1 /* -----------------------------------------------------------------------------
2 * $Id: Storage.c,v 1.41 2001/07/23 17:23:20 simonmar 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 nat alloc_blocks; /* number of allocate()d blocks since GC */
33 nat alloc_blocks_lim; /* approximate limit on alloc_blocks */
35 StgPtr alloc_Hp = NULL; /* next free byte in small_alloc_list */
36 StgPtr alloc_HpLim = NULL; /* end of block at small_alloc_list */
38 generation *generations; /* all the generations */
39 generation *g0; /* generation 0, for convenience */
40 generation *oldest_gen; /* oldest generation, for convenience */
41 step *g0s0; /* generation 0, step 0, for convenience */
43 lnat total_allocated = 0; /* total memory allocated during run */
46 * Storage manager mutex: protects all the above state from
47 * simultaneous access by two STG threads.
50 pthread_mutex_t sm_mutex = PTHREAD_MUTEX_INITIALIZER;
56 static void *stgAllocForGMP (size_t size_in_bytes);
57 static void *stgReallocForGMP (void *ptr, size_t old_size, size_t new_size);
58 static void stgDeallocForGMP (void *ptr, size_t size);
67 /* If we're doing heap profiling, we want a two-space heap with a
68 * fixed-size allocation area so that we get roughly even-spaced
72 /* As an experiment, try a 2 generation collector
75 #if defined(PROFILING) || defined(DEBUG)
76 if (RtsFlags.ProfFlags.doHeapProfile) {
77 RtsFlags.GcFlags.generations = 1;
78 RtsFlags.GcFlags.steps = 1;
79 RtsFlags.GcFlags.oldGenFactor = 0;
80 RtsFlags.GcFlags.heapSizeSuggestion = 0;
84 if (RtsFlags.GcFlags.heapSizeSuggestion >
85 RtsFlags.GcFlags.maxHeapSize) {
86 RtsFlags.GcFlags.maxHeapSize = RtsFlags.GcFlags.heapSizeSuggestion;
91 /* allocate generation info array */
92 generations = (generation *)stgMallocBytes(RtsFlags.GcFlags.generations
93 * sizeof(struct _generation),
96 /* Initialise all generations */
97 for(g = 0; g < RtsFlags.GcFlags.generations; g++) {
98 gen = &generations[g];
100 gen->mut_list = END_MUT_LIST;
101 gen->mut_once_list = END_MUT_LIST;
102 gen->collections = 0;
103 gen->failed_promotions = 0;
107 /* A couple of convenience pointers */
108 g0 = &generations[0];
109 oldest_gen = &generations[RtsFlags.GcFlags.generations-1];
111 /* Allocate step structures in each generation */
112 if (RtsFlags.GcFlags.generations > 1) {
113 /* Only for multiple-generations */
115 /* Oldest generation: one step */
116 oldest_gen->n_steps = 1;
118 stgMallocBytes(1 * sizeof(struct _step), "initStorage: last step");
120 /* set up all except the oldest generation with 2 steps */
121 for(g = 0; g < RtsFlags.GcFlags.generations-1; g++) {
122 generations[g].n_steps = RtsFlags.GcFlags.steps;
123 generations[g].steps =
124 stgMallocBytes (RtsFlags.GcFlags.steps * sizeof(struct _step),
125 "initStorage: steps");
129 /* single generation, i.e. a two-space collector */
131 g0->steps = stgMallocBytes (sizeof(struct _step), "initStorage: steps");
134 /* Initialise all steps */
135 for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
136 for (s = 0; s < generations[g].n_steps; s++) {
137 stp = &generations[g].steps[s];
141 stp->gen = &generations[g];
148 stp->large_objects = NULL;
149 stp->new_large_objects = NULL;
150 stp->scavenged_large_objects = NULL;
151 stp->is_compacted = 0;
155 /* Set up the destination pointers in each younger gen. step */
156 for (g = 0; g < RtsFlags.GcFlags.generations-1; g++) {
157 for (s = 0; s < generations[g].n_steps-1; s++) {
158 generations[g].steps[s].to = &generations[g].steps[s+1];
160 generations[g].steps[s].to = &generations[g+1].steps[0];
163 /* The oldest generation has one step and it is compacted. */
164 if (RtsFlags.GcFlags.compact) {
165 oldest_gen->steps[0].is_compacted = 1;
167 oldest_gen->steps[0].to = &oldest_gen->steps[0];
169 /* generation 0 is special: that's the nursery */
170 generations[0].max_blocks = 0;
172 /* G0S0: the allocation area. Policy: keep the allocation area
173 * small to begin with, even if we have a large suggested heap
174 * size. Reason: we're going to do a major collection first, and we
175 * don't want it to be a big one. This vague idea is borne out by
176 * rigorous experimental evidence.
178 g0s0 = &generations[0].steps[0];
182 weak_ptr_list = NULL;
185 /* initialise the allocate() interface */
186 small_alloc_list = NULL;
187 large_alloc_list = NULL;
189 alloc_blocks_lim = RtsFlags.GcFlags.minAllocAreaSize;
191 /* Tell GNU multi-precision pkg about our custom alloc functions */
192 mp_set_memory_functions(stgAllocForGMP, stgReallocForGMP, stgDeallocForGMP);
195 pthread_mutex_init(&sm_mutex, NULL);
198 IF_DEBUG(gc, statDescribeGens());
204 stat_exit(calcAllocated());
207 /* -----------------------------------------------------------------------------
210 The entry code for every CAF does the following:
212 - builds a CAF_BLACKHOLE in the heap
213 - pushes an update frame pointing to the CAF_BLACKHOLE
214 - invokes UPD_CAF(), which:
215 - calls newCaf, below
216 - updates the CAF with a static indirection to the CAF_BLACKHOLE
218 Why do we build a BLACKHOLE in the heap rather than just updating
219 the thunk directly? It's so that we only need one kind of update
220 frame - otherwise we'd need a static version of the update frame too.
222 newCaf() does the following:
224 - it puts the CAF on the oldest generation's mut-once list.
225 This is so that we can treat the CAF as a root when collecting
228 For GHCI, we have additional requirements when dealing with CAFs:
230 - we must *retain* all dynamically-loaded CAFs ever entered,
231 just in case we need them again.
232 - we must be able to *revert* CAFs that have been evaluated, to
233 their pre-evaluated form.
235 To do this, we use an additional CAF list. When newCaf() is
236 called on a dynamically-loaded CAF, we add it to the CAF list
237 instead of the old-generation mutable list, and save away its
238 old info pointer (in caf->saved_info) for later reversion.
240 To revert all the CAFs, we traverse the CAF list and reset the
241 info pointer to caf->saved_info, then throw away the CAF list.
242 (see GC.c:revertCAFs()).
246 -------------------------------------------------------------------------- */
249 newCAF(StgClosure* caf)
251 /* Put this CAF on the mutable list for the old generation.
252 * This is a HACK - the IND_STATIC closure doesn't really have
253 * a mut_link field, but we pretend it has - in fact we re-use
254 * the STATIC_LINK field for the time being, because when we
255 * come to do a major GC we won't need the mut_link field
256 * any more and can use it as a STATIC_LINK.
258 ACQUIRE_LOCK(&sm_mutex);
260 if (is_dynamically_loaded_rwdata_ptr((StgPtr)caf)) {
261 ((StgIndStatic *)caf)->saved_info = (StgInfoTable *)caf->header.info;
262 ((StgIndStatic *)caf)->static_link = caf_list;
265 ((StgIndStatic *)caf)->saved_info = NULL;
266 ((StgMutClosure *)caf)->mut_link = oldest_gen->mut_once_list;
267 oldest_gen->mut_once_list = (StgMutClosure *)caf;
270 RELEASE_LOCK(&sm_mutex);
273 /* If we are PAR or DIST then we never forget a CAF */
275 //belch("<##> Globalising CAF %08x %s",caf,info_type(caf));
276 newGA=makeGlobal(caf,rtsTrue); /*given full weight*/
282 /* -----------------------------------------------------------------------------
284 -------------------------------------------------------------------------- */
287 allocNurseries( void )
296 for (cap = free_capabilities; cap != NULL; cap = cap->link) {
297 cap->rNursery = allocNursery(NULL, RtsFlags.GcFlags.minAllocAreaSize);
298 cap->rCurrentNursery = cap->rNursery;
299 for (bd = cap->rNursery; bd != NULL; bd = bd->link) {
300 bd->u.back = (bdescr *)cap;
303 /* Set the back links to be equal to the Capability,
304 * so we can do slightly better informed locking.
308 nursery_blocks = RtsFlags.GcFlags.minAllocAreaSize;
309 g0s0->blocks = allocNursery(NULL, nursery_blocks);
310 g0s0->n_blocks = nursery_blocks;
311 g0s0->to_blocks = NULL;
312 g0s0->n_to_blocks = 0;
313 MainRegTable.rNursery = g0s0->blocks;
314 MainRegTable.rCurrentNursery = g0s0->blocks;
315 /* hp, hpLim, hp_bd, to_space etc. aren't used in G0S0 */
320 resetNurseries( void )
326 /* All tasks must be stopped */
327 ASSERT(n_free_capabilities == RtsFlags.ParFlags.nNodes);
329 for (cap = free_capabilities; cap != NULL; cap = cap->link) {
330 for (bd = cap->rNursery; bd; bd = bd->link) {
331 bd->free = bd->start;
332 ASSERT(bd->gen_no == 0);
333 ASSERT(bd->step == g0s0);
334 IF_DEBUG(sanity,memset(bd->start, 0xaa, BLOCK_SIZE));
336 cap->rCurrentNursery = cap->rNursery;
339 for (bd = g0s0->blocks; bd; bd = bd->link) {
340 bd->free = bd->start;
341 ASSERT(bd->gen_no == 0);
342 ASSERT(bd->step == g0s0);
343 IF_DEBUG(sanity,memset(bd->start, 0xaa, BLOCK_SIZE));
345 MainRegTable.rNursery = g0s0->blocks;
346 MainRegTable.rCurrentNursery = g0s0->blocks;
351 allocNursery (bdescr *last_bd, nat blocks)
356 /* Allocate a nursery */
357 for (i=0; i < blocks; i++) {
363 bd->free = bd->start;
370 resizeNursery ( nat blocks )
375 barf("resizeNursery: can't resize in SMP mode");
378 if (nursery_blocks == blocks) {
379 ASSERT(g0s0->n_blocks == blocks);
383 else if (nursery_blocks < blocks) {
384 IF_DEBUG(gc, fprintf(stderr, "Increasing size of nursery to %d blocks\n",
386 g0s0->blocks = allocNursery(g0s0->blocks, blocks-nursery_blocks);
392 IF_DEBUG(gc, fprintf(stderr, "Decreasing size of nursery to %d blocks\n",
394 for (bd = g0s0->blocks; nursery_blocks > blocks; nursery_blocks--) {
402 g0s0->n_blocks = nursery_blocks = blocks;
405 /* -----------------------------------------------------------------------------
406 The allocate() interface
408 allocate(n) always succeeds, and returns a chunk of memory n words
409 long. n can be larger than the size of a block if necessary, in
410 which case a contiguous block group will be allocated.
411 -------------------------------------------------------------------------- */
419 ACQUIRE_LOCK(&sm_mutex);
421 TICK_ALLOC_HEAP_NOCTR(n);
424 /* big allocation (>LARGE_OBJECT_THRESHOLD) */
425 /* ToDo: allocate directly into generation 1 */
426 if (n >= LARGE_OBJECT_THRESHOLD/sizeof(W_)) {
427 nat req_blocks = (lnat)BLOCK_ROUND_UP(n*sizeof(W_)) / BLOCK_SIZE;
428 bd = allocGroup(req_blocks);
429 dbl_link_onto(bd, &g0s0->large_objects);
432 bd->flags = BF_LARGE;
433 bd->free = bd->start;
434 /* don't add these blocks to alloc_blocks, since we're assuming
435 * that large objects are likely to remain live for quite a while
436 * (eg. running threads), so garbage collecting early won't make
439 alloc_blocks += req_blocks;
440 RELEASE_LOCK(&sm_mutex);
443 /* small allocation (<LARGE_OBJECT_THRESHOLD) */
444 } else if (small_alloc_list == NULL || alloc_Hp + n > alloc_HpLim) {
445 if (small_alloc_list) {
446 small_alloc_list->free = alloc_Hp;
449 bd->link = small_alloc_list;
450 small_alloc_list = bd;
454 alloc_Hp = bd->start;
455 alloc_HpLim = bd->start + BLOCK_SIZE_W;
461 RELEASE_LOCK(&sm_mutex);
465 lnat allocated_bytes(void)
467 return (alloc_blocks * BLOCK_SIZE_W - (alloc_HpLim - alloc_Hp));
470 /* -----------------------------------------------------------------------------
471 Allocation functions for GMP.
473 These all use the allocate() interface - we can't have any garbage
474 collection going on during a gmp operation, so we use allocate()
475 which always succeeds. The gmp operations which might need to
476 allocate will ask the storage manager (via doYouWantToGC()) whether
477 a garbage collection is required, in case we get into a loop doing
478 only allocate() style allocation.
479 -------------------------------------------------------------------------- */
482 stgAllocForGMP (size_t size_in_bytes)
485 nat data_size_in_words, total_size_in_words;
487 /* should be a multiple of sizeof(StgWord) (whole no. of limbs) */
488 ASSERT(size_in_bytes % sizeof(W_) == 0);
490 data_size_in_words = size_in_bytes / sizeof(W_);
491 total_size_in_words = sizeofW(StgArrWords) + data_size_in_words;
493 /* allocate and fill it in. */
494 arr = (StgArrWords *)allocate(total_size_in_words);
495 SET_ARR_HDR(arr, &stg_ARR_WORDS_info, CCCS, data_size_in_words);
497 /* and return a ptr to the goods inside the array */
498 return(BYTE_ARR_CTS(arr));
502 stgReallocForGMP (void *ptr, size_t old_size, size_t new_size)
504 void *new_stuff_ptr = stgAllocForGMP(new_size);
506 char *p = (char *) ptr;
507 char *q = (char *) new_stuff_ptr;
509 for (; i < old_size; i++, p++, q++) {
513 return(new_stuff_ptr);
517 stgDeallocForGMP (void *ptr STG_UNUSED,
518 size_t size STG_UNUSED)
520 /* easy for us: the garbage collector does the dealloc'n */
523 /* -----------------------------------------------------------------------------
525 * -------------------------------------------------------------------------- */
527 /* -----------------------------------------------------------------------------
530 * Approximate how much we've allocated: number of blocks in the
531 * nursery + blocks allocated via allocate() - unused nusery blocks.
532 * This leaves a little slop at the end of each block, and doesn't
533 * take into account large objects (ToDo).
534 * -------------------------------------------------------------------------- */
537 calcAllocated( void )
545 /* All tasks must be stopped. Can't assert that all the
546 capabilities are owned by the scheduler, though: one or more
547 tasks might have been stopped while they were running (non-main)
549 /* ASSERT(n_free_capabilities == RtsFlags.ParFlags.nNodes); */
552 n_free_capabilities * RtsFlags.GcFlags.minAllocAreaSize * BLOCK_SIZE_W
555 for (cap = free_capabilities; cap != NULL; cap = cap->link) {
556 for ( bd = cap->rCurrentNursery->link; bd != NULL; bd = bd->link ) {
557 allocated -= BLOCK_SIZE_W;
559 if (cap->rCurrentNursery->free < cap->rCurrentNursery->start
561 allocated -= (cap->rCurrentNursery->start + BLOCK_SIZE_W)
562 - cap->rCurrentNursery->free;
567 bdescr *current_nursery = MainRegTable.rCurrentNursery;
569 allocated = (nursery_blocks * BLOCK_SIZE_W) + allocated_bytes();
570 for ( bd = current_nursery->link; bd != NULL; bd = bd->link ) {
571 allocated -= BLOCK_SIZE_W;
573 if (current_nursery->free < current_nursery->start + BLOCK_SIZE_W) {
574 allocated -= (current_nursery->start + BLOCK_SIZE_W)
575 - current_nursery->free;
579 total_allocated += allocated;
583 /* Approximate the amount of live data in the heap. To be called just
584 * after garbage collection (see GarbageCollect()).
593 if (RtsFlags.GcFlags.generations == 1) {
594 live = (g0s0->n_to_blocks - 1) * BLOCK_SIZE_W +
595 ((lnat)g0s0->hp_bd->free - (lnat)g0s0->hp_bd->start) / sizeof(W_);
599 for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
600 for (s = 0; s < generations[g].n_steps; s++) {
601 /* approximate amount of live data (doesn't take into account slop
602 * at end of each block).
604 if (g == 0 && s == 0) {
607 stp = &generations[g].steps[s];
608 live += (stp->n_blocks - 1) * BLOCK_SIZE_W;
609 if (stp->hp_bd != NULL) {
610 live += ((lnat)stp->hp_bd->free - (lnat)stp->hp_bd->start)
618 /* Approximate the number of blocks that will be needed at the next
619 * garbage collection.
621 * Assume: all data currently live will remain live. Steps that will
622 * be collected next time will therefore need twice as many blocks
623 * since all the data will be copied.
632 for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
633 for (s = 0; s < generations[g].n_steps; s++) {
634 if (g == 0 && s == 0) { continue; }
635 stp = &generations[g].steps[s];
636 if (generations[g].steps[0].n_blocks > generations[g].max_blocks
637 && stp->is_compacted == 0) {
638 needed += 2 * stp->n_blocks;
640 needed += stp->n_blocks;
647 /* -----------------------------------------------------------------------------
650 memInventory() checks for memory leaks by counting up all the
651 blocks we know about and comparing that to the number of blocks
652 allegedly floating around in the system.
653 -------------------------------------------------------------------------- */
663 lnat total_blocks = 0, free_blocks = 0;
665 /* count the blocks we current have */
667 for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
668 for (s = 0; s < generations[g].n_steps; s++) {
669 stp = &generations[g].steps[s];
670 total_blocks += stp->n_blocks;
671 if (RtsFlags.GcFlags.generations == 1) {
672 /* two-space collector has a to-space too :-) */
673 total_blocks += g0s0->n_to_blocks;
675 for (bd = stp->large_objects; bd; bd = bd->link) {
676 total_blocks += bd->blocks;
677 /* hack for megablock groups: they have an extra block or two in
678 the second and subsequent megablocks where the block
679 descriptors would normally go.
681 if (bd->blocks > BLOCKS_PER_MBLOCK) {
682 total_blocks -= (MBLOCK_SIZE / BLOCK_SIZE - BLOCKS_PER_MBLOCK)
683 * (bd->blocks/(MBLOCK_SIZE/BLOCK_SIZE));
689 /* any blocks held by allocate() */
690 for (bd = small_alloc_list; bd; bd = bd->link) {
691 total_blocks += bd->blocks;
693 for (bd = large_alloc_list; bd; bd = bd->link) {
694 total_blocks += bd->blocks;
697 /* count the blocks on the free list */
698 free_blocks = countFreeList();
700 if (total_blocks + free_blocks != mblocks_allocated *
702 fprintf(stderr, "Blocks: %ld live + %ld free = %ld total (%ld around)\n",
703 total_blocks, free_blocks, total_blocks + free_blocks,
704 mblocks_allocated * BLOCKS_PER_MBLOCK);
707 ASSERT(total_blocks + free_blocks == mblocks_allocated * BLOCKS_PER_MBLOCK);
711 countBlocks(bdescr *bd)
714 for (n=0; bd != NULL; bd=bd->link) {
720 /* Full heap sanity check. */
726 if (RtsFlags.GcFlags.generations == 1) {
727 checkHeap(g0s0->to_blocks);
728 checkChain(g0s0->large_objects);
731 for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
732 for (s = 0; s < generations[g].n_steps; s++) {
733 if (g == 0 && s == 0) { continue; }
734 checkHeap(generations[g].steps[s].blocks);
735 ASSERT(countBlocks(generations[g].steps[s].blocks)
736 == generations[g].steps[s].n_blocks);
737 checkChain(generations[g].steps[s].large_objects);
739 checkMutableList(generations[g].mut_list, g);
740 checkMutOnceList(generations[g].mut_once_list, g);
744 checkFreeListSanity();