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"
53 #include <string.h> // for memset()
56 /* -----------------------------------------------------------------------------
58 -------------------------------------------------------------------------- */
60 /* STATIC OBJECT LIST.
63 * We maintain a linked list of static objects that are still live.
64 * The requirements for this list are:
66 * - we need to scan the list while adding to it, in order to
67 * scavenge all the static objects (in the same way that
68 * breadth-first scavenging works for dynamic objects).
70 * - we need to be able to tell whether an object is already on
71 * the list, to break loops.
73 * Each static object has a "static link field", which we use for
74 * linking objects on to the list. We use a stack-type list, consing
75 * objects on the front as they are added (this means that the
76 * scavenge phase is depth-first, not breadth-first, but that
79 * A separate list is kept for objects that have been scavenged
80 * already - this is so that we can zero all the marks afterwards.
82 * An object is on the list if its static link field is non-zero; this
83 * means that we have to mark the end of the list with '1', not NULL.
85 * Extra notes for generational GC:
87 * Each generation has a static object list associated with it. When
88 * collecting generations up to N, we treat the static object lists
89 * from generations > N as roots.
91 * We build up a static object list while collecting generations 0..N,
92 * which is then appended to the static object list of generation N+1.
95 /* N is the oldest generation being collected, where the generations
96 * are numbered starting at 0. A major GC (indicated by the major_gc
97 * flag) is when we're collecting all generations. We only attempt to
98 * deal with static objects and GC CAFs when doing a major GC.
103 /* Data used for allocation area sizing.
105 static lnat g0s0_pcnt_kept = 30; // percentage of g0s0 live at last minor GC
115 /* Thread-local data for each GC thread
117 gc_thread **gc_threads = NULL;
118 // gc_thread *gct = NULL; // this thread's gct TODO: make thread-local
120 // Number of threads running in *this* GC. Affects how many
121 // step->todos[] lists we have to look in to find work.
125 long copied; // *words* copied & scavenged during this GC
128 SpinLock recordMutableGen_sync;
133 /* -----------------------------------------------------------------------------
134 Static function declarations
135 -------------------------------------------------------------------------- */
137 static void mark_root (void *user, StgClosure **root);
138 static void zero_static_object_list (StgClosure* first_static);
139 static nat initialise_N (rtsBool force_major_gc);
140 static void alloc_gc_threads (void);
141 static void init_collected_gen (nat g, nat threads);
142 static void init_uncollected_gen (nat g, nat threads);
143 static void init_gc_thread (gc_thread *t);
144 static void update_task_list (void);
145 static void resize_generations (void);
146 static void resize_nursery (void);
147 static void start_gc_threads (void);
148 static void scavenge_until_all_done (void);
149 static nat inc_running (void);
150 static nat dec_running (void);
151 static void wakeup_gc_threads (nat n_threads);
152 static void shutdown_gc_threads (nat n_threads);
154 #if 0 && defined(DEBUG)
155 static void gcCAFs (void);
158 /* -----------------------------------------------------------------------------
159 The mark bitmap & stack.
160 -------------------------------------------------------------------------- */
162 #define MARK_STACK_BLOCKS 4
164 bdescr *mark_stack_bdescr;
169 // Flag and pointers used for falling back to a linear scan when the
170 // mark stack overflows.
171 rtsBool mark_stack_overflowed;
172 bdescr *oldgen_scan_bd;
175 /* -----------------------------------------------------------------------------
176 GarbageCollect: the main entry point to the garbage collector.
178 Locks held: all capabilities are held throughout GarbageCollect().
179 -------------------------------------------------------------------------- */
182 GarbageCollect ( rtsBool force_major_gc )
186 lnat live, allocated, max_copied, avg_copied, slop;
187 gc_thread *saved_gct;
190 // necessary if we stole a callee-saves register for gct:
194 CostCentreStack *prev_CCS;
199 #if defined(RTS_USER_SIGNALS)
200 if (RtsFlags.MiscFlags.install_signal_handlers) {
206 ASSERT(sizeof(step_workspace) == 16 * sizeof(StgWord));
207 // otherwise adjust the padding in step_workspace.
209 // tell the stats department that we've started a GC
212 // tell the STM to discard any cached closures it's hoping to re-use
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 // Update pointers from capabilities (probably just the spark queues)
381 updateCapabilitiesPostGC();
383 // Now see which stable names are still alive.
387 // We call processHeapClosureForDead() on every closure destroyed during
388 // the current garbage collection, so we invoke LdvCensusForDead().
389 if (RtsFlags.ProfFlags.doHeapProfile == HEAP_BY_LDV
390 || RtsFlags.ProfFlags.bioSelector != NULL)
394 // NO MORE EVACUATION AFTER THIS POINT!
396 // Two-space collector: free the old to-space.
397 // g0s0->old_blocks is the old nursery
398 // g0s0->blocks is to-space from the previous GC
399 if (RtsFlags.GcFlags.generations == 1) {
400 if (g0s0->blocks != NULL) {
401 freeChain(g0s0->blocks);
406 // For each workspace, in each thread, move the copied blocks to the step
412 for (t = 0; t < n_gc_threads; t++) {
416 if (RtsFlags.GcFlags.generations == 1) {
421 for (; s < total_steps; s++) {
424 // Push the final block
426 push_scanned_block(ws->todo_bd, ws);
429 ASSERT(gct->scan_bd == NULL);
430 ASSERT(countBlocks(ws->scavd_list) == ws->n_scavd_blocks);
433 for (bd = ws->scavd_list; bd != NULL; bd = bd->link) {
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;
445 // Add all the partial blocks *after* we've added all the full
446 // blocks. This is so that we can grab the partial blocks back
447 // again and try to fill them up in the next GC.
448 for (t = 0; t < n_gc_threads; t++) {
452 if (RtsFlags.GcFlags.generations == 1) {
457 for (; s < total_steps; s++) {
461 for (bd = ws->part_list; bd != NULL; bd = next) {
463 if (bd->free == bd->start) {
465 ws->part_list = next;
472 ws->step->n_words += bd->free - bd->start;
477 prev->link = ws->step->blocks;
478 ws->step->blocks = ws->part_list;
480 ws->step->n_blocks += ws->n_part_blocks;
482 ASSERT(countBlocks(ws->step->blocks) == ws->step->n_blocks);
483 ASSERT(countOccupied(ws->step->blocks) == ws->step->n_words);
488 // Finally: compaction of the oldest generation.
489 if (major_gc && oldest_gen->steps[0].is_compacted) {
490 compact(gct->scavenged_static_objects);
493 IF_DEBUG(sanity, checkGlobalTSOList(rtsFalse));
495 /* run through all the generations/steps and tidy up
502 for (i=0; i < n_gc_threads; i++) {
503 if (n_gc_threads > 1) {
504 trace(TRACE_gc,"thread %d:", i);
505 trace(TRACE_gc," copied %ld", gc_threads[i]->copied * sizeof(W_));
506 trace(TRACE_gc," scanned %ld", gc_threads[i]->scanned * sizeof(W_));
507 trace(TRACE_gc," any_work %ld", gc_threads[i]->any_work);
508 trace(TRACE_gc," no_work %ld", gc_threads[i]->no_work);
509 trace(TRACE_gc," scav_find_work %ld", gc_threads[i]->scav_find_work);
511 copied += gc_threads[i]->copied;
512 max_copied = stg_max(gc_threads[i]->copied, max_copied);
514 if (n_gc_threads == 1) {
522 for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
525 generations[g].collections++; // for stats
526 if (n_gc_threads > 1) generations[g].par_collections++;
529 // Count the mutable list as bytes "copied" for the purposes of
530 // stats. Every mutable list is copied during every GC.
532 nat mut_list_size = 0;
533 for (bd = generations[g].mut_list; bd != NULL; bd = bd->link) {
534 mut_list_size += bd->free - bd->start;
536 copied += mut_list_size;
539 "mut_list_size: %lu (%d vars, %d arrays, %d MVARs, %d others)",
540 (unsigned long)(mut_list_size * sizeof(W_)),
541 mutlist_MUTVARS, mutlist_MUTARRS, mutlist_MVARS, mutlist_OTHERS);
544 for (s = 0; s < generations[g].n_steps; s++) {
546 stp = &generations[g].steps[s];
548 // for generations we collected...
551 /* free old memory and shift to-space into from-space for all
552 * the collected steps (except the allocation area). These
553 * freed blocks will probaby be quickly recycled.
555 if (!(g == 0 && s == 0 && RtsFlags.GcFlags.generations > 1)) {
556 if (stp->is_compacted)
558 // tack the new blocks on the end of the existing blocks
559 if (stp->old_blocks != NULL) {
560 for (bd = stp->old_blocks; bd != NULL; bd = next) {
561 stp->n_words += bd->free - bd->start;
563 // NB. this step might not be compacted next
564 // time, so reset the BF_COMPACTED flags.
565 // They are set before GC if we're going to
566 // compact. (search for BF_COMPACTED above).
567 bd->flags &= ~BF_COMPACTED;
569 // between GCs, all blocks in the heap except
570 // for the nursery have the BF_EVACUATED flag set.
571 bd->flags |= BF_EVACUATED;
575 bd->link = stp->blocks;
578 stp->blocks = stp->old_blocks;
580 // add the new blocks to the block tally
581 stp->n_blocks += stp->n_old_blocks;
582 ASSERT(countBlocks(stp->blocks) == stp->n_blocks);
583 ASSERT(countOccupied(stp->blocks) == stp->n_words);
587 freeChain(stp->old_blocks);
589 stp->old_blocks = NULL;
590 stp->n_old_blocks = 0;
593 /* LARGE OBJECTS. The current live large objects are chained on
594 * scavenged_large, having been moved during garbage
595 * collection from large_objects. Any objects left on
596 * large_objects list are therefore dead, so we free them here.
598 for (bd = stp->large_objects; bd != NULL; bd = next) {
604 stp->large_objects = stp->scavenged_large_objects;
605 stp->n_large_blocks = stp->n_scavenged_large_blocks;
608 else // for older generations...
610 /* For older generations, we need to append the
611 * scavenged_large_object list (i.e. large objects that have been
612 * promoted during this GC) to the large_object list for that step.
614 for (bd = stp->scavenged_large_objects; bd; bd = next) {
616 dbl_link_onto(bd, &stp->large_objects);
619 // add the new blocks we promoted during this GC
620 stp->n_large_blocks += stp->n_scavenged_large_blocks;
625 // update the max size of older generations after a major GC
626 resize_generations();
628 // Calculate the amount of live data for stats.
629 live = calcLiveWords();
631 // Free the small objects allocated via allocate(), since this will
632 // all have been copied into G0S1 now.
633 if (RtsFlags.GcFlags.generations > 1) {
634 if (g0s0->blocks != NULL) {
635 freeChain(g0s0->blocks);
642 alloc_blocks_lim = RtsFlags.GcFlags.minAllocAreaSize;
644 // Start a new pinned_object_block
645 pinned_object_block = NULL;
647 // Free the mark stack.
648 if (mark_stack_bdescr != NULL) {
649 freeGroup(mark_stack_bdescr);
653 for (g = 0; g <= N; g++) {
654 for (s = 0; s < generations[g].n_steps; s++) {
655 stp = &generations[g].steps[s];
656 if (stp->bitmap != NULL) {
657 freeGroup(stp->bitmap);
665 // mark the garbage collected CAFs as dead
666 #if 0 && defined(DEBUG) // doesn't work at the moment
667 if (major_gc) { gcCAFs(); }
671 // resetStaticObjectForRetainerProfiling() must be called before
673 if (n_gc_threads > 1) {
674 barf("profiling is currently broken with multi-threaded GC");
675 // ToDo: fix the gct->scavenged_static_objects below
677 resetStaticObjectForRetainerProfiling(gct->scavenged_static_objects);
680 // zero the scavenged static object list
683 for (i = 0; i < n_gc_threads; i++) {
684 zero_static_object_list(gc_threads[i]->scavenged_static_objects);
691 // start any pending finalizers
693 scheduleFinalizers(last_free_capability, old_weak_ptr_list);
696 // send exceptions to any threads which were about to die
698 resurrectThreads(resurrected_threads);
699 performPendingThrowTos(exception_threads);
702 // Update the stable pointer hash table.
703 updateStablePtrTable(major_gc);
705 // check sanity after GC
706 IF_DEBUG(sanity, checkSanity());
708 // extra GC trace info
709 if (traceClass(TRACE_gc|DEBUG_gc)) statDescribeGens();
712 // symbol-table based profiling
713 /* heapCensus(to_blocks); */ /* ToDo */
716 // restore enclosing cost centre
722 // check for memory leaks if DEBUG is on
723 memInventory(traceClass(DEBUG_gc));
726 #ifdef RTS_GTK_FRONTPANEL
727 if (RtsFlags.GcFlags.frontpanel) {
728 updateFrontPanelAfterGC( N, live );
732 // ok, GC over: tell the stats department what happened.
733 slop = calcLiveBlocks() * BLOCK_SIZE_W - live;
734 stat_endGC(allocated, live, copied, N, max_copied, avg_copied, slop);
736 #if defined(RTS_USER_SIGNALS)
737 if (RtsFlags.MiscFlags.install_signal_handlers) {
738 // unblock signals again
739 unblockUserSignals();
748 /* -----------------------------------------------------------------------------
749 Figure out which generation to collect, initialise N and major_gc.
751 Also returns the total number of blocks in generations that will be
753 -------------------------------------------------------------------------- */
756 initialise_N (rtsBool force_major_gc)
759 nat s, blocks, blocks_total;
764 if (force_major_gc) {
765 N = RtsFlags.GcFlags.generations - 1;
770 for (g = RtsFlags.GcFlags.generations - 1; g >= 0; g--) {
772 for (s = 0; s < generations[g].n_steps; s++) {
773 blocks += generations[g].steps[s].n_words / BLOCK_SIZE_W;
774 blocks += generations[g].steps[s].n_large_blocks;
776 if (blocks >= generations[g].max_blocks) {
780 blocks_total += blocks;
784 blocks_total += countNurseryBlocks();
786 major_gc = (N == RtsFlags.GcFlags.generations-1);
790 /* -----------------------------------------------------------------------------
791 Initialise the gc_thread structures.
792 -------------------------------------------------------------------------- */
795 alloc_gc_thread (int n)
801 t = stgMallocBytes(sizeof(gc_thread) + total_steps * sizeof(step_workspace),
806 initCondition(&t->wake_cond);
807 initMutex(&t->wake_mutex);
808 t->wakeup = rtsTrue; // starts true, so we can wait for the
809 // thread to start up, see wakeup_gc_threads
814 t->free_blocks = NULL;
823 for (s = 0; s < total_steps; s++)
826 ws->step = &all_steps[s];
827 ASSERT(s == ws->step->abs_no);
831 ws->buffer_todo_bd = NULL;
833 ws->part_list = NULL;
834 ws->n_part_blocks = 0;
836 ws->scavd_list = NULL;
837 ws->n_scavd_blocks = 0;
845 alloc_gc_threads (void)
847 if (gc_threads == NULL) {
848 #if defined(THREADED_RTS)
850 gc_threads = stgMallocBytes (RtsFlags.ParFlags.gcThreads *
854 for (i = 0; i < RtsFlags.ParFlags.gcThreads; i++) {
855 gc_threads[i] = alloc_gc_thread(i);
858 gc_threads = stgMallocBytes (sizeof(gc_thread*),
861 gc_threads[0] = alloc_gc_thread(0);
866 /* ----------------------------------------------------------------------------
868 ------------------------------------------------------------------------- */
870 static nat gc_running_threads;
872 #if defined(THREADED_RTS)
873 static Mutex gc_running_mutex;
880 ACQUIRE_LOCK(&gc_running_mutex);
881 n_running = ++gc_running_threads;
882 RELEASE_LOCK(&gc_running_mutex);
883 ASSERT(n_running <= n_gc_threads);
891 ACQUIRE_LOCK(&gc_running_mutex);
892 ASSERT(n_gc_threads != 0);
893 n_running = --gc_running_threads;
894 RELEASE_LOCK(&gc_running_mutex);
908 // scavenge objects in compacted generation
909 if (mark_stack_overflowed || oldgen_scan_bd != NULL ||
910 (mark_stack_bdescr != NULL && !mark_stack_empty())) {
914 // Check for global work in any step. We don't need to check for
915 // local work, because we have already exited scavenge_loop(),
916 // which means there is no local work for this thread.
917 for (s = total_steps-1; s >= 0; s--) {
918 if (s == 0 && RtsFlags.GcFlags.generations > 1) {
922 if (ws->todo_large_objects) return rtsTrue;
923 if (ws->step->todos) return rtsTrue;
932 scavenge_until_all_done (void)
936 debugTrace(DEBUG_gc, "GC thread %d working", gct->thread_index);
939 #if defined(THREADED_RTS)
940 if (n_gc_threads > 1) {
949 // scavenge_loop() only exits when there's no work to do
952 debugTrace(DEBUG_gc, "GC thread %d idle (%d still running)",
953 gct->thread_index, r);
955 while (gc_running_threads != 0) {
961 // any_work() does not remove the work from the queue, it
962 // just checks for the presence of work. If we find any,
963 // then we increment gc_running_threads and go back to
964 // scavenge_loop() to perform any pending work.
967 // All threads are now stopped
968 debugTrace(DEBUG_gc, "GC thread %d finished.", gct->thread_index);
971 #if defined(THREADED_RTS)
973 // gc_thread_work(): Scavenge until there's no work left to do and all
974 // the running threads are idle.
977 gc_thread_work (void)
979 // gc_running_threads has already been incremented for us; this is
980 // a worker thread and the main thread bumped gc_running_threads
981 // before waking us up.
983 // Every thread evacuates some roots.
985 markSomeCapabilities(mark_root, gct, gct->thread_index, n_gc_threads);
987 scavenge_until_all_done();
992 gc_thread_mainloop (void)
996 // Wait until we're told to wake up
997 ACQUIRE_LOCK(&gct->wake_mutex);
998 gct->wakeup = rtsFalse;
999 while (!gct->wakeup) {
1000 debugTrace(DEBUG_gc, "GC thread %d standing by...",
1002 waitCondition(&gct->wake_cond, &gct->wake_mutex);
1004 RELEASE_LOCK(&gct->wake_mutex);
1005 if (gct->exit) break;
1008 // start performance counters in this thread...
1009 if (gct->papi_events == -1) {
1010 papi_init_eventset(&gct->papi_events);
1012 papi_thread_start_gc1_count(gct->papi_events);
1018 // count events in this thread towards the GC totals
1019 papi_thread_stop_gc1_count(gct->papi_events);
1025 #if defined(THREADED_RTS)
1027 gc_thread_entry (gc_thread *my_gct)
1030 debugTrace(DEBUG_gc, "GC thread %d starting...", gct->thread_index);
1031 gct->id = osThreadId();
1032 gc_thread_mainloop();
1037 start_gc_threads (void)
1039 #if defined(THREADED_RTS)
1042 static rtsBool done = rtsFalse;
1044 gc_running_threads = 0;
1045 initMutex(&gc_running_mutex);
1048 // Start from 1: the main thread is 0
1049 for (i = 1; i < RtsFlags.ParFlags.gcThreads; i++) {
1050 createOSThread(&id, (OSThreadProc*)&gc_thread_entry,
1059 wakeup_gc_threads (nat n_threads USED_IF_THREADS)
1061 #if defined(THREADED_RTS)
1063 for (i=1; i < n_threads; i++) {
1065 debugTrace(DEBUG_gc, "waking up gc thread %d", i);
1067 ACQUIRE_LOCK(&gc_threads[i]->wake_mutex);
1068 if (gc_threads[i]->wakeup) {
1069 RELEASE_LOCK(&gc_threads[i]->wake_mutex);
1075 gc_threads[i]->wakeup = rtsTrue;
1076 signalCondition(&gc_threads[i]->wake_cond);
1077 RELEASE_LOCK(&gc_threads[i]->wake_mutex);
1082 // After GC is complete, we must wait for all GC threads to enter the
1083 // standby state, otherwise they may still be executing inside
1084 // any_work(), and may even remain awake until the next GC starts.
1086 shutdown_gc_threads (nat n_threads USED_IF_THREADS)
1088 #if defined(THREADED_RTS)
1091 for (i=1; i < n_threads; i++) {
1093 ACQUIRE_LOCK(&gc_threads[i]->wake_mutex);
1094 wakeup = gc_threads[i]->wakeup;
1095 // wakeup is false while the thread is waiting
1096 RELEASE_LOCK(&gc_threads[i]->wake_mutex);
1102 /* ----------------------------------------------------------------------------
1103 Initialise a generation that is to be collected
1104 ------------------------------------------------------------------------- */
1107 init_collected_gen (nat g, nat n_threads)
1114 // Throw away the current mutable list. Invariant: the mutable
1115 // list always has at least one block; this means we can avoid a
1116 // check for NULL in recordMutable().
1118 freeChain(generations[g].mut_list);
1119 generations[g].mut_list = allocBlock();
1120 for (i = 0; i < n_capabilities; i++) {
1121 freeChain(capabilities[i].mut_lists[g]);
1122 capabilities[i].mut_lists[g] = allocBlock();
1126 for (s = 0; s < generations[g].n_steps; s++) {
1128 stp = &generations[g].steps[s];
1129 ASSERT(stp->gen_no == g);
1131 // we'll construct a new list of threads in this step
1132 // during GC, throw away the current list.
1133 stp->old_threads = stp->threads;
1134 stp->threads = END_TSO_QUEUE;
1136 // generation 0, step 0 doesn't need to-space
1137 if (g == 0 && s == 0 && RtsFlags.GcFlags.generations > 1) {
1141 // deprecate the existing blocks
1142 stp->old_blocks = stp->blocks;
1143 stp->n_old_blocks = stp->n_blocks;
1148 // we don't have any to-be-scavenged blocks yet
1150 stp->todos_last = NULL;
1153 // initialise the large object queues.
1154 stp->scavenged_large_objects = NULL;
1155 stp->n_scavenged_large_blocks = 0;
1157 // mark the small objects as from-space
1158 for (bd = stp->old_blocks; bd; bd = bd->link) {
1159 bd->flags &= ~BF_EVACUATED;
1162 // mark the large objects as from-space
1163 for (bd = stp->large_objects; bd; bd = bd->link) {
1164 bd->flags &= ~BF_EVACUATED;
1167 // for a compacted step, we need to allocate the bitmap
1168 if (stp->is_compacted) {
1169 nat bitmap_size; // in bytes
1170 bdescr *bitmap_bdescr;
1173 bitmap_size = stp->n_old_blocks * BLOCK_SIZE / (sizeof(W_)*BITS_PER_BYTE);
1175 if (bitmap_size > 0) {
1176 bitmap_bdescr = allocGroup((lnat)BLOCK_ROUND_UP(bitmap_size)
1178 stp->bitmap = bitmap_bdescr;
1179 bitmap = bitmap_bdescr->start;
1181 debugTrace(DEBUG_gc, "bitmap_size: %d, bitmap: %p",
1182 bitmap_size, bitmap);
1184 // don't forget to fill it with zeros!
1185 memset(bitmap, 0, bitmap_size);
1187 // For each block in this step, point to its bitmap from the
1188 // block descriptor.
1189 for (bd=stp->old_blocks; bd != NULL; bd = bd->link) {
1190 bd->u.bitmap = bitmap;
1191 bitmap += BLOCK_SIZE_W / (sizeof(W_)*BITS_PER_BYTE);
1193 // Also at this point we set the BF_COMPACTED flag
1194 // for this block. The invariant is that
1195 // BF_COMPACTED is always unset, except during GC
1196 // when it is set on those blocks which will be
1198 bd->flags |= BF_COMPACTED;
1204 // For each GC thread, for each step, allocate a "todo" block to
1205 // store evacuated objects to be scavenged, and a block to store
1206 // evacuated objects that do not need to be scavenged.
1207 for (t = 0; t < n_threads; t++) {
1208 for (s = 0; s < generations[g].n_steps; s++) {
1210 // we don't copy objects into g0s0, unless -G0
1211 if (g==0 && s==0 && RtsFlags.GcFlags.generations > 1) continue;
1213 ws = &gc_threads[t]->steps[g * RtsFlags.GcFlags.steps + s];
1215 ws->todo_large_objects = NULL;
1217 ws->part_list = NULL;
1218 ws->n_part_blocks = 0;
1220 // allocate the first to-space block; extra blocks will be
1221 // chained on as necessary.
1223 ws->buffer_todo_bd = NULL;
1224 alloc_todo_block(ws,0);
1226 ws->scavd_list = NULL;
1227 ws->n_scavd_blocks = 0;
1233 /* ----------------------------------------------------------------------------
1234 Initialise a generation that is *not* to be collected
1235 ------------------------------------------------------------------------- */
1238 init_uncollected_gen (nat g, nat threads)
1245 for (s = 0; s < generations[g].n_steps; s++) {
1246 stp = &generations[g].steps[s];
1247 stp->scavenged_large_objects = NULL;
1248 stp->n_scavenged_large_blocks = 0;
1251 for (s = 0; s < generations[g].n_steps; s++) {
1253 stp = &generations[g].steps[s];
1255 for (t = 0; t < threads; t++) {
1256 ws = &gc_threads[t]->steps[g * RtsFlags.GcFlags.steps + s];
1258 ws->buffer_todo_bd = NULL;
1259 ws->todo_large_objects = NULL;
1261 ws->part_list = NULL;
1262 ws->n_part_blocks = 0;
1264 ws->scavd_list = NULL;
1265 ws->n_scavd_blocks = 0;
1267 // If the block at the head of the list in this generation
1268 // is less than 3/4 full, then use it as a todo block.
1269 if (stp->blocks && isPartiallyFull(stp->blocks))
1271 ws->todo_bd = stp->blocks;
1272 ws->todo_free = ws->todo_bd->free;
1273 ws->todo_lim = ws->todo_bd->start + BLOCK_SIZE_W;
1274 stp->blocks = stp->blocks->link;
1276 stp->n_words -= ws->todo_bd->free - ws->todo_bd->start;
1277 ws->todo_bd->link = NULL;
1278 // we must scan from the current end point.
1279 ws->todo_bd->u.scan = ws->todo_bd->free;
1284 alloc_todo_block(ws,0);
1288 // deal out any more partial blocks to the threads' part_lists
1290 while (stp->blocks && isPartiallyFull(stp->blocks))
1293 stp->blocks = bd->link;
1294 ws = &gc_threads[t]->steps[g * RtsFlags.GcFlags.steps + s];
1295 bd->link = ws->part_list;
1297 ws->n_part_blocks += 1;
1298 bd->u.scan = bd->free;
1300 stp->n_words -= bd->free - bd->start;
1302 if (t == n_gc_threads) t = 0;
1307 // Move the private mutable lists from each capability onto the
1308 // main mutable list for the generation.
1309 for (i = 0; i < n_capabilities; i++) {
1310 for (bd = capabilities[i].mut_lists[g];
1311 bd->link != NULL; bd = bd->link) {
1314 bd->link = generations[g].mut_list;
1315 generations[g].mut_list = capabilities[i].mut_lists[g];
1316 capabilities[i].mut_lists[g] = allocBlock();
1320 /* -----------------------------------------------------------------------------
1321 Initialise a gc_thread before GC
1322 -------------------------------------------------------------------------- */
1325 init_gc_thread (gc_thread *t)
1327 t->static_objects = END_OF_STATIC_LIST;
1328 t->scavenged_static_objects = END_OF_STATIC_LIST;
1331 t->failed_to_evac = rtsFalse;
1332 t->eager_promotion = rtsTrue;
1333 t->thunk_selector_depth = 0;
1338 t->scav_find_work = 0;
1341 /* -----------------------------------------------------------------------------
1342 Function we pass to evacuate roots.
1343 -------------------------------------------------------------------------- */
1346 mark_root(void *user, StgClosure **root)
1348 // we stole a register for gct, but this function is called from
1349 // *outside* the GC where the register variable is not in effect,
1350 // so we need to save and restore it here. NB. only call
1351 // mark_root() from the main GC thread, otherwise gct will be
1353 gc_thread *saved_gct;
1362 /* -----------------------------------------------------------------------------
1363 Initialising the static object & mutable lists
1364 -------------------------------------------------------------------------- */
1367 zero_static_object_list(StgClosure* first_static)
1371 const StgInfoTable *info;
1373 for (p = first_static; p != END_OF_STATIC_LIST; p = link) {
1375 link = *STATIC_LINK(info, p);
1376 *STATIC_LINK(info,p) = NULL;
1380 /* ----------------------------------------------------------------------------
1381 Update the pointers from the task list
1383 These are treated as weak pointers because we want to allow a main
1384 thread to get a BlockedOnDeadMVar exception in the same way as any
1385 other thread. Note that the threads should all have been retained
1386 by GC by virtue of being on the all_threads list, we're just
1387 updating pointers here.
1388 ------------------------------------------------------------------------- */
1391 update_task_list (void)
1395 for (task = all_tasks; task != NULL; task = task->all_link) {
1396 if (!task->stopped && task->tso) {
1397 ASSERT(task->tso->bound == task);
1398 tso = (StgTSO *) isAlive((StgClosure *)task->tso);
1400 barf("task %p: main thread %d has been GC'd",
1413 /* ----------------------------------------------------------------------------
1414 Reset the sizes of the older generations when we do a major
1417 CURRENT STRATEGY: make all generations except zero the same size.
1418 We have to stay within the maximum heap size, and leave a certain
1419 percentage of the maximum heap size available to allocate into.
1420 ------------------------------------------------------------------------- */
1423 resize_generations (void)
1427 if (major_gc && RtsFlags.GcFlags.generations > 1) {
1428 nat live, size, min_alloc;
1429 nat max = RtsFlags.GcFlags.maxHeapSize;
1430 nat gens = RtsFlags.GcFlags.generations;
1432 // live in the oldest generations
1433 live = (oldest_gen->steps[0].n_words + BLOCK_SIZE_W - 1) / BLOCK_SIZE_W+
1434 oldest_gen->steps[0].n_large_blocks;
1436 // default max size for all generations except zero
1437 size = stg_max(live * RtsFlags.GcFlags.oldGenFactor,
1438 RtsFlags.GcFlags.minOldGenSize);
1440 // minimum size for generation zero
1441 min_alloc = stg_max((RtsFlags.GcFlags.pcFreeHeap * max) / 200,
1442 RtsFlags.GcFlags.minAllocAreaSize);
1444 // Auto-enable compaction when the residency reaches a
1445 // certain percentage of the maximum heap size (default: 30%).
1446 if (RtsFlags.GcFlags.generations > 1 &&
1447 (RtsFlags.GcFlags.compact ||
1449 oldest_gen->steps[0].n_blocks >
1450 (RtsFlags.GcFlags.compactThreshold * max) / 100))) {
1451 oldest_gen->steps[0].is_compacted = 1;
1452 // debugBelch("compaction: on\n", live);
1454 oldest_gen->steps[0].is_compacted = 0;
1455 // debugBelch("compaction: off\n", live);
1458 // if we're going to go over the maximum heap size, reduce the
1459 // size of the generations accordingly. The calculation is
1460 // different if compaction is turned on, because we don't need
1461 // to double the space required to collect the old generation.
1464 // this test is necessary to ensure that the calculations
1465 // below don't have any negative results - we're working
1466 // with unsigned values here.
1467 if (max < min_alloc) {
1471 if (oldest_gen->steps[0].is_compacted) {
1472 if ( (size + (size - 1) * (gens - 2) * 2) + min_alloc > max ) {
1473 size = (max - min_alloc) / ((gens - 1) * 2 - 1);
1476 if ( (size * (gens - 1) * 2) + min_alloc > max ) {
1477 size = (max - min_alloc) / ((gens - 1) * 2);
1487 debugBelch("live: %d, min_alloc: %d, size : %d, max = %d\n", live,
1488 min_alloc, size, max);
1491 for (g = 0; g < gens; g++) {
1492 generations[g].max_blocks = size;
1497 /* -----------------------------------------------------------------------------
1498 Calculate the new size of the nursery, and resize it.
1499 -------------------------------------------------------------------------- */
1502 resize_nursery (void)
1504 if (RtsFlags.GcFlags.generations == 1)
1505 { // Two-space collector:
1508 /* set up a new nursery. Allocate a nursery size based on a
1509 * function of the amount of live data (by default a factor of 2)
1510 * Use the blocks from the old nursery if possible, freeing up any
1513 * If we get near the maximum heap size, then adjust our nursery
1514 * size accordingly. If the nursery is the same size as the live
1515 * data (L), then we need 3L bytes. We can reduce the size of the
1516 * nursery to bring the required memory down near 2L bytes.
1518 * A normal 2-space collector would need 4L bytes to give the same
1519 * performance we get from 3L bytes, reducing to the same
1520 * performance at 2L bytes.
1522 blocks = g0s0->n_blocks;
1524 if ( RtsFlags.GcFlags.maxHeapSize != 0 &&
1525 blocks * RtsFlags.GcFlags.oldGenFactor * 2 >
1526 RtsFlags.GcFlags.maxHeapSize )
1528 long adjusted_blocks; // signed on purpose
1531 adjusted_blocks = (RtsFlags.GcFlags.maxHeapSize - 2 * blocks);
1533 debugTrace(DEBUG_gc, "near maximum heap size of 0x%x blocks, blocks = %d, adjusted to %ld",
1534 RtsFlags.GcFlags.maxHeapSize, blocks, adjusted_blocks);
1536 pc_free = adjusted_blocks * 100 / RtsFlags.GcFlags.maxHeapSize;
1537 if (pc_free < RtsFlags.GcFlags.pcFreeHeap) /* might even * be < 0 */
1541 blocks = adjusted_blocks;
1545 blocks *= RtsFlags.GcFlags.oldGenFactor;
1546 if (blocks < RtsFlags.GcFlags.minAllocAreaSize)
1548 blocks = RtsFlags.GcFlags.minAllocAreaSize;
1551 resizeNurseries(blocks);
1553 else // Generational collector
1556 * If the user has given us a suggested heap size, adjust our
1557 * allocation area to make best use of the memory available.
1559 if (RtsFlags.GcFlags.heapSizeSuggestion)
1562 nat needed = calcNeeded(); // approx blocks needed at next GC
1564 /* Guess how much will be live in generation 0 step 0 next time.
1565 * A good approximation is obtained by finding the
1566 * percentage of g0s0 that was live at the last minor GC.
1568 * We have an accurate figure for the amount of copied data in
1569 * 'copied', but we must convert this to a number of blocks, with
1570 * a small adjustment for estimated slop at the end of a block
1575 g0s0_pcnt_kept = ((copied / (BLOCK_SIZE_W - 10)) * 100)
1576 / countNurseryBlocks();
1579 /* Estimate a size for the allocation area based on the
1580 * information available. We might end up going slightly under
1581 * or over the suggested heap size, but we should be pretty
1584 * Formula: suggested - needed
1585 * ----------------------------
1586 * 1 + g0s0_pcnt_kept/100
1588 * where 'needed' is the amount of memory needed at the next
1589 * collection for collecting all steps except g0s0.
1592 (((long)RtsFlags.GcFlags.heapSizeSuggestion - (long)needed) * 100) /
1593 (100 + (long)g0s0_pcnt_kept);
1595 if (blocks < (long)RtsFlags.GcFlags.minAllocAreaSize) {
1596 blocks = RtsFlags.GcFlags.minAllocAreaSize;
1599 resizeNurseries((nat)blocks);
1603 // we might have added extra large blocks to the nursery, so
1604 // resize back to minAllocAreaSize again.
1605 resizeNurseriesFixed(RtsFlags.GcFlags.minAllocAreaSize);
1610 /* -----------------------------------------------------------------------------
1611 Sanity code for CAF garbage collection.
1613 With DEBUG turned on, we manage a CAF list in addition to the SRT
1614 mechanism. After GC, we run down the CAF list and blackhole any
1615 CAFs which have been garbage collected. This means we get an error
1616 whenever the program tries to enter a garbage collected CAF.
1618 Any garbage collected CAFs are taken off the CAF list at the same
1620 -------------------------------------------------------------------------- */
1622 #if 0 && defined(DEBUG)
1629 const StgInfoTable *info;
1640 ASSERT(info->type == IND_STATIC);
1642 if (STATIC_LINK(info,p) == NULL) {
1643 debugTrace(DEBUG_gccafs, "CAF gc'd at 0x%04lx", (long)p);
1645 SET_INFO(p,&stg_BLACKHOLE_info);
1646 p = STATIC_LINK2(info,p);
1650 pp = &STATIC_LINK2(info,p);
1657 debugTrace(DEBUG_gccafs, "%d CAFs live", i);