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"
19 #include "OSThreads.h"
20 #include "LdvProfile.h"
25 #include "BlockAlloc.h"
31 #include "ParTicky.h" // ToDo: move into Rts.h
32 #include "RtsSignals.h"
36 #if defined(RTS_GTK_FRONTPANEL)
37 #include "FrontPanel.h"
40 #include "RetainerProfile.h"
41 #include "RaiseAsync.h"
54 #include <string.h> // for memset()
57 /* -----------------------------------------------------------------------------
59 -------------------------------------------------------------------------- */
61 /* STATIC OBJECT LIST.
64 * We maintain a linked list of static objects that are still live.
65 * The requirements for this list are:
67 * - we need to scan the list while adding to it, in order to
68 * scavenge all the static objects (in the same way that
69 * breadth-first scavenging works for dynamic objects).
71 * - we need to be able to tell whether an object is already on
72 * the list, to break loops.
74 * Each static object has a "static link field", which we use for
75 * linking objects on to the list. We use a stack-type list, consing
76 * objects on the front as they are added (this means that the
77 * scavenge phase is depth-first, not breadth-first, but that
80 * A separate list is kept for objects that have been scavenged
81 * already - this is so that we can zero all the marks afterwards.
83 * An object is on the list if its static link field is non-zero; this
84 * means that we have to mark the end of the list with '1', not NULL.
86 * Extra notes for generational GC:
88 * Each generation has a static object list associated with it. When
89 * collecting generations up to N, we treat the static object lists
90 * from generations > N as roots.
92 * We build up a static object list while collecting generations 0..N,
93 * which is then appended to the static object list of generation N+1.
96 /* N is the oldest generation being collected, where the generations
97 * are numbered starting at 0. A major GC (indicated by the major_gc
98 * flag) is when we're collecting all generations. We only attempt to
99 * deal with static objects and GC CAFs when doing a major GC.
104 /* Data used for allocation area sizing.
106 static lnat g0s0_pcnt_kept = 30; // percentage of g0s0 live at last minor GC
116 /* Thread-local data for each GC thread
118 gc_thread **gc_threads = NULL;
120 #if !defined(THREADED_RTS)
121 StgWord8 the_gc_thread[sizeof(gc_thread) + 64 * sizeof(step_workspace)];
124 // Number of threads running in *this* GC. Affects how many
125 // step->todos[] lists we have to look in to find work.
129 long copied; // *words* copied & scavenged during this GC
131 rtsBool work_stealing;
135 /* -----------------------------------------------------------------------------
136 Static function declarations
137 -------------------------------------------------------------------------- */
139 static void mark_root (void *user, StgClosure **root);
140 static void zero_static_object_list (StgClosure* first_static);
141 static nat initialise_N (rtsBool force_major_gc);
142 static void init_collected_gen (nat g, nat threads);
143 static void init_uncollected_gen (nat g, nat threads);
144 static void init_gc_thread (gc_thread *t);
145 static void update_task_list (void);
146 static void resize_generations (void);
147 static void resize_nursery (void);
148 static void start_gc_threads (void);
149 static void scavenge_until_all_done (void);
150 static StgWord inc_running (void);
151 static StgWord dec_running (void);
152 static void wakeup_gc_threads (nat n_threads, nat me);
153 static void shutdown_gc_threads (nat n_threads, nat me);
155 #if 0 && defined(DEBUG)
156 static void gcCAFs (void);
159 /* -----------------------------------------------------------------------------
160 The mark bitmap & stack.
161 -------------------------------------------------------------------------- */
163 #define MARK_STACK_BLOCKS 4
165 bdescr *mark_stack_bdescr;
170 // Flag and pointers used for falling back to a linear scan when the
171 // mark stack overflows.
172 rtsBool mark_stack_overflowed;
173 bdescr *oldgen_scan_bd;
176 /* -----------------------------------------------------------------------------
177 GarbageCollect: the main entry point to the garbage collector.
179 Locks held: all capabilities are held throughout GarbageCollect().
180 -------------------------------------------------------------------------- */
183 GarbageCollect (rtsBool force_major_gc,
184 nat gc_type USED_IF_THREADS,
189 lnat live, allocated, max_copied, avg_copied, slop;
190 gc_thread *saved_gct;
193 // necessary if we stole a callee-saves register for gct:
197 CostCentreStack *prev_CCS;
202 #if defined(RTS_USER_SIGNALS)
203 if (RtsFlags.MiscFlags.install_signal_handlers) {
209 ASSERT(sizeof(step_workspace) == 16 * sizeof(StgWord));
210 // otherwise adjust the padding in step_workspace.
212 // tell the stats department that we've started a GC
215 // tell the STM to discard any cached closures it's hoping to re-use
218 // lock the StablePtr table
227 // attribute any costs to CCS_GC
233 /* Approximate how much we allocated.
234 * Todo: only when generating stats?
236 allocated = calcAllocated();
238 /* Figure out which generation to collect
240 n = initialise_N(force_major_gc);
242 #if defined(THREADED_RTS)
243 work_stealing = RtsFlags.ParFlags.parGcLoadBalancing;
244 // It's not always a good idea to do load balancing in parallel
245 // GC. In particular, for a parallel program we don't want to
246 // lose locality by moving cached data into another CPU's cache
247 // (this effect can be quite significant).
249 // We could have a more complex way to deterimine whether to do
250 // work stealing or not, e.g. it might be a good idea to do it
251 // if the heap is big. For now, we just turn it on or off with
255 /* Start threads, so they can be spinning up while we finish initialisation.
259 #if defined(THREADED_RTS)
260 /* How many threads will be participating in this GC?
261 * We don't try to parallelise minor GCs (unless the user asks for
262 * it with +RTS -gn0), or mark/compact/sweep GC.
264 if (gc_type == PENDING_GC_PAR) {
265 n_gc_threads = RtsFlags.ParFlags.nNodes;
273 debugTrace(DEBUG_gc, "GC (gen %d): %d KB to collect, %ld MB in use, using %d thread(s)",
274 N, n * (BLOCK_SIZE / 1024), mblocks_allocated, n_gc_threads);
276 #ifdef RTS_GTK_FRONTPANEL
277 if (RtsFlags.GcFlags.frontpanel) {
278 updateFrontPanelBeforeGC(N);
283 // check for memory leaks if DEBUG is on
284 memInventory(traceClass(DEBUG_gc));
287 // check stack sanity *before* GC
288 IF_DEBUG(sanity, checkFreeListSanity());
289 IF_DEBUG(sanity, checkMutableLists(rtsTrue));
291 // Initialise all our gc_thread structures
292 for (t = 0; t < n_gc_threads; t++) {
293 init_gc_thread(gc_threads[t]);
296 // Initialise all the generations/steps that we're collecting.
297 for (g = 0; g <= N; g++) {
298 init_collected_gen(g,n_gc_threads);
301 // Initialise all the generations/steps that we're *not* collecting.
302 for (g = N+1; g < RtsFlags.GcFlags.generations; g++) {
303 init_uncollected_gen(g,n_gc_threads);
306 /* Allocate a mark stack if we're doing a major collection.
308 if (major_gc && oldest_gen->steps[0].mark) {
309 nat mark_stack_blocks;
310 mark_stack_blocks = stg_max(MARK_STACK_BLOCKS,
311 oldest_gen->steps[0].n_old_blocks / 100);
312 mark_stack_bdescr = allocGroup(mark_stack_blocks);
313 mark_stack = (StgPtr *)mark_stack_bdescr->start;
314 mark_sp = mark_stack;
315 mark_splim = mark_stack + (mark_stack_blocks * BLOCK_SIZE_W);
317 mark_stack_bdescr = NULL;
320 // this is the main thread
322 if (n_gc_threads == 1) {
323 SET_GCT(gc_threads[0]);
325 SET_GCT(gc_threads[cap->no]);
328 SET_GCT(gc_threads[0]);
331 /* -----------------------------------------------------------------------
332 * follow all the roots that we know about:
335 // the main thread is running: this prevents any other threads from
336 // exiting prematurely, so we can start them now.
337 // NB. do this after the mutable lists have been saved above, otherwise
338 // the other GC threads will be writing into the old mutable lists.
340 wakeup_gc_threads(n_gc_threads, gct->thread_index);
342 // Mutable lists from each generation > N
343 // we want to *scavenge* these roots, not evacuate them: they're not
344 // going to move in this GC.
345 // Also do them in reverse generation order, for the usual reason:
346 // namely to reduce the likelihood of spurious old->new pointers.
348 for (g = RtsFlags.GcFlags.generations-1; g > N; g--) {
349 scavenge_mutable_list(generations[g].saved_mut_list, &generations[g]);
350 freeChain_sync(generations[g].saved_mut_list);
351 generations[g].saved_mut_list = NULL;
355 // scavenge the capability-private mutable lists. This isn't part
356 // of markSomeCapabilities() because markSomeCapabilities() can only
357 // call back into the GC via mark_root() (due to the gct register
359 if (n_gc_threads == 1) {
360 for (n = 0; n < n_capabilities; n++) {
361 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 // if any blackholes are alive, make the threads that wait on
400 if (traverseBlackholeQueue()) {
405 // must be last... invariant is that everything is fully
406 // scavenged at this point.
407 if (traverseWeakPtrList()) { // returns rtsTrue if evaced something
412 // If we get to here, there's really nothing left to do.
416 shutdown_gc_threads(n_gc_threads, gct->thread_index);
418 // Update pointers from the Task list
421 // Now see which stable names are still alive.
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 // g0s0->old_blocks is the old nursery
436 // g0s0->blocks is to-space from the previous GC
437 if (RtsFlags.GcFlags.generations == 1) {
438 if (g0s0->blocks != NULL) {
439 freeChain(g0s0->blocks);
444 // For each workspace, in each thread, move the copied blocks to the step
450 for (t = 0; t < n_gc_threads; t++) {
454 if (RtsFlags.GcFlags.generations == 1) {
459 for (; s < total_steps; s++) {
462 // Push the final block
464 push_scanned_block(ws->todo_bd, ws);
467 ASSERT(gct->scan_bd == NULL);
468 ASSERT(countBlocks(ws->scavd_list) == ws->n_scavd_blocks);
471 for (bd = ws->scavd_list; bd != NULL; bd = bd->link) {
472 ws->step->n_words += bd->free - bd->start;
476 prev->link = ws->step->blocks;
477 ws->step->blocks = ws->scavd_list;
479 ws->step->n_blocks += ws->n_scavd_blocks;
483 // Add all the partial blocks *after* we've added all the full
484 // blocks. This is so that we can grab the partial blocks back
485 // again and try to fill them up in the next GC.
486 for (t = 0; t < n_gc_threads; t++) {
490 if (RtsFlags.GcFlags.generations == 1) {
495 for (; s < total_steps; s++) {
499 for (bd = ws->part_list; bd != NULL; bd = next) {
501 if (bd->free == bd->start) {
503 ws->part_list = next;
510 ws->step->n_words += bd->free - bd->start;
515 prev->link = ws->step->blocks;
516 ws->step->blocks = ws->part_list;
518 ws->step->n_blocks += ws->n_part_blocks;
520 ASSERT(countBlocks(ws->step->blocks) == ws->step->n_blocks);
521 ASSERT(countOccupied(ws->step->blocks) == ws->step->n_words);
526 // Finally: compact or sweep the oldest generation.
527 if (major_gc && oldest_gen->steps[0].mark) {
528 if (oldest_gen->steps[0].compact)
529 compact(gct->scavenged_static_objects);
531 sweep(&oldest_gen->steps[0]);
534 /* run through all the generations/steps and tidy up
541 for (i=0; i < n_gc_threads; i++) {
542 if (n_gc_threads > 1) {
543 debugTrace(DEBUG_gc,"thread %d:", i);
544 debugTrace(DEBUG_gc," copied %ld", gc_threads[i]->copied * sizeof(W_));
545 debugTrace(DEBUG_gc," scanned %ld", gc_threads[i]->scanned * sizeof(W_));
546 debugTrace(DEBUG_gc," any_work %ld", gc_threads[i]->any_work);
547 debugTrace(DEBUG_gc," no_work %ld", gc_threads[i]->no_work);
548 debugTrace(DEBUG_gc," scav_find_work %ld", gc_threads[i]->scav_find_work);
550 copied += gc_threads[i]->copied;
551 max_copied = stg_max(gc_threads[i]->copied, max_copied);
553 if (n_gc_threads == 1) {
561 for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
564 generations[g].collections++; // for stats
565 if (n_gc_threads > 1) generations[g].par_collections++;
568 // Count the mutable list as bytes "copied" for the purposes of
569 // stats. Every mutable list is copied during every GC.
571 nat mut_list_size = 0;
572 for (bd = generations[g].mut_list; bd != NULL; bd = bd->link) {
573 mut_list_size += bd->free - bd->start;
575 for (n = 0; n < n_capabilities; n++) {
576 for (bd = capabilities[n].mut_lists[g];
577 bd != NULL; bd = bd->link) {
578 mut_list_size += bd->free - bd->start;
581 copied += mut_list_size;
584 "mut_list_size: %lu (%d vars, %d arrays, %d MVARs, %d others)",
585 (unsigned long)(mut_list_size * sizeof(W_)),
586 mutlist_MUTVARS, mutlist_MUTARRS, mutlist_MVARS, mutlist_OTHERS);
589 for (s = 0; s < generations[g].n_steps; s++) {
591 stp = &generations[g].steps[s];
593 // for generations we collected...
596 /* free old memory and shift to-space into from-space for all
597 * the collected steps (except the allocation area). These
598 * freed blocks will probaby be quickly recycled.
600 if (!(g == 0 && s == 0 && RtsFlags.GcFlags.generations > 1)) {
603 // tack the new blocks on the end of the existing blocks
604 if (stp->old_blocks != NULL) {
607 for (bd = stp->old_blocks; bd != NULL; bd = next) {
611 if (!(bd->flags & BF_MARKED))
614 stp->old_blocks = next;
623 stp->n_words += bd->free - bd->start;
625 // NB. this step might not be compacted next
626 // time, so reset the BF_MARKED flags.
627 // They are set before GC if we're going to
628 // compact. (search for BF_MARKED above).
629 bd->flags &= ~BF_MARKED;
631 // between GCs, all blocks in the heap except
632 // for the nursery have the BF_EVACUATED flag set.
633 bd->flags |= BF_EVACUATED;
640 prev->link = stp->blocks;
641 stp->blocks = stp->old_blocks;
644 // add the new blocks to the block tally
645 stp->n_blocks += stp->n_old_blocks;
646 ASSERT(countBlocks(stp->blocks) == stp->n_blocks);
647 ASSERT(countOccupied(stp->blocks) == stp->n_words);
651 freeChain(stp->old_blocks);
653 stp->old_blocks = NULL;
654 stp->n_old_blocks = 0;
657 /* LARGE OBJECTS. The current live large objects are chained on
658 * scavenged_large, having been moved during garbage
659 * collection from large_objects. Any objects left on
660 * large_objects list are therefore dead, so we free them here.
662 for (bd = stp->large_objects; bd != NULL; bd = next) {
668 stp->large_objects = stp->scavenged_large_objects;
669 stp->n_large_blocks = stp->n_scavenged_large_blocks;
672 else // for older generations...
674 /* For older generations, we need to append the
675 * scavenged_large_object list (i.e. large objects that have been
676 * promoted during this GC) to the large_object list for that step.
678 for (bd = stp->scavenged_large_objects; bd; bd = next) {
680 dbl_link_onto(bd, &stp->large_objects);
683 // add the new blocks we promoted during this GC
684 stp->n_large_blocks += stp->n_scavenged_large_blocks;
689 // update the max size of older generations after a major GC
690 resize_generations();
692 // Calculate the amount of live data for stats.
693 live = calcLiveWords();
695 // Free the small objects allocated via allocate(), since this will
696 // all have been copied into G0S1 now.
697 if (RtsFlags.GcFlags.generations > 1) {
698 if (g0s0->blocks != NULL) {
699 freeChain(g0s0->blocks);
706 alloc_blocks_lim = RtsFlags.GcFlags.minAllocAreaSize;
708 // Start a new pinned_object_block
709 pinned_object_block = NULL;
711 // Free the mark stack.
712 if (mark_stack_bdescr != NULL) {
713 freeGroup(mark_stack_bdescr);
717 for (g = 0; g <= N; g++) {
718 for (s = 0; s < generations[g].n_steps; s++) {
719 stp = &generations[g].steps[s];
720 if (stp->bitmap != NULL) {
721 freeGroup(stp->bitmap);
729 // mark the garbage collected CAFs as dead
730 #if 0 && defined(DEBUG) // doesn't work at the moment
731 if (major_gc) { gcCAFs(); }
735 // resetStaticObjectForRetainerProfiling() must be called before
737 if (n_gc_threads > 1) {
738 barf("profiling is currently broken with multi-threaded GC");
739 // ToDo: fix the gct->scavenged_static_objects below
741 resetStaticObjectForRetainerProfiling(gct->scavenged_static_objects);
744 // zero the scavenged static object list
747 for (i = 0; i < n_gc_threads; i++) {
748 zero_static_object_list(gc_threads[i]->scavenged_static_objects);
755 // start any pending finalizers
757 scheduleFinalizers(cap, old_weak_ptr_list);
760 // send exceptions to any threads which were about to die
762 resurrectThreads(resurrected_threads);
763 performPendingThrowTos(exception_threads);
766 // Update the stable pointer hash table.
767 updateStablePtrTable(major_gc);
769 // check sanity after GC
770 IF_DEBUG(sanity, checkSanity());
772 // extra GC trace info
773 IF_DEBUG(gc, statDescribeGens());
776 // symbol-table based profiling
777 /* heapCensus(to_blocks); */ /* ToDo */
780 // restore enclosing cost centre
786 // check for memory leaks if DEBUG is on
787 memInventory(traceClass(DEBUG_gc));
790 #ifdef RTS_GTK_FRONTPANEL
791 if (RtsFlags.GcFlags.frontpanel) {
792 updateFrontPanelAfterGC( N, live );
796 // ok, GC over: tell the stats department what happened.
797 slop = calcLiveBlocks() * BLOCK_SIZE_W - live;
798 stat_endGC(allocated, live, copied, N, max_copied, avg_copied, slop);
800 // unlock the StablePtr table
803 // Guess which generation we'll collect *next* time
804 initialise_N(force_major_gc);
806 #if defined(RTS_USER_SIGNALS)
807 if (RtsFlags.MiscFlags.install_signal_handlers) {
808 // unblock signals again
809 unblockUserSignals();
818 /* -----------------------------------------------------------------------------
819 Figure out which generation to collect, initialise N and major_gc.
821 Also returns the total number of blocks in generations that will be
823 -------------------------------------------------------------------------- */
826 initialise_N (rtsBool force_major_gc)
829 nat s, blocks, blocks_total;
834 if (force_major_gc) {
835 N = RtsFlags.GcFlags.generations - 1;
840 for (g = RtsFlags.GcFlags.generations - 1; g >= 0; g--) {
842 for (s = 0; s < generations[g].n_steps; s++) {
843 blocks += generations[g].steps[s].n_words / BLOCK_SIZE_W;
844 blocks += generations[g].steps[s].n_large_blocks;
846 if (blocks >= generations[g].max_blocks) {
850 blocks_total += blocks;
854 blocks_total += countNurseryBlocks();
856 major_gc = (N == RtsFlags.GcFlags.generations-1);
860 /* -----------------------------------------------------------------------------
861 Initialise the gc_thread structures.
862 -------------------------------------------------------------------------- */
864 #define GC_THREAD_INACTIVE 0
865 #define GC_THREAD_STANDING_BY 1
866 #define GC_THREAD_RUNNING 2
867 #define GC_THREAD_WAITING_TO_CONTINUE 3
870 new_gc_thread (nat n, gc_thread *t)
877 initSpinLock(&t->gc_spin);
878 initSpinLock(&t->mut_spin);
879 ACQUIRE_SPIN_LOCK(&t->gc_spin);
880 t->wakeup = GC_THREAD_INACTIVE; // starts true, so we can wait for the
881 // thread to start up, see wakeup_gc_threads
885 t->free_blocks = NULL;
894 for (s = 0; s < total_steps; s++)
897 ws->step = &all_steps[s];
898 ASSERT(s == ws->step->abs_no);
902 ws->todo_q = newWSDeque(128);
903 ws->todo_overflow = NULL;
904 ws->n_todo_overflow = 0;
906 ws->part_list = NULL;
907 ws->n_part_blocks = 0;
909 ws->scavd_list = NULL;
910 ws->n_scavd_blocks = 0;
918 if (gc_threads == NULL) {
919 #if defined(THREADED_RTS)
921 gc_threads = stgMallocBytes (RtsFlags.ParFlags.nNodes *
925 for (i = 0; i < RtsFlags.ParFlags.nNodes; i++) {
927 stgMallocBytes(sizeof(gc_thread) + total_steps * sizeof(step_workspace),
930 new_gc_thread(i, gc_threads[i]);
933 gc_threads = stgMallocBytes (sizeof(gc_thread*),"alloc_gc_threads");
935 new_gc_thread(0,gc_threads[0]);
943 if (gc_threads != NULL) {
944 #if defined(THREADED_RTS)
946 for (i = 0; i < RtsFlags.ParFlags.nNodes; i++) {
947 stgFree (gc_threads[i]);
949 stgFree (gc_threads);
951 stgFree (gc_threads);
957 /* ----------------------------------------------------------------------------
959 ------------------------------------------------------------------------- */
961 static volatile StgWord gc_running_threads;
967 new = atomic_inc(&gc_running_threads);
968 ASSERT(new <= n_gc_threads);
975 ASSERT(gc_running_threads != 0);
976 return atomic_dec(&gc_running_threads);
989 // scavenge objects in compacted generation
990 if (mark_stack_overflowed || oldgen_scan_bd != NULL ||
991 (mark_stack_bdescr != NULL && !mark_stack_empty())) {
995 // Check for global work in any step. We don't need to check for
996 // local work, because we have already exited scavenge_loop(),
997 // which means there is no local work for this thread.
998 for (s = total_steps-1; s >= 0; s--) {
999 if (s == 0 && RtsFlags.GcFlags.generations > 1) {
1002 ws = &gct->steps[s];
1003 if (ws->todo_large_objects) return rtsTrue;
1004 if (!looksEmptyWSDeque(ws->todo_q)) return rtsTrue;
1005 if (ws->todo_overflow) return rtsTrue;
1008 #if defined(THREADED_RTS)
1009 if (work_stealing) {
1011 // look for work to steal
1012 for (n = 0; n < n_gc_threads; n++) {
1013 if (n == gct->thread_index) continue;
1014 for (s = total_steps-1; s >= 0; s--) {
1015 ws = &gc_threads[n]->steps[s];
1016 if (!looksEmptyWSDeque(ws->todo_q)) return rtsTrue;
1028 scavenge_until_all_done (void)
1032 debugTrace(DEBUG_gc, "GC thread %d working", gct->thread_index);
1035 #if defined(THREADED_RTS)
1036 if (n_gc_threads > 1) {
1045 // scavenge_loop() only exits when there's no work to do
1048 debugTrace(DEBUG_gc, "GC thread %d idle (%d still running)",
1049 gct->thread_index, 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 // All threads are now stopped
1064 debugTrace(DEBUG_gc, "GC thread %d finished.", gct->thread_index);
1067 #if defined(THREADED_RTS)
1070 gcWorkerThread (Capability *cap)
1072 cap->in_gc = rtsTrue;
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 // count events in this thread towards the GC totals
1101 papi_thread_stop_gc1_count(gct->papi_events);
1104 // Wait until we're told to continue
1105 RELEASE_SPIN_LOCK(&gct->gc_spin);
1106 gct->wakeup = GC_THREAD_WAITING_TO_CONTINUE;
1107 debugTrace(DEBUG_gc, "GC thread %d waiting to continue...",
1109 ACQUIRE_SPIN_LOCK(&gct->mut_spin);
1110 debugTrace(DEBUG_gc, "GC thread %d on my way...", gct->thread_index);
1116 waitForGcThreads (Capability *cap USED_IF_THREADS)
1118 #if defined(THREADED_RTS)
1119 nat n_threads = RtsFlags.ParFlags.nNodes;
1122 rtsBool retry = rtsTrue;
1125 for (i=0; i < n_threads; i++) {
1126 if (i == me) continue;
1127 if (gc_threads[i]->wakeup != GC_THREAD_STANDING_BY) {
1128 prodCapability(&capabilities[i], cap->running_task);
1131 for (j=0; j < 10000000; j++) {
1133 for (i=0; i < n_threads; i++) {
1134 if (i == me) continue;
1136 setContextSwitches();
1137 if (gc_threads[i]->wakeup != GC_THREAD_STANDING_BY) {
1148 start_gc_threads (void)
1150 #if defined(THREADED_RTS)
1151 gc_running_threads = 0;
1156 wakeup_gc_threads (nat n_threads USED_IF_THREADS, nat me USED_IF_THREADS)
1158 #if defined(THREADED_RTS)
1160 for (i=0; i < n_threads; i++) {
1161 if (i == me) continue;
1163 debugTrace(DEBUG_gc, "waking up gc thread %d", i);
1164 if (gc_threads[i]->wakeup != GC_THREAD_STANDING_BY) barf("wakeup_gc_threads");
1166 gc_threads[i]->wakeup = GC_THREAD_RUNNING;
1167 ACQUIRE_SPIN_LOCK(&gc_threads[i]->mut_spin);
1168 RELEASE_SPIN_LOCK(&gc_threads[i]->gc_spin);
1173 // After GC is complete, we must wait for all GC threads to enter the
1174 // standby state, otherwise they may still be executing inside
1175 // any_work(), and may even remain awake until the next GC starts.
1177 shutdown_gc_threads (nat n_threads USED_IF_THREADS, nat me USED_IF_THREADS)
1179 #if defined(THREADED_RTS)
1181 for (i=0; i < n_threads; i++) {
1182 if (i == me) continue;
1183 while (gc_threads[i]->wakeup != GC_THREAD_WAITING_TO_CONTINUE) { write_barrier(); }
1189 releaseGCThreads (Capability *cap USED_IF_THREADS)
1191 #if defined(THREADED_RTS)
1192 nat n_threads = RtsFlags.ParFlags.nNodes;
1195 for (i=0; i < n_threads; i++) {
1196 if (i == me) continue;
1197 if (gc_threads[i]->wakeup != GC_THREAD_WAITING_TO_CONTINUE)
1198 barf("releaseGCThreads");
1200 gc_threads[i]->wakeup = GC_THREAD_INACTIVE;
1201 ACQUIRE_SPIN_LOCK(&gc_threads[i]->gc_spin);
1202 RELEASE_SPIN_LOCK(&gc_threads[i]->mut_spin);
1207 /* ----------------------------------------------------------------------------
1208 Initialise a generation that is to be collected
1209 ------------------------------------------------------------------------- */
1212 init_collected_gen (nat g, nat n_threads)
1219 // Throw away the current mutable list. Invariant: the mutable
1220 // list always has at least one block; this means we can avoid a
1221 // check for NULL in recordMutable().
1223 freeChain(generations[g].mut_list);
1224 generations[g].mut_list = allocBlock();
1225 for (i = 0; i < n_capabilities; i++) {
1226 freeChain(capabilities[i].mut_lists[g]);
1227 capabilities[i].mut_lists[g] = allocBlock();
1231 for (s = 0; s < generations[g].n_steps; s++) {
1233 stp = &generations[g].steps[s];
1234 ASSERT(stp->gen_no == g);
1236 // we'll construct a new list of threads in this step
1237 // during GC, throw away the current list.
1238 stp->old_threads = stp->threads;
1239 stp->threads = END_TSO_QUEUE;
1241 // generation 0, step 0 doesn't need to-space
1242 if (g == 0 && s == 0 && RtsFlags.GcFlags.generations > 1) {
1246 // deprecate the existing blocks
1247 stp->old_blocks = stp->blocks;
1248 stp->n_old_blocks = stp->n_blocks;
1252 stp->live_estimate = 0;
1254 // initialise the large object queues.
1255 stp->scavenged_large_objects = NULL;
1256 stp->n_scavenged_large_blocks = 0;
1258 // mark the small objects as from-space
1259 for (bd = stp->old_blocks; bd; bd = bd->link) {
1260 bd->flags &= ~BF_EVACUATED;
1263 // mark the large objects as from-space
1264 for (bd = stp->large_objects; bd; bd = bd->link) {
1265 bd->flags &= ~BF_EVACUATED;
1268 // for a compacted step, we need to allocate the bitmap
1270 nat bitmap_size; // in bytes
1271 bdescr *bitmap_bdescr;
1274 bitmap_size = stp->n_old_blocks * BLOCK_SIZE / (sizeof(W_)*BITS_PER_BYTE);
1276 if (bitmap_size > 0) {
1277 bitmap_bdescr = allocGroup((lnat)BLOCK_ROUND_UP(bitmap_size)
1279 stp->bitmap = bitmap_bdescr;
1280 bitmap = bitmap_bdescr->start;
1282 debugTrace(DEBUG_gc, "bitmap_size: %d, bitmap: %p",
1283 bitmap_size, bitmap);
1285 // don't forget to fill it with zeros!
1286 memset(bitmap, 0, bitmap_size);
1288 // For each block in this step, point to its bitmap from the
1289 // block descriptor.
1290 for (bd=stp->old_blocks; bd != NULL; bd = bd->link) {
1291 bd->u.bitmap = bitmap;
1292 bitmap += BLOCK_SIZE_W / (sizeof(W_)*BITS_PER_BYTE);
1294 // Also at this point we set the BF_MARKED flag
1295 // for this block. The invariant is that
1296 // BF_MARKED is always unset, except during GC
1297 // when it is set on those blocks which will be
1299 if (!(bd->flags & BF_FRAGMENTED)) {
1300 bd->flags |= BF_MARKED;
1307 // For each GC thread, for each step, allocate a "todo" block to
1308 // store evacuated objects to be scavenged, and a block to store
1309 // evacuated objects that do not need to be scavenged.
1310 for (t = 0; t < n_threads; t++) {
1311 for (s = 0; s < generations[g].n_steps; s++) {
1313 // we don't copy objects into g0s0, unless -G0
1314 if (g==0 && s==0 && RtsFlags.GcFlags.generations > 1) continue;
1316 ws = &gc_threads[t]->steps[g * RtsFlags.GcFlags.steps + s];
1318 ws->todo_large_objects = NULL;
1320 ws->part_list = NULL;
1321 ws->n_part_blocks = 0;
1323 // allocate the first to-space block; extra blocks will be
1324 // chained on as necessary.
1326 ASSERT(looksEmptyWSDeque(ws->todo_q));
1327 alloc_todo_block(ws,0);
1329 ws->todo_overflow = NULL;
1330 ws->n_todo_overflow = 0;
1332 ws->scavd_list = NULL;
1333 ws->n_scavd_blocks = 0;
1339 /* ----------------------------------------------------------------------------
1340 Initialise a generation that is *not* to be collected
1341 ------------------------------------------------------------------------- */
1344 init_uncollected_gen (nat g, nat threads)
1351 // save the current mutable lists for this generation, and
1352 // allocate a fresh block for each one. We'll traverse these
1353 // mutable lists as roots early on in the GC.
1354 generations[g].saved_mut_list = generations[g].mut_list;
1355 generations[g].mut_list = allocBlock();
1356 for (n = 0; n < n_capabilities; n++) {
1357 capabilities[n].saved_mut_lists[g] = capabilities[n].mut_lists[g];
1358 capabilities[n].mut_lists[g] = allocBlock();
1361 for (s = 0; s < generations[g].n_steps; s++) {
1362 stp = &generations[g].steps[s];
1363 stp->scavenged_large_objects = NULL;
1364 stp->n_scavenged_large_blocks = 0;
1367 for (s = 0; s < generations[g].n_steps; s++) {
1369 stp = &generations[g].steps[s];
1371 for (t = 0; t < threads; t++) {
1372 ws = &gc_threads[t]->steps[g * RtsFlags.GcFlags.steps + s];
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 (stp->blocks && isPartiallyFull(stp->blocks))
1387 ws->todo_bd = stp->blocks;
1388 ws->todo_free = ws->todo_bd->free;
1389 ws->todo_lim = ws->todo_bd->start + BLOCK_SIZE_W;
1390 stp->blocks = stp->blocks->link;
1392 stp->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 (stp->blocks && isPartiallyFull(stp->blocks))
1409 stp->blocks = bd->link;
1410 ws = &gc_threads[t]->steps[g * RtsFlags.GcFlags.steps + s];
1411 bd->link = ws->part_list;
1413 ws->n_part_blocks += 1;
1414 bd->u.scan = bd->free;
1416 stp->n_words -= bd->free - bd->start;
1418 if (t == n_gc_threads) t = 0;
1423 /* -----------------------------------------------------------------------------
1424 Initialise a gc_thread before GC
1425 -------------------------------------------------------------------------- */
1428 init_gc_thread (gc_thread *t)
1430 t->static_objects = END_OF_STATIC_LIST;
1431 t->scavenged_static_objects = END_OF_STATIC_LIST;
1433 t->mut_lists = capabilities[t->thread_index].mut_lists;
1435 t->failed_to_evac = rtsFalse;
1436 t->eager_promotion = rtsTrue;
1437 t->thunk_selector_depth = 0;
1442 t->scav_find_work = 0;
1445 /* -----------------------------------------------------------------------------
1446 Function we pass to evacuate roots.
1447 -------------------------------------------------------------------------- */
1450 mark_root(void *user USED_IF_THREADS, StgClosure **root)
1452 // we stole a register for gct, but this function is called from
1453 // *outside* the GC where the register variable is not in effect,
1454 // so we need to save and restore it here. NB. only call
1455 // mark_root() from the main GC thread, otherwise gct will be
1457 gc_thread *saved_gct;
1466 /* -----------------------------------------------------------------------------
1467 Initialising the static object & mutable lists
1468 -------------------------------------------------------------------------- */
1471 zero_static_object_list(StgClosure* first_static)
1475 const StgInfoTable *info;
1477 for (p = first_static; p != END_OF_STATIC_LIST; p = link) {
1479 link = *STATIC_LINK(info, p);
1480 *STATIC_LINK(info,p) = NULL;
1484 /* ----------------------------------------------------------------------------
1485 Update the pointers from the task list
1487 These are treated as weak pointers because we want to allow a main
1488 thread to get a BlockedOnDeadMVar exception in the same way as any
1489 other thread. Note that the threads should all have been retained
1490 by GC by virtue of being on the all_threads list, we're just
1491 updating pointers here.
1492 ------------------------------------------------------------------------- */
1495 update_task_list (void)
1499 for (task = all_tasks; task != NULL; task = task->all_link) {
1500 if (!task->stopped && task->tso) {
1501 ASSERT(task->tso->bound == task);
1502 tso = (StgTSO *) isAlive((StgClosure *)task->tso);
1504 barf("task %p: main thread %d has been GC'd",
1517 /* ----------------------------------------------------------------------------
1518 Reset the sizes of the older generations when we do a major
1521 CURRENT STRATEGY: make all generations except zero the same size.
1522 We have to stay within the maximum heap size, and leave a certain
1523 percentage of the maximum heap size available to allocate into.
1524 ------------------------------------------------------------------------- */
1527 resize_generations (void)
1531 if (major_gc && RtsFlags.GcFlags.generations > 1) {
1532 nat live, size, min_alloc, words;
1533 nat max = RtsFlags.GcFlags.maxHeapSize;
1534 nat gens = RtsFlags.GcFlags.generations;
1536 // live in the oldest generations
1537 if (oldest_gen->steps[0].live_estimate != 0) {
1538 words = oldest_gen->steps[0].live_estimate;
1540 words = oldest_gen->steps[0].n_words;
1542 live = (words + BLOCK_SIZE_W - 1) / BLOCK_SIZE_W +
1543 oldest_gen->steps[0].n_large_blocks;
1545 // default max size for all generations except zero
1546 size = stg_max(live * RtsFlags.GcFlags.oldGenFactor,
1547 RtsFlags.GcFlags.minOldGenSize);
1549 // minimum size for generation zero
1550 min_alloc = stg_max((RtsFlags.GcFlags.pcFreeHeap * max) / 200,
1551 RtsFlags.GcFlags.minAllocAreaSize);
1553 // Auto-enable compaction when the residency reaches a
1554 // certain percentage of the maximum heap size (default: 30%).
1555 if (RtsFlags.GcFlags.generations > 1 &&
1556 (RtsFlags.GcFlags.compact ||
1558 oldest_gen->steps[0].n_blocks >
1559 (RtsFlags.GcFlags.compactThreshold * max) / 100))) {
1560 oldest_gen->steps[0].mark = 1;
1561 oldest_gen->steps[0].compact = 1;
1562 // debugBelch("compaction: on\n", live);
1564 oldest_gen->steps[0].mark = 0;
1565 oldest_gen->steps[0].compact = 0;
1566 // debugBelch("compaction: off\n", live);
1569 if (RtsFlags.GcFlags.sweep) {
1570 oldest_gen->steps[0].mark = 1;
1573 // if we're going to go over the maximum heap size, reduce the
1574 // size of the generations accordingly. The calculation is
1575 // different if compaction is turned on, because we don't need
1576 // to double the space required to collect the old generation.
1579 // this test is necessary to ensure that the calculations
1580 // below don't have any negative results - we're working
1581 // with unsigned values here.
1582 if (max < min_alloc) {
1586 if (oldest_gen->steps[0].compact) {
1587 if ( (size + (size - 1) * (gens - 2) * 2) + min_alloc > max ) {
1588 size = (max - min_alloc) / ((gens - 1) * 2 - 1);
1591 if ( (size * (gens - 1) * 2) + min_alloc > max ) {
1592 size = (max - min_alloc) / ((gens - 1) * 2);
1602 debugBelch("live: %d, min_alloc: %d, size : %d, max = %d\n", live,
1603 min_alloc, size, max);
1606 for (g = 0; g < gens; g++) {
1607 generations[g].max_blocks = size;
1612 /* -----------------------------------------------------------------------------
1613 Calculate the new size of the nursery, and resize it.
1614 -------------------------------------------------------------------------- */
1617 resize_nursery (void)
1619 if (RtsFlags.GcFlags.generations == 1)
1620 { // Two-space collector:
1623 /* set up a new nursery. Allocate a nursery size based on a
1624 * function of the amount of live data (by default a factor of 2)
1625 * Use the blocks from the old nursery if possible, freeing up any
1628 * If we get near the maximum heap size, then adjust our nursery
1629 * size accordingly. If the nursery is the same size as the live
1630 * data (L), then we need 3L bytes. We can reduce the size of the
1631 * nursery to bring the required memory down near 2L bytes.
1633 * A normal 2-space collector would need 4L bytes to give the same
1634 * performance we get from 3L bytes, reducing to the same
1635 * performance at 2L bytes.
1637 blocks = g0s0->n_blocks;
1639 if ( RtsFlags.GcFlags.maxHeapSize != 0 &&
1640 blocks * RtsFlags.GcFlags.oldGenFactor * 2 >
1641 RtsFlags.GcFlags.maxHeapSize )
1643 long adjusted_blocks; // signed on purpose
1646 adjusted_blocks = (RtsFlags.GcFlags.maxHeapSize - 2 * blocks);
1648 debugTrace(DEBUG_gc, "near maximum heap size of 0x%x blocks, blocks = %d, adjusted to %ld",
1649 RtsFlags.GcFlags.maxHeapSize, blocks, adjusted_blocks);
1651 pc_free = adjusted_blocks * 100 / RtsFlags.GcFlags.maxHeapSize;
1652 if (pc_free < RtsFlags.GcFlags.pcFreeHeap) /* might even * be < 0 */
1656 blocks = adjusted_blocks;
1660 blocks *= RtsFlags.GcFlags.oldGenFactor;
1661 if (blocks < RtsFlags.GcFlags.minAllocAreaSize)
1663 blocks = RtsFlags.GcFlags.minAllocAreaSize;
1666 resizeNurseries(blocks);
1668 else // Generational collector
1671 * If the user has given us a suggested heap size, adjust our
1672 * allocation area to make best use of the memory available.
1674 if (RtsFlags.GcFlags.heapSizeSuggestion)
1677 nat needed = calcNeeded(); // approx blocks needed at next GC
1679 /* Guess how much will be live in generation 0 step 0 next time.
1680 * A good approximation is obtained by finding the
1681 * percentage of g0s0 that was live at the last minor GC.
1683 * We have an accurate figure for the amount of copied data in
1684 * 'copied', but we must convert this to a number of blocks, with
1685 * a small adjustment for estimated slop at the end of a block
1690 g0s0_pcnt_kept = ((copied / (BLOCK_SIZE_W - 10)) * 100)
1691 / countNurseryBlocks();
1694 /* Estimate a size for the allocation area based on the
1695 * information available. We might end up going slightly under
1696 * or over the suggested heap size, but we should be pretty
1699 * Formula: suggested - needed
1700 * ----------------------------
1701 * 1 + g0s0_pcnt_kept/100
1703 * where 'needed' is the amount of memory needed at the next
1704 * collection for collecting all steps except g0s0.
1707 (((long)RtsFlags.GcFlags.heapSizeSuggestion - (long)needed) * 100) /
1708 (100 + (long)g0s0_pcnt_kept);
1710 if (blocks < (long)RtsFlags.GcFlags.minAllocAreaSize) {
1711 blocks = RtsFlags.GcFlags.minAllocAreaSize;
1714 resizeNurseries((nat)blocks);
1718 // we might have added extra large blocks to the nursery, so
1719 // resize back to minAllocAreaSize again.
1720 resizeNurseriesFixed(RtsFlags.GcFlags.minAllocAreaSize);
1725 /* -----------------------------------------------------------------------------
1726 Sanity code for CAF garbage collection.
1728 With DEBUG turned on, we manage a CAF list in addition to the SRT
1729 mechanism. After GC, we run down the CAF list and blackhole any
1730 CAFs which have been garbage collected. This means we get an error
1731 whenever the program tries to enter a garbage collected CAF.
1733 Any garbage collected CAFs are taken off the CAF list at the same
1735 -------------------------------------------------------------------------- */
1737 #if 0 && defined(DEBUG)
1744 const StgInfoTable *info;
1755 ASSERT(info->type == IND_STATIC);
1757 if (STATIC_LINK(info,p) == NULL) {
1758 debugTrace(DEBUG_gccafs, "CAF gc'd at 0x%04lx", (long)p);
1760 SET_INFO(p,&stg_BLACKHOLE_info);
1761 p = STATIC_LINK2(info,p);
1765 pp = &STATIC_LINK2(info,p);
1772 debugTrace(DEBUG_gccafs, "%d CAFs live", i);