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=380fdeb862bdb8fbb64b79cae0594803dd7fde7c;hp=d37893bd7c3a9862ad89fb9d88b94337d0355f13;hb=bec0592fc0d51e8bd5038d47bde1db65ef147871;hpb=e1c283de301355cb8fa3f7d4fa28a6dd0af501eb 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 d37893bd7c..380fdeb862 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 @@ -13,8 +13,10 @@ 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; @@ -25,6 +27,7 @@ import org.opendaylight.controller.cluster.access.concepts.Request; 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; @@ -40,14 +43,38 @@ 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 static final FiniteDuration REQUEST_TIMEOUT_DURATION = FiniteDuration.apply(REQUEST_TIMEOUT_NANOS, - TimeUnit.NANOSECONDS); + /** + * 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); + + // Emit a debug entry if we sleep for more that this amount + private static final long DEBUG_DELAY_NANOS = TimeUnit.MILLISECONDS.toNanos(100); + + // Upper bound on the time a thread is forced to sleep to keep queue size under control + private static final long MAX_DELAY_SECONDS = 5; + private static final long MAX_DELAY_NANOS = TimeUnit.SECONDS.toNanos(MAX_DELAY_SECONDS); private final Lock lock = new ReentrantLock(); private final ClientActorContext context; @@ -58,6 +85,11 @@ public abstract class AbstractClientConnection { @GuardedBy("lock") private boolean haveTimer; + /** + * Time reference when we saw any activity from the backend. + */ + private long lastReceivedTicks; + private volatile RequestException poisoned; // Do not allow subclassing outside of this package @@ -66,6 +98,7 @@ public abstract class AbstractClientConnection { this.context = Preconditions.checkNotNull(context); this.cookie = Preconditions.checkNotNull(cookie); this.queue = Preconditions.checkNotNull(queue); + this.lastReceivedTicks = currentTime(); } // Do not allow subclassing outside of this package @@ -73,6 +106,7 @@ public abstract class AbstractClientConnection { this.context = oldConnection.context; this.cookie = oldConnection.cookie; this.queue = new TransmitQueue.Halted(targetQueueSize); + this.lastReceivedTicks = oldConnection.lastReceivedTicks; } public final ClientActorContext context() { @@ -103,13 +137,7 @@ public abstract class AbstractClientConnection { */ 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); - } + sendEntry(new ConnectionEntry(request, callback, now), now); } /** @@ -129,16 +157,48 @@ public abstract class AbstractClientConnection { enqueueEntry(new ConnectionEntry(request, callback, enqueuedTicks), currentTime()); } + public 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(); + } + } + 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) { setForwarder(forwarder); + + /* + * The process of replaying all messages may have taken a significant chunk of time, depending on type + * of messages, queue depth and available processing power. In extreme situations this may have already + * exceeded BACKEND_ALIVE_TIMEOUT_NANOS, in which case we are running the risk of not making reasonable forward + * progress before we start a reconnect cycle. + * + * Note that the timer is armed after we have sent the first message, hence we should be seeing a response + * from the backend before we see a timeout, simply due to how the mailbox operates. + * + * At any rate, reset the timestamp once we complete reconnection (which an atomic transition from the + * perspective of outside world), as that makes it a bit easier to reason about timing of events. + */ + lastReceivedTicks = currentTime(); lock.unlock(); } @@ -148,30 +208,34 @@ public abstract class AbstractClientConnection { } @GuardedBy("lock") - abstract ClientActorBehavior lockedReconnect(ClientActorBehavior current); + abstract ClientActorBehavior lockedReconnect(ClientActorBehavior current, + RequestException runtimeRequestException); - final long enqueueEntry(final ConnectionEntry entry, final long now) { - lock.lock(); + final void sendEntry(final ConnectionEntry entry, final long now) { + long delay = enqueueEntry(entry, now); try { - final RequestException maybePoison = poisoned; - if (maybePoison != null) { - throw new IllegalStateException("Connection " + this + " has been poisoned", maybePoison); + if (delay >= DEBUG_DELAY_NANOS) { + if (delay > MAX_DELAY_NANOS) { + LOG.info("Capping {} throttle delay from {} to {} seconds", this, + TimeUnit.NANOSECONDS.toSeconds(delay), MAX_DELAY_SECONDS, new Throwable()); + delay = MAX_DELAY_NANOS; + } + if (LOG.isDebugEnabled()) { + LOG.debug("{}: Sleeping for {}ms on connection {}", context.persistenceId(), + TimeUnit.NANOSECONDS.toMillis(delay), this); + } } - - if (queue.isEmpty()) { - // The queue is becoming non-empty, schedule a timer - scheduleTimer(REQUEST_TIMEOUT_DURATION); - } - return queue.enqueue(entry, now); - } finally { - lock.unlock(); + TimeUnit.NANOSECONDS.sleep(delay); + } catch (InterruptedException e) { + Thread.currentThread().interrupt(); + LOG.debug("Interrupted after sleeping {}ns", e, currentTime() - now); } } - final ClientActorBehavior reconnect(final ClientActorBehavior current) { + final ClientActorBehavior reconnect(final ClientActorBehavior current, final RequestException cause) { lock.lock(); try { - return lockedReconnect(current); + return lockedReconnect(current, cause); } finally { lock.unlock(); } @@ -183,17 +247,23 @@ public abstract class AbstractClientConnection { * @param delay Delay, in nanoseconds */ @GuardedBy("lock") - private void scheduleTimer(final FiniteDuration delay) { + private void scheduleTimer(final long delay) { if (haveTimer) { - LOG.debug("{}: timer already scheduled", context.persistenceId()); + LOG.debug("{}: timer already scheduled on {}", context.persistenceId(), this); return; } if (queue.hasSuccessor()) { - LOG.debug("{}: connection has successor, not scheduling timer", context.persistenceId()); + LOG.debug("{}: connection {} has a successor, not scheduling timer", context.persistenceId(), this); return; } - LOG.debug("{}: scheduling timeout in {}", context.persistenceId(), delay); - context.executeInActor(this::runTimer, delay); + + // 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("{}: connection {} scheduling timeout in {}", context.persistenceId(), this, dur); + context.executeInActor(this::runTimer, dur); haveTimer = true; } @@ -206,13 +276,17 @@ public abstract class AbstractClientConnection { */ @VisibleForTesting final ClientActorBehavior runTimer(final ClientActorBehavior current) { - final Optional delay; + final Optional delay; lock.lock(); try { haveTimer = false; final long now = currentTime(); + + LOG.debug("{}: running timer on {}", context.persistenceId(), this); + // 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, @@ -229,12 +303,16 @@ public abstract class AbstractClientConnection { delay = lockedCheckTimeout(now); if (delay == null) { // We have timed out. There is no point in scheduling a timer - return lockedReconnect(current); + LOG.debug("{}: connection {} timed out", context.persistenceId(), this); + 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()); + } else { + LOG.debug("{}: not scheduling timeout on {}", context.persistenceId(), this); } } finally { lock.unlock(); @@ -244,7 +322,7 @@ public abstract class AbstractClientConnection { } @VisibleForTesting - final Optional checkTimeout(final long now) { + final Optional checkTimeout(final long now) { lock.lock(); try { return lockedCheckTimeout(now); @@ -253,6 +331,10 @@ public abstract class AbstractClientConnection { } } + 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() @@ -262,19 +344,41 @@ 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()) { + LOG.debug("{}: connection {} is empty", context.persistenceId(), this); return Optional.empty(); } - final long beenOpen = now - head.getEnqueuedTicks(); - if (beenOpen >= REQUEST_TIMEOUT_NANOS) { - LOG.debug("Connection {} has a request not completed for {} nanoseconds, timing out", this, beenOpen); + final long backendSilentTicks = backendSilentTicks(now); + if (backendSilentTicks >= BACKEND_ALIVE_TIMEOUT_NANOS) { + LOG.debug("{}: Connection {} has not seen activity from backend for {} nanoseconds, timing out", + context.persistenceId(), this, backendSilentTicks); return null; } - return Optional.of(FiniteDuration.apply(REQUEST_TIMEOUT_NANOS - beenOpen, 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 entry {}", context.persistenceId(), this, head); + + final double time = (beenOpen * 1.0) / 1_000_000_000; + head.complete(head.getRequest().toRequestFailure( + new RequestTimeoutException("Timed out after " + time + "seconds"))); + } + + LOG.debug("Connection {} timed out {} tasks", this, tasksTimedOut); + if (tasksTimedOut != 0) { + queue.tryTransmit(now); + } + + return Optional.empty(); } final void poison(final RequestException cause) { @@ -288,10 +392,14 @@ public abstract class AbstractClientConnection { @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; @@ -299,6 +407,7 @@ public abstract class AbstractClientConnection { final void receiveResponse(final ResponseEnvelope envelope) { final long now = currentTime(); + lastReceivedTicks = now; final Optional maybeEntry; lock.lock();