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.Comparator; 028import java.util.Deque; 029import java.util.HashSet; 030import java.util.List; 031import java.util.PriorityQueue; 032import java.util.Set; 033import java.util.concurrent.ConcurrentHashMap; 034import java.util.concurrent.CopyOnWriteArrayList; 035import java.util.concurrent.ExecutorService; 036import java.util.concurrent.Executors; 037import java.util.concurrent.LinkedBlockingQueue; 038import java.util.concurrent.ThreadFactory; 039import java.util.concurrent.ThreadPoolExecutor; 040import java.util.concurrent.TimeUnit; 041import java.util.concurrent.atomic.AtomicBoolean; 042import java.util.concurrent.atomic.AtomicInteger; 043import java.util.concurrent.atomic.AtomicLong; 044import java.util.stream.Collectors; 045import java.util.stream.Stream; 046import org.apache.hadoop.conf.Configuration; 047import org.apache.hadoop.hbase.HConstants; 048import org.apache.hadoop.hbase.exceptions.IllegalArgumentIOException; 049import org.apache.hadoop.hbase.log.HBaseMarkers; 050import org.apache.hadoop.hbase.procedure2.Procedure.LockState; 051import org.apache.hadoop.hbase.procedure2.store.ProcedureStore; 052import org.apache.hadoop.hbase.procedure2.store.ProcedureStore.ProcedureIterator; 053import org.apache.hadoop.hbase.procedure2.store.ProcedureStore.ProcedureStoreListener; 054import org.apache.hadoop.hbase.procedure2.trace.ProcedureSpanBuilder; 055import org.apache.hadoop.hbase.procedure2.util.StringUtils; 056import org.apache.hadoop.hbase.security.User; 057import org.apache.hadoop.hbase.trace.TraceUtil; 058import org.apache.hadoop.hbase.util.EnvironmentEdgeManager; 059import org.apache.hadoop.hbase.util.IdLock; 060import org.apache.hadoop.hbase.util.NonceKey; 061import org.apache.hadoop.hbase.util.Threads; 062import org.apache.yetus.audience.InterfaceAudience; 063import org.slf4j.Logger; 064import org.slf4j.LoggerFactory; 065 066import org.apache.hbase.thirdparty.com.google.common.base.Preconditions; 067import org.apache.hbase.thirdparty.com.google.common.util.concurrent.ThreadFactoryBuilder; 068 069import org.apache.hadoop.hbase.shaded.protobuf.generated.ProcedureProtos.ProcedureState; 070 071/** 072 * Thread Pool that executes the submitted procedures. The executor has a ProcedureStore associated. 073 * Each operation is logged and on restart the pending procedures are resumed. Unless the Procedure 074 * code throws an error (e.g. invalid user input) the procedure will complete (at some point in 075 * time), On restart the pending procedures are resumed and the once failed will be rolledback. The 076 * user can add procedures to the executor via submitProcedure(proc) check for the finished state 077 * via isFinished(procId) and get the result via getResult(procId) 078 */ 079@InterfaceAudience.Private 080public class ProcedureExecutor<TEnvironment> { 081 private static final Logger LOG = LoggerFactory.getLogger(ProcedureExecutor.class); 082 083 public static final String CHECK_OWNER_SET_CONF_KEY = "hbase.procedure.check.owner.set"; 084 private static final boolean DEFAULT_CHECK_OWNER_SET = false; 085 086 public static final String WORKER_KEEP_ALIVE_TIME_CONF_KEY = 087 "hbase.procedure.worker.keep.alive.time.msec"; 088 private static final long DEFAULT_WORKER_KEEP_ALIVE_TIME = TimeUnit.MINUTES.toMillis(1); 089 090 public static final String EVICT_TTL_CONF_KEY = "hbase.procedure.cleaner.evict.ttl"; 091 static final int DEFAULT_EVICT_TTL = 15 * 60000; // 15min 092 093 public static final String EVICT_ACKED_TTL_CONF_KEY = "hbase.procedure.cleaner.acked.evict.ttl"; 094 static final int DEFAULT_ACKED_EVICT_TTL = 5 * 60000; // 5min 095 096 /** 097 * {@link #testing} is non-null when ProcedureExecutor is being tested. Tests will try to break PE 098 * having it fail at various junctures. When non-null, testing is set to an instance of the below 099 * internal {@link Testing} class with flags set for the particular test. 100 */ 101 volatile Testing testing = null; 102 103 /** 104 * Class with parameters describing how to fail/die when in testing-context. 105 */ 106 public static class Testing { 107 protected volatile boolean killIfHasParent = true; 108 protected volatile boolean killIfSuspended = false; 109 110 /** 111 * Kill the PE BEFORE we store state to the WAL. Good for figuring out if a Procedure is 112 * persisting all the state it needs to recover after a crash. 113 */ 114 protected volatile boolean killBeforeStoreUpdate = false; 115 protected volatile boolean toggleKillBeforeStoreUpdate = false; 116 117 /** 118 * Set when we want to fail AFTER state has been stored into the WAL. Rarely used. HBASE-20978 119 * is about a case where memory-state was being set after store to WAL where a crash could cause 120 * us to get stuck. This flag allows killing at what was a vulnerable time. 121 */ 122 protected volatile boolean killAfterStoreUpdate = false; 123 protected volatile boolean toggleKillAfterStoreUpdate = false; 124 125 protected volatile boolean killBeforeStoreUpdateInRollback = false; 126 protected volatile boolean toggleKillBeforeStoreUpdateInRollback = false; 127 128 protected boolean shouldKillBeforeStoreUpdate() { 129 final boolean kill = this.killBeforeStoreUpdate; 130 if (this.toggleKillBeforeStoreUpdate) { 131 this.killBeforeStoreUpdate = !kill; 132 LOG.warn("Toggle KILL before store update to: " + this.killBeforeStoreUpdate); 133 } 134 return kill; 135 } 136 137 protected boolean shouldKillBeforeStoreUpdate(boolean isSuspended, boolean hasParent) { 138 if (isSuspended && !killIfSuspended) { 139 return false; 140 } 141 if (hasParent && !killIfHasParent) { 142 return false; 143 } 144 return shouldKillBeforeStoreUpdate(); 145 } 146 147 protected boolean shouldKillAfterStoreUpdate() { 148 final boolean kill = this.killAfterStoreUpdate; 149 if (this.toggleKillAfterStoreUpdate) { 150 this.killAfterStoreUpdate = !kill; 151 LOG.warn("Toggle KILL after store update to: " + this.killAfterStoreUpdate); 152 } 153 return kill; 154 } 155 156 protected boolean shouldKillAfterStoreUpdate(final boolean isSuspended) { 157 return (isSuspended && !killIfSuspended) ? false : shouldKillAfterStoreUpdate(); 158 } 159 160 protected boolean shouldKillBeforeStoreUpdateInRollback() { 161 final boolean kill = this.killBeforeStoreUpdateInRollback; 162 if (this.toggleKillBeforeStoreUpdateInRollback) { 163 this.killBeforeStoreUpdateInRollback = !kill; 164 LOG.warn("Toggle KILL before store update in rollback to: " 165 + this.killBeforeStoreUpdateInRollback); 166 } 167 return kill; 168 } 169 } 170 171 public interface ProcedureExecutorListener { 172 void procedureLoaded(long procId); 173 174 void procedureAdded(long procId); 175 176 void procedureFinished(long procId); 177 } 178 179 /** 180 * Map the the procId returned by submitProcedure(), the Root-ProcID, to the Procedure. Once a 181 * Root-Procedure completes (success or failure), the result will be added to this map. The user 182 * of ProcedureExecutor should call getResult(procId) to get the result. 183 */ 184 private final ConcurrentHashMap<Long, CompletedProcedureRetainer<TEnvironment>> completed = 185 new ConcurrentHashMap<>(); 186 187 /** 188 * Map the the procId returned by submitProcedure(), the Root-ProcID, to the RootProcedureState. 189 * The RootProcedureState contains the execution stack of the Root-Procedure, It is added to the 190 * map by submitProcedure() and removed on procedure completion. 191 */ 192 private final ConcurrentHashMap<Long, RootProcedureState<TEnvironment>> rollbackStack = 193 new ConcurrentHashMap<>(); 194 195 /** 196 * Helper map to lookup the live procedures by ID. This map contains every procedure. 197 * root-procedures and subprocedures. 198 */ 199 private final ConcurrentHashMap<Long, Procedure<TEnvironment>> procedures = 200 new ConcurrentHashMap<>(); 201 202 /** 203 * Helper map to lookup whether the procedure already issued from the same client. This map 204 * contains every root procedure. 205 */ 206 private final ConcurrentHashMap<NonceKey, Long> nonceKeysToProcIdsMap = new ConcurrentHashMap<>(); 207 208 private final CopyOnWriteArrayList<ProcedureExecutorListener> listeners = 209 new CopyOnWriteArrayList<>(); 210 211 private Configuration conf; 212 213 /** 214 * Created in the {@link #init(int, boolean)} method. Destroyed in {@link #join()} (FIX! Doing 215 * resource handling rather than observing in a #join is unexpected). Overridden when we do the 216 * ProcedureTestingUtility.testRecoveryAndDoubleExecution trickery (Should be ok). 217 */ 218 private ThreadGroup threadGroup; 219 220 /** 221 * Created in the {@link #init(int, boolean)} method. Terminated in {@link #join()} (FIX! Doing 222 * resource handling rather than observing in a #join is unexpected). Overridden when we do the 223 * ProcedureTestingUtility.testRecoveryAndDoubleExecution trickery (Should be ok). 224 */ 225 private CopyOnWriteArrayList<WorkerThread> workerThreads; 226 227 /** 228 * Created in the {@link #init(int, boolean)} method. Terminated in {@link #join()} (FIX! Doing 229 * resource handling rather than observing in a #join is unexpected). Overridden when we do the 230 * ProcedureTestingUtility.testRecoveryAndDoubleExecution trickery (Should be ok). 231 */ 232 private TimeoutExecutorThread<TEnvironment> timeoutExecutor; 233 234 /** 235 * WorkerMonitor check for stuck workers and new worker thread when necessary, for example if 236 * there is no worker to assign meta, it will new worker thread for it, so it is very important. 237 * TimeoutExecutor execute many tasks like DeadServerMetricRegionChore RegionInTransitionChore and 238 * so on, some tasks may execute for a long time so will block other tasks like WorkerMonitor, so 239 * use a dedicated thread for executing WorkerMonitor. 240 */ 241 private TimeoutExecutorThread<TEnvironment> workerMonitorExecutor; 242 243 private ExecutorService forceUpdateExecutor; 244 245 // A thread pool for executing some asynchronous tasks for procedures, you can find references to 246 // getAsyncTaskExecutor to see the usage 247 private ExecutorService asyncTaskExecutor; 248 249 private int corePoolSize; 250 private int maxPoolSize; 251 252 private volatile long keepAliveTime; 253 254 /** 255 * Scheduler/Queue that contains runnable procedures. 256 */ 257 private final ProcedureScheduler scheduler; 258 259 private final AtomicLong lastProcId = new AtomicLong(-1); 260 private final AtomicLong workerId = new AtomicLong(0); 261 private final AtomicInteger activeExecutorCount = new AtomicInteger(0); 262 private final AtomicBoolean running = new AtomicBoolean(false); 263 private final TEnvironment environment; 264 private final ProcedureStore store; 265 266 private final boolean checkOwnerSet; 267 268 // To prevent concurrent execution of the same procedure. 269 // For some rare cases, especially if the procedure uses ProcedureEvent, it is possible that the 270 // procedure is woken up before we finish the suspend which causes the same procedures to be 271 // executed in parallel. This does lead to some problems, see HBASE-20939&HBASE-20949, and is also 272 // a bit confusing to the developers. So here we introduce this lock to prevent the concurrent 273 // execution of the same procedure. 274 private final IdLock procExecutionLock = new IdLock(); 275 276 public ProcedureExecutor(final Configuration conf, final TEnvironment environment, 277 final ProcedureStore store) { 278 this(conf, environment, store, new SimpleProcedureScheduler()); 279 } 280 281 private boolean isRootFinished(Procedure<?> proc) { 282 Procedure<?> rootProc = procedures.get(proc.getRootProcId()); 283 return rootProc == null || rootProc.isFinished(); 284 } 285 286 private void forceUpdateProcedure(long procId) throws IOException { 287 IdLock.Entry lockEntry = procExecutionLock.getLockEntry(procId); 288 try { 289 Procedure<TEnvironment> proc = procedures.get(procId); 290 if (proc != null) { 291 if (proc.isFinished() && proc.hasParent() && isRootFinished(proc)) { 292 LOG.debug("Procedure {} has already been finished and parent is succeeded," 293 + " skip force updating", proc); 294 return; 295 } 296 } else { 297 CompletedProcedureRetainer<TEnvironment> retainer = completed.get(procId); 298 if (retainer == null || retainer.getProcedure() instanceof FailedProcedure) { 299 LOG.debug("No pending procedure with id = {}, skip force updating.", procId); 300 return; 301 } 302 long evictTtl = conf.getInt(EVICT_TTL_CONF_KEY, DEFAULT_EVICT_TTL); 303 long evictAckTtl = conf.getInt(EVICT_ACKED_TTL_CONF_KEY, DEFAULT_ACKED_EVICT_TTL); 304 if (retainer.isExpired(EnvironmentEdgeManager.currentTime(), evictTtl, evictAckTtl)) { 305 LOG.debug("Procedure {} has already been finished and expired, skip force updating", 306 procId); 307 return; 308 } 309 proc = retainer.getProcedure(); 310 } 311 LOG.debug("Force update procedure {}", proc); 312 store.update(proc); 313 } finally { 314 procExecutionLock.releaseLockEntry(lockEntry); 315 } 316 } 317 318 public ProcedureExecutor(final Configuration conf, final TEnvironment environment, 319 final ProcedureStore store, final ProcedureScheduler scheduler) { 320 this.environment = environment; 321 this.scheduler = scheduler; 322 this.store = store; 323 this.conf = conf; 324 this.checkOwnerSet = conf.getBoolean(CHECK_OWNER_SET_CONF_KEY, DEFAULT_CHECK_OWNER_SET); 325 refreshConfiguration(conf); 326 } 327 328 private void load(final boolean abortOnCorruption) throws IOException { 329 Preconditions.checkArgument(completed.isEmpty(), "completed not empty: %s", completed); 330 Preconditions.checkArgument(rollbackStack.isEmpty(), "rollback state not empty: %s", 331 rollbackStack); 332 Preconditions.checkArgument(procedures.isEmpty(), "procedure map not empty: %s", procedures); 333 Preconditions.checkArgument(scheduler.size() == 0, "scheduler queue not empty: %s", scheduler); 334 335 store.load(new ProcedureStore.ProcedureLoader() { 336 @Override 337 public void setMaxProcId(long maxProcId) { 338 assert lastProcId.get() < 0 : "expected only one call to setMaxProcId()"; 339 lastProcId.set(maxProcId); 340 } 341 342 @Override 343 public void load(ProcedureIterator procIter) throws IOException { 344 loadProcedures(procIter); 345 } 346 347 @Override 348 public void handleCorrupted(ProcedureIterator procIter) throws IOException { 349 int corruptedCount = 0; 350 while (procIter.hasNext()) { 351 Procedure<?> proc = procIter.next(); 352 LOG.error("Corrupt " + proc); 353 corruptedCount++; 354 } 355 if (abortOnCorruption && corruptedCount > 0) { 356 throw new IOException("found " + corruptedCount + " corrupted procedure(s) on replay"); 357 } 358 } 359 }); 360 } 361 362 private void restoreLock(Procedure<TEnvironment> proc, Set<Long> restored) { 363 proc.restoreLock(getEnvironment()); 364 restored.add(proc.getProcId()); 365 } 366 367 private void restoreLocks(Deque<Procedure<TEnvironment>> stack, Set<Long> restored) { 368 while (!stack.isEmpty()) { 369 restoreLock(stack.pop(), restored); 370 } 371 } 372 373 // Restore the locks for all the procedures. 374 // Notice that we need to restore the locks starting from the root proc, otherwise there will be 375 // problem that a sub procedure may hold the exclusive lock first and then we are stuck when 376 // calling the acquireLock method for the parent procedure. 377 // The algorithm is straight-forward: 378 // 1. Use a set to record the procedures which locks have already been restored. 379 // 2. Use a stack to store the hierarchy of the procedures 380 // 3. For all the procedure, we will first try to find its parent and push it into the stack, 381 // unless 382 // a. We have no parent, i.e, we are the root procedure 383 // b. The lock has already been restored(by checking the set introduced in #1) 384 // then we start to pop the stack and call acquireLock for each procedure. 385 // Notice that this should be done for all procedures, not only the ones in runnableList. 386 private void restoreLocks() { 387 Set<Long> restored = new HashSet<>(); 388 Deque<Procedure<TEnvironment>> stack = new ArrayDeque<>(); 389 procedures.values().forEach(proc -> { 390 for (;;) { 391 if (restored.contains(proc.getProcId())) { 392 restoreLocks(stack, restored); 393 return; 394 } 395 if (!proc.hasParent()) { 396 restoreLock(proc, restored); 397 restoreLocks(stack, restored); 398 return; 399 } 400 stack.push(proc); 401 proc = procedures.get(proc.getParentProcId()); 402 } 403 }); 404 } 405 406 private void initializeStacks(ProcedureIterator procIter, 407 List<Procedure<TEnvironment>> runnableList, List<Procedure<TEnvironment>> failedList, 408 List<Procedure<TEnvironment>> waitingList, List<Procedure<TEnvironment>> waitingTimeoutList) 409 throws IOException { 410 procIter.reset(); 411 while (procIter.hasNext()) { 412 if (procIter.isNextFinished()) { 413 procIter.skipNext(); 414 continue; 415 } 416 417 @SuppressWarnings("unchecked") 418 Procedure<TEnvironment> proc = procIter.next(); 419 assert !(proc.isFinished() && !proc.hasParent()) : "unexpected completed proc=" + proc; 420 LOG.debug("Loading {}", proc); 421 Long rootProcId = getRootProcedureId(proc); 422 // The orphan procedures will be passed to handleCorrupted, so add an assert here 423 assert rootProcId != null; 424 425 if (proc.hasParent()) { 426 Procedure<TEnvironment> parent = procedures.get(proc.getParentProcId()); 427 if (parent != null && !proc.isFinished()) { 428 parent.incChildrenLatch(); 429 } 430 } 431 432 RootProcedureState<TEnvironment> procStack = rollbackStack.get(rootProcId); 433 procStack.loadStack(proc); 434 435 proc.setRootProcId(rootProcId); 436 switch (proc.getState()) { 437 case RUNNABLE: 438 runnableList.add(proc); 439 break; 440 case WAITING: 441 waitingList.add(proc); 442 break; 443 case WAITING_TIMEOUT: 444 waitingTimeoutList.add(proc); 445 break; 446 case FAILED: 447 failedList.add(proc); 448 break; 449 case ROLLEDBACK: 450 case INITIALIZING: 451 String msg = "Unexpected " + proc.getState() + " state for " + proc; 452 LOG.error(msg); 453 throw new UnsupportedOperationException(msg); 454 default: 455 break; 456 } 457 } 458 rollbackStack.forEach((rootProcId, procStack) -> { 459 if (procStack.getSubproceduresStack() != null) { 460 // if we have already record some stack ids, it means we support rollback 461 procStack.setRollbackSupported(true); 462 } else { 463 // otherwise, test the root procedure to see if we support rollback 464 procStack.setRollbackSupported(procedures.get(rootProcId).isRollbackSupported()); 465 } 466 }); 467 } 468 469 private void processWaitingProcedures(List<Procedure<TEnvironment>> waitingList, 470 List<Procedure<TEnvironment>> runnableList) { 471 waitingList.forEach(proc -> { 472 if (!proc.hasChildren()) { 473 // Normally, WAITING procedures should be waken by its children. But, there is a case that, 474 // all the children are successful and before they can wake up their parent procedure, the 475 // master was killed. So, during recovering the procedures from ProcedureWal, its children 476 // are not loaded because of their SUCCESS state. So we need to continue to run this WAITING 477 // procedure. But before executing, we need to set its state to RUNNABLE, otherwise, a 478 // exception will throw: 479 // Preconditions.checkArgument(procedure.getState() == ProcedureState.RUNNABLE, 480 // "NOT RUNNABLE! " + procedure.toString()); 481 proc.setState(ProcedureState.RUNNABLE); 482 runnableList.add(proc); 483 } else { 484 proc.afterReplay(getEnvironment()); 485 } 486 }); 487 } 488 489 private void processWaitingTimeoutProcedures(List<Procedure<TEnvironment>> waitingTimeoutList) { 490 waitingTimeoutList.forEach(proc -> { 491 proc.afterReplay(getEnvironment()); 492 timeoutExecutor.add(proc); 493 }); 494 } 495 496 private void pushProceduresAfterLoad(List<Procedure<TEnvironment>> runnableList, 497 List<Procedure<TEnvironment>> failedList) { 498 failedList.forEach(scheduler::addBack); 499 runnableList.forEach(p -> { 500 p.afterReplay(getEnvironment()); 501 if (!p.hasParent()) { 502 sendProcedureLoadedNotification(p.getProcId()); 503 } 504 scheduler.addBack(p); 505 }); 506 } 507 508 private void loadProcedures(ProcedureIterator procIter) throws IOException { 509 // 1. Build the rollback stack 510 int runnableCount = 0; 511 int failedCount = 0; 512 int waitingCount = 0; 513 int waitingTimeoutCount = 0; 514 while (procIter.hasNext()) { 515 boolean finished = procIter.isNextFinished(); 516 @SuppressWarnings("unchecked") 517 Procedure<TEnvironment> proc = procIter.next(); 518 NonceKey nonceKey = proc.getNonceKey(); 519 long procId = proc.getProcId(); 520 521 if (finished) { 522 completed.put(proc.getProcId(), new CompletedProcedureRetainer<>(proc)); 523 LOG.debug("Completed {}", proc); 524 } else { 525 if (!proc.hasParent()) { 526 assert !proc.isFinished() : "unexpected finished procedure"; 527 rollbackStack.put(proc.getProcId(), new RootProcedureState<>()); 528 } 529 530 // add the procedure to the map 531 proc.beforeReplay(getEnvironment()); 532 procedures.put(proc.getProcId(), proc); 533 switch (proc.getState()) { 534 case RUNNABLE: 535 runnableCount++; 536 break; 537 case FAILED: 538 failedCount++; 539 break; 540 case WAITING: 541 waitingCount++; 542 break; 543 case WAITING_TIMEOUT: 544 waitingTimeoutCount++; 545 break; 546 default: 547 break; 548 } 549 } 550 551 if (nonceKey != null) { 552 nonceKeysToProcIdsMap.put(nonceKey, procId); // add the nonce to the map 553 } 554 } 555 556 // 2. Initialize the stacks: In the old implementation, for procedures in FAILED state, we will 557 // push it into the ProcedureScheduler directly to execute the rollback. But this does not work 558 // after we introduce the restore lock stage. For now, when we acquire a xlock, we will remove 559 // the queue from runQueue in scheduler, and then when a procedure which has lock access, for 560 // example, a sub procedure of the procedure which has the xlock, is pushed into the scheduler, 561 // we will add the queue back to let the workers poll from it. The assumption here is that, the 562 // procedure which has the xlock should have been polled out already, so when loading we can not 563 // add the procedure to scheduler first and then call acquireLock, since the procedure is still 564 // in the queue, and since we will remove the queue from runQueue, then no one can poll it out, 565 // then there is a dead lock 566 List<Procedure<TEnvironment>> runnableList = new ArrayList<>(runnableCount); 567 List<Procedure<TEnvironment>> failedList = new ArrayList<>(failedCount); 568 List<Procedure<TEnvironment>> waitingList = new ArrayList<>(waitingCount); 569 List<Procedure<TEnvironment>> waitingTimeoutList = new ArrayList<>(waitingTimeoutCount); 570 571 initializeStacks(procIter, runnableList, failedList, waitingList, waitingTimeoutList); 572 573 // 3. Check the waiting procedures to see if some of them can be added to runnable. 574 processWaitingProcedures(waitingList, runnableList); 575 576 // 4. restore locks 577 restoreLocks(); 578 579 // 5. Push the procedures to the timeout executor 580 processWaitingTimeoutProcedures(waitingTimeoutList); 581 582 // 6. Push the procedure to the scheduler 583 pushProceduresAfterLoad(runnableList, failedList); 584 // After all procedures put into the queue, signal the worker threads. 585 // Otherwise, there is a race condition. See HBASE-21364. 586 scheduler.signalAll(); 587 } 588 589 /** 590 * Initialize the procedure executor, but do not start workers. We will start them later. 591 * <p/> 592 * It calls ProcedureStore.recoverLease() and ProcedureStore.load() to recover the lease, and 593 * ensure a single executor, and start the procedure replay to resume and recover the previous 594 * pending and in-progress procedures. 595 * @param numThreads number of threads available for procedure execution. 596 * @param abortOnCorruption true if you want to abort your service in case a corrupted procedure 597 * is found on replay. otherwise false. 598 */ 599 public void init(int numThreads, boolean abortOnCorruption) throws IOException { 600 // We have numThreads executor + one timer thread used for timing out 601 // procedures and triggering periodic procedures. 602 this.corePoolSize = numThreads; 603 this.maxPoolSize = 10 * numThreads; 604 LOG.info("Starting {} core workers (bigger of cpus/4 or 16) with max (burst) worker count={}", 605 corePoolSize, maxPoolSize); 606 607 this.threadGroup = new ThreadGroup("PEWorkerGroup"); 608 this.timeoutExecutor = new TimeoutExecutorThread<>(this, threadGroup, "ProcExecTimeout"); 609 this.workerMonitorExecutor = new TimeoutExecutorThread<>(this, threadGroup, "WorkerMonitor"); 610 ThreadFactory backingThreadFactory = new ThreadFactory() { 611 612 @Override 613 public Thread newThread(Runnable r) { 614 return new Thread(threadGroup, r); 615 } 616 }; 617 int size = Math.max(2, Runtime.getRuntime().availableProcessors()); 618 ThreadPoolExecutor executor = 619 new ThreadPoolExecutor(size, size, 1, TimeUnit.MINUTES, new LinkedBlockingQueue<Runnable>(), 620 new ThreadFactoryBuilder().setDaemon(true) 621 .setNameFormat(getClass().getSimpleName() + "-Async-Task-Executor-%d") 622 .setThreadFactory(backingThreadFactory).build()); 623 executor.allowCoreThreadTimeOut(true); 624 this.asyncTaskExecutor = executor; 625 forceUpdateExecutor = Executors.newFixedThreadPool(1, new ThreadFactoryBuilder().setDaemon(true) 626 .setNameFormat("Force-Update-PEWorker-%d").setThreadFactory(backingThreadFactory).build()); 627 store.registerListener(new ProcedureStoreListener() { 628 629 @Override 630 public void forceUpdate(long[] procIds) { 631 Arrays.stream(procIds).forEach(procId -> forceUpdateExecutor.execute(() -> { 632 try { 633 forceUpdateProcedure(procId); 634 } catch (IOException e) { 635 LOG.warn("Failed to force update procedure with pid={}", procId); 636 } 637 })); 638 } 639 }); 640 641 // Create the workers 642 workerId.set(0); 643 workerThreads = new CopyOnWriteArrayList<>(); 644 for (int i = 0; i < corePoolSize; ++i) { 645 workerThreads.add(new WorkerThread(threadGroup)); 646 } 647 648 long st, et; 649 650 // Acquire the store lease. 651 st = System.nanoTime(); 652 store.recoverLease(); 653 et = System.nanoTime(); 654 LOG.info("Recovered {} lease in {}", store.getClass().getSimpleName(), 655 StringUtils.humanTimeDiff(TimeUnit.NANOSECONDS.toMillis(et - st))); 656 657 // start the procedure scheduler 658 scheduler.start(); 659 660 // TODO: Split in two steps. 661 // TODO: Handle corrupted procedures (currently just a warn) 662 // The first one will make sure that we have the latest id, 663 // so we can start the threads and accept new procedures. 664 // The second step will do the actual load of old procedures. 665 st = System.nanoTime(); 666 load(abortOnCorruption); 667 et = System.nanoTime(); 668 LOG.info("Loaded {} in {}", store.getClass().getSimpleName(), 669 StringUtils.humanTimeDiff(TimeUnit.NANOSECONDS.toMillis(et - st))); 670 } 671 672 /** 673 * Start the workers. 674 */ 675 public void startWorkers() throws IOException { 676 if (!running.compareAndSet(false, true)) { 677 LOG.warn("Already running"); 678 return; 679 } 680 // Start the executors. Here we must have the lastProcId set. 681 LOG.trace("Start workers {}", workerThreads.size()); 682 timeoutExecutor.start(); 683 workerMonitorExecutor.start(); 684 for (WorkerThread worker : workerThreads) { 685 worker.start(); 686 } 687 688 // Internal chores 689 workerMonitorExecutor.add(new WorkerMonitor()); 690 691 // Add completed cleaner chore 692 addChore(new CompletedProcedureCleaner<>(conf, store, procExecutionLock, completed, 693 nonceKeysToProcIdsMap)); 694 } 695 696 public void stop() { 697 // it is possible that we fail in init, while loading procedures, so we will not set running to 698 // true but we should have already started the ProcedureScheduler, and also the two 699 // ExecutorServices, so here we do not check running state, just stop them 700 running.set(false); 701 LOG.info("Stopping"); 702 scheduler.stop(); 703 timeoutExecutor.sendStopSignal(); 704 workerMonitorExecutor.sendStopSignal(); 705 forceUpdateExecutor.shutdown(); 706 asyncTaskExecutor.shutdown(); 707 } 708 709 public void join() { 710 assert !isRunning() : "expected not running"; 711 712 // stop the timeout executor 713 timeoutExecutor.awaitTermination(); 714 // stop the work monitor executor 715 workerMonitorExecutor.awaitTermination(); 716 717 // stop the worker threads 718 for (WorkerThread worker : workerThreads) { 719 worker.awaitTermination(); 720 } 721 try { 722 if (!forceUpdateExecutor.awaitTermination(5, TimeUnit.SECONDS)) { 723 LOG.warn("There are still pending tasks in forceUpdateExecutor"); 724 } 725 } catch (InterruptedException e) { 726 LOG.warn("interrupted while waiting for forceUpdateExecutor termination", e); 727 Thread.currentThread().interrupt(); 728 } 729 try { 730 if (!asyncTaskExecutor.awaitTermination(5, TimeUnit.SECONDS)) { 731 LOG.warn("There are still pending tasks in asyncTaskExecutor"); 732 } 733 } catch (InterruptedException e) { 734 LOG.warn("interrupted while waiting for asyncTaskExecutor termination", e); 735 Thread.currentThread().interrupt(); 736 } 737 738 // log the still active threads, ThreadGroup.destroy is deprecated in JDK17 and it is not 739 // necessary for us to must destroy it here, so we just do a check and log 740 if (threadGroup.activeCount() > 0) { 741 LOG.error("There are still active thread in group {}, see STDOUT", threadGroup); 742 threadGroup.list(); 743 } 744 745 // reset the in-memory state for testing 746 completed.clear(); 747 rollbackStack.clear(); 748 procedures.clear(); 749 nonceKeysToProcIdsMap.clear(); 750 scheduler.clear(); 751 lastProcId.set(-1); 752 } 753 754 public void refreshConfiguration(final Configuration conf) { 755 this.conf = conf; 756 setKeepAliveTime(conf.getLong(WORKER_KEEP_ALIVE_TIME_CONF_KEY, DEFAULT_WORKER_KEEP_ALIVE_TIME), 757 TimeUnit.MILLISECONDS); 758 } 759 760 // ========================================================================== 761 // Accessors 762 // ========================================================================== 763 public boolean isRunning() { 764 return running.get(); 765 } 766 767 /** Returns the current number of worker threads. */ 768 public int getWorkerThreadCount() { 769 return workerThreads.size(); 770 } 771 772 /** Returns the core pool size settings. */ 773 public int getCorePoolSize() { 774 return corePoolSize; 775 } 776 777 public int getActiveExecutorCount() { 778 return activeExecutorCount.get(); 779 } 780 781 public TEnvironment getEnvironment() { 782 return this.environment; 783 } 784 785 public ProcedureStore getStore() { 786 return this.store; 787 } 788 789 ProcedureScheduler getScheduler() { 790 return scheduler; 791 } 792 793 public void setKeepAliveTime(final long keepAliveTime, final TimeUnit timeUnit) { 794 this.keepAliveTime = timeUnit.toMillis(keepAliveTime); 795 this.scheduler.signalAll(); 796 } 797 798 public long getKeepAliveTime(final TimeUnit timeUnit) { 799 return timeUnit.convert(keepAliveTime, TimeUnit.MILLISECONDS); 800 } 801 802 // ========================================================================== 803 // Submit/Remove Chores 804 // ========================================================================== 805 806 /** 807 * Add a chore procedure to the executor 808 * @param chore the chore to add 809 */ 810 public void addChore(@Nullable ProcedureInMemoryChore<TEnvironment> chore) { 811 if (chore == null) { 812 return; 813 } 814 chore.setState(ProcedureState.WAITING_TIMEOUT); 815 timeoutExecutor.add(chore); 816 } 817 818 /** 819 * Remove a chore procedure from the executor 820 * @param chore the chore to remove 821 * @return whether the chore is removed, or it will be removed later 822 */ 823 public boolean removeChore(@Nullable ProcedureInMemoryChore<TEnvironment> chore) { 824 if (chore == null) { 825 return true; 826 } 827 chore.setState(ProcedureState.SUCCESS); 828 return timeoutExecutor.remove(chore); 829 } 830 831 // ========================================================================== 832 // Nonce Procedure helpers 833 // ========================================================================== 834 /** 835 * Create a NonceKey from the specified nonceGroup and nonce. 836 * @param nonceGroup the group to use for the {@link NonceKey} 837 * @param nonce the nonce to use in the {@link NonceKey} 838 * @return the generated NonceKey 839 */ 840 public NonceKey createNonceKey(final long nonceGroup, final long nonce) { 841 return (nonce == HConstants.NO_NONCE) ? null : new NonceKey(nonceGroup, nonce); 842 } 843 844 /** 845 * Register a nonce for a procedure that is going to be submitted. A procId will be reserved and 846 * on submitProcedure(), the procedure with the specified nonce will take the reserved ProcId. If 847 * someone already reserved the nonce, this method will return the procId reserved, otherwise an 848 * invalid procId will be returned. and the caller should procede and submit the procedure. 849 * @param nonceKey A unique identifier for this operation from the client or process. 850 * @return the procId associated with the nonce, if any otherwise an invalid procId. 851 */ 852 public long registerNonce(final NonceKey nonceKey) { 853 if (nonceKey == null) { 854 return -1; 855 } 856 857 // check if we have already a Reserved ID for the nonce 858 Long oldProcId = nonceKeysToProcIdsMap.get(nonceKey); 859 if (oldProcId == null) { 860 // reserve a new Procedure ID, this will be associated with the nonce 861 // and the procedure submitted with the specified nonce will use this ID. 862 final long newProcId = nextProcId(); 863 oldProcId = nonceKeysToProcIdsMap.putIfAbsent(nonceKey, newProcId); 864 if (oldProcId == null) { 865 return -1; 866 } 867 } 868 869 // we found a registered nonce, but the procedure may not have been submitted yet. 870 // since the client expect the procedure to be submitted, spin here until it is. 871 final boolean traceEnabled = LOG.isTraceEnabled(); 872 while ( 873 isRunning() && !(procedures.containsKey(oldProcId) || completed.containsKey(oldProcId)) 874 && nonceKeysToProcIdsMap.containsKey(nonceKey) 875 ) { 876 if (traceEnabled) { 877 LOG.trace("Waiting for pid=" + oldProcId.longValue() + " to be submitted"); 878 } 879 Threads.sleep(100); 880 } 881 return oldProcId.longValue(); 882 } 883 884 /** 885 * Remove the NonceKey if the procedure was not submitted to the executor. 886 * @param nonceKey A unique identifier for this operation from the client or process. 887 */ 888 public void unregisterNonceIfProcedureWasNotSubmitted(final NonceKey nonceKey) { 889 if (nonceKey == null) { 890 return; 891 } 892 893 final Long procId = nonceKeysToProcIdsMap.get(nonceKey); 894 if (procId == null) { 895 return; 896 } 897 898 // if the procedure was not submitted, remove the nonce 899 if (!(procedures.containsKey(procId) || completed.containsKey(procId))) { 900 nonceKeysToProcIdsMap.remove(nonceKey); 901 } 902 } 903 904 /** 905 * If the failure failed before submitting it, we may want to give back the same error to the 906 * requests with the same nonceKey. 907 * @param nonceKey A unique identifier for this operation from the client or process 908 * @param procName name of the procedure, used to inform the user 909 * @param procOwner name of the owner of the procedure, used to inform the user 910 * @param exception the failure to report to the user 911 */ 912 public void setFailureResultForNonce(NonceKey nonceKey, String procName, User procOwner, 913 IOException exception) { 914 if (nonceKey == null) { 915 return; 916 } 917 918 Long procId = nonceKeysToProcIdsMap.get(nonceKey); 919 if (procId == null || completed.containsKey(procId)) { 920 return; 921 } 922 923 completed.computeIfAbsent(procId, (key) -> { 924 Procedure<TEnvironment> proc = 925 new FailedProcedure<>(procId.longValue(), procName, procOwner, nonceKey, exception); 926 927 return new CompletedProcedureRetainer<>(proc); 928 }); 929 } 930 931 // ========================================================================== 932 // Submit/Abort Procedure 933 // ========================================================================== 934 /** 935 * Add a new root-procedure to the executor. 936 * @param proc the new procedure to execute. 937 * @return the procedure id, that can be used to monitor the operation 938 */ 939 public long submitProcedure(Procedure<TEnvironment> proc) { 940 return submitProcedure(proc, null); 941 } 942 943 /** 944 * Bypass a procedure. If the procedure is set to bypass, all the logic in execute/rollback will 945 * be ignored and it will return success, whatever. It is used to recover buggy stuck procedures, 946 * releasing the lock resources and letting other procedures run. Bypassing one procedure (and its 947 * ancestors will be bypassed automatically) may leave the cluster in a middle state, e.g. region 948 * not assigned, or some hdfs files left behind. After getting rid of those stuck procedures, the 949 * operators may have to do some clean up on hdfs or schedule some assign procedures to let region 950 * online. DO AT YOUR OWN RISK. 951 * <p> 952 * A procedure can be bypassed only if 1. The procedure is in state of RUNNABLE, WAITING, 953 * WAITING_TIMEOUT or it is a root procedure without any child. 2. No other worker thread is 954 * executing it 3. No child procedure has been submitted 955 * <p> 956 * If all the requirements are meet, the procedure and its ancestors will be bypassed and 957 * persisted to WAL. 958 * <p> 959 * If the procedure is in WAITING state, will set it to RUNNABLE add it to run queue. TODO: What 960 * about WAITING_TIMEOUT? 961 * @param pids the procedure id 962 * @param lockWait time to wait lock 963 * @param force if force set to true, we will bypass the procedure even if it is executing. 964 * This is for procedures which can't break out during executing(due to bug, 965 * mostly) In this case, bypassing the procedure is not enough, since it is 966 * already stuck there. We need to restart the master after bypassing, and 967 * letting the problematic procedure to execute wth bypass=true, so in that 968 * condition, the procedure can be successfully bypassed. 969 * @param recursive We will do an expensive search for children of each pid. EXPENSIVE! 970 * @return true if bypass success 971 * @throws IOException IOException 972 */ 973 public List<Boolean> bypassProcedure(List<Long> pids, long lockWait, boolean force, 974 boolean recursive) throws IOException { 975 List<Boolean> result = new ArrayList<Boolean>(pids.size()); 976 for (long pid : pids) { 977 result.add(bypassProcedure(pid, lockWait, force, recursive)); 978 } 979 return result; 980 } 981 982 boolean bypassProcedure(long pid, long lockWait, boolean override, boolean recursive) 983 throws IOException { 984 Preconditions.checkArgument(lockWait > 0, "lockWait should be positive"); 985 final Procedure<TEnvironment> procedure = getProcedure(pid); 986 if (procedure == null) { 987 LOG.debug("Procedure pid={} does not exist, skipping bypass", pid); 988 return false; 989 } 990 991 LOG.debug("Begin bypass {} with lockWait={}, override={}, recursive={}", procedure, lockWait, 992 override, recursive); 993 IdLock.Entry lockEntry = procExecutionLock.tryLockEntry(procedure.getProcId(), lockWait); 994 if (lockEntry == null && !override) { 995 LOG.debug("Waited {} ms, but {} is still running, skipping bypass with force={}", lockWait, 996 procedure, override); 997 return false; 998 } else if (lockEntry == null) { 999 LOG.debug("Waited {} ms, but {} is still running, begin bypass with force={}", lockWait, 1000 procedure, override); 1001 } 1002 try { 1003 // check whether the procedure is already finished 1004 if (procedure.isFinished()) { 1005 LOG.debug("{} is already finished, skipping bypass", procedure); 1006 return false; 1007 } 1008 1009 if (procedure.hasChildren()) { 1010 if (recursive) { 1011 // EXPENSIVE. Checks each live procedure of which there could be many!!! 1012 // Is there another way to get children of a procedure? 1013 LOG.info("Recursive bypass on children of pid={}", procedure.getProcId()); 1014 this.procedures.forEachValue(1 /* Single-threaded */, 1015 // Transformer 1016 v -> v.getParentProcId() == procedure.getProcId() ? v : null, 1017 // Consumer 1018 v -> { 1019 try { 1020 bypassProcedure(v.getProcId(), lockWait, override, recursive); 1021 } catch (IOException e) { 1022 LOG.warn("Recursive bypass of pid={}", v.getProcId(), e); 1023 } 1024 }); 1025 } else { 1026 LOG.debug("{} has children, skipping bypass", procedure); 1027 return false; 1028 } 1029 } 1030 1031 // If the procedure has no parent or no child, we are safe to bypass it in whatever state 1032 if ( 1033 procedure.hasParent() && procedure.getState() != ProcedureState.RUNNABLE 1034 && procedure.getState() != ProcedureState.WAITING 1035 && procedure.getState() != ProcedureState.WAITING_TIMEOUT 1036 ) { 1037 LOG.debug("Bypassing procedures in RUNNABLE, WAITING and WAITING_TIMEOUT states " 1038 + "(with no parent), {}", procedure); 1039 // Question: how is the bypass done here? 1040 return false; 1041 } 1042 1043 // Now, the procedure is not finished, and no one can execute it since we take the lock now 1044 // And we can be sure that its ancestor is not running too, since their child has not 1045 // finished yet 1046 Procedure<TEnvironment> current = procedure; 1047 while (current != null) { 1048 LOG.debug("Bypassing {}", current); 1049 current.bypass(getEnvironment()); 1050 store.update(current); 1051 long parentID = current.getParentProcId(); 1052 current = getProcedure(parentID); 1053 } 1054 1055 // wake up waiting procedure, already checked there is no child 1056 if (procedure.getState() == ProcedureState.WAITING) { 1057 procedure.setState(ProcedureState.RUNNABLE); 1058 store.update(procedure); 1059 } 1060 1061 // If state of procedure is WAITING_TIMEOUT, we can directly submit it to the scheduler. 1062 // Instead we should remove it from timeout Executor queue and tranfer its state to RUNNABLE 1063 if (procedure.getState() == ProcedureState.WAITING_TIMEOUT) { 1064 LOG.debug("transform procedure {} from WAITING_TIMEOUT to RUNNABLE", procedure); 1065 if (timeoutExecutor.remove(procedure)) { 1066 LOG.debug("removed procedure {} from timeoutExecutor", procedure); 1067 timeoutExecutor.executeTimedoutProcedure(procedure); 1068 } 1069 } else if (lockEntry != null) { 1070 scheduler.addFront(procedure); 1071 LOG.debug("Bypassing {} and its ancestors successfully, adding to queue", procedure); 1072 } else { 1073 // If we don't have the lock, we can't re-submit the queue, 1074 // since it is already executing. To get rid of the stuck situation, we 1075 // need to restart the master. With the procedure set to bypass, the procedureExecutor 1076 // will bypass it and won't get stuck again. 1077 LOG.debug("Bypassing {} and its ancestors successfully, but since it is already running, " 1078 + "skipping add to queue", procedure); 1079 } 1080 return true; 1081 1082 } finally { 1083 if (lockEntry != null) { 1084 procExecutionLock.releaseLockEntry(lockEntry); 1085 } 1086 } 1087 } 1088 1089 /** 1090 * Add a new root-procedure to the executor. 1091 * @param proc the new procedure to execute. 1092 * @param nonceKey the registered unique identifier for this operation from the client or process. 1093 * @return the procedure id, that can be used to monitor the operation 1094 */ 1095 @edu.umd.cs.findbugs.annotations.SuppressWarnings(value = "NP_NULL_ON_SOME_PATH", 1096 justification = "FindBugs is blind to the check-for-null") 1097 public long submitProcedure(Procedure<TEnvironment> proc, NonceKey nonceKey) { 1098 Preconditions.checkArgument(lastProcId.get() >= 0); 1099 1100 prepareProcedure(proc); 1101 1102 final Long currentProcId; 1103 if (nonceKey != null) { 1104 currentProcId = nonceKeysToProcIdsMap.get(nonceKey); 1105 Preconditions.checkArgument(currentProcId != null, 1106 "Expected nonceKey=" + nonceKey + " to be reserved, use registerNonce(); proc=" + proc); 1107 } else { 1108 currentProcId = nextProcId(); 1109 } 1110 1111 // Initialize the procedure 1112 proc.setNonceKey(nonceKey); 1113 proc.setProcId(currentProcId.longValue()); 1114 1115 // Commit the transaction 1116 store.insert(proc, null); 1117 LOG.debug("Stored {}", proc); 1118 1119 // Add the procedure to the executor 1120 return pushProcedure(proc); 1121 } 1122 1123 /** 1124 * Add a set of new root-procedure to the executor. 1125 * @param procs the new procedures to execute. 1126 */ 1127 // TODO: Do we need to take nonces here? 1128 public void submitProcedures(Procedure<TEnvironment>[] procs) { 1129 Preconditions.checkArgument(lastProcId.get() >= 0); 1130 if (procs == null || procs.length <= 0) { 1131 return; 1132 } 1133 1134 // Prepare procedure 1135 for (int i = 0; i < procs.length; ++i) { 1136 prepareProcedure(procs[i]).setProcId(nextProcId()); 1137 } 1138 1139 // Commit the transaction 1140 store.insert(procs); 1141 if (LOG.isDebugEnabled()) { 1142 LOG.debug("Stored " + Arrays.toString(procs)); 1143 } 1144 1145 // Add the procedure to the executor 1146 for (int i = 0; i < procs.length; ++i) { 1147 pushProcedure(procs[i]); 1148 } 1149 } 1150 1151 private Procedure<TEnvironment> prepareProcedure(Procedure<TEnvironment> proc) { 1152 Preconditions.checkArgument(proc.getState() == ProcedureState.INITIALIZING); 1153 Preconditions.checkArgument(!proc.hasParent(), "unexpected parent", proc); 1154 if (this.checkOwnerSet) { 1155 Preconditions.checkArgument(proc.hasOwner(), "missing owner"); 1156 } 1157 return proc; 1158 } 1159 1160 private long pushProcedure(Procedure<TEnvironment> proc) { 1161 final long currentProcId = proc.getProcId(); 1162 1163 // Update metrics on start of a procedure 1164 proc.updateMetricsOnSubmit(getEnvironment()); 1165 1166 // Create the rollback stack for the procedure 1167 RootProcedureState<TEnvironment> stack = new RootProcedureState<>(); 1168 stack.setRollbackSupported(proc.isRollbackSupported()); 1169 rollbackStack.put(currentProcId, stack); 1170 1171 // Submit the new subprocedures 1172 assert !procedures.containsKey(currentProcId); 1173 procedures.put(currentProcId, proc); 1174 sendProcedureAddedNotification(currentProcId); 1175 scheduler.addBack(proc); 1176 return proc.getProcId(); 1177 } 1178 1179 /** 1180 * Send an abort notification the specified procedure. Depending on the procedure implementation 1181 * the abort can be considered or ignored. 1182 * @param procId the procedure to abort 1183 * @return true if the procedure exists and has received the abort, otherwise false. 1184 */ 1185 public boolean abort(long procId) { 1186 return abort(procId, true); 1187 } 1188 1189 /** 1190 * Send an abort notification to the specified procedure. Depending on the procedure 1191 * implementation, the abort can be considered or ignored. 1192 * @param procId the procedure to abort 1193 * @param mayInterruptIfRunning if the proc completed at least one step, should it be aborted? 1194 * @return true if the procedure exists and has received the abort, otherwise false. 1195 */ 1196 public boolean abort(long procId, boolean mayInterruptIfRunning) { 1197 Procedure<TEnvironment> proc = procedures.get(procId); 1198 if (proc != null) { 1199 if (!mayInterruptIfRunning && proc.wasExecuted()) { 1200 return false; 1201 } 1202 return proc.abort(getEnvironment()); 1203 } 1204 return false; 1205 } 1206 1207 // ========================================================================== 1208 // Executor query helpers 1209 // ========================================================================== 1210 public Procedure<TEnvironment> getProcedure(final long procId) { 1211 return procedures.get(procId); 1212 } 1213 1214 public <T extends Procedure<TEnvironment>> T getProcedure(Class<T> clazz, long procId) { 1215 Procedure<TEnvironment> proc = getProcedure(procId); 1216 if (clazz.isInstance(proc)) { 1217 return clazz.cast(proc); 1218 } 1219 return null; 1220 } 1221 1222 public Procedure<TEnvironment> getResult(long procId) { 1223 CompletedProcedureRetainer<TEnvironment> retainer = completed.get(procId); 1224 if (retainer == null) { 1225 return null; 1226 } else { 1227 return retainer.getProcedure(); 1228 } 1229 } 1230 1231 /** 1232 * Return true if the procedure is finished. The state may be "completed successfully" or "failed 1233 * and rolledback". Use getResult() to check the state or get the result data. 1234 * @param procId the ID of the procedure to check 1235 * @return true if the procedure execution is finished, otherwise false. 1236 */ 1237 public boolean isFinished(final long procId) { 1238 return !procedures.containsKey(procId); 1239 } 1240 1241 /** 1242 * Return true if the procedure is started. 1243 * @param procId the ID of the procedure to check 1244 * @return true if the procedure execution is started, otherwise false. 1245 */ 1246 public boolean isStarted(long procId) { 1247 Procedure<?> proc = procedures.get(procId); 1248 if (proc == null) { 1249 return completed.get(procId) != null; 1250 } 1251 return proc.wasExecuted(); 1252 } 1253 1254 /** 1255 * Mark the specified completed procedure, as ready to remove. 1256 * @param procId the ID of the procedure to remove 1257 */ 1258 public void removeResult(long procId) { 1259 CompletedProcedureRetainer<TEnvironment> retainer = completed.get(procId); 1260 if (retainer == null) { 1261 assert !procedures.containsKey(procId) : "pid=" + procId + " is still running"; 1262 LOG.debug("pid={} already removed by the cleaner.", procId); 1263 return; 1264 } 1265 1266 // The CompletedProcedureCleaner will take care of deletion, once the TTL is expired. 1267 retainer.setClientAckTime(EnvironmentEdgeManager.currentTime()); 1268 } 1269 1270 public Procedure<TEnvironment> getResultOrProcedure(long procId) { 1271 CompletedProcedureRetainer<TEnvironment> retainer = completed.get(procId); 1272 if (retainer == null) { 1273 return procedures.get(procId); 1274 } else { 1275 return retainer.getProcedure(); 1276 } 1277 } 1278 1279 /** 1280 * Check if the user is this procedure's owner 1281 * @param procId the target procedure 1282 * @param user the user 1283 * @return true if the user is the owner of the procedure, false otherwise or the owner is 1284 * unknown. 1285 */ 1286 public boolean isProcedureOwner(long procId, User user) { 1287 if (user == null) { 1288 return false; 1289 } 1290 final Procedure<TEnvironment> runningProc = procedures.get(procId); 1291 if (runningProc != null) { 1292 return runningProc.getOwner().equals(user.getShortName()); 1293 } 1294 1295 final CompletedProcedureRetainer<TEnvironment> retainer = completed.get(procId); 1296 if (retainer != null) { 1297 return retainer.getProcedure().getOwner().equals(user.getShortName()); 1298 } 1299 1300 // Procedure either does not exist or has already completed and got cleaned up. 1301 // At this time, we cannot check the owner of the procedure 1302 return false; 1303 } 1304 1305 /** 1306 * Should only be used when starting up, where the procedure workers have not been started. 1307 * <p/> 1308 * If the procedure works has been started, the return values maybe changed when you are 1309 * processing it so usually this is not safe. Use {@link #getProcedures()} below for most cases as 1310 * it will do a copy, and also include the finished procedures. 1311 */ 1312 public Collection<Procedure<TEnvironment>> getActiveProceduresNoCopy() { 1313 return procedures.values(); 1314 } 1315 1316 /** 1317 * Get procedures. 1318 * @return the procedures in a list 1319 */ 1320 public List<Procedure<TEnvironment>> getProcedures() { 1321 List<Procedure<TEnvironment>> procedureList = 1322 new ArrayList<>(procedures.size() + completed.size()); 1323 procedureList.addAll(procedures.values()); 1324 // Note: The procedure could show up twice in the list with different state, as 1325 // it could complete after we walk through procedures list and insert into 1326 // procedureList - it is ok, as we will use the information in the Procedure 1327 // to figure it out; to prevent this would increase the complexity of the logic. 1328 completed.values().stream().map(CompletedProcedureRetainer::getProcedure) 1329 .forEach(procedureList::add); 1330 return procedureList; 1331 } 1332 1333 // ========================================================================== 1334 // Listeners helpers 1335 // ========================================================================== 1336 public void registerListener(ProcedureExecutorListener listener) { 1337 this.listeners.add(listener); 1338 } 1339 1340 public boolean unregisterListener(ProcedureExecutorListener listener) { 1341 return this.listeners.remove(listener); 1342 } 1343 1344 private void sendProcedureLoadedNotification(final long procId) { 1345 if (!this.listeners.isEmpty()) { 1346 for (ProcedureExecutorListener listener : this.listeners) { 1347 try { 1348 listener.procedureLoaded(procId); 1349 } catch (Throwable e) { 1350 LOG.error("Listener " + listener + " had an error: " + e.getMessage(), e); 1351 } 1352 } 1353 } 1354 } 1355 1356 private void sendProcedureAddedNotification(final long procId) { 1357 if (!this.listeners.isEmpty()) { 1358 for (ProcedureExecutorListener listener : this.listeners) { 1359 try { 1360 listener.procedureAdded(procId); 1361 } catch (Throwable e) { 1362 LOG.error("Listener " + listener + " had an error: " + e.getMessage(), e); 1363 } 1364 } 1365 } 1366 } 1367 1368 private void sendProcedureFinishedNotification(final long procId) { 1369 if (!this.listeners.isEmpty()) { 1370 for (ProcedureExecutorListener listener : this.listeners) { 1371 try { 1372 listener.procedureFinished(procId); 1373 } catch (Throwable e) { 1374 LOG.error("Listener " + listener + " had an error: " + e.getMessage(), e); 1375 } 1376 } 1377 } 1378 } 1379 1380 // ========================================================================== 1381 // Procedure IDs helpers 1382 // ========================================================================== 1383 private long nextProcId() { 1384 long procId = lastProcId.incrementAndGet(); 1385 if (procId < 0) { 1386 while (!lastProcId.compareAndSet(procId, 0)) { 1387 procId = lastProcId.get(); 1388 if (procId >= 0) { 1389 break; 1390 } 1391 } 1392 while (procedures.containsKey(procId)) { 1393 procId = lastProcId.incrementAndGet(); 1394 } 1395 } 1396 assert procId >= 0 : "Invalid procId " + procId; 1397 return procId; 1398 } 1399 1400 protected long getLastProcId() { 1401 return lastProcId.get(); 1402 } 1403 1404 public Set<Long> getActiveProcIds() { 1405 return procedures.keySet(); 1406 } 1407 1408 Long getRootProcedureId(Procedure<TEnvironment> proc) { 1409 return Procedure.getRootProcedureId(procedures, proc); 1410 } 1411 1412 // ========================================================================== 1413 // Executions 1414 // ========================================================================== 1415 private void executeProcedure(Procedure<TEnvironment> proc) { 1416 if (proc.isFinished()) { 1417 LOG.debug("{} is already finished, skipping execution", proc); 1418 return; 1419 } 1420 final Long rootProcId = getRootProcedureId(proc); 1421 if (rootProcId == null) { 1422 // The 'proc' was ready to run but the root procedure was rolledback 1423 LOG.warn("Rollback because parent is done/rolledback proc=" + proc); 1424 executeRollback(proc); 1425 return; 1426 } 1427 1428 RootProcedureState<TEnvironment> procStack = rollbackStack.get(rootProcId); 1429 if (procStack == null) { 1430 LOG.warn("RootProcedureState is null for " + proc.getProcId()); 1431 return; 1432 } 1433 do { 1434 // Try to acquire the execution 1435 if (!procStack.acquire(proc)) { 1436 if (procStack.setRollback()) { 1437 // we have the 'rollback-lock' we can start rollingback 1438 switch (executeRollback(rootProcId, procStack)) { 1439 case LOCK_ACQUIRED: 1440 break; 1441 case LOCK_YIELD_WAIT: 1442 procStack.unsetRollback(); 1443 scheduler.yield(proc); 1444 break; 1445 case LOCK_EVENT_WAIT: 1446 LOG.info("LOCK_EVENT_WAIT rollback..." + proc); 1447 procStack.unsetRollback(); 1448 break; 1449 default: 1450 throw new UnsupportedOperationException(); 1451 } 1452 } else { 1453 // if we can't rollback means that some child is still running. 1454 // the rollback will be executed after all the children are done. 1455 // If the procedure was never executed, remove and mark it as rolledback. 1456 if (!proc.wasExecuted()) { 1457 switch (executeRollback(proc)) { 1458 case LOCK_ACQUIRED: 1459 break; 1460 case LOCK_YIELD_WAIT: 1461 scheduler.yield(proc); 1462 break; 1463 case LOCK_EVENT_WAIT: 1464 LOG.info("LOCK_EVENT_WAIT can't rollback child running?..." + proc); 1465 break; 1466 default: 1467 throw new UnsupportedOperationException(); 1468 } 1469 } 1470 } 1471 break; 1472 } 1473 1474 // Execute the procedure 1475 assert proc.getState() == ProcedureState.RUNNABLE : proc; 1476 // Note that lock is NOT about concurrency but rather about ensuring 1477 // ownership of a procedure of an entity such as a region or table 1478 LockState lockState = acquireLock(proc); 1479 switch (lockState) { 1480 case LOCK_ACQUIRED: 1481 execProcedure(procStack, proc); 1482 break; 1483 case LOCK_YIELD_WAIT: 1484 LOG.info(lockState + " " + proc); 1485 scheduler.yield(proc); 1486 break; 1487 case LOCK_EVENT_WAIT: 1488 // Someone will wake us up when the lock is available 1489 LOG.debug(lockState + " " + proc); 1490 break; 1491 default: 1492 throw new UnsupportedOperationException(); 1493 } 1494 procStack.release(proc); 1495 1496 if (proc.isSuccess()) { 1497 // update metrics on finishing the procedure 1498 proc.updateMetricsOnFinish(getEnvironment(), proc.elapsedTime(), true); 1499 LOG.info("Finished " + proc + " in " + StringUtils.humanTimeDiff(proc.elapsedTime())); 1500 // Finalize the procedure state 1501 if (proc.getProcId() == rootProcId) { 1502 procedureFinished(proc); 1503 } else { 1504 execCompletionCleanup(proc); 1505 } 1506 break; 1507 } 1508 } while (procStack.isFailed()); 1509 } 1510 1511 private LockState acquireLock(Procedure<TEnvironment> proc) { 1512 TEnvironment env = getEnvironment(); 1513 // if holdLock is true, then maybe we already have the lock, so just return LOCK_ACQUIRED if 1514 // hasLock is true. 1515 if (proc.hasLock()) { 1516 return LockState.LOCK_ACQUIRED; 1517 } 1518 return proc.doAcquireLock(env, store); 1519 } 1520 1521 private void releaseLock(Procedure<TEnvironment> proc, boolean force) { 1522 TEnvironment env = getEnvironment(); 1523 // For how the framework works, we know that we will always have the lock 1524 // when we call releaseLock(), so we can avoid calling proc.hasLock() 1525 if (force || !proc.holdLock(env) || proc.isFinished()) { 1526 proc.doReleaseLock(env, store); 1527 } 1528 } 1529 1530 // Returning null means we have already held the execution lock, so you do not need to get the 1531 // lock entry for releasing 1532 private IdLock.Entry getLockEntryForRollback(long procId) { 1533 // Hold the execution lock if it is not held by us. The IdLock is not reentrant so we need 1534 // this check, as the worker will hold the lock before executing a procedure. This is the only 1535 // place where we may hold two procedure execution locks, and there is a fence in the 1536 // RootProcedureState where we can make sure that only one worker can execute the rollback of 1537 // a RootProcedureState, so there is no dead lock problem. And the lock here is necessary to 1538 // prevent race between us and the force update thread. 1539 if (!procExecutionLock.isHeldByCurrentThread(procId)) { 1540 try { 1541 return procExecutionLock.getLockEntry(procId); 1542 } catch (IOException e) { 1543 // can only happen if interrupted, so not a big deal to propagate it 1544 throw new UncheckedIOException(e); 1545 } 1546 } 1547 return null; 1548 } 1549 1550 private void executeUnexpectedRollback(Procedure<TEnvironment> rootProc, 1551 RootProcedureState<TEnvironment> procStack) { 1552 if (procStack.getSubprocs() != null) { 1553 // comparing proc id in reverse order, so we will delete later procedures first, otherwise we 1554 // may delete parent procedure first and if we fail in the middle of this operation, when 1555 // loading we will find some orphan procedures 1556 PriorityQueue<Procedure<TEnvironment>> pq = 1557 new PriorityQueue<>(procStack.getSubprocs().size(), 1558 Comparator.<Procedure<TEnvironment>> comparingLong(Procedure::getProcId).reversed()); 1559 pq.addAll(procStack.getSubprocs()); 1560 for (;;) { 1561 Procedure<TEnvironment> subproc = pq.poll(); 1562 if (subproc == null) { 1563 break; 1564 } 1565 if (!procedures.containsKey(subproc.getProcId())) { 1566 // this means it has already been rolledback 1567 continue; 1568 } 1569 IdLock.Entry lockEntry = getLockEntryForRollback(subproc.getProcId()); 1570 try { 1571 cleanupAfterRollbackOneStep(subproc); 1572 execCompletionCleanup(subproc); 1573 } finally { 1574 if (lockEntry != null) { 1575 procExecutionLock.releaseLockEntry(lockEntry); 1576 } 1577 } 1578 } 1579 } 1580 IdLock.Entry lockEntry = getLockEntryForRollback(rootProc.getProcId()); 1581 try { 1582 cleanupAfterRollbackOneStep(rootProc); 1583 } finally { 1584 if (lockEntry != null) { 1585 procExecutionLock.releaseLockEntry(lockEntry); 1586 } 1587 } 1588 } 1589 1590 private LockState executeNormalRollback(Procedure<TEnvironment> rootProc, 1591 RootProcedureState<TEnvironment> procStack) { 1592 List<Procedure<TEnvironment>> subprocStack = procStack.getSubproceduresStack(); 1593 assert subprocStack != null : "Called rollback with no steps executed rootProc=" + rootProc; 1594 1595 int stackTail = subprocStack.size(); 1596 while (stackTail-- > 0) { 1597 Procedure<TEnvironment> proc = subprocStack.get(stackTail); 1598 IdLock.Entry lockEntry = getLockEntryForRollback(proc.getProcId()); 1599 try { 1600 // For the sub procedures which are successfully finished, we do not rollback them. 1601 // Typically, if we want to rollback a procedure, we first need to rollback it, and then 1602 // recursively rollback its ancestors. The state changes which are done by sub procedures 1603 // should be handled by parent procedures when rolling back. For example, when rolling back 1604 // a MergeTableProcedure, we will schedule new procedures to bring the offline regions 1605 // online, instead of rolling back the original procedures which offlined the regions(in 1606 // fact these procedures can not be rolled back...). 1607 if (proc.isSuccess()) { 1608 // Just do the cleanup work, without actually executing the rollback 1609 subprocStack.remove(stackTail); 1610 cleanupAfterRollbackOneStep(proc); 1611 continue; 1612 } 1613 LockState lockState = acquireLock(proc); 1614 if (lockState != LockState.LOCK_ACQUIRED) { 1615 // can't take a lock on the procedure, add the root-proc back on the 1616 // queue waiting for the lock availability 1617 return lockState; 1618 } 1619 1620 lockState = executeRollback(proc); 1621 releaseLock(proc, false); 1622 boolean abortRollback = lockState != LockState.LOCK_ACQUIRED; 1623 abortRollback |= !isRunning() || !store.isRunning(); 1624 1625 // allows to kill the executor before something is stored to the wal. 1626 // useful to test the procedure recovery. 1627 if (abortRollback) { 1628 return lockState; 1629 } 1630 1631 subprocStack.remove(stackTail); 1632 1633 // if the procedure is kind enough to pass the slot to someone else, yield 1634 // if the proc is already finished, do not yield 1635 if (!proc.isFinished() && proc.isYieldAfterExecutionStep(getEnvironment())) { 1636 return LockState.LOCK_YIELD_WAIT; 1637 } 1638 1639 if (proc != rootProc) { 1640 execCompletionCleanup(proc); 1641 } 1642 } finally { 1643 if (lockEntry != null) { 1644 procExecutionLock.releaseLockEntry(lockEntry); 1645 } 1646 } 1647 } 1648 return LockState.LOCK_ACQUIRED; 1649 } 1650 1651 /** 1652 * Execute the rollback of the full procedure stack. Once the procedure is rolledback, the 1653 * root-procedure will be visible as finished to user, and the result will be the fatal exception. 1654 */ 1655 private LockState executeRollback(long rootProcId, RootProcedureState<TEnvironment> procStack) { 1656 Procedure<TEnvironment> rootProc = procedures.get(rootProcId); 1657 RemoteProcedureException exception = rootProc.getException(); 1658 // TODO: This needs doc. The root proc doesn't have an exception. Maybe we are 1659 // rolling back because the subprocedure does. Clarify. 1660 if (exception == null) { 1661 exception = procStack.getException(); 1662 rootProc.setFailure(exception); 1663 store.update(rootProc); 1664 } 1665 1666 if (procStack.isRollbackSupported()) { 1667 LockState lockState = executeNormalRollback(rootProc, procStack); 1668 if (lockState != LockState.LOCK_ACQUIRED) { 1669 return lockState; 1670 } 1671 } else { 1672 // the procedure does not support rollback, so typically we should not reach here, this 1673 // usually means there are code bugs, let's just wait all the subprocedures to finish and then 1674 // mark the root procedure as failure. 1675 LOG.error(HBaseMarkers.FATAL, 1676 "Root Procedure {} does not support rollback but the execution failed" 1677 + " and try to rollback, code bug?", 1678 rootProc, exception); 1679 executeUnexpectedRollback(rootProc, procStack); 1680 } 1681 1682 IdLock.Entry lockEntry = getLockEntryForRollback(rootProc.getProcId()); 1683 try { 1684 // Finalize the procedure state 1685 LOG.info("Rolled back {} exec-time={}", rootProc, 1686 StringUtils.humanTimeDiff(rootProc.elapsedTime())); 1687 procedureFinished(rootProc); 1688 } finally { 1689 if (lockEntry != null) { 1690 procExecutionLock.releaseLockEntry(lockEntry); 1691 } 1692 } 1693 1694 return LockState.LOCK_ACQUIRED; 1695 } 1696 1697 private void cleanupAfterRollbackOneStep(Procedure<TEnvironment> proc) { 1698 if (testing != null && testing.shouldKillBeforeStoreUpdateInRollback()) { 1699 kill("TESTING: Kill BEFORE store update in rollback: " + proc); 1700 } 1701 if (proc.removeStackIndex()) { 1702 if (!proc.isSuccess()) { 1703 proc.setState(ProcedureState.ROLLEDBACK); 1704 } 1705 1706 // update metrics on finishing the procedure (fail) 1707 proc.updateMetricsOnFinish(getEnvironment(), proc.elapsedTime(), false); 1708 1709 if (proc.hasParent()) { 1710 store.delete(proc.getProcId()); 1711 procedures.remove(proc.getProcId()); 1712 } else { 1713 final long[] childProcIds = rollbackStack.get(proc.getProcId()).getSubprocedureIds(); 1714 if (childProcIds != null) { 1715 store.delete(proc, childProcIds); 1716 } else { 1717 store.update(proc); 1718 } 1719 } 1720 } else { 1721 store.update(proc); 1722 } 1723 } 1724 1725 /** 1726 * Execute the rollback of the procedure step. It updates the store with the new state (stack 1727 * index) or will remove completly the procedure in case it is a child. 1728 */ 1729 private LockState executeRollback(Procedure<TEnvironment> proc) { 1730 try { 1731 proc.doRollback(getEnvironment()); 1732 } catch (IOException e) { 1733 LOG.debug("Roll back attempt failed for {}", proc, e); 1734 return LockState.LOCK_YIELD_WAIT; 1735 } catch (InterruptedException e) { 1736 handleInterruptedException(proc, e); 1737 return LockState.LOCK_YIELD_WAIT; 1738 } catch (Throwable e) { 1739 // Catch NullPointerExceptions or similar errors... 1740 LOG.error(HBaseMarkers.FATAL, "CODE-BUG: Uncaught runtime exception for " + proc, e); 1741 } 1742 1743 cleanupAfterRollbackOneStep(proc); 1744 1745 return LockState.LOCK_ACQUIRED; 1746 } 1747 1748 private void yieldProcedure(Procedure<TEnvironment> proc) { 1749 releaseLock(proc, false); 1750 scheduler.yield(proc); 1751 } 1752 1753 /** 1754 * Executes <code>procedure</code> 1755 * <ul> 1756 * <li>Calls the doExecute() of the procedure 1757 * <li>If the procedure execution didn't fail (i.e. valid user input) 1758 * <ul> 1759 * <li>...and returned subprocedures 1760 * <ul> 1761 * <li>The subprocedures are initialized. 1762 * <li>The subprocedures are added to the store 1763 * <li>The subprocedures are added to the runnable queue 1764 * <li>The procedure is now in a WAITING state, waiting for the subprocedures to complete 1765 * </ul> 1766 * </li> 1767 * <li>...if there are no subprocedure 1768 * <ul> 1769 * <li>the procedure completed successfully 1770 * <li>if there is a parent (WAITING) 1771 * <li>the parent state will be set to RUNNABLE 1772 * </ul> 1773 * </li> 1774 * </ul> 1775 * </li> 1776 * <li>In case of failure 1777 * <ul> 1778 * <li>The store is updated with the new state</li> 1779 * <li>The executor (caller of this method) will start the rollback of the procedure</li> 1780 * </ul> 1781 * </li> 1782 * </ul> 1783 */ 1784 private void execProcedure(RootProcedureState<TEnvironment> procStack, 1785 Procedure<TEnvironment> procedure) { 1786 Preconditions.checkArgument(procedure.getState() == ProcedureState.RUNNABLE, 1787 "NOT RUNNABLE! " + procedure.toString()); 1788 1789 // Procedures can suspend themselves. They skip out by throwing a ProcedureSuspendedException. 1790 // The exception is caught below and then we hurry to the exit without disturbing state. The 1791 // idea is that the processing of this procedure will be unsuspended later by an external event 1792 // such the report of a region open. 1793 boolean suspended = false; 1794 1795 // Whether to 're-' -execute; run through the loop again. 1796 boolean reExecute = false; 1797 1798 Procedure<TEnvironment>[] subprocs = null; 1799 do { 1800 reExecute = false; 1801 procedure.resetPersistence(); 1802 try { 1803 subprocs = procedure.doExecute(getEnvironment()); 1804 if (subprocs != null && subprocs.length == 0) { 1805 subprocs = null; 1806 } 1807 } catch (ProcedureSuspendedException e) { 1808 LOG.trace("Suspend {}", procedure); 1809 suspended = true; 1810 } catch (ProcedureYieldException e) { 1811 LOG.trace("Yield {}", procedure, e); 1812 yieldProcedure(procedure); 1813 return; 1814 } catch (InterruptedException e) { 1815 LOG.trace("Yield interrupt {}", procedure, e); 1816 handleInterruptedException(procedure, e); 1817 yieldProcedure(procedure); 1818 return; 1819 } catch (Throwable e) { 1820 // Catch NullPointerExceptions or similar errors... 1821 String msg = "CODE-BUG: Uncaught runtime exception: " + procedure; 1822 LOG.error(msg, e); 1823 procedure.setFailure(new RemoteProcedureException(msg, e)); 1824 } 1825 1826 if (!procedure.isFailed()) { 1827 if (subprocs != null) { 1828 if (subprocs.length == 1 && subprocs[0] == procedure) { 1829 // Procedure returned itself. Quick-shortcut for a state machine-like procedure; 1830 // i.e. we go around this loop again rather than go back out on the scheduler queue. 1831 subprocs = null; 1832 reExecute = true; 1833 LOG.trace("Short-circuit to next step on pid={}", procedure.getProcId()); 1834 } else { 1835 // Yield the current procedure, and make the subprocedure runnable 1836 // subprocs may come back 'null'. 1837 subprocs = initializeChildren(procStack, procedure, subprocs); 1838 LOG.info("Initialized subprocedures=" + (subprocs == null 1839 ? null 1840 : Stream.of(subprocs).map(e -> "{" + e.toString() + "}").collect(Collectors.toList()) 1841 .toString())); 1842 } 1843 } else if (procedure.getState() == ProcedureState.WAITING_TIMEOUT) { 1844 LOG.trace("Added to timeoutExecutor {}", procedure); 1845 timeoutExecutor.add(procedure); 1846 } else if (!suspended) { 1847 // No subtask, so we are done 1848 procedure.setState(ProcedureState.SUCCESS); 1849 } 1850 } 1851 1852 // allows to kill the executor before something is stored to the wal. 1853 // useful to test the procedure recovery. 1854 if ( 1855 testing != null && testing.shouldKillBeforeStoreUpdate(suspended, procedure.hasParent()) 1856 ) { 1857 kill("TESTING: Kill BEFORE store update: " + procedure); 1858 } 1859 1860 // TODO: The code here doesn't check if store is running before persisting to the store as 1861 // it relies on the method call below to throw RuntimeException to wind up the stack and 1862 // executor thread to stop. The statement following the method call below seems to check if 1863 // store is not running, to prevent scheduling children procedures, re-execution or yield 1864 // of this procedure. This may need more scrutiny and subsequent cleanup in future 1865 // 1866 // Commit the transaction even if a suspend (state may have changed). Note this append 1867 // can take a bunch of time to complete. 1868 if (procedure.needPersistence()) { 1869 // Add the procedure to the stack 1870 // See HBASE-28210 on why we need synchronized here 1871 boolean needUpdateStoreOutsideLock = false; 1872 synchronized (procStack) { 1873 if (procStack.addRollbackStep(procedure)) { 1874 updateStoreOnExec(procStack, procedure, subprocs); 1875 } else { 1876 needUpdateStoreOutsideLock = true; 1877 } 1878 } 1879 // this is an optimization if we do not need to maintain rollback step, as all subprocedures 1880 // of the same root procedure share the same root procedure state, if we can only update 1881 // store under the above lock, the sub procedures of the same root procedure can only be 1882 // persistent sequentially, which will have a bad performance. See HBASE-28212 for more 1883 // details. 1884 if (needUpdateStoreOutsideLock) { 1885 updateStoreOnExec(procStack, procedure, subprocs); 1886 } 1887 } 1888 1889 // if the store is not running we are aborting 1890 if (!store.isRunning()) { 1891 return; 1892 } 1893 // if the procedure is kind enough to pass the slot to someone else, yield 1894 if ( 1895 procedure.isRunnable() && !suspended 1896 && procedure.isYieldAfterExecutionStep(getEnvironment()) 1897 ) { 1898 yieldProcedure(procedure); 1899 return; 1900 } 1901 1902 assert (reExecute && subprocs == null) || !reExecute; 1903 } while (reExecute); 1904 1905 // Allows to kill the executor after something is stored to the WAL but before the below 1906 // state settings are done -- in particular the one on the end where we make parent 1907 // RUNNABLE again when its children are done; see countDownChildren. 1908 if (testing != null && testing.shouldKillAfterStoreUpdate(suspended)) { 1909 kill("TESTING: Kill AFTER store update: " + procedure); 1910 } 1911 1912 // Submit the new subprocedures 1913 if (subprocs != null && !procedure.isFailed()) { 1914 submitChildrenProcedures(subprocs); 1915 } 1916 1917 // we need to log the release lock operation before waking up the parent procedure, as there 1918 // could be race that the parent procedure may call updateStoreOnExec ahead of us and remove all 1919 // the sub procedures from store and cause problems... 1920 releaseLock(procedure, false); 1921 1922 // if the procedure is complete and has a parent, count down the children latch. 1923 // If 'suspended', do nothing to change state -- let other threads handle unsuspend event. 1924 if (!suspended && procedure.isFinished() && procedure.hasParent()) { 1925 countDownChildren(procStack, procedure); 1926 } 1927 } 1928 1929 private void kill(String msg) { 1930 LOG.debug(msg); 1931 stop(); 1932 throw new RuntimeException(msg); 1933 } 1934 1935 private Procedure<TEnvironment>[] initializeChildren(RootProcedureState<TEnvironment> procStack, 1936 Procedure<TEnvironment> procedure, Procedure<TEnvironment>[] subprocs) { 1937 assert subprocs != null : "expected subprocedures"; 1938 final long rootProcId = getRootProcedureId(procedure); 1939 for (int i = 0; i < subprocs.length; ++i) { 1940 Procedure<TEnvironment> subproc = subprocs[i]; 1941 if (subproc == null) { 1942 String msg = "subproc[" + i + "] is null, aborting the procedure"; 1943 procedure 1944 .setFailure(new RemoteProcedureException(msg, new IllegalArgumentIOException(msg))); 1945 return null; 1946 } 1947 1948 assert subproc.getState() == ProcedureState.INITIALIZING : subproc; 1949 subproc.setParentProcId(procedure.getProcId()); 1950 subproc.setRootProcId(rootProcId); 1951 subproc.setProcId(nextProcId()); 1952 procStack.addSubProcedure(subproc); 1953 } 1954 1955 if (!procedure.isFailed()) { 1956 procedure.setChildrenLatch(subprocs.length); 1957 switch (procedure.getState()) { 1958 case RUNNABLE: 1959 procedure.setState(ProcedureState.WAITING); 1960 break; 1961 case WAITING_TIMEOUT: 1962 timeoutExecutor.add(procedure); 1963 break; 1964 default: 1965 break; 1966 } 1967 } 1968 return subprocs; 1969 } 1970 1971 private void submitChildrenProcedures(Procedure<TEnvironment>[] subprocs) { 1972 for (int i = 0; i < subprocs.length; ++i) { 1973 Procedure<TEnvironment> subproc = subprocs[i]; 1974 subproc.updateMetricsOnSubmit(getEnvironment()); 1975 assert !procedures.containsKey(subproc.getProcId()); 1976 procedures.put(subproc.getProcId(), subproc); 1977 scheduler.addFront(subproc); 1978 } 1979 } 1980 1981 private void countDownChildren(RootProcedureState<TEnvironment> procStack, 1982 Procedure<TEnvironment> procedure) { 1983 Procedure<TEnvironment> parent = procedures.get(procedure.getParentProcId()); 1984 if (parent == null) { 1985 assert procStack.isRollingback(); 1986 return; 1987 } 1988 1989 // If this procedure is the last child awake the parent procedure 1990 if (parent.tryRunnable()) { 1991 // If we succeeded in making the parent runnable -- i.e. all of its 1992 // children have completed, move parent to front of the queue. 1993 store.update(parent); 1994 scheduler.addFront(parent); 1995 LOG.info("Finished subprocedure pid={}, resume processing ppid={}", procedure.getProcId(), 1996 parent.getProcId()); 1997 return; 1998 } 1999 } 2000 2001 private void updateStoreOnExec(RootProcedureState<TEnvironment> procStack, 2002 Procedure<TEnvironment> procedure, Procedure<TEnvironment>[] subprocs) { 2003 if (subprocs != null && !procedure.isFailed()) { 2004 if (LOG.isTraceEnabled()) { 2005 LOG.trace("Stored " + procedure + ", children " + Arrays.toString(subprocs)); 2006 } 2007 store.insert(procedure, subprocs); 2008 } else { 2009 LOG.trace("Store update {}", procedure); 2010 if (procedure.isFinished() && !procedure.hasParent()) { 2011 // remove child procedures 2012 final long[] childProcIds = procStack.getSubprocedureIds(); 2013 if (childProcIds != null) { 2014 store.delete(procedure, childProcIds); 2015 for (int i = 0; i < childProcIds.length; ++i) { 2016 procedures.remove(childProcIds[i]); 2017 } 2018 } else { 2019 store.update(procedure); 2020 } 2021 } else { 2022 store.update(procedure); 2023 } 2024 } 2025 } 2026 2027 private void handleInterruptedException(Procedure<TEnvironment> proc, InterruptedException e) { 2028 LOG.trace("Interrupt during {}. suspend and retry it later.", proc, e); 2029 // NOTE: We don't call Thread.currentThread().interrupt() 2030 // because otherwise all the subsequent calls e.g. Thread.sleep() will throw 2031 // the InterruptedException. If the master is going down, we will be notified 2032 // and the executor/store will be stopped. 2033 // (The interrupted procedure will be retried on the next run) 2034 } 2035 2036 private void execCompletionCleanup(Procedure<TEnvironment> proc) { 2037 final TEnvironment env = getEnvironment(); 2038 if (proc.hasLock()) { 2039 LOG.warn("Usually this should not happen, we will release the lock before if the procedure" 2040 + " is finished, even if the holdLock is true, arrive here means we have some holes where" 2041 + " we do not release the lock. And the releaseLock below may fail since the procedure may" 2042 + " have already been deleted from the procedure store."); 2043 releaseLock(proc, true); 2044 } 2045 try { 2046 proc.completionCleanup(env); 2047 } catch (Throwable e) { 2048 // Catch NullPointerExceptions or similar errors... 2049 LOG.error("CODE-BUG: uncatched runtime exception for procedure: " + proc, e); 2050 } 2051 } 2052 2053 private void procedureFinished(Procedure<TEnvironment> proc) { 2054 // call the procedure completion cleanup handler 2055 execCompletionCleanup(proc); 2056 2057 CompletedProcedureRetainer<TEnvironment> retainer = new CompletedProcedureRetainer<>(proc); 2058 2059 // update the executor internal state maps 2060 if (!proc.shouldWaitClientAck(getEnvironment())) { 2061 retainer.setClientAckTime(0); 2062 } 2063 2064 completed.put(proc.getProcId(), retainer); 2065 rollbackStack.remove(proc.getProcId()); 2066 procedures.remove(proc.getProcId()); 2067 2068 // call the runnableSet completion cleanup handler 2069 try { 2070 scheduler.completionCleanup(proc); 2071 } catch (Throwable e) { 2072 // Catch NullPointerExceptions or similar errors... 2073 LOG.error("CODE-BUG: uncatched runtime exception for completion cleanup: {}", proc, e); 2074 } 2075 2076 // Notify the listeners 2077 sendProcedureFinishedNotification(proc.getProcId()); 2078 } 2079 2080 RootProcedureState<TEnvironment> getProcStack(long rootProcId) { 2081 return rollbackStack.get(rootProcId); 2082 } 2083 2084 ProcedureScheduler getProcedureScheduler() { 2085 return scheduler; 2086 } 2087 2088 int getCompletedSize() { 2089 return completed.size(); 2090 } 2091 2092 public IdLock getProcExecutionLock() { 2093 return procExecutionLock; 2094 } 2095 2096 /** 2097 * Get a thread pool for executing some asynchronous tasks 2098 */ 2099 public ExecutorService getAsyncTaskExecutor() { 2100 return asyncTaskExecutor; 2101 } 2102 2103 // ========================================================================== 2104 // Worker Thread 2105 // ========================================================================== 2106 private class WorkerThread extends StoppableThread { 2107 private final AtomicLong executionStartTime = new AtomicLong(Long.MAX_VALUE); 2108 private volatile Procedure<TEnvironment> activeProcedure; 2109 2110 public WorkerThread(ThreadGroup group) { 2111 this(group, "PEWorker-"); 2112 } 2113 2114 protected WorkerThread(ThreadGroup group, String prefix) { 2115 super(group, prefix + workerId.incrementAndGet()); 2116 setDaemon(true); 2117 } 2118 2119 @Override 2120 public void sendStopSignal() { 2121 scheduler.signalAll(); 2122 } 2123 2124 /** 2125 * Encapsulates execution of the current {@link #activeProcedure} for easy tracing. 2126 */ 2127 private long runProcedure() throws IOException { 2128 final Procedure<TEnvironment> proc = this.activeProcedure; 2129 int activeCount = activeExecutorCount.incrementAndGet(); 2130 int runningCount = store.setRunningProcedureCount(activeCount); 2131 LOG.trace("Execute pid={} runningCount={}, activeCount={}", proc.getProcId(), runningCount, 2132 activeCount); 2133 executionStartTime.set(EnvironmentEdgeManager.currentTime()); 2134 IdLock.Entry lockEntry = procExecutionLock.getLockEntry(proc.getProcId()); 2135 try { 2136 executeProcedure(proc); 2137 } catch (AssertionError e) { 2138 LOG.info("ASSERT pid=" + proc.getProcId(), e); 2139 throw e; 2140 } finally { 2141 procExecutionLock.releaseLockEntry(lockEntry); 2142 activeCount = activeExecutorCount.decrementAndGet(); 2143 runningCount = store.setRunningProcedureCount(activeCount); 2144 LOG.trace("Halt pid={} runningCount={}, activeCount={}", proc.getProcId(), runningCount, 2145 activeCount); 2146 this.activeProcedure = null; 2147 executionStartTime.set(Long.MAX_VALUE); 2148 } 2149 return EnvironmentEdgeManager.currentTime(); 2150 } 2151 2152 @Override 2153 public void run() { 2154 long lastUpdate = EnvironmentEdgeManager.currentTime(); 2155 try { 2156 while (isRunning() && keepAlive(lastUpdate)) { 2157 @SuppressWarnings("unchecked") 2158 Procedure<TEnvironment> proc = scheduler.poll(keepAliveTime, TimeUnit.MILLISECONDS); 2159 if (proc == null) { 2160 continue; 2161 } 2162 this.activeProcedure = proc; 2163 lastUpdate = TraceUtil.trace(this::runProcedure, new ProcedureSpanBuilder(proc)); 2164 } 2165 } catch (Throwable t) { 2166 LOG.warn("Worker terminating UNNATURALLY {}", this.activeProcedure, t); 2167 } finally { 2168 LOG.trace("Worker terminated."); 2169 } 2170 workerThreads.remove(this); 2171 } 2172 2173 @Override 2174 public String toString() { 2175 Procedure<?> p = this.activeProcedure; 2176 return getName() + "(pid=" + (p == null ? Procedure.NO_PROC_ID : p.getProcId() + ")"); 2177 } 2178 2179 /** Returns the time since the current procedure is running */ 2180 public long getCurrentRunTime() { 2181 return EnvironmentEdgeManager.currentTime() - executionStartTime.get(); 2182 } 2183 2184 // core worker never timeout 2185 protected boolean keepAlive(long lastUpdate) { 2186 return true; 2187 } 2188 } 2189 2190 // A worker thread which can be added when core workers are stuck. Will timeout after 2191 // keepAliveTime if there is no procedure to run. 2192 private final class KeepAliveWorkerThread extends WorkerThread { 2193 public KeepAliveWorkerThread(ThreadGroup group) { 2194 super(group, "KeepAlivePEWorker-"); 2195 } 2196 2197 @Override 2198 protected boolean keepAlive(long lastUpdate) { 2199 return EnvironmentEdgeManager.currentTime() - lastUpdate < keepAliveTime; 2200 } 2201 } 2202 2203 // ---------------------------------------------------------------------------- 2204 // TODO-MAYBE: Should we provide a InlineChore to notify the store with the 2205 // full set of procedures pending and completed to write a compacted 2206 // version of the log (in case is a log)? 2207 // In theory no, procedures are have a short life, so at some point the store 2208 // will have the tracker saying everything is in the last log. 2209 // ---------------------------------------------------------------------------- 2210 2211 private final class WorkerMonitor extends InlineChore { 2212 public static final String WORKER_MONITOR_INTERVAL_CONF_KEY = 2213 "hbase.procedure.worker.monitor.interval.msec"; 2214 private static final int DEFAULT_WORKER_MONITOR_INTERVAL = 5000; // 5sec 2215 2216 public static final String WORKER_STUCK_THRESHOLD_CONF_KEY = 2217 "hbase.procedure.worker.stuck.threshold.msec"; 2218 private static final int DEFAULT_WORKER_STUCK_THRESHOLD = 10000; // 10sec 2219 2220 public static final String WORKER_ADD_STUCK_PERCENTAGE_CONF_KEY = 2221 "hbase.procedure.worker.add.stuck.percentage"; 2222 private static final float DEFAULT_WORKER_ADD_STUCK_PERCENTAGE = 0.5f; // 50% stuck 2223 2224 private float addWorkerStuckPercentage = DEFAULT_WORKER_ADD_STUCK_PERCENTAGE; 2225 private int timeoutInterval = DEFAULT_WORKER_MONITOR_INTERVAL; 2226 private int stuckThreshold = DEFAULT_WORKER_STUCK_THRESHOLD; 2227 2228 public WorkerMonitor() { 2229 refreshConfig(); 2230 } 2231 2232 @Override 2233 public void run() { 2234 final int stuckCount = checkForStuckWorkers(); 2235 checkThreadCount(stuckCount); 2236 2237 // refresh interval (poor man dynamic conf update) 2238 refreshConfig(); 2239 } 2240 2241 private int checkForStuckWorkers() { 2242 // check if any of the worker is stuck 2243 int stuckCount = 0; 2244 for (WorkerThread worker : workerThreads) { 2245 if (worker.getCurrentRunTime() < stuckThreshold) { 2246 continue; 2247 } 2248 2249 // WARN the worker is stuck 2250 stuckCount++; 2251 LOG.warn("Worker stuck {}, run time {}", worker, 2252 StringUtils.humanTimeDiff(worker.getCurrentRunTime())); 2253 } 2254 return stuckCount; 2255 } 2256 2257 private void checkThreadCount(final int stuckCount) { 2258 // nothing to do if there are no runnable tasks 2259 if (stuckCount < 1 || !scheduler.hasRunnables()) { 2260 return; 2261 } 2262 2263 // add a new thread if the worker stuck percentage exceed the threshold limit 2264 // and every handler is active. 2265 final float stuckPerc = ((float) stuckCount) / workerThreads.size(); 2266 // let's add new worker thread more aggressively, as they will timeout finally if there is no 2267 // work to do. 2268 if (stuckPerc >= addWorkerStuckPercentage && workerThreads.size() < maxPoolSize) { 2269 final KeepAliveWorkerThread worker = new KeepAliveWorkerThread(threadGroup); 2270 workerThreads.add(worker); 2271 worker.start(); 2272 LOG.debug("Added new worker thread {}", worker); 2273 } 2274 } 2275 2276 private void refreshConfig() { 2277 addWorkerStuckPercentage = 2278 conf.getFloat(WORKER_ADD_STUCK_PERCENTAGE_CONF_KEY, DEFAULT_WORKER_ADD_STUCK_PERCENTAGE); 2279 timeoutInterval = 2280 conf.getInt(WORKER_MONITOR_INTERVAL_CONF_KEY, DEFAULT_WORKER_MONITOR_INTERVAL); 2281 stuckThreshold = conf.getInt(WORKER_STUCK_THRESHOLD_CONF_KEY, DEFAULT_WORKER_STUCK_THRESHOLD); 2282 } 2283 2284 @Override 2285 public int getTimeoutInterval() { 2286 return timeoutInterval; 2287 } 2288 } 2289}