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()
55 /* -----------------------------------------------------------------------------
57 -------------------------------------------------------------------------- */
59 /* STATIC OBJECT LIST.
62 * We maintain a linked list of static objects that are still live.
63 * The requirements for this list are:
65 * - we need to scan the list while adding to it, in order to
66 * scavenge all the static objects (in the same way that
67 * breadth-first scavenging works for dynamic objects).
69 * - we need to be able to tell whether an object is already on
70 * the list, to break loops.
72 * Each static object has a "static link field", which we use for
73 * linking objects on to the list. We use a stack-type list, consing
74 * objects on the front as they are added (this means that the
75 * scavenge phase is depth-first, not breadth-first, but that
78 * A separate list is kept for objects that have been scavenged
79 * already - this is so that we can zero all the marks afterwards.
81 * An object is on the list if its static link field is non-zero; this
82 * means that we have to mark the end of the list with '1', not NULL.
84 * Extra notes for generational GC:
86 * Each generation has a static object list associated with it. When
87 * collecting generations up to N, we treat the static object lists
88 * from generations > N as roots.
90 * We build up a static object list while collecting generations 0..N,
91 * which is then appended to the static object list of generation N+1.
93 StgClosure* static_objects; // live static objects
94 StgClosure* scavenged_static_objects; // static objects scavenged so far
96 SpinLock static_objects_sync;
99 /* N is the oldest generation being collected, where the generations
100 * are numbered starting at 0. A major GC (indicated by the major_gc
101 * flag) is when we're collecting all generations. We only attempt to
102 * deal with static objects and GC CAFs when doing a major GC.
107 /* Data used for allocation area sizing.
109 static lnat g0s0_pcnt_kept = 30; // percentage of g0s0 live at last minor GC
119 /* Thread-local data for each GC thread
121 gc_thread *gc_threads = NULL;
122 // gc_thread *gct = NULL; // this thread's gct TODO: make thread-local
124 // Number of threads running in *this* GC. Affects how many
125 // step->todos[] lists we have to look in to find work.
129 long copied; // *words* copied & scavenged during this GC
132 SpinLock recordMutableGen_sync;
135 /* -----------------------------------------------------------------------------
136 Static function declarations
137 -------------------------------------------------------------------------- */
139 static void mark_root (StgClosure **root);
140 static void zero_static_object_list (StgClosure* first_static);
141 static void initialise_N (rtsBool force_major_gc);
142 static void alloc_gc_threads (void);
143 static void init_collected_gen (nat g, nat threads);
144 static void init_uncollected_gen (nat g, nat threads);
145 static void init_gc_thread (gc_thread *t);
146 static void update_task_list (void);
147 static void resize_generations (void);
148 static void resize_nursery (void);
149 static void start_gc_threads (void);
150 static void gc_thread_work (void);
151 static nat inc_running (void);
152 static nat dec_running (void);
153 static void wakeup_gc_threads (nat n_threads);
155 #if 0 && defined(DEBUG)
156 static void gcCAFs (void);
159 /* -----------------------------------------------------------------------------
160 The mark bitmap & stack.
161 -------------------------------------------------------------------------- */
163 #define MARK_STACK_BLOCKS 4
165 bdescr *mark_stack_bdescr;
170 // Flag and pointers used for falling back to a linear scan when the
171 // mark stack overflows.
172 rtsBool mark_stack_overflowed;
173 bdescr *oldgen_scan_bd;
176 /* -----------------------------------------------------------------------------
177 GarbageCollect: the main entry point to the garbage collector.
179 Locks held: all capabilities are held throughout GarbageCollect().
180 -------------------------------------------------------------------------- */
183 GarbageCollect ( rtsBool force_major_gc )
187 lnat live, allocated;
188 lnat oldgen_saved_blocks = 0;
189 gc_thread *saved_gct;
192 // necessary if we stole a callee-saves register for gct:
196 CostCentreStack *prev_CCS;
201 debugTrace(DEBUG_gc, "starting GC");
203 #if defined(RTS_USER_SIGNALS)
204 if (RtsFlags.MiscFlags.install_signal_handlers) {
210 // tell the STM to discard any cached closures it's hoping to re-use
213 // tell the stats department that we've started a GC
217 // check for memory leaks if DEBUG is on
227 // attribute any costs to CCS_GC
233 /* Approximate how much we allocated.
234 * Todo: only when generating stats?
236 allocated = calcAllocated();
238 /* Figure out which generation to collect
240 initialise_N(force_major_gc);
242 /* Allocate + initialise the gc_thread structures.
246 /* Start threads, so they can be spinning up while we finish initialisation.
250 /* How many threads will be participating in this GC?
251 * We don't try to parallelise minor GC.
253 #if defined(THREADED_RTS)
257 n_gc_threads = RtsFlags.ParFlags.gcThreads;
263 #ifdef RTS_GTK_FRONTPANEL
264 if (RtsFlags.GcFlags.frontpanel) {
265 updateFrontPanelBeforeGC(N);
269 // check stack sanity *before* GC (ToDo: check all threads)
270 IF_DEBUG(sanity, checkFreeListSanity());
272 /* Initialise the static object lists
274 static_objects = END_OF_STATIC_LIST;
275 scavenged_static_objects = END_OF_STATIC_LIST;
278 initSpinLock(&static_objects_sync);
279 initSpinLock(&recordMutableGen_sync);
280 initSpinLock(&gc_alloc_block_sync);
283 // Initialise all the generations/steps that we're collecting.
284 for (g = 0; g <= N; g++) {
285 init_collected_gen(g,n_gc_threads);
288 // Initialise all the generations/steps that we're *not* collecting.
289 for (g = N+1; g < RtsFlags.GcFlags.generations; g++) {
290 init_uncollected_gen(g,n_gc_threads);
293 /* Allocate a mark stack if we're doing a major collection.
296 mark_stack_bdescr = allocGroup(MARK_STACK_BLOCKS);
297 mark_stack = (StgPtr *)mark_stack_bdescr->start;
298 mark_sp = mark_stack;
299 mark_splim = mark_stack + (MARK_STACK_BLOCKS * BLOCK_SIZE_W);
301 mark_stack_bdescr = NULL;
304 // Initialise all our gc_thread structures
305 for (t = 0; t < n_gc_threads; t++) {
306 init_gc_thread(&gc_threads[t]);
309 // the main thread is running: this prevents any other threads from
310 // exiting prematurely, so we can start them now.
312 wakeup_gc_threads(n_gc_threads);
317 // this is the main thread
318 gct = &gc_threads[0];
320 /* -----------------------------------------------------------------------
321 * follow all the roots that we know about:
322 * - mutable lists from each generation > N
323 * we want to *scavenge* these roots, not evacuate them: they're not
324 * going to move in this GC.
325 * Also do them in reverse generation order, for the usual reason:
326 * namely to reduce the likelihood of spurious old->new pointers.
329 for (g = RtsFlags.GcFlags.generations-1; g > N; g--) {
330 generations[g].saved_mut_list = generations[g].mut_list;
331 generations[g].mut_list = allocBlock();
332 // mut_list always has at least one block.
334 for (g = RtsFlags.GcFlags.generations-1; g > N; g--) {
335 scavenge_mutable_list(&generations[g]);
339 // follow roots from the CAF list (used by GHCi)
343 // follow all the roots that the application knows about.
347 #if defined(RTS_USER_SIGNALS)
348 // mark the signal handlers (signals should be already blocked)
349 markSignalHandlers(mark_root);
352 // Mark the weak pointer list, and prepare to detect dead weak pointers.
356 // Mark the stable pointer table.
357 markStablePtrTable(mark_root);
359 /* -------------------------------------------------------------------------
360 * Repeatedly scavenge all the areas we know about until there's no
361 * more scavenging to be done.
366 // The other threads are now stopped. We might recurse back to
367 // here, but from now on this is the only thread.
369 // if any blackholes are alive, make the threads that wait on
371 if (traverseBlackholeQueue()) {
376 // must be last... invariant is that everything is fully
377 // scavenged at this point.
378 if (traverseWeakPtrList()) { // returns rtsTrue if evaced something
383 // If we get to here, there's really nothing left to do.
387 // Update pointers from the Task list
390 // Now see which stable names are still alive.
394 // We call processHeapClosureForDead() on every closure destroyed during
395 // the current garbage collection, so we invoke LdvCensusForDead().
396 if (RtsFlags.ProfFlags.doHeapProfile == HEAP_BY_LDV
397 || RtsFlags.ProfFlags.bioSelector != NULL)
401 // NO MORE EVACUATION AFTER THIS POINT!
402 // Finally: compaction of the oldest generation.
403 if (major_gc && oldest_gen->steps[0].is_compacted) {
404 // save number of blocks for stats
405 oldgen_saved_blocks = oldest_gen->steps[0].n_old_blocks;
409 IF_DEBUG(sanity, checkGlobalTSOList(rtsFalse));
411 // Two-space collector: free the old to-space.
412 // g0s0->old_blocks is the old nursery
413 // g0s0->blocks is to-space from the previous GC
414 if (RtsFlags.GcFlags.generations == 1) {
415 if (g0s0->blocks != NULL) {
416 freeChain(g0s0->blocks);
421 // For each workspace, in each thread:
422 // * clear the BF_EVACUATED flag from each copied block
423 // * move the copied blocks to the step
429 for (t = 0; t < n_gc_threads; t++) {
430 thr = &gc_threads[t];
432 for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
433 for (s = 0; s < generations[g].n_steps; s++) {
434 ws = &thr->steps[g][s];
435 if (g==0 && s==0) continue;
438 // ASSERT( ws->scan_bd == ws->todo_bd );
439 ASSERT( ws->scan_bd ? ws->scan == ws->scan_bd->free : 1 );
441 // Push the final block
442 if (ws->scan_bd) { push_scan_block(ws->scan_bd, ws); }
444 ASSERT(countBlocks(ws->scavd_list) == ws->n_scavd_blocks);
446 prev = ws->scavd_list;
447 for (bd = ws->scavd_list; bd != NULL; bd = bd->link) {
448 bd->flags &= ~BF_EVACUATED; // now from-space
451 prev->link = ws->stp->blocks;
452 ws->stp->blocks = ws->scavd_list;
453 ws->stp->n_blocks += ws->n_scavd_blocks;
454 ASSERT(countBlocks(ws->stp->blocks) == ws->stp->n_blocks);
460 // Two-space collector: swap the semi-spaces around.
461 // Currently: g0s0->old_blocks is the old nursery
462 // g0s0->blocks is to-space from this GC
463 // We want these the other way around.
464 if (RtsFlags.GcFlags.generations == 1) {
465 bdescr *nursery_blocks = g0s0->old_blocks;
466 nat n_nursery_blocks = g0s0->n_old_blocks;
467 g0s0->old_blocks = g0s0->blocks;
468 g0s0->n_old_blocks = g0s0->n_blocks;
469 g0s0->blocks = nursery_blocks;
470 g0s0->n_blocks = n_nursery_blocks;
473 /* run through all the generations/steps and tidy up
475 for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
478 generations[g].collections++; // for stats
481 // Count the mutable list as bytes "copied" for the purposes of
482 // stats. Every mutable list is copied during every GC.
484 nat mut_list_size = 0;
485 for (bd = generations[g].mut_list; bd != NULL; bd = bd->link) {
486 mut_list_size += bd->free - bd->start;
488 copied += mut_list_size;
491 "mut_list_size: %lu (%d vars, %d arrays, %d MVARs, %d others)",
492 (unsigned long)(mut_list_size * sizeof(W_)),
493 mutlist_MUTVARS, mutlist_MUTARRS, mutlist_MVARS, mutlist_OTHERS);
496 for (s = 0; s < generations[g].n_steps; s++) {
498 stp = &generations[g].steps[s];
500 // for generations we collected...
503 /* free old memory and shift to-space into from-space for all
504 * the collected steps (except the allocation area). These
505 * freed blocks will probaby be quickly recycled.
507 if (!(g == 0 && s == 0 && RtsFlags.GcFlags.generations > 1)) {
508 if (stp->is_compacted)
510 // for a compacted step, just shift the new to-space
511 // onto the front of the now-compacted existing blocks.
512 for (bd = stp->blocks; bd != NULL; bd = bd->link) {
513 bd->flags &= ~BF_EVACUATED; // now from-space
515 // tack the new blocks on the end of the existing blocks
516 if (stp->old_blocks != NULL) {
517 for (bd = stp->old_blocks; bd != NULL; bd = next) {
518 // NB. this step might not be compacted next
519 // time, so reset the BF_COMPACTED flags.
520 // They are set before GC if we're going to
521 // compact. (search for BF_COMPACTED above).
522 bd->flags &= ~BF_COMPACTED;
525 bd->link = stp->blocks;
528 stp->blocks = stp->old_blocks;
530 // add the new blocks to the block tally
531 stp->n_blocks += stp->n_old_blocks;
532 ASSERT(countBlocks(stp->blocks) == stp->n_blocks);
536 freeChain(stp->old_blocks);
538 stp->old_blocks = NULL;
539 stp->n_old_blocks = 0;
542 /* LARGE OBJECTS. The current live large objects are chained on
543 * scavenged_large, having been moved during garbage
544 * collection from large_objects. Any objects left on
545 * large_objects list are therefore dead, so we free them here.
547 for (bd = stp->large_objects; bd != NULL; bd = next) {
553 // update the count of blocks used by large objects
554 for (bd = stp->scavenged_large_objects; bd != NULL; bd = bd->link) {
555 bd->flags &= ~BF_EVACUATED;
557 stp->large_objects = stp->scavenged_large_objects;
558 stp->n_large_blocks = stp->n_scavenged_large_blocks;
561 else // for older generations...
563 /* For older generations, we need to append the
564 * scavenged_large_object list (i.e. large objects that have been
565 * promoted during this GC) to the large_object list for that step.
567 for (bd = stp->scavenged_large_objects; bd; bd = next) {
569 bd->flags &= ~BF_EVACUATED;
570 dbl_link_onto(bd, &stp->large_objects);
573 // add the new blocks we promoted during this GC
574 stp->n_large_blocks += stp->n_scavenged_large_blocks;
579 // update the max size of older generations after a major GC
580 resize_generations();
582 // Guess the amount of live data for stats.
583 live = calcLiveBlocks() * BLOCK_SIZE_W;
584 debugTrace(DEBUG_gc, "Slop: %ldKB",
585 (live - calcLiveWords()) / (1024/sizeof(W_)));
587 // Free the small objects allocated via allocate(), since this will
588 // all have been copied into G0S1 now.
589 if (RtsFlags.GcFlags.generations > 1) {
590 if (g0s0->blocks != NULL) {
591 freeChain(g0s0->blocks);
597 alloc_blocks_lim = RtsFlags.GcFlags.minAllocAreaSize;
599 // Start a new pinned_object_block
600 pinned_object_block = NULL;
602 // Free the mark stack.
603 if (mark_stack_bdescr != NULL) {
604 freeGroup(mark_stack_bdescr);
608 for (g = 0; g <= N; g++) {
609 for (s = 0; s < generations[g].n_steps; s++) {
610 stp = &generations[g].steps[s];
611 if (stp->bitmap != NULL) {
612 freeGroup(stp->bitmap);
620 // mark the garbage collected CAFs as dead
621 #if 0 && defined(DEBUG) // doesn't work at the moment
622 if (major_gc) { gcCAFs(); }
626 // resetStaticObjectForRetainerProfiling() must be called before
628 resetStaticObjectForRetainerProfiling();
631 // zero the scavenged static object list
633 zero_static_object_list(scavenged_static_objects);
639 // start any pending finalizers
641 scheduleFinalizers(last_free_capability, old_weak_ptr_list);
644 // send exceptions to any threads which were about to die
646 resurrectThreads(resurrected_threads);
649 // Update the stable pointer hash table.
650 updateStablePtrTable(major_gc);
652 // check sanity after GC
653 IF_DEBUG(sanity, checkSanity());
655 // extra GC trace info
656 IF_DEBUG(gc, statDescribeGens());
659 // symbol-table based profiling
660 /* heapCensus(to_blocks); */ /* ToDo */
663 // restore enclosing cost centre
669 // check for memory leaks if DEBUG is on
673 #ifdef RTS_GTK_FRONTPANEL
674 if (RtsFlags.GcFlags.frontpanel) {
675 updateFrontPanelAfterGC( N, live );
679 // ok, GC over: tell the stats department what happened.
680 stat_endGC(allocated, live, copied, N);
682 #if defined(RTS_USER_SIGNALS)
683 if (RtsFlags.MiscFlags.install_signal_handlers) {
684 // unblock signals again
685 unblockUserSignals();
694 /* -----------------------------------------------------------------------------
695 * Mark all nodes pointed to by sparks in the spark queues (for GC) Does an
696 * implicit slide i.e. after marking all sparks are at the beginning of the
697 * spark pool and the spark pool only contains sparkable closures
698 * -------------------------------------------------------------------------- */
702 markSparkQueue (evac_fn evac, Capability *cap)
704 StgClosure **sparkp, **to_sparkp;
705 nat n, pruned_sparks; // stats only
708 PAR_TICKY_MARK_SPARK_QUEUE_START();
713 pool = &(cap->r.rSparks);
715 ASSERT_SPARK_POOL_INVARIANTS(pool);
717 #if defined(PARALLEL_HASKELL)
724 to_sparkp = pool->hd;
725 while (sparkp != pool->tl) {
726 ASSERT(*sparkp!=NULL);
727 ASSERT(LOOKS_LIKE_CLOSURE_PTR(((StgClosure *)*sparkp)));
728 // ToDo?: statistics gathering here (also for GUM!)
729 if (closure_SHOULD_SPARK(*sparkp)) {
731 *to_sparkp++ = *sparkp;
732 if (to_sparkp == pool->lim) {
733 to_sparkp = pool->base;
740 if (sparkp == pool->lim) {
744 pool->tl = to_sparkp;
746 PAR_TICKY_MARK_SPARK_QUEUE_END(n);
748 #if defined(PARALLEL_HASKELL)
749 debugTrace(DEBUG_sched,
750 "marked %d sparks and pruned %d sparks on [%x]",
751 n, pruned_sparks, mytid);
753 debugTrace(DEBUG_sched,
754 "marked %d sparks and pruned %d sparks",
758 debugTrace(DEBUG_sched,
759 "new spark queue len=%d; (hd=%p; tl=%p)\n",
760 sparkPoolSize(pool), pool->hd, pool->tl);
764 /* ---------------------------------------------------------------------------
765 Where are the roots that we know about?
767 - all the threads on the runnable queue
768 - all the threads on the blocked queue
769 - all the threads on the sleeping queue
770 - all the thread currently executing a _ccall_GC
771 - all the "main threads"
773 ------------------------------------------------------------------------ */
776 GetRoots( evac_fn evac )
782 // Each GC thread is responsible for following roots from the
783 // Capability of the same number. There will usually be the same
784 // or fewer Capabilities as GC threads, but just in case there
785 // are more, we mark every Capability whose number is the GC
786 // thread's index plus a multiple of the number of GC threads.
787 for (i = gct->thread_index; i < n_capabilities; i += n_gc_threads) {
788 cap = &capabilities[i];
789 evac((StgClosure **)(void *)&cap->run_queue_hd);
790 evac((StgClosure **)(void *)&cap->run_queue_tl);
791 #if defined(THREADED_RTS)
792 evac((StgClosure **)(void *)&cap->wakeup_queue_hd);
793 evac((StgClosure **)(void *)&cap->wakeup_queue_tl);
795 for (task = cap->suspended_ccalling_tasks; task != NULL;
797 debugTrace(DEBUG_sched,
798 "evac'ing suspended TSO %lu", (unsigned long)task->suspended_tso->id);
799 evac((StgClosure **)(void *)&task->suspended_tso);
802 #if defined(THREADED_RTS)
803 markSparkQueue(evac,cap);
807 #if !defined(THREADED_RTS)
808 evac((StgClosure **)(void *)&blocked_queue_hd);
809 evac((StgClosure **)(void *)&blocked_queue_tl);
810 evac((StgClosure **)(void *)&sleeping_queue);
814 /* -----------------------------------------------------------------------------
815 isAlive determines whether the given closure is still alive (after
816 a garbage collection) or not. It returns the new address of the
817 closure if it is alive, or NULL otherwise.
819 NOTE: Use it before compaction only!
820 It untags and (if needed) retags pointers to closures.
821 -------------------------------------------------------------------------- */
825 isAlive(StgClosure *p)
827 const StgInfoTable *info;
833 /* The tag and the pointer are split, to be merged later when needed. */
834 tag = GET_CLOSURE_TAG(p);
835 q = UNTAG_CLOSURE(p);
837 ASSERT(LOOKS_LIKE_CLOSURE_PTR(q));
840 // ignore static closures
842 // ToDo: for static closures, check the static link field.
843 // Problem here is that we sometimes don't set the link field, eg.
844 // for static closures with an empty SRT or CONSTR_STATIC_NOCAFs.
846 if (!HEAP_ALLOCED(q)) {
850 // ignore closures in generations that we're not collecting.
852 if (bd->gen_no > N) {
856 // if it's a pointer into to-space, then we're done
857 if (bd->flags & BF_EVACUATED) {
861 // large objects use the evacuated flag
862 if (bd->flags & BF_LARGE) {
866 // check the mark bit for compacted steps
867 if ((bd->flags & BF_COMPACTED) && is_marked((P_)q,bd)) {
871 switch (info->type) {
876 case IND_OLDGEN: // rely on compatible layout with StgInd
877 case IND_OLDGEN_PERM:
878 // follow indirections
879 p = ((StgInd *)q)->indirectee;
884 return ((StgEvacuated *)q)->evacuee;
887 if (((StgTSO *)q)->what_next == ThreadRelocated) {
888 p = (StgClosure *)((StgTSO *)q)->link;
900 /* -----------------------------------------------------------------------------
901 Figure out which generation to collect, initialise N and major_gc.
902 -------------------------------------------------------------------------- */
905 initialise_N (rtsBool force_major_gc)
909 if (force_major_gc) {
910 N = RtsFlags.GcFlags.generations - 1;
914 for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
915 if (generations[g].steps[0].n_blocks +
916 generations[g].steps[0].n_large_blocks
917 >= generations[g].max_blocks) {
921 major_gc = (N == RtsFlags.GcFlags.generations-1);
925 /* -----------------------------------------------------------------------------
926 Initialise the gc_thread structures.
927 -------------------------------------------------------------------------- */
930 alloc_gc_thread (gc_thread *t, int n)
937 initCondition(&t->wake_cond);
938 initMutex(&t->wake_mutex);
939 t->wakeup = rtsFalse;
944 t->free_blocks = NULL;
953 t->steps = stgMallocBytes(RtsFlags.GcFlags.generations *
954 sizeof(step_workspace *),
955 "initialise_gc_thread");
957 for (g = 0; g < RtsFlags.GcFlags.generations; g++)
959 t->steps[g] = stgMallocBytes(generations[g].n_steps *
960 sizeof(step_workspace),
961 "initialise_gc_thread/2");
963 for (s = 0; s < generations[g].n_steps; s++)
965 ws = &t->steps[g][s];
966 ws->stp = &generations[g].steps[s];
973 ws->buffer_todo_bd = NULL;
975 ws->scavd_list = NULL;
976 ws->n_scavd_blocks = 0;
983 alloc_gc_threads (void)
985 if (gc_threads == NULL) {
986 #if defined(THREADED_RTS)
988 gc_threads = stgMallocBytes (RtsFlags.ParFlags.gcThreads *
992 for (i = 0; i < RtsFlags.ParFlags.gcThreads; i++) {
993 alloc_gc_thread(&gc_threads[i], i);
996 gc_threads = stgMallocBytes (sizeof(gc_thread),
999 alloc_gc_thread(gc_threads, 0);
1004 /* ----------------------------------------------------------------------------
1006 ------------------------------------------------------------------------- */
1008 static nat gc_running_threads;
1010 #if defined(THREADED_RTS)
1011 static Mutex gc_running_mutex;
1018 ACQUIRE_LOCK(&gc_running_mutex);
1019 n_running = ++gc_running_threads;
1020 RELEASE_LOCK(&gc_running_mutex);
1028 ACQUIRE_LOCK(&gc_running_mutex);
1029 n_running = --gc_running_threads;
1030 RELEASE_LOCK(&gc_running_mutex);
1035 // gc_thread_work(): Scavenge until there's no work left to do and all
1036 // the running threads are idle.
1039 gc_thread_work (void)
1043 debugTrace(DEBUG_gc, "GC thread %d working", gct->thread_index);
1045 // gc_running_threads has already been incremented for us; either
1046 // this is the main thread and we incremented it inside
1047 // GarbageCollect(), or this is a worker thread and the main
1048 // thread bumped gc_running_threads before waking us up.
1050 // Every thread evacuates some roots.
1052 GetRoots(mark_root);
1056 // scavenge_loop() only exits when there's no work to do
1059 debugTrace(DEBUG_gc, "GC thread %d idle (%d still running)",
1060 gct->thread_index, r);
1062 while (gc_running_threads != 0) {
1067 // any_work() does not remove the work from the queue, it
1068 // just checks for the presence of work. If we find any,
1069 // then we increment gc_running_threads and go back to
1070 // scavenge_loop() to perform any pending work.
1073 // All threads are now stopped
1074 debugTrace(DEBUG_gc, "GC thread %d finished.", gct->thread_index);
1078 #if defined(THREADED_RTS)
1080 gc_thread_mainloop (void)
1082 while (!gct->exit) {
1084 // Wait until we're told to wake up
1085 ACQUIRE_LOCK(&gct->wake_mutex);
1086 while (!gct->wakeup) {
1087 debugTrace(DEBUG_gc, "GC thread %d standing by...",
1089 waitCondition(&gct->wake_cond, &gct->wake_mutex);
1091 RELEASE_LOCK(&gct->wake_mutex);
1092 gct->wakeup = rtsFalse;
1093 if (gct->exit) break;
1096 // start performance counters in this thread...
1097 if (gct->papi_events == -1) {
1098 papi_init_eventset(&gct->papi_events);
1100 papi_thread_start_gc1_count(gct->papi_events);
1106 // count events in this thread towards the GC totals
1107 papi_thread_stop_gc1_count(gct->papi_events);
1113 #if defined(THREADED_RTS)
1115 gc_thread_entry (gc_thread *my_gct)
1118 debugTrace(DEBUG_gc, "GC thread %d starting...", gct->thread_index);
1119 gct->id = osThreadId();
1120 gc_thread_mainloop();
1125 start_gc_threads (void)
1127 #if defined(THREADED_RTS)
1130 static rtsBool done = rtsFalse;
1132 gc_running_threads = 0;
1133 initMutex(&gc_running_mutex);
1136 // Start from 1: the main thread is 0
1137 for (i = 1; i < RtsFlags.ParFlags.gcThreads; i++) {
1138 createOSThread(&id, (OSThreadProc*)&gc_thread_entry,
1147 wakeup_gc_threads (nat n_threads USED_IF_THREADS)
1149 #if defined(THREADED_RTS)
1151 for (i=1; i < n_threads; i++) {
1153 ACQUIRE_LOCK(&gc_threads[i].wake_mutex);
1154 gc_threads[i].wakeup = rtsTrue;
1155 signalCondition(&gc_threads[i].wake_cond);
1156 RELEASE_LOCK(&gc_threads[i].wake_mutex);
1161 /* ----------------------------------------------------------------------------
1162 Initialise a generation that is to be collected
1163 ------------------------------------------------------------------------- */
1166 init_collected_gen (nat g, nat n_threads)
1173 // Throw away the current mutable list. Invariant: the mutable
1174 // list always has at least one block; this means we can avoid a
1175 // check for NULL in recordMutable().
1177 freeChain(generations[g].mut_list);
1178 generations[g].mut_list = allocBlock();
1179 for (i = 0; i < n_capabilities; i++) {
1180 freeChain(capabilities[i].mut_lists[g]);
1181 capabilities[i].mut_lists[g] = allocBlock();
1185 for (s = 0; s < generations[g].n_steps; s++) {
1187 // generation 0, step 0 doesn't need to-space
1188 if (g == 0 && s == 0 && RtsFlags.GcFlags.generations > 1) {
1192 stp = &generations[g].steps[s];
1193 ASSERT(stp->gen_no == g);
1195 // deprecate the existing blocks
1196 stp->old_blocks = stp->blocks;
1197 stp->n_old_blocks = stp->n_blocks;
1201 // we don't have any to-be-scavenged blocks yet
1205 // initialise the large object queues.
1206 stp->scavenged_large_objects = NULL;
1207 stp->n_scavenged_large_blocks = 0;
1209 // mark the large objects as not evacuated yet
1210 for (bd = stp->large_objects; bd; bd = bd->link) {
1211 bd->flags &= ~BF_EVACUATED;
1214 // for a compacted step, we need to allocate the bitmap
1215 if (stp->is_compacted) {
1216 nat bitmap_size; // in bytes
1217 bdescr *bitmap_bdescr;
1220 bitmap_size = stp->n_old_blocks * BLOCK_SIZE / (sizeof(W_)*BITS_PER_BYTE);
1222 if (bitmap_size > 0) {
1223 bitmap_bdescr = allocGroup((lnat)BLOCK_ROUND_UP(bitmap_size)
1225 stp->bitmap = bitmap_bdescr;
1226 bitmap = bitmap_bdescr->start;
1228 debugTrace(DEBUG_gc, "bitmap_size: %d, bitmap: %p",
1229 bitmap_size, bitmap);
1231 // don't forget to fill it with zeros!
1232 memset(bitmap, 0, bitmap_size);
1234 // For each block in this step, point to its bitmap from the
1235 // block descriptor.
1236 for (bd=stp->old_blocks; bd != NULL; bd = bd->link) {
1237 bd->u.bitmap = bitmap;
1238 bitmap += BLOCK_SIZE_W / (sizeof(W_)*BITS_PER_BYTE);
1240 // Also at this point we set the BF_COMPACTED flag
1241 // for this block. The invariant is that
1242 // BF_COMPACTED is always unset, except during GC
1243 // when it is set on those blocks which will be
1245 bd->flags |= BF_COMPACTED;
1251 // For each GC thread, for each step, allocate a "todo" block to
1252 // store evacuated objects to be scavenged, and a block to store
1253 // evacuated objects that do not need to be scavenged.
1254 for (t = 0; t < n_threads; t++) {
1255 for (s = 0; s < generations[g].n_steps; s++) {
1257 // we don't copy objects into g0s0, unless -G0
1258 if (g==0 && s==0 && RtsFlags.GcFlags.generations > 1) continue;
1260 ws = &gc_threads[t].steps[g][s];
1265 ws->todo_large_objects = NULL;
1267 // allocate the first to-space block; extra blocks will be
1268 // chained on as necessary.
1270 ws->buffer_todo_bd = NULL;
1271 gc_alloc_todo_block(ws);
1273 ws->scavd_list = NULL;
1274 ws->n_scavd_blocks = 0;
1280 /* ----------------------------------------------------------------------------
1281 Initialise a generation that is *not* to be collected
1282 ------------------------------------------------------------------------- */
1285 init_uncollected_gen (nat g, nat threads)
1292 for (s = 0; s < generations[g].n_steps; s++) {
1293 stp = &generations[g].steps[s];
1294 stp->scavenged_large_objects = NULL;
1295 stp->n_scavenged_large_blocks = 0;
1298 for (t = 0; t < threads; t++) {
1299 for (s = 0; s < generations[g].n_steps; s++) {
1301 ws = &gc_threads[t].steps[g][s];
1304 ws->buffer_todo_bd = NULL;
1305 ws->todo_large_objects = NULL;
1307 ws->scavd_list = NULL;
1308 ws->n_scavd_blocks = 0;
1310 // If the block at the head of the list in this generation
1311 // is less than 3/4 full, then use it as a todo block.
1312 if (stp->blocks && isPartiallyFull(stp->blocks))
1314 ws->todo_bd = stp->blocks;
1315 ws->todo_free = ws->todo_bd->free;
1316 ws->todo_lim = ws->todo_bd->start + BLOCK_SIZE_W;
1317 stp->blocks = stp->blocks->link;
1319 ws->todo_bd->link = NULL;
1321 // this block is also the scan block; we must scan
1322 // from the current end point.
1323 ws->scan_bd = ws->todo_bd;
1324 ws->scan = ws->scan_bd->free;
1326 // subtract the contents of this block from the stats,
1327 // because we'll count the whole block later.
1328 copied -= ws->scan_bd->free - ws->scan_bd->start;
1335 gc_alloc_todo_block(ws);
1340 // Move the private mutable lists from each capability onto the
1341 // main mutable list for the generation.
1342 for (i = 0; i < n_capabilities; i++) {
1343 for (bd = capabilities[i].mut_lists[g];
1344 bd->link != NULL; bd = bd->link) {
1347 bd->link = generations[g].mut_list;
1348 generations[g].mut_list = capabilities[i].mut_lists[g];
1349 capabilities[i].mut_lists[g] = allocBlock();
1353 /* -----------------------------------------------------------------------------
1354 Initialise a gc_thread before GC
1355 -------------------------------------------------------------------------- */
1358 init_gc_thread (gc_thread *t)
1361 t->failed_to_evac = rtsFalse;
1362 t->eager_promotion = rtsTrue;
1363 t->thunk_selector_depth = 0;
1366 /* -----------------------------------------------------------------------------
1367 Function we pass to GetRoots to evacuate roots.
1368 -------------------------------------------------------------------------- */
1371 mark_root(StgClosure **root)
1376 /* -----------------------------------------------------------------------------
1377 Initialising the static object & mutable lists
1378 -------------------------------------------------------------------------- */
1381 zero_static_object_list(StgClosure* first_static)
1385 const StgInfoTable *info;
1387 for (p = first_static; p != END_OF_STATIC_LIST; p = link) {
1389 link = *STATIC_LINK(info, p);
1390 *STATIC_LINK(info,p) = NULL;
1394 /* -----------------------------------------------------------------------------
1396 -------------------------------------------------------------------------- */
1403 for (c = (StgIndStatic *)revertible_caf_list; c != NULL;
1404 c = (StgIndStatic *)c->static_link)
1406 SET_INFO(c, c->saved_info);
1407 c->saved_info = NULL;
1408 // could, but not necessary: c->static_link = NULL;
1410 revertible_caf_list = NULL;
1414 markCAFs( evac_fn evac )
1418 for (c = (StgIndStatic *)caf_list; c != NULL;
1419 c = (StgIndStatic *)c->static_link)
1421 evac(&c->indirectee);
1423 for (c = (StgIndStatic *)revertible_caf_list; c != NULL;
1424 c = (StgIndStatic *)c->static_link)
1426 evac(&c->indirectee);
1430 /* ----------------------------------------------------------------------------
1431 Update the pointers from the task list
1433 These are treated as weak pointers because we want to allow a main
1434 thread to get a BlockedOnDeadMVar exception in the same way as any
1435 other thread. Note that the threads should all have been retained
1436 by GC by virtue of being on the all_threads list, we're just
1437 updating pointers here.
1438 ------------------------------------------------------------------------- */
1441 update_task_list (void)
1445 for (task = all_tasks; task != NULL; task = task->all_link) {
1446 if (!task->stopped && task->tso) {
1447 ASSERT(task->tso->bound == task);
1448 tso = (StgTSO *) isAlive((StgClosure *)task->tso);
1450 barf("task %p: main thread %d has been GC'd",
1463 /* ----------------------------------------------------------------------------
1464 Reset the sizes of the older generations when we do a major
1467 CURRENT STRATEGY: make all generations except zero the same size.
1468 We have to stay within the maximum heap size, and leave a certain
1469 percentage of the maximum heap size available to allocate into.
1470 ------------------------------------------------------------------------- */
1473 resize_generations (void)
1477 if (major_gc && RtsFlags.GcFlags.generations > 1) {
1478 nat live, size, min_alloc;
1479 nat max = RtsFlags.GcFlags.maxHeapSize;
1480 nat gens = RtsFlags.GcFlags.generations;
1482 // live in the oldest generations
1483 live = oldest_gen->steps[0].n_blocks +
1484 oldest_gen->steps[0].n_large_blocks;
1486 // default max size for all generations except zero
1487 size = stg_max(live * RtsFlags.GcFlags.oldGenFactor,
1488 RtsFlags.GcFlags.minOldGenSize);
1490 // minimum size for generation zero
1491 min_alloc = stg_max((RtsFlags.GcFlags.pcFreeHeap * max) / 200,
1492 RtsFlags.GcFlags.minAllocAreaSize);
1494 // Auto-enable compaction when the residency reaches a
1495 // certain percentage of the maximum heap size (default: 30%).
1496 if (RtsFlags.GcFlags.generations > 1 &&
1497 (RtsFlags.GcFlags.compact ||
1499 oldest_gen->steps[0].n_blocks >
1500 (RtsFlags.GcFlags.compactThreshold * max) / 100))) {
1501 oldest_gen->steps[0].is_compacted = 1;
1502 // debugBelch("compaction: on\n", live);
1504 oldest_gen->steps[0].is_compacted = 0;
1505 // debugBelch("compaction: off\n", live);
1508 // if we're going to go over the maximum heap size, reduce the
1509 // size of the generations accordingly. The calculation is
1510 // different if compaction is turned on, because we don't need
1511 // to double the space required to collect the old generation.
1514 // this test is necessary to ensure that the calculations
1515 // below don't have any negative results - we're working
1516 // with unsigned values here.
1517 if (max < min_alloc) {
1521 if (oldest_gen->steps[0].is_compacted) {
1522 if ( (size + (size - 1) * (gens - 2) * 2) + min_alloc > max ) {
1523 size = (max - min_alloc) / ((gens - 1) * 2 - 1);
1526 if ( (size * (gens - 1) * 2) + min_alloc > max ) {
1527 size = (max - min_alloc) / ((gens - 1) * 2);
1537 debugBelch("live: %d, min_alloc: %d, size : %d, max = %d\n", live,
1538 min_alloc, size, max);
1541 for (g = 0; g < gens; g++) {
1542 generations[g].max_blocks = size;
1547 /* -----------------------------------------------------------------------------
1548 Calculate the new size of the nursery, and resize it.
1549 -------------------------------------------------------------------------- */
1552 resize_nursery (void)
1554 if (RtsFlags.GcFlags.generations == 1)
1555 { // Two-space collector:
1558 /* set up a new nursery. Allocate a nursery size based on a
1559 * function of the amount of live data (by default a factor of 2)
1560 * Use the blocks from the old nursery if possible, freeing up any
1563 * If we get near the maximum heap size, then adjust our nursery
1564 * size accordingly. If the nursery is the same size as the live
1565 * data (L), then we need 3L bytes. We can reduce the size of the
1566 * nursery to bring the required memory down near 2L bytes.
1568 * A normal 2-space collector would need 4L bytes to give the same
1569 * performance we get from 3L bytes, reducing to the same
1570 * performance at 2L bytes.
1572 blocks = g0s0->n_old_blocks;
1574 if ( RtsFlags.GcFlags.maxHeapSize != 0 &&
1575 blocks * RtsFlags.GcFlags.oldGenFactor * 2 >
1576 RtsFlags.GcFlags.maxHeapSize )
1578 long adjusted_blocks; // signed on purpose
1581 adjusted_blocks = (RtsFlags.GcFlags.maxHeapSize - 2 * blocks);
1583 debugTrace(DEBUG_gc, "near maximum heap size of 0x%x blocks, blocks = %d, adjusted to %ld",
1584 RtsFlags.GcFlags.maxHeapSize, blocks, adjusted_blocks);
1586 pc_free = adjusted_blocks * 100 / RtsFlags.GcFlags.maxHeapSize;
1587 if (pc_free < RtsFlags.GcFlags.pcFreeHeap) /* might even * be < 0 */
1591 blocks = adjusted_blocks;
1595 blocks *= RtsFlags.GcFlags.oldGenFactor;
1596 if (blocks < RtsFlags.GcFlags.minAllocAreaSize)
1598 blocks = RtsFlags.GcFlags.minAllocAreaSize;
1601 resizeNurseries(blocks);
1603 else // Generational collector
1606 * If the user has given us a suggested heap size, adjust our
1607 * allocation area to make best use of the memory available.
1609 if (RtsFlags.GcFlags.heapSizeSuggestion)
1612 nat needed = calcNeeded(); // approx blocks needed at next GC
1614 /* Guess how much will be live in generation 0 step 0 next time.
1615 * A good approximation is obtained by finding the
1616 * percentage of g0s0 that was live at the last minor GC.
1618 * We have an accurate figure for the amount of copied data in
1619 * 'copied', but we must convert this to a number of blocks, with
1620 * a small adjustment for estimated slop at the end of a block
1625 g0s0_pcnt_kept = ((copied / (BLOCK_SIZE_W - 10)) * 100)
1626 / countNurseryBlocks();
1629 /* Estimate a size for the allocation area based on the
1630 * information available. We might end up going slightly under
1631 * or over the suggested heap size, but we should be pretty
1634 * Formula: suggested - needed
1635 * ----------------------------
1636 * 1 + g0s0_pcnt_kept/100
1638 * where 'needed' is the amount of memory needed at the next
1639 * collection for collecting all steps except g0s0.
1642 (((long)RtsFlags.GcFlags.heapSizeSuggestion - (long)needed) * 100) /
1643 (100 + (long)g0s0_pcnt_kept);
1645 if (blocks < (long)RtsFlags.GcFlags.minAllocAreaSize) {
1646 blocks = RtsFlags.GcFlags.minAllocAreaSize;
1649 resizeNurseries((nat)blocks);
1653 // we might have added extra large blocks to the nursery, so
1654 // resize back to minAllocAreaSize again.
1655 resizeNurseriesFixed(RtsFlags.GcFlags.minAllocAreaSize);
1660 /* -----------------------------------------------------------------------------
1661 Sanity code for CAF garbage collection.
1663 With DEBUG turned on, we manage a CAF list in addition to the SRT
1664 mechanism. After GC, we run down the CAF list and blackhole any
1665 CAFs which have been garbage collected. This means we get an error
1666 whenever the program tries to enter a garbage collected CAF.
1668 Any garbage collected CAFs are taken off the CAF list at the same
1670 -------------------------------------------------------------------------- */
1672 #if 0 && defined(DEBUG)
1679 const StgInfoTable *info;
1690 ASSERT(info->type == IND_STATIC);
1692 if (STATIC_LINK(info,p) == NULL) {
1693 debugTrace(DEBUG_gccafs, "CAF gc'd at 0x%04lx", (long)p);
1695 SET_INFO(p,&stg_BLACKHOLE_info);
1696 p = STATIC_LINK2(info,p);
1700 pp = &STATIC_LINK2(info,p);
1707 debugTrace(DEBUG_gccafs, "%d CAFs live", i);