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, max_copied, avg_copied, slop;
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 // Push the final block
424 push_scanned_block(ws->todo_bd, ws);
427 ASSERT(gct->scan_bd == NULL);
428 ASSERT(countBlocks(ws->scavd_list) == ws->n_scavd_blocks);
431 for (bd = ws->scavd_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->step->blocks;
438 ws->step->blocks = ws->scavd_list;
440 ws->step->n_blocks += ws->n_scavd_blocks;
443 for (bd = ws->part_list; bd != NULL; bd = next) {
445 if (bd->free == bd->start) {
447 ws->part_list = next;
454 bd->flags &= ~BF_EVACUATED; // now from-space
455 ws->step->n_words += bd->free - bd->start;
460 prev->link = ws->step->blocks;
461 ws->step->blocks = ws->part_list;
463 ws->step->n_blocks += ws->n_part_blocks;
465 ASSERT(countBlocks(ws->step->blocks) == ws->step->n_blocks);
466 ASSERT(countOccupied(ws->step->blocks) == ws->step->n_words);
471 // Two-space collector: swap the semi-spaces around.
472 // Currently: g0s0->old_blocks is the old nursery
473 // g0s0->blocks is to-space from this GC
474 // We want these the other way around.
475 if (RtsFlags.GcFlags.generations == 1) {
476 bdescr *nursery_blocks = g0s0->old_blocks;
477 nat n_nursery_blocks = g0s0->n_old_blocks;
478 g0s0->old_blocks = g0s0->blocks;
479 g0s0->n_old_blocks = g0s0->n_blocks;
480 g0s0->blocks = nursery_blocks;
481 g0s0->n_blocks = n_nursery_blocks;
484 /* run through all the generations/steps and tidy up
491 for (i=0; i < n_gc_threads; i++) {
492 if (n_gc_threads > 1) {
493 trace(TRACE_gc,"thread %d:", i);
494 trace(TRACE_gc," copied %ld", gc_threads[i]->copied * sizeof(W_));
495 trace(TRACE_gc," scanned %ld", gc_threads[i]->scanned * sizeof(W_));
496 trace(TRACE_gc," any_work %ld", gc_threads[i]->any_work);
497 trace(TRACE_gc," no_work %ld", gc_threads[i]->no_work);
498 trace(TRACE_gc," scav_find_work %ld", gc_threads[i]->scav_find_work);
500 copied += gc_threads[i]->copied;
501 max_copied = stg_max(gc_threads[i]->copied, max_copied);
503 if (n_gc_threads == 1) {
511 for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
514 generations[g].collections++; // for stats
515 if (n_gc_threads > 1) generations[g].par_collections++;
518 // Count the mutable list as bytes "copied" for the purposes of
519 // stats. Every mutable list is copied during every GC.
521 nat mut_list_size = 0;
522 for (bd = generations[g].mut_list; bd != NULL; bd = bd->link) {
523 mut_list_size += bd->free - bd->start;
525 copied += mut_list_size;
528 "mut_list_size: %lu (%d vars, %d arrays, %d MVARs, %d others)",
529 (unsigned long)(mut_list_size * sizeof(W_)),
530 mutlist_MUTVARS, mutlist_MUTARRS, mutlist_MVARS, mutlist_OTHERS);
533 for (s = 0; s < generations[g].n_steps; s++) {
535 stp = &generations[g].steps[s];
537 // for generations we collected...
540 /* free old memory and shift to-space into from-space for all
541 * the collected steps (except the allocation area). These
542 * freed blocks will probaby be quickly recycled.
544 if (!(g == 0 && s == 0 && RtsFlags.GcFlags.generations > 1)) {
545 if (stp->is_compacted)
547 // for a compacted step, just shift the new to-space
548 // onto the front of the now-compacted existing blocks.
549 for (bd = stp->blocks; bd != NULL; bd = bd->link) {
550 bd->flags &= ~BF_EVACUATED; // now from-space
551 stp->n_words += bd->free - bd->start;
553 // tack the new blocks on the end of the existing blocks
554 if (stp->old_blocks != NULL) {
555 for (bd = stp->old_blocks; bd != NULL; bd = next) {
556 // NB. this step might not be compacted next
557 // time, so reset the BF_COMPACTED flags.
558 // They are set before GC if we're going to
559 // compact. (search for BF_COMPACTED above).
560 bd->flags &= ~BF_COMPACTED;
563 bd->link = stp->blocks;
566 stp->blocks = stp->old_blocks;
568 // add the new blocks to the block tally
569 stp->n_blocks += stp->n_old_blocks;
570 ASSERT(countBlocks(stp->blocks) == stp->n_blocks);
571 ASSERT(countOccupied(stp->blocks) == stp->n_words);
575 freeChain(stp->old_blocks);
577 stp->old_blocks = NULL;
578 stp->n_old_blocks = 0;
581 /* LARGE OBJECTS. The current live large objects are chained on
582 * scavenged_large, having been moved during garbage
583 * collection from large_objects. Any objects left on
584 * large_objects list are therefore dead, so we free them here.
586 for (bd = stp->large_objects; bd != NULL; bd = next) {
592 // update the count of blocks used by large objects
593 for (bd = stp->scavenged_large_objects; bd != NULL; bd = bd->link) {
594 bd->flags &= ~BF_EVACUATED;
596 stp->large_objects = stp->scavenged_large_objects;
597 stp->n_large_blocks = stp->n_scavenged_large_blocks;
600 else // for older generations...
602 /* For older generations, we need to append the
603 * scavenged_large_object list (i.e. large objects that have been
604 * promoted during this GC) to the large_object list for that step.
606 for (bd = stp->scavenged_large_objects; bd; bd = next) {
608 bd->flags &= ~BF_EVACUATED;
609 dbl_link_onto(bd, &stp->large_objects);
612 // add the new blocks we promoted during this GC
613 stp->n_large_blocks += stp->n_scavenged_large_blocks;
618 // update the max size of older generations after a major GC
619 resize_generations();
621 // Calculate the amount of live data for stats.
622 live = calcLiveWords();
624 // Free the small objects allocated via allocate(), since this will
625 // all have been copied into G0S1 now.
626 if (RtsFlags.GcFlags.generations > 1) {
627 if (g0s0->blocks != NULL) {
628 freeChain(g0s0->blocks);
635 alloc_blocks_lim = RtsFlags.GcFlags.minAllocAreaSize;
637 // Start a new pinned_object_block
638 pinned_object_block = NULL;
640 // Free the mark stack.
641 if (mark_stack_bdescr != NULL) {
642 freeGroup(mark_stack_bdescr);
646 for (g = 0; g <= N; g++) {
647 for (s = 0; s < generations[g].n_steps; s++) {
648 stp = &generations[g].steps[s];
649 if (stp->bitmap != NULL) {
650 freeGroup(stp->bitmap);
658 // mark the garbage collected CAFs as dead
659 #if 0 && defined(DEBUG) // doesn't work at the moment
660 if (major_gc) { gcCAFs(); }
664 // resetStaticObjectForRetainerProfiling() must be called before
666 if (n_gc_threads > 1) {
667 barf("profiling is currently broken with multi-threaded GC");
668 // ToDo: fix the gct->scavenged_static_objects below
670 resetStaticObjectForRetainerProfiling(gct->scavenged_static_objects);
673 // zero the scavenged static object list
676 for (i = 0; i < n_gc_threads; i++) {
677 zero_static_object_list(gc_threads[i]->scavenged_static_objects);
684 // start any pending finalizers
686 scheduleFinalizers(last_free_capability, old_weak_ptr_list);
689 // send exceptions to any threads which were about to die
691 resurrectThreads(resurrected_threads);
694 // Update the stable pointer hash table.
695 updateStablePtrTable(major_gc);
697 // check sanity after GC
698 IF_DEBUG(sanity, checkSanity());
700 // extra GC trace info
701 if (traceClass(TRACE_gc|DEBUG_gc)) statDescribeGens();
704 // symbol-table based profiling
705 /* heapCensus(to_blocks); */ /* ToDo */
708 // restore enclosing cost centre
714 // check for memory leaks if DEBUG is on
715 memInventory(traceClass(DEBUG_gc));
718 #ifdef RTS_GTK_FRONTPANEL
719 if (RtsFlags.GcFlags.frontpanel) {
720 updateFrontPanelAfterGC( N, live );
724 // ok, GC over: tell the stats department what happened.
725 slop = calcLiveBlocks() * BLOCK_SIZE_W - live;
726 stat_endGC(allocated, live, copied, N, max_copied, avg_copied, slop);
728 #if defined(RTS_USER_SIGNALS)
729 if (RtsFlags.MiscFlags.install_signal_handlers) {
730 // unblock signals again
731 unblockUserSignals();
740 /* -----------------------------------------------------------------------------
741 * Mark all nodes pointed to by sparks in the spark queues (for GC) Does an
742 * implicit slide i.e. after marking all sparks are at the beginning of the
743 * spark pool and the spark pool only contains sparkable closures
744 * -------------------------------------------------------------------------- */
748 markSparkQueue (evac_fn evac, Capability *cap)
750 StgClosure **sparkp, **to_sparkp;
751 nat n, pruned_sparks; // stats only
754 PAR_TICKY_MARK_SPARK_QUEUE_START();
759 pool = &(cap->r.rSparks);
761 ASSERT_SPARK_POOL_INVARIANTS(pool);
763 #if defined(PARALLEL_HASKELL)
770 to_sparkp = pool->hd;
771 while (sparkp != pool->tl) {
772 ASSERT(*sparkp!=NULL);
773 ASSERT(LOOKS_LIKE_CLOSURE_PTR(((StgClosure *)*sparkp)));
774 // ToDo?: statistics gathering here (also for GUM!)
775 if (closure_SHOULD_SPARK(*sparkp)) {
777 *to_sparkp++ = *sparkp;
778 if (to_sparkp == pool->lim) {
779 to_sparkp = pool->base;
786 if (sparkp == pool->lim) {
790 pool->tl = to_sparkp;
792 PAR_TICKY_MARK_SPARK_QUEUE_END(n);
794 #if defined(PARALLEL_HASKELL)
795 debugTrace(DEBUG_sched,
796 "marked %d sparks and pruned %d sparks on [%x]",
797 n, pruned_sparks, mytid);
799 debugTrace(DEBUG_sched,
800 "marked %d sparks and pruned %d sparks",
804 debugTrace(DEBUG_sched,
805 "new spark queue len=%d; (hd=%p; tl=%p)\n",
806 sparkPoolSize(pool), pool->hd, pool->tl);
810 /* ---------------------------------------------------------------------------
811 Where are the roots that we know about?
813 - all the threads on the runnable queue
814 - all the threads on the blocked queue
815 - all the threads on the sleeping queue
816 - all the thread currently executing a _ccall_GC
817 - all the "main threads"
819 ------------------------------------------------------------------------ */
822 GetRoots( evac_fn evac )
828 // Each GC thread is responsible for following roots from the
829 // Capability of the same number. There will usually be the same
830 // or fewer Capabilities as GC threads, but just in case there
831 // are more, we mark every Capability whose number is the GC
832 // thread's index plus a multiple of the number of GC threads.
833 for (i = gct->thread_index; i < n_capabilities; i += n_gc_threads) {
834 cap = &capabilities[i];
835 evac((StgClosure **)(void *)&cap->run_queue_hd);
836 evac((StgClosure **)(void *)&cap->run_queue_tl);
837 #if defined(THREADED_RTS)
838 evac((StgClosure **)(void *)&cap->wakeup_queue_hd);
839 evac((StgClosure **)(void *)&cap->wakeup_queue_tl);
841 for (task = cap->suspended_ccalling_tasks; task != NULL;
843 debugTrace(DEBUG_sched,
844 "evac'ing suspended TSO %lu", (unsigned long)task->suspended_tso->id);
845 evac((StgClosure **)(void *)&task->suspended_tso);
848 #if defined(THREADED_RTS)
849 markSparkQueue(evac,cap);
853 #if !defined(THREADED_RTS)
854 evac((StgClosure **)(void *)&blocked_queue_hd);
855 evac((StgClosure **)(void *)&blocked_queue_tl);
856 evac((StgClosure **)(void *)&sleeping_queue);
860 /* -----------------------------------------------------------------------------
861 isAlive determines whether the given closure is still alive (after
862 a garbage collection) or not. It returns the new address of the
863 closure if it is alive, or NULL otherwise.
865 NOTE: Use it before compaction only!
866 It untags and (if needed) retags pointers to closures.
867 -------------------------------------------------------------------------- */
871 isAlive(StgClosure *p)
873 const StgInfoTable *info;
879 /* The tag and the pointer are split, to be merged later when needed. */
880 tag = GET_CLOSURE_TAG(p);
881 q = UNTAG_CLOSURE(p);
883 ASSERT(LOOKS_LIKE_CLOSURE_PTR(q));
886 // ignore static closures
888 // ToDo: for static closures, check the static link field.
889 // Problem here is that we sometimes don't set the link field, eg.
890 // for static closures with an empty SRT or CONSTR_STATIC_NOCAFs.
892 if (!HEAP_ALLOCED(q)) {
896 // ignore closures in generations that we're not collecting.
898 if (bd->gen_no > N) {
902 // if it's a pointer into to-space, then we're done
903 if (bd->flags & BF_EVACUATED) {
907 // large objects use the evacuated flag
908 if (bd->flags & BF_LARGE) {
912 // check the mark bit for compacted steps
913 if ((bd->flags & BF_COMPACTED) && is_marked((P_)q,bd)) {
917 switch (info->type) {
922 case IND_OLDGEN: // rely on compatible layout with StgInd
923 case IND_OLDGEN_PERM:
924 // follow indirections
925 p = ((StgInd *)q)->indirectee;
930 return ((StgEvacuated *)q)->evacuee;
933 if (((StgTSO *)q)->what_next == ThreadRelocated) {
934 p = (StgClosure *)((StgTSO *)q)->link;
946 /* -----------------------------------------------------------------------------
947 Figure out which generation to collect, initialise N and major_gc.
949 Also returns the total number of blocks in generations that will be
951 -------------------------------------------------------------------------- */
954 initialise_N (rtsBool force_major_gc)
957 nat s, blocks, blocks_total;
962 if (force_major_gc) {
963 N = RtsFlags.GcFlags.generations - 1;
968 for (g = RtsFlags.GcFlags.generations - 1; g >= 0; g--) {
970 for (s = 0; s < generations[g].n_steps; s++) {
971 blocks += generations[g].steps[s].n_words / BLOCK_SIZE_W;
972 blocks += generations[g].steps[s].n_large_blocks;
974 if (blocks >= generations[g].max_blocks) {
978 blocks_total += blocks;
982 blocks_total += countNurseryBlocks();
984 major_gc = (N == RtsFlags.GcFlags.generations-1);
988 /* -----------------------------------------------------------------------------
989 Initialise the gc_thread structures.
990 -------------------------------------------------------------------------- */
993 alloc_gc_thread (int n)
999 t = stgMallocBytes(sizeof(gc_thread) + total_steps * sizeof(step_workspace),
1004 initCondition(&t->wake_cond);
1005 initMutex(&t->wake_mutex);
1006 t->wakeup = rtsTrue; // starts true, so we can wait for the
1007 // thread to start up, see wakeup_gc_threads
1011 t->thread_index = n;
1012 t->free_blocks = NULL;
1018 t->papi_events = -1;
1021 for (s = 0; s < total_steps; s++)
1024 ws->step = &all_steps[s];
1025 ASSERT(s == ws->step->abs_no);
1029 ws->buffer_todo_bd = NULL;
1031 ws->part_list = NULL;
1032 ws->n_part_blocks = 0;
1034 ws->scavd_list = NULL;
1035 ws->n_scavd_blocks = 0;
1043 alloc_gc_threads (void)
1045 if (gc_threads == NULL) {
1046 #if defined(THREADED_RTS)
1048 gc_threads = stgMallocBytes (RtsFlags.ParFlags.gcThreads *
1050 "alloc_gc_threads");
1052 for (i = 0; i < RtsFlags.ParFlags.gcThreads; i++) {
1053 gc_threads[i] = alloc_gc_thread(i);
1056 gc_threads = stgMallocBytes (sizeof(gc_thread*),
1057 "alloc_gc_threads");
1059 gc_threads[0] = alloc_gc_thread(0);
1064 /* ----------------------------------------------------------------------------
1066 ------------------------------------------------------------------------- */
1068 static nat gc_running_threads;
1070 #if defined(THREADED_RTS)
1071 static Mutex gc_running_mutex;
1078 ACQUIRE_LOCK(&gc_running_mutex);
1079 n_running = ++gc_running_threads;
1080 RELEASE_LOCK(&gc_running_mutex);
1081 ASSERT(n_running <= n_gc_threads);
1089 ACQUIRE_LOCK(&gc_running_mutex);
1090 ASSERT(n_gc_threads != 0);
1091 n_running = --gc_running_threads;
1092 RELEASE_LOCK(&gc_running_mutex);
1097 // gc_thread_work(): Scavenge until there's no work left to do and all
1098 // the running threads are idle.
1101 gc_thread_work (void)
1105 debugTrace(DEBUG_gc, "GC thread %d working", gct->thread_index);
1107 // gc_running_threads has already been incremented for us; either
1108 // this is the main thread and we incremented it inside
1109 // GarbageCollect(), or this is a worker thread and the main
1110 // thread bumped gc_running_threads before waking us up.
1112 // Every thread evacuates some roots.
1114 GetRoots(mark_root);
1118 // scavenge_loop() only exits when there's no work to do
1121 debugTrace(DEBUG_gc, "GC thread %d idle (%d still running)",
1122 gct->thread_index, r);
1124 while (gc_running_threads != 0) {
1130 // any_work() does not remove the work from the queue, it
1131 // just checks for the presence of work. If we find any,
1132 // then we increment gc_running_threads and go back to
1133 // scavenge_loop() to perform any pending work.
1136 // All threads are now stopped
1137 debugTrace(DEBUG_gc, "GC thread %d finished.", gct->thread_index);
1141 #if defined(THREADED_RTS)
1143 gc_thread_mainloop (void)
1145 while (!gct->exit) {
1147 // Wait until we're told to wake up
1148 ACQUIRE_LOCK(&gct->wake_mutex);
1149 gct->wakeup = rtsFalse;
1150 while (!gct->wakeup) {
1151 debugTrace(DEBUG_gc, "GC thread %d standing by...",
1153 waitCondition(&gct->wake_cond, &gct->wake_mutex);
1155 RELEASE_LOCK(&gct->wake_mutex);
1156 if (gct->exit) break;
1159 // start performance counters in this thread...
1160 if (gct->papi_events == -1) {
1161 papi_init_eventset(&gct->papi_events);
1163 papi_thread_start_gc1_count(gct->papi_events);
1169 // count events in this thread towards the GC totals
1170 papi_thread_stop_gc1_count(gct->papi_events);
1176 #if defined(THREADED_RTS)
1178 gc_thread_entry (gc_thread *my_gct)
1181 debugTrace(DEBUG_gc, "GC thread %d starting...", gct->thread_index);
1182 gct->id = osThreadId();
1183 gc_thread_mainloop();
1188 start_gc_threads (void)
1190 #if defined(THREADED_RTS)
1193 static rtsBool done = rtsFalse;
1195 gc_running_threads = 0;
1196 initMutex(&gc_running_mutex);
1199 // Start from 1: the main thread is 0
1200 for (i = 1; i < RtsFlags.ParFlags.gcThreads; i++) {
1201 createOSThread(&id, (OSThreadProc*)&gc_thread_entry,
1210 wakeup_gc_threads (nat n_threads USED_IF_THREADS)
1212 #if defined(THREADED_RTS)
1214 for (i=1; i < n_threads; i++) {
1216 debugTrace(DEBUG_gc, "waking up gc thread %d", i);
1218 ACQUIRE_LOCK(&gc_threads[i]->wake_mutex);
1219 if (gc_threads[i]->wakeup) {
1220 RELEASE_LOCK(&gc_threads[i]->wake_mutex);
1226 gc_threads[i]->wakeup = rtsTrue;
1227 signalCondition(&gc_threads[i]->wake_cond);
1228 RELEASE_LOCK(&gc_threads[i]->wake_mutex);
1233 // After GC is complete, we must wait for all GC threads to enter the
1234 // standby state, otherwise they may still be executing inside
1235 // any_work(), and may even remain awake until the next GC starts.
1237 shutdown_gc_threads (nat n_threads USED_IF_THREADS)
1239 #if defined(THREADED_RTS)
1242 for (i=1; i < n_threads; i++) {
1244 ACQUIRE_LOCK(&gc_threads[i]->wake_mutex);
1245 wakeup = gc_threads[i]->wakeup;
1246 // wakeup is false while the thread is waiting
1247 RELEASE_LOCK(&gc_threads[i]->wake_mutex);
1253 /* ----------------------------------------------------------------------------
1254 Initialise a generation that is to be collected
1255 ------------------------------------------------------------------------- */
1258 init_collected_gen (nat g, nat n_threads)
1265 // Throw away the current mutable list. Invariant: the mutable
1266 // list always has at least one block; this means we can avoid a
1267 // check for NULL in recordMutable().
1269 freeChain(generations[g].mut_list);
1270 generations[g].mut_list = allocBlock();
1271 for (i = 0; i < n_capabilities; i++) {
1272 freeChain(capabilities[i].mut_lists[g]);
1273 capabilities[i].mut_lists[g] = allocBlock();
1277 for (s = 0; s < generations[g].n_steps; s++) {
1279 // generation 0, step 0 doesn't need to-space
1280 if (g == 0 && s == 0 && RtsFlags.GcFlags.generations > 1) {
1284 stp = &generations[g].steps[s];
1285 ASSERT(stp->gen_no == g);
1287 // deprecate the existing blocks
1288 stp->old_blocks = stp->blocks;
1289 stp->n_old_blocks = stp->n_blocks;
1294 // we don't have any to-be-scavenged blocks yet
1296 stp->todos_last = NULL;
1299 // initialise the large object queues.
1300 stp->scavenged_large_objects = NULL;
1301 stp->n_scavenged_large_blocks = 0;
1303 // mark the large objects as not evacuated yet
1304 for (bd = stp->large_objects; bd; bd = bd->link) {
1305 bd->flags &= ~BF_EVACUATED;
1308 // for a compacted step, we need to allocate the bitmap
1309 if (stp->is_compacted) {
1310 nat bitmap_size; // in bytes
1311 bdescr *bitmap_bdescr;
1314 bitmap_size = stp->n_old_blocks * BLOCK_SIZE / (sizeof(W_)*BITS_PER_BYTE);
1316 if (bitmap_size > 0) {
1317 bitmap_bdescr = allocGroup((lnat)BLOCK_ROUND_UP(bitmap_size)
1319 stp->bitmap = bitmap_bdescr;
1320 bitmap = bitmap_bdescr->start;
1322 debugTrace(DEBUG_gc, "bitmap_size: %d, bitmap: %p",
1323 bitmap_size, bitmap);
1325 // don't forget to fill it with zeros!
1326 memset(bitmap, 0, bitmap_size);
1328 // For each block in this step, point to its bitmap from the
1329 // block descriptor.
1330 for (bd=stp->old_blocks; bd != NULL; bd = bd->link) {
1331 bd->u.bitmap = bitmap;
1332 bitmap += BLOCK_SIZE_W / (sizeof(W_)*BITS_PER_BYTE);
1334 // Also at this point we set the BF_COMPACTED flag
1335 // for this block. The invariant is that
1336 // BF_COMPACTED is always unset, except during GC
1337 // when it is set on those blocks which will be
1339 bd->flags |= BF_COMPACTED;
1345 // For each GC thread, for each step, allocate a "todo" block to
1346 // store evacuated objects to be scavenged, and a block to store
1347 // evacuated objects that do not need to be scavenged.
1348 for (t = 0; t < n_threads; t++) {
1349 for (s = 0; s < generations[g].n_steps; s++) {
1351 // we don't copy objects into g0s0, unless -G0
1352 if (g==0 && s==0 && RtsFlags.GcFlags.generations > 1) continue;
1354 ws = &gc_threads[t]->steps[g * RtsFlags.GcFlags.steps + s];
1356 ws->todo_large_objects = NULL;
1358 ws->part_list = NULL;
1359 ws->n_part_blocks = 0;
1361 // allocate the first to-space block; extra blocks will be
1362 // chained on as necessary.
1364 ws->buffer_todo_bd = NULL;
1365 alloc_todo_block(ws,0);
1367 ws->scavd_list = NULL;
1368 ws->n_scavd_blocks = 0;
1374 /* ----------------------------------------------------------------------------
1375 Initialise a generation that is *not* to be collected
1376 ------------------------------------------------------------------------- */
1379 init_uncollected_gen (nat g, nat threads)
1386 for (s = 0; s < generations[g].n_steps; s++) {
1387 stp = &generations[g].steps[s];
1388 stp->scavenged_large_objects = NULL;
1389 stp->n_scavenged_large_blocks = 0;
1392 for (t = 0; t < threads; t++) {
1393 for (s = 0; s < generations[g].n_steps; s++) {
1395 ws = &gc_threads[t]->steps[g * RtsFlags.GcFlags.steps + s];
1398 ws->buffer_todo_bd = NULL;
1399 ws->todo_large_objects = NULL;
1401 ws->part_list = NULL;
1402 ws->n_part_blocks = 0;
1404 ws->scavd_list = NULL;
1405 ws->n_scavd_blocks = 0;
1407 // If the block at the head of the list in this generation
1408 // is less than 3/4 full, then use it as a todo block.
1409 if (stp->blocks && isPartiallyFull(stp->blocks))
1411 ws->todo_bd = stp->blocks;
1412 ws->todo_free = ws->todo_bd->free;
1413 ws->todo_lim = ws->todo_bd->start + BLOCK_SIZE_W;
1414 stp->blocks = stp->blocks->link;
1416 stp->n_words -= ws->todo_bd->free - ws->todo_bd->start;
1417 ws->todo_bd->link = NULL;
1418 // we must scan from the current end point.
1419 ws->todo_bd->u.scan = ws->todo_bd->free;
1424 alloc_todo_block(ws,0);
1429 // Move the private mutable lists from each capability onto the
1430 // main mutable list for the generation.
1431 for (i = 0; i < n_capabilities; i++) {
1432 for (bd = capabilities[i].mut_lists[g];
1433 bd->link != NULL; bd = bd->link) {
1436 bd->link = generations[g].mut_list;
1437 generations[g].mut_list = capabilities[i].mut_lists[g];
1438 capabilities[i].mut_lists[g] = allocBlock();
1442 /* -----------------------------------------------------------------------------
1443 Initialise a gc_thread before GC
1444 -------------------------------------------------------------------------- */
1447 init_gc_thread (gc_thread *t)
1449 t->static_objects = END_OF_STATIC_LIST;
1450 t->scavenged_static_objects = END_OF_STATIC_LIST;
1453 t->failed_to_evac = rtsFalse;
1454 t->eager_promotion = rtsTrue;
1455 t->thunk_selector_depth = 0;
1460 t->scav_find_work = 0;
1463 /* -----------------------------------------------------------------------------
1464 Function we pass to GetRoots to evacuate roots.
1465 -------------------------------------------------------------------------- */
1468 mark_root(StgClosure **root)
1473 /* -----------------------------------------------------------------------------
1474 Initialising the static object & mutable lists
1475 -------------------------------------------------------------------------- */
1478 zero_static_object_list(StgClosure* first_static)
1482 const StgInfoTable *info;
1484 for (p = first_static; p != END_OF_STATIC_LIST; p = link) {
1486 link = *STATIC_LINK(info, p);
1487 *STATIC_LINK(info,p) = NULL;
1491 /* -----------------------------------------------------------------------------
1493 -------------------------------------------------------------------------- */
1500 for (c = (StgIndStatic *)revertible_caf_list; c != NULL;
1501 c = (StgIndStatic *)c->static_link)
1503 SET_INFO(c, c->saved_info);
1504 c->saved_info = NULL;
1505 // could, but not necessary: c->static_link = NULL;
1507 revertible_caf_list = NULL;
1511 markCAFs( evac_fn evac )
1515 for (c = (StgIndStatic *)caf_list; c != NULL;
1516 c = (StgIndStatic *)c->static_link)
1518 evac(&c->indirectee);
1520 for (c = (StgIndStatic *)revertible_caf_list; c != NULL;
1521 c = (StgIndStatic *)c->static_link)
1523 evac(&c->indirectee);
1527 /* ----------------------------------------------------------------------------
1528 Update the pointers from the task list
1530 These are treated as weak pointers because we want to allow a main
1531 thread to get a BlockedOnDeadMVar exception in the same way as any
1532 other thread. Note that the threads should all have been retained
1533 by GC by virtue of being on the all_threads list, we're just
1534 updating pointers here.
1535 ------------------------------------------------------------------------- */
1538 update_task_list (void)
1542 for (task = all_tasks; task != NULL; task = task->all_link) {
1543 if (!task->stopped && task->tso) {
1544 ASSERT(task->tso->bound == task);
1545 tso = (StgTSO *) isAlive((StgClosure *)task->tso);
1547 barf("task %p: main thread %d has been GC'd",
1560 /* ----------------------------------------------------------------------------
1561 Reset the sizes of the older generations when we do a major
1564 CURRENT STRATEGY: make all generations except zero the same size.
1565 We have to stay within the maximum heap size, and leave a certain
1566 percentage of the maximum heap size available to allocate into.
1567 ------------------------------------------------------------------------- */
1570 resize_generations (void)
1574 if (major_gc && RtsFlags.GcFlags.generations > 1) {
1575 nat live, size, min_alloc;
1576 nat max = RtsFlags.GcFlags.maxHeapSize;
1577 nat gens = RtsFlags.GcFlags.generations;
1579 // live in the oldest generations
1580 live = (oldest_gen->steps[0].n_words + BLOCK_SIZE_W - 1) / BLOCK_SIZE_W+
1581 oldest_gen->steps[0].n_large_blocks;
1583 // default max size for all generations except zero
1584 size = stg_max(live * RtsFlags.GcFlags.oldGenFactor,
1585 RtsFlags.GcFlags.minOldGenSize);
1587 // minimum size for generation zero
1588 min_alloc = stg_max((RtsFlags.GcFlags.pcFreeHeap * max) / 200,
1589 RtsFlags.GcFlags.minAllocAreaSize);
1591 // Auto-enable compaction when the residency reaches a
1592 // certain percentage of the maximum heap size (default: 30%).
1593 if (RtsFlags.GcFlags.generations > 1 &&
1594 (RtsFlags.GcFlags.compact ||
1596 oldest_gen->steps[0].n_blocks >
1597 (RtsFlags.GcFlags.compactThreshold * max) / 100))) {
1598 oldest_gen->steps[0].is_compacted = 1;
1599 // debugBelch("compaction: on\n", live);
1601 oldest_gen->steps[0].is_compacted = 0;
1602 // debugBelch("compaction: off\n", live);
1605 // if we're going to go over the maximum heap size, reduce the
1606 // size of the generations accordingly. The calculation is
1607 // different if compaction is turned on, because we don't need
1608 // to double the space required to collect the old generation.
1611 // this test is necessary to ensure that the calculations
1612 // below don't have any negative results - we're working
1613 // with unsigned values here.
1614 if (max < min_alloc) {
1618 if (oldest_gen->steps[0].is_compacted) {
1619 if ( (size + (size - 1) * (gens - 2) * 2) + min_alloc > max ) {
1620 size = (max - min_alloc) / ((gens - 1) * 2 - 1);
1623 if ( (size * (gens - 1) * 2) + min_alloc > max ) {
1624 size = (max - min_alloc) / ((gens - 1) * 2);
1634 debugBelch("live: %d, min_alloc: %d, size : %d, max = %d\n", live,
1635 min_alloc, size, max);
1638 for (g = 0; g < gens; g++) {
1639 generations[g].max_blocks = size;
1644 /* -----------------------------------------------------------------------------
1645 Calculate the new size of the nursery, and resize it.
1646 -------------------------------------------------------------------------- */
1649 resize_nursery (void)
1651 if (RtsFlags.GcFlags.generations == 1)
1652 { // Two-space collector:
1655 /* set up a new nursery. Allocate a nursery size based on a
1656 * function of the amount of live data (by default a factor of 2)
1657 * Use the blocks from the old nursery if possible, freeing up any
1660 * If we get near the maximum heap size, then adjust our nursery
1661 * size accordingly. If the nursery is the same size as the live
1662 * data (L), then we need 3L bytes. We can reduce the size of the
1663 * nursery to bring the required memory down near 2L bytes.
1665 * A normal 2-space collector would need 4L bytes to give the same
1666 * performance we get from 3L bytes, reducing to the same
1667 * performance at 2L bytes.
1669 blocks = g0s0->n_old_blocks;
1671 if ( RtsFlags.GcFlags.maxHeapSize != 0 &&
1672 blocks * RtsFlags.GcFlags.oldGenFactor * 2 >
1673 RtsFlags.GcFlags.maxHeapSize )
1675 long adjusted_blocks; // signed on purpose
1678 adjusted_blocks = (RtsFlags.GcFlags.maxHeapSize - 2 * blocks);
1680 debugTrace(DEBUG_gc, "near maximum heap size of 0x%x blocks, blocks = %d, adjusted to %ld",
1681 RtsFlags.GcFlags.maxHeapSize, blocks, adjusted_blocks);
1683 pc_free = adjusted_blocks * 100 / RtsFlags.GcFlags.maxHeapSize;
1684 if (pc_free < RtsFlags.GcFlags.pcFreeHeap) /* might even * be < 0 */
1688 blocks = adjusted_blocks;
1692 blocks *= RtsFlags.GcFlags.oldGenFactor;
1693 if (blocks < RtsFlags.GcFlags.minAllocAreaSize)
1695 blocks = RtsFlags.GcFlags.minAllocAreaSize;
1698 resizeNurseries(blocks);
1700 else // Generational collector
1703 * If the user has given us a suggested heap size, adjust our
1704 * allocation area to make best use of the memory available.
1706 if (RtsFlags.GcFlags.heapSizeSuggestion)
1709 nat needed = calcNeeded(); // approx blocks needed at next GC
1711 /* Guess how much will be live in generation 0 step 0 next time.
1712 * A good approximation is obtained by finding the
1713 * percentage of g0s0 that was live at the last minor GC.
1715 * We have an accurate figure for the amount of copied data in
1716 * 'copied', but we must convert this to a number of blocks, with
1717 * a small adjustment for estimated slop at the end of a block
1722 g0s0_pcnt_kept = ((copied / (BLOCK_SIZE_W - 10)) * 100)
1723 / countNurseryBlocks();
1726 /* Estimate a size for the allocation area based on the
1727 * information available. We might end up going slightly under
1728 * or over the suggested heap size, but we should be pretty
1731 * Formula: suggested - needed
1732 * ----------------------------
1733 * 1 + g0s0_pcnt_kept/100
1735 * where 'needed' is the amount of memory needed at the next
1736 * collection for collecting all steps except g0s0.
1739 (((long)RtsFlags.GcFlags.heapSizeSuggestion - (long)needed) * 100) /
1740 (100 + (long)g0s0_pcnt_kept);
1742 if (blocks < (long)RtsFlags.GcFlags.minAllocAreaSize) {
1743 blocks = RtsFlags.GcFlags.minAllocAreaSize;
1746 resizeNurseries((nat)blocks);
1750 // we might have added extra large blocks to the nursery, so
1751 // resize back to minAllocAreaSize again.
1752 resizeNurseriesFixed(RtsFlags.GcFlags.minAllocAreaSize);
1757 /* -----------------------------------------------------------------------------
1758 Sanity code for CAF garbage collection.
1760 With DEBUG turned on, we manage a CAF list in addition to the SRT
1761 mechanism. After GC, we run down the CAF list and blackhole any
1762 CAFs which have been garbage collected. This means we get an error
1763 whenever the program tries to enter a garbage collected CAF.
1765 Any garbage collected CAFs are taken off the CAF list at the same
1767 -------------------------------------------------------------------------- */
1769 #if 0 && defined(DEBUG)
1776 const StgInfoTable *info;
1787 ASSERT(info->type == IND_STATIC);
1789 if (STATIC_LINK(info,p) == NULL) {
1790 debugTrace(DEBUG_gccafs, "CAF gc'd at 0x%04lx", (long)p);
1792 SET_INFO(p,&stg_BLACKHOLE_info);
1793 p = STATIC_LINK2(info,p);
1797 pp = &STATIC_LINK2(info,p);
1804 debugTrace(DEBUG_gccafs, "%d CAFs live", i);