summaryrefslogtreecommitdiff
path: root/rts/sm/Storage.c
blob: dcc5b3a3c740ee74ae93b94b4cfd65360327a2a3 (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
/* -----------------------------------------------------------------------------
 *
 * (c) The GHC Team, 1998-2012
 *
 * Storage manager front end
 *
 * Documentation on the architecture of the Storage Manager can be
 * found in the online commentary:
 *
 *   http://ghc.haskell.org/trac/ghc/wiki/Commentary/Rts/Storage
 *
 * ---------------------------------------------------------------------------*/

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

#include "Storage.h"
#include "GCThread.h"
#include "RtsUtils.h"
#include "Stats.h"
#include "BlockAlloc.h"
#include "Weak.h"
#include "Sanity.h"
#include "Arena.h"
#include "Capability.h"
#include "Schedule.h"
#include "RetainerProfile.h"        // for counting memory blocks (memInventory)
#include "OSMem.h"
#include "Trace.h"
#include "GC.h"
#include "Evac.h"
#if defined(ios_HOST_OS)
#include "Hash.h"
#endif

#include <string.h>

#include "ffi.h"

/*
 * All these globals require sm_mutex to access in THREADED_RTS mode.
 */
StgIndStatic  *dyn_caf_list        = NULL;
StgIndStatic  *debug_caf_list      = NULL;
StgIndStatic  *revertible_caf_list = NULL;
bool           keepCAFs;

W_ large_alloc_lim;    /* GC if n_large_blocks in any nursery
                        * reaches this. */

bdescr *exec_block;

generation *generations = NULL; /* all the generations */
generation *g0          = NULL; /* generation 0, for convenience */
generation *oldest_gen  = NULL; /* oldest generation, for convenience */

/*
 * Array of nurseries, size == n_capabilities
 *
 * nursery[i] belongs to NUMA node (i % n_numa_nodes)
 * This is chosen to be the same convention as capabilities[i], so
 * that when not using nursery chunks (+RTS -n), we just map
 * capabilities to nurseries 1:1.
 */
nursery *nurseries = NULL;
uint32_t n_nurseries;

/*
 * When we are using nursery chunks, we need a separate next_nursery
 * pointer for each NUMA node.
 */
volatile StgWord next_nursery[MAX_NUMA_NODES];

#if defined(THREADED_RTS)
/*
 * Storage manager mutex:  protects all the above state from
 * simultaneous access by two STG threads.
 */
Mutex sm_mutex;
#endif

static void allocNurseries (uint32_t from, uint32_t to);
static void assignNurseriesToCapabilities (uint32_t from, uint32_t to);

static void
initGeneration (generation *gen, int g)
{
    gen->no = g;
    gen->collections = 0;
    gen->par_collections = 0;
    gen->failed_promotions = 0;
    gen->max_blocks = 0;
    gen->blocks = NULL;
    gen->n_blocks = 0;
    gen->n_words = 0;
    gen->live_estimate = 0;
    gen->old_blocks = NULL;
    gen->n_old_blocks = 0;
    gen->large_objects = NULL;
    gen->n_large_blocks = 0;
    gen->n_large_words = 0;
    gen->n_new_large_words = 0;
    gen->compact_objects = NULL;
    gen->n_compact_blocks = 0;
    gen->compact_blocks_in_import = NULL;
    gen->n_compact_blocks_in_import = 0;
    gen->scavenged_large_objects = NULL;
    gen->n_scavenged_large_blocks = 0;
    gen->live_compact_objects = NULL;
    gen->n_live_compact_blocks = 0;
    gen->compact_blocks_in_import = NULL;
    gen->n_compact_blocks_in_import = 0;
    gen->mark = 0;
    gen->compact = 0;
    gen->bitmap = NULL;
#if defined(THREADED_RTS)
    initSpinLock(&gen->sync);
#endif
    gen->threads = END_TSO_QUEUE;
    gen->old_threads = END_TSO_QUEUE;
    gen->weak_ptr_list = NULL;
    gen->old_weak_ptr_list = NULL;
}

void
initStorage (void)
{
  uint32_t g, n;

  if (generations != NULL) {
      // multi-init protection
      return;
  }

  initMBlocks();

  /* Sanity check to make sure the LOOKS_LIKE_ macros appear to be
   * doing something reasonable.
   */
  /* We use the NOT_NULL variant or gcc warns that the test is always true */
  ASSERT(LOOKS_LIKE_INFO_PTR_NOT_NULL((StgWord)&stg_BLOCKING_QUEUE_CLEAN_info));
  ASSERT(LOOKS_LIKE_CLOSURE_PTR(&stg_dummy_ret_closure));
  ASSERT(!HEAP_ALLOCED(&stg_dummy_ret_closure));

  initBlockAllocator();

#if defined(THREADED_RTS)
  initMutex(&sm_mutex);
#endif

  ACQUIRE_SM_LOCK;

  /* allocate generation info array */
  generations = (generation *)stgMallocBytes(RtsFlags.GcFlags.generations
                                             * sizeof(struct generation_),
                                             "initStorage: gens");

  /* Initialise all generations */
  for(g = 0; g < RtsFlags.GcFlags.generations; g++) {
      initGeneration(&generations[g], g);
  }

  /* A couple of convenience pointers */
  g0 = &generations[0];
  oldest_gen = &generations[RtsFlags.GcFlags.generations-1];

  /* Set up the destination pointers in each younger gen. step */
  for (g = 0; g < RtsFlags.GcFlags.generations-1; g++) {
      generations[g].to = &generations[g+1];
  }
  oldest_gen->to = oldest_gen;

  /* The oldest generation has one step. */
  if (RtsFlags.GcFlags.compact || RtsFlags.GcFlags.sweep) {
      if (RtsFlags.GcFlags.generations == 1) {
          errorBelch("WARNING: compact/sweep is incompatible with -G1; disabled");
      } else {
          oldest_gen->mark = 1;
          if (RtsFlags.GcFlags.compact)
              oldest_gen->compact = 1;
      }
  }

  generations[0].max_blocks = 0;

  dyn_caf_list = (StgIndStatic*)END_OF_CAF_LIST;
  debug_caf_list = (StgIndStatic*)END_OF_CAF_LIST;
  revertible_caf_list = (StgIndStatic*)END_OF_CAF_LIST;

  if (RtsFlags.GcFlags.largeAllocLim > 0) {
      large_alloc_lim = RtsFlags.GcFlags.largeAllocLim * BLOCK_SIZE_W;
  } else {
      large_alloc_lim = RtsFlags.GcFlags.minAllocAreaSize * BLOCK_SIZE_W;
  }

  exec_block = NULL;

#if defined(THREADED_RTS)
  initSpinLock(&gc_alloc_block_sync);
#endif
  N = 0;

  for (n = 0; n < n_numa_nodes; n++) {
      next_nursery[n] = n;
  }
  storageAddCapabilities(0, n_capabilities);

  IF_DEBUG(gc, statDescribeGens());

  RELEASE_SM_LOCK;

  traceEventHeapInfo(CAPSET_HEAP_DEFAULT,
                     RtsFlags.GcFlags.generations,
                     RtsFlags.GcFlags.maxHeapSize * BLOCK_SIZE,
                     RtsFlags.GcFlags.minAllocAreaSize * BLOCK_SIZE,
                     MBLOCK_SIZE,
                     BLOCK_SIZE);
}

void storageAddCapabilities (uint32_t from, uint32_t to)
{
    uint32_t n, g, i, new_n_nurseries;
    nursery *old_nurseries;

    if (RtsFlags.GcFlags.nurseryChunkSize == 0) {
        new_n_nurseries = to;
    } else {
        memcount total_alloc = to * RtsFlags.GcFlags.minAllocAreaSize;
        new_n_nurseries =
            stg_max(to, total_alloc / RtsFlags.GcFlags.nurseryChunkSize);
    }

    old_nurseries = nurseries;
    if (from > 0) {
        nurseries = stgReallocBytes(nurseries,
                                    new_n_nurseries * sizeof(struct nursery_),
                                    "storageAddCapabilities");
    } else {
        nurseries = stgMallocBytes(new_n_nurseries * sizeof(struct nursery_),
                                   "storageAddCapabilities");
    }

    // we've moved the nurseries, so we have to update the rNursery
    // pointers from the Capabilities.
    for (i = 0; i < from; i++) {
        uint32_t index = capabilities[i]->r.rNursery - old_nurseries;
        capabilities[i]->r.rNursery = &nurseries[index];
    }

    /* The allocation area.  Policy: keep the allocation area
     * small to begin with, even if we have a large suggested heap
     * size.  Reason: we're going to do a major collection first, and we
     * don't want it to be a big one.  This vague idea is borne out by
     * rigorous experimental evidence.
     */
    allocNurseries(n_nurseries, new_n_nurseries);
    n_nurseries = new_n_nurseries;

    /*
     * Assign each of the new capabilities a nursery.  Remember to start from
     * next_nursery, because we may have already consumed some of the earlier
     * nurseries.
     */
    assignNurseriesToCapabilities(from,to);

    // allocate a block for each mut list
    for (n = from; n < to; n++) {
        for (g = 1; g < RtsFlags.GcFlags.generations; g++) {
            capabilities[n]->mut_lists[g] =
                allocBlockOnNode(capNoToNumaNode(n));
        }
    }

#if defined(THREADED_RTS) && defined(llvm_CC_FLAVOR) && (CC_SUPPORTS_TLS == 0)
    newThreadLocalKey(&gctKey);
#endif

    initGcThreads(from, to);
}


void
exitStorage (void)
{
    updateNurseriesStats();
    stat_exit();
}

void
freeStorage (bool free_heap)
{
    stgFree(generations);
    if (free_heap) freeAllMBlocks();
#if defined(THREADED_RTS)
    closeMutex(&sm_mutex);
#endif
    stgFree(nurseries);
#if defined(THREADED_RTS) && defined(llvm_CC_FLAVOR) && (CC_SUPPORTS_TLS == 0)
    freeThreadLocalKey(&gctKey);
#endif
    freeGcThreads();
}

/* -----------------------------------------------------------------------------
   Note [CAF management].

   The entry code for every CAF does the following:

      - calls newCAF, which builds a CAF_BLACKHOLE on the heap and atomically
        updates the CAF with IND_STATIC pointing to the CAF_BLACKHOLE

      - if newCAF returns zero, it re-enters the CAF (see Note [atomic
        CAF entry])

      - pushes an update frame pointing to the CAF_BLACKHOLE

   Why do we build a BLACKHOLE in the heap rather than just updating
   the thunk directly?  It's so that we only need one kind of update
   frame - otherwise we'd need a static version of the update frame
   too, and various other parts of the RTS that deal with update
   frames would also need special cases for static update frames.

   newCAF() does the following:

      - atomically locks the CAF (see [atomic CAF entry])

      - it builds a CAF_BLACKHOLE on the heap

      - it updates the CAF with an IND_STATIC pointing to the
        CAF_BLACKHOLE, atomically.

      - it puts the CAF on the oldest generation's mutable list.
        This is so that we treat the CAF as a root when collecting
        younger generations.

      - links the CAF onto the CAF list (see below)

   ------------------
   Note [atomic CAF entry]

   With THREADED_RTS, newCAF() is required to be atomic (see
   #5558). This is because if two threads happened to enter the same
   CAF simultaneously, they would create two distinct CAF_BLACKHOLEs,
   and so the normal threadPaused() machinery for detecting duplicate
   evaluation will not detect this.  Hence in lockCAF() below, we
   atomically lock the CAF with WHITEHOLE before updating it with
   IND_STATIC, and return zero if another thread locked the CAF first.
   In the event that we lost the race, CAF entry code will re-enter
   the CAF and block on the other thread's CAF_BLACKHOLE.

   ------------------
   Note [GHCi CAFs]

   For GHCI, we have additional requirements when dealing with CAFs:

      - we must *retain* all dynamically-loaded CAFs ever entered,
        just in case we need them again.
      - we must be able to *revert* CAFs that have been evaluated, to
        their pre-evaluated form.

      To do this, we use an additional CAF list.  When newCAF() is
      called on a dynamically-loaded CAF, we add it to the CAF list
      instead of the old-generation mutable list, and save away its
      old info pointer (in caf->saved_info) for later reversion.

      To revert all the CAFs, we traverse the CAF list and reset the
      info pointer to caf->saved_info, then throw away the CAF list.
      (see GC.c:revertCAFs()).

      -- SDM 29/1/01

   -------------------------------------------------------------------------- */

STATIC_INLINE StgInd *
lockCAF (StgRegTable *reg, StgIndStatic *caf)
{
    const StgInfoTable *orig_info;
    Capability *cap = regTableToCapability(reg);
    StgInd *bh;

    orig_info = caf->header.info;

#if defined(THREADED_RTS)
    const StgInfoTable *cur_info;

    if (orig_info == &stg_IND_STATIC_info ||
        orig_info == &stg_WHITEHOLE_info) {
        // already claimed by another thread; re-enter the CAF
        return NULL;
    }

    cur_info = (const StgInfoTable *)
        cas((StgVolatilePtr)&caf->header.info,
            (StgWord)orig_info,
            (StgWord)&stg_WHITEHOLE_info);

    if (cur_info != orig_info) {
        // already claimed by another thread; re-enter the CAF
        return NULL;
    }

    // successfully claimed by us; overwrite with IND_STATIC
#endif

    // For the benefit of revertCAFs(), save the original info pointer
    caf->saved_info = orig_info;

    // Allocate the blackhole indirection closure
    bh = (StgInd *)allocate(cap, sizeofW(*bh));
    SET_HDR(bh, &stg_CAF_BLACKHOLE_info, caf->header.prof.ccs);
    bh->indirectee = (StgClosure *)cap->r.rCurrentTSO;

    caf->indirectee = (StgClosure *)bh;
    write_barrier();
    SET_INFO((StgClosure*)caf,&stg_IND_STATIC_info);

    return bh;
}

StgInd *
newCAF(StgRegTable *reg, StgIndStatic *caf)
{
    StgInd *bh;

    bh = lockCAF(reg, caf);
    if (!bh) return NULL;

    if(keepCAFs)
    {
        // Note [dyn_caf_list]
        // If we are in GHCi _and_ we are using dynamic libraries,
        // then we can't redirect newCAF calls to newRetainedCAF (see below),
        // so we make newCAF behave almost like newRetainedCAF.
        // The dynamic libraries might be used by both the interpreted
        // program and GHCi itself, so they must not be reverted.
        // This also means that in GHCi with dynamic libraries, CAFs are not
        // garbage collected. If this turns out to be a problem, we could
        // do another hack here and do an address range test on caf to figure
        // out whether it is from a dynamic library.

        ACQUIRE_SM_LOCK; // dyn_caf_list is global, locked by sm_mutex
        caf->static_link = (StgClosure*)dyn_caf_list;
        dyn_caf_list = (StgIndStatic*)((StgWord)caf | STATIC_FLAG_LIST);
        RELEASE_SM_LOCK;
    }
    else
    {
        // Put this CAF on the mutable list for the old generation.
        if (oldest_gen->no != 0) {
            recordMutableCap((StgClosure*)caf,
                             regTableToCapability(reg), oldest_gen->no);
        }

#if defined(DEBUG)
        // In the DEBUG rts, we keep track of live CAFs by chaining them
        // onto a list debug_caf_list.  This is so that we can tell if we
        // ever enter a GC'd CAF, and emit a suitable barf().
        //
        // The saved_info field of the CAF is used as the link field for
        // debug_caf_list, because this field is only used by newDynCAF
        // for revertible CAFs, and we don't put those on the
        // debug_caf_list.
        ACQUIRE_SM_LOCK; // debug_caf_list is global, locked by sm_mutex
        ((StgIndStatic *)caf)->saved_info = (const StgInfoTable*)debug_caf_list;
        debug_caf_list = (StgIndStatic*)caf;
        RELEASE_SM_LOCK;
#endif
    }

    return bh;
}

// External API for setting the keepCAFs flag. see #3900.
void
setKeepCAFs (void)
{
    keepCAFs = 1;
}

// An alternate version of newCAF which is used for dynamically loaded
// object code in GHCi.  In this case we want to retain *all* CAFs in
// the object code, because they might be demanded at any time from an
// expression evaluated on the command line.
// Also, GHCi might want to revert CAFs, so we add these to the
// revertible_caf_list.
//
// The linker hackily arranges that references to newCAF from dynamic
// code end up pointing to newRetainedCAF.
//
StgInd* newRetainedCAF (StgRegTable *reg, StgIndStatic *caf)
{
    StgInd *bh;

    bh = lockCAF(reg, caf);
    if (!bh) return NULL;

    ACQUIRE_SM_LOCK;

    caf->static_link = (StgClosure*)revertible_caf_list;
    revertible_caf_list = (StgIndStatic*)((StgWord)caf | STATIC_FLAG_LIST);

    RELEASE_SM_LOCK;

    return bh;
}

// If we are using loadObj/unloadObj in the linker, then we want to
//
//  - retain all CAFs in statically linked code (keepCAFs == true),
//    because we might link a new object that uses any of these CAFs.
//
//  - GC CAFs in dynamically-linked code, so that we can detect when
//    a dynamically-linked object is unloadable.
//
// So for this case, we set keepCAFs to true, and link newCAF to newGCdCAF
// for dynamically-linked code.
//
StgInd* newGCdCAF (StgRegTable *reg, StgIndStatic *caf)
{
    StgInd *bh;

    bh = lockCAF(reg, caf);
    if (!bh) return NULL;

    // Put this CAF on the mutable list for the old generation.
    if (oldest_gen->no != 0) {
        recordMutableCap((StgClosure*)caf,
                         regTableToCapability(reg), oldest_gen->no);
    }

    return bh;
}

/* -----------------------------------------------------------------------------
   Nursery management.
   -------------------------------------------------------------------------- */

static bdescr *
allocNursery (uint32_t node, bdescr *tail, W_ blocks)
{
    bdescr *bd = NULL;
    W_ i, n;

    // We allocate the nursery as a single contiguous block and then
    // divide it into single blocks manually.  This way we guarantee
    // that the nursery blocks are adjacent, so that the processor's
    // automatic prefetching works across nursery blocks.  This is a
    // tiny optimisation (~0.5%), but it's free.

    while (blocks > 0) {
        n = stg_min(BLOCKS_PER_MBLOCK, blocks);
        // allocLargeChunk will prefer large chunks, but will pick up
        // small chunks if there are any available.  We must allow
        // single blocks here to avoid fragmentation (#7257)
        bd = allocLargeChunkOnNode(node, 1, n);
        n = bd->blocks;
        blocks -= n;

        for (i = 0; i < n; i++) {
            initBdescr(&bd[i], g0, g0);

            bd[i].blocks = 1;
            bd[i].flags = 0;

            if (i > 0) {
                bd[i].u.back = &bd[i-1];
            } else {
                bd[i].u.back = NULL;
            }

            if (i+1 < n) {
                bd[i].link = &bd[i+1];
            } else {
                bd[i].link = tail;
                if (tail != NULL) {
                    tail->u.back = &bd[i];
                }
            }

            bd[i].free = bd[i].start;
        }

        tail = &bd[0];
    }

    return &bd[0];
}

STATIC_INLINE void
assignNurseryToCapability (Capability *cap, uint32_t n)
{
    ASSERT(n < n_nurseries);
    cap->r.rNursery = &nurseries[n];
    cap->r.rCurrentNursery = nurseries[n].blocks;
    newNurseryBlock(nurseries[n].blocks);
    cap->r.rCurrentAlloc   = NULL;
    ASSERT(cap->r.rCurrentNursery->node == cap->node);
}

/*
 * Give each Capability a nursery from the pool. No need to do atomic increments
 * here, everything must be stopped to call this function.
 */
static void
assignNurseriesToCapabilities (uint32_t from, uint32_t to)
{
    uint32_t i, node;

    for (i = from; i < to; i++) {
        node = capabilities[i]->node;
        assignNurseryToCapability(capabilities[i], next_nursery[node]);
        next_nursery[node] += n_numa_nodes;
    }
}

static void
allocNurseries (uint32_t from, uint32_t to)
{
    uint32_t i;
    memcount n_blocks;

    if (RtsFlags.GcFlags.nurseryChunkSize) {
        n_blocks = RtsFlags.GcFlags.nurseryChunkSize;
    } else {
        n_blocks = RtsFlags.GcFlags.minAllocAreaSize;
    }

    for (i = from; i < to; i++) {
        nurseries[i].blocks = allocNursery(capNoToNumaNode(i), NULL, n_blocks);
        nurseries[i].n_blocks = n_blocks;
    }
}

void
resetNurseries (void)
{
    uint32_t n;

    for (n = 0; n < n_numa_nodes; n++) {
        next_nursery[n] = n;
    }
    assignNurseriesToCapabilities(0, n_capabilities);

#if defined(DEBUG)
    bdescr *bd;
    for (n = 0; n < n_nurseries; n++) {
        for (bd = nurseries[n].blocks; bd; bd = bd->link) {
            ASSERT(bd->gen_no == 0);
            ASSERT(bd->gen == g0);
            ASSERT(bd->node == capNoToNumaNode(n));
            IF_DEBUG(sanity, memset(bd->start, 0xaa, BLOCK_SIZE));
        }
    }
#endif
}

W_
countNurseryBlocks (void)
{
    uint32_t i;
    W_ blocks = 0;

    for (i = 0; i < n_nurseries; i++) {
        blocks += nurseries[i].n_blocks;
    }
    return blocks;
}

//
// Resize each of the nurseries to the specified size.
//
static void
resizeNurseriesEach (W_ blocks)
{
    uint32_t i, node;
    bdescr *bd;
    W_ nursery_blocks;
    nursery *nursery;

    for (i = 0; i < n_nurseries; i++) {
        nursery = &nurseries[i];
        nursery_blocks = nursery->n_blocks;
        if (nursery_blocks == blocks) continue;

        node = capNoToNumaNode(i);
        if (nursery_blocks < blocks) {
            debugTrace(DEBUG_gc, "increasing size of nursery to %d blocks",
                       blocks);
            nursery->blocks = allocNursery(node, nursery->blocks,
                                           blocks-nursery_blocks);
        }
        else
        {
            bdescr *next_bd;

            debugTrace(DEBUG_gc, "decreasing size of nursery to %d blocks",
                       blocks);

            bd = nursery->blocks;
            while (nursery_blocks > blocks) {
                next_bd = bd->link;
                next_bd->u.back = NULL;
                nursery_blocks -= bd->blocks; // might be a large block
                freeGroup(bd);
                bd = next_bd;
            }
            nursery->blocks = bd;
            // might have gone just under, by freeing a large block, so make
            // up the difference.
            if (nursery_blocks < blocks) {
                nursery->blocks = allocNursery(node, nursery->blocks,
                                               blocks-nursery_blocks);
            }
        }
        nursery->n_blocks = blocks;
        ASSERT(countBlocks(nursery->blocks) == nursery->n_blocks);
    }
}

void
resizeNurseriesFixed (void)
{
    uint32_t blocks;

    if (RtsFlags.GcFlags.nurseryChunkSize) {
        blocks = RtsFlags.GcFlags.nurseryChunkSize;
    } else {
        blocks = RtsFlags.GcFlags.minAllocAreaSize;
    }

    resizeNurseriesEach(blocks);
}

//
// Resize the nurseries to the total specified size.
//
void
resizeNurseries (W_ blocks)
{
    // If there are multiple nurseries, then we just divide the number
    // of available blocks between them.
    resizeNurseriesEach(blocks / n_nurseries);
}

bool
getNewNursery (Capability *cap)
{
    StgWord i;
    uint32_t node = cap->node;
    uint32_t n;

    for(;;) {
        i = next_nursery[node];
        if (i < n_nurseries) {
            if (cas(&next_nursery[node], i, i+n_numa_nodes) == i) {
                assignNurseryToCapability(cap, i);
                return true;
            }
        } else if (n_numa_nodes > 1) {
            // Try to find an unused nursery chunk on other nodes.  We'll get
            // remote memory, but the rationale is that avoiding GC is better
            // than avoiding remote memory access.
            bool lost = false;
            for (n = 0; n < n_numa_nodes; n++) {
                if (n == node) continue;
                i = next_nursery[n];
                if (i < n_nurseries) {
                    if (cas(&next_nursery[n], i, i+n_numa_nodes) == i) {
                        assignNurseryToCapability(cap, i);
                        return true;
                    } else {
                        lost = true; /* lost a race */
                    }
                }
            }
            if (!lost) return false;
        } else {
            return false;
        }
    }
}

/* -----------------------------------------------------------------------------
   move_STACK is called to update the TSO structure after it has been
   moved from one place to another.
   -------------------------------------------------------------------------- */

void
move_STACK (StgStack *src, StgStack *dest)
{
    ptrdiff_t diff;

    // relocate the stack pointer...
    diff = (StgPtr)dest - (StgPtr)src; // In *words*
    dest->sp = (StgPtr)dest->sp + diff;
}

STATIC_INLINE void
accountAllocation(Capability *cap, W_ n)
{
    TICK_ALLOC_HEAP_NOCTR(WDS(n));
    CCS_ALLOC(cap->r.rCCCS,n);
    if (cap->r.rCurrentTSO != NULL) {
        // cap->r.rCurrentTSO->alloc_limit -= n*sizeof(W_)
        ASSIGN_Int64((W_*)&(cap->r.rCurrentTSO->alloc_limit),
                     (PK_Int64((W_*)&(cap->r.rCurrentTSO->alloc_limit))
                      - n*sizeof(W_)));
    }

}

/* -----------------------------------------------------------------------------
   StgPtr allocate (Capability *cap, W_ n)

   Allocates an area of memory n *words* large, from the nursery of
   the supplied Capability, or from the global block pool if the area
   requested is larger than LARGE_OBJECT_THRESHOLD.  Memory is not
   allocated from the current nursery block, so as not to interfere
   with Hp/HpLim.

   The address of the allocated memory is returned. allocate() never
   fails; if it returns, the returned value is a valid address.  If
   the nursery is already full, then another block is allocated from
   the global block pool.  If we need to get memory from the OS and
   that operation fails, then the whole process will be killed.
   -------------------------------------------------------------------------- */

/*
 * Allocate some n words of heap memory; terminating
 * on heap overflow
 */
StgPtr
allocate (Capability *cap, W_ n)
{
    StgPtr p = allocateMightFail(cap, n);
    if (p == NULL) {
        reportHeapOverflow();
        // heapOverflow() doesn't exit (see #2592), but we aren't
        // in a position to do a clean shutdown here: we
        // either have to allocate the memory or exit now.
        // Allocating the memory would be bad, because the user
        // has requested that we not exceed maxHeapSize, so we
        // just exit.
        stg_exit(EXIT_HEAPOVERFLOW);
    }
    return p;
}

/*
 * Allocate some n words of heap memory; returning NULL
 * on heap overflow
 */
StgPtr
allocateMightFail (Capability *cap, W_ n)
{
    bdescr *bd;
    StgPtr p;

    if (n >= LARGE_OBJECT_THRESHOLD/sizeof(W_)) {
        // The largest number of words such that
        // the computation of req_blocks will not overflow.
        W_ max_words = (HS_WORD_MAX & ~(BLOCK_SIZE-1)) / sizeof(W_);
        W_ req_blocks;

        if (n > max_words)
            req_blocks = HS_WORD_MAX; // signal overflow below
        else
            req_blocks = (W_)BLOCK_ROUND_UP(n*sizeof(W_)) / BLOCK_SIZE;

        // Attempting to allocate an object larger than maxHeapSize
        // should definitely be disallowed.  (bug #1791)
        if ((RtsFlags.GcFlags.maxHeapSize > 0 &&
             req_blocks >= RtsFlags.GcFlags.maxHeapSize) ||
            req_blocks >= HS_INT32_MAX)   // avoid overflow when
                                          // calling allocGroup() below
        {
            return NULL;
        }

        // Only credit allocation after we've passed the size check above
        accountAllocation(cap, n);

        ACQUIRE_SM_LOCK
        bd = allocGroupOnNode(cap->node,req_blocks);
        dbl_link_onto(bd, &g0->large_objects);
        g0->n_large_blocks += bd->blocks; // might be larger than req_blocks
        g0->n_new_large_words += n;
        RELEASE_SM_LOCK;
        initBdescr(bd, g0, g0);
        bd->flags = BF_LARGE;
        bd->free = bd->start + n;
        cap->total_allocated += n;
        return bd->start;
    }

    /* small allocation (<LARGE_OBJECT_THRESHOLD) */

    accountAllocation(cap, n);
    bd = cap->r.rCurrentAlloc;
    if (bd == NULL || bd->free + n > bd->start + BLOCK_SIZE_W) {

        if (bd) finishedNurseryBlock(cap,bd);

        // The CurrentAlloc block is full, we need to find another
        // one.  First, we try taking the next block from the
        // nursery:
        bd = cap->r.rCurrentNursery->link;

        if (bd == NULL) {
            // The nursery is empty: allocate a fresh block (we can't
            // fail here).
            ACQUIRE_SM_LOCK;
            bd = allocBlockOnNode(cap->node);
            cap->r.rNursery->n_blocks++;
            RELEASE_SM_LOCK;
            initBdescr(bd, g0, g0);
            bd->flags = 0;
            // If we had to allocate a new block, then we'll GC
            // pretty quickly now, because MAYBE_GC() will
            // notice that CurrentNursery->link is NULL.
        } else {
            newNurseryBlock(bd);
            // we have a block in the nursery: take it and put
            // it at the *front* of the nursery list, and use it
            // to allocate() from.
            //
            // Previously the nursery looked like this:
            //
            //           CurrentNursery
            //                  /
            //                +-+    +-+
            // nursery -> ... |A| -> |B| -> ...
            //                +-+    +-+
            //
            // After doing this, it looks like this:
            //
            //                      CurrentNursery
            //                            /
            //            +-+           +-+
            // nursery -> |B| -> ... -> |A| -> ...
            //            +-+           +-+
            //             |
            //             CurrentAlloc
            //
            // The point is to get the block out of the way of the
            // advancing CurrentNursery pointer, while keeping it
            // on the nursery list so we don't lose track of it.
            cap->r.rCurrentNursery->link = bd->link;
            if (bd->link != NULL) {
                bd->link->u.back = cap->r.rCurrentNursery;
            }
        }
        dbl_link_onto(bd, &cap->r.rNursery->blocks);
        cap->r.rCurrentAlloc = bd;
        IF_DEBUG(sanity, checkNurserySanity(cap->r.rNursery));
    }
    p = bd->free;
    bd->free += n;

    IF_DEBUG(sanity, ASSERT(*((StgWord8*)p) == 0xaa));
    return p;
}

/* ---------------------------------------------------------------------------
   Allocate a fixed/pinned object.

   We allocate small pinned objects into a single block, allocating a
   new block when the current one overflows.  The block is chained
   onto the large_object_list of generation 0.

   NOTE: The GC can't in general handle pinned objects.  This
   interface is only safe to use for ByteArrays, which have no
   pointers and don't require scavenging.  It works because the
   block's descriptor has the BF_LARGE flag set, so the block is
   treated as a large object and chained onto various lists, rather
   than the individual objects being copied.  However, when it comes
   to scavenge the block, the GC will only scavenge the first object.
   The reason is that the GC can't linearly scan a block of pinned
   objects at the moment (doing so would require using the
   mostly-copying techniques).  But since we're restricting ourselves
   to pinned ByteArrays, not scavenging is ok.

   This function is called by newPinnedByteArray# which immediately
   fills the allocated memory with a MutableByteArray#. Note that
   this returns NULL on heap overflow.
   ------------------------------------------------------------------------- */

StgPtr
allocatePinned (Capability *cap, W_ n)
{
    StgPtr p;
    bdescr *bd;

    // If the request is for a large object, then allocate()
    // will give us a pinned object anyway.
    if (n >= LARGE_OBJECT_THRESHOLD/sizeof(W_)) {
        p = allocateMightFail(cap, n);
        if (p == NULL) {
            return NULL;
        } else {
            Bdescr(p)->flags |= BF_PINNED;
            return p;
        }
    }

    accountAllocation(cap, n);
    bd = cap->pinned_object_block;

    // If we don't have a block of pinned objects yet, or the current
    // one isn't large enough to hold the new object, get a new one.
    if (bd == NULL || (bd->free + n) > (bd->start + BLOCK_SIZE_W)) {

        // stash the old block on cap->pinned_object_blocks.  On the
        // next GC cycle these objects will be moved to
        // g0->large_objects.
        if (bd != NULL) {
            // add it to the allocation stats when the block is full
            finishedNurseryBlock(cap, bd);
            dbl_link_onto(bd, &cap->pinned_object_blocks);
        }

        // We need to find another block.  We could just allocate one,
        // but that means taking a global lock and we really want to
        // avoid that (benchmarks that allocate a lot of pinned
        // objects scale really badly if we do this).
        //
        // So first, we try taking the next block from the nursery, in
        // the same way as allocate().
        bd = cap->r.rCurrentNursery->link;
        if (bd == NULL) {
            // The nursery is empty: allocate a fresh block (we can't fail
            // here).
            ACQUIRE_SM_LOCK;
            bd = allocBlockOnNode(cap->node);
            RELEASE_SM_LOCK;
            initBdescr(bd, g0, g0);
        } else {
            newNurseryBlock(bd);
            // we have a block in the nursery: steal it
            cap->r.rCurrentNursery->link = bd->link;
            if (bd->link != NULL) {
                bd->link->u.back = cap->r.rCurrentNursery;
            }
            cap->r.rNursery->n_blocks -= bd->blocks;
        }

        cap->pinned_object_block = bd;
        bd->flags  = BF_PINNED | BF_LARGE | BF_EVACUATED;

        // The pinned_object_block remains attached to the capability
        // until it is full, even if a GC occurs.  We want this
        // behaviour because otherwise the unallocated portion of the
        // block would be forever slop, and under certain workloads
        // (allocating a few ByteStrings per GC) we accumulate a lot
        // of slop.
        //
        // So, the pinned_object_block is initially marked
        // BF_EVACUATED so the GC won't touch it.  When it is full,
        // we place it on the large_objects list, and at the start of
        // the next GC the BF_EVACUATED flag will be cleared, and the
        // block will be promoted as usual (if anything in it is
        // live).
    }

    p = bd->free;
    bd->free += n;
    return p;
}

/* -----------------------------------------------------------------------------
   Write Barriers
   -------------------------------------------------------------------------- */

/*
   This is the write barrier for MUT_VARs, a.k.a. IORefs.  A
   MUT_VAR_CLEAN object is not on the mutable list; a MUT_VAR_DIRTY
   is.  When written to, a MUT_VAR_CLEAN turns into a MUT_VAR_DIRTY
   and is put on the mutable list.
*/
void
dirty_MUT_VAR(StgRegTable *reg, StgClosure *p)
{
    Capability *cap = regTableToCapability(reg);
    if (p->header.info == &stg_MUT_VAR_CLEAN_info) {
        p->header.info = &stg_MUT_VAR_DIRTY_info;
        recordClosureMutated(cap,p);
    }
}

void
dirty_TVAR(Capability *cap, StgTVar *p)
{
    if (p->header.info == &stg_TVAR_CLEAN_info) {
        p->header.info = &stg_TVAR_DIRTY_info;
        recordClosureMutated(cap,(StgClosure*)p);
    }
}

// Setting a TSO's link field with a write barrier.
// It is *not* necessary to call this function when
//    * setting the link field to END_TSO_QUEUE
//    * setting the link field of the currently running TSO, as it
//      will already be dirty.
void
setTSOLink (Capability *cap, StgTSO *tso, StgTSO *target)
{
    if (tso->dirty == 0) {
        tso->dirty = 1;
        recordClosureMutated(cap,(StgClosure*)tso);
    }
    tso->_link = target;
}

void
setTSOPrev (Capability *cap, StgTSO *tso, StgTSO *target)
{
    if (tso->dirty == 0) {
        tso->dirty = 1;
        recordClosureMutated(cap,(StgClosure*)tso);
    }
    tso->block_info.prev = target;
}

void
dirty_TSO (Capability *cap, StgTSO *tso)
{
    if (tso->dirty == 0) {
        tso->dirty = 1;
        recordClosureMutated(cap,(StgClosure*)tso);
    }
}

void
dirty_STACK (Capability *cap, StgStack *stack)
{
    if (stack->dirty == 0) {
        stack->dirty = 1;
        recordClosureMutated(cap,(StgClosure*)stack);
    }
}

/*
   This is the write barrier for MVARs.  An MVAR_CLEAN objects is not
   on the mutable list; a MVAR_DIRTY is.  When written to, a
   MVAR_CLEAN turns into a MVAR_DIRTY and is put on the mutable list.
   The check for MVAR_CLEAN is inlined at the call site for speed,
   this really does make a difference on concurrency-heavy benchmarks
   such as Chaneneos and cheap-concurrency.
*/
void
dirty_MVAR(StgRegTable *reg, StgClosure *p)
{
    recordClosureMutated(regTableToCapability(reg),p);
}

/* -----------------------------------------------------------------------------
 * Stats and stuff
 * -------------------------------------------------------------------------- */

/* -----------------------------------------------------------------------------
 * Note [allocation accounting]
 *
 *   - When cap->r.rCurrentNusery moves to a new block in the nursery,
 *     we add the size of the used portion of the previous block to
 *     cap->total_allocated. (see finishedNurseryBlock())
 *
 *   - When we start a GC, the allocated portion of CurrentNursery and
 *     CurrentAlloc are added to cap->total_allocated. (see
 *     updateNurseriesStats())
 *
 * -------------------------------------------------------------------------- */

//
// Calculate the total allocated memory since the start of the
// program.  Also emits events reporting the per-cap allocation
// totals.
//
StgWord
calcTotalAllocated (void)
{
    W_ tot_alloc = 0;
    W_ n;

    for (n = 0; n < n_capabilities; n++) {
        tot_alloc += capabilities[n]->total_allocated;

        traceEventHeapAllocated(capabilities[n],
                                CAPSET_HEAP_DEFAULT,
                                capabilities[n]->total_allocated * sizeof(W_));
    }

    return tot_alloc;
}

//
// Update the per-cap total_allocated numbers with an approximation of
// the amount of memory used in each cap's nursery.
//
void
updateNurseriesStats (void)
{
    uint32_t i;
    bdescr *bd;

    for (i = 0; i < n_capabilities; i++) {
        // The current nursery block and the current allocate block have not
        // yet been accounted for in cap->total_allocated, so we add them here.
        bd = capabilities[i]->r.rCurrentNursery;
        if (bd) finishedNurseryBlock(capabilities[i], bd);
        bd = capabilities[i]->r.rCurrentAlloc;
        if (bd) finishedNurseryBlock(capabilities[i], bd);
    }
}

W_ countOccupied (bdescr *bd)
{
    W_ words;

    words = 0;
    for (; bd != NULL; bd = bd->link) {
        ASSERT(bd->free <= bd->start + bd->blocks * BLOCK_SIZE_W);
        words += bd->free - bd->start;
    }
    return words;
}

W_ genLiveWords (generation *gen)
{
    return gen->n_words + gen->n_large_words +
        gen->n_compact_blocks * BLOCK_SIZE_W;
}

W_ genLiveBlocks (generation *gen)
{
    return gen->n_blocks + gen->n_large_blocks + gen->n_compact_blocks;
}

W_ gcThreadLiveWords (uint32_t i, uint32_t g)
{
    W_ a, b, c;

    a = countOccupied(gc_threads[i]->gens[g].todo_bd);
    b = gc_threads[i]->gens[g].n_part_words;
    c = gc_threads[i]->gens[g].n_scavd_words;

//    debugBelch("cap %d, g%d, %ld %ld %ld\n", i, g, a, b, c);
    return a + b + c;
}

W_ gcThreadLiveBlocks (uint32_t i, uint32_t g)
{
    W_ blocks;

    blocks  = countBlocks(gc_threads[i]->gens[g].todo_bd);
    blocks += gc_threads[i]->gens[g].n_part_blocks;
    blocks += gc_threads[i]->gens[g].n_scavd_blocks;

    return blocks;
}

/* Determine which generation will be collected next, and approximate
 * the maximum amount of memory that will be required to do the GC,
 * taking into account data that will be copied, and the space needed
 * to store bitmaps and the mark stack.  Note: blocks_needed does not
 * include the blocks in the nursery.
 *
 * Assume: all data currently live will remain live.  Generations
 * that will be collected next time will therefore need twice as many
 * blocks since all the data will be copied.
 */
extern W_
calcNeeded (bool force_major, memcount *blocks_needed)
{
    W_ needed = 0;
    uint32_t N;

    if (force_major) {
        N = RtsFlags.GcFlags.generations - 1;
    } else {
        N = 0;
    }

    for (uint32_t g = 0; g < RtsFlags.GcFlags.generations; g++) {
        generation *gen = &generations[g];

        W_ blocks = gen->n_blocks // or: gen->n_words / BLOCK_SIZE_W (?)
                  + gen->n_large_blocks
                  + gen->n_compact_blocks;

        // we need at least this much space
        needed += blocks;

        // are we collecting this gen?
        if (g == 0 || // always collect gen 0
            blocks > gen->max_blocks)
        {
            N = stg_max(N,g);

            // we will collect this gen next time
            if (gen->mark) {
                //  bitmap:
                needed += gen->n_blocks / BITS_IN(W_);
                //  mark stack:
                needed += gen->n_blocks / 100;
            }
            if (gen->compact) {
                continue; // no additional space needed for compaction
            } else {
                needed += gen->n_blocks;
            }
        }
    }

    if (blocks_needed != NULL) {
        *blocks_needed = needed;
    }
    return N;
}

StgWord calcTotalLargeObjectsW (void)
{
    uint32_t g;
    StgWord totalW = 0;

    for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
        totalW += generations[g].n_large_words;
    }
    return totalW;
}

StgWord calcTotalCompactW (void)
{
    uint32_t g;
    StgWord totalW = 0;

    for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
        totalW += generations[g].n_compact_blocks * BLOCK_SIZE_W;
    }
    return totalW;
}

/* ----------------------------------------------------------------------------
   Executable memory

   Executable memory must be managed separately from non-executable
   memory.  Most OSs these days require you to jump through hoops to
   dynamically allocate executable memory, due to various security
   measures.

   Here we provide a small memory allocator for executable memory.
   Memory is managed with a page granularity; we allocate linearly
   in the page, and when the page is emptied (all objects on the page
   are free) we free the page again, not forgetting to make it
   non-executable.

   TODO: The inability to handle objects bigger than BLOCK_SIZE_W means that
         the linker cannot use allocateExec for loading object code files
         on Windows. Once allocateExec can handle larger objects, the linker
         should be modified to use allocateExec instead of VirtualAlloc.
   ------------------------------------------------------------------------- */

#if (defined(arm_HOST_ARCH) || defined(aarch64_HOST_ARCH)) && defined(ios_HOST_OS)
#include <libkern/OSCacheControl.h>
#endif

#if defined(__clang__)
/* clang defines __clear_cache as a builtin on some platforms.
 * For example on armv7-linux-androideabi. The type slightly
 * differs from gcc.
 */
extern void __clear_cache(void * begin, void * end);
#elif defined(__GNUC__)
/* __clear_cache is a libgcc function.
 * It existed before __builtin___clear_cache was introduced.
 * See Trac #8562.
 */
extern void __clear_cache(char * begin, char * end);
#endif /* __GNUC__ */

/* On ARM and other platforms, we need to flush the cache after
   writing code into memory, so the processor reliably sees it. */
void flushExec (W_ len, AdjustorExecutable exec_addr)
{
#if defined(i386_HOST_ARCH) || defined(x86_64_HOST_ARCH)
  /* x86 doesn't need to do anything, so just suppress some warnings. */
  (void)len;
  (void)exec_addr;
#elif (defined(arm_HOST_ARCH) || defined(aarch64_HOST_ARCH)) && defined(ios_HOST_OS)
  /* On iOS we need to use the special 'sys_icache_invalidate' call. */
  sys_icache_invalidate(exec_addr, len);
#elif defined(__clang__)
  unsigned char* begin = (unsigned char*)exec_addr;
  unsigned char* end   = begin + len;
# if __has_builtin(__builtin___clear_cache)
  __builtin___clear_cache((void*)begin, (void*)end);
# else
  __clear_cache((void*)begin, (void*)end);
# endif
#elif defined(__GNUC__)
  /* For all other platforms, fall back to a libgcc builtin. */
  unsigned char* begin = (unsigned char*)exec_addr;
  unsigned char* end   = begin + len;
  /* __builtin___clear_cache is supported since GNU C 4.3.6.
   * We pick 4.4 to simplify condition a bit.
   */
# if __GNUC__ > 4 || (__GNUC__ == 4 && __GNUC_MINOR__ >= 4)
  __builtin___clear_cache((void*)begin, (void*)end);
# else
  __clear_cache((void*)begin, (void*)end);
# endif
#else
#error Missing support to flush the instruction cache
#endif
}

#if defined(linux_HOST_OS)

// On Linux we need to use libffi for allocating executable memory,
// because it knows how to work around the restrictions put in place
// by SELinux.

AdjustorWritable allocateExec (W_ bytes, AdjustorExecutable *exec_ret)
{
    void **ret, **exec;
    ACQUIRE_SM_LOCK;
    ret = ffi_closure_alloc (sizeof(void *) + (size_t)bytes, (void**)&exec);
    RELEASE_SM_LOCK;
    if (ret == NULL) return ret;
    *ret = ret; // save the address of the writable mapping, for freeExec().
    *exec_ret = exec + 1;
    return (ret + 1);
}

// freeExec gets passed the executable address, not the writable address.
void freeExec (AdjustorExecutable addr)
{
    AdjustorWritable writable;
    writable = *((void**)addr - 1);
    ACQUIRE_SM_LOCK;
    ffi_closure_free (writable);
    RELEASE_SM_LOCK
}

#elif defined(ios_HOST_OS)

static HashTable* allocatedExecs;

AdjustorWritable allocateExec(W_ bytes, AdjustorExecutable *exec_ret)
{
    AdjustorWritable writ;
    ffi_closure* cl;
    if (bytes != sizeof(ffi_closure)) {
        barf("allocateExec: for ffi_closure only");
    }
    ACQUIRE_SM_LOCK;
    cl = writ = ffi_closure_alloc((size_t)bytes, exec_ret);
    if (cl != NULL) {
        if (allocatedExecs == NULL) {
            allocatedExecs = allocHashTable();
        }
        insertHashTable(allocatedExecs, (StgWord)*exec_ret, writ);
    }
    RELEASE_SM_LOCK;
    return writ;
}

AdjustorWritable execToWritable(AdjustorExecutable exec)
{
    AdjustorWritable writ;
    ACQUIRE_SM_LOCK;
    if (allocatedExecs == NULL ||
       (writ = lookupHashTable(allocatedExecs, (StgWord)exec)) == NULL) {
        RELEASE_SM_LOCK;
        barf("execToWritable: not found");
    }
    RELEASE_SM_LOCK;
    return writ;
}

void freeExec(AdjustorExecutable exec)
{
    AdjustorWritable writ;
    ffi_closure* cl;
    cl = writ = execToWritable(exec);
    ACQUIRE_SM_LOCK;
    removeHashTable(allocatedExecs, (StgWord)exec, writ);
    ffi_closure_free(cl);
    RELEASE_SM_LOCK
}

#else

AdjustorWritable allocateExec (W_ bytes, AdjustorExecutable *exec_ret)
{
    void *ret;
    W_ n;

    ACQUIRE_SM_LOCK;

    // round up to words.
    n  = (bytes + sizeof(W_) + 1) / sizeof(W_);

    if (n+1 > BLOCK_SIZE_W) {
        barf("allocateExec: can't handle large objects");
    }

    if (exec_block == NULL ||
        exec_block->free + n + 1 > exec_block->start + BLOCK_SIZE_W) {
        bdescr *bd;
        W_ pagesize = getPageSize();
        bd = allocGroup(stg_max(1, pagesize / BLOCK_SIZE));
        debugTrace(DEBUG_gc, "allocate exec block %p", bd->start);
        bd->gen_no = 0;
        bd->flags = BF_EXEC;
        bd->link = exec_block;
        if (exec_block != NULL) {
            exec_block->u.back = bd;
        }
        bd->u.back = NULL;
        setExecutable(bd->start, bd->blocks * BLOCK_SIZE, true);
        exec_block = bd;
    }
    *(exec_block->free) = n;  // store the size of this chunk
    exec_block->gen_no += n;  // gen_no stores the number of words allocated
    ret = exec_block->free + 1;
    exec_block->free += n + 1;

    RELEASE_SM_LOCK
    *exec_ret = ret;
    return ret;
}

void freeExec (void *addr)
{
    StgPtr p = (StgPtr)addr - 1;
    bdescr *bd = Bdescr((StgPtr)p);

    if ((bd->flags & BF_EXEC) == 0) {
        barf("freeExec: not executable");
    }

    if (*(StgPtr)p == 0) {
        barf("freeExec: already free?");
    }

    ACQUIRE_SM_LOCK;

    bd->gen_no -= *(StgPtr)p;
    *(StgPtr)p = 0;

    if (bd->gen_no == 0) {
        // Free the block if it is empty, but not if it is the block at
        // the head of the queue.
        if (bd != exec_block) {
            debugTrace(DEBUG_gc, "free exec block %p", bd->start);
            dbl_link_remove(bd, &exec_block);
            setExecutable(bd->start, bd->blocks * BLOCK_SIZE, false);
            freeGroup(bd);
        } else {
            bd->free = bd->start;
        }
    }

    RELEASE_SM_LOCK
}

#endif /* switch(HOST_OS) */

#if defined(DEBUG)

// handy function for use in gdb, because Bdescr() is inlined.
extern bdescr *_bdescr (StgPtr p);

bdescr *
_bdescr (StgPtr p)
{
    return Bdescr(p);
}

#endif