f0506cd77cd8037a8774b8c723e2153f770f98f6
[ghc.git] / rts / sm / Storage.c
1 /* -----------------------------------------------------------------------------
2 *
3 * (c) The GHC Team, 1998-2008
4 *
5 * Storage manager front end
6 *
7 * Documentation on the architecture of the Storage Manager can be
8 * found in the online commentary:
9 *
10 * http://hackage.haskell.org/trac/ghc/wiki/Commentary/Rts/Storage
11 *
12 * ---------------------------------------------------------------------------*/
13
14 #include "PosixSource.h"
15 #include "Rts.h"
16
17 #include "Storage.h"
18 #include "RtsUtils.h"
19 #include "Stats.h"
20 #include "BlockAlloc.h"
21 #include "Weak.h"
22 #include "Sanity.h"
23 #include "Arena.h"
24 #include "Capability.h"
25 #include "Schedule.h"
26 #include "RetainerProfile.h" // for counting memory blocks (memInventory)
27 #include "OSMem.h"
28 #include "Trace.h"
29 #include "GC.h"
30 #include "Evac.h"
31
32 #include <string.h>
33
34 #include "ffi.h"
35
36 /*
37 * All these globals require sm_mutex to access in THREADED_RTS mode.
38 */
39 StgClosure *caf_list = NULL;
40 StgClosure *revertible_caf_list = NULL;
41 rtsBool keepCAFs;
42
43 bdescr *pinned_object_block; /* allocate pinned objects into this block */
44 nat alloc_blocks; /* number of allocate()d blocks since GC */
45 nat alloc_blocks_lim; /* approximate limit on alloc_blocks */
46
47 static bdescr *exec_block;
48
49 generation *generations = NULL; /* all the generations */
50 generation *g0 = NULL; /* generation 0, for convenience */
51 generation *oldest_gen = NULL; /* oldest generation, for convenience */
52 step *g0s0 = NULL; /* generation 0, step 0, for convenience */
53
54 nat total_steps = 0;
55 step *all_steps = NULL; /* single array of steps */
56
57 ullong total_allocated = 0; /* total memory allocated during run */
58
59 nat n_nurseries = 0; /* == RtsFlags.ParFlags.nNodes, convenience */
60 step *nurseries = NULL; /* array of nurseries, >1 only if THREADED_RTS */
61
62 #ifdef THREADED_RTS
63 /*
64 * Storage manager mutex: protects all the above state from
65 * simultaneous access by two STG threads.
66 */
67 Mutex sm_mutex;
68 #endif
69
70 static void allocNurseries ( void );
71
72 static void
73 initStep (step *stp, int g, int s)
74 {
75 stp->no = s;
76 stp->abs_no = RtsFlags.GcFlags.steps * g + s;
77 stp->blocks = NULL;
78 stp->n_blocks = 0;
79 stp->n_words = 0;
80 stp->live_estimate = 0;
81 stp->old_blocks = NULL;
82 stp->n_old_blocks = 0;
83 stp->gen = &generations[g];
84 stp->gen_no = g;
85 stp->large_objects = NULL;
86 stp->n_large_blocks = 0;
87 stp->scavenged_large_objects = NULL;
88 stp->n_scavenged_large_blocks = 0;
89 stp->mark = 0;
90 stp->compact = 0;
91 stp->bitmap = NULL;
92 #ifdef THREADED_RTS
93 initSpinLock(&stp->sync_large_objects);
94 #endif
95 stp->threads = END_TSO_QUEUE;
96 stp->old_threads = END_TSO_QUEUE;
97 }
98
99 void
100 initStorage( void )
101 {
102 nat g, s;
103 generation *gen;
104
105 if (generations != NULL) {
106 // multi-init protection
107 return;
108 }
109
110 initMBlocks();
111
112 /* Sanity check to make sure the LOOKS_LIKE_ macros appear to be
113 * doing something reasonable.
114 */
115 /* We use the NOT_NULL variant or gcc warns that the test is always true */
116 ASSERT(LOOKS_LIKE_INFO_PTR_NOT_NULL((StgWord)&stg_BLACKHOLE_info));
117 ASSERT(LOOKS_LIKE_CLOSURE_PTR(&stg_dummy_ret_closure));
118 ASSERT(!HEAP_ALLOCED(&stg_dummy_ret_closure));
119
120 if (RtsFlags.GcFlags.maxHeapSize != 0 &&
121 RtsFlags.GcFlags.heapSizeSuggestion >
122 RtsFlags.GcFlags.maxHeapSize) {
123 RtsFlags.GcFlags.maxHeapSize = RtsFlags.GcFlags.heapSizeSuggestion;
124 }
125
126 if (RtsFlags.GcFlags.maxHeapSize != 0 &&
127 RtsFlags.GcFlags.minAllocAreaSize >
128 RtsFlags.GcFlags.maxHeapSize) {
129 errorBelch("maximum heap size (-M) is smaller than minimum alloc area size (-A)");
130 RtsFlags.GcFlags.minAllocAreaSize = RtsFlags.GcFlags.maxHeapSize;
131 }
132
133 initBlockAllocator();
134
135 #if defined(THREADED_RTS)
136 initMutex(&sm_mutex);
137 #endif
138
139 ACQUIRE_SM_LOCK;
140
141 /* allocate generation info array */
142 generations = (generation *)stgMallocBytes(RtsFlags.GcFlags.generations
143 * sizeof(struct generation_),
144 "initStorage: gens");
145
146 /* allocate all the steps into an array. It is important that we do
147 it this way, because we need the invariant that two step pointers
148 can be directly compared to see which is the oldest.
149 Remember that the last generation has only one step. */
150 total_steps = 1 + (RtsFlags.GcFlags.generations - 1) * RtsFlags.GcFlags.steps;
151 all_steps = stgMallocBytes(total_steps * sizeof(struct step_),
152 "initStorage: steps");
153
154 /* Initialise all generations */
155 for(g = 0; g < RtsFlags.GcFlags.generations; g++) {
156 gen = &generations[g];
157 gen->no = g;
158 gen->mut_list = allocBlock();
159 gen->collections = 0;
160 gen->par_collections = 0;
161 gen->failed_promotions = 0;
162 gen->max_blocks = 0;
163 }
164
165 /* A couple of convenience pointers */
166 g0 = &generations[0];
167 oldest_gen = &generations[RtsFlags.GcFlags.generations-1];
168
169 /* Allocate step structures in each generation */
170 if (RtsFlags.GcFlags.generations > 1) {
171 /* Only for multiple-generations */
172
173 /* Oldest generation: one step */
174 oldest_gen->n_steps = 1;
175 oldest_gen->steps = all_steps + (RtsFlags.GcFlags.generations - 1)
176 * RtsFlags.GcFlags.steps;
177
178 /* set up all except the oldest generation with 2 steps */
179 for(g = 0; g < RtsFlags.GcFlags.generations-1; g++) {
180 generations[g].n_steps = RtsFlags.GcFlags.steps;
181 generations[g].steps = all_steps + g * RtsFlags.GcFlags.steps;
182 }
183
184 } else {
185 /* single generation, i.e. a two-space collector */
186 g0->n_steps = 1;
187 g0->steps = all_steps;
188 }
189
190 #ifdef THREADED_RTS
191 n_nurseries = n_capabilities;
192 #else
193 n_nurseries = 1;
194 #endif
195 nurseries = stgMallocBytes (n_nurseries * sizeof(struct step_),
196 "initStorage: nurseries");
197
198 /* Initialise all steps */
199 for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
200 for (s = 0; s < generations[g].n_steps; s++) {
201 initStep(&generations[g].steps[s], g, s);
202 }
203 }
204
205 for (s = 0; s < n_nurseries; s++) {
206 initStep(&nurseries[s], 0, s);
207 }
208
209 /* Set up the destination pointers in each younger gen. step */
210 for (g = 0; g < RtsFlags.GcFlags.generations-1; g++) {
211 for (s = 0; s < generations[g].n_steps-1; s++) {
212 generations[g].steps[s].to = &generations[g].steps[s+1];
213 }
214 generations[g].steps[s].to = &generations[g+1].steps[0];
215 }
216 oldest_gen->steps[0].to = &oldest_gen->steps[0];
217
218 for (s = 0; s < n_nurseries; s++) {
219 nurseries[s].to = generations[0].steps[0].to;
220 }
221
222 /* The oldest generation has one step. */
223 if (RtsFlags.GcFlags.compact || RtsFlags.GcFlags.sweep) {
224 if (RtsFlags.GcFlags.generations == 1) {
225 errorBelch("WARNING: compact/sweep is incompatible with -G1; disabled");
226 } else {
227 oldest_gen->steps[0].mark = 1;
228 if (RtsFlags.GcFlags.compact)
229 oldest_gen->steps[0].compact = 1;
230 }
231 }
232
233 generations[0].max_blocks = 0;
234 g0s0 = &generations[0].steps[0];
235
236 /* The allocation area. Policy: keep the allocation area
237 * small to begin with, even if we have a large suggested heap
238 * size. Reason: we're going to do a major collection first, and we
239 * don't want it to be a big one. This vague idea is borne out by
240 * rigorous experimental evidence.
241 */
242 allocNurseries();
243
244 weak_ptr_list = NULL;
245 caf_list = NULL;
246 revertible_caf_list = NULL;
247
248 /* initialise the allocate() interface */
249 alloc_blocks = 0;
250 alloc_blocks_lim = RtsFlags.GcFlags.minAllocAreaSize;
251
252 exec_block = NULL;
253
254 #ifdef THREADED_RTS
255 initSpinLock(&gc_alloc_block_sync);
256 whitehole_spin = 0;
257 #endif
258
259 N = 0;
260
261 initGcThreads();
262
263 IF_DEBUG(gc, statDescribeGens());
264
265 RELEASE_SM_LOCK;
266 }
267
268 void
269 exitStorage (void)
270 {
271 stat_exit(calcAllocated());
272 }
273
274 void
275 freeStorage (void)
276 {
277 stgFree(g0s0); // frees all the steps
278 stgFree(generations);
279 freeAllMBlocks();
280 #if defined(THREADED_RTS)
281 closeMutex(&sm_mutex);
282 #endif
283 stgFree(nurseries);
284 freeGcThreads();
285 }
286
287 /* -----------------------------------------------------------------------------
288 CAF management.
289
290 The entry code for every CAF does the following:
291
292 - builds a CAF_BLACKHOLE in the heap
293 - pushes an update frame pointing to the CAF_BLACKHOLE
294 - invokes UPD_CAF(), which:
295 - calls newCaf, below
296 - updates the CAF with a static indirection to the CAF_BLACKHOLE
297
298 Why do we build a BLACKHOLE in the heap rather than just updating
299 the thunk directly? It's so that we only need one kind of update
300 frame - otherwise we'd need a static version of the update frame too.
301
302 newCaf() does the following:
303
304 - it puts the CAF on the oldest generation's mut-once list.
305 This is so that we can treat the CAF as a root when collecting
306 younger generations.
307
308 For GHCI, we have additional requirements when dealing with CAFs:
309
310 - we must *retain* all dynamically-loaded CAFs ever entered,
311 just in case we need them again.
312 - we must be able to *revert* CAFs that have been evaluated, to
313 their pre-evaluated form.
314
315 To do this, we use an additional CAF list. When newCaf() is
316 called on a dynamically-loaded CAF, we add it to the CAF list
317 instead of the old-generation mutable list, and save away its
318 old info pointer (in caf->saved_info) for later reversion.
319
320 To revert all the CAFs, we traverse the CAF list and reset the
321 info pointer to caf->saved_info, then throw away the CAF list.
322 (see GC.c:revertCAFs()).
323
324 -- SDM 29/1/01
325
326 -------------------------------------------------------------------------- */
327
328 void
329 newCAF(StgClosure* caf)
330 {
331 ACQUIRE_SM_LOCK;
332
333 #ifdef DYNAMIC
334 if(keepCAFs)
335 {
336 // HACK:
337 // If we are in GHCi _and_ we are using dynamic libraries,
338 // then we can't redirect newCAF calls to newDynCAF (see below),
339 // so we make newCAF behave almost like newDynCAF.
340 // The dynamic libraries might be used by both the interpreted
341 // program and GHCi itself, so they must not be reverted.
342 // This also means that in GHCi with dynamic libraries, CAFs are not
343 // garbage collected. If this turns out to be a problem, we could
344 // do another hack here and do an address range test on caf to figure
345 // out whether it is from a dynamic library.
346 ((StgIndStatic *)caf)->saved_info = (StgInfoTable *)caf->header.info;
347 ((StgIndStatic *)caf)->static_link = caf_list;
348 caf_list = caf;
349 }
350 else
351 #endif
352 {
353 /* Put this CAF on the mutable list for the old generation.
354 * This is a HACK - the IND_STATIC closure doesn't really have
355 * a mut_link field, but we pretend it has - in fact we re-use
356 * the STATIC_LINK field for the time being, because when we
357 * come to do a major GC we won't need the mut_link field
358 * any more and can use it as a STATIC_LINK.
359 */
360 ((StgIndStatic *)caf)->saved_info = NULL;
361 recordMutableGen(caf, oldest_gen->no);
362 }
363
364 RELEASE_SM_LOCK;
365 }
366
367 // An alternate version of newCaf which is used for dynamically loaded
368 // object code in GHCi. In this case we want to retain *all* CAFs in
369 // the object code, because they might be demanded at any time from an
370 // expression evaluated on the command line.
371 // Also, GHCi might want to revert CAFs, so we add these to the
372 // revertible_caf_list.
373 //
374 // The linker hackily arranges that references to newCaf from dynamic
375 // code end up pointing to newDynCAF.
376 void
377 newDynCAF(StgClosure *caf)
378 {
379 ACQUIRE_SM_LOCK;
380
381 ((StgIndStatic *)caf)->saved_info = (StgInfoTable *)caf->header.info;
382 ((StgIndStatic *)caf)->static_link = revertible_caf_list;
383 revertible_caf_list = caf;
384
385 RELEASE_SM_LOCK;
386 }
387
388 /* -----------------------------------------------------------------------------
389 Nursery management.
390 -------------------------------------------------------------------------- */
391
392 static bdescr *
393 allocNursery (step *stp, bdescr *tail, nat blocks)
394 {
395 bdescr *bd;
396 nat i;
397
398 // Allocate a nursery: we allocate fresh blocks one at a time and
399 // cons them on to the front of the list, not forgetting to update
400 // the back pointer on the tail of the list to point to the new block.
401 for (i=0; i < blocks; i++) {
402 // @LDV profiling
403 /*
404 processNursery() in LdvProfile.c assumes that every block group in
405 the nursery contains only a single block. So, if a block group is
406 given multiple blocks, change processNursery() accordingly.
407 */
408 bd = allocBlock();
409 bd->link = tail;
410 // double-link the nursery: we might need to insert blocks
411 if (tail != NULL) {
412 tail->u.back = bd;
413 }
414 bd->step = stp;
415 bd->gen_no = 0;
416 bd->flags = 0;
417 bd->free = bd->start;
418 tail = bd;
419 }
420 tail->u.back = NULL;
421 return tail;
422 }
423
424 static void
425 assignNurseriesToCapabilities (void)
426 {
427 #ifdef THREADED_RTS
428 nat i;
429
430 for (i = 0; i < n_nurseries; i++) {
431 capabilities[i].r.rNursery = &nurseries[i];
432 capabilities[i].r.rCurrentNursery = nurseries[i].blocks;
433 capabilities[i].r.rCurrentAlloc = NULL;
434 }
435 #else /* THREADED_RTS */
436 MainCapability.r.rNursery = &nurseries[0];
437 MainCapability.r.rCurrentNursery = nurseries[0].blocks;
438 MainCapability.r.rCurrentAlloc = NULL;
439 #endif
440 }
441
442 static void
443 allocNurseries( void )
444 {
445 nat i;
446
447 for (i = 0; i < n_nurseries; i++) {
448 nurseries[i].blocks =
449 allocNursery(&nurseries[i], NULL,
450 RtsFlags.GcFlags.minAllocAreaSize);
451 nurseries[i].n_blocks = RtsFlags.GcFlags.minAllocAreaSize;
452 nurseries[i].old_blocks = NULL;
453 nurseries[i].n_old_blocks = 0;
454 }
455 assignNurseriesToCapabilities();
456 }
457
458 void
459 resetNurseries( void )
460 {
461 nat i;
462 bdescr *bd;
463 step *stp;
464
465 for (i = 0; i < n_nurseries; i++) {
466 stp = &nurseries[i];
467 for (bd = stp->blocks; bd; bd = bd->link) {
468 bd->free = bd->start;
469 ASSERT(bd->gen_no == 0);
470 ASSERT(bd->step == stp);
471 IF_DEBUG(sanity,memset(bd->start, 0xaa, BLOCK_SIZE));
472 }
473 }
474 assignNurseriesToCapabilities();
475 }
476
477 lnat
478 countNurseryBlocks (void)
479 {
480 nat i;
481 lnat blocks = 0;
482
483 for (i = 0; i < n_nurseries; i++) {
484 blocks += nurseries[i].n_blocks;
485 }
486 return blocks;
487 }
488
489 static void
490 resizeNursery ( step *stp, nat blocks )
491 {
492 bdescr *bd;
493 nat nursery_blocks;
494
495 nursery_blocks = stp->n_blocks;
496 if (nursery_blocks == blocks) return;
497
498 if (nursery_blocks < blocks) {
499 debugTrace(DEBUG_gc, "increasing size of nursery to %d blocks",
500 blocks);
501 stp->blocks = allocNursery(stp, stp->blocks, blocks-nursery_blocks);
502 }
503 else {
504 bdescr *next_bd;
505
506 debugTrace(DEBUG_gc, "decreasing size of nursery to %d blocks",
507 blocks);
508
509 bd = stp->blocks;
510 while (nursery_blocks > blocks) {
511 next_bd = bd->link;
512 next_bd->u.back = NULL;
513 nursery_blocks -= bd->blocks; // might be a large block
514 freeGroup(bd);
515 bd = next_bd;
516 }
517 stp->blocks = bd;
518 // might have gone just under, by freeing a large block, so make
519 // up the difference.
520 if (nursery_blocks < blocks) {
521 stp->blocks = allocNursery(stp, stp->blocks, blocks-nursery_blocks);
522 }
523 }
524
525 stp->n_blocks = blocks;
526 ASSERT(countBlocks(stp->blocks) == stp->n_blocks);
527 }
528
529 //
530 // Resize each of the nurseries to the specified size.
531 //
532 void
533 resizeNurseriesFixed (nat blocks)
534 {
535 nat i;
536 for (i = 0; i < n_nurseries; i++) {
537 resizeNursery(&nurseries[i], blocks);
538 }
539 }
540
541 //
542 // Resize the nurseries to the total specified size.
543 //
544 void
545 resizeNurseries (nat blocks)
546 {
547 // If there are multiple nurseries, then we just divide the number
548 // of available blocks between them.
549 resizeNurseriesFixed(blocks / n_nurseries);
550 }
551
552
553 /* -----------------------------------------------------------------------------
554 move_TSO is called to update the TSO structure after it has been
555 moved from one place to another.
556 -------------------------------------------------------------------------- */
557
558 void
559 move_TSO (StgTSO *src, StgTSO *dest)
560 {
561 ptrdiff_t diff;
562
563 // relocate the stack pointer...
564 diff = (StgPtr)dest - (StgPtr)src; // In *words*
565 dest->sp = (StgPtr)dest->sp + diff;
566 }
567
568 /* -----------------------------------------------------------------------------
569 The allocate() interface
570
571 allocateInGen() function allocates memory directly into a specific
572 generation. It always succeeds, and returns a chunk of memory n
573 words long. n can be larger than the size of a block if necessary,
574 in which case a contiguous block group will be allocated.
575
576 allocate(n) is equivalent to allocateInGen(g0).
577 -------------------------------------------------------------------------- */
578
579 StgPtr
580 allocateInGen (generation *g, lnat n)
581 {
582 step *stp;
583 bdescr *bd;
584 StgPtr ret;
585
586 ACQUIRE_SM_LOCK;
587
588 TICK_ALLOC_HEAP_NOCTR(n);
589 CCS_ALLOC(CCCS,n);
590
591 stp = &g->steps[0];
592
593 if (n >= LARGE_OBJECT_THRESHOLD/sizeof(W_))
594 {
595 lnat req_blocks = (lnat)BLOCK_ROUND_UP(n*sizeof(W_)) / BLOCK_SIZE;
596
597 // Attempting to allocate an object larger than maxHeapSize
598 // should definitely be disallowed. (bug #1791)
599 if (RtsFlags.GcFlags.maxHeapSize > 0 &&
600 req_blocks >= RtsFlags.GcFlags.maxHeapSize) {
601 heapOverflow();
602 // heapOverflow() doesn't exit (see #2592), but we aren't
603 // in a position to do a clean shutdown here: we
604 // either have to allocate the memory or exit now.
605 // Allocating the memory would be bad, because the user
606 // has requested that we not exceed maxHeapSize, so we
607 // just exit.
608 stg_exit(EXIT_HEAPOVERFLOW);
609 }
610
611 bd = allocGroup(req_blocks);
612 dbl_link_onto(bd, &stp->large_objects);
613 stp->n_large_blocks += bd->blocks; // might be larger than req_blocks
614 alloc_blocks += bd->blocks;
615 bd->gen_no = g->no;
616 bd->step = stp;
617 bd->flags = BF_LARGE;
618 bd->free = bd->start + n;
619 ret = bd->start;
620 }
621 else
622 {
623 // small allocation (<LARGE_OBJECT_THRESHOLD) */
624 bd = stp->blocks;
625 if (bd == NULL || bd->free + n > bd->start + BLOCK_SIZE_W) {
626 bd = allocBlock();
627 bd->gen_no = g->no;
628 bd->step = stp;
629 bd->flags = 0;
630 bd->link = stp->blocks;
631 stp->blocks = bd;
632 stp->n_blocks++;
633 alloc_blocks++;
634 }
635 ret = bd->free;
636 bd->free += n;
637 }
638
639 RELEASE_SM_LOCK;
640
641 return ret;
642 }
643
644 StgPtr
645 allocate (lnat n)
646 {
647 return allocateInGen(g0,n);
648 }
649
650 lnat
651 allocatedBytes( void )
652 {
653 lnat allocated;
654
655 allocated = alloc_blocks * BLOCK_SIZE_W;
656 if (pinned_object_block != NULL) {
657 allocated -= (pinned_object_block->start + BLOCK_SIZE_W) -
658 pinned_object_block->free;
659 }
660
661 return allocated;
662 }
663
664 // split N blocks off the front of the given bdescr, returning the
665 // new block group. We treat the remainder as if it
666 // had been freshly allocated in generation 0.
667 bdescr *
668 splitLargeBlock (bdescr *bd, nat blocks)
669 {
670 bdescr *new_bd;
671
672 // subtract the original number of blocks from the counter first
673 bd->step->n_large_blocks -= bd->blocks;
674
675 new_bd = splitBlockGroup (bd, blocks);
676
677 dbl_link_onto(new_bd, &g0s0->large_objects);
678 g0s0->n_large_blocks += new_bd->blocks;
679 new_bd->gen_no = g0s0->no;
680 new_bd->step = g0s0;
681 new_bd->flags = BF_LARGE;
682 new_bd->free = bd->free;
683 ASSERT(new_bd->free <= new_bd->start + new_bd->blocks * BLOCK_SIZE_W);
684
685 // add the new number of blocks to the counter. Due to the gaps
686 // for block descriptor, new_bd->blocks + bd->blocks might not be
687 // equal to the original bd->blocks, which is why we do it this way.
688 bd->step->n_large_blocks += bd->blocks;
689
690 return new_bd;
691 }
692
693 /* -----------------------------------------------------------------------------
694 allocateLocal()
695
696 This allocates memory in the current thread - it is intended for
697 use primarily from STG-land where we have a Capability. It is
698 better than allocate() because it doesn't require taking the
699 sm_mutex lock in the common case.
700
701 Memory is allocated directly from the nursery if possible (but not
702 from the current nursery block, so as not to interfere with
703 Hp/HpLim).
704 -------------------------------------------------------------------------- */
705
706 StgPtr
707 allocateLocal (Capability *cap, lnat n)
708 {
709 bdescr *bd;
710 StgPtr p;
711
712 if (n >= LARGE_OBJECT_THRESHOLD/sizeof(W_)) {
713 return allocateInGen(g0,n);
714 }
715
716 /* small allocation (<LARGE_OBJECT_THRESHOLD) */
717
718 TICK_ALLOC_HEAP_NOCTR(n);
719 CCS_ALLOC(CCCS,n);
720
721 bd = cap->r.rCurrentAlloc;
722 if (bd == NULL || bd->free + n > bd->start + BLOCK_SIZE_W) {
723
724 // The CurrentAlloc block is full, we need to find another
725 // one. First, we try taking the next block from the
726 // nursery:
727 bd = cap->r.rCurrentNursery->link;
728
729 if (bd == NULL || bd->free + n > bd->start + BLOCK_SIZE_W) {
730 // The nursery is empty, or the next block is already
731 // full: allocate a fresh block (we can't fail here).
732 ACQUIRE_SM_LOCK;
733 bd = allocBlock();
734 cap->r.rNursery->n_blocks++;
735 RELEASE_SM_LOCK;
736 bd->gen_no = 0;
737 bd->step = cap->r.rNursery;
738 bd->flags = 0;
739 // NO: alloc_blocks++;
740 // calcAllocated() uses the size of the nursery, and we've
741 // already bumpted nursery->n_blocks above. We'll GC
742 // pretty quickly now anyway, because MAYBE_GC() will
743 // notice that CurrentNursery->link is NULL.
744 } else {
745 // we have a block in the nursery: take it and put
746 // it at the *front* of the nursery list, and use it
747 // to allocate() from.
748 cap->r.rCurrentNursery->link = bd->link;
749 if (bd->link != NULL) {
750 bd->link->u.back = cap->r.rCurrentNursery;
751 }
752 }
753 dbl_link_onto(bd, &cap->r.rNursery->blocks);
754 cap->r.rCurrentAlloc = bd;
755 IF_DEBUG(sanity, checkNurserySanity(cap->r.rNursery));
756 }
757 p = bd->free;
758 bd->free += n;
759 return p;
760 }
761
762 /* ---------------------------------------------------------------------------
763 Allocate a fixed/pinned object.
764
765 We allocate small pinned objects into a single block, allocating a
766 new block when the current one overflows. The block is chained
767 onto the large_object_list of generation 0 step 0.
768
769 NOTE: The GC can't in general handle pinned objects. This
770 interface is only safe to use for ByteArrays, which have no
771 pointers and don't require scavenging. It works because the
772 block's descriptor has the BF_LARGE flag set, so the block is
773 treated as a large object and chained onto various lists, rather
774 than the individual objects being copied. However, when it comes
775 to scavenge the block, the GC will only scavenge the first object.
776 The reason is that the GC can't linearly scan a block of pinned
777 objects at the moment (doing so would require using the
778 mostly-copying techniques). But since we're restricting ourselves
779 to pinned ByteArrays, not scavenging is ok.
780
781 This function is called by newPinnedByteArray# which immediately
782 fills the allocated memory with a MutableByteArray#.
783 ------------------------------------------------------------------------- */
784
785 StgPtr
786 allocatePinned( lnat n )
787 {
788 StgPtr p;
789 bdescr *bd = pinned_object_block;
790
791 // If the request is for a large object, then allocate()
792 // will give us a pinned object anyway.
793 if (n >= LARGE_OBJECT_THRESHOLD/sizeof(W_)) {
794 p = allocate(n);
795 Bdescr(p)->flags |= BF_PINNED;
796 return p;
797 }
798
799 ACQUIRE_SM_LOCK;
800
801 TICK_ALLOC_HEAP_NOCTR(n);
802 CCS_ALLOC(CCCS,n);
803
804 // If we don't have a block of pinned objects yet, or the current
805 // one isn't large enough to hold the new object, allocate a new one.
806 if (bd == NULL || (bd->free + n) > (bd->start + BLOCK_SIZE_W)) {
807 pinned_object_block = bd = allocBlock();
808 dbl_link_onto(bd, &g0s0->large_objects);
809 g0s0->n_large_blocks++;
810 bd->gen_no = 0;
811 bd->step = g0s0;
812 bd->flags = BF_PINNED | BF_LARGE;
813 bd->free = bd->start;
814 alloc_blocks++;
815 }
816
817 p = bd->free;
818 bd->free += n;
819 RELEASE_SM_LOCK;
820 return p;
821 }
822
823 /* -----------------------------------------------------------------------------
824 Write Barriers
825 -------------------------------------------------------------------------- */
826
827 /*
828 This is the write barrier for MUT_VARs, a.k.a. IORefs. A
829 MUT_VAR_CLEAN object is not on the mutable list; a MUT_VAR_DIRTY
830 is. When written to, a MUT_VAR_CLEAN turns into a MUT_VAR_DIRTY
831 and is put on the mutable list.
832 */
833 void
834 dirty_MUT_VAR(StgRegTable *reg, StgClosure *p)
835 {
836 Capability *cap = regTableToCapability(reg);
837 bdescr *bd;
838 if (p->header.info == &stg_MUT_VAR_CLEAN_info) {
839 p->header.info = &stg_MUT_VAR_DIRTY_info;
840 bd = Bdescr((StgPtr)p);
841 if (bd->gen_no > 0) recordMutableCap(p,cap,bd->gen_no);
842 }
843 }
844
845 // Setting a TSO's link field with a write barrier.
846 // It is *not* necessary to call this function when
847 // * setting the link field to END_TSO_QUEUE
848 // * putting a TSO on the blackhole_queue
849 // * setting the link field of the currently running TSO, as it
850 // will already be dirty.
851 void
852 setTSOLink (Capability *cap, StgTSO *tso, StgTSO *target)
853 {
854 bdescr *bd;
855 if (tso->dirty == 0 && (tso->flags & TSO_LINK_DIRTY) == 0) {
856 tso->flags |= TSO_LINK_DIRTY;
857 bd = Bdescr((StgPtr)tso);
858 if (bd->gen_no > 0) recordMutableCap((StgClosure*)tso,cap,bd->gen_no);
859 }
860 tso->_link = target;
861 }
862
863 void
864 dirty_TSO (Capability *cap, StgTSO *tso)
865 {
866 bdescr *bd;
867 if (tso->dirty == 0 && (tso->flags & TSO_LINK_DIRTY) == 0) {
868 bd = Bdescr((StgPtr)tso);
869 if (bd->gen_no > 0) recordMutableCap((StgClosure*)tso,cap,bd->gen_no);
870 }
871 tso->dirty = 1;
872 }
873
874 /*
875 This is the write barrier for MVARs. An MVAR_CLEAN objects is not
876 on the mutable list; a MVAR_DIRTY is. When written to, a
877 MVAR_CLEAN turns into a MVAR_DIRTY and is put on the mutable list.
878 The check for MVAR_CLEAN is inlined at the call site for speed,
879 this really does make a difference on concurrency-heavy benchmarks
880 such as Chaneneos and cheap-concurrency.
881 */
882 void
883 dirty_MVAR(StgRegTable *reg, StgClosure *p)
884 {
885 Capability *cap = regTableToCapability(reg);
886 bdescr *bd;
887 bd = Bdescr((StgPtr)p);
888 if (bd->gen_no > 0) recordMutableCap(p,cap,bd->gen_no);
889 }
890
891 /* -----------------------------------------------------------------------------
892 * Stats and stuff
893 * -------------------------------------------------------------------------- */
894
895 /* -----------------------------------------------------------------------------
896 * calcAllocated()
897 *
898 * Approximate how much we've allocated: number of blocks in the
899 * nursery + blocks allocated via allocate() - unused nusery blocks.
900 * This leaves a little slop at the end of each block, and doesn't
901 * take into account large objects (ToDo).
902 * -------------------------------------------------------------------------- */
903
904 lnat
905 calcAllocated( void )
906 {
907 nat allocated;
908 bdescr *bd;
909
910 allocated = allocatedBytes();
911 allocated += countNurseryBlocks() * BLOCK_SIZE_W;
912
913 {
914 #ifdef THREADED_RTS
915 nat i;
916 for (i = 0; i < n_nurseries; i++) {
917 Capability *cap;
918 for ( bd = capabilities[i].r.rCurrentNursery->link;
919 bd != NULL; bd = bd->link ) {
920 allocated -= BLOCK_SIZE_W;
921 }
922 cap = &capabilities[i];
923 if (cap->r.rCurrentNursery->free <
924 cap->r.rCurrentNursery->start + BLOCK_SIZE_W) {
925 allocated -= (cap->r.rCurrentNursery->start + BLOCK_SIZE_W)
926 - cap->r.rCurrentNursery->free;
927 }
928 }
929 #else
930 bdescr *current_nursery = MainCapability.r.rCurrentNursery;
931
932 for ( bd = current_nursery->link; bd != NULL; bd = bd->link ) {
933 allocated -= BLOCK_SIZE_W;
934 }
935 if (current_nursery->free < current_nursery->start + BLOCK_SIZE_W) {
936 allocated -= (current_nursery->start + BLOCK_SIZE_W)
937 - current_nursery->free;
938 }
939 #endif
940 }
941
942 total_allocated += allocated;
943 return allocated;
944 }
945
946 /* Approximate the amount of live data in the heap. To be called just
947 * after garbage collection (see GarbageCollect()).
948 */
949 lnat
950 calcLiveBlocks(void)
951 {
952 nat g, s;
953 lnat live = 0;
954 step *stp;
955
956 if (RtsFlags.GcFlags.generations == 1) {
957 return g0s0->n_large_blocks + g0s0->n_blocks;
958 }
959
960 for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
961 for (s = 0; s < generations[g].n_steps; s++) {
962 /* approximate amount of live data (doesn't take into account slop
963 * at end of each block).
964 */
965 if (g == 0 && s == 0) {
966 continue;
967 }
968 stp = &generations[g].steps[s];
969 live += stp->n_large_blocks + stp->n_blocks;
970 }
971 }
972 return live;
973 }
974
975 lnat
976 countOccupied(bdescr *bd)
977 {
978 lnat words;
979
980 words = 0;
981 for (; bd != NULL; bd = bd->link) {
982 ASSERT(bd->free <= bd->start + bd->blocks * BLOCK_SIZE_W);
983 words += bd->free - bd->start;
984 }
985 return words;
986 }
987
988 // Return an accurate count of the live data in the heap, excluding
989 // generation 0.
990 lnat
991 calcLiveWords(void)
992 {
993 nat g, s;
994 lnat live;
995 step *stp;
996
997 if (RtsFlags.GcFlags.generations == 1) {
998 return g0s0->n_words + countOccupied(g0s0->large_objects);
999 }
1000
1001 live = 0;
1002 for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
1003 for (s = 0; s < generations[g].n_steps; s++) {
1004 if (g == 0 && s == 0) continue;
1005 stp = &generations[g].steps[s];
1006 live += stp->n_words + countOccupied(stp->large_objects);
1007 }
1008 }
1009 return live;
1010 }
1011
1012 /* Approximate the number of blocks that will be needed at the next
1013 * garbage collection.
1014 *
1015 * Assume: all data currently live will remain live. Steps that will
1016 * be collected next time will therefore need twice as many blocks
1017 * since all the data will be copied.
1018 */
1019 extern lnat
1020 calcNeeded(void)
1021 {
1022 lnat needed = 0;
1023 nat g, s;
1024 step *stp;
1025
1026 for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
1027 for (s = 0; s < generations[g].n_steps; s++) {
1028 if (g == 0 && s == 0) { continue; }
1029 stp = &generations[g].steps[s];
1030
1031 // we need at least this much space
1032 needed += stp->n_blocks + stp->n_large_blocks;
1033
1034 // any additional space needed to collect this gen next time?
1035 if (g == 0 || // always collect gen 0
1036 (generations[g].steps[0].n_blocks +
1037 generations[g].steps[0].n_large_blocks
1038 > generations[g].max_blocks)) {
1039 // we will collect this gen next time
1040 if (stp->mark) {
1041 // bitmap:
1042 needed += stp->n_blocks / BITS_IN(W_);
1043 // mark stack:
1044 needed += stp->n_blocks / 100;
1045 }
1046 if (stp->compact) {
1047 continue; // no additional space needed for compaction
1048 } else {
1049 needed += stp->n_blocks;
1050 }
1051 }
1052 }
1053 }
1054 return needed;
1055 }
1056
1057 /* ----------------------------------------------------------------------------
1058 Executable memory
1059
1060 Executable memory must be managed separately from non-executable
1061 memory. Most OSs these days require you to jump through hoops to
1062 dynamically allocate executable memory, due to various security
1063 measures.
1064
1065 Here we provide a small memory allocator for executable memory.
1066 Memory is managed with a page granularity; we allocate linearly
1067 in the page, and when the page is emptied (all objects on the page
1068 are free) we free the page again, not forgetting to make it
1069 non-executable.
1070
1071 TODO: The inability to handle objects bigger than BLOCK_SIZE_W means that
1072 the linker cannot use allocateExec for loading object code files
1073 on Windows. Once allocateExec can handle larger objects, the linker
1074 should be modified to use allocateExec instead of VirtualAlloc.
1075 ------------------------------------------------------------------------- */
1076
1077 #if defined(linux_HOST_OS)
1078
1079 // On Linux we need to use libffi for allocating executable memory,
1080 // because it knows how to work around the restrictions put in place
1081 // by SELinux.
1082
1083 void *allocateExec (nat bytes, void **exec_ret)
1084 {
1085 void **ret, **exec;
1086 ACQUIRE_SM_LOCK;
1087 ret = ffi_closure_alloc (sizeof(void *) + (size_t)bytes, (void**)&exec);
1088 RELEASE_SM_LOCK;
1089 if (ret == NULL) return ret;
1090 *ret = ret; // save the address of the writable mapping, for freeExec().
1091 *exec_ret = exec + 1;
1092 return (ret + 1);
1093 }
1094
1095 // freeExec gets passed the executable address, not the writable address.
1096 void freeExec (void *addr)
1097 {
1098 void *writable;
1099 writable = *((void**)addr - 1);
1100 ACQUIRE_SM_LOCK;
1101 ffi_closure_free (writable);
1102 RELEASE_SM_LOCK
1103 }
1104
1105 #else
1106
1107 void *allocateExec (nat bytes, void **exec_ret)
1108 {
1109 void *ret;
1110 nat n;
1111
1112 ACQUIRE_SM_LOCK;
1113
1114 // round up to words.
1115 n = (bytes + sizeof(W_) + 1) / sizeof(W_);
1116
1117 if (n+1 > BLOCK_SIZE_W) {
1118 barf("allocateExec: can't handle large objects");
1119 }
1120
1121 if (exec_block == NULL ||
1122 exec_block->free + n + 1 > exec_block->start + BLOCK_SIZE_W) {
1123 bdescr *bd;
1124 lnat pagesize = getPageSize();
1125 bd = allocGroup(stg_max(1, pagesize / BLOCK_SIZE));
1126 debugTrace(DEBUG_gc, "allocate exec block %p", bd->start);
1127 bd->gen_no = 0;
1128 bd->flags = BF_EXEC;
1129 bd->link = exec_block;
1130 if (exec_block != NULL) {
1131 exec_block->u.back = bd;
1132 }
1133 bd->u.back = NULL;
1134 setExecutable(bd->start, bd->blocks * BLOCK_SIZE, rtsTrue);
1135 exec_block = bd;
1136 }
1137 *(exec_block->free) = n; // store the size of this chunk
1138 exec_block->gen_no += n; // gen_no stores the number of words allocated
1139 ret = exec_block->free + 1;
1140 exec_block->free += n + 1;
1141
1142 RELEASE_SM_LOCK
1143 *exec_ret = ret;
1144 return ret;
1145 }
1146
1147 void freeExec (void *addr)
1148 {
1149 StgPtr p = (StgPtr)addr - 1;
1150 bdescr *bd = Bdescr((StgPtr)p);
1151
1152 if ((bd->flags & BF_EXEC) == 0) {
1153 barf("freeExec: not executable");
1154 }
1155
1156 if (*(StgPtr)p == 0) {
1157 barf("freeExec: already free?");
1158 }
1159
1160 ACQUIRE_SM_LOCK;
1161
1162 bd->gen_no -= *(StgPtr)p;
1163 *(StgPtr)p = 0;
1164
1165 if (bd->gen_no == 0) {
1166 // Free the block if it is empty, but not if it is the block at
1167 // the head of the queue.
1168 if (bd != exec_block) {
1169 debugTrace(DEBUG_gc, "free exec block %p", bd->start);
1170 dbl_link_remove(bd, &exec_block);
1171 setExecutable(bd->start, bd->blocks * BLOCK_SIZE, rtsFalse);
1172 freeGroup(bd);
1173 } else {
1174 bd->free = bd->start;
1175 }
1176 }
1177
1178 RELEASE_SM_LOCK
1179 }
1180
1181 #endif /* mingw32_HOST_OS */
1182
1183 /* -----------------------------------------------------------------------------
1184 Debugging
1185
1186 memInventory() checks for memory leaks by counting up all the
1187 blocks we know about and comparing that to the number of blocks
1188 allegedly floating around in the system.
1189 -------------------------------------------------------------------------- */
1190
1191 #ifdef DEBUG
1192
1193 // Useful for finding partially full blocks in gdb
1194 void findSlop(bdescr *bd);
1195 void findSlop(bdescr *bd)
1196 {
1197 lnat slop;
1198
1199 for (; bd != NULL; bd = bd->link) {
1200 slop = (bd->blocks * BLOCK_SIZE_W) - (bd->free - bd->start);
1201 if (slop > (1024/sizeof(W_))) {
1202 debugBelch("block at %p (bdescr %p) has %ldKB slop\n",
1203 bd->start, bd, slop / (1024/sizeof(W_)));
1204 }
1205 }
1206 }
1207
1208 nat
1209 countBlocks(bdescr *bd)
1210 {
1211 nat n;
1212 for (n=0; bd != NULL; bd=bd->link) {
1213 n += bd->blocks;
1214 }
1215 return n;
1216 }
1217
1218 // (*1) Just like countBlocks, except that we adjust the count for a
1219 // megablock group so that it doesn't include the extra few blocks
1220 // that would be taken up by block descriptors in the second and
1221 // subsequent megablock. This is so we can tally the count with the
1222 // number of blocks allocated in the system, for memInventory().
1223 static nat
1224 countAllocdBlocks(bdescr *bd)
1225 {
1226 nat n;
1227 for (n=0; bd != NULL; bd=bd->link) {
1228 n += bd->blocks;
1229 // hack for megablock groups: see (*1) above
1230 if (bd->blocks > BLOCKS_PER_MBLOCK) {
1231 n -= (MBLOCK_SIZE / BLOCK_SIZE - BLOCKS_PER_MBLOCK)
1232 * (bd->blocks/(MBLOCK_SIZE/BLOCK_SIZE));
1233 }
1234 }
1235 return n;
1236 }
1237
1238 static lnat
1239 stepBlocks (step *stp)
1240 {
1241 ASSERT(countBlocks(stp->blocks) == stp->n_blocks);
1242 ASSERT(countBlocks(stp->large_objects) == stp->n_large_blocks);
1243 return stp->n_blocks + stp->n_old_blocks +
1244 countAllocdBlocks(stp->large_objects);
1245 }
1246
1247 // If memInventory() calculates that we have a memory leak, this
1248 // function will try to find the block(s) that are leaking by marking
1249 // all the ones that we know about, and search through memory to find
1250 // blocks that are not marked. In the debugger this can help to give
1251 // us a clue about what kind of block leaked. In the future we might
1252 // annotate blocks with their allocation site to give more helpful
1253 // info.
1254 static void
1255 findMemoryLeak (void)
1256 {
1257 nat g, s, i;
1258 for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
1259 for (i = 0; i < n_capabilities; i++) {
1260 markBlocks(capabilities[i].mut_lists[g]);
1261 }
1262 markBlocks(generations[g].mut_list);
1263 for (s = 0; s < generations[g].n_steps; s++) {
1264 markBlocks(generations[g].steps[s].blocks);
1265 markBlocks(generations[g].steps[s].large_objects);
1266 }
1267 }
1268
1269 for (i = 0; i < n_nurseries; i++) {
1270 markBlocks(nurseries[i].blocks);
1271 markBlocks(nurseries[i].large_objects);
1272 }
1273
1274 #ifdef PROFILING
1275 // TODO:
1276 // if (RtsFlags.ProfFlags.doHeapProfile == HEAP_BY_RETAINER) {
1277 // markRetainerBlocks();
1278 // }
1279 #endif
1280
1281 // count the blocks allocated by the arena allocator
1282 // TODO:
1283 // markArenaBlocks();
1284
1285 // count the blocks containing executable memory
1286 markBlocks(exec_block);
1287
1288 reportUnmarkedBlocks();
1289 }
1290
1291
1292 void
1293 memInventory (rtsBool show)
1294 {
1295 nat g, s, i;
1296 step *stp;
1297 lnat gen_blocks[RtsFlags.GcFlags.generations];
1298 lnat nursery_blocks, retainer_blocks,
1299 arena_blocks, exec_blocks;
1300 lnat live_blocks = 0, free_blocks = 0;
1301 rtsBool leak;
1302
1303 // count the blocks we current have
1304
1305 for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
1306 gen_blocks[g] = 0;
1307 for (i = 0; i < n_capabilities; i++) {
1308 gen_blocks[g] += countBlocks(capabilities[i].mut_lists[g]);
1309 }
1310 gen_blocks[g] += countAllocdBlocks(generations[g].mut_list);
1311 for (s = 0; s < generations[g].n_steps; s++) {
1312 stp = &generations[g].steps[s];
1313 gen_blocks[g] += stepBlocks(stp);
1314 }
1315 }
1316
1317 nursery_blocks = 0;
1318 for (i = 0; i < n_nurseries; i++) {
1319 nursery_blocks += stepBlocks(&nurseries[i]);
1320 }
1321
1322 retainer_blocks = 0;
1323 #ifdef PROFILING
1324 if (RtsFlags.ProfFlags.doHeapProfile == HEAP_BY_RETAINER) {
1325 retainer_blocks = retainerStackBlocks();
1326 }
1327 #endif
1328
1329 // count the blocks allocated by the arena allocator
1330 arena_blocks = arenaBlocks();
1331
1332 // count the blocks containing executable memory
1333 exec_blocks = countAllocdBlocks(exec_block);
1334
1335 /* count the blocks on the free list */
1336 free_blocks = countFreeList();
1337
1338 live_blocks = 0;
1339 for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
1340 live_blocks += gen_blocks[g];
1341 }
1342 live_blocks += nursery_blocks +
1343 + retainer_blocks + arena_blocks + exec_blocks;
1344
1345 #define MB(n) (((n) * BLOCK_SIZE_W) / ((1024*1024)/sizeof(W_)))
1346
1347 leak = live_blocks + free_blocks != mblocks_allocated * BLOCKS_PER_MBLOCK;
1348
1349 if (show || leak)
1350 {
1351 if (leak) {
1352 debugBelch("Memory leak detected:\n");
1353 } else {
1354 debugBelch("Memory inventory:\n");
1355 }
1356 for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
1357 debugBelch(" gen %d blocks : %5lu blocks (%lu MB)\n", g,
1358 gen_blocks[g], MB(gen_blocks[g]));
1359 }
1360 debugBelch(" nursery : %5lu blocks (%lu MB)\n",
1361 nursery_blocks, MB(nursery_blocks));
1362 debugBelch(" retainer : %5lu blocks (%lu MB)\n",
1363 retainer_blocks, MB(retainer_blocks));
1364 debugBelch(" arena blocks : %5lu blocks (%lu MB)\n",
1365 arena_blocks, MB(arena_blocks));
1366 debugBelch(" exec : %5lu blocks (%lu MB)\n",
1367 exec_blocks, MB(exec_blocks));
1368 debugBelch(" free : %5lu blocks (%lu MB)\n",
1369 free_blocks, MB(free_blocks));
1370 debugBelch(" total : %5lu blocks (%lu MB)\n",
1371 live_blocks + free_blocks, MB(live_blocks+free_blocks));
1372 if (leak) {
1373 debugBelch("\n in system : %5lu blocks (%lu MB)\n",
1374 mblocks_allocated * BLOCKS_PER_MBLOCK, mblocks_allocated);
1375 }
1376 }
1377
1378 if (leak) {
1379 debugBelch("\n");
1380 findMemoryLeak();
1381 }
1382 ASSERT(n_alloc_blocks == live_blocks);
1383 ASSERT(!leak);
1384 }
1385
1386
1387 /* Full heap sanity check. */
1388 void
1389 checkSanity( void )
1390 {
1391 nat g, s;
1392
1393 if (RtsFlags.GcFlags.generations == 1) {
1394 checkHeap(g0s0->blocks);
1395 checkLargeObjects(g0s0->large_objects);
1396 } else {
1397
1398 for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
1399 for (s = 0; s < generations[g].n_steps; s++) {
1400 if (g == 0 && s == 0) { continue; }
1401 ASSERT(countBlocks(generations[g].steps[s].blocks)
1402 == generations[g].steps[s].n_blocks);
1403 ASSERT(countBlocks(generations[g].steps[s].large_objects)
1404 == generations[g].steps[s].n_large_blocks);
1405 checkHeap(generations[g].steps[s].blocks);
1406 checkLargeObjects(generations[g].steps[s].large_objects);
1407 }
1408 }
1409
1410 for (s = 0; s < n_nurseries; s++) {
1411 ASSERT(countBlocks(nurseries[s].blocks)
1412 == nurseries[s].n_blocks);
1413 ASSERT(countBlocks(nurseries[s].large_objects)
1414 == nurseries[s].n_large_blocks);
1415 }
1416
1417 checkFreeListSanity();
1418 }
1419
1420 #if defined(THREADED_RTS)
1421 // check the stacks too in threaded mode, because we don't do a
1422 // full heap sanity check in this case (see checkHeap())
1423 checkMutableLists(rtsTrue);
1424 #else
1425 checkMutableLists(rtsFalse);
1426 #endif
1427 }
1428
1429 /* Nursery sanity check */
1430 void
1431 checkNurserySanity( step *stp )
1432 {
1433 bdescr *bd, *prev;
1434 nat blocks = 0;
1435
1436 prev = NULL;
1437 for (bd = stp->blocks; bd != NULL; bd = bd->link) {
1438 ASSERT(bd->u.back == prev);
1439 prev = bd;
1440 blocks += bd->blocks;
1441 }
1442 ASSERT(blocks == stp->n_blocks);
1443 }
1444
1445 // handy function for use in gdb, because Bdescr() is inlined.
1446 extern bdescr *_bdescr( StgPtr p );
1447
1448 bdescr *
1449 _bdescr( StgPtr p )
1450 {
1451 return Bdescr(p);
1452 }
1453
1454 #endif