X-Git-Url: http://git.megacz.com/?a=blobdiff_plain;f=ghc%2Frts%2FMBlock.c;h=fa8fd49d886daa62d1f8e71d7d8c31b7cc32959e;hb=1dfac5c8e457dccde541c2d38e702cb1567ed661;hp=9f0e89b5e8621a905364c138d95024cfe1910f82;hpb=5cc57129f941a8197e58f865cd93491c3a72a078;p=ghc-hetmet.git diff --git a/ghc/rts/MBlock.c b/ghc/rts/MBlock.c index 9f0e89b..fa8fd49 100644 --- a/ghc/rts/MBlock.c +++ b/ghc/rts/MBlock.c @@ -1,5 +1,4 @@ /* ----------------------------------------------------------------------------- - * $Id: MBlock.c,v 1.31 2002/10/21 12:42:31 simonmar Exp $ * * (c) The GHC Team 1998-1999 * @@ -21,13 +20,16 @@ #ifdef HAVE_STDLIB_H #include #endif +#ifdef HAVE_STRING_H +#include +#endif #ifdef HAVE_UNISTD_H #include #endif #ifdef HAVE_SYS_TYPES_H #include #endif -#ifndef mingw32_TARGET_OS +#ifndef mingw32_HOST_OS # ifdef HAVE_SYS_MMAN_H # include # endif @@ -38,6 +40,9 @@ #if HAVE_WINDOWS_H #include #endif +#if darwin_HOST_OS +#include +#endif #include @@ -47,12 +52,61 @@ lnat mblocks_allocated = 0; The MBlock Map: provides our implementation of HEAP_ALLOCED() -------------------------------------------------------------------------- */ -StgWord8 mblock_map[4096]; // initially all zeros +#if SIZEOF_VOID_P == 4 +StgWord8 mblock_map[MBLOCK_MAP_SIZE]; // initially all zeros +#elif SIZEOF_VOID_P == 8 +static MBlockMap dummy_mblock_map; +MBlockMap *mblock_cache = &dummy_mblock_map; +int mblock_map_count = 0; +MBlockMap **mblock_maps = NULL; + +static MBlockMap * +findMBlockMap(void *p) +{ + int i; + StgWord32 hi = (StgWord32) (((StgWord)p) >> 32); + for( i = 0; i < mblock_map_count; i++ ) + { + if(mblock_maps[i]->addrHigh32 == hi) + { + return mblock_maps[i]; + } + } + return NULL; +} + +StgBool +slowIsHeapAlloced(void *p) +{ + MBlockMap *map = findMBlockMap(p); + if(map) + { + mblock_cache = map; + return map->mblocks[MBLOCK_MAP_ENTRY(p)]; + } + else + return 0; +} +#endif static void -mblockIsHeap (void *p) +markHeapAlloced(void *p) { - mblock_map[((StgWord)p & ~MBLOCK_MASK) >> MBLOCK_SHIFT] = 1; +#if SIZEOF_VOID_P == 4 + mblock_map[MBLOCK_MAP_ENTRY(p)] = 1; +#elif SIZEOF_VOID_P == 8 + MBlockMap *map = findMBlockMap(p); + if(map == NULL) + { + mblock_map_count++; + mblock_maps = realloc(mblock_maps, + sizeof(MBlockMap*) * mblock_map_count); + map = mblock_maps[mblock_map_count-1] = calloc(1,sizeof(MBlockMap)); + map->addrHigh32 = (StgWord32) (((StgWord)p) >> 32); + } + map->mblocks[MBLOCK_MAP_ENTRY(p)] = 1; + mblock_cache = map; +#endif } /* ----------------------------------------------------------------------------- @@ -80,44 +134,86 @@ getMBlock(void) chunk, on the grounds that this is aligned and likely to be free. If it turns out that we were wrong, we have to munmap() and try again using the general method. + + Note on posix_memalign(): this interface is available on recent + systems and appears to provide exactly what we want. However, it + turns out not to be as good as our mmap() implementation, because + it wastes extra space (using double the address space, in a test on + x86_64/Linux). The problem seems to be that posix_memalign() + returns memory that can be free()'d, so the library must store + extra information along with the allocated block, thus messing up + the alignment. Hence, we don't use posix_memalign() for now. + -------------------------------------------------------------------------- */ -#if !defined(mingw32_TARGET_OS) && !defined(cygwin32_TARGET_OS) +#if !defined(mingw32_HOST_OS) && !defined(cygwin32_HOST_OS) // A wrapper around mmap(), to abstract away from OS differences in // the mmap() interface. static void * -my_mmap (void *addr, int size) +my_mmap (void *addr, lnat size) { void *ret; -#ifdef solaris2_TARGET_OS +#if defined(solaris2_HOST_OS) || defined(irix_HOST_OS) { int fd = open("/dev/zero",O_RDONLY); - ret = mmap(addr, size, PROT_READ | PROT_WRITE, - MAP_FIXED | MAP_PRIVATE, fd, 0); + ret = mmap(addr, size, PROT_READ | PROT_WRITE, MAP_PRIVATE, fd, 0); close(fd); } -#elif hpux_TARGET_OS +#elif hpux_HOST_OS ret = mmap(addr, size, PROT_READ | PROT_WRITE, MAP_ANONYMOUS | MAP_PRIVATE, -1, 0); -#elif darwin_TARGET_OS - ret = mmap(addr, size, PROT_READ | PROT_WRITE, - MAP_FIXED | MAP_ANON | MAP_PRIVATE, -1, 0); +#elif darwin_HOST_OS + // Without MAP_FIXED, Apple's mmap ignores addr. + // With MAP_FIXED, it overwrites already mapped regions, whic + // mmap(0, ... MAP_FIXED ...) is worst of all: It unmaps the program text + // and replaces it with zeroes, causing instant death. + // This behaviour seems to be conformant with IEEE Std 1003.1-2001. + // Let's just use the underlying Mach Microkernel calls directly, + // they're much nicer. + + kern_return_t err; + ret = addr; + if(addr) // try to allocate at adress + err = vm_allocate(mach_task_self(),(vm_address_t*) &ret, size, FALSE); + if(!addr || err) // try to allocate anywhere + err = vm_allocate(mach_task_self(),(vm_address_t*) &ret, size, TRUE); + + if(err) { + // don't know what the error codes mean exactly, assume it's + // not our problem though. + errorBelch("memory allocation failed (requested %lu bytes)", size); + stg_exit(EXIT_FAILURE); + } else { + vm_protect(mach_task_self(),ret,size,FALSE,VM_PROT_READ|VM_PROT_WRITE); + } #else - ret = mmap(addr, size, PROT_READ | PROT_WRITE, + ret = mmap(addr, size, PROT_READ | PROT_WRITE | PROT_EXEC, MAP_ANON | MAP_PRIVATE, -1, 0); #endif + if (ret == (void *)-1) { + if (errno == ENOMEM || + (errno == EINVAL && sizeof(void*)==4 && size >= 0xc0000000)) { + // If we request more than 3Gig, then we get EINVAL + // instead of ENOMEM (at least on Linux). + errorBelch("out of memory (requested %lu bytes)", size); + stg_exit(EXIT_FAILURE); + } else { + barf("getMBlock: mmap: %s", strerror(errno)); + } + } + return ret; -} +} // Implements the general case: allocate a chunk of memory of 'size' // mblocks. static void * -gen_map_mblocks (int size) +gen_map_mblocks (lnat size) { int slop; void *ret; @@ -126,20 +222,29 @@ gen_map_mblocks (int size) // it (unmap the rest). size += MBLOCK_SIZE; ret = my_mmap(0, size); - if (ret == (void *)-1) { - barf("gen_map_mblocks: mmap failed"); - } // unmap the slop bits around the chunk we allocated slop = (W_)ret & MBLOCK_MASK; - + if (munmap(ret, MBLOCK_SIZE - slop) == -1) { - barf("gen_map_mblocks: munmap failed"); + barf("gen_map_mblocks: munmap failed"); } if (slop > 0 && munmap(ret+size-slop, slop) == -1) { - barf("gen_map_mblocks: munmap failed"); + barf("gen_map_mblocks: munmap failed"); } - + + // ToDo: if we happened to get an aligned block, then don't + // unmap the excess, just use it. For this to work, you + // need to keep in mind the following: + // * Calling my_mmap() with an 'addr' arg pointing to + // already my_mmap()ed space is OK and won't fail. + // * If my_mmap() can't satisfy the request at the + // given 'next_request' address in getMBlocks(), that + // you unmap the extra mblock mmap()ed here (or simply + // satisfy yourself that the slop introduced isn't worth + // salvaging.) + // + // next time, try after the block we just got. ret += MBLOCK_SIZE - slop; return ret; @@ -162,22 +267,13 @@ getMBlocks(nat n) ret = gen_map_mblocks(size); } else { ret = my_mmap(next_request, size); - - if (ret == (void *)-1) { - if (errno == ENOMEM) { - belch("out of memory (requested %d bytes)", n * BLOCK_SIZE); - stg_exit(EXIT_FAILURE); - } else { - barf("getMBlock: mmap failed"); - } - } if (((W_)ret & MBLOCK_MASK) != 0) { // misaligned block! -#ifdef DEBUG - belch("getMBlock: misaligned block %p returned when allocating %d megablock(s) at %p", ret, n, next_request); +#if 0 // defined(DEBUG) + errorBelch("warning: getMBlock: misaligned block %p returned when allocating %d megablock(s) at %p", ret, n, next_request); #endif - + // unmap this block... if (munmap(ret, size) == -1) { barf("getMBlock: munmap failed"); @@ -188,13 +284,14 @@ getMBlocks(nat n) } // Next time, we'll try to allocate right after the block we just got. + // ToDo: check that we haven't already grabbed the memory at next_request next_request = ret + size; - IF_DEBUG(gc,fprintf(stderr,"Allocated %d megablock(s) at %p\n",n,ret)); + IF_DEBUG(gc,debugBelch("Allocated %d megablock(s) at %p\n",n,ret)); // fill in the table for (i = 0; i < n; i++) { - mblockIsHeap( ret + i * MBLOCK_SIZE ); + markHeapAlloced( ret + i * MBLOCK_SIZE ); } mblocks_allocated += n; @@ -202,7 +299,13 @@ getMBlocks(nat n) return ret; } -#else /* defined(mingw32_TARGET_OS) || defined(cygwin32_TARGET_OS) */ +void +freeAllMBlocks(void) +{ + /* XXX Do something here */ +} + +#else /* defined(mingw32_HOST_OS) || defined(cygwin32_HOST_OS) */ /* On Win32 platforms we make use of the two-phased virtual memory API @@ -219,8 +322,10 @@ getMBlocks(nat n) our case). */ -char* base_non_committed = (char*)0; -char* end_non_committed = (char*)0; +static char* base_non_committed = (char*)0; +static char* end_non_committed = (char*)0; + +static void *membase; /* Default is to reserve 256M of VM to minimise the slop cost. */ #define SIZE_RESERVED_POOL ( 256 * 1024 * 1024 ) @@ -234,12 +339,19 @@ getMBlocks(nat n) static char* base_mblocks = (char*)0; static char* next_request = (char*)0; void* ret = (void*)0; + nat i; lnat size = MBLOCK_SIZE * n; if ( (base_non_committed == 0) || (next_request + size > end_non_committed) ) { if (base_non_committed) { - barf("RTS exhausted max heap size (%d bytes)\n", size_reserved_pool); + /* Tacky, but if no user-provided -M option is in effect, + * set it to the default (==256M) in time for the heap overflow PSA. + */ + if (RtsFlags.GcFlags.maxHeapSize == 0) { + RtsFlags.GcFlags.maxHeapSize = size_reserved_pool / BLOCK_SIZE; + } + heapOverflow(); } if (RtsFlags.GcFlags.maxHeapSize != 0) { size_reserved_pool = BLOCK_SIZE * RtsFlags.GcFlags.maxHeapSize; @@ -252,20 +364,21 @@ getMBlocks(nat n) , MEM_RESERVE , PAGE_READWRITE ); + membase = base_non_committed; if ( base_non_committed == 0 ) { - fprintf(stderr, "getMBlocks: VirtualAlloc failed with: %ld\n", GetLastError()); - ret=(void*)-1; + errorBelch("getMBlocks: VirtualAlloc MEM_RESERVE %lu failed with: %ld\n", size_reserved_pool, GetLastError()); + ret=(void*)-1; } else { end_non_committed = (char*)base_non_committed + (unsigned long)size_reserved_pool; /* The returned pointer is not aligned on a mega-block boundary. Make it. */ base_mblocks = (char*)((unsigned long)base_non_committed & (unsigned long)~MBLOCK_MASK) + MBLOCK_SIZE; # if 0 - fprintf(stderr, "getMBlocks: Dropping %d bytes off of 256M chunk\n", - (unsigned)base_mblocks - (unsigned)base_non_committed); + debugBelch("getMBlocks: Dropping %d bytes off of 256M chunk\n", + (unsigned)base_mblocks - (unsigned)base_non_committed); # endif if ( ((char*)base_mblocks + size) > end_non_committed ) { - fprintf(stderr, "getMBlocks: oops, committed too small a region to start with."); + debugBelch("getMBlocks: oops, committed too small a region to start with."); ret=(void*)-1; } else { next_request = base_mblocks; @@ -276,7 +389,7 @@ getMBlocks(nat n) if ( ret != (void*)-1 ) { ret = VirtualAlloc(next_request, size, MEM_COMMIT, PAGE_READWRITE); if (ret == NULL) { - fprintf(stderr, "getMBlocks: VirtualAlloc failed with: %ld\n", GetLastError()); + debugBelch("getMBlocks: VirtualAlloc MEM_COMMIT %lu failed with: %ld\n", size, GetLastError()); ret=(void*)-1; } } @@ -289,19 +402,31 @@ getMBlocks(nat n) barf("getMBlocks: unknown memory allocation failure on Win32."); } - IF_DEBUG(gc,fprintf(stderr,"Allocated %d megablock(s) at 0x%x\n",n,(nat)ret)); + IF_DEBUG(gc,debugBelch("Allocated %d megablock(s) at 0x%x\n",n,(nat)ret)); next_request = (char*)next_request + size; mblocks_allocated += n; // fill in the table for (i = 0; i < n; i++) { - mblockIsHeap( ret + i * MBLOCK_SIZE ); + markHeapAlloced( ret + i * MBLOCK_SIZE ); } return ret; } +void +freeAllMBlocks(void) +{ + BOOL rc; + + rc = VirtualFree(membase, 0, MEM_RELEASE); + + if (rc == FALSE) { + debugBelch("freeAllMBlocks: VirtualFree failed with: %ld\n", GetLastError()); + } +} + /* Hand back the physical memory that is allocated to a mega-block. ToDo: chain the released mega block onto some list so that getMBlocks() can get at it. @@ -318,7 +443,7 @@ freeMBlock(void* p, nat n) if (rc == FALSE) { # ifdef DEBUG - fprintf(stderr, "freeMBlocks: VirtualFree failed with: %d\n", GetLastError()); + debugBelch("freeMBlocks: VirtualFree failed with: %d\n", GetLastError()); # endif }