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_large_objects);
109 stp->threads = END_TSO_QUEUE;
110 stp->old_threads = END_TSO_QUEUE;
119 if (generations != NULL) {
120 // multi-init protection
126 /* Sanity check to make sure the LOOKS_LIKE_ macros appear to be
127 * doing something reasonable.
129 /* We use the NOT_NULL variant or gcc warns that the test is always true */
130 ASSERT(LOOKS_LIKE_INFO_PTR_NOT_NULL((StgWord)&stg_BLACKHOLE_info));
131 ASSERT(LOOKS_LIKE_CLOSURE_PTR(&stg_dummy_ret_closure));
132 ASSERT(!HEAP_ALLOCED(&stg_dummy_ret_closure));
134 if (RtsFlags.GcFlags.maxHeapSize != 0 &&
135 RtsFlags.GcFlags.heapSizeSuggestion >
136 RtsFlags.GcFlags.maxHeapSize) {
137 RtsFlags.GcFlags.maxHeapSize = RtsFlags.GcFlags.heapSizeSuggestion;
140 if (RtsFlags.GcFlags.maxHeapSize != 0 &&
141 RtsFlags.GcFlags.minAllocAreaSize >
142 RtsFlags.GcFlags.maxHeapSize) {
143 errorBelch("maximum heap size (-M) is smaller than minimum alloc area size (-A)");
144 RtsFlags.GcFlags.minAllocAreaSize = RtsFlags.GcFlags.maxHeapSize;
147 initBlockAllocator();
149 #if defined(THREADED_RTS)
150 initMutex(&sm_mutex);
151 initMutex(&atomic_modify_mutvar_mutex);
156 /* allocate generation info array */
157 generations = (generation *)stgMallocBytes(RtsFlags.GcFlags.generations
158 * sizeof(struct generation_),
159 "initStorage: gens");
161 /* allocate all the steps into an array. It is important that we do
162 it this way, because we need the invariant that two step pointers
163 can be directly compared to see which is the oldest.
164 Remember that the last generation has only one step. */
165 total_steps = 1 + (RtsFlags.GcFlags.generations - 1) * RtsFlags.GcFlags.steps;
166 all_steps = stgMallocBytes(total_steps * sizeof(struct step_),
167 "initStorage: steps");
169 /* Initialise all generations */
170 for(g = 0; g < RtsFlags.GcFlags.generations; g++) {
171 gen = &generations[g];
173 gen->mut_list = allocBlock();
174 gen->collections = 0;
175 gen->par_collections = 0;
176 gen->failed_promotions = 0;
180 /* A couple of convenience pointers */
181 g0 = &generations[0];
182 oldest_gen = &generations[RtsFlags.GcFlags.generations-1];
184 /* Allocate step structures in each generation */
185 if (RtsFlags.GcFlags.generations > 1) {
186 /* Only for multiple-generations */
188 /* Oldest generation: one step */
189 oldest_gen->n_steps = 1;
190 oldest_gen->steps = all_steps + (RtsFlags.GcFlags.generations - 1)
191 * RtsFlags.GcFlags.steps;
193 /* set up all except the oldest generation with 2 steps */
194 for(g = 0; g < RtsFlags.GcFlags.generations-1; g++) {
195 generations[g].n_steps = RtsFlags.GcFlags.steps;
196 generations[g].steps = all_steps + g * RtsFlags.GcFlags.steps;
200 /* single generation, i.e. a two-space collector */
202 g0->steps = all_steps;
206 n_nurseries = n_capabilities;
210 nurseries = stgMallocBytes (n_nurseries * sizeof(struct step_),
211 "initStorage: nurseries");
213 /* Initialise all steps */
214 for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
215 for (s = 0; s < generations[g].n_steps; s++) {
216 initStep(&generations[g].steps[s], g, s);
220 for (s = 0; s < n_nurseries; s++) {
221 initStep(&nurseries[s], 0, s);
224 /* Set up the destination pointers in each younger gen. step */
225 for (g = 0; g < RtsFlags.GcFlags.generations-1; g++) {
226 for (s = 0; s < generations[g].n_steps-1; s++) {
227 generations[g].steps[s].to = &generations[g].steps[s+1];
229 generations[g].steps[s].to = &generations[g+1].steps[0];
231 oldest_gen->steps[0].to = &oldest_gen->steps[0];
233 for (s = 0; s < n_nurseries; s++) {
234 nurseries[s].to = generations[0].steps[0].to;
237 /* The oldest generation has one step. */
238 if (RtsFlags.GcFlags.compact || RtsFlags.GcFlags.sweep) {
239 if (RtsFlags.GcFlags.generations == 1) {
240 errorBelch("WARNING: compact/sweep is incompatible with -G1; disabled");
242 oldest_gen->steps[0].mark = 1;
243 if (RtsFlags.GcFlags.compact)
244 oldest_gen->steps[0].compact = 1;
248 generations[0].max_blocks = 0;
249 g0s0 = &generations[0].steps[0];
251 /* The allocation area. Policy: keep the allocation area
252 * small to begin with, even if we have a large suggested heap
253 * size. Reason: we're going to do a major collection first, and we
254 * don't want it to be a big one. This vague idea is borne out by
255 * rigorous experimental evidence.
259 weak_ptr_list = NULL;
261 revertible_caf_list = NULL;
263 /* initialise the allocate() interface */
265 alloc_blocks_lim = RtsFlags.GcFlags.minAllocAreaSize;
269 /* Tell GNU multi-precision pkg about our custom alloc functions */
270 mp_set_memory_functions(stgAllocForGMP, stgReallocForGMP, stgDeallocForGMP);
273 initSpinLock(&gc_alloc_block_sync);
281 IF_DEBUG(gc, statDescribeGens());
289 stat_exit(calcAllocated());
295 stgFree(g0s0); // frees all the steps
296 stgFree(generations);
298 #if defined(THREADED_RTS)
299 closeMutex(&sm_mutex);
300 closeMutex(&atomic_modify_mutvar_mutex);
305 /* -----------------------------------------------------------------------------
308 The entry code for every CAF does the following:
310 - builds a CAF_BLACKHOLE in the heap
311 - pushes an update frame pointing to the CAF_BLACKHOLE
312 - invokes UPD_CAF(), which:
313 - calls newCaf, below
314 - updates the CAF with a static indirection to the CAF_BLACKHOLE
316 Why do we build a BLACKHOLE in the heap rather than just updating
317 the thunk directly? It's so that we only need one kind of update
318 frame - otherwise we'd need a static version of the update frame too.
320 newCaf() does the following:
322 - it puts the CAF on the oldest generation's mut-once list.
323 This is so that we can treat the CAF as a root when collecting
326 For GHCI, we have additional requirements when dealing with CAFs:
328 - we must *retain* all dynamically-loaded CAFs ever entered,
329 just in case we need them again.
330 - we must be able to *revert* CAFs that have been evaluated, to
331 their pre-evaluated form.
333 To do this, we use an additional CAF list. When newCaf() is
334 called on a dynamically-loaded CAF, we add it to the CAF list
335 instead of the old-generation mutable list, and save away its
336 old info pointer (in caf->saved_info) for later reversion.
338 To revert all the CAFs, we traverse the CAF list and reset the
339 info pointer to caf->saved_info, then throw away the CAF list.
340 (see GC.c:revertCAFs()).
344 -------------------------------------------------------------------------- */
347 newCAF(StgClosure* caf)
354 // If we are in GHCi _and_ we are using dynamic libraries,
355 // then we can't redirect newCAF calls to newDynCAF (see below),
356 // so we make newCAF behave almost like newDynCAF.
357 // The dynamic libraries might be used by both the interpreted
358 // program and GHCi itself, so they must not be reverted.
359 // This also means that in GHCi with dynamic libraries, CAFs are not
360 // garbage collected. If this turns out to be a problem, we could
361 // do another hack here and do an address range test on caf to figure
362 // out whether it is from a dynamic library.
363 ((StgIndStatic *)caf)->saved_info = (StgInfoTable *)caf->header.info;
364 ((StgIndStatic *)caf)->static_link = caf_list;
369 /* Put this CAF on the mutable list for the old generation.
370 * This is a HACK - the IND_STATIC closure doesn't really have
371 * a mut_link field, but we pretend it has - in fact we re-use
372 * the STATIC_LINK field for the time being, because when we
373 * come to do a major GC we won't need the mut_link field
374 * any more and can use it as a STATIC_LINK.
376 ((StgIndStatic *)caf)->saved_info = NULL;
377 recordMutableGen(caf, oldest_gen->no);
383 // An alternate version of newCaf which is used for dynamically loaded
384 // object code in GHCi. In this case we want to retain *all* CAFs in
385 // the object code, because they might be demanded at any time from an
386 // expression evaluated on the command line.
387 // Also, GHCi might want to revert CAFs, so we add these to the
388 // revertible_caf_list.
390 // The linker hackily arranges that references to newCaf from dynamic
391 // code end up pointing to newDynCAF.
393 newDynCAF(StgClosure *caf)
397 ((StgIndStatic *)caf)->saved_info = (StgInfoTable *)caf->header.info;
398 ((StgIndStatic *)caf)->static_link = revertible_caf_list;
399 revertible_caf_list = caf;
404 /* -----------------------------------------------------------------------------
406 -------------------------------------------------------------------------- */
409 allocNursery (step *stp, bdescr *tail, nat blocks)
414 // Allocate a nursery: we allocate fresh blocks one at a time and
415 // cons them on to the front of the list, not forgetting to update
416 // the back pointer on the tail of the list to point to the new block.
417 for (i=0; i < blocks; i++) {
420 processNursery() in LdvProfile.c assumes that every block group in
421 the nursery contains only a single block. So, if a block group is
422 given multiple blocks, change processNursery() accordingly.
426 // double-link the nursery: we might need to insert blocks
433 bd->free = bd->start;
441 assignNurseriesToCapabilities (void)
446 for (i = 0; i < n_nurseries; i++) {
447 capabilities[i].r.rNursery = &nurseries[i];
448 capabilities[i].r.rCurrentNursery = nurseries[i].blocks;
449 capabilities[i].r.rCurrentAlloc = NULL;
451 #else /* THREADED_RTS */
452 MainCapability.r.rNursery = &nurseries[0];
453 MainCapability.r.rCurrentNursery = nurseries[0].blocks;
454 MainCapability.r.rCurrentAlloc = NULL;
459 allocNurseries( void )
463 for (i = 0; i < n_nurseries; i++) {
464 nurseries[i].blocks =
465 allocNursery(&nurseries[i], NULL,
466 RtsFlags.GcFlags.minAllocAreaSize);
467 nurseries[i].n_blocks = RtsFlags.GcFlags.minAllocAreaSize;
468 nurseries[i].old_blocks = NULL;
469 nurseries[i].n_old_blocks = 0;
471 assignNurseriesToCapabilities();
475 resetNurseries( void )
481 for (i = 0; i < n_nurseries; i++) {
483 for (bd = stp->blocks; bd; bd = bd->link) {
484 bd->free = bd->start;
485 ASSERT(bd->gen_no == 0);
486 ASSERT(bd->step == stp);
487 IF_DEBUG(sanity,memset(bd->start, 0xaa, BLOCK_SIZE));
490 assignNurseriesToCapabilities();
494 countNurseryBlocks (void)
499 for (i = 0; i < n_nurseries; i++) {
500 blocks += nurseries[i].n_blocks;
506 resizeNursery ( step *stp, nat blocks )
511 nursery_blocks = stp->n_blocks;
512 if (nursery_blocks == blocks) return;
514 if (nursery_blocks < blocks) {
515 debugTrace(DEBUG_gc, "increasing size of nursery to %d blocks",
517 stp->blocks = allocNursery(stp, stp->blocks, blocks-nursery_blocks);
522 debugTrace(DEBUG_gc, "decreasing size of nursery to %d blocks",
526 while (nursery_blocks > blocks) {
528 next_bd->u.back = NULL;
529 nursery_blocks -= bd->blocks; // might be a large block
534 // might have gone just under, by freeing a large block, so make
535 // up the difference.
536 if (nursery_blocks < blocks) {
537 stp->blocks = allocNursery(stp, stp->blocks, blocks-nursery_blocks);
541 stp->n_blocks = blocks;
542 ASSERT(countBlocks(stp->blocks) == stp->n_blocks);
546 // Resize each of the nurseries to the specified size.
549 resizeNurseriesFixed (nat blocks)
552 for (i = 0; i < n_nurseries; i++) {
553 resizeNursery(&nurseries[i], blocks);
558 // Resize the nurseries to the total specified size.
561 resizeNurseries (nat blocks)
563 // If there are multiple nurseries, then we just divide the number
564 // of available blocks between them.
565 resizeNurseriesFixed(blocks / n_nurseries);
569 /* -----------------------------------------------------------------------------
570 move_TSO is called to update the TSO structure after it has been
571 moved from one place to another.
572 -------------------------------------------------------------------------- */
575 move_TSO (StgTSO *src, StgTSO *dest)
579 // relocate the stack pointer...
580 diff = (StgPtr)dest - (StgPtr)src; // In *words*
581 dest->sp = (StgPtr)dest->sp + diff;
584 /* -----------------------------------------------------------------------------
585 The allocate() interface
587 allocateInGen() function allocates memory directly into a specific
588 generation. It always succeeds, and returns a chunk of memory n
589 words long. n can be larger than the size of a block if necessary,
590 in which case a contiguous block group will be allocated.
592 allocate(n) is equivalent to allocateInGen(g0).
593 -------------------------------------------------------------------------- */
596 allocateInGen (generation *g, lnat n)
604 TICK_ALLOC_HEAP_NOCTR(n);
609 if (n >= LARGE_OBJECT_THRESHOLD/sizeof(W_))
611 lnat req_blocks = (lnat)BLOCK_ROUND_UP(n*sizeof(W_)) / BLOCK_SIZE;
613 // Attempting to allocate an object larger than maxHeapSize
614 // should definitely be disallowed. (bug #1791)
615 if (RtsFlags.GcFlags.maxHeapSize > 0 &&
616 req_blocks >= RtsFlags.GcFlags.maxHeapSize) {
618 // heapOverflow() doesn't exit (see #2592), but we aren't
619 // in a position to do a clean shutdown here: we
620 // either have to allocate the memory or exit now.
621 // Allocating the memory would be bad, because the user
622 // has requested that we not exceed maxHeapSize, so we
624 stg_exit(EXIT_HEAPOVERFLOW);
627 bd = allocGroup(req_blocks);
628 dbl_link_onto(bd, &stp->large_objects);
629 stp->n_large_blocks += bd->blocks; // might be larger than req_blocks
630 alloc_blocks += bd->blocks;
633 bd->flags = BF_LARGE;
634 bd->free = bd->start + n;
639 // small allocation (<LARGE_OBJECT_THRESHOLD) */
641 if (bd == NULL || bd->free + n > bd->start + BLOCK_SIZE_W) {
646 bd->link = stp->blocks;
663 return allocateInGen(g0,n);
667 allocatedBytes( void )
671 allocated = alloc_blocks * BLOCK_SIZE_W;
672 if (pinned_object_block != NULL) {
673 allocated -= (pinned_object_block->start + BLOCK_SIZE_W) -
674 pinned_object_block->free;
680 // split N blocks off the front of the given bdescr, returning the
681 // new block group. We treat the remainder as if it
682 // had been freshly allocated in generation 0.
684 splitLargeBlock (bdescr *bd, nat blocks)
688 // subtract the original number of blocks from the counter first
689 bd->step->n_large_blocks -= bd->blocks;
691 new_bd = splitBlockGroup (bd, blocks);
693 dbl_link_onto(new_bd, &g0s0->large_objects);
694 g0s0->n_large_blocks += new_bd->blocks;
695 new_bd->gen_no = g0s0->no;
697 new_bd->flags = BF_LARGE;
698 new_bd->free = bd->free;
699 ASSERT(new_bd->free <= new_bd->start + new_bd->blocks * BLOCK_SIZE_W);
701 // add the new number of blocks to the counter. Due to the gaps
702 // for block descriptor, new_bd->blocks + bd->blocks might not be
703 // equal to the original bd->blocks, which is why we do it this way.
704 bd->step->n_large_blocks += bd->blocks;
709 /* -----------------------------------------------------------------------------
712 This allocates memory in the current thread - it is intended for
713 use primarily from STG-land where we have a Capability. It is
714 better than allocate() because it doesn't require taking the
715 sm_mutex lock in the common case.
717 Memory is allocated directly from the nursery if possible (but not
718 from the current nursery block, so as not to interfere with
720 -------------------------------------------------------------------------- */
723 allocateLocal (Capability *cap, lnat n)
728 if (n >= LARGE_OBJECT_THRESHOLD/sizeof(W_)) {
729 return allocateInGen(g0,n);
732 /* small allocation (<LARGE_OBJECT_THRESHOLD) */
734 TICK_ALLOC_HEAP_NOCTR(n);
737 bd = cap->r.rCurrentAlloc;
738 if (bd == NULL || bd->free + n > bd->start + BLOCK_SIZE_W) {
740 // The CurrentAlloc block is full, we need to find another
741 // one. First, we try taking the next block from the
743 bd = cap->r.rCurrentNursery->link;
745 if (bd == NULL || bd->free + n > bd->start + BLOCK_SIZE_W) {
746 // The nursery is empty, or the next block is already
747 // full: allocate a fresh block (we can't fail here).
750 cap->r.rNursery->n_blocks++;
753 bd->step = cap->r.rNursery;
755 // NO: alloc_blocks++;
756 // calcAllocated() uses the size of the nursery, and we've
757 // already bumpted nursery->n_blocks above. We'll GC
758 // pretty quickly now anyway, because MAYBE_GC() will
759 // notice that CurrentNursery->link is NULL.
761 // we have a block in the nursery: take it and put
762 // it at the *front* of the nursery list, and use it
763 // to allocate() from.
764 cap->r.rCurrentNursery->link = bd->link;
765 if (bd->link != NULL) {
766 bd->link->u.back = cap->r.rCurrentNursery;
769 dbl_link_onto(bd, &cap->r.rNursery->blocks);
770 cap->r.rCurrentAlloc = bd;
771 IF_DEBUG(sanity, checkNurserySanity(cap->r.rNursery));
778 /* ---------------------------------------------------------------------------
779 Allocate a fixed/pinned object.
781 We allocate small pinned objects into a single block, allocating a
782 new block when the current one overflows. The block is chained
783 onto the large_object_list of generation 0 step 0.
785 NOTE: The GC can't in general handle pinned objects. This
786 interface is only safe to use for ByteArrays, which have no
787 pointers and don't require scavenging. It works because the
788 block's descriptor has the BF_LARGE flag set, so the block is
789 treated as a large object and chained onto various lists, rather
790 than the individual objects being copied. However, when it comes
791 to scavenge the block, the GC will only scavenge the first object.
792 The reason is that the GC can't linearly scan a block of pinned
793 objects at the moment (doing so would require using the
794 mostly-copying techniques). But since we're restricting ourselves
795 to pinned ByteArrays, not scavenging is ok.
797 This function is called by newPinnedByteArray# which immediately
798 fills the allocated memory with a MutableByteArray#.
799 ------------------------------------------------------------------------- */
802 allocatePinned( lnat n )
805 bdescr *bd = pinned_object_block;
807 // If the request is for a large object, then allocate()
808 // will give us a pinned object anyway.
809 if (n >= LARGE_OBJECT_THRESHOLD/sizeof(W_)) {
811 Bdescr(p)->flags |= BF_PINNED;
817 TICK_ALLOC_HEAP_NOCTR(n);
820 // If we don't have a block of pinned objects yet, or the current
821 // one isn't large enough to hold the new object, allocate a new one.
822 if (bd == NULL || (bd->free + n) > (bd->start + BLOCK_SIZE_W)) {
823 pinned_object_block = bd = allocBlock();
824 dbl_link_onto(bd, &g0s0->large_objects);
825 g0s0->n_large_blocks++;
828 bd->flags = BF_PINNED | BF_LARGE;
829 bd->free = bd->start;
839 /* -----------------------------------------------------------------------------
841 -------------------------------------------------------------------------- */
844 This is the write barrier for MUT_VARs, a.k.a. IORefs. A
845 MUT_VAR_CLEAN object is not on the mutable list; a MUT_VAR_DIRTY
846 is. When written to, a MUT_VAR_CLEAN turns into a MUT_VAR_DIRTY
847 and is put on the mutable list.
850 dirty_MUT_VAR(StgRegTable *reg, StgClosure *p)
852 Capability *cap = regTableToCapability(reg);
854 if (p->header.info == &stg_MUT_VAR_CLEAN_info) {
855 p->header.info = &stg_MUT_VAR_DIRTY_info;
856 bd = Bdescr((StgPtr)p);
857 if (bd->gen_no > 0) recordMutableCap(p,cap,bd->gen_no);
861 // Setting a TSO's link field with a write barrier.
862 // It is *not* necessary to call this function when
863 // * setting the link field to END_TSO_QUEUE
864 // * putting a TSO on the blackhole_queue
865 // * setting the link field of the currently running TSO, as it
866 // will already be dirty.
868 setTSOLink (Capability *cap, StgTSO *tso, StgTSO *target)
871 if ((tso->flags & (TSO_DIRTY|TSO_LINK_DIRTY)) == 0) {
872 tso->flags |= TSO_LINK_DIRTY;
873 bd = Bdescr((StgPtr)tso);
874 if (bd->gen_no > 0) recordMutableCap((StgClosure*)tso,cap,bd->gen_no);
880 dirty_TSO (Capability *cap, StgTSO *tso)
883 if ((tso->flags & (TSO_DIRTY|TSO_LINK_DIRTY)) == 0) {
884 bd = Bdescr((StgPtr)tso);
885 if (bd->gen_no > 0) recordMutableCap((StgClosure*)tso,cap,bd->gen_no);
887 tso->flags |= TSO_DIRTY;
891 This is the write barrier for MVARs. An MVAR_CLEAN objects is not
892 on the mutable list; a MVAR_DIRTY is. When written to, a
893 MVAR_CLEAN turns into a MVAR_DIRTY and is put on the mutable list.
894 The check for MVAR_CLEAN is inlined at the call site for speed,
895 this really does make a difference on concurrency-heavy benchmarks
896 such as Chaneneos and cheap-concurrency.
899 dirty_MVAR(StgRegTable *reg, StgClosure *p)
901 Capability *cap = regTableToCapability(reg);
903 bd = Bdescr((StgPtr)p);
904 if (bd->gen_no > 0) recordMutableCap(p,cap,bd->gen_no);
907 /* -----------------------------------------------------------------------------
908 Allocation functions for GMP.
910 These all use the allocate() interface - we can't have any garbage
911 collection going on during a gmp operation, so we use allocate()
912 which always succeeds. The gmp operations which might need to
913 allocate will ask the storage manager (via doYouWantToGC()) whether
914 a garbage collection is required, in case we get into a loop doing
915 only allocate() style allocation.
916 -------------------------------------------------------------------------- */
919 stgAllocForGMP (size_t size_in_bytes)
922 nat data_size_in_words, total_size_in_words;
924 /* round up to a whole number of words */
925 data_size_in_words = (size_in_bytes + sizeof(W_) + 1) / sizeof(W_);
926 total_size_in_words = sizeofW(StgArrWords) + data_size_in_words;
928 /* allocate and fill it in. */
929 #if defined(THREADED_RTS)
930 arr = (StgArrWords *)allocateLocal(myTask()->cap, total_size_in_words);
932 arr = (StgArrWords *)allocateLocal(&MainCapability, total_size_in_words);
934 SET_ARR_HDR(arr, &stg_ARR_WORDS_info, CCCS, data_size_in_words);
936 /* and return a ptr to the goods inside the array */
941 stgReallocForGMP (void *ptr, size_t old_size, size_t new_size)
944 void *new_stuff_ptr = stgAllocForGMP(new_size);
946 char *p = (char *) ptr;
947 char *q = (char *) new_stuff_ptr;
949 min_size = old_size < new_size ? old_size : new_size;
950 for (; i < min_size; i++, p++, q++) {
954 return(new_stuff_ptr);
958 stgDeallocForGMP (void *ptr STG_UNUSED,
959 size_t size STG_UNUSED)
961 /* easy for us: the garbage collector does the dealloc'n */
964 /* -----------------------------------------------------------------------------
966 * -------------------------------------------------------------------------- */
968 /* -----------------------------------------------------------------------------
971 * Approximate how much we've allocated: number of blocks in the
972 * nursery + blocks allocated via allocate() - unused nusery blocks.
973 * This leaves a little slop at the end of each block, and doesn't
974 * take into account large objects (ToDo).
975 * -------------------------------------------------------------------------- */
978 calcAllocated( void )
983 allocated = allocatedBytes();
984 allocated += countNurseryBlocks() * BLOCK_SIZE_W;
989 for (i = 0; i < n_nurseries; i++) {
991 for ( bd = capabilities[i].r.rCurrentNursery->link;
992 bd != NULL; bd = bd->link ) {
993 allocated -= BLOCK_SIZE_W;
995 cap = &capabilities[i];
996 if (cap->r.rCurrentNursery->free <
997 cap->r.rCurrentNursery->start + BLOCK_SIZE_W) {
998 allocated -= (cap->r.rCurrentNursery->start + BLOCK_SIZE_W)
999 - cap->r.rCurrentNursery->free;
1003 bdescr *current_nursery = MainCapability.r.rCurrentNursery;
1005 for ( bd = current_nursery->link; bd != NULL; bd = bd->link ) {
1006 allocated -= BLOCK_SIZE_W;
1008 if (current_nursery->free < current_nursery->start + BLOCK_SIZE_W) {
1009 allocated -= (current_nursery->start + BLOCK_SIZE_W)
1010 - current_nursery->free;
1015 total_allocated += allocated;
1019 /* Approximate the amount of live data in the heap. To be called just
1020 * after garbage collection (see GarbageCollect()).
1023 calcLiveBlocks(void)
1029 if (RtsFlags.GcFlags.generations == 1) {
1030 return g0s0->n_large_blocks + g0s0->n_blocks;
1033 for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
1034 for (s = 0; s < generations[g].n_steps; s++) {
1035 /* approximate amount of live data (doesn't take into account slop
1036 * at end of each block).
1038 if (g == 0 && s == 0) {
1041 stp = &generations[g].steps[s];
1042 live += stp->n_large_blocks + stp->n_blocks;
1049 countOccupied(bdescr *bd)
1054 for (; bd != NULL; bd = bd->link) {
1055 ASSERT(bd->free <= bd->start + bd->blocks * BLOCK_SIZE_W);
1056 words += bd->free - bd->start;
1061 // Return an accurate count of the live data in the heap, excluding
1070 if (RtsFlags.GcFlags.generations == 1) {
1071 return g0s0->n_words + countOccupied(g0s0->large_objects);
1075 for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
1076 for (s = 0; s < generations[g].n_steps; s++) {
1077 if (g == 0 && s == 0) continue;
1078 stp = &generations[g].steps[s];
1079 live += stp->n_words + countOccupied(stp->large_objects);
1085 /* Approximate the number of blocks that will be needed at the next
1086 * garbage collection.
1088 * Assume: all data currently live will remain live. Steps that will
1089 * be collected next time will therefore need twice as many blocks
1090 * since all the data will be copied.
1099 for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
1100 for (s = 0; s < generations[g].n_steps; s++) {
1101 if (g == 0 && s == 0) { continue; }
1102 stp = &generations[g].steps[s];
1104 // we need at least this much space
1105 needed += stp->n_blocks + stp->n_large_blocks;
1107 // any additional space needed to collect this gen next time?
1108 if (g == 0 || // always collect gen 0
1109 (generations[g].steps[0].n_blocks +
1110 generations[g].steps[0].n_large_blocks
1111 > generations[g].max_blocks)) {
1112 // we will collect this gen next time
1115 needed += stp->n_blocks / BITS_IN(W_);
1117 needed += stp->n_blocks / 100;
1120 continue; // no additional space needed for compaction
1122 needed += stp->n_blocks;
1130 /* ----------------------------------------------------------------------------
1133 Executable memory must be managed separately from non-executable
1134 memory. Most OSs these days require you to jump through hoops to
1135 dynamically allocate executable memory, due to various security
1138 Here we provide a small memory allocator for executable memory.
1139 Memory is managed with a page granularity; we allocate linearly
1140 in the page, and when the page is emptied (all objects on the page
1141 are free) we free the page again, not forgetting to make it
1144 TODO: The inability to handle objects bigger than BLOCK_SIZE_W means that
1145 the linker cannot use allocateExec for loading object code files
1146 on Windows. Once allocateExec can handle larger objects, the linker
1147 should be modified to use allocateExec instead of VirtualAlloc.
1148 ------------------------------------------------------------------------- */
1150 #if defined(linux_HOST_OS)
1152 // On Linux we need to use libffi for allocating executable memory,
1153 // because it knows how to work around the restrictions put in place
1156 void *allocateExec (nat bytes, void **exec_ret)
1160 ret = ffi_closure_alloc (sizeof(void *) + (size_t)bytes, (void**)&exec);
1162 if (ret == NULL) return ret;
1163 *ret = ret; // save the address of the writable mapping, for freeExec().
1164 *exec_ret = exec + 1;
1168 // freeExec gets passed the executable address, not the writable address.
1169 void freeExec (void *addr)
1172 writable = *((void**)addr - 1);
1174 ffi_closure_free (writable);
1180 void *allocateExec (nat bytes, void **exec_ret)
1187 // round up to words.
1188 n = (bytes + sizeof(W_) + 1) / sizeof(W_);
1190 if (n+1 > BLOCK_SIZE_W) {
1191 barf("allocateExec: can't handle large objects");
1194 if (exec_block == NULL ||
1195 exec_block->free + n + 1 > exec_block->start + BLOCK_SIZE_W) {
1197 lnat pagesize = getPageSize();
1198 bd = allocGroup(stg_max(1, pagesize / BLOCK_SIZE));
1199 debugTrace(DEBUG_gc, "allocate exec block %p", bd->start);
1201 bd->flags = BF_EXEC;
1202 bd->link = exec_block;
1203 if (exec_block != NULL) {
1204 exec_block->u.back = bd;
1207 setExecutable(bd->start, bd->blocks * BLOCK_SIZE, rtsTrue);
1210 *(exec_block->free) = n; // store the size of this chunk
1211 exec_block->gen_no += n; // gen_no stores the number of words allocated
1212 ret = exec_block->free + 1;
1213 exec_block->free += n + 1;
1220 void freeExec (void *addr)
1222 StgPtr p = (StgPtr)addr - 1;
1223 bdescr *bd = Bdescr((StgPtr)p);
1225 if ((bd->flags & BF_EXEC) == 0) {
1226 barf("freeExec: not executable");
1229 if (*(StgPtr)p == 0) {
1230 barf("freeExec: already free?");
1235 bd->gen_no -= *(StgPtr)p;
1238 if (bd->gen_no == 0) {
1239 // Free the block if it is empty, but not if it is the block at
1240 // the head of the queue.
1241 if (bd != exec_block) {
1242 debugTrace(DEBUG_gc, "free exec block %p", bd->start);
1243 dbl_link_remove(bd, &exec_block);
1244 setExecutable(bd->start, bd->blocks * BLOCK_SIZE, rtsFalse);
1247 bd->free = bd->start;
1254 #endif /* mingw32_HOST_OS */
1256 /* -----------------------------------------------------------------------------
1259 memInventory() checks for memory leaks by counting up all the
1260 blocks we know about and comparing that to the number of blocks
1261 allegedly floating around in the system.
1262 -------------------------------------------------------------------------- */
1266 // Useful for finding partially full blocks in gdb
1267 void findSlop(bdescr *bd);
1268 void findSlop(bdescr *bd)
1272 for (; bd != NULL; bd = bd->link) {
1273 slop = (bd->blocks * BLOCK_SIZE_W) - (bd->free - bd->start);
1274 if (slop > (1024/sizeof(W_))) {
1275 debugBelch("block at %p (bdescr %p) has %ldKB slop\n",
1276 bd->start, bd, slop / (1024/sizeof(W_)));
1282 countBlocks(bdescr *bd)
1285 for (n=0; bd != NULL; bd=bd->link) {
1291 // (*1) Just like countBlocks, except that we adjust the count for a
1292 // megablock group so that it doesn't include the extra few blocks
1293 // that would be taken up by block descriptors in the second and
1294 // subsequent megablock. This is so we can tally the count with the
1295 // number of blocks allocated in the system, for memInventory().
1297 countAllocdBlocks(bdescr *bd)
1300 for (n=0; bd != NULL; bd=bd->link) {
1302 // hack for megablock groups: see (*1) above
1303 if (bd->blocks > BLOCKS_PER_MBLOCK) {
1304 n -= (MBLOCK_SIZE / BLOCK_SIZE - BLOCKS_PER_MBLOCK)
1305 * (bd->blocks/(MBLOCK_SIZE/BLOCK_SIZE));
1312 stepBlocks (step *stp)
1314 ASSERT(countBlocks(stp->blocks) == stp->n_blocks);
1315 ASSERT(countBlocks(stp->large_objects) == stp->n_large_blocks);
1316 return stp->n_blocks + stp->n_old_blocks +
1317 countAllocdBlocks(stp->large_objects);
1320 // If memInventory() calculates that we have a memory leak, this
1321 // function will try to find the block(s) that are leaking by marking
1322 // all the ones that we know about, and search through memory to find
1323 // blocks that are not marked. In the debugger this can help to give
1324 // us a clue about what kind of block leaked. In the future we might
1325 // annotate blocks with their allocation site to give more helpful
1328 findMemoryLeak (void)
1331 for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
1332 for (i = 0; i < n_capabilities; i++) {
1333 markBlocks(capabilities[i].mut_lists[g]);
1335 markBlocks(generations[g].mut_list);
1336 for (s = 0; s < generations[g].n_steps; s++) {
1337 markBlocks(generations[g].steps[s].blocks);
1338 markBlocks(generations[g].steps[s].large_objects);
1342 for (i = 0; i < n_nurseries; i++) {
1343 markBlocks(nurseries[i].blocks);
1344 markBlocks(nurseries[i].large_objects);
1349 // if (RtsFlags.ProfFlags.doHeapProfile == HEAP_BY_RETAINER) {
1350 // markRetainerBlocks();
1354 // count the blocks allocated by the arena allocator
1356 // markArenaBlocks();
1358 // count the blocks containing executable memory
1359 markBlocks(exec_block);
1361 reportUnmarkedBlocks();
1366 memInventory (rtsBool show)
1370 lnat gen_blocks[RtsFlags.GcFlags.generations];
1371 lnat nursery_blocks, retainer_blocks,
1372 arena_blocks, exec_blocks;
1373 lnat live_blocks = 0, free_blocks = 0;
1376 // count the blocks we current have
1378 for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
1380 for (i = 0; i < n_capabilities; i++) {
1381 gen_blocks[g] += countBlocks(capabilities[i].mut_lists[g]);
1383 gen_blocks[g] += countAllocdBlocks(generations[g].mut_list);
1384 for (s = 0; s < generations[g].n_steps; s++) {
1385 stp = &generations[g].steps[s];
1386 gen_blocks[g] += stepBlocks(stp);
1391 for (i = 0; i < n_nurseries; i++) {
1392 nursery_blocks += stepBlocks(&nurseries[i]);
1395 retainer_blocks = 0;
1397 if (RtsFlags.ProfFlags.doHeapProfile == HEAP_BY_RETAINER) {
1398 retainer_blocks = retainerStackBlocks();
1402 // count the blocks allocated by the arena allocator
1403 arena_blocks = arenaBlocks();
1405 // count the blocks containing executable memory
1406 exec_blocks = countAllocdBlocks(exec_block);
1408 /* count the blocks on the free list */
1409 free_blocks = countFreeList();
1412 for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
1413 live_blocks += gen_blocks[g];
1415 live_blocks += nursery_blocks +
1416 + retainer_blocks + arena_blocks + exec_blocks;
1418 #define MB(n) (((n) * BLOCK_SIZE_W) / ((1024*1024)/sizeof(W_)))
1420 leak = live_blocks + free_blocks != mblocks_allocated * BLOCKS_PER_MBLOCK;
1425 debugBelch("Memory leak detected:\n");
1427 debugBelch("Memory inventory:\n");
1429 for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
1430 debugBelch(" gen %d blocks : %5lu blocks (%lu MB)\n", g,
1431 gen_blocks[g], MB(gen_blocks[g]));
1433 debugBelch(" nursery : %5lu blocks (%lu MB)\n",
1434 nursery_blocks, MB(nursery_blocks));
1435 debugBelch(" retainer : %5lu blocks (%lu MB)\n",
1436 retainer_blocks, MB(retainer_blocks));
1437 debugBelch(" arena blocks : %5lu blocks (%lu MB)\n",
1438 arena_blocks, MB(arena_blocks));
1439 debugBelch(" exec : %5lu blocks (%lu MB)\n",
1440 exec_blocks, MB(exec_blocks));
1441 debugBelch(" free : %5lu blocks (%lu MB)\n",
1442 free_blocks, MB(free_blocks));
1443 debugBelch(" total : %5lu blocks (%lu MB)\n",
1444 live_blocks + free_blocks, MB(live_blocks+free_blocks));
1446 debugBelch("\n in system : %5lu blocks (%lu MB)\n",
1447 mblocks_allocated * BLOCKS_PER_MBLOCK, mblocks_allocated);
1455 ASSERT(n_alloc_blocks == live_blocks);
1460 /* Full heap sanity check. */
1466 if (RtsFlags.GcFlags.generations == 1) {
1467 checkHeap(g0s0->blocks);
1468 checkLargeObjects(g0s0->large_objects);
1471 for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
1472 for (s = 0; s < generations[g].n_steps; s++) {
1473 if (g == 0 && s == 0) { continue; }
1474 ASSERT(countBlocks(generations[g].steps[s].blocks)
1475 == generations[g].steps[s].n_blocks);
1476 ASSERT(countBlocks(generations[g].steps[s].large_objects)
1477 == generations[g].steps[s].n_large_blocks);
1478 checkHeap(generations[g].steps[s].blocks);
1479 checkLargeObjects(generations[g].steps[s].large_objects);
1483 for (s = 0; s < n_nurseries; s++) {
1484 ASSERT(countBlocks(nurseries[s].blocks)
1485 == nurseries[s].n_blocks);
1486 ASSERT(countBlocks(nurseries[s].large_objects)
1487 == nurseries[s].n_large_blocks);
1490 checkFreeListSanity();
1493 #if defined(THREADED_RTS)
1494 // check the stacks too in threaded mode, because we don't do a
1495 // full heap sanity check in this case (see checkHeap())
1496 checkMutableLists(rtsTrue);
1498 checkMutableLists(rtsFalse);
1502 /* Nursery sanity check */
1504 checkNurserySanity( step *stp )
1510 for (bd = stp->blocks; bd != NULL; bd = bd->link) {
1511 ASSERT(bd->u.back == prev);
1513 blocks += bd->blocks;
1515 ASSERT(blocks == stp->n_blocks);
1518 // handy function for use in gdb, because Bdescr() is inlined.
1519 extern bdescr *_bdescr( StgPtr p );