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
120 /* Sanity check to make sure the LOOKS_LIKE_ macros appear to be
121 * doing something reasonable.
123 ASSERT(LOOKS_LIKE_INFO_PTR(&stg_BLACKHOLE_info));
124 ASSERT(LOOKS_LIKE_CLOSURE_PTR(&stg_dummy_ret_closure));
125 ASSERT(!HEAP_ALLOCED(&stg_dummy_ret_closure));
127 if (RtsFlags.GcFlags.maxHeapSize != 0 &&
128 RtsFlags.GcFlags.heapSizeSuggestion >
129 RtsFlags.GcFlags.maxHeapSize) {
130 RtsFlags.GcFlags.maxHeapSize = RtsFlags.GcFlags.heapSizeSuggestion;
133 if (RtsFlags.GcFlags.maxHeapSize != 0 &&
134 RtsFlags.GcFlags.minAllocAreaSize >
135 RtsFlags.GcFlags.maxHeapSize) {
136 errorBelch("maximum heap size (-M) is smaller than minimum alloc area size (-A)");
137 RtsFlags.GcFlags.minAllocAreaSize = RtsFlags.GcFlags.maxHeapSize;
140 initBlockAllocator();
142 #if defined(THREADED_RTS)
143 initMutex(&sm_mutex);
144 initMutex(&atomic_modify_mutvar_mutex);
149 /* allocate generation info array */
150 generations = (generation *)stgMallocBytes(RtsFlags.GcFlags.generations
151 * sizeof(struct generation_),
152 "initStorage: gens");
154 /* Initialise all generations */
155 for(g = 0; g < RtsFlags.GcFlags.generations; g++) {
156 gen = &generations[g];
158 gen->mut_list = allocBlock();
159 gen->collections = 0;
160 gen->failed_promotions = 0;
164 /* A couple of convenience pointers */
165 g0 = &generations[0];
166 oldest_gen = &generations[RtsFlags.GcFlags.generations-1];
168 /* Allocate step structures in each generation */
169 if (RtsFlags.GcFlags.generations > 1) {
170 /* Only for multiple-generations */
172 /* Oldest generation: one step */
173 oldest_gen->n_steps = 1;
175 stgMallocBytes(1 * sizeof(struct step_), "initStorage: last step");
177 /* set up all except the oldest generation with 2 steps */
178 for(g = 0; g < RtsFlags.GcFlags.generations-1; g++) {
179 generations[g].n_steps = RtsFlags.GcFlags.steps;
180 generations[g].steps =
181 stgMallocBytes (RtsFlags.GcFlags.steps * sizeof(struct step_),
182 "initStorage: steps");
186 /* single generation, i.e. a two-space collector */
188 g0->steps = stgMallocBytes (sizeof(struct step_), "initStorage: steps");
192 n_nurseries = n_capabilities;
193 nurseries = stgMallocBytes (n_nurseries * sizeof(struct step_),
194 "initStorage: nurseries");
197 nurseries = g0->steps; // just share nurseries[0] with g0s0
200 /* Initialise all steps */
201 for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
202 for (s = 0; s < generations[g].n_steps; s++) {
203 initStep(&generations[g].steps[s], g, s);
208 for (s = 0; s < n_nurseries; s++) {
209 initStep(&nurseries[s], 0, s);
213 /* Set up the destination pointers in each younger gen. step */
214 for (g = 0; g < RtsFlags.GcFlags.generations-1; g++) {
215 for (s = 0; s < generations[g].n_steps-1; s++) {
216 generations[g].steps[s].to = &generations[g].steps[s+1];
218 generations[g].steps[s].to = &generations[g+1].steps[0];
220 oldest_gen->steps[0].to = &oldest_gen->steps[0];
223 for (s = 0; s < n_nurseries; s++) {
224 nurseries[s].to = generations[0].steps[0].to;
228 /* The oldest generation has one step. */
229 if (RtsFlags.GcFlags.compact) {
230 if (RtsFlags.GcFlags.generations == 1) {
231 errorBelch("WARNING: compaction is incompatible with -G1; disabled");
233 oldest_gen->steps[0].is_compacted = 1;
238 if (RtsFlags.GcFlags.generations == 1) {
239 errorBelch("-G1 is incompatible with -threaded");
240 stg_exit(EXIT_FAILURE);
244 /* generation 0 is special: that's the nursery */
245 generations[0].max_blocks = 0;
247 /* G0S0: the allocation area. Policy: keep the allocation area
248 * small to begin with, even if we have a large suggested heap
249 * size. Reason: we're going to do a major collection first, and we
250 * don't want it to be a big one. This vague idea is borne out by
251 * rigorous experimental evidence.
253 g0s0 = &generations[0].steps[0];
257 weak_ptr_list = NULL;
259 revertible_caf_list = NULL;
261 /* initialise the allocate() interface */
262 small_alloc_list = NULL;
264 alloc_blocks_lim = RtsFlags.GcFlags.minAllocAreaSize;
266 /* Tell GNU multi-precision pkg about our custom alloc functions */
267 mp_set_memory_functions(stgAllocForGMP, stgReallocForGMP, stgDeallocForGMP);
269 IF_DEBUG(gc, statDescribeGens());
277 stat_exit(calcAllocated());
285 for(g = 0; g < RtsFlags.GcFlags.generations; g++)
286 stgFree(generations[g].steps);
287 stgFree(generations);
289 #if defined(THREADED_RTS)
290 closeMutex(&sm_mutex);
291 closeMutex(&atomic_modify_mutvar_mutex);
296 /* -----------------------------------------------------------------------------
299 The entry code for every CAF does the following:
301 - builds a CAF_BLACKHOLE in the heap
302 - pushes an update frame pointing to the CAF_BLACKHOLE
303 - invokes UPD_CAF(), which:
304 - calls newCaf, below
305 - updates the CAF with a static indirection to the CAF_BLACKHOLE
307 Why do we build a BLACKHOLE in the heap rather than just updating
308 the thunk directly? It's so that we only need one kind of update
309 frame - otherwise we'd need a static version of the update frame too.
311 newCaf() does the following:
313 - it puts the CAF on the oldest generation's mut-once list.
314 This is so that we can treat the CAF as a root when collecting
317 For GHCI, we have additional requirements when dealing with CAFs:
319 - we must *retain* all dynamically-loaded CAFs ever entered,
320 just in case we need them again.
321 - we must be able to *revert* CAFs that have been evaluated, to
322 their pre-evaluated form.
324 To do this, we use an additional CAF list. When newCaf() is
325 called on a dynamically-loaded CAF, we add it to the CAF list
326 instead of the old-generation mutable list, and save away its
327 old info pointer (in caf->saved_info) for later reversion.
329 To revert all the CAFs, we traverse the CAF list and reset the
330 info pointer to caf->saved_info, then throw away the CAF list.
331 (see GC.c:revertCAFs()).
335 -------------------------------------------------------------------------- */
338 newCAF(StgClosure* caf)
345 // If we are in GHCi _and_ we are using dynamic libraries,
346 // then we can't redirect newCAF calls to newDynCAF (see below),
347 // so we make newCAF behave almost like newDynCAF.
348 // The dynamic libraries might be used by both the interpreted
349 // program and GHCi itself, so they must not be reverted.
350 // This also means that in GHCi with dynamic libraries, CAFs are not
351 // garbage collected. If this turns out to be a problem, we could
352 // do another hack here and do an address range test on caf to figure
353 // out whether it is from a dynamic library.
354 ((StgIndStatic *)caf)->saved_info = (StgInfoTable *)caf->header.info;
355 ((StgIndStatic *)caf)->static_link = caf_list;
360 /* Put this CAF on the mutable list for the old generation.
361 * This is a HACK - the IND_STATIC closure doesn't really have
362 * a mut_link field, but we pretend it has - in fact we re-use
363 * the STATIC_LINK field for the time being, because when we
364 * come to do a major GC we won't need the mut_link field
365 * any more and can use it as a STATIC_LINK.
367 ((StgIndStatic *)caf)->saved_info = NULL;
368 recordMutableGen(caf, oldest_gen);
374 // An alternate version of newCaf which is used for dynamically loaded
375 // object code in GHCi. In this case we want to retain *all* CAFs in
376 // the object code, because they might be demanded at any time from an
377 // expression evaluated on the command line.
378 // Also, GHCi might want to revert CAFs, so we add these to the
379 // revertible_caf_list.
381 // The linker hackily arranges that references to newCaf from dynamic
382 // code end up pointing to newDynCAF.
384 newDynCAF(StgClosure *caf)
388 ((StgIndStatic *)caf)->saved_info = (StgInfoTable *)caf->header.info;
389 ((StgIndStatic *)caf)->static_link = revertible_caf_list;
390 revertible_caf_list = caf;
395 /* -----------------------------------------------------------------------------
397 -------------------------------------------------------------------------- */
400 allocNursery (step *stp, bdescr *tail, nat blocks)
405 // Allocate a nursery: we allocate fresh blocks one at a time and
406 // cons them on to the front of the list, not forgetting to update
407 // the back pointer on the tail of the list to point to the new block.
408 for (i=0; i < blocks; i++) {
411 processNursery() in LdvProfile.c assumes that every block group in
412 the nursery contains only a single block. So, if a block group is
413 given multiple blocks, change processNursery() accordingly.
417 // double-link the nursery: we might need to insert blocks
424 bd->free = bd->start;
432 assignNurseriesToCapabilities (void)
437 for (i = 0; i < n_nurseries; i++) {
438 capabilities[i].r.rNursery = &nurseries[i];
439 capabilities[i].r.rCurrentNursery = nurseries[i].blocks;
440 capabilities[i].r.rCurrentAlloc = NULL;
442 #else /* THREADED_RTS */
443 MainCapability.r.rNursery = &nurseries[0];
444 MainCapability.r.rCurrentNursery = nurseries[0].blocks;
445 MainCapability.r.rCurrentAlloc = NULL;
450 allocNurseries( void )
454 for (i = 0; i < n_nurseries; i++) {
455 nurseries[i].blocks =
456 allocNursery(&nurseries[i], NULL,
457 RtsFlags.GcFlags.minAllocAreaSize);
458 nurseries[i].n_blocks = RtsFlags.GcFlags.minAllocAreaSize;
459 nurseries[i].old_blocks = NULL;
460 nurseries[i].n_old_blocks = 0;
462 assignNurseriesToCapabilities();
466 resetNurseries( void )
472 for (i = 0; i < n_nurseries; i++) {
474 for (bd = stp->blocks; bd; bd = bd->link) {
475 bd->free = bd->start;
476 ASSERT(bd->gen_no == 0);
477 ASSERT(bd->step == stp);
478 IF_DEBUG(sanity,memset(bd->start, 0xaa, BLOCK_SIZE));
481 assignNurseriesToCapabilities();
485 countNurseryBlocks (void)
490 for (i = 0; i < n_nurseries; i++) {
491 blocks += nurseries[i].n_blocks;
497 resizeNursery ( step *stp, nat blocks )
502 nursery_blocks = stp->n_blocks;
503 if (nursery_blocks == blocks) return;
505 if (nursery_blocks < blocks) {
506 debugTrace(DEBUG_gc, "increasing size of nursery to %d blocks",
508 stp->blocks = allocNursery(stp, stp->blocks, blocks-nursery_blocks);
513 debugTrace(DEBUG_gc, "decreasing size of nursery to %d blocks",
517 while (nursery_blocks > blocks) {
519 next_bd->u.back = NULL;
520 nursery_blocks -= bd->blocks; // might be a large block
525 // might have gone just under, by freeing a large block, so make
526 // up the difference.
527 if (nursery_blocks < blocks) {
528 stp->blocks = allocNursery(stp, stp->blocks, blocks-nursery_blocks);
532 stp->n_blocks = blocks;
533 ASSERT(countBlocks(stp->blocks) == stp->n_blocks);
537 // Resize each of the nurseries to the specified size.
540 resizeNurseriesFixed (nat blocks)
543 for (i = 0; i < n_nurseries; i++) {
544 resizeNursery(&nurseries[i], blocks);
549 // Resize the nurseries to the total specified size.
552 resizeNurseries (nat blocks)
554 // If there are multiple nurseries, then we just divide the number
555 // of available blocks between them.
556 resizeNurseriesFixed(blocks / n_nurseries);
559 /* -----------------------------------------------------------------------------
560 The allocate() interface
562 allocate(n) always succeeds, and returns a chunk of memory n words
563 long. n can be larger than the size of a block if necessary, in
564 which case a contiguous block group will be allocated.
565 -------------------------------------------------------------------------- */
575 TICK_ALLOC_HEAP_NOCTR(n);
578 /* big allocation (>LARGE_OBJECT_THRESHOLD) */
579 /* ToDo: allocate directly into generation 1 */
580 if (n >= LARGE_OBJECT_THRESHOLD/sizeof(W_)) {
581 nat req_blocks = (lnat)BLOCK_ROUND_UP(n*sizeof(W_)) / BLOCK_SIZE;
582 bd = allocGroup(req_blocks);
583 dbl_link_onto(bd, &g0s0->large_objects);
584 g0s0->n_large_blocks += bd->blocks; // might be larger than req_blocks
587 bd->flags = BF_LARGE;
588 bd->free = bd->start + n;
589 alloc_blocks += req_blocks;
593 /* small allocation (<LARGE_OBJECT_THRESHOLD) */
594 } else if (small_alloc_list == NULL || alloc_Hp + n > alloc_HpLim) {
595 if (small_alloc_list) {
596 small_alloc_list->free = alloc_Hp;
599 bd->link = small_alloc_list;
600 small_alloc_list = bd;
604 alloc_Hp = bd->start;
605 alloc_HpLim = bd->start + BLOCK_SIZE_W;
616 allocatedBytes( void )
620 allocated = alloc_blocks * BLOCK_SIZE_W - (alloc_HpLim - alloc_Hp);
621 if (pinned_object_block != NULL) {
622 allocated -= (pinned_object_block->start + BLOCK_SIZE_W) -
623 pinned_object_block->free;
630 tidyAllocateLists (void)
632 if (small_alloc_list != NULL) {
633 ASSERT(alloc_Hp >= small_alloc_list->start &&
634 alloc_Hp <= small_alloc_list->start + BLOCK_SIZE);
635 small_alloc_list->free = alloc_Hp;
639 /* -----------------------------------------------------------------------------
642 This allocates memory in the current thread - it is intended for
643 use primarily from STG-land where we have a Capability. It is
644 better than allocate() because it doesn't require taking the
645 sm_mutex lock in the common case.
647 Memory is allocated directly from the nursery if possible (but not
648 from the current nursery block, so as not to interfere with
650 -------------------------------------------------------------------------- */
653 allocateLocal (Capability *cap, nat n)
658 TICK_ALLOC_HEAP_NOCTR(n);
661 /* big allocation (>LARGE_OBJECT_THRESHOLD) */
662 /* ToDo: allocate directly into generation 1 */
663 if (n >= LARGE_OBJECT_THRESHOLD/sizeof(W_)) {
664 nat req_blocks = (lnat)BLOCK_ROUND_UP(n*sizeof(W_)) / BLOCK_SIZE;
666 bd = allocGroup(req_blocks);
667 dbl_link_onto(bd, &g0s0->large_objects);
668 g0s0->n_large_blocks += bd->blocks; // might be larger than req_blocks
671 bd->flags = BF_LARGE;
672 bd->free = bd->start + n;
673 alloc_blocks += req_blocks;
677 /* small allocation (<LARGE_OBJECT_THRESHOLD) */
680 bd = cap->r.rCurrentAlloc;
681 if (bd == NULL || bd->free + n > bd->start + BLOCK_SIZE_W) {
683 // The CurrentAlloc block is full, we need to find another
684 // one. First, we try taking the next block from the
686 bd = cap->r.rCurrentNursery->link;
688 if (bd == NULL || bd->free + n > bd->start + BLOCK_SIZE_W) {
689 // The nursery is empty, or the next block is already
690 // full: allocate a fresh block (we can't fail here).
693 cap->r.rNursery->n_blocks++;
696 bd->step = cap->r.rNursery;
699 // we have a block in the nursery: take it and put
700 // it at the *front* of the nursery list, and use it
701 // to allocate() from.
702 cap->r.rCurrentNursery->link = bd->link;
703 if (bd->link != NULL) {
704 bd->link->u.back = cap->r.rCurrentNursery;
707 dbl_link_onto(bd, &cap->r.rNursery->blocks);
708 cap->r.rCurrentAlloc = bd;
709 IF_DEBUG(sanity, checkNurserySanity(cap->r.rNursery));
717 /* ---------------------------------------------------------------------------
718 Allocate a fixed/pinned object.
720 We allocate small pinned objects into a single block, allocating a
721 new block when the current one overflows. The block is chained
722 onto the large_object_list of generation 0 step 0.
724 NOTE: The GC can't in general handle pinned objects. This
725 interface is only safe to use for ByteArrays, which have no
726 pointers and don't require scavenging. It works because the
727 block's descriptor has the BF_LARGE flag set, so the block is
728 treated as a large object and chained onto various lists, rather
729 than the individual objects being copied. However, when it comes
730 to scavenge the block, the GC will only scavenge the first object.
731 The reason is that the GC can't linearly scan a block of pinned
732 objects at the moment (doing so would require using the
733 mostly-copying techniques). But since we're restricting ourselves
734 to pinned ByteArrays, not scavenging is ok.
736 This function is called by newPinnedByteArray# which immediately
737 fills the allocated memory with a MutableByteArray#.
738 ------------------------------------------------------------------------- */
741 allocatePinned( nat n )
744 bdescr *bd = pinned_object_block;
746 // If the request is for a large object, then allocate()
747 // will give us a pinned object anyway.
748 if (n >= LARGE_OBJECT_THRESHOLD/sizeof(W_)) {
754 TICK_ALLOC_HEAP_NOCTR(n);
757 // we always return 8-byte aligned memory. bd->free must be
758 // 8-byte aligned to begin with, so we just round up n to
759 // the nearest multiple of 8 bytes.
760 if (sizeof(StgWord) == 4) {
764 // If we don't have a block of pinned objects yet, or the current
765 // one isn't large enough to hold the new object, allocate a new one.
766 if (bd == NULL || (bd->free + n) > (bd->start + BLOCK_SIZE_W)) {
767 pinned_object_block = bd = allocBlock();
768 dbl_link_onto(bd, &g0s0->large_objects);
769 g0s0->n_large_blocks++;
772 bd->flags = BF_PINNED | BF_LARGE;
773 bd->free = bd->start;
783 /* -----------------------------------------------------------------------------
784 This is the write barrier for MUT_VARs, a.k.a. IORefs. A
785 MUT_VAR_CLEAN object is not on the mutable list; a MUT_VAR_DIRTY
786 is. When written to, a MUT_VAR_CLEAN turns into a MUT_VAR_DIRTY
787 and is put on the mutable list.
788 -------------------------------------------------------------------------- */
791 dirty_MUT_VAR(StgRegTable *reg, StgClosure *p)
793 Capability *cap = regTableToCapability(reg);
795 if (p->header.info == &stg_MUT_VAR_CLEAN_info) {
796 p->header.info = &stg_MUT_VAR_DIRTY_info;
797 bd = Bdescr((StgPtr)p);
798 if (bd->gen_no > 0) recordMutableCap(p,cap,bd->gen_no);
802 /* -----------------------------------------------------------------------------
803 Allocation functions for GMP.
805 These all use the allocate() interface - we can't have any garbage
806 collection going on during a gmp operation, so we use allocate()
807 which always succeeds. The gmp operations which might need to
808 allocate will ask the storage manager (via doYouWantToGC()) whether
809 a garbage collection is required, in case we get into a loop doing
810 only allocate() style allocation.
811 -------------------------------------------------------------------------- */
814 stgAllocForGMP (size_t size_in_bytes)
817 nat data_size_in_words, total_size_in_words;
819 /* round up to a whole number of words */
820 data_size_in_words = (size_in_bytes + sizeof(W_) + 1) / sizeof(W_);
821 total_size_in_words = sizeofW(StgArrWords) + data_size_in_words;
823 /* allocate and fill it in. */
824 #if defined(THREADED_RTS)
825 arr = (StgArrWords *)allocateLocal(myTask()->cap, total_size_in_words);
827 arr = (StgArrWords *)allocateLocal(&MainCapability, total_size_in_words);
829 SET_ARR_HDR(arr, &stg_ARR_WORDS_info, CCCS, data_size_in_words);
831 /* and return a ptr to the goods inside the array */
836 stgReallocForGMP (void *ptr, size_t old_size, size_t new_size)
838 void *new_stuff_ptr = stgAllocForGMP(new_size);
840 char *p = (char *) ptr;
841 char *q = (char *) new_stuff_ptr;
843 for (; i < old_size; i++, p++, q++) {
847 return(new_stuff_ptr);
851 stgDeallocForGMP (void *ptr STG_UNUSED,
852 size_t size STG_UNUSED)
854 /* easy for us: the garbage collector does the dealloc'n */
857 /* -----------------------------------------------------------------------------
859 * -------------------------------------------------------------------------- */
861 /* -----------------------------------------------------------------------------
864 * Approximate how much we've allocated: number of blocks in the
865 * nursery + blocks allocated via allocate() - unused nusery blocks.
866 * This leaves a little slop at the end of each block, and doesn't
867 * take into account large objects (ToDo).
868 * -------------------------------------------------------------------------- */
871 calcAllocated( void )
876 allocated = allocatedBytes();
877 allocated += countNurseryBlocks() * BLOCK_SIZE_W;
882 for (i = 0; i < n_nurseries; i++) {
884 for ( bd = capabilities[i].r.rCurrentNursery->link;
885 bd != NULL; bd = bd->link ) {
886 allocated -= BLOCK_SIZE_W;
888 cap = &capabilities[i];
889 if (cap->r.rCurrentNursery->free <
890 cap->r.rCurrentNursery->start + BLOCK_SIZE_W) {
891 allocated -= (cap->r.rCurrentNursery->start + BLOCK_SIZE_W)
892 - cap->r.rCurrentNursery->free;
896 bdescr *current_nursery = MainCapability.r.rCurrentNursery;
898 for ( bd = current_nursery->link; bd != NULL; bd = bd->link ) {
899 allocated -= BLOCK_SIZE_W;
901 if (current_nursery->free < current_nursery->start + BLOCK_SIZE_W) {
902 allocated -= (current_nursery->start + BLOCK_SIZE_W)
903 - current_nursery->free;
908 total_allocated += allocated;
912 /* Approximate the amount of live data in the heap. To be called just
913 * after garbage collection (see GarbageCollect()).
922 if (RtsFlags.GcFlags.generations == 1) {
923 return (g0s0->n_large_blocks + g0s0->n_blocks) * BLOCK_SIZE_W;
926 for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
927 for (s = 0; s < generations[g].n_steps; s++) {
928 /* approximate amount of live data (doesn't take into account slop
929 * at end of each block).
931 if (g == 0 && s == 0) {
934 stp = &generations[g].steps[s];
935 live += (stp->n_large_blocks + stp->n_blocks) * BLOCK_SIZE_W;
941 /* Approximate the number of blocks that will be needed at the next
942 * garbage collection.
944 * Assume: all data currently live will remain live. Steps that will
945 * be collected next time will therefore need twice as many blocks
946 * since all the data will be copied.
955 for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
956 for (s = 0; s < generations[g].n_steps; s++) {
957 if (g == 0 && s == 0) { continue; }
958 stp = &generations[g].steps[s];
959 if (generations[g].steps[0].n_blocks +
960 generations[g].steps[0].n_large_blocks
961 > generations[g].max_blocks
962 && stp->is_compacted == 0) {
963 needed += 2 * stp->n_blocks;
965 needed += stp->n_blocks;
972 /* ----------------------------------------------------------------------------
975 Executable memory must be managed separately from non-executable
976 memory. Most OSs these days require you to jump through hoops to
977 dynamically allocate executable memory, due to various security
980 Here we provide a small memory allocator for executable memory.
981 Memory is managed with a page granularity; we allocate linearly
982 in the page, and when the page is emptied (all objects on the page
983 are free) we free the page again, not forgetting to make it
986 TODO: The inability to handle objects bigger than BLOCK_SIZE_W means that
987 the linker cannot use allocateExec for loading object code files
988 on Windows. Once allocateExec can handle larger objects, the linker
989 should be modified to use allocateExec instead of VirtualAlloc.
990 ------------------------------------------------------------------------- */
992 static bdescr *exec_block;
994 void *allocateExec (nat bytes)
1001 // round up to words.
1002 n = (bytes + sizeof(W_) + 1) / sizeof(W_);
1004 if (n+1 > BLOCK_SIZE_W) {
1005 barf("allocateExec: can't handle large objects");
1008 if (exec_block == NULL ||
1009 exec_block->free + n + 1 > exec_block->start + BLOCK_SIZE_W) {
1011 lnat pagesize = getPageSize();
1012 bd = allocGroup(stg_max(1, pagesize / BLOCK_SIZE));
1013 debugTrace(DEBUG_gc, "allocate exec block %p", bd->start);
1015 bd->flags = BF_EXEC;
1016 bd->link = exec_block;
1017 if (exec_block != NULL) {
1018 exec_block->u.back = bd;
1021 setExecutable(bd->start, bd->blocks * BLOCK_SIZE, rtsTrue);
1024 *(exec_block->free) = n; // store the size of this chunk
1025 exec_block->gen_no += n; // gen_no stores the number of words allocated
1026 ret = exec_block->free + 1;
1027 exec_block->free += n + 1;
1033 void freeExec (void *addr)
1035 StgPtr p = (StgPtr)addr - 1;
1036 bdescr *bd = Bdescr((StgPtr)p);
1038 if ((bd->flags & BF_EXEC) == 0) {
1039 barf("freeExec: not executable");
1042 if (*(StgPtr)p == 0) {
1043 barf("freeExec: already free?");
1048 bd->gen_no -= *(StgPtr)p;
1051 if (bd->gen_no == 0) {
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 != exec_block) {
1055 debugTrace(DEBUG_gc, "free exec block %p", bd->start);
1056 dbl_link_remove(bd, &exec_block);
1057 setExecutable(bd->start, bd->blocks * BLOCK_SIZE, rtsFalse);
1060 bd->free = bd->start;
1067 /* -----------------------------------------------------------------------------
1070 memInventory() checks for memory leaks by counting up all the
1071 blocks we know about and comparing that to the number of blocks
1072 allegedly floating around in the system.
1073 -------------------------------------------------------------------------- */
1078 countBlocks(bdescr *bd)
1081 for (n=0; bd != NULL; bd=bd->link) {
1087 // (*1) Just like countBlocks, except that we adjust the count for a
1088 // megablock group so that it doesn't include the extra few blocks
1089 // that would be taken up by block descriptors in the second and
1090 // subsequent megablock. This is so we can tally the count with the
1091 // number of blocks allocated in the system, for memInventory().
1093 countAllocdBlocks(bdescr *bd)
1096 for (n=0; bd != NULL; bd=bd->link) {
1098 // hack for megablock groups: see (*1) above
1099 if (bd->blocks > BLOCKS_PER_MBLOCK) {
1100 n -= (MBLOCK_SIZE / BLOCK_SIZE - BLOCKS_PER_MBLOCK)
1101 * (bd->blocks/(MBLOCK_SIZE/BLOCK_SIZE));
1108 stepBlocks (step *stp)
1110 ASSERT(countBlocks(stp->blocks) == stp->n_blocks);
1111 ASSERT(countBlocks(stp->large_objects) == stp->n_large_blocks);
1112 return stp->n_blocks + stp->n_old_blocks +
1113 countAllocdBlocks(stp->large_objects);
1121 lnat gen_blocks[RtsFlags.GcFlags.generations];
1122 lnat nursery_blocks, allocate_blocks, retainer_blocks,
1123 arena_blocks, exec_blocks;
1124 lnat live_blocks = 0, free_blocks = 0;
1126 // count the blocks we current have
1128 for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
1130 for (i = 0; i < n_capabilities; i++) {
1131 gen_blocks[g] += countBlocks(capabilities[i].mut_lists[g]);
1133 gen_blocks[g] += countAllocdBlocks(generations[g].mut_list);
1134 for (s = 0; s < generations[g].n_steps; s++) {
1135 #if !defined(THREADED_RTS)
1136 // We put pinned object blocks in g0s0, so better count
1137 // blocks there too.
1138 if (g==0 && s==0) continue;
1140 stp = &generations[g].steps[s];
1141 gen_blocks[g] += stepBlocks(stp);
1146 for (i = 0; i < n_nurseries; i++) {
1147 nursery_blocks += stepBlocks(&nurseries[i]);
1150 /* any blocks held by allocate() */
1151 allocate_blocks = countAllocdBlocks(small_alloc_list);
1153 retainer_blocks = 0;
1155 if (RtsFlags.ProfFlags.doHeapProfile == HEAP_BY_RETAINER) {
1156 retainer_blocks = retainerStackBlocks();
1160 // count the blocks allocated by the arena allocator
1161 arena_blocks = arenaBlocks();
1163 // count the blocks containing executable memory
1164 exec_blocks = countAllocdBlocks(exec_block);
1166 /* count the blocks on the free list */
1167 free_blocks = countFreeList();
1170 for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
1171 live_blocks += gen_blocks[g];
1173 live_blocks += nursery_blocks + allocate_blocks
1174 + retainer_blocks + arena_blocks + exec_blocks;
1176 if (live_blocks + free_blocks != mblocks_allocated * BLOCKS_PER_MBLOCK)
1178 debugBelch("Memory leak detected\n");
1179 for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
1180 debugBelch(" gen %d blocks : %4lu\n", g, gen_blocks[g]);
1182 debugBelch(" nursery : %4lu\n", nursery_blocks);
1183 debugBelch(" allocate() : %4lu\n", allocate_blocks);
1184 debugBelch(" retainer : %4lu\n", retainer_blocks);
1185 debugBelch(" arena blocks : %4lu\n", arena_blocks);
1186 debugBelch(" exec : %4lu\n", exec_blocks);
1187 debugBelch(" free : %4lu\n", free_blocks);
1188 debugBelch(" total : %4lu\n\n", live_blocks + free_blocks);
1189 debugBelch(" in system : %4lu\n", mblocks_allocated * BLOCKS_PER_MBLOCK);
1195 /* Full heap sanity check. */
1201 if (RtsFlags.GcFlags.generations == 1) {
1202 checkHeap(g0s0->blocks);
1203 checkChain(g0s0->large_objects);
1206 for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
1207 for (s = 0; s < generations[g].n_steps; s++) {
1208 if (g == 0 && s == 0) { continue; }
1209 ASSERT(countBlocks(generations[g].steps[s].blocks)
1210 == generations[g].steps[s].n_blocks);
1211 ASSERT(countBlocks(generations[g].steps[s].large_objects)
1212 == generations[g].steps[s].n_large_blocks);
1213 checkHeap(generations[g].steps[s].blocks);
1214 checkChain(generations[g].steps[s].large_objects);
1216 checkMutableList(generations[g].mut_list, g);
1221 for (s = 0; s < n_nurseries; s++) {
1222 ASSERT(countBlocks(nurseries[s].blocks)
1223 == nurseries[s].n_blocks);
1224 ASSERT(countBlocks(nurseries[s].large_objects)
1225 == nurseries[s].n_large_blocks);
1228 checkFreeListSanity();
1232 /* Nursery sanity check */
1234 checkNurserySanity( step *stp )
1240 for (bd = stp->blocks; bd != NULL; bd = bd->link) {
1241 ASSERT(bd->u.back == prev);
1243 blocks += bd->blocks;
1245 ASSERT(blocks == stp->n_blocks);
1248 // handy function for use in gdb, because Bdescr() is inlined.
1249 extern bdescr *_bdescr( StgPtr p );