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->n_new_large_blocks = 0;
83 gen->mut_list = allocBlock();
84 gen->scavenged_large_objects = NULL;
85 gen->n_scavenged_large_blocks = 0;
90 initSpinLock(&gen->sync_large_objects);
92 gen->threads = END_TSO_QUEUE;
93 gen->old_threads = END_TSO_QUEUE;
101 if (generations != NULL) {
102 // multi-init protection
108 /* Sanity check to make sure the LOOKS_LIKE_ macros appear to be
109 * doing something reasonable.
111 /* We use the NOT_NULL variant or gcc warns that the test is always true */
112 ASSERT(LOOKS_LIKE_INFO_PTR_NOT_NULL((StgWord)&stg_BLACKHOLE_info));
113 ASSERT(LOOKS_LIKE_CLOSURE_PTR(&stg_dummy_ret_closure));
114 ASSERT(!HEAP_ALLOCED(&stg_dummy_ret_closure));
116 if (RtsFlags.GcFlags.maxHeapSize != 0 &&
117 RtsFlags.GcFlags.heapSizeSuggestion >
118 RtsFlags.GcFlags.maxHeapSize) {
119 RtsFlags.GcFlags.maxHeapSize = RtsFlags.GcFlags.heapSizeSuggestion;
122 if (RtsFlags.GcFlags.maxHeapSize != 0 &&
123 RtsFlags.GcFlags.minAllocAreaSize >
124 RtsFlags.GcFlags.maxHeapSize) {
125 errorBelch("maximum heap size (-M) is smaller than minimum alloc area size (-A)");
126 RtsFlags.GcFlags.minAllocAreaSize = RtsFlags.GcFlags.maxHeapSize;
129 initBlockAllocator();
131 #if defined(THREADED_RTS)
132 initMutex(&sm_mutex);
137 /* allocate generation info array */
138 generations = (generation *)stgMallocBytes(RtsFlags.GcFlags.generations
139 * sizeof(struct generation_),
140 "initStorage: gens");
142 /* Initialise all generations */
143 for(g = 0; g < RtsFlags.GcFlags.generations; g++) {
144 initGeneration(&generations[g], g);
147 /* A couple of convenience pointers */
148 g0 = &generations[0];
149 oldest_gen = &generations[RtsFlags.GcFlags.generations-1];
151 nurseries = stgMallocBytes(n_capabilities * sizeof(struct nursery_),
152 "initStorage: nurseries");
154 /* Set up the destination pointers in each younger gen. step */
155 for (g = 0; g < RtsFlags.GcFlags.generations-1; g++) {
156 generations[g].to = &generations[g+1];
158 oldest_gen->to = oldest_gen;
160 /* The oldest generation has one step. */
161 if (RtsFlags.GcFlags.compact || RtsFlags.GcFlags.sweep) {
162 if (RtsFlags.GcFlags.generations == 1) {
163 errorBelch("WARNING: compact/sweep is incompatible with -G1; disabled");
165 oldest_gen->mark = 1;
166 if (RtsFlags.GcFlags.compact)
167 oldest_gen->compact = 1;
171 generations[0].max_blocks = 0;
173 /* The allocation area. Policy: keep the allocation area
174 * small to begin with, even if we have a large suggested heap
175 * size. Reason: we're going to do a major collection first, and we
176 * don't want it to be a big one. This vague idea is borne out by
177 * rigorous experimental evidence.
181 weak_ptr_list = NULL;
183 revertible_caf_list = NULL;
185 /* initialise the allocate() interface */
186 alloc_blocks_lim = RtsFlags.GcFlags.minAllocAreaSize;
191 initSpinLock(&gc_alloc_block_sync);
199 IF_DEBUG(gc, statDescribeGens());
207 stat_exit(calcAllocated());
213 stgFree(generations);
215 #if defined(THREADED_RTS)
216 closeMutex(&sm_mutex);
222 /* -----------------------------------------------------------------------------
225 The entry code for every CAF does the following:
227 - builds a CAF_BLACKHOLE in the heap
228 - pushes an update frame pointing to the CAF_BLACKHOLE
229 - invokes UPD_CAF(), which:
230 - calls newCaf, below
231 - updates the CAF with a static indirection to the CAF_BLACKHOLE
233 Why do we build a BLACKHOLE in the heap rather than just updating
234 the thunk directly? It's so that we only need one kind of update
235 frame - otherwise we'd need a static version of the update frame too.
237 newCaf() does the following:
239 - it puts the CAF on the oldest generation's mut-once list.
240 This is so that we can treat the CAF as a root when collecting
243 For GHCI, we have additional requirements when dealing with CAFs:
245 - we must *retain* all dynamically-loaded CAFs ever entered,
246 just in case we need them again.
247 - we must be able to *revert* CAFs that have been evaluated, to
248 their pre-evaluated form.
250 To do this, we use an additional CAF list. When newCaf() is
251 called on a dynamically-loaded CAF, we add it to the CAF list
252 instead of the old-generation mutable list, and save away its
253 old info pointer (in caf->saved_info) for later reversion.
255 To revert all the CAFs, we traverse the CAF list and reset the
256 info pointer to caf->saved_info, then throw away the CAF list.
257 (see GC.c:revertCAFs()).
261 -------------------------------------------------------------------------- */
264 newCAF(StgClosure* caf)
272 // If we are in GHCi _and_ we are using dynamic libraries,
273 // then we can't redirect newCAF calls to newDynCAF (see below),
274 // so we make newCAF behave almost like newDynCAF.
275 // The dynamic libraries might be used by both the interpreted
276 // program and GHCi itself, so they must not be reverted.
277 // This also means that in GHCi with dynamic libraries, CAFs are not
278 // garbage collected. If this turns out to be a problem, we could
279 // do another hack here and do an address range test on caf to figure
280 // out whether it is from a dynamic library.
281 ((StgIndStatic *)caf)->saved_info = (StgInfoTable *)caf->header.info;
282 ((StgIndStatic *)caf)->static_link = caf_list;
288 /* Put this CAF on the mutable list for the old generation.
289 * This is a HACK - the IND_STATIC closure doesn't really have
290 * a mut_link field, but we pretend it has - in fact we re-use
291 * the STATIC_LINK field for the time being, because when we
292 * come to do a major GC we won't need the mut_link field
293 * any more and can use it as a STATIC_LINK.
295 ((StgIndStatic *)caf)->saved_info = NULL;
296 recordMutableGen(caf, oldest_gen->no);
302 // An alternate version of newCaf which is used for dynamically loaded
303 // object code in GHCi. In this case we want to retain *all* CAFs in
304 // the object code, because they might be demanded at any time from an
305 // expression evaluated on the command line.
306 // Also, GHCi might want to revert CAFs, so we add these to the
307 // revertible_caf_list.
309 // The linker hackily arranges that references to newCaf from dynamic
310 // code end up pointing to newDynCAF.
312 newDynCAF(StgClosure *caf)
316 ((StgIndStatic *)caf)->saved_info = (StgInfoTable *)caf->header.info;
317 ((StgIndStatic *)caf)->static_link = revertible_caf_list;
318 revertible_caf_list = caf;
323 /* -----------------------------------------------------------------------------
325 -------------------------------------------------------------------------- */
328 allocNursery (bdescr *tail, nat blocks)
333 // Allocate a nursery: we allocate fresh blocks one at a time and
334 // cons them on to the front of the list, not forgetting to update
335 // the back pointer on the tail of the list to point to the new block.
336 for (i=0; i < blocks; i++) {
339 processNursery() in LdvProfile.c assumes that every block group in
340 the nursery contains only a single block. So, if a block group is
341 given multiple blocks, change processNursery() accordingly.
345 // double-link the nursery: we might need to insert blocks
349 initBdescr(bd, g0, g0);
351 bd->free = bd->start;
359 assignNurseriesToCapabilities (void)
363 for (i = 0; i < n_capabilities; i++) {
364 capabilities[i].r.rNursery = &nurseries[i];
365 capabilities[i].r.rCurrentNursery = nurseries[i].blocks;
366 capabilities[i].r.rCurrentAlloc = NULL;
371 allocNurseries( void )
375 for (i = 0; i < n_capabilities; i++) {
376 nurseries[i].blocks =
377 allocNursery(NULL, RtsFlags.GcFlags.minAllocAreaSize);
378 nurseries[i].n_blocks =
379 RtsFlags.GcFlags.minAllocAreaSize;
381 assignNurseriesToCapabilities();
385 resetNurseries( void )
390 for (i = 0; i < n_capabilities; i++) {
391 for (bd = nurseries[i].blocks; bd; bd = bd->link) {
392 bd->free = bd->start;
393 ASSERT(bd->gen_no == 0);
394 ASSERT(bd->gen == g0);
395 IF_DEBUG(sanity,memset(bd->start, 0xaa, BLOCK_SIZE));
398 assignNurseriesToCapabilities();
402 countNurseryBlocks (void)
407 for (i = 0; i < n_capabilities; i++) {
408 blocks += nurseries[i].n_blocks;
414 resizeNursery ( nursery *nursery, nat blocks )
419 nursery_blocks = nursery->n_blocks;
420 if (nursery_blocks == blocks) return;
422 if (nursery_blocks < blocks) {
423 debugTrace(DEBUG_gc, "increasing size of nursery to %d blocks",
425 nursery->blocks = allocNursery(nursery->blocks, blocks-nursery_blocks);
430 debugTrace(DEBUG_gc, "decreasing size of nursery to %d blocks",
433 bd = nursery->blocks;
434 while (nursery_blocks > blocks) {
436 next_bd->u.back = NULL;
437 nursery_blocks -= bd->blocks; // might be a large block
441 nursery->blocks = bd;
442 // might have gone just under, by freeing a large block, so make
443 // up the difference.
444 if (nursery_blocks < blocks) {
445 nursery->blocks = allocNursery(nursery->blocks, blocks-nursery_blocks);
449 nursery->n_blocks = blocks;
450 ASSERT(countBlocks(nursery->blocks) == nursery->n_blocks);
454 // Resize each of the nurseries to the specified size.
457 resizeNurseriesFixed (nat blocks)
460 for (i = 0; i < n_capabilities; i++) {
461 resizeNursery(&nurseries[i], blocks);
466 // Resize the nurseries to the total specified size.
469 resizeNurseries (nat blocks)
471 // If there are multiple nurseries, then we just divide the number
472 // of available blocks between them.
473 resizeNurseriesFixed(blocks / n_capabilities);
477 /* -----------------------------------------------------------------------------
478 move_TSO is called to update the TSO structure after it has been
479 moved from one place to another.
480 -------------------------------------------------------------------------- */
483 move_TSO (StgTSO *src, StgTSO *dest)
487 // relocate the stack pointer...
488 diff = (StgPtr)dest - (StgPtr)src; // In *words*
489 dest->sp = (StgPtr)dest->sp + diff;
492 /* -----------------------------------------------------------------------------
493 split N blocks off the front of the given bdescr, returning the
494 new block group. We add the remainder to the large_blocks list
495 in the same step as the original block.
496 -------------------------------------------------------------------------- */
499 splitLargeBlock (bdescr *bd, nat blocks)
505 ASSERT(countBlocks(bd->gen->large_objects) == bd->gen->n_large_blocks);
507 // subtract the original number of blocks from the counter first
508 bd->gen->n_large_blocks -= bd->blocks;
510 new_bd = splitBlockGroup (bd, blocks);
511 initBdescr(new_bd, bd->gen, bd->gen->to);
512 new_bd->flags = BF_LARGE | (bd->flags & BF_EVACUATED);
513 // if new_bd is in an old generation, we have to set BF_EVACUATED
514 new_bd->free = bd->free;
515 dbl_link_onto(new_bd, &bd->gen->large_objects);
517 ASSERT(new_bd->free <= new_bd->start + new_bd->blocks * BLOCK_SIZE_W);
519 // add the new number of blocks to the counter. Due to the gaps
520 // for block descriptors, new_bd->blocks + bd->blocks might not be
521 // equal to the original bd->blocks, which is why we do it this way.
522 bd->gen->n_large_blocks += bd->blocks + new_bd->blocks;
524 ASSERT(countBlocks(bd->gen->large_objects) == bd->gen->n_large_blocks);
531 /* -----------------------------------------------------------------------------
534 This allocates memory in the current thread - it is intended for
535 use primarily from STG-land where we have a Capability. It is
536 better than allocate() because it doesn't require taking the
537 sm_mutex lock in the common case.
539 Memory is allocated directly from the nursery if possible (but not
540 from the current nursery block, so as not to interfere with
542 -------------------------------------------------------------------------- */
545 allocate (Capability *cap, lnat n)
550 if (n >= LARGE_OBJECT_THRESHOLD/sizeof(W_)) {
551 lnat req_blocks = (lnat)BLOCK_ROUND_UP(n*sizeof(W_)) / BLOCK_SIZE;
553 // Attempting to allocate an object larger than maxHeapSize
554 // should definitely be disallowed. (bug #1791)
555 if (RtsFlags.GcFlags.maxHeapSize > 0 &&
556 req_blocks >= RtsFlags.GcFlags.maxHeapSize) {
558 // heapOverflow() doesn't exit (see #2592), but we aren't
559 // in a position to do a clean shutdown here: we
560 // either have to allocate the memory or exit now.
561 // Allocating the memory would be bad, because the user
562 // has requested that we not exceed maxHeapSize, so we
564 stg_exit(EXIT_HEAPOVERFLOW);
568 bd = allocGroup(req_blocks);
569 dbl_link_onto(bd, &g0->large_objects);
570 g0->n_large_blocks += bd->blocks; // might be larger than req_blocks
571 g0->n_new_large_blocks += bd->blocks;
573 initBdescr(bd, g0, g0);
574 bd->flags = BF_LARGE;
575 bd->free = bd->start + n;
579 /* small allocation (<LARGE_OBJECT_THRESHOLD) */
581 TICK_ALLOC_HEAP_NOCTR(n);
584 bd = cap->r.rCurrentAlloc;
585 if (bd == NULL || bd->free + n > bd->start + BLOCK_SIZE_W) {
587 // The CurrentAlloc block is full, we need to find another
588 // one. First, we try taking the next block from the
590 bd = cap->r.rCurrentNursery->link;
592 if (bd == NULL || bd->free + n > bd->start + BLOCK_SIZE_W) {
593 // The nursery is empty, or the next block is already
594 // full: allocate a fresh block (we can't fail here).
597 cap->r.rNursery->n_blocks++;
599 initBdescr(bd, g0, g0);
601 // If we had to allocate a new block, then we'll GC
602 // pretty quickly now, because MAYBE_GC() will
603 // notice that CurrentNursery->link is NULL.
605 // we have a block in the nursery: take it and put
606 // it at the *front* of the nursery list, and use it
607 // to allocate() from.
608 cap->r.rCurrentNursery->link = bd->link;
609 if (bd->link != NULL) {
610 bd->link->u.back = cap->r.rCurrentNursery;
613 dbl_link_onto(bd, &cap->r.rNursery->blocks);
614 cap->r.rCurrentAlloc = bd;
615 IF_DEBUG(sanity, checkNurserySanity(cap->r.rNursery));
622 /* ---------------------------------------------------------------------------
623 Allocate a fixed/pinned object.
625 We allocate small pinned objects into a single block, allocating a
626 new block when the current one overflows. The block is chained
627 onto the large_object_list of generation 0.
629 NOTE: The GC can't in general handle pinned objects. This
630 interface is only safe to use for ByteArrays, which have no
631 pointers and don't require scavenging. It works because the
632 block's descriptor has the BF_LARGE flag set, so the block is
633 treated as a large object and chained onto various lists, rather
634 than the individual objects being copied. However, when it comes
635 to scavenge the block, the GC will only scavenge the first object.
636 The reason is that the GC can't linearly scan a block of pinned
637 objects at the moment (doing so would require using the
638 mostly-copying techniques). But since we're restricting ourselves
639 to pinned ByteArrays, not scavenging is ok.
641 This function is called by newPinnedByteArray# which immediately
642 fills the allocated memory with a MutableByteArray#.
643 ------------------------------------------------------------------------- */
646 allocatePinned (Capability *cap, lnat n)
651 // If the request is for a large object, then allocate()
652 // will give us a pinned object anyway.
653 if (n >= LARGE_OBJECT_THRESHOLD/sizeof(W_)) {
654 p = allocate(cap, n);
655 Bdescr(p)->flags |= BF_PINNED;
659 TICK_ALLOC_HEAP_NOCTR(n);
662 bd = cap->pinned_object_block;
664 // If we don't have a block of pinned objects yet, or the current
665 // one isn't large enough to hold the new object, allocate a new one.
666 if (bd == NULL || (bd->free + n) > (bd->start + BLOCK_SIZE_W)) {
668 cap->pinned_object_block = bd = allocBlock();
669 dbl_link_onto(bd, &g0->large_objects);
670 g0->n_large_blocks++;
671 g0->n_new_large_blocks++;
673 initBdescr(bd, g0, g0);
674 bd->flags = BF_PINNED | BF_LARGE;
675 bd->free = bd->start;
683 /* -----------------------------------------------------------------------------
685 -------------------------------------------------------------------------- */
688 This is the write barrier for MUT_VARs, a.k.a. IORefs. A
689 MUT_VAR_CLEAN object is not on the mutable list; a MUT_VAR_DIRTY
690 is. When written to, a MUT_VAR_CLEAN turns into a MUT_VAR_DIRTY
691 and is put on the mutable list.
694 dirty_MUT_VAR(StgRegTable *reg, StgClosure *p)
696 Capability *cap = regTableToCapability(reg);
698 if (p->header.info == &stg_MUT_VAR_CLEAN_info) {
699 p->header.info = &stg_MUT_VAR_DIRTY_info;
700 bd = Bdescr((StgPtr)p);
701 if (bd->gen_no > 0) recordMutableCap(p,cap,bd->gen_no);
705 // Setting a TSO's link field with a write barrier.
706 // It is *not* necessary to call this function when
707 // * setting the link field to END_TSO_QUEUE
708 // * putting a TSO on the blackhole_queue
709 // * setting the link field of the currently running TSO, as it
710 // will already be dirty.
712 setTSOLink (Capability *cap, StgTSO *tso, StgTSO *target)
715 if (tso->dirty == 0 && (tso->flags & TSO_LINK_DIRTY) == 0) {
716 tso->flags |= TSO_LINK_DIRTY;
717 bd = Bdescr((StgPtr)tso);
718 if (bd->gen_no > 0) recordMutableCap((StgClosure*)tso,cap,bd->gen_no);
724 dirty_TSO (Capability *cap, StgTSO *tso)
727 if (tso->dirty == 0 && (tso->flags & TSO_LINK_DIRTY) == 0) {
728 bd = Bdescr((StgPtr)tso);
729 if (bd->gen_no > 0) recordMutableCap((StgClosure*)tso,cap,bd->gen_no);
735 This is the write barrier for MVARs. An MVAR_CLEAN objects is not
736 on the mutable list; a MVAR_DIRTY is. When written to, a
737 MVAR_CLEAN turns into a MVAR_DIRTY and is put on the mutable list.
738 The check for MVAR_CLEAN is inlined at the call site for speed,
739 this really does make a difference on concurrency-heavy benchmarks
740 such as Chaneneos and cheap-concurrency.
743 dirty_MVAR(StgRegTable *reg, StgClosure *p)
745 Capability *cap = regTableToCapability(reg);
747 bd = Bdescr((StgPtr)p);
748 if (bd->gen_no > 0) recordMutableCap(p,cap,bd->gen_no);
751 /* -----------------------------------------------------------------------------
753 * -------------------------------------------------------------------------- */
755 /* -----------------------------------------------------------------------------
758 * Approximate how much we've allocated: number of blocks in the
759 * nursery + blocks allocated via allocate() - unused nusery blocks.
760 * This leaves a little slop at the end of each block.
761 * -------------------------------------------------------------------------- */
764 calcAllocated( void )
770 allocated = countNurseryBlocks() * BLOCK_SIZE_W;
772 for (i = 0; i < n_capabilities; i++) {
774 for ( bd = capabilities[i].r.rCurrentNursery->link;
775 bd != NULL; bd = bd->link ) {
776 allocated -= BLOCK_SIZE_W;
778 cap = &capabilities[i];
779 if (cap->r.rCurrentNursery->free <
780 cap->r.rCurrentNursery->start + BLOCK_SIZE_W) {
781 allocated -= (cap->r.rCurrentNursery->start + BLOCK_SIZE_W)
782 - cap->r.rCurrentNursery->free;
784 if (cap->pinned_object_block != NULL) {
785 allocated -= (cap->pinned_object_block->start + BLOCK_SIZE_W) -
786 cap->pinned_object_block->free;
790 allocated += g0->n_new_large_blocks * BLOCK_SIZE_W;
792 total_allocated += allocated;
796 /* Approximate the amount of live data in the heap. To be called just
797 * after garbage collection (see GarbageCollect()).
799 lnat calcLiveBlocks (void)
805 for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
806 /* approximate amount of live data (doesn't take into account slop
807 * at end of each block).
809 gen = &generations[g];
810 live += gen->n_large_blocks + gen->n_blocks;
815 lnat countOccupied (bdescr *bd)
820 for (; bd != NULL; bd = bd->link) {
821 ASSERT(bd->free <= bd->start + bd->blocks * BLOCK_SIZE_W);
822 words += bd->free - bd->start;
827 // Return an accurate count of the live data in the heap, excluding
829 lnat calcLiveWords (void)
836 for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
837 gen = &generations[g];
838 live += gen->n_words + countOccupied(gen->large_objects);
843 /* Approximate the number of blocks that will be needed at the next
844 * garbage collection.
846 * Assume: all data currently live will remain live. Generationss
847 * that will be collected next time will therefore need twice as many
848 * blocks since all the data will be copied.
857 for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
858 gen = &generations[g];
860 // we need at least this much space
861 needed += gen->n_blocks + gen->n_large_blocks;
863 // any additional space needed to collect this gen next time?
864 if (g == 0 || // always collect gen 0
865 (gen->n_blocks + gen->n_large_blocks > gen->max_blocks)) {
866 // we will collect this gen next time
869 needed += gen->n_blocks / BITS_IN(W_);
871 needed += gen->n_blocks / 100;
874 continue; // no additional space needed for compaction
876 needed += gen->n_blocks;
883 /* ----------------------------------------------------------------------------
886 Executable memory must be managed separately from non-executable
887 memory. Most OSs these days require you to jump through hoops to
888 dynamically allocate executable memory, due to various security
891 Here we provide a small memory allocator for executable memory.
892 Memory is managed with a page granularity; we allocate linearly
893 in the page, and when the page is emptied (all objects on the page
894 are free) we free the page again, not forgetting to make it
897 TODO: The inability to handle objects bigger than BLOCK_SIZE_W means that
898 the linker cannot use allocateExec for loading object code files
899 on Windows. Once allocateExec can handle larger objects, the linker
900 should be modified to use allocateExec instead of VirtualAlloc.
901 ------------------------------------------------------------------------- */
903 #if defined(linux_HOST_OS)
905 // On Linux we need to use libffi for allocating executable memory,
906 // because it knows how to work around the restrictions put in place
909 void *allocateExec (nat bytes, void **exec_ret)
913 ret = ffi_closure_alloc (sizeof(void *) + (size_t)bytes, (void**)&exec);
915 if (ret == NULL) return ret;
916 *ret = ret; // save the address of the writable mapping, for freeExec().
917 *exec_ret = exec + 1;
921 // freeExec gets passed the executable address, not the writable address.
922 void freeExec (void *addr)
925 writable = *((void**)addr - 1);
927 ffi_closure_free (writable);
933 void *allocateExec (nat bytes, void **exec_ret)
940 // round up to words.
941 n = (bytes + sizeof(W_) + 1) / sizeof(W_);
943 if (n+1 > BLOCK_SIZE_W) {
944 barf("allocateExec: can't handle large objects");
947 if (exec_block == NULL ||
948 exec_block->free + n + 1 > exec_block->start + BLOCK_SIZE_W) {
950 lnat pagesize = getPageSize();
951 bd = allocGroup(stg_max(1, pagesize / BLOCK_SIZE));
952 debugTrace(DEBUG_gc, "allocate exec block %p", bd->start);
955 bd->link = exec_block;
956 if (exec_block != NULL) {
957 exec_block->u.back = bd;
960 setExecutable(bd->start, bd->blocks * BLOCK_SIZE, rtsTrue);
963 *(exec_block->free) = n; // store the size of this chunk
964 exec_block->gen_no += n; // gen_no stores the number of words allocated
965 ret = exec_block->free + 1;
966 exec_block->free += n + 1;
973 void freeExec (void *addr)
975 StgPtr p = (StgPtr)addr - 1;
976 bdescr *bd = Bdescr((StgPtr)p);
978 if ((bd->flags & BF_EXEC) == 0) {
979 barf("freeExec: not executable");
982 if (*(StgPtr)p == 0) {
983 barf("freeExec: already free?");
988 bd->gen_no -= *(StgPtr)p;
991 if (bd->gen_no == 0) {
992 // Free the block if it is empty, but not if it is the block at
993 // the head of the queue.
994 if (bd != exec_block) {
995 debugTrace(DEBUG_gc, "free exec block %p", bd->start);
996 dbl_link_remove(bd, &exec_block);
997 setExecutable(bd->start, bd->blocks * BLOCK_SIZE, rtsFalse);
1000 bd->free = bd->start;
1007 #endif /* mingw32_HOST_OS */
1011 // handy function for use in gdb, because Bdescr() is inlined.
1012 extern bdescr *_bdescr( StgPtr p );