X-Git-Url: https://git.opendaylight.org/gerrit/gitweb?p=controller.git;a=blobdiff_plain;f=opendaylight%2Fmd-sal%2Fcds-access-client%2Fsrc%2Fmain%2Fjava%2Forg%2Fopendaylight%2Fcontroller%2Fcluster%2Faccess%2Fclient%2FAbstractClientConnection.java;h=da016bae885be4ea9fc15035f8e05f6e62c6d058;hp=170b1507a91254a6c7c13e2297158a6914e62b59;hb=b74c6012092e47430a8f4d6f4ddeb1d3e2b1b7df;hpb=320a4e5cd2d9d80468a3f82798744f2035488218 diff --git a/opendaylight/md-sal/cds-access-client/src/main/java/org/opendaylight/controller/cluster/access/client/AbstractClientConnection.java b/opendaylight/md-sal/cds-access-client/src/main/java/org/opendaylight/controller/cluster/access/client/AbstractClientConnection.java index 170b1507a9..da016bae88 100644 --- a/opendaylight/md-sal/cds-access-client/src/main/java/org/opendaylight/controller/cluster/access/client/AbstractClientConnection.java +++ b/opendaylight/md-sal/cds-access-client/src/main/java/org/opendaylight/controller/cluster/access/client/AbstractClientConnection.java @@ -9,13 +9,15 @@ package org.opendaylight.controller.cluster.access.client; 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 com.google.common.base.Verify; import edu.umd.cs.findbugs.annotations.SuppressFBWarnings; -import java.util.ArrayDeque; +import java.util.Collection; import java.util.Optional; -import java.util.Queue; import java.util.concurrent.TimeUnit; +import java.util.concurrent.locks.Lock; +import java.util.concurrent.locks.ReentrantLock; import java.util.function.Consumer; import javax.annotation.Nonnull; import javax.annotation.concurrent.GuardedBy; @@ -39,36 +41,63 @@ import scala.concurrent.duration.FiniteDuration; public abstract class AbstractClientConnection { 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.SECONDS.toNanos(30); + static final long REQUEST_TIMEOUT_NANOS = TimeUnit.MINUTES.toNanos(2); - private final Queue pending; + /** + * No progress timeout. A client fails to make any forward progress in this time, it will terminate itself. + */ + @VisibleForTesting + static final long NO_PROGRESS_TIMEOUT_NANOS = TimeUnit.MINUTES.toNanos(15); + + private final Lock lock = new ReentrantLock(); private final ClientActorContext context; + @GuardedBy("lock") + private final TransmitQueue queue; private final Long cookie; - private volatile ReconnectForwarder successor; + @GuardedBy("lock") + private boolean haveTimer; + + /** + * Time reference when we saw any activity from the backend. + */ + private long lastReceivedTicks; + private volatile RequestException poisoned; - private long lastProgress; - private AbstractClientConnection(final ClientActorContext context, final Long cookie, - final Queue pending) { + // Do not allow subclassing outside of this package + AbstractClientConnection(final ClientActorContext context, final Long cookie, + final TransmitQueue queue) { this.context = Preconditions.checkNotNull(context); this.cookie = Preconditions.checkNotNull(cookie); - this.pending = Preconditions.checkNotNull(pending); - this.lastProgress = readTime(); - } - - // Do not allow subclassing outside of this package - AbstractClientConnection(final ClientActorContext context, final Long cookie) { - this(context, cookie, new ArrayDeque<>(1)); + this.queue = Preconditions.checkNotNull(queue); + this.lastReceivedTicks = currentTime(); } // Do not allow subclassing outside of this package - AbstractClientConnection(final AbstractClientConnection oldConnection) { - this(oldConnection.context, oldConnection.cookie, oldConnection.pending); + AbstractClientConnection(final AbstractClientConnection oldConnection, final int targetQueueSize) { + this.context = oldConnection.context; + this.cookie = oldConnection.cookie; + this.queue = new TransmitQueue.Halted(targetQueueSize); + this.lastReceivedTicks = oldConnection.lastReceivedTicks; } public final ClientActorContext context() { @@ -83,67 +112,94 @@ public abstract class AbstractClientConnection { return context.self(); } - final long readTime() { + public final long currentTime() { return context.ticker().read(); } - final Queue pending() { - return pending; + /** + * Send a request to the backend and invoke a specified callback when it finishes. This method is safe to invoke + * from any thread. + * + *

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> callback) { + final long now = currentTime(); + final long delay = enqueueEntry(new ConnectionEntry(request, callback, now), now); + try { + 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. * + *

+ * 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 sendRequest(final Request request, final Consumer> callback) { - Preconditions.checkState(poisoned == null, "Connection %s has been poisoned", this); + public final void enqueueRequest(final Request request, final Consumer> callback, + final long enqueuedTicks) { + enqueueEntry(new ConnectionEntry(request, callback, enqueuedTicks), currentTime()); + } - final ReconnectForwarder beforeQueue = successor; - final ConnectionEntry entry = new ConnectionEntry(request, callback, readTime()); - if (beforeQueue != null) { - LOG.trace("Forwarding entry {} from {} to {}", entry, this, beforeQueue); - beforeQueue.forwardEntry(entry); - return; - } + public abstract Optional getBackendInfo(); - enqueueEntry(entry); + final Collection startReplay() { + lock.lock(); + return queue.drain(); + } - final ReconnectForwarder afterQueue = successor; - if (afterQueue != null) { - synchronized (this) { - spliceToSuccessor(afterQueue); - } - } + @GuardedBy("lock") + final void finishReplay(final ReconnectForwarder forwarder) { + setForwarder(forwarder); + lock.unlock(); } - public final synchronized void setForwarder(final ReconnectForwarder forwarder) { - Verify.verify(successor == null, "Successor {} already set on connection {}", successor, this); - successor = Preconditions.checkNotNull(forwarder); - LOG.debug("Connection {} superseded by {}, splicing queue", this, successor); - spliceToSuccessor(forwarder); + @GuardedBy("lock") + final void setForwarder(final ReconnectForwarder forwarder) { + queue.setForwarder(forwarder, currentTime()); } - public abstract Optional getBackendInfo(); + @GuardedBy("lock") + abstract ClientActorBehavior lockedReconnect(ClientActorBehavior current); - @GuardedBy("this") - void spliceToSuccessor(final ReconnectForwarder successor) { - ConnectionEntry entry = pending.poll(); - while (entry != null) { - successor.forwardEntry(entry); - entry = pending.poll(); - } - } + 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); + } - final ConnectionEntry dequeEntry() { - lastProgress = readTime(); - return pending.poll(); + 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(); + } } - void enqueueEntry(final ConnectionEntry entry) { - pending.add(entry); + final ClientActorBehavior reconnect(final ClientActorBehavior current) { + lock.lock(); + try { + return lockedReconnect(current); + } finally { + lock.unlock(); + } } /** @@ -151,53 +207,82 @@ public abstract class AbstractClientConnection { * * @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; } /** - * Check queue timeouts and return true if a timeout has occurred. + * Check this queue for timeout and initiate reconnection if that happened. If the queue has not made progress + * in {@link #NO_PROGRESS_TIMEOUT_NANOS} nanoseconds, it will be aborted. * - * @return True if a timeout occurred - * @throws NoProgressException if the queue failed to make progress for an extended - * time. + * @param current Current behavior + * @return Next behavior to use */ @VisibleForTesting final ClientActorBehavior runTimer(final ClientActorBehavior current) { - final long now = readTime(); - - if (!isEmpty()) { - final long ticksSinceProgress = now - lastProgress; + final Optional delay; + + lock.lock(); + try { + 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)); - poison(new NoProgressException(ticksSinceProgress)); + lockedPoison(new NoProgressException(ticksSinceProgress)); current.removeConnection(this); return current; } - } - // Requests are always scheduled in sequence, hence checking for timeout is relatively straightforward. - // Note we use also inquire about the delay, so we can re-schedule if needed, hence the unusual tri-state - // return convention. - final Optional delay = checkTimeout(now); - if (delay == null) { - // We have timed out. There is no point in scheduling a timer - return reconnectConnection(current); - } + // Requests are always scheduled in sequence, hence checking for timeout is relatively straightforward. + // Note we use also inquire about the delay, so we can re-schedule if needed, hence the unusual tri-state + // return convention. + delay = lockedCheckTimeout(now); + if (delay == null) { + // We have timed out. There is no point in scheduling a timer + return lockedReconnect(current); + } - if (delay.isPresent()) { - // If there is new delay, schedule a timer - scheduleTimer(delay.get()); + if (delay.isPresent()) { + // If there is new delay, schedule a timer + scheduleTimer(delay.get()); + } + } finally { + lock.unlock(); } return current; } - boolean isEmpty() { - return pending.isEmpty(); + @VisibleForTesting + final Optional checkTimeout(final long now) { + lock.lock(); + try { + return lockedCheckTimeout(now); + } finally { + lock.unlock(); + } } /* @@ -208,44 +293,54 @@ public abstract class AbstractClientConnection { */ @SuppressFBWarnings(value = "NP_OPTIONAL_RETURN_NULL", justification = "Returning null Optional is documented in the API contract.") - final Optional checkTimeout(final ConnectionEntry head, final long now) { - if (head == null) { + @GuardedBy("lock") + private Optional 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 = now - lastReceivedTicks; + 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); + } - /* - * We are using tri-state return here to indicate one of three conditions: - * - if there is no timeout to schedule, return Optional.empty() - * - if there is a timeout to schedule, return a non-empty optional - * - if this connections has timed out, return null - */ - @SuppressFBWarnings(value = "NP_OPTIONAL_RETURN_NULL", - justification = "Returning null Optional is documented in the API contract.") - Optional checkTimeout(final long now) { - return checkTimeout(pending.peek(), now); + 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(); } - static void poisonQueue(final Queue queue, final RequestException cause) { - for (ConnectionEntry e : queue) { - final Request request = e.getRequest(); - LOG.trace("Poisoning request {}", request, cause); - e.complete(request.toRequestFailure(cause)); + final void poison(final RequestException cause) { + lock.lock(); + try { + lockedPoison(cause); + } finally { + lock.unlock(); } - queue.clear(); } - void poison(final RequestException cause) { + @GuardedBy("lock") + private void lockedPoison(final RequestException cause) { poisoned = cause; - poisonQueue(pending, cause); + queue.poison(cause); } @VisibleForTesting @@ -253,7 +348,31 @@ public abstract class AbstractClientConnection { return poisoned; } - abstract ClientActorBehavior reconnectConnection(ClientActorBehavior current); + final void receiveResponse(final ResponseEnvelope envelope) { + final long now = currentTime(); + lastReceivedTicks = now; + + final Optional maybeEntry; + lock.lock(); + try { + maybeEntry = queue.complete(envelope, now); + } finally { + lock.unlock(); + } + + if (maybeEntry.isPresent()) { + final TransmittedConnectionEntry entry = maybeEntry.get(); + LOG.debug("Completing {} with {}", entry, envelope); + entry.complete(envelope.getMessage()); + } + } + + @Override + public final String toString() { + return addToStringAttributes(MoreObjects.toStringHelper(this).omitNullValues()).toString(); + } - abstract void receiveResponse(final ResponseEnvelope envelope); + ToStringHelper addToStringAttributes(final ToStringHelper toStringHelper) { + return toStringHelper.add("client", context.getIdentifier()).add("cookie", cookie).add("poisoned", poisoned); + } }