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(rtsFalse/* don't count the nursery yet */);
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_words = 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 // Start a new pinned_object_block
678 for (n = 0; n < n_capabilities; n++) {
679 capabilities[n].pinned_object_block = NULL;
682 // Free the mark stack.
683 if (mark_stack_top_bd != NULL) {
684 debugTrace(DEBUG_gc, "mark stack: %d blocks",
685 countBlocks(mark_stack_top_bd));
686 freeChain(mark_stack_top_bd);
690 for (g = 0; g <= N; g++) {
691 gen = &generations[g];
692 if (gen->bitmap != NULL) {
693 freeGroup(gen->bitmap);
698 // Reset the nursery: make the blocks empty
699 allocated += clearNurseries();
705 // mark the garbage collected CAFs as dead
706 #if 0 && defined(DEBUG) // doesn't work at the moment
707 if (major_gc) { gcCAFs(); }
711 // resetStaticObjectForRetainerProfiling() must be called before
713 if (n_gc_threads > 1) {
714 barf("profiling is currently broken with multi-threaded GC");
715 // ToDo: fix the gct->scavenged_static_objects below
717 resetStaticObjectForRetainerProfiling(gct->scavenged_static_objects);
720 // zero the scavenged static object list
723 for (i = 0; i < n_gc_threads; i++) {
724 zero_static_object_list(gc_threads[i]->scavenged_static_objects);
728 // send exceptions to any threads which were about to die
730 resurrectThreads(resurrected_threads);
733 // Update the stable pointer hash table.
734 updateStablePtrTable(major_gc);
736 // unlock the StablePtr table. Must be before scheduleFinalizers(),
737 // because a finalizer may call hs_free_fun_ptr() or
738 // hs_free_stable_ptr(), both of which access the StablePtr table.
741 // Start any pending finalizers. Must be after
742 // updateStablePtrTable() and stablePtrPostGC() (see #4221).
744 scheduleFinalizers(cap, old_weak_ptr_list);
749 need = BLOCKS_TO_MBLOCKS(n_alloc_blocks);
750 got = mblocks_allocated;
751 /* If the amount of data remains constant, next major GC we'll
752 require (F+1)*need. We leave (F+2)*need in order to reduce
753 repeated deallocation and reallocation. */
754 need = (RtsFlags.GcFlags.oldGenFactor + 2) * need;
756 returnMemoryToOS(got - need);
760 // check sanity after GC
761 IF_DEBUG(sanity, checkSanity(rtsTrue));
763 // extra GC trace info
764 IF_DEBUG(gc, statDescribeGens());
767 // symbol-table based profiling
768 /* heapCensus(to_blocks); */ /* ToDo */
771 // restore enclosing cost centre
777 // check for memory leaks if DEBUG is on
778 memInventory(DEBUG_gc);
781 #ifdef RTS_GTK_FRONTPANEL
782 if (RtsFlags.GcFlags.frontpanel) {
783 updateFrontPanelAfterGC( N, live );
787 // ok, GC over: tell the stats department what happened.
788 slop = calcLiveBlocks() * BLOCK_SIZE_W - live;
789 stat_endGC(allocated, live, copied, N, max_copied, avg_copied, slop);
791 // Guess which generation we'll collect *next* time
792 initialise_N(force_major_gc);
794 #if defined(RTS_USER_SIGNALS)
795 if (RtsFlags.MiscFlags.install_signal_handlers) {
796 // unblock signals again
797 unblockUserSignals();
806 /* -----------------------------------------------------------------------------
807 Figure out which generation to collect, initialise N and major_gc.
809 Also returns the total number of blocks in generations that will be
811 -------------------------------------------------------------------------- */
814 initialise_N (rtsBool force_major_gc)
817 nat blocks, blocks_total;
822 if (force_major_gc) {
823 N = RtsFlags.GcFlags.generations - 1;
828 for (g = RtsFlags.GcFlags.generations - 1; g >= 0; g--) {
830 blocks = generations[g].n_words / BLOCK_SIZE_W
831 + generations[g].n_large_blocks;
833 if (blocks >= generations[g].max_blocks) {
837 blocks_total += blocks;
841 blocks_total += countNurseryBlocks();
843 major_gc = (N == RtsFlags.GcFlags.generations-1);
847 /* -----------------------------------------------------------------------------
848 Initialise the gc_thread structures.
849 -------------------------------------------------------------------------- */
851 #define GC_THREAD_INACTIVE 0
852 #define GC_THREAD_STANDING_BY 1
853 #define GC_THREAD_RUNNING 2
854 #define GC_THREAD_WAITING_TO_CONTINUE 3
857 new_gc_thread (nat n, gc_thread *t)
864 initSpinLock(&t->gc_spin);
865 initSpinLock(&t->mut_spin);
866 ACQUIRE_SPIN_LOCK(&t->gc_spin);
867 t->wakeup = GC_THREAD_INACTIVE; // starts true, so we can wait for the
868 // thread to start up, see wakeup_gc_threads
872 t->free_blocks = NULL;
881 for (g = 0; g < RtsFlags.GcFlags.generations; g++)
884 ws->gen = &generations[g];
885 ASSERT(g == ws->gen->no);
889 ws->todo_q = newWSDeque(128);
890 ws->todo_overflow = NULL;
891 ws->n_todo_overflow = 0;
893 ws->part_list = NULL;
894 ws->n_part_blocks = 0;
896 ws->scavd_list = NULL;
897 ws->n_scavd_blocks = 0;
905 if (gc_threads == NULL) {
906 #if defined(THREADED_RTS)
908 gc_threads = stgMallocBytes (RtsFlags.ParFlags.nNodes *
912 for (i = 0; i < RtsFlags.ParFlags.nNodes; i++) {
914 stgMallocBytes(sizeof(gc_thread) +
915 RtsFlags.GcFlags.generations * sizeof(gen_workspace),
918 new_gc_thread(i, gc_threads[i]);
921 gc_threads = stgMallocBytes (sizeof(gc_thread*),"alloc_gc_threads");
923 new_gc_thread(0,gc_threads[0]);
932 if (gc_threads != NULL) {
933 #if defined(THREADED_RTS)
935 for (i = 0; i < n_capabilities; i++) {
936 for (g = 0; g < RtsFlags.GcFlags.generations; g++)
938 freeWSDeque(gc_threads[i]->gens[g].todo_q);
940 stgFree (gc_threads[i]);
942 stgFree (gc_threads);
944 for (g = 0; g < RtsFlags.GcFlags.generations; g++)
946 freeWSDeque(gc_threads[0]->gens[g].todo_q);
948 stgFree (gc_threads);
954 /* ----------------------------------------------------------------------------
956 ------------------------------------------------------------------------- */
958 static volatile StgWord gc_running_threads;
964 new = atomic_inc(&gc_running_threads);
965 ASSERT(new <= n_gc_threads);
972 ASSERT(gc_running_threads != 0);
973 return atomic_dec(&gc_running_threads);
986 // scavenge objects in compacted generation
987 if (mark_stack_bd != NULL && !mark_stack_empty()) {
991 // Check for global work in any step. We don't need to check for
992 // local work, because we have already exited scavenge_loop(),
993 // which means there is no local work for this thread.
994 for (g = 0; g < (int)RtsFlags.GcFlags.generations; g++) {
996 if (ws->todo_large_objects) return rtsTrue;
997 if (!looksEmptyWSDeque(ws->todo_q)) return rtsTrue;
998 if (ws->todo_overflow) return rtsTrue;
1001 #if defined(THREADED_RTS)
1002 if (work_stealing) {
1004 // look for work to steal
1005 for (n = 0; n < n_gc_threads; n++) {
1006 if (n == gct->thread_index) continue;
1007 for (g = RtsFlags.GcFlags.generations-1; g >= 0; g--) {
1008 ws = &gc_threads[n]->gens[g];
1009 if (!looksEmptyWSDeque(ws->todo_q)) return rtsTrue;
1016 #if defined(THREADED_RTS)
1024 scavenge_until_all_done (void)
1030 traceEventGcWork(&capabilities[gct->thread_index]);
1032 #if defined(THREADED_RTS)
1033 if (n_gc_threads > 1) {
1042 // scavenge_loop() only exits when there's no work to do
1045 traceEventGcIdle(&capabilities[gct->thread_index]);
1047 debugTrace(DEBUG_gc, "%d GC threads still running", r);
1049 while (gc_running_threads != 0) {
1055 // any_work() does not remove the work from the queue, it
1056 // just checks for the presence of work. If we find any,
1057 // then we increment gc_running_threads and go back to
1058 // scavenge_loop() to perform any pending work.
1061 traceEventGcDone(&capabilities[gct->thread_index]);
1064 #if defined(THREADED_RTS)
1067 gcWorkerThread (Capability *cap)
1069 gc_thread *saved_gct;
1071 // necessary if we stole a callee-saves register for gct:
1074 gct = gc_threads[cap->no];
1075 gct->id = osThreadId();
1077 // Wait until we're told to wake up
1078 RELEASE_SPIN_LOCK(&gct->mut_spin);
1079 gct->wakeup = GC_THREAD_STANDING_BY;
1080 debugTrace(DEBUG_gc, "GC thread %d standing by...", gct->thread_index);
1081 ACQUIRE_SPIN_LOCK(&gct->gc_spin);
1084 // start performance counters in this thread...
1085 if (gct->papi_events == -1) {
1086 papi_init_eventset(&gct->papi_events);
1088 papi_thread_start_gc1_count(gct->papi_events);
1091 // Every thread evacuates some roots.
1093 markSomeCapabilities(mark_root, gct, gct->thread_index, n_gc_threads,
1094 rtsTrue/*prune sparks*/);
1095 scavenge_capability_mut_lists(&capabilities[gct->thread_index]);
1097 scavenge_until_all_done();
1100 // Now that the whole heap is marked, we discard any sparks that
1101 // were found to be unreachable. The main GC thread is currently
1102 // marking heap reachable via weak pointers, so it is
1103 // non-deterministic whether a spark will be retained if it is
1104 // only reachable via weak pointers. To fix this problem would
1105 // require another GC barrier, which is too high a price.
1106 pruneSparkQueue(cap);
1110 // count events in this thread towards the GC totals
1111 papi_thread_stop_gc1_count(gct->papi_events);
1114 // Wait until we're told to continue
1115 RELEASE_SPIN_LOCK(&gct->gc_spin);
1116 gct->wakeup = GC_THREAD_WAITING_TO_CONTINUE;
1117 debugTrace(DEBUG_gc, "GC thread %d waiting to continue...",
1119 ACQUIRE_SPIN_LOCK(&gct->mut_spin);
1120 debugTrace(DEBUG_gc, "GC thread %d on my way...", gct->thread_index);
1127 #if defined(THREADED_RTS)
1130 waitForGcThreads (Capability *cap USED_IF_THREADS)
1132 const nat n_threads = RtsFlags.ParFlags.nNodes;
1133 const nat me = cap->no;
1135 rtsBool retry = rtsTrue;
1138 for (i=0; i < n_threads; i++) {
1139 if (i == me) continue;
1140 if (gc_threads[i]->wakeup != GC_THREAD_STANDING_BY) {
1141 prodCapability(&capabilities[i], cap->running_task);
1144 for (j=0; j < 10; j++) {
1146 for (i=0; i < n_threads; i++) {
1147 if (i == me) continue;
1149 setContextSwitches();
1150 if (gc_threads[i]->wakeup != GC_THREAD_STANDING_BY) {
1160 #endif // THREADED_RTS
1163 start_gc_threads (void)
1165 #if defined(THREADED_RTS)
1166 gc_running_threads = 0;
1171 wakeup_gc_threads (nat n_threads USED_IF_THREADS, nat me USED_IF_THREADS)
1173 #if defined(THREADED_RTS)
1175 for (i=0; i < n_threads; i++) {
1176 if (i == me) continue;
1178 debugTrace(DEBUG_gc, "waking up gc thread %d", i);
1179 if (gc_threads[i]->wakeup != GC_THREAD_STANDING_BY) barf("wakeup_gc_threads");
1181 gc_threads[i]->wakeup = GC_THREAD_RUNNING;
1182 ACQUIRE_SPIN_LOCK(&gc_threads[i]->mut_spin);
1183 RELEASE_SPIN_LOCK(&gc_threads[i]->gc_spin);
1188 // After GC is complete, we must wait for all GC threads to enter the
1189 // standby state, otherwise they may still be executing inside
1190 // any_work(), and may even remain awake until the next GC starts.
1192 shutdown_gc_threads (nat n_threads USED_IF_THREADS, nat me USED_IF_THREADS)
1194 #if defined(THREADED_RTS)
1196 for (i=0; i < n_threads; i++) {
1197 if (i == me) continue;
1198 while (gc_threads[i]->wakeup != GC_THREAD_WAITING_TO_CONTINUE) { write_barrier(); }
1203 #if defined(THREADED_RTS)
1205 releaseGCThreads (Capability *cap USED_IF_THREADS)
1207 const nat n_threads = RtsFlags.ParFlags.nNodes;
1208 const nat me = cap->no;
1210 for (i=0; i < n_threads; i++) {
1211 if (i == me) continue;
1212 if (gc_threads[i]->wakeup != GC_THREAD_WAITING_TO_CONTINUE)
1213 barf("releaseGCThreads");
1215 gc_threads[i]->wakeup = GC_THREAD_INACTIVE;
1216 ACQUIRE_SPIN_LOCK(&gc_threads[i]->gc_spin);
1217 RELEASE_SPIN_LOCK(&gc_threads[i]->mut_spin);
1222 /* ----------------------------------------------------------------------------
1223 Initialise a generation that is to be collected
1224 ------------------------------------------------------------------------- */
1227 init_collected_gen (nat g, nat n_threads)
1234 // Throw away the current mutable list. Invariant: the mutable
1235 // list always has at least one block; this means we can avoid a
1236 // check for NULL in recordMutable().
1238 freeChain(generations[g].mut_list);
1239 generations[g].mut_list = allocBlock();
1240 for (i = 0; i < n_capabilities; i++) {
1241 freeChain(capabilities[i].mut_lists[g]);
1242 capabilities[i].mut_lists[g] = allocBlock();
1246 gen = &generations[g];
1247 ASSERT(gen->no == g);
1249 // we'll construct a new list of threads in this step
1250 // during GC, throw away the current list.
1251 gen->old_threads = gen->threads;
1252 gen->threads = END_TSO_QUEUE;
1254 // deprecate the existing blocks
1255 gen->old_blocks = gen->blocks;
1256 gen->n_old_blocks = gen->n_blocks;
1260 gen->live_estimate = 0;
1262 // initialise the large object queues.
1263 gen->scavenged_large_objects = NULL;
1264 gen->n_scavenged_large_blocks = 0;
1266 // mark the small objects as from-space
1267 for (bd = gen->old_blocks; bd; bd = bd->link) {
1268 bd->flags &= ~BF_EVACUATED;
1271 // mark the large objects as from-space
1272 for (bd = gen->large_objects; bd; bd = bd->link) {
1273 bd->flags &= ~BF_EVACUATED;
1276 // for a compacted generation, we need to allocate the bitmap
1278 nat bitmap_size; // in bytes
1279 bdescr *bitmap_bdescr;
1282 bitmap_size = gen->n_old_blocks * BLOCK_SIZE / (sizeof(W_)*BITS_PER_BYTE);
1284 if (bitmap_size > 0) {
1285 bitmap_bdescr = allocGroup((lnat)BLOCK_ROUND_UP(bitmap_size)
1287 gen->bitmap = bitmap_bdescr;
1288 bitmap = bitmap_bdescr->start;
1290 debugTrace(DEBUG_gc, "bitmap_size: %d, bitmap: %p",
1291 bitmap_size, bitmap);
1293 // don't forget to fill it with zeros!
1294 memset(bitmap, 0, bitmap_size);
1296 // For each block in this step, point to its bitmap from the
1297 // block descriptor.
1298 for (bd=gen->old_blocks; bd != NULL; bd = bd->link) {
1299 bd->u.bitmap = bitmap;
1300 bitmap += BLOCK_SIZE_W / (sizeof(W_)*BITS_PER_BYTE);
1302 // Also at this point we set the BF_MARKED flag
1303 // for this block. The invariant is that
1304 // BF_MARKED is always unset, except during GC
1305 // when it is set on those blocks which will be
1307 if (!(bd->flags & BF_FRAGMENTED)) {
1308 bd->flags |= BF_MARKED;
1311 // BF_SWEPT should be marked only for blocks that are being
1312 // collected in sweep()
1313 bd->flags &= ~BF_SWEPT;
1318 // For each GC thread, for each step, allocate a "todo" block to
1319 // store evacuated objects to be scavenged, and a block to store
1320 // evacuated objects that do not need to be scavenged.
1321 for (t = 0; t < n_threads; t++) {
1322 ws = &gc_threads[t]->gens[g];
1324 ws->todo_large_objects = NULL;
1326 ws->part_list = NULL;
1327 ws->n_part_blocks = 0;
1329 // allocate the first to-space block; extra blocks will be
1330 // chained on as necessary.
1332 ASSERT(looksEmptyWSDeque(ws->todo_q));
1333 alloc_todo_block(ws,0);
1335 ws->todo_overflow = NULL;
1336 ws->n_todo_overflow = 0;
1338 ws->scavd_list = NULL;
1339 ws->n_scavd_blocks = 0;
1344 /* ----------------------------------------------------------------------------
1345 Initialise a generation that is *not* to be collected
1346 ------------------------------------------------------------------------- */
1349 init_uncollected_gen (nat g, nat threads)
1356 // save the current mutable lists for this generation, and
1357 // allocate a fresh block for each one. We'll traverse these
1358 // mutable lists as roots early on in the GC.
1359 generations[g].saved_mut_list = generations[g].mut_list;
1360 generations[g].mut_list = allocBlock();
1361 for (n = 0; n < n_capabilities; n++) {
1362 capabilities[n].saved_mut_lists[g] = capabilities[n].mut_lists[g];
1363 capabilities[n].mut_lists[g] = allocBlock();
1366 gen = &generations[g];
1368 gen->scavenged_large_objects = NULL;
1369 gen->n_scavenged_large_blocks = 0;
1371 for (t = 0; t < threads; t++) {
1372 ws = &gc_threads[t]->gens[g];
1374 ASSERT(looksEmptyWSDeque(ws->todo_q));
1375 ws->todo_large_objects = NULL;
1377 ws->part_list = NULL;
1378 ws->n_part_blocks = 0;
1380 ws->scavd_list = NULL;
1381 ws->n_scavd_blocks = 0;
1383 // If the block at the head of the list in this generation
1384 // is less than 3/4 full, then use it as a todo block.
1385 if (gen->blocks && isPartiallyFull(gen->blocks))
1387 ws->todo_bd = gen->blocks;
1388 ws->todo_free = ws->todo_bd->free;
1389 ws->todo_lim = ws->todo_bd->start + BLOCK_SIZE_W;
1390 gen->blocks = gen->blocks->link;
1392 gen->n_words -= ws->todo_bd->free - ws->todo_bd->start;
1393 ws->todo_bd->link = NULL;
1394 // we must scan from the current end point.
1395 ws->todo_bd->u.scan = ws->todo_bd->free;
1400 alloc_todo_block(ws,0);
1404 // deal out any more partial blocks to the threads' part_lists
1406 while (gen->blocks && isPartiallyFull(gen->blocks))
1409 gen->blocks = bd->link;
1410 ws = &gc_threads[t]->gens[g];
1411 bd->link = ws->part_list;
1413 ws->n_part_blocks += 1;
1414 bd->u.scan = bd->free;
1416 gen->n_words -= bd->free - bd->start;
1418 if (t == n_gc_threads) t = 0;
1422 /* -----------------------------------------------------------------------------
1423 Initialise a gc_thread before GC
1424 -------------------------------------------------------------------------- */
1427 init_gc_thread (gc_thread *t)
1429 t->static_objects = END_OF_STATIC_LIST;
1430 t->scavenged_static_objects = END_OF_STATIC_LIST;
1432 t->mut_lists = capabilities[t->thread_index].mut_lists;
1434 t->failed_to_evac = rtsFalse;
1435 t->eager_promotion = rtsTrue;
1436 t->thunk_selector_depth = 0;
1441 t->scav_find_work = 0;
1444 /* -----------------------------------------------------------------------------
1445 Function we pass to evacuate roots.
1446 -------------------------------------------------------------------------- */
1449 mark_root(void *user USED_IF_THREADS, StgClosure **root)
1451 // we stole a register for gct, but this function is called from
1452 // *outside* the GC where the register variable is not in effect,
1453 // so we need to save and restore it here. NB. only call
1454 // mark_root() from the main GC thread, otherwise gct will be
1456 gc_thread *saved_gct;
1465 /* -----------------------------------------------------------------------------
1466 Initialising the static object & mutable lists
1467 -------------------------------------------------------------------------- */
1470 zero_static_object_list(StgClosure* first_static)
1474 const StgInfoTable *info;
1476 for (p = first_static; p != END_OF_STATIC_LIST; p = link) {
1478 link = *STATIC_LINK(info, p);
1479 *STATIC_LINK(info,p) = NULL;
1483 /* ----------------------------------------------------------------------------
1484 Reset the sizes of the older generations when we do a major
1487 CURRENT STRATEGY: make all generations except zero the same size.
1488 We have to stay within the maximum heap size, and leave a certain
1489 percentage of the maximum heap size available to allocate into.
1490 ------------------------------------------------------------------------- */
1493 resize_generations (void)
1497 if (major_gc && RtsFlags.GcFlags.generations > 1) {
1498 nat live, size, min_alloc, words;
1499 const nat max = RtsFlags.GcFlags.maxHeapSize;
1500 const nat gens = RtsFlags.GcFlags.generations;
1502 // live in the oldest generations
1503 if (oldest_gen->live_estimate != 0) {
1504 words = oldest_gen->live_estimate;
1506 words = oldest_gen->n_words;
1508 live = (words + BLOCK_SIZE_W - 1) / BLOCK_SIZE_W +
1509 oldest_gen->n_large_blocks;
1511 // default max size for all generations except zero
1512 size = stg_max(live * RtsFlags.GcFlags.oldGenFactor,
1513 RtsFlags.GcFlags.minOldGenSize);
1515 if (RtsFlags.GcFlags.heapSizeSuggestionAuto) {
1516 RtsFlags.GcFlags.heapSizeSuggestion = size;
1519 // minimum size for generation zero
1520 min_alloc = stg_max((RtsFlags.GcFlags.pcFreeHeap * max) / 200,
1521 RtsFlags.GcFlags.minAllocAreaSize);
1523 // Auto-enable compaction when the residency reaches a
1524 // certain percentage of the maximum heap size (default: 30%).
1525 if (RtsFlags.GcFlags.compact ||
1527 oldest_gen->n_blocks >
1528 (RtsFlags.GcFlags.compactThreshold * max) / 100)) {
1529 oldest_gen->mark = 1;
1530 oldest_gen->compact = 1;
1531 // debugBelch("compaction: on\n", live);
1533 oldest_gen->mark = 0;
1534 oldest_gen->compact = 0;
1535 // debugBelch("compaction: off\n", live);
1538 if (RtsFlags.GcFlags.sweep) {
1539 oldest_gen->mark = 1;
1542 // if we're going to go over the maximum heap size, reduce the
1543 // size of the generations accordingly. The calculation is
1544 // different if compaction is turned on, because we don't need
1545 // to double the space required to collect the old generation.
1548 // this test is necessary to ensure that the calculations
1549 // below don't have any negative results - we're working
1550 // with unsigned values here.
1551 if (max < min_alloc) {
1555 if (oldest_gen->compact) {
1556 if ( (size + (size - 1) * (gens - 2) * 2) + min_alloc > max ) {
1557 size = (max - min_alloc) / ((gens - 1) * 2 - 1);
1560 if ( (size * (gens - 1) * 2) + min_alloc > max ) {
1561 size = (max - min_alloc) / ((gens - 1) * 2);
1571 debugBelch("live: %d, min_alloc: %d, size : %d, max = %d\n", live,
1572 min_alloc, size, max);
1575 for (g = 0; g < gens; g++) {
1576 generations[g].max_blocks = size;
1581 /* -----------------------------------------------------------------------------
1582 Calculate the new size of the nursery, and resize it.
1583 -------------------------------------------------------------------------- */
1586 resize_nursery (void)
1588 const lnat min_nursery = RtsFlags.GcFlags.minAllocAreaSize * n_capabilities;
1590 if (RtsFlags.GcFlags.generations == 1)
1591 { // Two-space collector:
1594 /* set up a new nursery. Allocate a nursery size based on a
1595 * function of the amount of live data (by default a factor of 2)
1596 * Use the blocks from the old nursery if possible, freeing up any
1599 * If we get near the maximum heap size, then adjust our nursery
1600 * size accordingly. If the nursery is the same size as the live
1601 * data (L), then we need 3L bytes. We can reduce the size of the
1602 * nursery to bring the required memory down near 2L bytes.
1604 * A normal 2-space collector would need 4L bytes to give the same
1605 * performance we get from 3L bytes, reducing to the same
1606 * performance at 2L bytes.
1608 blocks = generations[0].n_blocks;
1610 if ( RtsFlags.GcFlags.maxHeapSize != 0 &&
1611 blocks * RtsFlags.GcFlags.oldGenFactor * 2 >
1612 RtsFlags.GcFlags.maxHeapSize )
1614 long adjusted_blocks; // signed on purpose
1617 adjusted_blocks = (RtsFlags.GcFlags.maxHeapSize - 2 * blocks);
1619 debugTrace(DEBUG_gc, "near maximum heap size of 0x%x blocks, blocks = %d, adjusted to %ld",
1620 RtsFlags.GcFlags.maxHeapSize, blocks, adjusted_blocks);
1622 pc_free = adjusted_blocks * 100 / RtsFlags.GcFlags.maxHeapSize;
1623 if (pc_free < RtsFlags.GcFlags.pcFreeHeap) /* might even * be < 0 */
1627 blocks = adjusted_blocks;
1631 blocks *= RtsFlags.GcFlags.oldGenFactor;
1632 if (blocks < min_nursery)
1634 blocks = min_nursery;
1637 resizeNurseries(blocks);
1639 else // Generational collector
1642 * If the user has given us a suggested heap size, adjust our
1643 * allocation area to make best use of the memory available.
1645 if (RtsFlags.GcFlags.heapSizeSuggestion)
1648 const nat needed = calcNeeded(); // approx blocks needed at next GC
1650 /* Guess how much will be live in generation 0 step 0 next time.
1651 * A good approximation is obtained by finding the
1652 * percentage of g0 that was live at the last minor GC.
1654 * We have an accurate figure for the amount of copied data in
1655 * 'copied', but we must convert this to a number of blocks, with
1656 * a small adjustment for estimated slop at the end of a block
1661 g0_pcnt_kept = ((copied / (BLOCK_SIZE_W - 10)) * 100)
1662 / countNurseryBlocks();
1665 /* Estimate a size for the allocation area based on the
1666 * information available. We might end up going slightly under
1667 * or over the suggested heap size, but we should be pretty
1670 * Formula: suggested - needed
1671 * ----------------------------
1672 * 1 + g0_pcnt_kept/100
1674 * where 'needed' is the amount of memory needed at the next
1675 * collection for collecting all gens except g0.
1678 (((long)RtsFlags.GcFlags.heapSizeSuggestion - (long)needed) * 100) /
1679 (100 + (long)g0_pcnt_kept);
1681 if (blocks < (long)min_nursery) {
1682 blocks = min_nursery;
1685 resizeNurseries((nat)blocks);
1689 // we might have added extra large blocks to the nursery, so
1690 // resize back to minAllocAreaSize again.
1691 resizeNurseriesFixed(RtsFlags.GcFlags.minAllocAreaSize);
1696 /* -----------------------------------------------------------------------------
1697 Sanity code for CAF garbage collection.
1699 With DEBUG turned on, we manage a CAF list in addition to the SRT
1700 mechanism. After GC, we run down the CAF list and blackhole any
1701 CAFs which have been garbage collected. This means we get an error
1702 whenever the program tries to enter a garbage collected CAF.
1704 Any garbage collected CAFs are taken off the CAF list at the same
1706 -------------------------------------------------------------------------- */
1708 #if 0 && defined(DEBUG)
1715 const StgInfoTable *info;
1726 ASSERT(info->type == IND_STATIC);
1728 if (STATIC_LINK(info,p) == NULL) {
1729 debugTrace(DEBUG_gccafs, "CAF gc'd at 0x%04lx", (long)p);
1731 SET_INFO(p,&stg_BLACKHOLE_info);
1732 p = STATIC_LINK2(info,p);
1736 pp = &STATIC_LINK2(info,p);
1743 debugTrace(DEBUG_gccafs, "%d CAFs live", i);