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.
73 * This mutex is used by atomicModifyMutVar# only
75 Mutex atomic_modify_mutvar_mutex;
82 static void *stgAllocForGMP (size_t size_in_bytes);
83 static void *stgReallocForGMP (void *ptr, size_t old_size, size_t new_size);
84 static void stgDeallocForGMP (void *ptr, size_t size);
87 initStep (step *stp, int g, int s)
90 stp->abs_no = RtsFlags.GcFlags.steps * g + s;
94 stp->live_estimate = 0;
95 stp->old_blocks = NULL;
96 stp->n_old_blocks = 0;
97 stp->gen = &generations[g];
99 stp->large_objects = NULL;
100 stp->n_large_blocks = 0;
101 stp->scavenged_large_objects = NULL;
102 stp->n_scavenged_large_blocks = 0;
107 initSpinLock(&stp->sync_todo);
108 initSpinLock(&stp->sync_large_objects);
110 stp->threads = END_TSO_QUEUE;
111 stp->old_threads = END_TSO_QUEUE;
120 if (generations != NULL) {
121 // multi-init protection
127 /* Sanity check to make sure the LOOKS_LIKE_ macros appear to be
128 * doing something reasonable.
130 /* We use the NOT_NULL variant or gcc warns that the test is always true */
131 ASSERT(LOOKS_LIKE_INFO_PTR_NOT_NULL((StgWord)&stg_BLACKHOLE_info));
132 ASSERT(LOOKS_LIKE_CLOSURE_PTR(&stg_dummy_ret_closure));
133 ASSERT(!HEAP_ALLOCED(&stg_dummy_ret_closure));
135 if (RtsFlags.GcFlags.maxHeapSize != 0 &&
136 RtsFlags.GcFlags.heapSizeSuggestion >
137 RtsFlags.GcFlags.maxHeapSize) {
138 RtsFlags.GcFlags.maxHeapSize = RtsFlags.GcFlags.heapSizeSuggestion;
141 if (RtsFlags.GcFlags.maxHeapSize != 0 &&
142 RtsFlags.GcFlags.minAllocAreaSize >
143 RtsFlags.GcFlags.maxHeapSize) {
144 errorBelch("maximum heap size (-M) is smaller than minimum alloc area size (-A)");
145 RtsFlags.GcFlags.minAllocAreaSize = RtsFlags.GcFlags.maxHeapSize;
148 initBlockAllocator();
150 #if defined(THREADED_RTS)
151 initMutex(&sm_mutex);
152 initMutex(&atomic_modify_mutvar_mutex);
157 /* allocate generation info array */
158 generations = (generation *)stgMallocBytes(RtsFlags.GcFlags.generations
159 * sizeof(struct generation_),
160 "initStorage: gens");
162 /* allocate all the steps into an array. It is important that we do
163 it this way, because we need the invariant that two step pointers
164 can be directly compared to see which is the oldest.
165 Remember that the last generation has only one step. */
166 total_steps = 1 + (RtsFlags.GcFlags.generations - 1) * RtsFlags.GcFlags.steps;
167 all_steps = stgMallocBytes(total_steps * sizeof(struct step_),
168 "initStorage: steps");
170 /* Initialise all generations */
171 for(g = 0; g < RtsFlags.GcFlags.generations; g++) {
172 gen = &generations[g];
174 gen->mut_list = allocBlock();
175 gen->collections = 0;
176 gen->par_collections = 0;
177 gen->failed_promotions = 0;
181 /* A couple of convenience pointers */
182 g0 = &generations[0];
183 oldest_gen = &generations[RtsFlags.GcFlags.generations-1];
185 /* Allocate step structures in each generation */
186 if (RtsFlags.GcFlags.generations > 1) {
187 /* Only for multiple-generations */
189 /* Oldest generation: one step */
190 oldest_gen->n_steps = 1;
191 oldest_gen->steps = all_steps + (RtsFlags.GcFlags.generations - 1)
192 * RtsFlags.GcFlags.steps;
194 /* set up all except the oldest generation with 2 steps */
195 for(g = 0; g < RtsFlags.GcFlags.generations-1; g++) {
196 generations[g].n_steps = RtsFlags.GcFlags.steps;
197 generations[g].steps = all_steps + g * RtsFlags.GcFlags.steps;
201 /* single generation, i.e. a two-space collector */
203 g0->steps = all_steps;
207 n_nurseries = n_capabilities;
211 nurseries = stgMallocBytes (n_nurseries * sizeof(struct step_),
212 "initStorage: nurseries");
214 /* Initialise all steps */
215 for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
216 for (s = 0; s < generations[g].n_steps; s++) {
217 initStep(&generations[g].steps[s], g, s);
221 for (s = 0; s < n_nurseries; s++) {
222 initStep(&nurseries[s], 0, s);
225 /* Set up the destination pointers in each younger gen. step */
226 for (g = 0; g < RtsFlags.GcFlags.generations-1; g++) {
227 for (s = 0; s < generations[g].n_steps-1; s++) {
228 generations[g].steps[s].to = &generations[g].steps[s+1];
230 generations[g].steps[s].to = &generations[g+1].steps[0];
232 oldest_gen->steps[0].to = &oldest_gen->steps[0];
234 for (s = 0; s < n_nurseries; s++) {
235 nurseries[s].to = generations[0].steps[0].to;
238 /* The oldest generation has one step. */
239 if (RtsFlags.GcFlags.compact || RtsFlags.GcFlags.sweep) {
240 if (RtsFlags.GcFlags.generations == 1) {
241 errorBelch("WARNING: compact/sweep is incompatible with -G1; disabled");
243 oldest_gen->steps[0].mark = 1;
244 if (RtsFlags.GcFlags.compact)
245 oldest_gen->steps[0].compact = 1;
249 generations[0].max_blocks = 0;
250 g0s0 = &generations[0].steps[0];
252 /* The allocation area. Policy: keep the allocation area
253 * small to begin with, even if we have a large suggested heap
254 * size. Reason: we're going to do a major collection first, and we
255 * don't want it to be a big one. This vague idea is borne out by
256 * rigorous experimental evidence.
260 weak_ptr_list = NULL;
262 revertible_caf_list = NULL;
264 /* initialise the allocate() interface */
266 alloc_blocks_lim = RtsFlags.GcFlags.minAllocAreaSize;
270 /* Tell GNU multi-precision pkg about our custom alloc functions */
271 mp_set_memory_functions(stgAllocForGMP, stgReallocForGMP, stgDeallocForGMP);
274 initSpinLock(&gc_alloc_block_sync);
275 initSpinLock(&recordMutableGen_sync);
279 IF_DEBUG(gc, statDescribeGens());
287 stat_exit(calcAllocated());
293 stgFree(g0s0); // frees all the steps
294 stgFree(generations);
296 #if defined(THREADED_RTS)
297 closeMutex(&sm_mutex);
298 closeMutex(&atomic_modify_mutvar_mutex);
303 /* -----------------------------------------------------------------------------
306 The entry code for every CAF does the following:
308 - builds a CAF_BLACKHOLE in the heap
309 - pushes an update frame pointing to the CAF_BLACKHOLE
310 - invokes UPD_CAF(), which:
311 - calls newCaf, below
312 - updates the CAF with a static indirection to the CAF_BLACKHOLE
314 Why do we build a BLACKHOLE in the heap rather than just updating
315 the thunk directly? It's so that we only need one kind of update
316 frame - otherwise we'd need a static version of the update frame too.
318 newCaf() does the following:
320 - it puts the CAF on the oldest generation's mut-once list.
321 This is so that we can treat the CAF as a root when collecting
324 For GHCI, we have additional requirements when dealing with CAFs:
326 - we must *retain* all dynamically-loaded CAFs ever entered,
327 just in case we need them again.
328 - we must be able to *revert* CAFs that have been evaluated, to
329 their pre-evaluated form.
331 To do this, we use an additional CAF list. When newCaf() is
332 called on a dynamically-loaded CAF, we add it to the CAF list
333 instead of the old-generation mutable list, and save away its
334 old info pointer (in caf->saved_info) for later reversion.
336 To revert all the CAFs, we traverse the CAF list and reset the
337 info pointer to caf->saved_info, then throw away the CAF list.
338 (see GC.c:revertCAFs()).
342 -------------------------------------------------------------------------- */
345 newCAF(StgClosure* caf)
352 // If we are in GHCi _and_ we are using dynamic libraries,
353 // then we can't redirect newCAF calls to newDynCAF (see below),
354 // so we make newCAF behave almost like newDynCAF.
355 // The dynamic libraries might be used by both the interpreted
356 // program and GHCi itself, so they must not be reverted.
357 // This also means that in GHCi with dynamic libraries, CAFs are not
358 // garbage collected. If this turns out to be a problem, we could
359 // do another hack here and do an address range test on caf to figure
360 // out whether it is from a dynamic library.
361 ((StgIndStatic *)caf)->saved_info = (StgInfoTable *)caf->header.info;
362 ((StgIndStatic *)caf)->static_link = caf_list;
367 /* Put this CAF on the mutable list for the old generation.
368 * This is a HACK - the IND_STATIC closure doesn't really have
369 * a mut_link field, but we pretend it has - in fact we re-use
370 * the STATIC_LINK field for the time being, because when we
371 * come to do a major GC we won't need the mut_link field
372 * any more and can use it as a STATIC_LINK.
374 ((StgIndStatic *)caf)->saved_info = NULL;
375 recordMutableGen(caf, oldest_gen);
381 // An alternate version of newCaf which is used for dynamically loaded
382 // object code in GHCi. In this case we want to retain *all* CAFs in
383 // the object code, because they might be demanded at any time from an
384 // expression evaluated on the command line.
385 // Also, GHCi might want to revert CAFs, so we add these to the
386 // revertible_caf_list.
388 // The linker hackily arranges that references to newCaf from dynamic
389 // code end up pointing to newDynCAF.
391 newDynCAF(StgClosure *caf)
395 ((StgIndStatic *)caf)->saved_info = (StgInfoTable *)caf->header.info;
396 ((StgIndStatic *)caf)->static_link = revertible_caf_list;
397 revertible_caf_list = caf;
402 /* -----------------------------------------------------------------------------
404 -------------------------------------------------------------------------- */
407 allocNursery (step *stp, bdescr *tail, nat blocks)
412 // Allocate a nursery: we allocate fresh blocks one at a time and
413 // cons them on to the front of the list, not forgetting to update
414 // the back pointer on the tail of the list to point to the new block.
415 for (i=0; i < blocks; i++) {
418 processNursery() in LdvProfile.c assumes that every block group in
419 the nursery contains only a single block. So, if a block group is
420 given multiple blocks, change processNursery() accordingly.
424 // double-link the nursery: we might need to insert blocks
431 bd->free = bd->start;
439 assignNurseriesToCapabilities (void)
444 for (i = 0; i < n_nurseries; i++) {
445 capabilities[i].r.rNursery = &nurseries[i];
446 capabilities[i].r.rCurrentNursery = nurseries[i].blocks;
447 capabilities[i].r.rCurrentAlloc = NULL;
449 #else /* THREADED_RTS */
450 MainCapability.r.rNursery = &nurseries[0];
451 MainCapability.r.rCurrentNursery = nurseries[0].blocks;
452 MainCapability.r.rCurrentAlloc = NULL;
457 allocNurseries( void )
461 for (i = 0; i < n_nurseries; i++) {
462 nurseries[i].blocks =
463 allocNursery(&nurseries[i], NULL,
464 RtsFlags.GcFlags.minAllocAreaSize);
465 nurseries[i].n_blocks = RtsFlags.GcFlags.minAllocAreaSize;
466 nurseries[i].old_blocks = NULL;
467 nurseries[i].n_old_blocks = 0;
469 assignNurseriesToCapabilities();
473 resetNurseries( void )
479 for (i = 0; i < n_nurseries; i++) {
481 for (bd = stp->blocks; bd; bd = bd->link) {
482 bd->free = bd->start;
483 ASSERT(bd->gen_no == 0);
484 ASSERT(bd->step == stp);
485 IF_DEBUG(sanity,memset(bd->start, 0xaa, BLOCK_SIZE));
488 assignNurseriesToCapabilities();
492 countNurseryBlocks (void)
497 for (i = 0; i < n_nurseries; i++) {
498 blocks += nurseries[i].n_blocks;
504 resizeNursery ( step *stp, nat blocks )
509 nursery_blocks = stp->n_blocks;
510 if (nursery_blocks == blocks) return;
512 if (nursery_blocks < blocks) {
513 debugTrace(DEBUG_gc, "increasing size of nursery to %d blocks",
515 stp->blocks = allocNursery(stp, stp->blocks, blocks-nursery_blocks);
520 debugTrace(DEBUG_gc, "decreasing size of nursery to %d blocks",
524 while (nursery_blocks > blocks) {
526 next_bd->u.back = NULL;
527 nursery_blocks -= bd->blocks; // might be a large block
532 // might have gone just under, by freeing a large block, so make
533 // up the difference.
534 if (nursery_blocks < blocks) {
535 stp->blocks = allocNursery(stp, stp->blocks, blocks-nursery_blocks);
539 stp->n_blocks = blocks;
540 ASSERT(countBlocks(stp->blocks) == stp->n_blocks);
544 // Resize each of the nurseries to the specified size.
547 resizeNurseriesFixed (nat blocks)
550 for (i = 0; i < n_nurseries; i++) {
551 resizeNursery(&nurseries[i], blocks);
556 // Resize the nurseries to the total specified size.
559 resizeNurseries (nat blocks)
561 // If there are multiple nurseries, then we just divide the number
562 // of available blocks between them.
563 resizeNurseriesFixed(blocks / n_nurseries);
567 /* -----------------------------------------------------------------------------
568 move_TSO is called to update the TSO structure after it has been
569 moved from one place to another.
570 -------------------------------------------------------------------------- */
573 move_TSO (StgTSO *src, StgTSO *dest)
577 // relocate the stack pointer...
578 diff = (StgPtr)dest - (StgPtr)src; // In *words*
579 dest->sp = (StgPtr)dest->sp + diff;
582 /* -----------------------------------------------------------------------------
583 The allocate() interface
585 allocateInGen() function allocates memory directly into a specific
586 generation. It always succeeds, and returns a chunk of memory n
587 words long. n can be larger than the size of a block if necessary,
588 in which case a contiguous block group will be allocated.
590 allocate(n) is equivalent to allocateInGen(g0).
591 -------------------------------------------------------------------------- */
594 allocateInGen (generation *g, lnat n)
602 TICK_ALLOC_HEAP_NOCTR(n);
607 if (n >= LARGE_OBJECT_THRESHOLD/sizeof(W_))
609 lnat req_blocks = (lnat)BLOCK_ROUND_UP(n*sizeof(W_)) / BLOCK_SIZE;
611 // Attempting to allocate an object larger than maxHeapSize
612 // should definitely be disallowed. (bug #1791)
613 if (RtsFlags.GcFlags.maxHeapSize > 0 &&
614 req_blocks >= RtsFlags.GcFlags.maxHeapSize) {
618 bd = allocGroup(req_blocks);
619 dbl_link_onto(bd, &stp->large_objects);
620 stp->n_large_blocks += bd->blocks; // might be larger than req_blocks
621 alloc_blocks += bd->blocks;
624 bd->flags = BF_LARGE;
625 bd->free = bd->start + n;
630 // small allocation (<LARGE_OBJECT_THRESHOLD) */
632 if (bd == NULL || bd->free + n > bd->start + BLOCK_SIZE_W) {
637 bd->link = stp->blocks;
654 return allocateInGen(g0,n);
658 allocatedBytes( void )
662 allocated = alloc_blocks * BLOCK_SIZE_W;
663 if (pinned_object_block != NULL) {
664 allocated -= (pinned_object_block->start + BLOCK_SIZE_W) -
665 pinned_object_block->free;
671 // split N blocks off the front of the given bdescr, returning the
672 // new block group. We treat the remainder as if it
673 // had been freshly allocated in generation 0.
675 splitLargeBlock (bdescr *bd, nat blocks)
679 // subtract the original number of blocks from the counter first
680 bd->step->n_large_blocks -= bd->blocks;
682 new_bd = splitBlockGroup (bd, blocks);
684 dbl_link_onto(new_bd, &g0s0->large_objects);
685 g0s0->n_large_blocks += new_bd->blocks;
686 new_bd->gen_no = g0s0->no;
688 new_bd->flags = BF_LARGE;
689 new_bd->free = bd->free;
690 ASSERT(new_bd->free <= new_bd->start + new_bd->blocks * BLOCK_SIZE_W);
692 // add the new number of blocks to the counter. Due to the gaps
693 // for block descriptor, new_bd->blocks + bd->blocks might not be
694 // equal to the original bd->blocks, which is why we do it this way.
695 bd->step->n_large_blocks += bd->blocks;
700 /* -----------------------------------------------------------------------------
703 This allocates memory in the current thread - it is intended for
704 use primarily from STG-land where we have a Capability. It is
705 better than allocate() because it doesn't require taking the
706 sm_mutex lock in the common case.
708 Memory is allocated directly from the nursery if possible (but not
709 from the current nursery block, so as not to interfere with
711 -------------------------------------------------------------------------- */
714 allocateLocal (Capability *cap, lnat n)
719 if (n >= LARGE_OBJECT_THRESHOLD/sizeof(W_)) {
720 return allocateInGen(g0,n);
723 /* small allocation (<LARGE_OBJECT_THRESHOLD) */
725 TICK_ALLOC_HEAP_NOCTR(n);
728 bd = cap->r.rCurrentAlloc;
729 if (bd == NULL || bd->free + n > bd->start + BLOCK_SIZE_W) {
731 // The CurrentAlloc block is full, we need to find another
732 // one. First, we try taking the next block from the
734 bd = cap->r.rCurrentNursery->link;
736 if (bd == NULL || bd->free + n > bd->start + BLOCK_SIZE_W) {
737 // The nursery is empty, or the next block is already
738 // full: allocate a fresh block (we can't fail here).
741 cap->r.rNursery->n_blocks++;
744 bd->step = cap->r.rNursery;
746 // NO: alloc_blocks++;
747 // calcAllocated() uses the size of the nursery, and we've
748 // already bumpted nursery->n_blocks above. We'll GC
749 // pretty quickly now anyway, because MAYBE_GC() will
750 // notice that CurrentNursery->link is NULL.
752 // we have a block in the nursery: take it and put
753 // it at the *front* of the nursery list, and use it
754 // to allocate() from.
755 cap->r.rCurrentNursery->link = bd->link;
756 if (bd->link != NULL) {
757 bd->link->u.back = cap->r.rCurrentNursery;
760 dbl_link_onto(bd, &cap->r.rNursery->blocks);
761 cap->r.rCurrentAlloc = bd;
762 IF_DEBUG(sanity, checkNurserySanity(cap->r.rNursery));
769 /* ---------------------------------------------------------------------------
770 Allocate a fixed/pinned object.
772 We allocate small pinned objects into a single block, allocating a
773 new block when the current one overflows. The block is chained
774 onto the large_object_list of generation 0 step 0.
776 NOTE: The GC can't in general handle pinned objects. This
777 interface is only safe to use for ByteArrays, which have no
778 pointers and don't require scavenging. It works because the
779 block's descriptor has the BF_LARGE flag set, so the block is
780 treated as a large object and chained onto various lists, rather
781 than the individual objects being copied. However, when it comes
782 to scavenge the block, the GC will only scavenge the first object.
783 The reason is that the GC can't linearly scan a block of pinned
784 objects at the moment (doing so would require using the
785 mostly-copying techniques). But since we're restricting ourselves
786 to pinned ByteArrays, not scavenging is ok.
788 This function is called by newPinnedByteArray# which immediately
789 fills the allocated memory with a MutableByteArray#.
790 ------------------------------------------------------------------------- */
793 allocatePinned( lnat n )
796 bdescr *bd = pinned_object_block;
798 // If the request is for a large object, then allocate()
799 // will give us a pinned object anyway.
800 if (n >= LARGE_OBJECT_THRESHOLD/sizeof(W_)) {
806 TICK_ALLOC_HEAP_NOCTR(n);
809 // we always return 8-byte aligned memory. bd->free must be
810 // 8-byte aligned to begin with, so we just round up n to
811 // the nearest multiple of 8 bytes.
812 if (sizeof(StgWord) == 4) {
816 // If we don't have a block of pinned objects yet, or the current
817 // one isn't large enough to hold the new object, allocate a new one.
818 if (bd == NULL || (bd->free + n) > (bd->start + BLOCK_SIZE_W)) {
819 pinned_object_block = bd = allocBlock();
820 dbl_link_onto(bd, &g0s0->large_objects);
821 g0s0->n_large_blocks++;
824 bd->flags = BF_PINNED | BF_LARGE;
825 bd->free = bd->start;
835 /* -----------------------------------------------------------------------------
837 -------------------------------------------------------------------------- */
840 This is the write barrier for MUT_VARs, a.k.a. IORefs. A
841 MUT_VAR_CLEAN object is not on the mutable list; a MUT_VAR_DIRTY
842 is. When written to, a MUT_VAR_CLEAN turns into a MUT_VAR_DIRTY
843 and is put on the mutable list.
846 dirty_MUT_VAR(StgRegTable *reg, StgClosure *p)
848 Capability *cap = regTableToCapability(reg);
850 if (p->header.info == &stg_MUT_VAR_CLEAN_info) {
851 p->header.info = &stg_MUT_VAR_DIRTY_info;
852 bd = Bdescr((StgPtr)p);
853 if (bd->gen_no > 0) recordMutableCap(p,cap,bd->gen_no);
857 // Setting a TSO's link field with a write barrier.
858 // It is *not* necessary to call this function when
859 // * setting the link field to END_TSO_QUEUE
860 // * putting a TSO on the blackhole_queue
861 // * setting the link field of the currently running TSO, as it
862 // will already be dirty.
864 setTSOLink (Capability *cap, StgTSO *tso, StgTSO *target)
867 if ((tso->flags & (TSO_DIRTY|TSO_LINK_DIRTY)) == 0) {
868 tso->flags |= TSO_LINK_DIRTY;
869 bd = Bdescr((StgPtr)tso);
870 if (bd->gen_no > 0) recordMutableCap((StgClosure*)tso,cap,bd->gen_no);
876 dirty_TSO (Capability *cap, StgTSO *tso)
879 if ((tso->flags & (TSO_DIRTY|TSO_LINK_DIRTY)) == 0) {
880 bd = Bdescr((StgPtr)tso);
881 if (bd->gen_no > 0) recordMutableCap((StgClosure*)tso,cap,bd->gen_no);
883 tso->flags |= TSO_DIRTY;
887 This is the write barrier for MVARs. An MVAR_CLEAN objects is not
888 on the mutable list; a MVAR_DIRTY is. When written to, a
889 MVAR_CLEAN turns into a MVAR_DIRTY and is put on the mutable list.
890 The check for MVAR_CLEAN is inlined at the call site for speed,
891 this really does make a difference on concurrency-heavy benchmarks
892 such as Chaneneos and cheap-concurrency.
895 dirty_MVAR(StgRegTable *reg, StgClosure *p)
897 Capability *cap = regTableToCapability(reg);
899 bd = Bdescr((StgPtr)p);
900 if (bd->gen_no > 0) recordMutableCap(p,cap,bd->gen_no);
903 /* -----------------------------------------------------------------------------
904 Allocation functions for GMP.
906 These all use the allocate() interface - we can't have any garbage
907 collection going on during a gmp operation, so we use allocate()
908 which always succeeds. The gmp operations which might need to
909 allocate will ask the storage manager (via doYouWantToGC()) whether
910 a garbage collection is required, in case we get into a loop doing
911 only allocate() style allocation.
912 -------------------------------------------------------------------------- */
915 stgAllocForGMP (size_t size_in_bytes)
918 nat data_size_in_words, total_size_in_words;
920 /* round up to a whole number of words */
921 data_size_in_words = (size_in_bytes + sizeof(W_) + 1) / sizeof(W_);
922 total_size_in_words = sizeofW(StgArrWords) + data_size_in_words;
924 /* allocate and fill it in. */
925 #if defined(THREADED_RTS)
926 arr = (StgArrWords *)allocateLocal(myTask()->cap, total_size_in_words);
928 arr = (StgArrWords *)allocateLocal(&MainCapability, total_size_in_words);
930 SET_ARR_HDR(arr, &stg_ARR_WORDS_info, CCCS, data_size_in_words);
932 /* and return a ptr to the goods inside the array */
937 stgReallocForGMP (void *ptr, size_t old_size, size_t new_size)
940 void *new_stuff_ptr = stgAllocForGMP(new_size);
942 char *p = (char *) ptr;
943 char *q = (char *) new_stuff_ptr;
945 min_size = old_size < new_size ? old_size : new_size;
946 for (; i < min_size; i++, p++, q++) {
950 return(new_stuff_ptr);
954 stgDeallocForGMP (void *ptr STG_UNUSED,
955 size_t size STG_UNUSED)
957 /* easy for us: the garbage collector does the dealloc'n */
960 /* -----------------------------------------------------------------------------
962 * -------------------------------------------------------------------------- */
964 /* -----------------------------------------------------------------------------
967 * Approximate how much we've allocated: number of blocks in the
968 * nursery + blocks allocated via allocate() - unused nusery blocks.
969 * This leaves a little slop at the end of each block, and doesn't
970 * take into account large objects (ToDo).
971 * -------------------------------------------------------------------------- */
974 calcAllocated( void )
979 allocated = allocatedBytes();
980 allocated += countNurseryBlocks() * BLOCK_SIZE_W;
985 for (i = 0; i < n_nurseries; i++) {
987 for ( bd = capabilities[i].r.rCurrentNursery->link;
988 bd != NULL; bd = bd->link ) {
989 allocated -= BLOCK_SIZE_W;
991 cap = &capabilities[i];
992 if (cap->r.rCurrentNursery->free <
993 cap->r.rCurrentNursery->start + BLOCK_SIZE_W) {
994 allocated -= (cap->r.rCurrentNursery->start + BLOCK_SIZE_W)
995 - cap->r.rCurrentNursery->free;
999 bdescr *current_nursery = MainCapability.r.rCurrentNursery;
1001 for ( bd = current_nursery->link; bd != NULL; bd = bd->link ) {
1002 allocated -= BLOCK_SIZE_W;
1004 if (current_nursery->free < current_nursery->start + BLOCK_SIZE_W) {
1005 allocated -= (current_nursery->start + BLOCK_SIZE_W)
1006 - current_nursery->free;
1011 total_allocated += allocated;
1015 /* Approximate the amount of live data in the heap. To be called just
1016 * after garbage collection (see GarbageCollect()).
1019 calcLiveBlocks(void)
1025 if (RtsFlags.GcFlags.generations == 1) {
1026 return g0s0->n_large_blocks + g0s0->n_blocks;
1029 for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
1030 for (s = 0; s < generations[g].n_steps; s++) {
1031 /* approximate amount of live data (doesn't take into account slop
1032 * at end of each block).
1034 if (g == 0 && s == 0) {
1037 stp = &generations[g].steps[s];
1038 live += stp->n_large_blocks + stp->n_blocks;
1045 countOccupied(bdescr *bd)
1050 for (; bd != NULL; bd = bd->link) {
1051 ASSERT(bd->free <= bd->start + bd->blocks * BLOCK_SIZE_W);
1052 words += bd->free - bd->start;
1057 // Return an accurate count of the live data in the heap, excluding
1066 if (RtsFlags.GcFlags.generations == 1) {
1067 return g0s0->n_words + countOccupied(g0s0->large_objects);
1071 for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
1072 for (s = 0; s < generations[g].n_steps; s++) {
1073 if (g == 0 && s == 0) continue;
1074 stp = &generations[g].steps[s];
1075 live += stp->n_words + countOccupied(stp->large_objects);
1081 /* Approximate the number of blocks that will be needed at the next
1082 * garbage collection.
1084 * Assume: all data currently live will remain live. Steps that will
1085 * be collected next time will therefore need twice as many blocks
1086 * since all the data will be copied.
1095 for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
1096 for (s = 0; s < generations[g].n_steps; s++) {
1097 if (g == 0 && s == 0) { continue; }
1098 stp = &generations[g].steps[s];
1100 // we need at least this much space
1101 needed += stp->n_blocks + stp->n_large_blocks;
1103 // any additional space needed to collect this gen next time?
1104 if (g == 0 || // always collect gen 0
1105 (generations[g].steps[0].n_blocks +
1106 generations[g].steps[0].n_large_blocks
1107 > generations[g].max_blocks)) {
1108 // we will collect this gen next time
1111 needed += stp->n_blocks / BITS_IN(W_);
1113 needed += stp->n_blocks / 100;
1116 continue; // no additional space needed for compaction
1118 needed += stp->n_blocks;
1126 /* ----------------------------------------------------------------------------
1129 Executable memory must be managed separately from non-executable
1130 memory. Most OSs these days require you to jump through hoops to
1131 dynamically allocate executable memory, due to various security
1134 Here we provide a small memory allocator for executable memory.
1135 Memory is managed with a page granularity; we allocate linearly
1136 in the page, and when the page is emptied (all objects on the page
1137 are free) we free the page again, not forgetting to make it
1140 TODO: The inability to handle objects bigger than BLOCK_SIZE_W means that
1141 the linker cannot use allocateExec for loading object code files
1142 on Windows. Once allocateExec can handle larger objects, the linker
1143 should be modified to use allocateExec instead of VirtualAlloc.
1144 ------------------------------------------------------------------------- */
1146 #if defined(linux_HOST_OS)
1148 // On Linux we need to use libffi for allocating executable memory,
1149 // because it knows how to work around the restrictions put in place
1152 void *allocateExec (nat bytes, void **exec_ret)
1156 ret = ffi_closure_alloc (sizeof(void *) + (size_t)bytes, (void**)&exec);
1158 if (ret == NULL) return ret;
1159 *ret = ret; // save the address of the writable mapping, for freeExec().
1160 *exec_ret = exec + 1;
1164 // freeExec gets passed the executable address, not the writable address.
1165 void freeExec (void *addr)
1168 writable = *((void**)addr - 1);
1170 ffi_closure_free (writable);
1176 void *allocateExec (nat bytes, void **exec_ret)
1183 // round up to words.
1184 n = (bytes + sizeof(W_) + 1) / sizeof(W_);
1186 if (n+1 > BLOCK_SIZE_W) {
1187 barf("allocateExec: can't handle large objects");
1190 if (exec_block == NULL ||
1191 exec_block->free + n + 1 > exec_block->start + BLOCK_SIZE_W) {
1193 lnat pagesize = getPageSize();
1194 bd = allocGroup(stg_max(1, pagesize / BLOCK_SIZE));
1195 debugTrace(DEBUG_gc, "allocate exec block %p", bd->start);
1197 bd->flags = BF_EXEC;
1198 bd->link = exec_block;
1199 if (exec_block != NULL) {
1200 exec_block->u.back = bd;
1203 setExecutable(bd->start, bd->blocks * BLOCK_SIZE, rtsTrue);
1206 *(exec_block->free) = n; // store the size of this chunk
1207 exec_block->gen_no += n; // gen_no stores the number of words allocated
1208 ret = exec_block->free + 1;
1209 exec_block->free += n + 1;
1216 void freeExec (void *addr)
1218 StgPtr p = (StgPtr)addr - 1;
1219 bdescr *bd = Bdescr((StgPtr)p);
1221 if ((bd->flags & BF_EXEC) == 0) {
1222 barf("freeExec: not executable");
1225 if (*(StgPtr)p == 0) {
1226 barf("freeExec: already free?");
1231 bd->gen_no -= *(StgPtr)p;
1234 if (bd->gen_no == 0) {
1235 // Free the block if it is empty, but not if it is the block at
1236 // the head of the queue.
1237 if (bd != exec_block) {
1238 debugTrace(DEBUG_gc, "free exec block %p", bd->start);
1239 dbl_link_remove(bd, &exec_block);
1240 setExecutable(bd->start, bd->blocks * BLOCK_SIZE, rtsFalse);
1243 bd->free = bd->start;
1250 #endif /* mingw32_HOST_OS */
1252 /* -----------------------------------------------------------------------------
1255 memInventory() checks for memory leaks by counting up all the
1256 blocks we know about and comparing that to the number of blocks
1257 allegedly floating around in the system.
1258 -------------------------------------------------------------------------- */
1262 // Useful for finding partially full blocks in gdb
1263 void findSlop(bdescr *bd);
1264 void findSlop(bdescr *bd)
1268 for (; bd != NULL; bd = bd->link) {
1269 slop = (bd->blocks * BLOCK_SIZE_W) - (bd->free - bd->start);
1270 if (slop > (1024/sizeof(W_))) {
1271 debugBelch("block at %p (bdescr %p) has %ldKB slop\n",
1272 bd->start, bd, slop / (1024/sizeof(W_)));
1278 countBlocks(bdescr *bd)
1281 for (n=0; bd != NULL; bd=bd->link) {
1287 // (*1) Just like countBlocks, except that we adjust the count for a
1288 // megablock group so that it doesn't include the extra few blocks
1289 // that would be taken up by block descriptors in the second and
1290 // subsequent megablock. This is so we can tally the count with the
1291 // number of blocks allocated in the system, for memInventory().
1293 countAllocdBlocks(bdescr *bd)
1296 for (n=0; bd != NULL; bd=bd->link) {
1298 // hack for megablock groups: see (*1) above
1299 if (bd->blocks > BLOCKS_PER_MBLOCK) {
1300 n -= (MBLOCK_SIZE / BLOCK_SIZE - BLOCKS_PER_MBLOCK)
1301 * (bd->blocks/(MBLOCK_SIZE/BLOCK_SIZE));
1308 stepBlocks (step *stp)
1310 ASSERT(countBlocks(stp->blocks) == stp->n_blocks);
1311 ASSERT(countBlocks(stp->large_objects) == stp->n_large_blocks);
1312 return stp->n_blocks + stp->n_old_blocks +
1313 countAllocdBlocks(stp->large_objects);
1316 // If memInventory() calculates that we have a memory leak, this
1317 // function will try to find the block(s) that are leaking by marking
1318 // all the ones that we know about, and search through memory to find
1319 // blocks that are not marked. In the debugger this can help to give
1320 // us a clue about what kind of block leaked. In the future we might
1321 // annotate blocks with their allocation site to give more helpful
1324 findMemoryLeak (void)
1327 for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
1328 for (i = 0; i < n_capabilities; i++) {
1329 markBlocks(capabilities[i].mut_lists[g]);
1331 markBlocks(generations[g].mut_list);
1332 for (s = 0; s < generations[g].n_steps; s++) {
1333 markBlocks(generations[g].steps[s].blocks);
1334 markBlocks(generations[g].steps[s].large_objects);
1338 for (i = 0; i < n_nurseries; i++) {
1339 markBlocks(nurseries[i].blocks);
1340 markBlocks(nurseries[i].large_objects);
1345 // if (RtsFlags.ProfFlags.doHeapProfile == HEAP_BY_RETAINER) {
1346 // markRetainerBlocks();
1350 // count the blocks allocated by the arena allocator
1352 // markArenaBlocks();
1354 // count the blocks containing executable memory
1355 markBlocks(exec_block);
1357 reportUnmarkedBlocks();
1362 memInventory (rtsBool show)
1366 lnat gen_blocks[RtsFlags.GcFlags.generations];
1367 lnat nursery_blocks, retainer_blocks,
1368 arena_blocks, exec_blocks;
1369 lnat live_blocks = 0, free_blocks = 0;
1372 // count the blocks we current have
1374 for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
1376 for (i = 0; i < n_capabilities; i++) {
1377 gen_blocks[g] += countBlocks(capabilities[i].mut_lists[g]);
1379 gen_blocks[g] += countAllocdBlocks(generations[g].mut_list);
1380 for (s = 0; s < generations[g].n_steps; s++) {
1381 stp = &generations[g].steps[s];
1382 gen_blocks[g] += stepBlocks(stp);
1387 for (i = 0; i < n_nurseries; i++) {
1388 nursery_blocks += stepBlocks(&nurseries[i]);
1391 retainer_blocks = 0;
1393 if (RtsFlags.ProfFlags.doHeapProfile == HEAP_BY_RETAINER) {
1394 retainer_blocks = retainerStackBlocks();
1398 // count the blocks allocated by the arena allocator
1399 arena_blocks = arenaBlocks();
1401 // count the blocks containing executable memory
1402 exec_blocks = countAllocdBlocks(exec_block);
1404 /* count the blocks on the free list */
1405 free_blocks = countFreeList();
1408 for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
1409 live_blocks += gen_blocks[g];
1411 live_blocks += nursery_blocks +
1412 + retainer_blocks + arena_blocks + exec_blocks;
1414 #define MB(n) (((n) * BLOCK_SIZE_W) / ((1024*1024)/sizeof(W_)))
1416 leak = live_blocks + free_blocks != mblocks_allocated * BLOCKS_PER_MBLOCK;
1421 debugBelch("Memory leak detected:\n");
1423 debugBelch("Memory inventory:\n");
1425 for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
1426 debugBelch(" gen %d blocks : %5lu blocks (%lu MB)\n", g,
1427 gen_blocks[g], MB(gen_blocks[g]));
1429 debugBelch(" nursery : %5lu blocks (%lu MB)\n",
1430 nursery_blocks, MB(nursery_blocks));
1431 debugBelch(" retainer : %5lu blocks (%lu MB)\n",
1432 retainer_blocks, MB(retainer_blocks));
1433 debugBelch(" arena blocks : %5lu blocks (%lu MB)\n",
1434 arena_blocks, MB(arena_blocks));
1435 debugBelch(" exec : %5lu blocks (%lu MB)\n",
1436 exec_blocks, MB(exec_blocks));
1437 debugBelch(" free : %5lu blocks (%lu MB)\n",
1438 free_blocks, MB(free_blocks));
1439 debugBelch(" total : %5lu blocks (%lu MB)\n",
1440 live_blocks + free_blocks, MB(live_blocks+free_blocks));
1442 debugBelch("\n in system : %5lu blocks (%lu MB)\n",
1443 mblocks_allocated * BLOCKS_PER_MBLOCK, mblocks_allocated);
1451 ASSERT(n_alloc_blocks == live_blocks);
1456 /* Full heap sanity check. */
1462 if (RtsFlags.GcFlags.generations == 1) {
1463 checkHeap(g0s0->blocks);
1464 checkChain(g0s0->large_objects);
1467 for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
1468 for (s = 0; s < generations[g].n_steps; s++) {
1469 if (g == 0 && s == 0) { continue; }
1470 ASSERT(countBlocks(generations[g].steps[s].blocks)
1471 == generations[g].steps[s].n_blocks);
1472 ASSERT(countBlocks(generations[g].steps[s].large_objects)
1473 == generations[g].steps[s].n_large_blocks);
1474 checkHeap(generations[g].steps[s].blocks);
1475 checkChain(generations[g].steps[s].large_objects);
1477 checkMutableList(generations[g].mut_list, g);
1482 for (s = 0; s < n_nurseries; s++) {
1483 ASSERT(countBlocks(nurseries[s].blocks)
1484 == nurseries[s].n_blocks);
1485 ASSERT(countBlocks(nurseries[s].large_objects)
1486 == nurseries[s].n_large_blocks);
1489 checkFreeListSanity();
1492 #if defined(THREADED_RTS)
1493 // check the stacks too in threaded mode, because we don't do a
1494 // full heap sanity check in this case (see checkHeap())
1495 checkGlobalTSOList(rtsTrue);
1497 checkGlobalTSOList(rtsFalse);
1501 /* Nursery sanity check */
1503 checkNurserySanity( step *stp )
1509 for (bd = stp->blocks; bd != NULL; bd = bd->link) {
1510 ASSERT(bd->u.back == prev);
1512 blocks += bd->blocks;
1514 ASSERT(blocks == stp->n_blocks);
1517 // handy function for use in gdb, because Bdescr() is inlined.
1518 extern bdescr *_bdescr( StgPtr p );