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();
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 // Mutable lists from each generation > N
331 // we want to *scavenge* these roots, not evacuate them: they're not
332 // going to move in this GC.
333 // Also do them in reverse generation order, for the usual reason:
334 // namely to reduce the likelihood of spurious old->new pointers.
336 for (g = RtsFlags.GcFlags.generations-1; g > N; g--) {
337 #if defined(THREADED_RTS)
338 if (n_gc_threads > 1) {
339 scavenge_mutable_list(generations[g].saved_mut_list, &generations[g]);
341 scavenge_mutable_list1(generations[g].saved_mut_list, &generations[g]);
344 scavenge_mutable_list(generations[g].saved_mut_list, &generations[g]);
346 freeChain_sync(generations[g].saved_mut_list);
347 generations[g].saved_mut_list = NULL;
351 // scavenge the capability-private mutable lists. This isn't part
352 // of markSomeCapabilities() because markSomeCapabilities() can only
353 // call back into the GC via mark_root() (due to the gct register
355 if (n_gc_threads == 1) {
356 for (n = 0; n < n_capabilities; n++) {
357 #if defined(THREADED_RTS)
358 scavenge_capability_mut_Lists1(&capabilities[n]);
360 scavenge_capability_mut_lists(&capabilities[n]);
364 scavenge_capability_mut_lists(&capabilities[gct->thread_index]);
367 // follow roots from the CAF list (used by GHCi)
369 markCAFs(mark_root, gct);
371 // follow all the roots that the application knows about.
373 markSomeCapabilities(mark_root, gct, gct->thread_index, n_gc_threads,
374 rtsTrue/*prune sparks*/);
376 #if defined(RTS_USER_SIGNALS)
377 // mark the signal handlers (signals should be already blocked)
378 markSignalHandlers(mark_root, gct);
381 // Mark the weak pointer list, and prepare to detect dead weak pointers.
385 // Mark the stable pointer table.
386 markStablePtrTable(mark_root, gct);
388 /* -------------------------------------------------------------------------
389 * Repeatedly scavenge all the areas we know about until there's no
390 * more scavenging to be done.
394 scavenge_until_all_done();
395 // The other threads are now stopped. We might recurse back to
396 // here, but from now on this is the only thread.
398 // must be last... invariant is that everything is fully
399 // scavenged at this point.
400 if (traverseWeakPtrList()) { // returns rtsTrue if evaced something
405 // If we get to here, there's really nothing left to do.
409 shutdown_gc_threads(n_gc_threads, gct->thread_index);
411 // Now see which stable names are still alive.
415 if (n_gc_threads == 1) {
416 for (n = 0; n < n_capabilities; n++) {
417 pruneSparkQueue(&capabilities[n]);
420 pruneSparkQueue(&capabilities[gct->thread_index]);
425 // We call processHeapClosureForDead() on every closure destroyed during
426 // the current garbage collection, so we invoke LdvCensusForDead().
427 if (RtsFlags.ProfFlags.doHeapProfile == HEAP_BY_LDV
428 || RtsFlags.ProfFlags.bioSelector != NULL)
432 // NO MORE EVACUATION AFTER THIS POINT!
434 // Two-space collector: free the old to-space.
435 // g0->old_blocks is the old nursery
436 // g0->blocks is to-space from the previous GC
437 if (RtsFlags.GcFlags.generations == 1) {
438 if (g0->blocks != NULL) {
439 freeChain(g0->blocks);
444 // For each workspace, in each thread, move the copied blocks to the step
450 for (t = 0; t < n_gc_threads; t++) {
453 for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
456 // Push the final block
458 push_scanned_block(ws->todo_bd, ws);
461 ASSERT(gct->scan_bd == NULL);
462 ASSERT(countBlocks(ws->scavd_list) == ws->n_scavd_blocks);
465 for (bd = ws->scavd_list; bd != NULL; bd = bd->link) {
466 ws->gen->n_words += bd->free - bd->start;
470 prev->link = ws->gen->blocks;
471 ws->gen->blocks = ws->scavd_list;
473 ws->gen->n_blocks += ws->n_scavd_blocks;
477 // Add all the partial blocks *after* we've added all the full
478 // blocks. This is so that we can grab the partial blocks back
479 // again and try to fill them up in the next GC.
480 for (t = 0; t < n_gc_threads; t++) {
483 for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
487 for (bd = ws->part_list; bd != NULL; bd = next) {
489 if (bd->free == bd->start) {
491 ws->part_list = next;
498 ws->gen->n_words += bd->free - bd->start;
503 prev->link = ws->gen->blocks;
504 ws->gen->blocks = ws->part_list;
506 ws->gen->n_blocks += ws->n_part_blocks;
508 ASSERT(countBlocks(ws->gen->blocks) == ws->gen->n_blocks);
509 ASSERT(countOccupied(ws->gen->blocks) == ws->gen->n_words);
514 // Finally: compact or sweep the oldest generation.
515 if (major_gc && oldest_gen->mark) {
516 if (oldest_gen->compact)
517 compact(gct->scavenged_static_objects);
522 /* run through all the generations/steps and tidy up
529 for (i=0; i < n_gc_threads; i++) {
530 if (n_gc_threads > 1) {
531 debugTrace(DEBUG_gc,"thread %d:", i);
532 debugTrace(DEBUG_gc," copied %ld", gc_threads[i]->copied * sizeof(W_));
533 debugTrace(DEBUG_gc," scanned %ld", gc_threads[i]->scanned * sizeof(W_));
534 debugTrace(DEBUG_gc," any_work %ld", gc_threads[i]->any_work);
535 debugTrace(DEBUG_gc," no_work %ld", gc_threads[i]->no_work);
536 debugTrace(DEBUG_gc," scav_find_work %ld", gc_threads[i]->scav_find_work);
538 copied += gc_threads[i]->copied;
539 max_copied = stg_max(gc_threads[i]->copied, max_copied);
541 if (n_gc_threads == 1) {
549 for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
552 generations[g].collections++; // for stats
553 if (n_gc_threads > 1) generations[g].par_collections++;
556 // Count the mutable list as bytes "copied" for the purposes of
557 // stats. Every mutable list is copied during every GC.
559 nat mut_list_size = 0;
560 for (bd = generations[g].mut_list; bd != NULL; bd = bd->link) {
561 mut_list_size += bd->free - bd->start;
563 for (n = 0; n < n_capabilities; n++) {
564 for (bd = capabilities[n].mut_lists[g];
565 bd != NULL; bd = bd->link) {
566 mut_list_size += bd->free - bd->start;
569 copied += mut_list_size;
572 "mut_list_size: %lu (%d vars, %d arrays, %d MVARs, %d others)",
573 (unsigned long)(mut_list_size * sizeof(W_)),
574 mutlist_MUTVARS, mutlist_MUTARRS, mutlist_MVARS, mutlist_OTHERS);
578 gen = &generations[g];
580 // for generations we collected...
583 /* free old memory and shift to-space into from-space for all
584 * the collected steps (except the allocation area). These
585 * freed blocks will probaby be quickly recycled.
589 // tack the new blocks on the end of the existing blocks
590 if (gen->old_blocks != NULL) {
593 for (bd = gen->old_blocks; bd != NULL; bd = next) {
597 if (!(bd->flags & BF_MARKED))
600 gen->old_blocks = next;
609 gen->n_words += bd->free - bd->start;
611 // NB. this step might not be compacted next
612 // time, so reset the BF_MARKED flags.
613 // They are set before GC if we're going to
614 // compact. (search for BF_MARKED above).
615 bd->flags &= ~BF_MARKED;
617 // between GCs, all blocks in the heap except
618 // for the nursery have the BF_EVACUATED flag set.
619 bd->flags |= BF_EVACUATED;
626 prev->link = gen->blocks;
627 gen->blocks = gen->old_blocks;
630 // add the new blocks to the block tally
631 gen->n_blocks += gen->n_old_blocks;
632 ASSERT(countBlocks(gen->blocks) == gen->n_blocks);
633 ASSERT(countOccupied(gen->blocks) == gen->n_words);
637 freeChain(gen->old_blocks);
640 gen->old_blocks = NULL;
641 gen->n_old_blocks = 0;
643 /* LARGE OBJECTS. The current live large objects are chained on
644 * scavenged_large, having been moved during garbage
645 * collection from large_objects. Any objects left on the
646 * large_objects list are therefore dead, so we free them here.
648 freeChain(gen->large_objects);
649 gen->large_objects = gen->scavenged_large_objects;
650 gen->n_large_blocks = gen->n_scavenged_large_blocks;
651 gen->n_new_large_blocks = 0;
652 ASSERT(countBlocks(gen->large_objects) == gen->n_large_blocks);
654 else // for generations > N
656 /* For older generations, we need to append the
657 * scavenged_large_object list (i.e. large objects that have been
658 * promoted during this GC) to the large_object list for that step.
660 for (bd = gen->scavenged_large_objects; bd; bd = next) {
662 dbl_link_onto(bd, &gen->large_objects);
665 // add the new blocks we promoted during this GC
666 gen->n_large_blocks += gen->n_scavenged_large_blocks;
667 ASSERT(countBlocks(gen->large_objects) == gen->n_large_blocks);
669 } // for all generations
671 // update the max size of older generations after a major GC
672 resize_generations();
674 // Calculate the amount of live data for stats.
675 live = calcLiveWords();
677 // Free the small objects allocated via allocate(), since this will
678 // all have been copied into G0S1 now.
679 alloc_blocks_lim = RtsFlags.GcFlags.minAllocAreaSize;
681 // Start a new pinned_object_block
682 for (n = 0; n < n_capabilities; n++) {
683 capabilities[n].pinned_object_block = NULL;
686 // Free the mark stack.
687 if (mark_stack_top_bd != NULL) {
688 debugTrace(DEBUG_gc, "mark stack: %d blocks",
689 countBlocks(mark_stack_top_bd));
690 freeChain(mark_stack_top_bd);
694 for (g = 0; g <= N; g++) {
695 gen = &generations[g];
696 if (gen->bitmap != NULL) {
697 freeGroup(gen->bitmap);
704 // mark the garbage collected CAFs as dead
705 #if 0 && defined(DEBUG) // doesn't work at the moment
706 if (major_gc) { gcCAFs(); }
710 // resetStaticObjectForRetainerProfiling() must be called before
712 if (n_gc_threads > 1) {
713 barf("profiling is currently broken with multi-threaded GC");
714 // ToDo: fix the gct->scavenged_static_objects below
716 resetStaticObjectForRetainerProfiling(gct->scavenged_static_objects);
719 // zero the scavenged static object list
722 for (i = 0; i < n_gc_threads; i++) {
723 zero_static_object_list(gc_threads[i]->scavenged_static_objects);
730 // send exceptions to any threads which were about to die
732 resurrectThreads(resurrected_threads);
735 // Update the stable pointer hash table.
736 updateStablePtrTable(major_gc);
738 // unlock the StablePtr table. Must be before scheduleFinalizers(),
739 // because a finalizer may call hs_free_fun_ptr() or
740 // hs_free_stable_ptr(), both of which access the StablePtr table.
743 // Start any pending finalizers. Must be after
744 // updateStablePtrTable() and stablePtrPostGC() (see #4221).
746 scheduleFinalizers(cap, old_weak_ptr_list);
749 // check sanity after GC
750 IF_DEBUG(sanity, checkSanity(rtsTrue));
752 // extra GC trace info
753 IF_DEBUG(gc, statDescribeGens());
756 // symbol-table based profiling
757 /* heapCensus(to_blocks); */ /* ToDo */
760 // restore enclosing cost centre
766 // check for memory leaks if DEBUG is on
767 memInventory(DEBUG_gc);
770 #ifdef RTS_GTK_FRONTPANEL
771 if (RtsFlags.GcFlags.frontpanel) {
772 updateFrontPanelAfterGC( N, live );
776 // ok, GC over: tell the stats department what happened.
777 slop = calcLiveBlocks() * BLOCK_SIZE_W - live;
778 stat_endGC(allocated, live, copied, N, max_copied, avg_copied, slop);
780 // Guess which generation we'll collect *next* time
781 initialise_N(force_major_gc);
783 #if defined(RTS_USER_SIGNALS)
784 if (RtsFlags.MiscFlags.install_signal_handlers) {
785 // unblock signals again
786 unblockUserSignals();
795 /* -----------------------------------------------------------------------------
796 Figure out which generation to collect, initialise N and major_gc.
798 Also returns the total number of blocks in generations that will be
800 -------------------------------------------------------------------------- */
803 initialise_N (rtsBool force_major_gc)
806 nat blocks, blocks_total;
811 if (force_major_gc) {
812 N = RtsFlags.GcFlags.generations - 1;
817 for (g = RtsFlags.GcFlags.generations - 1; g >= 0; g--) {
819 blocks = generations[g].n_words / BLOCK_SIZE_W
820 + generations[g].n_large_blocks;
822 if (blocks >= generations[g].max_blocks) {
826 blocks_total += blocks;
830 blocks_total += countNurseryBlocks();
832 major_gc = (N == RtsFlags.GcFlags.generations-1);
836 /* -----------------------------------------------------------------------------
837 Initialise the gc_thread structures.
838 -------------------------------------------------------------------------- */
840 #define GC_THREAD_INACTIVE 0
841 #define GC_THREAD_STANDING_BY 1
842 #define GC_THREAD_RUNNING 2
843 #define GC_THREAD_WAITING_TO_CONTINUE 3
846 new_gc_thread (nat n, gc_thread *t)
853 initSpinLock(&t->gc_spin);
854 initSpinLock(&t->mut_spin);
855 ACQUIRE_SPIN_LOCK(&t->gc_spin);
856 t->wakeup = GC_THREAD_INACTIVE; // starts true, so we can wait for the
857 // thread to start up, see wakeup_gc_threads
861 t->free_blocks = NULL;
870 for (g = 0; g < RtsFlags.GcFlags.generations; g++)
873 ws->gen = &generations[g];
874 ASSERT(g == ws->gen->no);
878 ws->todo_q = newWSDeque(128);
879 ws->todo_overflow = NULL;
880 ws->n_todo_overflow = 0;
882 ws->part_list = NULL;
883 ws->n_part_blocks = 0;
885 ws->scavd_list = NULL;
886 ws->n_scavd_blocks = 0;
894 if (gc_threads == NULL) {
895 #if defined(THREADED_RTS)
897 gc_threads = stgMallocBytes (RtsFlags.ParFlags.nNodes *
901 for (i = 0; i < RtsFlags.ParFlags.nNodes; i++) {
903 stgMallocBytes(sizeof(gc_thread) +
904 RtsFlags.GcFlags.generations * sizeof(gen_workspace),
907 new_gc_thread(i, gc_threads[i]);
910 gc_threads = stgMallocBytes (sizeof(gc_thread*),"alloc_gc_threads");
912 new_gc_thread(0,gc_threads[0]);
921 if (gc_threads != NULL) {
922 #if defined(THREADED_RTS)
924 for (i = 0; i < n_capabilities; i++) {
925 for (g = 0; g < RtsFlags.GcFlags.generations; g++)
927 freeWSDeque(gc_threads[i]->gens[g].todo_q);
929 stgFree (gc_threads[i]);
931 stgFree (gc_threads);
933 for (g = 0; g < RtsFlags.GcFlags.generations; g++)
935 freeWSDeque(gc_threads[0]->gens[g].todo_q);
937 stgFree (gc_threads);
943 /* ----------------------------------------------------------------------------
945 ------------------------------------------------------------------------- */
947 static volatile StgWord gc_running_threads;
953 new = atomic_inc(&gc_running_threads);
954 ASSERT(new <= n_gc_threads);
961 ASSERT(gc_running_threads != 0);
962 return atomic_dec(&gc_running_threads);
975 // scavenge objects in compacted generation
976 if (mark_stack_bd != NULL && !mark_stack_empty()) {
980 // Check for global work in any step. We don't need to check for
981 // local work, because we have already exited scavenge_loop(),
982 // which means there is no local work for this thread.
983 for (g = 0; g < (int)RtsFlags.GcFlags.generations; g++) {
985 if (ws->todo_large_objects) return rtsTrue;
986 if (!looksEmptyWSDeque(ws->todo_q)) return rtsTrue;
987 if (ws->todo_overflow) return rtsTrue;
990 #if defined(THREADED_RTS)
993 // look for work to steal
994 for (n = 0; n < n_gc_threads; n++) {
995 if (n == gct->thread_index) continue;
996 for (g = RtsFlags.GcFlags.generations-1; g >= 0; g--) {
997 ws = &gc_threads[n]->gens[g];
998 if (!looksEmptyWSDeque(ws->todo_q)) return rtsTrue;
1005 #if defined(THREADED_RTS)
1013 scavenge_until_all_done (void)
1019 traceEventGcWork(&capabilities[gct->thread_index]);
1021 #if defined(THREADED_RTS)
1022 if (n_gc_threads > 1) {
1031 // scavenge_loop() only exits when there's no work to do
1034 traceEventGcIdle(&capabilities[gct->thread_index]);
1036 debugTrace(DEBUG_gc, "%d GC threads still running", r);
1038 while (gc_running_threads != 0) {
1044 // any_work() does not remove the work from the queue, it
1045 // just checks for the presence of work. If we find any,
1046 // then we increment gc_running_threads and go back to
1047 // scavenge_loop() to perform any pending work.
1050 traceEventGcDone(&capabilities[gct->thread_index]);
1053 #if defined(THREADED_RTS)
1056 gcWorkerThread (Capability *cap)
1058 gc_thread *saved_gct;
1060 // necessary if we stole a callee-saves register for gct:
1063 gct = gc_threads[cap->no];
1064 gct->id = osThreadId();
1066 // Wait until we're told to wake up
1067 RELEASE_SPIN_LOCK(&gct->mut_spin);
1068 gct->wakeup = GC_THREAD_STANDING_BY;
1069 debugTrace(DEBUG_gc, "GC thread %d standing by...", gct->thread_index);
1070 ACQUIRE_SPIN_LOCK(&gct->gc_spin);
1073 // start performance counters in this thread...
1074 if (gct->papi_events == -1) {
1075 papi_init_eventset(&gct->papi_events);
1077 papi_thread_start_gc1_count(gct->papi_events);
1080 // Every thread evacuates some roots.
1082 markSomeCapabilities(mark_root, gct, gct->thread_index, n_gc_threads,
1083 rtsTrue/*prune sparks*/);
1084 scavenge_capability_mut_lists(&capabilities[gct->thread_index]);
1086 scavenge_until_all_done();
1089 // Now that the whole heap is marked, we discard any sparks that
1090 // were found to be unreachable. The main GC thread is currently
1091 // marking heap reachable via weak pointers, so it is
1092 // non-deterministic whether a spark will be retained if it is
1093 // only reachable via weak pointers. To fix this problem would
1094 // require another GC barrier, which is too high a price.
1095 pruneSparkQueue(cap);
1099 // count events in this thread towards the GC totals
1100 papi_thread_stop_gc1_count(gct->papi_events);
1103 // Wait until we're told to continue
1104 RELEASE_SPIN_LOCK(&gct->gc_spin);
1105 gct->wakeup = GC_THREAD_WAITING_TO_CONTINUE;
1106 debugTrace(DEBUG_gc, "GC thread %d waiting to continue...",
1108 ACQUIRE_SPIN_LOCK(&gct->mut_spin);
1109 debugTrace(DEBUG_gc, "GC thread %d on my way...", gct->thread_index);
1116 #if defined(THREADED_RTS)
1119 waitForGcThreads (Capability *cap USED_IF_THREADS)
1121 const nat n_threads = RtsFlags.ParFlags.nNodes;
1122 const nat me = cap->no;
1124 rtsBool retry = rtsTrue;
1127 for (i=0; i < n_threads; i++) {
1128 if (i == me) continue;
1129 if (gc_threads[i]->wakeup != GC_THREAD_STANDING_BY) {
1130 prodCapability(&capabilities[i], cap->running_task);
1133 for (j=0; j < 10; j++) {
1135 for (i=0; i < n_threads; i++) {
1136 if (i == me) continue;
1138 setContextSwitches();
1139 if (gc_threads[i]->wakeup != GC_THREAD_STANDING_BY) {
1149 #endif // THREADED_RTS
1152 start_gc_threads (void)
1154 #if defined(THREADED_RTS)
1155 gc_running_threads = 0;
1160 wakeup_gc_threads (nat n_threads USED_IF_THREADS, nat me USED_IF_THREADS)
1162 #if defined(THREADED_RTS)
1164 for (i=0; i < n_threads; i++) {
1165 if (i == me) continue;
1167 debugTrace(DEBUG_gc, "waking up gc thread %d", i);
1168 if (gc_threads[i]->wakeup != GC_THREAD_STANDING_BY) barf("wakeup_gc_threads");
1170 gc_threads[i]->wakeup = GC_THREAD_RUNNING;
1171 ACQUIRE_SPIN_LOCK(&gc_threads[i]->mut_spin);
1172 RELEASE_SPIN_LOCK(&gc_threads[i]->gc_spin);
1177 // After GC is complete, we must wait for all GC threads to enter the
1178 // standby state, otherwise they may still be executing inside
1179 // any_work(), and may even remain awake until the next GC starts.
1181 shutdown_gc_threads (nat n_threads USED_IF_THREADS, nat me USED_IF_THREADS)
1183 #if defined(THREADED_RTS)
1185 for (i=0; i < n_threads; i++) {
1186 if (i == me) continue;
1187 while (gc_threads[i]->wakeup != GC_THREAD_WAITING_TO_CONTINUE) { write_barrier(); }
1192 #if defined(THREADED_RTS)
1194 releaseGCThreads (Capability *cap USED_IF_THREADS)
1196 const nat n_threads = RtsFlags.ParFlags.nNodes;
1197 const nat me = cap->no;
1199 for (i=0; i < n_threads; i++) {
1200 if (i == me) continue;
1201 if (gc_threads[i]->wakeup != GC_THREAD_WAITING_TO_CONTINUE)
1202 barf("releaseGCThreads");
1204 gc_threads[i]->wakeup = GC_THREAD_INACTIVE;
1205 ACQUIRE_SPIN_LOCK(&gc_threads[i]->gc_spin);
1206 RELEASE_SPIN_LOCK(&gc_threads[i]->mut_spin);
1211 /* ----------------------------------------------------------------------------
1212 Initialise a generation that is to be collected
1213 ------------------------------------------------------------------------- */
1216 init_collected_gen (nat g, nat n_threads)
1223 // Throw away the current mutable list. Invariant: the mutable
1224 // list always has at least one block; this means we can avoid a
1225 // check for NULL in recordMutable().
1227 freeChain(generations[g].mut_list);
1228 generations[g].mut_list = allocBlock();
1229 for (i = 0; i < n_capabilities; i++) {
1230 freeChain(capabilities[i].mut_lists[g]);
1231 capabilities[i].mut_lists[g] = allocBlock();
1235 gen = &generations[g];
1236 ASSERT(gen->no == g);
1238 // we'll construct a new list of threads in this step
1239 // during GC, throw away the current list.
1240 gen->old_threads = gen->threads;
1241 gen->threads = END_TSO_QUEUE;
1243 // deprecate the existing blocks
1244 gen->old_blocks = gen->blocks;
1245 gen->n_old_blocks = gen->n_blocks;
1249 gen->live_estimate = 0;
1251 // initialise the large object queues.
1252 gen->scavenged_large_objects = NULL;
1253 gen->n_scavenged_large_blocks = 0;
1255 // mark the small objects as from-space
1256 for (bd = gen->old_blocks; bd; bd = bd->link) {
1257 bd->flags &= ~BF_EVACUATED;
1260 // mark the large objects as from-space
1261 for (bd = gen->large_objects; bd; bd = bd->link) {
1262 bd->flags &= ~BF_EVACUATED;
1265 // for a compacted generation, we need to allocate the bitmap
1267 nat bitmap_size; // in bytes
1268 bdescr *bitmap_bdescr;
1271 bitmap_size = gen->n_old_blocks * BLOCK_SIZE / (sizeof(W_)*BITS_PER_BYTE);
1273 if (bitmap_size > 0) {
1274 bitmap_bdescr = allocGroup((lnat)BLOCK_ROUND_UP(bitmap_size)
1276 gen->bitmap = bitmap_bdescr;
1277 bitmap = bitmap_bdescr->start;
1279 debugTrace(DEBUG_gc, "bitmap_size: %d, bitmap: %p",
1280 bitmap_size, bitmap);
1282 // don't forget to fill it with zeros!
1283 memset(bitmap, 0, bitmap_size);
1285 // For each block in this step, point to its bitmap from the
1286 // block descriptor.
1287 for (bd=gen->old_blocks; bd != NULL; bd = bd->link) {
1288 bd->u.bitmap = bitmap;
1289 bitmap += BLOCK_SIZE_W / (sizeof(W_)*BITS_PER_BYTE);
1291 // Also at this point we set the BF_MARKED flag
1292 // for this block. The invariant is that
1293 // BF_MARKED is always unset, except during GC
1294 // when it is set on those blocks which will be
1296 if (!(bd->flags & BF_FRAGMENTED)) {
1297 bd->flags |= BF_MARKED;
1303 // For each GC thread, for each step, allocate a "todo" block to
1304 // store evacuated objects to be scavenged, and a block to store
1305 // evacuated objects that do not need to be scavenged.
1306 for (t = 0; t < n_threads; t++) {
1307 ws = &gc_threads[t]->gens[g];
1309 ws->todo_large_objects = NULL;
1311 ws->part_list = NULL;
1312 ws->n_part_blocks = 0;
1314 // allocate the first to-space block; extra blocks will be
1315 // chained on as necessary.
1317 ASSERT(looksEmptyWSDeque(ws->todo_q));
1318 alloc_todo_block(ws,0);
1320 ws->todo_overflow = NULL;
1321 ws->n_todo_overflow = 0;
1323 ws->scavd_list = NULL;
1324 ws->n_scavd_blocks = 0;
1329 /* ----------------------------------------------------------------------------
1330 Initialise a generation that is *not* to be collected
1331 ------------------------------------------------------------------------- */
1334 init_uncollected_gen (nat g, nat threads)
1341 // save the current mutable lists for this generation, and
1342 // allocate a fresh block for each one. We'll traverse these
1343 // mutable lists as roots early on in the GC.
1344 generations[g].saved_mut_list = generations[g].mut_list;
1345 generations[g].mut_list = allocBlock();
1346 for (n = 0; n < n_capabilities; n++) {
1347 capabilities[n].saved_mut_lists[g] = capabilities[n].mut_lists[g];
1348 capabilities[n].mut_lists[g] = allocBlock();
1351 gen = &generations[g];
1353 gen->scavenged_large_objects = NULL;
1354 gen->n_scavenged_large_blocks = 0;
1356 for (t = 0; t < threads; t++) {
1357 ws = &gc_threads[t]->gens[g];
1359 ASSERT(looksEmptyWSDeque(ws->todo_q));
1360 ws->todo_large_objects = NULL;
1362 ws->part_list = NULL;
1363 ws->n_part_blocks = 0;
1365 ws->scavd_list = NULL;
1366 ws->n_scavd_blocks = 0;
1368 // If the block at the head of the list in this generation
1369 // is less than 3/4 full, then use it as a todo block.
1370 if (gen->blocks && isPartiallyFull(gen->blocks))
1372 ws->todo_bd = gen->blocks;
1373 ws->todo_free = ws->todo_bd->free;
1374 ws->todo_lim = ws->todo_bd->start + BLOCK_SIZE_W;
1375 gen->blocks = gen->blocks->link;
1377 gen->n_words -= ws->todo_bd->free - ws->todo_bd->start;
1378 ws->todo_bd->link = NULL;
1379 // we must scan from the current end point.
1380 ws->todo_bd->u.scan = ws->todo_bd->free;
1385 alloc_todo_block(ws,0);
1389 // deal out any more partial blocks to the threads' part_lists
1391 while (gen->blocks && isPartiallyFull(gen->blocks))
1394 gen->blocks = bd->link;
1395 ws = &gc_threads[t]->gens[g];
1396 bd->link = ws->part_list;
1398 ws->n_part_blocks += 1;
1399 bd->u.scan = bd->free;
1401 gen->n_words -= bd->free - bd->start;
1403 if (t == n_gc_threads) t = 0;
1407 /* -----------------------------------------------------------------------------
1408 Initialise a gc_thread before GC
1409 -------------------------------------------------------------------------- */
1412 init_gc_thread (gc_thread *t)
1414 t->static_objects = END_OF_STATIC_LIST;
1415 t->scavenged_static_objects = END_OF_STATIC_LIST;
1417 t->mut_lists = capabilities[t->thread_index].mut_lists;
1419 t->failed_to_evac = rtsFalse;
1420 t->eager_promotion = rtsTrue;
1421 t->thunk_selector_depth = 0;
1426 t->scav_find_work = 0;
1429 /* -----------------------------------------------------------------------------
1430 Function we pass to evacuate roots.
1431 -------------------------------------------------------------------------- */
1434 mark_root(void *user USED_IF_THREADS, StgClosure **root)
1436 // we stole a register for gct, but this function is called from
1437 // *outside* the GC where the register variable is not in effect,
1438 // so we need to save and restore it here. NB. only call
1439 // mark_root() from the main GC thread, otherwise gct will be
1441 gc_thread *saved_gct;
1450 /* -----------------------------------------------------------------------------
1451 Initialising the static object & mutable lists
1452 -------------------------------------------------------------------------- */
1455 zero_static_object_list(StgClosure* first_static)
1459 const StgInfoTable *info;
1461 for (p = first_static; p != END_OF_STATIC_LIST; p = link) {
1463 link = *STATIC_LINK(info, p);
1464 *STATIC_LINK(info,p) = NULL;
1468 /* ----------------------------------------------------------------------------
1469 Reset the sizes of the older generations when we do a major
1472 CURRENT STRATEGY: make all generations except zero the same size.
1473 We have to stay within the maximum heap size, and leave a certain
1474 percentage of the maximum heap size available to allocate into.
1475 ------------------------------------------------------------------------- */
1478 resize_generations (void)
1482 if (major_gc && RtsFlags.GcFlags.generations > 1) {
1483 nat live, size, min_alloc, words;
1484 const nat max = RtsFlags.GcFlags.maxHeapSize;
1485 const nat gens = RtsFlags.GcFlags.generations;
1487 // live in the oldest generations
1488 if (oldest_gen->live_estimate != 0) {
1489 words = oldest_gen->live_estimate;
1491 words = oldest_gen->n_words;
1493 live = (words + BLOCK_SIZE_W - 1) / BLOCK_SIZE_W +
1494 oldest_gen->n_large_blocks;
1496 // default max size for all generations except zero
1497 size = stg_max(live * RtsFlags.GcFlags.oldGenFactor,
1498 RtsFlags.GcFlags.minOldGenSize);
1500 if (RtsFlags.GcFlags.heapSizeSuggestionAuto) {
1501 RtsFlags.GcFlags.heapSizeSuggestion = size;
1504 // minimum size for generation zero
1505 min_alloc = stg_max((RtsFlags.GcFlags.pcFreeHeap * max) / 200,
1506 RtsFlags.GcFlags.minAllocAreaSize);
1508 // Auto-enable compaction when the residency reaches a
1509 // certain percentage of the maximum heap size (default: 30%).
1510 if (RtsFlags.GcFlags.compact ||
1512 oldest_gen->n_blocks >
1513 (RtsFlags.GcFlags.compactThreshold * max) / 100)) {
1514 oldest_gen->mark = 1;
1515 oldest_gen->compact = 1;
1516 // debugBelch("compaction: on\n", live);
1518 oldest_gen->mark = 0;
1519 oldest_gen->compact = 0;
1520 // debugBelch("compaction: off\n", live);
1523 if (RtsFlags.GcFlags.sweep) {
1524 oldest_gen->mark = 1;
1527 // if we're going to go over the maximum heap size, reduce the
1528 // size of the generations accordingly. The calculation is
1529 // different if compaction is turned on, because we don't need
1530 // to double the space required to collect the old generation.
1533 // this test is necessary to ensure that the calculations
1534 // below don't have any negative results - we're working
1535 // with unsigned values here.
1536 if (max < min_alloc) {
1540 if (oldest_gen->compact) {
1541 if ( (size + (size - 1) * (gens - 2) * 2) + min_alloc > max ) {
1542 size = (max - min_alloc) / ((gens - 1) * 2 - 1);
1545 if ( (size * (gens - 1) * 2) + min_alloc > max ) {
1546 size = (max - min_alloc) / ((gens - 1) * 2);
1556 debugBelch("live: %d, min_alloc: %d, size : %d, max = %d\n", live,
1557 min_alloc, size, max);
1560 for (g = 0; g < gens; g++) {
1561 generations[g].max_blocks = size;
1566 /* -----------------------------------------------------------------------------
1567 Calculate the new size of the nursery, and resize it.
1568 -------------------------------------------------------------------------- */
1571 resize_nursery (void)
1573 const lnat min_nursery = RtsFlags.GcFlags.minAllocAreaSize * n_capabilities;
1575 if (RtsFlags.GcFlags.generations == 1)
1576 { // Two-space collector:
1579 /* set up a new nursery. Allocate a nursery size based on a
1580 * function of the amount of live data (by default a factor of 2)
1581 * Use the blocks from the old nursery if possible, freeing up any
1584 * If we get near the maximum heap size, then adjust our nursery
1585 * size accordingly. If the nursery is the same size as the live
1586 * data (L), then we need 3L bytes. We can reduce the size of the
1587 * nursery to bring the required memory down near 2L bytes.
1589 * A normal 2-space collector would need 4L bytes to give the same
1590 * performance we get from 3L bytes, reducing to the same
1591 * performance at 2L bytes.
1593 blocks = generations[0].n_blocks;
1595 if ( RtsFlags.GcFlags.maxHeapSize != 0 &&
1596 blocks * RtsFlags.GcFlags.oldGenFactor * 2 >
1597 RtsFlags.GcFlags.maxHeapSize )
1599 long adjusted_blocks; // signed on purpose
1602 adjusted_blocks = (RtsFlags.GcFlags.maxHeapSize - 2 * blocks);
1604 debugTrace(DEBUG_gc, "near maximum heap size of 0x%x blocks, blocks = %d, adjusted to %ld",
1605 RtsFlags.GcFlags.maxHeapSize, blocks, adjusted_blocks);
1607 pc_free = adjusted_blocks * 100 / RtsFlags.GcFlags.maxHeapSize;
1608 if (pc_free < RtsFlags.GcFlags.pcFreeHeap) /* might even * be < 0 */
1612 blocks = adjusted_blocks;
1616 blocks *= RtsFlags.GcFlags.oldGenFactor;
1617 if (blocks < min_nursery)
1619 blocks = min_nursery;
1622 resizeNurseries(blocks);
1624 else // Generational collector
1627 * If the user has given us a suggested heap size, adjust our
1628 * allocation area to make best use of the memory available.
1630 if (RtsFlags.GcFlags.heapSizeSuggestion)
1633 const nat needed = calcNeeded(); // approx blocks needed at next GC
1635 /* Guess how much will be live in generation 0 step 0 next time.
1636 * A good approximation is obtained by finding the
1637 * percentage of g0 that was live at the last minor GC.
1639 * We have an accurate figure for the amount of copied data in
1640 * 'copied', but we must convert this to a number of blocks, with
1641 * a small adjustment for estimated slop at the end of a block
1646 g0_pcnt_kept = ((copied / (BLOCK_SIZE_W - 10)) * 100)
1647 / countNurseryBlocks();
1650 /* Estimate a size for the allocation area based on the
1651 * information available. We might end up going slightly under
1652 * or over the suggested heap size, but we should be pretty
1655 * Formula: suggested - needed
1656 * ----------------------------
1657 * 1 + g0_pcnt_kept/100
1659 * where 'needed' is the amount of memory needed at the next
1660 * collection for collecting all gens except g0.
1663 (((long)RtsFlags.GcFlags.heapSizeSuggestion - (long)needed) * 100) /
1664 (100 + (long)g0_pcnt_kept);
1666 if (blocks < (long)min_nursery) {
1667 blocks = min_nursery;
1670 resizeNurseries((nat)blocks);
1674 // we might have added extra large blocks to the nursery, so
1675 // resize back to minAllocAreaSize again.
1676 resizeNurseriesFixed(RtsFlags.GcFlags.minAllocAreaSize);
1681 /* -----------------------------------------------------------------------------
1682 Sanity code for CAF garbage collection.
1684 With DEBUG turned on, we manage a CAF list in addition to the SRT
1685 mechanism. After GC, we run down the CAF list and blackhole any
1686 CAFs which have been garbage collected. This means we get an error
1687 whenever the program tries to enter a garbage collected CAF.
1689 Any garbage collected CAFs are taken off the CAF list at the same
1691 -------------------------------------------------------------------------- */
1693 #if 0 && defined(DEBUG)
1700 const StgInfoTable *info;
1711 ASSERT(info->type == IND_STATIC);
1713 if (STATIC_LINK(info,p) == NULL) {
1714 debugTrace(DEBUG_gccafs, "CAF gc'd at 0x%04lx", (long)p);
1716 SET_INFO(p,&stg_BLACKHOLE_info);
1717 p = STATIC_LINK2(info,p);
1721 pp = &STATIC_LINK2(info,p);
1728 debugTrace(DEBUG_gccafs, "%d CAFs live", i);