1 /* -----------------------------------------------------------------------------
3 * (c) The GHC Team, 1998-2006
5 * Storage manager front end
7 * Documentation on the architecture of the Storage Manager can be
8 * found in the online commentary:
10 * http://hackage.haskell.org/trac/ghc/wiki/Commentary/Rts/Storage
12 * ---------------------------------------------------------------------------*/
14 #include "PosixSource.h"
20 #include "BlockAlloc.h"
25 #include "OSThreads.h"
26 #include "Capability.h"
29 #include "RetainerProfile.h" // for counting memory blocks (memInventory)
37 * All these globals require sm_mutex to access in THREADED_RTS mode.
39 StgClosure *caf_list = NULL;
40 StgClosure *revertible_caf_list = NULL;
43 bdescr *small_alloc_list; /* allocate()d small objects */
44 bdescr *pinned_object_block; /* allocate pinned objects into this block */
45 nat alloc_blocks; /* number of allocate()d blocks since GC */
46 nat alloc_blocks_lim; /* approximate limit on alloc_blocks */
48 StgPtr alloc_Hp = NULL; /* next free byte in small_alloc_list */
49 StgPtr alloc_HpLim = NULL; /* end of block at small_alloc_list */
51 generation *generations = NULL; /* all the generations */
52 generation *g0 = NULL; /* generation 0, for convenience */
53 generation *oldest_gen = NULL; /* oldest generation, for convenience */
54 step *g0s0 = NULL; /* generation 0, step 0, for convenience */
56 ullong total_allocated = 0; /* total memory allocated during run */
58 nat n_nurseries = 0; /* == RtsFlags.ParFlags.nNodes, convenience */
59 step *nurseries = NULL; /* array of nurseries, >1 only if THREADED_RTS */
63 * Storage manager mutex: protects all the above state from
64 * simultaneous access by two STG threads.
68 * This mutex is used by atomicModifyMutVar# only
70 Mutex atomic_modify_mutvar_mutex;
77 static void *stgAllocForGMP (size_t size_in_bytes);
78 static void *stgReallocForGMP (void *ptr, size_t old_size, size_t new_size);
79 static void stgDeallocForGMP (void *ptr, size_t size);
82 initStep (step *stp, int g, int s)
87 stp->old_blocks = NULL;
88 stp->n_old_blocks = 0;
89 stp->gen = &generations[g];
95 stp->scavd_hpLim = NULL;
98 stp->large_objects = NULL;
99 stp->n_large_blocks = 0;
100 stp->new_large_objects = NULL;
101 stp->scavenged_large_objects = NULL;
102 stp->n_scavenged_large_blocks = 0;
103 stp->is_compacted = 0;
113 if (generations != NULL) {
114 // multi-init protection
118 /* Sanity check to make sure the LOOKS_LIKE_ macros appear to be
119 * doing something reasonable.
121 ASSERT(LOOKS_LIKE_INFO_PTR(&stg_BLACKHOLE_info));
122 ASSERT(LOOKS_LIKE_CLOSURE_PTR(&stg_dummy_ret_closure));
123 ASSERT(!HEAP_ALLOCED(&stg_dummy_ret_closure));
125 if (RtsFlags.GcFlags.maxHeapSize != 0 &&
126 RtsFlags.GcFlags.heapSizeSuggestion >
127 RtsFlags.GcFlags.maxHeapSize) {
128 RtsFlags.GcFlags.maxHeapSize = RtsFlags.GcFlags.heapSizeSuggestion;
131 if (RtsFlags.GcFlags.maxHeapSize != 0 &&
132 RtsFlags.GcFlags.minAllocAreaSize >
133 RtsFlags.GcFlags.maxHeapSize) {
134 errorBelch("maximum heap size (-M) is smaller than minimum alloc area size (-A)");
135 RtsFlags.GcFlags.minAllocAreaSize = RtsFlags.GcFlags.maxHeapSize;
138 initBlockAllocator();
140 #if defined(THREADED_RTS)
141 initMutex(&sm_mutex);
142 initMutex(&atomic_modify_mutvar_mutex);
147 /* allocate generation info array */
148 generations = (generation *)stgMallocBytes(RtsFlags.GcFlags.generations
149 * sizeof(struct generation_),
150 "initStorage: gens");
152 /* Initialise all generations */
153 for(g = 0; g < RtsFlags.GcFlags.generations; g++) {
154 gen = &generations[g];
156 gen->mut_list = allocBlock();
157 gen->collections = 0;
158 gen->failed_promotions = 0;
162 /* A couple of convenience pointers */
163 g0 = &generations[0];
164 oldest_gen = &generations[RtsFlags.GcFlags.generations-1];
166 /* Allocate step structures in each generation */
167 if (RtsFlags.GcFlags.generations > 1) {
168 /* Only for multiple-generations */
170 /* Oldest generation: one step */
171 oldest_gen->n_steps = 1;
173 stgMallocBytes(1 * sizeof(struct step_), "initStorage: last step");
175 /* set up all except the oldest generation with 2 steps */
176 for(g = 0; g < RtsFlags.GcFlags.generations-1; g++) {
177 generations[g].n_steps = RtsFlags.GcFlags.steps;
178 generations[g].steps =
179 stgMallocBytes (RtsFlags.GcFlags.steps * sizeof(struct step_),
180 "initStorage: steps");
184 /* single generation, i.e. a two-space collector */
186 g0->steps = stgMallocBytes (sizeof(struct step_), "initStorage: steps");
190 n_nurseries = n_capabilities;
191 nurseries = stgMallocBytes (n_nurseries * sizeof(struct step_),
192 "initStorage: nurseries");
195 nurseries = g0->steps; // just share nurseries[0] with g0s0
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);
206 for (s = 0; s < n_nurseries; s++) {
207 initStep(&nurseries[s], 0, s);
211 /* Set up the destination pointers in each younger gen. step */
212 for (g = 0; g < RtsFlags.GcFlags.generations-1; g++) {
213 for (s = 0; s < generations[g].n_steps-1; s++) {
214 generations[g].steps[s].to = &generations[g].steps[s+1];
216 generations[g].steps[s].to = &generations[g+1].steps[0];
218 oldest_gen->steps[0].to = &oldest_gen->steps[0];
221 for (s = 0; s < n_nurseries; s++) {
222 nurseries[s].to = generations[0].steps[0].to;
226 /* The oldest generation has one step. */
227 if (RtsFlags.GcFlags.compact) {
228 if (RtsFlags.GcFlags.generations == 1) {
229 errorBelch("WARNING: compaction is incompatible with -G1; disabled");
231 oldest_gen->steps[0].is_compacted = 1;
236 if (RtsFlags.GcFlags.generations == 1) {
237 errorBelch("-G1 is incompatible with -threaded");
238 stg_exit(EXIT_FAILURE);
242 /* generation 0 is special: that's the nursery */
243 generations[0].max_blocks = 0;
245 /* G0S0: the allocation area. Policy: keep the allocation area
246 * small to begin with, even if we have a large suggested heap
247 * size. Reason: we're going to do a major collection first, and we
248 * don't want it to be a big one. This vague idea is borne out by
249 * rigorous experimental evidence.
251 g0s0 = &generations[0].steps[0];
255 weak_ptr_list = NULL;
257 revertible_caf_list = NULL;
259 /* initialise the allocate() interface */
260 small_alloc_list = NULL;
262 alloc_blocks_lim = RtsFlags.GcFlags.minAllocAreaSize;
264 /* Tell GNU multi-precision pkg about our custom alloc functions */
265 mp_set_memory_functions(stgAllocForGMP, stgReallocForGMP, stgDeallocForGMP);
267 IF_DEBUG(gc, statDescribeGens());
275 stat_exit(calcAllocated());
283 for(g = 0; g < RtsFlags.GcFlags.generations; g++)
284 stgFree(generations[g].steps);
285 stgFree(generations);
287 #if defined(THREADED_RTS)
288 closeMutex(&sm_mutex);
289 closeMutex(&atomic_modify_mutvar_mutex);
294 /* -----------------------------------------------------------------------------
297 The entry code for every CAF does the following:
299 - builds a CAF_BLACKHOLE in the heap
300 - pushes an update frame pointing to the CAF_BLACKHOLE
301 - invokes UPD_CAF(), which:
302 - calls newCaf, below
303 - updates the CAF with a static indirection to the CAF_BLACKHOLE
305 Why do we build a BLACKHOLE in the heap rather than just updating
306 the thunk directly? It's so that we only need one kind of update
307 frame - otherwise we'd need a static version of the update frame too.
309 newCaf() does the following:
311 - it puts the CAF on the oldest generation's mut-once list.
312 This is so that we can treat the CAF as a root when collecting
315 For GHCI, we have additional requirements when dealing with CAFs:
317 - we must *retain* all dynamically-loaded CAFs ever entered,
318 just in case we need them again.
319 - we must be able to *revert* CAFs that have been evaluated, to
320 their pre-evaluated form.
322 To do this, we use an additional CAF list. When newCaf() is
323 called on a dynamically-loaded CAF, we add it to the CAF list
324 instead of the old-generation mutable list, and save away its
325 old info pointer (in caf->saved_info) for later reversion.
327 To revert all the CAFs, we traverse the CAF list and reset the
328 info pointer to caf->saved_info, then throw away the CAF list.
329 (see GC.c:revertCAFs()).
333 -------------------------------------------------------------------------- */
336 newCAF(StgClosure* caf)
343 // If we are in GHCi _and_ we are using dynamic libraries,
344 // then we can't redirect newCAF calls to newDynCAF (see below),
345 // so we make newCAF behave almost like newDynCAF.
346 // The dynamic libraries might be used by both the interpreted
347 // program and GHCi itself, so they must not be reverted.
348 // This also means that in GHCi with dynamic libraries, CAFs are not
349 // garbage collected. If this turns out to be a problem, we could
350 // do another hack here and do an address range test on caf to figure
351 // out whether it is from a dynamic library.
352 ((StgIndStatic *)caf)->saved_info = (StgInfoTable *)caf->header.info;
353 ((StgIndStatic *)caf)->static_link = caf_list;
358 /* Put this CAF on the mutable list for the old generation.
359 * This is a HACK - the IND_STATIC closure doesn't really have
360 * a mut_link field, but we pretend it has - in fact we re-use
361 * the STATIC_LINK field for the time being, because when we
362 * come to do a major GC we won't need the mut_link field
363 * any more and can use it as a STATIC_LINK.
365 ((StgIndStatic *)caf)->saved_info = NULL;
366 recordMutableGen(caf, oldest_gen);
372 // An alternate version of newCaf which is used for dynamically loaded
373 // object code in GHCi. In this case we want to retain *all* CAFs in
374 // the object code, because they might be demanded at any time from an
375 // expression evaluated on the command line.
376 // Also, GHCi might want to revert CAFs, so we add these to the
377 // revertible_caf_list.
379 // The linker hackily arranges that references to newCaf from dynamic
380 // code end up pointing to newDynCAF.
382 newDynCAF(StgClosure *caf)
386 ((StgIndStatic *)caf)->saved_info = (StgInfoTable *)caf->header.info;
387 ((StgIndStatic *)caf)->static_link = revertible_caf_list;
388 revertible_caf_list = caf;
393 /* -----------------------------------------------------------------------------
395 -------------------------------------------------------------------------- */
398 allocNursery (step *stp, bdescr *tail, nat blocks)
403 // Allocate a nursery: we allocate fresh blocks one at a time and
404 // cons them on to the front of the list, not forgetting to update
405 // the back pointer on the tail of the list to point to the new block.
406 for (i=0; i < blocks; i++) {
409 processNursery() in LdvProfile.c assumes that every block group in
410 the nursery contains only a single block. So, if a block group is
411 given multiple blocks, change processNursery() accordingly.
415 // double-link the nursery: we might need to insert blocks
422 bd->free = bd->start;
430 assignNurseriesToCapabilities (void)
435 for (i = 0; i < n_nurseries; i++) {
436 capabilities[i].r.rNursery = &nurseries[i];
437 capabilities[i].r.rCurrentNursery = nurseries[i].blocks;
438 capabilities[i].r.rCurrentAlloc = NULL;
440 #else /* THREADED_RTS */
441 MainCapability.r.rNursery = &nurseries[0];
442 MainCapability.r.rCurrentNursery = nurseries[0].blocks;
443 MainCapability.r.rCurrentAlloc = NULL;
448 allocNurseries( void )
452 for (i = 0; i < n_nurseries; i++) {
453 nurseries[i].blocks =
454 allocNursery(&nurseries[i], NULL,
455 RtsFlags.GcFlags.minAllocAreaSize);
456 nurseries[i].n_blocks = RtsFlags.GcFlags.minAllocAreaSize;
457 nurseries[i].old_blocks = NULL;
458 nurseries[i].n_old_blocks = 0;
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 += bd->blocks; // might be larger than 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 allocatedBytes( 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 += bd->blocks; // might be larger than 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);
767 g0s0->n_large_blocks++;
770 bd->flags = BF_PINNED | BF_LARGE;
771 bd->free = bd->start;
781 /* -----------------------------------------------------------------------------
782 This is the write barrier for MUT_VARs, a.k.a. IORefs. A
783 MUT_VAR_CLEAN object is not on the mutable list; a MUT_VAR_DIRTY
784 is. When written to, a MUT_VAR_CLEAN turns into a MUT_VAR_DIRTY
785 and is put on the mutable list.
786 -------------------------------------------------------------------------- */
789 dirty_MUT_VAR(StgRegTable *reg, StgClosure *p)
791 Capability *cap = regTableToCapability(reg);
793 if (p->header.info == &stg_MUT_VAR_CLEAN_info) {
794 p->header.info = &stg_MUT_VAR_DIRTY_info;
795 bd = Bdescr((StgPtr)p);
796 if (bd->gen_no > 0) recordMutableCap(p,cap,bd->gen_no);
800 /* -----------------------------------------------------------------------------
801 Allocation functions for GMP.
803 These all use the allocate() interface - we can't have any garbage
804 collection going on during a gmp operation, so we use allocate()
805 which always succeeds. The gmp operations which might need to
806 allocate will ask the storage manager (via doYouWantToGC()) whether
807 a garbage collection is required, in case we get into a loop doing
808 only allocate() style allocation.
809 -------------------------------------------------------------------------- */
812 stgAllocForGMP (size_t size_in_bytes)
815 nat data_size_in_words, total_size_in_words;
817 /* round up to a whole number of words */
818 data_size_in_words = (size_in_bytes + sizeof(W_) + 1) / sizeof(W_);
819 total_size_in_words = sizeofW(StgArrWords) + data_size_in_words;
821 /* allocate and fill it in. */
822 #if defined(THREADED_RTS)
823 arr = (StgArrWords *)allocateLocal(myTask()->cap, total_size_in_words);
825 arr = (StgArrWords *)allocateLocal(&MainCapability, total_size_in_words);
827 SET_ARR_HDR(arr, &stg_ARR_WORDS_info, CCCS, data_size_in_words);
829 /* and return a ptr to the goods inside the array */
834 stgReallocForGMP (void *ptr, size_t old_size, size_t new_size)
836 void *new_stuff_ptr = stgAllocForGMP(new_size);
838 char *p = (char *) ptr;
839 char *q = (char *) new_stuff_ptr;
841 for (; i < old_size; i++, p++, q++) {
845 return(new_stuff_ptr);
849 stgDeallocForGMP (void *ptr STG_UNUSED,
850 size_t size STG_UNUSED)
852 /* easy for us: the garbage collector does the dealloc'n */
855 /* -----------------------------------------------------------------------------
857 * -------------------------------------------------------------------------- */
859 /* -----------------------------------------------------------------------------
862 * Approximate how much we've allocated: number of blocks in the
863 * nursery + blocks allocated via allocate() - unused nusery blocks.
864 * This leaves a little slop at the end of each block, and doesn't
865 * take into account large objects (ToDo).
866 * -------------------------------------------------------------------------- */
869 calcAllocated( void )
874 allocated = allocatedBytes();
875 allocated += countNurseryBlocks() * BLOCK_SIZE_W;
880 for (i = 0; i < n_nurseries; i++) {
882 for ( bd = capabilities[i].r.rCurrentNursery->link;
883 bd != NULL; bd = bd->link ) {
884 allocated -= BLOCK_SIZE_W;
886 cap = &capabilities[i];
887 if (cap->r.rCurrentNursery->free <
888 cap->r.rCurrentNursery->start + BLOCK_SIZE_W) {
889 allocated -= (cap->r.rCurrentNursery->start + BLOCK_SIZE_W)
890 - cap->r.rCurrentNursery->free;
894 bdescr *current_nursery = MainCapability.r.rCurrentNursery;
896 for ( bd = current_nursery->link; bd != NULL; bd = bd->link ) {
897 allocated -= BLOCK_SIZE_W;
899 if (current_nursery->free < current_nursery->start + BLOCK_SIZE_W) {
900 allocated -= (current_nursery->start + BLOCK_SIZE_W)
901 - current_nursery->free;
906 total_allocated += allocated;
910 /* Approximate the amount of live data in the heap. To be called just
911 * after garbage collection (see GarbageCollect()).
920 if (RtsFlags.GcFlags.generations == 1) {
921 return (g0s0->n_large_blocks + g0s0->n_blocks) * BLOCK_SIZE_W;
924 for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
925 for (s = 0; s < generations[g].n_steps; s++) {
926 /* approximate amount of live data (doesn't take into account slop
927 * at end of each block).
929 if (g == 0 && s == 0) {
932 stp = &generations[g].steps[s];
933 live += (stp->n_large_blocks + stp->n_blocks) * BLOCK_SIZE_W;
939 /* Approximate the number of blocks that will be needed at the next
940 * garbage collection.
942 * Assume: all data currently live will remain live. Steps that will
943 * be collected next time will therefore need twice as many blocks
944 * since all the data will be copied.
953 for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
954 for (s = 0; s < generations[g].n_steps; s++) {
955 if (g == 0 && s == 0) { continue; }
956 stp = &generations[g].steps[s];
957 if (generations[g].steps[0].n_blocks +
958 generations[g].steps[0].n_large_blocks
959 > generations[g].max_blocks
960 && stp->is_compacted == 0) {
961 needed += 2 * stp->n_blocks;
963 needed += stp->n_blocks;
970 /* ----------------------------------------------------------------------------
973 Executable memory must be managed separately from non-executable
974 memory. Most OSs these days require you to jump through hoops to
975 dynamically allocate executable memory, due to various security
978 Here we provide a small memory allocator for executable memory.
979 Memory is managed with a page granularity; we allocate linearly
980 in the page, and when the page is emptied (all objects on the page
981 are free) we free the page again, not forgetting to make it
984 TODO: The inability to handle objects bigger than BLOCK_SIZE_W means that
985 the linker cannot use allocateExec for loading object code files
986 on Windows. Once allocateExec can handle larger objects, the linker
987 should be modified to use allocateExec instead of VirtualAlloc.
988 ------------------------------------------------------------------------- */
990 static bdescr *exec_block;
992 void *allocateExec (nat bytes)
999 // round up to words.
1000 n = (bytes + sizeof(W_) + 1) / sizeof(W_);
1002 if (n+1 > BLOCK_SIZE_W) {
1003 barf("allocateExec: can't handle large objects");
1006 if (exec_block == NULL ||
1007 exec_block->free + n + 1 > exec_block->start + BLOCK_SIZE_W) {
1009 lnat pagesize = getPageSize();
1010 bd = allocGroup(stg_max(1, pagesize / BLOCK_SIZE));
1011 debugTrace(DEBUG_gc, "allocate exec block %p", bd->start);
1013 bd->flags = BF_EXEC;
1014 bd->link = exec_block;
1015 if (exec_block != NULL) {
1016 exec_block->u.back = bd;
1019 setExecutable(bd->start, bd->blocks * BLOCK_SIZE, rtsTrue);
1022 *(exec_block->free) = n; // store the size of this chunk
1023 exec_block->gen_no += n; // gen_no stores the number of words allocated
1024 ret = exec_block->free + 1;
1025 exec_block->free += n + 1;
1031 void freeExec (void *addr)
1033 StgPtr p = (StgPtr)addr - 1;
1034 bdescr *bd = Bdescr((StgPtr)p);
1036 if ((bd->flags & BF_EXEC) == 0) {
1037 barf("freeExec: not executable");
1040 if (*(StgPtr)p == 0) {
1041 barf("freeExec: already free?");
1046 bd->gen_no -= *(StgPtr)p;
1049 // Free the block if it is empty, but not if it is the block at
1050 // the head of the queue.
1051 if (bd->gen_no == 0 && bd != exec_block) {
1052 debugTrace(DEBUG_gc, "free exec block %p", bd->start);
1053 dbl_link_remove(bd, &exec_block);
1054 setExecutable(bd->start, bd->blocks * BLOCK_SIZE, rtsFalse);
1061 /* -----------------------------------------------------------------------------
1064 memInventory() checks for memory leaks by counting up all the
1065 blocks we know about and comparing that to the number of blocks
1066 allegedly floating around in the system.
1067 -------------------------------------------------------------------------- */
1072 countBlocks(bdescr *bd)
1075 for (n=0; bd != NULL; bd=bd->link) {
1081 // (*1) Just like countBlocks, except that we adjust the count for a
1082 // megablock group so that it doesn't include the extra few blocks
1083 // that would be taken up by block descriptors in the second and
1084 // subsequent megablock. This is so we can tally the count with the
1085 // number of blocks allocated in the system, for memInventory().
1087 countAllocdBlocks(bdescr *bd)
1090 for (n=0; bd != NULL; bd=bd->link) {
1092 // hack for megablock groups: see (*1) above
1093 if (bd->blocks > BLOCKS_PER_MBLOCK) {
1094 n -= (MBLOCK_SIZE / BLOCK_SIZE - BLOCKS_PER_MBLOCK)
1095 * (bd->blocks/(MBLOCK_SIZE/BLOCK_SIZE));
1102 stepBlocks (step *stp)
1104 ASSERT(countBlocks(stp->blocks) == stp->n_blocks);
1105 ASSERT(countBlocks(stp->large_objects) == stp->n_large_blocks);
1106 return stp->n_blocks + stp->n_old_blocks +
1107 countAllocdBlocks(stp->large_objects);
1115 lnat gen_blocks[RtsFlags.GcFlags.generations];
1116 lnat nursery_blocks, allocate_blocks, retainer_blocks,
1117 arena_blocks, exec_blocks;
1118 lnat live_blocks = 0, free_blocks = 0;
1120 // count the blocks we current have
1122 for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
1124 for (i = 0; i < n_capabilities; i++) {
1125 gen_blocks[g] += countBlocks(capabilities[i].mut_lists[g]);
1127 gen_blocks[g] += countAllocdBlocks(generations[g].mut_list);
1128 for (s = 0; s < generations[g].n_steps; s++) {
1129 #if !defined(THREADED_RTS)
1130 // We put pinned object blocks in g0s0, so better count
1131 // blocks there too.
1132 if (g==0 && s==0) continue;
1134 stp = &generations[g].steps[s];
1135 gen_blocks[g] += stepBlocks(stp);
1140 for (i = 0; i < n_nurseries; i++) {
1141 nursery_blocks += stepBlocks(&nurseries[i]);
1144 /* any blocks held by allocate() */
1145 allocate_blocks = countAllocdBlocks(small_alloc_list);
1147 retainer_blocks = 0;
1149 if (RtsFlags.ProfFlags.doHeapProfile == HEAP_BY_RETAINER) {
1150 retainer_blocks = retainerStackBlocks();
1154 // count the blocks allocated by the arena allocator
1155 arena_blocks = arenaBlocks();
1157 // count the blocks containing executable memory
1158 exec_blocks = countAllocdBlocks(exec_block);
1160 /* count the blocks on the free list */
1161 free_blocks = countFreeList();
1164 for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
1165 live_blocks += gen_blocks[g];
1167 live_blocks += nursery_blocks + allocate_blocks
1168 + retainer_blocks + arena_blocks + exec_blocks;
1170 if (live_blocks + free_blocks != mblocks_allocated * BLOCKS_PER_MBLOCK)
1172 debugBelch("Memory leak detected\n");
1173 for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
1174 debugBelch(" gen %d blocks : %4lu\n", g, gen_blocks[g]);
1176 debugBelch(" nursery : %4lu\n", nursery_blocks);
1177 debugBelch(" allocate() : %4lu\n", allocate_blocks);
1178 debugBelch(" retainer : %4lu\n", retainer_blocks);
1179 debugBelch(" arena blocks : %4lu\n", arena_blocks);
1180 debugBelch(" exec : %4lu\n", exec_blocks);
1181 debugBelch(" free : %4lu\n", free_blocks);
1182 debugBelch(" total : %4lu\n\n", live_blocks + free_blocks);
1183 debugBelch(" in system : %4lu\n", mblocks_allocated * BLOCKS_PER_MBLOCK);
1189 /* Full heap sanity check. */
1195 if (RtsFlags.GcFlags.generations == 1) {
1196 checkHeap(g0s0->blocks);
1197 checkChain(g0s0->large_objects);
1200 for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
1201 for (s = 0; s < generations[g].n_steps; s++) {
1202 if (g == 0 && s == 0) { continue; }
1203 ASSERT(countBlocks(generations[g].steps[s].blocks)
1204 == generations[g].steps[s].n_blocks);
1205 ASSERT(countBlocks(generations[g].steps[s].large_objects)
1206 == generations[g].steps[s].n_large_blocks);
1207 checkHeap(generations[g].steps[s].blocks);
1208 checkChain(generations[g].steps[s].large_objects);
1210 checkMutableList(generations[g].mut_list, g);
1215 for (s = 0; s < n_nurseries; s++) {
1216 ASSERT(countBlocks(nurseries[s].blocks)
1217 == nurseries[s].n_blocks);
1218 ASSERT(countBlocks(nurseries[s].large_objects)
1219 == nurseries[s].n_large_blocks);
1222 checkFreeListSanity();
1226 /* Nursery sanity check */
1228 checkNurserySanity( step *stp )
1234 for (bd = stp->blocks; bd != NULL; bd = bd->link) {
1235 ASSERT(bd->u.back == prev);
1237 blocks += bd->blocks;
1239 ASSERT(blocks == stp->n_blocks);
1242 // handy function for use in gdb, because Bdescr() is inlined.
1243 extern bdescr *_bdescr( StgPtr p );