1 /* -----------------------------------------------------------------------------
2 * $Id: Storage.c,v 1.48 2001/08/09 12:46:06 sewardj Exp $
4 * (c) The GHC Team, 1998-1999
6 * Storage manager front end
8 * ---------------------------------------------------------------------------*/
15 #include "BlockAlloc.h"
22 #include "StoragePriv.h"
25 nat nursery_blocks; /* number of blocks in the nursery */
28 StgClosure *caf_list = NULL;
30 bdescr *small_alloc_list; /* allocate()d small objects */
31 bdescr *large_alloc_list; /* allocate()d large objects */
32 bdescr *pinned_object_block; /* allocate pinned objects into this block */
33 nat alloc_blocks; /* number of allocate()d blocks since GC */
34 nat alloc_blocks_lim; /* approximate limit on alloc_blocks */
36 StgPtr alloc_Hp = NULL; /* next free byte in small_alloc_list */
37 StgPtr alloc_HpLim = NULL; /* end of block at small_alloc_list */
39 generation *generations; /* all the generations */
40 generation *g0; /* generation 0, for convenience */
41 generation *oldest_gen; /* oldest generation, for convenience */
42 step *g0s0; /* generation 0, step 0, for convenience */
44 lnat total_allocated = 0; /* total memory allocated during run */
47 * Storage manager mutex: protects all the above state from
48 * simultaneous access by two STG threads.
51 pthread_mutex_t sm_mutex = PTHREAD_MUTEX_INITIALIZER;
57 static void *stgAllocForGMP (size_t size_in_bytes);
58 static void *stgReallocForGMP (void *ptr, size_t old_size, size_t new_size);
59 static void stgDeallocForGMP (void *ptr, size_t size);
68 /* If we're doing heap profiling, we want a two-space heap with a
69 * fixed-size allocation area so that we get roughly even-spaced
73 /* As an experiment, try a 2 generation collector
76 #if defined(PROFILING) || defined(DEBUG)
77 if (RtsFlags.ProfFlags.doHeapProfile) {
78 RtsFlags.GcFlags.generations = 1;
79 RtsFlags.GcFlags.steps = 1;
80 RtsFlags.GcFlags.oldGenFactor = 0;
81 RtsFlags.GcFlags.heapSizeSuggestion = 0;
85 if (RtsFlags.GcFlags.maxHeapSize != 0 &&
86 RtsFlags.GcFlags.heapSizeSuggestion >
87 RtsFlags.GcFlags.maxHeapSize) {
88 RtsFlags.GcFlags.maxHeapSize = RtsFlags.GcFlags.heapSizeSuggestion;
93 /* allocate generation info array */
94 generations = (generation *)stgMallocBytes(RtsFlags.GcFlags.generations
95 * sizeof(struct _generation),
98 /* Initialise all generations */
99 for(g = 0; g < RtsFlags.GcFlags.generations; g++) {
100 gen = &generations[g];
102 gen->mut_list = END_MUT_LIST;
103 gen->mut_once_list = END_MUT_LIST;
104 gen->collections = 0;
105 gen->failed_promotions = 0;
109 /* A couple of convenience pointers */
110 g0 = &generations[0];
111 oldest_gen = &generations[RtsFlags.GcFlags.generations-1];
113 /* Allocate step structures in each generation */
114 if (RtsFlags.GcFlags.generations > 1) {
115 /* Only for multiple-generations */
117 /* Oldest generation: one step */
118 oldest_gen->n_steps = 1;
120 stgMallocBytes(1 * sizeof(struct _step), "initStorage: last step");
122 /* set up all except the oldest generation with 2 steps */
123 for(g = 0; g < RtsFlags.GcFlags.generations-1; g++) {
124 generations[g].n_steps = RtsFlags.GcFlags.steps;
125 generations[g].steps =
126 stgMallocBytes (RtsFlags.GcFlags.steps * sizeof(struct _step),
127 "initStorage: steps");
131 /* single generation, i.e. a two-space collector */
133 g0->steps = stgMallocBytes (sizeof(struct _step), "initStorage: steps");
136 /* Initialise all steps */
137 for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
138 for (s = 0; s < generations[g].n_steps; s++) {
139 stp = &generations[g].steps[s];
143 stp->gen = &generations[g];
150 stp->large_objects = NULL;
151 stp->n_large_blocks = 0;
152 stp->new_large_objects = NULL;
153 stp->scavenged_large_objects = NULL;
154 stp->n_scavenged_large_blocks = 0;
155 stp->is_compacted = 0;
160 /* Set up the destination pointers in each younger gen. step */
161 for (g = 0; g < RtsFlags.GcFlags.generations-1; g++) {
162 for (s = 0; s < generations[g].n_steps-1; s++) {
163 generations[g].steps[s].to = &generations[g].steps[s+1];
165 generations[g].steps[s].to = &generations[g+1].steps[0];
168 /* The oldest generation has one step and it is compacted. */
169 if (RtsFlags.GcFlags.compact) {
170 oldest_gen->steps[0].is_compacted = 1;
172 oldest_gen->steps[0].to = &oldest_gen->steps[0];
174 /* generation 0 is special: that's the nursery */
175 generations[0].max_blocks = 0;
177 /* G0S0: the allocation area. Policy: keep the allocation area
178 * small to begin with, even if we have a large suggested heap
179 * size. Reason: we're going to do a major collection first, and we
180 * don't want it to be a big one. This vague idea is borne out by
181 * rigorous experimental evidence.
183 g0s0 = &generations[0].steps[0];
187 weak_ptr_list = NULL;
190 /* initialise the allocate() interface */
191 small_alloc_list = NULL;
192 large_alloc_list = NULL;
194 alloc_blocks_lim = RtsFlags.GcFlags.minAllocAreaSize;
196 /* Tell GNU multi-precision pkg about our custom alloc functions */
197 mp_set_memory_functions(stgAllocForGMP, stgReallocForGMP, stgDeallocForGMP);
200 pthread_mutex_init(&sm_mutex, NULL);
203 IF_DEBUG(gc, statDescribeGens());
209 stat_exit(calcAllocated());
212 /* -----------------------------------------------------------------------------
215 The entry code for every CAF does the following:
217 - builds a CAF_BLACKHOLE in the heap
218 - pushes an update frame pointing to the CAF_BLACKHOLE
219 - invokes UPD_CAF(), which:
220 - calls newCaf, below
221 - updates the CAF with a static indirection to the CAF_BLACKHOLE
223 Why do we build a BLACKHOLE in the heap rather than just updating
224 the thunk directly? It's so that we only need one kind of update
225 frame - otherwise we'd need a static version of the update frame too.
227 newCaf() does the following:
229 - it puts the CAF on the oldest generation's mut-once list.
230 This is so that we can treat the CAF as a root when collecting
233 For GHCI, we have additional requirements when dealing with CAFs:
235 - we must *retain* all dynamically-loaded CAFs ever entered,
236 just in case we need them again.
237 - we must be able to *revert* CAFs that have been evaluated, to
238 their pre-evaluated form.
240 To do this, we use an additional CAF list. When newCaf() is
241 called on a dynamically-loaded CAF, we add it to the CAF list
242 instead of the old-generation mutable list, and save away its
243 old info pointer (in caf->saved_info) for later reversion.
245 To revert all the CAFs, we traverse the CAF list and reset the
246 info pointer to caf->saved_info, then throw away the CAF list.
247 (see GC.c:revertCAFs()).
251 -------------------------------------------------------------------------- */
254 newCAF(StgClosure* caf)
256 /* Put this CAF on the mutable list for the old generation.
257 * This is a HACK - the IND_STATIC closure doesn't really have
258 * a mut_link field, but we pretend it has - in fact we re-use
259 * the STATIC_LINK field for the time being, because when we
260 * come to do a major GC we won't need the mut_link field
261 * any more and can use it as a STATIC_LINK.
263 ACQUIRE_LOCK(&sm_mutex);
265 if (is_dynamically_loaded_rwdata_ptr((StgPtr)caf)) {
266 ((StgIndStatic *)caf)->saved_info = (StgInfoTable *)caf->header.info;
267 ((StgIndStatic *)caf)->static_link = caf_list;
270 ((StgIndStatic *)caf)->saved_info = NULL;
271 ((StgMutClosure *)caf)->mut_link = oldest_gen->mut_once_list;
272 oldest_gen->mut_once_list = (StgMutClosure *)caf;
275 RELEASE_LOCK(&sm_mutex);
278 /* If we are PAR or DIST then we never forget a CAF */
280 //belch("<##> Globalising CAF %08x %s",caf,info_type(caf));
281 newGA=makeGlobal(caf,rtsTrue); /*given full weight*/
287 /* -----------------------------------------------------------------------------
289 -------------------------------------------------------------------------- */
292 allocNurseries( void )
301 for (cap = free_capabilities; cap != NULL; cap = cap->link) {
302 cap->rNursery = allocNursery(NULL, RtsFlags.GcFlags.minAllocAreaSize);
303 cap->rCurrentNursery = cap->rNursery;
304 for (bd = cap->rNursery; bd != NULL; bd = bd->link) {
305 bd->u.back = (bdescr *)cap;
308 /* Set the back links to be equal to the Capability,
309 * so we can do slightly better informed locking.
313 nursery_blocks = RtsFlags.GcFlags.minAllocAreaSize;
314 g0s0->blocks = allocNursery(NULL, nursery_blocks);
315 g0s0->n_blocks = nursery_blocks;
316 g0s0->to_blocks = NULL;
317 g0s0->n_to_blocks = 0;
318 MainRegTable.rNursery = g0s0->blocks;
319 MainRegTable.rCurrentNursery = g0s0->blocks;
320 /* hp, hpLim, hp_bd, to_space etc. aren't used in G0S0 */
325 resetNurseries( void )
331 /* All tasks must be stopped */
332 ASSERT(n_free_capabilities == RtsFlags.ParFlags.nNodes);
334 for (cap = free_capabilities; cap != NULL; cap = cap->link) {
335 for (bd = cap->rNursery; bd; bd = bd->link) {
336 bd->free = bd->start;
337 ASSERT(bd->gen_no == 0);
338 ASSERT(bd->step == g0s0);
339 IF_DEBUG(sanity,memset(bd->start, 0xaa, BLOCK_SIZE));
341 cap->rCurrentNursery = cap->rNursery;
344 for (bd = g0s0->blocks; bd; bd = bd->link) {
345 bd->free = bd->start;
346 ASSERT(bd->gen_no == 0);
347 ASSERT(bd->step == g0s0);
348 IF_DEBUG(sanity,memset(bd->start, 0xaa, BLOCK_SIZE));
350 MainRegTable.rNursery = g0s0->blocks;
351 MainRegTable.rCurrentNursery = g0s0->blocks;
356 allocNursery (bdescr *last_bd, nat blocks)
361 /* Allocate a nursery */
362 for (i=0; i < blocks; i++) {
368 bd->free = bd->start;
375 resizeNursery ( nat blocks )
380 barf("resizeNursery: can't resize in SMP mode");
383 if (nursery_blocks == blocks) {
384 ASSERT(g0s0->n_blocks == blocks);
388 else if (nursery_blocks < blocks) {
389 IF_DEBUG(gc, fprintf(stderr, "Increasing size of nursery to %d blocks\n",
391 g0s0->blocks = allocNursery(g0s0->blocks, blocks-nursery_blocks);
397 IF_DEBUG(gc, fprintf(stderr, "Decreasing size of nursery to %d blocks\n",
399 for (bd = g0s0->blocks; nursery_blocks > blocks; nursery_blocks--) {
407 g0s0->n_blocks = nursery_blocks = blocks;
410 /* -----------------------------------------------------------------------------
411 The allocate() interface
413 allocate(n) always succeeds, and returns a chunk of memory n words
414 long. n can be larger than the size of a block if necessary, in
415 which case a contiguous block group will be allocated.
416 -------------------------------------------------------------------------- */
424 ACQUIRE_LOCK(&sm_mutex);
426 TICK_ALLOC_HEAP_NOCTR(n);
429 /* big allocation (>LARGE_OBJECT_THRESHOLD) */
430 /* ToDo: allocate directly into generation 1 */
431 if (n >= LARGE_OBJECT_THRESHOLD/sizeof(W_)) {
432 nat req_blocks = (lnat)BLOCK_ROUND_UP(n*sizeof(W_)) / BLOCK_SIZE;
433 bd = allocGroup(req_blocks);
434 dbl_link_onto(bd, &g0s0->large_objects);
437 bd->flags = BF_LARGE;
438 bd->free = bd->start;
439 /* don't add these blocks to alloc_blocks, since we're assuming
440 * that large objects are likely to remain live for quite a while
441 * (eg. running threads), so garbage collecting early won't make
444 alloc_blocks += req_blocks;
445 RELEASE_LOCK(&sm_mutex);
448 /* small allocation (<LARGE_OBJECT_THRESHOLD) */
449 } else if (small_alloc_list == NULL || alloc_Hp + n > alloc_HpLim) {
450 if (small_alloc_list) {
451 small_alloc_list->free = alloc_Hp;
454 bd->link = small_alloc_list;
455 small_alloc_list = bd;
459 alloc_Hp = bd->start;
460 alloc_HpLim = bd->start + BLOCK_SIZE_W;
466 RELEASE_LOCK(&sm_mutex);
471 allocated_bytes( void )
473 return (alloc_blocks * BLOCK_SIZE_W - (alloc_HpLim - alloc_Hp));
476 /* ---------------------------------------------------------------------------
477 Allocate a fixed/pinned object.
479 We allocate small pinned objects into a single block, allocating a
480 new block when the current one overflows. The block is chained
481 onto the large_object_list of generation 0 step 0.
483 NOTE: The GC can't in general handle pinned objects. This
484 interface is only safe to use for ByteArrays, which have no
485 pointers and don't require scavenging. It works because the
486 block's descriptor has the BF_LARGE flag set, so the block is
487 treated as a large object and chained onto various lists, rather
488 than the individual objects being copied. However, when it comes
489 to scavenge the block, the GC will only scavenge the first object.
490 The reason is that the GC can't linearly scan a block of pinned
491 objects at the moment (doing so would require using the
492 mostly-copying techniques). But since we're restricting ourselves
493 to pinned ByteArrays, not scavenging is ok.
495 This function is called by newPinnedByteArray# which immediately
496 fills the allocated memory with a MutableByteArray#.
497 ------------------------------------------------------------------------- */
500 allocatePinned( nat n )
503 bdescr *bd = pinned_object_block;
505 ACQUIRE_LOCK(&sm_mutex);
507 TICK_ALLOC_HEAP_NOCTR(n);
510 // If the request is for a large object, then allocate()
511 // will give us a pinned object anyway.
512 if (n >= LARGE_OBJECT_THRESHOLD/sizeof(W_)) {
513 RELEASE_LOCK(&sm_mutex);
517 // If we don't have a block of pinned objects yet, or the current
518 // one isn't large enough to hold the new object, allocate a new one.
519 if (bd == NULL || (bd->free + n) > (bd->start + BLOCK_SIZE_W)) {
520 pinned_object_block = bd = allocBlock();
521 dbl_link_onto(bd, &g0s0->large_objects);
524 bd->flags = BF_LARGE;
525 bd->free = bd->start;
531 RELEASE_LOCK(&sm_mutex);
535 /* -----------------------------------------------------------------------------
536 Allocation functions for GMP.
538 These all use the allocate() interface - we can't have any garbage
539 collection going on during a gmp operation, so we use allocate()
540 which always succeeds. The gmp operations which might need to
541 allocate will ask the storage manager (via doYouWantToGC()) whether
542 a garbage collection is required, in case we get into a loop doing
543 only allocate() style allocation.
544 -------------------------------------------------------------------------- */
547 stgAllocForGMP (size_t size_in_bytes)
550 nat data_size_in_words, total_size_in_words;
552 /* should be a multiple of sizeof(StgWord) (whole no. of limbs) */
553 ASSERT(size_in_bytes % sizeof(W_) == 0);
555 data_size_in_words = size_in_bytes / sizeof(W_);
556 total_size_in_words = sizeofW(StgArrWords) + data_size_in_words;
558 /* allocate and fill it in. */
559 arr = (StgArrWords *)allocate(total_size_in_words);
560 SET_ARR_HDR(arr, &stg_ARR_WORDS_info, CCCS, data_size_in_words);
562 /* and return a ptr to the goods inside the array */
563 return(BYTE_ARR_CTS(arr));
567 stgReallocForGMP (void *ptr, size_t old_size, size_t new_size)
569 void *new_stuff_ptr = stgAllocForGMP(new_size);
571 char *p = (char *) ptr;
572 char *q = (char *) new_stuff_ptr;
574 for (; i < old_size; i++, p++, q++) {
578 return(new_stuff_ptr);
582 stgDeallocForGMP (void *ptr STG_UNUSED,
583 size_t size STG_UNUSED)
585 /* easy for us: the garbage collector does the dealloc'n */
588 /* -----------------------------------------------------------------------------
590 * -------------------------------------------------------------------------- */
592 /* -----------------------------------------------------------------------------
595 * Approximate how much we've allocated: number of blocks in the
596 * nursery + blocks allocated via allocate() - unused nusery blocks.
597 * This leaves a little slop at the end of each block, and doesn't
598 * take into account large objects (ToDo).
599 * -------------------------------------------------------------------------- */
602 calcAllocated( void )
610 /* All tasks must be stopped. Can't assert that all the
611 capabilities are owned by the scheduler, though: one or more
612 tasks might have been stopped while they were running (non-main)
614 /* ASSERT(n_free_capabilities == RtsFlags.ParFlags.nNodes); */
617 n_free_capabilities * RtsFlags.GcFlags.minAllocAreaSize * BLOCK_SIZE_W
620 for (cap = free_capabilities; cap != NULL; cap = cap->link) {
621 for ( bd = cap->rCurrentNursery->link; bd != NULL; bd = bd->link ) {
622 allocated -= BLOCK_SIZE_W;
624 if (cap->rCurrentNursery->free < cap->rCurrentNursery->start
626 allocated -= (cap->rCurrentNursery->start + BLOCK_SIZE_W)
627 - cap->rCurrentNursery->free;
632 bdescr *current_nursery = MainRegTable.rCurrentNursery;
634 allocated = (nursery_blocks * BLOCK_SIZE_W) + allocated_bytes();
635 for ( bd = current_nursery->link; bd != NULL; bd = bd->link ) {
636 allocated -= BLOCK_SIZE_W;
638 if (current_nursery->free < current_nursery->start + BLOCK_SIZE_W) {
639 allocated -= (current_nursery->start + BLOCK_SIZE_W)
640 - current_nursery->free;
644 total_allocated += allocated;
648 /* Approximate the amount of live data in the heap. To be called just
649 * after garbage collection (see GarbageCollect()).
658 if (RtsFlags.GcFlags.generations == 1) {
659 live = (g0s0->n_to_blocks - 1) * BLOCK_SIZE_W +
660 ((lnat)g0s0->hp_bd->free - (lnat)g0s0->hp_bd->start) / sizeof(W_);
664 for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
665 for (s = 0; s < generations[g].n_steps; s++) {
666 /* approximate amount of live data (doesn't take into account slop
667 * at end of each block).
669 if (g == 0 && s == 0) {
672 stp = &generations[g].steps[s];
673 live += (stp->n_large_blocks + stp->n_blocks - 1) * BLOCK_SIZE_W;
674 if (stp->hp_bd != NULL) {
675 live += ((lnat)stp->hp_bd->free - (lnat)stp->hp_bd->start)
683 /* Approximate the number of blocks that will be needed at the next
684 * garbage collection.
686 * Assume: all data currently live will remain live. Steps that will
687 * be collected next time will therefore need twice as many blocks
688 * since all the data will be copied.
697 for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
698 for (s = 0; s < generations[g].n_steps; s++) {
699 if (g == 0 && s == 0) { continue; }
700 stp = &generations[g].steps[s];
701 if (generations[g].steps[0].n_blocks +
702 generations[g].steps[0].n_large_blocks
703 > generations[g].max_blocks
704 && stp->is_compacted == 0) {
705 needed += 2 * stp->n_blocks;
707 needed += stp->n_blocks;
714 /* -----------------------------------------------------------------------------
717 memInventory() checks for memory leaks by counting up all the
718 blocks we know about and comparing that to the number of blocks
719 allegedly floating around in the system.
720 -------------------------------------------------------------------------- */
730 lnat total_blocks = 0, free_blocks = 0;
732 /* count the blocks we current have */
734 for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
735 for (s = 0; s < generations[g].n_steps; s++) {
736 stp = &generations[g].steps[s];
737 total_blocks += stp->n_blocks;
738 if (RtsFlags.GcFlags.generations == 1) {
739 /* two-space collector has a to-space too :-) */
740 total_blocks += g0s0->n_to_blocks;
742 for (bd = stp->large_objects; bd; bd = bd->link) {
743 total_blocks += bd->blocks;
744 /* hack for megablock groups: they have an extra block or two in
745 the second and subsequent megablocks where the block
746 descriptors would normally go.
748 if (bd->blocks > BLOCKS_PER_MBLOCK) {
749 total_blocks -= (MBLOCK_SIZE / BLOCK_SIZE - BLOCKS_PER_MBLOCK)
750 * (bd->blocks/(MBLOCK_SIZE/BLOCK_SIZE));
756 /* any blocks held by allocate() */
757 for (bd = small_alloc_list; bd; bd = bd->link) {
758 total_blocks += bd->blocks;
760 for (bd = large_alloc_list; bd; bd = bd->link) {
761 total_blocks += bd->blocks;
764 /* count the blocks on the free list */
765 free_blocks = countFreeList();
767 if (total_blocks + free_blocks != mblocks_allocated *
769 fprintf(stderr, "Blocks: %ld live + %ld free = %ld total (%ld around)\n",
770 total_blocks, free_blocks, total_blocks + free_blocks,
771 mblocks_allocated * BLOCKS_PER_MBLOCK);
774 ASSERT(total_blocks + free_blocks == mblocks_allocated * BLOCKS_PER_MBLOCK);
778 countBlocks(bdescr *bd)
781 for (n=0; bd != NULL; bd=bd->link) {
787 /* Full heap sanity check. */
793 if (RtsFlags.GcFlags.generations == 1) {
794 checkHeap(g0s0->to_blocks);
795 checkChain(g0s0->large_objects);
798 for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
799 for (s = 0; s < generations[g].n_steps; s++) {
800 if (g == 0 && s == 0) { continue; }
801 checkHeap(generations[g].steps[s].blocks);
802 checkChain(generations[g].steps[s].large_objects);
803 ASSERT(countBlocks(generations[g].steps[s].blocks)
804 == generations[g].steps[s].n_blocks);
805 ASSERT(countBlocks(generations[g].steps[s].large_objects)
806 == generations[g].steps[s].n_large_blocks);
808 checkMutableList(generations[g].mut_list, g);
809 checkMutOnceList(generations[g].mut_once_list, g);
813 checkFreeListSanity();