1 /* -----------------------------------------------------------------------------
2 * $Id: Storage.c,v 1.30 2000/12/11 12:37:00 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());
199 /* -----------------------------------------------------------------------------
200 Setting the heap size. This function is callable from Haskell (GHC
201 uses it to implement the -H<size> option).
202 -------------------------------------------------------------------------- */
205 setHeapSize( HsInt size )
207 RtsFlags.GcFlags.heapSizeSuggestion = size / BLOCK_SIZE;
208 if (RtsFlags.GcFlags.heapSizeSuggestion >
209 RtsFlags.GcFlags.maxHeapSize) {
210 RtsFlags.GcFlags.maxHeapSize = RtsFlags.GcFlags.heapSizeSuggestion;
214 /* -----------------------------------------------------------------------------
216 -------------------------------------------------------------------------- */
219 newCAF(StgClosure* caf)
221 /* Put this CAF on the mutable list for the old generation.
222 * This is a HACK - the IND_STATIC closure doesn't really have
223 * a mut_link field, but we pretend it has - in fact we re-use
224 * the STATIC_LINK field for the time being, because when we
225 * come to do a major GC we won't need the mut_link field
226 * any more and can use it as a STATIC_LINK.
228 ACQUIRE_LOCK(&sm_mutex);
230 ASSERT( ((StgMutClosure*)caf)->mut_link == NULL );
231 ((StgMutClosure *)caf)->mut_link = oldest_gen->mut_once_list;
232 oldest_gen->mut_once_list = (StgMutClosure *)caf;
235 /* If we're Hugs, we also have to put it in the CAF table, so that
236 the CAF can be reverted. When reverting, CAFs created by compiled
237 code are recorded in the CAF table, which lives outside the
238 heap, in mallocville. CAFs created by interpreted code are
239 chained together via the link fields in StgCAFs, and are not
240 recorded in the CAF table.
242 ASSERT( get_itbl(caf)->type == THUNK_STATIC );
243 addToECafTable ( caf, get_itbl(caf) );
246 RELEASE_LOCK(&sm_mutex);
251 newCAF_made_by_Hugs(StgCAF* caf)
253 ACQUIRE_LOCK(&sm_mutex);
255 ASSERT( get_itbl(caf)->type == CAF_ENTERED );
256 recordOldToNewPtrs((StgMutClosure*)caf);
257 caf->link = ecafList;
258 ecafList = caf->link;
260 RELEASE_LOCK(&sm_mutex);
265 /* These initialisations are critical for correct operation
266 on the first call of addToECafTable.
268 StgCAF* ecafList = END_ECAF_LIST;
269 StgCAFTabEntry* ecafTable = NULL;
270 StgInt usedECafTable = 0;
271 StgInt sizeECafTable = 0;
274 void clearECafTable ( void )
279 void addToECafTable ( StgClosure* closure, StgInfoTable* origItbl )
283 if (usedECafTable == sizeECafTable) {
284 /* Make the initial table size be 8 */
286 if (sizeECafTable == 0) sizeECafTable = 8;
287 et2 = stgMallocBytes (
288 sizeECafTable * sizeof(StgCAFTabEntry),
290 for (i = 0; i < usedECafTable; i++)
291 et2[i] = ecafTable[i];
292 if (ecafTable) free(ecafTable);
295 ecafTable[usedECafTable].closure = closure;
296 ecafTable[usedECafTable].origItbl = origItbl;
301 /* -----------------------------------------------------------------------------
303 -------------------------------------------------------------------------- */
306 allocNurseries( void )
315 for (cap = free_capabilities; cap != NULL; cap = cap->link) {
316 cap->rNursery = allocNursery(NULL, RtsFlags.GcFlags.minAllocAreaSize);
317 cap->rCurrentNursery = cap->rNursery;
318 for (bd = cap->rNursery; bd != NULL; bd = bd->link) {
319 bd->back = (bdescr *)cap;
322 /* Set the back links to be equal to the Capability,
323 * so we can do slightly better informed locking.
327 nursery_blocks = RtsFlags.GcFlags.minAllocAreaSize;
328 g0s0->blocks = allocNursery(NULL, nursery_blocks);
329 g0s0->n_blocks = nursery_blocks;
330 g0s0->to_space = NULL;
331 MainRegTable.rNursery = g0s0->blocks;
332 MainRegTable.rCurrentNursery = g0s0->blocks;
333 /* hp, hpLim, hp_bd, to_space etc. aren't used in G0S0 */
338 resetNurseries( void )
344 /* All tasks must be stopped */
345 ASSERT(n_free_capabilities == RtsFlags.ParFlags.nNodes);
347 for (cap = free_capabilities; cap != NULL; cap = cap->link) {
348 for (bd = cap->rNursery; bd; bd = bd->link) {
349 bd->free = bd->start;
350 ASSERT(bd->gen == g0);
351 ASSERT(bd->step == g0s0);
352 IF_DEBUG(sanity,memset(bd->start, 0xaa, BLOCK_SIZE));
354 cap->rCurrentNursery = cap->rNursery;
357 for (bd = g0s0->blocks; bd; bd = bd->link) {
358 bd->free = bd->start;
359 ASSERT(bd->gen == g0);
360 ASSERT(bd->step == g0s0);
361 IF_DEBUG(sanity,memset(bd->start, 0xaa, BLOCK_SIZE));
363 MainRegTable.rNursery = g0s0->blocks;
364 MainRegTable.rCurrentNursery = g0s0->blocks;
369 allocNursery (bdescr *last_bd, nat blocks)
374 /* Allocate a nursery */
375 for (i=0; i < blocks; i++) {
381 bd->free = bd->start;
388 resizeNursery ( nat blocks )
393 barf("resizeNursery: can't resize in SMP mode");
396 if (nursery_blocks == blocks) {
397 ASSERT(g0s0->n_blocks == blocks);
401 else if (nursery_blocks < blocks) {
402 IF_DEBUG(gc, fprintf(stderr, "Increasing size of nursery to %d blocks\n",
404 g0s0->blocks = allocNursery(g0s0->blocks, blocks-nursery_blocks);
410 IF_DEBUG(gc, fprintf(stderr, "Decreasing size of nursery to %d blocks\n",
412 for (bd = g0s0->blocks; nursery_blocks > blocks; nursery_blocks--) {
420 g0s0->n_blocks = nursery_blocks = blocks;
423 /* -----------------------------------------------------------------------------
424 The allocate() interface
426 allocate(n) always succeeds, and returns a chunk of memory n words
427 long. n can be larger than the size of a block if necessary, in
428 which case a contiguous block group will be allocated.
429 -------------------------------------------------------------------------- */
437 ACQUIRE_LOCK(&sm_mutex);
439 TICK_ALLOC_HEAP_NOCTR(n);
442 /* big allocation (>LARGE_OBJECT_THRESHOLD) */
443 /* ToDo: allocate directly into generation 1 */
444 if (n >= LARGE_OBJECT_THRESHOLD/sizeof(W_)) {
445 nat req_blocks = (lnat)BLOCK_ROUND_UP(n*sizeof(W_)) / BLOCK_SIZE;
446 bd = allocGroup(req_blocks);
447 dbl_link_onto(bd, &g0s0->large_objects);
451 bd->free = bd->start;
452 /* don't add these blocks to alloc_blocks, since we're assuming
453 * that large objects are likely to remain live for quite a while
454 * (eg. running threads), so garbage collecting early won't make
457 alloc_blocks += req_blocks;
458 RELEASE_LOCK(&sm_mutex);
461 /* small allocation (<LARGE_OBJECT_THRESHOLD) */
462 } else if (small_alloc_list == NULL || alloc_Hp + n > alloc_HpLim) {
463 if (small_alloc_list) {
464 small_alloc_list->free = alloc_Hp;
467 bd->link = small_alloc_list;
468 small_alloc_list = bd;
472 alloc_Hp = bd->start;
473 alloc_HpLim = bd->start + BLOCK_SIZE_W;
479 RELEASE_LOCK(&sm_mutex);
483 lnat allocated_bytes(void)
485 return (alloc_blocks * BLOCK_SIZE_W - (alloc_HpLim - alloc_Hp));
488 /* -----------------------------------------------------------------------------
489 Allocation functions for GMP.
491 These all use the allocate() interface - we can't have any garbage
492 collection going on during a gmp operation, so we use allocate()
493 which always succeeds. The gmp operations which might need to
494 allocate will ask the storage manager (via doYouWantToGC()) whether
495 a garbage collection is required, in case we get into a loop doing
496 only allocate() style allocation.
497 -------------------------------------------------------------------------- */
500 stgAllocForGMP (size_t size_in_bytes)
503 nat data_size_in_words, total_size_in_words;
505 /* should be a multiple of sizeof(StgWord) (whole no. of limbs) */
506 ASSERT(size_in_bytes % sizeof(W_) == 0);
508 data_size_in_words = size_in_bytes / sizeof(W_);
509 total_size_in_words = sizeofW(StgArrWords) + data_size_in_words;
511 /* allocate and fill it in. */
512 arr = (StgArrWords *)allocate(total_size_in_words);
513 SET_ARR_HDR(arr, &stg_ARR_WORDS_info, CCCS, data_size_in_words);
515 /* and return a ptr to the goods inside the array */
516 return(BYTE_ARR_CTS(arr));
520 stgReallocForGMP (void *ptr, size_t old_size, size_t new_size)
522 void *new_stuff_ptr = stgAllocForGMP(new_size);
524 char *p = (char *) ptr;
525 char *q = (char *) new_stuff_ptr;
527 for (; i < old_size; i++, p++, q++) {
531 return(new_stuff_ptr);
535 stgDeallocForGMP (void *ptr STG_UNUSED,
536 size_t size STG_UNUSED)
538 /* easy for us: the garbage collector does the dealloc'n */
541 /* -----------------------------------------------------------------------------
543 * -------------------------------------------------------------------------- */
545 /* -----------------------------------------------------------------------------
548 * Approximate how much we've allocated: number of blocks in the
549 * nursery + blocks allocated via allocate() - unused nusery blocks.
550 * This leaves a little slop at the end of each block, and doesn't
551 * take into account large objects (ToDo).
552 * -------------------------------------------------------------------------- */
555 calcAllocated( void )
563 /* All tasks must be stopped. Can't assert that all the
564 capabilities are owned by the scheduler, though: one or more
565 tasks might have been stopped while they were running (non-main)
567 /* ASSERT(n_free_capabilities == RtsFlags.ParFlags.nNodes); */
570 n_free_capabilities * RtsFlags.GcFlags.minAllocAreaSize * BLOCK_SIZE_W
573 for (cap = free_capabilities; cap != NULL; cap = cap->link) {
574 for ( bd = cap->rCurrentNursery->link; bd != NULL; bd = bd->link ) {
575 allocated -= BLOCK_SIZE_W;
577 if (cap->rCurrentNursery->free < cap->rCurrentNursery->start
579 allocated -= (cap->rCurrentNursery->start + BLOCK_SIZE_W)
580 - cap->rCurrentNursery->free;
585 bdescr *current_nursery = MainRegTable.rCurrentNursery;
587 allocated = (nursery_blocks * BLOCK_SIZE_W) + allocated_bytes();
588 for ( bd = current_nursery->link; bd != NULL; bd = bd->link ) {
589 allocated -= BLOCK_SIZE_W;
591 if (current_nursery->free < current_nursery->start + BLOCK_SIZE_W) {
592 allocated -= (current_nursery->start + BLOCK_SIZE_W)
593 - current_nursery->free;
597 total_allocated += allocated;
601 /* Approximate the amount of live data in the heap. To be called just
602 * after garbage collection (see GarbageCollect()).
611 if (RtsFlags.GcFlags.generations == 1) {
612 live = (g0s0->to_blocks - 1) * BLOCK_SIZE_W +
613 ((lnat)g0s0->hp_bd->free - (lnat)g0s0->hp_bd->start) / sizeof(W_);
617 for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
618 for (s = 0; s < generations[g].n_steps; s++) {
619 /* approximate amount of live data (doesn't take into account slop
620 * at end of each block).
622 if (g == 0 && s == 0) {
625 step = &generations[g].steps[s];
626 live += (step->n_blocks - 1) * BLOCK_SIZE_W +
627 ((lnat)step->hp_bd->free - (lnat)step->hp_bd->start) / sizeof(W_);
633 /* Approximate the number of blocks that will be needed at the next
634 * garbage collection.
636 * Assume: all data currently live will remain live. Steps that will
637 * be collected next time will therefore need twice as many blocks
638 * since all the data will be copied.
647 for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
648 for (s = 0; s < generations[g].n_steps; s++) {
649 if (g == 0 && s == 0) { continue; }
650 step = &generations[g].steps[s];
651 if (generations[g].steps[0].n_blocks > generations[g].max_blocks) {
652 needed += 2 * step->n_blocks;
654 needed += step->n_blocks;
661 /* -----------------------------------------------------------------------------
664 memInventory() checks for memory leaks by counting up all the
665 blocks we know about and comparing that to the number of blocks
666 allegedly floating around in the system.
667 -------------------------------------------------------------------------- */
677 lnat total_blocks = 0, free_blocks = 0;
679 /* count the blocks we current have */
681 for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
682 for (s = 0; s < generations[g].n_steps; s++) {
683 step = &generations[g].steps[s];
684 total_blocks += step->n_blocks;
685 if (RtsFlags.GcFlags.generations == 1) {
686 /* two-space collector has a to-space too :-) */
687 total_blocks += g0s0->to_blocks;
689 for (bd = step->large_objects; bd; bd = bd->link) {
690 total_blocks += bd->blocks;
691 /* hack for megablock groups: they have an extra block or two in
692 the second and subsequent megablocks where the block
693 descriptors would normally go.
695 if (bd->blocks > BLOCKS_PER_MBLOCK) {
696 total_blocks -= (MBLOCK_SIZE / BLOCK_SIZE - BLOCKS_PER_MBLOCK)
697 * (bd->blocks/(MBLOCK_SIZE/BLOCK_SIZE));
703 /* any blocks held by allocate() */
704 for (bd = small_alloc_list; bd; bd = bd->link) {
705 total_blocks += bd->blocks;
707 for (bd = large_alloc_list; bd; bd = bd->link) {
708 total_blocks += bd->blocks;
711 /* count the blocks on the free list */
712 free_blocks = countFreeList();
714 ASSERT(total_blocks + free_blocks == mblocks_allocated * BLOCKS_PER_MBLOCK);
717 if (total_blocks + free_blocks != mblocks_allocated *
719 fprintf(stderr, "Blocks: %ld live + %ld free = %ld total (%ld around)\n",
720 total_blocks, free_blocks, total_blocks + free_blocks,
721 mblocks_allocated * BLOCKS_PER_MBLOCK);
726 /* Full heap sanity check. */
733 if (RtsFlags.GcFlags.generations == 1) {
734 checkHeap(g0s0->to_space, NULL);
735 checkChain(g0s0->large_objects);
738 for (g = 0; g <= N; g++) {
739 for (s = 0; s < generations[g].n_steps; s++) {
740 if (g == 0 && s == 0) { continue; }
741 checkHeap(generations[g].steps[s].blocks, NULL);
744 for (g = N+1; g < RtsFlags.GcFlags.generations; g++) {
745 for (s = 0; s < generations[g].n_steps; s++) {
746 checkHeap(generations[g].steps[s].blocks,
747 generations[g].steps[s].blocks->start);
748 checkChain(generations[g].steps[s].large_objects);
751 checkFreeListSanity();