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 // Do not allow subclassing outside of this package
96 AbstractClientConnection(final ClientActorContext context, final Long cookie,
97 final TransmitQueue queue) {
98 this.context = Preconditions.checkNotNull(context);
99 this.cookie = Preconditions.checkNotNull(cookie);
100 this.queue = Preconditions.checkNotNull(queue);
101 this.lastReceivedTicks = currentTime();
104 // Do not allow subclassing outside of this package
105 AbstractClientConnection(final AbstractClientConnection<T> oldConnection, final int targetQueueSize) {
106 this.context = oldConnection.context;
107 this.cookie = oldConnection.cookie;
108 this.queue = new TransmitQueue.Halted(targetQueueSize);
109 this.lastReceivedTicks = oldConnection.lastReceivedTicks;
112 public final ClientActorContext context() {
116 public final @Nonnull Long cookie() {
120 public final ActorRef localActor() {
121 return context.self();
124 public final long currentTime() {
125 return context.ticker().read();
129 * Send a request to the backend and invoke a specified callback when it finishes. This method is safe to invoke
132 * <p>This method may put the caller thread to sleep in order to throttle the request rate.
133 * The callback may be called before the sleep finishes.
135 * @param request Request to send
136 * @param callback Callback to invoke
138 public final void sendRequest(final Request<?, ?> request, final Consumer<Response<?, ?>> callback) {
139 final long now = currentTime();
140 sendEntry(new ConnectionEntry(request, callback, now), now);
144 * Send a request to the backend and invoke a specified callback when it finishes. This method is safe to invoke
148 * Note that unlike {@link #sendRequest(Request, Consumer)}, this method does not exert backpressure, hence it
149 * should never be called from an application thread.
151 * @param request Request to send
152 * @param callback Callback to invoke
153 * @param enqueuedTicks Time (according to {@link #currentTime()} of request enqueue
155 public final void enqueueRequest(final Request<?, ?> request, final Consumer<Response<?, ?>> callback,
156 final long enqueuedTicks) {
157 enqueueEntry(new ConnectionEntry(request, callback, enqueuedTicks), currentTime());
160 public final long enqueueEntry(final ConnectionEntry entry, final long now) {
163 final RequestException maybePoison = poisoned;
164 if (maybePoison != null) {
165 throw new IllegalStateException("Connection " + this + " has been poisoned", maybePoison);
168 if (queue.isEmpty()) {
169 // The queue is becoming non-empty, schedule a timer.
170 scheduleTimer(entry.getEnqueuedTicks() + REQUEST_TIMEOUT_NANOS - now);
172 return queue.enqueue(entry, now);
178 public abstract Optional<T> getBackendInfo();
180 final Collection<ConnectionEntry> startReplay() {
182 return queue.drain();
186 final void finishReplay(final ReconnectForwarder forwarder) {
187 setForwarder(forwarder);
190 * The process of replaying all messages may have taken a significant chunk of time, depending on type
191 * of messages, queue depth and available processing power. In extreme situations this may have already
192 * exceeded BACKEND_ALIVE_TIMEOUT_NANOS, in which case we are running the risk of not making reasonable forward
193 * progress before we start a reconnect cycle.
195 * Note that the timer is armed after we have sent the first message, hence we should be seeing a response
196 * from the backend before we see a timeout, simply due to how the mailbox operates.
198 * At any rate, reset the timestamp once we complete reconnection (which an atomic transition from the
199 * perspective of outside world), as that makes it a bit easier to reason about timing of events.
201 lastReceivedTicks = currentTime();
206 final void setForwarder(final ReconnectForwarder forwarder) {
207 queue.setForwarder(forwarder, currentTime());
211 abstract ClientActorBehavior<T> lockedReconnect(ClientActorBehavior<T> current,
212 RequestException runtimeRequestException);
214 final void sendEntry(final ConnectionEntry entry, final long now) {
215 long delay = enqueueEntry(entry, now);
217 if (delay >= DEBUG_DELAY_NANOS) {
218 if (delay > MAX_DELAY_NANOS) {
219 LOG.info("Capping {} throttle delay from {} to {} seconds", this,
220 TimeUnit.NANOSECONDS.toSeconds(delay), MAX_DELAY_SECONDS);
221 delay = MAX_DELAY_NANOS;
223 if (LOG.isDebugEnabled()) {
224 LOG.debug("{}: Sleeping for {}ms on connection {}", context.persistenceId(),
225 TimeUnit.NANOSECONDS.toMillis(delay), this);
228 TimeUnit.NANOSECONDS.sleep(delay);
229 } catch (InterruptedException e) {
230 Thread.currentThread().interrupt();
231 LOG.debug("Interrupted after sleeping {}ns", e, currentTime() - now);
235 final ClientActorBehavior<T> reconnect(final ClientActorBehavior<T> current, final RequestException cause) {
238 return lockedReconnect(current, cause);
245 * Schedule a timer to fire on the actor thread after a delay.
247 * @param delay Delay, in nanoseconds
250 private void scheduleTimer(final long delay) {
252 LOG.debug("{}: timer already scheduled on {}", context.persistenceId(), this);
255 if (queue.hasSuccessor()) {
256 LOG.debug("{}: connection {} has a successor, not scheduling timer", context.persistenceId(), this);
260 // If the delay is negative, we need to schedule an action immediately. While the caller could have checked
261 // for that condition and take appropriate action, but this is more convenient and less error-prone.
262 final long normalized = delay <= 0 ? 0 : Math.min(delay, BACKEND_ALIVE_TIMEOUT_NANOS);
264 final FiniteDuration dur = FiniteDuration.fromNanos(normalized);
265 LOG.debug("{}: connection {} scheduling timeout in {}", context.persistenceId(), this, dur);
266 context.executeInActor(this::runTimer, dur);
271 * Check this queue for timeout and initiate reconnection if that happened. If the queue has not made progress
272 * in {@link #NO_PROGRESS_TIMEOUT_NANOS} nanoseconds, it will be aborted.
274 * @param current Current behavior
275 * @return Next behavior to use
278 final ClientActorBehavior<T> runTimer(final ClientActorBehavior<T> current) {
279 final Optional<Long> delay;
284 final long now = currentTime();
286 LOG.debug("{}: running timer on {}", context.persistenceId(), this);
288 // The following line is only reliable when queue is not forwarding, but such state should not last long.
289 // FIXME: BUG-8422: this may not be accurate w.r.t. replayed entries
290 final long ticksSinceProgress = queue.ticksStalling(now);
291 if (ticksSinceProgress >= NO_PROGRESS_TIMEOUT_NANOS) {
292 LOG.error("Queue {} has not seen progress in {} seconds, failing all requests", this,
293 TimeUnit.NANOSECONDS.toSeconds(ticksSinceProgress));
295 lockedPoison(new NoProgressException(ticksSinceProgress));
296 current.removeConnection(this);
300 // Requests are always scheduled in sequence, hence checking for timeout is relatively straightforward.
301 // Note we use also inquire about the delay, so we can re-schedule if needed, hence the unusual tri-state
302 // return convention.
303 delay = lockedCheckTimeout(now);
305 // We have timed out. There is no point in scheduling a timer
306 LOG.debug("{}: connection {} timed out", context.persistenceId(), this);
307 return lockedReconnect(current, new RuntimeRequestException("Backend connection timed out",
308 new TimeoutException()));
311 if (delay.isPresent()) {
312 // If there is new delay, schedule a timer
313 scheduleTimer(delay.get());
315 LOG.debug("{}: not scheduling timeout on {}", context.persistenceId(), this);
325 final Optional<Long> checkTimeout(final long now) {
328 return lockedCheckTimeout(now);
334 long backendSilentTicks(final long now) {
335 return now - lastReceivedTicks;
339 * We are using tri-state return here to indicate one of three conditions:
340 * - if there is no timeout to schedule, return Optional.empty()
341 * - if there is a timeout to schedule, return a non-empty optional
342 * - if this connections has timed out, return null
344 @SuppressFBWarnings(value = "NP_OPTIONAL_RETURN_NULL",
345 justification = "Returning null Optional is documented in the API contract.")
347 private Optional<Long> lockedCheckTimeout(final long now) {
348 if (queue.isEmpty()) {
349 LOG.debug("{}: connection {} is empty", context.persistenceId(), this);
350 return Optional.empty();
353 final long backendSilentTicks = backendSilentTicks(now);
354 if (backendSilentTicks >= BACKEND_ALIVE_TIMEOUT_NANOS) {
355 LOG.debug("{}: Connection {} has not seen activity from backend for {} nanoseconds, timing out",
356 context.persistenceId(), this, backendSilentTicks);
360 int tasksTimedOut = 0;
361 for (ConnectionEntry head = queue.peek(); head != null; head = queue.peek()) {
362 final long beenOpen = now - head.getEnqueuedTicks();
363 if (beenOpen < REQUEST_TIMEOUT_NANOS) {
364 return Optional.of(REQUEST_TIMEOUT_NANOS - beenOpen);
369 LOG.debug("{}: Connection {} timed out entry {}", context.persistenceId(), this, head);
370 head.complete(head.getRequest().toRequestFailure(
371 new RequestTimeoutException("Timed out after " + beenOpen + "ns")));
374 LOG.debug("Connection {} timed out {} tasks", this, tasksTimedOut);
375 if (tasksTimedOut != 0) {
376 queue.tryTransmit(now);
379 return Optional.empty();
382 final void poison(final RequestException cause) {
392 private void lockedPoison(final RequestException cause) {
393 poisoned = enrichPoison(cause);
397 RequestException enrichPoison(final RequestException ex) {
402 final RequestException poisoned() {
406 final void receiveResponse(final ResponseEnvelope<?> envelope) {
407 final long now = currentTime();
408 lastReceivedTicks = now;
410 final Optional<TransmittedConnectionEntry> maybeEntry;
413 maybeEntry = queue.complete(envelope, now);
418 if (maybeEntry.isPresent()) {
419 final TransmittedConnectionEntry entry = maybeEntry.get();
420 LOG.debug("Completing {} with {}", entry, envelope);
421 entry.complete(envelope.getMessage());
426 public final String toString() {
427 return addToStringAttributes(MoreObjects.toStringHelper(this).omitNullValues()).toString();
430 ToStringHelper addToStringAttributes(final ToStringHelper toStringHelper) {
431 return toStringHelper.add("client", context.getIdentifier()).add("cookie", cookie).add("poisoned", poisoned);