org.apache.cassandra.net.OutboundConnection Maven / Gradle / Ivy
Go to download
Show more of this group Show more artifacts with this name
Show all versions of cassandra-all Show documentation
Show all versions of cassandra-all Show documentation
The Apache Cassandra Project develops a highly scalable second-generation distributed database, bringing together Dynamo's fully distributed design and Bigtable's ColumnFamily-based data model.
/*
* Licensed to the Apache Software Foundation (ASF) under one
* or more contributor license agreements. See the NOTICE file
* distributed with this work for additional information
* regarding copyright ownership. The ASF licenses this file
* to you under the Apache License, Version 2.0 (the
* "License"); you may not use this file except in compliance
* with the License. You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
*/
package org.apache.cassandra.net;
import java.io.IOException;
import java.net.ConnectException;
import java.net.InetSocketAddress;
import java.nio.channels.ClosedChannelException;
import java.util.Objects;
import java.util.concurrent.CountDownLatch;
import java.util.concurrent.ExecutorService;
import java.util.concurrent.TimeUnit;
import java.util.concurrent.atomic.AtomicInteger;
import java.util.concurrent.atomic.AtomicLongFieldUpdater;
import java.util.concurrent.atomic.AtomicReference;
import java.util.concurrent.atomic.AtomicReferenceFieldUpdater;
import javax.annotation.Nullable;
import com.google.common.annotations.VisibleForTesting;
import com.google.common.util.concurrent.Uninterruptibles;
import org.slf4j.Logger;
import org.slf4j.LoggerFactory;
import io.netty.channel.Channel;
import io.netty.channel.ChannelFuture;
import io.netty.channel.ChannelHandlerContext;
import io.netty.channel.ChannelInboundHandlerAdapter;
import io.netty.channel.EventLoop;
import io.netty.channel.unix.Errors;
import io.netty.util.concurrent.Future;
import io.netty.util.concurrent.Promise;
import io.netty.util.concurrent.PromiseNotifier;
import io.netty.util.concurrent.SucceededFuture;
import org.apache.cassandra.config.DatabaseDescriptor;
import org.apache.cassandra.io.util.DataOutputBufferFixed;
import org.apache.cassandra.net.OutboundConnectionInitiator.Result.MessagingSuccess;
import org.apache.cassandra.tracing.Tracing;
import org.apache.cassandra.utils.FBUtilities;
import org.apache.cassandra.utils.JVMStabilityInspector;
import org.apache.cassandra.utils.NoSpamLogger;
import static java.lang.Math.max;
import static java.lang.Math.min;
import static java.util.concurrent.TimeUnit.MILLISECONDS;
import static org.apache.cassandra.net.MessagingService.current_version;
import static org.apache.cassandra.net.OutboundConnectionInitiator.*;
import static org.apache.cassandra.net.OutboundConnections.LARGE_MESSAGE_THRESHOLD;
import static org.apache.cassandra.net.ResourceLimits.*;
import static org.apache.cassandra.net.ResourceLimits.Outcome.*;
import static org.apache.cassandra.net.SocketFactory.*;
import static org.apache.cassandra.utils.FBUtilities.prettyPrintMemory;
import static org.apache.cassandra.utils.MonotonicClock.approxTime;
import static org.apache.cassandra.utils.Throwables.isCausedBy;
/**
* Represents a connection type to a peer, and handles the state transistions on the connection and the netty {@link Channel}.
* The underlying socket is not opened until explicitly requested (by sending a message).
*
* TODO: complete this description
*
* Aside from a few administrative methods, the main entry point to sending a message is {@link #enqueue(Message)}.
* Any thread may send a message (enqueueing it to {@link #queue}), but only one thread may consume messages from this
* queue. There is a single delivery thread - either the event loop, or a companion thread - that has logical ownership
* of the queue, but other threads may temporarily take ownership in order to perform book keeping, pruning, etc.,
* to ensure system stability.
*
* {@link Delivery#run()} is the main entry point for consuming messages from the queue, and executes either on the event
* loop or on a non-dedicated companion thread. This processing is activated via {@link Delivery#execute()}.
*
* Almost all internal state maintenance on this class occurs on the eventLoop, a single threaded executor which is
* assigned in the constructor. Further details are outlined below in the class. Some behaviours require coordination
* between the eventLoop and the companion thread (if any). Some minimal set of behaviours are permitted to occur on
* producers to ensure the connection remains healthy and does not overcommit resources.
*
* All methods are safe to invoke from any thread unless otherwise stated.
*/
@SuppressWarnings({ "WeakerAccess", "FieldMayBeFinal", "NonAtomicOperationOnVolatileField", "SameParameterValue" })
public class OutboundConnection
{
static final Logger logger = LoggerFactory.getLogger(OutboundConnection.class);
private static final NoSpamLogger noSpamLogger = NoSpamLogger.getLogger(logger, 30L, TimeUnit.SECONDS);
private static final AtomicLongFieldUpdater submittedUpdater = AtomicLongFieldUpdater.newUpdater(OutboundConnection.class, "submittedCount");
private static final AtomicLongFieldUpdater pendingCountAndBytesUpdater = AtomicLongFieldUpdater.newUpdater(OutboundConnection.class, "pendingCountAndBytes");
private static final AtomicLongFieldUpdater overloadedCountUpdater = AtomicLongFieldUpdater.newUpdater(OutboundConnection.class, "overloadedCount");
private static final AtomicLongFieldUpdater overloadedBytesUpdater = AtomicLongFieldUpdater.newUpdater(OutboundConnection.class, "overloadedBytes");
private static final AtomicReferenceFieldUpdater closingUpdater = AtomicReferenceFieldUpdater.newUpdater(OutboundConnection.class, Future.class, "closing");
private static final AtomicReferenceFieldUpdater scheduledCloseUpdater = AtomicReferenceFieldUpdater.newUpdater(OutboundConnection.class, Future.class, "scheduledClose");
private final EventLoop eventLoop;
private final Delivery delivery;
private final OutboundMessageCallbacks callbacks;
private final OutboundDebugCallbacks debug;
@VisibleForTesting
final OutboundMessageQueue queue;
/** the number of bytes we permit to queue to the network without acquiring any shared resource permits */
private final long pendingCapacityInBytes;
/** the number of messages and bytes queued for flush to the network,
* including those that are being flushed but have not been completed,
* packed into a long (top 20 bits for count, bottom 42 for bytes)*/
private volatile long pendingCountAndBytes = 0;
/** global shared limits that we use only if our local limits are exhausted;
* we allocate from here whenever queueSize > queueCapacity */
private final EndpointAndGlobal reserveCapacityInBytes;
/** Used in logging statements to lazily build a human-readable number of pending bytes. */
private final Object readablePendingBytes =
new Object() { @Override public String toString() { return prettyPrintMemory(pendingBytes()); } };
/** Used in logging statements to lazily build a human-readable number of reserve endpoint bytes in use. */
private final Object readableReserveEndpointUsing =
new Object() { @Override public String toString() { return prettyPrintMemory(reserveCapacityInBytes.endpoint.using()); } };
/** Used in logging statements to lazily build a human-readable number of reserve global bytes in use. */
private final Object readableReserveGlobalUsing =
new Object() { @Override public String toString() { return prettyPrintMemory(reserveCapacityInBytes.global.using()); } };
private volatile long submittedCount = 0; // updated with cas
private volatile long overloadedCount = 0; // updated with cas
private volatile long overloadedBytes = 0; // updated with cas
private long expiredCount = 0; // updated with queue lock held
private long expiredBytes = 0; // updated with queue lock held
private long errorCount = 0; // updated only by delivery thread
private long errorBytes = 0; // updated by delivery thread only
private long sentCount; // updated by delivery thread only
private long sentBytes; // updated by delivery thread only
private long successfulConnections; // updated by event loop only
private long connectionAttempts; // updated by event loop only
private static final int pendingByteBits = 42;
private static boolean isMaxPendingCount(long pendingCountAndBytes)
{
return (pendingCountAndBytes & (-1L << pendingByteBits)) == (-1L << pendingByteBits);
}
private static int pendingCount(long pendingCountAndBytes)
{
return (int) (pendingCountAndBytes >>> pendingByteBits);
}
private static long pendingBytes(long pendingCountAndBytes)
{
return pendingCountAndBytes & (-1L >>> (64 - pendingByteBits));
}
private static long pendingCountAndBytes(long pendingCount, long pendingBytes)
{
return (pendingCount << pendingByteBits) | pendingBytes;
}
private final ConnectionType type;
/**
* Contains the base settings for this connection, _including_ any defaults filled in.
*
*/
private OutboundConnectionSettings template;
private static class State
{
static final State CLOSED = new State(Kind.CLOSED);
enum Kind { ESTABLISHED, CONNECTING, DORMANT, CLOSED }
final Kind kind;
State(Kind kind)
{
this.kind = kind;
}
boolean isEstablished() { return kind == Kind.ESTABLISHED; }
boolean isConnecting() { return kind == Kind.CONNECTING; }
boolean isDisconnected() { return kind == Kind.CONNECTING || kind == Kind.DORMANT; }
boolean isClosed() { return kind == Kind.CLOSED; }
Established established() { return (Established) this; }
Connecting connecting() { return (Connecting) this; }
Disconnected disconnected() { return (Disconnected) this; }
}
/**
* We have successfully negotiated a channel, and believe it to still be valid.
*
* Before using this, we should check isConnected() to check the Channel hasn't
* become invalid.
*/
private static class Established extends State
{
final int messagingVersion;
final Channel channel;
final FrameEncoder.PayloadAllocator payloadAllocator;
final OutboundConnectionSettings settings;
Established(int messagingVersion, Channel channel, FrameEncoder.PayloadAllocator payloadAllocator, OutboundConnectionSettings settings)
{
super(Kind.ESTABLISHED);
this.messagingVersion = messagingVersion;
this.channel = channel;
this.payloadAllocator = payloadAllocator;
this.settings = settings;
}
boolean isConnected() { return channel.isOpen(); }
}
private static class Disconnected extends State
{
/** Periodic message expiry scheduled while we are disconnected; this will be cancelled and cleared each time we connect */
final Future maintenance;
Disconnected(Kind kind, Future maintenance)
{
super(kind);
this.maintenance = maintenance;
}
public static Disconnected dormant(Future maintenance)
{
return new Disconnected(Kind.DORMANT, maintenance);
}
}
private static class Connecting extends Disconnected
{
/**
* Currently (or scheduled to) (re)connect; this may be cancelled (if closing) or waited on (for delivery)
*
* - The work managed by this future is partially performed asynchronously, not necessarily on the eventLoop.
* - It is only completed on the eventLoop
* - It may not be executing, but might be scheduled to be submitted if {@link #scheduled} is not null
*/
final Future> attempt;
/**
* If we are retrying to connect with some delay, this represents the scheduled inititation of another attempt
*/
@Nullable
final Future scheduled;
/**
* true iff we are retrying to connect after some failure (immediately or following a delay)
*/
final boolean isFailingToConnect;
Connecting(Disconnected previous, Future> attempt)
{
this(previous, attempt, null);
}
Connecting(Disconnected previous, Future> attempt, Future scheduled)
{
super(Kind.CONNECTING, previous.maintenance);
this.attempt = attempt;
this.scheduled = scheduled;
this.isFailingToConnect = scheduled != null || (previous.isConnecting() && previous.connecting().isFailingToConnect);
}
/**
* Cancel the connection attempt
*
* No cleanup is needed here, as {@link #attempt} is only completed on the eventLoop,
* so we have either already invoked the callbacks and are no longer in {@link #state},
* or the {@link OutboundConnectionInitiator} will handle our successful cancellation
* when it comes to complete, by closing the channel (if we could not cancel it before then)
*/
void cancel()
{
if (scheduled != null)
scheduled.cancel(true);
// we guarantee that attempt is only ever completed by the eventLoop
boolean cancelled = attempt.cancel(true);
assert cancelled;
}
}
private volatile State state;
/** The connection is being permanently closed */
private volatile Future closing;
/** The connection is being permanently closed in the near future */
private volatile Future scheduledClose;
OutboundConnection(ConnectionType type, OutboundConnectionSettings settings, EndpointAndGlobal reserveCapacityInBytes)
{
this.template = settings.withDefaults(ConnectionCategory.MESSAGING);
this.type = type;
this.eventLoop = template.socketFactory.defaultGroup().next();
this.pendingCapacityInBytes = template.applicationSendQueueCapacityInBytes;
this.reserveCapacityInBytes = reserveCapacityInBytes;
this.callbacks = template.callbacks;
this.debug = template.debug;
this.queue = new OutboundMessageQueue(approxTime, this::onExpired);
this.delivery = type == ConnectionType.LARGE_MESSAGES
? new LargeMessageDelivery(template.socketFactory.synchronousWorkExecutor)
: new EventLoopDelivery();
setDisconnected();
}
/**
* This is the main entry point for enqueuing a message to be sent to the remote peer.
*/
public void enqueue(Message message) throws ClosedChannelException
{
if (isClosing())
throw new ClosedChannelException();
final int canonicalSize = canonicalSize(message);
if (canonicalSize > DatabaseDescriptor.getInternodeMaxMessageSizeInBytes())
throw new Message.OversizedMessageException(canonicalSize);
submittedUpdater.incrementAndGet(this);
switch (acquireCapacity(canonicalSize))
{
case INSUFFICIENT_ENDPOINT:
// if we're overloaded to one endpoint, we may be accumulating expirable messages, so
// attempt an expiry to see if this makes room for our newer message.
// this is an optimisation only; messages will be expired on ~100ms cycle, and by Delivery when it runs
if (queue.maybePruneExpired() && SUCCESS == acquireCapacity(canonicalSize))
break;
case INSUFFICIENT_GLOBAL:
onOverloaded(message);
return;
}
queue.add(message);
delivery.execute();
// we might race with the channel closing; if this happens, to ensure this message eventually arrives
// we need to remove ourselves from the queue and throw a ClosedChannelException, so that another channel
// can be opened in our place to try and send on.
if (isClosing() && queue.remove(message))
{
releaseCapacity(1, canonicalSize);
throw new ClosedChannelException();
}
}
/**
* Try to acquire the necessary resource permits for a number of pending bytes for this connection.
*
* Since the owner limit is shared amongst multiple connections, our semantics cannot be super trivial.
* Were they per-connection, we could simply perform an atomic increment of the queue size, then
* allocate any excess we need in the reserve, and on release free everything we see from both.
* Since we are coordinating two independent atomic variables we have to track every byte we allocate in reserve
* and ensure it is matched by a corresponding released byte. We also need to be sure we do not permit another
* releasing thread to release reserve bytes we have not yet - and may never - actually reserve.
*
* As such, we have to first check if we would need reserve bytes, then allocate them *before* we increment our
* queue size. We only increment the queue size if the reserve bytes are definitely not needed, or we could first
* obtain them. If in the process of obtaining any reserve bytes the queue size changes, we have some bytes that are
* reserved for us, but may be a different number to that we need. So we must continue to track these.
*
* In the happy path, this is still efficient as we simply CAS
*/
private Outcome acquireCapacity(long bytes)
{
return acquireCapacity(1, bytes);
}
private Outcome acquireCapacity(long count, long bytes)
{
long increment = pendingCountAndBytes(count, bytes);
long unusedClaimedReserve = 0;
Outcome outcome = null;
loop: while (true)
{
long current = pendingCountAndBytes;
if (isMaxPendingCount(current))
{
outcome = INSUFFICIENT_ENDPOINT;
break;
}
long next = current + increment;
if (pendingBytes(next) <= pendingCapacityInBytes)
{
if (pendingCountAndBytesUpdater.compareAndSet(this, current, next))
{
outcome = SUCCESS;
break;
}
continue;
}
State state = this.state;
if (state.isConnecting() && state.connecting().isFailingToConnect)
{
outcome = INSUFFICIENT_ENDPOINT;
break;
}
long requiredReserve = min(bytes, pendingBytes(next) - pendingCapacityInBytes);
if (unusedClaimedReserve < requiredReserve)
{
long extraGlobalReserve = requiredReserve - unusedClaimedReserve;
switch (outcome = reserveCapacityInBytes.tryAllocate(extraGlobalReserve))
{
case INSUFFICIENT_ENDPOINT:
case INSUFFICIENT_GLOBAL:
break loop;
case SUCCESS:
unusedClaimedReserve += extraGlobalReserve;
}
}
if (pendingCountAndBytesUpdater.compareAndSet(this, current, next))
{
unusedClaimedReserve -= requiredReserve;
break;
}
}
if (unusedClaimedReserve > 0)
reserveCapacityInBytes.release(unusedClaimedReserve);
return outcome;
}
/**
* Mark a number of pending bytes as flushed to the network, releasing their capacity for new outbound messages.
*/
private void releaseCapacity(long count, long bytes)
{
long decrement = pendingCountAndBytes(count, bytes);
long prev = pendingCountAndBytesUpdater.getAndAdd(this, -decrement);
if (pendingBytes(prev) > pendingCapacityInBytes)
{
long excess = min(pendingBytes(prev) - pendingCapacityInBytes, bytes);
reserveCapacityInBytes.release(excess);
}
}
private void onOverloaded(Message message)
{
overloadedCountUpdater.incrementAndGet(this);
int canonicalSize = canonicalSize(message);
overloadedBytesUpdater.addAndGet(this, canonicalSize);
noSpamLogger.warn("{} overloaded; dropping {} message (queue: {} local, {} endpoint, {} global)",
this, FBUtilities.prettyPrintMemory(canonicalSize),
readablePendingBytes, readableReserveEndpointUsing, readableReserveGlobalUsing);
callbacks.onOverloaded(message, template.to);
}
/**
* Take any necessary cleanup action after a message has been selected to be discarded from the queue.
*
* Only to be invoked while holding OutboundMessageQueue.WithLock
*/
private boolean onExpired(Message message)
{
noSpamLogger.warn("{} dropping message of type {} whose timeout expired before reaching the network", id(), message.verb());
releaseCapacity(1, canonicalSize(message));
expiredCount += 1;
expiredBytes += canonicalSize(message);
callbacks.onExpired(message, template.to);
return true;
}
/**
* Take any necessary cleanup action after a message has been selected to be discarded from the queue.
*
* Only to be invoked by the delivery thread
*/
private void onFailedSerialize(Message message, int messagingVersion, int bytesWrittenToNetwork, Throwable t)
{
logger.warn("{} dropping message of type {} due to error", id(), message.verb(), t);
JVMStabilityInspector.inspectThrowable(t);
releaseCapacity(1, canonicalSize(message));
errorCount += 1;
errorBytes += message.serializedSize(messagingVersion);
callbacks.onFailedSerialize(message, template.to, messagingVersion, bytesWrittenToNetwork, t);
}
/**
* Take any necessary cleanup action after a message has been selected to be discarded from the queue on close.
* Note that this is only for messages that were queued prior to closing without graceful flush, OR
* for those that are unceremoniously dropped when we decide close has been trying to complete for too long.
*/
private void onClosed(Message message)
{
releaseCapacity(1, canonicalSize(message));
callbacks.onDiscardOnClose(message, template.to);
}
/**
* Delivery bundles the following:
*
* - the work that is necessary to actually deliver messages safely, and handle any exceptional states
* - the ability to schedule delivery for some time in the future
* - the ability to schedule some non-delivery work to happen some time in the future, that is guaranteed
* NOT to coincide with delivery for its duration, including any data that is being flushed (e.g. for closing channels)
* - this feature is *not* efficient, and should only be used for infrequent operations
*/
private abstract class Delivery extends AtomicInteger implements Runnable
{
final ExecutorService executor;
// the AtomicInteger we extend always contains some combination of these bit flags, representing our current run state
/** Not running, and will not be scheduled again until transitioned to a new state */
private static final int STOPPED = 0;
/** Currently executing (may only be scheduled to execute, or may be about to terminate);
* will stop at end of this run, without rescheduling */
private static final int EXECUTING = 1;
/** Another execution has been requested; a new execution will begin some time after this state is taken */
private static final int EXECUTE_AGAIN = 2;
/** We are currently executing and will submit another execution before we terminate */
private static final int EXECUTING_AGAIN = EXECUTING | EXECUTE_AGAIN;
/** Will begin a new execution some time after this state is taken, but only once some condition is met.
* This state will initially be taken in tandem with EXECUTING, but if delivery completes without clearing
* the state, the condition will be held on its own until {@link #executeAgain} is invoked */
private static final int WAITING_TO_EXECUTE = 4;
/**
* Force all task execution to stop, once any currently in progress work is completed
*/
private volatile boolean terminated;
/**
* Is there asynchronous delivery work in progress.
*
* This temporarily prevents any {@link #stopAndRun} work from being performed.
* Once both inProgress and stopAndRun are set we perform no more delivery work until one is unset,
* to ensure we eventually run stopAndRun.
*
* This should be updated and read only on the Delivery thread.
*/
private boolean inProgress = false;
/**
* Request a task's execution while there is no delivery work in progress.
*
* This is to permit cleanly tearing down a connection without interrupting any messages that might be in flight.
* If stopAndRun is set, we should not enter doRun() until a corresponding setInProgress(false) occurs.
*/
final AtomicReference stopAndRun = new AtomicReference<>();
Delivery(ExecutorService executor)
{
this.executor = executor;
}
/**
* Ensure that any messages or stopAndRun that were queued prior to this invocation will be seen by at least
* one future invocation of the delivery task, unless delivery has already been terminated.
*/
public void execute()
{
if (get() < EXECUTE_AGAIN && STOPPED == getAndUpdate(i -> i == STOPPED ? EXECUTING: i | EXECUTE_AGAIN))
executor.execute(this);
}
private boolean isExecuting(int state)
{
return 0 != (state & EXECUTING);
}
/**
* This method is typically invoked after WAITING_TO_EXECUTE is set.
*
* However WAITING_TO_EXECUTE does not need to be set; all this method needs to ensure is that
* delivery unconditionally performs one new execution promptly.
*/
void executeAgain()
{
// if we are already executing, set EXECUTING_AGAIN and leave scheduling to the currently running one.
// otherwise, set ourselves unconditionally to EXECUTING and schedule ourselves immediately
if (!isExecuting(getAndUpdate(i -> !isExecuting(i) ? EXECUTING : EXECUTING_AGAIN)))
executor.execute(this);
}
/**
* Invoke this when we cannot make further progress now, but we guarantee that we will execute later when we can.
* This simply communicates to {@link #run} that we should not schedule ourselves again, just unset the EXECUTING bit.
*/
void promiseToExecuteLater()
{
set(EXECUTING | WAITING_TO_EXECUTE);
}
/**
* Called when exiting {@link #run} to schedule another run if necessary.
*
* If we are currently executing, we only reschedule if the present state is EXECUTING_AGAIN.
* If this is the case, we clear the EXECUTE_AGAIN bit (setting ourselves to EXECUTING), and reschedule.
* Otherwise, we clear the EXECUTING bit and terminate, which will set us to either STOPPED or WAITING_TO_EXECUTE
* (or possibly WAITING_TO_EXECUTE | EXECUTE_AGAIN, which is logically the same as WAITING_TO_EXECUTE)
*/
private void maybeExecuteAgain()
{
if (EXECUTING_AGAIN == getAndUpdate(i -> i == EXECUTING_AGAIN ? EXECUTING : (i & ~EXECUTING)))
executor.execute(this);
}
/**
* No more tasks or delivery will be executed, once any in progress complete.
*/
public void terminate()
{
terminated = true;
}
/**
* Only to be invoked by the Delivery task.
*
* If true, indicates that we have begun asynchronous delivery work, so that
* we cannot safely stopAndRun until it completes.
*
* Once it completes, we ensure any stopAndRun task has a chance to execute
* by ensuring delivery is scheduled.
*
* If stopAndRun is also set, we should not enter doRun() until a corresponding
* setInProgress(false) occurs.
*/
void setInProgress(boolean inProgress)
{
boolean wasInProgress = this.inProgress;
this.inProgress = inProgress;
if (!inProgress && wasInProgress)
executeAgain();
}
/**
* Perform some delivery work.
*
* Must never be invoked directly, only via {@link #execute()}
*/
public void run()
{
/* do/while handling setup for {@link #doRun()}, and repeat invocations thereof */
while (true)
{
if (terminated)
return;
if (null != stopAndRun.get())
{
// if we have an external request to perform, attempt it - if no async delivery is in progress
if (inProgress)
{
// if we are in progress, we cannot do anything;
// so, exit and rely on setInProgress(false) executing us
// (which must happen later, since it must happen on this thread)
promiseToExecuteLater();
break;
}
stopAndRun.getAndSet(null).run();
}
State state = OutboundConnection.this.state;
if (!state.isEstablished() || !state.established().isConnected())
{
// if we have messages yet to deliver, or a task to run, we need to reconnect and try again
// we try to reconnect before running another stopAndRun so that we do not infinite loop in close
if (hasPending() || null != stopAndRun.get())
{
promiseToExecuteLater();
requestConnect().addListener(f -> executeAgain());
}
break;
}
if (!doRun(state.established()))
break;
}
maybeExecuteAgain();
}
/**
* @return true if we should run again immediately;
* always false for eventLoop executor, as want to service other channels
*/
abstract boolean doRun(Established established);
/**
* Schedule a task to run later on the delivery thread while delivery is not in progress,
* i.e. there are no bytes in flight to the network buffer.
*
* Does not guarantee to run promptly if there is no current connection to the remote host.
* May wait until a new connection is established, or a connection timeout elapses, before executing.
*
* Update the shared atomic property containing work we want to interrupt message processing to perform,
* the invoke schedule() to be certain it gets run.
*/
void stopAndRun(Runnable run)
{
stopAndRun.accumulateAndGet(run, OutboundConnection::andThen);
execute();
}
/**
* Schedule a task to run on the eventLoop, guaranteeing that delivery will not occur while the task is performed.
*/
abstract void stopAndRunOnEventLoop(Runnable run);
}
/**
* Delivery that runs entirely on the eventLoop
*
* Since this has single threaded access to most of its environment, it can be simple and efficient, however
* it must also have bounded run time, and limit its resource consumption to ensure other channels serviced by the
* eventLoop can also make progress.
*
* This operates on modest buffers, no larger than the {@link OutboundConnections#LARGE_MESSAGE_THRESHOLD} and
* filling at most one at a time before writing (potentially asynchronously) to the socket.
*
* We track the number of bytes we have in flight, ensuring no more than a user-defined maximum at any one time.
*/
class EventLoopDelivery extends Delivery
{
private int flushingBytes;
private boolean isWritable = true;
EventLoopDelivery()
{
super(eventLoop);
}
/**
* {@link Delivery#doRun}
*
* Since we are on the eventLoop, in order to ensure other channels are serviced
* we never return true to request another run immediately.
*
* If there is more work to be done, we submit ourselves for execution once the eventLoop has time.
*/
@SuppressWarnings("resource")
boolean doRun(Established established)
{
if (!isWritable)
return false;
// pendingBytes is updated before queue.size() (which triggers notEmpty, and begins delivery),
// so it is safe to use it here to exit delivery
// this number is inaccurate for old versions, but we don't mind terribly - we'll send at least one message,
// and get round to it eventually (though we could add a fudge factor for some room for older versions)
int maxSendBytes = (int) min(pendingBytes() - flushingBytes, LARGE_MESSAGE_THRESHOLD);
if (maxSendBytes == 0)
return false;
OutboundConnectionSettings settings = established.settings;
int messagingVersion = established.messagingVersion;
FrameEncoder.Payload sending = null;
int canonicalSize = 0; // number of bytes we must use for our resource accounting
int sendingBytes = 0;
int sendingCount = 0;
try (OutboundMessageQueue.WithLock withLock = queue.lockOrCallback(approxTime.now(), this::execute))
{
if (withLock == null)
return false; // we failed to acquire the queue lock, so return; we will be scheduled again when the lock is available
sending = established.payloadAllocator.allocate(true, maxSendBytes);
DataOutputBufferFixed out = new DataOutputBufferFixed(sending.buffer);
Message next;
while ( null != (next = withLock.peek()) )
{
try
{
int messageSize = next.serializedSize(messagingVersion);
// actual message size for this version is larger than permitted maximum
if (messageSize > DatabaseDescriptor.getInternodeMaxMessageSizeInBytes())
throw new Message.OversizedMessageException(messageSize);
if (messageSize > sending.remaining())
{
// if we don't have enough room to serialize the next message, we have either
// 1) run out of room after writing some messages successfully; this might mean that we are
// overflowing our highWaterMark, or that we have just filled our buffer
// 2) we have a message that is too large for this connection; this can happen if a message's
// size was calculated for the wrong messaging version when enqueued.
// In this case we want to write it anyway, so simply allocate a large enough buffer.
if (sendingBytes > 0)
break;
sending.release();
sending = null; // set to null to prevent double-release if we fail to allocate our new buffer
sending = established.payloadAllocator.allocate(true, messageSize);
//noinspection IOResourceOpenedButNotSafelyClosed
out = new DataOutputBufferFixed(sending.buffer);
}
Tracing.instance.traceOutgoingMessage(next, messageSize, settings.connectTo);
Message.serializer.serialize(next, out, messagingVersion);
if (sending.length() != sendingBytes + messageSize)
throw new InvalidSerializedSizeException(next.verb(), messageSize, sending.length() - sendingBytes);
canonicalSize += canonicalSize(next);
sendingCount += 1;
sendingBytes += messageSize;
}
catch (Throwable t)
{
onFailedSerialize(next, messagingVersion, 0, t);
assert sending != null;
// reset the buffer to ignore the message we failed to serialize
sending.trim(sendingBytes);
}
withLock.removeHead(next);
}
if (0 == sendingBytes)
return false;
sending.finish();
debug.onSendSmallFrame(sendingCount, sendingBytes);
ChannelFuture flushResult = AsyncChannelPromise.writeAndFlush(established.channel, sending);
sending = null;
if (flushResult.isSuccess())
{
sentCount += sendingCount;
sentBytes += sendingBytes;
debug.onSentSmallFrame(sendingCount, sendingBytes);
}
else
{
flushingBytes += canonicalSize;
setInProgress(true);
boolean hasOverflowed = flushingBytes >= settings.flushHighWaterMark;
if (hasOverflowed)
{
isWritable = false;
promiseToExecuteLater();
}
int releaseBytesFinal = canonicalSize;
int sendingBytesFinal = sendingBytes;
int sendingCountFinal = sendingCount;
flushResult.addListener(future -> {
releaseCapacity(sendingCountFinal, releaseBytesFinal);
flushingBytes -= releaseBytesFinal;
if (flushingBytes == 0)
setInProgress(false);
if (!isWritable && flushingBytes <= settings.flushLowWaterMark)
{
isWritable = true;
executeAgain();
}
if (future.isSuccess())
{
sentCount += sendingCountFinal;
sentBytes += sendingBytesFinal;
debug.onSentSmallFrame(sendingCountFinal, sendingBytesFinal);
}
else
{
errorCount += sendingCountFinal;
errorBytes += sendingBytesFinal;
invalidateChannel(established, future.cause());
debug.onFailedSmallFrame(sendingCountFinal, sendingBytesFinal);
}
});
canonicalSize = 0;
}
}
catch (Throwable t)
{
errorCount += sendingCount;
errorBytes += sendingBytes;
invalidateChannel(established, t);
}
finally
{
if (canonicalSize > 0)
releaseCapacity(sendingCount, canonicalSize);
if (sending != null)
sending.release();
if (pendingBytes() > flushingBytes && isWritable)
execute();
}
return false;
}
void stopAndRunOnEventLoop(Runnable run)
{
stopAndRun(run);
}
}
/**
* Delivery that coordinates between the eventLoop and another (non-dedicated) thread
*
* This is to service messages that are too large to fully serialize on the eventLoop, as they could block
* prompt service of other requests. Since our serializers assume blocking IO, the easiest approach is to
* ensure a companion thread performs blocking IO that, under the hood, is serviced by async IO on the eventLoop.
*
* Most of the work here is handed off to {@link AsyncChannelOutputPlus}, with our main job being coordinating
* when and what we should run.
*
* To avoid allocating a huge number of threads across a cluster, we utilise the shared methods of {@link Delivery}
* to ensure that only one run() is actually scheduled to run at a time - this permits us to use any {@link ExecutorService}
* as a backing, with the number of threads defined only by the maximum concurrency needed to deliver all large messages.
* We use a shared caching {@link java.util.concurrent.ThreadPoolExecutor}, and rename the Threads that service
* our connection on entry and exit.
*/
class LargeMessageDelivery extends Delivery
{
static final int DEFAULT_BUFFER_SIZE = 32 * 1024;
LargeMessageDelivery(ExecutorService executor)
{
super(executor);
}
/**
* A simple wrapper of {@link Delivery#run} to set the current Thread name for the duration of its execution.
*/
public void run()
{
String threadName, priorThreadName = null;
try
{
priorThreadName = Thread.currentThread().getName();
threadName = "Messaging-OUT-" + template.from() + "->" + template.to + '-' + type;
Thread.currentThread().setName(threadName);
super.run();
}
finally
{
if (priorThreadName != null)
Thread.currentThread().setName(priorThreadName);
}
}
@SuppressWarnings({ "resource", "RedundantSuppression" }) // make eclipse warnings go away
boolean doRun(Established established)
{
Message send = queue.tryPoll(approxTime.now(), this::execute);
if (send == null)
return false;
AsyncMessageOutputPlus out = null;
try
{
int messageSize = send.serializedSize(established.messagingVersion);
out = new AsyncMessageOutputPlus(established.channel, DEFAULT_BUFFER_SIZE, messageSize, established.payloadAllocator);
// actual message size for this version is larger than permitted maximum
if (messageSize > DatabaseDescriptor.getInternodeMaxMessageSizeInBytes())
throw new Message.OversizedMessageException(messageSize);
Tracing.instance.traceOutgoingMessage(send, messageSize, established.settings.connectTo);
Message.serializer.serialize(send, out, established.messagingVersion);
if (out.position() != messageSize)
throw new InvalidSerializedSizeException(send.verb(), messageSize, out.position());
out.close();
sentCount += 1;
sentBytes += messageSize;
releaseCapacity(1, canonicalSize(send));
return hasPending();
}
catch (Throwable t)
{
boolean tryAgain = true;
if (out != null)
{
out.discard();
if (out.flushed() > 0 ||
isCausedBy(t, cause -> isConnectionReset(cause)
|| cause instanceof Errors.NativeIoException
|| cause instanceof AsyncChannelOutputPlus.FlushException))
{
// close the channel, and wait for eventLoop to execute
disconnectNow(established).awaitUninterruptibly();
tryAgain = false;
try
{
// after closing, wait until we are signalled about the in flight writes;
// this ensures flushedToNetwork() is correct below
out.waitUntilFlushed(0, 0);
}
catch (Throwable ignore)
{
// irrelevant
}
}
}
onFailedSerialize(send, established.messagingVersion, out == null ? 0 : (int) out.flushedToNetwork(), t);
return tryAgain;
}
}
void stopAndRunOnEventLoop(Runnable run)
{
stopAndRun(() -> {
try
{
runOnEventLoop(run).await();
}
catch (InterruptedException e)
{
throw new RuntimeException(e);
}
});
}
}
/*
* Size used for capacity enforcement purposes. Using current messaging version no matter what the peer's version is.
*/
private int canonicalSize(Message message)
{
return message.serializedSize(current_version);
}
private void invalidateChannel(Established established, Throwable cause)
{
JVMStabilityInspector.inspectThrowable(cause);
if (state != established)
return; // do nothing; channel already invalidated
if (isCausedByConnectionReset(cause))
logger.info("{} channel closed by provider", id(), cause);
else
logger.error("{} channel in potentially inconsistent state after error; closing", id(), cause);
disconnectNow(established);
}
/**
* Attempt to open a new channel to the remote endpoint.
*
* Most of the actual work is performed by OutboundConnectionInitiator, this method just manages
* our book keeping on either success or failure.
*
* This method is only to be invoked by the eventLoop, and the inner class' methods should only be evaluated by the eventtLoop
*/
Future initiate()
{
class Initiate
{
/**
* If we fail to connect, we want to try and connect again before any messages timeout.
* However, we update this each time to ensure we do not retry unreasonably often, and settle on a periodicity
* that might lead to timeouts in some aggressive systems.
*/
long retryRateMillis = DatabaseDescriptor.getMinRpcTimeout(MILLISECONDS) / 2;
// our connection settings, possibly updated on retry
int messagingVersion = template.endpointToVersion().get(template.to);
OutboundConnectionSettings settings;
/**
* If we failed for any reason, try again
*/
void onFailure(Throwable cause)
{
if (cause instanceof ConnectException)
noSpamLogger.info("{} failed to connect", id(), cause);
else
noSpamLogger.error("{} failed to connect", id(), cause);
JVMStabilityInspector.inspectThrowable(cause);
if (hasPending())
{
Promise> result = new AsyncPromise<>(eventLoop);
state = new Connecting(state.disconnected(), result, eventLoop.schedule(() -> attempt(result), max(100, retryRateMillis), MILLISECONDS));
retryRateMillis = min(1000, retryRateMillis * 2);
}
else
{
// this Initiate will be discarded
state = Disconnected.dormant(state.disconnected().maintenance);
}
}
void onCompletedHandshake(Result result)
{
switch (result.outcome)
{
case SUCCESS:
// it is expected that close, if successful, has already cancelled us; so we do not need to worry about leaking connections
assert !state.isClosed();
MessagingSuccess success = result.success();
debug.onConnect(success.messagingVersion, settings);
state.disconnected().maintenance.cancel(false);
FrameEncoder.PayloadAllocator payloadAllocator = success.allocator;
Channel channel = success.channel;
Established established = new Established(messagingVersion, channel, payloadAllocator, settings);
state = established;
channel.pipeline().addLast("handleExceptionalStates", new ChannelInboundHandlerAdapter() {
@Override
public void channelInactive(ChannelHandlerContext ctx)
{
disconnectNow(established);
ctx.fireChannelInactive();
}
@Override
public void exceptionCaught(ChannelHandlerContext ctx, Throwable cause)
{
try
{
invalidateChannel(established, cause);
}
catch (Throwable t)
{
logger.error("Unexpected exception in {}.exceptionCaught", this.getClass().getSimpleName(), t);
}
}
});
++successfulConnections;
logger.info("{} successfully connected, version = {}, framing = {}, encryption = {}",
id(true),
success.messagingVersion,
settings.framing,
encryptionConnectionSummary(channel));
break;
case RETRY:
if (logger.isTraceEnabled())
logger.trace("{} incorrect legacy peer version predicted; reconnecting", id());
// the messaging version we connected with was incorrect; try again with the one supplied by the remote host
messagingVersion = result.retry().withMessagingVersion;
settings.endpointToVersion.set(settings.to, messagingVersion);
initiate();
break;
case INCOMPATIBLE:
// we cannot communicate with this peer given its messaging version; mark this as any other failure, and continue trying
Throwable t = new IOException(String.format("Incompatible peer: %s, messaging version: %s",
settings.to, result.incompatible().maxMessagingVersion));
t.fillInStackTrace();
onFailure(t);
break;
default:
throw new AssertionError();
}
}
/**
* Initiate all the actions required to establish a working, valid connection. This includes
* opening the socket, negotiating the internode messaging handshake, and setting up the working
* Netty {@link Channel}. However, this method will not block for all those actions: it will only
* kick off the connection attempt, setting the @{link #connecting} future to track its completion.
*
* Note: this should only be invoked on the event loop.
*/
private void attempt(Promise> result)
{
++connectionAttempts;
/*
* Re-evaluate messagingVersion before re-attempting the connection in case
* endpointToVersion were updated. This happens if the outbound connection
* is made before the endpointToVersion table is initially constructed or out
* of date (e.g. if outbound connections are established for gossip
* as a result of an inbound connection) and can result in the wrong outbound
* port being selected if configured with enable_legacy_ssl_storage_port=true.
*/
int knownMessagingVersion = messagingVersion();
if (knownMessagingVersion != messagingVersion)
{
logger.trace("Endpoint version changed from {} to {} since connection initialized, updating.",
messagingVersion, knownMessagingVersion);
messagingVersion = knownMessagingVersion;
}
settings = template;
if (messagingVersion > settings.acceptVersions.max)
messagingVersion = settings.acceptVersions.max;
// ensure we connect to the correct SSL port
settings = settings.withLegacyPortIfNecessary(messagingVersion);
initiateMessaging(eventLoop, type, settings, messagingVersion, result)
.addListener(future -> {
if (future.isCancelled())
return;
if (future.isSuccess()) //noinspection unchecked
onCompletedHandshake((Result) future.getNow());
else
onFailure(future.cause());
});
}
Future> initiate()
{
Promise> result = new AsyncPromise<>(eventLoop);
state = new Connecting(state.disconnected(), result);
attempt(result);
return result;
}
}
return new Initiate().initiate();
}
/**
* Returns a future that completes when we are _maybe_ reconnected.
*
* The connection attempt is guaranteed to have completed (successfully or not) by the time any listeners are invoked,
* so if a reconnection attempt is needed, it is already scheduled.
*/
private Future requestConnect()
{
// we may race with updates to this variable, but this is fine, since we only guarantee that we see a value
// that did at some point represent an active connection attempt - if it is stale, it will have been completed
// and the caller can retry (or utilise the successfully established connection)
{
State state = this.state;
if (state.isConnecting())
return state.connecting().attempt;
}
Promise © 2015 - 2024 Weber Informatics LLC | Privacy Policy