1 /* -----------------------------------------------------------------------------
3 * (c) The GHC Team 2001-2008
5 * Compacting garbage collector
7 * Documentation on the architecture of the Garbage Collector can be
8 * found in the online commentary:
10 * http://hackage.haskell.org/trac/ghc/wiki/Commentary/Rts/Storage/GC
12 * ---------------------------------------------------------------------------*/
14 #include "PosixSource.h"
18 #include "OSThreads.h"
19 #include "BlockAlloc.h"
27 // Turn off inlining when debugging - it obfuscates things
30 # define STATIC_INLINE static
33 /* ----------------------------------------------------------------------------
34 Threading / unthreading pointers.
36 The basic idea here is to chain together all the fields pointing at
37 a particular object, with the root of the chain in the object's
38 info table field. The original contents of the info pointer goes
39 at the end of the chain.
41 Adding a new field to the chain is a matter of swapping the
42 contents of the field with the contents of the object's info table
45 To unthread the chain, we walk down it updating all the fields on
46 the chain with the new location of the object. We stop when we
47 reach the info pointer at the end.
49 The main difficulty here is that we need to be able to identify the
50 info pointer at the end of the chain. We can't use the low bits of
51 the pointer for this; they are already being used for
52 pointer-tagging. What's more, we need to retain the
53 pointer-tagging tag bits on each pointer during the
54 threading/unthreading process.
56 Our solution is as follows:
57 - an info pointer (chain length zero) is identified by having tag 0
58 - in a threaded chain of length > 0:
59 - the pointer-tagging tag bits are attached to the info pointer
60 - the first entry in the chain has tag 1
61 - second and subsequent entries in the chain have tag 2
63 This exploits the fact that the tag on each pointer to a given
64 closure is normally the same (if they are not the same, then
65 presumably the tag is not essential and it therefore doesn't matter
66 if we throw away some of the tags).
67 ------------------------------------------------------------------------- */
70 thread (StgClosure **p)
78 q = (StgPtr)UNTAG_CLOSURE(q0);
80 // It doesn't look like a closure at the moment, because the info
81 // ptr is possibly threaded:
82 // ASSERT(LOOKS_LIKE_CLOSURE_PTR(q));
84 if (HEAP_ALLOCED(q)) {
87 if (bd->flags & BF_MARKED)
90 switch (GET_CLOSURE_TAG((StgClosure *)iptr))
93 // this is the info pointer; we are creating a new chain.
94 // save the original tag at the end of the chain.
95 *p = (StgClosure *)((StgWord)iptr + GET_CLOSURE_TAG(q0));
100 // this is a chain of length 1 or more
101 *p = (StgClosure *)iptr;
110 thread_root (void *user STG_UNUSED, StgClosure **p)
115 // This version of thread() takes a (void *), used to circumvent
116 // warnings from gcc about pointer punning and strict aliasing.
117 STATIC_INLINE void thread_ (void *p) { thread((StgClosure **)p); }
120 unthread( StgPtr p, StgWord free )
127 switch (GET_CLOSURE_TAG((StgClosure *)q))
130 // nothing to do; the chain is length zero
134 r = *q0; // r is the info ptr, tagged with the pointer-tag
136 *p = (StgWord)UNTAG_CLOSURE((StgClosure *)r);
149 // Traverse a threaded chain and pull out the info pointer at the end.
150 // The info pointer is also tagged with the appropriate pointer tag
151 // for this closure, which should be attached to the pointer
152 // subsequently passed to unthread().
153 STATIC_INLINE StgWord
154 get_threaded_info( StgPtr p )
158 q = (W_)GET_INFO(UNTAG_CLOSURE((StgClosure *)p));
161 switch (GET_CLOSURE_TAG((StgClosure *)q))
164 ASSERT(LOOKS_LIKE_INFO_PTR(q));
168 StgWord r = *(StgPtr)(q-1);
169 ASSERT(LOOKS_LIKE_INFO_PTR(UNTAG_CLOSURE((StgClosure *)r)));
176 barf("get_threaded_info");
180 // A word-aligned memmove will be faster for small objects than libc's or gcc's.
181 // Remember, the two regions *might* overlap, but: to <= from.
183 move(StgPtr to, StgPtr from, nat size)
185 for(; size > 0; --size) {
191 thread_static( StgClosure* p )
193 const StgInfoTable *info;
195 // keep going until we've threaded all the objects on the linked
197 while (p != END_OF_STATIC_LIST) {
200 switch (info->type) {
203 thread(&((StgInd *)p)->indirectee);
204 p = *IND_STATIC_LINK(p);
208 p = *THUNK_STATIC_LINK(p);
211 p = *FUN_STATIC_LINK(p);
214 p = *STATIC_LINK(info,p);
218 barf("thread_static: strange closure %d", (int)(info->type));
225 thread_large_bitmap( StgPtr p, StgLargeBitmap *large_bitmap, nat size )
231 bitmap = large_bitmap->bitmap[b];
232 for (i = 0; i < size; ) {
233 if ((bitmap & 1) == 0) {
234 thread((StgClosure **)p);
238 if (i % BITS_IN(W_) == 0) {
240 bitmap = large_bitmap->bitmap[b];
242 bitmap = bitmap >> 1;
248 thread_arg_block (StgFunInfoTable *fun_info, StgClosure **args)
255 switch (fun_info->f.fun_type) {
257 bitmap = BITMAP_BITS(fun_info->f.b.bitmap);
258 size = BITMAP_SIZE(fun_info->f.b.bitmap);
261 size = GET_FUN_LARGE_BITMAP(fun_info)->size;
262 thread_large_bitmap(p, GET_FUN_LARGE_BITMAP(fun_info), size);
266 bitmap = BITMAP_BITS(stg_arg_bitmaps[fun_info->f.fun_type]);
267 size = BITMAP_SIZE(stg_arg_bitmaps[fun_info->f.fun_type]);
270 if ((bitmap & 1) == 0) {
271 thread((StgClosure **)p);
274 bitmap = bitmap >> 1;
283 thread_stack(StgPtr p, StgPtr stack_end)
285 const StgRetInfoTable* info;
289 // highly similar to scavenge_stack, but we do pointer threading here.
291 while (p < stack_end) {
293 // *p must be the info pointer of an activation
294 // record. All activation records have 'bitmap' style layout
297 info = get_ret_itbl((StgClosure *)p);
299 switch (info->i.type) {
301 // Dynamic bitmap: the mask is stored on the stack
305 dyn = ((StgRetDyn *)p)->liveness;
307 // traverse the bitmap first
308 bitmap = RET_DYN_LIVENESS(dyn);
309 p = (P_)&((StgRetDyn *)p)->payload[0];
310 size = RET_DYN_BITMAP_SIZE;
312 if ((bitmap & 1) == 0) {
313 thread((StgClosure **)p);
316 bitmap = bitmap >> 1;
320 // skip over the non-ptr words
321 p += RET_DYN_NONPTRS(dyn) + RET_DYN_NONPTR_REGS_SIZE;
323 // follow the ptr words
324 for (size = RET_DYN_PTRS(dyn); size > 0; size--) {
325 thread((StgClosure **)p);
331 // small bitmap (<= 32 entries, or 64 on a 64-bit machine)
332 case CATCH_RETRY_FRAME:
333 case CATCH_STM_FRAME:
334 case ATOMICALLY_FRAME:
339 bitmap = BITMAP_BITS(info->i.layout.bitmap);
340 size = BITMAP_SIZE(info->i.layout.bitmap);
342 // NOTE: the payload starts immediately after the info-ptr, we
343 // don't have an StgHeader in the same sense as a heap closure.
345 if ((bitmap & 1) == 0) {
346 thread((StgClosure **)p);
349 bitmap = bitmap >> 1;
360 thread((StgClosure **)p);
362 size = BCO_BITMAP_SIZE(bco);
363 thread_large_bitmap(p, BCO_BITMAP(bco), size);
368 // large bitmap (> 32 entries, or 64 on a 64-bit machine)
371 size = GET_LARGE_BITMAP(&info->i)->size;
372 thread_large_bitmap(p, GET_LARGE_BITMAP(&info->i), size);
378 StgRetFun *ret_fun = (StgRetFun *)p;
379 StgFunInfoTable *fun_info;
381 fun_info = FUN_INFO_PTR_TO_STRUCT(UNTAG_CLOSURE((StgClosure *)
382 get_threaded_info((StgPtr)ret_fun->fun)));
383 // *before* threading it!
384 thread(&ret_fun->fun);
385 p = thread_arg_block(fun_info, ret_fun->payload);
390 barf("thread_stack: weird activation record found on stack: %d",
391 (int)(info->i.type));
397 thread_PAP_payload (StgClosure *fun, StgClosure **payload, StgWord size)
401 StgFunInfoTable *fun_info;
403 fun_info = FUN_INFO_PTR_TO_STRUCT(UNTAG_CLOSURE((StgClosure *)
404 get_threaded_info((StgPtr)fun)));
405 ASSERT(fun_info->i.type != PAP);
409 switch (fun_info->f.fun_type) {
411 bitmap = BITMAP_BITS(fun_info->f.b.bitmap);
414 thread_large_bitmap(p, GET_FUN_LARGE_BITMAP(fun_info), size);
418 thread_large_bitmap((StgPtr)payload, BCO_BITMAP(fun), size);
422 bitmap = BITMAP_BITS(stg_arg_bitmaps[fun_info->f.fun_type]);
425 if ((bitmap & 1) == 0) {
426 thread((StgClosure **)p);
429 bitmap = bitmap >> 1;
439 thread_PAP (StgPAP *pap)
442 p = thread_PAP_payload(pap->fun, pap->payload, pap->n_args);
448 thread_AP (StgAP *ap)
451 p = thread_PAP_payload(ap->fun, ap->payload, ap->n_args);
457 thread_AP_STACK (StgAP_STACK *ap)
460 thread_stack((P_)ap->payload, (P_)ap->payload + ap->size);
461 return (P_)ap + sizeofW(StgAP_STACK) + ap->size;
465 thread_TSO (StgTSO *tso)
467 thread_(&tso->_link);
468 thread_(&tso->global_link);
470 if ( tso->why_blocked == BlockedOnMVar
471 || tso->why_blocked == BlockedOnBlackHole
472 || tso->why_blocked == BlockedOnException
474 thread_(&tso->block_info.closure);
476 thread_(&tso->blocked_exceptions);
480 thread_stack(tso->sp, &(tso->stack[tso->stack_size]));
481 return (StgPtr)tso + tso_sizeW(tso);
486 update_fwd_large( bdescr *bd )
489 const StgInfoTable* info;
491 for (; bd != NULL; bd = bd->link) {
494 info = get_itbl((StgClosure *)p);
496 switch (info->type) {
502 case MUT_ARR_PTRS_CLEAN:
503 case MUT_ARR_PTRS_DIRTY:
504 case MUT_ARR_PTRS_FROZEN:
505 case MUT_ARR_PTRS_FROZEN0:
510 next = p + mut_arr_ptrs_sizeW((StgMutArrPtrs*)p);
511 for (p = (P_)((StgMutArrPtrs *)p)->payload; p < next; p++) {
512 thread((StgClosure **)p);
518 thread_TSO((StgTSO *)p);
522 thread_AP_STACK((StgAP_STACK *)p);
526 thread_PAP((StgPAP *)p);
532 StgTRecChunk *tc = (StgTRecChunk *)p;
533 TRecEntry *e = &(tc -> entries[0]);
534 thread_(&tc->prev_chunk);
535 for (i = 0; i < tc -> next_entry_idx; i ++, e++ ) {
537 thread(&e->expected_value);
538 thread(&e->new_value);
544 barf("update_fwd_large: unknown/strange object %d", (int)(info->type));
549 // ToDo: too big to inline
550 static /* STATIC_INLINE */ StgPtr
551 thread_obj (StgInfoTable *info, StgPtr p)
553 switch (info->type) {
555 return p + sizeofW(StgThunk) + 1;
559 return p + sizeofW(StgHeader) + 1;
563 thread(&((StgClosure *)p)->payload[0]);
564 return p + sizeofW(StgHeader) + 1;
567 thread(&((StgThunk *)p)->payload[0]);
568 return p + sizeofW(StgThunk) + 1;
571 return p + sizeofW(StgThunk) + 2;
575 return p + sizeofW(StgHeader) + 2;
578 thread(&((StgThunk *)p)->payload[0]);
579 return p + sizeofW(StgThunk) + 2;
583 thread(&((StgClosure *)p)->payload[0]);
584 return p + sizeofW(StgHeader) + 2;
587 thread(&((StgThunk *)p)->payload[0]);
588 thread(&((StgThunk *)p)->payload[1]);
589 return p + sizeofW(StgThunk) + 2;
593 thread(&((StgClosure *)p)->payload[0]);
594 thread(&((StgClosure *)p)->payload[1]);
595 return p + sizeofW(StgHeader) + 2;
598 StgBCO *bco = (StgBCO *)p;
599 thread_(&bco->instrs);
600 thread_(&bco->literals);
602 return p + bco_sizeW(bco);
609 end = (P_)((StgThunk *)p)->payload +
610 info->layout.payload.ptrs;
611 for (p = (P_)((StgThunk *)p)->payload; p < end; p++) {
612 thread((StgClosure **)p);
614 return p + info->layout.payload.nptrs;
624 case SE_CAF_BLACKHOLE:
630 end = (P_)((StgClosure *)p)->payload +
631 info->layout.payload.ptrs;
632 for (p = (P_)((StgClosure *)p)->payload; p < end; p++) {
633 thread((StgClosure **)p);
635 return p + info->layout.payload.nptrs;
640 StgWeak *w = (StgWeak *)p;
643 thread(&w->finalizer);
644 if (w->link != NULL) {
647 return p + sizeofW(StgWeak);
653 StgMVar *mvar = (StgMVar *)p;
654 thread_(&mvar->head);
655 thread_(&mvar->tail);
656 thread(&mvar->value);
657 return p + sizeofW(StgMVar);
661 case IND_OLDGEN_PERM:
662 thread(&((StgInd *)p)->indirectee);
663 return p + sizeofW(StgInd);
667 StgSelector *s = (StgSelector *)p;
668 thread(&s->selectee);
669 return p + THUNK_SELECTOR_sizeW();
673 return thread_AP_STACK((StgAP_STACK *)p);
676 return thread_PAP((StgPAP *)p);
679 return thread_AP((StgAP *)p);
682 return p + arr_words_sizeW((StgArrWords *)p);
684 case MUT_ARR_PTRS_CLEAN:
685 case MUT_ARR_PTRS_DIRTY:
686 case MUT_ARR_PTRS_FROZEN:
687 case MUT_ARR_PTRS_FROZEN0:
692 next = p + mut_arr_ptrs_sizeW((StgMutArrPtrs*)p);
693 for (p = (P_)((StgMutArrPtrs *)p)->payload; p < next; p++) {
694 thread((StgClosure **)p);
700 return thread_TSO((StgTSO *)p);
702 case TVAR_WATCH_QUEUE:
704 StgTVarWatchQueue *wq = (StgTVarWatchQueue *)p;
705 thread_(&wq->closure);
706 thread_(&wq->next_queue_entry);
707 thread_(&wq->prev_queue_entry);
708 return p + sizeofW(StgTVarWatchQueue);
713 StgTVar *tvar = (StgTVar *)p;
714 thread((void *)&tvar->current_value);
715 thread((void *)&tvar->first_watch_queue_entry);
716 return p + sizeofW(StgTVar);
721 StgTRecHeader *trec = (StgTRecHeader *)p;
722 thread_(&trec->enclosing_trec);
723 thread_(&trec->current_chunk);
724 thread_(&trec->invariants_to_check);
725 return p + sizeofW(StgTRecHeader);
731 StgTRecChunk *tc = (StgTRecChunk *)p;
732 TRecEntry *e = &(tc -> entries[0]);
733 thread_(&tc->prev_chunk);
734 for (i = 0; i < tc -> next_entry_idx; i ++, e++ ) {
736 thread(&e->expected_value);
737 thread(&e->new_value);
739 return p + sizeofW(StgTRecChunk);
742 case ATOMIC_INVARIANT:
744 StgAtomicInvariant *invariant = (StgAtomicInvariant *)p;
745 thread_(&invariant->code);
746 thread_(&invariant->last_execution);
747 return p + sizeofW(StgAtomicInvariant);
750 case INVARIANT_CHECK_QUEUE:
752 StgInvariantCheckQueue *queue = (StgInvariantCheckQueue *)p;
753 thread_(&queue->invariant);
754 thread_(&queue->my_execution);
755 thread_(&queue->next_queue_entry);
756 return p + sizeofW(StgInvariantCheckQueue);
760 barf("update_fwd: unknown/strange object %d", (int)(info->type));
766 update_fwd( bdescr *blocks )
774 // cycle through all the blocks in the step
775 for (; bd != NULL; bd = bd->link) {
778 // linearly scan the objects in this block
779 while (p < bd->free) {
780 ASSERT(LOOKS_LIKE_CLOSURE_PTR(p));
781 info = get_itbl((StgClosure *)p);
782 p = thread_obj(info, p);
788 update_fwd_compact( bdescr *blocks )
794 bdescr *bd, *free_bd;
801 free = free_bd->start;
803 // cycle through all the blocks in the step
804 for (; bd != NULL; bd = bd->link) {
807 while (p < bd->free ) {
809 while ( p < bd->free && !is_marked(p,bd) ) {
818 m = * ((StgPtr)bd->u.bitmap + ((p - bd->start) / (BITS_IN(StgWord))));
819 m >>= ((p - bd->start) & (BITS_IN(StgWord) - 1));
821 while ( p < bd->free ) {
826 if (((StgWord)p & (sizeof(W_) * BITS_IN(StgWord))) == 0) {
834 // Problem: we need to know the destination for this cell
835 // in order to unthread its info pointer. But we can't
836 // know the destination without the size, because we may
837 // spill into the next block. So we have to run down the
838 // threaded list and get the info ptr first.
840 // ToDo: one possible avenue of attack is to use the fact
841 // that if (p&BLOCK_MASK) >= (free&BLOCK_MASK), then we
842 // definitely have enough room. Also see bug #1147.
843 iptr = get_threaded_info(p);
844 info = INFO_PTR_TO_STRUCT(UNTAG_CLOSURE((StgClosure *)iptr));
848 p = thread_obj(info, p);
851 if (free + size > free_bd->start + BLOCK_SIZE_W) {
852 // unset the next bit in the bitmap to indicate that
853 // this object needs to be pushed into the next
854 // block. This saves us having to run down the
855 // threaded info pointer list twice during the next pass.
857 free_bd = free_bd->link;
858 free = free_bd->start;
860 ASSERT(is_marked(q+1,bd));
863 unthread(q,(StgWord)free + GET_CLOSURE_TAG((StgClosure *)iptr));
873 update_bkwd_compact( step *stp )
879 bdescr *bd, *free_bd;
881 nat size, free_blocks;
884 bd = free_bd = stp->old_blocks;
885 free = free_bd->start;
888 // cycle through all the blocks in the step
889 for (; bd != NULL; bd = bd->link) {
892 while (p < bd->free ) {
894 while ( p < bd->free && !is_marked(p,bd) ) {
903 m = * ((StgPtr)bd->u.bitmap + ((p - bd->start) / (BITS_IN(StgWord))));
904 m >>= ((p - bd->start) & (BITS_IN(StgWord) - 1));
906 while ( p < bd->free ) {
911 if (((StgWord)p & (sizeof(W_) * BITS_IN(StgWord))) == 0) {
919 if (!is_marked(p+1,bd)) {
920 // don't forget to update the free ptr in the block desc.
921 free_bd->free = free;
922 free_bd = free_bd->link;
923 free = free_bd->start;
927 iptr = get_threaded_info(p);
928 unthread(p, (StgWord)free + GET_CLOSURE_TAG((StgClosure *)iptr));
929 ASSERT(LOOKS_LIKE_INFO_PTR(((StgClosure *)p)->header.info));
930 info = get_itbl((StgClosure *)p);
931 size = closure_sizeW_((StgClosure *)p,info);
938 if (info->type == TSO) {
939 move_TSO((StgTSO *)p, (StgTSO *)free);
950 // free the remaining blocks and count what's left.
951 free_bd->free = free;
952 if (free_bd->link != NULL) {
953 freeChain(free_bd->link);
954 free_bd->link = NULL;
961 compact(StgClosure *static_objects)
966 // 1. thread the roots
967 markCapabilities((evac_fn)thread_root, NULL);
969 // the weak pointer lists...
970 if (weak_ptr_list != NULL) {
971 thread((void *)&weak_ptr_list);
973 if (old_weak_ptr_list != NULL) {
974 thread((void *)&old_weak_ptr_list); // tmp
978 for (g = 1; g < RtsFlags.GcFlags.generations; g++) {
981 for (bd = generations[g].mut_list; bd != NULL; bd = bd->link) {
982 for (p = bd->start; p < bd->free; p++) {
983 thread((StgClosure **)p);
988 // the global thread list
989 for (s = 0; s < total_steps; s++) {
990 thread((void *)&all_steps[s].threads);
993 // any threads resurrected during this GC
994 thread((void *)&resurrected_threads);
996 // the blackhole queue
997 thread((void *)&blackhole_queue);
1002 for (task = all_tasks; task != NULL; task = task->all_link) {
1004 thread_(&task->tso);
1009 // the static objects
1010 thread_static(static_objects /* ToDo: ok? */);
1012 // the stable pointer table
1013 threadStablePtrTable((evac_fn)thread_root, NULL);
1015 // the CAF list (used by GHCi)
1016 markCAFs((evac_fn)thread_root, NULL);
1018 // 2. update forward ptrs
1019 for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
1020 for (s = 0; s < generations[g].n_steps; s++) {
1021 if (g==0 && s ==0) continue;
1022 stp = &generations[g].steps[s];
1023 debugTrace(DEBUG_gc, "update_fwd: %d.%d",
1024 stp->gen->no, stp->no);
1026 update_fwd(stp->blocks);
1027 update_fwd_large(stp->scavenged_large_objects);
1028 if (g == RtsFlags.GcFlags.generations-1 && stp->old_blocks != NULL) {
1029 debugTrace(DEBUG_gc, "update_fwd: %d.%d (compact)",
1030 stp->gen->no, stp->no);
1031 update_fwd_compact(stp->old_blocks);
1036 // 3. update backward ptrs
1037 stp = &oldest_gen->steps[0];
1038 if (stp->old_blocks != NULL) {
1039 blocks = update_bkwd_compact(stp);
1040 debugTrace(DEBUG_gc,
1041 "update_bkwd: %d.%d (compact, old: %d blocks, now %d blocks)",
1042 stp->gen->no, stp->no,
1043 stp->n_old_blocks, blocks);
1044 stp->n_old_blocks = blocks;