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;
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 debugTrace(DEBUG_gc, "starting GC");
200 #if defined(RTS_USER_SIGNALS)
201 if (RtsFlags.MiscFlags.install_signal_handlers) {
207 // tell the stats department that we've started a GC
210 // tell the STM to discard any cached closures it's hoping to re-use
219 // attribute any costs to CCS_GC
225 /* Approximate how much we allocated.
226 * Todo: only when generating stats?
228 allocated = calcAllocated();
230 /* Figure out which generation to collect
232 n = initialise_N(force_major_gc);
234 /* Allocate + initialise the gc_thread structures.
238 /* Start threads, so they can be spinning up while we finish initialisation.
242 /* How many threads will be participating in this GC?
243 * We don't try to parallelise minor GC.
245 #if defined(THREADED_RTS)
246 if (n < (4*1024*1024 / BLOCK_SIZE)) {
249 n_gc_threads = RtsFlags.ParFlags.gcThreads;
251 trace(TRACE_gc|DEBUG_gc, "GC (gen %d): %dKB to collect, using %d thread(s)",
252 N, n * (BLOCK_SIZE / 1024), n_gc_threads);
257 #ifdef RTS_GTK_FRONTPANEL
258 if (RtsFlags.GcFlags.frontpanel) {
259 updateFrontPanelBeforeGC(N);
264 // check for memory leaks if DEBUG is on
265 memInventory(traceClass(DEBUG_gc));
268 // check stack sanity *before* GC (ToDo: check all threads)
269 IF_DEBUG(sanity, checkFreeListSanity());
271 // Initialise all our gc_thread structures
272 for (t = 0; t < n_gc_threads; t++) {
273 init_gc_thread(gc_threads[t]);
276 // Initialise all the generations/steps that we're collecting.
277 for (g = 0; g <= N; g++) {
278 init_collected_gen(g,n_gc_threads);
281 // Initialise all the generations/steps that we're *not* collecting.
282 for (g = N+1; g < RtsFlags.GcFlags.generations; g++) {
283 init_uncollected_gen(g,n_gc_threads);
286 /* Allocate a mark stack if we're doing a major collection.
289 mark_stack_bdescr = allocGroup(MARK_STACK_BLOCKS);
290 mark_stack = (StgPtr *)mark_stack_bdescr->start;
291 mark_sp = mark_stack;
292 mark_splim = mark_stack + (MARK_STACK_BLOCKS * BLOCK_SIZE_W);
294 mark_stack_bdescr = NULL;
297 // this is the main thread
300 /* -----------------------------------------------------------------------
301 * follow all the roots that we know about:
302 * - mutable lists from each generation > N
303 * we want to *scavenge* these roots, not evacuate them: they're not
304 * going to move in this GC.
305 * Also do them in reverse generation order, for the usual reason:
306 * namely to reduce the likelihood of spurious old->new pointers.
308 for (g = RtsFlags.GcFlags.generations-1; g > N; g--) {
309 generations[g].saved_mut_list = generations[g].mut_list;
310 generations[g].mut_list = allocBlock();
311 // mut_list always has at least one block.
314 // the main thread is running: this prevents any other threads from
315 // exiting prematurely, so we can start them now.
316 // NB. do this after the mutable lists have been saved above, otherwise
317 // the other GC threads will be writing into the old mutable lists.
319 wakeup_gc_threads(n_gc_threads);
321 for (g = RtsFlags.GcFlags.generations-1; g > N; g--) {
322 scavenge_mutable_list(&generations[g]);
325 // follow roots from the CAF list (used by GHCi)
329 // follow all the roots that the application knows about.
333 #if defined(RTS_USER_SIGNALS)
334 // mark the signal handlers (signals should be already blocked)
335 markSignalHandlers(mark_root);
338 // Mark the weak pointer list, and prepare to detect dead weak pointers.
342 // Mark the stable pointer table.
343 markStablePtrTable(mark_root);
345 /* -------------------------------------------------------------------------
346 * Repeatedly scavenge all the areas we know about until there's no
347 * more scavenging to be done.
352 // The other threads are now stopped. We might recurse back to
353 // here, but from now on this is the only thread.
355 // if any blackholes are alive, make the threads that wait on
357 if (traverseBlackholeQueue()) {
362 // must be last... invariant is that everything is fully
363 // scavenged at this point.
364 if (traverseWeakPtrList()) { // returns rtsTrue if evaced something
369 // If we get to here, there's really nothing left to do.
373 shutdown_gc_threads(n_gc_threads);
375 // Update pointers from the Task list
378 // Now see which stable names are still alive.
382 // We call processHeapClosureForDead() on every closure destroyed during
383 // the current garbage collection, so we invoke LdvCensusForDead().
384 if (RtsFlags.ProfFlags.doHeapProfile == HEAP_BY_LDV
385 || RtsFlags.ProfFlags.bioSelector != NULL)
389 // NO MORE EVACUATION AFTER THIS POINT!
390 // Finally: compaction of the oldest generation.
391 if (major_gc && oldest_gen->steps[0].is_compacted) {
392 // save number of blocks for stats
393 oldgen_saved_blocks = oldest_gen->steps[0].n_old_blocks;
397 IF_DEBUG(sanity, checkGlobalTSOList(rtsFalse));
399 // Two-space collector: free the old to-space.
400 // g0s0->old_blocks is the old nursery
401 // g0s0->blocks is to-space from the previous GC
402 if (RtsFlags.GcFlags.generations == 1) {
403 if (g0s0->blocks != NULL) {
404 freeChain(g0s0->blocks);
409 // For each workspace, in each thread:
410 // * clear the BF_EVACUATED flag from each copied block
411 // * move the copied blocks to the step
417 for (t = 0; t < n_gc_threads; t++) {
421 for (s = 1; s < total_steps; s++) {
424 // ASSERT( ws->scan_bd == ws->todo_bd );
425 ASSERT( ws->scan_bd ? ws->scan == ws->scan_bd->free : 1 );
427 // Push the final block
428 if (ws->scan_bd) { push_scan_block(ws->scan_bd, ws); }
430 ASSERT(countBlocks(ws->scavd_list) == ws->n_scavd_blocks);
432 prev = ws->scavd_list;
433 for (bd = ws->scavd_list; bd != NULL; bd = bd->link) {
434 bd->flags &= ~BF_EVACUATED; // now from-space
437 prev->link = ws->stp->blocks;
438 ws->stp->blocks = ws->scavd_list;
439 ws->stp->n_blocks += ws->n_scavd_blocks;
440 ASSERT(countBlocks(ws->stp->blocks) == ws->stp->n_blocks);
445 // Two-space collector: swap the semi-spaces around.
446 // Currently: g0s0->old_blocks is the old nursery
447 // g0s0->blocks is to-space from this GC
448 // We want these the other way around.
449 if (RtsFlags.GcFlags.generations == 1) {
450 bdescr *nursery_blocks = g0s0->old_blocks;
451 nat n_nursery_blocks = g0s0->n_old_blocks;
452 g0s0->old_blocks = g0s0->blocks;
453 g0s0->n_old_blocks = g0s0->n_blocks;
454 g0s0->blocks = nursery_blocks;
455 g0s0->n_blocks = n_nursery_blocks;
458 /* run through all the generations/steps and tidy up
463 for (i=0; i < n_gc_threads; i++) {
464 if (n_gc_threads > 1) {
465 trace(TRACE_gc,"thread %d:", i);
466 trace(TRACE_gc," copied %ld", gc_threads[i]->copied * sizeof(W_));
467 trace(TRACE_gc," any_work %ld", gc_threads[i]->any_work);
468 trace(TRACE_gc," no_work %ld", gc_threads[i]->no_work);
469 trace(TRACE_gc," scav_global_work %ld", gc_threads[i]->scav_global_work);
470 trace(TRACE_gc," scav_local_work %ld", gc_threads[i]->scav_local_work);
472 copied += gc_threads[i]->copied;
476 for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
479 generations[g].collections++; // for stats
482 // Count the mutable list as bytes "copied" for the purposes of
483 // stats. Every mutable list is copied during every GC.
485 nat mut_list_size = 0;
486 for (bd = generations[g].mut_list; bd != NULL; bd = bd->link) {
487 mut_list_size += bd->free - bd->start;
489 copied += mut_list_size;
492 "mut_list_size: %lu (%d vars, %d arrays, %d MVARs, %d others)",
493 (unsigned long)(mut_list_size * sizeof(W_)),
494 mutlist_MUTVARS, mutlist_MUTARRS, mutlist_MVARS, mutlist_OTHERS);
497 for (s = 0; s < generations[g].n_steps; s++) {
499 stp = &generations[g].steps[s];
501 // for generations we collected...
504 /* free old memory and shift to-space into from-space for all
505 * the collected steps (except the allocation area). These
506 * freed blocks will probaby be quickly recycled.
508 if (!(g == 0 && s == 0 && RtsFlags.GcFlags.generations > 1)) {
509 if (stp->is_compacted)
511 // for a compacted step, just shift the new to-space
512 // onto the front of the now-compacted existing blocks.
513 for (bd = stp->blocks; bd != NULL; bd = bd->link) {
514 bd->flags &= ~BF_EVACUATED; // now from-space
516 // tack the new blocks on the end of the existing blocks
517 if (stp->old_blocks != NULL) {
518 for (bd = stp->old_blocks; bd != NULL; bd = next) {
519 // NB. this step might not be compacted next
520 // time, so reset the BF_COMPACTED flags.
521 // They are set before GC if we're going to
522 // compact. (search for BF_COMPACTED above).
523 bd->flags &= ~BF_COMPACTED;
526 bd->link = stp->blocks;
529 stp->blocks = stp->old_blocks;
531 // add the new blocks to the block tally
532 stp->n_blocks += stp->n_old_blocks;
533 ASSERT(countBlocks(stp->blocks) == stp->n_blocks);
537 freeChain(stp->old_blocks);
539 stp->old_blocks = NULL;
540 stp->n_old_blocks = 0;
543 /* LARGE OBJECTS. The current live large objects are chained on
544 * scavenged_large, having been moved during garbage
545 * collection from large_objects. Any objects left on
546 * large_objects list are therefore dead, so we free them here.
548 for (bd = stp->large_objects; bd != NULL; bd = next) {
554 // update the count of blocks used by large objects
555 for (bd = stp->scavenged_large_objects; bd != NULL; bd = bd->link) {
556 bd->flags &= ~BF_EVACUATED;
558 stp->large_objects = stp->scavenged_large_objects;
559 stp->n_large_blocks = stp->n_scavenged_large_blocks;
562 else // for older generations...
564 /* For older generations, we need to append the
565 * scavenged_large_object list (i.e. large objects that have been
566 * promoted during this GC) to the large_object list for that step.
568 for (bd = stp->scavenged_large_objects; bd; bd = next) {
570 bd->flags &= ~BF_EVACUATED;
571 dbl_link_onto(bd, &stp->large_objects);
574 // add the new blocks we promoted during this GC
575 stp->n_large_blocks += stp->n_scavenged_large_blocks;
580 // update the max size of older generations after a major GC
581 resize_generations();
583 // Guess the amount of live data for stats.
584 live = calcLiveBlocks() * BLOCK_SIZE_W;
585 debugTrace(DEBUG_gc, "Slop: %ldKB",
586 (live - calcLiveWords()) / (1024/sizeof(W_)));
588 // Free the small objects allocated via allocate(), since this will
589 // all have been copied into G0S1 now.
590 if (RtsFlags.GcFlags.generations > 1) {
591 if (g0s0->blocks != NULL) {
592 freeChain(g0s0->blocks);
598 alloc_blocks_lim = RtsFlags.GcFlags.minAllocAreaSize;
600 // Start a new pinned_object_block
601 pinned_object_block = NULL;
603 // Free the mark stack.
604 if (mark_stack_bdescr != NULL) {
605 freeGroup(mark_stack_bdescr);
609 for (g = 0; g <= N; g++) {
610 for (s = 0; s < generations[g].n_steps; s++) {
611 stp = &generations[g].steps[s];
612 if (stp->bitmap != NULL) {
613 freeGroup(stp->bitmap);
621 // mark the garbage collected CAFs as dead
622 #if 0 && defined(DEBUG) // doesn't work at the moment
623 if (major_gc) { gcCAFs(); }
627 // resetStaticObjectForRetainerProfiling() must be called before
629 if (n_gc_threads > 1) {
630 barf("profiling is currently broken with multi-threaded GC");
631 // ToDo: fix the gct->scavenged_static_objects below
633 resetStaticObjectForRetainerProfiling(gct->scavenged_static_objects);
636 // zero the scavenged static object list
639 for (i = 0; i < n_gc_threads; i++) {
640 zero_static_object_list(gc_threads[i]->scavenged_static_objects);
647 // start any pending finalizers
649 scheduleFinalizers(last_free_capability, old_weak_ptr_list);
652 // send exceptions to any threads which were about to die
654 resurrectThreads(resurrected_threads);
657 // Update the stable pointer hash table.
658 updateStablePtrTable(major_gc);
660 // check sanity after GC
661 IF_DEBUG(sanity, checkSanity());
663 // extra GC trace info
664 if (traceClass(TRACE_gc|DEBUG_gc)) statDescribeGens();
667 // symbol-table based profiling
668 /* heapCensus(to_blocks); */ /* ToDo */
671 // restore enclosing cost centre
677 // check for memory leaks if DEBUG is on
678 memInventory(traceClass(DEBUG_gc));
681 #ifdef RTS_GTK_FRONTPANEL
682 if (RtsFlags.GcFlags.frontpanel) {
683 updateFrontPanelAfterGC( N, live );
687 // ok, GC over: tell the stats department what happened.
688 stat_endGC(allocated, live, copied, N);
690 #if defined(RTS_USER_SIGNALS)
691 if (RtsFlags.MiscFlags.install_signal_handlers) {
692 // unblock signals again
693 unblockUserSignals();
702 /* -----------------------------------------------------------------------------
703 * Mark all nodes pointed to by sparks in the spark queues (for GC) Does an
704 * implicit slide i.e. after marking all sparks are at the beginning of the
705 * spark pool and the spark pool only contains sparkable closures
706 * -------------------------------------------------------------------------- */
710 markSparkQueue (evac_fn evac, Capability *cap)
712 StgClosure **sparkp, **to_sparkp;
713 nat n, pruned_sparks; // stats only
716 PAR_TICKY_MARK_SPARK_QUEUE_START();
721 pool = &(cap->r.rSparks);
723 ASSERT_SPARK_POOL_INVARIANTS(pool);
725 #if defined(PARALLEL_HASKELL)
732 to_sparkp = pool->hd;
733 while (sparkp != pool->tl) {
734 ASSERT(*sparkp!=NULL);
735 ASSERT(LOOKS_LIKE_CLOSURE_PTR(((StgClosure *)*sparkp)));
736 // ToDo?: statistics gathering here (also for GUM!)
737 if (closure_SHOULD_SPARK(*sparkp)) {
739 *to_sparkp++ = *sparkp;
740 if (to_sparkp == pool->lim) {
741 to_sparkp = pool->base;
748 if (sparkp == pool->lim) {
752 pool->tl = to_sparkp;
754 PAR_TICKY_MARK_SPARK_QUEUE_END(n);
756 #if defined(PARALLEL_HASKELL)
757 debugTrace(DEBUG_sched,
758 "marked %d sparks and pruned %d sparks on [%x]",
759 n, pruned_sparks, mytid);
761 debugTrace(DEBUG_sched,
762 "marked %d sparks and pruned %d sparks",
766 debugTrace(DEBUG_sched,
767 "new spark queue len=%d; (hd=%p; tl=%p)\n",
768 sparkPoolSize(pool), pool->hd, pool->tl);
772 /* ---------------------------------------------------------------------------
773 Where are the roots that we know about?
775 - all the threads on the runnable queue
776 - all the threads on the blocked queue
777 - all the threads on the sleeping queue
778 - all the thread currently executing a _ccall_GC
779 - all the "main threads"
781 ------------------------------------------------------------------------ */
784 GetRoots( evac_fn evac )
790 // Each GC thread is responsible for following roots from the
791 // Capability of the same number. There will usually be the same
792 // or fewer Capabilities as GC threads, but just in case there
793 // are more, we mark every Capability whose number is the GC
794 // thread's index plus a multiple of the number of GC threads.
795 for (i = gct->thread_index; i < n_capabilities; i += n_gc_threads) {
796 cap = &capabilities[i];
797 evac((StgClosure **)(void *)&cap->run_queue_hd);
798 evac((StgClosure **)(void *)&cap->run_queue_tl);
799 #if defined(THREADED_RTS)
800 evac((StgClosure **)(void *)&cap->wakeup_queue_hd);
801 evac((StgClosure **)(void *)&cap->wakeup_queue_tl);
803 for (task = cap->suspended_ccalling_tasks; task != NULL;
805 debugTrace(DEBUG_sched,
806 "evac'ing suspended TSO %lu", (unsigned long)task->suspended_tso->id);
807 evac((StgClosure **)(void *)&task->suspended_tso);
810 #if defined(THREADED_RTS)
811 markSparkQueue(evac,cap);
815 #if !defined(THREADED_RTS)
816 evac((StgClosure **)(void *)&blocked_queue_hd);
817 evac((StgClosure **)(void *)&blocked_queue_tl);
818 evac((StgClosure **)(void *)&sleeping_queue);
822 /* -----------------------------------------------------------------------------
823 isAlive determines whether the given closure is still alive (after
824 a garbage collection) or not. It returns the new address of the
825 closure if it is alive, or NULL otherwise.
827 NOTE: Use it before compaction only!
828 It untags and (if needed) retags pointers to closures.
829 -------------------------------------------------------------------------- */
833 isAlive(StgClosure *p)
835 const StgInfoTable *info;
841 /* The tag and the pointer are split, to be merged later when needed. */
842 tag = GET_CLOSURE_TAG(p);
843 q = UNTAG_CLOSURE(p);
845 ASSERT(LOOKS_LIKE_CLOSURE_PTR(q));
848 // ignore static closures
850 // ToDo: for static closures, check the static link field.
851 // Problem here is that we sometimes don't set the link field, eg.
852 // for static closures with an empty SRT or CONSTR_STATIC_NOCAFs.
854 if (!HEAP_ALLOCED(q)) {
858 // ignore closures in generations that we're not collecting.
860 if (bd->gen_no > N) {
864 // if it's a pointer into to-space, then we're done
865 if (bd->flags & BF_EVACUATED) {
869 // large objects use the evacuated flag
870 if (bd->flags & BF_LARGE) {
874 // check the mark bit for compacted steps
875 if ((bd->flags & BF_COMPACTED) && is_marked((P_)q,bd)) {
879 switch (info->type) {
884 case IND_OLDGEN: // rely on compatible layout with StgInd
885 case IND_OLDGEN_PERM:
886 // follow indirections
887 p = ((StgInd *)q)->indirectee;
892 return ((StgEvacuated *)q)->evacuee;
895 if (((StgTSO *)q)->what_next == ThreadRelocated) {
896 p = (StgClosure *)((StgTSO *)q)->link;
908 /* -----------------------------------------------------------------------------
909 Figure out which generation to collect, initialise N and major_gc.
911 Also returns the total number of blocks in generations that will be
913 -------------------------------------------------------------------------- */
916 initialise_N (rtsBool force_major_gc)
919 nat s, blocks, blocks_total;
924 if (force_major_gc) {
925 N = RtsFlags.GcFlags.generations - 1;
930 for (g = RtsFlags.GcFlags.generations - 1; g >= 0; g--) {
932 for (s = 0; s < generations[g].n_steps; s++) {
933 blocks += generations[g].steps[s].n_blocks;
934 blocks += generations[g].steps[s].n_large_blocks;
936 if (blocks >= generations[g].max_blocks) {
940 blocks_total += blocks;
944 blocks_total += countNurseryBlocks();
946 major_gc = (N == RtsFlags.GcFlags.generations-1);
950 /* -----------------------------------------------------------------------------
951 Initialise the gc_thread structures.
952 -------------------------------------------------------------------------- */
955 alloc_gc_thread (int n)
961 t = stgMallocBytes(sizeof(gc_thread) + total_steps * sizeof(step_workspace),
966 initCondition(&t->wake_cond);
967 initMutex(&t->wake_mutex);
968 t->wakeup = rtsFalse;
973 t->free_blocks = NULL;
982 for (s = 0; s < total_steps; s++)
985 ws->stp = &all_steps[s];
986 ASSERT(s == ws->stp->abs_no);
993 ws->buffer_todo_bd = NULL;
995 ws->scavd_list = NULL;
996 ws->n_scavd_blocks = 0;
1004 alloc_gc_threads (void)
1006 if (gc_threads == NULL) {
1007 #if defined(THREADED_RTS)
1009 gc_threads = stgMallocBytes (RtsFlags.ParFlags.gcThreads *
1011 "alloc_gc_threads");
1013 for (i = 0; i < RtsFlags.ParFlags.gcThreads; i++) {
1014 gc_threads[i] = alloc_gc_thread(i);
1017 gc_threads = stgMallocBytes (sizeof(gc_thread*),
1018 "alloc_gc_threads");
1020 gc_threads[0] = alloc_gc_thread(0);
1025 /* ----------------------------------------------------------------------------
1027 ------------------------------------------------------------------------- */
1029 static nat gc_running_threads;
1031 #if defined(THREADED_RTS)
1032 static Mutex gc_running_mutex;
1039 ACQUIRE_LOCK(&gc_running_mutex);
1040 n_running = ++gc_running_threads;
1041 RELEASE_LOCK(&gc_running_mutex);
1042 ASSERT(n_running <= n_gc_threads);
1050 ACQUIRE_LOCK(&gc_running_mutex);
1051 ASSERT(n_gc_threads != 0);
1052 n_running = --gc_running_threads;
1053 RELEASE_LOCK(&gc_running_mutex);
1058 // gc_thread_work(): Scavenge until there's no work left to do and all
1059 // the running threads are idle.
1062 gc_thread_work (void)
1066 debugTrace(DEBUG_gc, "GC thread %d working", gct->thread_index);
1068 // gc_running_threads has already been incremented for us; either
1069 // this is the main thread and we incremented it inside
1070 // GarbageCollect(), or this is a worker thread and the main
1071 // thread bumped gc_running_threads before waking us up.
1073 // Every thread evacuates some roots.
1075 GetRoots(mark_root);
1079 // scavenge_loop() only exits when there's no work to do
1082 debugTrace(DEBUG_gc, "GC thread %d idle (%d still running)",
1083 gct->thread_index, r);
1085 while (gc_running_threads != 0) {
1091 // any_work() does not remove the work from the queue, it
1092 // just checks for the presence of work. If we find any,
1093 // then we increment gc_running_threads and go back to
1094 // scavenge_loop() to perform any pending work.
1097 // All threads are now stopped
1098 debugTrace(DEBUG_gc, "GC thread %d finished.", gct->thread_index);
1102 #if defined(THREADED_RTS)
1104 gc_thread_mainloop (void)
1106 while (!gct->exit) {
1108 // Wait until we're told to wake up
1109 ACQUIRE_LOCK(&gct->wake_mutex);
1110 gct->wakeup = rtsFalse;
1111 while (!gct->wakeup) {
1112 debugTrace(DEBUG_gc, "GC thread %d standing by...",
1114 waitCondition(&gct->wake_cond, &gct->wake_mutex);
1116 RELEASE_LOCK(&gct->wake_mutex);
1117 if (gct->exit) break;
1120 // start performance counters in this thread...
1121 if (gct->papi_events == -1) {
1122 papi_init_eventset(&gct->papi_events);
1124 papi_thread_start_gc1_count(gct->papi_events);
1130 // count events in this thread towards the GC totals
1131 papi_thread_stop_gc1_count(gct->papi_events);
1137 #if defined(THREADED_RTS)
1139 gc_thread_entry (gc_thread *my_gct)
1142 debugTrace(DEBUG_gc, "GC thread %d starting...", gct->thread_index);
1143 gct->id = osThreadId();
1144 gc_thread_mainloop();
1149 start_gc_threads (void)
1151 #if defined(THREADED_RTS)
1154 static rtsBool done = rtsFalse;
1156 gc_running_threads = 0;
1157 initMutex(&gc_running_mutex);
1160 // Start from 1: the main thread is 0
1161 for (i = 1; i < RtsFlags.ParFlags.gcThreads; i++) {
1162 createOSThread(&id, (OSThreadProc*)&gc_thread_entry,
1171 wakeup_gc_threads (nat n_threads USED_IF_THREADS)
1173 #if defined(THREADED_RTS)
1175 for (i=1; i < n_threads; i++) {
1177 ACQUIRE_LOCK(&gc_threads[i]->wake_mutex);
1178 gc_threads[i]->wakeup = rtsTrue;
1179 signalCondition(&gc_threads[i]->wake_cond);
1180 RELEASE_LOCK(&gc_threads[i]->wake_mutex);
1185 // After GC is complete, we must wait for all GC threads to enter the
1186 // standby state, otherwise they may still be executing inside
1187 // any_work(), and may even remain awake until the next GC starts.
1189 shutdown_gc_threads (nat n_threads USED_IF_THREADS)
1191 #if defined(THREADED_RTS)
1194 for (i=1; i < n_threads; i++) {
1196 ACQUIRE_LOCK(&gc_threads[i]->wake_mutex);
1197 wakeup = gc_threads[i]->wakeup;
1198 // wakeup is false while the thread is waiting
1199 RELEASE_LOCK(&gc_threads[i]->wake_mutex);
1205 /* ----------------------------------------------------------------------------
1206 Initialise a generation that is to be collected
1207 ------------------------------------------------------------------------- */
1210 init_collected_gen (nat g, nat n_threads)
1217 // Throw away the current mutable list. Invariant: the mutable
1218 // list always has at least one block; this means we can avoid a
1219 // check for NULL in recordMutable().
1221 freeChain(generations[g].mut_list);
1222 generations[g].mut_list = allocBlock();
1223 for (i = 0; i < n_capabilities; i++) {
1224 freeChain(capabilities[i].mut_lists[g]);
1225 capabilities[i].mut_lists[g] = allocBlock();
1229 for (s = 0; s < generations[g].n_steps; s++) {
1231 // generation 0, step 0 doesn't need to-space
1232 if (g == 0 && s == 0 && RtsFlags.GcFlags.generations > 1) {
1236 stp = &generations[g].steps[s];
1237 ASSERT(stp->gen_no == g);
1239 // deprecate the existing blocks
1240 stp->old_blocks = stp->blocks;
1241 stp->n_old_blocks = stp->n_blocks;
1245 // we don't have any to-be-scavenged blocks yet
1247 stp->todos_last = NULL;
1250 // initialise the large object queues.
1251 stp->scavenged_large_objects = NULL;
1252 stp->n_scavenged_large_blocks = 0;
1254 // mark the large objects as not evacuated yet
1255 for (bd = stp->large_objects; bd; bd = bd->link) {
1256 bd->flags &= ~BF_EVACUATED;
1259 // for a compacted step, we need to allocate the bitmap
1260 if (stp->is_compacted) {
1261 nat bitmap_size; // in bytes
1262 bdescr *bitmap_bdescr;
1265 bitmap_size = stp->n_old_blocks * BLOCK_SIZE / (sizeof(W_)*BITS_PER_BYTE);
1267 if (bitmap_size > 0) {
1268 bitmap_bdescr = allocGroup((lnat)BLOCK_ROUND_UP(bitmap_size)
1270 stp->bitmap = bitmap_bdescr;
1271 bitmap = bitmap_bdescr->start;
1273 debugTrace(DEBUG_gc, "bitmap_size: %d, bitmap: %p",
1274 bitmap_size, bitmap);
1276 // don't forget to fill it with zeros!
1277 memset(bitmap, 0, bitmap_size);
1279 // For each block in this step, point to its bitmap from the
1280 // block descriptor.
1281 for (bd=stp->old_blocks; bd != NULL; bd = bd->link) {
1282 bd->u.bitmap = bitmap;
1283 bitmap += BLOCK_SIZE_W / (sizeof(W_)*BITS_PER_BYTE);
1285 // Also at this point we set the BF_COMPACTED flag
1286 // for this block. The invariant is that
1287 // BF_COMPACTED is always unset, except during GC
1288 // when it is set on those blocks which will be
1290 bd->flags |= BF_COMPACTED;
1296 // For each GC thread, for each step, allocate a "todo" block to
1297 // store evacuated objects to be scavenged, and a block to store
1298 // evacuated objects that do not need to be scavenged.
1299 for (t = 0; t < n_threads; t++) {
1300 for (s = 0; s < generations[g].n_steps; s++) {
1302 // we don't copy objects into g0s0, unless -G0
1303 if (g==0 && s==0 && RtsFlags.GcFlags.generations > 1) continue;
1305 ws = &gc_threads[t]->steps[g * RtsFlags.GcFlags.steps + s];
1310 ws->todo_large_objects = NULL;
1312 // allocate the first to-space block; extra blocks will be
1313 // chained on as necessary.
1315 ws->buffer_todo_bd = NULL;
1316 gc_alloc_todo_block(ws);
1318 ws->scavd_list = NULL;
1319 ws->n_scavd_blocks = 0;
1325 /* ----------------------------------------------------------------------------
1326 Initialise a generation that is *not* to be collected
1327 ------------------------------------------------------------------------- */
1330 init_uncollected_gen (nat g, nat threads)
1337 for (s = 0; s < generations[g].n_steps; s++) {
1338 stp = &generations[g].steps[s];
1339 stp->scavenged_large_objects = NULL;
1340 stp->n_scavenged_large_blocks = 0;
1343 for (t = 0; t < threads; t++) {
1344 for (s = 0; s < generations[g].n_steps; s++) {
1346 ws = &gc_threads[t]->steps[g * RtsFlags.GcFlags.steps + s];
1349 ws->buffer_todo_bd = NULL;
1350 ws->todo_large_objects = NULL;
1352 ws->scavd_list = NULL;
1353 ws->n_scavd_blocks = 0;
1355 // If the block at the head of the list in this generation
1356 // is less than 3/4 full, then use it as a todo block.
1357 if (stp->blocks && isPartiallyFull(stp->blocks))
1359 ws->todo_bd = stp->blocks;
1360 ws->todo_free = ws->todo_bd->free;
1361 ws->todo_lim = ws->todo_bd->start + BLOCK_SIZE_W;
1362 stp->blocks = stp->blocks->link;
1364 ws->todo_bd->link = NULL;
1366 // this block is also the scan block; we must scan
1367 // from the current end point.
1368 ws->scan_bd = ws->todo_bd;
1369 ws->scan = ws->scan_bd->free;
1371 // subtract the contents of this block from the stats,
1372 // because we'll count the whole block later.
1373 copied -= ws->scan_bd->free - ws->scan_bd->start;
1380 gc_alloc_todo_block(ws);
1385 // Move the private mutable lists from each capability onto the
1386 // main mutable list for the generation.
1387 for (i = 0; i < n_capabilities; i++) {
1388 for (bd = capabilities[i].mut_lists[g];
1389 bd->link != NULL; bd = bd->link) {
1392 bd->link = generations[g].mut_list;
1393 generations[g].mut_list = capabilities[i].mut_lists[g];
1394 capabilities[i].mut_lists[g] = allocBlock();
1398 /* -----------------------------------------------------------------------------
1399 Initialise a gc_thread before GC
1400 -------------------------------------------------------------------------- */
1403 init_gc_thread (gc_thread *t)
1405 t->static_objects = END_OF_STATIC_LIST;
1406 t->scavenged_static_objects = END_OF_STATIC_LIST;
1408 t->failed_to_evac = rtsFalse;
1409 t->eager_promotion = rtsTrue;
1410 t->thunk_selector_depth = 0;
1414 t->scav_global_work = 0;
1415 t->scav_local_work = 0;
1419 /* -----------------------------------------------------------------------------
1420 Function we pass to GetRoots to evacuate roots.
1421 -------------------------------------------------------------------------- */
1424 mark_root(StgClosure **root)
1429 /* -----------------------------------------------------------------------------
1430 Initialising the static object & mutable lists
1431 -------------------------------------------------------------------------- */
1434 zero_static_object_list(StgClosure* first_static)
1438 const StgInfoTable *info;
1440 for (p = first_static; p != END_OF_STATIC_LIST; p = link) {
1442 link = *STATIC_LINK(info, p);
1443 *STATIC_LINK(info,p) = NULL;
1447 /* -----------------------------------------------------------------------------
1449 -------------------------------------------------------------------------- */
1456 for (c = (StgIndStatic *)revertible_caf_list; c != NULL;
1457 c = (StgIndStatic *)c->static_link)
1459 SET_INFO(c, c->saved_info);
1460 c->saved_info = NULL;
1461 // could, but not necessary: c->static_link = NULL;
1463 revertible_caf_list = NULL;
1467 markCAFs( evac_fn evac )
1471 for (c = (StgIndStatic *)caf_list; c != NULL;
1472 c = (StgIndStatic *)c->static_link)
1474 evac(&c->indirectee);
1476 for (c = (StgIndStatic *)revertible_caf_list; c != NULL;
1477 c = (StgIndStatic *)c->static_link)
1479 evac(&c->indirectee);
1483 /* ----------------------------------------------------------------------------
1484 Update the pointers from the task list
1486 These are treated as weak pointers because we want to allow a main
1487 thread to get a BlockedOnDeadMVar exception in the same way as any
1488 other thread. Note that the threads should all have been retained
1489 by GC by virtue of being on the all_threads list, we're just
1490 updating pointers here.
1491 ------------------------------------------------------------------------- */
1494 update_task_list (void)
1498 for (task = all_tasks; task != NULL; task = task->all_link) {
1499 if (!task->stopped && task->tso) {
1500 ASSERT(task->tso->bound == task);
1501 tso = (StgTSO *) isAlive((StgClosure *)task->tso);
1503 barf("task %p: main thread %d has been GC'd",
1516 /* ----------------------------------------------------------------------------
1517 Reset the sizes of the older generations when we do a major
1520 CURRENT STRATEGY: make all generations except zero the same size.
1521 We have to stay within the maximum heap size, and leave a certain
1522 percentage of the maximum heap size available to allocate into.
1523 ------------------------------------------------------------------------- */
1526 resize_generations (void)
1530 if (major_gc && RtsFlags.GcFlags.generations > 1) {
1531 nat live, size, min_alloc;
1532 nat max = RtsFlags.GcFlags.maxHeapSize;
1533 nat gens = RtsFlags.GcFlags.generations;
1535 // live in the oldest generations
1536 live = oldest_gen->steps[0].n_blocks +
1537 oldest_gen->steps[0].n_large_blocks;
1539 // default max size for all generations except zero
1540 size = stg_max(live * RtsFlags.GcFlags.oldGenFactor,
1541 RtsFlags.GcFlags.minOldGenSize);
1543 // minimum size for generation zero
1544 min_alloc = stg_max((RtsFlags.GcFlags.pcFreeHeap * max) / 200,
1545 RtsFlags.GcFlags.minAllocAreaSize);
1547 // Auto-enable compaction when the residency reaches a
1548 // certain percentage of the maximum heap size (default: 30%).
1549 if (RtsFlags.GcFlags.generations > 1 &&
1550 (RtsFlags.GcFlags.compact ||
1552 oldest_gen->steps[0].n_blocks >
1553 (RtsFlags.GcFlags.compactThreshold * max) / 100))) {
1554 oldest_gen->steps[0].is_compacted = 1;
1555 // debugBelch("compaction: on\n", live);
1557 oldest_gen->steps[0].is_compacted = 0;
1558 // debugBelch("compaction: off\n", live);
1561 // if we're going to go over the maximum heap size, reduce the
1562 // size of the generations accordingly. The calculation is
1563 // different if compaction is turned on, because we don't need
1564 // to double the space required to collect the old generation.
1567 // this test is necessary to ensure that the calculations
1568 // below don't have any negative results - we're working
1569 // with unsigned values here.
1570 if (max < min_alloc) {
1574 if (oldest_gen->steps[0].is_compacted) {
1575 if ( (size + (size - 1) * (gens - 2) * 2) + min_alloc > max ) {
1576 size = (max - min_alloc) / ((gens - 1) * 2 - 1);
1579 if ( (size * (gens - 1) * 2) + min_alloc > max ) {
1580 size = (max - min_alloc) / ((gens - 1) * 2);
1590 debugBelch("live: %d, min_alloc: %d, size : %d, max = %d\n", live,
1591 min_alloc, size, max);
1594 for (g = 0; g < gens; g++) {
1595 generations[g].max_blocks = size;
1600 /* -----------------------------------------------------------------------------
1601 Calculate the new size of the nursery, and resize it.
1602 -------------------------------------------------------------------------- */
1605 resize_nursery (void)
1607 if (RtsFlags.GcFlags.generations == 1)
1608 { // Two-space collector:
1611 /* set up a new nursery. Allocate a nursery size based on a
1612 * function of the amount of live data (by default a factor of 2)
1613 * Use the blocks from the old nursery if possible, freeing up any
1616 * If we get near the maximum heap size, then adjust our nursery
1617 * size accordingly. If the nursery is the same size as the live
1618 * data (L), then we need 3L bytes. We can reduce the size of the
1619 * nursery to bring the required memory down near 2L bytes.
1621 * A normal 2-space collector would need 4L bytes to give the same
1622 * performance we get from 3L bytes, reducing to the same
1623 * performance at 2L bytes.
1625 blocks = g0s0->n_old_blocks;
1627 if ( RtsFlags.GcFlags.maxHeapSize != 0 &&
1628 blocks * RtsFlags.GcFlags.oldGenFactor * 2 >
1629 RtsFlags.GcFlags.maxHeapSize )
1631 long adjusted_blocks; // signed on purpose
1634 adjusted_blocks = (RtsFlags.GcFlags.maxHeapSize - 2 * blocks);
1636 debugTrace(DEBUG_gc, "near maximum heap size of 0x%x blocks, blocks = %d, adjusted to %ld",
1637 RtsFlags.GcFlags.maxHeapSize, blocks, adjusted_blocks);
1639 pc_free = adjusted_blocks * 100 / RtsFlags.GcFlags.maxHeapSize;
1640 if (pc_free < RtsFlags.GcFlags.pcFreeHeap) /* might even * be < 0 */
1644 blocks = adjusted_blocks;
1648 blocks *= RtsFlags.GcFlags.oldGenFactor;
1649 if (blocks < RtsFlags.GcFlags.minAllocAreaSize)
1651 blocks = RtsFlags.GcFlags.minAllocAreaSize;
1654 resizeNurseries(blocks);
1656 else // Generational collector
1659 * If the user has given us a suggested heap size, adjust our
1660 * allocation area to make best use of the memory available.
1662 if (RtsFlags.GcFlags.heapSizeSuggestion)
1665 nat needed = calcNeeded(); // approx blocks needed at next GC
1667 /* Guess how much will be live in generation 0 step 0 next time.
1668 * A good approximation is obtained by finding the
1669 * percentage of g0s0 that was live at the last minor GC.
1671 * We have an accurate figure for the amount of copied data in
1672 * 'copied', but we must convert this to a number of blocks, with
1673 * a small adjustment for estimated slop at the end of a block
1678 g0s0_pcnt_kept = ((copied / (BLOCK_SIZE_W - 10)) * 100)
1679 / countNurseryBlocks();
1682 /* Estimate a size for the allocation area based on the
1683 * information available. We might end up going slightly under
1684 * or over the suggested heap size, but we should be pretty
1687 * Formula: suggested - needed
1688 * ----------------------------
1689 * 1 + g0s0_pcnt_kept/100
1691 * where 'needed' is the amount of memory needed at the next
1692 * collection for collecting all steps except g0s0.
1695 (((long)RtsFlags.GcFlags.heapSizeSuggestion - (long)needed) * 100) /
1696 (100 + (long)g0s0_pcnt_kept);
1698 if (blocks < (long)RtsFlags.GcFlags.minAllocAreaSize) {
1699 blocks = RtsFlags.GcFlags.minAllocAreaSize;
1702 resizeNurseries((nat)blocks);
1706 // we might have added extra large blocks to the nursery, so
1707 // resize back to minAllocAreaSize again.
1708 resizeNurseriesFixed(RtsFlags.GcFlags.minAllocAreaSize);
1713 /* -----------------------------------------------------------------------------
1714 Sanity code for CAF garbage collection.
1716 With DEBUG turned on, we manage a CAF list in addition to the SRT
1717 mechanism. After GC, we run down the CAF list and blackhole any
1718 CAFs which have been garbage collected. This means we get an error
1719 whenever the program tries to enter a garbage collected CAF.
1721 Any garbage collected CAFs are taken off the CAF list at the same
1723 -------------------------------------------------------------------------- */
1725 #if 0 && defined(DEBUG)
1732 const StgInfoTable *info;
1743 ASSERT(info->type == IND_STATIC);
1745 if (STATIC_LINK(info,p) == NULL) {
1746 debugTrace(DEBUG_gccafs, "CAF gc'd at 0x%04lx", (long)p);
1748 SET_INFO(p,&stg_BLACKHOLE_info);
1749 p = STATIC_LINK2(info,p);
1753 pp = &STATIC_LINK2(info,p);
1760 debugTrace(DEBUG_gccafs, "%d CAFs live", i);