001/* 002 * Licensed to the Apache Software Foundation (ASF) under one 003 * or more contributor license agreements. See the NOTICE file 004 * distributed with this work for additional information 005 * regarding copyright ownership. The ASF licenses this file 006 * to you under the Apache License, Version 2.0 (the 007 * "License"); you may not use this file except in compliance 008 * with the License. You may obtain a copy of the License at 009 * 010 * http://www.apache.org/licenses/LICENSE-2.0 011 * 012 * Unless required by applicable law or agreed to in writing, software 013 * distributed under the License is distributed on an "AS IS" BASIS, 014 * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 015 * See the License for the specific language governing permissions and 016 * limitations under the License. 017 */ 018package org.apache.hadoop.hbase.procedure2; 019 020import edu.umd.cs.findbugs.annotations.Nullable; 021import java.io.IOException; 022import java.io.UncheckedIOException; 023import java.util.ArrayDeque; 024import java.util.ArrayList; 025import java.util.Arrays; 026import java.util.Collection; 027import java.util.Deque; 028import java.util.HashSet; 029import java.util.List; 030import java.util.Set; 031import java.util.concurrent.ConcurrentHashMap; 032import java.util.concurrent.CopyOnWriteArrayList; 033import java.util.concurrent.Executor; 034import java.util.concurrent.Executors; 035import java.util.concurrent.TimeUnit; 036import java.util.concurrent.atomic.AtomicBoolean; 037import java.util.concurrent.atomic.AtomicInteger; 038import java.util.concurrent.atomic.AtomicLong; 039import java.util.stream.Collectors; 040import java.util.stream.Stream; 041import org.apache.hadoop.conf.Configuration; 042import org.apache.hadoop.hbase.HConstants; 043import org.apache.hadoop.hbase.exceptions.IllegalArgumentIOException; 044import org.apache.hadoop.hbase.log.HBaseMarkers; 045import org.apache.hadoop.hbase.procedure2.Procedure.LockState; 046import org.apache.hadoop.hbase.procedure2.store.ProcedureStore; 047import org.apache.hadoop.hbase.procedure2.store.ProcedureStore.ProcedureIterator; 048import org.apache.hadoop.hbase.procedure2.store.ProcedureStore.ProcedureStoreListener; 049import org.apache.hadoop.hbase.procedure2.trace.ProcedureSpanBuilder; 050import org.apache.hadoop.hbase.procedure2.util.StringUtils; 051import org.apache.hadoop.hbase.security.User; 052import org.apache.hadoop.hbase.trace.TraceUtil; 053import org.apache.hadoop.hbase.util.EnvironmentEdgeManager; 054import org.apache.hadoop.hbase.util.IdLock; 055import org.apache.hadoop.hbase.util.NonceKey; 056import org.apache.hadoop.hbase.util.Threads; 057import org.apache.yetus.audience.InterfaceAudience; 058import org.slf4j.Logger; 059import org.slf4j.LoggerFactory; 060 061import org.apache.hbase.thirdparty.com.google.common.base.Preconditions; 062import org.apache.hbase.thirdparty.com.google.common.util.concurrent.ThreadFactoryBuilder; 063 064import org.apache.hadoop.hbase.shaded.protobuf.generated.ProcedureProtos.ProcedureState; 065 066/** 067 * Thread Pool that executes the submitted procedures. The executor has a ProcedureStore associated. 068 * Each operation is logged and on restart the pending procedures are resumed. Unless the Procedure 069 * code throws an error (e.g. invalid user input) the procedure will complete (at some point in 070 * time), On restart the pending procedures are resumed and the once failed will be rolledback. The 071 * user can add procedures to the executor via submitProcedure(proc) check for the finished state 072 * via isFinished(procId) and get the result via getResult(procId) 073 */ 074@InterfaceAudience.Private 075public class ProcedureExecutor<TEnvironment> { 076 private static final Logger LOG = LoggerFactory.getLogger(ProcedureExecutor.class); 077 078 public static final String CHECK_OWNER_SET_CONF_KEY = "hbase.procedure.check.owner.set"; 079 private static final boolean DEFAULT_CHECK_OWNER_SET = false; 080 081 public static final String WORKER_KEEP_ALIVE_TIME_CONF_KEY = 082 "hbase.procedure.worker.keep.alive.time.msec"; 083 private static final long DEFAULT_WORKER_KEEP_ALIVE_TIME = TimeUnit.MINUTES.toMillis(1); 084 085 public static final String EVICT_TTL_CONF_KEY = "hbase.procedure.cleaner.evict.ttl"; 086 static final int DEFAULT_EVICT_TTL = 15 * 60000; // 15min 087 088 public static final String EVICT_ACKED_TTL_CONF_KEY = "hbase.procedure.cleaner.acked.evict.ttl"; 089 static final int DEFAULT_ACKED_EVICT_TTL = 5 * 60000; // 5min 090 091 /** 092 * {@link #testing} is non-null when ProcedureExecutor is being tested. Tests will try to break PE 093 * having it fail at various junctures. When non-null, testing is set to an instance of the below 094 * internal {@link Testing} class with flags set for the particular test. 095 */ 096 volatile Testing testing = null; 097 098 /** 099 * Class with parameters describing how to fail/die when in testing-context. 100 */ 101 public static class Testing { 102 protected volatile boolean killIfHasParent = true; 103 protected volatile boolean killIfSuspended = false; 104 105 /** 106 * Kill the PE BEFORE we store state to the WAL. Good for figuring out if a Procedure is 107 * persisting all the state it needs to recover after a crash. 108 */ 109 protected volatile boolean killBeforeStoreUpdate = false; 110 protected volatile boolean toggleKillBeforeStoreUpdate = false; 111 112 /** 113 * Set when we want to fail AFTER state has been stored into the WAL. Rarely used. HBASE-20978 114 * is about a case where memory-state was being set after store to WAL where a crash could cause 115 * us to get stuck. This flag allows killing at what was a vulnerable time. 116 */ 117 protected volatile boolean killAfterStoreUpdate = false; 118 protected volatile boolean toggleKillAfterStoreUpdate = false; 119 120 protected boolean shouldKillBeforeStoreUpdate() { 121 final boolean kill = this.killBeforeStoreUpdate; 122 if (this.toggleKillBeforeStoreUpdate) { 123 this.killBeforeStoreUpdate = !kill; 124 LOG.warn("Toggle KILL before store update to: " + this.killBeforeStoreUpdate); 125 } 126 return kill; 127 } 128 129 protected boolean shouldKillBeforeStoreUpdate(boolean isSuspended, boolean hasParent) { 130 if (isSuspended && !killIfSuspended) { 131 return false; 132 } 133 if (hasParent && !killIfHasParent) { 134 return false; 135 } 136 return shouldKillBeforeStoreUpdate(); 137 } 138 139 protected boolean shouldKillAfterStoreUpdate() { 140 final boolean kill = this.killAfterStoreUpdate; 141 if (this.toggleKillAfterStoreUpdate) { 142 this.killAfterStoreUpdate = !kill; 143 LOG.warn("Toggle KILL after store update to: " + this.killAfterStoreUpdate); 144 } 145 return kill; 146 } 147 148 protected boolean shouldKillAfterStoreUpdate(final boolean isSuspended) { 149 return (isSuspended && !killIfSuspended) ? false : shouldKillAfterStoreUpdate(); 150 } 151 } 152 153 public interface ProcedureExecutorListener { 154 void procedureLoaded(long procId); 155 156 void procedureAdded(long procId); 157 158 void procedureFinished(long procId); 159 } 160 161 /** 162 * Map the the procId returned by submitProcedure(), the Root-ProcID, to the Procedure. Once a 163 * Root-Procedure completes (success or failure), the result will be added to this map. The user 164 * of ProcedureExecutor should call getResult(procId) to get the result. 165 */ 166 private final ConcurrentHashMap<Long, CompletedProcedureRetainer<TEnvironment>> completed = 167 new ConcurrentHashMap<>(); 168 169 /** 170 * Map the the procId returned by submitProcedure(), the Root-ProcID, to the RootProcedureState. 171 * The RootProcedureState contains the execution stack of the Root-Procedure, It is added to the 172 * map by submitProcedure() and removed on procedure completion. 173 */ 174 private final ConcurrentHashMap<Long, RootProcedureState<TEnvironment>> rollbackStack = 175 new ConcurrentHashMap<>(); 176 177 /** 178 * Helper map to lookup the live procedures by ID. This map contains every procedure. 179 * root-procedures and subprocedures. 180 */ 181 private final ConcurrentHashMap<Long, Procedure<TEnvironment>> procedures = 182 new ConcurrentHashMap<>(); 183 184 /** 185 * Helper map to lookup whether the procedure already issued from the same client. This map 186 * contains every root procedure. 187 */ 188 private final ConcurrentHashMap<NonceKey, Long> nonceKeysToProcIdsMap = new ConcurrentHashMap<>(); 189 190 private final CopyOnWriteArrayList<ProcedureExecutorListener> listeners = 191 new CopyOnWriteArrayList<>(); 192 193 private Configuration conf; 194 195 /** 196 * Created in the {@link #init(int, boolean)} method. Destroyed in {@link #join()} (FIX! Doing 197 * resource handling rather than observing in a #join is unexpected). Overridden when we do the 198 * ProcedureTestingUtility.testRecoveryAndDoubleExecution trickery (Should be ok). 199 */ 200 private ThreadGroup threadGroup; 201 202 /** 203 * Created in the {@link #init(int, boolean)} method. Terminated in {@link #join()} (FIX! Doing 204 * resource handling rather than observing in a #join is unexpected). Overridden when we do the 205 * ProcedureTestingUtility.testRecoveryAndDoubleExecution trickery (Should be ok). 206 */ 207 private CopyOnWriteArrayList<WorkerThread> workerThreads; 208 209 /** 210 * Created in the {@link #init(int, boolean)} method. Terminated in {@link #join()} (FIX! Doing 211 * resource handling rather than observing in a #join is unexpected). Overridden when we do the 212 * ProcedureTestingUtility.testRecoveryAndDoubleExecution trickery (Should be ok). 213 */ 214 private TimeoutExecutorThread<TEnvironment> timeoutExecutor; 215 216 /** 217 * WorkerMonitor check for stuck workers and new worker thread when necessary, for example if 218 * there is no worker to assign meta, it will new worker thread for it, so it is very important. 219 * TimeoutExecutor execute many tasks like DeadServerMetricRegionChore RegionInTransitionChore and 220 * so on, some tasks may execute for a long time so will block other tasks like WorkerMonitor, so 221 * use a dedicated thread for executing WorkerMonitor. 222 */ 223 private TimeoutExecutorThread<TEnvironment> workerMonitorExecutor; 224 225 private int corePoolSize; 226 private int maxPoolSize; 227 228 private volatile long keepAliveTime; 229 230 /** 231 * Scheduler/Queue that contains runnable procedures. 232 */ 233 private final ProcedureScheduler scheduler; 234 235 private final Executor forceUpdateExecutor = Executors.newSingleThreadExecutor( 236 new ThreadFactoryBuilder().setDaemon(true).setNameFormat("Force-Update-PEWorker-%d").build()); 237 238 private final AtomicLong lastProcId = new AtomicLong(-1); 239 private final AtomicLong workerId = new AtomicLong(0); 240 private final AtomicInteger activeExecutorCount = new AtomicInteger(0); 241 private final AtomicBoolean running = new AtomicBoolean(false); 242 private final TEnvironment environment; 243 private final ProcedureStore store; 244 245 private final boolean checkOwnerSet; 246 247 // To prevent concurrent execution of the same procedure. 248 // For some rare cases, especially if the procedure uses ProcedureEvent, it is possible that the 249 // procedure is woken up before we finish the suspend which causes the same procedures to be 250 // executed in parallel. This does lead to some problems, see HBASE-20939&HBASE-20949, and is also 251 // a bit confusing to the developers. So here we introduce this lock to prevent the concurrent 252 // execution of the same procedure. 253 private final IdLock procExecutionLock = new IdLock(); 254 255 public ProcedureExecutor(final Configuration conf, final TEnvironment environment, 256 final ProcedureStore store) { 257 this(conf, environment, store, new SimpleProcedureScheduler()); 258 } 259 260 private boolean isRootFinished(Procedure<?> proc) { 261 Procedure<?> rootProc = procedures.get(proc.getRootProcId()); 262 return rootProc == null || rootProc.isFinished(); 263 } 264 265 private void forceUpdateProcedure(long procId) throws IOException { 266 IdLock.Entry lockEntry = procExecutionLock.getLockEntry(procId); 267 try { 268 Procedure<TEnvironment> proc = procedures.get(procId); 269 if (proc != null) { 270 if (proc.isFinished() && proc.hasParent() && isRootFinished(proc)) { 271 LOG.debug("Procedure {} has already been finished and parent is succeeded," 272 + " skip force updating", proc); 273 return; 274 } 275 } else { 276 CompletedProcedureRetainer<TEnvironment> retainer = completed.get(procId); 277 if (retainer == null || retainer.getProcedure() instanceof FailedProcedure) { 278 LOG.debug("No pending procedure with id = {}, skip force updating.", procId); 279 return; 280 } 281 long evictTtl = conf.getInt(EVICT_TTL_CONF_KEY, DEFAULT_EVICT_TTL); 282 long evictAckTtl = conf.getInt(EVICT_ACKED_TTL_CONF_KEY, DEFAULT_ACKED_EVICT_TTL); 283 if (retainer.isExpired(EnvironmentEdgeManager.currentTime(), evictTtl, evictAckTtl)) { 284 LOG.debug("Procedure {} has already been finished and expired, skip force updating", 285 procId); 286 return; 287 } 288 proc = retainer.getProcedure(); 289 } 290 LOG.debug("Force update procedure {}", proc); 291 store.update(proc); 292 } finally { 293 procExecutionLock.releaseLockEntry(lockEntry); 294 } 295 } 296 297 public ProcedureExecutor(final Configuration conf, final TEnvironment environment, 298 final ProcedureStore store, final ProcedureScheduler scheduler) { 299 this.environment = environment; 300 this.scheduler = scheduler; 301 this.store = store; 302 this.conf = conf; 303 this.checkOwnerSet = conf.getBoolean(CHECK_OWNER_SET_CONF_KEY, DEFAULT_CHECK_OWNER_SET); 304 refreshConfiguration(conf); 305 store.registerListener(new ProcedureStoreListener() { 306 307 @Override 308 public void forceUpdate(long[] procIds) { 309 Arrays.stream(procIds).forEach(procId -> forceUpdateExecutor.execute(() -> { 310 try { 311 forceUpdateProcedure(procId); 312 } catch (IOException e) { 313 LOG.warn("Failed to force update procedure with pid={}", procId); 314 } 315 })); 316 } 317 }); 318 } 319 320 private void load(final boolean abortOnCorruption) throws IOException { 321 Preconditions.checkArgument(completed.isEmpty(), "completed not empty"); 322 Preconditions.checkArgument(rollbackStack.isEmpty(), "rollback state not empty"); 323 Preconditions.checkArgument(procedures.isEmpty(), "procedure map not empty"); 324 Preconditions.checkArgument(scheduler.size() == 0, "run queue not empty"); 325 326 store.load(new ProcedureStore.ProcedureLoader() { 327 @Override 328 public void setMaxProcId(long maxProcId) { 329 assert lastProcId.get() < 0 : "expected only one call to setMaxProcId()"; 330 lastProcId.set(maxProcId); 331 } 332 333 @Override 334 public void load(ProcedureIterator procIter) throws IOException { 335 loadProcedures(procIter, abortOnCorruption); 336 } 337 338 @Override 339 public void handleCorrupted(ProcedureIterator procIter) throws IOException { 340 int corruptedCount = 0; 341 while (procIter.hasNext()) { 342 Procedure<?> proc = procIter.next(); 343 LOG.error("Corrupt " + proc); 344 corruptedCount++; 345 } 346 if (abortOnCorruption && corruptedCount > 0) { 347 throw new IOException("found " + corruptedCount + " corrupted procedure(s) on replay"); 348 } 349 } 350 }); 351 } 352 353 private void restoreLock(Procedure<TEnvironment> proc, Set<Long> restored) { 354 proc.restoreLock(getEnvironment()); 355 restored.add(proc.getProcId()); 356 } 357 358 private void restoreLocks(Deque<Procedure<TEnvironment>> stack, Set<Long> restored) { 359 while (!stack.isEmpty()) { 360 restoreLock(stack.pop(), restored); 361 } 362 } 363 364 // Restore the locks for all the procedures. 365 // Notice that we need to restore the locks starting from the root proc, otherwise there will be 366 // problem that a sub procedure may hold the exclusive lock first and then we are stuck when 367 // calling the acquireLock method for the parent procedure. 368 // The algorithm is straight-forward: 369 // 1. Use a set to record the procedures which locks have already been restored. 370 // 2. Use a stack to store the hierarchy of the procedures 371 // 3. For all the procedure, we will first try to find its parent and push it into the stack, 372 // unless 373 // a. We have no parent, i.e, we are the root procedure 374 // b. The lock has already been restored(by checking the set introduced in #1) 375 // then we start to pop the stack and call acquireLock for each procedure. 376 // Notice that this should be done for all procedures, not only the ones in runnableList. 377 private void restoreLocks() { 378 Set<Long> restored = new HashSet<>(); 379 Deque<Procedure<TEnvironment>> stack = new ArrayDeque<>(); 380 procedures.values().forEach(proc -> { 381 for (;;) { 382 if (restored.contains(proc.getProcId())) { 383 restoreLocks(stack, restored); 384 return; 385 } 386 if (!proc.hasParent()) { 387 restoreLock(proc, restored); 388 restoreLocks(stack, restored); 389 return; 390 } 391 stack.push(proc); 392 proc = procedures.get(proc.getParentProcId()); 393 } 394 }); 395 } 396 397 private void loadProcedures(ProcedureIterator procIter, boolean abortOnCorruption) 398 throws IOException { 399 // 1. Build the rollback stack 400 int runnableCount = 0; 401 int failedCount = 0; 402 int waitingCount = 0; 403 int waitingTimeoutCount = 0; 404 while (procIter.hasNext()) { 405 boolean finished = procIter.isNextFinished(); 406 @SuppressWarnings("unchecked") 407 Procedure<TEnvironment> proc = procIter.next(); 408 NonceKey nonceKey = proc.getNonceKey(); 409 long procId = proc.getProcId(); 410 411 if (finished) { 412 completed.put(proc.getProcId(), new CompletedProcedureRetainer<>(proc)); 413 LOG.debug("Completed {}", proc); 414 } else { 415 if (!proc.hasParent()) { 416 assert !proc.isFinished() : "unexpected finished procedure"; 417 rollbackStack.put(proc.getProcId(), new RootProcedureState<>()); 418 } 419 420 // add the procedure to the map 421 proc.beforeReplay(getEnvironment()); 422 procedures.put(proc.getProcId(), proc); 423 switch (proc.getState()) { 424 case RUNNABLE: 425 runnableCount++; 426 break; 427 case FAILED: 428 failedCount++; 429 break; 430 case WAITING: 431 waitingCount++; 432 break; 433 case WAITING_TIMEOUT: 434 waitingTimeoutCount++; 435 break; 436 default: 437 break; 438 } 439 } 440 441 if (nonceKey != null) { 442 nonceKeysToProcIdsMap.put(nonceKey, procId); // add the nonce to the map 443 } 444 } 445 446 // 2. Initialize the stacks: In the old implementation, for procedures in FAILED state, we will 447 // push it into the ProcedureScheduler directly to execute the rollback. But this does not work 448 // after we introduce the restore lock stage. For now, when we acquire a xlock, we will remove 449 // the queue from runQueue in scheduler, and then when a procedure which has lock access, for 450 // example, a sub procedure of the procedure which has the xlock, is pushed into the scheduler, 451 // we will add the queue back to let the workers poll from it. The assumption here is that, the 452 // procedure which has the xlock should have been polled out already, so when loading we can not 453 // add the procedure to scheduler first and then call acquireLock, since the procedure is still 454 // in the queue, and since we will remove the queue from runQueue, then no one can poll it out, 455 // then there is a dead lock 456 List<Procedure<TEnvironment>> runnableList = new ArrayList<>(runnableCount); 457 List<Procedure<TEnvironment>> failedList = new ArrayList<>(failedCount); 458 List<Procedure<TEnvironment>> waitingList = new ArrayList<>(waitingCount); 459 List<Procedure<TEnvironment>> waitingTimeoutList = new ArrayList<>(waitingTimeoutCount); 460 procIter.reset(); 461 while (procIter.hasNext()) { 462 if (procIter.isNextFinished()) { 463 procIter.skipNext(); 464 continue; 465 } 466 467 @SuppressWarnings("unchecked") 468 Procedure<TEnvironment> proc = procIter.next(); 469 assert !(proc.isFinished() && !proc.hasParent()) : "unexpected completed proc=" + proc; 470 LOG.debug("Loading {}", proc); 471 Long rootProcId = getRootProcedureId(proc); 472 // The orphan procedures will be passed to handleCorrupted, so add an assert here 473 assert rootProcId != null; 474 475 if (proc.hasParent()) { 476 Procedure<TEnvironment> parent = procedures.get(proc.getParentProcId()); 477 if (parent != null && !proc.isFinished()) { 478 parent.incChildrenLatch(); 479 } 480 } 481 482 RootProcedureState<TEnvironment> procStack = rollbackStack.get(rootProcId); 483 procStack.loadStack(proc); 484 485 proc.setRootProcId(rootProcId); 486 switch (proc.getState()) { 487 case RUNNABLE: 488 runnableList.add(proc); 489 break; 490 case WAITING: 491 waitingList.add(proc); 492 break; 493 case WAITING_TIMEOUT: 494 waitingTimeoutList.add(proc); 495 break; 496 case FAILED: 497 failedList.add(proc); 498 break; 499 case ROLLEDBACK: 500 case INITIALIZING: 501 String msg = "Unexpected " + proc.getState() + " state for " + proc; 502 LOG.error(msg); 503 throw new UnsupportedOperationException(msg); 504 default: 505 break; 506 } 507 } 508 509 // 3. Check the waiting procedures to see if some of them can be added to runnable. 510 waitingList.forEach(proc -> { 511 if (!proc.hasChildren()) { 512 // Normally, WAITING procedures should be waken by its children. But, there is a case that, 513 // all the children are successful and before they can wake up their parent procedure, the 514 // master was killed. So, during recovering the procedures from ProcedureWal, its children 515 // are not loaded because of their SUCCESS state. So we need to continue to run this WAITING 516 // procedure. But before executing, we need to set its state to RUNNABLE, otherwise, a 517 // exception will throw: 518 // Preconditions.checkArgument(procedure.getState() == ProcedureState.RUNNABLE, 519 // "NOT RUNNABLE! " + procedure.toString()); 520 proc.setState(ProcedureState.RUNNABLE); 521 runnableList.add(proc); 522 } else { 523 proc.afterReplay(getEnvironment()); 524 } 525 }); 526 // 4. restore locks 527 restoreLocks(); 528 529 // 5. Push the procedures to the timeout executor 530 waitingTimeoutList.forEach(proc -> { 531 proc.afterReplay(getEnvironment()); 532 timeoutExecutor.add(proc); 533 }); 534 535 // 6. Push the procedure to the scheduler 536 failedList.forEach(scheduler::addBack); 537 runnableList.forEach(p -> { 538 p.afterReplay(getEnvironment()); 539 if (!p.hasParent()) { 540 sendProcedureLoadedNotification(p.getProcId()); 541 } 542 scheduler.addBack(p); 543 }); 544 // After all procedures put into the queue, signal the worker threads. 545 // Otherwise, there is a race condition. See HBASE-21364. 546 scheduler.signalAll(); 547 } 548 549 /** 550 * Initialize the procedure executor, but do not start workers. We will start them later. 551 * <p/> 552 * It calls ProcedureStore.recoverLease() and ProcedureStore.load() to recover the lease, and 553 * ensure a single executor, and start the procedure replay to resume and recover the previous 554 * pending and in-progress procedures. 555 * @param numThreads number of threads available for procedure execution. 556 * @param abortOnCorruption true if you want to abort your service in case a corrupted procedure 557 * is found on replay. otherwise false. 558 */ 559 public void init(int numThreads, boolean abortOnCorruption) throws IOException { 560 // We have numThreads executor + one timer thread used for timing out 561 // procedures and triggering periodic procedures. 562 this.corePoolSize = numThreads; 563 this.maxPoolSize = 10 * numThreads; 564 LOG.info("Starting {} core workers (bigger of cpus/4 or 16) with max (burst) worker count={}", 565 corePoolSize, maxPoolSize); 566 567 this.threadGroup = new ThreadGroup("PEWorkerGroup"); 568 this.timeoutExecutor = new TimeoutExecutorThread<>(this, threadGroup, "ProcExecTimeout"); 569 this.workerMonitorExecutor = new TimeoutExecutorThread<>(this, threadGroup, "WorkerMonitor"); 570 571 // Create the workers 572 workerId.set(0); 573 workerThreads = new CopyOnWriteArrayList<>(); 574 for (int i = 0; i < corePoolSize; ++i) { 575 workerThreads.add(new WorkerThread(threadGroup)); 576 } 577 578 long st, et; 579 580 // Acquire the store lease. 581 st = System.nanoTime(); 582 store.recoverLease(); 583 et = System.nanoTime(); 584 LOG.info("Recovered {} lease in {}", store.getClass().getSimpleName(), 585 StringUtils.humanTimeDiff(TimeUnit.NANOSECONDS.toMillis(et - st))); 586 587 // start the procedure scheduler 588 scheduler.start(); 589 590 // TODO: Split in two steps. 591 // TODO: Handle corrupted procedures (currently just a warn) 592 // The first one will make sure that we have the latest id, 593 // so we can start the threads and accept new procedures. 594 // The second step will do the actual load of old procedures. 595 st = System.nanoTime(); 596 load(abortOnCorruption); 597 et = System.nanoTime(); 598 LOG.info("Loaded {} in {}", store.getClass().getSimpleName(), 599 StringUtils.humanTimeDiff(TimeUnit.NANOSECONDS.toMillis(et - st))); 600 } 601 602 /** 603 * Start the workers. 604 */ 605 public void startWorkers() throws IOException { 606 if (!running.compareAndSet(false, true)) { 607 LOG.warn("Already running"); 608 return; 609 } 610 // Start the executors. Here we must have the lastProcId set. 611 LOG.trace("Start workers {}", workerThreads.size()); 612 timeoutExecutor.start(); 613 workerMonitorExecutor.start(); 614 for (WorkerThread worker : workerThreads) { 615 worker.start(); 616 } 617 618 // Internal chores 619 workerMonitorExecutor.add(new WorkerMonitor()); 620 621 // Add completed cleaner chore 622 addChore(new CompletedProcedureCleaner<>(conf, store, procExecutionLock, completed, 623 nonceKeysToProcIdsMap)); 624 } 625 626 public void stop() { 627 if (!running.getAndSet(false)) { 628 return; 629 } 630 631 LOG.info("Stopping"); 632 scheduler.stop(); 633 timeoutExecutor.sendStopSignal(); 634 workerMonitorExecutor.sendStopSignal(); 635 } 636 637 public void join() { 638 assert !isRunning() : "expected not running"; 639 640 // stop the timeout executor 641 timeoutExecutor.awaitTermination(); 642 // stop the work monitor executor 643 workerMonitorExecutor.awaitTermination(); 644 645 // stop the worker threads 646 for (WorkerThread worker : workerThreads) { 647 worker.awaitTermination(); 648 } 649 650 // Destroy the Thread Group for the executors 651 // TODO: Fix. #join is not place to destroy resources. 652 try { 653 threadGroup.destroy(); 654 } catch (IllegalThreadStateException e) { 655 LOG.error("ThreadGroup {} contains running threads; {}: See STDOUT", this.threadGroup, 656 e.getMessage()); 657 // This dumps list of threads on STDOUT. 658 this.threadGroup.list(); 659 } 660 661 // reset the in-memory state for testing 662 completed.clear(); 663 rollbackStack.clear(); 664 procedures.clear(); 665 nonceKeysToProcIdsMap.clear(); 666 scheduler.clear(); 667 lastProcId.set(-1); 668 } 669 670 public void refreshConfiguration(final Configuration conf) { 671 this.conf = conf; 672 setKeepAliveTime(conf.getLong(WORKER_KEEP_ALIVE_TIME_CONF_KEY, DEFAULT_WORKER_KEEP_ALIVE_TIME), 673 TimeUnit.MILLISECONDS); 674 } 675 676 // ========================================================================== 677 // Accessors 678 // ========================================================================== 679 public boolean isRunning() { 680 return running.get(); 681 } 682 683 /** Returns the current number of worker threads. */ 684 public int getWorkerThreadCount() { 685 return workerThreads.size(); 686 } 687 688 /** Returns the core pool size settings. */ 689 public int getCorePoolSize() { 690 return corePoolSize; 691 } 692 693 public int getActiveExecutorCount() { 694 return activeExecutorCount.get(); 695 } 696 697 public TEnvironment getEnvironment() { 698 return this.environment; 699 } 700 701 public ProcedureStore getStore() { 702 return this.store; 703 } 704 705 ProcedureScheduler getScheduler() { 706 return scheduler; 707 } 708 709 public void setKeepAliveTime(final long keepAliveTime, final TimeUnit timeUnit) { 710 this.keepAliveTime = timeUnit.toMillis(keepAliveTime); 711 this.scheduler.signalAll(); 712 } 713 714 public long getKeepAliveTime(final TimeUnit timeUnit) { 715 return timeUnit.convert(keepAliveTime, TimeUnit.MILLISECONDS); 716 } 717 718 // ========================================================================== 719 // Submit/Remove Chores 720 // ========================================================================== 721 722 /** 723 * Add a chore procedure to the executor 724 * @param chore the chore to add 725 */ 726 public void addChore(@Nullable ProcedureInMemoryChore<TEnvironment> chore) { 727 if (chore == null) { 728 return; 729 } 730 chore.setState(ProcedureState.WAITING_TIMEOUT); 731 timeoutExecutor.add(chore); 732 } 733 734 /** 735 * Remove a chore procedure from the executor 736 * @param chore the chore to remove 737 * @return whether the chore is removed, or it will be removed later 738 */ 739 public boolean removeChore(@Nullable ProcedureInMemoryChore<TEnvironment> chore) { 740 if (chore == null) { 741 return true; 742 } 743 chore.setState(ProcedureState.SUCCESS); 744 return timeoutExecutor.remove(chore); 745 } 746 747 // ========================================================================== 748 // Nonce Procedure helpers 749 // ========================================================================== 750 /** 751 * Create a NonceKey from the specified nonceGroup and nonce. 752 * @param nonceGroup the group to use for the {@link NonceKey} 753 * @param nonce the nonce to use in the {@link NonceKey} 754 * @return the generated NonceKey 755 */ 756 public NonceKey createNonceKey(final long nonceGroup, final long nonce) { 757 return (nonce == HConstants.NO_NONCE) ? null : new NonceKey(nonceGroup, nonce); 758 } 759 760 /** 761 * Register a nonce for a procedure that is going to be submitted. A procId will be reserved and 762 * on submitProcedure(), the procedure with the specified nonce will take the reserved ProcId. If 763 * someone already reserved the nonce, this method will return the procId reserved, otherwise an 764 * invalid procId will be returned. and the caller should procede and submit the procedure. 765 * @param nonceKey A unique identifier for this operation from the client or process. 766 * @return the procId associated with the nonce, if any otherwise an invalid procId. 767 */ 768 public long registerNonce(final NonceKey nonceKey) { 769 if (nonceKey == null) { 770 return -1; 771 } 772 773 // check if we have already a Reserved ID for the nonce 774 Long oldProcId = nonceKeysToProcIdsMap.get(nonceKey); 775 if (oldProcId == null) { 776 // reserve a new Procedure ID, this will be associated with the nonce 777 // and the procedure submitted with the specified nonce will use this ID. 778 final long newProcId = nextProcId(); 779 oldProcId = nonceKeysToProcIdsMap.putIfAbsent(nonceKey, newProcId); 780 if (oldProcId == null) { 781 return -1; 782 } 783 } 784 785 // we found a registered nonce, but the procedure may not have been submitted yet. 786 // since the client expect the procedure to be submitted, spin here until it is. 787 final boolean traceEnabled = LOG.isTraceEnabled(); 788 while ( 789 isRunning() && !(procedures.containsKey(oldProcId) || completed.containsKey(oldProcId)) 790 && nonceKeysToProcIdsMap.containsKey(nonceKey) 791 ) { 792 if (traceEnabled) { 793 LOG.trace("Waiting for pid=" + oldProcId.longValue() + " to be submitted"); 794 } 795 Threads.sleep(100); 796 } 797 return oldProcId.longValue(); 798 } 799 800 /** 801 * Remove the NonceKey if the procedure was not submitted to the executor. 802 * @param nonceKey A unique identifier for this operation from the client or process. 803 */ 804 public void unregisterNonceIfProcedureWasNotSubmitted(final NonceKey nonceKey) { 805 if (nonceKey == null) { 806 return; 807 } 808 809 final Long procId = nonceKeysToProcIdsMap.get(nonceKey); 810 if (procId == null) { 811 return; 812 } 813 814 // if the procedure was not submitted, remove the nonce 815 if (!(procedures.containsKey(procId) || completed.containsKey(procId))) { 816 nonceKeysToProcIdsMap.remove(nonceKey); 817 } 818 } 819 820 /** 821 * If the failure failed before submitting it, we may want to give back the same error to the 822 * requests with the same nonceKey. 823 * @param nonceKey A unique identifier for this operation from the client or process 824 * @param procName name of the procedure, used to inform the user 825 * @param procOwner name of the owner of the procedure, used to inform the user 826 * @param exception the failure to report to the user 827 */ 828 public void setFailureResultForNonce(NonceKey nonceKey, String procName, User procOwner, 829 IOException exception) { 830 if (nonceKey == null) { 831 return; 832 } 833 834 Long procId = nonceKeysToProcIdsMap.get(nonceKey); 835 if (procId == null || completed.containsKey(procId)) { 836 return; 837 } 838 839 Procedure<TEnvironment> proc = 840 new FailedProcedure<>(procId.longValue(), procName, procOwner, nonceKey, exception); 841 842 completed.putIfAbsent(procId, new CompletedProcedureRetainer<>(proc)); 843 } 844 845 // ========================================================================== 846 // Submit/Abort Procedure 847 // ========================================================================== 848 /** 849 * Add a new root-procedure to the executor. 850 * @param proc the new procedure to execute. 851 * @return the procedure id, that can be used to monitor the operation 852 */ 853 public long submitProcedure(Procedure<TEnvironment> proc) { 854 return submitProcedure(proc, null); 855 } 856 857 /** 858 * Bypass a procedure. If the procedure is set to bypass, all the logic in execute/rollback will 859 * be ignored and it will return success, whatever. It is used to recover buggy stuck procedures, 860 * releasing the lock resources and letting other procedures run. Bypassing one procedure (and its 861 * ancestors will be bypassed automatically) may leave the cluster in a middle state, e.g. region 862 * not assigned, or some hdfs files left behind. After getting rid of those stuck procedures, the 863 * operators may have to do some clean up on hdfs or schedule some assign procedures to let region 864 * online. DO AT YOUR OWN RISK. 865 * <p> 866 * A procedure can be bypassed only if 1. The procedure is in state of RUNNABLE, WAITING, 867 * WAITING_TIMEOUT or it is a root procedure without any child. 2. No other worker thread is 868 * executing it 3. No child procedure has been submitted 869 * <p> 870 * If all the requirements are meet, the procedure and its ancestors will be bypassed and 871 * persisted to WAL. 872 * <p> 873 * If the procedure is in WAITING state, will set it to RUNNABLE add it to run queue. TODO: What 874 * about WAITING_TIMEOUT? 875 * @param pids the procedure id 876 * @param lockWait time to wait lock 877 * @param force if force set to true, we will bypass the procedure even if it is executing. 878 * This is for procedures which can't break out during executing(due to bug, 879 * mostly) In this case, bypassing the procedure is not enough, since it is 880 * already stuck there. We need to restart the master after bypassing, and 881 * letting the problematic procedure to execute wth bypass=true, so in that 882 * condition, the procedure can be successfully bypassed. 883 * @param recursive We will do an expensive search for children of each pid. EXPENSIVE! 884 * @return true if bypass success 885 * @throws IOException IOException 886 */ 887 public List<Boolean> bypassProcedure(List<Long> pids, long lockWait, boolean force, 888 boolean recursive) throws IOException { 889 List<Boolean> result = new ArrayList<Boolean>(pids.size()); 890 for (long pid : pids) { 891 result.add(bypassProcedure(pid, lockWait, force, recursive)); 892 } 893 return result; 894 } 895 896 boolean bypassProcedure(long pid, long lockWait, boolean override, boolean recursive) 897 throws IOException { 898 Preconditions.checkArgument(lockWait > 0, "lockWait should be positive"); 899 final Procedure<TEnvironment> procedure = getProcedure(pid); 900 if (procedure == null) { 901 LOG.debug("Procedure pid={} does not exist, skipping bypass", pid); 902 return false; 903 } 904 905 LOG.debug("Begin bypass {} with lockWait={}, override={}, recursive={}", procedure, lockWait, 906 override, recursive); 907 IdLock.Entry lockEntry = procExecutionLock.tryLockEntry(procedure.getProcId(), lockWait); 908 if (lockEntry == null && !override) { 909 LOG.debug("Waited {} ms, but {} is still running, skipping bypass with force={}", lockWait, 910 procedure, override); 911 return false; 912 } else if (lockEntry == null) { 913 LOG.debug("Waited {} ms, but {} is still running, begin bypass with force={}", lockWait, 914 procedure, override); 915 } 916 try { 917 // check whether the procedure is already finished 918 if (procedure.isFinished()) { 919 LOG.debug("{} is already finished, skipping bypass", procedure); 920 return false; 921 } 922 923 if (procedure.hasChildren()) { 924 if (recursive) { 925 // EXPENSIVE. Checks each live procedure of which there could be many!!! 926 // Is there another way to get children of a procedure? 927 LOG.info("Recursive bypass on children of pid={}", procedure.getProcId()); 928 this.procedures.forEachValue(1 /* Single-threaded */, 929 // Transformer 930 v -> v.getParentProcId() == procedure.getProcId() ? v : null, 931 // Consumer 932 v -> { 933 try { 934 bypassProcedure(v.getProcId(), lockWait, override, recursive); 935 } catch (IOException e) { 936 LOG.warn("Recursive bypass of pid={}", v.getProcId(), e); 937 } 938 }); 939 } else { 940 LOG.debug("{} has children, skipping bypass", procedure); 941 return false; 942 } 943 } 944 945 // If the procedure has no parent or no child, we are safe to bypass it in whatever state 946 if ( 947 procedure.hasParent() && procedure.getState() != ProcedureState.RUNNABLE 948 && procedure.getState() != ProcedureState.WAITING 949 && procedure.getState() != ProcedureState.WAITING_TIMEOUT 950 ) { 951 LOG.debug("Bypassing procedures in RUNNABLE, WAITING and WAITING_TIMEOUT states " 952 + "(with no parent), {}", procedure); 953 // Question: how is the bypass done here? 954 return false; 955 } 956 957 // Now, the procedure is not finished, and no one can execute it since we take the lock now 958 // And we can be sure that its ancestor is not running too, since their child has not 959 // finished yet 960 Procedure<TEnvironment> current = procedure; 961 while (current != null) { 962 LOG.debug("Bypassing {}", current); 963 current.bypass(getEnvironment()); 964 store.update(current); 965 long parentID = current.getParentProcId(); 966 current = getProcedure(parentID); 967 } 968 969 // wake up waiting procedure, already checked there is no child 970 if (procedure.getState() == ProcedureState.WAITING) { 971 procedure.setState(ProcedureState.RUNNABLE); 972 store.update(procedure); 973 } 974 975 // If state of procedure is WAITING_TIMEOUT, we can directly submit it to the scheduler. 976 // Instead we should remove it from timeout Executor queue and tranfer its state to RUNNABLE 977 if (procedure.getState() == ProcedureState.WAITING_TIMEOUT) { 978 LOG.debug("transform procedure {} from WAITING_TIMEOUT to RUNNABLE", procedure); 979 if (timeoutExecutor.remove(procedure)) { 980 LOG.debug("removed procedure {} from timeoutExecutor", procedure); 981 timeoutExecutor.executeTimedoutProcedure(procedure); 982 } 983 } else if (lockEntry != null) { 984 scheduler.addFront(procedure); 985 LOG.debug("Bypassing {} and its ancestors successfully, adding to queue", procedure); 986 } else { 987 // If we don't have the lock, we can't re-submit the queue, 988 // since it is already executing. To get rid of the stuck situation, we 989 // need to restart the master. With the procedure set to bypass, the procedureExecutor 990 // will bypass it and won't get stuck again. 991 LOG.debug("Bypassing {} and its ancestors successfully, but since it is already running, " 992 + "skipping add to queue", procedure); 993 } 994 return true; 995 996 } finally { 997 if (lockEntry != null) { 998 procExecutionLock.releaseLockEntry(lockEntry); 999 } 1000 } 1001 } 1002 1003 /** 1004 * Add a new root-procedure to the executor. 1005 * @param proc the new procedure to execute. 1006 * @param nonceKey the registered unique identifier for this operation from the client or process. 1007 * @return the procedure id, that can be used to monitor the operation 1008 */ 1009 @edu.umd.cs.findbugs.annotations.SuppressWarnings(value = "NP_NULL_ON_SOME_PATH", 1010 justification = "FindBugs is blind to the check-for-null") 1011 public long submitProcedure(Procedure<TEnvironment> proc, NonceKey nonceKey) { 1012 Preconditions.checkArgument(lastProcId.get() >= 0); 1013 1014 prepareProcedure(proc); 1015 1016 final Long currentProcId; 1017 if (nonceKey != null) { 1018 currentProcId = nonceKeysToProcIdsMap.get(nonceKey); 1019 Preconditions.checkArgument(currentProcId != null, 1020 "Expected nonceKey=" + nonceKey + " to be reserved, use registerNonce(); proc=" + proc); 1021 } else { 1022 currentProcId = nextProcId(); 1023 } 1024 1025 // Initialize the procedure 1026 proc.setNonceKey(nonceKey); 1027 proc.setProcId(currentProcId.longValue()); 1028 1029 // Commit the transaction 1030 store.insert(proc, null); 1031 LOG.debug("Stored {}", proc); 1032 1033 // Add the procedure to the executor 1034 return pushProcedure(proc); 1035 } 1036 1037 /** 1038 * Add a set of new root-procedure to the executor. 1039 * @param procs the new procedures to execute. 1040 */ 1041 // TODO: Do we need to take nonces here? 1042 public void submitProcedures(Procedure<TEnvironment>[] procs) { 1043 Preconditions.checkArgument(lastProcId.get() >= 0); 1044 if (procs == null || procs.length <= 0) { 1045 return; 1046 } 1047 1048 // Prepare procedure 1049 for (int i = 0; i < procs.length; ++i) { 1050 prepareProcedure(procs[i]).setProcId(nextProcId()); 1051 } 1052 1053 // Commit the transaction 1054 store.insert(procs); 1055 if (LOG.isDebugEnabled()) { 1056 LOG.debug("Stored " + Arrays.toString(procs)); 1057 } 1058 1059 // Add the procedure to the executor 1060 for (int i = 0; i < procs.length; ++i) { 1061 pushProcedure(procs[i]); 1062 } 1063 } 1064 1065 private Procedure<TEnvironment> prepareProcedure(Procedure<TEnvironment> proc) { 1066 Preconditions.checkArgument(proc.getState() == ProcedureState.INITIALIZING); 1067 Preconditions.checkArgument(!proc.hasParent(), "unexpected parent", proc); 1068 if (this.checkOwnerSet) { 1069 Preconditions.checkArgument(proc.hasOwner(), "missing owner"); 1070 } 1071 return proc; 1072 } 1073 1074 private long pushProcedure(Procedure<TEnvironment> proc) { 1075 final long currentProcId = proc.getProcId(); 1076 1077 // Update metrics on start of a procedure 1078 proc.updateMetricsOnSubmit(getEnvironment()); 1079 1080 // Create the rollback stack for the procedure 1081 RootProcedureState<TEnvironment> stack = new RootProcedureState<>(); 1082 rollbackStack.put(currentProcId, stack); 1083 1084 // Submit the new subprocedures 1085 assert !procedures.containsKey(currentProcId); 1086 procedures.put(currentProcId, proc); 1087 sendProcedureAddedNotification(currentProcId); 1088 scheduler.addBack(proc); 1089 return proc.getProcId(); 1090 } 1091 1092 /** 1093 * Send an abort notification the specified procedure. Depending on the procedure implementation 1094 * the abort can be considered or ignored. 1095 * @param procId the procedure to abort 1096 * @return true if the procedure exists and has received the abort, otherwise false. 1097 */ 1098 public boolean abort(long procId) { 1099 return abort(procId, true); 1100 } 1101 1102 /** 1103 * Send an abort notification to the specified procedure. Depending on the procedure 1104 * implementation, the abort can be considered or ignored. 1105 * @param procId the procedure to abort 1106 * @param mayInterruptIfRunning if the proc completed at least one step, should it be aborted? 1107 * @return true if the procedure exists and has received the abort, otherwise false. 1108 */ 1109 public boolean abort(long procId, boolean mayInterruptIfRunning) { 1110 Procedure<TEnvironment> proc = procedures.get(procId); 1111 if (proc != null) { 1112 if (!mayInterruptIfRunning && proc.wasExecuted()) { 1113 return false; 1114 } 1115 return proc.abort(getEnvironment()); 1116 } 1117 return false; 1118 } 1119 1120 // ========================================================================== 1121 // Executor query helpers 1122 // ========================================================================== 1123 public Procedure<TEnvironment> getProcedure(final long procId) { 1124 return procedures.get(procId); 1125 } 1126 1127 public <T extends Procedure<TEnvironment>> T getProcedure(Class<T> clazz, long procId) { 1128 Procedure<TEnvironment> proc = getProcedure(procId); 1129 if (clazz.isInstance(proc)) { 1130 return clazz.cast(proc); 1131 } 1132 return null; 1133 } 1134 1135 public Procedure<TEnvironment> getResult(long procId) { 1136 CompletedProcedureRetainer<TEnvironment> retainer = completed.get(procId); 1137 if (retainer == null) { 1138 return null; 1139 } else { 1140 return retainer.getProcedure(); 1141 } 1142 } 1143 1144 /** 1145 * Return true if the procedure is finished. The state may be "completed successfully" or "failed 1146 * and rolledback". Use getResult() to check the state or get the result data. 1147 * @param procId the ID of the procedure to check 1148 * @return true if the procedure execution is finished, otherwise false. 1149 */ 1150 public boolean isFinished(final long procId) { 1151 return !procedures.containsKey(procId); 1152 } 1153 1154 /** 1155 * Return true if the procedure is started. 1156 * @param procId the ID of the procedure to check 1157 * @return true if the procedure execution is started, otherwise false. 1158 */ 1159 public boolean isStarted(long procId) { 1160 Procedure<?> proc = procedures.get(procId); 1161 if (proc == null) { 1162 return completed.get(procId) != null; 1163 } 1164 return proc.wasExecuted(); 1165 } 1166 1167 /** 1168 * Mark the specified completed procedure, as ready to remove. 1169 * @param procId the ID of the procedure to remove 1170 */ 1171 public void removeResult(long procId) { 1172 CompletedProcedureRetainer<TEnvironment> retainer = completed.get(procId); 1173 if (retainer == null) { 1174 assert !procedures.containsKey(procId) : "pid=" + procId + " is still running"; 1175 LOG.debug("pid={} already removed by the cleaner.", procId); 1176 return; 1177 } 1178 1179 // The CompletedProcedureCleaner will take care of deletion, once the TTL is expired. 1180 retainer.setClientAckTime(EnvironmentEdgeManager.currentTime()); 1181 } 1182 1183 public Procedure<TEnvironment> getResultOrProcedure(long procId) { 1184 CompletedProcedureRetainer<TEnvironment> retainer = completed.get(procId); 1185 if (retainer == null) { 1186 return procedures.get(procId); 1187 } else { 1188 return retainer.getProcedure(); 1189 } 1190 } 1191 1192 /** 1193 * Check if the user is this procedure's owner 1194 * @param procId the target procedure 1195 * @param user the user 1196 * @return true if the user is the owner of the procedure, false otherwise or the owner is 1197 * unknown. 1198 */ 1199 public boolean isProcedureOwner(long procId, User user) { 1200 if (user == null) { 1201 return false; 1202 } 1203 final Procedure<TEnvironment> runningProc = procedures.get(procId); 1204 if (runningProc != null) { 1205 return runningProc.getOwner().equals(user.getShortName()); 1206 } 1207 1208 final CompletedProcedureRetainer<TEnvironment> retainer = completed.get(procId); 1209 if (retainer != null) { 1210 return retainer.getProcedure().getOwner().equals(user.getShortName()); 1211 } 1212 1213 // Procedure either does not exist or has already completed and got cleaned up. 1214 // At this time, we cannot check the owner of the procedure 1215 return false; 1216 } 1217 1218 /** 1219 * Should only be used when starting up, where the procedure workers have not been started. 1220 * <p/> 1221 * If the procedure works has been started, the return values maybe changed when you are 1222 * processing it so usually this is not safe. Use {@link #getProcedures()} below for most cases as 1223 * it will do a copy, and also include the finished procedures. 1224 */ 1225 public Collection<Procedure<TEnvironment>> getActiveProceduresNoCopy() { 1226 return procedures.values(); 1227 } 1228 1229 /** 1230 * Get procedures. 1231 * @return the procedures in a list 1232 */ 1233 public List<Procedure<TEnvironment>> getProcedures() { 1234 List<Procedure<TEnvironment>> procedureList = 1235 new ArrayList<>(procedures.size() + completed.size()); 1236 procedureList.addAll(procedures.values()); 1237 // Note: The procedure could show up twice in the list with different state, as 1238 // it could complete after we walk through procedures list and insert into 1239 // procedureList - it is ok, as we will use the information in the Procedure 1240 // to figure it out; to prevent this would increase the complexity of the logic. 1241 completed.values().stream().map(CompletedProcedureRetainer::getProcedure) 1242 .forEach(procedureList::add); 1243 return procedureList; 1244 } 1245 1246 // ========================================================================== 1247 // Listeners helpers 1248 // ========================================================================== 1249 public void registerListener(ProcedureExecutorListener listener) { 1250 this.listeners.add(listener); 1251 } 1252 1253 public boolean unregisterListener(ProcedureExecutorListener listener) { 1254 return this.listeners.remove(listener); 1255 } 1256 1257 private void sendProcedureLoadedNotification(final long procId) { 1258 if (!this.listeners.isEmpty()) { 1259 for (ProcedureExecutorListener listener : this.listeners) { 1260 try { 1261 listener.procedureLoaded(procId); 1262 } catch (Throwable e) { 1263 LOG.error("Listener " + listener + " had an error: " + e.getMessage(), e); 1264 } 1265 } 1266 } 1267 } 1268 1269 private void sendProcedureAddedNotification(final long procId) { 1270 if (!this.listeners.isEmpty()) { 1271 for (ProcedureExecutorListener listener : this.listeners) { 1272 try { 1273 listener.procedureAdded(procId); 1274 } catch (Throwable e) { 1275 LOG.error("Listener " + listener + " had an error: " + e.getMessage(), e); 1276 } 1277 } 1278 } 1279 } 1280 1281 private void sendProcedureFinishedNotification(final long procId) { 1282 if (!this.listeners.isEmpty()) { 1283 for (ProcedureExecutorListener listener : this.listeners) { 1284 try { 1285 listener.procedureFinished(procId); 1286 } catch (Throwable e) { 1287 LOG.error("Listener " + listener + " had an error: " + e.getMessage(), e); 1288 } 1289 } 1290 } 1291 } 1292 1293 // ========================================================================== 1294 // Procedure IDs helpers 1295 // ========================================================================== 1296 private long nextProcId() { 1297 long procId = lastProcId.incrementAndGet(); 1298 if (procId < 0) { 1299 while (!lastProcId.compareAndSet(procId, 0)) { 1300 procId = lastProcId.get(); 1301 if (procId >= 0) { 1302 break; 1303 } 1304 } 1305 while (procedures.containsKey(procId)) { 1306 procId = lastProcId.incrementAndGet(); 1307 } 1308 } 1309 assert procId >= 0 : "Invalid procId " + procId; 1310 return procId; 1311 } 1312 1313 protected long getLastProcId() { 1314 return lastProcId.get(); 1315 } 1316 1317 public Set<Long> getActiveProcIds() { 1318 return procedures.keySet(); 1319 } 1320 1321 Long getRootProcedureId(Procedure<TEnvironment> proc) { 1322 return Procedure.getRootProcedureId(procedures, proc); 1323 } 1324 1325 // ========================================================================== 1326 // Executions 1327 // ========================================================================== 1328 private void executeProcedure(Procedure<TEnvironment> proc) { 1329 if (proc.isFinished()) { 1330 LOG.debug("{} is already finished, skipping execution", proc); 1331 return; 1332 } 1333 final Long rootProcId = getRootProcedureId(proc); 1334 if (rootProcId == null) { 1335 // The 'proc' was ready to run but the root procedure was rolledback 1336 LOG.warn("Rollback because parent is done/rolledback proc=" + proc); 1337 executeRollback(proc); 1338 return; 1339 } 1340 1341 RootProcedureState<TEnvironment> procStack = rollbackStack.get(rootProcId); 1342 if (procStack == null) { 1343 LOG.warn("RootProcedureState is null for " + proc.getProcId()); 1344 return; 1345 } 1346 do { 1347 // Try to acquire the execution 1348 if (!procStack.acquire(proc)) { 1349 if (procStack.setRollback()) { 1350 // we have the 'rollback-lock' we can start rollingback 1351 switch (executeRollback(rootProcId, procStack)) { 1352 case LOCK_ACQUIRED: 1353 break; 1354 case LOCK_YIELD_WAIT: 1355 procStack.unsetRollback(); 1356 scheduler.yield(proc); 1357 break; 1358 case LOCK_EVENT_WAIT: 1359 LOG.info("LOCK_EVENT_WAIT rollback..." + proc); 1360 procStack.unsetRollback(); 1361 break; 1362 default: 1363 throw new UnsupportedOperationException(); 1364 } 1365 } else { 1366 // if we can't rollback means that some child is still running. 1367 // the rollback will be executed after all the children are done. 1368 // If the procedure was never executed, remove and mark it as rolledback. 1369 if (!proc.wasExecuted()) { 1370 switch (executeRollback(proc)) { 1371 case LOCK_ACQUIRED: 1372 break; 1373 case LOCK_YIELD_WAIT: 1374 scheduler.yield(proc); 1375 break; 1376 case LOCK_EVENT_WAIT: 1377 LOG.info("LOCK_EVENT_WAIT can't rollback child running?..." + proc); 1378 break; 1379 default: 1380 throw new UnsupportedOperationException(); 1381 } 1382 } 1383 } 1384 break; 1385 } 1386 1387 // Execute the procedure 1388 assert proc.getState() == ProcedureState.RUNNABLE : proc; 1389 // Note that lock is NOT about concurrency but rather about ensuring 1390 // ownership of a procedure of an entity such as a region or table 1391 LockState lockState = acquireLock(proc); 1392 switch (lockState) { 1393 case LOCK_ACQUIRED: 1394 execProcedure(procStack, proc); 1395 break; 1396 case LOCK_YIELD_WAIT: 1397 LOG.info(lockState + " " + proc); 1398 scheduler.yield(proc); 1399 break; 1400 case LOCK_EVENT_WAIT: 1401 // Someone will wake us up when the lock is available 1402 LOG.debug(lockState + " " + proc); 1403 break; 1404 default: 1405 throw new UnsupportedOperationException(); 1406 } 1407 procStack.release(proc); 1408 1409 if (proc.isSuccess()) { 1410 // update metrics on finishing the procedure 1411 proc.updateMetricsOnFinish(getEnvironment(), proc.elapsedTime(), true); 1412 LOG.info("Finished " + proc + " in " + StringUtils.humanTimeDiff(proc.elapsedTime())); 1413 // Finalize the procedure state 1414 if (proc.getProcId() == rootProcId) { 1415 procedureFinished(proc); 1416 } else { 1417 execCompletionCleanup(proc); 1418 } 1419 break; 1420 } 1421 } while (procStack.isFailed()); 1422 } 1423 1424 private LockState acquireLock(Procedure<TEnvironment> proc) { 1425 TEnvironment env = getEnvironment(); 1426 // if holdLock is true, then maybe we already have the lock, so just return LOCK_ACQUIRED if 1427 // hasLock is true. 1428 if (proc.hasLock()) { 1429 return LockState.LOCK_ACQUIRED; 1430 } 1431 return proc.doAcquireLock(env, store); 1432 } 1433 1434 private void releaseLock(Procedure<TEnvironment> proc, boolean force) { 1435 TEnvironment env = getEnvironment(); 1436 // For how the framework works, we know that we will always have the lock 1437 // when we call releaseLock(), so we can avoid calling proc.hasLock() 1438 if (force || !proc.holdLock(env) || proc.isFinished()) { 1439 proc.doReleaseLock(env, store); 1440 } 1441 } 1442 1443 /** 1444 * Execute the rollback of the full procedure stack. Once the procedure is rolledback, the 1445 * root-procedure will be visible as finished to user, and the result will be the fatal exception. 1446 */ 1447 private LockState executeRollback(long rootProcId, RootProcedureState<TEnvironment> procStack) { 1448 Procedure<TEnvironment> rootProc = procedures.get(rootProcId); 1449 RemoteProcedureException exception = rootProc.getException(); 1450 // TODO: This needs doc. The root proc doesn't have an exception. Maybe we are 1451 // rolling back because the subprocedure does. Clarify. 1452 if (exception == null) { 1453 exception = procStack.getException(); 1454 rootProc.setFailure(exception); 1455 store.update(rootProc); 1456 } 1457 1458 List<Procedure<TEnvironment>> subprocStack = procStack.getSubproceduresStack(); 1459 assert subprocStack != null : "Called rollback with no steps executed rootProc=" + rootProc; 1460 1461 int stackTail = subprocStack.size(); 1462 while (stackTail-- > 0) { 1463 Procedure<TEnvironment> proc = subprocStack.get(stackTail); 1464 IdLock.Entry lockEntry = null; 1465 // Hold the execution lock if it is not held by us. The IdLock is not reentrant so we need 1466 // this check, as the worker will hold the lock before executing a procedure. This is the only 1467 // place where we may hold two procedure execution locks, and there is a fence in the 1468 // RootProcedureState where we can make sure that only one worker can execute the rollback of 1469 // a RootProcedureState, so there is no dead lock problem. And the lock here is necessary to 1470 // prevent race between us and the force update thread. 1471 if (!procExecutionLock.isHeldByCurrentThread(proc.getProcId())) { 1472 try { 1473 lockEntry = procExecutionLock.getLockEntry(proc.getProcId()); 1474 } catch (IOException e) { 1475 // can only happen if interrupted, so not a big deal to propagate it 1476 throw new UncheckedIOException(e); 1477 } 1478 } 1479 try { 1480 // For the sub procedures which are successfully finished, we do not rollback them. 1481 // Typically, if we want to rollback a procedure, we first need to rollback it, and then 1482 // recursively rollback its ancestors. The state changes which are done by sub procedures 1483 // should be handled by parent procedures when rolling back. For example, when rolling back 1484 // a MergeTableProcedure, we will schedule new procedures to bring the offline regions 1485 // online, instead of rolling back the original procedures which offlined the regions(in 1486 // fact these procedures can not be rolled back...). 1487 if (proc.isSuccess()) { 1488 // Just do the cleanup work, without actually executing the rollback 1489 subprocStack.remove(stackTail); 1490 cleanupAfterRollbackOneStep(proc); 1491 continue; 1492 } 1493 LockState lockState = acquireLock(proc); 1494 if (lockState != LockState.LOCK_ACQUIRED) { 1495 // can't take a lock on the procedure, add the root-proc back on the 1496 // queue waiting for the lock availability 1497 return lockState; 1498 } 1499 1500 lockState = executeRollback(proc); 1501 releaseLock(proc, false); 1502 boolean abortRollback = lockState != LockState.LOCK_ACQUIRED; 1503 abortRollback |= !isRunning() || !store.isRunning(); 1504 1505 // allows to kill the executor before something is stored to the wal. 1506 // useful to test the procedure recovery. 1507 if (abortRollback) { 1508 return lockState; 1509 } 1510 1511 subprocStack.remove(stackTail); 1512 1513 // if the procedure is kind enough to pass the slot to someone else, yield 1514 // if the proc is already finished, do not yield 1515 if (!proc.isFinished() && proc.isYieldAfterExecutionStep(getEnvironment())) { 1516 return LockState.LOCK_YIELD_WAIT; 1517 } 1518 1519 if (proc != rootProc) { 1520 execCompletionCleanup(proc); 1521 } 1522 } finally { 1523 if (lockEntry != null) { 1524 procExecutionLock.releaseLockEntry(lockEntry); 1525 } 1526 } 1527 } 1528 1529 // Finalize the procedure state 1530 LOG.info("Rolled back {} exec-time={}", rootProc, 1531 StringUtils.humanTimeDiff(rootProc.elapsedTime())); 1532 procedureFinished(rootProc); 1533 return LockState.LOCK_ACQUIRED; 1534 } 1535 1536 private void cleanupAfterRollbackOneStep(Procedure<TEnvironment> proc) { 1537 if (proc.removeStackIndex()) { 1538 if (!proc.isSuccess()) { 1539 proc.setState(ProcedureState.ROLLEDBACK); 1540 } 1541 1542 // update metrics on finishing the procedure (fail) 1543 proc.updateMetricsOnFinish(getEnvironment(), proc.elapsedTime(), false); 1544 1545 if (proc.hasParent()) { 1546 store.delete(proc.getProcId()); 1547 procedures.remove(proc.getProcId()); 1548 } else { 1549 final long[] childProcIds = rollbackStack.get(proc.getProcId()).getSubprocedureIds(); 1550 if (childProcIds != null) { 1551 store.delete(proc, childProcIds); 1552 } else { 1553 store.update(proc); 1554 } 1555 } 1556 } else { 1557 store.update(proc); 1558 } 1559 } 1560 1561 /** 1562 * Execute the rollback of the procedure step. It updates the store with the new state (stack 1563 * index) or will remove completly the procedure in case it is a child. 1564 */ 1565 private LockState executeRollback(Procedure<TEnvironment> proc) { 1566 try { 1567 proc.doRollback(getEnvironment()); 1568 } catch (IOException e) { 1569 LOG.debug("Roll back attempt failed for {}", proc, e); 1570 return LockState.LOCK_YIELD_WAIT; 1571 } catch (InterruptedException e) { 1572 handleInterruptedException(proc, e); 1573 return LockState.LOCK_YIELD_WAIT; 1574 } catch (Throwable e) { 1575 // Catch NullPointerExceptions or similar errors... 1576 LOG.error(HBaseMarkers.FATAL, "CODE-BUG: Uncaught runtime exception for " + proc, e); 1577 } 1578 1579 // allows to kill the executor before something is stored to the wal. 1580 // useful to test the procedure recovery. 1581 if (testing != null && testing.shouldKillBeforeStoreUpdate()) { 1582 String msg = "TESTING: Kill before store update"; 1583 LOG.debug(msg); 1584 stop(); 1585 throw new RuntimeException(msg); 1586 } 1587 1588 cleanupAfterRollbackOneStep(proc); 1589 1590 return LockState.LOCK_ACQUIRED; 1591 } 1592 1593 private void yieldProcedure(Procedure<TEnvironment> proc) { 1594 releaseLock(proc, false); 1595 scheduler.yield(proc); 1596 } 1597 1598 /** 1599 * Executes <code>procedure</code> 1600 * <ul> 1601 * <li>Calls the doExecute() of the procedure 1602 * <li>If the procedure execution didn't fail (i.e. valid user input) 1603 * <ul> 1604 * <li>...and returned subprocedures 1605 * <ul> 1606 * <li>The subprocedures are initialized. 1607 * <li>The subprocedures are added to the store 1608 * <li>The subprocedures are added to the runnable queue 1609 * <li>The procedure is now in a WAITING state, waiting for the subprocedures to complete 1610 * </ul> 1611 * </li> 1612 * <li>...if there are no subprocedure 1613 * <ul> 1614 * <li>the procedure completed successfully 1615 * <li>if there is a parent (WAITING) 1616 * <li>the parent state will be set to RUNNABLE 1617 * </ul> 1618 * </li> 1619 * </ul> 1620 * </li> 1621 * <li>In case of failure 1622 * <ul> 1623 * <li>The store is updated with the new state</li> 1624 * <li>The executor (caller of this method) will start the rollback of the procedure</li> 1625 * </ul> 1626 * </li> 1627 * </ul> 1628 */ 1629 private void execProcedure(RootProcedureState<TEnvironment> procStack, 1630 Procedure<TEnvironment> procedure) { 1631 Preconditions.checkArgument(procedure.getState() == ProcedureState.RUNNABLE, 1632 "NOT RUNNABLE! " + procedure.toString()); 1633 1634 // Procedures can suspend themselves. They skip out by throwing a ProcedureSuspendedException. 1635 // The exception is caught below and then we hurry to the exit without disturbing state. The 1636 // idea is that the processing of this procedure will be unsuspended later by an external event 1637 // such the report of a region open. 1638 boolean suspended = false; 1639 1640 // Whether to 're-' -execute; run through the loop again. 1641 boolean reExecute = false; 1642 1643 Procedure<TEnvironment>[] subprocs = null; 1644 do { 1645 reExecute = false; 1646 procedure.resetPersistence(); 1647 try { 1648 subprocs = procedure.doExecute(getEnvironment()); 1649 if (subprocs != null && subprocs.length == 0) { 1650 subprocs = null; 1651 } 1652 } catch (ProcedureSuspendedException e) { 1653 LOG.trace("Suspend {}", procedure); 1654 suspended = true; 1655 } catch (ProcedureYieldException e) { 1656 LOG.trace("Yield {}", procedure, e); 1657 yieldProcedure(procedure); 1658 return; 1659 } catch (InterruptedException e) { 1660 LOG.trace("Yield interrupt {}", procedure, e); 1661 handleInterruptedException(procedure, e); 1662 yieldProcedure(procedure); 1663 return; 1664 } catch (Throwable e) { 1665 // Catch NullPointerExceptions or similar errors... 1666 String msg = "CODE-BUG: Uncaught runtime exception: " + procedure; 1667 LOG.error(msg, e); 1668 procedure.setFailure(new RemoteProcedureException(msg, e)); 1669 } 1670 1671 if (!procedure.isFailed()) { 1672 if (subprocs != null) { 1673 if (subprocs.length == 1 && subprocs[0] == procedure) { 1674 // Procedure returned itself. Quick-shortcut for a state machine-like procedure; 1675 // i.e. we go around this loop again rather than go back out on the scheduler queue. 1676 subprocs = null; 1677 reExecute = true; 1678 LOG.trace("Short-circuit to next step on pid={}", procedure.getProcId()); 1679 } else { 1680 // Yield the current procedure, and make the subprocedure runnable 1681 // subprocs may come back 'null'. 1682 subprocs = initializeChildren(procStack, procedure, subprocs); 1683 LOG.info("Initialized subprocedures=" + (subprocs == null 1684 ? null 1685 : Stream.of(subprocs).map(e -> "{" + e.toString() + "}").collect(Collectors.toList()) 1686 .toString())); 1687 } 1688 } else if (procedure.getState() == ProcedureState.WAITING_TIMEOUT) { 1689 LOG.trace("Added to timeoutExecutor {}", procedure); 1690 timeoutExecutor.add(procedure); 1691 } else if (!suspended) { 1692 // No subtask, so we are done 1693 procedure.setState(ProcedureState.SUCCESS); 1694 } 1695 } 1696 1697 // Add the procedure to the stack 1698 procStack.addRollbackStep(procedure); 1699 1700 // allows to kill the executor before something is stored to the wal. 1701 // useful to test the procedure recovery. 1702 if ( 1703 testing != null && testing.shouldKillBeforeStoreUpdate(suspended, procedure.hasParent()) 1704 ) { 1705 kill("TESTING: Kill BEFORE store update: " + procedure); 1706 } 1707 1708 // TODO: The code here doesn't check if store is running before persisting to the store as 1709 // it relies on the method call below to throw RuntimeException to wind up the stack and 1710 // executor thread to stop. The statement following the method call below seems to check if 1711 // store is not running, to prevent scheduling children procedures, re-execution or yield 1712 // of this procedure. This may need more scrutiny and subsequent cleanup in future 1713 // 1714 // Commit the transaction even if a suspend (state may have changed). Note this append 1715 // can take a bunch of time to complete. 1716 if (procedure.needPersistence()) { 1717 updateStoreOnExec(procStack, procedure, subprocs); 1718 } 1719 1720 // if the store is not running we are aborting 1721 if (!store.isRunning()) { 1722 return; 1723 } 1724 // if the procedure is kind enough to pass the slot to someone else, yield 1725 if ( 1726 procedure.isRunnable() && !suspended 1727 && procedure.isYieldAfterExecutionStep(getEnvironment()) 1728 ) { 1729 yieldProcedure(procedure); 1730 return; 1731 } 1732 1733 assert (reExecute && subprocs == null) || !reExecute; 1734 } while (reExecute); 1735 1736 // Allows to kill the executor after something is stored to the WAL but before the below 1737 // state settings are done -- in particular the one on the end where we make parent 1738 // RUNNABLE again when its children are done; see countDownChildren. 1739 if (testing != null && testing.shouldKillAfterStoreUpdate(suspended)) { 1740 kill("TESTING: Kill AFTER store update: " + procedure); 1741 } 1742 1743 // Submit the new subprocedures 1744 if (subprocs != null && !procedure.isFailed()) { 1745 submitChildrenProcedures(subprocs); 1746 } 1747 1748 // we need to log the release lock operation before waking up the parent procedure, as there 1749 // could be race that the parent procedure may call updateStoreOnExec ahead of us and remove all 1750 // the sub procedures from store and cause problems... 1751 releaseLock(procedure, false); 1752 1753 // if the procedure is complete and has a parent, count down the children latch. 1754 // If 'suspended', do nothing to change state -- let other threads handle unsuspend event. 1755 if (!suspended && procedure.isFinished() && procedure.hasParent()) { 1756 countDownChildren(procStack, procedure); 1757 } 1758 } 1759 1760 private void kill(String msg) { 1761 LOG.debug(msg); 1762 stop(); 1763 throw new RuntimeException(msg); 1764 } 1765 1766 private Procedure<TEnvironment>[] initializeChildren(RootProcedureState<TEnvironment> procStack, 1767 Procedure<TEnvironment> procedure, Procedure<TEnvironment>[] subprocs) { 1768 assert subprocs != null : "expected subprocedures"; 1769 final long rootProcId = getRootProcedureId(procedure); 1770 for (int i = 0; i < subprocs.length; ++i) { 1771 Procedure<TEnvironment> subproc = subprocs[i]; 1772 if (subproc == null) { 1773 String msg = "subproc[" + i + "] is null, aborting the procedure"; 1774 procedure 1775 .setFailure(new RemoteProcedureException(msg, new IllegalArgumentIOException(msg))); 1776 return null; 1777 } 1778 1779 assert subproc.getState() == ProcedureState.INITIALIZING : subproc; 1780 subproc.setParentProcId(procedure.getProcId()); 1781 subproc.setRootProcId(rootProcId); 1782 subproc.setProcId(nextProcId()); 1783 procStack.addSubProcedure(subproc); 1784 } 1785 1786 if (!procedure.isFailed()) { 1787 procedure.setChildrenLatch(subprocs.length); 1788 switch (procedure.getState()) { 1789 case RUNNABLE: 1790 procedure.setState(ProcedureState.WAITING); 1791 break; 1792 case WAITING_TIMEOUT: 1793 timeoutExecutor.add(procedure); 1794 break; 1795 default: 1796 break; 1797 } 1798 } 1799 return subprocs; 1800 } 1801 1802 private void submitChildrenProcedures(Procedure<TEnvironment>[] subprocs) { 1803 for (int i = 0; i < subprocs.length; ++i) { 1804 Procedure<TEnvironment> subproc = subprocs[i]; 1805 subproc.updateMetricsOnSubmit(getEnvironment()); 1806 assert !procedures.containsKey(subproc.getProcId()); 1807 procedures.put(subproc.getProcId(), subproc); 1808 scheduler.addFront(subproc); 1809 } 1810 } 1811 1812 private void countDownChildren(RootProcedureState<TEnvironment> procStack, 1813 Procedure<TEnvironment> procedure) { 1814 Procedure<TEnvironment> parent = procedures.get(procedure.getParentProcId()); 1815 if (parent == null) { 1816 assert procStack.isRollingback(); 1817 return; 1818 } 1819 1820 // If this procedure is the last child awake the parent procedure 1821 if (parent.tryRunnable()) { 1822 // If we succeeded in making the parent runnable -- i.e. all of its 1823 // children have completed, move parent to front of the queue. 1824 store.update(parent); 1825 scheduler.addFront(parent); 1826 LOG.info("Finished subprocedure pid={}, resume processing ppid={}", procedure.getProcId(), 1827 parent.getProcId()); 1828 return; 1829 } 1830 } 1831 1832 private void updateStoreOnExec(RootProcedureState<TEnvironment> procStack, 1833 Procedure<TEnvironment> procedure, Procedure<TEnvironment>[] subprocs) { 1834 if (subprocs != null && !procedure.isFailed()) { 1835 if (LOG.isTraceEnabled()) { 1836 LOG.trace("Stored " + procedure + ", children " + Arrays.toString(subprocs)); 1837 } 1838 store.insert(procedure, subprocs); 1839 } else { 1840 LOG.trace("Store update {}", procedure); 1841 if (procedure.isFinished() && !procedure.hasParent()) { 1842 // remove child procedures 1843 final long[] childProcIds = procStack.getSubprocedureIds(); 1844 if (childProcIds != null) { 1845 store.delete(procedure, childProcIds); 1846 for (int i = 0; i < childProcIds.length; ++i) { 1847 procedures.remove(childProcIds[i]); 1848 } 1849 } else { 1850 store.update(procedure); 1851 } 1852 } else { 1853 store.update(procedure); 1854 } 1855 } 1856 } 1857 1858 private void handleInterruptedException(Procedure<TEnvironment> proc, InterruptedException e) { 1859 LOG.trace("Interrupt during {}. suspend and retry it later.", proc, e); 1860 // NOTE: We don't call Thread.currentThread().interrupt() 1861 // because otherwise all the subsequent calls e.g. Thread.sleep() will throw 1862 // the InterruptedException. If the master is going down, we will be notified 1863 // and the executor/store will be stopped. 1864 // (The interrupted procedure will be retried on the next run) 1865 } 1866 1867 private void execCompletionCleanup(Procedure<TEnvironment> proc) { 1868 final TEnvironment env = getEnvironment(); 1869 if (proc.hasLock()) { 1870 LOG.warn("Usually this should not happen, we will release the lock before if the procedure" 1871 + " is finished, even if the holdLock is true, arrive here means we have some holes where" 1872 + " we do not release the lock. And the releaseLock below may fail since the procedure may" 1873 + " have already been deleted from the procedure store."); 1874 releaseLock(proc, true); 1875 } 1876 try { 1877 proc.completionCleanup(env); 1878 } catch (Throwable e) { 1879 // Catch NullPointerExceptions or similar errors... 1880 LOG.error("CODE-BUG: uncatched runtime exception for procedure: " + proc, e); 1881 } 1882 } 1883 1884 private void procedureFinished(Procedure<TEnvironment> proc) { 1885 // call the procedure completion cleanup handler 1886 execCompletionCleanup(proc); 1887 1888 CompletedProcedureRetainer<TEnvironment> retainer = new CompletedProcedureRetainer<>(proc); 1889 1890 // update the executor internal state maps 1891 if (!proc.shouldWaitClientAck(getEnvironment())) { 1892 retainer.setClientAckTime(0); 1893 } 1894 1895 completed.put(proc.getProcId(), retainer); 1896 rollbackStack.remove(proc.getProcId()); 1897 procedures.remove(proc.getProcId()); 1898 1899 // call the runnableSet completion cleanup handler 1900 try { 1901 scheduler.completionCleanup(proc); 1902 } catch (Throwable e) { 1903 // Catch NullPointerExceptions or similar errors... 1904 LOG.error("CODE-BUG: uncatched runtime exception for completion cleanup: {}", proc, e); 1905 } 1906 1907 // Notify the listeners 1908 sendProcedureFinishedNotification(proc.getProcId()); 1909 } 1910 1911 RootProcedureState<TEnvironment> getProcStack(long rootProcId) { 1912 return rollbackStack.get(rootProcId); 1913 } 1914 1915 ProcedureScheduler getProcedureScheduler() { 1916 return scheduler; 1917 } 1918 1919 int getCompletedSize() { 1920 return completed.size(); 1921 } 1922 1923 public IdLock getProcExecutionLock() { 1924 return procExecutionLock; 1925 } 1926 1927 // ========================================================================== 1928 // Worker Thread 1929 // ========================================================================== 1930 private class WorkerThread extends StoppableThread { 1931 private final AtomicLong executionStartTime = new AtomicLong(Long.MAX_VALUE); 1932 private volatile Procedure<TEnvironment> activeProcedure; 1933 1934 public WorkerThread(ThreadGroup group) { 1935 this(group, "PEWorker-"); 1936 } 1937 1938 protected WorkerThread(ThreadGroup group, String prefix) { 1939 super(group, prefix + workerId.incrementAndGet()); 1940 setDaemon(true); 1941 } 1942 1943 @Override 1944 public void sendStopSignal() { 1945 scheduler.signalAll(); 1946 } 1947 1948 /** 1949 * Encapsulates execution of the current {@link #activeProcedure} for easy tracing. 1950 */ 1951 private long runProcedure() throws IOException { 1952 final Procedure<TEnvironment> proc = this.activeProcedure; 1953 int activeCount = activeExecutorCount.incrementAndGet(); 1954 int runningCount = store.setRunningProcedureCount(activeCount); 1955 LOG.trace("Execute pid={} runningCount={}, activeCount={}", proc.getProcId(), runningCount, 1956 activeCount); 1957 executionStartTime.set(EnvironmentEdgeManager.currentTime()); 1958 IdLock.Entry lockEntry = procExecutionLock.getLockEntry(proc.getProcId()); 1959 try { 1960 executeProcedure(proc); 1961 } catch (AssertionError e) { 1962 LOG.info("ASSERT pid=" + proc.getProcId(), e); 1963 throw e; 1964 } finally { 1965 procExecutionLock.releaseLockEntry(lockEntry); 1966 activeCount = activeExecutorCount.decrementAndGet(); 1967 runningCount = store.setRunningProcedureCount(activeCount); 1968 LOG.trace("Halt pid={} runningCount={}, activeCount={}", proc.getProcId(), runningCount, 1969 activeCount); 1970 this.activeProcedure = null; 1971 executionStartTime.set(Long.MAX_VALUE); 1972 } 1973 return EnvironmentEdgeManager.currentTime(); 1974 } 1975 1976 @Override 1977 public void run() { 1978 long lastUpdate = EnvironmentEdgeManager.currentTime(); 1979 try { 1980 while (isRunning() && keepAlive(lastUpdate)) { 1981 @SuppressWarnings("unchecked") 1982 Procedure<TEnvironment> proc = scheduler.poll(keepAliveTime, TimeUnit.MILLISECONDS); 1983 if (proc == null) { 1984 continue; 1985 } 1986 this.activeProcedure = proc; 1987 lastUpdate = TraceUtil.trace(this::runProcedure, new ProcedureSpanBuilder(proc)); 1988 } 1989 } catch (Throwable t) { 1990 LOG.warn("Worker terminating UNNATURALLY {}", this.activeProcedure, t); 1991 } finally { 1992 LOG.trace("Worker terminated."); 1993 } 1994 workerThreads.remove(this); 1995 } 1996 1997 @Override 1998 public String toString() { 1999 Procedure<?> p = this.activeProcedure; 2000 return getName() + "(pid=" + (p == null ? Procedure.NO_PROC_ID : p.getProcId() + ")"); 2001 } 2002 2003 /** Returns the time since the current procedure is running */ 2004 public long getCurrentRunTime() { 2005 return EnvironmentEdgeManager.currentTime() - executionStartTime.get(); 2006 } 2007 2008 // core worker never timeout 2009 protected boolean keepAlive(long lastUpdate) { 2010 return true; 2011 } 2012 } 2013 2014 // A worker thread which can be added when core workers are stuck. Will timeout after 2015 // keepAliveTime if there is no procedure to run. 2016 private final class KeepAliveWorkerThread extends WorkerThread { 2017 public KeepAliveWorkerThread(ThreadGroup group) { 2018 super(group, "KeepAlivePEWorker-"); 2019 } 2020 2021 @Override 2022 protected boolean keepAlive(long lastUpdate) { 2023 return EnvironmentEdgeManager.currentTime() - lastUpdate < keepAliveTime; 2024 } 2025 } 2026 2027 // ---------------------------------------------------------------------------- 2028 // TODO-MAYBE: Should we provide a InlineChore to notify the store with the 2029 // full set of procedures pending and completed to write a compacted 2030 // version of the log (in case is a log)? 2031 // In theory no, procedures are have a short life, so at some point the store 2032 // will have the tracker saying everything is in the last log. 2033 // ---------------------------------------------------------------------------- 2034 2035 private final class WorkerMonitor extends InlineChore { 2036 public static final String WORKER_MONITOR_INTERVAL_CONF_KEY = 2037 "hbase.procedure.worker.monitor.interval.msec"; 2038 private static final int DEFAULT_WORKER_MONITOR_INTERVAL = 5000; // 5sec 2039 2040 public static final String WORKER_STUCK_THRESHOLD_CONF_KEY = 2041 "hbase.procedure.worker.stuck.threshold.msec"; 2042 private static final int DEFAULT_WORKER_STUCK_THRESHOLD = 10000; // 10sec 2043 2044 public static final String WORKER_ADD_STUCK_PERCENTAGE_CONF_KEY = 2045 "hbase.procedure.worker.add.stuck.percentage"; 2046 private static final float DEFAULT_WORKER_ADD_STUCK_PERCENTAGE = 0.5f; // 50% stuck 2047 2048 private float addWorkerStuckPercentage = DEFAULT_WORKER_ADD_STUCK_PERCENTAGE; 2049 private int timeoutInterval = DEFAULT_WORKER_MONITOR_INTERVAL; 2050 private int stuckThreshold = DEFAULT_WORKER_STUCK_THRESHOLD; 2051 2052 public WorkerMonitor() { 2053 refreshConfig(); 2054 } 2055 2056 @Override 2057 public void run() { 2058 final int stuckCount = checkForStuckWorkers(); 2059 checkThreadCount(stuckCount); 2060 2061 // refresh interval (poor man dynamic conf update) 2062 refreshConfig(); 2063 } 2064 2065 private int checkForStuckWorkers() { 2066 // check if any of the worker is stuck 2067 int stuckCount = 0; 2068 for (WorkerThread worker : workerThreads) { 2069 if (worker.getCurrentRunTime() < stuckThreshold) { 2070 continue; 2071 } 2072 2073 // WARN the worker is stuck 2074 stuckCount++; 2075 LOG.warn("Worker stuck {}, run time {}", worker, 2076 StringUtils.humanTimeDiff(worker.getCurrentRunTime())); 2077 } 2078 return stuckCount; 2079 } 2080 2081 private void checkThreadCount(final int stuckCount) { 2082 // nothing to do if there are no runnable tasks 2083 if (stuckCount < 1 || !scheduler.hasRunnables()) { 2084 return; 2085 } 2086 2087 // add a new thread if the worker stuck percentage exceed the threshold limit 2088 // and every handler is active. 2089 final float stuckPerc = ((float) stuckCount) / workerThreads.size(); 2090 // let's add new worker thread more aggressively, as they will timeout finally if there is no 2091 // work to do. 2092 if (stuckPerc >= addWorkerStuckPercentage && workerThreads.size() < maxPoolSize) { 2093 final KeepAliveWorkerThread worker = new KeepAliveWorkerThread(threadGroup); 2094 workerThreads.add(worker); 2095 worker.start(); 2096 LOG.debug("Added new worker thread {}", worker); 2097 } 2098 } 2099 2100 private void refreshConfig() { 2101 addWorkerStuckPercentage = 2102 conf.getFloat(WORKER_ADD_STUCK_PERCENTAGE_CONF_KEY, DEFAULT_WORKER_ADD_STUCK_PERCENTAGE); 2103 timeoutInterval = 2104 conf.getInt(WORKER_MONITOR_INTERVAL_CONF_KEY, DEFAULT_WORKER_MONITOR_INTERVAL); 2105 stuckThreshold = conf.getInt(WORKER_STUCK_THRESHOLD_CONF_KEY, DEFAULT_WORKER_STUCK_THRESHOLD); 2106 } 2107 2108 @Override 2109 public int getTimeoutInterval() { 2110 return timeoutInterval; 2111 } 2112 } 2113}