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"
48 #include "MarkStack.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 g0_pcnt_kept = 30; // percentage of g0 live at last minor GC
115 /* Thread-local data for each GC thread
117 gc_thread **gc_threads = NULL;
119 #if !defined(THREADED_RTS)
120 StgWord8 the_gc_thread[sizeof(gc_thread) + 64 * sizeof(gen_workspace)];
123 // Number of threads running in *this* GC. Affects how many
124 // step->todos[] lists we have to look in to find work.
128 long copied; // *words* copied & scavenged during this GC
130 rtsBool work_stealing;
134 /* -----------------------------------------------------------------------------
135 Static function declarations
136 -------------------------------------------------------------------------- */
138 static void mark_root (void *user, StgClosure **root);
139 static void zero_static_object_list (StgClosure* first_static);
140 static nat initialise_N (rtsBool force_major_gc);
141 static void prepare_collected_gen (generation *gen);
142 static void prepare_uncollected_gen (generation *gen);
143 static void init_gc_thread (gc_thread *t);
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 me);
151 static void shutdown_gc_threads (nat me);
152 static void collect_gct_blocks (void);
154 #if 0 && defined(DEBUG)
155 static void gcCAFs (void);
158 /* -----------------------------------------------------------------------------
160 -------------------------------------------------------------------------- */
162 bdescr *mark_stack_top_bd; // topmost block in the mark stack
163 bdescr *mark_stack_bd; // current block in the mark stack
164 StgPtr mark_sp; // pointer to the next unallocated mark stack entry
166 /* -----------------------------------------------------------------------------
167 GarbageCollect: the main entry point to the garbage collector.
169 Locks held: all capabilities are held throughout GarbageCollect().
170 -------------------------------------------------------------------------- */
173 GarbageCollect (rtsBool force_major_gc,
174 nat gc_type USED_IF_THREADS,
179 lnat live_blocks, live_words, allocated, max_copied, avg_copied;
180 gc_thread *saved_gct;
183 // necessary if we stole a callee-saves register for gct:
187 CostCentreStack *prev_CCS;
192 #if defined(RTS_USER_SIGNALS)
193 if (RtsFlags.MiscFlags.install_signal_handlers) {
199 ASSERT(sizeof(gen_workspace) == 16 * sizeof(StgWord));
200 // otherwise adjust the padding in gen_workspace.
202 // this is the main thread
203 SET_GCT(gc_threads[cap->no]);
205 // tell the stats department that we've started a GC
208 // lock the StablePtr table
217 // attribute any costs to CCS_GC
223 /* Approximate how much we allocated.
224 * Todo: only when generating stats?
226 allocated = calcAllocated(rtsFalse/* don't count the nursery yet */);
228 /* Figure out which generation to collect
230 n = initialise_N(force_major_gc);
232 #if defined(THREADED_RTS)
233 work_stealing = RtsFlags.ParFlags.parGcLoadBalancingEnabled &&
234 N >= RtsFlags.ParFlags.parGcLoadBalancingGen;
235 // It's not always a good idea to do load balancing in parallel
236 // GC. In particular, for a parallel program we don't want to
237 // lose locality by moving cached data into another CPU's cache
238 // (this effect can be quite significant).
240 // We could have a more complex way to deterimine whether to do
241 // work stealing or not, e.g. it might be a good idea to do it
242 // if the heap is big. For now, we just turn it on or off with
246 /* Start threads, so they can be spinning up while we finish initialisation.
250 #if defined(THREADED_RTS)
251 /* How many threads will be participating in this GC?
252 * We don't try to parallelise minor GCs (unless the user asks for
253 * it with +RTS -gn0), or mark/compact/sweep GC.
255 if (gc_type == PENDING_GC_PAR) {
256 n_gc_threads = RtsFlags.ParFlags.nNodes;
264 debugTrace(DEBUG_gc, "GC (gen %d): %d KB to collect, %ld MB in use, using %d thread(s)",
265 N, n * (BLOCK_SIZE / 1024), mblocks_allocated, n_gc_threads);
267 #ifdef RTS_GTK_FRONTPANEL
268 if (RtsFlags.GcFlags.frontpanel) {
269 updateFrontPanelBeforeGC(N);
274 // check for memory leaks if DEBUG is on
275 memInventory(DEBUG_gc);
278 // check sanity *before* GC
279 IF_DEBUG(sanity, checkSanity(rtsFalse /* before GC */, major_gc));
281 // Initialise all the generations/steps that we're collecting.
282 for (g = 0; g <= N; g++) {
283 prepare_collected_gen(&generations[g]);
285 // Initialise all the generations/steps that we're *not* collecting.
286 for (g = N+1; g < RtsFlags.GcFlags.generations; g++) {
287 prepare_uncollected_gen(&generations[g]);
290 // Prepare this gc_thread
293 /* Allocate a mark stack if we're doing a major collection.
295 if (major_gc && oldest_gen->mark) {
296 mark_stack_bd = allocBlock();
297 mark_stack_top_bd = mark_stack_bd;
298 mark_stack_bd->link = NULL;
299 mark_stack_bd->u.back = NULL;
300 mark_sp = mark_stack_bd->start;
302 mark_stack_bd = NULL;
303 mark_stack_top_bd = NULL;
307 /* -----------------------------------------------------------------------
308 * follow all the roots that we know about:
311 // the main thread is running: this prevents any other threads from
312 // exiting prematurely, so we can start them now.
313 // NB. do this after the mutable lists have been saved above, otherwise
314 // the other GC threads will be writing into the old mutable lists.
316 wakeup_gc_threads(gct->thread_index);
318 traceEventGcWork(gct->cap);
320 // scavenge the capability-private mutable lists. This isn't part
321 // of markSomeCapabilities() because markSomeCapabilities() can only
322 // call back into the GC via mark_root() (due to the gct register
324 if (n_gc_threads == 1) {
325 for (n = 0; n < n_capabilities; n++) {
326 #if defined(THREADED_RTS)
327 scavenge_capability_mut_Lists1(&capabilities[n]);
329 scavenge_capability_mut_lists(&capabilities[n]);
333 scavenge_capability_mut_lists(gct->cap);
336 // follow roots from the CAF list (used by GHCi)
337 gct->evac_gen_no = 0;
338 markCAFs(mark_root, gct);
340 // follow all the roots that the application knows about.
341 gct->evac_gen_no = 0;
342 if (n_gc_threads == 1) {
343 for (n = 0; n < n_capabilities; n++) {
344 markCapability(mark_root, gct, &capabilities[n],
345 rtsTrue/*don't mark sparks*/);
348 markCapability(mark_root, gct, cap, rtsTrue/*don't mark sparks*/);
351 markScheduler(mark_root, gct);
353 #if defined(RTS_USER_SIGNALS)
354 // mark the signal handlers (signals should be already blocked)
355 markSignalHandlers(mark_root, gct);
358 // Mark the weak pointer list, and prepare to detect dead weak pointers.
362 // Mark the stable pointer table.
363 markStablePtrTable(mark_root, gct);
365 /* -------------------------------------------------------------------------
366 * Repeatedly scavenge all the areas we know about until there's no
367 * more scavenging to be done.
371 scavenge_until_all_done();
372 // The other threads are now stopped. We might recurse back to
373 // here, but from now on this is the only thread.
375 // must be last... invariant is that everything is fully
376 // scavenged at this point.
377 if (traverseWeakPtrList()) { // returns rtsTrue if evaced something
382 // If we get to here, there's really nothing left to do.
386 shutdown_gc_threads(gct->thread_index);
388 // Now see which stable names are still alive.
392 if (n_gc_threads == 1) {
393 for (n = 0; n < n_capabilities; n++) {
394 pruneSparkQueue(&capabilities[n]);
397 pruneSparkQueue(gct->cap);
402 // We call processHeapClosureForDead() on every closure destroyed during
403 // the current garbage collection, so we invoke LdvCensusForDead().
404 if (RtsFlags.ProfFlags.doHeapProfile == HEAP_BY_LDV
405 || RtsFlags.ProfFlags.bioSelector != NULL)
409 // NO MORE EVACUATION AFTER THIS POINT!
411 // Two-space collector: free the old to-space.
412 // g0->old_blocks is the old nursery
413 // g0->blocks is to-space from the previous GC
414 if (RtsFlags.GcFlags.generations == 1) {
415 if (g0->blocks != NULL) {
416 freeChain(g0->blocks);
421 // Finally: compact or sweep the oldest generation.
422 if (major_gc && oldest_gen->mark) {
423 if (oldest_gen->compact)
424 compact(gct->scavenged_static_objects);
434 for (i=0; i < n_gc_threads; i++) {
435 if (n_gc_threads > 1) {
436 debugTrace(DEBUG_gc,"thread %d:", i);
437 debugTrace(DEBUG_gc," copied %ld", gc_threads[i]->copied * sizeof(W_));
438 debugTrace(DEBUG_gc," scanned %ld", gc_threads[i]->scanned * sizeof(W_));
439 debugTrace(DEBUG_gc," any_work %ld", gc_threads[i]->any_work);
440 debugTrace(DEBUG_gc," no_work %ld", gc_threads[i]->no_work);
441 debugTrace(DEBUG_gc," scav_find_work %ld", gc_threads[i]->scav_find_work);
443 copied += gc_threads[i]->copied;
444 max_copied = stg_max(gc_threads[i]->copied, max_copied);
446 if (n_gc_threads == 1) {
454 // Run through all the generations/steps and tidy up.
456 // - count the amount of "live" data (live_words, live_blocks)
457 // - count the amount of "copied" data in this GC (copied)
459 // - make to-space the new from-space (set BF_EVACUATED on all blocks)
464 for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
467 generations[g].collections++; // for stats
468 if (n_gc_threads > 1) generations[g].par_collections++;
471 // Count the mutable list as bytes "copied" for the purposes of
472 // stats. Every mutable list is copied during every GC.
474 nat mut_list_size = 0;
475 for (n = 0; n < n_capabilities; n++) {
476 mut_list_size += countOccupied(capabilities[n].mut_lists[g]);
478 copied += mut_list_size;
481 "mut_list_size: %lu (%d vars, %d arrays, %d MVARs, %d others)",
482 (unsigned long)(mut_list_size * sizeof(W_)),
483 mutlist_MUTVARS, mutlist_MUTARRS, mutlist_MVARS, mutlist_OTHERS);
487 gen = &generations[g];
489 // for generations we collected...
492 /* free old memory and shift to-space into from-space for all
493 * the collected steps (except the allocation area). These
494 * freed blocks will probaby be quickly recycled.
498 // tack the new blocks on the end of the existing blocks
499 if (gen->old_blocks != NULL) {
502 for (bd = gen->old_blocks; bd != NULL; bd = next) {
506 if (!(bd->flags & BF_MARKED))
509 gen->old_blocks = next;
518 gen->n_words += bd->free - bd->start;
520 // NB. this step might not be compacted next
521 // time, so reset the BF_MARKED flags.
522 // They are set before GC if we're going to
523 // compact. (search for BF_MARKED above).
524 bd->flags &= ~BF_MARKED;
526 // between GCs, all blocks in the heap except
527 // for the nursery have the BF_EVACUATED flag set.
528 bd->flags |= BF_EVACUATED;
535 prev->link = gen->blocks;
536 gen->blocks = gen->old_blocks;
539 // add the new blocks to the block tally
540 gen->n_blocks += gen->n_old_blocks;
541 ASSERT(countBlocks(gen->blocks) == gen->n_blocks);
542 ASSERT(countOccupied(gen->blocks) == gen->n_words);
546 freeChain(gen->old_blocks);
549 gen->old_blocks = NULL;
550 gen->n_old_blocks = 0;
552 /* LARGE OBJECTS. The current live large objects are chained on
553 * scavenged_large, having been moved during garbage
554 * collection from large_objects. Any objects left on the
555 * large_objects list are therefore dead, so we free them here.
557 freeChain(gen->large_objects);
558 gen->large_objects = gen->scavenged_large_objects;
559 gen->n_large_blocks = gen->n_scavenged_large_blocks;
560 gen->n_new_large_words = 0;
562 else // for generations > N
564 /* For older generations, we need to append the
565 * scavenged_large_object list (i.e. large objects that have been
566 * promoted during this GC) to the large_object list for that step.
568 for (bd = gen->scavenged_large_objects; bd; bd = next) {
570 dbl_link_onto(bd, &gen->large_objects);
573 // add the new blocks we promoted during this GC
574 gen->n_large_blocks += gen->n_scavenged_large_blocks;
577 ASSERT(countBlocks(gen->large_objects) == gen->n_large_blocks);
579 gen->scavenged_large_objects = NULL;
580 gen->n_scavenged_large_blocks = 0;
583 live_words += genLiveWords(gen);
584 live_blocks += genLiveBlocks(gen);
586 // add in the partial blocks in the gen_workspaces, but ignore gen 0
587 // if this is a local GC (we can't count another capability's part_list)
590 for (i = 0; i < n_capabilities; i++) {
591 live_words += gcThreadLiveWords(i, gen->no);
592 live_blocks += gcThreadLiveBlocks(i, gen->no);
595 } // for all generations
597 // update the max size of older generations after a major GC
598 resize_generations();
600 // Start a new pinned_object_block
601 for (n = 0; n < n_capabilities; n++) {
602 capabilities[n].pinned_object_block = NULL;
605 // Free the mark stack.
606 if (mark_stack_top_bd != NULL) {
607 debugTrace(DEBUG_gc, "mark stack: %d blocks",
608 countBlocks(mark_stack_top_bd));
609 freeChain(mark_stack_top_bd);
613 for (g = 0; g <= N; g++) {
614 gen = &generations[g];
615 if (gen->bitmap != NULL) {
616 freeGroup(gen->bitmap);
621 // Reset the nursery: make the blocks empty
622 allocated += clearNurseries();
628 // mark the garbage collected CAFs as dead
629 #if 0 && defined(DEBUG) // doesn't work at the moment
630 if (major_gc) { gcCAFs(); }
634 // resetStaticObjectForRetainerProfiling() must be called before
636 if (n_gc_threads > 1) {
637 barf("profiling is currently broken with multi-threaded GC");
638 // ToDo: fix the gct->scavenged_static_objects below
640 resetStaticObjectForRetainerProfiling(gct->scavenged_static_objects);
643 // zero the scavenged static object list
646 for (i = 0; i < n_gc_threads; i++) {
647 zero_static_object_list(gc_threads[i]->scavenged_static_objects);
651 // Update the stable pointer hash table.
652 updateStablePtrTable(major_gc);
654 // unlock the StablePtr table. Must be before scheduleFinalizers(),
655 // because a finalizer may call hs_free_fun_ptr() or
656 // hs_free_stable_ptr(), both of which access the StablePtr table.
659 // Start any pending finalizers. Must be after
660 // updateStablePtrTable() and stablePtrPostGC() (see #4221).
662 scheduleFinalizers(cap, old_weak_ptr_list);
665 // check sanity after GC
666 // before resurrectThreads(), because that might overwrite some
667 // closures, which will cause problems with THREADED where we don't
669 IF_DEBUG(sanity, checkSanity(rtsTrue /* after GC */, major_gc));
671 // send exceptions to any threads which were about to die
673 resurrectThreads(resurrected_threads);
678 need = BLOCKS_TO_MBLOCKS(n_alloc_blocks);
679 got = mblocks_allocated;
680 /* If the amount of data remains constant, next major GC we'll
681 require (F+1)*need. We leave (F+2)*need in order to reduce
682 repeated deallocation and reallocation. */
683 need = (RtsFlags.GcFlags.oldGenFactor + 2) * need;
685 returnMemoryToOS(got - need);
689 // extra GC trace info
690 IF_DEBUG(gc, statDescribeGens());
693 // symbol-table based profiling
694 /* heapCensus(to_blocks); */ /* ToDo */
697 // restore enclosing cost centre
703 // check for memory leaks if DEBUG is on
704 memInventory(DEBUG_gc);
707 #ifdef RTS_GTK_FRONTPANEL
708 if (RtsFlags.GcFlags.frontpanel) {
709 updateFrontPanelAfterGC( N, live );
713 // ok, GC over: tell the stats department what happened.
714 stat_endGC(gct, allocated, live_words,
715 copied, N, max_copied, avg_copied,
716 live_blocks * BLOCK_SIZE_W - live_words /* slop */);
718 // Guess which generation we'll collect *next* time
719 initialise_N(force_major_gc);
721 #if defined(RTS_USER_SIGNALS)
722 if (RtsFlags.MiscFlags.install_signal_handlers) {
723 // unblock signals again
724 unblockUserSignals();
733 /* -----------------------------------------------------------------------------
734 Figure out which generation to collect, initialise N and major_gc.
736 Also returns the total number of blocks in generations that will be
738 -------------------------------------------------------------------------- */
741 initialise_N (rtsBool force_major_gc)
744 nat blocks, blocks_total;
749 if (force_major_gc) {
750 N = RtsFlags.GcFlags.generations - 1;
755 for (g = RtsFlags.GcFlags.generations - 1; g >= 0; g--) {
757 blocks = generations[g].n_words / BLOCK_SIZE_W
758 + generations[g].n_large_blocks;
760 if (blocks >= generations[g].max_blocks) {
764 blocks_total += blocks;
768 blocks_total += countNurseryBlocks();
770 major_gc = (N == RtsFlags.GcFlags.generations-1);
774 /* -----------------------------------------------------------------------------
775 Initialise the gc_thread structures.
776 -------------------------------------------------------------------------- */
778 #define GC_THREAD_INACTIVE 0
779 #define GC_THREAD_STANDING_BY 1
780 #define GC_THREAD_RUNNING 2
781 #define GC_THREAD_WAITING_TO_CONTINUE 3
784 new_gc_thread (nat n, gc_thread *t)
789 t->cap = &capabilities[n];
793 initSpinLock(&t->gc_spin);
794 initSpinLock(&t->mut_spin);
795 ACQUIRE_SPIN_LOCK(&t->gc_spin);
796 t->wakeup = GC_THREAD_INACTIVE; // starts true, so we can wait for the
797 // thread to start up, see wakeup_gc_threads
801 t->free_blocks = NULL;
810 for (g = 0; g < RtsFlags.GcFlags.generations; g++)
813 ws->gen = &generations[g];
814 ASSERT(g == ws->gen->no);
818 // alloc_todo_block(ws,0);
819 // but can't, because it uses gct which isn't set up at this point.
820 // Hence, allocate a block for todo_bd manually:
822 bdescr *bd = allocBlock(); // no lock, locks aren't initialised yet
823 initBdescr(bd, ws->gen, ws->gen->to);
824 bd->flags = BF_EVACUATED;
825 bd->u.scan = bd->free = bd->start;
828 ws->todo_free = bd->free;
829 ws->todo_lim = bd->start + BLOCK_SIZE_W;
832 ws->todo_q = newWSDeque(128);
833 ws->todo_overflow = NULL;
834 ws->n_todo_overflow = 0;
835 ws->todo_large_objects = NULL;
837 ws->part_list = NULL;
838 ws->n_part_blocks = 0;
840 ws->scavd_list = NULL;
841 ws->n_scavd_blocks = 0;
849 if (gc_threads == NULL) {
850 #if defined(THREADED_RTS)
852 gc_threads = stgMallocBytes (RtsFlags.ParFlags.nNodes *
856 for (i = 0; i < RtsFlags.ParFlags.nNodes; i++) {
858 stgMallocBytes(sizeof(gc_thread) +
859 RtsFlags.GcFlags.generations * sizeof(gen_workspace),
862 new_gc_thread(i, gc_threads[i]);
865 gc_threads = stgMallocBytes (sizeof(gc_thread*),"alloc_gc_threads");
867 new_gc_thread(0,gc_threads[0]);
876 if (gc_threads != NULL) {
877 #if defined(THREADED_RTS)
879 for (i = 0; i < n_capabilities; i++) {
880 for (g = 0; g < RtsFlags.GcFlags.generations; g++)
882 freeWSDeque(gc_threads[i]->gens[g].todo_q);
884 stgFree (gc_threads[i]);
886 stgFree (gc_threads);
888 for (g = 0; g < RtsFlags.GcFlags.generations; g++)
890 freeWSDeque(gc_threads[0]->gens[g].todo_q);
892 stgFree (gc_threads);
898 /* ----------------------------------------------------------------------------
900 ------------------------------------------------------------------------- */
902 static volatile StgWord gc_running_threads;
908 new = atomic_inc(&gc_running_threads);
909 ASSERT(new <= n_gc_threads);
916 ASSERT(gc_running_threads != 0);
917 return atomic_dec(&gc_running_threads);
930 // scavenge objects in compacted generation
931 if (mark_stack_bd != NULL && !mark_stack_empty()) {
935 // Check for global work in any step. We don't need to check for
936 // local work, because we have already exited scavenge_loop(),
937 // which means there is no local work for this thread.
938 for (g = 0; g < (int)RtsFlags.GcFlags.generations; g++) {
940 if (ws->todo_large_objects) return rtsTrue;
941 if (!looksEmptyWSDeque(ws->todo_q)) return rtsTrue;
942 if (ws->todo_overflow) return rtsTrue;
945 #if defined(THREADED_RTS)
948 // look for work to steal
949 for (n = 0; n < n_gc_threads; n++) {
950 if (n == gct->thread_index) continue;
951 for (g = RtsFlags.GcFlags.generations-1; g >= 0; g--) {
952 ws = &gc_threads[n]->gens[g];
953 if (!looksEmptyWSDeque(ws->todo_q)) return rtsTrue;
960 #if defined(THREADED_RTS)
968 scavenge_until_all_done (void)
974 #if defined(THREADED_RTS)
975 if (n_gc_threads > 1) {
984 collect_gct_blocks();
986 // scavenge_loop() only exits when there's no work to do
989 traceEventGcIdle(gct->cap);
991 debugTrace(DEBUG_gc, "%d GC threads still running", r);
993 while (gc_running_threads != 0) {
997 traceEventGcWork(gct->cap);
1000 // any_work() does not remove the work from the queue, it
1001 // just checks for the presence of work. If we find any,
1002 // then we increment gc_running_threads and go back to
1003 // scavenge_loop() to perform any pending work.
1006 traceEventGcDone(gct->cap);
1009 #if defined(THREADED_RTS)
1012 gcWorkerThread (Capability *cap)
1014 gc_thread *saved_gct;
1016 // necessary if we stole a callee-saves register for gct:
1019 gct = gc_threads[cap->no];
1020 gct->id = osThreadId();
1022 stat_gcWorkerThreadStart(gct);
1024 // Wait until we're told to wake up
1025 RELEASE_SPIN_LOCK(&gct->mut_spin);
1026 gct->wakeup = GC_THREAD_STANDING_BY;
1027 debugTrace(DEBUG_gc, "GC thread %d standing by...", gct->thread_index);
1028 ACQUIRE_SPIN_LOCK(&gct->gc_spin);
1031 // start performance counters in this thread...
1032 if (gct->papi_events == -1) {
1033 papi_init_eventset(&gct->papi_events);
1035 papi_thread_start_gc1_count(gct->papi_events);
1038 init_gc_thread(gct);
1040 traceEventGcWork(gct->cap);
1042 // Every thread evacuates some roots.
1043 gct->evac_gen_no = 0;
1044 markCapability(mark_root, gct, cap, rtsTrue/*prune sparks*/);
1045 scavenge_capability_mut_lists(cap);
1047 scavenge_until_all_done();
1050 // Now that the whole heap is marked, we discard any sparks that
1051 // were found to be unreachable. The main GC thread is currently
1052 // marking heap reachable via weak pointers, so it is
1053 // non-deterministic whether a spark will be retained if it is
1054 // only reachable via weak pointers. To fix this problem would
1055 // require another GC barrier, which is too high a price.
1056 pruneSparkQueue(cap);
1060 // count events in this thread towards the GC totals
1061 papi_thread_stop_gc1_count(gct->papi_events);
1064 // Wait until we're told to continue
1065 RELEASE_SPIN_LOCK(&gct->gc_spin);
1066 gct->wakeup = GC_THREAD_WAITING_TO_CONTINUE;
1067 debugTrace(DEBUG_gc, "GC thread %d waiting to continue...",
1069 ACQUIRE_SPIN_LOCK(&gct->mut_spin);
1070 debugTrace(DEBUG_gc, "GC thread %d on my way...", gct->thread_index);
1072 // record the time spent doing GC in the Task structure
1073 stat_gcWorkerThreadDone(gct);
1080 #if defined(THREADED_RTS)
1083 waitForGcThreads (Capability *cap USED_IF_THREADS)
1085 const nat n_threads = RtsFlags.ParFlags.nNodes;
1086 const nat me = cap->no;
1088 rtsBool retry = rtsTrue;
1091 for (i=0; i < n_threads; i++) {
1092 if (i == me) continue;
1093 if (gc_threads[i]->wakeup != GC_THREAD_STANDING_BY) {
1094 prodCapability(&capabilities[i], cap->running_task);
1097 for (j=0; j < 10; j++) {
1099 for (i=0; i < n_threads; i++) {
1100 if (i == me) continue;
1102 setContextSwitches();
1103 if (gc_threads[i]->wakeup != GC_THREAD_STANDING_BY) {
1113 #endif // THREADED_RTS
1116 start_gc_threads (void)
1118 #if defined(THREADED_RTS)
1119 gc_running_threads = 0;
1124 wakeup_gc_threads (nat me USED_IF_THREADS)
1126 #if defined(THREADED_RTS)
1129 if (n_gc_threads == 1) return;
1131 for (i=0; i < n_gc_threads; i++) {
1132 if (i == me) continue;
1134 debugTrace(DEBUG_gc, "waking up gc thread %d", i);
1135 if (gc_threads[i]->wakeup != GC_THREAD_STANDING_BY) barf("wakeup_gc_threads");
1137 gc_threads[i]->wakeup = GC_THREAD_RUNNING;
1138 ACQUIRE_SPIN_LOCK(&gc_threads[i]->mut_spin);
1139 RELEASE_SPIN_LOCK(&gc_threads[i]->gc_spin);
1144 // After GC is complete, we must wait for all GC threads to enter the
1145 // standby state, otherwise they may still be executing inside
1146 // any_work(), and may even remain awake until the next GC starts.
1148 shutdown_gc_threads (nat me USED_IF_THREADS)
1150 #if defined(THREADED_RTS)
1153 if (n_gc_threads == 1) return;
1155 for (i=0; i < n_gc_threads; i++) {
1156 if (i == me) continue;
1157 while (gc_threads[i]->wakeup != GC_THREAD_WAITING_TO_CONTINUE) { write_barrier(); }
1162 #if defined(THREADED_RTS)
1164 releaseGCThreads (Capability *cap USED_IF_THREADS)
1166 const nat n_threads = RtsFlags.ParFlags.nNodes;
1167 const nat me = cap->no;
1169 for (i=0; i < n_threads; i++) {
1170 if (i == me) continue;
1171 if (gc_threads[i]->wakeup != GC_THREAD_WAITING_TO_CONTINUE)
1172 barf("releaseGCThreads");
1174 gc_threads[i]->wakeup = GC_THREAD_INACTIVE;
1175 ACQUIRE_SPIN_LOCK(&gc_threads[i]->gc_spin);
1176 RELEASE_SPIN_LOCK(&gc_threads[i]->mut_spin);
1181 /* ----------------------------------------------------------------------------
1182 Initialise a generation that is to be collected
1183 ------------------------------------------------------------------------- */
1186 prepare_collected_gen (generation *gen)
1192 // Throw away the current mutable list. Invariant: the mutable
1193 // list always has at least one block; this means we can avoid a
1194 // check for NULL in recordMutable().
1197 for (i = 0; i < n_capabilities; i++) {
1198 freeChain(capabilities[i].mut_lists[g]);
1199 capabilities[i].mut_lists[g] = allocBlock();
1203 gen = &generations[g];
1204 ASSERT(gen->no == g);
1206 // we'll construct a new list of threads in this step
1207 // during GC, throw away the current list.
1208 gen->old_threads = gen->threads;
1209 gen->threads = END_TSO_QUEUE;
1211 // deprecate the existing blocks
1212 gen->old_blocks = gen->blocks;
1213 gen->n_old_blocks = gen->n_blocks;
1217 gen->live_estimate = 0;
1219 // initialise the large object queues.
1220 ASSERT(gen->scavenged_large_objects == NULL);
1221 ASSERT(gen->n_scavenged_large_blocks == 0);
1223 // grab all the partial blocks stashed in the gc_thread workspaces and
1224 // move them to the old_blocks list of this gen.
1225 for (n = 0; n < n_capabilities; n++) {
1226 ws = &gc_threads[n]->gens[gen->no];
1228 for (bd = ws->part_list; bd != NULL; bd = next) {
1230 bd->link = gen->old_blocks;
1231 gen->old_blocks = bd;
1232 gen->n_old_blocks += bd->blocks;
1234 ws->part_list = NULL;
1235 ws->n_part_blocks = 0;
1237 ASSERT(ws->scavd_list == NULL);
1238 ASSERT(ws->n_scavd_blocks == 0);
1240 if (ws->todo_free != ws->todo_bd->start) {
1241 ws->todo_bd->free = ws->todo_free;
1242 ws->todo_bd->link = gen->old_blocks;
1243 gen->old_blocks = ws->todo_bd;
1244 gen->n_old_blocks += ws->todo_bd->blocks;
1245 alloc_todo_block(ws,0); // always has one block.
1249 // mark the small objects as from-space
1250 for (bd = gen->old_blocks; bd; bd = bd->link) {
1251 bd->flags &= ~BF_EVACUATED;
1254 // mark the large objects as from-space
1255 for (bd = gen->large_objects; bd; bd = bd->link) {
1256 bd->flags &= ~BF_EVACUATED;
1259 // for a compacted generation, we need to allocate the bitmap
1261 nat bitmap_size; // in bytes
1262 bdescr *bitmap_bdescr;
1265 bitmap_size = gen->n_old_blocks * BLOCK_SIZE / (sizeof(W_)*BITS_PER_BYTE);
1267 if (bitmap_size > 0) {
1268 bitmap_bdescr = allocGroup((lnat)BLOCK_ROUND_UP(bitmap_size)
1270 gen->bitmap = bitmap_bdescr;
1271 bitmap = bitmap_bdescr->start;
1273 debugTrace(DEBUG_gc, "bitmap_size: %d, bitmap: %p",
1274 bitmap_size, bitmap);
1276 // don't forget to fill it with zeros!
1277 memset(bitmap, 0, bitmap_size);
1279 // For each block in this step, point to its bitmap from the
1280 // block descriptor.
1281 for (bd=gen->old_blocks; bd != NULL; bd = bd->link) {
1282 bd->u.bitmap = bitmap;
1283 bitmap += BLOCK_SIZE_W / (sizeof(W_)*BITS_PER_BYTE);
1285 // Also at this point we set the BF_MARKED flag
1286 // for this block. The invariant is that
1287 // BF_MARKED is always unset, except during GC
1288 // when it is set on those blocks which will be
1290 if (!(bd->flags & BF_FRAGMENTED)) {
1291 bd->flags |= BF_MARKED;
1294 // BF_SWEPT should be marked only for blocks that are being
1295 // collected in sweep()
1296 bd->flags &= ~BF_SWEPT;
1303 /* ----------------------------------------------------------------------------
1304 Save the mutable lists in saved_mut_lists
1305 ------------------------------------------------------------------------- */
1308 stash_mut_list (Capability *cap, nat gen_no)
1310 cap->saved_mut_lists[gen_no] = cap->mut_lists[gen_no];
1311 cap->mut_lists[gen_no] = allocBlock_sync();
1314 /* ----------------------------------------------------------------------------
1315 Initialise a generation that is *not* to be collected
1316 ------------------------------------------------------------------------- */
1319 prepare_uncollected_gen (generation *gen)
1324 ASSERT(gen->no > 0);
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 for (i = 0; i < n_capabilities; i++) {
1330 stash_mut_list(&capabilities[i], gen->no);
1333 ASSERT(gen->scavenged_large_objects == NULL);
1334 ASSERT(gen->n_scavenged_large_blocks == 0);
1337 /* -----------------------------------------------------------------------------
1338 Collect the completed blocks from a GC thread and attach them to
1340 -------------------------------------------------------------------------- */
1343 collect_gct_blocks (void)
1349 for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
1352 // there may still be a block attached to ws->todo_bd;
1353 // leave it there to use next time.
1355 if (ws->scavd_list != NULL) {
1356 ACQUIRE_SPIN_LOCK(&ws->gen->sync);
1358 ASSERT(gct->scan_bd == NULL);
1359 ASSERT(countBlocks(ws->scavd_list) == ws->n_scavd_blocks);
1362 for (bd = ws->scavd_list; bd != NULL; bd = bd->link) {
1363 ws->gen->n_words += bd->free - bd->start;
1367 prev->link = ws->gen->blocks;
1368 ws->gen->blocks = ws->scavd_list;
1370 ws->gen->n_blocks += ws->n_scavd_blocks;
1372 ws->scavd_list = NULL;
1373 ws->n_scavd_blocks = 0;
1375 RELEASE_SPIN_LOCK(&ws->gen->sync);
1380 /* -----------------------------------------------------------------------------
1381 Initialise a gc_thread before GC
1382 -------------------------------------------------------------------------- */
1385 init_gc_thread (gc_thread *t)
1387 t->static_objects = END_OF_STATIC_LIST;
1388 t->scavenged_static_objects = END_OF_STATIC_LIST;
1390 t->mut_lists = t->cap->mut_lists;
1392 t->failed_to_evac = rtsFalse;
1393 t->eager_promotion = rtsTrue;
1394 t->thunk_selector_depth = 0;
1399 t->scav_find_work = 0;
1402 /* -----------------------------------------------------------------------------
1403 Function we pass to evacuate roots.
1404 -------------------------------------------------------------------------- */
1407 mark_root(void *user USED_IF_THREADS, StgClosure **root)
1409 // we stole a register for gct, but this function is called from
1410 // *outside* the GC where the register variable is not in effect,
1411 // so we need to save and restore it here. NB. only call
1412 // mark_root() from the main GC thread, otherwise gct will be
1414 gc_thread *saved_gct;
1423 /* -----------------------------------------------------------------------------
1424 Initialising the static object & mutable lists
1425 -------------------------------------------------------------------------- */
1428 zero_static_object_list(StgClosure* first_static)
1432 const StgInfoTable *info;
1434 for (p = first_static; p != END_OF_STATIC_LIST; p = link) {
1436 link = *STATIC_LINK(info, p);
1437 *STATIC_LINK(info,p) = NULL;
1441 /* ----------------------------------------------------------------------------
1442 Reset the sizes of the older generations when we do a major
1445 CURRENT STRATEGY: make all generations except zero the same size.
1446 We have to stay within the maximum heap size, and leave a certain
1447 percentage of the maximum heap size available to allocate into.
1448 ------------------------------------------------------------------------- */
1451 resize_generations (void)
1455 if (major_gc && RtsFlags.GcFlags.generations > 1) {
1456 nat live, size, min_alloc, words;
1457 const nat max = RtsFlags.GcFlags.maxHeapSize;
1458 const nat gens = RtsFlags.GcFlags.generations;
1460 // live in the oldest generations
1461 if (oldest_gen->live_estimate != 0) {
1462 words = oldest_gen->live_estimate;
1464 words = oldest_gen->n_words;
1466 live = (words + BLOCK_SIZE_W - 1) / BLOCK_SIZE_W +
1467 oldest_gen->n_large_blocks;
1469 // default max size for all generations except zero
1470 size = stg_max(live * RtsFlags.GcFlags.oldGenFactor,
1471 RtsFlags.GcFlags.minOldGenSize);
1473 if (RtsFlags.GcFlags.heapSizeSuggestionAuto) {
1474 RtsFlags.GcFlags.heapSizeSuggestion = size;
1477 // minimum size for generation zero
1478 min_alloc = stg_max((RtsFlags.GcFlags.pcFreeHeap * max) / 200,
1479 RtsFlags.GcFlags.minAllocAreaSize);
1481 // Auto-enable compaction when the residency reaches a
1482 // certain percentage of the maximum heap size (default: 30%).
1483 if (RtsFlags.GcFlags.compact ||
1485 oldest_gen->n_blocks >
1486 (RtsFlags.GcFlags.compactThreshold * max) / 100)) {
1487 oldest_gen->mark = 1;
1488 oldest_gen->compact = 1;
1489 // debugBelch("compaction: on\n", live);
1491 oldest_gen->mark = 0;
1492 oldest_gen->compact = 0;
1493 // debugBelch("compaction: off\n", live);
1496 if (RtsFlags.GcFlags.sweep) {
1497 oldest_gen->mark = 1;
1500 // if we're going to go over the maximum heap size, reduce the
1501 // size of the generations accordingly. The calculation is
1502 // different if compaction is turned on, because we don't need
1503 // to double the space required to collect the old generation.
1506 // this test is necessary to ensure that the calculations
1507 // below don't have any negative results - we're working
1508 // with unsigned values here.
1509 if (max < min_alloc) {
1513 if (oldest_gen->compact) {
1514 if ( (size + (size - 1) * (gens - 2) * 2) + min_alloc > max ) {
1515 size = (max - min_alloc) / ((gens - 1) * 2 - 1);
1518 if ( (size * (gens - 1) * 2) + min_alloc > max ) {
1519 size = (max - min_alloc) / ((gens - 1) * 2);
1529 debugBelch("live: %d, min_alloc: %d, size : %d, max = %d\n", live,
1530 min_alloc, size, max);
1533 for (g = 0; g < gens; g++) {
1534 generations[g].max_blocks = size;
1539 /* -----------------------------------------------------------------------------
1540 Calculate the new size of the nursery, and resize it.
1541 -------------------------------------------------------------------------- */
1544 resize_nursery (void)
1546 const lnat min_nursery = RtsFlags.GcFlags.minAllocAreaSize * n_capabilities;
1548 if (RtsFlags.GcFlags.generations == 1)
1549 { // Two-space collector:
1552 /* set up a new nursery. Allocate a nursery size based on a
1553 * function of the amount of live data (by default a factor of 2)
1554 * Use the blocks from the old nursery if possible, freeing up any
1557 * If we get near the maximum heap size, then adjust our nursery
1558 * size accordingly. If the nursery is the same size as the live
1559 * data (L), then we need 3L bytes. We can reduce the size of the
1560 * nursery to bring the required memory down near 2L bytes.
1562 * A normal 2-space collector would need 4L bytes to give the same
1563 * performance we get from 3L bytes, reducing to the same
1564 * performance at 2L bytes.
1566 blocks = generations[0].n_blocks;
1568 if ( RtsFlags.GcFlags.maxHeapSize != 0 &&
1569 blocks * RtsFlags.GcFlags.oldGenFactor * 2 >
1570 RtsFlags.GcFlags.maxHeapSize )
1572 long adjusted_blocks; // signed on purpose
1575 adjusted_blocks = (RtsFlags.GcFlags.maxHeapSize - 2 * blocks);
1577 debugTrace(DEBUG_gc, "near maximum heap size of 0x%x blocks, blocks = %d, adjusted to %ld",
1578 RtsFlags.GcFlags.maxHeapSize, blocks, adjusted_blocks);
1580 pc_free = adjusted_blocks * 100 / RtsFlags.GcFlags.maxHeapSize;
1581 if (pc_free < RtsFlags.GcFlags.pcFreeHeap) /* might even * be < 0 */
1585 blocks = adjusted_blocks;
1589 blocks *= RtsFlags.GcFlags.oldGenFactor;
1590 if (blocks < min_nursery)
1592 blocks = min_nursery;
1595 resizeNurseries(blocks);
1597 else // Generational collector
1600 * If the user has given us a suggested heap size, adjust our
1601 * allocation area to make best use of the memory available.
1603 if (RtsFlags.GcFlags.heapSizeSuggestion)
1606 const nat needed = calcNeeded(); // approx blocks needed at next GC
1608 /* Guess how much will be live in generation 0 step 0 next time.
1609 * A good approximation is obtained by finding the
1610 * percentage of g0 that was live at the last minor GC.
1612 * We have an accurate figure for the amount of copied data in
1613 * 'copied', but we must convert this to a number of blocks, with
1614 * a small adjustment for estimated slop at the end of a block
1619 g0_pcnt_kept = ((copied / (BLOCK_SIZE_W - 10)) * 100)
1620 / countNurseryBlocks();
1623 /* Estimate a size for the allocation area based on the
1624 * information available. We might end up going slightly under
1625 * or over the suggested heap size, but we should be pretty
1628 * Formula: suggested - needed
1629 * ----------------------------
1630 * 1 + g0_pcnt_kept/100
1632 * where 'needed' is the amount of memory needed at the next
1633 * collection for collecting all gens except g0.
1636 (((long)RtsFlags.GcFlags.heapSizeSuggestion - (long)needed) * 100) /
1637 (100 + (long)g0_pcnt_kept);
1639 if (blocks < (long)min_nursery) {
1640 blocks = min_nursery;
1643 resizeNurseries((nat)blocks);
1647 // we might have added extra large blocks to the nursery, so
1648 // resize back to minAllocAreaSize again.
1649 resizeNurseriesFixed(RtsFlags.GcFlags.minAllocAreaSize);
1654 /* -----------------------------------------------------------------------------
1655 Sanity code for CAF garbage collection.
1657 With DEBUG turned on, we manage a CAF list in addition to the SRT
1658 mechanism. After GC, we run down the CAF list and blackhole any
1659 CAFs which have been garbage collected. This means we get an error
1660 whenever the program tries to enter a garbage collected CAF.
1662 Any garbage collected CAFs are taken off the CAF list at the same
1664 -------------------------------------------------------------------------- */
1666 #if 0 && defined(DEBUG)
1673 const StgInfoTable *info;
1684 ASSERT(info->type == IND_STATIC);
1686 if (STATIC_LINK(info,p) == NULL) {
1687 debugTrace(DEBUG_gccafs, "CAF gc'd at 0x%04lx", (long)p);
1689 SET_INFO(p,&stg_BLACKHOLE_info);
1690 p = STATIC_LINK2(info,p);
1694 pp = &STATIC_LINK2(info,p);
1701 debugTrace(DEBUG_gccafs, "%d CAFs live", i);