1 /* -----------------------------------------------------------------------------
2 * $Id: Storage.c,v 1.21 1999/11/09 15:46:59 simonmar Exp $
4 * (c) The GHC Team, 1998-1999
6 * Storage manager front end
8 * ---------------------------------------------------------------------------*/
15 #include "BlockAlloc.h"
23 #include "StoragePriv.h"
26 nat nursery_blocks; /* number of blocks in the nursery */
29 StgClosure *caf_list = NULL;
31 bdescr *small_alloc_list; /* allocate()d small objects */
32 bdescr *large_alloc_list; /* allocate()d large objects */
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 */
45 * Storage manager mutex: protects all the above state from
46 * simultaneous access by two STG threads.
49 pthread_mutex_t sm_mutex = PTHREAD_MUTEX_INITIALIZER;
55 static void *stgAllocForGMP (size_t size_in_bytes);
56 static void *stgReallocForGMP (void *ptr, size_t old_size, size_t new_size);
57 static void stgDeallocForGMP (void *ptr, size_t size);
66 /* If we're doing heap profiling, we want a two-space heap with a
67 * fixed-size allocation area so that we get roughly even-spaced
70 #if defined(PROFILING) || defined(DEBUG)
71 if (RtsFlags.ProfFlags.doHeapProfile) {
72 RtsFlags.GcFlags.generations = 1;
73 RtsFlags.GcFlags.steps = 1;
74 RtsFlags.GcFlags.oldGenFactor = 0;
75 RtsFlags.GcFlags.heapSizeSuggestion = 0;
79 if (RtsFlags.GcFlags.heapSizeSuggestion >
80 RtsFlags.GcFlags.maxHeapSize) {
81 RtsFlags.GcFlags.maxHeapSize = RtsFlags.GcFlags.heapSizeSuggestion;
86 /* allocate generation info array */
87 generations = (generation *)stgMallocBytes(RtsFlags.GcFlags.generations
88 * sizeof(struct _generation),
91 /* Initialise all generations */
92 for(g = 0; g < RtsFlags.GcFlags.generations; g++) {
93 gen = &generations[g];
95 gen->mut_list = END_MUT_LIST;
96 gen->mut_once_list = END_MUT_LIST;
98 gen->failed_promotions = 0;
102 /* A couple of convenience pointers */
103 g0 = &generations[0];
104 oldest_gen = &generations[RtsFlags.GcFlags.generations-1];
106 /* Allocate step structures in each generation */
107 if (RtsFlags.GcFlags.generations > 1) {
108 /* Only for multiple-generations */
110 /* Oldest generation: one step */
111 oldest_gen->n_steps = 1;
113 stgMallocBytes(1 * sizeof(struct _step), "initStorage: last step");
115 /* set up all except the oldest generation with 2 steps */
116 for(g = 0; g < RtsFlags.GcFlags.generations-1; g++) {
117 generations[g].n_steps = RtsFlags.GcFlags.steps;
118 generations[g].steps =
119 stgMallocBytes (RtsFlags.GcFlags.steps * sizeof(struct _step),
120 "initStorage: steps");
124 /* single generation, i.e. a two-space collector */
126 g0->steps = stgMallocBytes (sizeof(struct _step), "initStorage: steps");
129 /* Initialise all steps */
130 for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
131 for (s = 0; s < generations[g].n_steps; s++) {
132 step = &generations[g].steps[s];
136 step->gen = &generations[g];
141 step->scan_bd = NULL;
142 step->large_objects = NULL;
143 step->new_large_objects = NULL;
144 step->scavenged_large_objects = NULL;
148 /* Set up the destination pointers in each younger gen. step */
149 for (g = 0; g < RtsFlags.GcFlags.generations-1; g++) {
150 for (s = 0; s < generations[g].n_steps-1; s++) {
151 generations[g].steps[s].to = &generations[g].steps[s+1];
153 generations[g].steps[s].to = &generations[g+1].steps[0];
156 /* The oldest generation has one step and its destination is the
158 oldest_gen->steps[0].to = &oldest_gen->steps[0];
160 /* generation 0 is special: that's the nursery */
161 generations[0].max_blocks = 0;
163 /* G0S0: the allocation area. Policy: keep the allocation area
164 * small to begin with, even if we have a large suggested heap
165 * size. Reason: we're going to do a major collection first, and we
166 * don't want it to be a big one. This vague idea is borne out by
167 * rigorous experimental evidence.
169 g0s0 = &generations[0].steps[0];
173 weak_ptr_list = NULL;
176 /* initialise the allocate() interface */
177 small_alloc_list = NULL;
178 large_alloc_list = NULL;
180 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);
188 pthread_mutex_init(&sm_mutex, NULL);
191 IF_DEBUG(gc, stat_describe_gens());
197 stat_exit(calcAllocated());
201 newCAF(StgClosure* caf)
203 /* Put this CAF on the mutable list for the old generation.
204 * This is a HACK - the IND_STATIC closure doesn't really have
205 * a mut_link field, but we pretend it has - in fact we re-use
206 * the STATIC_LINK field for the time being, because when we
207 * come to do a major GC we won't need the mut_link field
208 * any more and can use it as a STATIC_LINK.
210 ACQUIRE_LOCK(&sm_mutex);
211 ((StgMutClosure *)caf)->mut_link = oldest_gen->mut_once_list;
212 oldest_gen->mut_once_list = (StgMutClosure *)caf;
216 const StgInfoTable *info;
218 info = get_itbl(caf);
219 ASSERT(info->type == IND_STATIC);
221 STATIC_LINK2(info,caf) = caf_list;
226 RELEASE_LOCK(&sm_mutex);
229 /* -----------------------------------------------------------------------------
231 -------------------------------------------------------------------------- */
234 allocNurseries( void )
243 for (cap = free_capabilities; cap != NULL; cap = cap->link) {
244 cap->rNursery = allocNursery(NULL, RtsFlags.GcFlags.minAllocAreaSize);
245 cap->rCurrentNursery = cap->rNursery;
246 for (bd = cap->rNursery; bd != NULL; bd = bd->link) {
247 bd->back = (bdescr *)cap;
250 /* Set the back links to be equal to the Capability,
251 * so we can do slightly better informed locking.
255 nursery_blocks = RtsFlags.GcFlags.minAllocAreaSize;
256 g0s0->blocks = allocNursery(NULL, nursery_blocks);
257 g0s0->n_blocks = nursery_blocks;
258 g0s0->to_space = NULL;
259 MainRegTable.rNursery = g0s0->blocks;
260 MainRegTable.rCurrentNursery = g0s0->blocks;
261 /* hp, hpLim, hp_bd, to_space etc. aren't used in G0S0 */
266 resetNurseries( void )
272 /* All tasks must be stopped */
273 ASSERT(n_free_capabilities == RtsFlags.ConcFlags.nNodes);
275 for (cap = free_capabilities; cap != NULL; cap = cap->link) {
276 for (bd = cap->rNursery; bd; bd = bd->link) {
277 bd->free = bd->start;
278 ASSERT(bd->gen == g0);
279 ASSERT(bd->step == g0s0);
280 IF_DEBUG(sanity,memset(bd->start, 0xaa, BLOCK_SIZE));
282 cap->rCurrentNursery = cap->rNursery;
285 for (bd = g0s0->blocks; bd; bd = bd->link) {
286 bd->free = bd->start;
287 ASSERT(bd->gen == g0);
288 ASSERT(bd->step == g0s0);
289 IF_DEBUG(sanity,memset(bd->start, 0xaa, BLOCK_SIZE));
291 MainRegTable.rNursery = g0s0->blocks;
292 MainRegTable.rCurrentNursery = g0s0->blocks;
297 allocNursery (bdescr *last_bd, nat blocks)
302 /* Allocate a nursery */
303 for (i=0; i < blocks; i++) {
309 bd->free = bd->start;
316 resizeNursery ( nat blocks )
321 barf("resizeNursery: can't resize in SMP mode");
324 if (nursery_blocks == blocks) {
325 ASSERT(g0s0->n_blocks == blocks);
329 else if (nursery_blocks < blocks) {
330 IF_DEBUG(gc, fprintf(stderr, "Increasing size of nursery to %d blocks\n",
332 g0s0->blocks = allocNursery(g0s0->blocks, blocks-nursery_blocks);
338 IF_DEBUG(gc, fprintf(stderr, "Decreasing size of nursery to %d blocks\n",
340 for (bd = g0s0->blocks; nursery_blocks > blocks; nursery_blocks--) {
348 g0s0->n_blocks = nursery_blocks = blocks;
351 /* -----------------------------------------------------------------------------
352 The allocate() interface
354 allocate(n) always succeeds, and returns a chunk of memory n words
355 long. n can be larger than the size of a block if necessary, in
356 which case a contiguous block group will be allocated.
357 -------------------------------------------------------------------------- */
365 ACQUIRE_LOCK(&sm_mutex);
367 TICK_ALLOC_HEAP_NOCTR(n);
370 /* big allocation (>LARGE_OBJECT_THRESHOLD) */
371 /* ToDo: allocate directly into generation 1 */
372 if (n >= LARGE_OBJECT_THRESHOLD/sizeof(W_)) {
373 nat req_blocks = (lnat)BLOCK_ROUND_UP(n*sizeof(W_)) / BLOCK_SIZE;
374 bd = allocGroup(req_blocks);
375 dbl_link_onto(bd, &g0s0->large_objects);
379 bd->free = bd->start;
380 /* don't add these blocks to alloc_blocks, since we're assuming
381 * that large objects are likely to remain live for quite a while
382 * (eg. running threads), so garbage collecting early won't make
385 RELEASE_LOCK(&sm_mutex);
388 /* small allocation (<LARGE_OBJECT_THRESHOLD) */
389 } else if (small_alloc_list == NULL || alloc_Hp + n > alloc_HpLim) {
390 if (small_alloc_list) {
391 small_alloc_list->free = alloc_Hp;
394 bd->link = small_alloc_list;
395 small_alloc_list = bd;
399 alloc_Hp = bd->start;
400 alloc_HpLim = bd->start + BLOCK_SIZE_W;
406 RELEASE_LOCK(&sm_mutex);
410 lnat allocated_bytes(void)
412 return (alloc_blocks * BLOCK_SIZE_W - (alloc_HpLim - alloc_Hp));
415 /* -----------------------------------------------------------------------------
416 Allocation functions for GMP.
418 These all use the allocate() interface - we can't have any garbage
419 collection going on during a gmp operation, so we use allocate()
420 which always succeeds. The gmp operations which might need to
421 allocate will ask the storage manager (via doYouWantToGC()) whether
422 a garbage collection is required, in case we get into a loop doing
423 only allocate() style allocation.
424 -------------------------------------------------------------------------- */
427 stgAllocForGMP (size_t size_in_bytes)
430 nat data_size_in_words, total_size_in_words;
432 /* should be a multiple of sizeof(StgWord) (whole no. of limbs) */
433 ASSERT(size_in_bytes % sizeof(W_) == 0);
435 data_size_in_words = size_in_bytes / sizeof(W_);
436 total_size_in_words = sizeofW(StgArrWords) + data_size_in_words;
438 /* allocate and fill it in. */
439 arr = (StgArrWords *)allocate(total_size_in_words);
440 SET_ARR_HDR(arr, &ARR_WORDS_info, CCCS, data_size_in_words);
442 /* and return a ptr to the goods inside the array */
443 return(BYTE_ARR_CTS(arr));
447 stgReallocForGMP (void *ptr, size_t old_size, size_t new_size)
449 void *new_stuff_ptr = stgAllocForGMP(new_size);
451 char *p = (char *) ptr;
452 char *q = (char *) new_stuff_ptr;
454 for (; i < old_size; i++, p++, q++) {
458 return(new_stuff_ptr);
462 stgDeallocForGMP (void *ptr STG_UNUSED,
463 size_t size STG_UNUSED)
465 /* easy for us: the garbage collector does the dealloc'n */
468 /* -----------------------------------------------------------------------------
470 * -------------------------------------------------------------------------- */
472 /* -----------------------------------------------------------------------------
475 * Approximate how much we've allocated: number of blocks in the
476 * nursery + blocks allocated via allocate() - unused nusery blocks.
477 * This leaves a little slop at the end of each block, and doesn't
478 * take into account large objects (ToDo).
479 * -------------------------------------------------------------------------- */
482 calcAllocated( void )
490 /* All tasks must be stopped. Can't assert that all the
491 capabilities are owned by the scheduler, though: one or more
492 tasks might have been stopped while they were running (non-main)
494 /* ASSERT(n_free_capabilities == RtsFlags.ConcFlags.nNodes); */
497 n_free_capabilities * RtsFlags.GcFlags.minAllocAreaSize * BLOCK_SIZE_W
500 for (cap = free_capabilities; cap != NULL; cap = cap->link) {
501 for ( bd = cap->rCurrentNursery->link; bd != NULL; bd = bd->link ) {
502 allocated -= BLOCK_SIZE_W;
504 if (cap->rCurrentNursery->free < cap->rCurrentNursery->start
506 allocated -= (cap->rCurrentNursery->start + BLOCK_SIZE_W)
507 - cap->rCurrentNursery->free;
512 bdescr *current_nursery = MainRegTable.rCurrentNursery;
514 allocated = (nursery_blocks * BLOCK_SIZE_W) + allocated_bytes();
515 for ( bd = current_nursery->link; bd != NULL; bd = bd->link ) {
516 allocated -= BLOCK_SIZE_W;
518 if (current_nursery->free < current_nursery->start + BLOCK_SIZE_W) {
519 allocated -= (current_nursery->start + BLOCK_SIZE_W)
520 - current_nursery->free;
527 /* Approximate the amount of live data in the heap. To be called just
528 * after garbage collection (see GarbageCollect()).
537 if (RtsFlags.GcFlags.generations == 1) {
538 live = (g0s0->to_blocks - 1) * BLOCK_SIZE_W +
539 ((lnat)g0s0->hp_bd->free - (lnat)g0s0->hp_bd->start) / sizeof(W_);
543 for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
544 for (s = 0; s < generations[g].n_steps; s++) {
545 /* approximate amount of live data (doesn't take into account slop
546 * at end of each block).
548 if (g == 0 && s == 0) {
551 step = &generations[g].steps[s];
552 live += (step->n_blocks - 1) * BLOCK_SIZE_W +
553 ((lnat)step->hp_bd->free - (lnat)step->hp_bd->start) / sizeof(W_);
559 /* Approximate the number of blocks that will be needed at the next
560 * garbage collection.
562 * Assume: all data currently live will remain live. Steps that will
563 * be collected next time will therefore need twice as many blocks
564 * since all the data will be copied.
573 for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
574 for (s = 0; s < generations[g].n_steps; s++) {
575 if (g == 0 && s == 0) { continue; }
576 step = &generations[g].steps[s];
577 if (generations[g].steps[0].n_blocks > generations[g].max_blocks) {
578 needed += 2 * step->n_blocks;
580 needed += step->n_blocks;
587 /* -----------------------------------------------------------------------------
590 memInventory() checks for memory leaks by counting up all the
591 blocks we know about and comparing that to the number of blocks
592 allegedly floating around in the system.
593 -------------------------------------------------------------------------- */
603 lnat total_blocks = 0, free_blocks = 0;
605 /* count the blocks we current have */
607 for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
608 for (s = 0; s < generations[g].n_steps; s++) {
609 step = &generations[g].steps[s];
610 total_blocks += step->n_blocks;
611 if (RtsFlags.GcFlags.generations == 1) {
612 /* two-space collector has a to-space too :-) */
613 total_blocks += g0s0->to_blocks;
615 for (bd = step->large_objects; bd; bd = bd->link) {
616 total_blocks += bd->blocks;
617 /* hack for megablock groups: they have an extra block or two in
618 the second and subsequent megablocks where the block
619 descriptors would normally go.
621 if (bd->blocks > BLOCKS_PER_MBLOCK) {
622 total_blocks -= (MBLOCK_SIZE / BLOCK_SIZE - BLOCKS_PER_MBLOCK)
623 * (bd->blocks/(MBLOCK_SIZE/BLOCK_SIZE));
629 /* any blocks held by allocate() */
630 for (bd = small_alloc_list; bd; bd = bd->link) {
631 total_blocks += bd->blocks;
633 for (bd = large_alloc_list; bd; bd = bd->link) {
634 total_blocks += bd->blocks;
637 /* count the blocks on the free list */
638 free_blocks = countFreeList();
640 ASSERT(total_blocks + free_blocks == mblocks_allocated * BLOCKS_PER_MBLOCK);
643 if (total_blocks + free_blocks != mblocks_allocated *
645 fprintf(stderr, "Blocks: %ld live + %ld free = %ld total (%ld around)\n",
646 total_blocks, free_blocks, total_blocks + free_blocks,
647 mblocks_allocated * BLOCKS_PER_MBLOCK);
652 /* Full heap sanity check. */
659 if (RtsFlags.GcFlags.generations == 1) {
660 checkHeap(g0s0->to_space, NULL);
661 checkChain(g0s0->large_objects);
664 for (g = 0; g <= N; g++) {
665 for (s = 0; s < generations[g].n_steps; s++) {
666 if (g == 0 && s == 0) { continue; }
667 checkHeap(generations[g].steps[s].blocks, NULL);
670 for (g = N+1; g < RtsFlags.GcFlags.generations; g++) {
671 for (s = 0; s < generations[g].n_steps; s++) {
672 checkHeap(generations[g].steps[s].blocks,
673 generations[g].steps[s].blocks->start);
674 checkChain(generations[g].steps[s].large_objects);
677 checkFreeListSanity();