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 g0_pcnt_kept = 30; // percentage of g0 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(gen_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 resize_generations (void);
144 static void resize_nursery (void);
145 static void start_gc_threads (void);
146 static void scavenge_until_all_done (void);
147 static StgWord inc_running (void);
148 static StgWord dec_running (void);
149 static void wakeup_gc_threads (nat n_threads, nat me);
150 static void shutdown_gc_threads (nat n_threads, nat me);
152 #if 0 && defined(DEBUG)
153 static void gcCAFs (void);
156 /* -----------------------------------------------------------------------------
158 -------------------------------------------------------------------------- */
160 bdescr *mark_stack_top_bd; // topmost block in the mark stack
161 bdescr *mark_stack_bd; // current block in the mark stack
162 StgPtr mark_sp; // pointer to the next unallocated mark stack entry
164 /* -----------------------------------------------------------------------------
165 GarbageCollect: the main entry point to the garbage collector.
167 Locks held: all capabilities are held throughout GarbageCollect().
168 -------------------------------------------------------------------------- */
171 GarbageCollect (rtsBool force_major_gc,
172 nat gc_type USED_IF_THREADS,
177 lnat live, allocated, max_copied, avg_copied, slop;
178 gc_thread *saved_gct;
181 // necessary if we stole a callee-saves register for gct:
185 CostCentreStack *prev_CCS;
190 #if defined(RTS_USER_SIGNALS)
191 if (RtsFlags.MiscFlags.install_signal_handlers) {
197 ASSERT(sizeof(gen_workspace) == 16 * sizeof(StgWord));
198 // otherwise adjust the padding in gen_workspace.
200 // tell the stats department that we've started a GC
203 // tell the STM to discard any cached closures it's hoping to re-use
206 // lock the StablePtr table
215 // attribute any costs to CCS_GC
221 /* Approximate how much we allocated.
222 * Todo: only when generating stats?
224 allocated = calcAllocated(rtsFalse/* don't count the nursery yet */);
226 /* Figure out which generation to collect
228 n = initialise_N(force_major_gc);
230 #if defined(THREADED_RTS)
231 work_stealing = RtsFlags.ParFlags.parGcLoadBalancingEnabled &&
232 N >= RtsFlags.ParFlags.parGcLoadBalancingGen;
233 // It's not always a good idea to do load balancing in parallel
234 // GC. In particular, for a parallel program we don't want to
235 // lose locality by moving cached data into another CPU's cache
236 // (this effect can be quite significant).
238 // We could have a more complex way to deterimine whether to do
239 // work stealing or not, e.g. it might be a good idea to do it
240 // if the heap is big. For now, we just turn it on or off with
244 /* Start threads, so they can be spinning up while we finish initialisation.
248 #if defined(THREADED_RTS)
249 /* How many threads will be participating in this GC?
250 * We don't try to parallelise minor GCs (unless the user asks for
251 * it with +RTS -gn0), or mark/compact/sweep GC.
253 if (gc_type == PENDING_GC_PAR) {
254 n_gc_threads = RtsFlags.ParFlags.nNodes;
262 debugTrace(DEBUG_gc, "GC (gen %d): %d KB to collect, %ld MB in use, using %d thread(s)",
263 N, n * (BLOCK_SIZE / 1024), mblocks_allocated, n_gc_threads);
265 #ifdef RTS_GTK_FRONTPANEL
266 if (RtsFlags.GcFlags.frontpanel) {
267 updateFrontPanelBeforeGC(N);
272 // check for memory leaks if DEBUG is on
273 memInventory(DEBUG_gc);
276 // check sanity *before* GC
277 IF_DEBUG(sanity, checkSanity(rtsTrue));
279 // Initialise all our gc_thread structures
280 for (t = 0; t < n_gc_threads; t++) {
281 init_gc_thread(gc_threads[t]);
284 // Initialise all the generations/steps that we're collecting.
285 for (g = 0; g <= N; g++) {
286 init_collected_gen(g,n_gc_threads);
289 // Initialise all the generations/steps that we're *not* collecting.
290 for (g = N+1; g < RtsFlags.GcFlags.generations; g++) {
291 init_uncollected_gen(g,n_gc_threads);
294 /* Allocate a mark stack if we're doing a major collection.
296 if (major_gc && oldest_gen->mark) {
297 mark_stack_bd = allocBlock();
298 mark_stack_top_bd = mark_stack_bd;
299 mark_stack_bd->link = NULL;
300 mark_stack_bd->u.back = NULL;
301 mark_sp = mark_stack_bd->start;
303 mark_stack_bd = NULL;
304 mark_stack_top_bd = NULL;
308 // this is the main thread
310 if (n_gc_threads == 1) {
311 SET_GCT(gc_threads[0]);
313 SET_GCT(gc_threads[cap->no]);
316 SET_GCT(gc_threads[0]);
319 /* -----------------------------------------------------------------------
320 * follow all the roots that we know about:
323 // the main thread is running: this prevents any other threads from
324 // exiting prematurely, so we can start them now.
325 // NB. do this after the mutable lists have been saved above, otherwise
326 // the other GC threads will be writing into the old mutable lists.
328 wakeup_gc_threads(n_gc_threads, gct->thread_index);
330 // scavenge the capability-private mutable lists. This isn't part
331 // of markSomeCapabilities() because markSomeCapabilities() can only
332 // call back into the GC via mark_root() (due to the gct register
334 if (n_gc_threads == 1) {
335 for (n = 0; n < n_capabilities; n++) {
336 #if defined(THREADED_RTS)
337 scavenge_capability_mut_Lists1(&capabilities[n]);
339 scavenge_capability_mut_lists(&capabilities[n]);
343 scavenge_capability_mut_lists(&capabilities[gct->thread_index]);
346 // follow roots from the CAF list (used by GHCi)
347 gct->evac_gen_no = 0;
348 markCAFs(mark_root, gct);
350 // follow all the roots that the application knows about.
351 gct->evac_gen_no = 0;
352 markSomeCapabilities(mark_root, gct, gct->thread_index, n_gc_threads,
353 rtsTrue/*prune sparks*/);
355 #if defined(RTS_USER_SIGNALS)
356 // mark the signal handlers (signals should be already blocked)
357 markSignalHandlers(mark_root, gct);
360 // Mark the weak pointer list, and prepare to detect dead weak pointers.
364 // Mark the stable pointer table.
365 markStablePtrTable(mark_root, gct);
367 /* -------------------------------------------------------------------------
368 * Repeatedly scavenge all the areas we know about until there's no
369 * more scavenging to be done.
373 scavenge_until_all_done();
374 // The other threads are now stopped. We might recurse back to
375 // here, but from now on this is the only thread.
377 // must be last... invariant is that everything is fully
378 // scavenged at this point.
379 if (traverseWeakPtrList()) { // returns rtsTrue if evaced something
384 // If we get to here, there's really nothing left to do.
388 shutdown_gc_threads(n_gc_threads, gct->thread_index);
390 // Now see which stable names are still alive.
394 if (n_gc_threads == 1) {
395 for (n = 0; n < n_capabilities; n++) {
396 pruneSparkQueue(&capabilities[n]);
399 pruneSparkQueue(&capabilities[gct->thread_index]);
404 // We call processHeapClosureForDead() on every closure destroyed during
405 // the current garbage collection, so we invoke LdvCensusForDead().
406 if (RtsFlags.ProfFlags.doHeapProfile == HEAP_BY_LDV
407 || RtsFlags.ProfFlags.bioSelector != NULL)
411 // NO MORE EVACUATION AFTER THIS POINT!
413 // Two-space collector: free the old to-space.
414 // g0->old_blocks is the old nursery
415 // g0->blocks is to-space from the previous GC
416 if (RtsFlags.GcFlags.generations == 1) {
417 if (g0->blocks != NULL) {
418 freeChain(g0->blocks);
423 // For each workspace, in each thread, move the copied blocks to the step
429 for (t = 0; t < n_gc_threads; t++) {
432 for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
435 // Push the final block
437 push_scanned_block(ws->todo_bd, ws);
440 ASSERT(gct->scan_bd == NULL);
441 ASSERT(countBlocks(ws->scavd_list) == ws->n_scavd_blocks);
444 for (bd = ws->scavd_list; bd != NULL; bd = bd->link) {
445 ws->gen->n_words += bd->free - bd->start;
449 prev->link = ws->gen->blocks;
450 ws->gen->blocks = ws->scavd_list;
452 ws->gen->n_blocks += ws->n_scavd_blocks;
456 // Add all the partial blocks *after* we've added all the full
457 // blocks. This is so that we can grab the partial blocks back
458 // again and try to fill them up in the next GC.
459 for (t = 0; t < n_gc_threads; t++) {
462 for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
466 for (bd = ws->part_list; bd != NULL; bd = next) {
468 if (bd->free == bd->start) {
470 ws->part_list = next;
477 ws->gen->n_words += bd->free - bd->start;
482 prev->link = ws->gen->blocks;
483 ws->gen->blocks = ws->part_list;
485 ws->gen->n_blocks += ws->n_part_blocks;
487 ASSERT(countBlocks(ws->gen->blocks) == ws->gen->n_blocks);
488 ASSERT(countOccupied(ws->gen->blocks) == ws->gen->n_words);
493 // Finally: compact or sweep the oldest generation.
494 if (major_gc && oldest_gen->mark) {
495 if (oldest_gen->compact)
496 compact(gct->scavenged_static_objects);
501 /* run through all the generations/steps and tidy up
508 for (i=0; i < n_gc_threads; i++) {
509 if (n_gc_threads > 1) {
510 debugTrace(DEBUG_gc,"thread %d:", i);
511 debugTrace(DEBUG_gc," copied %ld", gc_threads[i]->copied * sizeof(W_));
512 debugTrace(DEBUG_gc," scanned %ld", gc_threads[i]->scanned * sizeof(W_));
513 debugTrace(DEBUG_gc," any_work %ld", gc_threads[i]->any_work);
514 debugTrace(DEBUG_gc," no_work %ld", gc_threads[i]->no_work);
515 debugTrace(DEBUG_gc," scav_find_work %ld", gc_threads[i]->scav_find_work);
517 copied += gc_threads[i]->copied;
518 max_copied = stg_max(gc_threads[i]->copied, max_copied);
520 if (n_gc_threads == 1) {
528 for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
531 generations[g].collections++; // for stats
532 if (n_gc_threads > 1) generations[g].par_collections++;
535 // Count the mutable list as bytes "copied" for the purposes of
536 // stats. Every mutable list is copied during every GC.
538 nat mut_list_size = 0;
539 for (n = 0; n < n_capabilities; n++) {
540 mut_list_size += countOccupied(capabilities[n].mut_lists[g]);
542 copied += mut_list_size;
545 "mut_list_size: %lu (%d vars, %d arrays, %d MVARs, %d others)",
546 (unsigned long)(mut_list_size * sizeof(W_)),
547 mutlist_MUTVARS, mutlist_MUTARRS, mutlist_MVARS, mutlist_OTHERS);
551 gen = &generations[g];
553 // for generations we collected...
556 /* free old memory and shift to-space into from-space for all
557 * the collected steps (except the allocation area). These
558 * freed blocks will probaby be quickly recycled.
562 // tack the new blocks on the end of the existing blocks
563 if (gen->old_blocks != NULL) {
566 for (bd = gen->old_blocks; bd != NULL; bd = next) {
570 if (!(bd->flags & BF_MARKED))
573 gen->old_blocks = next;
582 gen->n_words += bd->free - bd->start;
584 // NB. this step might not be compacted next
585 // time, so reset the BF_MARKED flags.
586 // They are set before GC if we're going to
587 // compact. (search for BF_MARKED above).
588 bd->flags &= ~BF_MARKED;
590 // between GCs, all blocks in the heap except
591 // for the nursery have the BF_EVACUATED flag set.
592 bd->flags |= BF_EVACUATED;
599 prev->link = gen->blocks;
600 gen->blocks = gen->old_blocks;
603 // add the new blocks to the block tally
604 gen->n_blocks += gen->n_old_blocks;
605 ASSERT(countBlocks(gen->blocks) == gen->n_blocks);
606 ASSERT(countOccupied(gen->blocks) == gen->n_words);
610 freeChain(gen->old_blocks);
613 gen->old_blocks = NULL;
614 gen->n_old_blocks = 0;
616 /* LARGE OBJECTS. The current live large objects are chained on
617 * scavenged_large, having been moved during garbage
618 * collection from large_objects. Any objects left on the
619 * large_objects list are therefore dead, so we free them here.
621 freeChain(gen->large_objects);
622 gen->large_objects = gen->scavenged_large_objects;
623 gen->n_large_blocks = gen->n_scavenged_large_blocks;
624 gen->n_new_large_words = 0;
625 ASSERT(countBlocks(gen->large_objects) == gen->n_large_blocks);
627 else // for generations > N
629 /* For older generations, we need to append the
630 * scavenged_large_object list (i.e. large objects that have been
631 * promoted during this GC) to the large_object list for that step.
633 for (bd = gen->scavenged_large_objects; bd; bd = next) {
635 dbl_link_onto(bd, &gen->large_objects);
638 // add the new blocks we promoted during this GC
639 gen->n_large_blocks += gen->n_scavenged_large_blocks;
640 ASSERT(countBlocks(gen->large_objects) == gen->n_large_blocks);
642 } // for all generations
644 // update the max size of older generations after a major GC
645 resize_generations();
647 // Calculate the amount of live data for stats.
648 live = calcLiveWords();
650 // Start a new pinned_object_block
651 for (n = 0; n < n_capabilities; n++) {
652 capabilities[n].pinned_object_block = NULL;
655 // Free the mark stack.
656 if (mark_stack_top_bd != NULL) {
657 debugTrace(DEBUG_gc, "mark stack: %d blocks",
658 countBlocks(mark_stack_top_bd));
659 freeChain(mark_stack_top_bd);
663 for (g = 0; g <= N; g++) {
664 gen = &generations[g];
665 if (gen->bitmap != NULL) {
666 freeGroup(gen->bitmap);
671 // Reset the nursery: make the blocks empty
672 allocated += clearNurseries();
678 // mark the garbage collected CAFs as dead
679 #if 0 && defined(DEBUG) // doesn't work at the moment
680 if (major_gc) { gcCAFs(); }
684 // resetStaticObjectForRetainerProfiling() must be called before
686 if (n_gc_threads > 1) {
687 barf("profiling is currently broken with multi-threaded GC");
688 // ToDo: fix the gct->scavenged_static_objects below
690 resetStaticObjectForRetainerProfiling(gct->scavenged_static_objects);
693 // zero the scavenged static object list
696 for (i = 0; i < n_gc_threads; i++) {
697 zero_static_object_list(gc_threads[i]->scavenged_static_objects);
701 // send exceptions to any threads which were about to die
703 resurrectThreads(resurrected_threads);
706 // Update the stable pointer hash table.
707 updateStablePtrTable(major_gc);
709 // unlock the StablePtr table. Must be before scheduleFinalizers(),
710 // because a finalizer may call hs_free_fun_ptr() or
711 // hs_free_stable_ptr(), both of which access the StablePtr table.
714 // Start any pending finalizers. Must be after
715 // updateStablePtrTable() and stablePtrPostGC() (see #4221).
717 scheduleFinalizers(cap, old_weak_ptr_list);
722 need = BLOCKS_TO_MBLOCKS(n_alloc_blocks);
723 got = mblocks_allocated;
724 /* If the amount of data remains constant, next major GC we'll
725 require (F+1)*need. We leave (F+2)*need in order to reduce
726 repeated deallocation and reallocation. */
727 need = (RtsFlags.GcFlags.oldGenFactor + 2) * need;
729 returnMemoryToOS(got - need);
733 // check sanity after GC
734 IF_DEBUG(sanity, checkSanity(rtsTrue));
736 // extra GC trace info
737 IF_DEBUG(gc, statDescribeGens());
740 // symbol-table based profiling
741 /* heapCensus(to_blocks); */ /* ToDo */
744 // restore enclosing cost centre
750 // check for memory leaks if DEBUG is on
751 memInventory(DEBUG_gc);
754 #ifdef RTS_GTK_FRONTPANEL
755 if (RtsFlags.GcFlags.frontpanel) {
756 updateFrontPanelAfterGC( N, live );
760 // ok, GC over: tell the stats department what happened.
761 slop = calcLiveBlocks() * BLOCK_SIZE_W - live;
762 stat_endGC(allocated, live, copied, N, max_copied, avg_copied, slop);
764 // Guess which generation we'll collect *next* time
765 initialise_N(force_major_gc);
767 #if defined(RTS_USER_SIGNALS)
768 if (RtsFlags.MiscFlags.install_signal_handlers) {
769 // unblock signals again
770 unblockUserSignals();
779 /* -----------------------------------------------------------------------------
780 Figure out which generation to collect, initialise N and major_gc.
782 Also returns the total number of blocks in generations that will be
784 -------------------------------------------------------------------------- */
787 initialise_N (rtsBool force_major_gc)
790 nat blocks, blocks_total;
795 if (force_major_gc) {
796 N = RtsFlags.GcFlags.generations - 1;
801 for (g = RtsFlags.GcFlags.generations - 1; g >= 0; g--) {
803 blocks = generations[g].n_words / BLOCK_SIZE_W
804 + generations[g].n_large_blocks;
806 if (blocks >= generations[g].max_blocks) {
810 blocks_total += blocks;
814 blocks_total += countNurseryBlocks();
816 major_gc = (N == RtsFlags.GcFlags.generations-1);
820 /* -----------------------------------------------------------------------------
821 Initialise the gc_thread structures.
822 -------------------------------------------------------------------------- */
824 #define GC_THREAD_INACTIVE 0
825 #define GC_THREAD_STANDING_BY 1
826 #define GC_THREAD_RUNNING 2
827 #define GC_THREAD_WAITING_TO_CONTINUE 3
830 new_gc_thread (nat n, gc_thread *t)
837 initSpinLock(&t->gc_spin);
838 initSpinLock(&t->mut_spin);
839 ACQUIRE_SPIN_LOCK(&t->gc_spin);
840 t->wakeup = GC_THREAD_INACTIVE; // starts true, so we can wait for the
841 // thread to start up, see wakeup_gc_threads
845 t->free_blocks = NULL;
854 for (g = 0; g < RtsFlags.GcFlags.generations; g++)
857 ws->gen = &generations[g];
858 ASSERT(g == ws->gen->no);
862 ws->todo_q = newWSDeque(128);
863 ws->todo_overflow = NULL;
864 ws->n_todo_overflow = 0;
866 ws->part_list = NULL;
867 ws->n_part_blocks = 0;
869 ws->scavd_list = NULL;
870 ws->n_scavd_blocks = 0;
878 if (gc_threads == NULL) {
879 #if defined(THREADED_RTS)
881 gc_threads = stgMallocBytes (RtsFlags.ParFlags.nNodes *
885 for (i = 0; i < RtsFlags.ParFlags.nNodes; i++) {
887 stgMallocBytes(sizeof(gc_thread) +
888 RtsFlags.GcFlags.generations * sizeof(gen_workspace),
891 new_gc_thread(i, gc_threads[i]);
894 gc_threads = stgMallocBytes (sizeof(gc_thread*),"alloc_gc_threads");
896 new_gc_thread(0,gc_threads[0]);
905 if (gc_threads != NULL) {
906 #if defined(THREADED_RTS)
908 for (i = 0; i < n_capabilities; i++) {
909 for (g = 0; g < RtsFlags.GcFlags.generations; g++)
911 freeWSDeque(gc_threads[i]->gens[g].todo_q);
913 stgFree (gc_threads[i]);
915 stgFree (gc_threads);
917 for (g = 0; g < RtsFlags.GcFlags.generations; g++)
919 freeWSDeque(gc_threads[0]->gens[g].todo_q);
921 stgFree (gc_threads);
927 /* ----------------------------------------------------------------------------
929 ------------------------------------------------------------------------- */
931 static volatile StgWord gc_running_threads;
937 new = atomic_inc(&gc_running_threads);
938 ASSERT(new <= n_gc_threads);
945 ASSERT(gc_running_threads != 0);
946 return atomic_dec(&gc_running_threads);
959 // scavenge objects in compacted generation
960 if (mark_stack_bd != NULL && !mark_stack_empty()) {
964 // Check for global work in any step. We don't need to check for
965 // local work, because we have already exited scavenge_loop(),
966 // which means there is no local work for this thread.
967 for (g = 0; g < (int)RtsFlags.GcFlags.generations; g++) {
969 if (ws->todo_large_objects) return rtsTrue;
970 if (!looksEmptyWSDeque(ws->todo_q)) return rtsTrue;
971 if (ws->todo_overflow) return rtsTrue;
974 #if defined(THREADED_RTS)
977 // look for work to steal
978 for (n = 0; n < n_gc_threads; n++) {
979 if (n == gct->thread_index) continue;
980 for (g = RtsFlags.GcFlags.generations-1; g >= 0; g--) {
981 ws = &gc_threads[n]->gens[g];
982 if (!looksEmptyWSDeque(ws->todo_q)) return rtsTrue;
989 #if defined(THREADED_RTS)
997 scavenge_until_all_done (void)
1003 traceEventGcWork(&capabilities[gct->thread_index]);
1005 #if defined(THREADED_RTS)
1006 if (n_gc_threads > 1) {
1015 // scavenge_loop() only exits when there's no work to do
1018 traceEventGcIdle(&capabilities[gct->thread_index]);
1020 debugTrace(DEBUG_gc, "%d GC threads still running", r);
1022 while (gc_running_threads != 0) {
1028 // any_work() does not remove the work from the queue, it
1029 // just checks for the presence of work. If we find any,
1030 // then we increment gc_running_threads and go back to
1031 // scavenge_loop() to perform any pending work.
1034 traceEventGcDone(&capabilities[gct->thread_index]);
1037 #if defined(THREADED_RTS)
1040 gcWorkerThread (Capability *cap)
1042 gc_thread *saved_gct;
1044 // necessary if we stole a callee-saves register for gct:
1047 gct = gc_threads[cap->no];
1048 gct->id = osThreadId();
1050 // Wait until we're told to wake up
1051 RELEASE_SPIN_LOCK(&gct->mut_spin);
1052 gct->wakeup = GC_THREAD_STANDING_BY;
1053 debugTrace(DEBUG_gc, "GC thread %d standing by...", gct->thread_index);
1054 ACQUIRE_SPIN_LOCK(&gct->gc_spin);
1057 // start performance counters in this thread...
1058 if (gct->papi_events == -1) {
1059 papi_init_eventset(&gct->papi_events);
1061 papi_thread_start_gc1_count(gct->papi_events);
1064 // Every thread evacuates some roots.
1065 gct->evac_gen_no = 0;
1066 markSomeCapabilities(mark_root, gct, gct->thread_index, n_gc_threads,
1067 rtsTrue/*prune sparks*/);
1068 scavenge_capability_mut_lists(&capabilities[gct->thread_index]);
1070 scavenge_until_all_done();
1073 // Now that the whole heap is marked, we discard any sparks that
1074 // were found to be unreachable. The main GC thread is currently
1075 // marking heap reachable via weak pointers, so it is
1076 // non-deterministic whether a spark will be retained if it is
1077 // only reachable via weak pointers. To fix this problem would
1078 // require another GC barrier, which is too high a price.
1079 pruneSparkQueue(cap);
1083 // count events in this thread towards the GC totals
1084 papi_thread_stop_gc1_count(gct->papi_events);
1087 // Wait until we're told to continue
1088 RELEASE_SPIN_LOCK(&gct->gc_spin);
1089 gct->wakeup = GC_THREAD_WAITING_TO_CONTINUE;
1090 debugTrace(DEBUG_gc, "GC thread %d waiting to continue...",
1092 ACQUIRE_SPIN_LOCK(&gct->mut_spin);
1093 debugTrace(DEBUG_gc, "GC thread %d on my way...", gct->thread_index);
1100 #if defined(THREADED_RTS)
1103 waitForGcThreads (Capability *cap USED_IF_THREADS)
1105 const nat n_threads = RtsFlags.ParFlags.nNodes;
1106 const nat me = cap->no;
1108 rtsBool retry = rtsTrue;
1111 for (i=0; i < n_threads; i++) {
1112 if (i == me) continue;
1113 if (gc_threads[i]->wakeup != GC_THREAD_STANDING_BY) {
1114 prodCapability(&capabilities[i], cap->running_task);
1117 for (j=0; j < 10; j++) {
1119 for (i=0; i < n_threads; i++) {
1120 if (i == me) continue;
1122 setContextSwitches();
1123 if (gc_threads[i]->wakeup != GC_THREAD_STANDING_BY) {
1133 #endif // THREADED_RTS
1136 start_gc_threads (void)
1138 #if defined(THREADED_RTS)
1139 gc_running_threads = 0;
1144 wakeup_gc_threads (nat n_threads USED_IF_THREADS, nat me USED_IF_THREADS)
1146 #if defined(THREADED_RTS)
1148 for (i=0; i < n_threads; i++) {
1149 if (i == me) continue;
1151 debugTrace(DEBUG_gc, "waking up gc thread %d", i);
1152 if (gc_threads[i]->wakeup != GC_THREAD_STANDING_BY) barf("wakeup_gc_threads");
1154 gc_threads[i]->wakeup = GC_THREAD_RUNNING;
1155 ACQUIRE_SPIN_LOCK(&gc_threads[i]->mut_spin);
1156 RELEASE_SPIN_LOCK(&gc_threads[i]->gc_spin);
1161 // After GC is complete, we must wait for all GC threads to enter the
1162 // standby state, otherwise they may still be executing inside
1163 // any_work(), and may even remain awake until the next GC starts.
1165 shutdown_gc_threads (nat n_threads USED_IF_THREADS, nat me USED_IF_THREADS)
1167 #if defined(THREADED_RTS)
1169 for (i=0; i < n_threads; i++) {
1170 if (i == me) continue;
1171 while (gc_threads[i]->wakeup != GC_THREAD_WAITING_TO_CONTINUE) { write_barrier(); }
1176 #if defined(THREADED_RTS)
1178 releaseGCThreads (Capability *cap USED_IF_THREADS)
1180 const nat n_threads = RtsFlags.ParFlags.nNodes;
1181 const nat me = cap->no;
1183 for (i=0; i < n_threads; i++) {
1184 if (i == me) continue;
1185 if (gc_threads[i]->wakeup != GC_THREAD_WAITING_TO_CONTINUE)
1186 barf("releaseGCThreads");
1188 gc_threads[i]->wakeup = GC_THREAD_INACTIVE;
1189 ACQUIRE_SPIN_LOCK(&gc_threads[i]->gc_spin);
1190 RELEASE_SPIN_LOCK(&gc_threads[i]->mut_spin);
1195 /* ----------------------------------------------------------------------------
1196 Initialise a generation that is to be collected
1197 ------------------------------------------------------------------------- */
1200 init_collected_gen (nat g, nat n_threads)
1207 // Throw away the current mutable list. Invariant: the mutable
1208 // list always has at least one block; this means we can avoid a
1209 // check for NULL in recordMutable().
1211 for (i = 0; i < n_capabilities; i++) {
1212 freeChain(capabilities[i].mut_lists[g]);
1213 capabilities[i].mut_lists[g] = allocBlock();
1217 gen = &generations[g];
1218 ASSERT(gen->no == g);
1220 // we'll construct a new list of threads in this step
1221 // during GC, throw away the current list.
1222 gen->old_threads = gen->threads;
1223 gen->threads = END_TSO_QUEUE;
1225 // deprecate the existing blocks
1226 gen->old_blocks = gen->blocks;
1227 gen->n_old_blocks = gen->n_blocks;
1231 gen->live_estimate = 0;
1233 // initialise the large object queues.
1234 gen->scavenged_large_objects = NULL;
1235 gen->n_scavenged_large_blocks = 0;
1237 // mark the small objects as from-space
1238 for (bd = gen->old_blocks; bd; bd = bd->link) {
1239 bd->flags &= ~BF_EVACUATED;
1242 // mark the large objects as from-space
1243 for (bd = gen->large_objects; bd; bd = bd->link) {
1244 bd->flags &= ~BF_EVACUATED;
1247 // for a compacted generation, we need to allocate the bitmap
1249 nat bitmap_size; // in bytes
1250 bdescr *bitmap_bdescr;
1253 bitmap_size = gen->n_old_blocks * BLOCK_SIZE / (sizeof(W_)*BITS_PER_BYTE);
1255 if (bitmap_size > 0) {
1256 bitmap_bdescr = allocGroup((lnat)BLOCK_ROUND_UP(bitmap_size)
1258 gen->bitmap = bitmap_bdescr;
1259 bitmap = bitmap_bdescr->start;
1261 debugTrace(DEBUG_gc, "bitmap_size: %d, bitmap: %p",
1262 bitmap_size, bitmap);
1264 // don't forget to fill it with zeros!
1265 memset(bitmap, 0, bitmap_size);
1267 // For each block in this step, point to its bitmap from the
1268 // block descriptor.
1269 for (bd=gen->old_blocks; bd != NULL; bd = bd->link) {
1270 bd->u.bitmap = bitmap;
1271 bitmap += BLOCK_SIZE_W / (sizeof(W_)*BITS_PER_BYTE);
1273 // Also at this point we set the BF_MARKED flag
1274 // for this block. The invariant is that
1275 // BF_MARKED is always unset, except during GC
1276 // when it is set on those blocks which will be
1278 if (!(bd->flags & BF_FRAGMENTED)) {
1279 bd->flags |= BF_MARKED;
1282 // BF_SWEPT should be marked only for blocks that are being
1283 // collected in sweep()
1284 bd->flags &= ~BF_SWEPT;
1289 // For each GC thread, for each step, allocate a "todo" block to
1290 // store evacuated objects to be scavenged, and a block to store
1291 // evacuated objects that do not need to be scavenged.
1292 for (t = 0; t < n_threads; t++) {
1293 ws = &gc_threads[t]->gens[g];
1295 ws->todo_large_objects = NULL;
1297 ws->part_list = NULL;
1298 ws->n_part_blocks = 0;
1300 // allocate the first to-space block; extra blocks will be
1301 // chained on as necessary.
1303 ASSERT(looksEmptyWSDeque(ws->todo_q));
1304 alloc_todo_block(ws,0);
1306 ws->todo_overflow = NULL;
1307 ws->n_todo_overflow = 0;
1309 ws->scavd_list = NULL;
1310 ws->n_scavd_blocks = 0;
1315 /* ----------------------------------------------------------------------------
1316 Initialise a generation that is *not* to be collected
1317 ------------------------------------------------------------------------- */
1320 init_uncollected_gen (nat g, nat threads)
1327 // save the current mutable lists for this generation, and
1328 // allocate a fresh block for each one. We'll traverse these
1329 // mutable lists as roots early on in the GC.
1330 for (n = 0; n < n_capabilities; n++) {
1331 capabilities[n].saved_mut_lists[g] = capabilities[n].mut_lists[g];
1332 capabilities[n].mut_lists[g] = allocBlock();
1335 gen = &generations[g];
1337 gen->scavenged_large_objects = NULL;
1338 gen->n_scavenged_large_blocks = 0;
1340 for (t = 0; t < threads; t++) {
1341 ws = &gc_threads[t]->gens[g];
1343 ASSERT(looksEmptyWSDeque(ws->todo_q));
1344 ws->todo_large_objects = NULL;
1346 ws->part_list = NULL;
1347 ws->n_part_blocks = 0;
1349 ws->scavd_list = NULL;
1350 ws->n_scavd_blocks = 0;
1352 // If the block at the head of the list in this generation
1353 // is less than 3/4 full, then use it as a todo block.
1354 if (gen->blocks && isPartiallyFull(gen->blocks))
1356 ws->todo_bd = gen->blocks;
1357 ws->todo_free = ws->todo_bd->free;
1358 ws->todo_lim = ws->todo_bd->start + BLOCK_SIZE_W;
1359 gen->blocks = gen->blocks->link;
1361 gen->n_words -= ws->todo_bd->free - ws->todo_bd->start;
1362 ws->todo_bd->link = NULL;
1363 // we must scan from the current end point.
1364 ws->todo_bd->u.scan = ws->todo_bd->free;
1369 alloc_todo_block(ws,0);
1373 // deal out any more partial blocks to the threads' part_lists
1375 while (gen->blocks && isPartiallyFull(gen->blocks))
1378 gen->blocks = bd->link;
1379 ws = &gc_threads[t]->gens[g];
1380 bd->link = ws->part_list;
1382 ws->n_part_blocks += 1;
1383 bd->u.scan = bd->free;
1385 gen->n_words -= bd->free - bd->start;
1387 if (t == n_gc_threads) t = 0;
1391 /* -----------------------------------------------------------------------------
1392 Initialise a gc_thread before GC
1393 -------------------------------------------------------------------------- */
1396 init_gc_thread (gc_thread *t)
1398 t->static_objects = END_OF_STATIC_LIST;
1399 t->scavenged_static_objects = END_OF_STATIC_LIST;
1401 t->mut_lists = capabilities[t->thread_index].mut_lists;
1403 t->failed_to_evac = rtsFalse;
1404 t->eager_promotion = rtsTrue;
1405 t->thunk_selector_depth = 0;
1410 t->scav_find_work = 0;
1413 /* -----------------------------------------------------------------------------
1414 Function we pass to evacuate roots.
1415 -------------------------------------------------------------------------- */
1418 mark_root(void *user USED_IF_THREADS, StgClosure **root)
1420 // we stole a register for gct, but this function is called from
1421 // *outside* the GC where the register variable is not in effect,
1422 // so we need to save and restore it here. NB. only call
1423 // mark_root() from the main GC thread, otherwise gct will be
1425 gc_thread *saved_gct;
1434 /* -----------------------------------------------------------------------------
1435 Initialising the static object & mutable lists
1436 -------------------------------------------------------------------------- */
1439 zero_static_object_list(StgClosure* first_static)
1443 const StgInfoTable *info;
1445 for (p = first_static; p != END_OF_STATIC_LIST; p = link) {
1447 link = *STATIC_LINK(info, p);
1448 *STATIC_LINK(info,p) = NULL;
1452 /* ----------------------------------------------------------------------------
1453 Reset the sizes of the older generations when we do a major
1456 CURRENT STRATEGY: make all generations except zero the same size.
1457 We have to stay within the maximum heap size, and leave a certain
1458 percentage of the maximum heap size available to allocate into.
1459 ------------------------------------------------------------------------- */
1462 resize_generations (void)
1466 if (major_gc && RtsFlags.GcFlags.generations > 1) {
1467 nat live, size, min_alloc, words;
1468 const nat max = RtsFlags.GcFlags.maxHeapSize;
1469 const nat gens = RtsFlags.GcFlags.generations;
1471 // live in the oldest generations
1472 if (oldest_gen->live_estimate != 0) {
1473 words = oldest_gen->live_estimate;
1475 words = oldest_gen->n_words;
1477 live = (words + BLOCK_SIZE_W - 1) / BLOCK_SIZE_W +
1478 oldest_gen->n_large_blocks;
1480 // default max size for all generations except zero
1481 size = stg_max(live * RtsFlags.GcFlags.oldGenFactor,
1482 RtsFlags.GcFlags.minOldGenSize);
1484 if (RtsFlags.GcFlags.heapSizeSuggestionAuto) {
1485 RtsFlags.GcFlags.heapSizeSuggestion = size;
1488 // minimum size for generation zero
1489 min_alloc = stg_max((RtsFlags.GcFlags.pcFreeHeap * max) / 200,
1490 RtsFlags.GcFlags.minAllocAreaSize);
1492 // Auto-enable compaction when the residency reaches a
1493 // certain percentage of the maximum heap size (default: 30%).
1494 if (RtsFlags.GcFlags.compact ||
1496 oldest_gen->n_blocks >
1497 (RtsFlags.GcFlags.compactThreshold * max) / 100)) {
1498 oldest_gen->mark = 1;
1499 oldest_gen->compact = 1;
1500 // debugBelch("compaction: on\n", live);
1502 oldest_gen->mark = 0;
1503 oldest_gen->compact = 0;
1504 // debugBelch("compaction: off\n", live);
1507 if (RtsFlags.GcFlags.sweep) {
1508 oldest_gen->mark = 1;
1511 // if we're going to go over the maximum heap size, reduce the
1512 // size of the generations accordingly. The calculation is
1513 // different if compaction is turned on, because we don't need
1514 // to double the space required to collect the old generation.
1517 // this test is necessary to ensure that the calculations
1518 // below don't have any negative results - we're working
1519 // with unsigned values here.
1520 if (max < min_alloc) {
1524 if (oldest_gen->compact) {
1525 if ( (size + (size - 1) * (gens - 2) * 2) + min_alloc > max ) {
1526 size = (max - min_alloc) / ((gens - 1) * 2 - 1);
1529 if ( (size * (gens - 1) * 2) + min_alloc > max ) {
1530 size = (max - min_alloc) / ((gens - 1) * 2);
1540 debugBelch("live: %d, min_alloc: %d, size : %d, max = %d\n", live,
1541 min_alloc, size, max);
1544 for (g = 0; g < gens; g++) {
1545 generations[g].max_blocks = size;
1550 /* -----------------------------------------------------------------------------
1551 Calculate the new size of the nursery, and resize it.
1552 -------------------------------------------------------------------------- */
1555 resize_nursery (void)
1557 const lnat min_nursery = RtsFlags.GcFlags.minAllocAreaSize * n_capabilities;
1559 if (RtsFlags.GcFlags.generations == 1)
1560 { // Two-space collector:
1563 /* set up a new nursery. Allocate a nursery size based on a
1564 * function of the amount of live data (by default a factor of 2)
1565 * Use the blocks from the old nursery if possible, freeing up any
1568 * If we get near the maximum heap size, then adjust our nursery
1569 * size accordingly. If the nursery is the same size as the live
1570 * data (L), then we need 3L bytes. We can reduce the size of the
1571 * nursery to bring the required memory down near 2L bytes.
1573 * A normal 2-space collector would need 4L bytes to give the same
1574 * performance we get from 3L bytes, reducing to the same
1575 * performance at 2L bytes.
1577 blocks = generations[0].n_blocks;
1579 if ( RtsFlags.GcFlags.maxHeapSize != 0 &&
1580 blocks * RtsFlags.GcFlags.oldGenFactor * 2 >
1581 RtsFlags.GcFlags.maxHeapSize )
1583 long adjusted_blocks; // signed on purpose
1586 adjusted_blocks = (RtsFlags.GcFlags.maxHeapSize - 2 * blocks);
1588 debugTrace(DEBUG_gc, "near maximum heap size of 0x%x blocks, blocks = %d, adjusted to %ld",
1589 RtsFlags.GcFlags.maxHeapSize, blocks, adjusted_blocks);
1591 pc_free = adjusted_blocks * 100 / RtsFlags.GcFlags.maxHeapSize;
1592 if (pc_free < RtsFlags.GcFlags.pcFreeHeap) /* might even * be < 0 */
1596 blocks = adjusted_blocks;
1600 blocks *= RtsFlags.GcFlags.oldGenFactor;
1601 if (blocks < min_nursery)
1603 blocks = min_nursery;
1606 resizeNurseries(blocks);
1608 else // Generational collector
1611 * If the user has given us a suggested heap size, adjust our
1612 * allocation area to make best use of the memory available.
1614 if (RtsFlags.GcFlags.heapSizeSuggestion)
1617 const nat needed = calcNeeded(); // approx blocks needed at next GC
1619 /* Guess how much will be live in generation 0 step 0 next time.
1620 * A good approximation is obtained by finding the
1621 * percentage of g0 that was live at the last minor GC.
1623 * We have an accurate figure for the amount of copied data in
1624 * 'copied', but we must convert this to a number of blocks, with
1625 * a small adjustment for estimated slop at the end of a block
1630 g0_pcnt_kept = ((copied / (BLOCK_SIZE_W - 10)) * 100)
1631 / countNurseryBlocks();
1634 /* Estimate a size for the allocation area based on the
1635 * information available. We might end up going slightly under
1636 * or over the suggested heap size, but we should be pretty
1639 * Formula: suggested - needed
1640 * ----------------------------
1641 * 1 + g0_pcnt_kept/100
1643 * where 'needed' is the amount of memory needed at the next
1644 * collection for collecting all gens except g0.
1647 (((long)RtsFlags.GcFlags.heapSizeSuggestion - (long)needed) * 100) /
1648 (100 + (long)g0_pcnt_kept);
1650 if (blocks < (long)min_nursery) {
1651 blocks = min_nursery;
1654 resizeNurseries((nat)blocks);
1658 // we might have added extra large blocks to the nursery, so
1659 // resize back to minAllocAreaSize again.
1660 resizeNurseriesFixed(RtsFlags.GcFlags.minAllocAreaSize);
1665 /* -----------------------------------------------------------------------------
1666 Sanity code for CAF garbage collection.
1668 With DEBUG turned on, we manage a CAF list in addition to the SRT
1669 mechanism. After GC, we run down the CAF list and blackhole any
1670 CAFs which have been garbage collected. This means we get an error
1671 whenever the program tries to enter a garbage collected CAF.
1673 Any garbage collected CAFs are taken off the CAF list at the same
1675 -------------------------------------------------------------------------- */
1677 #if 0 && defined(DEBUG)
1684 const StgInfoTable *info;
1695 ASSERT(info->type == IND_STATIC);
1697 if (STATIC_LINK(info,p) == NULL) {
1698 debugTrace(DEBUG_gccafs, "CAF gc'd at 0x%04lx", (long)p);
1700 SET_INFO(p,&stg_BLACKHOLE_info);
1701 p = STATIC_LINK2(info,p);
1705 pp = &STATIC_LINK2(info,p);
1712 debugTrace(DEBUG_gccafs, "%d CAFs live", i);