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_find_work %ld", gc_threads[i]->scav_find_work);
471 copied += gc_threads[i]->copied;
475 for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
478 generations[g].collections++; // for stats
481 // Count the mutable list as bytes "copied" for the purposes of
482 // stats. Every mutable list is copied during every GC.
484 nat mut_list_size = 0;
485 for (bd = generations[g].mut_list; bd != NULL; bd = bd->link) {
486 mut_list_size += bd->free - bd->start;
488 copied += mut_list_size;
491 "mut_list_size: %lu (%d vars, %d arrays, %d MVARs, %d others)",
492 (unsigned long)(mut_list_size * sizeof(W_)),
493 mutlist_MUTVARS, mutlist_MUTARRS, mutlist_MVARS, mutlist_OTHERS);
496 for (s = 0; s < generations[g].n_steps; s++) {
498 stp = &generations[g].steps[s];
500 // for generations we collected...
503 /* free old memory and shift to-space into from-space for all
504 * the collected steps (except the allocation area). These
505 * freed blocks will probaby be quickly recycled.
507 if (!(g == 0 && s == 0 && RtsFlags.GcFlags.generations > 1)) {
508 if (stp->is_compacted)
510 // for a compacted step, just shift the new to-space
511 // onto the front of the now-compacted existing blocks.
512 for (bd = stp->blocks; bd != NULL; bd = bd->link) {
513 bd->flags &= ~BF_EVACUATED; // now from-space
515 // tack the new blocks on the end of the existing blocks
516 if (stp->old_blocks != NULL) {
517 for (bd = stp->old_blocks; bd != NULL; bd = next) {
518 // NB. this step might not be compacted next
519 // time, so reset the BF_COMPACTED flags.
520 // They are set before GC if we're going to
521 // compact. (search for BF_COMPACTED above).
522 bd->flags &= ~BF_COMPACTED;
525 bd->link = stp->blocks;
528 stp->blocks = stp->old_blocks;
530 // add the new blocks to the block tally
531 stp->n_blocks += stp->n_old_blocks;
532 ASSERT(countBlocks(stp->blocks) == stp->n_blocks);
536 freeChain(stp->old_blocks);
538 stp->old_blocks = NULL;
539 stp->n_old_blocks = 0;
542 /* LARGE OBJECTS. The current live large objects are chained on
543 * scavenged_large, having been moved during garbage
544 * collection from large_objects. Any objects left on
545 * large_objects list are therefore dead, so we free them here.
547 for (bd = stp->large_objects; bd != NULL; bd = next) {
553 // update the count of blocks used by large objects
554 for (bd = stp->scavenged_large_objects; bd != NULL; bd = bd->link) {
555 bd->flags &= ~BF_EVACUATED;
557 stp->large_objects = stp->scavenged_large_objects;
558 stp->n_large_blocks = stp->n_scavenged_large_blocks;
561 else // for older generations...
563 /* For older generations, we need to append the
564 * scavenged_large_object list (i.e. large objects that have been
565 * promoted during this GC) to the large_object list for that step.
567 for (bd = stp->scavenged_large_objects; bd; bd = next) {
569 bd->flags &= ~BF_EVACUATED;
570 dbl_link_onto(bd, &stp->large_objects);
573 // add the new blocks we promoted during this GC
574 stp->n_large_blocks += stp->n_scavenged_large_blocks;
579 // update the max size of older generations after a major GC
580 resize_generations();
582 // Guess the amount of live data for stats.
583 live = calcLiveBlocks() * BLOCK_SIZE_W;
584 debugTrace(DEBUG_gc, "Slop: %ldKB",
585 (live - calcLiveWords()) / (1024/sizeof(W_)));
587 // Free the small objects allocated via allocate(), since this will
588 // all have been copied into G0S1 now.
589 if (RtsFlags.GcFlags.generations > 1) {
590 if (g0s0->blocks != NULL) {
591 freeChain(g0s0->blocks);
597 alloc_blocks_lim = RtsFlags.GcFlags.minAllocAreaSize;
599 // Start a new pinned_object_block
600 pinned_object_block = NULL;
602 // Free the mark stack.
603 if (mark_stack_bdescr != NULL) {
604 freeGroup(mark_stack_bdescr);
608 for (g = 0; g <= N; g++) {
609 for (s = 0; s < generations[g].n_steps; s++) {
610 stp = &generations[g].steps[s];
611 if (stp->bitmap != NULL) {
612 freeGroup(stp->bitmap);
620 // mark the garbage collected CAFs as dead
621 #if 0 && defined(DEBUG) // doesn't work at the moment
622 if (major_gc) { gcCAFs(); }
626 // resetStaticObjectForRetainerProfiling() must be called before
628 if (n_gc_threads > 1) {
629 barf("profiling is currently broken with multi-threaded GC");
630 // ToDo: fix the gct->scavenged_static_objects below
632 resetStaticObjectForRetainerProfiling(gct->scavenged_static_objects);
635 // zero the scavenged static object list
638 for (i = 0; i < n_gc_threads; i++) {
639 zero_static_object_list(gc_threads[i]->scavenged_static_objects);
646 // start any pending finalizers
648 scheduleFinalizers(last_free_capability, old_weak_ptr_list);
651 // send exceptions to any threads which were about to die
653 resurrectThreads(resurrected_threads);
656 // Update the stable pointer hash table.
657 updateStablePtrTable(major_gc);
659 // check sanity after GC
660 IF_DEBUG(sanity, checkSanity());
662 // extra GC trace info
663 if (traceClass(TRACE_gc|DEBUG_gc)) statDescribeGens();
666 // symbol-table based profiling
667 /* heapCensus(to_blocks); */ /* ToDo */
670 // restore enclosing cost centre
676 // check for memory leaks if DEBUG is on
677 memInventory(traceClass(DEBUG_gc));
680 #ifdef RTS_GTK_FRONTPANEL
681 if (RtsFlags.GcFlags.frontpanel) {
682 updateFrontPanelAfterGC( N, live );
686 // ok, GC over: tell the stats department what happened.
687 stat_endGC(allocated, live, copied, N);
689 #if defined(RTS_USER_SIGNALS)
690 if (RtsFlags.MiscFlags.install_signal_handlers) {
691 // unblock signals again
692 unblockUserSignals();
701 /* -----------------------------------------------------------------------------
702 * Mark all nodes pointed to by sparks in the spark queues (for GC) Does an
703 * implicit slide i.e. after marking all sparks are at the beginning of the
704 * spark pool and the spark pool only contains sparkable closures
705 * -------------------------------------------------------------------------- */
709 markSparkQueue (evac_fn evac, Capability *cap)
711 StgClosure **sparkp, **to_sparkp;
712 nat n, pruned_sparks; // stats only
715 PAR_TICKY_MARK_SPARK_QUEUE_START();
720 pool = &(cap->r.rSparks);
722 ASSERT_SPARK_POOL_INVARIANTS(pool);
724 #if defined(PARALLEL_HASKELL)
731 to_sparkp = pool->hd;
732 while (sparkp != pool->tl) {
733 ASSERT(*sparkp!=NULL);
734 ASSERT(LOOKS_LIKE_CLOSURE_PTR(((StgClosure *)*sparkp)));
735 // ToDo?: statistics gathering here (also for GUM!)
736 if (closure_SHOULD_SPARK(*sparkp)) {
738 *to_sparkp++ = *sparkp;
739 if (to_sparkp == pool->lim) {
740 to_sparkp = pool->base;
747 if (sparkp == pool->lim) {
751 pool->tl = to_sparkp;
753 PAR_TICKY_MARK_SPARK_QUEUE_END(n);
755 #if defined(PARALLEL_HASKELL)
756 debugTrace(DEBUG_sched,
757 "marked %d sparks and pruned %d sparks on [%x]",
758 n, pruned_sparks, mytid);
760 debugTrace(DEBUG_sched,
761 "marked %d sparks and pruned %d sparks",
765 debugTrace(DEBUG_sched,
766 "new spark queue len=%d; (hd=%p; tl=%p)\n",
767 sparkPoolSize(pool), pool->hd, pool->tl);
771 /* ---------------------------------------------------------------------------
772 Where are the roots that we know about?
774 - all the threads on the runnable queue
775 - all the threads on the blocked queue
776 - all the threads on the sleeping queue
777 - all the thread currently executing a _ccall_GC
778 - all the "main threads"
780 ------------------------------------------------------------------------ */
783 GetRoots( evac_fn evac )
789 // Each GC thread is responsible for following roots from the
790 // Capability of the same number. There will usually be the same
791 // or fewer Capabilities as GC threads, but just in case there
792 // are more, we mark every Capability whose number is the GC
793 // thread's index plus a multiple of the number of GC threads.
794 for (i = gct->thread_index; i < n_capabilities; i += n_gc_threads) {
795 cap = &capabilities[i];
796 evac((StgClosure **)(void *)&cap->run_queue_hd);
797 evac((StgClosure **)(void *)&cap->run_queue_tl);
798 #if defined(THREADED_RTS)
799 evac((StgClosure **)(void *)&cap->wakeup_queue_hd);
800 evac((StgClosure **)(void *)&cap->wakeup_queue_tl);
802 for (task = cap->suspended_ccalling_tasks; task != NULL;
804 debugTrace(DEBUG_sched,
805 "evac'ing suspended TSO %lu", (unsigned long)task->suspended_tso->id);
806 evac((StgClosure **)(void *)&task->suspended_tso);
809 #if defined(THREADED_RTS)
810 markSparkQueue(evac,cap);
814 #if !defined(THREADED_RTS)
815 evac((StgClosure **)(void *)&blocked_queue_hd);
816 evac((StgClosure **)(void *)&blocked_queue_tl);
817 evac((StgClosure **)(void *)&sleeping_queue);
821 /* -----------------------------------------------------------------------------
822 isAlive determines whether the given closure is still alive (after
823 a garbage collection) or not. It returns the new address of the
824 closure if it is alive, or NULL otherwise.
826 NOTE: Use it before compaction only!
827 It untags and (if needed) retags pointers to closures.
828 -------------------------------------------------------------------------- */
832 isAlive(StgClosure *p)
834 const StgInfoTable *info;
840 /* The tag and the pointer are split, to be merged later when needed. */
841 tag = GET_CLOSURE_TAG(p);
842 q = UNTAG_CLOSURE(p);
844 ASSERT(LOOKS_LIKE_CLOSURE_PTR(q));
847 // ignore static closures
849 // ToDo: for static closures, check the static link field.
850 // Problem here is that we sometimes don't set the link field, eg.
851 // for static closures with an empty SRT or CONSTR_STATIC_NOCAFs.
853 if (!HEAP_ALLOCED(q)) {
857 // ignore closures in generations that we're not collecting.
859 if (bd->gen_no > N) {
863 // if it's a pointer into to-space, then we're done
864 if (bd->flags & BF_EVACUATED) {
868 // large objects use the evacuated flag
869 if (bd->flags & BF_LARGE) {
873 // check the mark bit for compacted steps
874 if ((bd->flags & BF_COMPACTED) && is_marked((P_)q,bd)) {
878 switch (info->type) {
883 case IND_OLDGEN: // rely on compatible layout with StgInd
884 case IND_OLDGEN_PERM:
885 // follow indirections
886 p = ((StgInd *)q)->indirectee;
891 return ((StgEvacuated *)q)->evacuee;
894 if (((StgTSO *)q)->what_next == ThreadRelocated) {
895 p = (StgClosure *)((StgTSO *)q)->link;
907 /* -----------------------------------------------------------------------------
908 Figure out which generation to collect, initialise N and major_gc.
910 Also returns the total number of blocks in generations that will be
912 -------------------------------------------------------------------------- */
915 initialise_N (rtsBool force_major_gc)
918 nat s, blocks, blocks_total;
923 if (force_major_gc) {
924 N = RtsFlags.GcFlags.generations - 1;
929 for (g = RtsFlags.GcFlags.generations - 1; g >= 0; g--) {
931 for (s = 0; s < generations[g].n_steps; s++) {
932 blocks += generations[g].steps[s].n_blocks;
933 blocks += generations[g].steps[s].n_large_blocks;
935 if (blocks >= generations[g].max_blocks) {
939 blocks_total += blocks;
943 blocks_total += countNurseryBlocks();
945 major_gc = (N == RtsFlags.GcFlags.generations-1);
949 /* -----------------------------------------------------------------------------
950 Initialise the gc_thread structures.
951 -------------------------------------------------------------------------- */
954 alloc_gc_thread (int n)
960 t = stgMallocBytes(sizeof(gc_thread) + total_steps * sizeof(step_workspace),
965 initCondition(&t->wake_cond);
966 initMutex(&t->wake_mutex);
967 t->wakeup = rtsFalse;
972 t->free_blocks = NULL;
981 for (s = 0; s < total_steps; s++)
984 ws->stp = &all_steps[s];
985 ASSERT(s == ws->stp->abs_no);
992 ws->buffer_todo_bd = NULL;
994 ws->scavd_list = NULL;
995 ws->n_scavd_blocks = 0;
1003 alloc_gc_threads (void)
1005 if (gc_threads == NULL) {
1006 #if defined(THREADED_RTS)
1008 gc_threads = stgMallocBytes (RtsFlags.ParFlags.gcThreads *
1010 "alloc_gc_threads");
1012 for (i = 0; i < RtsFlags.ParFlags.gcThreads; i++) {
1013 gc_threads[i] = alloc_gc_thread(i);
1016 gc_threads = stgMallocBytes (sizeof(gc_thread*),
1017 "alloc_gc_threads");
1019 gc_threads[0] = alloc_gc_thread(0);
1024 /* ----------------------------------------------------------------------------
1026 ------------------------------------------------------------------------- */
1028 static nat gc_running_threads;
1030 #if defined(THREADED_RTS)
1031 static Mutex gc_running_mutex;
1038 ACQUIRE_LOCK(&gc_running_mutex);
1039 n_running = ++gc_running_threads;
1040 RELEASE_LOCK(&gc_running_mutex);
1041 ASSERT(n_running <= n_gc_threads);
1049 ACQUIRE_LOCK(&gc_running_mutex);
1050 ASSERT(n_gc_threads != 0);
1051 n_running = --gc_running_threads;
1052 RELEASE_LOCK(&gc_running_mutex);
1057 // gc_thread_work(): Scavenge until there's no work left to do and all
1058 // the running threads are idle.
1061 gc_thread_work (void)
1065 debugTrace(DEBUG_gc, "GC thread %d working", gct->thread_index);
1067 // gc_running_threads has already been incremented for us; either
1068 // this is the main thread and we incremented it inside
1069 // GarbageCollect(), or this is a worker thread and the main
1070 // thread bumped gc_running_threads before waking us up.
1072 // Every thread evacuates some roots.
1074 GetRoots(mark_root);
1078 // scavenge_loop() only exits when there's no work to do
1081 debugTrace(DEBUG_gc, "GC thread %d idle (%d still running)",
1082 gct->thread_index, r);
1084 while (gc_running_threads != 0) {
1090 // any_work() does not remove the work from the queue, it
1091 // just checks for the presence of work. If we find any,
1092 // then we increment gc_running_threads and go back to
1093 // scavenge_loop() to perform any pending work.
1096 // All threads are now stopped
1097 debugTrace(DEBUG_gc, "GC thread %d finished.", gct->thread_index);
1101 #if defined(THREADED_RTS)
1103 gc_thread_mainloop (void)
1105 while (!gct->exit) {
1107 // Wait until we're told to wake up
1108 ACQUIRE_LOCK(&gct->wake_mutex);
1109 gct->wakeup = rtsFalse;
1110 while (!gct->wakeup) {
1111 debugTrace(DEBUG_gc, "GC thread %d standing by...",
1113 waitCondition(&gct->wake_cond, &gct->wake_mutex);
1115 RELEASE_LOCK(&gct->wake_mutex);
1116 if (gct->exit) break;
1119 // start performance counters in this thread...
1120 if (gct->papi_events == -1) {
1121 papi_init_eventset(&gct->papi_events);
1123 papi_thread_start_gc1_count(gct->papi_events);
1129 // count events in this thread towards the GC totals
1130 papi_thread_stop_gc1_count(gct->papi_events);
1136 #if defined(THREADED_RTS)
1138 gc_thread_entry (gc_thread *my_gct)
1141 debugTrace(DEBUG_gc, "GC thread %d starting...", gct->thread_index);
1142 gct->id = osThreadId();
1143 gc_thread_mainloop();
1148 start_gc_threads (void)
1150 #if defined(THREADED_RTS)
1153 static rtsBool done = rtsFalse;
1155 gc_running_threads = 0;
1156 initMutex(&gc_running_mutex);
1159 // Start from 1: the main thread is 0
1160 for (i = 1; i < RtsFlags.ParFlags.gcThreads; i++) {
1161 createOSThread(&id, (OSThreadProc*)&gc_thread_entry,
1170 wakeup_gc_threads (nat n_threads USED_IF_THREADS)
1172 #if defined(THREADED_RTS)
1174 for (i=1; i < n_threads; i++) {
1176 ACQUIRE_LOCK(&gc_threads[i]->wake_mutex);
1177 gc_threads[i]->wakeup = rtsTrue;
1178 signalCondition(&gc_threads[i]->wake_cond);
1179 RELEASE_LOCK(&gc_threads[i]->wake_mutex);
1184 // After GC is complete, we must wait for all GC threads to enter the
1185 // standby state, otherwise they may still be executing inside
1186 // any_work(), and may even remain awake until the next GC starts.
1188 shutdown_gc_threads (nat n_threads USED_IF_THREADS)
1190 #if defined(THREADED_RTS)
1193 for (i=1; i < n_threads; i++) {
1195 ACQUIRE_LOCK(&gc_threads[i]->wake_mutex);
1196 wakeup = gc_threads[i]->wakeup;
1197 // wakeup is false while the thread is waiting
1198 RELEASE_LOCK(&gc_threads[i]->wake_mutex);
1204 /* ----------------------------------------------------------------------------
1205 Initialise a generation that is to be collected
1206 ------------------------------------------------------------------------- */
1209 init_collected_gen (nat g, nat n_threads)
1216 // Throw away the current mutable list. Invariant: the mutable
1217 // list always has at least one block; this means we can avoid a
1218 // check for NULL in recordMutable().
1220 freeChain(generations[g].mut_list);
1221 generations[g].mut_list = allocBlock();
1222 for (i = 0; i < n_capabilities; i++) {
1223 freeChain(capabilities[i].mut_lists[g]);
1224 capabilities[i].mut_lists[g] = allocBlock();
1228 for (s = 0; s < generations[g].n_steps; s++) {
1230 // generation 0, step 0 doesn't need to-space
1231 if (g == 0 && s == 0 && RtsFlags.GcFlags.generations > 1) {
1235 stp = &generations[g].steps[s];
1236 ASSERT(stp->gen_no == g);
1238 // deprecate the existing blocks
1239 stp->old_blocks = stp->blocks;
1240 stp->n_old_blocks = stp->n_blocks;
1244 // we don't have any to-be-scavenged blocks yet
1246 stp->todos_last = NULL;
1249 // initialise the large object queues.
1250 stp->scavenged_large_objects = NULL;
1251 stp->n_scavenged_large_blocks = 0;
1253 // mark the large objects as not evacuated yet
1254 for (bd = stp->large_objects; bd; bd = bd->link) {
1255 bd->flags &= ~BF_EVACUATED;
1258 // for a compacted step, we need to allocate the bitmap
1259 if (stp->is_compacted) {
1260 nat bitmap_size; // in bytes
1261 bdescr *bitmap_bdescr;
1264 bitmap_size = stp->n_old_blocks * BLOCK_SIZE / (sizeof(W_)*BITS_PER_BYTE);
1266 if (bitmap_size > 0) {
1267 bitmap_bdescr = allocGroup((lnat)BLOCK_ROUND_UP(bitmap_size)
1269 stp->bitmap = bitmap_bdescr;
1270 bitmap = bitmap_bdescr->start;
1272 debugTrace(DEBUG_gc, "bitmap_size: %d, bitmap: %p",
1273 bitmap_size, bitmap);
1275 // don't forget to fill it with zeros!
1276 memset(bitmap, 0, bitmap_size);
1278 // For each block in this step, point to its bitmap from the
1279 // block descriptor.
1280 for (bd=stp->old_blocks; bd != NULL; bd = bd->link) {
1281 bd->u.bitmap = bitmap;
1282 bitmap += BLOCK_SIZE_W / (sizeof(W_)*BITS_PER_BYTE);
1284 // Also at this point we set the BF_COMPACTED flag
1285 // for this block. The invariant is that
1286 // BF_COMPACTED is always unset, except during GC
1287 // when it is set on those blocks which will be
1289 bd->flags |= BF_COMPACTED;
1295 // For each GC thread, for each step, allocate a "todo" block to
1296 // store evacuated objects to be scavenged, and a block to store
1297 // evacuated objects that do not need to be scavenged.
1298 for (t = 0; t < n_threads; t++) {
1299 for (s = 0; s < generations[g].n_steps; s++) {
1301 // we don't copy objects into g0s0, unless -G0
1302 if (g==0 && s==0 && RtsFlags.GcFlags.generations > 1) continue;
1304 ws = &gc_threads[t]->steps[g * RtsFlags.GcFlags.steps + s];
1309 ws->todo_large_objects = NULL;
1311 // allocate the first to-space block; extra blocks will be
1312 // chained on as necessary.
1314 ws->buffer_todo_bd = NULL;
1315 gc_alloc_todo_block(ws);
1317 ws->scavd_list = NULL;
1318 ws->n_scavd_blocks = 0;
1324 /* ----------------------------------------------------------------------------
1325 Initialise a generation that is *not* to be collected
1326 ------------------------------------------------------------------------- */
1329 init_uncollected_gen (nat g, nat threads)
1336 for (s = 0; s < generations[g].n_steps; s++) {
1337 stp = &generations[g].steps[s];
1338 stp->scavenged_large_objects = NULL;
1339 stp->n_scavenged_large_blocks = 0;
1342 for (t = 0; t < threads; t++) {
1343 for (s = 0; s < generations[g].n_steps; s++) {
1345 ws = &gc_threads[t]->steps[g * RtsFlags.GcFlags.steps + s];
1348 ws->buffer_todo_bd = NULL;
1349 ws->todo_large_objects = NULL;
1351 ws->scavd_list = NULL;
1352 ws->n_scavd_blocks = 0;
1354 // If the block at the head of the list in this generation
1355 // is less than 3/4 full, then use it as a todo block.
1356 if (stp->blocks && isPartiallyFull(stp->blocks))
1358 ws->todo_bd = stp->blocks;
1359 ws->todo_free = ws->todo_bd->free;
1360 ws->todo_lim = ws->todo_bd->start + BLOCK_SIZE_W;
1361 stp->blocks = stp->blocks->link;
1363 ws->todo_bd->link = NULL;
1365 // this block is also the scan block; we must scan
1366 // from the current end point.
1367 ws->scan_bd = ws->todo_bd;
1368 ws->scan = ws->scan_bd->free;
1370 // subtract the contents of this block from the stats,
1371 // because we'll count the whole block later.
1372 copied -= ws->scan_bd->free - ws->scan_bd->start;
1379 gc_alloc_todo_block(ws);
1384 // Move the private mutable lists from each capability onto the
1385 // main mutable list for the generation.
1386 for (i = 0; i < n_capabilities; i++) {
1387 for (bd = capabilities[i].mut_lists[g];
1388 bd->link != NULL; bd = bd->link) {
1391 bd->link = generations[g].mut_list;
1392 generations[g].mut_list = capabilities[i].mut_lists[g];
1393 capabilities[i].mut_lists[g] = allocBlock();
1397 /* -----------------------------------------------------------------------------
1398 Initialise a gc_thread before GC
1399 -------------------------------------------------------------------------- */
1402 init_gc_thread (gc_thread *t)
1404 t->static_objects = END_OF_STATIC_LIST;
1405 t->scavenged_static_objects = END_OF_STATIC_LIST;
1407 t->failed_to_evac = rtsFalse;
1408 t->eager_promotion = rtsTrue;
1409 t->thunk_selector_depth = 0;
1413 t->scav_find_work = 0;
1417 /* -----------------------------------------------------------------------------
1418 Function we pass to GetRoots to evacuate roots.
1419 -------------------------------------------------------------------------- */
1422 mark_root(StgClosure **root)
1427 /* -----------------------------------------------------------------------------
1428 Initialising the static object & mutable lists
1429 -------------------------------------------------------------------------- */
1432 zero_static_object_list(StgClosure* first_static)
1436 const StgInfoTable *info;
1438 for (p = first_static; p != END_OF_STATIC_LIST; p = link) {
1440 link = *STATIC_LINK(info, p);
1441 *STATIC_LINK(info,p) = NULL;
1445 /* -----------------------------------------------------------------------------
1447 -------------------------------------------------------------------------- */
1454 for (c = (StgIndStatic *)revertible_caf_list; c != NULL;
1455 c = (StgIndStatic *)c->static_link)
1457 SET_INFO(c, c->saved_info);
1458 c->saved_info = NULL;
1459 // could, but not necessary: c->static_link = NULL;
1461 revertible_caf_list = NULL;
1465 markCAFs( evac_fn evac )
1469 for (c = (StgIndStatic *)caf_list; c != NULL;
1470 c = (StgIndStatic *)c->static_link)
1472 evac(&c->indirectee);
1474 for (c = (StgIndStatic *)revertible_caf_list; c != NULL;
1475 c = (StgIndStatic *)c->static_link)
1477 evac(&c->indirectee);
1481 /* ----------------------------------------------------------------------------
1482 Update the pointers from the task list
1484 These are treated as weak pointers because we want to allow a main
1485 thread to get a BlockedOnDeadMVar exception in the same way as any
1486 other thread. Note that the threads should all have been retained
1487 by GC by virtue of being on the all_threads list, we're just
1488 updating pointers here.
1489 ------------------------------------------------------------------------- */
1492 update_task_list (void)
1496 for (task = all_tasks; task != NULL; task = task->all_link) {
1497 if (!task->stopped && task->tso) {
1498 ASSERT(task->tso->bound == task);
1499 tso = (StgTSO *) isAlive((StgClosure *)task->tso);
1501 barf("task %p: main thread %d has been GC'd",
1514 /* ----------------------------------------------------------------------------
1515 Reset the sizes of the older generations when we do a major
1518 CURRENT STRATEGY: make all generations except zero the same size.
1519 We have to stay within the maximum heap size, and leave a certain
1520 percentage of the maximum heap size available to allocate into.
1521 ------------------------------------------------------------------------- */
1524 resize_generations (void)
1528 if (major_gc && RtsFlags.GcFlags.generations > 1) {
1529 nat live, size, min_alloc;
1530 nat max = RtsFlags.GcFlags.maxHeapSize;
1531 nat gens = RtsFlags.GcFlags.generations;
1533 // live in the oldest generations
1534 live = oldest_gen->steps[0].n_blocks +
1535 oldest_gen->steps[0].n_large_blocks;
1537 // default max size for all generations except zero
1538 size = stg_max(live * RtsFlags.GcFlags.oldGenFactor,
1539 RtsFlags.GcFlags.minOldGenSize);
1541 // minimum size for generation zero
1542 min_alloc = stg_max((RtsFlags.GcFlags.pcFreeHeap * max) / 200,
1543 RtsFlags.GcFlags.minAllocAreaSize);
1545 // Auto-enable compaction when the residency reaches a
1546 // certain percentage of the maximum heap size (default: 30%).
1547 if (RtsFlags.GcFlags.generations > 1 &&
1548 (RtsFlags.GcFlags.compact ||
1550 oldest_gen->steps[0].n_blocks >
1551 (RtsFlags.GcFlags.compactThreshold * max) / 100))) {
1552 oldest_gen->steps[0].is_compacted = 1;
1553 // debugBelch("compaction: on\n", live);
1555 oldest_gen->steps[0].is_compacted = 0;
1556 // debugBelch("compaction: off\n", live);
1559 // if we're going to go over the maximum heap size, reduce the
1560 // size of the generations accordingly. The calculation is
1561 // different if compaction is turned on, because we don't need
1562 // to double the space required to collect the old generation.
1565 // this test is necessary to ensure that the calculations
1566 // below don't have any negative results - we're working
1567 // with unsigned values here.
1568 if (max < min_alloc) {
1572 if (oldest_gen->steps[0].is_compacted) {
1573 if ( (size + (size - 1) * (gens - 2) * 2) + min_alloc > max ) {
1574 size = (max - min_alloc) / ((gens - 1) * 2 - 1);
1577 if ( (size * (gens - 1) * 2) + min_alloc > max ) {
1578 size = (max - min_alloc) / ((gens - 1) * 2);
1588 debugBelch("live: %d, min_alloc: %d, size : %d, max = %d\n", live,
1589 min_alloc, size, max);
1592 for (g = 0; g < gens; g++) {
1593 generations[g].max_blocks = size;
1598 /* -----------------------------------------------------------------------------
1599 Calculate the new size of the nursery, and resize it.
1600 -------------------------------------------------------------------------- */
1603 resize_nursery (void)
1605 if (RtsFlags.GcFlags.generations == 1)
1606 { // Two-space collector:
1609 /* set up a new nursery. Allocate a nursery size based on a
1610 * function of the amount of live data (by default a factor of 2)
1611 * Use the blocks from the old nursery if possible, freeing up any
1614 * If we get near the maximum heap size, then adjust our nursery
1615 * size accordingly. If the nursery is the same size as the live
1616 * data (L), then we need 3L bytes. We can reduce the size of the
1617 * nursery to bring the required memory down near 2L bytes.
1619 * A normal 2-space collector would need 4L bytes to give the same
1620 * performance we get from 3L bytes, reducing to the same
1621 * performance at 2L bytes.
1623 blocks = g0s0->n_old_blocks;
1625 if ( RtsFlags.GcFlags.maxHeapSize != 0 &&
1626 blocks * RtsFlags.GcFlags.oldGenFactor * 2 >
1627 RtsFlags.GcFlags.maxHeapSize )
1629 long adjusted_blocks; // signed on purpose
1632 adjusted_blocks = (RtsFlags.GcFlags.maxHeapSize - 2 * blocks);
1634 debugTrace(DEBUG_gc, "near maximum heap size of 0x%x blocks, blocks = %d, adjusted to %ld",
1635 RtsFlags.GcFlags.maxHeapSize, blocks, adjusted_blocks);
1637 pc_free = adjusted_blocks * 100 / RtsFlags.GcFlags.maxHeapSize;
1638 if (pc_free < RtsFlags.GcFlags.pcFreeHeap) /* might even * be < 0 */
1642 blocks = adjusted_blocks;
1646 blocks *= RtsFlags.GcFlags.oldGenFactor;
1647 if (blocks < RtsFlags.GcFlags.minAllocAreaSize)
1649 blocks = RtsFlags.GcFlags.minAllocAreaSize;
1652 resizeNurseries(blocks);
1654 else // Generational collector
1657 * If the user has given us a suggested heap size, adjust our
1658 * allocation area to make best use of the memory available.
1660 if (RtsFlags.GcFlags.heapSizeSuggestion)
1663 nat needed = calcNeeded(); // approx blocks needed at next GC
1665 /* Guess how much will be live in generation 0 step 0 next time.
1666 * A good approximation is obtained by finding the
1667 * percentage of g0s0 that was live at the last minor GC.
1669 * We have an accurate figure for the amount of copied data in
1670 * 'copied', but we must convert this to a number of blocks, with
1671 * a small adjustment for estimated slop at the end of a block
1676 g0s0_pcnt_kept = ((copied / (BLOCK_SIZE_W - 10)) * 100)
1677 / countNurseryBlocks();
1680 /* Estimate a size for the allocation area based on the
1681 * information available. We might end up going slightly under
1682 * or over the suggested heap size, but we should be pretty
1685 * Formula: suggested - needed
1686 * ----------------------------
1687 * 1 + g0s0_pcnt_kept/100
1689 * where 'needed' is the amount of memory needed at the next
1690 * collection for collecting all steps except g0s0.
1693 (((long)RtsFlags.GcFlags.heapSizeSuggestion - (long)needed) * 100) /
1694 (100 + (long)g0s0_pcnt_kept);
1696 if (blocks < (long)RtsFlags.GcFlags.minAllocAreaSize) {
1697 blocks = RtsFlags.GcFlags.minAllocAreaSize;
1700 resizeNurseries((nat)blocks);
1704 // we might have added extra large blocks to the nursery, so
1705 // resize back to minAllocAreaSize again.
1706 resizeNurseriesFixed(RtsFlags.GcFlags.minAllocAreaSize);
1711 /* -----------------------------------------------------------------------------
1712 Sanity code for CAF garbage collection.
1714 With DEBUG turned on, we manage a CAF list in addition to the SRT
1715 mechanism. After GC, we run down the CAF list and blackhole any
1716 CAFs which have been garbage collected. This means we get an error
1717 whenever the program tries to enter a garbage collected CAF.
1719 Any garbage collected CAFs are taken off the CAF list at the same
1721 -------------------------------------------------------------------------- */
1723 #if 0 && defined(DEBUG)
1730 const StgInfoTable *info;
1741 ASSERT(info->type == IND_STATIC);
1743 if (STATIC_LINK(info,p) == NULL) {
1744 debugTrace(DEBUG_gccafs, "CAF gc'd at 0x%04lx", (long)p);
1746 SET_INFO(p,&stg_BLACKHOLE_info);
1747 p = STATIC_LINK2(info,p);
1751 pp = &STATIC_LINK2(info,p);
1758 debugTrace(DEBUG_gccafs, "%d CAFs live", i);