X-Git-Url: http://git.megacz.com/?a=blobdiff_plain;f=ghc%2Frts%2FStorage.c;h=367530fcb6a0467b109b9e891b0cc49769e0196b;hb=423d477bfecd490de1449c59325c8776f91d7aac;hp=8cba094f5f5bcb55df74bb00ab2dd2e47823b8b2;hpb=c9477f2a34961d1f1fd1b72f5cc6fa4618f70794;p=ghc-hetmet.git diff --git a/ghc/rts/Storage.c b/ghc/rts/Storage.c index 8cba094..367530f 100644 --- a/ghc/rts/Storage.c +++ b/ghc/rts/Storage.c @@ -1,5 +1,5 @@ /* ----------------------------------------------------------------------------- - * $Id: Storage.c,v 1.50 2001/08/30 10:21:40 simonmar Exp $ + * $Id: Storage.c,v 1.84 2004/08/13 13:11:01 simonmar Exp $ * * (c) The GHC Team, 1998-1999 * @@ -17,19 +17,20 @@ #include "MBlock.h" #include "Weak.h" #include "Sanity.h" +#include "Arena.h" #include "Storage.h" #include "Schedule.h" -#include "StoragePriv.h" +#include "OSThreads.h" -#ifndef SMP -nat nursery_blocks; /* number of blocks in the nursery */ -#endif +#include "RetainerProfile.h" // for counting memory blocks (memInventory) + +#include +#include StgClosure *caf_list = NULL; 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 */ @@ -37,19 +38,19 @@ 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 */ -lnat total_allocated = 0; /* total memory allocated during run */ +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 -pthread_mutex_t sm_mutex = PTHREAD_MUTEX_INITIALIZER; +Mutex sm_mutex = INIT_MUTEX_VAR; #endif /* @@ -66,31 +67,37 @@ initStorage( void ) step *stp; generation *gen; - /* If we're doing heap profiling, we want a two-space heap with a - * fixed-size allocation area so that we get roughly even-spaced - * samples. - */ - - /* As an experiment, try a 2 generation collector - */ - -#if defined(PROFILING) || defined(DEBUG) - if (RtsFlags.ProfFlags.doHeapProfile) { - RtsFlags.GcFlags.generations = 1; - RtsFlags.GcFlags.steps = 1; - RtsFlags.GcFlags.oldGenFactor = 0; - RtsFlags.GcFlags.heapSizeSuggestion = 0; + if (generations != NULL) { + // multi-init protection + return; } -#endif + /* 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) { + prog_belch("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), @@ -140,6 +147,7 @@ initStorage( void ) 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; @@ -194,17 +202,12 @@ initStorage( void ) /* initialise the allocate() interface */ small_alloc_list = NULL; - large_alloc_list = NULL; alloc_blocks = 0; alloc_blocks_lim = RtsFlags.GcFlags.minAllocAreaSize; /* Tell GNU multi-precision pkg about our custom alloc functions */ mp_set_memory_functions(stgAllocForGMP, stgReallocForGMP, stgDeallocForGMP); -#ifdef SMP - pthread_mutex_init(&sm_mutex, NULL); -#endif - IF_DEBUG(gc, statDescribeGens()); } @@ -265,19 +268,13 @@ newCAF(StgClosure* caf) * come to do a major GC we won't need the mut_link field * any more and can use it as a STATIC_LINK. */ - ACQUIRE_LOCK(&sm_mutex); + ACQUIRE_SM_LOCK; - if (is_dynamically_loaded_rwdata_ptr((StgPtr)caf)) { - ((StgIndStatic *)caf)->saved_info = (StgInfoTable *)caf->header.info; - ((StgIndStatic *)caf)->static_link = caf_list; - caf_list = caf; - } else { - ((StgIndStatic *)caf)->saved_info = NULL; - ((StgMutClosure *)caf)->mut_link = oldest_gen->mut_once_list; - oldest_gen->mut_once_list = (StgMutClosure *)caf; - } + ((StgIndStatic *)caf)->saved_info = NULL; + ((StgMutClosure *)caf)->mut_link = oldest_gen->mut_once_list; + oldest_gen->mut_once_list = (StgMutClosure *)caf; - RELEASE_LOCK(&sm_mutex); + RELEASE_SM_LOCK; #ifdef PAR /* If we are PAR or DIST then we never forget a CAF */ @@ -289,6 +286,25 @@ newCAF(StgClosure* caf) #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. +// +// 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 = caf_list; + caf_list = caf; + + RELEASE_SM_LOCK; +} + /* ----------------------------------------------------------------------------- Nursery management. -------------------------------------------------------------------------- */ @@ -297,31 +313,28 @@ 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->rNursery = allocNursery(NULL, RtsFlags.GcFlags.minAllocAreaSize); - cap->rCurrentNursery = cap->rNursery; - for (bd = cap->rNursery; bd != NULL; bd = bd->link) { - bd->u.back = (bdescr *)cap; - } - } + 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 */ - nursery_blocks = RtsFlags.GcFlags.minAllocAreaSize; - g0s0->blocks = allocNursery(NULL, nursery_blocks); - g0s0->n_blocks = nursery_blocks; + g0s0->blocks = allocNursery(NULL, RtsFlags.GcFlags.minAllocAreaSize); + g0s0->n_blocks = RtsFlags.GcFlags.minAllocAreaSize; g0s0->to_blocks = NULL; g0s0->n_to_blocks = 0; - MainRegTable.rNursery = g0s0->blocks; - MainRegTable.rCurrentNursery = g0s0->blocks; + MainCapability.r.rNursery = g0s0->blocks; + MainCapability.r.rCurrentNursery = g0s0->blocks; /* hp, hpLim, hp_bd, to_space etc. aren't used in G0S0 */ #endif } @@ -337,13 +350,13 @@ resetNurseries( void ) ASSERT(n_free_capabilities == RtsFlags.ParFlags.nNodes); for (cap = free_capabilities; cap != NULL; cap = cap->link) { - for (bd = cap->rNursery; bd; bd = bd->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->rCurrentNursery = cap->rNursery; + cap->r.rCurrentNursery = cap->r.rNursery; } #else for (bd = g0s0->blocks; bd; bd = bd->link) { @@ -352,41 +365,55 @@ resetNurseries( void ) ASSERT(bd->step == g0s0); IF_DEBUG(sanity,memset(bd->start, 0xaa, BLOCK_SIZE)); } - MainRegTable.rNursery = g0s0->blocks; - MainRegTable.rCurrentNursery = g0s0->blocks; + MainCapability.r.rNursery = g0s0->blocks; + MainCapability.r.rCurrentNursery = g0s0->blocks; #endif } bdescr * -allocNursery (bdescr *last_bd, nat blocks) +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_no = 0; bd->flags = 0; bd->free = bd->start; - last_bd = bd; + tail = bd; } - return last_bd; + tail->u.back = NULL; + return tail; } 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; } @@ -401,15 +428,25 @@ resizeNursery ( nat blocks ) IF_DEBUG(gc, fprintf(stderr, "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; + g0s0->n_blocks = blocks; + ASSERT(countBlocks(g0s0->blocks) == g0s0->n_blocks); } /* ----------------------------------------------------------------------------- @@ -426,7 +463,7 @@ allocate( nat n ) bdescr *bd; StgPtr p; - ACQUIRE_LOCK(&sm_mutex); + ACQUIRE_SM_LOCK; TICK_ALLOC_HEAP_NOCTR(n); CCS_ALLOC(CCCS,n); @@ -437,17 +474,13 @@ 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); + g0s0->n_large_blocks += req_blocks; bd->gen_no = 0; bd->step = g0s0; bd->flags = BF_LARGE; - 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->free = bd->start + n; alloc_blocks += req_blocks; - RELEASE_LOCK(&sm_mutex); + RELEASE_SM_LOCK; return bd->start; /* small allocation (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; + } } /* --------------------------------------------------------------------------- @@ -507,18 +558,24 @@ allocatePinned( nat n ) StgPtr p; bdescr *bd = pinned_object_block; - ACQUIRE_LOCK(&sm_mutex); - - TICK_ALLOC_HEAP_NOCTR(n); - CCS_ALLOC(CCCS,n); - // If the request is for a large object, then allocate() // will give us a pinned object anyway. if (n >= LARGE_OBJECT_THRESHOLD/sizeof(W_)) { - RELEASE_LOCK(&sm_mutex); 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)) { @@ -526,14 +583,14 @@ allocatePinned( nat n ) dbl_link_onto(bd, &g0s0->large_objects); bd->gen_no = 0; bd->step = g0s0; - bd->flags = BF_LARGE; + bd->flags = BF_PINNED | BF_LARGE; bd->free = bd->start; alloc_blocks++; } p = bd->free; bd->free += n; - RELEASE_LOCK(&sm_mutex); + RELEASE_SM_LOCK; return p; } @@ -554,10 +611,8 @@ 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. */ @@ -565,7 +620,7 @@ stgAllocForGMP (size_t size_in_bytes) 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 * @@ -623,20 +678,20 @@ calcAllocated( void ) + allocated_bytes(); for (cap = free_capabilities; cap != NULL; cap = cap->link) { - for ( bd = cap->rCurrentNursery->link; bd != NULL; bd = bd->link ) { + for ( bd = cap->r.rCurrentNursery->link; bd != NULL; bd = bd->link ) { allocated -= BLOCK_SIZE_W; } - if (cap->rCurrentNursery->free < cap->rCurrentNursery->start + if (cap->r.rCurrentNursery->free < cap->r.rCurrentNursery->start + BLOCK_SIZE_W) { - allocated -= (cap->rCurrentNursery->start + BLOCK_SIZE_W) - - cap->rCurrentNursery->free; + allocated -= (cap->r.rCurrentNursery->start + BLOCK_SIZE_W) + - cap->r.rCurrentNursery->free; } } #else /* !SMP */ - bdescr *current_nursery = MainRegTable.rCurrentNursery; + bdescr *current_nursery = MainCapability.r.rCurrentNursery; - allocated = (nursery_blocks * BLOCK_SIZE_W) + allocated_bytes(); + allocated = (g0s0->n_blocks * BLOCK_SIZE_W) + allocated_bytes(); for ( bd = current_nursery->link; bd != NULL; bd = bd->link ) { allocated -= BLOCK_SIZE_W; } @@ -762,10 +817,16 @@ memInventory(void) 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(); @@ -779,7 +840,8 @@ memInventory(void) ASSERT(total_blocks + free_blocks == mblocks_allocated * BLOCKS_PER_MBLOCK); } -static nat + +nat countBlocks(bdescr *bd) { nat n; @@ -802,13 +864,13 @@ checkSanity( void ) for (g = 0; g < RtsFlags.GcFlags.generations; g++) { for (s = 0; s < generations[g].n_steps; s++) { - if (g == 0 && s == 0) { continue; } - checkHeap(generations[g].steps[s].blocks); - checkChain(generations[g].steps[s].large_objects); 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); checkMutOnceList(generations[g].mut_once_list, g); @@ -819,4 +881,13 @@ checkSanity( void ) } } +// handy function for use in gdb, because Bdescr() is inlined. +extern bdescr *_bdescr( StgPtr p ); + +bdescr * +_bdescr( StgPtr p ) +{ + return Bdescr(p); +} + #endif