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=0366e7ace2496c0f5edd5433f237ca09db569a01;hb=b74c6012092e47430a8f4d6f4ddeb1d3e2b1b7df;hpb=b4d95acff78952020e9fbde4372d13b461fd7469 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 0366e7ace2..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 @@ -9,8 +9,11 @@ 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 edu.umd.cs.findbugs.annotations.SuppressFBWarnings; +import java.util.Collection; import java.util.Optional; import java.util.concurrent.TimeUnit; import java.util.concurrent.locks.Lock; @@ -38,11 +41,31 @@ 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); + + /** + * 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; @@ -50,8 +73,15 @@ public abstract class AbstractClientConnection { private final TransmitQueue queue; private final Long cookie; + @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; // Do not allow subclassing outside of this package AbstractClientConnection(final ClientActorContext context, final Long cookie, @@ -59,15 +89,15 @@ public abstract class AbstractClientConnection { 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 oldConnection) { + AbstractClientConnection(final AbstractClientConnection 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() { @@ -82,58 +112,91 @@ public abstract class AbstractClientConnection { 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. * + *

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 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(); + 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 enqueueRequest(final Request request, final Consumer> callback, + final long enqueuedTicks) { + enqueueEntry(new ConnectionEntry(request, callback, enqueuedTicks), currentTime()); + } + public abstract Optional getBackendInfo(); - final Iterable startReplay() { + final Collection 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 reconnectConnection(ClientActorBehavior current); + abstract ClientActorBehavior lockedReconnect(ClientActorBehavior current); - 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 reconnect(final ClientActorBehavior current) { lock.lock(); try { - queue.enqueue(entry, now); + return lockedReconnect(current); } finally { lock.unlock(); } @@ -144,9 +207,25 @@ 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; } /** @@ -158,21 +237,22 @@ public abstract class AbstractClientConnection { */ @VisibleForTesting final ClientActorBehavior runTimer(final ClientActorBehavior current) { - final Optional delay; + final Optional 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. @@ -181,22 +261,22 @@ public abstract class AbstractClientConnection { delay = lockedCheckTimeout(now); if (delay == null) { // We have timed out. There is no point in scheduling a timer - return reconnectConnection(current); + return lockedReconnect(current); + } + + 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 checkTimeout(final long now) { + final Optional checkTimeout(final long now) { lock.lock(); try { return lockedCheckTimeout(now); @@ -214,19 +294,38 @@ public abstract class AbstractClientConnection { @SuppressFBWarnings(value = "NP_OPTIONAL_RETURN_NULL", justification = "Returning null Optional is documented in the API contract.") @GuardedBy("lock") - private Optional lockedCheckTimeout(final long now) { - final ConnectionEntry head = queue.peek(); - if (head == null) { + 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); + } + + 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) { @@ -250,15 +349,30 @@ public abstract class AbstractClientConnection { } final void receiveResponse(final ResponseEnvelope envelope) { - final long now = readTime(); + final long now = currentTime(); + lastReceivedTicks = now; + final Optional maybeEntry; lock.lock(); try { - queue.complete(envelope, now); + maybeEntry = queue.complete(envelope, now); } finally { lock.unlock(); } - lastProgress = readTime(); + 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(); + } + + ToStringHelper addToStringAttributes(final ToStringHelper toStringHelper) { + return toStringHelper.add("client", context.getIdentifier()).add("cookie", cookie).add("poisoned", poisoned); } }