junction/details/LeapFrog.h

   1 /*------------------------------------------------------------------------
   2   Junction: Concurrent data structures in C++
   3   Copyright (c) 2016 Jeff Preshing
   4
   5   Distributed under the Simplified BSD License.
   6   Original location: https://github.com/preshing/junction
   7
   8   This software is distributed WITHOUT ANY WARRANTY; without even the
   9   implied warranty of MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.
  10   See the LICENSE file for more information.
  11 ------------------------------------------------------------------------*/
  12
  13 #ifndef JUNCTION_DETAILS_LEAPFROG_H
  14 #define JUNCTION_DETAILS_LEAPFROG_H
  15
  16 #include <junction/Core.h>
  17 #include <turf/Atomic.h>
  18 #include <turf/Mutex.h>
  19 #include <turf/ManualResetEvent.h>
  20 #include <turf/Util.h>
  21 #include <junction/MapTraits.h>
  22 #include <turf/Trace.h>
  23 #include <turf/Heap.h>
  24 #include <junction/SimpleJobCoordinator.h>
  25 #include <junction/QSBR.h>
  26
  27 namespace junction {
  28 namespace details {
  29
  30 TURF_TRACE_DECLARE(LeapFrog, 33)
  31
  32 template<class Map>
  33 struct LeapFrog {
  34     typedef typename Map::Hash Hash;
  35     typedef typename Map::Value Value;
  36     typedef typename Map::KeyTraits KeyTraits;
  37     typedef typename Map::ValueTraits ValueTraits;
  38
  39     static const ureg InitialSize = 8;
  40     static const ureg TableMigrationUnitSize = 32;
  41     static const ureg LinearSearchLimit = 128;
  42     static const ureg CellsInUseSample = LinearSearchLimit;
  43     TURF_STATIC_ASSERT(LinearSearchLimit > 0 && LinearSearchLimit < 256); // Must fit in CellGroup::links
  44     TURF_STATIC_ASSERT(CellsInUseSample > 0 && CellsInUseSample <= LinearSearchLimit); // Limit sample to failed search chain
  45
  46     struct Cell {
  47         turf::Atomic<Hash> hash;
  48         turf::Atomic<Value> value;
  49     };
  50
  51     struct CellGroup {
  52         // Every cell in the table actually represents a bucket of cells, all linked together in a probe chain.
  53         // Each cell in the probe chain is located within the table itself.
  54         // "deltas" determines the index of the next cell in the probe chain.
  55         // The first cell in the chain is the one that was hashed. It may or may not actually belong in the bucket.
  56         // The "second" cell in the chain is given by deltas 0 - 3. It's guaranteed to belong in the bucket.
  57         // All subsequent cells in the chain is given by deltas 4 - 7. Also guaranteed to belong in the bucket.
  58         turf::Atomic<u8> deltas[8];
  59         Cell cells[4];
  60     };
  61
  62     struct Table {
  63         const ureg sizeMask;                    // a power of two minus one
  64         turf::Mutex mutex;                      // to DCLI the TableMigration (stored in the jobCoordinator)
  65         SimpleJobCoordinator jobCoordinator;    // makes all blocked threads participate in the migration
  66
  67         Table(ureg sizeMask) : sizeMask(sizeMask) {
  68         }
  69
  70         static Table* create(ureg tableSize) {
  71             TURF_ASSERT(turf::util::isPowerOf2(tableSize));
  72             TURF_ASSERT(tableSize >= 4);
  73             ureg numGroups = tableSize >> 2;
  74             Table* table = (Table*) TURF_HEAP.alloc(sizeof(Table) + sizeof(CellGroup) * numGroups);
  75             new(table) Table(tableSize - 1);
  76             for (ureg i = 0; i < numGroups; i++) {
  77                 CellGroup* group = table->getCellGroups() + i;
  78                 for (ureg j = 0; j < 4; j++) {
  79                     group->deltas[j].storeNonatomic(0);
  80                     group->deltas[j + 4].storeNonatomic(0);
  81                     group->cells[j].hash.storeNonatomic(KeyTraits::NullHash);
  82                     group->cells[j].value.storeNonatomic(Value(ValueTraits::NullValue));
  83                 }
  84             }
  85             return table;
  86         }
  87
  88         void destroy() {
  89             this->Table::~Table();
  90             TURF_HEAP.free(this);
  91         }
  92
  93         CellGroup* getCellGroups() const {
  94             return (CellGroup*) (this + 1);
  95         }
  96
  97         ureg getNumMigrationUnits() const {
  98             return sizeMask / TableMigrationUnitSize + 1;
  99         }
 100     };
 101
 102     class TableMigration : public SimpleJobCoordinator::Job {
 103     public:
 104         struct Source {
 105             Table* table;
 106             turf::Atomic<ureg> sourceIndex;
 107         };
 108
 109         Map& m_map;
 110         Table* m_destination;
 111         turf::Atomic<ureg> m_workerStatus;          // number of workers + end flag
 112         turf::Atomic<bool> m_overflowed;
 113         turf::Atomic<sreg> m_unitsRemaining;
 114         ureg m_numSources;
 115
 116         TableMigration(Map& map) : m_map(map) {
 117         }
 118
 119         static TableMigration* create(Map& map, ureg numSources) {
 120             TableMigration* migration = (TableMigration*) TURF_HEAP.alloc(sizeof(TableMigration) + sizeof(TableMigration::Source) * numSources);
 121             new(migration) TableMigration(map);
 122             migration->m_workerStatus.storeNonatomic(0);
 123             migration->m_overflowed.storeNonatomic(false);
 124             migration->m_unitsRemaining.storeNonatomic(0);
 125             migration->m_numSources = numSources;
 126             // Caller is responsible for filling in sources & destination
 127             return migration;
 128         }
 129
 130         virtual ~TableMigration() TURF_OVERRIDE {
 131         }
 132
 133         void destroy() {
 134             // Destroy all source tables.
 135             for (ureg i = 0; i < m_numSources; i++)
 136                 if (getSources()[i].table)
 137                     getSources()[i].table->destroy();
 138             // Delete the migration object itself.
 139             this->TableMigration::~TableMigration();
 140             TURF_HEAP.free(this);
 141         }
 142
 143         Source* getSources() const {
 144             return (Source*) (this + 1);
 145         }
 146
 147         bool migrateRange(Table* srcTable, ureg startIdx);
 148         virtual void run() TURF_OVERRIDE;
 149     };
 150
 151     static Cell* find(Hash hash, Table* table) {
 152         TURF_TRACE(LeapFrog, 0, "[find] called", uptr(table), hash);
 153         TURF_ASSERT(table);
 154         TURF_ASSERT(hash != KeyTraits::NullHash);
 155         ureg sizeMask = table->sizeMask;
 156         // Optimistically check hashed cell even though it might belong to another bucket
 157         ureg idx = hash & sizeMask;
 158         CellGroup* group = table->getCellGroups() + (idx >> 2);
 159         Cell* cell = group->cells + (idx & 3);
 160         Hash probeHash = cell->hash.load(turf::Relaxed);
 161         if (probeHash == hash) {
 162             TURF_TRACE(LeapFrog, 1, "[find] found existing cell optimistically", uptr(table), idx);
 163             return cell;
 164         } else if (probeHash == KeyTraits::NullHash) {
 165             return cell = NULL;
 166         }
 167         // Follow probe chain for our bucket
 168         u8 delta = group->deltas[idx & 3].load(turf::Relaxed);
 169         while (delta) {
 170             idx = (idx + delta) & sizeMask;
 171             group = table->getCellGroups() + (idx >> 2);
 172             cell = group->cells + (idx & 3);
 173             Hash probeHash = cell->hash.load(turf::Relaxed);
 174             // Note: probeHash might actually be NULL due to memory reordering of a concurrent insert,
 175             // but we don't check for it. We just follow the probe chain.
 176             if (probeHash == hash) {
 177                 TURF_TRACE(LeapFrog, 2, "[find] found existing cell", uptr(table), idx);
 178                 return cell;
 179             }
 180             delta = group->deltas[(idx & 3) + 4].load(turf::Relaxed);
 181         }
 182         // End of probe chain, not found
 183         return NULL;
 184     }
 185
 186     // FIXME: Possible optimization: Dedicated insert for migration? It wouldn't check for InsertResult_AlreadyFound.
 187     enum InsertResult {
 188         InsertResult_AlreadyFound,
 189         InsertResult_InsertedNew,
 190         InsertResult_Overflow
 191     };
 192     static InsertResult insert(Hash hash, Table* table, Cell*& cell, ureg& overflowIdx) {
 193         TURF_TRACE(LeapFrog, 3, "[insert] called", uptr(table), hash);
 194         TURF_ASSERT(table);
 195         TURF_ASSERT(hash != KeyTraits::NullHash);
 196         ureg sizeMask = table->sizeMask;
 197         ureg idx = hash;
 198
 199         // Check hashed cell first, though it may not even belong to the bucket.
 200         CellGroup* group = table->getCellGroups() + ((idx & sizeMask) >> 2);
 201         cell = group->cells + (idx & 3);
 202         Hash probeHash = cell->hash.load(turf::Relaxed);
 203         if (probeHash == KeyTraits::NullHash) {
 204             if (cell->hash.compareExchangeStrong(probeHash, hash, turf::Relaxed)) {
 205                 TURF_TRACE(LeapFrog, 4, "[insert] reserved first cell", uptr(table), idx);
 206                 // There are no links to set. We're done.
 207                 return InsertResult_InsertedNew;
 208             } else {
 209                 TURF_TRACE(LeapFrog, 5, "[insert] race to reserve first cell", uptr(table), idx);
 210                 // Fall through to check if it was the same hash...
 211             }
 212         }
 213         if (probeHash == hash) {
 214             TURF_TRACE(LeapFrog, 6, "[insert] found in first cell", uptr(table), idx);
 215             return InsertResult_AlreadyFound;
 216         }
 217
 218         // Follow the link chain for this bucket.
 219         ureg maxIdx = idx + sizeMask;
 220         ureg linkLevel = 0;
 221         turf::Atomic<u8>* prevLink;
 222         for (;;) {
 223         followLink:
 224             prevLink = group->deltas + ((idx & 3) + linkLevel);
 225             linkLevel = 4;
 226             u8 probeDelta = prevLink->load(turf::Relaxed);
 227             if (probeDelta) {
 228                 idx += probeDelta;
 229                 // Check the hash for this cell.
 230                 group = table->getCellGroups() + ((idx & sizeMask) >> 2);
 231                 cell = group->cells + (idx & 3);
 232                 probeHash = cell->hash.load(turf::Relaxed);
 233                 if (probeHash == KeyTraits::NullHash) {
 234                     // Cell was linked, but hash is not visible yet.
 235                     // We could avoid this case (and guarantee it's visible) using acquire & release, but instead,
 236                     // just poll until it becomes visible.
 237                     TURF_TRACE(LeapFrog, 7, "[insert] race to read hash", uptr(table), idx);
 238                     do {
 239                         probeHash = cell->hash.load(turf::Acquire);
 240                     } while (probeHash == KeyTraits::NullHash);
 241                 }
 242                 TURF_ASSERT(((probeHash ^ hash) & sizeMask) == 0);   // Only hashes in same bucket can be linked
 243                 if (probeHash == hash) {
 244                     TURF_TRACE(LeapFrog, 8, "[insert] found in probe chain", uptr(table), idx);
 245                     return InsertResult_AlreadyFound;
 246                 }
 247             } else {
 248                 // Reached the end of the link chain for this bucket.
 249                 // Switch to linear probing until we reserve a new cell or find a late-arriving cell in the same bucket.
 250                 ureg prevLinkIdx = idx;
 251                 TURF_ASSERT(sreg(maxIdx - idx) >= 0);    // Nobody would have linked an idx that's out of range.
 252                 ureg linearProbesRemaining = turf::util::min(maxIdx - idx, LinearSearchLimit);
 253                 while (linearProbesRemaining-- > 0) {
 254                     idx++;
 255                     group = table->getCellGroups() + ((idx & sizeMask) >> 2);
 256                     cell = group->cells + (idx & 3);
 257                     probeHash = cell->hash.load(turf::Relaxed);
 258                     if (probeHash == KeyTraits::NullHash) {
 259                         // It's an empty cell. Try to reserve it.
 260                         if (cell->hash.compareExchangeStrong(probeHash, hash, turf::Relaxed)) {
 261                             // Success. We've reserved the cell. Link it to previous cell in same bucket.
 262                             TURF_TRACE(LeapFrog, 9, "[insert] reserved cell", uptr(table), idx);
 263                             TURF_ASSERT(probeDelta == 0);
 264                             u8 desiredDelta = idx - prevLinkIdx;
 265 #if TURF_WITH_ASSERTS
 266                             probeDelta = prevLink->exchange(desiredDelta, turf::Relaxed);
 267                             TURF_ASSERT(probeDelta == 0 || probeDelta == desiredDelta);
 268 #else
 269                             prevLink->store(desiredDelta, turf::Relaxed);
 270 #endif
 271                             return InsertResult_InsertedNew;
 272                         } else {
 273                             TURF_TRACE(LeapFrog, 10, "[insert] race to reserve cell", uptr(table), idx);
 274                             // Fall through to check if it's the same hash...
 275                         }
 276                     }
 277                     Hash x = (probeHash ^ hash);
 278                     // Check for same hash.
 279                     if (!x) {
 280                         TURF_TRACE(LeapFrog, 11, "[insert] found outside probe chain", uptr(table), idx);
 281                         return InsertResult_AlreadyFound;
 282                     }
 283                     // Check for same bucket.
 284                     if ((x & sizeMask) == 0) {
 285                         TURF_TRACE(LeapFrog, 12, "[insert] found late-arriving cell in same bucket", uptr(table), idx);
 286                         // Attempt to set the link on behalf of the late-arriving cell.
 287                         // This is usually redundant, but if we don't attempt to set the late-arriving cell's link here,
 288                         // there's no guarantee that our own link chain will be well-formed by the time this function returns.
 289                         // (Indeed, subsequent lookups sometimes failed during testing, for this exact reason.)
 290                         u8 desiredDelta = idx - prevLinkIdx;
 291 #if TURF_WITH_ASSERTS
 292                         probeDelta = prevLink->exchange(desiredDelta, turf::Relaxed);
 293                         TURF_ASSERT(probeDelta == 0 || probeDelta == desiredDelta);
 294                         if (probeDelta == 0)
 295                             TURF_TRACE(LeapFrog, 13, "[insert] set link on behalf of late-arriving cell", uptr(table), idx);
 296 #else
 297                         prevLink->store(desiredDelta, turf::Relaxed);
 298 #endif
 299                         goto followLink;  // Try to follow link chain for the bucket again.
 300                     }
 301                     // Continue linear search...
 302                 }
 303                 // Table is too full to insert.
 304                 overflowIdx = idx + 1;
 305                 TURF_TRACE(LeapFrog, 14, "[insert] overflow", uptr(table), overflowIdx);
 306                 return InsertResult_Overflow;
 307             }
 308         }
 309     }
 310
 311     static void beginTableMigrationToSize(Map& map, Table* table, ureg nextTableSize) {
 312         // Create new migration by DCLI.
 313         TURF_TRACE(LeapFrog, 15, "[beginTableMigrationToSize] called", 0, 0);
 314         SimpleJobCoordinator::Job* job = table->jobCoordinator.loadConsume();
 315         if (job) {
 316             TURF_TRACE(LeapFrog, 16, "[beginTableMigrationToSize] new migration already exists", 0, 0);
 317         } else {
 318             turf::LockGuard<turf::Mutex> guard(table->mutex);
 319             job = table->jobCoordinator.loadConsume();  // Non-atomic would be sufficient, but that's OK.
 320             if (job) {
 321                 TURF_TRACE(LeapFrog, 17, "[beginTableMigrationToSize] new migration already exists (double-checked)", 0, 0);
 322             } else {
 323                 // Create new migration.
 324                 TableMigration* migration = TableMigration::create(map, 1);
 325                 migration->m_unitsRemaining.storeNonatomic(table->getNumMigrationUnits());
 326                 migration->getSources()[0].table = table;
 327                 migration->getSources()[0].sourceIndex.storeNonatomic(0);
 328                 migration->m_destination = Table::create(nextTableSize);
 329                 // Publish the new migration.
 330                 table->jobCoordinator.storeRelease(migration);
 331             }
 332         }
 333     }
 334
 335     static void beginTableMigration(Map& map, Table* table, ureg overflowIdx) {
 336         // Estimate number of cells in use based on a small sample.
 337         ureg sizeMask = table->sizeMask;
 338         ureg idx = overflowIdx - CellsInUseSample;
 339         ureg inUseCells = 0;
 340         for (ureg linearProbesRemaining = CellsInUseSample; linearProbesRemaining > 0; linearProbesRemaining--) {
 341             CellGroup* group = table->getCellGroups() + ((idx & sizeMask) >> 2);
 342             Cell* cell = group->cells + (idx & 3);
 343             Value value = cell->value.load(turf::Relaxed);
 344             if (value == Value(ValueTraits::Redirect)) {
 345                 // Another thread kicked off the jobCoordinator. The caller will participate upon return.
 346                 TURF_TRACE(LeapFrog, 18, "[beginTableMigration] redirected while determining table size", 0, 0);
 347                 return;
 348             }
 349             if (value != Value(ValueTraits::NullValue))
 350                 inUseCells++;
 351             idx++;
 352         }
 353         float inUseRatio = float(inUseCells) / CellsInUseSample;
 354         float estimatedInUse = (sizeMask + 1) * inUseRatio;
 355         ureg nextTableSize = turf::util::roundUpPowerOf2(ureg(estimatedInUse * 2));
 356         beginTableMigrationToSize(map, table, nextTableSize);
 357     }
 358 }; // LeapFrog
 359
 360 template<class Map>
 361 bool LeapFrog<Map>::TableMigration::migrateRange(Table* srcTable, ureg startIdx) {
 362     ureg srcSizeMask = srcTable->sizeMask;
 363     ureg endIdx = turf::util::min(startIdx + TableMigrationUnitSize, srcSizeMask + 1);
 364     // Iterate over source range.
 365     for (ureg srcIdx = startIdx; srcIdx < endIdx; srcIdx++) {
 366         CellGroup* srcGroup = srcTable->getCellGroups() + ((srcIdx & srcSizeMask) >> 2);
 367         Cell* srcCell = srcGroup->cells + (srcIdx & 3);
 368         Hash srcHash;
 369         Value srcValue;
 370         // Fetch the srcHash and srcValue.
 371         for (;;) {
 372             srcHash = srcCell->hash.load(turf::Relaxed);
 373             if (srcHash == KeyTraits::NullHash) {
 374                 // An unused cell. Try to put a Redirect marker in its value.
 375                 srcValue = srcCell->value.compareExchange(Value(ValueTraits::NullValue), Value(ValueTraits::Redirect), turf::Relaxed);
 376                 if (srcValue == Value(ValueTraits::Redirect)) {
 377                     // srcValue is already marked Redirect due to previous incomplete migration.
 378                     TURF_TRACE(LeapFrog, 19, "[migrateRange] empty cell already redirected", uptr(srcTable), srcIdx);
 379                     break;
 380                 }
 381                 if (srcValue == Value(ValueTraits::NullValue))
 382                     break;  // Redirect has been placed. Break inner loop, continue outer loop.
 383                 TURF_TRACE(LeapFrog, 20, "[migrateRange] race to insert key", uptr(srcTable), srcIdx);
 384                 // Otherwise, somebody just claimed the cell. Read srcHash again...
 385             } else {
 386                 // Check for deleted/uninitialized value.
 387                 srcValue = srcCell->value.load(turf::Relaxed);
 388                 if (srcValue == Value(ValueTraits::NullValue)) {
 389                     // Try to put a Redirect marker.
 390                     if (srcCell->value.compareExchangeStrong(srcValue, Value(ValueTraits::Redirect), turf::Relaxed))
 391                         break;  // Redirect has been placed. Break inner loop, continue outer loop.
 392                     TURF_TRACE(LeapFrog, 21, "[migrateRange] race to insert value", uptr(srcTable), srcIdx);
 393                     if (srcValue == Value(ValueTraits::Redirect)) {
 394                         // FIXME: I don't think this will happen. Investigate & change to assert
 395                         TURF_TRACE(LeapFrog, 22, "[migrateRange] race inserted Redirect", uptr(srcTable), srcIdx);
 396                         break;
 397                     }
 398                 } else if (srcValue == Value(ValueTraits::Redirect)) {
 399                     // srcValue is already marked Redirect due to previous incomplete migration.
 400                     TURF_TRACE(LeapFrog, 23, "[migrateRange] in-use cell already redirected", uptr(srcTable), srcIdx);
 401                     break;
 402                 }
 403
 404                 // We've got a key/value pair to migrate.
 405                 // Reserve a destination cell in the destination.
 406                 TURF_ASSERT(srcHash != KeyTraits::NullHash);
 407                 TURF_ASSERT(srcValue != Value(ValueTraits::NullValue));
 408                 TURF_ASSERT(srcValue != Value(ValueTraits::Redirect));
 409                 Cell* dstCell;
 410                 ureg overflowIdx;
 411                 InsertResult result = insert(srcHash, m_destination, dstCell, overflowIdx);
 412                 // During migration, a hash can only exist in one place among all the source tables,
 413                 // and it is only migrated by one thread. Therefore, the hash will never already exist
 414                 // in the destination table:
 415                 TURF_ASSERT(result != InsertResult_AlreadyFound);
 416                 if (result == InsertResult_Overflow) {
 417                     // Destination overflow.
 418                     // This can happen for several reasons. For example, the source table could have
 419                     // existed of all deleted cells when it overflowed, resulting in a small destination
 420                     // table size, but then another thread could re-insert all the same hashes
 421                     // before the migration completed.
 422                     // Caller will cancel the current migration and begin a new one.
 423                     return false;
 424                 }
 425                 // Migrate the old value to the new cell.
 426                 for (;;) {
 427                     // Copy srcValue to the destination.
 428                     dstCell->value.store(srcValue, turf::Relaxed);
 429                     // Try to place a Redirect marker in srcValue.
 430                     Value doubleCheckedSrcValue = srcCell->value.compareExchange(srcValue, Value(ValueTraits::Redirect), turf::Relaxed);
 431                     TURF_ASSERT(doubleCheckedSrcValue != Value(ValueTraits::Redirect)); // Only one thread can redirect a cell at a time.
 432                     if (doubleCheckedSrcValue == srcValue) {
 433                         // No racing writes to the src. We've successfully placed the Redirect marker.
 434                         // srcValue was non-NULL when we decided to migrate it, but it may have changed to NULL
 435                         // by a late-arriving erase.
 436                         if (srcValue == Value(ValueTraits::NullValue))
 437                             TURF_TRACE(LeapFrog, 24, "[migrateRange] racing update was erase", uptr(srcTable), srcIdx);
 438                         break;
 439                     }
 440                     // There was a late-arriving write (or erase) to the src. Migrate the new value and try again.
 441                     TURF_TRACE(LeapFrog, 25, "[migrateRange] race to update migrated value", uptr(srcTable), srcIdx);
 442                     srcValue = doubleCheckedSrcValue;
 443                 }
 444                 // Cell successfully migrated. Proceed to next source cell.
 445                 break;
 446             }
 447         }
 448     }
 449     // Range has been migrated successfully.
 450     return true;
 451 }
 452
 453 template <class Map>
 454 void LeapFrog<Map>::TableMigration::run() {
 455     // Conditionally increment the shared # of workers.
 456     ureg probeStatus = m_workerStatus.load(turf::Relaxed);
 457     do {
 458         if (probeStatus & 1) {
 459             // End flag is already set, so do nothing.
 460             TURF_TRACE(LeapFrog, 26, "[TableMigration::run] already ended", uptr(this), 0);
 461             return;
 462         }
 463     } while (!m_workerStatus.compareExchangeWeak(probeStatus, probeStatus + 2, turf::Relaxed, turf::Relaxed));
 464     // # of workers has been incremented, and the end flag is clear.
 465     TURF_ASSERT((probeStatus & 1) == 0);
 466
 467     // Iterate over all source tables.
 468     for (ureg s = 0; s < m_numSources; s++) {
 469         Source& source = getSources()[s];
 470         // Loop over all migration units in this source table.
 471         for (;;) {
 472             if (m_workerStatus.load(turf::Relaxed) & 1) {
 473                 TURF_TRACE(LeapFrog, 27, "[TableMigration::run] detected end flag set", uptr(this), 0);
 474                 goto endMigration;
 475             }
 476             ureg startIdx = source.sourceIndex.fetchAdd(TableMigrationUnitSize, turf::Relaxed);
 477             if (startIdx >= source.table->sizeMask + 1)
 478                 break;   // No more migration units in this table. Try next source table.
 479             bool overflowed = !migrateRange(source.table, startIdx);
 480             if (overflowed) {
 481                 // *** FAILED MIGRATION ***
 482                 // TableMigration failed due to destination table overflow.
 483                 // No other thread can declare the migration successful at this point, because *this* unit will never complete, hence m_unitsRemaining won't reach zero.
 484                 // However, multiple threads can independently detect a failed migration at the same time.
 485                 TURF_TRACE(LeapFrog, 28, "[TableMigration::run] destination overflow", uptr(source.table), uptr(startIdx));
 486                 // The reason we store overflowed in a shared variable is because we can must flush all the worker threads before
 487                 // we can safely deal with the overflow. Therefore, the thread that detects the failure is often different from the thread
 488                 // that deals with it.
 489                 bool oldOverflowed = m_overflowed.exchange(overflowed, turf::Relaxed);
 490                 if (oldOverflowed)
 491                     TURF_TRACE(LeapFrog, 29, "[TableMigration::run] race to set m_overflowed", uptr(overflowed), uptr(oldOverflowed));
 492                 m_workerStatus.fetchOr(1, turf::Relaxed);
 493                 goto endMigration;
 494             }
 495             sreg prevRemaining = m_unitsRemaining.fetchSub(1, turf::Relaxed);
 496             TURF_ASSERT(prevRemaining > 0);
 497             if (prevRemaining == 1) {
 498                 // *** SUCCESSFUL MIGRATION ***
 499                 // That was the last chunk to migrate.
 500                 m_workerStatus.fetchOr(1, turf::Relaxed);
 501                 goto endMigration;
 502             }
 503         }
 504     }
 505     TURF_TRACE(LeapFrog, 30, "[TableMigration::run] out of migration units", uptr(this), 0);
 506
 507 endMigration:
 508     // Decrement the shared # of workers.
 509     probeStatus = m_workerStatus.fetchSub(2, turf::AcquireRelease);    // AcquireRelease makes all previous writes visible to the last worker thread.
 510     if (probeStatus >= 4) {
 511         // There are other workers remaining. Return here so that only the very last worker will proceed.
 512         TURF_TRACE(LeapFrog, 31, "[TableMigration::run] not the last worker", uptr(this), uptr(probeStatus));
 513         return;
 514     }
 515
 516     // We're the very last worker thread.
 517     // Perform the appropriate post-migration step depending on whether the migration succeeded or failed.
 518     TURF_ASSERT(probeStatus == 3);
 519     bool overflowed = m_overflowed.loadNonatomic();  // No racing writes at this point
 520     if (!overflowed) {
 521         // The migration succeeded. This is the most likely outcome. Publish the new subtree.
 522         m_map.publishTableMigration(this);
 523         // End the jobCoodinator.
 524         getSources()[0].table->jobCoordinator.end();
 525     } else {
 526         // The migration failed due to the overflow of the destination table.
 527         Table* origTable = getSources()[0].table;
 528         turf::LockGuard<turf::Mutex> guard(origTable->mutex);
 529         SimpleJobCoordinator::Job* checkedJob = origTable->jobCoordinator.loadConsume();
 530         if (checkedJob != this) {
 531             TURF_TRACE(LeapFrog, 32, "[TableMigration::run] a new TableMigration was already started", uptr(origTable), uptr(checkedJob));
 532         } else {
 533             TableMigration* migration = TableMigration::create(m_map, m_numSources + 1);
 534             // Double the destination table size.
 535             migration->m_destination = Table::create((m_destination->sizeMask + 1) * 2);
 536             // Transfer source tables to the new migration.
 537             for (ureg i = 0; i < m_numSources; i++) {
 538                 migration->getSources()[i].table = getSources()[i].table;
 539                 getSources()[i].table = NULL;
 540                 migration->getSources()[i].sourceIndex.storeNonatomic(0);
 541             }
 542             migration->getSources()[m_numSources].table = m_destination;
 543             migration->getSources()[m_numSources].sourceIndex.storeNonatomic(0);
 544             // Calculate total number of migration units to move.
 545             ureg unitsRemaining = 0;
 546             for (ureg s = 0; s < migration->m_numSources; s++)
 547                 unitsRemaining += migration->getSources()[s].table->getNumMigrationUnits();
 548             migration->m_unitsRemaining.storeNonatomic(unitsRemaining);
 549             // Publish the new migration.
 550             origTable->jobCoordinator.storeRelease(migration);
 551         }
 552     }
 553
 554     // We're done with this TableMigration. Queue it for GC.
 555     DefaultQSBR.enqueue(&TableMigration::destroy, this);
 556 }
 557
 558 } // namespace details
 559 } // namespace junction
 560
 561 #endif // JUNCTION_DETAILS_LEAPFROG_H