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;
131 /* -----------------------------------------------------------------------------
132 Static function declarations
133 -------------------------------------------------------------------------- */
135 static void mark_root (void *user, StgClosure **root);
136 static void zero_static_object_list (StgClosure* first_static);
137 static nat initialise_N (rtsBool force_major_gc);
138 static void alloc_gc_threads (void);
139 static void init_collected_gen (nat g, nat threads);
140 static void init_uncollected_gen (nat g, nat threads);
141 static void init_gc_thread (gc_thread *t);
142 static void update_task_list (void);
143 static void resize_generations (void);
144 static void resize_nursery (void);
145 static void start_gc_threads (void);
146 static void scavenge_until_all_done (void);
147 static nat inc_running (void);
148 static nat dec_running (void);
149 static void wakeup_gc_threads (nat n_threads);
150 static void shutdown_gc_threads (nat n_threads);
152 #if 0 && defined(DEBUG)
153 static void gcCAFs (void);
156 /* -----------------------------------------------------------------------------
157 The mark bitmap & stack.
158 -------------------------------------------------------------------------- */
160 #define MARK_STACK_BLOCKS 4
162 bdescr *mark_stack_bdescr;
167 // Flag and pointers used for falling back to a linear scan when the
168 // mark stack overflows.
169 rtsBool mark_stack_overflowed;
170 bdescr *oldgen_scan_bd;
173 /* -----------------------------------------------------------------------------
174 GarbageCollect: the main entry point to the garbage collector.
176 Locks held: all capabilities are held throughout GarbageCollect().
177 -------------------------------------------------------------------------- */
180 GarbageCollect ( rtsBool force_major_gc )
184 lnat live, allocated, max_copied, avg_copied, slop;
185 gc_thread *saved_gct;
188 // necessary if we stole a callee-saves register for gct:
192 CostCentreStack *prev_CCS;
197 #if defined(RTS_USER_SIGNALS)
198 if (RtsFlags.MiscFlags.install_signal_handlers) {
204 ASSERT(sizeof(step_workspace) == 16 * sizeof(StgWord));
205 // otherwise adjust the padding in step_workspace.
207 // tell the stats department that we've started a GC
210 // tell the STM to discard any cached closures it's hoping to re-use
219 // attribute any costs to CCS_GC
225 /* Approximate how much we allocated.
226 * Todo: only when generating stats?
228 allocated = calcAllocated();
230 /* Figure out which generation to collect
232 n = initialise_N(force_major_gc);
234 /* Allocate + initialise the gc_thread structures.
238 /* Start threads, so they can be spinning up while we finish initialisation.
242 /* How many threads will be participating in this GC?
243 * We don't try to parallelise minor GC.
245 #if defined(THREADED_RTS)
246 if (n < (4*1024*1024 / BLOCK_SIZE)) {
249 n_gc_threads = RtsFlags.ParFlags.gcThreads;
254 trace(TRACE_gc|DEBUG_gc, "GC (gen %d): %d KB to collect, %ld MB in use, using %d thread(s)",
255 N, n * (BLOCK_SIZE / 1024), mblocks_allocated, n_gc_threads);
257 #ifdef RTS_GTK_FRONTPANEL
258 if (RtsFlags.GcFlags.frontpanel) {
259 updateFrontPanelBeforeGC(N);
264 // check for memory leaks if DEBUG is on
265 memInventory(traceClass(DEBUG_gc));
268 // check stack sanity *before* GC (ToDo: check all threads)
269 IF_DEBUG(sanity, checkFreeListSanity());
271 // Initialise all our gc_thread structures
272 for (t = 0; t < n_gc_threads; t++) {
273 init_gc_thread(gc_threads[t]);
276 // Initialise all the generations/steps that we're collecting.
277 for (g = 0; g <= N; g++) {
278 init_collected_gen(g,n_gc_threads);
281 // Initialise all the generations/steps that we're *not* collecting.
282 for (g = N+1; g < RtsFlags.GcFlags.generations; g++) {
283 init_uncollected_gen(g,n_gc_threads);
286 /* Allocate a mark stack if we're doing a major collection.
289 mark_stack_bdescr = allocGroup(MARK_STACK_BLOCKS);
290 mark_stack = (StgPtr *)mark_stack_bdescr->start;
291 mark_sp = mark_stack;
292 mark_splim = mark_stack + (MARK_STACK_BLOCKS * BLOCK_SIZE_W);
294 mark_stack_bdescr = NULL;
297 // this is the main thread
300 /* -----------------------------------------------------------------------
301 * follow all the roots that we know about:
302 * - mutable lists from each generation > N
303 * we want to *scavenge* these roots, not evacuate them: they're not
304 * going to move in this GC.
305 * Also do them in reverse generation order, for the usual reason:
306 * namely to reduce the likelihood of spurious old->new pointers.
308 for (g = RtsFlags.GcFlags.generations-1; g > N; g--) {
309 generations[g].saved_mut_list = generations[g].mut_list;
310 generations[g].mut_list = allocBlock();
311 // mut_list always has at least one block.
314 // the main thread is running: this prevents any other threads from
315 // exiting prematurely, so we can start them now.
316 // NB. do this after the mutable lists have been saved above, otherwise
317 // the other GC threads will be writing into the old mutable lists.
319 wakeup_gc_threads(n_gc_threads);
321 for (g = RtsFlags.GcFlags.generations-1; g > N; g--) {
322 scavenge_mutable_list(&generations[g]);
325 // follow roots from the CAF list (used by GHCi)
327 markCAFs(mark_root, gct);
329 // follow all the roots that the application knows about.
331 markSomeCapabilities(mark_root, gct, gct->thread_index, n_gc_threads);
333 #if defined(RTS_USER_SIGNALS)
334 // mark the signal handlers (signals should be already blocked)
335 markSignalHandlers(mark_root, gct);
338 // Mark the weak pointer list, and prepare to detect dead weak pointers.
342 // Mark the stable pointer table.
343 markStablePtrTable(mark_root, gct);
345 /* -------------------------------------------------------------------------
346 * Repeatedly scavenge all the areas we know about until there's no
347 * more scavenging to be done.
351 scavenge_until_all_done();
352 // The other threads are now stopped. We might recurse back to
353 // here, but from now on this is the only thread.
355 // if any blackholes are alive, make the threads that wait on
357 if (traverseBlackholeQueue()) {
362 // must be last... invariant is that everything is fully
363 // scavenged at this point.
364 if (traverseWeakPtrList()) { // returns rtsTrue if evaced something
369 // If we get to here, there's really nothing left to do.
373 shutdown_gc_threads(n_gc_threads);
375 // Update pointers from the Task list
378 // Update pointers from capabilities (probably just the spark queues)
379 updateCapabilitiesPostGC();
381 // Now see which stable names are still alive.
385 // We call processHeapClosureForDead() on every closure destroyed during
386 // the current garbage collection, so we invoke LdvCensusForDead().
387 if (RtsFlags.ProfFlags.doHeapProfile == HEAP_BY_LDV
388 || RtsFlags.ProfFlags.bioSelector != NULL)
392 // NO MORE EVACUATION AFTER THIS POINT!
394 // Two-space collector: free the old to-space.
395 // g0s0->old_blocks is the old nursery
396 // g0s0->blocks is to-space from the previous GC
397 if (RtsFlags.GcFlags.generations == 1) {
398 if (g0s0->blocks != NULL) {
399 freeChain(g0s0->blocks);
404 // For each workspace, in each thread, move the copied blocks to the step
410 for (t = 0; t < n_gc_threads; t++) {
414 if (RtsFlags.GcFlags.generations == 1) {
419 for (; s < total_steps; s++) {
422 // Push the final block
424 push_scanned_block(ws->todo_bd, ws);
427 ASSERT(gct->scan_bd == NULL);
428 ASSERT(countBlocks(ws->scavd_list) == ws->n_scavd_blocks);
431 for (bd = ws->scavd_list; bd != NULL; bd = bd->link) {
432 ws->step->n_words += bd->free - bd->start;
436 prev->link = ws->step->blocks;
437 ws->step->blocks = ws->scavd_list;
439 ws->step->n_blocks += ws->n_scavd_blocks;
443 // Add all the partial blocks *after* we've added all the full
444 // blocks. This is so that we can grab the partial blocks back
445 // again and try to fill them up in the next GC.
446 for (t = 0; t < n_gc_threads; t++) {
450 if (RtsFlags.GcFlags.generations == 1) {
455 for (; s < total_steps; s++) {
459 for (bd = ws->part_list; bd != NULL; bd = next) {
461 if (bd->free == bd->start) {
463 ws->part_list = next;
470 ws->step->n_words += bd->free - bd->start;
475 prev->link = ws->step->blocks;
476 ws->step->blocks = ws->part_list;
478 ws->step->n_blocks += ws->n_part_blocks;
480 ASSERT(countBlocks(ws->step->blocks) == ws->step->n_blocks);
481 ASSERT(countOccupied(ws->step->blocks) == ws->step->n_words);
486 // Finally: compaction of the oldest generation.
487 if (major_gc && oldest_gen->steps[0].is_compacted) {
488 compact(gct->scavenged_static_objects);
491 IF_DEBUG(sanity, checkGlobalTSOList(rtsFalse));
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 // tack the new blocks on the end of the existing blocks
557 if (stp->old_blocks != NULL) {
558 for (bd = stp->old_blocks; bd != NULL; bd = next) {
559 stp->n_words += bd->free - bd->start;
561 // NB. this step might not be compacted next
562 // time, so reset the BF_COMPACTED flags.
563 // They are set before GC if we're going to
564 // compact. (search for BF_COMPACTED above).
565 bd->flags &= ~BF_COMPACTED;
567 // between GCs, all blocks in the heap except
568 // for the nursery have the BF_EVACUATED flag set.
569 bd->flags |= BF_EVACUATED;
573 bd->link = stp->blocks;
576 stp->blocks = stp->old_blocks;
578 // add the new blocks to the block tally
579 stp->n_blocks += stp->n_old_blocks;
580 ASSERT(countBlocks(stp->blocks) == stp->n_blocks);
581 ASSERT(countOccupied(stp->blocks) == stp->n_words);
585 freeChain(stp->old_blocks);
587 stp->old_blocks = NULL;
588 stp->n_old_blocks = 0;
591 /* LARGE OBJECTS. The current live large objects are chained on
592 * scavenged_large, having been moved during garbage
593 * collection from large_objects. Any objects left on
594 * large_objects list are therefore dead, so we free them here.
596 for (bd = stp->large_objects; bd != NULL; bd = next) {
602 stp->large_objects = stp->scavenged_large_objects;
603 stp->n_large_blocks = stp->n_scavenged_large_blocks;
606 else // for older generations...
608 /* For older generations, we need to append the
609 * scavenged_large_object list (i.e. large objects that have been
610 * promoted during this GC) to the large_object list for that step.
612 for (bd = stp->scavenged_large_objects; bd; bd = next) {
614 dbl_link_onto(bd, &stp->large_objects);
617 // add the new blocks we promoted during this GC
618 stp->n_large_blocks += stp->n_scavenged_large_blocks;
623 // update the max size of older generations after a major GC
624 resize_generations();
626 // Calculate the amount of live data for stats.
627 live = calcLiveWords();
629 // Free the small objects allocated via allocate(), since this will
630 // all have been copied into G0S1 now.
631 if (RtsFlags.GcFlags.generations > 1) {
632 if (g0s0->blocks != NULL) {
633 freeChain(g0s0->blocks);
640 alloc_blocks_lim = RtsFlags.GcFlags.minAllocAreaSize;
642 // Start a new pinned_object_block
643 pinned_object_block = NULL;
645 // Free the mark stack.
646 if (mark_stack_bdescr != NULL) {
647 freeGroup(mark_stack_bdescr);
651 for (g = 0; g <= N; g++) {
652 for (s = 0; s < generations[g].n_steps; s++) {
653 stp = &generations[g].steps[s];
654 if (stp->bitmap != NULL) {
655 freeGroup(stp->bitmap);
663 // mark the garbage collected CAFs as dead
664 #if 0 && defined(DEBUG) // doesn't work at the moment
665 if (major_gc) { gcCAFs(); }
669 // resetStaticObjectForRetainerProfiling() must be called before
671 if (n_gc_threads > 1) {
672 barf("profiling is currently broken with multi-threaded GC");
673 // ToDo: fix the gct->scavenged_static_objects below
675 resetStaticObjectForRetainerProfiling(gct->scavenged_static_objects);
678 // zero the scavenged static object list
681 for (i = 0; i < n_gc_threads; i++) {
682 zero_static_object_list(gc_threads[i]->scavenged_static_objects);
689 // start any pending finalizers
691 scheduleFinalizers(last_free_capability, old_weak_ptr_list);
694 // send exceptions to any threads which were about to die
696 resurrectThreads(resurrected_threads);
697 performPendingThrowTos(exception_threads);
700 // Update the stable pointer hash table.
701 updateStablePtrTable(major_gc);
703 // check sanity after GC
704 IF_DEBUG(sanity, checkSanity());
706 // extra GC trace info
707 if (traceClass(TRACE_gc|DEBUG_gc)) statDescribeGens();
710 // symbol-table based profiling
711 /* heapCensus(to_blocks); */ /* ToDo */
714 // restore enclosing cost centre
720 // check for memory leaks if DEBUG is on
721 memInventory(traceClass(DEBUG_gc));
724 #ifdef RTS_GTK_FRONTPANEL
725 if (RtsFlags.GcFlags.frontpanel) {
726 updateFrontPanelAfterGC( N, live );
730 // ok, GC over: tell the stats department what happened.
731 slop = calcLiveBlocks() * BLOCK_SIZE_W - live;
732 stat_endGC(allocated, live, copied, N, max_copied, avg_copied, slop);
734 #if defined(RTS_USER_SIGNALS)
735 if (RtsFlags.MiscFlags.install_signal_handlers) {
736 // unblock signals again
737 unblockUserSignals();
746 /* -----------------------------------------------------------------------------
747 Figure out which generation to collect, initialise N and major_gc.
749 Also returns the total number of blocks in generations that will be
751 -------------------------------------------------------------------------- */
754 initialise_N (rtsBool force_major_gc)
757 nat s, blocks, blocks_total;
762 if (force_major_gc) {
763 N = RtsFlags.GcFlags.generations - 1;
768 for (g = RtsFlags.GcFlags.generations - 1; g >= 0; g--) {
770 for (s = 0; s < generations[g].n_steps; s++) {
771 blocks += generations[g].steps[s].n_words / BLOCK_SIZE_W;
772 blocks += generations[g].steps[s].n_large_blocks;
774 if (blocks >= generations[g].max_blocks) {
778 blocks_total += blocks;
782 blocks_total += countNurseryBlocks();
784 major_gc = (N == RtsFlags.GcFlags.generations-1);
788 /* -----------------------------------------------------------------------------
789 Initialise the gc_thread structures.
790 -------------------------------------------------------------------------- */
793 alloc_gc_thread (int n)
799 t = stgMallocBytes(sizeof(gc_thread) + total_steps * sizeof(step_workspace),
804 initCondition(&t->wake_cond);
805 initMutex(&t->wake_mutex);
806 t->wakeup = rtsTrue; // starts true, so we can wait for the
807 // thread to start up, see wakeup_gc_threads
812 t->free_blocks = NULL;
821 for (s = 0; s < total_steps; s++)
824 ws->step = &all_steps[s];
825 ASSERT(s == ws->step->abs_no);
829 ws->buffer_todo_bd = NULL;
831 ws->part_list = NULL;
832 ws->n_part_blocks = 0;
834 ws->scavd_list = NULL;
835 ws->n_scavd_blocks = 0;
843 alloc_gc_threads (void)
845 if (gc_threads == NULL) {
846 #if defined(THREADED_RTS)
848 gc_threads = stgMallocBytes (RtsFlags.ParFlags.gcThreads *
852 for (i = 0; i < RtsFlags.ParFlags.gcThreads; i++) {
853 gc_threads[i] = alloc_gc_thread(i);
856 gc_threads = stgMallocBytes (sizeof(gc_thread*),
859 gc_threads[0] = alloc_gc_thread(0);
864 /* ----------------------------------------------------------------------------
866 ------------------------------------------------------------------------- */
868 static nat gc_running_threads;
870 #if defined(THREADED_RTS)
871 static Mutex gc_running_mutex;
878 ACQUIRE_LOCK(&gc_running_mutex);
879 n_running = ++gc_running_threads;
880 RELEASE_LOCK(&gc_running_mutex);
881 ASSERT(n_running <= n_gc_threads);
889 ACQUIRE_LOCK(&gc_running_mutex);
890 ASSERT(n_gc_threads != 0);
891 n_running = --gc_running_threads;
892 RELEASE_LOCK(&gc_running_mutex);
897 scavenge_until_all_done (void)
901 debugTrace(DEBUG_gc, "GC thread %d working", gct->thread_index);
905 // scavenge_loop() only exits when there's no work to do
908 debugTrace(DEBUG_gc, "GC thread %d idle (%d still running)",
909 gct->thread_index, r);
911 while (gc_running_threads != 0) {
917 // any_work() does not remove the work from the queue, it
918 // just checks for the presence of work. If we find any,
919 // then we increment gc_running_threads and go back to
920 // scavenge_loop() to perform any pending work.
923 // All threads are now stopped
924 debugTrace(DEBUG_gc, "GC thread %d finished.", gct->thread_index);
927 #if defined(THREADED_RTS)
929 // gc_thread_work(): Scavenge until there's no work left to do and all
930 // the running threads are idle.
933 gc_thread_work (void)
935 // gc_running_threads has already been incremented for us; this is
936 // a worker thread and the main thread bumped gc_running_threads
937 // before waking us up.
939 // Every thread evacuates some roots.
941 markSomeCapabilities(mark_root, gct, gct->thread_index, n_gc_threads);
943 scavenge_until_all_done();
948 gc_thread_mainloop (void)
952 // Wait until we're told to wake up
953 ACQUIRE_LOCK(&gct->wake_mutex);
954 gct->wakeup = rtsFalse;
955 while (!gct->wakeup) {
956 debugTrace(DEBUG_gc, "GC thread %d standing by...",
958 waitCondition(&gct->wake_cond, &gct->wake_mutex);
960 RELEASE_LOCK(&gct->wake_mutex);
961 if (gct->exit) break;
964 // start performance counters in this thread...
965 if (gct->papi_events == -1) {
966 papi_init_eventset(&gct->papi_events);
968 papi_thread_start_gc1_count(gct->papi_events);
974 // count events in this thread towards the GC totals
975 papi_thread_stop_gc1_count(gct->papi_events);
981 #if defined(THREADED_RTS)
983 gc_thread_entry (gc_thread *my_gct)
986 debugTrace(DEBUG_gc, "GC thread %d starting...", gct->thread_index);
987 gct->id = osThreadId();
988 gc_thread_mainloop();
993 start_gc_threads (void)
995 #if defined(THREADED_RTS)
998 static rtsBool done = rtsFalse;
1000 gc_running_threads = 0;
1001 initMutex(&gc_running_mutex);
1004 // Start from 1: the main thread is 0
1005 for (i = 1; i < RtsFlags.ParFlags.gcThreads; i++) {
1006 createOSThread(&id, (OSThreadProc*)&gc_thread_entry,
1015 wakeup_gc_threads (nat n_threads USED_IF_THREADS)
1017 #if defined(THREADED_RTS)
1019 for (i=1; i < n_threads; i++) {
1021 debugTrace(DEBUG_gc, "waking up gc thread %d", i);
1023 ACQUIRE_LOCK(&gc_threads[i]->wake_mutex);
1024 if (gc_threads[i]->wakeup) {
1025 RELEASE_LOCK(&gc_threads[i]->wake_mutex);
1031 gc_threads[i]->wakeup = rtsTrue;
1032 signalCondition(&gc_threads[i]->wake_cond);
1033 RELEASE_LOCK(&gc_threads[i]->wake_mutex);
1038 // After GC is complete, we must wait for all GC threads to enter the
1039 // standby state, otherwise they may still be executing inside
1040 // any_work(), and may even remain awake until the next GC starts.
1042 shutdown_gc_threads (nat n_threads USED_IF_THREADS)
1044 #if defined(THREADED_RTS)
1047 for (i=1; i < n_threads; i++) {
1049 ACQUIRE_LOCK(&gc_threads[i]->wake_mutex);
1050 wakeup = gc_threads[i]->wakeup;
1051 // wakeup is false while the thread is waiting
1052 RELEASE_LOCK(&gc_threads[i]->wake_mutex);
1058 /* ----------------------------------------------------------------------------
1059 Initialise a generation that is to be collected
1060 ------------------------------------------------------------------------- */
1063 init_collected_gen (nat g, nat n_threads)
1070 // Throw away the current mutable list. Invariant: the mutable
1071 // list always has at least one block; this means we can avoid a
1072 // check for NULL in recordMutable().
1074 freeChain(generations[g].mut_list);
1075 generations[g].mut_list = allocBlock();
1076 for (i = 0; i < n_capabilities; i++) {
1077 freeChain(capabilities[i].mut_lists[g]);
1078 capabilities[i].mut_lists[g] = allocBlock();
1082 for (s = 0; s < generations[g].n_steps; s++) {
1084 stp = &generations[g].steps[s];
1085 ASSERT(stp->gen_no == g);
1087 // we'll construct a new list of threads in this step
1088 // during GC, throw away the current list.
1089 stp->old_threads = stp->threads;
1090 stp->threads = END_TSO_QUEUE;
1092 // generation 0, step 0 doesn't need to-space
1093 if (g == 0 && s == 0 && RtsFlags.GcFlags.generations > 1) {
1097 // deprecate the existing blocks
1098 stp->old_blocks = stp->blocks;
1099 stp->n_old_blocks = stp->n_blocks;
1104 // we don't have any to-be-scavenged blocks yet
1106 stp->todos_last = NULL;
1109 // initialise the large object queues.
1110 stp->scavenged_large_objects = NULL;
1111 stp->n_scavenged_large_blocks = 0;
1113 // mark the small objects as from-space
1114 for (bd = stp->old_blocks; bd; bd = bd->link) {
1115 bd->flags &= ~BF_EVACUATED;
1118 // mark the large objects as from-space
1119 for (bd = stp->large_objects; bd; bd = bd->link) {
1120 bd->flags &= ~BF_EVACUATED;
1123 // for a compacted step, we need to allocate the bitmap
1124 if (stp->is_compacted) {
1125 nat bitmap_size; // in bytes
1126 bdescr *bitmap_bdescr;
1129 bitmap_size = stp->n_old_blocks * BLOCK_SIZE / (sizeof(W_)*BITS_PER_BYTE);
1131 if (bitmap_size > 0) {
1132 bitmap_bdescr = allocGroup((lnat)BLOCK_ROUND_UP(bitmap_size)
1134 stp->bitmap = bitmap_bdescr;
1135 bitmap = bitmap_bdescr->start;
1137 debugTrace(DEBUG_gc, "bitmap_size: %d, bitmap: %p",
1138 bitmap_size, bitmap);
1140 // don't forget to fill it with zeros!
1141 memset(bitmap, 0, bitmap_size);
1143 // For each block in this step, point to its bitmap from the
1144 // block descriptor.
1145 for (bd=stp->old_blocks; bd != NULL; bd = bd->link) {
1146 bd->u.bitmap = bitmap;
1147 bitmap += BLOCK_SIZE_W / (sizeof(W_)*BITS_PER_BYTE);
1149 // Also at this point we set the BF_COMPACTED flag
1150 // for this block. The invariant is that
1151 // BF_COMPACTED is always unset, except during GC
1152 // when it is set on those blocks which will be
1154 bd->flags |= BF_COMPACTED;
1160 // For each GC thread, for each step, allocate a "todo" block to
1161 // store evacuated objects to be scavenged, and a block to store
1162 // evacuated objects that do not need to be scavenged.
1163 for (t = 0; t < n_threads; t++) {
1164 for (s = 0; s < generations[g].n_steps; s++) {
1166 // we don't copy objects into g0s0, unless -G0
1167 if (g==0 && s==0 && RtsFlags.GcFlags.generations > 1) continue;
1169 ws = &gc_threads[t]->steps[g * RtsFlags.GcFlags.steps + s];
1171 ws->todo_large_objects = NULL;
1173 ws->part_list = NULL;
1174 ws->n_part_blocks = 0;
1176 // allocate the first to-space block; extra blocks will be
1177 // chained on as necessary.
1179 ws->buffer_todo_bd = NULL;
1180 alloc_todo_block(ws,0);
1182 ws->scavd_list = NULL;
1183 ws->n_scavd_blocks = 0;
1189 /* ----------------------------------------------------------------------------
1190 Initialise a generation that is *not* to be collected
1191 ------------------------------------------------------------------------- */
1194 init_uncollected_gen (nat g, nat threads)
1201 for (s = 0; s < generations[g].n_steps; s++) {
1202 stp = &generations[g].steps[s];
1203 stp->scavenged_large_objects = NULL;
1204 stp->n_scavenged_large_blocks = 0;
1207 for (s = 0; s < generations[g].n_steps; s++) {
1209 stp = &generations[g].steps[s];
1211 for (t = 0; t < threads; t++) {
1212 ws = &gc_threads[t]->steps[g * RtsFlags.GcFlags.steps + s];
1214 ws->buffer_todo_bd = NULL;
1215 ws->todo_large_objects = NULL;
1217 ws->part_list = NULL;
1218 ws->n_part_blocks = 0;
1220 ws->scavd_list = NULL;
1221 ws->n_scavd_blocks = 0;
1223 // If the block at the head of the list in this generation
1224 // is less than 3/4 full, then use it as a todo block.
1225 if (stp->blocks && isPartiallyFull(stp->blocks))
1227 ws->todo_bd = stp->blocks;
1228 ws->todo_free = ws->todo_bd->free;
1229 ws->todo_lim = ws->todo_bd->start + BLOCK_SIZE_W;
1230 stp->blocks = stp->blocks->link;
1232 stp->n_words -= ws->todo_bd->free - ws->todo_bd->start;
1233 ws->todo_bd->link = NULL;
1234 // we must scan from the current end point.
1235 ws->todo_bd->u.scan = ws->todo_bd->free;
1240 alloc_todo_block(ws,0);
1244 // deal out any more partial blocks to the threads' part_lists
1246 while (stp->blocks && isPartiallyFull(stp->blocks))
1249 stp->blocks = bd->link;
1250 ws = &gc_threads[t]->steps[g * RtsFlags.GcFlags.steps + s];
1251 bd->link = ws->part_list;
1253 ws->n_part_blocks += 1;
1254 bd->u.scan = bd->free;
1256 stp->n_words -= bd->free - bd->start;
1258 if (t == n_gc_threads) t = 0;
1263 // Move the private mutable lists from each capability onto the
1264 // main mutable list for the generation.
1265 for (i = 0; i < n_capabilities; i++) {
1266 for (bd = capabilities[i].mut_lists[g];
1267 bd->link != NULL; bd = bd->link) {
1270 bd->link = generations[g].mut_list;
1271 generations[g].mut_list = capabilities[i].mut_lists[g];
1272 capabilities[i].mut_lists[g] = allocBlock();
1276 /* -----------------------------------------------------------------------------
1277 Initialise a gc_thread before GC
1278 -------------------------------------------------------------------------- */
1281 init_gc_thread (gc_thread *t)
1283 t->static_objects = END_OF_STATIC_LIST;
1284 t->scavenged_static_objects = END_OF_STATIC_LIST;
1287 t->failed_to_evac = rtsFalse;
1288 t->eager_promotion = rtsTrue;
1289 t->thunk_selector_depth = 0;
1294 t->scav_find_work = 0;
1297 /* -----------------------------------------------------------------------------
1298 Function we pass to evacuate roots.
1299 -------------------------------------------------------------------------- */
1302 mark_root(void *user, StgClosure **root)
1304 // we stole a register for gct, but this function is called from
1305 // *outside* the GC where the register variable is not in effect,
1306 // so we need to save and restore it here. NB. only call
1307 // mark_root() from the main GC thread, otherwise gct will be
1309 gc_thread *saved_gct;
1318 /* -----------------------------------------------------------------------------
1319 Initialising the static object & mutable lists
1320 -------------------------------------------------------------------------- */
1323 zero_static_object_list(StgClosure* first_static)
1327 const StgInfoTable *info;
1329 for (p = first_static; p != END_OF_STATIC_LIST; p = link) {
1331 link = *STATIC_LINK(info, p);
1332 *STATIC_LINK(info,p) = NULL;
1336 /* ----------------------------------------------------------------------------
1337 Update the pointers from the task list
1339 These are treated as weak pointers because we want to allow a main
1340 thread to get a BlockedOnDeadMVar exception in the same way as any
1341 other thread. Note that the threads should all have been retained
1342 by GC by virtue of being on the all_threads list, we're just
1343 updating pointers here.
1344 ------------------------------------------------------------------------- */
1347 update_task_list (void)
1351 for (task = all_tasks; task != NULL; task = task->all_link) {
1352 if (!task->stopped && task->tso) {
1353 ASSERT(task->tso->bound == task);
1354 tso = (StgTSO *) isAlive((StgClosure *)task->tso);
1356 barf("task %p: main thread %d has been GC'd",
1369 /* ----------------------------------------------------------------------------
1370 Reset the sizes of the older generations when we do a major
1373 CURRENT STRATEGY: make all generations except zero the same size.
1374 We have to stay within the maximum heap size, and leave a certain
1375 percentage of the maximum heap size available to allocate into.
1376 ------------------------------------------------------------------------- */
1379 resize_generations (void)
1383 if (major_gc && RtsFlags.GcFlags.generations > 1) {
1384 nat live, size, min_alloc;
1385 nat max = RtsFlags.GcFlags.maxHeapSize;
1386 nat gens = RtsFlags.GcFlags.generations;
1388 // live in the oldest generations
1389 live = (oldest_gen->steps[0].n_words + BLOCK_SIZE_W - 1) / BLOCK_SIZE_W+
1390 oldest_gen->steps[0].n_large_blocks;
1392 // default max size for all generations except zero
1393 size = stg_max(live * RtsFlags.GcFlags.oldGenFactor,
1394 RtsFlags.GcFlags.minOldGenSize);
1396 // minimum size for generation zero
1397 min_alloc = stg_max((RtsFlags.GcFlags.pcFreeHeap * max) / 200,
1398 RtsFlags.GcFlags.minAllocAreaSize);
1400 // Auto-enable compaction when the residency reaches a
1401 // certain percentage of the maximum heap size (default: 30%).
1402 if (RtsFlags.GcFlags.generations > 1 &&
1403 (RtsFlags.GcFlags.compact ||
1405 oldest_gen->steps[0].n_blocks >
1406 (RtsFlags.GcFlags.compactThreshold * max) / 100))) {
1407 oldest_gen->steps[0].is_compacted = 1;
1408 // debugBelch("compaction: on\n", live);
1410 oldest_gen->steps[0].is_compacted = 0;
1411 // debugBelch("compaction: off\n", live);
1414 // if we're going to go over the maximum heap size, reduce the
1415 // size of the generations accordingly. The calculation is
1416 // different if compaction is turned on, because we don't need
1417 // to double the space required to collect the old generation.
1420 // this test is necessary to ensure that the calculations
1421 // below don't have any negative results - we're working
1422 // with unsigned values here.
1423 if (max < min_alloc) {
1427 if (oldest_gen->steps[0].is_compacted) {
1428 if ( (size + (size - 1) * (gens - 2) * 2) + min_alloc > max ) {
1429 size = (max - min_alloc) / ((gens - 1) * 2 - 1);
1432 if ( (size * (gens - 1) * 2) + min_alloc > max ) {
1433 size = (max - min_alloc) / ((gens - 1) * 2);
1443 debugBelch("live: %d, min_alloc: %d, size : %d, max = %d\n", live,
1444 min_alloc, size, max);
1447 for (g = 0; g < gens; g++) {
1448 generations[g].max_blocks = size;
1453 /* -----------------------------------------------------------------------------
1454 Calculate the new size of the nursery, and resize it.
1455 -------------------------------------------------------------------------- */
1458 resize_nursery (void)
1460 if (RtsFlags.GcFlags.generations == 1)
1461 { // Two-space collector:
1464 /* set up a new nursery. Allocate a nursery size based on a
1465 * function of the amount of live data (by default a factor of 2)
1466 * Use the blocks from the old nursery if possible, freeing up any
1469 * If we get near the maximum heap size, then adjust our nursery
1470 * size accordingly. If the nursery is the same size as the live
1471 * data (L), then we need 3L bytes. We can reduce the size of the
1472 * nursery to bring the required memory down near 2L bytes.
1474 * A normal 2-space collector would need 4L bytes to give the same
1475 * performance we get from 3L bytes, reducing to the same
1476 * performance at 2L bytes.
1478 blocks = g0s0->n_blocks;
1480 if ( RtsFlags.GcFlags.maxHeapSize != 0 &&
1481 blocks * RtsFlags.GcFlags.oldGenFactor * 2 >
1482 RtsFlags.GcFlags.maxHeapSize )
1484 long adjusted_blocks; // signed on purpose
1487 adjusted_blocks = (RtsFlags.GcFlags.maxHeapSize - 2 * blocks);
1489 debugTrace(DEBUG_gc, "near maximum heap size of 0x%x blocks, blocks = %d, adjusted to %ld",
1490 RtsFlags.GcFlags.maxHeapSize, blocks, adjusted_blocks);
1492 pc_free = adjusted_blocks * 100 / RtsFlags.GcFlags.maxHeapSize;
1493 if (pc_free < RtsFlags.GcFlags.pcFreeHeap) /* might even * be < 0 */
1497 blocks = adjusted_blocks;
1501 blocks *= RtsFlags.GcFlags.oldGenFactor;
1502 if (blocks < RtsFlags.GcFlags.minAllocAreaSize)
1504 blocks = RtsFlags.GcFlags.minAllocAreaSize;
1507 resizeNurseries(blocks);
1509 else // Generational collector
1512 * If the user has given us a suggested heap size, adjust our
1513 * allocation area to make best use of the memory available.
1515 if (RtsFlags.GcFlags.heapSizeSuggestion)
1518 nat needed = calcNeeded(); // approx blocks needed at next GC
1520 /* Guess how much will be live in generation 0 step 0 next time.
1521 * A good approximation is obtained by finding the
1522 * percentage of g0s0 that was live at the last minor GC.
1524 * We have an accurate figure for the amount of copied data in
1525 * 'copied', but we must convert this to a number of blocks, with
1526 * a small adjustment for estimated slop at the end of a block
1531 g0s0_pcnt_kept = ((copied / (BLOCK_SIZE_W - 10)) * 100)
1532 / countNurseryBlocks();
1535 /* Estimate a size for the allocation area based on the
1536 * information available. We might end up going slightly under
1537 * or over the suggested heap size, but we should be pretty
1540 * Formula: suggested - needed
1541 * ----------------------------
1542 * 1 + g0s0_pcnt_kept/100
1544 * where 'needed' is the amount of memory needed at the next
1545 * collection for collecting all steps except g0s0.
1548 (((long)RtsFlags.GcFlags.heapSizeSuggestion - (long)needed) * 100) /
1549 (100 + (long)g0s0_pcnt_kept);
1551 if (blocks < (long)RtsFlags.GcFlags.minAllocAreaSize) {
1552 blocks = RtsFlags.GcFlags.minAllocAreaSize;
1555 resizeNurseries((nat)blocks);
1559 // we might have added extra large blocks to the nursery, so
1560 // resize back to minAllocAreaSize again.
1561 resizeNurseriesFixed(RtsFlags.GcFlags.minAllocAreaSize);
1566 /* -----------------------------------------------------------------------------
1567 Sanity code for CAF garbage collection.
1569 With DEBUG turned on, we manage a CAF list in addition to the SRT
1570 mechanism. After GC, we run down the CAF list and blackhole any
1571 CAFs which have been garbage collected. This means we get an error
1572 whenever the program tries to enter a garbage collected CAF.
1574 Any garbage collected CAFs are taken off the CAF list at the same
1576 -------------------------------------------------------------------------- */
1578 #if 0 && defined(DEBUG)
1585 const StgInfoTable *info;
1596 ASSERT(info->type == IND_STATIC);
1598 if (STATIC_LINK(info,p) == NULL) {
1599 debugTrace(DEBUG_gccafs, "CAF gc'd at 0x%04lx", (long)p);
1601 SET_INFO(p,&stg_BLACKHOLE_info);
1602 p = STATIC_LINK2(info,p);
1606 pp = &STATIC_LINK2(info,p);
1613 debugTrace(DEBUG_gccafs, "%d CAFs live", i);