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"
25 #include "BlockAlloc.h"
29 #include "RtsSignals.h"
31 #if defined(RTS_GTK_FRONTPANEL)
32 #include "FrontPanel.h"
35 #include "RetainerProfile.h"
36 #include "LdvProfile.h"
37 #include "RaiseAsync.h"
47 #include "MarkStack.h"
52 #include <string.h> // for memset()
55 /* -----------------------------------------------------------------------------
57 -------------------------------------------------------------------------- */
59 /* STATIC OBJECT LIST.
62 * We maintain a linked list of static objects that are still live.
63 * The requirements for this list are:
65 * - we need to scan the list while adding to it, in order to
66 * scavenge all the static objects (in the same way that
67 * breadth-first scavenging works for dynamic objects).
69 * - we need to be able to tell whether an object is already on
70 * the list, to break loops.
72 * Each static object has a "static link field", which we use for
73 * linking objects on to the list. We use a stack-type list, consing
74 * objects on the front as they are added (this means that the
75 * scavenge phase is depth-first, not breadth-first, but that
78 * A separate list is kept for objects that have been scavenged
79 * already - this is so that we can zero all the marks afterwards.
81 * An object is on the list if its static link field is non-zero; this
82 * means that we have to mark the end of the list with '1', not NULL.
84 * Extra notes for generational GC:
86 * Each generation has a static object list associated with it. When
87 * collecting generations up to N, we treat the static object lists
88 * from generations > N as roots.
90 * We build up a static object list while collecting generations 0..N,
91 * which is then appended to the static object list of generation N+1.
94 /* N is the oldest generation being collected, where the generations
95 * are numbered starting at 0. A major GC (indicated by the major_gc
96 * flag) is when we're collecting all generations. We only attempt to
97 * deal with static objects and GC CAFs when doing a major GC.
102 /* Data used for allocation area sizing.
104 static lnat g0s0_pcnt_kept = 30; // percentage of g0s0 live at last minor GC
114 /* Thread-local data for each GC thread
116 gc_thread **gc_threads = NULL;
118 #if !defined(THREADED_RTS)
119 StgWord8 the_gc_thread[sizeof(gc_thread) + 64 * sizeof(step_workspace)];
122 // Number of threads running in *this* GC. Affects how many
123 // step->todos[] lists we have to look in to find work.
127 long copied; // *words* copied & scavenged during this GC
129 rtsBool work_stealing;
133 /* -----------------------------------------------------------------------------
134 Static function declarations
135 -------------------------------------------------------------------------- */
137 static void mark_root (void *user, StgClosure **root);
138 static void zero_static_object_list (StgClosure* first_static);
139 static nat initialise_N (rtsBool force_major_gc);
140 static void 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 scavenge_until_all_done (void);
148 static StgWord inc_running (void);
149 static StgWord dec_running (void);
150 static void wakeup_gc_threads (nat n_threads, nat me);
151 static void shutdown_gc_threads (nat n_threads, nat me);
153 #if 0 && defined(DEBUG)
154 static void gcCAFs (void);
157 /* -----------------------------------------------------------------------------
159 -------------------------------------------------------------------------- */
161 bdescr *mark_stack_top_bd; // topmost block in the mark stack
162 bdescr *mark_stack_bd; // current block in the mark stack
163 StgPtr mark_sp; // pointer to the next unallocated mark stack entry
165 /* -----------------------------------------------------------------------------
166 GarbageCollect: the main entry point to the garbage collector.
168 Locks held: all capabilities are held throughout GarbageCollect().
169 -------------------------------------------------------------------------- */
172 GarbageCollect (rtsBool force_major_gc,
173 nat gc_type USED_IF_THREADS,
178 lnat live, allocated, max_copied, avg_copied, slop;
179 gc_thread *saved_gct;
182 // necessary if we stole a callee-saves register for gct:
186 CostCentreStack *prev_CCS;
191 #if defined(RTS_USER_SIGNALS)
192 if (RtsFlags.MiscFlags.install_signal_handlers) {
198 ASSERT(sizeof(step_workspace) == 16 * sizeof(StgWord));
199 // otherwise adjust the padding in step_workspace.
201 // tell the stats department that we've started a GC
204 // tell the STM to discard any cached closures it's hoping to re-use
207 // lock the StablePtr table
216 // attribute any costs to CCS_GC
222 /* Approximate how much we allocated.
223 * Todo: only when generating stats?
225 allocated = calcAllocated();
227 /* Figure out which generation to collect
229 n = initialise_N(force_major_gc);
231 #if defined(THREADED_RTS)
232 work_stealing = RtsFlags.ParFlags.parGcLoadBalancingEnabled &&
233 N >= RtsFlags.ParFlags.parGcLoadBalancingGen;
234 // It's not always a good idea to do load balancing in parallel
235 // GC. In particular, for a parallel program we don't want to
236 // lose locality by moving cached data into another CPU's cache
237 // (this effect can be quite significant).
239 // We could have a more complex way to deterimine whether to do
240 // work stealing or not, e.g. it might be a good idea to do it
241 // if the heap is big. For now, we just turn it on or off with
245 /* Start threads, so they can be spinning up while we finish initialisation.
249 #if defined(THREADED_RTS)
250 /* How many threads will be participating in this GC?
251 * We don't try to parallelise minor GCs (unless the user asks for
252 * it with +RTS -gn0), or mark/compact/sweep GC.
254 if (gc_type == PENDING_GC_PAR) {
255 n_gc_threads = RtsFlags.ParFlags.nNodes;
263 debugTrace(DEBUG_gc, "GC (gen %d): %d KB to collect, %ld MB in use, using %d thread(s)",
264 N, n * (BLOCK_SIZE / 1024), mblocks_allocated, n_gc_threads);
266 #ifdef RTS_GTK_FRONTPANEL
267 if (RtsFlags.GcFlags.frontpanel) {
268 updateFrontPanelBeforeGC(N);
273 // check for memory leaks if DEBUG is on
274 memInventory(DEBUG_gc);
277 // check stack sanity *before* GC
278 IF_DEBUG(sanity, checkFreeListSanity());
279 IF_DEBUG(sanity, checkMutableLists(rtsTrue));
281 // Initialise all our gc_thread structures
282 for (t = 0; t < n_gc_threads; t++) {
283 init_gc_thread(gc_threads[t]);
286 // Initialise all the generations/steps that we're collecting.
287 for (g = 0; g <= N; g++) {
288 init_collected_gen(g,n_gc_threads);
291 // Initialise all the generations/steps that we're *not* collecting.
292 for (g = N+1; g < RtsFlags.GcFlags.generations; g++) {
293 init_uncollected_gen(g,n_gc_threads);
296 /* Allocate a mark stack if we're doing a major collection.
298 if (major_gc && oldest_gen->steps[0].mark) {
299 mark_stack_bd = allocBlock();
300 mark_stack_top_bd = mark_stack_bd;
301 mark_stack_bd->link = NULL;
302 mark_stack_bd->u.back = NULL;
303 mark_sp = mark_stack_bd->start;
305 mark_stack_bd = NULL;
306 mark_stack_top_bd = NULL;
310 // this is the main thread
312 if (n_gc_threads == 1) {
313 SET_GCT(gc_threads[0]);
315 SET_GCT(gc_threads[cap->no]);
318 SET_GCT(gc_threads[0]);
321 /* -----------------------------------------------------------------------
322 * follow all the roots that we know about:
325 // the main thread is running: this prevents any other threads from
326 // exiting prematurely, so we can start them now.
327 // NB. do this after the mutable lists have been saved above, otherwise
328 // the other GC threads will be writing into the old mutable lists.
330 wakeup_gc_threads(n_gc_threads, gct->thread_index);
332 // Mutable lists from each generation > N
333 // we want to *scavenge* these roots, not evacuate them: they're not
334 // going to move in this GC.
335 // Also do them in reverse generation order, for the usual reason:
336 // namely to reduce the likelihood of spurious old->new pointers.
338 for (g = RtsFlags.GcFlags.generations-1; g > N; g--) {
339 scavenge_mutable_list(generations[g].saved_mut_list, &generations[g]);
340 freeChain_sync(generations[g].saved_mut_list);
341 generations[g].saved_mut_list = NULL;
345 // scavenge the capability-private mutable lists. This isn't part
346 // of markSomeCapabilities() because markSomeCapabilities() can only
347 // call back into the GC via mark_root() (due to the gct register
349 if (n_gc_threads == 1) {
350 for (n = 0; n < n_capabilities; n++) {
351 scavenge_capability_mut_lists(&capabilities[n]);
354 scavenge_capability_mut_lists(&capabilities[gct->thread_index]);
357 // follow roots from the CAF list (used by GHCi)
359 markCAFs(mark_root, gct);
361 // follow all the roots that the application knows about.
363 markSomeCapabilities(mark_root, gct, gct->thread_index, n_gc_threads,
364 rtsTrue/*prune sparks*/);
366 #if defined(RTS_USER_SIGNALS)
367 // mark the signal handlers (signals should be already blocked)
368 markSignalHandlers(mark_root, gct);
371 // Mark the weak pointer list, and prepare to detect dead weak pointers.
375 // Mark the stable pointer table.
376 markStablePtrTable(mark_root, gct);
378 /* -------------------------------------------------------------------------
379 * Repeatedly scavenge all the areas we know about until there's no
380 * more scavenging to be done.
384 scavenge_until_all_done();
385 // The other threads are now stopped. We might recurse back to
386 // here, but from now on this is the only thread.
388 // if any blackholes are alive, make the threads that wait on
390 if (traverseBlackholeQueue()) {
395 // must be last... invariant is that everything is fully
396 // scavenged at this point.
397 if (traverseWeakPtrList()) { // returns rtsTrue if evaced something
402 // If we get to here, there's really nothing left to do.
406 shutdown_gc_threads(n_gc_threads, gct->thread_index);
408 // Update pointers from the Task list
411 // Now see which stable names are still alive.
415 // We call processHeapClosureForDead() on every closure destroyed during
416 // the current garbage collection, so we invoke LdvCensusForDead().
417 if (RtsFlags.ProfFlags.doHeapProfile == HEAP_BY_LDV
418 || RtsFlags.ProfFlags.bioSelector != NULL)
422 // NO MORE EVACUATION AFTER THIS POINT!
424 // Two-space collector: free the old to-space.
425 // g0s0->old_blocks is the old nursery
426 // g0s0->blocks is to-space from the previous GC
427 if (RtsFlags.GcFlags.generations == 1) {
428 if (g0s0->blocks != NULL) {
429 freeChain(g0s0->blocks);
434 // For each workspace, in each thread, move the copied blocks to the step
440 for (t = 0; t < n_gc_threads; t++) {
444 if (RtsFlags.GcFlags.generations == 1) {
449 for (; s < total_steps; s++) {
452 // Push the final block
454 push_scanned_block(ws->todo_bd, ws);
457 ASSERT(gct->scan_bd == NULL);
458 ASSERT(countBlocks(ws->scavd_list) == ws->n_scavd_blocks);
461 for (bd = ws->scavd_list; bd != NULL; bd = bd->link) {
462 ws->step->n_words += bd->free - bd->start;
466 prev->link = ws->step->blocks;
467 ws->step->blocks = ws->scavd_list;
469 ws->step->n_blocks += ws->n_scavd_blocks;
473 // Add all the partial blocks *after* we've added all the full
474 // blocks. This is so that we can grab the partial blocks back
475 // again and try to fill them up in the next GC.
476 for (t = 0; t < n_gc_threads; t++) {
480 if (RtsFlags.GcFlags.generations == 1) {
485 for (; s < total_steps; s++) {
489 for (bd = ws->part_list; bd != NULL; bd = next) {
491 if (bd->free == bd->start) {
493 ws->part_list = next;
500 ws->step->n_words += bd->free - bd->start;
505 prev->link = ws->step->blocks;
506 ws->step->blocks = ws->part_list;
508 ws->step->n_blocks += ws->n_part_blocks;
510 ASSERT(countBlocks(ws->step->blocks) == ws->step->n_blocks);
511 ASSERT(countOccupied(ws->step->blocks) == ws->step->n_words);
516 // Finally: compact or sweep the oldest generation.
517 if (major_gc && oldest_gen->steps[0].mark) {
518 if (oldest_gen->steps[0].compact)
519 compact(gct->scavenged_static_objects);
521 sweep(&oldest_gen->steps[0]);
524 /* run through all the generations/steps and tidy up
531 for (i=0; i < n_gc_threads; i++) {
532 if (n_gc_threads > 1) {
533 debugTrace(DEBUG_gc,"thread %d:", i);
534 debugTrace(DEBUG_gc," copied %ld", gc_threads[i]->copied * sizeof(W_));
535 debugTrace(DEBUG_gc," scanned %ld", gc_threads[i]->scanned * sizeof(W_));
536 debugTrace(DEBUG_gc," any_work %ld", gc_threads[i]->any_work);
537 debugTrace(DEBUG_gc," no_work %ld", gc_threads[i]->no_work);
538 debugTrace(DEBUG_gc," scav_find_work %ld", gc_threads[i]->scav_find_work);
540 copied += gc_threads[i]->copied;
541 max_copied = stg_max(gc_threads[i]->copied, max_copied);
543 if (n_gc_threads == 1) {
551 for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
554 generations[g].collections++; // for stats
555 if (n_gc_threads > 1) generations[g].par_collections++;
558 // Count the mutable list as bytes "copied" for the purposes of
559 // stats. Every mutable list is copied during every GC.
561 nat mut_list_size = 0;
562 for (bd = generations[g].mut_list; bd != NULL; bd = bd->link) {
563 mut_list_size += bd->free - bd->start;
565 for (n = 0; n < n_capabilities; n++) {
566 for (bd = capabilities[n].mut_lists[g];
567 bd != NULL; bd = bd->link) {
568 mut_list_size += bd->free - bd->start;
571 copied += mut_list_size;
574 "mut_list_size: %lu (%d vars, %d arrays, %d MVARs, %d others)",
575 (unsigned long)(mut_list_size * sizeof(W_)),
576 mutlist_MUTVARS, mutlist_MUTARRS, mutlist_MVARS, mutlist_OTHERS);
579 for (s = 0; s < generations[g].n_steps; s++) {
581 stp = &generations[g].steps[s];
583 // for generations we collected...
586 /* free old memory and shift to-space into from-space for all
587 * the collected steps (except the allocation area). These
588 * freed blocks will probaby be quickly recycled.
590 if (!(g == 0 && s == 0 && RtsFlags.GcFlags.generations > 1)) {
593 // tack the new blocks on the end of the existing blocks
594 if (stp->old_blocks != NULL) {
597 for (bd = stp->old_blocks; bd != NULL; bd = next) {
601 if (!(bd->flags & BF_MARKED))
604 stp->old_blocks = next;
613 stp->n_words += bd->free - bd->start;
615 // NB. this step might not be compacted next
616 // time, so reset the BF_MARKED flags.
617 // They are set before GC if we're going to
618 // compact. (search for BF_MARKED above).
619 bd->flags &= ~BF_MARKED;
621 // between GCs, all blocks in the heap except
622 // for the nursery have the BF_EVACUATED flag set.
623 bd->flags |= BF_EVACUATED;
630 prev->link = stp->blocks;
631 stp->blocks = stp->old_blocks;
634 // add the new blocks to the block tally
635 stp->n_blocks += stp->n_old_blocks;
636 ASSERT(countBlocks(stp->blocks) == stp->n_blocks);
637 ASSERT(countOccupied(stp->blocks) == stp->n_words);
641 freeChain(stp->old_blocks);
643 stp->old_blocks = NULL;
644 stp->n_old_blocks = 0;
647 /* LARGE OBJECTS. The current live large objects are chained on
648 * scavenged_large, having been moved during garbage
649 * collection from large_objects. Any objects left on
650 * large_objects list are therefore dead, so we free them here.
652 for (bd = stp->large_objects; bd != NULL; bd = next) {
658 stp->large_objects = stp->scavenged_large_objects;
659 stp->n_large_blocks = stp->n_scavenged_large_blocks;
662 else // for older generations...
664 /* For older generations, we need to append the
665 * scavenged_large_object list (i.e. large objects that have been
666 * promoted during this GC) to the large_object list for that step.
668 for (bd = stp->scavenged_large_objects; bd; bd = next) {
670 dbl_link_onto(bd, &stp->large_objects);
673 // add the new blocks we promoted during this GC
674 stp->n_large_blocks += stp->n_scavenged_large_blocks;
679 // update the max size of older generations after a major GC
680 resize_generations();
682 // Calculate the amount of live data for stats.
683 live = calcLiveWords();
685 // Free the small objects allocated via allocate(), since this will
686 // all have been copied into G0S1 now.
687 if (RtsFlags.GcFlags.generations > 1) {
688 if (g0s0->blocks != NULL) {
689 freeChain(g0s0->blocks);
696 alloc_blocks_lim = RtsFlags.GcFlags.minAllocAreaSize;
698 // Start a new pinned_object_block
699 pinned_object_block = NULL;
701 // Free the mark stack.
702 if (mark_stack_top_bd != NULL) {
703 debugTrace(DEBUG_gc, "mark stack: %d blocks",
704 countBlocks(mark_stack_top_bd));
705 freeChain(mark_stack_top_bd);
709 for (g = 0; g <= N; g++) {
710 for (s = 0; s < generations[g].n_steps; s++) {
711 stp = &generations[g].steps[s];
712 if (stp->bitmap != NULL) {
713 freeGroup(stp->bitmap);
721 // mark the garbage collected CAFs as dead
722 #if 0 && defined(DEBUG) // doesn't work at the moment
723 if (major_gc) { gcCAFs(); }
727 // resetStaticObjectForRetainerProfiling() must be called before
729 if (n_gc_threads > 1) {
730 barf("profiling is currently broken with multi-threaded GC");
731 // ToDo: fix the gct->scavenged_static_objects below
733 resetStaticObjectForRetainerProfiling(gct->scavenged_static_objects);
736 // zero the scavenged static object list
739 for (i = 0; i < n_gc_threads; i++) {
740 zero_static_object_list(gc_threads[i]->scavenged_static_objects);
747 // start any pending finalizers
749 scheduleFinalizers(cap, old_weak_ptr_list);
752 // send exceptions to any threads which were about to die
754 resurrectThreads(resurrected_threads);
755 performPendingThrowTos(exception_threads);
758 // Update the stable pointer hash table.
759 updateStablePtrTable(major_gc);
761 // check sanity after GC
762 IF_DEBUG(sanity, checkSanity());
764 // extra GC trace info
765 IF_DEBUG(gc, statDescribeGens());
768 // symbol-table based profiling
769 /* heapCensus(to_blocks); */ /* ToDo */
772 // restore enclosing cost centre
778 // check for memory leaks if DEBUG is on
779 memInventory(DEBUG_gc);
782 #ifdef RTS_GTK_FRONTPANEL
783 if (RtsFlags.GcFlags.frontpanel) {
784 updateFrontPanelAfterGC( N, live );
788 // ok, GC over: tell the stats department what happened.
789 slop = calcLiveBlocks() * BLOCK_SIZE_W - live;
790 stat_endGC(allocated, live, copied, N, max_copied, avg_copied, slop);
792 // unlock the StablePtr table
795 // Guess which generation we'll collect *next* time
796 initialise_N(force_major_gc);
798 #if defined(RTS_USER_SIGNALS)
799 if (RtsFlags.MiscFlags.install_signal_handlers) {
800 // unblock signals again
801 unblockUserSignals();
810 /* -----------------------------------------------------------------------------
811 Figure out which generation to collect, initialise N and major_gc.
813 Also returns the total number of blocks in generations that will be
815 -------------------------------------------------------------------------- */
818 initialise_N (rtsBool force_major_gc)
821 nat s, blocks, blocks_total;
826 if (force_major_gc) {
827 N = RtsFlags.GcFlags.generations - 1;
832 for (g = RtsFlags.GcFlags.generations - 1; g >= 0; g--) {
834 for (s = 0; s < generations[g].n_steps; s++) {
835 blocks += generations[g].steps[s].n_words / BLOCK_SIZE_W;
836 blocks += generations[g].steps[s].n_large_blocks;
838 if (blocks >= generations[g].max_blocks) {
842 blocks_total += blocks;
846 blocks_total += countNurseryBlocks();
848 major_gc = (N == RtsFlags.GcFlags.generations-1);
852 /* -----------------------------------------------------------------------------
853 Initialise the gc_thread structures.
854 -------------------------------------------------------------------------- */
856 #define GC_THREAD_INACTIVE 0
857 #define GC_THREAD_STANDING_BY 1
858 #define GC_THREAD_RUNNING 2
859 #define GC_THREAD_WAITING_TO_CONTINUE 3
862 new_gc_thread (nat n, gc_thread *t)
869 initSpinLock(&t->gc_spin);
870 initSpinLock(&t->mut_spin);
871 ACQUIRE_SPIN_LOCK(&t->gc_spin);
872 t->wakeup = GC_THREAD_INACTIVE; // starts true, so we can wait for the
873 // thread to start up, see wakeup_gc_threads
877 t->free_blocks = NULL;
886 for (s = 0; s < total_steps; s++)
889 ws->step = &all_steps[s];
890 ASSERT(s == ws->step->abs_no);
894 ws->todo_q = newWSDeque(128);
895 ws->todo_overflow = NULL;
896 ws->n_todo_overflow = 0;
898 ws->part_list = NULL;
899 ws->n_part_blocks = 0;
901 ws->scavd_list = NULL;
902 ws->n_scavd_blocks = 0;
910 if (gc_threads == NULL) {
911 #if defined(THREADED_RTS)
913 gc_threads = stgMallocBytes (RtsFlags.ParFlags.nNodes *
917 for (i = 0; i < RtsFlags.ParFlags.nNodes; i++) {
919 stgMallocBytes(sizeof(gc_thread) + total_steps * sizeof(step_workspace),
922 new_gc_thread(i, gc_threads[i]);
925 gc_threads = stgMallocBytes (sizeof(gc_thread*),"alloc_gc_threads");
927 new_gc_thread(0,gc_threads[0]);
935 if (gc_threads != NULL) {
936 #if defined(THREADED_RTS)
938 for (i = 0; i < RtsFlags.ParFlags.nNodes; i++) {
939 stgFree (gc_threads[i]);
941 stgFree (gc_threads);
943 stgFree (gc_threads);
949 /* ----------------------------------------------------------------------------
951 ------------------------------------------------------------------------- */
953 static volatile StgWord gc_running_threads;
959 new = atomic_inc(&gc_running_threads);
960 ASSERT(new <= n_gc_threads);
967 ASSERT(gc_running_threads != 0);
968 return atomic_dec(&gc_running_threads);
981 // scavenge objects in compacted generation
982 if (mark_stack_bd != NULL && !mark_stack_empty()) {
986 // Check for global work in any step. We don't need to check for
987 // local work, because we have already exited scavenge_loop(),
988 // which means there is no local work for this thread.
989 for (s = total_steps-1; s >= 0; s--) {
990 if (s == 0 && RtsFlags.GcFlags.generations > 1) {
994 if (ws->todo_large_objects) return rtsTrue;
995 if (!looksEmptyWSDeque(ws->todo_q)) return rtsTrue;
996 if (ws->todo_overflow) return rtsTrue;
999 #if defined(THREADED_RTS)
1000 if (work_stealing) {
1002 // look for work to steal
1003 for (n = 0; n < n_gc_threads; n++) {
1004 if (n == gct->thread_index) continue;
1005 for (s = total_steps-1; s >= 0; s--) {
1006 ws = &gc_threads[n]->steps[s];
1007 if (!looksEmptyWSDeque(ws->todo_q)) return rtsTrue;
1019 scavenge_until_all_done (void)
1025 traceEvent(&capabilities[gct->thread_index], EVENT_GC_WORK);
1027 #if defined(THREADED_RTS)
1028 if (n_gc_threads > 1) {
1037 // scavenge_loop() only exits when there's no work to do
1040 traceEvent(&capabilities[gct->thread_index], EVENT_GC_IDLE);
1042 debugTrace(DEBUG_gc, "%d GC threads still running", r);
1044 while (gc_running_threads != 0) {
1050 // any_work() does not remove the work from the queue, it
1051 // just checks for the presence of work. If we find any,
1052 // then we increment gc_running_threads and go back to
1053 // scavenge_loop() to perform any pending work.
1056 traceEvent(&capabilities[gct->thread_index], EVENT_GC_DONE);
1059 #if defined(THREADED_RTS)
1062 gcWorkerThread (Capability *cap)
1064 gc_thread *saved_gct;
1066 // necessary if we stole a callee-saves register for gct:
1069 cap->in_gc = rtsTrue;
1071 gct = gc_threads[cap->no];
1072 gct->id = osThreadId();
1074 // Wait until we're told to wake up
1075 RELEASE_SPIN_LOCK(&gct->mut_spin);
1076 gct->wakeup = GC_THREAD_STANDING_BY;
1077 debugTrace(DEBUG_gc, "GC thread %d standing by...", gct->thread_index);
1078 ACQUIRE_SPIN_LOCK(&gct->gc_spin);
1081 // start performance counters in this thread...
1082 if (gct->papi_events == -1) {
1083 papi_init_eventset(&gct->papi_events);
1085 papi_thread_start_gc1_count(gct->papi_events);
1088 // Every thread evacuates some roots.
1090 markSomeCapabilities(mark_root, gct, gct->thread_index, n_gc_threads,
1091 rtsTrue/*prune sparks*/);
1092 scavenge_capability_mut_lists(&capabilities[gct->thread_index]);
1094 scavenge_until_all_done();
1097 // count events in this thread towards the GC totals
1098 papi_thread_stop_gc1_count(gct->papi_events);
1101 // Wait until we're told to continue
1102 RELEASE_SPIN_LOCK(&gct->gc_spin);
1103 gct->wakeup = GC_THREAD_WAITING_TO_CONTINUE;
1104 debugTrace(DEBUG_gc, "GC thread %d waiting to continue...",
1106 ACQUIRE_SPIN_LOCK(&gct->mut_spin);
1107 debugTrace(DEBUG_gc, "GC thread %d on my way...", gct->thread_index);
1114 #if defined(THREADED_RTS)
1117 waitForGcThreads (Capability *cap USED_IF_THREADS)
1119 nat n_threads = RtsFlags.ParFlags.nNodes;
1122 rtsBool retry = rtsTrue;
1125 for (i=0; i < n_threads; i++) {
1126 if (i == me) continue;
1127 if (gc_threads[i]->wakeup != GC_THREAD_STANDING_BY) {
1128 prodCapability(&capabilities[i], cap->running_task);
1131 for (j=0; j < 10; j++) {
1133 for (i=0; i < n_threads; i++) {
1134 if (i == me) continue;
1136 setContextSwitches();
1137 if (gc_threads[i]->wakeup != GC_THREAD_STANDING_BY) {
1147 #endif // THREADED_RTS
1150 start_gc_threads (void)
1152 #if defined(THREADED_RTS)
1153 gc_running_threads = 0;
1158 wakeup_gc_threads (nat n_threads USED_IF_THREADS, nat me USED_IF_THREADS)
1160 #if defined(THREADED_RTS)
1162 for (i=0; i < n_threads; i++) {
1163 if (i == me) continue;
1165 debugTrace(DEBUG_gc, "waking up gc thread %d", i);
1166 if (gc_threads[i]->wakeup != GC_THREAD_STANDING_BY) barf("wakeup_gc_threads");
1168 gc_threads[i]->wakeup = GC_THREAD_RUNNING;
1169 ACQUIRE_SPIN_LOCK(&gc_threads[i]->mut_spin);
1170 RELEASE_SPIN_LOCK(&gc_threads[i]->gc_spin);
1175 // After GC is complete, we must wait for all GC threads to enter the
1176 // standby state, otherwise they may still be executing inside
1177 // any_work(), and may even remain awake until the next GC starts.
1179 shutdown_gc_threads (nat n_threads USED_IF_THREADS, nat me USED_IF_THREADS)
1181 #if defined(THREADED_RTS)
1183 for (i=0; i < n_threads; i++) {
1184 if (i == me) continue;
1185 while (gc_threads[i]->wakeup != GC_THREAD_WAITING_TO_CONTINUE) { write_barrier(); }
1190 #if defined(THREADED_RTS)
1192 releaseGCThreads (Capability *cap USED_IF_THREADS)
1194 nat n_threads = RtsFlags.ParFlags.nNodes;
1197 for (i=0; i < n_threads; i++) {
1198 if (i == me) continue;
1199 if (gc_threads[i]->wakeup != GC_THREAD_WAITING_TO_CONTINUE)
1200 barf("releaseGCThreads");
1202 gc_threads[i]->wakeup = GC_THREAD_INACTIVE;
1203 ACQUIRE_SPIN_LOCK(&gc_threads[i]->gc_spin);
1204 RELEASE_SPIN_LOCK(&gc_threads[i]->mut_spin);
1209 /* ----------------------------------------------------------------------------
1210 Initialise a generation that is to be collected
1211 ------------------------------------------------------------------------- */
1214 init_collected_gen (nat g, nat n_threads)
1221 // Throw away the current mutable list. Invariant: the mutable
1222 // list always has at least one block; this means we can avoid a
1223 // check for NULL in recordMutable().
1225 freeChain(generations[g].mut_list);
1226 generations[g].mut_list = allocBlock();
1227 for (i = 0; i < n_capabilities; i++) {
1228 freeChain(capabilities[i].mut_lists[g]);
1229 capabilities[i].mut_lists[g] = allocBlock();
1233 for (s = 0; s < generations[g].n_steps; s++) {
1235 stp = &generations[g].steps[s];
1236 ASSERT(stp->gen_no == g);
1238 // we'll construct a new list of threads in this step
1239 // during GC, throw away the current list.
1240 stp->old_threads = stp->threads;
1241 stp->threads = END_TSO_QUEUE;
1243 // generation 0, step 0 doesn't need to-space
1244 if (g == 0 && s == 0 && RtsFlags.GcFlags.generations > 1) {
1248 // deprecate the existing blocks
1249 stp->old_blocks = stp->blocks;
1250 stp->n_old_blocks = stp->n_blocks;
1254 stp->live_estimate = 0;
1256 // initialise the large object queues.
1257 stp->scavenged_large_objects = NULL;
1258 stp->n_scavenged_large_blocks = 0;
1260 // mark the small objects as from-space
1261 for (bd = stp->old_blocks; bd; bd = bd->link) {
1262 bd->flags &= ~BF_EVACUATED;
1265 // mark the large objects as from-space
1266 for (bd = stp->large_objects; bd; bd = bd->link) {
1267 bd->flags &= ~BF_EVACUATED;
1270 // for a compacted step, we need to allocate the bitmap
1272 nat bitmap_size; // in bytes
1273 bdescr *bitmap_bdescr;
1276 bitmap_size = stp->n_old_blocks * BLOCK_SIZE / (sizeof(W_)*BITS_PER_BYTE);
1278 if (bitmap_size > 0) {
1279 bitmap_bdescr = allocGroup((lnat)BLOCK_ROUND_UP(bitmap_size)
1281 stp->bitmap = bitmap_bdescr;
1282 bitmap = bitmap_bdescr->start;
1284 debugTrace(DEBUG_gc, "bitmap_size: %d, bitmap: %p",
1285 bitmap_size, bitmap);
1287 // don't forget to fill it with zeros!
1288 memset(bitmap, 0, bitmap_size);
1290 // For each block in this step, point to its bitmap from the
1291 // block descriptor.
1292 for (bd=stp->old_blocks; bd != NULL; bd = bd->link) {
1293 bd->u.bitmap = bitmap;
1294 bitmap += BLOCK_SIZE_W / (sizeof(W_)*BITS_PER_BYTE);
1296 // Also at this point we set the BF_MARKED flag
1297 // for this block. The invariant is that
1298 // BF_MARKED is always unset, except during GC
1299 // when it is set on those blocks which will be
1301 if (!(bd->flags & BF_FRAGMENTED)) {
1302 bd->flags |= BF_MARKED;
1309 // For each GC thread, for each step, allocate a "todo" block to
1310 // store evacuated objects to be scavenged, and a block to store
1311 // evacuated objects that do not need to be scavenged.
1312 for (t = 0; t < n_threads; t++) {
1313 for (s = 0; s < generations[g].n_steps; s++) {
1315 // we don't copy objects into g0s0, unless -G0
1316 if (g==0 && s==0 && RtsFlags.GcFlags.generations > 1) continue;
1318 ws = &gc_threads[t]->steps[g * RtsFlags.GcFlags.steps + s];
1320 ws->todo_large_objects = NULL;
1322 ws->part_list = NULL;
1323 ws->n_part_blocks = 0;
1325 // allocate the first to-space block; extra blocks will be
1326 // chained on as necessary.
1328 ASSERT(looksEmptyWSDeque(ws->todo_q));
1329 alloc_todo_block(ws,0);
1331 ws->todo_overflow = NULL;
1332 ws->n_todo_overflow = 0;
1334 ws->scavd_list = NULL;
1335 ws->n_scavd_blocks = 0;
1341 /* ----------------------------------------------------------------------------
1342 Initialise a generation that is *not* to be collected
1343 ------------------------------------------------------------------------- */
1346 init_uncollected_gen (nat g, nat threads)
1353 // save the current mutable lists for this generation, and
1354 // allocate a fresh block for each one. We'll traverse these
1355 // mutable lists as roots early on in the GC.
1356 generations[g].saved_mut_list = generations[g].mut_list;
1357 generations[g].mut_list = allocBlock();
1358 for (n = 0; n < n_capabilities; n++) {
1359 capabilities[n].saved_mut_lists[g] = capabilities[n].mut_lists[g];
1360 capabilities[n].mut_lists[g] = allocBlock();
1363 for (s = 0; s < generations[g].n_steps; s++) {
1364 stp = &generations[g].steps[s];
1365 stp->scavenged_large_objects = NULL;
1366 stp->n_scavenged_large_blocks = 0;
1369 for (s = 0; s < generations[g].n_steps; s++) {
1371 stp = &generations[g].steps[s];
1373 for (t = 0; t < threads; t++) {
1374 ws = &gc_threads[t]->steps[g * RtsFlags.GcFlags.steps + s];
1376 ASSERT(looksEmptyWSDeque(ws->todo_q));
1377 ws->todo_large_objects = NULL;
1379 ws->part_list = NULL;
1380 ws->n_part_blocks = 0;
1382 ws->scavd_list = NULL;
1383 ws->n_scavd_blocks = 0;
1385 // If the block at the head of the list in this generation
1386 // is less than 3/4 full, then use it as a todo block.
1387 if (stp->blocks && isPartiallyFull(stp->blocks))
1389 ws->todo_bd = stp->blocks;
1390 ws->todo_free = ws->todo_bd->free;
1391 ws->todo_lim = ws->todo_bd->start + BLOCK_SIZE_W;
1392 stp->blocks = stp->blocks->link;
1394 stp->n_words -= ws->todo_bd->free - ws->todo_bd->start;
1395 ws->todo_bd->link = NULL;
1396 // we must scan from the current end point.
1397 ws->todo_bd->u.scan = ws->todo_bd->free;
1402 alloc_todo_block(ws,0);
1406 // deal out any more partial blocks to the threads' part_lists
1408 while (stp->blocks && isPartiallyFull(stp->blocks))
1411 stp->blocks = bd->link;
1412 ws = &gc_threads[t]->steps[g * RtsFlags.GcFlags.steps + s];
1413 bd->link = ws->part_list;
1415 ws->n_part_blocks += 1;
1416 bd->u.scan = bd->free;
1418 stp->n_words -= bd->free - bd->start;
1420 if (t == n_gc_threads) t = 0;
1425 /* -----------------------------------------------------------------------------
1426 Initialise a gc_thread before GC
1427 -------------------------------------------------------------------------- */
1430 init_gc_thread (gc_thread *t)
1432 t->static_objects = END_OF_STATIC_LIST;
1433 t->scavenged_static_objects = END_OF_STATIC_LIST;
1435 t->mut_lists = capabilities[t->thread_index].mut_lists;
1437 t->failed_to_evac = rtsFalse;
1438 t->eager_promotion = rtsTrue;
1439 t->thunk_selector_depth = 0;
1444 t->scav_find_work = 0;
1447 /* -----------------------------------------------------------------------------
1448 Function we pass to evacuate roots.
1449 -------------------------------------------------------------------------- */
1452 mark_root(void *user USED_IF_THREADS, StgClosure **root)
1454 // we stole a register for gct, but this function is called from
1455 // *outside* the GC where the register variable is not in effect,
1456 // so we need to save and restore it here. NB. only call
1457 // mark_root() from the main GC thread, otherwise gct will be
1459 gc_thread *saved_gct;
1468 /* -----------------------------------------------------------------------------
1469 Initialising the static object & mutable lists
1470 -------------------------------------------------------------------------- */
1473 zero_static_object_list(StgClosure* first_static)
1477 const StgInfoTable *info;
1479 for (p = first_static; p != END_OF_STATIC_LIST; p = link) {
1481 link = *STATIC_LINK(info, p);
1482 *STATIC_LINK(info,p) = NULL;
1486 /* ----------------------------------------------------------------------------
1487 Update the pointers from the task list
1489 These are treated as weak pointers because we want to allow a main
1490 thread to get a BlockedOnDeadMVar exception in the same way as any
1491 other thread. Note that the threads should all have been retained
1492 by GC by virtue of being on the all_threads list, we're just
1493 updating pointers here.
1494 ------------------------------------------------------------------------- */
1497 update_task_list (void)
1501 for (task = all_tasks; task != NULL; task = task->all_link) {
1502 if (!task->stopped && task->tso) {
1503 ASSERT(task->tso->bound == task);
1504 tso = (StgTSO *) isAlive((StgClosure *)task->tso);
1506 barf("task %p: main thread %d has been GC'd",
1519 /* ----------------------------------------------------------------------------
1520 Reset the sizes of the older generations when we do a major
1523 CURRENT STRATEGY: make all generations except zero the same size.
1524 We have to stay within the maximum heap size, and leave a certain
1525 percentage of the maximum heap size available to allocate into.
1526 ------------------------------------------------------------------------- */
1529 resize_generations (void)
1533 if (major_gc && RtsFlags.GcFlags.generations > 1) {
1534 nat live, size, min_alloc, words;
1535 nat max = RtsFlags.GcFlags.maxHeapSize;
1536 nat gens = RtsFlags.GcFlags.generations;
1538 // live in the oldest generations
1539 if (oldest_gen->steps[0].live_estimate != 0) {
1540 words = oldest_gen->steps[0].live_estimate;
1542 words = oldest_gen->steps[0].n_words;
1544 live = (words + BLOCK_SIZE_W - 1) / BLOCK_SIZE_W +
1545 oldest_gen->steps[0].n_large_blocks;
1547 // default max size for all generations except zero
1548 size = stg_max(live * RtsFlags.GcFlags.oldGenFactor,
1549 RtsFlags.GcFlags.minOldGenSize);
1551 // minimum size for generation zero
1552 min_alloc = stg_max((RtsFlags.GcFlags.pcFreeHeap * max) / 200,
1553 RtsFlags.GcFlags.minAllocAreaSize);
1555 // Auto-enable compaction when the residency reaches a
1556 // certain percentage of the maximum heap size (default: 30%).
1557 if (RtsFlags.GcFlags.generations > 1 &&
1558 (RtsFlags.GcFlags.compact ||
1560 oldest_gen->steps[0].n_blocks >
1561 (RtsFlags.GcFlags.compactThreshold * max) / 100))) {
1562 oldest_gen->steps[0].mark = 1;
1563 oldest_gen->steps[0].compact = 1;
1564 // debugBelch("compaction: on\n", live);
1566 oldest_gen->steps[0].mark = 0;
1567 oldest_gen->steps[0].compact = 0;
1568 // debugBelch("compaction: off\n", live);
1571 if (RtsFlags.GcFlags.sweep) {
1572 oldest_gen->steps[0].mark = 1;
1575 // if we're going to go over the maximum heap size, reduce the
1576 // size of the generations accordingly. The calculation is
1577 // different if compaction is turned on, because we don't need
1578 // to double the space required to collect the old generation.
1581 // this test is necessary to ensure that the calculations
1582 // below don't have any negative results - we're working
1583 // with unsigned values here.
1584 if (max < min_alloc) {
1588 if (oldest_gen->steps[0].compact) {
1589 if ( (size + (size - 1) * (gens - 2) * 2) + min_alloc > max ) {
1590 size = (max - min_alloc) / ((gens - 1) * 2 - 1);
1593 if ( (size * (gens - 1) * 2) + min_alloc > max ) {
1594 size = (max - min_alloc) / ((gens - 1) * 2);
1604 debugBelch("live: %d, min_alloc: %d, size : %d, max = %d\n", live,
1605 min_alloc, size, max);
1608 for (g = 0; g < gens; g++) {
1609 generations[g].max_blocks = size;
1614 /* -----------------------------------------------------------------------------
1615 Calculate the new size of the nursery, and resize it.
1616 -------------------------------------------------------------------------- */
1619 resize_nursery (void)
1621 lnat min_nursery = RtsFlags.GcFlags.minAllocAreaSize * n_capabilities;
1623 if (RtsFlags.GcFlags.generations == 1)
1624 { // Two-space collector:
1627 /* set up a new nursery. Allocate a nursery size based on a
1628 * function of the amount of live data (by default a factor of 2)
1629 * Use the blocks from the old nursery if possible, freeing up any
1632 * If we get near the maximum heap size, then adjust our nursery
1633 * size accordingly. If the nursery is the same size as the live
1634 * data (L), then we need 3L bytes. We can reduce the size of the
1635 * nursery to bring the required memory down near 2L bytes.
1637 * A normal 2-space collector would need 4L bytes to give the same
1638 * performance we get from 3L bytes, reducing to the same
1639 * performance at 2L bytes.
1641 blocks = g0s0->n_blocks;
1643 if ( RtsFlags.GcFlags.maxHeapSize != 0 &&
1644 blocks * RtsFlags.GcFlags.oldGenFactor * 2 >
1645 RtsFlags.GcFlags.maxHeapSize )
1647 long adjusted_blocks; // signed on purpose
1650 adjusted_blocks = (RtsFlags.GcFlags.maxHeapSize - 2 * blocks);
1652 debugTrace(DEBUG_gc, "near maximum heap size of 0x%x blocks, blocks = %d, adjusted to %ld",
1653 RtsFlags.GcFlags.maxHeapSize, blocks, adjusted_blocks);
1655 pc_free = adjusted_blocks * 100 / RtsFlags.GcFlags.maxHeapSize;
1656 if (pc_free < RtsFlags.GcFlags.pcFreeHeap) /* might even * be < 0 */
1660 blocks = adjusted_blocks;
1664 blocks *= RtsFlags.GcFlags.oldGenFactor;
1665 if (blocks < min_nursery)
1667 blocks = min_nursery;
1670 resizeNurseries(blocks);
1672 else // Generational collector
1675 * If the user has given us a suggested heap size, adjust our
1676 * allocation area to make best use of the memory available.
1678 if (RtsFlags.GcFlags.heapSizeSuggestion)
1681 nat needed = calcNeeded(); // approx blocks needed at next GC
1683 /* Guess how much will be live in generation 0 step 0 next time.
1684 * A good approximation is obtained by finding the
1685 * percentage of g0s0 that was live at the last minor GC.
1687 * We have an accurate figure for the amount of copied data in
1688 * 'copied', but we must convert this to a number of blocks, with
1689 * a small adjustment for estimated slop at the end of a block
1694 g0s0_pcnt_kept = ((copied / (BLOCK_SIZE_W - 10)) * 100)
1695 / countNurseryBlocks();
1698 /* Estimate a size for the allocation area based on the
1699 * information available. We might end up going slightly under
1700 * or over the suggested heap size, but we should be pretty
1703 * Formula: suggested - needed
1704 * ----------------------------
1705 * 1 + g0s0_pcnt_kept/100
1707 * where 'needed' is the amount of memory needed at the next
1708 * collection for collecting all steps except g0s0.
1711 (((long)RtsFlags.GcFlags.heapSizeSuggestion - (long)needed) * 100) /
1712 (100 + (long)g0s0_pcnt_kept);
1714 if (blocks < (long)min_nursery) {
1715 blocks = min_nursery;
1718 resizeNurseries((nat)blocks);
1722 // we might have added extra large blocks to the nursery, so
1723 // resize back to minAllocAreaSize again.
1724 resizeNurseriesFixed(RtsFlags.GcFlags.minAllocAreaSize);
1729 /* -----------------------------------------------------------------------------
1730 Sanity code for CAF garbage collection.
1732 With DEBUG turned on, we manage a CAF list in addition to the SRT
1733 mechanism. After GC, we run down the CAF list and blackhole any
1734 CAFs which have been garbage collected. This means we get an error
1735 whenever the program tries to enter a garbage collected CAF.
1737 Any garbage collected CAFs are taken off the CAF list at the same
1739 -------------------------------------------------------------------------- */
1741 #if 0 && defined(DEBUG)
1748 const StgInfoTable *info;
1759 ASSERT(info->type == IND_STATIC);
1761 if (STATIC_LINK(info,p) == NULL) {
1762 debugTrace(DEBUG_gccafs, "CAF gc'd at 0x%04lx", (long)p);
1764 SET_INFO(p,&stg_BLACKHOLE_info);
1765 p = STATIC_LINK2(info,p);
1769 pp = &STATIC_LINK2(info,p);
1776 debugTrace(DEBUG_gccafs, "%d CAFs live", i);