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 update_task_list (void);
144 static void resize_generations (void);
145 static void resize_nursery (void);
146 static void start_gc_threads (void);
147 static void scavenge_until_all_done (void);
148 static StgWord inc_running (void);
149 static StgWord dec_running (void);
150 static void wakeup_gc_threads (nat n_threads, nat me);
151 static void shutdown_gc_threads (nat n_threads, nat me);
153 #if 0 && defined(DEBUG)
154 static void gcCAFs (void);
157 /* -----------------------------------------------------------------------------
159 -------------------------------------------------------------------------- */
161 bdescr *mark_stack_top_bd; // topmost block in the mark stack
162 bdescr *mark_stack_bd; // current block in the mark stack
163 StgPtr mark_sp; // pointer to the next unallocated mark stack entry
165 /* -----------------------------------------------------------------------------
166 GarbageCollect: the main entry point to the garbage collector.
168 Locks held: all capabilities are held throughout GarbageCollect().
169 -------------------------------------------------------------------------- */
172 GarbageCollect (rtsBool force_major_gc,
173 nat gc_type USED_IF_THREADS,
178 lnat live, allocated, max_copied, avg_copied, slop;
179 gc_thread *saved_gct;
182 // necessary if we stole a callee-saves register for gct:
186 CostCentreStack *prev_CCS;
191 #if defined(RTS_USER_SIGNALS)
192 if (RtsFlags.MiscFlags.install_signal_handlers) {
198 ASSERT(sizeof(gen_workspace) == 16 * sizeof(StgWord));
199 // otherwise adjust the padding in gen_workspace.
201 // tell the stats department that we've started a GC
204 // tell the STM to discard any cached closures it's hoping to re-use
207 // lock the StablePtr table
216 // attribute any costs to CCS_GC
222 /* Approximate how much we allocated.
223 * Todo: only when generating stats?
225 allocated = calcAllocated();
227 /* Figure out which generation to collect
229 n = initialise_N(force_major_gc);
231 #if defined(THREADED_RTS)
232 work_stealing = RtsFlags.ParFlags.parGcLoadBalancingEnabled &&
233 N >= RtsFlags.ParFlags.parGcLoadBalancingGen;
234 // It's not always a good idea to do load balancing in parallel
235 // GC. In particular, for a parallel program we don't want to
236 // lose locality by moving cached data into another CPU's cache
237 // (this effect can be quite significant).
239 // We could have a more complex way to deterimine whether to do
240 // work stealing or not, e.g. it might be a good idea to do it
241 // if the heap is big. For now, we just turn it on or off with
245 /* Start threads, so they can be spinning up while we finish initialisation.
249 #if defined(THREADED_RTS)
250 /* How many threads will be participating in this GC?
251 * We don't try to parallelise minor GCs (unless the user asks for
252 * it with +RTS -gn0), or mark/compact/sweep GC.
254 if (gc_type == PENDING_GC_PAR) {
255 n_gc_threads = RtsFlags.ParFlags.nNodes;
263 debugTrace(DEBUG_gc, "GC (gen %d): %d KB to collect, %ld MB in use, using %d thread(s)",
264 N, n * (BLOCK_SIZE / 1024), mblocks_allocated, n_gc_threads);
266 #ifdef RTS_GTK_FRONTPANEL
267 if (RtsFlags.GcFlags.frontpanel) {
268 updateFrontPanelBeforeGC(N);
273 // check for memory leaks if DEBUG is on
274 memInventory(DEBUG_gc);
277 // check sanity *before* GC
278 IF_DEBUG(sanity, checkSanity(rtsTrue));
280 // Initialise all our gc_thread structures
281 for (t = 0; t < n_gc_threads; t++) {
282 init_gc_thread(gc_threads[t]);
285 // Initialise all the generations/steps that we're collecting.
286 for (g = 0; g <= N; g++) {
287 init_collected_gen(g,n_gc_threads);
290 // Initialise all the generations/steps that we're *not* collecting.
291 for (g = N+1; g < RtsFlags.GcFlags.generations; g++) {
292 init_uncollected_gen(g,n_gc_threads);
295 /* Allocate a mark stack if we're doing a major collection.
297 if (major_gc && oldest_gen->mark) {
298 mark_stack_bd = allocBlock();
299 mark_stack_top_bd = mark_stack_bd;
300 mark_stack_bd->link = NULL;
301 mark_stack_bd->u.back = NULL;
302 mark_sp = mark_stack_bd->start;
304 mark_stack_bd = NULL;
305 mark_stack_top_bd = NULL;
309 // this is the main thread
311 if (n_gc_threads == 1) {
312 SET_GCT(gc_threads[0]);
314 SET_GCT(gc_threads[cap->no]);
317 SET_GCT(gc_threads[0]);
320 /* -----------------------------------------------------------------------
321 * follow all the roots that we know about:
324 // the main thread is running: this prevents any other threads from
325 // exiting prematurely, so we can start them now.
326 // NB. do this after the mutable lists have been saved above, otherwise
327 // the other GC threads will be writing into the old mutable lists.
329 wakeup_gc_threads(n_gc_threads, gct->thread_index);
331 // Mutable lists from each generation > N
332 // we want to *scavenge* these roots, not evacuate them: they're not
333 // going to move in this GC.
334 // Also do them in reverse generation order, for the usual reason:
335 // namely to reduce the likelihood of spurious old->new pointers.
337 for (g = RtsFlags.GcFlags.generations-1; g > N; g--) {
338 #if defined(THREADED_RTS)
339 if (n_gc_threads > 1) {
340 scavenge_mutable_list(generations[g].saved_mut_list, &generations[g]);
342 scavenge_mutable_list1(generations[g].saved_mut_list, &generations[g]);
345 scavenge_mutable_list(generations[g].saved_mut_list, &generations[g]);
347 freeChain_sync(generations[g].saved_mut_list);
348 generations[g].saved_mut_list = NULL;
352 // scavenge the capability-private mutable lists. This isn't part
353 // of markSomeCapabilities() because markSomeCapabilities() can only
354 // call back into the GC via mark_root() (due to the gct register
356 if (n_gc_threads == 1) {
357 for (n = 0; n < n_capabilities; n++) {
358 #if defined(THREADED_RTS)
359 scavenge_capability_mut_Lists1(&capabilities[n]);
361 scavenge_capability_mut_lists(&capabilities[n]);
365 scavenge_capability_mut_lists(&capabilities[gct->thread_index]);
368 // follow roots from the CAF list (used by GHCi)
370 markCAFs(mark_root, gct);
372 // follow all the roots that the application knows about.
374 markSomeCapabilities(mark_root, gct, gct->thread_index, n_gc_threads,
375 rtsTrue/*prune sparks*/);
377 #if defined(RTS_USER_SIGNALS)
378 // mark the signal handlers (signals should be already blocked)
379 markSignalHandlers(mark_root, gct);
382 // Mark the weak pointer list, and prepare to detect dead weak pointers.
386 // Mark the stable pointer table.
387 markStablePtrTable(mark_root, gct);
389 /* -------------------------------------------------------------------------
390 * Repeatedly scavenge all the areas we know about until there's no
391 * more scavenging to be done.
395 scavenge_until_all_done();
396 // The other threads are now stopped. We might recurse back to
397 // here, but from now on this is the only thread.
399 // if any blackholes are alive, make the threads that wait on
401 if (traverseBlackholeQueue()) {
406 // must be last... invariant is that everything is fully
407 // scavenged at this point.
408 if (traverseWeakPtrList()) { // returns rtsTrue if evaced something
413 // If we get to here, there's really nothing left to do.
417 shutdown_gc_threads(n_gc_threads, gct->thread_index);
419 // Update pointers from the Task list
422 // Now see which stable names are still alive.
426 // We call processHeapClosureForDead() on every closure destroyed during
427 // the current garbage collection, so we invoke LdvCensusForDead().
428 if (RtsFlags.ProfFlags.doHeapProfile == HEAP_BY_LDV
429 || RtsFlags.ProfFlags.bioSelector != NULL)
433 // NO MORE EVACUATION AFTER THIS POINT!
435 // Two-space collector: free the old to-space.
436 // g0->old_blocks is the old nursery
437 // g0->blocks is to-space from the previous GC
438 if (RtsFlags.GcFlags.generations == 1) {
439 if (g0->blocks != NULL) {
440 freeChain(g0->blocks);
445 // For each workspace, in each thread, move the copied blocks to the step
451 for (t = 0; t < n_gc_threads; t++) {
454 for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
457 // Push the final block
459 push_scanned_block(ws->todo_bd, ws);
462 ASSERT(gct->scan_bd == NULL);
463 ASSERT(countBlocks(ws->scavd_list) == ws->n_scavd_blocks);
466 for (bd = ws->scavd_list; bd != NULL; bd = bd->link) {
467 ws->gen->n_words += bd->free - bd->start;
471 prev->link = ws->gen->blocks;
472 ws->gen->blocks = ws->scavd_list;
474 ws->gen->n_blocks += ws->n_scavd_blocks;
478 // Add all the partial blocks *after* we've added all the full
479 // blocks. This is so that we can grab the partial blocks back
480 // again and try to fill them up in the next GC.
481 for (t = 0; t < n_gc_threads; t++) {
484 for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
488 for (bd = ws->part_list; bd != NULL; bd = next) {
490 if (bd->free == bd->start) {
492 ws->part_list = next;
499 ws->gen->n_words += bd->free - bd->start;
504 prev->link = ws->gen->blocks;
505 ws->gen->blocks = ws->part_list;
507 ws->gen->n_blocks += ws->n_part_blocks;
509 ASSERT(countBlocks(ws->gen->blocks) == ws->gen->n_blocks);
510 ASSERT(countOccupied(ws->gen->blocks) == ws->gen->n_words);
515 // Finally: compact or sweep the oldest generation.
516 if (major_gc && oldest_gen->mark) {
517 if (oldest_gen->compact)
518 compact(gct->scavenged_static_objects);
523 /* run through all the generations/steps and tidy up
530 for (i=0; i < n_gc_threads; i++) {
531 if (n_gc_threads > 1) {
532 debugTrace(DEBUG_gc,"thread %d:", i);
533 debugTrace(DEBUG_gc," copied %ld", gc_threads[i]->copied * sizeof(W_));
534 debugTrace(DEBUG_gc," scanned %ld", gc_threads[i]->scanned * sizeof(W_));
535 debugTrace(DEBUG_gc," any_work %ld", gc_threads[i]->any_work);
536 debugTrace(DEBUG_gc," no_work %ld", gc_threads[i]->no_work);
537 debugTrace(DEBUG_gc," scav_find_work %ld", gc_threads[i]->scav_find_work);
539 copied += gc_threads[i]->copied;
540 max_copied = stg_max(gc_threads[i]->copied, max_copied);
542 if (n_gc_threads == 1) {
550 for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
553 generations[g].collections++; // for stats
554 if (n_gc_threads > 1) generations[g].par_collections++;
557 // Count the mutable list as bytes "copied" for the purposes of
558 // stats. Every mutable list is copied during every GC.
560 nat mut_list_size = 0;
561 for (bd = generations[g].mut_list; bd != NULL; bd = bd->link) {
562 mut_list_size += bd->free - bd->start;
564 for (n = 0; n < n_capabilities; n++) {
565 for (bd = capabilities[n].mut_lists[g];
566 bd != NULL; bd = bd->link) {
567 mut_list_size += bd->free - bd->start;
570 copied += mut_list_size;
573 "mut_list_size: %lu (%d vars, %d arrays, %d MVARs, %d others)",
574 (unsigned long)(mut_list_size * sizeof(W_)),
575 mutlist_MUTVARS, mutlist_MUTARRS, mutlist_MVARS, mutlist_OTHERS);
579 gen = &generations[g];
581 // for generations we collected...
584 /* free old memory and shift to-space into from-space for all
585 * the collected steps (except the allocation area). These
586 * freed blocks will probaby be quickly recycled.
590 // tack the new blocks on the end of the existing blocks
591 if (gen->old_blocks != NULL) {
594 for (bd = gen->old_blocks; bd != NULL; bd = next) {
598 if (!(bd->flags & BF_MARKED))
601 gen->old_blocks = next;
610 gen->n_words += bd->free - bd->start;
612 // NB. this step might not be compacted next
613 // time, so reset the BF_MARKED flags.
614 // They are set before GC if we're going to
615 // compact. (search for BF_MARKED above).
616 bd->flags &= ~BF_MARKED;
618 // between GCs, all blocks in the heap except
619 // for the nursery have the BF_EVACUATED flag set.
620 bd->flags |= BF_EVACUATED;
627 prev->link = gen->blocks;
628 gen->blocks = gen->old_blocks;
631 // add the new blocks to the block tally
632 gen->n_blocks += gen->n_old_blocks;
633 ASSERT(countBlocks(gen->blocks) == gen->n_blocks);
634 ASSERT(countOccupied(gen->blocks) == gen->n_words);
638 freeChain(gen->old_blocks);
641 gen->old_blocks = NULL;
642 gen->n_old_blocks = 0;
644 /* LARGE OBJECTS. The current live large objects are chained on
645 * scavenged_large, having been moved during garbage
646 * collection from large_objects. Any objects left on the
647 * large_objects list are therefore dead, so we free them here.
649 freeChain(gen->large_objects);
650 gen->large_objects = gen->scavenged_large_objects;
651 gen->n_large_blocks = gen->n_scavenged_large_blocks;
652 gen->n_new_large_blocks = 0;
653 ASSERT(countBlocks(gen->large_objects) == gen->n_large_blocks);
655 else // for generations > N
657 /* For older generations, we need to append the
658 * scavenged_large_object list (i.e. large objects that have been
659 * promoted during this GC) to the large_object list for that step.
661 for (bd = gen->scavenged_large_objects; bd; bd = next) {
663 dbl_link_onto(bd, &gen->large_objects);
666 // add the new blocks we promoted during this GC
667 gen->n_large_blocks += gen->n_scavenged_large_blocks;
668 ASSERT(countBlocks(gen->large_objects) == gen->n_large_blocks);
670 } // for all generations
672 // update the max size of older generations after a major GC
673 resize_generations();
675 // Calculate the amount of live data for stats.
676 live = calcLiveWords();
678 // Free the small objects allocated via allocate(), since this will
679 // all have been copied into G0S1 now.
680 alloc_blocks_lim = RtsFlags.GcFlags.minAllocAreaSize;
682 // Start a new pinned_object_block
683 for (n = 0; n < n_capabilities; n++) {
684 capabilities[n].pinned_object_block = NULL;
687 // Free the mark stack.
688 if (mark_stack_top_bd != NULL) {
689 debugTrace(DEBUG_gc, "mark stack: %d blocks",
690 countBlocks(mark_stack_top_bd));
691 freeChain(mark_stack_top_bd);
695 for (g = 0; g <= N; g++) {
696 gen = &generations[g];
697 if (gen->bitmap != NULL) {
698 freeGroup(gen->bitmap);
705 // mark the garbage collected CAFs as dead
706 #if 0 && defined(DEBUG) // doesn't work at the moment
707 if (major_gc) { gcCAFs(); }
711 // resetStaticObjectForRetainerProfiling() must be called before
713 if (n_gc_threads > 1) {
714 barf("profiling is currently broken with multi-threaded GC");
715 // ToDo: fix the gct->scavenged_static_objects below
717 resetStaticObjectForRetainerProfiling(gct->scavenged_static_objects);
720 // zero the scavenged static object list
723 for (i = 0; i < n_gc_threads; i++) {
724 zero_static_object_list(gc_threads[i]->scavenged_static_objects);
731 // start any pending finalizers
733 scheduleFinalizers(cap, old_weak_ptr_list);
736 // send exceptions to any threads which were about to die
738 resurrectThreads(resurrected_threads);
739 performPendingThrowTos(exception_threads);
742 // Update the stable pointer hash table.
743 updateStablePtrTable(major_gc);
745 // check sanity after GC
746 IF_DEBUG(sanity, checkSanity(rtsTrue));
748 // extra GC trace info
749 IF_DEBUG(gc, statDescribeGens());
752 // symbol-table based profiling
753 /* heapCensus(to_blocks); */ /* ToDo */
756 // restore enclosing cost centre
762 // check for memory leaks if DEBUG is on
763 memInventory(DEBUG_gc);
766 #ifdef RTS_GTK_FRONTPANEL
767 if (RtsFlags.GcFlags.frontpanel) {
768 updateFrontPanelAfterGC( N, live );
772 // ok, GC over: tell the stats department what happened.
773 slop = calcLiveBlocks() * BLOCK_SIZE_W - live;
774 stat_endGC(allocated, live, copied, N, max_copied, avg_copied, slop);
776 // unlock the StablePtr table
779 // Guess which generation we'll collect *next* time
780 initialise_N(force_major_gc);
782 #if defined(RTS_USER_SIGNALS)
783 if (RtsFlags.MiscFlags.install_signal_handlers) {
784 // unblock signals again
785 unblockUserSignals();
794 /* -----------------------------------------------------------------------------
795 Figure out which generation to collect, initialise N and major_gc.
797 Also returns the total number of blocks in generations that will be
799 -------------------------------------------------------------------------- */
802 initialise_N (rtsBool force_major_gc)
805 nat blocks, blocks_total;
810 if (force_major_gc) {
811 N = RtsFlags.GcFlags.generations - 1;
816 for (g = RtsFlags.GcFlags.generations - 1; g >= 0; g--) {
818 blocks = generations[g].n_words / BLOCK_SIZE_W
819 + generations[g].n_large_blocks;
821 if (blocks >= generations[g].max_blocks) {
825 blocks_total += blocks;
829 blocks_total += countNurseryBlocks();
831 major_gc = (N == RtsFlags.GcFlags.generations-1);
835 /* -----------------------------------------------------------------------------
836 Initialise the gc_thread structures.
837 -------------------------------------------------------------------------- */
839 #define GC_THREAD_INACTIVE 0
840 #define GC_THREAD_STANDING_BY 1
841 #define GC_THREAD_RUNNING 2
842 #define GC_THREAD_WAITING_TO_CONTINUE 3
845 new_gc_thread (nat n, gc_thread *t)
852 initSpinLock(&t->gc_spin);
853 initSpinLock(&t->mut_spin);
854 ACQUIRE_SPIN_LOCK(&t->gc_spin);
855 t->wakeup = GC_THREAD_INACTIVE; // starts true, so we can wait for the
856 // thread to start up, see wakeup_gc_threads
860 t->free_blocks = NULL;
869 for (g = 0; g < RtsFlags.GcFlags.generations; g++)
872 ws->gen = &generations[g];
873 ASSERT(g == ws->gen->no);
877 ws->todo_q = newWSDeque(128);
878 ws->todo_overflow = NULL;
879 ws->n_todo_overflow = 0;
881 ws->part_list = NULL;
882 ws->n_part_blocks = 0;
884 ws->scavd_list = NULL;
885 ws->n_scavd_blocks = 0;
893 if (gc_threads == NULL) {
894 #if defined(THREADED_RTS)
896 gc_threads = stgMallocBytes (RtsFlags.ParFlags.nNodes *
900 for (i = 0; i < RtsFlags.ParFlags.nNodes; i++) {
902 stgMallocBytes(sizeof(gc_thread) +
903 RtsFlags.GcFlags.generations * sizeof(gen_workspace),
906 new_gc_thread(i, gc_threads[i]);
909 gc_threads = stgMallocBytes (sizeof(gc_thread*),"alloc_gc_threads");
911 new_gc_thread(0,gc_threads[0]);
920 if (gc_threads != NULL) {
921 #if defined(THREADED_RTS)
923 for (i = 0; i < n_capabilities; i++) {
924 for (g = 0; g < RtsFlags.GcFlags.generations; g++)
926 freeWSDeque(gc_threads[i]->gens[g].todo_q);
928 stgFree (gc_threads[i]);
930 stgFree (gc_threads);
932 for (g = 0; g < RtsFlags.GcFlags.generations; g++)
934 freeWSDeque(gc_threads[0]->gens[g].todo_q);
936 stgFree (gc_threads);
942 /* ----------------------------------------------------------------------------
944 ------------------------------------------------------------------------- */
946 static volatile StgWord gc_running_threads;
952 new = atomic_inc(&gc_running_threads);
953 ASSERT(new <= n_gc_threads);
960 ASSERT(gc_running_threads != 0);
961 return atomic_dec(&gc_running_threads);
974 // scavenge objects in compacted generation
975 if (mark_stack_bd != NULL && !mark_stack_empty()) {
979 // Check for global work in any step. We don't need to check for
980 // local work, because we have already exited scavenge_loop(),
981 // which means there is no local work for this thread.
982 for (g = 0; g < (int)RtsFlags.GcFlags.generations; g++) {
984 if (ws->todo_large_objects) return rtsTrue;
985 if (!looksEmptyWSDeque(ws->todo_q)) return rtsTrue;
986 if (ws->todo_overflow) return rtsTrue;
989 #if defined(THREADED_RTS)
992 // look for work to steal
993 for (n = 0; n < n_gc_threads; n++) {
994 if (n == gct->thread_index) continue;
995 for (g = RtsFlags.GcFlags.generations-1; g >= 0; g--) {
996 ws = &gc_threads[n]->gens[g];
997 if (!looksEmptyWSDeque(ws->todo_q)) return rtsTrue;
1009 scavenge_until_all_done (void)
1015 traceEvent(&capabilities[gct->thread_index], EVENT_GC_WORK);
1017 #if defined(THREADED_RTS)
1018 if (n_gc_threads > 1) {
1027 // scavenge_loop() only exits when there's no work to do
1030 traceEvent(&capabilities[gct->thread_index], EVENT_GC_IDLE);
1032 debugTrace(DEBUG_gc, "%d GC threads still running", r);
1034 while (gc_running_threads != 0) {
1040 // any_work() does not remove the work from the queue, it
1041 // just checks for the presence of work. If we find any,
1042 // then we increment gc_running_threads and go back to
1043 // scavenge_loop() to perform any pending work.
1046 traceEvent(&capabilities[gct->thread_index], EVENT_GC_DONE);
1049 #if defined(THREADED_RTS)
1052 gcWorkerThread (Capability *cap)
1054 gc_thread *saved_gct;
1056 // necessary if we stole a callee-saves register for gct:
1059 gct = gc_threads[cap->no];
1060 gct->id = osThreadId();
1062 // Wait until we're told to wake up
1063 RELEASE_SPIN_LOCK(&gct->mut_spin);
1064 gct->wakeup = GC_THREAD_STANDING_BY;
1065 debugTrace(DEBUG_gc, "GC thread %d standing by...", gct->thread_index);
1066 ACQUIRE_SPIN_LOCK(&gct->gc_spin);
1069 // start performance counters in this thread...
1070 if (gct->papi_events == -1) {
1071 papi_init_eventset(&gct->papi_events);
1073 papi_thread_start_gc1_count(gct->papi_events);
1076 // Every thread evacuates some roots.
1078 markSomeCapabilities(mark_root, gct, gct->thread_index, n_gc_threads,
1079 rtsTrue/*prune sparks*/);
1080 scavenge_capability_mut_lists(&capabilities[gct->thread_index]);
1082 scavenge_until_all_done();
1085 // count events in this thread towards the GC totals
1086 papi_thread_stop_gc1_count(gct->papi_events);
1089 // Wait until we're told to continue
1090 RELEASE_SPIN_LOCK(&gct->gc_spin);
1091 gct->wakeup = GC_THREAD_WAITING_TO_CONTINUE;
1092 debugTrace(DEBUG_gc, "GC thread %d waiting to continue...",
1094 ACQUIRE_SPIN_LOCK(&gct->mut_spin);
1095 debugTrace(DEBUG_gc, "GC thread %d on my way...", gct->thread_index);
1102 #if defined(THREADED_RTS)
1105 waitForGcThreads (Capability *cap USED_IF_THREADS)
1107 nat n_threads = RtsFlags.ParFlags.nNodes;
1110 rtsBool retry = rtsTrue;
1113 for (i=0; i < n_threads; i++) {
1114 if (i == me) continue;
1115 if (gc_threads[i]->wakeup != GC_THREAD_STANDING_BY) {
1116 prodCapability(&capabilities[i], cap->running_task);
1119 for (j=0; j < 10; j++) {
1121 for (i=0; i < n_threads; i++) {
1122 if (i == me) continue;
1124 setContextSwitches();
1125 if (gc_threads[i]->wakeup != GC_THREAD_STANDING_BY) {
1135 #endif // THREADED_RTS
1138 start_gc_threads (void)
1140 #if defined(THREADED_RTS)
1141 gc_running_threads = 0;
1146 wakeup_gc_threads (nat n_threads USED_IF_THREADS, nat me USED_IF_THREADS)
1148 #if defined(THREADED_RTS)
1150 for (i=0; i < n_threads; i++) {
1151 if (i == me) continue;
1153 debugTrace(DEBUG_gc, "waking up gc thread %d", i);
1154 if (gc_threads[i]->wakeup != GC_THREAD_STANDING_BY) barf("wakeup_gc_threads");
1156 gc_threads[i]->wakeup = GC_THREAD_RUNNING;
1157 ACQUIRE_SPIN_LOCK(&gc_threads[i]->mut_spin);
1158 RELEASE_SPIN_LOCK(&gc_threads[i]->gc_spin);
1163 // After GC is complete, we must wait for all GC threads to enter the
1164 // standby state, otherwise they may still be executing inside
1165 // any_work(), and may even remain awake until the next GC starts.
1167 shutdown_gc_threads (nat n_threads USED_IF_THREADS, nat me USED_IF_THREADS)
1169 #if defined(THREADED_RTS)
1171 for (i=0; i < n_threads; i++) {
1172 if (i == me) continue;
1173 while (gc_threads[i]->wakeup != GC_THREAD_WAITING_TO_CONTINUE) { write_barrier(); }
1178 #if defined(THREADED_RTS)
1180 releaseGCThreads (Capability *cap USED_IF_THREADS)
1182 nat n_threads = RtsFlags.ParFlags.nNodes;
1185 for (i=0; i < n_threads; i++) {
1186 if (i == me) continue;
1187 if (gc_threads[i]->wakeup != GC_THREAD_WAITING_TO_CONTINUE)
1188 barf("releaseGCThreads");
1190 gc_threads[i]->wakeup = GC_THREAD_INACTIVE;
1191 ACQUIRE_SPIN_LOCK(&gc_threads[i]->gc_spin);
1192 RELEASE_SPIN_LOCK(&gc_threads[i]->mut_spin);
1197 /* ----------------------------------------------------------------------------
1198 Initialise a generation that is to be collected
1199 ------------------------------------------------------------------------- */
1202 init_collected_gen (nat g, nat n_threads)
1209 // Throw away the current mutable list. Invariant: the mutable
1210 // list always has at least one block; this means we can avoid a
1211 // check for NULL in recordMutable().
1213 freeChain(generations[g].mut_list);
1214 generations[g].mut_list = allocBlock();
1215 for (i = 0; i < n_capabilities; i++) {
1216 freeChain(capabilities[i].mut_lists[g]);
1217 capabilities[i].mut_lists[g] = allocBlock();
1221 gen = &generations[g];
1222 ASSERT(gen->no == g);
1224 // we'll construct a new list of threads in this step
1225 // during GC, throw away the current list.
1226 gen->old_threads = gen->threads;
1227 gen->threads = END_TSO_QUEUE;
1229 // deprecate the existing blocks
1230 gen->old_blocks = gen->blocks;
1231 gen->n_old_blocks = gen->n_blocks;
1235 gen->live_estimate = 0;
1237 // initialise the large object queues.
1238 gen->scavenged_large_objects = NULL;
1239 gen->n_scavenged_large_blocks = 0;
1241 // mark the small objects as from-space
1242 for (bd = gen->old_blocks; bd; bd = bd->link) {
1243 bd->flags &= ~BF_EVACUATED;
1246 // mark the large objects as from-space
1247 for (bd = gen->large_objects; bd; bd = bd->link) {
1248 bd->flags &= ~BF_EVACUATED;
1251 // for a compacted generation, we need to allocate the bitmap
1253 nat bitmap_size; // in bytes
1254 bdescr *bitmap_bdescr;
1257 bitmap_size = gen->n_old_blocks * BLOCK_SIZE / (sizeof(W_)*BITS_PER_BYTE);
1259 if (bitmap_size > 0) {
1260 bitmap_bdescr = allocGroup((lnat)BLOCK_ROUND_UP(bitmap_size)
1262 gen->bitmap = bitmap_bdescr;
1263 bitmap = bitmap_bdescr->start;
1265 debugTrace(DEBUG_gc, "bitmap_size: %d, bitmap: %p",
1266 bitmap_size, bitmap);
1268 // don't forget to fill it with zeros!
1269 memset(bitmap, 0, bitmap_size);
1271 // For each block in this step, point to its bitmap from the
1272 // block descriptor.
1273 for (bd=gen->old_blocks; bd != NULL; bd = bd->link) {
1274 bd->u.bitmap = bitmap;
1275 bitmap += BLOCK_SIZE_W / (sizeof(W_)*BITS_PER_BYTE);
1277 // Also at this point we set the BF_MARKED flag
1278 // for this block. The invariant is that
1279 // BF_MARKED is always unset, except during GC
1280 // when it is set on those blocks which will be
1282 if (!(bd->flags & BF_FRAGMENTED)) {
1283 bd->flags |= BF_MARKED;
1289 // For each GC thread, for each step, allocate a "todo" block to
1290 // store evacuated objects to be scavenged, and a block to store
1291 // evacuated objects that do not need to be scavenged.
1292 for (t = 0; t < n_threads; t++) {
1293 ws = &gc_threads[t]->gens[g];
1295 ws->todo_large_objects = NULL;
1297 ws->part_list = NULL;
1298 ws->n_part_blocks = 0;
1300 // allocate the first to-space block; extra blocks will be
1301 // chained on as necessary.
1303 ASSERT(looksEmptyWSDeque(ws->todo_q));
1304 alloc_todo_block(ws,0);
1306 ws->todo_overflow = NULL;
1307 ws->n_todo_overflow = 0;
1309 ws->scavd_list = NULL;
1310 ws->n_scavd_blocks = 0;
1315 /* ----------------------------------------------------------------------------
1316 Initialise a generation that is *not* to be collected
1317 ------------------------------------------------------------------------- */
1320 init_uncollected_gen (nat g, nat threads)
1327 // save the current mutable lists for this generation, and
1328 // allocate a fresh block for each one. We'll traverse these
1329 // mutable lists as roots early on in the GC.
1330 generations[g].saved_mut_list = generations[g].mut_list;
1331 generations[g].mut_list = allocBlock();
1332 for (n = 0; n < n_capabilities; n++) {
1333 capabilities[n].saved_mut_lists[g] = capabilities[n].mut_lists[g];
1334 capabilities[n].mut_lists[g] = allocBlock();
1337 gen = &generations[g];
1339 gen->scavenged_large_objects = NULL;
1340 gen->n_scavenged_large_blocks = 0;
1342 for (t = 0; t < threads; t++) {
1343 ws = &gc_threads[t]->gens[g];
1345 ASSERT(looksEmptyWSDeque(ws->todo_q));
1346 ws->todo_large_objects = NULL;
1348 ws->part_list = NULL;
1349 ws->n_part_blocks = 0;
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 (gen->blocks && isPartiallyFull(gen->blocks))
1358 ws->todo_bd = gen->blocks;
1359 ws->todo_free = ws->todo_bd->free;
1360 ws->todo_lim = ws->todo_bd->start + BLOCK_SIZE_W;
1361 gen->blocks = gen->blocks->link;
1363 gen->n_words -= ws->todo_bd->free - ws->todo_bd->start;
1364 ws->todo_bd->link = NULL;
1365 // we must scan from the current end point.
1366 ws->todo_bd->u.scan = ws->todo_bd->free;
1371 alloc_todo_block(ws,0);
1375 // deal out any more partial blocks to the threads' part_lists
1377 while (gen->blocks && isPartiallyFull(gen->blocks))
1380 gen->blocks = bd->link;
1381 ws = &gc_threads[t]->gens[g];
1382 bd->link = ws->part_list;
1384 ws->n_part_blocks += 1;
1385 bd->u.scan = bd->free;
1387 gen->n_words -= bd->free - bd->start;
1389 if (t == n_gc_threads) t = 0;
1393 /* -----------------------------------------------------------------------------
1394 Initialise a gc_thread before GC
1395 -------------------------------------------------------------------------- */
1398 init_gc_thread (gc_thread *t)
1400 t->static_objects = END_OF_STATIC_LIST;
1401 t->scavenged_static_objects = END_OF_STATIC_LIST;
1403 t->mut_lists = capabilities[t->thread_index].mut_lists;
1405 t->failed_to_evac = rtsFalse;
1406 t->eager_promotion = rtsTrue;
1407 t->thunk_selector_depth = 0;
1412 t->scav_find_work = 0;
1415 /* -----------------------------------------------------------------------------
1416 Function we pass to evacuate roots.
1417 -------------------------------------------------------------------------- */
1420 mark_root(void *user USED_IF_THREADS, StgClosure **root)
1422 // we stole a register for gct, but this function is called from
1423 // *outside* the GC where the register variable is not in effect,
1424 // so we need to save and restore it here. NB. only call
1425 // mark_root() from the main GC thread, otherwise gct will be
1427 gc_thread *saved_gct;
1436 /* -----------------------------------------------------------------------------
1437 Initialising the static object & mutable lists
1438 -------------------------------------------------------------------------- */
1441 zero_static_object_list(StgClosure* first_static)
1445 const StgInfoTable *info;
1447 for (p = first_static; p != END_OF_STATIC_LIST; p = link) {
1449 link = *STATIC_LINK(info, p);
1450 *STATIC_LINK(info,p) = NULL;
1454 /* ----------------------------------------------------------------------------
1455 Update the pointers from the task list
1457 These are treated as weak pointers because we want to allow a main
1458 thread to get a BlockedOnDeadMVar exception in the same way as any
1459 other thread. Note that the threads should all have been retained
1460 by GC by virtue of being on the all_threads list, we're just
1461 updating pointers here.
1462 ------------------------------------------------------------------------- */
1465 update_task_list (void)
1469 for (task = all_tasks; task != NULL; task = task->all_link) {
1470 if (!task->stopped && task->tso) {
1471 ASSERT(task->tso->bound == task);
1472 tso = (StgTSO *) isAlive((StgClosure *)task->tso);
1474 barf("task %p: main thread %d has been GC'd",
1487 /* ----------------------------------------------------------------------------
1488 Reset the sizes of the older generations when we do a major
1491 CURRENT STRATEGY: make all generations except zero the same size.
1492 We have to stay within the maximum heap size, and leave a certain
1493 percentage of the maximum heap size available to allocate into.
1494 ------------------------------------------------------------------------- */
1497 resize_generations (void)
1501 if (major_gc && RtsFlags.GcFlags.generations > 1) {
1502 nat live, size, min_alloc, words;
1503 nat max = RtsFlags.GcFlags.maxHeapSize;
1504 nat gens = RtsFlags.GcFlags.generations;
1506 // live in the oldest generations
1507 if (oldest_gen->live_estimate != 0) {
1508 words = oldest_gen->live_estimate;
1510 words = oldest_gen->n_words;
1512 live = (words + BLOCK_SIZE_W - 1) / BLOCK_SIZE_W +
1513 oldest_gen->n_large_blocks;
1515 // default max size for all generations except zero
1516 size = stg_max(live * RtsFlags.GcFlags.oldGenFactor,
1517 RtsFlags.GcFlags.minOldGenSize);
1519 if (RtsFlags.GcFlags.heapSizeSuggestionAuto) {
1520 RtsFlags.GcFlags.heapSizeSuggestion = size;
1523 // minimum size for generation zero
1524 min_alloc = stg_max((RtsFlags.GcFlags.pcFreeHeap * max) / 200,
1525 RtsFlags.GcFlags.minAllocAreaSize);
1527 // Auto-enable compaction when the residency reaches a
1528 // certain percentage of the maximum heap size (default: 30%).
1529 if (RtsFlags.GcFlags.generations > 1 &&
1530 (RtsFlags.GcFlags.compact ||
1532 oldest_gen->n_blocks >
1533 (RtsFlags.GcFlags.compactThreshold * max) / 100))) {
1534 oldest_gen->mark = 1;
1535 oldest_gen->compact = 1;
1536 // debugBelch("compaction: on\n", live);
1538 oldest_gen->mark = 0;
1539 oldest_gen->compact = 0;
1540 // debugBelch("compaction: off\n", live);
1543 if (RtsFlags.GcFlags.sweep) {
1544 oldest_gen->mark = 1;
1547 // if we're going to go over the maximum heap size, reduce the
1548 // size of the generations accordingly. The calculation is
1549 // different if compaction is turned on, because we don't need
1550 // to double the space required to collect the old generation.
1553 // this test is necessary to ensure that the calculations
1554 // below don't have any negative results - we're working
1555 // with unsigned values here.
1556 if (max < min_alloc) {
1560 if (oldest_gen->compact) {
1561 if ( (size + (size - 1) * (gens - 2) * 2) + min_alloc > max ) {
1562 size = (max - min_alloc) / ((gens - 1) * 2 - 1);
1565 if ( (size * (gens - 1) * 2) + min_alloc > max ) {
1566 size = (max - min_alloc) / ((gens - 1) * 2);
1576 debugBelch("live: %d, min_alloc: %d, size : %d, max = %d\n", live,
1577 min_alloc, size, max);
1580 for (g = 0; g < gens; g++) {
1581 generations[g].max_blocks = size;
1586 /* -----------------------------------------------------------------------------
1587 Calculate the new size of the nursery, and resize it.
1588 -------------------------------------------------------------------------- */
1591 resize_nursery (void)
1593 lnat min_nursery = RtsFlags.GcFlags.minAllocAreaSize * n_capabilities;
1595 if (RtsFlags.GcFlags.generations == 1)
1596 { // Two-space collector:
1599 /* set up a new nursery. Allocate a nursery size based on a
1600 * function of the amount of live data (by default a factor of 2)
1601 * Use the blocks from the old nursery if possible, freeing up any
1604 * If we get near the maximum heap size, then adjust our nursery
1605 * size accordingly. If the nursery is the same size as the live
1606 * data (L), then we need 3L bytes. We can reduce the size of the
1607 * nursery to bring the required memory down near 2L bytes.
1609 * A normal 2-space collector would need 4L bytes to give the same
1610 * performance we get from 3L bytes, reducing to the same
1611 * performance at 2L bytes.
1613 blocks = generations[0].n_blocks;
1615 if ( RtsFlags.GcFlags.maxHeapSize != 0 &&
1616 blocks * RtsFlags.GcFlags.oldGenFactor * 2 >
1617 RtsFlags.GcFlags.maxHeapSize )
1619 long adjusted_blocks; // signed on purpose
1622 adjusted_blocks = (RtsFlags.GcFlags.maxHeapSize - 2 * blocks);
1624 debugTrace(DEBUG_gc, "near maximum heap size of 0x%x blocks, blocks = %d, adjusted to %ld",
1625 RtsFlags.GcFlags.maxHeapSize, blocks, adjusted_blocks);
1627 pc_free = adjusted_blocks * 100 / RtsFlags.GcFlags.maxHeapSize;
1628 if (pc_free < RtsFlags.GcFlags.pcFreeHeap) /* might even * be < 0 */
1632 blocks = adjusted_blocks;
1636 blocks *= RtsFlags.GcFlags.oldGenFactor;
1637 if (blocks < min_nursery)
1639 blocks = min_nursery;
1642 resizeNurseries(blocks);
1644 else // Generational collector
1647 * If the user has given us a suggested heap size, adjust our
1648 * allocation area to make best use of the memory available.
1650 if (RtsFlags.GcFlags.heapSizeSuggestion)
1653 nat needed = calcNeeded(); // approx blocks needed at next GC
1655 /* Guess how much will be live in generation 0 step 0 next time.
1656 * A good approximation is obtained by finding the
1657 * percentage of g0 that was live at the last minor GC.
1659 * We have an accurate figure for the amount of copied data in
1660 * 'copied', but we must convert this to a number of blocks, with
1661 * a small adjustment for estimated slop at the end of a block
1666 g0_pcnt_kept = ((copied / (BLOCK_SIZE_W - 10)) * 100)
1667 / countNurseryBlocks();
1670 /* Estimate a size for the allocation area based on the
1671 * information available. We might end up going slightly under
1672 * or over the suggested heap size, but we should be pretty
1675 * Formula: suggested - needed
1676 * ----------------------------
1677 * 1 + g0_pcnt_kept/100
1679 * where 'needed' is the amount of memory needed at the next
1680 * collection for collecting all gens except g0.
1683 (((long)RtsFlags.GcFlags.heapSizeSuggestion - (long)needed) * 100) /
1684 (100 + (long)g0_pcnt_kept);
1686 if (blocks < (long)min_nursery) {
1687 blocks = min_nursery;
1690 resizeNurseries((nat)blocks);
1694 // we might have added extra large blocks to the nursery, so
1695 // resize back to minAllocAreaSize again.
1696 resizeNurseriesFixed(RtsFlags.GcFlags.minAllocAreaSize);
1701 /* -----------------------------------------------------------------------------
1702 Sanity code for CAF garbage collection.
1704 With DEBUG turned on, we manage a CAF list in addition to the SRT
1705 mechanism. After GC, we run down the CAF list and blackhole any
1706 CAFs which have been garbage collected. This means we get an error
1707 whenever the program tries to enter a garbage collected CAF.
1709 Any garbage collected CAFs are taken off the CAF list at the same
1711 -------------------------------------------------------------------------- */
1713 #if 0 && defined(DEBUG)
1720 const StgInfoTable *info;
1731 ASSERT(info->type == IND_STATIC);
1733 if (STATIC_LINK(info,p) == NULL) {
1734 debugTrace(DEBUG_gccafs, "CAF gc'd at 0x%04lx", (long)p);
1736 SET_INFO(p,&stg_BLACKHOLE_info);
1737 p = STATIC_LINK2(info,p);
1741 pp = &STATIC_LINK2(info,p);
1748 debugTrace(DEBUG_gccafs, "%d CAFs live", i);