Allow the number of capabilities to be increased at runtime (#3729)
[ghc.git] / rts / sm / GC.c
1 /* -----------------------------------------------------------------------------
2 *
3 * (c) The GHC Team 1998-2008
4 *
5 * Generational garbage collector
6 *
7 * Documentation on the architecture of the Garbage Collector can be
8 * found in the online commentary:
9 *
10 * http://hackage.haskell.org/trac/ghc/wiki/Commentary/Rts/Storage/GC
11 *
12 * ---------------------------------------------------------------------------*/
13
14 #include "PosixSource.h"
15 #include "Rts.h"
16 #include "HsFFI.h"
17
18 #include "Storage.h"
19 #include "RtsUtils.h"
20 #include "Apply.h"
21 #include "Updates.h"
22 #include "Stats.h"
23 #include "Schedule.h"
24 #include "Sanity.h"
25 #include "BlockAlloc.h"
26 #include "ProfHeap.h"
27 #include "Weak.h"
28 #include "Prelude.h"
29 #include "RtsSignals.h"
30 #include "STM.h"
31 #if defined(RTS_GTK_FRONTPANEL)
32 #include "FrontPanel.h"
33 #endif
34 #include "Trace.h"
35 #include "RetainerProfile.h"
36 #include "LdvProfile.h"
37 #include "RaiseAsync.h"
38 #include "Papi.h"
39 #include "Stable.h"
40
41 #include "GC.h"
42 #include "GCThread.h"
43 #include "GCTDecl.h"
44 #include "Compact.h"
45 #include "Evac.h"
46 #include "Scav.h"
47 #include "GCUtils.h"
48 #include "MarkStack.h"
49 #include "MarkWeak.h"
50 #include "Sparks.h"
51 #include "Sweep.h"
52
53 #include <string.h> // for memset()
54 #include <unistd.h>
55
56 /* -----------------------------------------------------------------------------
57 Global variables
58 -------------------------------------------------------------------------- */
59
60 /* STATIC OBJECT LIST.
61 *
62 * During GC:
63 * We maintain a linked list of static objects that are still live.
64 * The requirements for this list are:
65 *
66 * - we need to scan the list while adding to it, in order to
67 * scavenge all the static objects (in the same way that
68 * breadth-first scavenging works for dynamic objects).
69 *
70 * - we need to be able to tell whether an object is already on
71 * the list, to break loops.
72 *
73 * Each static object has a "static link field", which we use for
74 * linking objects on to the list. We use a stack-type list, consing
75 * objects on the front as they are added (this means that the
76 * scavenge phase is depth-first, not breadth-first, but that
77 * shouldn't matter).
78 *
79 * A separate list is kept for objects that have been scavenged
80 * already - this is so that we can zero all the marks afterwards.
81 *
82 * An object is on the list if its static link field is non-zero; this
83 * means that we have to mark the end of the list with '1', not NULL.
84 *
85 * Extra notes for generational GC:
86 *
87 * Each generation has a static object list associated with it. When
88 * collecting generations up to N, we treat the static object lists
89 * from generations > N as roots.
90 *
91 * We build up a static object list while collecting generations 0..N,
92 * which is then appended to the static object list of generation N+1.
93 */
94
95 /* N is the oldest generation being collected, where the generations
96 * are numbered starting at 0. A major GC (indicated by the major_gc
97 * flag) is when we're collecting all generations. We only attempt to
98 * deal with static objects and GC CAFs when doing a major GC.
99 */
100 nat N;
101 rtsBool major_gc;
102
103 /* Data used for allocation area sizing.
104 */
105 static lnat g0_pcnt_kept = 30; // percentage of g0 live at last minor GC
106
107 /* Mut-list stats */
108 #ifdef DEBUG
109 nat mutlist_MUTVARS,
110 mutlist_MUTARRS,
111 mutlist_MVARS,
112 mutlist_OTHERS;
113 #endif
114
115 /* Thread-local data for each GC thread
116 */
117 gc_thread **gc_threads = NULL;
118
119 #if !defined(THREADED_RTS)
120 StgWord8 the_gc_thread[sizeof(gc_thread) + 64 * sizeof(gen_workspace)];
121 #endif
122
123 // Number of threads running in *this* GC. Affects how many
124 // step->todos[] lists we have to look in to find work.
125 nat n_gc_threads;
126
127 // For stats:
128 long copied; // *words* copied & scavenged during this GC
129
130 rtsBool work_stealing;
131
132 DECLARE_GCT
133
134 /* -----------------------------------------------------------------------------
135 Static function declarations
136 -------------------------------------------------------------------------- */
137
138 static void mark_root (void *user, StgClosure **root);
139 static void zero_static_object_list (StgClosure* first_static);
140 static nat initialise_N (rtsBool force_major_gc);
141 static void prepare_collected_gen (generation *gen);
142 static void prepare_uncollected_gen (generation *gen);
143 static void init_gc_thread (gc_thread *t);
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 me);
151 static void shutdown_gc_threads (nat me);
152 static void collect_gct_blocks (void);
153
154 #if 0 && defined(DEBUG)
155 static void gcCAFs (void);
156 #endif
157
158 /* -----------------------------------------------------------------------------
159 The mark stack.
160 -------------------------------------------------------------------------- */
161
162 bdescr *mark_stack_top_bd; // topmost block in the mark stack
163 bdescr *mark_stack_bd; // current block in the mark stack
164 StgPtr mark_sp; // pointer to the next unallocated mark stack entry
165
166 /* -----------------------------------------------------------------------------
167 GarbageCollect: the main entry point to the garbage collector.
168
169 Locks held: all capabilities are held throughout GarbageCollect().
170 -------------------------------------------------------------------------- */
171
172 void
173 GarbageCollect (rtsBool force_major_gc,
174 rtsBool do_heap_census,
175 nat gc_type USED_IF_THREADS,
176 Capability *cap)
177 {
178 bdescr *bd;
179 generation *gen;
180 lnat live_blocks, live_words, allocated, max_copied, avg_copied;
181 #if defined(THREADED_RTS)
182 gc_thread *saved_gct;
183 #endif
184 nat g, n;
185
186 // necessary if we stole a callee-saves register for gct:
187 #if defined(THREADED_RTS)
188 saved_gct = gct;
189 #endif
190
191 #ifdef PROFILING
192 CostCentreStack *save_CCS[n_capabilities];
193 #endif
194
195 ACQUIRE_SM_LOCK;
196
197 #if defined(RTS_USER_SIGNALS)
198 if (RtsFlags.MiscFlags.install_signal_handlers) {
199 // block signals
200 blockUserSignals();
201 }
202 #endif
203
204 ASSERT(sizeof(gen_workspace) == 16 * sizeof(StgWord));
205 // otherwise adjust the padding in gen_workspace.
206
207 // this is the main thread
208 SET_GCT(gc_threads[cap->no]);
209
210 // tell the stats department that we've started a GC
211 stat_startGC(gct);
212
213 // lock the StablePtr table
214 stablePtrPreGC();
215
216 #ifdef DEBUG
217 mutlist_MUTVARS = 0;
218 mutlist_MUTARRS = 0;
219 mutlist_OTHERS = 0;
220 #endif
221
222 // attribute any costs to CCS_GC
223 #ifdef PROFILING
224 for (n = 0; n < n_capabilities; n++) {
225 save_CCS[n] = capabilities[n].r.rCCCS;
226 capabilities[n].r.rCCCS = CCS_GC;
227 }
228 #endif
229
230 /* Approximate how much we allocated.
231 * Todo: only when generating stats?
232 */
233 allocated = calcAllocated(rtsFalse/* don't count the nursery yet */);
234
235 /* Figure out which generation to collect
236 */
237 n = initialise_N(force_major_gc);
238
239 #if defined(THREADED_RTS)
240 work_stealing = RtsFlags.ParFlags.parGcLoadBalancingEnabled &&
241 N >= RtsFlags.ParFlags.parGcLoadBalancingGen;
242 // It's not always a good idea to do load balancing in parallel
243 // GC. In particular, for a parallel program we don't want to
244 // lose locality by moving cached data into another CPU's cache
245 // (this effect can be quite significant).
246 //
247 // We could have a more complex way to deterimine whether to do
248 // work stealing or not, e.g. it might be a good idea to do it
249 // if the heap is big. For now, we just turn it on or off with
250 // a flag.
251 #endif
252
253 /* Start threads, so they can be spinning up while we finish initialisation.
254 */
255 start_gc_threads();
256
257 #if defined(THREADED_RTS)
258 /* How many threads will be participating in this GC?
259 * We don't try to parallelise minor GCs (unless the user asks for
260 * it with +RTS -gn0), or mark/compact/sweep GC.
261 */
262 if (gc_type == SYNC_GC_PAR) {
263 n_gc_threads = n_capabilities;
264 } else {
265 n_gc_threads = 1;
266 }
267 #else
268 n_gc_threads = 1;
269 #endif
270
271 debugTrace(DEBUG_gc, "GC (gen %d): %d KB to collect, %ld MB in use, using %d thread(s)",
272 N, n * (BLOCK_SIZE / 1024), mblocks_allocated, n_gc_threads);
273
274 #ifdef RTS_GTK_FRONTPANEL
275 if (RtsFlags.GcFlags.frontpanel) {
276 updateFrontPanelBeforeGC(N);
277 }
278 #endif
279
280 #ifdef DEBUG
281 // check for memory leaks if DEBUG is on
282 memInventory(DEBUG_gc);
283 #endif
284
285 // check sanity *before* GC
286 IF_DEBUG(sanity, checkSanity(rtsFalse /* before GC */, major_gc));
287
288 // Initialise all the generations/steps that we're collecting.
289 for (g = 0; g <= N; g++) {
290 prepare_collected_gen(&generations[g]);
291 }
292 // Initialise all the generations/steps that we're *not* collecting.
293 for (g = N+1; g < RtsFlags.GcFlags.generations; g++) {
294 prepare_uncollected_gen(&generations[g]);
295 }
296
297 // Prepare this gc_thread
298 init_gc_thread(gct);
299
300 /* Allocate a mark stack if we're doing a major collection.
301 */
302 if (major_gc && oldest_gen->mark) {
303 mark_stack_bd = allocBlock();
304 mark_stack_top_bd = mark_stack_bd;
305 mark_stack_bd->link = NULL;
306 mark_stack_bd->u.back = NULL;
307 mark_sp = mark_stack_bd->start;
308 } else {
309 mark_stack_bd = NULL;
310 mark_stack_top_bd = NULL;
311 mark_sp = NULL;
312 }
313
314 /* -----------------------------------------------------------------------
315 * follow all the roots that we know about:
316 */
317
318 // the main thread is running: this prevents any other threads from
319 // exiting prematurely, so we can start them now.
320 // NB. do this after the mutable lists have been saved above, otherwise
321 // the other GC threads will be writing into the old mutable lists.
322 inc_running();
323 wakeup_gc_threads(gct->thread_index);
324
325 traceEventGcWork(gct->cap);
326
327 // scavenge the capability-private mutable lists. This isn't part
328 // of markSomeCapabilities() because markSomeCapabilities() can only
329 // call back into the GC via mark_root() (due to the gct register
330 // variable).
331 if (n_gc_threads == 1) {
332 for (n = 0; n < n_capabilities; n++) {
333 #if defined(THREADED_RTS)
334 scavenge_capability_mut_Lists1(&capabilities[n]);
335 #else
336 scavenge_capability_mut_lists(&capabilities[n]);
337 #endif
338 }
339 } else {
340 scavenge_capability_mut_lists(gct->cap);
341 }
342
343 // follow roots from the CAF list (used by GHCi)
344 gct->evac_gen_no = 0;
345 markCAFs(mark_root, gct);
346
347 // follow all the roots that the application knows about.
348 gct->evac_gen_no = 0;
349 if (n_gc_threads == 1) {
350 for (n = 0; n < n_capabilities; n++) {
351 markCapability(mark_root, gct, &capabilities[n],
352 rtsTrue/*don't mark sparks*/);
353 }
354 } else {
355 markCapability(mark_root, gct, cap, rtsTrue/*don't mark sparks*/);
356 }
357
358 markScheduler(mark_root, gct);
359
360 #if defined(RTS_USER_SIGNALS)
361 // mark the signal handlers (signals should be already blocked)
362 markSignalHandlers(mark_root, gct);
363 #endif
364
365 // Mark the weak pointer list, and prepare to detect dead weak pointers.
366 markWeakPtrList();
367 initWeakForGC();
368
369 // Mark the stable pointer table.
370 markStablePtrTable(mark_root, gct);
371
372 /* -------------------------------------------------------------------------
373 * Repeatedly scavenge all the areas we know about until there's no
374 * more scavenging to be done.
375 */
376 for (;;)
377 {
378 scavenge_until_all_done();
379 // The other threads are now stopped. We might recurse back to
380 // here, but from now on this is the only thread.
381
382 // must be last... invariant is that everything is fully
383 // scavenged at this point.
384 if (traverseWeakPtrList()) { // returns rtsTrue if evaced something
385 inc_running();
386 continue;
387 }
388
389 // If we get to here, there's really nothing left to do.
390 break;
391 }
392
393 shutdown_gc_threads(gct->thread_index);
394
395 // Now see which stable names are still alive.
396 gcStablePtrTable();
397
398 #ifdef THREADED_RTS
399 if (n_gc_threads == 1) {
400 for (n = 0; n < n_capabilities; n++) {
401 pruneSparkQueue(&capabilities[n]);
402 }
403 } else {
404 pruneSparkQueue(gct->cap);
405 }
406 #endif
407
408 #ifdef PROFILING
409 // We call processHeapClosureForDead() on every closure destroyed during
410 // the current garbage collection, so we invoke LdvCensusForDead().
411 if (RtsFlags.ProfFlags.doHeapProfile == HEAP_BY_LDV
412 || RtsFlags.ProfFlags.bioSelector != NULL) {
413 RELEASE_SM_LOCK; // LdvCensusForDead may need to take the lock
414 LdvCensusForDead(N);
415 ACQUIRE_SM_LOCK;
416 }
417 #endif
418
419 // NO MORE EVACUATION AFTER THIS POINT!
420
421 // Finally: compact or sweep the oldest generation.
422 if (major_gc && oldest_gen->mark) {
423 if (oldest_gen->compact)
424 compact(gct->scavenged_static_objects);
425 else
426 sweep(oldest_gen);
427 }
428
429 copied = 0;
430 max_copied = 0;
431 avg_copied = 0;
432 {
433 nat i;
434 for (i=0; i < n_gc_threads; i++) {
435 if (n_gc_threads > 1) {
436 debugTrace(DEBUG_gc,"thread %d:", i);
437 debugTrace(DEBUG_gc," copied %ld", gc_threads[i]->copied * sizeof(W_));
438 debugTrace(DEBUG_gc," scanned %ld", gc_threads[i]->scanned * sizeof(W_));
439 debugTrace(DEBUG_gc," any_work %ld", gc_threads[i]->any_work);
440 debugTrace(DEBUG_gc," no_work %ld", gc_threads[i]->no_work);
441 debugTrace(DEBUG_gc," scav_find_work %ld", gc_threads[i]->scav_find_work);
442 }
443 copied += gc_threads[i]->copied;
444 max_copied = stg_max(gc_threads[i]->copied, max_copied);
445 }
446 if (n_gc_threads == 1) {
447 max_copied = 0;
448 avg_copied = 0;
449 } else {
450 avg_copied = copied;
451 }
452 }
453
454 // Run through all the generations/steps and tidy up.
455 // We're going to:
456 // - count the amount of "live" data (live_words, live_blocks)
457 // - count the amount of "copied" data in this GC (copied)
458 // - free from-space
459 // - make to-space the new from-space (set BF_EVACUATED on all blocks)
460 //
461 live_words = 0;
462 live_blocks = 0;
463
464 for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
465
466 if (g == N) {
467 generations[g].collections++; // for stats
468 if (n_gc_threads > 1) generations[g].par_collections++;
469 }
470
471 // Count the mutable list as bytes "copied" for the purposes of
472 // stats. Every mutable list is copied during every GC.
473 if (g > 0) {
474 nat mut_list_size = 0;
475 for (n = 0; n < n_capabilities; n++) {
476 mut_list_size += countOccupied(capabilities[n].mut_lists[g]);
477 }
478 copied += mut_list_size;
479
480 debugTrace(DEBUG_gc,
481 "mut_list_size: %lu (%d vars, %d arrays, %d MVARs, %d others)",
482 (unsigned long)(mut_list_size * sizeof(W_)),
483 mutlist_MUTVARS, mutlist_MUTARRS, mutlist_MVARS, mutlist_OTHERS);
484 }
485
486 bdescr *next, *prev;
487 gen = &generations[g];
488
489 // for generations we collected...
490 if (g <= N) {
491
492 /* free old memory and shift to-space into from-space for all
493 * the collected steps (except the allocation area). These
494 * freed blocks will probaby be quickly recycled.
495 */
496 if (gen->mark)
497 {
498 // tack the new blocks on the end of the existing blocks
499 if (gen->old_blocks != NULL) {
500
501 prev = NULL;
502 for (bd = gen->old_blocks; bd != NULL; bd = next) {
503
504 next = bd->link;
505
506 if (!(bd->flags & BF_MARKED))
507 {
508 if (prev == NULL) {
509 gen->old_blocks = next;
510 } else {
511 prev->link = next;
512 }
513 freeGroup(bd);
514 gen->n_old_blocks--;
515 }
516 else
517 {
518 gen->n_words += bd->free - bd->start;
519
520 // NB. this step might not be compacted next
521 // time, so reset the BF_MARKED flags.
522 // They are set before GC if we're going to
523 // compact. (search for BF_MARKED above).
524 bd->flags &= ~BF_MARKED;
525
526 // between GCs, all blocks in the heap except
527 // for the nursery have the BF_EVACUATED flag set.
528 bd->flags |= BF_EVACUATED;
529
530 prev = bd;
531 }
532 }
533
534 if (prev != NULL) {
535 prev->link = gen->blocks;
536 gen->blocks = gen->old_blocks;
537 }
538 }
539 // add the new blocks to the block tally
540 gen->n_blocks += gen->n_old_blocks;
541 ASSERT(countBlocks(gen->blocks) == gen->n_blocks);
542 ASSERT(countOccupied(gen->blocks) == gen->n_words);
543 }
544 else // not copacted
545 {
546 freeChain(gen->old_blocks);
547 }
548
549 gen->old_blocks = NULL;
550 gen->n_old_blocks = 0;
551
552 /* LARGE OBJECTS. The current live large objects are chained on
553 * scavenged_large, having been moved during garbage
554 * collection from large_objects. Any objects left on the
555 * large_objects list are therefore dead, so we free them here.
556 */
557 freeChain(gen->large_objects);
558 gen->large_objects = gen->scavenged_large_objects;
559 gen->n_large_blocks = gen->n_scavenged_large_blocks;
560 gen->n_new_large_words = 0;
561 }
562 else // for generations > N
563 {
564 /* For older generations, we need to append the
565 * scavenged_large_object list (i.e. large objects that have been
566 * promoted during this GC) to the large_object list for that step.
567 */
568 for (bd = gen->scavenged_large_objects; bd; bd = next) {
569 next = bd->link;
570 dbl_link_onto(bd, &gen->large_objects);
571 }
572
573 // add the new blocks we promoted during this GC
574 gen->n_large_blocks += gen->n_scavenged_large_blocks;
575 }
576
577 ASSERT(countBlocks(gen->large_objects) == gen->n_large_blocks);
578
579 gen->scavenged_large_objects = NULL;
580 gen->n_scavenged_large_blocks = 0;
581
582 // Count "live" data
583 live_words += genLiveWords(gen);
584 live_blocks += genLiveBlocks(gen);
585
586 // add in the partial blocks in the gen_workspaces, but ignore gen 0
587 // if this is a local GC (we can't count another capability's part_list)
588 {
589 nat i;
590 for (i = 0; i < n_capabilities; i++) {
591 live_words += gcThreadLiveWords(i, gen->no);
592 live_blocks += gcThreadLiveBlocks(i, gen->no);
593 }
594 }
595 } // for all generations
596
597 // update the max size of older generations after a major GC
598 resize_generations();
599
600 // Free the mark stack.
601 if (mark_stack_top_bd != NULL) {
602 debugTrace(DEBUG_gc, "mark stack: %d blocks",
603 countBlocks(mark_stack_top_bd));
604 freeChain(mark_stack_top_bd);
605 }
606
607 // Free any bitmaps.
608 for (g = 0; g <= N; g++) {
609 gen = &generations[g];
610 if (gen->bitmap != NULL) {
611 freeGroup(gen->bitmap);
612 gen->bitmap = NULL;
613 }
614 }
615
616 // Reset the nursery: make the blocks empty
617 allocated += clearNurseries();
618
619 resize_nursery();
620
621 resetNurseries();
622
623 // mark the garbage collected CAFs as dead
624 #if 0 && defined(DEBUG) // doesn't work at the moment
625 if (major_gc) { gcCAFs(); }
626 #endif
627
628 #ifdef PROFILING
629 // resetStaticObjectForRetainerProfiling() must be called before
630 // zeroing below.
631
632 // ToDo: fix the gct->scavenged_static_objects below
633 resetStaticObjectForRetainerProfiling(gct->scavenged_static_objects);
634 #endif
635
636 // zero the scavenged static object list
637 if (major_gc) {
638 nat i;
639 if (n_gc_threads == 1) {
640 zero_static_object_list(gct->scavenged_static_objects);
641 } else {
642 for (i = 0; i < n_gc_threads; i++) {
643 zero_static_object_list(gc_threads[i]->scavenged_static_objects);
644 }
645 }
646 }
647
648 // Update the stable pointer hash table.
649 updateStablePtrTable(major_gc);
650
651 // unlock the StablePtr table. Must be before scheduleFinalizers(),
652 // because a finalizer may call hs_free_fun_ptr() or
653 // hs_free_stable_ptr(), both of which access the StablePtr table.
654 stablePtrPostGC();
655
656 // Start any pending finalizers. Must be after
657 // updateStablePtrTable() and stablePtrPostGC() (see #4221).
658 RELEASE_SM_LOCK;
659 scheduleFinalizers(cap, old_weak_ptr_list);
660 ACQUIRE_SM_LOCK;
661
662 // check sanity after GC
663 // before resurrectThreads(), because that might overwrite some
664 // closures, which will cause problems with THREADED where we don't
665 // fill slop.
666 IF_DEBUG(sanity, checkSanity(rtsTrue /* after GC */, major_gc));
667
668 // If a heap census is due, we need to do it before
669 // resurrectThreads(), for the same reason as checkSanity above:
670 // resurrectThreads() will overwrite some closures and leave slop
671 // behind.
672 if (do_heap_census) {
673 debugTrace(DEBUG_sched, "performing heap census");
674 RELEASE_SM_LOCK;
675 heapCensus(gct->gc_start_cpu);
676 ACQUIRE_SM_LOCK;
677 }
678
679 // send exceptions to any threads which were about to die
680 RELEASE_SM_LOCK;
681 resurrectThreads(resurrected_threads);
682 ACQUIRE_SM_LOCK;
683
684 if (major_gc) {
685 nat need, got;
686 need = BLOCKS_TO_MBLOCKS(n_alloc_blocks);
687 got = mblocks_allocated;
688 /* If the amount of data remains constant, next major GC we'll
689 require (F+1)*need. We leave (F+2)*need in order to reduce
690 repeated deallocation and reallocation. */
691 need = (RtsFlags.GcFlags.oldGenFactor + 2) * need;
692 if (got > need) {
693 returnMemoryToOS(got - need);
694 }
695 }
696
697 // extra GC trace info
698 IF_DEBUG(gc, statDescribeGens());
699
700 #ifdef DEBUG
701 // symbol-table based profiling
702 /* heapCensus(to_blocks); */ /* ToDo */
703 #endif
704
705 // restore enclosing cost centre
706 #ifdef PROFILING
707 for (n = 0; n < n_capabilities; n++) {
708 capabilities[n].r.rCCCS = save_CCS[n];
709 }
710 #endif
711
712 #ifdef DEBUG
713 // check for memory leaks if DEBUG is on
714 memInventory(DEBUG_gc);
715 #endif
716
717 #ifdef RTS_GTK_FRONTPANEL
718 if (RtsFlags.GcFlags.frontpanel) {
719 updateFrontPanelAfterGC( N, live );
720 }
721 #endif
722
723 // ok, GC over: tell the stats department what happened.
724 stat_endGC(gct, allocated, live_words,
725 copied, N, max_copied, avg_copied,
726 live_blocks * BLOCK_SIZE_W - live_words /* slop */);
727
728 // Guess which generation we'll collect *next* time
729 initialise_N(force_major_gc);
730
731 #if defined(RTS_USER_SIGNALS)
732 if (RtsFlags.MiscFlags.install_signal_handlers) {
733 // unblock signals again
734 unblockUserSignals();
735 }
736 #endif
737
738 RELEASE_SM_LOCK;
739
740 SET_GCT(saved_gct);
741 }
742
743 /* -----------------------------------------------------------------------------
744 Figure out which generation to collect, initialise N and major_gc.
745
746 Also returns the total number of blocks in generations that will be
747 collected.
748 -------------------------------------------------------------------------- */
749
750 static nat
751 initialise_N (rtsBool force_major_gc)
752 {
753 int g;
754 nat blocks, blocks_total;
755
756 blocks = 0;
757 blocks_total = 0;
758
759 if (force_major_gc) {
760 N = RtsFlags.GcFlags.generations - 1;
761 } else {
762 N = 0;
763 }
764
765 for (g = RtsFlags.GcFlags.generations - 1; g >= 0; g--) {
766
767 blocks = generations[g].n_words / BLOCK_SIZE_W
768 + generations[g].n_large_blocks;
769
770 if (blocks >= generations[g].max_blocks) {
771 N = stg_max(N,g);
772 }
773 if ((nat)g <= N) {
774 blocks_total += blocks;
775 }
776 }
777
778 blocks_total += countNurseryBlocks();
779
780 major_gc = (N == RtsFlags.GcFlags.generations-1);
781 return blocks_total;
782 }
783
784 /* -----------------------------------------------------------------------------
785 Initialise the gc_thread structures.
786 -------------------------------------------------------------------------- */
787
788 #define GC_THREAD_INACTIVE 0
789 #define GC_THREAD_STANDING_BY 1
790 #define GC_THREAD_RUNNING 2
791 #define GC_THREAD_WAITING_TO_CONTINUE 3
792
793 static void
794 new_gc_thread (nat n, gc_thread *t)
795 {
796 nat g;
797 gen_workspace *ws;
798
799 t->cap = &capabilities[n];
800
801 #ifdef THREADED_RTS
802 t->id = 0;
803 initSpinLock(&t->gc_spin);
804 initSpinLock(&t->mut_spin);
805 ACQUIRE_SPIN_LOCK(&t->gc_spin);
806 t->wakeup = GC_THREAD_INACTIVE; // starts true, so we can wait for the
807 // thread to start up, see wakeup_gc_threads
808 #endif
809
810 t->thread_index = n;
811 t->free_blocks = NULL;
812 t->gc_count = 0;
813
814 init_gc_thread(t);
815
816 #ifdef USE_PAPI
817 t->papi_events = -1;
818 #endif
819
820 for (g = 0; g < RtsFlags.GcFlags.generations; g++)
821 {
822 ws = &t->gens[g];
823 ws->gen = &generations[g];
824 ASSERT(g == ws->gen->no);
825 ws->my_gct = t;
826
827 // We want to call
828 // alloc_todo_block(ws,0);
829 // but can't, because it uses gct which isn't set up at this point.
830 // Hence, allocate a block for todo_bd manually:
831 {
832 bdescr *bd = allocBlock(); // no lock, locks aren't initialised yet
833 initBdescr(bd, ws->gen, ws->gen->to);
834 bd->flags = BF_EVACUATED;
835 bd->u.scan = bd->free = bd->start;
836
837 ws->todo_bd = bd;
838 ws->todo_free = bd->free;
839 ws->todo_lim = bd->start + BLOCK_SIZE_W;
840 }
841
842 ws->todo_q = newWSDeque(128);
843 ws->todo_overflow = NULL;
844 ws->n_todo_overflow = 0;
845 ws->todo_large_objects = NULL;
846
847 ws->part_list = NULL;
848 ws->n_part_blocks = 0;
849
850 ws->scavd_list = NULL;
851 ws->n_scavd_blocks = 0;
852 }
853 }
854
855
856 void
857 initGcThreads (nat from USED_IF_THREADS, nat to USED_IF_THREADS)
858 {
859 #if defined(THREADED_RTS)
860 nat i;
861
862 if (from > 0) {
863 gc_threads = stgReallocBytes (gc_threads, to * sizeof(gc_thread*),
864 "initGcThreads");
865 } else {
866 gc_threads = stgMallocBytes (to * sizeof(gc_thread*),
867 "initGcThreads");
868 }
869
870 // We have to update the gct->cap pointers to point to the new
871 // Capability array now.
872 for (i = 0; i < from; i++) {
873 gc_threads[i]->cap = &capabilities[gc_threads[i]->cap->no];
874 }
875
876 for (i = from; i < to; i++) {
877 gc_threads[i] =
878 stgMallocBytes(sizeof(gc_thread) +
879 RtsFlags.GcFlags.generations * sizeof(gen_workspace),
880 "alloc_gc_threads");
881
882 new_gc_thread(i, gc_threads[i]);
883 }
884 #else
885 ASSERT(from == 0 && to == 1);
886 gc_threads = stgMallocBytes (sizeof(gc_thread*),"alloc_gc_threads");
887 gc_threads[0] = gct;
888 new_gc_thread(0,gc_threads[0]);
889 #endif
890 }
891
892 void
893 freeGcThreads (void)
894 {
895 nat g;
896 if (gc_threads != NULL) {
897 #if defined(THREADED_RTS)
898 nat i;
899 for (i = 0; i < n_capabilities; i++) {
900 for (g = 0; g < RtsFlags.GcFlags.generations; g++)
901 {
902 freeWSDeque(gc_threads[i]->gens[g].todo_q);
903 }
904 stgFree (gc_threads[i]);
905 }
906 stgFree (gc_threads);
907 #else
908 for (g = 0; g < RtsFlags.GcFlags.generations; g++)
909 {
910 freeWSDeque(gc_threads[0]->gens[g].todo_q);
911 }
912 stgFree (gc_threads);
913 #endif
914 gc_threads = NULL;
915 }
916 }
917
918 /* ----------------------------------------------------------------------------
919 Start GC threads
920 ------------------------------------------------------------------------- */
921
922 static volatile StgWord gc_running_threads;
923
924 static StgWord
925 inc_running (void)
926 {
927 StgWord new;
928 new = atomic_inc(&gc_running_threads);
929 ASSERT(new <= n_gc_threads);
930 return new;
931 }
932
933 static StgWord
934 dec_running (void)
935 {
936 ASSERT(gc_running_threads != 0);
937 return atomic_dec(&gc_running_threads);
938 }
939
940 static rtsBool
941 any_work (void)
942 {
943 int g;
944 gen_workspace *ws;
945
946 gct->any_work++;
947
948 write_barrier();
949
950 // scavenge objects in compacted generation
951 if (mark_stack_bd != NULL && !mark_stack_empty()) {
952 return rtsTrue;
953 }
954
955 // Check for global work in any step. We don't need to check for
956 // local work, because we have already exited scavenge_loop(),
957 // which means there is no local work for this thread.
958 for (g = 0; g < (int)RtsFlags.GcFlags.generations; g++) {
959 ws = &gct->gens[g];
960 if (ws->todo_large_objects) return rtsTrue;
961 if (!looksEmptyWSDeque(ws->todo_q)) return rtsTrue;
962 if (ws->todo_overflow) return rtsTrue;
963 }
964
965 #if defined(THREADED_RTS)
966 if (work_stealing) {
967 nat n;
968 // look for work to steal
969 for (n = 0; n < n_gc_threads; n++) {
970 if (n == gct->thread_index) continue;
971 for (g = RtsFlags.GcFlags.generations-1; g >= 0; g--) {
972 ws = &gc_threads[n]->gens[g];
973 if (!looksEmptyWSDeque(ws->todo_q)) return rtsTrue;
974 }
975 }
976 }
977 #endif
978
979 gct->no_work++;
980 #if defined(THREADED_RTS)
981 yieldThread();
982 #endif
983
984 return rtsFalse;
985 }
986
987 static void
988 scavenge_until_all_done (void)
989 {
990 DEBUG_ONLY( nat r );
991
992
993 loop:
994 #if defined(THREADED_RTS)
995 if (n_gc_threads > 1) {
996 scavenge_loop();
997 } else {
998 scavenge_loop1();
999 }
1000 #else
1001 scavenge_loop();
1002 #endif
1003
1004 collect_gct_blocks();
1005
1006 // scavenge_loop() only exits when there's no work to do
1007
1008 #ifdef DEBUG
1009 r = dec_running();
1010 #else
1011 dec_running();
1012 #endif
1013
1014 traceEventGcIdle(gct->cap);
1015
1016 debugTrace(DEBUG_gc, "%d GC threads still running", r);
1017
1018 while (gc_running_threads != 0) {
1019 // usleep(1);
1020 if (any_work()) {
1021 inc_running();
1022 traceEventGcWork(gct->cap);
1023 goto loop;
1024 }
1025 // any_work() does not remove the work from the queue, it
1026 // just checks for the presence of work. If we find any,
1027 // then we increment gc_running_threads and go back to
1028 // scavenge_loop() to perform any pending work.
1029 }
1030
1031 traceEventGcDone(gct->cap);
1032 }
1033
1034 #if defined(THREADED_RTS)
1035
1036 void
1037 gcWorkerThread (Capability *cap)
1038 {
1039 gc_thread *saved_gct;
1040
1041 // necessary if we stole a callee-saves register for gct:
1042 saved_gct = gct;
1043
1044 SET_GCT(gc_threads[cap->no]);
1045 gct->id = osThreadId();
1046
1047 stat_gcWorkerThreadStart(gct);
1048
1049 // Wait until we're told to wake up
1050 RELEASE_SPIN_LOCK(&gct->mut_spin);
1051 gct->wakeup = GC_THREAD_STANDING_BY;
1052 debugTrace(DEBUG_gc, "GC thread %d standing by...", gct->thread_index);
1053 ACQUIRE_SPIN_LOCK(&gct->gc_spin);
1054
1055 #ifdef USE_PAPI
1056 // start performance counters in this thread...
1057 if (gct->papi_events == -1) {
1058 papi_init_eventset(&gct->papi_events);
1059 }
1060 papi_thread_start_gc1_count(gct->papi_events);
1061 #endif
1062
1063 init_gc_thread(gct);
1064
1065 traceEventGcWork(gct->cap);
1066
1067 // Every thread evacuates some roots.
1068 gct->evac_gen_no = 0;
1069 markCapability(mark_root, gct, cap, rtsTrue/*prune sparks*/);
1070 scavenge_capability_mut_lists(cap);
1071
1072 scavenge_until_all_done();
1073
1074 #ifdef THREADED_RTS
1075 // Now that the whole heap is marked, we discard any sparks that
1076 // were found to be unreachable. The main GC thread is currently
1077 // marking heap reachable via weak pointers, so it is
1078 // non-deterministic whether a spark will be retained if it is
1079 // only reachable via weak pointers. To fix this problem would
1080 // require another GC barrier, which is too high a price.
1081 pruneSparkQueue(cap);
1082 #endif
1083
1084 #ifdef USE_PAPI
1085 // count events in this thread towards the GC totals
1086 papi_thread_stop_gc1_count(gct->papi_events);
1087 #endif
1088
1089 // Wait until we're told to continue
1090 RELEASE_SPIN_LOCK(&gct->gc_spin);
1091 gct->wakeup = GC_THREAD_WAITING_TO_CONTINUE;
1092 debugTrace(DEBUG_gc, "GC thread %d waiting to continue...",
1093 gct->thread_index);
1094 ACQUIRE_SPIN_LOCK(&gct->mut_spin);
1095 debugTrace(DEBUG_gc, "GC thread %d on my way...", gct->thread_index);
1096
1097 // record the time spent doing GC in the Task structure
1098 stat_gcWorkerThreadDone(gct);
1099
1100 SET_GCT(saved_gct);
1101 }
1102
1103 #endif
1104
1105 #if defined(THREADED_RTS)
1106
1107 void
1108 waitForGcThreads (Capability *cap USED_IF_THREADS)
1109 {
1110 const nat n_threads = n_capabilities;
1111 const nat me = cap->no;
1112 nat i, j;
1113 rtsBool retry = rtsTrue;
1114
1115 while(retry) {
1116 for (i=0; i < n_threads; i++) {
1117 if (i == me) continue;
1118 if (gc_threads[i]->wakeup != GC_THREAD_STANDING_BY) {
1119 prodCapability(&capabilities[i], cap->running_task);
1120 }
1121 }
1122 for (j=0; j < 10; j++) {
1123 retry = rtsFalse;
1124 for (i=0; i < n_threads; i++) {
1125 if (i == me) continue;
1126 write_barrier();
1127 interruptAllCapabilities();
1128 if (gc_threads[i]->wakeup != GC_THREAD_STANDING_BY) {
1129 retry = rtsTrue;
1130 }
1131 }
1132 if (!retry) break;
1133 yieldThread();
1134 }
1135 }
1136 }
1137
1138 #endif // THREADED_RTS
1139
1140 static void
1141 start_gc_threads (void)
1142 {
1143 #if defined(THREADED_RTS)
1144 gc_running_threads = 0;
1145 #endif
1146 }
1147
1148 static void
1149 wakeup_gc_threads (nat me USED_IF_THREADS)
1150 {
1151 #if defined(THREADED_RTS)
1152 nat i;
1153
1154 if (n_gc_threads == 1) return;
1155
1156 for (i=0; i < n_gc_threads; i++) {
1157 if (i == me) continue;
1158 inc_running();
1159 debugTrace(DEBUG_gc, "waking up gc thread %d", i);
1160 if (gc_threads[i]->wakeup != GC_THREAD_STANDING_BY) barf("wakeup_gc_threads");
1161
1162 gc_threads[i]->wakeup = GC_THREAD_RUNNING;
1163 ACQUIRE_SPIN_LOCK(&gc_threads[i]->mut_spin);
1164 RELEASE_SPIN_LOCK(&gc_threads[i]->gc_spin);
1165 }
1166 #endif
1167 }
1168
1169 // After GC is complete, we must wait for all GC threads to enter the
1170 // standby state, otherwise they may still be executing inside
1171 // any_work(), and may even remain awake until the next GC starts.
1172 static void
1173 shutdown_gc_threads (nat me USED_IF_THREADS)
1174 {
1175 #if defined(THREADED_RTS)
1176 nat i;
1177
1178 if (n_gc_threads == 1) return;
1179
1180 for (i=0; i < n_gc_threads; i++) {
1181 if (i == me) continue;
1182 while (gc_threads[i]->wakeup != GC_THREAD_WAITING_TO_CONTINUE) { write_barrier(); }
1183 }
1184 #endif
1185 }
1186
1187 #if defined(THREADED_RTS)
1188 void
1189 releaseGCThreads (Capability *cap USED_IF_THREADS)
1190 {
1191 const nat n_threads = n_capabilities;
1192 const nat me = cap->no;
1193 nat i;
1194 for (i=0; i < n_threads; i++) {
1195 if (i == me) continue;
1196 if (gc_threads[i]->wakeup != GC_THREAD_WAITING_TO_CONTINUE)
1197 barf("releaseGCThreads");
1198
1199 gc_threads[i]->wakeup = GC_THREAD_INACTIVE;
1200 ACQUIRE_SPIN_LOCK(&gc_threads[i]->gc_spin);
1201 RELEASE_SPIN_LOCK(&gc_threads[i]->mut_spin);
1202 }
1203 }
1204 #endif
1205
1206 /* ----------------------------------------------------------------------------
1207 Initialise a generation that is to be collected
1208 ------------------------------------------------------------------------- */
1209
1210 static void
1211 prepare_collected_gen (generation *gen)
1212 {
1213 nat i, g, n;
1214 gen_workspace *ws;
1215 bdescr *bd, *next;
1216
1217 // Throw away the current mutable list. Invariant: the mutable
1218 // list always has at least one block; this means we can avoid a
1219 // check for NULL in recordMutable().
1220 g = gen->no;
1221 if (g != 0) {
1222 for (i = 0; i < n_capabilities; i++) {
1223 freeChain(capabilities[i].mut_lists[g]);
1224 capabilities[i].mut_lists[g] = allocBlock();
1225 }
1226 }
1227
1228 gen = &generations[g];
1229 ASSERT(gen->no == g);
1230
1231 // we'll construct a new list of threads in this step
1232 // during GC, throw away the current list.
1233 gen->old_threads = gen->threads;
1234 gen->threads = END_TSO_QUEUE;
1235
1236 // deprecate the existing blocks
1237 gen->old_blocks = gen->blocks;
1238 gen->n_old_blocks = gen->n_blocks;
1239 gen->blocks = NULL;
1240 gen->n_blocks = 0;
1241 gen->n_words = 0;
1242 gen->live_estimate = 0;
1243
1244 // initialise the large object queues.
1245 ASSERT(gen->scavenged_large_objects == NULL);
1246 ASSERT(gen->n_scavenged_large_blocks == 0);
1247
1248 // grab all the partial blocks stashed in the gc_thread workspaces and
1249 // move them to the old_blocks list of this gen.
1250 for (n = 0; n < n_capabilities; n++) {
1251 ws = &gc_threads[n]->gens[gen->no];
1252
1253 for (bd = ws->part_list; bd != NULL; bd = next) {
1254 next = bd->link;
1255 bd->link = gen->old_blocks;
1256 gen->old_blocks = bd;
1257 gen->n_old_blocks += bd->blocks;
1258 }
1259 ws->part_list = NULL;
1260 ws->n_part_blocks = 0;
1261
1262 ASSERT(ws->scavd_list == NULL);
1263 ASSERT(ws->n_scavd_blocks == 0);
1264
1265 if (ws->todo_free != ws->todo_bd->start) {
1266 ws->todo_bd->free = ws->todo_free;
1267 ws->todo_bd->link = gen->old_blocks;
1268 gen->old_blocks = ws->todo_bd;
1269 gen->n_old_blocks += ws->todo_bd->blocks;
1270 alloc_todo_block(ws,0); // always has one block.
1271 }
1272 }
1273
1274 // mark the small objects as from-space
1275 for (bd = gen->old_blocks; bd; bd = bd->link) {
1276 bd->flags &= ~BF_EVACUATED;
1277 }
1278
1279 // mark the large objects as from-space
1280 for (bd = gen->large_objects; bd; bd = bd->link) {
1281 bd->flags &= ~BF_EVACUATED;
1282 }
1283
1284 // for a compacted generation, we need to allocate the bitmap
1285 if (gen->mark) {
1286 lnat bitmap_size; // in bytes
1287 bdescr *bitmap_bdescr;
1288 StgWord *bitmap;
1289
1290 bitmap_size = gen->n_old_blocks * BLOCK_SIZE / (sizeof(W_)*BITS_PER_BYTE);
1291
1292 if (bitmap_size > 0) {
1293 bitmap_bdescr = allocGroup((lnat)BLOCK_ROUND_UP(bitmap_size)
1294 / BLOCK_SIZE);
1295 gen->bitmap = bitmap_bdescr;
1296 bitmap = bitmap_bdescr->start;
1297
1298 debugTrace(DEBUG_gc, "bitmap_size: %d, bitmap: %p",
1299 bitmap_size, bitmap);
1300
1301 // don't forget to fill it with zeros!
1302 memset(bitmap, 0, bitmap_size);
1303
1304 // For each block in this step, point to its bitmap from the
1305 // block descriptor.
1306 for (bd=gen->old_blocks; bd != NULL; bd = bd->link) {
1307 bd->u.bitmap = bitmap;
1308 bitmap += BLOCK_SIZE_W / (sizeof(W_)*BITS_PER_BYTE);
1309
1310 // Also at this point we set the BF_MARKED flag
1311 // for this block. The invariant is that
1312 // BF_MARKED is always unset, except during GC
1313 // when it is set on those blocks which will be
1314 // compacted.
1315 if (!(bd->flags & BF_FRAGMENTED)) {
1316 bd->flags |= BF_MARKED;
1317 }
1318
1319 // BF_SWEPT should be marked only for blocks that are being
1320 // collected in sweep()
1321 bd->flags &= ~BF_SWEPT;
1322 }
1323 }
1324 }
1325 }
1326
1327
1328 /* ----------------------------------------------------------------------------
1329 Save the mutable lists in saved_mut_lists
1330 ------------------------------------------------------------------------- */
1331
1332 static void
1333 stash_mut_list (Capability *cap, nat gen_no)
1334 {
1335 cap->saved_mut_lists[gen_no] = cap->mut_lists[gen_no];
1336 cap->mut_lists[gen_no] = allocBlock_sync();
1337 }
1338
1339 /* ----------------------------------------------------------------------------
1340 Initialise a generation that is *not* to be collected
1341 ------------------------------------------------------------------------- */
1342
1343 static void
1344 prepare_uncollected_gen (generation *gen)
1345 {
1346 nat i;
1347
1348
1349 ASSERT(gen->no > 0);
1350
1351 // save the current mutable lists for this generation, and
1352 // allocate a fresh block for each one. We'll traverse these
1353 // mutable lists as roots early on in the GC.
1354 for (i = 0; i < n_capabilities; i++) {
1355 stash_mut_list(&capabilities[i], gen->no);
1356 }
1357
1358 ASSERT(gen->scavenged_large_objects == NULL);
1359 ASSERT(gen->n_scavenged_large_blocks == 0);
1360 }
1361
1362 /* -----------------------------------------------------------------------------
1363 Collect the completed blocks from a GC thread and attach them to
1364 the generation.
1365 -------------------------------------------------------------------------- */
1366
1367 static void
1368 collect_gct_blocks (void)
1369 {
1370 nat g;
1371 gen_workspace *ws;
1372 bdescr *bd, *prev;
1373
1374 for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
1375 ws = &gct->gens[g];
1376
1377 // there may still be a block attached to ws->todo_bd;
1378 // leave it there to use next time.
1379
1380 if (ws->scavd_list != NULL) {
1381 ACQUIRE_SPIN_LOCK(&ws->gen->sync);
1382
1383 ASSERT(gct->scan_bd == NULL);
1384 ASSERT(countBlocks(ws->scavd_list) == ws->n_scavd_blocks);
1385
1386 prev = NULL;
1387 for (bd = ws->scavd_list; bd != NULL; bd = bd->link) {
1388 ws->gen->n_words += bd->free - bd->start;
1389 prev = bd;
1390 }
1391 if (prev != NULL) {
1392 prev->link = ws->gen->blocks;
1393 ws->gen->blocks = ws->scavd_list;
1394 }
1395 ws->gen->n_blocks += ws->n_scavd_blocks;
1396
1397 ws->scavd_list = NULL;
1398 ws->n_scavd_blocks = 0;
1399
1400 RELEASE_SPIN_LOCK(&ws->gen->sync);
1401 }
1402 }
1403 }
1404
1405 /* -----------------------------------------------------------------------------
1406 Initialise a gc_thread before GC
1407 -------------------------------------------------------------------------- */
1408
1409 static void
1410 init_gc_thread (gc_thread *t)
1411 {
1412 t->static_objects = END_OF_STATIC_LIST;
1413 t->scavenged_static_objects = END_OF_STATIC_LIST;
1414 t->scan_bd = NULL;
1415 t->mut_lists = t->cap->mut_lists;
1416 t->evac_gen_no = 0;
1417 t->failed_to_evac = rtsFalse;
1418 t->eager_promotion = rtsTrue;
1419 t->thunk_selector_depth = 0;
1420 t->copied = 0;
1421 t->scanned = 0;
1422 t->any_work = 0;
1423 t->no_work = 0;
1424 t->scav_find_work = 0;
1425 }
1426
1427 /* -----------------------------------------------------------------------------
1428 Function we pass to evacuate roots.
1429 -------------------------------------------------------------------------- */
1430
1431 static void
1432 mark_root(void *user USED_IF_THREADS, StgClosure **root)
1433 {
1434 // we stole a register for gct, but this function is called from
1435 // *outside* the GC where the register variable is not in effect,
1436 // so we need to save and restore it here. NB. only call
1437 // mark_root() from the main GC thread, otherwise gct will be
1438 // incorrect.
1439 #if defined(THREADED_RTS)
1440 gc_thread *saved_gct;
1441 saved_gct = gct;
1442 #endif
1443 SET_GCT(user);
1444
1445 evacuate(root);
1446
1447 SET_GCT(saved_gct);
1448 }
1449
1450 /* -----------------------------------------------------------------------------
1451 Initialising the static object & mutable lists
1452 -------------------------------------------------------------------------- */
1453
1454 static void
1455 zero_static_object_list(StgClosure* first_static)
1456 {
1457 StgClosure* p;
1458 StgClosure* link;
1459 const StgInfoTable *info;
1460
1461 for (p = first_static; p != END_OF_STATIC_LIST; p = link) {
1462 info = get_itbl(p);
1463 link = *STATIC_LINK(info, p);
1464 *STATIC_LINK(info,p) = NULL;
1465 }
1466 }
1467
1468 /* ----------------------------------------------------------------------------
1469 Reset the sizes of the older generations when we do a major
1470 collection.
1471
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 ------------------------------------------------------------------------- */
1476
1477 static void
1478 resize_generations (void)
1479 {
1480 nat g;
1481
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;
1486
1487 // live in the oldest generations
1488 if (oldest_gen->live_estimate != 0) {
1489 words = oldest_gen->live_estimate;
1490 } else {
1491 words = oldest_gen->n_words;
1492 }
1493 live = (words + BLOCK_SIZE_W - 1) / BLOCK_SIZE_W +
1494 oldest_gen->n_large_blocks;
1495
1496 // default max size for all generations except zero
1497 size = stg_max(live * RtsFlags.GcFlags.oldGenFactor,
1498 RtsFlags.GcFlags.minOldGenSize);
1499
1500 if (RtsFlags.GcFlags.heapSizeSuggestionAuto) {
1501 RtsFlags.GcFlags.heapSizeSuggestion = size;
1502 }
1503
1504 // minimum size for generation zero
1505 min_alloc = stg_max((RtsFlags.GcFlags.pcFreeHeap * max) / 200,
1506 RtsFlags.GcFlags.minAllocAreaSize);
1507
1508 // Auto-enable compaction when the residency reaches a
1509 // certain percentage of the maximum heap size (default: 30%).
1510 if (RtsFlags.GcFlags.compact ||
1511 (max > 0 &&
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);
1517 } else {
1518 oldest_gen->mark = 0;
1519 oldest_gen->compact = 0;
1520 // debugBelch("compaction: off\n", live);
1521 }
1522
1523 if (RtsFlags.GcFlags.sweep) {
1524 oldest_gen->mark = 1;
1525 }
1526
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.
1531 if (max != 0) {
1532
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) {
1537 heapOverflow();
1538 }
1539
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);
1543 }
1544 } else {
1545 if ( (size * (gens - 1) * 2) + min_alloc > max ) {
1546 size = (max - min_alloc) / ((gens - 1) * 2);
1547 }
1548 }
1549
1550 if (size < live) {
1551 heapOverflow();
1552 }
1553 }
1554
1555 #if 0
1556 debugBelch("live: %d, min_alloc: %d, size : %d, max = %d\n", live,
1557 min_alloc, size, max);
1558 #endif
1559
1560 for (g = 0; g < gens; g++) {
1561 generations[g].max_blocks = size;
1562 }
1563 }
1564 }
1565
1566 /* -----------------------------------------------------------------------------
1567 Calculate the new size of the nursery, and resize it.
1568 -------------------------------------------------------------------------- */
1569
1570 static void
1571 resize_nursery (void)
1572 {
1573 const lnat min_nursery = RtsFlags.GcFlags.minAllocAreaSize * n_capabilities;
1574
1575 if (RtsFlags.GcFlags.generations == 1)
1576 { // Two-space collector:
1577 nat blocks;
1578
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
1582 * left over blocks.
1583 *
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.
1588 *
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.
1592 */
1593 blocks = generations[0].n_blocks;
1594
1595 if ( RtsFlags.GcFlags.maxHeapSize != 0 &&
1596 blocks * RtsFlags.GcFlags.oldGenFactor * 2 >
1597 RtsFlags.GcFlags.maxHeapSize )
1598 {
1599 long adjusted_blocks; // signed on purpose
1600 int pc_free;
1601
1602 adjusted_blocks = (RtsFlags.GcFlags.maxHeapSize - 2 * blocks);
1603
1604 debugTrace(DEBUG_gc, "near maximum heap size of 0x%x blocks, blocks = %d, adjusted to %ld",
1605 RtsFlags.GcFlags.maxHeapSize, blocks, adjusted_blocks);
1606
1607 pc_free = adjusted_blocks * 100 / RtsFlags.GcFlags.maxHeapSize;
1608 if (pc_free < RtsFlags.GcFlags.pcFreeHeap) /* might even * be < 0 */
1609 {
1610 heapOverflow();
1611 }
1612 blocks = adjusted_blocks;
1613 }
1614 else
1615 {
1616 blocks *= RtsFlags.GcFlags.oldGenFactor;
1617 if (blocks < min_nursery)
1618 {
1619 blocks = min_nursery;
1620 }
1621 }
1622 resizeNurseries(blocks);
1623 }
1624 else // Generational collector
1625 {
1626 /*
1627 * If the user has given us a suggested heap size, adjust our
1628 * allocation area to make best use of the memory available.
1629 */
1630 if (RtsFlags.GcFlags.heapSizeSuggestion)
1631 {
1632 long blocks;
1633 const nat needed = calcNeeded(); // approx blocks needed at next GC
1634
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.
1638 *
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
1642 * (- 10 words).
1643 */
1644 if (N == 0)
1645 {
1646 g0_pcnt_kept = ((copied / (BLOCK_SIZE_W - 10)) * 100)
1647 / countNurseryBlocks();
1648 }
1649
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
1653 * close on average.
1654 *
1655 * Formula: suggested - needed
1656 * ----------------------------
1657 * 1 + g0_pcnt_kept/100
1658 *
1659 * where 'needed' is the amount of memory needed at the next
1660 * collection for collecting all gens except g0.
1661 */
1662 blocks =
1663 (((long)RtsFlags.GcFlags.heapSizeSuggestion - (long)needed) * 100) /
1664 (100 + (long)g0_pcnt_kept);
1665
1666 if (blocks < (long)min_nursery) {
1667 blocks = min_nursery;
1668 }
1669
1670 resizeNurseries((nat)blocks);
1671 }
1672 else
1673 {
1674 // we might have added extra large blocks to the nursery, so
1675 // resize back to minAllocAreaSize again.
1676 resizeNurseriesFixed(RtsFlags.GcFlags.minAllocAreaSize);
1677 }
1678 }
1679 }
1680
1681 /* -----------------------------------------------------------------------------
1682 Sanity code for CAF garbage collection.
1683
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.
1688
1689 Any garbage collected CAFs are taken off the CAF list at the same
1690 time.
1691 -------------------------------------------------------------------------- */
1692
1693 #if 0 && defined(DEBUG)
1694
1695 static void
1696 gcCAFs(void)
1697 {
1698 StgClosure* p;
1699 StgClosure** pp;
1700 const StgInfoTable *info;
1701 nat i;
1702
1703 i = 0;
1704 p = caf_list;
1705 pp = &caf_list;
1706
1707 while (p != NULL) {
1708
1709 info = get_itbl(p);
1710
1711 ASSERT(info->type == IND_STATIC);
1712
1713 if (STATIC_LINK(info,p) == NULL) {
1714 debugTrace(DEBUG_gccafs, "CAF gc'd at 0x%04lx", (long)p);
1715 // black hole it
1716 SET_INFO(p,&stg_BLACKHOLE_info);
1717 p = STATIC_LINK2(info,p);
1718 *pp = p;
1719 }
1720 else {
1721 pp = &STATIC_LINK2(info,p);
1722 p = *pp;
1723 i++;
1724 }
1725
1726 }
1727
1728 debugTrace(DEBUG_gccafs, "%d CAFs live", i);
1729 }
1730 #endif