keepalives tell any intermediate devices that the connection remains alive, which helps with overzealous firewalls that are
killing idle connections. keepalives are enabled by default in Elasticsearch, but use system defaults for their
configuration, which often times do not have reasonable defaults (e.g. 7200s for TCP_KEEP_IDLE) in the context of
distributed systems such as Elasticsearch.
This PR sets the socket-level keep_alive options for network.tcp.{keep_idle,keep_interval} to 5 minutes on configurations
that support it (>= Java 11 & (MacOS || Linux)) and where the system defaults are set to something higher than 5
minutes. This helps keep the connections alive while not interfering with system defaults or user-specified settings
unless they are deemed to be set too high by providing better out-of-the-box defaults.
* Remove eclipse conditionals
We used to have some meta projects with a `-test` prefix because
historically eclipse could not distinguish between test and main
source-sets and could only use a single classpath.
This is no longer the case for the past few Eclipse versions.
This PR adds the necessary configuration to correctly categorize source
folders and libraries.
With this change eclipse can import projects, and the visibility rules
are correct e.x. auto compete doesn't offer classes from test code or
`testCompile` dependencies when editing classes in `main`.
Unfortunately the cyclic dependency detection in Eclipse doesn't seem to
take the difference between test and non test source sets into account,
but since we are checking this in Gradle anyhow, it's safe to set to
`warning` in the settings. Unfortunately there is no setting to ignore
it.
This might cause problems when building since Eclipse will probably not
know the right order to build things in so more wirk might be necesarry.
Due to recent changes in the nio transport, a failure to bind the server
channel has started to be logged at an error level. This exception leads
to an automatic retry on a different port, so it should only be logged
at a trace level.
Uses JDK 11's per-socket configuration of TCP keepalive (supported on Linux and Mac), see
https://bugs.openjdk.java.net/browse/JDK-8194298, and exposes these as transport settings.
By default, these options are disabled for now (i.e. fall-back to OS behavior), but we would like
to explore whether we can enable them by default, in particular to force keepalive configurations
that are better tuned for running ES.
Currently in the transport-nio work we connect and bind channels on the
a thread before the channel is registered with a selector. Additionally,
it is at this point that we set all the socket options. This commit
moves these operations onto the event-loop after the channel has been
registered with a selector. It attempts to set the socket options for a
non-server channel at registration time. If that fails, it will attempt
to set the options after the channel is connected. This should fix
#41071.
Currently, we do not handle READ or WRITE events until the channel
connection process is complete. However, the external write queue path
allows a write to be attempted when the conneciton is not complete. This
commit closes the loophole and only queues write operations when the
connection process is not complete.
Registering a channel with a selector is a required operation for the
channel to be handled properly. Currently, we mix the registeration with
other setup operations (ip filtering, SSL initiation, etc). However, a
fail to register is fatal. This PR modifies how registeration occurs to
immediately close the channel if it fails.
There are still two clear loopholes for how a user can interact with a
channel even if registration fails. 1. through the exception handler.
2. through the channel accepted callback. These can perhaps be improved
in the future. For now, this PR prevents writes from proceeding if the
channel is not registered.
* Fix Exceptions in EventHandler#postHandling Breaking Select Loop
* We can run into the `write` path for SSL channels when they are not fully registered (if registration fails and a close message is attempted to be written) and thus into NPEs from missing selection keys
* This is a quick fix to quiet down tests, a cleaner solution will be incoming for #44343
* Relates #44343
* Remove Redundant Setting of OP_WRITE Interest
* We shouldn't have to set OP_WRITE interest before running into a partial write. Since setting OP_WRITE is handled by the `eventHandler.postHandling` logic, I think we can simply remove this operation and simplify/remove tests that were testing the setting of the write interest
* Use atomic boolean to guard wakeups
* Don't trigger wakeups from the select loops thread itself for registering and closing channels
* Don't needlessly queue writes
Co-authored-by: Tim Brooks <tim@uncontended.net>
Currently nio implements ip filtering at the channel context level. This
is kind of a hack as the application logic should be implemented at the
handler level. This commit moves the ip filtering into a channel
handler. This requires adding an indicator to the channel handler to
show when a channel should be closed.
Currently, when the SSLEngine needs to produce handshake or close data,
we must manually call the nonApplicationWrite method. However, this data
is only required when something triggers the need (starting handshake,
reading from the wire, initiating close, etc). As we have a dedicated
outbound buffer, this data can be produced automatically. Additionally,
with this refactoring, we combine handshake and application mode into a
single mode. This is necessary as there are non-application messages that
are sent post handshake in TLS 1.3. Finally, this commit modifies the
SSLDriver tests to test against TLS 1.3.
This is related to #27260. Currently we have a single read buffer that
is no larger than a single TLS packet. This prevents us from reading
multiple TLS packets in a single socket read call. This commit modifies
our TLS work to support reading similar to the plaintext case. The data
will be copied to a (potentially) recycled TLS packet-sized buffer for
interaction with the SSLEngine.
This is related to #27260. Currently there is a setting
http.read_timeout that allows users to define a read timeout for the
http transport. This commit implements support for this functionality
with the transport-nio plugin. The behavior here is that a repeating
task will be scheduled for the interval defined. If there have been
no requests received since the last run and there are no inflight
requests, the channel will be closed.
This is related to #27260. Currently for the SSLDriver we allocate a
dedicated network write buffer and encrypt the data into that buffer one
buffer at a time. This requires constantly switching between encrypting
and flushing. This commit adds a dedicated outbound buffer for SSL
operations that will internally allocate new packet sized buffers as
they are need (for writing encrypted data). This allows us to totally
encrypt an operation before writing it to the network. Eventually it can
be hooked up to buffer recycling.
This commit also backports the following commit:
Handle WRAP ops during SSL read
It is possible that a WRAP operation can occur while decrypting
handshake data in TLS 1.3. The SSLDriver does not currently handle this
well as it does not have access to the outbound buffer during read call.
This commit moves the buffer into the Driver to fix this issue. Data
wrapped during a read call will be queued for writing after the read
call is complete.
This is related to #27260. Currently for the SSLDriver we allocate a
dedicated network write buffer and encrypt the data into that buffer one
buffer at a time. This requires constantly switching between encrypting
and flushing. This commit adds a dedicated outbound buffer for SSL
operations that will internally allocate new packet sized buffers as
they are need (for writing encrypted data). This allows us to totally
encrypt an operation before writing it to the network. Eventually it can
be hooked up to buffer recycling.
Recently we have had a number of test issues related to blocking
activity occuring on the io thread. This commit adds a log warning for
when handling event takes a >150 milliseconds. This is implemented
for the MockNioTransport which is the transport used in
ESIntegTestCase.
* A few warnings could be observed in test logs about `NoSuchElementException` being thrown in `InboundChannelBuffer#sliceBuffersTo`.
These were the result of calls to this method after the relevant channel and hence the buffer was closed already as a result of a failed IO operation.
* Fixed by adding the necessary guard statements to break out in these cases. I don't think there is a need here to do any additional error handling since `eventHandler.postHandling(channelContext);` at the end of the `processKey`
call in the main selection loop handles closing channels and invoking callbacks for writes that failed to go through already.
This commit introduces a NetworkMessage class. This class has two
subclasses - InboundMessage and OutboundMessage. These messages can
be serialized and deserialized independent of the transport. This allows
more granular testing. Additionally, the serialization mechanism is now
a simple Supplier. This builds the framework to eventually move the
serialization of transport messages to the network thread. This is the
one serialization component that is not currently performed on the
network thread (transport deserialization and http serialization and
deserialization are all on the network thread).
Currently we create dedicated network threads for both the http and
transport implementations. Since these these threads should never
perform blocking operations, these threads could be shared. This commit
modifies the nio-transport to have 0 http workers be default. If the
default configs are used, this will cause the http transport to be run
on the transport worker threads. The http worker setting will still exist
in case the user would like to configure dedicated workers. Additionally,
this commmit deletes dedicated acceptor threads. We have never had these
for the netty transport and they can be added back if a need is
determined in the future.
Currently we read and write 64KB at a time in the nio libraries. As a
single byte buffer per event loop thread does not consume much memory,
there is little reason to not increase it further. This commit increases
the buffer to 256KB but still limits a single write to 64KB. The write
limit could be increased, but too high of a write limit will lead to
copying more data (if all the data is not flushed and needs to be copied
on the next call). This is something to explore in the future.
Closing a channel using TLS/SSL requires reading and writing a
CLOSE_NOTIFY message (for pre-1.3 TLS versions). Many implementations do
not actually send the CLOSE_NOTIFY message, which means we are depending
on the TCP close from the other side to ensure channels are closed. In
case there is an issue with this, we need a timeout. This commit adds a
timeout to the channel close process for TLS secured channels.
As part of this change, we need a timer service. We could use the
generic Elasticsearch timeout threadpool. However, it would be nice to
have a local to the nio event loop timer service dedicated to network needs. In
the future this service could support read timeouts, connect timeouts,
request timeouts, etc. This commit adds a basic priority queue backed
service. Since our timeout volume (channel closes) is very low, this
should be fine. However, this can be updated to something more efficient
in the future if needed (timer wheel). Everything being local to the event loop
thread makes the logic simple as no locking or synchronization is necessary.
`PageCacheRecycler` is the class that creates and holds pages of arrays
for various uses. `BigArrays` is just one user of these pages. This
commit moves the constants that define the page sizes for the recycler
to be on the recycler class.
This is related to #27260. In Elasticsearch all of the messages that we
serialize to write to the network are composed of heap bytes. When you
read or write to a nio socket in java, the heap memory you passed down
must be copied to/from direct memory. The JVM internally does some
buffering of the direct memory, however it is essentially unbounded.
This commit introduces a simple mechanism of buffering and copying the
memory in transport-nio. Each network event loop is given a 64kb
DirectByteBuffer. When we go to read we use this buffer and copy the
data after the read. Additionally, when we go to write, we copy the data
to the direct memory before calling write. 64KB is chosen as this is the
default receive buffer size we use for transport-netty4
(NETTY_RECEIVE_PREDICTOR_SIZE).
Since we only have one buffer per thread, we could afford larger.
However, if we the buffer is large and not all of the data is flushed in
a write call, we will do excess copies. This is something we can
explore in the future.
This commit is related to #32517. It allows an "sni_server_name"
attribute on a DiscoveryNode to be propagated to the server using
the TLS SNI extentsion. Prior to this commit, this functionality
was only support for the netty transport. This commit adds this
functionality to the security nio transport.
This change cleans up "unused variable" warnings. There are several cases were we
most likely want to suppress the warnings (especially in the client documentation test
where the snippets contain many unused variables). In a lot of cases the unused
variables can just be deleted though.
This is related to #27260. It adds the SecurityNioTransport to the
security plugin. Additionally, it adds support for ip filtering. And it
randomly uses the nio transport in security integration tests.
Currently we set local addresses on the creation time of a NioChannel.
However, this may return null as the local address may not have been
set yet. An example is the local address has not been set on a client
channel as the connection process is not yet complete.
This PR modifies the getter to set the local field if it is currently null.
Currently, when we open a new channel, we pass it an
InboundChannelBuffer. The channel buffer is preallocated a single 16kb
page. However, there is no guarantee that this channel will be read from
anytime soon. Instead, this commit does not preallocate that page. That
page will be allocated when we receive a read event.
This is related to #28898. This PR implements pooling of bytes arrays
when reading from the wire in the http server transport. In order to do
this, we must integrate with netty reference counting. That manner in
which this PR implements this is making Pages in InboundChannelBuffer
reference counted. When we accessing the underlying page to pass to
netty, we retain the page. When netty releases its bytebuf, it releases
the underlying pages we have passed to it.
This is related to #27260. Currently when we queue a write with a
channel we set OP_WRITE and wait until the next selection loop to flush
the write. However, if the channel does not have a pending write, it
is probably ready to flush. This PR implements an optimistic flush logic
that will attempt this flush.
Yannick Welsch
Alpar Torok
Tim Brooks
Armin Braun
Albert Zaharovits
Christoph Büscher
Tanguy Leroux
Jason Tedor