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)
274 // If we are in GHCi _and_ we are using dynamic libraries,
275 // then we can't redirect newCAF calls to newDynCAF (see below),
276 // so we make newCAF behave almost like newDynCAF.
277 // The dynamic libraries might be used by both the interpreted
278 // program and GHCi itself, so they must not be reverted.
279 // This also means that in GHCi with dynamic libraries, CAFs are not
280 // garbage collected. If this turns out to be a problem, we could
281 // do another hack here and do an address range test on caf to figure
282 // out whether it is from a dynamic library.
283 ((StgIndStatic *)caf)->saved_info = (StgInfoTable *)caf->header.info;
285 ACQUIRE_SM_LOCK; // caf_list is global, locked by sm_mutex
286 ((StgIndStatic *)caf)->static_link = caf_list;
292 // Put this CAF on the mutable list for the old generation.
293 ((StgIndStatic *)caf)->saved_info = NULL;
294 recordMutableCap(caf, regTableToCapability(reg), oldest_gen->no);
298 // External API for setting the keepCAFs flag. see #3900.
305 // An alternate version of newCaf which is used for dynamically loaded
306 // object code in GHCi. In this case we want to retain *all* CAFs in
307 // the object code, because they might be demanded at any time from an
308 // expression evaluated on the command line.
309 // Also, GHCi might want to revert CAFs, so we add these to the
310 // revertible_caf_list.
312 // The linker hackily arranges that references to newCaf from dynamic
313 // code end up pointing to newDynCAF.
315 newDynCAF (StgRegTable *reg STG_UNUSED, StgClosure *caf)
319 ((StgIndStatic *)caf)->saved_info = (StgInfoTable *)caf->header.info;
320 ((StgIndStatic *)caf)->static_link = revertible_caf_list;
321 revertible_caf_list = caf;
326 /* -----------------------------------------------------------------------------
328 -------------------------------------------------------------------------- */
331 allocNursery (bdescr *tail, nat blocks)
336 // Allocate a nursery: we allocate fresh blocks one at a time and
337 // cons them on to the front of the list, not forgetting to update
338 // the back pointer on the tail of the list to point to the new block.
339 for (i=0; i < blocks; i++) {
342 processNursery() in LdvProfile.c assumes that every block group in
343 the nursery contains only a single block. So, if a block group is
344 given multiple blocks, change processNursery() accordingly.
348 // double-link the nursery: we might need to insert blocks
352 initBdescr(bd, g0, g0);
354 bd->free = bd->start;
362 assignNurseriesToCapabilities (void)
366 for (i = 0; i < n_capabilities; i++) {
367 capabilities[i].r.rNursery = &nurseries[i];
368 capabilities[i].r.rCurrentNursery = nurseries[i].blocks;
369 capabilities[i].r.rCurrentAlloc = NULL;
374 allocNurseries( void )
378 for (i = 0; i < n_capabilities; i++) {
379 nurseries[i].blocks =
380 allocNursery(NULL, RtsFlags.GcFlags.minAllocAreaSize);
381 nurseries[i].n_blocks =
382 RtsFlags.GcFlags.minAllocAreaSize;
384 assignNurseriesToCapabilities();
388 resetNurseries( void )
393 for (i = 0; i < n_capabilities; i++) {
394 for (bd = nurseries[i].blocks; bd; bd = bd->link) {
395 bd->free = bd->start;
396 ASSERT(bd->gen_no == 0);
397 ASSERT(bd->gen == g0);
398 IF_DEBUG(sanity,memset(bd->start, 0xaa, BLOCK_SIZE));
401 assignNurseriesToCapabilities();
405 countNurseryBlocks (void)
410 for (i = 0; i < n_capabilities; i++) {
411 blocks += nurseries[i].n_blocks;
417 resizeNursery ( nursery *nursery, nat blocks )
422 nursery_blocks = nursery->n_blocks;
423 if (nursery_blocks == blocks) return;
425 if (nursery_blocks < blocks) {
426 debugTrace(DEBUG_gc, "increasing size of nursery to %d blocks",
428 nursery->blocks = allocNursery(nursery->blocks, blocks-nursery_blocks);
433 debugTrace(DEBUG_gc, "decreasing size of nursery to %d blocks",
436 bd = nursery->blocks;
437 while (nursery_blocks > blocks) {
439 next_bd->u.back = NULL;
440 nursery_blocks -= bd->blocks; // might be a large block
444 nursery->blocks = bd;
445 // might have gone just under, by freeing a large block, so make
446 // up the difference.
447 if (nursery_blocks < blocks) {
448 nursery->blocks = allocNursery(nursery->blocks, blocks-nursery_blocks);
452 nursery->n_blocks = blocks;
453 ASSERT(countBlocks(nursery->blocks) == nursery->n_blocks);
457 // Resize each of the nurseries to the specified size.
460 resizeNurseriesFixed (nat blocks)
463 for (i = 0; i < n_capabilities; i++) {
464 resizeNursery(&nurseries[i], blocks);
469 // Resize the nurseries to the total specified size.
472 resizeNurseries (nat blocks)
474 // If there are multiple nurseries, then we just divide the number
475 // of available blocks between them.
476 resizeNurseriesFixed(blocks / n_capabilities);
480 /* -----------------------------------------------------------------------------
481 move_TSO is called to update the TSO structure after it has been
482 moved from one place to another.
483 -------------------------------------------------------------------------- */
486 move_TSO (StgTSO *src, StgTSO *dest)
490 // relocate the stack pointer...
491 diff = (StgPtr)dest - (StgPtr)src; // In *words*
492 dest->sp = (StgPtr)dest->sp + diff;
495 /* -----------------------------------------------------------------------------
496 split N blocks off the front of the given bdescr, returning the
497 new block group. We add the remainder to the large_blocks list
498 in the same step as the original block.
499 -------------------------------------------------------------------------- */
502 splitLargeBlock (bdescr *bd, nat blocks)
508 ASSERT(countBlocks(bd->gen->large_objects) == bd->gen->n_large_blocks);
510 // subtract the original number of blocks from the counter first
511 bd->gen->n_large_blocks -= bd->blocks;
513 new_bd = splitBlockGroup (bd, blocks);
514 initBdescr(new_bd, bd->gen, bd->gen->to);
515 new_bd->flags = BF_LARGE | (bd->flags & BF_EVACUATED);
516 // if new_bd is in an old generation, we have to set BF_EVACUATED
517 new_bd->free = bd->free;
518 dbl_link_onto(new_bd, &bd->gen->large_objects);
520 ASSERT(new_bd->free <= new_bd->start + new_bd->blocks * BLOCK_SIZE_W);
522 // add the new number of blocks to the counter. Due to the gaps
523 // for block descriptors, new_bd->blocks + bd->blocks might not be
524 // equal to the original bd->blocks, which is why we do it this way.
525 bd->gen->n_large_blocks += bd->blocks + new_bd->blocks;
527 ASSERT(countBlocks(bd->gen->large_objects) == bd->gen->n_large_blocks);
534 /* -----------------------------------------------------------------------------
537 This allocates memory in the current thread - it is intended for
538 use primarily from STG-land where we have a Capability. It is
539 better than allocate() because it doesn't require taking the
540 sm_mutex lock in the common case.
542 Memory is allocated directly from the nursery if possible (but not
543 from the current nursery block, so as not to interfere with
545 -------------------------------------------------------------------------- */
548 allocate (Capability *cap, lnat n)
553 if (n >= LARGE_OBJECT_THRESHOLD/sizeof(W_)) {
554 lnat req_blocks = (lnat)BLOCK_ROUND_UP(n*sizeof(W_)) / BLOCK_SIZE;
556 // Attempting to allocate an object larger than maxHeapSize
557 // should definitely be disallowed. (bug #1791)
558 if (RtsFlags.GcFlags.maxHeapSize > 0 &&
559 req_blocks >= RtsFlags.GcFlags.maxHeapSize) {
561 // heapOverflow() doesn't exit (see #2592), but we aren't
562 // in a position to do a clean shutdown here: we
563 // either have to allocate the memory or exit now.
564 // Allocating the memory would be bad, because the user
565 // has requested that we not exceed maxHeapSize, so we
567 stg_exit(EXIT_HEAPOVERFLOW);
571 bd = allocGroup(req_blocks);
572 dbl_link_onto(bd, &g0->large_objects);
573 g0->n_large_blocks += bd->blocks; // might be larger than req_blocks
574 g0->n_new_large_blocks += bd->blocks;
576 initBdescr(bd, g0, g0);
577 bd->flags = BF_LARGE;
578 bd->free = bd->start + n;
582 /* small allocation (<LARGE_OBJECT_THRESHOLD) */
584 TICK_ALLOC_HEAP_NOCTR(n);
587 bd = cap->r.rCurrentAlloc;
588 if (bd == NULL || bd->free + n > bd->start + BLOCK_SIZE_W) {
590 // The CurrentAlloc block is full, we need to find another
591 // one. First, we try taking the next block from the
593 bd = cap->r.rCurrentNursery->link;
595 if (bd == NULL || bd->free + n > bd->start + BLOCK_SIZE_W) {
596 // The nursery is empty, or the next block is already
597 // full: allocate a fresh block (we can't fail here).
600 cap->r.rNursery->n_blocks++;
602 initBdescr(bd, g0, g0);
604 // If we had to allocate a new block, then we'll GC
605 // pretty quickly now, because MAYBE_GC() will
606 // notice that CurrentNursery->link is NULL.
608 // we have a block in the nursery: take it and put
609 // it at the *front* of the nursery list, and use it
610 // to allocate() from.
611 cap->r.rCurrentNursery->link = bd->link;
612 if (bd->link != NULL) {
613 bd->link->u.back = cap->r.rCurrentNursery;
616 dbl_link_onto(bd, &cap->r.rNursery->blocks);
617 cap->r.rCurrentAlloc = bd;
618 IF_DEBUG(sanity, checkNurserySanity(cap->r.rNursery));
623 IF_DEBUG(sanity, ASSERT(*((StgWord8*)p) == 0xaa));
627 /* ---------------------------------------------------------------------------
628 Allocate a fixed/pinned object.
630 We allocate small pinned objects into a single block, allocating a
631 new block when the current one overflows. The block is chained
632 onto the large_object_list of generation 0.
634 NOTE: The GC can't in general handle pinned objects. This
635 interface is only safe to use for ByteArrays, which have no
636 pointers and don't require scavenging. It works because the
637 block's descriptor has the BF_LARGE flag set, so the block is
638 treated as a large object and chained onto various lists, rather
639 than the individual objects being copied. However, when it comes
640 to scavenge the block, the GC will only scavenge the first object.
641 The reason is that the GC can't linearly scan a block of pinned
642 objects at the moment (doing so would require using the
643 mostly-copying techniques). But since we're restricting ourselves
644 to pinned ByteArrays, not scavenging is ok.
646 This function is called by newPinnedByteArray# which immediately
647 fills the allocated memory with a MutableByteArray#.
648 ------------------------------------------------------------------------- */
651 allocatePinned (Capability *cap, lnat n)
656 // If the request is for a large object, then allocate()
657 // will give us a pinned object anyway.
658 if (n >= LARGE_OBJECT_THRESHOLD/sizeof(W_)) {
659 p = allocate(cap, n);
660 Bdescr(p)->flags |= BF_PINNED;
664 TICK_ALLOC_HEAP_NOCTR(n);
667 bd = cap->pinned_object_block;
669 // If we don't have a block of pinned objects yet, or the current
670 // one isn't large enough to hold the new object, allocate a new one.
671 if (bd == NULL || (bd->free + n) > (bd->start + BLOCK_SIZE_W)) {
673 cap->pinned_object_block = bd = allocBlock();
674 dbl_link_onto(bd, &g0->large_objects);
675 g0->n_large_blocks++;
676 g0->n_new_large_blocks++;
678 initBdescr(bd, g0, g0);
679 bd->flags = BF_PINNED | BF_LARGE;
680 bd->free = bd->start;
688 /* -----------------------------------------------------------------------------
690 -------------------------------------------------------------------------- */
693 This is the write barrier for MUT_VARs, a.k.a. IORefs. A
694 MUT_VAR_CLEAN object is not on the mutable list; a MUT_VAR_DIRTY
695 is. When written to, a MUT_VAR_CLEAN turns into a MUT_VAR_DIRTY
696 and is put on the mutable list.
699 dirty_MUT_VAR(StgRegTable *reg, StgClosure *p)
701 Capability *cap = regTableToCapability(reg);
703 if (p->header.info == &stg_MUT_VAR_CLEAN_info) {
704 p->header.info = &stg_MUT_VAR_DIRTY_info;
705 bd = Bdescr((StgPtr)p);
706 if (bd->gen_no > 0) recordMutableCap(p,cap,bd->gen_no);
710 // Setting a TSO's link field with a write barrier.
711 // It is *not* necessary to call this function when
712 // * setting the link field to END_TSO_QUEUE
713 // * putting a TSO on the blackhole_queue
714 // * setting the link field of the currently running TSO, as it
715 // will already be dirty.
717 setTSOLink (Capability *cap, StgTSO *tso, StgTSO *target)
720 if (tso->dirty == 0 && (tso->flags & TSO_LINK_DIRTY) == 0) {
721 tso->flags |= TSO_LINK_DIRTY;
722 bd = Bdescr((StgPtr)tso);
723 if (bd->gen_no > 0) recordMutableCap((StgClosure*)tso,cap,bd->gen_no);
729 dirty_TSO (Capability *cap, StgTSO *tso)
732 if (tso->dirty == 0 && (tso->flags & TSO_LINK_DIRTY) == 0) {
733 bd = Bdescr((StgPtr)tso);
734 if (bd->gen_no > 0) recordMutableCap((StgClosure*)tso,cap,bd->gen_no);
740 This is the write barrier for MVARs. An MVAR_CLEAN objects is not
741 on the mutable list; a MVAR_DIRTY is. When written to, a
742 MVAR_CLEAN turns into a MVAR_DIRTY and is put on the mutable list.
743 The check for MVAR_CLEAN is inlined at the call site for speed,
744 this really does make a difference on concurrency-heavy benchmarks
745 such as Chaneneos and cheap-concurrency.
748 dirty_MVAR(StgRegTable *reg, StgClosure *p)
750 Capability *cap = regTableToCapability(reg);
752 bd = Bdescr((StgPtr)p);
753 if (bd->gen_no > 0) recordMutableCap(p,cap,bd->gen_no);
756 /* -----------------------------------------------------------------------------
758 * -------------------------------------------------------------------------- */
760 /* -----------------------------------------------------------------------------
763 * Approximate how much we've allocated: number of blocks in the
764 * nursery + blocks allocated via allocate() - unused nusery blocks.
765 * This leaves a little slop at the end of each block.
766 * -------------------------------------------------------------------------- */
769 calcAllocated( void )
775 allocated = countNurseryBlocks() * BLOCK_SIZE_W;
777 for (i = 0; i < n_capabilities; i++) {
779 for ( bd = capabilities[i].r.rCurrentNursery->link;
780 bd != NULL; bd = bd->link ) {
781 allocated -= BLOCK_SIZE_W;
783 cap = &capabilities[i];
784 if (cap->r.rCurrentNursery->free <
785 cap->r.rCurrentNursery->start + BLOCK_SIZE_W) {
786 allocated -= (cap->r.rCurrentNursery->start + BLOCK_SIZE_W)
787 - cap->r.rCurrentNursery->free;
789 if (cap->pinned_object_block != NULL) {
790 allocated -= (cap->pinned_object_block->start + BLOCK_SIZE_W) -
791 cap->pinned_object_block->free;
795 allocated += g0->n_new_large_blocks * BLOCK_SIZE_W;
800 /* Approximate the amount of live data in the heap. To be called just
801 * after garbage collection (see GarbageCollect()).
803 lnat calcLiveBlocks (void)
809 for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
810 /* approximate amount of live data (doesn't take into account slop
811 * at end of each block).
813 gen = &generations[g];
814 live += gen->n_large_blocks + gen->n_blocks;
819 lnat countOccupied (bdescr *bd)
824 for (; bd != NULL; bd = bd->link) {
825 ASSERT(bd->free <= bd->start + bd->blocks * BLOCK_SIZE_W);
826 words += bd->free - bd->start;
831 // Return an accurate count of the live data in the heap, excluding
833 lnat calcLiveWords (void)
840 for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
841 gen = &generations[g];
842 live += gen->n_words + countOccupied(gen->large_objects);
847 /* Approximate the number of blocks that will be needed at the next
848 * garbage collection.
850 * Assume: all data currently live will remain live. Generationss
851 * that will be collected next time will therefore need twice as many
852 * blocks since all the data will be copied.
861 for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
862 gen = &generations[g];
864 // we need at least this much space
865 needed += gen->n_blocks + gen->n_large_blocks;
867 // any additional space needed to collect this gen next time?
868 if (g == 0 || // always collect gen 0
869 (gen->n_blocks + gen->n_large_blocks > gen->max_blocks)) {
870 // we will collect this gen next time
873 needed += gen->n_blocks / BITS_IN(W_);
875 needed += gen->n_blocks / 100;
878 continue; // no additional space needed for compaction
880 needed += gen->n_blocks;
887 /* ----------------------------------------------------------------------------
890 Executable memory must be managed separately from non-executable
891 memory. Most OSs these days require you to jump through hoops to
892 dynamically allocate executable memory, due to various security
895 Here we provide a small memory allocator for executable memory.
896 Memory is managed with a page granularity; we allocate linearly
897 in the page, and when the page is emptied (all objects on the page
898 are free) we free the page again, not forgetting to make it
901 TODO: The inability to handle objects bigger than BLOCK_SIZE_W means that
902 the linker cannot use allocateExec for loading object code files
903 on Windows. Once allocateExec can handle larger objects, the linker
904 should be modified to use allocateExec instead of VirtualAlloc.
905 ------------------------------------------------------------------------- */
907 #if defined(linux_HOST_OS)
909 // On Linux we need to use libffi for allocating executable memory,
910 // because it knows how to work around the restrictions put in place
913 void *allocateExec (nat bytes, void **exec_ret)
917 ret = ffi_closure_alloc (sizeof(void *) + (size_t)bytes, (void**)&exec);
919 if (ret == NULL) return ret;
920 *ret = ret; // save the address of the writable mapping, for freeExec().
921 *exec_ret = exec + 1;
925 // freeExec gets passed the executable address, not the writable address.
926 void freeExec (void *addr)
929 writable = *((void**)addr - 1);
931 ffi_closure_free (writable);
937 void *allocateExec (nat bytes, void **exec_ret)
944 // round up to words.
945 n = (bytes + sizeof(W_) + 1) / sizeof(W_);
947 if (n+1 > BLOCK_SIZE_W) {
948 barf("allocateExec: can't handle large objects");
951 if (exec_block == NULL ||
952 exec_block->free + n + 1 > exec_block->start + BLOCK_SIZE_W) {
954 lnat pagesize = getPageSize();
955 bd = allocGroup(stg_max(1, pagesize / BLOCK_SIZE));
956 debugTrace(DEBUG_gc, "allocate exec block %p", bd->start);
959 bd->link = exec_block;
960 if (exec_block != NULL) {
961 exec_block->u.back = bd;
964 setExecutable(bd->start, bd->blocks * BLOCK_SIZE, rtsTrue);
967 *(exec_block->free) = n; // store the size of this chunk
968 exec_block->gen_no += n; // gen_no stores the number of words allocated
969 ret = exec_block->free + 1;
970 exec_block->free += n + 1;
977 void freeExec (void *addr)
979 StgPtr p = (StgPtr)addr - 1;
980 bdescr *bd = Bdescr((StgPtr)p);
982 if ((bd->flags & BF_EXEC) == 0) {
983 barf("freeExec: not executable");
986 if (*(StgPtr)p == 0) {
987 barf("freeExec: already free?");
992 bd->gen_no -= *(StgPtr)p;
995 if (bd->gen_no == 0) {
996 // Free the block if it is empty, but not if it is the block at
997 // the head of the queue.
998 if (bd != exec_block) {
999 debugTrace(DEBUG_gc, "free exec block %p", bd->start);
1000 dbl_link_remove(bd, &exec_block);
1001 setExecutable(bd->start, bd->blocks * BLOCK_SIZE, rtsFalse);
1004 bd->free = bd->start;
1011 #endif /* mingw32_HOST_OS */
1015 // handy function for use in gdb, because Bdescr() is inlined.
1016 extern bdescr *_bdescr( StgPtr p );