X-Git-Url: http://git.megacz.com/?a=blobdiff_plain;ds=sidebyside;f=ghc%2Frts%2FStorage.c;h=f4e3bab15def353ea766a6b1126ffcaae29ed65f;hb=928e00e2a994559bf0eeb56dff22c7a3cf5c68be;hp=b2878c2ce14f866936cb4c218eee5fe4d1fb083b;hpb=ec0984a97cec59d0403ae1f23e23f100666e8148;p=ghc-hetmet.git diff --git a/ghc/rts/Storage.c b/ghc/rts/Storage.c index b2878c2..f4e3bab 100644 --- a/ghc/rts/Storage.c +++ b/ghc/rts/Storage.c @@ -22,11 +22,13 @@ #include "Storage.h" #include "Schedule.h" #include "RetainerProfile.h" // for counting memory blocks (memInventory) -#include "StoragePriv.h" #include #include +/* + * All these globals require sm_mutex to access in SMP mode. + */ StgClosure *caf_list = NULL; StgClosure *revertible_caf_list = NULL; rtsBool keepCAFs; @@ -69,13 +71,16 @@ initStep (step *stp, int g, int s) { stp->no = s; stp->blocks = NULL; - stp->n_to_blocks = 0; stp->n_blocks = 0; + stp->old_blocks = NULL; + stp->n_old_blocks = 0; stp->gen = &generations[g]; stp->gen_no = g; stp->hp = NULL; stp->hpLim = NULL; stp->hp_bd = NULL; + stp->scavd_hp = NULL; + stp->scavd_hpLim = NULL; stp->scan = NULL; stp->scan_bd = NULL; stp->large_objects = NULL; @@ -217,11 +222,6 @@ initStorage( void ) errorBelch("-G1 is incompatible with SMP"); stg_exit(1); } - // No -H, for now - if (RtsFlags.GcFlags.heapSizeSuggestion > 0) { - errorBelch("-H is incompatible with SMP"); - stg_exit(1); - } #endif /* generation 0 is special: that's the nursery */ @@ -411,10 +411,12 @@ assignNurseriesToCapabilities (void) for (i = 0; i < n_nurseries; i++) { capabilities[i].r.rNursery = &nurseries[i]; capabilities[i].r.rCurrentNursery = nurseries[i].blocks; + capabilities[i].r.rCurrentAlloc = NULL; } #else /* SMP */ MainCapability.r.rNursery = &nurseries[0]; MainCapability.r.rCurrentNursery = nurseries[0].blocks; + MainCapability.r.rCurrentAlloc = NULL; #endif } @@ -428,8 +430,8 @@ allocNurseries( void ) allocNursery(&nurseries[i], NULL, RtsFlags.GcFlags.minAllocAreaSize); nurseries[i].n_blocks = RtsFlags.GcFlags.minAllocAreaSize; - nurseries[i].to_blocks = NULL; - nurseries[i].n_to_blocks = 0; + nurseries[i].old_blocks = NULL; + nurseries[i].n_old_blocks = 0; /* hp, hpLim, hp_bd, to_space etc. aren't used in the nursery */ } assignNurseriesToCapabilities(); @@ -510,7 +512,7 @@ resizeNursery ( step *stp, nat blocks ) // Resize each of the nurseries to the specified size. // void -resizeNurseries (nat blocks) +resizeNurseriesFixed (nat blocks) { nat i; for (i = 0; i < n_nurseries; i++) { @@ -518,6 +520,17 @@ resizeNurseries (nat blocks) } } +// +// Resize the nurseries to the total specified size. +// +void +resizeNurseries (nat blocks) +{ + // If there are multiple nurseries, then we just divide the number + // of available blocks between them. + resizeNurseriesFixed(blocks / n_nurseries); +} + /* ----------------------------------------------------------------------------- The allocate() interface @@ -529,49 +542,49 @@ resizeNurseries (nat blocks) StgPtr allocate( nat n ) { - bdescr *bd; - StgPtr p; + bdescr *bd; + StgPtr p; - ACQUIRE_SM_LOCK; + ACQUIRE_SM_LOCK; - TICK_ALLOC_HEAP_NOCTR(n); - CCS_ALLOC(CCCS,n); - - /* big allocation (>LARGE_OBJECT_THRESHOLD) */ - /* ToDo: allocate directly into generation 1 */ - if (n >= LARGE_OBJECT_THRESHOLD/sizeof(W_)) { - 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 + n; - alloc_blocks += req_blocks; - RELEASE_SM_LOCK; - return bd->start; + TICK_ALLOC_HEAP_NOCTR(n); + CCS_ALLOC(CCCS,n); - /* small allocation ( alloc_HpLim) { - if (small_alloc_list) { - small_alloc_list->free = alloc_Hp; + /* big allocation (>LARGE_OBJECT_THRESHOLD) */ + /* ToDo: allocate directly into generation 1 */ + if (n >= LARGE_OBJECT_THRESHOLD/sizeof(W_)) { + 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 + n; + alloc_blocks += req_blocks; + RELEASE_SM_LOCK; + return bd->start; + + /* small allocation ( alloc_HpLim) { + if (small_alloc_list) { + small_alloc_list->free = alloc_Hp; + } + bd = allocBlock(); + bd->link = small_alloc_list; + small_alloc_list = bd; + bd->gen_no = 0; + bd->step = g0s0; + bd->flags = 0; + alloc_Hp = bd->start; + alloc_HpLim = bd->start + BLOCK_SIZE_W; + alloc_blocks++; } - bd = allocBlock(); - bd->link = small_alloc_list; - small_alloc_list = bd; - bd->gen_no = 0; - bd->step = g0s0; - 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; + + p = alloc_Hp; + alloc_Hp += n; + RELEASE_SM_LOCK; + return p; } lnat @@ -598,6 +611,84 @@ tidyAllocateLists (void) } } +/* ----------------------------------------------------------------------------- + allocateLocal() + + This allocates memory in the current thread - it is intended for + use primarily from STG-land where we have a Capability. It is + better than allocate() because it doesn't require taking the + sm_mutex lock in the common case. + + Memory is allocated directly from the nursery if possible (but not + from the current nursery block, so as not to interfere with + Hp/HpLim). + -------------------------------------------------------------------------- */ + +StgPtr +allocateLocal( StgRegTable *reg, nat n ) +{ + bdescr *bd; + StgPtr p; + + TICK_ALLOC_HEAP_NOCTR(n); + CCS_ALLOC(CCCS,n); + + /* big allocation (>LARGE_OBJECT_THRESHOLD) */ + /* ToDo: allocate directly into generation 1 */ + if (n >= LARGE_OBJECT_THRESHOLD/sizeof(W_)) { + nat req_blocks = (lnat)BLOCK_ROUND_UP(n*sizeof(W_)) / BLOCK_SIZE; + ACQUIRE_SM_LOCK; + 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 + n; + alloc_blocks += req_blocks; + RELEASE_SM_LOCK; + return bd->start; + + /* small allocation (rCurrentAlloc; + if (bd == NULL || bd->free + n > bd->start + BLOCK_SIZE_W) { + + // The CurrentAlloc block is full, we need to find another + // one. First, we try taking the next block from the + // nursery: + bd = reg->rCurrentNursery->link; + + if (bd == NULL || bd->free + n > bd->start + BLOCK_SIZE_W) { + // The nursery is empty, or the next block is already + // full: allocate a fresh block (we can't fail here). + ACQUIRE_SM_LOCK; + bd = allocBlock(); + reg->rNursery->n_blocks++; + RELEASE_SM_LOCK; + bd->gen_no = 0; + bd->step = g0s0; + bd->flags = 0; + } else { + // we have a block in the nursery: take it and put + // it at the *front* of the nursery list, and use it + // to allocate() from. + reg->rCurrentNursery->link = bd->link; + if (bd->link != NULL) { + bd->link->u.back = reg->rCurrentNursery; + } + } + dbl_link_onto(bd, ®->rNursery->blocks); + reg->rCurrentAlloc = bd; + IF_DEBUG(sanity, checkNurserySanity(reg->rNursery)); + } + } + p = bd->free; + bd->free += n; + return p; +} + /* --------------------------------------------------------------------------- Allocate a fixed/pinned object. @@ -685,7 +776,11 @@ stgAllocForGMP (size_t size_in_bytes) total_size_in_words = sizeofW(StgArrWords) + data_size_in_words; /* allocate and fill it in. */ - arr = (StgArrWords *)allocate(total_size_in_words); +#if defined(SMP) + arr = (StgArrWords *)allocateLocal(&(myCapability()->r), total_size_in_words); +#else + arr = (StgArrWords *)allocateLocal(&MainCapability.r, total_size_in_words); +#endif SET_ARR_HDR(arr, &stg_ARR_WORDS_info, CCCS, data_size_in_words); /* and return a ptr to the goods inside the array */ @@ -732,17 +827,16 @@ calcAllocated( void ) { nat allocated; bdescr *bd; - nat i; allocated = allocated_bytes(); - for (i = 0; i < n_nurseries; i++) { - allocated += nurseries[i].n_blocks * BLOCK_SIZE_W; - } + allocated += countNurseryBlocks() * BLOCK_SIZE_W; + { #ifdef SMP + nat i; for (i = 0; i < n_nurseries; i++) { Capability *cap; - for ( bd = capabilities[i].r.rCurrentNursery; + for ( bd = capabilities[i].r.rCurrentNursery->link; bd != NULL; bd = bd->link ) { allocated -= BLOCK_SIZE_W; } @@ -764,6 +858,7 @@ calcAllocated( void ) - current_nursery->free; } #endif + } total_allocated += allocated; return allocated; @@ -780,7 +875,7 @@ calcLive(void) step *stp; if (RtsFlags.GcFlags.generations == 1) { - live = (g0s0->n_to_blocks - 1) * BLOCK_SIZE_W + + live = (g0s0->n_blocks - 1) * BLOCK_SIZE_W + ((lnat)g0s0->hp_bd->free - (lnat)g0s0->hp_bd->start) / sizeof(W_); return live; } @@ -799,6 +894,9 @@ calcLive(void) live += ((lnat)stp->hp_bd->free - (lnat)stp->hp_bd->start) / sizeof(W_); } + if (stp->scavd_hp != NULL) { + live -= (P_)(BLOCK_ROUND_UP(stp->scavd_hp)) - stp->scavd_hp; + } } } return live; @@ -893,7 +991,7 @@ memInventory(void) if (RtsFlags.GcFlags.generations == 1) { /* two-space collector has a to-space too :-) */ - total_blocks += g0s0->n_to_blocks; + total_blocks += g0s0->n_old_blocks; } /* any blocks held by allocate() */ @@ -941,7 +1039,7 @@ checkSanity( void ) nat g, s; if (RtsFlags.GcFlags.generations == 1) { - checkHeap(g0s0->to_blocks); + checkHeap(g0s0->blocks); checkChain(g0s0->large_objects); } else { @@ -961,16 +1059,32 @@ checkSanity( void ) } for (s = 0; s < n_nurseries; 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); + ASSERT(countBlocks(nurseries[s].blocks) + == nurseries[s].n_blocks); + ASSERT(countBlocks(nurseries[s].large_objects) + == nurseries[s].n_large_blocks); } checkFreeListSanity(); } } +/* Nursery sanity check */ +void +checkNurserySanity( step *stp ) +{ + bdescr *bd, *prev; + nat blocks = 0; + + prev = NULL; + for (bd = stp->blocks; bd != NULL; bd = bd->link) { + ASSERT(bd->u.back == prev); + prev = bd; + blocks += bd->blocks; + } + ASSERT(blocks == stp->n_blocks); +} + // handy function for use in gdb, because Bdescr() is inlined. extern bdescr *_bdescr( StgPtr p );