1 /* -----------------------------------------------------------------------------
2 * $Id: Storage.c,v 1.50 2001/08/30 10:21:40 simonmar 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 if (RtsFlags.GcFlags.generations == 1) {
172 belch("WARNING: compaction is incompatible with -G1; disabled");
174 oldest_gen->steps[0].is_compacted = 1;
177 oldest_gen->steps[0].to = &oldest_gen->steps[0];
179 /* generation 0 is special: that's the nursery */
180 generations[0].max_blocks = 0;
182 /* G0S0: the allocation area. Policy: keep the allocation area
183 * small to begin with, even if we have a large suggested heap
184 * size. Reason: we're going to do a major collection first, and we
185 * don't want it to be a big one. This vague idea is borne out by
186 * rigorous experimental evidence.
188 g0s0 = &generations[0].steps[0];
192 weak_ptr_list = NULL;
195 /* initialise the allocate() interface */
196 small_alloc_list = NULL;
197 large_alloc_list = NULL;
199 alloc_blocks_lim = RtsFlags.GcFlags.minAllocAreaSize;
201 /* Tell GNU multi-precision pkg about our custom alloc functions */
202 mp_set_memory_functions(stgAllocForGMP, stgReallocForGMP, stgDeallocForGMP);
205 pthread_mutex_init(&sm_mutex, NULL);
208 IF_DEBUG(gc, statDescribeGens());
214 stat_exit(calcAllocated());
217 /* -----------------------------------------------------------------------------
220 The entry code for every CAF does the following:
222 - builds a CAF_BLACKHOLE in the heap
223 - pushes an update frame pointing to the CAF_BLACKHOLE
224 - invokes UPD_CAF(), which:
225 - calls newCaf, below
226 - updates the CAF with a static indirection to the CAF_BLACKHOLE
228 Why do we build a BLACKHOLE in the heap rather than just updating
229 the thunk directly? It's so that we only need one kind of update
230 frame - otherwise we'd need a static version of the update frame too.
232 newCaf() does the following:
234 - it puts the CAF on the oldest generation's mut-once list.
235 This is so that we can treat the CAF as a root when collecting
238 For GHCI, we have additional requirements when dealing with CAFs:
240 - we must *retain* all dynamically-loaded CAFs ever entered,
241 just in case we need them again.
242 - we must be able to *revert* CAFs that have been evaluated, to
243 their pre-evaluated form.
245 To do this, we use an additional CAF list. When newCaf() is
246 called on a dynamically-loaded CAF, we add it to the CAF list
247 instead of the old-generation mutable list, and save away its
248 old info pointer (in caf->saved_info) for later reversion.
250 To revert all the CAFs, we traverse the CAF list and reset the
251 info pointer to caf->saved_info, then throw away the CAF list.
252 (see GC.c:revertCAFs()).
256 -------------------------------------------------------------------------- */
259 newCAF(StgClosure* caf)
261 /* Put this CAF on the mutable list for the old generation.
262 * This is a HACK - the IND_STATIC closure doesn't really have
263 * a mut_link field, but we pretend it has - in fact we re-use
264 * the STATIC_LINK field for the time being, because when we
265 * come to do a major GC we won't need the mut_link field
266 * any more and can use it as a STATIC_LINK.
268 ACQUIRE_LOCK(&sm_mutex);
270 if (is_dynamically_loaded_rwdata_ptr((StgPtr)caf)) {
271 ((StgIndStatic *)caf)->saved_info = (StgInfoTable *)caf->header.info;
272 ((StgIndStatic *)caf)->static_link = caf_list;
275 ((StgIndStatic *)caf)->saved_info = NULL;
276 ((StgMutClosure *)caf)->mut_link = oldest_gen->mut_once_list;
277 oldest_gen->mut_once_list = (StgMutClosure *)caf;
280 RELEASE_LOCK(&sm_mutex);
283 /* If we are PAR or DIST then we never forget a CAF */
285 //belch("<##> Globalising CAF %08x %s",caf,info_type(caf));
286 newGA=makeGlobal(caf,rtsTrue); /*given full weight*/
292 /* -----------------------------------------------------------------------------
294 -------------------------------------------------------------------------- */
297 allocNurseries( void )
306 for (cap = free_capabilities; cap != NULL; cap = cap->link) {
307 cap->rNursery = allocNursery(NULL, RtsFlags.GcFlags.minAllocAreaSize);
308 cap->rCurrentNursery = cap->rNursery;
309 for (bd = cap->rNursery; bd != NULL; bd = bd->link) {
310 bd->u.back = (bdescr *)cap;
313 /* Set the back links to be equal to the Capability,
314 * so we can do slightly better informed locking.
318 nursery_blocks = RtsFlags.GcFlags.minAllocAreaSize;
319 g0s0->blocks = allocNursery(NULL, nursery_blocks);
320 g0s0->n_blocks = nursery_blocks;
321 g0s0->to_blocks = NULL;
322 g0s0->n_to_blocks = 0;
323 MainRegTable.rNursery = g0s0->blocks;
324 MainRegTable.rCurrentNursery = g0s0->blocks;
325 /* hp, hpLim, hp_bd, to_space etc. aren't used in G0S0 */
330 resetNurseries( void )
336 /* All tasks must be stopped */
337 ASSERT(n_free_capabilities == RtsFlags.ParFlags.nNodes);
339 for (cap = free_capabilities; cap != NULL; cap = cap->link) {
340 for (bd = cap->rNursery; bd; bd = bd->link) {
341 bd->free = bd->start;
342 ASSERT(bd->gen_no == 0);
343 ASSERT(bd->step == g0s0);
344 IF_DEBUG(sanity,memset(bd->start, 0xaa, BLOCK_SIZE));
346 cap->rCurrentNursery = cap->rNursery;
349 for (bd = g0s0->blocks; bd; bd = bd->link) {
350 bd->free = bd->start;
351 ASSERT(bd->gen_no == 0);
352 ASSERT(bd->step == g0s0);
353 IF_DEBUG(sanity,memset(bd->start, 0xaa, BLOCK_SIZE));
355 MainRegTable.rNursery = g0s0->blocks;
356 MainRegTable.rCurrentNursery = g0s0->blocks;
361 allocNursery (bdescr *last_bd, nat blocks)
366 /* Allocate a nursery */
367 for (i=0; i < blocks; i++) {
373 bd->free = bd->start;
380 resizeNursery ( nat blocks )
385 barf("resizeNursery: can't resize in SMP mode");
388 if (nursery_blocks == blocks) {
389 ASSERT(g0s0->n_blocks == blocks);
393 else if (nursery_blocks < blocks) {
394 IF_DEBUG(gc, fprintf(stderr, "Increasing size of nursery to %d blocks\n",
396 g0s0->blocks = allocNursery(g0s0->blocks, blocks-nursery_blocks);
402 IF_DEBUG(gc, fprintf(stderr, "Decreasing size of nursery to %d blocks\n",
404 for (bd = g0s0->blocks; nursery_blocks > blocks; nursery_blocks--) {
412 g0s0->n_blocks = nursery_blocks = blocks;
415 /* -----------------------------------------------------------------------------
416 The allocate() interface
418 allocate(n) always succeeds, and returns a chunk of memory n words
419 long. n can be larger than the size of a block if necessary, in
420 which case a contiguous block group will be allocated.
421 -------------------------------------------------------------------------- */
429 ACQUIRE_LOCK(&sm_mutex);
431 TICK_ALLOC_HEAP_NOCTR(n);
434 /* big allocation (>LARGE_OBJECT_THRESHOLD) */
435 /* ToDo: allocate directly into generation 1 */
436 if (n >= LARGE_OBJECT_THRESHOLD/sizeof(W_)) {
437 nat req_blocks = (lnat)BLOCK_ROUND_UP(n*sizeof(W_)) / BLOCK_SIZE;
438 bd = allocGroup(req_blocks);
439 dbl_link_onto(bd, &g0s0->large_objects);
442 bd->flags = BF_LARGE;
443 bd->free = bd->start;
444 /* don't add these blocks to alloc_blocks, since we're assuming
445 * that large objects are likely to remain live for quite a while
446 * (eg. running threads), so garbage collecting early won't make
449 alloc_blocks += req_blocks;
450 RELEASE_LOCK(&sm_mutex);
453 /* small allocation (<LARGE_OBJECT_THRESHOLD) */
454 } else if (small_alloc_list == NULL || alloc_Hp + n > alloc_HpLim) {
455 if (small_alloc_list) {
456 small_alloc_list->free = alloc_Hp;
459 bd->link = small_alloc_list;
460 small_alloc_list = bd;
464 alloc_Hp = bd->start;
465 alloc_HpLim = bd->start + BLOCK_SIZE_W;
471 RELEASE_LOCK(&sm_mutex);
476 allocated_bytes( void )
478 return (alloc_blocks * BLOCK_SIZE_W - (alloc_HpLim - alloc_Hp));
481 /* ---------------------------------------------------------------------------
482 Allocate a fixed/pinned object.
484 We allocate small pinned objects into a single block, allocating a
485 new block when the current one overflows. The block is chained
486 onto the large_object_list of generation 0 step 0.
488 NOTE: The GC can't in general handle pinned objects. This
489 interface is only safe to use for ByteArrays, which have no
490 pointers and don't require scavenging. It works because the
491 block's descriptor has the BF_LARGE flag set, so the block is
492 treated as a large object and chained onto various lists, rather
493 than the individual objects being copied. However, when it comes
494 to scavenge the block, the GC will only scavenge the first object.
495 The reason is that the GC can't linearly scan a block of pinned
496 objects at the moment (doing so would require using the
497 mostly-copying techniques). But since we're restricting ourselves
498 to pinned ByteArrays, not scavenging is ok.
500 This function is called by newPinnedByteArray# which immediately
501 fills the allocated memory with a MutableByteArray#.
502 ------------------------------------------------------------------------- */
505 allocatePinned( nat n )
508 bdescr *bd = pinned_object_block;
510 ACQUIRE_LOCK(&sm_mutex);
512 TICK_ALLOC_HEAP_NOCTR(n);
515 // If the request is for a large object, then allocate()
516 // will give us a pinned object anyway.
517 if (n >= LARGE_OBJECT_THRESHOLD/sizeof(W_)) {
518 RELEASE_LOCK(&sm_mutex);
522 // If we don't have a block of pinned objects yet, or the current
523 // one isn't large enough to hold the new object, allocate a new one.
524 if (bd == NULL || (bd->free + n) > (bd->start + BLOCK_SIZE_W)) {
525 pinned_object_block = bd = allocBlock();
526 dbl_link_onto(bd, &g0s0->large_objects);
529 bd->flags = BF_LARGE;
530 bd->free = bd->start;
536 RELEASE_LOCK(&sm_mutex);
540 /* -----------------------------------------------------------------------------
541 Allocation functions for GMP.
543 These all use the allocate() interface - we can't have any garbage
544 collection going on during a gmp operation, so we use allocate()
545 which always succeeds. The gmp operations which might need to
546 allocate will ask the storage manager (via doYouWantToGC()) whether
547 a garbage collection is required, in case we get into a loop doing
548 only allocate() style allocation.
549 -------------------------------------------------------------------------- */
552 stgAllocForGMP (size_t size_in_bytes)
555 nat data_size_in_words, total_size_in_words;
557 /* should be a multiple of sizeof(StgWord) (whole no. of limbs) */
558 ASSERT(size_in_bytes % sizeof(W_) == 0);
560 data_size_in_words = size_in_bytes / sizeof(W_);
561 total_size_in_words = sizeofW(StgArrWords) + data_size_in_words;
563 /* allocate and fill it in. */
564 arr = (StgArrWords *)allocate(total_size_in_words);
565 SET_ARR_HDR(arr, &stg_ARR_WORDS_info, CCCS, data_size_in_words);
567 /* and return a ptr to the goods inside the array */
568 return(BYTE_ARR_CTS(arr));
572 stgReallocForGMP (void *ptr, size_t old_size, size_t new_size)
574 void *new_stuff_ptr = stgAllocForGMP(new_size);
576 char *p = (char *) ptr;
577 char *q = (char *) new_stuff_ptr;
579 for (; i < old_size; i++, p++, q++) {
583 return(new_stuff_ptr);
587 stgDeallocForGMP (void *ptr STG_UNUSED,
588 size_t size STG_UNUSED)
590 /* easy for us: the garbage collector does the dealloc'n */
593 /* -----------------------------------------------------------------------------
595 * -------------------------------------------------------------------------- */
597 /* -----------------------------------------------------------------------------
600 * Approximate how much we've allocated: number of blocks in the
601 * nursery + blocks allocated via allocate() - unused nusery blocks.
602 * This leaves a little slop at the end of each block, and doesn't
603 * take into account large objects (ToDo).
604 * -------------------------------------------------------------------------- */
607 calcAllocated( void )
615 /* All tasks must be stopped. Can't assert that all the
616 capabilities are owned by the scheduler, though: one or more
617 tasks might have been stopped while they were running (non-main)
619 /* ASSERT(n_free_capabilities == RtsFlags.ParFlags.nNodes); */
622 n_free_capabilities * RtsFlags.GcFlags.minAllocAreaSize * BLOCK_SIZE_W
625 for (cap = free_capabilities; cap != NULL; cap = cap->link) {
626 for ( bd = cap->rCurrentNursery->link; bd != NULL; bd = bd->link ) {
627 allocated -= BLOCK_SIZE_W;
629 if (cap->rCurrentNursery->free < cap->rCurrentNursery->start
631 allocated -= (cap->rCurrentNursery->start + BLOCK_SIZE_W)
632 - cap->rCurrentNursery->free;
637 bdescr *current_nursery = MainRegTable.rCurrentNursery;
639 allocated = (nursery_blocks * BLOCK_SIZE_W) + allocated_bytes();
640 for ( bd = current_nursery->link; bd != NULL; bd = bd->link ) {
641 allocated -= BLOCK_SIZE_W;
643 if (current_nursery->free < current_nursery->start + BLOCK_SIZE_W) {
644 allocated -= (current_nursery->start + BLOCK_SIZE_W)
645 - current_nursery->free;
649 total_allocated += allocated;
653 /* Approximate the amount of live data in the heap. To be called just
654 * after garbage collection (see GarbageCollect()).
663 if (RtsFlags.GcFlags.generations == 1) {
664 live = (g0s0->n_to_blocks - 1) * BLOCK_SIZE_W +
665 ((lnat)g0s0->hp_bd->free - (lnat)g0s0->hp_bd->start) / sizeof(W_);
669 for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
670 for (s = 0; s < generations[g].n_steps; s++) {
671 /* approximate amount of live data (doesn't take into account slop
672 * at end of each block).
674 if (g == 0 && s == 0) {
677 stp = &generations[g].steps[s];
678 live += (stp->n_large_blocks + stp->n_blocks - 1) * BLOCK_SIZE_W;
679 if (stp->hp_bd != NULL) {
680 live += ((lnat)stp->hp_bd->free - (lnat)stp->hp_bd->start)
688 /* Approximate the number of blocks that will be needed at the next
689 * garbage collection.
691 * Assume: all data currently live will remain live. Steps that will
692 * be collected next time will therefore need twice as many blocks
693 * since all the data will be copied.
702 for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
703 for (s = 0; s < generations[g].n_steps; s++) {
704 if (g == 0 && s == 0) { continue; }
705 stp = &generations[g].steps[s];
706 if (generations[g].steps[0].n_blocks +
707 generations[g].steps[0].n_large_blocks
708 > generations[g].max_blocks
709 && stp->is_compacted == 0) {
710 needed += 2 * stp->n_blocks;
712 needed += stp->n_blocks;
719 /* -----------------------------------------------------------------------------
722 memInventory() checks for memory leaks by counting up all the
723 blocks we know about and comparing that to the number of blocks
724 allegedly floating around in the system.
725 -------------------------------------------------------------------------- */
735 lnat total_blocks = 0, free_blocks = 0;
737 /* count the blocks we current have */
739 for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
740 for (s = 0; s < generations[g].n_steps; s++) {
741 stp = &generations[g].steps[s];
742 total_blocks += stp->n_blocks;
743 if (RtsFlags.GcFlags.generations == 1) {
744 /* two-space collector has a to-space too :-) */
745 total_blocks += g0s0->n_to_blocks;
747 for (bd = stp->large_objects; bd; bd = bd->link) {
748 total_blocks += bd->blocks;
749 /* hack for megablock groups: they have an extra block or two in
750 the second and subsequent megablocks where the block
751 descriptors would normally go.
753 if (bd->blocks > BLOCKS_PER_MBLOCK) {
754 total_blocks -= (MBLOCK_SIZE / BLOCK_SIZE - BLOCKS_PER_MBLOCK)
755 * (bd->blocks/(MBLOCK_SIZE/BLOCK_SIZE));
761 /* any blocks held by allocate() */
762 for (bd = small_alloc_list; bd; bd = bd->link) {
763 total_blocks += bd->blocks;
765 for (bd = large_alloc_list; bd; bd = bd->link) {
766 total_blocks += bd->blocks;
769 /* count the blocks on the free list */
770 free_blocks = countFreeList();
772 if (total_blocks + free_blocks != mblocks_allocated *
774 fprintf(stderr, "Blocks: %ld live + %ld free = %ld total (%ld around)\n",
775 total_blocks, free_blocks, total_blocks + free_blocks,
776 mblocks_allocated * BLOCKS_PER_MBLOCK);
779 ASSERT(total_blocks + free_blocks == mblocks_allocated * BLOCKS_PER_MBLOCK);
783 countBlocks(bdescr *bd)
786 for (n=0; bd != NULL; bd=bd->link) {
792 /* Full heap sanity check. */
798 if (RtsFlags.GcFlags.generations == 1) {
799 checkHeap(g0s0->to_blocks);
800 checkChain(g0s0->large_objects);
803 for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
804 for (s = 0; s < generations[g].n_steps; s++) {
805 if (g == 0 && s == 0) { continue; }
806 checkHeap(generations[g].steps[s].blocks);
807 checkChain(generations[g].steps[s].large_objects);
808 ASSERT(countBlocks(generations[g].steps[s].blocks)
809 == generations[g].steps[s].n_blocks);
810 ASSERT(countBlocks(generations[g].steps[s].large_objects)
811 == generations[g].steps[s].n_large_blocks);
813 checkMutableList(generations[g].mut_list, g);
814 checkMutOnceList(generations[g].mut_once_list, g);
818 checkFreeListSanity();