X-Git-Url: http://git.megacz.com/?a=blobdiff_plain;ds=sidebyside;f=rts%2Fsm%2FStorage.c;h=97615e9d1b5b0df06c9c4179b1a4081238b21413;hb=a2a67cd520b9841114d69a87a423dabcb3b4368e;hp=d37a076ab624699e41864186f3b0406f9d7c4339;hpb=b339c8b1d0f239031802555b454062e9430ec8bb;p=ghc-hetmet.git diff --git a/rts/sm/Storage.c b/rts/sm/Storage.c index d37a076..97615e9 100644 --- a/rts/sm/Storage.c +++ b/rts/sm/Storage.c @@ -13,18 +13,15 @@ #include "PosixSource.h" #include "Rts.h" + +#include "Storage.h" #include "RtsUtils.h" -#include "RtsFlags.h" #include "Stats.h" -#include "Hooks.h" #include "BlockAlloc.h" -#include "MBlock.h" #include "Weak.h" #include "Sanity.h" #include "Arena.h" -#include "OSThreads.h" #include "Capability.h" -#include "Storage.h" #include "Schedule.h" #include "RetainerProfile.h" // for counting memory blocks (memInventory) #include "OSMem.h" @@ -32,9 +29,10 @@ #include "GC.h" #include "Evac.h" -#include #include +#include "ffi.h" + /* * All these globals require sm_mutex to access in THREADED_RTS mode. */ @@ -46,6 +44,8 @@ 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 */ +static bdescr *exec_block; + generation *generations = NULL; /* all the generations */ generation *g0 = NULL; /* generation 0, for convenience */ generation *oldest_gen = NULL; /* oldest generation, for convenience */ @@ -65,19 +65,9 @@ step *nurseries = NULL; /* array of nurseries, >1 only if THREADED_RTS * * simultaneous access by two STG threads. */ Mutex sm_mutex; -/* - * This mutex is used by atomicModifyMutVar# only - */ -Mutex atomic_modify_mutvar_mutex; #endif - -/* - * Forward references - */ -static void *stgAllocForGMP (size_t size_in_bytes); -static void *stgReallocForGMP (void *ptr, size_t old_size, size_t new_size); -static void stgDeallocForGMP (void *ptr, size_t size); +static void allocNurseries ( void ); static void initStep (step *stp, int g, int s) @@ -87,6 +77,7 @@ initStep (step *stp, int g, int s) stp->blocks = NULL; stp->n_blocks = 0; stp->n_words = 0; + stp->live_estimate = 0; stp->old_blocks = NULL; stp->n_old_blocks = 0; stp->gen = &generations[g]; @@ -95,10 +86,10 @@ initStep (step *stp, int g, int s) stp->n_large_blocks = 0; stp->scavenged_large_objects = NULL; stp->n_scavenged_large_blocks = 0; - stp->is_compacted = 0; + stp->mark = 0; + stp->compact = 0; stp->bitmap = NULL; #ifdef THREADED_RTS - initSpinLock(&stp->sync_todo); initSpinLock(&stp->sync_large_objects); #endif stp->threads = END_TSO_QUEUE; @@ -122,7 +113,7 @@ initStorage( void ) * doing something reasonable. */ /* We use the NOT_NULL variant or gcc warns that the test is always true */ - ASSERT(LOOKS_LIKE_INFO_PTR_NOT_NULL(&stg_BLACKHOLE_info)); + ASSERT(LOOKS_LIKE_INFO_PTR_NOT_NULL((StgWord)&stg_BLACKHOLE_info)); ASSERT(LOOKS_LIKE_CLOSURE_PTR(&stg_dummy_ret_closure)); ASSERT(!HEAP_ALLOCED(&stg_dummy_ret_closure)); @@ -143,7 +134,6 @@ initStorage( void ) #if defined(THREADED_RTS) initMutex(&sm_mutex); - initMutex(&atomic_modify_mutvar_mutex); #endif ACQUIRE_SM_LOCK; @@ -230,11 +220,13 @@ initStorage( void ) } /* The oldest generation has one step. */ - if (RtsFlags.GcFlags.compact) { + if (RtsFlags.GcFlags.compact || RtsFlags.GcFlags.sweep) { if (RtsFlags.GcFlags.generations == 1) { - errorBelch("WARNING: compaction is incompatible with -G1; disabled"); + errorBelch("WARNING: compact/sweep is incompatible with -G1; disabled"); } else { - oldest_gen->steps[0].is_compacted = 1; + oldest_gen->steps[0].mark = 1; + if (RtsFlags.GcFlags.compact) + oldest_gen->steps[0].compact = 1; } } @@ -257,15 +249,17 @@ initStorage( void ) 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); + exec_block = NULL; #ifdef THREADED_RTS initSpinLock(&gc_alloc_block_sync); - initSpinLock(&recordMutableGen_sync); whitehole_spin = 0; #endif + N = 0; + + initGcThreads(); + IF_DEBUG(gc, statDescribeGens()); RELEASE_SM_LOCK; @@ -285,9 +279,9 @@ freeStorage (void) freeAllMBlocks(); #if defined(THREADED_RTS) closeMutex(&sm_mutex); - closeMutex(&atomic_modify_mutvar_mutex); #endif stgFree(nurseries); + freeGcThreads(); } /* ----------------------------------------------------------------------------- @@ -362,7 +356,7 @@ newCAF(StgClosure* caf) * any more and can use it as a STATIC_LINK. */ ((StgIndStatic *)caf)->saved_info = NULL; - recordMutableGen(caf, oldest_gen); + recordMutableGen(caf, oldest_gen->no); } RELEASE_SM_LOCK; @@ -443,7 +437,7 @@ assignNurseriesToCapabilities (void) #endif } -void +static void allocNurseries( void ) { nat i; @@ -581,7 +575,7 @@ move_TSO (StgTSO *src, StgTSO *dest) -------------------------------------------------------------------------- */ StgPtr -allocateInGen (generation *g, nat n) +allocateInGen (generation *g, lnat n) { step *stp; bdescr *bd; @@ -596,18 +590,26 @@ allocateInGen (generation *g, nat n) if (n >= LARGE_OBJECT_THRESHOLD/sizeof(W_)) { - nat req_blocks = (lnat)BLOCK_ROUND_UP(n*sizeof(W_)) / BLOCK_SIZE; + lnat req_blocks = (lnat)BLOCK_ROUND_UP(n*sizeof(W_)) / BLOCK_SIZE; // Attempting to allocate an object larger than maxHeapSize // should definitely be disallowed. (bug #1791) if (RtsFlags.GcFlags.maxHeapSize > 0 && req_blocks >= RtsFlags.GcFlags.maxHeapSize) { heapOverflow(); + // heapOverflow() doesn't exit (see #2592), but we aren't + // in a position to do a clean shutdown here: we + // either have to allocate the memory or exit now. + // Allocating the memory would be bad, because the user + // has requested that we not exceed maxHeapSize, so we + // just exit. + stg_exit(EXIT_HEAPOVERFLOW); } bd = allocGroup(req_blocks); dbl_link_onto(bd, &stp->large_objects); stp->n_large_blocks += bd->blocks; // might be larger than req_blocks + alloc_blocks += bd->blocks; bd->gen_no = g->no; bd->step = stp; bd->flags = BF_LARGE; @@ -638,7 +640,7 @@ allocateInGen (generation *g, nat n) } StgPtr -allocate (nat n) +allocate (lnat n) { return allocateInGen(g0,n); } @@ -657,8 +659,8 @@ allocatedBytes( void ) return allocated; } -// split N blocks off the start of the given bdescr, returning the -// remainder as a new block group. We treat the remainder as if it +// split N blocks off the front of the given bdescr, returning the +// new block group. We treat the remainder as if it // had been freshly allocated in generation 0. bdescr * splitLargeBlock (bdescr *bd, nat blocks) @@ -676,6 +678,7 @@ splitLargeBlock (bdescr *bd, nat blocks) new_bd->step = g0s0; new_bd->flags = BF_LARGE; new_bd->free = bd->free; + ASSERT(new_bd->free <= new_bd->start + new_bd->blocks * BLOCK_SIZE_W); // add the new number of blocks to the counter. Due to the gaps // for block descriptor, new_bd->blocks + bd->blocks might not be @@ -683,7 +686,7 @@ splitLargeBlock (bdescr *bd, nat blocks) bd->step->n_large_blocks += bd->blocks; return new_bd; -} +} /* ----------------------------------------------------------------------------- allocateLocal() @@ -699,7 +702,7 @@ splitLargeBlock (bdescr *bd, nat blocks) -------------------------------------------------------------------------- */ StgPtr -allocateLocal (Capability *cap, nat n) +allocateLocal (Capability *cap, lnat n) { bdescr *bd; StgPtr p; @@ -733,7 +736,9 @@ allocateLocal (Capability *cap, nat n) bd->flags = 0; // NO: alloc_blocks++; // calcAllocated() uses the size of the nursery, and we've - // already bumpted nursery->n_blocks above. + // already bumpted nursery->n_blocks above. We'll GC + // pretty quickly now anyway, because MAYBE_GC() will + // notice that CurrentNursery->link is NULL. } else { // we have a block in the nursery: take it and put // it at the *front* of the nursery list, and use it @@ -776,7 +781,7 @@ allocateLocal (Capability *cap, nat n) ------------------------------------------------------------------------- */ StgPtr -allocatePinned( nat n ) +allocatePinned( lnat n ) { StgPtr p; bdescr *bd = pinned_object_block; @@ -784,7 +789,9 @@ allocatePinned( nat 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_)) { - return allocate(n); + p = allocate(n); + Bdescr(p)->flags |= BF_PINNED; + return p; } ACQUIRE_SM_LOCK; @@ -792,13 +799,6 @@ allocatePinned( nat n ) 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)) { @@ -887,61 +887,6 @@ dirty_MVAR(StgRegTable *reg, StgClosure *p) } /* ----------------------------------------------------------------------------- - Allocation functions for GMP. - - These all use the allocate() interface - we can't have any garbage - collection going on during a gmp operation, so we use allocate() - which always succeeds. The gmp operations which might need to - allocate will ask the storage manager (via doYouWantToGC()) whether - a garbage collection is required, in case we get into a loop doing - only allocate() style allocation. - -------------------------------------------------------------------------- */ - -static void * -stgAllocForGMP (size_t size_in_bytes) -{ - StgArrWords* arr; - nat data_size_in_words, total_size_in_words; - - /* 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. */ -#if defined(THREADED_RTS) - arr = (StgArrWords *)allocateLocal(myTask()->cap, total_size_in_words); -#else - arr = (StgArrWords *)allocateLocal(&MainCapability, 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 */ - return arr->payload; -} - -static void * -stgReallocForGMP (void *ptr, size_t old_size, size_t new_size) -{ - void *new_stuff_ptr = stgAllocForGMP(new_size); - nat i = 0; - char *p = (char *) ptr; - char *q = (char *) new_stuff_ptr; - - for (; i < old_size; i++, p++, q++) { - *q = *p; - } - - return(new_stuff_ptr); -} - -static void -stgDeallocForGMP (void *ptr STG_UNUSED, - size_t size STG_UNUSED) -{ - /* easy for us: the garbage collector does the dealloc'n */ -} - -/* ----------------------------------------------------------------------------- * Stats and stuff * -------------------------------------------------------------------------- */ @@ -1032,6 +977,7 @@ countOccupied(bdescr *bd) words = 0; for (; bd != NULL; bd = bd->link) { + ASSERT(bd->free <= bd->start + bd->blocks * BLOCK_SIZE_W); words += bd->free - bd->start; } return words; @@ -1079,14 +1025,27 @@ calcNeeded(void) for (s = 0; s < generations[g].n_steps; s++) { if (g == 0 && s == 0) { continue; } stp = &generations[g].steps[s]; + + // we need at least this much space + needed += stp->n_blocks + stp->n_large_blocks; + + // any additional space needed to collect this gen next time? if (g == 0 || // always collect gen 0 (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 + stp->n_large_blocks; - } else { - needed += stp->n_blocks + stp->n_large_blocks; + > generations[g].max_blocks)) { + // we will collect this gen next time + if (stp->mark) { + // bitmap: + needed += stp->n_blocks / BITS_IN(W_); + // mark stack: + needed += stp->n_blocks / 100; + } + if (stp->compact) { + continue; // no additional space needed for compaction + } else { + needed += stp->n_blocks; + } } } } @@ -1113,9 +1072,37 @@ calcNeeded(void) should be modified to use allocateExec instead of VirtualAlloc. ------------------------------------------------------------------------- */ -static bdescr *exec_block; +#if defined(linux_HOST_OS) + +// On Linux we need to use libffi for allocating executable memory, +// because it knows how to work around the restrictions put in place +// by SELinux. -void *allocateExec (nat bytes) +void *allocateExec (nat bytes, void **exec_ret) +{ + void **ret, **exec; + ACQUIRE_SM_LOCK; + ret = ffi_closure_alloc (sizeof(void *) + (size_t)bytes, (void**)&exec); + RELEASE_SM_LOCK; + if (ret == NULL) return ret; + *ret = ret; // save the address of the writable mapping, for freeExec(). + *exec_ret = exec + 1; + return (ret + 1); +} + +// freeExec gets passed the executable address, not the writable address. +void freeExec (void *addr) +{ + void *writable; + writable = *((void**)addr - 1); + ACQUIRE_SM_LOCK; + ffi_closure_free (writable); + RELEASE_SM_LOCK +} + +#else + +void *allocateExec (nat bytes, void **exec_ret) { void *ret; nat n; @@ -1151,6 +1138,7 @@ void *allocateExec (nat bytes) exec_block->free += n + 1; RELEASE_SM_LOCK + *exec_ret = ret; return ret; } @@ -1188,6 +1176,8 @@ void freeExec (void *addr) RELEASE_SM_LOCK } +#endif /* mingw32_HOST_OS */ + /* ----------------------------------------------------------------------------- Debugging @@ -1252,6 +1242,51 @@ stepBlocks (step *stp) countAllocdBlocks(stp->large_objects); } +// If memInventory() calculates that we have a memory leak, this +// function will try to find the block(s) that are leaking by marking +// all the ones that we know about, and search through memory to find +// blocks that are not marked. In the debugger this can help to give +// us a clue about what kind of block leaked. In the future we might +// annotate blocks with their allocation site to give more helpful +// info. +static void +findMemoryLeak (void) +{ + nat g, s, i; + for (g = 0; g < RtsFlags.GcFlags.generations; g++) { + for (i = 0; i < n_capabilities; i++) { + markBlocks(capabilities[i].mut_lists[g]); + } + markBlocks(generations[g].mut_list); + for (s = 0; s < generations[g].n_steps; s++) { + markBlocks(generations[g].steps[s].blocks); + markBlocks(generations[g].steps[s].large_objects); + } + } + + for (i = 0; i < n_nurseries; i++) { + markBlocks(nurseries[i].blocks); + markBlocks(nurseries[i].large_objects); + } + +#ifdef PROFILING + // TODO: + // if (RtsFlags.ProfFlags.doHeapProfile == HEAP_BY_RETAINER) { + // markRetainerBlocks(); + // } +#endif + + // count the blocks allocated by the arena allocator + // TODO: + // markArenaBlocks(); + + // count the blocks containing executable memory + markBlocks(exec_block); + + reportUnmarkedBlocks(); +} + + void memInventory (rtsBool show) { @@ -1308,6 +1343,7 @@ memInventory (rtsBool show) #define MB(n) (((n) * BLOCK_SIZE_W) / ((1024*1024)/sizeof(W_))) leak = live_blocks + free_blocks != mblocks_allocated * BLOCKS_PER_MBLOCK; + if (show || leak) { if (leak) { @@ -1336,6 +1372,13 @@ memInventory (rtsBool show) mblocks_allocated * BLOCKS_PER_MBLOCK, mblocks_allocated); } } + + if (leak) { + debugBelch("\n"); + findMemoryLeak(); + } + ASSERT(n_alloc_blocks == live_blocks); + ASSERT(!leak); } @@ -1347,7 +1390,7 @@ checkSanity( void ) if (RtsFlags.GcFlags.generations == 1) { checkHeap(g0s0->blocks); - checkChain(g0s0->large_objects); + checkLargeObjects(g0s0->large_objects); } else { for (g = 0; g < RtsFlags.GcFlags.generations; g++) { @@ -1358,10 +1401,7 @@ checkSanity( void ) ASSERT(countBlocks(generations[g].steps[s].large_objects) == generations[g].steps[s].n_large_blocks); checkHeap(generations[g].steps[s].blocks); - checkChain(generations[g].steps[s].large_objects); - if (g > 0) { - checkMutableList(generations[g].mut_list, g); - } + checkLargeObjects(generations[g].steps[s].large_objects); } } @@ -1374,6 +1414,14 @@ checkSanity( void ) checkFreeListSanity(); } + +#if defined(THREADED_RTS) + // check the stacks too in threaded mode, because we don't do a + // full heap sanity check in this case (see checkHeap()) + checkMutableLists(rtsTrue); +#else + checkMutableLists(rtsFalse); +#endif } /* Nursery sanity check */