2 * Copyright (c) 2016 Cisco Systems, Inc. and others. All rights reserved.
4 * This program and the accompanying materials are made available under the
5 * terms of the Eclipse Public License v1.0 which accompanies this distribution,
6 * and is available at http://www.eclipse.org/legal/epl-v10.html
8 package org.opendaylight.controller.cluster.access.client;
10 import akka.actor.ActorRef;
11 import com.google.common.annotations.VisibleForTesting;
12 import com.google.common.base.Preconditions;
13 import com.google.common.base.Verify;
14 import edu.umd.cs.findbugs.annotations.SuppressFBWarnings;
15 import java.util.ArrayDeque;
16 import java.util.Collection;
17 import java.util.Deque;
18 import java.util.Iterator;
19 import java.util.Optional;
20 import java.util.Queue;
21 import javax.annotation.concurrent.NotThreadSafe;
22 import org.opendaylight.controller.cluster.access.concepts.Request;
23 import org.opendaylight.controller.cluster.access.concepts.RequestEnvelope;
24 import org.opendaylight.controller.cluster.access.concepts.RequestException;
25 import org.opendaylight.controller.cluster.access.concepts.Response;
26 import org.opendaylight.controller.cluster.access.concepts.ResponseEnvelope;
27 import org.slf4j.Logger;
28 import org.slf4j.LoggerFactory;
31 * This queue is internally split into two queues for performance reasons, both memory efficiency and copy
35 * Entries are always appended to the end, but then they are transmitted to the remote end and do not necessarily
36 * complete in the order in which they were sent -- hence the head of the queue does not increase linearly,
37 * but can involve spurious removals of non-head entries.
40 * For memory efficiency we want to pre-allocate both queues -- which points to ArrayDeque, but that is very
41 * inefficient when entries are removed from the middle. In the typical case we expect the number of in-flight
42 * entries to be an order of magnitude lower than the number of enqueued entries, hence the split.
45 * Note that in transient case of reconnect, when the backend gives us a lower number of maximum in-flight entries
46 * than the previous incarnation, we may end up still moving the pending queue -- but that is a very exceptional
47 * scenario, hence we consciously ignore it to keep the design relatively simple.
50 * This class is not thread-safe, as it is expected to be guarded by {@link AbstractClientConnection}.
52 * @author Robert Varga
55 abstract class TransmitQueue {
56 static final class Halted extends TransmitQueue {
57 // For ConnectingClientConnection.
58 Halted(final int targetDepth) {
62 // For ReconnectingClientConnection.
63 Halted(final TransmitQueue oldQueue, final long now) {
68 int canTransmitCount(final int inflightSize) {
73 TransmittedConnectionEntry transmit(final ConnectionEntry entry, final long now) {
74 throw new UnsupportedOperationException("Attempted to transmit on a halted queue");
78 static final class Transmitting extends TransmitQueue {
79 private final BackendInfo backend;
80 private long nextTxSequence;
82 // For ConnectedClientConnection.
83 Transmitting(final TransmitQueue oldQueue, final int targetDepth, final BackendInfo backend, final long now) {
84 super(oldQueue, targetDepth, now);
85 this.backend = Preconditions.checkNotNull(backend);
89 int canTransmitCount(final int inflightSize) {
90 return backend.getMaxMessages() - inflightSize;
94 TransmittedConnectionEntry transmit(final ConnectionEntry entry, final long now) {
95 final RequestEnvelope env = new RequestEnvelope(entry.getRequest().toVersion(backend.getVersion()),
96 backend.getSessionId(), nextTxSequence++);
98 final TransmittedConnectionEntry ret = new TransmittedConnectionEntry(entry, env.getSessionId(),
99 env.getTxSequence(), now);
100 backend.getActor().tell(env, ActorRef.noSender());
105 private static final Logger LOG = LoggerFactory.getLogger(TransmitQueue.class);
107 private final Deque<TransmittedConnectionEntry> inflight = new ArrayDeque<>();
108 private final Deque<ConnectionEntry> pending = new ArrayDeque<>();
109 private final AveragingProgressTracker tracker; // Cannot be just ProgressTracker as we are inheriting limits.
110 private ReconnectForwarder successor;
113 * Construct initial transmitting queue.
115 TransmitQueue(final int targetDepth) {
116 tracker = new AveragingProgressTracker(targetDepth);
120 * Construct new transmitting queue while inheriting timing data from the previous transmit queue instance.
122 TransmitQueue(final TransmitQueue oldQueue, final int targetDepth, final long now) {
123 tracker = new AveragingProgressTracker(oldQueue.tracker, targetDepth, now);
127 * Construct new transmitting queue while inheriting timing and size data from the previous transmit queue instance.
129 TransmitQueue(final TransmitQueue oldQueue, final long now) {
130 tracker = new AveragingProgressTracker(oldQueue.tracker, now);
134 * Cancel the accumulated sum of delays as we expect the new backend to work now.
136 void cancelDebt(final long now) {
137 tracker.cancelDebt(now);
141 * Drain the contents of the connection into a list. This will leave the queue empty and allow further entries
142 * to be added to it during replay. When we set the successor all entries enqueued between when this methods
143 * returns and the successor is set will be replayed to the successor.
145 * @return Collection of entries present in the queue.
147 final Collection<ConnectionEntry> drain() {
148 final Collection<ConnectionEntry> ret = new ArrayDeque<>(inflight.size() + pending.size());
149 ret.addAll(inflight);
156 final long ticksStalling(final long now) {
157 return tracker.ticksStalling(now);
160 final boolean hasSuccessor() {
161 return successor != null;
164 // If a matching request was found, this will track a task was closed.
165 final Optional<TransmittedConnectionEntry> complete(final ResponseEnvelope<?> envelope, final long now) {
166 Optional<TransmittedConnectionEntry> maybeEntry = findMatchingEntry(inflight, envelope);
167 if (maybeEntry == null) {
168 LOG.debug("Request for {} not found in inflight queue, checking pending queue", envelope);
169 maybeEntry = findMatchingEntry(pending, envelope);
172 if (maybeEntry == null || !maybeEntry.isPresent()) {
173 LOG.warn("No request matching {} found, ignoring response", envelope);
174 return Optional.empty();
177 final TransmittedConnectionEntry entry = maybeEntry.get();
178 tracker.closeTask(now, entry.getEnqueuedTicks(), entry.getTxTicks(), envelope.getExecutionTimeNanos());
180 // We have freed up a slot, try to transmit something
183 return Optional.of(entry);
186 final void tryTransmit(final long now) {
187 final int toSend = canTransmitCount(inflight.size());
188 if (toSend > 0 && !pending.isEmpty()) {
189 transmitEntries(toSend, now);
193 private void transmitEntries(final int maxTransmit, final long now) {
194 for (int i = 0; i < maxTransmit; ++i) {
195 final ConnectionEntry e = pending.poll();
197 LOG.debug("Queue {} transmitted {} requests", this, i);
201 transmitEntry(e, now);
204 LOG.debug("Queue {} transmitted {} requests", this, maxTransmit);
207 private void transmitEntry(final ConnectionEntry entry, final long now) {
208 LOG.debug("Queue {} transmitting entry {}", this, entry);
209 // We are not thread-safe and are supposed to be externally-guarded,
210 // hence send-before-record should be fine.
211 // This needs to be revisited if the external guards are lowered.
212 inflight.addLast(transmit(entry, now));
216 * Enqueue an entry, possibly also transmitting it.
218 * @return Delay to be forced on the calling thread, in nanoseconds.
220 final long enqueue(final ConnectionEntry entry, final long now) {
221 if (successor != null) {
222 // This call will pay the enqueuing price, hence the caller does not have to
223 successor.forwardEntry(entry, now);
227 // XXX: we should place a guard against incorrect entry sequences:
228 // entry.getEnqueueTicks() should have non-negative difference from the last entry present in the queues
230 // Reserve an entry before we do anything that can fail
231 final long delay = tracker.openTask(now);
234 * This is defensive to make sure we do not do the wrong thing here and reorder messages if we ever happen
235 * to have available send slots and non-empty pending queue.
237 final int toSend = canTransmitCount(inflight.size());
239 LOG.trace("Queue is at capacity, delayed sending of request {}", entry.getRequest());
240 pending.addLast(entry);
244 if (pending.isEmpty()) {
245 transmitEntry(entry, now);
249 pending.addLast(entry);
250 transmitEntries(toSend, now);
255 * Return the number of entries which can be transmitted assuming the supplied in-flight queue size.
257 abstract int canTransmitCount(int inflightSize);
259 abstract TransmittedConnectionEntry transmit(ConnectionEntry entry, long now);
261 final boolean isEmpty() {
262 return inflight.isEmpty() && pending.isEmpty();
265 final ConnectionEntry peek() {
266 final ConnectionEntry ret = inflight.peek();
271 return pending.peek();
274 final void poison(final RequestException cause) {
275 poisonQueue(inflight, cause);
276 poisonQueue(pending, cause);
279 final void setForwarder(final ReconnectForwarder forwarder, final long now) {
280 Verify.verify(successor == null, "Successor %s already set on connection %s", successor, this);
281 successor = Preconditions.checkNotNull(forwarder);
282 LOG.debug("Connection {} superseded by {}, splicing queue", this, successor);
285 * We need to account for entries which have been added between the time drain() was called and this method
286 * is invoked. Since the old connection is visible during replay and some entries may have completed on the
287 * replay thread, there was an avenue for this to happen.
290 ConnectionEntry entry = inflight.poll();
291 while (entry != null) {
292 successor.replayEntry(entry, now);
293 entry = inflight.poll();
297 entry = pending.poll();
298 while (entry != null) {
299 successor.replayEntry(entry, now);
300 entry = pending.poll();
304 LOG.debug("Connection {} queue spliced {} messages", this, count);
307 final void remove(final long now) {
308 final TransmittedConnectionEntry txe = inflight.poll();
310 final ConnectionEntry entry = pending.pop();
311 tracker.closeTask(now, entry.getEnqueuedTicks(), 0, 0);
313 tracker.closeTask(now, txe.getEnqueuedTicks(), txe.getTxTicks(), 0);
318 Deque<TransmittedConnectionEntry> getInflight() {
323 Deque<ConnectionEntry> getPending() {
328 * We are using tri-state return here to indicate one of three conditions:
329 * - if a matching entry is found, return an Optional containing it
330 * - if a matching entry is not found, but it makes sense to keep looking at other queues, return null
331 * - if a conflicting entry is encountered, indicating we should ignore this request, return an empty Optional
333 @SuppressFBWarnings(value = "NP_OPTIONAL_RETURN_NULL",
334 justification = "Returning null Optional is documented in the API contract.")
335 private static Optional<TransmittedConnectionEntry> findMatchingEntry(final Queue<? extends ConnectionEntry> queue,
336 final ResponseEnvelope<?> envelope) {
337 // Try to find the request in a queue. Responses may legally come back in a different order, hence we need
338 // to use an iterator
339 final Iterator<? extends ConnectionEntry> it = queue.iterator();
340 while (it.hasNext()) {
341 final ConnectionEntry e = it.next();
342 final Request<?, ?> request = e.getRequest();
343 final Response<?, ?> response = envelope.getMessage();
345 // First check for matching target, or move to next entry
346 if (!request.getTarget().equals(response.getTarget())) {
350 // Sanity-check logical sequence, ignore any out-of-order messages
351 if (request.getSequence() != response.getSequence()) {
352 LOG.debug("Expecting sequence {}, ignoring response {}", request.getSequence(), envelope);
353 return Optional.empty();
356 // Check if the entry has (ever) been transmitted
357 if (!(e instanceof TransmittedConnectionEntry)) {
358 return Optional.empty();
361 final TransmittedConnectionEntry te = (TransmittedConnectionEntry) e;
363 // Now check session match
364 if (envelope.getSessionId() != te.getSessionId()) {
365 LOG.debug("Expecting session {}, ignoring response {}", te.getSessionId(), envelope);
366 return Optional.empty();
368 if (envelope.getTxSequence() != te.getTxSequence()) {
369 LOG.warn("Expecting txSequence {}, ignoring response {}", te.getTxSequence(), envelope);
370 return Optional.empty();
373 LOG.debug("Completing request {} with {}", request, envelope);
375 return Optional.of(te);
381 private static void poisonQueue(final Queue<? extends ConnectionEntry> queue, final RequestException cause) {
382 for (ConnectionEntry e : queue) {
383 final Request<?, ?> request = e.getRequest();
384 LOG.trace("Poisoning request {}", request, cause);
385 e.complete(request.toRequestFailure(cause));