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 // Mark the weak pointer list, and prepare to detect dead weak pointers.
351 // Mark the stable pointer table.
352 markStablePtrTable(mark_root);
354 /* -------------------------------------------------------------------------
355 * Repeatedly scavenge all the areas we know about until there's no
356 * more scavenging to be done.
361 // The other threads are now stopped. We might recurse back to
362 // here, but from now on this is the only thread.
364 // if any blackholes are alive, make the threads that wait on
366 if (traverseBlackholeQueue()) {
371 // must be last... invariant is that everything is fully
372 // scavenged at this point.
373 if (traverseWeakPtrList()) { // returns rtsTrue if evaced something
378 // If we get to here, there's really nothing left to do.
382 // Update pointers from the Task list
385 // Now see which stable names are still alive.
389 // We call processHeapClosureForDead() on every closure destroyed during
390 // the current garbage collection, so we invoke LdvCensusForDead().
391 if (RtsFlags.ProfFlags.doHeapProfile == HEAP_BY_LDV
392 || RtsFlags.ProfFlags.bioSelector != NULL)
396 // NO MORE EVACUATION AFTER THIS POINT!
397 // Finally: compaction of the oldest generation.
398 if (major_gc && oldest_gen->steps[0].is_compacted) {
399 // save number of blocks for stats
400 oldgen_saved_blocks = oldest_gen->steps[0].n_old_blocks;
404 IF_DEBUG(sanity, checkGlobalTSOList(rtsFalse));
406 // Two-space collector: free the old to-space.
407 // g0s0->old_blocks is the old nursery
408 // g0s0->blocks is to-space from the previous GC
409 if (RtsFlags.GcFlags.generations == 1) {
410 if (g0s0->blocks != NULL) {
411 freeChain(g0s0->blocks);
416 // For each workspace, in each thread:
417 // * clear the BF_EVACUATED flag from each copied block
418 // * move the copied blocks to the step
424 for (t = 0; t < n_gc_threads; t++) {
425 thr = &gc_threads[t];
427 for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
428 for (s = 0; s < generations[g].n_steps; s++) {
429 ws = &thr->steps[g][s];
430 if (g==0 && s==0) continue;
433 // ASSERT( ws->scan_bd == ws->todo_bd );
434 ASSERT( ws->scan_bd ? ws->scan == ws->scan_bd->free : 1 );
436 // Push the final block
437 if (ws->scan_bd) { push_scan_block(ws->scan_bd, ws); }
439 ASSERT(countBlocks(ws->scavd_list) == ws->n_scavd_blocks);
441 prev = ws->scavd_list;
442 for (bd = ws->scavd_list; bd != NULL; bd = bd->link) {
443 bd->flags &= ~BF_EVACUATED; // now from-space
446 prev->link = ws->stp->blocks;
447 ws->stp->blocks = ws->scavd_list;
448 ws->stp->n_blocks += ws->n_scavd_blocks;
449 ASSERT(countBlocks(ws->stp->blocks) == ws->stp->n_blocks);
455 // Two-space collector: swap the semi-spaces around.
456 // Currently: g0s0->old_blocks is the old nursery
457 // g0s0->blocks is to-space from this GC
458 // We want these the other way around.
459 if (RtsFlags.GcFlags.generations == 1) {
460 bdescr *nursery_blocks = g0s0->old_blocks;
461 nat n_nursery_blocks = g0s0->n_old_blocks;
462 g0s0->old_blocks = g0s0->blocks;
463 g0s0->n_old_blocks = g0s0->n_blocks;
464 g0s0->blocks = nursery_blocks;
465 g0s0->n_blocks = n_nursery_blocks;
468 /* run through all the generations/steps and tidy up
470 for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
473 generations[g].collections++; // for stats
476 // Count the mutable list as bytes "copied" for the purposes of
477 // stats. Every mutable list is copied during every GC.
479 nat mut_list_size = 0;
480 for (bd = generations[g].mut_list; bd != NULL; bd = bd->link) {
481 mut_list_size += bd->free - bd->start;
483 copied += mut_list_size;
486 "mut_list_size: %lu (%d vars, %d arrays, %d MVARs, %d others)",
487 (unsigned long)(mut_list_size * sizeof(W_)),
488 mutlist_MUTVARS, mutlist_MUTARRS, mutlist_MVARS, mutlist_OTHERS);
491 for (s = 0; s < generations[g].n_steps; s++) {
493 stp = &generations[g].steps[s];
495 // for generations we collected...
498 /* free old memory and shift to-space into from-space for all
499 * the collected steps (except the allocation area). These
500 * freed blocks will probaby be quickly recycled.
502 if (!(g == 0 && s == 0 && RtsFlags.GcFlags.generations > 1)) {
503 if (stp->is_compacted)
505 // for a compacted step, just shift the new to-space
506 // onto the front of the now-compacted existing blocks.
507 for (bd = stp->blocks; bd != NULL; bd = bd->link) {
508 bd->flags &= ~BF_EVACUATED; // now from-space
510 // tack the new blocks on the end of the existing blocks
511 if (stp->old_blocks != NULL) {
512 for (bd = stp->old_blocks; bd != NULL; bd = next) {
513 // NB. this step might not be compacted next
514 // time, so reset the BF_COMPACTED flags.
515 // They are set before GC if we're going to
516 // compact. (search for BF_COMPACTED above).
517 bd->flags &= ~BF_COMPACTED;
520 bd->link = stp->blocks;
523 stp->blocks = stp->old_blocks;
525 // add the new blocks to the block tally
526 stp->n_blocks += stp->n_old_blocks;
527 ASSERT(countBlocks(stp->blocks) == stp->n_blocks);
531 freeChain(stp->old_blocks);
533 stp->old_blocks = NULL;
534 stp->n_old_blocks = 0;
537 /* LARGE OBJECTS. The current live large objects are chained on
538 * scavenged_large, having been moved during garbage
539 * collection from large_objects. Any objects left on
540 * large_objects list are therefore dead, so we free them here.
542 for (bd = stp->large_objects; bd != NULL; bd = next) {
548 // update the count of blocks used by large objects
549 for (bd = stp->scavenged_large_objects; bd != NULL; bd = bd->link) {
550 bd->flags &= ~BF_EVACUATED;
552 stp->large_objects = stp->scavenged_large_objects;
553 stp->n_large_blocks = stp->n_scavenged_large_blocks;
556 else // for older generations...
558 /* For older generations, we need to append the
559 * scavenged_large_object list (i.e. large objects that have been
560 * promoted during this GC) to the large_object list for that step.
562 for (bd = stp->scavenged_large_objects; bd; bd = next) {
564 bd->flags &= ~BF_EVACUATED;
565 dbl_link_onto(bd, &stp->large_objects);
568 // add the new blocks we promoted during this GC
569 stp->n_large_blocks += stp->n_scavenged_large_blocks;
574 // update the max size of older generations after a major GC
575 resize_generations();
577 // Guess the amount of live data for stats.
580 // Free the small objects allocated via allocate(), since this will
581 // all have been copied into G0S1 now.
582 if (RtsFlags.GcFlags.generations > 1) {
583 if (g0s0->blocks != NULL) {
584 freeChain(g0s0->blocks);
590 alloc_blocks_lim = RtsFlags.GcFlags.minAllocAreaSize;
592 // Start a new pinned_object_block
593 pinned_object_block = NULL;
595 // Free the mark stack.
596 if (mark_stack_bdescr != NULL) {
597 freeGroup(mark_stack_bdescr);
601 for (g = 0; g <= N; g++) {
602 for (s = 0; s < generations[g].n_steps; s++) {
603 stp = &generations[g].steps[s];
604 if (stp->bitmap != NULL) {
605 freeGroup(stp->bitmap);
613 // mark the garbage collected CAFs as dead
614 #if 0 && defined(DEBUG) // doesn't work at the moment
615 if (major_gc) { gcCAFs(); }
619 // resetStaticObjectForRetainerProfiling() must be called before
621 resetStaticObjectForRetainerProfiling();
624 // zero the scavenged static object list
626 zero_static_object_list(scavenged_static_objects);
632 // start any pending finalizers
634 scheduleFinalizers(last_free_capability, old_weak_ptr_list);
637 // send exceptions to any threads which were about to die
639 resurrectThreads(resurrected_threads);
642 // Update the stable pointer hash table.
643 updateStablePtrTable(major_gc);
645 // check sanity after GC
646 IF_DEBUG(sanity, checkSanity());
648 // extra GC trace info
649 IF_DEBUG(gc, statDescribeGens());
652 // symbol-table based profiling
653 /* heapCensus(to_blocks); */ /* ToDo */
656 // restore enclosing cost centre
662 // check for memory leaks if DEBUG is on
666 #ifdef RTS_GTK_FRONTPANEL
667 if (RtsFlags.GcFlags.frontpanel) {
668 updateFrontPanelAfterGC( N, live );
672 // ok, GC over: tell the stats department what happened.
673 stat_endGC(allocated, live, copied, N);
675 #if defined(RTS_USER_SIGNALS)
676 if (RtsFlags.MiscFlags.install_signal_handlers) {
677 // unblock signals again
678 unblockUserSignals();
687 /* ---------------------------------------------------------------------------
688 Where are the roots that we know about?
690 - all the threads on the runnable queue
691 - all the threads on the blocked queue
692 - all the threads on the sleeping queue
693 - all the thread currently executing a _ccall_GC
694 - all the "main threads"
696 ------------------------------------------------------------------------ */
699 GetRoots( evac_fn evac )
705 for (i = 0; i < n_capabilities; i++) {
706 cap = &capabilities[i];
707 evac((StgClosure **)(void *)&cap->run_queue_hd);
708 evac((StgClosure **)(void *)&cap->run_queue_tl);
709 #if defined(THREADED_RTS)
710 evac((StgClosure **)(void *)&cap->wakeup_queue_hd);
711 evac((StgClosure **)(void *)&cap->wakeup_queue_tl);
713 for (task = cap->suspended_ccalling_tasks; task != NULL;
715 debugTrace(DEBUG_sched,
716 "evac'ing suspended TSO %lu", (unsigned long)task->suspended_tso->id);
717 evac((StgClosure **)(void *)&task->suspended_tso);
722 #if !defined(THREADED_RTS)
723 evac((StgClosure **)(void *)&blocked_queue_hd);
724 evac((StgClosure **)(void *)&blocked_queue_tl);
725 evac((StgClosure **)(void *)&sleeping_queue);
728 // evac((StgClosure **)&blackhole_queue);
730 #if defined(THREADED_RTS)
731 markSparkQueue(evac);
734 #if defined(RTS_USER_SIGNALS)
735 // mark the signal handlers (signals should be already blocked)
736 markSignalHandlers(evac);
740 /* -----------------------------------------------------------------------------
741 isAlive determines whether the given closure is still alive (after
742 a garbage collection) or not. It returns the new address of the
743 closure if it is alive, or NULL otherwise.
745 NOTE: Use it before compaction only!
746 It untags and (if needed) retags pointers to closures.
747 -------------------------------------------------------------------------- */
751 isAlive(StgClosure *p)
753 const StgInfoTable *info;
759 /* The tag and the pointer are split, to be merged later when needed. */
760 tag = GET_CLOSURE_TAG(p);
761 q = UNTAG_CLOSURE(p);
763 ASSERT(LOOKS_LIKE_CLOSURE_PTR(q));
766 // ignore static closures
768 // ToDo: for static closures, check the static link field.
769 // Problem here is that we sometimes don't set the link field, eg.
770 // for static closures with an empty SRT or CONSTR_STATIC_NOCAFs.
772 if (!HEAP_ALLOCED(q)) {
776 // ignore closures in generations that we're not collecting.
778 if (bd->gen_no > N) {
782 // if it's a pointer into to-space, then we're done
783 if (bd->flags & BF_EVACUATED) {
787 // large objects use the evacuated flag
788 if (bd->flags & BF_LARGE) {
792 // check the mark bit for compacted steps
793 if ((bd->flags & BF_COMPACTED) && is_marked((P_)q,bd)) {
797 switch (info->type) {
802 case IND_OLDGEN: // rely on compatible layout with StgInd
803 case IND_OLDGEN_PERM:
804 // follow indirections
805 p = ((StgInd *)q)->indirectee;
810 return ((StgEvacuated *)q)->evacuee;
813 if (((StgTSO *)q)->what_next == ThreadRelocated) {
814 p = (StgClosure *)((StgTSO *)q)->link;
826 /* -----------------------------------------------------------------------------
827 Figure out which generation to collect, initialise N and major_gc.
828 -------------------------------------------------------------------------- */
831 initialise_N (rtsBool force_major_gc)
835 if (force_major_gc) {
836 N = RtsFlags.GcFlags.generations - 1;
840 for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
841 if (generations[g].steps[0].n_blocks +
842 generations[g].steps[0].n_large_blocks
843 >= generations[g].max_blocks) {
847 major_gc = (N == RtsFlags.GcFlags.generations-1);
851 /* -----------------------------------------------------------------------------
852 Initialise the gc_thread structures.
853 -------------------------------------------------------------------------- */
856 alloc_gc_thread (gc_thread *t, int n)
863 initCondition(&t->wake_cond);
864 initMutex(&t->wake_mutex);
865 t->wakeup = rtsFalse;
870 t->free_blocks = NULL;
879 t->steps = stgMallocBytes(RtsFlags.GcFlags.generations *
880 sizeof(step_workspace *),
881 "initialise_gc_thread");
883 for (g = 0; g < RtsFlags.GcFlags.generations; g++)
885 t->steps[g] = stgMallocBytes(generations[g].n_steps *
886 sizeof(step_workspace),
887 "initialise_gc_thread/2");
889 for (s = 0; s < generations[g].n_steps; s++)
891 ws = &t->steps[g][s];
892 ws->stp = &generations[g].steps[s];
899 ws->buffer_todo_bd = NULL;
901 ws->scavd_list = NULL;
902 ws->n_scavd_blocks = 0;
909 alloc_gc_threads (void)
911 if (gc_threads == NULL) {
912 #if defined(THREADED_RTS)
914 gc_threads = stgMallocBytes (RtsFlags.ParFlags.gcThreads *
918 for (i = 0; i < RtsFlags.ParFlags.gcThreads; i++) {
919 alloc_gc_thread(&gc_threads[i], i);
922 gc_threads = stgMallocBytes (sizeof(gc_thread),
925 alloc_gc_thread(gc_threads, 0);
930 /* ----------------------------------------------------------------------------
932 ------------------------------------------------------------------------- */
934 static nat gc_running_threads;
936 #if defined(THREADED_RTS)
937 static Mutex gc_running_mutex;
944 ACQUIRE_LOCK(&gc_running_mutex);
945 n_running = ++gc_running_threads;
946 RELEASE_LOCK(&gc_running_mutex);
954 ACQUIRE_LOCK(&gc_running_mutex);
955 n_running = --gc_running_threads;
956 RELEASE_LOCK(&gc_running_mutex);
961 // gc_thread_work(): Scavenge until there's no work left to do and all
962 // the running threads are idle.
965 gc_thread_work (void)
969 debugTrace(DEBUG_gc, "GC thread %d working", gct->thread_index);
971 // gc_running_threads has already been incremented for us; either
972 // this is the main thread and we incremented it inside
973 // GarbageCollect(), or this is a worker thread and the main
974 // thread bumped gc_running_threads before waking us up.
978 // scavenge_loop() only exits when there's no work to do
981 debugTrace(DEBUG_gc, "GC thread %d idle (%d still running)",
982 gct->thread_index, r);
984 while (gc_running_threads != 0) {
989 // any_work() does not remove the work from the queue, it
990 // just checks for the presence of work. If we find any,
991 // then we increment gc_running_threads and go back to
992 // scavenge_loop() to perform any pending work.
995 // All threads are now stopped
996 debugTrace(DEBUG_gc, "GC thread %d finished.", gct->thread_index);
1000 #if defined(THREADED_RTS)
1002 gc_thread_mainloop (void)
1004 while (!gct->exit) {
1006 // Wait until we're told to wake up
1007 ACQUIRE_LOCK(&gct->wake_mutex);
1008 while (!gct->wakeup) {
1009 debugTrace(DEBUG_gc, "GC thread %d standing by...",
1011 waitCondition(&gct->wake_cond, &gct->wake_mutex);
1013 RELEASE_LOCK(&gct->wake_mutex);
1014 gct->wakeup = rtsFalse;
1015 if (gct->exit) break;
1018 // start performance counters in this thread...
1019 if (gct->papi_events == -1) {
1020 papi_init_eventset(&gct->papi_events);
1022 papi_thread_start_gc1_count(gct->papi_events);
1028 // count events in this thread towards the GC totals
1029 papi_thread_stop_gc1_count(gct->papi_events);
1035 #if defined(THREADED_RTS)
1037 gc_thread_entry (gc_thread *my_gct)
1040 debugTrace(DEBUG_gc, "GC thread %d starting...", gct->thread_index);
1041 gct->id = osThreadId();
1042 gc_thread_mainloop();
1047 start_gc_threads (void)
1049 #if defined(THREADED_RTS)
1052 static rtsBool done = rtsFalse;
1054 gc_running_threads = 0;
1055 initMutex(&gc_running_mutex);
1058 // Start from 1: the main thread is 0
1059 for (i = 1; i < RtsFlags.ParFlags.gcThreads; i++) {
1060 createOSThread(&id, (OSThreadProc*)&gc_thread_entry,
1069 wakeup_gc_threads (nat n_threads USED_IF_THREADS)
1071 #if defined(THREADED_RTS)
1073 for (i=1; i < n_threads; i++) {
1075 ACQUIRE_LOCK(&gc_threads[i].wake_mutex);
1076 gc_threads[i].wakeup = rtsTrue;
1077 signalCondition(&gc_threads[i].wake_cond);
1078 RELEASE_LOCK(&gc_threads[i].wake_mutex);
1083 /* ----------------------------------------------------------------------------
1084 Initialise a generation that is to be collected
1085 ------------------------------------------------------------------------- */
1088 init_collected_gen (nat g, nat n_threads)
1095 // Throw away the current mutable list. Invariant: the mutable
1096 // list always has at least one block; this means we can avoid a
1097 // check for NULL in recordMutable().
1099 freeChain(generations[g].mut_list);
1100 generations[g].mut_list = allocBlock();
1101 for (i = 0; i < n_capabilities; i++) {
1102 freeChain(capabilities[i].mut_lists[g]);
1103 capabilities[i].mut_lists[g] = allocBlock();
1107 for (s = 0; s < generations[g].n_steps; s++) {
1109 // generation 0, step 0 doesn't need to-space
1110 if (g == 0 && s == 0 && RtsFlags.GcFlags.generations > 1) {
1114 stp = &generations[g].steps[s];
1115 ASSERT(stp->gen_no == g);
1117 // deprecate the existing blocks
1118 stp->old_blocks = stp->blocks;
1119 stp->n_old_blocks = stp->n_blocks;
1123 // we don't have any to-be-scavenged blocks yet
1127 // initialise the large object queues.
1128 stp->scavenged_large_objects = NULL;
1129 stp->n_scavenged_large_blocks = 0;
1131 // mark the large objects as not evacuated yet
1132 for (bd = stp->large_objects; bd; bd = bd->link) {
1133 bd->flags &= ~BF_EVACUATED;
1136 // for a compacted step, we need to allocate the bitmap
1137 if (stp->is_compacted) {
1138 nat bitmap_size; // in bytes
1139 bdescr *bitmap_bdescr;
1142 bitmap_size = stp->n_old_blocks * BLOCK_SIZE / (sizeof(W_)*BITS_PER_BYTE);
1144 if (bitmap_size > 0) {
1145 bitmap_bdescr = allocGroup((lnat)BLOCK_ROUND_UP(bitmap_size)
1147 stp->bitmap = bitmap_bdescr;
1148 bitmap = bitmap_bdescr->start;
1150 debugTrace(DEBUG_gc, "bitmap_size: %d, bitmap: %p",
1151 bitmap_size, bitmap);
1153 // don't forget to fill it with zeros!
1154 memset(bitmap, 0, bitmap_size);
1156 // For each block in this step, point to its bitmap from the
1157 // block descriptor.
1158 for (bd=stp->old_blocks; bd != NULL; bd = bd->link) {
1159 bd->u.bitmap = bitmap;
1160 bitmap += BLOCK_SIZE_W / (sizeof(W_)*BITS_PER_BYTE);
1162 // Also at this point we set the BF_COMPACTED flag
1163 // for this block. The invariant is that
1164 // BF_COMPACTED is always unset, except during GC
1165 // when it is set on those blocks which will be
1167 bd->flags |= BF_COMPACTED;
1173 // For each GC thread, for each step, allocate a "todo" block to
1174 // store evacuated objects to be scavenged, and a block to store
1175 // evacuated objects that do not need to be scavenged.
1176 for (t = 0; t < n_threads; t++) {
1177 for (s = 0; s < generations[g].n_steps; s++) {
1179 // we don't copy objects into g0s0, unless -G0
1180 if (g==0 && s==0 && RtsFlags.GcFlags.generations > 1) continue;
1182 ws = &gc_threads[t].steps[g][s];
1187 ws->todo_large_objects = NULL;
1189 // allocate the first to-space block; extra blocks will be
1190 // chained on as necessary.
1192 ws->buffer_todo_bd = NULL;
1193 gc_alloc_todo_block(ws);
1195 ws->scavd_list = NULL;
1196 ws->n_scavd_blocks = 0;
1202 /* ----------------------------------------------------------------------------
1203 Initialise a generation that is *not* to be collected
1204 ------------------------------------------------------------------------- */
1207 init_uncollected_gen (nat g, nat threads)
1214 for (s = 0; s < generations[g].n_steps; s++) {
1215 stp = &generations[g].steps[s];
1216 stp->scavenged_large_objects = NULL;
1217 stp->n_scavenged_large_blocks = 0;
1220 for (t = 0; t < threads; t++) {
1221 for (s = 0; s < generations[g].n_steps; s++) {
1223 ws = &gc_threads[t].steps[g][s];
1226 ws->buffer_todo_bd = NULL;
1227 ws->todo_large_objects = NULL;
1229 ws->scavd_list = NULL;
1230 ws->n_scavd_blocks = 0;
1232 // If the block at the head of the list in this generation
1233 // is less than 3/4 full, then use it as a todo block.
1234 if (isPartiallyFull(stp->blocks))
1236 ws->todo_bd = stp->blocks;
1237 ws->todo_free = ws->todo_bd->free;
1238 ws->todo_lim = ws->todo_bd->start + BLOCK_SIZE_W;
1239 stp->blocks = stp->blocks->link;
1241 ws->todo_bd->link = NULL;
1243 // this block is also the scan block; we must scan
1244 // from the current end point.
1245 ws->scan_bd = ws->todo_bd;
1246 ws->scan = ws->scan_bd->free;
1248 // subtract the contents of this block from the stats,
1249 // because we'll count the whole block later.
1250 copied -= ws->scan_bd->free - ws->scan_bd->start;
1257 gc_alloc_todo_block(ws);
1262 // Move the private mutable lists from each capability onto the
1263 // main mutable list for the generation.
1264 for (i = 0; i < n_capabilities; i++) {
1265 for (bd = capabilities[i].mut_lists[g];
1266 bd->link != NULL; bd = bd->link) {
1269 bd->link = generations[g].mut_list;
1270 generations[g].mut_list = capabilities[i].mut_lists[g];
1271 capabilities[i].mut_lists[g] = allocBlock();
1275 /* -----------------------------------------------------------------------------
1276 Initialise a gc_thread before GC
1277 -------------------------------------------------------------------------- */
1280 init_gc_thread (gc_thread *t)
1283 t->failed_to_evac = rtsFalse;
1284 t->eager_promotion = rtsTrue;
1285 t->thunk_selector_depth = 0;
1288 /* -----------------------------------------------------------------------------
1289 Function we pass to GetRoots to evacuate roots.
1290 -------------------------------------------------------------------------- */
1293 mark_root(StgClosure **root)
1298 /* -----------------------------------------------------------------------------
1299 Initialising the static object & mutable lists
1300 -------------------------------------------------------------------------- */
1303 zero_static_object_list(StgClosure* first_static)
1307 const StgInfoTable *info;
1309 for (p = first_static; p != END_OF_STATIC_LIST; p = link) {
1311 link = *STATIC_LINK(info, p);
1312 *STATIC_LINK(info,p) = NULL;
1316 /* -----------------------------------------------------------------------------
1318 -------------------------------------------------------------------------- */
1325 for (c = (StgIndStatic *)revertible_caf_list; c != NULL;
1326 c = (StgIndStatic *)c->static_link)
1328 SET_INFO(c, c->saved_info);
1329 c->saved_info = NULL;
1330 // could, but not necessary: c->static_link = NULL;
1332 revertible_caf_list = NULL;
1336 markCAFs( evac_fn evac )
1340 for (c = (StgIndStatic *)caf_list; c != NULL;
1341 c = (StgIndStatic *)c->static_link)
1343 evac(&c->indirectee);
1345 for (c = (StgIndStatic *)revertible_caf_list; c != NULL;
1346 c = (StgIndStatic *)c->static_link)
1348 evac(&c->indirectee);
1352 /* ----------------------------------------------------------------------------
1353 Update the pointers from the task list
1355 These are treated as weak pointers because we want to allow a main
1356 thread to get a BlockedOnDeadMVar exception in the same way as any
1357 other thread. Note that the threads should all have been retained
1358 by GC by virtue of being on the all_threads list, we're just
1359 updating pointers here.
1360 ------------------------------------------------------------------------- */
1363 update_task_list (void)
1367 for (task = all_tasks; task != NULL; task = task->all_link) {
1368 if (!task->stopped && task->tso) {
1369 ASSERT(task->tso->bound == task);
1370 tso = (StgTSO *) isAlive((StgClosure *)task->tso);
1372 barf("task %p: main thread %d has been GC'd",
1385 /* ----------------------------------------------------------------------------
1386 Reset the sizes of the older generations when we do a major
1389 CURRENT STRATEGY: make all generations except zero the same size.
1390 We have to stay within the maximum heap size, and leave a certain
1391 percentage of the maximum heap size available to allocate into.
1392 ------------------------------------------------------------------------- */
1395 resize_generations (void)
1399 if (major_gc && RtsFlags.GcFlags.generations > 1) {
1400 nat live, size, min_alloc;
1401 nat max = RtsFlags.GcFlags.maxHeapSize;
1402 nat gens = RtsFlags.GcFlags.generations;
1404 // live in the oldest generations
1405 live = oldest_gen->steps[0].n_blocks +
1406 oldest_gen->steps[0].n_large_blocks;
1408 // default max size for all generations except zero
1409 size = stg_max(live * RtsFlags.GcFlags.oldGenFactor,
1410 RtsFlags.GcFlags.minOldGenSize);
1412 // minimum size for generation zero
1413 min_alloc = stg_max((RtsFlags.GcFlags.pcFreeHeap * max) / 200,
1414 RtsFlags.GcFlags.minAllocAreaSize);
1416 // Auto-enable compaction when the residency reaches a
1417 // certain percentage of the maximum heap size (default: 30%).
1418 if (RtsFlags.GcFlags.generations > 1 &&
1419 (RtsFlags.GcFlags.compact ||
1421 oldest_gen->steps[0].n_blocks >
1422 (RtsFlags.GcFlags.compactThreshold * max) / 100))) {
1423 oldest_gen->steps[0].is_compacted = 1;
1424 // debugBelch("compaction: on\n", live);
1426 oldest_gen->steps[0].is_compacted = 0;
1427 // debugBelch("compaction: off\n", live);
1430 // if we're going to go over the maximum heap size, reduce the
1431 // size of the generations accordingly. The calculation is
1432 // different if compaction is turned on, because we don't need
1433 // to double the space required to collect the old generation.
1436 // this test is necessary to ensure that the calculations
1437 // below don't have any negative results - we're working
1438 // with unsigned values here.
1439 if (max < min_alloc) {
1443 if (oldest_gen->steps[0].is_compacted) {
1444 if ( (size + (size - 1) * (gens - 2) * 2) + min_alloc > max ) {
1445 size = (max - min_alloc) / ((gens - 1) * 2 - 1);
1448 if ( (size * (gens - 1) * 2) + min_alloc > max ) {
1449 size = (max - min_alloc) / ((gens - 1) * 2);
1459 debugBelch("live: %d, min_alloc: %d, size : %d, max = %d\n", live,
1460 min_alloc, size, max);
1463 for (g = 0; g < gens; g++) {
1464 generations[g].max_blocks = size;
1469 /* -----------------------------------------------------------------------------
1470 Calculate the new size of the nursery, and resize it.
1471 -------------------------------------------------------------------------- */
1474 resize_nursery (void)
1476 if (RtsFlags.GcFlags.generations == 1)
1477 { // Two-space collector:
1480 /* set up a new nursery. Allocate a nursery size based on a
1481 * function of the amount of live data (by default a factor of 2)
1482 * Use the blocks from the old nursery if possible, freeing up any
1485 * If we get near the maximum heap size, then adjust our nursery
1486 * size accordingly. If the nursery is the same size as the live
1487 * data (L), then we need 3L bytes. We can reduce the size of the
1488 * nursery to bring the required memory down near 2L bytes.
1490 * A normal 2-space collector would need 4L bytes to give the same
1491 * performance we get from 3L bytes, reducing to the same
1492 * performance at 2L bytes.
1494 blocks = g0s0->n_old_blocks;
1496 if ( RtsFlags.GcFlags.maxHeapSize != 0 &&
1497 blocks * RtsFlags.GcFlags.oldGenFactor * 2 >
1498 RtsFlags.GcFlags.maxHeapSize )
1500 long adjusted_blocks; // signed on purpose
1503 adjusted_blocks = (RtsFlags.GcFlags.maxHeapSize - 2 * blocks);
1505 debugTrace(DEBUG_gc, "near maximum heap size of 0x%x blocks, blocks = %d, adjusted to %ld",
1506 RtsFlags.GcFlags.maxHeapSize, blocks, adjusted_blocks);
1508 pc_free = adjusted_blocks * 100 / RtsFlags.GcFlags.maxHeapSize;
1509 if (pc_free < RtsFlags.GcFlags.pcFreeHeap) /* might even * be < 0 */
1513 blocks = adjusted_blocks;
1517 blocks *= RtsFlags.GcFlags.oldGenFactor;
1518 if (blocks < RtsFlags.GcFlags.minAllocAreaSize)
1520 blocks = RtsFlags.GcFlags.minAllocAreaSize;
1523 resizeNurseries(blocks);
1525 else // Generational collector
1528 * If the user has given us a suggested heap size, adjust our
1529 * allocation area to make best use of the memory available.
1531 if (RtsFlags.GcFlags.heapSizeSuggestion)
1534 nat needed = calcNeeded(); // approx blocks needed at next GC
1536 /* Guess how much will be live in generation 0 step 0 next time.
1537 * A good approximation is obtained by finding the
1538 * percentage of g0s0 that was live at the last minor GC.
1540 * We have an accurate figure for the amount of copied data in
1541 * 'copied', but we must convert this to a number of blocks, with
1542 * a small adjustment for estimated slop at the end of a block
1547 g0s0_pcnt_kept = ((copied / (BLOCK_SIZE_W - 10)) * 100)
1548 / countNurseryBlocks();
1551 /* Estimate a size for the allocation area based on the
1552 * information available. We might end up going slightly under
1553 * or over the suggested heap size, but we should be pretty
1556 * Formula: suggested - needed
1557 * ----------------------------
1558 * 1 + g0s0_pcnt_kept/100
1560 * where 'needed' is the amount of memory needed at the next
1561 * collection for collecting all steps except g0s0.
1564 (((long)RtsFlags.GcFlags.heapSizeSuggestion - (long)needed) * 100) /
1565 (100 + (long)g0s0_pcnt_kept);
1567 if (blocks < (long)RtsFlags.GcFlags.minAllocAreaSize) {
1568 blocks = RtsFlags.GcFlags.minAllocAreaSize;
1571 resizeNurseries((nat)blocks);
1575 // we might have added extra large blocks to the nursery, so
1576 // resize back to minAllocAreaSize again.
1577 resizeNurseriesFixed(RtsFlags.GcFlags.minAllocAreaSize);
1582 /* -----------------------------------------------------------------------------
1583 Sanity code for CAF garbage collection.
1585 With DEBUG turned on, we manage a CAF list in addition to the SRT
1586 mechanism. After GC, we run down the CAF list and blackhole any
1587 CAFs which have been garbage collected. This means we get an error
1588 whenever the program tries to enter a garbage collected CAF.
1590 Any garbage collected CAFs are taken off the CAF list at the same
1592 -------------------------------------------------------------------------- */
1594 #if 0 && defined(DEBUG)
1601 const StgInfoTable *info;
1612 ASSERT(info->type == IND_STATIC);
1614 if (STATIC_LINK(info,p) == NULL) {
1615 debugTrace(DEBUG_gccafs, "CAF gc'd at 0x%04lx", (long)p);
1617 SET_INFO(p,&stg_BLACKHOLE_info);
1618 p = STATIC_LINK2(info,p);
1622 pp = &STATIC_LINK2(info,p);
1629 debugTrace(DEBUG_gccafs, "%d CAFs live", i);