Storage.c 24.7 KB
Newer Older
1
/* -----------------------------------------------------------------------------
sof's avatar
sof committed
2
 * $Id: Storage.c,v 1.58 2002/01/24 01:45:55 sof Exp $
3
4
 *
 * (c) The GHC Team, 1998-1999
5
6
7
8
9
 *
 * Storage manager front end
 *
 * ---------------------------------------------------------------------------*/

10
#include "PosixSource.h"
11
12
13
14
15
16
#include "Rts.h"
#include "RtsUtils.h"
#include "RtsFlags.h"
#include "Stats.h"
#include "Hooks.h"
#include "BlockAlloc.h"
17
#include "MBlock.h"
18
#include "Weak.h"
19
#include "Sanity.h"
20
#include "Arena.h"
21
22

#include "Storage.h"
23
#include "Schedule.h"
24
25
#include "StoragePriv.h"

26
27
#include "RetainerProfile.h"	// for counting memory blocks (memInventory)

28
29
30
31
StgClosure    *caf_list         = NULL;

bdescr *small_alloc_list;	/* allocate()d small objects */
bdescr *large_alloc_list;	/* allocate()d large objects */
32
bdescr *pinned_object_block;    /* allocate pinned objects into this block */
33
34
35
36
37
38
nat alloc_blocks;		/* number of allocate()d blocks since GC */
nat alloc_blocks_lim;		/* approximate limit on alloc_blocks */

StgPtr alloc_Hp    = NULL;	/* next free byte in small_alloc_list */
StgPtr alloc_HpLim = NULL;	/* end of block at small_alloc_list   */

39
40
41
42
43
generation *generations;	/* all the generations */
generation *g0;			/* generation 0, for convenience */
generation *oldest_gen;		/* oldest generation, for convenience */
step *g0s0;			/* generation 0, step 0, for convenience */

44
45
lnat total_allocated = 0;	/* total memory allocated during run */

46
47
48
49
50
51
52
53
/*
 * Storage manager mutex:  protects all the above state from
 * simultaneous access by two STG threads.
 */
#ifdef SMP
pthread_mutex_t sm_mutex = PTHREAD_MUTEX_INITIALIZER;
#endif

54
55
56
57
58
59
60
61
/*
 * Forward references
 */
static void *stgAllocForGMP   (size_t size_in_bytes);
static void *stgReallocForGMP (void *ptr, size_t old_size, size_t new_size);
static void  stgDeallocForGMP (void *ptr, size_t size);

void
62
initStorage( void )
63
{
64
  nat g, s;
65
  step *stp;
66
  generation *gen;
67

68
69
  if (RtsFlags.GcFlags.maxHeapSize != 0 &&
      RtsFlags.GcFlags.heapSizeSuggestion > 
70
      RtsFlags.GcFlags.maxHeapSize) {
71
    RtsFlags.GcFlags.maxHeapSize = RtsFlags.GcFlags.heapSizeSuggestion;
72
73
  }

74
75
76
77
78
79
80
  if (RtsFlags.GcFlags.maxHeapSize != 0 &&
      RtsFlags.GcFlags.minAllocAreaSize > 
      RtsFlags.GcFlags.maxHeapSize) {
      prog_belch("maximum heap size (-M) is smaller than minimum alloc area size (-A)");
      exit(1);
  }

81
82
  initBlockAllocator();
  
83
84
85
86
87
  /* allocate generation info array */
  generations = (generation *)stgMallocBytes(RtsFlags.GcFlags.generations 
					     * sizeof(struct _generation),
					     "initStorage: gens");

88
  /* Initialise all generations */
89
  for(g = 0; g < RtsFlags.GcFlags.generations; g++) {
90
91
92
    gen = &generations[g];
    gen->no = g;
    gen->mut_list = END_MUT_LIST;
93
    gen->mut_once_list = END_MUT_LIST;
94
95
    gen->collections = 0;
    gen->failed_promotions = 0;
96
    gen->max_blocks = 0;
97
98
  }

99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
  /* A couple of convenience pointers */
  g0 = &generations[0];
  oldest_gen = &generations[RtsFlags.GcFlags.generations-1];

  /* Allocate step structures in each generation */
  if (RtsFlags.GcFlags.generations > 1) {
    /* Only for multiple-generations */

    /* Oldest generation: one step */
    oldest_gen->n_steps = 1;
    oldest_gen->steps = 
      stgMallocBytes(1 * sizeof(struct _step), "initStorage: last step");

    /* set up all except the oldest generation with 2 steps */
    for(g = 0; g < RtsFlags.GcFlags.generations-1; g++) {
114
115
116
117
      generations[g].n_steps = RtsFlags.GcFlags.steps;
      generations[g].steps  = 
	stgMallocBytes (RtsFlags.GcFlags.steps * sizeof(struct _step),
			"initStorage: steps");
118
119
120
121
122
123
    }
    
  } else {
    /* single generation, i.e. a two-space collector */
    g0->n_steps = 1;
    g0->steps = stgMallocBytes (sizeof(struct _step), "initStorage: steps");
124
125
  }

126
127
  /* Initialise all steps */
  for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
128
    for (s = 0; s < generations[g].n_steps; s++) {
129
130
131
132
133
      stp = &generations[g].steps[s];
      stp->no = s;
      stp->blocks = NULL;
      stp->n_blocks = 0;
      stp->gen = &generations[g];
134
      stp->gen_no = g;
135
136
137
138
139
140
      stp->hp = NULL;
      stp->hpLim = NULL;
      stp->hp_bd = NULL;
      stp->scan = NULL;
      stp->scan_bd = NULL;
      stp->large_objects = NULL;
141
      stp->n_large_blocks = 0;
142
143
      stp->new_large_objects = NULL;
      stp->scavenged_large_objects = NULL;
144
      stp->n_scavenged_large_blocks = 0;
145
      stp->is_compacted = 0;
146
      stp->bitmap = NULL;
147
148
149
    }
  }
  
150
151
  /* Set up the destination pointers in each younger gen. step */
  for (g = 0; g < RtsFlags.GcFlags.generations-1; g++) {
152
153
    for (s = 0; s < generations[g].n_steps-1; s++) {
      generations[g].steps[s].to = &generations[g].steps[s+1];
154
    }
155
    generations[g].steps[s].to = &generations[g+1].steps[0];
156
157
  }
  
158
159
  /* The oldest generation has one step and it is compacted. */
  if (RtsFlags.GcFlags.compact) {
160
161
162
163
164
      if (RtsFlags.GcFlags.generations == 1) {
	  belch("WARNING: compaction is incompatible with -G1; disabled");
      } else {
	  oldest_gen->steps[0].is_compacted = 1;
      }
165
  }
166
  oldest_gen->steps[0].to = &oldest_gen->steps[0];
167
168
169
170

  /* generation 0 is special: that's the nursery */
  generations[0].max_blocks = 0;

171
172
173
174
175
176
  /* G0S0: 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.
   */
177
178
179
  g0s0 = &generations[0].steps[0];

  allocNurseries();
180
181
182
183
184
185
186
187
188
189
190
191

  weak_ptr_list = NULL;
  caf_list = NULL;
   
  /* initialise the allocate() interface */
  small_alloc_list = NULL;
  large_alloc_list = NULL;
  alloc_blocks = 0;
  alloc_blocks_lim = RtsFlags.GcFlags.minAllocAreaSize;

  /* Tell GNU multi-precision pkg about our custom alloc functions */
  mp_set_memory_functions(stgAllocForGMP, stgReallocForGMP, stgDeallocForGMP);
192

193
194
195
196
#ifdef SMP
  pthread_mutex_init(&sm_mutex, NULL);
#endif

197
  IF_DEBUG(gc, statDescribeGens());
198
199
}

200
201
202
void
exitStorage (void)
{
203
    stat_exit(calcAllocated());
204
205
}

206
207
/* -----------------------------------------------------------------------------
   CAF management.
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

   The entry code for every CAF does the following:
     
      - builds a CAF_BLACKHOLE in the heap
      - pushes an update frame pointing to the CAF_BLACKHOLE
      - invokes UPD_CAF(), which:
          - calls newCaf, below
	  - updates the CAF with a static indirection 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.

   newCaf() does the following:
       
      - it puts the CAF on the oldest generation's mut-once list.
        This is so that we can treat the CAF as a root when collecting
	younger generations.

   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

245
246
   -------------------------------------------------------------------------- */

247
248
249
250
251
252
253
254
255
256
257
void
newCAF(StgClosure* caf)
{
  /* Put this CAF on the mutable list for the old generation.
   * This is a HACK - the IND_STATIC closure doesn't really have
   * a mut_link field, but we pretend it has - in fact we re-use
   * the STATIC_LINK field for the time being, because when we
   * come to do a major GC we won't need the mut_link field
   * any more and can use it as a STATIC_LINK.
   */
  ACQUIRE_LOCK(&sm_mutex);
258

259
  if (is_dynamically_loaded_rwdata_ptr((StgPtr)caf)) {
260
261
      ((StgIndStatic *)caf)->saved_info  = (StgInfoTable *)caf->header.info;
      ((StgIndStatic *)caf)->static_link = caf_list;
262
      caf_list = caf;
263
264
265
266
  } else {
      ((StgIndStatic *)caf)->saved_info = NULL;
      ((StgMutClosure *)caf)->mut_link = oldest_gen->mut_once_list;
      oldest_gen->mut_once_list = (StgMutClosure *)caf;
267
  }
268

269
  RELEASE_LOCK(&sm_mutex);
270
271
272
273
274
275
276
277

#ifdef PAR
  /* If we are PAR or DIST then  we never forget a CAF */
  { globalAddr *newGA;
    //belch("<##> Globalising CAF %08x %s",caf,info_type(caf));
    newGA=makeGlobal(caf,rtsTrue); /*given full weight*/
    ASSERT(newGA);
  } 
sof's avatar
sof committed
278
#endif /* PAR */
279
280
281
282
283
284
285
286
287
288
289
290
}

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

void
allocNurseries( void )
{ 
#ifdef SMP
  {
    Capability *cap;
291
292
    bdescr *bd;

293
294
295
    g0s0->blocks = NULL;
    g0s0->n_blocks = 0;
    for (cap = free_capabilities; cap != NULL; cap = cap->link) {
sof's avatar
sof committed
296
297
298
      cap->r.rNursery = allocNursery(NULL, RtsFlags.GcFlags.minAllocAreaSize);
      cap->r.rCurrentNursery = cap->r.rNursery;
      for (bd = cap->r.rNursery; bd != NULL; bd = bd->link) {
299
	bd->u.back = (bdescr *)cap;
300
      }
301
    }
302
303
304
    /* Set the back links to be equal to the Capability,
     * so we can do slightly better informed locking.
     */
305
306
  }
#else /* SMP */
307
308
  g0s0->blocks      = allocNursery(NULL, RtsFlags.GcFlags.minAllocAreaSize);
  g0s0->n_blocks    = RtsFlags.GcFlags.minAllocAreaSize;
309
310
  g0s0->to_blocks   = NULL;
  g0s0->n_to_blocks = 0;
311
312
  MainCapability.r.rNursery        = g0s0->blocks;
  MainCapability.r.rCurrentNursery = g0s0->blocks;
313
314
315
316
317
318
319
320
321
322
323
324
  /* hp, hpLim, hp_bd, to_space etc. aren't used in G0S0 */
#endif
}
      
void
resetNurseries( void )
{
  bdescr *bd;
#ifdef SMP
  Capability *cap;
  
  /* All tasks must be stopped */
325
  ASSERT(n_free_capabilities == RtsFlags.ParFlags.nNodes);
326
327

  for (cap = free_capabilities; cap != NULL; cap = cap->link) {
sof's avatar
sof committed
328
    for (bd = cap->r.rNursery; bd; bd = bd->link) {
329
      bd->free = bd->start;
330
      ASSERT(bd->gen_no == 0);
331
332
333
      ASSERT(bd->step == g0s0);
      IF_DEBUG(sanity,memset(bd->start, 0xaa, BLOCK_SIZE));
    }
sof's avatar
sof committed
334
    cap->r.rCurrentNursery = cap->r.rNursery;
335
336
337
338
  }
#else
  for (bd = g0s0->blocks; bd; bd = bd->link) {
    bd->free = bd->start;
339
    ASSERT(bd->gen_no == 0);
340
341
342
    ASSERT(bd->step == g0s0);
    IF_DEBUG(sanity,memset(bd->start, 0xaa, BLOCK_SIZE));
  }
343
344
  MainCapability.r.rNursery = g0s0->blocks;
  MainCapability.r.rCurrentNursery = g0s0->blocks;
345
346
347
348
#endif
}

bdescr *
349
allocNursery (bdescr *tail, nat blocks)
350
{
351
  bdescr *bd;
352
353
  nat i;

354
355
356
  // Allocate a nursery: we allocate fresh blocks one at a time and
  // cons them on to the front of the list, not forgetting to update
  // the back pointer on the tail of the list to point to the new block.
357
  for (i=0; i < blocks; i++) {
358
359
360
361
362
363
    // @LDV profiling
    /*
      processNursery() in LdvProfile.c assumes that every block group in
      the nursery contains only a single block. So, if a block group is
      given multiple blocks, change processNursery() accordingly.
     */
364
    bd = allocBlock();
365
366
367
368
369
    bd->link = tail;
    // double-link the nursery: we might need to insert blocks
    if (tail != NULL) {
	tail->u.back = bd;
    }
370
    bd->step = g0s0;
371
    bd->gen_no = 0;
372
    bd->flags = 0;
373
    bd->free = bd->start;
374
    tail = bd;
375
  }
376
377
  tail->u.back = NULL;
  return tail;
378
379
}

380
void
381
382
383
resizeNursery ( nat blocks )
{
  bdescr *bd;
384
  nat nursery_blocks;
385

386
387
388
389
#ifdef SMP
  barf("resizeNursery: can't resize in SMP mode");
#endif

390
  nursery_blocks = g0s0->n_blocks;
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
  if (nursery_blocks == blocks) {
    return;
  }

  else if (nursery_blocks < blocks) {
    IF_DEBUG(gc, fprintf(stderr, "Increasing size of nursery to %d blocks\n", 
			 blocks));
    g0s0->blocks = allocNursery(g0s0->blocks, blocks-nursery_blocks);
  } 

  else {
    bdescr *next_bd;
    
    IF_DEBUG(gc, fprintf(stderr, "Decreasing size of nursery to %d blocks\n", 
			 blocks));
406
407
408
409
410
411
412
413

    bd = g0s0->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;
414
415
    }
    g0s0->blocks = bd;
416
417
418
419
420
    // might have gone just under, by freeing a large block, so make
    // up the difference.
    if (nursery_blocks < blocks) {
	g0s0->blocks = allocNursery(g0s0->blocks, blocks-nursery_blocks);
    }
421
422
  }
  
423
424
  g0s0->n_blocks = blocks;
  ASSERT(countBlocks(g0s0->blocks) == g0s0->n_blocks);
425
426
}

427
428
429
430
431
432
433
434
435
/* -----------------------------------------------------------------------------
   The allocate() interface

   allocate(n) always succeeds, and returns a chunk of memory n words
   long.  n can be larger than the size of a block if necessary, in
   which case a contiguous block group will be allocated.
   -------------------------------------------------------------------------- */

StgPtr
436
allocate( nat n )
437
438
439
440
{
  bdescr *bd;
  StgPtr p;

441
442
  ACQUIRE_LOCK(&sm_mutex);

443
  TICK_ALLOC_HEAP_NOCTR(n);
444
445
446
  CCS_ALLOC(CCCS,n);

  /* big allocation (>LARGE_OBJECT_THRESHOLD) */
447
  /* ToDo: allocate directly into generation 1 */
448
449
450
  if (n >= LARGE_OBJECT_THRESHOLD/sizeof(W_)) {
    nat req_blocks =  (lnat)BLOCK_ROUND_UP(n*sizeof(W_)) / BLOCK_SIZE;
    bd = allocGroup(req_blocks);
451
    dbl_link_onto(bd, &g0s0->large_objects);
452
    bd->gen_no  = 0;
453
    bd->step = g0s0;
454
    bd->flags = BF_LARGE;
455
    bd->free = bd->start;
456
457
458
459
460
    /* don't add these blocks to alloc_blocks, since we're assuming
     * that large objects are likely to remain live for quite a while
     * (eg. running threads), so garbage collecting early won't make
     * much difference.
     */
461
    alloc_blocks += req_blocks;
462
    RELEASE_LOCK(&sm_mutex);
463
464
465
466
467
468
469
470
471
472
    return bd->start;

  /* small allocation (<LARGE_OBJECT_THRESHOLD) */
  } else if (small_alloc_list == NULL || alloc_Hp + n > alloc_HpLim) {
    if (small_alloc_list) {
      small_alloc_list->free = alloc_Hp;
    }
    bd = allocBlock();
    bd->link = small_alloc_list;
    small_alloc_list = bd;
473
    bd->gen_no = 0;
474
    bd->step = g0s0;
475
    bd->flags = 0;
476
477
478
479
    alloc_Hp = bd->start;
    alloc_HpLim = bd->start + BLOCK_SIZE_W;
    alloc_blocks++;
  }
480

481
482
  p = alloc_Hp;
  alloc_Hp += n;
483
  RELEASE_LOCK(&sm_mutex);
484
485
486
  return p;
}

487
488
lnat
allocated_bytes( void )
489
490
491
492
{
  return (alloc_blocks * BLOCK_SIZE_W - (alloc_HpLim - alloc_Hp));
}

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
/* ---------------------------------------------------------------------------
   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 step 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#.
   ------------------------------------------------------------------------- */

StgPtr
allocatePinned( nat n )
{
    StgPtr p;
    bdescr *bd = pinned_object_block;

    ACQUIRE_LOCK(&sm_mutex);
    
    TICK_ALLOC_HEAP_NOCTR(n);
    CCS_ALLOC(CCCS,n);

    // If the request is for a large object, then allocate()
    // will give us a pinned object anyway.
    if (n >= LARGE_OBJECT_THRESHOLD/sizeof(W_)) {
	RELEASE_LOCK(&sm_mutex);
	return allocate(n);
    }

    // If we don't have a block of pinned objects yet, or the current
    // one isn't large enough to hold the new object, allocate a new one.
    if (bd == NULL || (bd->free + n) > (bd->start + BLOCK_SIZE_W)) {
	pinned_object_block = bd = allocBlock();
	dbl_link_onto(bd, &g0s0->large_objects);
	bd->gen_no = 0;
	bd->step   = g0s0;
	bd->flags  = BF_LARGE;
	bd->free   = bd->start;
	alloc_blocks++;
    }

    p = bd->free;
    bd->free += n;
    RELEASE_LOCK(&sm_mutex);
    return p;
}

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
/* -----------------------------------------------------------------------------
   Allocation functions for GMP.

   These all use the allocate() interface - we can't have any garbage
   collection going on during a gmp operation, so we use allocate()
   which always succeeds.  The gmp operations which might need to
   allocate will ask the storage manager (via doYouWantToGC()) whether
   a garbage collection is required, in case we get into a loop doing
   only allocate() style allocation.
   -------------------------------------------------------------------------- */

static void *
stgAllocForGMP (size_t size_in_bytes)
{
  StgArrWords* arr;
  nat data_size_in_words, total_size_in_words;
  
  /* should be a multiple of sizeof(StgWord) (whole no. of limbs) */
  ASSERT(size_in_bytes % sizeof(W_) == 0);
  
  data_size_in_words  = size_in_bytes / sizeof(W_);
  total_size_in_words = sizeofW(StgArrWords) + data_size_in_words;
  
  /* allocate and fill it in. */
  arr = (StgArrWords *)allocate(total_size_in_words);
577
  SET_ARR_HDR(arr, &stg_ARR_WORDS_info, CCCS, data_size_in_words);
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
  
  /* and return a ptr to the goods inside the array */
  return(BYTE_ARR_CTS(arr));
}

static void *
stgReallocForGMP (void *ptr, size_t old_size, size_t new_size)
{
    void *new_stuff_ptr = stgAllocForGMP(new_size);
    nat i = 0;
    char *p = (char *) ptr;
    char *q = (char *) new_stuff_ptr;

    for (; i < old_size; i++, p++, q++) {
	*q = *p;
    }

    return(new_stuff_ptr);
}

static void
stgDeallocForGMP (void *ptr STG_UNUSED, 
		  size_t size STG_UNUSED)
{
    /* easy for us: the garbage collector does the dealloc'n */
}
604

605
/* -----------------------------------------------------------------------------
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
 * Stats and stuff
 * -------------------------------------------------------------------------- */

/* -----------------------------------------------------------------------------
 * calcAllocated()
 *
 * Approximate how much we've allocated: number of blocks in the
 * nursery + blocks allocated via allocate() - unused nusery blocks.
 * This leaves a little slop at the end of each block, and doesn't
 * take into account large objects (ToDo).
 * -------------------------------------------------------------------------- */

lnat
calcAllocated( void )
{
  nat allocated;
  bdescr *bd;

#ifdef SMP
  Capability *cap;

627
628
629
630
  /* All tasks must be stopped.  Can't assert that all the
     capabilities are owned by the scheduler, though: one or more
     tasks might have been stopped while they were running (non-main)
     threads. */
631
  /*  ASSERT(n_free_capabilities == RtsFlags.ParFlags.nNodes); */
632
633
634
635
636
637

  allocated = 
    n_free_capabilities * RtsFlags.GcFlags.minAllocAreaSize * BLOCK_SIZE_W
    + allocated_bytes();

  for (cap = free_capabilities; cap != NULL; cap = cap->link) {
sof's avatar
sof committed
638
    for ( bd = cap->r.rCurrentNursery->link; bd != NULL; bd = bd->link ) {
639
640
      allocated -= BLOCK_SIZE_W;
    }
sof's avatar
sof committed
641
    if (cap->r.rCurrentNursery->free < cap->r.rCurrentNursery->start 
642
	+ BLOCK_SIZE_W) {
sof's avatar
sof committed
643
644
      allocated -= (cap->r.rCurrentNursery->start + BLOCK_SIZE_W)
	- cap->r.rCurrentNursery->free;
645
646
647
648
    }
  }

#else /* !SMP */
649
  bdescr *current_nursery = MainCapability.r.rCurrentNursery;
650

651
  allocated = (g0s0->n_blocks * BLOCK_SIZE_W) + allocated_bytes();
652
653
654
655
656
657
658
659
660
  for ( bd = current_nursery->link; bd != NULL; bd = bd->link ) {
    allocated -= BLOCK_SIZE_W;
  }
  if (current_nursery->free < current_nursery->start + BLOCK_SIZE_W) {
    allocated -= (current_nursery->start + BLOCK_SIZE_W)
      - current_nursery->free;
  }
#endif

661
  total_allocated += allocated;
662
663
  return allocated;
}  
664
665
666
667
668
669
670
671
672

/* Approximate the amount of live data in the heap.  To be called just
 * after garbage collection (see GarbageCollect()).
 */
extern lnat 
calcLive(void)
{
  nat g, s;
  lnat live = 0;
673
  step *stp;
674
675

  if (RtsFlags.GcFlags.generations == 1) {
676
    live = (g0s0->n_to_blocks - 1) * BLOCK_SIZE_W + 
677
      ((lnat)g0s0->hp_bd->free - (lnat)g0s0->hp_bd->start) / sizeof(W_);
678
    return live;
679
680
681
682
683
  }

  for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
    for (s = 0; s < generations[g].n_steps; s++) {
      /* approximate amount of live data (doesn't take into account slop
684
685
       * at end of each block).
       */
686
687
688
      if (g == 0 && s == 0) { 
	  continue; 
      }
689
      stp = &generations[g].steps[s];
690
      live += (stp->n_large_blocks + stp->n_blocks - 1) * BLOCK_SIZE_W;
691
692
693
694
      if (stp->hp_bd != NULL) {
	  live += ((lnat)stp->hp_bd->free - (lnat)stp->hp_bd->start) 
	      / sizeof(W_);
      }
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
    }
  }
  return live;
}

/* Approximate the number of blocks that will be needed at the next
 * garbage collection.
 *
 * Assume: all data currently live will remain live.  Steps that will
 * be collected next time will therefore need twice as many blocks
 * since all the data will be copied.
 */
extern lnat 
calcNeeded(void)
{
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
    lnat needed = 0;
    nat g, s;
    step *stp;
    
    for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
	for (s = 0; s < generations[g].n_steps; s++) {
	    if (g == 0 && s == 0) { continue; }
	    stp = &generations[g].steps[s];
	    if (generations[g].steps[0].n_blocks +
		generations[g].steps[0].n_large_blocks 
		> generations[g].max_blocks
		&& stp->is_compacted == 0) {
		needed += 2 * stp->n_blocks;
	    } else {
		needed += stp->n_blocks;
	    }
	}
727
    }
728
    return needed;
729
730
}

731
732
733
734
735
736
737
738
739
740
/* -----------------------------------------------------------------------------
   Debugging

   memInventory() checks for memory leaks by counting up all the
   blocks we know about and comparing that to the number of blocks
   allegedly floating around in the system.
   -------------------------------------------------------------------------- */

#ifdef DEBUG

741
void
742
743
744
memInventory(void)
{
  nat g, s;
745
  step *stp;
746
747
748
749
  bdescr *bd;
  lnat total_blocks = 0, free_blocks = 0;

  /* count the blocks we current have */
750

751
752
  for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
    for (s = 0; s < generations[g].n_steps; s++) {
753
754
      stp = &generations[g].steps[s];
      total_blocks += stp->n_blocks;
755
756
      if (RtsFlags.GcFlags.generations == 1) {
	/* two-space collector has a to-space too :-) */
757
	total_blocks += g0s0->n_to_blocks;
758
      }
759
      for (bd = stp->large_objects; bd; bd = bd->link) {
760
761
762
763
764
765
766
	total_blocks += bd->blocks;
	/* hack for megablock groups: they have an extra block or two in
	   the second and subsequent megablocks where the block
	   descriptors would normally go.
	*/
	if (bd->blocks > BLOCKS_PER_MBLOCK) {
	  total_blocks -= (MBLOCK_SIZE / BLOCK_SIZE - BLOCKS_PER_MBLOCK)
767
	                  * (bd->blocks/(MBLOCK_SIZE/BLOCK_SIZE));
768
769
770
771
772
773
774
775
776
777
778
779
	}
      }
    }
  }

  /* any blocks held by allocate() */
  for (bd = small_alloc_list; bd; bd = bd->link) {
    total_blocks += bd->blocks;
  }
  for (bd = large_alloc_list; bd; bd = bd->link) {
    total_blocks += bd->blocks;
  }
780
781
782
783
784
785
786
787

#ifdef PROFILING
  if (RtsFlags.ProfFlags.doHeapProfile == HEAP_BY_RETAINER) {
    for (bd = firstStack; bd != NULL; bd = bd->link) 
      total_blocks += bd->blocks;
  }
#endif

788
789
790
  // count the blocks allocated by the arena allocator
  total_blocks += arenaBlocks();

791
792
793
794
795
796
797
798
799
800
  /* count the blocks on the free list */
  free_blocks = countFreeList();

  if (total_blocks + free_blocks != mblocks_allocated *
      BLOCKS_PER_MBLOCK) {
    fprintf(stderr, "Blocks: %ld live + %ld free  = %ld total (%ld around)\n",
	    total_blocks, free_blocks, total_blocks + free_blocks,
	    mblocks_allocated * BLOCKS_PER_MBLOCK);
  }

801
802
  ASSERT(total_blocks + free_blocks == mblocks_allocated * BLOCKS_PER_MBLOCK);
}
803

804
805

nat
806
countBlocks(bdescr *bd)
807
{
808
809
    nat n;
    for (n=0; bd != NULL; bd=bd->link) {
810
	n += bd->blocks;
811
    }
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
    return n;
}

/* Full heap sanity check. */
void
checkSanity( void )
{
    nat g, s;

    if (RtsFlags.GcFlags.generations == 1) {
	checkHeap(g0s0->to_blocks);
	checkChain(g0s0->large_objects);
    } else {
	
	for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
	    for (s = 0; s < generations[g].n_steps; s++) {
		ASSERT(countBlocks(generations[g].steps[s].blocks)
		       == generations[g].steps[s].n_blocks);
830
831
		ASSERT(countBlocks(generations[g].steps[s].large_objects)
		       == generations[g].steps[s].n_large_blocks);
832
833
834
		if (g == 0 && s == 0) { continue; }
		checkHeap(generations[g].steps[s].blocks);
		checkChain(generations[g].steps[s].large_objects);
835
836
837
838
839
840
841
		if (g > 0) {
		    checkMutableList(generations[g].mut_list, g);
		    checkMutOnceList(generations[g].mut_once_list, g);
		}
	    }
	}
	checkFreeListSanity();
842
843
844
    }
}

845
846
847
848
849
850
851
852
853
// handy function for use in gdb, because Bdescr() is inlined.
extern bdescr *_bdescr( StgPtr p );

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

854
#endif