1 /* -----------------------------------------------------------------------------
3 * (c) The GHC Team, 1998-2006
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"
25 #include "OSThreads.h"
26 #include "Capability.h"
29 #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 bdescr *pinned_object_block; /* allocate pinned objects into this block */
44 nat alloc_blocks; /* number of allocate()d blocks since GC */
45 nat alloc_blocks_lim; /* approximate limit on alloc_blocks */
47 generation *generations = NULL; /* all the generations */
48 generation *g0 = NULL; /* generation 0, for convenience */
49 generation *oldest_gen = NULL; /* oldest generation, for convenience */
50 step *g0s0 = NULL; /* generation 0, step 0, for convenience */
52 ullong total_allocated = 0; /* total memory allocated during run */
54 nat n_nurseries = 0; /* == RtsFlags.ParFlags.nNodes, convenience */
55 step *nurseries = NULL; /* array of nurseries, >1 only if THREADED_RTS */
59 * Storage manager mutex: protects all the above state from
60 * simultaneous access by two STG threads.
64 * This mutex is used by atomicModifyMutVar# only
66 Mutex atomic_modify_mutvar_mutex;
73 static void *stgAllocForGMP (size_t size_in_bytes);
74 static void *stgReallocForGMP (void *ptr, size_t old_size, size_t new_size);
75 static void stgDeallocForGMP (void *ptr, size_t size);
78 initStep (step *stp, int g, int s)
83 stp->old_blocks = NULL;
84 stp->n_old_blocks = 0;
85 stp->gen = &generations[g];
91 stp->scavd_hpLim = NULL;
94 stp->large_objects = NULL;
95 stp->n_large_blocks = 0;
96 stp->new_large_objects = NULL;
97 stp->scavenged_large_objects = NULL;
98 stp->n_scavenged_large_blocks = 0;
99 stp->is_compacted = 0;
109 if (generations != NULL) {
110 // multi-init protection
116 /* Sanity check to make sure the LOOKS_LIKE_ macros appear to be
117 * doing something reasonable.
119 ASSERT(LOOKS_LIKE_INFO_PTR(&stg_BLACKHOLE_info));
120 ASSERT(LOOKS_LIKE_CLOSURE_PTR(&stg_dummy_ret_closure));
121 ASSERT(!HEAP_ALLOCED(&stg_dummy_ret_closure));
123 if (RtsFlags.GcFlags.maxHeapSize != 0 &&
124 RtsFlags.GcFlags.heapSizeSuggestion >
125 RtsFlags.GcFlags.maxHeapSize) {
126 RtsFlags.GcFlags.maxHeapSize = RtsFlags.GcFlags.heapSizeSuggestion;
129 if (RtsFlags.GcFlags.maxHeapSize != 0 &&
130 RtsFlags.GcFlags.minAllocAreaSize >
131 RtsFlags.GcFlags.maxHeapSize) {
132 errorBelch("maximum heap size (-M) is smaller than minimum alloc area size (-A)");
133 RtsFlags.GcFlags.minAllocAreaSize = RtsFlags.GcFlags.maxHeapSize;
136 initBlockAllocator();
138 #if defined(THREADED_RTS)
139 initMutex(&sm_mutex);
140 initMutex(&atomic_modify_mutvar_mutex);
145 /* allocate generation info array */
146 generations = (generation *)stgMallocBytes(RtsFlags.GcFlags.generations
147 * sizeof(struct generation_),
148 "initStorage: gens");
150 /* Initialise all generations */
151 for(g = 0; g < RtsFlags.GcFlags.generations; g++) {
152 gen = &generations[g];
154 gen->mut_list = allocBlock();
155 gen->collections = 0;
156 gen->failed_promotions = 0;
160 /* A couple of convenience pointers */
161 g0 = &generations[0];
162 oldest_gen = &generations[RtsFlags.GcFlags.generations-1];
164 /* Allocate step structures in each generation */
165 if (RtsFlags.GcFlags.generations > 1) {
166 /* Only for multiple-generations */
168 /* Oldest generation: one step */
169 oldest_gen->n_steps = 1;
171 stgMallocBytes(1 * sizeof(struct step_), "initStorage: last step");
173 /* set up all except the oldest generation with 2 steps */
174 for(g = 0; g < RtsFlags.GcFlags.generations-1; g++) {
175 generations[g].n_steps = RtsFlags.GcFlags.steps;
176 generations[g].steps =
177 stgMallocBytes (RtsFlags.GcFlags.steps * sizeof(struct step_),
178 "initStorage: steps");
182 /* single generation, i.e. a two-space collector */
184 g0->steps = stgMallocBytes (sizeof(struct step_), "initStorage: steps");
188 n_nurseries = n_capabilities;
192 nurseries = stgMallocBytes (n_nurseries * sizeof(struct step_),
193 "initStorage: nurseries");
195 /* Initialise all steps */
196 for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
197 for (s = 0; s < generations[g].n_steps; s++) {
198 initStep(&generations[g].steps[s], g, s);
202 for (s = 0; s < n_nurseries; s++) {
203 initStep(&nurseries[s], 0, s);
206 /* Set up the destination pointers in each younger gen. step */
207 for (g = 0; g < RtsFlags.GcFlags.generations-1; g++) {
208 for (s = 0; s < generations[g].n_steps-1; s++) {
209 generations[g].steps[s].to = &generations[g].steps[s+1];
211 generations[g].steps[s].to = &generations[g+1].steps[0];
213 oldest_gen->steps[0].to = &oldest_gen->steps[0];
215 for (s = 0; s < n_nurseries; s++) {
216 nurseries[s].to = generations[0].steps[0].to;
219 /* The oldest generation has one step. */
220 if (RtsFlags.GcFlags.compact) {
221 if (RtsFlags.GcFlags.generations == 1) {
222 errorBelch("WARNING: compaction is incompatible with -G1; disabled");
224 oldest_gen->steps[0].is_compacted = 1;
228 generations[0].max_blocks = 0;
229 g0s0 = &generations[0].steps[0];
231 /* The allocation area. Policy: keep the allocation area
232 * small to begin with, even if we have a large suggested heap
233 * size. Reason: we're going to do a major collection first, and we
234 * don't want it to be a big one. This vague idea is borne out by
235 * rigorous experimental evidence.
239 weak_ptr_list = NULL;
241 revertible_caf_list = NULL;
243 /* initialise the allocate() interface */
245 alloc_blocks_lim = RtsFlags.GcFlags.minAllocAreaSize;
247 /* Tell GNU multi-precision pkg about our custom alloc functions */
248 mp_set_memory_functions(stgAllocForGMP, stgReallocForGMP, stgDeallocForGMP);
250 IF_DEBUG(gc, statDescribeGens());
258 stat_exit(calcAllocated());
266 for(g = 0; g < RtsFlags.GcFlags.generations; g++)
267 stgFree(generations[g].steps);
268 stgFree(generations);
270 #if defined(THREADED_RTS)
271 closeMutex(&sm_mutex);
272 closeMutex(&atomic_modify_mutvar_mutex);
277 /* -----------------------------------------------------------------------------
280 The entry code for every CAF does the following:
282 - builds a CAF_BLACKHOLE in the heap
283 - pushes an update frame pointing to the CAF_BLACKHOLE
284 - invokes UPD_CAF(), which:
285 - calls newCaf, below
286 - updates the CAF with a static indirection to the CAF_BLACKHOLE
288 Why do we build a BLACKHOLE in the heap rather than just updating
289 the thunk directly? It's so that we only need one kind of update
290 frame - otherwise we'd need a static version of the update frame too.
292 newCaf() does the following:
294 - it puts the CAF on the oldest generation's mut-once list.
295 This is so that we can treat the CAF as a root when collecting
298 For GHCI, we have additional requirements when dealing with CAFs:
300 - we must *retain* all dynamically-loaded CAFs ever entered,
301 just in case we need them again.
302 - we must be able to *revert* CAFs that have been evaluated, to
303 their pre-evaluated form.
305 To do this, we use an additional CAF list. When newCaf() is
306 called on a dynamically-loaded CAF, we add it to the CAF list
307 instead of the old-generation mutable list, and save away its
308 old info pointer (in caf->saved_info) for later reversion.
310 To revert all the CAFs, we traverse the CAF list and reset the
311 info pointer to caf->saved_info, then throw away the CAF list.
312 (see GC.c:revertCAFs()).
316 -------------------------------------------------------------------------- */
319 newCAF(StgClosure* caf)
326 // If we are in GHCi _and_ we are using dynamic libraries,
327 // then we can't redirect newCAF calls to newDynCAF (see below),
328 // so we make newCAF behave almost like newDynCAF.
329 // The dynamic libraries might be used by both the interpreted
330 // program and GHCi itself, so they must not be reverted.
331 // This also means that in GHCi with dynamic libraries, CAFs are not
332 // garbage collected. If this turns out to be a problem, we could
333 // do another hack here and do an address range test on caf to figure
334 // out whether it is from a dynamic library.
335 ((StgIndStatic *)caf)->saved_info = (StgInfoTable *)caf->header.info;
336 ((StgIndStatic *)caf)->static_link = caf_list;
341 /* Put this CAF on the mutable list for the old generation.
342 * This is a HACK - the IND_STATIC closure doesn't really have
343 * a mut_link field, but we pretend it has - in fact we re-use
344 * the STATIC_LINK field for the time being, because when we
345 * come to do a major GC we won't need the mut_link field
346 * any more and can use it as a STATIC_LINK.
348 ((StgIndStatic *)caf)->saved_info = NULL;
349 recordMutableGen(caf, oldest_gen);
355 // An alternate version of newCaf which is used for dynamically loaded
356 // object code in GHCi. In this case we want to retain *all* CAFs in
357 // the object code, because they might be demanded at any time from an
358 // expression evaluated on the command line.
359 // Also, GHCi might want to revert CAFs, so we add these to the
360 // revertible_caf_list.
362 // The linker hackily arranges that references to newCaf from dynamic
363 // code end up pointing to newDynCAF.
365 newDynCAF(StgClosure *caf)
369 ((StgIndStatic *)caf)->saved_info = (StgInfoTable *)caf->header.info;
370 ((StgIndStatic *)caf)->static_link = revertible_caf_list;
371 revertible_caf_list = caf;
376 /* -----------------------------------------------------------------------------
378 -------------------------------------------------------------------------- */
381 allocNursery (step *stp, bdescr *tail, nat blocks)
386 // Allocate a nursery: we allocate fresh blocks one at a time and
387 // cons them on to the front of the list, not forgetting to update
388 // the back pointer on the tail of the list to point to the new block.
389 for (i=0; i < blocks; i++) {
392 processNursery() in LdvProfile.c assumes that every block group in
393 the nursery contains only a single block. So, if a block group is
394 given multiple blocks, change processNursery() accordingly.
398 // double-link the nursery: we might need to insert blocks
405 bd->free = bd->start;
413 assignNurseriesToCapabilities (void)
418 for (i = 0; i < n_nurseries; i++) {
419 capabilities[i].r.rNursery = &nurseries[i];
420 capabilities[i].r.rCurrentNursery = nurseries[i].blocks;
421 capabilities[i].r.rCurrentAlloc = NULL;
423 #else /* THREADED_RTS */
424 MainCapability.r.rNursery = &nurseries[0];
425 MainCapability.r.rCurrentNursery = nurseries[0].blocks;
426 MainCapability.r.rCurrentAlloc = NULL;
431 allocNurseries( void )
435 for (i = 0; i < n_nurseries; i++) {
436 nurseries[i].blocks =
437 allocNursery(&nurseries[i], NULL,
438 RtsFlags.GcFlags.minAllocAreaSize);
439 nurseries[i].n_blocks = RtsFlags.GcFlags.minAllocAreaSize;
440 nurseries[i].old_blocks = NULL;
441 nurseries[i].n_old_blocks = 0;
443 assignNurseriesToCapabilities();
447 resetNurseries( void )
453 for (i = 0; i < n_nurseries; i++) {
455 for (bd = stp->blocks; bd; bd = bd->link) {
456 bd->free = bd->start;
457 ASSERT(bd->gen_no == 0);
458 ASSERT(bd->step == stp);
459 IF_DEBUG(sanity,memset(bd->start, 0xaa, BLOCK_SIZE));
462 assignNurseriesToCapabilities();
466 countNurseryBlocks (void)
471 for (i = 0; i < n_nurseries; i++) {
472 blocks += nurseries[i].n_blocks;
478 resizeNursery ( step *stp, nat blocks )
483 nursery_blocks = stp->n_blocks;
484 if (nursery_blocks == blocks) return;
486 if (nursery_blocks < blocks) {
487 debugTrace(DEBUG_gc, "increasing size of nursery to %d blocks",
489 stp->blocks = allocNursery(stp, stp->blocks, blocks-nursery_blocks);
494 debugTrace(DEBUG_gc, "decreasing size of nursery to %d blocks",
498 while (nursery_blocks > blocks) {
500 next_bd->u.back = NULL;
501 nursery_blocks -= bd->blocks; // might be a large block
506 // might have gone just under, by freeing a large block, so make
507 // up the difference.
508 if (nursery_blocks < blocks) {
509 stp->blocks = allocNursery(stp, stp->blocks, blocks-nursery_blocks);
513 stp->n_blocks = blocks;
514 ASSERT(countBlocks(stp->blocks) == stp->n_blocks);
518 // Resize each of the nurseries to the specified size.
521 resizeNurseriesFixed (nat blocks)
524 for (i = 0; i < n_nurseries; i++) {
525 resizeNursery(&nurseries[i], blocks);
530 // Resize the nurseries to the total specified size.
533 resizeNurseries (nat blocks)
535 // If there are multiple nurseries, then we just divide the number
536 // of available blocks between them.
537 resizeNurseriesFixed(blocks / n_nurseries);
540 /* -----------------------------------------------------------------------------
541 The allocate() interface
543 allocateInGen() function allocates memory directly into a specific
544 generation. It always succeeds, and returns a chunk of memory n
545 words long. n can be larger than the size of a block if necessary,
546 in which case a contiguous block group will be allocated.
548 allocate(n) is equivalent to allocateInGen(g0).
549 -------------------------------------------------------------------------- */
552 allocateInGen (generation *g, nat n)
560 TICK_ALLOC_HEAP_NOCTR(n);
565 if (n >= LARGE_OBJECT_THRESHOLD/sizeof(W_))
567 nat req_blocks = (lnat)BLOCK_ROUND_UP(n*sizeof(W_)) / BLOCK_SIZE;
568 bd = allocGroup(req_blocks);
569 dbl_link_onto(bd, &stp->large_objects);
570 stp->n_large_blocks += bd->blocks; // might be larger than req_blocks
573 bd->flags = BF_LARGE;
574 bd->free = bd->start + n;
579 // small allocation (<LARGE_OBJECT_THRESHOLD) */
581 if (bd == NULL || bd->free + n > bd->start + BLOCK_SIZE_W) {
586 bd->link = stp->blocks;
603 return allocateInGen(g0,n);
607 allocatedBytes( void )
611 allocated = alloc_blocks * BLOCK_SIZE_W;
612 if (pinned_object_block != NULL) {
613 allocated -= (pinned_object_block->start + BLOCK_SIZE_W) -
614 pinned_object_block->free;
620 /* -----------------------------------------------------------------------------
623 This allocates memory in the current thread - it is intended for
624 use primarily from STG-land where we have a Capability. It is
625 better than allocate() because it doesn't require taking the
626 sm_mutex lock in the common case.
628 Memory is allocated directly from the nursery if possible (but not
629 from the current nursery block, so as not to interfere with
631 -------------------------------------------------------------------------- */
634 allocateLocal (Capability *cap, nat n)
639 if (n >= LARGE_OBJECT_THRESHOLD/sizeof(W_)) {
640 return allocateInGen(g0,n);
643 /* small allocation (<LARGE_OBJECT_THRESHOLD) */
645 TICK_ALLOC_HEAP_NOCTR(n);
648 bd = cap->r.rCurrentAlloc;
649 if (bd == NULL || bd->free + n > bd->start + BLOCK_SIZE_W) {
651 // The CurrentAlloc block is full, we need to find another
652 // one. First, we try taking the next block from the
654 bd = cap->r.rCurrentNursery->link;
656 if (bd == NULL || bd->free + n > bd->start + BLOCK_SIZE_W) {
657 // The nursery is empty, or the next block is already
658 // full: allocate a fresh block (we can't fail here).
661 cap->r.rNursery->n_blocks++;
664 bd->step = cap->r.rNursery;
666 // NO: alloc_blocks++;
667 // calcAllocated() uses the size of the nursery, and we've
668 // already bumpted nursery->n_blocks above.
670 // we have a block in the nursery: take it and put
671 // it at the *front* of the nursery list, and use it
672 // to allocate() from.
673 cap->r.rCurrentNursery->link = bd->link;
674 if (bd->link != NULL) {
675 bd->link->u.back = cap->r.rCurrentNursery;
678 dbl_link_onto(bd, &cap->r.rNursery->blocks);
679 cap->r.rCurrentAlloc = bd;
680 IF_DEBUG(sanity, checkNurserySanity(cap->r.rNursery));
687 /* ---------------------------------------------------------------------------
688 Allocate a fixed/pinned object.
690 We allocate small pinned objects into a single block, allocating a
691 new block when the current one overflows. The block is chained
692 onto the large_object_list of generation 0 step 0.
694 NOTE: The GC can't in general handle pinned objects. This
695 interface is only safe to use for ByteArrays, which have no
696 pointers and don't require scavenging. It works because the
697 block's descriptor has the BF_LARGE flag set, so the block is
698 treated as a large object and chained onto various lists, rather
699 than the individual objects being copied. However, when it comes
700 to scavenge the block, the GC will only scavenge the first object.
701 The reason is that the GC can't linearly scan a block of pinned
702 objects at the moment (doing so would require using the
703 mostly-copying techniques). But since we're restricting ourselves
704 to pinned ByteArrays, not scavenging is ok.
706 This function is called by newPinnedByteArray# which immediately
707 fills the allocated memory with a MutableByteArray#.
708 ------------------------------------------------------------------------- */
711 allocatePinned( nat n )
714 bdescr *bd = pinned_object_block;
716 // If the request is for a large object, then allocate()
717 // will give us a pinned object anyway.
718 if (n >= LARGE_OBJECT_THRESHOLD/sizeof(W_)) {
724 TICK_ALLOC_HEAP_NOCTR(n);
727 // we always return 8-byte aligned memory. bd->free must be
728 // 8-byte aligned to begin with, so we just round up n to
729 // the nearest multiple of 8 bytes.
730 if (sizeof(StgWord) == 4) {
734 // If we don't have a block of pinned objects yet, or the current
735 // one isn't large enough to hold the new object, allocate a new one.
736 if (bd == NULL || (bd->free + n) > (bd->start + BLOCK_SIZE_W)) {
737 pinned_object_block = bd = allocBlock();
738 dbl_link_onto(bd, &g0s0->large_objects);
739 g0s0->n_large_blocks++;
742 bd->flags = BF_PINNED | BF_LARGE;
743 bd->free = bd->start;
753 /* -----------------------------------------------------------------------------
755 -------------------------------------------------------------------------- */
758 This is the write barrier for MUT_VARs, a.k.a. IORefs. A
759 MUT_VAR_CLEAN object is not on the mutable list; a MUT_VAR_DIRTY
760 is. When written to, a MUT_VAR_CLEAN turns into a MUT_VAR_DIRTY
761 and is put on the mutable list.
764 dirty_MUT_VAR(StgRegTable *reg, StgClosure *p)
766 Capability *cap = regTableToCapability(reg);
768 if (p->header.info == &stg_MUT_VAR_CLEAN_info) {
769 p->header.info = &stg_MUT_VAR_DIRTY_info;
770 bd = Bdescr((StgPtr)p);
771 if (bd->gen_no > 0) recordMutableCap(p,cap,bd->gen_no);
776 This is the write barrier for MVARs. An MVAR_CLEAN objects is not
777 on the mutable list; a MVAR_DIRTY is. When written to, a
778 MVAR_CLEAN turns into a MVAR_DIRTY and is put on the mutable list.
779 The check for MVAR_CLEAN is inlined at the call site for speed,
780 this really does make a difference on concurrency-heavy benchmarks
781 such as Chaneneos and cheap-concurrency.
784 dirty_MVAR(StgRegTable *reg, StgClosure *p)
786 Capability *cap = regTableToCapability(reg);
788 bd = Bdescr((StgPtr)p);
789 if (bd->gen_no > 0) recordMutableCap(p,cap,bd->gen_no);
792 /* -----------------------------------------------------------------------------
793 Allocation functions for GMP.
795 These all use the allocate() interface - we can't have any garbage
796 collection going on during a gmp operation, so we use allocate()
797 which always succeeds. The gmp operations which might need to
798 allocate will ask the storage manager (via doYouWantToGC()) whether
799 a garbage collection is required, in case we get into a loop doing
800 only allocate() style allocation.
801 -------------------------------------------------------------------------- */
804 stgAllocForGMP (size_t size_in_bytes)
807 nat data_size_in_words, total_size_in_words;
809 /* round up to a whole number of words */
810 data_size_in_words = (size_in_bytes + sizeof(W_) + 1) / sizeof(W_);
811 total_size_in_words = sizeofW(StgArrWords) + data_size_in_words;
813 /* allocate and fill it in. */
814 #if defined(THREADED_RTS)
815 arr = (StgArrWords *)allocateLocal(myTask()->cap, total_size_in_words);
817 arr = (StgArrWords *)allocateLocal(&MainCapability, total_size_in_words);
819 SET_ARR_HDR(arr, &stg_ARR_WORDS_info, CCCS, data_size_in_words);
821 /* and return a ptr to the goods inside the array */
826 stgReallocForGMP (void *ptr, size_t old_size, size_t new_size)
828 void *new_stuff_ptr = stgAllocForGMP(new_size);
830 char *p = (char *) ptr;
831 char *q = (char *) new_stuff_ptr;
833 for (; i < old_size; i++, p++, q++) {
837 return(new_stuff_ptr);
841 stgDeallocForGMP (void *ptr STG_UNUSED,
842 size_t size STG_UNUSED)
844 /* easy for us: the garbage collector does the dealloc'n */
847 /* -----------------------------------------------------------------------------
849 * -------------------------------------------------------------------------- */
851 /* -----------------------------------------------------------------------------
854 * Approximate how much we've allocated: number of blocks in the
855 * nursery + blocks allocated via allocate() - unused nusery blocks.
856 * This leaves a little slop at the end of each block, and doesn't
857 * take into account large objects (ToDo).
858 * -------------------------------------------------------------------------- */
861 calcAllocated( void )
866 allocated = allocatedBytes();
867 allocated += countNurseryBlocks() * BLOCK_SIZE_W;
872 for (i = 0; i < n_nurseries; i++) {
874 for ( bd = capabilities[i].r.rCurrentNursery->link;
875 bd != NULL; bd = bd->link ) {
876 allocated -= BLOCK_SIZE_W;
878 cap = &capabilities[i];
879 if (cap->r.rCurrentNursery->free <
880 cap->r.rCurrentNursery->start + BLOCK_SIZE_W) {
881 allocated -= (cap->r.rCurrentNursery->start + BLOCK_SIZE_W)
882 - cap->r.rCurrentNursery->free;
886 bdescr *current_nursery = MainCapability.r.rCurrentNursery;
888 for ( bd = current_nursery->link; bd != NULL; bd = bd->link ) {
889 allocated -= BLOCK_SIZE_W;
891 if (current_nursery->free < current_nursery->start + BLOCK_SIZE_W) {
892 allocated -= (current_nursery->start + BLOCK_SIZE_W)
893 - current_nursery->free;
898 total_allocated += allocated;
902 /* Approximate the amount of live data in the heap. To be called just
903 * after garbage collection (see GarbageCollect()).
912 if (RtsFlags.GcFlags.generations == 1) {
913 return (g0s0->n_large_blocks + g0s0->n_blocks) * BLOCK_SIZE_W;
916 for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
917 for (s = 0; s < generations[g].n_steps; s++) {
918 /* approximate amount of live data (doesn't take into account slop
919 * at end of each block).
921 if (g == 0 && s == 0) {
924 stp = &generations[g].steps[s];
925 live += (stp->n_large_blocks + stp->n_blocks) * BLOCK_SIZE_W;
931 /* Approximate the number of blocks that will be needed at the next
932 * garbage collection.
934 * Assume: all data currently live will remain live. Steps that will
935 * be collected next time will therefore need twice as many blocks
936 * since all the data will be copied.
945 for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
946 for (s = 0; s < generations[g].n_steps; s++) {
947 if (g == 0 && s == 0) { continue; }
948 stp = &generations[g].steps[s];
949 if (generations[g].steps[0].n_blocks +
950 generations[g].steps[0].n_large_blocks
951 > generations[g].max_blocks
952 && stp->is_compacted == 0) {
953 needed += 2 * stp->n_blocks;
955 needed += stp->n_blocks;
962 /* ----------------------------------------------------------------------------
965 Executable memory must be managed separately from non-executable
966 memory. Most OSs these days require you to jump through hoops to
967 dynamically allocate executable memory, due to various security
970 Here we provide a small memory allocator for executable memory.
971 Memory is managed with a page granularity; we allocate linearly
972 in the page, and when the page is emptied (all objects on the page
973 are free) we free the page again, not forgetting to make it
976 TODO: The inability to handle objects bigger than BLOCK_SIZE_W means that
977 the linker cannot use allocateExec for loading object code files
978 on Windows. Once allocateExec can handle larger objects, the linker
979 should be modified to use allocateExec instead of VirtualAlloc.
980 ------------------------------------------------------------------------- */
982 static bdescr *exec_block;
984 void *allocateExec (nat bytes)
991 // round up to words.
992 n = (bytes + sizeof(W_) + 1) / sizeof(W_);
994 if (n+1 > BLOCK_SIZE_W) {
995 barf("allocateExec: can't handle large objects");
998 if (exec_block == NULL ||
999 exec_block->free + n + 1 > exec_block->start + BLOCK_SIZE_W) {
1001 lnat pagesize = getPageSize();
1002 bd = allocGroup(stg_max(1, pagesize / BLOCK_SIZE));
1003 debugTrace(DEBUG_gc, "allocate exec block %p", bd->start);
1005 bd->flags = BF_EXEC;
1006 bd->link = exec_block;
1007 if (exec_block != NULL) {
1008 exec_block->u.back = bd;
1011 setExecutable(bd->start, bd->blocks * BLOCK_SIZE, rtsTrue);
1014 *(exec_block->free) = n; // store the size of this chunk
1015 exec_block->gen_no += n; // gen_no stores the number of words allocated
1016 ret = exec_block->free + 1;
1017 exec_block->free += n + 1;
1023 void freeExec (void *addr)
1025 StgPtr p = (StgPtr)addr - 1;
1026 bdescr *bd = Bdescr((StgPtr)p);
1028 if ((bd->flags & BF_EXEC) == 0) {
1029 barf("freeExec: not executable");
1032 if (*(StgPtr)p == 0) {
1033 barf("freeExec: already free?");
1038 bd->gen_no -= *(StgPtr)p;
1041 if (bd->gen_no == 0) {
1042 // Free the block if it is empty, but not if it is the block at
1043 // the head of the queue.
1044 if (bd != exec_block) {
1045 debugTrace(DEBUG_gc, "free exec block %p", bd->start);
1046 dbl_link_remove(bd, &exec_block);
1047 setExecutable(bd->start, bd->blocks * BLOCK_SIZE, rtsFalse);
1050 bd->free = bd->start;
1057 /* -----------------------------------------------------------------------------
1060 memInventory() checks for memory leaks by counting up all the
1061 blocks we know about and comparing that to the number of blocks
1062 allegedly floating around in the system.
1063 -------------------------------------------------------------------------- */
1068 countBlocks(bdescr *bd)
1071 for (n=0; bd != NULL; bd=bd->link) {
1077 // (*1) Just like countBlocks, except that we adjust the count for a
1078 // megablock group so that it doesn't include the extra few blocks
1079 // that would be taken up by block descriptors in the second and
1080 // subsequent megablock. This is so we can tally the count with the
1081 // number of blocks allocated in the system, for memInventory().
1083 countAllocdBlocks(bdescr *bd)
1086 for (n=0; bd != NULL; bd=bd->link) {
1088 // hack for megablock groups: see (*1) above
1089 if (bd->blocks > BLOCKS_PER_MBLOCK) {
1090 n -= (MBLOCK_SIZE / BLOCK_SIZE - BLOCKS_PER_MBLOCK)
1091 * (bd->blocks/(MBLOCK_SIZE/BLOCK_SIZE));
1098 stepBlocks (step *stp)
1100 ASSERT(countBlocks(stp->blocks) == stp->n_blocks);
1101 ASSERT(countBlocks(stp->large_objects) == stp->n_large_blocks);
1102 return stp->n_blocks + stp->n_old_blocks +
1103 countAllocdBlocks(stp->large_objects);
1111 lnat gen_blocks[RtsFlags.GcFlags.generations];
1112 lnat nursery_blocks, retainer_blocks,
1113 arena_blocks, exec_blocks;
1114 lnat live_blocks = 0, free_blocks = 0;
1116 // count the blocks we current have
1118 for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
1120 for (i = 0; i < n_capabilities; i++) {
1121 gen_blocks[g] += countBlocks(capabilities[i].mut_lists[g]);
1123 gen_blocks[g] += countAllocdBlocks(generations[g].mut_list);
1124 for (s = 0; s < generations[g].n_steps; s++) {
1125 stp = &generations[g].steps[s];
1126 gen_blocks[g] += stepBlocks(stp);
1131 for (i = 0; i < n_nurseries; i++) {
1132 nursery_blocks += stepBlocks(&nurseries[i]);
1135 retainer_blocks = 0;
1137 if (RtsFlags.ProfFlags.doHeapProfile == HEAP_BY_RETAINER) {
1138 retainer_blocks = retainerStackBlocks();
1142 // count the blocks allocated by the arena allocator
1143 arena_blocks = arenaBlocks();
1145 // count the blocks containing executable memory
1146 exec_blocks = countAllocdBlocks(exec_block);
1148 /* count the blocks on the free list */
1149 free_blocks = countFreeList();
1152 for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
1153 live_blocks += gen_blocks[g];
1155 live_blocks += nursery_blocks +
1156 + retainer_blocks + arena_blocks + exec_blocks;
1158 if (live_blocks + free_blocks != mblocks_allocated * BLOCKS_PER_MBLOCK)
1160 debugBelch("Memory leak detected\n");
1161 for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
1162 debugBelch(" gen %d blocks : %4lu\n", g, gen_blocks[g]);
1164 debugBelch(" nursery : %4lu\n", nursery_blocks);
1165 debugBelch(" retainer : %4lu\n", retainer_blocks);
1166 debugBelch(" arena blocks : %4lu\n", arena_blocks);
1167 debugBelch(" exec : %4lu\n", exec_blocks);
1168 debugBelch(" free : %4lu\n", free_blocks);
1169 debugBelch(" total : %4lu\n\n", live_blocks + free_blocks);
1170 debugBelch(" in system : %4lu\n", mblocks_allocated * BLOCKS_PER_MBLOCK);
1176 /* Full heap sanity check. */
1182 if (RtsFlags.GcFlags.generations == 1) {
1183 checkHeap(g0s0->blocks);
1184 checkChain(g0s0->large_objects);
1187 for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
1188 for (s = 0; s < generations[g].n_steps; s++) {
1189 if (g == 0 && s == 0) { continue; }
1190 ASSERT(countBlocks(generations[g].steps[s].blocks)
1191 == generations[g].steps[s].n_blocks);
1192 ASSERT(countBlocks(generations[g].steps[s].large_objects)
1193 == generations[g].steps[s].n_large_blocks);
1194 checkHeap(generations[g].steps[s].blocks);
1195 checkChain(generations[g].steps[s].large_objects);
1197 checkMutableList(generations[g].mut_list, g);
1202 for (s = 0; s < n_nurseries; s++) {
1203 ASSERT(countBlocks(nurseries[s].blocks)
1204 == nurseries[s].n_blocks);
1205 ASSERT(countBlocks(nurseries[s].large_objects)
1206 == nurseries[s].n_large_blocks);
1209 checkFreeListSanity();
1213 /* Nursery sanity check */
1215 checkNurserySanity( step *stp )
1221 for (bd = stp->blocks; bd != NULL; bd = bd->link) {
1222 ASSERT(bd->u.back == prev);
1224 blocks += bd->blocks;
1226 ASSERT(blocks == stp->n_blocks);
1229 // handy function for use in gdb, because Bdescr() is inlined.
1230 extern bdescr *_bdescr( StgPtr p );