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 nat alloc_blocks_lim; /* GC if n_large_blocks in any nursery
46 static bdescr *exec_block;
48 generation *generations = NULL; /* all the generations */
49 generation *g0 = NULL; /* generation 0, for convenience */
50 generation *oldest_gen = NULL; /* oldest generation, for convenience */
53 step *all_steps = NULL; /* single array of steps */
55 ullong total_allocated = 0; /* total memory allocated during run */
57 nat n_nurseries = 0; /* == RtsFlags.ParFlags.nNodes, convenience */
58 step *nurseries = NULL; /* array of nurseries, >1 only if THREADED_RTS */
62 * Storage manager mutex: protects all the above state from
63 * simultaneous access by two STG threads.
68 static void allocNurseries ( void );
71 initStep (step *stp, int g, int s)
74 stp->abs_no = RtsFlags.GcFlags.steps * g + s;
78 stp->live_estimate = 0;
79 stp->old_blocks = NULL;
80 stp->n_old_blocks = 0;
81 stp->gen = &generations[g];
83 stp->large_objects = NULL;
84 stp->n_large_blocks = 0;
85 stp->scavenged_large_objects = NULL;
86 stp->n_scavenged_large_blocks = 0;
91 initSpinLock(&stp->sync_large_objects);
93 stp->threads = END_TSO_QUEUE;
94 stp->old_threads = END_TSO_QUEUE;
103 if (generations != NULL) {
104 // multi-init protection
110 /* Sanity check to make sure the LOOKS_LIKE_ macros appear to be
111 * doing something reasonable.
113 /* We use the NOT_NULL variant or gcc warns that the test is always true */
114 ASSERT(LOOKS_LIKE_INFO_PTR_NOT_NULL((StgWord)&stg_BLACKHOLE_info));
115 ASSERT(LOOKS_LIKE_CLOSURE_PTR(&stg_dummy_ret_closure));
116 ASSERT(!HEAP_ALLOCED(&stg_dummy_ret_closure));
118 if (RtsFlags.GcFlags.maxHeapSize != 0 &&
119 RtsFlags.GcFlags.heapSizeSuggestion >
120 RtsFlags.GcFlags.maxHeapSize) {
121 RtsFlags.GcFlags.maxHeapSize = RtsFlags.GcFlags.heapSizeSuggestion;
124 if (RtsFlags.GcFlags.maxHeapSize != 0 &&
125 RtsFlags.GcFlags.minAllocAreaSize >
126 RtsFlags.GcFlags.maxHeapSize) {
127 errorBelch("maximum heap size (-M) is smaller than minimum alloc area size (-A)");
128 RtsFlags.GcFlags.minAllocAreaSize = RtsFlags.GcFlags.maxHeapSize;
131 initBlockAllocator();
133 #if defined(THREADED_RTS)
134 initMutex(&sm_mutex);
139 /* allocate generation info array */
140 generations = (generation *)stgMallocBytes(RtsFlags.GcFlags.generations
141 * sizeof(struct generation_),
142 "initStorage: gens");
144 /* Initialise all generations */
145 for(g = 0; g < RtsFlags.GcFlags.generations; g++) {
146 gen = &generations[g];
148 gen->mut_list = allocBlock();
149 gen->collections = 0;
150 gen->par_collections = 0;
151 gen->failed_promotions = 0;
155 /* A couple of convenience pointers */
156 g0 = &generations[0];
157 oldest_gen = &generations[RtsFlags.GcFlags.generations-1];
159 /* allocate all the steps into an array. It is important that we do
160 it this way, because we need the invariant that two step pointers
161 can be directly compared to see which is the oldest.
162 Remember that the last generation has only one step. */
163 total_steps = 1 + (RtsFlags.GcFlags.generations - 1) * RtsFlags.GcFlags.steps;
164 all_steps = stgMallocBytes(total_steps * sizeof(struct step_),
165 "initStorage: steps");
167 /* Allocate step structures in each generation */
168 if (RtsFlags.GcFlags.generations > 1) {
169 /* Only for multiple-generations */
171 /* Oldest generation: one step */
172 oldest_gen->n_steps = 1;
173 oldest_gen->steps = all_steps + (RtsFlags.GcFlags.generations - 1)
174 * RtsFlags.GcFlags.steps;
176 /* set up all except the oldest generation with 2 steps */
177 for(g = 0; g < RtsFlags.GcFlags.generations-1; g++) {
178 generations[g].n_steps = RtsFlags.GcFlags.steps;
179 generations[g].steps = all_steps + g * RtsFlags.GcFlags.steps;
183 /* single generation, i.e. a two-space collector */
185 g0->steps = all_steps;
188 n_nurseries = n_capabilities;
189 nurseries = stgMallocBytes (n_nurseries * sizeof(struct step_),
190 "initStorage: nurseries");
192 /* Initialise all steps */
193 for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
194 for (s = 0; s < generations[g].n_steps; s++) {
195 initStep(&generations[g].steps[s], g, s);
199 for (s = 0; s < n_nurseries; s++) {
200 initStep(&nurseries[s], 0, s);
203 /* Set up the destination pointers in each younger gen. step */
204 for (g = 0; g < RtsFlags.GcFlags.generations-1; g++) {
205 for (s = 0; s < generations[g].n_steps-1; s++) {
206 generations[g].steps[s].to = &generations[g].steps[s+1];
208 generations[g].steps[s].to = &generations[g+1].steps[0];
210 oldest_gen->steps[0].to = &oldest_gen->steps[0];
212 for (s = 0; s < n_nurseries; s++) {
213 nurseries[s].to = generations[0].steps[0].to;
216 /* The oldest generation has one step. */
217 if (RtsFlags.GcFlags.compact || RtsFlags.GcFlags.sweep) {
218 if (RtsFlags.GcFlags.generations == 1) {
219 errorBelch("WARNING: compact/sweep is incompatible with -G1; disabled");
221 oldest_gen->steps[0].mark = 1;
222 if (RtsFlags.GcFlags.compact)
223 oldest_gen->steps[0].compact = 1;
227 generations[0].max_blocks = 0;
229 /* The allocation area. Policy: keep the allocation area
230 * small to begin with, even if we have a large suggested heap
231 * size. Reason: we're going to do a major collection first, and we
232 * don't want it to be a big one. This vague idea is borne out by
233 * rigorous experimental evidence.
237 weak_ptr_list = NULL;
239 revertible_caf_list = NULL;
241 /* initialise the allocate() interface */
242 alloc_blocks_lim = RtsFlags.GcFlags.minAllocAreaSize;
247 initSpinLock(&gc_alloc_block_sync);
255 IF_DEBUG(gc, statDescribeGens());
263 stat_exit(calcAllocated());
269 stgFree(all_steps); // frees all the steps
270 stgFree(generations);
272 #if defined(THREADED_RTS)
273 closeMutex(&sm_mutex);
279 /* -----------------------------------------------------------------------------
282 The entry code for every CAF does the following:
284 - builds a CAF_BLACKHOLE in the heap
285 - pushes an update frame pointing to the CAF_BLACKHOLE
286 - invokes UPD_CAF(), which:
287 - calls newCaf, below
288 - updates the CAF with a static indirection to the CAF_BLACKHOLE
290 Why do we build a BLACKHOLE in the heap rather than just updating
291 the thunk directly? It's so that we only need one kind of update
292 frame - otherwise we'd need a static version of the update frame too.
294 newCaf() does the following:
296 - it puts the CAF on the oldest generation's mut-once list.
297 This is so that we can treat the CAF as a root when collecting
300 For GHCI, we have additional requirements when dealing with CAFs:
302 - we must *retain* all dynamically-loaded CAFs ever entered,
303 just in case we need them again.
304 - we must be able to *revert* CAFs that have been evaluated, to
305 their pre-evaluated form.
307 To do this, we use an additional CAF list. When newCaf() is
308 called on a dynamically-loaded CAF, we add it to the CAF list
309 instead of the old-generation mutable list, and save away its
310 old info pointer (in caf->saved_info) for later reversion.
312 To revert all the CAFs, we traverse the CAF list and reset the
313 info pointer to caf->saved_info, then throw away the CAF list.
314 (see GC.c:revertCAFs()).
318 -------------------------------------------------------------------------- */
321 newCAF(StgClosure* caf)
329 // If we are in GHCi _and_ we are using dynamic libraries,
330 // then we can't redirect newCAF calls to newDynCAF (see below),
331 // so we make newCAF behave almost like newDynCAF.
332 // The dynamic libraries might be used by both the interpreted
333 // program and GHCi itself, so they must not be reverted.
334 // This also means that in GHCi with dynamic libraries, CAFs are not
335 // garbage collected. If this turns out to be a problem, we could
336 // do another hack here and do an address range test on caf to figure
337 // out whether it is from a dynamic library.
338 ((StgIndStatic *)caf)->saved_info = (StgInfoTable *)caf->header.info;
339 ((StgIndStatic *)caf)->static_link = caf_list;
345 /* Put this CAF on the mutable list for the old generation.
346 * This is a HACK - the IND_STATIC closure doesn't really have
347 * a mut_link field, but we pretend it has - in fact we re-use
348 * the STATIC_LINK field for the time being, because when we
349 * come to do a major GC we won't need the mut_link field
350 * any more and can use it as a STATIC_LINK.
352 ((StgIndStatic *)caf)->saved_info = NULL;
353 recordMutableGen(caf, oldest_gen->no);
359 // An alternate version of newCaf which is used for dynamically loaded
360 // object code in GHCi. In this case we want to retain *all* CAFs in
361 // the object code, because they might be demanded at any time from an
362 // expression evaluated on the command line.
363 // Also, GHCi might want to revert CAFs, so we add these to the
364 // revertible_caf_list.
366 // The linker hackily arranges that references to newCaf from dynamic
367 // code end up pointing to newDynCAF.
369 newDynCAF(StgClosure *caf)
373 ((StgIndStatic *)caf)->saved_info = (StgInfoTable *)caf->header.info;
374 ((StgIndStatic *)caf)->static_link = revertible_caf_list;
375 revertible_caf_list = caf;
380 /* -----------------------------------------------------------------------------
382 -------------------------------------------------------------------------- */
385 allocNursery (step *stp, bdescr *tail, nat blocks)
390 // Allocate a nursery: we allocate fresh blocks one at a time and
391 // cons them on to the front of the list, not forgetting to update
392 // the back pointer on the tail of the list to point to the new block.
393 for (i=0; i < blocks; i++) {
396 processNursery() in LdvProfile.c assumes that every block group in
397 the nursery contains only a single block. So, if a block group is
398 given multiple blocks, change processNursery() accordingly.
402 // double-link the nursery: we might need to insert blocks
408 bd->free = bd->start;
416 assignNurseriesToCapabilities (void)
420 for (i = 0; i < n_nurseries; i++) {
421 capabilities[i].r.rNursery = &nurseries[i];
422 capabilities[i].r.rCurrentNursery = nurseries[i].blocks;
423 capabilities[i].r.rCurrentAlloc = NULL;
428 allocNurseries( void )
432 for (i = 0; i < n_nurseries; i++) {
433 nurseries[i].blocks =
434 allocNursery(&nurseries[i], NULL,
435 RtsFlags.GcFlags.minAllocAreaSize);
436 nurseries[i].n_blocks = RtsFlags.GcFlags.minAllocAreaSize;
437 nurseries[i].old_blocks = NULL;
438 nurseries[i].n_old_blocks = 0;
440 assignNurseriesToCapabilities();
444 resetNurseries( void )
450 for (i = 0; i < n_nurseries; i++) {
452 for (bd = stp->blocks; bd; bd = bd->link) {
453 bd->free = bd->start;
454 ASSERT(bd->gen_no == 0);
455 ASSERT(bd->step == stp);
456 IF_DEBUG(sanity,memset(bd->start, 0xaa, BLOCK_SIZE));
458 // these large objects are dead, since we have just GC'd
459 freeChain(stp->large_objects);
460 stp->large_objects = NULL;
461 stp->n_large_blocks = 0;
463 assignNurseriesToCapabilities();
467 countNurseryBlocks (void)
472 for (i = 0; i < n_nurseries; i++) {
473 blocks += nurseries[i].n_blocks;
474 blocks += nurseries[i].n_large_blocks;
480 resizeNursery ( step *stp, nat blocks )
485 nursery_blocks = stp->n_blocks;
486 if (nursery_blocks == blocks) return;
488 if (nursery_blocks < blocks) {
489 debugTrace(DEBUG_gc, "increasing size of nursery to %d blocks",
491 stp->blocks = allocNursery(stp, stp->blocks, blocks-nursery_blocks);
496 debugTrace(DEBUG_gc, "decreasing size of nursery to %d blocks",
500 while (nursery_blocks > blocks) {
502 next_bd->u.back = NULL;
503 nursery_blocks -= bd->blocks; // might be a large block
508 // might have gone just under, by freeing a large block, so make
509 // up the difference.
510 if (nursery_blocks < blocks) {
511 stp->blocks = allocNursery(stp, stp->blocks, blocks-nursery_blocks);
515 stp->n_blocks = blocks;
516 ASSERT(countBlocks(stp->blocks) == stp->n_blocks);
520 // Resize each of the nurseries to the specified size.
523 resizeNurseriesFixed (nat blocks)
526 for (i = 0; i < n_nurseries; i++) {
527 resizeNursery(&nurseries[i], blocks);
532 // Resize the nurseries to the total specified size.
535 resizeNurseries (nat blocks)
537 // If there are multiple nurseries, then we just divide the number
538 // of available blocks between them.
539 resizeNurseriesFixed(blocks / n_nurseries);
543 /* -----------------------------------------------------------------------------
544 move_TSO is called to update the TSO structure after it has been
545 moved from one place to another.
546 -------------------------------------------------------------------------- */
549 move_TSO (StgTSO *src, StgTSO *dest)
553 // relocate the stack pointer...
554 diff = (StgPtr)dest - (StgPtr)src; // In *words*
555 dest->sp = (StgPtr)dest->sp + diff;
558 /* -----------------------------------------------------------------------------
559 split N blocks off the front of the given bdescr, returning the
560 new block group. We add the remainder to the large_blocks list
561 in the same step as the original block.
562 -------------------------------------------------------------------------- */
565 splitLargeBlock (bdescr *bd, nat blocks)
571 ASSERT(countBlocks(bd->step->large_objects) == bd->step->n_large_blocks);
573 // subtract the original number of blocks from the counter first
574 bd->step->n_large_blocks -= bd->blocks;
576 new_bd = splitBlockGroup (bd, blocks);
577 initBdescr(new_bd, bd->step);
578 new_bd->flags = BF_LARGE | (bd->flags & BF_EVACUATED);
579 // if new_bd is in an old generation, we have to set BF_EVACUATED
580 new_bd->free = bd->free;
581 dbl_link_onto(new_bd, &bd->step->large_objects);
583 ASSERT(new_bd->free <= new_bd->start + new_bd->blocks * BLOCK_SIZE_W);
585 // add the new number of blocks to the counter. Due to the gaps
586 // for block descriptors, new_bd->blocks + bd->blocks might not be
587 // equal to the original bd->blocks, which is why we do it this way.
588 bd->step->n_large_blocks += bd->blocks + new_bd->blocks;
590 ASSERT(countBlocks(bd->step->large_objects) == bd->step->n_large_blocks);
597 /* -----------------------------------------------------------------------------
600 This allocates memory in the current thread - it is intended for
601 use primarily from STG-land where we have a Capability. It is
602 better than allocate() because it doesn't require taking the
603 sm_mutex lock in the common case.
605 Memory is allocated directly from the nursery if possible (but not
606 from the current nursery block, so as not to interfere with
608 -------------------------------------------------------------------------- */
611 allocate (Capability *cap, lnat n)
617 if (n >= LARGE_OBJECT_THRESHOLD/sizeof(W_)) {
618 lnat req_blocks = (lnat)BLOCK_ROUND_UP(n*sizeof(W_)) / BLOCK_SIZE;
620 // Attempting to allocate an object larger than maxHeapSize
621 // should definitely be disallowed. (bug #1791)
622 if (RtsFlags.GcFlags.maxHeapSize > 0 &&
623 req_blocks >= RtsFlags.GcFlags.maxHeapSize) {
625 // heapOverflow() doesn't exit (see #2592), but we aren't
626 // in a position to do a clean shutdown here: we
627 // either have to allocate the memory or exit now.
628 // Allocating the memory would be bad, because the user
629 // has requested that we not exceed maxHeapSize, so we
631 stg_exit(EXIT_HEAPOVERFLOW);
634 stp = &nurseries[cap->no];
636 bd = allocGroup(req_blocks);
637 dbl_link_onto(bd, &stp->large_objects);
638 stp->n_large_blocks += bd->blocks; // might be larger than req_blocks
640 bd->flags = BF_LARGE;
641 bd->free = bd->start + n;
645 /* small allocation (<LARGE_OBJECT_THRESHOLD) */
647 TICK_ALLOC_HEAP_NOCTR(n);
650 bd = cap->r.rCurrentAlloc;
651 if (bd == NULL || bd->free + n > bd->start + BLOCK_SIZE_W) {
653 // The CurrentAlloc block is full, we need to find another
654 // one. First, we try taking the next block from the
656 bd = cap->r.rCurrentNursery->link;
658 if (bd == NULL || bd->free + n > bd->start + BLOCK_SIZE_W) {
659 // The nursery is empty, or the next block is already
660 // full: allocate a fresh block (we can't fail here).
663 cap->r.rNursery->n_blocks++;
665 initBdescr(bd, cap->r.rNursery);
667 // If we had to allocate a new block, then we'll GC
668 // pretty quickly now, because MAYBE_GC() will
669 // notice that CurrentNursery->link is NULL.
671 // we have a block in the nursery: take it and put
672 // it at the *front* of the nursery list, and use it
673 // to allocate() from.
674 cap->r.rCurrentNursery->link = bd->link;
675 if (bd->link != NULL) {
676 bd->link->u.back = cap->r.rCurrentNursery;
679 dbl_link_onto(bd, &cap->r.rNursery->blocks);
680 cap->r.rCurrentAlloc = bd;
681 IF_DEBUG(sanity, checkNurserySanity(cap->r.rNursery));
688 /* ---------------------------------------------------------------------------
689 Allocate a fixed/pinned object.
691 We allocate small pinned objects into a single block, allocating a
692 new block when the current one overflows. The block is chained
693 onto the large_object_list of generation 0 step 0.
695 NOTE: The GC can't in general handle pinned objects. This
696 interface is only safe to use for ByteArrays, which have no
697 pointers and don't require scavenging. It works because the
698 block's descriptor has the BF_LARGE flag set, so the block is
699 treated as a large object and chained onto various lists, rather
700 than the individual objects being copied. However, when it comes
701 to scavenge the block, the GC will only scavenge the first object.
702 The reason is that the GC can't linearly scan a block of pinned
703 objects at the moment (doing so would require using the
704 mostly-copying techniques). But since we're restricting ourselves
705 to pinned ByteArrays, not scavenging is ok.
707 This function is called by newPinnedByteArray# which immediately
708 fills the allocated memory with a MutableByteArray#.
709 ------------------------------------------------------------------------- */
712 allocatePinned (Capability *cap, lnat n)
718 // If the request is for a large object, then allocate()
719 // will give us a pinned object anyway.
720 if (n >= LARGE_OBJECT_THRESHOLD/sizeof(W_)) {
721 p = allocate(cap, n);
722 Bdescr(p)->flags |= BF_PINNED;
726 TICK_ALLOC_HEAP_NOCTR(n);
729 bd = cap->pinned_object_block;
731 // If we don't have a block of pinned objects yet, or the current
732 // one isn't large enough to hold the new object, allocate a new one.
733 if (bd == NULL || (bd->free + n) > (bd->start + BLOCK_SIZE_W)) {
735 cap->pinned_object_block = bd = allocBlock();
737 stp = &nurseries[cap->no];
738 dbl_link_onto(bd, &stp->large_objects);
739 stp->n_large_blocks++;
741 bd->flags = BF_PINNED | BF_LARGE;
742 bd->free = bd->start;
750 /* -----------------------------------------------------------------------------
752 -------------------------------------------------------------------------- */
755 This is the write barrier for MUT_VARs, a.k.a. IORefs. A
756 MUT_VAR_CLEAN object is not on the mutable list; a MUT_VAR_DIRTY
757 is. When written to, a MUT_VAR_CLEAN turns into a MUT_VAR_DIRTY
758 and is put on the mutable list.
761 dirty_MUT_VAR(StgRegTable *reg, StgClosure *p)
763 Capability *cap = regTableToCapability(reg);
765 if (p->header.info == &stg_MUT_VAR_CLEAN_info) {
766 p->header.info = &stg_MUT_VAR_DIRTY_info;
767 bd = Bdescr((StgPtr)p);
768 if (bd->gen_no > 0) recordMutableCap(p,cap,bd->gen_no);
772 // Setting a TSO's link field with a write barrier.
773 // It is *not* necessary to call this function when
774 // * setting the link field to END_TSO_QUEUE
775 // * putting a TSO on the blackhole_queue
776 // * setting the link field of the currently running TSO, as it
777 // will already be dirty.
779 setTSOLink (Capability *cap, StgTSO *tso, StgTSO *target)
782 if (tso->dirty == 0 && (tso->flags & TSO_LINK_DIRTY) == 0) {
783 tso->flags |= TSO_LINK_DIRTY;
784 bd = Bdescr((StgPtr)tso);
785 if (bd->gen_no > 0) recordMutableCap((StgClosure*)tso,cap,bd->gen_no);
791 dirty_TSO (Capability *cap, StgTSO *tso)
794 if (tso->dirty == 0 && (tso->flags & TSO_LINK_DIRTY) == 0) {
795 bd = Bdescr((StgPtr)tso);
796 if (bd->gen_no > 0) recordMutableCap((StgClosure*)tso,cap,bd->gen_no);
802 This is the write barrier for MVARs. An MVAR_CLEAN objects is not
803 on the mutable list; a MVAR_DIRTY is. When written to, a
804 MVAR_CLEAN turns into a MVAR_DIRTY and is put on the mutable list.
805 The check for MVAR_CLEAN is inlined at the call site for speed,
806 this really does make a difference on concurrency-heavy benchmarks
807 such as Chaneneos and cheap-concurrency.
810 dirty_MVAR(StgRegTable *reg, StgClosure *p)
812 Capability *cap = regTableToCapability(reg);
814 bd = Bdescr((StgPtr)p);
815 if (bd->gen_no > 0) recordMutableCap(p,cap,bd->gen_no);
818 /* -----------------------------------------------------------------------------
820 * -------------------------------------------------------------------------- */
822 /* -----------------------------------------------------------------------------
825 * Approximate how much we've allocated: number of blocks in the
826 * nursery + blocks allocated via allocate() - unused nusery blocks.
827 * This leaves a little slop at the end of each block, and doesn't
828 * take into account large objects (ToDo).
829 * -------------------------------------------------------------------------- */
832 calcAllocated( void )
838 allocated = countNurseryBlocks() * BLOCK_SIZE_W;
840 for (i = 0; i < n_capabilities; i++) {
842 for ( bd = capabilities[i].r.rCurrentNursery->link;
843 bd != NULL; bd = bd->link ) {
844 allocated -= BLOCK_SIZE_W;
846 cap = &capabilities[i];
847 if (cap->r.rCurrentNursery->free <
848 cap->r.rCurrentNursery->start + BLOCK_SIZE_W) {
849 allocated -= (cap->r.rCurrentNursery->start + BLOCK_SIZE_W)
850 - cap->r.rCurrentNursery->free;
852 if (cap->pinned_object_block != NULL) {
853 allocated -= (cap->pinned_object_block->start + BLOCK_SIZE_W) -
854 cap->pinned_object_block->free;
858 total_allocated += allocated;
862 /* Approximate the amount of live data in the heap. To be called just
863 * after garbage collection (see GarbageCollect()).
872 for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
873 for (s = 0; s < generations[g].n_steps; s++) {
874 /* approximate amount of live data (doesn't take into account slop
875 * at end of each block).
877 if (g == 0 && s == 0 && RtsFlags.GcFlags.generations > 1) {
880 stp = &generations[g].steps[s];
881 live += stp->n_large_blocks + stp->n_blocks;
888 countOccupied(bdescr *bd)
893 for (; bd != NULL; bd = bd->link) {
894 ASSERT(bd->free <= bd->start + bd->blocks * BLOCK_SIZE_W);
895 words += bd->free - bd->start;
900 // Return an accurate count of the live data in the heap, excluding
910 for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
911 for (s = 0; s < generations[g].n_steps; s++) {
912 if (g == 0 && s == 0 && RtsFlags.GcFlags.generations > 1) continue;
913 stp = &generations[g].steps[s];
914 live += stp->n_words + countOccupied(stp->large_objects);
920 /* Approximate the number of blocks that will be needed at the next
921 * garbage collection.
923 * Assume: all data currently live will remain live. Steps that will
924 * be collected next time will therefore need twice as many blocks
925 * since all the data will be copied.
934 for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
935 for (s = 0; s < generations[g].n_steps; s++) {
936 if (g == 0 && s == 0) { continue; }
937 stp = &generations[g].steps[s];
939 // we need at least this much space
940 needed += stp->n_blocks + stp->n_large_blocks;
942 // any additional space needed to collect this gen next time?
943 if (g == 0 || // always collect gen 0
944 (generations[g].steps[0].n_blocks +
945 generations[g].steps[0].n_large_blocks
946 > generations[g].max_blocks)) {
947 // we will collect this gen next time
950 needed += stp->n_blocks / BITS_IN(W_);
952 needed += stp->n_blocks / 100;
955 continue; // no additional space needed for compaction
957 needed += stp->n_blocks;
965 /* ----------------------------------------------------------------------------
968 Executable memory must be managed separately from non-executable
969 memory. Most OSs these days require you to jump through hoops to
970 dynamically allocate executable memory, due to various security
973 Here we provide a small memory allocator for executable memory.
974 Memory is managed with a page granularity; we allocate linearly
975 in the page, and when the page is emptied (all objects on the page
976 are free) we free the page again, not forgetting to make it
979 TODO: The inability to handle objects bigger than BLOCK_SIZE_W means that
980 the linker cannot use allocateExec for loading object code files
981 on Windows. Once allocateExec can handle larger objects, the linker
982 should be modified to use allocateExec instead of VirtualAlloc.
983 ------------------------------------------------------------------------- */
985 #if defined(linux_HOST_OS)
987 // On Linux we need to use libffi for allocating executable memory,
988 // because it knows how to work around the restrictions put in place
991 void *allocateExec (nat bytes, void **exec_ret)
995 ret = ffi_closure_alloc (sizeof(void *) + (size_t)bytes, (void**)&exec);
997 if (ret == NULL) return ret;
998 *ret = ret; // save the address of the writable mapping, for freeExec().
999 *exec_ret = exec + 1;
1003 // freeExec gets passed the executable address, not the writable address.
1004 void freeExec (void *addr)
1007 writable = *((void**)addr - 1);
1009 ffi_closure_free (writable);
1015 void *allocateExec (nat bytes, void **exec_ret)
1022 // round up to words.
1023 n = (bytes + sizeof(W_) + 1) / sizeof(W_);
1025 if (n+1 > BLOCK_SIZE_W) {
1026 barf("allocateExec: can't handle large objects");
1029 if (exec_block == NULL ||
1030 exec_block->free + n + 1 > exec_block->start + BLOCK_SIZE_W) {
1032 lnat pagesize = getPageSize();
1033 bd = allocGroup(stg_max(1, pagesize / BLOCK_SIZE));
1034 debugTrace(DEBUG_gc, "allocate exec block %p", bd->start);
1036 bd->flags = BF_EXEC;
1037 bd->link = exec_block;
1038 if (exec_block != NULL) {
1039 exec_block->u.back = bd;
1042 setExecutable(bd->start, bd->blocks * BLOCK_SIZE, rtsTrue);
1045 *(exec_block->free) = n; // store the size of this chunk
1046 exec_block->gen_no += n; // gen_no stores the number of words allocated
1047 ret = exec_block->free + 1;
1048 exec_block->free += n + 1;
1055 void freeExec (void *addr)
1057 StgPtr p = (StgPtr)addr - 1;
1058 bdescr *bd = Bdescr((StgPtr)p);
1060 if ((bd->flags & BF_EXEC) == 0) {
1061 barf("freeExec: not executable");
1064 if (*(StgPtr)p == 0) {
1065 barf("freeExec: already free?");
1070 bd->gen_no -= *(StgPtr)p;
1073 if (bd->gen_no == 0) {
1074 // Free the block if it is empty, but not if it is the block at
1075 // the head of the queue.
1076 if (bd != exec_block) {
1077 debugTrace(DEBUG_gc, "free exec block %p", bd->start);
1078 dbl_link_remove(bd, &exec_block);
1079 setExecutable(bd->start, bd->blocks * BLOCK_SIZE, rtsFalse);
1082 bd->free = bd->start;
1089 #endif /* mingw32_HOST_OS */
1091 /* -----------------------------------------------------------------------------
1094 memInventory() checks for memory leaks by counting up all the
1095 blocks we know about and comparing that to the number of blocks
1096 allegedly floating around in the system.
1097 -------------------------------------------------------------------------- */
1101 // Useful for finding partially full blocks in gdb
1102 void findSlop(bdescr *bd);
1103 void findSlop(bdescr *bd)
1107 for (; bd != NULL; bd = bd->link) {
1108 slop = (bd->blocks * BLOCK_SIZE_W) - (bd->free - bd->start);
1109 if (slop > (1024/sizeof(W_))) {
1110 debugBelch("block at %p (bdescr %p) has %ldKB slop\n",
1111 bd->start, bd, slop / (1024/sizeof(W_)));
1117 countBlocks(bdescr *bd)
1120 for (n=0; bd != NULL; bd=bd->link) {
1126 // (*1) Just like countBlocks, except that we adjust the count for a
1127 // megablock group so that it doesn't include the extra few blocks
1128 // that would be taken up by block descriptors in the second and
1129 // subsequent megablock. This is so we can tally the count with the
1130 // number of blocks allocated in the system, for memInventory().
1132 countAllocdBlocks(bdescr *bd)
1135 for (n=0; bd != NULL; bd=bd->link) {
1137 // hack for megablock groups: see (*1) above
1138 if (bd->blocks > BLOCKS_PER_MBLOCK) {
1139 n -= (MBLOCK_SIZE / BLOCK_SIZE - BLOCKS_PER_MBLOCK)
1140 * (bd->blocks/(MBLOCK_SIZE/BLOCK_SIZE));
1147 stepBlocks (step *stp)
1149 ASSERT(countBlocks(stp->blocks) == stp->n_blocks);
1150 ASSERT(countBlocks(stp->large_objects) == stp->n_large_blocks);
1151 return stp->n_blocks + stp->n_old_blocks +
1152 countAllocdBlocks(stp->large_objects);
1155 // If memInventory() calculates that we have a memory leak, this
1156 // function will try to find the block(s) that are leaking by marking
1157 // all the ones that we know about, and search through memory to find
1158 // blocks that are not marked. In the debugger this can help to give
1159 // us a clue about what kind of block leaked. In the future we might
1160 // annotate blocks with their allocation site to give more helpful
1163 findMemoryLeak (void)
1166 for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
1167 for (i = 0; i < n_capabilities; i++) {
1168 markBlocks(capabilities[i].mut_lists[g]);
1170 markBlocks(generations[g].mut_list);
1171 for (s = 0; s < generations[g].n_steps; s++) {
1172 markBlocks(generations[g].steps[s].blocks);
1173 markBlocks(generations[g].steps[s].large_objects);
1177 for (i = 0; i < n_nurseries; i++) {
1178 markBlocks(nurseries[i].blocks);
1179 markBlocks(nurseries[i].large_objects);
1184 // if (RtsFlags.ProfFlags.doHeapProfile == HEAP_BY_RETAINER) {
1185 // markRetainerBlocks();
1189 // count the blocks allocated by the arena allocator
1191 // markArenaBlocks();
1193 // count the blocks containing executable memory
1194 markBlocks(exec_block);
1196 reportUnmarkedBlocks();
1201 memInventory (rtsBool show)
1205 lnat gen_blocks[RtsFlags.GcFlags.generations];
1206 lnat nursery_blocks, retainer_blocks,
1207 arena_blocks, exec_blocks;
1208 lnat live_blocks = 0, free_blocks = 0;
1211 // count the blocks we current have
1213 for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
1215 for (i = 0; i < n_capabilities; i++) {
1216 gen_blocks[g] += countBlocks(capabilities[i].mut_lists[g]);
1218 gen_blocks[g] += countAllocdBlocks(generations[g].mut_list);
1219 for (s = 0; s < generations[g].n_steps; s++) {
1220 stp = &generations[g].steps[s];
1221 gen_blocks[g] += stepBlocks(stp);
1226 for (i = 0; i < n_nurseries; i++) {
1227 nursery_blocks += stepBlocks(&nurseries[i]);
1230 retainer_blocks = 0;
1232 if (RtsFlags.ProfFlags.doHeapProfile == HEAP_BY_RETAINER) {
1233 retainer_blocks = retainerStackBlocks();
1237 // count the blocks allocated by the arena allocator
1238 arena_blocks = arenaBlocks();
1240 // count the blocks containing executable memory
1241 exec_blocks = countAllocdBlocks(exec_block);
1243 /* count the blocks on the free list */
1244 free_blocks = countFreeList();
1247 for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
1248 live_blocks += gen_blocks[g];
1250 live_blocks += nursery_blocks +
1251 + retainer_blocks + arena_blocks + exec_blocks;
1253 #define MB(n) (((n) * BLOCK_SIZE_W) / ((1024*1024)/sizeof(W_)))
1255 leak = live_blocks + free_blocks != mblocks_allocated * BLOCKS_PER_MBLOCK;
1260 debugBelch("Memory leak detected:\n");
1262 debugBelch("Memory inventory:\n");
1264 for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
1265 debugBelch(" gen %d blocks : %5lu blocks (%lu MB)\n", g,
1266 gen_blocks[g], MB(gen_blocks[g]));
1268 debugBelch(" nursery : %5lu blocks (%lu MB)\n",
1269 nursery_blocks, MB(nursery_blocks));
1270 debugBelch(" retainer : %5lu blocks (%lu MB)\n",
1271 retainer_blocks, MB(retainer_blocks));
1272 debugBelch(" arena blocks : %5lu blocks (%lu MB)\n",
1273 arena_blocks, MB(arena_blocks));
1274 debugBelch(" exec : %5lu blocks (%lu MB)\n",
1275 exec_blocks, MB(exec_blocks));
1276 debugBelch(" free : %5lu blocks (%lu MB)\n",
1277 free_blocks, MB(free_blocks));
1278 debugBelch(" total : %5lu blocks (%lu MB)\n",
1279 live_blocks + free_blocks, MB(live_blocks+free_blocks));
1281 debugBelch("\n in system : %5lu blocks (%lu MB)\n",
1282 mblocks_allocated * BLOCKS_PER_MBLOCK, mblocks_allocated);
1290 ASSERT(n_alloc_blocks == live_blocks);
1295 /* Full heap sanity check. */
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 && RtsFlags.GcFlags.generations > 1) {
1306 ASSERT(countBlocks(generations[g].steps[s].blocks)
1307 == generations[g].steps[s].n_blocks);
1308 ASSERT(countBlocks(generations[g].steps[s].large_objects)
1309 == generations[g].steps[s].n_large_blocks);
1310 checkHeap(generations[g].steps[s].blocks);
1311 checkLargeObjects(generations[g].steps[s].large_objects);
1315 for (s = 0; s < n_nurseries; s++) {
1316 ASSERT(countBlocks(nurseries[s].blocks)
1317 == nurseries[s].n_blocks);
1318 ASSERT(countBlocks(nurseries[s].large_objects)
1319 == nurseries[s].n_large_blocks);
1322 checkFreeListSanity();
1324 #if defined(THREADED_RTS)
1325 // check the stacks too in threaded mode, because we don't do a
1326 // full heap sanity check in this case (see checkHeap())
1327 checkMutableLists(rtsTrue);
1329 checkMutableLists(rtsFalse);
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 );