2 * Copyright (c) 2016 Cisco Systems, Inc. and others. All rights reserved.
4 * This program and the accompanying materials are made available under the
5 * terms of the Eclipse Public License v1.0 which accompanies this distribution,
6 * and is available at http://www.eclipse.org/legal/epl-v10.html
8 package org.opendaylight.controller.cluster.access.client;
10 import akka.actor.ActorRef;
11 import com.google.common.annotations.VisibleForTesting;
12 import com.google.common.base.MoreObjects;
13 import com.google.common.base.MoreObjects.ToStringHelper;
14 import com.google.common.base.Preconditions;
15 import edu.umd.cs.findbugs.annotations.SuppressFBWarnings;
16 import java.util.Collection;
17 import java.util.Optional;
18 import java.util.concurrent.TimeUnit;
19 import java.util.concurrent.TimeoutException;
20 import java.util.concurrent.locks.Lock;
21 import java.util.concurrent.locks.ReentrantLock;
22 import java.util.function.Consumer;
23 import javax.annotation.Nonnull;
24 import javax.annotation.concurrent.GuardedBy;
25 import javax.annotation.concurrent.NotThreadSafe;
26 import org.opendaylight.controller.cluster.access.concepts.Request;
27 import org.opendaylight.controller.cluster.access.concepts.RequestException;
28 import org.opendaylight.controller.cluster.access.concepts.Response;
29 import org.opendaylight.controller.cluster.access.concepts.ResponseEnvelope;
30 import org.opendaylight.controller.cluster.access.concepts.RuntimeRequestException;
31 import org.slf4j.Logger;
32 import org.slf4j.LoggerFactory;
33 import scala.concurrent.duration.FiniteDuration;
36 * Base class for a connection to the backend. Responsible to queueing and dispatch of requests toward the backend.
37 * Can be in three conceptual states: Connecting, Connected and Reconnecting, which are represented by public final
38 * classes exposed from this package.
40 * @author Robert Varga
43 public abstract class AbstractClientConnection<T extends BackendInfo> {
44 private static final Logger LOG = LoggerFactory.getLogger(AbstractClientConnection.class);
47 * Timers involved in communication with the backend. There are three tiers which are spaced out to allow for
48 * recovery at each tier. Keep these constants in nanoseconds, as that prevents unnecessary conversions in the fast
52 * Backend aliveness timer. This is reset whenever we receive a response from the backend and kept armed whenever
53 * we have an outstanding request. If when this time expires, we tear down this connection and attempt to reconnect
56 public static final long DEFAULT_BACKEND_ALIVE_TIMEOUT_NANOS = TimeUnit.SECONDS.toNanos(30);
59 * Request timeout. If the request fails to complete within this time since it was originally enqueued, we time
62 public static final long DEFAULT_REQUEST_TIMEOUT_NANOS = TimeUnit.MINUTES.toNanos(2);
65 * No progress timeout. A client fails to make any forward progress in this time, it will terminate itself.
67 public static final long DEFAULT_NO_PROGRESS_TIMEOUT_NANOS = TimeUnit.MINUTES.toNanos(15);
69 // Emit a debug entry if we sleep for more that this amount
70 private static final long DEBUG_DELAY_NANOS = TimeUnit.MILLISECONDS.toNanos(100);
72 // Upper bound on the time a thread is forced to sleep to keep queue size under control
73 private static final long MAX_DELAY_SECONDS = 5;
74 private static final long MAX_DELAY_NANOS = TimeUnit.SECONDS.toNanos(MAX_DELAY_SECONDS);
76 private final Lock lock = new ReentrantLock();
77 private final ClientActorContext context;
79 private final TransmitQueue queue;
80 private final Long cookie;
81 private final String backendName;
84 private boolean haveTimer;
87 * Time reference when we saw any activity from the backend.
89 private long lastReceivedTicks;
91 private volatile RequestException poisoned;
93 // Private constructor to avoid code duplication.
94 private AbstractClientConnection(final AbstractClientConnection<T> oldConn, final TransmitQueue newQueue,
95 final String backendName) {
96 this.context = Preconditions.checkNotNull(oldConn.context);
97 this.cookie = Preconditions.checkNotNull(oldConn.cookie);
98 this.backendName = Preconditions.checkNotNull(backendName);
99 this.queue = Preconditions.checkNotNull(newQueue);
100 // Will be updated in finishReplay if needed.
101 this.lastReceivedTicks = oldConn.lastReceivedTicks;
104 // This constructor is only to be called by ConnectingClientConnection constructor.
105 // Do not allow subclassing outside of this package
106 AbstractClientConnection(final ClientActorContext context, final Long cookie, final String backendName,
107 final int queueDepth) {
108 this.context = Preconditions.checkNotNull(context);
109 this.cookie = Preconditions.checkNotNull(cookie);
110 this.backendName = Preconditions.checkNotNull(backendName);
111 this.queue = new TransmitQueue.Halted(queueDepth);
112 this.lastReceivedTicks = currentTime();
115 // This constructor is only to be called (indirectly) by ReconnectingClientConnection constructor.
116 // Do not allow subclassing outside of this package
117 AbstractClientConnection(final AbstractClientConnection<T> oldConn) {
118 this(oldConn, new TransmitQueue.Halted(oldConn.queue, oldConn.currentTime()), oldConn.backendName);
121 // This constructor is only to be called (indirectly) by ConnectedClientConnection constructor.
122 // Do not allow subclassing outside of this package
123 AbstractClientConnection(final AbstractClientConnection<T> oldConn, final T newBackend,
124 final int queueDepth) {
125 this(oldConn, new TransmitQueue.Transmitting(oldConn.queue, queueDepth, newBackend, oldConn.currentTime(),
126 Preconditions.checkNotNull(oldConn.context).messageSlicer()), newBackend.getName());
129 public final ClientActorContext context() {
133 public final @Nonnull Long cookie() {
137 public final ActorRef localActor() {
138 return context.self();
141 public final long currentTime() {
142 return context.ticker().read();
146 * Send a request to the backend and invoke a specified callback when it finishes. This method is safe to invoke
149 * <p>This method may put the caller thread to sleep in order to throttle the request rate.
150 * The callback may be called before the sleep finishes.
152 * @param request Request to send
153 * @param callback Callback to invoke
155 public final void sendRequest(final Request<?, ?> request, final Consumer<Response<?, ?>> callback) {
156 final long now = currentTime();
157 sendEntry(new ConnectionEntry(request, callback, now), now);
161 * Send a request to the backend and invoke a specified callback when it finishes. This method is safe to invoke
165 * Note that unlike {@link #sendRequest(Request, Consumer)}, this method does not exert backpressure, hence it
166 * should never be called from an application thread.
168 * @param request Request to send
169 * @param callback Callback to invoke
170 * @param enqueuedTicks Time (according to {@link #currentTime()} of request enqueue
172 public final void enqueueRequest(final Request<?, ?> request, final Consumer<Response<?, ?>> callback,
173 final long enqueuedTicks) {
174 enqueueEntry(new ConnectionEntry(request, callback, enqueuedTicks), currentTime());
177 private long enqueueOrForward(final ConnectionEntry entry, final long now) {
180 commonEnqueue(entry, now);
181 return queue.enqueueOrForward(entry, now);
188 * Enqueue an entry, possibly also transmitting it.
190 public final void enqueueEntry(final ConnectionEntry entry, final long now) {
193 commonEnqueue(entry, now);
194 queue.enqueueOrReplay(entry, now);
201 private void commonEnqueue(final ConnectionEntry entry, final long now) {
202 final RequestException maybePoison = poisoned;
203 if (maybePoison != null) {
204 throw new IllegalStateException("Connection " + this + " has been poisoned", maybePoison);
207 if (queue.isEmpty()) {
208 // The queue is becoming non-empty, schedule a timer.
209 scheduleTimer(entry.getEnqueuedTicks() + context.config().getRequestTimeout() - now);
213 // To be called from ClientActorBehavior on ConnectedClientConnection after entries are replayed.
214 final void cancelDebt() {
215 queue.cancelDebt(currentTime());
218 public abstract Optional<T> getBackendInfo();
220 final Collection<ConnectionEntry> startReplay() {
222 return queue.drain();
226 final void finishReplay(final ReconnectForwarder forwarder) {
227 setForwarder(forwarder);
230 * The process of replaying all messages may have taken a significant chunk of time, depending on type
231 * of messages, queue depth and available processing power. In extreme situations this may have already
232 * exceeded BACKEND_ALIVE_TIMEOUT_NANOS, in which case we are running the risk of not making reasonable forward
233 * progress before we start a reconnect cycle.
235 * Note that the timer is armed after we have sent the first message, hence we should be seeing a response
236 * from the backend before we see a timeout, simply due to how the mailbox operates.
238 * At any rate, reset the timestamp once we complete reconnection (which an atomic transition from the
239 * perspective of outside world), as that makes it a bit easier to reason about timing of events.
241 lastReceivedTicks = currentTime();
246 final void setForwarder(final ReconnectForwarder forwarder) {
247 queue.setForwarder(forwarder, currentTime());
251 abstract ClientActorBehavior<T> lockedReconnect(ClientActorBehavior<T> current,
252 RequestException runtimeRequestException);
254 final void sendEntry(final ConnectionEntry entry, final long now) {
255 long delay = enqueueOrForward(entry, now);
257 if (delay >= DEBUG_DELAY_NANOS) {
258 if (delay > MAX_DELAY_NANOS) {
259 LOG.info("Capping {} throttle delay from {} to {} seconds", this,
260 TimeUnit.NANOSECONDS.toSeconds(delay), MAX_DELAY_SECONDS, new Throwable());
261 delay = MAX_DELAY_NANOS;
263 if (LOG.isDebugEnabled()) {
264 LOG.debug("{}: Sleeping for {}ms on connection {}", context.persistenceId(),
265 TimeUnit.NANOSECONDS.toMillis(delay), this);
268 TimeUnit.NANOSECONDS.sleep(delay);
269 } catch (InterruptedException e) {
270 Thread.currentThread().interrupt();
271 LOG.debug("Interrupted after sleeping {}ns", currentTime() - now, e);
275 final ClientActorBehavior<T> reconnect(final ClientActorBehavior<T> current, final RequestException cause) {
278 return lockedReconnect(current, cause);
285 * Schedule a timer to fire on the actor thread after a delay.
287 * @param delay Delay, in nanoseconds
290 private void scheduleTimer(final long delay) {
292 LOG.debug("{}: timer already scheduled on {}", context.persistenceId(), this);
295 if (queue.hasSuccessor()) {
296 LOG.debug("{}: connection {} has a successor, not scheduling timer", context.persistenceId(), this);
300 // If the delay is negative, we need to schedule an action immediately. While the caller could have checked
301 // for that condition and take appropriate action, but this is more convenient and less error-prone.
302 final long normalized = delay <= 0 ? 0 : Math.min(delay, context.config().getBackendAlivenessTimerInterval());
304 final FiniteDuration dur = FiniteDuration.fromNanos(normalized);
305 LOG.debug("{}: connection {} scheduling timeout in {}", context.persistenceId(), this, dur);
306 context.executeInActor(this::runTimer, dur);
311 * Check this queue for timeout and initiate reconnection if that happened. If the queue has not made progress
312 * in {@link #DEFAULT_NO_PROGRESS_TIMEOUT_NANOS} nanoseconds, it will be aborted.
314 * @param current Current behavior
315 * @return Next behavior to use
318 final ClientActorBehavior<T> runTimer(final ClientActorBehavior<T> current) {
319 final Optional<Long> delay;
324 final long now = currentTime();
326 LOG.debug("{}: running timer on {}", context.persistenceId(), this);
328 // The following line is only reliable when queue is not forwarding, but such state should not last long.
329 // FIXME: BUG-8422: this may not be accurate w.r.t. replayed entries
330 final long ticksSinceProgress = queue.ticksStalling(now);
331 if (ticksSinceProgress >= context.config().getNoProgressTimeout()) {
332 LOG.error("Queue {} has not seen progress in {} seconds, failing all requests", this,
333 TimeUnit.NANOSECONDS.toSeconds(ticksSinceProgress));
335 lockedPoison(new NoProgressException(ticksSinceProgress));
336 current.removeConnection(this);
340 // Requests are always scheduled in sequence, hence checking for timeout is relatively straightforward.
341 // Note we use also inquire about the delay, so we can re-schedule if needed, hence the unusual tri-state
342 // return convention.
343 delay = lockedCheckTimeout(now);
345 // We have timed out. There is no point in scheduling a timer
346 LOG.debug("{}: connection {} timed out", context.persistenceId(), this);
347 return lockedReconnect(current, new RuntimeRequestException("Backend connection timed out",
348 new TimeoutException()));
351 if (delay.isPresent()) {
352 // If there is new delay, schedule a timer
353 scheduleTimer(delay.get());
355 LOG.debug("{}: not scheduling timeout on {}", context.persistenceId(), this);
365 final Optional<Long> checkTimeout(final long now) {
368 return lockedCheckTimeout(now);
374 long backendSilentTicks(final long now) {
375 return now - lastReceivedTicks;
379 * We are using tri-state return here to indicate one of three conditions:
380 * - if there is no timeout to schedule, return Optional.empty()
381 * - if there is a timeout to schedule, return a non-empty optional
382 * - if this connections has timed out, return null
384 @SuppressFBWarnings(value = "NP_OPTIONAL_RETURN_NULL",
385 justification = "Returning null Optional is documented in the API contract.")
387 private Optional<Long> lockedCheckTimeout(final long now) {
388 if (queue.isEmpty()) {
389 LOG.debug("{}: connection {} is empty", context.persistenceId(), this);
390 return Optional.empty();
393 final long backendSilentTicks = backendSilentTicks(now);
394 if (backendSilentTicks >= context.config().getBackendAlivenessTimerInterval()) {
395 LOG.debug("{}: Connection {} has not seen activity from backend for {} nanoseconds, timing out",
396 context.persistenceId(), this, backendSilentTicks);
400 int tasksTimedOut = 0;
401 for (ConnectionEntry head = queue.peek(); head != null; head = queue.peek()) {
402 final long beenOpen = now - head.getEnqueuedTicks();
403 final long requestTimeout = context.config().getRequestTimeout();
404 if (beenOpen < requestTimeout) {
405 return Optional.of(requestTimeout - beenOpen);
410 LOG.debug("{}: Connection {} timed out entry {}", context.persistenceId(), this, head);
412 timeoutEntry(head, beenOpen);
415 LOG.debug("Connection {} timed out {} tasks", this, tasksTimedOut);
416 if (tasksTimedOut != 0) {
417 queue.tryTransmit(now);
420 return Optional.empty();
423 private void timeoutEntry(final ConnectionEntry entry, final long beenOpen) {
424 // Timeouts needs to be re-scheduled on actor thread because we are holding the lock on the current queue,
425 // which may be the tail of a successor chain. This is a problem if the callback attempts to send a request
426 // because that will attempt to lock the chain from the start, potentially causing a deadlock if there is
427 // a concurrent attempt to transmit.
428 context.executeInActor(current -> {
429 final double time = beenOpen * 1.0 / 1_000_000_000;
430 entry.complete(entry.getRequest().toRequestFailure(
431 new RequestTimeoutException(entry.getRequest() + " timed out after " + time
432 + " seconds. The backend for " + backendName + " is not available.")));
437 final void poison(final RequestException cause) {
447 private void lockedPoison(final RequestException cause) {
448 poisoned = enrichPoison(cause);
452 RequestException enrichPoison(final RequestException ex) {
457 final RequestException poisoned() {
461 void receiveResponse(final ResponseEnvelope<?> envelope) {
462 final long now = currentTime();
463 lastReceivedTicks = now;
465 final Optional<TransmittedConnectionEntry> maybeEntry;
468 maybeEntry = queue.complete(envelope, now);
473 if (maybeEntry.isPresent()) {
474 final TransmittedConnectionEntry entry = maybeEntry.get();
475 LOG.debug("Completing {} with {}", entry, envelope);
476 entry.complete(envelope.getMessage());
481 public final String toString() {
482 return addToStringAttributes(MoreObjects.toStringHelper(this).omitNullValues()).toString();
485 ToStringHelper addToStringAttributes(final ToStringHelper toStringHelper) {
486 return toStringHelper.add("client", context.getIdentifier()).add("cookie", cookie).add("poisoned", poisoned);