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_BLOCKING_QUEUE_CLEAN_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;
180 caf_list = END_OF_STATIC_LIST;
181 revertible_caf_list = END_OF_STATIC_LIST;
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());
216 freeStorage (rtsBool free_heap)
218 stgFree(generations);
219 if (free_heap) freeAllMBlocks();
220 #if defined(THREADED_RTS)
221 closeMutex(&sm_mutex);
227 /* -----------------------------------------------------------------------------
230 The entry code for every CAF does the following:
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
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.
241 newCaf() does the following:
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
247 For GHCI, we have additional requirements when dealing with CAFs:
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.
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.
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()).
265 -------------------------------------------------------------------------- */
268 newCAF(StgRegTable *reg, StgClosure* caf)
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;
284 ACQUIRE_SM_LOCK; // caf_list is global, locked by sm_mutex
285 ((StgIndStatic *)caf)->static_link = caf_list;
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);
299 // External API for setting the keepCAFs flag. see #3900.
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.
313 // The linker hackily arranges that references to newCaf from dynamic
314 // code end up pointing to newDynCAF.
316 newDynCAF (StgRegTable *reg STG_UNUSED, StgClosure *caf)
320 ((StgIndStatic *)caf)->saved_info = (StgInfoTable *)caf->header.info;
321 ((StgIndStatic *)caf)->static_link = revertible_caf_list;
322 revertible_caf_list = caf;
327 /* -----------------------------------------------------------------------------
329 -------------------------------------------------------------------------- */
332 allocNursery (bdescr *tail, nat blocks)
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.
344 n = stg_min(blocks, BLOCKS_PER_MBLOCK);
348 for (i = 0; i < n; i++) {
349 initBdescr(&bd[i], g0, g0);
355 bd[i].u.back = &bd[i-1];
361 bd[i].link = &bd[i+1];
365 tail->u.back = &bd[i];
369 bd[i].free = bd[i].start;
379 assignNurseriesToCapabilities (void)
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;
391 allocNurseries( void )
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;
401 assignNurseriesToCapabilities();
405 resetNurseries( void )
410 for (i = 0; i < n_capabilities; i++) {
411 for (bd = nurseries[i].blocks; bd; bd = bd->link) {
412 bd->free = bd->start;
413 ASSERT(bd->gen_no == 0);
414 ASSERT(bd->gen == g0);
415 IF_DEBUG(sanity,memset(bd->start, 0xaa, BLOCK_SIZE));
418 assignNurseriesToCapabilities();
422 countNurseryBlocks (void)
427 for (i = 0; i < n_capabilities; i++) {
428 blocks += nurseries[i].n_blocks;
434 resizeNursery ( nursery *nursery, nat blocks )
439 nursery_blocks = nursery->n_blocks;
440 if (nursery_blocks == blocks) return;
442 if (nursery_blocks < blocks) {
443 debugTrace(DEBUG_gc, "increasing size of nursery to %d blocks",
445 nursery->blocks = allocNursery(nursery->blocks, blocks-nursery_blocks);
450 debugTrace(DEBUG_gc, "decreasing size of nursery to %d blocks",
453 bd = nursery->blocks;
454 while (nursery_blocks > blocks) {
456 next_bd->u.back = NULL;
457 nursery_blocks -= bd->blocks; // might be a large block
461 nursery->blocks = bd;
462 // might have gone just under, by freeing a large block, so make
463 // up the difference.
464 if (nursery_blocks < blocks) {
465 nursery->blocks = allocNursery(nursery->blocks, blocks-nursery_blocks);
469 nursery->n_blocks = blocks;
470 ASSERT(countBlocks(nursery->blocks) == nursery->n_blocks);
474 // Resize each of the nurseries to the specified size.
477 resizeNurseriesFixed (nat blocks)
480 for (i = 0; i < n_capabilities; i++) {
481 resizeNursery(&nurseries[i], blocks);
486 // Resize the nurseries to the total specified size.
489 resizeNurseries (nat blocks)
491 // If there are multiple nurseries, then we just divide the number
492 // of available blocks between them.
493 resizeNurseriesFixed(blocks / n_capabilities);
497 /* -----------------------------------------------------------------------------
498 move_TSO is called to update the TSO structure after it has been
499 moved from one place to another.
500 -------------------------------------------------------------------------- */
503 move_TSO (StgTSO *src, StgTSO *dest)
507 // relocate the stack pointer...
508 diff = (StgPtr)dest - (StgPtr)src; // In *words*
509 dest->sp = (StgPtr)dest->sp + diff;
512 /* -----------------------------------------------------------------------------
513 split N blocks off the front of the given bdescr, returning the
514 new block group. We add the remainder to the large_blocks list
515 in the same step as the original block.
516 -------------------------------------------------------------------------- */
519 splitLargeBlock (bdescr *bd, nat blocks)
525 ASSERT(countBlocks(bd->gen->large_objects) == bd->gen->n_large_blocks);
527 // subtract the original number of blocks from the counter first
528 bd->gen->n_large_blocks -= bd->blocks;
530 new_bd = splitBlockGroup (bd, blocks);
531 initBdescr(new_bd, bd->gen, bd->gen->to);
532 new_bd->flags = BF_LARGE | (bd->flags & BF_EVACUATED);
533 // if new_bd is in an old generation, we have to set BF_EVACUATED
534 new_bd->free = bd->free;
535 dbl_link_onto(new_bd, &bd->gen->large_objects);
537 ASSERT(new_bd->free <= new_bd->start + new_bd->blocks * BLOCK_SIZE_W);
539 // add the new number of blocks to the counter. Due to the gaps
540 // for block descriptors, new_bd->blocks + bd->blocks might not be
541 // equal to the original bd->blocks, which is why we do it this way.
542 bd->gen->n_large_blocks += bd->blocks + new_bd->blocks;
544 ASSERT(countBlocks(bd->gen->large_objects) == bd->gen->n_large_blocks);
551 /* -----------------------------------------------------------------------------
554 This allocates memory in the current thread - it is intended for
555 use primarily from STG-land where we have a Capability. It is
556 better than allocate() because it doesn't require taking the
557 sm_mutex lock in the common case.
559 Memory is allocated directly from the nursery if possible (but not
560 from the current nursery block, so as not to interfere with
562 -------------------------------------------------------------------------- */
565 allocate (Capability *cap, lnat n)
570 if (n >= LARGE_OBJECT_THRESHOLD/sizeof(W_)) {
571 lnat req_blocks = (lnat)BLOCK_ROUND_UP(n*sizeof(W_)) / BLOCK_SIZE;
573 // Attempting to allocate an object larger than maxHeapSize
574 // should definitely be disallowed. (bug #1791)
575 if (RtsFlags.GcFlags.maxHeapSize > 0 &&
576 req_blocks >= RtsFlags.GcFlags.maxHeapSize) {
578 // heapOverflow() doesn't exit (see #2592), but we aren't
579 // in a position to do a clean shutdown here: we
580 // either have to allocate the memory or exit now.
581 // Allocating the memory would be bad, because the user
582 // has requested that we not exceed maxHeapSize, so we
584 stg_exit(EXIT_HEAPOVERFLOW);
588 bd = allocGroup(req_blocks);
589 dbl_link_onto(bd, &g0->large_objects);
590 g0->n_large_blocks += bd->blocks; // might be larger than req_blocks
591 g0->n_new_large_blocks += bd->blocks;
593 initBdescr(bd, g0, g0);
594 bd->flags = BF_LARGE;
595 bd->free = bd->start + n;
599 /* small allocation (<LARGE_OBJECT_THRESHOLD) */
601 TICK_ALLOC_HEAP_NOCTR(n);
604 bd = cap->r.rCurrentAlloc;
605 if (bd == NULL || bd->free + n > bd->start + BLOCK_SIZE_W) {
607 // The CurrentAlloc block is full, we need to find another
608 // one. First, we try taking the next block from the
610 bd = cap->r.rCurrentNursery->link;
612 if (bd == NULL || bd->free + n > bd->start + BLOCK_SIZE_W) {
613 // The nursery is empty, or the next block is already
614 // full: allocate a fresh block (we can't fail here).
617 cap->r.rNursery->n_blocks++;
619 initBdescr(bd, g0, g0);
621 // If we had to allocate a new block, then we'll GC
622 // pretty quickly now, because MAYBE_GC() will
623 // notice that CurrentNursery->link is NULL.
625 // we have a block in the nursery: take it and put
626 // it at the *front* of the nursery list, and use it
627 // to allocate() from.
628 cap->r.rCurrentNursery->link = bd->link;
629 if (bd->link != NULL) {
630 bd->link->u.back = cap->r.rCurrentNursery;
633 dbl_link_onto(bd, &cap->r.rNursery->blocks);
634 cap->r.rCurrentAlloc = bd;
635 IF_DEBUG(sanity, checkNurserySanity(cap->r.rNursery));
640 IF_DEBUG(sanity, ASSERT(*((StgWord8*)p) == 0xaa));
644 /* ---------------------------------------------------------------------------
645 Allocate a fixed/pinned object.
647 We allocate small pinned objects into a single block, allocating a
648 new block when the current one overflows. The block is chained
649 onto the large_object_list of generation 0.
651 NOTE: The GC can't in general handle pinned objects. This
652 interface is only safe to use for ByteArrays, which have no
653 pointers and don't require scavenging. It works because the
654 block's descriptor has the BF_LARGE flag set, so the block is
655 treated as a large object and chained onto various lists, rather
656 than the individual objects being copied. However, when it comes
657 to scavenge the block, the GC will only scavenge the first object.
658 The reason is that the GC can't linearly scan a block of pinned
659 objects at the moment (doing so would require using the
660 mostly-copying techniques). But since we're restricting ourselves
661 to pinned ByteArrays, not scavenging is ok.
663 This function is called by newPinnedByteArray# which immediately
664 fills the allocated memory with a MutableByteArray#.
665 ------------------------------------------------------------------------- */
668 allocatePinned (Capability *cap, lnat n)
673 // If the request is for a large object, then allocate()
674 // will give us a pinned object anyway.
675 if (n >= LARGE_OBJECT_THRESHOLD/sizeof(W_)) {
676 p = allocate(cap, n);
677 Bdescr(p)->flags |= BF_PINNED;
681 TICK_ALLOC_HEAP_NOCTR(n);
684 bd = cap->pinned_object_block;
686 // If we don't have a block of pinned objects yet, or the current
687 // one isn't large enough to hold the new object, allocate a new one.
688 if (bd == NULL || (bd->free + n) > (bd->start + BLOCK_SIZE_W)) {
690 cap->pinned_object_block = bd = allocBlock();
691 dbl_link_onto(bd, &g0->large_objects);
692 g0->n_large_blocks++;
693 g0->n_new_large_blocks++;
695 initBdescr(bd, g0, g0);
696 bd->flags = BF_PINNED | BF_LARGE;
697 bd->free = bd->start;
705 /* -----------------------------------------------------------------------------
707 -------------------------------------------------------------------------- */
710 This is the write barrier for MUT_VARs, a.k.a. IORefs. A
711 MUT_VAR_CLEAN object is not on the mutable list; a MUT_VAR_DIRTY
712 is. When written to, a MUT_VAR_CLEAN turns into a MUT_VAR_DIRTY
713 and is put on the mutable list.
716 dirty_MUT_VAR(StgRegTable *reg, StgClosure *p)
718 Capability *cap = regTableToCapability(reg);
719 if (p->header.info == &stg_MUT_VAR_CLEAN_info) {
720 p->header.info = &stg_MUT_VAR_DIRTY_info;
721 recordClosureMutated(cap,p);
725 // Setting a TSO's link field with a write barrier.
726 // It is *not* necessary to call this function when
727 // * setting the link field to END_TSO_QUEUE
728 // * putting a TSO on the blackhole_queue
729 // * setting the link field of the currently running TSO, as it
730 // will already be dirty.
732 setTSOLink (Capability *cap, StgTSO *tso, StgTSO *target)
734 if (tso->dirty == 0 && (tso->flags & TSO_LINK_DIRTY) == 0) {
735 tso->flags |= TSO_LINK_DIRTY;
736 recordClosureMutated(cap,(StgClosure*)tso);
742 setTSOPrev (Capability *cap, StgTSO *tso, StgTSO *target)
744 if (tso->dirty == 0 && (tso->flags & TSO_LINK_DIRTY) == 0) {
745 tso->flags |= TSO_LINK_DIRTY;
746 recordClosureMutated(cap,(StgClosure*)tso);
748 tso->block_info.prev = target;
752 dirty_TSO (Capability *cap, StgTSO *tso)
754 if (tso->dirty == 0 && (tso->flags & TSO_LINK_DIRTY) == 0) {
755 recordClosureMutated(cap,(StgClosure*)tso);
761 This is the write barrier for MVARs. An MVAR_CLEAN objects is not
762 on the mutable list; a MVAR_DIRTY is. When written to, a
763 MVAR_CLEAN turns into a MVAR_DIRTY and is put on the mutable list.
764 The check for MVAR_CLEAN is inlined at the call site for speed,
765 this really does make a difference on concurrency-heavy benchmarks
766 such as Chaneneos and cheap-concurrency.
769 dirty_MVAR(StgRegTable *reg, StgClosure *p)
771 recordClosureMutated(regTableToCapability(reg),p);
774 /* -----------------------------------------------------------------------------
776 * -------------------------------------------------------------------------- */
778 /* -----------------------------------------------------------------------------
781 * Approximate how much we've allocated: number of blocks in the
782 * nursery + blocks allocated via allocate() - unused nusery blocks.
783 * This leaves a little slop at the end of each block.
784 * -------------------------------------------------------------------------- */
787 calcAllocated( void )
793 allocated = countNurseryBlocks() * BLOCK_SIZE_W;
795 for (i = 0; i < n_capabilities; i++) {
797 for ( bd = capabilities[i].r.rCurrentNursery->link;
798 bd != NULL; bd = bd->link ) {
799 allocated -= BLOCK_SIZE_W;
801 cap = &capabilities[i];
802 if (cap->r.rCurrentNursery->free <
803 cap->r.rCurrentNursery->start + BLOCK_SIZE_W) {
804 allocated -= (cap->r.rCurrentNursery->start + BLOCK_SIZE_W)
805 - cap->r.rCurrentNursery->free;
807 if (cap->pinned_object_block != NULL) {
808 allocated -= (cap->pinned_object_block->start + BLOCK_SIZE_W) -
809 cap->pinned_object_block->free;
813 allocated += g0->n_new_large_blocks * BLOCK_SIZE_W;
818 /* Approximate the amount of live data in the heap. To be called just
819 * after garbage collection (see GarbageCollect()).
821 lnat calcLiveBlocks (void)
827 for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
828 /* approximate amount of live data (doesn't take into account slop
829 * at end of each block).
831 gen = &generations[g];
832 live += gen->n_large_blocks + gen->n_blocks;
837 lnat countOccupied (bdescr *bd)
842 for (; bd != NULL; bd = bd->link) {
843 ASSERT(bd->free <= bd->start + bd->blocks * BLOCK_SIZE_W);
844 words += bd->free - bd->start;
849 // Return an accurate count of the live data in the heap, excluding
851 lnat calcLiveWords (void)
858 for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
859 gen = &generations[g];
860 live += gen->n_words + countOccupied(gen->large_objects);
865 /* Approximate the number of blocks that will be needed at the next
866 * garbage collection.
868 * Assume: all data currently live will remain live. Generationss
869 * that will be collected next time will therefore need twice as many
870 * blocks since all the data will be copied.
879 for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
880 gen = &generations[g];
882 // we need at least this much space
883 needed += gen->n_blocks + gen->n_large_blocks;
885 // any additional space needed to collect this gen next time?
886 if (g == 0 || // always collect gen 0
887 (gen->n_blocks + gen->n_large_blocks > gen->max_blocks)) {
888 // we will collect this gen next time
891 needed += gen->n_blocks / BITS_IN(W_);
893 needed += gen->n_blocks / 100;
896 continue; // no additional space needed for compaction
898 needed += gen->n_blocks;
905 /* ----------------------------------------------------------------------------
908 Executable memory must be managed separately from non-executable
909 memory. Most OSs these days require you to jump through hoops to
910 dynamically allocate executable memory, due to various security
913 Here we provide a small memory allocator for executable memory.
914 Memory is managed with a page granularity; we allocate linearly
915 in the page, and when the page is emptied (all objects on the page
916 are free) we free the page again, not forgetting to make it
919 TODO: The inability to handle objects bigger than BLOCK_SIZE_W means that
920 the linker cannot use allocateExec for loading object code files
921 on Windows. Once allocateExec can handle larger objects, the linker
922 should be modified to use allocateExec instead of VirtualAlloc.
923 ------------------------------------------------------------------------- */
925 #if defined(linux_HOST_OS)
927 // On Linux we need to use libffi for allocating executable memory,
928 // because it knows how to work around the restrictions put in place
931 void *allocateExec (nat bytes, void **exec_ret)
935 ret = ffi_closure_alloc (sizeof(void *) + (size_t)bytes, (void**)&exec);
937 if (ret == NULL) return ret;
938 *ret = ret; // save the address of the writable mapping, for freeExec().
939 *exec_ret = exec + 1;
943 // freeExec gets passed the executable address, not the writable address.
944 void freeExec (void *addr)
947 writable = *((void**)addr - 1);
949 ffi_closure_free (writable);
955 void *allocateExec (nat bytes, void **exec_ret)
962 // round up to words.
963 n = (bytes + sizeof(W_) + 1) / sizeof(W_);
965 if (n+1 > BLOCK_SIZE_W) {
966 barf("allocateExec: can't handle large objects");
969 if (exec_block == NULL ||
970 exec_block->free + n + 1 > exec_block->start + BLOCK_SIZE_W) {
972 lnat pagesize = getPageSize();
973 bd = allocGroup(stg_max(1, pagesize / BLOCK_SIZE));
974 debugTrace(DEBUG_gc, "allocate exec block %p", bd->start);
977 bd->link = exec_block;
978 if (exec_block != NULL) {
979 exec_block->u.back = bd;
982 setExecutable(bd->start, bd->blocks * BLOCK_SIZE, rtsTrue);
985 *(exec_block->free) = n; // store the size of this chunk
986 exec_block->gen_no += n; // gen_no stores the number of words allocated
987 ret = exec_block->free + 1;
988 exec_block->free += n + 1;
995 void freeExec (void *addr)
997 StgPtr p = (StgPtr)addr - 1;
998 bdescr *bd = Bdescr((StgPtr)p);
1000 if ((bd->flags & BF_EXEC) == 0) {
1001 barf("freeExec: not executable");
1004 if (*(StgPtr)p == 0) {
1005 barf("freeExec: already free?");
1010 bd->gen_no -= *(StgPtr)p;
1013 if (bd->gen_no == 0) {
1014 // Free the block if it is empty, but not if it is the block at
1015 // the head of the queue.
1016 if (bd != exec_block) {
1017 debugTrace(DEBUG_gc, "free exec block %p", bd->start);
1018 dbl_link_remove(bd, &exec_block);
1019 setExecutable(bd->start, bd->blocks * BLOCK_SIZE, rtsFalse);
1022 bd->free = bd->start;
1029 #endif /* mingw32_HOST_OS */
1033 // handy function for use in gdb, because Bdescr() is inlined.
1034 extern bdescr *_bdescr( StgPtr p );