import akka.actor.ActorRef;
import com.google.common.annotations.VisibleForTesting;
+import com.google.common.base.MoreObjects;
+import com.google.common.base.MoreObjects.ToStringHelper;
import com.google.common.base.Preconditions;
import edu.umd.cs.findbugs.annotations.SuppressFBWarnings;
+import java.util.Collection;
import java.util.Optional;
import java.util.concurrent.TimeUnit;
+import java.util.concurrent.TimeoutException;
import java.util.concurrent.locks.Lock;
import java.util.concurrent.locks.ReentrantLock;
import java.util.function.Consumer;
import org.opendaylight.controller.cluster.access.concepts.RequestException;
import org.opendaylight.controller.cluster.access.concepts.Response;
import org.opendaylight.controller.cluster.access.concepts.ResponseEnvelope;
+import org.opendaylight.controller.cluster.access.concepts.RuntimeRequestException;
import org.slf4j.Logger;
import org.slf4j.LoggerFactory;
import scala.concurrent.duration.FiniteDuration;
public abstract class AbstractClientConnection<T extends BackendInfo> {
private static final Logger LOG = LoggerFactory.getLogger(AbstractClientConnection.class);
- // Keep these constants in nanoseconds, as that prevents unnecessary conversions in the fast path
+ /*
+ * Timers involved in communication with the backend. There are three tiers which are spaced out to allow for
+ * recovery at each tier. Keep these constants in nanoseconds, as that prevents unnecessary conversions in the fast
+ * path.
+ */
+ /**
+ * Backend aliveness timer. This is reset whenever we receive a response from the backend and kept armed whenever
+ * we have an outstanding request. If when this time expires, we tear down this connection and attept to reconnect
+ * it.
+ */
@VisibleForTesting
- static final long NO_PROGRESS_TIMEOUT_NANOS = TimeUnit.MINUTES.toNanos(15);
+ static final long BACKEND_ALIVE_TIMEOUT_NANOS = TimeUnit.SECONDS.toNanos(30);
+
+ /**
+ * Request timeout. If the request fails to complete within this time since it was originally enqueued, we time
+ * the request out.
+ */
+ @VisibleForTesting
+ static final long REQUEST_TIMEOUT_NANOS = TimeUnit.MINUTES.toNanos(2);
+
+ /**
+ * No progress timeout. A client fails to make any forward progress in this time, it will terminate itself.
+ */
@VisibleForTesting
- static final long REQUEST_TIMEOUT_NANOS = TimeUnit.SECONDS.toNanos(30);
+ static final long NO_PROGRESS_TIMEOUT_NANOS = TimeUnit.MINUTES.toNanos(15);
+
+ // Emit a debug entry if we sleep for more that this amount
+ private static final long DEBUG_DELAY_NANOS = TimeUnit.MILLISECONDS.toNanos(100);
private final Lock lock = new ReentrantLock();
private final ClientActorContext context;
private final TransmitQueue queue;
private final Long cookie;
- // Updated from actor thread only
- private long lastProgress;
+ @GuardedBy("lock")
+ private boolean haveTimer;
+
+ /**
+ * Time reference when we saw any activity from the backend.
+ */
+ private long lastReceivedTicks;
private volatile RequestException poisoned;
this.context = Preconditions.checkNotNull(context);
this.cookie = Preconditions.checkNotNull(cookie);
this.queue = Preconditions.checkNotNull(queue);
- this.lastProgress = readTime();
+ this.lastReceivedTicks = currentTime();
}
// Do not allow subclassing outside of this package
- AbstractClientConnection(final AbstractClientConnection<T> oldConnection) {
+ AbstractClientConnection(final AbstractClientConnection<T> oldConnection, final int targetQueueSize) {
this.context = oldConnection.context;
this.cookie = oldConnection.cookie;
- this.lastProgress = oldConnection.lastProgress;
- this.queue = new TransmitQueue.Halted();
+ this.queue = new TransmitQueue.Halted(targetQueueSize);
+ this.lastReceivedTicks = oldConnection.lastReceivedTicks;
}
public final ClientActorContext context() {
return context.self();
}
+ public final long currentTime() {
+ return context.ticker().read();
+ }
+
/**
* Send a request to the backend and invoke a specified callback when it finishes. This method is safe to invoke
* from any thread.
*
+ * <p>This method may put the caller thread to sleep in order to throttle the request rate.
+ * The callback may be called before the sleep finishes.
+ *
* @param request Request to send
* @param callback Callback to invoke
*/
public final void sendRequest(final Request<?, ?> request, final Consumer<Response<?, ?>> callback) {
- final RequestException maybePoison = poisoned;
- if (maybePoison != null) {
- throw new IllegalStateException("Connection " + this + " has been poisoned", maybePoison);
- }
-
- final ConnectionEntry entry = new ConnectionEntry(request, callback, readTime());
-
- lock.lock();
+ final long now = currentTime();
+ final long delay = enqueueEntry(new ConnectionEntry(request, callback, now), now);
try {
- queue.enqueue(entry, entry.getEnqueuedTicks());
- } finally {
- lock.unlock();
+ if (delay >= DEBUG_DELAY_NANOS && LOG.isDebugEnabled()) {
+ LOG.debug("Sleeping for {}ms", TimeUnit.NANOSECONDS.toMillis(delay));
+ }
+ TimeUnit.NANOSECONDS.sleep(delay);
+ } catch (InterruptedException e) {
+ Thread.currentThread().interrupt();
+ LOG.debug("Interrupted after sleeping {}ns", e, currentTime() - now);
}
}
+ /**
+ * Send a request to the backend and invoke a specified callback when it finishes. This method is safe to invoke
+ * from any thread.
+ *
+ * <p>
+ * Note that unlike {@link #sendRequest(Request, Consumer)}, this method does not exert backpressure, hence it
+ * should never be called from an application thread.
+ *
+ * @param request Request to send
+ * @param callback Callback to invoke
+ * @param enqueuedTicks Time (according to {@link #currentTime()} of request enqueue
+ */
+ public final void enqueueRequest(final Request<?, ?> request, final Consumer<Response<?, ?>> callback,
+ final long enqueuedTicks) {
+ enqueueEntry(new ConnectionEntry(request, callback, enqueuedTicks), currentTime());
+ }
+
public abstract Optional<T> getBackendInfo();
- final Iterable<ConnectionEntry> startReplay() {
+ final Collection<ConnectionEntry> startReplay() {
lock.lock();
- return queue.asIterable();
+ return queue.drain();
}
@GuardedBy("lock")
final void finishReplay(final ReconnectForwarder forwarder) {
- queue.setForwarder(forwarder, readTime());
+ setForwarder(forwarder);
lock.unlock();
}
@GuardedBy("lock")
final void setForwarder(final ReconnectForwarder forwarder) {
- queue.setForwarder(forwarder, readTime());
+ queue.setForwarder(forwarder, currentTime());
}
@GuardedBy("lock")
- abstract ClientActorBehavior<T> reconnectConnection(ClientActorBehavior<T> current);
+ abstract ClientActorBehavior<T> lockedReconnect(ClientActorBehavior<T> current,
+ RequestException runtimeRequestException);
- private long readTime() {
- return context.ticker().read();
+ final long enqueueEntry(final ConnectionEntry entry, final long now) {
+ lock.lock();
+ try {
+ final RequestException maybePoison = poisoned;
+ if (maybePoison != null) {
+ throw new IllegalStateException("Connection " + this + " has been poisoned", maybePoison);
+ }
+
+ if (queue.isEmpty()) {
+ // The queue is becoming non-empty, schedule a timer.
+ scheduleTimer(entry.getEnqueuedTicks() + REQUEST_TIMEOUT_NANOS - now);
+ }
+ return queue.enqueue(entry, now);
+ } finally {
+ lock.unlock();
+ }
}
- final void enqueueEntry(final ConnectionEntry entry, final long now) {
+ final ClientActorBehavior<T> reconnect(final ClientActorBehavior<T> current, final RequestException cause) {
lock.lock();
try {
- queue.enqueue(entry, now);
+ return lockedReconnect(current, cause);
} finally {
lock.unlock();
}
*
* @param delay Delay, in nanoseconds
*/
- private void scheduleTimer(final FiniteDuration delay) {
- LOG.debug("{}: scheduling timeout in {}", context.persistenceId(), delay);
- context.executeInActor(this::runTimer, delay);
+ @GuardedBy("lock")
+ private void scheduleTimer(final long delay) {
+ if (haveTimer) {
+ LOG.debug("{}: timer already scheduled", context.persistenceId());
+ return;
+ }
+ if (queue.hasSuccessor()) {
+ LOG.debug("{}: connection has successor, not scheduling timer", context.persistenceId());
+ return;
+ }
+
+ // If the delay is negative, we need to schedule an action immediately. While the caller could have checked
+ // for that condition and take appropriate action, but this is more convenient and less error-prone.
+ final long normalized = delay <= 0 ? 0 : Math.min(delay, BACKEND_ALIVE_TIMEOUT_NANOS);
+
+ final FiniteDuration dur = FiniteDuration.fromNanos(normalized);
+ LOG.debug("{}: scheduling timeout in {}", context.persistenceId(), dur);
+ context.executeInActor(this::runTimer, dur);
+ haveTimer = true;
}
/**
*/
@VisibleForTesting
final ClientActorBehavior<T> runTimer(final ClientActorBehavior<T> current) {
- final Optional<FiniteDuration> delay;
+ final Optional<Long> delay;
lock.lock();
try {
- final long now = readTime();
- if (!queue.isEmpty()) {
- final long ticksSinceProgress = now - lastProgress;
- if (ticksSinceProgress >= NO_PROGRESS_TIMEOUT_NANOS) {
- LOG.error("Queue {} has not seen progress in {} seconds, failing all requests", this,
- TimeUnit.NANOSECONDS.toSeconds(ticksSinceProgress));
-
- lockedPoison(new NoProgressException(ticksSinceProgress));
- current.removeConnection(this);
- return current;
- }
+ haveTimer = false;
+ final long now = currentTime();
+ // The following line is only reliable when queue is not forwarding, but such state should not last long.
+ // FIXME: BUG-8422: this may not be accurate w.r.t. replayed entries
+ final long ticksSinceProgress = queue.ticksStalling(now);
+ if (ticksSinceProgress >= NO_PROGRESS_TIMEOUT_NANOS) {
+ LOG.error("Queue {} has not seen progress in {} seconds, failing all requests", this,
+ TimeUnit.NANOSECONDS.toSeconds(ticksSinceProgress));
+
+ lockedPoison(new NoProgressException(ticksSinceProgress));
+ current.removeConnection(this);
+ return current;
}
// Requests are always scheduled in sequence, hence checking for timeout is relatively straightforward.
delay = lockedCheckTimeout(now);
if (delay == null) {
// We have timed out. There is no point in scheduling a timer
- return reconnectConnection(current);
+ return lockedReconnect(current, new RuntimeRequestException("Backend connection timed out",
+ new TimeoutException()));
+ }
+
+ if (delay.isPresent()) {
+ // If there is new delay, schedule a timer
+ scheduleTimer(delay.get());
}
} finally {
lock.unlock();
}
- if (delay.isPresent()) {
- // If there is new delay, schedule a timer
- scheduleTimer(delay.get());
- }
-
return current;
}
@VisibleForTesting
- final Optional<FiniteDuration> checkTimeout(final long now) {
+ final Optional<Long> checkTimeout(final long now) {
lock.lock();
try {
return lockedCheckTimeout(now);
}
}
+ long backendSilentTicks(final long now) {
+ return now - lastReceivedTicks;
+ }
+
/*
* We are using tri-state return here to indicate one of three conditions:
* - if there is no timeout to schedule, return Optional.empty()
@SuppressFBWarnings(value = "NP_OPTIONAL_RETURN_NULL",
justification = "Returning null Optional is documented in the API contract.")
@GuardedBy("lock")
- private Optional<FiniteDuration> lockedCheckTimeout(final long now) {
- final ConnectionEntry head = queue.peek();
- if (head == null) {
+ private Optional<Long> lockedCheckTimeout(final long now) {
+ if (queue.isEmpty()) {
return Optional.empty();
}
- final long delay = head.getEnqueuedTicks() - now + REQUEST_TIMEOUT_NANOS;
- if (delay <= 0) {
- LOG.debug("Connection {} timed out", this);
+ final long backendSilentTicks = backendSilentTicks(now);
+ if (backendSilentTicks >= BACKEND_ALIVE_TIMEOUT_NANOS) {
+ LOG.debug("Connection {} has not seen activity from backend for {} nanoseconds, timing out", this,
+ backendSilentTicks);
return null;
}
- return Optional.of(FiniteDuration.apply(delay, TimeUnit.NANOSECONDS));
+ int tasksTimedOut = 0;
+ for (ConnectionEntry head = queue.peek(); head != null; head = queue.peek()) {
+ final long beenOpen = now - head.getEnqueuedTicks();
+ if (beenOpen < REQUEST_TIMEOUT_NANOS) {
+ return Optional.of(REQUEST_TIMEOUT_NANOS - beenOpen);
+ }
+
+ tasksTimedOut++;
+ queue.remove(now);
+ LOG.debug("Connection {} timed out entryt {}", this, head);
+ head.complete(head.getRequest().toRequestFailure(
+ new RequestTimeoutException("Timed out after " + beenOpen + "ns")));
+ }
+
+ LOG.debug("Connection {} timed out {} tasks", this, tasksTimedOut);
+ if (tasksTimedOut != 0) {
+ queue.tryTransmit(now);
+ }
+
+ return Optional.empty();
}
final void poison(final RequestException cause) {
@GuardedBy("lock")
private void lockedPoison(final RequestException cause) {
- poisoned = cause;
+ poisoned = enrichPoison(cause);
queue.poison(cause);
}
+ RequestException enrichPoison(final RequestException ex) {
+ return ex;
+ }
+
@VisibleForTesting
final RequestException poisoned() {
return poisoned;
}
final void receiveResponse(final ResponseEnvelope<?> envelope) {
- final long now = readTime();
+ final long now = currentTime();
+ lastReceivedTicks = now;
final Optional<TransmittedConnectionEntry> maybeEntry;
lock.lock();
LOG.debug("Completing {} with {}", entry, envelope);
entry.complete(envelope.getMessage());
}
+ }
+
+ @Override
+ public final String toString() {
+ return addToStringAttributes(MoreObjects.toStringHelper(this).omitNullValues()).toString();
+ }
- lastProgress = readTime();
+ ToStringHelper addToStringAttributes(final ToStringHelper toStringHelper) {
+ return toStringHelper.add("client", context.getIdentifier()).add("cookie", cookie).add("poisoned", poisoned);
}
}