1 /* -----------------------------------------------------------------------------
3 * (c) The GHC Team, 1998-2008
5 * Storage manager front end
7 * Documentation on the architecture of the Storage Manager can be
8 * found in the online commentary:
10 * http://hackage.haskell.org/trac/ghc/wiki/Commentary/Rts/Storage
12 * ---------------------------------------------------------------------------*/
14 #include "PosixSource.h"
20 #include "BlockAlloc.h"
24 #include "Capability.h"
26 #include "RetainerProfile.h" // for counting memory blocks (memInventory)
37 * All these globals require sm_mutex to access in THREADED_RTS mode.
39 StgClosure *caf_list = NULL;
40 StgClosure *revertible_caf_list = NULL;
43 bdescr *pinned_object_block; /* allocate pinned objects into this block */
44 nat alloc_blocks; /* number of allocate()d blocks since GC */
45 nat alloc_blocks_lim; /* approximate limit on alloc_blocks */
47 static bdescr *exec_block;
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 */
55 step *all_steps = NULL; /* single array of steps */
57 ullong total_allocated = 0; /* total memory allocated during run */
59 nat n_nurseries = 0; /* == RtsFlags.ParFlags.nNodes, convenience */
60 step *nurseries = NULL; /* array of nurseries, >1 only if THREADED_RTS */
64 * Storage manager mutex: protects all the above state from
65 * simultaneous access by two STG threads.
70 static void allocNurseries ( void );
73 initStep (step *stp, int g, int s)
76 stp->abs_no = RtsFlags.GcFlags.steps * g + s;
80 stp->live_estimate = 0;
81 stp->old_blocks = NULL;
82 stp->n_old_blocks = 0;
83 stp->gen = &generations[g];
85 stp->large_objects = NULL;
86 stp->n_large_blocks = 0;
87 stp->scavenged_large_objects = NULL;
88 stp->n_scavenged_large_blocks = 0;
93 initSpinLock(&stp->sync_large_objects);
95 stp->threads = END_TSO_QUEUE;
96 stp->old_threads = END_TSO_QUEUE;
105 if (generations != NULL) {
106 // multi-init protection
112 /* Sanity check to make sure the LOOKS_LIKE_ macros appear to be
113 * doing something reasonable.
115 /* We use the NOT_NULL variant or gcc warns that the test is always true */
116 ASSERT(LOOKS_LIKE_INFO_PTR_NOT_NULL((StgWord)&stg_BLACKHOLE_info));
117 ASSERT(LOOKS_LIKE_CLOSURE_PTR(&stg_dummy_ret_closure));
118 ASSERT(!HEAP_ALLOCED(&stg_dummy_ret_closure));
120 if (RtsFlags.GcFlags.maxHeapSize != 0 &&
121 RtsFlags.GcFlags.heapSizeSuggestion >
122 RtsFlags.GcFlags.maxHeapSize) {
123 RtsFlags.GcFlags.maxHeapSize = RtsFlags.GcFlags.heapSizeSuggestion;
126 if (RtsFlags.GcFlags.maxHeapSize != 0 &&
127 RtsFlags.GcFlags.minAllocAreaSize >
128 RtsFlags.GcFlags.maxHeapSize) {
129 errorBelch("maximum heap size (-M) is smaller than minimum alloc area size (-A)");
130 RtsFlags.GcFlags.minAllocAreaSize = RtsFlags.GcFlags.maxHeapSize;
133 initBlockAllocator();
135 #if defined(THREADED_RTS)
136 initMutex(&sm_mutex);
141 /* allocate generation info array */
142 generations = (generation *)stgMallocBytes(RtsFlags.GcFlags.generations
143 * sizeof(struct generation_),
144 "initStorage: gens");
146 /* allocate all the steps into an array. It is important that we do
147 it this way, because we need the invariant that two step pointers
148 can be directly compared to see which is the oldest.
149 Remember that the last generation has only one step. */
150 total_steps = 1 + (RtsFlags.GcFlags.generations - 1) * RtsFlags.GcFlags.steps;
151 all_steps = stgMallocBytes(total_steps * sizeof(struct step_),
152 "initStorage: steps");
154 /* Initialise all generations */
155 for(g = 0; g < RtsFlags.GcFlags.generations; g++) {
156 gen = &generations[g];
158 gen->mut_list = allocBlock();
159 gen->collections = 0;
160 gen->par_collections = 0;
161 gen->failed_promotions = 0;
165 /* A couple of convenience pointers */
166 g0 = &generations[0];
167 oldest_gen = &generations[RtsFlags.GcFlags.generations-1];
169 /* Allocate step structures in each generation */
170 if (RtsFlags.GcFlags.generations > 1) {
171 /* Only for multiple-generations */
173 /* Oldest generation: one step */
174 oldest_gen->n_steps = 1;
175 oldest_gen->steps = all_steps + (RtsFlags.GcFlags.generations - 1)
176 * RtsFlags.GcFlags.steps;
178 /* set up all except the oldest generation with 2 steps */
179 for(g = 0; g < RtsFlags.GcFlags.generations-1; g++) {
180 generations[g].n_steps = RtsFlags.GcFlags.steps;
181 generations[g].steps = all_steps + g * RtsFlags.GcFlags.steps;
185 /* single generation, i.e. a two-space collector */
187 g0->steps = all_steps;
191 n_nurseries = n_capabilities;
195 nurseries = stgMallocBytes (n_nurseries * sizeof(struct step_),
196 "initStorage: nurseries");
198 /* Initialise all steps */
199 for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
200 for (s = 0; s < generations[g].n_steps; s++) {
201 initStep(&generations[g].steps[s], g, s);
205 for (s = 0; s < n_nurseries; s++) {
206 initStep(&nurseries[s], 0, s);
209 /* Set up the destination pointers in each younger gen. step */
210 for (g = 0; g < RtsFlags.GcFlags.generations-1; g++) {
211 for (s = 0; s < generations[g].n_steps-1; s++) {
212 generations[g].steps[s].to = &generations[g].steps[s+1];
214 generations[g].steps[s].to = &generations[g+1].steps[0];
216 oldest_gen->steps[0].to = &oldest_gen->steps[0];
218 for (s = 0; s < n_nurseries; s++) {
219 nurseries[s].to = generations[0].steps[0].to;
222 /* The oldest generation has one step. */
223 if (RtsFlags.GcFlags.compact || RtsFlags.GcFlags.sweep) {
224 if (RtsFlags.GcFlags.generations == 1) {
225 errorBelch("WARNING: compact/sweep is incompatible with -G1; disabled");
227 oldest_gen->steps[0].mark = 1;
228 if (RtsFlags.GcFlags.compact)
229 oldest_gen->steps[0].compact = 1;
233 generations[0].max_blocks = 0;
234 g0s0 = &generations[0].steps[0];
236 /* The allocation area. Policy: keep the allocation area
237 * small to begin with, even if we have a large suggested heap
238 * size. Reason: we're going to do a major collection first, and we
239 * don't want it to be a big one. This vague idea is borne out by
240 * rigorous experimental evidence.
244 weak_ptr_list = NULL;
246 revertible_caf_list = NULL;
248 /* initialise the allocate() interface */
250 alloc_blocks_lim = RtsFlags.GcFlags.minAllocAreaSize;
255 initSpinLock(&gc_alloc_block_sync);
263 IF_DEBUG(gc, statDescribeGens());
271 stat_exit(calcAllocated());
277 stgFree(g0s0); // frees all the steps
278 stgFree(generations);
280 #if defined(THREADED_RTS)
281 closeMutex(&sm_mutex);
287 /* -----------------------------------------------------------------------------
290 The entry code for every CAF does the following:
292 - builds a CAF_BLACKHOLE in the heap
293 - pushes an update frame pointing to the CAF_BLACKHOLE
294 - invokes UPD_CAF(), which:
295 - calls newCaf, below
296 - updates the CAF with a static indirection to the CAF_BLACKHOLE
298 Why do we build a BLACKHOLE in the heap rather than just updating
299 the thunk directly? It's so that we only need one kind of update
300 frame - otherwise we'd need a static version of the update frame too.
302 newCaf() does the following:
304 - it puts the CAF on the oldest generation's mut-once list.
305 This is so that we can treat the CAF as a root when collecting
308 For GHCI, we have additional requirements when dealing with CAFs:
310 - we must *retain* all dynamically-loaded CAFs ever entered,
311 just in case we need them again.
312 - we must be able to *revert* CAFs that have been evaluated, to
313 their pre-evaluated form.
315 To do this, we use an additional CAF list. When newCaf() is
316 called on a dynamically-loaded CAF, we add it to the CAF list
317 instead of the old-generation mutable list, and save away its
318 old info pointer (in caf->saved_info) for later reversion.
320 To revert all the CAFs, we traverse the CAF list and reset the
321 info pointer to caf->saved_info, then throw away the CAF list.
322 (see GC.c:revertCAFs()).
326 -------------------------------------------------------------------------- */
329 newCAF(StgClosure* caf)
336 // If we are in GHCi _and_ we are using dynamic libraries,
337 // then we can't redirect newCAF calls to newDynCAF (see below),
338 // so we make newCAF behave almost like newDynCAF.
339 // The dynamic libraries might be used by both the interpreted
340 // program and GHCi itself, so they must not be reverted.
341 // This also means that in GHCi with dynamic libraries, CAFs are not
342 // garbage collected. If this turns out to be a problem, we could
343 // do another hack here and do an address range test on caf to figure
344 // out whether it is from a dynamic library.
345 ((StgIndStatic *)caf)->saved_info = (StgInfoTable *)caf->header.info;
346 ((StgIndStatic *)caf)->static_link = caf_list;
351 /* Put this CAF on the mutable list for the old generation.
352 * This is a HACK - the IND_STATIC closure doesn't really have
353 * a mut_link field, but we pretend it has - in fact we re-use
354 * the STATIC_LINK field for the time being, because when we
355 * come to do a major GC we won't need the mut_link field
356 * any more and can use it as a STATIC_LINK.
358 ((StgIndStatic *)caf)->saved_info = NULL;
359 recordMutableGen(caf, oldest_gen->no);
365 // An alternate version of newCaf which is used for dynamically loaded
366 // object code in GHCi. In this case we want to retain *all* CAFs in
367 // the object code, because they might be demanded at any time from an
368 // expression evaluated on the command line.
369 // Also, GHCi might want to revert CAFs, so we add these to the
370 // revertible_caf_list.
372 // The linker hackily arranges that references to newCaf from dynamic
373 // code end up pointing to newDynCAF.
375 newDynCAF(StgClosure *caf)
379 ((StgIndStatic *)caf)->saved_info = (StgInfoTable *)caf->header.info;
380 ((StgIndStatic *)caf)->static_link = revertible_caf_list;
381 revertible_caf_list = caf;
386 /* -----------------------------------------------------------------------------
388 -------------------------------------------------------------------------- */
391 allocNursery (step *stp, bdescr *tail, nat blocks)
396 // Allocate a nursery: we allocate fresh blocks one at a time and
397 // cons them on to the front of the list, not forgetting to update
398 // the back pointer on the tail of the list to point to the new block.
399 for (i=0; i < blocks; i++) {
402 processNursery() in LdvProfile.c assumes that every block group in
403 the nursery contains only a single block. So, if a block group is
404 given multiple blocks, change processNursery() accordingly.
408 // double-link the nursery: we might need to insert blocks
415 bd->free = bd->start;
423 assignNurseriesToCapabilities (void)
428 for (i = 0; i < n_nurseries; i++) {
429 capabilities[i].r.rNursery = &nurseries[i];
430 capabilities[i].r.rCurrentNursery = nurseries[i].blocks;
431 capabilities[i].r.rCurrentAlloc = NULL;
433 #else /* THREADED_RTS */
434 MainCapability.r.rNursery = &nurseries[0];
435 MainCapability.r.rCurrentNursery = nurseries[0].blocks;
436 MainCapability.r.rCurrentAlloc = NULL;
441 allocNurseries( void )
445 for (i = 0; i < n_nurseries; i++) {
446 nurseries[i].blocks =
447 allocNursery(&nurseries[i], NULL,
448 RtsFlags.GcFlags.minAllocAreaSize);
449 nurseries[i].n_blocks = RtsFlags.GcFlags.minAllocAreaSize;
450 nurseries[i].old_blocks = NULL;
451 nurseries[i].n_old_blocks = 0;
453 assignNurseriesToCapabilities();
457 resetNurseries( void )
463 for (i = 0; i < n_nurseries; i++) {
465 for (bd = stp->blocks; bd; bd = bd->link) {
466 bd->free = bd->start;
467 ASSERT(bd->gen_no == 0);
468 ASSERT(bd->step == stp);
469 IF_DEBUG(sanity,memset(bd->start, 0xaa, BLOCK_SIZE));
472 assignNurseriesToCapabilities();
476 countNurseryBlocks (void)
481 for (i = 0; i < n_nurseries; i++) {
482 blocks += nurseries[i].n_blocks;
488 resizeNursery ( step *stp, nat blocks )
493 nursery_blocks = stp->n_blocks;
494 if (nursery_blocks == blocks) return;
496 if (nursery_blocks < blocks) {
497 debugTrace(DEBUG_gc, "increasing size of nursery to %d blocks",
499 stp->blocks = allocNursery(stp, stp->blocks, blocks-nursery_blocks);
504 debugTrace(DEBUG_gc, "decreasing size of nursery to %d blocks",
508 while (nursery_blocks > blocks) {
510 next_bd->u.back = NULL;
511 nursery_blocks -= bd->blocks; // might be a large block
516 // might have gone just under, by freeing a large block, so make
517 // up the difference.
518 if (nursery_blocks < blocks) {
519 stp->blocks = allocNursery(stp, stp->blocks, blocks-nursery_blocks);
523 stp->n_blocks = blocks;
524 ASSERT(countBlocks(stp->blocks) == stp->n_blocks);
528 // Resize each of the nurseries to the specified size.
531 resizeNurseriesFixed (nat blocks)
534 for (i = 0; i < n_nurseries; i++) {
535 resizeNursery(&nurseries[i], blocks);
540 // Resize the nurseries to the total specified size.
543 resizeNurseries (nat blocks)
545 // If there are multiple nurseries, then we just divide the number
546 // of available blocks between them.
547 resizeNurseriesFixed(blocks / n_nurseries);
551 /* -----------------------------------------------------------------------------
552 move_TSO is called to update the TSO structure after it has been
553 moved from one place to another.
554 -------------------------------------------------------------------------- */
557 move_TSO (StgTSO *src, StgTSO *dest)
561 // relocate the stack pointer...
562 diff = (StgPtr)dest - (StgPtr)src; // In *words*
563 dest->sp = (StgPtr)dest->sp + diff;
566 /* -----------------------------------------------------------------------------
567 The allocate() interface
569 allocateInGen() function allocates memory directly into a specific
570 generation. It always succeeds, and returns a chunk of memory n
571 words long. n can be larger than the size of a block if necessary,
572 in which case a contiguous block group will be allocated.
574 allocate(n) is equivalent to allocateInGen(g0).
575 -------------------------------------------------------------------------- */
578 allocateInGen (generation *g, lnat n)
586 TICK_ALLOC_HEAP_NOCTR(n);
591 if (n >= LARGE_OBJECT_THRESHOLD/sizeof(W_))
593 lnat req_blocks = (lnat)BLOCK_ROUND_UP(n*sizeof(W_)) / BLOCK_SIZE;
595 // Attempting to allocate an object larger than maxHeapSize
596 // should definitely be disallowed. (bug #1791)
597 if (RtsFlags.GcFlags.maxHeapSize > 0 &&
598 req_blocks >= RtsFlags.GcFlags.maxHeapSize) {
600 // heapOverflow() doesn't exit (see #2592), but we aren't
601 // in a position to do a clean shutdown here: we
602 // either have to allocate the memory or exit now.
603 // Allocating the memory would be bad, because the user
604 // has requested that we not exceed maxHeapSize, so we
606 stg_exit(EXIT_HEAPOVERFLOW);
609 bd = allocGroup(req_blocks);
610 dbl_link_onto(bd, &stp->large_objects);
611 stp->n_large_blocks += bd->blocks; // might be larger than req_blocks
612 alloc_blocks += bd->blocks;
615 bd->flags = BF_LARGE;
616 bd->free = bd->start + n;
621 // small allocation (<LARGE_OBJECT_THRESHOLD) */
623 if (bd == NULL || bd->free + n > bd->start + BLOCK_SIZE_W) {
628 bd->link = stp->blocks;
645 return allocateInGen(g0,n);
649 allocatedBytes( void )
653 allocated = alloc_blocks * BLOCK_SIZE_W;
654 if (pinned_object_block != NULL) {
655 allocated -= (pinned_object_block->start + BLOCK_SIZE_W) -
656 pinned_object_block->free;
662 // split N blocks off the front of the given bdescr, returning the
663 // new block group. We treat the remainder as if it
664 // had been freshly allocated in generation 0.
666 splitLargeBlock (bdescr *bd, nat blocks)
670 // subtract the original number of blocks from the counter first
671 bd->step->n_large_blocks -= bd->blocks;
673 new_bd = splitBlockGroup (bd, blocks);
675 dbl_link_onto(new_bd, &g0s0->large_objects);
676 g0s0->n_large_blocks += new_bd->blocks;
677 new_bd->gen_no = g0s0->no;
679 new_bd->flags = BF_LARGE;
680 new_bd->free = bd->free;
681 ASSERT(new_bd->free <= new_bd->start + new_bd->blocks * BLOCK_SIZE_W);
683 // add the new number of blocks to the counter. Due to the gaps
684 // for block descriptor, new_bd->blocks + bd->blocks might not be
685 // equal to the original bd->blocks, which is why we do it this way.
686 bd->step->n_large_blocks += bd->blocks;
691 /* -----------------------------------------------------------------------------
694 This allocates memory in the current thread - it is intended for
695 use primarily from STG-land where we have a Capability. It is
696 better than allocate() because it doesn't require taking the
697 sm_mutex lock in the common case.
699 Memory is allocated directly from the nursery if possible (but not
700 from the current nursery block, so as not to interfere with
702 -------------------------------------------------------------------------- */
705 allocateLocal (Capability *cap, lnat n)
710 if (n >= LARGE_OBJECT_THRESHOLD/sizeof(W_)) {
711 return allocateInGen(g0,n);
714 /* small allocation (<LARGE_OBJECT_THRESHOLD) */
716 TICK_ALLOC_HEAP_NOCTR(n);
719 bd = cap->r.rCurrentAlloc;
720 if (bd == NULL || bd->free + n > bd->start + BLOCK_SIZE_W) {
722 // The CurrentAlloc block is full, we need to find another
723 // one. First, we try taking the next block from the
725 bd = cap->r.rCurrentNursery->link;
727 if (bd == NULL || bd->free + n > bd->start + BLOCK_SIZE_W) {
728 // The nursery is empty, or the next block is already
729 // full: allocate a fresh block (we can't fail here).
732 cap->r.rNursery->n_blocks++;
735 bd->step = cap->r.rNursery;
737 // NO: alloc_blocks++;
738 // calcAllocated() uses the size of the nursery, and we've
739 // already bumpted nursery->n_blocks above. We'll GC
740 // pretty quickly now anyway, because MAYBE_GC() will
741 // notice that CurrentNursery->link is NULL.
743 // we have a block in the nursery: take it and put
744 // it at the *front* of the nursery list, and use it
745 // to allocate() from.
746 cap->r.rCurrentNursery->link = bd->link;
747 if (bd->link != NULL) {
748 bd->link->u.back = cap->r.rCurrentNursery;
751 dbl_link_onto(bd, &cap->r.rNursery->blocks);
752 cap->r.rCurrentAlloc = bd;
753 IF_DEBUG(sanity, checkNurserySanity(cap->r.rNursery));
760 /* ---------------------------------------------------------------------------
761 Allocate a fixed/pinned object.
763 We allocate small pinned objects into a single block, allocating a
764 new block when the current one overflows. The block is chained
765 onto the large_object_list of generation 0 step 0.
767 NOTE: The GC can't in general handle pinned objects. This
768 interface is only safe to use for ByteArrays, which have no
769 pointers and don't require scavenging. It works because the
770 block's descriptor has the BF_LARGE flag set, so the block is
771 treated as a large object and chained onto various lists, rather
772 than the individual objects being copied. However, when it comes
773 to scavenge the block, the GC will only scavenge the first object.
774 The reason is that the GC can't linearly scan a block of pinned
775 objects at the moment (doing so would require using the
776 mostly-copying techniques). But since we're restricting ourselves
777 to pinned ByteArrays, not scavenging is ok.
779 This function is called by newPinnedByteArray# which immediately
780 fills the allocated memory with a MutableByteArray#.
781 ------------------------------------------------------------------------- */
784 allocatePinned( lnat n )
787 bdescr *bd = pinned_object_block;
789 // If the request is for a large object, then allocate()
790 // will give us a pinned object anyway.
791 if (n >= LARGE_OBJECT_THRESHOLD/sizeof(W_)) {
793 Bdescr(p)->flags |= BF_PINNED;
799 TICK_ALLOC_HEAP_NOCTR(n);
802 // If we don't have a block of pinned objects yet, or the current
803 // one isn't large enough to hold the new object, allocate a new one.
804 if (bd == NULL || (bd->free + n) > (bd->start + BLOCK_SIZE_W)) {
805 pinned_object_block = bd = allocBlock();
806 dbl_link_onto(bd, &g0s0->large_objects);
807 g0s0->n_large_blocks++;
810 bd->flags = BF_PINNED | BF_LARGE;
811 bd->free = bd->start;
821 /* -----------------------------------------------------------------------------
823 -------------------------------------------------------------------------- */
826 This is the write barrier for MUT_VARs, a.k.a. IORefs. A
827 MUT_VAR_CLEAN object is not on the mutable list; a MUT_VAR_DIRTY
828 is. When written to, a MUT_VAR_CLEAN turns into a MUT_VAR_DIRTY
829 and is put on the mutable list.
832 dirty_MUT_VAR(StgRegTable *reg, StgClosure *p)
834 Capability *cap = regTableToCapability(reg);
836 if (p->header.info == &stg_MUT_VAR_CLEAN_info) {
837 p->header.info = &stg_MUT_VAR_DIRTY_info;
838 bd = Bdescr((StgPtr)p);
839 if (bd->gen_no > 0) recordMutableCap(p,cap,bd->gen_no);
843 // Setting a TSO's link field with a write barrier.
844 // It is *not* necessary to call this function when
845 // * setting the link field to END_TSO_QUEUE
846 // * putting a TSO on the blackhole_queue
847 // * setting the link field of the currently running TSO, as it
848 // will already be dirty.
850 setTSOLink (Capability *cap, StgTSO *tso, StgTSO *target)
853 if (tso->dirty == 0 && (tso->flags & TSO_LINK_DIRTY) == 0) {
854 tso->flags |= TSO_LINK_DIRTY;
855 bd = Bdescr((StgPtr)tso);
856 if (bd->gen_no > 0) recordMutableCap((StgClosure*)tso,cap,bd->gen_no);
862 dirty_TSO (Capability *cap, StgTSO *tso)
865 if (tso->dirty == 0 && (tso->flags & TSO_LINK_DIRTY) == 0) {
866 bd = Bdescr((StgPtr)tso);
867 if (bd->gen_no > 0) recordMutableCap((StgClosure*)tso,cap,bd->gen_no);
873 This is the write barrier for MVARs. An MVAR_CLEAN objects is not
874 on the mutable list; a MVAR_DIRTY is. When written to, a
875 MVAR_CLEAN turns into a MVAR_DIRTY and is put on the mutable list.
876 The check for MVAR_CLEAN is inlined at the call site for speed,
877 this really does make a difference on concurrency-heavy benchmarks
878 such as Chaneneos and cheap-concurrency.
881 dirty_MVAR(StgRegTable *reg, StgClosure *p)
883 Capability *cap = regTableToCapability(reg);
885 bd = Bdescr((StgPtr)p);
886 if (bd->gen_no > 0) recordMutableCap(p,cap,bd->gen_no);
889 /* -----------------------------------------------------------------------------
891 * -------------------------------------------------------------------------- */
893 /* -----------------------------------------------------------------------------
896 * Approximate how much we've allocated: number of blocks in the
897 * nursery + blocks allocated via allocate() - unused nusery blocks.
898 * This leaves a little slop at the end of each block, and doesn't
899 * take into account large objects (ToDo).
900 * -------------------------------------------------------------------------- */
903 calcAllocated( void )
908 allocated = allocatedBytes();
909 allocated += countNurseryBlocks() * BLOCK_SIZE_W;
914 for (i = 0; i < n_nurseries; i++) {
916 for ( bd = capabilities[i].r.rCurrentNursery->link;
917 bd != NULL; bd = bd->link ) {
918 allocated -= BLOCK_SIZE_W;
920 cap = &capabilities[i];
921 if (cap->r.rCurrentNursery->free <
922 cap->r.rCurrentNursery->start + BLOCK_SIZE_W) {
923 allocated -= (cap->r.rCurrentNursery->start + BLOCK_SIZE_W)
924 - cap->r.rCurrentNursery->free;
928 bdescr *current_nursery = MainCapability.r.rCurrentNursery;
930 for ( bd = current_nursery->link; bd != NULL; bd = bd->link ) {
931 allocated -= BLOCK_SIZE_W;
933 if (current_nursery->free < current_nursery->start + BLOCK_SIZE_W) {
934 allocated -= (current_nursery->start + BLOCK_SIZE_W)
935 - current_nursery->free;
940 total_allocated += allocated;
944 /* Approximate the amount of live data in the heap. To be called just
945 * after garbage collection (see GarbageCollect()).
954 if (RtsFlags.GcFlags.generations == 1) {
955 return g0s0->n_large_blocks + g0s0->n_blocks;
958 for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
959 for (s = 0; s < generations[g].n_steps; s++) {
960 /* approximate amount of live data (doesn't take into account slop
961 * at end of each block).
963 if (g == 0 && s == 0) {
966 stp = &generations[g].steps[s];
967 live += stp->n_large_blocks + stp->n_blocks;
974 countOccupied(bdescr *bd)
979 for (; bd != NULL; bd = bd->link) {
980 ASSERT(bd->free <= bd->start + bd->blocks * BLOCK_SIZE_W);
981 words += bd->free - bd->start;
986 // Return an accurate count of the live data in the heap, excluding
995 if (RtsFlags.GcFlags.generations == 1) {
996 return g0s0->n_words + countOccupied(g0s0->large_objects);
1000 for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
1001 for (s = 0; s < generations[g].n_steps; s++) {
1002 if (g == 0 && s == 0) continue;
1003 stp = &generations[g].steps[s];
1004 live += stp->n_words + countOccupied(stp->large_objects);
1010 /* Approximate the number of blocks that will be needed at the next
1011 * garbage collection.
1013 * Assume: all data currently live will remain live. Steps that will
1014 * be collected next time will therefore need twice as many blocks
1015 * since all the data will be copied.
1024 for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
1025 for (s = 0; s < generations[g].n_steps; s++) {
1026 if (g == 0 && s == 0) { continue; }
1027 stp = &generations[g].steps[s];
1029 // we need at least this much space
1030 needed += stp->n_blocks + stp->n_large_blocks;
1032 // any additional space needed to collect this gen next time?
1033 if (g == 0 || // always collect gen 0
1034 (generations[g].steps[0].n_blocks +
1035 generations[g].steps[0].n_large_blocks
1036 > generations[g].max_blocks)) {
1037 // we will collect this gen next time
1040 needed += stp->n_blocks / BITS_IN(W_);
1042 needed += stp->n_blocks / 100;
1045 continue; // no additional space needed for compaction
1047 needed += stp->n_blocks;
1055 /* ----------------------------------------------------------------------------
1058 Executable memory must be managed separately from non-executable
1059 memory. Most OSs these days require you to jump through hoops to
1060 dynamically allocate executable memory, due to various security
1063 Here we provide a small memory allocator for executable memory.
1064 Memory is managed with a page granularity; we allocate linearly
1065 in the page, and when the page is emptied (all objects on the page
1066 are free) we free the page again, not forgetting to make it
1069 TODO: The inability to handle objects bigger than BLOCK_SIZE_W means that
1070 the linker cannot use allocateExec for loading object code files
1071 on Windows. Once allocateExec can handle larger objects, the linker
1072 should be modified to use allocateExec instead of VirtualAlloc.
1073 ------------------------------------------------------------------------- */
1075 #if defined(linux_HOST_OS)
1077 // On Linux we need to use libffi for allocating executable memory,
1078 // because it knows how to work around the restrictions put in place
1081 void *allocateExec (nat bytes, void **exec_ret)
1085 ret = ffi_closure_alloc (sizeof(void *) + (size_t)bytes, (void**)&exec);
1087 if (ret == NULL) return ret;
1088 *ret = ret; // save the address of the writable mapping, for freeExec().
1089 *exec_ret = exec + 1;
1093 // freeExec gets passed the executable address, not the writable address.
1094 void freeExec (void *addr)
1097 writable = *((void**)addr - 1);
1099 ffi_closure_free (writable);
1105 void *allocateExec (nat bytes, void **exec_ret)
1112 // round up to words.
1113 n = (bytes + sizeof(W_) + 1) / sizeof(W_);
1115 if (n+1 > BLOCK_SIZE_W) {
1116 barf("allocateExec: can't handle large objects");
1119 if (exec_block == NULL ||
1120 exec_block->free + n + 1 > exec_block->start + BLOCK_SIZE_W) {
1122 lnat pagesize = getPageSize();
1123 bd = allocGroup(stg_max(1, pagesize / BLOCK_SIZE));
1124 debugTrace(DEBUG_gc, "allocate exec block %p", bd->start);
1126 bd->flags = BF_EXEC;
1127 bd->link = exec_block;
1128 if (exec_block != NULL) {
1129 exec_block->u.back = bd;
1132 setExecutable(bd->start, bd->blocks * BLOCK_SIZE, rtsTrue);
1135 *(exec_block->free) = n; // store the size of this chunk
1136 exec_block->gen_no += n; // gen_no stores the number of words allocated
1137 ret = exec_block->free + 1;
1138 exec_block->free += n + 1;
1145 void freeExec (void *addr)
1147 StgPtr p = (StgPtr)addr - 1;
1148 bdescr *bd = Bdescr((StgPtr)p);
1150 if ((bd->flags & BF_EXEC) == 0) {
1151 barf("freeExec: not executable");
1154 if (*(StgPtr)p == 0) {
1155 barf("freeExec: already free?");
1160 bd->gen_no -= *(StgPtr)p;
1163 if (bd->gen_no == 0) {
1164 // Free the block if it is empty, but not if it is the block at
1165 // the head of the queue.
1166 if (bd != exec_block) {
1167 debugTrace(DEBUG_gc, "free exec block %p", bd->start);
1168 dbl_link_remove(bd, &exec_block);
1169 setExecutable(bd->start, bd->blocks * BLOCK_SIZE, rtsFalse);
1172 bd->free = bd->start;
1179 #endif /* mingw32_HOST_OS */
1181 /* -----------------------------------------------------------------------------
1184 memInventory() checks for memory leaks by counting up all the
1185 blocks we know about and comparing that to the number of blocks
1186 allegedly floating around in the system.
1187 -------------------------------------------------------------------------- */
1191 // Useful for finding partially full blocks in gdb
1192 void findSlop(bdescr *bd);
1193 void findSlop(bdescr *bd)
1197 for (; bd != NULL; bd = bd->link) {
1198 slop = (bd->blocks * BLOCK_SIZE_W) - (bd->free - bd->start);
1199 if (slop > (1024/sizeof(W_))) {
1200 debugBelch("block at %p (bdescr %p) has %ldKB slop\n",
1201 bd->start, bd, slop / (1024/sizeof(W_)));
1207 countBlocks(bdescr *bd)
1210 for (n=0; bd != NULL; bd=bd->link) {
1216 // (*1) Just like countBlocks, except that we adjust the count for a
1217 // megablock group so that it doesn't include the extra few blocks
1218 // that would be taken up by block descriptors in the second and
1219 // subsequent megablock. This is so we can tally the count with the
1220 // number of blocks allocated in the system, for memInventory().
1222 countAllocdBlocks(bdescr *bd)
1225 for (n=0; bd != NULL; bd=bd->link) {
1227 // hack for megablock groups: see (*1) above
1228 if (bd->blocks > BLOCKS_PER_MBLOCK) {
1229 n -= (MBLOCK_SIZE / BLOCK_SIZE - BLOCKS_PER_MBLOCK)
1230 * (bd->blocks/(MBLOCK_SIZE/BLOCK_SIZE));
1237 stepBlocks (step *stp)
1239 ASSERT(countBlocks(stp->blocks) == stp->n_blocks);
1240 ASSERT(countBlocks(stp->large_objects) == stp->n_large_blocks);
1241 return stp->n_blocks + stp->n_old_blocks +
1242 countAllocdBlocks(stp->large_objects);
1245 // If memInventory() calculates that we have a memory leak, this
1246 // function will try to find the block(s) that are leaking by marking
1247 // all the ones that we know about, and search through memory to find
1248 // blocks that are not marked. In the debugger this can help to give
1249 // us a clue about what kind of block leaked. In the future we might
1250 // annotate blocks with their allocation site to give more helpful
1253 findMemoryLeak (void)
1256 for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
1257 for (i = 0; i < n_capabilities; i++) {
1258 markBlocks(capabilities[i].mut_lists[g]);
1260 markBlocks(generations[g].mut_list);
1261 for (s = 0; s < generations[g].n_steps; s++) {
1262 markBlocks(generations[g].steps[s].blocks);
1263 markBlocks(generations[g].steps[s].large_objects);
1267 for (i = 0; i < n_nurseries; i++) {
1268 markBlocks(nurseries[i].blocks);
1269 markBlocks(nurseries[i].large_objects);
1274 // if (RtsFlags.ProfFlags.doHeapProfile == HEAP_BY_RETAINER) {
1275 // markRetainerBlocks();
1279 // count the blocks allocated by the arena allocator
1281 // markArenaBlocks();
1283 // count the blocks containing executable memory
1284 markBlocks(exec_block);
1286 reportUnmarkedBlocks();
1291 memInventory (rtsBool show)
1295 lnat gen_blocks[RtsFlags.GcFlags.generations];
1296 lnat nursery_blocks, retainer_blocks,
1297 arena_blocks, exec_blocks;
1298 lnat live_blocks = 0, free_blocks = 0;
1301 // count the blocks we current have
1303 for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
1305 for (i = 0; i < n_capabilities; i++) {
1306 gen_blocks[g] += countBlocks(capabilities[i].mut_lists[g]);
1308 gen_blocks[g] += countAllocdBlocks(generations[g].mut_list);
1309 for (s = 0; s < generations[g].n_steps; s++) {
1310 stp = &generations[g].steps[s];
1311 gen_blocks[g] += stepBlocks(stp);
1316 for (i = 0; i < n_nurseries; i++) {
1317 nursery_blocks += stepBlocks(&nurseries[i]);
1320 retainer_blocks = 0;
1322 if (RtsFlags.ProfFlags.doHeapProfile == HEAP_BY_RETAINER) {
1323 retainer_blocks = retainerStackBlocks();
1327 // count the blocks allocated by the arena allocator
1328 arena_blocks = arenaBlocks();
1330 // count the blocks containing executable memory
1331 exec_blocks = countAllocdBlocks(exec_block);
1333 /* count the blocks on the free list */
1334 free_blocks = countFreeList();
1337 for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
1338 live_blocks += gen_blocks[g];
1340 live_blocks += nursery_blocks +
1341 + retainer_blocks + arena_blocks + exec_blocks;
1343 #define MB(n) (((n) * BLOCK_SIZE_W) / ((1024*1024)/sizeof(W_)))
1345 leak = live_blocks + free_blocks != mblocks_allocated * BLOCKS_PER_MBLOCK;
1350 debugBelch("Memory leak detected:\n");
1352 debugBelch("Memory inventory:\n");
1354 for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
1355 debugBelch(" gen %d blocks : %5lu blocks (%lu MB)\n", g,
1356 gen_blocks[g], MB(gen_blocks[g]));
1358 debugBelch(" nursery : %5lu blocks (%lu MB)\n",
1359 nursery_blocks, MB(nursery_blocks));
1360 debugBelch(" retainer : %5lu blocks (%lu MB)\n",
1361 retainer_blocks, MB(retainer_blocks));
1362 debugBelch(" arena blocks : %5lu blocks (%lu MB)\n",
1363 arena_blocks, MB(arena_blocks));
1364 debugBelch(" exec : %5lu blocks (%lu MB)\n",
1365 exec_blocks, MB(exec_blocks));
1366 debugBelch(" free : %5lu blocks (%lu MB)\n",
1367 free_blocks, MB(free_blocks));
1368 debugBelch(" total : %5lu blocks (%lu MB)\n",
1369 live_blocks + free_blocks, MB(live_blocks+free_blocks));
1371 debugBelch("\n in system : %5lu blocks (%lu MB)\n",
1372 mblocks_allocated * BLOCKS_PER_MBLOCK, mblocks_allocated);
1380 ASSERT(n_alloc_blocks == live_blocks);
1385 /* Full heap sanity check. */
1391 if (RtsFlags.GcFlags.generations == 1) {
1392 checkHeap(g0s0->blocks);
1393 checkLargeObjects(g0s0->large_objects);
1396 for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
1397 for (s = 0; s < generations[g].n_steps; s++) {
1398 if (g == 0 && s == 0) { continue; }
1399 ASSERT(countBlocks(generations[g].steps[s].blocks)
1400 == generations[g].steps[s].n_blocks);
1401 ASSERT(countBlocks(generations[g].steps[s].large_objects)
1402 == generations[g].steps[s].n_large_blocks);
1403 checkHeap(generations[g].steps[s].blocks);
1404 checkLargeObjects(generations[g].steps[s].large_objects);
1408 for (s = 0; s < n_nurseries; s++) {
1409 ASSERT(countBlocks(nurseries[s].blocks)
1410 == nurseries[s].n_blocks);
1411 ASSERT(countBlocks(nurseries[s].large_objects)
1412 == nurseries[s].n_large_blocks);
1415 checkFreeListSanity();
1418 #if defined(THREADED_RTS)
1419 // check the stacks too in threaded mode, because we don't do a
1420 // full heap sanity check in this case (see checkHeap())
1421 checkMutableLists(rtsTrue);
1423 checkMutableLists(rtsFalse);
1427 /* Nursery sanity check */
1429 checkNurserySanity( step *stp )
1435 for (bd = stp->blocks; bd != NULL; bd = bd->link) {
1436 ASSERT(bd->u.back == prev);
1438 blocks += bd->blocks;
1440 ASSERT(blocks == stp->n_blocks);
1443 // handy function for use in gdb, because Bdescr() is inlined.
1444 extern bdescr *_bdescr( StgPtr p );