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 // Finally: compact or sweep the oldest generation.
412 if (major_gc && oldest_gen->mark) {
413 if (oldest_gen->compact)
414 compact(gct->scavenged_static_objects);
424 for (i=0; i < n_gc_threads; i++) {
425 if (n_gc_threads > 1) {
426 debugTrace(DEBUG_gc,"thread %d:", i);
427 debugTrace(DEBUG_gc," copied %ld", gc_threads[i]->copied * sizeof(W_));
428 debugTrace(DEBUG_gc," scanned %ld", gc_threads[i]->scanned * sizeof(W_));
429 debugTrace(DEBUG_gc," any_work %ld", gc_threads[i]->any_work);
430 debugTrace(DEBUG_gc," no_work %ld", gc_threads[i]->no_work);
431 debugTrace(DEBUG_gc," scav_find_work %ld", gc_threads[i]->scav_find_work);
433 copied += gc_threads[i]->copied;
434 max_copied = stg_max(gc_threads[i]->copied, max_copied);
436 if (n_gc_threads == 1) {
444 // Run through all the generations/steps and tidy up.
446 // - count the amount of "live" data (live_words, live_blocks)
447 // - count the amount of "copied" data in this GC (copied)
449 // - make to-space the new from-space (set BF_EVACUATED on all blocks)
454 for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
457 generations[g].collections++; // for stats
458 if (n_gc_threads > 1) generations[g].par_collections++;
461 // Count the mutable list as bytes "copied" for the purposes of
462 // stats. Every mutable list is copied during every GC.
464 nat mut_list_size = 0;
465 for (n = 0; n < n_capabilities; n++) {
466 mut_list_size += countOccupied(capabilities[n].mut_lists[g]);
468 copied += mut_list_size;
471 "mut_list_size: %lu (%d vars, %d arrays, %d MVARs, %d others)",
472 (unsigned long)(mut_list_size * sizeof(W_)),
473 mutlist_MUTVARS, mutlist_MUTARRS, mutlist_MVARS, mutlist_OTHERS);
477 gen = &generations[g];
479 // for generations we collected...
482 /* free old memory and shift to-space into from-space for all
483 * the collected steps (except the allocation area). These
484 * freed blocks will probaby be quickly recycled.
488 // tack the new blocks on the end of the existing blocks
489 if (gen->old_blocks != NULL) {
492 for (bd = gen->old_blocks; bd != NULL; bd = next) {
496 if (!(bd->flags & BF_MARKED))
499 gen->old_blocks = next;
508 gen->n_words += bd->free - bd->start;
510 // NB. this step might not be compacted next
511 // time, so reset the BF_MARKED flags.
512 // They are set before GC if we're going to
513 // compact. (search for BF_MARKED above).
514 bd->flags &= ~BF_MARKED;
516 // between GCs, all blocks in the heap except
517 // for the nursery have the BF_EVACUATED flag set.
518 bd->flags |= BF_EVACUATED;
525 prev->link = gen->blocks;
526 gen->blocks = gen->old_blocks;
529 // add the new blocks to the block tally
530 gen->n_blocks += gen->n_old_blocks;
531 ASSERT(countBlocks(gen->blocks) == gen->n_blocks);
532 ASSERT(countOccupied(gen->blocks) == gen->n_words);
536 freeChain(gen->old_blocks);
539 gen->old_blocks = NULL;
540 gen->n_old_blocks = 0;
542 /* LARGE OBJECTS. The current live large objects are chained on
543 * scavenged_large, having been moved during garbage
544 * collection from large_objects. Any objects left on the
545 * large_objects list are therefore dead, so we free them here.
547 freeChain(gen->large_objects);
548 gen->large_objects = gen->scavenged_large_objects;
549 gen->n_large_blocks = gen->n_scavenged_large_blocks;
550 gen->n_new_large_words = 0;
552 else // for generations > N
554 /* For older generations, we need to append the
555 * scavenged_large_object list (i.e. large objects that have been
556 * promoted during this GC) to the large_object list for that step.
558 for (bd = gen->scavenged_large_objects; bd; bd = next) {
560 dbl_link_onto(bd, &gen->large_objects);
563 // add the new blocks we promoted during this GC
564 gen->n_large_blocks += gen->n_scavenged_large_blocks;
567 ASSERT(countBlocks(gen->large_objects) == gen->n_large_blocks);
569 gen->scavenged_large_objects = NULL;
570 gen->n_scavenged_large_blocks = 0;
573 live_words += genLiveWords(gen);
574 live_blocks += genLiveBlocks(gen);
576 // add in the partial blocks in the gen_workspaces, but ignore gen 0
577 // if this is a local GC (we can't count another capability's part_list)
580 for (i = 0; i < n_capabilities; i++) {
581 live_words += gcThreadLiveWords(i, gen->no);
582 live_blocks += gcThreadLiveBlocks(i, gen->no);
585 } // for all generations
587 // update the max size of older generations after a major GC
588 resize_generations();
590 // Free the mark stack.
591 if (mark_stack_top_bd != NULL) {
592 debugTrace(DEBUG_gc, "mark stack: %d blocks",
593 countBlocks(mark_stack_top_bd));
594 freeChain(mark_stack_top_bd);
598 for (g = 0; g <= N; g++) {
599 gen = &generations[g];
600 if (gen->bitmap != NULL) {
601 freeGroup(gen->bitmap);
606 // Reset the nursery: make the blocks empty
607 allocated += clearNurseries();
613 // mark the garbage collected CAFs as dead
614 #if 0 && defined(DEBUG) // doesn't work at the moment
615 if (major_gc) { gcCAFs(); }
619 // resetStaticObjectForRetainerProfiling() must be called before
621 if (n_gc_threads > 1) {
622 barf("profiling is currently broken with multi-threaded GC");
623 // ToDo: fix the gct->scavenged_static_objects below
625 resetStaticObjectForRetainerProfiling(gct->scavenged_static_objects);
628 // zero the scavenged static object list
631 if (n_gc_threads == 1) {
632 zero_static_object_list(gct->scavenged_static_objects);
634 for (i = 0; i < n_gc_threads; i++) {
635 zero_static_object_list(gc_threads[i]->scavenged_static_objects);
640 // Update the stable pointer hash table.
641 updateStablePtrTable(major_gc);
643 // unlock the StablePtr table. Must be before scheduleFinalizers(),
644 // because a finalizer may call hs_free_fun_ptr() or
645 // hs_free_stable_ptr(), both of which access the StablePtr table.
648 // Start any pending finalizers. Must be after
649 // updateStablePtrTable() and stablePtrPostGC() (see #4221).
651 scheduleFinalizers(cap, old_weak_ptr_list);
654 // check sanity after GC
655 // before resurrectThreads(), because that might overwrite some
656 // closures, which will cause problems with THREADED where we don't
658 IF_DEBUG(sanity, checkSanity(rtsTrue /* after GC */, major_gc));
660 // send exceptions to any threads which were about to die
662 resurrectThreads(resurrected_threads);
667 need = BLOCKS_TO_MBLOCKS(n_alloc_blocks);
668 got = mblocks_allocated;
669 /* If the amount of data remains constant, next major GC we'll
670 require (F+1)*need. We leave (F+2)*need in order to reduce
671 repeated deallocation and reallocation. */
672 need = (RtsFlags.GcFlags.oldGenFactor + 2) * need;
674 returnMemoryToOS(got - need);
678 // extra GC trace info
679 IF_DEBUG(gc, statDescribeGens());
682 // symbol-table based profiling
683 /* heapCensus(to_blocks); */ /* ToDo */
686 // restore enclosing cost centre
692 // check for memory leaks if DEBUG is on
693 memInventory(DEBUG_gc);
696 #ifdef RTS_GTK_FRONTPANEL
697 if (RtsFlags.GcFlags.frontpanel) {
698 updateFrontPanelAfterGC( N, live );
702 // ok, GC over: tell the stats department what happened.
703 stat_endGC(gct, allocated, live_words,
704 copied, N, max_copied, avg_copied,
705 live_blocks * BLOCK_SIZE_W - live_words /* slop */);
707 // Guess which generation we'll collect *next* time
708 initialise_N(force_major_gc);
710 #if defined(RTS_USER_SIGNALS)
711 if (RtsFlags.MiscFlags.install_signal_handlers) {
712 // unblock signals again
713 unblockUserSignals();
722 /* -----------------------------------------------------------------------------
723 Figure out which generation to collect, initialise N and major_gc.
725 Also returns the total number of blocks in generations that will be
727 -------------------------------------------------------------------------- */
730 initialise_N (rtsBool force_major_gc)
733 nat blocks, blocks_total;
738 if (force_major_gc) {
739 N = RtsFlags.GcFlags.generations - 1;
744 for (g = RtsFlags.GcFlags.generations - 1; g >= 0; g--) {
746 blocks = generations[g].n_words / BLOCK_SIZE_W
747 + generations[g].n_large_blocks;
749 if (blocks >= generations[g].max_blocks) {
753 blocks_total += blocks;
757 blocks_total += countNurseryBlocks();
759 major_gc = (N == RtsFlags.GcFlags.generations-1);
763 /* -----------------------------------------------------------------------------
764 Initialise the gc_thread structures.
765 -------------------------------------------------------------------------- */
767 #define GC_THREAD_INACTIVE 0
768 #define GC_THREAD_STANDING_BY 1
769 #define GC_THREAD_RUNNING 2
770 #define GC_THREAD_WAITING_TO_CONTINUE 3
773 new_gc_thread (nat n, gc_thread *t)
778 t->cap = &capabilities[n];
782 initSpinLock(&t->gc_spin);
783 initSpinLock(&t->mut_spin);
784 ACQUIRE_SPIN_LOCK(&t->gc_spin);
785 t->wakeup = GC_THREAD_INACTIVE; // starts true, so we can wait for the
786 // thread to start up, see wakeup_gc_threads
790 t->free_blocks = NULL;
799 for (g = 0; g < RtsFlags.GcFlags.generations; g++)
802 ws->gen = &generations[g];
803 ASSERT(g == ws->gen->no);
807 // alloc_todo_block(ws,0);
808 // but can't, because it uses gct which isn't set up at this point.
809 // Hence, allocate a block for todo_bd manually:
811 bdescr *bd = allocBlock(); // no lock, locks aren't initialised yet
812 initBdescr(bd, ws->gen, ws->gen->to);
813 bd->flags = BF_EVACUATED;
814 bd->u.scan = bd->free = bd->start;
817 ws->todo_free = bd->free;
818 ws->todo_lim = bd->start + BLOCK_SIZE_W;
821 ws->todo_q = newWSDeque(128);
822 ws->todo_overflow = NULL;
823 ws->n_todo_overflow = 0;
824 ws->todo_large_objects = NULL;
826 ws->part_list = NULL;
827 ws->n_part_blocks = 0;
829 ws->scavd_list = NULL;
830 ws->n_scavd_blocks = 0;
838 if (gc_threads == NULL) {
839 #if defined(THREADED_RTS)
841 gc_threads = stgMallocBytes (RtsFlags.ParFlags.nNodes *
845 for (i = 0; i < RtsFlags.ParFlags.nNodes; i++) {
847 stgMallocBytes(sizeof(gc_thread) +
848 RtsFlags.GcFlags.generations * sizeof(gen_workspace),
851 new_gc_thread(i, gc_threads[i]);
854 gc_threads = stgMallocBytes (sizeof(gc_thread*),"alloc_gc_threads");
856 new_gc_thread(0,gc_threads[0]);
865 if (gc_threads != NULL) {
866 #if defined(THREADED_RTS)
868 for (i = 0; i < n_capabilities; i++) {
869 for (g = 0; g < RtsFlags.GcFlags.generations; g++)
871 freeWSDeque(gc_threads[i]->gens[g].todo_q);
873 stgFree (gc_threads[i]);
875 stgFree (gc_threads);
877 for (g = 0; g < RtsFlags.GcFlags.generations; g++)
879 freeWSDeque(gc_threads[0]->gens[g].todo_q);
881 stgFree (gc_threads);
887 /* ----------------------------------------------------------------------------
889 ------------------------------------------------------------------------- */
891 static volatile StgWord gc_running_threads;
897 new = atomic_inc(&gc_running_threads);
898 ASSERT(new <= n_gc_threads);
905 ASSERT(gc_running_threads != 0);
906 return atomic_dec(&gc_running_threads);
919 // scavenge objects in compacted generation
920 if (mark_stack_bd != NULL && !mark_stack_empty()) {
924 // Check for global work in any step. We don't need to check for
925 // local work, because we have already exited scavenge_loop(),
926 // which means there is no local work for this thread.
927 for (g = 0; g < (int)RtsFlags.GcFlags.generations; g++) {
929 if (ws->todo_large_objects) return rtsTrue;
930 if (!looksEmptyWSDeque(ws->todo_q)) return rtsTrue;
931 if (ws->todo_overflow) return rtsTrue;
934 #if defined(THREADED_RTS)
937 // look for work to steal
938 for (n = 0; n < n_gc_threads; n++) {
939 if (n == gct->thread_index) continue;
940 for (g = RtsFlags.GcFlags.generations-1; g >= 0; g--) {
941 ws = &gc_threads[n]->gens[g];
942 if (!looksEmptyWSDeque(ws->todo_q)) return rtsTrue;
949 #if defined(THREADED_RTS)
957 scavenge_until_all_done (void)
963 #if defined(THREADED_RTS)
964 if (n_gc_threads > 1) {
973 collect_gct_blocks();
975 // scavenge_loop() only exits when there's no work to do
978 traceEventGcIdle(gct->cap);
980 debugTrace(DEBUG_gc, "%d GC threads still running", r);
982 while (gc_running_threads != 0) {
986 traceEventGcWork(gct->cap);
989 // any_work() does not remove the work from the queue, it
990 // just checks for the presence of work. If we find any,
991 // then we increment gc_running_threads and go back to
992 // scavenge_loop() to perform any pending work.
995 traceEventGcDone(gct->cap);
998 #if defined(THREADED_RTS)
1001 gcWorkerThread (Capability *cap)
1003 gc_thread *saved_gct;
1005 // necessary if we stole a callee-saves register for gct:
1008 gct = gc_threads[cap->no];
1009 gct->id = osThreadId();
1011 stat_gcWorkerThreadStart(gct);
1013 // Wait until we're told to wake up
1014 RELEASE_SPIN_LOCK(&gct->mut_spin);
1015 gct->wakeup = GC_THREAD_STANDING_BY;
1016 debugTrace(DEBUG_gc, "GC thread %d standing by...", gct->thread_index);
1017 ACQUIRE_SPIN_LOCK(&gct->gc_spin);
1020 // start performance counters in this thread...
1021 if (gct->papi_events == -1) {
1022 papi_init_eventset(&gct->papi_events);
1024 papi_thread_start_gc1_count(gct->papi_events);
1027 init_gc_thread(gct);
1029 traceEventGcWork(gct->cap);
1031 // Every thread evacuates some roots.
1032 gct->evac_gen_no = 0;
1033 markCapability(mark_root, gct, cap, rtsTrue/*prune sparks*/);
1034 scavenge_capability_mut_lists(cap);
1036 scavenge_until_all_done();
1039 // Now that the whole heap is marked, we discard any sparks that
1040 // were found to be unreachable. The main GC thread is currently
1041 // marking heap reachable via weak pointers, so it is
1042 // non-deterministic whether a spark will be retained if it is
1043 // only reachable via weak pointers. To fix this problem would
1044 // require another GC barrier, which is too high a price.
1045 pruneSparkQueue(cap);
1049 // count events in this thread towards the GC totals
1050 papi_thread_stop_gc1_count(gct->papi_events);
1053 // Wait until we're told to continue
1054 RELEASE_SPIN_LOCK(&gct->gc_spin);
1055 gct->wakeup = GC_THREAD_WAITING_TO_CONTINUE;
1056 debugTrace(DEBUG_gc, "GC thread %d waiting to continue...",
1058 ACQUIRE_SPIN_LOCK(&gct->mut_spin);
1059 debugTrace(DEBUG_gc, "GC thread %d on my way...", gct->thread_index);
1061 // record the time spent doing GC in the Task structure
1062 stat_gcWorkerThreadDone(gct);
1069 #if defined(THREADED_RTS)
1072 waitForGcThreads (Capability *cap USED_IF_THREADS)
1074 const nat n_threads = RtsFlags.ParFlags.nNodes;
1075 const nat me = cap->no;
1077 rtsBool retry = rtsTrue;
1080 for (i=0; i < n_threads; i++) {
1081 if (i == me) continue;
1082 if (gc_threads[i]->wakeup != GC_THREAD_STANDING_BY) {
1083 prodCapability(&capabilities[i], cap->running_task);
1086 for (j=0; j < 10; j++) {
1088 for (i=0; i < n_threads; i++) {
1089 if (i == me) continue;
1091 setContextSwitches();
1092 if (gc_threads[i]->wakeup != GC_THREAD_STANDING_BY) {
1102 #endif // THREADED_RTS
1105 start_gc_threads (void)
1107 #if defined(THREADED_RTS)
1108 gc_running_threads = 0;
1113 wakeup_gc_threads (nat me USED_IF_THREADS)
1115 #if defined(THREADED_RTS)
1118 if (n_gc_threads == 1) return;
1120 for (i=0; i < n_gc_threads; i++) {
1121 if (i == me) continue;
1123 debugTrace(DEBUG_gc, "waking up gc thread %d", i);
1124 if (gc_threads[i]->wakeup != GC_THREAD_STANDING_BY) barf("wakeup_gc_threads");
1126 gc_threads[i]->wakeup = GC_THREAD_RUNNING;
1127 ACQUIRE_SPIN_LOCK(&gc_threads[i]->mut_spin);
1128 RELEASE_SPIN_LOCK(&gc_threads[i]->gc_spin);
1133 // After GC is complete, we must wait for all GC threads to enter the
1134 // standby state, otherwise they may still be executing inside
1135 // any_work(), and may even remain awake until the next GC starts.
1137 shutdown_gc_threads (nat me USED_IF_THREADS)
1139 #if defined(THREADED_RTS)
1142 if (n_gc_threads == 1) return;
1144 for (i=0; i < n_gc_threads; i++) {
1145 if (i == me) continue;
1146 while (gc_threads[i]->wakeup != GC_THREAD_WAITING_TO_CONTINUE) { write_barrier(); }
1151 #if defined(THREADED_RTS)
1153 releaseGCThreads (Capability *cap USED_IF_THREADS)
1155 const nat n_threads = RtsFlags.ParFlags.nNodes;
1156 const nat me = cap->no;
1158 for (i=0; i < n_threads; i++) {
1159 if (i == me) continue;
1160 if (gc_threads[i]->wakeup != GC_THREAD_WAITING_TO_CONTINUE)
1161 barf("releaseGCThreads");
1163 gc_threads[i]->wakeup = GC_THREAD_INACTIVE;
1164 ACQUIRE_SPIN_LOCK(&gc_threads[i]->gc_spin);
1165 RELEASE_SPIN_LOCK(&gc_threads[i]->mut_spin);
1170 /* ----------------------------------------------------------------------------
1171 Initialise a generation that is to be collected
1172 ------------------------------------------------------------------------- */
1175 prepare_collected_gen (generation *gen)
1181 // Throw away the current mutable list. Invariant: the mutable
1182 // list always has at least one block; this means we can avoid a
1183 // check for NULL in recordMutable().
1186 for (i = 0; i < n_capabilities; i++) {
1187 freeChain(capabilities[i].mut_lists[g]);
1188 capabilities[i].mut_lists[g] = allocBlock();
1192 gen = &generations[g];
1193 ASSERT(gen->no == g);
1195 // we'll construct a new list of threads in this step
1196 // during GC, throw away the current list.
1197 gen->old_threads = gen->threads;
1198 gen->threads = END_TSO_QUEUE;
1200 // deprecate the existing blocks
1201 gen->old_blocks = gen->blocks;
1202 gen->n_old_blocks = gen->n_blocks;
1206 gen->live_estimate = 0;
1208 // initialise the large object queues.
1209 ASSERT(gen->scavenged_large_objects == NULL);
1210 ASSERT(gen->n_scavenged_large_blocks == 0);
1212 // grab all the partial blocks stashed in the gc_thread workspaces and
1213 // move them to the old_blocks list of this gen.
1214 for (n = 0; n < n_capabilities; n++) {
1215 ws = &gc_threads[n]->gens[gen->no];
1217 for (bd = ws->part_list; bd != NULL; bd = next) {
1219 bd->link = gen->old_blocks;
1220 gen->old_blocks = bd;
1221 gen->n_old_blocks += bd->blocks;
1223 ws->part_list = NULL;
1224 ws->n_part_blocks = 0;
1226 ASSERT(ws->scavd_list == NULL);
1227 ASSERT(ws->n_scavd_blocks == 0);
1229 if (ws->todo_free != ws->todo_bd->start) {
1230 ws->todo_bd->free = ws->todo_free;
1231 ws->todo_bd->link = gen->old_blocks;
1232 gen->old_blocks = ws->todo_bd;
1233 gen->n_old_blocks += ws->todo_bd->blocks;
1234 alloc_todo_block(ws,0); // always has one block.
1238 // mark the small objects as from-space
1239 for (bd = gen->old_blocks; bd; bd = bd->link) {
1240 bd->flags &= ~BF_EVACUATED;
1243 // mark the large objects as from-space
1244 for (bd = gen->large_objects; bd; bd = bd->link) {
1245 bd->flags &= ~BF_EVACUATED;
1248 // for a compacted generation, we need to allocate the bitmap
1250 lnat bitmap_size; // in bytes
1251 bdescr *bitmap_bdescr;
1254 bitmap_size = gen->n_old_blocks * BLOCK_SIZE / (sizeof(W_)*BITS_PER_BYTE);
1256 if (bitmap_size > 0) {
1257 bitmap_bdescr = allocGroup((lnat)BLOCK_ROUND_UP(bitmap_size)
1259 gen->bitmap = bitmap_bdescr;
1260 bitmap = bitmap_bdescr->start;
1262 debugTrace(DEBUG_gc, "bitmap_size: %d, bitmap: %p",
1263 bitmap_size, bitmap);
1265 // don't forget to fill it with zeros!
1266 memset(bitmap, 0, bitmap_size);
1268 // For each block in this step, point to its bitmap from the
1269 // block descriptor.
1270 for (bd=gen->old_blocks; bd != NULL; bd = bd->link) {
1271 bd->u.bitmap = bitmap;
1272 bitmap += BLOCK_SIZE_W / (sizeof(W_)*BITS_PER_BYTE);
1274 // Also at this point we set the BF_MARKED flag
1275 // for this block. The invariant is that
1276 // BF_MARKED is always unset, except during GC
1277 // when it is set on those blocks which will be
1279 if (!(bd->flags & BF_FRAGMENTED)) {
1280 bd->flags |= BF_MARKED;
1283 // BF_SWEPT should be marked only for blocks that are being
1284 // collected in sweep()
1285 bd->flags &= ~BF_SWEPT;
1292 /* ----------------------------------------------------------------------------
1293 Save the mutable lists in saved_mut_lists
1294 ------------------------------------------------------------------------- */
1297 stash_mut_list (Capability *cap, nat gen_no)
1299 cap->saved_mut_lists[gen_no] = cap->mut_lists[gen_no];
1300 cap->mut_lists[gen_no] = allocBlock_sync();
1303 /* ----------------------------------------------------------------------------
1304 Initialise a generation that is *not* to be collected
1305 ------------------------------------------------------------------------- */
1308 prepare_uncollected_gen (generation *gen)
1313 ASSERT(gen->no > 0);
1315 // save the current mutable lists for this generation, and
1316 // allocate a fresh block for each one. We'll traverse these
1317 // mutable lists as roots early on in the GC.
1318 for (i = 0; i < n_capabilities; i++) {
1319 stash_mut_list(&capabilities[i], gen->no);
1322 ASSERT(gen->scavenged_large_objects == NULL);
1323 ASSERT(gen->n_scavenged_large_blocks == 0);
1326 /* -----------------------------------------------------------------------------
1327 Collect the completed blocks from a GC thread and attach them to
1329 -------------------------------------------------------------------------- */
1332 collect_gct_blocks (void)
1338 for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
1341 // there may still be a block attached to ws->todo_bd;
1342 // leave it there to use next time.
1344 if (ws->scavd_list != NULL) {
1345 ACQUIRE_SPIN_LOCK(&ws->gen->sync);
1347 ASSERT(gct->scan_bd == NULL);
1348 ASSERT(countBlocks(ws->scavd_list) == ws->n_scavd_blocks);
1351 for (bd = ws->scavd_list; bd != NULL; bd = bd->link) {
1352 ws->gen->n_words += bd->free - bd->start;
1356 prev->link = ws->gen->blocks;
1357 ws->gen->blocks = ws->scavd_list;
1359 ws->gen->n_blocks += ws->n_scavd_blocks;
1361 ws->scavd_list = NULL;
1362 ws->n_scavd_blocks = 0;
1364 RELEASE_SPIN_LOCK(&ws->gen->sync);
1369 /* -----------------------------------------------------------------------------
1370 Initialise a gc_thread before GC
1371 -------------------------------------------------------------------------- */
1374 init_gc_thread (gc_thread *t)
1376 t->static_objects = END_OF_STATIC_LIST;
1377 t->scavenged_static_objects = END_OF_STATIC_LIST;
1379 t->mut_lists = t->cap->mut_lists;
1381 t->failed_to_evac = rtsFalse;
1382 t->eager_promotion = rtsTrue;
1383 t->thunk_selector_depth = 0;
1388 t->scav_find_work = 0;
1391 /* -----------------------------------------------------------------------------
1392 Function we pass to evacuate roots.
1393 -------------------------------------------------------------------------- */
1396 mark_root(void *user USED_IF_THREADS, StgClosure **root)
1398 // we stole a register for gct, but this function is called from
1399 // *outside* the GC where the register variable is not in effect,
1400 // so we need to save and restore it here. NB. only call
1401 // mark_root() from the main GC thread, otherwise gct will be
1403 gc_thread *saved_gct;
1412 /* -----------------------------------------------------------------------------
1413 Initialising the static object & mutable lists
1414 -------------------------------------------------------------------------- */
1417 zero_static_object_list(StgClosure* first_static)
1421 const StgInfoTable *info;
1423 for (p = first_static; p != END_OF_STATIC_LIST; p = link) {
1425 link = *STATIC_LINK(info, p);
1426 *STATIC_LINK(info,p) = NULL;
1430 /* ----------------------------------------------------------------------------
1431 Reset the sizes of the older generations when we do a major
1434 CURRENT STRATEGY: make all generations except zero the same size.
1435 We have to stay within the maximum heap size, and leave a certain
1436 percentage of the maximum heap size available to allocate into.
1437 ------------------------------------------------------------------------- */
1440 resize_generations (void)
1444 if (major_gc && RtsFlags.GcFlags.generations > 1) {
1445 nat live, size, min_alloc, words;
1446 const nat max = RtsFlags.GcFlags.maxHeapSize;
1447 const nat gens = RtsFlags.GcFlags.generations;
1449 // live in the oldest generations
1450 if (oldest_gen->live_estimate != 0) {
1451 words = oldest_gen->live_estimate;
1453 words = oldest_gen->n_words;
1455 live = (words + BLOCK_SIZE_W - 1) / BLOCK_SIZE_W +
1456 oldest_gen->n_large_blocks;
1458 // default max size for all generations except zero
1459 size = stg_max(live * RtsFlags.GcFlags.oldGenFactor,
1460 RtsFlags.GcFlags.minOldGenSize);
1462 if (RtsFlags.GcFlags.heapSizeSuggestionAuto) {
1463 RtsFlags.GcFlags.heapSizeSuggestion = size;
1466 // minimum size for generation zero
1467 min_alloc = stg_max((RtsFlags.GcFlags.pcFreeHeap * max) / 200,
1468 RtsFlags.GcFlags.minAllocAreaSize);
1470 // Auto-enable compaction when the residency reaches a
1471 // certain percentage of the maximum heap size (default: 30%).
1472 if (RtsFlags.GcFlags.compact ||
1474 oldest_gen->n_blocks >
1475 (RtsFlags.GcFlags.compactThreshold * max) / 100)) {
1476 oldest_gen->mark = 1;
1477 oldest_gen->compact = 1;
1478 // debugBelch("compaction: on\n", live);
1480 oldest_gen->mark = 0;
1481 oldest_gen->compact = 0;
1482 // debugBelch("compaction: off\n", live);
1485 if (RtsFlags.GcFlags.sweep) {
1486 oldest_gen->mark = 1;
1489 // if we're going to go over the maximum heap size, reduce the
1490 // size of the generations accordingly. The calculation is
1491 // different if compaction is turned on, because we don't need
1492 // to double the space required to collect the old generation.
1495 // this test is necessary to ensure that the calculations
1496 // below don't have any negative results - we're working
1497 // with unsigned values here.
1498 if (max < min_alloc) {
1502 if (oldest_gen->compact) {
1503 if ( (size + (size - 1) * (gens - 2) * 2) + min_alloc > max ) {
1504 size = (max - min_alloc) / ((gens - 1) * 2 - 1);
1507 if ( (size * (gens - 1) * 2) + min_alloc > max ) {
1508 size = (max - min_alloc) / ((gens - 1) * 2);
1518 debugBelch("live: %d, min_alloc: %d, size : %d, max = %d\n", live,
1519 min_alloc, size, max);
1522 for (g = 0; g < gens; g++) {
1523 generations[g].max_blocks = size;
1528 /* -----------------------------------------------------------------------------
1529 Calculate the new size of the nursery, and resize it.
1530 -------------------------------------------------------------------------- */
1533 resize_nursery (void)
1535 const lnat min_nursery = RtsFlags.GcFlags.minAllocAreaSize * n_capabilities;
1537 if (RtsFlags.GcFlags.generations == 1)
1538 { // Two-space collector:
1541 /* set up a new nursery. Allocate a nursery size based on a
1542 * function of the amount of live data (by default a factor of 2)
1543 * Use the blocks from the old nursery if possible, freeing up any
1546 * If we get near the maximum heap size, then adjust our nursery
1547 * size accordingly. If the nursery is the same size as the live
1548 * data (L), then we need 3L bytes. We can reduce the size of the
1549 * nursery to bring the required memory down near 2L bytes.
1551 * A normal 2-space collector would need 4L bytes to give the same
1552 * performance we get from 3L bytes, reducing to the same
1553 * performance at 2L bytes.
1555 blocks = generations[0].n_blocks;
1557 if ( RtsFlags.GcFlags.maxHeapSize != 0 &&
1558 blocks * RtsFlags.GcFlags.oldGenFactor * 2 >
1559 RtsFlags.GcFlags.maxHeapSize )
1561 long adjusted_blocks; // signed on purpose
1564 adjusted_blocks = (RtsFlags.GcFlags.maxHeapSize - 2 * blocks);
1566 debugTrace(DEBUG_gc, "near maximum heap size of 0x%x blocks, blocks = %d, adjusted to %ld",
1567 RtsFlags.GcFlags.maxHeapSize, blocks, adjusted_blocks);
1569 pc_free = adjusted_blocks * 100 / RtsFlags.GcFlags.maxHeapSize;
1570 if (pc_free < RtsFlags.GcFlags.pcFreeHeap) /* might even * be < 0 */
1574 blocks = adjusted_blocks;
1578 blocks *= RtsFlags.GcFlags.oldGenFactor;
1579 if (blocks < min_nursery)
1581 blocks = min_nursery;
1584 resizeNurseries(blocks);
1586 else // Generational collector
1589 * If the user has given us a suggested heap size, adjust our
1590 * allocation area to make best use of the memory available.
1592 if (RtsFlags.GcFlags.heapSizeSuggestion)
1595 const nat needed = calcNeeded(); // approx blocks needed at next GC
1597 /* Guess how much will be live in generation 0 step 0 next time.
1598 * A good approximation is obtained by finding the
1599 * percentage of g0 that was live at the last minor GC.
1601 * We have an accurate figure for the amount of copied data in
1602 * 'copied', but we must convert this to a number of blocks, with
1603 * a small adjustment for estimated slop at the end of a block
1608 g0_pcnt_kept = ((copied / (BLOCK_SIZE_W - 10)) * 100)
1609 / countNurseryBlocks();
1612 /* Estimate a size for the allocation area based on the
1613 * information available. We might end up going slightly under
1614 * or over the suggested heap size, but we should be pretty
1617 * Formula: suggested - needed
1618 * ----------------------------
1619 * 1 + g0_pcnt_kept/100
1621 * where 'needed' is the amount of memory needed at the next
1622 * collection for collecting all gens except g0.
1625 (((long)RtsFlags.GcFlags.heapSizeSuggestion - (long)needed) * 100) /
1626 (100 + (long)g0_pcnt_kept);
1628 if (blocks < (long)min_nursery) {
1629 blocks = min_nursery;
1632 resizeNurseries((nat)blocks);
1636 // we might have added extra large blocks to the nursery, so
1637 // resize back to minAllocAreaSize again.
1638 resizeNurseriesFixed(RtsFlags.GcFlags.minAllocAreaSize);
1643 /* -----------------------------------------------------------------------------
1644 Sanity code for CAF garbage collection.
1646 With DEBUG turned on, we manage a CAF list in addition to the SRT
1647 mechanism. After GC, we run down the CAF list and blackhole any
1648 CAFs which have been garbage collected. This means we get an error
1649 whenever the program tries to enter a garbage collected CAF.
1651 Any garbage collected CAFs are taken off the CAF list at the same
1653 -------------------------------------------------------------------------- */
1655 #if 0 && defined(DEBUG)
1662 const StgInfoTable *info;
1673 ASSERT(info->type == IND_STATIC);
1675 if (STATIC_LINK(info,p) == NULL) {
1676 debugTrace(DEBUG_gccafs, "CAF gc'd at 0x%04lx", (long)p);
1678 SET_INFO(p,&stg_BLACKHOLE_info);
1679 p = STATIC_LINK2(info,p);
1683 pp = &STATIC_LINK2(info,p);
1690 debugTrace(DEBUG_gccafs, "%d CAFs live", i);