Motivation:
Sometimes it's useful to be able to create a Epoll*Channel from an existing file descriptor. This is especially helpful if you integrade some c/jni code.
Modifications:
- Add extra constructor to Epoll*Channel implementations that take a FileDescriptor as an argument
- Make Rename EpollFileDescriptor to NativeFileDescriptor and make it public
- Also ensure we obtain the correct remote/local address when create a Channel from a FileDescriptor
Result:
It's now possible to create a FileDescriptor and instance a Epoll*Channel via it.
Motivation:
If SO_LINGER is used shutdownOutput() and close() syscalls will block until either all data was send or until the timeout exceed. This is a problem when we try to execute them on the EventLoop as this means the EventLoop may be blocked and so can not process any other I/O.
Modifications:
- Add AbstractUnsafe.closeExecutor() which returns null by default and use this Executor for close if not null.
- Override the closeExecutor() in NioSocketChannel and EpollSocketChannel and return GlobalEventExecutor.INSTANCE if getSoLinger() > 0
- use closeExecutor() in shutdownInput(...) in NioSocketChannel and EpollSocketChannel
Result:
No more blocking of the EventLoop if SO_LINGER is used and shutdownOutput() or close() is called.
Motivation:
Some of the methods are frequently called and so should be inlined if possible.
Modifications:
Give the compiler a hint that we want to inline these methods.
Result:
Better performance if inlined.
Motivation:
Older linux kernels have problems handling a large value for epoll_wait(...) and so wait for ever.
Modifications:
Adjust timeout on the fly if a too big value is passed in.
Result:
Correctly works also on older kernels.
Motivation:
The writeSpinCount was ignored in the epoll transport and it just kept on trying writing. This could cause unnessary cpu spinning if a slow remote peer was reading the data very very slow.
Modification:
- Correctly take writeSpinCount into account when writing.
Result:
Less cpu spinning when writing to a slow remote peer.
Motivation:
Fix regression introduced by 585ce1593f, which missed to set EPOLLRDHUP for all stream channels.
Modifications:
Correctly set EPOLLRDHUP for all stream channels in the AbstractEpollStreamChannel constructor.
Result:
No more test failures in EpollDomain*Channel tests.
Motivation:
Before we used a long[] to store the ready events, this had a few problems and limitations:
- An extra loop was needed to translate between epoll_event and our long
- JNI may need to do extra memory copy if the JVM not supports pinning
- More branches
Modifications:
- Introduce a EpollEventArray which allows to directly write in a struct epoll_event* and pass it to epoll_wait.
Result:
Better speed when using native transport, as shown in the benchmark.
Before:
[xxx@xxx wrk]$ ./wrk -H 'Connection: keep-alive' -d 120 -c 256 -t 16 -s scripts/pipeline-many.lua http://xxx:8080/plaintext
Running 2m test @ http://xxx:8080/plaintext
16 threads and 256 connections
Thread Stats Avg Stdev Max +/- Stdev
Latency 14.56ms 8.64ms 117.15ms 80.58%
Req/Sec 286.17k 38.71k 421.48k 68.17%
546324329 requests in 2.00m, 73.78GB read
Requests/sec: 4553438.39
Transfer/sec: 629.66MB
After:
[xxx@xxx wrk]$ ./wrk -H 'Connection: keep-alive' -d 120 -c 256 -t 16 -s scripts/pipeline-many.lua http://xxx:8080/plaintext
Running 2m test @ http://xxx:8080/plaintext
16 threads and 256 connections
Thread Stats Avg Stdev Max +/- Stdev
Latency 14.12ms 8.69ms 100.40ms 83.08%
Req/Sec 294.79k 40.23k 472.70k 66.75%
555997226 requests in 2.00m, 75.08GB read
Requests/sec: 4634343.40
Transfer/sec: 640.85MB
Motivation:
Netty uses edge-triggered epoll by default for performance reasons. The downside here is that a messagesPerRead limit can not be enforced correctly, as we need to consume everything from the channel when notified.
Modification:
- Allow to switch epoll modes before channel is registered
- Some refactoring to share more code
Result:
It's now possible to switch epoll mode.
Motiviation:
When using domain sockets on linux it is supported to recv and send file descriptors. This can be used to pass around for example sockets.
Modifications:
- Add support for recv and send file descriptors when using EpollDomainSocketChannel.
- Allow to obtain the file descriptor for an Epoll*Channel so it can be send via domain sockets.
Result:
recv and send of file descriptors is supported now.
Motivation:
Using Unix Domain Sockets can be very useful when communication should take place on the same host and has less overhead then using loopback. We should support this with the native epoll transport.
Modifications:
- Add support for Unix Domain Sockets.
- Adjust testsuite to be able to reuse tests.
Result:
Unix Domain Sockets are now support when using native epoll transport.
Motivation:
At the moment the max number of events that can be handled per epoll wakup was set during construction.
Modifications:
- Automatically increase the max number of events to handle
Result:
Better performance when a lot of events need to be handled without adjusting the code.
Motivation:
The current way how the guard against overflow when generating the nextId() is pretty slow once an overflow happened.
Modifications:
Once a possible overflow is detected all ids used by the EpollEventLoop are scrubed and re-assigned to the registered Channels. This way we only need to do extra work each time an overflow is detected.
Result:
More consistent performance even after the first overflow was detected.
Motivation:
On Linux, you can gather various metrics using getsockopt(..., TCP_INFO,
...).
Modifications:
Add EpollSocketChannel.tcpInfo() which returns EpollTcpInfo that exposes
all metrics exposed via getsockopt(..., TCP_INFO, ...)
Result:
TCP_INFO support implemented
Motivation:
In the native transport we use getpeername to obtain the remote address from the file descriptor. This may fail for various reasons in which case NULL is returned.
Modifications:
- Check for null when try to obtain remote / local address
Result:
No more NPE
Related: #3274
Motivation:
channelReadComplete() event is not triggered after reading successfully
in EpollDatagramChannel.
Modifications:
- Trigger exceptionCaught() event for read failure only once for less
noise
- Trigger channelReadComplete() event at the end of the read.
Result:
Fix#3274
Rebased and cleaned-up based on the work by @normanmaurer
Motivation:
Currently, IOExceptions and ClosedChannelExceptions are thrown from
inside the JNI methods. Instantiation of Java objects inside JNI code is
an expensive operation, needless to say about filling stack trace for
every instantiation of an exception.
Modifications:
Change most JNI methods to return a negative value on failure so that
the exceptions are instantiated outside the native code.
Also, pre-instantiate some commonly-thrown exceptions for better
performance.
Result:
Performance gain
Motivation:
So far, we generated and deployed test JARs to Maven repositories. The
deployed JAR had the classifier 'test-jar'. The test JAR is consumed by
transport-native-epoll as a test dependency.
The problem is, when netty-transport-native-epoll pulls the test JAR as
a dependency, that Maven resolves its transitive dependencies at
'compile' and 'runtime' scope only, which is incorrect.
I was bitten by this problem recently while trying to add a new
dependency to netty-testsuite. Because I added a new dependency at the
'test' scope, the new dependency was not pulled transitively by
transport-native-epoll and caused an unexpected build failure.
- d6160208c3
- bf77bb4c3a
Modifications:
- Move all classes in netty-testsuite from src/test to src/main
- Update the 'compile' scope dependencies of netty-testsuite
- Override the test directory configuration properties of the surefire
plugin
- Do not generate the test JAR anymore
- Update the dependency of netty-transport-native-epoll
Result:
It is less error-prone to add a new dependency to netty-testsuite.
Motivation:
Everytime a new connection is accepted via EpollSocketServerChannel it will create a new EpollSocketChannel that needs to get the remote and local addresses in the constructor. The current implementation uses new InetSocketAddress(String, int) to create these. This is quite slow due the implementation in oracle and openjdk.
Modifications:
Encode all needed informations into a byte array before return from jni layer and then use new InetSocketAddress(InetAddress, int) to create the socket addresses. This allows to create the InetAddress via a byte[] and so reduce the overhead, this is done either by using InetAddress.getByteAddress(byte[]) or by Inet6Address.getByteAddress(String, byte[], int).
Result:
Reduce performance overhead while accept new connections with native transport
Motivation:
So far, our TLS renegotiation test did not test changing cipher suite
during renegotiation explicitly.
Modifications:
- Switch the cipher suite during renegotiation
Result:
We are now sure the cipher suite change works.
Motivation:
We only provided a constructor in DefaultFileRegion that takes a FileChannel which means the File itself needs to get opened on construction. This has the problem that if you want to write a lot of Files very fast you may end up with may open FD's even if they are not needed yet. This can lead to hit the open FD limit of the OS.
Modifications:
Add a new constructor to DefaultFileRegion which allows to construct it from a File. The FileChannel will only be obtained when transferTo(...) is called or the DefaultFileRegion is explicit open'ed via open() (this is needed for the native epoll transport)
Result:
Less resource usage when writing a lot of DefaultFileRegion.
Related: #3125
Motivation:
We did not expose a way to initiate TLS renegotiation and to get
notified when the renegotiation is done.
Modifications:
- Add SslHandler.renegotiate() so that a user can initiate TLS
renegotiation and get the future that's notified on completion
- Make SslHandler.handshakeFuture() return the future for the most
recent handshake so that a user can get the future of the last
renegotiation
- Add the test for renegotiation to SocketSslEchoTest
Result:
Both client-initiated and server-initiated renegotiations are now
supported properly.
Motivation:
JDK's exception messages triggered by a connection attempt failure do
not contain the related remote address in its message. We currently
append the remote address to ConnectException's message, but I found
that we need to cover more exception types such as SocketException.
Modifications:
- Add AbstractUnsafe.annotateConnectException() to de-duplicate the
code that appends the remote address
Result:
- Less duplication
- A transport implementor can annotate connection attempt failure
message more easily
Motivation:
We use malloc(1) in the on JNI_OnLoad method but never free the allocated memory. This means we have a tiny memory leak of 1 byte.
Modifications:
Call free(...) on previous allocated memory.
Result:
Fix memory leak
Motiviation:
If sendmmsg is already defined then the native epoll module failed to build because of conflicting definitions.
The mmsghdr type was also redefined on systems that already supported this structure.
Modifications:
Provide a way so that systems which already define sendmmsg and mmsghdr can build
Provide a way so that systems which don't define sendmmsg and mmsghdr can build
Result:
The native EPOLL module can build in more environments
Motivation:
In linux it is possible to write more then one buffer withone syscall when sending datagram messages.
Modifications:
Not copy CompositeByteBuf if it only contains direct buffers.
Result:
More performance due less overhead for copy.
Motivation:
On linux with glibc >= 2.14 it is possible to send multiple DatagramPackets with one syscall. This can be a huge performance win and so we should support it in our native transport.
Modification:
- Add support for sendmmsg by reuse IovArray
- Factor out ThreadLocal support of IovArray to IovArrayThreadLocal for better separation as we use IovArray also without ThreadLocal in NativeDatagramPacketArray now
- Introduce NativeDatagramPacketArray which is used for sendmmsg(...)
- Implement sendmmsg(...) via jni
- Expand DatagramUnicastTest to test also sendmmsg(...)
Result:
Netty now automatically use sendmmsg(...) if it is supported and we have more then 1 DatagramPacket in the ChannelOutboundBuffer and flush() is called.
Motivation:
On linux it is possible to use the sendMsg(...) system call to write multiple buffers with one system call when using datagram/udp.
Modifications:
- Implement the needed changes and make use of sendMsg(...) if possible for max performance
- Add tests that test sending datagram packets with all kind of different ByteBuf implementations.
Result:
Performance improvement when using CompoisteByteBuf and EpollDatagramChannel.
Motivation:
InetAddress.getByName(...) uses exceptions for control flow when try to parse IPv4-mapped-on-IPv6 addresses. This is quite expensive.
Modifications:
Detect IPv4-mapped-on-IPv6 addresses in the JNI level and convert to IPv4 addresses before pass to InetAddress.getByName(...) (via InetSocketAddress constructor).
Result:
Eliminate performance problem causes by exception creation when parsing IPv4-mapped-on-IPv6 addresses.
Motivation:
In EpollSocketchannel.doWriteFileRegion(...) we need to make sure we write until sendFile(...) returns either 0 or all is written. Otherwise we may not get notified once the Channel is writable again.
This is the case as we use EPOLL_ET.
Modifications:
Always write until either sendFile returns 0 or all is written.
Result:
No more hangs when writing DefaultFileRegion can happen.
Motivation:
There were no way to efficient write a CompositeByteBuf as we always did a memory copy to a direct buffer in this case. This is not needed as we can just write a CompositeByteBuf as long as all the components are buffers with a memory address.
Modifications:
- Write CompositeByteBuf which contains only direct buffers without memory copy
- Also handle CompositeByteBuf that have more components then 1024.
Result:
More efficient writing of CompositeByteBuf.
Related issue: #2764
Motivation:
EpollSocketChannel.writeFileRegion() does not handle the case where the
position of a FileRegion is non-zero properly.
Modifications:
- Improve SocketFileRegionTest so that it tests the cases where the file
transfer begins from the middle of the file
- Add another jlong parameter named 'base_off' so that we can take the
position of a FileRegion into account
Result:
Improved test passes. Corruption is gone.
Motivation:
While trying to merge our ChannelOutboundBuffer changes we've made last
week, I realized that we have quite a bit of conflicting changes at 4.1
and master. It was primarily because we added
ChannelOutboundBuffer.beforeAdd() and moved some logic there, such as
direct buffer conversion.
However, this is not possible with the changes we've made for 4.0. We
made ChannelOutboundBuffer final for example.
Maintaining multiple branch is already getting painful and having
different core will make it even worse, so I think we should keep the
differences between 4.0 and other branches minimal.
Modifications:
- Move ChannelOutboundBuffer.safeRelease() to ReferenceCountUtil
- Add ByteBufUtil.threadLocalBuffer()
- Backported from ThreadLocalPooledDirectByteBuf
- Make most methods in AbstractUnsafe final
- Add AbstractChannel.filterOutboundMessage() so that a transport can
convert a message to another (e.g. heap -> off-heap), and also
reject unsupported messages
- Move all direct buffer conversions to filterOutboundMessage()
- Move all type checks to filterOutboundMessage()
- Move AbstractChannel.checkEOF() to OioByteStreamChannel, because it's
the only place it is used at all
- Remove ChannelOutboundBuffer.current(Object), because it's not used
anymore
- Add protected direct buffer conversion methods to AbstractNioChannel
and AbstractEpollChannel so that they can be used by their subtypes
- Update all transport implementations according to the changes above
Result:
- The missing extension point in 4.0 has been added.
- AbstractChannel.filterOutboundMessage()
- Thanks to the new extension point, we moved all transport-specific
logic from ChannelOutboundBuffer to each transport implementation
- We can copy most of the transport implementations in 4.0 to 4.1 and
master now, so that we have much less merge conflict when we modify
the core.
Related issue: #2733
Motivation:
Unlike OpenSsl, Epoll lacks a couple useful availability checker
methods:
- ensureAvailability()
- unavailabilityCause()
Modifications:
Add missing methods
Result:
More ways to check the availability and to get the cause of
unavailability programatically.
Motivation:
When a ChannelOutboundBuffer contains a series of entries whose messages
are all empty buffers, EpollSocketChannel sometimes fails to remove
them. As a result, the result of the write(EmptyByteBuf) is never
notified, making the user application hang.
Modifications:
- Add ChannelOutboundBuffer.removeBytes(long) method that updates the
progress of the entries and removes them as much as the specified
number of written bytes. It also updates the reader index of
partially flushed buffer.
- Make both NioSocketChannel and EpollSocketChannel use it to reduce
code duplication
- Replace EpollSocketChannel.updateOutboundBuffer()
- Refactor EpollSocketChannel.doWrite() for simplicity
- Split doWrite() into doWriteSingle() and doWriteMultiple()
- Do not add a zero-length buffer to IovArray
- Do not perform any real I/O when the size of IovArray is 0
Result:
Another regression is gone.
- ChannelOutboundBuffer.Entry.buffers -> bufs for consistency
- Make Native.IOV_MAX final because it's a constant
- Naming changes
- FlushedMessageProcessor -> MessageProcessor just in case we can
reuse it for unflushed messages in the future
- Add ChannelOutboundBuffer.Entry.recycle() that does not return the
next entry, and use it wherever possible
- Javadoc clean-up
Motivation:
While benchmarking the native transport, I noticed that gathering write
is not as fast as expected. It was due to the fact that we have to do a
lot of array copies to put the buffer addresses into the iovec struct
array.
Modifications:
Introduce a new class called IovArray, which allows to fill buffers
directly into an off-heap array of iovec structs, so that it can be
passed over to JNI without any extra array copies.
Result:
Big performance improvement when doing gathering writes:
Before:
[nmaurer@xxx]~% wrk/wrk -H 'Host: localhost' -H 'Accept: text/html,application/xhtml+xml,application/xml;q=0.9,*/*;q=0.8' -H 'Connection: keep-alive' -d 120 -c 256 -t 16 --pipeline 256 http://xxx:8080/plaintext
Running 2m test @ http://xxx:8080/plaintext
16 threads and 256 connections
Thread Stats Avg Stdev Max +/- Stdev
Latency 23.44ms 16.37ms 259.57ms 91.77%
Req/Sec 181.99k 31.69k 304.60k 78.12%
346544071 requests in 2.00m, 46.48GB read
Requests/sec: 2887885.09
Transfer/sec: 396.59MB
After:
[nmaurer@xxx]~% wrk/wrk -H 'Host: localhost' -H 'Accept: text/html,application/xhtml+xml,application/xml;q=0.9,*/*;q=0.8' -H 'Connection: keep-alive' -d 120 -c 256 -t 16 --pipeline 256 http://xxx:8080/plaintext
Running 2m test @ http://xxx:8080/plaintext
16 threads and 256 connections
Thread Stats Avg Stdev Max +/- Stdev
Latency 21.93ms 16.33ms 305.73ms 92.34%
Req/Sec 194.56k 33.75k 309.33k 77.04%
369617503 requests in 2.00m, 49.57GB read
Requests/sec: 3080169.65
Transfer/sec: 423.00MB
Motivation:
73dfd7c01b introduced various test
failures because:
- EpollSocketChannel.doWrite() raised a NullPointerException when
notifying the write progress.
- ChannelOutboundBuffer.nioBuffers() did not expand the internal array
when the pending entries contained more than 1024 buffers, dropping
the remainder.
Modifications:
- Fix the NPE in EpollSocketChannel by removing an unnecessary progress
update
- Expand the thread-local buffer array if there is not enough room,
which was the original behavior dropped by the offending commit
Result:
Regression is gone.
Motivation:
We sometimes not use the correct exception message when throw it from the native code.
Modifications:
Fixed the message.
Result:
Correct message in exception
