X-Git-Url: http://git.megacz.com/?a=blobdiff_plain;f=ghc%2Frts%2FStorage.c;h=45d94aea01f38fd3c771f4748175b0f15bca1dce;hb=abde5fdfcaa5f363d9cf8cf0a10aa281c40887fe;hp=1f080c58a75493d08b3d2293428d67e033fbf918;hpb=f939293c9e2b8eead06935196500734b1240aee8;p=ghc-hetmet.git diff --git a/ghc/rts/Storage.c b/ghc/rts/Storage.c index 1f080c5..45d94ae 100644 --- a/ghc/rts/Storage.c +++ b/ghc/rts/Storage.c @@ -1,10 +1,12 @@ /* ----------------------------------------------------------------------------- - * $Id: Storage.c,v 1.8 1999/01/28 15:04:02 simonm Exp $ + * + * (c) The GHC Team, 1998-2004 * * Storage manager front end * * ---------------------------------------------------------------------------*/ +#include "PosixSource.h" #include "Rts.h" #include "RtsUtils.h" #include "RtsFlags.h" @@ -12,30 +14,45 @@ #include "Hooks.h" #include "BlockAlloc.h" #include "MBlock.h" -#include "gmp.h" #include "Weak.h" #include "Sanity.h" +#include "Arena.h" #include "Storage.h" -#include "StoragePriv.h" +#include "Schedule.h" +#include "OSThreads.h" + +#include "RetainerProfile.h" // for counting memory blocks (memInventory) -bdescr *current_nursery; /* next available nursery block, or NULL */ -nat nursery_blocks; /* number of blocks in the nursery */ +#include +#include StgClosure *caf_list = NULL; +StgClosure *revertible_caf_list = NULL; +rtsBool keepCAFs; bdescr *small_alloc_list; /* allocate()d small objects */ -bdescr *large_alloc_list; /* allocate()d large objects */ +bdescr *pinned_object_block; /* allocate pinned objects into this block */ nat alloc_blocks; /* number of allocate()d blocks since GC */ nat alloc_blocks_lim; /* approximate limit on alloc_blocks */ StgPtr alloc_Hp = NULL; /* next free byte in small_alloc_list */ StgPtr alloc_HpLim = NULL; /* end of block at small_alloc_list */ -generation *generations; /* all the generations */ -generation *g0; /* generation 0, for convenience */ -generation *oldest_gen; /* oldest generation, for convenience */ -step *g0s0; /* generation 0, step 0, for convenience */ +generation *generations = NULL; /* all the generations */ +generation *g0 = NULL; /* generation 0, for convenience */ +generation *oldest_gen = NULL; /* oldest generation, for convenience */ +step *g0s0 = NULL; /* generation 0, step 0, for convenience */ + +ullong total_allocated = 0; /* total memory allocated during run */ + +/* + * Storage manager mutex: protects all the above state from + * simultaneous access by two STG threads. + */ +#ifdef SMP +Mutex sm_mutex = INIT_MUTEX_VAR; +#endif /* * Forward references @@ -45,19 +62,43 @@ static void *stgReallocForGMP (void *ptr, size_t old_size, size_t new_size); static void stgDeallocForGMP (void *ptr, size_t size); void -initStorage (void) +initStorage( void ) { nat g, s; - step *step; + step *stp; generation *gen; - if (RtsFlags.GcFlags.heapSizeSuggestion > + if (generations != NULL) { + // multi-init protection + return; + } + + /* Sanity check to make sure the LOOKS_LIKE_ macros appear to be + * doing something reasonable. + */ + ASSERT(LOOKS_LIKE_INFO_PTR(&stg_BLACKHOLE_info)); + ASSERT(LOOKS_LIKE_CLOSURE_PTR(&stg_dummy_ret_closure)); + ASSERT(!HEAP_ALLOCED(&stg_dummy_ret_closure)); + + if (RtsFlags.GcFlags.maxHeapSize != 0 && + RtsFlags.GcFlags.heapSizeSuggestion > + RtsFlags.GcFlags.maxHeapSize) { + RtsFlags.GcFlags.maxHeapSize = RtsFlags.GcFlags.heapSizeSuggestion; + } + + if (RtsFlags.GcFlags.maxHeapSize != 0 && + RtsFlags.GcFlags.minAllocAreaSize > RtsFlags.GcFlags.maxHeapSize) { - barf("Suggested heap size (-H) is larger than max. heap size (-M)\n"); + errorBelch("maximum heap size (-M) is smaller than minimum alloc area size (-A)"); + exit(1); } initBlockAllocator(); +#if defined(SMP) + initMutex(&sm_mutex); +#endif + /* allocate generation info array */ generations = (generation *)stgMallocBytes(RtsFlags.GcFlags.generations * sizeof(struct _generation), @@ -67,10 +108,10 @@ initStorage (void) for(g = 0; g < RtsFlags.GcFlags.generations; g++) { gen = &generations[g]; gen->no = g; - gen->mut_list = END_MUT_LIST; + gen->mut_list = allocBlock(); gen->collections = 0; gen->failed_promotions = 0; - gen->max_blocks = RtsFlags.GcFlags.minOldGenSize; + gen->max_blocks = 0; } /* A couple of convenience pointers */ @@ -103,17 +144,25 @@ initStorage (void) /* Initialise all steps */ for (g = 0; g < RtsFlags.GcFlags.generations; g++) { for (s = 0; s < generations[g].n_steps; s++) { - step = &generations[g].steps[s]; - step->no = s; - step->blocks = NULL; - step->n_blocks = 0; - step->gen = &generations[g]; - step->hp = NULL; - step->hpLim = NULL; - step->hp_bd = NULL; - step->large_objects = NULL; - step->new_large_objects = NULL; - step->scavenged_large_objects = NULL; + stp = &generations[g].steps[s]; + stp->no = s; + stp->blocks = NULL; + stp->n_to_blocks = 0; + stp->n_blocks = 0; + stp->gen = &generations[g]; + stp->gen_no = g; + stp->hp = NULL; + stp->hpLim = NULL; + stp->hp_bd = NULL; + stp->scan = NULL; + stp->scan_bd = NULL; + stp->large_objects = NULL; + stp->n_large_blocks = 0; + stp->new_large_objects = NULL; + stp->scavenged_large_objects = NULL; + stp->n_scavenged_large_blocks = 0; + stp->is_compacted = 0; + stp->bitmap = NULL; } } @@ -125,70 +174,272 @@ initStorage (void) generations[g].steps[s].to = &generations[g+1].steps[0]; } - /* The oldest generation has one step and its destination is the - * same step. */ + /* The oldest generation has one step and it is compacted. */ + if (RtsFlags.GcFlags.compact) { + if (RtsFlags.GcFlags.generations == 1) { + errorBelch("WARNING: compaction is incompatible with -G1; disabled"); + } else { + oldest_gen->steps[0].is_compacted = 1; + } + } oldest_gen->steps[0].to = &oldest_gen->steps[0]; /* generation 0 is special: that's the nursery */ generations[0].max_blocks = 0; - /* G0S0: the allocation area */ - step = &generations[0].steps[0]; - g0s0 = step; - step->blocks = allocNursery(NULL, RtsFlags.GcFlags.minAllocAreaSize); - step->n_blocks = RtsFlags.GcFlags.minAllocAreaSize; - nursery_blocks = RtsFlags.GcFlags.minAllocAreaSize; - current_nursery = step->blocks; - /* hp, hpLim, hp_bd, to_space etc. aren't used in G0S0 */ + /* G0S0: the allocation area. Policy: keep the allocation area + * small to begin with, even if we have a large suggested heap + * size. Reason: we're going to do a major collection first, and we + * don't want it to be a big one. This vague idea is borne out by + * rigorous experimental evidence. + */ + g0s0 = &generations[0].steps[0]; + + allocNurseries(); weak_ptr_list = NULL; caf_list = NULL; + revertible_caf_list = NULL; /* initialise the allocate() interface */ small_alloc_list = NULL; - large_alloc_list = NULL; alloc_blocks = 0; alloc_blocks_lim = RtsFlags.GcFlags.minAllocAreaSize; -#ifdef COMPILER /* Tell GNU multi-precision pkg about our custom alloc functions */ mp_set_memory_functions(stgAllocForGMP, stgReallocForGMP, stgDeallocForGMP); -#endif - IF_DEBUG(gc, stat_describe_gens()); + IF_DEBUG(gc, statDescribeGens()); +} + +void +exitStorage (void) +{ + stat_exit(calcAllocated()); +} + +/* ----------------------------------------------------------------------------- + CAF management. + + The entry code for every CAF does the following: + + - builds a CAF_BLACKHOLE in the heap + - pushes an update frame pointing to the CAF_BLACKHOLE + - invokes UPD_CAF(), which: + - calls newCaf, below + - updates the CAF with a static indirection to the CAF_BLACKHOLE + + Why do we build a BLACKHOLE in the heap rather than just updating + the thunk directly? It's so that we only need one kind of update + frame - otherwise we'd need a static version of the update frame too. + + newCaf() does the following: + + - it puts the CAF on the oldest generation's mut-once list. + This is so that we can treat the CAF as a root when collecting + younger generations. + + For GHCI, we have additional requirements when dealing with CAFs: + + - we must *retain* all dynamically-loaded CAFs ever entered, + just in case we need them again. + - we must be able to *revert* CAFs that have been evaluated, to + their pre-evaluated form. + + To do this, we use an additional CAF list. When newCaf() is + called on a dynamically-loaded CAF, we add it to the CAF list + instead of the old-generation mutable list, and save away its + old info pointer (in caf->saved_info) for later reversion. + + To revert all the CAFs, we traverse the CAF list and reset the + info pointer to caf->saved_info, then throw away the CAF list. + (see GC.c:revertCAFs()). + + -- SDM 29/1/01 + + -------------------------------------------------------------------------- */ + +void +newCAF(StgClosure* caf) +{ + ACQUIRE_SM_LOCK; + + if(keepCAFs) + { + // HACK: + // If we are in GHCi _and_ we are using dynamic libraries, + // then we can't redirect newCAF calls to newDynCAF (see below), + // so we make newCAF behave almost like newDynCAF. + // The dynamic libraries might be used by both the interpreted + // program and GHCi itself, so they must not be reverted. + // This also means that in GHCi with dynamic libraries, CAFs are not + // garbage collected. If this turns out to be a problem, we could + // do another hack here and do an address range test on caf to figure + // out whether it is from a dynamic library. + ((StgIndStatic *)caf)->saved_info = (StgInfoTable *)caf->header.info; + ((StgIndStatic *)caf)->static_link = caf_list; + caf_list = caf; + } + else + { + /* Put this CAF on the mutable list for the old generation. + * This is a HACK - the IND_STATIC closure doesn't really have + * a mut_link field, but we pretend it has - in fact we re-use + * the STATIC_LINK field for the time being, because when we + * come to do a major GC we won't need the mut_link field + * any more and can use it as a STATIC_LINK. + */ + ((StgIndStatic *)caf)->saved_info = NULL; + recordMutableGen(caf, oldest_gen); + } + + RELEASE_SM_LOCK; + +#ifdef PAR + /* If we are PAR or DIST then we never forget a CAF */ + { globalAddr *newGA; + //debugBelch("<##> Globalising CAF %08x %s",caf,info_type(caf)); + newGA=makeGlobal(caf,rtsTrue); /*given full weight*/ + ASSERT(newGA); + } +#endif /* PAR */ +} + +// An alternate version of newCaf which is used for dynamically loaded +// object code in GHCi. In this case we want to retain *all* CAFs in +// the object code, because they might be demanded at any time from an +// expression evaluated on the command line. +// Also, GHCi might want to revert CAFs, so we add these to the +// revertible_caf_list. +// +// The linker hackily arranges that references to newCaf from dynamic +// code end up pointing to newDynCAF. +void +newDynCAF(StgClosure *caf) +{ + ACQUIRE_SM_LOCK; + + ((StgIndStatic *)caf)->saved_info = (StgInfoTable *)caf->header.info; + ((StgIndStatic *)caf)->static_link = revertible_caf_list; + revertible_caf_list = caf; + + RELEASE_SM_LOCK; +} + +/* ----------------------------------------------------------------------------- + Nursery management. + -------------------------------------------------------------------------- */ + +void +allocNurseries( void ) +{ +#ifdef SMP + Capability *cap; + bdescr *bd; + + g0s0->blocks = NULL; + g0s0->n_blocks = 0; + for (cap = free_capabilities; cap != NULL; cap = cap->link) { + cap->r.rNursery = allocNursery(NULL, RtsFlags.GcFlags.minAllocAreaSize); + cap->r.rCurrentNursery = cap->r.rNursery; + /* Set the back links to be equal to the Capability, + * so we can do slightly better informed locking. + */ + for (bd = cap->r.rNursery; bd != NULL; bd = bd->link) { + bd->u.back = (bdescr *)cap; + } + } +#else /* SMP */ + g0s0->blocks = allocNursery(NULL, RtsFlags.GcFlags.minAllocAreaSize); + g0s0->n_blocks = RtsFlags.GcFlags.minAllocAreaSize; + g0s0->to_blocks = NULL; + g0s0->n_to_blocks = 0; + MainCapability.r.rNursery = g0s0->blocks; + MainCapability.r.rCurrentNursery = g0s0->blocks; + /* hp, hpLim, hp_bd, to_space etc. aren't used in G0S0 */ +#endif +} + +void +resetNurseries( void ) +{ + bdescr *bd; +#ifdef SMP + Capability *cap; + + /* All tasks must be stopped */ + ASSERT(n_free_capabilities == RtsFlags.ParFlags.nNodes); + + for (cap = free_capabilities; cap != NULL; cap = cap->link) { + for (bd = cap->r.rNursery; bd; bd = bd->link) { + bd->free = bd->start; + ASSERT(bd->gen_no == 0); + ASSERT(bd->step == g0s0); + IF_DEBUG(sanity,memset(bd->start, 0xaa, BLOCK_SIZE)); + } + cap->r.rCurrentNursery = cap->r.rNursery; + } +#else + for (bd = g0s0->blocks; bd; bd = bd->link) { + bd->free = bd->start; + ASSERT(bd->gen_no == 0); + ASSERT(bd->step == g0s0); + IF_DEBUG(sanity,memset(bd->start, 0xaa, BLOCK_SIZE)); + } + MainCapability.r.rNursery = g0s0->blocks; + MainCapability.r.rCurrentNursery = g0s0->blocks; +#endif } -extern bdescr * -allocNursery (bdescr *last_bd, nat blocks) +bdescr * +allocNursery (bdescr *tail, nat blocks) { bdescr *bd; nat i; - /* Allocate a nursery */ + // Allocate a nursery: we allocate fresh blocks one at a time and + // cons them on to the front of the list, not forgetting to update + // the back pointer on the tail of the list to point to the new block. for (i=0; i < blocks; i++) { + // @LDV profiling + /* + processNursery() in LdvProfile.c assumes that every block group in + the nursery contains only a single block. So, if a block group is + given multiple blocks, change processNursery() accordingly. + */ bd = allocBlock(); - bd->link = last_bd; + bd->link = tail; + // double-link the nursery: we might need to insert blocks + if (tail != NULL) { + tail->u.back = bd; + } bd->step = g0s0; - bd->gen = g0; - bd->evacuated = 0; + bd->gen_no = 0; + bd->flags = 0; bd->free = bd->start; - last_bd = bd; + tail = bd; } - return last_bd; + tail->u.back = NULL; + return tail; } -extern void +void resizeNursery ( nat blocks ) { bdescr *bd; + nat nursery_blocks; +#ifdef SMP + barf("resizeNursery: can't resize in SMP mode"); +#endif + + nursery_blocks = g0s0->n_blocks; if (nursery_blocks == blocks) { - ASSERT(g0s0->n_blocks == blocks); return; } else if (nursery_blocks < blocks) { - IF_DEBUG(gc, fprintf(stderr, "Increasing size of nursery to %d blocks\n", + IF_DEBUG(gc, debugBelch("Increasing size of nursery to %d blocks\n", blocks)); g0s0->blocks = allocNursery(g0s0->blocks, blocks-nursery_blocks); } @@ -196,77 +447,27 @@ resizeNursery ( nat blocks ) else { bdescr *next_bd; - IF_DEBUG(gc, fprintf(stderr, "Decreasing size of nursery to %d blocks\n", + IF_DEBUG(gc, debugBelch("Decreasing size of nursery to %d blocks\n", blocks)); - for (bd = g0s0->blocks; nursery_blocks > blocks; nursery_blocks--) { - next_bd = bd->link; - freeGroup(bd); - bd = next_bd; + + bd = g0s0->blocks; + while (nursery_blocks > blocks) { + next_bd = bd->link; + next_bd->u.back = NULL; + nursery_blocks -= bd->blocks; // might be a large block + freeGroup(bd); + bd = next_bd; } g0s0->blocks = bd; + // might have gone just under, by freeing a large block, so make + // up the difference. + if (nursery_blocks < blocks) { + g0s0->blocks = allocNursery(g0s0->blocks, blocks-nursery_blocks); + } } - g0s0->n_blocks = nursery_blocks = blocks; -} - -void -exitStorage (void) -{ - lnat allocated; - bdescr *bd; - - /* Return code ignored for now */ - /* ToDo: allocation figure is slightly wrong (see also GarbageCollect()) */ - allocated = (nursery_blocks * BLOCK_SIZE_W) + allocated_bytes(); - for ( bd = current_nursery->link; bd != NULL; bd = bd->link ) { - allocated -= BLOCK_SIZE_W; - } - stat_exit(allocated); -} - -void -recordMutable(StgMutClosure *p) -{ - bdescr *bd; - - ASSERT(closure_MUTABLE(p)); - - bd = Bdescr((P_)p); - - /* no need to bother in generation 0 */ - if (bd->gen == g0) { - return; - } - - if (p->mut_link == NULL) { - p->mut_link = bd->gen->mut_list; - bd->gen->mut_list = p; - } -} - -void -newCAF(StgClosure* caf) -{ - /* Put this CAF on the mutable list for the old generation. - * This is a HACK - the IND_STATIC closure doesn't really have - * a mut_link field, but we pretend it has - in fact we re-use - * the STATIC_LINK field for the time being, because when we - * come to do a major GC we won't need the mut_link field - * any more and can use it as a STATIC_LINK. - */ - ((StgMutClosure *)caf)->mut_link = oldest_gen->mut_list; - oldest_gen->mut_list = (StgMutClosure *)caf; - -#ifdef DEBUG - { - const StgInfoTable *info; - - info = get_itbl(caf); - ASSERT(info->type == IND_STATIC); - STATIC_LINK2(info,caf) = caf_list; - caf_list = caf; - } -#endif + g0s0->n_blocks = blocks; + ASSERT(countBlocks(g0s0->blocks) == g0s0->n_blocks); } /* ----------------------------------------------------------------------------- @@ -278,12 +479,14 @@ newCAF(StgClosure* caf) -------------------------------------------------------------------------- */ StgPtr -allocate(nat n) +allocate( nat n ) { bdescr *bd; StgPtr p; - TICK_ALLOC_HEAP(n); + ACQUIRE_SM_LOCK; + + TICK_ALLOC_HEAP_NOCTR(n); CCS_ALLOC(CCCS,n); /* big allocation (>LARGE_OBJECT_THRESHOLD) */ @@ -292,15 +495,13 @@ allocate(nat n) nat req_blocks = (lnat)BLOCK_ROUND_UP(n*sizeof(W_)) / BLOCK_SIZE; bd = allocGroup(req_blocks); dbl_link_onto(bd, &g0s0->large_objects); - bd->gen = g0; + g0s0->n_large_blocks += req_blocks; + bd->gen_no = 0; bd->step = g0s0; - bd->evacuated = 0; - bd->free = bd->start; - /* don't add these blocks to alloc_blocks, since we're assuming - * that large objects are likely to remain live for quite a while - * (eg. running threads), so garbage collecting early won't make - * much difference. - */ + bd->flags = BF_LARGE; + bd->free = bd->start + n; + alloc_blocks += req_blocks; + RELEASE_SM_LOCK; return bd->start; /* small allocation (link = small_alloc_list; small_alloc_list = bd; - bd->gen = g0; + bd->gen_no = 0; bd->step = g0s0; - bd->evacuated = 0; + bd->flags = 0; alloc_Hp = bd->start; alloc_HpLim = bd->start + BLOCK_SIZE_W; alloc_blocks++; } - + p = alloc_Hp; alloc_Hp += n; + RELEASE_SM_LOCK; return p; } -lnat allocated_bytes(void) +lnat +allocated_bytes( void ) { - return (alloc_blocks * BLOCK_SIZE_W - (alloc_HpLim - alloc_Hp)); + lnat allocated; + + allocated = alloc_blocks * BLOCK_SIZE_W - (alloc_HpLim - alloc_Hp); + if (pinned_object_block != NULL) { + allocated -= (pinned_object_block->start + BLOCK_SIZE_W) - + pinned_object_block->free; + } + + return allocated; +} + +void +tidyAllocateLists (void) +{ + if (small_alloc_list != NULL) { + ASSERT(alloc_Hp >= small_alloc_list->start && + alloc_Hp <= small_alloc_list->start + BLOCK_SIZE); + small_alloc_list->free = alloc_Hp; + } +} + +/* --------------------------------------------------------------------------- + Allocate a fixed/pinned object. + + We allocate small pinned objects into a single block, allocating a + new block when the current one overflows. The block is chained + onto the large_object_list of generation 0 step 0. + + NOTE: The GC can't in general handle pinned objects. This + interface is only safe to use for ByteArrays, which have no + pointers and don't require scavenging. It works because the + block's descriptor has the BF_LARGE flag set, so the block is + treated as a large object and chained onto various lists, rather + than the individual objects being copied. However, when it comes + to scavenge the block, the GC will only scavenge the first object. + The reason is that the GC can't linearly scan a block of pinned + objects at the moment (doing so would require using the + mostly-copying techniques). But since we're restricting ourselves + to pinned ByteArrays, not scavenging is ok. + + This function is called by newPinnedByteArray# which immediately + fills the allocated memory with a MutableByteArray#. + ------------------------------------------------------------------------- */ + +StgPtr +allocatePinned( nat n ) +{ + StgPtr p; + bdescr *bd = pinned_object_block; + + // If the request is for a large object, then allocate() + // will give us a pinned object anyway. + if (n >= LARGE_OBJECT_THRESHOLD/sizeof(W_)) { + return allocate(n); + } + + ACQUIRE_SM_LOCK; + + TICK_ALLOC_HEAP_NOCTR(n); + CCS_ALLOC(CCCS,n); + + // we always return 8-byte aligned memory. bd->free must be + // 8-byte aligned to begin with, so we just round up n to + // the nearest multiple of 8 bytes. + if (sizeof(StgWord) == 4) { + n = (n+1) & ~1; + } + + // If we don't have a block of pinned objects yet, or the current + // one isn't large enough to hold the new object, allocate a new one. + if (bd == NULL || (bd->free + n) > (bd->start + BLOCK_SIZE_W)) { + pinned_object_block = bd = allocBlock(); + dbl_link_onto(bd, &g0s0->large_objects); + bd->gen_no = 0; + bd->step = g0s0; + bd->flags = BF_PINNED | BF_LARGE; + bd->free = bd->start; + alloc_blocks++; + } + + p = bd->free; + bd->free += n; + RELEASE_SM_LOCK; + return p; } /* ----------------------------------------------------------------------------- @@ -346,18 +632,16 @@ stgAllocForGMP (size_t size_in_bytes) StgArrWords* arr; nat data_size_in_words, total_size_in_words; - /* should be a multiple of sizeof(StgWord) (whole no. of limbs) */ - ASSERT(size_in_bytes % sizeof(W_) == 0); - - data_size_in_words = size_in_bytes / sizeof(W_); + /* round up to a whole number of words */ + data_size_in_words = (size_in_bytes + sizeof(W_) + 1) / sizeof(W_); total_size_in_words = sizeofW(StgArrWords) + data_size_in_words; /* allocate and fill it in. */ arr = (StgArrWords *)allocate(total_size_in_words); - SET_ARR_HDR(arr, &ARR_WORDS_info, CCCS, data_size_in_words); + SET_ARR_HDR(arr, &stg_ARR_WORDS_info, CCCS, data_size_in_words); /* and return a ptr to the goods inside the array */ - return(BYTE_ARR_CTS(arr)); + return arr->payload; } static void * @@ -383,8 +667,64 @@ stgDeallocForGMP (void *ptr STG_UNUSED, } /* ----------------------------------------------------------------------------- - Stats and stuff - -------------------------------------------------------------------------- */ + * Stats and stuff + * -------------------------------------------------------------------------- */ + +/* ----------------------------------------------------------------------------- + * calcAllocated() + * + * Approximate how much we've allocated: number of blocks in the + * nursery + blocks allocated via allocate() - unused nusery blocks. + * This leaves a little slop at the end of each block, and doesn't + * take into account large objects (ToDo). + * -------------------------------------------------------------------------- */ + +lnat +calcAllocated( void ) +{ + nat allocated; + bdescr *bd; + +#ifdef SMP + Capability *cap; + + /* All tasks must be stopped. Can't assert that all the + capabilities are owned by the scheduler, though: one or more + tasks might have been stopped while they were running (non-main) + threads. */ + /* ASSERT(n_free_capabilities == RtsFlags.ParFlags.nNodes); */ + + allocated = + n_free_capabilities * RtsFlags.GcFlags.minAllocAreaSize * BLOCK_SIZE_W + + allocated_bytes(); + + for (cap = free_capabilities; cap != NULL; cap = cap->link) { + for ( bd = cap->r.rCurrentNursery->link; bd != NULL; bd = bd->link ) { + allocated -= BLOCK_SIZE_W; + } + if (cap->r.rCurrentNursery->free < cap->r.rCurrentNursery->start + + BLOCK_SIZE_W) { + allocated -= (cap->r.rCurrentNursery->start + BLOCK_SIZE_W) + - cap->r.rCurrentNursery->free; + } + } + +#else /* !SMP */ + bdescr *current_nursery = MainCapability.r.rCurrentNursery; + + allocated = (g0s0->n_blocks * BLOCK_SIZE_W) + allocated_bytes(); + for ( bd = current_nursery->link; bd != NULL; bd = bd->link ) { + allocated -= BLOCK_SIZE_W; + } + if (current_nursery->free < current_nursery->start + BLOCK_SIZE_W) { + allocated -= (current_nursery->start + BLOCK_SIZE_W) + - current_nursery->free; + } +#endif + + total_allocated += allocated; + return allocated; +} /* Approximate the amount of live data in the heap. To be called just * after garbage collection (see GarbageCollect()). @@ -394,24 +734,28 @@ calcLive(void) { nat g, s; lnat live = 0; - step *step; + step *stp; if (RtsFlags.GcFlags.generations == 1) { - live = g0s0->to_blocks * BLOCK_SIZE_W + + live = (g0s0->n_to_blocks - 1) * BLOCK_SIZE_W + ((lnat)g0s0->hp_bd->free - (lnat)g0s0->hp_bd->start) / sizeof(W_); + return live; } for (g = 0; g < RtsFlags.GcFlags.generations; g++) { for (s = 0; s < generations[g].n_steps; s++) { /* approximate amount of live data (doesn't take into account slop - * at end of each block). - */ + * at end of each block). + */ if (g == 0 && s == 0) { continue; } - step = &generations[g].steps[s]; - live += step->n_blocks * BLOCK_SIZE_W + - ((lnat)step->hp_bd->free -(lnat)step->hp_bd->start) / sizeof(W_); + stp = &generations[g].steps[s]; + live += (stp->n_large_blocks + stp->n_blocks - 1) * BLOCK_SIZE_W; + if (stp->hp_bd != NULL) { + live += ((lnat)stp->hp_bd->free - (lnat)stp->hp_bd->start) + / sizeof(W_); + } } } return live; @@ -427,22 +771,25 @@ calcLive(void) extern lnat calcNeeded(void) { - lnat needed = 0; - nat g, s; - step *step; - - for (g = 0; g < RtsFlags.GcFlags.generations; g++) { - for (s = 0; s < generations[g].n_steps; s++) { - if (g == 0 && s == 0) { continue; } - step = &generations[g].steps[s]; - if (generations[g].steps[0].n_blocks > generations[g].max_blocks) { - needed += 2 * step->n_blocks; - } else { - needed += step->n_blocks; - } + lnat needed = 0; + nat g, s; + step *stp; + + for (g = 0; g < RtsFlags.GcFlags.generations; g++) { + for (s = 0; s < generations[g].n_steps; s++) { + if (g == 0 && s == 0) { continue; } + stp = &generations[g].steps[s]; + if (generations[g].steps[0].n_blocks + + generations[g].steps[0].n_large_blocks + > generations[g].max_blocks + && stp->is_compacted == 0) { + needed += 2 * stp->n_blocks; + } else { + needed += stp->n_blocks; + } + } } - } - return needed; + return needed; } /* ----------------------------------------------------------------------------- @@ -455,88 +802,115 @@ calcNeeded(void) #ifdef DEBUG -extern void +void memInventory(void) { nat g, s; - step *step; + step *stp; bdescr *bd; lnat total_blocks = 0, free_blocks = 0; /* count the blocks we current have */ for (g = 0; g < RtsFlags.GcFlags.generations; g++) { - for (s = 0; s < generations[g].n_steps; s++) { - step = &generations[g].steps[s]; - total_blocks += step->n_blocks; - if (RtsFlags.GcFlags.generations == 1) { - /* two-space collector has a to-space too :-) */ - total_blocks += g0s0->to_blocks; + for (bd = generations[g].mut_list; bd != NULL; bd = bd->link) { + total_blocks += bd->blocks; } - for (bd = step->large_objects; bd; bd = bd->link) { - total_blocks += bd->blocks; - /* hack for megablock groups: they have an extra block or two in - the second and subsequent megablocks where the block - descriptors would normally go. - */ - if (bd->blocks > BLOCKS_PER_MBLOCK) { - total_blocks -= (MBLOCK_SIZE / BLOCK_SIZE - BLOCKS_PER_MBLOCK) - * bd->blocks/(MBLOCK_SIZE/BLOCK_SIZE); - } + for (s = 0; s < generations[g].n_steps; s++) { + stp = &generations[g].steps[s]; + total_blocks += stp->n_blocks; + if (RtsFlags.GcFlags.generations == 1) { + /* two-space collector has a to-space too :-) */ + total_blocks += g0s0->n_to_blocks; + } + for (bd = stp->large_objects; bd; bd = bd->link) { + total_blocks += bd->blocks; + /* hack for megablock groups: they have an extra block or two in + the second and subsequent megablocks where the block + descriptors would normally go. + */ + if (bd->blocks > BLOCKS_PER_MBLOCK) { + total_blocks -= (MBLOCK_SIZE / BLOCK_SIZE - BLOCKS_PER_MBLOCK) + * (bd->blocks/(MBLOCK_SIZE/BLOCK_SIZE)); + } + } } - } } /* any blocks held by allocate() */ for (bd = small_alloc_list; bd; bd = bd->link) { total_blocks += bd->blocks; } - for (bd = large_alloc_list; bd; bd = bd->link) { - total_blocks += bd->blocks; + +#ifdef PROFILING + if (RtsFlags.ProfFlags.doHeapProfile == HEAP_BY_RETAINER) { + total_blocks += retainerStackBlocks(); } - +#endif + + // count the blocks allocated by the arena allocator + total_blocks += arenaBlocks(); + /* count the blocks on the free list */ free_blocks = countFreeList(); - ASSERT(total_blocks + free_blocks == mblocks_allocated * BLOCKS_PER_MBLOCK); - -#if 0 if (total_blocks + free_blocks != mblocks_allocated * BLOCKS_PER_MBLOCK) { - fprintf(stderr, "Blocks: %ld live + %ld free = %ld total (%ld around)\n", + debugBelch("Blocks: %ld live + %ld free = %ld total (%ld around)\n", total_blocks, free_blocks, total_blocks + free_blocks, mblocks_allocated * BLOCKS_PER_MBLOCK); } -#endif + + ASSERT(total_blocks + free_blocks == mblocks_allocated * BLOCKS_PER_MBLOCK); } -/* Full heap sanity check. */ -extern void -checkSanity(nat N) +nat +countBlocks(bdescr *bd) { - nat g, s; - - if (RtsFlags.GcFlags.generations == 1) { - checkHeap(g0s0->to_space, NULL); - checkChain(g0s0->large_objects); - } else { - - for (g = 0; g <= N; g++) { - for (s = 0; s < generations[g].n_steps; s++) { - if (g == 0 && s == 0) { continue; } - checkHeap(generations[g].steps[s].blocks, NULL); - } + nat n; + for (n=0; bd != NULL; bd=bd->link) { + n += bd->blocks; } - for (g = N+1; g < RtsFlags.GcFlags.generations; g++) { - for (s = 0; s < generations[g].n_steps; s++) { - checkHeap(generations[g].steps[s].blocks, - generations[g].steps[s].blocks->start); - checkChain(generations[g].steps[s].large_objects); - } + return n; +} + +/* Full heap sanity check. */ +void +checkSanity( void ) +{ + nat g, s; + + if (RtsFlags.GcFlags.generations == 1) { + checkHeap(g0s0->to_blocks); + checkChain(g0s0->large_objects); + } else { + + for (g = 0; g < RtsFlags.GcFlags.generations; g++) { + for (s = 0; s < generations[g].n_steps; s++) { + ASSERT(countBlocks(generations[g].steps[s].blocks) + == generations[g].steps[s].n_blocks); + ASSERT(countBlocks(generations[g].steps[s].large_objects) + == generations[g].steps[s].n_large_blocks); + if (g == 0 && s == 0) { continue; } + checkHeap(generations[g].steps[s].blocks); + checkChain(generations[g].steps[s].large_objects); + if (g > 0) { + checkMutableList(generations[g].mut_list, g); + } + } + } + checkFreeListSanity(); } - checkFreeListSanity(); - } +} + +// handy function for use in gdb, because Bdescr() is inlined. +extern bdescr *_bdescr( StgPtr p ); + +bdescr * +_bdescr( StgPtr p ) +{ + return Bdescr(p); } #endif