1 /* -----------------------------------------------------------------------------
3 * (c) The GHC Team 1998-2006
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 gc_thread_work (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 // tell the stats department that we've started a GC
209 // tell the STM to discard any cached closures it's hoping to re-use
218 // attribute any costs to CCS_GC
224 /* Approximate how much we allocated.
225 * Todo: only when generating stats?
227 allocated = calcAllocated();
229 /* Figure out which generation to collect
231 n = initialise_N(force_major_gc);
233 /* Allocate + initialise the gc_thread structures.
237 /* Start threads, so they can be spinning up while we finish initialisation.
241 /* How many threads will be participating in this GC?
242 * We don't try to parallelise minor GC.
244 #if defined(THREADED_RTS)
245 if (n < (4*1024*1024 / BLOCK_SIZE)) {
248 n_gc_threads = RtsFlags.ParFlags.gcThreads;
253 trace(TRACE_gc|DEBUG_gc, "GC (gen %d): %dKB to collect, using %d thread(s)",
254 N, n * (BLOCK_SIZE / 1024), n_gc_threads);
256 #ifdef RTS_GTK_FRONTPANEL
257 if (RtsFlags.GcFlags.frontpanel) {
258 updateFrontPanelBeforeGC(N);
263 // check for memory leaks if DEBUG is on
264 memInventory(traceClass(DEBUG_gc));
267 // check stack sanity *before* GC (ToDo: check all threads)
268 IF_DEBUG(sanity, checkFreeListSanity());
270 // Initialise all our gc_thread structures
271 for (t = 0; t < n_gc_threads; t++) {
272 init_gc_thread(gc_threads[t]);
275 // Initialise all the generations/steps that we're collecting.
276 for (g = 0; g <= N; g++) {
277 init_collected_gen(g,n_gc_threads);
280 // Initialise all the generations/steps that we're *not* collecting.
281 for (g = N+1; g < RtsFlags.GcFlags.generations; g++) {
282 init_uncollected_gen(g,n_gc_threads);
285 /* Allocate a mark stack if we're doing a major collection.
288 mark_stack_bdescr = allocGroup(MARK_STACK_BLOCKS);
289 mark_stack = (StgPtr *)mark_stack_bdescr->start;
290 mark_sp = mark_stack;
291 mark_splim = mark_stack + (MARK_STACK_BLOCKS * BLOCK_SIZE_W);
293 mark_stack_bdescr = NULL;
296 // this is the main thread
299 /* -----------------------------------------------------------------------
300 * follow all the roots that we know about:
301 * - mutable lists from each generation > N
302 * we want to *scavenge* these roots, not evacuate them: they're not
303 * going to move in this GC.
304 * Also do them in reverse generation order, for the usual reason:
305 * namely to reduce the likelihood of spurious old->new pointers.
307 for (g = RtsFlags.GcFlags.generations-1; g > N; g--) {
308 generations[g].saved_mut_list = generations[g].mut_list;
309 generations[g].mut_list = allocBlock();
310 // mut_list always has at least one block.
313 // the main thread is running: this prevents any other threads from
314 // exiting prematurely, so we can start them now.
315 // NB. do this after the mutable lists have been saved above, otherwise
316 // the other GC threads will be writing into the old mutable lists.
318 wakeup_gc_threads(n_gc_threads);
320 for (g = RtsFlags.GcFlags.generations-1; g > N; g--) {
321 scavenge_mutable_list(&generations[g]);
324 // follow roots from the CAF list (used by GHCi)
326 markCAFs(mark_root, gct);
328 // follow all the roots that the application knows about.
330 markSomeCapabilities(mark_root, gct, gct->thread_index, n_gc_threads);
332 #if defined(RTS_USER_SIGNALS)
333 // mark the signal handlers (signals should be already blocked)
334 markSignalHandlers(mark_root, gct);
337 // Mark the weak pointer list, and prepare to detect dead weak pointers.
341 // Mark the stable pointer table.
342 markStablePtrTable(mark_root, gct);
344 /* -------------------------------------------------------------------------
345 * Repeatedly scavenge all the areas we know about until there's no
346 * more scavenging to be done.
351 // The other threads are now stopped. We might recurse back to
352 // here, but from now on this is the only thread.
354 // if any blackholes are alive, make the threads that wait on
356 if (traverseBlackholeQueue()) {
361 // must be last... invariant is that everything is fully
362 // scavenged at this point.
363 if (traverseWeakPtrList()) { // returns rtsTrue if evaced something
368 // If we get to here, there's really nothing left to do.
372 shutdown_gc_threads(n_gc_threads);
374 // Update pointers from the Task list
377 // Now see which stable names are still alive.
381 // We call processHeapClosureForDead() on every closure destroyed during
382 // the current garbage collection, so we invoke LdvCensusForDead().
383 if (RtsFlags.ProfFlags.doHeapProfile == HEAP_BY_LDV
384 || RtsFlags.ProfFlags.bioSelector != NULL)
388 // NO MORE EVACUATION AFTER THIS POINT!
389 // Finally: compaction of the oldest generation.
390 if (major_gc && oldest_gen->steps[0].is_compacted) {
391 // save number of blocks for stats
392 oldgen_saved_blocks = oldest_gen->steps[0].n_old_blocks;
393 compact(gct->scavenged_static_objects);
396 IF_DEBUG(sanity, checkGlobalTSOList(rtsFalse));
398 // Two-space collector: free the old to-space.
399 // g0s0->old_blocks is the old nursery
400 // g0s0->blocks is to-space from the previous GC
401 if (RtsFlags.GcFlags.generations == 1) {
402 if (g0s0->blocks != NULL) {
403 freeChain(g0s0->blocks);
408 // For each workspace, in each thread:
409 // * clear the BF_EVACUATED flag from each copied block
410 // * move the copied blocks to the step
416 for (t = 0; t < n_gc_threads; t++) {
420 for (s = 1; s < total_steps; s++) {
423 // Push the final block
425 push_scanned_block(ws->todo_bd, ws);
428 ASSERT(gct->scan_bd == NULL);
429 ASSERT(countBlocks(ws->scavd_list) == ws->n_scavd_blocks);
432 for (bd = ws->scavd_list; bd != NULL; bd = bd->link) {
433 bd->flags &= ~BF_EVACUATED; // now from-space
434 ws->step->n_words += bd->free - bd->start;
438 prev->link = ws->step->blocks;
439 ws->step->blocks = ws->scavd_list;
441 ws->step->n_blocks += ws->n_scavd_blocks;
444 for (bd = ws->part_list; bd != NULL; bd = next) {
446 if (bd->free == bd->start) {
448 ws->part_list = next;
455 bd->flags &= ~BF_EVACUATED; // now from-space
456 ws->step->n_words += bd->free - bd->start;
461 prev->link = ws->step->blocks;
462 ws->step->blocks = ws->part_list;
464 ws->step->n_blocks += ws->n_part_blocks;
466 ASSERT(countBlocks(ws->step->blocks) == ws->step->n_blocks);
467 ASSERT(countOccupied(ws->step->blocks) == ws->step->n_words);
472 // Two-space collector: swap the semi-spaces around.
473 // Currently: g0s0->old_blocks is the old nursery
474 // g0s0->blocks is to-space from this GC
475 // We want these the other way around.
476 if (RtsFlags.GcFlags.generations == 1) {
477 bdescr *nursery_blocks = g0s0->old_blocks;
478 nat n_nursery_blocks = g0s0->n_old_blocks;
479 g0s0->old_blocks = g0s0->blocks;
480 g0s0->n_old_blocks = g0s0->n_blocks;
481 g0s0->blocks = nursery_blocks;
482 g0s0->n_blocks = n_nursery_blocks;
485 /* run through all the generations/steps and tidy up
492 for (i=0; i < n_gc_threads; i++) {
493 if (n_gc_threads > 1) {
494 trace(TRACE_gc,"thread %d:", i);
495 trace(TRACE_gc," copied %ld", gc_threads[i]->copied * sizeof(W_));
496 trace(TRACE_gc," scanned %ld", gc_threads[i]->scanned * sizeof(W_));
497 trace(TRACE_gc," any_work %ld", gc_threads[i]->any_work);
498 trace(TRACE_gc," no_work %ld", gc_threads[i]->no_work);
499 trace(TRACE_gc," scav_find_work %ld", gc_threads[i]->scav_find_work);
501 copied += gc_threads[i]->copied;
502 max_copied = stg_max(gc_threads[i]->copied, max_copied);
504 if (n_gc_threads == 1) {
512 for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
515 generations[g].collections++; // for stats
516 if (n_gc_threads > 1) generations[g].par_collections++;
519 // Count the mutable list as bytes "copied" for the purposes of
520 // stats. Every mutable list is copied during every GC.
522 nat mut_list_size = 0;
523 for (bd = generations[g].mut_list; bd != NULL; bd = bd->link) {
524 mut_list_size += bd->free - bd->start;
526 copied += mut_list_size;
529 "mut_list_size: %lu (%d vars, %d arrays, %d MVARs, %d others)",
530 (unsigned long)(mut_list_size * sizeof(W_)),
531 mutlist_MUTVARS, mutlist_MUTARRS, mutlist_MVARS, mutlist_OTHERS);
534 for (s = 0; s < generations[g].n_steps; s++) {
536 stp = &generations[g].steps[s];
538 // for generations we collected...
541 /* free old memory and shift to-space into from-space for all
542 * the collected steps (except the allocation area). These
543 * freed blocks will probaby be quickly recycled.
545 if (!(g == 0 && s == 0 && RtsFlags.GcFlags.generations > 1)) {
546 if (stp->is_compacted)
548 // for a compacted step, just shift the new to-space
549 // onto the front of the now-compacted existing blocks.
550 for (bd = stp->blocks; bd != NULL; bd = bd->link) {
551 bd->flags &= ~BF_EVACUATED; // now from-space
552 stp->n_words += bd->free - bd->start;
554 // tack the new blocks on the end of the existing blocks
555 if (stp->old_blocks != NULL) {
556 for (bd = stp->old_blocks; bd != NULL; bd = next) {
557 // NB. this step might not be compacted next
558 // time, so reset the BF_COMPACTED flags.
559 // They are set before GC if we're going to
560 // compact. (search for BF_COMPACTED above).
561 bd->flags &= ~BF_COMPACTED;
564 bd->link = stp->blocks;
567 stp->blocks = stp->old_blocks;
569 // add the new blocks to the block tally
570 stp->n_blocks += stp->n_old_blocks;
571 ASSERT(countBlocks(stp->blocks) == stp->n_blocks);
572 ASSERT(countOccupied(stp->blocks) == stp->n_words);
576 freeChain(stp->old_blocks);
578 stp->old_blocks = NULL;
579 stp->n_old_blocks = 0;
582 /* LARGE OBJECTS. The current live large objects are chained on
583 * scavenged_large, having been moved during garbage
584 * collection from large_objects. Any objects left on
585 * large_objects list are therefore dead, so we free them here.
587 for (bd = stp->large_objects; bd != NULL; bd = next) {
593 // update the count of blocks used by large objects
594 for (bd = stp->scavenged_large_objects; bd != NULL; bd = bd->link) {
595 bd->flags &= ~BF_EVACUATED;
597 stp->large_objects = stp->scavenged_large_objects;
598 stp->n_large_blocks = stp->n_scavenged_large_blocks;
601 else // for older generations...
603 /* For older generations, we need to append the
604 * scavenged_large_object list (i.e. large objects that have been
605 * promoted during this GC) to the large_object list for that step.
607 for (bd = stp->scavenged_large_objects; bd; bd = next) {
609 bd->flags &= ~BF_EVACUATED;
610 dbl_link_onto(bd, &stp->large_objects);
613 // add the new blocks we promoted during this GC
614 stp->n_large_blocks += stp->n_scavenged_large_blocks;
619 // update the max size of older generations after a major GC
620 resize_generations();
622 // Calculate the amount of live data for stats.
623 live = calcLiveWords();
625 // Free the small objects allocated via allocate(), since this will
626 // all have been copied into G0S1 now.
627 if (RtsFlags.GcFlags.generations > 1) {
628 if (g0s0->blocks != NULL) {
629 freeChain(g0s0->blocks);
636 alloc_blocks_lim = RtsFlags.GcFlags.minAllocAreaSize;
638 // Start a new pinned_object_block
639 pinned_object_block = NULL;
641 // Free the mark stack.
642 if (mark_stack_bdescr != NULL) {
643 freeGroup(mark_stack_bdescr);
647 for (g = 0; g <= N; g++) {
648 for (s = 0; s < generations[g].n_steps; s++) {
649 stp = &generations[g].steps[s];
650 if (stp->bitmap != NULL) {
651 freeGroup(stp->bitmap);
659 // mark the garbage collected CAFs as dead
660 #if 0 && defined(DEBUG) // doesn't work at the moment
661 if (major_gc) { gcCAFs(); }
665 // resetStaticObjectForRetainerProfiling() must be called before
667 if (n_gc_threads > 1) {
668 barf("profiling is currently broken with multi-threaded GC");
669 // ToDo: fix the gct->scavenged_static_objects below
671 resetStaticObjectForRetainerProfiling(gct->scavenged_static_objects);
674 // zero the scavenged static object list
677 for (i = 0; i < n_gc_threads; i++) {
678 zero_static_object_list(gc_threads[i]->scavenged_static_objects);
685 // start any pending finalizers
687 scheduleFinalizers(last_free_capability, old_weak_ptr_list);
690 // send exceptions to any threads which were about to die
692 resurrectThreads(resurrected_threads);
695 // Update the stable pointer hash table.
696 updateStablePtrTable(major_gc);
698 // check sanity after GC
699 IF_DEBUG(sanity, checkSanity());
701 // extra GC trace info
702 if (traceClass(TRACE_gc|DEBUG_gc)) statDescribeGens();
705 // symbol-table based profiling
706 /* heapCensus(to_blocks); */ /* ToDo */
709 // restore enclosing cost centre
715 // check for memory leaks if DEBUG is on
716 memInventory(traceClass(DEBUG_gc));
719 #ifdef RTS_GTK_FRONTPANEL
720 if (RtsFlags.GcFlags.frontpanel) {
721 updateFrontPanelAfterGC( N, live );
725 // ok, GC over: tell the stats department what happened.
726 slop = calcLiveBlocks() * BLOCK_SIZE_W - live;
727 stat_endGC(allocated, live, copied, N, max_copied, avg_copied, slop);
729 #if defined(RTS_USER_SIGNALS)
730 if (RtsFlags.MiscFlags.install_signal_handlers) {
731 // unblock signals again
732 unblockUserSignals();
741 /* -----------------------------------------------------------------------------
742 Figure out which generation to collect, initialise N and major_gc.
744 Also returns the total number of blocks in generations that will be
746 -------------------------------------------------------------------------- */
749 initialise_N (rtsBool force_major_gc)
752 nat s, blocks, blocks_total;
757 if (force_major_gc) {
758 N = RtsFlags.GcFlags.generations - 1;
763 for (g = RtsFlags.GcFlags.generations - 1; g >= 0; g--) {
765 for (s = 0; s < generations[g].n_steps; s++) {
766 blocks += generations[g].steps[s].n_words / BLOCK_SIZE_W;
767 blocks += generations[g].steps[s].n_large_blocks;
769 if (blocks >= generations[g].max_blocks) {
773 blocks_total += blocks;
777 blocks_total += countNurseryBlocks();
779 major_gc = (N == RtsFlags.GcFlags.generations-1);
783 /* -----------------------------------------------------------------------------
784 Initialise the gc_thread structures.
785 -------------------------------------------------------------------------- */
788 alloc_gc_thread (int n)
794 t = stgMallocBytes(sizeof(gc_thread) + total_steps * sizeof(step_workspace),
799 initCondition(&t->wake_cond);
800 initMutex(&t->wake_mutex);
801 t->wakeup = rtsTrue; // starts true, so we can wait for the
802 // thread to start up, see wakeup_gc_threads
807 t->free_blocks = NULL;
816 for (s = 0; s < total_steps; s++)
819 ws->step = &all_steps[s];
820 ASSERT(s == ws->step->abs_no);
824 ws->buffer_todo_bd = NULL;
826 ws->part_list = NULL;
827 ws->n_part_blocks = 0;
829 ws->scavd_list = NULL;
830 ws->n_scavd_blocks = 0;
838 alloc_gc_threads (void)
840 if (gc_threads == NULL) {
841 #if defined(THREADED_RTS)
843 gc_threads = stgMallocBytes (RtsFlags.ParFlags.gcThreads *
847 for (i = 0; i < RtsFlags.ParFlags.gcThreads; i++) {
848 gc_threads[i] = alloc_gc_thread(i);
851 gc_threads = stgMallocBytes (sizeof(gc_thread*),
854 gc_threads[0] = alloc_gc_thread(0);
859 /* ----------------------------------------------------------------------------
861 ------------------------------------------------------------------------- */
863 static nat gc_running_threads;
865 #if defined(THREADED_RTS)
866 static Mutex gc_running_mutex;
873 ACQUIRE_LOCK(&gc_running_mutex);
874 n_running = ++gc_running_threads;
875 RELEASE_LOCK(&gc_running_mutex);
876 ASSERT(n_running <= n_gc_threads);
884 ACQUIRE_LOCK(&gc_running_mutex);
885 ASSERT(n_gc_threads != 0);
886 n_running = --gc_running_threads;
887 RELEASE_LOCK(&gc_running_mutex);
892 // gc_thread_work(): Scavenge until there's no work left to do and all
893 // the running threads are idle.
896 gc_thread_work (void)
900 debugTrace(DEBUG_gc, "GC thread %d working", gct->thread_index);
902 // gc_running_threads has already been incremented for us; either
903 // this is the main thread and we incremented it inside
904 // GarbageCollect(), or this is a worker thread and the main
905 // thread bumped gc_running_threads before waking us up.
907 // Every thread evacuates some roots.
909 markSomeCapabilities(mark_root, gct, gct->thread_index, n_gc_threads);
913 // scavenge_loop() only exits when there's no work to do
916 debugTrace(DEBUG_gc, "GC thread %d idle (%d still running)",
917 gct->thread_index, r);
919 while (gc_running_threads != 0) {
925 // any_work() does not remove the work from the queue, it
926 // just checks for the presence of work. If we find any,
927 // then we increment gc_running_threads and go back to
928 // scavenge_loop() to perform any pending work.
931 // All threads are now stopped
932 debugTrace(DEBUG_gc, "GC thread %d finished.", gct->thread_index);
936 #if defined(THREADED_RTS)
938 gc_thread_mainloop (void)
942 // Wait until we're told to wake up
943 ACQUIRE_LOCK(&gct->wake_mutex);
944 gct->wakeup = rtsFalse;
945 while (!gct->wakeup) {
946 debugTrace(DEBUG_gc, "GC thread %d standing by...",
948 waitCondition(&gct->wake_cond, &gct->wake_mutex);
950 RELEASE_LOCK(&gct->wake_mutex);
951 if (gct->exit) break;
954 // start performance counters in this thread...
955 if (gct->papi_events == -1) {
956 papi_init_eventset(&gct->papi_events);
958 papi_thread_start_gc1_count(gct->papi_events);
964 // count events in this thread towards the GC totals
965 papi_thread_stop_gc1_count(gct->papi_events);
971 #if defined(THREADED_RTS)
973 gc_thread_entry (gc_thread *my_gct)
976 debugTrace(DEBUG_gc, "GC thread %d starting...", gct->thread_index);
977 gct->id = osThreadId();
978 gc_thread_mainloop();
983 start_gc_threads (void)
985 #if defined(THREADED_RTS)
988 static rtsBool done = rtsFalse;
990 gc_running_threads = 0;
991 initMutex(&gc_running_mutex);
994 // Start from 1: the main thread is 0
995 for (i = 1; i < RtsFlags.ParFlags.gcThreads; i++) {
996 createOSThread(&id, (OSThreadProc*)&gc_thread_entry,
1005 wakeup_gc_threads (nat n_threads USED_IF_THREADS)
1007 #if defined(THREADED_RTS)
1009 for (i=1; i < n_threads; i++) {
1011 debugTrace(DEBUG_gc, "waking up gc thread %d", i);
1013 ACQUIRE_LOCK(&gc_threads[i]->wake_mutex);
1014 if (gc_threads[i]->wakeup) {
1015 RELEASE_LOCK(&gc_threads[i]->wake_mutex);
1021 gc_threads[i]->wakeup = rtsTrue;
1022 signalCondition(&gc_threads[i]->wake_cond);
1023 RELEASE_LOCK(&gc_threads[i]->wake_mutex);
1028 // After GC is complete, we must wait for all GC threads to enter the
1029 // standby state, otherwise they may still be executing inside
1030 // any_work(), and may even remain awake until the next GC starts.
1032 shutdown_gc_threads (nat n_threads USED_IF_THREADS)
1034 #if defined(THREADED_RTS)
1037 for (i=1; i < n_threads; i++) {
1039 ACQUIRE_LOCK(&gc_threads[i]->wake_mutex);
1040 wakeup = gc_threads[i]->wakeup;
1041 // wakeup is false while the thread is waiting
1042 RELEASE_LOCK(&gc_threads[i]->wake_mutex);
1048 /* ----------------------------------------------------------------------------
1049 Initialise a generation that is to be collected
1050 ------------------------------------------------------------------------- */
1053 init_collected_gen (nat g, nat n_threads)
1060 // Throw away the current mutable list. Invariant: the mutable
1061 // list always has at least one block; this means we can avoid a
1062 // check for NULL in recordMutable().
1064 freeChain(generations[g].mut_list);
1065 generations[g].mut_list = allocBlock();
1066 for (i = 0; i < n_capabilities; i++) {
1067 freeChain(capabilities[i].mut_lists[g]);
1068 capabilities[i].mut_lists[g] = allocBlock();
1072 for (s = 0; s < generations[g].n_steps; s++) {
1074 // generation 0, step 0 doesn't need to-space
1075 if (g == 0 && s == 0 && RtsFlags.GcFlags.generations > 1) {
1079 stp = &generations[g].steps[s];
1080 ASSERT(stp->gen_no == g);
1082 // deprecate the existing blocks
1083 stp->old_blocks = stp->blocks;
1084 stp->n_old_blocks = stp->n_blocks;
1089 // we don't have any to-be-scavenged blocks yet
1091 stp->todos_last = NULL;
1094 // initialise the large object queues.
1095 stp->scavenged_large_objects = NULL;
1096 stp->n_scavenged_large_blocks = 0;
1098 // mark the large objects as not evacuated yet
1099 for (bd = stp->large_objects; bd; bd = bd->link) {
1100 bd->flags &= ~BF_EVACUATED;
1103 // for a compacted step, we need to allocate the bitmap
1104 if (stp->is_compacted) {
1105 nat bitmap_size; // in bytes
1106 bdescr *bitmap_bdescr;
1109 bitmap_size = stp->n_old_blocks * BLOCK_SIZE / (sizeof(W_)*BITS_PER_BYTE);
1111 if (bitmap_size > 0) {
1112 bitmap_bdescr = allocGroup((lnat)BLOCK_ROUND_UP(bitmap_size)
1114 stp->bitmap = bitmap_bdescr;
1115 bitmap = bitmap_bdescr->start;
1117 debugTrace(DEBUG_gc, "bitmap_size: %d, bitmap: %p",
1118 bitmap_size, bitmap);
1120 // don't forget to fill it with zeros!
1121 memset(bitmap, 0, bitmap_size);
1123 // For each block in this step, point to its bitmap from the
1124 // block descriptor.
1125 for (bd=stp->old_blocks; bd != NULL; bd = bd->link) {
1126 bd->u.bitmap = bitmap;
1127 bitmap += BLOCK_SIZE_W / (sizeof(W_)*BITS_PER_BYTE);
1129 // Also at this point we set the BF_COMPACTED flag
1130 // for this block. The invariant is that
1131 // BF_COMPACTED is always unset, except during GC
1132 // when it is set on those blocks which will be
1134 bd->flags |= BF_COMPACTED;
1140 // For each GC thread, for each step, allocate a "todo" block to
1141 // store evacuated objects to be scavenged, and a block to store
1142 // evacuated objects that do not need to be scavenged.
1143 for (t = 0; t < n_threads; t++) {
1144 for (s = 0; s < generations[g].n_steps; s++) {
1146 // we don't copy objects into g0s0, unless -G0
1147 if (g==0 && s==0 && RtsFlags.GcFlags.generations > 1) continue;
1149 ws = &gc_threads[t]->steps[g * RtsFlags.GcFlags.steps + s];
1151 ws->todo_large_objects = NULL;
1153 ws->part_list = NULL;
1154 ws->n_part_blocks = 0;
1156 // allocate the first to-space block; extra blocks will be
1157 // chained on as necessary.
1159 ws->buffer_todo_bd = NULL;
1160 alloc_todo_block(ws,0);
1162 ws->scavd_list = NULL;
1163 ws->n_scavd_blocks = 0;
1169 /* ----------------------------------------------------------------------------
1170 Initialise a generation that is *not* to be collected
1171 ------------------------------------------------------------------------- */
1174 init_uncollected_gen (nat g, nat threads)
1181 for (s = 0; s < generations[g].n_steps; s++) {
1182 stp = &generations[g].steps[s];
1183 stp->scavenged_large_objects = NULL;
1184 stp->n_scavenged_large_blocks = 0;
1187 for (t = 0; t < threads; t++) {
1188 for (s = 0; s < generations[g].n_steps; s++) {
1190 ws = &gc_threads[t]->steps[g * RtsFlags.GcFlags.steps + s];
1193 ws->buffer_todo_bd = NULL;
1194 ws->todo_large_objects = NULL;
1196 ws->part_list = NULL;
1197 ws->n_part_blocks = 0;
1199 ws->scavd_list = NULL;
1200 ws->n_scavd_blocks = 0;
1202 // If the block at the head of the list in this generation
1203 // is less than 3/4 full, then use it as a todo block.
1204 if (stp->blocks && isPartiallyFull(stp->blocks))
1206 ws->todo_bd = stp->blocks;
1207 ws->todo_free = ws->todo_bd->free;
1208 ws->todo_lim = ws->todo_bd->start + BLOCK_SIZE_W;
1209 stp->blocks = stp->blocks->link;
1211 stp->n_words -= ws->todo_bd->free - ws->todo_bd->start;
1212 ws->todo_bd->link = NULL;
1213 // we must scan from the current end point.
1214 ws->todo_bd->u.scan = ws->todo_bd->free;
1219 alloc_todo_block(ws,0);
1224 // Move the private mutable lists from each capability onto the
1225 // main mutable list for the generation.
1226 for (i = 0; i < n_capabilities; i++) {
1227 for (bd = capabilities[i].mut_lists[g];
1228 bd->link != NULL; bd = bd->link) {
1231 bd->link = generations[g].mut_list;
1232 generations[g].mut_list = capabilities[i].mut_lists[g];
1233 capabilities[i].mut_lists[g] = allocBlock();
1237 /* -----------------------------------------------------------------------------
1238 Initialise a gc_thread before GC
1239 -------------------------------------------------------------------------- */
1242 init_gc_thread (gc_thread *t)
1244 t->static_objects = END_OF_STATIC_LIST;
1245 t->scavenged_static_objects = END_OF_STATIC_LIST;
1248 t->failed_to_evac = rtsFalse;
1249 t->eager_promotion = rtsTrue;
1250 t->thunk_selector_depth = 0;
1255 t->scav_find_work = 0;
1258 /* -----------------------------------------------------------------------------
1259 Function we pass to evacuate roots.
1260 -------------------------------------------------------------------------- */
1263 mark_root(void *user, StgClosure **root)
1265 // we stole a register for gct, but this function is called from
1266 // *outside* the GC where the register variable is not in effect,
1267 // so we need to save and restore it here. NB. only call
1268 // mark_root() from the main GC thread, otherwise gct will be
1270 gc_thread *saved_gct;
1279 /* -----------------------------------------------------------------------------
1280 Initialising the static object & mutable lists
1281 -------------------------------------------------------------------------- */
1284 zero_static_object_list(StgClosure* first_static)
1288 const StgInfoTable *info;
1290 for (p = first_static; p != END_OF_STATIC_LIST; p = link) {
1292 link = *STATIC_LINK(info, p);
1293 *STATIC_LINK(info,p) = NULL;
1297 /* ----------------------------------------------------------------------------
1298 Update the pointers from the task list
1300 These are treated as weak pointers because we want to allow a main
1301 thread to get a BlockedOnDeadMVar exception in the same way as any
1302 other thread. Note that the threads should all have been retained
1303 by GC by virtue of being on the all_threads list, we're just
1304 updating pointers here.
1305 ------------------------------------------------------------------------- */
1308 update_task_list (void)
1312 for (task = all_tasks; task != NULL; task = task->all_link) {
1313 if (!task->stopped && task->tso) {
1314 ASSERT(task->tso->bound == task);
1315 tso = (StgTSO *) isAlive((StgClosure *)task->tso);
1317 barf("task %p: main thread %d has been GC'd",
1330 /* ----------------------------------------------------------------------------
1331 Reset the sizes of the older generations when we do a major
1334 CURRENT STRATEGY: make all generations except zero the same size.
1335 We have to stay within the maximum heap size, and leave a certain
1336 percentage of the maximum heap size available to allocate into.
1337 ------------------------------------------------------------------------- */
1340 resize_generations (void)
1344 if (major_gc && RtsFlags.GcFlags.generations > 1) {
1345 nat live, size, min_alloc;
1346 nat max = RtsFlags.GcFlags.maxHeapSize;
1347 nat gens = RtsFlags.GcFlags.generations;
1349 // live in the oldest generations
1350 live = (oldest_gen->steps[0].n_words + BLOCK_SIZE_W - 1) / BLOCK_SIZE_W+
1351 oldest_gen->steps[0].n_large_blocks;
1353 // default max size for all generations except zero
1354 size = stg_max(live * RtsFlags.GcFlags.oldGenFactor,
1355 RtsFlags.GcFlags.minOldGenSize);
1357 // minimum size for generation zero
1358 min_alloc = stg_max((RtsFlags.GcFlags.pcFreeHeap * max) / 200,
1359 RtsFlags.GcFlags.minAllocAreaSize);
1361 // Auto-enable compaction when the residency reaches a
1362 // certain percentage of the maximum heap size (default: 30%).
1363 if (RtsFlags.GcFlags.generations > 1 &&
1364 (RtsFlags.GcFlags.compact ||
1366 oldest_gen->steps[0].n_blocks >
1367 (RtsFlags.GcFlags.compactThreshold * max) / 100))) {
1368 oldest_gen->steps[0].is_compacted = 1;
1369 // debugBelch("compaction: on\n", live);
1371 oldest_gen->steps[0].is_compacted = 0;
1372 // debugBelch("compaction: off\n", live);
1375 // if we're going to go over the maximum heap size, reduce the
1376 // size of the generations accordingly. The calculation is
1377 // different if compaction is turned on, because we don't need
1378 // to double the space required to collect the old generation.
1381 // this test is necessary to ensure that the calculations
1382 // below don't have any negative results - we're working
1383 // with unsigned values here.
1384 if (max < min_alloc) {
1388 if (oldest_gen->steps[0].is_compacted) {
1389 if ( (size + (size - 1) * (gens - 2) * 2) + min_alloc > max ) {
1390 size = (max - min_alloc) / ((gens - 1) * 2 - 1);
1393 if ( (size * (gens - 1) * 2) + min_alloc > max ) {
1394 size = (max - min_alloc) / ((gens - 1) * 2);
1404 debugBelch("live: %d, min_alloc: %d, size : %d, max = %d\n", live,
1405 min_alloc, size, max);
1408 for (g = 0; g < gens; g++) {
1409 generations[g].max_blocks = size;
1414 /* -----------------------------------------------------------------------------
1415 Calculate the new size of the nursery, and resize it.
1416 -------------------------------------------------------------------------- */
1419 resize_nursery (void)
1421 if (RtsFlags.GcFlags.generations == 1)
1422 { // Two-space collector:
1425 /* set up a new nursery. Allocate a nursery size based on a
1426 * function of the amount of live data (by default a factor of 2)
1427 * Use the blocks from the old nursery if possible, freeing up any
1430 * If we get near the maximum heap size, then adjust our nursery
1431 * size accordingly. If the nursery is the same size as the live
1432 * data (L), then we need 3L bytes. We can reduce the size of the
1433 * nursery to bring the required memory down near 2L bytes.
1435 * A normal 2-space collector would need 4L bytes to give the same
1436 * performance we get from 3L bytes, reducing to the same
1437 * performance at 2L bytes.
1439 blocks = g0s0->n_old_blocks;
1441 if ( RtsFlags.GcFlags.maxHeapSize != 0 &&
1442 blocks * RtsFlags.GcFlags.oldGenFactor * 2 >
1443 RtsFlags.GcFlags.maxHeapSize )
1445 long adjusted_blocks; // signed on purpose
1448 adjusted_blocks = (RtsFlags.GcFlags.maxHeapSize - 2 * blocks);
1450 debugTrace(DEBUG_gc, "near maximum heap size of 0x%x blocks, blocks = %d, adjusted to %ld",
1451 RtsFlags.GcFlags.maxHeapSize, blocks, adjusted_blocks);
1453 pc_free = adjusted_blocks * 100 / RtsFlags.GcFlags.maxHeapSize;
1454 if (pc_free < RtsFlags.GcFlags.pcFreeHeap) /* might even * be < 0 */
1458 blocks = adjusted_blocks;
1462 blocks *= RtsFlags.GcFlags.oldGenFactor;
1463 if (blocks < RtsFlags.GcFlags.minAllocAreaSize)
1465 blocks = RtsFlags.GcFlags.minAllocAreaSize;
1468 resizeNurseries(blocks);
1470 else // Generational collector
1473 * If the user has given us a suggested heap size, adjust our
1474 * allocation area to make best use of the memory available.
1476 if (RtsFlags.GcFlags.heapSizeSuggestion)
1479 nat needed = calcNeeded(); // approx blocks needed at next GC
1481 /* Guess how much will be live in generation 0 step 0 next time.
1482 * A good approximation is obtained by finding the
1483 * percentage of g0s0 that was live at the last minor GC.
1485 * We have an accurate figure for the amount of copied data in
1486 * 'copied', but we must convert this to a number of blocks, with
1487 * a small adjustment for estimated slop at the end of a block
1492 g0s0_pcnt_kept = ((copied / (BLOCK_SIZE_W - 10)) * 100)
1493 / countNurseryBlocks();
1496 /* Estimate a size for the allocation area based on the
1497 * information available. We might end up going slightly under
1498 * or over the suggested heap size, but we should be pretty
1501 * Formula: suggested - needed
1502 * ----------------------------
1503 * 1 + g0s0_pcnt_kept/100
1505 * where 'needed' is the amount of memory needed at the next
1506 * collection for collecting all steps except g0s0.
1509 (((long)RtsFlags.GcFlags.heapSizeSuggestion - (long)needed) * 100) /
1510 (100 + (long)g0s0_pcnt_kept);
1512 if (blocks < (long)RtsFlags.GcFlags.minAllocAreaSize) {
1513 blocks = RtsFlags.GcFlags.minAllocAreaSize;
1516 resizeNurseries((nat)blocks);
1520 // we might have added extra large blocks to the nursery, so
1521 // resize back to minAllocAreaSize again.
1522 resizeNurseriesFixed(RtsFlags.GcFlags.minAllocAreaSize);
1527 /* -----------------------------------------------------------------------------
1528 Sanity code for CAF garbage collection.
1530 With DEBUG turned on, we manage a CAF list in addition to the SRT
1531 mechanism. After GC, we run down the CAF list and blackhole any
1532 CAFs which have been garbage collected. This means we get an error
1533 whenever the program tries to enter a garbage collected CAF.
1535 Any garbage collected CAFs are taken off the CAF list at the same
1537 -------------------------------------------------------------------------- */
1539 #if 0 && defined(DEBUG)
1546 const StgInfoTable *info;
1557 ASSERT(info->type == IND_STATIC);
1559 if (STATIC_LINK(info,p) == NULL) {
1560 debugTrace(DEBUG_gccafs, "CAF gc'd at 0x%04lx", (long)p);
1562 SET_INFO(p,&stg_BLACKHOLE_info);
1563 p = STATIC_LINK2(info,p);
1567 pp = &STATIC_LINK2(info,p);
1574 debugTrace(DEBUG_gccafs, "%d CAFs live", i);