1 /* -----------------------------------------------------------------------------
2 * $Id: Storage.c,v 1.47 2001/08/09 12:12:23 sewardj 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 bdescr *pinned_object_block; /* allocate pinned objects into this block */
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 */
44 lnat total_allocated = 0; /* total memory allocated during run */
47 * Storage manager mutex: protects all the above state from
48 * simultaneous access by two STG threads.
51 pthread_mutex_t sm_mutex = PTHREAD_MUTEX_INITIALIZER;
57 static void *stgAllocForGMP (size_t size_in_bytes);
58 static void *stgReallocForGMP (void *ptr, size_t old_size, size_t new_size);
59 static void stgDeallocForGMP (void *ptr, size_t size);
68 /* If we're doing heap profiling, we want a two-space heap with a
69 * fixed-size allocation area so that we get roughly even-spaced
73 /* As an experiment, try a 2 generation collector
76 #if defined(PROFILING) || defined(DEBUG)
77 if (RtsFlags.ProfFlags.doHeapProfile) {
78 RtsFlags.GcFlags.generations = 1;
79 RtsFlags.GcFlags.steps = 1;
80 RtsFlags.GcFlags.oldGenFactor = 0;
81 RtsFlags.GcFlags.heapSizeSuggestion = 0;
85 if (RtsFlags.GcFlags.maxHeapSize != 0 &&
86 RtsFlags.GcFlags.heapSizeSuggestion >
87 RtsFlags.GcFlags.maxHeapSize) {
88 RtsFlags.GcFlags.maxHeapSize = RtsFlags.GcFlags.heapSizeSuggestion;
93 /* allocate generation info array */
94 generations = (generation *)stgMallocBytes(RtsFlags.GcFlags.generations
95 * sizeof(struct _generation),
98 /* Initialise all generations */
99 for(g = 0; g < RtsFlags.GcFlags.generations; g++) {
100 gen = &generations[g];
102 gen->mut_list = END_MUT_LIST;
103 gen->mut_once_list = END_MUT_LIST;
104 gen->collections = 0;
105 gen->failed_promotions = 0;
109 /* A couple of convenience pointers */
110 g0 = &generations[0];
111 oldest_gen = &generations[RtsFlags.GcFlags.generations-1];
113 /* Allocate step structures in each generation */
114 if (RtsFlags.GcFlags.generations > 1) {
115 /* Only for multiple-generations */
117 /* Oldest generation: one step */
118 oldest_gen->n_steps = 1;
120 stgMallocBytes(1 * sizeof(struct _step), "initStorage: last step");
122 /* set up all except the oldest generation with 2 steps */
123 for(g = 0; g < RtsFlags.GcFlags.generations-1; g++) {
124 generations[g].n_steps = RtsFlags.GcFlags.steps;
125 generations[g].steps =
126 stgMallocBytes (RtsFlags.GcFlags.steps * sizeof(struct _step),
127 "initStorage: steps");
131 /* single generation, i.e. a two-space collector */
133 g0->steps = stgMallocBytes (sizeof(struct _step), "initStorage: steps");
136 /* Initialise all steps */
137 for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
138 for (s = 0; s < generations[g].n_steps; s++) {
139 stp = &generations[g].steps[s];
143 stp->gen = &generations[g];
150 stp->large_objects = NULL;
151 stp->n_large_blocks = 0;
152 stp->new_large_objects = NULL;
153 stp->scavenged_large_objects = NULL;
154 stp->n_scavenged_large_blocks = 0;
155 stp->is_compacted = 0;
159 /* Set up the destination pointers in each younger gen. step */
160 for (g = 0; g < RtsFlags.GcFlags.generations-1; g++) {
161 for (s = 0; s < generations[g].n_steps-1; s++) {
162 generations[g].steps[s].to = &generations[g].steps[s+1];
164 generations[g].steps[s].to = &generations[g+1].steps[0];
167 /* The oldest generation has one step and it is compacted. */
168 if (RtsFlags.GcFlags.compact) {
169 oldest_gen->steps[0].is_compacted = 1;
171 oldest_gen->steps[0].to = &oldest_gen->steps[0];
173 /* generation 0 is special: that's the nursery */
174 generations[0].max_blocks = 0;
176 /* G0S0: the allocation area. Policy: keep the allocation area
177 * small to begin with, even if we have a large suggested heap
178 * size. Reason: we're going to do a major collection first, and we
179 * don't want it to be a big one. This vague idea is borne out by
180 * rigorous experimental evidence.
182 g0s0 = &generations[0].steps[0];
186 weak_ptr_list = NULL;
189 /* initialise the allocate() interface */
190 small_alloc_list = NULL;
191 large_alloc_list = NULL;
193 alloc_blocks_lim = RtsFlags.GcFlags.minAllocAreaSize;
195 /* Tell GNU multi-precision pkg about our custom alloc functions */
196 mp_set_memory_functions(stgAllocForGMP, stgReallocForGMP, stgDeallocForGMP);
199 pthread_mutex_init(&sm_mutex, NULL);
202 IF_DEBUG(gc, statDescribeGens());
208 stat_exit(calcAllocated());
211 /* -----------------------------------------------------------------------------
214 The entry code for every CAF does the following:
216 - builds a CAF_BLACKHOLE in the heap
217 - pushes an update frame pointing to the CAF_BLACKHOLE
218 - invokes UPD_CAF(), which:
219 - calls newCaf, below
220 - updates the CAF with a static indirection to the CAF_BLACKHOLE
222 Why do we build a BLACKHOLE in the heap rather than just updating
223 the thunk directly? It's so that we only need one kind of update
224 frame - otherwise we'd need a static version of the update frame too.
226 newCaf() does the following:
228 - it puts the CAF on the oldest generation's mut-once list.
229 This is so that we can treat the CAF as a root when collecting
232 For GHCI, we have additional requirements when dealing with CAFs:
234 - we must *retain* all dynamically-loaded CAFs ever entered,
235 just in case we need them again.
236 - we must be able to *revert* CAFs that have been evaluated, to
237 their pre-evaluated form.
239 To do this, we use an additional CAF list. When newCaf() is
240 called on a dynamically-loaded CAF, we add it to the CAF list
241 instead of the old-generation mutable list, and save away its
242 old info pointer (in caf->saved_info) for later reversion.
244 To revert all the CAFs, we traverse the CAF list and reset the
245 info pointer to caf->saved_info, then throw away the CAF list.
246 (see GC.c:revertCAFs()).
250 -------------------------------------------------------------------------- */
253 newCAF(StgClosure* caf)
255 /* Put this CAF on the mutable list for the old generation.
256 * This is a HACK - the IND_STATIC closure doesn't really have
257 * a mut_link field, but we pretend it has - in fact we re-use
258 * the STATIC_LINK field for the time being, because when we
259 * come to do a major GC we won't need the mut_link field
260 * any more and can use it as a STATIC_LINK.
262 ACQUIRE_LOCK(&sm_mutex);
264 if (is_dynamically_loaded_rwdata_ptr((StgPtr)caf)) {
265 ((StgIndStatic *)caf)->saved_info = (StgInfoTable *)caf->header.info;
266 ((StgIndStatic *)caf)->static_link = caf_list;
269 ((StgIndStatic *)caf)->saved_info = NULL;
270 ((StgMutClosure *)caf)->mut_link = oldest_gen->mut_once_list;
271 oldest_gen->mut_once_list = (StgMutClosure *)caf;
274 RELEASE_LOCK(&sm_mutex);
277 /* If we are PAR or DIST then we never forget a CAF */
279 //belch("<##> Globalising CAF %08x %s",caf,info_type(caf));
280 newGA=makeGlobal(caf,rtsTrue); /*given full weight*/
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->u.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_blocks = NULL;
316 g0s0->n_to_blocks = 0;
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_no == 0);
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_no == 0);
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);
436 bd->flags = BF_LARGE;
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);
470 allocated_bytes( void )
472 return (alloc_blocks * BLOCK_SIZE_W - (alloc_HpLim - alloc_Hp));
475 /* ---------------------------------------------------------------------------
476 Allocate a fixed/pinned object.
478 We allocate small pinned objects into a single block, allocating a
479 new block when the current one overflows. The block is chained
480 onto the large_object_list of generation 0 step 0.
482 NOTE: The GC can't in general handle pinned objects. This
483 interface is only safe to use for ByteArrays, which have no
484 pointers and don't require scavenging. It works because the
485 block's descriptor has the BF_LARGE flag set, so the block is
486 treated as a large object and chained onto various lists, rather
487 than the individual objects being copied. However, when it comes
488 to scavenge the block, the GC will only scavenge the first object.
489 The reason is that the GC can't linearly scan a block of pinned
490 objects at the moment (doing so would require using the
491 mostly-copying techniques). But since we're restricting ourselves
492 to pinned ByteArrays, not scavenging is ok.
494 This function is called by newPinnedByteArray# which immediately
495 fills the allocated memory with a MutableByteArray#.
496 ------------------------------------------------------------------------- */
499 allocatePinned( nat n )
502 bdescr *bd = pinned_object_block;
504 ACQUIRE_LOCK(&sm_mutex);
506 TICK_ALLOC_HEAP_NOCTR(n);
509 // If the request is for a large object, then allocate()
510 // will give us a pinned object anyway.
511 if (n >= LARGE_OBJECT_THRESHOLD/sizeof(W_)) {
512 RELEASE_LOCK(&sm_mutex);
516 // If we don't have a block of pinned objects yet, or the current
517 // one isn't large enough to hold the new object, allocate a new one.
518 if (bd == NULL || (bd->free + n) > (bd->start + BLOCK_SIZE_W)) {
519 pinned_object_block = bd = allocBlock();
520 dbl_link_onto(bd, &g0s0->large_objects);
523 bd->flags = BF_LARGE;
524 bd->free = bd->start;
530 RELEASE_LOCK(&sm_mutex);
534 /* -----------------------------------------------------------------------------
535 Allocation functions for GMP.
537 These all use the allocate() interface - we can't have any garbage
538 collection going on during a gmp operation, so we use allocate()
539 which always succeeds. The gmp operations which might need to
540 allocate will ask the storage manager (via doYouWantToGC()) whether
541 a garbage collection is required, in case we get into a loop doing
542 only allocate() style allocation.
543 -------------------------------------------------------------------------- */
546 stgAllocForGMP (size_t size_in_bytes)
549 nat data_size_in_words, total_size_in_words;
551 /* should be a multiple of sizeof(StgWord) (whole no. of limbs) */
552 ASSERT(size_in_bytes % sizeof(W_) == 0);
554 data_size_in_words = size_in_bytes / sizeof(W_);
555 total_size_in_words = sizeofW(StgArrWords) + data_size_in_words;
557 /* allocate and fill it in. */
558 arr = (StgArrWords *)allocate(total_size_in_words);
559 SET_ARR_HDR(arr, &stg_ARR_WORDS_info, CCCS, data_size_in_words);
561 /* and return a ptr to the goods inside the array */
562 return(BYTE_ARR_CTS(arr));
566 stgReallocForGMP (void *ptr, size_t old_size, size_t new_size)
568 void *new_stuff_ptr = stgAllocForGMP(new_size);
570 char *p = (char *) ptr;
571 char *q = (char *) new_stuff_ptr;
573 for (; i < old_size; i++, p++, q++) {
577 return(new_stuff_ptr);
581 stgDeallocForGMP (void *ptr STG_UNUSED,
582 size_t size STG_UNUSED)
584 /* easy for us: the garbage collector does the dealloc'n */
587 /* -----------------------------------------------------------------------------
589 * -------------------------------------------------------------------------- */
591 /* -----------------------------------------------------------------------------
594 * Approximate how much we've allocated: number of blocks in the
595 * nursery + blocks allocated via allocate() - unused nusery blocks.
596 * This leaves a little slop at the end of each block, and doesn't
597 * take into account large objects (ToDo).
598 * -------------------------------------------------------------------------- */
601 calcAllocated( void )
609 /* All tasks must be stopped. Can't assert that all the
610 capabilities are owned by the scheduler, though: one or more
611 tasks might have been stopped while they were running (non-main)
613 /* ASSERT(n_free_capabilities == RtsFlags.ParFlags.nNodes); */
616 n_free_capabilities * RtsFlags.GcFlags.minAllocAreaSize * BLOCK_SIZE_W
619 for (cap = free_capabilities; cap != NULL; cap = cap->link) {
620 for ( bd = cap->rCurrentNursery->link; bd != NULL; bd = bd->link ) {
621 allocated -= BLOCK_SIZE_W;
623 if (cap->rCurrentNursery->free < cap->rCurrentNursery->start
625 allocated -= (cap->rCurrentNursery->start + BLOCK_SIZE_W)
626 - cap->rCurrentNursery->free;
631 bdescr *current_nursery = MainRegTable.rCurrentNursery;
633 allocated = (nursery_blocks * BLOCK_SIZE_W) + allocated_bytes();
634 for ( bd = current_nursery->link; bd != NULL; bd = bd->link ) {
635 allocated -= BLOCK_SIZE_W;
637 if (current_nursery->free < current_nursery->start + BLOCK_SIZE_W) {
638 allocated -= (current_nursery->start + BLOCK_SIZE_W)
639 - current_nursery->free;
643 total_allocated += allocated;
647 /* Approximate the amount of live data in the heap. To be called just
648 * after garbage collection (see GarbageCollect()).
657 if (RtsFlags.GcFlags.generations == 1) {
658 live = (g0s0->n_to_blocks - 1) * BLOCK_SIZE_W +
659 ((lnat)g0s0->hp_bd->free - (lnat)g0s0->hp_bd->start) / sizeof(W_);
663 for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
664 for (s = 0; s < generations[g].n_steps; s++) {
665 /* approximate amount of live data (doesn't take into account slop
666 * at end of each block).
668 if (g == 0 && s == 0) {
671 stp = &generations[g].steps[s];
672 live += (stp->n_large_blocks + stp->n_blocks - 1) * BLOCK_SIZE_W;
673 if (stp->hp_bd != NULL) {
674 live += ((lnat)stp->hp_bd->free - (lnat)stp->hp_bd->start)
682 /* Approximate the number of blocks that will be needed at the next
683 * garbage collection.
685 * Assume: all data currently live will remain live. Steps that will
686 * be collected next time will therefore need twice as many blocks
687 * since all the data will be copied.
696 for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
697 for (s = 0; s < generations[g].n_steps; s++) {
698 if (g == 0 && s == 0) { continue; }
699 stp = &generations[g].steps[s];
700 if (generations[g].steps[0].n_blocks +
701 generations[g].steps[0].n_large_blocks
702 > generations[g].max_blocks
703 && stp->is_compacted == 0) {
704 needed += 2 * stp->n_blocks;
706 needed += stp->n_blocks;
713 /* -----------------------------------------------------------------------------
716 memInventory() checks for memory leaks by counting up all the
717 blocks we know about and comparing that to the number of blocks
718 allegedly floating around in the system.
719 -------------------------------------------------------------------------- */
729 lnat total_blocks = 0, free_blocks = 0;
731 /* count the blocks we current have */
733 for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
734 for (s = 0; s < generations[g].n_steps; s++) {
735 stp = &generations[g].steps[s];
736 total_blocks += stp->n_blocks;
737 if (RtsFlags.GcFlags.generations == 1) {
738 /* two-space collector has a to-space too :-) */
739 total_blocks += g0s0->n_to_blocks;
741 for (bd = stp->large_objects; bd; bd = bd->link) {
742 total_blocks += bd->blocks;
743 /* hack for megablock groups: they have an extra block or two in
744 the second and subsequent megablocks where the block
745 descriptors would normally go.
747 if (bd->blocks > BLOCKS_PER_MBLOCK) {
748 total_blocks -= (MBLOCK_SIZE / BLOCK_SIZE - BLOCKS_PER_MBLOCK)
749 * (bd->blocks/(MBLOCK_SIZE/BLOCK_SIZE));
755 /* any blocks held by allocate() */
756 for (bd = small_alloc_list; bd; bd = bd->link) {
757 total_blocks += bd->blocks;
759 for (bd = large_alloc_list; bd; bd = bd->link) {
760 total_blocks += bd->blocks;
763 /* count the blocks on the free list */
764 free_blocks = countFreeList();
766 if (total_blocks + free_blocks != mblocks_allocated *
768 fprintf(stderr, "Blocks: %ld live + %ld free = %ld total (%ld around)\n",
769 total_blocks, free_blocks, total_blocks + free_blocks,
770 mblocks_allocated * BLOCKS_PER_MBLOCK);
773 ASSERT(total_blocks + free_blocks == mblocks_allocated * BLOCKS_PER_MBLOCK);
777 countBlocks(bdescr *bd)
780 for (n=0; bd != NULL; bd=bd->link) {
786 /* Full heap sanity check. */
792 if (RtsFlags.GcFlags.generations == 1) {
793 checkHeap(g0s0->to_blocks);
794 checkChain(g0s0->large_objects);
797 for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
798 for (s = 0; s < generations[g].n_steps; s++) {
799 if (g == 0 && s == 0) { continue; }
800 checkHeap(generations[g].steps[s].blocks);
801 checkChain(generations[g].steps[s].large_objects);
802 ASSERT(countBlocks(generations[g].steps[s].blocks)
803 == generations[g].steps[s].n_blocks);
804 ASSERT(countBlocks(generations[g].steps[s].large_objects)
805 == generations[g].steps[s].n_large_blocks);
807 checkMutableList(generations[g].mut_list, g);
808 checkMutOnceList(generations[g].mut_once_list, g);
812 checkFreeListSanity();