1 /* -----------------------------------------------------------------------------
3 * (c) The GHC Team 1998-2008
5 * Generational garbage collector
7 * Documentation on the architecture of the Garbage Collector can be
8 * found in the online commentary:
10 * http://hackage.haskell.org/trac/ghc/wiki/Commentary/Rts/Storage/GC
12 * ---------------------------------------------------------------------------*/
14 // #include "PosixSource.h"
19 #include "OSThreads.h"
20 #include "LdvProfile.h"
25 #include "BlockAlloc.h"
31 #include "ParTicky.h" // ToDo: move into Rts.h
32 #include "RtsSignals.h"
36 #if defined(RTS_GTK_FRONTPANEL)
37 #include "FrontPanel.h"
40 #include "RetainerProfile.h"
41 #include "RaiseAsync.h"
54 #include <string.h> // for memset()
57 /* -----------------------------------------------------------------------------
59 -------------------------------------------------------------------------- */
61 /* STATIC OBJECT LIST.
64 * We maintain a linked list of static objects that are still live.
65 * The requirements for this list are:
67 * - we need to scan the list while adding to it, in order to
68 * scavenge all the static objects (in the same way that
69 * breadth-first scavenging works for dynamic objects).
71 * - we need to be able to tell whether an object is already on
72 * the list, to break loops.
74 * Each static object has a "static link field", which we use for
75 * linking objects on to the list. We use a stack-type list, consing
76 * objects on the front as they are added (this means that the
77 * scavenge phase is depth-first, not breadth-first, but that
80 * A separate list is kept for objects that have been scavenged
81 * already - this is so that we can zero all the marks afterwards.
83 * An object is on the list if its static link field is non-zero; this
84 * means that we have to mark the end of the list with '1', not NULL.
86 * Extra notes for generational GC:
88 * Each generation has a static object list associated with it. When
89 * collecting generations up to N, we treat the static object lists
90 * from generations > N as roots.
92 * We build up a static object list while collecting generations 0..N,
93 * which is then appended to the static object list of generation N+1.
96 /* N is the oldest generation being collected, where the generations
97 * are numbered starting at 0. A major GC (indicated by the major_gc
98 * flag) is when we're collecting all generations. We only attempt to
99 * deal with static objects and GC CAFs when doing a major GC.
104 /* Data used for allocation area sizing.
106 static lnat g0s0_pcnt_kept = 30; // percentage of g0s0 live at last minor GC
116 /* Thread-local data for each GC thread
118 gc_thread **gc_threads = NULL;
119 // gc_thread *gct = NULL; // this thread's gct TODO: make thread-local
121 // Number of threads running in *this* GC. Affects how many
122 // step->todos[] lists we have to look in to find work.
126 long copied; // *words* copied & scavenged during this GC
130 /* -----------------------------------------------------------------------------
131 Static function declarations
132 -------------------------------------------------------------------------- */
134 static void mark_root (void *user, StgClosure **root);
135 static void zero_static_object_list (StgClosure* first_static);
136 static nat initialise_N (rtsBool force_major_gc);
137 static void init_collected_gen (nat g, nat threads);
138 static void init_uncollected_gen (nat g, nat threads);
139 static void init_gc_thread (gc_thread *t);
140 static void update_task_list (void);
141 static void resize_generations (void);
142 static void resize_nursery (void);
143 static void start_gc_threads (void);
144 static void scavenge_until_all_done (void);
145 static nat inc_running (void);
146 static nat dec_running (void);
147 static void wakeup_gc_threads (nat n_threads, nat me);
148 static void shutdown_gc_threads (nat n_threads, nat me);
149 static void continue_gc_threads (nat n_threads, nat me);
151 #if 0 && defined(DEBUG)
152 static void gcCAFs (void);
155 /* -----------------------------------------------------------------------------
156 The mark bitmap & stack.
157 -------------------------------------------------------------------------- */
159 #define MARK_STACK_BLOCKS 4
161 bdescr *mark_stack_bdescr;
166 // Flag and pointers used for falling back to a linear scan when the
167 // mark stack overflows.
168 rtsBool mark_stack_overflowed;
169 bdescr *oldgen_scan_bd;
172 /* -----------------------------------------------------------------------------
173 GarbageCollect: the main entry point to the garbage collector.
175 Locks held: all capabilities are held throughout GarbageCollect().
176 -------------------------------------------------------------------------- */
179 GarbageCollect (rtsBool force_major_gc,
180 nat gc_type USED_IF_THREADS,
181 Capability *cap USED_IF_THREADS)
185 lnat live, allocated, max_copied, avg_copied, slop;
186 gc_thread *saved_gct;
189 // necessary if we stole a callee-saves register for gct:
193 CostCentreStack *prev_CCS;
198 #if defined(RTS_USER_SIGNALS)
199 if (RtsFlags.MiscFlags.install_signal_handlers) {
205 ASSERT(sizeof(step_workspace) == 16 * sizeof(StgWord));
206 // otherwise adjust the padding in step_workspace.
208 // tell the stats department that we've started a GC
211 // tell the STM to discard any cached closures it's hoping to re-use
220 // attribute any costs to CCS_GC
226 /* Approximate how much we allocated.
227 * Todo: only when generating stats?
229 allocated = calcAllocated();
231 /* Figure out which generation to collect
233 n = initialise_N(force_major_gc);
235 /* Start threads, so they can be spinning up while we finish initialisation.
239 #if defined(THREADED_RTS)
240 /* How many threads will be participating in this GC?
241 * We don't try to parallelise minor GCs (unless the user asks for
242 * it with +RTS -gn0), or mark/compact/sweep GC.
244 if (gc_type == PENDING_GC_PAR) {
245 n_gc_threads = RtsFlags.ParFlags.nNodes;
253 trace(TRACE_gc|DEBUG_gc, "GC (gen %d): %d KB to collect, %ld MB in use, using %d thread(s)",
254 N, n * (BLOCK_SIZE / 1024), mblocks_allocated, n_gc_threads);
256 #ifdef RTS_GTK_FRONTPANEL
257 if (RtsFlags.GcFlags.frontpanel) {
258 updateFrontPanelBeforeGC(N);
263 // check for memory leaks if DEBUG is on
264 memInventory(traceClass(DEBUG_gc));
267 // check stack sanity *before* GC
268 IF_DEBUG(sanity, checkFreeListSanity());
269 IF_DEBUG(sanity, checkMutableLists(rtsTrue));
271 // Initialise all our gc_thread structures
272 for (t = 0; t < n_gc_threads; t++) {
273 init_gc_thread(gc_threads[t]);
276 // Initialise all the generations/steps that we're collecting.
277 for (g = 0; g <= N; g++) {
278 init_collected_gen(g,n_gc_threads);
281 // Initialise all the generations/steps that we're *not* collecting.
282 for (g = N+1; g < RtsFlags.GcFlags.generations; g++) {
283 init_uncollected_gen(g,n_gc_threads);
286 /* Allocate a mark stack if we're doing a major collection.
288 if (major_gc && oldest_gen->steps[0].mark) {
289 nat mark_stack_blocks;
290 mark_stack_blocks = stg_max(MARK_STACK_BLOCKS,
291 oldest_gen->steps[0].n_old_blocks / 100);
292 mark_stack_bdescr = allocGroup(mark_stack_blocks);
293 mark_stack = (StgPtr *)mark_stack_bdescr->start;
294 mark_sp = mark_stack;
295 mark_splim = mark_stack + (mark_stack_blocks * BLOCK_SIZE_W);
297 mark_stack_bdescr = NULL;
300 // this is the main thread
302 if (n_gc_threads == 1) {
305 gct = gc_threads[cap->no];
311 /* -----------------------------------------------------------------------
312 * follow all the roots that we know about:
315 // the main thread is running: this prevents any other threads from
316 // exiting prematurely, so we can start them now.
317 // NB. do this after the mutable lists have been saved above, otherwise
318 // the other GC threads will be writing into the old mutable lists.
320 wakeup_gc_threads(n_gc_threads, gct->thread_index);
322 // Mutable lists from each generation > N
323 // we want to *scavenge* these roots, not evacuate them: they're not
324 // going to move in this GC.
325 // Also do them in reverse generation order, for the usual reason:
326 // namely to reduce the likelihood of spurious old->new pointers.
328 for (g = RtsFlags.GcFlags.generations-1; g > N; g--) {
329 scavenge_mutable_list(generations[g].saved_mut_list, &generations[g]);
330 freeChain_sync(generations[g].saved_mut_list);
331 generations[g].saved_mut_list = NULL;
335 // scavenge the capability-private mutable lists. This isn't part
336 // of markSomeCapabilities() because markSomeCapabilities() can only
337 // call back into the GC via mark_root() (due to the gct register
339 if (n_gc_threads == 1) {
340 for (n = 0; n < n_capabilities; n++) {
341 scavenge_capability_mut_lists(&capabilities[n]);
344 scavenge_capability_mut_lists(&capabilities[gct->thread_index]);
347 // follow roots from the CAF list (used by GHCi)
349 markCAFs(mark_root, gct);
351 // follow all the roots that the application knows about.
353 markSomeCapabilities(mark_root, gct, gct->thread_index, n_gc_threads,
354 rtsTrue/*prune sparks*/);
356 #if defined(RTS_USER_SIGNALS)
357 // mark the signal handlers (signals should be already blocked)
358 markSignalHandlers(mark_root, gct);
361 // Mark the weak pointer list, and prepare to detect dead weak pointers.
365 // Mark the stable pointer table.
366 markStablePtrTable(mark_root, gct);
368 /* -------------------------------------------------------------------------
369 * Repeatedly scavenge all the areas we know about until there's no
370 * more scavenging to be done.
374 scavenge_until_all_done();
375 // The other threads are now stopped. We might recurse back to
376 // here, but from now on this is the only thread.
378 // if any blackholes are alive, make the threads that wait on
380 if (traverseBlackholeQueue()) {
385 // must be last... invariant is that everything is fully
386 // scavenged at this point.
387 if (traverseWeakPtrList()) { // returns rtsTrue if evaced something
392 // If we get to here, there's really nothing left to do.
396 shutdown_gc_threads(n_gc_threads, gct->thread_index);
398 // Update pointers from the Task list
401 // Now see which stable names are still alive.
405 // We call processHeapClosureForDead() on every closure destroyed during
406 // the current garbage collection, so we invoke LdvCensusForDead().
407 if (RtsFlags.ProfFlags.doHeapProfile == HEAP_BY_LDV
408 || RtsFlags.ProfFlags.bioSelector != NULL)
412 // NO MORE EVACUATION AFTER THIS POINT!
414 // Two-space collector: free the old to-space.
415 // g0s0->old_blocks is the old nursery
416 // g0s0->blocks is to-space from the previous GC
417 if (RtsFlags.GcFlags.generations == 1) {
418 if (g0s0->blocks != NULL) {
419 freeChain(g0s0->blocks);
424 // For each workspace, in each thread, move the copied blocks to the step
430 for (t = 0; t < n_gc_threads; t++) {
434 if (RtsFlags.GcFlags.generations == 1) {
439 for (; s < total_steps; s++) {
442 // Push the final block
444 push_scanned_block(ws->todo_bd, ws);
447 ASSERT(gct->scan_bd == NULL);
448 ASSERT(countBlocks(ws->scavd_list) == ws->n_scavd_blocks);
451 for (bd = ws->scavd_list; bd != NULL; bd = bd->link) {
452 ws->step->n_words += bd->free - bd->start;
456 prev->link = ws->step->blocks;
457 ws->step->blocks = ws->scavd_list;
459 ws->step->n_blocks += ws->n_scavd_blocks;
463 // Add all the partial blocks *after* we've added all the full
464 // blocks. This is so that we can grab the partial blocks back
465 // again and try to fill them up in the next GC.
466 for (t = 0; t < n_gc_threads; t++) {
470 if (RtsFlags.GcFlags.generations == 1) {
475 for (; s < total_steps; s++) {
479 for (bd = ws->part_list; bd != NULL; bd = next) {
481 if (bd->free == bd->start) {
483 ws->part_list = next;
490 ws->step->n_words += bd->free - bd->start;
495 prev->link = ws->step->blocks;
496 ws->step->blocks = ws->part_list;
498 ws->step->n_blocks += ws->n_part_blocks;
500 ASSERT(countBlocks(ws->step->blocks) == ws->step->n_blocks);
501 ASSERT(countOccupied(ws->step->blocks) == ws->step->n_words);
506 // Finally: compact or sweep the oldest generation.
507 if (major_gc && oldest_gen->steps[0].mark) {
508 if (oldest_gen->steps[0].compact)
509 compact(gct->scavenged_static_objects);
511 sweep(&oldest_gen->steps[0]);
514 /* run through all the generations/steps and tidy up
521 for (i=0; i < n_gc_threads; i++) {
522 if (n_gc_threads > 1) {
523 trace(TRACE_gc,"thread %d:", i);
524 trace(TRACE_gc," copied %ld", gc_threads[i]->copied * sizeof(W_));
525 trace(TRACE_gc," scanned %ld", gc_threads[i]->scanned * sizeof(W_));
526 trace(TRACE_gc," any_work %ld", gc_threads[i]->any_work);
527 trace(TRACE_gc," no_work %ld", gc_threads[i]->no_work);
528 trace(TRACE_gc," scav_find_work %ld", gc_threads[i]->scav_find_work);
530 copied += gc_threads[i]->copied;
531 max_copied = stg_max(gc_threads[i]->copied, max_copied);
533 if (n_gc_threads == 1) {
541 for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
544 generations[g].collections++; // for stats
545 if (n_gc_threads > 1) generations[g].par_collections++;
548 // Count the mutable list as bytes "copied" for the purposes of
549 // stats. Every mutable list is copied during every GC.
551 nat mut_list_size = 0;
552 for (bd = generations[g].mut_list; bd != NULL; bd = bd->link) {
553 mut_list_size += bd->free - bd->start;
555 for (n = 0; n < n_capabilities; n++) {
556 for (bd = capabilities[n].mut_lists[g];
557 bd != NULL; bd = bd->link) {
558 mut_list_size += bd->free - bd->start;
561 copied += mut_list_size;
564 "mut_list_size: %lu (%d vars, %d arrays, %d MVARs, %d others)",
565 (unsigned long)(mut_list_size * sizeof(W_)),
566 mutlist_MUTVARS, mutlist_MUTARRS, mutlist_MVARS, mutlist_OTHERS);
569 for (s = 0; s < generations[g].n_steps; s++) {
571 stp = &generations[g].steps[s];
573 // for generations we collected...
576 /* free old memory and shift to-space into from-space for all
577 * the collected steps (except the allocation area). These
578 * freed blocks will probaby be quickly recycled.
580 if (!(g == 0 && s == 0 && RtsFlags.GcFlags.generations > 1)) {
583 // tack the new blocks on the end of the existing blocks
584 if (stp->old_blocks != NULL) {
587 for (bd = stp->old_blocks; bd != NULL; bd = next) {
591 if (!(bd->flags & BF_MARKED))
594 stp->old_blocks = next;
603 stp->n_words += bd->free - bd->start;
605 // NB. this step might not be compacted next
606 // time, so reset the BF_MARKED flags.
607 // They are set before GC if we're going to
608 // compact. (search for BF_MARKED above).
609 bd->flags &= ~BF_MARKED;
611 // between GCs, all blocks in the heap except
612 // for the nursery have the BF_EVACUATED flag set.
613 bd->flags |= BF_EVACUATED;
620 prev->link = stp->blocks;
621 stp->blocks = stp->old_blocks;
624 // add the new blocks to the block tally
625 stp->n_blocks += stp->n_old_blocks;
626 ASSERT(countBlocks(stp->blocks) == stp->n_blocks);
627 ASSERT(countOccupied(stp->blocks) == stp->n_words);
631 freeChain(stp->old_blocks);
633 stp->old_blocks = NULL;
634 stp->n_old_blocks = 0;
637 /* LARGE OBJECTS. The current live large objects are chained on
638 * scavenged_large, having been moved during garbage
639 * collection from large_objects. Any objects left on
640 * large_objects list are therefore dead, so we free them here.
642 for (bd = stp->large_objects; bd != NULL; bd = next) {
648 stp->large_objects = stp->scavenged_large_objects;
649 stp->n_large_blocks = stp->n_scavenged_large_blocks;
652 else // for older generations...
654 /* For older generations, we need to append the
655 * scavenged_large_object list (i.e. large objects that have been
656 * promoted during this GC) to the large_object list for that step.
658 for (bd = stp->scavenged_large_objects; bd; bd = next) {
660 dbl_link_onto(bd, &stp->large_objects);
663 // add the new blocks we promoted during this GC
664 stp->n_large_blocks += stp->n_scavenged_large_blocks;
669 // update the max size of older generations after a major GC
670 resize_generations();
672 // Calculate the amount of live data for stats.
673 live = calcLiveWords();
675 // Free the small objects allocated via allocate(), since this will
676 // all have been copied into G0S1 now.
677 if (RtsFlags.GcFlags.generations > 1) {
678 if (g0s0->blocks != NULL) {
679 freeChain(g0s0->blocks);
686 alloc_blocks_lim = RtsFlags.GcFlags.minAllocAreaSize;
688 // Start a new pinned_object_block
689 pinned_object_block = NULL;
691 // Free the mark stack.
692 if (mark_stack_bdescr != NULL) {
693 freeGroup(mark_stack_bdescr);
697 for (g = 0; g <= N; g++) {
698 for (s = 0; s < generations[g].n_steps; s++) {
699 stp = &generations[g].steps[s];
700 if (stp->bitmap != NULL) {
701 freeGroup(stp->bitmap);
709 // mark the garbage collected CAFs as dead
710 #if 0 && defined(DEBUG) // doesn't work at the moment
711 if (major_gc) { gcCAFs(); }
715 // resetStaticObjectForRetainerProfiling() must be called before
717 if (n_gc_threads > 1) {
718 barf("profiling is currently broken with multi-threaded GC");
719 // ToDo: fix the gct->scavenged_static_objects below
721 resetStaticObjectForRetainerProfiling(gct->scavenged_static_objects);
724 // zero the scavenged static object list
727 for (i = 0; i < n_gc_threads; i++) {
728 zero_static_object_list(gc_threads[i]->scavenged_static_objects);
735 // start any pending finalizers
737 scheduleFinalizers(last_free_capability, old_weak_ptr_list);
740 // send exceptions to any threads which were about to die
742 resurrectThreads(resurrected_threads);
743 performPendingThrowTos(exception_threads);
746 // Update the stable pointer hash table.
747 updateStablePtrTable(major_gc);
749 // check sanity after GC
750 IF_DEBUG(sanity, checkSanity());
752 // extra GC trace info
753 if (traceClass(TRACE_gc|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(traceClass(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();
790 continue_gc_threads(n_gc_threads, gct->thread_index);
797 /* -----------------------------------------------------------------------------
798 Figure out which generation to collect, initialise N and major_gc.
800 Also returns the total number of blocks in generations that will be
802 -------------------------------------------------------------------------- */
805 initialise_N (rtsBool force_major_gc)
808 nat s, blocks, blocks_total;
813 if (force_major_gc) {
814 N = RtsFlags.GcFlags.generations - 1;
819 for (g = RtsFlags.GcFlags.generations - 1; g >= 0; g--) {
821 for (s = 0; s < generations[g].n_steps; s++) {
822 blocks += generations[g].steps[s].n_words / BLOCK_SIZE_W;
823 blocks += generations[g].steps[s].n_large_blocks;
825 if (blocks >= generations[g].max_blocks) {
829 blocks_total += blocks;
833 blocks_total += countNurseryBlocks();
835 major_gc = (N == RtsFlags.GcFlags.generations-1);
839 /* -----------------------------------------------------------------------------
840 Initialise the gc_thread structures.
841 -------------------------------------------------------------------------- */
843 #define GC_THREAD_INACTIVE 0
844 #define GC_THREAD_STANDING_BY 1
845 #define GC_THREAD_RUNNING 2
846 #define GC_THREAD_WAITING_TO_CONTINUE 3
849 alloc_gc_thread (int n)
855 t = stgMallocBytes(sizeof(gc_thread) + total_steps * sizeof(step_workspace),
860 initSpinLock(&t->gc_spin);
861 initSpinLock(&t->mut_spin);
862 ACQUIRE_SPIN_LOCK(&t->gc_spin);
863 t->wakeup = GC_THREAD_INACTIVE; // starts true, so we can wait for the
864 // thread to start up, see wakeup_gc_threads
868 t->free_blocks = NULL;
877 for (s = 0; s < total_steps; s++)
880 ws->step = &all_steps[s];
881 ASSERT(s == ws->step->abs_no);
885 ws->buffer_todo_bd = NULL;
887 ws->part_list = NULL;
888 ws->n_part_blocks = 0;
890 ws->scavd_list = NULL;
891 ws->n_scavd_blocks = 0;
901 if (gc_threads == NULL) {
902 #if defined(THREADED_RTS)
904 gc_threads = stgMallocBytes (RtsFlags.ParFlags.nNodes *
908 for (i = 0; i < RtsFlags.ParFlags.nNodes; i++) {
909 gc_threads[i] = alloc_gc_thread(i);
912 gc_threads = stgMallocBytes (sizeof(gc_thread*),
915 gc_threads[0] = alloc_gc_thread(0);
920 /* ----------------------------------------------------------------------------
922 ------------------------------------------------------------------------- */
924 static nat gc_running_threads;
926 #if defined(THREADED_RTS)
927 static Mutex gc_running_mutex;
934 ACQUIRE_LOCK(&gc_running_mutex);
935 n_running = ++gc_running_threads;
936 RELEASE_LOCK(&gc_running_mutex);
937 ASSERT(n_running <= n_gc_threads);
945 ACQUIRE_LOCK(&gc_running_mutex);
946 ASSERT(n_gc_threads != 0);
947 n_running = --gc_running_threads;
948 RELEASE_LOCK(&gc_running_mutex);
962 // scavenge objects in compacted generation
963 if (mark_stack_overflowed || oldgen_scan_bd != NULL ||
964 (mark_stack_bdescr != NULL && !mark_stack_empty())) {
968 // Check for global work in any step. We don't need to check for
969 // local work, because we have already exited scavenge_loop(),
970 // which means there is no local work for this thread.
971 for (s = total_steps-1; s >= 0; s--) {
972 if (s == 0 && RtsFlags.GcFlags.generations > 1) {
976 if (ws->todo_large_objects) return rtsTrue;
977 if (ws->step->todos) return rtsTrue;
986 scavenge_until_all_done (void)
990 debugTrace(DEBUG_gc, "GC thread %d working", gct->thread_index);
993 #if defined(THREADED_RTS)
994 if (n_gc_threads > 1) {
1003 // scavenge_loop() only exits when there's no work to do
1006 debugTrace(DEBUG_gc, "GC thread %d idle (%d still running)",
1007 gct->thread_index, r);
1009 while (gc_running_threads != 0) {
1015 // any_work() does not remove the work from the queue, it
1016 // just checks for the presence of work. If we find any,
1017 // then we increment gc_running_threads and go back to
1018 // scavenge_loop() to perform any pending work.
1021 // All threads are now stopped
1022 debugTrace(DEBUG_gc, "GC thread %d finished.", gct->thread_index);
1025 #if defined(THREADED_RTS)
1028 gcWorkerThread (Capability *cap)
1030 cap->in_gc = rtsTrue;
1032 gct = gc_threads[cap->no];
1033 gct->id = osThreadId();
1035 // Wait until we're told to wake up
1036 RELEASE_SPIN_LOCK(&gct->mut_spin);
1037 gct->wakeup = GC_THREAD_STANDING_BY;
1038 debugTrace(DEBUG_gc, "GC thread %d standing by...", gct->thread_index);
1039 ACQUIRE_SPIN_LOCK(&gct->gc_spin);
1042 // start performance counters in this thread...
1043 if (gct->papi_events == -1) {
1044 papi_init_eventset(&gct->papi_events);
1046 papi_thread_start_gc1_count(gct->papi_events);
1049 // Every thread evacuates some roots.
1051 markSomeCapabilities(mark_root, gct, gct->thread_index, n_gc_threads,
1052 rtsTrue/*prune sparks*/);
1053 scavenge_capability_mut_lists(&capabilities[gct->thread_index]);
1055 scavenge_until_all_done();
1058 // count events in this thread towards the GC totals
1059 papi_thread_stop_gc1_count(gct->papi_events);
1062 // Wait until we're told to continue
1063 RELEASE_SPIN_LOCK(&gct->gc_spin);
1064 gct->wakeup = GC_THREAD_WAITING_TO_CONTINUE;
1065 debugTrace(DEBUG_gc, "GC thread %d waiting to continue...",
1067 ACQUIRE_SPIN_LOCK(&gct->mut_spin);
1068 debugTrace(DEBUG_gc, "GC thread %d on my way...", gct->thread_index);
1074 waitForGcThreads (Capability *cap USED_IF_THREADS)
1076 #if defined(THREADED_RTS)
1077 nat n_threads = RtsFlags.ParFlags.nNodes;
1080 rtsBool retry = rtsTrue;
1083 for (i=0; i < n_threads; i++) {
1084 if (i == me) continue;
1085 if (gc_threads[i]->wakeup != GC_THREAD_STANDING_BY) {
1086 prodCapability(&capabilities[i], cap->running_task);
1089 for (j=0; j < 10000000; j++) {
1091 for (i=0; i < n_threads; i++) {
1092 if (i == me) continue;
1094 setContextSwitches();
1095 if (gc_threads[i]->wakeup != GC_THREAD_STANDING_BY) {
1106 start_gc_threads (void)
1108 #if defined(THREADED_RTS)
1109 gc_running_threads = 0;
1110 initMutex(&gc_running_mutex);
1115 wakeup_gc_threads (nat n_threads USED_IF_THREADS, nat me USED_IF_THREADS)
1117 #if defined(THREADED_RTS)
1119 for (i=0; i < n_threads; i++) {
1120 if (i == me) continue;
1122 debugTrace(DEBUG_gc, "waking up gc thread %d", i);
1123 if (gc_threads[i]->wakeup != GC_THREAD_STANDING_BY) barf("wakeup_gc_threads");
1125 gc_threads[i]->wakeup = GC_THREAD_RUNNING;
1126 ACQUIRE_SPIN_LOCK(&gc_threads[i]->mut_spin);
1127 RELEASE_SPIN_LOCK(&gc_threads[i]->gc_spin);
1132 // After GC is complete, we must wait for all GC threads to enter the
1133 // standby state, otherwise they may still be executing inside
1134 // any_work(), and may even remain awake until the next GC starts.
1136 shutdown_gc_threads (nat n_threads USED_IF_THREADS, nat me USED_IF_THREADS)
1138 #if defined(THREADED_RTS)
1140 for (i=0; i < n_threads; i++) {
1141 if (i == me) continue;
1142 while (gc_threads[i]->wakeup != GC_THREAD_WAITING_TO_CONTINUE) { write_barrier(); }
1148 continue_gc_threads (nat n_threads USED_IF_THREADS, nat me USED_IF_THREADS)
1150 #if defined(THREADED_RTS)
1152 for (i=0; i < n_threads; i++) {
1153 if (i == me) continue;
1154 if (gc_threads[i]->wakeup != GC_THREAD_WAITING_TO_CONTINUE) barf("continue_gc_threads");
1156 gc_threads[i]->wakeup = GC_THREAD_INACTIVE;
1157 ACQUIRE_SPIN_LOCK(&gc_threads[i]->gc_spin);
1158 RELEASE_SPIN_LOCK(&gc_threads[i]->mut_spin);
1163 /* ----------------------------------------------------------------------------
1164 Initialise a generation that is to be collected
1165 ------------------------------------------------------------------------- */
1168 init_collected_gen (nat g, nat n_threads)
1175 // Throw away the current mutable list. Invariant: the mutable
1176 // list always has at least one block; this means we can avoid a
1177 // check for NULL in recordMutable().
1179 freeChain(generations[g].mut_list);
1180 generations[g].mut_list = allocBlock();
1181 for (i = 0; i < n_capabilities; i++) {
1182 freeChain(capabilities[i].mut_lists[g]);
1183 capabilities[i].mut_lists[g] = allocBlock();
1187 for (s = 0; s < generations[g].n_steps; s++) {
1189 stp = &generations[g].steps[s];
1190 ASSERT(stp->gen_no == g);
1192 // we'll construct a new list of threads in this step
1193 // during GC, throw away the current list.
1194 stp->old_threads = stp->threads;
1195 stp->threads = END_TSO_QUEUE;
1197 // generation 0, step 0 doesn't need to-space
1198 if (g == 0 && s == 0 && RtsFlags.GcFlags.generations > 1) {
1202 // deprecate the existing blocks
1203 stp->old_blocks = stp->blocks;
1204 stp->n_old_blocks = stp->n_blocks;
1208 stp->live_estimate = 0;
1210 // we don't have any to-be-scavenged blocks yet
1212 stp->todos_last = NULL;
1215 // initialise the large object queues.
1216 stp->scavenged_large_objects = NULL;
1217 stp->n_scavenged_large_blocks = 0;
1219 // mark the small objects as from-space
1220 for (bd = stp->old_blocks; bd; bd = bd->link) {
1221 bd->flags &= ~BF_EVACUATED;
1224 // mark the large objects as from-space
1225 for (bd = stp->large_objects; bd; bd = bd->link) {
1226 bd->flags &= ~BF_EVACUATED;
1229 // for a compacted step, we need to allocate the bitmap
1231 nat bitmap_size; // in bytes
1232 bdescr *bitmap_bdescr;
1235 bitmap_size = stp->n_old_blocks * BLOCK_SIZE / (sizeof(W_)*BITS_PER_BYTE);
1237 if (bitmap_size > 0) {
1238 bitmap_bdescr = allocGroup((lnat)BLOCK_ROUND_UP(bitmap_size)
1240 stp->bitmap = bitmap_bdescr;
1241 bitmap = bitmap_bdescr->start;
1243 debugTrace(DEBUG_gc, "bitmap_size: %d, bitmap: %p",
1244 bitmap_size, bitmap);
1246 // don't forget to fill it with zeros!
1247 memset(bitmap, 0, bitmap_size);
1249 // For each block in this step, point to its bitmap from the
1250 // block descriptor.
1251 for (bd=stp->old_blocks; bd != NULL; bd = bd->link) {
1252 bd->u.bitmap = bitmap;
1253 bitmap += BLOCK_SIZE_W / (sizeof(W_)*BITS_PER_BYTE);
1255 // Also at this point we set the BF_MARKED flag
1256 // for this block. The invariant is that
1257 // BF_MARKED is always unset, except during GC
1258 // when it is set on those blocks which will be
1260 if (!(bd->flags & BF_FRAGMENTED)) {
1261 bd->flags |= BF_MARKED;
1268 // For each GC thread, for each step, allocate a "todo" block to
1269 // store evacuated objects to be scavenged, and a block to store
1270 // evacuated objects that do not need to be scavenged.
1271 for (t = 0; t < n_threads; t++) {
1272 for (s = 0; s < generations[g].n_steps; s++) {
1274 // we don't copy objects into g0s0, unless -G0
1275 if (g==0 && s==0 && RtsFlags.GcFlags.generations > 1) continue;
1277 ws = &gc_threads[t]->steps[g * RtsFlags.GcFlags.steps + s];
1279 ws->todo_large_objects = NULL;
1281 ws->part_list = NULL;
1282 ws->n_part_blocks = 0;
1284 // allocate the first to-space block; extra blocks will be
1285 // chained on as necessary.
1287 ws->buffer_todo_bd = NULL;
1288 alloc_todo_block(ws,0);
1290 ws->scavd_list = NULL;
1291 ws->n_scavd_blocks = 0;
1297 /* ----------------------------------------------------------------------------
1298 Initialise a generation that is *not* to be collected
1299 ------------------------------------------------------------------------- */
1302 init_uncollected_gen (nat g, nat threads)
1309 // save the current mutable lists for this generation, and
1310 // allocate a fresh block for each one. We'll traverse these
1311 // mutable lists as roots early on in the GC.
1312 generations[g].saved_mut_list = generations[g].mut_list;
1313 generations[g].mut_list = allocBlock();
1314 for (n = 0; n < n_capabilities; n++) {
1315 capabilities[n].saved_mut_lists[g] = capabilities[n].mut_lists[g];
1316 capabilities[n].mut_lists[g] = allocBlock();
1319 for (s = 0; s < generations[g].n_steps; s++) {
1320 stp = &generations[g].steps[s];
1321 stp->scavenged_large_objects = NULL;
1322 stp->n_scavenged_large_blocks = 0;
1325 for (s = 0; s < generations[g].n_steps; s++) {
1327 stp = &generations[g].steps[s];
1329 for (t = 0; t < threads; t++) {
1330 ws = &gc_threads[t]->steps[g * RtsFlags.GcFlags.steps + s];
1332 ws->buffer_todo_bd = NULL;
1333 ws->todo_large_objects = NULL;
1335 ws->part_list = NULL;
1336 ws->n_part_blocks = 0;
1338 ws->scavd_list = NULL;
1339 ws->n_scavd_blocks = 0;
1341 // If the block at the head of the list in this generation
1342 // is less than 3/4 full, then use it as a todo block.
1343 if (stp->blocks && isPartiallyFull(stp->blocks))
1345 ws->todo_bd = stp->blocks;
1346 ws->todo_free = ws->todo_bd->free;
1347 ws->todo_lim = ws->todo_bd->start + BLOCK_SIZE_W;
1348 stp->blocks = stp->blocks->link;
1350 stp->n_words -= ws->todo_bd->free - ws->todo_bd->start;
1351 ws->todo_bd->link = NULL;
1352 // we must scan from the current end point.
1353 ws->todo_bd->u.scan = ws->todo_bd->free;
1358 alloc_todo_block(ws,0);
1362 // deal out any more partial blocks to the threads' part_lists
1364 while (stp->blocks && isPartiallyFull(stp->blocks))
1367 stp->blocks = bd->link;
1368 ws = &gc_threads[t]->steps[g * RtsFlags.GcFlags.steps + s];
1369 bd->link = ws->part_list;
1371 ws->n_part_blocks += 1;
1372 bd->u.scan = bd->free;
1374 stp->n_words -= bd->free - bd->start;
1376 if (t == n_gc_threads) t = 0;
1381 /* -----------------------------------------------------------------------------
1382 Initialise a gc_thread before GC
1383 -------------------------------------------------------------------------- */
1386 init_gc_thread (gc_thread *t)
1388 t->static_objects = END_OF_STATIC_LIST;
1389 t->scavenged_static_objects = END_OF_STATIC_LIST;
1391 t->mut_lists = capabilities[t->thread_index].mut_lists;
1393 t->failed_to_evac = rtsFalse;
1394 t->eager_promotion = rtsTrue;
1395 t->thunk_selector_depth = 0;
1400 t->scav_find_work = 0;
1403 /* -----------------------------------------------------------------------------
1404 Function we pass to evacuate roots.
1405 -------------------------------------------------------------------------- */
1408 mark_root(void *user, StgClosure **root)
1410 // we stole a register for gct, but this function is called from
1411 // *outside* the GC where the register variable is not in effect,
1412 // so we need to save and restore it here. NB. only call
1413 // mark_root() from the main GC thread, otherwise gct will be
1415 gc_thread *saved_gct;
1424 /* -----------------------------------------------------------------------------
1425 Initialising the static object & mutable lists
1426 -------------------------------------------------------------------------- */
1429 zero_static_object_list(StgClosure* first_static)
1433 const StgInfoTable *info;
1435 for (p = first_static; p != END_OF_STATIC_LIST; p = link) {
1437 link = *STATIC_LINK(info, p);
1438 *STATIC_LINK(info,p) = NULL;
1442 /* ----------------------------------------------------------------------------
1443 Update the pointers from the task list
1445 These are treated as weak pointers because we want to allow a main
1446 thread to get a BlockedOnDeadMVar exception in the same way as any
1447 other thread. Note that the threads should all have been retained
1448 by GC by virtue of being on the all_threads list, we're just
1449 updating pointers here.
1450 ------------------------------------------------------------------------- */
1453 update_task_list (void)
1457 for (task = all_tasks; task != NULL; task = task->all_link) {
1458 if (!task->stopped && task->tso) {
1459 ASSERT(task->tso->bound == task);
1460 tso = (StgTSO *) isAlive((StgClosure *)task->tso);
1462 barf("task %p: main thread %d has been GC'd",
1475 /* ----------------------------------------------------------------------------
1476 Reset the sizes of the older generations when we do a major
1479 CURRENT STRATEGY: make all generations except zero the same size.
1480 We have to stay within the maximum heap size, and leave a certain
1481 percentage of the maximum heap size available to allocate into.
1482 ------------------------------------------------------------------------- */
1485 resize_generations (void)
1489 if (major_gc && RtsFlags.GcFlags.generations > 1) {
1490 nat live, size, min_alloc, words;
1491 nat max = RtsFlags.GcFlags.maxHeapSize;
1492 nat gens = RtsFlags.GcFlags.generations;
1494 // live in the oldest generations
1495 if (oldest_gen->steps[0].live_estimate != 0) {
1496 words = oldest_gen->steps[0].live_estimate;
1498 words = oldest_gen->steps[0].n_words;
1500 live = (words + BLOCK_SIZE_W - 1) / BLOCK_SIZE_W +
1501 oldest_gen->steps[0].n_large_blocks;
1503 // default max size for all generations except zero
1504 size = stg_max(live * RtsFlags.GcFlags.oldGenFactor,
1505 RtsFlags.GcFlags.minOldGenSize);
1507 // minimum size for generation zero
1508 min_alloc = stg_max((RtsFlags.GcFlags.pcFreeHeap * max) / 200,
1509 RtsFlags.GcFlags.minAllocAreaSize);
1511 // Auto-enable compaction when the residency reaches a
1512 // certain percentage of the maximum heap size (default: 30%).
1513 if (RtsFlags.GcFlags.generations > 1 &&
1514 (RtsFlags.GcFlags.compact ||
1516 oldest_gen->steps[0].n_blocks >
1517 (RtsFlags.GcFlags.compactThreshold * max) / 100))) {
1518 oldest_gen->steps[0].mark = 1;
1519 oldest_gen->steps[0].compact = 1;
1520 // debugBelch("compaction: on\n", live);
1522 oldest_gen->steps[0].mark = 0;
1523 oldest_gen->steps[0].compact = 0;
1524 // debugBelch("compaction: off\n", live);
1527 if (RtsFlags.GcFlags.sweep) {
1528 oldest_gen->steps[0].mark = 1;
1531 // if we're going to go over the maximum heap size, reduce the
1532 // size of the generations accordingly. The calculation is
1533 // different if compaction is turned on, because we don't need
1534 // to double the space required to collect the old generation.
1537 // this test is necessary to ensure that the calculations
1538 // below don't have any negative results - we're working
1539 // with unsigned values here.
1540 if (max < min_alloc) {
1544 if (oldest_gen->steps[0].compact) {
1545 if ( (size + (size - 1) * (gens - 2) * 2) + min_alloc > max ) {
1546 size = (max - min_alloc) / ((gens - 1) * 2 - 1);
1549 if ( (size * (gens - 1) * 2) + min_alloc > max ) {
1550 size = (max - min_alloc) / ((gens - 1) * 2);
1560 debugBelch("live: %d, min_alloc: %d, size : %d, max = %d\n", live,
1561 min_alloc, size, max);
1564 for (g = 0; g < gens; g++) {
1565 generations[g].max_blocks = size;
1570 /* -----------------------------------------------------------------------------
1571 Calculate the new size of the nursery, and resize it.
1572 -------------------------------------------------------------------------- */
1575 resize_nursery (void)
1577 if (RtsFlags.GcFlags.generations == 1)
1578 { // Two-space collector:
1581 /* set up a new nursery. Allocate a nursery size based on a
1582 * function of the amount of live data (by default a factor of 2)
1583 * Use the blocks from the old nursery if possible, freeing up any
1586 * If we get near the maximum heap size, then adjust our nursery
1587 * size accordingly. If the nursery is the same size as the live
1588 * data (L), then we need 3L bytes. We can reduce the size of the
1589 * nursery to bring the required memory down near 2L bytes.
1591 * A normal 2-space collector would need 4L bytes to give the same
1592 * performance we get from 3L bytes, reducing to the same
1593 * performance at 2L bytes.
1595 blocks = g0s0->n_blocks;
1597 if ( RtsFlags.GcFlags.maxHeapSize != 0 &&
1598 blocks * RtsFlags.GcFlags.oldGenFactor * 2 >
1599 RtsFlags.GcFlags.maxHeapSize )
1601 long adjusted_blocks; // signed on purpose
1604 adjusted_blocks = (RtsFlags.GcFlags.maxHeapSize - 2 * blocks);
1606 debugTrace(DEBUG_gc, "near maximum heap size of 0x%x blocks, blocks = %d, adjusted to %ld",
1607 RtsFlags.GcFlags.maxHeapSize, blocks, adjusted_blocks);
1609 pc_free = adjusted_blocks * 100 / RtsFlags.GcFlags.maxHeapSize;
1610 if (pc_free < RtsFlags.GcFlags.pcFreeHeap) /* might even * be < 0 */
1614 blocks = adjusted_blocks;
1618 blocks *= RtsFlags.GcFlags.oldGenFactor;
1619 if (blocks < RtsFlags.GcFlags.minAllocAreaSize)
1621 blocks = RtsFlags.GcFlags.minAllocAreaSize;
1624 resizeNurseries(blocks);
1626 else // Generational collector
1629 * If the user has given us a suggested heap size, adjust our
1630 * allocation area to make best use of the memory available.
1632 if (RtsFlags.GcFlags.heapSizeSuggestion)
1635 nat needed = calcNeeded(); // approx blocks needed at next GC
1637 /* Guess how much will be live in generation 0 step 0 next time.
1638 * A good approximation is obtained by finding the
1639 * percentage of g0s0 that was live at the last minor GC.
1641 * We have an accurate figure for the amount of copied data in
1642 * 'copied', but we must convert this to a number of blocks, with
1643 * a small adjustment for estimated slop at the end of a block
1648 g0s0_pcnt_kept = ((copied / (BLOCK_SIZE_W - 10)) * 100)
1649 / countNurseryBlocks();
1652 /* Estimate a size for the allocation area based on the
1653 * information available. We might end up going slightly under
1654 * or over the suggested heap size, but we should be pretty
1657 * Formula: suggested - needed
1658 * ----------------------------
1659 * 1 + g0s0_pcnt_kept/100
1661 * where 'needed' is the amount of memory needed at the next
1662 * collection for collecting all steps except g0s0.
1665 (((long)RtsFlags.GcFlags.heapSizeSuggestion - (long)needed) * 100) /
1666 (100 + (long)g0s0_pcnt_kept);
1668 if (blocks < (long)RtsFlags.GcFlags.minAllocAreaSize) {
1669 blocks = RtsFlags.GcFlags.minAllocAreaSize;
1672 resizeNurseries((nat)blocks);
1676 // we might have added extra large blocks to the nursery, so
1677 // resize back to minAllocAreaSize again.
1678 resizeNurseriesFixed(RtsFlags.GcFlags.minAllocAreaSize);
1683 /* -----------------------------------------------------------------------------
1684 Sanity code for CAF garbage collection.
1686 With DEBUG turned on, we manage a CAF list in addition to the SRT
1687 mechanism. After GC, we run down the CAF list and blackhole any
1688 CAFs which have been garbage collected. This means we get an error
1689 whenever the program tries to enter a garbage collected CAF.
1691 Any garbage collected CAFs are taken off the CAF list at the same
1693 -------------------------------------------------------------------------- */
1695 #if 0 && defined(DEBUG)
1702 const StgInfoTable *info;
1713 ASSERT(info->type == IND_STATIC);
1715 if (STATIC_LINK(info,p) == NULL) {
1716 debugTrace(DEBUG_gccafs, "CAF gc'd at 0x%04lx", (long)p);
1718 SET_INFO(p,&stg_BLACKHOLE_info);
1719 p = STATIC_LINK2(info,p);
1723 pp = &STATIC_LINK2(info,p);
1730 debugTrace(DEBUG_gccafs, "%d CAFs live", i);