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)
32 * All these globals require sm_mutex to access in THREADED_RTS mode.
34 StgClosure *caf_list = NULL;
35 StgClosure *revertible_caf_list = NULL;
38 bdescr *small_alloc_list; /* allocate()d small objects */
39 bdescr *pinned_object_block; /* allocate pinned objects into this block */
40 nat alloc_blocks; /* number of allocate()d blocks since GC */
41 nat alloc_blocks_lim; /* approximate limit on alloc_blocks */
43 StgPtr alloc_Hp = NULL; /* next free byte in small_alloc_list */
44 StgPtr alloc_HpLim = NULL; /* end of block at small_alloc_list */
46 generation *generations = NULL; /* all the generations */
47 generation *g0 = NULL; /* generation 0, for convenience */
48 generation *oldest_gen = NULL; /* oldest generation, for convenience */
49 step *g0s0 = NULL; /* generation 0, step 0, for convenience */
51 ullong total_allocated = 0; /* total memory allocated during run */
53 nat n_nurseries = 0; /* == RtsFlags.ParFlags.nNodes, convenience */
54 step *nurseries = NULL; /* array of nurseries, >1 only if THREADED_RTS */
58 * Storage manager mutex: protects all the above state from
59 * simultaneous access by two STG threads.
63 * This mutex is used by atomicModifyMutVar# only
65 Mutex atomic_modify_mutvar_mutex;
72 static void *stgAllocForGMP (size_t size_in_bytes);
73 static void *stgReallocForGMP (void *ptr, size_t old_size, size_t new_size);
74 static void stgDeallocForGMP (void *ptr, size_t size);
77 initStep (step *stp, int g, int s)
82 stp->old_blocks = NULL;
83 stp->n_old_blocks = 0;
84 stp->gen = &generations[g];
90 stp->scavd_hpLim = NULL;
93 stp->large_objects = NULL;
94 stp->n_large_blocks = 0;
95 stp->new_large_objects = NULL;
96 stp->scavenged_large_objects = NULL;
97 stp->n_scavenged_large_blocks = 0;
98 stp->is_compacted = 0;
108 if (generations != NULL) {
109 // multi-init protection
113 /* Sanity check to make sure the LOOKS_LIKE_ macros appear to be
114 * doing something reasonable.
116 ASSERT(LOOKS_LIKE_INFO_PTR(&stg_BLACKHOLE_info));
117 ASSERT(LOOKS_LIKE_CLOSURE_PTR(&stg_dummy_ret_closure));
118 ASSERT(!HEAP_ALLOCED(&stg_dummy_ret_closure));
120 if (RtsFlags.GcFlags.maxHeapSize != 0 &&
121 RtsFlags.GcFlags.heapSizeSuggestion >
122 RtsFlags.GcFlags.maxHeapSize) {
123 RtsFlags.GcFlags.maxHeapSize = RtsFlags.GcFlags.heapSizeSuggestion;
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;
133 initBlockAllocator();
135 #if defined(THREADED_RTS)
136 initMutex(&sm_mutex);
137 initMutex(&atomic_modify_mutvar_mutex);
142 /* allocate generation info array */
143 generations = (generation *)stgMallocBytes(RtsFlags.GcFlags.generations
144 * sizeof(struct generation_),
145 "initStorage: gens");
147 /* Initialise all generations */
148 for(g = 0; g < RtsFlags.GcFlags.generations; g++) {
149 gen = &generations[g];
151 gen->mut_list = allocBlock();
152 gen->collections = 0;
153 gen->failed_promotions = 0;
157 /* A couple of convenience pointers */
158 g0 = &generations[0];
159 oldest_gen = &generations[RtsFlags.GcFlags.generations-1];
161 /* Allocate step structures in each generation */
162 if (RtsFlags.GcFlags.generations > 1) {
163 /* Only for multiple-generations */
165 /* Oldest generation: one step */
166 oldest_gen->n_steps = 1;
168 stgMallocBytes(1 * sizeof(struct step_), "initStorage: last step");
170 /* set up all except the oldest generation with 2 steps */
171 for(g = 0; g < RtsFlags.GcFlags.generations-1; g++) {
172 generations[g].n_steps = RtsFlags.GcFlags.steps;
173 generations[g].steps =
174 stgMallocBytes (RtsFlags.GcFlags.steps * sizeof(struct step_),
175 "initStorage: steps");
179 /* single generation, i.e. a two-space collector */
181 g0->steps = stgMallocBytes (sizeof(struct step_), "initStorage: steps");
185 n_nurseries = n_capabilities;
186 nurseries = stgMallocBytes (n_nurseries * sizeof(struct step_),
187 "initStorage: nurseries");
190 nurseries = g0->steps; // just share nurseries[0] with g0s0
193 /* Initialise all steps */
194 for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
195 for (s = 0; s < generations[g].n_steps; s++) {
196 initStep(&generations[g].steps[s], g, s);
201 for (s = 0; s < n_nurseries; s++) {
202 initStep(&nurseries[s], 0, s);
206 /* Set up the destination pointers in each younger gen. step */
207 for (g = 0; g < RtsFlags.GcFlags.generations-1; g++) {
208 for (s = 0; s < generations[g].n_steps-1; s++) {
209 generations[g].steps[s].to = &generations[g].steps[s+1];
211 generations[g].steps[s].to = &generations[g+1].steps[0];
213 oldest_gen->steps[0].to = &oldest_gen->steps[0];
216 for (s = 0; s < n_nurseries; s++) {
217 nurseries[s].to = generations[0].steps[0].to;
221 /* The oldest generation has one step. */
222 if (RtsFlags.GcFlags.compact) {
223 if (RtsFlags.GcFlags.generations == 1) {
224 errorBelch("WARNING: compaction is incompatible with -G1; disabled");
226 oldest_gen->steps[0].is_compacted = 1;
231 if (RtsFlags.GcFlags.generations == 1) {
232 errorBelch("-G1 is incompatible with -threaded");
233 stg_exit(EXIT_FAILURE);
237 /* generation 0 is special: that's the nursery */
238 generations[0].max_blocks = 0;
240 /* G0S0: the allocation area. Policy: keep the allocation area
241 * small to begin with, even if we have a large suggested heap
242 * size. Reason: we're going to do a major collection first, and we
243 * don't want it to be a big one. This vague idea is borne out by
244 * rigorous experimental evidence.
246 g0s0 = &generations[0].steps[0];
250 weak_ptr_list = NULL;
252 revertible_caf_list = NULL;
254 /* initialise the allocate() interface */
255 small_alloc_list = NULL;
257 alloc_blocks_lim = RtsFlags.GcFlags.minAllocAreaSize;
259 /* Tell GNU multi-precision pkg about our custom alloc functions */
260 mp_set_memory_functions(stgAllocForGMP, stgReallocForGMP, stgDeallocForGMP);
262 IF_DEBUG(gc, statDescribeGens());
270 stat_exit(calcAllocated());
278 for(g = 0; g < RtsFlags.GcFlags.generations; g++)
279 stgFree(generations[g].steps);
280 stgFree(generations);
284 /* -----------------------------------------------------------------------------
287 The entry code for every CAF does the following:
289 - builds a CAF_BLACKHOLE in the heap
290 - pushes an update frame pointing to the CAF_BLACKHOLE
291 - invokes UPD_CAF(), which:
292 - calls newCaf, below
293 - updates the CAF with a static indirection to the CAF_BLACKHOLE
295 Why do we build a BLACKHOLE in the heap rather than just updating
296 the thunk directly? It's so that we only need one kind of update
297 frame - otherwise we'd need a static version of the update frame too.
299 newCaf() does the following:
301 - it puts the CAF on the oldest generation's mut-once list.
302 This is so that we can treat the CAF as a root when collecting
305 For GHCI, we have additional requirements when dealing with CAFs:
307 - we must *retain* all dynamically-loaded CAFs ever entered,
308 just in case we need them again.
309 - we must be able to *revert* CAFs that have been evaluated, to
310 their pre-evaluated form.
312 To do this, we use an additional CAF list. When newCaf() is
313 called on a dynamically-loaded CAF, we add it to the CAF list
314 instead of the old-generation mutable list, and save away its
315 old info pointer (in caf->saved_info) for later reversion.
317 To revert all the CAFs, we traverse the CAF list and reset the
318 info pointer to caf->saved_info, then throw away the CAF list.
319 (see GC.c:revertCAFs()).
323 -------------------------------------------------------------------------- */
326 newCAF(StgClosure* caf)
333 // If we are in GHCi _and_ we are using dynamic libraries,
334 // then we can't redirect newCAF calls to newDynCAF (see below),
335 // so we make newCAF behave almost like newDynCAF.
336 // The dynamic libraries might be used by both the interpreted
337 // program and GHCi itself, so they must not be reverted.
338 // This also means that in GHCi with dynamic libraries, CAFs are not
339 // garbage collected. If this turns out to be a problem, we could
340 // do another hack here and do an address range test on caf to figure
341 // out whether it is from a dynamic library.
342 ((StgIndStatic *)caf)->saved_info = (StgInfoTable *)caf->header.info;
343 ((StgIndStatic *)caf)->static_link = caf_list;
348 /* Put this CAF on the mutable list for the old generation.
349 * This is a HACK - the IND_STATIC closure doesn't really have
350 * a mut_link field, but we pretend it has - in fact we re-use
351 * the STATIC_LINK field for the time being, because when we
352 * come to do a major GC we won't need the mut_link field
353 * any more and can use it as a STATIC_LINK.
355 ((StgIndStatic *)caf)->saved_info = NULL;
356 recordMutableGen(caf, oldest_gen);
362 /* If we are PAR or DIST then we never forget a CAF */
364 //debugBelch("<##> Globalising CAF %08x %s",caf,info_type(caf));
365 newGA=makeGlobal(caf,rtsTrue); /*given full weight*/
371 // An alternate version of newCaf which is used for dynamically loaded
372 // object code in GHCi. In this case we want to retain *all* CAFs in
373 // the object code, because they might be demanded at any time from an
374 // expression evaluated on the command line.
375 // Also, GHCi might want to revert CAFs, so we add these to the
376 // revertible_caf_list.
378 // The linker hackily arranges that references to newCaf from dynamic
379 // code end up pointing to newDynCAF.
381 newDynCAF(StgClosure *caf)
385 ((StgIndStatic *)caf)->saved_info = (StgInfoTable *)caf->header.info;
386 ((StgIndStatic *)caf)->static_link = revertible_caf_list;
387 revertible_caf_list = caf;
392 /* -----------------------------------------------------------------------------
394 -------------------------------------------------------------------------- */
397 allocNursery (step *stp, bdescr *tail, nat blocks)
402 // Allocate a nursery: we allocate fresh blocks one at a time and
403 // cons them on to the front of the list, not forgetting to update
404 // the back pointer on the tail of the list to point to the new block.
405 for (i=0; i < blocks; i++) {
408 processNursery() in LdvProfile.c assumes that every block group in
409 the nursery contains only a single block. So, if a block group is
410 given multiple blocks, change processNursery() accordingly.
414 // double-link the nursery: we might need to insert blocks
421 bd->free = bd->start;
429 assignNurseriesToCapabilities (void)
434 for (i = 0; i < n_nurseries; i++) {
435 capabilities[i].r.rNursery = &nurseries[i];
436 capabilities[i].r.rCurrentNursery = nurseries[i].blocks;
437 capabilities[i].r.rCurrentAlloc = NULL;
439 #else /* THREADED_RTS */
440 MainCapability.r.rNursery = &nurseries[0];
441 MainCapability.r.rCurrentNursery = nurseries[0].blocks;
442 MainCapability.r.rCurrentAlloc = NULL;
447 allocNurseries( void )
451 for (i = 0; i < n_nurseries; i++) {
452 nurseries[i].blocks =
453 allocNursery(&nurseries[i], NULL,
454 RtsFlags.GcFlags.minAllocAreaSize);
455 nurseries[i].n_blocks = RtsFlags.GcFlags.minAllocAreaSize;
456 nurseries[i].old_blocks = NULL;
457 nurseries[i].n_old_blocks = 0;
458 /* hp, hpLim, hp_bd, to_space etc. aren't used in the nursery */
460 assignNurseriesToCapabilities();
464 resetNurseries( void )
470 for (i = 0; i < n_nurseries; i++) {
472 for (bd = stp->blocks; bd; bd = bd->link) {
473 bd->free = bd->start;
474 ASSERT(bd->gen_no == 0);
475 ASSERT(bd->step == stp);
476 IF_DEBUG(sanity,memset(bd->start, 0xaa, BLOCK_SIZE));
479 assignNurseriesToCapabilities();
483 countNurseryBlocks (void)
488 for (i = 0; i < n_nurseries; i++) {
489 blocks += nurseries[i].n_blocks;
495 resizeNursery ( step *stp, nat blocks )
500 nursery_blocks = stp->n_blocks;
501 if (nursery_blocks == blocks) return;
503 if (nursery_blocks < blocks) {
504 debugTrace(DEBUG_gc, "increasing size of nursery to %d blocks",
506 stp->blocks = allocNursery(stp, stp->blocks, blocks-nursery_blocks);
511 debugTrace(DEBUG_gc, "decreasing size of nursery to %d blocks",
515 while (nursery_blocks > blocks) {
517 next_bd->u.back = NULL;
518 nursery_blocks -= bd->blocks; // might be a large block
523 // might have gone just under, by freeing a large block, so make
524 // up the difference.
525 if (nursery_blocks < blocks) {
526 stp->blocks = allocNursery(stp, stp->blocks, blocks-nursery_blocks);
530 stp->n_blocks = blocks;
531 ASSERT(countBlocks(stp->blocks) == stp->n_blocks);
535 // Resize each of the nurseries to the specified size.
538 resizeNurseriesFixed (nat blocks)
541 for (i = 0; i < n_nurseries; i++) {
542 resizeNursery(&nurseries[i], blocks);
547 // Resize the nurseries to the total specified size.
550 resizeNurseries (nat blocks)
552 // If there are multiple nurseries, then we just divide the number
553 // of available blocks between them.
554 resizeNurseriesFixed(blocks / n_nurseries);
557 /* -----------------------------------------------------------------------------
558 The allocate() interface
560 allocate(n) always succeeds, and returns a chunk of memory n words
561 long. n can be larger than the size of a block if necessary, in
562 which case a contiguous block group will be allocated.
563 -------------------------------------------------------------------------- */
573 TICK_ALLOC_HEAP_NOCTR(n);
576 /* big allocation (>LARGE_OBJECT_THRESHOLD) */
577 /* ToDo: allocate directly into generation 1 */
578 if (n >= LARGE_OBJECT_THRESHOLD/sizeof(W_)) {
579 nat req_blocks = (lnat)BLOCK_ROUND_UP(n*sizeof(W_)) / BLOCK_SIZE;
580 bd = allocGroup(req_blocks);
581 dbl_link_onto(bd, &g0s0->large_objects);
582 g0s0->n_large_blocks += req_blocks;
585 bd->flags = BF_LARGE;
586 bd->free = bd->start + n;
587 alloc_blocks += req_blocks;
591 /* small allocation (<LARGE_OBJECT_THRESHOLD) */
592 } else if (small_alloc_list == NULL || alloc_Hp + n > alloc_HpLim) {
593 if (small_alloc_list) {
594 small_alloc_list->free = alloc_Hp;
597 bd->link = small_alloc_list;
598 small_alloc_list = bd;
602 alloc_Hp = bd->start;
603 alloc_HpLim = bd->start + BLOCK_SIZE_W;
614 allocated_bytes( void )
618 allocated = alloc_blocks * BLOCK_SIZE_W - (alloc_HpLim - alloc_Hp);
619 if (pinned_object_block != NULL) {
620 allocated -= (pinned_object_block->start + BLOCK_SIZE_W) -
621 pinned_object_block->free;
628 tidyAllocateLists (void)
630 if (small_alloc_list != NULL) {
631 ASSERT(alloc_Hp >= small_alloc_list->start &&
632 alloc_Hp <= small_alloc_list->start + BLOCK_SIZE);
633 small_alloc_list->free = alloc_Hp;
637 /* -----------------------------------------------------------------------------
640 This allocates memory in the current thread - it is intended for
641 use primarily from STG-land where we have a Capability. It is
642 better than allocate() because it doesn't require taking the
643 sm_mutex lock in the common case.
645 Memory is allocated directly from the nursery if possible (but not
646 from the current nursery block, so as not to interfere with
648 -------------------------------------------------------------------------- */
651 allocateLocal (Capability *cap, nat n)
656 TICK_ALLOC_HEAP_NOCTR(n);
659 /* big allocation (>LARGE_OBJECT_THRESHOLD) */
660 /* ToDo: allocate directly into generation 1 */
661 if (n >= LARGE_OBJECT_THRESHOLD/sizeof(W_)) {
662 nat req_blocks = (lnat)BLOCK_ROUND_UP(n*sizeof(W_)) / BLOCK_SIZE;
664 bd = allocGroup(req_blocks);
665 dbl_link_onto(bd, &g0s0->large_objects);
666 g0s0->n_large_blocks += req_blocks;
669 bd->flags = BF_LARGE;
670 bd->free = bd->start + n;
671 alloc_blocks += req_blocks;
675 /* small allocation (<LARGE_OBJECT_THRESHOLD) */
678 bd = cap->r.rCurrentAlloc;
679 if (bd == NULL || bd->free + n > bd->start + BLOCK_SIZE_W) {
681 // The CurrentAlloc block is full, we need to find another
682 // one. First, we try taking the next block from the
684 bd = cap->r.rCurrentNursery->link;
686 if (bd == NULL || bd->free + n > bd->start + BLOCK_SIZE_W) {
687 // The nursery is empty, or the next block is already
688 // full: allocate a fresh block (we can't fail here).
691 cap->r.rNursery->n_blocks++;
694 bd->step = cap->r.rNursery;
697 // we have a block in the nursery: take it and put
698 // it at the *front* of the nursery list, and use it
699 // to allocate() from.
700 cap->r.rCurrentNursery->link = bd->link;
701 if (bd->link != NULL) {
702 bd->link->u.back = cap->r.rCurrentNursery;
705 dbl_link_onto(bd, &cap->r.rNursery->blocks);
706 cap->r.rCurrentAlloc = bd;
707 IF_DEBUG(sanity, checkNurserySanity(cap->r.rNursery));
715 /* ---------------------------------------------------------------------------
716 Allocate a fixed/pinned object.
718 We allocate small pinned objects into a single block, allocating a
719 new block when the current one overflows. The block is chained
720 onto the large_object_list of generation 0 step 0.
722 NOTE: The GC can't in general handle pinned objects. This
723 interface is only safe to use for ByteArrays, which have no
724 pointers and don't require scavenging. It works because the
725 block's descriptor has the BF_LARGE flag set, so the block is
726 treated as a large object and chained onto various lists, rather
727 than the individual objects being copied. However, when it comes
728 to scavenge the block, the GC will only scavenge the first object.
729 The reason is that the GC can't linearly scan a block of pinned
730 objects at the moment (doing so would require using the
731 mostly-copying techniques). But since we're restricting ourselves
732 to pinned ByteArrays, not scavenging is ok.
734 This function is called by newPinnedByteArray# which immediately
735 fills the allocated memory with a MutableByteArray#.
736 ------------------------------------------------------------------------- */
739 allocatePinned( nat n )
742 bdescr *bd = pinned_object_block;
744 // If the request is for a large object, then allocate()
745 // will give us a pinned object anyway.
746 if (n >= LARGE_OBJECT_THRESHOLD/sizeof(W_)) {
752 TICK_ALLOC_HEAP_NOCTR(n);
755 // we always return 8-byte aligned memory. bd->free must be
756 // 8-byte aligned to begin with, so we just round up n to
757 // the nearest multiple of 8 bytes.
758 if (sizeof(StgWord) == 4) {
762 // If we don't have a block of pinned objects yet, or the current
763 // one isn't large enough to hold the new object, allocate a new one.
764 if (bd == NULL || (bd->free + n) > (bd->start + BLOCK_SIZE_W)) {
765 pinned_object_block = bd = allocBlock();
766 dbl_link_onto(bd, &g0s0->large_objects);
769 bd->flags = BF_PINNED | BF_LARGE;
770 bd->free = bd->start;
780 /* -----------------------------------------------------------------------------
781 This is the write barrier for MUT_VARs, a.k.a. IORefs. A
782 MUT_VAR_CLEAN object is not on the mutable list; a MUT_VAR_DIRTY
783 is. When written to, a MUT_VAR_CLEAN turns into a MUT_VAR_DIRTY
784 and is put on the mutable list.
785 -------------------------------------------------------------------------- */
788 dirty_MUT_VAR(StgRegTable *reg, StgClosure *p)
790 Capability *cap = regTableToCapability(reg);
792 if (p->header.info == &stg_MUT_VAR_CLEAN_info) {
793 p->header.info = &stg_MUT_VAR_DIRTY_info;
794 bd = Bdescr((StgPtr)p);
795 if (bd->gen_no > 0) recordMutableCap(p,cap,bd->gen_no);
799 /* -----------------------------------------------------------------------------
800 Allocation functions for GMP.
802 These all use the allocate() interface - we can't have any garbage
803 collection going on during a gmp operation, so we use allocate()
804 which always succeeds. The gmp operations which might need to
805 allocate will ask the storage manager (via doYouWantToGC()) whether
806 a garbage collection is required, in case we get into a loop doing
807 only allocate() style allocation.
808 -------------------------------------------------------------------------- */
811 stgAllocForGMP (size_t size_in_bytes)
814 nat data_size_in_words, total_size_in_words;
816 /* round up to a whole number of words */
817 data_size_in_words = (size_in_bytes + sizeof(W_) + 1) / sizeof(W_);
818 total_size_in_words = sizeofW(StgArrWords) + data_size_in_words;
820 /* allocate and fill it in. */
821 #if defined(THREADED_RTS)
822 arr = (StgArrWords *)allocateLocal(myTask()->cap, total_size_in_words);
824 arr = (StgArrWords *)allocateLocal(&MainCapability, total_size_in_words);
826 SET_ARR_HDR(arr, &stg_ARR_WORDS_info, CCCS, data_size_in_words);
828 /* and return a ptr to the goods inside the array */
833 stgReallocForGMP (void *ptr, size_t old_size, size_t new_size)
835 void *new_stuff_ptr = stgAllocForGMP(new_size);
837 char *p = (char *) ptr;
838 char *q = (char *) new_stuff_ptr;
840 for (; i < old_size; i++, p++, q++) {
844 return(new_stuff_ptr);
848 stgDeallocForGMP (void *ptr STG_UNUSED,
849 size_t size STG_UNUSED)
851 /* easy for us: the garbage collector does the dealloc'n */
854 /* -----------------------------------------------------------------------------
856 * -------------------------------------------------------------------------- */
858 /* -----------------------------------------------------------------------------
861 * Approximate how much we've allocated: number of blocks in the
862 * nursery + blocks allocated via allocate() - unused nusery blocks.
863 * This leaves a little slop at the end of each block, and doesn't
864 * take into account large objects (ToDo).
865 * -------------------------------------------------------------------------- */
868 calcAllocated( void )
873 allocated = allocated_bytes();
874 allocated += countNurseryBlocks() * BLOCK_SIZE_W;
879 for (i = 0; i < n_nurseries; i++) {
881 for ( bd = capabilities[i].r.rCurrentNursery->link;
882 bd != NULL; bd = bd->link ) {
883 allocated -= BLOCK_SIZE_W;
885 cap = &capabilities[i];
886 if (cap->r.rCurrentNursery->free <
887 cap->r.rCurrentNursery->start + BLOCK_SIZE_W) {
888 allocated -= (cap->r.rCurrentNursery->start + BLOCK_SIZE_W)
889 - cap->r.rCurrentNursery->free;
893 bdescr *current_nursery = MainCapability.r.rCurrentNursery;
895 for ( bd = current_nursery->link; bd != NULL; bd = bd->link ) {
896 allocated -= BLOCK_SIZE_W;
898 if (current_nursery->free < current_nursery->start + BLOCK_SIZE_W) {
899 allocated -= (current_nursery->start + BLOCK_SIZE_W)
900 - current_nursery->free;
905 total_allocated += allocated;
909 /* Approximate the amount of live data in the heap. To be called just
910 * after garbage collection (see GarbageCollect()).
919 if (RtsFlags.GcFlags.generations == 1) {
920 live = (g0s0->n_blocks - 1) * BLOCK_SIZE_W +
921 ((lnat)g0s0->hp_bd->free - (lnat)g0s0->hp_bd->start) / sizeof(W_);
925 for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
926 for (s = 0; s < generations[g].n_steps; s++) {
927 /* approximate amount of live data (doesn't take into account slop
928 * at end of each block).
930 if (g == 0 && s == 0) {
933 stp = &generations[g].steps[s];
934 live += (stp->n_large_blocks + stp->n_blocks - 1) * BLOCK_SIZE_W;
935 if (stp->hp_bd != NULL) {
936 live += ((lnat)stp->hp_bd->free - (lnat)stp->hp_bd->start)
939 if (stp->scavd_hp != NULL) {
940 live -= (P_)(BLOCK_ROUND_UP(stp->scavd_hp)) - stp->scavd_hp;
947 /* Approximate the number of blocks that will be needed at the next
948 * garbage collection.
950 * Assume: all data currently live will remain live. Steps that will
951 * be collected next time will therefore need twice as many blocks
952 * since all the data will be copied.
961 for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
962 for (s = 0; s < generations[g].n_steps; s++) {
963 if (g == 0 && s == 0) { continue; }
964 stp = &generations[g].steps[s];
965 if (generations[g].steps[0].n_blocks +
966 generations[g].steps[0].n_large_blocks
967 > generations[g].max_blocks
968 && stp->is_compacted == 0) {
969 needed += 2 * stp->n_blocks;
971 needed += stp->n_blocks;
978 /* ----------------------------------------------------------------------------
981 Executable memory must be managed separately from non-executable
982 memory. Most OSs these days require you to jump through hoops to
983 dynamically allocate executable memory, due to various security
986 Here we provide a small memory allocator for executable memory.
987 Memory is managed with a page granularity; we allocate linearly
988 in the page, and when the page is emptied (all objects on the page
989 are free) we free the page again, not forgetting to make it
991 ------------------------------------------------------------------------- */
993 static bdescr *exec_block;
995 void *allocateExec (nat bytes)
1002 // round up to words.
1003 n = (bytes + sizeof(W_) + 1) / sizeof(W_);
1005 if (n+1 > BLOCK_SIZE_W) {
1006 barf("allocateExec: can't handle large objects");
1009 if (exec_block == NULL ||
1010 exec_block->free + n + 1 > exec_block->start + BLOCK_SIZE_W) {
1012 lnat pagesize = getPageSize();
1013 bd = allocGroup(stg_max(1, pagesize / BLOCK_SIZE));
1014 debugTrace(DEBUG_gc, "allocate exec block %p", bd->start);
1016 bd->flags = BF_EXEC;
1017 bd->link = exec_block;
1018 if (exec_block != NULL) {
1019 exec_block->u.back = bd;
1022 setExecutable(bd->start, bd->blocks * BLOCK_SIZE, rtsTrue);
1025 *(exec_block->free) = n; // store the size of this chunk
1026 exec_block->gen_no += n; // gen_no stores the number of words allocated
1027 ret = exec_block->free + 1;
1028 exec_block->free += n + 1;
1034 void freeExec (void *addr)
1036 StgPtr p = (StgPtr)addr - 1;
1037 bdescr *bd = Bdescr((StgPtr)p);
1039 if ((bd->flags & BF_EXEC) == 0) {
1040 barf("freeExec: not executable");
1043 if (*(StgPtr)p == 0) {
1044 barf("freeExec: already free?");
1049 bd->gen_no -= *(StgPtr)p;
1052 // Free the block if it is empty, but not if it is the block at
1053 // the head of the queue.
1054 if (bd->gen_no == 0 && bd != exec_block) {
1055 debugTrace(DEBUG_gc, "free exec block %p", bd->start);
1057 bd->u.back->link = bd->link;
1059 exec_block = bd->link;
1062 bd->link->u.back = bd->u.back;
1064 setExecutable(bd->start, bd->blocks * BLOCK_SIZE, rtsFalse);
1071 /* -----------------------------------------------------------------------------
1074 memInventory() checks for memory leaks by counting up all the
1075 blocks we know about and comparing that to the number of blocks
1076 allegedly floating around in the system.
1077 -------------------------------------------------------------------------- */
1082 stepBlocks (step *stp)
1087 total_blocks = stp->n_blocks;
1088 total_blocks += stp->n_old_blocks;
1089 for (bd = stp->large_objects; bd; bd = bd->link) {
1090 total_blocks += bd->blocks;
1091 /* hack for megablock groups: they have an extra block or two in
1092 the second and subsequent megablocks where the block
1093 descriptors would normally go.
1095 if (bd->blocks > BLOCKS_PER_MBLOCK) {
1096 total_blocks -= (MBLOCK_SIZE / BLOCK_SIZE - BLOCKS_PER_MBLOCK)
1097 * (bd->blocks/(MBLOCK_SIZE/BLOCK_SIZE));
1100 return total_blocks;
1109 lnat total_blocks = 0, free_blocks = 0;
1111 /* count the blocks we current have */
1113 for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
1114 for (i = 0; i < n_capabilities; i++) {
1115 for (bd = capabilities[i].mut_lists[g]; bd != NULL; bd = bd->link) {
1116 total_blocks += bd->blocks;
1119 for (bd = generations[g].mut_list; bd != NULL; bd = bd->link) {
1120 total_blocks += bd->blocks;
1122 for (s = 0; s < generations[g].n_steps; s++) {
1123 if (g==0 && s==0) continue;
1124 stp = &generations[g].steps[s];
1125 total_blocks += stepBlocks(stp);
1129 for (i = 0; i < n_nurseries; i++) {
1130 total_blocks += stepBlocks(&nurseries[i]);
1133 // We put pinned object blocks in g0s0, so better count blocks there too.
1134 total_blocks += stepBlocks(g0s0);
1137 /* any blocks held by allocate() */
1138 for (bd = small_alloc_list; bd; bd = bd->link) {
1139 total_blocks += bd->blocks;
1143 if (RtsFlags.ProfFlags.doHeapProfile == HEAP_BY_RETAINER) {
1144 total_blocks += retainerStackBlocks();
1148 // count the blocks allocated by the arena allocator
1149 total_blocks += arenaBlocks();
1151 // count the blocks containing executable memory
1152 for (bd = exec_block; bd; bd = bd->link) {
1153 total_blocks += bd->blocks;
1156 /* count the blocks on the free list */
1157 free_blocks = countFreeList();
1159 if (total_blocks + free_blocks != mblocks_allocated *
1160 BLOCKS_PER_MBLOCK) {
1161 debugBelch("Blocks: %ld live + %ld free = %ld total (%ld around)\n",
1162 total_blocks, free_blocks, total_blocks + free_blocks,
1163 mblocks_allocated * BLOCKS_PER_MBLOCK);
1166 ASSERT(total_blocks + free_blocks == mblocks_allocated * BLOCKS_PER_MBLOCK);
1171 countBlocks(bdescr *bd)
1174 for (n=0; bd != NULL; bd=bd->link) {
1180 /* Full heap sanity check. */
1186 if (RtsFlags.GcFlags.generations == 1) {
1187 checkHeap(g0s0->blocks);
1188 checkChain(g0s0->large_objects);
1191 for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
1192 for (s = 0; s < generations[g].n_steps; s++) {
1193 if (g == 0 && s == 0) { continue; }
1194 ASSERT(countBlocks(generations[g].steps[s].blocks)
1195 == generations[g].steps[s].n_blocks);
1196 ASSERT(countBlocks(generations[g].steps[s].large_objects)
1197 == generations[g].steps[s].n_large_blocks);
1198 checkHeap(generations[g].steps[s].blocks);
1199 checkChain(generations[g].steps[s].large_objects);
1201 checkMutableList(generations[g].mut_list, g);
1206 for (s = 0; s < n_nurseries; s++) {
1207 ASSERT(countBlocks(nurseries[s].blocks)
1208 == nurseries[s].n_blocks);
1209 ASSERT(countBlocks(nurseries[s].large_objects)
1210 == nurseries[s].n_large_blocks);
1213 checkFreeListSanity();
1217 /* Nursery sanity check */
1219 checkNurserySanity( step *stp )
1225 for (bd = stp->blocks; bd != NULL; bd = bd->link) {
1226 ASSERT(bd->u.back == prev);
1228 blocks += bd->blocks;
1230 ASSERT(blocks == stp->n_blocks);
1233 // handy function for use in gdb, because Bdescr() is inlined.
1234 extern bdescr *_bdescr( StgPtr p );