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 nursery *nurseries = NULL; /* array of nurseries, size == n_capabilities */
56 * Storage manager mutex: protects all the above state from
57 * simultaneous access by two STG threads.
62 static void allocNurseries ( void );
65 initGeneration (generation *gen, int g)
69 gen->par_collections = 0;
70 gen->failed_promotions = 0;
75 gen->live_estimate = 0;
76 gen->old_blocks = NULL;
77 gen->n_old_blocks = 0;
78 gen->large_objects = NULL;
79 gen->n_large_blocks = 0;
80 gen->n_new_large_blocks = 0;
81 gen->mut_list = allocBlock();
82 gen->scavenged_large_objects = NULL;
83 gen->n_scavenged_large_blocks = 0;
88 initSpinLock(&gen->sync_large_objects);
90 gen->threads = END_TSO_QUEUE;
91 gen->old_threads = END_TSO_QUEUE;
99 if (generations != NULL) {
100 // multi-init protection
106 /* Sanity check to make sure the LOOKS_LIKE_ macros appear to be
107 * doing something reasonable.
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_BLACKHOLE_info));
111 ASSERT(LOOKS_LIKE_CLOSURE_PTR(&stg_dummy_ret_closure));
112 ASSERT(!HEAP_ALLOCED(&stg_dummy_ret_closure));
114 if (RtsFlags.GcFlags.maxHeapSize != 0 &&
115 RtsFlags.GcFlags.heapSizeSuggestion >
116 RtsFlags.GcFlags.maxHeapSize) {
117 RtsFlags.GcFlags.maxHeapSize = RtsFlags.GcFlags.heapSizeSuggestion;
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;
127 initBlockAllocator();
129 #if defined(THREADED_RTS)
130 initMutex(&sm_mutex);
135 /* allocate generation info array */
136 generations = (generation *)stgMallocBytes(RtsFlags.GcFlags.generations
137 * sizeof(struct generation_),
138 "initStorage: gens");
140 /* Initialise all generations */
141 for(g = 0; g < RtsFlags.GcFlags.generations; g++) {
142 initGeneration(&generations[g], g);
145 /* A couple of convenience pointers */
146 g0 = &generations[0];
147 oldest_gen = &generations[RtsFlags.GcFlags.generations-1];
149 nurseries = stgMallocBytes(n_capabilities * sizeof(struct nursery_),
150 "initStorage: nurseries");
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];
156 oldest_gen->to = oldest_gen;
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");
163 oldest_gen->mark = 1;
164 if (RtsFlags.GcFlags.compact)
165 oldest_gen->compact = 1;
169 generations[0].max_blocks = 0;
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.
179 weak_ptr_list = NULL;
181 revertible_caf_list = NULL;
183 /* initialise the allocate() interface */
184 alloc_blocks_lim = RtsFlags.GcFlags.minAllocAreaSize;
189 initSpinLock(&gc_alloc_block_sync);
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();
204 IF_DEBUG(gc, statDescribeGens());
212 stat_exit(calcAllocated());
218 stgFree(generations);
220 #if defined(THREADED_RTS)
221 closeMutex(&sm_mutex);
227 /* -----------------------------------------------------------------------------
230 The entry code for every CAF does the following:
232 - builds a CAF_BLACKHOLE in the heap
233 - pushes an update frame pointing to the CAF_BLACKHOLE
234 - invokes UPD_CAF(), which:
235 - calls newCaf, below
236 - updates the CAF with a static indirection to the CAF_BLACKHOLE
238 Why do we build a BLACKHOLE in the heap rather than just updating
239 the thunk directly? It's so that we only need one kind of update
240 frame - otherwise we'd need a static version of the update frame too.
242 newCaf() does the following:
244 - it puts the CAF on the oldest generation's mutable list.
245 This is so that we treat the CAF as a root when collecting
248 For GHCI, we have additional requirements when dealing with CAFs:
250 - we must *retain* all dynamically-loaded CAFs ever entered,
251 just in case we need them again.
252 - we must be able to *revert* CAFs that have been evaluated, to
253 their pre-evaluated form.
255 To do this, we use an additional CAF list. When newCaf() is
256 called on a dynamically-loaded CAF, we add it to the CAF list
257 instead of the old-generation mutable list, and save away its
258 old info pointer (in caf->saved_info) for later reversion.
260 To revert all the CAFs, we traverse the CAF list and reset the
261 info pointer to caf->saved_info, then throw away the CAF list.
262 (see GC.c:revertCAFs()).
266 -------------------------------------------------------------------------- */
269 newCAF(StgRegTable *reg, StgClosure* caf)
275 // If we are in GHCi _and_ we are using dynamic libraries,
276 // then we can't redirect newCAF calls to newDynCAF (see below),
277 // so we make newCAF behave almost like newDynCAF.
278 // The dynamic libraries might be used by both the interpreted
279 // program and GHCi itself, so they must not be reverted.
280 // This also means that in GHCi with dynamic libraries, CAFs are not
281 // garbage collected. If this turns out to be a problem, we could
282 // do another hack here and do an address range test on caf to figure
283 // out whether it is from a dynamic library.
284 ((StgIndStatic *)caf)->saved_info = (StgInfoTable *)caf->header.info;
286 ACQUIRE_SM_LOCK; // caf_list is global, locked by sm_mutex
287 ((StgIndStatic *)caf)->static_link = caf_list;
294 // Put this CAF on the mutable list for the old generation.
295 ((StgIndStatic *)caf)->saved_info = NULL;
296 recordMutableCap(caf, regTableToCapability(reg), oldest_gen->no);
300 // An alternate version of newCaf which is used for dynamically loaded
301 // object code in GHCi. In this case we want to retain *all* CAFs in
302 // the object code, because they might be demanded at any time from an
303 // expression evaluated on the command line.
304 // Also, GHCi might want to revert CAFs, so we add these to the
305 // revertible_caf_list.
307 // The linker hackily arranges that references to newCaf from dynamic
308 // code end up pointing to newDynCAF.
310 newDynCAF (StgRegTable *reg STG_UNUSED, StgClosure *caf)
314 ((StgIndStatic *)caf)->saved_info = (StgInfoTable *)caf->header.info;
315 ((StgIndStatic *)caf)->static_link = revertible_caf_list;
316 revertible_caf_list = caf;
321 /* -----------------------------------------------------------------------------
323 -------------------------------------------------------------------------- */
326 allocNursery (bdescr *tail, nat blocks)
331 // Allocate a nursery: we allocate fresh blocks one at a time and
332 // cons them on to the front of the list, not forgetting to update
333 // the back pointer on the tail of the list to point to the new block.
334 for (i=0; i < blocks; i++) {
337 processNursery() in LdvProfile.c assumes that every block group in
338 the nursery contains only a single block. So, if a block group is
339 given multiple blocks, change processNursery() accordingly.
343 // double-link the nursery: we might need to insert blocks
347 initBdescr(bd, g0, g0);
349 bd->free = bd->start;
357 assignNurseriesToCapabilities (void)
361 for (i = 0; i < n_capabilities; i++) {
362 capabilities[i].r.rNursery = &nurseries[i];
363 capabilities[i].r.rCurrentNursery = nurseries[i].blocks;
364 capabilities[i].r.rCurrentAlloc = NULL;
369 allocNurseries( void )
373 for (i = 0; i < n_capabilities; i++) {
374 nurseries[i].blocks =
375 allocNursery(NULL, RtsFlags.GcFlags.minAllocAreaSize);
376 nurseries[i].n_blocks =
377 RtsFlags.GcFlags.minAllocAreaSize;
379 assignNurseriesToCapabilities();
383 resetNurseries( void )
388 for (i = 0; i < n_capabilities; i++) {
389 for (bd = nurseries[i].blocks; bd; bd = bd->link) {
390 bd->free = bd->start;
391 ASSERT(bd->gen_no == 0);
392 ASSERT(bd->gen == g0);
393 IF_DEBUG(sanity,memset(bd->start, 0xaa, BLOCK_SIZE));
396 assignNurseriesToCapabilities();
400 countNurseryBlocks (void)
405 for (i = 0; i < n_capabilities; i++) {
406 blocks += nurseries[i].n_blocks;
412 resizeNursery ( nursery *nursery, nat blocks )
417 nursery_blocks = nursery->n_blocks;
418 if (nursery_blocks == blocks) return;
420 if (nursery_blocks < blocks) {
421 debugTrace(DEBUG_gc, "increasing size of nursery to %d blocks",
423 nursery->blocks = allocNursery(nursery->blocks, blocks-nursery_blocks);
428 debugTrace(DEBUG_gc, "decreasing size of nursery to %d blocks",
431 bd = nursery->blocks;
432 while (nursery_blocks > blocks) {
434 next_bd->u.back = NULL;
435 nursery_blocks -= bd->blocks; // might be a large block
439 nursery->blocks = bd;
440 // might have gone just under, by freeing a large block, so make
441 // up the difference.
442 if (nursery_blocks < blocks) {
443 nursery->blocks = allocNursery(nursery->blocks, blocks-nursery_blocks);
447 nursery->n_blocks = blocks;
448 ASSERT(countBlocks(nursery->blocks) == nursery->n_blocks);
452 // Resize each of the nurseries to the specified size.
455 resizeNurseriesFixed (nat blocks)
458 for (i = 0; i < n_capabilities; i++) {
459 resizeNursery(&nurseries[i], blocks);
464 // Resize the nurseries to the total specified size.
467 resizeNurseries (nat blocks)
469 // If there are multiple nurseries, then we just divide the number
470 // of available blocks between them.
471 resizeNurseriesFixed(blocks / n_capabilities);
475 /* -----------------------------------------------------------------------------
476 move_TSO is called to update the TSO structure after it has been
477 moved from one place to another.
478 -------------------------------------------------------------------------- */
481 move_TSO (StgTSO *src, StgTSO *dest)
485 // relocate the stack pointer...
486 diff = (StgPtr)dest - (StgPtr)src; // In *words*
487 dest->sp = (StgPtr)dest->sp + diff;
490 /* -----------------------------------------------------------------------------
491 split N blocks off the front of the given bdescr, returning the
492 new block group. We add the remainder to the large_blocks list
493 in the same step as the original block.
494 -------------------------------------------------------------------------- */
497 splitLargeBlock (bdescr *bd, nat blocks)
503 ASSERT(countBlocks(bd->gen->large_objects) == bd->gen->n_large_blocks);
505 // subtract the original number of blocks from the counter first
506 bd->gen->n_large_blocks -= bd->blocks;
508 new_bd = splitBlockGroup (bd, blocks);
509 initBdescr(new_bd, bd->gen, bd->gen->to);
510 new_bd->flags = BF_LARGE | (bd->flags & BF_EVACUATED);
511 // if new_bd is in an old generation, we have to set BF_EVACUATED
512 new_bd->free = bd->free;
513 dbl_link_onto(new_bd, &bd->gen->large_objects);
515 ASSERT(new_bd->free <= new_bd->start + new_bd->blocks * BLOCK_SIZE_W);
517 // add the new number of blocks to the counter. Due to the gaps
518 // for block descriptors, new_bd->blocks + bd->blocks might not be
519 // equal to the original bd->blocks, which is why we do it this way.
520 bd->gen->n_large_blocks += bd->blocks + new_bd->blocks;
522 ASSERT(countBlocks(bd->gen->large_objects) == bd->gen->n_large_blocks);
529 /* -----------------------------------------------------------------------------
532 This allocates memory in the current thread - it is intended for
533 use primarily from STG-land where we have a Capability. It is
534 better than allocate() because it doesn't require taking the
535 sm_mutex lock in the common case.
537 Memory is allocated directly from the nursery if possible (but not
538 from the current nursery block, so as not to interfere with
540 -------------------------------------------------------------------------- */
543 allocate (Capability *cap, lnat n)
548 if (n >= LARGE_OBJECT_THRESHOLD/sizeof(W_)) {
549 lnat req_blocks = (lnat)BLOCK_ROUND_UP(n*sizeof(W_)) / BLOCK_SIZE;
551 // Attempting to allocate an object larger than maxHeapSize
552 // should definitely be disallowed. (bug #1791)
553 if (RtsFlags.GcFlags.maxHeapSize > 0 &&
554 req_blocks >= RtsFlags.GcFlags.maxHeapSize) {
556 // heapOverflow() doesn't exit (see #2592), but we aren't
557 // in a position to do a clean shutdown here: we
558 // either have to allocate the memory or exit now.
559 // Allocating the memory would be bad, because the user
560 // has requested that we not exceed maxHeapSize, so we
562 stg_exit(EXIT_HEAPOVERFLOW);
566 bd = allocGroup(req_blocks);
567 dbl_link_onto(bd, &g0->large_objects);
568 g0->n_large_blocks += bd->blocks; // might be larger than req_blocks
569 g0->n_new_large_blocks += bd->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++;
669 g0->n_new_large_blocks++;
671 initBdescr(bd, g0, g0);
672 bd->flags = BF_PINNED | BF_LARGE;
673 bd->free = bd->start;
681 /* -----------------------------------------------------------------------------
683 -------------------------------------------------------------------------- */
686 This is the write barrier for MUT_VARs, a.k.a. IORefs. A
687 MUT_VAR_CLEAN object is not on the mutable list; a MUT_VAR_DIRTY
688 is. When written to, a MUT_VAR_CLEAN turns into a MUT_VAR_DIRTY
689 and is put on the mutable list.
692 dirty_MUT_VAR(StgRegTable *reg, StgClosure *p)
694 Capability *cap = regTableToCapability(reg);
696 if (p->header.info == &stg_MUT_VAR_CLEAN_info) {
697 p->header.info = &stg_MUT_VAR_DIRTY_info;
698 bd = Bdescr((StgPtr)p);
699 if (bd->gen_no > 0) recordMutableCap(p,cap,bd->gen_no);
703 // Setting a TSO's link field with a write barrier.
704 // It is *not* necessary to call this function when
705 // * setting the link field to END_TSO_QUEUE
706 // * putting a TSO on the blackhole_queue
707 // * setting the link field of the currently running TSO, as it
708 // will already be dirty.
710 setTSOLink (Capability *cap, StgTSO *tso, StgTSO *target)
713 if (tso->dirty == 0 && (tso->flags & TSO_LINK_DIRTY) == 0) {
714 tso->flags |= TSO_LINK_DIRTY;
715 bd = Bdescr((StgPtr)tso);
716 if (bd->gen_no > 0) recordMutableCap((StgClosure*)tso,cap,bd->gen_no);
722 dirty_TSO (Capability *cap, StgTSO *tso)
725 if (tso->dirty == 0 && (tso->flags & TSO_LINK_DIRTY) == 0) {
726 bd = Bdescr((StgPtr)tso);
727 if (bd->gen_no > 0) recordMutableCap((StgClosure*)tso,cap,bd->gen_no);
733 This is the write barrier for MVARs. An MVAR_CLEAN objects is not
734 on the mutable list; a MVAR_DIRTY is. When written to, a
735 MVAR_CLEAN turns into a MVAR_DIRTY and is put on the mutable list.
736 The check for MVAR_CLEAN is inlined at the call site for speed,
737 this really does make a difference on concurrency-heavy benchmarks
738 such as Chaneneos and cheap-concurrency.
741 dirty_MVAR(StgRegTable *reg, StgClosure *p)
743 Capability *cap = regTableToCapability(reg);
745 bd = Bdescr((StgPtr)p);
746 if (bd->gen_no > 0) recordMutableCap(p,cap,bd->gen_no);
749 /* -----------------------------------------------------------------------------
751 * -------------------------------------------------------------------------- */
753 /* -----------------------------------------------------------------------------
756 * Approximate how much we've allocated: number of blocks in the
757 * nursery + blocks allocated via allocate() - unused nusery blocks.
758 * This leaves a little slop at the end of each block.
759 * -------------------------------------------------------------------------- */
762 calcAllocated( void )
768 allocated = countNurseryBlocks() * BLOCK_SIZE_W;
770 for (i = 0; i < n_capabilities; i++) {
772 for ( bd = capabilities[i].r.rCurrentNursery->link;
773 bd != NULL; bd = bd->link ) {
774 allocated -= BLOCK_SIZE_W;
776 cap = &capabilities[i];
777 if (cap->r.rCurrentNursery->free <
778 cap->r.rCurrentNursery->start + BLOCK_SIZE_W) {
779 allocated -= (cap->r.rCurrentNursery->start + BLOCK_SIZE_W)
780 - cap->r.rCurrentNursery->free;
782 if (cap->pinned_object_block != NULL) {
783 allocated -= (cap->pinned_object_block->start + BLOCK_SIZE_W) -
784 cap->pinned_object_block->free;
788 allocated += g0->n_new_large_blocks * BLOCK_SIZE_W;
793 /* Approximate the amount of live data in the heap. To be called just
794 * after garbage collection (see GarbageCollect()).
796 lnat calcLiveBlocks (void)
802 for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
803 /* approximate amount of live data (doesn't take into account slop
804 * at end of each block).
806 gen = &generations[g];
807 live += gen->n_large_blocks + gen->n_blocks;
812 lnat countOccupied (bdescr *bd)
817 for (; bd != NULL; bd = bd->link) {
818 ASSERT(bd->free <= bd->start + bd->blocks * BLOCK_SIZE_W);
819 words += bd->free - bd->start;
824 // Return an accurate count of the live data in the heap, excluding
826 lnat calcLiveWords (void)
833 for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
834 gen = &generations[g];
835 live += gen->n_words + countOccupied(gen->large_objects);
840 /* Approximate the number of blocks that will be needed at the next
841 * garbage collection.
843 * Assume: all data currently live will remain live. Generationss
844 * that will be collected next time will therefore need twice as many
845 * blocks since all the data will be copied.
854 for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
855 gen = &generations[g];
857 // we need at least this much space
858 needed += gen->n_blocks + gen->n_large_blocks;
860 // any additional space needed to collect this gen next time?
861 if (g == 0 || // always collect gen 0
862 (gen->n_blocks + gen->n_large_blocks > gen->max_blocks)) {
863 // we will collect this gen next time
866 needed += gen->n_blocks / BITS_IN(W_);
868 needed += gen->n_blocks / 100;
871 continue; // no additional space needed for compaction
873 needed += gen->n_blocks;
880 /* ----------------------------------------------------------------------------
883 Executable memory must be managed separately from non-executable
884 memory. Most OSs these days require you to jump through hoops to
885 dynamically allocate executable memory, due to various security
888 Here we provide a small memory allocator for executable memory.
889 Memory is managed with a page granularity; we allocate linearly
890 in the page, and when the page is emptied (all objects on the page
891 are free) we free the page again, not forgetting to make it
894 TODO: The inability to handle objects bigger than BLOCK_SIZE_W means that
895 the linker cannot use allocateExec for loading object code files
896 on Windows. Once allocateExec can handle larger objects, the linker
897 should be modified to use allocateExec instead of VirtualAlloc.
898 ------------------------------------------------------------------------- */
900 #if defined(linux_HOST_OS)
902 // On Linux we need to use libffi for allocating executable memory,
903 // because it knows how to work around the restrictions put in place
906 void *allocateExec (nat bytes, void **exec_ret)
910 ret = ffi_closure_alloc (sizeof(void *) + (size_t)bytes, (void**)&exec);
912 if (ret == NULL) return ret;
913 *ret = ret; // save the address of the writable mapping, for freeExec().
914 *exec_ret = exec + 1;
918 // freeExec gets passed the executable address, not the writable address.
919 void freeExec (void *addr)
922 writable = *((void**)addr - 1);
924 ffi_closure_free (writable);
930 void *allocateExec (nat bytes, void **exec_ret)
937 // round up to words.
938 n = (bytes + sizeof(W_) + 1) / sizeof(W_);
940 if (n+1 > BLOCK_SIZE_W) {
941 barf("allocateExec: can't handle large objects");
944 if (exec_block == NULL ||
945 exec_block->free + n + 1 > exec_block->start + BLOCK_SIZE_W) {
947 lnat pagesize = getPageSize();
948 bd = allocGroup(stg_max(1, pagesize / BLOCK_SIZE));
949 debugTrace(DEBUG_gc, "allocate exec block %p", bd->start);
952 bd->link = exec_block;
953 if (exec_block != NULL) {
954 exec_block->u.back = bd;
957 setExecutable(bd->start, bd->blocks * BLOCK_SIZE, rtsTrue);
960 *(exec_block->free) = n; // store the size of this chunk
961 exec_block->gen_no += n; // gen_no stores the number of words allocated
962 ret = exec_block->free + 1;
963 exec_block->free += n + 1;
970 void freeExec (void *addr)
972 StgPtr p = (StgPtr)addr - 1;
973 bdescr *bd = Bdescr((StgPtr)p);
975 if ((bd->flags & BF_EXEC) == 0) {
976 barf("freeExec: not executable");
979 if (*(StgPtr)p == 0) {
980 barf("freeExec: already free?");
985 bd->gen_no -= *(StgPtr)p;
988 if (bd->gen_no == 0) {
989 // Free the block if it is empty, but not if it is the block at
990 // the head of the queue.
991 if (bd != exec_block) {
992 debugTrace(DEBUG_gc, "free exec block %p", bd->start);
993 dbl_link_remove(bd, &exec_block);
994 setExecutable(bd->start, bd->blocks * BLOCK_SIZE, rtsFalse);
997 bd->free = bd->start;
1004 #endif /* mingw32_HOST_OS */
1008 // handy function for use in gdb, because Bdescr() is inlined.
1009 extern bdescr *_bdescr( StgPtr p );