c9908359b84216401f2ed1b6ce54622099431257
[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 "Compact.h"
44 #include "Evac.h"
45 #include "Scav.h"
46 #include "GCUtils.h"
47 #include "MarkStack.h"
48 #include "MarkWeak.h"
49 #include "Sparks.h"
50 #include "Sweep.h"
51
52 #include <string.h> // for memset()
53 #include <unistd.h>
54
55 /* -----------------------------------------------------------------------------
56 Global variables
57 -------------------------------------------------------------------------- */
58
59 /* STATIC OBJECT LIST.
60 *
61 * During GC:
62 * We maintain a linked list of static objects that are still live.
63 * The requirements for this list are:
64 *
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).
68 *
69 * - we need to be able to tell whether an object is already on
70 * the list, to break loops.
71 *
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
76 * shouldn't matter).
77 *
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.
80 *
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.
83 *
84 * Extra notes for generational GC:
85 *
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.
89 *
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.
92 */
93
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.
98 */
99 nat N;
100 rtsBool major_gc;
101
102 /* Data used for allocation area sizing.
103 */
104 static lnat g0_pcnt_kept = 30; // percentage of g0 live at last minor GC
105
106 /* Mut-list stats */
107 #ifdef DEBUG
108 nat mutlist_MUTVARS,
109 mutlist_MUTARRS,
110 mutlist_MVARS,
111 mutlist_OTHERS;
112 #endif
113
114 /* Thread-local data for each GC thread
115 */
116 gc_thread **gc_threads = NULL;
117
118 #if !defined(THREADED_RTS)
119 StgWord8 the_gc_thread[sizeof(gc_thread) + 64 * sizeof(gen_workspace)];
120 #endif
121
122 // Number of threads running in *this* GC. Affects how many
123 // step->todos[] lists we have to look in to find work.
124 nat n_gc_threads;
125
126 // For stats:
127 long copied; // *words* copied & scavenged during this GC
128
129 rtsBool work_stealing;
130
131 DECLARE_GCT
132
133 /* -----------------------------------------------------------------------------
134 Static function declarations
135 -------------------------------------------------------------------------- */
136
137 static void mark_root (void *user, StgClosure **root);
138 static void zero_static_object_list (StgClosure* first_static);
139 static nat initialise_N (rtsBool force_major_gc);
140 static void init_collected_gen (nat g, nat threads);
141 static void init_uncollected_gen (nat g, nat threads);
142 static void init_gc_thread (gc_thread *t);
143 static void resize_generations (void);
144 static void resize_nursery (void);
145 static void start_gc_threads (void);
146 static void scavenge_until_all_done (void);
147 static StgWord inc_running (void);
148 static StgWord dec_running (void);
149 static void wakeup_gc_threads (nat n_threads, nat me);
150 static void shutdown_gc_threads (nat n_threads, nat me);
151
152 #if 0 && defined(DEBUG)
153 static void gcCAFs (void);
154 #endif
155
156 /* -----------------------------------------------------------------------------
157 The mark stack.
158 -------------------------------------------------------------------------- */
159
160 bdescr *mark_stack_top_bd; // topmost block in the mark stack
161 bdescr *mark_stack_bd; // current block in the mark stack
162 StgPtr mark_sp; // pointer to the next unallocated mark stack entry
163
164 /* -----------------------------------------------------------------------------
165 GarbageCollect: the main entry point to the garbage collector.
166
167 Locks held: all capabilities are held throughout GarbageCollect().
168 -------------------------------------------------------------------------- */
169
170 void
171 GarbageCollect (rtsBool force_major_gc,
172 nat gc_type USED_IF_THREADS,
173 Capability *cap)
174 {
175 bdescr *bd;
176 generation *gen;
177 lnat live, allocated, max_copied, avg_copied, slop;
178 gc_thread *saved_gct;
179 nat g, t, n;
180
181 // necessary if we stole a callee-saves register for gct:
182 saved_gct = gct;
183
184 #ifdef PROFILING
185 CostCentreStack *prev_CCS;
186 #endif
187
188 ACQUIRE_SM_LOCK;
189
190 #if defined(RTS_USER_SIGNALS)
191 if (RtsFlags.MiscFlags.install_signal_handlers) {
192 // block signals
193 blockUserSignals();
194 }
195 #endif
196
197 ASSERT(sizeof(gen_workspace) == 16 * sizeof(StgWord));
198 // otherwise adjust the padding in gen_workspace.
199
200 // tell the stats department that we've started a GC
201 stat_startGC();
202
203 // tell the STM to discard any cached closures it's hoping to re-use
204 stmPreGCHook();
205
206 // lock the StablePtr table
207 stablePtrPreGC();
208
209 #ifdef DEBUG
210 mutlist_MUTVARS = 0;
211 mutlist_MUTARRS = 0;
212 mutlist_OTHERS = 0;
213 #endif
214
215 // attribute any costs to CCS_GC
216 #ifdef PROFILING
217 prev_CCS = CCCS;
218 CCCS = CCS_GC;
219 #endif
220
221 /* Approximate how much we allocated.
222 * Todo: only when generating stats?
223 */
224 allocated = calcAllocated(rtsFalse/* don't count the nursery yet */);
225
226 /* Figure out which generation to collect
227 */
228 n = initialise_N(force_major_gc);
229
230 #if defined(THREADED_RTS)
231 work_stealing = RtsFlags.ParFlags.parGcLoadBalancingEnabled &&
232 N >= RtsFlags.ParFlags.parGcLoadBalancingGen;
233 // It's not always a good idea to do load balancing in parallel
234 // GC. In particular, for a parallel program we don't want to
235 // lose locality by moving cached data into another CPU's cache
236 // (this effect can be quite significant).
237 //
238 // We could have a more complex way to deterimine whether to do
239 // work stealing or not, e.g. it might be a good idea to do it
240 // if the heap is big. For now, we just turn it on or off with
241 // a flag.
242 #endif
243
244 /* Start threads, so they can be spinning up while we finish initialisation.
245 */
246 start_gc_threads();
247
248 #if defined(THREADED_RTS)
249 /* How many threads will be participating in this GC?
250 * We don't try to parallelise minor GCs (unless the user asks for
251 * it with +RTS -gn0), or mark/compact/sweep GC.
252 */
253 if (gc_type == PENDING_GC_PAR) {
254 n_gc_threads = RtsFlags.ParFlags.nNodes;
255 } else {
256 n_gc_threads = 1;
257 }
258 #else
259 n_gc_threads = 1;
260 #endif
261
262 debugTrace(DEBUG_gc, "GC (gen %d): %d KB to collect, %ld MB in use, using %d thread(s)",
263 N, n * (BLOCK_SIZE / 1024), mblocks_allocated, n_gc_threads);
264
265 #ifdef RTS_GTK_FRONTPANEL
266 if (RtsFlags.GcFlags.frontpanel) {
267 updateFrontPanelBeforeGC(N);
268 }
269 #endif
270
271 #ifdef DEBUG
272 // check for memory leaks if DEBUG is on
273 memInventory(DEBUG_gc);
274 #endif
275
276 // check sanity *before* GC
277 IF_DEBUG(sanity, checkSanity(rtsTrue));
278
279 // Initialise all our gc_thread structures
280 for (t = 0; t < n_gc_threads; t++) {
281 init_gc_thread(gc_threads[t]);
282 }
283
284 // Initialise all the generations/steps that we're collecting.
285 for (g = 0; g <= N; g++) {
286 init_collected_gen(g,n_gc_threads);
287 }
288
289 // Initialise all the generations/steps that we're *not* collecting.
290 for (g = N+1; g < RtsFlags.GcFlags.generations; g++) {
291 init_uncollected_gen(g,n_gc_threads);
292 }
293
294 /* Allocate a mark stack if we're doing a major collection.
295 */
296 if (major_gc && oldest_gen->mark) {
297 mark_stack_bd = allocBlock();
298 mark_stack_top_bd = mark_stack_bd;
299 mark_stack_bd->link = NULL;
300 mark_stack_bd->u.back = NULL;
301 mark_sp = mark_stack_bd->start;
302 } else {
303 mark_stack_bd = NULL;
304 mark_stack_top_bd = NULL;
305 mark_sp = NULL;
306 }
307
308 // this is the main thread
309 #ifdef THREADED_RTS
310 if (n_gc_threads == 1) {
311 SET_GCT(gc_threads[0]);
312 } else {
313 SET_GCT(gc_threads[cap->no]);
314 }
315 #else
316 SET_GCT(gc_threads[0]);
317 #endif
318
319 /* -----------------------------------------------------------------------
320 * follow all the roots that we know about:
321 */
322
323 // the main thread is running: this prevents any other threads from
324 // exiting prematurely, so we can start them now.
325 // NB. do this after the mutable lists have been saved above, otherwise
326 // the other GC threads will be writing into the old mutable lists.
327 inc_running();
328 wakeup_gc_threads(n_gc_threads, gct->thread_index);
329
330 // scavenge the capability-private mutable lists. This isn't part
331 // of markSomeCapabilities() because markSomeCapabilities() can only
332 // call back into the GC via mark_root() (due to the gct register
333 // variable).
334 if (n_gc_threads == 1) {
335 for (n = 0; n < n_capabilities; n++) {
336 #if defined(THREADED_RTS)
337 scavenge_capability_mut_Lists1(&capabilities[n]);
338 #else
339 scavenge_capability_mut_lists(&capabilities[n]);
340 #endif
341 }
342 } else {
343 scavenge_capability_mut_lists(&capabilities[gct->thread_index]);
344 }
345
346 // follow roots from the CAF list (used by GHCi)
347 gct->evac_gen_no = 0;
348 markCAFs(mark_root, gct);
349
350 // follow all the roots that the application knows about.
351 gct->evac_gen_no = 0;
352 markSomeCapabilities(mark_root, gct, gct->thread_index, n_gc_threads,
353 rtsTrue/*prune sparks*/);
354
355 #if defined(RTS_USER_SIGNALS)
356 // mark the signal handlers (signals should be already blocked)
357 markSignalHandlers(mark_root, gct);
358 #endif
359
360 // Mark the weak pointer list, and prepare to detect dead weak pointers.
361 markWeakPtrList();
362 initWeakForGC();
363
364 // Mark the stable pointer table.
365 markStablePtrTable(mark_root, gct);
366
367 /* -------------------------------------------------------------------------
368 * Repeatedly scavenge all the areas we know about until there's no
369 * more scavenging to be done.
370 */
371 for (;;)
372 {
373 scavenge_until_all_done();
374 // The other threads are now stopped. We might recurse back to
375 // here, but from now on this is the only thread.
376
377 // must be last... invariant is that everything is fully
378 // scavenged at this point.
379 if (traverseWeakPtrList()) { // returns rtsTrue if evaced something
380 inc_running();
381 continue;
382 }
383
384 // If we get to here, there's really nothing left to do.
385 break;
386 }
387
388 shutdown_gc_threads(n_gc_threads, gct->thread_index);
389
390 // Now see which stable names are still alive.
391 gcStablePtrTable();
392
393 #ifdef THREADED_RTS
394 if (n_gc_threads == 1) {
395 for (n = 0; n < n_capabilities; n++) {
396 pruneSparkQueue(&capabilities[n]);
397 }
398 } else {
399 pruneSparkQueue(&capabilities[gct->thread_index]);
400 }
401 #endif
402
403 #ifdef PROFILING
404 // We call processHeapClosureForDead() on every closure destroyed during
405 // the current garbage collection, so we invoke LdvCensusForDead().
406 if (RtsFlags.ProfFlags.doHeapProfile == HEAP_BY_LDV
407 || RtsFlags.ProfFlags.bioSelector != NULL)
408 LdvCensusForDead(N);
409 #endif
410
411 // NO MORE EVACUATION AFTER THIS POINT!
412
413 // Two-space collector: free the old to-space.
414 // g0->old_blocks is the old nursery
415 // g0->blocks is to-space from the previous GC
416 if (RtsFlags.GcFlags.generations == 1) {
417 if (g0->blocks != NULL) {
418 freeChain(g0->blocks);
419 g0->blocks = NULL;
420 }
421 }
422
423 // For each workspace, in each thread, move the copied blocks to the step
424 {
425 gc_thread *thr;
426 gen_workspace *ws;
427 bdescr *prev, *next;
428
429 for (t = 0; t < n_gc_threads; t++) {
430 thr = gc_threads[t];
431
432 for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
433 ws = &thr->gens[g];
434
435 // Push the final block
436 if (ws->todo_bd) {
437 push_scanned_block(ws->todo_bd, ws);
438 }
439
440 ASSERT(gct->scan_bd == NULL);
441 ASSERT(countBlocks(ws->scavd_list) == ws->n_scavd_blocks);
442
443 prev = NULL;
444 for (bd = ws->scavd_list; bd != NULL; bd = bd->link) {
445 ws->gen->n_words += bd->free - bd->start;
446 prev = bd;
447 }
448 if (prev != NULL) {
449 prev->link = ws->gen->blocks;
450 ws->gen->blocks = ws->scavd_list;
451 }
452 ws->gen->n_blocks += ws->n_scavd_blocks;
453 }
454 }
455
456 // Add all the partial blocks *after* we've added all the full
457 // blocks. This is so that we can grab the partial blocks back
458 // again and try to fill them up in the next GC.
459 for (t = 0; t < n_gc_threads; t++) {
460 thr = gc_threads[t];
461
462 for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
463 ws = &thr->gens[g];
464
465 prev = NULL;
466 for (bd = ws->part_list; bd != NULL; bd = next) {
467 next = bd->link;
468 if (bd->free == bd->start) {
469 if (prev == NULL) {
470 ws->part_list = next;
471 } else {
472 prev->link = next;
473 }
474 freeGroup(bd);
475 ws->n_part_blocks--;
476 } else {
477 ws->gen->n_words += bd->free - bd->start;
478 prev = bd;
479 }
480 }
481 if (prev != NULL) {
482 prev->link = ws->gen->blocks;
483 ws->gen->blocks = ws->part_list;
484 }
485 ws->gen->n_blocks += ws->n_part_blocks;
486
487 ASSERT(countBlocks(ws->gen->blocks) == ws->gen->n_blocks);
488 ASSERT(countOccupied(ws->gen->blocks) == ws->gen->n_words);
489 }
490 }
491 }
492
493 // Finally: compact or sweep the oldest generation.
494 if (major_gc && oldest_gen->mark) {
495 if (oldest_gen->compact)
496 compact(gct->scavenged_static_objects);
497 else
498 sweep(oldest_gen);
499 }
500
501 /* run through all the generations/steps and tidy up
502 */
503 copied = 0;
504 max_copied = 0;
505 avg_copied = 0;
506 {
507 nat i;
508 for (i=0; i < n_gc_threads; i++) {
509 if (n_gc_threads > 1) {
510 debugTrace(DEBUG_gc,"thread %d:", i);
511 debugTrace(DEBUG_gc," copied %ld", gc_threads[i]->copied * sizeof(W_));
512 debugTrace(DEBUG_gc," scanned %ld", gc_threads[i]->scanned * sizeof(W_));
513 debugTrace(DEBUG_gc," any_work %ld", gc_threads[i]->any_work);
514 debugTrace(DEBUG_gc," no_work %ld", gc_threads[i]->no_work);
515 debugTrace(DEBUG_gc," scav_find_work %ld", gc_threads[i]->scav_find_work);
516 }
517 copied += gc_threads[i]->copied;
518 max_copied = stg_max(gc_threads[i]->copied, max_copied);
519 }
520 if (n_gc_threads == 1) {
521 max_copied = 0;
522 avg_copied = 0;
523 } else {
524 avg_copied = copied;
525 }
526 }
527
528 for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
529
530 if (g == N) {
531 generations[g].collections++; // for stats
532 if (n_gc_threads > 1) generations[g].par_collections++;
533 }
534
535 // Count the mutable list as bytes "copied" for the purposes of
536 // stats. Every mutable list is copied during every GC.
537 if (g > 0) {
538 nat mut_list_size = 0;
539 for (n = 0; n < n_capabilities; n++) {
540 mut_list_size += countOccupied(capabilities[n].mut_lists[g]);
541 }
542 copied += mut_list_size;
543
544 debugTrace(DEBUG_gc,
545 "mut_list_size: %lu (%d vars, %d arrays, %d MVARs, %d others)",
546 (unsigned long)(mut_list_size * sizeof(W_)),
547 mutlist_MUTVARS, mutlist_MUTARRS, mutlist_MVARS, mutlist_OTHERS);
548 }
549
550 bdescr *next, *prev;
551 gen = &generations[g];
552
553 // for generations we collected...
554 if (g <= N) {
555
556 /* free old memory and shift to-space into from-space for all
557 * the collected steps (except the allocation area). These
558 * freed blocks will probaby be quickly recycled.
559 */
560 if (gen->mark)
561 {
562 // tack the new blocks on the end of the existing blocks
563 if (gen->old_blocks != NULL) {
564
565 prev = NULL;
566 for (bd = gen->old_blocks; bd != NULL; bd = next) {
567
568 next = bd->link;
569
570 if (!(bd->flags & BF_MARKED))
571 {
572 if (prev == NULL) {
573 gen->old_blocks = next;
574 } else {
575 prev->link = next;
576 }
577 freeGroup(bd);
578 gen->n_old_blocks--;
579 }
580 else
581 {
582 gen->n_words += bd->free - bd->start;
583
584 // NB. this step might not be compacted next
585 // time, so reset the BF_MARKED flags.
586 // They are set before GC if we're going to
587 // compact. (search for BF_MARKED above).
588 bd->flags &= ~BF_MARKED;
589
590 // between GCs, all blocks in the heap except
591 // for the nursery have the BF_EVACUATED flag set.
592 bd->flags |= BF_EVACUATED;
593
594 prev = bd;
595 }
596 }
597
598 if (prev != NULL) {
599 prev->link = gen->blocks;
600 gen->blocks = gen->old_blocks;
601 }
602 }
603 // add the new blocks to the block tally
604 gen->n_blocks += gen->n_old_blocks;
605 ASSERT(countBlocks(gen->blocks) == gen->n_blocks);
606 ASSERT(countOccupied(gen->blocks) == gen->n_words);
607 }
608 else // not copacted
609 {
610 freeChain(gen->old_blocks);
611 }
612
613 gen->old_blocks = NULL;
614 gen->n_old_blocks = 0;
615
616 /* LARGE OBJECTS. The current live large objects are chained on
617 * scavenged_large, having been moved during garbage
618 * collection from large_objects. Any objects left on the
619 * large_objects list are therefore dead, so we free them here.
620 */
621 freeChain(gen->large_objects);
622 gen->large_objects = gen->scavenged_large_objects;
623 gen->n_large_blocks = gen->n_scavenged_large_blocks;
624 gen->n_new_large_words = 0;
625 ASSERT(countBlocks(gen->large_objects) == gen->n_large_blocks);
626 }
627 else // for generations > N
628 {
629 /* For older generations, we need to append the
630 * scavenged_large_object list (i.e. large objects that have been
631 * promoted during this GC) to the large_object list for that step.
632 */
633 for (bd = gen->scavenged_large_objects; bd; bd = next) {
634 next = bd->link;
635 dbl_link_onto(bd, &gen->large_objects);
636 }
637
638 // add the new blocks we promoted during this GC
639 gen->n_large_blocks += gen->n_scavenged_large_blocks;
640 ASSERT(countBlocks(gen->large_objects) == gen->n_large_blocks);
641 }
642 } // for all generations
643
644 // update the max size of older generations after a major GC
645 resize_generations();
646
647 // Calculate the amount of live data for stats.
648 live = calcLiveWords();
649
650 // Start a new pinned_object_block
651 for (n = 0; n < n_capabilities; n++) {
652 capabilities[n].pinned_object_block = NULL;
653 }
654
655 // Free the mark stack.
656 if (mark_stack_top_bd != NULL) {
657 debugTrace(DEBUG_gc, "mark stack: %d blocks",
658 countBlocks(mark_stack_top_bd));
659 freeChain(mark_stack_top_bd);
660 }
661
662 // Free any bitmaps.
663 for (g = 0; g <= N; g++) {
664 gen = &generations[g];
665 if (gen->bitmap != NULL) {
666 freeGroup(gen->bitmap);
667 gen->bitmap = NULL;
668 }
669 }
670
671 // Reset the nursery: make the blocks empty
672 allocated += clearNurseries();
673
674 resize_nursery();
675
676 resetNurseries();
677
678 // mark the garbage collected CAFs as dead
679 #if 0 && defined(DEBUG) // doesn't work at the moment
680 if (major_gc) { gcCAFs(); }
681 #endif
682
683 #ifdef PROFILING
684 // resetStaticObjectForRetainerProfiling() must be called before
685 // zeroing below.
686 if (n_gc_threads > 1) {
687 barf("profiling is currently broken with multi-threaded GC");
688 // ToDo: fix the gct->scavenged_static_objects below
689 }
690 resetStaticObjectForRetainerProfiling(gct->scavenged_static_objects);
691 #endif
692
693 // zero the scavenged static object list
694 if (major_gc) {
695 nat i;
696 for (i = 0; i < n_gc_threads; i++) {
697 zero_static_object_list(gc_threads[i]->scavenged_static_objects);
698 }
699 }
700
701 // send exceptions to any threads which were about to die
702 RELEASE_SM_LOCK;
703 resurrectThreads(resurrected_threads);
704 ACQUIRE_SM_LOCK;
705
706 // Update the stable pointer hash table.
707 updateStablePtrTable(major_gc);
708
709 // unlock the StablePtr table. Must be before scheduleFinalizers(),
710 // because a finalizer may call hs_free_fun_ptr() or
711 // hs_free_stable_ptr(), both of which access the StablePtr table.
712 stablePtrPostGC();
713
714 // Start any pending finalizers. Must be after
715 // updateStablePtrTable() and stablePtrPostGC() (see #4221).
716 RELEASE_SM_LOCK;
717 scheduleFinalizers(cap, old_weak_ptr_list);
718 ACQUIRE_SM_LOCK;
719
720 if (major_gc) {
721 nat need, got;
722 need = BLOCKS_TO_MBLOCKS(n_alloc_blocks);
723 got = mblocks_allocated;
724 /* If the amount of data remains constant, next major GC we'll
725 require (F+1)*need. We leave (F+2)*need in order to reduce
726 repeated deallocation and reallocation. */
727 need = (RtsFlags.GcFlags.oldGenFactor + 2) * need;
728 if (got > need) {
729 returnMemoryToOS(got - need);
730 }
731 }
732
733 // check sanity after GC
734 IF_DEBUG(sanity, checkSanity(rtsTrue));
735
736 // extra GC trace info
737 IF_DEBUG(gc, statDescribeGens());
738
739 #ifdef DEBUG
740 // symbol-table based profiling
741 /* heapCensus(to_blocks); */ /* ToDo */
742 #endif
743
744 // restore enclosing cost centre
745 #ifdef PROFILING
746 CCCS = prev_CCS;
747 #endif
748
749 #ifdef DEBUG
750 // check for memory leaks if DEBUG is on
751 memInventory(DEBUG_gc);
752 #endif
753
754 #ifdef RTS_GTK_FRONTPANEL
755 if (RtsFlags.GcFlags.frontpanel) {
756 updateFrontPanelAfterGC( N, live );
757 }
758 #endif
759
760 // ok, GC over: tell the stats department what happened.
761 slop = calcLiveBlocks() * BLOCK_SIZE_W - live;
762 stat_endGC(allocated, live, copied, N, max_copied, avg_copied, slop);
763
764 // Guess which generation we'll collect *next* time
765 initialise_N(force_major_gc);
766
767 #if defined(RTS_USER_SIGNALS)
768 if (RtsFlags.MiscFlags.install_signal_handlers) {
769 // unblock signals again
770 unblockUserSignals();
771 }
772 #endif
773
774 RELEASE_SM_LOCK;
775
776 SET_GCT(saved_gct);
777 }
778
779 /* -----------------------------------------------------------------------------
780 Figure out which generation to collect, initialise N and major_gc.
781
782 Also returns the total number of blocks in generations that will be
783 collected.
784 -------------------------------------------------------------------------- */
785
786 static nat
787 initialise_N (rtsBool force_major_gc)
788 {
789 int g;
790 nat blocks, blocks_total;
791
792 blocks = 0;
793 blocks_total = 0;
794
795 if (force_major_gc) {
796 N = RtsFlags.GcFlags.generations - 1;
797 } else {
798 N = 0;
799 }
800
801 for (g = RtsFlags.GcFlags.generations - 1; g >= 0; g--) {
802
803 blocks = generations[g].n_words / BLOCK_SIZE_W
804 + generations[g].n_large_blocks;
805
806 if (blocks >= generations[g].max_blocks) {
807 N = stg_max(N,g);
808 }
809 if ((nat)g <= N) {
810 blocks_total += blocks;
811 }
812 }
813
814 blocks_total += countNurseryBlocks();
815
816 major_gc = (N == RtsFlags.GcFlags.generations-1);
817 return blocks_total;
818 }
819
820 /* -----------------------------------------------------------------------------
821 Initialise the gc_thread structures.
822 -------------------------------------------------------------------------- */
823
824 #define GC_THREAD_INACTIVE 0
825 #define GC_THREAD_STANDING_BY 1
826 #define GC_THREAD_RUNNING 2
827 #define GC_THREAD_WAITING_TO_CONTINUE 3
828
829 static void
830 new_gc_thread (nat n, gc_thread *t)
831 {
832 nat g;
833 gen_workspace *ws;
834
835 #ifdef THREADED_RTS
836 t->id = 0;
837 initSpinLock(&t->gc_spin);
838 initSpinLock(&t->mut_spin);
839 ACQUIRE_SPIN_LOCK(&t->gc_spin);
840 t->wakeup = GC_THREAD_INACTIVE; // starts true, so we can wait for the
841 // thread to start up, see wakeup_gc_threads
842 #endif
843
844 t->thread_index = n;
845 t->free_blocks = NULL;
846 t->gc_count = 0;
847
848 init_gc_thread(t);
849
850 #ifdef USE_PAPI
851 t->papi_events = -1;
852 #endif
853
854 for (g = 0; g < RtsFlags.GcFlags.generations; g++)
855 {
856 ws = &t->gens[g];
857 ws->gen = &generations[g];
858 ASSERT(g == ws->gen->no);
859 ws->my_gct = t;
860
861 ws->todo_bd = NULL;
862 ws->todo_q = newWSDeque(128);
863 ws->todo_overflow = NULL;
864 ws->n_todo_overflow = 0;
865
866 ws->part_list = NULL;
867 ws->n_part_blocks = 0;
868
869 ws->scavd_list = NULL;
870 ws->n_scavd_blocks = 0;
871 }
872 }
873
874
875 void
876 initGcThreads (void)
877 {
878 if (gc_threads == NULL) {
879 #if defined(THREADED_RTS)
880 nat i;
881 gc_threads = stgMallocBytes (RtsFlags.ParFlags.nNodes *
882 sizeof(gc_thread*),
883 "alloc_gc_threads");
884
885 for (i = 0; i < RtsFlags.ParFlags.nNodes; i++) {
886 gc_threads[i] =
887 stgMallocBytes(sizeof(gc_thread) +
888 RtsFlags.GcFlags.generations * sizeof(gen_workspace),
889 "alloc_gc_threads");
890
891 new_gc_thread(i, gc_threads[i]);
892 }
893 #else
894 gc_threads = stgMallocBytes (sizeof(gc_thread*),"alloc_gc_threads");
895 gc_threads[0] = gct;
896 new_gc_thread(0,gc_threads[0]);
897 #endif
898 }
899 }
900
901 void
902 freeGcThreads (void)
903 {
904 nat g;
905 if (gc_threads != NULL) {
906 #if defined(THREADED_RTS)
907 nat i;
908 for (i = 0; i < n_capabilities; i++) {
909 for (g = 0; g < RtsFlags.GcFlags.generations; g++)
910 {
911 freeWSDeque(gc_threads[i]->gens[g].todo_q);
912 }
913 stgFree (gc_threads[i]);
914 }
915 stgFree (gc_threads);
916 #else
917 for (g = 0; g < RtsFlags.GcFlags.generations; g++)
918 {
919 freeWSDeque(gc_threads[0]->gens[g].todo_q);
920 }
921 stgFree (gc_threads);
922 #endif
923 gc_threads = NULL;
924 }
925 }
926
927 /* ----------------------------------------------------------------------------
928 Start GC threads
929 ------------------------------------------------------------------------- */
930
931 static volatile StgWord gc_running_threads;
932
933 static StgWord
934 inc_running (void)
935 {
936 StgWord new;
937 new = atomic_inc(&gc_running_threads);
938 ASSERT(new <= n_gc_threads);
939 return new;
940 }
941
942 static StgWord
943 dec_running (void)
944 {
945 ASSERT(gc_running_threads != 0);
946 return atomic_dec(&gc_running_threads);
947 }
948
949 static rtsBool
950 any_work (void)
951 {
952 int g;
953 gen_workspace *ws;
954
955 gct->any_work++;
956
957 write_barrier();
958
959 // scavenge objects in compacted generation
960 if (mark_stack_bd != NULL && !mark_stack_empty()) {
961 return rtsTrue;
962 }
963
964 // Check for global work in any step. We don't need to check for
965 // local work, because we have already exited scavenge_loop(),
966 // which means there is no local work for this thread.
967 for (g = 0; g < (int)RtsFlags.GcFlags.generations; g++) {
968 ws = &gct->gens[g];
969 if (ws->todo_large_objects) return rtsTrue;
970 if (!looksEmptyWSDeque(ws->todo_q)) return rtsTrue;
971 if (ws->todo_overflow) return rtsTrue;
972 }
973
974 #if defined(THREADED_RTS)
975 if (work_stealing) {
976 nat n;
977 // look for work to steal
978 for (n = 0; n < n_gc_threads; n++) {
979 if (n == gct->thread_index) continue;
980 for (g = RtsFlags.GcFlags.generations-1; g >= 0; g--) {
981 ws = &gc_threads[n]->gens[g];
982 if (!looksEmptyWSDeque(ws->todo_q)) return rtsTrue;
983 }
984 }
985 }
986 #endif
987
988 gct->no_work++;
989 #if defined(THREADED_RTS)
990 yieldThread();
991 #endif
992
993 return rtsFalse;
994 }
995
996 static void
997 scavenge_until_all_done (void)
998 {
999 nat r;
1000
1001
1002 loop:
1003 traceEventGcWork(&capabilities[gct->thread_index]);
1004
1005 #if defined(THREADED_RTS)
1006 if (n_gc_threads > 1) {
1007 scavenge_loop();
1008 } else {
1009 scavenge_loop1();
1010 }
1011 #else
1012 scavenge_loop();
1013 #endif
1014
1015 // scavenge_loop() only exits when there's no work to do
1016 r = dec_running();
1017
1018 traceEventGcIdle(&capabilities[gct->thread_index]);
1019
1020 debugTrace(DEBUG_gc, "%d GC threads still running", r);
1021
1022 while (gc_running_threads != 0) {
1023 // usleep(1);
1024 if (any_work()) {
1025 inc_running();
1026 goto loop;
1027 }
1028 // any_work() does not remove the work from the queue, it
1029 // just checks for the presence of work. If we find any,
1030 // then we increment gc_running_threads and go back to
1031 // scavenge_loop() to perform any pending work.
1032 }
1033
1034 traceEventGcDone(&capabilities[gct->thread_index]);
1035 }
1036
1037 #if defined(THREADED_RTS)
1038
1039 void
1040 gcWorkerThread (Capability *cap)
1041 {
1042 gc_thread *saved_gct;
1043
1044 // necessary if we stole a callee-saves register for gct:
1045 saved_gct = gct;
1046
1047 gct = gc_threads[cap->no];
1048 gct->id = osThreadId();
1049
1050 // Wait until we're told to wake up
1051 RELEASE_SPIN_LOCK(&gct->mut_spin);
1052 gct->wakeup = GC_THREAD_STANDING_BY;
1053 debugTrace(DEBUG_gc, "GC thread %d standing by...", gct->thread_index);
1054 ACQUIRE_SPIN_LOCK(&gct->gc_spin);
1055
1056 #ifdef USE_PAPI
1057 // start performance counters in this thread...
1058 if (gct->papi_events == -1) {
1059 papi_init_eventset(&gct->papi_events);
1060 }
1061 papi_thread_start_gc1_count(gct->papi_events);
1062 #endif
1063
1064 // Every thread evacuates some roots.
1065 gct->evac_gen_no = 0;
1066 markSomeCapabilities(mark_root, gct, gct->thread_index, n_gc_threads,
1067 rtsTrue/*prune sparks*/);
1068 scavenge_capability_mut_lists(&capabilities[gct->thread_index]);
1069
1070 scavenge_until_all_done();
1071
1072 #ifdef THREADED_RTS
1073 // Now that the whole heap is marked, we discard any sparks that
1074 // were found to be unreachable. The main GC thread is currently
1075 // marking heap reachable via weak pointers, so it is
1076 // non-deterministic whether a spark will be retained if it is
1077 // only reachable via weak pointers. To fix this problem would
1078 // require another GC barrier, which is too high a price.
1079 pruneSparkQueue(cap);
1080 #endif
1081
1082 #ifdef USE_PAPI
1083 // count events in this thread towards the GC totals
1084 papi_thread_stop_gc1_count(gct->papi_events);
1085 #endif
1086
1087 // Wait until we're told to continue
1088 RELEASE_SPIN_LOCK(&gct->gc_spin);
1089 gct->wakeup = GC_THREAD_WAITING_TO_CONTINUE;
1090 debugTrace(DEBUG_gc, "GC thread %d waiting to continue...",
1091 gct->thread_index);
1092 ACQUIRE_SPIN_LOCK(&gct->mut_spin);
1093 debugTrace(DEBUG_gc, "GC thread %d on my way...", gct->thread_index);
1094
1095 SET_GCT(saved_gct);
1096 }
1097
1098 #endif
1099
1100 #if defined(THREADED_RTS)
1101
1102 void
1103 waitForGcThreads (Capability *cap USED_IF_THREADS)
1104 {
1105 const nat n_threads = RtsFlags.ParFlags.nNodes;
1106 const nat me = cap->no;
1107 nat i, j;
1108 rtsBool retry = rtsTrue;
1109
1110 while(retry) {
1111 for (i=0; i < n_threads; i++) {
1112 if (i == me) continue;
1113 if (gc_threads[i]->wakeup != GC_THREAD_STANDING_BY) {
1114 prodCapability(&capabilities[i], cap->running_task);
1115 }
1116 }
1117 for (j=0; j < 10; j++) {
1118 retry = rtsFalse;
1119 for (i=0; i < n_threads; i++) {
1120 if (i == me) continue;
1121 write_barrier();
1122 setContextSwitches();
1123 if (gc_threads[i]->wakeup != GC_THREAD_STANDING_BY) {
1124 retry = rtsTrue;
1125 }
1126 }
1127 if (!retry) break;
1128 yieldThread();
1129 }
1130 }
1131 }
1132
1133 #endif // THREADED_RTS
1134
1135 static void
1136 start_gc_threads (void)
1137 {
1138 #if defined(THREADED_RTS)
1139 gc_running_threads = 0;
1140 #endif
1141 }
1142
1143 static void
1144 wakeup_gc_threads (nat n_threads USED_IF_THREADS, nat me USED_IF_THREADS)
1145 {
1146 #if defined(THREADED_RTS)
1147 nat i;
1148 for (i=0; i < n_threads; i++) {
1149 if (i == me) continue;
1150 inc_running();
1151 debugTrace(DEBUG_gc, "waking up gc thread %d", i);
1152 if (gc_threads[i]->wakeup != GC_THREAD_STANDING_BY) barf("wakeup_gc_threads");
1153
1154 gc_threads[i]->wakeup = GC_THREAD_RUNNING;
1155 ACQUIRE_SPIN_LOCK(&gc_threads[i]->mut_spin);
1156 RELEASE_SPIN_LOCK(&gc_threads[i]->gc_spin);
1157 }
1158 #endif
1159 }
1160
1161 // After GC is complete, we must wait for all GC threads to enter the
1162 // standby state, otherwise they may still be executing inside
1163 // any_work(), and may even remain awake until the next GC starts.
1164 static void
1165 shutdown_gc_threads (nat n_threads USED_IF_THREADS, nat me USED_IF_THREADS)
1166 {
1167 #if defined(THREADED_RTS)
1168 nat i;
1169 for (i=0; i < n_threads; i++) {
1170 if (i == me) continue;
1171 while (gc_threads[i]->wakeup != GC_THREAD_WAITING_TO_CONTINUE) { write_barrier(); }
1172 }
1173 #endif
1174 }
1175
1176 #if defined(THREADED_RTS)
1177 void
1178 releaseGCThreads (Capability *cap USED_IF_THREADS)
1179 {
1180 const nat n_threads = RtsFlags.ParFlags.nNodes;
1181 const nat me = cap->no;
1182 nat i;
1183 for (i=0; i < n_threads; i++) {
1184 if (i == me) continue;
1185 if (gc_threads[i]->wakeup != GC_THREAD_WAITING_TO_CONTINUE)
1186 barf("releaseGCThreads");
1187
1188 gc_threads[i]->wakeup = GC_THREAD_INACTIVE;
1189 ACQUIRE_SPIN_LOCK(&gc_threads[i]->gc_spin);
1190 RELEASE_SPIN_LOCK(&gc_threads[i]->mut_spin);
1191 }
1192 }
1193 #endif
1194
1195 /* ----------------------------------------------------------------------------
1196 Initialise a generation that is to be collected
1197 ------------------------------------------------------------------------- */
1198
1199 static void
1200 init_collected_gen (nat g, nat n_threads)
1201 {
1202 nat t, i;
1203 gen_workspace *ws;
1204 generation *gen;
1205 bdescr *bd;
1206
1207 // Throw away the current mutable list. Invariant: the mutable
1208 // list always has at least one block; this means we can avoid a
1209 // check for NULL in recordMutable().
1210 if (g != 0) {
1211 for (i = 0; i < n_capabilities; i++) {
1212 freeChain(capabilities[i].mut_lists[g]);
1213 capabilities[i].mut_lists[g] = allocBlock();
1214 }
1215 }
1216
1217 gen = &generations[g];
1218 ASSERT(gen->no == g);
1219
1220 // we'll construct a new list of threads in this step
1221 // during GC, throw away the current list.
1222 gen->old_threads = gen->threads;
1223 gen->threads = END_TSO_QUEUE;
1224
1225 // deprecate the existing blocks
1226 gen->old_blocks = gen->blocks;
1227 gen->n_old_blocks = gen->n_blocks;
1228 gen->blocks = NULL;
1229 gen->n_blocks = 0;
1230 gen->n_words = 0;
1231 gen->live_estimate = 0;
1232
1233 // initialise the large object queues.
1234 gen->scavenged_large_objects = NULL;
1235 gen->n_scavenged_large_blocks = 0;
1236
1237 // mark the small objects as from-space
1238 for (bd = gen->old_blocks; bd; bd = bd->link) {
1239 bd->flags &= ~BF_EVACUATED;
1240 }
1241
1242 // mark the large objects as from-space
1243 for (bd = gen->large_objects; bd; bd = bd->link) {
1244 bd->flags &= ~BF_EVACUATED;
1245 }
1246
1247 // for a compacted generation, we need to allocate the bitmap
1248 if (gen->mark) {
1249 nat bitmap_size; // in bytes
1250 bdescr *bitmap_bdescr;
1251 StgWord *bitmap;
1252
1253 bitmap_size = gen->n_old_blocks * BLOCK_SIZE / (sizeof(W_)*BITS_PER_BYTE);
1254
1255 if (bitmap_size > 0) {
1256 bitmap_bdescr = allocGroup((lnat)BLOCK_ROUND_UP(bitmap_size)
1257 / BLOCK_SIZE);
1258 gen->bitmap = bitmap_bdescr;
1259 bitmap = bitmap_bdescr->start;
1260
1261 debugTrace(DEBUG_gc, "bitmap_size: %d, bitmap: %p",
1262 bitmap_size, bitmap);
1263
1264 // don't forget to fill it with zeros!
1265 memset(bitmap, 0, bitmap_size);
1266
1267 // For each block in this step, point to its bitmap from the
1268 // block descriptor.
1269 for (bd=gen->old_blocks; bd != NULL; bd = bd->link) {
1270 bd->u.bitmap = bitmap;
1271 bitmap += BLOCK_SIZE_W / (sizeof(W_)*BITS_PER_BYTE);
1272
1273 // Also at this point we set the BF_MARKED flag
1274 // for this block. The invariant is that
1275 // BF_MARKED is always unset, except during GC
1276 // when it is set on those blocks which will be
1277 // compacted.
1278 if (!(bd->flags & BF_FRAGMENTED)) {
1279 bd->flags |= BF_MARKED;
1280 }
1281
1282 // BF_SWEPT should be marked only for blocks that are being
1283 // collected in sweep()
1284 bd->flags &= ~BF_SWEPT;
1285 }
1286 }
1287 }
1288
1289 // For each GC thread, for each step, allocate a "todo" block to
1290 // store evacuated objects to be scavenged, and a block to store
1291 // evacuated objects that do not need to be scavenged.
1292 for (t = 0; t < n_threads; t++) {
1293 ws = &gc_threads[t]->gens[g];
1294
1295 ws->todo_large_objects = NULL;
1296
1297 ws->part_list = NULL;
1298 ws->n_part_blocks = 0;
1299
1300 // allocate the first to-space block; extra blocks will be
1301 // chained on as necessary.
1302 ws->todo_bd = NULL;
1303 ASSERT(looksEmptyWSDeque(ws->todo_q));
1304 alloc_todo_block(ws,0);
1305
1306 ws->todo_overflow = NULL;
1307 ws->n_todo_overflow = 0;
1308
1309 ws->scavd_list = NULL;
1310 ws->n_scavd_blocks = 0;
1311 }
1312 }
1313
1314
1315 /* ----------------------------------------------------------------------------
1316 Initialise a generation that is *not* to be collected
1317 ------------------------------------------------------------------------- */
1318
1319 static void
1320 init_uncollected_gen (nat g, nat threads)
1321 {
1322 nat t, n;
1323 gen_workspace *ws;
1324 generation *gen;
1325 bdescr *bd;
1326
1327 // save the current mutable lists for this generation, and
1328 // allocate a fresh block for each one. We'll traverse these
1329 // mutable lists as roots early on in the GC.
1330 for (n = 0; n < n_capabilities; n++) {
1331 capabilities[n].saved_mut_lists[g] = capabilities[n].mut_lists[g];
1332 capabilities[n].mut_lists[g] = allocBlock();
1333 }
1334
1335 gen = &generations[g];
1336
1337 gen->scavenged_large_objects = NULL;
1338 gen->n_scavenged_large_blocks = 0;
1339
1340 for (t = 0; t < threads; t++) {
1341 ws = &gc_threads[t]->gens[g];
1342
1343 ASSERT(looksEmptyWSDeque(ws->todo_q));
1344 ws->todo_large_objects = NULL;
1345
1346 ws->part_list = NULL;
1347 ws->n_part_blocks = 0;
1348
1349 ws->scavd_list = NULL;
1350 ws->n_scavd_blocks = 0;
1351
1352 // If the block at the head of the list in this generation
1353 // is less than 3/4 full, then use it as a todo block.
1354 if (gen->blocks && isPartiallyFull(gen->blocks))
1355 {
1356 ws->todo_bd = gen->blocks;
1357 ws->todo_free = ws->todo_bd->free;
1358 ws->todo_lim = ws->todo_bd->start + BLOCK_SIZE_W;
1359 gen->blocks = gen->blocks->link;
1360 gen->n_blocks -= 1;
1361 gen->n_words -= ws->todo_bd->free - ws->todo_bd->start;
1362 ws->todo_bd->link = NULL;
1363 // we must scan from the current end point.
1364 ws->todo_bd->u.scan = ws->todo_bd->free;
1365 }
1366 else
1367 {
1368 ws->todo_bd = NULL;
1369 alloc_todo_block(ws,0);
1370 }
1371 }
1372
1373 // deal out any more partial blocks to the threads' part_lists
1374 t = 0;
1375 while (gen->blocks && isPartiallyFull(gen->blocks))
1376 {
1377 bd = gen->blocks;
1378 gen->blocks = bd->link;
1379 ws = &gc_threads[t]->gens[g];
1380 bd->link = ws->part_list;
1381 ws->part_list = bd;
1382 ws->n_part_blocks += 1;
1383 bd->u.scan = bd->free;
1384 gen->n_blocks -= 1;
1385 gen->n_words -= bd->free - bd->start;
1386 t++;
1387 if (t == n_gc_threads) t = 0;
1388 }
1389 }
1390
1391 /* -----------------------------------------------------------------------------
1392 Initialise a gc_thread before GC
1393 -------------------------------------------------------------------------- */
1394
1395 static void
1396 init_gc_thread (gc_thread *t)
1397 {
1398 t->static_objects = END_OF_STATIC_LIST;
1399 t->scavenged_static_objects = END_OF_STATIC_LIST;
1400 t->scan_bd = NULL;
1401 t->mut_lists = capabilities[t->thread_index].mut_lists;
1402 t->evac_gen_no = 0;
1403 t->failed_to_evac = rtsFalse;
1404 t->eager_promotion = rtsTrue;
1405 t->thunk_selector_depth = 0;
1406 t->copied = 0;
1407 t->scanned = 0;
1408 t->any_work = 0;
1409 t->no_work = 0;
1410 t->scav_find_work = 0;
1411 }
1412
1413 /* -----------------------------------------------------------------------------
1414 Function we pass to evacuate roots.
1415 -------------------------------------------------------------------------- */
1416
1417 static void
1418 mark_root(void *user USED_IF_THREADS, StgClosure **root)
1419 {
1420 // we stole a register for gct, but this function is called from
1421 // *outside* the GC where the register variable is not in effect,
1422 // so we need to save and restore it here. NB. only call
1423 // mark_root() from the main GC thread, otherwise gct will be
1424 // incorrect.
1425 gc_thread *saved_gct;
1426 saved_gct = gct;
1427 SET_GCT(user);
1428
1429 evacuate(root);
1430
1431 SET_GCT(saved_gct);
1432 }
1433
1434 /* -----------------------------------------------------------------------------
1435 Initialising the static object & mutable lists
1436 -------------------------------------------------------------------------- */
1437
1438 static void
1439 zero_static_object_list(StgClosure* first_static)
1440 {
1441 StgClosure* p;
1442 StgClosure* link;
1443 const StgInfoTable *info;
1444
1445 for (p = first_static; p != END_OF_STATIC_LIST; p = link) {
1446 info = get_itbl(p);
1447 link = *STATIC_LINK(info, p);
1448 *STATIC_LINK(info,p) = NULL;
1449 }
1450 }
1451
1452 /* ----------------------------------------------------------------------------
1453 Reset the sizes of the older generations when we do a major
1454 collection.
1455
1456 CURRENT STRATEGY: make all generations except zero the same size.
1457 We have to stay within the maximum heap size, and leave a certain
1458 percentage of the maximum heap size available to allocate into.
1459 ------------------------------------------------------------------------- */
1460
1461 static void
1462 resize_generations (void)
1463 {
1464 nat g;
1465
1466 if (major_gc && RtsFlags.GcFlags.generations > 1) {
1467 nat live, size, min_alloc, words;
1468 const nat max = RtsFlags.GcFlags.maxHeapSize;
1469 const nat gens = RtsFlags.GcFlags.generations;
1470
1471 // live in the oldest generations
1472 if (oldest_gen->live_estimate != 0) {
1473 words = oldest_gen->live_estimate;
1474 } else {
1475 words = oldest_gen->n_words;
1476 }
1477 live = (words + BLOCK_SIZE_W - 1) / BLOCK_SIZE_W +
1478 oldest_gen->n_large_blocks;
1479
1480 // default max size for all generations except zero
1481 size = stg_max(live * RtsFlags.GcFlags.oldGenFactor,
1482 RtsFlags.GcFlags.minOldGenSize);
1483
1484 if (RtsFlags.GcFlags.heapSizeSuggestionAuto) {
1485 RtsFlags.GcFlags.heapSizeSuggestion = size;
1486 }
1487
1488 // minimum size for generation zero
1489 min_alloc = stg_max((RtsFlags.GcFlags.pcFreeHeap * max) / 200,
1490 RtsFlags.GcFlags.minAllocAreaSize);
1491
1492 // Auto-enable compaction when the residency reaches a
1493 // certain percentage of the maximum heap size (default: 30%).
1494 if (RtsFlags.GcFlags.compact ||
1495 (max > 0 &&
1496 oldest_gen->n_blocks >
1497 (RtsFlags.GcFlags.compactThreshold * max) / 100)) {
1498 oldest_gen->mark = 1;
1499 oldest_gen->compact = 1;
1500 // debugBelch("compaction: on\n", live);
1501 } else {
1502 oldest_gen->mark = 0;
1503 oldest_gen->compact = 0;
1504 // debugBelch("compaction: off\n", live);
1505 }
1506
1507 if (RtsFlags.GcFlags.sweep) {
1508 oldest_gen->mark = 1;
1509 }
1510
1511 // if we're going to go over the maximum heap size, reduce the
1512 // size of the generations accordingly. The calculation is
1513 // different if compaction is turned on, because we don't need
1514 // to double the space required to collect the old generation.
1515 if (max != 0) {
1516
1517 // this test is necessary to ensure that the calculations
1518 // below don't have any negative results - we're working
1519 // with unsigned values here.
1520 if (max < min_alloc) {
1521 heapOverflow();
1522 }
1523
1524 if (oldest_gen->compact) {
1525 if ( (size + (size - 1) * (gens - 2) * 2) + min_alloc > max ) {
1526 size = (max - min_alloc) / ((gens - 1) * 2 - 1);
1527 }
1528 } else {
1529 if ( (size * (gens - 1) * 2) + min_alloc > max ) {
1530 size = (max - min_alloc) / ((gens - 1) * 2);
1531 }
1532 }
1533
1534 if (size < live) {
1535 heapOverflow();
1536 }
1537 }
1538
1539 #if 0
1540 debugBelch("live: %d, min_alloc: %d, size : %d, max = %d\n", live,
1541 min_alloc, size, max);
1542 #endif
1543
1544 for (g = 0; g < gens; g++) {
1545 generations[g].max_blocks = size;
1546 }
1547 }
1548 }
1549
1550 /* -----------------------------------------------------------------------------
1551 Calculate the new size of the nursery, and resize it.
1552 -------------------------------------------------------------------------- */
1553
1554 static void
1555 resize_nursery (void)
1556 {
1557 const lnat min_nursery = RtsFlags.GcFlags.minAllocAreaSize * n_capabilities;
1558
1559 if (RtsFlags.GcFlags.generations == 1)
1560 { // Two-space collector:
1561 nat blocks;
1562
1563 /* set up a new nursery. Allocate a nursery size based on a
1564 * function of the amount of live data (by default a factor of 2)
1565 * Use the blocks from the old nursery if possible, freeing up any
1566 * left over blocks.
1567 *
1568 * If we get near the maximum heap size, then adjust our nursery
1569 * size accordingly. If the nursery is the same size as the live
1570 * data (L), then we need 3L bytes. We can reduce the size of the
1571 * nursery to bring the required memory down near 2L bytes.
1572 *
1573 * A normal 2-space collector would need 4L bytes to give the same
1574 * performance we get from 3L bytes, reducing to the same
1575 * performance at 2L bytes.
1576 */
1577 blocks = generations[0].n_blocks;
1578
1579 if ( RtsFlags.GcFlags.maxHeapSize != 0 &&
1580 blocks * RtsFlags.GcFlags.oldGenFactor * 2 >
1581 RtsFlags.GcFlags.maxHeapSize )
1582 {
1583 long adjusted_blocks; // signed on purpose
1584 int pc_free;
1585
1586 adjusted_blocks = (RtsFlags.GcFlags.maxHeapSize - 2 * blocks);
1587
1588 debugTrace(DEBUG_gc, "near maximum heap size of 0x%x blocks, blocks = %d, adjusted to %ld",
1589 RtsFlags.GcFlags.maxHeapSize, blocks, adjusted_blocks);
1590
1591 pc_free = adjusted_blocks * 100 / RtsFlags.GcFlags.maxHeapSize;
1592 if (pc_free < RtsFlags.GcFlags.pcFreeHeap) /* might even * be < 0 */
1593 {
1594 heapOverflow();
1595 }
1596 blocks = adjusted_blocks;
1597 }
1598 else
1599 {
1600 blocks *= RtsFlags.GcFlags.oldGenFactor;
1601 if (blocks < min_nursery)
1602 {
1603 blocks = min_nursery;
1604 }
1605 }
1606 resizeNurseries(blocks);
1607 }
1608 else // Generational collector
1609 {
1610 /*
1611 * If the user has given us a suggested heap size, adjust our
1612 * allocation area to make best use of the memory available.
1613 */
1614 if (RtsFlags.GcFlags.heapSizeSuggestion)
1615 {
1616 long blocks;
1617 const nat needed = calcNeeded(); // approx blocks needed at next GC
1618
1619 /* Guess how much will be live in generation 0 step 0 next time.
1620 * A good approximation is obtained by finding the
1621 * percentage of g0 that was live at the last minor GC.
1622 *
1623 * We have an accurate figure for the amount of copied data in
1624 * 'copied', but we must convert this to a number of blocks, with
1625 * a small adjustment for estimated slop at the end of a block
1626 * (- 10 words).
1627 */
1628 if (N == 0)
1629 {
1630 g0_pcnt_kept = ((copied / (BLOCK_SIZE_W - 10)) * 100)
1631 / countNurseryBlocks();
1632 }
1633
1634 /* Estimate a size for the allocation area based on the
1635 * information available. We might end up going slightly under
1636 * or over the suggested heap size, but we should be pretty
1637 * close on average.
1638 *
1639 * Formula: suggested - needed
1640 * ----------------------------
1641 * 1 + g0_pcnt_kept/100
1642 *
1643 * where 'needed' is the amount of memory needed at the next
1644 * collection for collecting all gens except g0.
1645 */
1646 blocks =
1647 (((long)RtsFlags.GcFlags.heapSizeSuggestion - (long)needed) * 100) /
1648 (100 + (long)g0_pcnt_kept);
1649
1650 if (blocks < (long)min_nursery) {
1651 blocks = min_nursery;
1652 }
1653
1654 resizeNurseries((nat)blocks);
1655 }
1656 else
1657 {
1658 // we might have added extra large blocks to the nursery, so
1659 // resize back to minAllocAreaSize again.
1660 resizeNurseriesFixed(RtsFlags.GcFlags.minAllocAreaSize);
1661 }
1662 }
1663 }
1664
1665 /* -----------------------------------------------------------------------------
1666 Sanity code for CAF garbage collection.
1667
1668 With DEBUG turned on, we manage a CAF list in addition to the SRT
1669 mechanism. After GC, we run down the CAF list and blackhole any
1670 CAFs which have been garbage collected. This means we get an error
1671 whenever the program tries to enter a garbage collected CAF.
1672
1673 Any garbage collected CAFs are taken off the CAF list at the same
1674 time.
1675 -------------------------------------------------------------------------- */
1676
1677 #if 0 && defined(DEBUG)
1678
1679 static void
1680 gcCAFs(void)
1681 {
1682 StgClosure* p;
1683 StgClosure** pp;
1684 const StgInfoTable *info;
1685 nat i;
1686
1687 i = 0;
1688 p = caf_list;
1689 pp = &caf_list;
1690
1691 while (p != NULL) {
1692
1693 info = get_itbl(p);
1694
1695 ASSERT(info->type == IND_STATIC);
1696
1697 if (STATIC_LINK(info,p) == NULL) {
1698 debugTrace(DEBUG_gccafs, "CAF gc'd at 0x%04lx", (long)p);
1699 // black hole it
1700 SET_INFO(p,&stg_BLACKHOLE_info);
1701 p = STATIC_LINK2(info,p);
1702 *pp = p;
1703 }
1704 else {
1705 pp = &STATIC_LINK2(info,p);
1706 p = *pp;
1707 i++;
1708 }
1709
1710 }
1711
1712 debugTrace(DEBUG_gccafs, "%d CAFs live", i);
1713 }
1714 #endif