1 /* -----------------------------------------------------------------------------
3 * (c) The GHC Team 1998-2008
5 * Generational garbage collector
7 * Documentation on the architecture of the Garbage Collector can be
8 * found in the online commentary:
10 * http://hackage.haskell.org/trac/ghc/wiki/Commentary/Rts/Storage/GC
12 * ---------------------------------------------------------------------------*/
14 #include "PosixSource.h"
25 #include "BlockAlloc.h"
29 #include "RtsSignals.h"
31 #if defined(RTS_GTK_FRONTPANEL)
32 #include "FrontPanel.h"
35 #include "RetainerProfile.h"
36 #include "LdvProfile.h"
37 #include "RaiseAsync.h"
47 #include "MarkStack.h"
52 #include <string.h> // for memset()
55 /* -----------------------------------------------------------------------------
57 -------------------------------------------------------------------------- */
59 /* STATIC OBJECT LIST.
62 * We maintain a linked list of static objects that are still live.
63 * The requirements for this list are:
65 * - we need to scan the list while adding to it, in order to
66 * scavenge all the static objects (in the same way that
67 * breadth-first scavenging works for dynamic objects).
69 * - we need to be able to tell whether an object is already on
70 * the list, to break loops.
72 * Each static object has a "static link field", which we use for
73 * linking objects on to the list. We use a stack-type list, consing
74 * objects on the front as they are added (this means that the
75 * scavenge phase is depth-first, not breadth-first, but that
78 * A separate list is kept for objects that have been scavenged
79 * already - this is so that we can zero all the marks afterwards.
81 * An object is on the list if its static link field is non-zero; this
82 * means that we have to mark the end of the list with '1', not NULL.
84 * Extra notes for generational GC:
86 * Each generation has a static object list associated with it. When
87 * collecting generations up to N, we treat the static object lists
88 * from generations > N as roots.
90 * We build up a static object list while collecting generations 0..N,
91 * which is then appended to the static object list of generation N+1.
94 /* N is the oldest generation being collected, where the generations
95 * are numbered starting at 0. A major GC (indicated by the major_gc
96 * flag) is when we're collecting all generations. We only attempt to
97 * deal with static objects and GC CAFs when doing a major GC.
102 /* Data used for allocation area sizing.
104 static lnat g0s0_pcnt_kept = 30; // percentage of g0s0 live at last minor GC
114 /* Thread-local data for each GC thread
116 gc_thread **gc_threads = NULL;
118 #if !defined(THREADED_RTS)
119 StgWord8 the_gc_thread[sizeof(gc_thread) + 64 * sizeof(step_workspace)];
122 // Number of threads running in *this* GC. Affects how many
123 // step->todos[] lists we have to look in to find work.
127 long copied; // *words* copied & scavenged during this GC
129 rtsBool work_stealing;
133 /* -----------------------------------------------------------------------------
134 Static function declarations
135 -------------------------------------------------------------------------- */
137 static void mark_root (void *user, StgClosure **root);
138 static void zero_static_object_list (StgClosure* first_static);
139 static nat initialise_N (rtsBool force_major_gc);
140 static void init_collected_gen (nat g, nat threads);
141 static void init_uncollected_gen (nat g, nat threads);
142 static void init_gc_thread (gc_thread *t);
143 static void update_task_list (void);
144 static void resize_generations (void);
145 static void resize_nursery (void);
146 static void start_gc_threads (void);
147 static void scavenge_until_all_done (void);
148 static StgWord inc_running (void);
149 static StgWord dec_running (void);
150 static void wakeup_gc_threads (nat n_threads, nat me);
151 static void shutdown_gc_threads (nat n_threads, nat me);
153 #if 0 && defined(DEBUG)
154 static void gcCAFs (void);
157 /* -----------------------------------------------------------------------------
159 -------------------------------------------------------------------------- */
161 bdescr *mark_stack_top_bd; // topmost block in the mark stack
162 bdescr *mark_stack_bd; // current block in the mark stack
163 StgPtr mark_sp; // pointer to the next unallocated mark stack entry
165 /* -----------------------------------------------------------------------------
166 GarbageCollect: the main entry point to the garbage collector.
168 Locks held: all capabilities are held throughout GarbageCollect().
169 -------------------------------------------------------------------------- */
172 GarbageCollect (rtsBool force_major_gc,
173 nat gc_type USED_IF_THREADS,
178 lnat live, allocated, max_copied, avg_copied, slop;
179 gc_thread *saved_gct;
182 // necessary if we stole a callee-saves register for gct:
186 CostCentreStack *prev_CCS;
191 #if defined(RTS_USER_SIGNALS)
192 if (RtsFlags.MiscFlags.install_signal_handlers) {
198 ASSERT(sizeof(step_workspace) == 16 * sizeof(StgWord));
199 // otherwise adjust the padding in step_workspace.
201 // tell the stats department that we've started a GC
204 // tell the STM to discard any cached closures it's hoping to re-use
207 // lock the StablePtr table
216 // attribute any costs to CCS_GC
222 /* Approximate how much we allocated.
223 * Todo: only when generating stats?
225 allocated = calcAllocated();
227 /* Figure out which generation to collect
229 n = initialise_N(force_major_gc);
231 #if defined(THREADED_RTS)
232 work_stealing = RtsFlags.ParFlags.parGcLoadBalancingEnabled &&
233 N >= RtsFlags.ParFlags.parGcLoadBalancingGen;
234 // It's not always a good idea to do load balancing in parallel
235 // GC. In particular, for a parallel program we don't want to
236 // lose locality by moving cached data into another CPU's cache
237 // (this effect can be quite significant).
239 // We could have a more complex way to deterimine whether to do
240 // work stealing or not, e.g. it might be a good idea to do it
241 // if the heap is big. For now, we just turn it on or off with
245 /* Start threads, so they can be spinning up while we finish initialisation.
249 #if defined(THREADED_RTS)
250 /* How many threads will be participating in this GC?
251 * We don't try to parallelise minor GCs (unless the user asks for
252 * it with +RTS -gn0), or mark/compact/sweep GC.
254 if (gc_type == PENDING_GC_PAR) {
255 n_gc_threads = RtsFlags.ParFlags.nNodes;
263 debugTrace(DEBUG_gc, "GC (gen %d): %d KB to collect, %ld MB in use, using %d thread(s)",
264 N, n * (BLOCK_SIZE / 1024), mblocks_allocated, n_gc_threads);
266 #ifdef RTS_GTK_FRONTPANEL
267 if (RtsFlags.GcFlags.frontpanel) {
268 updateFrontPanelBeforeGC(N);
273 // check for memory leaks if DEBUG is on
274 memInventory(DEBUG_gc);
277 // check sanity *before* GC
278 IF_DEBUG(sanity, checkSanity(rtsTrue));
280 // Initialise all our gc_thread structures
281 for (t = 0; t < n_gc_threads; t++) {
282 init_gc_thread(gc_threads[t]);
285 // Initialise all the generations/steps that we're collecting.
286 for (g = 0; g <= N; g++) {
287 init_collected_gen(g,n_gc_threads);
290 // Initialise all the generations/steps that we're *not* collecting.
291 for (g = N+1; g < RtsFlags.GcFlags.generations; g++) {
292 init_uncollected_gen(g,n_gc_threads);
295 /* Allocate a mark stack if we're doing a major collection.
297 if (major_gc && oldest_gen->steps[0].mark) {
298 mark_stack_bd = allocBlock();
299 mark_stack_top_bd = mark_stack_bd;
300 mark_stack_bd->link = NULL;
301 mark_stack_bd->u.back = NULL;
302 mark_sp = mark_stack_bd->start;
304 mark_stack_bd = NULL;
305 mark_stack_top_bd = NULL;
309 // this is the main thread
311 if (n_gc_threads == 1) {
312 SET_GCT(gc_threads[0]);
314 SET_GCT(gc_threads[cap->no]);
317 SET_GCT(gc_threads[0]);
320 /* -----------------------------------------------------------------------
321 * follow all the roots that we know about:
324 // the main thread is running: this prevents any other threads from
325 // exiting prematurely, so we can start them now.
326 // NB. do this after the mutable lists have been saved above, otherwise
327 // the other GC threads will be writing into the old mutable lists.
329 wakeup_gc_threads(n_gc_threads, gct->thread_index);
331 // Mutable lists from each generation > N
332 // we want to *scavenge* these roots, not evacuate them: they're not
333 // going to move in this GC.
334 // Also do them in reverse generation order, for the usual reason:
335 // namely to reduce the likelihood of spurious old->new pointers.
337 for (g = RtsFlags.GcFlags.generations-1; g > N; g--) {
338 scavenge_mutable_list(generations[g].saved_mut_list, &generations[g]);
339 freeChain_sync(generations[g].saved_mut_list);
340 generations[g].saved_mut_list = NULL;
344 // scavenge the capability-private mutable lists. This isn't part
345 // of markSomeCapabilities() because markSomeCapabilities() can only
346 // call back into the GC via mark_root() (due to the gct register
348 if (n_gc_threads == 1) {
349 for (n = 0; n < n_capabilities; n++) {
350 scavenge_capability_mut_lists(&capabilities[n]);
353 scavenge_capability_mut_lists(&capabilities[gct->thread_index]);
356 // follow roots from the CAF list (used by GHCi)
358 markCAFs(mark_root, gct);
360 // follow all the roots that the application knows about.
362 markSomeCapabilities(mark_root, gct, gct->thread_index, n_gc_threads,
363 rtsTrue/*prune sparks*/);
365 #if defined(RTS_USER_SIGNALS)
366 // mark the signal handlers (signals should be already blocked)
367 markSignalHandlers(mark_root, gct);
370 // Mark the weak pointer list, and prepare to detect dead weak pointers.
374 // Mark the stable pointer table.
375 markStablePtrTable(mark_root, gct);
377 /* -------------------------------------------------------------------------
378 * Repeatedly scavenge all the areas we know about until there's no
379 * more scavenging to be done.
383 scavenge_until_all_done();
384 // The other threads are now stopped. We might recurse back to
385 // here, but from now on this is the only thread.
387 // if any blackholes are alive, make the threads that wait on
389 if (traverseBlackholeQueue()) {
394 // must be last... invariant is that everything is fully
395 // scavenged at this point.
396 if (traverseWeakPtrList()) { // returns rtsTrue if evaced something
401 // If we get to here, there's really nothing left to do.
405 shutdown_gc_threads(n_gc_threads, gct->thread_index);
407 // Update pointers from the Task list
410 // Now see which stable names are still alive.
414 // We call processHeapClosureForDead() on every closure destroyed during
415 // the current garbage collection, so we invoke LdvCensusForDead().
416 if (RtsFlags.ProfFlags.doHeapProfile == HEAP_BY_LDV
417 || RtsFlags.ProfFlags.bioSelector != NULL)
421 // NO MORE EVACUATION AFTER THIS POINT!
423 // Two-space collector: free the old to-space.
424 // g0s0->old_blocks is the old nursery
425 // g0s0->blocks is to-space from the previous GC
426 if (RtsFlags.GcFlags.generations == 1) {
427 if (g0->steps[0].blocks != NULL) {
428 freeChain(g0->steps[0].blocks);
429 g0->steps[0].blocks = NULL;
433 // For each workspace, in each thread, move the copied blocks to the step
439 for (t = 0; t < n_gc_threads; t++) {
443 if (RtsFlags.GcFlags.generations == 1) {
448 for (; s < total_steps; s++) {
451 // Push the final block
453 push_scanned_block(ws->todo_bd, ws);
456 ASSERT(gct->scan_bd == NULL);
457 ASSERT(countBlocks(ws->scavd_list) == ws->n_scavd_blocks);
460 for (bd = ws->scavd_list; bd != NULL; bd = bd->link) {
461 ws->step->n_words += bd->free - bd->start;
465 prev->link = ws->step->blocks;
466 ws->step->blocks = ws->scavd_list;
468 ws->step->n_blocks += ws->n_scavd_blocks;
472 // Add all the partial blocks *after* we've added all the full
473 // blocks. This is so that we can grab the partial blocks back
474 // again and try to fill them up in the next GC.
475 for (t = 0; t < n_gc_threads; t++) {
479 if (RtsFlags.GcFlags.generations == 1) {
484 for (; s < total_steps; s++) {
488 for (bd = ws->part_list; bd != NULL; bd = next) {
490 if (bd->free == bd->start) {
492 ws->part_list = next;
499 ws->step->n_words += bd->free - bd->start;
504 prev->link = ws->step->blocks;
505 ws->step->blocks = ws->part_list;
507 ws->step->n_blocks += ws->n_part_blocks;
509 ASSERT(countBlocks(ws->step->blocks) == ws->step->n_blocks);
510 ASSERT(countOccupied(ws->step->blocks) == ws->step->n_words);
515 // Finally: compact or sweep the oldest generation.
516 if (major_gc && oldest_gen->steps[0].mark) {
517 if (oldest_gen->steps[0].compact)
518 compact(gct->scavenged_static_objects);
520 sweep(&oldest_gen->steps[0]);
523 /* run through all the generations/steps and tidy up
530 for (i=0; i < n_gc_threads; i++) {
531 if (n_gc_threads > 1) {
532 debugTrace(DEBUG_gc,"thread %d:", i);
533 debugTrace(DEBUG_gc," copied %ld", gc_threads[i]->copied * sizeof(W_));
534 debugTrace(DEBUG_gc," scanned %ld", gc_threads[i]->scanned * sizeof(W_));
535 debugTrace(DEBUG_gc," any_work %ld", gc_threads[i]->any_work);
536 debugTrace(DEBUG_gc," no_work %ld", gc_threads[i]->no_work);
537 debugTrace(DEBUG_gc," scav_find_work %ld", gc_threads[i]->scav_find_work);
539 copied += gc_threads[i]->copied;
540 max_copied = stg_max(gc_threads[i]->copied, max_copied);
542 if (n_gc_threads == 1) {
550 for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
553 generations[g].collections++; // for stats
554 if (n_gc_threads > 1) generations[g].par_collections++;
557 // Count the mutable list as bytes "copied" for the purposes of
558 // stats. Every mutable list is copied during every GC.
560 nat mut_list_size = 0;
561 for (bd = generations[g].mut_list; bd != NULL; bd = bd->link) {
562 mut_list_size += bd->free - bd->start;
564 for (n = 0; n < n_capabilities; n++) {
565 for (bd = capabilities[n].mut_lists[g];
566 bd != NULL; bd = bd->link) {
567 mut_list_size += bd->free - bd->start;
570 copied += mut_list_size;
573 "mut_list_size: %lu (%d vars, %d arrays, %d MVARs, %d others)",
574 (unsigned long)(mut_list_size * sizeof(W_)),
575 mutlist_MUTVARS, mutlist_MUTARRS, mutlist_MVARS, mutlist_OTHERS);
578 for (s = 0; s < generations[g].n_steps; s++) {
580 stp = &generations[g].steps[s];
582 // for generations we collected...
585 /* free old memory and shift to-space into from-space for all
586 * the collected steps (except the allocation area). These
587 * freed blocks will probaby be quickly recycled.
589 if (!(g == 0 && s == 0 && RtsFlags.GcFlags.generations > 1)) {
592 // tack the new blocks on the end of the existing blocks
593 if (stp->old_blocks != NULL) {
596 for (bd = stp->old_blocks; bd != NULL; bd = next) {
600 if (!(bd->flags & BF_MARKED))
603 stp->old_blocks = next;
612 stp->n_words += bd->free - bd->start;
614 // NB. this step might not be compacted next
615 // time, so reset the BF_MARKED flags.
616 // They are set before GC if we're going to
617 // compact. (search for BF_MARKED above).
618 bd->flags &= ~BF_MARKED;
620 // between GCs, all blocks in the heap except
621 // for the nursery have the BF_EVACUATED flag set.
622 bd->flags |= BF_EVACUATED;
629 prev->link = stp->blocks;
630 stp->blocks = stp->old_blocks;
633 // add the new blocks to the block tally
634 stp->n_blocks += stp->n_old_blocks;
635 ASSERT(countBlocks(stp->blocks) == stp->n_blocks);
636 ASSERT(countOccupied(stp->blocks) == stp->n_words);
640 freeChain(stp->old_blocks);
642 stp->old_blocks = NULL;
643 stp->n_old_blocks = 0;
646 /* LARGE OBJECTS. The current live large objects are chained on
647 * scavenged_large, having been moved during garbage
648 * collection from large_objects. Any objects left on the
649 * large_objects list are therefore dead, so we free them here.
651 freeChain(stp->large_objects);
652 stp->large_objects = stp->scavenged_large_objects;
653 stp->n_large_blocks = stp->n_scavenged_large_blocks;
654 ASSERT(countBlocks(stp->large_objects) == stp->n_large_blocks);
656 else // for older generations...
658 /* For older generations, we need to append the
659 * scavenged_large_object list (i.e. large objects that have been
660 * promoted during this GC) to the large_object list for that step.
662 for (bd = stp->scavenged_large_objects; bd; bd = next) {
664 dbl_link_onto(bd, &stp->large_objects);
667 // add the new blocks we promoted during this GC
668 stp->n_large_blocks += stp->n_scavenged_large_blocks;
669 ASSERT(countBlocks(stp->large_objects) == stp->n_large_blocks);
674 // update the max size of older generations after a major GC
675 resize_generations();
677 // Calculate the amount of live data for stats.
678 live = calcLiveWords();
680 // Free the small objects allocated via allocate(), since this will
681 // all have been copied into G0S1 now.
682 if (RtsFlags.GcFlags.generations > 1) {
683 if (g0->steps[0].blocks != NULL) {
684 freeChain(g0->steps[0].blocks);
685 g0->steps[0].blocks = NULL;
687 g0->steps[0].n_blocks = 0;
688 g0->steps[0].n_words = 0;
690 alloc_blocks_lim = RtsFlags.GcFlags.minAllocAreaSize;
692 // Start a new pinned_object_block
693 for (n = 0; n < n_capabilities; n++) {
694 capabilities[n].pinned_object_block = NULL;
697 // Free the mark stack.
698 if (mark_stack_top_bd != NULL) {
699 debugTrace(DEBUG_gc, "mark stack: %d blocks",
700 countBlocks(mark_stack_top_bd));
701 freeChain(mark_stack_top_bd);
705 for (g = 0; g <= N; g++) {
706 for (s = 0; s < generations[g].n_steps; s++) {
707 stp = &generations[g].steps[s];
708 if (stp->bitmap != NULL) {
709 freeGroup(stp->bitmap);
717 // mark the garbage collected CAFs as dead
718 #if 0 && defined(DEBUG) // doesn't work at the moment
719 if (major_gc) { gcCAFs(); }
723 // resetStaticObjectForRetainerProfiling() must be called before
725 if (n_gc_threads > 1) {
726 barf("profiling is currently broken with multi-threaded GC");
727 // ToDo: fix the gct->scavenged_static_objects below
729 resetStaticObjectForRetainerProfiling(gct->scavenged_static_objects);
732 // zero the scavenged static object list
735 for (i = 0; i < n_gc_threads; i++) {
736 zero_static_object_list(gc_threads[i]->scavenged_static_objects);
743 // start any pending finalizers
745 scheduleFinalizers(cap, old_weak_ptr_list);
748 // send exceptions to any threads which were about to die
750 resurrectThreads(resurrected_threads);
751 performPendingThrowTos(exception_threads);
754 // Update the stable pointer hash table.
755 updateStablePtrTable(major_gc);
757 // check sanity after GC
758 IF_DEBUG(sanity, checkSanity(rtsTrue));
760 // extra GC trace info
761 IF_DEBUG(gc, statDescribeGens());
764 // symbol-table based profiling
765 /* heapCensus(to_blocks); */ /* ToDo */
768 // restore enclosing cost centre
774 // check for memory leaks if DEBUG is on
775 memInventory(DEBUG_gc);
778 #ifdef RTS_GTK_FRONTPANEL
779 if (RtsFlags.GcFlags.frontpanel) {
780 updateFrontPanelAfterGC( N, live );
784 // ok, GC over: tell the stats department what happened.
785 slop = calcLiveBlocks() * BLOCK_SIZE_W - live;
786 stat_endGC(allocated, live, copied, N, max_copied, avg_copied, slop);
788 // unlock the StablePtr table
791 // Guess which generation we'll collect *next* time
792 initialise_N(force_major_gc);
794 #if defined(RTS_USER_SIGNALS)
795 if (RtsFlags.MiscFlags.install_signal_handlers) {
796 // unblock signals again
797 unblockUserSignals();
806 /* -----------------------------------------------------------------------------
807 Figure out which generation to collect, initialise N and major_gc.
809 Also returns the total number of blocks in generations that will be
811 -------------------------------------------------------------------------- */
814 initialise_N (rtsBool force_major_gc)
817 nat s, blocks, blocks_total;
822 if (force_major_gc) {
823 N = RtsFlags.GcFlags.generations - 1;
828 for (g = RtsFlags.GcFlags.generations - 1; g >= 0; g--) {
830 for (s = 0; s < generations[g].n_steps; s++) {
831 blocks += generations[g].steps[s].n_words / BLOCK_SIZE_W;
832 blocks += generations[g].steps[s].n_large_blocks;
834 if (blocks >= generations[g].max_blocks) {
838 blocks_total += blocks;
842 blocks_total += countNurseryBlocks();
844 major_gc = (N == RtsFlags.GcFlags.generations-1);
848 /* -----------------------------------------------------------------------------
849 Initialise the gc_thread structures.
850 -------------------------------------------------------------------------- */
852 #define GC_THREAD_INACTIVE 0
853 #define GC_THREAD_STANDING_BY 1
854 #define GC_THREAD_RUNNING 2
855 #define GC_THREAD_WAITING_TO_CONTINUE 3
858 new_gc_thread (nat n, gc_thread *t)
865 initSpinLock(&t->gc_spin);
866 initSpinLock(&t->mut_spin);
867 ACQUIRE_SPIN_LOCK(&t->gc_spin);
868 t->wakeup = GC_THREAD_INACTIVE; // starts true, so we can wait for the
869 // thread to start up, see wakeup_gc_threads
873 t->free_blocks = NULL;
882 for (s = 0; s < total_steps; s++)
885 ws->step = &all_steps[s];
886 ASSERT(s == ws->step->abs_no);
890 ws->todo_q = newWSDeque(128);
891 ws->todo_overflow = NULL;
892 ws->n_todo_overflow = 0;
894 ws->part_list = NULL;
895 ws->n_part_blocks = 0;
897 ws->scavd_list = NULL;
898 ws->n_scavd_blocks = 0;
906 if (gc_threads == NULL) {
907 #if defined(THREADED_RTS)
909 gc_threads = stgMallocBytes (RtsFlags.ParFlags.nNodes *
913 for (i = 0; i < RtsFlags.ParFlags.nNodes; i++) {
915 stgMallocBytes(sizeof(gc_thread) + total_steps * sizeof(step_workspace),
918 new_gc_thread(i, gc_threads[i]);
921 gc_threads = stgMallocBytes (sizeof(gc_thread*),"alloc_gc_threads");
923 new_gc_thread(0,gc_threads[0]);
932 if (gc_threads != NULL) {
933 #if defined(THREADED_RTS)
935 for (i = 0; i < n_capabilities; i++) {
936 for (s = 0; s < total_steps; s++)
938 freeWSDeque(gc_threads[i]->steps[s].todo_q);
940 stgFree (gc_threads[i]);
942 stgFree (gc_threads);
944 for (s = 0; s < total_steps; s++)
946 freeWSDeque(gc_threads[0]->steps[s].todo_q);
948 stgFree (gc_threads);
954 /* ----------------------------------------------------------------------------
956 ------------------------------------------------------------------------- */
958 static volatile StgWord gc_running_threads;
964 new = atomic_inc(&gc_running_threads);
965 ASSERT(new <= n_gc_threads);
972 ASSERT(gc_running_threads != 0);
973 return atomic_dec(&gc_running_threads);
986 // scavenge objects in compacted generation
987 if (mark_stack_bd != NULL && !mark_stack_empty()) {
991 // Check for global work in any step. We don't need to check for
992 // local work, because we have already exited scavenge_loop(),
993 // which means there is no local work for this thread.
994 for (s = total_steps-1; s >= 0; s--) {
995 if (s == 0 && RtsFlags.GcFlags.generations > 1) {
999 if (ws->todo_large_objects) return rtsTrue;
1000 if (!looksEmptyWSDeque(ws->todo_q)) return rtsTrue;
1001 if (ws->todo_overflow) return rtsTrue;
1004 #if defined(THREADED_RTS)
1005 if (work_stealing) {
1007 // look for work to steal
1008 for (n = 0; n < n_gc_threads; n++) {
1009 if (n == gct->thread_index) continue;
1010 for (s = total_steps-1; s >= 0; s--) {
1011 ws = &gc_threads[n]->steps[s];
1012 if (!looksEmptyWSDeque(ws->todo_q)) return rtsTrue;
1024 scavenge_until_all_done (void)
1030 traceEvent(&capabilities[gct->thread_index], EVENT_GC_WORK);
1032 #if defined(THREADED_RTS)
1033 if (n_gc_threads > 1) {
1042 // scavenge_loop() only exits when there's no work to do
1045 traceEvent(&capabilities[gct->thread_index], EVENT_GC_IDLE);
1047 debugTrace(DEBUG_gc, "%d GC threads still running", r);
1049 while (gc_running_threads != 0) {
1055 // any_work() does not remove the work from the queue, it
1056 // just checks for the presence of work. If we find any,
1057 // then we increment gc_running_threads and go back to
1058 // scavenge_loop() to perform any pending work.
1061 traceEvent(&capabilities[gct->thread_index], EVENT_GC_DONE);
1064 #if defined(THREADED_RTS)
1067 gcWorkerThread (Capability *cap)
1069 gc_thread *saved_gct;
1071 // necessary if we stole a callee-saves register for gct:
1074 cap->in_gc = rtsTrue;
1076 gct = gc_threads[cap->no];
1077 gct->id = osThreadId();
1079 // Wait until we're told to wake up
1080 RELEASE_SPIN_LOCK(&gct->mut_spin);
1081 gct->wakeup = GC_THREAD_STANDING_BY;
1082 debugTrace(DEBUG_gc, "GC thread %d standing by...", gct->thread_index);
1083 ACQUIRE_SPIN_LOCK(&gct->gc_spin);
1086 // start performance counters in this thread...
1087 if (gct->papi_events == -1) {
1088 papi_init_eventset(&gct->papi_events);
1090 papi_thread_start_gc1_count(gct->papi_events);
1093 // Every thread evacuates some roots.
1095 markSomeCapabilities(mark_root, gct, gct->thread_index, n_gc_threads,
1096 rtsTrue/*prune sparks*/);
1097 scavenge_capability_mut_lists(&capabilities[gct->thread_index]);
1099 scavenge_until_all_done();
1102 // count events in this thread towards the GC totals
1103 papi_thread_stop_gc1_count(gct->papi_events);
1106 // Wait until we're told to continue
1107 RELEASE_SPIN_LOCK(&gct->gc_spin);
1108 gct->wakeup = GC_THREAD_WAITING_TO_CONTINUE;
1109 debugTrace(DEBUG_gc, "GC thread %d waiting to continue...",
1111 ACQUIRE_SPIN_LOCK(&gct->mut_spin);
1112 debugTrace(DEBUG_gc, "GC thread %d on my way...", gct->thread_index);
1119 #if defined(THREADED_RTS)
1122 waitForGcThreads (Capability *cap USED_IF_THREADS)
1124 nat n_threads = RtsFlags.ParFlags.nNodes;
1127 rtsBool retry = rtsTrue;
1130 for (i=0; i < n_threads; i++) {
1131 if (i == me) continue;
1132 if (gc_threads[i]->wakeup != GC_THREAD_STANDING_BY) {
1133 prodCapability(&capabilities[i], cap->running_task);
1136 for (j=0; j < 10; j++) {
1138 for (i=0; i < n_threads; i++) {
1139 if (i == me) continue;
1141 setContextSwitches();
1142 if (gc_threads[i]->wakeup != GC_THREAD_STANDING_BY) {
1152 #endif // THREADED_RTS
1155 start_gc_threads (void)
1157 #if defined(THREADED_RTS)
1158 gc_running_threads = 0;
1163 wakeup_gc_threads (nat n_threads USED_IF_THREADS, nat me USED_IF_THREADS)
1165 #if defined(THREADED_RTS)
1167 for (i=0; i < n_threads; i++) {
1168 if (i == me) continue;
1170 debugTrace(DEBUG_gc, "waking up gc thread %d", i);
1171 if (gc_threads[i]->wakeup != GC_THREAD_STANDING_BY) barf("wakeup_gc_threads");
1173 gc_threads[i]->wakeup = GC_THREAD_RUNNING;
1174 ACQUIRE_SPIN_LOCK(&gc_threads[i]->mut_spin);
1175 RELEASE_SPIN_LOCK(&gc_threads[i]->gc_spin);
1180 // After GC is complete, we must wait for all GC threads to enter the
1181 // standby state, otherwise they may still be executing inside
1182 // any_work(), and may even remain awake until the next GC starts.
1184 shutdown_gc_threads (nat n_threads USED_IF_THREADS, nat me USED_IF_THREADS)
1186 #if defined(THREADED_RTS)
1188 for (i=0; i < n_threads; i++) {
1189 if (i == me) continue;
1190 while (gc_threads[i]->wakeup != GC_THREAD_WAITING_TO_CONTINUE) { write_barrier(); }
1195 #if defined(THREADED_RTS)
1197 releaseGCThreads (Capability *cap USED_IF_THREADS)
1199 nat n_threads = RtsFlags.ParFlags.nNodes;
1202 for (i=0; i < n_threads; i++) {
1203 if (i == me) continue;
1204 if (gc_threads[i]->wakeup != GC_THREAD_WAITING_TO_CONTINUE)
1205 barf("releaseGCThreads");
1207 gc_threads[i]->wakeup = GC_THREAD_INACTIVE;
1208 ACQUIRE_SPIN_LOCK(&gc_threads[i]->gc_spin);
1209 RELEASE_SPIN_LOCK(&gc_threads[i]->mut_spin);
1214 /* ----------------------------------------------------------------------------
1215 Initialise a generation that is to be collected
1216 ------------------------------------------------------------------------- */
1219 init_collected_gen (nat g, nat n_threads)
1226 // Throw away the current mutable list. Invariant: the mutable
1227 // list always has at least one block; this means we can avoid a
1228 // check for NULL in recordMutable().
1230 freeChain(generations[g].mut_list);
1231 generations[g].mut_list = allocBlock();
1232 for (i = 0; i < n_capabilities; i++) {
1233 freeChain(capabilities[i].mut_lists[g]);
1234 capabilities[i].mut_lists[g] = allocBlock();
1239 for (i = 0; i < n_capabilities; i++) {
1240 stp = &nurseries[i];
1241 stp->old_threads = stp->threads;
1242 stp->threads = END_TSO_QUEUE;
1246 for (s = 0; s < generations[g].n_steps; s++) {
1248 // generation 0, step 0 doesn't need to-space, unless -G1
1249 if (g == 0 && s == 0 && RtsFlags.GcFlags.generations > 1) {
1253 stp = &generations[g].steps[s];
1254 ASSERT(stp->gen_no == g);
1256 // we'll construct a new list of threads in this step
1257 // during GC, throw away the current list.
1258 stp->old_threads = stp->threads;
1259 stp->threads = END_TSO_QUEUE;
1261 // deprecate the existing blocks
1262 stp->old_blocks = stp->blocks;
1263 stp->n_old_blocks = stp->n_blocks;
1267 stp->live_estimate = 0;
1269 // initialise the large object queues.
1270 stp->scavenged_large_objects = NULL;
1271 stp->n_scavenged_large_blocks = 0;
1273 // mark the small objects as from-space
1274 for (bd = stp->old_blocks; bd; bd = bd->link) {
1275 bd->flags &= ~BF_EVACUATED;
1278 // mark the large objects as from-space
1279 for (bd = stp->large_objects; bd; bd = bd->link) {
1280 bd->flags &= ~BF_EVACUATED;
1283 // for a compacted step, we need to allocate the bitmap
1285 nat bitmap_size; // in bytes
1286 bdescr *bitmap_bdescr;
1289 bitmap_size = stp->n_old_blocks * BLOCK_SIZE / (sizeof(W_)*BITS_PER_BYTE);
1291 if (bitmap_size > 0) {
1292 bitmap_bdescr = allocGroup((lnat)BLOCK_ROUND_UP(bitmap_size)
1294 stp->bitmap = bitmap_bdescr;
1295 bitmap = bitmap_bdescr->start;
1297 debugTrace(DEBUG_gc, "bitmap_size: %d, bitmap: %p",
1298 bitmap_size, bitmap);
1300 // don't forget to fill it with zeros!
1301 memset(bitmap, 0, bitmap_size);
1303 // For each block in this step, point to its bitmap from the
1304 // block descriptor.
1305 for (bd=stp->old_blocks; bd != NULL; bd = bd->link) {
1306 bd->u.bitmap = bitmap;
1307 bitmap += BLOCK_SIZE_W / (sizeof(W_)*BITS_PER_BYTE);
1309 // Also at this point we set the BF_MARKED flag
1310 // for this block. The invariant is that
1311 // BF_MARKED is always unset, except during GC
1312 // when it is set on those blocks which will be
1314 if (!(bd->flags & BF_FRAGMENTED)) {
1315 bd->flags |= BF_MARKED;
1322 // For each GC thread, for each step, allocate a "todo" block to
1323 // store evacuated objects to be scavenged, and a block to store
1324 // evacuated objects that do not need to be scavenged.
1325 for (t = 0; t < n_threads; t++) {
1326 for (s = 0; s < generations[g].n_steps; s++) {
1328 // we don't copy objects into g0s0, unless -G0
1329 if (g==0 && s==0 && RtsFlags.GcFlags.generations > 1) continue;
1331 ws = &gc_threads[t]->steps[g * RtsFlags.GcFlags.steps + s];
1333 ws->todo_large_objects = NULL;
1335 ws->part_list = NULL;
1336 ws->n_part_blocks = 0;
1338 // allocate the first to-space block; extra blocks will be
1339 // chained on as necessary.
1341 ASSERT(looksEmptyWSDeque(ws->todo_q));
1342 alloc_todo_block(ws,0);
1344 ws->todo_overflow = NULL;
1345 ws->n_todo_overflow = 0;
1347 ws->scavd_list = NULL;
1348 ws->n_scavd_blocks = 0;
1354 /* ----------------------------------------------------------------------------
1355 Initialise a generation that is *not* to be collected
1356 ------------------------------------------------------------------------- */
1359 init_uncollected_gen (nat g, nat threads)
1366 // save the current mutable lists for this generation, and
1367 // allocate a fresh block for each one. We'll traverse these
1368 // mutable lists as roots early on in the GC.
1369 generations[g].saved_mut_list = generations[g].mut_list;
1370 generations[g].mut_list = allocBlock();
1371 for (n = 0; n < n_capabilities; n++) {
1372 capabilities[n].saved_mut_lists[g] = capabilities[n].mut_lists[g];
1373 capabilities[n].mut_lists[g] = allocBlock();
1376 for (s = 0; s < generations[g].n_steps; s++) {
1377 stp = &generations[g].steps[s];
1378 stp->scavenged_large_objects = NULL;
1379 stp->n_scavenged_large_blocks = 0;
1382 for (s = 0; s < generations[g].n_steps; s++) {
1384 stp = &generations[g].steps[s];
1386 for (t = 0; t < threads; t++) {
1387 ws = &gc_threads[t]->steps[g * RtsFlags.GcFlags.steps + s];
1389 ASSERT(looksEmptyWSDeque(ws->todo_q));
1390 ws->todo_large_objects = NULL;
1392 ws->part_list = NULL;
1393 ws->n_part_blocks = 0;
1395 ws->scavd_list = NULL;
1396 ws->n_scavd_blocks = 0;
1398 // If the block at the head of the list in this generation
1399 // is less than 3/4 full, then use it as a todo block.
1400 if (stp->blocks && isPartiallyFull(stp->blocks))
1402 ws->todo_bd = stp->blocks;
1403 ws->todo_free = ws->todo_bd->free;
1404 ws->todo_lim = ws->todo_bd->start + BLOCK_SIZE_W;
1405 stp->blocks = stp->blocks->link;
1407 stp->n_words -= ws->todo_bd->free - ws->todo_bd->start;
1408 ws->todo_bd->link = NULL;
1409 // we must scan from the current end point.
1410 ws->todo_bd->u.scan = ws->todo_bd->free;
1415 alloc_todo_block(ws,0);
1419 // deal out any more partial blocks to the threads' part_lists
1421 while (stp->blocks && isPartiallyFull(stp->blocks))
1424 stp->blocks = bd->link;
1425 ws = &gc_threads[t]->steps[g * RtsFlags.GcFlags.steps + s];
1426 bd->link = ws->part_list;
1428 ws->n_part_blocks += 1;
1429 bd->u.scan = bd->free;
1431 stp->n_words -= bd->free - bd->start;
1433 if (t == n_gc_threads) t = 0;
1438 /* -----------------------------------------------------------------------------
1439 Initialise a gc_thread before GC
1440 -------------------------------------------------------------------------- */
1443 init_gc_thread (gc_thread *t)
1445 t->static_objects = END_OF_STATIC_LIST;
1446 t->scavenged_static_objects = END_OF_STATIC_LIST;
1448 t->mut_lists = capabilities[t->thread_index].mut_lists;
1450 t->failed_to_evac = rtsFalse;
1451 t->eager_promotion = rtsTrue;
1452 t->thunk_selector_depth = 0;
1457 t->scav_find_work = 0;
1460 /* -----------------------------------------------------------------------------
1461 Function we pass to evacuate roots.
1462 -------------------------------------------------------------------------- */
1465 mark_root(void *user USED_IF_THREADS, StgClosure **root)
1467 // we stole a register for gct, but this function is called from
1468 // *outside* the GC where the register variable is not in effect,
1469 // so we need to save and restore it here. NB. only call
1470 // mark_root() from the main GC thread, otherwise gct will be
1472 gc_thread *saved_gct;
1481 /* -----------------------------------------------------------------------------
1482 Initialising the static object & mutable lists
1483 -------------------------------------------------------------------------- */
1486 zero_static_object_list(StgClosure* first_static)
1490 const StgInfoTable *info;
1492 for (p = first_static; p != END_OF_STATIC_LIST; p = link) {
1494 link = *STATIC_LINK(info, p);
1495 *STATIC_LINK(info,p) = NULL;
1499 /* ----------------------------------------------------------------------------
1500 Update the pointers from the task list
1502 These are treated as weak pointers because we want to allow a main
1503 thread to get a BlockedOnDeadMVar exception in the same way as any
1504 other thread. Note that the threads should all have been retained
1505 by GC by virtue of being on the all_threads list, we're just
1506 updating pointers here.
1507 ------------------------------------------------------------------------- */
1510 update_task_list (void)
1514 for (task = all_tasks; task != NULL; task = task->all_link) {
1515 if (!task->stopped && task->tso) {
1516 ASSERT(task->tso->bound == task);
1517 tso = (StgTSO *) isAlive((StgClosure *)task->tso);
1519 barf("task %p: main thread %d has been GC'd",
1532 /* ----------------------------------------------------------------------------
1533 Reset the sizes of the older generations when we do a major
1536 CURRENT STRATEGY: make all generations except zero the same size.
1537 We have to stay within the maximum heap size, and leave a certain
1538 percentage of the maximum heap size available to allocate into.
1539 ------------------------------------------------------------------------- */
1542 resize_generations (void)
1546 if (major_gc && RtsFlags.GcFlags.generations > 1) {
1547 nat live, size, min_alloc, words;
1548 nat max = RtsFlags.GcFlags.maxHeapSize;
1549 nat gens = RtsFlags.GcFlags.generations;
1551 // live in the oldest generations
1552 if (oldest_gen->steps[0].live_estimate != 0) {
1553 words = oldest_gen->steps[0].live_estimate;
1555 words = oldest_gen->steps[0].n_words;
1557 live = (words + BLOCK_SIZE_W - 1) / BLOCK_SIZE_W +
1558 oldest_gen->steps[0].n_large_blocks;
1560 // default max size for all generations except zero
1561 size = stg_max(live * RtsFlags.GcFlags.oldGenFactor,
1562 RtsFlags.GcFlags.minOldGenSize);
1564 if (RtsFlags.GcFlags.heapSizeSuggestionAuto) {
1565 RtsFlags.GcFlags.heapSizeSuggestion = size;
1568 // minimum size for generation zero
1569 min_alloc = stg_max((RtsFlags.GcFlags.pcFreeHeap * max) / 200,
1570 RtsFlags.GcFlags.minAllocAreaSize);
1572 // Auto-enable compaction when the residency reaches a
1573 // certain percentage of the maximum heap size (default: 30%).
1574 if (RtsFlags.GcFlags.generations > 1 &&
1575 (RtsFlags.GcFlags.compact ||
1577 oldest_gen->steps[0].n_blocks >
1578 (RtsFlags.GcFlags.compactThreshold * max) / 100))) {
1579 oldest_gen->steps[0].mark = 1;
1580 oldest_gen->steps[0].compact = 1;
1581 // debugBelch("compaction: on\n", live);
1583 oldest_gen->steps[0].mark = 0;
1584 oldest_gen->steps[0].compact = 0;
1585 // debugBelch("compaction: off\n", live);
1588 if (RtsFlags.GcFlags.sweep) {
1589 oldest_gen->steps[0].mark = 1;
1592 // if we're going to go over the maximum heap size, reduce the
1593 // size of the generations accordingly. The calculation is
1594 // different if compaction is turned on, because we don't need
1595 // to double the space required to collect the old generation.
1598 // this test is necessary to ensure that the calculations
1599 // below don't have any negative results - we're working
1600 // with unsigned values here.
1601 if (max < min_alloc) {
1605 if (oldest_gen->steps[0].compact) {
1606 if ( (size + (size - 1) * (gens - 2) * 2) + min_alloc > max ) {
1607 size = (max - min_alloc) / ((gens - 1) * 2 - 1);
1610 if ( (size * (gens - 1) * 2) + min_alloc > max ) {
1611 size = (max - min_alloc) / ((gens - 1) * 2);
1621 debugBelch("live: %d, min_alloc: %d, size : %d, max = %d\n", live,
1622 min_alloc, size, max);
1625 for (g = 0; g < gens; g++) {
1626 generations[g].max_blocks = size;
1631 /* -----------------------------------------------------------------------------
1632 Calculate the new size of the nursery, and resize it.
1633 -------------------------------------------------------------------------- */
1636 resize_nursery (void)
1638 lnat min_nursery = RtsFlags.GcFlags.minAllocAreaSize * n_capabilities;
1640 if (RtsFlags.GcFlags.generations == 1)
1641 { // Two-space collector:
1644 /* set up a new nursery. Allocate a nursery size based on a
1645 * function of the amount of live data (by default a factor of 2)
1646 * Use the blocks from the old nursery if possible, freeing up any
1649 * If we get near the maximum heap size, then adjust our nursery
1650 * size accordingly. If the nursery is the same size as the live
1651 * data (L), then we need 3L bytes. We can reduce the size of the
1652 * nursery to bring the required memory down near 2L bytes.
1654 * A normal 2-space collector would need 4L bytes to give the same
1655 * performance we get from 3L bytes, reducing to the same
1656 * performance at 2L bytes.
1658 blocks = generations[0].steps[0].n_blocks;
1660 if ( RtsFlags.GcFlags.maxHeapSize != 0 &&
1661 blocks * RtsFlags.GcFlags.oldGenFactor * 2 >
1662 RtsFlags.GcFlags.maxHeapSize )
1664 long adjusted_blocks; // signed on purpose
1667 adjusted_blocks = (RtsFlags.GcFlags.maxHeapSize - 2 * blocks);
1669 debugTrace(DEBUG_gc, "near maximum heap size of 0x%x blocks, blocks = %d, adjusted to %ld",
1670 RtsFlags.GcFlags.maxHeapSize, blocks, adjusted_blocks);
1672 pc_free = adjusted_blocks * 100 / RtsFlags.GcFlags.maxHeapSize;
1673 if (pc_free < RtsFlags.GcFlags.pcFreeHeap) /* might even * be < 0 */
1677 blocks = adjusted_blocks;
1681 blocks *= RtsFlags.GcFlags.oldGenFactor;
1682 if (blocks < min_nursery)
1684 blocks = min_nursery;
1687 resizeNurseries(blocks);
1689 else // Generational collector
1692 * If the user has given us a suggested heap size, adjust our
1693 * allocation area to make best use of the memory available.
1695 if (RtsFlags.GcFlags.heapSizeSuggestion)
1698 nat needed = calcNeeded(); // approx blocks needed at next GC
1700 /* Guess how much will be live in generation 0 step 0 next time.
1701 * A good approximation is obtained by finding the
1702 * percentage of g0s0 that was live at the last minor GC.
1704 * We have an accurate figure for the amount of copied data in
1705 * 'copied', but we must convert this to a number of blocks, with
1706 * a small adjustment for estimated slop at the end of a block
1711 g0s0_pcnt_kept = ((copied / (BLOCK_SIZE_W - 10)) * 100)
1712 / countNurseryBlocks();
1715 /* Estimate a size for the allocation area based on the
1716 * information available. We might end up going slightly under
1717 * or over the suggested heap size, but we should be pretty
1720 * Formula: suggested - needed
1721 * ----------------------------
1722 * 1 + g0s0_pcnt_kept/100
1724 * where 'needed' is the amount of memory needed at the next
1725 * collection for collecting all steps except g0s0.
1728 (((long)RtsFlags.GcFlags.heapSizeSuggestion - (long)needed) * 100) /
1729 (100 + (long)g0s0_pcnt_kept);
1731 if (blocks < (long)min_nursery) {
1732 blocks = min_nursery;
1735 resizeNurseries((nat)blocks);
1739 // we might have added extra large blocks to the nursery, so
1740 // resize back to minAllocAreaSize again.
1741 resizeNurseriesFixed(RtsFlags.GcFlags.minAllocAreaSize);
1746 /* -----------------------------------------------------------------------------
1747 Sanity code for CAF garbage collection.
1749 With DEBUG turned on, we manage a CAF list in addition to the SRT
1750 mechanism. After GC, we run down the CAF list and blackhole any
1751 CAFs which have been garbage collected. This means we get an error
1752 whenever the program tries to enter a garbage collected CAF.
1754 Any garbage collected CAFs are taken off the CAF list at the same
1756 -------------------------------------------------------------------------- */
1758 #if 0 && defined(DEBUG)
1765 const StgInfoTable *info;
1776 ASSERT(info->type == IND_STATIC);
1778 if (STATIC_LINK(info,p) == NULL) {
1779 debugTrace(DEBUG_gccafs, "CAF gc'd at 0x%04lx", (long)p);
1781 SET_INFO(p,&stg_BLACKHOLE_info);
1782 p = STATIC_LINK2(info,p);
1786 pp = &STATIC_LINK2(info,p);
1793 debugTrace(DEBUG_gccafs, "%d CAFs live", i);