Reorganisation to fix problems related to the gct register variable
[ghc.git] / rts / sm / GC.c
1 /* -----------------------------------------------------------------------------
2 *
3 * (c) The GHC Team 1998-2006
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 "RtsFlags.h"
17 #include "RtsUtils.h"
18 #include "Apply.h"
19 #include "OSThreads.h"
20 #include "LdvProfile.h"
21 #include "Updates.h"
22 #include "Stats.h"
23 #include "Schedule.h"
24 #include "Sanity.h"
25 #include "BlockAlloc.h"
26 #include "MBlock.h"
27 #include "ProfHeap.h"
28 #include "SchedAPI.h"
29 #include "Weak.h"
30 #include "Prelude.h"
31 #include "ParTicky.h" // ToDo: move into Rts.h
32 #include "RtsSignals.h"
33 #include "STM.h"
34 #include "HsFFI.h"
35 #include "Linker.h"
36 #if defined(RTS_GTK_FRONTPANEL)
37 #include "FrontPanel.h"
38 #endif
39 #include "Trace.h"
40 #include "RetainerProfile.h"
41 #include "RaiseAsync.h"
42 #include "Sparks.h"
43 #include "Papi.h"
44
45 #include "GC.h"
46 #include "GCThread.h"
47 #include "Compact.h"
48 #include "Evac.h"
49 #include "Scav.h"
50 #include "GCUtils.h"
51 #include "MarkWeak.h"
52 #include "Sparks.h"
53
54 #include <string.h> // for memset()
55 #include <unistd.h>
56
57 /* -----------------------------------------------------------------------------
58 Global variables
59 -------------------------------------------------------------------------- */
60
61 /* STATIC OBJECT LIST.
62 *
63 * During GC:
64 * We maintain a linked list of static objects that are still live.
65 * The requirements for this list are:
66 *
67 * - we need to scan the list while adding to it, in order to
68 * scavenge all the static objects (in the same way that
69 * breadth-first scavenging works for dynamic objects).
70 *
71 * - we need to be able to tell whether an object is already on
72 * the list, to break loops.
73 *
74 * Each static object has a "static link field", which we use for
75 * linking objects on to the list. We use a stack-type list, consing
76 * objects on the front as they are added (this means that the
77 * scavenge phase is depth-first, not breadth-first, but that
78 * shouldn't matter).
79 *
80 * A separate list is kept for objects that have been scavenged
81 * already - this is so that we can zero all the marks afterwards.
82 *
83 * An object is on the list if its static link field is non-zero; this
84 * means that we have to mark the end of the list with '1', not NULL.
85 *
86 * Extra notes for generational GC:
87 *
88 * Each generation has a static object list associated with it. When
89 * collecting generations up to N, we treat the static object lists
90 * from generations > N as roots.
91 *
92 * We build up a static object list while collecting generations 0..N,
93 * which is then appended to the static object list of generation N+1.
94 */
95
96 /* N is the oldest generation being collected, where the generations
97 * are numbered starting at 0. A major GC (indicated by the major_gc
98 * flag) is when we're collecting all generations. We only attempt to
99 * deal with static objects and GC CAFs when doing a major GC.
100 */
101 nat N;
102 rtsBool major_gc;
103
104 /* Data used for allocation area sizing.
105 */
106 static lnat g0s0_pcnt_kept = 30; // percentage of g0s0 live at last minor GC
107
108 /* Mut-list stats */
109 #ifdef DEBUG
110 nat mutlist_MUTVARS,
111 mutlist_MUTARRS,
112 mutlist_MVARS,
113 mutlist_OTHERS;
114 #endif
115
116 /* Thread-local data for each GC thread
117 */
118 gc_thread **gc_threads = NULL;
119 // gc_thread *gct = NULL; // this thread's gct TODO: make thread-local
120
121 // Number of threads running in *this* GC. Affects how many
122 // step->todos[] lists we have to look in to find work.
123 nat n_gc_threads;
124
125 // For stats:
126 long copied; // *words* copied & scavenged during this GC
127
128 #ifdef THREADED_RTS
129 SpinLock recordMutableGen_sync;
130 #endif
131
132 /* -----------------------------------------------------------------------------
133 Static function declarations
134 -------------------------------------------------------------------------- */
135
136 static void mark_root (void *user, StgClosure **root);
137 static void zero_static_object_list (StgClosure* first_static);
138 static nat initialise_N (rtsBool force_major_gc);
139 static void alloc_gc_threads (void);
140 static void init_collected_gen (nat g, nat threads);
141 static void init_uncollected_gen (nat g, nat threads);
142 static void init_gc_thread (gc_thread *t);
143 static void update_task_list (void);
144 static void resize_generations (void);
145 static void resize_nursery (void);
146 static void start_gc_threads (void);
147 static void gc_thread_work (void);
148 static nat inc_running (void);
149 static nat dec_running (void);
150 static void wakeup_gc_threads (nat n_threads);
151 static void shutdown_gc_threads (nat n_threads);
152
153 #if 0 && defined(DEBUG)
154 static void gcCAFs (void);
155 #endif
156
157 /* -----------------------------------------------------------------------------
158 The mark bitmap & stack.
159 -------------------------------------------------------------------------- */
160
161 #define MARK_STACK_BLOCKS 4
162
163 bdescr *mark_stack_bdescr;
164 StgPtr *mark_stack;
165 StgPtr *mark_sp;
166 StgPtr *mark_splim;
167
168 // Flag and pointers used for falling back to a linear scan when the
169 // mark stack overflows.
170 rtsBool mark_stack_overflowed;
171 bdescr *oldgen_scan_bd;
172 StgPtr oldgen_scan;
173
174 /* -----------------------------------------------------------------------------
175 GarbageCollect: the main entry point to the garbage collector.
176
177 Locks held: all capabilities are held throughout GarbageCollect().
178 -------------------------------------------------------------------------- */
179
180 void
181 GarbageCollect ( rtsBool force_major_gc )
182 {
183 bdescr *bd;
184 step *stp;
185 lnat live, allocated, max_copied, avg_copied, slop;
186 lnat oldgen_saved_blocks = 0;
187 gc_thread *saved_gct;
188 nat g, s, t, n;
189
190 // necessary if we stole a callee-saves register for gct:
191 saved_gct = gct;
192
193 #ifdef PROFILING
194 CostCentreStack *prev_CCS;
195 #endif
196
197 ACQUIRE_SM_LOCK;
198
199 #if defined(RTS_USER_SIGNALS)
200 if (RtsFlags.MiscFlags.install_signal_handlers) {
201 // block signals
202 blockUserSignals();
203 }
204 #endif
205
206 // tell the stats department that we've started a GC
207 stat_startGC();
208
209 // tell the STM to discard any cached closures it's hoping to re-use
210 stmPreGCHook();
211
212 #ifdef DEBUG
213 mutlist_MUTVARS = 0;
214 mutlist_MUTARRS = 0;
215 mutlist_OTHERS = 0;
216 #endif
217
218 // attribute any costs to CCS_GC
219 #ifdef PROFILING
220 prev_CCS = CCCS;
221 CCCS = CCS_GC;
222 #endif
223
224 /* Approximate how much we allocated.
225 * Todo: only when generating stats?
226 */
227 allocated = calcAllocated();
228
229 /* Figure out which generation to collect
230 */
231 n = initialise_N(force_major_gc);
232
233 /* Allocate + initialise the gc_thread structures.
234 */
235 alloc_gc_threads();
236
237 /* Start threads, so they can be spinning up while we finish initialisation.
238 */
239 start_gc_threads();
240
241 /* How many threads will be participating in this GC?
242 * We don't try to parallelise minor GC.
243 */
244 #if defined(THREADED_RTS)
245 if (n < (4*1024*1024 / BLOCK_SIZE)) {
246 n_gc_threads = 1;
247 } else {
248 n_gc_threads = RtsFlags.ParFlags.gcThreads;
249 }
250 #else
251 n_gc_threads = 1;
252 #endif
253 trace(TRACE_gc|DEBUG_gc, "GC (gen %d): %dKB to collect, using %d thread(s)",
254 N, n * (BLOCK_SIZE / 1024), n_gc_threads);
255
256 #ifdef RTS_GTK_FRONTPANEL
257 if (RtsFlags.GcFlags.frontpanel) {
258 updateFrontPanelBeforeGC(N);
259 }
260 #endif
261
262 #ifdef DEBUG
263 // check for memory leaks if DEBUG is on
264 memInventory(traceClass(DEBUG_gc));
265 #endif
266
267 // check stack sanity *before* GC (ToDo: check all threads)
268 IF_DEBUG(sanity, checkFreeListSanity());
269
270 // Initialise all our gc_thread structures
271 for (t = 0; t < n_gc_threads; t++) {
272 init_gc_thread(gc_threads[t]);
273 }
274
275 // Initialise all the generations/steps that we're collecting.
276 for (g = 0; g <= N; g++) {
277 init_collected_gen(g,n_gc_threads);
278 }
279
280 // Initialise all the generations/steps that we're *not* collecting.
281 for (g = N+1; g < RtsFlags.GcFlags.generations; g++) {
282 init_uncollected_gen(g,n_gc_threads);
283 }
284
285 /* Allocate a mark stack if we're doing a major collection.
286 */
287 if (major_gc) {
288 mark_stack_bdescr = allocGroup(MARK_STACK_BLOCKS);
289 mark_stack = (StgPtr *)mark_stack_bdescr->start;
290 mark_sp = mark_stack;
291 mark_splim = mark_stack + (MARK_STACK_BLOCKS * BLOCK_SIZE_W);
292 } else {
293 mark_stack_bdescr = NULL;
294 }
295
296 // this is the main thread
297 gct = gc_threads[0];
298
299 /* -----------------------------------------------------------------------
300 * follow all the roots that we know about:
301 * - mutable lists from each generation > N
302 * we want to *scavenge* these roots, not evacuate them: they're not
303 * going to move in this GC.
304 * Also do them in reverse generation order, for the usual reason:
305 * namely to reduce the likelihood of spurious old->new pointers.
306 */
307 for (g = RtsFlags.GcFlags.generations-1; g > N; g--) {
308 generations[g].saved_mut_list = generations[g].mut_list;
309 generations[g].mut_list = allocBlock();
310 // mut_list always has at least one block.
311 }
312
313 // the main thread is running: this prevents any other threads from
314 // exiting prematurely, so we can start them now.
315 // NB. do this after the mutable lists have been saved above, otherwise
316 // the other GC threads will be writing into the old mutable lists.
317 inc_running();
318 wakeup_gc_threads(n_gc_threads);
319
320 for (g = RtsFlags.GcFlags.generations-1; g > N; g--) {
321 scavenge_mutable_list(&generations[g]);
322 }
323
324 // follow roots from the CAF list (used by GHCi)
325 gct->evac_step = 0;
326 markCAFs(mark_root, gct);
327
328 // follow all the roots that the application knows about.
329 gct->evac_step = 0;
330 markSomeCapabilities(mark_root, gct, gct->thread_index, n_gc_threads);
331
332 #if defined(RTS_USER_SIGNALS)
333 // mark the signal handlers (signals should be already blocked)
334 markSignalHandlers(mark_root, gct);
335 #endif
336
337 // Mark the weak pointer list, and prepare to detect dead weak pointers.
338 markWeakPtrList();
339 initWeakForGC();
340
341 // Mark the stable pointer table.
342 markStablePtrTable(mark_root, gct);
343
344 /* -------------------------------------------------------------------------
345 * Repeatedly scavenge all the areas we know about until there's no
346 * more scavenging to be done.
347 */
348 for (;;)
349 {
350 gc_thread_work();
351 // The other threads are now stopped. We might recurse back to
352 // here, but from now on this is the only thread.
353
354 // if any blackholes are alive, make the threads that wait on
355 // them alive too.
356 if (traverseBlackholeQueue()) {
357 inc_running();
358 continue;
359 }
360
361 // must be last... invariant is that everything is fully
362 // scavenged at this point.
363 if (traverseWeakPtrList()) { // returns rtsTrue if evaced something
364 inc_running();
365 continue;
366 }
367
368 // If we get to here, there's really nothing left to do.
369 break;
370 }
371
372 shutdown_gc_threads(n_gc_threads);
373
374 // Update pointers from the Task list
375 update_task_list();
376
377 // Now see which stable names are still alive.
378 gcStablePtrTable();
379
380 #ifdef PROFILING
381 // We call processHeapClosureForDead() on every closure destroyed during
382 // the current garbage collection, so we invoke LdvCensusForDead().
383 if (RtsFlags.ProfFlags.doHeapProfile == HEAP_BY_LDV
384 || RtsFlags.ProfFlags.bioSelector != NULL)
385 LdvCensusForDead(N);
386 #endif
387
388 // NO MORE EVACUATION AFTER THIS POINT!
389 // Finally: compaction of the oldest generation.
390 if (major_gc && oldest_gen->steps[0].is_compacted) {
391 // save number of blocks for stats
392 oldgen_saved_blocks = oldest_gen->steps[0].n_old_blocks;
393 compact(gct->scavenged_static_objects);
394 }
395
396 IF_DEBUG(sanity, checkGlobalTSOList(rtsFalse));
397
398 // Two-space collector: free the old to-space.
399 // g0s0->old_blocks is the old nursery
400 // g0s0->blocks is to-space from the previous GC
401 if (RtsFlags.GcFlags.generations == 1) {
402 if (g0s0->blocks != NULL) {
403 freeChain(g0s0->blocks);
404 g0s0->blocks = NULL;
405 }
406 }
407
408 // For each workspace, in each thread:
409 // * clear the BF_EVACUATED flag from each copied block
410 // * move the copied blocks to the step
411 {
412 gc_thread *thr;
413 step_workspace *ws;
414 bdescr *prev, *next;
415
416 for (t = 0; t < n_gc_threads; t++) {
417 thr = gc_threads[t];
418
419 // not step 0
420 for (s = 1; s < total_steps; s++) {
421 ws = &thr->steps[s];
422
423 // Push the final block
424 if (ws->todo_bd) {
425 push_scanned_block(ws->todo_bd, ws);
426 }
427
428 ASSERT(gct->scan_bd == NULL);
429 ASSERT(countBlocks(ws->scavd_list) == ws->n_scavd_blocks);
430
431 prev = NULL;
432 for (bd = ws->scavd_list; bd != NULL; bd = bd->link) {
433 bd->flags &= ~BF_EVACUATED; // now from-space
434 ws->step->n_words += bd->free - bd->start;
435 prev = bd;
436 }
437 if (prev != NULL) {
438 prev->link = ws->step->blocks;
439 ws->step->blocks = ws->scavd_list;
440 }
441 ws->step->n_blocks += ws->n_scavd_blocks;
442
443 prev = NULL;
444 for (bd = ws->part_list; bd != NULL; bd = next) {
445 next = bd->link;
446 if (bd->free == bd->start) {
447 if (prev == NULL) {
448 ws->part_list = next;
449 } else {
450 prev->link = next;
451 }
452 freeGroup(bd);
453 ws->n_part_blocks--;
454 } else {
455 bd->flags &= ~BF_EVACUATED; // now from-space
456 ws->step->n_words += bd->free - bd->start;
457 prev = bd;
458 }
459 }
460 if (prev != NULL) {
461 prev->link = ws->step->blocks;
462 ws->step->blocks = ws->part_list;
463 }
464 ws->step->n_blocks += ws->n_part_blocks;
465
466 ASSERT(countBlocks(ws->step->blocks) == ws->step->n_blocks);
467 ASSERT(countOccupied(ws->step->blocks) == ws->step->n_words);
468 }
469 }
470 }
471
472 // Two-space collector: swap the semi-spaces around.
473 // Currently: g0s0->old_blocks is the old nursery
474 // g0s0->blocks is to-space from this GC
475 // We want these the other way around.
476 if (RtsFlags.GcFlags.generations == 1) {
477 bdescr *nursery_blocks = g0s0->old_blocks;
478 nat n_nursery_blocks = g0s0->n_old_blocks;
479 g0s0->old_blocks = g0s0->blocks;
480 g0s0->n_old_blocks = g0s0->n_blocks;
481 g0s0->blocks = nursery_blocks;
482 g0s0->n_blocks = n_nursery_blocks;
483 }
484
485 /* run through all the generations/steps and tidy up
486 */
487 copied = 0;
488 max_copied = 0;
489 avg_copied = 0;
490 {
491 nat i;
492 for (i=0; i < n_gc_threads; i++) {
493 if (n_gc_threads > 1) {
494 trace(TRACE_gc,"thread %d:", i);
495 trace(TRACE_gc," copied %ld", gc_threads[i]->copied * sizeof(W_));
496 trace(TRACE_gc," scanned %ld", gc_threads[i]->scanned * sizeof(W_));
497 trace(TRACE_gc," any_work %ld", gc_threads[i]->any_work);
498 trace(TRACE_gc," no_work %ld", gc_threads[i]->no_work);
499 trace(TRACE_gc," scav_find_work %ld", gc_threads[i]->scav_find_work);
500 }
501 copied += gc_threads[i]->copied;
502 max_copied = stg_max(gc_threads[i]->copied, max_copied);
503 }
504 if (n_gc_threads == 1) {
505 max_copied = 0;
506 avg_copied = 0;
507 } else {
508 avg_copied = copied;
509 }
510 }
511
512 for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
513
514 if (g == N) {
515 generations[g].collections++; // for stats
516 if (n_gc_threads > 1) generations[g].par_collections++;
517 }
518
519 // Count the mutable list as bytes "copied" for the purposes of
520 // stats. Every mutable list is copied during every GC.
521 if (g > 0) {
522 nat mut_list_size = 0;
523 for (bd = generations[g].mut_list; bd != NULL; bd = bd->link) {
524 mut_list_size += bd->free - bd->start;
525 }
526 copied += mut_list_size;
527
528 debugTrace(DEBUG_gc,
529 "mut_list_size: %lu (%d vars, %d arrays, %d MVARs, %d others)",
530 (unsigned long)(mut_list_size * sizeof(W_)),
531 mutlist_MUTVARS, mutlist_MUTARRS, mutlist_MVARS, mutlist_OTHERS);
532 }
533
534 for (s = 0; s < generations[g].n_steps; s++) {
535 bdescr *next;
536 stp = &generations[g].steps[s];
537
538 // for generations we collected...
539 if (g <= N) {
540
541 /* free old memory and shift to-space into from-space for all
542 * the collected steps (except the allocation area). These
543 * freed blocks will probaby be quickly recycled.
544 */
545 if (!(g == 0 && s == 0 && RtsFlags.GcFlags.generations > 1)) {
546 if (stp->is_compacted)
547 {
548 // for a compacted step, just shift the new to-space
549 // onto the front of the now-compacted existing blocks.
550 for (bd = stp->blocks; bd != NULL; bd = bd->link) {
551 bd->flags &= ~BF_EVACUATED; // now from-space
552 stp->n_words += bd->free - bd->start;
553 }
554 // tack the new blocks on the end of the existing blocks
555 if (stp->old_blocks != NULL) {
556 for (bd = stp->old_blocks; bd != NULL; bd = next) {
557 // NB. this step might not be compacted next
558 // time, so reset the BF_COMPACTED flags.
559 // They are set before GC if we're going to
560 // compact. (search for BF_COMPACTED above).
561 bd->flags &= ~BF_COMPACTED;
562 next = bd->link;
563 if (next == NULL) {
564 bd->link = stp->blocks;
565 }
566 }
567 stp->blocks = stp->old_blocks;
568 }
569 // add the new blocks to the block tally
570 stp->n_blocks += stp->n_old_blocks;
571 ASSERT(countBlocks(stp->blocks) == stp->n_blocks);
572 ASSERT(countOccupied(stp->blocks) == stp->n_words);
573 }
574 else // not copacted
575 {
576 freeChain(stp->old_blocks);
577 }
578 stp->old_blocks = NULL;
579 stp->n_old_blocks = 0;
580 }
581
582 /* LARGE OBJECTS. The current live large objects are chained on
583 * scavenged_large, having been moved during garbage
584 * collection from large_objects. Any objects left on
585 * large_objects list are therefore dead, so we free them here.
586 */
587 for (bd = stp->large_objects; bd != NULL; bd = next) {
588 next = bd->link;
589 freeGroup(bd);
590 bd = next;
591 }
592
593 // update the count of blocks used by large objects
594 for (bd = stp->scavenged_large_objects; bd != NULL; bd = bd->link) {
595 bd->flags &= ~BF_EVACUATED;
596 }
597 stp->large_objects = stp->scavenged_large_objects;
598 stp->n_large_blocks = stp->n_scavenged_large_blocks;
599
600 }
601 else // for older generations...
602 {
603 /* For older generations, we need to append the
604 * scavenged_large_object list (i.e. large objects that have been
605 * promoted during this GC) to the large_object list for that step.
606 */
607 for (bd = stp->scavenged_large_objects; bd; bd = next) {
608 next = bd->link;
609 bd->flags &= ~BF_EVACUATED;
610 dbl_link_onto(bd, &stp->large_objects);
611 }
612
613 // add the new blocks we promoted during this GC
614 stp->n_large_blocks += stp->n_scavenged_large_blocks;
615 }
616 }
617 }
618
619 // update the max size of older generations after a major GC
620 resize_generations();
621
622 // Calculate the amount of live data for stats.
623 live = calcLiveWords();
624
625 // Free the small objects allocated via allocate(), since this will
626 // all have been copied into G0S1 now.
627 if (RtsFlags.GcFlags.generations > 1) {
628 if (g0s0->blocks != NULL) {
629 freeChain(g0s0->blocks);
630 g0s0->blocks = NULL;
631 }
632 g0s0->n_blocks = 0;
633 g0s0->n_words = 0;
634 }
635 alloc_blocks = 0;
636 alloc_blocks_lim = RtsFlags.GcFlags.minAllocAreaSize;
637
638 // Start a new pinned_object_block
639 pinned_object_block = NULL;
640
641 // Free the mark stack.
642 if (mark_stack_bdescr != NULL) {
643 freeGroup(mark_stack_bdescr);
644 }
645
646 // Free any bitmaps.
647 for (g = 0; g <= N; g++) {
648 for (s = 0; s < generations[g].n_steps; s++) {
649 stp = &generations[g].steps[s];
650 if (stp->bitmap != NULL) {
651 freeGroup(stp->bitmap);
652 stp->bitmap = NULL;
653 }
654 }
655 }
656
657 resize_nursery();
658
659 // mark the garbage collected CAFs as dead
660 #if 0 && defined(DEBUG) // doesn't work at the moment
661 if (major_gc) { gcCAFs(); }
662 #endif
663
664 #ifdef PROFILING
665 // resetStaticObjectForRetainerProfiling() must be called before
666 // zeroing below.
667 if (n_gc_threads > 1) {
668 barf("profiling is currently broken with multi-threaded GC");
669 // ToDo: fix the gct->scavenged_static_objects below
670 }
671 resetStaticObjectForRetainerProfiling(gct->scavenged_static_objects);
672 #endif
673
674 // zero the scavenged static object list
675 if (major_gc) {
676 nat i;
677 for (i = 0; i < n_gc_threads; i++) {
678 zero_static_object_list(gc_threads[i]->scavenged_static_objects);
679 }
680 }
681
682 // Reset the nursery
683 resetNurseries();
684
685 // start any pending finalizers
686 RELEASE_SM_LOCK;
687 scheduleFinalizers(last_free_capability, old_weak_ptr_list);
688 ACQUIRE_SM_LOCK;
689
690 // send exceptions to any threads which were about to die
691 RELEASE_SM_LOCK;
692 resurrectThreads(resurrected_threads);
693 ACQUIRE_SM_LOCK;
694
695 // Update the stable pointer hash table.
696 updateStablePtrTable(major_gc);
697
698 // check sanity after GC
699 IF_DEBUG(sanity, checkSanity());
700
701 // extra GC trace info
702 if (traceClass(TRACE_gc|DEBUG_gc)) statDescribeGens();
703
704 #ifdef DEBUG
705 // symbol-table based profiling
706 /* heapCensus(to_blocks); */ /* ToDo */
707 #endif
708
709 // restore enclosing cost centre
710 #ifdef PROFILING
711 CCCS = prev_CCS;
712 #endif
713
714 #ifdef DEBUG
715 // check for memory leaks if DEBUG is on
716 memInventory(traceClass(DEBUG_gc));
717 #endif
718
719 #ifdef RTS_GTK_FRONTPANEL
720 if (RtsFlags.GcFlags.frontpanel) {
721 updateFrontPanelAfterGC( N, live );
722 }
723 #endif
724
725 // ok, GC over: tell the stats department what happened.
726 slop = calcLiveBlocks() * BLOCK_SIZE_W - live;
727 stat_endGC(allocated, live, copied, N, max_copied, avg_copied, slop);
728
729 #if defined(RTS_USER_SIGNALS)
730 if (RtsFlags.MiscFlags.install_signal_handlers) {
731 // unblock signals again
732 unblockUserSignals();
733 }
734 #endif
735
736 RELEASE_SM_LOCK;
737
738 gct = saved_gct;
739 }
740
741 /* -----------------------------------------------------------------------------
742 Figure out which generation to collect, initialise N and major_gc.
743
744 Also returns the total number of blocks in generations that will be
745 collected.
746 -------------------------------------------------------------------------- */
747
748 static nat
749 initialise_N (rtsBool force_major_gc)
750 {
751 int g;
752 nat s, blocks, blocks_total;
753
754 blocks = 0;
755 blocks_total = 0;
756
757 if (force_major_gc) {
758 N = RtsFlags.GcFlags.generations - 1;
759 } else {
760 N = 0;
761 }
762
763 for (g = RtsFlags.GcFlags.generations - 1; g >= 0; g--) {
764 blocks = 0;
765 for (s = 0; s < generations[g].n_steps; s++) {
766 blocks += generations[g].steps[s].n_words / BLOCK_SIZE_W;
767 blocks += generations[g].steps[s].n_large_blocks;
768 }
769 if (blocks >= generations[g].max_blocks) {
770 N = stg_max(N,g);
771 }
772 if ((nat)g <= N) {
773 blocks_total += blocks;
774 }
775 }
776
777 blocks_total += countNurseryBlocks();
778
779 major_gc = (N == RtsFlags.GcFlags.generations-1);
780 return blocks_total;
781 }
782
783 /* -----------------------------------------------------------------------------
784 Initialise the gc_thread structures.
785 -------------------------------------------------------------------------- */
786
787 static gc_thread *
788 alloc_gc_thread (int n)
789 {
790 nat s;
791 step_workspace *ws;
792 gc_thread *t;
793
794 t = stgMallocBytes(sizeof(gc_thread) + total_steps * sizeof(step_workspace),
795 "alloc_gc_thread");
796
797 #ifdef THREADED_RTS
798 t->id = 0;
799 initCondition(&t->wake_cond);
800 initMutex(&t->wake_mutex);
801 t->wakeup = rtsTrue; // starts true, so we can wait for the
802 // thread to start up, see wakeup_gc_threads
803 t->exit = rtsFalse;
804 #endif
805
806 t->thread_index = n;
807 t->free_blocks = NULL;
808 t->gc_count = 0;
809
810 init_gc_thread(t);
811
812 #ifdef USE_PAPI
813 t->papi_events = -1;
814 #endif
815
816 for (s = 0; s < total_steps; s++)
817 {
818 ws = &t->steps[s];
819 ws->step = &all_steps[s];
820 ASSERT(s == ws->step->abs_no);
821 ws->gct = t;
822
823 ws->todo_bd = NULL;
824 ws->buffer_todo_bd = NULL;
825
826 ws->part_list = NULL;
827 ws->n_part_blocks = 0;
828
829 ws->scavd_list = NULL;
830 ws->n_scavd_blocks = 0;
831 }
832
833 return t;
834 }
835
836
837 static void
838 alloc_gc_threads (void)
839 {
840 if (gc_threads == NULL) {
841 #if defined(THREADED_RTS)
842 nat i;
843 gc_threads = stgMallocBytes (RtsFlags.ParFlags.gcThreads *
844 sizeof(gc_thread*),
845 "alloc_gc_threads");
846
847 for (i = 0; i < RtsFlags.ParFlags.gcThreads; i++) {
848 gc_threads[i] = alloc_gc_thread(i);
849 }
850 #else
851 gc_threads = stgMallocBytes (sizeof(gc_thread*),
852 "alloc_gc_threads");
853
854 gc_threads[0] = alloc_gc_thread(0);
855 #endif
856 }
857 }
858
859 /* ----------------------------------------------------------------------------
860 Start GC threads
861 ------------------------------------------------------------------------- */
862
863 static nat gc_running_threads;
864
865 #if defined(THREADED_RTS)
866 static Mutex gc_running_mutex;
867 #endif
868
869 static nat
870 inc_running (void)
871 {
872 nat n_running;
873 ACQUIRE_LOCK(&gc_running_mutex);
874 n_running = ++gc_running_threads;
875 RELEASE_LOCK(&gc_running_mutex);
876 ASSERT(n_running <= n_gc_threads);
877 return n_running;
878 }
879
880 static nat
881 dec_running (void)
882 {
883 nat n_running;
884 ACQUIRE_LOCK(&gc_running_mutex);
885 ASSERT(n_gc_threads != 0);
886 n_running = --gc_running_threads;
887 RELEASE_LOCK(&gc_running_mutex);
888 return n_running;
889 }
890
891 //
892 // gc_thread_work(): Scavenge until there's no work left to do and all
893 // the running threads are idle.
894 //
895 static void
896 gc_thread_work (void)
897 {
898 nat r;
899
900 debugTrace(DEBUG_gc, "GC thread %d working", gct->thread_index);
901
902 // gc_running_threads has already been incremented for us; either
903 // this is the main thread and we incremented it inside
904 // GarbageCollect(), or this is a worker thread and the main
905 // thread bumped gc_running_threads before waking us up.
906
907 // Every thread evacuates some roots.
908 gct->evac_step = 0;
909 markSomeCapabilities(mark_root, gct, gct->thread_index, n_gc_threads);
910
911 loop:
912 scavenge_loop();
913 // scavenge_loop() only exits when there's no work to do
914 r = dec_running();
915
916 debugTrace(DEBUG_gc, "GC thread %d idle (%d still running)",
917 gct->thread_index, r);
918
919 while (gc_running_threads != 0) {
920 usleep(1);
921 if (any_work()) {
922 inc_running();
923 goto loop;
924 }
925 // any_work() does not remove the work from the queue, it
926 // just checks for the presence of work. If we find any,
927 // then we increment gc_running_threads and go back to
928 // scavenge_loop() to perform any pending work.
929 }
930
931 // All threads are now stopped
932 debugTrace(DEBUG_gc, "GC thread %d finished.", gct->thread_index);
933 }
934
935
936 #if defined(THREADED_RTS)
937 static void
938 gc_thread_mainloop (void)
939 {
940 while (!gct->exit) {
941
942 // Wait until we're told to wake up
943 ACQUIRE_LOCK(&gct->wake_mutex);
944 gct->wakeup = rtsFalse;
945 while (!gct->wakeup) {
946 debugTrace(DEBUG_gc, "GC thread %d standing by...",
947 gct->thread_index);
948 waitCondition(&gct->wake_cond, &gct->wake_mutex);
949 }
950 RELEASE_LOCK(&gct->wake_mutex);
951 if (gct->exit) break;
952
953 #ifdef USE_PAPI
954 // start performance counters in this thread...
955 if (gct->papi_events == -1) {
956 papi_init_eventset(&gct->papi_events);
957 }
958 papi_thread_start_gc1_count(gct->papi_events);
959 #endif
960
961 gc_thread_work();
962
963 #ifdef USE_PAPI
964 // count events in this thread towards the GC totals
965 papi_thread_stop_gc1_count(gct->papi_events);
966 #endif
967 }
968 }
969 #endif
970
971 #if defined(THREADED_RTS)
972 static void
973 gc_thread_entry (gc_thread *my_gct)
974 {
975 gct = my_gct;
976 debugTrace(DEBUG_gc, "GC thread %d starting...", gct->thread_index);
977 gct->id = osThreadId();
978 gc_thread_mainloop();
979 }
980 #endif
981
982 static void
983 start_gc_threads (void)
984 {
985 #if defined(THREADED_RTS)
986 nat i;
987 OSThreadId id;
988 static rtsBool done = rtsFalse;
989
990 gc_running_threads = 0;
991 initMutex(&gc_running_mutex);
992
993 if (!done) {
994 // Start from 1: the main thread is 0
995 for (i = 1; i < RtsFlags.ParFlags.gcThreads; i++) {
996 createOSThread(&id, (OSThreadProc*)&gc_thread_entry,
997 gc_threads[i]);
998 }
999 done = rtsTrue;
1000 }
1001 #endif
1002 }
1003
1004 static void
1005 wakeup_gc_threads (nat n_threads USED_IF_THREADS)
1006 {
1007 #if defined(THREADED_RTS)
1008 nat i;
1009 for (i=1; i < n_threads; i++) {
1010 inc_running();
1011 debugTrace(DEBUG_gc, "waking up gc thread %d", i);
1012 do {
1013 ACQUIRE_LOCK(&gc_threads[i]->wake_mutex);
1014 if (gc_threads[i]->wakeup) {
1015 RELEASE_LOCK(&gc_threads[i]->wake_mutex);
1016 continue;
1017 } else {
1018 break;
1019 }
1020 } while (1);
1021 gc_threads[i]->wakeup = rtsTrue;
1022 signalCondition(&gc_threads[i]->wake_cond);
1023 RELEASE_LOCK(&gc_threads[i]->wake_mutex);
1024 }
1025 #endif
1026 }
1027
1028 // After GC is complete, we must wait for all GC threads to enter the
1029 // standby state, otherwise they may still be executing inside
1030 // any_work(), and may even remain awake until the next GC starts.
1031 static void
1032 shutdown_gc_threads (nat n_threads USED_IF_THREADS)
1033 {
1034 #if defined(THREADED_RTS)
1035 nat i;
1036 rtsBool wakeup;
1037 for (i=1; i < n_threads; i++) {
1038 do {
1039 ACQUIRE_LOCK(&gc_threads[i]->wake_mutex);
1040 wakeup = gc_threads[i]->wakeup;
1041 // wakeup is false while the thread is waiting
1042 RELEASE_LOCK(&gc_threads[i]->wake_mutex);
1043 } while (wakeup);
1044 }
1045 #endif
1046 }
1047
1048 /* ----------------------------------------------------------------------------
1049 Initialise a generation that is to be collected
1050 ------------------------------------------------------------------------- */
1051
1052 static void
1053 init_collected_gen (nat g, nat n_threads)
1054 {
1055 nat s, t, i;
1056 step_workspace *ws;
1057 step *stp;
1058 bdescr *bd;
1059
1060 // Throw away the current mutable list. Invariant: the mutable
1061 // list always has at least one block; this means we can avoid a
1062 // check for NULL in recordMutable().
1063 if (g != 0) {
1064 freeChain(generations[g].mut_list);
1065 generations[g].mut_list = allocBlock();
1066 for (i = 0; i < n_capabilities; i++) {
1067 freeChain(capabilities[i].mut_lists[g]);
1068 capabilities[i].mut_lists[g] = allocBlock();
1069 }
1070 }
1071
1072 for (s = 0; s < generations[g].n_steps; s++) {
1073
1074 // generation 0, step 0 doesn't need to-space
1075 if (g == 0 && s == 0 && RtsFlags.GcFlags.generations > 1) {
1076 continue;
1077 }
1078
1079 stp = &generations[g].steps[s];
1080 ASSERT(stp->gen_no == g);
1081
1082 // deprecate the existing blocks
1083 stp->old_blocks = stp->blocks;
1084 stp->n_old_blocks = stp->n_blocks;
1085 stp->blocks = NULL;
1086 stp->n_blocks = 0;
1087 stp->n_words = 0;
1088
1089 // we don't have any to-be-scavenged blocks yet
1090 stp->todos = NULL;
1091 stp->todos_last = NULL;
1092 stp->n_todos = 0;
1093
1094 // initialise the large object queues.
1095 stp->scavenged_large_objects = NULL;
1096 stp->n_scavenged_large_blocks = 0;
1097
1098 // mark the large objects as not evacuated yet
1099 for (bd = stp->large_objects; bd; bd = bd->link) {
1100 bd->flags &= ~BF_EVACUATED;
1101 }
1102
1103 // for a compacted step, we need to allocate the bitmap
1104 if (stp->is_compacted) {
1105 nat bitmap_size; // in bytes
1106 bdescr *bitmap_bdescr;
1107 StgWord *bitmap;
1108
1109 bitmap_size = stp->n_old_blocks * BLOCK_SIZE / (sizeof(W_)*BITS_PER_BYTE);
1110
1111 if (bitmap_size > 0) {
1112 bitmap_bdescr = allocGroup((lnat)BLOCK_ROUND_UP(bitmap_size)
1113 / BLOCK_SIZE);
1114 stp->bitmap = bitmap_bdescr;
1115 bitmap = bitmap_bdescr->start;
1116
1117 debugTrace(DEBUG_gc, "bitmap_size: %d, bitmap: %p",
1118 bitmap_size, bitmap);
1119
1120 // don't forget to fill it with zeros!
1121 memset(bitmap, 0, bitmap_size);
1122
1123 // For each block in this step, point to its bitmap from the
1124 // block descriptor.
1125 for (bd=stp->old_blocks; bd != NULL; bd = bd->link) {
1126 bd->u.bitmap = bitmap;
1127 bitmap += BLOCK_SIZE_W / (sizeof(W_)*BITS_PER_BYTE);
1128
1129 // Also at this point we set the BF_COMPACTED flag
1130 // for this block. The invariant is that
1131 // BF_COMPACTED is always unset, except during GC
1132 // when it is set on those blocks which will be
1133 // compacted.
1134 bd->flags |= BF_COMPACTED;
1135 }
1136 }
1137 }
1138 }
1139
1140 // For each GC thread, for each step, allocate a "todo" block to
1141 // store evacuated objects to be scavenged, and a block to store
1142 // evacuated objects that do not need to be scavenged.
1143 for (t = 0; t < n_threads; t++) {
1144 for (s = 0; s < generations[g].n_steps; s++) {
1145
1146 // we don't copy objects into g0s0, unless -G0
1147 if (g==0 && s==0 && RtsFlags.GcFlags.generations > 1) continue;
1148
1149 ws = &gc_threads[t]->steps[g * RtsFlags.GcFlags.steps + s];
1150
1151 ws->todo_large_objects = NULL;
1152
1153 ws->part_list = NULL;
1154 ws->n_part_blocks = 0;
1155
1156 // allocate the first to-space block; extra blocks will be
1157 // chained on as necessary.
1158 ws->todo_bd = NULL;
1159 ws->buffer_todo_bd = NULL;
1160 alloc_todo_block(ws,0);
1161
1162 ws->scavd_list = NULL;
1163 ws->n_scavd_blocks = 0;
1164 }
1165 }
1166 }
1167
1168
1169 /* ----------------------------------------------------------------------------
1170 Initialise a generation that is *not* to be collected
1171 ------------------------------------------------------------------------- */
1172
1173 static void
1174 init_uncollected_gen (nat g, nat threads)
1175 {
1176 nat s, t, i;
1177 step_workspace *ws;
1178 step *stp;
1179 bdescr *bd;
1180
1181 for (s = 0; s < generations[g].n_steps; s++) {
1182 stp = &generations[g].steps[s];
1183 stp->scavenged_large_objects = NULL;
1184 stp->n_scavenged_large_blocks = 0;
1185 }
1186
1187 for (t = 0; t < threads; t++) {
1188 for (s = 0; s < generations[g].n_steps; s++) {
1189
1190 ws = &gc_threads[t]->steps[g * RtsFlags.GcFlags.steps + s];
1191 stp = ws->step;
1192
1193 ws->buffer_todo_bd = NULL;
1194 ws->todo_large_objects = NULL;
1195
1196 ws->part_list = NULL;
1197 ws->n_part_blocks = 0;
1198
1199 ws->scavd_list = NULL;
1200 ws->n_scavd_blocks = 0;
1201
1202 // If the block at the head of the list in this generation
1203 // is less than 3/4 full, then use it as a todo block.
1204 if (stp->blocks && isPartiallyFull(stp->blocks))
1205 {
1206 ws->todo_bd = stp->blocks;
1207 ws->todo_free = ws->todo_bd->free;
1208 ws->todo_lim = ws->todo_bd->start + BLOCK_SIZE_W;
1209 stp->blocks = stp->blocks->link;
1210 stp->n_blocks -= 1;
1211 stp->n_words -= ws->todo_bd->free - ws->todo_bd->start;
1212 ws->todo_bd->link = NULL;
1213 // we must scan from the current end point.
1214 ws->todo_bd->u.scan = ws->todo_bd->free;
1215 }
1216 else
1217 {
1218 ws->todo_bd = NULL;
1219 alloc_todo_block(ws,0);
1220 }
1221 }
1222 }
1223
1224 // Move the private mutable lists from each capability onto the
1225 // main mutable list for the generation.
1226 for (i = 0; i < n_capabilities; i++) {
1227 for (bd = capabilities[i].mut_lists[g];
1228 bd->link != NULL; bd = bd->link) {
1229 /* nothing */
1230 }
1231 bd->link = generations[g].mut_list;
1232 generations[g].mut_list = capabilities[i].mut_lists[g];
1233 capabilities[i].mut_lists[g] = allocBlock();
1234 }
1235 }
1236
1237 /* -----------------------------------------------------------------------------
1238 Initialise a gc_thread before GC
1239 -------------------------------------------------------------------------- */
1240
1241 static void
1242 init_gc_thread (gc_thread *t)
1243 {
1244 t->static_objects = END_OF_STATIC_LIST;
1245 t->scavenged_static_objects = END_OF_STATIC_LIST;
1246 t->scan_bd = NULL;
1247 t->evac_step = 0;
1248 t->failed_to_evac = rtsFalse;
1249 t->eager_promotion = rtsTrue;
1250 t->thunk_selector_depth = 0;
1251 t->copied = 0;
1252 t->scanned = 0;
1253 t->any_work = 0;
1254 t->no_work = 0;
1255 t->scav_find_work = 0;
1256 }
1257
1258 /* -----------------------------------------------------------------------------
1259 Function we pass to evacuate roots.
1260 -------------------------------------------------------------------------- */
1261
1262 static void
1263 mark_root(void *user, StgClosure **root)
1264 {
1265 // we stole a register for gct, but this function is called from
1266 // *outside* the GC where the register variable is not in effect,
1267 // so we need to save and restore it here. NB. only call
1268 // mark_root() from the main GC thread, otherwise gct will be
1269 // incorrect.
1270 gc_thread *saved_gct;
1271 saved_gct = gct;
1272 gct = user;
1273
1274 evacuate(root);
1275
1276 gct = saved_gct;
1277 }
1278
1279 /* -----------------------------------------------------------------------------
1280 Initialising the static object & mutable lists
1281 -------------------------------------------------------------------------- */
1282
1283 static void
1284 zero_static_object_list(StgClosure* first_static)
1285 {
1286 StgClosure* p;
1287 StgClosure* link;
1288 const StgInfoTable *info;
1289
1290 for (p = first_static; p != END_OF_STATIC_LIST; p = link) {
1291 info = get_itbl(p);
1292 link = *STATIC_LINK(info, p);
1293 *STATIC_LINK(info,p) = NULL;
1294 }
1295 }
1296
1297 /* ----------------------------------------------------------------------------
1298 Update the pointers from the task list
1299
1300 These are treated as weak pointers because we want to allow a main
1301 thread to get a BlockedOnDeadMVar exception in the same way as any
1302 other thread. Note that the threads should all have been retained
1303 by GC by virtue of being on the all_threads list, we're just
1304 updating pointers here.
1305 ------------------------------------------------------------------------- */
1306
1307 static void
1308 update_task_list (void)
1309 {
1310 Task *task;
1311 StgTSO *tso;
1312 for (task = all_tasks; task != NULL; task = task->all_link) {
1313 if (!task->stopped && task->tso) {
1314 ASSERT(task->tso->bound == task);
1315 tso = (StgTSO *) isAlive((StgClosure *)task->tso);
1316 if (tso == NULL) {
1317 barf("task %p: main thread %d has been GC'd",
1318 #ifdef THREADED_RTS
1319 (void *)task->id,
1320 #else
1321 (void *)task,
1322 #endif
1323 task->tso->id);
1324 }
1325 task->tso = tso;
1326 }
1327 }
1328 }
1329
1330 /* ----------------------------------------------------------------------------
1331 Reset the sizes of the older generations when we do a major
1332 collection.
1333
1334 CURRENT STRATEGY: make all generations except zero the same size.
1335 We have to stay within the maximum heap size, and leave a certain
1336 percentage of the maximum heap size available to allocate into.
1337 ------------------------------------------------------------------------- */
1338
1339 static void
1340 resize_generations (void)
1341 {
1342 nat g;
1343
1344 if (major_gc && RtsFlags.GcFlags.generations > 1) {
1345 nat live, size, min_alloc;
1346 nat max = RtsFlags.GcFlags.maxHeapSize;
1347 nat gens = RtsFlags.GcFlags.generations;
1348
1349 // live in the oldest generations
1350 live = (oldest_gen->steps[0].n_words + BLOCK_SIZE_W - 1) / BLOCK_SIZE_W+
1351 oldest_gen->steps[0].n_large_blocks;
1352
1353 // default max size for all generations except zero
1354 size = stg_max(live * RtsFlags.GcFlags.oldGenFactor,
1355 RtsFlags.GcFlags.minOldGenSize);
1356
1357 // minimum size for generation zero
1358 min_alloc = stg_max((RtsFlags.GcFlags.pcFreeHeap * max) / 200,
1359 RtsFlags.GcFlags.minAllocAreaSize);
1360
1361 // Auto-enable compaction when the residency reaches a
1362 // certain percentage of the maximum heap size (default: 30%).
1363 if (RtsFlags.GcFlags.generations > 1 &&
1364 (RtsFlags.GcFlags.compact ||
1365 (max > 0 &&
1366 oldest_gen->steps[0].n_blocks >
1367 (RtsFlags.GcFlags.compactThreshold * max) / 100))) {
1368 oldest_gen->steps[0].is_compacted = 1;
1369 // debugBelch("compaction: on\n", live);
1370 } else {
1371 oldest_gen->steps[0].is_compacted = 0;
1372 // debugBelch("compaction: off\n", live);
1373 }
1374
1375 // if we're going to go over the maximum heap size, reduce the
1376 // size of the generations accordingly. The calculation is
1377 // different if compaction is turned on, because we don't need
1378 // to double the space required to collect the old generation.
1379 if (max != 0) {
1380
1381 // this test is necessary to ensure that the calculations
1382 // below don't have any negative results - we're working
1383 // with unsigned values here.
1384 if (max < min_alloc) {
1385 heapOverflow();
1386 }
1387
1388 if (oldest_gen->steps[0].is_compacted) {
1389 if ( (size + (size - 1) * (gens - 2) * 2) + min_alloc > max ) {
1390 size = (max - min_alloc) / ((gens - 1) * 2 - 1);
1391 }
1392 } else {
1393 if ( (size * (gens - 1) * 2) + min_alloc > max ) {
1394 size = (max - min_alloc) / ((gens - 1) * 2);
1395 }
1396 }
1397
1398 if (size < live) {
1399 heapOverflow();
1400 }
1401 }
1402
1403 #if 0
1404 debugBelch("live: %d, min_alloc: %d, size : %d, max = %d\n", live,
1405 min_alloc, size, max);
1406 #endif
1407
1408 for (g = 0; g < gens; g++) {
1409 generations[g].max_blocks = size;
1410 }
1411 }
1412 }
1413
1414 /* -----------------------------------------------------------------------------
1415 Calculate the new size of the nursery, and resize it.
1416 -------------------------------------------------------------------------- */
1417
1418 static void
1419 resize_nursery (void)
1420 {
1421 if (RtsFlags.GcFlags.generations == 1)
1422 { // Two-space collector:
1423 nat blocks;
1424
1425 /* set up a new nursery. Allocate a nursery size based on a
1426 * function of the amount of live data (by default a factor of 2)
1427 * Use the blocks from the old nursery if possible, freeing up any
1428 * left over blocks.
1429 *
1430 * If we get near the maximum heap size, then adjust our nursery
1431 * size accordingly. If the nursery is the same size as the live
1432 * data (L), then we need 3L bytes. We can reduce the size of the
1433 * nursery to bring the required memory down near 2L bytes.
1434 *
1435 * A normal 2-space collector would need 4L bytes to give the same
1436 * performance we get from 3L bytes, reducing to the same
1437 * performance at 2L bytes.
1438 */
1439 blocks = g0s0->n_old_blocks;
1440
1441 if ( RtsFlags.GcFlags.maxHeapSize != 0 &&
1442 blocks * RtsFlags.GcFlags.oldGenFactor * 2 >
1443 RtsFlags.GcFlags.maxHeapSize )
1444 {
1445 long adjusted_blocks; // signed on purpose
1446 int pc_free;
1447
1448 adjusted_blocks = (RtsFlags.GcFlags.maxHeapSize - 2 * blocks);
1449
1450 debugTrace(DEBUG_gc, "near maximum heap size of 0x%x blocks, blocks = %d, adjusted to %ld",
1451 RtsFlags.GcFlags.maxHeapSize, blocks, adjusted_blocks);
1452
1453 pc_free = adjusted_blocks * 100 / RtsFlags.GcFlags.maxHeapSize;
1454 if (pc_free < RtsFlags.GcFlags.pcFreeHeap) /* might even * be < 0 */
1455 {
1456 heapOverflow();
1457 }
1458 blocks = adjusted_blocks;
1459 }
1460 else
1461 {
1462 blocks *= RtsFlags.GcFlags.oldGenFactor;
1463 if (blocks < RtsFlags.GcFlags.minAllocAreaSize)
1464 {
1465 blocks = RtsFlags.GcFlags.minAllocAreaSize;
1466 }
1467 }
1468 resizeNurseries(blocks);
1469 }
1470 else // Generational collector
1471 {
1472 /*
1473 * If the user has given us a suggested heap size, adjust our
1474 * allocation area to make best use of the memory available.
1475 */
1476 if (RtsFlags.GcFlags.heapSizeSuggestion)
1477 {
1478 long blocks;
1479 nat needed = calcNeeded(); // approx blocks needed at next GC
1480
1481 /* Guess how much will be live in generation 0 step 0 next time.
1482 * A good approximation is obtained by finding the
1483 * percentage of g0s0 that was live at the last minor GC.
1484 *
1485 * We have an accurate figure for the amount of copied data in
1486 * 'copied', but we must convert this to a number of blocks, with
1487 * a small adjustment for estimated slop at the end of a block
1488 * (- 10 words).
1489 */
1490 if (N == 0)
1491 {
1492 g0s0_pcnt_kept = ((copied / (BLOCK_SIZE_W - 10)) * 100)
1493 / countNurseryBlocks();
1494 }
1495
1496 /* Estimate a size for the allocation area based on the
1497 * information available. We might end up going slightly under
1498 * or over the suggested heap size, but we should be pretty
1499 * close on average.
1500 *
1501 * Formula: suggested - needed
1502 * ----------------------------
1503 * 1 + g0s0_pcnt_kept/100
1504 *
1505 * where 'needed' is the amount of memory needed at the next
1506 * collection for collecting all steps except g0s0.
1507 */
1508 blocks =
1509 (((long)RtsFlags.GcFlags.heapSizeSuggestion - (long)needed) * 100) /
1510 (100 + (long)g0s0_pcnt_kept);
1511
1512 if (blocks < (long)RtsFlags.GcFlags.minAllocAreaSize) {
1513 blocks = RtsFlags.GcFlags.minAllocAreaSize;
1514 }
1515
1516 resizeNurseries((nat)blocks);
1517 }
1518 else
1519 {
1520 // we might have added extra large blocks to the nursery, so
1521 // resize back to minAllocAreaSize again.
1522 resizeNurseriesFixed(RtsFlags.GcFlags.minAllocAreaSize);
1523 }
1524 }
1525 }
1526
1527 /* -----------------------------------------------------------------------------
1528 Sanity code for CAF garbage collection.
1529
1530 With DEBUG turned on, we manage a CAF list in addition to the SRT
1531 mechanism. After GC, we run down the CAF list and blackhole any
1532 CAFs which have been garbage collected. This means we get an error
1533 whenever the program tries to enter a garbage collected CAF.
1534
1535 Any garbage collected CAFs are taken off the CAF list at the same
1536 time.
1537 -------------------------------------------------------------------------- */
1538
1539 #if 0 && defined(DEBUG)
1540
1541 static void
1542 gcCAFs(void)
1543 {
1544 StgClosure* p;
1545 StgClosure** pp;
1546 const StgInfoTable *info;
1547 nat i;
1548
1549 i = 0;
1550 p = caf_list;
1551 pp = &caf_list;
1552
1553 while (p != NULL) {
1554
1555 info = get_itbl(p);
1556
1557 ASSERT(info->type == IND_STATIC);
1558
1559 if (STATIC_LINK(info,p) == NULL) {
1560 debugTrace(DEBUG_gccafs, "CAF gc'd at 0x%04lx", (long)p);
1561 // black hole it
1562 SET_INFO(p,&stg_BLACKHOLE_info);
1563 p = STATIC_LINK2(info,p);
1564 *pp = p;
1565 }
1566 else {
1567 pp = &STATIC_LINK2(info,p);
1568 p = *pp;
1569 i++;
1570 }
1571
1572 }
1573
1574 debugTrace(DEBUG_gccafs, "%d CAFs live", i);
1575 }
1576 #endif