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"
51 #include <string.h> // for memset()
54 /* -----------------------------------------------------------------------------
56 -------------------------------------------------------------------------- */
58 /* STATIC OBJECT LIST.
61 * We maintain a linked list of static objects that are still live.
62 * The requirements for this list are:
64 * - we need to scan the list while adding to it, in order to
65 * scavenge all the static objects (in the same way that
66 * breadth-first scavenging works for dynamic objects).
68 * - we need to be able to tell whether an object is already on
69 * the list, to break loops.
71 * Each static object has a "static link field", which we use for
72 * linking objects on to the list. We use a stack-type list, consing
73 * objects on the front as they are added (this means that the
74 * scavenge phase is depth-first, not breadth-first, but that
77 * A separate list is kept for objects that have been scavenged
78 * already - this is so that we can zero all the marks afterwards.
80 * An object is on the list if its static link field is non-zero; this
81 * means that we have to mark the end of the list with '1', not NULL.
83 * Extra notes for generational GC:
85 * Each generation has a static object list associated with it. When
86 * collecting generations up to N, we treat the static object lists
87 * from generations > N as roots.
89 * We build up a static object list while collecting generations 0..N,
90 * which is then appended to the static object list of generation N+1.
93 /* N is the oldest generation being collected, where the generations
94 * are numbered starting at 0. A major GC (indicated by the major_gc
95 * flag) is when we're collecting all generations. We only attempt to
96 * deal with static objects and GC CAFs when doing a major GC.
101 /* Data used for allocation area sizing.
103 static lnat g0s0_pcnt_kept = 30; // percentage of g0s0 live at last minor GC
113 /* Thread-local data for each GC thread
115 gc_thread **gc_threads = NULL;
117 #if !defined(THREADED_RTS)
118 StgWord8 the_gc_thread[sizeof(gc_thread) + 64 * sizeof(step_workspace)];
121 // Number of threads running in *this* GC. Affects how many
122 // step->todos[] lists we have to look in to find work.
126 long copied; // *words* copied & scavenged during this GC
128 rtsBool work_stealing;
132 /* -----------------------------------------------------------------------------
133 Static function declarations
134 -------------------------------------------------------------------------- */
136 static void mark_root (void *user, StgClosure **root);
137 static void zero_static_object_list (StgClosure* first_static);
138 static nat initialise_N (rtsBool force_major_gc);
139 static void init_collected_gen (nat g, nat threads);
140 static void init_uncollected_gen (nat g, nat threads);
141 static void init_gc_thread (gc_thread *t);
142 static void update_task_list (void);
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 /* -----------------------------------------------------------------------------
157 The mark bitmap & stack.
158 -------------------------------------------------------------------------- */
160 #define MARK_STACK_BLOCKS 4
162 bdescr *mark_stack_bdescr;
167 // Flag and pointers used for falling back to a linear scan when the
168 // mark stack overflows.
169 rtsBool mark_stack_overflowed;
170 bdescr *oldgen_scan_bd;
173 /* -----------------------------------------------------------------------------
174 GarbageCollect: the main entry point to the garbage collector.
176 Locks held: all capabilities are held throughout GarbageCollect().
177 -------------------------------------------------------------------------- */
180 GarbageCollect (rtsBool force_major_gc,
181 nat gc_type USED_IF_THREADS,
186 lnat live, allocated, max_copied, avg_copied, slop;
187 gc_thread *saved_gct;
190 // necessary if we stole a callee-saves register for gct:
194 CostCentreStack *prev_CCS;
199 #if defined(RTS_USER_SIGNALS)
200 if (RtsFlags.MiscFlags.install_signal_handlers) {
206 ASSERT(sizeof(step_workspace) == 16 * sizeof(StgWord));
207 // otherwise adjust the padding in step_workspace.
209 // tell the stats department that we've started a GC
212 // tell the STM to discard any cached closures it's hoping to re-use
215 // lock the StablePtr table
224 // attribute any costs to CCS_GC
230 /* Approximate how much we allocated.
231 * Todo: only when generating stats?
233 allocated = calcAllocated();
235 /* Figure out which generation to collect
237 n = initialise_N(force_major_gc);
239 #if defined(THREADED_RTS)
240 work_stealing = RtsFlags.ParFlags.parGcLoadBalancing;
241 // It's not always a good idea to do load balancing in parallel
242 // GC. In particular, for a parallel program we don't want to
243 // lose locality by moving cached data into another CPU's cache
244 // (this effect can be quite significant).
246 // We could have a more complex way to deterimine whether to do
247 // work stealing or not, e.g. it might be a good idea to do it
248 // if the heap is big. For now, we just turn it on or off with
252 /* Start threads, so they can be spinning up while we finish initialisation.
256 #if defined(THREADED_RTS)
257 /* How many threads will be participating in this GC?
258 * We don't try to parallelise minor GCs (unless the user asks for
259 * it with +RTS -gn0), or mark/compact/sweep GC.
261 if (gc_type == PENDING_GC_PAR) {
262 n_gc_threads = RtsFlags.ParFlags.nNodes;
270 debugTrace(DEBUG_gc, "GC (gen %d): %d KB to collect, %ld MB in use, using %d thread(s)",
271 N, n * (BLOCK_SIZE / 1024), mblocks_allocated, n_gc_threads);
273 #ifdef RTS_GTK_FRONTPANEL
274 if (RtsFlags.GcFlags.frontpanel) {
275 updateFrontPanelBeforeGC(N);
280 // check for memory leaks if DEBUG is on
281 memInventory(traceClass(DEBUG_gc));
284 // check stack sanity *before* GC
285 IF_DEBUG(sanity, checkFreeListSanity());
286 IF_DEBUG(sanity, checkMutableLists(rtsTrue));
288 // Initialise all our gc_thread structures
289 for (t = 0; t < n_gc_threads; t++) {
290 init_gc_thread(gc_threads[t]);
293 // Initialise all the generations/steps that we're collecting.
294 for (g = 0; g <= N; g++) {
295 init_collected_gen(g,n_gc_threads);
298 // Initialise all the generations/steps that we're *not* collecting.
299 for (g = N+1; g < RtsFlags.GcFlags.generations; g++) {
300 init_uncollected_gen(g,n_gc_threads);
303 /* Allocate a mark stack if we're doing a major collection.
305 if (major_gc && oldest_gen->steps[0].mark) {
306 nat mark_stack_blocks;
307 mark_stack_blocks = stg_max(MARK_STACK_BLOCKS,
308 oldest_gen->steps[0].n_old_blocks / 100);
309 mark_stack_bdescr = allocGroup(mark_stack_blocks);
310 mark_stack = (StgPtr *)mark_stack_bdescr->start;
311 mark_sp = mark_stack;
312 mark_splim = mark_stack + (mark_stack_blocks * BLOCK_SIZE_W);
314 mark_stack_bdescr = NULL;
317 // this is the main thread
319 if (n_gc_threads == 1) {
320 SET_GCT(gc_threads[0]);
322 SET_GCT(gc_threads[cap->no]);
325 SET_GCT(gc_threads[0]);
328 /* -----------------------------------------------------------------------
329 * follow all the roots that we know about:
332 // the main thread is running: this prevents any other threads from
333 // exiting prematurely, so we can start them now.
334 // NB. do this after the mutable lists have been saved above, otherwise
335 // the other GC threads will be writing into the old mutable lists.
337 wakeup_gc_threads(n_gc_threads, gct->thread_index);
339 // Mutable lists from each generation > N
340 // we want to *scavenge* these roots, not evacuate them: they're not
341 // going to move in this GC.
342 // Also do them in reverse generation order, for the usual reason:
343 // namely to reduce the likelihood of spurious old->new pointers.
345 for (g = RtsFlags.GcFlags.generations-1; g > N; g--) {
346 scavenge_mutable_list(generations[g].saved_mut_list, &generations[g]);
347 freeChain_sync(generations[g].saved_mut_list);
348 generations[g].saved_mut_list = NULL;
352 // scavenge the capability-private mutable lists. This isn't part
353 // of markSomeCapabilities() because markSomeCapabilities() can only
354 // call back into the GC via mark_root() (due to the gct register
356 if (n_gc_threads == 1) {
357 for (n = 0; n < n_capabilities; n++) {
358 scavenge_capability_mut_lists(&capabilities[n]);
361 scavenge_capability_mut_lists(&capabilities[gct->thread_index]);
364 // follow roots from the CAF list (used by GHCi)
366 markCAFs(mark_root, gct);
368 // follow all the roots that the application knows about.
370 markSomeCapabilities(mark_root, gct, gct->thread_index, n_gc_threads,
371 rtsTrue/*prune sparks*/);
373 #if defined(RTS_USER_SIGNALS)
374 // mark the signal handlers (signals should be already blocked)
375 markSignalHandlers(mark_root, gct);
378 // Mark the weak pointer list, and prepare to detect dead weak pointers.
382 // Mark the stable pointer table.
383 markStablePtrTable(mark_root, gct);
385 /* -------------------------------------------------------------------------
386 * Repeatedly scavenge all the areas we know about until there's no
387 * more scavenging to be done.
391 scavenge_until_all_done();
392 // The other threads are now stopped. We might recurse back to
393 // here, but from now on this is the only thread.
395 // if any blackholes are alive, make the threads that wait on
397 if (traverseBlackholeQueue()) {
402 // must be last... invariant is that everything is fully
403 // scavenged at this point.
404 if (traverseWeakPtrList()) { // returns rtsTrue if evaced something
409 // If we get to here, there's really nothing left to do.
413 shutdown_gc_threads(n_gc_threads, gct->thread_index);
415 // Update pointers from the Task list
418 // Now see which stable names are still alive.
422 // We call processHeapClosureForDead() on every closure destroyed during
423 // the current garbage collection, so we invoke LdvCensusForDead().
424 if (RtsFlags.ProfFlags.doHeapProfile == HEAP_BY_LDV
425 || RtsFlags.ProfFlags.bioSelector != NULL)
429 // NO MORE EVACUATION AFTER THIS POINT!
431 // Two-space collector: free the old to-space.
432 // g0s0->old_blocks is the old nursery
433 // g0s0->blocks is to-space from the previous GC
434 if (RtsFlags.GcFlags.generations == 1) {
435 if (g0s0->blocks != NULL) {
436 freeChain(g0s0->blocks);
441 // For each workspace, in each thread, move the copied blocks to the step
447 for (t = 0; t < n_gc_threads; t++) {
451 if (RtsFlags.GcFlags.generations == 1) {
456 for (; s < total_steps; s++) {
459 // Push the final block
461 push_scanned_block(ws->todo_bd, ws);
464 ASSERT(gct->scan_bd == NULL);
465 ASSERT(countBlocks(ws->scavd_list) == ws->n_scavd_blocks);
468 for (bd = ws->scavd_list; bd != NULL; bd = bd->link) {
469 ws->step->n_words += bd->free - bd->start;
473 prev->link = ws->step->blocks;
474 ws->step->blocks = ws->scavd_list;
476 ws->step->n_blocks += ws->n_scavd_blocks;
480 // Add all the partial blocks *after* we've added all the full
481 // blocks. This is so that we can grab the partial blocks back
482 // again and try to fill them up in the next GC.
483 for (t = 0; t < n_gc_threads; t++) {
487 if (RtsFlags.GcFlags.generations == 1) {
492 for (; s < total_steps; s++) {
496 for (bd = ws->part_list; bd != NULL; bd = next) {
498 if (bd->free == bd->start) {
500 ws->part_list = next;
507 ws->step->n_words += bd->free - bd->start;
512 prev->link = ws->step->blocks;
513 ws->step->blocks = ws->part_list;
515 ws->step->n_blocks += ws->n_part_blocks;
517 ASSERT(countBlocks(ws->step->blocks) == ws->step->n_blocks);
518 ASSERT(countOccupied(ws->step->blocks) == ws->step->n_words);
523 // Finally: compact or sweep the oldest generation.
524 if (major_gc && oldest_gen->steps[0].mark) {
525 if (oldest_gen->steps[0].compact)
526 compact(gct->scavenged_static_objects);
528 sweep(&oldest_gen->steps[0]);
531 /* run through all the generations/steps and tidy up
538 for (i=0; i < n_gc_threads; i++) {
539 if (n_gc_threads > 1) {
540 debugTrace(DEBUG_gc,"thread %d:", i);
541 debugTrace(DEBUG_gc," copied %ld", gc_threads[i]->copied * sizeof(W_));
542 debugTrace(DEBUG_gc," scanned %ld", gc_threads[i]->scanned * sizeof(W_));
543 debugTrace(DEBUG_gc," any_work %ld", gc_threads[i]->any_work);
544 debugTrace(DEBUG_gc," no_work %ld", gc_threads[i]->no_work);
545 debugTrace(DEBUG_gc," scav_find_work %ld", gc_threads[i]->scav_find_work);
547 copied += gc_threads[i]->copied;
548 max_copied = stg_max(gc_threads[i]->copied, max_copied);
550 if (n_gc_threads == 1) {
558 for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
561 generations[g].collections++; // for stats
562 if (n_gc_threads > 1) generations[g].par_collections++;
565 // Count the mutable list as bytes "copied" for the purposes of
566 // stats. Every mutable list is copied during every GC.
568 nat mut_list_size = 0;
569 for (bd = generations[g].mut_list; bd != NULL; bd = bd->link) {
570 mut_list_size += bd->free - bd->start;
572 for (n = 0; n < n_capabilities; n++) {
573 for (bd = capabilities[n].mut_lists[g];
574 bd != NULL; bd = bd->link) {
575 mut_list_size += bd->free - bd->start;
578 copied += mut_list_size;
581 "mut_list_size: %lu (%d vars, %d arrays, %d MVARs, %d others)",
582 (unsigned long)(mut_list_size * sizeof(W_)),
583 mutlist_MUTVARS, mutlist_MUTARRS, mutlist_MVARS, mutlist_OTHERS);
586 for (s = 0; s < generations[g].n_steps; s++) {
588 stp = &generations[g].steps[s];
590 // for generations we collected...
593 /* free old memory and shift to-space into from-space for all
594 * the collected steps (except the allocation area). These
595 * freed blocks will probaby be quickly recycled.
597 if (!(g == 0 && s == 0 && RtsFlags.GcFlags.generations > 1)) {
600 // tack the new blocks on the end of the existing blocks
601 if (stp->old_blocks != NULL) {
604 for (bd = stp->old_blocks; bd != NULL; bd = next) {
608 if (!(bd->flags & BF_MARKED))
611 stp->old_blocks = next;
620 stp->n_words += bd->free - bd->start;
622 // NB. this step might not be compacted next
623 // time, so reset the BF_MARKED flags.
624 // They are set before GC if we're going to
625 // compact. (search for BF_MARKED above).
626 bd->flags &= ~BF_MARKED;
628 // between GCs, all blocks in the heap except
629 // for the nursery have the BF_EVACUATED flag set.
630 bd->flags |= BF_EVACUATED;
637 prev->link = stp->blocks;
638 stp->blocks = stp->old_blocks;
641 // add the new blocks to the block tally
642 stp->n_blocks += stp->n_old_blocks;
643 ASSERT(countBlocks(stp->blocks) == stp->n_blocks);
644 ASSERT(countOccupied(stp->blocks) == stp->n_words);
648 freeChain(stp->old_blocks);
650 stp->old_blocks = NULL;
651 stp->n_old_blocks = 0;
654 /* LARGE OBJECTS. The current live large objects are chained on
655 * scavenged_large, having been moved during garbage
656 * collection from large_objects. Any objects left on
657 * large_objects list are therefore dead, so we free them here.
659 for (bd = stp->large_objects; bd != NULL; bd = next) {
665 stp->large_objects = stp->scavenged_large_objects;
666 stp->n_large_blocks = stp->n_scavenged_large_blocks;
669 else // for older generations...
671 /* For older generations, we need to append the
672 * scavenged_large_object list (i.e. large objects that have been
673 * promoted during this GC) to the large_object list for that step.
675 for (bd = stp->scavenged_large_objects; bd; bd = next) {
677 dbl_link_onto(bd, &stp->large_objects);
680 // add the new blocks we promoted during this GC
681 stp->n_large_blocks += stp->n_scavenged_large_blocks;
686 // update the max size of older generations after a major GC
687 resize_generations();
689 // Calculate the amount of live data for stats.
690 live = calcLiveWords();
692 // Free the small objects allocated via allocate(), since this will
693 // all have been copied into G0S1 now.
694 if (RtsFlags.GcFlags.generations > 1) {
695 if (g0s0->blocks != NULL) {
696 freeChain(g0s0->blocks);
703 alloc_blocks_lim = RtsFlags.GcFlags.minAllocAreaSize;
705 // Start a new pinned_object_block
706 pinned_object_block = NULL;
708 // Free the mark stack.
709 if (mark_stack_bdescr != NULL) {
710 freeGroup(mark_stack_bdescr);
714 for (g = 0; g <= N; g++) {
715 for (s = 0; s < generations[g].n_steps; s++) {
716 stp = &generations[g].steps[s];
717 if (stp->bitmap != NULL) {
718 freeGroup(stp->bitmap);
726 // mark the garbage collected CAFs as dead
727 #if 0 && defined(DEBUG) // doesn't work at the moment
728 if (major_gc) { gcCAFs(); }
732 // resetStaticObjectForRetainerProfiling() must be called before
734 if (n_gc_threads > 1) {
735 barf("profiling is currently broken with multi-threaded GC");
736 // ToDo: fix the gct->scavenged_static_objects below
738 resetStaticObjectForRetainerProfiling(gct->scavenged_static_objects);
741 // zero the scavenged static object list
744 for (i = 0; i < n_gc_threads; i++) {
745 zero_static_object_list(gc_threads[i]->scavenged_static_objects);
752 // start any pending finalizers
754 scheduleFinalizers(cap, old_weak_ptr_list);
757 // send exceptions to any threads which were about to die
759 resurrectThreads(resurrected_threads);
760 performPendingThrowTos(exception_threads);
763 // Update the stable pointer hash table.
764 updateStablePtrTable(major_gc);
766 // check sanity after GC
767 IF_DEBUG(sanity, checkSanity());
769 // extra GC trace info
770 IF_DEBUG(gc, statDescribeGens());
773 // symbol-table based profiling
774 /* heapCensus(to_blocks); */ /* ToDo */
777 // restore enclosing cost centre
783 // check for memory leaks if DEBUG is on
784 memInventory(traceClass(DEBUG_gc));
787 #ifdef RTS_GTK_FRONTPANEL
788 if (RtsFlags.GcFlags.frontpanel) {
789 updateFrontPanelAfterGC( N, live );
793 // ok, GC over: tell the stats department what happened.
794 slop = calcLiveBlocks() * BLOCK_SIZE_W - live;
795 stat_endGC(allocated, live, copied, N, max_copied, avg_copied, slop);
797 // unlock the StablePtr table
800 // Guess which generation we'll collect *next* time
801 initialise_N(force_major_gc);
803 #if defined(RTS_USER_SIGNALS)
804 if (RtsFlags.MiscFlags.install_signal_handlers) {
805 // unblock signals again
806 unblockUserSignals();
815 /* -----------------------------------------------------------------------------
816 Figure out which generation to collect, initialise N and major_gc.
818 Also returns the total number of blocks in generations that will be
820 -------------------------------------------------------------------------- */
823 initialise_N (rtsBool force_major_gc)
826 nat s, blocks, blocks_total;
831 if (force_major_gc) {
832 N = RtsFlags.GcFlags.generations - 1;
837 for (g = RtsFlags.GcFlags.generations - 1; g >= 0; g--) {
839 for (s = 0; s < generations[g].n_steps; s++) {
840 blocks += generations[g].steps[s].n_words / BLOCK_SIZE_W;
841 blocks += generations[g].steps[s].n_large_blocks;
843 if (blocks >= generations[g].max_blocks) {
847 blocks_total += blocks;
851 blocks_total += countNurseryBlocks();
853 major_gc = (N == RtsFlags.GcFlags.generations-1);
857 /* -----------------------------------------------------------------------------
858 Initialise the gc_thread structures.
859 -------------------------------------------------------------------------- */
861 #define GC_THREAD_INACTIVE 0
862 #define GC_THREAD_STANDING_BY 1
863 #define GC_THREAD_RUNNING 2
864 #define GC_THREAD_WAITING_TO_CONTINUE 3
867 new_gc_thread (nat n, gc_thread *t)
874 initSpinLock(&t->gc_spin);
875 initSpinLock(&t->mut_spin);
876 ACQUIRE_SPIN_LOCK(&t->gc_spin);
877 t->wakeup = GC_THREAD_INACTIVE; // starts true, so we can wait for the
878 // thread to start up, see wakeup_gc_threads
882 t->free_blocks = NULL;
891 for (s = 0; s < total_steps; s++)
894 ws->step = &all_steps[s];
895 ASSERT(s == ws->step->abs_no);
899 ws->todo_q = newWSDeque(128);
900 ws->todo_overflow = NULL;
901 ws->n_todo_overflow = 0;
903 ws->part_list = NULL;
904 ws->n_part_blocks = 0;
906 ws->scavd_list = NULL;
907 ws->n_scavd_blocks = 0;
915 if (gc_threads == NULL) {
916 #if defined(THREADED_RTS)
918 gc_threads = stgMallocBytes (RtsFlags.ParFlags.nNodes *
922 for (i = 0; i < RtsFlags.ParFlags.nNodes; i++) {
924 stgMallocBytes(sizeof(gc_thread) + total_steps * sizeof(step_workspace),
927 new_gc_thread(i, gc_threads[i]);
930 gc_threads = stgMallocBytes (sizeof(gc_thread*),"alloc_gc_threads");
932 new_gc_thread(0,gc_threads[0]);
940 if (gc_threads != NULL) {
941 #if defined(THREADED_RTS)
943 for (i = 0; i < RtsFlags.ParFlags.nNodes; i++) {
944 stgFree (gc_threads[i]);
946 stgFree (gc_threads);
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_overflowed || oldgen_scan_bd != NULL ||
988 (mark_stack_bdescr != NULL && !mark_stack_empty())) {
992 // Check for global work in any step. We don't need to check for
993 // local work, because we have already exited scavenge_loop(),
994 // which means there is no local work for this thread.
995 for (s = total_steps-1; s >= 0; s--) {
996 if (s == 0 && RtsFlags.GcFlags.generations > 1) {
1000 if (ws->todo_large_objects) return rtsTrue;
1001 if (!looksEmptyWSDeque(ws->todo_q)) return rtsTrue;
1002 if (ws->todo_overflow) return rtsTrue;
1005 #if defined(THREADED_RTS)
1006 if (work_stealing) {
1008 // look for work to steal
1009 for (n = 0; n < n_gc_threads; n++) {
1010 if (n == gct->thread_index) continue;
1011 for (s = total_steps-1; s >= 0; s--) {
1012 ws = &gc_threads[n]->steps[s];
1013 if (!looksEmptyWSDeque(ws->todo_q)) return rtsTrue;
1025 scavenge_until_all_done (void)
1029 debugTrace(DEBUG_gc, "GC thread %d working", 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 debugTrace(DEBUG_gc, "GC thread %d idle (%d still running)",
1046 gct->thread_index, r);
1048 while (gc_running_threads != 0) {
1054 // any_work() does not remove the work from the queue, it
1055 // just checks for the presence of work. If we find any,
1056 // then we increment gc_running_threads and go back to
1057 // scavenge_loop() to perform any pending work.
1060 // All threads are now stopped
1061 debugTrace(DEBUG_gc, "GC thread %d finished.", 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 cap->in_gc = rtsTrue;
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 // count events in this thread towards the GC totals
1103 papi_thread_stop_gc1_count(gct->papi_events);
1106 // Wait until we're told to continue
1107 RELEASE_SPIN_LOCK(&gct->gc_spin);
1108 gct->wakeup = GC_THREAD_WAITING_TO_CONTINUE;
1109 debugTrace(DEBUG_gc, "GC thread %d waiting to continue...",
1111 ACQUIRE_SPIN_LOCK(&gct->mut_spin);
1112 debugTrace(DEBUG_gc, "GC thread %d on my way...", gct->thread_index);
1119 #if defined(THREADED_RTS)
1122 waitForGcThreads (Capability *cap USED_IF_THREADS)
1124 nat n_threads = RtsFlags.ParFlags.nNodes;
1127 rtsBool retry = rtsTrue;
1130 for (i=0; i < n_threads; i++) {
1131 if (i == me) continue;
1132 if (gc_threads[i]->wakeup != GC_THREAD_STANDING_BY) {
1133 prodCapability(&capabilities[i], cap->running_task);
1136 for (j=0; j < 10000000; j++) {
1138 for (i=0; i < n_threads; i++) {
1139 if (i == me) continue;
1141 setContextSwitches();
1142 if (gc_threads[i]->wakeup != GC_THREAD_STANDING_BY) {
1151 #endif // THREADED_RTS
1154 start_gc_threads (void)
1156 #if defined(THREADED_RTS)
1157 gc_running_threads = 0;
1162 wakeup_gc_threads (nat n_threads USED_IF_THREADS, nat me USED_IF_THREADS)
1164 #if defined(THREADED_RTS)
1166 for (i=0; i < n_threads; i++) {
1167 if (i == me) continue;
1169 debugTrace(DEBUG_gc, "waking up gc thread %d", i);
1170 if (gc_threads[i]->wakeup != GC_THREAD_STANDING_BY) barf("wakeup_gc_threads");
1172 gc_threads[i]->wakeup = GC_THREAD_RUNNING;
1173 ACQUIRE_SPIN_LOCK(&gc_threads[i]->mut_spin);
1174 RELEASE_SPIN_LOCK(&gc_threads[i]->gc_spin);
1179 // After GC is complete, we must wait for all GC threads to enter the
1180 // standby state, otherwise they may still be executing inside
1181 // any_work(), and may even remain awake until the next GC starts.
1183 shutdown_gc_threads (nat n_threads USED_IF_THREADS, nat me USED_IF_THREADS)
1185 #if defined(THREADED_RTS)
1187 for (i=0; i < n_threads; i++) {
1188 if (i == me) continue;
1189 while (gc_threads[i]->wakeup != GC_THREAD_WAITING_TO_CONTINUE) { write_barrier(); }
1194 #if defined(THREADED_RTS)
1196 releaseGCThreads (Capability *cap USED_IF_THREADS)
1198 nat n_threads = RtsFlags.ParFlags.nNodes;
1201 for (i=0; i < n_threads; i++) {
1202 if (i == me) continue;
1203 if (gc_threads[i]->wakeup != GC_THREAD_WAITING_TO_CONTINUE)
1204 barf("releaseGCThreads");
1206 gc_threads[i]->wakeup = GC_THREAD_INACTIVE;
1207 ACQUIRE_SPIN_LOCK(&gc_threads[i]->gc_spin);
1208 RELEASE_SPIN_LOCK(&gc_threads[i]->mut_spin);
1213 /* ----------------------------------------------------------------------------
1214 Initialise a generation that is to be collected
1215 ------------------------------------------------------------------------- */
1218 init_collected_gen (nat g, nat n_threads)
1225 // Throw away the current mutable list. Invariant: the mutable
1226 // list always has at least one block; this means we can avoid a
1227 // check for NULL in recordMutable().
1229 freeChain(generations[g].mut_list);
1230 generations[g].mut_list = allocBlock();
1231 for (i = 0; i < n_capabilities; i++) {
1232 freeChain(capabilities[i].mut_lists[g]);
1233 capabilities[i].mut_lists[g] = allocBlock();
1237 for (s = 0; s < generations[g].n_steps; s++) {
1239 stp = &generations[g].steps[s];
1240 ASSERT(stp->gen_no == g);
1242 // we'll construct a new list of threads in this step
1243 // during GC, throw away the current list.
1244 stp->old_threads = stp->threads;
1245 stp->threads = END_TSO_QUEUE;
1247 // generation 0, step 0 doesn't need to-space
1248 if (g == 0 && s == 0 && RtsFlags.GcFlags.generations > 1) {
1252 // deprecate the existing blocks
1253 stp->old_blocks = stp->blocks;
1254 stp->n_old_blocks = stp->n_blocks;
1258 stp->live_estimate = 0;
1260 // initialise the large object queues.
1261 stp->scavenged_large_objects = NULL;
1262 stp->n_scavenged_large_blocks = 0;
1264 // mark the small objects as from-space
1265 for (bd = stp->old_blocks; bd; bd = bd->link) {
1266 bd->flags &= ~BF_EVACUATED;
1269 // mark the large objects as from-space
1270 for (bd = stp->large_objects; bd; bd = bd->link) {
1271 bd->flags &= ~BF_EVACUATED;
1274 // for a compacted step, we need to allocate the bitmap
1276 nat bitmap_size; // in bytes
1277 bdescr *bitmap_bdescr;
1280 bitmap_size = stp->n_old_blocks * BLOCK_SIZE / (sizeof(W_)*BITS_PER_BYTE);
1282 if (bitmap_size > 0) {
1283 bitmap_bdescr = allocGroup((lnat)BLOCK_ROUND_UP(bitmap_size)
1285 stp->bitmap = bitmap_bdescr;
1286 bitmap = bitmap_bdescr->start;
1288 debugTrace(DEBUG_gc, "bitmap_size: %d, bitmap: %p",
1289 bitmap_size, bitmap);
1291 // don't forget to fill it with zeros!
1292 memset(bitmap, 0, bitmap_size);
1294 // For each block in this step, point to its bitmap from the
1295 // block descriptor.
1296 for (bd=stp->old_blocks; bd != NULL; bd = bd->link) {
1297 bd->u.bitmap = bitmap;
1298 bitmap += BLOCK_SIZE_W / (sizeof(W_)*BITS_PER_BYTE);
1300 // Also at this point we set the BF_MARKED flag
1301 // for this block. The invariant is that
1302 // BF_MARKED is always unset, except during GC
1303 // when it is set on those blocks which will be
1305 if (!(bd->flags & BF_FRAGMENTED)) {
1306 bd->flags |= BF_MARKED;
1313 // For each GC thread, for each step, allocate a "todo" block to
1314 // store evacuated objects to be scavenged, and a block to store
1315 // evacuated objects that do not need to be scavenged.
1316 for (t = 0; t < n_threads; t++) {
1317 for (s = 0; s < generations[g].n_steps; s++) {
1319 // we don't copy objects into g0s0, unless -G0
1320 if (g==0 && s==0 && RtsFlags.GcFlags.generations > 1) continue;
1322 ws = &gc_threads[t]->steps[g * RtsFlags.GcFlags.steps + s];
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;
1345 /* ----------------------------------------------------------------------------
1346 Initialise a generation that is *not* to be collected
1347 ------------------------------------------------------------------------- */
1350 init_uncollected_gen (nat g, nat threads)
1357 // save the current mutable lists for this generation, and
1358 // allocate a fresh block for each one. We'll traverse these
1359 // mutable lists as roots early on in the GC.
1360 generations[g].saved_mut_list = generations[g].mut_list;
1361 generations[g].mut_list = allocBlock();
1362 for (n = 0; n < n_capabilities; n++) {
1363 capabilities[n].saved_mut_lists[g] = capabilities[n].mut_lists[g];
1364 capabilities[n].mut_lists[g] = allocBlock();
1367 for (s = 0; s < generations[g].n_steps; s++) {
1368 stp = &generations[g].steps[s];
1369 stp->scavenged_large_objects = NULL;
1370 stp->n_scavenged_large_blocks = 0;
1373 for (s = 0; s < generations[g].n_steps; s++) {
1375 stp = &generations[g].steps[s];
1377 for (t = 0; t < threads; t++) {
1378 ws = &gc_threads[t]->steps[g * RtsFlags.GcFlags.steps + s];
1380 ASSERT(looksEmptyWSDeque(ws->todo_q));
1381 ws->todo_large_objects = NULL;
1383 ws->part_list = NULL;
1384 ws->n_part_blocks = 0;
1386 ws->scavd_list = NULL;
1387 ws->n_scavd_blocks = 0;
1389 // If the block at the head of the list in this generation
1390 // is less than 3/4 full, then use it as a todo block.
1391 if (stp->blocks && isPartiallyFull(stp->blocks))
1393 ws->todo_bd = stp->blocks;
1394 ws->todo_free = ws->todo_bd->free;
1395 ws->todo_lim = ws->todo_bd->start + BLOCK_SIZE_W;
1396 stp->blocks = stp->blocks->link;
1398 stp->n_words -= ws->todo_bd->free - ws->todo_bd->start;
1399 ws->todo_bd->link = NULL;
1400 // we must scan from the current end point.
1401 ws->todo_bd->u.scan = ws->todo_bd->free;
1406 alloc_todo_block(ws,0);
1410 // deal out any more partial blocks to the threads' part_lists
1412 while (stp->blocks && isPartiallyFull(stp->blocks))
1415 stp->blocks = bd->link;
1416 ws = &gc_threads[t]->steps[g * RtsFlags.GcFlags.steps + s];
1417 bd->link = ws->part_list;
1419 ws->n_part_blocks += 1;
1420 bd->u.scan = bd->free;
1422 stp->n_words -= bd->free - bd->start;
1424 if (t == n_gc_threads) t = 0;
1429 /* -----------------------------------------------------------------------------
1430 Initialise a gc_thread before GC
1431 -------------------------------------------------------------------------- */
1434 init_gc_thread (gc_thread *t)
1436 t->static_objects = END_OF_STATIC_LIST;
1437 t->scavenged_static_objects = END_OF_STATIC_LIST;
1439 t->mut_lists = capabilities[t->thread_index].mut_lists;
1441 t->failed_to_evac = rtsFalse;
1442 t->eager_promotion = rtsTrue;
1443 t->thunk_selector_depth = 0;
1448 t->scav_find_work = 0;
1451 /* -----------------------------------------------------------------------------
1452 Function we pass to evacuate roots.
1453 -------------------------------------------------------------------------- */
1456 mark_root(void *user USED_IF_THREADS, StgClosure **root)
1458 // we stole a register for gct, but this function is called from
1459 // *outside* the GC where the register variable is not in effect,
1460 // so we need to save and restore it here. NB. only call
1461 // mark_root() from the main GC thread, otherwise gct will be
1463 gc_thread *saved_gct;
1472 /* -----------------------------------------------------------------------------
1473 Initialising the static object & mutable lists
1474 -------------------------------------------------------------------------- */
1477 zero_static_object_list(StgClosure* first_static)
1481 const StgInfoTable *info;
1483 for (p = first_static; p != END_OF_STATIC_LIST; p = link) {
1485 link = *STATIC_LINK(info, p);
1486 *STATIC_LINK(info,p) = NULL;
1490 /* ----------------------------------------------------------------------------
1491 Update the pointers from the task list
1493 These are treated as weak pointers because we want to allow a main
1494 thread to get a BlockedOnDeadMVar exception in the same way as any
1495 other thread. Note that the threads should all have been retained
1496 by GC by virtue of being on the all_threads list, we're just
1497 updating pointers here.
1498 ------------------------------------------------------------------------- */
1501 update_task_list (void)
1505 for (task = all_tasks; task != NULL; task = task->all_link) {
1506 if (!task->stopped && task->tso) {
1507 ASSERT(task->tso->bound == task);
1508 tso = (StgTSO *) isAlive((StgClosure *)task->tso);
1510 barf("task %p: main thread %d has been GC'd",
1523 /* ----------------------------------------------------------------------------
1524 Reset the sizes of the older generations when we do a major
1527 CURRENT STRATEGY: make all generations except zero the same size.
1528 We have to stay within the maximum heap size, and leave a certain
1529 percentage of the maximum heap size available to allocate into.
1530 ------------------------------------------------------------------------- */
1533 resize_generations (void)
1537 if (major_gc && RtsFlags.GcFlags.generations > 1) {
1538 nat live, size, min_alloc, words;
1539 nat max = RtsFlags.GcFlags.maxHeapSize;
1540 nat gens = RtsFlags.GcFlags.generations;
1542 // live in the oldest generations
1543 if (oldest_gen->steps[0].live_estimate != 0) {
1544 words = oldest_gen->steps[0].live_estimate;
1546 words = oldest_gen->steps[0].n_words;
1548 live = (words + BLOCK_SIZE_W - 1) / BLOCK_SIZE_W +
1549 oldest_gen->steps[0].n_large_blocks;
1551 // default max size for all generations except zero
1552 size = stg_max(live * RtsFlags.GcFlags.oldGenFactor,
1553 RtsFlags.GcFlags.minOldGenSize);
1555 // minimum size for generation zero
1556 min_alloc = stg_max((RtsFlags.GcFlags.pcFreeHeap * max) / 200,
1557 RtsFlags.GcFlags.minAllocAreaSize);
1559 // Auto-enable compaction when the residency reaches a
1560 // certain percentage of the maximum heap size (default: 30%).
1561 if (RtsFlags.GcFlags.generations > 1 &&
1562 (RtsFlags.GcFlags.compact ||
1564 oldest_gen->steps[0].n_blocks >
1565 (RtsFlags.GcFlags.compactThreshold * max) / 100))) {
1566 oldest_gen->steps[0].mark = 1;
1567 oldest_gen->steps[0].compact = 1;
1568 // debugBelch("compaction: on\n", live);
1570 oldest_gen->steps[0].mark = 0;
1571 oldest_gen->steps[0].compact = 0;
1572 // debugBelch("compaction: off\n", live);
1575 if (RtsFlags.GcFlags.sweep) {
1576 oldest_gen->steps[0].mark = 1;
1579 // if we're going to go over the maximum heap size, reduce the
1580 // size of the generations accordingly. The calculation is
1581 // different if compaction is turned on, because we don't need
1582 // to double the space required to collect the old generation.
1585 // this test is necessary to ensure that the calculations
1586 // below don't have any negative results - we're working
1587 // with unsigned values here.
1588 if (max < min_alloc) {
1592 if (oldest_gen->steps[0].compact) {
1593 if ( (size + (size - 1) * (gens - 2) * 2) + min_alloc > max ) {
1594 size = (max - min_alloc) / ((gens - 1) * 2 - 1);
1597 if ( (size * (gens - 1) * 2) + min_alloc > max ) {
1598 size = (max - min_alloc) / ((gens - 1) * 2);
1608 debugBelch("live: %d, min_alloc: %d, size : %d, max = %d\n", live,
1609 min_alloc, size, max);
1612 for (g = 0; g < gens; g++) {
1613 generations[g].max_blocks = size;
1618 /* -----------------------------------------------------------------------------
1619 Calculate the new size of the nursery, and resize it.
1620 -------------------------------------------------------------------------- */
1623 resize_nursery (void)
1625 if (RtsFlags.GcFlags.generations == 1)
1626 { // Two-space collector:
1629 /* set up a new nursery. Allocate a nursery size based on a
1630 * function of the amount of live data (by default a factor of 2)
1631 * Use the blocks from the old nursery if possible, freeing up any
1634 * If we get near the maximum heap size, then adjust our nursery
1635 * size accordingly. If the nursery is the same size as the live
1636 * data (L), then we need 3L bytes. We can reduce the size of the
1637 * nursery to bring the required memory down near 2L bytes.
1639 * A normal 2-space collector would need 4L bytes to give the same
1640 * performance we get from 3L bytes, reducing to the same
1641 * performance at 2L bytes.
1643 blocks = g0s0->n_blocks;
1645 if ( RtsFlags.GcFlags.maxHeapSize != 0 &&
1646 blocks * RtsFlags.GcFlags.oldGenFactor * 2 >
1647 RtsFlags.GcFlags.maxHeapSize )
1649 long adjusted_blocks; // signed on purpose
1652 adjusted_blocks = (RtsFlags.GcFlags.maxHeapSize - 2 * blocks);
1654 debugTrace(DEBUG_gc, "near maximum heap size of 0x%x blocks, blocks = %d, adjusted to %ld",
1655 RtsFlags.GcFlags.maxHeapSize, blocks, adjusted_blocks);
1657 pc_free = adjusted_blocks * 100 / RtsFlags.GcFlags.maxHeapSize;
1658 if (pc_free < RtsFlags.GcFlags.pcFreeHeap) /* might even * be < 0 */
1662 blocks = adjusted_blocks;
1666 blocks *= RtsFlags.GcFlags.oldGenFactor;
1667 if (blocks < RtsFlags.GcFlags.minAllocAreaSize)
1669 blocks = RtsFlags.GcFlags.minAllocAreaSize;
1672 resizeNurseries(blocks);
1674 else // Generational collector
1677 * If the user has given us a suggested heap size, adjust our
1678 * allocation area to make best use of the memory available.
1680 if (RtsFlags.GcFlags.heapSizeSuggestion)
1683 nat needed = calcNeeded(); // approx blocks needed at next GC
1685 /* Guess how much will be live in generation 0 step 0 next time.
1686 * A good approximation is obtained by finding the
1687 * percentage of g0s0 that was live at the last minor GC.
1689 * We have an accurate figure for the amount of copied data in
1690 * 'copied', but we must convert this to a number of blocks, with
1691 * a small adjustment for estimated slop at the end of a block
1696 g0s0_pcnt_kept = ((copied / (BLOCK_SIZE_W - 10)) * 100)
1697 / countNurseryBlocks();
1700 /* Estimate a size for the allocation area based on the
1701 * information available. We might end up going slightly under
1702 * or over the suggested heap size, but we should be pretty
1705 * Formula: suggested - needed
1706 * ----------------------------
1707 * 1 + g0s0_pcnt_kept/100
1709 * where 'needed' is the amount of memory needed at the next
1710 * collection for collecting all steps except g0s0.
1713 (((long)RtsFlags.GcFlags.heapSizeSuggestion - (long)needed) * 100) /
1714 (100 + (long)g0s0_pcnt_kept);
1716 if (blocks < (long)RtsFlags.GcFlags.minAllocAreaSize) {
1717 blocks = RtsFlags.GcFlags.minAllocAreaSize;
1720 resizeNurseries((nat)blocks);
1724 // we might have added extra large blocks to the nursery, so
1725 // resize back to minAllocAreaSize again.
1726 resizeNurseriesFixed(RtsFlags.GcFlags.minAllocAreaSize);
1731 /* -----------------------------------------------------------------------------
1732 Sanity code for CAF garbage collection.
1734 With DEBUG turned on, we manage a CAF list in addition to the SRT
1735 mechanism. After GC, we run down the CAF list and blackhole any
1736 CAFs which have been garbage collected. This means we get an error
1737 whenever the program tries to enter a garbage collected CAF.
1739 Any garbage collected CAFs are taken off the CAF list at the same
1741 -------------------------------------------------------------------------- */
1743 #if 0 && defined(DEBUG)
1750 const StgInfoTable *info;
1761 ASSERT(info->type == IND_STATIC);
1763 if (STATIC_LINK(info,p) == NULL) {
1764 debugTrace(DEBUG_gccafs, "CAF gc'd at 0x%04lx", (long)p);
1766 SET_INFO(p,&stg_BLACKHOLE_info);
1767 p = STATIC_LINK2(info,p);
1771 pp = &STATIC_LINK2(info,p);
1778 debugTrace(DEBUG_gccafs, "%d CAFs live", i);