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 attept to reconnect
57 static final long BACKEND_ALIVE_TIMEOUT_NANOS = TimeUnit.SECONDS.toNanos(30);
60 * Request timeout. If the request fails to complete within this time since it was originally enqueued, we time
64 static final long REQUEST_TIMEOUT_NANOS = TimeUnit.MINUTES.toNanos(2);
67 * No progress timeout. A client fails to make any forward progress in this time, it will terminate itself.
70 static final long NO_PROGRESS_TIMEOUT_NANOS = TimeUnit.MINUTES.toNanos(15);
72 // Emit a debug entry if we sleep for more that this amount
73 private static final long DEBUG_DELAY_NANOS = TimeUnit.MILLISECONDS.toNanos(100);
75 // Upper bound on the time a thread is forced to sleep to keep queue size under control
76 private static final long MAX_DELAY_SECONDS = 5;
77 private static final long MAX_DELAY_NANOS = TimeUnit.SECONDS.toNanos(MAX_DELAY_SECONDS);
79 private final Lock lock = new ReentrantLock();
80 private final ClientActorContext context;
82 private final TransmitQueue queue;
83 private final Long cookie;
86 private boolean haveTimer;
89 * Time reference when we saw any activity from the backend.
91 private long lastReceivedTicks;
93 private volatile RequestException poisoned;
95 // Private constructor to avoid code duplication.
96 private AbstractClientConnection(final AbstractClientConnection<T> oldConn, final TransmitQueue newQueue) {
97 this.context = Preconditions.checkNotNull(oldConn.context);
98 this.cookie = Preconditions.checkNotNull(oldConn.cookie);
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 int queueDepth) {
107 this.context = Preconditions.checkNotNull(context);
108 this.cookie = Preconditions.checkNotNull(cookie);
109 this.queue = new TransmitQueue.Halted(queueDepth);
110 this.lastReceivedTicks = currentTime();
113 // This constructor is only to be called (indirectly) by ReconnectingClientConnection constructor.
114 // Do not allow subclassing outside of this package
115 AbstractClientConnection(final AbstractClientConnection<T> oldConn) {
116 this(oldConn, new TransmitQueue.Halted(oldConn.queue, oldConn.currentTime()));
119 // This constructor is only to be called (indirectly) by ConnectedClientConnection constructor.
120 // Do not allow subclassing outside of this package
121 AbstractClientConnection(final AbstractClientConnection<T> oldConn, final T newBackend, final int queueDepth) {
122 this(oldConn, new TransmitQueue.Transmitting(oldConn.queue, queueDepth, newBackend, oldConn.currentTime()));
125 public final ClientActorContext context() {
129 public final @Nonnull Long cookie() {
133 public final ActorRef localActor() {
134 return context.self();
137 public final long currentTime() {
138 return context.ticker().read();
142 * Send a request to the backend and invoke a specified callback when it finishes. This method is safe to invoke
145 * <p>This method may put the caller thread to sleep in order to throttle the request rate.
146 * The callback may be called before the sleep finishes.
148 * @param request Request to send
149 * @param callback Callback to invoke
151 public final void sendRequest(final Request<?, ?> request, final Consumer<Response<?, ?>> callback) {
152 final long now = currentTime();
153 sendEntry(new ConnectionEntry(request, callback, now), now);
157 * Send a request to the backend and invoke a specified callback when it finishes. This method is safe to invoke
161 * Note that unlike {@link #sendRequest(Request, Consumer)}, this method does not exert backpressure, hence it
162 * should never be called from an application thread.
164 * @param request Request to send
165 * @param callback Callback to invoke
166 * @param enqueuedTicks Time (according to {@link #currentTime()} of request enqueue
168 public final void enqueueRequest(final Request<?, ?> request, final Consumer<Response<?, ?>> callback,
169 final long enqueuedTicks) {
170 enqueueEntry(new ConnectionEntry(request, callback, enqueuedTicks), currentTime());
173 private long enqueueOrForward(final ConnectionEntry entry, final long now) {
176 commonEnqueue(entry, now);
177 return queue.enqueueOrForward(entry, now);
184 * Enqueue an entry, possibly also transmitting it.
186 public final void enqueueEntry(final ConnectionEntry entry, final long now) {
189 commonEnqueue(entry, now);
190 queue.enqueueOrReplay(entry, now);
197 private void commonEnqueue(final ConnectionEntry entry, final long now) {
198 final RequestException maybePoison = poisoned;
199 if (maybePoison != null) {
200 throw new IllegalStateException("Connection " + this + " has been poisoned", maybePoison);
203 if (queue.isEmpty()) {
204 // The queue is becoming non-empty, schedule a timer.
205 scheduleTimer(entry.getEnqueuedTicks() + REQUEST_TIMEOUT_NANOS - now);
209 // To be called from ClientActorBehavior on ConnectedClientConnection after entries are replayed.
210 final void cancelDebt() {
211 queue.cancelDebt(currentTime());
214 public abstract Optional<T> getBackendInfo();
216 final Collection<ConnectionEntry> startReplay() {
218 return queue.drain();
222 final void finishReplay(final ReconnectForwarder forwarder) {
223 setForwarder(forwarder);
226 * The process of replaying all messages may have taken a significant chunk of time, depending on type
227 * of messages, queue depth and available processing power. In extreme situations this may have already
228 * exceeded BACKEND_ALIVE_TIMEOUT_NANOS, in which case we are running the risk of not making reasonable forward
229 * progress before we start a reconnect cycle.
231 * Note that the timer is armed after we have sent the first message, hence we should be seeing a response
232 * from the backend before we see a timeout, simply due to how the mailbox operates.
234 * At any rate, reset the timestamp once we complete reconnection (which an atomic transition from the
235 * perspective of outside world), as that makes it a bit easier to reason about timing of events.
237 lastReceivedTicks = currentTime();
242 final void setForwarder(final ReconnectForwarder forwarder) {
243 queue.setForwarder(forwarder, currentTime());
247 abstract ClientActorBehavior<T> lockedReconnect(ClientActorBehavior<T> current,
248 RequestException runtimeRequestException);
250 final void sendEntry(final ConnectionEntry entry, final long now) {
251 long delay = enqueueOrForward(entry, now);
253 if (delay >= DEBUG_DELAY_NANOS) {
254 if (delay > MAX_DELAY_NANOS) {
255 LOG.info("Capping {} throttle delay from {} to {} seconds", this,
256 TimeUnit.NANOSECONDS.toSeconds(delay), MAX_DELAY_SECONDS, new Throwable());
257 delay = MAX_DELAY_NANOS;
259 if (LOG.isDebugEnabled()) {
260 LOG.debug("{}: Sleeping for {}ms on connection {}", context.persistenceId(),
261 TimeUnit.NANOSECONDS.toMillis(delay), this);
264 TimeUnit.NANOSECONDS.sleep(delay);
265 } catch (InterruptedException e) {
266 Thread.currentThread().interrupt();
267 LOG.debug("Interrupted after sleeping {}ns", e, currentTime() - now);
271 final ClientActorBehavior<T> reconnect(final ClientActorBehavior<T> current, final RequestException cause) {
274 return lockedReconnect(current, cause);
281 * Schedule a timer to fire on the actor thread after a delay.
283 * @param delay Delay, in nanoseconds
286 private void scheduleTimer(final long delay) {
288 LOG.debug("{}: timer already scheduled on {}", context.persistenceId(), this);
291 if (queue.hasSuccessor()) {
292 LOG.debug("{}: connection {} has a successor, not scheduling timer", context.persistenceId(), this);
296 // If the delay is negative, we need to schedule an action immediately. While the caller could have checked
297 // for that condition and take appropriate action, but this is more convenient and less error-prone.
298 final long normalized = delay <= 0 ? 0 : Math.min(delay, BACKEND_ALIVE_TIMEOUT_NANOS);
300 final FiniteDuration dur = FiniteDuration.fromNanos(normalized);
301 LOG.debug("{}: connection {} scheduling timeout in {}", context.persistenceId(), this, dur);
302 context.executeInActor(this::runTimer, dur);
307 * Check this queue for timeout and initiate reconnection if that happened. If the queue has not made progress
308 * in {@link #NO_PROGRESS_TIMEOUT_NANOS} nanoseconds, it will be aborted.
310 * @param current Current behavior
311 * @return Next behavior to use
314 final ClientActorBehavior<T> runTimer(final ClientActorBehavior<T> current) {
315 final Optional<Long> delay;
320 final long now = currentTime();
322 LOG.debug("{}: running timer on {}", context.persistenceId(), this);
324 // The following line is only reliable when queue is not forwarding, but such state should not last long.
325 // FIXME: BUG-8422: this may not be accurate w.r.t. replayed entries
326 final long ticksSinceProgress = queue.ticksStalling(now);
327 if (ticksSinceProgress >= NO_PROGRESS_TIMEOUT_NANOS) {
328 LOG.error("Queue {} has not seen progress in {} seconds, failing all requests", this,
329 TimeUnit.NANOSECONDS.toSeconds(ticksSinceProgress));
331 lockedPoison(new NoProgressException(ticksSinceProgress));
332 current.removeConnection(this);
336 // Requests are always scheduled in sequence, hence checking for timeout is relatively straightforward.
337 // Note we use also inquire about the delay, so we can re-schedule if needed, hence the unusual tri-state
338 // return convention.
339 delay = lockedCheckTimeout(now);
341 // We have timed out. There is no point in scheduling a timer
342 LOG.debug("{}: connection {} timed out", context.persistenceId(), this);
343 return lockedReconnect(current, new RuntimeRequestException("Backend connection timed out",
344 new TimeoutException()));
347 if (delay.isPresent()) {
348 // If there is new delay, schedule a timer
349 scheduleTimer(delay.get());
351 LOG.debug("{}: not scheduling timeout on {}", context.persistenceId(), this);
361 final Optional<Long> checkTimeout(final long now) {
364 return lockedCheckTimeout(now);
370 long backendSilentTicks(final long now) {
371 return now - lastReceivedTicks;
375 * We are using tri-state return here to indicate one of three conditions:
376 * - if there is no timeout to schedule, return Optional.empty()
377 * - if there is a timeout to schedule, return a non-empty optional
378 * - if this connections has timed out, return null
380 @SuppressFBWarnings(value = "NP_OPTIONAL_RETURN_NULL",
381 justification = "Returning null Optional is documented in the API contract.")
383 private Optional<Long> lockedCheckTimeout(final long now) {
384 if (queue.isEmpty()) {
385 LOG.debug("{}: connection {} is empty", context.persistenceId(), this);
386 return Optional.empty();
389 final long backendSilentTicks = backendSilentTicks(now);
390 if (backendSilentTicks >= BACKEND_ALIVE_TIMEOUT_NANOS) {
391 LOG.debug("{}: Connection {} has not seen activity from backend for {} nanoseconds, timing out",
392 context.persistenceId(), this, backendSilentTicks);
396 int tasksTimedOut = 0;
397 for (ConnectionEntry head = queue.peek(); head != null; head = queue.peek()) {
398 final long beenOpen = now - head.getEnqueuedTicks();
399 if (beenOpen < REQUEST_TIMEOUT_NANOS) {
400 return Optional.of(REQUEST_TIMEOUT_NANOS - beenOpen);
405 LOG.debug("{}: Connection {} timed out entry {}", context.persistenceId(), this, head);
407 timeoutEntry(head, beenOpen);
410 LOG.debug("Connection {} timed out {} tasks", this, tasksTimedOut);
411 if (tasksTimedOut != 0) {
412 queue.tryTransmit(now);
415 return Optional.empty();
418 private void timeoutEntry(final ConnectionEntry entry, final long beenOpen) {
419 // Timeouts needs to be re-scheduled on actor thread because we are holding the lock on the current queue,
420 // which may be the tail of a successor chain. This is a problem if the callback attempts to send a request
421 // because that will attempt to lock the chain from the start, potentially causing a deadlock if there is
422 // a concurrent attempt to transmit.
423 context.executeInActor(current -> {
424 final double time = beenOpen * 1.0 / 1_000_000_000;
425 entry.complete(entry.getRequest().toRequestFailure(
426 new RequestTimeoutException("Timed out after " + time + "seconds")));
431 final void poison(final RequestException cause) {
441 private void lockedPoison(final RequestException cause) {
442 poisoned = enrichPoison(cause);
446 RequestException enrichPoison(final RequestException ex) {
451 final RequestException poisoned() {
455 void receiveResponse(final ResponseEnvelope<?> envelope) {
456 final long now = currentTime();
457 lastReceivedTicks = now;
459 final Optional<TransmittedConnectionEntry> maybeEntry;
462 maybeEntry = queue.complete(envelope, now);
467 if (maybeEntry.isPresent()) {
468 final TransmittedConnectionEntry entry = maybeEntry.get();
469 LOG.debug("Completing {} with {}", entry, envelope);
470 entry.complete(envelope.getMessage());
475 public final String toString() {
476 return addToStringAttributes(MoreObjects.toStringHelper(this).omitNullValues()).toString();
479 ToStringHelper addToStringAttributes(final ToStringHelper toStringHelper) {
480 return toStringHelper.add("client", context.getIdentifier()).add("cookie", cookie).add("poisoned", poisoned);