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 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 // Allocate a nursery: we allocate fresh blocks one at a time and
338 // cons them on to the front of the list, not forgetting to update
339 // the back pointer on the tail of the list to point to the new block.
340 for (i=0; i < blocks; i++) {
343 processNursery() in LdvProfile.c assumes that every block group in
344 the nursery contains only a single block. So, if a block group is
345 given multiple blocks, change processNursery() accordingly.
349 // double-link the nursery: we might need to insert blocks
353 initBdescr(bd, g0, g0);
355 bd->free = bd->start;
363 assignNurseriesToCapabilities (void)
367 for (i = 0; i < n_capabilities; i++) {
368 capabilities[i].r.rNursery = &nurseries[i];
369 capabilities[i].r.rCurrentNursery = nurseries[i].blocks;
370 capabilities[i].r.rCurrentAlloc = NULL;
375 allocNurseries( void )
379 for (i = 0; i < n_capabilities; i++) {
380 nurseries[i].blocks =
381 allocNursery(NULL, RtsFlags.GcFlags.minAllocAreaSize);
382 nurseries[i].n_blocks =
383 RtsFlags.GcFlags.minAllocAreaSize;
385 assignNurseriesToCapabilities();
389 resetNurseries( void )
394 for (i = 0; i < n_capabilities; i++) {
395 for (bd = nurseries[i].blocks; bd; bd = bd->link) {
396 bd->free = bd->start;
397 ASSERT(bd->gen_no == 0);
398 ASSERT(bd->gen == g0);
399 IF_DEBUG(sanity,memset(bd->start, 0xaa, BLOCK_SIZE));
402 assignNurseriesToCapabilities();
406 countNurseryBlocks (void)
411 for (i = 0; i < n_capabilities; i++) {
412 blocks += nurseries[i].n_blocks;
418 resizeNursery ( nursery *nursery, nat blocks )
423 nursery_blocks = nursery->n_blocks;
424 if (nursery_blocks == blocks) return;
426 if (nursery_blocks < blocks) {
427 debugTrace(DEBUG_gc, "increasing size of nursery to %d blocks",
429 nursery->blocks = allocNursery(nursery->blocks, blocks-nursery_blocks);
434 debugTrace(DEBUG_gc, "decreasing size of nursery to %d blocks",
437 bd = nursery->blocks;
438 while (nursery_blocks > blocks) {
440 next_bd->u.back = NULL;
441 nursery_blocks -= bd->blocks; // might be a large block
445 nursery->blocks = bd;
446 // might have gone just under, by freeing a large block, so make
447 // up the difference.
448 if (nursery_blocks < blocks) {
449 nursery->blocks = allocNursery(nursery->blocks, blocks-nursery_blocks);
453 nursery->n_blocks = blocks;
454 ASSERT(countBlocks(nursery->blocks) == nursery->n_blocks);
458 // Resize each of the nurseries to the specified size.
461 resizeNurseriesFixed (nat blocks)
464 for (i = 0; i < n_capabilities; i++) {
465 resizeNursery(&nurseries[i], blocks);
470 // Resize the nurseries to the total specified size.
473 resizeNurseries (nat blocks)
475 // If there are multiple nurseries, then we just divide the number
476 // of available blocks between them.
477 resizeNurseriesFixed(blocks / n_capabilities);
481 /* -----------------------------------------------------------------------------
482 move_TSO is called to update the TSO structure after it has been
483 moved from one place to another.
484 -------------------------------------------------------------------------- */
487 move_TSO (StgTSO *src, StgTSO *dest)
491 // relocate the stack pointer...
492 diff = (StgPtr)dest - (StgPtr)src; // In *words*
493 dest->sp = (StgPtr)dest->sp + diff;
496 /* -----------------------------------------------------------------------------
497 split N blocks off the front of the given bdescr, returning the
498 new block group. We add the remainder to the large_blocks list
499 in the same step as the original block.
500 -------------------------------------------------------------------------- */
503 splitLargeBlock (bdescr *bd, nat blocks)
509 ASSERT(countBlocks(bd->gen->large_objects) == bd->gen->n_large_blocks);
511 // subtract the original number of blocks from the counter first
512 bd->gen->n_large_blocks -= bd->blocks;
514 new_bd = splitBlockGroup (bd, blocks);
515 initBdescr(new_bd, bd->gen, bd->gen->to);
516 new_bd->flags = BF_LARGE | (bd->flags & BF_EVACUATED);
517 // if new_bd is in an old generation, we have to set BF_EVACUATED
518 new_bd->free = bd->free;
519 dbl_link_onto(new_bd, &bd->gen->large_objects);
521 ASSERT(new_bd->free <= new_bd->start + new_bd->blocks * BLOCK_SIZE_W);
523 // add the new number of blocks to the counter. Due to the gaps
524 // for block descriptors, new_bd->blocks + bd->blocks might not be
525 // equal to the original bd->blocks, which is why we do it this way.
526 bd->gen->n_large_blocks += bd->blocks + new_bd->blocks;
528 ASSERT(countBlocks(bd->gen->large_objects) == bd->gen->n_large_blocks);
535 /* -----------------------------------------------------------------------------
538 This allocates memory in the current thread - it is intended for
539 use primarily from STG-land where we have a Capability. It is
540 better than allocate() because it doesn't require taking the
541 sm_mutex lock in the common case.
543 Memory is allocated directly from the nursery if possible (but not
544 from the current nursery block, so as not to interfere with
546 -------------------------------------------------------------------------- */
549 allocate (Capability *cap, lnat n)
554 if (n >= LARGE_OBJECT_THRESHOLD/sizeof(W_)) {
555 lnat req_blocks = (lnat)BLOCK_ROUND_UP(n*sizeof(W_)) / BLOCK_SIZE;
557 // Attempting to allocate an object larger than maxHeapSize
558 // should definitely be disallowed. (bug #1791)
559 if (RtsFlags.GcFlags.maxHeapSize > 0 &&
560 req_blocks >= RtsFlags.GcFlags.maxHeapSize) {
562 // heapOverflow() doesn't exit (see #2592), but we aren't
563 // in a position to do a clean shutdown here: we
564 // either have to allocate the memory or exit now.
565 // Allocating the memory would be bad, because the user
566 // has requested that we not exceed maxHeapSize, so we
568 stg_exit(EXIT_HEAPOVERFLOW);
572 bd = allocGroup(req_blocks);
573 dbl_link_onto(bd, &g0->large_objects);
574 g0->n_large_blocks += bd->blocks; // might be larger than req_blocks
575 g0->n_new_large_blocks += bd->blocks;
577 initBdescr(bd, g0, g0);
578 bd->flags = BF_LARGE;
579 bd->free = bd->start + n;
583 /* small allocation (<LARGE_OBJECT_THRESHOLD) */
585 TICK_ALLOC_HEAP_NOCTR(n);
588 bd = cap->r.rCurrentAlloc;
589 if (bd == NULL || bd->free + n > bd->start + BLOCK_SIZE_W) {
591 // The CurrentAlloc block is full, we need to find another
592 // one. First, we try taking the next block from the
594 bd = cap->r.rCurrentNursery->link;
596 if (bd == NULL || bd->free + n > bd->start + BLOCK_SIZE_W) {
597 // The nursery is empty, or the next block is already
598 // full: allocate a fresh block (we can't fail here).
601 cap->r.rNursery->n_blocks++;
603 initBdescr(bd, g0, g0);
605 // If we had to allocate a new block, then we'll GC
606 // pretty quickly now, because MAYBE_GC() will
607 // notice that CurrentNursery->link is NULL.
609 // we have a block in the nursery: take it and put
610 // it at the *front* of the nursery list, and use it
611 // to allocate() from.
612 cap->r.rCurrentNursery->link = bd->link;
613 if (bd->link != NULL) {
614 bd->link->u.back = cap->r.rCurrentNursery;
617 dbl_link_onto(bd, &cap->r.rNursery->blocks);
618 cap->r.rCurrentAlloc = bd;
619 IF_DEBUG(sanity, checkNurserySanity(cap->r.rNursery));
624 IF_DEBUG(sanity, ASSERT(*((StgWord8*)p) == 0xaa));
628 /* ---------------------------------------------------------------------------
629 Allocate a fixed/pinned object.
631 We allocate small pinned objects into a single block, allocating a
632 new block when the current one overflows. The block is chained
633 onto the large_object_list of generation 0.
635 NOTE: The GC can't in general handle pinned objects. This
636 interface is only safe to use for ByteArrays, which have no
637 pointers and don't require scavenging. It works because the
638 block's descriptor has the BF_LARGE flag set, so the block is
639 treated as a large object and chained onto various lists, rather
640 than the individual objects being copied. However, when it comes
641 to scavenge the block, the GC will only scavenge the first object.
642 The reason is that the GC can't linearly scan a block of pinned
643 objects at the moment (doing so would require using the
644 mostly-copying techniques). But since we're restricting ourselves
645 to pinned ByteArrays, not scavenging is ok.
647 This function is called by newPinnedByteArray# which immediately
648 fills the allocated memory with a MutableByteArray#.
649 ------------------------------------------------------------------------- */
652 allocatePinned (Capability *cap, lnat n)
657 // If the request is for a large object, then allocate()
658 // will give us a pinned object anyway.
659 if (n >= LARGE_OBJECT_THRESHOLD/sizeof(W_)) {
660 p = allocate(cap, n);
661 Bdescr(p)->flags |= BF_PINNED;
665 TICK_ALLOC_HEAP_NOCTR(n);
668 bd = cap->pinned_object_block;
670 // If we don't have a block of pinned objects yet, or the current
671 // one isn't large enough to hold the new object, allocate a new one.
672 if (bd == NULL || (bd->free + n) > (bd->start + BLOCK_SIZE_W)) {
674 cap->pinned_object_block = bd = allocBlock();
675 dbl_link_onto(bd, &g0->large_objects);
676 g0->n_large_blocks++;
677 g0->n_new_large_blocks++;
679 initBdescr(bd, g0, g0);
680 bd->flags = BF_PINNED | BF_LARGE;
681 bd->free = bd->start;
689 /* -----------------------------------------------------------------------------
691 -------------------------------------------------------------------------- */
694 This is the write barrier for MUT_VARs, a.k.a. IORefs. A
695 MUT_VAR_CLEAN object is not on the mutable list; a MUT_VAR_DIRTY
696 is. When written to, a MUT_VAR_CLEAN turns into a MUT_VAR_DIRTY
697 and is put on the mutable list.
700 dirty_MUT_VAR(StgRegTable *reg, StgClosure *p)
702 Capability *cap = regTableToCapability(reg);
703 if (p->header.info == &stg_MUT_VAR_CLEAN_info) {
704 p->header.info = &stg_MUT_VAR_DIRTY_info;
705 recordClosureMutated(cap,p);
709 // Setting a TSO's link field with a write barrier.
710 // It is *not* necessary to call this function when
711 // * setting the link field to END_TSO_QUEUE
712 // * putting a TSO on the blackhole_queue
713 // * setting the link field of the currently running TSO, as it
714 // will already be dirty.
716 setTSOLink (Capability *cap, StgTSO *tso, StgTSO *target)
718 if (tso->dirty == 0 && (tso->flags & TSO_LINK_DIRTY) == 0) {
719 tso->flags |= TSO_LINK_DIRTY;
720 recordClosureMutated(cap,(StgClosure*)tso);
726 setTSOPrev (Capability *cap, StgTSO *tso, StgTSO *target)
728 if (tso->dirty == 0 && (tso->flags & TSO_LINK_DIRTY) == 0) {
729 tso->flags |= TSO_LINK_DIRTY;
730 recordClosureMutated(cap,(StgClosure*)tso);
732 tso->block_info.prev = target;
736 dirty_TSO (Capability *cap, StgTSO *tso)
738 if (tso->dirty == 0 && (tso->flags & TSO_LINK_DIRTY) == 0) {
739 recordClosureMutated(cap,(StgClosure*)tso);
745 This is the write barrier for MVARs. An MVAR_CLEAN objects is not
746 on the mutable list; a MVAR_DIRTY is. When written to, a
747 MVAR_CLEAN turns into a MVAR_DIRTY and is put on the mutable list.
748 The check for MVAR_CLEAN is inlined at the call site for speed,
749 this really does make a difference on concurrency-heavy benchmarks
750 such as Chaneneos and cheap-concurrency.
753 dirty_MVAR(StgRegTable *reg, StgClosure *p)
755 recordClosureMutated(regTableToCapability(reg),p);
758 /* -----------------------------------------------------------------------------
760 * -------------------------------------------------------------------------- */
762 /* -----------------------------------------------------------------------------
765 * Approximate how much we've allocated: number of blocks in the
766 * nursery + blocks allocated via allocate() - unused nusery blocks.
767 * This leaves a little slop at the end of each block.
768 * -------------------------------------------------------------------------- */
771 calcAllocated( void )
777 allocated = countNurseryBlocks() * BLOCK_SIZE_W;
779 for (i = 0; i < n_capabilities; i++) {
781 for ( bd = capabilities[i].r.rCurrentNursery->link;
782 bd != NULL; bd = bd->link ) {
783 allocated -= BLOCK_SIZE_W;
785 cap = &capabilities[i];
786 if (cap->r.rCurrentNursery->free <
787 cap->r.rCurrentNursery->start + BLOCK_SIZE_W) {
788 allocated -= (cap->r.rCurrentNursery->start + BLOCK_SIZE_W)
789 - cap->r.rCurrentNursery->free;
791 if (cap->pinned_object_block != NULL) {
792 allocated -= (cap->pinned_object_block->start + BLOCK_SIZE_W) -
793 cap->pinned_object_block->free;
797 allocated += g0->n_new_large_blocks * BLOCK_SIZE_W;
802 /* Approximate the amount of live data in the heap. To be called just
803 * after garbage collection (see GarbageCollect()).
805 lnat calcLiveBlocks (void)
811 for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
812 /* approximate amount of live data (doesn't take into account slop
813 * at end of each block).
815 gen = &generations[g];
816 live += gen->n_large_blocks + gen->n_blocks;
821 lnat countOccupied (bdescr *bd)
826 for (; bd != NULL; bd = bd->link) {
827 ASSERT(bd->free <= bd->start + bd->blocks * BLOCK_SIZE_W);
828 words += bd->free - bd->start;
833 // Return an accurate count of the live data in the heap, excluding
835 lnat calcLiveWords (void)
842 for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
843 gen = &generations[g];
844 live += gen->n_words + countOccupied(gen->large_objects);
849 /* Approximate the number of blocks that will be needed at the next
850 * garbage collection.
852 * Assume: all data currently live will remain live. Generationss
853 * that will be collected next time will therefore need twice as many
854 * blocks since all the data will be copied.
863 for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
864 gen = &generations[g];
866 // we need at least this much space
867 needed += gen->n_blocks + gen->n_large_blocks;
869 // any additional space needed to collect this gen next time?
870 if (g == 0 || // always collect gen 0
871 (gen->n_blocks + gen->n_large_blocks > gen->max_blocks)) {
872 // we will collect this gen next time
875 needed += gen->n_blocks / BITS_IN(W_);
877 needed += gen->n_blocks / 100;
880 continue; // no additional space needed for compaction
882 needed += gen->n_blocks;
889 /* ----------------------------------------------------------------------------
892 Executable memory must be managed separately from non-executable
893 memory. Most OSs these days require you to jump through hoops to
894 dynamically allocate executable memory, due to various security
897 Here we provide a small memory allocator for executable memory.
898 Memory is managed with a page granularity; we allocate linearly
899 in the page, and when the page is emptied (all objects on the page
900 are free) we free the page again, not forgetting to make it
903 TODO: The inability to handle objects bigger than BLOCK_SIZE_W means that
904 the linker cannot use allocateExec for loading object code files
905 on Windows. Once allocateExec can handle larger objects, the linker
906 should be modified to use allocateExec instead of VirtualAlloc.
907 ------------------------------------------------------------------------- */
909 #if defined(linux_HOST_OS)
911 // On Linux we need to use libffi for allocating executable memory,
912 // because it knows how to work around the restrictions put in place
915 void *allocateExec (nat bytes, void **exec_ret)
919 ret = ffi_closure_alloc (sizeof(void *) + (size_t)bytes, (void**)&exec);
921 if (ret == NULL) return ret;
922 *ret = ret; // save the address of the writable mapping, for freeExec().
923 *exec_ret = exec + 1;
927 // freeExec gets passed the executable address, not the writable address.
928 void freeExec (void *addr)
931 writable = *((void**)addr - 1);
933 ffi_closure_free (writable);
939 void *allocateExec (nat bytes, void **exec_ret)
946 // round up to words.
947 n = (bytes + sizeof(W_) + 1) / sizeof(W_);
949 if (n+1 > BLOCK_SIZE_W) {
950 barf("allocateExec: can't handle large objects");
953 if (exec_block == NULL ||
954 exec_block->free + n + 1 > exec_block->start + BLOCK_SIZE_W) {
956 lnat pagesize = getPageSize();
957 bd = allocGroup(stg_max(1, pagesize / BLOCK_SIZE));
958 debugTrace(DEBUG_gc, "allocate exec block %p", bd->start);
961 bd->link = exec_block;
962 if (exec_block != NULL) {
963 exec_block->u.back = bd;
966 setExecutable(bd->start, bd->blocks * BLOCK_SIZE, rtsTrue);
969 *(exec_block->free) = n; // store the size of this chunk
970 exec_block->gen_no += n; // gen_no stores the number of words allocated
971 ret = exec_block->free + 1;
972 exec_block->free += n + 1;
979 void freeExec (void *addr)
981 StgPtr p = (StgPtr)addr - 1;
982 bdescr *bd = Bdescr((StgPtr)p);
984 if ((bd->flags & BF_EXEC) == 0) {
985 barf("freeExec: not executable");
988 if (*(StgPtr)p == 0) {
989 barf("freeExec: already free?");
994 bd->gen_no -= *(StgPtr)p;
997 if (bd->gen_no == 0) {
998 // Free the block if it is empty, but not if it is the block at
999 // the head of the queue.
1000 if (bd != exec_block) {
1001 debugTrace(DEBUG_gc, "free exec block %p", bd->start);
1002 dbl_link_remove(bd, &exec_block);
1003 setExecutable(bd->start, bd->blocks * BLOCK_SIZE, rtsFalse);
1006 bd->free = bd->start;
1013 #endif /* mingw32_HOST_OS */
1017 // handy function for use in gdb, because Bdescr() is inlined.
1018 extern bdescr *_bdescr( StgPtr p );