1 /* -----------------------------------------------------------------------------
3 * (c) The GHC Team 1998-2008
5 * Generational garbage collector
7 * Documentation on the architecture of the Garbage Collector can be
8 * found in the online commentary:
10 * http://hackage.haskell.org/trac/ghc/wiki/Commentary/Rts/Storage/GC
12 * ---------------------------------------------------------------------------*/
14 #include "PosixSource.h"
25 #include "BlockAlloc.h"
29 #include "RtsSignals.h"
31 #if defined(RTS_GTK_FRONTPANEL)
32 #include "FrontPanel.h"
35 #include "RetainerProfile.h"
36 #include "LdvProfile.h"
37 #include "RaiseAsync.h"
47 #include "MarkStack.h"
52 #include <string.h> // for memset()
55 /* -----------------------------------------------------------------------------
57 -------------------------------------------------------------------------- */
59 /* STATIC OBJECT LIST.
62 * We maintain a linked list of static objects that are still live.
63 * The requirements for this list are:
65 * - we need to scan the list while adding to it, in order to
66 * scavenge all the static objects (in the same way that
67 * breadth-first scavenging works for dynamic objects).
69 * - we need to be able to tell whether an object is already on
70 * the list, to break loops.
72 * Each static object has a "static link field", which we use for
73 * linking objects on to the list. We use a stack-type list, consing
74 * objects on the front as they are added (this means that the
75 * scavenge phase is depth-first, not breadth-first, but that
78 * A separate list is kept for objects that have been scavenged
79 * already - this is so that we can zero all the marks afterwards.
81 * An object is on the list if its static link field is non-zero; this
82 * means that we have to mark the end of the list with '1', not NULL.
84 * Extra notes for generational GC:
86 * Each generation has a static object list associated with it. When
87 * collecting generations up to N, we treat the static object lists
88 * from generations > N as roots.
90 * We build up a static object list while collecting generations 0..N,
91 * which is then appended to the static object list of generation N+1.
94 /* N is the oldest generation being collected, where the generations
95 * are numbered starting at 0. A major GC (indicated by the major_gc
96 * flag) is when we're collecting all generations. We only attempt to
97 * deal with static objects and GC CAFs when doing a major GC.
102 /* Data used for allocation area sizing.
104 static lnat g0_pcnt_kept = 30; // percentage of g0 live at last minor GC
114 /* Thread-local data for each GC thread
116 gc_thread **gc_threads = NULL;
118 #if !defined(THREADED_RTS)
119 StgWord8 the_gc_thread[sizeof(gc_thread) + 64 * sizeof(gen_workspace)];
122 // Number of threads running in *this* GC. Affects how many
123 // step->todos[] lists we have to look in to find work.
127 long copied; // *words* copied & scavenged during this GC
129 rtsBool work_stealing;
133 /* -----------------------------------------------------------------------------
134 Static function declarations
135 -------------------------------------------------------------------------- */
137 static void mark_root (void *user, StgClosure **root);
138 static void zero_static_object_list (StgClosure* first_static);
139 static nat initialise_N (rtsBool force_major_gc);
140 static void init_collected_gen (nat g, nat threads);
141 static void init_uncollected_gen (nat g, nat threads);
142 static void init_gc_thread (gc_thread *t);
143 static void resize_generations (void);
144 static void resize_nursery (void);
145 static void start_gc_threads (void);
146 static void scavenge_until_all_done (void);
147 static StgWord inc_running (void);
148 static StgWord dec_running (void);
149 static void wakeup_gc_threads (nat n_threads, nat me);
150 static void shutdown_gc_threads (nat n_threads, nat me);
152 #if 0 && defined(DEBUG)
153 static void gcCAFs (void);
156 /* -----------------------------------------------------------------------------
158 -------------------------------------------------------------------------- */
160 bdescr *mark_stack_top_bd; // topmost block in the mark stack
161 bdescr *mark_stack_bd; // current block in the mark stack
162 StgPtr mark_sp; // pointer to the next unallocated mark stack entry
164 /* -----------------------------------------------------------------------------
165 GarbageCollect: the main entry point to the garbage collector.
167 Locks held: all capabilities are held throughout GarbageCollect().
168 -------------------------------------------------------------------------- */
171 GarbageCollect (rtsBool force_major_gc,
172 nat gc_type USED_IF_THREADS,
177 lnat live, allocated, max_copied, avg_copied, slop;
178 gc_thread *saved_gct;
181 // necessary if we stole a callee-saves register for gct:
185 CostCentreStack *prev_CCS;
190 #if defined(RTS_USER_SIGNALS)
191 if (RtsFlags.MiscFlags.install_signal_handlers) {
197 ASSERT(sizeof(gen_workspace) == 16 * sizeof(StgWord));
198 // otherwise adjust the padding in gen_workspace.
200 // tell the stats department that we've started a GC
203 // tell the STM to discard any cached closures it's hoping to re-use
206 // lock the StablePtr table
215 // attribute any costs to CCS_GC
221 /* Approximate how much we allocated.
222 * Todo: only when generating stats?
224 allocated = calcAllocated();
226 /* Figure out which generation to collect
228 n = initialise_N(force_major_gc);
230 #if defined(THREADED_RTS)
231 work_stealing = RtsFlags.ParFlags.parGcLoadBalancingEnabled &&
232 N >= RtsFlags.ParFlags.parGcLoadBalancingGen;
233 // It's not always a good idea to do load balancing in parallel
234 // GC. In particular, for a parallel program we don't want to
235 // lose locality by moving cached data into another CPU's cache
236 // (this effect can be quite significant).
238 // We could have a more complex way to deterimine whether to do
239 // work stealing or not, e.g. it might be a good idea to do it
240 // if the heap is big. For now, we just turn it on or off with
244 /* Start threads, so they can be spinning up while we finish initialisation.
248 #if defined(THREADED_RTS)
249 /* How many threads will be participating in this GC?
250 * We don't try to parallelise minor GCs (unless the user asks for
251 * it with +RTS -gn0), or mark/compact/sweep GC.
253 if (gc_type == PENDING_GC_PAR) {
254 n_gc_threads = RtsFlags.ParFlags.nNodes;
262 debugTrace(DEBUG_gc, "GC (gen %d): %d KB to collect, %ld MB in use, using %d thread(s)",
263 N, n * (BLOCK_SIZE / 1024), mblocks_allocated, n_gc_threads);
265 #ifdef RTS_GTK_FRONTPANEL
266 if (RtsFlags.GcFlags.frontpanel) {
267 updateFrontPanelBeforeGC(N);
272 // check for memory leaks if DEBUG is on
273 memInventory(DEBUG_gc);
276 // check sanity *before* GC
277 IF_DEBUG(sanity, checkSanity(rtsTrue));
279 // Initialise all our gc_thread structures
280 for (t = 0; t < n_gc_threads; t++) {
281 init_gc_thread(gc_threads[t]);
284 // Initialise all the generations/steps that we're collecting.
285 for (g = 0; g <= N; g++) {
286 init_collected_gen(g,n_gc_threads);
289 // Initialise all the generations/steps that we're *not* collecting.
290 for (g = N+1; g < RtsFlags.GcFlags.generations; g++) {
291 init_uncollected_gen(g,n_gc_threads);
294 /* Allocate a mark stack if we're doing a major collection.
296 if (major_gc && oldest_gen->mark) {
297 mark_stack_bd = allocBlock();
298 mark_stack_top_bd = mark_stack_bd;
299 mark_stack_bd->link = NULL;
300 mark_stack_bd->u.back = NULL;
301 mark_sp = mark_stack_bd->start;
303 mark_stack_bd = NULL;
304 mark_stack_top_bd = NULL;
308 // this is the main thread
310 if (n_gc_threads == 1) {
311 SET_GCT(gc_threads[0]);
313 SET_GCT(gc_threads[cap->no]);
316 SET_GCT(gc_threads[0]);
319 /* -----------------------------------------------------------------------
320 * follow all the roots that we know about:
323 // the main thread is running: this prevents any other threads from
324 // exiting prematurely, so we can start them now.
325 // NB. do this after the mutable lists have been saved above, otherwise
326 // the other GC threads will be writing into the old mutable lists.
328 wakeup_gc_threads(n_gc_threads, gct->thread_index);
330 // Mutable lists from each generation > N
331 // we want to *scavenge* these roots, not evacuate them: they're not
332 // going to move in this GC.
333 // Also do them in reverse generation order, for the usual reason:
334 // namely to reduce the likelihood of spurious old->new pointers.
336 for (g = RtsFlags.GcFlags.generations-1; g > N; g--) {
337 #if defined(THREADED_RTS)
338 if (n_gc_threads > 1) {
339 scavenge_mutable_list(generations[g].saved_mut_list, &generations[g]);
341 scavenge_mutable_list1(generations[g].saved_mut_list, &generations[g]);
344 scavenge_mutable_list(generations[g].saved_mut_list, &generations[g]);
346 freeChain_sync(generations[g].saved_mut_list);
347 generations[g].saved_mut_list = NULL;
351 // scavenge the capability-private mutable lists. This isn't part
352 // of markSomeCapabilities() because markSomeCapabilities() can only
353 // call back into the GC via mark_root() (due to the gct register
355 if (n_gc_threads == 1) {
356 for (n = 0; n < n_capabilities; n++) {
357 #if defined(THREADED_RTS)
358 scavenge_capability_mut_Lists1(&capabilities[n]);
360 scavenge_capability_mut_lists(&capabilities[n]);
364 scavenge_capability_mut_lists(&capabilities[gct->thread_index]);
367 // follow roots from the CAF list (used by GHCi)
369 markCAFs(mark_root, gct);
371 // follow all the roots that the application knows about.
373 markSomeCapabilities(mark_root, gct, gct->thread_index, n_gc_threads,
374 rtsTrue/*prune sparks*/);
376 #if defined(RTS_USER_SIGNALS)
377 // mark the signal handlers (signals should be already blocked)
378 markSignalHandlers(mark_root, gct);
381 // Mark the weak pointer list, and prepare to detect dead weak pointers.
385 // Mark the stable pointer table.
386 markStablePtrTable(mark_root, gct);
388 /* -------------------------------------------------------------------------
389 * Repeatedly scavenge all the areas we know about until there's no
390 * more scavenging to be done.
394 scavenge_until_all_done();
395 // The other threads are now stopped. We might recurse back to
396 // here, but from now on this is the only thread.
398 // must be last... invariant is that everything is fully
399 // scavenged at this point.
400 if (traverseWeakPtrList()) { // returns rtsTrue if evaced something
405 // If we get to here, there's really nothing left to do.
409 shutdown_gc_threads(n_gc_threads, gct->thread_index);
411 // Now see which stable names are still alive.
415 if (n_gc_threads == 1) {
416 for (n = 0; n < n_capabilities; n++) {
417 pruneSparkQueue(&capabilities[n]);
420 pruneSparkQueue(&capabilities[gct->thread_index]);
425 // We call processHeapClosureForDead() on every closure destroyed during
426 // the current garbage collection, so we invoke LdvCensusForDead().
427 if (RtsFlags.ProfFlags.doHeapProfile == HEAP_BY_LDV
428 || RtsFlags.ProfFlags.bioSelector != NULL)
432 // NO MORE EVACUATION AFTER THIS POINT!
434 // Two-space collector: free the old to-space.
435 // g0->old_blocks is the old nursery
436 // g0->blocks is to-space from the previous GC
437 if (RtsFlags.GcFlags.generations == 1) {
438 if (g0->blocks != NULL) {
439 freeChain(g0->blocks);
444 // For each workspace, in each thread, move the copied blocks to the step
450 for (t = 0; t < n_gc_threads; t++) {
453 for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
456 // Push the final block
458 push_scanned_block(ws->todo_bd, ws);
461 ASSERT(gct->scan_bd == NULL);
462 ASSERT(countBlocks(ws->scavd_list) == ws->n_scavd_blocks);
465 for (bd = ws->scavd_list; bd != NULL; bd = bd->link) {
466 ws->gen->n_words += bd->free - bd->start;
470 prev->link = ws->gen->blocks;
471 ws->gen->blocks = ws->scavd_list;
473 ws->gen->n_blocks += ws->n_scavd_blocks;
477 // Add all the partial blocks *after* we've added all the full
478 // blocks. This is so that we can grab the partial blocks back
479 // again and try to fill them up in the next GC.
480 for (t = 0; t < n_gc_threads; t++) {
483 for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
487 for (bd = ws->part_list; bd != NULL; bd = next) {
489 if (bd->free == bd->start) {
491 ws->part_list = next;
498 ws->gen->n_words += bd->free - bd->start;
503 prev->link = ws->gen->blocks;
504 ws->gen->blocks = ws->part_list;
506 ws->gen->n_blocks += ws->n_part_blocks;
508 ASSERT(countBlocks(ws->gen->blocks) == ws->gen->n_blocks);
509 ASSERT(countOccupied(ws->gen->blocks) == ws->gen->n_words);
514 // Finally: compact or sweep the oldest generation.
515 if (major_gc && oldest_gen->mark) {
516 if (oldest_gen->compact)
517 compact(gct->scavenged_static_objects);
522 /* run through all the generations/steps and tidy up
529 for (i=0; i < n_gc_threads; i++) {
530 if (n_gc_threads > 1) {
531 debugTrace(DEBUG_gc,"thread %d:", i);
532 debugTrace(DEBUG_gc," copied %ld", gc_threads[i]->copied * sizeof(W_));
533 debugTrace(DEBUG_gc," scanned %ld", gc_threads[i]->scanned * sizeof(W_));
534 debugTrace(DEBUG_gc," any_work %ld", gc_threads[i]->any_work);
535 debugTrace(DEBUG_gc," no_work %ld", gc_threads[i]->no_work);
536 debugTrace(DEBUG_gc," scav_find_work %ld", gc_threads[i]->scav_find_work);
538 copied += gc_threads[i]->copied;
539 max_copied = stg_max(gc_threads[i]->copied, max_copied);
541 if (n_gc_threads == 1) {
549 for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
552 generations[g].collections++; // for stats
553 if (n_gc_threads > 1) generations[g].par_collections++;
556 // Count the mutable list as bytes "copied" for the purposes of
557 // stats. Every mutable list is copied during every GC.
559 nat mut_list_size = 0;
560 for (bd = generations[g].mut_list; bd != NULL; bd = bd->link) {
561 mut_list_size += bd->free - bd->start;
563 for (n = 0; n < n_capabilities; n++) {
564 for (bd = capabilities[n].mut_lists[g];
565 bd != NULL; bd = bd->link) {
566 mut_list_size += bd->free - bd->start;
569 copied += mut_list_size;
572 "mut_list_size: %lu (%d vars, %d arrays, %d MVARs, %d others)",
573 (unsigned long)(mut_list_size * sizeof(W_)),
574 mutlist_MUTVARS, mutlist_MUTARRS, mutlist_MVARS, mutlist_OTHERS);
578 gen = &generations[g];
580 // for generations we collected...
583 /* free old memory and shift to-space into from-space for all
584 * the collected steps (except the allocation area). These
585 * freed blocks will probaby be quickly recycled.
589 // tack the new blocks on the end of the existing blocks
590 if (gen->old_blocks != NULL) {
593 for (bd = gen->old_blocks; bd != NULL; bd = next) {
597 if (!(bd->flags & BF_MARKED))
600 gen->old_blocks = next;
609 gen->n_words += bd->free - bd->start;
611 // NB. this step might not be compacted next
612 // time, so reset the BF_MARKED flags.
613 // They are set before GC if we're going to
614 // compact. (search for BF_MARKED above).
615 bd->flags &= ~BF_MARKED;
617 // between GCs, all blocks in the heap except
618 // for the nursery have the BF_EVACUATED flag set.
619 bd->flags |= BF_EVACUATED;
626 prev->link = gen->blocks;
627 gen->blocks = gen->old_blocks;
630 // add the new blocks to the block tally
631 gen->n_blocks += gen->n_old_blocks;
632 ASSERT(countBlocks(gen->blocks) == gen->n_blocks);
633 ASSERT(countOccupied(gen->blocks) == gen->n_words);
637 freeChain(gen->old_blocks);
640 gen->old_blocks = NULL;
641 gen->n_old_blocks = 0;
643 /* LARGE OBJECTS. The current live large objects are chained on
644 * scavenged_large, having been moved during garbage
645 * collection from large_objects. Any objects left on the
646 * large_objects list are therefore dead, so we free them here.
648 freeChain(gen->large_objects);
649 gen->large_objects = gen->scavenged_large_objects;
650 gen->n_large_blocks = gen->n_scavenged_large_blocks;
651 gen->n_new_large_blocks = 0;
652 ASSERT(countBlocks(gen->large_objects) == gen->n_large_blocks);
654 else // for generations > N
656 /* For older generations, we need to append the
657 * scavenged_large_object list (i.e. large objects that have been
658 * promoted during this GC) to the large_object list for that step.
660 for (bd = gen->scavenged_large_objects; bd; bd = next) {
662 dbl_link_onto(bd, &gen->large_objects);
665 // add the new blocks we promoted during this GC
666 gen->n_large_blocks += gen->n_scavenged_large_blocks;
667 ASSERT(countBlocks(gen->large_objects) == gen->n_large_blocks);
669 } // for all generations
671 // update the max size of older generations after a major GC
672 resize_generations();
674 // Calculate the amount of live data for stats.
675 live = calcLiveWords();
677 // Free the small objects allocated via allocate(), since this will
678 // all have been copied into G0S1 now.
679 alloc_blocks_lim = RtsFlags.GcFlags.minAllocAreaSize;
681 // Start a new pinned_object_block
682 for (n = 0; n < n_capabilities; n++) {
683 capabilities[n].pinned_object_block = NULL;
686 // Free the mark stack.
687 if (mark_stack_top_bd != NULL) {
688 debugTrace(DEBUG_gc, "mark stack: %d blocks",
689 countBlocks(mark_stack_top_bd));
690 freeChain(mark_stack_top_bd);
694 for (g = 0; g <= N; g++) {
695 gen = &generations[g];
696 if (gen->bitmap != NULL) {
697 freeGroup(gen->bitmap);
704 // mark the garbage collected CAFs as dead
705 #if 0 && defined(DEBUG) // doesn't work at the moment
706 if (major_gc) { gcCAFs(); }
710 // resetStaticObjectForRetainerProfiling() must be called before
712 if (n_gc_threads > 1) {
713 barf("profiling is currently broken with multi-threaded GC");
714 // ToDo: fix the gct->scavenged_static_objects below
716 resetStaticObjectForRetainerProfiling(gct->scavenged_static_objects);
719 // zero the scavenged static object list
722 for (i = 0; i < n_gc_threads; i++) {
723 zero_static_object_list(gc_threads[i]->scavenged_static_objects);
730 // send exceptions to any threads which were about to die
732 resurrectThreads(resurrected_threads);
735 // Update the stable pointer hash table.
736 updateStablePtrTable(major_gc);
738 // unlock the StablePtr table. Must be before scheduleFinalizers(),
739 // because a finalizer may call hs_free_fun_ptr() or
740 // hs_free_stable_ptr(), both of which access the StablePtr table.
743 // Start any pending finalizers. Must be after
744 // updateStablePtrTable() and stablePtrPostGC() (see #4221).
746 scheduleFinalizers(cap, old_weak_ptr_list);
751 need = BLOCKS_TO_MBLOCKS(n_alloc_blocks);
752 got = mblocks_allocated;
753 /* If the amount of data remains constant, next major GC we'll
754 require (F+1)*need. We leave (F+2)*need in order to reduce
755 repeated deallocation and reallocation. */
756 need = (RtsFlags.GcFlags.oldGenFactor + 2) * need;
758 returnMemoryToOS(got - need);
762 // check sanity after GC
763 IF_DEBUG(sanity, checkSanity(rtsTrue));
765 // extra GC trace info
766 IF_DEBUG(gc, statDescribeGens());
769 // symbol-table based profiling
770 /* heapCensus(to_blocks); */ /* ToDo */
773 // restore enclosing cost centre
779 // check for memory leaks if DEBUG is on
780 memInventory(DEBUG_gc);
783 #ifdef RTS_GTK_FRONTPANEL
784 if (RtsFlags.GcFlags.frontpanel) {
785 updateFrontPanelAfterGC( N, live );
789 // ok, GC over: tell the stats department what happened.
790 slop = calcLiveBlocks() * BLOCK_SIZE_W - live;
791 stat_endGC(allocated, live, copied, N, max_copied, avg_copied, slop);
793 // Guess which generation we'll collect *next* time
794 initialise_N(force_major_gc);
796 #if defined(RTS_USER_SIGNALS)
797 if (RtsFlags.MiscFlags.install_signal_handlers) {
798 // unblock signals again
799 unblockUserSignals();
808 /* -----------------------------------------------------------------------------
809 Figure out which generation to collect, initialise N and major_gc.
811 Also returns the total number of blocks in generations that will be
813 -------------------------------------------------------------------------- */
816 initialise_N (rtsBool force_major_gc)
819 nat blocks, blocks_total;
824 if (force_major_gc) {
825 N = RtsFlags.GcFlags.generations - 1;
830 for (g = RtsFlags.GcFlags.generations - 1; g >= 0; g--) {
832 blocks = generations[g].n_words / BLOCK_SIZE_W
833 + generations[g].n_large_blocks;
835 if (blocks >= generations[g].max_blocks) {
839 blocks_total += blocks;
843 blocks_total += countNurseryBlocks();
845 major_gc = (N == RtsFlags.GcFlags.generations-1);
849 /* -----------------------------------------------------------------------------
850 Initialise the gc_thread structures.
851 -------------------------------------------------------------------------- */
853 #define GC_THREAD_INACTIVE 0
854 #define GC_THREAD_STANDING_BY 1
855 #define GC_THREAD_RUNNING 2
856 #define GC_THREAD_WAITING_TO_CONTINUE 3
859 new_gc_thread (nat n, gc_thread *t)
866 initSpinLock(&t->gc_spin);
867 initSpinLock(&t->mut_spin);
868 ACQUIRE_SPIN_LOCK(&t->gc_spin);
869 t->wakeup = GC_THREAD_INACTIVE; // starts true, so we can wait for the
870 // thread to start up, see wakeup_gc_threads
874 t->free_blocks = NULL;
883 for (g = 0; g < RtsFlags.GcFlags.generations; g++)
886 ws->gen = &generations[g];
887 ASSERT(g == ws->gen->no);
891 ws->todo_q = newWSDeque(128);
892 ws->todo_overflow = NULL;
893 ws->n_todo_overflow = 0;
895 ws->part_list = NULL;
896 ws->n_part_blocks = 0;
898 ws->scavd_list = NULL;
899 ws->n_scavd_blocks = 0;
907 if (gc_threads == NULL) {
908 #if defined(THREADED_RTS)
910 gc_threads = stgMallocBytes (RtsFlags.ParFlags.nNodes *
914 for (i = 0; i < RtsFlags.ParFlags.nNodes; i++) {
916 stgMallocBytes(sizeof(gc_thread) +
917 RtsFlags.GcFlags.generations * sizeof(gen_workspace),
920 new_gc_thread(i, gc_threads[i]);
923 gc_threads = stgMallocBytes (sizeof(gc_thread*),"alloc_gc_threads");
925 new_gc_thread(0,gc_threads[0]);
934 if (gc_threads != NULL) {
935 #if defined(THREADED_RTS)
937 for (i = 0; i < n_capabilities; i++) {
938 for (g = 0; g < RtsFlags.GcFlags.generations; g++)
940 freeWSDeque(gc_threads[i]->gens[g].todo_q);
942 stgFree (gc_threads[i]);
944 stgFree (gc_threads);
946 for (g = 0; g < RtsFlags.GcFlags.generations; g++)
948 freeWSDeque(gc_threads[0]->gens[g].todo_q);
950 stgFree (gc_threads);
956 /* ----------------------------------------------------------------------------
958 ------------------------------------------------------------------------- */
960 static volatile StgWord gc_running_threads;
966 new = atomic_inc(&gc_running_threads);
967 ASSERT(new <= n_gc_threads);
974 ASSERT(gc_running_threads != 0);
975 return atomic_dec(&gc_running_threads);
988 // scavenge objects in compacted generation
989 if (mark_stack_bd != NULL && !mark_stack_empty()) {
993 // Check for global work in any step. We don't need to check for
994 // local work, because we have already exited scavenge_loop(),
995 // which means there is no local work for this thread.
996 for (g = 0; g < (int)RtsFlags.GcFlags.generations; g++) {
998 if (ws->todo_large_objects) return rtsTrue;
999 if (!looksEmptyWSDeque(ws->todo_q)) return rtsTrue;
1000 if (ws->todo_overflow) return rtsTrue;
1003 #if defined(THREADED_RTS)
1004 if (work_stealing) {
1006 // look for work to steal
1007 for (n = 0; n < n_gc_threads; n++) {
1008 if (n == gct->thread_index) continue;
1009 for (g = RtsFlags.GcFlags.generations-1; g >= 0; g--) {
1010 ws = &gc_threads[n]->gens[g];
1011 if (!looksEmptyWSDeque(ws->todo_q)) return rtsTrue;
1018 #if defined(THREADED_RTS)
1026 scavenge_until_all_done (void)
1032 traceEventGcWork(&capabilities[gct->thread_index]);
1034 #if defined(THREADED_RTS)
1035 if (n_gc_threads > 1) {
1044 // scavenge_loop() only exits when there's no work to do
1047 traceEventGcIdle(&capabilities[gct->thread_index]);
1049 debugTrace(DEBUG_gc, "%d GC threads still running", r);
1051 while (gc_running_threads != 0) {
1057 // any_work() does not remove the work from the queue, it
1058 // just checks for the presence of work. If we find any,
1059 // then we increment gc_running_threads and go back to
1060 // scavenge_loop() to perform any pending work.
1063 traceEventGcDone(&capabilities[gct->thread_index]);
1066 #if defined(THREADED_RTS)
1069 gcWorkerThread (Capability *cap)
1071 gc_thread *saved_gct;
1073 // necessary if we stole a callee-saves register for gct:
1076 gct = gc_threads[cap->no];
1077 gct->id = osThreadId();
1079 // Wait until we're told to wake up
1080 RELEASE_SPIN_LOCK(&gct->mut_spin);
1081 gct->wakeup = GC_THREAD_STANDING_BY;
1082 debugTrace(DEBUG_gc, "GC thread %d standing by...", gct->thread_index);
1083 ACQUIRE_SPIN_LOCK(&gct->gc_spin);
1086 // start performance counters in this thread...
1087 if (gct->papi_events == -1) {
1088 papi_init_eventset(&gct->papi_events);
1090 papi_thread_start_gc1_count(gct->papi_events);
1093 // Every thread evacuates some roots.
1095 markSomeCapabilities(mark_root, gct, gct->thread_index, n_gc_threads,
1096 rtsTrue/*prune sparks*/);
1097 scavenge_capability_mut_lists(&capabilities[gct->thread_index]);
1099 scavenge_until_all_done();
1102 // Now that the whole heap is marked, we discard any sparks that
1103 // were found to be unreachable. The main GC thread is currently
1104 // marking heap reachable via weak pointers, so it is
1105 // non-deterministic whether a spark will be retained if it is
1106 // only reachable via weak pointers. To fix this problem would
1107 // require another GC barrier, which is too high a price.
1108 pruneSparkQueue(cap);
1112 // count events in this thread towards the GC totals
1113 papi_thread_stop_gc1_count(gct->papi_events);
1116 // Wait until we're told to continue
1117 RELEASE_SPIN_LOCK(&gct->gc_spin);
1118 gct->wakeup = GC_THREAD_WAITING_TO_CONTINUE;
1119 debugTrace(DEBUG_gc, "GC thread %d waiting to continue...",
1121 ACQUIRE_SPIN_LOCK(&gct->mut_spin);
1122 debugTrace(DEBUG_gc, "GC thread %d on my way...", gct->thread_index);
1129 #if defined(THREADED_RTS)
1132 waitForGcThreads (Capability *cap USED_IF_THREADS)
1134 const nat n_threads = RtsFlags.ParFlags.nNodes;
1135 const nat me = cap->no;
1137 rtsBool retry = rtsTrue;
1140 for (i=0; i < n_threads; i++) {
1141 if (i == me) continue;
1142 if (gc_threads[i]->wakeup != GC_THREAD_STANDING_BY) {
1143 prodCapability(&capabilities[i], cap->running_task);
1146 for (j=0; j < 10; j++) {
1148 for (i=0; i < n_threads; i++) {
1149 if (i == me) continue;
1151 setContextSwitches();
1152 if (gc_threads[i]->wakeup != GC_THREAD_STANDING_BY) {
1162 #endif // THREADED_RTS
1165 start_gc_threads (void)
1167 #if defined(THREADED_RTS)
1168 gc_running_threads = 0;
1173 wakeup_gc_threads (nat n_threads USED_IF_THREADS, nat me USED_IF_THREADS)
1175 #if defined(THREADED_RTS)
1177 for (i=0; i < n_threads; i++) {
1178 if (i == me) continue;
1180 debugTrace(DEBUG_gc, "waking up gc thread %d", i);
1181 if (gc_threads[i]->wakeup != GC_THREAD_STANDING_BY) barf("wakeup_gc_threads");
1183 gc_threads[i]->wakeup = GC_THREAD_RUNNING;
1184 ACQUIRE_SPIN_LOCK(&gc_threads[i]->mut_spin);
1185 RELEASE_SPIN_LOCK(&gc_threads[i]->gc_spin);
1190 // After GC is complete, we must wait for all GC threads to enter the
1191 // standby state, otherwise they may still be executing inside
1192 // any_work(), and may even remain awake until the next GC starts.
1194 shutdown_gc_threads (nat n_threads USED_IF_THREADS, nat me USED_IF_THREADS)
1196 #if defined(THREADED_RTS)
1198 for (i=0; i < n_threads; i++) {
1199 if (i == me) continue;
1200 while (gc_threads[i]->wakeup != GC_THREAD_WAITING_TO_CONTINUE) { write_barrier(); }
1205 #if defined(THREADED_RTS)
1207 releaseGCThreads (Capability *cap USED_IF_THREADS)
1209 const nat n_threads = RtsFlags.ParFlags.nNodes;
1210 const nat me = cap->no;
1212 for (i=0; i < n_threads; i++) {
1213 if (i == me) continue;
1214 if (gc_threads[i]->wakeup != GC_THREAD_WAITING_TO_CONTINUE)
1215 barf("releaseGCThreads");
1217 gc_threads[i]->wakeup = GC_THREAD_INACTIVE;
1218 ACQUIRE_SPIN_LOCK(&gc_threads[i]->gc_spin);
1219 RELEASE_SPIN_LOCK(&gc_threads[i]->mut_spin);
1224 /* ----------------------------------------------------------------------------
1225 Initialise a generation that is to be collected
1226 ------------------------------------------------------------------------- */
1229 init_collected_gen (nat g, nat n_threads)
1236 // Throw away the current mutable list. Invariant: the mutable
1237 // list always has at least one block; this means we can avoid a
1238 // check for NULL in recordMutable().
1240 freeChain(generations[g].mut_list);
1241 generations[g].mut_list = allocBlock();
1242 for (i = 0; i < n_capabilities; i++) {
1243 freeChain(capabilities[i].mut_lists[g]);
1244 capabilities[i].mut_lists[g] = allocBlock();
1248 gen = &generations[g];
1249 ASSERT(gen->no == g);
1251 // we'll construct a new list of threads in this step
1252 // during GC, throw away the current list.
1253 gen->old_threads = gen->threads;
1254 gen->threads = END_TSO_QUEUE;
1256 // deprecate the existing blocks
1257 gen->old_blocks = gen->blocks;
1258 gen->n_old_blocks = gen->n_blocks;
1262 gen->live_estimate = 0;
1264 // initialise the large object queues.
1265 gen->scavenged_large_objects = NULL;
1266 gen->n_scavenged_large_blocks = 0;
1268 // mark the small objects as from-space
1269 for (bd = gen->old_blocks; bd; bd = bd->link) {
1270 bd->flags &= ~BF_EVACUATED;
1273 // mark the large objects as from-space
1274 for (bd = gen->large_objects; bd; bd = bd->link) {
1275 bd->flags &= ~BF_EVACUATED;
1278 // for a compacted generation, we need to allocate the bitmap
1280 nat bitmap_size; // in bytes
1281 bdescr *bitmap_bdescr;
1284 bitmap_size = gen->n_old_blocks * BLOCK_SIZE / (sizeof(W_)*BITS_PER_BYTE);
1286 if (bitmap_size > 0) {
1287 bitmap_bdescr = allocGroup((lnat)BLOCK_ROUND_UP(bitmap_size)
1289 gen->bitmap = bitmap_bdescr;
1290 bitmap = bitmap_bdescr->start;
1292 debugTrace(DEBUG_gc, "bitmap_size: %d, bitmap: %p",
1293 bitmap_size, bitmap);
1295 // don't forget to fill it with zeros!
1296 memset(bitmap, 0, bitmap_size);
1298 // For each block in this step, point to its bitmap from the
1299 // block descriptor.
1300 for (bd=gen->old_blocks; bd != NULL; bd = bd->link) {
1301 bd->u.bitmap = bitmap;
1302 bitmap += BLOCK_SIZE_W / (sizeof(W_)*BITS_PER_BYTE);
1304 // Also at this point we set the BF_MARKED flag
1305 // for this block. The invariant is that
1306 // BF_MARKED is always unset, except during GC
1307 // when it is set on those blocks which will be
1309 if (!(bd->flags & BF_FRAGMENTED)) {
1310 bd->flags |= BF_MARKED;
1313 // BF_SWEPT should be marked only for blocks that are being
1314 // collected in sweep()
1315 bd->flags &= ~BF_SWEPT;
1320 // For each GC thread, for each step, allocate a "todo" block to
1321 // store evacuated objects to be scavenged, and a block to store
1322 // evacuated objects that do not need to be scavenged.
1323 for (t = 0; t < n_threads; t++) {
1324 ws = &gc_threads[t]->gens[g];
1326 ws->todo_large_objects = NULL;
1328 ws->part_list = NULL;
1329 ws->n_part_blocks = 0;
1331 // allocate the first to-space block; extra blocks will be
1332 // chained on as necessary.
1334 ASSERT(looksEmptyWSDeque(ws->todo_q));
1335 alloc_todo_block(ws,0);
1337 ws->todo_overflow = NULL;
1338 ws->n_todo_overflow = 0;
1340 ws->scavd_list = NULL;
1341 ws->n_scavd_blocks = 0;
1346 /* ----------------------------------------------------------------------------
1347 Initialise a generation that is *not* to be collected
1348 ------------------------------------------------------------------------- */
1351 init_uncollected_gen (nat g, nat threads)
1358 // save the current mutable lists for this generation, and
1359 // allocate a fresh block for each one. We'll traverse these
1360 // mutable lists as roots early on in the GC.
1361 generations[g].saved_mut_list = generations[g].mut_list;
1362 generations[g].mut_list = allocBlock();
1363 for (n = 0; n < n_capabilities; n++) {
1364 capabilities[n].saved_mut_lists[g] = capabilities[n].mut_lists[g];
1365 capabilities[n].mut_lists[g] = allocBlock();
1368 gen = &generations[g];
1370 gen->scavenged_large_objects = NULL;
1371 gen->n_scavenged_large_blocks = 0;
1373 for (t = 0; t < threads; t++) {
1374 ws = &gc_threads[t]->gens[g];
1376 ASSERT(looksEmptyWSDeque(ws->todo_q));
1377 ws->todo_large_objects = NULL;
1379 ws->part_list = NULL;
1380 ws->n_part_blocks = 0;
1382 ws->scavd_list = NULL;
1383 ws->n_scavd_blocks = 0;
1385 // If the block at the head of the list in this generation
1386 // is less than 3/4 full, then use it as a todo block.
1387 if (gen->blocks && isPartiallyFull(gen->blocks))
1389 ws->todo_bd = gen->blocks;
1390 ws->todo_free = ws->todo_bd->free;
1391 ws->todo_lim = ws->todo_bd->start + BLOCK_SIZE_W;
1392 gen->blocks = gen->blocks->link;
1394 gen->n_words -= ws->todo_bd->free - ws->todo_bd->start;
1395 ws->todo_bd->link = NULL;
1396 // we must scan from the current end point.
1397 ws->todo_bd->u.scan = ws->todo_bd->free;
1402 alloc_todo_block(ws,0);
1406 // deal out any more partial blocks to the threads' part_lists
1408 while (gen->blocks && isPartiallyFull(gen->blocks))
1411 gen->blocks = bd->link;
1412 ws = &gc_threads[t]->gens[g];
1413 bd->link = ws->part_list;
1415 ws->n_part_blocks += 1;
1416 bd->u.scan = bd->free;
1418 gen->n_words -= bd->free - bd->start;
1420 if (t == n_gc_threads) t = 0;
1424 /* -----------------------------------------------------------------------------
1425 Initialise a gc_thread before GC
1426 -------------------------------------------------------------------------- */
1429 init_gc_thread (gc_thread *t)
1431 t->static_objects = END_OF_STATIC_LIST;
1432 t->scavenged_static_objects = END_OF_STATIC_LIST;
1434 t->mut_lists = capabilities[t->thread_index].mut_lists;
1436 t->failed_to_evac = rtsFalse;
1437 t->eager_promotion = rtsTrue;
1438 t->thunk_selector_depth = 0;
1443 t->scav_find_work = 0;
1446 /* -----------------------------------------------------------------------------
1447 Function we pass to evacuate roots.
1448 -------------------------------------------------------------------------- */
1451 mark_root(void *user USED_IF_THREADS, StgClosure **root)
1453 // we stole a register for gct, but this function is called from
1454 // *outside* the GC where the register variable is not in effect,
1455 // so we need to save and restore it here. NB. only call
1456 // mark_root() from the main GC thread, otherwise gct will be
1458 gc_thread *saved_gct;
1467 /* -----------------------------------------------------------------------------
1468 Initialising the static object & mutable lists
1469 -------------------------------------------------------------------------- */
1472 zero_static_object_list(StgClosure* first_static)
1476 const StgInfoTable *info;
1478 for (p = first_static; p != END_OF_STATIC_LIST; p = link) {
1480 link = *STATIC_LINK(info, p);
1481 *STATIC_LINK(info,p) = NULL;
1485 /* ----------------------------------------------------------------------------
1486 Reset the sizes of the older generations when we do a major
1489 CURRENT STRATEGY: make all generations except zero the same size.
1490 We have to stay within the maximum heap size, and leave a certain
1491 percentage of the maximum heap size available to allocate into.
1492 ------------------------------------------------------------------------- */
1495 resize_generations (void)
1499 if (major_gc && RtsFlags.GcFlags.generations > 1) {
1500 nat live, size, min_alloc, words;
1501 const nat max = RtsFlags.GcFlags.maxHeapSize;
1502 const nat gens = RtsFlags.GcFlags.generations;
1504 // live in the oldest generations
1505 if (oldest_gen->live_estimate != 0) {
1506 words = oldest_gen->live_estimate;
1508 words = oldest_gen->n_words;
1510 live = (words + BLOCK_SIZE_W - 1) / BLOCK_SIZE_W +
1511 oldest_gen->n_large_blocks;
1513 // default max size for all generations except zero
1514 size = stg_max(live * RtsFlags.GcFlags.oldGenFactor,
1515 RtsFlags.GcFlags.minOldGenSize);
1517 if (RtsFlags.GcFlags.heapSizeSuggestionAuto) {
1518 RtsFlags.GcFlags.heapSizeSuggestion = size;
1521 // minimum size for generation zero
1522 min_alloc = stg_max((RtsFlags.GcFlags.pcFreeHeap * max) / 200,
1523 RtsFlags.GcFlags.minAllocAreaSize);
1525 // Auto-enable compaction when the residency reaches a
1526 // certain percentage of the maximum heap size (default: 30%).
1527 if (RtsFlags.GcFlags.compact ||
1529 oldest_gen->n_blocks >
1530 (RtsFlags.GcFlags.compactThreshold * max) / 100)) {
1531 oldest_gen->mark = 1;
1532 oldest_gen->compact = 1;
1533 // debugBelch("compaction: on\n", live);
1535 oldest_gen->mark = 0;
1536 oldest_gen->compact = 0;
1537 // debugBelch("compaction: off\n", live);
1540 if (RtsFlags.GcFlags.sweep) {
1541 oldest_gen->mark = 1;
1544 // if we're going to go over the maximum heap size, reduce the
1545 // size of the generations accordingly. The calculation is
1546 // different if compaction is turned on, because we don't need
1547 // to double the space required to collect the old generation.
1550 // this test is necessary to ensure that the calculations
1551 // below don't have any negative results - we're working
1552 // with unsigned values here.
1553 if (max < min_alloc) {
1557 if (oldest_gen->compact) {
1558 if ( (size + (size - 1) * (gens - 2) * 2) + min_alloc > max ) {
1559 size = (max - min_alloc) / ((gens - 1) * 2 - 1);
1562 if ( (size * (gens - 1) * 2) + min_alloc > max ) {
1563 size = (max - min_alloc) / ((gens - 1) * 2);
1573 debugBelch("live: %d, min_alloc: %d, size : %d, max = %d\n", live,
1574 min_alloc, size, max);
1577 for (g = 0; g < gens; g++) {
1578 generations[g].max_blocks = size;
1583 /* -----------------------------------------------------------------------------
1584 Calculate the new size of the nursery, and resize it.
1585 -------------------------------------------------------------------------- */
1588 resize_nursery (void)
1590 const lnat min_nursery = RtsFlags.GcFlags.minAllocAreaSize * n_capabilities;
1592 if (RtsFlags.GcFlags.generations == 1)
1593 { // Two-space collector:
1596 /* set up a new nursery. Allocate a nursery size based on a
1597 * function of the amount of live data (by default a factor of 2)
1598 * Use the blocks from the old nursery if possible, freeing up any
1601 * If we get near the maximum heap size, then adjust our nursery
1602 * size accordingly. If the nursery is the same size as the live
1603 * data (L), then we need 3L bytes. We can reduce the size of the
1604 * nursery to bring the required memory down near 2L bytes.
1606 * A normal 2-space collector would need 4L bytes to give the same
1607 * performance we get from 3L bytes, reducing to the same
1608 * performance at 2L bytes.
1610 blocks = generations[0].n_blocks;
1612 if ( RtsFlags.GcFlags.maxHeapSize != 0 &&
1613 blocks * RtsFlags.GcFlags.oldGenFactor * 2 >
1614 RtsFlags.GcFlags.maxHeapSize )
1616 long adjusted_blocks; // signed on purpose
1619 adjusted_blocks = (RtsFlags.GcFlags.maxHeapSize - 2 * blocks);
1621 debugTrace(DEBUG_gc, "near maximum heap size of 0x%x blocks, blocks = %d, adjusted to %ld",
1622 RtsFlags.GcFlags.maxHeapSize, blocks, adjusted_blocks);
1624 pc_free = adjusted_blocks * 100 / RtsFlags.GcFlags.maxHeapSize;
1625 if (pc_free < RtsFlags.GcFlags.pcFreeHeap) /* might even * be < 0 */
1629 blocks = adjusted_blocks;
1633 blocks *= RtsFlags.GcFlags.oldGenFactor;
1634 if (blocks < min_nursery)
1636 blocks = min_nursery;
1639 resizeNurseries(blocks);
1641 else // Generational collector
1644 * If the user has given us a suggested heap size, adjust our
1645 * allocation area to make best use of the memory available.
1647 if (RtsFlags.GcFlags.heapSizeSuggestion)
1650 const nat needed = calcNeeded(); // approx blocks needed at next GC
1652 /* Guess how much will be live in generation 0 step 0 next time.
1653 * A good approximation is obtained by finding the
1654 * percentage of g0 that was live at the last minor GC.
1656 * We have an accurate figure for the amount of copied data in
1657 * 'copied', but we must convert this to a number of blocks, with
1658 * a small adjustment for estimated slop at the end of a block
1663 g0_pcnt_kept = ((copied / (BLOCK_SIZE_W - 10)) * 100)
1664 / countNurseryBlocks();
1667 /* Estimate a size for the allocation area based on the
1668 * information available. We might end up going slightly under
1669 * or over the suggested heap size, but we should be pretty
1672 * Formula: suggested - needed
1673 * ----------------------------
1674 * 1 + g0_pcnt_kept/100
1676 * where 'needed' is the amount of memory needed at the next
1677 * collection for collecting all gens except g0.
1680 (((long)RtsFlags.GcFlags.heapSizeSuggestion - (long)needed) * 100) /
1681 (100 + (long)g0_pcnt_kept);
1683 if (blocks < (long)min_nursery) {
1684 blocks = min_nursery;
1687 resizeNurseries((nat)blocks);
1691 // we might have added extra large blocks to the nursery, so
1692 // resize back to minAllocAreaSize again.
1693 resizeNurseriesFixed(RtsFlags.GcFlags.minAllocAreaSize);
1698 /* -----------------------------------------------------------------------------
1699 Sanity code for CAF garbage collection.
1701 With DEBUG turned on, we manage a CAF list in addition to the SRT
1702 mechanism. After GC, we run down the CAF list and blackhole any
1703 CAFs which have been garbage collected. This means we get an error
1704 whenever the program tries to enter a garbage collected CAF.
1706 Any garbage collected CAFs are taken off the CAF list at the same
1708 -------------------------------------------------------------------------- */
1710 #if 0 && defined(DEBUG)
1717 const StgInfoTable *info;
1728 ASSERT(info->type == IND_STATIC);
1730 if (STATIC_LINK(info,p) == NULL) {
1731 debugTrace(DEBUG_gccafs, "CAF gc'd at 0x%04lx", (long)p);
1733 SET_INFO(p,&stg_BLACKHOLE_info);
1734 p = STATIC_LINK2(info,p);
1738 pp = &STATIC_LINK2(info,p);
1745 debugTrace(DEBUG_gccafs, "%d CAFs live", i);