1 /* -----------------------------------------------------------------------------
2 * $Id: Storage.c,v 1.84 2004/08/13 13:11:01 simonmar Exp $
4 * (c) The GHC Team, 1998-1999
6 * Storage manager front end
8 * ---------------------------------------------------------------------------*/
10 #include "PosixSource.h"
16 #include "BlockAlloc.h"
24 #include "OSThreads.h"
26 #include "RetainerProfile.h" // for counting memory blocks (memInventory)
31 StgClosure *caf_list = NULL;
33 bdescr *small_alloc_list; /* allocate()d small objects */
34 bdescr *pinned_object_block; /* allocate pinned objects into this block */
35 nat alloc_blocks; /* number of allocate()d blocks since GC */
36 nat alloc_blocks_lim; /* approximate limit on alloc_blocks */
38 StgPtr alloc_Hp = NULL; /* next free byte in small_alloc_list */
39 StgPtr alloc_HpLim = NULL; /* end of block at small_alloc_list */
41 generation *generations = NULL; /* all the generations */
42 generation *g0 = NULL; /* generation 0, for convenience */
43 generation *oldest_gen = NULL; /* oldest generation, for convenience */
44 step *g0s0 = NULL; /* generation 0, step 0, for convenience */
46 ullong total_allocated = 0; /* total memory allocated during run */
49 * Storage manager mutex: protects all the above state from
50 * simultaneous access by two STG threads.
53 Mutex sm_mutex = INIT_MUTEX_VAR;
59 static void *stgAllocForGMP (size_t size_in_bytes);
60 static void *stgReallocForGMP (void *ptr, size_t old_size, size_t new_size);
61 static void stgDeallocForGMP (void *ptr, size_t size);
70 if (generations != NULL) {
71 // multi-init protection
75 /* Sanity check to make sure the LOOKS_LIKE_ macros appear to be
76 * doing something reasonable.
78 ASSERT(LOOKS_LIKE_INFO_PTR(&stg_BLACKHOLE_info));
79 ASSERT(LOOKS_LIKE_CLOSURE_PTR(&stg_dummy_ret_closure));
80 ASSERT(!HEAP_ALLOCED(&stg_dummy_ret_closure));
82 if (RtsFlags.GcFlags.maxHeapSize != 0 &&
83 RtsFlags.GcFlags.heapSizeSuggestion >
84 RtsFlags.GcFlags.maxHeapSize) {
85 RtsFlags.GcFlags.maxHeapSize = RtsFlags.GcFlags.heapSizeSuggestion;
88 if (RtsFlags.GcFlags.maxHeapSize != 0 &&
89 RtsFlags.GcFlags.minAllocAreaSize >
90 RtsFlags.GcFlags.maxHeapSize) {
91 prog_belch("maximum heap size (-M) is smaller than minimum alloc area size (-A)");
101 /* allocate generation info array */
102 generations = (generation *)stgMallocBytes(RtsFlags.GcFlags.generations
103 * sizeof(struct _generation),
104 "initStorage: gens");
106 /* Initialise all generations */
107 for(g = 0; g < RtsFlags.GcFlags.generations; g++) {
108 gen = &generations[g];
110 gen->mut_list = END_MUT_LIST;
111 gen->mut_once_list = END_MUT_LIST;
112 gen->collections = 0;
113 gen->failed_promotions = 0;
117 /* A couple of convenience pointers */
118 g0 = &generations[0];
119 oldest_gen = &generations[RtsFlags.GcFlags.generations-1];
121 /* Allocate step structures in each generation */
122 if (RtsFlags.GcFlags.generations > 1) {
123 /* Only for multiple-generations */
125 /* Oldest generation: one step */
126 oldest_gen->n_steps = 1;
128 stgMallocBytes(1 * sizeof(struct _step), "initStorage: last step");
130 /* set up all except the oldest generation with 2 steps */
131 for(g = 0; g < RtsFlags.GcFlags.generations-1; g++) {
132 generations[g].n_steps = RtsFlags.GcFlags.steps;
133 generations[g].steps =
134 stgMallocBytes (RtsFlags.GcFlags.steps * sizeof(struct _step),
135 "initStorage: steps");
139 /* single generation, i.e. a two-space collector */
141 g0->steps = stgMallocBytes (sizeof(struct _step), "initStorage: steps");
144 /* Initialise all steps */
145 for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
146 for (s = 0; s < generations[g].n_steps; s++) {
147 stp = &generations[g].steps[s];
150 stp->n_to_blocks = 0;
152 stp->gen = &generations[g];
159 stp->large_objects = NULL;
160 stp->n_large_blocks = 0;
161 stp->new_large_objects = NULL;
162 stp->scavenged_large_objects = NULL;
163 stp->n_scavenged_large_blocks = 0;
164 stp->is_compacted = 0;
169 /* Set up the destination pointers in each younger gen. step */
170 for (g = 0; g < RtsFlags.GcFlags.generations-1; g++) {
171 for (s = 0; s < generations[g].n_steps-1; s++) {
172 generations[g].steps[s].to = &generations[g].steps[s+1];
174 generations[g].steps[s].to = &generations[g+1].steps[0];
177 /* The oldest generation has one step and it is compacted. */
178 if (RtsFlags.GcFlags.compact) {
179 if (RtsFlags.GcFlags.generations == 1) {
180 belch("WARNING: compaction is incompatible with -G1; disabled");
182 oldest_gen->steps[0].is_compacted = 1;
185 oldest_gen->steps[0].to = &oldest_gen->steps[0];
187 /* generation 0 is special: that's the nursery */
188 generations[0].max_blocks = 0;
190 /* G0S0: the allocation area. Policy: keep the allocation area
191 * small to begin with, even if we have a large suggested heap
192 * size. Reason: we're going to do a major collection first, and we
193 * don't want it to be a big one. This vague idea is borne out by
194 * rigorous experimental evidence.
196 g0s0 = &generations[0].steps[0];
200 weak_ptr_list = NULL;
203 /* initialise the allocate() interface */
204 small_alloc_list = NULL;
206 alloc_blocks_lim = RtsFlags.GcFlags.minAllocAreaSize;
208 /* Tell GNU multi-precision pkg about our custom alloc functions */
209 mp_set_memory_functions(stgAllocForGMP, stgReallocForGMP, stgDeallocForGMP);
211 IF_DEBUG(gc, statDescribeGens());
217 stat_exit(calcAllocated());
220 /* -----------------------------------------------------------------------------
223 The entry code for every CAF does the following:
225 - builds a CAF_BLACKHOLE in the heap
226 - pushes an update frame pointing to the CAF_BLACKHOLE
227 - invokes UPD_CAF(), which:
228 - calls newCaf, below
229 - updates the CAF with a static indirection to the CAF_BLACKHOLE
231 Why do we build a BLACKHOLE in the heap rather than just updating
232 the thunk directly? It's so that we only need one kind of update
233 frame - otherwise we'd need a static version of the update frame too.
235 newCaf() does the following:
237 - it puts the CAF on the oldest generation's mut-once list.
238 This is so that we can treat the CAF as a root when collecting
241 For GHCI, we have additional requirements when dealing with CAFs:
243 - we must *retain* all dynamically-loaded CAFs ever entered,
244 just in case we need them again.
245 - we must be able to *revert* CAFs that have been evaluated, to
246 their pre-evaluated form.
248 To do this, we use an additional CAF list. When newCaf() is
249 called on a dynamically-loaded CAF, we add it to the CAF list
250 instead of the old-generation mutable list, and save away its
251 old info pointer (in caf->saved_info) for later reversion.
253 To revert all the CAFs, we traverse the CAF list and reset the
254 info pointer to caf->saved_info, then throw away the CAF list.
255 (see GC.c:revertCAFs()).
259 -------------------------------------------------------------------------- */
262 newCAF(StgClosure* caf)
264 /* Put this CAF on the mutable list for the old generation.
265 * This is a HACK - the IND_STATIC closure doesn't really have
266 * a mut_link field, but we pretend it has - in fact we re-use
267 * the STATIC_LINK field for the time being, because when we
268 * come to do a major GC we won't need the mut_link field
269 * any more and can use it as a STATIC_LINK.
273 ((StgIndStatic *)caf)->saved_info = NULL;
274 ((StgMutClosure *)caf)->mut_link = oldest_gen->mut_once_list;
275 oldest_gen->mut_once_list = (StgMutClosure *)caf;
280 /* If we are PAR or DIST then we never forget a CAF */
282 //belch("<##> Globalising CAF %08x %s",caf,info_type(caf));
283 newGA=makeGlobal(caf,rtsTrue); /*given full weight*/
289 // An alternate version of newCaf which is used for dynamically loaded
290 // object code in GHCi. In this case we want to retain *all* CAFs in
291 // the object code, because they might be demanded at any time from an
292 // expression evaluated on the command line.
294 // The linker hackily arranges that references to newCaf from dynamic
295 // code end up pointing to newDynCAF.
297 newDynCAF(StgClosure *caf)
301 ((StgIndStatic *)caf)->saved_info = (StgInfoTable *)caf->header.info;
302 ((StgIndStatic *)caf)->static_link = caf_list;
308 /* -----------------------------------------------------------------------------
310 -------------------------------------------------------------------------- */
313 allocNurseries( void )
321 for (cap = free_capabilities; cap != NULL; cap = cap->link) {
322 cap->r.rNursery = allocNursery(NULL, RtsFlags.GcFlags.minAllocAreaSize);
323 cap->r.rCurrentNursery = cap->r.rNursery;
324 /* Set the back links to be equal to the Capability,
325 * so we can do slightly better informed locking.
327 for (bd = cap->r.rNursery; bd != NULL; bd = bd->link) {
328 bd->u.back = (bdescr *)cap;
332 g0s0->blocks = allocNursery(NULL, RtsFlags.GcFlags.minAllocAreaSize);
333 g0s0->n_blocks = RtsFlags.GcFlags.minAllocAreaSize;
334 g0s0->to_blocks = NULL;
335 g0s0->n_to_blocks = 0;
336 MainCapability.r.rNursery = g0s0->blocks;
337 MainCapability.r.rCurrentNursery = g0s0->blocks;
338 /* hp, hpLim, hp_bd, to_space etc. aren't used in G0S0 */
343 resetNurseries( void )
349 /* All tasks must be stopped */
350 ASSERT(n_free_capabilities == RtsFlags.ParFlags.nNodes);
352 for (cap = free_capabilities; cap != NULL; cap = cap->link) {
353 for (bd = cap->r.rNursery; bd; bd = bd->link) {
354 bd->free = bd->start;
355 ASSERT(bd->gen_no == 0);
356 ASSERT(bd->step == g0s0);
357 IF_DEBUG(sanity,memset(bd->start, 0xaa, BLOCK_SIZE));
359 cap->r.rCurrentNursery = cap->r.rNursery;
362 for (bd = g0s0->blocks; bd; bd = bd->link) {
363 bd->free = bd->start;
364 ASSERT(bd->gen_no == 0);
365 ASSERT(bd->step == g0s0);
366 IF_DEBUG(sanity,memset(bd->start, 0xaa, BLOCK_SIZE));
368 MainCapability.r.rNursery = g0s0->blocks;
369 MainCapability.r.rCurrentNursery = g0s0->blocks;
374 allocNursery (bdescr *tail, nat blocks)
379 // Allocate a nursery: we allocate fresh blocks one at a time and
380 // cons them on to the front of the list, not forgetting to update
381 // the back pointer on the tail of the list to point to the new block.
382 for (i=0; i < blocks; i++) {
385 processNursery() in LdvProfile.c assumes that every block group in
386 the nursery contains only a single block. So, if a block group is
387 given multiple blocks, change processNursery() accordingly.
391 // double-link the nursery: we might need to insert blocks
398 bd->free = bd->start;
406 resizeNursery ( nat blocks )
412 barf("resizeNursery: can't resize in SMP mode");
415 nursery_blocks = g0s0->n_blocks;
416 if (nursery_blocks == blocks) {
420 else if (nursery_blocks < blocks) {
421 IF_DEBUG(gc, fprintf(stderr, "Increasing size of nursery to %d blocks\n",
423 g0s0->blocks = allocNursery(g0s0->blocks, blocks-nursery_blocks);
429 IF_DEBUG(gc, fprintf(stderr, "Decreasing size of nursery to %d blocks\n",
433 while (nursery_blocks > blocks) {
435 next_bd->u.back = NULL;
436 nursery_blocks -= bd->blocks; // might be a large block
441 // might have gone just under, by freeing a large block, so make
442 // up the difference.
443 if (nursery_blocks < blocks) {
444 g0s0->blocks = allocNursery(g0s0->blocks, blocks-nursery_blocks);
448 g0s0->n_blocks = blocks;
449 ASSERT(countBlocks(g0s0->blocks) == g0s0->n_blocks);
452 /* -----------------------------------------------------------------------------
453 The allocate() interface
455 allocate(n) always succeeds, and returns a chunk of memory n words
456 long. n can be larger than the size of a block if necessary, in
457 which case a contiguous block group will be allocated.
458 -------------------------------------------------------------------------- */
468 TICK_ALLOC_HEAP_NOCTR(n);
471 /* big allocation (>LARGE_OBJECT_THRESHOLD) */
472 /* ToDo: allocate directly into generation 1 */
473 if (n >= LARGE_OBJECT_THRESHOLD/sizeof(W_)) {
474 nat req_blocks = (lnat)BLOCK_ROUND_UP(n*sizeof(W_)) / BLOCK_SIZE;
475 bd = allocGroup(req_blocks);
476 dbl_link_onto(bd, &g0s0->large_objects);
477 g0s0->n_large_blocks += req_blocks;
480 bd->flags = BF_LARGE;
481 bd->free = bd->start + n;
482 alloc_blocks += req_blocks;
486 /* small allocation (<LARGE_OBJECT_THRESHOLD) */
487 } else if (small_alloc_list == NULL || alloc_Hp + n > alloc_HpLim) {
488 if (small_alloc_list) {
489 small_alloc_list->free = alloc_Hp;
492 bd->link = small_alloc_list;
493 small_alloc_list = bd;
497 alloc_Hp = bd->start;
498 alloc_HpLim = bd->start + BLOCK_SIZE_W;
509 allocated_bytes( void )
513 allocated = alloc_blocks * BLOCK_SIZE_W - (alloc_HpLim - alloc_Hp);
514 if (pinned_object_block != NULL) {
515 allocated -= (pinned_object_block->start + BLOCK_SIZE_W) -
516 pinned_object_block->free;
523 tidyAllocateLists (void)
525 if (small_alloc_list != NULL) {
526 ASSERT(alloc_Hp >= small_alloc_list->start &&
527 alloc_Hp <= small_alloc_list->start + BLOCK_SIZE);
528 small_alloc_list->free = alloc_Hp;
532 /* ---------------------------------------------------------------------------
533 Allocate a fixed/pinned object.
535 We allocate small pinned objects into a single block, allocating a
536 new block when the current one overflows. The block is chained
537 onto the large_object_list of generation 0 step 0.
539 NOTE: The GC can't in general handle pinned objects. This
540 interface is only safe to use for ByteArrays, which have no
541 pointers and don't require scavenging. It works because the
542 block's descriptor has the BF_LARGE flag set, so the block is
543 treated as a large object and chained onto various lists, rather
544 than the individual objects being copied. However, when it comes
545 to scavenge the block, the GC will only scavenge the first object.
546 The reason is that the GC can't linearly scan a block of pinned
547 objects at the moment (doing so would require using the
548 mostly-copying techniques). But since we're restricting ourselves
549 to pinned ByteArrays, not scavenging is ok.
551 This function is called by newPinnedByteArray# which immediately
552 fills the allocated memory with a MutableByteArray#.
553 ------------------------------------------------------------------------- */
556 allocatePinned( nat n )
559 bdescr *bd = pinned_object_block;
561 // If the request is for a large object, then allocate()
562 // will give us a pinned object anyway.
563 if (n >= LARGE_OBJECT_THRESHOLD/sizeof(W_)) {
569 TICK_ALLOC_HEAP_NOCTR(n);
572 // we always return 8-byte aligned memory. bd->free must be
573 // 8-byte aligned to begin with, so we just round up n to
574 // the nearest multiple of 8 bytes.
575 if (sizeof(StgWord) == 4) {
579 // If we don't have a block of pinned objects yet, or the current
580 // one isn't large enough to hold the new object, allocate a new one.
581 if (bd == NULL || (bd->free + n) > (bd->start + BLOCK_SIZE_W)) {
582 pinned_object_block = bd = allocBlock();
583 dbl_link_onto(bd, &g0s0->large_objects);
586 bd->flags = BF_PINNED | BF_LARGE;
587 bd->free = bd->start;
597 /* -----------------------------------------------------------------------------
598 Allocation functions for GMP.
600 These all use the allocate() interface - we can't have any garbage
601 collection going on during a gmp operation, so we use allocate()
602 which always succeeds. The gmp operations which might need to
603 allocate will ask the storage manager (via doYouWantToGC()) whether
604 a garbage collection is required, in case we get into a loop doing
605 only allocate() style allocation.
606 -------------------------------------------------------------------------- */
609 stgAllocForGMP (size_t size_in_bytes)
612 nat data_size_in_words, total_size_in_words;
614 /* round up to a whole number of words */
615 data_size_in_words = (size_in_bytes + sizeof(W_) + 1) / sizeof(W_);
616 total_size_in_words = sizeofW(StgArrWords) + data_size_in_words;
618 /* allocate and fill it in. */
619 arr = (StgArrWords *)allocate(total_size_in_words);
620 SET_ARR_HDR(arr, &stg_ARR_WORDS_info, CCCS, data_size_in_words);
622 /* and return a ptr to the goods inside the array */
627 stgReallocForGMP (void *ptr, size_t old_size, size_t new_size)
629 void *new_stuff_ptr = stgAllocForGMP(new_size);
631 char *p = (char *) ptr;
632 char *q = (char *) new_stuff_ptr;
634 for (; i < old_size; i++, p++, q++) {
638 return(new_stuff_ptr);
642 stgDeallocForGMP (void *ptr STG_UNUSED,
643 size_t size STG_UNUSED)
645 /* easy for us: the garbage collector does the dealloc'n */
648 /* -----------------------------------------------------------------------------
650 * -------------------------------------------------------------------------- */
652 /* -----------------------------------------------------------------------------
655 * Approximate how much we've allocated: number of blocks in the
656 * nursery + blocks allocated via allocate() - unused nusery blocks.
657 * This leaves a little slop at the end of each block, and doesn't
658 * take into account large objects (ToDo).
659 * -------------------------------------------------------------------------- */
662 calcAllocated( void )
670 /* All tasks must be stopped. Can't assert that all the
671 capabilities are owned by the scheduler, though: one or more
672 tasks might have been stopped while they were running (non-main)
674 /* ASSERT(n_free_capabilities == RtsFlags.ParFlags.nNodes); */
677 n_free_capabilities * RtsFlags.GcFlags.minAllocAreaSize * BLOCK_SIZE_W
680 for (cap = free_capabilities; cap != NULL; cap = cap->link) {
681 for ( bd = cap->r.rCurrentNursery->link; bd != NULL; bd = bd->link ) {
682 allocated -= BLOCK_SIZE_W;
684 if (cap->r.rCurrentNursery->free < cap->r.rCurrentNursery->start
686 allocated -= (cap->r.rCurrentNursery->start + BLOCK_SIZE_W)
687 - cap->r.rCurrentNursery->free;
692 bdescr *current_nursery = MainCapability.r.rCurrentNursery;
694 allocated = (g0s0->n_blocks * BLOCK_SIZE_W) + allocated_bytes();
695 for ( bd = current_nursery->link; bd != NULL; bd = bd->link ) {
696 allocated -= BLOCK_SIZE_W;
698 if (current_nursery->free < current_nursery->start + BLOCK_SIZE_W) {
699 allocated -= (current_nursery->start + BLOCK_SIZE_W)
700 - current_nursery->free;
704 total_allocated += allocated;
708 /* Approximate the amount of live data in the heap. To be called just
709 * after garbage collection (see GarbageCollect()).
718 if (RtsFlags.GcFlags.generations == 1) {
719 live = (g0s0->n_to_blocks - 1) * BLOCK_SIZE_W +
720 ((lnat)g0s0->hp_bd->free - (lnat)g0s0->hp_bd->start) / sizeof(W_);
724 for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
725 for (s = 0; s < generations[g].n_steps; s++) {
726 /* approximate amount of live data (doesn't take into account slop
727 * at end of each block).
729 if (g == 0 && s == 0) {
732 stp = &generations[g].steps[s];
733 live += (stp->n_large_blocks + stp->n_blocks - 1) * BLOCK_SIZE_W;
734 if (stp->hp_bd != NULL) {
735 live += ((lnat)stp->hp_bd->free - (lnat)stp->hp_bd->start)
743 /* Approximate the number of blocks that will be needed at the next
744 * garbage collection.
746 * Assume: all data currently live will remain live. Steps that will
747 * be collected next time will therefore need twice as many blocks
748 * since all the data will be copied.
757 for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
758 for (s = 0; s < generations[g].n_steps; s++) {
759 if (g == 0 && s == 0) { continue; }
760 stp = &generations[g].steps[s];
761 if (generations[g].steps[0].n_blocks +
762 generations[g].steps[0].n_large_blocks
763 > generations[g].max_blocks
764 && stp->is_compacted == 0) {
765 needed += 2 * stp->n_blocks;
767 needed += stp->n_blocks;
774 /* -----------------------------------------------------------------------------
777 memInventory() checks for memory leaks by counting up all the
778 blocks we know about and comparing that to the number of blocks
779 allegedly floating around in the system.
780 -------------------------------------------------------------------------- */
790 lnat total_blocks = 0, free_blocks = 0;
792 /* count the blocks we current have */
794 for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
795 for (s = 0; s < generations[g].n_steps; s++) {
796 stp = &generations[g].steps[s];
797 total_blocks += stp->n_blocks;
798 if (RtsFlags.GcFlags.generations == 1) {
799 /* two-space collector has a to-space too :-) */
800 total_blocks += g0s0->n_to_blocks;
802 for (bd = stp->large_objects; bd; bd = bd->link) {
803 total_blocks += bd->blocks;
804 /* hack for megablock groups: they have an extra block or two in
805 the second and subsequent megablocks where the block
806 descriptors would normally go.
808 if (bd->blocks > BLOCKS_PER_MBLOCK) {
809 total_blocks -= (MBLOCK_SIZE / BLOCK_SIZE - BLOCKS_PER_MBLOCK)
810 * (bd->blocks/(MBLOCK_SIZE/BLOCK_SIZE));
816 /* any blocks held by allocate() */
817 for (bd = small_alloc_list; bd; bd = bd->link) {
818 total_blocks += bd->blocks;
822 if (RtsFlags.ProfFlags.doHeapProfile == HEAP_BY_RETAINER) {
823 total_blocks += retainerStackBlocks();
827 // count the blocks allocated by the arena allocator
828 total_blocks += arenaBlocks();
830 /* count the blocks on the free list */
831 free_blocks = countFreeList();
833 if (total_blocks + free_blocks != mblocks_allocated *
835 fprintf(stderr, "Blocks: %ld live + %ld free = %ld total (%ld around)\n",
836 total_blocks, free_blocks, total_blocks + free_blocks,
837 mblocks_allocated * BLOCKS_PER_MBLOCK);
840 ASSERT(total_blocks + free_blocks == mblocks_allocated * BLOCKS_PER_MBLOCK);
845 countBlocks(bdescr *bd)
848 for (n=0; bd != NULL; bd=bd->link) {
854 /* Full heap sanity check. */
860 if (RtsFlags.GcFlags.generations == 1) {
861 checkHeap(g0s0->to_blocks);
862 checkChain(g0s0->large_objects);
865 for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
866 for (s = 0; s < generations[g].n_steps; s++) {
867 ASSERT(countBlocks(generations[g].steps[s].blocks)
868 == generations[g].steps[s].n_blocks);
869 ASSERT(countBlocks(generations[g].steps[s].large_objects)
870 == generations[g].steps[s].n_large_blocks);
871 if (g == 0 && s == 0) { continue; }
872 checkHeap(generations[g].steps[s].blocks);
873 checkChain(generations[g].steps[s].large_objects);
875 checkMutableList(generations[g].mut_list, g);
876 checkMutOnceList(generations[g].mut_once_list, g);
880 checkFreeListSanity();
884 // handy function for use in gdb, because Bdescr() is inlined.
885 extern bdescr *_bdescr( StgPtr p );