1 /* -----------------------------------------------------------------------------
3 * (c) The GHC Team 1998-2008
5 * Generational garbage collector
7 * Documentation on the architecture of the Garbage Collector can be
8 * found in the online commentary:
10 * http://hackage.haskell.org/trac/ghc/wiki/Commentary/Rts/Storage/GC
12 * ---------------------------------------------------------------------------*/
14 // #include "PosixSource.h"
19 #include "OSThreads.h"
20 #include "LdvProfile.h"
25 #include "BlockAlloc.h"
31 #include "ParTicky.h" // ToDo: move into Rts.h
32 #include "RtsSignals.h"
36 #if defined(RTS_GTK_FRONTPANEL)
37 #include "FrontPanel.h"
40 #include "RetainerProfile.h"
41 #include "RaiseAsync.h"
54 #include <string.h> // for memset()
57 /* -----------------------------------------------------------------------------
59 -------------------------------------------------------------------------- */
61 /* STATIC OBJECT LIST.
64 * We maintain a linked list of static objects that are still live.
65 * The requirements for this list are:
67 * - we need to scan the list while adding to it, in order to
68 * scavenge all the static objects (in the same way that
69 * breadth-first scavenging works for dynamic objects).
71 * - we need to be able to tell whether an object is already on
72 * the list, to break loops.
74 * Each static object has a "static link field", which we use for
75 * linking objects on to the list. We use a stack-type list, consing
76 * objects on the front as they are added (this means that the
77 * scavenge phase is depth-first, not breadth-first, but that
80 * A separate list is kept for objects that have been scavenged
81 * already - this is so that we can zero all the marks afterwards.
83 * An object is on the list if its static link field is non-zero; this
84 * means that we have to mark the end of the list with '1', not NULL.
86 * Extra notes for generational GC:
88 * Each generation has a static object list associated with it. When
89 * collecting generations up to N, we treat the static object lists
90 * from generations > N as roots.
92 * We build up a static object list while collecting generations 0..N,
93 * which is then appended to the static object list of generation N+1.
96 /* N is the oldest generation being collected, where the generations
97 * are numbered starting at 0. A major GC (indicated by the major_gc
98 * flag) is when we're collecting all generations. We only attempt to
99 * deal with static objects and GC CAFs when doing a major GC.
104 /* Data used for allocation area sizing.
106 static lnat g0s0_pcnt_kept = 30; // percentage of g0s0 live at last minor GC
116 /* Thread-local data for each GC thread
118 gc_thread **gc_threads = NULL;
119 // gc_thread *gct = NULL; // this thread's gct TODO: make thread-local
121 // Number of threads running in *this* GC. Affects how many
122 // step->todos[] lists we have to look in to find work.
126 long copied; // *words* copied & scavenged during this GC
129 SpinLock recordMutableGen_sync;
132 /* -----------------------------------------------------------------------------
133 Static function declarations
134 -------------------------------------------------------------------------- */
136 static void mark_root (void *user, StgClosure **root);
137 static void zero_static_object_list (StgClosure* first_static);
138 static nat initialise_N (rtsBool force_major_gc);
139 static void alloc_gc_threads (void);
140 static void init_collected_gen (nat g, nat threads);
141 static void init_uncollected_gen (nat g, nat threads);
142 static void init_gc_thread (gc_thread *t);
143 static void update_task_list (void);
144 static void resize_generations (void);
145 static void resize_nursery (void);
146 static void start_gc_threads (void);
147 static void 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 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): %dKB to collect, using %d thread(s)",
257 N, n * (BLOCK_SIZE / 1024), 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.
354 // The other threads are now stopped. We might recurse back to
355 // here, but from now on this is the only thread.
357 // if any blackholes are alive, make the threads that wait on
359 if (traverseBlackholeQueue()) {
364 // must be last... invariant is that everything is fully
365 // scavenged at this point.
366 if (traverseWeakPtrList()) { // returns rtsTrue if evaced something
371 // If we get to here, there's really nothing left to do.
375 shutdown_gc_threads(n_gc_threads);
377 // Update pointers from the Task list
380 // Now see which stable names are still alive.
384 // We call processHeapClosureForDead() on every closure destroyed during
385 // the current garbage collection, so we invoke LdvCensusForDead().
386 if (RtsFlags.ProfFlags.doHeapProfile == HEAP_BY_LDV
387 || RtsFlags.ProfFlags.bioSelector != NULL)
391 // NO MORE EVACUATION AFTER THIS POINT!
392 // Finally: compaction of the oldest generation.
393 if (major_gc && oldest_gen->steps[0].is_compacted) {
394 // save number of blocks for stats
395 oldgen_saved_blocks = oldest_gen->steps[0].n_old_blocks;
396 compact(gct->scavenged_static_objects);
399 IF_DEBUG(sanity, checkGlobalTSOList(rtsFalse));
401 // Two-space collector: free the old to-space.
402 // g0s0->old_blocks is the old nursery
403 // g0s0->blocks is to-space from the previous GC
404 if (RtsFlags.GcFlags.generations == 1) {
405 if (g0s0->blocks != NULL) {
406 freeChain(g0s0->blocks);
411 // For each workspace, in each thread:
412 // * clear the BF_EVACUATED flag from each copied block
413 // * move the copied blocks to the step
419 for (t = 0; t < n_gc_threads; t++) {
423 if (RtsFlags.GcFlags.generations == 1) {
428 for (; s < total_steps; s++) {
431 // Push the final block
433 push_scanned_block(ws->todo_bd, ws);
436 ASSERT(gct->scan_bd == NULL);
437 ASSERT(countBlocks(ws->scavd_list) == ws->n_scavd_blocks);
440 for (bd = ws->scavd_list; bd != NULL; bd = bd->link) {
441 bd->flags &= ~BF_EVACUATED; // now from-space
442 ws->step->n_words += bd->free - bd->start;
446 prev->link = ws->step->blocks;
447 ws->step->blocks = ws->scavd_list;
449 ws->step->n_blocks += ws->n_scavd_blocks;
452 for (bd = ws->part_list; bd != NULL; bd = next) {
454 if (bd->free == bd->start) {
456 ws->part_list = next;
463 bd->flags &= ~BF_EVACUATED; // now from-space
464 ws->step->n_words += bd->free - bd->start;
469 prev->link = ws->step->blocks;
470 ws->step->blocks = ws->part_list;
472 ws->step->n_blocks += ws->n_part_blocks;
474 ASSERT(countBlocks(ws->step->blocks) == ws->step->n_blocks);
475 ASSERT(countOccupied(ws->step->blocks) == ws->step->n_words);
480 /* run through all the generations/steps and tidy up
487 for (i=0; i < n_gc_threads; i++) {
488 if (n_gc_threads > 1) {
489 trace(TRACE_gc,"thread %d:", i);
490 trace(TRACE_gc," copied %ld", gc_threads[i]->copied * sizeof(W_));
491 trace(TRACE_gc," scanned %ld", gc_threads[i]->scanned * sizeof(W_));
492 trace(TRACE_gc," any_work %ld", gc_threads[i]->any_work);
493 trace(TRACE_gc," no_work %ld", gc_threads[i]->no_work);
494 trace(TRACE_gc," scav_find_work %ld", gc_threads[i]->scav_find_work);
496 copied += gc_threads[i]->copied;
497 max_copied = stg_max(gc_threads[i]->copied, max_copied);
499 if (n_gc_threads == 1) {
507 for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
510 generations[g].collections++; // for stats
511 if (n_gc_threads > 1) generations[g].par_collections++;
514 // Count the mutable list as bytes "copied" for the purposes of
515 // stats. Every mutable list is copied during every GC.
517 nat mut_list_size = 0;
518 for (bd = generations[g].mut_list; bd != NULL; bd = bd->link) {
519 mut_list_size += bd->free - bd->start;
521 copied += mut_list_size;
524 "mut_list_size: %lu (%d vars, %d arrays, %d MVARs, %d others)",
525 (unsigned long)(mut_list_size * sizeof(W_)),
526 mutlist_MUTVARS, mutlist_MUTARRS, mutlist_MVARS, mutlist_OTHERS);
529 for (s = 0; s < generations[g].n_steps; s++) {
531 stp = &generations[g].steps[s];
533 // for generations we collected...
536 /* free old memory and shift to-space into from-space for all
537 * the collected steps (except the allocation area). These
538 * freed blocks will probaby be quickly recycled.
540 if (!(g == 0 && s == 0 && RtsFlags.GcFlags.generations > 1)) {
541 if (stp->is_compacted)
543 // for a compacted step, just shift the new to-space
544 // onto the front of the now-compacted existing blocks.
545 for (bd = stp->blocks; bd != NULL; bd = bd->link) {
546 bd->flags &= ~BF_EVACUATED; // now from-space
547 stp->n_words += bd->free - bd->start;
549 // tack the new blocks on the end of the existing blocks
550 if (stp->old_blocks != NULL) {
551 for (bd = stp->old_blocks; bd != NULL; bd = next) {
552 // NB. this step might not be compacted next
553 // time, so reset the BF_COMPACTED flags.
554 // They are set before GC if we're going to
555 // compact. (search for BF_COMPACTED above).
556 bd->flags &= ~BF_COMPACTED;
559 bd->link = stp->blocks;
562 stp->blocks = stp->old_blocks;
564 // add the new blocks to the block tally
565 stp->n_blocks += stp->n_old_blocks;
566 ASSERT(countBlocks(stp->blocks) == stp->n_blocks);
567 ASSERT(countOccupied(stp->blocks) == stp->n_words);
571 freeChain(stp->old_blocks);
573 stp->old_blocks = NULL;
574 stp->n_old_blocks = 0;
577 /* LARGE OBJECTS. The current live large objects are chained on
578 * scavenged_large, having been moved during garbage
579 * collection from large_objects. Any objects left on
580 * large_objects list are therefore dead, so we free them here.
582 for (bd = stp->large_objects; bd != NULL; bd = next) {
588 // update the count of blocks used by large objects
589 for (bd = stp->scavenged_large_objects; bd != NULL; bd = bd->link) {
590 bd->flags &= ~BF_EVACUATED;
592 stp->large_objects = stp->scavenged_large_objects;
593 stp->n_large_blocks = stp->n_scavenged_large_blocks;
596 else // for older generations...
598 /* For older generations, we need to append the
599 * scavenged_large_object list (i.e. large objects that have been
600 * promoted during this GC) to the large_object list for that step.
602 for (bd = stp->scavenged_large_objects; bd; bd = next) {
604 bd->flags &= ~BF_EVACUATED;
605 dbl_link_onto(bd, &stp->large_objects);
608 // add the new blocks we promoted during this GC
609 stp->n_large_blocks += stp->n_scavenged_large_blocks;
614 // update the max size of older generations after a major GC
615 resize_generations();
617 // Calculate the amount of live data for stats.
618 live = calcLiveWords();
620 // Free the small objects allocated via allocate(), since this will
621 // all have been copied into G0S1 now.
622 if (RtsFlags.GcFlags.generations > 1) {
623 if (g0s0->blocks != NULL) {
624 freeChain(g0s0->blocks);
631 alloc_blocks_lim = RtsFlags.GcFlags.minAllocAreaSize;
633 // Start a new pinned_object_block
634 pinned_object_block = NULL;
636 // Free the mark stack.
637 if (mark_stack_bdescr != NULL) {
638 freeGroup(mark_stack_bdescr);
642 for (g = 0; g <= N; g++) {
643 for (s = 0; s < generations[g].n_steps; s++) {
644 stp = &generations[g].steps[s];
645 if (stp->bitmap != NULL) {
646 freeGroup(stp->bitmap);
654 // mark the garbage collected CAFs as dead
655 #if 0 && defined(DEBUG) // doesn't work at the moment
656 if (major_gc) { gcCAFs(); }
660 // resetStaticObjectForRetainerProfiling() must be called before
662 if (n_gc_threads > 1) {
663 barf("profiling is currently broken with multi-threaded GC");
664 // ToDo: fix the gct->scavenged_static_objects below
666 resetStaticObjectForRetainerProfiling(gct->scavenged_static_objects);
669 // zero the scavenged static object list
672 for (i = 0; i < n_gc_threads; i++) {
673 zero_static_object_list(gc_threads[i]->scavenged_static_objects);
680 // start any pending finalizers
682 scheduleFinalizers(last_free_capability, old_weak_ptr_list);
685 // send exceptions to any threads which were about to die
687 resurrectThreads(resurrected_threads);
690 // Update the stable pointer hash table.
691 updateStablePtrTable(major_gc);
693 // check sanity after GC
694 IF_DEBUG(sanity, checkSanity());
696 // extra GC trace info
697 if (traceClass(TRACE_gc|DEBUG_gc)) statDescribeGens();
700 // symbol-table based profiling
701 /* heapCensus(to_blocks); */ /* ToDo */
704 // restore enclosing cost centre
710 // check for memory leaks if DEBUG is on
711 memInventory(traceClass(DEBUG_gc));
714 #ifdef RTS_GTK_FRONTPANEL
715 if (RtsFlags.GcFlags.frontpanel) {
716 updateFrontPanelAfterGC( N, live );
720 // ok, GC over: tell the stats department what happened.
721 slop = calcLiveBlocks() * BLOCK_SIZE_W - live;
722 stat_endGC(allocated, live, copied, N, max_copied, avg_copied, slop);
724 #if defined(RTS_USER_SIGNALS)
725 if (RtsFlags.MiscFlags.install_signal_handlers) {
726 // unblock signals again
727 unblockUserSignals();
736 /* -----------------------------------------------------------------------------
737 Figure out which generation to collect, initialise N and major_gc.
739 Also returns the total number of blocks in generations that will be
741 -------------------------------------------------------------------------- */
744 initialise_N (rtsBool force_major_gc)
747 nat s, blocks, blocks_total;
752 if (force_major_gc) {
753 N = RtsFlags.GcFlags.generations - 1;
758 for (g = RtsFlags.GcFlags.generations - 1; g >= 0; g--) {
760 for (s = 0; s < generations[g].n_steps; s++) {
761 blocks += generations[g].steps[s].n_words / BLOCK_SIZE_W;
762 blocks += generations[g].steps[s].n_large_blocks;
764 if (blocks >= generations[g].max_blocks) {
768 blocks_total += blocks;
772 blocks_total += countNurseryBlocks();
774 major_gc = (N == RtsFlags.GcFlags.generations-1);
778 /* -----------------------------------------------------------------------------
779 Initialise the gc_thread structures.
780 -------------------------------------------------------------------------- */
783 alloc_gc_thread (int n)
789 t = stgMallocBytes(sizeof(gc_thread) + total_steps * sizeof(step_workspace),
794 initCondition(&t->wake_cond);
795 initMutex(&t->wake_mutex);
796 t->wakeup = rtsTrue; // starts true, so we can wait for the
797 // thread to start up, see wakeup_gc_threads
802 t->free_blocks = NULL;
811 for (s = 0; s < total_steps; s++)
814 ws->step = &all_steps[s];
815 ASSERT(s == ws->step->abs_no);
819 ws->buffer_todo_bd = NULL;
821 ws->part_list = NULL;
822 ws->n_part_blocks = 0;
824 ws->scavd_list = NULL;
825 ws->n_scavd_blocks = 0;
833 alloc_gc_threads (void)
835 if (gc_threads == NULL) {
836 #if defined(THREADED_RTS)
838 gc_threads = stgMallocBytes (RtsFlags.ParFlags.gcThreads *
842 for (i = 0; i < RtsFlags.ParFlags.gcThreads; i++) {
843 gc_threads[i] = alloc_gc_thread(i);
846 gc_threads = stgMallocBytes (sizeof(gc_thread*),
849 gc_threads[0] = alloc_gc_thread(0);
854 /* ----------------------------------------------------------------------------
856 ------------------------------------------------------------------------- */
858 static nat gc_running_threads;
860 #if defined(THREADED_RTS)
861 static Mutex gc_running_mutex;
868 ACQUIRE_LOCK(&gc_running_mutex);
869 n_running = ++gc_running_threads;
870 RELEASE_LOCK(&gc_running_mutex);
871 ASSERT(n_running <= n_gc_threads);
879 ACQUIRE_LOCK(&gc_running_mutex);
880 ASSERT(n_gc_threads != 0);
881 n_running = --gc_running_threads;
882 RELEASE_LOCK(&gc_running_mutex);
887 // gc_thread_work(): Scavenge until there's no work left to do and all
888 // the running threads are idle.
891 gc_thread_work (void)
895 debugTrace(DEBUG_gc, "GC thread %d working", gct->thread_index);
897 // gc_running_threads has already been incremented for us; either
898 // this is the main thread and we incremented it inside
899 // GarbageCollect(), or this is a worker thread and the main
900 // thread bumped gc_running_threads before waking us up.
902 // Every thread evacuates some roots.
904 markSomeCapabilities(mark_root, gct, gct->thread_index, n_gc_threads);
908 // scavenge_loop() only exits when there's no work to do
911 debugTrace(DEBUG_gc, "GC thread %d idle (%d still running)",
912 gct->thread_index, r);
914 while (gc_running_threads != 0) {
920 // any_work() does not remove the work from the queue, it
921 // just checks for the presence of work. If we find any,
922 // then we increment gc_running_threads and go back to
923 // scavenge_loop() to perform any pending work.
926 // All threads are now stopped
927 debugTrace(DEBUG_gc, "GC thread %d finished.", gct->thread_index);
931 #if defined(THREADED_RTS)
933 gc_thread_mainloop (void)
937 // Wait until we're told to wake up
938 ACQUIRE_LOCK(&gct->wake_mutex);
939 gct->wakeup = rtsFalse;
940 while (!gct->wakeup) {
941 debugTrace(DEBUG_gc, "GC thread %d standing by...",
943 waitCondition(&gct->wake_cond, &gct->wake_mutex);
945 RELEASE_LOCK(&gct->wake_mutex);
946 if (gct->exit) break;
949 // start performance counters in this thread...
950 if (gct->papi_events == -1) {
951 papi_init_eventset(&gct->papi_events);
953 papi_thread_start_gc1_count(gct->papi_events);
959 // count events in this thread towards the GC totals
960 papi_thread_stop_gc1_count(gct->papi_events);
966 #if defined(THREADED_RTS)
968 gc_thread_entry (gc_thread *my_gct)
971 debugTrace(DEBUG_gc, "GC thread %d starting...", gct->thread_index);
972 gct->id = osThreadId();
973 gc_thread_mainloop();
978 start_gc_threads (void)
980 #if defined(THREADED_RTS)
983 static rtsBool done = rtsFalse;
985 gc_running_threads = 0;
986 initMutex(&gc_running_mutex);
989 // Start from 1: the main thread is 0
990 for (i = 1; i < RtsFlags.ParFlags.gcThreads; i++) {
991 createOSThread(&id, (OSThreadProc*)&gc_thread_entry,
1000 wakeup_gc_threads (nat n_threads USED_IF_THREADS)
1002 #if defined(THREADED_RTS)
1004 for (i=1; i < n_threads; i++) {
1006 debugTrace(DEBUG_gc, "waking up gc thread %d", i);
1008 ACQUIRE_LOCK(&gc_threads[i]->wake_mutex);
1009 if (gc_threads[i]->wakeup) {
1010 RELEASE_LOCK(&gc_threads[i]->wake_mutex);
1016 gc_threads[i]->wakeup = rtsTrue;
1017 signalCondition(&gc_threads[i]->wake_cond);
1018 RELEASE_LOCK(&gc_threads[i]->wake_mutex);
1023 // After GC is complete, we must wait for all GC threads to enter the
1024 // standby state, otherwise they may still be executing inside
1025 // any_work(), and may even remain awake until the next GC starts.
1027 shutdown_gc_threads (nat n_threads USED_IF_THREADS)
1029 #if defined(THREADED_RTS)
1032 for (i=1; i < n_threads; i++) {
1034 ACQUIRE_LOCK(&gc_threads[i]->wake_mutex);
1035 wakeup = gc_threads[i]->wakeup;
1036 // wakeup is false while the thread is waiting
1037 RELEASE_LOCK(&gc_threads[i]->wake_mutex);
1043 /* ----------------------------------------------------------------------------
1044 Initialise a generation that is to be collected
1045 ------------------------------------------------------------------------- */
1048 init_collected_gen (nat g, nat n_threads)
1055 // Throw away the current mutable list. Invariant: the mutable
1056 // list always has at least one block; this means we can avoid a
1057 // check for NULL in recordMutable().
1059 freeChain(generations[g].mut_list);
1060 generations[g].mut_list = allocBlock();
1061 for (i = 0; i < n_capabilities; i++) {
1062 freeChain(capabilities[i].mut_lists[g]);
1063 capabilities[i].mut_lists[g] = allocBlock();
1067 for (s = 0; s < generations[g].n_steps; s++) {
1069 // generation 0, step 0 doesn't need to-space
1070 if (g == 0 && s == 0 && RtsFlags.GcFlags.generations > 1) {
1074 stp = &generations[g].steps[s];
1075 ASSERT(stp->gen_no == g);
1077 // deprecate the existing blocks
1078 stp->old_blocks = stp->blocks;
1079 stp->n_old_blocks = stp->n_blocks;
1084 // we don't have any to-be-scavenged blocks yet
1086 stp->todos_last = NULL;
1089 // initialise the large object queues.
1090 stp->scavenged_large_objects = NULL;
1091 stp->n_scavenged_large_blocks = 0;
1093 // mark the large objects as not evacuated yet
1094 for (bd = stp->large_objects; bd; bd = bd->link) {
1095 bd->flags &= ~BF_EVACUATED;
1098 // for a compacted step, we need to allocate the bitmap
1099 if (stp->is_compacted) {
1100 nat bitmap_size; // in bytes
1101 bdescr *bitmap_bdescr;
1104 bitmap_size = stp->n_old_blocks * BLOCK_SIZE / (sizeof(W_)*BITS_PER_BYTE);
1106 if (bitmap_size > 0) {
1107 bitmap_bdescr = allocGroup((lnat)BLOCK_ROUND_UP(bitmap_size)
1109 stp->bitmap = bitmap_bdescr;
1110 bitmap = bitmap_bdescr->start;
1112 debugTrace(DEBUG_gc, "bitmap_size: %d, bitmap: %p",
1113 bitmap_size, bitmap);
1115 // don't forget to fill it with zeros!
1116 memset(bitmap, 0, bitmap_size);
1118 // For each block in this step, point to its bitmap from the
1119 // block descriptor.
1120 for (bd=stp->old_blocks; bd != NULL; bd = bd->link) {
1121 bd->u.bitmap = bitmap;
1122 bitmap += BLOCK_SIZE_W / (sizeof(W_)*BITS_PER_BYTE);
1124 // Also at this point we set the BF_COMPACTED flag
1125 // for this block. The invariant is that
1126 // BF_COMPACTED is always unset, except during GC
1127 // when it is set on those blocks which will be
1129 bd->flags |= BF_COMPACTED;
1135 // For each GC thread, for each step, allocate a "todo" block to
1136 // store evacuated objects to be scavenged, and a block to store
1137 // evacuated objects that do not need to be scavenged.
1138 for (t = 0; t < n_threads; t++) {
1139 for (s = 0; s < generations[g].n_steps; s++) {
1141 // we don't copy objects into g0s0, unless -G0
1142 if (g==0 && s==0 && RtsFlags.GcFlags.generations > 1) continue;
1144 ws = &gc_threads[t]->steps[g * RtsFlags.GcFlags.steps + s];
1146 ws->todo_large_objects = NULL;
1148 ws->part_list = NULL;
1149 ws->n_part_blocks = 0;
1151 // allocate the first to-space block; extra blocks will be
1152 // chained on as necessary.
1154 ws->buffer_todo_bd = NULL;
1155 alloc_todo_block(ws,0);
1157 ws->scavd_list = NULL;
1158 ws->n_scavd_blocks = 0;
1164 /* ----------------------------------------------------------------------------
1165 Initialise a generation that is *not* to be collected
1166 ------------------------------------------------------------------------- */
1169 init_uncollected_gen (nat g, nat threads)
1176 for (s = 0; s < generations[g].n_steps; s++) {
1177 stp = &generations[g].steps[s];
1178 stp->scavenged_large_objects = NULL;
1179 stp->n_scavenged_large_blocks = 0;
1182 for (t = 0; t < threads; t++) {
1183 for (s = 0; s < generations[g].n_steps; s++) {
1185 ws = &gc_threads[t]->steps[g * RtsFlags.GcFlags.steps + s];
1188 ws->buffer_todo_bd = NULL;
1189 ws->todo_large_objects = NULL;
1191 ws->part_list = NULL;
1192 ws->n_part_blocks = 0;
1194 ws->scavd_list = NULL;
1195 ws->n_scavd_blocks = 0;
1197 // If the block at the head of the list in this generation
1198 // is less than 3/4 full, then use it as a todo block.
1199 if (stp->blocks && isPartiallyFull(stp->blocks))
1201 ws->todo_bd = stp->blocks;
1202 ws->todo_free = ws->todo_bd->free;
1203 ws->todo_lim = ws->todo_bd->start + BLOCK_SIZE_W;
1204 stp->blocks = stp->blocks->link;
1206 stp->n_words -= ws->todo_bd->free - ws->todo_bd->start;
1207 ws->todo_bd->link = NULL;
1208 // we must scan from the current end point.
1209 ws->todo_bd->u.scan = ws->todo_bd->free;
1214 alloc_todo_block(ws,0);
1219 // Move the private mutable lists from each capability onto the
1220 // main mutable list for the generation.
1221 for (i = 0; i < n_capabilities; i++) {
1222 for (bd = capabilities[i].mut_lists[g];
1223 bd->link != NULL; bd = bd->link) {
1226 bd->link = generations[g].mut_list;
1227 generations[g].mut_list = capabilities[i].mut_lists[g];
1228 capabilities[i].mut_lists[g] = allocBlock();
1232 /* -----------------------------------------------------------------------------
1233 Initialise a gc_thread before GC
1234 -------------------------------------------------------------------------- */
1237 init_gc_thread (gc_thread *t)
1239 t->static_objects = END_OF_STATIC_LIST;
1240 t->scavenged_static_objects = END_OF_STATIC_LIST;
1243 t->failed_to_evac = rtsFalse;
1244 t->eager_promotion = rtsTrue;
1245 t->thunk_selector_depth = 0;
1250 t->scav_find_work = 0;
1253 /* -----------------------------------------------------------------------------
1254 Function we pass to evacuate roots.
1255 -------------------------------------------------------------------------- */
1258 mark_root(void *user, StgClosure **root)
1260 // we stole a register for gct, but this function is called from
1261 // *outside* the GC where the register variable is not in effect,
1262 // so we need to save and restore it here. NB. only call
1263 // mark_root() from the main GC thread, otherwise gct will be
1265 gc_thread *saved_gct;
1274 /* -----------------------------------------------------------------------------
1275 Initialising the static object & mutable lists
1276 -------------------------------------------------------------------------- */
1279 zero_static_object_list(StgClosure* first_static)
1283 const StgInfoTable *info;
1285 for (p = first_static; p != END_OF_STATIC_LIST; p = link) {
1287 link = *STATIC_LINK(info, p);
1288 *STATIC_LINK(info,p) = NULL;
1292 /* ----------------------------------------------------------------------------
1293 Update the pointers from the task list
1295 These are treated as weak pointers because we want to allow a main
1296 thread to get a BlockedOnDeadMVar exception in the same way as any
1297 other thread. Note that the threads should all have been retained
1298 by GC by virtue of being on the all_threads list, we're just
1299 updating pointers here.
1300 ------------------------------------------------------------------------- */
1303 update_task_list (void)
1307 for (task = all_tasks; task != NULL; task = task->all_link) {
1308 if (!task->stopped && task->tso) {
1309 ASSERT(task->tso->bound == task);
1310 tso = (StgTSO *) isAlive((StgClosure *)task->tso);
1312 barf("task %p: main thread %d has been GC'd",
1325 /* ----------------------------------------------------------------------------
1326 Reset the sizes of the older generations when we do a major
1329 CURRENT STRATEGY: make all generations except zero the same size.
1330 We have to stay within the maximum heap size, and leave a certain
1331 percentage of the maximum heap size available to allocate into.
1332 ------------------------------------------------------------------------- */
1335 resize_generations (void)
1339 if (major_gc && RtsFlags.GcFlags.generations > 1) {
1340 nat live, size, min_alloc;
1341 nat max = RtsFlags.GcFlags.maxHeapSize;
1342 nat gens = RtsFlags.GcFlags.generations;
1344 // live in the oldest generations
1345 live = (oldest_gen->steps[0].n_words + BLOCK_SIZE_W - 1) / BLOCK_SIZE_W+
1346 oldest_gen->steps[0].n_large_blocks;
1348 // default max size for all generations except zero
1349 size = stg_max(live * RtsFlags.GcFlags.oldGenFactor,
1350 RtsFlags.GcFlags.minOldGenSize);
1352 // minimum size for generation zero
1353 min_alloc = stg_max((RtsFlags.GcFlags.pcFreeHeap * max) / 200,
1354 RtsFlags.GcFlags.minAllocAreaSize);
1356 // Auto-enable compaction when the residency reaches a
1357 // certain percentage of the maximum heap size (default: 30%).
1358 if (RtsFlags.GcFlags.generations > 1 &&
1359 (RtsFlags.GcFlags.compact ||
1361 oldest_gen->steps[0].n_blocks >
1362 (RtsFlags.GcFlags.compactThreshold * max) / 100))) {
1363 oldest_gen->steps[0].is_compacted = 1;
1364 // debugBelch("compaction: on\n", live);
1366 oldest_gen->steps[0].is_compacted = 0;
1367 // debugBelch("compaction: off\n", live);
1370 // if we're going to go over the maximum heap size, reduce the
1371 // size of the generations accordingly. The calculation is
1372 // different if compaction is turned on, because we don't need
1373 // to double the space required to collect the old generation.
1376 // this test is necessary to ensure that the calculations
1377 // below don't have any negative results - we're working
1378 // with unsigned values here.
1379 if (max < min_alloc) {
1383 if (oldest_gen->steps[0].is_compacted) {
1384 if ( (size + (size - 1) * (gens - 2) * 2) + min_alloc > max ) {
1385 size = (max - min_alloc) / ((gens - 1) * 2 - 1);
1388 if ( (size * (gens - 1) * 2) + min_alloc > max ) {
1389 size = (max - min_alloc) / ((gens - 1) * 2);
1399 debugBelch("live: %d, min_alloc: %d, size : %d, max = %d\n", live,
1400 min_alloc, size, max);
1403 for (g = 0; g < gens; g++) {
1404 generations[g].max_blocks = size;
1409 /* -----------------------------------------------------------------------------
1410 Calculate the new size of the nursery, and resize it.
1411 -------------------------------------------------------------------------- */
1414 resize_nursery (void)
1416 if (RtsFlags.GcFlags.generations == 1)
1417 { // Two-space collector:
1420 /* set up a new nursery. Allocate a nursery size based on a
1421 * function of the amount of live data (by default a factor of 2)
1422 * Use the blocks from the old nursery if possible, freeing up any
1425 * If we get near the maximum heap size, then adjust our nursery
1426 * size accordingly. If the nursery is the same size as the live
1427 * data (L), then we need 3L bytes. We can reduce the size of the
1428 * nursery to bring the required memory down near 2L bytes.
1430 * A normal 2-space collector would need 4L bytes to give the same
1431 * performance we get from 3L bytes, reducing to the same
1432 * performance at 2L bytes.
1434 blocks = g0s0->n_blocks;
1436 if ( RtsFlags.GcFlags.maxHeapSize != 0 &&
1437 blocks * RtsFlags.GcFlags.oldGenFactor * 2 >
1438 RtsFlags.GcFlags.maxHeapSize )
1440 long adjusted_blocks; // signed on purpose
1443 adjusted_blocks = (RtsFlags.GcFlags.maxHeapSize - 2 * blocks);
1445 debugTrace(DEBUG_gc, "near maximum heap size of 0x%x blocks, blocks = %d, adjusted to %ld",
1446 RtsFlags.GcFlags.maxHeapSize, blocks, adjusted_blocks);
1448 pc_free = adjusted_blocks * 100 / RtsFlags.GcFlags.maxHeapSize;
1449 if (pc_free < RtsFlags.GcFlags.pcFreeHeap) /* might even * be < 0 */
1453 blocks = adjusted_blocks;
1457 blocks *= RtsFlags.GcFlags.oldGenFactor;
1458 if (blocks < RtsFlags.GcFlags.minAllocAreaSize)
1460 blocks = RtsFlags.GcFlags.minAllocAreaSize;
1463 resizeNurseries(blocks);
1465 else // Generational collector
1468 * If the user has given us a suggested heap size, adjust our
1469 * allocation area to make best use of the memory available.
1471 if (RtsFlags.GcFlags.heapSizeSuggestion)
1474 nat needed = calcNeeded(); // approx blocks needed at next GC
1476 /* Guess how much will be live in generation 0 step 0 next time.
1477 * A good approximation is obtained by finding the
1478 * percentage of g0s0 that was live at the last minor GC.
1480 * We have an accurate figure for the amount of copied data in
1481 * 'copied', but we must convert this to a number of blocks, with
1482 * a small adjustment for estimated slop at the end of a block
1487 g0s0_pcnt_kept = ((copied / (BLOCK_SIZE_W - 10)) * 100)
1488 / countNurseryBlocks();
1491 /* Estimate a size for the allocation area based on the
1492 * information available. We might end up going slightly under
1493 * or over the suggested heap size, but we should be pretty
1496 * Formula: suggested - needed
1497 * ----------------------------
1498 * 1 + g0s0_pcnt_kept/100
1500 * where 'needed' is the amount of memory needed at the next
1501 * collection for collecting all steps except g0s0.
1504 (((long)RtsFlags.GcFlags.heapSizeSuggestion - (long)needed) * 100) /
1505 (100 + (long)g0s0_pcnt_kept);
1507 if (blocks < (long)RtsFlags.GcFlags.minAllocAreaSize) {
1508 blocks = RtsFlags.GcFlags.minAllocAreaSize;
1511 resizeNurseries((nat)blocks);
1515 // we might have added extra large blocks to the nursery, so
1516 // resize back to minAllocAreaSize again.
1517 resizeNurseriesFixed(RtsFlags.GcFlags.minAllocAreaSize);
1522 /* -----------------------------------------------------------------------------
1523 Sanity code for CAF garbage collection.
1525 With DEBUG turned on, we manage a CAF list in addition to the SRT
1526 mechanism. After GC, we run down the CAF list and blackhole any
1527 CAFs which have been garbage collected. This means we get an error
1528 whenever the program tries to enter a garbage collected CAF.
1530 Any garbage collected CAFs are taken off the CAF list at the same
1532 -------------------------------------------------------------------------- */
1534 #if 0 && defined(DEBUG)
1541 const StgInfoTable *info;
1552 ASSERT(info->type == IND_STATIC);
1554 if (STATIC_LINK(info,p) == NULL) {
1555 debugTrace(DEBUG_gccafs, "CAF gc'd at 0x%04lx", (long)p);
1557 SET_INFO(p,&stg_BLACKHOLE_info);
1558 p = STATIC_LINK2(info,p);
1562 pp = &STATIC_LINK2(info,p);
1569 debugTrace(DEBUG_gccafs, "%d CAFs live", i);