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);
431 for (bd = ws->scavd_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->step->blocks;
438 ws->step->blocks = ws->scavd_list;
440 ws->step->n_blocks += ws->n_scavd_blocks;
443 for (bd = ws->part_list; bd != NULL; bd = next) {
445 if (bd->free == bd->start) {
447 ws->part_list = next;
454 bd->flags &= ~BF_EVACUATED; // now from-space
455 ws->step->n_words += bd->free - bd->start;
460 prev->link = ws->step->blocks;
461 ws->step->blocks = ws->part_list;
463 ws->step->n_blocks += ws->n_part_blocks;
465 ASSERT(countBlocks(ws->step->blocks) == ws->step->n_blocks);
466 ASSERT(countOccupied(ws->step->blocks) == ws->step->n_words);
471 // Two-space collector: swap the semi-spaces around.
472 // Currently: g0s0->old_blocks is the old nursery
473 // g0s0->blocks is to-space from this GC
474 // We want these the other way around.
475 if (RtsFlags.GcFlags.generations == 1) {
476 bdescr *nursery_blocks = g0s0->old_blocks;
477 nat n_nursery_blocks = g0s0->n_old_blocks;
478 g0s0->old_blocks = g0s0->blocks;
479 g0s0->n_old_blocks = g0s0->n_blocks;
480 g0s0->blocks = nursery_blocks;
481 g0s0->n_blocks = n_nursery_blocks;
484 /* run through all the generations/steps and tidy up
491 for (i=0; i < n_gc_threads; i++) {
492 if (n_gc_threads > 1) {
493 trace(TRACE_gc,"thread %d:", i);
494 trace(TRACE_gc," copied %ld", gc_threads[i]->copied * sizeof(W_));
495 trace(TRACE_gc," scanned %ld", gc_threads[i]->scanned * sizeof(W_));
496 trace(TRACE_gc," any_work %ld", gc_threads[i]->any_work);
497 trace(TRACE_gc," no_work %ld", gc_threads[i]->no_work);
498 trace(TRACE_gc," scav_find_work %ld", gc_threads[i]->scav_find_work);
500 copied += gc_threads[i]->copied;
501 max_copied = stg_max(gc_threads[i]->copied, max_copied);
503 if (n_gc_threads == 1) {
511 for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
514 generations[g].collections++; // for stats
515 if (n_gc_threads > 1) generations[g].par_collections++;
518 // Count the mutable list as bytes "copied" for the purposes of
519 // stats. Every mutable list is copied during every GC.
521 nat mut_list_size = 0;
522 for (bd = generations[g].mut_list; bd != NULL; bd = bd->link) {
523 mut_list_size += bd->free - bd->start;
525 copied += mut_list_size;
528 "mut_list_size: %lu (%d vars, %d arrays, %d MVARs, %d others)",
529 (unsigned long)(mut_list_size * sizeof(W_)),
530 mutlist_MUTVARS, mutlist_MUTARRS, mutlist_MVARS, mutlist_OTHERS);
533 for (s = 0; s < generations[g].n_steps; s++) {
535 stp = &generations[g].steps[s];
537 // for generations we collected...
540 /* free old memory and shift to-space into from-space for all
541 * the collected steps (except the allocation area). These
542 * freed blocks will probaby be quickly recycled.
544 if (!(g == 0 && s == 0 && RtsFlags.GcFlags.generations > 1)) {
545 if (stp->is_compacted)
547 // for a compacted step, just shift the new to-space
548 // onto the front of the now-compacted existing blocks.
549 for (bd = stp->blocks; bd != NULL; bd = bd->link) {
550 bd->flags &= ~BF_EVACUATED; // now from-space
551 stp->n_words += bd->free - bd->start;
553 // tack the new blocks on the end of the existing blocks
554 if (stp->old_blocks != NULL) {
555 for (bd = stp->old_blocks; bd != NULL; bd = next) {
556 // NB. this step might not be compacted next
557 // time, so reset the BF_COMPACTED flags.
558 // They are set before GC if we're going to
559 // compact. (search for BF_COMPACTED above).
560 bd->flags &= ~BF_COMPACTED;
563 bd->link = stp->blocks;
566 stp->blocks = stp->old_blocks;
568 // add the new blocks to the block tally
569 stp->n_blocks += stp->n_old_blocks;
570 ASSERT(countBlocks(stp->blocks) == stp->n_blocks);
571 ASSERT(countOccupied(stp->blocks) == stp->n_words);
575 freeChain(stp->old_blocks);
577 stp->old_blocks = NULL;
578 stp->n_old_blocks = 0;
581 /* LARGE OBJECTS. The current live large objects are chained on
582 * scavenged_large, having been moved during garbage
583 * collection from large_objects. Any objects left on
584 * large_objects list are therefore dead, so we free them here.
586 for (bd = stp->large_objects; bd != NULL; bd = next) {
592 // update the count of blocks used by large objects
593 for (bd = stp->scavenged_large_objects; bd != NULL; bd = bd->link) {
594 bd->flags &= ~BF_EVACUATED;
596 stp->large_objects = stp->scavenged_large_objects;
597 stp->n_large_blocks = stp->n_scavenged_large_blocks;
600 else // for older generations...
602 /* For older generations, we need to append the
603 * scavenged_large_object list (i.e. large objects that have been
604 * promoted during this GC) to the large_object list for that step.
606 for (bd = stp->scavenged_large_objects; bd; bd = next) {
608 bd->flags &= ~BF_EVACUATED;
609 dbl_link_onto(bd, &stp->large_objects);
612 // add the new blocks we promoted during this GC
613 stp->n_large_blocks += stp->n_scavenged_large_blocks;
618 // update the max size of older generations after a major GC
619 resize_generations();
621 // Calculate the amount of live data for stats.
622 live = calcLiveWords();
624 // Free the small objects allocated via allocate(), since this will
625 // all have been copied into G0S1 now.
626 if (RtsFlags.GcFlags.generations > 1) {
627 if (g0s0->blocks != NULL) {
628 freeChain(g0s0->blocks);
635 alloc_blocks_lim = RtsFlags.GcFlags.minAllocAreaSize;
637 // Start a new pinned_object_block
638 pinned_object_block = NULL;
640 // Free the mark stack.
641 if (mark_stack_bdescr != NULL) {
642 freeGroup(mark_stack_bdescr);
646 for (g = 0; g <= N; g++) {
647 for (s = 0; s < generations[g].n_steps; s++) {
648 stp = &generations[g].steps[s];
649 if (stp->bitmap != NULL) {
650 freeGroup(stp->bitmap);
658 // mark the garbage collected CAFs as dead
659 #if 0 && defined(DEBUG) // doesn't work at the moment
660 if (major_gc) { gcCAFs(); }
664 // resetStaticObjectForRetainerProfiling() must be called before
666 if (n_gc_threads > 1) {
667 barf("profiling is currently broken with multi-threaded GC");
668 // ToDo: fix the gct->scavenged_static_objects below
670 resetStaticObjectForRetainerProfiling(gct->scavenged_static_objects);
673 // zero the scavenged static object list
676 for (i = 0; i < n_gc_threads; i++) {
677 zero_static_object_list(gc_threads[i]->scavenged_static_objects);
684 // start any pending finalizers
686 scheduleFinalizers(last_free_capability, old_weak_ptr_list);
689 // send exceptions to any threads which were about to die
691 resurrectThreads(resurrected_threads);
694 // Update the stable pointer hash table.
695 updateStablePtrTable(major_gc);
697 // check sanity after GC
698 IF_DEBUG(sanity, checkSanity());
700 // extra GC trace info
701 if (traceClass(TRACE_gc|DEBUG_gc)) statDescribeGens();
704 // symbol-table based profiling
705 /* heapCensus(to_blocks); */ /* ToDo */
708 // restore enclosing cost centre
714 // check for memory leaks if DEBUG is on
715 memInventory(traceClass(DEBUG_gc));
718 #ifdef RTS_GTK_FRONTPANEL
719 if (RtsFlags.GcFlags.frontpanel) {
720 updateFrontPanelAfterGC( N, live );
724 // ok, GC over: tell the stats department what happened.
725 stat_endGC(allocated, live, copied, N, max_copied, avg_copied);
727 #if defined(RTS_USER_SIGNALS)
728 if (RtsFlags.MiscFlags.install_signal_handlers) {
729 // unblock signals again
730 unblockUserSignals();
739 /* -----------------------------------------------------------------------------
740 * Mark all nodes pointed to by sparks in the spark queues (for GC) Does an
741 * implicit slide i.e. after marking all sparks are at the beginning of the
742 * spark pool and the spark pool only contains sparkable closures
743 * -------------------------------------------------------------------------- */
747 markSparkQueue (evac_fn evac, Capability *cap)
749 StgClosure **sparkp, **to_sparkp;
750 nat n, pruned_sparks; // stats only
753 PAR_TICKY_MARK_SPARK_QUEUE_START();
758 pool = &(cap->r.rSparks);
760 ASSERT_SPARK_POOL_INVARIANTS(pool);
762 #if defined(PARALLEL_HASKELL)
769 to_sparkp = pool->hd;
770 while (sparkp != pool->tl) {
771 ASSERT(*sparkp!=NULL);
772 ASSERT(LOOKS_LIKE_CLOSURE_PTR(((StgClosure *)*sparkp)));
773 // ToDo?: statistics gathering here (also for GUM!)
774 if (closure_SHOULD_SPARK(*sparkp)) {
776 *to_sparkp++ = *sparkp;
777 if (to_sparkp == pool->lim) {
778 to_sparkp = pool->base;
785 if (sparkp == pool->lim) {
789 pool->tl = to_sparkp;
791 PAR_TICKY_MARK_SPARK_QUEUE_END(n);
793 #if defined(PARALLEL_HASKELL)
794 debugTrace(DEBUG_sched,
795 "marked %d sparks and pruned %d sparks on [%x]",
796 n, pruned_sparks, mytid);
798 debugTrace(DEBUG_sched,
799 "marked %d sparks and pruned %d sparks",
803 debugTrace(DEBUG_sched,
804 "new spark queue len=%d; (hd=%p; tl=%p)\n",
805 sparkPoolSize(pool), pool->hd, pool->tl);
809 /* ---------------------------------------------------------------------------
810 Where are the roots that we know about?
812 - all the threads on the runnable queue
813 - all the threads on the blocked queue
814 - all the threads on the sleeping queue
815 - all the thread currently executing a _ccall_GC
816 - all the "main threads"
818 ------------------------------------------------------------------------ */
821 GetRoots( evac_fn evac )
827 // Each GC thread is responsible for following roots from the
828 // Capability of the same number. There will usually be the same
829 // or fewer Capabilities as GC threads, but just in case there
830 // are more, we mark every Capability whose number is the GC
831 // thread's index plus a multiple of the number of GC threads.
832 for (i = gct->thread_index; i < n_capabilities; i += n_gc_threads) {
833 cap = &capabilities[i];
834 evac((StgClosure **)(void *)&cap->run_queue_hd);
835 evac((StgClosure **)(void *)&cap->run_queue_tl);
836 #if defined(THREADED_RTS)
837 evac((StgClosure **)(void *)&cap->wakeup_queue_hd);
838 evac((StgClosure **)(void *)&cap->wakeup_queue_tl);
840 for (task = cap->suspended_ccalling_tasks; task != NULL;
842 debugTrace(DEBUG_sched,
843 "evac'ing suspended TSO %lu", (unsigned long)task->suspended_tso->id);
844 evac((StgClosure **)(void *)&task->suspended_tso);
847 #if defined(THREADED_RTS)
848 markSparkQueue(evac,cap);
852 #if !defined(THREADED_RTS)
853 evac((StgClosure **)(void *)&blocked_queue_hd);
854 evac((StgClosure **)(void *)&blocked_queue_tl);
855 evac((StgClosure **)(void *)&sleeping_queue);
859 /* -----------------------------------------------------------------------------
860 isAlive determines whether the given closure is still alive (after
861 a garbage collection) or not. It returns the new address of the
862 closure if it is alive, or NULL otherwise.
864 NOTE: Use it before compaction only!
865 It untags and (if needed) retags pointers to closures.
866 -------------------------------------------------------------------------- */
870 isAlive(StgClosure *p)
872 const StgInfoTable *info;
878 /* The tag and the pointer are split, to be merged later when needed. */
879 tag = GET_CLOSURE_TAG(p);
880 q = UNTAG_CLOSURE(p);
882 ASSERT(LOOKS_LIKE_CLOSURE_PTR(q));
885 // ignore static closures
887 // ToDo: for static closures, check the static link field.
888 // Problem here is that we sometimes don't set the link field, eg.
889 // for static closures with an empty SRT or CONSTR_STATIC_NOCAFs.
891 if (!HEAP_ALLOCED(q)) {
895 // ignore closures in generations that we're not collecting.
897 if (bd->gen_no > N) {
901 // if it's a pointer into to-space, then we're done
902 if (bd->flags & BF_EVACUATED) {
906 // large objects use the evacuated flag
907 if (bd->flags & BF_LARGE) {
911 // check the mark bit for compacted steps
912 if ((bd->flags & BF_COMPACTED) && is_marked((P_)q,bd)) {
916 switch (info->type) {
921 case IND_OLDGEN: // rely on compatible layout with StgInd
922 case IND_OLDGEN_PERM:
923 // follow indirections
924 p = ((StgInd *)q)->indirectee;
929 return ((StgEvacuated *)q)->evacuee;
932 if (((StgTSO *)q)->what_next == ThreadRelocated) {
933 p = (StgClosure *)((StgTSO *)q)->link;
945 /* -----------------------------------------------------------------------------
946 Figure out which generation to collect, initialise N and major_gc.
948 Also returns the total number of blocks in generations that will be
950 -------------------------------------------------------------------------- */
953 initialise_N (rtsBool force_major_gc)
956 nat s, blocks, blocks_total;
961 if (force_major_gc) {
962 N = RtsFlags.GcFlags.generations - 1;
967 for (g = RtsFlags.GcFlags.generations - 1; g >= 0; g--) {
969 for (s = 0; s < generations[g].n_steps; s++) {
970 blocks += generations[g].steps[s].n_words / BLOCK_SIZE_W;
971 blocks += generations[g].steps[s].n_large_blocks;
973 if (blocks >= generations[g].max_blocks) {
977 blocks_total += blocks;
981 blocks_total += countNurseryBlocks();
983 major_gc = (N == RtsFlags.GcFlags.generations-1);
987 /* -----------------------------------------------------------------------------
988 Initialise the gc_thread structures.
989 -------------------------------------------------------------------------- */
992 alloc_gc_thread (int n)
998 t = stgMallocBytes(sizeof(gc_thread) + total_steps * sizeof(step_workspace),
1003 initCondition(&t->wake_cond);
1004 initMutex(&t->wake_mutex);
1005 t->wakeup = rtsTrue; // starts true, so we can wait for the
1006 // thread to start up, see wakeup_gc_threads
1010 t->thread_index = n;
1011 t->free_blocks = NULL;
1017 t->papi_events = -1;
1020 for (s = 0; s < total_steps; s++)
1023 ws->step = &all_steps[s];
1024 ASSERT(s == ws->step->abs_no);
1028 ws->buffer_todo_bd = NULL;
1030 ws->part_list = NULL;
1031 ws->n_part_blocks = 0;
1033 ws->scavd_list = NULL;
1034 ws->n_scavd_blocks = 0;
1042 alloc_gc_threads (void)
1044 if (gc_threads == NULL) {
1045 #if defined(THREADED_RTS)
1047 gc_threads = stgMallocBytes (RtsFlags.ParFlags.gcThreads *
1049 "alloc_gc_threads");
1051 for (i = 0; i < RtsFlags.ParFlags.gcThreads; i++) {
1052 gc_threads[i] = alloc_gc_thread(i);
1055 gc_threads = stgMallocBytes (sizeof(gc_thread*),
1056 "alloc_gc_threads");
1058 gc_threads[0] = alloc_gc_thread(0);
1063 /* ----------------------------------------------------------------------------
1065 ------------------------------------------------------------------------- */
1067 static nat gc_running_threads;
1069 #if defined(THREADED_RTS)
1070 static Mutex gc_running_mutex;
1077 ACQUIRE_LOCK(&gc_running_mutex);
1078 n_running = ++gc_running_threads;
1079 RELEASE_LOCK(&gc_running_mutex);
1080 ASSERT(n_running <= n_gc_threads);
1088 ACQUIRE_LOCK(&gc_running_mutex);
1089 ASSERT(n_gc_threads != 0);
1090 n_running = --gc_running_threads;
1091 RELEASE_LOCK(&gc_running_mutex);
1096 // gc_thread_work(): Scavenge until there's no work left to do and all
1097 // the running threads are idle.
1100 gc_thread_work (void)
1104 debugTrace(DEBUG_gc, "GC thread %d working", gct->thread_index);
1106 // gc_running_threads has already been incremented for us; either
1107 // this is the main thread and we incremented it inside
1108 // GarbageCollect(), or this is a worker thread and the main
1109 // thread bumped gc_running_threads before waking us up.
1111 // Every thread evacuates some roots.
1113 GetRoots(mark_root);
1117 // scavenge_loop() only exits when there's no work to do
1120 debugTrace(DEBUG_gc, "GC thread %d idle (%d still running)",
1121 gct->thread_index, r);
1123 while (gc_running_threads != 0) {
1129 // any_work() does not remove the work from the queue, it
1130 // just checks for the presence of work. If we find any,
1131 // then we increment gc_running_threads and go back to
1132 // scavenge_loop() to perform any pending work.
1135 // All threads are now stopped
1136 debugTrace(DEBUG_gc, "GC thread %d finished.", gct->thread_index);
1140 #if defined(THREADED_RTS)
1142 gc_thread_mainloop (void)
1144 while (!gct->exit) {
1146 // Wait until we're told to wake up
1147 ACQUIRE_LOCK(&gct->wake_mutex);
1148 gct->wakeup = rtsFalse;
1149 while (!gct->wakeup) {
1150 debugTrace(DEBUG_gc, "GC thread %d standing by...",
1152 waitCondition(&gct->wake_cond, &gct->wake_mutex);
1154 RELEASE_LOCK(&gct->wake_mutex);
1155 if (gct->exit) break;
1158 // start performance counters in this thread...
1159 if (gct->papi_events == -1) {
1160 papi_init_eventset(&gct->papi_events);
1162 papi_thread_start_gc1_count(gct->papi_events);
1168 // count events in this thread towards the GC totals
1169 papi_thread_stop_gc1_count(gct->papi_events);
1175 #if defined(THREADED_RTS)
1177 gc_thread_entry (gc_thread *my_gct)
1180 debugTrace(DEBUG_gc, "GC thread %d starting...", gct->thread_index);
1181 gct->id = osThreadId();
1182 gc_thread_mainloop();
1187 start_gc_threads (void)
1189 #if defined(THREADED_RTS)
1192 static rtsBool done = rtsFalse;
1194 gc_running_threads = 0;
1195 initMutex(&gc_running_mutex);
1198 // Start from 1: the main thread is 0
1199 for (i = 1; i < RtsFlags.ParFlags.gcThreads; i++) {
1200 createOSThread(&id, (OSThreadProc*)&gc_thread_entry,
1209 wakeup_gc_threads (nat n_threads USED_IF_THREADS)
1211 #if defined(THREADED_RTS)
1213 for (i=1; i < n_threads; i++) {
1215 debugTrace(DEBUG_gc, "waking up gc thread %d", i);
1217 ACQUIRE_LOCK(&gc_threads[i]->wake_mutex);
1218 if (gc_threads[i]->wakeup) {
1219 RELEASE_LOCK(&gc_threads[i]->wake_mutex);
1225 gc_threads[i]->wakeup = rtsTrue;
1226 signalCondition(&gc_threads[i]->wake_cond);
1227 RELEASE_LOCK(&gc_threads[i]->wake_mutex);
1232 // After GC is complete, we must wait for all GC threads to enter the
1233 // standby state, otherwise they may still be executing inside
1234 // any_work(), and may even remain awake until the next GC starts.
1236 shutdown_gc_threads (nat n_threads USED_IF_THREADS)
1238 #if defined(THREADED_RTS)
1241 for (i=1; i < n_threads; i++) {
1243 ACQUIRE_LOCK(&gc_threads[i]->wake_mutex);
1244 wakeup = gc_threads[i]->wakeup;
1245 // wakeup is false while the thread is waiting
1246 RELEASE_LOCK(&gc_threads[i]->wake_mutex);
1252 /* ----------------------------------------------------------------------------
1253 Initialise a generation that is to be collected
1254 ------------------------------------------------------------------------- */
1257 init_collected_gen (nat g, nat n_threads)
1264 // Throw away the current mutable list. Invariant: the mutable
1265 // list always has at least one block; this means we can avoid a
1266 // check for NULL in recordMutable().
1268 freeChain(generations[g].mut_list);
1269 generations[g].mut_list = allocBlock();
1270 for (i = 0; i < n_capabilities; i++) {
1271 freeChain(capabilities[i].mut_lists[g]);
1272 capabilities[i].mut_lists[g] = allocBlock();
1276 for (s = 0; s < generations[g].n_steps; s++) {
1278 // generation 0, step 0 doesn't need to-space
1279 if (g == 0 && s == 0 && RtsFlags.GcFlags.generations > 1) {
1283 stp = &generations[g].steps[s];
1284 ASSERT(stp->gen_no == g);
1286 // deprecate the existing blocks
1287 stp->old_blocks = stp->blocks;
1288 stp->n_old_blocks = stp->n_blocks;
1293 // we don't have any to-be-scavenged blocks yet
1295 stp->todos_last = NULL;
1298 // initialise the large object queues.
1299 stp->scavenged_large_objects = NULL;
1300 stp->n_scavenged_large_blocks = 0;
1302 // mark the large objects as not evacuated yet
1303 for (bd = stp->large_objects; bd; bd = bd->link) {
1304 bd->flags &= ~BF_EVACUATED;
1307 // for a compacted step, we need to allocate the bitmap
1308 if (stp->is_compacted) {
1309 nat bitmap_size; // in bytes
1310 bdescr *bitmap_bdescr;
1313 bitmap_size = stp->n_old_blocks * BLOCK_SIZE / (sizeof(W_)*BITS_PER_BYTE);
1315 if (bitmap_size > 0) {
1316 bitmap_bdescr = allocGroup((lnat)BLOCK_ROUND_UP(bitmap_size)
1318 stp->bitmap = bitmap_bdescr;
1319 bitmap = bitmap_bdescr->start;
1321 debugTrace(DEBUG_gc, "bitmap_size: %d, bitmap: %p",
1322 bitmap_size, bitmap);
1324 // don't forget to fill it with zeros!
1325 memset(bitmap, 0, bitmap_size);
1327 // For each block in this step, point to its bitmap from the
1328 // block descriptor.
1329 for (bd=stp->old_blocks; bd != NULL; bd = bd->link) {
1330 bd->u.bitmap = bitmap;
1331 bitmap += BLOCK_SIZE_W / (sizeof(W_)*BITS_PER_BYTE);
1333 // Also at this point we set the BF_COMPACTED flag
1334 // for this block. The invariant is that
1335 // BF_COMPACTED is always unset, except during GC
1336 // when it is set on those blocks which will be
1338 bd->flags |= BF_COMPACTED;
1344 // For each GC thread, for each step, allocate a "todo" block to
1345 // store evacuated objects to be scavenged, and a block to store
1346 // evacuated objects that do not need to be scavenged.
1347 for (t = 0; t < n_threads; t++) {
1348 for (s = 0; s < generations[g].n_steps; s++) {
1350 // we don't copy objects into g0s0, unless -G0
1351 if (g==0 && s==0 && RtsFlags.GcFlags.generations > 1) continue;
1353 ws = &gc_threads[t]->steps[g * RtsFlags.GcFlags.steps + s];
1355 ws->todo_large_objects = NULL;
1357 ws->part_list = NULL;
1358 ws->n_part_blocks = 0;
1360 // allocate the first to-space block; extra blocks will be
1361 // chained on as necessary.
1363 ws->buffer_todo_bd = NULL;
1364 alloc_todo_block(ws,0);
1366 ws->scavd_list = NULL;
1367 ws->n_scavd_blocks = 0;
1373 /* ----------------------------------------------------------------------------
1374 Initialise a generation that is *not* to be collected
1375 ------------------------------------------------------------------------- */
1378 init_uncollected_gen (nat g, nat threads)
1385 for (s = 0; s < generations[g].n_steps; s++) {
1386 stp = &generations[g].steps[s];
1387 stp->scavenged_large_objects = NULL;
1388 stp->n_scavenged_large_blocks = 0;
1391 for (t = 0; t < threads; t++) {
1392 for (s = 0; s < generations[g].n_steps; s++) {
1394 ws = &gc_threads[t]->steps[g * RtsFlags.GcFlags.steps + s];
1397 ws->buffer_todo_bd = NULL;
1398 ws->todo_large_objects = NULL;
1400 ws->part_list = NULL;
1401 ws->n_part_blocks = 0;
1403 ws->scavd_list = NULL;
1404 ws->n_scavd_blocks = 0;
1406 // If the block at the head of the list in this generation
1407 // is less than 3/4 full, then use it as a todo block.
1408 if (stp->blocks && isPartiallyFull(stp->blocks))
1410 ws->todo_bd = stp->blocks;
1411 ws->todo_free = ws->todo_bd->free;
1412 ws->todo_lim = ws->todo_bd->start + BLOCK_SIZE_W;
1413 stp->blocks = stp->blocks->link;
1415 stp->n_words -= ws->todo_bd->free - ws->todo_bd->start;
1416 ws->todo_bd->link = NULL;
1417 // we must scan from the current end point.
1418 ws->todo_bd->u.scan = ws->todo_bd->free;
1423 alloc_todo_block(ws,0);
1428 // Move the private mutable lists from each capability onto the
1429 // main mutable list for the generation.
1430 for (i = 0; i < n_capabilities; i++) {
1431 for (bd = capabilities[i].mut_lists[g];
1432 bd->link != NULL; bd = bd->link) {
1435 bd->link = generations[g].mut_list;
1436 generations[g].mut_list = capabilities[i].mut_lists[g];
1437 capabilities[i].mut_lists[g] = allocBlock();
1441 /* -----------------------------------------------------------------------------
1442 Initialise a gc_thread before GC
1443 -------------------------------------------------------------------------- */
1446 init_gc_thread (gc_thread *t)
1448 t->static_objects = END_OF_STATIC_LIST;
1449 t->scavenged_static_objects = END_OF_STATIC_LIST;
1452 t->failed_to_evac = rtsFalse;
1453 t->eager_promotion = rtsTrue;
1454 t->thunk_selector_depth = 0;
1459 t->scav_find_work = 0;
1462 /* -----------------------------------------------------------------------------
1463 Function we pass to GetRoots to evacuate roots.
1464 -------------------------------------------------------------------------- */
1467 mark_root(StgClosure **root)
1472 /* -----------------------------------------------------------------------------
1473 Initialising the static object & mutable lists
1474 -------------------------------------------------------------------------- */
1477 zero_static_object_list(StgClosure* first_static)
1481 const StgInfoTable *info;
1483 for (p = first_static; p != END_OF_STATIC_LIST; p = link) {
1485 link = *STATIC_LINK(info, p);
1486 *STATIC_LINK(info,p) = NULL;
1490 /* -----------------------------------------------------------------------------
1492 -------------------------------------------------------------------------- */
1499 for (c = (StgIndStatic *)revertible_caf_list; c != NULL;
1500 c = (StgIndStatic *)c->static_link)
1502 SET_INFO(c, c->saved_info);
1503 c->saved_info = NULL;
1504 // could, but not necessary: c->static_link = NULL;
1506 revertible_caf_list = NULL;
1510 markCAFs( evac_fn evac )
1514 for (c = (StgIndStatic *)caf_list; c != NULL;
1515 c = (StgIndStatic *)c->static_link)
1517 evac(&c->indirectee);
1519 for (c = (StgIndStatic *)revertible_caf_list; c != NULL;
1520 c = (StgIndStatic *)c->static_link)
1522 evac(&c->indirectee);
1526 /* ----------------------------------------------------------------------------
1527 Update the pointers from the task list
1529 These are treated as weak pointers because we want to allow a main
1530 thread to get a BlockedOnDeadMVar exception in the same way as any
1531 other thread. Note that the threads should all have been retained
1532 by GC by virtue of being on the all_threads list, we're just
1533 updating pointers here.
1534 ------------------------------------------------------------------------- */
1537 update_task_list (void)
1541 for (task = all_tasks; task != NULL; task = task->all_link) {
1542 if (!task->stopped && task->tso) {
1543 ASSERT(task->tso->bound == task);
1544 tso = (StgTSO *) isAlive((StgClosure *)task->tso);
1546 barf("task %p: main thread %d has been GC'd",
1559 /* ----------------------------------------------------------------------------
1560 Reset the sizes of the older generations when we do a major
1563 CURRENT STRATEGY: make all generations except zero the same size.
1564 We have to stay within the maximum heap size, and leave a certain
1565 percentage of the maximum heap size available to allocate into.
1566 ------------------------------------------------------------------------- */
1569 resize_generations (void)
1573 if (major_gc && RtsFlags.GcFlags.generations > 1) {
1574 nat live, size, min_alloc;
1575 nat max = RtsFlags.GcFlags.maxHeapSize;
1576 nat gens = RtsFlags.GcFlags.generations;
1578 // live in the oldest generations
1579 live = (oldest_gen->steps[0].n_words + BLOCK_SIZE_W - 1) / BLOCK_SIZE_W+
1580 oldest_gen->steps[0].n_large_blocks;
1582 // default max size for all generations except zero
1583 size = stg_max(live * RtsFlags.GcFlags.oldGenFactor,
1584 RtsFlags.GcFlags.minOldGenSize);
1586 // minimum size for generation zero
1587 min_alloc = stg_max((RtsFlags.GcFlags.pcFreeHeap * max) / 200,
1588 RtsFlags.GcFlags.minAllocAreaSize);
1590 // Auto-enable compaction when the residency reaches a
1591 // certain percentage of the maximum heap size (default: 30%).
1592 if (RtsFlags.GcFlags.generations > 1 &&
1593 (RtsFlags.GcFlags.compact ||
1595 oldest_gen->steps[0].n_blocks >
1596 (RtsFlags.GcFlags.compactThreshold * max) / 100))) {
1597 oldest_gen->steps[0].is_compacted = 1;
1598 // debugBelch("compaction: on\n", live);
1600 oldest_gen->steps[0].is_compacted = 0;
1601 // debugBelch("compaction: off\n", live);
1604 // if we're going to go over the maximum heap size, reduce the
1605 // size of the generations accordingly. The calculation is
1606 // different if compaction is turned on, because we don't need
1607 // to double the space required to collect the old generation.
1610 // this test is necessary to ensure that the calculations
1611 // below don't have any negative results - we're working
1612 // with unsigned values here.
1613 if (max < min_alloc) {
1617 if (oldest_gen->steps[0].is_compacted) {
1618 if ( (size + (size - 1) * (gens - 2) * 2) + min_alloc > max ) {
1619 size = (max - min_alloc) / ((gens - 1) * 2 - 1);
1622 if ( (size * (gens - 1) * 2) + min_alloc > max ) {
1623 size = (max - min_alloc) / ((gens - 1) * 2);
1633 debugBelch("live: %d, min_alloc: %d, size : %d, max = %d\n", live,
1634 min_alloc, size, max);
1637 for (g = 0; g < gens; g++) {
1638 generations[g].max_blocks = size;
1643 /* -----------------------------------------------------------------------------
1644 Calculate the new size of the nursery, and resize it.
1645 -------------------------------------------------------------------------- */
1648 resize_nursery (void)
1650 if (RtsFlags.GcFlags.generations == 1)
1651 { // Two-space collector:
1654 /* set up a new nursery. Allocate a nursery size based on a
1655 * function of the amount of live data (by default a factor of 2)
1656 * Use the blocks from the old nursery if possible, freeing up any
1659 * If we get near the maximum heap size, then adjust our nursery
1660 * size accordingly. If the nursery is the same size as the live
1661 * data (L), then we need 3L bytes. We can reduce the size of the
1662 * nursery to bring the required memory down near 2L bytes.
1664 * A normal 2-space collector would need 4L bytes to give the same
1665 * performance we get from 3L bytes, reducing to the same
1666 * performance at 2L bytes.
1668 blocks = g0s0->n_old_blocks;
1670 if ( RtsFlags.GcFlags.maxHeapSize != 0 &&
1671 blocks * RtsFlags.GcFlags.oldGenFactor * 2 >
1672 RtsFlags.GcFlags.maxHeapSize )
1674 long adjusted_blocks; // signed on purpose
1677 adjusted_blocks = (RtsFlags.GcFlags.maxHeapSize - 2 * blocks);
1679 debugTrace(DEBUG_gc, "near maximum heap size of 0x%x blocks, blocks = %d, adjusted to %ld",
1680 RtsFlags.GcFlags.maxHeapSize, blocks, adjusted_blocks);
1682 pc_free = adjusted_blocks * 100 / RtsFlags.GcFlags.maxHeapSize;
1683 if (pc_free < RtsFlags.GcFlags.pcFreeHeap) /* might even * be < 0 */
1687 blocks = adjusted_blocks;
1691 blocks *= RtsFlags.GcFlags.oldGenFactor;
1692 if (blocks < RtsFlags.GcFlags.minAllocAreaSize)
1694 blocks = RtsFlags.GcFlags.minAllocAreaSize;
1697 resizeNurseries(blocks);
1699 else // Generational collector
1702 * If the user has given us a suggested heap size, adjust our
1703 * allocation area to make best use of the memory available.
1705 if (RtsFlags.GcFlags.heapSizeSuggestion)
1708 nat needed = calcNeeded(); // approx blocks needed at next GC
1710 /* Guess how much will be live in generation 0 step 0 next time.
1711 * A good approximation is obtained by finding the
1712 * percentage of g0s0 that was live at the last minor GC.
1714 * We have an accurate figure for the amount of copied data in
1715 * 'copied', but we must convert this to a number of blocks, with
1716 * a small adjustment for estimated slop at the end of a block
1721 g0s0_pcnt_kept = ((copied / (BLOCK_SIZE_W - 10)) * 100)
1722 / countNurseryBlocks();
1725 /* Estimate a size for the allocation area based on the
1726 * information available. We might end up going slightly under
1727 * or over the suggested heap size, but we should be pretty
1730 * Formula: suggested - needed
1731 * ----------------------------
1732 * 1 + g0s0_pcnt_kept/100
1734 * where 'needed' is the amount of memory needed at the next
1735 * collection for collecting all steps except g0s0.
1738 (((long)RtsFlags.GcFlags.heapSizeSuggestion - (long)needed) * 100) /
1739 (100 + (long)g0s0_pcnt_kept);
1741 if (blocks < (long)RtsFlags.GcFlags.minAllocAreaSize) {
1742 blocks = RtsFlags.GcFlags.minAllocAreaSize;
1745 resizeNurseries((nat)blocks);
1749 // we might have added extra large blocks to the nursery, so
1750 // resize back to minAllocAreaSize again.
1751 resizeNurseriesFixed(RtsFlags.GcFlags.minAllocAreaSize);
1756 /* -----------------------------------------------------------------------------
1757 Sanity code for CAF garbage collection.
1759 With DEBUG turned on, we manage a CAF list in addition to the SRT
1760 mechanism. After GC, we run down the CAF list and blackhole any
1761 CAFs which have been garbage collected. This means we get an error
1762 whenever the program tries to enter a garbage collected CAF.
1764 Any garbage collected CAFs are taken off the CAF list at the same
1766 -------------------------------------------------------------------------- */
1768 #if 0 && defined(DEBUG)
1775 const StgInfoTable *info;
1786 ASSERT(info->type == IND_STATIC);
1788 if (STATIC_LINK(info,p) == NULL) {
1789 debugTrace(DEBUG_gccafs, "CAF gc'd at 0x%04lx", (long)p);
1791 SET_INFO(p,&stg_BLACKHOLE_info);
1792 p = STATIC_LINK2(info,p);
1796 pp = &STATIC_LINK2(info,p);
1803 debugTrace(DEBUG_gccafs, "%d CAFs live", i);