1 /* -----------------------------------------------------------------------------
3 * (c) The GHC Team, 1998-2004
5 * Storage manager front end
7 * ---------------------------------------------------------------------------*/
9 #include "PosixSource.h"
15 #include "BlockAlloc.h"
23 #include "OSThreads.h"
25 #include "RetainerProfile.h" // for counting memory blocks (memInventory)
30 StgClosure *caf_list = NULL;
32 bdescr *small_alloc_list; /* allocate()d small 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 = NULL; /* all the generations */
41 generation *g0 = NULL; /* generation 0, for convenience */
42 generation *oldest_gen = NULL; /* oldest generation, for convenience */
43 step *g0s0 = NULL; /* generation 0, step 0, for convenience */
45 ullong 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 Mutex sm_mutex = INIT_MUTEX_VAR;
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 (generations != NULL) {
70 // multi-init protection
74 /* Sanity check to make sure the LOOKS_LIKE_ macros appear to be
75 * doing something reasonable.
77 ASSERT(LOOKS_LIKE_INFO_PTR(&stg_BLACKHOLE_info));
78 ASSERT(LOOKS_LIKE_CLOSURE_PTR(&stg_dummy_ret_closure));
79 ASSERT(!HEAP_ALLOCED(&stg_dummy_ret_closure));
81 if (RtsFlags.GcFlags.maxHeapSize != 0 &&
82 RtsFlags.GcFlags.heapSizeSuggestion >
83 RtsFlags.GcFlags.maxHeapSize) {
84 RtsFlags.GcFlags.maxHeapSize = RtsFlags.GcFlags.heapSizeSuggestion;
87 if (RtsFlags.GcFlags.maxHeapSize != 0 &&
88 RtsFlags.GcFlags.minAllocAreaSize >
89 RtsFlags.GcFlags.maxHeapSize) {
90 errorBelch("maximum heap size (-M) is smaller than minimum alloc area size (-A)");
100 /* allocate generation info array */
101 generations = (generation *)stgMallocBytes(RtsFlags.GcFlags.generations
102 * sizeof(struct _generation),
103 "initStorage: gens");
105 /* Initialise all generations */
106 for(g = 0; g < RtsFlags.GcFlags.generations; g++) {
107 gen = &generations[g];
109 gen->mut_list = END_MUT_LIST;
110 gen->mut_once_list = END_MUT_LIST;
111 gen->collections = 0;
112 gen->failed_promotions = 0;
116 /* A couple of convenience pointers */
117 g0 = &generations[0];
118 oldest_gen = &generations[RtsFlags.GcFlags.generations-1];
120 /* Allocate step structures in each generation */
121 if (RtsFlags.GcFlags.generations > 1) {
122 /* Only for multiple-generations */
124 /* Oldest generation: one step */
125 oldest_gen->n_steps = 1;
127 stgMallocBytes(1 * sizeof(struct _step), "initStorage: last step");
129 /* set up all except the oldest generation with 2 steps */
130 for(g = 0; g < RtsFlags.GcFlags.generations-1; g++) {
131 generations[g].n_steps = RtsFlags.GcFlags.steps;
132 generations[g].steps =
133 stgMallocBytes (RtsFlags.GcFlags.steps * sizeof(struct _step),
134 "initStorage: steps");
138 /* single generation, i.e. a two-space collector */
140 g0->steps = stgMallocBytes (sizeof(struct _step), "initStorage: steps");
143 /* Initialise all steps */
144 for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
145 for (s = 0; s < generations[g].n_steps; s++) {
146 stp = &generations[g].steps[s];
149 stp->n_to_blocks = 0;
151 stp->gen = &generations[g];
158 stp->large_objects = NULL;
159 stp->n_large_blocks = 0;
160 stp->new_large_objects = NULL;
161 stp->scavenged_large_objects = NULL;
162 stp->n_scavenged_large_blocks = 0;
163 stp->is_compacted = 0;
168 /* Set up the destination pointers in each younger gen. step */
169 for (g = 0; g < RtsFlags.GcFlags.generations-1; g++) {
170 for (s = 0; s < generations[g].n_steps-1; s++) {
171 generations[g].steps[s].to = &generations[g].steps[s+1];
173 generations[g].steps[s].to = &generations[g+1].steps[0];
176 /* The oldest generation has one step and it is compacted. */
177 if (RtsFlags.GcFlags.compact) {
178 if (RtsFlags.GcFlags.generations == 1) {
179 errorBelch("WARNING: compaction is incompatible with -G1; disabled");
181 oldest_gen->steps[0].is_compacted = 1;
184 oldest_gen->steps[0].to = &oldest_gen->steps[0];
186 /* generation 0 is special: that's the nursery */
187 generations[0].max_blocks = 0;
189 /* G0S0: the allocation area. Policy: keep the allocation area
190 * small to begin with, even if we have a large suggested heap
191 * size. Reason: we're going to do a major collection first, and we
192 * don't want it to be a big one. This vague idea is borne out by
193 * rigorous experimental evidence.
195 g0s0 = &generations[0].steps[0];
199 weak_ptr_list = NULL;
202 /* initialise the allocate() interface */
203 small_alloc_list = NULL;
205 alloc_blocks_lim = RtsFlags.GcFlags.minAllocAreaSize;
207 /* Tell GNU multi-precision pkg about our custom alloc functions */
208 mp_set_memory_functions(stgAllocForGMP, stgReallocForGMP, stgDeallocForGMP);
210 IF_DEBUG(gc, statDescribeGens());
216 stat_exit(calcAllocated());
219 /* -----------------------------------------------------------------------------
222 The entry code for every CAF does the following:
224 - builds a CAF_BLACKHOLE in the heap
225 - pushes an update frame pointing to the CAF_BLACKHOLE
226 - invokes UPD_CAF(), which:
227 - calls newCaf, below
228 - updates the CAF with a static indirection to the CAF_BLACKHOLE
230 Why do we build a BLACKHOLE in the heap rather than just updating
231 the thunk directly? It's so that we only need one kind of update
232 frame - otherwise we'd need a static version of the update frame too.
234 newCaf() does the following:
236 - it puts the CAF on the oldest generation's mut-once list.
237 This is so that we can treat the CAF as a root when collecting
240 For GHCI, we have additional requirements when dealing with CAFs:
242 - we must *retain* all dynamically-loaded CAFs ever entered,
243 just in case we need them again.
244 - we must be able to *revert* CAFs that have been evaluated, to
245 their pre-evaluated form.
247 To do this, we use an additional CAF list. When newCaf() is
248 called on a dynamically-loaded CAF, we add it to the CAF list
249 instead of the old-generation mutable list, and save away its
250 old info pointer (in caf->saved_info) for later reversion.
252 To revert all the CAFs, we traverse the CAF list and reset the
253 info pointer to caf->saved_info, then throw away the CAF list.
254 (see GC.c:revertCAFs()).
258 -------------------------------------------------------------------------- */
261 newCAF(StgClosure* caf)
263 /* Put this CAF on the mutable list for the old generation.
264 * This is a HACK - the IND_STATIC closure doesn't really have
265 * a mut_link field, but we pretend it has - in fact we re-use
266 * the STATIC_LINK field for the time being, because when we
267 * come to do a major GC we won't need the mut_link field
268 * any more and can use it as a STATIC_LINK.
272 ((StgIndStatic *)caf)->saved_info = NULL;
273 ((StgMutClosure *)caf)->mut_link = oldest_gen->mut_once_list;
274 oldest_gen->mut_once_list = (StgMutClosure *)caf;
279 /* If we are PAR or DIST then we never forget a CAF */
281 //debugBelch("<##> Globalising CAF %08x %s",caf,info_type(caf));
282 newGA=makeGlobal(caf,rtsTrue); /*given full weight*/
288 // An alternate version of newCaf which is used for dynamically loaded
289 // object code in GHCi. In this case we want to retain *all* CAFs in
290 // the object code, because they might be demanded at any time from an
291 // expression evaluated on the command line.
293 // The linker hackily arranges that references to newCaf from dynamic
294 // code end up pointing to newDynCAF.
296 newDynCAF(StgClosure *caf)
300 ((StgIndStatic *)caf)->saved_info = (StgInfoTable *)caf->header.info;
301 ((StgIndStatic *)caf)->static_link = caf_list;
307 /* -----------------------------------------------------------------------------
309 -------------------------------------------------------------------------- */
312 allocNurseries( void )
320 for (cap = free_capabilities; cap != NULL; cap = cap->link) {
321 cap->r.rNursery = allocNursery(NULL, RtsFlags.GcFlags.minAllocAreaSize);
322 cap->r.rCurrentNursery = cap->r.rNursery;
323 /* Set the back links to be equal to the Capability,
324 * so we can do slightly better informed locking.
326 for (bd = cap->r.rNursery; bd != NULL; bd = bd->link) {
327 bd->u.back = (bdescr *)cap;
331 g0s0->blocks = allocNursery(NULL, RtsFlags.GcFlags.minAllocAreaSize);
332 g0s0->n_blocks = RtsFlags.GcFlags.minAllocAreaSize;
333 g0s0->to_blocks = NULL;
334 g0s0->n_to_blocks = 0;
335 MainCapability.r.rNursery = g0s0->blocks;
336 MainCapability.r.rCurrentNursery = g0s0->blocks;
337 /* hp, hpLim, hp_bd, to_space etc. aren't used in G0S0 */
342 resetNurseries( void )
348 /* All tasks must be stopped */
349 ASSERT(n_free_capabilities == RtsFlags.ParFlags.nNodes);
351 for (cap = free_capabilities; cap != NULL; cap = cap->link) {
352 for (bd = cap->r.rNursery; bd; bd = bd->link) {
353 bd->free = bd->start;
354 ASSERT(bd->gen_no == 0);
355 ASSERT(bd->step == g0s0);
356 IF_DEBUG(sanity,memset(bd->start, 0xaa, BLOCK_SIZE));
358 cap->r.rCurrentNursery = cap->r.rNursery;
361 for (bd = g0s0->blocks; bd; bd = bd->link) {
362 bd->free = bd->start;
363 ASSERT(bd->gen_no == 0);
364 ASSERT(bd->step == g0s0);
365 IF_DEBUG(sanity,memset(bd->start, 0xaa, BLOCK_SIZE));
367 MainCapability.r.rNursery = g0s0->blocks;
368 MainCapability.r.rCurrentNursery = g0s0->blocks;
373 allocNursery (bdescr *tail, nat blocks)
378 // Allocate a nursery: we allocate fresh blocks one at a time and
379 // cons them on to the front of the list, not forgetting to update
380 // the back pointer on the tail of the list to point to the new block.
381 for (i=0; i < blocks; i++) {
384 processNursery() in LdvProfile.c assumes that every block group in
385 the nursery contains only a single block. So, if a block group is
386 given multiple blocks, change processNursery() accordingly.
390 // double-link the nursery: we might need to insert blocks
397 bd->free = bd->start;
405 resizeNursery ( nat blocks )
411 barf("resizeNursery: can't resize in SMP mode");
414 nursery_blocks = g0s0->n_blocks;
415 if (nursery_blocks == blocks) {
419 else if (nursery_blocks < blocks) {
420 IF_DEBUG(gc, debugBelch("Increasing size of nursery to %d blocks\n",
422 g0s0->blocks = allocNursery(g0s0->blocks, blocks-nursery_blocks);
428 IF_DEBUG(gc, debugBelch("Decreasing size of nursery to %d blocks\n",
432 while (nursery_blocks > blocks) {
434 next_bd->u.back = NULL;
435 nursery_blocks -= bd->blocks; // might be a large block
440 // might have gone just under, by freeing a large block, so make
441 // up the difference.
442 if (nursery_blocks < blocks) {
443 g0s0->blocks = allocNursery(g0s0->blocks, blocks-nursery_blocks);
447 g0s0->n_blocks = blocks;
448 ASSERT(countBlocks(g0s0->blocks) == g0s0->n_blocks);
451 /* -----------------------------------------------------------------------------
452 The allocate() interface
454 allocate(n) always succeeds, and returns a chunk of memory n words
455 long. n can be larger than the size of a block if necessary, in
456 which case a contiguous block group will be allocated.
457 -------------------------------------------------------------------------- */
467 TICK_ALLOC_HEAP_NOCTR(n);
470 /* big allocation (>LARGE_OBJECT_THRESHOLD) */
471 /* ToDo: allocate directly into generation 1 */
472 if (n >= LARGE_OBJECT_THRESHOLD/sizeof(W_)) {
473 nat req_blocks = (lnat)BLOCK_ROUND_UP(n*sizeof(W_)) / BLOCK_SIZE;
474 bd = allocGroup(req_blocks);
475 dbl_link_onto(bd, &g0s0->large_objects);
476 g0s0->n_large_blocks += req_blocks;
479 bd->flags = BF_LARGE;
480 bd->free = bd->start + n;
481 alloc_blocks += req_blocks;
485 /* small allocation (<LARGE_OBJECT_THRESHOLD) */
486 } else if (small_alloc_list == NULL || alloc_Hp + n > alloc_HpLim) {
487 if (small_alloc_list) {
488 small_alloc_list->free = alloc_Hp;
491 bd->link = small_alloc_list;
492 small_alloc_list = bd;
496 alloc_Hp = bd->start;
497 alloc_HpLim = bd->start + BLOCK_SIZE_W;
508 allocated_bytes( void )
512 allocated = alloc_blocks * BLOCK_SIZE_W - (alloc_HpLim - alloc_Hp);
513 if (pinned_object_block != NULL) {
514 allocated -= (pinned_object_block->start + BLOCK_SIZE_W) -
515 pinned_object_block->free;
522 tidyAllocateLists (void)
524 if (small_alloc_list != NULL) {
525 ASSERT(alloc_Hp >= small_alloc_list->start &&
526 alloc_Hp <= small_alloc_list->start + BLOCK_SIZE);
527 small_alloc_list->free = alloc_Hp;
531 /* ---------------------------------------------------------------------------
532 Allocate a fixed/pinned object.
534 We allocate small pinned objects into a single block, allocating a
535 new block when the current one overflows. The block is chained
536 onto the large_object_list of generation 0 step 0.
538 NOTE: The GC can't in general handle pinned objects. This
539 interface is only safe to use for ByteArrays, which have no
540 pointers and don't require scavenging. It works because the
541 block's descriptor has the BF_LARGE flag set, so the block is
542 treated as a large object and chained onto various lists, rather
543 than the individual objects being copied. However, when it comes
544 to scavenge the block, the GC will only scavenge the first object.
545 The reason is that the GC can't linearly scan a block of pinned
546 objects at the moment (doing so would require using the
547 mostly-copying techniques). But since we're restricting ourselves
548 to pinned ByteArrays, not scavenging is ok.
550 This function is called by newPinnedByteArray# which immediately
551 fills the allocated memory with a MutableByteArray#.
552 ------------------------------------------------------------------------- */
555 allocatePinned( nat n )
558 bdescr *bd = pinned_object_block;
560 // If the request is for a large object, then allocate()
561 // will give us a pinned object anyway.
562 if (n >= LARGE_OBJECT_THRESHOLD/sizeof(W_)) {
568 TICK_ALLOC_HEAP_NOCTR(n);
571 // we always return 8-byte aligned memory. bd->free must be
572 // 8-byte aligned to begin with, so we just round up n to
573 // the nearest multiple of 8 bytes.
574 if (sizeof(StgWord) == 4) {
578 // If we don't have a block of pinned objects yet, or the current
579 // one isn't large enough to hold the new object, allocate a new one.
580 if (bd == NULL || (bd->free + n) > (bd->start + BLOCK_SIZE_W)) {
581 pinned_object_block = bd = allocBlock();
582 dbl_link_onto(bd, &g0s0->large_objects);
585 bd->flags = BF_PINNED | BF_LARGE;
586 bd->free = bd->start;
596 /* -----------------------------------------------------------------------------
597 Allocation functions for GMP.
599 These all use the allocate() interface - we can't have any garbage
600 collection going on during a gmp operation, so we use allocate()
601 which always succeeds. The gmp operations which might need to
602 allocate will ask the storage manager (via doYouWantToGC()) whether
603 a garbage collection is required, in case we get into a loop doing
604 only allocate() style allocation.
605 -------------------------------------------------------------------------- */
608 stgAllocForGMP (size_t size_in_bytes)
611 nat data_size_in_words, total_size_in_words;
613 /* round up to a whole number of words */
614 data_size_in_words = (size_in_bytes + sizeof(W_) + 1) / sizeof(W_);
615 total_size_in_words = sizeofW(StgArrWords) + data_size_in_words;
617 /* allocate and fill it in. */
618 arr = (StgArrWords *)allocate(total_size_in_words);
619 SET_ARR_HDR(arr, &stg_ARR_WORDS_info, CCCS, data_size_in_words);
621 /* and return a ptr to the goods inside the array */
626 stgReallocForGMP (void *ptr, size_t old_size, size_t new_size)
628 void *new_stuff_ptr = stgAllocForGMP(new_size);
630 char *p = (char *) ptr;
631 char *q = (char *) new_stuff_ptr;
633 for (; i < old_size; i++, p++, q++) {
637 return(new_stuff_ptr);
641 stgDeallocForGMP (void *ptr STG_UNUSED,
642 size_t size STG_UNUSED)
644 /* easy for us: the garbage collector does the dealloc'n */
647 /* -----------------------------------------------------------------------------
649 * -------------------------------------------------------------------------- */
651 /* -----------------------------------------------------------------------------
654 * Approximate how much we've allocated: number of blocks in the
655 * nursery + blocks allocated via allocate() - unused nusery blocks.
656 * This leaves a little slop at the end of each block, and doesn't
657 * take into account large objects (ToDo).
658 * -------------------------------------------------------------------------- */
661 calcAllocated( void )
669 /* All tasks must be stopped. Can't assert that all the
670 capabilities are owned by the scheduler, though: one or more
671 tasks might have been stopped while they were running (non-main)
673 /* ASSERT(n_free_capabilities == RtsFlags.ParFlags.nNodes); */
676 n_free_capabilities * RtsFlags.GcFlags.minAllocAreaSize * BLOCK_SIZE_W
679 for (cap = free_capabilities; cap != NULL; cap = cap->link) {
680 for ( bd = cap->r.rCurrentNursery->link; bd != NULL; bd = bd->link ) {
681 allocated -= BLOCK_SIZE_W;
683 if (cap->r.rCurrentNursery->free < cap->r.rCurrentNursery->start
685 allocated -= (cap->r.rCurrentNursery->start + BLOCK_SIZE_W)
686 - cap->r.rCurrentNursery->free;
691 bdescr *current_nursery = MainCapability.r.rCurrentNursery;
693 allocated = (g0s0->n_blocks * BLOCK_SIZE_W) + allocated_bytes();
694 for ( bd = current_nursery->link; bd != NULL; bd = bd->link ) {
695 allocated -= BLOCK_SIZE_W;
697 if (current_nursery->free < current_nursery->start + BLOCK_SIZE_W) {
698 allocated -= (current_nursery->start + BLOCK_SIZE_W)
699 - current_nursery->free;
703 total_allocated += allocated;
707 /* Approximate the amount of live data in the heap. To be called just
708 * after garbage collection (see GarbageCollect()).
717 if (RtsFlags.GcFlags.generations == 1) {
718 live = (g0s0->n_to_blocks - 1) * BLOCK_SIZE_W +
719 ((lnat)g0s0->hp_bd->free - (lnat)g0s0->hp_bd->start) / sizeof(W_);
723 for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
724 for (s = 0; s < generations[g].n_steps; s++) {
725 /* approximate amount of live data (doesn't take into account slop
726 * at end of each block).
728 if (g == 0 && s == 0) {
731 stp = &generations[g].steps[s];
732 live += (stp->n_large_blocks + stp->n_blocks - 1) * BLOCK_SIZE_W;
733 if (stp->hp_bd != NULL) {
734 live += ((lnat)stp->hp_bd->free - (lnat)stp->hp_bd->start)
742 /* Approximate the number of blocks that will be needed at the next
743 * garbage collection.
745 * Assume: all data currently live will remain live. Steps that will
746 * be collected next time will therefore need twice as many blocks
747 * since all the data will be copied.
756 for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
757 for (s = 0; s < generations[g].n_steps; s++) {
758 if (g == 0 && s == 0) { continue; }
759 stp = &generations[g].steps[s];
760 if (generations[g].steps[0].n_blocks +
761 generations[g].steps[0].n_large_blocks
762 > generations[g].max_blocks
763 && stp->is_compacted == 0) {
764 needed += 2 * stp->n_blocks;
766 needed += stp->n_blocks;
773 /* -----------------------------------------------------------------------------
776 memInventory() checks for memory leaks by counting up all the
777 blocks we know about and comparing that to the number of blocks
778 allegedly floating around in the system.
779 -------------------------------------------------------------------------- */
789 lnat total_blocks = 0, free_blocks = 0;
791 /* count the blocks we current have */
793 for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
794 for (s = 0; s < generations[g].n_steps; s++) {
795 stp = &generations[g].steps[s];
796 total_blocks += stp->n_blocks;
797 if (RtsFlags.GcFlags.generations == 1) {
798 /* two-space collector has a to-space too :-) */
799 total_blocks += g0s0->n_to_blocks;
801 for (bd = stp->large_objects; bd; bd = bd->link) {
802 total_blocks += bd->blocks;
803 /* hack for megablock groups: they have an extra block or two in
804 the second and subsequent megablocks where the block
805 descriptors would normally go.
807 if (bd->blocks > BLOCKS_PER_MBLOCK) {
808 total_blocks -= (MBLOCK_SIZE / BLOCK_SIZE - BLOCKS_PER_MBLOCK)
809 * (bd->blocks/(MBLOCK_SIZE/BLOCK_SIZE));
815 /* any blocks held by allocate() */
816 for (bd = small_alloc_list; bd; bd = bd->link) {
817 total_blocks += bd->blocks;
821 if (RtsFlags.ProfFlags.doHeapProfile == HEAP_BY_RETAINER) {
822 total_blocks += retainerStackBlocks();
826 // count the blocks allocated by the arena allocator
827 total_blocks += arenaBlocks();
829 /* count the blocks on the free list */
830 free_blocks = countFreeList();
832 if (total_blocks + free_blocks != mblocks_allocated *
834 debugBelch("Blocks: %ld live + %ld free = %ld total (%ld around)\n",
835 total_blocks, free_blocks, total_blocks + free_blocks,
836 mblocks_allocated * BLOCKS_PER_MBLOCK);
839 ASSERT(total_blocks + free_blocks == mblocks_allocated * BLOCKS_PER_MBLOCK);
844 countBlocks(bdescr *bd)
847 for (n=0; bd != NULL; bd=bd->link) {
853 /* Full heap sanity check. */
859 if (RtsFlags.GcFlags.generations == 1) {
860 checkHeap(g0s0->to_blocks);
861 checkChain(g0s0->large_objects);
864 for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
865 for (s = 0; s < generations[g].n_steps; s++) {
866 ASSERT(countBlocks(generations[g].steps[s].blocks)
867 == generations[g].steps[s].n_blocks);
868 ASSERT(countBlocks(generations[g].steps[s].large_objects)
869 == generations[g].steps[s].n_large_blocks);
870 if (g == 0 && s == 0) { continue; }
871 checkHeap(generations[g].steps[s].blocks);
872 checkChain(generations[g].steps[s].large_objects);
874 checkMutableList(generations[g].mut_list, g);
875 checkMutOnceList(generations[g].mut_once_list, g);
879 checkFreeListSanity();
883 // handy function for use in gdb, because Bdescr() is inlined.
884 extern bdescr *_bdescr( StgPtr p );