junction/details/Grampa.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_GRAMPA_H
  14 #define JUNCTION_DETAILS_GRAMPA_H
  15
  16 #include <junction/Core.h>
  17 #include <turf/Atomic.h>
  18 #include <junction/striped/Mutex.h>
  19 #include <junction/striped/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 #include <memory.h>
  27
  28 namespace junction {
  29 namespace details {
  30
  31 #if JUNCTION_TRACK_GRAMPA_STATS
  32 struct GrampaCounter {
  33     turf::Atomic<ureg> total;
  34     turf::Atomic<sreg> current;
  35
  36     void increment() {
  37         total.fetchAdd(1, turf::Relaxed);
  38         current.fetchAdd(1, turf::Relaxed);
  39     }
  40
  41     void decrement() {
  42         current.fetchSub(1, turf::Relaxed);
  43     }
  44 };
  45
  46 struct GrampaStats {
  47     GrampaCounter numTables;
  48     GrampaCounter numTableMigrations;
  49     GrampaCounter numFlatTrees;
  50     GrampaCounter numFlatTreeMigrations;
  51
  52     static GrampaStats Instance;    // Zero-initialized
  53 };
  54 #endif
  55
  56 TURF_TRACE_DECLARE(Grampa, 37)
  57
  58 template<class Map>
  59 struct Grampa {
  60     typedef typename Map::Hash Hash;
  61     typedef typename Map::Value Value;
  62     typedef typename Map::KeyTraits KeyTraits;
  63     typedef typename Map::ValueTraits ValueTraits;
  64
  65     static const ureg RedirectFlatTree = 1;
  66     static const ureg InitialSize = 8;
  67     static const ureg TableMigrationUnitSize = 32;
  68     static const ureg FlatTreeMigrationUnitSize = 32;
  69     static const ureg LinearSearchLimit = 128;
  70     static const ureg CellsInUseSample = LinearSearchLimit;
  71     TURF_STATIC_ASSERT(LinearSearchLimit > 0 && LinearSearchLimit < 256); // Must fit in CellGroup::links
  72     TURF_STATIC_ASSERT(CellsInUseSample > 0 && CellsInUseSample <= LinearSearchLimit); // Limit sample to failed search chain
  73
  74     static const ureg MinTableSize = 8;
  75     static const ureg LeafSizeBits = 10;
  76     static const ureg LeafSize = (ureg(1) << LeafSizeBits);
  77
  78     struct Cell {
  79         // If value == Redirect, threads participate in the jobCoordinator.
  80         turf::Atomic<Hash> hash;
  81         turf::Atomic<Value> value;
  82     };
  83
  84     struct CellGroup {
  85         // Every cell in the table actually represents a bucket of cells, all linked together in a probe chain.
  86         // Each cell in the probe chain is located within the table itself.
  87         // "deltas" determines the index of the next cell in the probe chain.
  88         // The first cell in the chain is the one that was hashed. It may or may not actually belong in the bucket.
  89         // The "second" cell in the chain is given by deltas 0 - 3. It's guaranteed to belong in the bucket.
  90         // All subsequent cells in the chain is given by deltas 4 - 7. Also guaranteed to belong in the bucket.
  91         turf::Atomic<u8> deltas[8];
  92         Cell cells[4];
  93     };
  94
  95     struct Table {
  96         // unsafeRangeShift determines how many slots are occupied by this Table in the flattree.
  97         // The range of hashes stored in this table is given by (1 << shift).
  98         // eg. If the entire map is stored in a single table, then Table::shift == HASH_BITS.
  99         // If the entire map is stored in two tables, then Table::shift == (HASH_BITS - 1) for each table.
 100         // FlatTree::shift is always <= Table::shift for all the tables it contains.
 101         const ureg sizeMask;                            // a power of two minus one
 102         const Hash baseHash;
 103         const ureg unsafeRangeShift;
 104         junction::striped::ManualResetEvent isPublished;    // To prevent publishing a subtree before its parent is published (happened in testing)
 105         junction::striped::Mutex mutex;                     // to DCLI the TableMigration (stored in the jobCoordinator)
 106         SimpleJobCoordinator jobCoordinator;            // makes all blocked threads participate in the migration
 107
 108         Table(ureg sizeMask, Hash baseHash, ureg unsafeRangeShift) : sizeMask(sizeMask), baseHash(baseHash), unsafeRangeShift(unsafeRangeShift) {
 109         }
 110
 111         static Table* create(ureg tableSize, ureg baseHash, ureg unsafeShift) {
 112             TURF_ASSERT(turf::util::isPowerOf2(tableSize));
 113             TURF_ASSERT(unsafeShift > 0 && unsafeShift <= sizeof(Hash) * 8);
 114             TURF_ASSERT(tableSize >= 4);
 115             ureg numGroups = tableSize >> 2;
 116             Table* table = (Table*) TURF_HEAP.alloc(sizeof(Table) + sizeof(CellGroup) * numGroups);
 117             new(table) Table(tableSize - 1, baseHash, (u8) unsafeShift);
 118             for (ureg i = 0; i < numGroups; i++) {
 119                 CellGroup* group = table->getCellGroups() + i;
 120                 for (ureg j = 0; j < 4; j++) {
 121                     group->deltas[j].storeNonatomic(0);
 122                     group->deltas[j + 4].storeNonatomic(0);
 123                     group->cells[j].hash.storeNonatomic(KeyTraits::NullHash);
 124                     group->cells[j].value.storeNonatomic(Value(ValueTraits::NullValue));
 125                 }
 126             }
 127 #if JUNCTION_TRACK_GRAMPA_STATS
 128             GrampaStats::Instance.numTables.increment();
 129 #endif
 130             return table;
 131         }
 132
 133         void destroy() {
 134 #if JUNCTION_TRACK_GRAMPA_STATS
 135             GrampaStats::Instance.numTables.decrement();
 136 #endif
 137             this->Table::~Table();
 138             TURF_HEAP.free(this);
 139         }
 140
 141         CellGroup* getCellGroups() const {
 142             return (CellGroup*) (this + 1);
 143         }
 144
 145         ureg getNumMigrationUnits() const {
 146             return sizeMask / TableMigrationUnitSize + 1;
 147         }
 148     };
 149
 150     class TableMigration : public SimpleJobCoordinator::Job {
 151     public:
 152         struct Source {
 153             Table* table;
 154             turf::Atomic<ureg> sourceIndex;
 155         };
 156
 157         Map& m_map;
 158         Hash m_baseHash;                           // The lowest possible hash value in this subtree; determines index in flattree.
 159         // If m_numDestinations == 1, m_shift == 0.
 160         // Otherwise, m_shift tells (indirectly) the size of the flattree in which our subtree would exactly fit: 1 << (HASH_BITS - m_shift).
 161         // This ensures that m_shift is always less than sizeof(Hash) * 8, so that shifting by m_shift is not undefined behavior.
 162         // To determine the subtree index for a hash during migration, we use: (hash >> m_shift) & (m_numDestinations - 1)
 163         // A mask is used since we are only migrating a subtree -- not necessarily the entire map.
 164         ureg m_safeShift;
 165         turf::Atomic<ureg> m_workerStatus;          // number of workers + end flag
 166         turf::Atomic<sreg> m_overflowTableIndex;
 167         turf::Atomic<sreg> m_unitsRemaining;
 168         ureg m_numSources;
 169         ureg m_numDestinations;                 // The size of the subtree being created. Some table pointers may be repeated.
 170
 171         TableMigration(Map& map) : m_map(map) {
 172         }
 173
 174         static TableMigration* create(Map& map, ureg numSources, ureg numDestinations) {
 175             TableMigration* migration = (TableMigration*) TURF_HEAP.alloc(sizeof(TableMigration) + sizeof(TableMigration::Source) * numSources + sizeof(Table*) * numDestinations);
 176             new(migration) TableMigration(map);
 177             migration->m_workerStatus.storeNonatomic(0);
 178             migration->m_overflowTableIndex.storeNonatomic(-1);
 179             migration->m_unitsRemaining.storeNonatomic(0);
 180             migration->m_numSources = numSources;
 181             migration->m_numDestinations = numDestinations;
 182             // Caller is responsible for filling in source & destination pointers
 183 #if JUNCTION_TRACK_GRAMPA_STATS
 184             GrampaStats::Instance.numTableMigrations.increment();
 185 #endif
 186             return migration;
 187         }
 188
 189         virtual ~TableMigration() TURF_OVERRIDE {
 190         }
 191
 192         void destroy() {
 193 #if JUNCTION_TRACK_GRAMPA_STATS
 194             GrampaStats::Instance.numTableMigrations.decrement();
 195 #endif
 196             // Destroy all source tables.
 197             for (ureg i = 0; i < m_numSources; i++)
 198                 if (getSources()[i].table)
 199                     getSources()[i].table->destroy();
 200             // Delete the migration object itself.
 201             this->TableMigration::~TableMigration();
 202             TURF_HEAP.free(this);
 203         }
 204
 205         ureg getUnsafeShift() const {
 206             return m_safeShift ? m_safeShift : (sizeof(Hash) * 8);
 207         }
 208
 209         Source* getSources() const {
 210             return (Source*) (this + 1);
 211         }
 212
 213         Table** getDestinations() const {
 214             return (Table**) (getSources() + m_numSources);
 215         }
 216
 217         sreg migrateRange(Table* srcTable, ureg startIdx);
 218         virtual void run() TURF_OVERRIDE;
 219     };
 220
 221     class FlatTreeMigration;
 222
 223     struct FlatTree {
 224         // The size of the flattree is 1 << 64 - HASH_BITS.
 225         // Or, stated another way, (Hash(-1) >> shift) + 1.
 226         // To determine the flattree index for a given hash, we simply use: (hash >> shift)
 227         // Smaller shift == more significant bits used as an index == bigger flattree.
 228         // For example, the simplest flattree has only two entries, and only the most significant
 229         // bit of each hash is used as the flattree index. In that case, shift == HASH_BITS - 1.
 230         // Each time the flattree doubles in size, shift decreases by 1.
 231         const ureg safeShift;
 232         junction::striped::Mutex mutex;
 233         FlatTreeMigration* migration;   // Protected by mutex
 234
 235         FlatTree(ureg safeShift) : safeShift(safeShift), migration(NULL) {
 236             // A FlatTree always has at least two tables, so the shift is always safe.
 237             TURF_ASSERT(safeShift < sizeof(Hash) * 8);
 238         }
 239
 240         static FlatTree* create(ureg safeShift) {
 241             // A flattree always has at least two tables, so the shift is always safe.
 242             TURF_ASSERT(safeShift < sizeof(Hash) * 8);
 243             ureg numLeaves = (Hash(-1) >> safeShift) + 1;
 244             FlatTree* flatTree = (FlatTree*) TURF_HEAP.alloc(sizeof(FlatTree) + sizeof(turf::Atomic<Table*>) * numLeaves);
 245             new(flatTree) FlatTree(safeShift);
 246             // Caller will initialize flatTree->getTables()
 247 #if JUNCTION_TRACK_GRAMPA_STATS
 248             GrampaStats::Instance.numFlatTrees.increment();
 249 #endif
 250             return flatTree;
 251         }
 252
 253         void destroy() {
 254 #if JUNCTION_TRACK_GRAMPA_STATS
 255             GrampaStats::Instance.numFlatTrees.decrement();
 256 #endif
 257             this->FlatTree::~FlatTree();
 258             TURF_HEAP.free(this);
 259         }
 260
 261         turf::Atomic<Table*>* getTables() const {
 262             return (turf::Atomic<Table*>*) (this + 1);
 263         }
 264
 265         ureg getSize() const {
 266             return (Hash(-1) >> safeShift) + 1;
 267         }
 268
 269         ureg getNumMigrationUnits() const {
 270             ureg sizeMask = Hash(-1) >> safeShift;
 271             return sizeMask / FlatTreeMigrationUnitSize + 1;
 272         }
 273     };
 274
 275     class FlatTreeMigration : public SimpleJobCoordinator::Job {
 276     public:
 277         Map& m_map;
 278         FlatTree* m_source;
 279         FlatTree* m_destination;
 280         turf::Atomic<ureg> m_workerStatus;
 281         turf::Atomic<ureg> m_sourceIndex;
 282         turf::Atomic<sreg> m_unitsRemaining;
 283         junction::striped::ManualResetEvent m_completed;
 284
 285         FlatTreeMigration(Map& map, FlatTree* flatTree, ureg shift) : m_map(map) {
 286             m_source = flatTree;
 287             m_destination = FlatTree::create(shift);
 288             m_workerStatus.storeNonatomic(0);
 289             m_sourceIndex.storeNonatomic(0);
 290             m_unitsRemaining.storeNonatomic(flatTree->getNumMigrationUnits());
 291 #if JUNCTION_TRACK_GRAMPA_STATS
 292             GrampaStats::Instance.numFlatTreeMigrations.increment();
 293 #endif
 294         }
 295
 296         virtual ~FlatTreeMigration() TURF_OVERRIDE {
 297 #if JUNCTION_TRACK_GRAMPA_STATS
 298             GrampaStats::Instance.numFlatTreeMigrations.decrement();
 299 #endif
 300             // Delete source flattree.
 301             m_source->destroy();
 302         }
 303
 304         void destroy() {
 305             delete this;
 306         }
 307
 308         virtual void run() TURF_OVERRIDE;
 309     };
 310
 311     static void garbageCollectTable(Table* table) {
 312         TURF_ASSERT(table);
 313         DefaultQSBR.enqueue(&Table::destroy, table);
 314     }
 315
 316     static void garbageCollectFlatTree(FlatTree* flatTree) {
 317         TURF_ASSERT(flatTree);
 318         DefaultQSBR.enqueue(&FlatTree::destroy, flatTree);
 319     }
 320
 321     static Cell* find(Hash hash, Table* table, ureg sizeMask) {
 322         TURF_TRACE(Grampa, 0, "[find] called", uptr(table), hash);
 323         TURF_ASSERT(table);
 324         TURF_ASSERT(hash != KeyTraits::NullHash);
 325         // Optimistically check hashed cell even though it might belong to another bucket
 326         ureg idx = hash & sizeMask;
 327         CellGroup* group = table->getCellGroups() + (idx >> 2);
 328         Cell* cell = group->cells + (idx & 3);
 329         Hash probeHash = cell->hash.load(turf::Relaxed);
 330         if (probeHash == hash) {
 331             TURF_TRACE(Grampa, 1, "[find] found existing cell optimistically", uptr(table), idx);
 332             return cell;
 333         } else if (probeHash == KeyTraits::NullHash) {
 334             return cell = NULL;
 335         }
 336         // Follow probe chain for our bucket
 337         u8 delta = group->deltas[idx & 3].load(turf::Relaxed);
 338         while (delta) {
 339             idx = (idx + delta) & sizeMask;
 340             group = table->getCellGroups() + (idx >> 2);
 341             cell = group->cells + (idx & 3);
 342             Hash probeHash = cell->hash.load(turf::Relaxed);
 343             // Note: probeHash might actually be NULL due to memory reordering of a concurrent insert,
 344             // but we don't check for it. We just follow the probe chain.
 345             if (probeHash == hash) {
 346                 TURF_TRACE(Grampa, 2, "[find] found existing cell", uptr(table), idx);
 347                 return cell;
 348             }
 349             delta = group->deltas[(idx & 3) + 4].load(turf::Relaxed);
 350         }
 351         // End of probe chain, not found
 352         return NULL;
 353     }
 354
 355     // FIXME: Possible optimization: Dedicated insert for migration? It wouldn't check for InsertResult_AlreadyFound.
 356     enum InsertResult {
 357         InsertResult_AlreadyFound,
 358         InsertResult_InsertedNew,
 359         InsertResult_Overflow
 360     };
 361     static InsertResult insert(Hash hash, Table* table, ureg sizeMask, Cell*& cell, ureg& overflowIdx) {
 362         TURF_TRACE(Grampa, 3, "[insert] called", uptr(table), hash);
 363         TURF_ASSERT(table);
 364         TURF_ASSERT(hash != KeyTraits::NullHash);
 365         ureg idx = hash;
 366
 367         // Check hashed cell first, though it may not even belong to the bucket.
 368         CellGroup* group = table->getCellGroups() + ((idx & sizeMask) >> 2);
 369         cell = group->cells + (idx & 3);
 370         Hash probeHash = cell->hash.load(turf::Relaxed);
 371         if (probeHash == KeyTraits::NullHash) {
 372             if (cell->hash.compareExchangeStrong(probeHash, hash, turf::Relaxed)) {
 373                 TURF_TRACE(Grampa, 4, "[insert] reserved first cell", uptr(table), idx);
 374                 // There are no links to set. We're done.
 375                 return InsertResult_InsertedNew;
 376             } else {
 377                 TURF_TRACE(Grampa, 5, "[insert] race to reserve first cell", uptr(table), idx);
 378                 // Fall through to check if it was the same hash...
 379             }
 380         }
 381         if (probeHash == hash) {
 382             TURF_TRACE(Grampa, 6, "[insert] found in first cell", uptr(table), idx);
 383             return InsertResult_AlreadyFound;
 384         }
 385
 386         // Follow the link chain for this bucket.
 387         ureg maxIdx = idx + sizeMask;
 388         ureg linkLevel = 0;
 389         turf::Atomic<u8>* prevLink;
 390         for (;;) {
 391         followLink:
 392             prevLink = group->deltas + ((idx & 3) + linkLevel);
 393             linkLevel = 4;
 394             u8 probeDelta = prevLink->load(turf::Relaxed);
 395             if (probeDelta) {
 396                 idx += probeDelta;
 397                 // Check the hash for this cell.
 398                 group = table->getCellGroups() + ((idx & sizeMask) >> 2);
 399                 cell = group->cells + (idx & 3);
 400                 probeHash = cell->hash.load(turf::Relaxed);
 401                 if (probeHash == KeyTraits::NullHash) {
 402                     // Cell was linked, but hash is not visible yet.
 403                     // We could avoid this case (and guarantee it's visible) using acquire & release, but instead,
 404                     // just poll until it becomes visible.
 405                     TURF_TRACE(Grampa, 7, "[insert] race to read hash", uptr(table), idx);
 406                     do {
 407                         probeHash = cell->hash.load(turf::Acquire);
 408                     } while (probeHash == KeyTraits::NullHash);
 409                 }
 410                 TURF_ASSERT(((probeHash ^ hash) & sizeMask) == 0);   // Only hashes in same bucket can be linked
 411                 if (probeHash == hash) {
 412                     TURF_TRACE(Grampa, 8, "[insert] found in probe chain", uptr(table), idx);
 413                     return InsertResult_AlreadyFound;
 414                 }
 415             } else {
 416                 // Reached the end of the link chain for this bucket.
 417                 // Switch to linear probing until we reserve a new cell or find a late-arriving cell in the same bucket.
 418                 ureg prevLinkIdx = idx;
 419                 TURF_ASSERT(sreg(maxIdx - idx) >= 0);    // Nobody would have linked an idx that's out of range.
 420                 ureg linearProbesRemaining = turf::util::min(maxIdx - idx, LinearSearchLimit);
 421                 while (linearProbesRemaining-- > 0) {
 422                     idx++;
 423                     group = table->getCellGroups() + ((idx & sizeMask) >> 2);
 424                     cell = group->cells + (idx & 3);
 425                     probeHash = cell->hash.load(turf::Relaxed);
 426                     if (probeHash == KeyTraits::NullHash) {
 427                         // It's an empty cell. Try to reserve it.
 428                         if (cell->hash.compareExchangeStrong(probeHash, hash, turf::Relaxed)) {
 429                             // Success. We've reserved the cell. Link it to previous cell in same bucket.
 430                             TURF_TRACE(Grampa, 9, "[insert] reserved cell", uptr(table), idx);
 431                             TURF_ASSERT(probeDelta == 0);
 432                             u8 desiredDelta = idx - prevLinkIdx;
 433                             // Note: another thread could actually set the link on our behalf (see below).
 434 #if TURF_WITH_ASSERTS
 435                             probeDelta = prevLink->exchange(desiredDelta, turf::Relaxed);
 436                             TURF_ASSERT(probeDelta == 0 || probeDelta == desiredDelta);
 437 #else
 438                             prevLink->store(desiredDelta, turf::Relaxed);
 439 #endif
 440                             return InsertResult_InsertedNew;
 441                         } else {
 442                             TURF_TRACE(Grampa, 10, "[insert] race to reserve cell", uptr(table), idx);
 443                             // Fall through to check if it's the same hash...
 444                         }
 445                     }
 446                     Hash x = (probeHash ^ hash);
 447                     // Check for same hash.
 448                     if (!x) {
 449                         TURF_TRACE(Grampa, 11, "[insert] found outside probe chain", uptr(table), idx);
 450                         return InsertResult_AlreadyFound;
 451                     }
 452                     // Check for same bucket.
 453                     if ((x & sizeMask) == 0) {
 454                         TURF_TRACE(Grampa, 12, "[insert] found late-arriving cell in same bucket", uptr(table), idx);
 455                         // Attempt to set the link on behalf of the late-arriving cell.
 456                         // This is usually redundant, but if we don't attempt to set the late-arriving cell's link here,
 457                         // there's no guarantee that our own link chain will be well-formed by the time this function returns.
 458                         // (Indeed, subsequent lookups sometimes failed during testing, for this exact reason.)
 459                         u8 desiredDelta = idx - prevLinkIdx;
 460 #if TURF_WITH_ASSERTS
 461                         probeDelta = prevLink->exchange(desiredDelta, turf::Relaxed);
 462                         TURF_ASSERT(probeDelta == 0 || probeDelta == desiredDelta);
 463                         if (probeDelta == 0)
 464                             TURF_TRACE(Grampa, 13, "[insert] set link on behalf of late-arriving cell", uptr(table), idx);
 465 #else
 466                         prevLink->store(desiredDelta, turf::Relaxed);
 467 #endif
 468                         goto followLink;  // Try to follow link chain for the bucket again.
 469                     }
 470                     // Continue linear search...
 471                 }
 472                 // Table is too full to insert.
 473                 overflowIdx = idx + 1;
 474                 TURF_TRACE(Grampa, 14, "[insert] overflow", uptr(table), overflowIdx);
 475                 return InsertResult_Overflow;
 476             }
 477         }
 478     }
 479
 480     static void beginTableMigrationToSize(Map& map, Table* table, ureg nextTableSize, ureg splitShift) {
 481         // Create new migration by DCLI.
 482         TURF_TRACE(Grampa, 15, "[beginTableMigrationToSize] called", 0, 0);
 483         SimpleJobCoordinator::Job* job = table->jobCoordinator.loadConsume();
 484         if (job) {
 485             TURF_TRACE(Grampa, 16, "[beginTableMigrationToSize] new migration already exists", 0, 0);
 486         } else {
 487             turf::LockGuard<junction::striped::Mutex> guard(table->mutex);
 488             job = table->jobCoordinator.loadConsume();  // Non-atomic would be sufficient, but that's OK.
 489             if (job) {
 490                 TURF_TRACE(Grampa, 17, "[beginTableMigrationToSize] new migration already exists (double-checked)", 0, 0);
 491             } else {
 492                 // Create new migration.
 493                 ureg numDestinations = ureg(1) << splitShift;
 494                 TableMigration* migration = TableMigration::create(map, 1, numDestinations);
 495                 migration->m_baseHash = table->baseHash;
 496                 ureg migrationShift = table->unsafeRangeShift - splitShift;
 497                 migration->m_safeShift = (migrationShift < sizeof(Hash) * 8) ? migrationShift : 0;
 498                 migration->m_unitsRemaining.storeNonatomic(table->getNumMigrationUnits());
 499                 migration->getSources()[0].table = table;
 500                 migration->getSources()[0].sourceIndex.storeNonatomic(0);
 501                 ureg subRangeShift = table->unsafeRangeShift - splitShift;      // subRangeShift is also "unsafe" (possibly represents entire range)
 502                 ureg hashOffsetDelta = subRangeShift < (sizeof(Hash) * 8) ? (ureg(1) << subRangeShift) : 0;
 503                 for (ureg i = 0; i < numDestinations; i++) {
 504                     migration->getDestinations()[i] = Table::create(nextTableSize, table->baseHash + hashOffsetDelta * i, subRangeShift);
 505                 }
 506                 // Publish the new migration.
 507                 table->jobCoordinator.storeRelease(migration);
 508             }
 509         }
 510     }
 511
 512     static void beginTableMigration(Map& map, Table* table, ureg overflowIdx) {
 513         // Estimate number of cells in use based on a small sample.
 514         ureg sizeMask = table->sizeMask;
 515         ureg idx = overflowIdx - CellsInUseSample;
 516         ureg inUseCells = 0;
 517         for (ureg linearProbesRemaining = CellsInUseSample; linearProbesRemaining > 0; linearProbesRemaining--) {
 518             CellGroup* group = table->getCellGroups() + ((idx & sizeMask) >> 2);
 519             Cell* cell = group->cells + (idx & 3);
 520             Value value = cell->value.load(turf::Relaxed);
 521             if (value == Value(ValueTraits::Redirect)) {
 522                 // Another thread kicked off the jobCoordinator. The caller will participate upon return.
 523                 TURF_TRACE(Grampa, 18, "[beginTableMigration] redirected while determining table size", 0, 0);
 524                 return;
 525             }
 526             if (value != Value(ValueTraits::NullValue))
 527                 inUseCells++;
 528             idx++;
 529         }
 530         float inUseRatio = float(inUseCells) / CellsInUseSample;
 531         float estimatedInUse = (sizeMask + 1) * inUseRatio;
 532         ureg nextTableSize = turf::util::roundUpPowerOf2(ureg(estimatedInUse * 2));
 533         // FIXME: Support migrating to smaller tables.
 534         nextTableSize = turf::util::max(nextTableSize, sizeMask + 1);
 535         // Split into multiple tables if necessary.
 536         ureg splitShift = 0;
 537         while (nextTableSize > LeafSize) {
 538             splitShift++;
 539             nextTableSize >>= 1;
 540         }
 541         beginTableMigrationToSize(map, table, nextTableSize, splitShift);
 542     }
 543
 544     static FlatTreeMigration* createFlatTreeMigration(Map& map, FlatTree* flatTree, ureg shift) {
 545         turf::LockGuard<junction::striped::Mutex> guard(flatTree->mutex);
 546         if (!flatTree->migration) {
 547             flatTree->migration = new FlatTreeMigration(map, flatTree, shift);
 548         }
 549         return flatTree->migration;
 550     }
 551
 552     static FlatTreeMigration* getExistingFlatTreeMigration(FlatTree* flatTree) {
 553         turf::LockGuard<junction::striped::Mutex> guard(flatTree->mutex);
 554         TURF_ASSERT(flatTree->migration);     // Must already exist!
 555         return flatTree->migration;
 556     }
 557 }; // Grampa
 558
 559 // Return index of the destination table that overflowed, or -1 if none
 560 template<class Map>
 561 sreg Grampa<Map>::TableMigration::migrateRange(Table* srcTable, ureg startIdx) {
 562     ureg srcSizeMask = srcTable->sizeMask;
 563     ureg safeShift = m_safeShift;
 564     Table** dstLeafs = getDestinations();
 565     ureg dstLeafMask = m_numDestinations - 1;
 566     ureg endIdx = turf::util::min(startIdx + TableMigrationUnitSize, srcSizeMask + 1);
 567     // Iterate over source range.
 568     for (ureg srcIdx = startIdx; srcIdx < endIdx; srcIdx++) {
 569         CellGroup* srcGroup = srcTable->getCellGroups() + ((srcIdx & srcSizeMask) >> 2);
 570         Cell* srcCell = srcGroup->cells + (srcIdx & 3);
 571         Hash srcHash;
 572         Value srcValue;
 573         // Fetch the srcHash and srcValue.
 574         for (;;) {
 575             srcHash = srcCell->hash.load(turf::Relaxed);
 576             if (srcHash == KeyTraits::NullHash) {
 577                 // An unused cell. Try to put a Redirect marker in its value.
 578                 srcValue = srcCell->value.compareExchange(Value(ValueTraits::NullValue), Value(ValueTraits::Redirect), turf::Relaxed);
 579                 if (srcValue == Value(ValueTraits::Redirect)) {
 580                     // srcValue is already marked Redirect due to previous incomplete migration.
 581                     TURF_TRACE(Grampa, 19, "[migrateRange] empty cell already redirected", uptr(srcTable), srcIdx);
 582                     break;
 583                 }
 584                 if (srcValue == Value(ValueTraits::NullValue))
 585                     break;  // Redirect has been placed. Break inner loop, continue outer loop.
 586                 TURF_TRACE(Grampa, 20, "[migrateRange] race to insert key", uptr(srcTable), srcIdx);
 587                 // Otherwise, somebody just claimed the cell. Read srcHash again...
 588             } else {
 589                 // Check for deleted/uninitialized value.
 590                 srcValue = srcCell->value.load(turf::Relaxed);
 591                 if (srcValue == Value(ValueTraits::NullValue)) {
 592                     // Try to put a Redirect marker.
 593                     if (srcCell->value.compareExchangeStrong(srcValue, Value(ValueTraits::Redirect), turf::Relaxed))
 594                         break;  // Redirect has been placed. Break inner loop, continue outer loop.
 595                     TURF_TRACE(Grampa, 21, "[migrateRange] race to insert value", uptr(srcTable), srcIdx);
 596                     if (srcValue == Value(ValueTraits::Redirect)) {
 597                         // FIXME: I don't think this will happen. Investigate & change to assert
 598                         TURF_TRACE(Grampa, 22, "[migrateRange] race inserted Redirect", uptr(srcTable), srcIdx);
 599                         break;
 600                     }
 601                 } else if (srcValue == Value(ValueTraits::Redirect)) {
 602                     // srcValue is already marked Redirect due to previous incomplete migration.
 603                     TURF_TRACE(Grampa, 23, "[migrateRange] in-use cell already redirected", uptr(srcTable), srcIdx);
 604                     break;
 605                 }
 606
 607                 // We've got a key/value pair to migrate.
 608                 // Reserve a destination cell in dstTable.
 609                 TURF_ASSERT(srcHash != KeyTraits::NullHash);
 610                 TURF_ASSERT(srcValue != Value(ValueTraits::NullValue));
 611                 TURF_ASSERT(srcValue != Value(ValueTraits::Redirect));
 612                 ureg destLeafIndex = (srcHash >> safeShift) & dstLeafMask;
 613                 Table* dstLeaf = dstLeafs[destLeafIndex];
 614                 Cell* dstCell;
 615                 ureg overflowIdx;
 616                 InsertResult result = insert(srcHash, dstLeaf, dstLeaf->sizeMask, dstCell, overflowIdx);
 617                 // During migration, a hash can only exist in one place among all the source tables,
 618                 // and it is only migrated by one thread. Therefore, the hash will never already exist
 619                 // in the destination table:
 620                 TURF_ASSERT(result != InsertResult_AlreadyFound);
 621                 if (result == InsertResult_Overflow) {
 622                     // Destination overflow.
 623                     // This can happen for several reasons. For example, the source table could have
 624                     // existed of all deleted cells when it overflowed, resulting in a small destination
 625                     // table size, but then another thread could re-insert all the same hashes
 626                     // before the migration completed.
 627                     // Caller will cancel the current migration and begin a new one.
 628                     return destLeafIndex;
 629                 }
 630                 // Migrate the old value to the new cell.
 631                 for (;;) {
 632                     // Copy srcValue to the destination.
 633                     dstCell->value.store(srcValue, turf::Relaxed);
 634                     // Try to place a Redirect marker in srcValue.
 635                     Value doubleCheckedSrcValue = srcCell->value.compareExchange(srcValue, Value(ValueTraits::Redirect), turf::Relaxed);
 636                     TURF_ASSERT(doubleCheckedSrcValue != Value(ValueTraits::Redirect)); // Only one thread can redirect a cell at a time.
 637                     if (doubleCheckedSrcValue == srcValue) {
 638                         // No racing writes to the src. We've successfully placed the Redirect marker.
 639                         // srcValue was non-NULL when we decided to migrate it, but it may have changed to NULL
 640                         // by a late-arriving erase.
 641                         if (srcValue == Value(ValueTraits::NullValue))
 642                             TURF_TRACE(Grampa, 24, "[migrateRange] racing update was erase", uptr(srcTable), srcIdx);
 643                         break;
 644                     }
 645                     // There was a late-arriving write (or erase) to the src. Migrate the new value and try again.
 646                     TURF_TRACE(Grampa, 25, "[migrateRange] race to update migrated value", uptr(srcTable), srcIdx);
 647                     srcValue = doubleCheckedSrcValue;
 648                 }
 649                 // Cell successfully migrated. Proceed to next source cell.
 650                 break;
 651             }
 652         }
 653     }
 654     // Range has been migrated successfully.
 655     return -1;
 656 }
 657
 658 template <class Map>
 659 void Grampa<Map>::TableMigration::run() {
 660     // Conditionally increment the shared # of workers.
 661     ureg probeStatus = m_workerStatus.load(turf::Relaxed);
 662     do {
 663         if (probeStatus & 1) {
 664             // End flag is already set, so do nothing.
 665             TURF_TRACE(Grampa, 26, "[TableMigration::run] already ended", uptr(this), 0);
 666             return;
 667         }
 668     } while (!m_workerStatus.compareExchangeWeak(probeStatus, probeStatus + 2, turf::Relaxed, turf::Relaxed));
 669     // # of workers has been incremented, and the end flag is clear.
 670     TURF_ASSERT((probeStatus & 1) == 0);
 671
 672     // Iterate over all source tables.
 673     Source* sources = getSources();
 674     for (ureg s = 0; s < m_numSources; s++) {
 675         Source& source = sources[s];
 676         // Loop over all migration units in this source table.
 677         for (;;) {
 678             if (m_workerStatus.load(turf::Relaxed) & 1) {
 679                 TURF_TRACE(Grampa, 27, "[TableMigration::run] detected end flag set", uptr(this), 0);
 680                 goto endMigration;
 681             }
 682             ureg startIdx = source.sourceIndex.fetchAdd(TableMigrationUnitSize, turf::Relaxed);
 683             if (startIdx >= source.table->sizeMask + 1)
 684                 break;   // No more migration units in this table. Try next source table.
 685             sreg overflowTableIndex = migrateRange(source.table, startIdx);
 686             if  (overflowTableIndex >= 0) {
 687                 // *** FAILED MIGRATION ***
 688                 // TableMigration failed due to destination table overflow.
 689                 // No other thread can declare the migration successful at this point, because *this* unit will never complete, hence m_unitsRemaining won't reach zero.
 690                 // However, multiple threads can independently detect a failed migration at the same time.
 691                 TURF_TRACE(Grampa, 28, "[TableMigration::run] destination overflow", uptr(source.table), uptr(startIdx));
 692                 // The reason we store overflowTableIndex in a shared variable is because we must flush all the worker threads before
 693                 // we can safely deal with the overflow. Therefore, the thread that detects the failure is often different from the thread
 694                 // that deals with it.
 695                 // Store overflowTableIndex unconditionally; racing writes should be rare, and it doesn't matter which one wins.
 696                 sreg oldIndex = m_overflowTableIndex.exchange(overflowTableIndex, turf::Relaxed);
 697                 if (oldIndex >= 0)
 698                     TURF_TRACE(Grampa, 29, "[TableMigration::run] race to set m_overflowTableIndex", uptr(overflowTableIndex), uptr(oldIndex));
 699                 m_workerStatus.fetchOr(1, turf::Relaxed);
 700                 goto endMigration;
 701             }
 702             sreg prevRemaining = m_unitsRemaining.fetchSub(1, turf::Relaxed);
 703             TURF_ASSERT(prevRemaining > 0);
 704             if (prevRemaining == 1) {
 705                 // *** SUCCESSFUL MIGRATION ***
 706                 // That was the last chunk to migrate.
 707                 m_workerStatus.fetchOr(1, turf::Relaxed);
 708                 goto endMigration;
 709             }
 710         }
 711     }
 712     TURF_TRACE(Grampa, 30, "[TableMigration::run] out of migration units", uptr(this), 0);
 713
 714 endMigration:
 715     // Decrement the shared # of workers.
 716     probeStatus = m_workerStatus.fetchSub(2, turf::AcquireRelease);  // Ensure all modifications are visible to the thread that will publish
 717     if (probeStatus >= 4) {
 718         // There are other workers remaining. Return here so that only the very last worker will proceed.
 719         TURF_TRACE(Grampa, 31, "[TableMigration::run] not the last worker", uptr(this), uptr(probeStatus));
 720         return;
 721     }
 722
 723     // We're the very last worker thread.
 724     // Perform the appropriate post-migration step depending on whether the migration succeeded or failed.
 725     TURF_ASSERT(probeStatus == 3);
 726     sreg overflowTableIndex = m_overflowTableIndex.loadNonatomic();  // No racing writes at this point
 727     if (overflowTableIndex < 0) {
 728         // The migration succeeded. This is the most likely outcome. Publish the new subtree.
 729         m_map.publishTableMigration(this);
 730         // End the jobCoodinator.
 731         sources[0].table->jobCoordinator.end();
 732     } else {
 733         // The migration failed due to the overflow of a destination table.
 734         Table* origTable = sources[0].table;
 735         ureg count = ureg(1) << (origTable->unsafeRangeShift - getUnsafeShift());
 736         ureg lo = overflowTableIndex & ~(count - 1);
 737         TURF_ASSERT(lo + count <= m_numDestinations);
 738         turf::LockGuard<junction::striped::Mutex> guard(origTable->mutex);
 739         SimpleJobCoordinator::Job* checkedJob = origTable->jobCoordinator.loadConsume();
 740         if (checkedJob != this) {
 741             TURF_TRACE(Grampa, 32, "[TableMigration::run] a new TableMigration was already started", uptr(origTable), uptr(checkedJob));
 742         } else {
 743             TableMigration* migration;
 744             Table* overflowedTable = getDestinations()[overflowTableIndex];
 745             if (overflowedTable->sizeMask + 1 < LeafSize) {
 746                 // The entire map is contained in a small table.
 747                 TURF_TRACE(Grampa, 33, "[TableMigration::run] overflow occured in a small map", uptr(origTable), uptr(checkedJob));
 748                 TURF_ASSERT(overflowedTable->unsafeRangeShift == sizeof(Hash) * 8);
 749                 TURF_ASSERT(overflowedTable->baseHash == 0);
 750                 TURF_ASSERT(m_numDestinations == 1);
 751                 TURF_ASSERT(m_baseHash == 0);
 752                 migration = TableMigration::create(m_map, m_numSources + 1, 1);
 753                 migration->m_baseHash = 0;
 754                 migration->m_safeShift = 0;
 755                 // Double the destination table size.
 756                 migration->getDestinations()[0] = Table::create((overflowedTable->sizeMask + 1) * 2, overflowedTable->baseHash, overflowedTable->unsafeRangeShift);
 757             } else {
 758                 // The overflowed table is already the size of a leaf. Split it into two ranges.
 759                 if (count == 1) {
 760                     TURF_TRACE(Grampa, 34, "[TableMigration::run] doubling subtree size after failure", uptr(origTable), uptr(checkedJob));
 761                     migration = TableMigration::create(m_map, m_numSources + 1, m_numDestinations * 2);
 762                     migration->m_baseHash = m_baseHash;
 763                     migration->m_safeShift = getUnsafeShift() - 1;
 764                     for (ureg i = 0; i < m_numDestinations; i++) {
 765                         migration->getDestinations()[i * 2] = getDestinations()[i];
 766                         migration->getDestinations()[i * 2 + 1] = getDestinations()[i];
 767                     }
 768                     count = 2;
 769                 } else {
 770                     TURF_TRACE(Grampa, 35, "[TableMigration::run] keeping same subtree size after failure", uptr(origTable), uptr(checkedJob));
 771                     migration = TableMigration::create(m_map, m_numSources + 1, m_numDestinations);
 772                     migration->m_baseHash = m_baseHash;
 773                     migration->m_safeShift = m_safeShift;
 774                     memcpy(migration->getDestinations(), getDestinations(), m_numDestinations * sizeof(Table*));
 775                 }
 776                 Table* splitTable1 = Table::create(LeafSize, origTable->baseHash, origTable->unsafeRangeShift - 1);
 777                 ureg i = 0;
 778                 for (; i < count / 2; i++) {
 779                     migration->getDestinations()[lo + i] = splitTable1;
 780                 }
 781                 ureg halfNumHashes = ureg(1) << (origTable->unsafeRangeShift - 1);
 782                 Table* splitTable2 = Table::create(LeafSize, origTable->baseHash + halfNumHashes, origTable->unsafeRangeShift - 1);
 783                 for (; i < count; i++) {
 784                     migration->getDestinations()[lo + i] = splitTable2;
 785                 }
 786             }
 787             // Transfer source tables to the new migration.
 788             for (ureg i = 0; i < m_numSources; i++) {
 789                 migration->getSources()[i].table = getSources()[i].table;
 790                 migration->getSources()[i].sourceIndex.storeNonatomic(0);
 791                 getSources()[i].table = NULL;
 792             }
 793             migration->getSources()[m_numSources].table = overflowedTable;
 794             migration->getSources()[m_numSources].sourceIndex.storeNonatomic(0);
 795             // Calculate total number of migration units to move.
 796             ureg unitsRemaining = 0;
 797             for (ureg s = 0; s < migration->m_numSources; s++)
 798                 unitsRemaining += migration->getSources()[s].table->getNumMigrationUnits();
 799             migration->m_unitsRemaining.storeNonatomic(unitsRemaining);
 800             // Publish the new migration.
 801             origTable->jobCoordinator.storeRelease(migration);
 802         }
 803     }
 804
 805     // We're done with this TableMigration. Queue it for GC.
 806     DefaultQSBR.enqueue(&TableMigration::destroy, this);
 807 }
 808
 809 template<class Map>
 810 void Grampa<Map>::FlatTreeMigration::run() {
 811     // Conditionally increment the shared # of workers.
 812     ureg probeStatus = m_workerStatus.load(turf::Relaxed);
 813     do {
 814         if (probeStatus & 1) {
 815             // End flag is already set, so do nothing.
 816             TURF_TRACE(Grampa, 36, "[FlatTreeMigration::run] already ended", uptr(this), 0);
 817             return;
 818         }
 819     } while (!m_workerStatus.compareExchangeWeak(probeStatus, probeStatus + 2, turf::Relaxed, turf::Relaxed));
 820     // # of workers has been incremented, and the end flag is clear.
 821     TURF_ASSERT((probeStatus & 1) == 0);
 822
 823     // Loop over all migration units
 824     ureg srcSize = (Hash(-1) >> m_source->safeShift) + 1;
 825     // FIXME: Support migration to smaller flattrees
 826     TURF_ASSERT(m_destination->safeShift < m_source->safeShift);
 827     ureg repeat = ureg(1) << (m_source->safeShift - m_destination->safeShift);
 828     for (;;) {
 829         ureg srcStart = m_sourceIndex.fetchAdd(FlatTreeMigrationUnitSize, turf::Relaxed);
 830         if (srcStart >= srcSize)
 831             break;      // No more migration units in this flattree.
 832         // Migrate this range
 833         ureg srcEnd = turf::util::min(srcSize, srcStart + FlatTreeMigrationUnitSize);
 834         ureg dst = srcStart * repeat;
 835         for (ureg src = srcStart; src < srcEnd; src++) {
 836             // Pointers in the source table can be changed at any time due to concurrent subtree publishing,
 837             // so we need to exchange them with Redirect markers.
 838             Table* t = m_source->getTables()[src].exchange((Table*) RedirectFlatTree, turf::Relaxed);
 839             TURF_ASSERT(uptr(t) != RedirectFlatTree);
 840             for (ureg r = repeat; r > 0; r--) {
 841                 m_destination->getTables()[dst].storeNonatomic(t);
 842                 dst++;
 843             }
 844         }
 845         // Decrement m_unitsRemaining
 846         sreg prevRemaining = m_unitsRemaining.fetchSub(1, turf::Relaxed);
 847         if (prevRemaining == 1) {
 848             // *** SUCCESSFUL MIGRATION ***
 849             // That was the last chunk to migrate.
 850             m_workerStatus.fetchOr(1, turf::Relaxed);
 851             break;
 852         }
 853     }
 854
 855     // Decrement the shared # of workers.
 856     probeStatus = m_workerStatus.fetchSub(2, turf::AcquireRelease);    // AcquireRelease makes all previous writes visible to the last worker thread.
 857     if (probeStatus >= 4) {
 858         // There are other workers remaining. Return here so that only the very last worker will proceed.
 859         return;
 860     }
 861
 862     // We're the very last worker thread.
 863     // Publish the new flattree.
 864     TURF_ASSERT(probeStatus == 3);      // End flag must be set
 865     m_map.publishFlatTreeMigration(this);
 866     m_completed.signal();
 867
 868     // We're done with this FlatTreeMigration. Queue it for GC.
 869     DefaultQSBR.enqueue(&FlatTreeMigration::destroy, this);
 870 }
 871
 872 } // namespace details
 873 } // namespace junction
 874
 875 #endif // JUNCTION_DETAILS_GRAMPA_H