1 /* -----------------------------------------------------------------------------
3 * (c) The GHC Team, 1998-2004
5 * Storage manager front end
7 * ---------------------------------------------------------------------------*/
9 #include "PosixSource.h"
15 #include "BlockAlloc.h"
20 #include "OSThreads.h"
21 #include "Capability.h"
24 #include "RetainerProfile.h" // for counting memory blocks (memInventory)
30 * All these globals require sm_mutex to access in SMP mode.
32 StgClosure *caf_list = NULL;
33 StgClosure *revertible_caf_list = NULL;
36 bdescr *small_alloc_list; /* allocate()d small objects */
37 bdescr *pinned_object_block; /* allocate pinned objects into this block */
38 nat alloc_blocks; /* number of allocate()d blocks since GC */
39 nat alloc_blocks_lim; /* approximate limit on alloc_blocks */
41 StgPtr alloc_Hp = NULL; /* next free byte in small_alloc_list */
42 StgPtr alloc_HpLim = NULL; /* end of block at small_alloc_list */
44 generation *generations = NULL; /* all the generations */
45 generation *g0 = NULL; /* generation 0, for convenience */
46 generation *oldest_gen = NULL; /* oldest generation, for convenience */
47 step *g0s0 = NULL; /* generation 0, step 0, for convenience */
49 ullong total_allocated = 0; /* total memory allocated during run */
51 nat n_nurseries = 0; /* == RtsFlags.ParFlags.nNodes, convenience */
52 step *nurseries = NULL; /* array of nurseries, >1 only if SMP */
55 * Storage manager mutex: protects all the above state from
56 * simultaneous access by two STG threads.
59 Mutex sm_mutex = INIT_MUTEX_VAR;
65 static void *stgAllocForGMP (size_t size_in_bytes);
66 static void *stgReallocForGMP (void *ptr, size_t old_size, size_t new_size);
67 static void stgDeallocForGMP (void *ptr, size_t size);
70 initStep (step *stp, int g, int s)
76 stp->gen = &generations[g];
83 stp->large_objects = NULL;
84 stp->n_large_blocks = 0;
85 stp->new_large_objects = NULL;
86 stp->scavenged_large_objects = NULL;
87 stp->n_scavenged_large_blocks = 0;
88 stp->is_compacted = 0;
98 if (generations != NULL) {
99 // multi-init protection
103 /* Sanity check to make sure the LOOKS_LIKE_ macros appear to be
104 * doing something reasonable.
106 ASSERT(LOOKS_LIKE_INFO_PTR(&stg_BLACKHOLE_info));
107 ASSERT(LOOKS_LIKE_CLOSURE_PTR(&stg_dummy_ret_closure));
108 ASSERT(!HEAP_ALLOCED(&stg_dummy_ret_closure));
110 if (RtsFlags.GcFlags.maxHeapSize != 0 &&
111 RtsFlags.GcFlags.heapSizeSuggestion >
112 RtsFlags.GcFlags.maxHeapSize) {
113 RtsFlags.GcFlags.maxHeapSize = RtsFlags.GcFlags.heapSizeSuggestion;
116 if (RtsFlags.GcFlags.maxHeapSize != 0 &&
117 RtsFlags.GcFlags.minAllocAreaSize >
118 RtsFlags.GcFlags.maxHeapSize) {
119 errorBelch("maximum heap size (-M) is smaller than minimum alloc area size (-A)");
123 initBlockAllocator();
126 initMutex(&sm_mutex);
129 /* allocate generation info array */
130 generations = (generation *)stgMallocBytes(RtsFlags.GcFlags.generations
131 * sizeof(struct generation_),
132 "initStorage: gens");
134 /* Initialise all generations */
135 for(g = 0; g < RtsFlags.GcFlags.generations; g++) {
136 gen = &generations[g];
138 gen->mut_list = allocBlock();
139 gen->collections = 0;
140 gen->failed_promotions = 0;
144 /* A couple of convenience pointers */
145 g0 = &generations[0];
146 oldest_gen = &generations[RtsFlags.GcFlags.generations-1];
148 /* Allocate step structures in each generation */
149 if (RtsFlags.GcFlags.generations > 1) {
150 /* Only for multiple-generations */
152 /* Oldest generation: one step */
153 oldest_gen->n_steps = 1;
155 stgMallocBytes(1 * sizeof(struct step_), "initStorage: last step");
157 /* set up all except the oldest generation with 2 steps */
158 for(g = 0; g < RtsFlags.GcFlags.generations-1; g++) {
159 generations[g].n_steps = RtsFlags.GcFlags.steps;
160 generations[g].steps =
161 stgMallocBytes (RtsFlags.GcFlags.steps * sizeof(struct step_),
162 "initStorage: steps");
166 /* single generation, i.e. a two-space collector */
168 g0->steps = stgMallocBytes (sizeof(struct step_), "initStorage: steps");
172 n_nurseries = RtsFlags.ParFlags.nNodes;
173 nurseries = stgMallocBytes (n_nurseries * sizeof(struct step_),
174 "initStorage: nurseries");
177 nurseries = g0->steps; // just share nurseries[0] with g0s0
180 /* Initialise all steps */
181 for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
182 for (s = 0; s < generations[g].n_steps; s++) {
183 initStep(&generations[g].steps[s], g, s);
188 for (s = 0; s < n_nurseries; s++) {
189 initStep(&nurseries[s], 0, s);
193 /* Set up the destination pointers in each younger gen. step */
194 for (g = 0; g < RtsFlags.GcFlags.generations-1; g++) {
195 for (s = 0; s < generations[g].n_steps-1; s++) {
196 generations[g].steps[s].to = &generations[g].steps[s+1];
198 generations[g].steps[s].to = &generations[g+1].steps[0];
200 oldest_gen->steps[0].to = &oldest_gen->steps[0];
203 for (s = 0; s < n_nurseries; s++) {
204 nurseries[s].to = generations[0].steps[0].to;
208 /* The oldest generation has one step. */
209 if (RtsFlags.GcFlags.compact) {
210 if (RtsFlags.GcFlags.generations == 1) {
211 errorBelch("WARNING: compaction is incompatible with -G1; disabled");
213 oldest_gen->steps[0].is_compacted = 1;
218 if (RtsFlags.GcFlags.generations == 1) {
219 errorBelch("-G1 is incompatible with SMP");
224 /* generation 0 is special: that's the nursery */
225 generations[0].max_blocks = 0;
227 /* G0S0: the allocation area. Policy: keep the allocation area
228 * small to begin with, even if we have a large suggested heap
229 * size. Reason: we're going to do a major collection first, and we
230 * don't want it to be a big one. This vague idea is borne out by
231 * rigorous experimental evidence.
233 g0s0 = &generations[0].steps[0];
237 weak_ptr_list = NULL;
239 revertible_caf_list = NULL;
241 /* initialise the allocate() interface */
242 small_alloc_list = NULL;
244 alloc_blocks_lim = RtsFlags.GcFlags.minAllocAreaSize;
246 /* Tell GNU multi-precision pkg about our custom alloc functions */
247 mp_set_memory_functions(stgAllocForGMP, stgReallocForGMP, stgDeallocForGMP);
249 IF_DEBUG(gc, statDescribeGens());
255 stat_exit(calcAllocated());
258 /* -----------------------------------------------------------------------------
261 The entry code for every CAF does the following:
263 - builds a CAF_BLACKHOLE in the heap
264 - pushes an update frame pointing to the CAF_BLACKHOLE
265 - invokes UPD_CAF(), which:
266 - calls newCaf, below
267 - updates the CAF with a static indirection to the CAF_BLACKHOLE
269 Why do we build a BLACKHOLE in the heap rather than just updating
270 the thunk directly? It's so that we only need one kind of update
271 frame - otherwise we'd need a static version of the update frame too.
273 newCaf() does the following:
275 - it puts the CAF on the oldest generation's mut-once list.
276 This is so that we can treat the CAF as a root when collecting
279 For GHCI, we have additional requirements when dealing with CAFs:
281 - we must *retain* all dynamically-loaded CAFs ever entered,
282 just in case we need them again.
283 - we must be able to *revert* CAFs that have been evaluated, to
284 their pre-evaluated form.
286 To do this, we use an additional CAF list. When newCaf() is
287 called on a dynamically-loaded CAF, we add it to the CAF list
288 instead of the old-generation mutable list, and save away its
289 old info pointer (in caf->saved_info) for later reversion.
291 To revert all the CAFs, we traverse the CAF list and reset the
292 info pointer to caf->saved_info, then throw away the CAF list.
293 (see GC.c:revertCAFs()).
297 -------------------------------------------------------------------------- */
300 newCAF(StgClosure* caf)
307 // If we are in GHCi _and_ we are using dynamic libraries,
308 // then we can't redirect newCAF calls to newDynCAF (see below),
309 // so we make newCAF behave almost like newDynCAF.
310 // The dynamic libraries might be used by both the interpreted
311 // program and GHCi itself, so they must not be reverted.
312 // This also means that in GHCi with dynamic libraries, CAFs are not
313 // garbage collected. If this turns out to be a problem, we could
314 // do another hack here and do an address range test on caf to figure
315 // out whether it is from a dynamic library.
316 ((StgIndStatic *)caf)->saved_info = (StgInfoTable *)caf->header.info;
317 ((StgIndStatic *)caf)->static_link = caf_list;
322 /* Put this CAF on the mutable list for the old generation.
323 * This is a HACK - the IND_STATIC closure doesn't really have
324 * a mut_link field, but we pretend it has - in fact we re-use
325 * the STATIC_LINK field for the time being, because when we
326 * come to do a major GC we won't need the mut_link field
327 * any more and can use it as a STATIC_LINK.
329 ((StgIndStatic *)caf)->saved_info = NULL;
330 recordMutableGen(caf, oldest_gen);
336 /* If we are PAR or DIST then we never forget a CAF */
338 //debugBelch("<##> Globalising CAF %08x %s",caf,info_type(caf));
339 newGA=makeGlobal(caf,rtsTrue); /*given full weight*/
345 // An alternate version of newCaf which is used for dynamically loaded
346 // object code in GHCi. In this case we want to retain *all* CAFs in
347 // the object code, because they might be demanded at any time from an
348 // expression evaluated on the command line.
349 // Also, GHCi might want to revert CAFs, so we add these to the
350 // revertible_caf_list.
352 // The linker hackily arranges that references to newCaf from dynamic
353 // code end up pointing to newDynCAF.
355 newDynCAF(StgClosure *caf)
359 ((StgIndStatic *)caf)->saved_info = (StgInfoTable *)caf->header.info;
360 ((StgIndStatic *)caf)->static_link = revertible_caf_list;
361 revertible_caf_list = caf;
366 /* -----------------------------------------------------------------------------
368 -------------------------------------------------------------------------- */
371 allocNursery (step *stp, bdescr *tail, nat blocks)
376 // Allocate a nursery: we allocate fresh blocks one at a time and
377 // cons them on to the front of the list, not forgetting to update
378 // the back pointer on the tail of the list to point to the new block.
379 for (i=0; i < blocks; i++) {
382 processNursery() in LdvProfile.c assumes that every block group in
383 the nursery contains only a single block. So, if a block group is
384 given multiple blocks, change processNursery() accordingly.
388 // double-link the nursery: we might need to insert blocks
395 bd->free = bd->start;
403 assignNurseriesToCapabilities (void)
408 for (i = 0; i < n_nurseries; i++) {
409 capabilities[i].r.rNursery = &nurseries[i];
410 capabilities[i].r.rCurrentNursery = nurseries[i].blocks;
411 capabilities[i].r.rCurrentAlloc = NULL;
414 MainCapability.r.rNursery = &nurseries[0];
415 MainCapability.r.rCurrentNursery = nurseries[0].blocks;
416 MainCapability.r.rCurrentAlloc = NULL;
421 allocNurseries( void )
425 for (i = 0; i < n_nurseries; i++) {
426 nurseries[i].blocks =
427 allocNursery(&nurseries[i], NULL,
428 RtsFlags.GcFlags.minAllocAreaSize);
429 nurseries[i].n_blocks = RtsFlags.GcFlags.minAllocAreaSize;
430 nurseries[i].to_blocks = NULL;
431 nurseries[i].n_to_blocks = 0;
432 /* hp, hpLim, hp_bd, to_space etc. aren't used in the nursery */
434 assignNurseriesToCapabilities();
438 resetNurseries( void )
444 for (i = 0; i < n_nurseries; i++) {
446 for (bd = stp->blocks; bd; bd = bd->link) {
447 bd->free = bd->start;
448 ASSERT(bd->gen_no == 0);
449 ASSERT(bd->step == stp);
450 IF_DEBUG(sanity,memset(bd->start, 0xaa, BLOCK_SIZE));
453 assignNurseriesToCapabilities();
457 countNurseryBlocks (void)
462 for (i = 0; i < n_nurseries; i++) {
463 blocks += nurseries[i].n_blocks;
469 resizeNursery ( step *stp, nat blocks )
474 nursery_blocks = stp->n_blocks;
475 if (nursery_blocks == blocks) return;
477 if (nursery_blocks < blocks) {
478 IF_DEBUG(gc, debugBelch("Increasing size of nursery to %d blocks\n",
480 stp->blocks = allocNursery(stp, stp->blocks, blocks-nursery_blocks);
485 IF_DEBUG(gc, debugBelch("Decreasing size of nursery to %d blocks\n",
489 while (nursery_blocks > blocks) {
491 next_bd->u.back = NULL;
492 nursery_blocks -= bd->blocks; // might be a large block
497 // might have gone just under, by freeing a large block, so make
498 // up the difference.
499 if (nursery_blocks < blocks) {
500 stp->blocks = allocNursery(stp, stp->blocks, blocks-nursery_blocks);
504 stp->n_blocks = blocks;
505 ASSERT(countBlocks(stp->blocks) == stp->n_blocks);
509 // Resize each of the nurseries to the specified size.
512 resizeNurseriesFixed (nat blocks)
515 for (i = 0; i < n_nurseries; i++) {
516 resizeNursery(&nurseries[i], blocks);
521 // Resize the nurseries to the total specified size.
524 resizeNurseries (nat blocks)
526 // If there are multiple nurseries, then we just divide the number
527 // of available blocks between them.
528 resizeNurseriesFixed(blocks / n_nurseries);
531 /* -----------------------------------------------------------------------------
532 The allocate() interface
534 allocate(n) always succeeds, and returns a chunk of memory n words
535 long. n can be larger than the size of a block if necessary, in
536 which case a contiguous block group will be allocated.
537 -------------------------------------------------------------------------- */
547 TICK_ALLOC_HEAP_NOCTR(n);
550 /* big allocation (>LARGE_OBJECT_THRESHOLD) */
551 /* ToDo: allocate directly into generation 1 */
552 if (n >= LARGE_OBJECT_THRESHOLD/sizeof(W_)) {
553 nat req_blocks = (lnat)BLOCK_ROUND_UP(n*sizeof(W_)) / BLOCK_SIZE;
554 bd = allocGroup(req_blocks);
555 dbl_link_onto(bd, &g0s0->large_objects);
556 g0s0->n_large_blocks += req_blocks;
559 bd->flags = BF_LARGE;
560 bd->free = bd->start + n;
561 alloc_blocks += req_blocks;
565 /* small allocation (<LARGE_OBJECT_THRESHOLD) */
566 } else if (small_alloc_list == NULL || alloc_Hp + n > alloc_HpLim) {
567 if (small_alloc_list) {
568 small_alloc_list->free = alloc_Hp;
571 bd->link = small_alloc_list;
572 small_alloc_list = bd;
576 alloc_Hp = bd->start;
577 alloc_HpLim = bd->start + BLOCK_SIZE_W;
588 allocated_bytes( void )
592 allocated = alloc_blocks * BLOCK_SIZE_W - (alloc_HpLim - alloc_Hp);
593 if (pinned_object_block != NULL) {
594 allocated -= (pinned_object_block->start + BLOCK_SIZE_W) -
595 pinned_object_block->free;
602 tidyAllocateLists (void)
604 if (small_alloc_list != NULL) {
605 ASSERT(alloc_Hp >= small_alloc_list->start &&
606 alloc_Hp <= small_alloc_list->start + BLOCK_SIZE);
607 small_alloc_list->free = alloc_Hp;
611 /* -----------------------------------------------------------------------------
614 This allocates memory in the current thread - it is intended for
615 use primarily from STG-land where we have a Capability. It is
616 better than allocate() because it doesn't require taking the
617 sm_mutex lock in the common case.
619 Memory is allocated directly from the nursery if possible (but not
620 from the current nursery block, so as not to interfere with
622 -------------------------------------------------------------------------- */
625 allocateLocal( StgRegTable *reg, nat n )
630 TICK_ALLOC_HEAP_NOCTR(n);
633 /* big allocation (>LARGE_OBJECT_THRESHOLD) */
634 /* ToDo: allocate directly into generation 1 */
635 if (n >= LARGE_OBJECT_THRESHOLD/sizeof(W_)) {
636 nat req_blocks = (lnat)BLOCK_ROUND_UP(n*sizeof(W_)) / BLOCK_SIZE;
638 bd = allocGroup(req_blocks);
639 dbl_link_onto(bd, &g0s0->large_objects);
640 g0s0->n_large_blocks += req_blocks;
643 bd->flags = BF_LARGE;
644 bd->free = bd->start + n;
645 alloc_blocks += req_blocks;
649 /* small allocation (<LARGE_OBJECT_THRESHOLD) */
652 bd = reg->rCurrentAlloc;
653 if (bd == NULL || bd->free + n > bd->start + BLOCK_SIZE_W) {
655 // The CurrentAlloc block is full, we need to find another
656 // one. First, we try taking the next block from the
658 bd = reg->rCurrentNursery->link;
660 if (bd == NULL || bd->free + n > bd->start + BLOCK_SIZE_W) {
661 // The nursery is empty, or the next block is already
662 // full: allocate a fresh block (we can't fail here).
671 // we have a block in the nursery: take it and put
672 // it at the *front* of the nursery list, and use it
673 // to allocate() from.
674 reg->rCurrentNursery->link = bd->link;
676 bd->link = reg->rNursery->blocks;
677 reg->rNursery->blocks = bd;
679 reg->rCurrentAlloc = bd;
687 /* ---------------------------------------------------------------------------
688 Allocate a fixed/pinned object.
690 We allocate small pinned objects into a single block, allocating a
691 new block when the current one overflows. The block is chained
692 onto the large_object_list of generation 0 step 0.
694 NOTE: The GC can't in general handle pinned objects. This
695 interface is only safe to use for ByteArrays, which have no
696 pointers and don't require scavenging. It works because the
697 block's descriptor has the BF_LARGE flag set, so the block is
698 treated as a large object and chained onto various lists, rather
699 than the individual objects being copied. However, when it comes
700 to scavenge the block, the GC will only scavenge the first object.
701 The reason is that the GC can't linearly scan a block of pinned
702 objects at the moment (doing so would require using the
703 mostly-copying techniques). But since we're restricting ourselves
704 to pinned ByteArrays, not scavenging is ok.
706 This function is called by newPinnedByteArray# which immediately
707 fills the allocated memory with a MutableByteArray#.
708 ------------------------------------------------------------------------- */
711 allocatePinned( nat n )
714 bdescr *bd = pinned_object_block;
716 // If the request is for a large object, then allocate()
717 // will give us a pinned object anyway.
718 if (n >= LARGE_OBJECT_THRESHOLD/sizeof(W_)) {
724 TICK_ALLOC_HEAP_NOCTR(n);
727 // we always return 8-byte aligned memory. bd->free must be
728 // 8-byte aligned to begin with, so we just round up n to
729 // the nearest multiple of 8 bytes.
730 if (sizeof(StgWord) == 4) {
734 // If we don't have a block of pinned objects yet, or the current
735 // one isn't large enough to hold the new object, allocate a new one.
736 if (bd == NULL || (bd->free + n) > (bd->start + BLOCK_SIZE_W)) {
737 pinned_object_block = bd = allocBlock();
738 dbl_link_onto(bd, &g0s0->large_objects);
741 bd->flags = BF_PINNED | BF_LARGE;
742 bd->free = bd->start;
752 /* -----------------------------------------------------------------------------
753 Allocation functions for GMP.
755 These all use the allocate() interface - we can't have any garbage
756 collection going on during a gmp operation, so we use allocate()
757 which always succeeds. The gmp operations which might need to
758 allocate will ask the storage manager (via doYouWantToGC()) whether
759 a garbage collection is required, in case we get into a loop doing
760 only allocate() style allocation.
761 -------------------------------------------------------------------------- */
764 stgAllocForGMP (size_t size_in_bytes)
767 nat data_size_in_words, total_size_in_words;
769 /* round up to a whole number of words */
770 data_size_in_words = (size_in_bytes + sizeof(W_) + 1) / sizeof(W_);
771 total_size_in_words = sizeofW(StgArrWords) + data_size_in_words;
773 /* allocate and fill it in. */
775 arr = (StgArrWords *)allocateLocal(&(myCapability()->r), total_size_in_words);
777 arr = (StgArrWords *)allocateLocal(&MainCapability.r, total_size_in_words);
779 SET_ARR_HDR(arr, &stg_ARR_WORDS_info, CCCS, data_size_in_words);
781 /* and return a ptr to the goods inside the array */
786 stgReallocForGMP (void *ptr, size_t old_size, size_t new_size)
788 void *new_stuff_ptr = stgAllocForGMP(new_size);
790 char *p = (char *) ptr;
791 char *q = (char *) new_stuff_ptr;
793 for (; i < old_size; i++, p++, q++) {
797 return(new_stuff_ptr);
801 stgDeallocForGMP (void *ptr STG_UNUSED,
802 size_t size STG_UNUSED)
804 /* easy for us: the garbage collector does the dealloc'n */
807 /* -----------------------------------------------------------------------------
809 * -------------------------------------------------------------------------- */
811 /* -----------------------------------------------------------------------------
814 * Approximate how much we've allocated: number of blocks in the
815 * nursery + blocks allocated via allocate() - unused nusery blocks.
816 * This leaves a little slop at the end of each block, and doesn't
817 * take into account large objects (ToDo).
818 * -------------------------------------------------------------------------- */
821 calcAllocated( void )
827 allocated = allocated_bytes();
828 allocated += countNurseryBlocks() * BLOCK_SIZE_W;
831 for (i = 0; i < n_nurseries; i++) {
833 for ( bd = capabilities[i].r.rCurrentNursery->link;
834 bd != NULL; bd = bd->link ) {
835 allocated -= BLOCK_SIZE_W;
837 cap = &capabilities[i];
838 if (cap->r.rCurrentNursery->free <
839 cap->r.rCurrentNursery->start + BLOCK_SIZE_W) {
840 allocated -= (cap->r.rCurrentNursery->start + BLOCK_SIZE_W)
841 - cap->r.rCurrentNursery->free;
845 bdescr *current_nursery = MainCapability.r.rCurrentNursery;
847 for ( bd = current_nursery->link; bd != NULL; bd = bd->link ) {
848 allocated -= BLOCK_SIZE_W;
850 if (current_nursery->free < current_nursery->start + BLOCK_SIZE_W) {
851 allocated -= (current_nursery->start + BLOCK_SIZE_W)
852 - current_nursery->free;
856 total_allocated += allocated;
860 /* Approximate the amount of live data in the heap. To be called just
861 * after garbage collection (see GarbageCollect()).
870 if (RtsFlags.GcFlags.generations == 1) {
871 live = (g0s0->n_to_blocks - 1) * BLOCK_SIZE_W +
872 ((lnat)g0s0->hp_bd->free - (lnat)g0s0->hp_bd->start) / sizeof(W_);
876 for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
877 for (s = 0; s < generations[g].n_steps; s++) {
878 /* approximate amount of live data (doesn't take into account slop
879 * at end of each block).
881 if (g == 0 && s == 0) {
884 stp = &generations[g].steps[s];
885 live += (stp->n_large_blocks + stp->n_blocks - 1) * BLOCK_SIZE_W;
886 if (stp->hp_bd != NULL) {
887 live += ((lnat)stp->hp_bd->free - (lnat)stp->hp_bd->start)
895 /* Approximate the number of blocks that will be needed at the next
896 * garbage collection.
898 * Assume: all data currently live will remain live. Steps that will
899 * be collected next time will therefore need twice as many blocks
900 * since all the data will be copied.
909 for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
910 for (s = 0; s < generations[g].n_steps; s++) {
911 if (g == 0 && s == 0) { continue; }
912 stp = &generations[g].steps[s];
913 if (generations[g].steps[0].n_blocks +
914 generations[g].steps[0].n_large_blocks
915 > generations[g].max_blocks
916 && stp->is_compacted == 0) {
917 needed += 2 * stp->n_blocks;
919 needed += stp->n_blocks;
926 /* -----------------------------------------------------------------------------
929 memInventory() checks for memory leaks by counting up all the
930 blocks we know about and comparing that to the number of blocks
931 allegedly floating around in the system.
932 -------------------------------------------------------------------------- */
937 stepBlocks (step *stp)
942 total_blocks = stp->n_blocks;
943 for (bd = stp->large_objects; bd; bd = bd->link) {
944 total_blocks += bd->blocks;
945 /* hack for megablock groups: they have an extra block or two in
946 the second and subsequent megablocks where the block
947 descriptors would normally go.
949 if (bd->blocks > BLOCKS_PER_MBLOCK) {
950 total_blocks -= (MBLOCK_SIZE / BLOCK_SIZE - BLOCKS_PER_MBLOCK)
951 * (bd->blocks/(MBLOCK_SIZE/BLOCK_SIZE));
963 lnat total_blocks = 0, free_blocks = 0;
965 /* count the blocks we current have */
967 for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
968 for (bd = generations[g].mut_list; bd != NULL; bd = bd->link) {
969 total_blocks += bd->blocks;
971 for (s = 0; s < generations[g].n_steps; s++) {
972 if (g==0 && s==0) continue;
973 stp = &generations[g].steps[s];
974 total_blocks += stepBlocks(stp);
978 for (i = 0; i < n_nurseries; i++) {
979 total_blocks += stepBlocks(&nurseries[i]);
982 if (RtsFlags.GcFlags.generations == 1) {
983 /* two-space collector has a to-space too :-) */
984 total_blocks += g0s0->n_to_blocks;
987 /* any blocks held by allocate() */
988 for (bd = small_alloc_list; bd; bd = bd->link) {
989 total_blocks += bd->blocks;
993 if (RtsFlags.ProfFlags.doHeapProfile == HEAP_BY_RETAINER) {
994 total_blocks += retainerStackBlocks();
998 // count the blocks allocated by the arena allocator
999 total_blocks += arenaBlocks();
1001 /* count the blocks on the free list */
1002 free_blocks = countFreeList();
1004 if (total_blocks + free_blocks != mblocks_allocated *
1005 BLOCKS_PER_MBLOCK) {
1006 debugBelch("Blocks: %ld live + %ld free = %ld total (%ld around)\n",
1007 total_blocks, free_blocks, total_blocks + free_blocks,
1008 mblocks_allocated * BLOCKS_PER_MBLOCK);
1011 ASSERT(total_blocks + free_blocks == mblocks_allocated * BLOCKS_PER_MBLOCK);
1016 countBlocks(bdescr *bd)
1019 for (n=0; bd != NULL; bd=bd->link) {
1025 /* Full heap sanity check. */
1031 if (RtsFlags.GcFlags.generations == 1) {
1032 checkHeap(g0s0->to_blocks);
1033 checkChain(g0s0->large_objects);
1036 for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
1037 for (s = 0; s < generations[g].n_steps; s++) {
1038 if (g == 0 && s == 0) { continue; }
1039 ASSERT(countBlocks(generations[g].steps[s].blocks)
1040 == generations[g].steps[s].n_blocks);
1041 ASSERT(countBlocks(generations[g].steps[s].large_objects)
1042 == generations[g].steps[s].n_large_blocks);
1043 checkHeap(generations[g].steps[s].blocks);
1044 checkChain(generations[g].steps[s].large_objects);
1046 checkMutableList(generations[g].mut_list, g);
1051 for (s = 0; s < n_nurseries; s++) {
1052 ASSERT(countBlocks(nurseries[s].blocks)
1053 == nurseries[s].n_blocks);
1054 ASSERT(countBlocks(nurseries[s].large_objects)
1055 == nurseries[s].n_large_blocks);
1058 checkFreeListSanity();
1062 // handy function for use in gdb, because Bdescr() is inlined.
1063 extern bdescr *_bdescr( StgPtr p );