1 /* -----------------------------------------------------------------------------
2 * $Id: Storage.c,v 1.35 2001/01/31 11:04:29 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 stp = &generations[g].steps[s];
137 stp->gen = &generations[g];
143 stp->large_objects = NULL;
144 stp->new_large_objects = NULL;
145 stp->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 /* -----------------------------------------------------------------------------
202 The entry code for every CAF does the following:
204 - builds a CAF_BLACKHOLE in the heap
205 - pushes an update frame pointing to the CAF_BLACKHOLE
206 - invokes UPD_CAF(), which:
207 - calls newCaf, below
208 - updates the CAF with a static indirection to the CAF_BLACKHOLE
210 Why do we build a BLACKHOLE in the heap rather than just updating
211 the thunk directly? It's so that we only need one kind of update
212 frame - otherwise we'd need a static version of the update frame too.
214 newCaf() does the following:
216 - it puts the CAF on the oldest generation's mut-once list.
217 This is so that we can treat the CAF as a root when collecting
220 For GHCI, we have additional requirements when dealing with CAFs:
222 - we must *retain* all dynamically-loaded CAFs ever entered,
223 just in case we need them again.
224 - we must be able to *revert* CAFs that have been evaluated, to
225 their pre-evaluated form.
227 To do this, we use an additional CAF list. When newCaf() is
228 called on a dynamically-loaded CAF, we add it to the CAF list
229 instead of the old-generation mutable list, and save away its
230 old info pointer (in caf->saved_info) for later reversion.
232 To revert all the CAFs, we traverse the CAF list and reset the
233 info pointer to caf->saved_info, then throw away the CAF list.
234 (see GC.c:revertCAFs()).
238 -------------------------------------------------------------------------- */
241 newCAF(StgClosure* caf)
243 /* Put this CAF on the mutable list for the old generation.
244 * This is a HACK - the IND_STATIC closure doesn't really have
245 * a mut_link field, but we pretend it has - in fact we re-use
246 * the STATIC_LINK field for the time being, because when we
247 * come to do a major GC we won't need the mut_link field
248 * any more and can use it as a STATIC_LINK.
250 ACQUIRE_LOCK(&sm_mutex);
253 if (is_dynamically_loaded_rwdata_ptr((StgPtr)caf)) {
254 ((StgIndStatic *)caf)->saved_info = (StgInfoTable *)caf->header.info;
255 ((StgIndStatic *)caf)->static_link = caf_list;
258 ((StgIndStatic *)caf)->saved_info = NULL;
259 ((StgMutClosure *)caf)->mut_link = oldest_gen->mut_once_list;
260 oldest_gen->mut_once_list = (StgMutClosure *)caf;
263 ASSERT( ((StgMutClosure*)caf)->mut_link == NULL );
264 ((StgMutClosure *)caf)->mut_link = oldest_gen->mut_once_list;
265 oldest_gen->mut_once_list = (StgMutClosure *)caf;
268 RELEASE_LOCK(&sm_mutex);
271 /* -----------------------------------------------------------------------------
273 -------------------------------------------------------------------------- */
276 allocNurseries( void )
285 for (cap = free_capabilities; cap != NULL; cap = cap->link) {
286 cap->rNursery = allocNursery(NULL, RtsFlags.GcFlags.minAllocAreaSize);
287 cap->rCurrentNursery = cap->rNursery;
288 for (bd = cap->rNursery; bd != NULL; bd = bd->link) {
289 bd->back = (bdescr *)cap;
292 /* Set the back links to be equal to the Capability,
293 * so we can do slightly better informed locking.
297 nursery_blocks = RtsFlags.GcFlags.minAllocAreaSize;
298 g0s0->blocks = allocNursery(NULL, nursery_blocks);
299 g0s0->n_blocks = nursery_blocks;
300 g0s0->to_space = NULL;
301 MainRegTable.rNursery = g0s0->blocks;
302 MainRegTable.rCurrentNursery = g0s0->blocks;
303 /* hp, hpLim, hp_bd, to_space etc. aren't used in G0S0 */
308 resetNurseries( void )
314 /* All tasks must be stopped */
315 ASSERT(n_free_capabilities == RtsFlags.ParFlags.nNodes);
317 for (cap = free_capabilities; cap != NULL; cap = cap->link) {
318 for (bd = cap->rNursery; bd; bd = bd->link) {
319 bd->free = bd->start;
320 ASSERT(bd->gen == g0);
321 ASSERT(bd->step == g0s0);
322 IF_DEBUG(sanity,memset(bd->start, 0xaa, BLOCK_SIZE));
324 cap->rCurrentNursery = cap->rNursery;
327 for (bd = g0s0->blocks; bd; bd = bd->link) {
328 bd->free = bd->start;
329 ASSERT(bd->gen == g0);
330 ASSERT(bd->step == g0s0);
331 IF_DEBUG(sanity,memset(bd->start, 0xaa, BLOCK_SIZE));
333 MainRegTable.rNursery = g0s0->blocks;
334 MainRegTable.rCurrentNursery = g0s0->blocks;
339 allocNursery (bdescr *last_bd, nat blocks)
344 /* Allocate a nursery */
345 for (i=0; i < blocks; i++) {
351 bd->free = bd->start;
358 resizeNursery ( nat blocks )
363 barf("resizeNursery: can't resize in SMP mode");
366 if (nursery_blocks == blocks) {
367 ASSERT(g0s0->n_blocks == blocks);
371 else if (nursery_blocks < blocks) {
372 IF_DEBUG(gc, fprintf(stderr, "Increasing size of nursery to %d blocks\n",
374 g0s0->blocks = allocNursery(g0s0->blocks, blocks-nursery_blocks);
380 IF_DEBUG(gc, fprintf(stderr, "Decreasing size of nursery to %d blocks\n",
382 for (bd = g0s0->blocks; nursery_blocks > blocks; nursery_blocks--) {
390 g0s0->n_blocks = nursery_blocks = blocks;
393 /* -----------------------------------------------------------------------------
394 The allocate() interface
396 allocate(n) always succeeds, and returns a chunk of memory n words
397 long. n can be larger than the size of a block if necessary, in
398 which case a contiguous block group will be allocated.
399 -------------------------------------------------------------------------- */
407 ACQUIRE_LOCK(&sm_mutex);
409 TICK_ALLOC_HEAP_NOCTR(n);
412 /* big allocation (>LARGE_OBJECT_THRESHOLD) */
413 /* ToDo: allocate directly into generation 1 */
414 if (n >= LARGE_OBJECT_THRESHOLD/sizeof(W_)) {
415 nat req_blocks = (lnat)BLOCK_ROUND_UP(n*sizeof(W_)) / BLOCK_SIZE;
416 bd = allocGroup(req_blocks);
417 dbl_link_onto(bd, &g0s0->large_objects);
421 bd->free = bd->start;
422 /* don't add these blocks to alloc_blocks, since we're assuming
423 * that large objects are likely to remain live for quite a while
424 * (eg. running threads), so garbage collecting early won't make
427 alloc_blocks += req_blocks;
428 RELEASE_LOCK(&sm_mutex);
431 /* small allocation (<LARGE_OBJECT_THRESHOLD) */
432 } else if (small_alloc_list == NULL || alloc_Hp + n > alloc_HpLim) {
433 if (small_alloc_list) {
434 small_alloc_list->free = alloc_Hp;
437 bd->link = small_alloc_list;
438 small_alloc_list = bd;
442 alloc_Hp = bd->start;
443 alloc_HpLim = bd->start + BLOCK_SIZE_W;
449 RELEASE_LOCK(&sm_mutex);
453 lnat allocated_bytes(void)
455 return (alloc_blocks * BLOCK_SIZE_W - (alloc_HpLim - alloc_Hp));
458 /* -----------------------------------------------------------------------------
459 Allocation functions for GMP.
461 These all use the allocate() interface - we can't have any garbage
462 collection going on during a gmp operation, so we use allocate()
463 which always succeeds. The gmp operations which might need to
464 allocate will ask the storage manager (via doYouWantToGC()) whether
465 a garbage collection is required, in case we get into a loop doing
466 only allocate() style allocation.
467 -------------------------------------------------------------------------- */
470 stgAllocForGMP (size_t size_in_bytes)
473 nat data_size_in_words, total_size_in_words;
475 /* should be a multiple of sizeof(StgWord) (whole no. of limbs) */
476 ASSERT(size_in_bytes % sizeof(W_) == 0);
478 data_size_in_words = size_in_bytes / sizeof(W_);
479 total_size_in_words = sizeofW(StgArrWords) + data_size_in_words;
481 /* allocate and fill it in. */
482 arr = (StgArrWords *)allocate(total_size_in_words);
483 SET_ARR_HDR(arr, &stg_ARR_WORDS_info, CCCS, data_size_in_words);
485 /* and return a ptr to the goods inside the array */
486 return(BYTE_ARR_CTS(arr));
490 stgReallocForGMP (void *ptr, size_t old_size, size_t new_size)
492 void *new_stuff_ptr = stgAllocForGMP(new_size);
494 char *p = (char *) ptr;
495 char *q = (char *) new_stuff_ptr;
497 for (; i < old_size; i++, p++, q++) {
501 return(new_stuff_ptr);
505 stgDeallocForGMP (void *ptr STG_UNUSED,
506 size_t size STG_UNUSED)
508 /* easy for us: the garbage collector does the dealloc'n */
511 /* -----------------------------------------------------------------------------
513 * -------------------------------------------------------------------------- */
515 /* -----------------------------------------------------------------------------
518 * Approximate how much we've allocated: number of blocks in the
519 * nursery + blocks allocated via allocate() - unused nusery blocks.
520 * This leaves a little slop at the end of each block, and doesn't
521 * take into account large objects (ToDo).
522 * -------------------------------------------------------------------------- */
525 calcAllocated( void )
533 /* All tasks must be stopped. Can't assert that all the
534 capabilities are owned by the scheduler, though: one or more
535 tasks might have been stopped while they were running (non-main)
537 /* ASSERT(n_free_capabilities == RtsFlags.ParFlags.nNodes); */
540 n_free_capabilities * RtsFlags.GcFlags.minAllocAreaSize * BLOCK_SIZE_W
543 for (cap = free_capabilities; cap != NULL; cap = cap->link) {
544 for ( bd = cap->rCurrentNursery->link; bd != NULL; bd = bd->link ) {
545 allocated -= BLOCK_SIZE_W;
547 if (cap->rCurrentNursery->free < cap->rCurrentNursery->start
549 allocated -= (cap->rCurrentNursery->start + BLOCK_SIZE_W)
550 - cap->rCurrentNursery->free;
555 bdescr *current_nursery = MainRegTable.rCurrentNursery;
557 allocated = (nursery_blocks * BLOCK_SIZE_W) + allocated_bytes();
558 for ( bd = current_nursery->link; bd != NULL; bd = bd->link ) {
559 allocated -= BLOCK_SIZE_W;
561 if (current_nursery->free < current_nursery->start + BLOCK_SIZE_W) {
562 allocated -= (current_nursery->start + BLOCK_SIZE_W)
563 - current_nursery->free;
567 total_allocated += allocated;
571 /* Approximate the amount of live data in the heap. To be called just
572 * after garbage collection (see GarbageCollect()).
581 if (RtsFlags.GcFlags.generations == 1) {
582 live = (g0s0->to_blocks - 1) * BLOCK_SIZE_W +
583 ((lnat)g0s0->hp_bd->free - (lnat)g0s0->hp_bd->start) / sizeof(W_);
587 for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
588 for (s = 0; s < generations[g].n_steps; s++) {
589 /* approximate amount of live data (doesn't take into account slop
590 * at end of each block).
592 if (g == 0 && s == 0) {
595 stp = &generations[g].steps[s];
596 live += (stp->n_blocks - 1) * BLOCK_SIZE_W +
597 ((lnat)stp->hp_bd->free - (lnat)stp->hp_bd->start) / sizeof(W_);
603 /* Approximate the number of blocks that will be needed at the next
604 * garbage collection.
606 * Assume: all data currently live will remain live. Steps that will
607 * be collected next time will therefore need twice as many blocks
608 * since all the data will be copied.
617 for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
618 for (s = 0; s < generations[g].n_steps; s++) {
619 if (g == 0 && s == 0) { continue; }
620 stp = &generations[g].steps[s];
621 if (generations[g].steps[0].n_blocks > generations[g].max_blocks) {
622 needed += 2 * stp->n_blocks;
624 needed += stp->n_blocks;
631 /* -----------------------------------------------------------------------------
634 memInventory() checks for memory leaks by counting up all the
635 blocks we know about and comparing that to the number of blocks
636 allegedly floating around in the system.
637 -------------------------------------------------------------------------- */
647 lnat total_blocks = 0, free_blocks = 0;
649 /* count the blocks we current have */
651 for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
652 for (s = 0; s < generations[g].n_steps; s++) {
653 stp = &generations[g].steps[s];
654 total_blocks += stp->n_blocks;
655 if (RtsFlags.GcFlags.generations == 1) {
656 /* two-space collector has a to-space too :-) */
657 total_blocks += g0s0->to_blocks;
659 for (bd = stp->large_objects; bd; bd = bd->link) {
660 total_blocks += bd->blocks;
661 /* hack for megablock groups: they have an extra block or two in
662 the second and subsequent megablocks where the block
663 descriptors would normally go.
665 if (bd->blocks > BLOCKS_PER_MBLOCK) {
666 total_blocks -= (MBLOCK_SIZE / BLOCK_SIZE - BLOCKS_PER_MBLOCK)
667 * (bd->blocks/(MBLOCK_SIZE/BLOCK_SIZE));
673 /* any blocks held by allocate() */
674 for (bd = small_alloc_list; bd; bd = bd->link) {
675 total_blocks += bd->blocks;
677 for (bd = large_alloc_list; bd; bd = bd->link) {
678 total_blocks += bd->blocks;
681 /* count the blocks on the free list */
682 free_blocks = countFreeList();
684 ASSERT(total_blocks + free_blocks == mblocks_allocated * BLOCKS_PER_MBLOCK);
687 if (total_blocks + free_blocks != mblocks_allocated *
689 fprintf(stderr, "Blocks: %ld live + %ld free = %ld total (%ld around)\n",
690 total_blocks, free_blocks, total_blocks + free_blocks,
691 mblocks_allocated * BLOCKS_PER_MBLOCK);
696 /* Full heap sanity check. */
703 if (RtsFlags.GcFlags.generations == 1) {
704 checkHeap(g0s0->to_space, NULL);
705 checkChain(g0s0->large_objects);
708 for (g = 0; g <= N; g++) {
709 for (s = 0; s < generations[g].n_steps; s++) {
710 if (g == 0 && s == 0) { continue; }
711 checkHeap(generations[g].steps[s].blocks, NULL);
714 for (g = N+1; g < RtsFlags.GcFlags.generations; g++) {
715 for (s = 0; s < generations[g].n_steps; s++) {
716 checkHeap(generations[g].steps[s].blocks,
717 generations[g].steps[s].blocks->start);
718 checkChain(generations[g].steps[s].large_objects);
721 checkFreeListSanity();