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;
134 /* -----------------------------------------------------------------------------
135 Static function declarations
136 -------------------------------------------------------------------------- */
138 static void mark_root (void *user, StgClosure **root);
139 static void zero_static_object_list (StgClosure* first_static);
140 static nat initialise_N (rtsBool force_major_gc);
141 static void alloc_gc_threads (void);
142 static void init_collected_gen (nat g, nat threads);
143 static void init_uncollected_gen (nat g, nat threads);
144 static void init_gc_thread (gc_thread *t);
145 static void update_task_list (void);
146 static void resize_generations (void);
147 static void resize_nursery (void);
148 static void start_gc_threads (void);
149 static void scavenge_until_all_done (void);
150 static nat inc_running (void);
151 static nat dec_running (void);
152 static void wakeup_gc_threads (nat n_threads);
153 static void shutdown_gc_threads (nat n_threads);
155 #if 0 && defined(DEBUG)
156 static void gcCAFs (void);
159 /* -----------------------------------------------------------------------------
160 The mark bitmap & stack.
161 -------------------------------------------------------------------------- */
163 #define MARK_STACK_BLOCKS 4
165 bdescr *mark_stack_bdescr;
170 // Flag and pointers used for falling back to a linear scan when the
171 // mark stack overflows.
172 rtsBool mark_stack_overflowed;
173 bdescr *oldgen_scan_bd;
176 /* -----------------------------------------------------------------------------
177 GarbageCollect: the main entry point to the garbage collector.
179 Locks held: all capabilities are held throughout GarbageCollect().
180 -------------------------------------------------------------------------- */
183 GarbageCollect ( rtsBool force_major_gc )
187 lnat live, allocated, max_copied, avg_copied, slop;
188 gc_thread *saved_gct;
191 // necessary if we stole a callee-saves register for gct:
195 CostCentreStack *prev_CCS;
200 #if defined(RTS_USER_SIGNALS)
201 if (RtsFlags.MiscFlags.install_signal_handlers) {
207 ASSERT(sizeof(step_workspace) == 16 * sizeof(StgWord));
208 // otherwise adjust the padding in step_workspace.
210 // tell the stats department that we've started a GC
213 // tell the STM to discard any cached closures it's hoping to re-use
222 // attribute any costs to CCS_GC
228 /* Approximate how much we allocated.
229 * Todo: only when generating stats?
231 allocated = calcAllocated();
233 /* Figure out which generation to collect
235 n = initialise_N(force_major_gc);
237 /* Allocate + initialise the gc_thread structures.
241 /* Start threads, so they can be spinning up while we finish initialisation.
245 /* How many threads will be participating in this GC?
246 * We don't try to parallelise minor GC, or mark/compact/sweep GC.
248 #if defined(THREADED_RTS)
249 if (n < (4*1024*1024 / BLOCK_SIZE) || oldest_gen->steps[0].mark) {
252 n_gc_threads = RtsFlags.ParFlags.gcThreads;
257 trace(TRACE_gc|DEBUG_gc, "GC (gen %d): %d KB to collect, %ld MB in use, using %d thread(s)",
258 N, n * (BLOCK_SIZE / 1024), mblocks_allocated, n_gc_threads);
260 #ifdef RTS_GTK_FRONTPANEL
261 if (RtsFlags.GcFlags.frontpanel) {
262 updateFrontPanelBeforeGC(N);
267 // check for memory leaks if DEBUG is on
268 memInventory(traceClass(DEBUG_gc));
271 // check stack sanity *before* GC
272 IF_DEBUG(sanity, checkFreeListSanity());
273 IF_DEBUG(sanity, checkMutableLists());
275 // Initialise all our gc_thread structures
276 for (t = 0; t < n_gc_threads; t++) {
277 init_gc_thread(gc_threads[t]);
280 // Initialise all the generations/steps that we're collecting.
281 for (g = 0; g <= N; g++) {
282 init_collected_gen(g,n_gc_threads);
285 // Initialise all the generations/steps that we're *not* collecting.
286 for (g = N+1; g < RtsFlags.GcFlags.generations; g++) {
287 init_uncollected_gen(g,n_gc_threads);
290 /* Allocate a mark stack if we're doing a major collection.
293 nat mark_stack_blocks;
294 mark_stack_blocks = stg_max(MARK_STACK_BLOCKS,
295 oldest_gen->steps[0].n_old_blocks / 100);
296 mark_stack_bdescr = allocGroup(mark_stack_blocks);
297 mark_stack = (StgPtr *)mark_stack_bdescr->start;
298 mark_sp = mark_stack;
299 mark_splim = mark_stack + (mark_stack_blocks * BLOCK_SIZE_W);
301 mark_stack_bdescr = NULL;
304 // this is the main thread
307 /* -----------------------------------------------------------------------
308 * follow all the roots that we know about:
309 * - mutable lists from each generation > N
310 * we want to *scavenge* these roots, not evacuate them: they're not
311 * going to move in this GC.
312 * Also do them in reverse generation order, for the usual reason:
313 * namely to reduce the likelihood of spurious old->new pointers.
315 for (g = RtsFlags.GcFlags.generations-1; g > N; g--) {
316 generations[g].saved_mut_list = generations[g].mut_list;
317 generations[g].mut_list = allocBlock();
318 // mut_list always has at least one block.
321 // the main thread is running: this prevents any other threads from
322 // exiting prematurely, so we can start them now.
323 // NB. do this after the mutable lists have been saved above, otherwise
324 // the other GC threads will be writing into the old mutable lists.
326 wakeup_gc_threads(n_gc_threads);
328 for (g = RtsFlags.GcFlags.generations-1; g > N; g--) {
329 scavenge_mutable_list(&generations[g]);
332 // follow roots from the CAF list (used by GHCi)
334 markCAFs(mark_root, gct);
336 // follow all the roots that the application knows about.
338 markSomeCapabilities(mark_root, gct, gct->thread_index, n_gc_threads);
340 #if defined(RTS_USER_SIGNALS)
341 // mark the signal handlers (signals should be already blocked)
342 markSignalHandlers(mark_root, gct);
345 // Mark the weak pointer list, and prepare to detect dead weak pointers.
349 // Mark the stable pointer table.
350 markStablePtrTable(mark_root, gct);
352 /* -------------------------------------------------------------------------
353 * Repeatedly scavenge all the areas we know about until there's no
354 * more scavenging to be done.
358 scavenge_until_all_done();
359 // The other threads are now stopped. We might recurse back to
360 // here, but from now on this is the only thread.
362 // if any blackholes are alive, make the threads that wait on
364 if (traverseBlackholeQueue()) {
369 // must be last... invariant is that everything is fully
370 // scavenged at this point.
371 if (traverseWeakPtrList()) { // returns rtsTrue if evaced something
376 // If we get to here, there's really nothing left to do.
380 shutdown_gc_threads(n_gc_threads);
382 // Update pointers from the Task list
385 // Now see which stable names are still alive.
389 // We call processHeapClosureForDead() on every closure destroyed during
390 // the current garbage collection, so we invoke LdvCensusForDead().
391 if (RtsFlags.ProfFlags.doHeapProfile == HEAP_BY_LDV
392 || RtsFlags.ProfFlags.bioSelector != NULL)
396 // NO MORE EVACUATION AFTER THIS POINT!
398 // Two-space collector: free the old to-space.
399 // g0s0->old_blocks is the old nursery
400 // g0s0->blocks is to-space from the previous GC
401 if (RtsFlags.GcFlags.generations == 1) {
402 if (g0s0->blocks != NULL) {
403 freeChain(g0s0->blocks);
408 // For each workspace, in each thread, move the copied blocks to the step
414 for (t = 0; t < n_gc_threads; t++) {
418 if (RtsFlags.GcFlags.generations == 1) {
423 for (; s < total_steps; s++) {
426 // Push the final block
428 push_scanned_block(ws->todo_bd, ws);
431 ASSERT(gct->scan_bd == NULL);
432 ASSERT(countBlocks(ws->scavd_list) == ws->n_scavd_blocks);
435 for (bd = ws->scavd_list; bd != NULL; bd = bd->link) {
436 ws->step->n_words += bd->free - bd->start;
440 prev->link = ws->step->blocks;
441 ws->step->blocks = ws->scavd_list;
443 ws->step->n_blocks += ws->n_scavd_blocks;
447 // Add all the partial blocks *after* we've added all the full
448 // blocks. This is so that we can grab the partial blocks back
449 // again and try to fill them up in the next GC.
450 for (t = 0; t < n_gc_threads; t++) {
454 if (RtsFlags.GcFlags.generations == 1) {
459 for (; s < total_steps; s++) {
463 for (bd = ws->part_list; bd != NULL; bd = next) {
465 if (bd->free == bd->start) {
467 ws->part_list = next;
474 ws->step->n_words += bd->free - bd->start;
479 prev->link = ws->step->blocks;
480 ws->step->blocks = ws->part_list;
482 ws->step->n_blocks += ws->n_part_blocks;
484 ASSERT(countBlocks(ws->step->blocks) == ws->step->n_blocks);
485 ASSERT(countOccupied(ws->step->blocks) == ws->step->n_words);
490 // Finally: compact or sweep the oldest generation.
491 if (major_gc && oldest_gen->steps[0].mark) {
492 if (oldest_gen->steps[0].compact)
493 compact(gct->scavenged_static_objects);
495 sweep(&oldest_gen->steps[0]);
498 /* run through all the generations/steps and tidy up
505 for (i=0; i < n_gc_threads; i++) {
506 if (n_gc_threads > 1) {
507 trace(TRACE_gc,"thread %d:", i);
508 trace(TRACE_gc," copied %ld", gc_threads[i]->copied * sizeof(W_));
509 trace(TRACE_gc," scanned %ld", gc_threads[i]->scanned * sizeof(W_));
510 trace(TRACE_gc," any_work %ld", gc_threads[i]->any_work);
511 trace(TRACE_gc," no_work %ld", gc_threads[i]->no_work);
512 trace(TRACE_gc," scav_find_work %ld", gc_threads[i]->scav_find_work);
514 copied += gc_threads[i]->copied;
515 max_copied = stg_max(gc_threads[i]->copied, max_copied);
517 if (n_gc_threads == 1) {
525 for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
528 generations[g].collections++; // for stats
529 if (n_gc_threads > 1) generations[g].par_collections++;
532 // Count the mutable list as bytes "copied" for the purposes of
533 // stats. Every mutable list is copied during every GC.
535 nat mut_list_size = 0;
536 for (bd = generations[g].mut_list; bd != NULL; bd = bd->link) {
537 mut_list_size += bd->free - bd->start;
539 copied += mut_list_size;
542 "mut_list_size: %lu (%d vars, %d arrays, %d MVARs, %d others)",
543 (unsigned long)(mut_list_size * sizeof(W_)),
544 mutlist_MUTVARS, mutlist_MUTARRS, mutlist_MVARS, mutlist_OTHERS);
547 for (s = 0; s < generations[g].n_steps; s++) {
549 stp = &generations[g].steps[s];
551 // for generations we collected...
554 /* free old memory and shift to-space into from-space for all
555 * the collected steps (except the allocation area). These
556 * freed blocks will probaby be quickly recycled.
558 if (!(g == 0 && s == 0 && RtsFlags.GcFlags.generations > 1)) {
561 // tack the new blocks on the end of the existing blocks
562 if (stp->old_blocks != NULL) {
565 for (bd = stp->old_blocks; bd != NULL; bd = next) {
569 if (!(bd->flags & BF_MARKED))
572 stp->old_blocks = next;
581 stp->n_words += bd->free - bd->start;
583 // NB. this step might not be compacted next
584 // time, so reset the BF_MARKED flags.
585 // They are set before GC if we're going to
586 // compact. (search for BF_MARKED above).
587 bd->flags &= ~BF_MARKED;
589 // between GCs, all blocks in the heap except
590 // for the nursery have the BF_EVACUATED flag set.
591 bd->flags |= BF_EVACUATED;
598 prev->link = stp->blocks;
599 stp->blocks = stp->old_blocks;
602 // add the new blocks to the block tally
603 stp->n_blocks += stp->n_old_blocks;
604 ASSERT(countBlocks(stp->blocks) == stp->n_blocks);
605 ASSERT(countOccupied(stp->blocks) == stp->n_words);
609 freeChain(stp->old_blocks);
611 stp->old_blocks = NULL;
612 stp->n_old_blocks = 0;
615 /* LARGE OBJECTS. The current live large objects are chained on
616 * scavenged_large, having been moved during garbage
617 * collection from large_objects. Any objects left on
618 * large_objects list are therefore dead, so we free them here.
620 for (bd = stp->large_objects; bd != NULL; bd = next) {
626 stp->large_objects = stp->scavenged_large_objects;
627 stp->n_large_blocks = stp->n_scavenged_large_blocks;
630 else // for older generations...
632 /* For older generations, we need to append the
633 * scavenged_large_object list (i.e. large objects that have been
634 * promoted during this GC) to the large_object list for that step.
636 for (bd = stp->scavenged_large_objects; bd; bd = next) {
638 dbl_link_onto(bd, &stp->large_objects);
641 // add the new blocks we promoted during this GC
642 stp->n_large_blocks += stp->n_scavenged_large_blocks;
647 // update the max size of older generations after a major GC
648 resize_generations();
650 // Calculate the amount of live data for stats.
651 live = calcLiveWords();
653 // Free the small objects allocated via allocate(), since this will
654 // all have been copied into G0S1 now.
655 if (RtsFlags.GcFlags.generations > 1) {
656 if (g0s0->blocks != NULL) {
657 freeChain(g0s0->blocks);
664 alloc_blocks_lim = RtsFlags.GcFlags.minAllocAreaSize;
666 // Start a new pinned_object_block
667 pinned_object_block = NULL;
669 // Free the mark stack.
670 if (mark_stack_bdescr != NULL) {
671 freeGroup(mark_stack_bdescr);
675 for (g = 0; g <= N; g++) {
676 for (s = 0; s < generations[g].n_steps; s++) {
677 stp = &generations[g].steps[s];
678 if (stp->bitmap != NULL) {
679 freeGroup(stp->bitmap);
687 // mark the garbage collected CAFs as dead
688 #if 0 && defined(DEBUG) // doesn't work at the moment
689 if (major_gc) { gcCAFs(); }
693 // resetStaticObjectForRetainerProfiling() must be called before
695 if (n_gc_threads > 1) {
696 barf("profiling is currently broken with multi-threaded GC");
697 // ToDo: fix the gct->scavenged_static_objects below
699 resetStaticObjectForRetainerProfiling(gct->scavenged_static_objects);
702 // zero the scavenged static object list
705 for (i = 0; i < n_gc_threads; i++) {
706 zero_static_object_list(gc_threads[i]->scavenged_static_objects);
713 // start any pending finalizers
715 scheduleFinalizers(last_free_capability, old_weak_ptr_list);
718 // send exceptions to any threads which were about to die
720 resurrectThreads(resurrected_threads);
721 performPendingThrowTos(exception_threads);
724 // Update the stable pointer hash table.
725 updateStablePtrTable(major_gc);
727 // Remove useless sparks from the spark pools
732 // check sanity after GC
733 IF_DEBUG(sanity, checkSanity());
735 // extra GC trace info
736 if (traceClass(TRACE_gc|DEBUG_gc)) statDescribeGens();
739 // symbol-table based profiling
740 /* heapCensus(to_blocks); */ /* ToDo */
743 // restore enclosing cost centre
749 // check for memory leaks if DEBUG is on
750 memInventory(traceClass(DEBUG_gc));
753 #ifdef RTS_GTK_FRONTPANEL
754 if (RtsFlags.GcFlags.frontpanel) {
755 updateFrontPanelAfterGC( N, live );
759 // ok, GC over: tell the stats department what happened.
760 slop = calcLiveBlocks() * BLOCK_SIZE_W - live;
761 stat_endGC(allocated, live, copied, N, max_copied, avg_copied, slop);
763 #if defined(RTS_USER_SIGNALS)
764 if (RtsFlags.MiscFlags.install_signal_handlers) {
765 // unblock signals again
766 unblockUserSignals();
775 /* -----------------------------------------------------------------------------
776 Figure out which generation to collect, initialise N and major_gc.
778 Also returns the total number of blocks in generations that will be
780 -------------------------------------------------------------------------- */
783 initialise_N (rtsBool force_major_gc)
786 nat s, blocks, blocks_total;
791 if (force_major_gc) {
792 N = RtsFlags.GcFlags.generations - 1;
797 for (g = RtsFlags.GcFlags.generations - 1; g >= 0; g--) {
799 for (s = 0; s < generations[g].n_steps; s++) {
800 blocks += generations[g].steps[s].n_words / BLOCK_SIZE_W;
801 blocks += generations[g].steps[s].n_large_blocks;
803 if (blocks >= generations[g].max_blocks) {
807 blocks_total += blocks;
811 blocks_total += countNurseryBlocks();
813 major_gc = (N == RtsFlags.GcFlags.generations-1);
817 /* -----------------------------------------------------------------------------
818 Initialise the gc_thread structures.
819 -------------------------------------------------------------------------- */
822 alloc_gc_thread (int n)
828 t = stgMallocBytes(sizeof(gc_thread) + total_steps * sizeof(step_workspace),
833 initCondition(&t->wake_cond);
834 initMutex(&t->wake_mutex);
835 t->wakeup = rtsTrue; // starts true, so we can wait for the
836 // thread to start up, see wakeup_gc_threads
841 t->free_blocks = NULL;
850 for (s = 0; s < total_steps; s++)
853 ws->step = &all_steps[s];
854 ASSERT(s == ws->step->abs_no);
858 ws->buffer_todo_bd = NULL;
860 ws->part_list = NULL;
861 ws->n_part_blocks = 0;
863 ws->scavd_list = NULL;
864 ws->n_scavd_blocks = 0;
872 alloc_gc_threads (void)
874 if (gc_threads == NULL) {
875 #if defined(THREADED_RTS)
877 gc_threads = stgMallocBytes (RtsFlags.ParFlags.gcThreads *
881 for (i = 0; i < RtsFlags.ParFlags.gcThreads; i++) {
882 gc_threads[i] = alloc_gc_thread(i);
885 gc_threads = stgMallocBytes (sizeof(gc_thread*),
888 gc_threads[0] = alloc_gc_thread(0);
893 /* ----------------------------------------------------------------------------
895 ------------------------------------------------------------------------- */
897 static nat gc_running_threads;
899 #if defined(THREADED_RTS)
900 static Mutex gc_running_mutex;
907 ACQUIRE_LOCK(&gc_running_mutex);
908 n_running = ++gc_running_threads;
909 RELEASE_LOCK(&gc_running_mutex);
910 ASSERT(n_running <= n_gc_threads);
918 ACQUIRE_LOCK(&gc_running_mutex);
919 ASSERT(n_gc_threads != 0);
920 n_running = --gc_running_threads;
921 RELEASE_LOCK(&gc_running_mutex);
935 // scavenge objects in compacted generation
936 if (mark_stack_overflowed || oldgen_scan_bd != NULL ||
937 (mark_stack_bdescr != NULL && !mark_stack_empty())) {
941 // Check for global work in any step. We don't need to check for
942 // local work, because we have already exited scavenge_loop(),
943 // which means there is no local work for this thread.
944 for (s = total_steps-1; s >= 0; s--) {
945 if (s == 0 && RtsFlags.GcFlags.generations > 1) {
949 if (ws->todo_large_objects) return rtsTrue;
950 if (ws->step->todos) return rtsTrue;
959 scavenge_until_all_done (void)
963 debugTrace(DEBUG_gc, "GC thread %d working", gct->thread_index);
966 #if defined(THREADED_RTS)
967 if (n_gc_threads > 1) {
976 // scavenge_loop() only exits when there's no work to do
979 debugTrace(DEBUG_gc, "GC thread %d idle (%d still running)",
980 gct->thread_index, r);
982 while (gc_running_threads != 0) {
988 // any_work() does not remove the work from the queue, it
989 // just checks for the presence of work. If we find any,
990 // then we increment gc_running_threads and go back to
991 // scavenge_loop() to perform any pending work.
994 // All threads are now stopped
995 debugTrace(DEBUG_gc, "GC thread %d finished.", gct->thread_index);
998 #if defined(THREADED_RTS)
1000 // gc_thread_work(): Scavenge until there's no work left to do and all
1001 // the running threads are idle.
1004 gc_thread_work (void)
1006 // gc_running_threads has already been incremented for us; this is
1007 // a worker thread and the main thread bumped gc_running_threads
1008 // before waking us up.
1010 // Every thread evacuates some roots.
1012 markSomeCapabilities(mark_root, gct, gct->thread_index, n_gc_threads);
1014 scavenge_until_all_done();
1019 gc_thread_mainloop (void)
1021 while (!gct->exit) {
1023 // Wait until we're told to wake up
1024 ACQUIRE_LOCK(&gct->wake_mutex);
1025 gct->wakeup = rtsFalse;
1026 while (!gct->wakeup) {
1027 debugTrace(DEBUG_gc, "GC thread %d standing by...",
1029 waitCondition(&gct->wake_cond, &gct->wake_mutex);
1031 RELEASE_LOCK(&gct->wake_mutex);
1032 if (gct->exit) break;
1035 // start performance counters in this thread...
1036 if (gct->papi_events == -1) {
1037 papi_init_eventset(&gct->papi_events);
1039 papi_thread_start_gc1_count(gct->papi_events);
1045 // count events in this thread towards the GC totals
1046 papi_thread_stop_gc1_count(gct->papi_events);
1052 #if defined(THREADED_RTS)
1054 gc_thread_entry (gc_thread *my_gct)
1057 debugTrace(DEBUG_gc, "GC thread %d starting...", gct->thread_index);
1058 gct->id = osThreadId();
1059 gc_thread_mainloop();
1064 start_gc_threads (void)
1066 #if defined(THREADED_RTS)
1069 static rtsBool done = rtsFalse;
1071 gc_running_threads = 0;
1072 initMutex(&gc_running_mutex);
1075 // Start from 1: the main thread is 0
1076 for (i = 1; i < RtsFlags.ParFlags.gcThreads; i++) {
1077 createOSThread(&id, (OSThreadProc*)&gc_thread_entry,
1086 wakeup_gc_threads (nat n_threads USED_IF_THREADS)
1088 #if defined(THREADED_RTS)
1090 for (i=1; i < n_threads; i++) {
1092 debugTrace(DEBUG_gc, "waking up gc thread %d", i);
1094 ACQUIRE_LOCK(&gc_threads[i]->wake_mutex);
1095 if (gc_threads[i]->wakeup) {
1096 RELEASE_LOCK(&gc_threads[i]->wake_mutex);
1102 gc_threads[i]->wakeup = rtsTrue;
1103 signalCondition(&gc_threads[i]->wake_cond);
1104 RELEASE_LOCK(&gc_threads[i]->wake_mutex);
1109 // After GC is complete, we must wait for all GC threads to enter the
1110 // standby state, otherwise they may still be executing inside
1111 // any_work(), and may even remain awake until the next GC starts.
1113 shutdown_gc_threads (nat n_threads USED_IF_THREADS)
1115 #if defined(THREADED_RTS)
1118 for (i=1; i < n_threads; i++) {
1120 ACQUIRE_LOCK(&gc_threads[i]->wake_mutex);
1121 wakeup = gc_threads[i]->wakeup;
1122 // wakeup is false while the thread is waiting
1123 RELEASE_LOCK(&gc_threads[i]->wake_mutex);
1129 /* ----------------------------------------------------------------------------
1130 Initialise a generation that is to be collected
1131 ------------------------------------------------------------------------- */
1134 init_collected_gen (nat g, nat n_threads)
1141 // Throw away the current mutable list. Invariant: the mutable
1142 // list always has at least one block; this means we can avoid a
1143 // check for NULL in recordMutable().
1145 freeChain(generations[g].mut_list);
1146 generations[g].mut_list = allocBlock();
1147 for (i = 0; i < n_capabilities; i++) {
1148 freeChain(capabilities[i].mut_lists[g]);
1149 capabilities[i].mut_lists[g] = allocBlock();
1153 for (s = 0; s < generations[g].n_steps; s++) {
1155 stp = &generations[g].steps[s];
1156 ASSERT(stp->gen_no == g);
1158 // we'll construct a new list of threads in this step
1159 // during GC, throw away the current list.
1160 stp->old_threads = stp->threads;
1161 stp->threads = END_TSO_QUEUE;
1163 // generation 0, step 0 doesn't need to-space
1164 if (g == 0 && s == 0 && RtsFlags.GcFlags.generations > 1) {
1168 // deprecate the existing blocks
1169 stp->old_blocks = stp->blocks;
1170 stp->n_old_blocks = stp->n_blocks;
1174 stp->live_estimate = 0;
1176 // we don't have any to-be-scavenged blocks yet
1178 stp->todos_last = NULL;
1181 // initialise the large object queues.
1182 stp->scavenged_large_objects = NULL;
1183 stp->n_scavenged_large_blocks = 0;
1185 // mark the small objects as from-space
1186 for (bd = stp->old_blocks; bd; bd = bd->link) {
1187 bd->flags &= ~BF_EVACUATED;
1190 // mark the large objects as from-space
1191 for (bd = stp->large_objects; bd; bd = bd->link) {
1192 bd->flags &= ~BF_EVACUATED;
1195 // for a compacted step, we need to allocate the bitmap
1197 nat bitmap_size; // in bytes
1198 bdescr *bitmap_bdescr;
1201 bitmap_size = stp->n_old_blocks * BLOCK_SIZE / (sizeof(W_)*BITS_PER_BYTE);
1203 if (bitmap_size > 0) {
1204 bitmap_bdescr = allocGroup((lnat)BLOCK_ROUND_UP(bitmap_size)
1206 stp->bitmap = bitmap_bdescr;
1207 bitmap = bitmap_bdescr->start;
1209 debugTrace(DEBUG_gc, "bitmap_size: %d, bitmap: %p",
1210 bitmap_size, bitmap);
1212 // don't forget to fill it with zeros!
1213 memset(bitmap, 0, bitmap_size);
1215 // For each block in this step, point to its bitmap from the
1216 // block descriptor.
1217 for (bd=stp->old_blocks; bd != NULL; bd = bd->link) {
1218 bd->u.bitmap = bitmap;
1219 bitmap += BLOCK_SIZE_W / (sizeof(W_)*BITS_PER_BYTE);
1221 // Also at this point we set the BF_MARKED flag
1222 // for this block. The invariant is that
1223 // BF_MARKED is always unset, except during GC
1224 // when it is set on those blocks which will be
1226 if (!(bd->flags & BF_FRAGMENTED)) {
1227 bd->flags |= BF_MARKED;
1234 // For each GC thread, for each step, allocate a "todo" block to
1235 // store evacuated objects to be scavenged, and a block to store
1236 // evacuated objects that do not need to be scavenged.
1237 for (t = 0; t < n_threads; t++) {
1238 for (s = 0; s < generations[g].n_steps; s++) {
1240 // we don't copy objects into g0s0, unless -G0
1241 if (g==0 && s==0 && RtsFlags.GcFlags.generations > 1) continue;
1243 ws = &gc_threads[t]->steps[g * RtsFlags.GcFlags.steps + s];
1245 ws->todo_large_objects = NULL;
1247 ws->part_list = NULL;
1248 ws->n_part_blocks = 0;
1250 // allocate the first to-space block; extra blocks will be
1251 // chained on as necessary.
1253 ws->buffer_todo_bd = NULL;
1254 alloc_todo_block(ws,0);
1256 ws->scavd_list = NULL;
1257 ws->n_scavd_blocks = 0;
1263 /* ----------------------------------------------------------------------------
1264 Initialise a generation that is *not* to be collected
1265 ------------------------------------------------------------------------- */
1268 init_uncollected_gen (nat g, nat threads)
1275 for (s = 0; s < generations[g].n_steps; s++) {
1276 stp = &generations[g].steps[s];
1277 stp->scavenged_large_objects = NULL;
1278 stp->n_scavenged_large_blocks = 0;
1281 for (s = 0; s < generations[g].n_steps; s++) {
1283 stp = &generations[g].steps[s];
1285 for (t = 0; t < threads; t++) {
1286 ws = &gc_threads[t]->steps[g * RtsFlags.GcFlags.steps + s];
1288 ws->buffer_todo_bd = NULL;
1289 ws->todo_large_objects = NULL;
1291 ws->part_list = NULL;
1292 ws->n_part_blocks = 0;
1294 ws->scavd_list = NULL;
1295 ws->n_scavd_blocks = 0;
1297 // If the block at the head of the list in this generation
1298 // is less than 3/4 full, then use it as a todo block.
1299 if (stp->blocks && isPartiallyFull(stp->blocks))
1301 ws->todo_bd = stp->blocks;
1302 ws->todo_free = ws->todo_bd->free;
1303 ws->todo_lim = ws->todo_bd->start + BLOCK_SIZE_W;
1304 stp->blocks = stp->blocks->link;
1306 stp->n_words -= ws->todo_bd->free - ws->todo_bd->start;
1307 ws->todo_bd->link = NULL;
1308 // we must scan from the current end point.
1309 ws->todo_bd->u.scan = ws->todo_bd->free;
1314 alloc_todo_block(ws,0);
1318 // deal out any more partial blocks to the threads' part_lists
1320 while (stp->blocks && isPartiallyFull(stp->blocks))
1323 stp->blocks = bd->link;
1324 ws = &gc_threads[t]->steps[g * RtsFlags.GcFlags.steps + s];
1325 bd->link = ws->part_list;
1327 ws->n_part_blocks += 1;
1328 bd->u.scan = bd->free;
1330 stp->n_words -= bd->free - bd->start;
1332 if (t == n_gc_threads) t = 0;
1337 // Move the private mutable lists from each capability onto the
1338 // main mutable list for the generation.
1339 for (i = 0; i < n_capabilities; i++) {
1340 for (bd = capabilities[i].mut_lists[g];
1341 bd->link != NULL; bd = bd->link) {
1344 bd->link = generations[g].mut_list;
1345 generations[g].mut_list = capabilities[i].mut_lists[g];
1346 capabilities[i].mut_lists[g] = allocBlock();
1350 /* -----------------------------------------------------------------------------
1351 Initialise a gc_thread before GC
1352 -------------------------------------------------------------------------- */
1355 init_gc_thread (gc_thread *t)
1357 t->static_objects = END_OF_STATIC_LIST;
1358 t->scavenged_static_objects = END_OF_STATIC_LIST;
1361 t->failed_to_evac = rtsFalse;
1362 t->eager_promotion = rtsTrue;
1363 t->thunk_selector_depth = 0;
1368 t->scav_find_work = 0;
1371 /* -----------------------------------------------------------------------------
1372 Function we pass to evacuate roots.
1373 -------------------------------------------------------------------------- */
1376 mark_root(void *user, StgClosure **root)
1378 // we stole a register for gct, but this function is called from
1379 // *outside* the GC where the register variable is not in effect,
1380 // so we need to save and restore it here. NB. only call
1381 // mark_root() from the main GC thread, otherwise gct will be
1383 gc_thread *saved_gct;
1392 /* -----------------------------------------------------------------------------
1393 Initialising the static object & mutable lists
1394 -------------------------------------------------------------------------- */
1397 zero_static_object_list(StgClosure* first_static)
1401 const StgInfoTable *info;
1403 for (p = first_static; p != END_OF_STATIC_LIST; p = link) {
1405 link = *STATIC_LINK(info, p);
1406 *STATIC_LINK(info,p) = NULL;
1410 /* ----------------------------------------------------------------------------
1411 Update the pointers from the task list
1413 These are treated as weak pointers because we want to allow a main
1414 thread to get a BlockedOnDeadMVar exception in the same way as any
1415 other thread. Note that the threads should all have been retained
1416 by GC by virtue of being on the all_threads list, we're just
1417 updating pointers here.
1418 ------------------------------------------------------------------------- */
1421 update_task_list (void)
1425 for (task = all_tasks; task != NULL; task = task->all_link) {
1426 if (!task->stopped && task->tso) {
1427 ASSERT(task->tso->bound == task);
1428 tso = (StgTSO *) isAlive((StgClosure *)task->tso);
1430 barf("task %p: main thread %d has been GC'd",
1443 /* ----------------------------------------------------------------------------
1444 Reset the sizes of the older generations when we do a major
1447 CURRENT STRATEGY: make all generations except zero the same size.
1448 We have to stay within the maximum heap size, and leave a certain
1449 percentage of the maximum heap size available to allocate into.
1450 ------------------------------------------------------------------------- */
1453 resize_generations (void)
1457 if (major_gc && RtsFlags.GcFlags.generations > 1) {
1458 nat live, size, min_alloc, words;
1459 nat max = RtsFlags.GcFlags.maxHeapSize;
1460 nat gens = RtsFlags.GcFlags.generations;
1462 // live in the oldest generations
1463 if (oldest_gen->steps[0].live_estimate != 0) {
1464 words = oldest_gen->steps[0].live_estimate;
1466 words = oldest_gen->steps[0].n_words;
1468 live = (words + BLOCK_SIZE_W - 1) / BLOCK_SIZE_W +
1469 oldest_gen->steps[0].n_large_blocks;
1471 // default max size for all generations except zero
1472 size = stg_max(live * RtsFlags.GcFlags.oldGenFactor,
1473 RtsFlags.GcFlags.minOldGenSize);
1475 // minimum size for generation zero
1476 min_alloc = stg_max((RtsFlags.GcFlags.pcFreeHeap * max) / 200,
1477 RtsFlags.GcFlags.minAllocAreaSize);
1479 // Auto-enable compaction when the residency reaches a
1480 // certain percentage of the maximum heap size (default: 30%).
1481 if (RtsFlags.GcFlags.generations > 1 &&
1482 (RtsFlags.GcFlags.compact ||
1484 oldest_gen->steps[0].n_blocks >
1485 (RtsFlags.GcFlags.compactThreshold * max) / 100))) {
1486 oldest_gen->steps[0].mark = 1;
1487 oldest_gen->steps[0].compact = 1;
1488 // debugBelch("compaction: on\n", live);
1490 oldest_gen->steps[0].mark = 0;
1491 oldest_gen->steps[0].compact = 0;
1492 // debugBelch("compaction: off\n", live);
1495 if (RtsFlags.GcFlags.sweep) {
1496 oldest_gen->steps[0].mark = 1;
1499 // if we're going to go over the maximum heap size, reduce the
1500 // size of the generations accordingly. The calculation is
1501 // different if compaction is turned on, because we don't need
1502 // to double the space required to collect the old generation.
1505 // this test is necessary to ensure that the calculations
1506 // below don't have any negative results - we're working
1507 // with unsigned values here.
1508 if (max < min_alloc) {
1512 if (oldest_gen->steps[0].compact) {
1513 if ( (size + (size - 1) * (gens - 2) * 2) + min_alloc > max ) {
1514 size = (max - min_alloc) / ((gens - 1) * 2 - 1);
1517 if ( (size * (gens - 1) * 2) + min_alloc > max ) {
1518 size = (max - min_alloc) / ((gens - 1) * 2);
1528 debugBelch("live: %d, min_alloc: %d, size : %d, max = %d\n", live,
1529 min_alloc, size, max);
1532 for (g = 0; g < gens; g++) {
1533 generations[g].max_blocks = size;
1538 /* -----------------------------------------------------------------------------
1539 Calculate the new size of the nursery, and resize it.
1540 -------------------------------------------------------------------------- */
1543 resize_nursery (void)
1545 if (RtsFlags.GcFlags.generations == 1)
1546 { // Two-space collector:
1549 /* set up a new nursery. Allocate a nursery size based on a
1550 * function of the amount of live data (by default a factor of 2)
1551 * Use the blocks from the old nursery if possible, freeing up any
1554 * If we get near the maximum heap size, then adjust our nursery
1555 * size accordingly. If the nursery is the same size as the live
1556 * data (L), then we need 3L bytes. We can reduce the size of the
1557 * nursery to bring the required memory down near 2L bytes.
1559 * A normal 2-space collector would need 4L bytes to give the same
1560 * performance we get from 3L bytes, reducing to the same
1561 * performance at 2L bytes.
1563 blocks = g0s0->n_blocks;
1565 if ( RtsFlags.GcFlags.maxHeapSize != 0 &&
1566 blocks * RtsFlags.GcFlags.oldGenFactor * 2 >
1567 RtsFlags.GcFlags.maxHeapSize )
1569 long adjusted_blocks; // signed on purpose
1572 adjusted_blocks = (RtsFlags.GcFlags.maxHeapSize - 2 * blocks);
1574 debugTrace(DEBUG_gc, "near maximum heap size of 0x%x blocks, blocks = %d, adjusted to %ld",
1575 RtsFlags.GcFlags.maxHeapSize, blocks, adjusted_blocks);
1577 pc_free = adjusted_blocks * 100 / RtsFlags.GcFlags.maxHeapSize;
1578 if (pc_free < RtsFlags.GcFlags.pcFreeHeap) /* might even * be < 0 */
1582 blocks = adjusted_blocks;
1586 blocks *= RtsFlags.GcFlags.oldGenFactor;
1587 if (blocks < RtsFlags.GcFlags.minAllocAreaSize)
1589 blocks = RtsFlags.GcFlags.minAllocAreaSize;
1592 resizeNurseries(blocks);
1594 else // Generational collector
1597 * If the user has given us a suggested heap size, adjust our
1598 * allocation area to make best use of the memory available.
1600 if (RtsFlags.GcFlags.heapSizeSuggestion)
1603 nat needed = calcNeeded(); // approx blocks needed at next GC
1605 /* Guess how much will be live in generation 0 step 0 next time.
1606 * A good approximation is obtained by finding the
1607 * percentage of g0s0 that was live at the last minor GC.
1609 * We have an accurate figure for the amount of copied data in
1610 * 'copied', but we must convert this to a number of blocks, with
1611 * a small adjustment for estimated slop at the end of a block
1616 g0s0_pcnt_kept = ((copied / (BLOCK_SIZE_W - 10)) * 100)
1617 / countNurseryBlocks();
1620 /* Estimate a size for the allocation area based on the
1621 * information available. We might end up going slightly under
1622 * or over the suggested heap size, but we should be pretty
1625 * Formula: suggested - needed
1626 * ----------------------------
1627 * 1 + g0s0_pcnt_kept/100
1629 * where 'needed' is the amount of memory needed at the next
1630 * collection for collecting all steps except g0s0.
1633 (((long)RtsFlags.GcFlags.heapSizeSuggestion - (long)needed) * 100) /
1634 (100 + (long)g0s0_pcnt_kept);
1636 if (blocks < (long)RtsFlags.GcFlags.minAllocAreaSize) {
1637 blocks = RtsFlags.GcFlags.minAllocAreaSize;
1640 resizeNurseries((nat)blocks);
1644 // we might have added extra large blocks to the nursery, so
1645 // resize back to minAllocAreaSize again.
1646 resizeNurseriesFixed(RtsFlags.GcFlags.minAllocAreaSize);
1651 /* -----------------------------------------------------------------------------
1652 Sanity code for CAF garbage collection.
1654 With DEBUG turned on, we manage a CAF list in addition to the SRT
1655 mechanism. After GC, we run down the CAF list and blackhole any
1656 CAFs which have been garbage collected. This means we get an error
1657 whenever the program tries to enter a garbage collected CAF.
1659 Any garbage collected CAFs are taken off the CAF list at the same
1661 -------------------------------------------------------------------------- */
1663 #if 0 && defined(DEBUG)
1670 const StgInfoTable *info;
1681 ASSERT(info->type == IND_STATIC);
1683 if (STATIC_LINK(info,p) == NULL) {
1684 debugTrace(DEBUG_gccafs, "CAF gc'd at 0x%04lx", (long)p);
1686 SET_INFO(p,&stg_BLACKHOLE_info);
1687 p = STATIC_LINK2(info,p);
1691 pp = &STATIC_LINK2(info,p);
1698 debugTrace(DEBUG_gccafs, "%d CAFs live", i);