to the application, however this may result in duplicate events being received from the selector in the event that the socket was closed.
Modifications:
- We should set readPending to false before each read attempt for all
transports besides NIO.
- Based upon the Javadocs it is possible that NIO may have spurious
wakeups [1]. In this case we should be more cautious and only set
readPending to false if data was actually read.
[1] https://docs.oracle.com/javase/7/docs/api/java/nio/channels/SelectionKey.html
That a selection key's ready set indicates that its channel is ready for some operation category is a hint, but not a guarantee, that an operation in such a category may be performed by a thread without causing the thread to block.
Result:
Notification from the selector (or simulated events from kqueue/epoll ET) in the event of socket closure.
Fixes https://github.com/netty/netty/issues/7255
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:
There are 2 motivations, the first depends on the second:
Loading Netty Epoll statically stopped working in 4.1.16, due to
`Native` always loading the arch specific shared object. In a
static binary, there is no arch specific SO.
Second, there are a ton of exceptions that can happen when loading
a native library. When loading native code, Netty tries a bunch of
different paths but a failure in any given may not be fatal.
Additionally: turning on debug logging is not always feasible so
exceptions get silently swallowed.
Modifications:
* Change Epoll and Kqueue to try the static load second
* Modify NativeLibraryLoader to record all the locations where
exceptions occur.
* Attempt to use `addSuppressed` from Java 7 if available.
Alternatives Considered:
An alternative would be to record log messages at each failure. If
all load attempts fail, the log messages are printed as warning,
else as debug. The problem with this is there is no `LogRecord` to
create like in java.util.logging. Buffering the args to
logger.log() at the end of the method loses the call site, and
changes the order of events to be confusing.
Another alternative is to teach NativeLibraryLoader about loading
the SO first, and then the static version. This would consolidate
the code fore Epoll, Kqueue, and TCNative. I think this is the
long term better option, but this PR is changing a lot already.
Someone else can take a crack at it later
Results:
Epoll Still Loads and easier debugging.
Motivation:
We used to check for version 6.8 but the latest is 6.9
Modifications:
Update version to 6.9 in the check.
Result:
Be able to cut a release on latest centos version
Motivation:
When SO_LINGER is used we run doClose() on the GlobalEventExecutor by default so we need to ensure we schedule all code that needs to be run on the EventLoop on the EventLoop in doClose. Beside this there are also threading issues when calling shutdownOutput(...)
Modifications:
- Schedule removal from EventLoop to the EventLoop
- Correctly handle shutdownOutput and shutdown in respect with threading-model
- Add unit tests
Result:
Fixes [#7159].
Motivation:
We need to ensure we use the correct osname in the Bundle-NativeCode declaration as declared in:
https://www.osgi.org/developer/specifications/reference/
Modifications:
Update osname to match the spec.
Result:
Correct Bundle-NativeCode entry in the MANIFEST
Motivation:
We should only try to load the native artifacts if the architecture we are currently running on is the same as the one the native libraries were compiled for.
Modifications:
Include architecture in native lib name and append the current arch when trying to load these. This will fail then if its not the same as the arch of the compiled arch.
Result:
Fixes [#7150].
Motivation:
PD and PD0 Both try to find and use Unsafe. If unavailable, they
try to log why and continue on. However, it is not always east to
enable this logging. Chaining exceptions together is much easier
to reach, and the original exception is relevant when Unsafe is
needed.
Modifications:
* Make PD log why PD0 could not be loaded with a trace level log
* Make PD0 remember why Unsafe wasn't available
* Expose unavailability cause through PD for higher level use.
* Make Epoll and KQueue include the reason when failing
Result:
Easier debugging in hard to reconfigure environments
Motivation:
If AutoClose is false and there is a IoException then AbstractChannel will not close the channel but instead just fail flushed element in the ChannelOutboundBuffer. AbstractChannel also notifies of writability changes, which may lead to an infinite loop if the peer has closed its read side of the socket because we will keep accepting more data but continuously fail because the peer isn't accepting writes.
Modifications:
- If the transport throws on a write we should acknowledge that the output side of the channel has been shutdown and cleanup. If the channel can't accept more data because it is full, and still healthy it is not expected to throw. However if the channel is not healthy it will throw and is not expected to accept any more writes. In this case we should shutdown the output for Channels that support this feature and otherwise just close.
- Connection-less protocols like UDP can remain the same because the channel may disconnected temporarily.
- Make sure AbstractUnsafe#shutdownOutput is called because the shutdown on the socket may throw an exception.
Result:
More correct handling of write failure when AutoClose is false.
Motivation:
As noticed in https://stackoverflow.com/questions/45700277/
compilation can fail if the definition of a method doesn't
match the declaration. It's easy enough to add this in, and make
it easy to compile.
Modifications:
Add JNIEXPORT to the entry points.
* On Windows this adds: `__declspec(dllexport)`
* On Mac this adds: `__attribute__((visibility("default")))`
* On Linux (GCC 4.2+) this adds: ` __attribute__((visibility("default")))`
* On other it doesn't add anything.
Result:
Easier compilation
Motivation:
KQueueEventLoop and EpollEventLoop implement different approaches to applying a timeout of their respective poll calls. Epoll attempts to ensure the desired timeout is satisfied at the java layer and at the JNI layer, but it should be sufficient to account for spurious wakups at the JNI layer. Epoll timeout granularity is also limited to milliseconds which may be too large for some latency sensitive applications.
Modifications:
- Make EpollEventLoop wait method look like KQueueEventLoop
- Epoll should support a finer timeout granularity via timerfd_create. We can hide most of these details behind the epollWait0 JNI call to avoid crossing additional JNI boundaries.
Result:
More consistent timeout approach between KQueue and Epoll.
Motivation:
The EPOLL transport uses EPOLLRDHUP to detect when the peer closes the write side of the socket. Currently KQueue is not able to mimic this behavior and the only way to detect if the peer has closed is to read. It may not always be appropriate to read for backpressure and other reasons at the application level.
Modifications:
- Support EVFILT_SOCK filter which provides notification when the peer closes the socket
Result:
KQueue transport has more consistent behavior with Epoll transport for detecting peer closure.
Motivation:
Due to an oversight (by myself), linking two JNI modules with
duplicate symbols fails in linking. This only seems to happen
some of the time (the behavior seems to be different between GCC
and Clang toolchains). For instance, including both netty tcnative
and netty epoll fails to link because of duplicate JNI_OnLoad
symobols.
Modification:
Do not define the JNI_OnLoad and JNI_OnUnload symbols when
compiling for static linkage, as indicated by the NETTY_BUILD_STATIC
preprocessor define. They are never directly called when
statically linked.
Result:
Able to statically compile epoll and tcnative code into a single
binary.
Motivation:
At the moment we try to load the library using multiple names which includes names using - but also _ . We should just use _ all the time.
Modifications:
Replace - with _
Result:
Fixes [#7069]
Motivation:
Implementations of DuplexChannel delegate the shutdownOutput to the underlying transport, but do not take any action on the ChannelOutboundBuffer. In the event of a write failure due to the underlying transport failing and application may attempt to shutdown the output and allow the read side the transport to finish and detect the close. However this may result in an issue where writes are failed, this generates a writability change, we continue to write more data, and this may lead to another writability change, and this loop may continue. Shutting down the output should fail all pending writes and not allow any future writes to avoid this scenario.
Modifications:
- Implementations of DuplexChannel should null out the ChannelOutboundBuffer and fail all pending writes
Result:
More controlled sequencing for shutting down the output side of a channel.
Motivation:
We did not correctly handle connect() and disconnect() in EpollDatagramChannel / KQueueDatagramChannel and so the behavior was different compared to NioDatagramChannel.
Modifications:
- Correct implement connect and disconnect methods
- Share connect and related code
- Add tests
Result:
EpollDatagramChannel / KQueueDatagramChannel also supports correctly connect() and disconnect() methods.
Motivation:
IP_TRANSPARENT support is not complete, the option can currently only be set on EpollServerSocket. Setting the option on an EpollSocket is also requires so as to be able to bind a socket to a non-local address as described in ip(7)
http://man7.org/linux/man-pages/man7/ip.7.html
"TProxy redirection with the iptables TPROXY target also
requires that this option be set on the redirected socket."
Modifications:
Added IP_TRANSPARENT socket option to EpollSocketChannelConfig
Result:
A redirecting socket can be created with a non-local IP address as required for TPROXY
Motivation:
We used an int[] to store all values that are returned in the struct for TCP_INFO which is not good enough as it uses usigned int values.
Modifications:
- Change int[] to long[] and correctly cast values.
Result:
No more truncated values.
Motivation:
`Epoll.ensureAvailability()` is called multiple times, once in
static initialization and in a couple of the constructors. This is
redundant and confusing to read.
Modifications:
Move `Epoll.ensureAvailability()` call into an instance initializer
and remove all other references. This ensures that every EELG
checks availability, while still delaying the check until
construction. This pattern is used when there are multiple ctors,
as in this class.
Result:
Easier to read code.
Motivation:
We should not try to detect a free port in tests put just use 0 when bind so there is no race in which the system my bind something to the port we choosen before.
Modifications:
- Remove the usage of TestUtils.getFreePort() in the testsuite
- Remove hack to workaround bind errors which will not happen anymore now
Result:
Less flacky tests.
Motivation:
When run the current testsuite on docker (mac) it will fail a few tests with:
io.netty.channel.AbstractChannel$AnnotatedConnectException: connect(..) failed: Cannot assign requested address: /0:0:0:0:0:0:0:0%0:46607
Caused by: java.net.ConnectException: connect(..) failed: Cannot assign requested address
Modifications:
Specify host explicit as done in other tests to only use ipv6 when really supported.
Result:
Build pass on docker as well
Motivation:
JCTools 2.0.2 provides an unbounded MPSC linked queue. Before we shaded JCTools we had our own unbounded MPSC linked queue and used it in various places but gave this up because there was no public equivalent available in JCTools at the time.
Modifications:
- Use JCTool's MPSC linked queue when no upper bound is specified
Result:
Fixes https://github.com/netty/netty/issues/5951
