1 /* -----------------------------------------------------------------------------
2 * $Id: Storage.c,v 1.83 2004/07/21 10:47:28 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"
25 #include "StoragePriv.h"
27 #include "RetainerProfile.h" // for counting memory blocks (memInventory)
32 StgClosure *caf_list = NULL;
34 bdescr *small_alloc_list; /* allocate()d small objects */
35 bdescr *pinned_object_block; /* allocate pinned objects into this block */
36 nat alloc_blocks; /* number of allocate()d blocks since GC */
37 nat alloc_blocks_lim; /* approximate limit on alloc_blocks */
39 StgPtr alloc_Hp = NULL; /* next free byte in small_alloc_list */
40 StgPtr alloc_HpLim = NULL; /* end of block at small_alloc_list */
42 generation *generations = NULL; /* all the generations */
43 generation *g0 = NULL; /* generation 0, for convenience */
44 generation *oldest_gen = NULL; /* oldest generation, for convenience */
45 step *g0s0 = NULL; /* generation 0, step 0, for convenience */
47 ullong total_allocated = 0; /* total memory allocated during run */
50 * Storage manager mutex: protects all the above state from
51 * simultaneous access by two STG threads.
54 Mutex sm_mutex = INIT_MUTEX_VAR;
60 static void *stgAllocForGMP (size_t size_in_bytes);
61 static void *stgReallocForGMP (void *ptr, size_t old_size, size_t new_size);
62 static void stgDeallocForGMP (void *ptr, size_t size);
71 if (generations != NULL) {
72 // multi-init protection
76 /* Sanity check to make sure the LOOKS_LIKE_ macros appear to be
77 * doing something reasonable.
79 ASSERT(LOOKS_LIKE_INFO_PTR(&stg_BLACKHOLE_info));
80 ASSERT(LOOKS_LIKE_CLOSURE_PTR(&stg_dummy_ret_closure));
81 ASSERT(!HEAP_ALLOCED(&stg_dummy_ret_closure));
83 if (RtsFlags.GcFlags.maxHeapSize != 0 &&
84 RtsFlags.GcFlags.heapSizeSuggestion >
85 RtsFlags.GcFlags.maxHeapSize) {
86 RtsFlags.GcFlags.maxHeapSize = RtsFlags.GcFlags.heapSizeSuggestion;
89 if (RtsFlags.GcFlags.maxHeapSize != 0 &&
90 RtsFlags.GcFlags.minAllocAreaSize >
91 RtsFlags.GcFlags.maxHeapSize) {
92 prog_belch("maximum heap size (-M) is smaller than minimum alloc area size (-A)");
102 /* allocate generation info array */
103 generations = (generation *)stgMallocBytes(RtsFlags.GcFlags.generations
104 * sizeof(struct _generation),
105 "initStorage: gens");
107 /* Initialise all generations */
108 for(g = 0; g < RtsFlags.GcFlags.generations; g++) {
109 gen = &generations[g];
111 gen->mut_list = END_MUT_LIST;
112 gen->mut_once_list = END_MUT_LIST;
113 gen->collections = 0;
114 gen->failed_promotions = 0;
118 /* A couple of convenience pointers */
119 g0 = &generations[0];
120 oldest_gen = &generations[RtsFlags.GcFlags.generations-1];
122 /* Allocate step structures in each generation */
123 if (RtsFlags.GcFlags.generations > 1) {
124 /* Only for multiple-generations */
126 /* Oldest generation: one step */
127 oldest_gen->n_steps = 1;
129 stgMallocBytes(1 * sizeof(struct _step), "initStorage: last step");
131 /* set up all except the oldest generation with 2 steps */
132 for(g = 0; g < RtsFlags.GcFlags.generations-1; g++) {
133 generations[g].n_steps = RtsFlags.GcFlags.steps;
134 generations[g].steps =
135 stgMallocBytes (RtsFlags.GcFlags.steps * sizeof(struct _step),
136 "initStorage: steps");
140 /* single generation, i.e. a two-space collector */
142 g0->steps = stgMallocBytes (sizeof(struct _step), "initStorage: steps");
145 /* Initialise all steps */
146 for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
147 for (s = 0; s < generations[g].n_steps; s++) {
148 stp = &generations[g].steps[s];
151 stp->n_to_blocks = 0;
153 stp->gen = &generations[g];
160 stp->large_objects = NULL;
161 stp->n_large_blocks = 0;
162 stp->new_large_objects = NULL;
163 stp->scavenged_large_objects = NULL;
164 stp->n_scavenged_large_blocks = 0;
165 stp->is_compacted = 0;
170 /* Set up the destination pointers in each younger gen. step */
171 for (g = 0; g < RtsFlags.GcFlags.generations-1; g++) {
172 for (s = 0; s < generations[g].n_steps-1; s++) {
173 generations[g].steps[s].to = &generations[g].steps[s+1];
175 generations[g].steps[s].to = &generations[g+1].steps[0];
178 /* The oldest generation has one step and it is compacted. */
179 if (RtsFlags.GcFlags.compact) {
180 if (RtsFlags.GcFlags.generations == 1) {
181 belch("WARNING: compaction is incompatible with -G1; disabled");
183 oldest_gen->steps[0].is_compacted = 1;
186 oldest_gen->steps[0].to = &oldest_gen->steps[0];
188 /* generation 0 is special: that's the nursery */
189 generations[0].max_blocks = 0;
191 /* G0S0: the allocation area. Policy: keep the allocation area
192 * small to begin with, even if we have a large suggested heap
193 * size. Reason: we're going to do a major collection first, and we
194 * don't want it to be a big one. This vague idea is borne out by
195 * rigorous experimental evidence.
197 g0s0 = &generations[0].steps[0];
201 weak_ptr_list = NULL;
204 /* initialise the allocate() interface */
205 small_alloc_list = NULL;
207 alloc_blocks_lim = RtsFlags.GcFlags.minAllocAreaSize;
209 /* Tell GNU multi-precision pkg about our custom alloc functions */
210 mp_set_memory_functions(stgAllocForGMP, stgReallocForGMP, stgDeallocForGMP);
212 IF_DEBUG(gc, statDescribeGens());
218 stat_exit(calcAllocated());
221 /* -----------------------------------------------------------------------------
224 The entry code for every CAF does the following:
226 - builds a CAF_BLACKHOLE in the heap
227 - pushes an update frame pointing to the CAF_BLACKHOLE
228 - invokes UPD_CAF(), which:
229 - calls newCaf, below
230 - updates the CAF with a static indirection to the CAF_BLACKHOLE
232 Why do we build a BLACKHOLE in the heap rather than just updating
233 the thunk directly? It's so that we only need one kind of update
234 frame - otherwise we'd need a static version of the update frame too.
236 newCaf() does the following:
238 - it puts the CAF on the oldest generation's mut-once list.
239 This is so that we can treat the CAF as a root when collecting
242 For GHCI, we have additional requirements when dealing with CAFs:
244 - we must *retain* all dynamically-loaded CAFs ever entered,
245 just in case we need them again.
246 - we must be able to *revert* CAFs that have been evaluated, to
247 their pre-evaluated form.
249 To do this, we use an additional CAF list. When newCaf() is
250 called on a dynamically-loaded CAF, we add it to the CAF list
251 instead of the old-generation mutable list, and save away its
252 old info pointer (in caf->saved_info) for later reversion.
254 To revert all the CAFs, we traverse the CAF list and reset the
255 info pointer to caf->saved_info, then throw away the CAF list.
256 (see GC.c:revertCAFs()).
260 -------------------------------------------------------------------------- */
263 newCAF(StgClosure* caf)
265 /* Put this CAF on the mutable list for the old generation.
266 * This is a HACK - the IND_STATIC closure doesn't really have
267 * a mut_link field, but we pretend it has - in fact we re-use
268 * the STATIC_LINK field for the time being, because when we
269 * come to do a major GC we won't need the mut_link field
270 * any more and can use it as a STATIC_LINK.
274 ((StgIndStatic *)caf)->saved_info = NULL;
275 ((StgMutClosure *)caf)->mut_link = oldest_gen->mut_once_list;
276 oldest_gen->mut_once_list = (StgMutClosure *)caf;
281 /* If we are PAR or DIST then we never forget a CAF */
283 //belch("<##> Globalising CAF %08x %s",caf,info_type(caf));
284 newGA=makeGlobal(caf,rtsTrue); /*given full weight*/
290 // An alternate version of newCaf which is used for dynamically loaded
291 // object code in GHCi. In this case we want to retain *all* CAFs in
292 // the object code, because they might be demanded at any time from an
293 // expression evaluated on the command line.
295 // The linker hackily arranges that references to newCaf from dynamic
296 // code end up pointing to newDynCAF.
298 newDynCAF(StgClosure *caf)
302 ((StgIndStatic *)caf)->saved_info = (StgInfoTable *)caf->header.info;
303 ((StgIndStatic *)caf)->static_link = caf_list;
309 /* -----------------------------------------------------------------------------
311 -------------------------------------------------------------------------- */
314 allocNurseries( void )
322 for (cap = free_capabilities; cap != NULL; cap = cap->link) {
323 cap->r.rNursery = allocNursery(NULL, RtsFlags.GcFlags.minAllocAreaSize);
324 cap->r.rCurrentNursery = cap->r.rNursery;
325 /* Set the back links to be equal to the Capability,
326 * so we can do slightly better informed locking.
328 for (bd = cap->r.rNursery; bd != NULL; bd = bd->link) {
329 bd->u.back = (bdescr *)cap;
333 g0s0->blocks = allocNursery(NULL, RtsFlags.GcFlags.minAllocAreaSize);
334 g0s0->n_blocks = RtsFlags.GcFlags.minAllocAreaSize;
335 g0s0->to_blocks = NULL;
336 g0s0->n_to_blocks = 0;
337 MainCapability.r.rNursery = g0s0->blocks;
338 MainCapability.r.rCurrentNursery = g0s0->blocks;
339 /* hp, hpLim, hp_bd, to_space etc. aren't used in G0S0 */
344 resetNurseries( void )
350 /* All tasks must be stopped */
351 ASSERT(n_free_capabilities == RtsFlags.ParFlags.nNodes);
353 for (cap = free_capabilities; cap != NULL; cap = cap->link) {
354 for (bd = cap->r.rNursery; bd; bd = bd->link) {
355 bd->free = bd->start;
356 ASSERT(bd->gen_no == 0);
357 ASSERT(bd->step == g0s0);
358 IF_DEBUG(sanity,memset(bd->start, 0xaa, BLOCK_SIZE));
360 cap->r.rCurrentNursery = cap->r.rNursery;
363 for (bd = g0s0->blocks; bd; bd = bd->link) {
364 bd->free = bd->start;
365 ASSERT(bd->gen_no == 0);
366 ASSERT(bd->step == g0s0);
367 IF_DEBUG(sanity,memset(bd->start, 0xaa, BLOCK_SIZE));
369 MainCapability.r.rNursery = g0s0->blocks;
370 MainCapability.r.rCurrentNursery = g0s0->blocks;
375 allocNursery (bdescr *tail, nat blocks)
380 // Allocate a nursery: we allocate fresh blocks one at a time and
381 // cons them on to the front of the list, not forgetting to update
382 // the back pointer on the tail of the list to point to the new block.
383 for (i=0; i < blocks; i++) {
386 processNursery() in LdvProfile.c assumes that every block group in
387 the nursery contains only a single block. So, if a block group is
388 given multiple blocks, change processNursery() accordingly.
392 // double-link the nursery: we might need to insert blocks
399 bd->free = bd->start;
407 resizeNursery ( nat blocks )
413 barf("resizeNursery: can't resize in SMP mode");
416 nursery_blocks = g0s0->n_blocks;
417 if (nursery_blocks == blocks) {
421 else if (nursery_blocks < blocks) {
422 IF_DEBUG(gc, fprintf(stderr, "Increasing size of nursery to %d blocks\n",
424 g0s0->blocks = allocNursery(g0s0->blocks, blocks-nursery_blocks);
430 IF_DEBUG(gc, fprintf(stderr, "Decreasing size of nursery to %d blocks\n",
434 while (nursery_blocks > blocks) {
436 next_bd->u.back = NULL;
437 nursery_blocks -= bd->blocks; // might be a large block
442 // might have gone just under, by freeing a large block, so make
443 // up the difference.
444 if (nursery_blocks < blocks) {
445 g0s0->blocks = allocNursery(g0s0->blocks, blocks-nursery_blocks);
449 g0s0->n_blocks = blocks;
450 ASSERT(countBlocks(g0s0->blocks) == g0s0->n_blocks);
453 /* -----------------------------------------------------------------------------
454 The allocate() interface
456 allocate(n) always succeeds, and returns a chunk of memory n words
457 long. n can be larger than the size of a block if necessary, in
458 which case a contiguous block group will be allocated.
459 -------------------------------------------------------------------------- */
469 TICK_ALLOC_HEAP_NOCTR(n);
472 /* big allocation (>LARGE_OBJECT_THRESHOLD) */
473 /* ToDo: allocate directly into generation 1 */
474 if (n >= LARGE_OBJECT_THRESHOLD/sizeof(W_)) {
475 nat req_blocks = (lnat)BLOCK_ROUND_UP(n*sizeof(W_)) / BLOCK_SIZE;
476 bd = allocGroup(req_blocks);
477 dbl_link_onto(bd, &g0s0->large_objects);
478 g0s0->n_large_blocks += req_blocks;
481 bd->flags = BF_LARGE;
482 bd->free = bd->start + n;
483 alloc_blocks += req_blocks;
487 /* small allocation (<LARGE_OBJECT_THRESHOLD) */
488 } else if (small_alloc_list == NULL || alloc_Hp + n > alloc_HpLim) {
489 if (small_alloc_list) {
490 small_alloc_list->free = alloc_Hp;
493 bd->link = small_alloc_list;
494 small_alloc_list = bd;
498 alloc_Hp = bd->start;
499 alloc_HpLim = bd->start + BLOCK_SIZE_W;
510 allocated_bytes( void )
514 allocated = alloc_blocks * BLOCK_SIZE_W - (alloc_HpLim - alloc_Hp);
515 if (pinned_object_block != NULL) {
516 allocated -= (pinned_object_block->start + BLOCK_SIZE_W) -
517 pinned_object_block->free;
524 tidyAllocateLists (void)
526 if (small_alloc_list != NULL) {
527 ASSERT(alloc_Hp >= small_alloc_list->start &&
528 alloc_Hp <= small_alloc_list->start + BLOCK_SIZE);
529 small_alloc_list->free = alloc_Hp;
533 /* ---------------------------------------------------------------------------
534 Allocate a fixed/pinned object.
536 We allocate small pinned objects into a single block, allocating a
537 new block when the current one overflows. The block is chained
538 onto the large_object_list of generation 0 step 0.
540 NOTE: The GC can't in general handle pinned objects. This
541 interface is only safe to use for ByteArrays, which have no
542 pointers and don't require scavenging. It works because the
543 block's descriptor has the BF_LARGE flag set, so the block is
544 treated as a large object and chained onto various lists, rather
545 than the individual objects being copied. However, when it comes
546 to scavenge the block, the GC will only scavenge the first object.
547 The reason is that the GC can't linearly scan a block of pinned
548 objects at the moment (doing so would require using the
549 mostly-copying techniques). But since we're restricting ourselves
550 to pinned ByteArrays, not scavenging is ok.
552 This function is called by newPinnedByteArray# which immediately
553 fills the allocated memory with a MutableByteArray#.
554 ------------------------------------------------------------------------- */
557 allocatePinned( nat n )
560 bdescr *bd = pinned_object_block;
562 // If the request is for a large object, then allocate()
563 // will give us a pinned object anyway.
564 if (n >= LARGE_OBJECT_THRESHOLD/sizeof(W_)) {
570 TICK_ALLOC_HEAP_NOCTR(n);
573 // we always return 8-byte aligned memory. bd->free must be
574 // 8-byte aligned to begin with, so we just round up n to
575 // the nearest multiple of 8 bytes.
576 if (sizeof(StgWord) == 4) {
580 // If we don't have a block of pinned objects yet, or the current
581 // one isn't large enough to hold the new object, allocate a new one.
582 if (bd == NULL || (bd->free + n) > (bd->start + BLOCK_SIZE_W)) {
583 pinned_object_block = bd = allocBlock();
584 dbl_link_onto(bd, &g0s0->large_objects);
587 bd->flags = BF_PINNED | BF_LARGE;
588 bd->free = bd->start;
598 /* -----------------------------------------------------------------------------
599 Allocation functions for GMP.
601 These all use the allocate() interface - we can't have any garbage
602 collection going on during a gmp operation, so we use allocate()
603 which always succeeds. The gmp operations which might need to
604 allocate will ask the storage manager (via doYouWantToGC()) whether
605 a garbage collection is required, in case we get into a loop doing
606 only allocate() style allocation.
607 -------------------------------------------------------------------------- */
610 stgAllocForGMP (size_t size_in_bytes)
613 nat data_size_in_words, total_size_in_words;
615 /* round up to a whole number of words */
616 data_size_in_words = (size_in_bytes + sizeof(W_) + 1) / sizeof(W_);
617 total_size_in_words = sizeofW(StgArrWords) + data_size_in_words;
619 /* allocate and fill it in. */
620 arr = (StgArrWords *)allocate(total_size_in_words);
621 SET_ARR_HDR(arr, &stg_ARR_WORDS_info, CCCS, data_size_in_words);
623 /* and return a ptr to the goods inside the array */
624 return(BYTE_ARR_CTS(arr));
628 stgReallocForGMP (void *ptr, size_t old_size, size_t new_size)
630 void *new_stuff_ptr = stgAllocForGMP(new_size);
632 char *p = (char *) ptr;
633 char *q = (char *) new_stuff_ptr;
635 for (; i < old_size; i++, p++, q++) {
639 return(new_stuff_ptr);
643 stgDeallocForGMP (void *ptr STG_UNUSED,
644 size_t size STG_UNUSED)
646 /* easy for us: the garbage collector does the dealloc'n */
649 /* -----------------------------------------------------------------------------
651 * -------------------------------------------------------------------------- */
653 /* -----------------------------------------------------------------------------
656 * Approximate how much we've allocated: number of blocks in the
657 * nursery + blocks allocated via allocate() - unused nusery blocks.
658 * This leaves a little slop at the end of each block, and doesn't
659 * take into account large objects (ToDo).
660 * -------------------------------------------------------------------------- */
663 calcAllocated( void )
671 /* All tasks must be stopped. Can't assert that all the
672 capabilities are owned by the scheduler, though: one or more
673 tasks might have been stopped while they were running (non-main)
675 /* ASSERT(n_free_capabilities == RtsFlags.ParFlags.nNodes); */
678 n_free_capabilities * RtsFlags.GcFlags.minAllocAreaSize * BLOCK_SIZE_W
681 for (cap = free_capabilities; cap != NULL; cap = cap->link) {
682 for ( bd = cap->r.rCurrentNursery->link; bd != NULL; bd = bd->link ) {
683 allocated -= BLOCK_SIZE_W;
685 if (cap->r.rCurrentNursery->free < cap->r.rCurrentNursery->start
687 allocated -= (cap->r.rCurrentNursery->start + BLOCK_SIZE_W)
688 - cap->r.rCurrentNursery->free;
693 bdescr *current_nursery = MainCapability.r.rCurrentNursery;
695 allocated = (g0s0->n_blocks * BLOCK_SIZE_W) + allocated_bytes();
696 for ( bd = current_nursery->link; bd != NULL; bd = bd->link ) {
697 allocated -= BLOCK_SIZE_W;
699 if (current_nursery->free < current_nursery->start + BLOCK_SIZE_W) {
700 allocated -= (current_nursery->start + BLOCK_SIZE_W)
701 - current_nursery->free;
705 total_allocated += allocated;
709 /* Approximate the amount of live data in the heap. To be called just
710 * after garbage collection (see GarbageCollect()).
719 if (RtsFlags.GcFlags.generations == 1) {
720 live = (g0s0->n_to_blocks - 1) * BLOCK_SIZE_W +
721 ((lnat)g0s0->hp_bd->free - (lnat)g0s0->hp_bd->start) / sizeof(W_);
725 for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
726 for (s = 0; s < generations[g].n_steps; s++) {
727 /* approximate amount of live data (doesn't take into account slop
728 * at end of each block).
730 if (g == 0 && s == 0) {
733 stp = &generations[g].steps[s];
734 live += (stp->n_large_blocks + stp->n_blocks - 1) * BLOCK_SIZE_W;
735 if (stp->hp_bd != NULL) {
736 live += ((lnat)stp->hp_bd->free - (lnat)stp->hp_bd->start)
744 /* Approximate the number of blocks that will be needed at the next
745 * garbage collection.
747 * Assume: all data currently live will remain live. Steps that will
748 * be collected next time will therefore need twice as many blocks
749 * since all the data will be copied.
758 for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
759 for (s = 0; s < generations[g].n_steps; s++) {
760 if (g == 0 && s == 0) { continue; }
761 stp = &generations[g].steps[s];
762 if (generations[g].steps[0].n_blocks +
763 generations[g].steps[0].n_large_blocks
764 > generations[g].max_blocks
765 && stp->is_compacted == 0) {
766 needed += 2 * stp->n_blocks;
768 needed += stp->n_blocks;
775 /* -----------------------------------------------------------------------------
778 memInventory() checks for memory leaks by counting up all the
779 blocks we know about and comparing that to the number of blocks
780 allegedly floating around in the system.
781 -------------------------------------------------------------------------- */
791 lnat total_blocks = 0, free_blocks = 0;
793 /* count the blocks we current have */
795 for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
796 for (s = 0; s < generations[g].n_steps; s++) {
797 stp = &generations[g].steps[s];
798 total_blocks += stp->n_blocks;
799 if (RtsFlags.GcFlags.generations == 1) {
800 /* two-space collector has a to-space too :-) */
801 total_blocks += g0s0->n_to_blocks;
803 for (bd = stp->large_objects; bd; bd = bd->link) {
804 total_blocks += bd->blocks;
805 /* hack for megablock groups: they have an extra block or two in
806 the second and subsequent megablocks where the block
807 descriptors would normally go.
809 if (bd->blocks > BLOCKS_PER_MBLOCK) {
810 total_blocks -= (MBLOCK_SIZE / BLOCK_SIZE - BLOCKS_PER_MBLOCK)
811 * (bd->blocks/(MBLOCK_SIZE/BLOCK_SIZE));
817 /* any blocks held by allocate() */
818 for (bd = small_alloc_list; bd; bd = bd->link) {
819 total_blocks += bd->blocks;
823 if (RtsFlags.ProfFlags.doHeapProfile == HEAP_BY_RETAINER) {
824 total_blocks += retainerStackBlocks();
828 // count the blocks allocated by the arena allocator
829 total_blocks += arenaBlocks();
831 /* count the blocks on the free list */
832 free_blocks = countFreeList();
834 if (total_blocks + free_blocks != mblocks_allocated *
836 fprintf(stderr, "Blocks: %ld live + %ld free = %ld total (%ld around)\n",
837 total_blocks, free_blocks, total_blocks + free_blocks,
838 mblocks_allocated * BLOCKS_PER_MBLOCK);
841 ASSERT(total_blocks + free_blocks == mblocks_allocated * BLOCKS_PER_MBLOCK);
846 countBlocks(bdescr *bd)
849 for (n=0; bd != NULL; bd=bd->link) {
855 /* Full heap sanity check. */
861 if (RtsFlags.GcFlags.generations == 1) {
862 checkHeap(g0s0->to_blocks);
863 checkChain(g0s0->large_objects);
866 for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
867 for (s = 0; s < generations[g].n_steps; s++) {
868 ASSERT(countBlocks(generations[g].steps[s].blocks)
869 == generations[g].steps[s].n_blocks);
870 ASSERT(countBlocks(generations[g].steps[s].large_objects)
871 == generations[g].steps[s].n_large_blocks);
872 if (g == 0 && s == 0) { continue; }
873 checkHeap(generations[g].steps[s].blocks);
874 checkChain(generations[g].steps[s].large_objects);
876 checkMutableList(generations[g].mut_list, g);
877 checkMutOnceList(generations[g].mut_once_list, g);
881 checkFreeListSanity();
885 // handy function for use in gdb, because Bdescr() is inlined.
886 extern bdescr *_bdescr( StgPtr p );