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 static java.util.Objects.requireNonNull;
12 import akka.actor.ActorRef;
13 import com.google.common.annotations.VisibleForTesting;
14 import com.google.common.base.MoreObjects;
15 import com.google.common.base.MoreObjects.ToStringHelper;
16 import edu.umd.cs.findbugs.annotations.SuppressFBWarnings;
17 import java.util.Collection;
18 import java.util.List;
19 import java.util.Optional;
20 import java.util.concurrent.TimeUnit;
21 import java.util.concurrent.TimeoutException;
22 import java.util.concurrent.locks.Lock;
23 import java.util.concurrent.locks.ReentrantLock;
24 import java.util.function.Consumer;
25 import org.checkerframework.checker.lock.qual.GuardedBy;
26 import org.checkerframework.checker.lock.qual.Holding;
27 import org.eclipse.jdt.annotation.NonNull;
28 import org.opendaylight.controller.cluster.access.concepts.Request;
29 import org.opendaylight.controller.cluster.access.concepts.RequestException;
30 import org.opendaylight.controller.cluster.access.concepts.Response;
31 import org.opendaylight.controller.cluster.access.concepts.ResponseEnvelope;
32 import org.opendaylight.controller.cluster.access.concepts.RuntimeRequestException;
33 import org.slf4j.Logger;
34 import org.slf4j.LoggerFactory;
35 import scala.concurrent.duration.FiniteDuration;
38 * Base class for a connection to the backend. Responsible to queueing and dispatch of requests toward the backend.
39 * Can be in three conceptual states: Connecting, Connected and Reconnecting, which are represented by public final
40 * classes exposed from this package. This class NOT thread-safe, not are its subclasses expected to be thread-safe.
42 * @author Robert Varga
44 public abstract class AbstractClientConnection<T extends BackendInfo> {
45 private static final Logger LOG = LoggerFactory.getLogger(AbstractClientConnection.class);
48 * Timers involved in communication with the backend. There are three tiers which are spaced out to allow for
49 * recovery at each tier. Keep these constants in nanoseconds, as that prevents unnecessary conversions in the fast
53 * Backend aliveness timer. This is reset whenever we receive a response from the backend and kept armed whenever
54 * we have an outstanding request. If when this time expires, we tear down this connection and attempt to reconnect
57 public static final long DEFAULT_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
63 public static final long DEFAULT_REQUEST_TIMEOUT_NANOS = TimeUnit.MINUTES.toNanos(2);
66 * No progress timeout. A client fails to make any forward progress in this time, it will terminate itself.
68 public static final long DEFAULT_NO_PROGRESS_TIMEOUT_NANOS = TimeUnit.MINUTES.toNanos(15);
70 // Emit a debug entry if we sleep for more that this amount
71 private static final long DEBUG_DELAY_NANOS = TimeUnit.MILLISECONDS.toNanos(100);
73 // Upper bound on the time a thread is forced to sleep to keep queue size under control
74 private static final long MAX_DELAY_SECONDS = 5;
75 private static final long MAX_DELAY_NANOS = TimeUnit.SECONDS.toNanos(MAX_DELAY_SECONDS);
77 private final Lock lock = new ReentrantLock();
78 private final ClientActorContext context;
80 private final TransmitQueue queue;
81 private final @NonNull Long cookie;
82 private final String backendName;
85 private boolean haveTimer;
88 * Time reference when we saw any activity from the backend.
90 private long lastReceivedTicks;
92 private volatile RequestException poisoned;
94 // Private constructor to avoid code duplication.
95 private AbstractClientConnection(final AbstractClientConnection<T> oldConn, final TransmitQueue newQueue,
96 final String backendName) {
97 this.context = requireNonNull(oldConn.context);
98 this.cookie = requireNonNull(oldConn.cookie);
99 this.backendName = requireNonNull(backendName);
100 this.queue = requireNonNull(newQueue);
101 // Will be updated in finishReplay if needed.
102 this.lastReceivedTicks = oldConn.lastReceivedTicks;
105 // This constructor is only to be called by ConnectingClientConnection constructor.
106 // Do not allow subclassing outside of this package
107 AbstractClientConnection(final ClientActorContext context, final Long cookie, final String backendName,
108 final int queueDepth) {
109 this.context = requireNonNull(context);
110 this.cookie = requireNonNull(cookie);
111 this.backendName = requireNonNull(backendName);
112 this.queue = new TransmitQueue.Halted(queueDepth);
113 this.lastReceivedTicks = currentTime();
116 // This constructor is only to be called (indirectly) by ReconnectingClientConnection constructor.
117 // Do not allow subclassing outside of this package
118 AbstractClientConnection(final AbstractClientConnection<T> oldConn) {
119 this(oldConn, new TransmitQueue.Halted(oldConn.queue, oldConn.currentTime()), oldConn.backendName);
122 // This constructor is only to be called (indirectly) by ConnectedClientConnection constructor.
123 // Do not allow subclassing outside of this package
124 AbstractClientConnection(final AbstractClientConnection<T> oldConn, final T newBackend,
125 final int queueDepth) {
126 this(oldConn, new TransmitQueue.Transmitting(oldConn.queue, queueDepth, newBackend, oldConn.currentTime(),
127 requireNonNull(oldConn.context).messageSlicer()), newBackend.getName());
130 public final ClientActorContext context() {
134 public final @NonNull Long cookie() {
138 public final ActorRef localActor() {
139 return context.self();
142 public final long currentTime() {
143 return context.ticker().read();
147 * Send a request to the backend and invoke a specified callback when it finishes. This method is safe to invoke
150 * <p>This method may put the caller thread to sleep in order to throttle the request rate.
151 * The callback may be called before the sleep finishes.
153 * @param request Request to send
154 * @param callback Callback to invoke
156 public final void sendRequest(final Request<?, ?> request, final Consumer<Response<?, ?>> callback) {
157 final long now = currentTime();
158 sendEntry(new ConnectionEntry(request, callback, now), now);
162 * Send a request to the backend and invoke a specified callback when it finishes. This method is safe to invoke
166 * Note that unlike {@link #sendRequest(Request, Consumer)}, this method does not exert backpressure, hence it
167 * should never be called from an application thread.
169 * @param request Request to send
170 * @param callback Callback to invoke
171 * @param enqueuedTicks Time (according to {@link #currentTime()} of request enqueue
173 public final void enqueueRequest(final Request<?, ?> request, final Consumer<Response<?, ?>> callback,
174 final long enqueuedTicks) {
175 enqueueEntry(new ConnectionEntry(request, callback, enqueuedTicks), currentTime());
178 private long enqueueOrForward(final ConnectionEntry entry, final long now) {
181 commonEnqueue(entry, now);
182 return queue.enqueueOrForward(entry, now);
189 * Enqueue an entry, possibly also transmitting it.
191 public final void enqueueEntry(final ConnectionEntry entry, final long now) {
194 commonEnqueue(entry, now);
195 queue.enqueueOrReplay(entry, now);
202 private void commonEnqueue(final ConnectionEntry entry, final long now) {
203 final RequestException maybePoison = poisoned;
204 if (maybePoison != null) {
205 throw new IllegalStateException("Connection " + this + " has been poisoned", maybePoison);
208 if (queue.isEmpty()) {
209 // The queue is becoming non-empty, schedule a timer.
210 scheduleTimer(entry.getEnqueuedTicks() + context.config().getRequestTimeout() - now);
214 // To be called from ClientActorBehavior on ConnectedClientConnection after entries are replayed.
215 final void cancelDebt() {
216 queue.cancelDebt(currentTime());
219 public abstract Optional<T> getBackendInfo();
221 final Collection<ConnectionEntry> startReplay() {
223 return queue.drain();
227 final void finishReplay(final ReconnectForwarder forwarder) {
228 setForwarder(forwarder);
231 * The process of replaying all messages may have taken a significant chunk of time, depending on type
232 * of messages, queue depth and available processing power. In extreme situations this may have already
233 * exceeded BACKEND_ALIVE_TIMEOUT_NANOS, in which case we are running the risk of not making reasonable forward
234 * progress before we start a reconnect cycle.
236 * Note that the timer is armed after we have sent the first message, hence we should be seeing a response
237 * from the backend before we see a timeout, simply due to how the mailbox operates.
239 * At any rate, reset the timestamp once we complete reconnection (which an atomic transition from the
240 * perspective of outside world), as that makes it a bit easier to reason about timing of events.
242 lastReceivedTicks = currentTime();
247 final void setForwarder(final ReconnectForwarder forwarder) {
248 queue.setForwarder(forwarder, currentTime());
252 abstract ClientActorBehavior<T> lockedReconnect(ClientActorBehavior<T> current,
253 RequestException runtimeRequestException);
255 final void sendEntry(final ConnectionEntry entry, final long now) {
256 long delay = enqueueOrForward(entry, now);
258 if (delay >= DEBUG_DELAY_NANOS) {
259 if (delay > MAX_DELAY_NANOS) {
260 LOG.info("Capping {} throttle delay from {} to {} seconds", this,
261 TimeUnit.NANOSECONDS.toSeconds(delay), MAX_DELAY_SECONDS, new Throwable());
262 delay = MAX_DELAY_NANOS;
264 if (LOG.isDebugEnabled()) {
265 LOG.debug("{}: Sleeping for {}ms on connection {}", context.persistenceId(),
266 TimeUnit.NANOSECONDS.toMillis(delay), this);
269 TimeUnit.NANOSECONDS.sleep(delay);
270 } catch (InterruptedException e) {
271 Thread.currentThread().interrupt();
272 LOG.debug("Interrupted after sleeping {}ns", currentTime() - now, e);
276 final ClientActorBehavior<T> reconnect(final ClientActorBehavior<T> current, final RequestException cause) {
279 return lockedReconnect(current, cause);
286 * Schedule a timer to fire on the actor thread after a delay.
288 * @param delay Delay, in nanoseconds
291 private void scheduleTimer(final long delay) {
293 LOG.debug("{}: timer already scheduled on {}", context.persistenceId(), this);
296 if (queue.hasSuccessor()) {
297 LOG.debug("{}: connection {} has a successor, not scheduling timer", context.persistenceId(), this);
301 // If the delay is negative, we need to schedule an action immediately. While the caller could have checked
302 // for that condition and take appropriate action, but this is more convenient and less error-prone.
303 final long normalized = delay <= 0 ? 0 : Math.min(delay, context.config().getBackendAlivenessTimerInterval());
305 final FiniteDuration dur = FiniteDuration.fromNanos(normalized);
306 LOG.debug("{}: connection {} scheduling timeout in {}", context.persistenceId(), this, dur);
307 context.executeInActor(this::runTimer, dur);
312 * Check this queue for timeout and initiate reconnection if that happened. If the queue has not made progress
313 * in {@link #DEFAULT_NO_PROGRESS_TIMEOUT_NANOS} nanoseconds, it will be aborted.
315 * @param current Current behavior
316 * @return Next behavior to use
319 final ClientActorBehavior<T> runTimer(final ClientActorBehavior<T> current) {
322 final List<ConnectionEntry> poisonEntries;
323 final NoProgressException poisonCause;
326 final long now = currentTime();
328 LOG.debug("{}: running timer on {}", context.persistenceId(), this);
330 // The following line is only reliable when queue is not forwarding, but such state should not last long.
331 // FIXME: BUG-8422: this may not be accurate w.r.t. replayed entries
332 final long ticksSinceProgress = queue.ticksStalling(now);
333 if (ticksSinceProgress < context.config().getNoProgressTimeout()) {
334 // Requests are always scheduled in sequence, hence checking for timeout is relatively straightforward.
335 // Note we use also inquire about the delay, so we can re-schedule if needed, hence the unusual
336 // tri-state return convention.
337 final Optional<Long> delay = lockedCheckTimeout(now);
339 // We have timed out. There is no point in scheduling a timer
340 LOG.debug("{}: connection {} timed out", context.persistenceId(), this);
341 return lockedReconnect(current, new RuntimeRequestException("Backend connection timed out",
342 new TimeoutException()));
345 if (delay.isPresent()) {
346 // If there is new delay, schedule a timer
347 scheduleTimer(delay.get());
349 LOG.debug("{}: not scheduling timeout on {}", context.persistenceId(), this);
355 LOG.error("Queue {} has not seen progress in {} seconds, failing all requests", this,
356 TimeUnit.NANOSECONDS.toSeconds(ticksSinceProgress));
357 poisonCause = new NoProgressException(ticksSinceProgress);
358 poisonEntries = lockedPoison(poisonCause);
359 current.removeConnection(this);
364 poison(poisonEntries, poisonCause);
369 final Optional<Long> checkTimeout(final long now) {
372 return lockedCheckTimeout(now);
378 long backendSilentTicks(final long now) {
379 return now - lastReceivedTicks;
383 * We are using tri-state return here to indicate one of three conditions:
384 * - if there is no timeout to schedule, return Optional.empty()
385 * - if there is a timeout to schedule, return a non-empty optional
386 * - if this connections has timed out, return null
388 @SuppressFBWarnings(value = "NP_OPTIONAL_RETURN_NULL",
389 justification = "Returning null Optional is documented in the API contract.")
391 private Optional<Long> lockedCheckTimeout(final long now) {
392 if (queue.isEmpty()) {
393 LOG.debug("{}: connection {} is empty", context.persistenceId(), this);
394 return Optional.empty();
397 final long backendSilentTicks = backendSilentTicks(now);
398 if (backendSilentTicks >= context.config().getBackendAlivenessTimerInterval()) {
399 LOG.debug("{}: Connection {} has not seen activity from backend for {} nanoseconds, timing out",
400 context.persistenceId(), this, backendSilentTicks);
404 int tasksTimedOut = 0;
405 for (ConnectionEntry head = queue.peek(); head != null; head = queue.peek()) {
406 final long beenOpen = now - head.getEnqueuedTicks();
407 final long requestTimeout = context.config().getRequestTimeout();
408 if (beenOpen < requestTimeout) {
409 return Optional.of(requestTimeout - beenOpen);
414 LOG.debug("{}: Connection {} timed out entry {}", context.persistenceId(), this, head);
416 timeoutEntry(head, beenOpen);
419 LOG.debug("Connection {} timed out {} tasks", this, tasksTimedOut);
420 if (tasksTimedOut != 0) {
421 queue.tryTransmit(now);
424 return Optional.empty();
427 private void timeoutEntry(final ConnectionEntry entry, final long beenOpen) {
428 // Timeouts needs to be re-scheduled on actor thread because we are holding the lock on the current queue,
429 // which may be the tail of a successor chain. This is a problem if the callback attempts to send a request
430 // because that will attempt to lock the chain from the start, potentially causing a deadlock if there is
431 // a concurrent attempt to transmit.
432 context.executeInActor(current -> {
433 final double time = beenOpen * 1.0 / 1_000_000_000;
434 entry.complete(entry.getRequest().toRequestFailure(
435 new RequestTimeoutException(entry.getRequest() + " timed out after " + time
436 + " seconds. The backend for " + backendName + " is not available.")));
441 final void poison(final RequestException cause) {
442 final List<ConnectionEntry> entries;
446 entries = lockedPoison(cause);
451 poison(entries, cause);
454 // Do not hold any locks while calling this
455 private static void poison(final Collection<? extends ConnectionEntry> entries, final RequestException cause) {
456 for (ConnectionEntry e : entries) {
457 final Request<?, ?> request = e.getRequest();
458 LOG.trace("Poisoning request {}", request, cause);
459 e.complete(request.toRequestFailure(cause));
464 private List<ConnectionEntry> lockedPoison(final RequestException cause) {
465 poisoned = enrichPoison(cause);
466 return queue.poison();
469 RequestException enrichPoison(final RequestException ex) {
474 final RequestException poisoned() {
478 void receiveResponse(final ResponseEnvelope<?> envelope) {
479 final long now = currentTime();
480 lastReceivedTicks = now;
482 final Optional<TransmittedConnectionEntry> maybeEntry;
485 maybeEntry = queue.complete(envelope, now);
490 if (maybeEntry.isPresent()) {
491 final TransmittedConnectionEntry entry = maybeEntry.get();
492 LOG.debug("Completing {} with {}", entry, envelope);
493 entry.complete(envelope.getMessage());
498 public final String toString() {
499 return addToStringAttributes(MoreObjects.toStringHelper(this).omitNullValues()).toString();
502 ToStringHelper addToStringAttributes(final ToStringHelper toStringHelper) {
503 return toStringHelper.add("client", context.getIdentifier()).add("cookie", cookie).add("poisoned", poisoned);