1 /* -----------------------------------------------------------------------------
3 * (c) The GHC Team 1998-2006
5 * Generational garbage collector
7 * Documentation on the architecture of the Garbage Collector can be
8 * found in the online commentary:
10 * http://hackage.haskell.org/trac/ghc/wiki/Commentary/Rts/Storage/GC
12 * ---------------------------------------------------------------------------*/
14 // #include "PosixSource.h"
19 #include "OSThreads.h"
20 #include "LdvProfile.h"
25 #include "BlockAlloc.h"
31 #include "ParTicky.h" // ToDo: move into Rts.h
32 #include "RtsSignals.h"
36 #if defined(RTS_GTK_FRONTPANEL)
37 #include "FrontPanel.h"
40 #include "RetainerProfile.h"
41 #include "RaiseAsync.h"
53 #include <string.h> // for memset()
56 /* -----------------------------------------------------------------------------
58 -------------------------------------------------------------------------- */
60 /* STATIC OBJECT LIST.
63 * We maintain a linked list of static objects that are still live.
64 * The requirements for this list are:
66 * - we need to scan the list while adding to it, in order to
67 * scavenge all the static objects (in the same way that
68 * breadth-first scavenging works for dynamic objects).
70 * - we need to be able to tell whether an object is already on
71 * the list, to break loops.
73 * Each static object has a "static link field", which we use for
74 * linking objects on to the list. We use a stack-type list, consing
75 * objects on the front as they are added (this means that the
76 * scavenge phase is depth-first, not breadth-first, but that
79 * A separate list is kept for objects that have been scavenged
80 * already - this is so that we can zero all the marks afterwards.
82 * An object is on the list if its static link field is non-zero; this
83 * means that we have to mark the end of the list with '1', not NULL.
85 * Extra notes for generational GC:
87 * Each generation has a static object list associated with it. When
88 * collecting generations up to N, we treat the static object lists
89 * from generations > N as roots.
91 * We build up a static object list while collecting generations 0..N,
92 * which is then appended to the static object list of generation N+1.
95 /* N is the oldest generation being collected, where the generations
96 * are numbered starting at 0. A major GC (indicated by the major_gc
97 * flag) is when we're collecting all generations. We only attempt to
98 * deal with static objects and GC CAFs when doing a major GC.
103 /* Data used for allocation area sizing.
105 static lnat g0s0_pcnt_kept = 30; // percentage of g0s0 live at last minor GC
115 /* Thread-local data for each GC thread
117 gc_thread **gc_threads = NULL;
118 // gc_thread *gct = NULL; // this thread's gct TODO: make thread-local
120 // Number of threads running in *this* GC. Affects how many
121 // step->todos[] lists we have to look in to find work.
125 long copied; // *words* copied & scavenged during this GC
128 SpinLock recordMutableGen_sync;
131 /* -----------------------------------------------------------------------------
132 Static function declarations
133 -------------------------------------------------------------------------- */
135 static void mark_root (StgClosure **root);
136 static void zero_static_object_list (StgClosure* first_static);
137 static nat initialise_N (rtsBool force_major_gc);
138 static void alloc_gc_threads (void);
139 static void init_collected_gen (nat g, nat threads);
140 static void init_uncollected_gen (nat g, nat threads);
141 static void init_gc_thread (gc_thread *t);
142 static void update_task_list (void);
143 static void resize_generations (void);
144 static void resize_nursery (void);
145 static void start_gc_threads (void);
146 static void gc_thread_work (void);
147 static nat inc_running (void);
148 static nat dec_running (void);
149 static void wakeup_gc_threads (nat n_threads);
150 static void shutdown_gc_threads (nat n_threads);
152 #if 0 && defined(DEBUG)
153 static void gcCAFs (void);
156 /* -----------------------------------------------------------------------------
157 The mark bitmap & stack.
158 -------------------------------------------------------------------------- */
160 #define MARK_STACK_BLOCKS 4
162 bdescr *mark_stack_bdescr;
167 // Flag and pointers used for falling back to a linear scan when the
168 // mark stack overflows.
169 rtsBool mark_stack_overflowed;
170 bdescr *oldgen_scan_bd;
173 /* -----------------------------------------------------------------------------
174 GarbageCollect: the main entry point to the garbage collector.
176 Locks held: all capabilities are held throughout GarbageCollect().
177 -------------------------------------------------------------------------- */
180 GarbageCollect ( rtsBool force_major_gc )
184 lnat live, allocated, max_copied, avg_copied;
185 lnat oldgen_saved_blocks = 0;
186 gc_thread *saved_gct;
189 // necessary if we stole a callee-saves register for gct:
193 CostCentreStack *prev_CCS;
198 #if defined(RTS_USER_SIGNALS)
199 if (RtsFlags.MiscFlags.install_signal_handlers) {
205 // tell the stats department that we've started a GC
208 // tell the STM to discard any cached closures it's hoping to re-use
217 // attribute any costs to CCS_GC
223 /* Approximate how much we allocated.
224 * Todo: only when generating stats?
226 allocated = calcAllocated();
228 /* Figure out which generation to collect
230 n = initialise_N(force_major_gc);
232 /* Allocate + initialise the gc_thread structures.
236 /* Start threads, so they can be spinning up while we finish initialisation.
240 /* How many threads will be participating in this GC?
241 * We don't try to parallelise minor GC.
243 #if defined(THREADED_RTS)
244 if (n < (4*1024*1024 / BLOCK_SIZE)) {
247 n_gc_threads = RtsFlags.ParFlags.gcThreads;
252 trace(TRACE_gc|DEBUG_gc, "GC (gen %d): %dKB to collect, using %d thread(s)",
253 N, n * (BLOCK_SIZE / 1024), n_gc_threads);
255 #ifdef RTS_GTK_FRONTPANEL
256 if (RtsFlags.GcFlags.frontpanel) {
257 updateFrontPanelBeforeGC(N);
262 // check for memory leaks if DEBUG is on
263 memInventory(traceClass(DEBUG_gc));
266 // check stack sanity *before* GC (ToDo: check all threads)
267 IF_DEBUG(sanity, checkFreeListSanity());
269 // Initialise all our gc_thread structures
270 for (t = 0; t < n_gc_threads; t++) {
271 init_gc_thread(gc_threads[t]);
274 // Initialise all the generations/steps that we're collecting.
275 for (g = 0; g <= N; g++) {
276 init_collected_gen(g,n_gc_threads);
279 // Initialise all the generations/steps that we're *not* collecting.
280 for (g = N+1; g < RtsFlags.GcFlags.generations; g++) {
281 init_uncollected_gen(g,n_gc_threads);
284 /* Allocate a mark stack if we're doing a major collection.
287 mark_stack_bdescr = allocGroup(MARK_STACK_BLOCKS);
288 mark_stack = (StgPtr *)mark_stack_bdescr->start;
289 mark_sp = mark_stack;
290 mark_splim = mark_stack + (MARK_STACK_BLOCKS * BLOCK_SIZE_W);
292 mark_stack_bdescr = NULL;
295 // this is the main thread
298 /* -----------------------------------------------------------------------
299 * follow all the roots that we know about:
300 * - mutable lists from each generation > N
301 * we want to *scavenge* these roots, not evacuate them: they're not
302 * going to move in this GC.
303 * Also do them in reverse generation order, for the usual reason:
304 * namely to reduce the likelihood of spurious old->new pointers.
306 for (g = RtsFlags.GcFlags.generations-1; g > N; g--) {
307 generations[g].saved_mut_list = generations[g].mut_list;
308 generations[g].mut_list = allocBlock();
309 // mut_list always has at least one block.
312 // the main thread is running: this prevents any other threads from
313 // exiting prematurely, so we can start them now.
314 // NB. do this after the mutable lists have been saved above, otherwise
315 // the other GC threads will be writing into the old mutable lists.
317 wakeup_gc_threads(n_gc_threads);
319 for (g = RtsFlags.GcFlags.generations-1; g > N; g--) {
320 scavenge_mutable_list(&generations[g]);
323 // follow roots from the CAF list (used by GHCi)
327 // follow all the roots that the application knows about.
331 #if defined(RTS_USER_SIGNALS)
332 // mark the signal handlers (signals should be already blocked)
333 markSignalHandlers(mark_root);
336 // Mark the weak pointer list, and prepare to detect dead weak pointers.
340 // Mark the stable pointer table.
341 markStablePtrTable(mark_root);
343 /* -------------------------------------------------------------------------
344 * Repeatedly scavenge all the areas we know about until there's no
345 * more scavenging to be done.
350 // The other threads are now stopped. We might recurse back to
351 // here, but from now on this is the only thread.
353 // if any blackholes are alive, make the threads that wait on
355 if (traverseBlackholeQueue()) {
360 // must be last... invariant is that everything is fully
361 // scavenged at this point.
362 if (traverseWeakPtrList()) { // returns rtsTrue if evaced something
367 // If we get to here, there's really nothing left to do.
371 shutdown_gc_threads(n_gc_threads);
373 // Update pointers from the Task list
376 // Now see which stable names are still alive.
380 // We call processHeapClosureForDead() on every closure destroyed during
381 // the current garbage collection, so we invoke LdvCensusForDead().
382 if (RtsFlags.ProfFlags.doHeapProfile == HEAP_BY_LDV
383 || RtsFlags.ProfFlags.bioSelector != NULL)
387 // NO MORE EVACUATION AFTER THIS POINT!
388 // Finally: compaction of the oldest generation.
389 if (major_gc && oldest_gen->steps[0].is_compacted) {
390 // save number of blocks for stats
391 oldgen_saved_blocks = oldest_gen->steps[0].n_old_blocks;
395 IF_DEBUG(sanity, checkGlobalTSOList(rtsFalse));
397 // Two-space collector: free the old to-space.
398 // g0s0->old_blocks is the old nursery
399 // g0s0->blocks is to-space from the previous GC
400 if (RtsFlags.GcFlags.generations == 1) {
401 if (g0s0->blocks != NULL) {
402 freeChain(g0s0->blocks);
407 // For each workspace, in each thread:
408 // * clear the BF_EVACUATED flag from each copied block
409 // * move the copied blocks to the step
415 for (t = 0; t < n_gc_threads; t++) {
419 for (s = 1; s < total_steps; s++) {
422 // Push the final block
424 push_scanned_block(ws->todo_bd, ws);
427 ASSERT(gct->scan_bd == NULL);
428 ASSERT(countBlocks(ws->scavd_list) == ws->n_scavd_blocks);
430 prev = ws->part_list;
431 for (bd = ws->part_list; bd != NULL; bd = bd->link) {
432 bd->flags &= ~BF_EVACUATED; // now from-space
433 ws->step->n_words += bd->free - bd->start;
437 prev->link = ws->scavd_list;
439 for (bd = ws->scavd_list; bd != NULL; bd = bd->link) {
440 bd->flags &= ~BF_EVACUATED; // now from-space
441 ws->step->n_words += bd->free - bd->start;
444 prev->link = ws->step->blocks;
445 if (ws->part_list != NULL) {
446 ws->step->blocks = ws->part_list;
448 ws->step->blocks = ws->scavd_list;
450 ws->step->n_blocks += ws->n_part_blocks;
451 ws->step->n_blocks += ws->n_scavd_blocks;
452 ASSERT(countBlocks(ws->step->blocks) == ws->step->n_blocks);
453 ASSERT(countOccupied(ws->step->blocks) == ws->step->n_words);
458 // Two-space collector: swap the semi-spaces around.
459 // Currently: g0s0->old_blocks is the old nursery
460 // g0s0->blocks is to-space from this GC
461 // We want these the other way around.
462 if (RtsFlags.GcFlags.generations == 1) {
463 bdescr *nursery_blocks = g0s0->old_blocks;
464 nat n_nursery_blocks = g0s0->n_old_blocks;
465 g0s0->old_blocks = g0s0->blocks;
466 g0s0->n_old_blocks = g0s0->n_blocks;
467 g0s0->blocks = nursery_blocks;
468 g0s0->n_blocks = n_nursery_blocks;
471 /* run through all the generations/steps and tidy up
478 for (i=0; i < n_gc_threads; i++) {
479 if (n_gc_threads > 1) {
480 trace(TRACE_gc,"thread %d:", i);
481 trace(TRACE_gc," copied %ld", gc_threads[i]->copied * sizeof(W_));
482 trace(TRACE_gc," scanned %ld", gc_threads[i]->scanned * sizeof(W_));
483 trace(TRACE_gc," any_work %ld", gc_threads[i]->any_work);
484 trace(TRACE_gc," no_work %ld", gc_threads[i]->no_work);
485 trace(TRACE_gc," scav_find_work %ld", gc_threads[i]->scav_find_work);
487 copied += gc_threads[i]->copied;
488 max_copied = stg_max(gc_threads[i]->copied, max_copied);
490 if (n_gc_threads == 1) {
498 for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
501 generations[g].collections++; // for stats
502 if (n_gc_threads > 1) generations[g].par_collections++;
505 // Count the mutable list as bytes "copied" for the purposes of
506 // stats. Every mutable list is copied during every GC.
508 nat mut_list_size = 0;
509 for (bd = generations[g].mut_list; bd != NULL; bd = bd->link) {
510 mut_list_size += bd->free - bd->start;
512 copied += mut_list_size;
515 "mut_list_size: %lu (%d vars, %d arrays, %d MVARs, %d others)",
516 (unsigned long)(mut_list_size * sizeof(W_)),
517 mutlist_MUTVARS, mutlist_MUTARRS, mutlist_MVARS, mutlist_OTHERS);
520 for (s = 0; s < generations[g].n_steps; s++) {
522 stp = &generations[g].steps[s];
524 // for generations we collected...
527 /* free old memory and shift to-space into from-space for all
528 * the collected steps (except the allocation area). These
529 * freed blocks will probaby be quickly recycled.
531 if (!(g == 0 && s == 0 && RtsFlags.GcFlags.generations > 1)) {
532 if (stp->is_compacted)
534 // for a compacted step, just shift the new to-space
535 // onto the front of the now-compacted existing blocks.
536 for (bd = stp->blocks; bd != NULL; bd = bd->link) {
537 bd->flags &= ~BF_EVACUATED; // now from-space
538 stp->n_words += bd->free - bd->start;
540 // tack the new blocks on the end of the existing blocks
541 if (stp->old_blocks != NULL) {
542 for (bd = stp->old_blocks; bd != NULL; bd = next) {
543 // NB. this step might not be compacted next
544 // time, so reset the BF_COMPACTED flags.
545 // They are set before GC if we're going to
546 // compact. (search for BF_COMPACTED above).
547 bd->flags &= ~BF_COMPACTED;
550 bd->link = stp->blocks;
553 stp->blocks = stp->old_blocks;
555 // add the new blocks to the block tally
556 stp->n_blocks += stp->n_old_blocks;
557 ASSERT(countBlocks(stp->blocks) == stp->n_blocks);
558 ASSERT(countOccupied(stp->blocks) == stp->n_words);
562 freeChain(stp->old_blocks);
564 stp->old_blocks = NULL;
565 stp->n_old_blocks = 0;
568 /* LARGE OBJECTS. The current live large objects are chained on
569 * scavenged_large, having been moved during garbage
570 * collection from large_objects. Any objects left on
571 * large_objects list are therefore dead, so we free them here.
573 for (bd = stp->large_objects; bd != NULL; bd = next) {
579 // update the count of blocks used by large objects
580 for (bd = stp->scavenged_large_objects; bd != NULL; bd = bd->link) {
581 bd->flags &= ~BF_EVACUATED;
583 stp->large_objects = stp->scavenged_large_objects;
584 stp->n_large_blocks = stp->n_scavenged_large_blocks;
587 else // for older generations...
589 /* For older generations, we need to append the
590 * scavenged_large_object list (i.e. large objects that have been
591 * promoted during this GC) to the large_object list for that step.
593 for (bd = stp->scavenged_large_objects; bd; bd = next) {
595 bd->flags &= ~BF_EVACUATED;
596 dbl_link_onto(bd, &stp->large_objects);
599 // add the new blocks we promoted during this GC
600 stp->n_large_blocks += stp->n_scavenged_large_blocks;
605 // update the max size of older generations after a major GC
606 resize_generations();
608 // Calculate the amount of live data for stats.
609 live = calcLiveWords();
611 // Free the small objects allocated via allocate(), since this will
612 // all have been copied into G0S1 now.
613 if (RtsFlags.GcFlags.generations > 1) {
614 if (g0s0->blocks != NULL) {
615 freeChain(g0s0->blocks);
622 alloc_blocks_lim = RtsFlags.GcFlags.minAllocAreaSize;
624 // Start a new pinned_object_block
625 pinned_object_block = NULL;
627 // Free the mark stack.
628 if (mark_stack_bdescr != NULL) {
629 freeGroup(mark_stack_bdescr);
633 for (g = 0; g <= N; g++) {
634 for (s = 0; s < generations[g].n_steps; s++) {
635 stp = &generations[g].steps[s];
636 if (stp->bitmap != NULL) {
637 freeGroup(stp->bitmap);
645 // mark the garbage collected CAFs as dead
646 #if 0 && defined(DEBUG) // doesn't work at the moment
647 if (major_gc) { gcCAFs(); }
651 // resetStaticObjectForRetainerProfiling() must be called before
653 if (n_gc_threads > 1) {
654 barf("profiling is currently broken with multi-threaded GC");
655 // ToDo: fix the gct->scavenged_static_objects below
657 resetStaticObjectForRetainerProfiling(gct->scavenged_static_objects);
660 // zero the scavenged static object list
663 for (i = 0; i < n_gc_threads; i++) {
664 zero_static_object_list(gc_threads[i]->scavenged_static_objects);
671 // start any pending finalizers
673 scheduleFinalizers(last_free_capability, old_weak_ptr_list);
676 // send exceptions to any threads which were about to die
678 resurrectThreads(resurrected_threads);
681 // Update the stable pointer hash table.
682 updateStablePtrTable(major_gc);
684 // check sanity after GC
685 IF_DEBUG(sanity, checkSanity());
687 // extra GC trace info
688 if (traceClass(TRACE_gc|DEBUG_gc)) statDescribeGens();
691 // symbol-table based profiling
692 /* heapCensus(to_blocks); */ /* ToDo */
695 // restore enclosing cost centre
701 // check for memory leaks if DEBUG is on
702 memInventory(traceClass(DEBUG_gc));
705 #ifdef RTS_GTK_FRONTPANEL
706 if (RtsFlags.GcFlags.frontpanel) {
707 updateFrontPanelAfterGC( N, live );
711 // ok, GC over: tell the stats department what happened.
712 stat_endGC(allocated, live, copied, N, max_copied, avg_copied);
714 #if defined(RTS_USER_SIGNALS)
715 if (RtsFlags.MiscFlags.install_signal_handlers) {
716 // unblock signals again
717 unblockUserSignals();
726 /* -----------------------------------------------------------------------------
727 * Mark all nodes pointed to by sparks in the spark queues (for GC) Does an
728 * implicit slide i.e. after marking all sparks are at the beginning of the
729 * spark pool and the spark pool only contains sparkable closures
730 * -------------------------------------------------------------------------- */
734 markSparkQueue (evac_fn evac, Capability *cap)
736 StgClosure **sparkp, **to_sparkp;
737 nat n, pruned_sparks; // stats only
740 PAR_TICKY_MARK_SPARK_QUEUE_START();
745 pool = &(cap->r.rSparks);
747 ASSERT_SPARK_POOL_INVARIANTS(pool);
749 #if defined(PARALLEL_HASKELL)
756 to_sparkp = pool->hd;
757 while (sparkp != pool->tl) {
758 ASSERT(*sparkp!=NULL);
759 ASSERT(LOOKS_LIKE_CLOSURE_PTR(((StgClosure *)*sparkp)));
760 // ToDo?: statistics gathering here (also for GUM!)
761 if (closure_SHOULD_SPARK(*sparkp)) {
763 *to_sparkp++ = *sparkp;
764 if (to_sparkp == pool->lim) {
765 to_sparkp = pool->base;
772 if (sparkp == pool->lim) {
776 pool->tl = to_sparkp;
778 PAR_TICKY_MARK_SPARK_QUEUE_END(n);
780 #if defined(PARALLEL_HASKELL)
781 debugTrace(DEBUG_sched,
782 "marked %d sparks and pruned %d sparks on [%x]",
783 n, pruned_sparks, mytid);
785 debugTrace(DEBUG_sched,
786 "marked %d sparks and pruned %d sparks",
790 debugTrace(DEBUG_sched,
791 "new spark queue len=%d; (hd=%p; tl=%p)\n",
792 sparkPoolSize(pool), pool->hd, pool->tl);
796 /* ---------------------------------------------------------------------------
797 Where are the roots that we know about?
799 - all the threads on the runnable queue
800 - all the threads on the blocked queue
801 - all the threads on the sleeping queue
802 - all the thread currently executing a _ccall_GC
803 - all the "main threads"
805 ------------------------------------------------------------------------ */
808 GetRoots( evac_fn evac )
814 // Each GC thread is responsible for following roots from the
815 // Capability of the same number. There will usually be the same
816 // or fewer Capabilities as GC threads, but just in case there
817 // are more, we mark every Capability whose number is the GC
818 // thread's index plus a multiple of the number of GC threads.
819 for (i = gct->thread_index; i < n_capabilities; i += n_gc_threads) {
820 cap = &capabilities[i];
821 evac((StgClosure **)(void *)&cap->run_queue_hd);
822 evac((StgClosure **)(void *)&cap->run_queue_tl);
823 #if defined(THREADED_RTS)
824 evac((StgClosure **)(void *)&cap->wakeup_queue_hd);
825 evac((StgClosure **)(void *)&cap->wakeup_queue_tl);
827 for (task = cap->suspended_ccalling_tasks; task != NULL;
829 debugTrace(DEBUG_sched,
830 "evac'ing suspended TSO %lu", (unsigned long)task->suspended_tso->id);
831 evac((StgClosure **)(void *)&task->suspended_tso);
834 #if defined(THREADED_RTS)
835 markSparkQueue(evac,cap);
839 #if !defined(THREADED_RTS)
840 evac((StgClosure **)(void *)&blocked_queue_hd);
841 evac((StgClosure **)(void *)&blocked_queue_tl);
842 evac((StgClosure **)(void *)&sleeping_queue);
846 /* -----------------------------------------------------------------------------
847 isAlive determines whether the given closure is still alive (after
848 a garbage collection) or not. It returns the new address of the
849 closure if it is alive, or NULL otherwise.
851 NOTE: Use it before compaction only!
852 It untags and (if needed) retags pointers to closures.
853 -------------------------------------------------------------------------- */
857 isAlive(StgClosure *p)
859 const StgInfoTable *info;
865 /* The tag and the pointer are split, to be merged later when needed. */
866 tag = GET_CLOSURE_TAG(p);
867 q = UNTAG_CLOSURE(p);
869 ASSERT(LOOKS_LIKE_CLOSURE_PTR(q));
872 // ignore static closures
874 // ToDo: for static closures, check the static link field.
875 // Problem here is that we sometimes don't set the link field, eg.
876 // for static closures with an empty SRT or CONSTR_STATIC_NOCAFs.
878 if (!HEAP_ALLOCED(q)) {
882 // ignore closures in generations that we're not collecting.
884 if (bd->gen_no > N) {
888 // if it's a pointer into to-space, then we're done
889 if (bd->flags & BF_EVACUATED) {
893 // large objects use the evacuated flag
894 if (bd->flags & BF_LARGE) {
898 // check the mark bit for compacted steps
899 if ((bd->flags & BF_COMPACTED) && is_marked((P_)q,bd)) {
903 switch (info->type) {
908 case IND_OLDGEN: // rely on compatible layout with StgInd
909 case IND_OLDGEN_PERM:
910 // follow indirections
911 p = ((StgInd *)q)->indirectee;
916 return ((StgEvacuated *)q)->evacuee;
919 if (((StgTSO *)q)->what_next == ThreadRelocated) {
920 p = (StgClosure *)((StgTSO *)q)->link;
932 /* -----------------------------------------------------------------------------
933 Figure out which generation to collect, initialise N and major_gc.
935 Also returns the total number of blocks in generations that will be
937 -------------------------------------------------------------------------- */
940 initialise_N (rtsBool force_major_gc)
943 nat s, blocks, blocks_total;
948 if (force_major_gc) {
949 N = RtsFlags.GcFlags.generations - 1;
954 for (g = RtsFlags.GcFlags.generations - 1; g >= 0; g--) {
956 for (s = 0; s < generations[g].n_steps; s++) {
957 blocks += generations[g].steps[s].n_words / BLOCK_SIZE_W;
958 blocks += generations[g].steps[s].n_large_blocks;
960 if (blocks >= generations[g].max_blocks) {
964 blocks_total += blocks;
968 blocks_total += countNurseryBlocks();
970 major_gc = (N == RtsFlags.GcFlags.generations-1);
974 /* -----------------------------------------------------------------------------
975 Initialise the gc_thread structures.
976 -------------------------------------------------------------------------- */
979 alloc_gc_thread (int n)
985 t = stgMallocBytes(sizeof(gc_thread) + total_steps * sizeof(step_workspace),
990 initCondition(&t->wake_cond);
991 initMutex(&t->wake_mutex);
992 t->wakeup = rtsTrue; // starts true, so we can wait for the
993 // thread to start up, see wakeup_gc_threads
998 t->free_blocks = NULL;
1004 t->papi_events = -1;
1007 for (s = 0; s < total_steps; s++)
1010 ws->step = &all_steps[s];
1011 ASSERT(s == ws->step->abs_no);
1015 ws->buffer_todo_bd = NULL;
1017 ws->part_list = NULL;
1018 ws->n_part_blocks = 0;
1020 ws->scavd_list = NULL;
1021 ws->n_scavd_blocks = 0;
1029 alloc_gc_threads (void)
1031 if (gc_threads == NULL) {
1032 #if defined(THREADED_RTS)
1034 gc_threads = stgMallocBytes (RtsFlags.ParFlags.gcThreads *
1036 "alloc_gc_threads");
1038 for (i = 0; i < RtsFlags.ParFlags.gcThreads; i++) {
1039 gc_threads[i] = alloc_gc_thread(i);
1042 gc_threads = stgMallocBytes (sizeof(gc_thread*),
1043 "alloc_gc_threads");
1045 gc_threads[0] = alloc_gc_thread(0);
1050 /* ----------------------------------------------------------------------------
1052 ------------------------------------------------------------------------- */
1054 static nat gc_running_threads;
1056 #if defined(THREADED_RTS)
1057 static Mutex gc_running_mutex;
1064 ACQUIRE_LOCK(&gc_running_mutex);
1065 n_running = ++gc_running_threads;
1066 RELEASE_LOCK(&gc_running_mutex);
1067 ASSERT(n_running <= n_gc_threads);
1075 ACQUIRE_LOCK(&gc_running_mutex);
1076 ASSERT(n_gc_threads != 0);
1077 n_running = --gc_running_threads;
1078 RELEASE_LOCK(&gc_running_mutex);
1083 // gc_thread_work(): Scavenge until there's no work left to do and all
1084 // the running threads are idle.
1087 gc_thread_work (void)
1091 debugTrace(DEBUG_gc, "GC thread %d working", gct->thread_index);
1093 // gc_running_threads has already been incremented for us; either
1094 // this is the main thread and we incremented it inside
1095 // GarbageCollect(), or this is a worker thread and the main
1096 // thread bumped gc_running_threads before waking us up.
1098 // Every thread evacuates some roots.
1100 GetRoots(mark_root);
1104 // scavenge_loop() only exits when there's no work to do
1107 debugTrace(DEBUG_gc, "GC thread %d idle (%d still running)",
1108 gct->thread_index, r);
1110 while (gc_running_threads != 0) {
1116 // any_work() does not remove the work from the queue, it
1117 // just checks for the presence of work. If we find any,
1118 // then we increment gc_running_threads and go back to
1119 // scavenge_loop() to perform any pending work.
1122 // All threads are now stopped
1123 debugTrace(DEBUG_gc, "GC thread %d finished.", gct->thread_index);
1127 #if defined(THREADED_RTS)
1129 gc_thread_mainloop (void)
1131 while (!gct->exit) {
1133 // Wait until we're told to wake up
1134 ACQUIRE_LOCK(&gct->wake_mutex);
1135 gct->wakeup = rtsFalse;
1136 while (!gct->wakeup) {
1137 debugTrace(DEBUG_gc, "GC thread %d standing by...",
1139 waitCondition(&gct->wake_cond, &gct->wake_mutex);
1141 RELEASE_LOCK(&gct->wake_mutex);
1142 if (gct->exit) break;
1145 // start performance counters in this thread...
1146 if (gct->papi_events == -1) {
1147 papi_init_eventset(&gct->papi_events);
1149 papi_thread_start_gc1_count(gct->papi_events);
1155 // count events in this thread towards the GC totals
1156 papi_thread_stop_gc1_count(gct->papi_events);
1162 #if defined(THREADED_RTS)
1164 gc_thread_entry (gc_thread *my_gct)
1167 debugTrace(DEBUG_gc, "GC thread %d starting...", gct->thread_index);
1168 gct->id = osThreadId();
1169 gc_thread_mainloop();
1174 start_gc_threads (void)
1176 #if defined(THREADED_RTS)
1179 static rtsBool done = rtsFalse;
1181 gc_running_threads = 0;
1182 initMutex(&gc_running_mutex);
1185 // Start from 1: the main thread is 0
1186 for (i = 1; i < RtsFlags.ParFlags.gcThreads; i++) {
1187 createOSThread(&id, (OSThreadProc*)&gc_thread_entry,
1196 wakeup_gc_threads (nat n_threads USED_IF_THREADS)
1198 #if defined(THREADED_RTS)
1200 for (i=1; i < n_threads; i++) {
1202 debugTrace(DEBUG_gc, "waking up gc thread %d", i);
1204 ACQUIRE_LOCK(&gc_threads[i]->wake_mutex);
1205 if (gc_threads[i]->wakeup) {
1206 RELEASE_LOCK(&gc_threads[i]->wake_mutex);
1212 gc_threads[i]->wakeup = rtsTrue;
1213 signalCondition(&gc_threads[i]->wake_cond);
1214 RELEASE_LOCK(&gc_threads[i]->wake_mutex);
1219 // After GC is complete, we must wait for all GC threads to enter the
1220 // standby state, otherwise they may still be executing inside
1221 // any_work(), and may even remain awake until the next GC starts.
1223 shutdown_gc_threads (nat n_threads USED_IF_THREADS)
1225 #if defined(THREADED_RTS)
1228 for (i=1; i < n_threads; i++) {
1230 ACQUIRE_LOCK(&gc_threads[i]->wake_mutex);
1231 wakeup = gc_threads[i]->wakeup;
1232 // wakeup is false while the thread is waiting
1233 RELEASE_LOCK(&gc_threads[i]->wake_mutex);
1239 /* ----------------------------------------------------------------------------
1240 Initialise a generation that is to be collected
1241 ------------------------------------------------------------------------- */
1244 init_collected_gen (nat g, nat n_threads)
1251 // Throw away the current mutable list. Invariant: the mutable
1252 // list always has at least one block; this means we can avoid a
1253 // check for NULL in recordMutable().
1255 freeChain(generations[g].mut_list);
1256 generations[g].mut_list = allocBlock();
1257 for (i = 0; i < n_capabilities; i++) {
1258 freeChain(capabilities[i].mut_lists[g]);
1259 capabilities[i].mut_lists[g] = allocBlock();
1263 for (s = 0; s < generations[g].n_steps; s++) {
1265 // generation 0, step 0 doesn't need to-space
1266 if (g == 0 && s == 0 && RtsFlags.GcFlags.generations > 1) {
1270 stp = &generations[g].steps[s];
1271 ASSERT(stp->gen_no == g);
1273 // deprecate the existing blocks
1274 stp->old_blocks = stp->blocks;
1275 stp->n_old_blocks = stp->n_blocks;
1280 // we don't have any to-be-scavenged blocks yet
1282 stp->todos_last = NULL;
1285 // initialise the large object queues.
1286 stp->scavenged_large_objects = NULL;
1287 stp->n_scavenged_large_blocks = 0;
1289 // mark the large objects as not evacuated yet
1290 for (bd = stp->large_objects; bd; bd = bd->link) {
1291 bd->flags &= ~BF_EVACUATED;
1294 // for a compacted step, we need to allocate the bitmap
1295 if (stp->is_compacted) {
1296 nat bitmap_size; // in bytes
1297 bdescr *bitmap_bdescr;
1300 bitmap_size = stp->n_old_blocks * BLOCK_SIZE / (sizeof(W_)*BITS_PER_BYTE);
1302 if (bitmap_size > 0) {
1303 bitmap_bdescr = allocGroup((lnat)BLOCK_ROUND_UP(bitmap_size)
1305 stp->bitmap = bitmap_bdescr;
1306 bitmap = bitmap_bdescr->start;
1308 debugTrace(DEBUG_gc, "bitmap_size: %d, bitmap: %p",
1309 bitmap_size, bitmap);
1311 // don't forget to fill it with zeros!
1312 memset(bitmap, 0, bitmap_size);
1314 // For each block in this step, point to its bitmap from the
1315 // block descriptor.
1316 for (bd=stp->old_blocks; bd != NULL; bd = bd->link) {
1317 bd->u.bitmap = bitmap;
1318 bitmap += BLOCK_SIZE_W / (sizeof(W_)*BITS_PER_BYTE);
1320 // Also at this point we set the BF_COMPACTED flag
1321 // for this block. The invariant is that
1322 // BF_COMPACTED is always unset, except during GC
1323 // when it is set on those blocks which will be
1325 bd->flags |= BF_COMPACTED;
1331 // For each GC thread, for each step, allocate a "todo" block to
1332 // store evacuated objects to be scavenged, and a block to store
1333 // evacuated objects that do not need to be scavenged.
1334 for (t = 0; t < n_threads; t++) {
1335 for (s = 0; s < generations[g].n_steps; s++) {
1337 // we don't copy objects into g0s0, unless -G0
1338 if (g==0 && s==0 && RtsFlags.GcFlags.generations > 1) continue;
1340 ws = &gc_threads[t]->steps[g * RtsFlags.GcFlags.steps + s];
1342 ws->todo_large_objects = NULL;
1344 ws->part_list = NULL;
1345 ws->n_part_blocks = 0;
1347 // allocate the first to-space block; extra blocks will be
1348 // chained on as necessary.
1350 ws->buffer_todo_bd = NULL;
1351 alloc_todo_block(ws,0);
1353 ws->scavd_list = NULL;
1354 ws->n_scavd_blocks = 0;
1360 /* ----------------------------------------------------------------------------
1361 Initialise a generation that is *not* to be collected
1362 ------------------------------------------------------------------------- */
1365 init_uncollected_gen (nat g, nat threads)
1372 for (s = 0; s < generations[g].n_steps; s++) {
1373 stp = &generations[g].steps[s];
1374 stp->scavenged_large_objects = NULL;
1375 stp->n_scavenged_large_blocks = 0;
1378 for (t = 0; t < threads; t++) {
1379 for (s = 0; s < generations[g].n_steps; s++) {
1381 ws = &gc_threads[t]->steps[g * RtsFlags.GcFlags.steps + s];
1384 ws->buffer_todo_bd = NULL;
1385 ws->todo_large_objects = NULL;
1387 ws->part_list = NULL;
1388 ws->n_part_blocks = 0;
1390 ws->scavd_list = NULL;
1391 ws->n_scavd_blocks = 0;
1393 // If the block at the head of the list in this generation
1394 // is less than 3/4 full, then use it as a todo block.
1395 if (stp->blocks && isPartiallyFull(stp->blocks))
1397 ws->todo_bd = stp->blocks;
1398 ws->todo_free = ws->todo_bd->free;
1399 ws->todo_lim = ws->todo_bd->start + BLOCK_SIZE_W;
1400 stp->blocks = stp->blocks->link;
1402 stp->n_words -= ws->todo_bd->free - ws->todo_bd->start;
1403 ws->todo_bd->link = NULL;
1404 // we must scan from the current end point.
1405 ws->todo_bd->u.scan = ws->todo_bd->free;
1410 alloc_todo_block(ws,0);
1415 // Move the private mutable lists from each capability onto the
1416 // main mutable list for the generation.
1417 for (i = 0; i < n_capabilities; i++) {
1418 for (bd = capabilities[i].mut_lists[g];
1419 bd->link != NULL; bd = bd->link) {
1422 bd->link = generations[g].mut_list;
1423 generations[g].mut_list = capabilities[i].mut_lists[g];
1424 capabilities[i].mut_lists[g] = allocBlock();
1428 /* -----------------------------------------------------------------------------
1429 Initialise a gc_thread before GC
1430 -------------------------------------------------------------------------- */
1433 init_gc_thread (gc_thread *t)
1435 t->static_objects = END_OF_STATIC_LIST;
1436 t->scavenged_static_objects = END_OF_STATIC_LIST;
1439 t->failed_to_evac = rtsFalse;
1440 t->eager_promotion = rtsTrue;
1441 t->thunk_selector_depth = 0;
1446 t->scav_find_work = 0;
1449 /* -----------------------------------------------------------------------------
1450 Function we pass to GetRoots to evacuate roots.
1451 -------------------------------------------------------------------------- */
1454 mark_root(StgClosure **root)
1459 /* -----------------------------------------------------------------------------
1460 Initialising the static object & mutable lists
1461 -------------------------------------------------------------------------- */
1464 zero_static_object_list(StgClosure* first_static)
1468 const StgInfoTable *info;
1470 for (p = first_static; p != END_OF_STATIC_LIST; p = link) {
1472 link = *STATIC_LINK(info, p);
1473 *STATIC_LINK(info,p) = NULL;
1477 /* -----------------------------------------------------------------------------
1479 -------------------------------------------------------------------------- */
1486 for (c = (StgIndStatic *)revertible_caf_list; c != NULL;
1487 c = (StgIndStatic *)c->static_link)
1489 SET_INFO(c, c->saved_info);
1490 c->saved_info = NULL;
1491 // could, but not necessary: c->static_link = NULL;
1493 revertible_caf_list = NULL;
1497 markCAFs( evac_fn evac )
1501 for (c = (StgIndStatic *)caf_list; c != NULL;
1502 c = (StgIndStatic *)c->static_link)
1504 evac(&c->indirectee);
1506 for (c = (StgIndStatic *)revertible_caf_list; c != NULL;
1507 c = (StgIndStatic *)c->static_link)
1509 evac(&c->indirectee);
1513 /* ----------------------------------------------------------------------------
1514 Update the pointers from the task list
1516 These are treated as weak pointers because we want to allow a main
1517 thread to get a BlockedOnDeadMVar exception in the same way as any
1518 other thread. Note that the threads should all have been retained
1519 by GC by virtue of being on the all_threads list, we're just
1520 updating pointers here.
1521 ------------------------------------------------------------------------- */
1524 update_task_list (void)
1528 for (task = all_tasks; task != NULL; task = task->all_link) {
1529 if (!task->stopped && task->tso) {
1530 ASSERT(task->tso->bound == task);
1531 tso = (StgTSO *) isAlive((StgClosure *)task->tso);
1533 barf("task %p: main thread %d has been GC'd",
1546 /* ----------------------------------------------------------------------------
1547 Reset the sizes of the older generations when we do a major
1550 CURRENT STRATEGY: make all generations except zero the same size.
1551 We have to stay within the maximum heap size, and leave a certain
1552 percentage of the maximum heap size available to allocate into.
1553 ------------------------------------------------------------------------- */
1556 resize_generations (void)
1560 if (major_gc && RtsFlags.GcFlags.generations > 1) {
1561 nat live, size, min_alloc;
1562 nat max = RtsFlags.GcFlags.maxHeapSize;
1563 nat gens = RtsFlags.GcFlags.generations;
1565 // live in the oldest generations
1566 live = (oldest_gen->steps[0].n_words + BLOCK_SIZE_W - 1) / BLOCK_SIZE_W+
1567 oldest_gen->steps[0].n_large_blocks;
1569 // default max size for all generations except zero
1570 size = stg_max(live * RtsFlags.GcFlags.oldGenFactor,
1571 RtsFlags.GcFlags.minOldGenSize);
1573 // minimum size for generation zero
1574 min_alloc = stg_max((RtsFlags.GcFlags.pcFreeHeap * max) / 200,
1575 RtsFlags.GcFlags.minAllocAreaSize);
1577 // Auto-enable compaction when the residency reaches a
1578 // certain percentage of the maximum heap size (default: 30%).
1579 if (RtsFlags.GcFlags.generations > 1 &&
1580 (RtsFlags.GcFlags.compact ||
1582 oldest_gen->steps[0].n_blocks >
1583 (RtsFlags.GcFlags.compactThreshold * max) / 100))) {
1584 oldest_gen->steps[0].is_compacted = 1;
1585 // debugBelch("compaction: on\n", live);
1587 oldest_gen->steps[0].is_compacted = 0;
1588 // debugBelch("compaction: off\n", live);
1591 // if we're going to go over the maximum heap size, reduce the
1592 // size of the generations accordingly. The calculation is
1593 // different if compaction is turned on, because we don't need
1594 // to double the space required to collect the old generation.
1597 // this test is necessary to ensure that the calculations
1598 // below don't have any negative results - we're working
1599 // with unsigned values here.
1600 if (max < min_alloc) {
1604 if (oldest_gen->steps[0].is_compacted) {
1605 if ( (size + (size - 1) * (gens - 2) * 2) + min_alloc > max ) {
1606 size = (max - min_alloc) / ((gens - 1) * 2 - 1);
1609 if ( (size * (gens - 1) * 2) + min_alloc > max ) {
1610 size = (max - min_alloc) / ((gens - 1) * 2);
1620 debugBelch("live: %d, min_alloc: %d, size : %d, max = %d\n", live,
1621 min_alloc, size, max);
1624 for (g = 0; g < gens; g++) {
1625 generations[g].max_blocks = size;
1630 /* -----------------------------------------------------------------------------
1631 Calculate the new size of the nursery, and resize it.
1632 -------------------------------------------------------------------------- */
1635 resize_nursery (void)
1637 if (RtsFlags.GcFlags.generations == 1)
1638 { // Two-space collector:
1641 /* set up a new nursery. Allocate a nursery size based on a
1642 * function of the amount of live data (by default a factor of 2)
1643 * Use the blocks from the old nursery if possible, freeing up any
1646 * If we get near the maximum heap size, then adjust our nursery
1647 * size accordingly. If the nursery is the same size as the live
1648 * data (L), then we need 3L bytes. We can reduce the size of the
1649 * nursery to bring the required memory down near 2L bytes.
1651 * A normal 2-space collector would need 4L bytes to give the same
1652 * performance we get from 3L bytes, reducing to the same
1653 * performance at 2L bytes.
1655 blocks = g0s0->n_old_blocks;
1657 if ( RtsFlags.GcFlags.maxHeapSize != 0 &&
1658 blocks * RtsFlags.GcFlags.oldGenFactor * 2 >
1659 RtsFlags.GcFlags.maxHeapSize )
1661 long adjusted_blocks; // signed on purpose
1664 adjusted_blocks = (RtsFlags.GcFlags.maxHeapSize - 2 * blocks);
1666 debugTrace(DEBUG_gc, "near maximum heap size of 0x%x blocks, blocks = %d, adjusted to %ld",
1667 RtsFlags.GcFlags.maxHeapSize, blocks, adjusted_blocks);
1669 pc_free = adjusted_blocks * 100 / RtsFlags.GcFlags.maxHeapSize;
1670 if (pc_free < RtsFlags.GcFlags.pcFreeHeap) /* might even * be < 0 */
1674 blocks = adjusted_blocks;
1678 blocks *= RtsFlags.GcFlags.oldGenFactor;
1679 if (blocks < RtsFlags.GcFlags.minAllocAreaSize)
1681 blocks = RtsFlags.GcFlags.minAllocAreaSize;
1684 resizeNurseries(blocks);
1686 else // Generational collector
1689 * If the user has given us a suggested heap size, adjust our
1690 * allocation area to make best use of the memory available.
1692 if (RtsFlags.GcFlags.heapSizeSuggestion)
1695 nat needed = calcNeeded(); // approx blocks needed at next GC
1697 /* Guess how much will be live in generation 0 step 0 next time.
1698 * A good approximation is obtained by finding the
1699 * percentage of g0s0 that was live at the last minor GC.
1701 * We have an accurate figure for the amount of copied data in
1702 * 'copied', but we must convert this to a number of blocks, with
1703 * a small adjustment for estimated slop at the end of a block
1708 g0s0_pcnt_kept = ((copied / (BLOCK_SIZE_W - 10)) * 100)
1709 / countNurseryBlocks();
1712 /* Estimate a size for the allocation area based on the
1713 * information available. We might end up going slightly under
1714 * or over the suggested heap size, but we should be pretty
1717 * Formula: suggested - needed
1718 * ----------------------------
1719 * 1 + g0s0_pcnt_kept/100
1721 * where 'needed' is the amount of memory needed at the next
1722 * collection for collecting all steps except g0s0.
1725 (((long)RtsFlags.GcFlags.heapSizeSuggestion - (long)needed) * 100) /
1726 (100 + (long)g0s0_pcnt_kept);
1728 if (blocks < (long)RtsFlags.GcFlags.minAllocAreaSize) {
1729 blocks = RtsFlags.GcFlags.minAllocAreaSize;
1732 resizeNurseries((nat)blocks);
1736 // we might have added extra large blocks to the nursery, so
1737 // resize back to minAllocAreaSize again.
1738 resizeNurseriesFixed(RtsFlags.GcFlags.minAllocAreaSize);
1743 /* -----------------------------------------------------------------------------
1744 Sanity code for CAF garbage collection.
1746 With DEBUG turned on, we manage a CAF list in addition to the SRT
1747 mechanism. After GC, we run down the CAF list and blackhole any
1748 CAFs which have been garbage collected. This means we get an error
1749 whenever the program tries to enter a garbage collected CAF.
1751 Any garbage collected CAFs are taken off the CAF list at the same
1753 -------------------------------------------------------------------------- */
1755 #if 0 && defined(DEBUG)
1762 const StgInfoTable *info;
1773 ASSERT(info->type == IND_STATIC);
1775 if (STATIC_LINK(info,p) == NULL) {
1776 debugTrace(DEBUG_gccafs, "CAF gc'd at 0x%04lx", (long)p);
1778 SET_INFO(p,&stg_BLACKHOLE_info);
1779 p = STATIC_LINK2(info,p);
1783 pp = &STATIC_LINK2(info,p);
1790 debugTrace(DEBUG_gccafs, "%d CAFs live", i);