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"
25 #include "OSThreads.h"
26 #include "Capability.h"
29 #include "RetainerProfile.h" // for counting memory blocks (memInventory)
41 * All these globals require sm_mutex to access in THREADED_RTS mode.
43 StgClosure *caf_list = NULL;
44 StgClosure *revertible_caf_list = NULL;
47 bdescr *pinned_object_block; /* allocate pinned objects into this block */
48 nat alloc_blocks; /* number of allocate()d blocks since GC */
49 nat alloc_blocks_lim; /* approximate limit on alloc_blocks */
51 static bdescr *exec_block;
53 generation *generations = NULL; /* all the generations */
54 generation *g0 = NULL; /* generation 0, for convenience */
55 generation *oldest_gen = NULL; /* oldest generation, for convenience */
56 step *g0s0 = NULL; /* generation 0, step 0, for convenience */
59 step *all_steps = NULL; /* single array of steps */
61 ullong total_allocated = 0; /* total memory allocated during run */
63 nat n_nurseries = 0; /* == RtsFlags.ParFlags.nNodes, convenience */
64 step *nurseries = NULL; /* array of nurseries, >1 only if THREADED_RTS */
68 * Storage manager mutex: protects all the above state from
69 * simultaneous access by two STG threads.
76 initStep (step *stp, int g, int s)
79 stp->abs_no = RtsFlags.GcFlags.steps * g + s;
83 stp->live_estimate = 0;
84 stp->old_blocks = NULL;
85 stp->n_old_blocks = 0;
86 stp->gen = &generations[g];
88 stp->large_objects = NULL;
89 stp->n_large_blocks = 0;
90 stp->scavenged_large_objects = NULL;
91 stp->n_scavenged_large_blocks = 0;
96 initSpinLock(&stp->sync_large_objects);
98 stp->threads = END_TSO_QUEUE;
99 stp->old_threads = END_TSO_QUEUE;
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((StgWord)&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);
144 /* allocate generation info array */
145 generations = (generation *)stgMallocBytes(RtsFlags.GcFlags.generations
146 * sizeof(struct generation_),
147 "initStorage: gens");
149 /* allocate all the steps into an array. It is important that we do
150 it this way, because we need the invariant that two step pointers
151 can be directly compared to see which is the oldest.
152 Remember that the last generation has only one step. */
153 total_steps = 1 + (RtsFlags.GcFlags.generations - 1) * RtsFlags.GcFlags.steps;
154 all_steps = stgMallocBytes(total_steps * sizeof(struct step_),
155 "initStorage: steps");
157 /* Initialise all generations */
158 for(g = 0; g < RtsFlags.GcFlags.generations; g++) {
159 gen = &generations[g];
161 gen->mut_list = allocBlock();
162 gen->collections = 0;
163 gen->par_collections = 0;
164 gen->failed_promotions = 0;
168 /* A couple of convenience pointers */
169 g0 = &generations[0];
170 oldest_gen = &generations[RtsFlags.GcFlags.generations-1];
172 /* Allocate step structures in each generation */
173 if (RtsFlags.GcFlags.generations > 1) {
174 /* Only for multiple-generations */
176 /* Oldest generation: one step */
177 oldest_gen->n_steps = 1;
178 oldest_gen->steps = all_steps + (RtsFlags.GcFlags.generations - 1)
179 * RtsFlags.GcFlags.steps;
181 /* set up all except the oldest generation with 2 steps */
182 for(g = 0; g < RtsFlags.GcFlags.generations-1; g++) {
183 generations[g].n_steps = RtsFlags.GcFlags.steps;
184 generations[g].steps = all_steps + g * RtsFlags.GcFlags.steps;
188 /* single generation, i.e. a two-space collector */
190 g0->steps = all_steps;
194 n_nurseries = n_capabilities;
198 nurseries = stgMallocBytes (n_nurseries * sizeof(struct step_),
199 "initStorage: nurseries");
201 /* Initialise all steps */
202 for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
203 for (s = 0; s < generations[g].n_steps; s++) {
204 initStep(&generations[g].steps[s], g, s);
208 for (s = 0; s < n_nurseries; s++) {
209 initStep(&nurseries[s], 0, s);
212 /* Set up the destination pointers in each younger gen. step */
213 for (g = 0; g < RtsFlags.GcFlags.generations-1; g++) {
214 for (s = 0; s < generations[g].n_steps-1; s++) {
215 generations[g].steps[s].to = &generations[g].steps[s+1];
217 generations[g].steps[s].to = &generations[g+1].steps[0];
219 oldest_gen->steps[0].to = &oldest_gen->steps[0];
221 for (s = 0; s < n_nurseries; s++) {
222 nurseries[s].to = generations[0].steps[0].to;
225 /* The oldest generation has one step. */
226 if (RtsFlags.GcFlags.compact || RtsFlags.GcFlags.sweep) {
227 if (RtsFlags.GcFlags.generations == 1) {
228 errorBelch("WARNING: compact/sweep is incompatible with -G1; disabled");
230 oldest_gen->steps[0].mark = 1;
231 if (RtsFlags.GcFlags.compact)
232 oldest_gen->steps[0].compact = 1;
236 generations[0].max_blocks = 0;
237 g0s0 = &generations[0].steps[0];
239 /* The allocation area. Policy: keep the allocation area
240 * small to begin with, even if we have a large suggested heap
241 * size. Reason: we're going to do a major collection first, and we
242 * don't want it to be a big one. This vague idea is borne out by
243 * rigorous experimental evidence.
247 weak_ptr_list = NULL;
249 revertible_caf_list = NULL;
251 /* initialise the allocate() interface */
253 alloc_blocks_lim = RtsFlags.GcFlags.minAllocAreaSize;
258 initSpinLock(&gc_alloc_block_sync);
266 IF_DEBUG(gc, statDescribeGens());
274 stat_exit(calcAllocated());
280 stgFree(g0s0); // frees all the steps
281 stgFree(generations);
283 #if defined(THREADED_RTS)
284 closeMutex(&sm_mutex);
289 /* -----------------------------------------------------------------------------
292 The entry code for every CAF does the following:
294 - builds a CAF_BLACKHOLE in the heap
295 - pushes an update frame pointing to the CAF_BLACKHOLE
296 - invokes UPD_CAF(), which:
297 - calls newCaf, below
298 - updates the CAF with a static indirection to the CAF_BLACKHOLE
300 Why do we build a BLACKHOLE in the heap rather than just updating
301 the thunk directly? It's so that we only need one kind of update
302 frame - otherwise we'd need a static version of the update frame too.
304 newCaf() does the following:
306 - it puts the CAF on the oldest generation's mut-once list.
307 This is so that we can treat the CAF as a root when collecting
310 For GHCI, we have additional requirements when dealing with CAFs:
312 - we must *retain* all dynamically-loaded CAFs ever entered,
313 just in case we need them again.
314 - we must be able to *revert* CAFs that have been evaluated, to
315 their pre-evaluated form.
317 To do this, we use an additional CAF list. When newCaf() is
318 called on a dynamically-loaded CAF, we add it to the CAF list
319 instead of the old-generation mutable list, and save away its
320 old info pointer (in caf->saved_info) for later reversion.
322 To revert all the CAFs, we traverse the CAF list and reset the
323 info pointer to caf->saved_info, then throw away the CAF list.
324 (see GC.c:revertCAFs()).
328 -------------------------------------------------------------------------- */
331 newCAF(StgClosure* caf)
338 // If we are in GHCi _and_ we are using dynamic libraries,
339 // then we can't redirect newCAF calls to newDynCAF (see below),
340 // so we make newCAF behave almost like newDynCAF.
341 // The dynamic libraries might be used by both the interpreted
342 // program and GHCi itself, so they must not be reverted.
343 // This also means that in GHCi with dynamic libraries, CAFs are not
344 // garbage collected. If this turns out to be a problem, we could
345 // do another hack here and do an address range test on caf to figure
346 // out whether it is from a dynamic library.
347 ((StgIndStatic *)caf)->saved_info = (StgInfoTable *)caf->header.info;
348 ((StgIndStatic *)caf)->static_link = caf_list;
353 /* Put this CAF on the mutable list for the old generation.
354 * This is a HACK - the IND_STATIC closure doesn't really have
355 * a mut_link field, but we pretend it has - in fact we re-use
356 * the STATIC_LINK field for the time being, because when we
357 * come to do a major GC we won't need the mut_link field
358 * any more and can use it as a STATIC_LINK.
360 ((StgIndStatic *)caf)->saved_info = NULL;
361 recordMutableGen(caf, oldest_gen->no);
367 // An alternate version of newCaf which is used for dynamically loaded
368 // object code in GHCi. In this case we want to retain *all* CAFs in
369 // the object code, because they might be demanded at any time from an
370 // expression evaluated on the command line.
371 // Also, GHCi might want to revert CAFs, so we add these to the
372 // revertible_caf_list.
374 // The linker hackily arranges that references to newCaf from dynamic
375 // code end up pointing to newDynCAF.
377 newDynCAF(StgClosure *caf)
381 ((StgIndStatic *)caf)->saved_info = (StgInfoTable *)caf->header.info;
382 ((StgIndStatic *)caf)->static_link = revertible_caf_list;
383 revertible_caf_list = caf;
388 /* -----------------------------------------------------------------------------
390 -------------------------------------------------------------------------- */
393 allocNursery (step *stp, bdescr *tail, nat blocks)
398 // Allocate a nursery: we allocate fresh blocks one at a time and
399 // cons them on to the front of the list, not forgetting to update
400 // the back pointer on the tail of the list to point to the new block.
401 for (i=0; i < blocks; i++) {
404 processNursery() in LdvProfile.c assumes that every block group in
405 the nursery contains only a single block. So, if a block group is
406 given multiple blocks, change processNursery() accordingly.
410 // double-link the nursery: we might need to insert blocks
417 bd->free = bd->start;
425 assignNurseriesToCapabilities (void)
430 for (i = 0; i < n_nurseries; i++) {
431 capabilities[i].r.rNursery = &nurseries[i];
432 capabilities[i].r.rCurrentNursery = nurseries[i].blocks;
433 capabilities[i].r.rCurrentAlloc = NULL;
435 #else /* THREADED_RTS */
436 MainCapability.r.rNursery = &nurseries[0];
437 MainCapability.r.rCurrentNursery = nurseries[0].blocks;
438 MainCapability.r.rCurrentAlloc = NULL;
443 allocNurseries( void )
447 for (i = 0; i < n_nurseries; i++) {
448 nurseries[i].blocks =
449 allocNursery(&nurseries[i], NULL,
450 RtsFlags.GcFlags.minAllocAreaSize);
451 nurseries[i].n_blocks = RtsFlags.GcFlags.minAllocAreaSize;
452 nurseries[i].old_blocks = NULL;
453 nurseries[i].n_old_blocks = 0;
455 assignNurseriesToCapabilities();
459 resetNurseries( void )
465 for (i = 0; i < n_nurseries; i++) {
467 for (bd = stp->blocks; bd; bd = bd->link) {
468 bd->free = bd->start;
469 ASSERT(bd->gen_no == 0);
470 ASSERT(bd->step == stp);
471 IF_DEBUG(sanity,memset(bd->start, 0xaa, BLOCK_SIZE));
474 assignNurseriesToCapabilities();
478 countNurseryBlocks (void)
483 for (i = 0; i < n_nurseries; i++) {
484 blocks += nurseries[i].n_blocks;
490 resizeNursery ( step *stp, nat blocks )
495 nursery_blocks = stp->n_blocks;
496 if (nursery_blocks == blocks) return;
498 if (nursery_blocks < blocks) {
499 debugTrace(DEBUG_gc, "increasing size of nursery to %d blocks",
501 stp->blocks = allocNursery(stp, stp->blocks, blocks-nursery_blocks);
506 debugTrace(DEBUG_gc, "decreasing size of nursery to %d blocks",
510 while (nursery_blocks > blocks) {
512 next_bd->u.back = NULL;
513 nursery_blocks -= bd->blocks; // might be a large block
518 // might have gone just under, by freeing a large block, so make
519 // up the difference.
520 if (nursery_blocks < blocks) {
521 stp->blocks = allocNursery(stp, stp->blocks, blocks-nursery_blocks);
525 stp->n_blocks = blocks;
526 ASSERT(countBlocks(stp->blocks) == stp->n_blocks);
530 // Resize each of the nurseries to the specified size.
533 resizeNurseriesFixed (nat blocks)
536 for (i = 0; i < n_nurseries; i++) {
537 resizeNursery(&nurseries[i], blocks);
542 // Resize the nurseries to the total specified size.
545 resizeNurseries (nat blocks)
547 // If there are multiple nurseries, then we just divide the number
548 // of available blocks between them.
549 resizeNurseriesFixed(blocks / n_nurseries);
553 /* -----------------------------------------------------------------------------
554 move_TSO is called to update the TSO structure after it has been
555 moved from one place to another.
556 -------------------------------------------------------------------------- */
559 move_TSO (StgTSO *src, StgTSO *dest)
563 // relocate the stack pointer...
564 diff = (StgPtr)dest - (StgPtr)src; // In *words*
565 dest->sp = (StgPtr)dest->sp + diff;
568 /* -----------------------------------------------------------------------------
569 The allocate() interface
571 allocateInGen() function allocates memory directly into a specific
572 generation. It always succeeds, and returns a chunk of memory n
573 words long. n can be larger than the size of a block if necessary,
574 in which case a contiguous block group will be allocated.
576 allocate(n) is equivalent to allocateInGen(g0).
577 -------------------------------------------------------------------------- */
580 allocateInGen (generation *g, lnat n)
588 TICK_ALLOC_HEAP_NOCTR(n);
593 if (n >= LARGE_OBJECT_THRESHOLD/sizeof(W_))
595 lnat req_blocks = (lnat)BLOCK_ROUND_UP(n*sizeof(W_)) / BLOCK_SIZE;
597 // Attempting to allocate an object larger than maxHeapSize
598 // should definitely be disallowed. (bug #1791)
599 if (RtsFlags.GcFlags.maxHeapSize > 0 &&
600 req_blocks >= RtsFlags.GcFlags.maxHeapSize) {
602 // heapOverflow() doesn't exit (see #2592), but we aren't
603 // in a position to do a clean shutdown here: we
604 // either have to allocate the memory or exit now.
605 // Allocating the memory would be bad, because the user
606 // has requested that we not exceed maxHeapSize, so we
608 stg_exit(EXIT_HEAPOVERFLOW);
611 bd = allocGroup(req_blocks);
612 dbl_link_onto(bd, &stp->large_objects);
613 stp->n_large_blocks += bd->blocks; // might be larger than req_blocks
614 alloc_blocks += bd->blocks;
617 bd->flags = BF_LARGE;
618 bd->free = bd->start + n;
623 // small allocation (<LARGE_OBJECT_THRESHOLD) */
625 if (bd == NULL || bd->free + n > bd->start + BLOCK_SIZE_W) {
630 bd->link = stp->blocks;
647 return allocateInGen(g0,n);
651 allocatedBytes( void )
655 allocated = alloc_blocks * BLOCK_SIZE_W;
656 if (pinned_object_block != NULL) {
657 allocated -= (pinned_object_block->start + BLOCK_SIZE_W) -
658 pinned_object_block->free;
664 // split N blocks off the front of the given bdescr, returning the
665 // new block group. We treat the remainder as if it
666 // had been freshly allocated in generation 0.
668 splitLargeBlock (bdescr *bd, nat blocks)
672 // subtract the original number of blocks from the counter first
673 bd->step->n_large_blocks -= bd->blocks;
675 new_bd = splitBlockGroup (bd, blocks);
677 dbl_link_onto(new_bd, &g0s0->large_objects);
678 g0s0->n_large_blocks += new_bd->blocks;
679 new_bd->gen_no = g0s0->no;
681 new_bd->flags = BF_LARGE;
682 new_bd->free = bd->free;
683 ASSERT(new_bd->free <= new_bd->start + new_bd->blocks * BLOCK_SIZE_W);
685 // add the new number of blocks to the counter. Due to the gaps
686 // for block descriptor, new_bd->blocks + bd->blocks might not be
687 // equal to the original bd->blocks, which is why we do it this way.
688 bd->step->n_large_blocks += bd->blocks;
693 /* -----------------------------------------------------------------------------
696 This allocates memory in the current thread - it is intended for
697 use primarily from STG-land where we have a Capability. It is
698 better than allocate() because it doesn't require taking the
699 sm_mutex lock in the common case.
701 Memory is allocated directly from the nursery if possible (but not
702 from the current nursery block, so as not to interfere with
704 -------------------------------------------------------------------------- */
707 allocateLocal (Capability *cap, lnat n)
712 if (n >= LARGE_OBJECT_THRESHOLD/sizeof(W_)) {
713 return allocateInGen(g0,n);
716 /* small allocation (<LARGE_OBJECT_THRESHOLD) */
718 TICK_ALLOC_HEAP_NOCTR(n);
721 bd = cap->r.rCurrentAlloc;
722 if (bd == NULL || bd->free + n > bd->start + BLOCK_SIZE_W) {
724 // The CurrentAlloc block is full, we need to find another
725 // one. First, we try taking the next block from the
727 bd = cap->r.rCurrentNursery->link;
729 if (bd == NULL || bd->free + n > bd->start + BLOCK_SIZE_W) {
730 // The nursery is empty, or the next block is already
731 // full: allocate a fresh block (we can't fail here).
734 cap->r.rNursery->n_blocks++;
737 bd->step = cap->r.rNursery;
739 // NO: alloc_blocks++;
740 // calcAllocated() uses the size of the nursery, and we've
741 // already bumpted nursery->n_blocks above. We'll GC
742 // pretty quickly now anyway, because MAYBE_GC() will
743 // notice that CurrentNursery->link is NULL.
745 // we have a block in the nursery: take it and put
746 // it at the *front* of the nursery list, and use it
747 // to allocate() from.
748 cap->r.rCurrentNursery->link = bd->link;
749 if (bd->link != NULL) {
750 bd->link->u.back = cap->r.rCurrentNursery;
753 dbl_link_onto(bd, &cap->r.rNursery->blocks);
754 cap->r.rCurrentAlloc = bd;
755 IF_DEBUG(sanity, checkNurserySanity(cap->r.rNursery));
762 /* ---------------------------------------------------------------------------
763 Allocate a fixed/pinned object.
765 We allocate small pinned objects into a single block, allocating a
766 new block when the current one overflows. The block is chained
767 onto the large_object_list of generation 0 step 0.
769 NOTE: The GC can't in general handle pinned objects. This
770 interface is only safe to use for ByteArrays, which have no
771 pointers and don't require scavenging. It works because the
772 block's descriptor has the BF_LARGE flag set, so the block is
773 treated as a large object and chained onto various lists, rather
774 than the individual objects being copied. However, when it comes
775 to scavenge the block, the GC will only scavenge the first object.
776 The reason is that the GC can't linearly scan a block of pinned
777 objects at the moment (doing so would require using the
778 mostly-copying techniques). But since we're restricting ourselves
779 to pinned ByteArrays, not scavenging is ok.
781 This function is called by newPinnedByteArray# which immediately
782 fills the allocated memory with a MutableByteArray#.
783 ------------------------------------------------------------------------- */
786 allocatePinned( lnat n )
789 bdescr *bd = pinned_object_block;
791 // If the request is for a large object, then allocate()
792 // will give us a pinned object anyway.
793 if (n >= LARGE_OBJECT_THRESHOLD/sizeof(W_)) {
795 Bdescr(p)->flags |= BF_PINNED;
801 TICK_ALLOC_HEAP_NOCTR(n);
804 // If we don't have a block of pinned objects yet, or the current
805 // one isn't large enough to hold the new object, allocate a new one.
806 if (bd == NULL || (bd->free + n) > (bd->start + BLOCK_SIZE_W)) {
807 pinned_object_block = bd = allocBlock();
808 dbl_link_onto(bd, &g0s0->large_objects);
809 g0s0->n_large_blocks++;
812 bd->flags = BF_PINNED | BF_LARGE;
813 bd->free = bd->start;
823 /* -----------------------------------------------------------------------------
825 -------------------------------------------------------------------------- */
828 This is the write barrier for MUT_VARs, a.k.a. IORefs. A
829 MUT_VAR_CLEAN object is not on the mutable list; a MUT_VAR_DIRTY
830 is. When written to, a MUT_VAR_CLEAN turns into a MUT_VAR_DIRTY
831 and is put on the mutable list.
834 dirty_MUT_VAR(StgRegTable *reg, StgClosure *p)
836 Capability *cap = regTableToCapability(reg);
838 if (p->header.info == &stg_MUT_VAR_CLEAN_info) {
839 p->header.info = &stg_MUT_VAR_DIRTY_info;
840 bd = Bdescr((StgPtr)p);
841 if (bd->gen_no > 0) recordMutableCap(p,cap,bd->gen_no);
845 // Setting a TSO's link field with a write barrier.
846 // It is *not* necessary to call this function when
847 // * setting the link field to END_TSO_QUEUE
848 // * putting a TSO on the blackhole_queue
849 // * setting the link field of the currently running TSO, as it
850 // will already be dirty.
852 setTSOLink (Capability *cap, StgTSO *tso, StgTSO *target)
855 if ((tso->flags & (TSO_DIRTY|TSO_LINK_DIRTY)) == 0) {
856 tso->flags |= TSO_LINK_DIRTY;
857 bd = Bdescr((StgPtr)tso);
858 if (bd->gen_no > 0) recordMutableCap((StgClosure*)tso,cap,bd->gen_no);
864 dirty_TSO (Capability *cap, StgTSO *tso)
867 if ((tso->flags & (TSO_DIRTY|TSO_LINK_DIRTY)) == 0) {
868 bd = Bdescr((StgPtr)tso);
869 if (bd->gen_no > 0) recordMutableCap((StgClosure*)tso,cap,bd->gen_no);
871 tso->flags |= TSO_DIRTY;
875 This is the write barrier for MVARs. An MVAR_CLEAN objects is not
876 on the mutable list; a MVAR_DIRTY is. When written to, a
877 MVAR_CLEAN turns into a MVAR_DIRTY and is put on the mutable list.
878 The check for MVAR_CLEAN is inlined at the call site for speed,
879 this really does make a difference on concurrency-heavy benchmarks
880 such as Chaneneos and cheap-concurrency.
883 dirty_MVAR(StgRegTable *reg, StgClosure *p)
885 Capability *cap = regTableToCapability(reg);
887 bd = Bdescr((StgPtr)p);
888 if (bd->gen_no > 0) recordMutableCap(p,cap,bd->gen_no);
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 ASSERT(bd->free <= bd->start + bd->blocks * BLOCK_SIZE_W);
983 words += bd->free - bd->start;
988 // Return an accurate count of the live data in the heap, excluding
997 if (RtsFlags.GcFlags.generations == 1) {
998 return g0s0->n_words + countOccupied(g0s0->large_objects);
1002 for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
1003 for (s = 0; s < generations[g].n_steps; s++) {
1004 if (g == 0 && s == 0) continue;
1005 stp = &generations[g].steps[s];
1006 live += stp->n_words + 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];
1031 // we need at least this much space
1032 needed += stp->n_blocks + stp->n_large_blocks;
1034 // any additional space needed to collect this gen next time?
1035 if (g == 0 || // always collect gen 0
1036 (generations[g].steps[0].n_blocks +
1037 generations[g].steps[0].n_large_blocks
1038 > generations[g].max_blocks)) {
1039 // we will collect this gen next time
1042 needed += stp->n_blocks / BITS_IN(W_);
1044 needed += stp->n_blocks / 100;
1047 continue; // no additional space needed for compaction
1049 needed += stp->n_blocks;
1057 /* ----------------------------------------------------------------------------
1060 Executable memory must be managed separately from non-executable
1061 memory. Most OSs these days require you to jump through hoops to
1062 dynamically allocate executable memory, due to various security
1065 Here we provide a small memory allocator for executable memory.
1066 Memory is managed with a page granularity; we allocate linearly
1067 in the page, and when the page is emptied (all objects on the page
1068 are free) we free the page again, not forgetting to make it
1071 TODO: The inability to handle objects bigger than BLOCK_SIZE_W means that
1072 the linker cannot use allocateExec for loading object code files
1073 on Windows. Once allocateExec can handle larger objects, the linker
1074 should be modified to use allocateExec instead of VirtualAlloc.
1075 ------------------------------------------------------------------------- */
1077 #if defined(linux_HOST_OS)
1079 // On Linux we need to use libffi for allocating executable memory,
1080 // because it knows how to work around the restrictions put in place
1083 void *allocateExec (nat bytes, void **exec_ret)
1087 ret = ffi_closure_alloc (sizeof(void *) + (size_t)bytes, (void**)&exec);
1089 if (ret == NULL) return ret;
1090 *ret = ret; // save the address of the writable mapping, for freeExec().
1091 *exec_ret = exec + 1;
1095 // freeExec gets passed the executable address, not the writable address.
1096 void freeExec (void *addr)
1099 writable = *((void**)addr - 1);
1101 ffi_closure_free (writable);
1107 void *allocateExec (nat bytes, void **exec_ret)
1114 // round up to words.
1115 n = (bytes + sizeof(W_) + 1) / sizeof(W_);
1117 if (n+1 > BLOCK_SIZE_W) {
1118 barf("allocateExec: can't handle large objects");
1121 if (exec_block == NULL ||
1122 exec_block->free + n + 1 > exec_block->start + BLOCK_SIZE_W) {
1124 lnat pagesize = getPageSize();
1125 bd = allocGroup(stg_max(1, pagesize / BLOCK_SIZE));
1126 debugTrace(DEBUG_gc, "allocate exec block %p", bd->start);
1128 bd->flags = BF_EXEC;
1129 bd->link = exec_block;
1130 if (exec_block != NULL) {
1131 exec_block->u.back = bd;
1134 setExecutable(bd->start, bd->blocks * BLOCK_SIZE, rtsTrue);
1137 *(exec_block->free) = n; // store the size of this chunk
1138 exec_block->gen_no += n; // gen_no stores the number of words allocated
1139 ret = exec_block->free + 1;
1140 exec_block->free += n + 1;
1147 void freeExec (void *addr)
1149 StgPtr p = (StgPtr)addr - 1;
1150 bdescr *bd = Bdescr((StgPtr)p);
1152 if ((bd->flags & BF_EXEC) == 0) {
1153 barf("freeExec: not executable");
1156 if (*(StgPtr)p == 0) {
1157 barf("freeExec: already free?");
1162 bd->gen_no -= *(StgPtr)p;
1165 if (bd->gen_no == 0) {
1166 // Free the block if it is empty, but not if it is the block at
1167 // the head of the queue.
1168 if (bd != exec_block) {
1169 debugTrace(DEBUG_gc, "free exec block %p", bd->start);
1170 dbl_link_remove(bd, &exec_block);
1171 setExecutable(bd->start, bd->blocks * BLOCK_SIZE, rtsFalse);
1174 bd->free = bd->start;
1181 #endif /* mingw32_HOST_OS */
1183 /* -----------------------------------------------------------------------------
1186 memInventory() checks for memory leaks by counting up all the
1187 blocks we know about and comparing that to the number of blocks
1188 allegedly floating around in the system.
1189 -------------------------------------------------------------------------- */
1193 // Useful for finding partially full blocks in gdb
1194 void findSlop(bdescr *bd);
1195 void findSlop(bdescr *bd)
1199 for (; bd != NULL; bd = bd->link) {
1200 slop = (bd->blocks * BLOCK_SIZE_W) - (bd->free - bd->start);
1201 if (slop > (1024/sizeof(W_))) {
1202 debugBelch("block at %p (bdescr %p) has %ldKB slop\n",
1203 bd->start, bd, slop / (1024/sizeof(W_)));
1209 countBlocks(bdescr *bd)
1212 for (n=0; bd != NULL; bd=bd->link) {
1218 // (*1) Just like countBlocks, except that we adjust the count for a
1219 // megablock group so that it doesn't include the extra few blocks
1220 // that would be taken up by block descriptors in the second and
1221 // subsequent megablock. This is so we can tally the count with the
1222 // number of blocks allocated in the system, for memInventory().
1224 countAllocdBlocks(bdescr *bd)
1227 for (n=0; bd != NULL; bd=bd->link) {
1229 // hack for megablock groups: see (*1) above
1230 if (bd->blocks > BLOCKS_PER_MBLOCK) {
1231 n -= (MBLOCK_SIZE / BLOCK_SIZE - BLOCKS_PER_MBLOCK)
1232 * (bd->blocks/(MBLOCK_SIZE/BLOCK_SIZE));
1239 stepBlocks (step *stp)
1241 ASSERT(countBlocks(stp->blocks) == stp->n_blocks);
1242 ASSERT(countBlocks(stp->large_objects) == stp->n_large_blocks);
1243 return stp->n_blocks + stp->n_old_blocks +
1244 countAllocdBlocks(stp->large_objects);
1247 // If memInventory() calculates that we have a memory leak, this
1248 // function will try to find the block(s) that are leaking by marking
1249 // all the ones that we know about, and search through memory to find
1250 // blocks that are not marked. In the debugger this can help to give
1251 // us a clue about what kind of block leaked. In the future we might
1252 // annotate blocks with their allocation site to give more helpful
1255 findMemoryLeak (void)
1258 for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
1259 for (i = 0; i < n_capabilities; i++) {
1260 markBlocks(capabilities[i].mut_lists[g]);
1262 markBlocks(generations[g].mut_list);
1263 for (s = 0; s < generations[g].n_steps; s++) {
1264 markBlocks(generations[g].steps[s].blocks);
1265 markBlocks(generations[g].steps[s].large_objects);
1269 for (i = 0; i < n_nurseries; i++) {
1270 markBlocks(nurseries[i].blocks);
1271 markBlocks(nurseries[i].large_objects);
1276 // if (RtsFlags.ProfFlags.doHeapProfile == HEAP_BY_RETAINER) {
1277 // markRetainerBlocks();
1281 // count the blocks allocated by the arena allocator
1283 // markArenaBlocks();
1285 // count the blocks containing executable memory
1286 markBlocks(exec_block);
1288 reportUnmarkedBlocks();
1293 memInventory (rtsBool show)
1297 lnat gen_blocks[RtsFlags.GcFlags.generations];
1298 lnat nursery_blocks, retainer_blocks,
1299 arena_blocks, exec_blocks;
1300 lnat live_blocks = 0, free_blocks = 0;
1303 // count the blocks we current have
1305 for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
1307 for (i = 0; i < n_capabilities; i++) {
1308 gen_blocks[g] += countBlocks(capabilities[i].mut_lists[g]);
1310 gen_blocks[g] += countAllocdBlocks(generations[g].mut_list);
1311 for (s = 0; s < generations[g].n_steps; s++) {
1312 stp = &generations[g].steps[s];
1313 gen_blocks[g] += stepBlocks(stp);
1318 for (i = 0; i < n_nurseries; i++) {
1319 nursery_blocks += stepBlocks(&nurseries[i]);
1322 retainer_blocks = 0;
1324 if (RtsFlags.ProfFlags.doHeapProfile == HEAP_BY_RETAINER) {
1325 retainer_blocks = retainerStackBlocks();
1329 // count the blocks allocated by the arena allocator
1330 arena_blocks = arenaBlocks();
1332 // count the blocks containing executable memory
1333 exec_blocks = countAllocdBlocks(exec_block);
1335 /* count the blocks on the free list */
1336 free_blocks = countFreeList();
1339 for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
1340 live_blocks += gen_blocks[g];
1342 live_blocks += nursery_blocks +
1343 + retainer_blocks + arena_blocks + exec_blocks;
1345 #define MB(n) (((n) * BLOCK_SIZE_W) / ((1024*1024)/sizeof(W_)))
1347 leak = live_blocks + free_blocks != mblocks_allocated * BLOCKS_PER_MBLOCK;
1352 debugBelch("Memory leak detected:\n");
1354 debugBelch("Memory inventory:\n");
1356 for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
1357 debugBelch(" gen %d blocks : %5lu blocks (%lu MB)\n", g,
1358 gen_blocks[g], MB(gen_blocks[g]));
1360 debugBelch(" nursery : %5lu blocks (%lu MB)\n",
1361 nursery_blocks, MB(nursery_blocks));
1362 debugBelch(" retainer : %5lu blocks (%lu MB)\n",
1363 retainer_blocks, MB(retainer_blocks));
1364 debugBelch(" arena blocks : %5lu blocks (%lu MB)\n",
1365 arena_blocks, MB(arena_blocks));
1366 debugBelch(" exec : %5lu blocks (%lu MB)\n",
1367 exec_blocks, MB(exec_blocks));
1368 debugBelch(" free : %5lu blocks (%lu MB)\n",
1369 free_blocks, MB(free_blocks));
1370 debugBelch(" total : %5lu blocks (%lu MB)\n",
1371 live_blocks + free_blocks, MB(live_blocks+free_blocks));
1373 debugBelch("\n in system : %5lu blocks (%lu MB)\n",
1374 mblocks_allocated * BLOCKS_PER_MBLOCK, mblocks_allocated);
1382 ASSERT(n_alloc_blocks == live_blocks);
1387 /* Full heap sanity check. */
1393 if (RtsFlags.GcFlags.generations == 1) {
1394 checkHeap(g0s0->blocks);
1395 checkLargeObjects(g0s0->large_objects);
1398 for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
1399 for (s = 0; s < generations[g].n_steps; s++) {
1400 if (g == 0 && s == 0) { continue; }
1401 ASSERT(countBlocks(generations[g].steps[s].blocks)
1402 == generations[g].steps[s].n_blocks);
1403 ASSERT(countBlocks(generations[g].steps[s].large_objects)
1404 == generations[g].steps[s].n_large_blocks);
1405 checkHeap(generations[g].steps[s].blocks);
1406 checkLargeObjects(generations[g].steps[s].large_objects);
1410 for (s = 0; s < n_nurseries; s++) {
1411 ASSERT(countBlocks(nurseries[s].blocks)
1412 == nurseries[s].n_blocks);
1413 ASSERT(countBlocks(nurseries[s].large_objects)
1414 == nurseries[s].n_large_blocks);
1417 checkFreeListSanity();
1420 #if defined(THREADED_RTS)
1421 // check the stacks too in threaded mode, because we don't do a
1422 // full heap sanity check in this case (see checkHeap())
1423 checkMutableLists(rtsTrue);
1425 checkMutableLists(rtsFalse);
1429 /* Nursery sanity check */
1431 checkNurserySanity( step *stp )
1437 for (bd = stp->blocks; bd != NULL; bd = bd->link) {
1438 ASSERT(bd->u.back == prev);
1440 blocks += bd->blocks;
1442 ASSERT(blocks == stp->n_blocks);
1445 // handy function for use in gdb, because Bdescr() is inlined.
1446 extern bdescr *_bdescr( StgPtr p );