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=af66369271c34c8eba953d6fc6ecf5c27637aa51;hp=98442256c6ecbcc9f63f54586d05611a80797d36;hb=634dfac8eead60f443bf75e749c70d1f2bb29198;hpb=c09801280b4c44f4ec26766e4d13b1a5d1f3ed59 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 98442256c6..af66369271 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 @@ -50,24 +50,21 @@ public abstract class AbstractClientConnection { */ /** * 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 + * we have an outstanding request. If when this time expires, we tear down this connection and attempt to reconnect * it. */ - @VisibleForTesting - static final long BACKEND_ALIVE_TIMEOUT_NANOS = TimeUnit.SECONDS.toNanos(30); + public static final long DEFAULT_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); + public static final long DEFAULT_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); + public static final long DEFAULT_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); @@ -92,21 +89,35 @@ public abstract class AbstractClientConnection { private volatile RequestException poisoned; + // Private constructor to avoid code duplication. + private AbstractClientConnection(final AbstractClientConnection oldConn, final TransmitQueue newQueue) { + this.context = Preconditions.checkNotNull(oldConn.context); + this.cookie = Preconditions.checkNotNull(oldConn.cookie); + this.queue = Preconditions.checkNotNull(newQueue); + // Will be updated in finishReplay if needed. + this.lastReceivedTicks = oldConn.lastReceivedTicks; + } + + // This constructor is only to be called by ConnectingClientConnection constructor. // Do not allow subclassing outside of this package - AbstractClientConnection(final ClientActorContext context, final Long cookie, - final TransmitQueue queue) { + AbstractClientConnection(final ClientActorContext context, final Long cookie, final int queueDepth) { this.context = Preconditions.checkNotNull(context); this.cookie = Preconditions.checkNotNull(cookie); - this.queue = Preconditions.checkNotNull(queue); + this.queue = new TransmitQueue.Halted(queueDepth); this.lastReceivedTicks = currentTime(); } + // This constructor is only to be called (indirectly) by ReconnectingClientConnection constructor. + // Do not allow subclassing outside of this package + AbstractClientConnection(final AbstractClientConnection oldConn) { + this(oldConn, new TransmitQueue.Halted(oldConn.queue, oldConn.currentTime())); + } + + // This constructor is only to be called (indirectly) by ConnectedClientConnection constructor. // Do not allow subclassing outside of this package - 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; + AbstractClientConnection(final AbstractClientConnection oldConn, final T newBackend, final int queueDepth) { + this(oldConn, new TransmitQueue.Transmitting(oldConn.queue, queueDepth, newBackend, oldConn.currentTime(), + Preconditions.checkNotNull(oldConn.context).messageSlicer())); } public final ClientActorContext context() { @@ -157,24 +168,47 @@ public abstract class AbstractClientConnection { enqueueEntry(new ConnectionEntry(request, callback, enqueuedTicks), currentTime()); } - public final long enqueueEntry(final ConnectionEntry entry, final long now) { + private long enqueueOrForward(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); - } + commonEnqueue(entry, now); + return queue.enqueueOrForward(entry, now); + } finally { + lock.unlock(); + } + } - if (queue.isEmpty()) { - // The queue is becoming non-empty, schedule a timer. - scheduleTimer(entry.getEnqueuedTicks() + REQUEST_TIMEOUT_NANOS - now); - } - return queue.enqueue(entry, now); + /** + * Enqueue an entry, possibly also transmitting it. + */ + public final void enqueueEntry(final ConnectionEntry entry, final long now) { + lock.lock(); + try { + commonEnqueue(entry, now); + queue.enqueueOrReplay(entry, now); } finally { lock.unlock(); } } + @GuardedBy("lock") + private void commonEnqueue(final ConnectionEntry entry, final long now) { + 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() + context.config().getRequestTimeout() - now); + } + } + + // To be called from ClientActorBehavior on ConnectedClientConnection after entries are replayed. + final void cancelDebt() { + queue.cancelDebt(currentTime()); + } + public abstract Optional getBackendInfo(); final Collection startReplay() { @@ -212,12 +246,12 @@ public abstract class AbstractClientConnection { RequestException runtimeRequestException); final void sendEntry(final ConnectionEntry entry, final long now) { - long delay = enqueueEntry(entry, now); + long delay = enqueueOrForward(entry, now); try { 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); + TimeUnit.NANOSECONDS.toSeconds(delay), MAX_DELAY_SECONDS, new Throwable()); delay = MAX_DELAY_NANOS; } if (LOG.isDebugEnabled()) { @@ -228,7 +262,7 @@ public abstract class AbstractClientConnection { TimeUnit.NANOSECONDS.sleep(delay); } catch (InterruptedException e) { Thread.currentThread().interrupt(); - LOG.debug("Interrupted after sleeping {}ns", e, currentTime() - now); + LOG.debug("Interrupted after sleeping {}ns", currentTime() - now, e); } } @@ -259,7 +293,7 @@ public abstract class AbstractClientConnection { // 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 long normalized = delay <= 0 ? 0 : Math.min(delay, context.config().getBackendAlivenessTimerInterval()); final FiniteDuration dur = FiniteDuration.fromNanos(normalized); LOG.debug("{}: connection {} scheduling timeout in {}", context.persistenceId(), this, dur); @@ -269,7 +303,7 @@ public abstract class AbstractClientConnection { /** * 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. + * in {@link #DEFAULT_NO_PROGRESS_TIMEOUT_NANOS} nanoseconds, it will be aborted. * * @param current Current behavior * @return Next behavior to use @@ -288,7 +322,7 @@ public abstract class AbstractClientConnection { // 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) { + if (ticksSinceProgress >= context.config().getNoProgressTimeout()) { LOG.error("Queue {} has not seen progress in {} seconds, failing all requests", this, TimeUnit.NANOSECONDS.toSeconds(ticksSinceProgress)); @@ -351,7 +385,7 @@ public abstract class AbstractClientConnection { } final long backendSilentTicks = backendSilentTicks(now); - if (backendSilentTicks >= BACKEND_ALIVE_TIMEOUT_NANOS) { + if (backendSilentTicks >= context.config().getBackendAlivenessTimerInterval()) { LOG.debug("{}: Connection {} has not seen activity from backend for {} nanoseconds, timing out", context.persistenceId(), this, backendSilentTicks); return null; @@ -360,15 +394,16 @@ public abstract class AbstractClientConnection { 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); + final long requestTimeout = context.config().getRequestTimeout(); + if (beenOpen < requestTimeout) { + return Optional.of(requestTimeout - beenOpen); } tasksTimedOut++; queue.remove(now); LOG.debug("{}: Connection {} timed out entry {}", context.persistenceId(), this, head); - head.complete(head.getRequest().toRequestFailure( - new RequestTimeoutException("Timed out after " + beenOpen + "ns"))); + + timeoutEntry(head, beenOpen); } LOG.debug("Connection {} timed out {} tasks", this, tasksTimedOut); @@ -379,6 +414,19 @@ public abstract class AbstractClientConnection { return Optional.empty(); } + private void timeoutEntry(final ConnectionEntry entry, final long beenOpen) { + // Timeouts needs to be re-scheduled on actor thread because we are holding the lock on the current queue, + // which may be the tail of a successor chain. This is a problem if the callback attempts to send a request + // because that will attempt to lock the chain from the start, potentially causing a deadlock if there is + // a concurrent attempt to transmit. + context.executeInActor(current -> { + final double time = beenOpen * 1.0 / 1_000_000_000; + entry.complete(entry.getRequest().toRequestFailure( + new RequestTimeoutException("Timed out after " + time + "seconds"))); + return current; + }); + } + final void poison(final RequestException cause) { lock.lock(); try { @@ -403,7 +451,7 @@ public abstract class AbstractClientConnection { return poisoned; } - final void receiveResponse(final ResponseEnvelope envelope) { + void receiveResponse(final ResponseEnvelope envelope) { final long now = currentTime(); lastReceivedTicks = now;