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 */
56 step *all_steps = NULL; /* single array of steps */
58 ullong total_allocated = 0; /* total memory allocated during run */
60 nat n_nurseries = 0; /* == RtsFlags.ParFlags.nNodes, convenience */
61 step *nurseries = NULL; /* array of nurseries, >1 only if THREADED_RTS */
65 * Storage manager mutex: protects all the above state from
66 * simultaneous access by two STG threads.
70 * This mutex is used by atomicModifyMutVar# only
72 Mutex atomic_modify_mutvar_mutex;
79 static void *stgAllocForGMP (size_t size_in_bytes);
80 static void *stgReallocForGMP (void *ptr, size_t old_size, size_t new_size);
81 static void stgDeallocForGMP (void *ptr, size_t size);
84 initStep (step *stp, int g, int s)
87 stp->abs_no = RtsFlags.GcFlags.steps * g + s;
91 stp->old_blocks = NULL;
92 stp->n_old_blocks = 0;
93 stp->gen = &generations[g];
95 stp->large_objects = NULL;
96 stp->n_large_blocks = 0;
97 stp->scavenged_large_objects = NULL;
98 stp->n_scavenged_large_blocks = 0;
99 stp->is_compacted = 0;
102 initSpinLock(&stp->sync_todo);
103 initSpinLock(&stp->sync_large_objects);
113 if (generations != NULL) {
114 // multi-init protection
120 /* Sanity check to make sure the LOOKS_LIKE_ macros appear to be
121 * doing something reasonable.
123 /* We use the NOT_NULL variant or gcc warns that the test is always true */
124 ASSERT(LOOKS_LIKE_INFO_PTR_NOT_NULL(&stg_BLACKHOLE_info));
125 ASSERT(LOOKS_LIKE_CLOSURE_PTR(&stg_dummy_ret_closure));
126 ASSERT(!HEAP_ALLOCED(&stg_dummy_ret_closure));
128 if (RtsFlags.GcFlags.maxHeapSize != 0 &&
129 RtsFlags.GcFlags.heapSizeSuggestion >
130 RtsFlags.GcFlags.maxHeapSize) {
131 RtsFlags.GcFlags.maxHeapSize = RtsFlags.GcFlags.heapSizeSuggestion;
134 if (RtsFlags.GcFlags.maxHeapSize != 0 &&
135 RtsFlags.GcFlags.minAllocAreaSize >
136 RtsFlags.GcFlags.maxHeapSize) {
137 errorBelch("maximum heap size (-M) is smaller than minimum alloc area size (-A)");
138 RtsFlags.GcFlags.minAllocAreaSize = RtsFlags.GcFlags.maxHeapSize;
141 initBlockAllocator();
143 #if defined(THREADED_RTS)
144 initMutex(&sm_mutex);
145 initMutex(&atomic_modify_mutvar_mutex);
150 /* allocate generation info array */
151 generations = (generation *)stgMallocBytes(RtsFlags.GcFlags.generations
152 * sizeof(struct generation_),
153 "initStorage: gens");
155 /* allocate all the steps into an array. It is important that we do
156 it this way, because we need the invariant that two step pointers
157 can be directly compared to see which is the oldest.
158 Remember that the last generation has only one step. */
159 total_steps = 1 + (RtsFlags.GcFlags.generations - 1) * RtsFlags.GcFlags.steps;
160 all_steps = stgMallocBytes(total_steps * sizeof(struct step_),
161 "initStorage: steps");
163 /* Initialise all generations */
164 for(g = 0; g < RtsFlags.GcFlags.generations; g++) {
165 gen = &generations[g];
167 gen->mut_list = allocBlock();
168 gen->collections = 0;
169 gen->par_collections = 0;
170 gen->failed_promotions = 0;
174 /* A couple of convenience pointers */
175 g0 = &generations[0];
176 oldest_gen = &generations[RtsFlags.GcFlags.generations-1];
178 /* Allocate step structures in each generation */
179 if (RtsFlags.GcFlags.generations > 1) {
180 /* Only for multiple-generations */
182 /* Oldest generation: one step */
183 oldest_gen->n_steps = 1;
184 oldest_gen->steps = all_steps + (RtsFlags.GcFlags.generations - 1)
185 * RtsFlags.GcFlags.steps;
187 /* set up all except the oldest generation with 2 steps */
188 for(g = 0; g < RtsFlags.GcFlags.generations-1; g++) {
189 generations[g].n_steps = RtsFlags.GcFlags.steps;
190 generations[g].steps = all_steps + g * RtsFlags.GcFlags.steps;
194 /* single generation, i.e. a two-space collector */
196 g0->steps = all_steps;
200 n_nurseries = n_capabilities;
204 nurseries = stgMallocBytes (n_nurseries * sizeof(struct step_),
205 "initStorage: nurseries");
207 /* Initialise all steps */
208 for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
209 for (s = 0; s < generations[g].n_steps; s++) {
210 initStep(&generations[g].steps[s], g, s);
214 for (s = 0; s < n_nurseries; s++) {
215 initStep(&nurseries[s], 0, s);
218 /* Set up the destination pointers in each younger gen. step */
219 for (g = 0; g < RtsFlags.GcFlags.generations-1; g++) {
220 for (s = 0; s < generations[g].n_steps-1; s++) {
221 generations[g].steps[s].to = &generations[g].steps[s+1];
223 generations[g].steps[s].to = &generations[g+1].steps[0];
225 oldest_gen->steps[0].to = &oldest_gen->steps[0];
227 for (s = 0; s < n_nurseries; s++) {
228 nurseries[s].to = generations[0].steps[0].to;
231 /* The oldest generation has one step. */
232 if (RtsFlags.GcFlags.compact) {
233 if (RtsFlags.GcFlags.generations == 1) {
234 errorBelch("WARNING: compaction is incompatible with -G1; disabled");
236 oldest_gen->steps[0].is_compacted = 1;
240 generations[0].max_blocks = 0;
241 g0s0 = &generations[0].steps[0];
243 /* The allocation area. Policy: keep the allocation area
244 * small to begin with, even if we have a large suggested heap
245 * size. Reason: we're going to do a major collection first, and we
246 * don't want it to be a big one. This vague idea is borne out by
247 * rigorous experimental evidence.
251 weak_ptr_list = NULL;
253 revertible_caf_list = NULL;
255 /* initialise the allocate() interface */
257 alloc_blocks_lim = RtsFlags.GcFlags.minAllocAreaSize;
259 /* Tell GNU multi-precision pkg about our custom alloc functions */
260 mp_set_memory_functions(stgAllocForGMP, stgReallocForGMP, stgDeallocForGMP);
263 initSpinLock(&gc_alloc_block_sync);
264 initSpinLock(&recordMutableGen_sync);
268 IF_DEBUG(gc, statDescribeGens());
276 stat_exit(calcAllocated());
282 stgFree(g0s0); // frees all the steps
283 stgFree(generations);
285 #if defined(THREADED_RTS)
286 closeMutex(&sm_mutex);
287 closeMutex(&atomic_modify_mutvar_mutex);
292 /* -----------------------------------------------------------------------------
295 The entry code for every CAF does the following:
297 - builds a CAF_BLACKHOLE in the heap
298 - pushes an update frame pointing to the CAF_BLACKHOLE
299 - invokes UPD_CAF(), which:
300 - calls newCaf, below
301 - updates the CAF with a static indirection to the CAF_BLACKHOLE
303 Why do we build a BLACKHOLE in the heap rather than just updating
304 the thunk directly? It's so that we only need one kind of update
305 frame - otherwise we'd need a static version of the update frame too.
307 newCaf() does the following:
309 - it puts the CAF on the oldest generation's mut-once list.
310 This is so that we can treat the CAF as a root when collecting
313 For GHCI, we have additional requirements when dealing with CAFs:
315 - we must *retain* all dynamically-loaded CAFs ever entered,
316 just in case we need them again.
317 - we must be able to *revert* CAFs that have been evaluated, to
318 their pre-evaluated form.
320 To do this, we use an additional CAF list. When newCaf() is
321 called on a dynamically-loaded CAF, we add it to the CAF list
322 instead of the old-generation mutable list, and save away its
323 old info pointer (in caf->saved_info) for later reversion.
325 To revert all the CAFs, we traverse the CAF list and reset the
326 info pointer to caf->saved_info, then throw away the CAF list.
327 (see GC.c:revertCAFs()).
331 -------------------------------------------------------------------------- */
334 newCAF(StgClosure* caf)
341 // If we are in GHCi _and_ we are using dynamic libraries,
342 // then we can't redirect newCAF calls to newDynCAF (see below),
343 // so we make newCAF behave almost like newDynCAF.
344 // The dynamic libraries might be used by both the interpreted
345 // program and GHCi itself, so they must not be reverted.
346 // This also means that in GHCi with dynamic libraries, CAFs are not
347 // garbage collected. If this turns out to be a problem, we could
348 // do another hack here and do an address range test on caf to figure
349 // out whether it is from a dynamic library.
350 ((StgIndStatic *)caf)->saved_info = (StgInfoTable *)caf->header.info;
351 ((StgIndStatic *)caf)->static_link = caf_list;
356 /* Put this CAF on the mutable list for the old generation.
357 * This is a HACK - the IND_STATIC closure doesn't really have
358 * a mut_link field, but we pretend it has - in fact we re-use
359 * the STATIC_LINK field for the time being, because when we
360 * come to do a major GC we won't need the mut_link field
361 * any more and can use it as a STATIC_LINK.
363 ((StgIndStatic *)caf)->saved_info = NULL;
364 recordMutableGen(caf, oldest_gen);
370 // An alternate version of newCaf which is used for dynamically loaded
371 // object code in GHCi. In this case we want to retain *all* CAFs in
372 // the object code, because they might be demanded at any time from an
373 // expression evaluated on the command line.
374 // Also, GHCi might want to revert CAFs, so we add these to the
375 // revertible_caf_list.
377 // The linker hackily arranges that references to newCaf from dynamic
378 // code end up pointing to newDynCAF.
380 newDynCAF(StgClosure *caf)
384 ((StgIndStatic *)caf)->saved_info = (StgInfoTable *)caf->header.info;
385 ((StgIndStatic *)caf)->static_link = revertible_caf_list;
386 revertible_caf_list = caf;
391 /* -----------------------------------------------------------------------------
393 -------------------------------------------------------------------------- */
396 allocNursery (step *stp, bdescr *tail, nat blocks)
401 // Allocate a nursery: we allocate fresh blocks one at a time and
402 // cons them on to the front of the list, not forgetting to update
403 // the back pointer on the tail of the list to point to the new block.
404 for (i=0; i < blocks; i++) {
407 processNursery() in LdvProfile.c assumes that every block group in
408 the nursery contains only a single block. So, if a block group is
409 given multiple blocks, change processNursery() accordingly.
413 // double-link the nursery: we might need to insert blocks
420 bd->free = bd->start;
428 assignNurseriesToCapabilities (void)
433 for (i = 0; i < n_nurseries; i++) {
434 capabilities[i].r.rNursery = &nurseries[i];
435 capabilities[i].r.rCurrentNursery = nurseries[i].blocks;
436 capabilities[i].r.rCurrentAlloc = NULL;
438 #else /* THREADED_RTS */
439 MainCapability.r.rNursery = &nurseries[0];
440 MainCapability.r.rCurrentNursery = nurseries[0].blocks;
441 MainCapability.r.rCurrentAlloc = NULL;
446 allocNurseries( void )
450 for (i = 0; i < n_nurseries; i++) {
451 nurseries[i].blocks =
452 allocNursery(&nurseries[i], NULL,
453 RtsFlags.GcFlags.minAllocAreaSize);
454 nurseries[i].n_blocks = RtsFlags.GcFlags.minAllocAreaSize;
455 nurseries[i].old_blocks = NULL;
456 nurseries[i].n_old_blocks = 0;
458 assignNurseriesToCapabilities();
462 resetNurseries( void )
468 for (i = 0; i < n_nurseries; i++) {
470 for (bd = stp->blocks; bd; bd = bd->link) {
471 bd->free = bd->start;
472 ASSERT(bd->gen_no == 0);
473 ASSERT(bd->step == stp);
474 IF_DEBUG(sanity,memset(bd->start, 0xaa, BLOCK_SIZE));
477 assignNurseriesToCapabilities();
481 countNurseryBlocks (void)
486 for (i = 0; i < n_nurseries; i++) {
487 blocks += nurseries[i].n_blocks;
493 resizeNursery ( step *stp, nat blocks )
498 nursery_blocks = stp->n_blocks;
499 if (nursery_blocks == blocks) return;
501 if (nursery_blocks < blocks) {
502 debugTrace(DEBUG_gc, "increasing size of nursery to %d blocks",
504 stp->blocks = allocNursery(stp, stp->blocks, blocks-nursery_blocks);
509 debugTrace(DEBUG_gc, "decreasing size of nursery to %d blocks",
513 while (nursery_blocks > blocks) {
515 next_bd->u.back = NULL;
516 nursery_blocks -= bd->blocks; // might be a large block
521 // might have gone just under, by freeing a large block, so make
522 // up the difference.
523 if (nursery_blocks < blocks) {
524 stp->blocks = allocNursery(stp, stp->blocks, blocks-nursery_blocks);
528 stp->n_blocks = blocks;
529 ASSERT(countBlocks(stp->blocks) == stp->n_blocks);
533 // Resize each of the nurseries to the specified size.
536 resizeNurseriesFixed (nat blocks)
539 for (i = 0; i < n_nurseries; i++) {
540 resizeNursery(&nurseries[i], blocks);
545 // Resize the nurseries to the total specified size.
548 resizeNurseries (nat blocks)
550 // If there are multiple nurseries, then we just divide the number
551 // of available blocks between them.
552 resizeNurseriesFixed(blocks / n_nurseries);
555 /* -----------------------------------------------------------------------------
556 The allocate() interface
558 allocateInGen() function allocates memory directly into a specific
559 generation. It always succeeds, and returns a chunk of memory n
560 words long. n can be larger than the size of a block if necessary,
561 in which case a contiguous block group will be allocated.
563 allocate(n) is equivalent to allocateInGen(g0).
564 -------------------------------------------------------------------------- */
567 allocateInGen (generation *g, nat n)
575 TICK_ALLOC_HEAP_NOCTR(n);
580 if (n >= LARGE_OBJECT_THRESHOLD/sizeof(W_))
582 nat req_blocks = (lnat)BLOCK_ROUND_UP(n*sizeof(W_)) / BLOCK_SIZE;
584 // Attempting to allocate an object larger than maxHeapSize
585 // should definitely be disallowed. (bug #1791)
586 if (RtsFlags.GcFlags.maxHeapSize > 0 &&
587 req_blocks >= RtsFlags.GcFlags.maxHeapSize) {
591 bd = allocGroup(req_blocks);
592 dbl_link_onto(bd, &stp->large_objects);
593 stp->n_large_blocks += bd->blocks; // might be larger than req_blocks
596 bd->flags = BF_LARGE;
597 bd->free = bd->start + n;
602 // small allocation (<LARGE_OBJECT_THRESHOLD) */
604 if (bd == NULL || bd->free + n > bd->start + BLOCK_SIZE_W) {
609 bd->link = stp->blocks;
626 return allocateInGen(g0,n);
630 allocatedBytes( void )
634 allocated = alloc_blocks * BLOCK_SIZE_W;
635 if (pinned_object_block != NULL) {
636 allocated -= (pinned_object_block->start + BLOCK_SIZE_W) -
637 pinned_object_block->free;
643 // split N blocks off the start of the given bdescr, returning the
644 // remainder as a new block group. We treat the remainder as if it
645 // had been freshly allocated in generation 0.
647 splitLargeBlock (bdescr *bd, nat blocks)
651 // subtract the original number of blocks from the counter first
652 bd->step->n_large_blocks -= bd->blocks;
654 new_bd = splitBlockGroup (bd, blocks);
656 dbl_link_onto(new_bd, &g0s0->large_objects);
657 g0s0->n_large_blocks += new_bd->blocks;
658 new_bd->gen_no = g0s0->no;
660 new_bd->flags = BF_LARGE;
661 new_bd->free = bd->free;
663 // add the new number of blocks to the counter. Due to the gaps
664 // for block descriptor, new_bd->blocks + bd->blocks might not be
665 // equal to the original bd->blocks, which is why we do it this way.
666 bd->step->n_large_blocks += bd->blocks;
671 /* -----------------------------------------------------------------------------
674 This allocates memory in the current thread - it is intended for
675 use primarily from STG-land where we have a Capability. It is
676 better than allocate() because it doesn't require taking the
677 sm_mutex lock in the common case.
679 Memory is allocated directly from the nursery if possible (but not
680 from the current nursery block, so as not to interfere with
682 -------------------------------------------------------------------------- */
685 allocateLocal (Capability *cap, nat n)
690 if (n >= LARGE_OBJECT_THRESHOLD/sizeof(W_)) {
691 return allocateInGen(g0,n);
694 /* small allocation (<LARGE_OBJECT_THRESHOLD) */
696 TICK_ALLOC_HEAP_NOCTR(n);
699 bd = cap->r.rCurrentAlloc;
700 if (bd == NULL || bd->free + n > bd->start + BLOCK_SIZE_W) {
702 // The CurrentAlloc block is full, we need to find another
703 // one. First, we try taking the next block from the
705 bd = cap->r.rCurrentNursery->link;
707 if (bd == NULL || bd->free + n > bd->start + BLOCK_SIZE_W) {
708 // The nursery is empty, or the next block is already
709 // full: allocate a fresh block (we can't fail here).
712 cap->r.rNursery->n_blocks++;
715 bd->step = cap->r.rNursery;
717 // NO: alloc_blocks++;
718 // calcAllocated() uses the size of the nursery, and we've
719 // already bumpted nursery->n_blocks above.
721 // we have a block in the nursery: take it and put
722 // it at the *front* of the nursery list, and use it
723 // to allocate() from.
724 cap->r.rCurrentNursery->link = bd->link;
725 if (bd->link != NULL) {
726 bd->link->u.back = cap->r.rCurrentNursery;
729 dbl_link_onto(bd, &cap->r.rNursery->blocks);
730 cap->r.rCurrentAlloc = bd;
731 IF_DEBUG(sanity, checkNurserySanity(cap->r.rNursery));
738 /* ---------------------------------------------------------------------------
739 Allocate a fixed/pinned object.
741 We allocate small pinned objects into a single block, allocating a
742 new block when the current one overflows. The block is chained
743 onto the large_object_list of generation 0 step 0.
745 NOTE: The GC can't in general handle pinned objects. This
746 interface is only safe to use for ByteArrays, which have no
747 pointers and don't require scavenging. It works because the
748 block's descriptor has the BF_LARGE flag set, so the block is
749 treated as a large object and chained onto various lists, rather
750 than the individual objects being copied. However, when it comes
751 to scavenge the block, the GC will only scavenge the first object.
752 The reason is that the GC can't linearly scan a block of pinned
753 objects at the moment (doing so would require using the
754 mostly-copying techniques). But since we're restricting ourselves
755 to pinned ByteArrays, not scavenging is ok.
757 This function is called by newPinnedByteArray# which immediately
758 fills the allocated memory with a MutableByteArray#.
759 ------------------------------------------------------------------------- */
762 allocatePinned( nat n )
765 bdescr *bd = pinned_object_block;
767 // If the request is for a large object, then allocate()
768 // will give us a pinned object anyway.
769 if (n >= LARGE_OBJECT_THRESHOLD/sizeof(W_)) {
775 TICK_ALLOC_HEAP_NOCTR(n);
778 // we always return 8-byte aligned memory. bd->free must be
779 // 8-byte aligned to begin with, so we just round up n to
780 // the nearest multiple of 8 bytes.
781 if (sizeof(StgWord) == 4) {
785 // If we don't have a block of pinned objects yet, or the current
786 // one isn't large enough to hold the new object, allocate a new one.
787 if (bd == NULL || (bd->free + n) > (bd->start + BLOCK_SIZE_W)) {
788 pinned_object_block = bd = allocBlock();
789 dbl_link_onto(bd, &g0s0->large_objects);
790 g0s0->n_large_blocks++;
793 bd->flags = BF_PINNED | BF_LARGE;
794 bd->free = bd->start;
804 /* -----------------------------------------------------------------------------
806 -------------------------------------------------------------------------- */
809 This is the write barrier for MUT_VARs, a.k.a. IORefs. A
810 MUT_VAR_CLEAN object is not on the mutable list; a MUT_VAR_DIRTY
811 is. When written to, a MUT_VAR_CLEAN turns into a MUT_VAR_DIRTY
812 and is put on the mutable list.
815 dirty_MUT_VAR(StgRegTable *reg, StgClosure *p)
817 Capability *cap = regTableToCapability(reg);
819 if (p->header.info == &stg_MUT_VAR_CLEAN_info) {
820 p->header.info = &stg_MUT_VAR_DIRTY_info;
821 bd = Bdescr((StgPtr)p);
822 if (bd->gen_no > 0) recordMutableCap(p,cap,bd->gen_no);
827 This is the write barrier for MVARs. An MVAR_CLEAN objects is not
828 on the mutable list; a MVAR_DIRTY is. When written to, a
829 MVAR_CLEAN turns into a MVAR_DIRTY and is put on the mutable list.
830 The check for MVAR_CLEAN is inlined at the call site for speed,
831 this really does make a difference on concurrency-heavy benchmarks
832 such as Chaneneos and cheap-concurrency.
835 dirty_MVAR(StgRegTable *reg, StgClosure *p)
837 Capability *cap = regTableToCapability(reg);
839 bd = Bdescr((StgPtr)p);
840 if (bd->gen_no > 0) recordMutableCap(p,cap,bd->gen_no);
843 /* -----------------------------------------------------------------------------
844 Allocation functions for GMP.
846 These all use the allocate() interface - we can't have any garbage
847 collection going on during a gmp operation, so we use allocate()
848 which always succeeds. The gmp operations which might need to
849 allocate will ask the storage manager (via doYouWantToGC()) whether
850 a garbage collection is required, in case we get into a loop doing
851 only allocate() style allocation.
852 -------------------------------------------------------------------------- */
855 stgAllocForGMP (size_t size_in_bytes)
858 nat data_size_in_words, total_size_in_words;
860 /* round up to a whole number of words */
861 data_size_in_words = (size_in_bytes + sizeof(W_) + 1) / sizeof(W_);
862 total_size_in_words = sizeofW(StgArrWords) + data_size_in_words;
864 /* allocate and fill it in. */
865 #if defined(THREADED_RTS)
866 arr = (StgArrWords *)allocateLocal(myTask()->cap, total_size_in_words);
868 arr = (StgArrWords *)allocateLocal(&MainCapability, total_size_in_words);
870 SET_ARR_HDR(arr, &stg_ARR_WORDS_info, CCCS, data_size_in_words);
872 /* and return a ptr to the goods inside the array */
877 stgReallocForGMP (void *ptr, size_t old_size, size_t new_size)
879 void *new_stuff_ptr = stgAllocForGMP(new_size);
881 char *p = (char *) ptr;
882 char *q = (char *) new_stuff_ptr;
884 for (; i < old_size; i++, p++, q++) {
888 return(new_stuff_ptr);
892 stgDeallocForGMP (void *ptr STG_UNUSED,
893 size_t size STG_UNUSED)
895 /* easy for us: the garbage collector does the dealloc'n */
898 /* -----------------------------------------------------------------------------
900 * -------------------------------------------------------------------------- */
902 /* -----------------------------------------------------------------------------
905 * Approximate how much we've allocated: number of blocks in the
906 * nursery + blocks allocated via allocate() - unused nusery blocks.
907 * This leaves a little slop at the end of each block, and doesn't
908 * take into account large objects (ToDo).
909 * -------------------------------------------------------------------------- */
912 calcAllocated( void )
917 allocated = allocatedBytes();
918 allocated += countNurseryBlocks() * BLOCK_SIZE_W;
923 for (i = 0; i < n_nurseries; i++) {
925 for ( bd = capabilities[i].r.rCurrentNursery->link;
926 bd != NULL; bd = bd->link ) {
927 allocated -= BLOCK_SIZE_W;
929 cap = &capabilities[i];
930 if (cap->r.rCurrentNursery->free <
931 cap->r.rCurrentNursery->start + BLOCK_SIZE_W) {
932 allocated -= (cap->r.rCurrentNursery->start + BLOCK_SIZE_W)
933 - cap->r.rCurrentNursery->free;
937 bdescr *current_nursery = MainCapability.r.rCurrentNursery;
939 for ( bd = current_nursery->link; bd != NULL; bd = bd->link ) {
940 allocated -= BLOCK_SIZE_W;
942 if (current_nursery->free < current_nursery->start + BLOCK_SIZE_W) {
943 allocated -= (current_nursery->start + BLOCK_SIZE_W)
944 - current_nursery->free;
949 total_allocated += allocated;
953 /* Approximate the amount of live data in the heap. To be called just
954 * after garbage collection (see GarbageCollect()).
963 if (RtsFlags.GcFlags.generations == 1) {
964 return g0s0->n_large_blocks + g0s0->n_blocks;
967 for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
968 for (s = 0; s < generations[g].n_steps; s++) {
969 /* approximate amount of live data (doesn't take into account slop
970 * at end of each block).
972 if (g == 0 && s == 0) {
975 stp = &generations[g].steps[s];
976 live += stp->n_large_blocks + stp->n_blocks;
983 countOccupied(bdescr *bd)
988 for (; bd != NULL; bd = bd->link) {
989 words += bd->free - bd->start;
994 // Return an accurate count of the live data in the heap, excluding
1003 if (RtsFlags.GcFlags.generations == 1) {
1004 return g0s0->n_words + countOccupied(g0s0->large_objects);
1008 for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
1009 for (s = 0; s < generations[g].n_steps; s++) {
1010 if (g == 0 && s == 0) continue;
1011 stp = &generations[g].steps[s];
1012 live += stp->n_words + countOccupied(stp->large_objects);
1018 /* Approximate the number of blocks that will be needed at the next
1019 * garbage collection.
1021 * Assume: all data currently live will remain live. Steps that will
1022 * be collected next time will therefore need twice as many blocks
1023 * since all the data will be copied.
1032 for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
1033 for (s = 0; s < generations[g].n_steps; s++) {
1034 if (g == 0 && s == 0) { continue; }
1035 stp = &generations[g].steps[s];
1036 if (g == 0 || // always collect gen 0
1037 (generations[g].steps[0].n_blocks +
1038 generations[g].steps[0].n_large_blocks
1039 > generations[g].max_blocks
1040 && stp->is_compacted == 0)) {
1041 needed += 2 * stp->n_blocks + stp->n_large_blocks;
1043 needed += stp->n_blocks + stp->n_large_blocks;
1050 /* ----------------------------------------------------------------------------
1053 Executable memory must be managed separately from non-executable
1054 memory. Most OSs these days require you to jump through hoops to
1055 dynamically allocate executable memory, due to various security
1058 Here we provide a small memory allocator for executable memory.
1059 Memory is managed with a page granularity; we allocate linearly
1060 in the page, and when the page is emptied (all objects on the page
1061 are free) we free the page again, not forgetting to make it
1064 TODO: The inability to handle objects bigger than BLOCK_SIZE_W means that
1065 the linker cannot use allocateExec for loading object code files
1066 on Windows. Once allocateExec can handle larger objects, the linker
1067 should be modified to use allocateExec instead of VirtualAlloc.
1068 ------------------------------------------------------------------------- */
1070 static bdescr *exec_block;
1072 void *allocateExec (nat bytes)
1079 // round up to words.
1080 n = (bytes + sizeof(W_) + 1) / sizeof(W_);
1082 if (n+1 > BLOCK_SIZE_W) {
1083 barf("allocateExec: can't handle large objects");
1086 if (exec_block == NULL ||
1087 exec_block->free + n + 1 > exec_block->start + BLOCK_SIZE_W) {
1089 lnat pagesize = getPageSize();
1090 bd = allocGroup(stg_max(1, pagesize / BLOCK_SIZE));
1091 debugTrace(DEBUG_gc, "allocate exec block %p", bd->start);
1093 bd->flags = BF_EXEC;
1094 bd->link = exec_block;
1095 if (exec_block != NULL) {
1096 exec_block->u.back = bd;
1099 setExecutable(bd->start, bd->blocks * BLOCK_SIZE, rtsTrue);
1102 *(exec_block->free) = n; // store the size of this chunk
1103 exec_block->gen_no += n; // gen_no stores the number of words allocated
1104 ret = exec_block->free + 1;
1105 exec_block->free += n + 1;
1111 void freeExec (void *addr)
1113 StgPtr p = (StgPtr)addr - 1;
1114 bdescr *bd = Bdescr((StgPtr)p);
1116 if ((bd->flags & BF_EXEC) == 0) {
1117 barf("freeExec: not executable");
1120 if (*(StgPtr)p == 0) {
1121 barf("freeExec: already free?");
1126 bd->gen_no -= *(StgPtr)p;
1129 if (bd->gen_no == 0) {
1130 // Free the block if it is empty, but not if it is the block at
1131 // the head of the queue.
1132 if (bd != exec_block) {
1133 debugTrace(DEBUG_gc, "free exec block %p", bd->start);
1134 dbl_link_remove(bd, &exec_block);
1135 setExecutable(bd->start, bd->blocks * BLOCK_SIZE, rtsFalse);
1138 bd->free = bd->start;
1145 /* -----------------------------------------------------------------------------
1148 memInventory() checks for memory leaks by counting up all the
1149 blocks we know about and comparing that to the number of blocks
1150 allegedly floating around in the system.
1151 -------------------------------------------------------------------------- */
1155 // Useful for finding partially full blocks in gdb
1156 void findSlop(bdescr *bd);
1157 void findSlop(bdescr *bd)
1161 for (; bd != NULL; bd = bd->link) {
1162 slop = (bd->blocks * BLOCK_SIZE_W) - (bd->free - bd->start);
1163 if (slop > (1024/sizeof(W_))) {
1164 debugBelch("block at %p (bdescr %p) has %ldKB slop\n",
1165 bd->start, bd, slop / (1024/sizeof(W_)));
1171 countBlocks(bdescr *bd)
1174 for (n=0; bd != NULL; bd=bd->link) {
1180 // (*1) Just like countBlocks, except that we adjust the count for a
1181 // megablock group so that it doesn't include the extra few blocks
1182 // that would be taken up by block descriptors in the second and
1183 // subsequent megablock. This is so we can tally the count with the
1184 // number of blocks allocated in the system, for memInventory().
1186 countAllocdBlocks(bdescr *bd)
1189 for (n=0; bd != NULL; bd=bd->link) {
1191 // hack for megablock groups: see (*1) above
1192 if (bd->blocks > BLOCKS_PER_MBLOCK) {
1193 n -= (MBLOCK_SIZE / BLOCK_SIZE - BLOCKS_PER_MBLOCK)
1194 * (bd->blocks/(MBLOCK_SIZE/BLOCK_SIZE));
1201 stepBlocks (step *stp)
1203 ASSERT(countBlocks(stp->blocks) == stp->n_blocks);
1204 ASSERT(countBlocks(stp->large_objects) == stp->n_large_blocks);
1205 return stp->n_blocks + stp->n_old_blocks +
1206 countAllocdBlocks(stp->large_objects);
1210 memInventory (rtsBool show)
1214 lnat gen_blocks[RtsFlags.GcFlags.generations];
1215 lnat nursery_blocks, retainer_blocks,
1216 arena_blocks, exec_blocks;
1217 lnat live_blocks = 0, free_blocks = 0;
1220 // count the blocks we current have
1222 for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
1224 for (i = 0; i < n_capabilities; i++) {
1225 gen_blocks[g] += countBlocks(capabilities[i].mut_lists[g]);
1227 gen_blocks[g] += countAllocdBlocks(generations[g].mut_list);
1228 for (s = 0; s < generations[g].n_steps; s++) {
1229 stp = &generations[g].steps[s];
1230 gen_blocks[g] += stepBlocks(stp);
1235 for (i = 0; i < n_nurseries; i++) {
1236 nursery_blocks += stepBlocks(&nurseries[i]);
1239 retainer_blocks = 0;
1241 if (RtsFlags.ProfFlags.doHeapProfile == HEAP_BY_RETAINER) {
1242 retainer_blocks = retainerStackBlocks();
1246 // count the blocks allocated by the arena allocator
1247 arena_blocks = arenaBlocks();
1249 // count the blocks containing executable memory
1250 exec_blocks = countAllocdBlocks(exec_block);
1252 /* count the blocks on the free list */
1253 free_blocks = countFreeList();
1256 for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
1257 live_blocks += gen_blocks[g];
1259 live_blocks += nursery_blocks +
1260 + retainer_blocks + arena_blocks + exec_blocks;
1262 #define MB(n) (((n) * BLOCK_SIZE_W) / ((1024*1024)/sizeof(W_)))
1264 leak = live_blocks + free_blocks != mblocks_allocated * BLOCKS_PER_MBLOCK;
1268 debugBelch("Memory leak detected:\n");
1270 debugBelch("Memory inventory:\n");
1272 for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
1273 debugBelch(" gen %d blocks : %5lu blocks (%lu MB)\n", g,
1274 gen_blocks[g], MB(gen_blocks[g]));
1276 debugBelch(" nursery : %5lu blocks (%lu MB)\n",
1277 nursery_blocks, MB(nursery_blocks));
1278 debugBelch(" retainer : %5lu blocks (%lu MB)\n",
1279 retainer_blocks, MB(retainer_blocks));
1280 debugBelch(" arena blocks : %5lu blocks (%lu MB)\n",
1281 arena_blocks, MB(arena_blocks));
1282 debugBelch(" exec : %5lu blocks (%lu MB)\n",
1283 exec_blocks, MB(exec_blocks));
1284 debugBelch(" free : %5lu blocks (%lu MB)\n",
1285 free_blocks, MB(free_blocks));
1286 debugBelch(" total : %5lu blocks (%lu MB)\n",
1287 live_blocks + free_blocks, MB(live_blocks+free_blocks));
1289 debugBelch("\n in system : %5lu blocks (%lu MB)\n",
1290 mblocks_allocated * BLOCKS_PER_MBLOCK, mblocks_allocated);
1296 /* Full heap sanity check. */
1302 if (RtsFlags.GcFlags.generations == 1) {
1303 checkHeap(g0s0->blocks);
1304 checkChain(g0s0->large_objects);
1307 for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
1308 for (s = 0; s < generations[g].n_steps; s++) {
1309 if (g == 0 && s == 0) { continue; }
1310 ASSERT(countBlocks(generations[g].steps[s].blocks)
1311 == generations[g].steps[s].n_blocks);
1312 ASSERT(countBlocks(generations[g].steps[s].large_objects)
1313 == generations[g].steps[s].n_large_blocks);
1314 checkHeap(generations[g].steps[s].blocks);
1315 checkChain(generations[g].steps[s].large_objects);
1317 checkMutableList(generations[g].mut_list, g);
1322 for (s = 0; s < n_nurseries; s++) {
1323 ASSERT(countBlocks(nurseries[s].blocks)
1324 == nurseries[s].n_blocks);
1325 ASSERT(countBlocks(nurseries[s].large_objects)
1326 == nurseries[s].n_large_blocks);
1329 checkFreeListSanity();
1333 /* Nursery sanity check */
1335 checkNurserySanity( step *stp )
1341 for (bd = stp->blocks; bd != NULL; bd = bd->link) {
1342 ASSERT(bd->u.back == prev);
1344 blocks += bd->blocks;
1346 ASSERT(blocks == stp->n_blocks);
1349 // handy function for use in gdb, because Bdescr() is inlined.
1350 extern bdescr *_bdescr( StgPtr p );