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 if (n_gc_threads == 1) {
647 zero_static_object_list(gct->scavenged_static_objects);
649 for (i = 0; i < n_gc_threads; i++) {
650 zero_static_object_list(gc_threads[i]->scavenged_static_objects);
655 // Update the stable pointer hash table.
656 updateStablePtrTable(major_gc);
658 // unlock the StablePtr table. Must be before scheduleFinalizers(),
659 // because a finalizer may call hs_free_fun_ptr() or
660 // hs_free_stable_ptr(), both of which access the StablePtr table.
663 // Start any pending finalizers. Must be after
664 // updateStablePtrTable() and stablePtrPostGC() (see #4221).
666 scheduleFinalizers(cap, old_weak_ptr_list);
669 // check sanity after GC
670 // before resurrectThreads(), because that might overwrite some
671 // closures, which will cause problems with THREADED where we don't
673 IF_DEBUG(sanity, checkSanity(rtsTrue /* after GC */, major_gc));
675 // send exceptions to any threads which were about to die
677 resurrectThreads(resurrected_threads);
682 need = BLOCKS_TO_MBLOCKS(n_alloc_blocks);
683 got = mblocks_allocated;
684 /* If the amount of data remains constant, next major GC we'll
685 require (F+1)*need. We leave (F+2)*need in order to reduce
686 repeated deallocation and reallocation. */
687 need = (RtsFlags.GcFlags.oldGenFactor + 2) * need;
689 returnMemoryToOS(got - need);
693 // extra GC trace info
694 IF_DEBUG(gc, statDescribeGens());
697 // symbol-table based profiling
698 /* heapCensus(to_blocks); */ /* ToDo */
701 // restore enclosing cost centre
707 // check for memory leaks if DEBUG is on
708 memInventory(DEBUG_gc);
711 #ifdef RTS_GTK_FRONTPANEL
712 if (RtsFlags.GcFlags.frontpanel) {
713 updateFrontPanelAfterGC( N, live );
717 // ok, GC over: tell the stats department what happened.
718 stat_endGC(gct, allocated, live_words,
719 copied, N, max_copied, avg_copied,
720 live_blocks * BLOCK_SIZE_W - live_words /* slop */);
722 // Guess which generation we'll collect *next* time
723 initialise_N(force_major_gc);
725 #if defined(RTS_USER_SIGNALS)
726 if (RtsFlags.MiscFlags.install_signal_handlers) {
727 // unblock signals again
728 unblockUserSignals();
737 /* -----------------------------------------------------------------------------
738 Figure out which generation to collect, initialise N and major_gc.
740 Also returns the total number of blocks in generations that will be
742 -------------------------------------------------------------------------- */
745 initialise_N (rtsBool force_major_gc)
748 nat blocks, blocks_total;
753 if (force_major_gc) {
754 N = RtsFlags.GcFlags.generations - 1;
759 for (g = RtsFlags.GcFlags.generations - 1; g >= 0; g--) {
761 blocks = generations[g].n_words / BLOCK_SIZE_W
762 + generations[g].n_large_blocks;
764 if (blocks >= generations[g].max_blocks) {
768 blocks_total += blocks;
772 blocks_total += countNurseryBlocks();
774 major_gc = (N == RtsFlags.GcFlags.generations-1);
778 /* -----------------------------------------------------------------------------
779 Initialise the gc_thread structures.
780 -------------------------------------------------------------------------- */
782 #define GC_THREAD_INACTIVE 0
783 #define GC_THREAD_STANDING_BY 1
784 #define GC_THREAD_RUNNING 2
785 #define GC_THREAD_WAITING_TO_CONTINUE 3
788 new_gc_thread (nat n, gc_thread *t)
793 t->cap = &capabilities[n];
797 initSpinLock(&t->gc_spin);
798 initSpinLock(&t->mut_spin);
799 ACQUIRE_SPIN_LOCK(&t->gc_spin);
800 t->wakeup = GC_THREAD_INACTIVE; // starts true, so we can wait for the
801 // thread to start up, see wakeup_gc_threads
805 t->free_blocks = NULL;
814 for (g = 0; g < RtsFlags.GcFlags.generations; g++)
817 ws->gen = &generations[g];
818 ASSERT(g == ws->gen->no);
822 // alloc_todo_block(ws,0);
823 // but can't, because it uses gct which isn't set up at this point.
824 // Hence, allocate a block for todo_bd manually:
826 bdescr *bd = allocBlock(); // no lock, locks aren't initialised yet
827 initBdescr(bd, ws->gen, ws->gen->to);
828 bd->flags = BF_EVACUATED;
829 bd->u.scan = bd->free = bd->start;
832 ws->todo_free = bd->free;
833 ws->todo_lim = bd->start + BLOCK_SIZE_W;
836 ws->todo_q = newWSDeque(128);
837 ws->todo_overflow = NULL;
838 ws->n_todo_overflow = 0;
839 ws->todo_large_objects = NULL;
841 ws->part_list = NULL;
842 ws->n_part_blocks = 0;
844 ws->scavd_list = NULL;
845 ws->n_scavd_blocks = 0;
853 if (gc_threads == NULL) {
854 #if defined(THREADED_RTS)
856 gc_threads = stgMallocBytes (RtsFlags.ParFlags.nNodes *
860 for (i = 0; i < RtsFlags.ParFlags.nNodes; i++) {
862 stgMallocBytes(sizeof(gc_thread) +
863 RtsFlags.GcFlags.generations * sizeof(gen_workspace),
866 new_gc_thread(i, gc_threads[i]);
869 gc_threads = stgMallocBytes (sizeof(gc_thread*),"alloc_gc_threads");
871 new_gc_thread(0,gc_threads[0]);
880 if (gc_threads != NULL) {
881 #if defined(THREADED_RTS)
883 for (i = 0; i < n_capabilities; i++) {
884 for (g = 0; g < RtsFlags.GcFlags.generations; g++)
886 freeWSDeque(gc_threads[i]->gens[g].todo_q);
888 stgFree (gc_threads[i]);
890 stgFree (gc_threads);
892 for (g = 0; g < RtsFlags.GcFlags.generations; g++)
894 freeWSDeque(gc_threads[0]->gens[g].todo_q);
896 stgFree (gc_threads);
902 /* ----------------------------------------------------------------------------
904 ------------------------------------------------------------------------- */
906 static volatile StgWord gc_running_threads;
912 new = atomic_inc(&gc_running_threads);
913 ASSERT(new <= n_gc_threads);
920 ASSERT(gc_running_threads != 0);
921 return atomic_dec(&gc_running_threads);
934 // scavenge objects in compacted generation
935 if (mark_stack_bd != NULL && !mark_stack_empty()) {
939 // Check for global work in any step. We don't need to check for
940 // local work, because we have already exited scavenge_loop(),
941 // which means there is no local work for this thread.
942 for (g = 0; g < (int)RtsFlags.GcFlags.generations; g++) {
944 if (ws->todo_large_objects) return rtsTrue;
945 if (!looksEmptyWSDeque(ws->todo_q)) return rtsTrue;
946 if (ws->todo_overflow) return rtsTrue;
949 #if defined(THREADED_RTS)
952 // look for work to steal
953 for (n = 0; n < n_gc_threads; n++) {
954 if (n == gct->thread_index) continue;
955 for (g = RtsFlags.GcFlags.generations-1; g >= 0; g--) {
956 ws = &gc_threads[n]->gens[g];
957 if (!looksEmptyWSDeque(ws->todo_q)) return rtsTrue;
964 #if defined(THREADED_RTS)
972 scavenge_until_all_done (void)
978 #if defined(THREADED_RTS)
979 if (n_gc_threads > 1) {
988 collect_gct_blocks();
990 // scavenge_loop() only exits when there's no work to do
993 traceEventGcIdle(gct->cap);
995 debugTrace(DEBUG_gc, "%d GC threads still running", r);
997 while (gc_running_threads != 0) {
1001 traceEventGcWork(gct->cap);
1004 // any_work() does not remove the work from the queue, it
1005 // just checks for the presence of work. If we find any,
1006 // then we increment gc_running_threads and go back to
1007 // scavenge_loop() to perform any pending work.
1010 traceEventGcDone(gct->cap);
1013 #if defined(THREADED_RTS)
1016 gcWorkerThread (Capability *cap)
1018 gc_thread *saved_gct;
1020 // necessary if we stole a callee-saves register for gct:
1023 gct = gc_threads[cap->no];
1024 gct->id = osThreadId();
1026 stat_gcWorkerThreadStart(gct);
1028 // Wait until we're told to wake up
1029 RELEASE_SPIN_LOCK(&gct->mut_spin);
1030 gct->wakeup = GC_THREAD_STANDING_BY;
1031 debugTrace(DEBUG_gc, "GC thread %d standing by...", gct->thread_index);
1032 ACQUIRE_SPIN_LOCK(&gct->gc_spin);
1035 // start performance counters in this thread...
1036 if (gct->papi_events == -1) {
1037 papi_init_eventset(&gct->papi_events);
1039 papi_thread_start_gc1_count(gct->papi_events);
1042 init_gc_thread(gct);
1044 traceEventGcWork(gct->cap);
1046 // Every thread evacuates some roots.
1047 gct->evac_gen_no = 0;
1048 markCapability(mark_root, gct, cap, rtsTrue/*prune sparks*/);
1049 scavenge_capability_mut_lists(cap);
1051 scavenge_until_all_done();
1054 // Now that the whole heap is marked, we discard any sparks that
1055 // were found to be unreachable. The main GC thread is currently
1056 // marking heap reachable via weak pointers, so it is
1057 // non-deterministic whether a spark will be retained if it is
1058 // only reachable via weak pointers. To fix this problem would
1059 // require another GC barrier, which is too high a price.
1060 pruneSparkQueue(cap);
1064 // count events in this thread towards the GC totals
1065 papi_thread_stop_gc1_count(gct->papi_events);
1068 // Wait until we're told to continue
1069 RELEASE_SPIN_LOCK(&gct->gc_spin);
1070 gct->wakeup = GC_THREAD_WAITING_TO_CONTINUE;
1071 debugTrace(DEBUG_gc, "GC thread %d waiting to continue...",
1073 ACQUIRE_SPIN_LOCK(&gct->mut_spin);
1074 debugTrace(DEBUG_gc, "GC thread %d on my way...", gct->thread_index);
1076 // record the time spent doing GC in the Task structure
1077 stat_gcWorkerThreadDone(gct);
1084 #if defined(THREADED_RTS)
1087 waitForGcThreads (Capability *cap USED_IF_THREADS)
1089 const nat n_threads = RtsFlags.ParFlags.nNodes;
1090 const nat me = cap->no;
1092 rtsBool retry = rtsTrue;
1095 for (i=0; i < n_threads; i++) {
1096 if (i == me) continue;
1097 if (gc_threads[i]->wakeup != GC_THREAD_STANDING_BY) {
1098 prodCapability(&capabilities[i], cap->running_task);
1101 for (j=0; j < 10; j++) {
1103 for (i=0; i < n_threads; i++) {
1104 if (i == me) continue;
1106 setContextSwitches();
1107 if (gc_threads[i]->wakeup != GC_THREAD_STANDING_BY) {
1117 #endif // THREADED_RTS
1120 start_gc_threads (void)
1122 #if defined(THREADED_RTS)
1123 gc_running_threads = 0;
1128 wakeup_gc_threads (nat me USED_IF_THREADS)
1130 #if defined(THREADED_RTS)
1133 if (n_gc_threads == 1) return;
1135 for (i=0; i < n_gc_threads; i++) {
1136 if (i == me) continue;
1138 debugTrace(DEBUG_gc, "waking up gc thread %d", i);
1139 if (gc_threads[i]->wakeup != GC_THREAD_STANDING_BY) barf("wakeup_gc_threads");
1141 gc_threads[i]->wakeup = GC_THREAD_RUNNING;
1142 ACQUIRE_SPIN_LOCK(&gc_threads[i]->mut_spin);
1143 RELEASE_SPIN_LOCK(&gc_threads[i]->gc_spin);
1148 // After GC is complete, we must wait for all GC threads to enter the
1149 // standby state, otherwise they may still be executing inside
1150 // any_work(), and may even remain awake until the next GC starts.
1152 shutdown_gc_threads (nat me USED_IF_THREADS)
1154 #if defined(THREADED_RTS)
1157 if (n_gc_threads == 1) return;
1159 for (i=0; i < n_gc_threads; i++) {
1160 if (i == me) continue;
1161 while (gc_threads[i]->wakeup != GC_THREAD_WAITING_TO_CONTINUE) { write_barrier(); }
1166 #if defined(THREADED_RTS)
1168 releaseGCThreads (Capability *cap USED_IF_THREADS)
1170 const nat n_threads = RtsFlags.ParFlags.nNodes;
1171 const nat me = cap->no;
1173 for (i=0; i < n_threads; i++) {
1174 if (i == me) continue;
1175 if (gc_threads[i]->wakeup != GC_THREAD_WAITING_TO_CONTINUE)
1176 barf("releaseGCThreads");
1178 gc_threads[i]->wakeup = GC_THREAD_INACTIVE;
1179 ACQUIRE_SPIN_LOCK(&gc_threads[i]->gc_spin);
1180 RELEASE_SPIN_LOCK(&gc_threads[i]->mut_spin);
1185 /* ----------------------------------------------------------------------------
1186 Initialise a generation that is to be collected
1187 ------------------------------------------------------------------------- */
1190 prepare_collected_gen (generation *gen)
1196 // Throw away the current mutable list. Invariant: the mutable
1197 // list always has at least one block; this means we can avoid a
1198 // check for NULL in recordMutable().
1201 for (i = 0; i < n_capabilities; i++) {
1202 freeChain(capabilities[i].mut_lists[g]);
1203 capabilities[i].mut_lists[g] = allocBlock();
1207 gen = &generations[g];
1208 ASSERT(gen->no == g);
1210 // we'll construct a new list of threads in this step
1211 // during GC, throw away the current list.
1212 gen->old_threads = gen->threads;
1213 gen->threads = END_TSO_QUEUE;
1215 // deprecate the existing blocks
1216 gen->old_blocks = gen->blocks;
1217 gen->n_old_blocks = gen->n_blocks;
1221 gen->live_estimate = 0;
1223 // initialise the large object queues.
1224 ASSERT(gen->scavenged_large_objects == NULL);
1225 ASSERT(gen->n_scavenged_large_blocks == 0);
1227 // grab all the partial blocks stashed in the gc_thread workspaces and
1228 // move them to the old_blocks list of this gen.
1229 for (n = 0; n < n_capabilities; n++) {
1230 ws = &gc_threads[n]->gens[gen->no];
1232 for (bd = ws->part_list; bd != NULL; bd = next) {
1234 bd->link = gen->old_blocks;
1235 gen->old_blocks = bd;
1236 gen->n_old_blocks += bd->blocks;
1238 ws->part_list = NULL;
1239 ws->n_part_blocks = 0;
1241 ASSERT(ws->scavd_list == NULL);
1242 ASSERT(ws->n_scavd_blocks == 0);
1244 if (ws->todo_free != ws->todo_bd->start) {
1245 ws->todo_bd->free = ws->todo_free;
1246 ws->todo_bd->link = gen->old_blocks;
1247 gen->old_blocks = ws->todo_bd;
1248 gen->n_old_blocks += ws->todo_bd->blocks;
1249 alloc_todo_block(ws,0); // always has one block.
1253 // mark the small objects as from-space
1254 for (bd = gen->old_blocks; bd; bd = bd->link) {
1255 bd->flags &= ~BF_EVACUATED;
1258 // mark the large objects as from-space
1259 for (bd = gen->large_objects; bd; bd = bd->link) {
1260 bd->flags &= ~BF_EVACUATED;
1263 // for a compacted generation, we need to allocate the bitmap
1265 nat bitmap_size; // in bytes
1266 bdescr *bitmap_bdescr;
1269 bitmap_size = gen->n_old_blocks * BLOCK_SIZE / (sizeof(W_)*BITS_PER_BYTE);
1271 if (bitmap_size > 0) {
1272 bitmap_bdescr = allocGroup((lnat)BLOCK_ROUND_UP(bitmap_size)
1274 gen->bitmap = bitmap_bdescr;
1275 bitmap = bitmap_bdescr->start;
1277 debugTrace(DEBUG_gc, "bitmap_size: %d, bitmap: %p",
1278 bitmap_size, bitmap);
1280 // don't forget to fill it with zeros!
1281 memset(bitmap, 0, bitmap_size);
1283 // For each block in this step, point to its bitmap from the
1284 // block descriptor.
1285 for (bd=gen->old_blocks; bd != NULL; bd = bd->link) {
1286 bd->u.bitmap = bitmap;
1287 bitmap += BLOCK_SIZE_W / (sizeof(W_)*BITS_PER_BYTE);
1289 // Also at this point we set the BF_MARKED flag
1290 // for this block. The invariant is that
1291 // BF_MARKED is always unset, except during GC
1292 // when it is set on those blocks which will be
1294 if (!(bd->flags & BF_FRAGMENTED)) {
1295 bd->flags |= BF_MARKED;
1298 // BF_SWEPT should be marked only for blocks that are being
1299 // collected in sweep()
1300 bd->flags &= ~BF_SWEPT;
1307 /* ----------------------------------------------------------------------------
1308 Save the mutable lists in saved_mut_lists
1309 ------------------------------------------------------------------------- */
1312 stash_mut_list (Capability *cap, nat gen_no)
1314 cap->saved_mut_lists[gen_no] = cap->mut_lists[gen_no];
1315 cap->mut_lists[gen_no] = allocBlock_sync();
1318 /* ----------------------------------------------------------------------------
1319 Initialise a generation that is *not* to be collected
1320 ------------------------------------------------------------------------- */
1323 prepare_uncollected_gen (generation *gen)
1328 ASSERT(gen->no > 0);
1330 // save the current mutable lists for this generation, and
1331 // allocate a fresh block for each one. We'll traverse these
1332 // mutable lists as roots early on in the GC.
1333 for (i = 0; i < n_capabilities; i++) {
1334 stash_mut_list(&capabilities[i], gen->no);
1337 ASSERT(gen->scavenged_large_objects == NULL);
1338 ASSERT(gen->n_scavenged_large_blocks == 0);
1341 /* -----------------------------------------------------------------------------
1342 Collect the completed blocks from a GC thread and attach them to
1344 -------------------------------------------------------------------------- */
1347 collect_gct_blocks (void)
1353 for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
1356 // there may still be a block attached to ws->todo_bd;
1357 // leave it there to use next time.
1359 if (ws->scavd_list != NULL) {
1360 ACQUIRE_SPIN_LOCK(&ws->gen->sync);
1362 ASSERT(gct->scan_bd == NULL);
1363 ASSERT(countBlocks(ws->scavd_list) == ws->n_scavd_blocks);
1366 for (bd = ws->scavd_list; bd != NULL; bd = bd->link) {
1367 ws->gen->n_words += bd->free - bd->start;
1371 prev->link = ws->gen->blocks;
1372 ws->gen->blocks = ws->scavd_list;
1374 ws->gen->n_blocks += ws->n_scavd_blocks;
1376 ws->scavd_list = NULL;
1377 ws->n_scavd_blocks = 0;
1379 RELEASE_SPIN_LOCK(&ws->gen->sync);
1384 /* -----------------------------------------------------------------------------
1385 Initialise a gc_thread before GC
1386 -------------------------------------------------------------------------- */
1389 init_gc_thread (gc_thread *t)
1391 t->static_objects = END_OF_STATIC_LIST;
1392 t->scavenged_static_objects = END_OF_STATIC_LIST;
1394 t->mut_lists = t->cap->mut_lists;
1396 t->failed_to_evac = rtsFalse;
1397 t->eager_promotion = rtsTrue;
1398 t->thunk_selector_depth = 0;
1403 t->scav_find_work = 0;
1406 /* -----------------------------------------------------------------------------
1407 Function we pass to evacuate roots.
1408 -------------------------------------------------------------------------- */
1411 mark_root(void *user USED_IF_THREADS, StgClosure **root)
1413 // we stole a register for gct, but this function is called from
1414 // *outside* the GC where the register variable is not in effect,
1415 // so we need to save and restore it here. NB. only call
1416 // mark_root() from the main GC thread, otherwise gct will be
1418 gc_thread *saved_gct;
1427 /* -----------------------------------------------------------------------------
1428 Initialising the static object & mutable lists
1429 -------------------------------------------------------------------------- */
1432 zero_static_object_list(StgClosure* first_static)
1436 const StgInfoTable *info;
1438 for (p = first_static; p != END_OF_STATIC_LIST; p = link) {
1440 link = *STATIC_LINK(info, p);
1441 *STATIC_LINK(info,p) = NULL;
1445 /* ----------------------------------------------------------------------------
1446 Reset the sizes of the older generations when we do a major
1449 CURRENT STRATEGY: make all generations except zero the same size.
1450 We have to stay within the maximum heap size, and leave a certain
1451 percentage of the maximum heap size available to allocate into.
1452 ------------------------------------------------------------------------- */
1455 resize_generations (void)
1459 if (major_gc && RtsFlags.GcFlags.generations > 1) {
1460 nat live, size, min_alloc, words;
1461 const nat max = RtsFlags.GcFlags.maxHeapSize;
1462 const nat gens = RtsFlags.GcFlags.generations;
1464 // live in the oldest generations
1465 if (oldest_gen->live_estimate != 0) {
1466 words = oldest_gen->live_estimate;
1468 words = oldest_gen->n_words;
1470 live = (words + BLOCK_SIZE_W - 1) / BLOCK_SIZE_W +
1471 oldest_gen->n_large_blocks;
1473 // default max size for all generations except zero
1474 size = stg_max(live * RtsFlags.GcFlags.oldGenFactor,
1475 RtsFlags.GcFlags.minOldGenSize);
1477 if (RtsFlags.GcFlags.heapSizeSuggestionAuto) {
1478 RtsFlags.GcFlags.heapSizeSuggestion = size;
1481 // minimum size for generation zero
1482 min_alloc = stg_max((RtsFlags.GcFlags.pcFreeHeap * max) / 200,
1483 RtsFlags.GcFlags.minAllocAreaSize);
1485 // Auto-enable compaction when the residency reaches a
1486 // certain percentage of the maximum heap size (default: 30%).
1487 if (RtsFlags.GcFlags.compact ||
1489 oldest_gen->n_blocks >
1490 (RtsFlags.GcFlags.compactThreshold * max) / 100)) {
1491 oldest_gen->mark = 1;
1492 oldest_gen->compact = 1;
1493 // debugBelch("compaction: on\n", live);
1495 oldest_gen->mark = 0;
1496 oldest_gen->compact = 0;
1497 // debugBelch("compaction: off\n", live);
1500 if (RtsFlags.GcFlags.sweep) {
1501 oldest_gen->mark = 1;
1504 // if we're going to go over the maximum heap size, reduce the
1505 // size of the generations accordingly. The calculation is
1506 // different if compaction is turned on, because we don't need
1507 // to double the space required to collect the old generation.
1510 // this test is necessary to ensure that the calculations
1511 // below don't have any negative results - we're working
1512 // with unsigned values here.
1513 if (max < min_alloc) {
1517 if (oldest_gen->compact) {
1518 if ( (size + (size - 1) * (gens - 2) * 2) + min_alloc > max ) {
1519 size = (max - min_alloc) / ((gens - 1) * 2 - 1);
1522 if ( (size * (gens - 1) * 2) + min_alloc > max ) {
1523 size = (max - min_alloc) / ((gens - 1) * 2);
1533 debugBelch("live: %d, min_alloc: %d, size : %d, max = %d\n", live,
1534 min_alloc, size, max);
1537 for (g = 0; g < gens; g++) {
1538 generations[g].max_blocks = size;
1543 /* -----------------------------------------------------------------------------
1544 Calculate the new size of the nursery, and resize it.
1545 -------------------------------------------------------------------------- */
1548 resize_nursery (void)
1550 const lnat min_nursery = RtsFlags.GcFlags.minAllocAreaSize * n_capabilities;
1552 if (RtsFlags.GcFlags.generations == 1)
1553 { // Two-space collector:
1556 /* set up a new nursery. Allocate a nursery size based on a
1557 * function of the amount of live data (by default a factor of 2)
1558 * Use the blocks from the old nursery if possible, freeing up any
1561 * If we get near the maximum heap size, then adjust our nursery
1562 * size accordingly. If the nursery is the same size as the live
1563 * data (L), then we need 3L bytes. We can reduce the size of the
1564 * nursery to bring the required memory down near 2L bytes.
1566 * A normal 2-space collector would need 4L bytes to give the same
1567 * performance we get from 3L bytes, reducing to the same
1568 * performance at 2L bytes.
1570 blocks = generations[0].n_blocks;
1572 if ( RtsFlags.GcFlags.maxHeapSize != 0 &&
1573 blocks * RtsFlags.GcFlags.oldGenFactor * 2 >
1574 RtsFlags.GcFlags.maxHeapSize )
1576 long adjusted_blocks; // signed on purpose
1579 adjusted_blocks = (RtsFlags.GcFlags.maxHeapSize - 2 * blocks);
1581 debugTrace(DEBUG_gc, "near maximum heap size of 0x%x blocks, blocks = %d, adjusted to %ld",
1582 RtsFlags.GcFlags.maxHeapSize, blocks, adjusted_blocks);
1584 pc_free = adjusted_blocks * 100 / RtsFlags.GcFlags.maxHeapSize;
1585 if (pc_free < RtsFlags.GcFlags.pcFreeHeap) /* might even * be < 0 */
1589 blocks = adjusted_blocks;
1593 blocks *= RtsFlags.GcFlags.oldGenFactor;
1594 if (blocks < min_nursery)
1596 blocks = min_nursery;
1599 resizeNurseries(blocks);
1601 else // Generational collector
1604 * If the user has given us a suggested heap size, adjust our
1605 * allocation area to make best use of the memory available.
1607 if (RtsFlags.GcFlags.heapSizeSuggestion)
1610 const nat needed = calcNeeded(); // approx blocks needed at next GC
1612 /* Guess how much will be live in generation 0 step 0 next time.
1613 * A good approximation is obtained by finding the
1614 * percentage of g0 that was live at the last minor GC.
1616 * We have an accurate figure for the amount of copied data in
1617 * 'copied', but we must convert this to a number of blocks, with
1618 * a small adjustment for estimated slop at the end of a block
1623 g0_pcnt_kept = ((copied / (BLOCK_SIZE_W - 10)) * 100)
1624 / countNurseryBlocks();
1627 /* Estimate a size for the allocation area based on the
1628 * information available. We might end up going slightly under
1629 * or over the suggested heap size, but we should be pretty
1632 * Formula: suggested - needed
1633 * ----------------------------
1634 * 1 + g0_pcnt_kept/100
1636 * where 'needed' is the amount of memory needed at the next
1637 * collection for collecting all gens except g0.
1640 (((long)RtsFlags.GcFlags.heapSizeSuggestion - (long)needed) * 100) /
1641 (100 + (long)g0_pcnt_kept);
1643 if (blocks < (long)min_nursery) {
1644 blocks = min_nursery;
1647 resizeNurseries((nat)blocks);
1651 // we might have added extra large blocks to the nursery, so
1652 // resize back to minAllocAreaSize again.
1653 resizeNurseriesFixed(RtsFlags.GcFlags.minAllocAreaSize);
1658 /* -----------------------------------------------------------------------------
1659 Sanity code for CAF garbage collection.
1661 With DEBUG turned on, we manage a CAF list in addition to the SRT
1662 mechanism. After GC, we run down the CAF list and blackhole any
1663 CAFs which have been garbage collected. This means we get an error
1664 whenever the program tries to enter a garbage collected CAF.
1666 Any garbage collected CAFs are taken off the CAF list at the same
1668 -------------------------------------------------------------------------- */
1670 #if 0 && defined(DEBUG)
1677 const StgInfoTable *info;
1688 ASSERT(info->type == IND_STATIC);
1690 if (STATIC_LINK(info,p) == NULL) {
1691 debugTrace(DEBUG_gccafs, "CAF gc'd at 0x%04lx", (long)p);
1693 SET_INFO(p,&stg_BLACKHOLE_info);
1694 p = STATIC_LINK2(info,p);
1698 pp = &STATIC_LINK2(info,p);
1705 debugTrace(DEBUG_gccafs, "%d CAFs live", i);