1 /* -----------------------------------------------------------------------------
2 * $Id: Storage.c,v 1.24 2000/04/14 15:18:07 sewardj 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;
182 /* Tell GNU multi-precision pkg about our custom alloc functions */
183 mp_set_memory_functions(stgAllocForGMP, stgReallocForGMP, stgDeallocForGMP);
186 pthread_mutex_init(&sm_mutex, NULL);
189 IF_DEBUG(gc, stat_describe_gens());
195 stat_exit(calcAllocated());
199 /* -----------------------------------------------------------------------------
201 -------------------------------------------------------------------------- */
204 newCAF(StgClosure* caf)
206 /* Put this CAF on the mutable list for the old generation.
207 * This is a HACK - the IND_STATIC closure doesn't really have
208 * a mut_link field, but we pretend it has - in fact we re-use
209 * the STATIC_LINK field for the time being, because when we
210 * come to do a major GC we won't need the mut_link field
211 * any more and can use it as a STATIC_LINK.
213 ACQUIRE_LOCK(&sm_mutex);
215 ASSERT( ((StgMutClosure*)caf)->mut_link == NULL );
216 ((StgMutClosure *)caf)->mut_link = oldest_gen->mut_once_list;
217 oldest_gen->mut_once_list = (StgMutClosure *)caf;
220 /* If we're Hugs, we also have to put it in the CAF table, so that
221 the CAF can be reverted. When reverting, CAFs created by compiled
222 code are recorded in the CAF table, which lives outside the
223 heap, in mallocville. CAFs created by interpreted code are
224 chained together via the link fields in StgCAFs, and are not
225 recorded in the CAF table.
227 ASSERT( get_itbl(caf)->type == THUNK_STATIC );
228 addToECafTable ( caf, get_itbl(caf) );
231 RELEASE_LOCK(&sm_mutex);
236 newCAF_made_by_Hugs(StgCAF* caf)
238 ACQUIRE_LOCK(&sm_mutex);
240 ASSERT( get_itbl(caf)->type == CAF_ENTERED );
241 recordOldToNewPtrs((StgMutClosure*)caf);
242 caf->link = ecafList;
243 ecafList = caf->link;
245 RELEASE_LOCK(&sm_mutex);
250 /* These initialisations are critical for correct operation
251 on the first call of addToECafTable.
253 StgCAF* ecafList = END_ECAF_LIST;
254 StgCAFTabEntry* ecafTable = NULL;
255 StgInt usedECafTable = 0;
256 StgInt sizeECafTable = 0;
259 void clearECafTable ( void )
264 void addToECafTable ( StgClosure* closure, StgInfoTable* origItbl )
268 if (usedECafTable == sizeECafTable) {
269 /* Make the initial table size be 8 */
271 if (sizeECafTable == 0) sizeECafTable = 8;
272 et2 = stgMallocBytes (
273 sizeECafTable * sizeof(StgCAFTabEntry),
275 for (i = 0; i < usedECafTable; i++)
276 et2[i] = ecafTable[i];
277 if (ecafTable) free(ecafTable);
280 ecafTable[usedECafTable].closure = closure;
281 ecafTable[usedECafTable].origItbl = origItbl;
286 /* -----------------------------------------------------------------------------
288 -------------------------------------------------------------------------- */
291 allocNurseries( void )
300 for (cap = free_capabilities; cap != NULL; cap = cap->link) {
301 cap->rNursery = allocNursery(NULL, RtsFlags.GcFlags.minAllocAreaSize);
302 cap->rCurrentNursery = cap->rNursery;
303 for (bd = cap->rNursery; bd != NULL; bd = bd->link) {
304 bd->back = (bdescr *)cap;
307 /* Set the back links to be equal to the Capability,
308 * so we can do slightly better informed locking.
312 nursery_blocks = RtsFlags.GcFlags.minAllocAreaSize;
313 g0s0->blocks = allocNursery(NULL, nursery_blocks);
314 g0s0->n_blocks = nursery_blocks;
315 g0s0->to_space = NULL;
316 MainRegTable.rNursery = g0s0->blocks;
317 MainRegTable.rCurrentNursery = g0s0->blocks;
318 /* hp, hpLim, hp_bd, to_space etc. aren't used in G0S0 */
323 resetNurseries( void )
329 /* All tasks must be stopped */
330 ASSERT(n_free_capabilities == RtsFlags.ParFlags.nNodes);
332 for (cap = free_capabilities; cap != NULL; cap = cap->link) {
333 for (bd = cap->rNursery; bd; bd = bd->link) {
334 bd->free = bd->start;
335 ASSERT(bd->gen == g0);
336 ASSERT(bd->step == g0s0);
337 IF_DEBUG(sanity,memset(bd->start, 0xaa, BLOCK_SIZE));
339 cap->rCurrentNursery = cap->rNursery;
342 for (bd = g0s0->blocks; bd; bd = bd->link) {
343 bd->free = bd->start;
344 ASSERT(bd->gen == g0);
345 ASSERT(bd->step == g0s0);
346 IF_DEBUG(sanity,memset(bd->start, 0xaa, BLOCK_SIZE));
348 MainRegTable.rNursery = g0s0->blocks;
349 MainRegTable.rCurrentNursery = g0s0->blocks;
354 allocNursery (bdescr *last_bd, nat blocks)
359 /* Allocate a nursery */
360 for (i=0; i < blocks; i++) {
366 bd->free = bd->start;
373 resizeNursery ( nat blocks )
378 barf("resizeNursery: can't resize in SMP mode");
381 if (nursery_blocks == blocks) {
382 ASSERT(g0s0->n_blocks == blocks);
386 else if (nursery_blocks < blocks) {
387 IF_DEBUG(gc, fprintf(stderr, "Increasing size of nursery to %d blocks\n",
389 g0s0->blocks = allocNursery(g0s0->blocks, blocks-nursery_blocks);
395 IF_DEBUG(gc, fprintf(stderr, "Decreasing size of nursery to %d blocks\n",
397 for (bd = g0s0->blocks; nursery_blocks > blocks; nursery_blocks--) {
405 g0s0->n_blocks = nursery_blocks = blocks;
408 /* -----------------------------------------------------------------------------
409 The allocate() interface
411 allocate(n) always succeeds, and returns a chunk of memory n words
412 long. n can be larger than the size of a block if necessary, in
413 which case a contiguous block group will be allocated.
414 -------------------------------------------------------------------------- */
422 ACQUIRE_LOCK(&sm_mutex);
424 TICK_ALLOC_HEAP_NOCTR(n);
427 /* big allocation (>LARGE_OBJECT_THRESHOLD) */
428 /* ToDo: allocate directly into generation 1 */
429 if (n >= LARGE_OBJECT_THRESHOLD/sizeof(W_)) {
430 nat req_blocks = (lnat)BLOCK_ROUND_UP(n*sizeof(W_)) / BLOCK_SIZE;
431 bd = allocGroup(req_blocks);
432 dbl_link_onto(bd, &g0s0->large_objects);
436 bd->free = bd->start;
437 /* don't add these blocks to alloc_blocks, since we're assuming
438 * that large objects are likely to remain live for quite a while
439 * (eg. running threads), so garbage collecting early won't make
442 RELEASE_LOCK(&sm_mutex);
445 /* small allocation (<LARGE_OBJECT_THRESHOLD) */
446 } else if (small_alloc_list == NULL || alloc_Hp + n > alloc_HpLim) {
447 if (small_alloc_list) {
448 small_alloc_list->free = alloc_Hp;
451 bd->link = small_alloc_list;
452 small_alloc_list = bd;
456 alloc_Hp = bd->start;
457 alloc_HpLim = bd->start + BLOCK_SIZE_W;
463 RELEASE_LOCK(&sm_mutex);
467 lnat allocated_bytes(void)
469 return (alloc_blocks * BLOCK_SIZE_W - (alloc_HpLim - alloc_Hp));
472 /* -----------------------------------------------------------------------------
473 Allocation functions for GMP.
475 These all use the allocate() interface - we can't have any garbage
476 collection going on during a gmp operation, so we use allocate()
477 which always succeeds. The gmp operations which might need to
478 allocate will ask the storage manager (via doYouWantToGC()) whether
479 a garbage collection is required, in case we get into a loop doing
480 only allocate() style allocation.
481 -------------------------------------------------------------------------- */
484 stgAllocForGMP (size_t size_in_bytes)
487 nat data_size_in_words, total_size_in_words;
489 /* should be a multiple of sizeof(StgWord) (whole no. of limbs) */
490 ASSERT(size_in_bytes % sizeof(W_) == 0);
492 data_size_in_words = size_in_bytes / sizeof(W_);
493 total_size_in_words = sizeofW(StgArrWords) + data_size_in_words;
495 /* allocate and fill it in. */
496 arr = (StgArrWords *)allocate(total_size_in_words);
497 SET_ARR_HDR(arr, &ARR_WORDS_info, CCCS, data_size_in_words);
499 /* and return a ptr to the goods inside the array */
500 return(BYTE_ARR_CTS(arr));
504 stgReallocForGMP (void *ptr, size_t old_size, size_t new_size)
506 void *new_stuff_ptr = stgAllocForGMP(new_size);
508 char *p = (char *) ptr;
509 char *q = (char *) new_stuff_ptr;
511 for (; i < old_size; i++, p++, q++) {
515 return(new_stuff_ptr);
519 stgDeallocForGMP (void *ptr STG_UNUSED,
520 size_t size STG_UNUSED)
522 /* easy for us: the garbage collector does the dealloc'n */
525 /* -----------------------------------------------------------------------------
527 * -------------------------------------------------------------------------- */
529 /* -----------------------------------------------------------------------------
532 * Approximate how much we've allocated: number of blocks in the
533 * nursery + blocks allocated via allocate() - unused nusery blocks.
534 * This leaves a little slop at the end of each block, and doesn't
535 * take into account large objects (ToDo).
536 * -------------------------------------------------------------------------- */
539 calcAllocated( void )
547 /* All tasks must be stopped. Can't assert that all the
548 capabilities are owned by the scheduler, though: one or more
549 tasks might have been stopped while they were running (non-main)
551 /* ASSERT(n_free_capabilities == RtsFlags.ParFlags.nNodes); */
554 n_free_capabilities * RtsFlags.GcFlags.minAllocAreaSize * BLOCK_SIZE_W
557 for (cap = free_capabilities; cap != NULL; cap = cap->link) {
558 for ( bd = cap->rCurrentNursery->link; bd != NULL; bd = bd->link ) {
559 allocated -= BLOCK_SIZE_W;
561 if (cap->rCurrentNursery->free < cap->rCurrentNursery->start
563 allocated -= (cap->rCurrentNursery->start + BLOCK_SIZE_W)
564 - cap->rCurrentNursery->free;
569 bdescr *current_nursery = MainRegTable.rCurrentNursery;
571 allocated = (nursery_blocks * BLOCK_SIZE_W) + allocated_bytes();
572 for ( bd = current_nursery->link; bd != NULL; bd = bd->link ) {
573 allocated -= BLOCK_SIZE_W;
575 if (current_nursery->free < current_nursery->start + BLOCK_SIZE_W) {
576 allocated -= (current_nursery->start + BLOCK_SIZE_W)
577 - current_nursery->free;
584 /* Approximate the amount of live data in the heap. To be called just
585 * after garbage collection (see GarbageCollect()).
594 if (RtsFlags.GcFlags.generations == 1) {
595 live = (g0s0->to_blocks - 1) * BLOCK_SIZE_W +
596 ((lnat)g0s0->hp_bd->free - (lnat)g0s0->hp_bd->start) / sizeof(W_);
600 for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
601 for (s = 0; s < generations[g].n_steps; s++) {
602 /* approximate amount of live data (doesn't take into account slop
603 * at end of each block).
605 if (g == 0 && s == 0) {
608 step = &generations[g].steps[s];
609 live += (step->n_blocks - 1) * BLOCK_SIZE_W +
610 ((lnat)step->hp_bd->free - (lnat)step->hp_bd->start) / sizeof(W_);
616 /* Approximate the number of blocks that will be needed at the next
617 * garbage collection.
619 * Assume: all data currently live will remain live. Steps that will
620 * be collected next time will therefore need twice as many blocks
621 * since all the data will be copied.
630 for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
631 for (s = 0; s < generations[g].n_steps; s++) {
632 if (g == 0 && s == 0) { continue; }
633 step = &generations[g].steps[s];
634 if (generations[g].steps[0].n_blocks > generations[g].max_blocks) {
635 needed += 2 * step->n_blocks;
637 needed += step->n_blocks;
644 /* -----------------------------------------------------------------------------
647 memInventory() checks for memory leaks by counting up all the
648 blocks we know about and comparing that to the number of blocks
649 allegedly floating around in the system.
650 -------------------------------------------------------------------------- */
660 lnat total_blocks = 0, free_blocks = 0;
662 /* count the blocks we current have */
664 for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
665 for (s = 0; s < generations[g].n_steps; s++) {
666 step = &generations[g].steps[s];
667 total_blocks += step->n_blocks;
668 if (RtsFlags.GcFlags.generations == 1) {
669 /* two-space collector has a to-space too :-) */
670 total_blocks += g0s0->to_blocks;
672 for (bd = step->large_objects; bd; bd = bd->link) {
673 total_blocks += bd->blocks;
674 /* hack for megablock groups: they have an extra block or two in
675 the second and subsequent megablocks where the block
676 descriptors would normally go.
678 if (bd->blocks > BLOCKS_PER_MBLOCK) {
679 total_blocks -= (MBLOCK_SIZE / BLOCK_SIZE - BLOCKS_PER_MBLOCK)
680 * (bd->blocks/(MBLOCK_SIZE/BLOCK_SIZE));
686 /* any blocks held by allocate() */
687 for (bd = small_alloc_list; bd; bd = bd->link) {
688 total_blocks += bd->blocks;
690 for (bd = large_alloc_list; bd; bd = bd->link) {
691 total_blocks += bd->blocks;
694 /* count the blocks on the free list */
695 free_blocks = countFreeList();
697 ASSERT(total_blocks + free_blocks == mblocks_allocated * BLOCKS_PER_MBLOCK);
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);
709 /* Full heap sanity check. */
715 fprintf(stderr, "--- checkSanity %d\n", N );
716 if (0&&RtsFlags.GcFlags.generations == 1) {
717 checkHeap(g0s0->to_space, NULL);
718 checkChain(g0s0->large_objects);
721 for (g = 0; g <= N; g++) {
722 for (s = 0; s < generations[g].n_steps; s++) {
723 if (g == 0 && s == 0) { continue; }
724 checkHeap(generations[g].steps[s].blocks, NULL);
727 for (g = N+1; g < RtsFlags.GcFlags.generations; g++) {
728 for (s = 0; s < generations[g].n_steps; s++) {
729 checkHeap(generations[g].steps[s].blocks,
730 generations[g].steps[s].blocks->start);
731 checkChain(generations[g].steps[s].large_objects);
734 checkFreeListSanity();