1 /* -----------------------------------------------------------------------------
2 * $Id: Storage.c,v 1.27 2000/11/01 11:41:47 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
71 #if defined(PROFILING) || defined(DEBUG)
72 if (RtsFlags.ProfFlags.doHeapProfile) {
73 RtsFlags.GcFlags.generations = 1;
74 RtsFlags.GcFlags.steps = 1;
75 RtsFlags.GcFlags.oldGenFactor = 0;
76 RtsFlags.GcFlags.heapSizeSuggestion = 0;
80 if (RtsFlags.GcFlags.heapSizeSuggestion >
81 RtsFlags.GcFlags.maxHeapSize) {
82 RtsFlags.GcFlags.maxHeapSize = RtsFlags.GcFlags.heapSizeSuggestion;
87 /* allocate generation info array */
88 generations = (generation *)stgMallocBytes(RtsFlags.GcFlags.generations
89 * sizeof(struct _generation),
92 /* Initialise all generations */
93 for(g = 0; g < RtsFlags.GcFlags.generations; g++) {
94 gen = &generations[g];
96 gen->mut_list = END_MUT_LIST;
97 gen->mut_once_list = END_MUT_LIST;
99 gen->failed_promotions = 0;
103 /* A couple of convenience pointers */
104 g0 = &generations[0];
105 oldest_gen = &generations[RtsFlags.GcFlags.generations-1];
107 /* Allocate step structures in each generation */
108 if (RtsFlags.GcFlags.generations > 1) {
109 /* Only for multiple-generations */
111 /* Oldest generation: one step */
112 oldest_gen->n_steps = 1;
114 stgMallocBytes(1 * sizeof(struct _step), "initStorage: last step");
116 /* set up all except the oldest generation with 2 steps */
117 for(g = 0; g < RtsFlags.GcFlags.generations-1; g++) {
118 generations[g].n_steps = RtsFlags.GcFlags.steps;
119 generations[g].steps =
120 stgMallocBytes (RtsFlags.GcFlags.steps * sizeof(struct _step),
121 "initStorage: steps");
125 /* single generation, i.e. a two-space collector */
127 g0->steps = stgMallocBytes (sizeof(struct _step), "initStorage: steps");
130 /* Initialise all steps */
131 for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
132 for (s = 0; s < generations[g].n_steps; s++) {
133 step = &generations[g].steps[s];
137 step->gen = &generations[g];
142 step->scan_bd = NULL;
143 step->large_objects = NULL;
144 step->new_large_objects = NULL;
145 step->scavenged_large_objects = NULL;
149 /* Set up the destination pointers in each younger gen. step */
150 for (g = 0; g < RtsFlags.GcFlags.generations-1; g++) {
151 for (s = 0; s < generations[g].n_steps-1; s++) {
152 generations[g].steps[s].to = &generations[g].steps[s+1];
154 generations[g].steps[s].to = &generations[g+1].steps[0];
157 /* The oldest generation has one step and its destination is the
159 oldest_gen->steps[0].to = &oldest_gen->steps[0];
161 /* generation 0 is special: that's the nursery */
162 generations[0].max_blocks = 0;
164 /* G0S0: the allocation area. Policy: keep the allocation area
165 * small to begin with, even if we have a large suggested heap
166 * size. Reason: we're going to do a major collection first, and we
167 * don't want it to be a big one. This vague idea is borne out by
168 * rigorous experimental evidence.
170 g0s0 = &generations[0].steps[0];
174 weak_ptr_list = NULL;
177 /* initialise the allocate() interface */
178 small_alloc_list = NULL;
179 large_alloc_list = NULL;
181 alloc_blocks_lim = RtsFlags.GcFlags.minAllocAreaSize;
183 /* Tell GNU multi-precision pkg about our custom alloc functions */
184 mp_set_memory_functions(stgAllocForGMP, stgReallocForGMP, stgDeallocForGMP);
187 pthread_mutex_init(&sm_mutex, NULL);
190 IF_DEBUG(gc, stat_describe_gens());
196 stat_exit(calcAllocated());
200 /* -----------------------------------------------------------------------------
202 -------------------------------------------------------------------------- */
205 newCAF(StgClosure* caf)
207 /* Put this CAF on the mutable list for the old generation.
208 * This is a HACK - the IND_STATIC closure doesn't really have
209 * a mut_link field, but we pretend it has - in fact we re-use
210 * the STATIC_LINK field for the time being, because when we
211 * come to do a major GC we won't need the mut_link field
212 * any more and can use it as a STATIC_LINK.
214 ACQUIRE_LOCK(&sm_mutex);
216 ASSERT( ((StgMutClosure*)caf)->mut_link == NULL );
217 ((StgMutClosure *)caf)->mut_link = oldest_gen->mut_once_list;
218 oldest_gen->mut_once_list = (StgMutClosure *)caf;
221 /* If we're Hugs, we also have to put it in the CAF table, so that
222 the CAF can be reverted. When reverting, CAFs created by compiled
223 code are recorded in the CAF table, which lives outside the
224 heap, in mallocville. CAFs created by interpreted code are
225 chained together via the link fields in StgCAFs, and are not
226 recorded in the CAF table.
228 ASSERT( get_itbl(caf)->type == THUNK_STATIC );
229 addToECafTable ( caf, get_itbl(caf) );
232 RELEASE_LOCK(&sm_mutex);
237 newCAF_made_by_Hugs(StgCAF* caf)
239 ACQUIRE_LOCK(&sm_mutex);
241 ASSERT( get_itbl(caf)->type == CAF_ENTERED );
242 recordOldToNewPtrs((StgMutClosure*)caf);
243 caf->link = ecafList;
244 ecafList = caf->link;
246 RELEASE_LOCK(&sm_mutex);
251 /* These initialisations are critical for correct operation
252 on the first call of addToECafTable.
254 StgCAF* ecafList = END_ECAF_LIST;
255 StgCAFTabEntry* ecafTable = NULL;
256 StgInt usedECafTable = 0;
257 StgInt sizeECafTable = 0;
260 void clearECafTable ( void )
265 void addToECafTable ( StgClosure* closure, StgInfoTable* origItbl )
269 if (usedECafTable == sizeECafTable) {
270 /* Make the initial table size be 8 */
272 if (sizeECafTable == 0) sizeECafTable = 8;
273 et2 = stgMallocBytes (
274 sizeECafTable * sizeof(StgCAFTabEntry),
276 for (i = 0; i < usedECafTable; i++)
277 et2[i] = ecafTable[i];
278 if (ecafTable) free(ecafTable);
281 ecafTable[usedECafTable].closure = closure;
282 ecafTable[usedECafTable].origItbl = origItbl;
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->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_space = NULL;
317 MainRegTable.rNursery = g0s0->blocks;
318 MainRegTable.rCurrentNursery = g0s0->blocks;
319 /* hp, hpLim, hp_bd, to_space etc. aren't used in G0S0 */
324 resetNurseries( void )
330 /* All tasks must be stopped */
331 ASSERT(n_free_capabilities == RtsFlags.ParFlags.nNodes);
333 for (cap = free_capabilities; cap != NULL; cap = cap->link) {
334 for (bd = cap->rNursery; bd; bd = bd->link) {
335 bd->free = bd->start;
336 ASSERT(bd->gen == g0);
337 ASSERT(bd->step == g0s0);
338 IF_DEBUG(sanity,memset(bd->start, 0xaa, BLOCK_SIZE));
340 cap->rCurrentNursery = cap->rNursery;
343 for (bd = g0s0->blocks; bd; bd = bd->link) {
344 bd->free = bd->start;
345 ASSERT(bd->gen == g0);
346 ASSERT(bd->step == g0s0);
347 IF_DEBUG(sanity,memset(bd->start, 0xaa, BLOCK_SIZE));
349 MainRegTable.rNursery = g0s0->blocks;
350 MainRegTable.rCurrentNursery = g0s0->blocks;
355 allocNursery (bdescr *last_bd, nat blocks)
360 /* Allocate a nursery */
361 for (i=0; i < blocks; i++) {
367 bd->free = bd->start;
374 resizeNursery ( nat blocks )
379 barf("resizeNursery: can't resize in SMP mode");
382 if (nursery_blocks == blocks) {
383 ASSERT(g0s0->n_blocks == blocks);
387 else if (nursery_blocks < blocks) {
388 IF_DEBUG(gc, fprintf(stderr, "Increasing size of nursery to %d blocks\n",
390 g0s0->blocks = allocNursery(g0s0->blocks, blocks-nursery_blocks);
396 IF_DEBUG(gc, fprintf(stderr, "Decreasing size of nursery to %d blocks\n",
398 for (bd = g0s0->blocks; nursery_blocks > blocks; nursery_blocks--) {
406 g0s0->n_blocks = nursery_blocks = blocks;
409 /* -----------------------------------------------------------------------------
410 The allocate() interface
412 allocate(n) always succeeds, and returns a chunk of memory n words
413 long. n can be larger than the size of a block if necessary, in
414 which case a contiguous block group will be allocated.
415 -------------------------------------------------------------------------- */
423 ACQUIRE_LOCK(&sm_mutex);
425 TICK_ALLOC_HEAP_NOCTR(n);
428 /* big allocation (>LARGE_OBJECT_THRESHOLD) */
429 /* ToDo: allocate directly into generation 1 */
430 if (n >= LARGE_OBJECT_THRESHOLD/sizeof(W_)) {
431 nat req_blocks = (lnat)BLOCK_ROUND_UP(n*sizeof(W_)) / BLOCK_SIZE;
432 bd = allocGroup(req_blocks);
433 dbl_link_onto(bd, &g0s0->large_objects);
437 bd->free = bd->start;
438 /* don't add these blocks to alloc_blocks, since we're assuming
439 * that large objects are likely to remain live for quite a while
440 * (eg. running threads), so garbage collecting early won't make
443 alloc_blocks += req_blocks;
444 RELEASE_LOCK(&sm_mutex);
447 /* small allocation (<LARGE_OBJECT_THRESHOLD) */
448 } else if (small_alloc_list == NULL || alloc_Hp + n > alloc_HpLim) {
449 if (small_alloc_list) {
450 small_alloc_list->free = alloc_Hp;
453 bd->link = small_alloc_list;
454 small_alloc_list = bd;
458 alloc_Hp = bd->start;
459 alloc_HpLim = bd->start + BLOCK_SIZE_W;
465 RELEASE_LOCK(&sm_mutex);
469 lnat allocated_bytes(void)
471 return (alloc_blocks * BLOCK_SIZE_W - (alloc_HpLim - alloc_Hp));
474 /* -----------------------------------------------------------------------------
475 Allocation functions for GMP.
477 These all use the allocate() interface - we can't have any garbage
478 collection going on during a gmp operation, so we use allocate()
479 which always succeeds. The gmp operations which might need to
480 allocate will ask the storage manager (via doYouWantToGC()) whether
481 a garbage collection is required, in case we get into a loop doing
482 only allocate() style allocation.
483 -------------------------------------------------------------------------- */
486 stgAllocForGMP (size_t size_in_bytes)
489 nat data_size_in_words, total_size_in_words;
491 /* should be a multiple of sizeof(StgWord) (whole no. of limbs) */
492 ASSERT(size_in_bytes % sizeof(W_) == 0);
494 data_size_in_words = size_in_bytes / sizeof(W_);
495 total_size_in_words = sizeofW(StgArrWords) + data_size_in_words;
497 /* allocate and fill it in. */
498 arr = (StgArrWords *)allocate(total_size_in_words);
499 SET_ARR_HDR(arr, &ARR_WORDS_info, CCCS, data_size_in_words);
501 /* and return a ptr to the goods inside the array */
502 return(BYTE_ARR_CTS(arr));
506 stgReallocForGMP (void *ptr, size_t old_size, size_t new_size)
508 void *new_stuff_ptr = stgAllocForGMP(new_size);
510 char *p = (char *) ptr;
511 char *q = (char *) new_stuff_ptr;
513 for (; i < old_size; i++, p++, q++) {
517 return(new_stuff_ptr);
521 stgDeallocForGMP (void *ptr STG_UNUSED,
522 size_t size STG_UNUSED)
524 /* easy for us: the garbage collector does the dealloc'n */
527 /* -----------------------------------------------------------------------------
529 * -------------------------------------------------------------------------- */
531 /* -----------------------------------------------------------------------------
534 * Approximate how much we've allocated: number of blocks in the
535 * nursery + blocks allocated via allocate() - unused nusery blocks.
536 * This leaves a little slop at the end of each block, and doesn't
537 * take into account large objects (ToDo).
538 * -------------------------------------------------------------------------- */
541 calcAllocated( void )
549 /* All tasks must be stopped. Can't assert that all the
550 capabilities are owned by the scheduler, though: one or more
551 tasks might have been stopped while they were running (non-main)
553 /* ASSERT(n_free_capabilities == RtsFlags.ParFlags.nNodes); */
556 n_free_capabilities * RtsFlags.GcFlags.minAllocAreaSize * BLOCK_SIZE_W
559 for (cap = free_capabilities; cap != NULL; cap = cap->link) {
560 for ( bd = cap->rCurrentNursery->link; bd != NULL; bd = bd->link ) {
561 allocated -= BLOCK_SIZE_W;
563 if (cap->rCurrentNursery->free < cap->rCurrentNursery->start
565 allocated -= (cap->rCurrentNursery->start + BLOCK_SIZE_W)
566 - cap->rCurrentNursery->free;
571 bdescr *current_nursery = MainRegTable.rCurrentNursery;
573 allocated = (nursery_blocks * BLOCK_SIZE_W) + allocated_bytes();
574 for ( bd = current_nursery->link; bd != NULL; bd = bd->link ) {
575 allocated -= BLOCK_SIZE_W;
577 if (current_nursery->free < current_nursery->start + BLOCK_SIZE_W) {
578 allocated -= (current_nursery->start + BLOCK_SIZE_W)
579 - current_nursery->free;
583 total_allocated += allocated;
587 /* Approximate the amount of live data in the heap. To be called just
588 * after garbage collection (see GarbageCollect()).
597 if (RtsFlags.GcFlags.generations == 1) {
598 live = (g0s0->to_blocks - 1) * BLOCK_SIZE_W +
599 ((lnat)g0s0->hp_bd->free - (lnat)g0s0->hp_bd->start) / sizeof(W_);
603 for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
604 for (s = 0; s < generations[g].n_steps; s++) {
605 /* approximate amount of live data (doesn't take into account slop
606 * at end of each block).
608 if (g == 0 && s == 0) {
611 step = &generations[g].steps[s];
612 live += (step->n_blocks - 1) * BLOCK_SIZE_W +
613 ((lnat)step->hp_bd->free - (lnat)step->hp_bd->start) / sizeof(W_);
619 /* Approximate the number of blocks that will be needed at the next
620 * garbage collection.
622 * Assume: all data currently live will remain live. Steps that will
623 * be collected next time will therefore need twice as many blocks
624 * since all the data will be copied.
633 for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
634 for (s = 0; s < generations[g].n_steps; s++) {
635 if (g == 0 && s == 0) { continue; }
636 step = &generations[g].steps[s];
637 if (generations[g].steps[0].n_blocks > generations[g].max_blocks) {
638 needed += 2 * step->n_blocks;
640 needed += step->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 step = &generations[g].steps[s];
670 total_blocks += step->n_blocks;
671 if (RtsFlags.GcFlags.generations == 1) {
672 /* two-space collector has a to-space too :-) */
673 total_blocks += g0s0->to_blocks;
675 for (bd = step->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 ASSERT(total_blocks + free_blocks == mblocks_allocated * BLOCKS_PER_MBLOCK);
703 if (total_blocks + free_blocks != mblocks_allocated *
705 fprintf(stderr, "Blocks: %ld live + %ld free = %ld total (%ld around)\n",
706 total_blocks, free_blocks, total_blocks + free_blocks,
707 mblocks_allocated * BLOCKS_PER_MBLOCK);
712 /* Full heap sanity check. */
719 if (RtsFlags.GcFlags.generations == 1) {
720 checkHeap(g0s0->to_space, NULL);
721 checkChain(g0s0->large_objects);
724 for (g = 0; g <= N; g++) {
725 for (s = 0; s < generations[g].n_steps; s++) {
726 if (g == 0 && s == 0) { continue; }
727 checkHeap(generations[g].steps[s].blocks, NULL);
730 for (g = N+1; g < RtsFlags.GcFlags.generations; g++) {
731 for (s = 0; s < generations[g].n_steps; s++) {
732 checkHeap(generations[g].steps[s].blocks,
733 generations[g].steps[s].blocks->start);
734 checkChain(generations[g].steps[s].large_objects);
737 checkFreeListSanity();