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)
40 * All these globals require sm_mutex to access in THREADED_RTS mode.
42 StgClosure *caf_list = NULL;
43 StgClosure *revertible_caf_list = NULL;
46 bdescr *pinned_object_block; /* allocate pinned objects into this block */
47 nat alloc_blocks; /* number of allocate()d blocks since GC */
48 nat alloc_blocks_lim; /* approximate limit on alloc_blocks */
50 generation *generations = NULL; /* all the generations */
51 generation *g0 = NULL; /* generation 0, for convenience */
52 generation *oldest_gen = NULL; /* oldest generation, for convenience */
53 step *g0s0 = NULL; /* generation 0, step 0, for convenience */
55 ullong total_allocated = 0; /* total memory allocated during run */
57 nat n_nurseries = 0; /* == RtsFlags.ParFlags.nNodes, convenience */
58 step *nurseries = NULL; /* array of nurseries, >1 only if THREADED_RTS */
62 * Storage manager mutex: protects all the above state from
63 * simultaneous access by two STG threads.
67 * This mutex is used by atomicModifyMutVar# only
69 Mutex atomic_modify_mutvar_mutex;
76 static void *stgAllocForGMP (size_t size_in_bytes);
77 static void *stgReallocForGMP (void *ptr, size_t old_size, size_t new_size);
78 static void stgDeallocForGMP (void *ptr, size_t size);
81 initStep (step *stp, int g, int s)
86 stp->old_blocks = NULL;
87 stp->n_old_blocks = 0;
88 stp->gen = &generations[g];
90 stp->large_objects = NULL;
91 stp->n_large_blocks = 0;
92 stp->scavenged_large_objects = NULL;
93 stp->n_scavenged_large_blocks = 0;
94 stp->is_compacted = 0;
97 initSpinLock(&stp->sync_todo);
98 initSpinLock(&stp->sync_large_objects);
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 /* We use the NOT_NULL variant or gcc warns that the test is always true */
120 ASSERT(LOOKS_LIKE_INFO_PTR_NOT_NULL(&stg_BLACKHOLE_info));
121 ASSERT(LOOKS_LIKE_CLOSURE_PTR(&stg_dummy_ret_closure));
122 ASSERT(!HEAP_ALLOCED(&stg_dummy_ret_closure));
124 if (RtsFlags.GcFlags.maxHeapSize != 0 &&
125 RtsFlags.GcFlags.heapSizeSuggestion >
126 RtsFlags.GcFlags.maxHeapSize) {
127 RtsFlags.GcFlags.maxHeapSize = RtsFlags.GcFlags.heapSizeSuggestion;
130 if (RtsFlags.GcFlags.maxHeapSize != 0 &&
131 RtsFlags.GcFlags.minAllocAreaSize >
132 RtsFlags.GcFlags.maxHeapSize) {
133 errorBelch("maximum heap size (-M) is smaller than minimum alloc area size (-A)");
134 RtsFlags.GcFlags.minAllocAreaSize = RtsFlags.GcFlags.maxHeapSize;
137 initBlockAllocator();
139 #if defined(THREADED_RTS)
140 initMutex(&sm_mutex);
141 initMutex(&atomic_modify_mutvar_mutex);
146 /* allocate generation info array */
147 generations = (generation *)stgMallocBytes(RtsFlags.GcFlags.generations
148 * sizeof(struct generation_),
149 "initStorage: gens");
151 /* allocate all the steps into an array. It is important that we do
152 it this way, because we need the invariant that two step pointers
153 can be directly compared to see which is the oldest.
154 Remember that the last generation has only one step. */
155 step_arr = stgMallocBytes(sizeof(struct step_)
156 * (1 + ((RtsFlags.GcFlags.generations - 1)
157 * RtsFlags.GcFlags.steps)),
158 "initStorage: steps");
160 /* Initialise all generations */
161 for(g = 0; g < RtsFlags.GcFlags.generations; g++) {
162 gen = &generations[g];
164 gen->mut_list = allocBlock();
165 gen->collections = 0;
166 gen->failed_promotions = 0;
170 /* A couple of convenience pointers */
171 g0 = &generations[0];
172 oldest_gen = &generations[RtsFlags.GcFlags.generations-1];
174 /* Allocate step structures in each generation */
175 if (RtsFlags.GcFlags.generations > 1) {
176 /* Only for multiple-generations */
178 /* Oldest generation: one step */
179 oldest_gen->n_steps = 1;
180 oldest_gen->steps = step_arr + (RtsFlags.GcFlags.generations - 1)
181 * RtsFlags.GcFlags.steps;
183 /* set up all except the oldest generation with 2 steps */
184 for(g = 0; g < RtsFlags.GcFlags.generations-1; g++) {
185 generations[g].n_steps = RtsFlags.GcFlags.steps;
186 generations[g].steps = step_arr + g * RtsFlags.GcFlags.steps;
190 /* single generation, i.e. a two-space collector */
192 g0->steps = step_arr;
196 n_nurseries = n_capabilities;
200 nurseries = stgMallocBytes (n_nurseries * sizeof(struct step_),
201 "initStorage: nurseries");
203 /* Initialise all steps */
204 for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
205 for (s = 0; s < generations[g].n_steps; s++) {
206 initStep(&generations[g].steps[s], g, s);
210 for (s = 0; s < n_nurseries; s++) {
211 initStep(&nurseries[s], 0, s);
214 /* Set up the destination pointers in each younger gen. step */
215 for (g = 0; g < RtsFlags.GcFlags.generations-1; g++) {
216 for (s = 0; s < generations[g].n_steps-1; s++) {
217 generations[g].steps[s].to = &generations[g].steps[s+1];
219 generations[g].steps[s].to = &generations[g+1].steps[0];
221 oldest_gen->steps[0].to = &oldest_gen->steps[0];
223 for (s = 0; s < n_nurseries; s++) {
224 nurseries[s].to = generations[0].steps[0].to;
227 /* The oldest generation has one step. */
228 if (RtsFlags.GcFlags.compact) {
229 if (RtsFlags.GcFlags.generations == 1) {
230 errorBelch("WARNING: compaction is incompatible with -G1; disabled");
232 oldest_gen->steps[0].is_compacted = 1;
236 generations[0].max_blocks = 0;
237 g0s0 = &generations[0].steps[0];
239 /* The allocation area. Policy: keep the allocation area
240 * small to begin with, even if we have a large suggested heap
241 * size. Reason: we're going to do a major collection first, and we
242 * don't want it to be a big one. This vague idea is borne out by
243 * rigorous experimental evidence.
247 weak_ptr_list = NULL;
249 revertible_caf_list = NULL;
251 /* initialise the allocate() interface */
253 alloc_blocks_lim = RtsFlags.GcFlags.minAllocAreaSize;
255 /* Tell GNU multi-precision pkg about our custom alloc functions */
256 mp_set_memory_functions(stgAllocForGMP, stgReallocForGMP, stgDeallocForGMP);
259 initSpinLock(&gc_alloc_block_sync);
260 initSpinLock(&static_objects_sync);
261 initSpinLock(&recordMutableGen_sync);
265 IF_DEBUG(gc, statDescribeGens());
273 stat_exit(calcAllocated());
279 stgFree(g0s0); // frees all the steps
280 stgFree(generations);
282 #if defined(THREADED_RTS)
283 closeMutex(&sm_mutex);
284 closeMutex(&atomic_modify_mutvar_mutex);
289 /* -----------------------------------------------------------------------------
292 The entry code for every CAF does the following:
294 - builds a CAF_BLACKHOLE in the heap
295 - pushes an update frame pointing to the CAF_BLACKHOLE
296 - invokes UPD_CAF(), which:
297 - calls newCaf, below
298 - updates the CAF with a static indirection to the CAF_BLACKHOLE
300 Why do we build a BLACKHOLE in the heap rather than just updating
301 the thunk directly? It's so that we only need one kind of update
302 frame - otherwise we'd need a static version of the update frame too.
304 newCaf() does the following:
306 - it puts the CAF on the oldest generation's mut-once list.
307 This is so that we can treat the CAF as a root when collecting
310 For GHCI, we have additional requirements when dealing with CAFs:
312 - we must *retain* all dynamically-loaded CAFs ever entered,
313 just in case we need them again.
314 - we must be able to *revert* CAFs that have been evaluated, to
315 their pre-evaluated form.
317 To do this, we use an additional CAF list. When newCaf() is
318 called on a dynamically-loaded CAF, we add it to the CAF list
319 instead of the old-generation mutable list, and save away its
320 old info pointer (in caf->saved_info) for later reversion.
322 To revert all the CAFs, we traverse the CAF list and reset the
323 info pointer to caf->saved_info, then throw away the CAF list.
324 (see GC.c:revertCAFs()).
328 -------------------------------------------------------------------------- */
331 newCAF(StgClosure* caf)
338 // If we are in GHCi _and_ we are using dynamic libraries,
339 // then we can't redirect newCAF calls to newDynCAF (see below),
340 // so we make newCAF behave almost like newDynCAF.
341 // The dynamic libraries might be used by both the interpreted
342 // program and GHCi itself, so they must not be reverted.
343 // This also means that in GHCi with dynamic libraries, CAFs are not
344 // garbage collected. If this turns out to be a problem, we could
345 // do another hack here and do an address range test on caf to figure
346 // out whether it is from a dynamic library.
347 ((StgIndStatic *)caf)->saved_info = (StgInfoTable *)caf->header.info;
348 ((StgIndStatic *)caf)->static_link = caf_list;
353 /* Put this CAF on the mutable list for the old generation.
354 * This is a HACK - the IND_STATIC closure doesn't really have
355 * a mut_link field, but we pretend it has - in fact we re-use
356 * the STATIC_LINK field for the time being, because when we
357 * come to do a major GC we won't need the mut_link field
358 * any more and can use it as a STATIC_LINK.
360 ((StgIndStatic *)caf)->saved_info = NULL;
361 recordMutableGen(caf, oldest_gen);
367 // An alternate version of newCaf which is used for dynamically loaded
368 // object code in GHCi. In this case we want to retain *all* CAFs in
369 // the object code, because they might be demanded at any time from an
370 // expression evaluated on the command line.
371 // Also, GHCi might want to revert CAFs, so we add these to the
372 // revertible_caf_list.
374 // The linker hackily arranges that references to newCaf from dynamic
375 // code end up pointing to newDynCAF.
377 newDynCAF(StgClosure *caf)
381 ((StgIndStatic *)caf)->saved_info = (StgInfoTable *)caf->header.info;
382 ((StgIndStatic *)caf)->static_link = revertible_caf_list;
383 revertible_caf_list = caf;
388 /* -----------------------------------------------------------------------------
390 -------------------------------------------------------------------------- */
393 allocNursery (step *stp, bdescr *tail, nat blocks)
398 // Allocate a nursery: we allocate fresh blocks one at a time and
399 // cons them on to the front of the list, not forgetting to update
400 // the back pointer on the tail of the list to point to the new block.
401 for (i=0; i < blocks; i++) {
404 processNursery() in LdvProfile.c assumes that every block group in
405 the nursery contains only a single block. So, if a block group is
406 given multiple blocks, change processNursery() accordingly.
410 // double-link the nursery: we might need to insert blocks
417 bd->free = bd->start;
425 assignNurseriesToCapabilities (void)
430 for (i = 0; i < n_nurseries; i++) {
431 capabilities[i].r.rNursery = &nurseries[i];
432 capabilities[i].r.rCurrentNursery = nurseries[i].blocks;
433 capabilities[i].r.rCurrentAlloc = NULL;
435 #else /* THREADED_RTS */
436 MainCapability.r.rNursery = &nurseries[0];
437 MainCapability.r.rCurrentNursery = nurseries[0].blocks;
438 MainCapability.r.rCurrentAlloc = NULL;
443 allocNurseries( void )
447 for (i = 0; i < n_nurseries; i++) {
448 nurseries[i].blocks =
449 allocNursery(&nurseries[i], NULL,
450 RtsFlags.GcFlags.minAllocAreaSize);
451 nurseries[i].n_blocks = RtsFlags.GcFlags.minAllocAreaSize;
452 nurseries[i].old_blocks = NULL;
453 nurseries[i].n_old_blocks = 0;
455 assignNurseriesToCapabilities();
459 resetNurseries( void )
465 for (i = 0; i < n_nurseries; i++) {
467 for (bd = stp->blocks; bd; bd = bd->link) {
468 bd->free = bd->start;
469 ASSERT(bd->gen_no == 0);
470 ASSERT(bd->step == stp);
471 IF_DEBUG(sanity,memset(bd->start, 0xaa, BLOCK_SIZE));
474 assignNurseriesToCapabilities();
478 countNurseryBlocks (void)
483 for (i = 0; i < n_nurseries; i++) {
484 blocks += nurseries[i].n_blocks;
490 resizeNursery ( step *stp, nat blocks )
495 nursery_blocks = stp->n_blocks;
496 if (nursery_blocks == blocks) return;
498 if (nursery_blocks < blocks) {
499 debugTrace(DEBUG_gc, "increasing size of nursery to %d blocks",
501 stp->blocks = allocNursery(stp, stp->blocks, blocks-nursery_blocks);
506 debugTrace(DEBUG_gc, "decreasing size of nursery to %d blocks",
510 while (nursery_blocks > blocks) {
512 next_bd->u.back = NULL;
513 nursery_blocks -= bd->blocks; // might be a large block
518 // might have gone just under, by freeing a large block, so make
519 // up the difference.
520 if (nursery_blocks < blocks) {
521 stp->blocks = allocNursery(stp, stp->blocks, blocks-nursery_blocks);
525 stp->n_blocks = blocks;
526 ASSERT(countBlocks(stp->blocks) == stp->n_blocks);
530 // Resize each of the nurseries to the specified size.
533 resizeNurseriesFixed (nat blocks)
536 for (i = 0; i < n_nurseries; i++) {
537 resizeNursery(&nurseries[i], blocks);
542 // Resize the nurseries to the total specified size.
545 resizeNurseries (nat blocks)
547 // If there are multiple nurseries, then we just divide the number
548 // of available blocks between them.
549 resizeNurseriesFixed(blocks / n_nurseries);
552 /* -----------------------------------------------------------------------------
553 The allocate() interface
555 allocateInGen() function allocates memory directly into a specific
556 generation. It always succeeds, and returns a chunk of memory n
557 words long. n can be larger than the size of a block if necessary,
558 in which case a contiguous block group will be allocated.
560 allocate(n) is equivalent to allocateInGen(g0).
561 -------------------------------------------------------------------------- */
564 allocateInGen (generation *g, nat n)
572 TICK_ALLOC_HEAP_NOCTR(n);
577 if (n >= LARGE_OBJECT_THRESHOLD/sizeof(W_))
579 nat req_blocks = (lnat)BLOCK_ROUND_UP(n*sizeof(W_)) / BLOCK_SIZE;
581 // Attempting to allocate an object larger than maxHeapSize
582 // should definitely be disallowed. (bug #1791)
583 if (RtsFlags.GcFlags.maxHeapSize > 0 &&
584 req_blocks >= RtsFlags.GcFlags.maxHeapSize) {
588 bd = allocGroup(req_blocks);
589 dbl_link_onto(bd, &stp->large_objects);
590 stp->n_large_blocks += bd->blocks; // might be larger than req_blocks
593 bd->flags = BF_LARGE;
594 bd->free = bd->start + n;
599 // small allocation (<LARGE_OBJECT_THRESHOLD) */
601 if (bd == NULL || bd->free + n > bd->start + BLOCK_SIZE_W) {
606 bd->link = stp->blocks;
623 return allocateInGen(g0,n);
627 allocatedBytes( void )
631 allocated = alloc_blocks * BLOCK_SIZE_W;
632 if (pinned_object_block != NULL) {
633 allocated -= (pinned_object_block->start + BLOCK_SIZE_W) -
634 pinned_object_block->free;
640 // split N blocks off the start of the given bdescr, returning the
641 // remainder as a new block group. We treat the remainder as if it
642 // had been freshly allocated in generation 0.
644 splitLargeBlock (bdescr *bd, nat blocks)
648 // subtract the original number of blocks from the counter first
649 bd->step->n_large_blocks -= bd->blocks;
651 new_bd = splitBlockGroup (bd, blocks);
653 dbl_link_onto(new_bd, &g0s0->large_objects);
654 g0s0->n_large_blocks += new_bd->blocks;
655 new_bd->gen_no = g0s0->no;
657 new_bd->flags = BF_LARGE;
658 new_bd->free = bd->free;
660 // add the new number of blocks to the counter. Due to the gaps
661 // for block descriptor, new_bd->blocks + bd->blocks might not be
662 // equal to the original bd->blocks, which is why we do it this way.
663 bd->step->n_large_blocks += bd->blocks;
668 /* -----------------------------------------------------------------------------
671 This allocates memory in the current thread - it is intended for
672 use primarily from STG-land where we have a Capability. It is
673 better than allocate() because it doesn't require taking the
674 sm_mutex lock in the common case.
676 Memory is allocated directly from the nursery if possible (but not
677 from the current nursery block, so as not to interfere with
679 -------------------------------------------------------------------------- */
682 allocateLocal (Capability *cap, nat n)
687 if (n >= LARGE_OBJECT_THRESHOLD/sizeof(W_)) {
688 return allocateInGen(g0,n);
691 /* small allocation (<LARGE_OBJECT_THRESHOLD) */
693 TICK_ALLOC_HEAP_NOCTR(n);
696 bd = cap->r.rCurrentAlloc;
697 if (bd == NULL || bd->free + n > bd->start + BLOCK_SIZE_W) {
699 // The CurrentAlloc block is full, we need to find another
700 // one. First, we try taking the next block from the
702 bd = cap->r.rCurrentNursery->link;
704 if (bd == NULL || bd->free + n > bd->start + BLOCK_SIZE_W) {
705 // The nursery is empty, or the next block is already
706 // full: allocate a fresh block (we can't fail here).
709 cap->r.rNursery->n_blocks++;
712 bd->step = cap->r.rNursery;
714 // NO: alloc_blocks++;
715 // calcAllocated() uses the size of the nursery, and we've
716 // already bumpted nursery->n_blocks above.
718 // we have a block in the nursery: take it and put
719 // it at the *front* of the nursery list, and use it
720 // to allocate() from.
721 cap->r.rCurrentNursery->link = bd->link;
722 if (bd->link != NULL) {
723 bd->link->u.back = cap->r.rCurrentNursery;
726 dbl_link_onto(bd, &cap->r.rNursery->blocks);
727 cap->r.rCurrentAlloc = bd;
728 IF_DEBUG(sanity, checkNurserySanity(cap->r.rNursery));
735 /* ---------------------------------------------------------------------------
736 Allocate a fixed/pinned object.
738 We allocate small pinned objects into a single block, allocating a
739 new block when the current one overflows. The block is chained
740 onto the large_object_list of generation 0 step 0.
742 NOTE: The GC can't in general handle pinned objects. This
743 interface is only safe to use for ByteArrays, which have no
744 pointers and don't require scavenging. It works because the
745 block's descriptor has the BF_LARGE flag set, so the block is
746 treated as a large object and chained onto various lists, rather
747 than the individual objects being copied. However, when it comes
748 to scavenge the block, the GC will only scavenge the first object.
749 The reason is that the GC can't linearly scan a block of pinned
750 objects at the moment (doing so would require using the
751 mostly-copying techniques). But since we're restricting ourselves
752 to pinned ByteArrays, not scavenging is ok.
754 This function is called by newPinnedByteArray# which immediately
755 fills the allocated memory with a MutableByteArray#.
756 ------------------------------------------------------------------------- */
759 allocatePinned( nat n )
762 bdescr *bd = pinned_object_block;
764 // If the request is for a large object, then allocate()
765 // will give us a pinned object anyway.
766 if (n >= LARGE_OBJECT_THRESHOLD/sizeof(W_)) {
772 TICK_ALLOC_HEAP_NOCTR(n);
775 // we always return 8-byte aligned memory. bd->free must be
776 // 8-byte aligned to begin with, so we just round up n to
777 // the nearest multiple of 8 bytes.
778 if (sizeof(StgWord) == 4) {
782 // If we don't have a block of pinned objects yet, or the current
783 // one isn't large enough to hold the new object, allocate a new one.
784 if (bd == NULL || (bd->free + n) > (bd->start + BLOCK_SIZE_W)) {
785 pinned_object_block = bd = allocBlock();
786 dbl_link_onto(bd, &g0s0->large_objects);
787 g0s0->n_large_blocks++;
790 bd->flags = BF_PINNED | BF_LARGE;
791 bd->free = bd->start;
801 /* -----------------------------------------------------------------------------
803 -------------------------------------------------------------------------- */
806 This is the write barrier for MUT_VARs, a.k.a. IORefs. A
807 MUT_VAR_CLEAN object is not on the mutable list; a MUT_VAR_DIRTY
808 is. When written to, a MUT_VAR_CLEAN turns into a MUT_VAR_DIRTY
809 and is put on the mutable list.
812 dirty_MUT_VAR(StgRegTable *reg, StgClosure *p)
814 Capability *cap = regTableToCapability(reg);
816 if (p->header.info == &stg_MUT_VAR_CLEAN_info) {
817 p->header.info = &stg_MUT_VAR_DIRTY_info;
818 bd = Bdescr((StgPtr)p);
819 if (bd->gen_no > 0) recordMutableCap(p,cap,bd->gen_no);
824 This is the write barrier for MVARs. An MVAR_CLEAN objects is not
825 on the mutable list; a MVAR_DIRTY is. When written to, a
826 MVAR_CLEAN turns into a MVAR_DIRTY and is put on the mutable list.
827 The check for MVAR_CLEAN is inlined at the call site for speed,
828 this really does make a difference on concurrency-heavy benchmarks
829 such as Chaneneos and cheap-concurrency.
832 dirty_MVAR(StgRegTable *reg, StgClosure *p)
834 Capability *cap = regTableToCapability(reg);
836 bd = Bdescr((StgPtr)p);
837 if (bd->gen_no > 0) recordMutableCap(p,cap,bd->gen_no);
840 /* -----------------------------------------------------------------------------
841 Allocation functions for GMP.
843 These all use the allocate() interface - we can't have any garbage
844 collection going on during a gmp operation, so we use allocate()
845 which always succeeds. The gmp operations which might need to
846 allocate will ask the storage manager (via doYouWantToGC()) whether
847 a garbage collection is required, in case we get into a loop doing
848 only allocate() style allocation.
849 -------------------------------------------------------------------------- */
852 stgAllocForGMP (size_t size_in_bytes)
855 nat data_size_in_words, total_size_in_words;
857 /* round up to a whole number of words */
858 data_size_in_words = (size_in_bytes + sizeof(W_) + 1) / sizeof(W_);
859 total_size_in_words = sizeofW(StgArrWords) + data_size_in_words;
861 /* allocate and fill it in. */
862 #if defined(THREADED_RTS)
863 arr = (StgArrWords *)allocateLocal(myTask()->cap, total_size_in_words);
865 arr = (StgArrWords *)allocateLocal(&MainCapability, total_size_in_words);
867 SET_ARR_HDR(arr, &stg_ARR_WORDS_info, CCCS, data_size_in_words);
869 /* and return a ptr to the goods inside the array */
874 stgReallocForGMP (void *ptr, size_t old_size, size_t new_size)
876 void *new_stuff_ptr = stgAllocForGMP(new_size);
878 char *p = (char *) ptr;
879 char *q = (char *) new_stuff_ptr;
881 for (; i < old_size; i++, p++, q++) {
885 return(new_stuff_ptr);
889 stgDeallocForGMP (void *ptr STG_UNUSED,
890 size_t size STG_UNUSED)
892 /* easy for us: the garbage collector does the dealloc'n */
895 /* -----------------------------------------------------------------------------
897 * -------------------------------------------------------------------------- */
899 /* -----------------------------------------------------------------------------
902 * Approximate how much we've allocated: number of blocks in the
903 * nursery + blocks allocated via allocate() - unused nusery blocks.
904 * This leaves a little slop at the end of each block, and doesn't
905 * take into account large objects (ToDo).
906 * -------------------------------------------------------------------------- */
909 calcAllocated( void )
914 allocated = allocatedBytes();
915 allocated += countNurseryBlocks() * BLOCK_SIZE_W;
920 for (i = 0; i < n_nurseries; i++) {
922 for ( bd = capabilities[i].r.rCurrentNursery->link;
923 bd != NULL; bd = bd->link ) {
924 allocated -= BLOCK_SIZE_W;
926 cap = &capabilities[i];
927 if (cap->r.rCurrentNursery->free <
928 cap->r.rCurrentNursery->start + BLOCK_SIZE_W) {
929 allocated -= (cap->r.rCurrentNursery->start + BLOCK_SIZE_W)
930 - cap->r.rCurrentNursery->free;
934 bdescr *current_nursery = MainCapability.r.rCurrentNursery;
936 for ( bd = current_nursery->link; bd != NULL; bd = bd->link ) {
937 allocated -= BLOCK_SIZE_W;
939 if (current_nursery->free < current_nursery->start + BLOCK_SIZE_W) {
940 allocated -= (current_nursery->start + BLOCK_SIZE_W)
941 - current_nursery->free;
946 total_allocated += allocated;
950 /* Approximate the amount of live data in the heap. To be called just
951 * after garbage collection (see GarbageCollect()).
960 if (RtsFlags.GcFlags.generations == 1) {
961 return g0s0->n_large_blocks + g0s0->n_blocks;
964 for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
965 for (s = 0; s < generations[g].n_steps; s++) {
966 /* approximate amount of live data (doesn't take into account slop
967 * at end of each block).
969 if (g == 0 && s == 0) {
972 stp = &generations[g].steps[s];
973 live += stp->n_large_blocks + stp->n_blocks;
980 countOccupied(bdescr *bd)
985 for (; bd != NULL; bd = bd->link) {
986 words += bd->free - bd->start;
991 // Return an accurate count of the live data in the heap, excluding
1000 if (RtsFlags.GcFlags.generations == 1) {
1001 return countOccupied(g0s0->blocks) + countOccupied(g0s0->large_objects);
1005 for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
1006 for (s = 0; s < generations[g].n_steps; s++) {
1007 if (g == 0 && s == 0) continue;
1008 stp = &generations[g].steps[s];
1009 live += countOccupied(stp->blocks) +
1010 countOccupied(stp->large_objects);
1016 /* Approximate the number of blocks that will be needed at the next
1017 * garbage collection.
1019 * Assume: all data currently live will remain live. Steps that will
1020 * be collected next time will therefore need twice as many blocks
1021 * since all the data will be copied.
1030 for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
1031 for (s = 0; s < generations[g].n_steps; s++) {
1032 if (g == 0 && s == 0) { continue; }
1033 stp = &generations[g].steps[s];
1034 if (g == 0 || // always collect gen 0
1035 (generations[g].steps[0].n_blocks +
1036 generations[g].steps[0].n_large_blocks
1037 > generations[g].max_blocks
1038 && stp->is_compacted == 0)) {
1039 needed += 2 * stp->n_blocks + stp->n_large_blocks;
1041 needed += stp->n_blocks + stp->n_large_blocks;
1048 /* ----------------------------------------------------------------------------
1051 Executable memory must be managed separately from non-executable
1052 memory. Most OSs these days require you to jump through hoops to
1053 dynamically allocate executable memory, due to various security
1056 Here we provide a small memory allocator for executable memory.
1057 Memory is managed with a page granularity; we allocate linearly
1058 in the page, and when the page is emptied (all objects on the page
1059 are free) we free the page again, not forgetting to make it
1062 TODO: The inability to handle objects bigger than BLOCK_SIZE_W means that
1063 the linker cannot use allocateExec for loading object code files
1064 on Windows. Once allocateExec can handle larger objects, the linker
1065 should be modified to use allocateExec instead of VirtualAlloc.
1066 ------------------------------------------------------------------------- */
1068 static bdescr *exec_block;
1070 void *allocateExec (nat bytes)
1077 // round up to words.
1078 n = (bytes + sizeof(W_) + 1) / sizeof(W_);
1080 if (n+1 > BLOCK_SIZE_W) {
1081 barf("allocateExec: can't handle large objects");
1084 if (exec_block == NULL ||
1085 exec_block->free + n + 1 > exec_block->start + BLOCK_SIZE_W) {
1087 lnat pagesize = getPageSize();
1088 bd = allocGroup(stg_max(1, pagesize / BLOCK_SIZE));
1089 debugTrace(DEBUG_gc, "allocate exec block %p", bd->start);
1091 bd->flags = BF_EXEC;
1092 bd->link = exec_block;
1093 if (exec_block != NULL) {
1094 exec_block->u.back = bd;
1097 setExecutable(bd->start, bd->blocks * BLOCK_SIZE, rtsTrue);
1100 *(exec_block->free) = n; // store the size of this chunk
1101 exec_block->gen_no += n; // gen_no stores the number of words allocated
1102 ret = exec_block->free + 1;
1103 exec_block->free += n + 1;
1109 void freeExec (void *addr)
1111 StgPtr p = (StgPtr)addr - 1;
1112 bdescr *bd = Bdescr((StgPtr)p);
1114 if ((bd->flags & BF_EXEC) == 0) {
1115 barf("freeExec: not executable");
1118 if (*(StgPtr)p == 0) {
1119 barf("freeExec: already free?");
1124 bd->gen_no -= *(StgPtr)p;
1127 if (bd->gen_no == 0) {
1128 // Free the block if it is empty, but not if it is the block at
1129 // the head of the queue.
1130 if (bd != exec_block) {
1131 debugTrace(DEBUG_gc, "free exec block %p", bd->start);
1132 dbl_link_remove(bd, &exec_block);
1133 setExecutable(bd->start, bd->blocks * BLOCK_SIZE, rtsFalse);
1136 bd->free = bd->start;
1143 /* -----------------------------------------------------------------------------
1146 memInventory() checks for memory leaks by counting up all the
1147 blocks we know about and comparing that to the number of blocks
1148 allegedly floating around in the system.
1149 -------------------------------------------------------------------------- */
1153 // Useful for finding partially full blocks in gdb
1154 void findSlop(bdescr *bd);
1155 void findSlop(bdescr *bd)
1159 for (; bd != NULL; bd = bd->link) {
1160 slop = (bd->blocks * BLOCK_SIZE_W) - (bd->free - bd->start);
1161 if (slop > (1024/sizeof(W_))) {
1162 debugBelch("block at %p (bdescr %p) has %ldKB slop\n",
1163 bd->start, bd, slop / (1024/sizeof(W_)));
1169 countBlocks(bdescr *bd)
1172 for (n=0; bd != NULL; bd=bd->link) {
1178 // (*1) Just like countBlocks, except that we adjust the count for a
1179 // megablock group so that it doesn't include the extra few blocks
1180 // that would be taken up by block descriptors in the second and
1181 // subsequent megablock. This is so we can tally the count with the
1182 // number of blocks allocated in the system, for memInventory().
1184 countAllocdBlocks(bdescr *bd)
1187 for (n=0; bd != NULL; bd=bd->link) {
1189 // hack for megablock groups: see (*1) above
1190 if (bd->blocks > BLOCKS_PER_MBLOCK) {
1191 n -= (MBLOCK_SIZE / BLOCK_SIZE - BLOCKS_PER_MBLOCK)
1192 * (bd->blocks/(MBLOCK_SIZE/BLOCK_SIZE));
1199 stepBlocks (step *stp)
1201 ASSERT(countBlocks(stp->blocks) == stp->n_blocks);
1202 ASSERT(countBlocks(stp->large_objects) == stp->n_large_blocks);
1203 return stp->n_blocks + stp->n_old_blocks +
1204 countAllocdBlocks(stp->large_objects);
1208 memInventory (rtsBool show)
1212 lnat gen_blocks[RtsFlags.GcFlags.generations];
1213 lnat nursery_blocks, retainer_blocks,
1214 arena_blocks, exec_blocks;
1215 lnat live_blocks = 0, free_blocks = 0;
1218 // count the blocks we current have
1220 for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
1222 for (i = 0; i < n_capabilities; i++) {
1223 gen_blocks[g] += countBlocks(capabilities[i].mut_lists[g]);
1225 gen_blocks[g] += countAllocdBlocks(generations[g].mut_list);
1226 for (s = 0; s < generations[g].n_steps; s++) {
1227 stp = &generations[g].steps[s];
1228 gen_blocks[g] += stepBlocks(stp);
1233 for (i = 0; i < n_nurseries; i++) {
1234 nursery_blocks += stepBlocks(&nurseries[i]);
1237 retainer_blocks = 0;
1239 if (RtsFlags.ProfFlags.doHeapProfile == HEAP_BY_RETAINER) {
1240 retainer_blocks = retainerStackBlocks();
1244 // count the blocks allocated by the arena allocator
1245 arena_blocks = arenaBlocks();
1247 // count the blocks containing executable memory
1248 exec_blocks = countAllocdBlocks(exec_block);
1250 /* count the blocks on the free list */
1251 free_blocks = countFreeList();
1254 for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
1255 live_blocks += gen_blocks[g];
1257 live_blocks += nursery_blocks +
1258 + retainer_blocks + arena_blocks + exec_blocks;
1260 #define MB(n) (((n) * BLOCK_SIZE_W) / ((1024*1024)/sizeof(W_)))
1262 leak = live_blocks + free_blocks != mblocks_allocated * BLOCKS_PER_MBLOCK;
1266 debugBelch("Memory leak detected:\n");
1268 debugBelch("Memory inventory:\n");
1270 for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
1271 debugBelch(" gen %d blocks : %5lu blocks (%lu MB)\n", g,
1272 gen_blocks[g], MB(gen_blocks[g]));
1274 debugBelch(" nursery : %5lu blocks (%lu MB)\n",
1275 nursery_blocks, MB(nursery_blocks));
1276 debugBelch(" retainer : %5lu blocks (%lu MB)\n",
1277 retainer_blocks, MB(retainer_blocks));
1278 debugBelch(" arena blocks : %5lu blocks (%lu MB)\n",
1279 arena_blocks, MB(arena_blocks));
1280 debugBelch(" exec : %5lu blocks (%lu MB)\n",
1281 exec_blocks, MB(exec_blocks));
1282 debugBelch(" free : %5lu blocks (%lu MB)\n",
1283 free_blocks, MB(free_blocks));
1284 debugBelch(" total : %5lu blocks (%lu MB)\n",
1285 live_blocks + free_blocks, MB(live_blocks+free_blocks));
1287 debugBelch("\n in system : %5lu blocks (%lu MB)\n",
1288 mblocks_allocated * BLOCKS_PER_MBLOCK, mblocks_allocated);
1294 /* Full heap sanity check. */
1300 if (RtsFlags.GcFlags.generations == 1) {
1301 checkHeap(g0s0->blocks);
1302 checkChain(g0s0->large_objects);
1305 for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
1306 for (s = 0; s < generations[g].n_steps; s++) {
1307 if (g == 0 && s == 0) { continue; }
1308 ASSERT(countBlocks(generations[g].steps[s].blocks)
1309 == generations[g].steps[s].n_blocks);
1310 ASSERT(countBlocks(generations[g].steps[s].large_objects)
1311 == generations[g].steps[s].n_large_blocks);
1312 checkHeap(generations[g].steps[s].blocks);
1313 checkChain(generations[g].steps[s].large_objects);
1315 checkMutableList(generations[g].mut_list, g);
1320 for (s = 0; s < n_nurseries; s++) {
1321 ASSERT(countBlocks(nurseries[s].blocks)
1322 == nurseries[s].n_blocks);
1323 ASSERT(countBlocks(nurseries[s].large_objects)
1324 == nurseries[s].n_large_blocks);
1327 checkFreeListSanity();
1331 /* Nursery sanity check */
1333 checkNurserySanity( step *stp )
1339 for (bd = stp->blocks; bd != NULL; bd = bd->link) {
1340 ASSERT(bd->u.back == prev);
1342 blocks += bd->blocks;
1344 ASSERT(blocks == stp->n_blocks);
1347 // handy function for use in gdb, because Bdescr() is inlined.
1348 extern bdescr *_bdescr( StgPtr p );