1 /* -----------------------------------------------------------------------------
3 * (c) The GHC Team 1998-2008
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"
25 #include "BlockAlloc.h"
29 #include "RtsSignals.h"
31 #if defined(RTS_GTK_FRONTPANEL)
32 #include "FrontPanel.h"
35 #include "RetainerProfile.h"
36 #include "LdvProfile.h"
37 #include "RaiseAsync.h"
47 #include "MarkStack.h"
52 #include <string.h> // for memset()
55 /* -----------------------------------------------------------------------------
57 -------------------------------------------------------------------------- */
59 /* STATIC OBJECT LIST.
62 * We maintain a linked list of static objects that are still live.
63 * The requirements for this list are:
65 * - we need to scan the list while adding to it, in order to
66 * scavenge all the static objects (in the same way that
67 * breadth-first scavenging works for dynamic objects).
69 * - we need to be able to tell whether an object is already on
70 * the list, to break loops.
72 * Each static object has a "static link field", which we use for
73 * linking objects on to the list. We use a stack-type list, consing
74 * objects on the front as they are added (this means that the
75 * scavenge phase is depth-first, not breadth-first, but that
78 * A separate list is kept for objects that have been scavenged
79 * already - this is so that we can zero all the marks afterwards.
81 * An object is on the list if its static link field is non-zero; this
82 * means that we have to mark the end of the list with '1', not NULL.
84 * Extra notes for generational GC:
86 * Each generation has a static object list associated with it. When
87 * collecting generations up to N, we treat the static object lists
88 * from generations > N as roots.
90 * We build up a static object list while collecting generations 0..N,
91 * which is then appended to the static object list of generation N+1.
94 /* N is the oldest generation being collected, where the generations
95 * are numbered starting at 0. A major GC (indicated by the major_gc
96 * flag) is when we're collecting all generations. We only attempt to
97 * deal with static objects and GC CAFs when doing a major GC.
102 /* Data used for allocation area sizing.
104 static lnat g0_pcnt_kept = 30; // percentage of g0 live at last minor GC
114 /* Thread-local data for each GC thread
116 gc_thread **gc_threads = NULL;
118 #if !defined(THREADED_RTS)
119 StgWord8 the_gc_thread[sizeof(gc_thread) + 64 * sizeof(gen_workspace)];
122 // Number of threads running in *this* GC. Affects how many
123 // step->todos[] lists we have to look in to find work.
127 long copied; // *words* copied & scavenged during this GC
129 rtsBool work_stealing;
133 /* -----------------------------------------------------------------------------
134 Static function declarations
135 -------------------------------------------------------------------------- */
137 static void mark_root (void *user, StgClosure **root);
138 static void zero_static_object_list (StgClosure* first_static);
139 static nat initialise_N (rtsBool force_major_gc);
140 static void init_collected_gen (nat g, nat threads);
141 static void init_uncollected_gen (nat g, nat threads);
142 static void init_gc_thread (gc_thread *t);
143 static void resize_generations (void);
144 static void resize_nursery (void);
145 static void start_gc_threads (void);
146 static void scavenge_until_all_done (void);
147 static StgWord inc_running (void);
148 static StgWord dec_running (void);
149 static void wakeup_gc_threads (nat n_threads, nat me);
150 static void shutdown_gc_threads (nat n_threads, nat me);
152 #if 0 && defined(DEBUG)
153 static void gcCAFs (void);
156 /* -----------------------------------------------------------------------------
158 -------------------------------------------------------------------------- */
160 bdescr *mark_stack_top_bd; // topmost block in the mark stack
161 bdescr *mark_stack_bd; // current block in the mark stack
162 StgPtr mark_sp; // pointer to the next unallocated mark stack entry
164 /* -----------------------------------------------------------------------------
165 GarbageCollect: the main entry point to the garbage collector.
167 Locks held: all capabilities are held throughout GarbageCollect().
168 -------------------------------------------------------------------------- */
171 GarbageCollect (rtsBool force_major_gc,
172 nat gc_type USED_IF_THREADS,
177 lnat live, allocated, max_copied, avg_copied, slop;
178 gc_thread *saved_gct;
181 // necessary if we stole a callee-saves register for gct:
185 CostCentreStack *prev_CCS;
190 #if defined(RTS_USER_SIGNALS)
191 if (RtsFlags.MiscFlags.install_signal_handlers) {
197 ASSERT(sizeof(gen_workspace) == 16 * sizeof(StgWord));
198 // otherwise adjust the padding in gen_workspace.
200 // tell the stats department that we've started a GC
203 // tell the STM to discard any cached closures it's hoping to re-use
206 // lock the StablePtr table
215 // attribute any costs to CCS_GC
221 /* Approximate how much we allocated.
222 * Todo: only when generating stats?
224 allocated = calcAllocated();
226 /* Figure out which generation to collect
228 n = initialise_N(force_major_gc);
230 #if defined(THREADED_RTS)
231 work_stealing = RtsFlags.ParFlags.parGcLoadBalancingEnabled &&
232 N >= RtsFlags.ParFlags.parGcLoadBalancingGen;
233 // It's not always a good idea to do load balancing in parallel
234 // GC. In particular, for a parallel program we don't want to
235 // lose locality by moving cached data into another CPU's cache
236 // (this effect can be quite significant).
238 // We could have a more complex way to deterimine whether to do
239 // work stealing or not, e.g. it might be a good idea to do it
240 // if the heap is big. For now, we just turn it on or off with
244 /* Start threads, so they can be spinning up while we finish initialisation.
248 #if defined(THREADED_RTS)
249 /* How many threads will be participating in this GC?
250 * We don't try to parallelise minor GCs (unless the user asks for
251 * it with +RTS -gn0), or mark/compact/sweep GC.
253 if (gc_type == PENDING_GC_PAR) {
254 n_gc_threads = RtsFlags.ParFlags.nNodes;
262 debugTrace(DEBUG_gc, "GC (gen %d): %d KB to collect, %ld MB in use, using %d thread(s)",
263 N, n * (BLOCK_SIZE / 1024), mblocks_allocated, n_gc_threads);
265 #ifdef RTS_GTK_FRONTPANEL
266 if (RtsFlags.GcFlags.frontpanel) {
267 updateFrontPanelBeforeGC(N);
272 // check for memory leaks if DEBUG is on
273 memInventory(DEBUG_gc);
276 // check sanity *before* GC
277 IF_DEBUG(sanity, checkSanity(rtsTrue));
279 // Initialise all our gc_thread structures
280 for (t = 0; t < n_gc_threads; t++) {
281 init_gc_thread(gc_threads[t]);
284 // Initialise all the generations/steps that we're collecting.
285 for (g = 0; g <= N; g++) {
286 init_collected_gen(g,n_gc_threads);
289 // Initialise all the generations/steps that we're *not* collecting.
290 for (g = N+1; g < RtsFlags.GcFlags.generations; g++) {
291 init_uncollected_gen(g,n_gc_threads);
294 /* Allocate a mark stack if we're doing a major collection.
296 if (major_gc && oldest_gen->mark) {
297 mark_stack_bd = allocBlock();
298 mark_stack_top_bd = mark_stack_bd;
299 mark_stack_bd->link = NULL;
300 mark_stack_bd->u.back = NULL;
301 mark_sp = mark_stack_bd->start;
303 mark_stack_bd = NULL;
304 mark_stack_top_bd = NULL;
308 // this is the main thread
310 if (n_gc_threads == 1) {
311 SET_GCT(gc_threads[0]);
313 SET_GCT(gc_threads[cap->no]);
316 SET_GCT(gc_threads[0]);
319 /* -----------------------------------------------------------------------
320 * follow all the roots that we know about:
323 // the main thread is running: this prevents any other threads from
324 // exiting prematurely, so we can start them now.
325 // NB. do this after the mutable lists have been saved above, otherwise
326 // the other GC threads will be writing into the old mutable lists.
328 wakeup_gc_threads(n_gc_threads, gct->thread_index);
330 // Mutable lists from each generation > N
331 // we want to *scavenge* these roots, not evacuate them: they're not
332 // going to move in this GC.
333 // Also do them in reverse generation order, for the usual reason:
334 // namely to reduce the likelihood of spurious old->new pointers.
336 for (g = RtsFlags.GcFlags.generations-1; g > N; g--) {
337 #if defined(THREADED_RTS)
338 if (n_gc_threads > 1) {
339 scavenge_mutable_list(generations[g].saved_mut_list, &generations[g]);
341 scavenge_mutable_list1(generations[g].saved_mut_list, &generations[g]);
344 scavenge_mutable_list(generations[g].saved_mut_list, &generations[g]);
346 freeChain_sync(generations[g].saved_mut_list);
347 generations[g].saved_mut_list = NULL;
351 // scavenge the capability-private mutable lists. This isn't part
352 // of markSomeCapabilities() because markSomeCapabilities() can only
353 // call back into the GC via mark_root() (due to the gct register
355 if (n_gc_threads == 1) {
356 for (n = 0; n < n_capabilities; n++) {
357 #if defined(THREADED_RTS)
358 scavenge_capability_mut_Lists1(&capabilities[n]);
360 scavenge_capability_mut_lists(&capabilities[n]);
364 scavenge_capability_mut_lists(&capabilities[gct->thread_index]);
367 // follow roots from the CAF list (used by GHCi)
369 markCAFs(mark_root, gct);
371 // follow all the roots that the application knows about.
373 markSomeCapabilities(mark_root, gct, gct->thread_index, n_gc_threads,
374 rtsTrue/*prune sparks*/);
376 #if defined(RTS_USER_SIGNALS)
377 // mark the signal handlers (signals should be already blocked)
378 markSignalHandlers(mark_root, gct);
381 // Mark the weak pointer list, and prepare to detect dead weak pointers.
385 // Mark the stable pointer table.
386 markStablePtrTable(mark_root, gct);
388 /* -------------------------------------------------------------------------
389 * Repeatedly scavenge all the areas we know about until there's no
390 * more scavenging to be done.
394 scavenge_until_all_done();
395 // The other threads are now stopped. We might recurse back to
396 // here, but from now on this is the only thread.
398 // if any blackholes are alive, make the threads that wait on
400 if (traverseBlackholeQueue()) {
405 // must be last... invariant is that everything is fully
406 // scavenged at this point.
407 if (traverseWeakPtrList()) { // returns rtsTrue if evaced something
412 // If we get to here, there's really nothing left to do.
416 shutdown_gc_threads(n_gc_threads, gct->thread_index);
418 // Now see which stable names are still alive.
422 // We call processHeapClosureForDead() on every closure destroyed during
423 // the current garbage collection, so we invoke LdvCensusForDead().
424 if (RtsFlags.ProfFlags.doHeapProfile == HEAP_BY_LDV
425 || RtsFlags.ProfFlags.bioSelector != NULL)
429 // NO MORE EVACUATION AFTER THIS POINT!
431 // Two-space collector: free the old to-space.
432 // g0->old_blocks is the old nursery
433 // g0->blocks is to-space from the previous GC
434 if (RtsFlags.GcFlags.generations == 1) {
435 if (g0->blocks != NULL) {
436 freeChain(g0->blocks);
441 // For each workspace, in each thread, move the copied blocks to the step
447 for (t = 0; t < n_gc_threads; t++) {
450 for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
453 // Push the final block
455 push_scanned_block(ws->todo_bd, ws);
458 ASSERT(gct->scan_bd == NULL);
459 ASSERT(countBlocks(ws->scavd_list) == ws->n_scavd_blocks);
462 for (bd = ws->scavd_list; bd != NULL; bd = bd->link) {
463 ws->gen->n_words += bd->free - bd->start;
467 prev->link = ws->gen->blocks;
468 ws->gen->blocks = ws->scavd_list;
470 ws->gen->n_blocks += ws->n_scavd_blocks;
474 // Add all the partial blocks *after* we've added all the full
475 // blocks. This is so that we can grab the partial blocks back
476 // again and try to fill them up in the next GC.
477 for (t = 0; t < n_gc_threads; t++) {
480 for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
484 for (bd = ws->part_list; bd != NULL; bd = next) {
486 if (bd->free == bd->start) {
488 ws->part_list = next;
495 ws->gen->n_words += bd->free - bd->start;
500 prev->link = ws->gen->blocks;
501 ws->gen->blocks = ws->part_list;
503 ws->gen->n_blocks += ws->n_part_blocks;
505 ASSERT(countBlocks(ws->gen->blocks) == ws->gen->n_blocks);
506 ASSERT(countOccupied(ws->gen->blocks) == ws->gen->n_words);
511 // Finally: compact or sweep the oldest generation.
512 if (major_gc && oldest_gen->mark) {
513 if (oldest_gen->compact)
514 compact(gct->scavenged_static_objects);
519 /* run through all the generations/steps and tidy up
526 for (i=0; i < n_gc_threads; i++) {
527 if (n_gc_threads > 1) {
528 debugTrace(DEBUG_gc,"thread %d:", i);
529 debugTrace(DEBUG_gc," copied %ld", gc_threads[i]->copied * sizeof(W_));
530 debugTrace(DEBUG_gc," scanned %ld", gc_threads[i]->scanned * sizeof(W_));
531 debugTrace(DEBUG_gc," any_work %ld", gc_threads[i]->any_work);
532 debugTrace(DEBUG_gc," no_work %ld", gc_threads[i]->no_work);
533 debugTrace(DEBUG_gc," scav_find_work %ld", gc_threads[i]->scav_find_work);
535 copied += gc_threads[i]->copied;
536 max_copied = stg_max(gc_threads[i]->copied, max_copied);
538 if (n_gc_threads == 1) {
546 for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
549 generations[g].collections++; // for stats
550 if (n_gc_threads > 1) generations[g].par_collections++;
553 // Count the mutable list as bytes "copied" for the purposes of
554 // stats. Every mutable list is copied during every GC.
556 nat mut_list_size = 0;
557 for (bd = generations[g].mut_list; bd != NULL; bd = bd->link) {
558 mut_list_size += bd->free - bd->start;
560 for (n = 0; n < n_capabilities; n++) {
561 for (bd = capabilities[n].mut_lists[g];
562 bd != NULL; bd = bd->link) {
563 mut_list_size += bd->free - bd->start;
566 copied += mut_list_size;
569 "mut_list_size: %lu (%d vars, %d arrays, %d MVARs, %d others)",
570 (unsigned long)(mut_list_size * sizeof(W_)),
571 mutlist_MUTVARS, mutlist_MUTARRS, mutlist_MVARS, mutlist_OTHERS);
575 gen = &generations[g];
577 // for generations we collected...
580 /* free old memory and shift to-space into from-space for all
581 * the collected steps (except the allocation area). These
582 * freed blocks will probaby be quickly recycled.
586 // tack the new blocks on the end of the existing blocks
587 if (gen->old_blocks != NULL) {
590 for (bd = gen->old_blocks; bd != NULL; bd = next) {
594 if (!(bd->flags & BF_MARKED))
597 gen->old_blocks = next;
606 gen->n_words += bd->free - bd->start;
608 // NB. this step might not be compacted next
609 // time, so reset the BF_MARKED flags.
610 // They are set before GC if we're going to
611 // compact. (search for BF_MARKED above).
612 bd->flags &= ~BF_MARKED;
614 // between GCs, all blocks in the heap except
615 // for the nursery have the BF_EVACUATED flag set.
616 bd->flags |= BF_EVACUATED;
623 prev->link = gen->blocks;
624 gen->blocks = gen->old_blocks;
627 // add the new blocks to the block tally
628 gen->n_blocks += gen->n_old_blocks;
629 ASSERT(countBlocks(gen->blocks) == gen->n_blocks);
630 ASSERT(countOccupied(gen->blocks) == gen->n_words);
634 freeChain(gen->old_blocks);
637 gen->old_blocks = NULL;
638 gen->n_old_blocks = 0;
640 /* LARGE OBJECTS. The current live large objects are chained on
641 * scavenged_large, having been moved during garbage
642 * collection from large_objects. Any objects left on the
643 * large_objects list are therefore dead, so we free them here.
645 freeChain(gen->large_objects);
646 gen->large_objects = gen->scavenged_large_objects;
647 gen->n_large_blocks = gen->n_scavenged_large_blocks;
648 gen->n_new_large_blocks = 0;
649 ASSERT(countBlocks(gen->large_objects) == gen->n_large_blocks);
651 else // for generations > N
653 /* For older generations, we need to append the
654 * scavenged_large_object list (i.e. large objects that have been
655 * promoted during this GC) to the large_object list for that step.
657 for (bd = gen->scavenged_large_objects; bd; bd = next) {
659 dbl_link_onto(bd, &gen->large_objects);
662 // add the new blocks we promoted during this GC
663 gen->n_large_blocks += gen->n_scavenged_large_blocks;
664 ASSERT(countBlocks(gen->large_objects) == gen->n_large_blocks);
666 } // for all generations
668 // update the max size of older generations after a major GC
669 resize_generations();
671 // Calculate the amount of live data for stats.
672 live = calcLiveWords();
674 // Free the small objects allocated via allocate(), since this will
675 // all have been copied into G0S1 now.
676 alloc_blocks_lim = RtsFlags.GcFlags.minAllocAreaSize;
678 // Start a new pinned_object_block
679 for (n = 0; n < n_capabilities; n++) {
680 capabilities[n].pinned_object_block = NULL;
683 // Free the mark stack.
684 if (mark_stack_top_bd != NULL) {
685 debugTrace(DEBUG_gc, "mark stack: %d blocks",
686 countBlocks(mark_stack_top_bd));
687 freeChain(mark_stack_top_bd);
691 for (g = 0; g <= N; g++) {
692 gen = &generations[g];
693 if (gen->bitmap != NULL) {
694 freeGroup(gen->bitmap);
701 // mark the garbage collected CAFs as dead
702 #if 0 && defined(DEBUG) // doesn't work at the moment
703 if (major_gc) { gcCAFs(); }
707 // resetStaticObjectForRetainerProfiling() must be called before
709 if (n_gc_threads > 1) {
710 barf("profiling is currently broken with multi-threaded GC");
711 // ToDo: fix the gct->scavenged_static_objects below
713 resetStaticObjectForRetainerProfiling(gct->scavenged_static_objects);
716 // zero the scavenged static object list
719 for (i = 0; i < n_gc_threads; i++) {
720 zero_static_object_list(gc_threads[i]->scavenged_static_objects);
727 // start any pending finalizers
729 scheduleFinalizers(cap, old_weak_ptr_list);
732 // send exceptions to any threads which were about to die
734 resurrectThreads(resurrected_threads);
735 performPendingThrowTos(exception_threads);
738 // Update the stable pointer hash table.
739 updateStablePtrTable(major_gc);
741 // check sanity after GC
742 IF_DEBUG(sanity, checkSanity(rtsTrue));
744 // extra GC trace info
745 IF_DEBUG(gc, statDescribeGens());
748 // symbol-table based profiling
749 /* heapCensus(to_blocks); */ /* ToDo */
752 // restore enclosing cost centre
758 // check for memory leaks if DEBUG is on
759 memInventory(DEBUG_gc);
762 #ifdef RTS_GTK_FRONTPANEL
763 if (RtsFlags.GcFlags.frontpanel) {
764 updateFrontPanelAfterGC( N, live );
768 // ok, GC over: tell the stats department what happened.
769 slop = calcLiveBlocks() * BLOCK_SIZE_W - live;
770 stat_endGC(allocated, live, copied, N, max_copied, avg_copied, slop);
772 // unlock the StablePtr table
775 // Guess which generation we'll collect *next* time
776 initialise_N(force_major_gc);
778 #if defined(RTS_USER_SIGNALS)
779 if (RtsFlags.MiscFlags.install_signal_handlers) {
780 // unblock signals again
781 unblockUserSignals();
790 /* -----------------------------------------------------------------------------
791 Figure out which generation to collect, initialise N and major_gc.
793 Also returns the total number of blocks in generations that will be
795 -------------------------------------------------------------------------- */
798 initialise_N (rtsBool force_major_gc)
801 nat blocks, blocks_total;
806 if (force_major_gc) {
807 N = RtsFlags.GcFlags.generations - 1;
812 for (g = RtsFlags.GcFlags.generations - 1; g >= 0; g--) {
814 blocks = generations[g].n_words / BLOCK_SIZE_W
815 + generations[g].n_large_blocks;
817 if (blocks >= generations[g].max_blocks) {
821 blocks_total += blocks;
825 blocks_total += countNurseryBlocks();
827 major_gc = (N == RtsFlags.GcFlags.generations-1);
831 /* -----------------------------------------------------------------------------
832 Initialise the gc_thread structures.
833 -------------------------------------------------------------------------- */
835 #define GC_THREAD_INACTIVE 0
836 #define GC_THREAD_STANDING_BY 1
837 #define GC_THREAD_RUNNING 2
838 #define GC_THREAD_WAITING_TO_CONTINUE 3
841 new_gc_thread (nat n, gc_thread *t)
848 initSpinLock(&t->gc_spin);
849 initSpinLock(&t->mut_spin);
850 ACQUIRE_SPIN_LOCK(&t->gc_spin);
851 t->wakeup = GC_THREAD_INACTIVE; // starts true, so we can wait for the
852 // thread to start up, see wakeup_gc_threads
856 t->free_blocks = NULL;
865 for (g = 0; g < RtsFlags.GcFlags.generations; g++)
868 ws->gen = &generations[g];
869 ASSERT(g == ws->gen->no);
873 ws->todo_q = newWSDeque(128);
874 ws->todo_overflow = NULL;
875 ws->n_todo_overflow = 0;
877 ws->part_list = NULL;
878 ws->n_part_blocks = 0;
880 ws->scavd_list = NULL;
881 ws->n_scavd_blocks = 0;
889 if (gc_threads == NULL) {
890 #if defined(THREADED_RTS)
892 gc_threads = stgMallocBytes (RtsFlags.ParFlags.nNodes *
896 for (i = 0; i < RtsFlags.ParFlags.nNodes; i++) {
898 stgMallocBytes(sizeof(gc_thread) +
899 RtsFlags.GcFlags.generations * sizeof(gen_workspace),
902 new_gc_thread(i, gc_threads[i]);
905 gc_threads = stgMallocBytes (sizeof(gc_thread*),"alloc_gc_threads");
907 new_gc_thread(0,gc_threads[0]);
916 if (gc_threads != NULL) {
917 #if defined(THREADED_RTS)
919 for (i = 0; i < n_capabilities; i++) {
920 for (g = 0; g < RtsFlags.GcFlags.generations; g++)
922 freeWSDeque(gc_threads[i]->gens[g].todo_q);
924 stgFree (gc_threads[i]);
926 stgFree (gc_threads);
928 for (g = 0; g < RtsFlags.GcFlags.generations; g++)
930 freeWSDeque(gc_threads[0]->gens[g].todo_q);
932 stgFree (gc_threads);
938 /* ----------------------------------------------------------------------------
940 ------------------------------------------------------------------------- */
942 static volatile StgWord gc_running_threads;
948 new = atomic_inc(&gc_running_threads);
949 ASSERT(new <= n_gc_threads);
956 ASSERT(gc_running_threads != 0);
957 return atomic_dec(&gc_running_threads);
970 // scavenge objects in compacted generation
971 if (mark_stack_bd != NULL && !mark_stack_empty()) {
975 // Check for global work in any step. We don't need to check for
976 // local work, because we have already exited scavenge_loop(),
977 // which means there is no local work for this thread.
978 for (g = 0; g < (int)RtsFlags.GcFlags.generations; g++) {
980 if (ws->todo_large_objects) return rtsTrue;
981 if (!looksEmptyWSDeque(ws->todo_q)) return rtsTrue;
982 if (ws->todo_overflow) return rtsTrue;
985 #if defined(THREADED_RTS)
988 // look for work to steal
989 for (n = 0; n < n_gc_threads; n++) {
990 if (n == gct->thread_index) continue;
991 for (g = RtsFlags.GcFlags.generations-1; g >= 0; g--) {
992 ws = &gc_threads[n]->gens[g];
993 if (!looksEmptyWSDeque(ws->todo_q)) return rtsTrue;
1000 #if defined(THREADED_RTS)
1008 scavenge_until_all_done (void)
1014 traceEventGcWork(&capabilities[gct->thread_index]);
1016 #if defined(THREADED_RTS)
1017 if (n_gc_threads > 1) {
1026 // scavenge_loop() only exits when there's no work to do
1029 traceEventGcIdle(&capabilities[gct->thread_index]);
1031 debugTrace(DEBUG_gc, "%d GC threads still running", r);
1033 while (gc_running_threads != 0) {
1039 // any_work() does not remove the work from the queue, it
1040 // just checks for the presence of work. If we find any,
1041 // then we increment gc_running_threads and go back to
1042 // scavenge_loop() to perform any pending work.
1045 traceEventGcDone(&capabilities[gct->thread_index]);
1048 #if defined(THREADED_RTS)
1051 gcWorkerThread (Capability *cap)
1053 gc_thread *saved_gct;
1055 // necessary if we stole a callee-saves register for gct:
1058 gct = gc_threads[cap->no];
1059 gct->id = osThreadId();
1061 // Wait until we're told to wake up
1062 RELEASE_SPIN_LOCK(&gct->mut_spin);
1063 gct->wakeup = GC_THREAD_STANDING_BY;
1064 debugTrace(DEBUG_gc, "GC thread %d standing by...", gct->thread_index);
1065 ACQUIRE_SPIN_LOCK(&gct->gc_spin);
1068 // start performance counters in this thread...
1069 if (gct->papi_events == -1) {
1070 papi_init_eventset(&gct->papi_events);
1072 papi_thread_start_gc1_count(gct->papi_events);
1075 // Every thread evacuates some roots.
1077 markSomeCapabilities(mark_root, gct, gct->thread_index, n_gc_threads,
1078 rtsTrue/*prune sparks*/);
1079 scavenge_capability_mut_lists(&capabilities[gct->thread_index]);
1081 scavenge_until_all_done();
1084 // count events in this thread towards the GC totals
1085 papi_thread_stop_gc1_count(gct->papi_events);
1088 // Wait until we're told to continue
1089 RELEASE_SPIN_LOCK(&gct->gc_spin);
1090 gct->wakeup = GC_THREAD_WAITING_TO_CONTINUE;
1091 debugTrace(DEBUG_gc, "GC thread %d waiting to continue...",
1093 ACQUIRE_SPIN_LOCK(&gct->mut_spin);
1094 debugTrace(DEBUG_gc, "GC thread %d on my way...", gct->thread_index);
1101 #if defined(THREADED_RTS)
1104 waitForGcThreads (Capability *cap USED_IF_THREADS)
1106 nat n_threads = RtsFlags.ParFlags.nNodes;
1109 rtsBool retry = rtsTrue;
1112 for (i=0; i < n_threads; i++) {
1113 if (i == me) continue;
1114 if (gc_threads[i]->wakeup != GC_THREAD_STANDING_BY) {
1115 prodCapability(&capabilities[i], cap->running_task);
1118 for (j=0; j < 10; j++) {
1120 for (i=0; i < n_threads; i++) {
1121 if (i == me) continue;
1123 setContextSwitches();
1124 if (gc_threads[i]->wakeup != GC_THREAD_STANDING_BY) {
1134 #endif // THREADED_RTS
1137 start_gc_threads (void)
1139 #if defined(THREADED_RTS)
1140 gc_running_threads = 0;
1145 wakeup_gc_threads (nat n_threads USED_IF_THREADS, nat me USED_IF_THREADS)
1147 #if defined(THREADED_RTS)
1149 for (i=0; i < n_threads; i++) {
1150 if (i == me) continue;
1152 debugTrace(DEBUG_gc, "waking up gc thread %d", i);
1153 if (gc_threads[i]->wakeup != GC_THREAD_STANDING_BY) barf("wakeup_gc_threads");
1155 gc_threads[i]->wakeup = GC_THREAD_RUNNING;
1156 ACQUIRE_SPIN_LOCK(&gc_threads[i]->mut_spin);
1157 RELEASE_SPIN_LOCK(&gc_threads[i]->gc_spin);
1162 // After GC is complete, we must wait for all GC threads to enter the
1163 // standby state, otherwise they may still be executing inside
1164 // any_work(), and may even remain awake until the next GC starts.
1166 shutdown_gc_threads (nat n_threads USED_IF_THREADS, nat me USED_IF_THREADS)
1168 #if defined(THREADED_RTS)
1170 for (i=0; i < n_threads; i++) {
1171 if (i == me) continue;
1172 while (gc_threads[i]->wakeup != GC_THREAD_WAITING_TO_CONTINUE) { write_barrier(); }
1177 #if defined(THREADED_RTS)
1179 releaseGCThreads (Capability *cap USED_IF_THREADS)
1181 nat n_threads = RtsFlags.ParFlags.nNodes;
1184 for (i=0; i < n_threads; i++) {
1185 if (i == me) continue;
1186 if (gc_threads[i]->wakeup != GC_THREAD_WAITING_TO_CONTINUE)
1187 barf("releaseGCThreads");
1189 gc_threads[i]->wakeup = GC_THREAD_INACTIVE;
1190 ACQUIRE_SPIN_LOCK(&gc_threads[i]->gc_spin);
1191 RELEASE_SPIN_LOCK(&gc_threads[i]->mut_spin);
1196 /* ----------------------------------------------------------------------------
1197 Initialise a generation that is to be collected
1198 ------------------------------------------------------------------------- */
1201 init_collected_gen (nat g, nat n_threads)
1208 // Throw away the current mutable list. Invariant: the mutable
1209 // list always has at least one block; this means we can avoid a
1210 // check for NULL in recordMutable().
1212 freeChain(generations[g].mut_list);
1213 generations[g].mut_list = allocBlock();
1214 for (i = 0; i < n_capabilities; i++) {
1215 freeChain(capabilities[i].mut_lists[g]);
1216 capabilities[i].mut_lists[g] = allocBlock();
1220 gen = &generations[g];
1221 ASSERT(gen->no == g);
1223 // we'll construct a new list of threads in this step
1224 // during GC, throw away the current list.
1225 gen->old_threads = gen->threads;
1226 gen->threads = END_TSO_QUEUE;
1228 // deprecate the existing blocks
1229 gen->old_blocks = gen->blocks;
1230 gen->n_old_blocks = gen->n_blocks;
1234 gen->live_estimate = 0;
1236 // initialise the large object queues.
1237 gen->scavenged_large_objects = NULL;
1238 gen->n_scavenged_large_blocks = 0;
1240 // mark the small objects as from-space
1241 for (bd = gen->old_blocks; bd; bd = bd->link) {
1242 bd->flags &= ~BF_EVACUATED;
1245 // mark the large objects as from-space
1246 for (bd = gen->large_objects; bd; bd = bd->link) {
1247 bd->flags &= ~BF_EVACUATED;
1250 // for a compacted generation, we need to allocate the bitmap
1252 nat bitmap_size; // in bytes
1253 bdescr *bitmap_bdescr;
1256 bitmap_size = gen->n_old_blocks * BLOCK_SIZE / (sizeof(W_)*BITS_PER_BYTE);
1258 if (bitmap_size > 0) {
1259 bitmap_bdescr = allocGroup((lnat)BLOCK_ROUND_UP(bitmap_size)
1261 gen->bitmap = bitmap_bdescr;
1262 bitmap = bitmap_bdescr->start;
1264 debugTrace(DEBUG_gc, "bitmap_size: %d, bitmap: %p",
1265 bitmap_size, bitmap);
1267 // don't forget to fill it with zeros!
1268 memset(bitmap, 0, bitmap_size);
1270 // For each block in this step, point to its bitmap from the
1271 // block descriptor.
1272 for (bd=gen->old_blocks; bd != NULL; bd = bd->link) {
1273 bd->u.bitmap = bitmap;
1274 bitmap += BLOCK_SIZE_W / (sizeof(W_)*BITS_PER_BYTE);
1276 // Also at this point we set the BF_MARKED flag
1277 // for this block. The invariant is that
1278 // BF_MARKED is always unset, except during GC
1279 // when it is set on those blocks which will be
1281 if (!(bd->flags & BF_FRAGMENTED)) {
1282 bd->flags |= BF_MARKED;
1288 // For each GC thread, for each step, allocate a "todo" block to
1289 // store evacuated objects to be scavenged, and a block to store
1290 // evacuated objects that do not need to be scavenged.
1291 for (t = 0; t < n_threads; t++) {
1292 ws = &gc_threads[t]->gens[g];
1294 ws->todo_large_objects = NULL;
1296 ws->part_list = NULL;
1297 ws->n_part_blocks = 0;
1299 // allocate the first to-space block; extra blocks will be
1300 // chained on as necessary.
1302 ASSERT(looksEmptyWSDeque(ws->todo_q));
1303 alloc_todo_block(ws,0);
1305 ws->todo_overflow = NULL;
1306 ws->n_todo_overflow = 0;
1308 ws->scavd_list = NULL;
1309 ws->n_scavd_blocks = 0;
1314 /* ----------------------------------------------------------------------------
1315 Initialise a generation that is *not* to be collected
1316 ------------------------------------------------------------------------- */
1319 init_uncollected_gen (nat g, nat threads)
1326 // save the current mutable lists for this generation, and
1327 // allocate a fresh block for each one. We'll traverse these
1328 // mutable lists as roots early on in the GC.
1329 generations[g].saved_mut_list = generations[g].mut_list;
1330 generations[g].mut_list = allocBlock();
1331 for (n = 0; n < n_capabilities; n++) {
1332 capabilities[n].saved_mut_lists[g] = capabilities[n].mut_lists[g];
1333 capabilities[n].mut_lists[g] = allocBlock();
1336 gen = &generations[g];
1338 gen->scavenged_large_objects = NULL;
1339 gen->n_scavenged_large_blocks = 0;
1341 for (t = 0; t < threads; t++) {
1342 ws = &gc_threads[t]->gens[g];
1344 ASSERT(looksEmptyWSDeque(ws->todo_q));
1345 ws->todo_large_objects = NULL;
1347 ws->part_list = NULL;
1348 ws->n_part_blocks = 0;
1350 ws->scavd_list = NULL;
1351 ws->n_scavd_blocks = 0;
1353 // If the block at the head of the list in this generation
1354 // is less than 3/4 full, then use it as a todo block.
1355 if (gen->blocks && isPartiallyFull(gen->blocks))
1357 ws->todo_bd = gen->blocks;
1358 ws->todo_free = ws->todo_bd->free;
1359 ws->todo_lim = ws->todo_bd->start + BLOCK_SIZE_W;
1360 gen->blocks = gen->blocks->link;
1362 gen->n_words -= ws->todo_bd->free - ws->todo_bd->start;
1363 ws->todo_bd->link = NULL;
1364 // we must scan from the current end point.
1365 ws->todo_bd->u.scan = ws->todo_bd->free;
1370 alloc_todo_block(ws,0);
1374 // deal out any more partial blocks to the threads' part_lists
1376 while (gen->blocks && isPartiallyFull(gen->blocks))
1379 gen->blocks = bd->link;
1380 ws = &gc_threads[t]->gens[g];
1381 bd->link = ws->part_list;
1383 ws->n_part_blocks += 1;
1384 bd->u.scan = bd->free;
1386 gen->n_words -= bd->free - bd->start;
1388 if (t == n_gc_threads) t = 0;
1392 /* -----------------------------------------------------------------------------
1393 Initialise a gc_thread before GC
1394 -------------------------------------------------------------------------- */
1397 init_gc_thread (gc_thread *t)
1399 t->static_objects = END_OF_STATIC_LIST;
1400 t->scavenged_static_objects = END_OF_STATIC_LIST;
1402 t->mut_lists = capabilities[t->thread_index].mut_lists;
1404 t->failed_to_evac = rtsFalse;
1405 t->eager_promotion = rtsTrue;
1406 t->thunk_selector_depth = 0;
1411 t->scav_find_work = 0;
1414 /* -----------------------------------------------------------------------------
1415 Function we pass to evacuate roots.
1416 -------------------------------------------------------------------------- */
1419 mark_root(void *user USED_IF_THREADS, StgClosure **root)
1421 // we stole a register for gct, but this function is called from
1422 // *outside* the GC where the register variable is not in effect,
1423 // so we need to save and restore it here. NB. only call
1424 // mark_root() from the main GC thread, otherwise gct will be
1426 gc_thread *saved_gct;
1435 /* -----------------------------------------------------------------------------
1436 Initialising the static object & mutable lists
1437 -------------------------------------------------------------------------- */
1440 zero_static_object_list(StgClosure* first_static)
1444 const StgInfoTable *info;
1446 for (p = first_static; p != END_OF_STATIC_LIST; p = link) {
1448 link = *STATIC_LINK(info, p);
1449 *STATIC_LINK(info,p) = NULL;
1453 /* ----------------------------------------------------------------------------
1454 Reset the sizes of the older generations when we do a major
1457 CURRENT STRATEGY: make all generations except zero the same size.
1458 We have to stay within the maximum heap size, and leave a certain
1459 percentage of the maximum heap size available to allocate into.
1460 ------------------------------------------------------------------------- */
1463 resize_generations (void)
1467 if (major_gc && RtsFlags.GcFlags.generations > 1) {
1468 nat live, size, min_alloc, words;
1469 nat max = RtsFlags.GcFlags.maxHeapSize;
1470 nat gens = RtsFlags.GcFlags.generations;
1472 // live in the oldest generations
1473 if (oldest_gen->live_estimate != 0) {
1474 words = oldest_gen->live_estimate;
1476 words = oldest_gen->n_words;
1478 live = (words + BLOCK_SIZE_W - 1) / BLOCK_SIZE_W +
1479 oldest_gen->n_large_blocks;
1481 // default max size for all generations except zero
1482 size = stg_max(live * RtsFlags.GcFlags.oldGenFactor,
1483 RtsFlags.GcFlags.minOldGenSize);
1485 if (RtsFlags.GcFlags.heapSizeSuggestionAuto) {
1486 RtsFlags.GcFlags.heapSizeSuggestion = size;
1489 // minimum size for generation zero
1490 min_alloc = stg_max((RtsFlags.GcFlags.pcFreeHeap * max) / 200,
1491 RtsFlags.GcFlags.minAllocAreaSize);
1493 // Auto-enable compaction when the residency reaches a
1494 // certain percentage of the maximum heap size (default: 30%).
1495 if (RtsFlags.GcFlags.generations > 1 &&
1496 (RtsFlags.GcFlags.compact ||
1498 oldest_gen->n_blocks >
1499 (RtsFlags.GcFlags.compactThreshold * max) / 100))) {
1500 oldest_gen->mark = 1;
1501 oldest_gen->compact = 1;
1502 // debugBelch("compaction: on\n", live);
1504 oldest_gen->mark = 0;
1505 oldest_gen->compact = 0;
1506 // debugBelch("compaction: off\n", live);
1509 if (RtsFlags.GcFlags.sweep) {
1510 oldest_gen->mark = 1;
1513 // if we're going to go over the maximum heap size, reduce the
1514 // size of the generations accordingly. The calculation is
1515 // different if compaction is turned on, because we don't need
1516 // to double the space required to collect the old generation.
1519 // this test is necessary to ensure that the calculations
1520 // below don't have any negative results - we're working
1521 // with unsigned values here.
1522 if (max < min_alloc) {
1526 if (oldest_gen->compact) {
1527 if ( (size + (size - 1) * (gens - 2) * 2) + min_alloc > max ) {
1528 size = (max - min_alloc) / ((gens - 1) * 2 - 1);
1531 if ( (size * (gens - 1) * 2) + min_alloc > max ) {
1532 size = (max - min_alloc) / ((gens - 1) * 2);
1542 debugBelch("live: %d, min_alloc: %d, size : %d, max = %d\n", live,
1543 min_alloc, size, max);
1546 for (g = 0; g < gens; g++) {
1547 generations[g].max_blocks = size;
1552 /* -----------------------------------------------------------------------------
1553 Calculate the new size of the nursery, and resize it.
1554 -------------------------------------------------------------------------- */
1557 resize_nursery (void)
1559 lnat min_nursery = RtsFlags.GcFlags.minAllocAreaSize * n_capabilities;
1561 if (RtsFlags.GcFlags.generations == 1)
1562 { // Two-space collector:
1565 /* set up a new nursery. Allocate a nursery size based on a
1566 * function of the amount of live data (by default a factor of 2)
1567 * Use the blocks from the old nursery if possible, freeing up any
1570 * If we get near the maximum heap size, then adjust our nursery
1571 * size accordingly. If the nursery is the same size as the live
1572 * data (L), then we need 3L bytes. We can reduce the size of the
1573 * nursery to bring the required memory down near 2L bytes.
1575 * A normal 2-space collector would need 4L bytes to give the same
1576 * performance we get from 3L bytes, reducing to the same
1577 * performance at 2L bytes.
1579 blocks = generations[0].n_blocks;
1581 if ( RtsFlags.GcFlags.maxHeapSize != 0 &&
1582 blocks * RtsFlags.GcFlags.oldGenFactor * 2 >
1583 RtsFlags.GcFlags.maxHeapSize )
1585 long adjusted_blocks; // signed on purpose
1588 adjusted_blocks = (RtsFlags.GcFlags.maxHeapSize - 2 * blocks);
1590 debugTrace(DEBUG_gc, "near maximum heap size of 0x%x blocks, blocks = %d, adjusted to %ld",
1591 RtsFlags.GcFlags.maxHeapSize, blocks, adjusted_blocks);
1593 pc_free = adjusted_blocks * 100 / RtsFlags.GcFlags.maxHeapSize;
1594 if (pc_free < RtsFlags.GcFlags.pcFreeHeap) /* might even * be < 0 */
1598 blocks = adjusted_blocks;
1602 blocks *= RtsFlags.GcFlags.oldGenFactor;
1603 if (blocks < min_nursery)
1605 blocks = min_nursery;
1608 resizeNurseries(blocks);
1610 else // Generational collector
1613 * If the user has given us a suggested heap size, adjust our
1614 * allocation area to make best use of the memory available.
1616 if (RtsFlags.GcFlags.heapSizeSuggestion)
1619 nat needed = calcNeeded(); // approx blocks needed at next GC
1621 /* Guess how much will be live in generation 0 step 0 next time.
1622 * A good approximation is obtained by finding the
1623 * percentage of g0 that was live at the last minor GC.
1625 * We have an accurate figure for the amount of copied data in
1626 * 'copied', but we must convert this to a number of blocks, with
1627 * a small adjustment for estimated slop at the end of a block
1632 g0_pcnt_kept = ((copied / (BLOCK_SIZE_W - 10)) * 100)
1633 / countNurseryBlocks();
1636 /* Estimate a size for the allocation area based on the
1637 * information available. We might end up going slightly under
1638 * or over the suggested heap size, but we should be pretty
1641 * Formula: suggested - needed
1642 * ----------------------------
1643 * 1 + g0_pcnt_kept/100
1645 * where 'needed' is the amount of memory needed at the next
1646 * collection for collecting all gens except g0.
1649 (((long)RtsFlags.GcFlags.heapSizeSuggestion - (long)needed) * 100) /
1650 (100 + (long)g0_pcnt_kept);
1652 if (blocks < (long)min_nursery) {
1653 blocks = min_nursery;
1656 resizeNurseries((nat)blocks);
1660 // we might have added extra large blocks to the nursery, so
1661 // resize back to minAllocAreaSize again.
1662 resizeNurseriesFixed(RtsFlags.GcFlags.minAllocAreaSize);
1667 /* -----------------------------------------------------------------------------
1668 Sanity code for CAF garbage collection.
1670 With DEBUG turned on, we manage a CAF list in addition to the SRT
1671 mechanism. After GC, we run down the CAF list and blackhole any
1672 CAFs which have been garbage collected. This means we get an error
1673 whenever the program tries to enter a garbage collected CAF.
1675 Any garbage collected CAFs are taken off the CAF list at the same
1677 -------------------------------------------------------------------------- */
1679 #if 0 && defined(DEBUG)
1686 const StgInfoTable *info;
1697 ASSERT(info->type == IND_STATIC);
1699 if (STATIC_LINK(info,p) == NULL) {
1700 debugTrace(DEBUG_gccafs, "CAF gc'd at 0x%04lx", (long)p);
1702 SET_INFO(p,&stg_BLACKHOLE_info);
1703 p = STATIC_LINK2(info,p);
1707 pp = &STATIC_LINK2(info,p);
1714 debugTrace(DEBUG_gccafs, "%d CAFs live", i);