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"
19 #include "BlockAlloc.h"
29 // Turn off inlining when debugging - it obfuscates things
32 # define STATIC_INLINE static
35 /* ----------------------------------------------------------------------------
36 Threading / unthreading pointers.
38 The basic idea here is to chain together all the fields pointing at
39 a particular object, with the root of the chain in the object's
40 info table field. The original contents of the info pointer goes
41 at the end of the chain.
43 Adding a new field to the chain is a matter of swapping the
44 contents of the field with the contents of the object's info table
47 To unthread the chain, we walk down it updating all the fields on
48 the chain with the new location of the object. We stop when we
49 reach the info pointer at the end.
51 The main difficulty here is that we need to be able to identify the
52 info pointer at the end of the chain. We can't use the low bits of
53 the pointer for this; they are already being used for
54 pointer-tagging. What's more, we need to retain the
55 pointer-tagging tag bits on each pointer during the
56 threading/unthreading process.
58 Our solution is as follows:
59 - an info pointer (chain length zero) is identified by having tag 0
60 - in a threaded chain of length > 0:
61 - the pointer-tagging tag bits are attached to the info pointer
62 - the first entry in the chain has tag 1
63 - second and subsequent entries in the chain have tag 2
65 This exploits the fact that the tag on each pointer to a given
66 closure is normally the same (if they are not the same, then
67 presumably the tag is not essential and it therefore doesn't matter
68 if we throw away some of the tags).
69 ------------------------------------------------------------------------- */
72 thread (StgClosure **p)
80 q = (StgPtr)UNTAG_CLOSURE(q0);
82 // It doesn't look like a closure at the moment, because the info
83 // ptr is possibly threaded:
84 // ASSERT(LOOKS_LIKE_CLOSURE_PTR(q));
86 if (HEAP_ALLOCED(q)) {
89 if (bd->flags & BF_MARKED)
92 switch (GET_CLOSURE_TAG((StgClosure *)iptr))
95 // this is the info pointer; we are creating a new chain.
96 // save the original tag at the end of the chain.
97 *p = (StgClosure *)((StgWord)iptr + GET_CLOSURE_TAG(q0));
102 // this is a chain of length 1 or more
103 *p = (StgClosure *)iptr;
112 thread_root (void *user STG_UNUSED, StgClosure **p)
117 // This version of thread() takes a (void *), used to circumvent
118 // warnings from gcc about pointer punning and strict aliasing.
119 STATIC_INLINE void thread_ (void *p) { thread((StgClosure **)p); }
122 unthread( StgPtr p, StgWord free )
129 switch (GET_CLOSURE_TAG((StgClosure *)q))
132 // nothing to do; the chain is length zero
136 r = *q0; // r is the info ptr, tagged with the pointer-tag
138 *p = (StgWord)UNTAG_CLOSURE((StgClosure *)r);
151 // Traverse a threaded chain and pull out the info pointer at the end.
152 // The info pointer is also tagged with the appropriate pointer tag
153 // for this closure, which should be attached to the pointer
154 // subsequently passed to unthread().
155 STATIC_INLINE StgWord
156 get_threaded_info( StgPtr p )
160 q = (W_)GET_INFO(UNTAG_CLOSURE((StgClosure *)p));
163 switch (GET_CLOSURE_TAG((StgClosure *)q))
166 ASSERT(LOOKS_LIKE_INFO_PTR(q));
170 StgWord r = *(StgPtr)(q-1);
171 ASSERT(LOOKS_LIKE_INFO_PTR((StgWord)UNTAG_CLOSURE((StgClosure *)r)));
178 barf("get_threaded_info");
182 // A word-aligned memmove will be faster for small objects than libc's or gcc's.
183 // Remember, the two regions *might* overlap, but: to <= from.
185 move(StgPtr to, StgPtr from, nat size)
187 for(; size > 0; --size) {
193 thread_static( StgClosure* p )
195 const StgInfoTable *info;
197 // keep going until we've threaded all the objects on the linked
199 while (p != END_OF_STATIC_LIST) {
202 switch (info->type) {
205 thread(&((StgInd *)p)->indirectee);
206 p = *IND_STATIC_LINK(p);
210 p = *THUNK_STATIC_LINK(p);
213 p = *FUN_STATIC_LINK(p);
216 p = *STATIC_LINK(info,p);
220 barf("thread_static: strange closure %d", (int)(info->type));
227 thread_large_bitmap( StgPtr p, StgLargeBitmap *large_bitmap, nat size )
233 bitmap = large_bitmap->bitmap[b];
234 for (i = 0; i < size; ) {
235 if ((bitmap & 1) == 0) {
236 thread((StgClosure **)p);
240 if (i % BITS_IN(W_) == 0) {
242 bitmap = large_bitmap->bitmap[b];
244 bitmap = bitmap >> 1;
250 thread_arg_block (StgFunInfoTable *fun_info, StgClosure **args)
257 switch (fun_info->f.fun_type) {
259 bitmap = BITMAP_BITS(fun_info->f.b.bitmap);
260 size = BITMAP_SIZE(fun_info->f.b.bitmap);
263 size = GET_FUN_LARGE_BITMAP(fun_info)->size;
264 thread_large_bitmap(p, GET_FUN_LARGE_BITMAP(fun_info), size);
268 bitmap = BITMAP_BITS(stg_arg_bitmaps[fun_info->f.fun_type]);
269 size = BITMAP_SIZE(stg_arg_bitmaps[fun_info->f.fun_type]);
272 if ((bitmap & 1) == 0) {
273 thread((StgClosure **)p);
276 bitmap = bitmap >> 1;
285 thread_stack(StgPtr p, StgPtr stack_end)
287 const StgRetInfoTable* info;
291 // highly similar to scavenge_stack, but we do pointer threading here.
293 while (p < stack_end) {
295 // *p must be the info pointer of an activation
296 // record. All activation records have 'bitmap' style layout
299 info = get_ret_itbl((StgClosure *)p);
301 switch (info->i.type) {
303 // Dynamic bitmap: the mask is stored on the stack
307 dyn = ((StgRetDyn *)p)->liveness;
309 // traverse the bitmap first
310 bitmap = RET_DYN_LIVENESS(dyn);
311 p = (P_)&((StgRetDyn *)p)->payload[0];
312 size = RET_DYN_BITMAP_SIZE;
314 if ((bitmap & 1) == 0) {
315 thread((StgClosure **)p);
318 bitmap = bitmap >> 1;
322 // skip over the non-ptr words
323 p += RET_DYN_NONPTRS(dyn) + RET_DYN_NONPTR_REGS_SIZE;
325 // follow the ptr words
326 for (size = RET_DYN_PTRS(dyn); size > 0; size--) {
327 thread((StgClosure **)p);
333 // small bitmap (<= 32 entries, or 64 on a 64-bit machine)
334 case CATCH_RETRY_FRAME:
335 case CATCH_STM_FRAME:
336 case ATOMICALLY_FRAME:
341 bitmap = BITMAP_BITS(info->i.layout.bitmap);
342 size = BITMAP_SIZE(info->i.layout.bitmap);
344 // NOTE: the payload starts immediately after the info-ptr, we
345 // don't have an StgHeader in the same sense as a heap closure.
347 if ((bitmap & 1) == 0) {
348 thread((StgClosure **)p);
351 bitmap = bitmap >> 1;
362 thread((StgClosure **)p);
364 size = BCO_BITMAP_SIZE(bco);
365 thread_large_bitmap(p, BCO_BITMAP(bco), size);
370 // large bitmap (> 32 entries, or 64 on a 64-bit machine)
373 size = GET_LARGE_BITMAP(&info->i)->size;
374 thread_large_bitmap(p, GET_LARGE_BITMAP(&info->i), size);
380 StgRetFun *ret_fun = (StgRetFun *)p;
381 StgFunInfoTable *fun_info;
383 fun_info = FUN_INFO_PTR_TO_STRUCT(UNTAG_CLOSURE((StgClosure *)
384 get_threaded_info((StgPtr)ret_fun->fun)));
385 // *before* threading it!
386 thread(&ret_fun->fun);
387 p = thread_arg_block(fun_info, ret_fun->payload);
392 barf("thread_stack: weird activation record found on stack: %d",
393 (int)(info->i.type));
399 thread_PAP_payload (StgClosure *fun, StgClosure **payload, StgWord size)
403 StgFunInfoTable *fun_info;
405 fun_info = FUN_INFO_PTR_TO_STRUCT(UNTAG_CLOSURE((StgClosure *)
406 get_threaded_info((StgPtr)fun)));
407 ASSERT(fun_info->i.type != PAP);
411 switch (fun_info->f.fun_type) {
413 bitmap = BITMAP_BITS(fun_info->f.b.bitmap);
416 thread_large_bitmap(p, GET_FUN_LARGE_BITMAP(fun_info), size);
420 thread_large_bitmap((StgPtr)payload, BCO_BITMAP(fun), size);
424 bitmap = BITMAP_BITS(stg_arg_bitmaps[fun_info->f.fun_type]);
427 if ((bitmap & 1) == 0) {
428 thread((StgClosure **)p);
431 bitmap = bitmap >> 1;
441 thread_PAP (StgPAP *pap)
444 p = thread_PAP_payload(pap->fun, pap->payload, pap->n_args);
450 thread_AP (StgAP *ap)
453 p = thread_PAP_payload(ap->fun, ap->payload, ap->n_args);
459 thread_AP_STACK (StgAP_STACK *ap)
462 thread_stack((P_)ap->payload, (P_)ap->payload + ap->size);
463 return (P_)ap + sizeofW(StgAP_STACK) + ap->size;
467 thread_TSO (StgTSO *tso)
469 thread_(&tso->_link);
470 thread_(&tso->global_link);
472 if ( tso->why_blocked == BlockedOnMVar
473 || tso->why_blocked == BlockedOnBlackHole
474 || tso->why_blocked == BlockedOnMsgThrowTo
475 || tso->why_blocked == BlockedOnMsgWakeup
477 thread_(&tso->block_info.closure);
479 thread_(&tso->blocked_exceptions);
483 thread_stack(tso->sp, &(tso->stack[tso->stack_size]));
484 return (StgPtr)tso + tso_sizeW(tso);
489 update_fwd_large( bdescr *bd )
492 const StgInfoTable* info;
494 for (; bd != NULL; bd = bd->link) {
496 // nothing to do in a pinned block; it might not even have an object
498 if (bd->flags & BF_PINNED) continue;
501 info = get_itbl((StgClosure *)p);
503 switch (info->type) {
509 case MUT_ARR_PTRS_CLEAN:
510 case MUT_ARR_PTRS_DIRTY:
511 case MUT_ARR_PTRS_FROZEN:
512 case MUT_ARR_PTRS_FROZEN0:
517 a = (StgMutArrPtrs*)p;
518 for (p = (P_)a->payload; p < (P_)&a->payload[a->ptrs]; p++) {
519 thread((StgClosure **)p);
525 thread_TSO((StgTSO *)p);
529 thread_AP_STACK((StgAP_STACK *)p);
533 thread_PAP((StgPAP *)p);
539 StgTRecChunk *tc = (StgTRecChunk *)p;
540 TRecEntry *e = &(tc -> entries[0]);
541 thread_(&tc->prev_chunk);
542 for (i = 0; i < tc -> next_entry_idx; i ++, e++ ) {
544 thread(&e->expected_value);
545 thread(&e->new_value);
551 barf("update_fwd_large: unknown/strange object %d", (int)(info->type));
556 // ToDo: too big to inline
557 static /* STATIC_INLINE */ StgPtr
558 thread_obj (StgInfoTable *info, StgPtr p)
560 switch (info->type) {
562 return p + sizeofW(StgThunk) + 1;
566 return p + sizeofW(StgHeader) + 1;
570 thread(&((StgClosure *)p)->payload[0]);
571 return p + sizeofW(StgHeader) + 1;
574 thread(&((StgThunk *)p)->payload[0]);
575 return p + sizeofW(StgThunk) + 1;
578 return p + sizeofW(StgThunk) + 2;
582 return p + sizeofW(StgHeader) + 2;
585 thread(&((StgThunk *)p)->payload[0]);
586 return p + sizeofW(StgThunk) + 2;
590 thread(&((StgClosure *)p)->payload[0]);
591 return p + sizeofW(StgHeader) + 2;
594 thread(&((StgThunk *)p)->payload[0]);
595 thread(&((StgThunk *)p)->payload[1]);
596 return p + sizeofW(StgThunk) + 2;
600 thread(&((StgClosure *)p)->payload[0]);
601 thread(&((StgClosure *)p)->payload[1]);
602 return p + sizeofW(StgHeader) + 2;
605 StgBCO *bco = (StgBCO *)p;
606 thread_(&bco->instrs);
607 thread_(&bco->literals);
609 return p + bco_sizeW(bco);
616 end = (P_)((StgThunk *)p)->payload +
617 info->layout.payload.ptrs;
618 for (p = (P_)((StgThunk *)p)->payload; p < end; p++) {
619 thread((StgClosure **)p);
621 return p + info->layout.payload.nptrs;
636 end = (P_)((StgClosure *)p)->payload +
637 info->layout.payload.ptrs;
638 for (p = (P_)((StgClosure *)p)->payload; p < end; p++) {
639 thread((StgClosure **)p);
641 return p + info->layout.payload.nptrs;
646 StgWeak *w = (StgWeak *)p;
647 thread(&w->cfinalizer);
650 thread(&w->finalizer);
651 if (w->link != NULL) {
654 return p + sizeofW(StgWeak);
660 StgMVar *mvar = (StgMVar *)p;
661 thread_(&mvar->head);
662 thread_(&mvar->tail);
663 thread(&mvar->value);
664 return p + sizeofW(StgMVar);
668 case IND_OLDGEN_PERM:
669 thread(&((StgInd *)p)->indirectee);
670 return p + sizeofW(StgInd);
674 StgSelector *s = (StgSelector *)p;
675 thread(&s->selectee);
676 return p + THUNK_SELECTOR_sizeW();
680 return thread_AP_STACK((StgAP_STACK *)p);
683 return thread_PAP((StgPAP *)p);
686 return thread_AP((StgAP *)p);
689 return p + arr_words_sizeW((StgArrWords *)p);
691 case MUT_ARR_PTRS_CLEAN:
692 case MUT_ARR_PTRS_DIRTY:
693 case MUT_ARR_PTRS_FROZEN:
694 case MUT_ARR_PTRS_FROZEN0:
699 a = (StgMutArrPtrs *)p;
700 for (p = (P_)a->payload; p < (P_)&a->payload[a->ptrs]; p++) {
701 thread((StgClosure **)p);
704 return (StgPtr)a + mut_arr_ptrs_sizeW(a);
708 return thread_TSO((StgTSO *)p);
713 StgTRecChunk *tc = (StgTRecChunk *)p;
714 TRecEntry *e = &(tc -> entries[0]);
715 thread_(&tc->prev_chunk);
716 for (i = 0; i < tc -> next_entry_idx; i ++, e++ ) {
718 thread(&e->expected_value);
719 thread(&e->new_value);
721 return p + sizeofW(StgTRecChunk);
725 barf("update_fwd: unknown/strange object %d", (int)(info->type));
731 update_fwd( bdescr *blocks )
739 // cycle through all the blocks in the step
740 for (; bd != NULL; bd = bd->link) {
743 // linearly scan the objects in this block
744 while (p < bd->free) {
745 ASSERT(LOOKS_LIKE_CLOSURE_PTR(p));
746 info = get_itbl((StgClosure *)p);
747 p = thread_obj(info, p);
753 update_fwd_compact( bdescr *blocks )
759 bdescr *bd, *free_bd;
766 free = free_bd->start;
768 // cycle through all the blocks in the step
769 for (; bd != NULL; bd = bd->link) {
772 while (p < bd->free ) {
774 while ( p < bd->free && !is_marked(p,bd) ) {
783 m = * ((StgPtr)bd->u.bitmap + ((p - bd->start) / (BITS_IN(StgWord))));
784 m >>= ((p - bd->start) & (BITS_IN(StgWord) - 1));
786 while ( p < bd->free ) {
791 if (((StgWord)p & (sizeof(W_) * BITS_IN(StgWord))) == 0) {
799 // Problem: we need to know the destination for this cell
800 // in order to unthread its info pointer. But we can't
801 // know the destination without the size, because we may
802 // spill into the next block. So we have to run down the
803 // threaded list and get the info ptr first.
805 // ToDo: one possible avenue of attack is to use the fact
806 // that if (p&BLOCK_MASK) >= (free&BLOCK_MASK), then we
807 // definitely have enough room. Also see bug #1147.
808 iptr = get_threaded_info(p);
809 info = INFO_PTR_TO_STRUCT(UNTAG_CLOSURE((StgClosure *)iptr));
813 p = thread_obj(info, p);
816 if (free + size > free_bd->start + BLOCK_SIZE_W) {
817 // set the next bit in the bitmap to indicate that
818 // this object needs to be pushed into the next
819 // block. This saves us having to run down the
820 // threaded info pointer list twice during the next pass.
822 free_bd = free_bd->link;
823 free = free_bd->start;
825 ASSERT(!is_marked(q+1,bd));
828 unthread(q,(StgWord)free + GET_CLOSURE_TAG((StgClosure *)iptr));
838 update_bkwd_compact( generation *gen )
844 bdescr *bd, *free_bd;
846 nat size, free_blocks;
849 bd = free_bd = gen->old_blocks;
850 free = free_bd->start;
853 // cycle through all the blocks in the step
854 for (; bd != NULL; bd = bd->link) {
857 while (p < bd->free ) {
859 while ( p < bd->free && !is_marked(p,bd) ) {
868 m = * ((StgPtr)bd->u.bitmap + ((p - bd->start) / (BITS_IN(StgWord))));
869 m >>= ((p - bd->start) & (BITS_IN(StgWord) - 1));
871 while ( p < bd->free ) {
876 if (((StgWord)p & (sizeof(W_) * BITS_IN(StgWord))) == 0) {
884 if (is_marked(p+1,bd)) {
885 // don't forget to update the free ptr in the block desc.
886 free_bd->free = free;
887 free_bd = free_bd->link;
888 free = free_bd->start;
892 iptr = get_threaded_info(p);
893 unthread(p, (StgWord)free + GET_CLOSURE_TAG((StgClosure *)iptr));
894 ASSERT(LOOKS_LIKE_INFO_PTR((StgWord)((StgClosure *)p)->header.info));
895 info = get_itbl((StgClosure *)p);
896 size = closure_sizeW_((StgClosure *)p,info);
903 if (info->type == TSO) {
904 move_TSO((StgTSO *)p, (StgTSO *)free);
915 // free the remaining blocks and count what's left.
916 free_bd->free = free;
917 if (free_bd->link != NULL) {
918 freeChain(free_bd->link);
919 free_bd->link = NULL;
926 compact(StgClosure *static_objects)
931 // 1. thread the roots
932 markCapabilities((evac_fn)thread_root, NULL);
934 // the weak pointer lists...
935 if (weak_ptr_list != NULL) {
936 thread((void *)&weak_ptr_list);
938 if (old_weak_ptr_list != NULL) {
939 thread((void *)&old_weak_ptr_list); // tmp
943 for (g = 1; g < RtsFlags.GcFlags.generations; g++) {
947 for (bd = generations[g].mut_list; bd != NULL; bd = bd->link) {
948 for (p = bd->start; p < bd->free; p++) {
949 thread((StgClosure **)p);
952 for (n = 0; n < n_capabilities; n++) {
953 for (bd = capabilities[n].mut_lists[g];
954 bd != NULL; bd = bd->link) {
955 for (p = bd->start; p < bd->free; p++) {
956 thread((StgClosure **)p);
962 // the global thread list
963 for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
964 thread((void *)&generations[g].threads);
967 // any threads resurrected during this GC
968 thread((void *)&resurrected_threads);
970 // the blackhole queue
971 thread((void *)&blackhole_queue);
977 for (task = all_tasks; task != NULL; task = task->all_link) {
978 for (incall = task->incall; incall != NULL;
979 incall = incall->prev_stack) {
981 thread_(&incall->tso);
987 // the static objects
988 thread_static(static_objects /* ToDo: ok? */);
990 // the stable pointer table
991 threadStablePtrTable((evac_fn)thread_root, NULL);
993 // the CAF list (used by GHCi)
994 markCAFs((evac_fn)thread_root, NULL);
996 // 2. update forward ptrs
997 for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
998 gen = &generations[g];
999 debugTrace(DEBUG_gc, "update_fwd: %d", g);
1001 update_fwd(gen->blocks);
1002 update_fwd_large(gen->scavenged_large_objects);
1003 if (g == RtsFlags.GcFlags.generations-1 && gen->old_blocks != NULL) {
1004 debugTrace(DEBUG_gc, "update_fwd: %d (compact)", g);
1005 update_fwd_compact(gen->old_blocks);
1009 // 3. update backward ptrs
1011 if (gen->old_blocks != NULL) {
1012 blocks = update_bkwd_compact(gen);
1013 debugTrace(DEBUG_gc,
1014 "update_bkwd: %d (compact, old: %d blocks, now %d blocks)",
1015 gen->no, gen->n_old_blocks, blocks);
1016 gen->n_old_blocks = blocks;