1 /* -----------------------------------------------------------------------------
3 * (c) The GHC Team, 1998-2008
5 * Storage manager front end
7 * Documentation on the architecture of the Storage Manager can be
8 * found in the online commentary:
10 * http://hackage.haskell.org/trac/ghc/wiki/Commentary/Rts/Storage
12 * ---------------------------------------------------------------------------*/
14 #include "PosixSource.h"
20 #include "BlockAlloc.h"
24 #include "Capability.h"
26 #include "RetainerProfile.h" // for counting memory blocks (memInventory)
37 * All these globals require sm_mutex to access in THREADED_RTS mode.
39 StgClosure *caf_list = NULL;
40 StgClosure *revertible_caf_list = NULL;
43 nat alloc_blocks_lim; /* GC if n_large_blocks in any nursery
48 generation *generations = NULL; /* all the generations */
49 generation *g0 = NULL; /* generation 0, for convenience */
50 generation *oldest_gen = NULL; /* oldest generation, for convenience */
52 ullong total_allocated = 0; /* total memory allocated during run */
54 nursery *nurseries = NULL; /* array of nurseries, size == n_capabilities */
58 * Storage manager mutex: protects all the above state from
59 * simultaneous access by two STG threads.
64 static void allocNurseries ( void );
67 initGeneration (generation *gen, int g)
71 gen->par_collections = 0;
72 gen->failed_promotions = 0;
77 gen->live_estimate = 0;
78 gen->old_blocks = NULL;
79 gen->n_old_blocks = 0;
80 gen->large_objects = NULL;
81 gen->n_large_blocks = 0;
82 gen->mut_list = allocBlock();
83 gen->scavenged_large_objects = NULL;
84 gen->n_scavenged_large_blocks = 0;
89 initSpinLock(&gen->sync_large_objects);
91 gen->threads = END_TSO_QUEUE;
92 gen->old_threads = END_TSO_QUEUE;
100 if (generations != NULL) {
101 // multi-init protection
107 /* Sanity check to make sure the LOOKS_LIKE_ macros appear to be
108 * doing something reasonable.
110 /* We use the NOT_NULL variant or gcc warns that the test is always true */
111 ASSERT(LOOKS_LIKE_INFO_PTR_NOT_NULL((StgWord)&stg_BLACKHOLE_info));
112 ASSERT(LOOKS_LIKE_CLOSURE_PTR(&stg_dummy_ret_closure));
113 ASSERT(!HEAP_ALLOCED(&stg_dummy_ret_closure));
115 if (RtsFlags.GcFlags.maxHeapSize != 0 &&
116 RtsFlags.GcFlags.heapSizeSuggestion >
117 RtsFlags.GcFlags.maxHeapSize) {
118 RtsFlags.GcFlags.maxHeapSize = RtsFlags.GcFlags.heapSizeSuggestion;
121 if (RtsFlags.GcFlags.maxHeapSize != 0 &&
122 RtsFlags.GcFlags.minAllocAreaSize >
123 RtsFlags.GcFlags.maxHeapSize) {
124 errorBelch("maximum heap size (-M) is smaller than minimum alloc area size (-A)");
125 RtsFlags.GcFlags.minAllocAreaSize = RtsFlags.GcFlags.maxHeapSize;
128 initBlockAllocator();
130 #if defined(THREADED_RTS)
131 initMutex(&sm_mutex);
136 /* allocate generation info array */
137 generations = (generation *)stgMallocBytes(RtsFlags.GcFlags.generations
138 * sizeof(struct generation_),
139 "initStorage: gens");
141 /* Initialise all generations */
142 for(g = 0; g < RtsFlags.GcFlags.generations; g++) {
143 initGeneration(&generations[g], g);
146 /* A couple of convenience pointers */
147 g0 = &generations[0];
148 oldest_gen = &generations[RtsFlags.GcFlags.generations-1];
150 nurseries = stgMallocBytes(n_capabilities * sizeof(struct nursery_),
151 "initStorage: nurseries");
153 /* Set up the destination pointers in each younger gen. step */
154 for (g = 0; g < RtsFlags.GcFlags.generations-1; g++) {
155 generations[g].to = &generations[g+1];
157 oldest_gen->to = oldest_gen;
159 /* The oldest generation has one step. */
160 if (RtsFlags.GcFlags.compact || RtsFlags.GcFlags.sweep) {
161 if (RtsFlags.GcFlags.generations == 1) {
162 errorBelch("WARNING: compact/sweep is incompatible with -G1; disabled");
164 oldest_gen->mark = 1;
165 if (RtsFlags.GcFlags.compact)
166 oldest_gen->compact = 1;
170 generations[0].max_blocks = 0;
172 /* The allocation area. Policy: keep the allocation area
173 * small to begin with, even if we have a large suggested heap
174 * size. Reason: we're going to do a major collection first, and we
175 * don't want it to be a big one. This vague idea is borne out by
176 * rigorous experimental evidence.
180 weak_ptr_list = NULL;
182 revertible_caf_list = NULL;
184 /* initialise the allocate() interface */
185 alloc_blocks_lim = RtsFlags.GcFlags.minAllocAreaSize;
190 initSpinLock(&gc_alloc_block_sync);
198 IF_DEBUG(gc, statDescribeGens());
206 stat_exit(calcAllocated());
212 stgFree(generations);
214 #if defined(THREADED_RTS)
215 closeMutex(&sm_mutex);
221 /* -----------------------------------------------------------------------------
224 The entry code for every CAF does the following:
226 - builds a CAF_BLACKHOLE in the heap
227 - pushes an update frame pointing to the CAF_BLACKHOLE
228 - invokes UPD_CAF(), which:
229 - calls newCaf, below
230 - updates the CAF with a static indirection to the CAF_BLACKHOLE
232 Why do we build a BLACKHOLE in the heap rather than just updating
233 the thunk directly? It's so that we only need one kind of update
234 frame - otherwise we'd need a static version of the update frame too.
236 newCaf() does the following:
238 - it puts the CAF on the oldest generation's mut-once list.
239 This is so that we can treat the CAF as a root when collecting
242 For GHCI, we have additional requirements when dealing with CAFs:
244 - we must *retain* all dynamically-loaded CAFs ever entered,
245 just in case we need them again.
246 - we must be able to *revert* CAFs that have been evaluated, to
247 their pre-evaluated form.
249 To do this, we use an additional CAF list. When newCaf() is
250 called on a dynamically-loaded CAF, we add it to the CAF list
251 instead of the old-generation mutable list, and save away its
252 old info pointer (in caf->saved_info) for later reversion.
254 To revert all the CAFs, we traverse the CAF list and reset the
255 info pointer to caf->saved_info, then throw away the CAF list.
256 (see GC.c:revertCAFs()).
260 -------------------------------------------------------------------------- */
263 newCAF(StgClosure* caf)
271 // If we are in GHCi _and_ we are using dynamic libraries,
272 // then we can't redirect newCAF calls to newDynCAF (see below),
273 // so we make newCAF behave almost like newDynCAF.
274 // The dynamic libraries might be used by both the interpreted
275 // program and GHCi itself, so they must not be reverted.
276 // This also means that in GHCi with dynamic libraries, CAFs are not
277 // garbage collected. If this turns out to be a problem, we could
278 // do another hack here and do an address range test on caf to figure
279 // out whether it is from a dynamic library.
280 ((StgIndStatic *)caf)->saved_info = (StgInfoTable *)caf->header.info;
281 ((StgIndStatic *)caf)->static_link = caf_list;
287 /* Put this CAF on the mutable list for the old generation.
288 * This is a HACK - the IND_STATIC closure doesn't really have
289 * a mut_link field, but we pretend it has - in fact we re-use
290 * the STATIC_LINK field for the time being, because when we
291 * come to do a major GC we won't need the mut_link field
292 * any more and can use it as a STATIC_LINK.
294 ((StgIndStatic *)caf)->saved_info = NULL;
295 recordMutableGen(caf, oldest_gen->no);
301 // An alternate version of newCaf which is used for dynamically loaded
302 // object code in GHCi. In this case we want to retain *all* CAFs in
303 // the object code, because they might be demanded at any time from an
304 // expression evaluated on the command line.
305 // Also, GHCi might want to revert CAFs, so we add these to the
306 // revertible_caf_list.
308 // The linker hackily arranges that references to newCaf from dynamic
309 // code end up pointing to newDynCAF.
311 newDynCAF(StgClosure *caf)
315 ((StgIndStatic *)caf)->saved_info = (StgInfoTable *)caf->header.info;
316 ((StgIndStatic *)caf)->static_link = revertible_caf_list;
317 revertible_caf_list = caf;
322 /* -----------------------------------------------------------------------------
324 -------------------------------------------------------------------------- */
327 allocNursery (bdescr *tail, nat blocks)
332 // Allocate a nursery: we allocate fresh blocks one at a time and
333 // cons them on to the front of the list, not forgetting to update
334 // the back pointer on the tail of the list to point to the new block.
335 for (i=0; i < blocks; i++) {
338 processNursery() in LdvProfile.c assumes that every block group in
339 the nursery contains only a single block. So, if a block group is
340 given multiple blocks, change processNursery() accordingly.
344 // double-link the nursery: we might need to insert blocks
348 initBdescr(bd, g0, g0);
350 bd->free = bd->start;
358 assignNurseriesToCapabilities (void)
362 for (i = 0; i < n_capabilities; i++) {
363 capabilities[i].r.rNursery = &nurseries[i];
364 capabilities[i].r.rCurrentNursery = nurseries[i].blocks;
365 capabilities[i].r.rCurrentAlloc = NULL;
370 allocNurseries( void )
374 for (i = 0; i < n_capabilities; i++) {
375 nurseries[i].blocks =
376 allocNursery(NULL, RtsFlags.GcFlags.minAllocAreaSize);
377 nurseries[i].n_blocks =
378 RtsFlags.GcFlags.minAllocAreaSize;
380 assignNurseriesToCapabilities();
384 resetNurseries( void )
389 for (i = 0; i < n_capabilities; i++) {
390 for (bd = nurseries[i].blocks; bd; bd = bd->link) {
391 bd->free = bd->start;
392 ASSERT(bd->gen_no == 0);
393 ASSERT(bd->gen == g0);
394 IF_DEBUG(sanity,memset(bd->start, 0xaa, BLOCK_SIZE));
397 assignNurseriesToCapabilities();
401 countNurseryBlocks (void)
406 for (i = 0; i < n_capabilities; i++) {
407 blocks += nurseries[i].n_blocks;
413 resizeNursery ( nursery *nursery, nat blocks )
418 nursery_blocks = nursery->n_blocks;
419 if (nursery_blocks == blocks) return;
421 if (nursery_blocks < blocks) {
422 debugTrace(DEBUG_gc, "increasing size of nursery to %d blocks",
424 nursery->blocks = allocNursery(nursery->blocks, blocks-nursery_blocks);
429 debugTrace(DEBUG_gc, "decreasing size of nursery to %d blocks",
432 bd = nursery->blocks;
433 while (nursery_blocks > blocks) {
435 next_bd->u.back = NULL;
436 nursery_blocks -= bd->blocks; // might be a large block
440 nursery->blocks = bd;
441 // might have gone just under, by freeing a large block, so make
442 // up the difference.
443 if (nursery_blocks < blocks) {
444 nursery->blocks = allocNursery(nursery->blocks, blocks-nursery_blocks);
448 nursery->n_blocks = blocks;
449 ASSERT(countBlocks(nursery->blocks) == nursery->n_blocks);
453 // Resize each of the nurseries to the specified size.
456 resizeNurseriesFixed (nat blocks)
459 for (i = 0; i < n_capabilities; i++) {
460 resizeNursery(&nurseries[i], blocks);
465 // Resize the nurseries to the total specified size.
468 resizeNurseries (nat blocks)
470 // If there are multiple nurseries, then we just divide the number
471 // of available blocks between them.
472 resizeNurseriesFixed(blocks / n_capabilities);
476 /* -----------------------------------------------------------------------------
477 move_TSO is called to update the TSO structure after it has been
478 moved from one place to another.
479 -------------------------------------------------------------------------- */
482 move_TSO (StgTSO *src, StgTSO *dest)
486 // relocate the stack pointer...
487 diff = (StgPtr)dest - (StgPtr)src; // In *words*
488 dest->sp = (StgPtr)dest->sp + diff;
491 /* -----------------------------------------------------------------------------
492 split N blocks off the front of the given bdescr, returning the
493 new block group. We add the remainder to the large_blocks list
494 in the same step as the original block.
495 -------------------------------------------------------------------------- */
498 splitLargeBlock (bdescr *bd, nat blocks)
504 ASSERT(countBlocks(bd->gen->large_objects) == bd->gen->n_large_blocks);
506 // subtract the original number of blocks from the counter first
507 bd->gen->n_large_blocks -= bd->blocks;
509 new_bd = splitBlockGroup (bd, blocks);
510 initBdescr(new_bd, bd->gen, bd->gen->to);
511 new_bd->flags = BF_LARGE | (bd->flags & BF_EVACUATED);
512 // if new_bd is in an old generation, we have to set BF_EVACUATED
513 new_bd->free = bd->free;
514 dbl_link_onto(new_bd, &bd->gen->large_objects);
516 ASSERT(new_bd->free <= new_bd->start + new_bd->blocks * BLOCK_SIZE_W);
518 // add the new number of blocks to the counter. Due to the gaps
519 // for block descriptors, new_bd->blocks + bd->blocks might not be
520 // equal to the original bd->blocks, which is why we do it this way.
521 bd->gen->n_large_blocks += bd->blocks + new_bd->blocks;
523 ASSERT(countBlocks(bd->gen->large_objects) == bd->gen->n_large_blocks);
530 /* -----------------------------------------------------------------------------
533 This allocates memory in the current thread - it is intended for
534 use primarily from STG-land where we have a Capability. It is
535 better than allocate() because it doesn't require taking the
536 sm_mutex lock in the common case.
538 Memory is allocated directly from the nursery if possible (but not
539 from the current nursery block, so as not to interfere with
541 -------------------------------------------------------------------------- */
544 allocate (Capability *cap, lnat n)
549 if (n >= LARGE_OBJECT_THRESHOLD/sizeof(W_)) {
550 lnat req_blocks = (lnat)BLOCK_ROUND_UP(n*sizeof(W_)) / BLOCK_SIZE;
552 // Attempting to allocate an object larger than maxHeapSize
553 // should definitely be disallowed. (bug #1791)
554 if (RtsFlags.GcFlags.maxHeapSize > 0 &&
555 req_blocks >= RtsFlags.GcFlags.maxHeapSize) {
557 // heapOverflow() doesn't exit (see #2592), but we aren't
558 // in a position to do a clean shutdown here: we
559 // either have to allocate the memory or exit now.
560 // Allocating the memory would be bad, because the user
561 // has requested that we not exceed maxHeapSize, so we
563 stg_exit(EXIT_HEAPOVERFLOW);
567 bd = allocGroup(req_blocks);
568 dbl_link_onto(bd, &g0->large_objects);
569 g0->n_large_blocks += bd->blocks; // might be larger than req_blocks
571 initBdescr(bd, g0, g0);
572 bd->flags = BF_LARGE;
573 bd->free = bd->start + n;
577 /* small allocation (<LARGE_OBJECT_THRESHOLD) */
579 TICK_ALLOC_HEAP_NOCTR(n);
582 bd = cap->r.rCurrentAlloc;
583 if (bd == NULL || bd->free + n > bd->start + BLOCK_SIZE_W) {
585 // The CurrentAlloc block is full, we need to find another
586 // one. First, we try taking the next block from the
588 bd = cap->r.rCurrentNursery->link;
590 if (bd == NULL || bd->free + n > bd->start + BLOCK_SIZE_W) {
591 // The nursery is empty, or the next block is already
592 // full: allocate a fresh block (we can't fail here).
595 cap->r.rNursery->n_blocks++;
597 initBdescr(bd, g0, g0);
599 // If we had to allocate a new block, then we'll GC
600 // pretty quickly now, because MAYBE_GC() will
601 // notice that CurrentNursery->link is NULL.
603 // we have a block in the nursery: take it and put
604 // it at the *front* of the nursery list, and use it
605 // to allocate() from.
606 cap->r.rCurrentNursery->link = bd->link;
607 if (bd->link != NULL) {
608 bd->link->u.back = cap->r.rCurrentNursery;
611 dbl_link_onto(bd, &cap->r.rNursery->blocks);
612 cap->r.rCurrentAlloc = bd;
613 IF_DEBUG(sanity, checkNurserySanity(cap->r.rNursery));
620 /* ---------------------------------------------------------------------------
621 Allocate a fixed/pinned object.
623 We allocate small pinned objects into a single block, allocating a
624 new block when the current one overflows. The block is chained
625 onto the large_object_list of generation 0.
627 NOTE: The GC can't in general handle pinned objects. This
628 interface is only safe to use for ByteArrays, which have no
629 pointers and don't require scavenging. It works because the
630 block's descriptor has the BF_LARGE flag set, so the block is
631 treated as a large object and chained onto various lists, rather
632 than the individual objects being copied. However, when it comes
633 to scavenge the block, the GC will only scavenge the first object.
634 The reason is that the GC can't linearly scan a block of pinned
635 objects at the moment (doing so would require using the
636 mostly-copying techniques). But since we're restricting ourselves
637 to pinned ByteArrays, not scavenging is ok.
639 This function is called by newPinnedByteArray# which immediately
640 fills the allocated memory with a MutableByteArray#.
641 ------------------------------------------------------------------------- */
644 allocatePinned (Capability *cap, lnat n)
649 // If the request is for a large object, then allocate()
650 // will give us a pinned object anyway.
651 if (n >= LARGE_OBJECT_THRESHOLD/sizeof(W_)) {
652 p = allocate(cap, n);
653 Bdescr(p)->flags |= BF_PINNED;
657 TICK_ALLOC_HEAP_NOCTR(n);
660 bd = cap->pinned_object_block;
662 // If we don't have a block of pinned objects yet, or the current
663 // one isn't large enough to hold the new object, allocate a new one.
664 if (bd == NULL || (bd->free + n) > (bd->start + BLOCK_SIZE_W)) {
666 cap->pinned_object_block = bd = allocBlock();
667 dbl_link_onto(bd, &g0->large_objects);
668 g0->n_large_blocks++;
670 initBdescr(bd, g0, g0);
671 bd->flags = BF_PINNED | BF_LARGE;
672 bd->free = bd->start;
680 /* -----------------------------------------------------------------------------
682 -------------------------------------------------------------------------- */
685 This is the write barrier for MUT_VARs, a.k.a. IORefs. A
686 MUT_VAR_CLEAN object is not on the mutable list; a MUT_VAR_DIRTY
687 is. When written to, a MUT_VAR_CLEAN turns into a MUT_VAR_DIRTY
688 and is put on the mutable list.
691 dirty_MUT_VAR(StgRegTable *reg, StgClosure *p)
693 Capability *cap = regTableToCapability(reg);
695 if (p->header.info == &stg_MUT_VAR_CLEAN_info) {
696 p->header.info = &stg_MUT_VAR_DIRTY_info;
697 bd = Bdescr((StgPtr)p);
698 if (bd->gen_no > 0) recordMutableCap(p,cap,bd->gen_no);
702 // Setting a TSO's link field with a write barrier.
703 // It is *not* necessary to call this function when
704 // * setting the link field to END_TSO_QUEUE
705 // * putting a TSO on the blackhole_queue
706 // * setting the link field of the currently running TSO, as it
707 // will already be dirty.
709 setTSOLink (Capability *cap, StgTSO *tso, StgTSO *target)
712 if (tso->dirty == 0 && (tso->flags & TSO_LINK_DIRTY) == 0) {
713 tso->flags |= TSO_LINK_DIRTY;
714 bd = Bdescr((StgPtr)tso);
715 if (bd->gen_no > 0) recordMutableCap((StgClosure*)tso,cap,bd->gen_no);
721 dirty_TSO (Capability *cap, StgTSO *tso)
724 if (tso->dirty == 0 && (tso->flags & TSO_LINK_DIRTY) == 0) {
725 bd = Bdescr((StgPtr)tso);
726 if (bd->gen_no > 0) recordMutableCap((StgClosure*)tso,cap,bd->gen_no);
732 This is the write barrier for MVARs. An MVAR_CLEAN objects is not
733 on the mutable list; a MVAR_DIRTY is. When written to, a
734 MVAR_CLEAN turns into a MVAR_DIRTY and is put on the mutable list.
735 The check for MVAR_CLEAN is inlined at the call site for speed,
736 this really does make a difference on concurrency-heavy benchmarks
737 such as Chaneneos and cheap-concurrency.
740 dirty_MVAR(StgRegTable *reg, StgClosure *p)
742 Capability *cap = regTableToCapability(reg);
744 bd = Bdescr((StgPtr)p);
745 if (bd->gen_no > 0) recordMutableCap(p,cap,bd->gen_no);
748 /* -----------------------------------------------------------------------------
750 * -------------------------------------------------------------------------- */
752 /* -----------------------------------------------------------------------------
755 * Approximate how much we've allocated: number of blocks in the
756 * nursery + blocks allocated via allocate() - unused nusery blocks.
757 * This leaves a little slop at the end of each block.
758 * -------------------------------------------------------------------------- */
761 calcAllocated( void )
767 allocated = countNurseryBlocks() * BLOCK_SIZE_W;
769 for (i = 0; i < n_capabilities; i++) {
771 for ( bd = capabilities[i].r.rCurrentNursery->link;
772 bd != NULL; bd = bd->link ) {
773 allocated -= BLOCK_SIZE_W;
775 cap = &capabilities[i];
776 if (cap->r.rCurrentNursery->free <
777 cap->r.rCurrentNursery->start + BLOCK_SIZE_W) {
778 allocated -= (cap->r.rCurrentNursery->start + BLOCK_SIZE_W)
779 - cap->r.rCurrentNursery->free;
781 if (cap->pinned_object_block != NULL) {
782 allocated -= (cap->pinned_object_block->start + BLOCK_SIZE_W) -
783 cap->pinned_object_block->free;
787 total_allocated += allocated;
791 /* Approximate the amount of live data in the heap. To be called just
792 * after garbage collection (see GarbageCollect()).
794 lnat calcLiveBlocks (void)
800 for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
801 /* approximate amount of live data (doesn't take into account slop
802 * at end of each block).
804 gen = &generations[g];
805 live += gen->n_large_blocks + gen->n_blocks;
810 lnat countOccupied (bdescr *bd)
815 for (; bd != NULL; bd = bd->link) {
816 ASSERT(bd->free <= bd->start + bd->blocks * BLOCK_SIZE_W);
817 words += bd->free - bd->start;
822 // Return an accurate count of the live data in the heap, excluding
824 lnat calcLiveWords (void)
831 for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
832 gen = &generations[g];
833 live += gen->n_words + countOccupied(gen->large_objects);
838 /* Approximate the number of blocks that will be needed at the next
839 * garbage collection.
841 * Assume: all data currently live will remain live. Generationss
842 * that will be collected next time will therefore need twice as many
843 * blocks since all the data will be copied.
852 for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
853 gen = &generations[g];
855 // we need at least this much space
856 needed += gen->n_blocks + gen->n_large_blocks;
858 // any additional space needed to collect this gen next time?
859 if (g == 0 || // always collect gen 0
860 (gen->n_blocks + gen->n_large_blocks > gen->max_blocks)) {
861 // we will collect this gen next time
864 needed += gen->n_blocks / BITS_IN(W_);
866 needed += gen->n_blocks / 100;
869 continue; // no additional space needed for compaction
871 needed += gen->n_blocks;
878 /* ----------------------------------------------------------------------------
881 Executable memory must be managed separately from non-executable
882 memory. Most OSs these days require you to jump through hoops to
883 dynamically allocate executable memory, due to various security
886 Here we provide a small memory allocator for executable memory.
887 Memory is managed with a page granularity; we allocate linearly
888 in the page, and when the page is emptied (all objects on the page
889 are free) we free the page again, not forgetting to make it
892 TODO: The inability to handle objects bigger than BLOCK_SIZE_W means that
893 the linker cannot use allocateExec for loading object code files
894 on Windows. Once allocateExec can handle larger objects, the linker
895 should be modified to use allocateExec instead of VirtualAlloc.
896 ------------------------------------------------------------------------- */
898 #if defined(linux_HOST_OS)
900 // On Linux we need to use libffi for allocating executable memory,
901 // because it knows how to work around the restrictions put in place
904 void *allocateExec (nat bytes, void **exec_ret)
908 ret = ffi_closure_alloc (sizeof(void *) + (size_t)bytes, (void**)&exec);
910 if (ret == NULL) return ret;
911 *ret = ret; // save the address of the writable mapping, for freeExec().
912 *exec_ret = exec + 1;
916 // freeExec gets passed the executable address, not the writable address.
917 void freeExec (void *addr)
920 writable = *((void**)addr - 1);
922 ffi_closure_free (writable);
928 void *allocateExec (nat bytes, void **exec_ret)
935 // round up to words.
936 n = (bytes + sizeof(W_) + 1) / sizeof(W_);
938 if (n+1 > BLOCK_SIZE_W) {
939 barf("allocateExec: can't handle large objects");
942 if (exec_block == NULL ||
943 exec_block->free + n + 1 > exec_block->start + BLOCK_SIZE_W) {
945 lnat pagesize = getPageSize();
946 bd = allocGroup(stg_max(1, pagesize / BLOCK_SIZE));
947 debugTrace(DEBUG_gc, "allocate exec block %p", bd->start);
950 bd->link = exec_block;
951 if (exec_block != NULL) {
952 exec_block->u.back = bd;
955 setExecutable(bd->start, bd->blocks * BLOCK_SIZE, rtsTrue);
958 *(exec_block->free) = n; // store the size of this chunk
959 exec_block->gen_no += n; // gen_no stores the number of words allocated
960 ret = exec_block->free + 1;
961 exec_block->free += n + 1;
968 void freeExec (void *addr)
970 StgPtr p = (StgPtr)addr - 1;
971 bdescr *bd = Bdescr((StgPtr)p);
973 if ((bd->flags & BF_EXEC) == 0) {
974 barf("freeExec: not executable");
977 if (*(StgPtr)p == 0) {
978 barf("freeExec: already free?");
983 bd->gen_no -= *(StgPtr)p;
986 if (bd->gen_no == 0) {
987 // Free the block if it is empty, but not if it is the block at
988 // the head of the queue.
989 if (bd != exec_block) {
990 debugTrace(DEBUG_gc, "free exec block %p", bd->start);
991 dbl_link_remove(bd, &exec_block);
992 setExecutable(bd->start, bd->blocks * BLOCK_SIZE, rtsFalse);
995 bd->free = bd->start;
1002 #endif /* mingw32_HOST_OS */
1006 // handy function for use in gdb, because Bdescr() is inlined.
1007 extern bdescr *_bdescr( StgPtr p );