junction/details/Linear.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_LINEAR_H
  14 #define JUNCTION_DETAILS_LINEAR_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 // Enable this to force migration overflows (for test purposes):
  28 #define JUNCTION_LINEAR_FORCE_MIGRATION_OVERFLOWS 0
  29
  30 namespace junction {
  31 namespace details {
  32
  33 TURF_TRACE_DECLARE(Linear, 26)
  34
  35 template <class Map>
  36 struct Linear {
  37     typedef typename Map::Hash Hash;
  38     typedef typename Map::Value Value;
  39     typedef typename Map::KeyTraits KeyTraits;
  40     typedef typename Map::ValueTraits ValueTraits;
  41
  42     static const ureg InitialSize = 8;
  43     static const ureg TableMigrationUnitSize = 32;
  44     static const ureg CellsInUseSample = 256;
  45
  46     struct Cell {
  47         turf::Atomic<Hash> hash;
  48         turf::Atomic<Value> value;
  49     };
  50
  51     struct Table {
  52         const ureg sizeMask; // a power of two minus one
  53         turf::Atomic<sreg> cellsRemaining;
  54         turf::Mutex mutex;                   // to DCLI the TableMigration (stored in the jobCoordinator)
  55         SimpleJobCoordinator jobCoordinator; // makes all blocked threads participate in the migration
  56
  57         Table(ureg sizeMask) : sizeMask(sizeMask), cellsRemaining(sreg(sizeMask * 0.75f)) {
  58         }
  59
  60         static Table* create(ureg tableSize) {
  61             TURF_ASSERT(turf::util::isPowerOf2(tableSize));
  62             Table* table = (Table*) TURF_HEAP.alloc(sizeof(Table) + sizeof(Cell) * tableSize);
  63             new (table) Table(tableSize - 1);
  64             for (ureg j = 0; j < tableSize; j++) {
  65                 table->getCells()[j].hash.storeNonatomic(KeyTraits::NullHash);
  66                 table->getCells()[j].value.storeNonatomic(Value(ValueTraits::NullValue));
  67             }
  68             return table;
  69         }
  70
  71         void destroy() {
  72             this->Table::~Table();
  73             TURF_HEAP.free(this);
  74         }
  75
  76         Cell* getCells() const {
  77             return (Cell*) (this + 1);
  78         }
  79
  80         ureg getNumMigrationUnits() const {
  81             return sizeMask / TableMigrationUnitSize + 1;
  82         }
  83     };
  84
  85     class TableMigration : public SimpleJobCoordinator::Job {
  86     public:
  87         struct Source {
  88             Table* table;
  89             turf::Atomic<ureg> sourceIndex;
  90         };
  91
  92         Map& m_map;
  93         Table* m_destination;
  94         turf::Atomic<ureg> m_workerStatus; // number of workers + end flag
  95         turf::Atomic<bool> m_overflowed;
  96         turf::Atomic<sreg> m_unitsRemaining;
  97         ureg m_numSources;
  98
  99         TableMigration(Map& map) : m_map(map) {
 100         }
 101
 102         static TableMigration* create(Map& map, ureg numSources) {
 103             TableMigration* migration =
 104                 (TableMigration*) TURF_HEAP.alloc(sizeof(TableMigration) + sizeof(TableMigration::Source) * numSources);
 105             new (migration) TableMigration(map);
 106             migration->m_workerStatus.storeNonatomic(0);
 107             migration->m_overflowed.storeNonatomic(false);
 108             migration->m_unitsRemaining.storeNonatomic(0);
 109             migration->m_numSources = numSources;
 110             // Caller is responsible for filling in sources & destination
 111             return migration;
 112         }
 113
 114         virtual ~TableMigration() TURF_OVERRIDE {
 115         }
 116
 117         void destroy() {
 118             // Destroy all source tables.
 119             for (ureg i = 0; i < m_numSources; i++)
 120                 if (getSources()[i].table)
 121                     getSources()[i].table->destroy();
 122             // Delete the migration object itself.
 123             this->TableMigration::~TableMigration();
 124             TURF_HEAP.free(this);
 125         }
 126
 127         Source* getSources() const {
 128             return (Source*) (this + 1);
 129         }
 130
 131         bool migrateRange(Table* srcTable, ureg startIdx);
 132         virtual void run() TURF_OVERRIDE;
 133     };
 134
 135     static Cell* find(Hash hash, Table* table) {
 136         TURF_TRACE(Linear, 0, "[find] called", uptr(table), hash);
 137         TURF_ASSERT(table);
 138         TURF_ASSERT(hash != KeyTraits::NullHash);
 139         ureg sizeMask = table->sizeMask;
 140         for (ureg idx = ureg(hash);; idx++) {
 141             idx &= sizeMask;
 142             Cell* cell = table->getCells() + idx;
 143             // Load the hash that was there.
 144             Hash probeHash = cell->hash.load(turf::Relaxed);
 145             if (probeHash == hash) {
 146                 TURF_TRACE(Linear, 1, "[find] found existing cell", uptr(table), idx);
 147                 return cell;
 148             } else if (probeHash == KeyTraits::NullHash) {
 149                 return NULL;
 150             }
 151         }
 152     }
 153
 154     // FIXME: Possible optimization: Dedicated insert for migration? It wouldn't check for InsertResult_AlreadyFound.
 155     enum InsertResult { InsertResult_AlreadyFound, InsertResult_InsertedNew, InsertResult_Overflow };
 156     static InsertResult insert(Hash hash, Table* table, Cell*& cell) {
 157         TURF_TRACE(Linear, 2, "[insert] called", uptr(table), hash);
 158         TURF_ASSERT(table);
 159         TURF_ASSERT(hash != KeyTraits::NullHash);
 160         ureg sizeMask = table->sizeMask;
 161
 162         for (ureg idx = ureg(hash);; idx++) {
 163             idx &= sizeMask;
 164             cell = table->getCells() + idx;
 165             // Load the existing hash.
 166             Hash probeHash = cell->hash.load(turf::Relaxed);
 167             if (probeHash == hash) {
 168                 TURF_TRACE(Linear, 3, "[insert] found existing cell", uptr(table), idx);
 169                 return InsertResult_AlreadyFound; // Key found in table. Return the existing cell.
 170             }
 171             if (probeHash == KeyTraits::NullHash) {
 172                 // It's an empty cell. Try to reserve it.
 173                 // But first, decrement cellsRemaining to ensure we have permission to create new cells.
 174                 s32 prevCellsRemaining = table->cellsRemaining.fetchSub(1, turf::Relaxed);
 175                 if (prevCellsRemaining <= 0) {
 176                     // Table is overpopulated.
 177                     TURF_TRACE(Linear, 4, "[insert] ran out of cellsRemaining", prevCellsRemaining, 0);
 178                     table->cellsRemaining.fetchAdd(1, turf::Relaxed); // Undo cellsRemaining decrement
 179                     return InsertResult_Overflow;
 180                 }
 181                 // Try to reserve this cell.
 182                 Hash prevHash = cell->hash.compareExchange(KeyTraits::NullHash, hash, turf::Relaxed);
 183                 if (prevHash == KeyTraits::NullHash) {
 184                     // Success. We reserved a new cell.
 185                     TURF_TRACE(Linear, 5, "[insert] reserved cell", prevCellsRemaining, idx);
 186                     return InsertResult_InsertedNew;
 187                 }
 188                 // There was a race and another thread reserved that cell from under us.
 189                 TURF_TRACE(Linear, 6, "[insert] detected race to reserve cell", ureg(hash), idx);
 190                 table->cellsRemaining.fetchAdd(1, turf::Relaxed); // Undo cellsRemaining decrement
 191                 if (prevHash == hash) {
 192                     TURF_TRACE(Linear, 7, "[insert] race reserved same hash", ureg(hash), idx);
 193                     return InsertResult_AlreadyFound; // They inserted the same key. Return the existing cell.
 194                 }
 195             }
 196             // Try again in the next cell.
 197         }
 198     }
 199
 200     static void beginTableMigrationToSize(Map& map, Table* table, ureg nextTableSize) {
 201         // Create new migration by DCLI.
 202         TURF_TRACE(Linear, 8, "[beginTableMigrationToSize] called", 0, 0);
 203         SimpleJobCoordinator::Job* job = table->jobCoordinator.loadConsume();
 204         if (job) {
 205             TURF_TRACE(Linear, 9, "[beginTableMigrationToSize] new migration already exists", 0, 0);
 206         } else {
 207             turf::LockGuard<turf::Mutex> guard(table->mutex);
 208             job = table->jobCoordinator.loadConsume(); // Non-atomic would be sufficient, but that's OK.
 209             if (job) {
 210                 TURF_TRACE(Linear, 10, "[beginTableMigrationToSize] new migration already exists (double-checked)", 0, 0);
 211             } else {
 212                 // Create new migration.
 213                 TableMigration* migration = TableMigration::create(map, 1);
 214                 migration->m_unitsRemaining.storeNonatomic(table->getNumMigrationUnits());
 215                 migration->getSources()[0].table = table;
 216                 migration->getSources()[0].sourceIndex.storeNonatomic(0);
 217                 migration->m_destination = Table::create(nextTableSize);
 218                 // Publish the new migration.
 219                 table->jobCoordinator.storeRelease(migration);
 220             }
 221         }
 222     }
 223
 224     static void beginTableMigration(Map& map, Table* table) {
 225         // Estimate number of cells in use based on a small sample.
 226         ureg idx = 0;
 227         ureg sampleSize = turf::util::min<ureg>(table->sizeMask + 1, CellsInUseSample);
 228         ureg inUseCells = 0;
 229         for (; idx < sampleSize; idx++) {
 230             Cell* cell = table->getCells() + idx;
 231             Value value = cell->value.load(turf::Relaxed);
 232             if (value == Value(ValueTraits::Redirect)) {
 233                 // Another thread kicked off the jobCoordinator. The caller will participate upon return.
 234                 TURF_TRACE(Linear, 11, "[beginTableMigration] redirected while determining table size", 0, 0);
 235                 return;
 236             }
 237             if (value != Value(ValueTraits::NullValue))
 238                 inUseCells++;
 239         }
 240         float inUseRatio = float(inUseCells) / sampleSize;
 241         float estimatedInUse = (table->sizeMask + 1) * inUseRatio;
 242 #if JUNCTION_LINEAR_FORCE_MIGRATION_OVERFLOWS
 243         // Periodically underestimate the number of cells in use.
 244         // This exercises the code that handles overflow during migration.
 245         static ureg counter = 1;
 246         if ((++counter & 3) == 0) {
 247             estimatedInUse /= 4;
 248         }
 249 #endif
 250         ureg nextTableSize = turf::util::max(InitialSize, turf::util::roundUpPowerOf2(ureg(estimatedInUse * 2)));
 251         beginTableMigrationToSize(map, table, nextTableSize);
 252     }
 253 }; // Linear
 254
 255 template <class Map>
 256 bool Linear<Map>::TableMigration::migrateRange(Table* srcTable, ureg startIdx) {
 257     ureg srcSizeMask = srcTable->sizeMask;
 258     ureg endIdx = turf::util::min(startIdx + TableMigrationUnitSize, srcSizeMask + 1);
 259     sreg valuesMigrated = 0;
 260     // Iterate over source range.
 261     for (ureg srcIdx = startIdx; srcIdx < endIdx; srcIdx++) {
 262         Cell* srcCell = srcTable->getCells() + (srcIdx & srcSizeMask);
 263         Hash srcHash;
 264         Value srcValue;
 265         // Fetch the srcHash and srcValue.
 266         for (;;) {
 267             srcHash = srcCell->hash.load(turf::Relaxed);
 268             if (srcHash == KeyTraits::NullHash) {
 269                 // An unused cell. Try to put a Redirect marker in its value.
 270                 srcValue =
 271                     srcCell->value.compareExchange(Value(ValueTraits::NullValue), Value(ValueTraits::Redirect), turf::Relaxed);
 272                 if (srcValue == Value(ValueTraits::Redirect)) {
 273                     // srcValue is already marked Redirect due to previous incomplete migration.
 274                     TURF_TRACE(Linear, 12, "[migrateRange] empty cell already redirected", uptr(srcTable), srcIdx);
 275                     break;
 276                 }
 277                 if (srcValue == Value(ValueTraits::NullValue))
 278                     break; // Redirect has been placed. Break inner loop, continue outer loop.
 279                 TURF_TRACE(Linear, 13, "[migrateRange] race to insert key", uptr(srcTable), srcIdx);
 280                 // Otherwise, somebody just claimed the cell. Read srcHash again...
 281             } else {
 282                 // Check for deleted/uninitialized value.
 283                 srcValue = srcCell->value.load(turf::Relaxed);
 284                 if (srcValue == Value(ValueTraits::NullValue)) {
 285                     // Try to put a Redirect marker.
 286                     if (srcCell->value.compareExchangeStrong(srcValue, Value(ValueTraits::Redirect), turf::Relaxed))
 287                         break; // Redirect has been placed. Break inner loop, continue outer loop.
 288                     TURF_TRACE(Linear, 14, "[migrateRange] race to insert value", uptr(srcTable), srcIdx);
 289                     if (srcValue == Value(ValueTraits::Redirect)) {
 290                         // FIXME: I don't think this will happen. Investigate & change to assert
 291                         TURF_TRACE(Linear, 15, "[migrateRange] race inserted Redirect", uptr(srcTable), srcIdx);
 292                         break;
 293                     }
 294                 } else if (srcValue == Value(ValueTraits::Redirect)) {
 295                     // srcValue is already marked Redirect due to previous incomplete migration.
 296                     TURF_TRACE(Linear, 16, "[migrateRange] in-use cell already redirected", uptr(srcTable), srcIdx);
 297                     break;
 298                 }
 299
 300                 // We've got a key/value pair to migrate.
 301                 // Reserve a destination cell in the destination.
 302                 TURF_ASSERT(srcHash != KeyTraits::NullHash);
 303                 TURF_ASSERT(srcValue != Value(ValueTraits::NullValue));
 304                 TURF_ASSERT(srcValue != Value(ValueTraits::Redirect));
 305                 Cell* dstCell;
 306                 InsertResult result = insert(srcHash, m_destination, dstCell);
 307                 // During migration, a hash can only exist in one place among all the source tables,
 308                 // and it is only migrated by one thread. Therefore, the hash will never already exist
 309                 // in the destination table:
 310                 TURF_ASSERT(result != InsertResult_AlreadyFound);
 311                 if (result == InsertResult_Overflow) {
 312                     // Destination overflow.
 313                     // This can happen for several reasons. For example, the source table could have
 314                     // existed of all deleted cells when it overflowed, resulting in a small destination
 315                     // table size, but then another thread could re-insert all the same hashes
 316                     // before the migration completed.
 317                     // Caller will cancel the current migration and begin a new one.
 318                     return false;
 319                 }
 320                 // Migrate the old value to the new cell.
 321                 for (;;) {
 322                     // Copy srcValue to the destination.
 323                     dstCell->value.store(srcValue, turf::Relaxed);
 324                     // Try to place a Redirect marker in srcValue.
 325                     Value doubleCheckedSrcValue =
 326                         srcCell->value.compareExchange(srcValue, Value(ValueTraits::Redirect), turf::Relaxed);
 327                     TURF_ASSERT(doubleCheckedSrcValue !=
 328                                 Value(ValueTraits::Redirect)); // Only one thread can redirect a cell at a time.
 329                     if (doubleCheckedSrcValue == srcValue) {
 330                         // No racing writes to the src. We've successfully placed the Redirect marker.
 331                         // srcValue was non-NULL when we decided to migrate it, but it may have changed to NULL
 332                         // by a late-arriving erase.
 333                         if (srcValue == Value(ValueTraits::NullValue))
 334                             TURF_TRACE(Linear, 17, "[migrateRange] racing update was erase", uptr(srcTable), srcIdx);
 335                         break;
 336                     }
 337                     // There was a late-arriving write (or erase) to the src. Migrate the new value and try again.
 338                     TURF_TRACE(Linear, 18, "[migrateRange] race to update migrated value", uptr(srcTable), srcIdx);
 339                     srcValue = doubleCheckedSrcValue;
 340                 }
 341                 // Cell successfully migrated. Proceed to next source cell.
 342                 break;
 343             }
 344         }
 345     }
 346     // Range has been migrated successfully.
 347     return true;
 348 }
 349
 350 template <class Map>
 351 void Linear<Map>::TableMigration::run() {
 352     // Conditionally increment the shared # of workers.
 353     ureg probeStatus = m_workerStatus.load(turf::Relaxed);
 354     do {
 355         if (probeStatus & 1) {
 356             // End flag is already set, so do nothing.
 357             TURF_TRACE(Linear, 19, "[TableMigration::run] already ended", uptr(this), 0);
 358             return;
 359         }
 360     } while (!m_workerStatus.compareExchangeWeak(probeStatus, probeStatus + 2, turf::Relaxed, turf::Relaxed));
 361     // # of workers has been incremented, and the end flag is clear.
 362     TURF_ASSERT((probeStatus & 1) == 0);
 363
 364     // Iterate over all source tables.
 365     for (ureg s = 0; s < m_numSources; s++) {
 366         Source& source = getSources()[s];
 367         // Loop over all migration units in this source table.
 368         for (;;) {
 369             if (m_workerStatus.load(turf::Relaxed) & 1) {
 370                 TURF_TRACE(Linear, 20, "[TableMigration::run] detected end flag set", uptr(this), 0);
 371                 goto endMigration;
 372             }
 373             ureg startIdx = source.sourceIndex.fetchAdd(TableMigrationUnitSize, turf::Relaxed);
 374             if (startIdx >= source.table->sizeMask + 1)
 375                 break; // No more migration units in this table. Try next source table.
 376             bool overflowed = !migrateRange(source.table, startIdx);
 377             if (overflowed) {
 378                 // *** FAILED MIGRATION ***
 379                 // TableMigration failed due to destination table overflow.
 380                 // No other thread can declare the migration successful at this point, because *this* unit will never complete,
 381                 // hence m_unitsRemaining won't reach zero.
 382                 // However, multiple threads can independently detect a failed migration at the same time.
 383                 TURF_TRACE(Linear, 21, "[TableMigration::run] destination overflow", uptr(source.table), uptr(startIdx));
 384                 // The reason we store overflowed in a shared variable is because we can must flush all the worker threads before
 385                 // we can safely deal with the overflow. Therefore, the thread that detects the failure is often different from
 386                 // the thread
 387                 // that deals with it.
 388                 bool oldOverflowed = m_overflowed.exchange(overflowed, turf::Relaxed);
 389                 if (oldOverflowed)
 390                     TURF_TRACE(Linear, 22, "[TableMigration::run] race to set m_overflowed", uptr(overflowed),
 391                                uptr(oldOverflowed));
 392                 m_workerStatus.fetchOr(1, turf::Relaxed);
 393                 goto endMigration;
 394             }
 395             sreg prevRemaining = m_unitsRemaining.fetchSub(1, turf::Relaxed);
 396             TURF_ASSERT(prevRemaining > 0);
 397             if (prevRemaining == 1) {
 398                 // *** SUCCESSFUL MIGRATION ***
 399                 // That was the last chunk to migrate.
 400                 m_workerStatus.fetchOr(1, turf::Relaxed);
 401                 goto endMigration;
 402             }
 403         }
 404     }
 405     TURF_TRACE(Linear, 23, "[TableMigration::run] out of migration units", uptr(this), 0);
 406
 407 endMigration:
 408     // Decrement the shared # of workers.
 409     probeStatus = m_workerStatus.fetchSub(
 410         2, turf::AcquireRelease); // AcquireRelease makes all previous writes visible to the last worker thread.
 411     if (probeStatus >= 4) {
 412         // There are other workers remaining. Return here so that only the very last worker will proceed.
 413         TURF_TRACE(Linear, 24, "[TableMigration::run] not the last worker", uptr(this), uptr(probeStatus));
 414         return;
 415     }
 416
 417     // We're the very last worker thread.
 418     // Perform the appropriate post-migration step depending on whether the migration succeeded or failed.
 419     TURF_ASSERT(probeStatus == 3);
 420     bool overflowed = m_overflowed.loadNonatomic(); // No racing writes at this point
 421     if (!overflowed) {
 422         // The migration succeeded. This is the most likely outcome. Publish the new subtree.
 423         m_map.publishTableMigration(this);
 424         // End the jobCoodinator.
 425         getSources()[0].table->jobCoordinator.end();
 426     } else {
 427         // The migration failed due to the overflow of the destination table.
 428         Table* origTable = getSources()[0].table;
 429         turf::LockGuard<turf::Mutex> guard(origTable->mutex);
 430         SimpleJobCoordinator::Job* checkedJob = origTable->jobCoordinator.loadConsume();
 431         if (checkedJob != this) {
 432             TURF_TRACE(Linear, 25, "[TableMigration::run] a new TableMigration was already started", uptr(origTable),
 433                        uptr(checkedJob));
 434         } else {
 435             TableMigration* migration = TableMigration::create(m_map, m_numSources + 1);
 436             // Double the destination table size.
 437             migration->m_destination = Table::create((m_destination->sizeMask + 1) * 2);
 438             // Transfer source tables to the new migration.
 439             for (ureg i = 0; i < m_numSources; i++) {
 440                 migration->getSources()[i].table = getSources()[i].table;
 441                 getSources()[i].table = NULL;
 442                 migration->getSources()[i].sourceIndex.storeNonatomic(0);
 443             }
 444             migration->getSources()[m_numSources].table = m_destination;
 445             migration->getSources()[m_numSources].sourceIndex.storeNonatomic(0);
 446             // Calculate total number of migration units to move.
 447             ureg unitsRemaining = 0;
 448             for (ureg s = 0; s < migration->m_numSources; s++)
 449                 unitsRemaining += migration->getSources()[s].table->getNumMigrationUnits();
 450             migration->m_unitsRemaining.storeNonatomic(unitsRemaining);
 451             // Publish the new migration.
 452             origTable->jobCoordinator.storeRelease(migration);
 453         }
 454     }
 455
 456     // We're done with this TableMigration. Queue it for GC.
 457     DefaultQSBR.enqueue(&TableMigration::destroy, this);
 458 }
 459
 460 } // namespace details
 461 } // namespace junction
 462
 463 #endif // JUNCTION_DETAILS_LINEAR_H