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)
39 * All these globals require sm_mutex to access in THREADED_RTS mode.
41 StgClosure *caf_list = NULL;
42 StgClosure *revertible_caf_list = NULL;
45 bdescr *pinned_object_block; /* allocate pinned objects into this block */
46 nat alloc_blocks; /* number of allocate()d blocks since GC */
47 nat alloc_blocks_lim; /* approximate limit on alloc_blocks */
49 generation *generations = NULL; /* all the generations */
50 generation *g0 = NULL; /* generation 0, for convenience */
51 generation *oldest_gen = NULL; /* oldest generation, for convenience */
52 step *g0s0 = NULL; /* generation 0, step 0, for convenience */
54 ullong total_allocated = 0; /* total memory allocated during run */
56 nat n_nurseries = 0; /* == RtsFlags.ParFlags.nNodes, convenience */
57 step *nurseries = NULL; /* array of nurseries, >1 only if THREADED_RTS */
61 * Storage manager mutex: protects all the above state from
62 * simultaneous access by two STG threads.
66 * This mutex is used by atomicModifyMutVar# only
68 Mutex atomic_modify_mutvar_mutex;
75 static void *stgAllocForGMP (size_t size_in_bytes);
76 static void *stgReallocForGMP (void *ptr, size_t old_size, size_t new_size);
77 static void stgDeallocForGMP (void *ptr, size_t size);
80 initStep (step *stp, int g, int s)
85 stp->old_blocks = NULL;
86 stp->n_old_blocks = 0;
87 stp->gen = &generations[g];
89 stp->large_objects = NULL;
90 stp->n_large_blocks = 0;
91 stp->scavenged_large_objects = NULL;
92 stp->n_scavenged_large_blocks = 0;
93 stp->is_compacted = 0;
96 initSpinLock(&stp->sync_todo);
97 initSpinLock(&stp->sync_large_objects);
108 if (generations != NULL) {
109 // multi-init protection
115 /* Sanity check to make sure the LOOKS_LIKE_ macros appear to be
116 * doing something reasonable.
118 /* We use the NOT_NULL variant or gcc warns that the test is always true */
119 ASSERT(LOOKS_LIKE_INFO_PTR_NOT_NULL(&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 /* allocate all the steps into an array. It is important that we do
151 it this way, because we need the invariant that two step pointers
152 can be directly compared to see which is the oldest.
153 Remember that the last generation has only one step. */
154 step_arr = stgMallocBytes(sizeof(struct step_)
155 * (1 + ((RtsFlags.GcFlags.generations - 1)
156 * RtsFlags.GcFlags.steps)),
157 "initStorage: steps");
159 /* Initialise all generations */
160 for(g = 0; g < RtsFlags.GcFlags.generations; g++) {
161 gen = &generations[g];
163 gen->mut_list = allocBlock();
164 gen->collections = 0;
165 gen->failed_promotions = 0;
169 /* A couple of convenience pointers */
170 g0 = &generations[0];
171 oldest_gen = &generations[RtsFlags.GcFlags.generations-1];
173 /* Allocate step structures in each generation */
174 if (RtsFlags.GcFlags.generations > 1) {
175 /* Only for multiple-generations */
177 /* Oldest generation: one step */
178 oldest_gen->n_steps = 1;
179 oldest_gen->steps = step_arr + (RtsFlags.GcFlags.generations - 1)
180 * RtsFlags.GcFlags.steps;
182 /* set up all except the oldest generation with 2 steps */
183 for(g = 0; g < RtsFlags.GcFlags.generations-1; g++) {
184 generations[g].n_steps = RtsFlags.GcFlags.steps;
185 generations[g].steps = step_arr + g * RtsFlags.GcFlags.steps;
189 /* single generation, i.e. a two-space collector */
191 g0->steps = step_arr;
195 n_nurseries = n_capabilities;
199 nurseries = stgMallocBytes (n_nurseries * sizeof(struct step_),
200 "initStorage: nurseries");
202 /* Initialise all steps */
203 for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
204 for (s = 0; s < generations[g].n_steps; s++) {
205 initStep(&generations[g].steps[s], g, s);
209 for (s = 0; s < n_nurseries; s++) {
210 initStep(&nurseries[s], 0, s);
213 /* Set up the destination pointers in each younger gen. step */
214 for (g = 0; g < RtsFlags.GcFlags.generations-1; g++) {
215 for (s = 0; s < generations[g].n_steps-1; s++) {
216 generations[g].steps[s].to = &generations[g].steps[s+1];
218 generations[g].steps[s].to = &generations[g+1].steps[0];
220 oldest_gen->steps[0].to = &oldest_gen->steps[0];
222 for (s = 0; s < n_nurseries; s++) {
223 nurseries[s].to = generations[0].steps[0].to;
226 /* The oldest generation has one step. */
227 if (RtsFlags.GcFlags.compact) {
228 if (RtsFlags.GcFlags.generations == 1) {
229 errorBelch("WARNING: compaction is incompatible with -G1; disabled");
231 oldest_gen->steps[0].is_compacted = 1;
235 generations[0].max_blocks = 0;
236 g0s0 = &generations[0].steps[0];
238 /* The allocation area. Policy: keep the allocation area
239 * small to begin with, even if we have a large suggested heap
240 * size. Reason: we're going to do a major collection first, and we
241 * don't want it to be a big one. This vague idea is borne out by
242 * rigorous experimental evidence.
246 weak_ptr_list = NULL;
248 revertible_caf_list = NULL;
250 /* initialise the allocate() interface */
252 alloc_blocks_lim = RtsFlags.GcFlags.minAllocAreaSize;
254 /* Tell GNU multi-precision pkg about our custom alloc functions */
255 mp_set_memory_functions(stgAllocForGMP, stgReallocForGMP, stgDeallocForGMP);
258 initSpinLock(&gc_alloc_block_sync);
261 IF_DEBUG(gc, statDescribeGens());
269 stat_exit(calcAllocated());
275 stgFree(g0s0); // frees all the steps
276 stgFree(generations);
278 #if defined(THREADED_RTS)
279 closeMutex(&sm_mutex);
280 closeMutex(&atomic_modify_mutvar_mutex);
285 /* -----------------------------------------------------------------------------
288 The entry code for every CAF does the following:
290 - builds a CAF_BLACKHOLE in the heap
291 - pushes an update frame pointing to the CAF_BLACKHOLE
292 - invokes UPD_CAF(), which:
293 - calls newCaf, below
294 - updates the CAF with a static indirection to the CAF_BLACKHOLE
296 Why do we build a BLACKHOLE in the heap rather than just updating
297 the thunk directly? It's so that we only need one kind of update
298 frame - otherwise we'd need a static version of the update frame too.
300 newCaf() does the following:
302 - it puts the CAF on the oldest generation's mut-once list.
303 This is so that we can treat the CAF as a root when collecting
306 For GHCI, we have additional requirements when dealing with CAFs:
308 - we must *retain* all dynamically-loaded CAFs ever entered,
309 just in case we need them again.
310 - we must be able to *revert* CAFs that have been evaluated, to
311 their pre-evaluated form.
313 To do this, we use an additional CAF list. When newCaf() is
314 called on a dynamically-loaded CAF, we add it to the CAF list
315 instead of the old-generation mutable list, and save away its
316 old info pointer (in caf->saved_info) for later reversion.
318 To revert all the CAFs, we traverse the CAF list and reset the
319 info pointer to caf->saved_info, then throw away the CAF list.
320 (see GC.c:revertCAFs()).
324 -------------------------------------------------------------------------- */
327 newCAF(StgClosure* caf)
334 // If we are in GHCi _and_ we are using dynamic libraries,
335 // then we can't redirect newCAF calls to newDynCAF (see below),
336 // so we make newCAF behave almost like newDynCAF.
337 // The dynamic libraries might be used by both the interpreted
338 // program and GHCi itself, so they must not be reverted.
339 // This also means that in GHCi with dynamic libraries, CAFs are not
340 // garbage collected. If this turns out to be a problem, we could
341 // do another hack here and do an address range test on caf to figure
342 // out whether it is from a dynamic library.
343 ((StgIndStatic *)caf)->saved_info = (StgInfoTable *)caf->header.info;
344 ((StgIndStatic *)caf)->static_link = caf_list;
349 /* Put this CAF on the mutable list for the old generation.
350 * This is a HACK - the IND_STATIC closure doesn't really have
351 * a mut_link field, but we pretend it has - in fact we re-use
352 * the STATIC_LINK field for the time being, because when we
353 * come to do a major GC we won't need the mut_link field
354 * any more and can use it as a STATIC_LINK.
356 ((StgIndStatic *)caf)->saved_info = NULL;
357 recordMutableGen(caf, oldest_gen);
363 // An alternate version of newCaf which is used for dynamically loaded
364 // object code in GHCi. In this case we want to retain *all* CAFs in
365 // the object code, because they might be demanded at any time from an
366 // expression evaluated on the command line.
367 // Also, GHCi might want to revert CAFs, so we add these to the
368 // revertible_caf_list.
370 // The linker hackily arranges that references to newCaf from dynamic
371 // code end up pointing to newDynCAF.
373 newDynCAF(StgClosure *caf)
377 ((StgIndStatic *)caf)->saved_info = (StgInfoTable *)caf->header.info;
378 ((StgIndStatic *)caf)->static_link = revertible_caf_list;
379 revertible_caf_list = caf;
384 /* -----------------------------------------------------------------------------
386 -------------------------------------------------------------------------- */
389 allocNursery (step *stp, bdescr *tail, nat blocks)
394 // Allocate a nursery: we allocate fresh blocks one at a time and
395 // cons them on to the front of the list, not forgetting to update
396 // the back pointer on the tail of the list to point to the new block.
397 for (i=0; i < blocks; i++) {
400 processNursery() in LdvProfile.c assumes that every block group in
401 the nursery contains only a single block. So, if a block group is
402 given multiple blocks, change processNursery() accordingly.
406 // double-link the nursery: we might need to insert blocks
413 bd->free = bd->start;
421 assignNurseriesToCapabilities (void)
426 for (i = 0; i < n_nurseries; i++) {
427 capabilities[i].r.rNursery = &nurseries[i];
428 capabilities[i].r.rCurrentNursery = nurseries[i].blocks;
429 capabilities[i].r.rCurrentAlloc = NULL;
431 #else /* THREADED_RTS */
432 MainCapability.r.rNursery = &nurseries[0];
433 MainCapability.r.rCurrentNursery = nurseries[0].blocks;
434 MainCapability.r.rCurrentAlloc = NULL;
439 allocNurseries( void )
443 for (i = 0; i < n_nurseries; i++) {
444 nurseries[i].blocks =
445 allocNursery(&nurseries[i], NULL,
446 RtsFlags.GcFlags.minAllocAreaSize);
447 nurseries[i].n_blocks = RtsFlags.GcFlags.minAllocAreaSize;
448 nurseries[i].old_blocks = NULL;
449 nurseries[i].n_old_blocks = 0;
451 assignNurseriesToCapabilities();
455 resetNurseries( void )
461 for (i = 0; i < n_nurseries; i++) {
463 for (bd = stp->blocks; bd; bd = bd->link) {
464 bd->free = bd->start;
465 ASSERT(bd->gen_no == 0);
466 ASSERT(bd->step == stp);
467 IF_DEBUG(sanity,memset(bd->start, 0xaa, BLOCK_SIZE));
470 assignNurseriesToCapabilities();
474 countNurseryBlocks (void)
479 for (i = 0; i < n_nurseries; i++) {
480 blocks += nurseries[i].n_blocks;
486 resizeNursery ( step *stp, nat blocks )
491 nursery_blocks = stp->n_blocks;
492 if (nursery_blocks == blocks) return;
494 if (nursery_blocks < blocks) {
495 debugTrace(DEBUG_gc, "increasing size of nursery to %d blocks",
497 stp->blocks = allocNursery(stp, stp->blocks, blocks-nursery_blocks);
502 debugTrace(DEBUG_gc, "decreasing size of nursery to %d blocks",
506 while (nursery_blocks > blocks) {
508 next_bd->u.back = NULL;
509 nursery_blocks -= bd->blocks; // might be a large block
514 // might have gone just under, by freeing a large block, so make
515 // up the difference.
516 if (nursery_blocks < blocks) {
517 stp->blocks = allocNursery(stp, stp->blocks, blocks-nursery_blocks);
521 stp->n_blocks = blocks;
522 ASSERT(countBlocks(stp->blocks) == stp->n_blocks);
526 // Resize each of the nurseries to the specified size.
529 resizeNurseriesFixed (nat blocks)
532 for (i = 0; i < n_nurseries; i++) {
533 resizeNursery(&nurseries[i], blocks);
538 // Resize the nurseries to the total specified size.
541 resizeNurseries (nat blocks)
543 // If there are multiple nurseries, then we just divide the number
544 // of available blocks between them.
545 resizeNurseriesFixed(blocks / n_nurseries);
548 /* -----------------------------------------------------------------------------
549 The allocate() interface
551 allocateInGen() function allocates memory directly into a specific
552 generation. It always succeeds, and returns a chunk of memory n
553 words long. n can be larger than the size of a block if necessary,
554 in which case a contiguous block group will be allocated.
556 allocate(n) is equivalent to allocateInGen(g0).
557 -------------------------------------------------------------------------- */
560 allocateInGen (generation *g, nat n)
568 TICK_ALLOC_HEAP_NOCTR(n);
573 if (n >= LARGE_OBJECT_THRESHOLD/sizeof(W_))
575 nat req_blocks = (lnat)BLOCK_ROUND_UP(n*sizeof(W_)) / BLOCK_SIZE;
577 // Attempting to allocate an object larger than maxHeapSize
578 // should definitely be disallowed. (bug #1791)
579 if (RtsFlags.GcFlags.maxHeapSize > 0 &&
580 req_blocks >= RtsFlags.GcFlags.maxHeapSize) {
584 bd = allocGroup(req_blocks);
585 dbl_link_onto(bd, &stp->large_objects);
586 stp->n_large_blocks += bd->blocks; // might be larger than req_blocks
589 bd->flags = BF_LARGE;
590 bd->free = bd->start + n;
595 // small allocation (<LARGE_OBJECT_THRESHOLD) */
597 if (bd == NULL || bd->free + n > bd->start + BLOCK_SIZE_W) {
602 bd->link = stp->blocks;
619 return allocateInGen(g0,n);
623 allocatedBytes( void )
627 allocated = alloc_blocks * BLOCK_SIZE_W;
628 if (pinned_object_block != NULL) {
629 allocated -= (pinned_object_block->start + BLOCK_SIZE_W) -
630 pinned_object_block->free;
636 // split N blocks off the start of the given bdescr, returning the
637 // remainder as a new block group. We treat the remainder as if it
638 // had been freshly allocated in generation 0.
640 splitLargeBlock (bdescr *bd, nat blocks)
644 // subtract the original number of blocks from the counter first
645 bd->step->n_large_blocks -= bd->blocks;
647 new_bd = splitBlockGroup (bd, blocks);
649 dbl_link_onto(new_bd, &g0s0->large_objects);
650 g0s0->n_large_blocks += new_bd->blocks;
651 new_bd->gen_no = g0s0->no;
653 new_bd->flags = BF_LARGE;
654 new_bd->free = bd->free;
656 // add the new number of blocks to the counter. Due to the gaps
657 // for block descriptor, new_bd->blocks + bd->blocks might not be
658 // equal to the original bd->blocks, which is why we do it this way.
659 bd->step->n_large_blocks += bd->blocks;
664 /* -----------------------------------------------------------------------------
667 This allocates memory in the current thread - it is intended for
668 use primarily from STG-land where we have a Capability. It is
669 better than allocate() because it doesn't require taking the
670 sm_mutex lock in the common case.
672 Memory is allocated directly from the nursery if possible (but not
673 from the current nursery block, so as not to interfere with
675 -------------------------------------------------------------------------- */
678 allocateLocal (Capability *cap, nat n)
683 if (n >= LARGE_OBJECT_THRESHOLD/sizeof(W_)) {
684 return allocateInGen(g0,n);
687 /* small allocation (<LARGE_OBJECT_THRESHOLD) */
689 TICK_ALLOC_HEAP_NOCTR(n);
692 bd = cap->r.rCurrentAlloc;
693 if (bd == NULL || bd->free + n > bd->start + BLOCK_SIZE_W) {
695 // The CurrentAlloc block is full, we need to find another
696 // one. First, we try taking the next block from the
698 bd = cap->r.rCurrentNursery->link;
700 if (bd == NULL || bd->free + n > bd->start + BLOCK_SIZE_W) {
701 // The nursery is empty, or the next block is already
702 // full: allocate a fresh block (we can't fail here).
705 cap->r.rNursery->n_blocks++;
708 bd->step = cap->r.rNursery;
710 // NO: alloc_blocks++;
711 // calcAllocated() uses the size of the nursery, and we've
712 // already bumpted nursery->n_blocks above.
714 // we have a block in the nursery: take it and put
715 // it at the *front* of the nursery list, and use it
716 // to allocate() from.
717 cap->r.rCurrentNursery->link = bd->link;
718 if (bd->link != NULL) {
719 bd->link->u.back = cap->r.rCurrentNursery;
722 dbl_link_onto(bd, &cap->r.rNursery->blocks);
723 cap->r.rCurrentAlloc = bd;
724 IF_DEBUG(sanity, checkNurserySanity(cap->r.rNursery));
731 /* ---------------------------------------------------------------------------
732 Allocate a fixed/pinned object.
734 We allocate small pinned objects into a single block, allocating a
735 new block when the current one overflows. The block is chained
736 onto the large_object_list of generation 0 step 0.
738 NOTE: The GC can't in general handle pinned objects. This
739 interface is only safe to use for ByteArrays, which have no
740 pointers and don't require scavenging. It works because the
741 block's descriptor has the BF_LARGE flag set, so the block is
742 treated as a large object and chained onto various lists, rather
743 than the individual objects being copied. However, when it comes
744 to scavenge the block, the GC will only scavenge the first object.
745 The reason is that the GC can't linearly scan a block of pinned
746 objects at the moment (doing so would require using the
747 mostly-copying techniques). But since we're restricting ourselves
748 to pinned ByteArrays, not scavenging is ok.
750 This function is called by newPinnedByteArray# which immediately
751 fills the allocated memory with a MutableByteArray#.
752 ------------------------------------------------------------------------- */
755 allocatePinned( nat n )
758 bdescr *bd = pinned_object_block;
760 // If the request is for a large object, then allocate()
761 // will give us a pinned object anyway.
762 if (n >= LARGE_OBJECT_THRESHOLD/sizeof(W_)) {
768 TICK_ALLOC_HEAP_NOCTR(n);
771 // we always return 8-byte aligned memory. bd->free must be
772 // 8-byte aligned to begin with, so we just round up n to
773 // the nearest multiple of 8 bytes.
774 if (sizeof(StgWord) == 4) {
778 // If we don't have a block of pinned objects yet, or the current
779 // one isn't large enough to hold the new object, allocate a new one.
780 if (bd == NULL || (bd->free + n) > (bd->start + BLOCK_SIZE_W)) {
781 pinned_object_block = bd = allocBlock();
782 dbl_link_onto(bd, &g0s0->large_objects);
783 g0s0->n_large_blocks++;
786 bd->flags = BF_PINNED | BF_LARGE;
787 bd->free = bd->start;
797 /* -----------------------------------------------------------------------------
799 -------------------------------------------------------------------------- */
802 This is the write barrier for MUT_VARs, a.k.a. IORefs. A
803 MUT_VAR_CLEAN object is not on the mutable list; a MUT_VAR_DIRTY
804 is. When written to, a MUT_VAR_CLEAN turns into a MUT_VAR_DIRTY
805 and is put on the mutable list.
808 dirty_MUT_VAR(StgRegTable *reg, StgClosure *p)
810 Capability *cap = regTableToCapability(reg);
812 if (p->header.info == &stg_MUT_VAR_CLEAN_info) {
813 p->header.info = &stg_MUT_VAR_DIRTY_info;
814 bd = Bdescr((StgPtr)p);
815 if (bd->gen_no > 0) recordMutableCap(p,cap,bd->gen_no);
820 This is the write barrier for MVARs. An MVAR_CLEAN objects is not
821 on the mutable list; a MVAR_DIRTY is. When written to, a
822 MVAR_CLEAN turns into a MVAR_DIRTY and is put on the mutable list.
823 The check for MVAR_CLEAN is inlined at the call site for speed,
824 this really does make a difference on concurrency-heavy benchmarks
825 such as Chaneneos and cheap-concurrency.
828 dirty_MVAR(StgRegTable *reg, StgClosure *p)
830 Capability *cap = regTableToCapability(reg);
832 bd = Bdescr((StgPtr)p);
833 if (bd->gen_no > 0) recordMutableCap(p,cap,bd->gen_no);
836 /* -----------------------------------------------------------------------------
837 Allocation functions for GMP.
839 These all use the allocate() interface - we can't have any garbage
840 collection going on during a gmp operation, so we use allocate()
841 which always succeeds. The gmp operations which might need to
842 allocate will ask the storage manager (via doYouWantToGC()) whether
843 a garbage collection is required, in case we get into a loop doing
844 only allocate() style allocation.
845 -------------------------------------------------------------------------- */
848 stgAllocForGMP (size_t size_in_bytes)
851 nat data_size_in_words, total_size_in_words;
853 /* round up to a whole number of words */
854 data_size_in_words = (size_in_bytes + sizeof(W_) + 1) / sizeof(W_);
855 total_size_in_words = sizeofW(StgArrWords) + data_size_in_words;
857 /* allocate and fill it in. */
858 #if defined(THREADED_RTS)
859 arr = (StgArrWords *)allocateLocal(myTask()->cap, total_size_in_words);
861 arr = (StgArrWords *)allocateLocal(&MainCapability, total_size_in_words);
863 SET_ARR_HDR(arr, &stg_ARR_WORDS_info, CCCS, data_size_in_words);
865 /* and return a ptr to the goods inside the array */
870 stgReallocForGMP (void *ptr, size_t old_size, size_t new_size)
872 void *new_stuff_ptr = stgAllocForGMP(new_size);
874 char *p = (char *) ptr;
875 char *q = (char *) new_stuff_ptr;
877 for (; i < old_size; i++, p++, q++) {
881 return(new_stuff_ptr);
885 stgDeallocForGMP (void *ptr STG_UNUSED,
886 size_t size STG_UNUSED)
888 /* easy for us: the garbage collector does the dealloc'n */
891 /* -----------------------------------------------------------------------------
893 * -------------------------------------------------------------------------- */
895 /* -----------------------------------------------------------------------------
898 * Approximate how much we've allocated: number of blocks in the
899 * nursery + blocks allocated via allocate() - unused nusery blocks.
900 * This leaves a little slop at the end of each block, and doesn't
901 * take into account large objects (ToDo).
902 * -------------------------------------------------------------------------- */
905 calcAllocated( void )
910 allocated = allocatedBytes();
911 allocated += countNurseryBlocks() * BLOCK_SIZE_W;
916 for (i = 0; i < n_nurseries; i++) {
918 for ( bd = capabilities[i].r.rCurrentNursery->link;
919 bd != NULL; bd = bd->link ) {
920 allocated -= BLOCK_SIZE_W;
922 cap = &capabilities[i];
923 if (cap->r.rCurrentNursery->free <
924 cap->r.rCurrentNursery->start + BLOCK_SIZE_W) {
925 allocated -= (cap->r.rCurrentNursery->start + BLOCK_SIZE_W)
926 - cap->r.rCurrentNursery->free;
930 bdescr *current_nursery = MainCapability.r.rCurrentNursery;
932 for ( bd = current_nursery->link; bd != NULL; bd = bd->link ) {
933 allocated -= BLOCK_SIZE_W;
935 if (current_nursery->free < current_nursery->start + BLOCK_SIZE_W) {
936 allocated -= (current_nursery->start + BLOCK_SIZE_W)
937 - current_nursery->free;
942 total_allocated += allocated;
946 /* Approximate the amount of live data in the heap. To be called just
947 * after garbage collection (see GarbageCollect()).
956 if (RtsFlags.GcFlags.generations == 1) {
957 return g0s0->n_large_blocks + g0s0->n_blocks;
960 for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
961 for (s = 0; s < generations[g].n_steps; s++) {
962 /* approximate amount of live data (doesn't take into account slop
963 * at end of each block).
965 if (g == 0 && s == 0) {
968 stp = &generations[g].steps[s];
969 live += stp->n_large_blocks + stp->n_blocks;
976 countOccupied(bdescr *bd)
981 for (; bd != NULL; bd = bd->link) {
982 words += bd->free - bd->start;
987 // Return an accurate count of the live data in the heap, excluding
996 if (RtsFlags.GcFlags.generations == 1) {
997 return countOccupied(g0s0->blocks) + countOccupied(g0s0->large_objects);
1001 for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
1002 for (s = 0; s < generations[g].n_steps; s++) {
1003 if (g == 0 && s == 0) continue;
1004 stp = &generations[g].steps[s];
1005 live += countOccupied(stp->blocks) +
1006 countOccupied(stp->large_objects);
1012 /* Approximate the number of blocks that will be needed at the next
1013 * garbage collection.
1015 * Assume: all data currently live will remain live. Steps that will
1016 * be collected next time will therefore need twice as many blocks
1017 * since all the data will be copied.
1026 for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
1027 for (s = 0; s < generations[g].n_steps; s++) {
1028 if (g == 0 && s == 0) { continue; }
1029 stp = &generations[g].steps[s];
1030 if (g == 0 || // always collect gen 0
1031 (generations[g].steps[0].n_blocks +
1032 generations[g].steps[0].n_large_blocks
1033 > generations[g].max_blocks
1034 && stp->is_compacted == 0)) {
1035 needed += 2 * stp->n_blocks + stp->n_large_blocks;
1037 needed += stp->n_blocks + stp->n_large_blocks;
1044 /* ----------------------------------------------------------------------------
1047 Executable memory must be managed separately from non-executable
1048 memory. Most OSs these days require you to jump through hoops to
1049 dynamically allocate executable memory, due to various security
1052 Here we provide a small memory allocator for executable memory.
1053 Memory is managed with a page granularity; we allocate linearly
1054 in the page, and when the page is emptied (all objects on the page
1055 are free) we free the page again, not forgetting to make it
1058 TODO: The inability to handle objects bigger than BLOCK_SIZE_W means that
1059 the linker cannot use allocateExec for loading object code files
1060 on Windows. Once allocateExec can handle larger objects, the linker
1061 should be modified to use allocateExec instead of VirtualAlloc.
1062 ------------------------------------------------------------------------- */
1064 static bdescr *exec_block;
1066 void *allocateExec (nat bytes)
1073 // round up to words.
1074 n = (bytes + sizeof(W_) + 1) / sizeof(W_);
1076 if (n+1 > BLOCK_SIZE_W) {
1077 barf("allocateExec: can't handle large objects");
1080 if (exec_block == NULL ||
1081 exec_block->free + n + 1 > exec_block->start + BLOCK_SIZE_W) {
1083 lnat pagesize = getPageSize();
1084 bd = allocGroup(stg_max(1, pagesize / BLOCK_SIZE));
1085 debugTrace(DEBUG_gc, "allocate exec block %p", bd->start);
1087 bd->flags = BF_EXEC;
1088 bd->link = exec_block;
1089 if (exec_block != NULL) {
1090 exec_block->u.back = bd;
1093 setExecutable(bd->start, bd->blocks * BLOCK_SIZE, rtsTrue);
1096 *(exec_block->free) = n; // store the size of this chunk
1097 exec_block->gen_no += n; // gen_no stores the number of words allocated
1098 ret = exec_block->free + 1;
1099 exec_block->free += n + 1;
1105 void freeExec (void *addr)
1107 StgPtr p = (StgPtr)addr - 1;
1108 bdescr *bd = Bdescr((StgPtr)p);
1110 if ((bd->flags & BF_EXEC) == 0) {
1111 barf("freeExec: not executable");
1114 if (*(StgPtr)p == 0) {
1115 barf("freeExec: already free?");
1120 bd->gen_no -= *(StgPtr)p;
1123 if (bd->gen_no == 0) {
1124 // Free the block if it is empty, but not if it is the block at
1125 // the head of the queue.
1126 if (bd != exec_block) {
1127 debugTrace(DEBUG_gc, "free exec block %p", bd->start);
1128 dbl_link_remove(bd, &exec_block);
1129 setExecutable(bd->start, bd->blocks * BLOCK_SIZE, rtsFalse);
1132 bd->free = bd->start;
1139 /* -----------------------------------------------------------------------------
1142 memInventory() checks for memory leaks by counting up all the
1143 blocks we know about and comparing that to the number of blocks
1144 allegedly floating around in the system.
1145 -------------------------------------------------------------------------- */
1149 // Useful for finding partially full blocks in gdb
1150 void findSlop(bdescr *bd);
1151 void findSlop(bdescr *bd)
1155 for (; bd != NULL; bd = bd->link) {
1156 slop = (bd->blocks * BLOCK_SIZE_W) - (bd->free - bd->start);
1157 if (slop > (1024/sizeof(W_))) {
1158 debugBelch("block at %p (bdescr %p) has %ldKB slop\n",
1159 bd->start, bd, slop / (1024/sizeof(W_)));
1165 countBlocks(bdescr *bd)
1168 for (n=0; bd != NULL; bd=bd->link) {
1174 // (*1) Just like countBlocks, except that we adjust the count for a
1175 // megablock group so that it doesn't include the extra few blocks
1176 // that would be taken up by block descriptors in the second and
1177 // subsequent megablock. This is so we can tally the count with the
1178 // number of blocks allocated in the system, for memInventory().
1180 countAllocdBlocks(bdescr *bd)
1183 for (n=0; bd != NULL; bd=bd->link) {
1185 // hack for megablock groups: see (*1) above
1186 if (bd->blocks > BLOCKS_PER_MBLOCK) {
1187 n -= (MBLOCK_SIZE / BLOCK_SIZE - BLOCKS_PER_MBLOCK)
1188 * (bd->blocks/(MBLOCK_SIZE/BLOCK_SIZE));
1195 stepBlocks (step *stp)
1197 ASSERT(countBlocks(stp->blocks) == stp->n_blocks);
1198 ASSERT(countBlocks(stp->large_objects) == stp->n_large_blocks);
1199 return stp->n_blocks + stp->n_old_blocks +
1200 countAllocdBlocks(stp->large_objects);
1204 memInventory (rtsBool show)
1208 lnat gen_blocks[RtsFlags.GcFlags.generations];
1209 lnat nursery_blocks, retainer_blocks,
1210 arena_blocks, exec_blocks;
1211 lnat live_blocks = 0, free_blocks = 0;
1214 // count the blocks we current have
1216 for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
1218 for (i = 0; i < n_capabilities; i++) {
1219 gen_blocks[g] += countBlocks(capabilities[i].mut_lists[g]);
1221 gen_blocks[g] += countAllocdBlocks(generations[g].mut_list);
1222 for (s = 0; s < generations[g].n_steps; s++) {
1223 stp = &generations[g].steps[s];
1224 gen_blocks[g] += stepBlocks(stp);
1229 for (i = 0; i < n_nurseries; i++) {
1230 nursery_blocks += stepBlocks(&nurseries[i]);
1233 retainer_blocks = 0;
1235 if (RtsFlags.ProfFlags.doHeapProfile == HEAP_BY_RETAINER) {
1236 retainer_blocks = retainerStackBlocks();
1240 // count the blocks allocated by the arena allocator
1241 arena_blocks = arenaBlocks();
1243 // count the blocks containing executable memory
1244 exec_blocks = countAllocdBlocks(exec_block);
1246 /* count the blocks on the free list */
1247 free_blocks = countFreeList();
1250 for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
1251 live_blocks += gen_blocks[g];
1253 live_blocks += nursery_blocks +
1254 + retainer_blocks + arena_blocks + exec_blocks;
1256 #define MB(n) (((n) * BLOCK_SIZE_W) / ((1024*1024)/sizeof(W_)))
1258 leak = live_blocks + free_blocks != mblocks_allocated * BLOCKS_PER_MBLOCK;
1262 debugBelch("Memory leak detected:\n");
1264 debugBelch("Memory inventory:\n");
1266 for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
1267 debugBelch(" gen %d blocks : %5lu blocks (%lu MB)\n", g,
1268 gen_blocks[g], MB(gen_blocks[g]));
1270 debugBelch(" nursery : %5lu blocks (%lu MB)\n",
1271 nursery_blocks, MB(nursery_blocks));
1272 debugBelch(" retainer : %5lu blocks (%lu MB)\n",
1273 retainer_blocks, MB(retainer_blocks));
1274 debugBelch(" arena blocks : %5lu blocks (%lu MB)\n",
1275 arena_blocks, MB(arena_blocks));
1276 debugBelch(" exec : %5lu blocks (%lu MB)\n",
1277 exec_blocks, MB(exec_blocks));
1278 debugBelch(" free : %5lu blocks (%lu MB)\n",
1279 free_blocks, MB(free_blocks));
1280 debugBelch(" total : %5lu blocks (%lu MB)\n",
1281 live_blocks + free_blocks, MB(live_blocks+free_blocks));
1283 debugBelch("\n in system : %5lu blocks (%lu MB)\n",
1284 mblocks_allocated * BLOCKS_PER_MBLOCK, mblocks_allocated);
1290 /* Full heap sanity check. */
1296 if (RtsFlags.GcFlags.generations == 1) {
1297 checkHeap(g0s0->blocks);
1298 checkChain(g0s0->large_objects);
1301 for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
1302 for (s = 0; s < generations[g].n_steps; s++) {
1303 if (g == 0 && s == 0) { continue; }
1304 ASSERT(countBlocks(generations[g].steps[s].blocks)
1305 == generations[g].steps[s].n_blocks);
1306 ASSERT(countBlocks(generations[g].steps[s].large_objects)
1307 == generations[g].steps[s].n_large_blocks);
1308 checkHeap(generations[g].steps[s].blocks);
1309 checkChain(generations[g].steps[s].large_objects);
1311 checkMutableList(generations[g].mut_list, g);
1316 for (s = 0; s < n_nurseries; s++) {
1317 ASSERT(countBlocks(nurseries[s].blocks)
1318 == nurseries[s].n_blocks);
1319 ASSERT(countBlocks(nurseries[s].large_objects)
1320 == nurseries[s].n_large_blocks);
1323 checkFreeListSanity();
1327 /* Nursery sanity check */
1329 checkNurserySanity( step *stp )
1335 for (bd = stp->blocks; bd != NULL; bd = bd->link) {
1336 ASSERT(bd->u.back == prev);
1338 blocks += bd->blocks;
1340 ASSERT(blocks == stp->n_blocks);
1343 // handy function for use in gdb, because Bdescr() is inlined.
1344 extern bdescr *_bdescr( StgPtr p );