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 THREADED_RTS 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 THREADED_RTS */
55 * Storage manager mutex: protects all the above state from
56 * simultaneous access by two STG threads.
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)
75 stp->old_blocks = NULL;
76 stp->n_old_blocks = 0;
77 stp->gen = &generations[g];
83 stp->scavd_hpLim = NULL;
86 stp->large_objects = NULL;
87 stp->n_large_blocks = 0;
88 stp->new_large_objects = NULL;
89 stp->scavenged_large_objects = NULL;
90 stp->n_scavenged_large_blocks = 0;
91 stp->is_compacted = 0;
101 if (generations != NULL) {
102 // multi-init protection
106 /* Sanity check to make sure the LOOKS_LIKE_ macros appear to be
107 * doing something reasonable.
109 ASSERT(LOOKS_LIKE_INFO_PTR(&stg_BLACKHOLE_info));
110 ASSERT(LOOKS_LIKE_CLOSURE_PTR(&stg_dummy_ret_closure));
111 ASSERT(!HEAP_ALLOCED(&stg_dummy_ret_closure));
113 if (RtsFlags.GcFlags.maxHeapSize != 0 &&
114 RtsFlags.GcFlags.heapSizeSuggestion >
115 RtsFlags.GcFlags.maxHeapSize) {
116 RtsFlags.GcFlags.maxHeapSize = RtsFlags.GcFlags.heapSizeSuggestion;
119 if (RtsFlags.GcFlags.maxHeapSize != 0 &&
120 RtsFlags.GcFlags.minAllocAreaSize >
121 RtsFlags.GcFlags.maxHeapSize) {
122 errorBelch("maximum heap size (-M) is smaller than minimum alloc area size (-A)");
126 initBlockAllocator();
128 #if defined(THREADED_RTS)
129 initMutex(&sm_mutex);
134 /* allocate generation info array */
135 generations = (generation *)stgMallocBytes(RtsFlags.GcFlags.generations
136 * sizeof(struct generation_),
137 "initStorage: gens");
139 /* Initialise all generations */
140 for(g = 0; g < RtsFlags.GcFlags.generations; g++) {
141 gen = &generations[g];
143 gen->mut_list = allocBlock();
144 gen->collections = 0;
145 gen->failed_promotions = 0;
149 /* A couple of convenience pointers */
150 g0 = &generations[0];
151 oldest_gen = &generations[RtsFlags.GcFlags.generations-1];
153 /* Allocate step structures in each generation */
154 if (RtsFlags.GcFlags.generations > 1) {
155 /* Only for multiple-generations */
157 /* Oldest generation: one step */
158 oldest_gen->n_steps = 1;
160 stgMallocBytes(1 * sizeof(struct step_), "initStorage: last step");
162 /* set up all except the oldest generation with 2 steps */
163 for(g = 0; g < RtsFlags.GcFlags.generations-1; g++) {
164 generations[g].n_steps = RtsFlags.GcFlags.steps;
165 generations[g].steps =
166 stgMallocBytes (RtsFlags.GcFlags.steps * sizeof(struct step_),
167 "initStorage: steps");
171 /* single generation, i.e. a two-space collector */
173 g0->steps = stgMallocBytes (sizeof(struct step_), "initStorage: steps");
177 n_nurseries = n_capabilities;
178 nurseries = stgMallocBytes (n_nurseries * sizeof(struct step_),
179 "initStorage: nurseries");
182 nurseries = g0->steps; // just share nurseries[0] with g0s0
185 /* Initialise all steps */
186 for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
187 for (s = 0; s < generations[g].n_steps; s++) {
188 initStep(&generations[g].steps[s], g, s);
193 for (s = 0; s < n_nurseries; s++) {
194 initStep(&nurseries[s], 0, s);
198 /* Set up the destination pointers in each younger gen. step */
199 for (g = 0; g < RtsFlags.GcFlags.generations-1; g++) {
200 for (s = 0; s < generations[g].n_steps-1; s++) {
201 generations[g].steps[s].to = &generations[g].steps[s+1];
203 generations[g].steps[s].to = &generations[g+1].steps[0];
205 oldest_gen->steps[0].to = &oldest_gen->steps[0];
208 for (s = 0; s < n_nurseries; s++) {
209 nurseries[s].to = generations[0].steps[0].to;
213 /* The oldest generation has one step. */
214 if (RtsFlags.GcFlags.compact) {
215 if (RtsFlags.GcFlags.generations == 1) {
216 errorBelch("WARNING: compaction is incompatible with -G1; disabled");
218 oldest_gen->steps[0].is_compacted = 1;
223 if (RtsFlags.GcFlags.generations == 1) {
224 errorBelch("-G1 is incompatible with -threaded");
225 stg_exit(EXIT_FAILURE);
229 /* generation 0 is special: that's the nursery */
230 generations[0].max_blocks = 0;
232 /* G0S0: the allocation area. Policy: keep the allocation area
233 * small to begin with, even if we have a large suggested heap
234 * size. Reason: we're going to do a major collection first, and we
235 * don't want it to be a big one. This vague idea is borne out by
236 * rigorous experimental evidence.
238 g0s0 = &generations[0].steps[0];
242 weak_ptr_list = NULL;
244 revertible_caf_list = NULL;
246 /* initialise the allocate() interface */
247 small_alloc_list = NULL;
249 alloc_blocks_lim = RtsFlags.GcFlags.minAllocAreaSize;
251 /* Tell GNU multi-precision pkg about our custom alloc functions */
252 mp_set_memory_functions(stgAllocForGMP, stgReallocForGMP, stgDeallocForGMP);
254 IF_DEBUG(gc, statDescribeGens());
262 stat_exit(calcAllocated());
265 /* -----------------------------------------------------------------------------
268 The entry code for every CAF does the following:
270 - builds a CAF_BLACKHOLE in the heap
271 - pushes an update frame pointing to the CAF_BLACKHOLE
272 - invokes UPD_CAF(), which:
273 - calls newCaf, below
274 - updates the CAF with a static indirection to the CAF_BLACKHOLE
276 Why do we build a BLACKHOLE in the heap rather than just updating
277 the thunk directly? It's so that we only need one kind of update
278 frame - otherwise we'd need a static version of the update frame too.
280 newCaf() does the following:
282 - it puts the CAF on the oldest generation's mut-once list.
283 This is so that we can treat the CAF as a root when collecting
286 For GHCI, we have additional requirements when dealing with CAFs:
288 - we must *retain* all dynamically-loaded CAFs ever entered,
289 just in case we need them again.
290 - we must be able to *revert* CAFs that have been evaluated, to
291 their pre-evaluated form.
293 To do this, we use an additional CAF list. When newCaf() is
294 called on a dynamically-loaded CAF, we add it to the CAF list
295 instead of the old-generation mutable list, and save away its
296 old info pointer (in caf->saved_info) for later reversion.
298 To revert all the CAFs, we traverse the CAF list and reset the
299 info pointer to caf->saved_info, then throw away the CAF list.
300 (see GC.c:revertCAFs()).
304 -------------------------------------------------------------------------- */
307 newCAF(StgClosure* caf)
314 // If we are in GHCi _and_ we are using dynamic libraries,
315 // then we can't redirect newCAF calls to newDynCAF (see below),
316 // so we make newCAF behave almost like newDynCAF.
317 // The dynamic libraries might be used by both the interpreted
318 // program and GHCi itself, so they must not be reverted.
319 // This also means that in GHCi with dynamic libraries, CAFs are not
320 // garbage collected. If this turns out to be a problem, we could
321 // do another hack here and do an address range test on caf to figure
322 // out whether it is from a dynamic library.
323 ((StgIndStatic *)caf)->saved_info = (StgInfoTable *)caf->header.info;
324 ((StgIndStatic *)caf)->static_link = caf_list;
329 /* Put this CAF on the mutable list for the old generation.
330 * This is a HACK - the IND_STATIC closure doesn't really have
331 * a mut_link field, but we pretend it has - in fact we re-use
332 * the STATIC_LINK field for the time being, because when we
333 * come to do a major GC we won't need the mut_link field
334 * any more and can use it as a STATIC_LINK.
336 ((StgIndStatic *)caf)->saved_info = NULL;
337 recordMutableGen(caf, oldest_gen);
343 /* If we are PAR or DIST then we never forget a CAF */
345 //debugBelch("<##> Globalising CAF %08x %s",caf,info_type(caf));
346 newGA=makeGlobal(caf,rtsTrue); /*given full weight*/
352 // An alternate version of newCaf which is used for dynamically loaded
353 // object code in GHCi. In this case we want to retain *all* CAFs in
354 // the object code, because they might be demanded at any time from an
355 // expression evaluated on the command line.
356 // Also, GHCi might want to revert CAFs, so we add these to the
357 // revertible_caf_list.
359 // The linker hackily arranges that references to newCaf from dynamic
360 // code end up pointing to newDynCAF.
362 newDynCAF(StgClosure *caf)
366 ((StgIndStatic *)caf)->saved_info = (StgInfoTable *)caf->header.info;
367 ((StgIndStatic *)caf)->static_link = revertible_caf_list;
368 revertible_caf_list = caf;
373 /* -----------------------------------------------------------------------------
375 -------------------------------------------------------------------------- */
378 allocNursery (step *stp, bdescr *tail, nat blocks)
383 // Allocate a nursery: we allocate fresh blocks one at a time and
384 // cons them on to the front of the list, not forgetting to update
385 // the back pointer on the tail of the list to point to the new block.
386 for (i=0; i < blocks; i++) {
389 processNursery() in LdvProfile.c assumes that every block group in
390 the nursery contains only a single block. So, if a block group is
391 given multiple blocks, change processNursery() accordingly.
395 // double-link the nursery: we might need to insert blocks
402 bd->free = bd->start;
410 assignNurseriesToCapabilities (void)
415 for (i = 0; i < n_nurseries; i++) {
416 capabilities[i].r.rNursery = &nurseries[i];
417 capabilities[i].r.rCurrentNursery = nurseries[i].blocks;
418 capabilities[i].r.rCurrentAlloc = NULL;
420 #else /* THREADED_RTS */
421 MainCapability.r.rNursery = &nurseries[0];
422 MainCapability.r.rCurrentNursery = nurseries[0].blocks;
423 MainCapability.r.rCurrentAlloc = NULL;
428 allocNurseries( void )
432 for (i = 0; i < n_nurseries; i++) {
433 nurseries[i].blocks =
434 allocNursery(&nurseries[i], NULL,
435 RtsFlags.GcFlags.minAllocAreaSize);
436 nurseries[i].n_blocks = RtsFlags.GcFlags.minAllocAreaSize;
437 nurseries[i].old_blocks = NULL;
438 nurseries[i].n_old_blocks = 0;
439 /* hp, hpLim, hp_bd, to_space etc. aren't used in the nursery */
441 assignNurseriesToCapabilities();
445 resetNurseries( void )
451 for (i = 0; i < n_nurseries; i++) {
453 for (bd = stp->blocks; bd; bd = bd->link) {
454 bd->free = bd->start;
455 ASSERT(bd->gen_no == 0);
456 ASSERT(bd->step == stp);
457 IF_DEBUG(sanity,memset(bd->start, 0xaa, BLOCK_SIZE));
460 assignNurseriesToCapabilities();
464 countNurseryBlocks (void)
469 for (i = 0; i < n_nurseries; i++) {
470 blocks += nurseries[i].n_blocks;
476 resizeNursery ( step *stp, nat blocks )
481 nursery_blocks = stp->n_blocks;
482 if (nursery_blocks == blocks) return;
484 if (nursery_blocks < blocks) {
485 IF_DEBUG(gc, debugBelch("Increasing size of nursery to %d blocks\n",
487 stp->blocks = allocNursery(stp, stp->blocks, blocks-nursery_blocks);
492 IF_DEBUG(gc, debugBelch("Decreasing size of nursery to %d blocks\n",
496 while (nursery_blocks > blocks) {
498 next_bd->u.back = NULL;
499 nursery_blocks -= bd->blocks; // might be a large block
504 // might have gone just under, by freeing a large block, so make
505 // up the difference.
506 if (nursery_blocks < blocks) {
507 stp->blocks = allocNursery(stp, stp->blocks, blocks-nursery_blocks);
511 stp->n_blocks = blocks;
512 ASSERT(countBlocks(stp->blocks) == stp->n_blocks);
516 // Resize each of the nurseries to the specified size.
519 resizeNurseriesFixed (nat blocks)
522 for (i = 0; i < n_nurseries; i++) {
523 resizeNursery(&nurseries[i], blocks);
528 // Resize the nurseries to the total specified size.
531 resizeNurseries (nat blocks)
533 // If there are multiple nurseries, then we just divide the number
534 // of available blocks between them.
535 resizeNurseriesFixed(blocks / n_nurseries);
538 /* -----------------------------------------------------------------------------
539 The allocate() interface
541 allocate(n) always succeeds, and returns a chunk of memory n words
542 long. n can be larger than the size of a block if necessary, in
543 which case a contiguous block group will be allocated.
544 -------------------------------------------------------------------------- */
554 TICK_ALLOC_HEAP_NOCTR(n);
557 /* big allocation (>LARGE_OBJECT_THRESHOLD) */
558 /* ToDo: allocate directly into generation 1 */
559 if (n >= LARGE_OBJECT_THRESHOLD/sizeof(W_)) {
560 nat req_blocks = (lnat)BLOCK_ROUND_UP(n*sizeof(W_)) / BLOCK_SIZE;
561 bd = allocGroup(req_blocks);
562 dbl_link_onto(bd, &g0s0->large_objects);
563 g0s0->n_large_blocks += req_blocks;
566 bd->flags = BF_LARGE;
567 bd->free = bd->start + n;
568 alloc_blocks += req_blocks;
572 /* small allocation (<LARGE_OBJECT_THRESHOLD) */
573 } else if (small_alloc_list == NULL || alloc_Hp + n > alloc_HpLim) {
574 if (small_alloc_list) {
575 small_alloc_list->free = alloc_Hp;
578 bd->link = small_alloc_list;
579 small_alloc_list = bd;
583 alloc_Hp = bd->start;
584 alloc_HpLim = bd->start + BLOCK_SIZE_W;
595 allocated_bytes( void )
599 allocated = alloc_blocks * BLOCK_SIZE_W - (alloc_HpLim - alloc_Hp);
600 if (pinned_object_block != NULL) {
601 allocated -= (pinned_object_block->start + BLOCK_SIZE_W) -
602 pinned_object_block->free;
609 tidyAllocateLists (void)
611 if (small_alloc_list != NULL) {
612 ASSERT(alloc_Hp >= small_alloc_list->start &&
613 alloc_Hp <= small_alloc_list->start + BLOCK_SIZE);
614 small_alloc_list->free = alloc_Hp;
618 /* -----------------------------------------------------------------------------
621 This allocates memory in the current thread - it is intended for
622 use primarily from STG-land where we have a Capability. It is
623 better than allocate() because it doesn't require taking the
624 sm_mutex lock in the common case.
626 Memory is allocated directly from the nursery if possible (but not
627 from the current nursery block, so as not to interfere with
629 -------------------------------------------------------------------------- */
632 allocateLocal (Capability *cap, nat n)
637 TICK_ALLOC_HEAP_NOCTR(n);
640 /* big allocation (>LARGE_OBJECT_THRESHOLD) */
641 /* ToDo: allocate directly into generation 1 */
642 if (n >= LARGE_OBJECT_THRESHOLD/sizeof(W_)) {
643 nat req_blocks = (lnat)BLOCK_ROUND_UP(n*sizeof(W_)) / BLOCK_SIZE;
645 bd = allocGroup(req_blocks);
646 dbl_link_onto(bd, &g0s0->large_objects);
647 g0s0->n_large_blocks += req_blocks;
650 bd->flags = BF_LARGE;
651 bd->free = bd->start + n;
652 alloc_blocks += req_blocks;
656 /* small allocation (<LARGE_OBJECT_THRESHOLD) */
659 bd = cap->r.rCurrentAlloc;
660 if (bd == NULL || bd->free + n > bd->start + BLOCK_SIZE_W) {
662 // The CurrentAlloc block is full, we need to find another
663 // one. First, we try taking the next block from the
665 bd = cap->r.rCurrentNursery->link;
667 if (bd == NULL || bd->free + n > bd->start + BLOCK_SIZE_W) {
668 // The nursery is empty, or the next block is already
669 // full: allocate a fresh block (we can't fail here).
672 cap->r.rNursery->n_blocks++;
675 bd->step = cap->r.rNursery;
678 // we have a block in the nursery: take it and put
679 // it at the *front* of the nursery list, and use it
680 // to allocate() from.
681 cap->r.rCurrentNursery->link = bd->link;
682 if (bd->link != NULL) {
683 bd->link->u.back = cap->r.rCurrentNursery;
686 dbl_link_onto(bd, &cap->r.rNursery->blocks);
687 cap->r.rCurrentAlloc = bd;
688 IF_DEBUG(sanity, checkNurserySanity(cap->r.rNursery));
696 /* ---------------------------------------------------------------------------
697 Allocate a fixed/pinned object.
699 We allocate small pinned objects into a single block, allocating a
700 new block when the current one overflows. The block is chained
701 onto the large_object_list of generation 0 step 0.
703 NOTE: The GC can't in general handle pinned objects. This
704 interface is only safe to use for ByteArrays, which have no
705 pointers and don't require scavenging. It works because the
706 block's descriptor has the BF_LARGE flag set, so the block is
707 treated as a large object and chained onto various lists, rather
708 than the individual objects being copied. However, when it comes
709 to scavenge the block, the GC will only scavenge the first object.
710 The reason is that the GC can't linearly scan a block of pinned
711 objects at the moment (doing so would require using the
712 mostly-copying techniques). But since we're restricting ourselves
713 to pinned ByteArrays, not scavenging is ok.
715 This function is called by newPinnedByteArray# which immediately
716 fills the allocated memory with a MutableByteArray#.
717 ------------------------------------------------------------------------- */
720 allocatePinned( nat n )
723 bdescr *bd = pinned_object_block;
725 // If the request is for a large object, then allocate()
726 // will give us a pinned object anyway.
727 if (n >= LARGE_OBJECT_THRESHOLD/sizeof(W_)) {
733 TICK_ALLOC_HEAP_NOCTR(n);
736 // we always return 8-byte aligned memory. bd->free must be
737 // 8-byte aligned to begin with, so we just round up n to
738 // the nearest multiple of 8 bytes.
739 if (sizeof(StgWord) == 4) {
743 // If we don't have a block of pinned objects yet, or the current
744 // one isn't large enough to hold the new object, allocate a new one.
745 if (bd == NULL || (bd->free + n) > (bd->start + BLOCK_SIZE_W)) {
746 pinned_object_block = bd = allocBlock();
747 dbl_link_onto(bd, &g0s0->large_objects);
750 bd->flags = BF_PINNED | BF_LARGE;
751 bd->free = bd->start;
761 /* -----------------------------------------------------------------------------
762 This is the write barrier for MUT_VARs, a.k.a. IORefs. A
763 MUT_VAR_CLEAN object is not on the mutable list; a MUT_VAR_DIRTY
764 is. When written to, a MUT_VAR_CLEAN turns into a MUT_VAR_DIRTY
765 and is put on the mutable list.
766 -------------------------------------------------------------------------- */
769 dirty_MUT_VAR(StgRegTable *reg, StgClosure *p)
771 Capability *cap = regTableToCapability(reg);
773 if (p->header.info == &stg_MUT_VAR_CLEAN_info) {
774 p->header.info = &stg_MUT_VAR_DIRTY_info;
776 if (bd->gen_no > 0) recordMutableCap(p,cap,bd->gen_no);
780 /* -----------------------------------------------------------------------------
781 Allocation functions for GMP.
783 These all use the allocate() interface - we can't have any garbage
784 collection going on during a gmp operation, so we use allocate()
785 which always succeeds. The gmp operations which might need to
786 allocate will ask the storage manager (via doYouWantToGC()) whether
787 a garbage collection is required, in case we get into a loop doing
788 only allocate() style allocation.
789 -------------------------------------------------------------------------- */
792 stgAllocForGMP (size_t size_in_bytes)
795 nat data_size_in_words, total_size_in_words;
797 /* round up to a whole number of words */
798 data_size_in_words = (size_in_bytes + sizeof(W_) + 1) / sizeof(W_);
799 total_size_in_words = sizeofW(StgArrWords) + data_size_in_words;
801 /* allocate and fill it in. */
802 #if defined(THREADED_RTS)
803 arr = (StgArrWords *)allocateLocal(myTask()->cap, total_size_in_words);
805 arr = (StgArrWords *)allocateLocal(&MainCapability, total_size_in_words);
807 SET_ARR_HDR(arr, &stg_ARR_WORDS_info, CCCS, data_size_in_words);
809 /* and return a ptr to the goods inside the array */
814 stgReallocForGMP (void *ptr, size_t old_size, size_t new_size)
816 void *new_stuff_ptr = stgAllocForGMP(new_size);
818 char *p = (char *) ptr;
819 char *q = (char *) new_stuff_ptr;
821 for (; i < old_size; i++, p++, q++) {
825 return(new_stuff_ptr);
829 stgDeallocForGMP (void *ptr STG_UNUSED,
830 size_t size STG_UNUSED)
832 /* easy for us: the garbage collector does the dealloc'n */
835 /* -----------------------------------------------------------------------------
837 * -------------------------------------------------------------------------- */
839 /* -----------------------------------------------------------------------------
842 * Approximate how much we've allocated: number of blocks in the
843 * nursery + blocks allocated via allocate() - unused nusery blocks.
844 * This leaves a little slop at the end of each block, and doesn't
845 * take into account large objects (ToDo).
846 * -------------------------------------------------------------------------- */
849 calcAllocated( void )
854 allocated = allocated_bytes();
855 allocated += countNurseryBlocks() * BLOCK_SIZE_W;
860 for (i = 0; i < n_nurseries; i++) {
862 for ( bd = capabilities[i].r.rCurrentNursery->link;
863 bd != NULL; bd = bd->link ) {
864 allocated -= BLOCK_SIZE_W;
866 cap = &capabilities[i];
867 if (cap->r.rCurrentNursery->free <
868 cap->r.rCurrentNursery->start + BLOCK_SIZE_W) {
869 allocated -= (cap->r.rCurrentNursery->start + BLOCK_SIZE_W)
870 - cap->r.rCurrentNursery->free;
874 bdescr *current_nursery = MainCapability.r.rCurrentNursery;
876 for ( bd = current_nursery->link; bd != NULL; bd = bd->link ) {
877 allocated -= BLOCK_SIZE_W;
879 if (current_nursery->free < current_nursery->start + BLOCK_SIZE_W) {
880 allocated -= (current_nursery->start + BLOCK_SIZE_W)
881 - current_nursery->free;
886 total_allocated += allocated;
890 /* Approximate the amount of live data in the heap. To be called just
891 * after garbage collection (see GarbageCollect()).
900 if (RtsFlags.GcFlags.generations == 1) {
901 live = (g0s0->n_blocks - 1) * BLOCK_SIZE_W +
902 ((lnat)g0s0->hp_bd->free - (lnat)g0s0->hp_bd->start) / sizeof(W_);
906 for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
907 for (s = 0; s < generations[g].n_steps; s++) {
908 /* approximate amount of live data (doesn't take into account slop
909 * at end of each block).
911 if (g == 0 && s == 0) {
914 stp = &generations[g].steps[s];
915 live += (stp->n_large_blocks + stp->n_blocks - 1) * BLOCK_SIZE_W;
916 if (stp->hp_bd != NULL) {
917 live += ((lnat)stp->hp_bd->free - (lnat)stp->hp_bd->start)
920 if (stp->scavd_hp != NULL) {
921 live -= (P_)(BLOCK_ROUND_UP(stp->scavd_hp)) - stp->scavd_hp;
928 /* Approximate the number of blocks that will be needed at the next
929 * garbage collection.
931 * Assume: all data currently live will remain live. Steps that will
932 * be collected next time will therefore need twice as many blocks
933 * since all the data will be copied.
942 for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
943 for (s = 0; s < generations[g].n_steps; s++) {
944 if (g == 0 && s == 0) { continue; }
945 stp = &generations[g].steps[s];
946 if (generations[g].steps[0].n_blocks +
947 generations[g].steps[0].n_large_blocks
948 > generations[g].max_blocks
949 && stp->is_compacted == 0) {
950 needed += 2 * stp->n_blocks;
952 needed += stp->n_blocks;
959 /* -----------------------------------------------------------------------------
962 memInventory() checks for memory leaks by counting up all the
963 blocks we know about and comparing that to the number of blocks
964 allegedly floating around in the system.
965 -------------------------------------------------------------------------- */
970 stepBlocks (step *stp)
975 total_blocks = stp->n_blocks;
976 total_blocks += stp->n_old_blocks;
977 for (bd = stp->large_objects; bd; bd = bd->link) {
978 total_blocks += bd->blocks;
979 /* hack for megablock groups: they have an extra block or two in
980 the second and subsequent megablocks where the block
981 descriptors would normally go.
983 if (bd->blocks > BLOCKS_PER_MBLOCK) {
984 total_blocks -= (MBLOCK_SIZE / BLOCK_SIZE - BLOCKS_PER_MBLOCK)
985 * (bd->blocks/(MBLOCK_SIZE/BLOCK_SIZE));
997 lnat total_blocks = 0, free_blocks = 0;
999 /* count the blocks we current have */
1001 for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
1002 for (i = 0; i < n_capabilities; i++) {
1003 for (bd = capabilities[i].mut_lists[g]; bd != NULL; bd = bd->link) {
1004 total_blocks += bd->blocks;
1007 for (bd = generations[g].mut_list; bd != NULL; bd = bd->link) {
1008 total_blocks += bd->blocks;
1010 for (s = 0; s < generations[g].n_steps; s++) {
1011 if (g==0 && s==0) continue;
1012 stp = &generations[g].steps[s];
1013 total_blocks += stepBlocks(stp);
1017 for (i = 0; i < n_nurseries; i++) {
1018 total_blocks += stepBlocks(&nurseries[i]);
1021 // We put pinned object blocks in g0s0, so better count blocks there too.
1022 total_blocks += stepBlocks(g0s0);
1025 /* any blocks held by allocate() */
1026 for (bd = small_alloc_list; bd; bd = bd->link) {
1027 total_blocks += bd->blocks;
1031 if (RtsFlags.ProfFlags.doHeapProfile == HEAP_BY_RETAINER) {
1032 total_blocks += retainerStackBlocks();
1036 // count the blocks allocated by the arena allocator
1037 total_blocks += arenaBlocks();
1039 /* count the blocks on the free list */
1040 free_blocks = countFreeList();
1042 if (total_blocks + free_blocks != mblocks_allocated *
1043 BLOCKS_PER_MBLOCK) {
1044 debugBelch("Blocks: %ld live + %ld free = %ld total (%ld around)\n",
1045 total_blocks, free_blocks, total_blocks + free_blocks,
1046 mblocks_allocated * BLOCKS_PER_MBLOCK);
1049 ASSERT(total_blocks + free_blocks == mblocks_allocated * BLOCKS_PER_MBLOCK);
1054 countBlocks(bdescr *bd)
1057 for (n=0; bd != NULL; bd=bd->link) {
1063 /* Full heap sanity check. */
1069 if (RtsFlags.GcFlags.generations == 1) {
1070 checkHeap(g0s0->blocks);
1071 checkChain(g0s0->large_objects);
1074 for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
1075 for (s = 0; s < generations[g].n_steps; s++) {
1076 if (g == 0 && s == 0) { continue; }
1077 ASSERT(countBlocks(generations[g].steps[s].blocks)
1078 == generations[g].steps[s].n_blocks);
1079 ASSERT(countBlocks(generations[g].steps[s].large_objects)
1080 == generations[g].steps[s].n_large_blocks);
1081 checkHeap(generations[g].steps[s].blocks);
1082 checkChain(generations[g].steps[s].large_objects);
1084 checkMutableList(generations[g].mut_list, g);
1089 for (s = 0; s < n_nurseries; s++) {
1090 ASSERT(countBlocks(nurseries[s].blocks)
1091 == nurseries[s].n_blocks);
1092 ASSERT(countBlocks(nurseries[s].large_objects)
1093 == nurseries[s].n_large_blocks);
1096 checkFreeListSanity();
1100 /* Nursery sanity check */
1102 checkNurserySanity( step *stp )
1108 for (bd = stp->blocks; bd != NULL; bd = bd->link) {
1109 ASSERT(bd->u.back == prev);
1111 blocks += bd->blocks;
1113 ASSERT(blocks == stp->n_blocks);
1116 // handy function for use in gdb, because Bdescr() is inlined.
1117 extern bdescr *_bdescr( StgPtr p );