1 /* -----------------------------------------------------------------------------
2 * $Id: Storage.c,v 1.49 2001/08/14 13:40:09 sewardj Exp $
4 * (c) The GHC Team, 1998-1999
6 * Storage manager front end
8 * ---------------------------------------------------------------------------*/
10 #include "PosixSource.h"
16 #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 bdescr *pinned_object_block; /* allocate pinned objects into this block */
34 nat alloc_blocks; /* number of allocate()d blocks since GC */
35 nat alloc_blocks_lim; /* approximate limit on alloc_blocks */
37 StgPtr alloc_Hp = NULL; /* next free byte in small_alloc_list */
38 StgPtr alloc_HpLim = NULL; /* end of block at small_alloc_list */
40 generation *generations; /* all the generations */
41 generation *g0; /* generation 0, for convenience */
42 generation *oldest_gen; /* oldest generation, for convenience */
43 step *g0s0; /* generation 0, step 0, for convenience */
45 lnat total_allocated = 0; /* total memory allocated during run */
48 * Storage manager mutex: protects all the above state from
49 * simultaneous access by two STG threads.
52 pthread_mutex_t sm_mutex = PTHREAD_MUTEX_INITIALIZER;
58 static void *stgAllocForGMP (size_t size_in_bytes);
59 static void *stgReallocForGMP (void *ptr, size_t old_size, size_t new_size);
60 static void stgDeallocForGMP (void *ptr, size_t size);
69 /* If we're doing heap profiling, we want a two-space heap with a
70 * fixed-size allocation area so that we get roughly even-spaced
74 /* As an experiment, try a 2 generation collector
77 #if defined(PROFILING) || defined(DEBUG)
78 if (RtsFlags.ProfFlags.doHeapProfile) {
79 RtsFlags.GcFlags.generations = 1;
80 RtsFlags.GcFlags.steps = 1;
81 RtsFlags.GcFlags.oldGenFactor = 0;
82 RtsFlags.GcFlags.heapSizeSuggestion = 0;
86 if (RtsFlags.GcFlags.maxHeapSize != 0 &&
87 RtsFlags.GcFlags.heapSizeSuggestion >
88 RtsFlags.GcFlags.maxHeapSize) {
89 RtsFlags.GcFlags.maxHeapSize = RtsFlags.GcFlags.heapSizeSuggestion;
94 /* allocate generation info array */
95 generations = (generation *)stgMallocBytes(RtsFlags.GcFlags.generations
96 * sizeof(struct _generation),
99 /* Initialise all generations */
100 for(g = 0; g < RtsFlags.GcFlags.generations; g++) {
101 gen = &generations[g];
103 gen->mut_list = END_MUT_LIST;
104 gen->mut_once_list = END_MUT_LIST;
105 gen->collections = 0;
106 gen->failed_promotions = 0;
110 /* A couple of convenience pointers */
111 g0 = &generations[0];
112 oldest_gen = &generations[RtsFlags.GcFlags.generations-1];
114 /* Allocate step structures in each generation */
115 if (RtsFlags.GcFlags.generations > 1) {
116 /* Only for multiple-generations */
118 /* Oldest generation: one step */
119 oldest_gen->n_steps = 1;
121 stgMallocBytes(1 * sizeof(struct _step), "initStorage: last step");
123 /* set up all except the oldest generation with 2 steps */
124 for(g = 0; g < RtsFlags.GcFlags.generations-1; g++) {
125 generations[g].n_steps = RtsFlags.GcFlags.steps;
126 generations[g].steps =
127 stgMallocBytes (RtsFlags.GcFlags.steps * sizeof(struct _step),
128 "initStorage: steps");
132 /* single generation, i.e. a two-space collector */
134 g0->steps = stgMallocBytes (sizeof(struct _step), "initStorage: steps");
137 /* Initialise all steps */
138 for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
139 for (s = 0; s < generations[g].n_steps; s++) {
140 stp = &generations[g].steps[s];
144 stp->gen = &generations[g];
151 stp->large_objects = NULL;
152 stp->n_large_blocks = 0;
153 stp->new_large_objects = NULL;
154 stp->scavenged_large_objects = NULL;
155 stp->n_scavenged_large_blocks = 0;
156 stp->is_compacted = 0;
161 /* Set up the destination pointers in each younger gen. step */
162 for (g = 0; g < RtsFlags.GcFlags.generations-1; g++) {
163 for (s = 0; s < generations[g].n_steps-1; s++) {
164 generations[g].steps[s].to = &generations[g].steps[s+1];
166 generations[g].steps[s].to = &generations[g+1].steps[0];
169 /* The oldest generation has one step and it is compacted. */
170 if (RtsFlags.GcFlags.compact) {
171 oldest_gen->steps[0].is_compacted = 1;
173 oldest_gen->steps[0].to = &oldest_gen->steps[0];
175 /* generation 0 is special: that's the nursery */
176 generations[0].max_blocks = 0;
178 /* G0S0: the allocation area. Policy: keep the allocation area
179 * small to begin with, even if we have a large suggested heap
180 * size. Reason: we're going to do a major collection first, and we
181 * don't want it to be a big one. This vague idea is borne out by
182 * rigorous experimental evidence.
184 g0s0 = &generations[0].steps[0];
188 weak_ptr_list = NULL;
191 /* initialise the allocate() interface */
192 small_alloc_list = NULL;
193 large_alloc_list = NULL;
195 alloc_blocks_lim = RtsFlags.GcFlags.minAllocAreaSize;
197 /* Tell GNU multi-precision pkg about our custom alloc functions */
198 mp_set_memory_functions(stgAllocForGMP, stgReallocForGMP, stgDeallocForGMP);
201 pthread_mutex_init(&sm_mutex, NULL);
204 IF_DEBUG(gc, statDescribeGens());
210 stat_exit(calcAllocated());
213 /* -----------------------------------------------------------------------------
216 The entry code for every CAF does the following:
218 - builds a CAF_BLACKHOLE in the heap
219 - pushes an update frame pointing to the CAF_BLACKHOLE
220 - invokes UPD_CAF(), which:
221 - calls newCaf, below
222 - updates the CAF with a static indirection to the CAF_BLACKHOLE
224 Why do we build a BLACKHOLE in the heap rather than just updating
225 the thunk directly? It's so that we only need one kind of update
226 frame - otherwise we'd need a static version of the update frame too.
228 newCaf() does the following:
230 - it puts the CAF on the oldest generation's mut-once list.
231 This is so that we can treat the CAF as a root when collecting
234 For GHCI, we have additional requirements when dealing with CAFs:
236 - we must *retain* all dynamically-loaded CAFs ever entered,
237 just in case we need them again.
238 - we must be able to *revert* CAFs that have been evaluated, to
239 their pre-evaluated form.
241 To do this, we use an additional CAF list. When newCaf() is
242 called on a dynamically-loaded CAF, we add it to the CAF list
243 instead of the old-generation mutable list, and save away its
244 old info pointer (in caf->saved_info) for later reversion.
246 To revert all the CAFs, we traverse the CAF list and reset the
247 info pointer to caf->saved_info, then throw away the CAF list.
248 (see GC.c:revertCAFs()).
252 -------------------------------------------------------------------------- */
255 newCAF(StgClosure* caf)
257 /* Put this CAF on the mutable list for the old generation.
258 * This is a HACK - the IND_STATIC closure doesn't really have
259 * a mut_link field, but we pretend it has - in fact we re-use
260 * the STATIC_LINK field for the time being, because when we
261 * come to do a major GC we won't need the mut_link field
262 * any more and can use it as a STATIC_LINK.
264 ACQUIRE_LOCK(&sm_mutex);
266 if (is_dynamically_loaded_rwdata_ptr((StgPtr)caf)) {
267 ((StgIndStatic *)caf)->saved_info = (StgInfoTable *)caf->header.info;
268 ((StgIndStatic *)caf)->static_link = caf_list;
271 ((StgIndStatic *)caf)->saved_info = NULL;
272 ((StgMutClosure *)caf)->mut_link = oldest_gen->mut_once_list;
273 oldest_gen->mut_once_list = (StgMutClosure *)caf;
276 RELEASE_LOCK(&sm_mutex);
279 /* If we are PAR or DIST then we never forget a CAF */
281 //belch("<##> Globalising CAF %08x %s",caf,info_type(caf));
282 newGA=makeGlobal(caf,rtsTrue); /*given full weight*/
288 /* -----------------------------------------------------------------------------
290 -------------------------------------------------------------------------- */
293 allocNurseries( void )
302 for (cap = free_capabilities; cap != NULL; cap = cap->link) {
303 cap->rNursery = allocNursery(NULL, RtsFlags.GcFlags.minAllocAreaSize);
304 cap->rCurrentNursery = cap->rNursery;
305 for (bd = cap->rNursery; bd != NULL; bd = bd->link) {
306 bd->u.back = (bdescr *)cap;
309 /* Set the back links to be equal to the Capability,
310 * so we can do slightly better informed locking.
314 nursery_blocks = RtsFlags.GcFlags.minAllocAreaSize;
315 g0s0->blocks = allocNursery(NULL, nursery_blocks);
316 g0s0->n_blocks = nursery_blocks;
317 g0s0->to_blocks = NULL;
318 g0s0->n_to_blocks = 0;
319 MainRegTable.rNursery = g0s0->blocks;
320 MainRegTable.rCurrentNursery = g0s0->blocks;
321 /* hp, hpLim, hp_bd, to_space etc. aren't used in G0S0 */
326 resetNurseries( void )
332 /* All tasks must be stopped */
333 ASSERT(n_free_capabilities == RtsFlags.ParFlags.nNodes);
335 for (cap = free_capabilities; cap != NULL; cap = cap->link) {
336 for (bd = cap->rNursery; bd; bd = bd->link) {
337 bd->free = bd->start;
338 ASSERT(bd->gen_no == 0);
339 ASSERT(bd->step == g0s0);
340 IF_DEBUG(sanity,memset(bd->start, 0xaa, BLOCK_SIZE));
342 cap->rCurrentNursery = cap->rNursery;
345 for (bd = g0s0->blocks; bd; bd = bd->link) {
346 bd->free = bd->start;
347 ASSERT(bd->gen_no == 0);
348 ASSERT(bd->step == g0s0);
349 IF_DEBUG(sanity,memset(bd->start, 0xaa, BLOCK_SIZE));
351 MainRegTable.rNursery = g0s0->blocks;
352 MainRegTable.rCurrentNursery = g0s0->blocks;
357 allocNursery (bdescr *last_bd, nat blocks)
362 /* Allocate a nursery */
363 for (i=0; i < blocks; i++) {
369 bd->free = bd->start;
376 resizeNursery ( nat blocks )
381 barf("resizeNursery: can't resize in SMP mode");
384 if (nursery_blocks == blocks) {
385 ASSERT(g0s0->n_blocks == blocks);
389 else if (nursery_blocks < blocks) {
390 IF_DEBUG(gc, fprintf(stderr, "Increasing size of nursery to %d blocks\n",
392 g0s0->blocks = allocNursery(g0s0->blocks, blocks-nursery_blocks);
398 IF_DEBUG(gc, fprintf(stderr, "Decreasing size of nursery to %d blocks\n",
400 for (bd = g0s0->blocks; nursery_blocks > blocks; nursery_blocks--) {
408 g0s0->n_blocks = nursery_blocks = blocks;
411 /* -----------------------------------------------------------------------------
412 The allocate() interface
414 allocate(n) always succeeds, and returns a chunk of memory n words
415 long. n can be larger than the size of a block if necessary, in
416 which case a contiguous block group will be allocated.
417 -------------------------------------------------------------------------- */
425 ACQUIRE_LOCK(&sm_mutex);
427 TICK_ALLOC_HEAP_NOCTR(n);
430 /* big allocation (>LARGE_OBJECT_THRESHOLD) */
431 /* ToDo: allocate directly into generation 1 */
432 if (n >= LARGE_OBJECT_THRESHOLD/sizeof(W_)) {
433 nat req_blocks = (lnat)BLOCK_ROUND_UP(n*sizeof(W_)) / BLOCK_SIZE;
434 bd = allocGroup(req_blocks);
435 dbl_link_onto(bd, &g0s0->large_objects);
438 bd->flags = BF_LARGE;
439 bd->free = bd->start;
440 /* don't add these blocks to alloc_blocks, since we're assuming
441 * that large objects are likely to remain live for quite a while
442 * (eg. running threads), so garbage collecting early won't make
445 alloc_blocks += req_blocks;
446 RELEASE_LOCK(&sm_mutex);
449 /* small allocation (<LARGE_OBJECT_THRESHOLD) */
450 } else if (small_alloc_list == NULL || alloc_Hp + n > alloc_HpLim) {
451 if (small_alloc_list) {
452 small_alloc_list->free = alloc_Hp;
455 bd->link = small_alloc_list;
456 small_alloc_list = bd;
460 alloc_Hp = bd->start;
461 alloc_HpLim = bd->start + BLOCK_SIZE_W;
467 RELEASE_LOCK(&sm_mutex);
472 allocated_bytes( void )
474 return (alloc_blocks * BLOCK_SIZE_W - (alloc_HpLim - alloc_Hp));
477 /* ---------------------------------------------------------------------------
478 Allocate a fixed/pinned object.
480 We allocate small pinned objects into a single block, allocating a
481 new block when the current one overflows. The block is chained
482 onto the large_object_list of generation 0 step 0.
484 NOTE: The GC can't in general handle pinned objects. This
485 interface is only safe to use for ByteArrays, which have no
486 pointers and don't require scavenging. It works because the
487 block's descriptor has the BF_LARGE flag set, so the block is
488 treated as a large object and chained onto various lists, rather
489 than the individual objects being copied. However, when it comes
490 to scavenge the block, the GC will only scavenge the first object.
491 The reason is that the GC can't linearly scan a block of pinned
492 objects at the moment (doing so would require using the
493 mostly-copying techniques). But since we're restricting ourselves
494 to pinned ByteArrays, not scavenging is ok.
496 This function is called by newPinnedByteArray# which immediately
497 fills the allocated memory with a MutableByteArray#.
498 ------------------------------------------------------------------------- */
501 allocatePinned( nat n )
504 bdescr *bd = pinned_object_block;
506 ACQUIRE_LOCK(&sm_mutex);
508 TICK_ALLOC_HEAP_NOCTR(n);
511 // If the request is for a large object, then allocate()
512 // will give us a pinned object anyway.
513 if (n >= LARGE_OBJECT_THRESHOLD/sizeof(W_)) {
514 RELEASE_LOCK(&sm_mutex);
518 // If we don't have a block of pinned objects yet, or the current
519 // one isn't large enough to hold the new object, allocate a new one.
520 if (bd == NULL || (bd->free + n) > (bd->start + BLOCK_SIZE_W)) {
521 pinned_object_block = bd = allocBlock();
522 dbl_link_onto(bd, &g0s0->large_objects);
525 bd->flags = BF_LARGE;
526 bd->free = bd->start;
532 RELEASE_LOCK(&sm_mutex);
536 /* -----------------------------------------------------------------------------
537 Allocation functions for GMP.
539 These all use the allocate() interface - we can't have any garbage
540 collection going on during a gmp operation, so we use allocate()
541 which always succeeds. The gmp operations which might need to
542 allocate will ask the storage manager (via doYouWantToGC()) whether
543 a garbage collection is required, in case we get into a loop doing
544 only allocate() style allocation.
545 -------------------------------------------------------------------------- */
548 stgAllocForGMP (size_t size_in_bytes)
551 nat data_size_in_words, total_size_in_words;
553 /* should be a multiple of sizeof(StgWord) (whole no. of limbs) */
554 ASSERT(size_in_bytes % sizeof(W_) == 0);
556 data_size_in_words = size_in_bytes / sizeof(W_);
557 total_size_in_words = sizeofW(StgArrWords) + data_size_in_words;
559 /* allocate and fill it in. */
560 arr = (StgArrWords *)allocate(total_size_in_words);
561 SET_ARR_HDR(arr, &stg_ARR_WORDS_info, CCCS, data_size_in_words);
563 /* and return a ptr to the goods inside the array */
564 return(BYTE_ARR_CTS(arr));
568 stgReallocForGMP (void *ptr, size_t old_size, size_t new_size)
570 void *new_stuff_ptr = stgAllocForGMP(new_size);
572 char *p = (char *) ptr;
573 char *q = (char *) new_stuff_ptr;
575 for (; i < old_size; i++, p++, q++) {
579 return(new_stuff_ptr);
583 stgDeallocForGMP (void *ptr STG_UNUSED,
584 size_t size STG_UNUSED)
586 /* easy for us: the garbage collector does the dealloc'n */
589 /* -----------------------------------------------------------------------------
591 * -------------------------------------------------------------------------- */
593 /* -----------------------------------------------------------------------------
596 * Approximate how much we've allocated: number of blocks in the
597 * nursery + blocks allocated via allocate() - unused nusery blocks.
598 * This leaves a little slop at the end of each block, and doesn't
599 * take into account large objects (ToDo).
600 * -------------------------------------------------------------------------- */
603 calcAllocated( void )
611 /* All tasks must be stopped. Can't assert that all the
612 capabilities are owned by the scheduler, though: one or more
613 tasks might have been stopped while they were running (non-main)
615 /* ASSERT(n_free_capabilities == RtsFlags.ParFlags.nNodes); */
618 n_free_capabilities * RtsFlags.GcFlags.minAllocAreaSize * BLOCK_SIZE_W
621 for (cap = free_capabilities; cap != NULL; cap = cap->link) {
622 for ( bd = cap->rCurrentNursery->link; bd != NULL; bd = bd->link ) {
623 allocated -= BLOCK_SIZE_W;
625 if (cap->rCurrentNursery->free < cap->rCurrentNursery->start
627 allocated -= (cap->rCurrentNursery->start + BLOCK_SIZE_W)
628 - cap->rCurrentNursery->free;
633 bdescr *current_nursery = MainRegTable.rCurrentNursery;
635 allocated = (nursery_blocks * BLOCK_SIZE_W) + allocated_bytes();
636 for ( bd = current_nursery->link; bd != NULL; bd = bd->link ) {
637 allocated -= BLOCK_SIZE_W;
639 if (current_nursery->free < current_nursery->start + BLOCK_SIZE_W) {
640 allocated -= (current_nursery->start + BLOCK_SIZE_W)
641 - current_nursery->free;
645 total_allocated += allocated;
649 /* Approximate the amount of live data in the heap. To be called just
650 * after garbage collection (see GarbageCollect()).
659 if (RtsFlags.GcFlags.generations == 1) {
660 live = (g0s0->n_to_blocks - 1) * BLOCK_SIZE_W +
661 ((lnat)g0s0->hp_bd->free - (lnat)g0s0->hp_bd->start) / sizeof(W_);
665 for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
666 for (s = 0; s < generations[g].n_steps; s++) {
667 /* approximate amount of live data (doesn't take into account slop
668 * at end of each block).
670 if (g == 0 && s == 0) {
673 stp = &generations[g].steps[s];
674 live += (stp->n_large_blocks + stp->n_blocks - 1) * BLOCK_SIZE_W;
675 if (stp->hp_bd != NULL) {
676 live += ((lnat)stp->hp_bd->free - (lnat)stp->hp_bd->start)
684 /* Approximate the number of blocks that will be needed at the next
685 * garbage collection.
687 * Assume: all data currently live will remain live. Steps that will
688 * be collected next time will therefore need twice as many blocks
689 * since all the data will be copied.
698 for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
699 for (s = 0; s < generations[g].n_steps; s++) {
700 if (g == 0 && s == 0) { continue; }
701 stp = &generations[g].steps[s];
702 if (generations[g].steps[0].n_blocks +
703 generations[g].steps[0].n_large_blocks
704 > generations[g].max_blocks
705 && stp->is_compacted == 0) {
706 needed += 2 * stp->n_blocks;
708 needed += stp->n_blocks;
715 /* -----------------------------------------------------------------------------
718 memInventory() checks for memory leaks by counting up all the
719 blocks we know about and comparing that to the number of blocks
720 allegedly floating around in the system.
721 -------------------------------------------------------------------------- */
731 lnat total_blocks = 0, free_blocks = 0;
733 /* count the blocks we current have */
735 for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
736 for (s = 0; s < generations[g].n_steps; s++) {
737 stp = &generations[g].steps[s];
738 total_blocks += stp->n_blocks;
739 if (RtsFlags.GcFlags.generations == 1) {
740 /* two-space collector has a to-space too :-) */
741 total_blocks += g0s0->n_to_blocks;
743 for (bd = stp->large_objects; bd; bd = bd->link) {
744 total_blocks += bd->blocks;
745 /* hack for megablock groups: they have an extra block or two in
746 the second and subsequent megablocks where the block
747 descriptors would normally go.
749 if (bd->blocks > BLOCKS_PER_MBLOCK) {
750 total_blocks -= (MBLOCK_SIZE / BLOCK_SIZE - BLOCKS_PER_MBLOCK)
751 * (bd->blocks/(MBLOCK_SIZE/BLOCK_SIZE));
757 /* any blocks held by allocate() */
758 for (bd = small_alloc_list; bd; bd = bd->link) {
759 total_blocks += bd->blocks;
761 for (bd = large_alloc_list; bd; bd = bd->link) {
762 total_blocks += bd->blocks;
765 /* count the blocks on the free list */
766 free_blocks = countFreeList();
768 if (total_blocks + free_blocks != mblocks_allocated *
770 fprintf(stderr, "Blocks: %ld live + %ld free = %ld total (%ld around)\n",
771 total_blocks, free_blocks, total_blocks + free_blocks,
772 mblocks_allocated * BLOCKS_PER_MBLOCK);
775 ASSERT(total_blocks + free_blocks == mblocks_allocated * BLOCKS_PER_MBLOCK);
779 countBlocks(bdescr *bd)
782 for (n=0; bd != NULL; bd=bd->link) {
788 /* Full heap sanity check. */
794 if (RtsFlags.GcFlags.generations == 1) {
795 checkHeap(g0s0->to_blocks);
796 checkChain(g0s0->large_objects);
799 for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
800 for (s = 0; s < generations[g].n_steps; s++) {
801 if (g == 0 && s == 0) { continue; }
802 checkHeap(generations[g].steps[s].blocks);
803 checkChain(generations[g].steps[s].large_objects);
804 ASSERT(countBlocks(generations[g].steps[s].blocks)
805 == generations[g].steps[s].n_blocks);
806 ASSERT(countBlocks(generations[g].steps[s].large_objects)
807 == generations[g].steps[s].n_large_blocks);
809 checkMutableList(generations[g].mut_list, g);
810 checkMutOnceList(generations[g].mut_once_list, g);
814 checkFreeListSanity();