GC refactoring and cleanup
[ghc.git] / rts / sm / Storage.c
1 /* -----------------------------------------------------------------------------
2 *
3 * (c) The GHC Team, 1998-2008
4 *
5 * Storage manager front end
6 *
7 * Documentation on the architecture of the Storage Manager can be
8 * found in the online commentary:
9 *
10 * http://hackage.haskell.org/trac/ghc/wiki/Commentary/Rts/Storage
11 *
12 * ---------------------------------------------------------------------------*/
13
14 #include "PosixSource.h"
15 #include "Rts.h"
16
17 #include "Storage.h"
18 #include "GCThread.h"
19 #include "RtsUtils.h"
20 #include "Stats.h"
21 #include "BlockAlloc.h"
22 #include "Weak.h"
23 #include "Sanity.h"
24 #include "Arena.h"
25 #include "Capability.h"
26 #include "Schedule.h"
27 #include "RetainerProfile.h" // for counting memory blocks (memInventory)
28 #include "OSMem.h"
29 #include "Trace.h"
30 #include "GC.h"
31 #include "Evac.h"
32
33 #include <string.h>
34
35 #include "ffi.h"
36
37 /*
38 * All these globals require sm_mutex to access in THREADED_RTS mode.
39 */
40 StgClosure *caf_list = NULL;
41 StgClosure *revertible_caf_list = NULL;
42 rtsBool keepCAFs;
43
44 nat large_alloc_lim; /* GC if n_large_blocks in any nursery
45 * reaches this. */
46
47 bdescr *exec_block;
48
49 generation *generations = NULL; /* all the generations */
50 generation *g0 = NULL; /* generation 0, for convenience */
51 generation *oldest_gen = NULL; /* oldest generation, for convenience */
52
53 nursery *nurseries = NULL; /* array of nurseries, size == n_capabilities */
54
55 #ifdef THREADED_RTS
56 /*
57 * Storage manager mutex: protects all the above state from
58 * simultaneous access by two STG threads.
59 */
60 Mutex sm_mutex;
61 #endif
62
63 static void allocNurseries ( void );
64
65 static void
66 initGeneration (generation *gen, int g)
67 {
68 gen->no = g;
69 gen->collections = 0;
70 gen->par_collections = 0;
71 gen->failed_promotions = 0;
72 gen->max_blocks = 0;
73 gen->blocks = NULL;
74 gen->n_blocks = 0;
75 gen->n_words = 0;
76 gen->live_estimate = 0;
77 gen->old_blocks = NULL;
78 gen->n_old_blocks = 0;
79 gen->large_objects = NULL;
80 gen->n_large_blocks = 0;
81 gen->n_new_large_words = 0;
82 gen->scavenged_large_objects = NULL;
83 gen->n_scavenged_large_blocks = 0;
84 gen->mark = 0;
85 gen->compact = 0;
86 gen->bitmap = NULL;
87 #ifdef THREADED_RTS
88 initSpinLock(&gen->sync);
89 #endif
90 gen->threads = END_TSO_QUEUE;
91 gen->old_threads = END_TSO_QUEUE;
92 }
93
94 void
95 initStorage( void )
96 {
97 nat g, n;
98
99 if (generations != NULL) {
100 // multi-init protection
101 return;
102 }
103
104 initMBlocks();
105
106 /* Sanity check to make sure the LOOKS_LIKE_ macros appear to be
107 * doing something reasonable.
108 */
109 /* We use the NOT_NULL variant or gcc warns that the test is always true */
110 ASSERT(LOOKS_LIKE_INFO_PTR_NOT_NULL((StgWord)&stg_BLOCKING_QUEUE_CLEAN_info));
111 ASSERT(LOOKS_LIKE_CLOSURE_PTR(&stg_dummy_ret_closure));
112 ASSERT(!HEAP_ALLOCED(&stg_dummy_ret_closure));
113
114 if (RtsFlags.GcFlags.maxHeapSize != 0 &&
115 RtsFlags.GcFlags.heapSizeSuggestion >
116 RtsFlags.GcFlags.maxHeapSize) {
117 RtsFlags.GcFlags.maxHeapSize = RtsFlags.GcFlags.heapSizeSuggestion;
118 }
119
120 if (RtsFlags.GcFlags.maxHeapSize != 0 &&
121 RtsFlags.GcFlags.minAllocAreaSize >
122 RtsFlags.GcFlags.maxHeapSize) {
123 errorBelch("maximum heap size (-M) is smaller than minimum alloc area size (-A)");
124 RtsFlags.GcFlags.minAllocAreaSize = RtsFlags.GcFlags.maxHeapSize;
125 }
126
127 initBlockAllocator();
128
129 #if defined(THREADED_RTS)
130 initMutex(&sm_mutex);
131 #endif
132
133 ACQUIRE_SM_LOCK;
134
135 /* allocate generation info array */
136 generations = (generation *)stgMallocBytes(RtsFlags.GcFlags.generations
137 * sizeof(struct generation_),
138 "initStorage: gens");
139
140 /* Initialise all generations */
141 for(g = 0; g < RtsFlags.GcFlags.generations; g++) {
142 initGeneration(&generations[g], g);
143 }
144
145 /* A couple of convenience pointers */
146 g0 = &generations[0];
147 oldest_gen = &generations[RtsFlags.GcFlags.generations-1];
148
149 nurseries = stgMallocBytes(n_capabilities * sizeof(struct nursery_),
150 "initStorage: nurseries");
151
152 /* Set up the destination pointers in each younger gen. step */
153 for (g = 0; g < RtsFlags.GcFlags.generations-1; g++) {
154 generations[g].to = &generations[g+1];
155 }
156 oldest_gen->to = oldest_gen;
157
158 /* The oldest generation has one step. */
159 if (RtsFlags.GcFlags.compact || RtsFlags.GcFlags.sweep) {
160 if (RtsFlags.GcFlags.generations == 1) {
161 errorBelch("WARNING: compact/sweep is incompatible with -G1; disabled");
162 } else {
163 oldest_gen->mark = 1;
164 if (RtsFlags.GcFlags.compact)
165 oldest_gen->compact = 1;
166 }
167 }
168
169 generations[0].max_blocks = 0;
170
171 /* The allocation area. Policy: keep the allocation area
172 * small to begin with, even if we have a large suggested heap
173 * size. Reason: we're going to do a major collection first, and we
174 * don't want it to be a big one. This vague idea is borne out by
175 * rigorous experimental evidence.
176 */
177 allocNurseries();
178
179 weak_ptr_list = NULL;
180 caf_list = END_OF_STATIC_LIST;
181 revertible_caf_list = END_OF_STATIC_LIST;
182
183 /* initialise the allocate() interface */
184 large_alloc_lim = RtsFlags.GcFlags.minAllocAreaSize * BLOCK_SIZE_W;
185
186 exec_block = NULL;
187
188 #ifdef THREADED_RTS
189 initSpinLock(&gc_alloc_block_sync);
190 whitehole_spin = 0;
191 #endif
192
193 N = 0;
194
195 // allocate a block for each mut list
196 for (n = 0; n < n_capabilities; n++) {
197 for (g = 1; g < RtsFlags.GcFlags.generations; g++) {
198 capabilities[n].mut_lists[g] = allocBlock();
199 }
200 }
201
202 initGcThreads();
203
204 IF_DEBUG(gc, statDescribeGens());
205
206 RELEASE_SM_LOCK;
207 }
208
209 void
210 exitStorage (void)
211 {
212 stat_exit(calcAllocated(rtsTrue));
213 }
214
215 void
216 freeStorage (rtsBool free_heap)
217 {
218 stgFree(generations);
219 if (free_heap) freeAllMBlocks();
220 #if defined(THREADED_RTS)
221 closeMutex(&sm_mutex);
222 #endif
223 stgFree(nurseries);
224 freeGcThreads();
225 }
226
227 /* -----------------------------------------------------------------------------
228 CAF management.
229
230 The entry code for every CAF does the following:
231
232 - builds a BLACKHOLE in the heap
233 - pushes an update frame pointing to the BLACKHOLE
234 - calls newCaf, below
235 - updates the CAF with a static indirection to the BLACKHOLE
236
237 Why do we build an BLACKHOLE in the heap rather than just updating
238 the thunk directly? It's so that we only need one kind of update
239 frame - otherwise we'd need a static version of the update frame too.
240
241 newCaf() does the following:
242
243 - it puts the CAF on the oldest generation's mutable list.
244 This is so that we treat the CAF as a root when collecting
245 younger generations.
246
247 For GHCI, we have additional requirements when dealing with CAFs:
248
249 - we must *retain* all dynamically-loaded CAFs ever entered,
250 just in case we need them again.
251 - we must be able to *revert* CAFs that have been evaluated, to
252 their pre-evaluated form.
253
254 To do this, we use an additional CAF list. When newCaf() is
255 called on a dynamically-loaded CAF, we add it to the CAF list
256 instead of the old-generation mutable list, and save away its
257 old info pointer (in caf->saved_info) for later reversion.
258
259 To revert all the CAFs, we traverse the CAF list and reset the
260 info pointer to caf->saved_info, then throw away the CAF list.
261 (see GC.c:revertCAFs()).
262
263 -- SDM 29/1/01
264
265 -------------------------------------------------------------------------- */
266
267 void
268 newCAF(StgRegTable *reg, StgClosure* caf)
269 {
270 if(keepCAFs)
271 {
272 // HACK:
273 // If we are in GHCi _and_ we are using dynamic libraries,
274 // then we can't redirect newCAF calls to newDynCAF (see below),
275 // so we make newCAF behave almost like newDynCAF.
276 // The dynamic libraries might be used by both the interpreted
277 // program and GHCi itself, so they must not be reverted.
278 // This also means that in GHCi with dynamic libraries, CAFs are not
279 // garbage collected. If this turns out to be a problem, we could
280 // do another hack here and do an address range test on caf to figure
281 // out whether it is from a dynamic library.
282 ((StgIndStatic *)caf)->saved_info = (StgInfoTable *)caf->header.info;
283
284 ACQUIRE_SM_LOCK; // caf_list is global, locked by sm_mutex
285 ((StgIndStatic *)caf)->static_link = caf_list;
286 caf_list = caf;
287 RELEASE_SM_LOCK;
288 }
289 else
290 {
291 // Put this CAF on the mutable list for the old generation.
292 ((StgIndStatic *)caf)->saved_info = NULL;
293 if (oldest_gen->no != 0) {
294 recordMutableCap(caf, regTableToCapability(reg), oldest_gen->no);
295 }
296 }
297 }
298
299 // External API for setting the keepCAFs flag. see #3900.
300 void
301 setKeepCAFs (void)
302 {
303 keepCAFs = 1;
304 }
305
306 // An alternate version of newCaf which is used for dynamically loaded
307 // object code in GHCi. In this case we want to retain *all* CAFs in
308 // the object code, because they might be demanded at any time from an
309 // expression evaluated on the command line.
310 // Also, GHCi might want to revert CAFs, so we add these to the
311 // revertible_caf_list.
312 //
313 // The linker hackily arranges that references to newCaf from dynamic
314 // code end up pointing to newDynCAF.
315 void
316 newDynCAF (StgRegTable *reg STG_UNUSED, StgClosure *caf)
317 {
318 ACQUIRE_SM_LOCK;
319
320 ((StgIndStatic *)caf)->saved_info = (StgInfoTable *)caf->header.info;
321 ((StgIndStatic *)caf)->static_link = revertible_caf_list;
322 revertible_caf_list = caf;
323
324 RELEASE_SM_LOCK;
325 }
326
327 /* -----------------------------------------------------------------------------
328 Nursery management.
329 -------------------------------------------------------------------------- */
330
331 static bdescr *
332 allocNursery (bdescr *tail, nat blocks)
333 {
334 bdescr *bd = NULL;
335 nat i, n;
336
337 // We allocate the nursery as a single contiguous block and then
338 // divide it into single blocks manually. This way we guarantee
339 // that the nursery blocks are adjacent, so that the processor's
340 // automatic prefetching works across nursery blocks. This is a
341 // tiny optimisation (~0.5%), but it's free.
342
343 while (blocks > 0) {
344 n = stg_min(blocks, BLOCKS_PER_MBLOCK);
345 blocks -= n;
346
347 bd = allocGroup(n);
348 for (i = 0; i < n; i++) {
349 initBdescr(&bd[i], g0, g0);
350
351 bd[i].blocks = 1;
352 bd[i].flags = 0;
353
354 if (i > 0) {
355 bd[i].u.back = &bd[i-1];
356 } else {
357 bd[i].u.back = NULL;
358 }
359
360 if (i+1 < n) {
361 bd[i].link = &bd[i+1];
362 } else {
363 bd[i].link = tail;
364 if (tail != NULL) {
365 tail->u.back = &bd[i];
366 }
367 }
368
369 bd[i].free = bd[i].start;
370 }
371
372 tail = &bd[0];
373 }
374
375 return &bd[0];
376 }
377
378 static void
379 assignNurseriesToCapabilities (void)
380 {
381 nat i;
382
383 for (i = 0; i < n_capabilities; i++) {
384 capabilities[i].r.rNursery = &nurseries[i];
385 capabilities[i].r.rCurrentNursery = nurseries[i].blocks;
386 capabilities[i].r.rCurrentAlloc = NULL;
387 }
388 }
389
390 static void
391 allocNurseries( void )
392 {
393 nat i;
394
395 for (i = 0; i < n_capabilities; i++) {
396 nurseries[i].blocks =
397 allocNursery(NULL, RtsFlags.GcFlags.minAllocAreaSize);
398 nurseries[i].n_blocks =
399 RtsFlags.GcFlags.minAllocAreaSize;
400 }
401 assignNurseriesToCapabilities();
402 }
403
404 lnat // words allocated
405 clearNurseries (void)
406 {
407 lnat allocated = 0;
408 nat i;
409 bdescr *bd;
410
411 for (i = 0; i < n_capabilities; i++) {
412 for (bd = nurseries[i].blocks; bd; bd = bd->link) {
413 allocated += (lnat)(bd->free - bd->start);
414 bd->free = bd->start;
415 ASSERT(bd->gen_no == 0);
416 ASSERT(bd->gen == g0);
417 IF_DEBUG(sanity,memset(bd->start, 0xaa, BLOCK_SIZE));
418 }
419 }
420
421 return allocated;
422 }
423
424 void
425 resetNurseries (void)
426 {
427 assignNurseriesToCapabilities();
428
429 }
430
431 lnat
432 countNurseryBlocks (void)
433 {
434 nat i;
435 lnat blocks = 0;
436
437 for (i = 0; i < n_capabilities; i++) {
438 blocks += nurseries[i].n_blocks;
439 }
440 return blocks;
441 }
442
443 static void
444 resizeNursery ( nursery *nursery, nat blocks )
445 {
446 bdescr *bd;
447 nat nursery_blocks;
448
449 nursery_blocks = nursery->n_blocks;
450 if (nursery_blocks == blocks) return;
451
452 if (nursery_blocks < blocks) {
453 debugTrace(DEBUG_gc, "increasing size of nursery to %d blocks",
454 blocks);
455 nursery->blocks = allocNursery(nursery->blocks, blocks-nursery_blocks);
456 }
457 else {
458 bdescr *next_bd;
459
460 debugTrace(DEBUG_gc, "decreasing size of nursery to %d blocks",
461 blocks);
462
463 bd = nursery->blocks;
464 while (nursery_blocks > blocks) {
465 next_bd = bd->link;
466 next_bd->u.back = NULL;
467 nursery_blocks -= bd->blocks; // might be a large block
468 freeGroup(bd);
469 bd = next_bd;
470 }
471 nursery->blocks = bd;
472 // might have gone just under, by freeing a large block, so make
473 // up the difference.
474 if (nursery_blocks < blocks) {
475 nursery->blocks = allocNursery(nursery->blocks, blocks-nursery_blocks);
476 }
477 }
478
479 nursery->n_blocks = blocks;
480 ASSERT(countBlocks(nursery->blocks) == nursery->n_blocks);
481 }
482
483 //
484 // Resize each of the nurseries to the specified size.
485 //
486 void
487 resizeNurseriesFixed (nat blocks)
488 {
489 nat i;
490 for (i = 0; i < n_capabilities; i++) {
491 resizeNursery(&nurseries[i], blocks);
492 }
493 }
494
495 //
496 // Resize the nurseries to the total specified size.
497 //
498 void
499 resizeNurseries (nat blocks)
500 {
501 // If there are multiple nurseries, then we just divide the number
502 // of available blocks between them.
503 resizeNurseriesFixed(blocks / n_capabilities);
504 }
505
506
507 /* -----------------------------------------------------------------------------
508 move_STACK is called to update the TSO structure after it has been
509 moved from one place to another.
510 -------------------------------------------------------------------------- */
511
512 void
513 move_STACK (StgStack *src, StgStack *dest)
514 {
515 ptrdiff_t diff;
516
517 // relocate the stack pointer...
518 diff = (StgPtr)dest - (StgPtr)src; // In *words*
519 dest->sp = (StgPtr)dest->sp + diff;
520 }
521
522 /* -----------------------------------------------------------------------------
523 allocate()
524
525 This allocates memory in the current thread - it is intended for
526 use primarily from STG-land where we have a Capability. It is
527 better than allocate() because it doesn't require taking the
528 sm_mutex lock in the common case.
529
530 Memory is allocated directly from the nursery if possible (but not
531 from the current nursery block, so as not to interfere with
532 Hp/HpLim).
533 -------------------------------------------------------------------------- */
534
535 StgPtr
536 allocate (Capability *cap, lnat n)
537 {
538 bdescr *bd;
539 StgPtr p;
540
541 if (n >= LARGE_OBJECT_THRESHOLD/sizeof(W_)) {
542 lnat req_blocks = (lnat)BLOCK_ROUND_UP(n*sizeof(W_)) / BLOCK_SIZE;
543
544 // Attempting to allocate an object larger than maxHeapSize
545 // should definitely be disallowed. (bug #1791)
546 if (RtsFlags.GcFlags.maxHeapSize > 0 &&
547 req_blocks >= RtsFlags.GcFlags.maxHeapSize) {
548 heapOverflow();
549 // heapOverflow() doesn't exit (see #2592), but we aren't
550 // in a position to do a clean shutdown here: we
551 // either have to allocate the memory or exit now.
552 // Allocating the memory would be bad, because the user
553 // has requested that we not exceed maxHeapSize, so we
554 // just exit.
555 stg_exit(EXIT_HEAPOVERFLOW);
556 }
557
558 ACQUIRE_SM_LOCK
559 bd = allocGroup(req_blocks);
560 dbl_link_onto(bd, &g0->large_objects);
561 g0->n_large_blocks += bd->blocks; // might be larger than req_blocks
562 g0->n_new_large_words += n;
563 RELEASE_SM_LOCK;
564 initBdescr(bd, g0, g0);
565 bd->flags = BF_LARGE;
566 bd->free = bd->start + n;
567 return bd->start;
568 }
569
570 /* small allocation (<LARGE_OBJECT_THRESHOLD) */
571
572 TICK_ALLOC_HEAP_NOCTR(n);
573 CCS_ALLOC(CCCS,n);
574
575 bd = cap->r.rCurrentAlloc;
576 if (bd == NULL || bd->free + n > bd->start + BLOCK_SIZE_W) {
577
578 // The CurrentAlloc block is full, we need to find another
579 // one. First, we try taking the next block from the
580 // nursery:
581 bd = cap->r.rCurrentNursery->link;
582
583 if (bd == NULL || bd->free + n > bd->start + BLOCK_SIZE_W) {
584 // The nursery is empty, or the next block is already
585 // full: allocate a fresh block (we can't fail here).
586 ACQUIRE_SM_LOCK;
587 bd = allocBlock();
588 cap->r.rNursery->n_blocks++;
589 RELEASE_SM_LOCK;
590 initBdescr(bd, g0, g0);
591 bd->flags = 0;
592 // If we had to allocate a new block, then we'll GC
593 // pretty quickly now, because MAYBE_GC() will
594 // notice that CurrentNursery->link is NULL.
595 } else {
596 // we have a block in the nursery: take it and put
597 // it at the *front* of the nursery list, and use it
598 // to allocate() from.
599 cap->r.rCurrentNursery->link = bd->link;
600 if (bd->link != NULL) {
601 bd->link->u.back = cap->r.rCurrentNursery;
602 }
603 }
604 dbl_link_onto(bd, &cap->r.rNursery->blocks);
605 cap->r.rCurrentAlloc = bd;
606 IF_DEBUG(sanity, checkNurserySanity(cap->r.rNursery));
607 }
608 p = bd->free;
609 bd->free += n;
610
611 IF_DEBUG(sanity, ASSERT(*((StgWord8*)p) == 0xaa));
612 return p;
613 }
614
615 /* ---------------------------------------------------------------------------
616 Allocate a fixed/pinned object.
617
618 We allocate small pinned objects into a single block, allocating a
619 new block when the current one overflows. The block is chained
620 onto the large_object_list of generation 0.
621
622 NOTE: The GC can't in general handle pinned objects. This
623 interface is only safe to use for ByteArrays, which have no
624 pointers and don't require scavenging. It works because the
625 block's descriptor has the BF_LARGE flag set, so the block is
626 treated as a large object and chained onto various lists, rather
627 than the individual objects being copied. However, when it comes
628 to scavenge the block, the GC will only scavenge the first object.
629 The reason is that the GC can't linearly scan a block of pinned
630 objects at the moment (doing so would require using the
631 mostly-copying techniques). But since we're restricting ourselves
632 to pinned ByteArrays, not scavenging is ok.
633
634 This function is called by newPinnedByteArray# which immediately
635 fills the allocated memory with a MutableByteArray#.
636 ------------------------------------------------------------------------- */
637
638 StgPtr
639 allocatePinned (Capability *cap, lnat n)
640 {
641 StgPtr p;
642 bdescr *bd;
643
644 // If the request is for a large object, then allocate()
645 // will give us a pinned object anyway.
646 if (n >= LARGE_OBJECT_THRESHOLD/sizeof(W_)) {
647 p = allocate(cap, n);
648 Bdescr(p)->flags |= BF_PINNED;
649 return p;
650 }
651
652 TICK_ALLOC_HEAP_NOCTR(n);
653 CCS_ALLOC(CCCS,n);
654
655 bd = cap->pinned_object_block;
656
657 // If we don't have a block of pinned objects yet, or the current
658 // one isn't large enough to hold the new object, allocate a new one.
659 if (bd == NULL || (bd->free + n) > (bd->start + BLOCK_SIZE_W)) {
660 ACQUIRE_SM_LOCK;
661 cap->pinned_object_block = bd = allocBlock();
662 dbl_link_onto(bd, &g0->large_objects);
663 g0->n_large_blocks++;
664 RELEASE_SM_LOCK;
665 initBdescr(bd, g0, g0);
666 bd->flags = BF_PINNED | BF_LARGE;
667 bd->free = bd->start;
668 }
669
670 g0->n_new_large_words += n;
671 p = bd->free;
672 bd->free += n;
673 return p;
674 }
675
676 /* -----------------------------------------------------------------------------
677 Write Barriers
678 -------------------------------------------------------------------------- */
679
680 /*
681 This is the write barrier for MUT_VARs, a.k.a. IORefs. A
682 MUT_VAR_CLEAN object is not on the mutable list; a MUT_VAR_DIRTY
683 is. When written to, a MUT_VAR_CLEAN turns into a MUT_VAR_DIRTY
684 and is put on the mutable list.
685 */
686 void
687 dirty_MUT_VAR(StgRegTable *reg, StgClosure *p)
688 {
689 Capability *cap = regTableToCapability(reg);
690 if (p->header.info == &stg_MUT_VAR_CLEAN_info) {
691 p->header.info = &stg_MUT_VAR_DIRTY_info;
692 recordClosureMutated(cap,p);
693 }
694 }
695
696 // Setting a TSO's link field with a write barrier.
697 // It is *not* necessary to call this function when
698 // * setting the link field to END_TSO_QUEUE
699 // * putting a TSO on the blackhole_queue
700 // * setting the link field of the currently running TSO, as it
701 // will already be dirty.
702 void
703 setTSOLink (Capability *cap, StgTSO *tso, StgTSO *target)
704 {
705 if (tso->dirty == 0) {
706 tso->dirty = 1;
707 recordClosureMutated(cap,(StgClosure*)tso);
708 }
709 tso->_link = target;
710 }
711
712 void
713 setTSOPrev (Capability *cap, StgTSO *tso, StgTSO *target)
714 {
715 if (tso->dirty == 0) {
716 tso->dirty = 1;
717 recordClosureMutated(cap,(StgClosure*)tso);
718 }
719 tso->block_info.prev = target;
720 }
721
722 void
723 dirty_TSO (Capability *cap, StgTSO *tso)
724 {
725 if (tso->dirty == 0) {
726 tso->dirty = 1;
727 recordClosureMutated(cap,(StgClosure*)tso);
728 }
729 }
730
731 void
732 dirty_STACK (Capability *cap, StgStack *stack)
733 {
734 if (stack->dirty == 0) {
735 stack->dirty = 1;
736 recordClosureMutated(cap,(StgClosure*)stack);
737 }
738 }
739
740 /*
741 This is the write barrier for MVARs. An MVAR_CLEAN objects is not
742 on the mutable list; a MVAR_DIRTY is. When written to, a
743 MVAR_CLEAN turns into a MVAR_DIRTY and is put on the mutable list.
744 The check for MVAR_CLEAN is inlined at the call site for speed,
745 this really does make a difference on concurrency-heavy benchmarks
746 such as Chaneneos and cheap-concurrency.
747 */
748 void
749 dirty_MVAR(StgRegTable *reg, StgClosure *p)
750 {
751 recordClosureMutated(regTableToCapability(reg),p);
752 }
753
754 /* -----------------------------------------------------------------------------
755 * Stats and stuff
756 * -------------------------------------------------------------------------- */
757
758 /* -----------------------------------------------------------------------------
759 * calcAllocated()
760 *
761 * Approximate how much we've allocated: number of blocks in the
762 * nursery + blocks allocated via allocate() - unused nusery blocks.
763 * This leaves a little slop at the end of each block.
764 * -------------------------------------------------------------------------- */
765
766 lnat
767 calcAllocated (rtsBool include_nurseries)
768 {
769 nat allocated = 0;
770 nat i;
771
772 // When called from GC.c, we already have the allocation count for
773 // the nursery from resetNurseries(), so we don't need to walk
774 // through these block lists again.
775 if (include_nurseries)
776 {
777 for (i = 0; i < n_capabilities; i++) {
778 allocated += countOccupied(nurseries[i].blocks);
779 }
780 }
781
782 // add in sizes of new large and pinned objects
783 allocated += g0->n_new_large_words;
784
785 return allocated;
786 }
787
788 lnat countOccupied (bdescr *bd)
789 {
790 lnat words;
791
792 words = 0;
793 for (; bd != NULL; bd = bd->link) {
794 ASSERT(bd->free <= bd->start + bd->blocks * BLOCK_SIZE_W);
795 words += bd->free - bd->start;
796 }
797 return words;
798 }
799
800 lnat genLiveWords (generation *gen)
801 {
802 return gen->n_words + countOccupied(gen->large_objects);
803 }
804
805 lnat genLiveBlocks (generation *gen)
806 {
807 return gen->n_blocks + gen->n_large_blocks;
808 }
809
810 lnat gcThreadLiveWords (nat i, nat g)
811 {
812 lnat words;
813
814 words = countOccupied(gc_threads[i]->gens[g].todo_bd);
815 words += countOccupied(gc_threads[i]->gens[g].part_list);
816 words += countOccupied(gc_threads[i]->gens[g].scavd_list);
817
818 return words;
819 }
820
821 lnat gcThreadLiveBlocks (nat i, nat g)
822 {
823 lnat blocks;
824
825 blocks = countBlocks(gc_threads[i]->gens[g].todo_bd);
826 blocks += gc_threads[i]->gens[g].n_part_blocks;
827 blocks += gc_threads[i]->gens[g].n_scavd_blocks;
828
829 return blocks;
830 }
831
832 // Return an accurate count of the live data in the heap, excluding
833 // generation 0.
834 lnat calcLiveWords (void)
835 {
836 nat g;
837 lnat live;
838 generation *gen;
839
840 live = 0;
841 for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
842 live += genLiveWords(&generations[g]);
843 }
844 return live;
845 }
846
847 lnat calcLiveBlocks (void)
848 {
849 nat g;
850 lnat live;
851
852 live = 0;
853 for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
854 live += genLiveBlocks(&generations[g]);
855 }
856 return live;
857 }
858
859 /* Approximate the number of blocks that will be needed at the next
860 * garbage collection.
861 *
862 * Assume: all data currently live will remain live. Generationss
863 * that will be collected next time will therefore need twice as many
864 * blocks since all the data will be copied.
865 */
866 extern lnat
867 calcNeeded(void)
868 {
869 lnat needed = 0;
870 nat g;
871 generation *gen;
872
873 for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
874 gen = &generations[g];
875
876 // we need at least this much space
877 needed += gen->n_blocks + gen->n_large_blocks;
878
879 // any additional space needed to collect this gen next time?
880 if (g == 0 || // always collect gen 0
881 (gen->n_blocks + gen->n_large_blocks > gen->max_blocks)) {
882 // we will collect this gen next time
883 if (gen->mark) {
884 // bitmap:
885 needed += gen->n_blocks / BITS_IN(W_);
886 // mark stack:
887 needed += gen->n_blocks / 100;
888 }
889 if (gen->compact) {
890 continue; // no additional space needed for compaction
891 } else {
892 needed += gen->n_blocks;
893 }
894 }
895 }
896 return needed;
897 }
898
899 /* ----------------------------------------------------------------------------
900 Executable memory
901
902 Executable memory must be managed separately from non-executable
903 memory. Most OSs these days require you to jump through hoops to
904 dynamically allocate executable memory, due to various security
905 measures.
906
907 Here we provide a small memory allocator for executable memory.
908 Memory is managed with a page granularity; we allocate linearly
909 in the page, and when the page is emptied (all objects on the page
910 are free) we free the page again, not forgetting to make it
911 non-executable.
912
913 TODO: The inability to handle objects bigger than BLOCK_SIZE_W means that
914 the linker cannot use allocateExec for loading object code files
915 on Windows. Once allocateExec can handle larger objects, the linker
916 should be modified to use allocateExec instead of VirtualAlloc.
917 ------------------------------------------------------------------------- */
918
919 #if defined(linux_HOST_OS)
920
921 // On Linux we need to use libffi for allocating executable memory,
922 // because it knows how to work around the restrictions put in place
923 // by SELinux.
924
925 void *allocateExec (nat bytes, void **exec_ret)
926 {
927 void **ret, **exec;
928 ACQUIRE_SM_LOCK;
929 ret = ffi_closure_alloc (sizeof(void *) + (size_t)bytes, (void**)&exec);
930 RELEASE_SM_LOCK;
931 if (ret == NULL) return ret;
932 *ret = ret; // save the address of the writable mapping, for freeExec().
933 *exec_ret = exec + 1;
934 return (ret + 1);
935 }
936
937 // freeExec gets passed the executable address, not the writable address.
938 void freeExec (void *addr)
939 {
940 void *writable;
941 writable = *((void**)addr - 1);
942 ACQUIRE_SM_LOCK;
943 ffi_closure_free (writable);
944 RELEASE_SM_LOCK
945 }
946
947 #else
948
949 void *allocateExec (nat bytes, void **exec_ret)
950 {
951 void *ret;
952 nat n;
953
954 ACQUIRE_SM_LOCK;
955
956 // round up to words.
957 n = (bytes + sizeof(W_) + 1) / sizeof(W_);
958
959 if (n+1 > BLOCK_SIZE_W) {
960 barf("allocateExec: can't handle large objects");
961 }
962
963 if (exec_block == NULL ||
964 exec_block->free + n + 1 > exec_block->start + BLOCK_SIZE_W) {
965 bdescr *bd;
966 lnat pagesize = getPageSize();
967 bd = allocGroup(stg_max(1, pagesize / BLOCK_SIZE));
968 debugTrace(DEBUG_gc, "allocate exec block %p", bd->start);
969 bd->gen_no = 0;
970 bd->flags = BF_EXEC;
971 bd->link = exec_block;
972 if (exec_block != NULL) {
973 exec_block->u.back = bd;
974 }
975 bd->u.back = NULL;
976 setExecutable(bd->start, bd->blocks * BLOCK_SIZE, rtsTrue);
977 exec_block = bd;
978 }
979 *(exec_block->free) = n; // store the size of this chunk
980 exec_block->gen_no += n; // gen_no stores the number of words allocated
981 ret = exec_block->free + 1;
982 exec_block->free += n + 1;
983
984 RELEASE_SM_LOCK
985 *exec_ret = ret;
986 return ret;
987 }
988
989 void freeExec (void *addr)
990 {
991 StgPtr p = (StgPtr)addr - 1;
992 bdescr *bd = Bdescr((StgPtr)p);
993
994 if ((bd->flags & BF_EXEC) == 0) {
995 barf("freeExec: not executable");
996 }
997
998 if (*(StgPtr)p == 0) {
999 barf("freeExec: already free?");
1000 }
1001
1002 ACQUIRE_SM_LOCK;
1003
1004 bd->gen_no -= *(StgPtr)p;
1005 *(StgPtr)p = 0;
1006
1007 if (bd->gen_no == 0) {
1008 // Free the block if it is empty, but not if it is the block at
1009 // the head of the queue.
1010 if (bd != exec_block) {
1011 debugTrace(DEBUG_gc, "free exec block %p", bd->start);
1012 dbl_link_remove(bd, &exec_block);
1013 setExecutable(bd->start, bd->blocks * BLOCK_SIZE, rtsFalse);
1014 freeGroup(bd);
1015 } else {
1016 bd->free = bd->start;
1017 }
1018 }
1019
1020 RELEASE_SM_LOCK
1021 }
1022
1023 #endif /* mingw32_HOST_OS */
1024
1025 #ifdef DEBUG
1026
1027 // handy function for use in gdb, because Bdescr() is inlined.
1028 extern bdescr *_bdescr( StgPtr p );
1029
1030 bdescr *
1031 _bdescr( StgPtr p )
1032 {
1033 return Bdescr(p);
1034 }
1035
1036 #endif