1 /* -----------------------------------------------------------------------------
2 * $Id: Storage.c,v 1.46 2001/08/08 14:14:08 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 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->new_large_objects = NULL;
152 stp->scavenged_large_objects = NULL;
153 stp->is_compacted = 0;
157 /* Set up the destination pointers in each younger gen. step */
158 for (g = 0; g < RtsFlags.GcFlags.generations-1; g++) {
159 for (s = 0; s < generations[g].n_steps-1; s++) {
160 generations[g].steps[s].to = &generations[g].steps[s+1];
162 generations[g].steps[s].to = &generations[g+1].steps[0];
165 /* The oldest generation has one step and it is compacted. */
166 if (RtsFlags.GcFlags.compact) {
167 oldest_gen->steps[0].is_compacted = 1;
169 oldest_gen->steps[0].to = &oldest_gen->steps[0];
171 /* generation 0 is special: that's the nursery */
172 generations[0].max_blocks = 0;
174 /* G0S0: the allocation area. Policy: keep the allocation area
175 * small to begin with, even if we have a large suggested heap
176 * size. Reason: we're going to do a major collection first, and we
177 * don't want it to be a big one. This vague idea is borne out by
178 * rigorous experimental evidence.
180 g0s0 = &generations[0].steps[0];
184 weak_ptr_list = NULL;
187 /* initialise the allocate() interface */
188 small_alloc_list = NULL;
189 large_alloc_list = NULL;
191 alloc_blocks_lim = RtsFlags.GcFlags.minAllocAreaSize;
193 /* Tell GNU multi-precision pkg about our custom alloc functions */
194 mp_set_memory_functions(stgAllocForGMP, stgReallocForGMP, stgDeallocForGMP);
197 pthread_mutex_init(&sm_mutex, NULL);
200 IF_DEBUG(gc, statDescribeGens());
206 stat_exit(calcAllocated());
209 /* -----------------------------------------------------------------------------
212 The entry code for every CAF does the following:
214 - builds a CAF_BLACKHOLE in the heap
215 - pushes an update frame pointing to the CAF_BLACKHOLE
216 - invokes UPD_CAF(), which:
217 - calls newCaf, below
218 - updates the CAF with a static indirection to the CAF_BLACKHOLE
220 Why do we build a BLACKHOLE in the heap rather than just updating
221 the thunk directly? It's so that we only need one kind of update
222 frame - otherwise we'd need a static version of the update frame too.
224 newCaf() does the following:
226 - it puts the CAF on the oldest generation's mut-once list.
227 This is so that we can treat the CAF as a root when collecting
230 For GHCI, we have additional requirements when dealing with CAFs:
232 - we must *retain* all dynamically-loaded CAFs ever entered,
233 just in case we need them again.
234 - we must be able to *revert* CAFs that have been evaluated, to
235 their pre-evaluated form.
237 To do this, we use an additional CAF list. When newCaf() is
238 called on a dynamically-loaded CAF, we add it to the CAF list
239 instead of the old-generation mutable list, and save away its
240 old info pointer (in caf->saved_info) for later reversion.
242 To revert all the CAFs, we traverse the CAF list and reset the
243 info pointer to caf->saved_info, then throw away the CAF list.
244 (see GC.c:revertCAFs()).
248 -------------------------------------------------------------------------- */
251 newCAF(StgClosure* caf)
253 /* Put this CAF on the mutable list for the old generation.
254 * This is a HACK - the IND_STATIC closure doesn't really have
255 * a mut_link field, but we pretend it has - in fact we re-use
256 * the STATIC_LINK field for the time being, because when we
257 * come to do a major GC we won't need the mut_link field
258 * any more and can use it as a STATIC_LINK.
260 ACQUIRE_LOCK(&sm_mutex);
262 if (is_dynamically_loaded_rwdata_ptr((StgPtr)caf)) {
263 ((StgIndStatic *)caf)->saved_info = (StgInfoTable *)caf->header.info;
264 ((StgIndStatic *)caf)->static_link = caf_list;
267 ((StgIndStatic *)caf)->saved_info = NULL;
268 ((StgMutClosure *)caf)->mut_link = oldest_gen->mut_once_list;
269 oldest_gen->mut_once_list = (StgMutClosure *)caf;
272 RELEASE_LOCK(&sm_mutex);
275 /* If we are PAR or DIST then we never forget a CAF */
277 //belch("<##> Globalising CAF %08x %s",caf,info_type(caf));
278 newGA=makeGlobal(caf,rtsTrue); /*given full weight*/
284 /* -----------------------------------------------------------------------------
286 -------------------------------------------------------------------------- */
289 allocNurseries( void )
298 for (cap = free_capabilities; cap != NULL; cap = cap->link) {
299 cap->rNursery = allocNursery(NULL, RtsFlags.GcFlags.minAllocAreaSize);
300 cap->rCurrentNursery = cap->rNursery;
301 for (bd = cap->rNursery; bd != NULL; bd = bd->link) {
302 bd->u.back = (bdescr *)cap;
305 /* Set the back links to be equal to the Capability,
306 * so we can do slightly better informed locking.
310 nursery_blocks = RtsFlags.GcFlags.minAllocAreaSize;
311 g0s0->blocks = allocNursery(NULL, nursery_blocks);
312 g0s0->n_blocks = nursery_blocks;
313 g0s0->to_blocks = NULL;
314 g0s0->n_to_blocks = 0;
315 MainRegTable.rNursery = g0s0->blocks;
316 MainRegTable.rCurrentNursery = g0s0->blocks;
317 /* hp, hpLim, hp_bd, to_space etc. aren't used in G0S0 */
322 resetNurseries( void )
328 /* All tasks must be stopped */
329 ASSERT(n_free_capabilities == RtsFlags.ParFlags.nNodes);
331 for (cap = free_capabilities; cap != NULL; cap = cap->link) {
332 for (bd = cap->rNursery; bd; bd = bd->link) {
333 bd->free = bd->start;
334 ASSERT(bd->gen_no == 0);
335 ASSERT(bd->step == g0s0);
336 IF_DEBUG(sanity,memset(bd->start, 0xaa, BLOCK_SIZE));
338 cap->rCurrentNursery = cap->rNursery;
341 for (bd = g0s0->blocks; bd; bd = bd->link) {
342 bd->free = bd->start;
343 ASSERT(bd->gen_no == 0);
344 ASSERT(bd->step == g0s0);
345 IF_DEBUG(sanity,memset(bd->start, 0xaa, BLOCK_SIZE));
347 MainRegTable.rNursery = g0s0->blocks;
348 MainRegTable.rCurrentNursery = g0s0->blocks;
353 allocNursery (bdescr *last_bd, nat blocks)
358 /* Allocate a nursery */
359 for (i=0; i < blocks; i++) {
365 bd->free = bd->start;
372 resizeNursery ( nat blocks )
377 barf("resizeNursery: can't resize in SMP mode");
380 if (nursery_blocks == blocks) {
381 ASSERT(g0s0->n_blocks == blocks);
385 else if (nursery_blocks < blocks) {
386 IF_DEBUG(gc, fprintf(stderr, "Increasing size of nursery to %d blocks\n",
388 g0s0->blocks = allocNursery(g0s0->blocks, blocks-nursery_blocks);
394 IF_DEBUG(gc, fprintf(stderr, "Decreasing size of nursery to %d blocks\n",
396 for (bd = g0s0->blocks; nursery_blocks > blocks; nursery_blocks--) {
404 g0s0->n_blocks = nursery_blocks = blocks;
407 /* -----------------------------------------------------------------------------
408 The allocate() interface
410 allocate(n) always succeeds, and returns a chunk of memory n words
411 long. n can be larger than the size of a block if necessary, in
412 which case a contiguous block group will be allocated.
413 -------------------------------------------------------------------------- */
421 ACQUIRE_LOCK(&sm_mutex);
423 TICK_ALLOC_HEAP_NOCTR(n);
426 /* big allocation (>LARGE_OBJECT_THRESHOLD) */
427 /* ToDo: allocate directly into generation 1 */
428 if (n >= LARGE_OBJECT_THRESHOLD/sizeof(W_)) {
429 nat req_blocks = (lnat)BLOCK_ROUND_UP(n*sizeof(W_)) / BLOCK_SIZE;
430 bd = allocGroup(req_blocks);
431 dbl_link_onto(bd, &g0s0->large_objects);
434 bd->flags = BF_LARGE;
435 bd->free = bd->start;
436 /* don't add these blocks to alloc_blocks, since we're assuming
437 * that large objects are likely to remain live for quite a while
438 * (eg. running threads), so garbage collecting early won't make
441 alloc_blocks += req_blocks;
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);
468 allocated_bytes( void )
470 return (alloc_blocks * BLOCK_SIZE_W - (alloc_HpLim - alloc_Hp));
473 /* ---------------------------------------------------------------------------
474 Allocate a fixed/pinned object.
476 We allocate small pinned objects into a single block, allocating a
477 new block when the current one overflows. The block is chained
478 onto the large_object_list of generation 0 step 0.
480 NOTE: The GC can't in general handle pinned objects. This
481 interface is only safe to use for ByteArrays, which have no
482 pointers and don't require scavenging. It works because the
483 block's descriptor has the BF_LARGE flag set, so the block is
484 treated as a large object and chained onto various lists, rather
485 than the individual objects being copied. However, when it comes
486 to scavenge the block, the GC will only scavenge the first object.
487 The reason is that the GC can't linearly scan a block of pinned
488 objects at the moment (doing so would require using the
489 mostly-copying techniques). But since we're restricting ourselves
490 to pinned ByteArrays, not scavenging is ok.
492 This function is called by newPinnedByteArray# which immediately
493 fills the allocated memory with a MutableByteArray#.
494 ------------------------------------------------------------------------- */
497 allocatePinned( nat n )
500 bdescr *bd = pinned_object_block;
502 ACQUIRE_LOCK(&sm_mutex);
504 TICK_ALLOC_HEAP_NOCTR(n);
507 // If the request is for a large object, then allocate()
508 // will give us a pinned object anyway.
509 if (n >= LARGE_OBJECT_THRESHOLD/sizeof(W_)) {
510 RELEASE_LOCK(&sm_mutex);
514 // If we don't have a block of pinned objects yet, or the current
515 // one isn't large enough to hold the new object, allocate a new one.
516 if (bd == NULL || (bd->free + n) > (bd->start + BLOCK_SIZE_W)) {
517 pinned_object_block = bd = allocBlock();
518 dbl_link_onto(bd, &g0s0->large_objects);
521 bd->flags = BF_LARGE;
522 bd->free = bd->start;
528 RELEASE_LOCK(&sm_mutex);
532 /* -----------------------------------------------------------------------------
533 Allocation functions for GMP.
535 These all use the allocate() interface - we can't have any garbage
536 collection going on during a gmp operation, so we use allocate()
537 which always succeeds. The gmp operations which might need to
538 allocate will ask the storage manager (via doYouWantToGC()) whether
539 a garbage collection is required, in case we get into a loop doing
540 only allocate() style allocation.
541 -------------------------------------------------------------------------- */
544 stgAllocForGMP (size_t size_in_bytes)
547 nat data_size_in_words, total_size_in_words;
549 /* should be a multiple of sizeof(StgWord) (whole no. of limbs) */
550 ASSERT(size_in_bytes % sizeof(W_) == 0);
552 data_size_in_words = size_in_bytes / sizeof(W_);
553 total_size_in_words = sizeofW(StgArrWords) + data_size_in_words;
555 /* allocate and fill it in. */
556 arr = (StgArrWords *)allocate(total_size_in_words);
557 SET_ARR_HDR(arr, &stg_ARR_WORDS_info, CCCS, data_size_in_words);
559 /* and return a ptr to the goods inside the array */
560 return(BYTE_ARR_CTS(arr));
564 stgReallocForGMP (void *ptr, size_t old_size, size_t new_size)
566 void *new_stuff_ptr = stgAllocForGMP(new_size);
568 char *p = (char *) ptr;
569 char *q = (char *) new_stuff_ptr;
571 for (; i < old_size; i++, p++, q++) {
575 return(new_stuff_ptr);
579 stgDeallocForGMP (void *ptr STG_UNUSED,
580 size_t size STG_UNUSED)
582 /* easy for us: the garbage collector does the dealloc'n */
585 /* -----------------------------------------------------------------------------
587 * -------------------------------------------------------------------------- */
589 /* -----------------------------------------------------------------------------
592 * Approximate how much we've allocated: number of blocks in the
593 * nursery + blocks allocated via allocate() - unused nusery blocks.
594 * This leaves a little slop at the end of each block, and doesn't
595 * take into account large objects (ToDo).
596 * -------------------------------------------------------------------------- */
599 calcAllocated( void )
607 /* All tasks must be stopped. Can't assert that all the
608 capabilities are owned by the scheduler, though: one or more
609 tasks might have been stopped while they were running (non-main)
611 /* ASSERT(n_free_capabilities == RtsFlags.ParFlags.nNodes); */
614 n_free_capabilities * RtsFlags.GcFlags.minAllocAreaSize * BLOCK_SIZE_W
617 for (cap = free_capabilities; cap != NULL; cap = cap->link) {
618 for ( bd = cap->rCurrentNursery->link; bd != NULL; bd = bd->link ) {
619 allocated -= BLOCK_SIZE_W;
621 if (cap->rCurrentNursery->free < cap->rCurrentNursery->start
623 allocated -= (cap->rCurrentNursery->start + BLOCK_SIZE_W)
624 - cap->rCurrentNursery->free;
629 bdescr *current_nursery = MainRegTable.rCurrentNursery;
631 allocated = (nursery_blocks * BLOCK_SIZE_W) + allocated_bytes();
632 for ( bd = current_nursery->link; bd != NULL; bd = bd->link ) {
633 allocated -= BLOCK_SIZE_W;
635 if (current_nursery->free < current_nursery->start + BLOCK_SIZE_W) {
636 allocated -= (current_nursery->start + BLOCK_SIZE_W)
637 - current_nursery->free;
641 total_allocated += allocated;
645 /* Approximate the amount of live data in the heap. To be called just
646 * after garbage collection (see GarbageCollect()).
655 if (RtsFlags.GcFlags.generations == 1) {
656 live = (g0s0->n_to_blocks - 1) * BLOCK_SIZE_W +
657 ((lnat)g0s0->hp_bd->free - (lnat)g0s0->hp_bd->start) / sizeof(W_);
661 for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
662 for (s = 0; s < generations[g].n_steps; s++) {
663 /* approximate amount of live data (doesn't take into account slop
664 * at end of each block).
666 if (g == 0 && s == 0) {
669 stp = &generations[g].steps[s];
670 live += (stp->n_large_blocks + stp->n_blocks - 1) * BLOCK_SIZE_W;
671 if (stp->hp_bd != NULL) {
672 live += ((lnat)stp->hp_bd->free - (lnat)stp->hp_bd->start)
680 /* Approximate the number of blocks that will be needed at the next
681 * garbage collection.
683 * Assume: all data currently live will remain live. Steps that will
684 * be collected next time will therefore need twice as many blocks
685 * since all the data will be copied.
694 for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
695 for (s = 0; s < generations[g].n_steps; s++) {
696 if (g == 0 && s == 0) { continue; }
697 stp = &generations[g].steps[s];
698 if (generations[g].steps[0].n_blocks +
699 generations[g].steps[0].n_large_blocks
700 > generations[g].max_blocks
701 && stp->is_compacted == 0) {
702 needed += 2 * stp->n_blocks;
704 needed += stp->n_blocks;
711 /* -----------------------------------------------------------------------------
714 memInventory() checks for memory leaks by counting up all the
715 blocks we know about and comparing that to the number of blocks
716 allegedly floating around in the system.
717 -------------------------------------------------------------------------- */
727 lnat total_blocks = 0, free_blocks = 0;
729 /* count the blocks we current have */
731 for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
732 for (s = 0; s < generations[g].n_steps; s++) {
733 stp = &generations[g].steps[s];
734 total_blocks += stp->n_blocks;
735 if (RtsFlags.GcFlags.generations == 1) {
736 /* two-space collector has a to-space too :-) */
737 total_blocks += g0s0->n_to_blocks;
739 for (bd = stp->large_objects; bd; bd = bd->link) {
740 total_blocks += bd->blocks;
741 /* hack for megablock groups: they have an extra block or two in
742 the second and subsequent megablocks where the block
743 descriptors would normally go.
745 if (bd->blocks > BLOCKS_PER_MBLOCK) {
746 total_blocks -= (MBLOCK_SIZE / BLOCK_SIZE - BLOCKS_PER_MBLOCK)
747 * (bd->blocks/(MBLOCK_SIZE/BLOCK_SIZE));
753 /* any blocks held by allocate() */
754 for (bd = small_alloc_list; bd; bd = bd->link) {
755 total_blocks += bd->blocks;
757 for (bd = large_alloc_list; bd; bd = bd->link) {
758 total_blocks += bd->blocks;
761 /* count the blocks on the free list */
762 free_blocks = countFreeList();
764 if (total_blocks + free_blocks != mblocks_allocated *
766 fprintf(stderr, "Blocks: %ld live + %ld free = %ld total (%ld around)\n",
767 total_blocks, free_blocks, total_blocks + free_blocks,
768 mblocks_allocated * BLOCKS_PER_MBLOCK);
771 ASSERT(total_blocks + free_blocks == mblocks_allocated * BLOCKS_PER_MBLOCK);
775 countBlocks(bdescr *bd)
778 for (n=0; bd != NULL; bd=bd->link) {
784 /* Full heap sanity check. */
790 if (RtsFlags.GcFlags.generations == 1) {
791 checkHeap(g0s0->to_blocks);
792 checkChain(g0s0->large_objects);
795 for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
796 for (s = 0; s < generations[g].n_steps; s++) {
797 if (g == 0 && s == 0) { continue; }
798 checkHeap(generations[g].steps[s].blocks);
799 checkChain(generations[g].steps[s].large_objects);
800 ASSERT(countBlocks(generations[g].steps[s].blocks)
801 == generations[g].steps[s].n_blocks);
802 ASSERT(countBlocks(generations[g].steps[s].large_objects)
803 == generations[g].steps[s].n_large_blocks);
805 checkMutableList(generations[g].mut_list, g);
806 checkMutOnceList(generations[g].mut_once_list, g);
810 checkFreeListSanity();