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;
119 // gc_thread *gct = NULL; // this thread's gct TODO: make thread-local
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
129 SpinLock recordMutableGen_sync;
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 alloc_gc_threads (void);
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 update_task_list (void);
144 static void resize_generations (void);
145 static void resize_nursery (void);
146 static void start_gc_threads (void);
147 static void scavenge_until_all_done (void);
148 static nat inc_running (void);
149 static nat dec_running (void);
150 static void wakeup_gc_threads (nat n_threads);
151 static void shutdown_gc_threads (nat n_threads);
153 #if 0 && defined(DEBUG)
154 static void gcCAFs (void);
157 /* -----------------------------------------------------------------------------
158 The mark bitmap & stack.
159 -------------------------------------------------------------------------- */
161 #define MARK_STACK_BLOCKS 4
163 bdescr *mark_stack_bdescr;
168 // Flag and pointers used for falling back to a linear scan when the
169 // mark stack overflows.
170 rtsBool mark_stack_overflowed;
171 bdescr *oldgen_scan_bd;
174 /* -----------------------------------------------------------------------------
175 GarbageCollect: the main entry point to the garbage collector.
177 Locks held: all capabilities are held throughout GarbageCollect().
178 -------------------------------------------------------------------------- */
181 GarbageCollect ( rtsBool force_major_gc )
185 lnat live, allocated, max_copied, avg_copied, slop;
186 lnat oldgen_saved_blocks = 0;
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
221 // attribute any costs to CCS_GC
227 /* Approximate how much we allocated.
228 * Todo: only when generating stats?
230 allocated = calcAllocated();
232 /* Figure out which generation to collect
234 n = initialise_N(force_major_gc);
236 /* Allocate + initialise the gc_thread structures.
240 /* Start threads, so they can be spinning up while we finish initialisation.
244 /* How many threads will be participating in this GC?
245 * We don't try to parallelise minor GC.
247 #if defined(THREADED_RTS)
248 if (n < (4*1024*1024 / BLOCK_SIZE)) {
251 n_gc_threads = RtsFlags.ParFlags.gcThreads;
256 trace(TRACE_gc|DEBUG_gc, "GC (gen %d): %d KB to collect, %ld MB in use, using %d thread(s)",
257 N, n * (BLOCK_SIZE / 1024), mblocks_allocated, n_gc_threads);
259 #ifdef RTS_GTK_FRONTPANEL
260 if (RtsFlags.GcFlags.frontpanel) {
261 updateFrontPanelBeforeGC(N);
266 // check for memory leaks if DEBUG is on
267 memInventory(traceClass(DEBUG_gc));
270 // check stack sanity *before* GC (ToDo: check all threads)
271 IF_DEBUG(sanity, checkFreeListSanity());
273 // Initialise all our gc_thread structures
274 for (t = 0; t < n_gc_threads; t++) {
275 init_gc_thread(gc_threads[t]);
278 // Initialise all the generations/steps that we're collecting.
279 for (g = 0; g <= N; g++) {
280 init_collected_gen(g,n_gc_threads);
283 // Initialise all the generations/steps that we're *not* collecting.
284 for (g = N+1; g < RtsFlags.GcFlags.generations; g++) {
285 init_uncollected_gen(g,n_gc_threads);
288 /* Allocate a mark stack if we're doing a major collection.
291 mark_stack_bdescr = allocGroup(MARK_STACK_BLOCKS);
292 mark_stack = (StgPtr *)mark_stack_bdescr->start;
293 mark_sp = mark_stack;
294 mark_splim = mark_stack + (MARK_STACK_BLOCKS * BLOCK_SIZE_W);
296 mark_stack_bdescr = NULL;
299 // this is the main thread
302 /* -----------------------------------------------------------------------
303 * follow all the roots that we know about:
304 * - mutable lists from each generation > N
305 * we want to *scavenge* these roots, not evacuate them: they're not
306 * going to move in this GC.
307 * Also do them in reverse generation order, for the usual reason:
308 * namely to reduce the likelihood of spurious old->new pointers.
310 for (g = RtsFlags.GcFlags.generations-1; g > N; g--) {
311 generations[g].saved_mut_list = generations[g].mut_list;
312 generations[g].mut_list = allocBlock();
313 // mut_list always has at least one block.
316 // the main thread is running: this prevents any other threads from
317 // exiting prematurely, so we can start them now.
318 // NB. do this after the mutable lists have been saved above, otherwise
319 // the other GC threads will be writing into the old mutable lists.
321 wakeup_gc_threads(n_gc_threads);
323 for (g = RtsFlags.GcFlags.generations-1; g > N; g--) {
324 scavenge_mutable_list(&generations[g]);
327 // follow roots from the CAF list (used by GHCi)
329 markCAFs(mark_root, gct);
331 // follow all the roots that the application knows about.
333 markSomeCapabilities(mark_root, gct, gct->thread_index, n_gc_threads);
335 #if defined(RTS_USER_SIGNALS)
336 // mark the signal handlers (signals should be already blocked)
337 markSignalHandlers(mark_root, gct);
340 // Mark the weak pointer list, and prepare to detect dead weak pointers.
344 // Mark the stable pointer table.
345 markStablePtrTable(mark_root, gct);
347 /* -------------------------------------------------------------------------
348 * Repeatedly scavenge all the areas we know about until there's no
349 * more scavenging to be done.
353 scavenge_until_all_done();
354 // The other threads are now stopped. We might recurse back to
355 // here, but from now on this is the only thread.
357 // if any blackholes are alive, make the threads that wait on
359 if (traverseBlackholeQueue()) {
364 // must be last... invariant is that everything is fully
365 // scavenged at this point.
366 if (traverseWeakPtrList()) { // returns rtsTrue if evaced something
371 // If we get to here, there's really nothing left to do.
375 shutdown_gc_threads(n_gc_threads);
377 // Update pointers from the Task list
380 // Now see which stable names are still alive.
384 // We call processHeapClosureForDead() on every closure destroyed during
385 // the current garbage collection, so we invoke LdvCensusForDead().
386 if (RtsFlags.ProfFlags.doHeapProfile == HEAP_BY_LDV
387 || RtsFlags.ProfFlags.bioSelector != NULL)
391 // NO MORE EVACUATION AFTER THIS POINT!
392 // Finally: compaction of the oldest generation.
393 if (major_gc && oldest_gen->steps[0].is_compacted) {
394 // save number of blocks for stats
395 oldgen_saved_blocks = oldest_gen->steps[0].n_old_blocks;
396 compact(gct->scavenged_static_objects);
399 IF_DEBUG(sanity, checkGlobalTSOList(rtsFalse));
401 // Two-space collector: free the old to-space.
402 // g0s0->old_blocks is the old nursery
403 // g0s0->blocks is to-space from the previous GC
404 if (RtsFlags.GcFlags.generations == 1) {
405 if (g0s0->blocks != NULL) {
406 freeChain(g0s0->blocks);
411 // For each workspace, in each thread, move the copied blocks to the step
417 for (t = 0; t < n_gc_threads; t++) {
421 if (RtsFlags.GcFlags.generations == 1) {
426 for (; s < total_steps; s++) {
429 // Push the final block
431 push_scanned_block(ws->todo_bd, ws);
434 ASSERT(gct->scan_bd == NULL);
435 ASSERT(countBlocks(ws->scavd_list) == ws->n_scavd_blocks);
438 for (bd = ws->scavd_list; bd != NULL; bd = bd->link) {
439 ws->step->n_words += bd->free - bd->start;
443 prev->link = ws->step->blocks;
444 ws->step->blocks = ws->scavd_list;
446 ws->step->n_blocks += ws->n_scavd_blocks;
450 // Add all the partial blocks *after* we've added all the full
451 // blocks. This is so that we can grab the partial blocks back
452 // again and try to fill them up in the next GC.
453 for (t = 0; t < n_gc_threads; t++) {
457 if (RtsFlags.GcFlags.generations == 1) {
462 for (; s < total_steps; s++) {
466 for (bd = ws->part_list; bd != NULL; bd = next) {
468 if (bd->free == bd->start) {
470 ws->part_list = next;
477 ws->step->n_words += bd->free - bd->start;
482 prev->link = ws->step->blocks;
483 ws->step->blocks = ws->part_list;
485 ws->step->n_blocks += ws->n_part_blocks;
487 ASSERT(countBlocks(ws->step->blocks) == ws->step->n_blocks);
488 ASSERT(countOccupied(ws->step->blocks) == ws->step->n_words);
493 /* run through all the generations/steps and tidy up
500 for (i=0; i < n_gc_threads; i++) {
501 if (n_gc_threads > 1) {
502 trace(TRACE_gc,"thread %d:", i);
503 trace(TRACE_gc," copied %ld", gc_threads[i]->copied * sizeof(W_));
504 trace(TRACE_gc," scanned %ld", gc_threads[i]->scanned * sizeof(W_));
505 trace(TRACE_gc," any_work %ld", gc_threads[i]->any_work);
506 trace(TRACE_gc," no_work %ld", gc_threads[i]->no_work);
507 trace(TRACE_gc," scav_find_work %ld", gc_threads[i]->scav_find_work);
509 copied += gc_threads[i]->copied;
510 max_copied = stg_max(gc_threads[i]->copied, max_copied);
512 if (n_gc_threads == 1) {
520 for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
523 generations[g].collections++; // for stats
524 if (n_gc_threads > 1) generations[g].par_collections++;
527 // Count the mutable list as bytes "copied" for the purposes of
528 // stats. Every mutable list is copied during every GC.
530 nat mut_list_size = 0;
531 for (bd = generations[g].mut_list; bd != NULL; bd = bd->link) {
532 mut_list_size += bd->free - bd->start;
534 copied += mut_list_size;
537 "mut_list_size: %lu (%d vars, %d arrays, %d MVARs, %d others)",
538 (unsigned long)(mut_list_size * sizeof(W_)),
539 mutlist_MUTVARS, mutlist_MUTARRS, mutlist_MVARS, mutlist_OTHERS);
542 for (s = 0; s < generations[g].n_steps; s++) {
544 stp = &generations[g].steps[s];
546 // for generations we collected...
549 /* free old memory and shift to-space into from-space for all
550 * the collected steps (except the allocation area). These
551 * freed blocks will probaby be quickly recycled.
553 if (!(g == 0 && s == 0 && RtsFlags.GcFlags.generations > 1)) {
554 if (stp->is_compacted)
556 // for a compacted step, just shift the new to-space
557 // onto the front of the now-compacted existing blocks.
558 for (bd = stp->blocks; bd != NULL; bd = bd->link) {
559 stp->n_words += bd->free - bd->start;
561 // tack the new blocks on the end of the existing blocks
562 if (stp->old_blocks != NULL) {
563 for (bd = stp->old_blocks; bd != NULL; bd = next) {
564 // NB. this step might not be compacted next
565 // time, so reset the BF_COMPACTED flags.
566 // They are set before GC if we're going to
567 // compact. (search for BF_COMPACTED above).
568 bd->flags &= ~BF_COMPACTED;
571 bd->link = stp->blocks;
574 stp->blocks = stp->old_blocks;
576 // add the new blocks to the block tally
577 stp->n_blocks += stp->n_old_blocks;
578 ASSERT(countBlocks(stp->blocks) == stp->n_blocks);
579 ASSERT(countOccupied(stp->blocks) == stp->n_words);
583 freeChain(stp->old_blocks);
585 stp->old_blocks = NULL;
586 stp->n_old_blocks = 0;
589 /* LARGE OBJECTS. The current live large objects are chained on
590 * scavenged_large, having been moved during garbage
591 * collection from large_objects. Any objects left on
592 * large_objects list are therefore dead, so we free them here.
594 for (bd = stp->large_objects; bd != NULL; bd = next) {
600 stp->large_objects = stp->scavenged_large_objects;
601 stp->n_large_blocks = stp->n_scavenged_large_blocks;
604 else // for older generations...
606 /* For older generations, we need to append the
607 * scavenged_large_object list (i.e. large objects that have been
608 * promoted during this GC) to the large_object list for that step.
610 for (bd = stp->scavenged_large_objects; bd; bd = next) {
612 dbl_link_onto(bd, &stp->large_objects);
615 // add the new blocks we promoted during this GC
616 stp->n_large_blocks += stp->n_scavenged_large_blocks;
621 // update the max size of older generations after a major GC
622 resize_generations();
624 // Calculate the amount of live data for stats.
625 live = calcLiveWords();
627 // Free the small objects allocated via allocate(), since this will
628 // all have been copied into G0S1 now.
629 if (RtsFlags.GcFlags.generations > 1) {
630 if (g0s0->blocks != NULL) {
631 freeChain(g0s0->blocks);
638 alloc_blocks_lim = RtsFlags.GcFlags.minAllocAreaSize;
640 // Start a new pinned_object_block
641 pinned_object_block = NULL;
643 // Free the mark stack.
644 if (mark_stack_bdescr != NULL) {
645 freeGroup(mark_stack_bdescr);
649 for (g = 0; g <= N; g++) {
650 for (s = 0; s < generations[g].n_steps; s++) {
651 stp = &generations[g].steps[s];
652 if (stp->bitmap != NULL) {
653 freeGroup(stp->bitmap);
661 // mark the garbage collected CAFs as dead
662 #if 0 && defined(DEBUG) // doesn't work at the moment
663 if (major_gc) { gcCAFs(); }
667 // resetStaticObjectForRetainerProfiling() must be called before
669 if (n_gc_threads > 1) {
670 barf("profiling is currently broken with multi-threaded GC");
671 // ToDo: fix the gct->scavenged_static_objects below
673 resetStaticObjectForRetainerProfiling(gct->scavenged_static_objects);
676 // zero the scavenged static object list
679 for (i = 0; i < n_gc_threads; i++) {
680 zero_static_object_list(gc_threads[i]->scavenged_static_objects);
687 // start any pending finalizers
689 scheduleFinalizers(last_free_capability, old_weak_ptr_list);
692 // send exceptions to any threads which were about to die
694 resurrectThreads(resurrected_threads);
697 // Update the stable pointer hash table.
698 updateStablePtrTable(major_gc);
700 // check sanity after GC
701 IF_DEBUG(sanity, checkSanity());
703 // extra GC trace info
704 if (traceClass(TRACE_gc|DEBUG_gc)) statDescribeGens();
707 // symbol-table based profiling
708 /* heapCensus(to_blocks); */ /* ToDo */
711 // restore enclosing cost centre
717 // check for memory leaks if DEBUG is on
718 memInventory(traceClass(DEBUG_gc));
721 #ifdef RTS_GTK_FRONTPANEL
722 if (RtsFlags.GcFlags.frontpanel) {
723 updateFrontPanelAfterGC( N, live );
727 // ok, GC over: tell the stats department what happened.
728 slop = calcLiveBlocks() * BLOCK_SIZE_W - live;
729 stat_endGC(allocated, live, copied, N, max_copied, avg_copied, slop);
731 #if defined(RTS_USER_SIGNALS)
732 if (RtsFlags.MiscFlags.install_signal_handlers) {
733 // unblock signals again
734 unblockUserSignals();
743 /* -----------------------------------------------------------------------------
744 Figure out which generation to collect, initialise N and major_gc.
746 Also returns the total number of blocks in generations that will be
748 -------------------------------------------------------------------------- */
751 initialise_N (rtsBool force_major_gc)
754 nat s, blocks, blocks_total;
759 if (force_major_gc) {
760 N = RtsFlags.GcFlags.generations - 1;
765 for (g = RtsFlags.GcFlags.generations - 1; g >= 0; g--) {
767 for (s = 0; s < generations[g].n_steps; s++) {
768 blocks += generations[g].steps[s].n_words / BLOCK_SIZE_W;
769 blocks += generations[g].steps[s].n_large_blocks;
771 if (blocks >= generations[g].max_blocks) {
775 blocks_total += blocks;
779 blocks_total += countNurseryBlocks();
781 major_gc = (N == RtsFlags.GcFlags.generations-1);
785 /* -----------------------------------------------------------------------------
786 Initialise the gc_thread structures.
787 -------------------------------------------------------------------------- */
790 alloc_gc_thread (int n)
796 t = stgMallocBytes(sizeof(gc_thread) + total_steps * sizeof(step_workspace),
801 initCondition(&t->wake_cond);
802 initMutex(&t->wake_mutex);
803 t->wakeup = rtsTrue; // starts true, so we can wait for the
804 // thread to start up, see wakeup_gc_threads
809 t->free_blocks = NULL;
818 for (s = 0; s < total_steps; s++)
821 ws->step = &all_steps[s];
822 ASSERT(s == ws->step->abs_no);
826 ws->buffer_todo_bd = NULL;
828 ws->part_list = NULL;
829 ws->n_part_blocks = 0;
831 ws->scavd_list = NULL;
832 ws->n_scavd_blocks = 0;
840 alloc_gc_threads (void)
842 if (gc_threads == NULL) {
843 #if defined(THREADED_RTS)
845 gc_threads = stgMallocBytes (RtsFlags.ParFlags.gcThreads *
849 for (i = 0; i < RtsFlags.ParFlags.gcThreads; i++) {
850 gc_threads[i] = alloc_gc_thread(i);
853 gc_threads = stgMallocBytes (sizeof(gc_thread*),
856 gc_threads[0] = alloc_gc_thread(0);
861 /* ----------------------------------------------------------------------------
863 ------------------------------------------------------------------------- */
865 static nat gc_running_threads;
867 #if defined(THREADED_RTS)
868 static Mutex gc_running_mutex;
875 ACQUIRE_LOCK(&gc_running_mutex);
876 n_running = ++gc_running_threads;
877 RELEASE_LOCK(&gc_running_mutex);
878 ASSERT(n_running <= n_gc_threads);
886 ACQUIRE_LOCK(&gc_running_mutex);
887 ASSERT(n_gc_threads != 0);
888 n_running = --gc_running_threads;
889 RELEASE_LOCK(&gc_running_mutex);
894 scavenge_until_all_done (void)
898 debugTrace(DEBUG_gc, "GC thread %d working", gct->thread_index);
902 // scavenge_loop() only exits when there's no work to do
905 debugTrace(DEBUG_gc, "GC thread %d idle (%d still running)",
906 gct->thread_index, r);
908 while (gc_running_threads != 0) {
914 // any_work() does not remove the work from the queue, it
915 // just checks for the presence of work. If we find any,
916 // then we increment gc_running_threads and go back to
917 // scavenge_loop() to perform any pending work.
920 // All threads are now stopped
921 debugTrace(DEBUG_gc, "GC thread %d finished.", gct->thread_index);
924 #if defined(THREADED_RTS)
926 // gc_thread_work(): Scavenge until there's no work left to do and all
927 // the running threads are idle.
930 gc_thread_work (void)
932 // gc_running_threads has already been incremented for us; this is
933 // a worker thread and the main thread bumped gc_running_threads
934 // before waking us up.
936 // Every thread evacuates some roots.
938 markSomeCapabilities(mark_root, gct, gct->thread_index, n_gc_threads);
940 scavenge_until_all_done();
945 gc_thread_mainloop (void)
949 // Wait until we're told to wake up
950 ACQUIRE_LOCK(&gct->wake_mutex);
951 gct->wakeup = rtsFalse;
952 while (!gct->wakeup) {
953 debugTrace(DEBUG_gc, "GC thread %d standing by...",
955 waitCondition(&gct->wake_cond, &gct->wake_mutex);
957 RELEASE_LOCK(&gct->wake_mutex);
958 if (gct->exit) break;
961 // start performance counters in this thread...
962 if (gct->papi_events == -1) {
963 papi_init_eventset(&gct->papi_events);
965 papi_thread_start_gc1_count(gct->papi_events);
971 // count events in this thread towards the GC totals
972 papi_thread_stop_gc1_count(gct->papi_events);
978 #if defined(THREADED_RTS)
980 gc_thread_entry (gc_thread *my_gct)
983 debugTrace(DEBUG_gc, "GC thread %d starting...", gct->thread_index);
984 gct->id = osThreadId();
985 gc_thread_mainloop();
990 start_gc_threads (void)
992 #if defined(THREADED_RTS)
995 static rtsBool done = rtsFalse;
997 gc_running_threads = 0;
998 initMutex(&gc_running_mutex);
1001 // Start from 1: the main thread is 0
1002 for (i = 1; i < RtsFlags.ParFlags.gcThreads; i++) {
1003 createOSThread(&id, (OSThreadProc*)&gc_thread_entry,
1012 wakeup_gc_threads (nat n_threads USED_IF_THREADS)
1014 #if defined(THREADED_RTS)
1016 for (i=1; i < n_threads; i++) {
1018 debugTrace(DEBUG_gc, "waking up gc thread %d", i);
1020 ACQUIRE_LOCK(&gc_threads[i]->wake_mutex);
1021 if (gc_threads[i]->wakeup) {
1022 RELEASE_LOCK(&gc_threads[i]->wake_mutex);
1028 gc_threads[i]->wakeup = rtsTrue;
1029 signalCondition(&gc_threads[i]->wake_cond);
1030 RELEASE_LOCK(&gc_threads[i]->wake_mutex);
1035 // After GC is complete, we must wait for all GC threads to enter the
1036 // standby state, otherwise they may still be executing inside
1037 // any_work(), and may even remain awake until the next GC starts.
1039 shutdown_gc_threads (nat n_threads USED_IF_THREADS)
1041 #if defined(THREADED_RTS)
1044 for (i=1; i < n_threads; i++) {
1046 ACQUIRE_LOCK(&gc_threads[i]->wake_mutex);
1047 wakeup = gc_threads[i]->wakeup;
1048 // wakeup is false while the thread is waiting
1049 RELEASE_LOCK(&gc_threads[i]->wake_mutex);
1055 /* ----------------------------------------------------------------------------
1056 Initialise a generation that is to be collected
1057 ------------------------------------------------------------------------- */
1060 init_collected_gen (nat g, nat n_threads)
1067 // Throw away the current mutable list. Invariant: the mutable
1068 // list always has at least one block; this means we can avoid a
1069 // check for NULL in recordMutable().
1071 freeChain(generations[g].mut_list);
1072 generations[g].mut_list = allocBlock();
1073 for (i = 0; i < n_capabilities; i++) {
1074 freeChain(capabilities[i].mut_lists[g]);
1075 capabilities[i].mut_lists[g] = allocBlock();
1079 for (s = 0; s < generations[g].n_steps; s++) {
1081 stp = &generations[g].steps[s];
1082 ASSERT(stp->gen_no == g);
1084 // we'll construct a new list of threads in this step
1085 // during GC, throw away the current list.
1086 stp->old_threads = stp->threads;
1087 stp->threads = END_TSO_QUEUE;
1089 // generation 0, step 0 doesn't need to-space
1090 if (g == 0 && s == 0 && RtsFlags.GcFlags.generations > 1) {
1094 // deprecate the existing blocks
1095 stp->old_blocks = stp->blocks;
1096 stp->n_old_blocks = stp->n_blocks;
1101 // we don't have any to-be-scavenged blocks yet
1103 stp->todos_last = NULL;
1106 // initialise the large object queues.
1107 stp->scavenged_large_objects = NULL;
1108 stp->n_scavenged_large_blocks = 0;
1110 // mark the small objects as from-space
1111 for (bd = stp->old_blocks; bd; bd = bd->link) {
1112 bd->flags &= ~BF_EVACUATED;
1115 // mark the large objects as from-space
1116 for (bd = stp->large_objects; bd; bd = bd->link) {
1117 bd->flags &= ~BF_EVACUATED;
1120 // for a compacted step, we need to allocate the bitmap
1121 if (stp->is_compacted) {
1122 nat bitmap_size; // in bytes
1123 bdescr *bitmap_bdescr;
1126 bitmap_size = stp->n_old_blocks * BLOCK_SIZE / (sizeof(W_)*BITS_PER_BYTE);
1128 if (bitmap_size > 0) {
1129 bitmap_bdescr = allocGroup((lnat)BLOCK_ROUND_UP(bitmap_size)
1131 stp->bitmap = bitmap_bdescr;
1132 bitmap = bitmap_bdescr->start;
1134 debugTrace(DEBUG_gc, "bitmap_size: %d, bitmap: %p",
1135 bitmap_size, bitmap);
1137 // don't forget to fill it with zeros!
1138 memset(bitmap, 0, bitmap_size);
1140 // For each block in this step, point to its bitmap from the
1141 // block descriptor.
1142 for (bd=stp->old_blocks; bd != NULL; bd = bd->link) {
1143 bd->u.bitmap = bitmap;
1144 bitmap += BLOCK_SIZE_W / (sizeof(W_)*BITS_PER_BYTE);
1146 // Also at this point we set the BF_COMPACTED flag
1147 // for this block. The invariant is that
1148 // BF_COMPACTED is always unset, except during GC
1149 // when it is set on those blocks which will be
1151 bd->flags |= BF_COMPACTED;
1157 // For each GC thread, for each step, allocate a "todo" block to
1158 // store evacuated objects to be scavenged, and a block to store
1159 // evacuated objects that do not need to be scavenged.
1160 for (t = 0; t < n_threads; t++) {
1161 for (s = 0; s < generations[g].n_steps; s++) {
1163 // we don't copy objects into g0s0, unless -G0
1164 if (g==0 && s==0 && RtsFlags.GcFlags.generations > 1) continue;
1166 ws = &gc_threads[t]->steps[g * RtsFlags.GcFlags.steps + s];
1168 ws->todo_large_objects = NULL;
1170 ws->part_list = NULL;
1171 ws->n_part_blocks = 0;
1173 // allocate the first to-space block; extra blocks will be
1174 // chained on as necessary.
1176 ws->buffer_todo_bd = NULL;
1177 alloc_todo_block(ws,0);
1179 ws->scavd_list = NULL;
1180 ws->n_scavd_blocks = 0;
1186 /* ----------------------------------------------------------------------------
1187 Initialise a generation that is *not* to be collected
1188 ------------------------------------------------------------------------- */
1191 init_uncollected_gen (nat g, nat threads)
1198 for (s = 0; s < generations[g].n_steps; s++) {
1199 stp = &generations[g].steps[s];
1200 stp->scavenged_large_objects = NULL;
1201 stp->n_scavenged_large_blocks = 0;
1204 for (s = 0; s < generations[g].n_steps; s++) {
1206 stp = &generations[g].steps[s];
1208 for (t = 0; t < threads; t++) {
1209 ws = &gc_threads[t]->steps[g * RtsFlags.GcFlags.steps + s];
1211 ws->buffer_todo_bd = NULL;
1212 ws->todo_large_objects = NULL;
1214 ws->part_list = NULL;
1215 ws->n_part_blocks = 0;
1217 ws->scavd_list = NULL;
1218 ws->n_scavd_blocks = 0;
1220 // If the block at the head of the list in this generation
1221 // is less than 3/4 full, then use it as a todo block.
1222 if (stp->blocks && isPartiallyFull(stp->blocks))
1224 ws->todo_bd = stp->blocks;
1225 ws->todo_free = ws->todo_bd->free;
1226 ws->todo_lim = ws->todo_bd->start + BLOCK_SIZE_W;
1227 stp->blocks = stp->blocks->link;
1229 stp->n_words -= ws->todo_bd->free - ws->todo_bd->start;
1230 ws->todo_bd->link = NULL;
1231 // we must scan from the current end point.
1232 ws->todo_bd->u.scan = ws->todo_bd->free;
1237 alloc_todo_block(ws,0);
1241 // deal out any more partial blocks to the threads' part_lists
1243 while (stp->blocks && isPartiallyFull(stp->blocks))
1246 stp->blocks = bd->link;
1247 ws = &gc_threads[t]->steps[g * RtsFlags.GcFlags.steps + s];
1248 bd->link = ws->part_list;
1250 ws->n_part_blocks += 1;
1251 bd->u.scan = bd->free;
1253 stp->n_words -= bd->free - bd->start;
1255 if (t == n_gc_threads) t = 0;
1260 // Move the private mutable lists from each capability onto the
1261 // main mutable list for the generation.
1262 for (i = 0; i < n_capabilities; i++) {
1263 for (bd = capabilities[i].mut_lists[g];
1264 bd->link != NULL; bd = bd->link) {
1267 bd->link = generations[g].mut_list;
1268 generations[g].mut_list = capabilities[i].mut_lists[g];
1269 capabilities[i].mut_lists[g] = allocBlock();
1273 /* -----------------------------------------------------------------------------
1274 Initialise a gc_thread before GC
1275 -------------------------------------------------------------------------- */
1278 init_gc_thread (gc_thread *t)
1280 t->static_objects = END_OF_STATIC_LIST;
1281 t->scavenged_static_objects = END_OF_STATIC_LIST;
1284 t->failed_to_evac = rtsFalse;
1285 t->eager_promotion = rtsTrue;
1286 t->thunk_selector_depth = 0;
1291 t->scav_find_work = 0;
1294 /* -----------------------------------------------------------------------------
1295 Function we pass to evacuate roots.
1296 -------------------------------------------------------------------------- */
1299 mark_root(void *user, StgClosure **root)
1301 // we stole a register for gct, but this function is called from
1302 // *outside* the GC where the register variable is not in effect,
1303 // so we need to save and restore it here. NB. only call
1304 // mark_root() from the main GC thread, otherwise gct will be
1306 gc_thread *saved_gct;
1315 /* -----------------------------------------------------------------------------
1316 Initialising the static object & mutable lists
1317 -------------------------------------------------------------------------- */
1320 zero_static_object_list(StgClosure* first_static)
1324 const StgInfoTable *info;
1326 for (p = first_static; p != END_OF_STATIC_LIST; p = link) {
1328 link = *STATIC_LINK(info, p);
1329 *STATIC_LINK(info,p) = NULL;
1333 /* ----------------------------------------------------------------------------
1334 Update the pointers from the task list
1336 These are treated as weak pointers because we want to allow a main
1337 thread to get a BlockedOnDeadMVar exception in the same way as any
1338 other thread. Note that the threads should all have been retained
1339 by GC by virtue of being on the all_threads list, we're just
1340 updating pointers here.
1341 ------------------------------------------------------------------------- */
1344 update_task_list (void)
1348 for (task = all_tasks; task != NULL; task = task->all_link) {
1349 if (!task->stopped && task->tso) {
1350 ASSERT(task->tso->bound == task);
1351 tso = (StgTSO *) isAlive((StgClosure *)task->tso);
1353 barf("task %p: main thread %d has been GC'd",
1366 /* ----------------------------------------------------------------------------
1367 Reset the sizes of the older generations when we do a major
1370 CURRENT STRATEGY: make all generations except zero the same size.
1371 We have to stay within the maximum heap size, and leave a certain
1372 percentage of the maximum heap size available to allocate into.
1373 ------------------------------------------------------------------------- */
1376 resize_generations (void)
1380 if (major_gc && RtsFlags.GcFlags.generations > 1) {
1381 nat live, size, min_alloc;
1382 nat max = RtsFlags.GcFlags.maxHeapSize;
1383 nat gens = RtsFlags.GcFlags.generations;
1385 // live in the oldest generations
1386 live = (oldest_gen->steps[0].n_words + BLOCK_SIZE_W - 1) / BLOCK_SIZE_W+
1387 oldest_gen->steps[0].n_large_blocks;
1389 // default max size for all generations except zero
1390 size = stg_max(live * RtsFlags.GcFlags.oldGenFactor,
1391 RtsFlags.GcFlags.minOldGenSize);
1393 // minimum size for generation zero
1394 min_alloc = stg_max((RtsFlags.GcFlags.pcFreeHeap * max) / 200,
1395 RtsFlags.GcFlags.minAllocAreaSize);
1397 // Auto-enable compaction when the residency reaches a
1398 // certain percentage of the maximum heap size (default: 30%).
1399 if (RtsFlags.GcFlags.generations > 1 &&
1400 (RtsFlags.GcFlags.compact ||
1402 oldest_gen->steps[0].n_blocks >
1403 (RtsFlags.GcFlags.compactThreshold * max) / 100))) {
1404 oldest_gen->steps[0].is_compacted = 1;
1405 // debugBelch("compaction: on\n", live);
1407 oldest_gen->steps[0].is_compacted = 0;
1408 // debugBelch("compaction: off\n", live);
1411 // if we're going to go over the maximum heap size, reduce the
1412 // size of the generations accordingly. The calculation is
1413 // different if compaction is turned on, because we don't need
1414 // to double the space required to collect the old generation.
1417 // this test is necessary to ensure that the calculations
1418 // below don't have any negative results - we're working
1419 // with unsigned values here.
1420 if (max < min_alloc) {
1424 if (oldest_gen->steps[0].is_compacted) {
1425 if ( (size + (size - 1) * (gens - 2) * 2) + min_alloc > max ) {
1426 size = (max - min_alloc) / ((gens - 1) * 2 - 1);
1429 if ( (size * (gens - 1) * 2) + min_alloc > max ) {
1430 size = (max - min_alloc) / ((gens - 1) * 2);
1440 debugBelch("live: %d, min_alloc: %d, size : %d, max = %d\n", live,
1441 min_alloc, size, max);
1444 for (g = 0; g < gens; g++) {
1445 generations[g].max_blocks = size;
1450 /* -----------------------------------------------------------------------------
1451 Calculate the new size of the nursery, and resize it.
1452 -------------------------------------------------------------------------- */
1455 resize_nursery (void)
1457 if (RtsFlags.GcFlags.generations == 1)
1458 { // Two-space collector:
1461 /* set up a new nursery. Allocate a nursery size based on a
1462 * function of the amount of live data (by default a factor of 2)
1463 * Use the blocks from the old nursery if possible, freeing up any
1466 * If we get near the maximum heap size, then adjust our nursery
1467 * size accordingly. If the nursery is the same size as the live
1468 * data (L), then we need 3L bytes. We can reduce the size of the
1469 * nursery to bring the required memory down near 2L bytes.
1471 * A normal 2-space collector would need 4L bytes to give the same
1472 * performance we get from 3L bytes, reducing to the same
1473 * performance at 2L bytes.
1475 blocks = g0s0->n_blocks;
1477 if ( RtsFlags.GcFlags.maxHeapSize != 0 &&
1478 blocks * RtsFlags.GcFlags.oldGenFactor * 2 >
1479 RtsFlags.GcFlags.maxHeapSize )
1481 long adjusted_blocks; // signed on purpose
1484 adjusted_blocks = (RtsFlags.GcFlags.maxHeapSize - 2 * blocks);
1486 debugTrace(DEBUG_gc, "near maximum heap size of 0x%x blocks, blocks = %d, adjusted to %ld",
1487 RtsFlags.GcFlags.maxHeapSize, blocks, adjusted_blocks);
1489 pc_free = adjusted_blocks * 100 / RtsFlags.GcFlags.maxHeapSize;
1490 if (pc_free < RtsFlags.GcFlags.pcFreeHeap) /* might even * be < 0 */
1494 blocks = adjusted_blocks;
1498 blocks *= RtsFlags.GcFlags.oldGenFactor;
1499 if (blocks < RtsFlags.GcFlags.minAllocAreaSize)
1501 blocks = RtsFlags.GcFlags.minAllocAreaSize;
1504 resizeNurseries(blocks);
1506 else // Generational collector
1509 * If the user has given us a suggested heap size, adjust our
1510 * allocation area to make best use of the memory available.
1512 if (RtsFlags.GcFlags.heapSizeSuggestion)
1515 nat needed = calcNeeded(); // approx blocks needed at next GC
1517 /* Guess how much will be live in generation 0 step 0 next time.
1518 * A good approximation is obtained by finding the
1519 * percentage of g0s0 that was live at the last minor GC.
1521 * We have an accurate figure for the amount of copied data in
1522 * 'copied', but we must convert this to a number of blocks, with
1523 * a small adjustment for estimated slop at the end of a block
1528 g0s0_pcnt_kept = ((copied / (BLOCK_SIZE_W - 10)) * 100)
1529 / countNurseryBlocks();
1532 /* Estimate a size for the allocation area based on the
1533 * information available. We might end up going slightly under
1534 * or over the suggested heap size, but we should be pretty
1537 * Formula: suggested - needed
1538 * ----------------------------
1539 * 1 + g0s0_pcnt_kept/100
1541 * where 'needed' is the amount of memory needed at the next
1542 * collection for collecting all steps except g0s0.
1545 (((long)RtsFlags.GcFlags.heapSizeSuggestion - (long)needed) * 100) /
1546 (100 + (long)g0s0_pcnt_kept);
1548 if (blocks < (long)RtsFlags.GcFlags.minAllocAreaSize) {
1549 blocks = RtsFlags.GcFlags.minAllocAreaSize;
1552 resizeNurseries((nat)blocks);
1556 // we might have added extra large blocks to the nursery, so
1557 // resize back to minAllocAreaSize again.
1558 resizeNurseriesFixed(RtsFlags.GcFlags.minAllocAreaSize);
1563 /* -----------------------------------------------------------------------------
1564 Sanity code for CAF garbage collection.
1566 With DEBUG turned on, we manage a CAF list in addition to the SRT
1567 mechanism. After GC, we run down the CAF list and blackhole any
1568 CAFs which have been garbage collected. This means we get an error
1569 whenever the program tries to enter a garbage collected CAF.
1571 Any garbage collected CAFs are taken off the CAF list at the same
1573 -------------------------------------------------------------------------- */
1575 #if 0 && defined(DEBUG)
1582 const StgInfoTable *info;
1593 ASSERT(info->type == IND_STATIC);
1595 if (STATIC_LINK(info,p) == NULL) {
1596 debugTrace(DEBUG_gccafs, "CAF gc'd at 0x%04lx", (long)p);
1598 SET_INFO(p,&stg_BLACKHOLE_info);
1599 p = STATIC_LINK2(info,p);
1603 pp = &STATIC_LINK2(info,p);
1610 debugTrace(DEBUG_gccafs, "%d CAFs live", i);