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());
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 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 recordMutableCap(caf, regTableToCapability(reg), oldest_gen->no);
297 // External API for setting the keepCAFs flag. see #3900.
304 // An alternate version of newCaf which is used for dynamically loaded
305 // object code in GHCi. In this case we want to retain *all* CAFs in
306 // the object code, because they might be demanded at any time from an
307 // expression evaluated on the command line.
308 // Also, GHCi might want to revert CAFs, so we add these to the
309 // revertible_caf_list.
311 // The linker hackily arranges that references to newCaf from dynamic
312 // code end up pointing to newDynCAF.
314 newDynCAF (StgRegTable *reg STG_UNUSED, StgClosure *caf)
318 ((StgIndStatic *)caf)->saved_info = (StgInfoTable *)caf->header.info;
319 ((StgIndStatic *)caf)->static_link = revertible_caf_list;
320 revertible_caf_list = caf;
325 /* -----------------------------------------------------------------------------
327 -------------------------------------------------------------------------- */
330 allocNursery (bdescr *tail, nat blocks)
335 // Allocate a nursery: we allocate fresh blocks one at a time and
336 // cons them on to the front of the list, not forgetting to update
337 // the back pointer on the tail of the list to point to the new block.
338 for (i=0; i < blocks; i++) {
341 processNursery() in LdvProfile.c assumes that every block group in
342 the nursery contains only a single block. So, if a block group is
343 given multiple blocks, change processNursery() accordingly.
347 // double-link the nursery: we might need to insert blocks
351 initBdescr(bd, g0, g0);
353 bd->free = bd->start;
361 assignNurseriesToCapabilities (void)
365 for (i = 0; i < n_capabilities; i++) {
366 capabilities[i].r.rNursery = &nurseries[i];
367 capabilities[i].r.rCurrentNursery = nurseries[i].blocks;
368 capabilities[i].r.rCurrentAlloc = NULL;
373 allocNurseries( void )
377 for (i = 0; i < n_capabilities; i++) {
378 nurseries[i].blocks =
379 allocNursery(NULL, RtsFlags.GcFlags.minAllocAreaSize);
380 nurseries[i].n_blocks =
381 RtsFlags.GcFlags.minAllocAreaSize;
383 assignNurseriesToCapabilities();
387 resetNurseries( void )
392 for (i = 0; i < n_capabilities; i++) {
393 for (bd = nurseries[i].blocks; bd; bd = bd->link) {
394 bd->free = bd->start;
395 ASSERT(bd->gen_no == 0);
396 ASSERT(bd->gen == g0);
397 IF_DEBUG(sanity,memset(bd->start, 0xaa, BLOCK_SIZE));
400 assignNurseriesToCapabilities();
404 countNurseryBlocks (void)
409 for (i = 0; i < n_capabilities; i++) {
410 blocks += nurseries[i].n_blocks;
416 resizeNursery ( nursery *nursery, nat blocks )
421 nursery_blocks = nursery->n_blocks;
422 if (nursery_blocks == blocks) return;
424 if (nursery_blocks < blocks) {
425 debugTrace(DEBUG_gc, "increasing size of nursery to %d blocks",
427 nursery->blocks = allocNursery(nursery->blocks, blocks-nursery_blocks);
432 debugTrace(DEBUG_gc, "decreasing size of nursery to %d blocks",
435 bd = nursery->blocks;
436 while (nursery_blocks > blocks) {
438 next_bd->u.back = NULL;
439 nursery_blocks -= bd->blocks; // might be a large block
443 nursery->blocks = bd;
444 // might have gone just under, by freeing a large block, so make
445 // up the difference.
446 if (nursery_blocks < blocks) {
447 nursery->blocks = allocNursery(nursery->blocks, blocks-nursery_blocks);
451 nursery->n_blocks = blocks;
452 ASSERT(countBlocks(nursery->blocks) == nursery->n_blocks);
456 // Resize each of the nurseries to the specified size.
459 resizeNurseriesFixed (nat blocks)
462 for (i = 0; i < n_capabilities; i++) {
463 resizeNursery(&nurseries[i], blocks);
468 // Resize the nurseries to the total specified size.
471 resizeNurseries (nat blocks)
473 // If there are multiple nurseries, then we just divide the number
474 // of available blocks between them.
475 resizeNurseriesFixed(blocks / n_capabilities);
479 /* -----------------------------------------------------------------------------
480 move_TSO is called to update the TSO structure after it has been
481 moved from one place to another.
482 -------------------------------------------------------------------------- */
485 move_TSO (StgTSO *src, StgTSO *dest)
489 // relocate the stack pointer...
490 diff = (StgPtr)dest - (StgPtr)src; // In *words*
491 dest->sp = (StgPtr)dest->sp + diff;
494 /* -----------------------------------------------------------------------------
495 split N blocks off the front of the given bdescr, returning the
496 new block group. We add the remainder to the large_blocks list
497 in the same step as the original block.
498 -------------------------------------------------------------------------- */
501 splitLargeBlock (bdescr *bd, nat blocks)
507 ASSERT(countBlocks(bd->gen->large_objects) == bd->gen->n_large_blocks);
509 // subtract the original number of blocks from the counter first
510 bd->gen->n_large_blocks -= bd->blocks;
512 new_bd = splitBlockGroup (bd, blocks);
513 initBdescr(new_bd, bd->gen, bd->gen->to);
514 new_bd->flags = BF_LARGE | (bd->flags & BF_EVACUATED);
515 // if new_bd is in an old generation, we have to set BF_EVACUATED
516 new_bd->free = bd->free;
517 dbl_link_onto(new_bd, &bd->gen->large_objects);
519 ASSERT(new_bd->free <= new_bd->start + new_bd->blocks * BLOCK_SIZE_W);
521 // add the new number of blocks to the counter. Due to the gaps
522 // for block descriptors, new_bd->blocks + bd->blocks might not be
523 // equal to the original bd->blocks, which is why we do it this way.
524 bd->gen->n_large_blocks += bd->blocks + new_bd->blocks;
526 ASSERT(countBlocks(bd->gen->large_objects) == bd->gen->n_large_blocks);
533 /* -----------------------------------------------------------------------------
536 This allocates memory in the current thread - it is intended for
537 use primarily from STG-land where we have a Capability. It is
538 better than allocate() because it doesn't require taking the
539 sm_mutex lock in the common case.
541 Memory is allocated directly from the nursery if possible (but not
542 from the current nursery block, so as not to interfere with
544 -------------------------------------------------------------------------- */
547 allocate (Capability *cap, lnat n)
552 if (n >= LARGE_OBJECT_THRESHOLD/sizeof(W_)) {
553 lnat req_blocks = (lnat)BLOCK_ROUND_UP(n*sizeof(W_)) / BLOCK_SIZE;
555 // Attempting to allocate an object larger than maxHeapSize
556 // should definitely be disallowed. (bug #1791)
557 if (RtsFlags.GcFlags.maxHeapSize > 0 &&
558 req_blocks >= RtsFlags.GcFlags.maxHeapSize) {
560 // heapOverflow() doesn't exit (see #2592), but we aren't
561 // in a position to do a clean shutdown here: we
562 // either have to allocate the memory or exit now.
563 // Allocating the memory would be bad, because the user
564 // has requested that we not exceed maxHeapSize, so we
566 stg_exit(EXIT_HEAPOVERFLOW);
570 bd = allocGroup(req_blocks);
571 dbl_link_onto(bd, &g0->large_objects);
572 g0->n_large_blocks += bd->blocks; // might be larger than req_blocks
573 g0->n_new_large_blocks += bd->blocks;
575 initBdescr(bd, g0, g0);
576 bd->flags = BF_LARGE;
577 bd->free = bd->start + n;
581 /* small allocation (<LARGE_OBJECT_THRESHOLD) */
583 TICK_ALLOC_HEAP_NOCTR(n);
586 bd = cap->r.rCurrentAlloc;
587 if (bd == NULL || bd->free + n > bd->start + BLOCK_SIZE_W) {
589 // The CurrentAlloc block is full, we need to find another
590 // one. First, we try taking the next block from the
592 bd = cap->r.rCurrentNursery->link;
594 if (bd == NULL || bd->free + n > bd->start + BLOCK_SIZE_W) {
595 // The nursery is empty, or the next block is already
596 // full: allocate a fresh block (we can't fail here).
599 cap->r.rNursery->n_blocks++;
601 initBdescr(bd, g0, g0);
603 // If we had to allocate a new block, then we'll GC
604 // pretty quickly now, because MAYBE_GC() will
605 // notice that CurrentNursery->link is NULL.
607 // we have a block in the nursery: take it and put
608 // it at the *front* of the nursery list, and use it
609 // to allocate() from.
610 cap->r.rCurrentNursery->link = bd->link;
611 if (bd->link != NULL) {
612 bd->link->u.back = cap->r.rCurrentNursery;
615 dbl_link_onto(bd, &cap->r.rNursery->blocks);
616 cap->r.rCurrentAlloc = bd;
617 IF_DEBUG(sanity, checkNurserySanity(cap->r.rNursery));
622 IF_DEBUG(sanity, ASSERT(*((StgWord8*)p) == 0xaa));
626 /* ---------------------------------------------------------------------------
627 Allocate a fixed/pinned object.
629 We allocate small pinned objects into a single block, allocating a
630 new block when the current one overflows. The block is chained
631 onto the large_object_list of generation 0.
633 NOTE: The GC can't in general handle pinned objects. This
634 interface is only safe to use for ByteArrays, which have no
635 pointers and don't require scavenging. It works because the
636 block's descriptor has the BF_LARGE flag set, so the block is
637 treated as a large object and chained onto various lists, rather
638 than the individual objects being copied. However, when it comes
639 to scavenge the block, the GC will only scavenge the first object.
640 The reason is that the GC can't linearly scan a block of pinned
641 objects at the moment (doing so would require using the
642 mostly-copying techniques). But since we're restricting ourselves
643 to pinned ByteArrays, not scavenging is ok.
645 This function is called by newPinnedByteArray# which immediately
646 fills the allocated memory with a MutableByteArray#.
647 ------------------------------------------------------------------------- */
650 allocatePinned (Capability *cap, lnat n)
655 // If the request is for a large object, then allocate()
656 // will give us a pinned object anyway.
657 if (n >= LARGE_OBJECT_THRESHOLD/sizeof(W_)) {
658 p = allocate(cap, n);
659 Bdescr(p)->flags |= BF_PINNED;
663 TICK_ALLOC_HEAP_NOCTR(n);
666 bd = cap->pinned_object_block;
668 // If we don't have a block of pinned objects yet, or the current
669 // one isn't large enough to hold the new object, allocate a new one.
670 if (bd == NULL || (bd->free + n) > (bd->start + BLOCK_SIZE_W)) {
672 cap->pinned_object_block = bd = allocBlock();
673 dbl_link_onto(bd, &g0->large_objects);
674 g0->n_large_blocks++;
675 g0->n_new_large_blocks++;
677 initBdescr(bd, g0, g0);
678 bd->flags = BF_PINNED | BF_LARGE;
679 bd->free = bd->start;
687 /* -----------------------------------------------------------------------------
689 -------------------------------------------------------------------------- */
692 This is the write barrier for MUT_VARs, a.k.a. IORefs. A
693 MUT_VAR_CLEAN object is not on the mutable list; a MUT_VAR_DIRTY
694 is. When written to, a MUT_VAR_CLEAN turns into a MUT_VAR_DIRTY
695 and is put on the mutable list.
698 dirty_MUT_VAR(StgRegTable *reg, StgClosure *p)
700 Capability *cap = regTableToCapability(reg);
701 if (p->header.info == &stg_MUT_VAR_CLEAN_info) {
702 p->header.info = &stg_MUT_VAR_DIRTY_info;
703 recordClosureMutated(cap,p);
707 // Setting a TSO's link field with a write barrier.
708 // It is *not* necessary to call this function when
709 // * setting the link field to END_TSO_QUEUE
710 // * putting a TSO on the blackhole_queue
711 // * setting the link field of the currently running TSO, as it
712 // will already be dirty.
714 setTSOLink (Capability *cap, StgTSO *tso, StgTSO *target)
716 if (tso->dirty == 0 && (tso->flags & TSO_LINK_DIRTY) == 0) {
717 tso->flags |= TSO_LINK_DIRTY;
718 recordClosureMutated(cap,(StgClosure*)tso);
724 setTSOPrev (Capability *cap, StgTSO *tso, StgTSO *target)
726 if (tso->dirty == 0 && (tso->flags & TSO_LINK_DIRTY) == 0) {
727 tso->flags |= TSO_LINK_DIRTY;
728 recordClosureMutated(cap,(StgClosure*)tso);
730 tso->block_info.prev = target;
734 dirty_TSO (Capability *cap, StgTSO *tso)
736 if (tso->dirty == 0 && (tso->flags & TSO_LINK_DIRTY) == 0) {
737 recordClosureMutated(cap,(StgClosure*)tso);
743 This is the write barrier for MVARs. An MVAR_CLEAN objects is not
744 on the mutable list; a MVAR_DIRTY is. When written to, a
745 MVAR_CLEAN turns into a MVAR_DIRTY and is put on the mutable list.
746 The check for MVAR_CLEAN is inlined at the call site for speed,
747 this really does make a difference on concurrency-heavy benchmarks
748 such as Chaneneos and cheap-concurrency.
751 dirty_MVAR(StgRegTable *reg, StgClosure *p)
753 recordClosureMutated(regTableToCapability(reg),p);
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 );