summaryrefslogtreecommitdiff
path: root/rts/sm/GC.c
blob: 0544fd29da9becd9dc62e5783552a914b6112528 (plain)
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
946
947
948
949
950
951
952
953
954
955
956
957
958
959
960
961
962
963
964
965
966
967
968
969
970
971
972
973
974
975
976
977
978
979
980
981
982
983
984
985
986
987
988
989
990
991
992
993
994
995
996
997
998
999
1000
1001
1002
1003
1004
1005
1006
1007
1008
1009
1010
1011
1012
1013
1014
1015
1016
1017
1018
1019
1020
1021
1022
1023
1024
1025
1026
1027
1028
1029
1030
1031
1032
1033
1034
1035
1036
1037
1038
1039
1040
1041
1042
1043
1044
1045
1046
1047
1048
1049
1050
1051
1052
1053
1054
1055
1056
1057
1058
1059
1060
1061
1062
1063
1064
1065
1066
1067
1068
1069
1070
1071
1072
1073
1074
1075
1076
1077
1078
1079
1080
1081
1082
1083
1084
1085
1086
1087
1088
1089
1090
1091
1092
1093
1094
1095
1096
1097
1098
1099
1100
1101
1102
1103
1104
1105
1106
1107
1108
1109
1110
1111
1112
1113
1114
1115
1116
1117
1118
1119
1120
1121
1122
1123
1124
1125
1126
1127
1128
1129
1130
1131
1132
1133
1134
1135
1136
1137
1138
1139
1140
1141
1142
1143
1144
1145
1146
1147
1148
1149
1150
1151
1152
1153
1154
1155
1156
1157
1158
1159
1160
1161
1162
1163
1164
1165
1166
1167
1168
1169
1170
1171
1172
1173
1174
1175
1176
1177
1178
1179
1180
1181
1182
1183
1184
1185
1186
1187
1188
1189
1190
1191
1192
1193
1194
1195
1196
1197
1198
1199
1200
1201
1202
1203
1204
1205
1206
1207
1208
1209
1210
1211
1212
1213
1214
1215
1216
1217
1218
1219
1220
1221
1222
1223
1224
1225
1226
1227
1228
1229
1230
1231
1232
1233
1234
1235
1236
1237
1238
1239
1240
1241
1242
1243
1244
1245
1246
1247
1248
1249
1250
1251
1252
1253
1254
1255
1256
1257
1258
1259
1260
1261
1262
1263
1264
1265
1266
1267
1268
1269
1270
1271
1272
1273
1274
1275
1276
1277
1278
1279
1280
1281
1282
1283
1284
1285
1286
1287
1288
1289
1290
1291
1292
1293
1294
1295
1296
1297
1298
1299
1300
1301
1302
1303
1304
1305
1306
1307
1308
1309
1310
1311
1312
1313
1314
1315
1316
1317
1318
1319
1320
1321
1322
1323
1324
1325
1326
1327
1328
1329
1330
1331
1332
1333
1334
1335
1336
1337
1338
1339
1340
1341
1342
1343
1344
1345
1346
1347
1348
1349
1350
1351
1352
1353
1354
1355
1356
1357
1358
1359
1360
1361
1362
1363
1364
1365
1366
1367
1368
1369
1370
1371
1372
1373
1374
1375
1376
1377
1378
1379
1380
1381
1382
1383
1384
1385
1386
1387
1388
1389
1390
1391
1392
1393
1394
1395
1396
1397
1398
1399
1400
1401
1402
1403
1404
1405
1406
1407
1408
1409
1410
1411
1412
1413
1414
1415
1416
1417
1418
1419
1420
1421
1422
1423
1424
1425
1426
1427
1428
1429
1430
1431
1432
1433
1434
1435
1436
1437
1438
1439
1440
1441
1442
1443
1444
1445
1446
1447
1448
1449
1450
1451
1452
1453
1454
1455
1456
1457
1458
1459
1460
1461
1462
1463
1464
1465
1466
1467
1468
1469
1470
1471
1472
1473
1474
1475
1476
1477
1478
1479
1480
1481
1482
1483
1484
1485
1486
1487
1488
1489
1490
1491
1492
1493
1494
1495
1496
1497
1498
1499
1500
1501
1502
1503
1504
1505
1506
1507
1508
1509
1510
1511
1512
1513
1514
1515
1516
1517
1518
1519
1520
1521
1522
1523
1524
1525
1526
1527
1528
1529
1530
1531
1532
1533
1534
1535
1536
1537
1538
1539
1540
1541
1542
1543
1544
1545
1546
1547
1548
1549
1550
1551
1552
1553
1554
1555
1556
1557
1558
1559
1560
1561
1562
1563
1564
1565
1566
1567
1568
1569
1570
1571
1572
1573
1574
1575
1576
1577
1578
1579
1580
1581
1582
1583
1584
1585
1586
1587
1588
1589
1590
1591
1592
1593
1594
1595
1596
1597
1598
1599
1600
1601
1602
1603
1604
1605
1606
1607
1608
1609
1610
1611
1612
1613
1614
1615
1616
1617
1618
1619
1620
1621
1622
1623
1624
1625
1626
1627
1628
1629
1630
1631
1632
1633
1634
1635
1636
1637
1638
1639
1640
1641
1642
1643
1644
1645
1646
1647
1648
1649
1650
1651
1652
1653
1654
1655
1656
1657
1658
1659
1660
1661
1662
1663
1664
1665
1666
1667
1668
1669
1670
1671
1672
1673
1674
1675
1676
1677
1678
1679
1680
1681
1682
1683
1684
1685
1686
1687
1688
1689
1690
1691
1692
1693
1694
1695
1696
1697
1698
1699
1700
1701
1702
1703
1704
1705
1706
1707
1708
1709
1710
1711
1712
1713
1714
1715
1716
1717
1718
1719
1720
1721
1722
1723
1724
1725
1726
1727
1728
1729
1730
1731
1732
1733
1734
1735
1736
1737
1738
1739
1740
1741
1742
1743
1744
1745
1746
1747
1748
1749
1750
1751
1752
1753
1754
1755
1756
1757
1758
1759
1760
1761
1762
1763
1764
1765
1766
1767
1768
1769
1770
1771
1772
1773
1774
1775
1776
1777
1778
1779
1780
1781
1782
1783
1784
1785
1786
1787
1788
1789
1790
1791
1792
1793
1794
1795
1796
1797
1798
1799
1800
1801
1802
1803
1804
1805
1806
1807
1808
1809
1810
1811
1812
1813
1814
1815
1816
1817
1818
1819
1820
1821
1822
1823
1824
1825
1826
1827
1828
1829
1830
1831
1832
1833
1834
1835
1836
1837
1838
1839
1840
1841
1842
1843
1844
1845
1846
1847
1848
1849
1850
1851
1852
1853
1854
1855
1856
1857
1858
1859
1860
1861
1862
1863
1864
1865
1866
1867
1868
1869
1870
1871
1872
1873
1874
1875
1876
1877
1878
1879
1880
1881
1882
1883
1884
1885
1886
1887
1888
1889
1890
1891
1892
1893
1894
1895
1896
1897
1898
1899
1900
1901
1902
1903
1904
1905
1906
1907
/* -----------------------------------------------------------------------------
 *
 * (c) The GHC Team 1998-2008
 *
 * Generational garbage collector
 *
 * Documentation on the architecture of the Garbage Collector can be
 * found in the online commentary:
 *
 *   http://ghc.haskell.org/trac/ghc/wiki/Commentary/Rts/Storage/GC
 *
 * ---------------------------------------------------------------------------*/

#include "PosixSource.h"
#include "Rts.h"
#include "HsFFI.h"

#include "GC.h"
#include "GCThread.h"
#include "GCTDecl.h"            // NB. before RtsSignals.h which
                                // clobbers REG_R1 on arm/Linux
#include "Compact.h"
#include "Evac.h"
#include "Scav.h"
#include "GCUtils.h"
#include "MarkStack.h"
#include "MarkWeak.h"
#include "Sparks.h"
#include "Sweep.h"

#include "Arena.h"
#include "Storage.h"
#include "RtsUtils.h"
#include "Apply.h"
#include "Updates.h"
#include "Stats.h"
#include "Schedule.h"
#include "Sanity.h"
#include "BlockAlloc.h"
#include "ProfHeap.h"
#include "Weak.h"
#include "Prelude.h"
#include "RtsSignals.h"
#include "STM.h"
#include "Trace.h"
#include "RetainerProfile.h"
#include "LdvProfile.h"
#include "RaiseAsync.h"
#include "StableName.h"
#include "StablePtr.h"
#include "CheckUnload.h"
#include "CNF.h"
#include "RtsFlags.h"

#if defined(PROFILING)
#include "RetainerProfile.h"
#endif

#include <string.h> // for memset()
#include <unistd.h>

/* -----------------------------------------------------------------------------
   Global variables
   -------------------------------------------------------------------------- */

/* STATIC OBJECT LIST.
 *
 * During GC:
 * We maintain a linked list of static objects that are still live.
 * The requirements for this list are:
 *
 *  - we need to scan the list while adding to it, in order to
 *    scavenge all the static objects (in the same way that
 *    breadth-first scavenging works for dynamic objects).
 *
 *  - we need to be able to tell whether an object is already on
 *    the list, to break loops.
 *
 * Each static object has a "static link field", which we use for
 * linking objects on to the list.  We use a stack-type list, consing
 * objects on the front as they are added (this means that the
 * scavenge phase is depth-first, not breadth-first, but that
 * shouldn't matter).
 *
 * A separate list is kept for objects that have been scavenged
 * already - this is so that we can zero all the marks afterwards.
 *
 * An object is on the list if its static link field is non-zero; this
 * means that we have to mark the end of the list with '1', not NULL.
 *
 * Extra notes for generational GC:
 *
 * Each generation has a static object list associated with it.  When
 * collecting generations up to N, we treat the static object lists
 * from generations > N as roots.
 *
 * We build up a static object list while collecting generations 0..N,
 * which is then appended to the static object list of generation N+1.
 *
 * See also: Note [STATIC_LINK fields] in Storage.h.
 */

/* N is the oldest generation being collected, where the generations
 * are numbered starting at 0.  A major GC (indicated by the major_gc
 * flag) is when we're collecting all generations.  We only attempt to
 * deal with static objects and GC CAFs when doing a major GC.
 */
uint32_t N;
bool major_gc;

/* Data used for allocation area sizing.
 */
static W_ g0_pcnt_kept = 30; // percentage of g0 live at last minor GC

/* Mut-list stats */
#if defined(DEBUG)
uint32_t mutlist_MUTVARS,
    mutlist_MUTARRS,
    mutlist_MVARS,
    mutlist_TVAR,
    mutlist_TVAR_WATCH_QUEUE,
    mutlist_TREC_CHUNK,
    mutlist_TREC_HEADER,
    mutlist_OTHERS;
#endif

/* Thread-local data for each GC thread
 */
gc_thread **gc_threads = NULL;

#if !defined(THREADED_RTS)
// Must be aligned to 64-bytes to meet stated 64-byte alignment of gen_workspace
StgWord8 the_gc_thread[sizeof(gc_thread) + 64 * sizeof(gen_workspace)]
    ATTRIBUTE_ALIGNED(64);
#endif

// Number of threads running in *this* GC.  Affects how many
// step->todos[] lists we have to look in to find work.
uint32_t n_gc_threads;

// For stats:
static long copied;        // *words* copied & scavenged during this GC

#if defined(PROF_SPIN) && defined(THREADED_RTS)
// spin and yield counts for the quasi-SpinLock in waitForGcThreads
volatile StgWord64 waitForGcThreads_spin = 0;
volatile StgWord64 waitForGcThreads_yield = 0;
volatile StgWord64 whitehole_gc_spin = 0;
#endif

bool work_stealing;

uint32_t static_flag = STATIC_FLAG_B;
uint32_t prev_static_flag = STATIC_FLAG_A;

DECLARE_GCT

/* -----------------------------------------------------------------------------
   Static function declarations
   -------------------------------------------------------------------------- */

static void mark_root               (void *user, StgClosure **root);
static void prepare_collected_gen   (generation *gen);
static void prepare_uncollected_gen (generation *gen);
static void init_gc_thread          (gc_thread *t);
static void resize_generations      (void);
static void resize_nursery          (void);
static void start_gc_threads        (void);
static void scavenge_until_all_done (void);
static StgWord inc_running          (void);
static StgWord dec_running          (void);
static void wakeup_gc_threads       (uint32_t me, bool idle_cap[]);
static void shutdown_gc_threads     (uint32_t me, bool idle_cap[]);
static void collect_gct_blocks      (void);
static void collect_pinned_object_blocks (void);
static void heapOverflow            (void);

#if defined(DEBUG)
static void gcCAFs                  (void);
#endif

/* -----------------------------------------------------------------------------
   The mark stack.
   -------------------------------------------------------------------------- */

bdescr *mark_stack_top_bd; // topmost block in the mark stack
bdescr *mark_stack_bd;     // current block in the mark stack
StgPtr mark_sp;            // pointer to the next unallocated mark stack entry

/* -----------------------------------------------------------------------------
   GarbageCollect: the main entry point to the garbage collector.

   The collect_gen parameter is gotten by calling calcNeeded().

   Locks held: all capabilities are held throughout GarbageCollect().
   -------------------------------------------------------------------------- */

void
GarbageCollect (uint32_t collect_gen,
                bool do_heap_census,
                uint32_t gc_type USED_IF_THREADS,
                Capability *cap,
                bool idle_cap[])
{
  bdescr *bd;
  generation *gen;
  StgWord live_blocks, live_words, par_max_copied, par_balanced_copied,
      gc_spin_spin, gc_spin_yield, mut_spin_spin, mut_spin_yield,
      any_work, no_work, scav_find_work;
#if defined(THREADED_RTS)
  gc_thread *saved_gct;
#endif
  uint32_t g, n;

  // necessary if we stole a callee-saves register for gct:
#if defined(THREADED_RTS)
  saved_gct = gct;
#endif

#if defined(PROFILING)
  CostCentreStack *save_CCS[n_capabilities];
#endif

  ACQUIRE_SM_LOCK;

#if defined(RTS_USER_SIGNALS)
  if (RtsFlags.MiscFlags.install_signal_handlers) {
    // block signals
    blockUserSignals();
  }
#endif

  ASSERT(sizeof(gen_workspace) == 16 * sizeof(StgWord));
  // otherwise adjust the padding in gen_workspace.

  // this is the main thread
  SET_GCT(gc_threads[cap->no]);

  // tell the stats department that we've started a GC
  stat_startGC(cap, gct);

  // Lock the StablePtr table. This prevents FFI calls manipulating
  // the table from occurring during GC.
  stablePtrLock();

#if defined(DEBUG)
  mutlist_MUTVARS = 0;
  mutlist_MUTARRS = 0;
  mutlist_MVARS = 0;
  mutlist_TVAR = 0;
  mutlist_TVAR_WATCH_QUEUE = 0;
  mutlist_TREC_CHUNK = 0;
  mutlist_TREC_HEADER = 0;
  mutlist_OTHERS = 0;
#endif

  // attribute any costs to CCS_GC
#if defined(PROFILING)
  for (n = 0; n < n_capabilities; n++) {
      save_CCS[n] = capabilities[n]->r.rCCCS;
      capabilities[n]->r.rCCCS = CCS_GC;
  }
#endif

  /* Figure out which generation to collect
   */
  N = collect_gen;
  major_gc = (N == RtsFlags.GcFlags.generations-1);

  if (major_gc) {
      prev_static_flag = static_flag;
      static_flag =
          static_flag == STATIC_FLAG_A ? STATIC_FLAG_B : STATIC_FLAG_A;
  }

#if defined(THREADED_RTS)
  work_stealing = RtsFlags.ParFlags.parGcLoadBalancingEnabled &&
                  N >= RtsFlags.ParFlags.parGcLoadBalancingGen;
      // It's not always a good idea to do load balancing in parallel
      // GC.  In particular, for a parallel program we don't want to
      // lose locality by moving cached data into another CPU's cache
      // (this effect can be quite significant).
      //
      // We could have a more complex way to deterimine whether to do
      // work stealing or not, e.g. it might be a good idea to do it
      // if the heap is big.  For now, we just turn it on or off with
      // a flag.
#endif

  /* Start threads, so they can be spinning up while we finish initialisation.
   */
  start_gc_threads();

#if defined(THREADED_RTS)
  /* How many threads will be participating in this GC?
   * We don't try to parallelise minor GCs (unless the user asks for
   * it with +RTS -gn0), or mark/compact/sweep GC.
   */
  if (gc_type == SYNC_GC_PAR) {
      n_gc_threads = n_capabilities;
  } else {
      n_gc_threads = 1;
  }
#else
  n_gc_threads = 1;
#endif

  debugTrace(DEBUG_gc, "GC (gen %d, using %d thread(s))",
             N, n_gc_threads);

#if defined(DEBUG)
  // check for memory leaks if DEBUG is on
  memInventory(DEBUG_gc);
#endif

  // do this *before* we start scavenging
  collectFreshWeakPtrs();

  // check sanity *before* GC
  IF_DEBUG(sanity, checkSanity(false /* before GC */, major_gc));

  // gather blocks allocated using allocatePinned() from each capability
  // and put them on the g0->large_object list.
  collect_pinned_object_blocks();

  // Initialise all the generations that we're collecting.
  for (g = 0; g <= N; g++) {
      prepare_collected_gen(&generations[g]);
  }
  // Initialise all the generations that we're *not* collecting.
  for (g = N+1; g < RtsFlags.GcFlags.generations; g++) {
      prepare_uncollected_gen(&generations[g]);
  }

  // Prepare this gc_thread
  init_gc_thread(gct);

  /* Allocate a mark stack if we're doing a major collection.
   */
  if (major_gc && oldest_gen->mark) {
      mark_stack_bd     = allocBlock();
      mark_stack_top_bd = mark_stack_bd;
      mark_stack_bd->link = NULL;
      mark_stack_bd->u.back = NULL;
      mark_sp           = mark_stack_bd->start;
  } else {
      mark_stack_bd     = NULL;
      mark_stack_top_bd = NULL;
      mark_sp           = NULL;
  }

  /* -----------------------------------------------------------------------
   * follow all the roots that we know about:
   */

  // the main thread is running: this prevents any other threads from
  // exiting prematurely, so we can start them now.
  // NB. do this after the mutable lists have been saved above, otherwise
  // the other GC threads will be writing into the old mutable lists.
  inc_running();
  wakeup_gc_threads(gct->thread_index, idle_cap);

  traceEventGcWork(gct->cap);

  // scavenge the capability-private mutable lists.  This isn't part
  // of markSomeCapabilities() because markSomeCapabilities() can only
  // call back into the GC via mark_root() (due to the gct register
  // variable).
  if (n_gc_threads == 1) {
      for (n = 0; n < n_capabilities; n++) {
#if defined(THREADED_RTS)
          scavenge_capability_mut_Lists1(capabilities[n]);
#else
          scavenge_capability_mut_lists(capabilities[n]);
#endif
      }
  } else {
      scavenge_capability_mut_lists(gct->cap);
      for (n = 0; n < n_capabilities; n++) {
          if (idle_cap[n]) {
              markCapability(mark_root, gct, capabilities[n],
                             true/*don't mark sparks*/);
              scavenge_capability_mut_lists(capabilities[n]);
          }
      }
  }

  // follow roots from the CAF list (used by GHCi)
  gct->evac_gen_no = 0;
  markCAFs(mark_root, gct);

  // follow all the roots that the application knows about.
  gct->evac_gen_no = 0;
  if (n_gc_threads == 1) {
      for (n = 0; n < n_capabilities; n++) {
          markCapability(mark_root, gct, capabilities[n],
                         true/*don't mark sparks*/);
      }
  } else {
      markCapability(mark_root, gct, cap, true/*don't mark sparks*/);
  }

  markScheduler(mark_root, gct);

  // Mark the weak pointer list, and prepare to detect dead weak pointers.
  markWeakPtrList();
  initWeakForGC();

  // Mark the stable pointer table.
  markStablePtrTable(mark_root, gct);

  // Remember old stable name addresses.
  rememberOldStableNameAddresses ();

  /* -------------------------------------------------------------------------
   * Repeatedly scavenge all the areas we know about until there's no
   * more scavenging to be done.
   */

  StgWeak *dead_weak_ptr_list = NULL;
  StgTSO *resurrected_threads = END_TSO_QUEUE;

  for (;;)
  {
      scavenge_until_all_done();

      // The other threads are now stopped.  We might recurse back to
      // here, but from now on this is the only thread.

      // must be last...  invariant is that everything is fully
      // scavenged at this point.
      if (traverseWeakPtrList(&dead_weak_ptr_list, &resurrected_threads)) { // returns true if evaced something
          inc_running();
          continue;
      }

      // If we get to here, there's really nothing left to do.
      break;
  }

  shutdown_gc_threads(gct->thread_index, idle_cap);

  // Now see which stable names are still alive.
  gcStableNameTable();

#if defined(THREADED_RTS)
  if (n_gc_threads == 1) {
      for (n = 0; n < n_capabilities; n++) {
          pruneSparkQueue(capabilities[n]);
      }
  } else {
      for (n = 0; n < n_capabilities; n++) {
          if (n == cap->no || idle_cap[n]) {
              pruneSparkQueue(capabilities[n]);
         }
      }
  }
#endif

#if defined(PROFILING)
  // We call processHeapClosureForDead() on every closure destroyed during
  // the current garbage collection, so we invoke LdvCensusForDead().
  if (RtsFlags.ProfFlags.doHeapProfile == HEAP_BY_LDV
      || RtsFlags.ProfFlags.bioSelector != NULL) {
      RELEASE_SM_LOCK; // LdvCensusForDead may need to take the lock
      LdvCensusForDead(N);
      ACQUIRE_SM_LOCK;
  }
#endif

  // NO MORE EVACUATION AFTER THIS POINT!

  // Finally: compact or sweep the oldest generation.
  if (major_gc && oldest_gen->mark) {
      if (oldest_gen->compact)
          compact(gct->scavenged_static_objects,
                  &dead_weak_ptr_list,
                  &resurrected_threads);
      else
          sweep(oldest_gen);
  }

  copied = 0;
  par_max_copied = 0;
  par_balanced_copied = 0;
  gc_spin_spin = 0;
  gc_spin_yield = 0;
  mut_spin_spin = 0;
  mut_spin_yield = 0;
  any_work = 0;
  no_work = 0;
  scav_find_work = 0;
  {
      uint32_t i;
      uint64_t par_balanced_copied_acc = 0;
      const gc_thread* thread;

      for (i=0; i < n_gc_threads; i++) {
          copied += gc_threads[i]->copied;
      }
      for (i=0; i < n_gc_threads; i++) {
          thread = gc_threads[i];
          if (n_gc_threads > 1) {
              debugTrace(DEBUG_gc,"thread %d:", i);
              debugTrace(DEBUG_gc,"   copied           %ld",
                         thread->copied * sizeof(W_));
              debugTrace(DEBUG_gc,"   scanned          %ld",
                         thread->scanned * sizeof(W_));
              debugTrace(DEBUG_gc,"   any_work         %ld",
                         thread->any_work);
              debugTrace(DEBUG_gc,"   no_work          %ld",
                         thread->no_work);
              debugTrace(DEBUG_gc,"   scav_find_work %ld",
                         thread->scav_find_work);

#if defined(THREADED_RTS) && defined(PROF_SPIN)
              gc_spin_spin += thread->gc_spin.spin;
              gc_spin_yield += thread->gc_spin.yield;
              mut_spin_spin += thread->mut_spin.spin;
              mut_spin_yield += thread->mut_spin.yield;
#endif

              any_work += thread->any_work;
              no_work += thread->no_work;
              scav_find_work += thread->scav_find_work;

              par_max_copied = stg_max(gc_threads[i]->copied, par_max_copied);
              par_balanced_copied_acc +=
                  stg_min(n_gc_threads * gc_threads[i]->copied, copied);
          }
      }
      if (n_gc_threads > 1) {
          // See Note [Work Balance] for an explanation of this computation
          par_balanced_copied =
              (par_balanced_copied_acc - copied + (n_gc_threads - 1) / 2) /
              (n_gc_threads - 1);
      }
  }

  // Run through all the generations and tidy up.
  // We're going to:
  //   - count the amount of "live" data (live_words, live_blocks)
  //   - count the amount of "copied" data in this GC (copied)
  //   - free from-space
  //   - make to-space the new from-space (set BF_EVACUATED on all blocks)
  //
  live_words = 0;
  live_blocks = 0;

  for (g = 0; g < RtsFlags.GcFlags.generations; g++) {

    if (g == N) {
      generations[g].collections++; // for stats
      if (n_gc_threads > 1) generations[g].par_collections++;
    }

    // Count the mutable list as bytes "copied" for the purposes of
    // stats.  Every mutable list is copied during every GC.
    if (g > 0) {
        W_ mut_list_size = 0;
        for (n = 0; n < n_capabilities; n++) {
            mut_list_size += countOccupied(capabilities[n]->mut_lists[g]);
        }
        copied +=  mut_list_size;

        debugTrace(DEBUG_gc,
                   "mut_list_size: %lu (%d vars, %d arrays, %d MVARs, %d TVARs, %d TVAR_WATCH_QUEUEs, %d TREC_CHUNKs, %d TREC_HEADERs, %d others)",
                   (unsigned long)(mut_list_size * sizeof(W_)),
                   mutlist_MUTVARS, mutlist_MUTARRS, mutlist_MVARS,
                   mutlist_TVAR, mutlist_TVAR_WATCH_QUEUE,
                   mutlist_TREC_CHUNK, mutlist_TREC_HEADER,
                   mutlist_OTHERS);
    }

    bdescr *next, *prev;
    gen = &generations[g];

    // for generations we collected...
    if (g <= N) {

        /* free old memory and shift to-space into from-space for all
         * the collected generations (except the allocation area).  These
         * freed blocks will probaby be quickly recycled.
         */
        if (gen->mark)
        {
            // tack the new blocks on the end of the existing blocks
            if (gen->old_blocks != NULL) {

                prev = NULL;
                for (bd = gen->old_blocks; bd != NULL; bd = next) {

                    next = bd->link;

                    if (!(bd->flags & BF_MARKED))
                    {
                        if (prev == NULL) {
                            gen->old_blocks = next;
                        } else {
                            prev->link = next;
                        }
                        freeGroup(bd);
                        gen->n_old_blocks--;
                    }
                    else
                    {
                        gen->n_words += bd->free - bd->start;

                        // NB. this step might not be compacted next
                        // time, so reset the BF_MARKED flags.
                        // They are set before GC if we're going to
                        // compact.  (search for BF_MARKED above).
                        bd->flags &= ~BF_MARKED;

                        // between GCs, all blocks in the heap except
                        // for the nursery have the BF_EVACUATED flag set.
                        bd->flags |= BF_EVACUATED;

                        prev = bd;
                    }
                }

                if (prev != NULL) {
                    prev->link = gen->blocks;
                    gen->blocks = gen->old_blocks;
                }
            }
            // add the new blocks to the block tally
            gen->n_blocks += gen->n_old_blocks;
            ASSERT(countBlocks(gen->blocks) == gen->n_blocks);
            ASSERT(countOccupied(gen->blocks) == gen->n_words);
        }
        else // not copacted
        {
            freeChain(gen->old_blocks);
        }

        gen->old_blocks = NULL;
        gen->n_old_blocks = 0;

        /* LARGE OBJECTS.  The current live large objects are chained on
         * scavenged_large, having been moved during garbage
         * collection from large_objects.  Any objects left on the
         * large_objects list are therefore dead, so we free them here.
         */
        freeChain(gen->large_objects);
        gen->large_objects  = gen->scavenged_large_objects;
        gen->n_large_blocks = gen->n_scavenged_large_blocks;
        gen->n_large_words  = countOccupied(gen->large_objects);
        gen->n_new_large_words = 0;

        /* COMPACT_NFDATA. The currently live compacts are chained
         * to live_compact_objects, quite like large objects. And
         * objects left on the compact_objects list are dead.
         *
         * We don't run a simple freeChain because want to give the
         * CNF module some chance to free memory that freeChain would
         * not see (namely blocks appended to a CNF through a compactResize).
         *
         * See Note [Compact Normal Forms] for details.
         */
        for (bd = gen->compact_objects; bd; bd = next) {
            next = bd->link;
            compactFree(((StgCompactNFDataBlock*)bd->start)->owner);
        }
        gen->compact_objects = gen->live_compact_objects;
        gen->n_compact_blocks = gen->n_live_compact_blocks;
    }
    else // for generations > N
    {
        /* For older generations, we need to append the
         * scavenged_large_object list (i.e. large objects that have been
         * promoted during this GC) to the large_object list for that step.
         */
        for (bd = gen->scavenged_large_objects; bd; bd = next) {
            next = bd->link;
            dbl_link_onto(bd, &gen->large_objects);
            gen->n_large_words += bd->free - bd->start;
        }

        // And same for compacts
        for (bd = gen->live_compact_objects; bd; bd = next) {
            next = bd->link;
            dbl_link_onto(bd, &gen->compact_objects);
        }

        // add the new blocks we promoted during this GC
        gen->n_large_blocks += gen->n_scavenged_large_blocks;
        gen->n_compact_blocks += gen->n_live_compact_blocks;
    }

    ASSERT(countBlocks(gen->large_objects) == gen->n_large_blocks);
    ASSERT(countOccupied(gen->large_objects) == gen->n_large_words);
    // We can run the same assertion on compact objects because there
    // is memory "the GC doesn't see" (directly), but which is still
    // accounted in gen->n_compact_blocks

    gen->scavenged_large_objects = NULL;
    gen->n_scavenged_large_blocks = 0;
    gen->live_compact_objects = NULL;
    gen->n_live_compact_blocks = 0;

    // Count "live" data
    live_words  += genLiveWords(gen);
    live_blocks += genLiveBlocks(gen);

    // add in the partial blocks in the gen_workspaces
    {
        uint32_t i;
        for (i = 0; i < n_capabilities; i++) {
            live_words  += gcThreadLiveWords(i, gen->no);
            live_blocks += gcThreadLiveBlocks(i, gen->no);
        }
    }
  } // for all generations

  // update the max size of older generations after a major GC
  resize_generations();

  // Free the mark stack.
  if (mark_stack_top_bd != NULL) {
      debugTrace(DEBUG_gc, "mark stack: %d blocks",
                 countBlocks(mark_stack_top_bd));
      freeChain(mark_stack_top_bd);
  }

  // Free any bitmaps.
  for (g = 0; g <= N; g++) {
      gen = &generations[g];
      if (gen->bitmap != NULL) {
          freeGroup(gen->bitmap);
          gen->bitmap = NULL;
      }
  }

  resize_nursery();

  resetNurseries();

 // mark the garbage collected CAFs as dead
#if defined(DEBUG)
  if (major_gc) { gcCAFs(); }
#endif

  // Update the stable name hash table
  updateStableNameTable(major_gc);

  // unlock the StablePtr table.  Must be before scheduleFinalizers(),
  // because a finalizer may call hs_free_fun_ptr() or
  // hs_free_stable_ptr(), both of which access the StablePtr table.
  stablePtrUnlock();

  // Must be after stablePtrUnlock(), because it might free stable ptrs.
  if (major_gc) {
      checkUnload (gct->scavenged_static_objects);
  }

#if defined(PROFILING)
  // resetStaticObjectForRetainerProfiling() must be called before
  // zeroing below.

  // ToDo: fix the gct->scavenged_static_objects below
  resetStaticObjectForRetainerProfiling(gct->scavenged_static_objects);
#endif

  // Start any pending finalizers.  Must be after
  // updateStableTables() and stableUnlock() (see #4221).
  RELEASE_SM_LOCK;
  scheduleFinalizers(cap, dead_weak_ptr_list);
  ACQUIRE_SM_LOCK;

  // check sanity after GC
  // before resurrectThreads(), because that might overwrite some
  // closures, which will cause problems with THREADED where we don't
  // fill slop.
  IF_DEBUG(sanity, checkSanity(true /* after GC */, major_gc));

  // If a heap census is due, we need to do it before
  // resurrectThreads(), for the same reason as checkSanity above:
  // resurrectThreads() will overwrite some closures and leave slop
  // behind.
  if (do_heap_census) {
      debugTrace(DEBUG_sched, "performing heap census");
      RELEASE_SM_LOCK;
      heapCensus(gct->gc_start_cpu);
      ACQUIRE_SM_LOCK;
  }

  // send exceptions to any threads which were about to die
  RELEASE_SM_LOCK;
  resurrectThreads(resurrected_threads);
  ACQUIRE_SM_LOCK;

  if (major_gc) {
      W_ need_prealloc, need_live, need, got;
      uint32_t i;

      need_live = 0;
      for (i = 0; i < RtsFlags.GcFlags.generations; i++) {
          need_live += genLiveBlocks(&generations[i]);
      }
      need_live = stg_max(RtsFlags.GcFlags.minOldGenSize, need_live);

      need_prealloc = 0;
      for (i = 0; i < n_nurseries; i++) {
          need_prealloc += nurseries[i].n_blocks;
      }
      need_prealloc += RtsFlags.GcFlags.largeAllocLim;
      need_prealloc += countAllocdBlocks(exec_block);
      need_prealloc += arenaBlocks();
#if defined(PROFILING)
      if (RtsFlags.ProfFlags.doHeapProfile == HEAP_BY_RETAINER) {
          need_prealloc = retainerStackBlocks();
      }
#endif

      /* If the amount of data remains constant, next major GC we'll
       * require (F+1)*live + prealloc. We leave (F+2)*live + prealloc
       * in order to reduce repeated deallocation and reallocation. #14702
       */
      need = need_prealloc + (RtsFlags.GcFlags.oldGenFactor + 2) * need_live;

      /* Also, if user set heap size, do not drop below it.
       */
      need = stg_max(RtsFlags.GcFlags.heapSizeSuggestion, need);

      /* But with a large nursery, the above estimate might exceed
       * maxHeapSize.  A large resident set size might make the OS
       * kill this process, or swap unnecessarily.  Therefore we
       * ensure that our estimate does not exceed maxHeapSize.
       */
      if (RtsFlags.GcFlags.maxHeapSize != 0) {
          need = stg_min(RtsFlags.GcFlags.maxHeapSize, need);
      }

      need = BLOCKS_TO_MBLOCKS(need);

      got = mblocks_allocated;

      if (got > need) {
          returnMemoryToOS(got - need);
      }
  }

  // extra GC trace info
  IF_DEBUG(gc, statDescribeGens());

#if defined(DEBUG)
  // symbol-table based profiling
  /*  heapCensus(to_blocks); */ /* ToDo */
#endif

  // restore enclosing cost centre
#if defined(PROFILING)
  for (n = 0; n < n_capabilities; n++) {
      capabilities[n]->r.rCCCS = save_CCS[n];
  }
#endif

#if defined(DEBUG)
  // check for memory leaks if DEBUG is on
  memInventory(DEBUG_gc);
#endif

  // ok, GC over: tell the stats department what happened.
  stat_endGC(cap, gct, live_words, copied,
             live_blocks * BLOCK_SIZE_W - live_words /* slop */,
             N, n_gc_threads, par_max_copied, par_balanced_copied,
             gc_spin_spin, gc_spin_yield, mut_spin_spin, mut_spin_yield,
             any_work, no_work, scav_find_work);

#if defined(RTS_USER_SIGNALS)
  if (RtsFlags.MiscFlags.install_signal_handlers) {
    // unblock signals again
    unblockUserSignals();
  }
#endif

  RELEASE_SM_LOCK;

  SET_GCT(saved_gct);
}

/* -----------------------------------------------------------------------------
   Heap overflow is indicated by setting a flag that the caller of
   GarbageCollect can check.  (not ideal, TODO: better)
   -------------------------------------------------------------------------- */

static void heapOverflow(void)
{
    heap_overflow = true;
}

/* -----------------------------------------------------------------------------
   Initialise the gc_thread structures.
   -------------------------------------------------------------------------- */

static void
new_gc_thread (uint32_t n, gc_thread *t)
{
    uint32_t g;
    gen_workspace *ws;

    t->cap = capabilities[n];

#if defined(THREADED_RTS)
    t->id = 0;
    initSpinLock(&t->gc_spin);
    initSpinLock(&t->mut_spin);
    ACQUIRE_SPIN_LOCK(&t->gc_spin);
    ACQUIRE_SPIN_LOCK(&t->mut_spin);
    t->wakeup = GC_THREAD_INACTIVE;  // starts true, so we can wait for the
                          // thread to start up, see wakeup_gc_threads
#endif

    t->thread_index = n;
    t->free_blocks = NULL;
    t->gc_count = 0;

    init_gc_thread(t);

    for (g = 0; g < RtsFlags.GcFlags.generations; g++)
    {
        ws = &t->gens[g];
        ws->gen = &generations[g];
        ASSERT(g == ws->gen->no);
        ws->my_gct = t;

        // We want to call
        //   alloc_todo_block(ws,0);
        // but can't, because it uses gct which isn't set up at this point.
        // Hence, allocate a block for todo_bd manually:
        {
            bdescr *bd = allocBlockOnNode(capNoToNumaNode(n));
                // no lock, locks aren't initialised yet
            initBdescr(bd, ws->gen, ws->gen->to);
            bd->flags = BF_EVACUATED;
            bd->u.scan = bd->free = bd->start;

            ws->todo_bd = bd;
            ws->todo_free = bd->free;
            ws->todo_lim = bd->start + BLOCK_SIZE_W;
        }

        ws->todo_q = newWSDeque(128);
        ws->todo_overflow = NULL;
        ws->n_todo_overflow = 0;
        ws->todo_large_objects = NULL;

        ws->part_list = NULL;
        ws->n_part_blocks = 0;
        ws->n_part_words = 0;

        ws->scavd_list = NULL;
        ws->n_scavd_blocks = 0;
        ws->n_scavd_words = 0;
    }
}


void
initGcThreads (uint32_t from USED_IF_THREADS, uint32_t to USED_IF_THREADS)
{
#if defined(THREADED_RTS)
    uint32_t i;

    if (from > 0) {
        gc_threads = stgReallocBytes (gc_threads, to * sizeof(gc_thread*),
                                      "initGcThreads");
    } else {
        gc_threads = stgMallocBytes (to * sizeof(gc_thread*),
                                     "initGcThreads");
    }

    for (i = from; i < to; i++) {
        gc_threads[i] =
            stgMallocBytes(sizeof(gc_thread) +
                           RtsFlags.GcFlags.generations * sizeof(gen_workspace),
                           "alloc_gc_threads");

        new_gc_thread(i, gc_threads[i]);
    }
#else
    ASSERT(from == 0 && to == 1);
    gc_threads = stgMallocBytes (sizeof(gc_thread*),"alloc_gc_threads");
    gc_threads[0] = gct;
    new_gc_thread(0,gc_threads[0]);
#endif
}

void
freeGcThreads (void)
{
    uint32_t g;
    if (gc_threads != NULL) {
#if defined(THREADED_RTS)
        uint32_t i;
        for (i = 0; i < n_capabilities; i++) {
            for (g = 0; g < RtsFlags.GcFlags.generations; g++)
            {
                freeWSDeque(gc_threads[i]->gens[g].todo_q);
            }
            stgFree (gc_threads[i]);
        }
        stgFree (gc_threads);
#else
        for (g = 0; g < RtsFlags.GcFlags.generations; g++)
        {
            freeWSDeque(gc_threads[0]->gens[g].todo_q);
        }
        stgFree (gc_threads);
#endif
        gc_threads = NULL;
    }
}

/* ----------------------------------------------------------------------------
   Start GC threads
   ------------------------------------------------------------------------- */

static volatile StgWord gc_running_threads;

static StgWord
inc_running (void)
{
    StgWord new;
    new = atomic_inc(&gc_running_threads, 1);
    ASSERT(new <= n_gc_threads);
    return new;
}

static StgWord
dec_running (void)
{
    ASSERT(gc_running_threads != 0);
    return atomic_dec(&gc_running_threads);
}

static bool
any_work (void)
{
    int g;
    gen_workspace *ws;

    gct->any_work++;

    write_barrier();

    // scavenge objects in compacted generation
    if (mark_stack_bd != NULL && !mark_stack_empty()) {
        return true;
    }

    // Check for global work in any gen.  We don't need to check for
    // local work, because we have already exited scavenge_loop(),
    // which means there is no local work for this thread.
    for (g = 0; g < (int)RtsFlags.GcFlags.generations; g++) {
        ws = &gct->gens[g];
        if (ws->todo_large_objects) return true;
        if (!looksEmptyWSDeque(ws->todo_q)) return true;
        if (ws->todo_overflow) return true;
    }

#if defined(THREADED_RTS)
    if (work_stealing) {
        uint32_t n;
        // look for work to steal
        for (n = 0; n < n_gc_threads; n++) {
            if (n == gct->thread_index) continue;
            for (g = RtsFlags.GcFlags.generations-1; g >= 0; g--) {
                ws = &gc_threads[n]->gens[g];
                if (!looksEmptyWSDeque(ws->todo_q)) return true;
            }
        }
    }
#endif

    gct->no_work++;
#if defined(THREADED_RTS)
    yieldThread();
#endif

    return false;
}

static void
scavenge_until_all_done (void)
{
    DEBUG_ONLY( uint32_t r );


loop:
#if defined(THREADED_RTS)
    if (n_gc_threads > 1) {
        scavenge_loop();
    } else {
        scavenge_loop1();
    }
#else
    scavenge_loop();
#endif

    collect_gct_blocks();

    // scavenge_loop() only exits when there's no work to do

#if defined(DEBUG)
    r = dec_running();
#else
    dec_running();
#endif

    traceEventGcIdle(gct->cap);

    debugTrace(DEBUG_gc, "%d GC threads still running", r);

    while (gc_running_threads != 0) {
        // usleep(1);
        if (any_work()) {
            inc_running();
            traceEventGcWork(gct->cap);
            goto loop;
        }
        // any_work() does not remove the work from the queue, it
        // just checks for the presence of work.  If we find any,
        // then we increment gc_running_threads and go back to
        // scavenge_loop() to perform any pending work.
    }

    traceEventGcDone(gct->cap);
}

#if defined(THREADED_RTS)

void
gcWorkerThread (Capability *cap)
{
    gc_thread *saved_gct;

    // necessary if we stole a callee-saves register for gct:
    saved_gct = gct;

    SET_GCT(gc_threads[cap->no]);
    gct->id = osThreadId();

    // Wait until we're told to wake up
    RELEASE_SPIN_LOCK(&gct->mut_spin);
    // yieldThread();
    //    Strangely, adding a yieldThread() here makes the CPU time
    //    measurements more accurate on Linux, perhaps because it syncs
    //    the CPU time across the multiple cores.  Without this, CPU time
    //    is heavily skewed towards GC rather than MUT.
    gct->wakeup = GC_THREAD_STANDING_BY;
    debugTrace(DEBUG_gc, "GC thread %d standing by...", gct->thread_index);
    ACQUIRE_SPIN_LOCK(&gct->gc_spin);

    init_gc_thread(gct);

    traceEventGcWork(gct->cap);

    // Every thread evacuates some roots.
    gct->evac_gen_no = 0;
    markCapability(mark_root, gct, cap, true/*prune sparks*/);
    scavenge_capability_mut_lists(cap);

    scavenge_until_all_done();

#if defined(THREADED_RTS)
    // Now that the whole heap is marked, we discard any sparks that
    // were found to be unreachable.  The main GC thread is currently
    // marking heap reachable via weak pointers, so it is
    // non-deterministic whether a spark will be retained if it is
    // only reachable via weak pointers.  To fix this problem would
    // require another GC barrier, which is too high a price.
    pruneSparkQueue(cap);
#endif

    // Wait until we're told to continue
    RELEASE_SPIN_LOCK(&gct->gc_spin);
    gct->wakeup = GC_THREAD_WAITING_TO_CONTINUE;
    debugTrace(DEBUG_gc, "GC thread %d waiting to continue...",
               gct->thread_index);
    ACQUIRE_SPIN_LOCK(&gct->mut_spin);
    debugTrace(DEBUG_gc, "GC thread %d on my way...", gct->thread_index);

    SET_GCT(saved_gct);
}

#endif

#if defined(THREADED_RTS)

void
waitForGcThreads (Capability *cap USED_IF_THREADS, bool idle_cap[])
{
    const uint32_t n_threads = n_capabilities;
    const uint32_t me = cap->no;
    uint32_t i, j;
    bool retry = true;
    Time t0, t1, t2;

    t0 = t1 = t2 = getProcessElapsedTime();

    while(retry) {
        for (i=0; i < n_threads; i++) {
            if (i == me || idle_cap[i]) continue;
            if (gc_threads[i]->wakeup != GC_THREAD_STANDING_BY) {
                prodCapability(capabilities[i], cap->running_task);
            }
        }
        for (j=0; j < 10; j++) {
            retry = false;
            for (i=0; i < n_threads; i++) {
                if (i == me || idle_cap[i]) continue;
                write_barrier();
                interruptCapability(capabilities[i]);
                if (gc_threads[i]->wakeup != GC_THREAD_STANDING_BY) {
                    retry = true;
                }
            }
            if (!retry) break;
#if defined(PROF_SPIN)
            waitForGcThreads_yield++;
#endif
            yieldThread();
        }

        t2 = getProcessElapsedTime();
        if (RtsFlags.GcFlags.longGCSync != 0 &&
            t2 - t1 > RtsFlags.GcFlags.longGCSync) {
            /* call this every longGCSync of delay */
            rtsConfig.longGCSync(cap->no, t2 - t0);
            t1 = t2;
        }
        if (retry) {
#if defined(PROF_SPIN)
            // This is a bit strange, we'll get more yields than spins.
            // I guess that means it's not a spin-lock at all, but these
            // numbers are still useful (I think).
            waitForGcThreads_spin++;
#endif
        }
    }

    if (RtsFlags.GcFlags.longGCSync != 0 &&
        t2 - t0 > RtsFlags.GcFlags.longGCSync) {
        rtsConfig.longGCSyncEnd(t2 - t0);
    }
}

#endif // THREADED_RTS

static void
start_gc_threads (void)
{
#if defined(THREADED_RTS)
    gc_running_threads = 0;
#endif
}

static void
wakeup_gc_threads (uint32_t me USED_IF_THREADS,
                   bool idle_cap[] USED_IF_THREADS)
{
#if defined(THREADED_RTS)
    uint32_t i;

    if (n_gc_threads == 1) return;

    for (i=0; i < n_gc_threads; i++) {
        if (i == me || idle_cap[i]) continue;
        inc_running();
        debugTrace(DEBUG_gc, "waking up gc thread %d", i);
        if (gc_threads[i]->wakeup != GC_THREAD_STANDING_BY)
            barf("wakeup_gc_threads");

        gc_threads[i]->wakeup = GC_THREAD_RUNNING;
        ACQUIRE_SPIN_LOCK(&gc_threads[i]->mut_spin);
        RELEASE_SPIN_LOCK(&gc_threads[i]->gc_spin);
    }
#endif
}

// After GC is complete, we must wait for all GC threads to enter the
// standby state, otherwise they may still be executing inside
// any_work(), and may even remain awake until the next GC starts.
static void
shutdown_gc_threads (uint32_t me USED_IF_THREADS,
                     bool idle_cap[] USED_IF_THREADS)
{
#if defined(THREADED_RTS)
    uint32_t i;

    if (n_gc_threads == 1) return;

    for (i=0; i < n_gc_threads; i++) {
        if (i == me || idle_cap[i]) continue;
        while (gc_threads[i]->wakeup != GC_THREAD_WAITING_TO_CONTINUE) {
            busy_wait_nop();
            write_barrier();
        }
    }
#endif
}

#if defined(THREADED_RTS)
void
releaseGCThreads (Capability *cap USED_IF_THREADS, bool idle_cap[])
{
    const uint32_t n_threads = n_capabilities;
    const uint32_t me = cap->no;
    uint32_t i;
    for (i=0; i < n_threads; i++) {
        if (i == me || idle_cap[i]) continue;
        if (gc_threads[i]->wakeup != GC_THREAD_WAITING_TO_CONTINUE)
            barf("releaseGCThreads");

        gc_threads[i]->wakeup = GC_THREAD_INACTIVE;
        ACQUIRE_SPIN_LOCK(&gc_threads[i]->gc_spin);
        RELEASE_SPIN_LOCK(&gc_threads[i]->mut_spin);
    }
}
#endif

/* ----------------------------------------------------------------------------
   Initialise a generation that is to be collected
   ------------------------------------------------------------------------- */

static void
prepare_collected_gen (generation *gen)
{
    uint32_t i, g, n;
    gen_workspace *ws;
    bdescr *bd, *next;

    // Throw away the current mutable list.  Invariant: the mutable
    // list always has at least one block; this means we can avoid a
    // check for NULL in recordMutable().
    g = gen->no;
    if (g != 0) {
        for (i = 0; i < n_capabilities; i++) {
            freeChain(capabilities[i]->mut_lists[g]);
            capabilities[i]->mut_lists[g] =
                allocBlockOnNode(capNoToNumaNode(i));
        }
    }

    gen = &generations[g];
    ASSERT(gen->no == g);

    // we'll construct a new list of threads in this step
    // during GC, throw away the current list.
    gen->old_threads = gen->threads;
    gen->threads = END_TSO_QUEUE;

    // deprecate the existing blocks
    gen->old_blocks   = gen->blocks;
    gen->n_old_blocks = gen->n_blocks;
    gen->blocks       = NULL;
    gen->n_blocks     = 0;
    gen->n_words      = 0;
    gen->live_estimate = 0;

    // initialise the large object queues.
    ASSERT(gen->scavenged_large_objects == NULL);
    ASSERT(gen->n_scavenged_large_blocks == 0);
    ASSERT(gen->live_compact_objects == NULL);
    ASSERT(gen->n_live_compact_blocks == 0);

    // grab all the partial blocks stashed in the gc_thread workspaces and
    // move them to the old_blocks list of this gen.
    for (n = 0; n < n_capabilities; n++) {
        ws = &gc_threads[n]->gens[gen->no];

        for (bd = ws->part_list; bd != NULL; bd = next) {
            next = bd->link;
            bd->link = gen->old_blocks;
            gen->old_blocks = bd;
            gen->n_old_blocks += bd->blocks;
        }
        ws->part_list = NULL;
        ws->n_part_blocks = 0;
        ws->n_part_words = 0;

        ASSERT(ws->scavd_list == NULL);
        ASSERT(ws->n_scavd_blocks == 0);
        ASSERT(ws->n_scavd_words == 0);

        if (ws->todo_free != ws->todo_bd->start) {
            ws->todo_bd->free = ws->todo_free;
            ws->todo_bd->link = gen->old_blocks;
            gen->old_blocks = ws->todo_bd;
            gen->n_old_blocks += ws->todo_bd->blocks;
            alloc_todo_block(ws,0); // always has one block.
        }
    }

    // mark the small objects as from-space
    for (bd = gen->old_blocks; bd; bd = bd->link) {
        bd->flags &= ~BF_EVACUATED;
    }

    // mark the large objects as from-space
    for (bd = gen->large_objects; bd; bd = bd->link) {
        bd->flags &= ~BF_EVACUATED;
    }

    // mark the compact objects as from-space
    for (bd = gen->compact_objects; bd; bd = bd->link) {
        bd->flags &= ~BF_EVACUATED;
    }

    // for a compacted generation, we need to allocate the bitmap
    if (gen->mark) {
        StgWord bitmap_size; // in bytes
        bdescr *bitmap_bdescr;
        StgWord *bitmap;

        bitmap_size = gen->n_old_blocks * BLOCK_SIZE / BITS_IN(W_);

        if (bitmap_size > 0) {
            bitmap_bdescr = allocGroup((StgWord)BLOCK_ROUND_UP(bitmap_size)
                                       / BLOCK_SIZE);
            gen->bitmap = bitmap_bdescr;
            bitmap = bitmap_bdescr->start;

            debugTrace(DEBUG_gc, "bitmap_size: %d, bitmap: %p",
                       bitmap_size, bitmap);

            // don't forget to fill it with zeros!
            memset(bitmap, 0, bitmap_size);

            // For each block in this step, point to its bitmap from the
            // block descriptor.
            for (bd=gen->old_blocks; bd != NULL; bd = bd->link) {
                bd->u.bitmap = bitmap;
                bitmap += BLOCK_SIZE_W / BITS_IN(W_);

                // Also at this point we set the BF_MARKED flag
                // for this block.  The invariant is that
                // BF_MARKED is always unset, except during GC
                // when it is set on those blocks which will be
                // compacted.
                if (!(bd->flags & BF_FRAGMENTED)) {
                    bd->flags |= BF_MARKED;
                }

                // BF_SWEPT should be marked only for blocks that are being
                // collected in sweep()
                bd->flags &= ~BF_SWEPT;
            }
        }
    }
}


/* ----------------------------------------------------------------------------
   Save the mutable lists in saved_mut_lists
   ------------------------------------------------------------------------- */

static void
stash_mut_list (Capability *cap, uint32_t gen_no)
{
    cap->saved_mut_lists[gen_no] = cap->mut_lists[gen_no];
    cap->mut_lists[gen_no] = allocBlockOnNode_sync(cap->node);
}

/* ----------------------------------------------------------------------------
   Initialise a generation that is *not* to be collected
   ------------------------------------------------------------------------- */

static void
prepare_uncollected_gen (generation *gen)
{
    uint32_t i;


    ASSERT(gen->no > 0);

    // save the current mutable lists for this generation, and
    // allocate a fresh block for each one.  We'll traverse these
    // mutable lists as roots early on in the GC.
    for (i = 0; i < n_capabilities; i++) {
        stash_mut_list(capabilities[i], gen->no);
    }

    ASSERT(gen->scavenged_large_objects == NULL);
    ASSERT(gen->n_scavenged_large_blocks == 0);
}

/* -----------------------------------------------------------------------------
   Collect the completed blocks from a GC thread and attach them to
   the generation.
   -------------------------------------------------------------------------- */

static void
collect_gct_blocks (void)
{
    uint32_t g;
    gen_workspace *ws;
    bdescr *bd, *prev;

    for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
        ws = &gct->gens[g];

        // there may still be a block attached to ws->todo_bd;
        // leave it there to use next time.

        if (ws->scavd_list != NULL) {
            ACQUIRE_SPIN_LOCK(&ws->gen->sync);

            ASSERT(gct->scan_bd == NULL);
            ASSERT(countBlocks(ws->scavd_list) == ws->n_scavd_blocks);

            prev = NULL;
            for (bd = ws->scavd_list; bd != NULL; bd = bd->link) {
                prev = bd;
            }
            if (prev != NULL) {
                prev->link = ws->gen->blocks;
                ws->gen->blocks = ws->scavd_list;
            }
            ws->gen->n_blocks += ws->n_scavd_blocks;
            ws->gen->n_words += ws->n_scavd_words;

            ws->scavd_list = NULL;
            ws->n_scavd_blocks = 0;
            ws->n_scavd_words = 0;

            RELEASE_SPIN_LOCK(&ws->gen->sync);
        }
    }
}

/* -----------------------------------------------------------------------------
   During mutation, any blocks that are filled by allocatePinned() are
   stashed on the local pinned_object_blocks list, to avoid needing to
   take a global lock.  Here we collect those blocks from the
   cap->pinned_object_blocks lists and put them on the
   main g0->large_object list.
   -------------------------------------------------------------------------- */

static void
collect_pinned_object_blocks (void)
{
    uint32_t n;
    bdescr *bd, *prev;

    for (n = 0; n < n_capabilities; n++) {
        prev = NULL;
        for (bd = capabilities[n]->pinned_object_blocks; bd != NULL; bd = bd->link) {
            prev = bd;
        }
        if (prev != NULL) {
            prev->link = g0->large_objects;
            if (g0->large_objects != NULL) {
                g0->large_objects->u.back = prev;
            }
            g0->large_objects = capabilities[n]->pinned_object_blocks;
            capabilities[n]->pinned_object_blocks = 0;
        }
    }
}

/* -----------------------------------------------------------------------------
   Initialise a gc_thread before GC
   -------------------------------------------------------------------------- */

static void
init_gc_thread (gc_thread *t)
{
    t->static_objects = END_OF_STATIC_OBJECT_LIST;
    t->scavenged_static_objects = END_OF_STATIC_OBJECT_LIST;
    t->scan_bd = NULL;
    t->mut_lists = t->cap->mut_lists;
    t->evac_gen_no = 0;
    t->failed_to_evac = false;
    t->eager_promotion = true;
    t->thunk_selector_depth = 0;
    t->copied = 0;
    t->scanned = 0;
    t->any_work = 0;
    t->no_work = 0;
    t->scav_find_work = 0;
}

/* -----------------------------------------------------------------------------
   Function we pass to evacuate roots.
   -------------------------------------------------------------------------- */

static void
mark_root(void *user USED_IF_THREADS, StgClosure **root)
{
    // we stole a register for gct, but this function is called from
    // *outside* the GC where the register variable is not in effect,
    // so we need to save and restore it here.  NB. only call
    // mark_root() from the main GC thread, otherwise gct will be
    // incorrect.
#if defined(THREADED_RTS)
    gc_thread *saved_gct;
    saved_gct = gct;
#endif
    SET_GCT(user);

    evacuate(root);

    SET_GCT(saved_gct);
}

/* ----------------------------------------------------------------------------
   Reset the sizes of the older generations when we do a major
   collection.

   CURRENT STRATEGY: make all generations except zero the same size.
   We have to stay within the maximum heap size, and leave a certain
   percentage of the maximum heap size available to allocate into.
   ------------------------------------------------------------------------- */

static void
resize_generations (void)
{
    uint32_t g;

    if (major_gc && RtsFlags.GcFlags.generations > 1) {
        W_ live, size, min_alloc, words;
        const W_ max  = RtsFlags.GcFlags.maxHeapSize;
        const W_ gens = RtsFlags.GcFlags.generations;

        // live in the oldest generations
        if (oldest_gen->live_estimate != 0) {
            words = oldest_gen->live_estimate;
        } else {
            words = oldest_gen->n_words;
        }
        live = (words + BLOCK_SIZE_W - 1) / BLOCK_SIZE_W +
            oldest_gen->n_large_blocks +
            oldest_gen->n_compact_blocks;

        // default max size for all generations except zero
        size = stg_max(live * RtsFlags.GcFlags.oldGenFactor,
                       RtsFlags.GcFlags.minOldGenSize);

        if (RtsFlags.GcFlags.heapSizeSuggestionAuto) {
            if (max > 0) {
                RtsFlags.GcFlags.heapSizeSuggestion = stg_min(max, size);
            } else {
                RtsFlags.GcFlags.heapSizeSuggestion = size;
            }
        }

        // minimum size for generation zero
        min_alloc = stg_max((RtsFlags.GcFlags.pcFreeHeap * max) / 200,
                            RtsFlags.GcFlags.minAllocAreaSize
                            * (W_)n_capabilities);

        // Auto-enable compaction when the residency reaches a
        // certain percentage of the maximum heap size (default: 30%).
        if (RtsFlags.GcFlags.compact ||
            (max > 0 &&
             oldest_gen->n_blocks >
             (RtsFlags.GcFlags.compactThreshold * max) / 100)) {
            oldest_gen->mark = 1;
            oldest_gen->compact = 1;
//        debugBelch("compaction: on\n", live);
        } else {
            oldest_gen->mark = 0;
            oldest_gen->compact = 0;
//        debugBelch("compaction: off\n", live);
        }

        if (RtsFlags.GcFlags.sweep) {
            oldest_gen->mark = 1;
        }

        // if we're going to go over the maximum heap size, reduce the
        // size of the generations accordingly.  The calculation is
        // different if compaction is turned on, because we don't need
        // to double the space required to collect the old generation.
        if (max != 0) {

            // this test is necessary to ensure that the calculations
            // below don't have any negative results - we're working
            // with unsigned values here.
            if (max < min_alloc) {
                heapOverflow();
            }

            if (oldest_gen->compact) {
                if ( (size + (size - 1) * (gens - 2) * 2) + min_alloc > max ) {
                    size = (max - min_alloc) / ((gens - 1) * 2 - 1);
                }
            } else {
                if ( (size * (gens - 1) * 2) + min_alloc > max ) {
                    size = (max - min_alloc) / ((gens - 1) * 2);
                }
            }

            if (size < live) {
                heapOverflow();
            }
        }

#if 0
        debugBelch("live: %d, min_alloc: %d, size : %d, max = %d\n", live,
                   min_alloc, size, max);
#endif

        for (g = 0; g < gens; g++) {
            generations[g].max_blocks = size;
        }
    }
}

/* -----------------------------------------------------------------------------
   Calculate the new size of the nursery, and resize it.
   -------------------------------------------------------------------------- */

static void
resize_nursery (void)
{
    const StgWord min_nursery =
      RtsFlags.GcFlags.minAllocAreaSize * (StgWord)n_capabilities;

    if (RtsFlags.GcFlags.generations == 1)
    {   // Two-space collector:
        W_ blocks;

        /* set up a new nursery.  Allocate a nursery size based on a
         * function of the amount of live data (by default a factor of 2)
         * Use the blocks from the old nursery if possible, freeing up any
         * left over blocks.
         *
         * If we get near the maximum heap size, then adjust our nursery
         * size accordingly.  If the nursery is the same size as the live
         * data (L), then we need 3L bytes.  We can reduce the size of the
         * nursery to bring the required memory down near 2L bytes.
         *
         * A normal 2-space collector would need 4L bytes to give the same
         * performance we get from 3L bytes, reducing to the same
         * performance at 2L bytes.
         */
        blocks = generations[0].n_blocks;

        if ( RtsFlags.GcFlags.maxHeapSize != 0 &&
             blocks * RtsFlags.GcFlags.oldGenFactor * 2 >
             RtsFlags.GcFlags.maxHeapSize )
        {
            long adjusted_blocks;  // signed on purpose
            int pc_free;

            adjusted_blocks = (RtsFlags.GcFlags.maxHeapSize - 2 * blocks);

            debugTrace(DEBUG_gc, "near maximum heap size of 0x%x blocks, blocks = %d, adjusted to %ld",
                       RtsFlags.GcFlags.maxHeapSize, blocks, adjusted_blocks);

            pc_free = adjusted_blocks * 100 / RtsFlags.GcFlags.maxHeapSize;
            if (pc_free < RtsFlags.GcFlags.pcFreeHeap) /* might even * be < 0 */
            {
                heapOverflow();
            }
            blocks = adjusted_blocks;
        }
        else
        {
            blocks *= RtsFlags.GcFlags.oldGenFactor;
            if (blocks < min_nursery)
            {
                blocks = min_nursery;
            }
        }
        resizeNurseries(blocks);
    }
    else  // Generational collector
    {
        /*
         * If the user has given us a suggested heap size, adjust our
         * allocation area to make best use of the memory available.
         */
        if (RtsFlags.GcFlags.heapSizeSuggestion)
        {
            long blocks;
            StgWord needed;

            calcNeeded(false, &needed); // approx blocks needed at next GC

            /* Guess how much will be live in generation 0 step 0 next time.
             * A good approximation is obtained by finding the
             * percentage of g0 that was live at the last minor GC.
             *
             * We have an accurate figure for the amount of copied data in
             * 'copied', but we must convert this to a number of blocks, with
             * a small adjustment for estimated slop at the end of a block
             * (- 10 words).
             */
            if (N == 0)
            {
                g0_pcnt_kept = ((copied / (BLOCK_SIZE_W - 10)) * 100)
                    / countNurseryBlocks();
            }

            /* Estimate a size for the allocation area based on the
             * information available.  We might end up going slightly under
             * or over the suggested heap size, but we should be pretty
             * close on average.
             *
             * Formula:            suggested - needed
             *                ----------------------------
             *                    1 + g0_pcnt_kept/100
             *
             * where 'needed' is the amount of memory needed at the next
             * collection for collecting all gens except g0.
             */
            blocks =
                (((long)RtsFlags.GcFlags.heapSizeSuggestion - (long)needed) * 100) /
                (100 + (long)g0_pcnt_kept);

            if (blocks < (long)min_nursery) {
                blocks = min_nursery;
            }

            resizeNurseries((W_)blocks);
        }
        else
        {
            // we might have added extra blocks to the nursery, so
            // resize back to the original size again.
            resizeNurseriesFixed();
        }
    }
}

/* -----------------------------------------------------------------------------
   Sanity code for CAF garbage collection.

   With DEBUG turned on, we manage a CAF list in addition to the SRT
   mechanism.  After GC, we run down the CAF list and make any
   CAFs which have been garbage collected GCD_CAF.  This means we get an error
   whenever the program tries to enter a garbage collected CAF.

   Any garbage collected CAFs are taken off the CAF list at the same
   time.
   -------------------------------------------------------------------------- */

#if defined(DEBUG)

static void gcCAFs(void)
{
    StgIndStatic *p, *prev;

    const StgInfoTable *info;
    uint32_t i;

    i = 0;
    p = debug_caf_list;
    prev = NULL;

    for (p = debug_caf_list; p != (StgIndStatic*)END_OF_CAF_LIST;
         p = (StgIndStatic*)p->saved_info) {

        info = get_itbl((StgClosure*)p);
        ASSERT(info->type == IND_STATIC);

        // See Note [STATIC_LINK fields] in Storage.h
        // This condition identifies CAFs that have just been GC'd and
        // don't have static_link==3 which means they should be ignored.
        if ((((StgWord)(p->static_link)&STATIC_BITS) | prev_static_flag) != 3) {
            debugTrace(DEBUG_gccafs, "CAF gc'd at 0x%p", p);
            SET_INFO((StgClosure*)p,&stg_GCD_CAF_info); // stub it
            if (prev == NULL) {
                debug_caf_list = (StgIndStatic*)p->saved_info;
            } else {
                prev->saved_info = p->saved_info;
            }
        } else {
            prev = p;
            i++;
        }
    }

    debugTrace(DEBUG_gccafs, "%d CAFs live", i);
}
#endif


/* -----------------------------------------------------------------------------
   The GC can leave some work for the mutator to do before the next
   GC, provided the work can be safely overlapped with mutation.  This
   can help reduce the GC pause time.

   The mutator can call doIdleGCWork() any time it likes, but
   preferably when it is idle.  It's safe for multiple capabilities to
   call doIdleGCWork().

   When 'all' is
     * false: doIdleGCWork() should only take a short, bounded, amount
       of time.
     * true: doIdleGCWork() will complete all the outstanding GC work.

   The return value is
     * true if there's more to do (only if 'all' is false).
     * false otherwise.
  -------------------------------------------------------------------------- */

bool doIdleGCWork(Capability *cap STG_UNUSED, bool all)
{
    return runSomeFinalizers(all);
}