1 /* -----------------------------------------------------------------------------
3 * (c) The GHC Team 1998-2006
5 * Generational 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 "OSThreads.h"
20 #include "LdvProfile.h"
25 #include "BlockAlloc.h"
31 #include "ParTicky.h" // ToDo: move into Rts.h
32 #include "RtsSignals.h"
36 #if defined(RTS_GTK_FRONTPANEL)
37 #include "FrontPanel.h"
40 #include "RetainerProfile.h"
41 #include "RaiseAsync.h"
53 #include <string.h> // for memset()
56 /* -----------------------------------------------------------------------------
58 -------------------------------------------------------------------------- */
60 /* STATIC OBJECT LIST.
63 * We maintain a linked list of static objects that are still live.
64 * The requirements for this list are:
66 * - we need to scan the list while adding to it, in order to
67 * scavenge all the static objects (in the same way that
68 * breadth-first scavenging works for dynamic objects).
70 * - we need to be able to tell whether an object is already on
71 * the list, to break loops.
73 * Each static object has a "static link field", which we use for
74 * linking objects on to the list. We use a stack-type list, consing
75 * objects on the front as they are added (this means that the
76 * scavenge phase is depth-first, not breadth-first, but that
79 * A separate list is kept for objects that have been scavenged
80 * already - this is so that we can zero all the marks afterwards.
82 * An object is on the list if its static link field is non-zero; this
83 * means that we have to mark the end of the list with '1', not NULL.
85 * Extra notes for generational GC:
87 * Each generation has a static object list associated with it. When
88 * collecting generations up to N, we treat the static object lists
89 * from generations > N as roots.
91 * We build up a static object list while collecting generations 0..N,
92 * which is then appended to the static object list of generation N+1.
95 /* N is the oldest generation being collected, where the generations
96 * are numbered starting at 0. A major GC (indicated by the major_gc
97 * flag) is when we're collecting all generations. We only attempt to
98 * deal with static objects and GC CAFs when doing a major GC.
103 /* Data used for allocation area sizing.
105 static lnat g0s0_pcnt_kept = 30; // percentage of g0s0 live at last minor GC
115 /* Thread-local data for each GC thread
117 gc_thread **gc_threads = NULL;
118 // gc_thread *gct = NULL; // this thread's gct TODO: make thread-local
120 // Number of threads running in *this* GC. Affects how many
121 // step->todos[] lists we have to look in to find work.
125 long copied; // *words* copied & scavenged during this GC
128 SpinLock recordMutableGen_sync;
131 /* -----------------------------------------------------------------------------
132 Static function declarations
133 -------------------------------------------------------------------------- */
135 static void mark_root (StgClosure **root);
136 static void zero_static_object_list (StgClosure* first_static);
137 static nat initialise_N (rtsBool force_major_gc);
138 static void alloc_gc_threads (void);
139 static void init_collected_gen (nat g, nat threads);
140 static void init_uncollected_gen (nat g, nat threads);
141 static void init_gc_thread (gc_thread *t);
142 static void update_task_list (void);
143 static void resize_generations (void);
144 static void resize_nursery (void);
145 static void start_gc_threads (void);
146 static void gc_thread_work (void);
147 static nat inc_running (void);
148 static nat dec_running (void);
149 static void wakeup_gc_threads (nat n_threads);
150 static void shutdown_gc_threads (nat n_threads);
152 #if 0 && defined(DEBUG)
153 static void gcCAFs (void);
156 /* -----------------------------------------------------------------------------
157 The mark bitmap & stack.
158 -------------------------------------------------------------------------- */
160 #define MARK_STACK_BLOCKS 4
162 bdescr *mark_stack_bdescr;
167 // Flag and pointers used for falling back to a linear scan when the
168 // mark stack overflows.
169 rtsBool mark_stack_overflowed;
170 bdescr *oldgen_scan_bd;
173 /* -----------------------------------------------------------------------------
174 GarbageCollect: the main entry point to the garbage collector.
176 Locks held: all capabilities are held throughout GarbageCollect().
177 -------------------------------------------------------------------------- */
180 GarbageCollect ( rtsBool force_major_gc )
184 lnat live, allocated;
185 lnat oldgen_saved_blocks = 0;
186 gc_thread *saved_gct;
189 // necessary if we stole a callee-saves register for gct:
193 CostCentreStack *prev_CCS;
198 #if defined(RTS_USER_SIGNALS)
199 if (RtsFlags.MiscFlags.install_signal_handlers) {
205 // tell the stats department that we've started a GC
208 // tell the STM to discard any cached closures it's hoping to re-use
217 // attribute any costs to CCS_GC
223 /* Approximate how much we allocated.
224 * Todo: only when generating stats?
226 allocated = calcAllocated();
228 /* Figure out which generation to collect
230 n = initialise_N(force_major_gc);
232 /* Allocate + initialise the gc_thread structures.
236 /* Start threads, so they can be spinning up while we finish initialisation.
240 /* How many threads will be participating in this GC?
241 * We don't try to parallelise minor GC.
243 #if defined(THREADED_RTS)
244 if (n < (4*1024*1024 / BLOCK_SIZE)) {
247 n_gc_threads = RtsFlags.ParFlags.gcThreads;
252 trace(TRACE_gc|DEBUG_gc, "GC (gen %d): %dKB to collect, using %d thread(s)",
253 N, n * (BLOCK_SIZE / 1024), n_gc_threads);
255 #ifdef RTS_GTK_FRONTPANEL
256 if (RtsFlags.GcFlags.frontpanel) {
257 updateFrontPanelBeforeGC(N);
262 // check for memory leaks if DEBUG is on
263 memInventory(traceClass(DEBUG_gc));
266 // check stack sanity *before* GC (ToDo: check all threads)
267 IF_DEBUG(sanity, checkFreeListSanity());
269 // Initialise all our gc_thread structures
270 for (t = 0; t < n_gc_threads; t++) {
271 init_gc_thread(gc_threads[t]);
274 // Initialise all the generations/steps that we're collecting.
275 for (g = 0; g <= N; g++) {
276 init_collected_gen(g,n_gc_threads);
279 // Initialise all the generations/steps that we're *not* collecting.
280 for (g = N+1; g < RtsFlags.GcFlags.generations; g++) {
281 init_uncollected_gen(g,n_gc_threads);
284 /* Allocate a mark stack if we're doing a major collection.
287 mark_stack_bdescr = allocGroup(MARK_STACK_BLOCKS);
288 mark_stack = (StgPtr *)mark_stack_bdescr->start;
289 mark_sp = mark_stack;
290 mark_splim = mark_stack + (MARK_STACK_BLOCKS * BLOCK_SIZE_W);
292 mark_stack_bdescr = NULL;
295 // this is the main thread
298 /* -----------------------------------------------------------------------
299 * follow all the roots that we know about:
300 * - mutable lists from each generation > N
301 * we want to *scavenge* these roots, not evacuate them: they're not
302 * going to move in this GC.
303 * Also do them in reverse generation order, for the usual reason:
304 * namely to reduce the likelihood of spurious old->new pointers.
306 for (g = RtsFlags.GcFlags.generations-1; g > N; g--) {
307 generations[g].saved_mut_list = generations[g].mut_list;
308 generations[g].mut_list = allocBlock();
309 // mut_list always has at least one block.
312 // the main thread is running: this prevents any other threads from
313 // exiting prematurely, so we can start them now.
314 // NB. do this after the mutable lists have been saved above, otherwise
315 // the other GC threads will be writing into the old mutable lists.
317 wakeup_gc_threads(n_gc_threads);
319 for (g = RtsFlags.GcFlags.generations-1; g > N; g--) {
320 scavenge_mutable_list(&generations[g]);
323 // follow roots from the CAF list (used by GHCi)
327 // follow all the roots that the application knows about.
331 #if defined(RTS_USER_SIGNALS)
332 // mark the signal handlers (signals should be already blocked)
333 markSignalHandlers(mark_root);
336 // Mark the weak pointer list, and prepare to detect dead weak pointers.
340 // Mark the stable pointer table.
341 markStablePtrTable(mark_root);
343 /* -------------------------------------------------------------------------
344 * Repeatedly scavenge all the areas we know about until there's no
345 * more scavenging to be done.
350 // The other threads are now stopped. We might recurse back to
351 // here, but from now on this is the only thread.
353 // if any blackholes are alive, make the threads that wait on
355 if (traverseBlackholeQueue()) {
360 // must be last... invariant is that everything is fully
361 // scavenged at this point.
362 if (traverseWeakPtrList()) { // returns rtsTrue if evaced something
367 // If we get to here, there's really nothing left to do.
371 shutdown_gc_threads(n_gc_threads);
373 // Update pointers from the Task list
376 // Now see which stable names are still alive.
380 // We call processHeapClosureForDead() on every closure destroyed during
381 // the current garbage collection, so we invoke LdvCensusForDead().
382 if (RtsFlags.ProfFlags.doHeapProfile == HEAP_BY_LDV
383 || RtsFlags.ProfFlags.bioSelector != NULL)
387 // NO MORE EVACUATION AFTER THIS POINT!
388 // Finally: compaction of the oldest generation.
389 if (major_gc && oldest_gen->steps[0].is_compacted) {
390 // save number of blocks for stats
391 oldgen_saved_blocks = oldest_gen->steps[0].n_old_blocks;
395 IF_DEBUG(sanity, checkGlobalTSOList(rtsFalse));
397 // Two-space collector: free the old to-space.
398 // g0s0->old_blocks is the old nursery
399 // g0s0->blocks is to-space from the previous GC
400 if (RtsFlags.GcFlags.generations == 1) {
401 if (g0s0->blocks != NULL) {
402 freeChain(g0s0->blocks);
407 // For each workspace, in each thread:
408 // * clear the BF_EVACUATED flag from each copied block
409 // * move the copied blocks to the step
415 for (t = 0; t < n_gc_threads; t++) {
419 for (s = 1; s < total_steps; s++) {
422 // ASSERT( ws->scan_bd == ws->todo_bd );
423 ASSERT( ws->scan_bd ? ws->scan_bd->u.scan == ws->scan_bd->free : 1 );
425 // Push the final block
426 if (ws->scan_bd) { push_scanned_block(ws->scan_bd, ws); }
428 ASSERT(countBlocks(ws->scavd_list) == ws->n_scavd_blocks);
430 prev = ws->part_list;
431 for (bd = ws->part_list; bd != NULL; bd = bd->link) {
432 bd->flags &= ~BF_EVACUATED; // now from-space
433 ws->step->n_words += bd->free - bd->start;
437 prev->link = ws->scavd_list;
439 for (bd = ws->scavd_list; bd != NULL; bd = bd->link) {
440 bd->flags &= ~BF_EVACUATED; // now from-space
441 ws->step->n_words += bd->free - bd->start;
444 prev->link = ws->step->blocks;
445 if (ws->part_list != NULL) {
446 ws->step->blocks = ws->part_list;
448 ws->step->blocks = ws->scavd_list;
450 ws->step->n_blocks += ws->n_part_blocks;
451 ws->step->n_blocks += ws->n_scavd_blocks;
452 ASSERT(countBlocks(ws->step->blocks) == ws->step->n_blocks);
453 ASSERT(countOccupied(ws->step->blocks) == ws->step->n_words);
458 // Two-space collector: swap the semi-spaces around.
459 // Currently: g0s0->old_blocks is the old nursery
460 // g0s0->blocks is to-space from this GC
461 // We want these the other way around.
462 if (RtsFlags.GcFlags.generations == 1) {
463 bdescr *nursery_blocks = g0s0->old_blocks;
464 nat n_nursery_blocks = g0s0->n_old_blocks;
465 g0s0->old_blocks = g0s0->blocks;
466 g0s0->n_old_blocks = g0s0->n_blocks;
467 g0s0->blocks = nursery_blocks;
468 g0s0->n_blocks = n_nursery_blocks;
471 /* run through all the generations/steps and tidy up
476 for (i=0; i < n_gc_threads; i++) {
477 if (n_gc_threads > 1) {
478 trace(TRACE_gc,"thread %d:", i);
479 trace(TRACE_gc," copied %ld", gc_threads[i]->copied * sizeof(W_));
480 trace(TRACE_gc," any_work %ld", gc_threads[i]->any_work);
481 trace(TRACE_gc," no_work %ld", gc_threads[i]->no_work);
482 trace(TRACE_gc," scav_find_work %ld", gc_threads[i]->scav_find_work);
484 copied += gc_threads[i]->copied;
488 for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
491 generations[g].collections++; // for stats
494 // Count the mutable list as bytes "copied" for the purposes of
495 // stats. Every mutable list is copied during every GC.
497 nat mut_list_size = 0;
498 for (bd = generations[g].mut_list; bd != NULL; bd = bd->link) {
499 mut_list_size += bd->free - bd->start;
501 copied += mut_list_size;
504 "mut_list_size: %lu (%d vars, %d arrays, %d MVARs, %d others)",
505 (unsigned long)(mut_list_size * sizeof(W_)),
506 mutlist_MUTVARS, mutlist_MUTARRS, mutlist_MVARS, mutlist_OTHERS);
509 for (s = 0; s < generations[g].n_steps; s++) {
511 stp = &generations[g].steps[s];
513 // for generations we collected...
516 /* free old memory and shift to-space into from-space for all
517 * the collected steps (except the allocation area). These
518 * freed blocks will probaby be quickly recycled.
520 if (!(g == 0 && s == 0 && RtsFlags.GcFlags.generations > 1)) {
521 if (stp->is_compacted)
523 // for a compacted step, just shift the new to-space
524 // onto the front of the now-compacted existing blocks.
525 for (bd = stp->blocks; bd != NULL; bd = bd->link) {
526 bd->flags &= ~BF_EVACUATED; // now from-space
527 stp->n_words += bd->free - bd->start;
529 // tack the new blocks on the end of the existing blocks
530 if (stp->old_blocks != NULL) {
531 for (bd = stp->old_blocks; bd != NULL; bd = next) {
532 // NB. this step might not be compacted next
533 // time, so reset the BF_COMPACTED flags.
534 // They are set before GC if we're going to
535 // compact. (search for BF_COMPACTED above).
536 bd->flags &= ~BF_COMPACTED;
539 bd->link = stp->blocks;
542 stp->blocks = stp->old_blocks;
544 // add the new blocks to the block tally
545 stp->n_blocks += stp->n_old_blocks;
546 ASSERT(countBlocks(stp->blocks) == stp->n_blocks);
547 ASSERT(countOccupied(stp->blocks) == stp->n_words);
551 freeChain(stp->old_blocks);
553 stp->old_blocks = NULL;
554 stp->n_old_blocks = 0;
557 /* LARGE OBJECTS. The current live large objects are chained on
558 * scavenged_large, having been moved during garbage
559 * collection from large_objects. Any objects left on
560 * large_objects list are therefore dead, so we free them here.
562 for (bd = stp->large_objects; bd != NULL; bd = next) {
568 // update the count of blocks used by large objects
569 for (bd = stp->scavenged_large_objects; bd != NULL; bd = bd->link) {
570 bd->flags &= ~BF_EVACUATED;
572 stp->large_objects = stp->scavenged_large_objects;
573 stp->n_large_blocks = stp->n_scavenged_large_blocks;
576 else // for older generations...
578 /* For older generations, we need to append the
579 * scavenged_large_object list (i.e. large objects that have been
580 * promoted during this GC) to the large_object list for that step.
582 for (bd = stp->scavenged_large_objects; bd; bd = next) {
584 bd->flags &= ~BF_EVACUATED;
585 dbl_link_onto(bd, &stp->large_objects);
588 // add the new blocks we promoted during this GC
589 stp->n_large_blocks += stp->n_scavenged_large_blocks;
594 // update the max size of older generations after a major GC
595 resize_generations();
597 // Calculate the amount of live data for stats.
598 live = calcLiveWords();
600 // Free the small objects allocated via allocate(), since this will
601 // all have been copied into G0S1 now.
602 if (RtsFlags.GcFlags.generations > 1) {
603 if (g0s0->blocks != NULL) {
604 freeChain(g0s0->blocks);
611 alloc_blocks_lim = RtsFlags.GcFlags.minAllocAreaSize;
613 // Start a new pinned_object_block
614 pinned_object_block = NULL;
616 // Free the mark stack.
617 if (mark_stack_bdescr != NULL) {
618 freeGroup(mark_stack_bdescr);
622 for (g = 0; g <= N; g++) {
623 for (s = 0; s < generations[g].n_steps; s++) {
624 stp = &generations[g].steps[s];
625 if (stp->bitmap != NULL) {
626 freeGroup(stp->bitmap);
634 // mark the garbage collected CAFs as dead
635 #if 0 && defined(DEBUG) // doesn't work at the moment
636 if (major_gc) { gcCAFs(); }
640 // resetStaticObjectForRetainerProfiling() must be called before
642 if (n_gc_threads > 1) {
643 barf("profiling is currently broken with multi-threaded GC");
644 // ToDo: fix the gct->scavenged_static_objects below
646 resetStaticObjectForRetainerProfiling(gct->scavenged_static_objects);
649 // zero the scavenged static object list
652 for (i = 0; i < n_gc_threads; i++) {
653 zero_static_object_list(gc_threads[i]->scavenged_static_objects);
660 // start any pending finalizers
662 scheduleFinalizers(last_free_capability, old_weak_ptr_list);
665 // send exceptions to any threads which were about to die
667 resurrectThreads(resurrected_threads);
670 // Update the stable pointer hash table.
671 updateStablePtrTable(major_gc);
673 // check sanity after GC
674 IF_DEBUG(sanity, checkSanity());
676 // extra GC trace info
677 if (traceClass(TRACE_gc|DEBUG_gc)) statDescribeGens();
680 // symbol-table based profiling
681 /* heapCensus(to_blocks); */ /* ToDo */
684 // restore enclosing cost centre
690 // check for memory leaks if DEBUG is on
691 memInventory(traceClass(DEBUG_gc));
694 #ifdef RTS_GTK_FRONTPANEL
695 if (RtsFlags.GcFlags.frontpanel) {
696 updateFrontPanelAfterGC( N, live );
700 // ok, GC over: tell the stats department what happened.
701 stat_endGC(allocated, live, copied, N);
703 #if defined(RTS_USER_SIGNALS)
704 if (RtsFlags.MiscFlags.install_signal_handlers) {
705 // unblock signals again
706 unblockUserSignals();
715 /* -----------------------------------------------------------------------------
716 * Mark all nodes pointed to by sparks in the spark queues (for GC) Does an
717 * implicit slide i.e. after marking all sparks are at the beginning of the
718 * spark pool and the spark pool only contains sparkable closures
719 * -------------------------------------------------------------------------- */
723 markSparkQueue (evac_fn evac, Capability *cap)
725 StgClosure **sparkp, **to_sparkp;
726 nat n, pruned_sparks; // stats only
729 PAR_TICKY_MARK_SPARK_QUEUE_START();
734 pool = &(cap->r.rSparks);
736 ASSERT_SPARK_POOL_INVARIANTS(pool);
738 #if defined(PARALLEL_HASKELL)
745 to_sparkp = pool->hd;
746 while (sparkp != pool->tl) {
747 ASSERT(*sparkp!=NULL);
748 ASSERT(LOOKS_LIKE_CLOSURE_PTR(((StgClosure *)*sparkp)));
749 // ToDo?: statistics gathering here (also for GUM!)
750 if (closure_SHOULD_SPARK(*sparkp)) {
752 *to_sparkp++ = *sparkp;
753 if (to_sparkp == pool->lim) {
754 to_sparkp = pool->base;
761 if (sparkp == pool->lim) {
765 pool->tl = to_sparkp;
767 PAR_TICKY_MARK_SPARK_QUEUE_END(n);
769 #if defined(PARALLEL_HASKELL)
770 debugTrace(DEBUG_sched,
771 "marked %d sparks and pruned %d sparks on [%x]",
772 n, pruned_sparks, mytid);
774 debugTrace(DEBUG_sched,
775 "marked %d sparks and pruned %d sparks",
779 debugTrace(DEBUG_sched,
780 "new spark queue len=%d; (hd=%p; tl=%p)\n",
781 sparkPoolSize(pool), pool->hd, pool->tl);
785 /* ---------------------------------------------------------------------------
786 Where are the roots that we know about?
788 - all the threads on the runnable queue
789 - all the threads on the blocked queue
790 - all the threads on the sleeping queue
791 - all the thread currently executing a _ccall_GC
792 - all the "main threads"
794 ------------------------------------------------------------------------ */
797 GetRoots( evac_fn evac )
803 // Each GC thread is responsible for following roots from the
804 // Capability of the same number. There will usually be the same
805 // or fewer Capabilities as GC threads, but just in case there
806 // are more, we mark every Capability whose number is the GC
807 // thread's index plus a multiple of the number of GC threads.
808 for (i = gct->thread_index; i < n_capabilities; i += n_gc_threads) {
809 cap = &capabilities[i];
810 evac((StgClosure **)(void *)&cap->run_queue_hd);
811 evac((StgClosure **)(void *)&cap->run_queue_tl);
812 #if defined(THREADED_RTS)
813 evac((StgClosure **)(void *)&cap->wakeup_queue_hd);
814 evac((StgClosure **)(void *)&cap->wakeup_queue_tl);
816 for (task = cap->suspended_ccalling_tasks; task != NULL;
818 debugTrace(DEBUG_sched,
819 "evac'ing suspended TSO %lu", (unsigned long)task->suspended_tso->id);
820 evac((StgClosure **)(void *)&task->suspended_tso);
823 #if defined(THREADED_RTS)
824 markSparkQueue(evac,cap);
828 #if !defined(THREADED_RTS)
829 evac((StgClosure **)(void *)&blocked_queue_hd);
830 evac((StgClosure **)(void *)&blocked_queue_tl);
831 evac((StgClosure **)(void *)&sleeping_queue);
835 /* -----------------------------------------------------------------------------
836 isAlive determines whether the given closure is still alive (after
837 a garbage collection) or not. It returns the new address of the
838 closure if it is alive, or NULL otherwise.
840 NOTE: Use it before compaction only!
841 It untags and (if needed) retags pointers to closures.
842 -------------------------------------------------------------------------- */
846 isAlive(StgClosure *p)
848 const StgInfoTable *info;
854 /* The tag and the pointer are split, to be merged later when needed. */
855 tag = GET_CLOSURE_TAG(p);
856 q = UNTAG_CLOSURE(p);
858 ASSERT(LOOKS_LIKE_CLOSURE_PTR(q));
861 // ignore static closures
863 // ToDo: for static closures, check the static link field.
864 // Problem here is that we sometimes don't set the link field, eg.
865 // for static closures with an empty SRT or CONSTR_STATIC_NOCAFs.
867 if (!HEAP_ALLOCED(q)) {
871 // ignore closures in generations that we're not collecting.
873 if (bd->gen_no > N) {
877 // if it's a pointer into to-space, then we're done
878 if (bd->flags & BF_EVACUATED) {
882 // large objects use the evacuated flag
883 if (bd->flags & BF_LARGE) {
887 // check the mark bit for compacted steps
888 if ((bd->flags & BF_COMPACTED) && is_marked((P_)q,bd)) {
892 switch (info->type) {
897 case IND_OLDGEN: // rely on compatible layout with StgInd
898 case IND_OLDGEN_PERM:
899 // follow indirections
900 p = ((StgInd *)q)->indirectee;
905 return ((StgEvacuated *)q)->evacuee;
908 if (((StgTSO *)q)->what_next == ThreadRelocated) {
909 p = (StgClosure *)((StgTSO *)q)->link;
921 /* -----------------------------------------------------------------------------
922 Figure out which generation to collect, initialise N and major_gc.
924 Also returns the total number of blocks in generations that will be
926 -------------------------------------------------------------------------- */
929 initialise_N (rtsBool force_major_gc)
932 nat s, blocks, blocks_total;
937 if (force_major_gc) {
938 N = RtsFlags.GcFlags.generations - 1;
943 for (g = RtsFlags.GcFlags.generations - 1; g >= 0; g--) {
945 for (s = 0; s < generations[g].n_steps; s++) {
946 blocks += generations[g].steps[s].n_words / BLOCK_SIZE_W;
947 blocks += generations[g].steps[s].n_large_blocks;
949 if (blocks >= generations[g].max_blocks) {
953 blocks_total += blocks;
957 blocks_total += countNurseryBlocks();
959 major_gc = (N == RtsFlags.GcFlags.generations-1);
963 /* -----------------------------------------------------------------------------
964 Initialise the gc_thread structures.
965 -------------------------------------------------------------------------- */
968 alloc_gc_thread (int n)
974 t = stgMallocBytes(sizeof(gc_thread) + total_steps * sizeof(step_workspace),
979 initCondition(&t->wake_cond);
980 initMutex(&t->wake_mutex);
981 t->wakeup = rtsFalse;
986 t->free_blocks = NULL;
995 for (s = 0; s < total_steps; s++)
998 ws->step = &all_steps[s];
999 ASSERT(s == ws->step->abs_no);
1005 ws->buffer_todo_bd = NULL;
1007 ws->part_list = NULL;
1008 ws->n_part_blocks = 0;
1010 ws->scavd_list = NULL;
1011 ws->n_scavd_blocks = 0;
1019 alloc_gc_threads (void)
1021 if (gc_threads == NULL) {
1022 #if defined(THREADED_RTS)
1024 gc_threads = stgMallocBytes (RtsFlags.ParFlags.gcThreads *
1026 "alloc_gc_threads");
1028 for (i = 0; i < RtsFlags.ParFlags.gcThreads; i++) {
1029 gc_threads[i] = alloc_gc_thread(i);
1032 gc_threads = stgMallocBytes (sizeof(gc_thread*),
1033 "alloc_gc_threads");
1035 gc_threads[0] = alloc_gc_thread(0);
1040 /* ----------------------------------------------------------------------------
1042 ------------------------------------------------------------------------- */
1044 static nat gc_running_threads;
1046 #if defined(THREADED_RTS)
1047 static Mutex gc_running_mutex;
1054 ACQUIRE_LOCK(&gc_running_mutex);
1055 n_running = ++gc_running_threads;
1056 RELEASE_LOCK(&gc_running_mutex);
1057 ASSERT(n_running <= n_gc_threads);
1065 ACQUIRE_LOCK(&gc_running_mutex);
1066 ASSERT(n_gc_threads != 0);
1067 n_running = --gc_running_threads;
1068 RELEASE_LOCK(&gc_running_mutex);
1073 // gc_thread_work(): Scavenge until there's no work left to do and all
1074 // the running threads are idle.
1077 gc_thread_work (void)
1081 debugTrace(DEBUG_gc, "GC thread %d working", gct->thread_index);
1083 // gc_running_threads has already been incremented for us; either
1084 // this is the main thread and we incremented it inside
1085 // GarbageCollect(), or this is a worker thread and the main
1086 // thread bumped gc_running_threads before waking us up.
1088 // Every thread evacuates some roots.
1090 GetRoots(mark_root);
1094 // scavenge_loop() only exits when there's no work to do
1097 debugTrace(DEBUG_gc, "GC thread %d idle (%d still running)",
1098 gct->thread_index, r);
1100 while (gc_running_threads != 0) {
1106 // any_work() does not remove the work from the queue, it
1107 // just checks for the presence of work. If we find any,
1108 // then we increment gc_running_threads and go back to
1109 // scavenge_loop() to perform any pending work.
1112 // All threads are now stopped
1113 debugTrace(DEBUG_gc, "GC thread %d finished.", gct->thread_index);
1117 #if defined(THREADED_RTS)
1119 gc_thread_mainloop (void)
1121 while (!gct->exit) {
1123 // Wait until we're told to wake up
1124 ACQUIRE_LOCK(&gct->wake_mutex);
1125 gct->wakeup = rtsFalse;
1126 while (!gct->wakeup) {
1127 debugTrace(DEBUG_gc, "GC thread %d standing by...",
1129 waitCondition(&gct->wake_cond, &gct->wake_mutex);
1131 RELEASE_LOCK(&gct->wake_mutex);
1132 if (gct->exit) break;
1135 // start performance counters in this thread...
1136 if (gct->papi_events == -1) {
1137 papi_init_eventset(&gct->papi_events);
1139 papi_thread_start_gc1_count(gct->papi_events);
1145 // count events in this thread towards the GC totals
1146 papi_thread_stop_gc1_count(gct->papi_events);
1152 #if defined(THREADED_RTS)
1154 gc_thread_entry (gc_thread *my_gct)
1157 debugTrace(DEBUG_gc, "GC thread %d starting...", gct->thread_index);
1158 gct->id = osThreadId();
1159 gc_thread_mainloop();
1164 start_gc_threads (void)
1166 #if defined(THREADED_RTS)
1169 static rtsBool done = rtsFalse;
1171 gc_running_threads = 0;
1172 initMutex(&gc_running_mutex);
1175 // Start from 1: the main thread is 0
1176 for (i = 1; i < RtsFlags.ParFlags.gcThreads; i++) {
1177 createOSThread(&id, (OSThreadProc*)&gc_thread_entry,
1186 wakeup_gc_threads (nat n_threads USED_IF_THREADS)
1188 #if defined(THREADED_RTS)
1190 for (i=1; i < n_threads; i++) {
1192 ACQUIRE_LOCK(&gc_threads[i]->wake_mutex);
1193 gc_threads[i]->wakeup = rtsTrue;
1194 signalCondition(&gc_threads[i]->wake_cond);
1195 RELEASE_LOCK(&gc_threads[i]->wake_mutex);
1200 // After GC is complete, we must wait for all GC threads to enter the
1201 // standby state, otherwise they may still be executing inside
1202 // any_work(), and may even remain awake until the next GC starts.
1204 shutdown_gc_threads (nat n_threads USED_IF_THREADS)
1206 #if defined(THREADED_RTS)
1209 for (i=1; i < n_threads; i++) {
1211 ACQUIRE_LOCK(&gc_threads[i]->wake_mutex);
1212 wakeup = gc_threads[i]->wakeup;
1213 // wakeup is false while the thread is waiting
1214 RELEASE_LOCK(&gc_threads[i]->wake_mutex);
1220 /* ----------------------------------------------------------------------------
1221 Initialise a generation that is to be collected
1222 ------------------------------------------------------------------------- */
1225 init_collected_gen (nat g, nat n_threads)
1232 // Throw away the current mutable list. Invariant: the mutable
1233 // list always has at least one block; this means we can avoid a
1234 // check for NULL in recordMutable().
1236 freeChain(generations[g].mut_list);
1237 generations[g].mut_list = allocBlock();
1238 for (i = 0; i < n_capabilities; i++) {
1239 freeChain(capabilities[i].mut_lists[g]);
1240 capabilities[i].mut_lists[g] = allocBlock();
1244 for (s = 0; s < generations[g].n_steps; s++) {
1246 // generation 0, step 0 doesn't need to-space
1247 if (g == 0 && s == 0 && RtsFlags.GcFlags.generations > 1) {
1251 stp = &generations[g].steps[s];
1252 ASSERT(stp->gen_no == g);
1254 // deprecate the existing blocks
1255 stp->old_blocks = stp->blocks;
1256 stp->n_old_blocks = stp->n_blocks;
1261 // we don't have any to-be-scavenged blocks yet
1263 stp->todos_last = NULL;
1266 // initialise the large object queues.
1267 stp->scavenged_large_objects = NULL;
1268 stp->n_scavenged_large_blocks = 0;
1270 // mark the large objects as not evacuated yet
1271 for (bd = stp->large_objects; bd; bd = bd->link) {
1272 bd->flags &= ~BF_EVACUATED;
1275 // for a compacted step, we need to allocate the bitmap
1276 if (stp->is_compacted) {
1277 nat bitmap_size; // in bytes
1278 bdescr *bitmap_bdescr;
1281 bitmap_size = stp->n_old_blocks * BLOCK_SIZE / (sizeof(W_)*BITS_PER_BYTE);
1283 if (bitmap_size > 0) {
1284 bitmap_bdescr = allocGroup((lnat)BLOCK_ROUND_UP(bitmap_size)
1286 stp->bitmap = bitmap_bdescr;
1287 bitmap = bitmap_bdescr->start;
1289 debugTrace(DEBUG_gc, "bitmap_size: %d, bitmap: %p",
1290 bitmap_size, bitmap);
1292 // don't forget to fill it with zeros!
1293 memset(bitmap, 0, bitmap_size);
1295 // For each block in this step, point to its bitmap from the
1296 // block descriptor.
1297 for (bd=stp->old_blocks; bd != NULL; bd = bd->link) {
1298 bd->u.bitmap = bitmap;
1299 bitmap += BLOCK_SIZE_W / (sizeof(W_)*BITS_PER_BYTE);
1301 // Also at this point we set the BF_COMPACTED flag
1302 // for this block. The invariant is that
1303 // BF_COMPACTED is always unset, except during GC
1304 // when it is set on those blocks which will be
1306 bd->flags |= BF_COMPACTED;
1312 // For each GC thread, for each step, allocate a "todo" block to
1313 // store evacuated objects to be scavenged, and a block to store
1314 // evacuated objects that do not need to be scavenged.
1315 for (t = 0; t < n_threads; t++) {
1316 for (s = 0; s < generations[g].n_steps; s++) {
1318 // we don't copy objects into g0s0, unless -G0
1319 if (g==0 && s==0 && RtsFlags.GcFlags.generations > 1) continue;
1321 ws = &gc_threads[t]->steps[g * RtsFlags.GcFlags.steps + s];
1325 ws->todo_large_objects = NULL;
1327 ws->part_list = NULL;
1328 ws->n_part_blocks = 0;
1330 // allocate the first to-space block; extra blocks will be
1331 // chained on as necessary.
1333 ws->buffer_todo_bd = NULL;
1334 alloc_todo_block(ws,0);
1336 ws->scavd_list = NULL;
1337 ws->n_scavd_blocks = 0;
1343 /* ----------------------------------------------------------------------------
1344 Initialise a generation that is *not* to be collected
1345 ------------------------------------------------------------------------- */
1348 init_uncollected_gen (nat g, nat threads)
1355 for (s = 0; s < generations[g].n_steps; s++) {
1356 stp = &generations[g].steps[s];
1357 stp->scavenged_large_objects = NULL;
1358 stp->n_scavenged_large_blocks = 0;
1361 for (t = 0; t < threads; t++) {
1362 for (s = 0; s < generations[g].n_steps; s++) {
1364 ws = &gc_threads[t]->steps[g * RtsFlags.GcFlags.steps + s];
1367 ws->buffer_todo_bd = NULL;
1368 ws->todo_large_objects = NULL;
1370 ws->part_list = NULL;
1371 ws->n_part_blocks = 0;
1373 ws->scavd_list = NULL;
1374 ws->n_scavd_blocks = 0;
1376 // If the block at the head of the list in this generation
1377 // is less than 3/4 full, then use it as a todo block.
1378 if (stp->blocks && isPartiallyFull(stp->blocks))
1380 ws->todo_bd = stp->blocks;
1381 ws->todo_free = ws->todo_bd->free;
1382 ws->todo_lim = ws->todo_bd->start + BLOCK_SIZE_W;
1383 stp->blocks = stp->blocks->link;
1385 stp->n_words -= ws->todo_bd->free - ws->todo_bd->start;
1386 ws->todo_bd->link = NULL;
1388 // this block is also the scan block; we must scan
1389 // from the current end point.
1390 ws->scan_bd = ws->todo_bd;
1391 ws->scan_bd->u.scan = ws->scan_bd->free;
1393 // subtract the contents of this block from the stats,
1394 // because we'll count the whole block later.
1395 copied -= ws->scan_bd->free - ws->scan_bd->start;
1401 alloc_todo_block(ws,0);
1406 // Move the private mutable lists from each capability onto the
1407 // main mutable list for the generation.
1408 for (i = 0; i < n_capabilities; i++) {
1409 for (bd = capabilities[i].mut_lists[g];
1410 bd->link != NULL; bd = bd->link) {
1413 bd->link = generations[g].mut_list;
1414 generations[g].mut_list = capabilities[i].mut_lists[g];
1415 capabilities[i].mut_lists[g] = allocBlock();
1419 /* -----------------------------------------------------------------------------
1420 Initialise a gc_thread before GC
1421 -------------------------------------------------------------------------- */
1424 init_gc_thread (gc_thread *t)
1426 t->static_objects = END_OF_STATIC_LIST;
1427 t->scavenged_static_objects = END_OF_STATIC_LIST;
1429 t->failed_to_evac = rtsFalse;
1430 t->eager_promotion = rtsTrue;
1431 t->thunk_selector_depth = 0;
1435 t->scav_find_work = 0;
1439 /* -----------------------------------------------------------------------------
1440 Function we pass to GetRoots to evacuate roots.
1441 -------------------------------------------------------------------------- */
1444 mark_root(StgClosure **root)
1449 /* -----------------------------------------------------------------------------
1450 Initialising the static object & mutable lists
1451 -------------------------------------------------------------------------- */
1454 zero_static_object_list(StgClosure* first_static)
1458 const StgInfoTable *info;
1460 for (p = first_static; p != END_OF_STATIC_LIST; p = link) {
1462 link = *STATIC_LINK(info, p);
1463 *STATIC_LINK(info,p) = NULL;
1467 /* -----------------------------------------------------------------------------
1469 -------------------------------------------------------------------------- */
1476 for (c = (StgIndStatic *)revertible_caf_list; c != NULL;
1477 c = (StgIndStatic *)c->static_link)
1479 SET_INFO(c, c->saved_info);
1480 c->saved_info = NULL;
1481 // could, but not necessary: c->static_link = NULL;
1483 revertible_caf_list = NULL;
1487 markCAFs( evac_fn evac )
1491 for (c = (StgIndStatic *)caf_list; c != NULL;
1492 c = (StgIndStatic *)c->static_link)
1494 evac(&c->indirectee);
1496 for (c = (StgIndStatic *)revertible_caf_list; c != NULL;
1497 c = (StgIndStatic *)c->static_link)
1499 evac(&c->indirectee);
1503 /* ----------------------------------------------------------------------------
1504 Update the pointers from the task list
1506 These are treated as weak pointers because we want to allow a main
1507 thread to get a BlockedOnDeadMVar exception in the same way as any
1508 other thread. Note that the threads should all have been retained
1509 by GC by virtue of being on the all_threads list, we're just
1510 updating pointers here.
1511 ------------------------------------------------------------------------- */
1514 update_task_list (void)
1518 for (task = all_tasks; task != NULL; task = task->all_link) {
1519 if (!task->stopped && task->tso) {
1520 ASSERT(task->tso->bound == task);
1521 tso = (StgTSO *) isAlive((StgClosure *)task->tso);
1523 barf("task %p: main thread %d has been GC'd",
1536 /* ----------------------------------------------------------------------------
1537 Reset the sizes of the older generations when we do a major
1540 CURRENT STRATEGY: make all generations except zero the same size.
1541 We have to stay within the maximum heap size, and leave a certain
1542 percentage of the maximum heap size available to allocate into.
1543 ------------------------------------------------------------------------- */
1546 resize_generations (void)
1550 if (major_gc && RtsFlags.GcFlags.generations > 1) {
1551 nat live, size, min_alloc;
1552 nat max = RtsFlags.GcFlags.maxHeapSize;
1553 nat gens = RtsFlags.GcFlags.generations;
1555 // live in the oldest generations
1556 live = (oldest_gen->steps[0].n_words + BLOCK_SIZE_W - 1) / BLOCK_SIZE_W+
1557 oldest_gen->steps[0].n_large_blocks;
1559 // default max size for all generations except zero
1560 size = stg_max(live * RtsFlags.GcFlags.oldGenFactor,
1561 RtsFlags.GcFlags.minOldGenSize);
1563 // minimum size for generation zero
1564 min_alloc = stg_max((RtsFlags.GcFlags.pcFreeHeap * max) / 200,
1565 RtsFlags.GcFlags.minAllocAreaSize);
1567 // Auto-enable compaction when the residency reaches a
1568 // certain percentage of the maximum heap size (default: 30%).
1569 if (RtsFlags.GcFlags.generations > 1 &&
1570 (RtsFlags.GcFlags.compact ||
1572 oldest_gen->steps[0].n_blocks >
1573 (RtsFlags.GcFlags.compactThreshold * max) / 100))) {
1574 oldest_gen->steps[0].is_compacted = 1;
1575 // debugBelch("compaction: on\n", live);
1577 oldest_gen->steps[0].is_compacted = 0;
1578 // debugBelch("compaction: off\n", live);
1581 // if we're going to go over the maximum heap size, reduce the
1582 // size of the generations accordingly. The calculation is
1583 // different if compaction is turned on, because we don't need
1584 // to double the space required to collect the old generation.
1587 // this test is necessary to ensure that the calculations
1588 // below don't have any negative results - we're working
1589 // with unsigned values here.
1590 if (max < min_alloc) {
1594 if (oldest_gen->steps[0].is_compacted) {
1595 if ( (size + (size - 1) * (gens - 2) * 2) + min_alloc > max ) {
1596 size = (max - min_alloc) / ((gens - 1) * 2 - 1);
1599 if ( (size * (gens - 1) * 2) + min_alloc > max ) {
1600 size = (max - min_alloc) / ((gens - 1) * 2);
1610 debugBelch("live: %d, min_alloc: %d, size : %d, max = %d\n", live,
1611 min_alloc, size, max);
1614 for (g = 0; g < gens; g++) {
1615 generations[g].max_blocks = size;
1620 /* -----------------------------------------------------------------------------
1621 Calculate the new size of the nursery, and resize it.
1622 -------------------------------------------------------------------------- */
1625 resize_nursery (void)
1627 if (RtsFlags.GcFlags.generations == 1)
1628 { // Two-space collector:
1631 /* set up a new nursery. Allocate a nursery size based on a
1632 * function of the amount of live data (by default a factor of 2)
1633 * Use the blocks from the old nursery if possible, freeing up any
1636 * If we get near the maximum heap size, then adjust our nursery
1637 * size accordingly. If the nursery is the same size as the live
1638 * data (L), then we need 3L bytes. We can reduce the size of the
1639 * nursery to bring the required memory down near 2L bytes.
1641 * A normal 2-space collector would need 4L bytes to give the same
1642 * performance we get from 3L bytes, reducing to the same
1643 * performance at 2L bytes.
1645 blocks = g0s0->n_old_blocks;
1647 if ( RtsFlags.GcFlags.maxHeapSize != 0 &&
1648 blocks * RtsFlags.GcFlags.oldGenFactor * 2 >
1649 RtsFlags.GcFlags.maxHeapSize )
1651 long adjusted_blocks; // signed on purpose
1654 adjusted_blocks = (RtsFlags.GcFlags.maxHeapSize - 2 * blocks);
1656 debugTrace(DEBUG_gc, "near maximum heap size of 0x%x blocks, blocks = %d, adjusted to %ld",
1657 RtsFlags.GcFlags.maxHeapSize, blocks, adjusted_blocks);
1659 pc_free = adjusted_blocks * 100 / RtsFlags.GcFlags.maxHeapSize;
1660 if (pc_free < RtsFlags.GcFlags.pcFreeHeap) /* might even * be < 0 */
1664 blocks = adjusted_blocks;
1668 blocks *= RtsFlags.GcFlags.oldGenFactor;
1669 if (blocks < RtsFlags.GcFlags.minAllocAreaSize)
1671 blocks = RtsFlags.GcFlags.minAllocAreaSize;
1674 resizeNurseries(blocks);
1676 else // Generational collector
1679 * If the user has given us a suggested heap size, adjust our
1680 * allocation area to make best use of the memory available.
1682 if (RtsFlags.GcFlags.heapSizeSuggestion)
1685 nat needed = calcNeeded(); // approx blocks needed at next GC
1687 /* Guess how much will be live in generation 0 step 0 next time.
1688 * A good approximation is obtained by finding the
1689 * percentage of g0s0 that was live at the last minor GC.
1691 * We have an accurate figure for the amount of copied data in
1692 * 'copied', but we must convert this to a number of blocks, with
1693 * a small adjustment for estimated slop at the end of a block
1698 g0s0_pcnt_kept = ((copied / (BLOCK_SIZE_W - 10)) * 100)
1699 / countNurseryBlocks();
1702 /* Estimate a size for the allocation area based on the
1703 * information available. We might end up going slightly under
1704 * or over the suggested heap size, but we should be pretty
1707 * Formula: suggested - needed
1708 * ----------------------------
1709 * 1 + g0s0_pcnt_kept/100
1711 * where 'needed' is the amount of memory needed at the next
1712 * collection for collecting all steps except g0s0.
1715 (((long)RtsFlags.GcFlags.heapSizeSuggestion - (long)needed) * 100) /
1716 (100 + (long)g0s0_pcnt_kept);
1718 if (blocks < (long)RtsFlags.GcFlags.minAllocAreaSize) {
1719 blocks = RtsFlags.GcFlags.minAllocAreaSize;
1722 resizeNurseries((nat)blocks);
1726 // we might have added extra large blocks to the nursery, so
1727 // resize back to minAllocAreaSize again.
1728 resizeNurseriesFixed(RtsFlags.GcFlags.minAllocAreaSize);
1733 /* -----------------------------------------------------------------------------
1734 Sanity code for CAF garbage collection.
1736 With DEBUG turned on, we manage a CAF list in addition to the SRT
1737 mechanism. After GC, we run down the CAF list and blackhole any
1738 CAFs which have been garbage collected. This means we get an error
1739 whenever the program tries to enter a garbage collected CAF.
1741 Any garbage collected CAFs are taken off the CAF list at the same
1743 -------------------------------------------------------------------------- */
1745 #if 0 && defined(DEBUG)
1752 const StgInfoTable *info;
1763 ASSERT(info->type == IND_STATIC);
1765 if (STATIC_LINK(info,p) == NULL) {
1766 debugTrace(DEBUG_gccafs, "CAF gc'd at 0x%04lx", (long)p);
1768 SET_INFO(p,&stg_BLACKHOLE_info);
1769 p = STATIC_LINK2(info,p);
1773 pp = &STATIC_LINK2(info,p);
1780 debugTrace(DEBUG_gccafs, "%d CAFs live", i);