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.
585 // Free the small objects allocated via allocate(), since this will
586 // all have been copied into G0S1 now.
587 if (RtsFlags.GcFlags.generations > 1) {
588 if (g0s0->blocks != NULL) {
589 freeChain(g0s0->blocks);
595 alloc_blocks_lim = RtsFlags.GcFlags.minAllocAreaSize;
597 // Start a new pinned_object_block
598 pinned_object_block = NULL;
600 // Free the mark stack.
601 if (mark_stack_bdescr != NULL) {
602 freeGroup(mark_stack_bdescr);
606 for (g = 0; g <= N; g++) {
607 for (s = 0; s < generations[g].n_steps; s++) {
608 stp = &generations[g].steps[s];
609 if (stp->bitmap != NULL) {
610 freeGroup(stp->bitmap);
618 // mark the garbage collected CAFs as dead
619 #if 0 && defined(DEBUG) // doesn't work at the moment
620 if (major_gc) { gcCAFs(); }
624 // resetStaticObjectForRetainerProfiling() must be called before
626 resetStaticObjectForRetainerProfiling();
629 // zero the scavenged static object list
631 zero_static_object_list(scavenged_static_objects);
637 // start any pending finalizers
639 scheduleFinalizers(last_free_capability, old_weak_ptr_list);
642 // send exceptions to any threads which were about to die
644 resurrectThreads(resurrected_threads);
647 // Update the stable pointer hash table.
648 updateStablePtrTable(major_gc);
650 // check sanity after GC
651 IF_DEBUG(sanity, checkSanity());
653 // extra GC trace info
654 IF_DEBUG(gc, statDescribeGens());
657 // symbol-table based profiling
658 /* heapCensus(to_blocks); */ /* ToDo */
661 // restore enclosing cost centre
667 // check for memory leaks if DEBUG is on
671 #ifdef RTS_GTK_FRONTPANEL
672 if (RtsFlags.GcFlags.frontpanel) {
673 updateFrontPanelAfterGC( N, live );
677 // ok, GC over: tell the stats department what happened.
678 stat_endGC(allocated, live, copied, N);
680 #if defined(RTS_USER_SIGNALS)
681 if (RtsFlags.MiscFlags.install_signal_handlers) {
682 // unblock signals again
683 unblockUserSignals();
692 /* ---------------------------------------------------------------------------
693 Where are the roots that we know about?
695 - all the threads on the runnable queue
696 - all the threads on the blocked queue
697 - all the threads on the sleeping queue
698 - all the thread currently executing a _ccall_GC
699 - all the "main threads"
701 ------------------------------------------------------------------------ */
704 GetRoots( evac_fn evac )
710 // Each GC thread is responsible for following roots from the
711 // Capability of the same number. There will usually be the same
712 // or fewer Capabilities as GC threads, but just in case there
713 // are more, we mark every Capability whose number is the GC
714 // thread's index plus a multiple of the number of GC threads.
715 for (i = gct->thread_index; i < n_capabilities; i += n_gc_threads) {
716 cap = &capabilities[i];
717 evac((StgClosure **)(void *)&cap->run_queue_hd);
718 evac((StgClosure **)(void *)&cap->run_queue_tl);
719 #if defined(THREADED_RTS)
720 evac((StgClosure **)(void *)&cap->wakeup_queue_hd);
721 evac((StgClosure **)(void *)&cap->wakeup_queue_tl);
723 for (task = cap->suspended_ccalling_tasks; task != NULL;
725 debugTrace(DEBUG_sched,
726 "evac'ing suspended TSO %lu", (unsigned long)task->suspended_tso->id);
727 evac((StgClosure **)(void *)&task->suspended_tso);
730 #if defined(THREADED_RTS)
731 markSparkQueue(evac,cap);
735 #if !defined(THREADED_RTS)
736 evac((StgClosure **)(void *)&blocked_queue_hd);
737 evac((StgClosure **)(void *)&blocked_queue_tl);
738 evac((StgClosure **)(void *)&sleeping_queue);
742 /* -----------------------------------------------------------------------------
743 isAlive determines whether the given closure is still alive (after
744 a garbage collection) or not. It returns the new address of the
745 closure if it is alive, or NULL otherwise.
747 NOTE: Use it before compaction only!
748 It untags and (if needed) retags pointers to closures.
749 -------------------------------------------------------------------------- */
753 isAlive(StgClosure *p)
755 const StgInfoTable *info;
761 /* The tag and the pointer are split, to be merged later when needed. */
762 tag = GET_CLOSURE_TAG(p);
763 q = UNTAG_CLOSURE(p);
765 ASSERT(LOOKS_LIKE_CLOSURE_PTR(q));
768 // ignore static closures
770 // ToDo: for static closures, check the static link field.
771 // Problem here is that we sometimes don't set the link field, eg.
772 // for static closures with an empty SRT or CONSTR_STATIC_NOCAFs.
774 if (!HEAP_ALLOCED(q)) {
778 // ignore closures in generations that we're not collecting.
780 if (bd->gen_no > N) {
784 // if it's a pointer into to-space, then we're done
785 if (bd->flags & BF_EVACUATED) {
789 // large objects use the evacuated flag
790 if (bd->flags & BF_LARGE) {
794 // check the mark bit for compacted steps
795 if ((bd->flags & BF_COMPACTED) && is_marked((P_)q,bd)) {
799 switch (info->type) {
804 case IND_OLDGEN: // rely on compatible layout with StgInd
805 case IND_OLDGEN_PERM:
806 // follow indirections
807 p = ((StgInd *)q)->indirectee;
812 return ((StgEvacuated *)q)->evacuee;
815 if (((StgTSO *)q)->what_next == ThreadRelocated) {
816 p = (StgClosure *)((StgTSO *)q)->link;
828 /* -----------------------------------------------------------------------------
829 Figure out which generation to collect, initialise N and major_gc.
830 -------------------------------------------------------------------------- */
833 initialise_N (rtsBool force_major_gc)
837 if (force_major_gc) {
838 N = RtsFlags.GcFlags.generations - 1;
842 for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
843 if (generations[g].steps[0].n_blocks +
844 generations[g].steps[0].n_large_blocks
845 >= generations[g].max_blocks) {
849 major_gc = (N == RtsFlags.GcFlags.generations-1);
853 /* -----------------------------------------------------------------------------
854 Initialise the gc_thread structures.
855 -------------------------------------------------------------------------- */
858 alloc_gc_thread (gc_thread *t, int n)
865 initCondition(&t->wake_cond);
866 initMutex(&t->wake_mutex);
867 t->wakeup = rtsFalse;
872 t->free_blocks = NULL;
881 t->steps = stgMallocBytes(RtsFlags.GcFlags.generations *
882 sizeof(step_workspace *),
883 "initialise_gc_thread");
885 for (g = 0; g < RtsFlags.GcFlags.generations; g++)
887 t->steps[g] = stgMallocBytes(generations[g].n_steps *
888 sizeof(step_workspace),
889 "initialise_gc_thread/2");
891 for (s = 0; s < generations[g].n_steps; s++)
893 ws = &t->steps[g][s];
894 ws->stp = &generations[g].steps[s];
901 ws->buffer_todo_bd = NULL;
903 ws->scavd_list = NULL;
904 ws->n_scavd_blocks = 0;
911 alloc_gc_threads (void)
913 if (gc_threads == NULL) {
914 #if defined(THREADED_RTS)
916 gc_threads = stgMallocBytes (RtsFlags.ParFlags.gcThreads *
920 for (i = 0; i < RtsFlags.ParFlags.gcThreads; i++) {
921 alloc_gc_thread(&gc_threads[i], i);
924 gc_threads = stgMallocBytes (sizeof(gc_thread),
927 alloc_gc_thread(gc_threads, 0);
932 /* ----------------------------------------------------------------------------
934 ------------------------------------------------------------------------- */
936 static nat gc_running_threads;
938 #if defined(THREADED_RTS)
939 static Mutex gc_running_mutex;
946 ACQUIRE_LOCK(&gc_running_mutex);
947 n_running = ++gc_running_threads;
948 RELEASE_LOCK(&gc_running_mutex);
956 ACQUIRE_LOCK(&gc_running_mutex);
957 n_running = --gc_running_threads;
958 RELEASE_LOCK(&gc_running_mutex);
963 // gc_thread_work(): Scavenge until there's no work left to do and all
964 // the running threads are idle.
967 gc_thread_work (void)
971 debugTrace(DEBUG_gc, "GC thread %d working", gct->thread_index);
973 // gc_running_threads has already been incremented for us; either
974 // this is the main thread and we incremented it inside
975 // GarbageCollect(), or this is a worker thread and the main
976 // thread bumped gc_running_threads before waking us up.
978 // Every thread evacuates some roots.
984 // scavenge_loop() only exits when there's no work to do
987 debugTrace(DEBUG_gc, "GC thread %d idle (%d still running)",
988 gct->thread_index, r);
990 while (gc_running_threads != 0) {
995 // any_work() does not remove the work from the queue, it
996 // just checks for the presence of work. If we find any,
997 // then we increment gc_running_threads and go back to
998 // scavenge_loop() to perform any pending work.
1001 // All threads are now stopped
1002 debugTrace(DEBUG_gc, "GC thread %d finished.", gct->thread_index);
1006 #if defined(THREADED_RTS)
1008 gc_thread_mainloop (void)
1010 while (!gct->exit) {
1012 // Wait until we're told to wake up
1013 ACQUIRE_LOCK(&gct->wake_mutex);
1014 while (!gct->wakeup) {
1015 debugTrace(DEBUG_gc, "GC thread %d standing by...",
1017 waitCondition(&gct->wake_cond, &gct->wake_mutex);
1019 RELEASE_LOCK(&gct->wake_mutex);
1020 gct->wakeup = rtsFalse;
1021 if (gct->exit) break;
1024 // start performance counters in this thread...
1025 if (gct->papi_events == -1) {
1026 papi_init_eventset(&gct->papi_events);
1028 papi_thread_start_gc1_count(gct->papi_events);
1034 // count events in this thread towards the GC totals
1035 papi_thread_stop_gc1_count(gct->papi_events);
1041 #if defined(THREADED_RTS)
1043 gc_thread_entry (gc_thread *my_gct)
1046 debugTrace(DEBUG_gc, "GC thread %d starting...", gct->thread_index);
1047 gct->id = osThreadId();
1048 gc_thread_mainloop();
1053 start_gc_threads (void)
1055 #if defined(THREADED_RTS)
1058 static rtsBool done = rtsFalse;
1060 gc_running_threads = 0;
1061 initMutex(&gc_running_mutex);
1064 // Start from 1: the main thread is 0
1065 for (i = 1; i < RtsFlags.ParFlags.gcThreads; i++) {
1066 createOSThread(&id, (OSThreadProc*)&gc_thread_entry,
1075 wakeup_gc_threads (nat n_threads USED_IF_THREADS)
1077 #if defined(THREADED_RTS)
1079 for (i=1; i < n_threads; i++) {
1081 ACQUIRE_LOCK(&gc_threads[i].wake_mutex);
1082 gc_threads[i].wakeup = rtsTrue;
1083 signalCondition(&gc_threads[i].wake_cond);
1084 RELEASE_LOCK(&gc_threads[i].wake_mutex);
1089 /* ----------------------------------------------------------------------------
1090 Initialise a generation that is to be collected
1091 ------------------------------------------------------------------------- */
1094 init_collected_gen (nat g, nat n_threads)
1101 // Throw away the current mutable list. Invariant: the mutable
1102 // list always has at least one block; this means we can avoid a
1103 // check for NULL in recordMutable().
1105 freeChain(generations[g].mut_list);
1106 generations[g].mut_list = allocBlock();
1107 for (i = 0; i < n_capabilities; i++) {
1108 freeChain(capabilities[i].mut_lists[g]);
1109 capabilities[i].mut_lists[g] = allocBlock();
1113 for (s = 0; s < generations[g].n_steps; s++) {
1115 // generation 0, step 0 doesn't need to-space
1116 if (g == 0 && s == 0 && RtsFlags.GcFlags.generations > 1) {
1120 stp = &generations[g].steps[s];
1121 ASSERT(stp->gen_no == g);
1123 // deprecate the existing blocks
1124 stp->old_blocks = stp->blocks;
1125 stp->n_old_blocks = stp->n_blocks;
1129 // we don't have any to-be-scavenged blocks yet
1133 // initialise the large object queues.
1134 stp->scavenged_large_objects = NULL;
1135 stp->n_scavenged_large_blocks = 0;
1137 // mark the large objects as not evacuated yet
1138 for (bd = stp->large_objects; bd; bd = bd->link) {
1139 bd->flags &= ~BF_EVACUATED;
1142 // for a compacted step, we need to allocate the bitmap
1143 if (stp->is_compacted) {
1144 nat bitmap_size; // in bytes
1145 bdescr *bitmap_bdescr;
1148 bitmap_size = stp->n_old_blocks * BLOCK_SIZE / (sizeof(W_)*BITS_PER_BYTE);
1150 if (bitmap_size > 0) {
1151 bitmap_bdescr = allocGroup((lnat)BLOCK_ROUND_UP(bitmap_size)
1153 stp->bitmap = bitmap_bdescr;
1154 bitmap = bitmap_bdescr->start;
1156 debugTrace(DEBUG_gc, "bitmap_size: %d, bitmap: %p",
1157 bitmap_size, bitmap);
1159 // don't forget to fill it with zeros!
1160 memset(bitmap, 0, bitmap_size);
1162 // For each block in this step, point to its bitmap from the
1163 // block descriptor.
1164 for (bd=stp->old_blocks; bd != NULL; bd = bd->link) {
1165 bd->u.bitmap = bitmap;
1166 bitmap += BLOCK_SIZE_W / (sizeof(W_)*BITS_PER_BYTE);
1168 // Also at this point we set the BF_COMPACTED flag
1169 // for this block. The invariant is that
1170 // BF_COMPACTED is always unset, except during GC
1171 // when it is set on those blocks which will be
1173 bd->flags |= BF_COMPACTED;
1179 // For each GC thread, for each step, allocate a "todo" block to
1180 // store evacuated objects to be scavenged, and a block to store
1181 // evacuated objects that do not need to be scavenged.
1182 for (t = 0; t < n_threads; t++) {
1183 for (s = 0; s < generations[g].n_steps; s++) {
1185 // we don't copy objects into g0s0, unless -G0
1186 if (g==0 && s==0 && RtsFlags.GcFlags.generations > 1) continue;
1188 ws = &gc_threads[t].steps[g][s];
1193 ws->todo_large_objects = NULL;
1195 // allocate the first to-space block; extra blocks will be
1196 // chained on as necessary.
1198 ws->buffer_todo_bd = NULL;
1199 gc_alloc_todo_block(ws);
1201 ws->scavd_list = NULL;
1202 ws->n_scavd_blocks = 0;
1208 /* ----------------------------------------------------------------------------
1209 Initialise a generation that is *not* to be collected
1210 ------------------------------------------------------------------------- */
1213 init_uncollected_gen (nat g, nat threads)
1220 for (s = 0; s < generations[g].n_steps; s++) {
1221 stp = &generations[g].steps[s];
1222 stp->scavenged_large_objects = NULL;
1223 stp->n_scavenged_large_blocks = 0;
1226 for (t = 0; t < threads; t++) {
1227 for (s = 0; s < generations[g].n_steps; s++) {
1229 ws = &gc_threads[t].steps[g][s];
1232 ws->buffer_todo_bd = NULL;
1233 ws->todo_large_objects = NULL;
1235 ws->scavd_list = NULL;
1236 ws->n_scavd_blocks = 0;
1238 // If the block at the head of the list in this generation
1239 // is less than 3/4 full, then use it as a todo block.
1240 if (isPartiallyFull(stp->blocks))
1242 ws->todo_bd = stp->blocks;
1243 ws->todo_free = ws->todo_bd->free;
1244 ws->todo_lim = ws->todo_bd->start + BLOCK_SIZE_W;
1245 stp->blocks = stp->blocks->link;
1247 ws->todo_bd->link = NULL;
1249 // this block is also the scan block; we must scan
1250 // from the current end point.
1251 ws->scan_bd = ws->todo_bd;
1252 ws->scan = ws->scan_bd->free;
1254 // subtract the contents of this block from the stats,
1255 // because we'll count the whole block later.
1256 copied -= ws->scan_bd->free - ws->scan_bd->start;
1263 gc_alloc_todo_block(ws);
1268 // Move the private mutable lists from each capability onto the
1269 // main mutable list for the generation.
1270 for (i = 0; i < n_capabilities; i++) {
1271 for (bd = capabilities[i].mut_lists[g];
1272 bd->link != NULL; bd = bd->link) {
1275 bd->link = generations[g].mut_list;
1276 generations[g].mut_list = capabilities[i].mut_lists[g];
1277 capabilities[i].mut_lists[g] = allocBlock();
1281 /* -----------------------------------------------------------------------------
1282 Initialise a gc_thread before GC
1283 -------------------------------------------------------------------------- */
1286 init_gc_thread (gc_thread *t)
1289 t->failed_to_evac = rtsFalse;
1290 t->eager_promotion = rtsTrue;
1291 t->thunk_selector_depth = 0;
1294 /* -----------------------------------------------------------------------------
1295 Function we pass to GetRoots to evacuate roots.
1296 -------------------------------------------------------------------------- */
1299 mark_root(StgClosure **root)
1304 /* -----------------------------------------------------------------------------
1305 Initialising the static object & mutable lists
1306 -------------------------------------------------------------------------- */
1309 zero_static_object_list(StgClosure* first_static)
1313 const StgInfoTable *info;
1315 for (p = first_static; p != END_OF_STATIC_LIST; p = link) {
1317 link = *STATIC_LINK(info, p);
1318 *STATIC_LINK(info,p) = NULL;
1322 /* -----------------------------------------------------------------------------
1324 -------------------------------------------------------------------------- */
1331 for (c = (StgIndStatic *)revertible_caf_list; c != NULL;
1332 c = (StgIndStatic *)c->static_link)
1334 SET_INFO(c, c->saved_info);
1335 c->saved_info = NULL;
1336 // could, but not necessary: c->static_link = NULL;
1338 revertible_caf_list = NULL;
1342 markCAFs( evac_fn evac )
1346 for (c = (StgIndStatic *)caf_list; c != NULL;
1347 c = (StgIndStatic *)c->static_link)
1349 evac(&c->indirectee);
1351 for (c = (StgIndStatic *)revertible_caf_list; c != NULL;
1352 c = (StgIndStatic *)c->static_link)
1354 evac(&c->indirectee);
1358 /* ----------------------------------------------------------------------------
1359 Update the pointers from the task list
1361 These are treated as weak pointers because we want to allow a main
1362 thread to get a BlockedOnDeadMVar exception in the same way as any
1363 other thread. Note that the threads should all have been retained
1364 by GC by virtue of being on the all_threads list, we're just
1365 updating pointers here.
1366 ------------------------------------------------------------------------- */
1369 update_task_list (void)
1373 for (task = all_tasks; task != NULL; task = task->all_link) {
1374 if (!task->stopped && task->tso) {
1375 ASSERT(task->tso->bound == task);
1376 tso = (StgTSO *) isAlive((StgClosure *)task->tso);
1378 barf("task %p: main thread %d has been GC'd",
1391 /* ----------------------------------------------------------------------------
1392 Reset the sizes of the older generations when we do a major
1395 CURRENT STRATEGY: make all generations except zero the same size.
1396 We have to stay within the maximum heap size, and leave a certain
1397 percentage of the maximum heap size available to allocate into.
1398 ------------------------------------------------------------------------- */
1401 resize_generations (void)
1405 if (major_gc && RtsFlags.GcFlags.generations > 1) {
1406 nat live, size, min_alloc;
1407 nat max = RtsFlags.GcFlags.maxHeapSize;
1408 nat gens = RtsFlags.GcFlags.generations;
1410 // live in the oldest generations
1411 live = oldest_gen->steps[0].n_blocks +
1412 oldest_gen->steps[0].n_large_blocks;
1414 // default max size for all generations except zero
1415 size = stg_max(live * RtsFlags.GcFlags.oldGenFactor,
1416 RtsFlags.GcFlags.minOldGenSize);
1418 // minimum size for generation zero
1419 min_alloc = stg_max((RtsFlags.GcFlags.pcFreeHeap * max) / 200,
1420 RtsFlags.GcFlags.minAllocAreaSize);
1422 // Auto-enable compaction when the residency reaches a
1423 // certain percentage of the maximum heap size (default: 30%).
1424 if (RtsFlags.GcFlags.generations > 1 &&
1425 (RtsFlags.GcFlags.compact ||
1427 oldest_gen->steps[0].n_blocks >
1428 (RtsFlags.GcFlags.compactThreshold * max) / 100))) {
1429 oldest_gen->steps[0].is_compacted = 1;
1430 // debugBelch("compaction: on\n", live);
1432 oldest_gen->steps[0].is_compacted = 0;
1433 // debugBelch("compaction: off\n", live);
1436 // if we're going to go over the maximum heap size, reduce the
1437 // size of the generations accordingly. The calculation is
1438 // different if compaction is turned on, because we don't need
1439 // to double the space required to collect the old generation.
1442 // this test is necessary to ensure that the calculations
1443 // below don't have any negative results - we're working
1444 // with unsigned values here.
1445 if (max < min_alloc) {
1449 if (oldest_gen->steps[0].is_compacted) {
1450 if ( (size + (size - 1) * (gens - 2) * 2) + min_alloc > max ) {
1451 size = (max - min_alloc) / ((gens - 1) * 2 - 1);
1454 if ( (size * (gens - 1) * 2) + min_alloc > max ) {
1455 size = (max - min_alloc) / ((gens - 1) * 2);
1465 debugBelch("live: %d, min_alloc: %d, size : %d, max = %d\n", live,
1466 min_alloc, size, max);
1469 for (g = 0; g < gens; g++) {
1470 generations[g].max_blocks = size;
1475 /* -----------------------------------------------------------------------------
1476 Calculate the new size of the nursery, and resize it.
1477 -------------------------------------------------------------------------- */
1480 resize_nursery (void)
1482 if (RtsFlags.GcFlags.generations == 1)
1483 { // Two-space collector:
1486 /* set up a new nursery. Allocate a nursery size based on a
1487 * function of the amount of live data (by default a factor of 2)
1488 * Use the blocks from the old nursery if possible, freeing up any
1491 * If we get near the maximum heap size, then adjust our nursery
1492 * size accordingly. If the nursery is the same size as the live
1493 * data (L), then we need 3L bytes. We can reduce the size of the
1494 * nursery to bring the required memory down near 2L bytes.
1496 * A normal 2-space collector would need 4L bytes to give the same
1497 * performance we get from 3L bytes, reducing to the same
1498 * performance at 2L bytes.
1500 blocks = g0s0->n_old_blocks;
1502 if ( RtsFlags.GcFlags.maxHeapSize != 0 &&
1503 blocks * RtsFlags.GcFlags.oldGenFactor * 2 >
1504 RtsFlags.GcFlags.maxHeapSize )
1506 long adjusted_blocks; // signed on purpose
1509 adjusted_blocks = (RtsFlags.GcFlags.maxHeapSize - 2 * blocks);
1511 debugTrace(DEBUG_gc, "near maximum heap size of 0x%x blocks, blocks = %d, adjusted to %ld",
1512 RtsFlags.GcFlags.maxHeapSize, blocks, adjusted_blocks);
1514 pc_free = adjusted_blocks * 100 / RtsFlags.GcFlags.maxHeapSize;
1515 if (pc_free < RtsFlags.GcFlags.pcFreeHeap) /* might even * be < 0 */
1519 blocks = adjusted_blocks;
1523 blocks *= RtsFlags.GcFlags.oldGenFactor;
1524 if (blocks < RtsFlags.GcFlags.minAllocAreaSize)
1526 blocks = RtsFlags.GcFlags.minAllocAreaSize;
1529 resizeNurseries(blocks);
1531 else // Generational collector
1534 * If the user has given us a suggested heap size, adjust our
1535 * allocation area to make best use of the memory available.
1537 if (RtsFlags.GcFlags.heapSizeSuggestion)
1540 nat needed = calcNeeded(); // approx blocks needed at next GC
1542 /* Guess how much will be live in generation 0 step 0 next time.
1543 * A good approximation is obtained by finding the
1544 * percentage of g0s0 that was live at the last minor GC.
1546 * We have an accurate figure for the amount of copied data in
1547 * 'copied', but we must convert this to a number of blocks, with
1548 * a small adjustment for estimated slop at the end of a block
1553 g0s0_pcnt_kept = ((copied / (BLOCK_SIZE_W - 10)) * 100)
1554 / countNurseryBlocks();
1557 /* Estimate a size for the allocation area based on the
1558 * information available. We might end up going slightly under
1559 * or over the suggested heap size, but we should be pretty
1562 * Formula: suggested - needed
1563 * ----------------------------
1564 * 1 + g0s0_pcnt_kept/100
1566 * where 'needed' is the amount of memory needed at the next
1567 * collection for collecting all steps except g0s0.
1570 (((long)RtsFlags.GcFlags.heapSizeSuggestion - (long)needed) * 100) /
1571 (100 + (long)g0s0_pcnt_kept);
1573 if (blocks < (long)RtsFlags.GcFlags.minAllocAreaSize) {
1574 blocks = RtsFlags.GcFlags.minAllocAreaSize;
1577 resizeNurseries((nat)blocks);
1581 // we might have added extra large blocks to the nursery, so
1582 // resize back to minAllocAreaSize again.
1583 resizeNurseriesFixed(RtsFlags.GcFlags.minAllocAreaSize);
1588 /* -----------------------------------------------------------------------------
1589 Sanity code for CAF garbage collection.
1591 With DEBUG turned on, we manage a CAF list in addition to the SRT
1592 mechanism. After GC, we run down the CAF list and blackhole any
1593 CAFs which have been garbage collected. This means we get an error
1594 whenever the program tries to enter a garbage collected CAF.
1596 Any garbage collected CAFs are taken off the CAF list at the same
1598 -------------------------------------------------------------------------- */
1600 #if 0 && defined(DEBUG)
1607 const StgInfoTable *info;
1618 ASSERT(info->type == IND_STATIC);
1620 if (STATIC_LINK(info,p) == NULL) {
1621 debugTrace(DEBUG_gccafs, "CAF gc'd at 0x%04lx", (long)p);
1623 SET_INFO(p,&stg_BLACKHOLE_info);
1624 p = STATIC_LINK2(info,p);
1628 pp = &STATIC_LINK2(info,p);
1635 debugTrace(DEBUG_gccafs, "%d CAFs live", i);