/* -*- Mode: C++; tab-width: 4; indent-tabs-mode: nil; c-basic-offset: 2 -*- */ /* This Source Code Form is subject to the terms of the Mozilla Public * License, v. 2.0. If a copy of the MPL was not distributed with this * file, You can obtain one at http://mozilla.org/MPL/2.0/. */ /* ** Thread safe versions of malloc, free, realloc, calloc and cfree. */ #include "primpl.h" #ifdef _PR_ZONE_ALLOCATOR /* ** The zone allocator code must use native mutexes and cannot ** use PRLocks because PR_NewLock calls PR_Calloc, resulting ** in cyclic dependency of initialization. */ #include union memBlkHdrUn; typedef struct MemoryZoneStr { union memBlkHdrUn *head; /* free list */ pthread_mutex_t lock; size_t blockSize; /* size of blocks on this free list */ PRUint32 locked; /* current state of lock */ PRUint32 contention; /* counter: had to wait for lock */ PRUint32 hits; /* allocated from free list */ PRUint32 misses; /* had to call malloc */ PRUint32 elements; /* on free list */ } MemoryZone; typedef union memBlkHdrUn { unsigned char filler[48]; /* fix the size of this beast */ struct memBlkHdrStr { union memBlkHdrUn *next; MemoryZone *zone; size_t blockSize; size_t requestedSize; PRUint32 magic; } s; } MemBlockHdr; #define MEM_ZONES 7 #define THREAD_POOLS 11 /* prime number for modulus */ #define ZONE_MAGIC 0x0BADC0DE static MemoryZone zones[MEM_ZONES][THREAD_POOLS]; static PRBool use_zone_allocator = PR_FALSE; static void pr_ZoneFree(void *ptr); void _PR_DestroyZones(void) { int i, j; if (!use_zone_allocator) { return; } for (j = 0; j < THREAD_POOLS; j++) { for (i = 0; i < MEM_ZONES; i++) { MemoryZone *mz = &zones[i][j]; pthread_mutex_destroy(&mz->lock); while (mz->head) { MemBlockHdr *hdr = mz->head; mz->head = hdr->s.next; /* unlink it */ free(hdr); mz->elements--; } } } use_zone_allocator = PR_FALSE; } /* ** pr_FindSymbolInProg ** ** Find the specified data symbol in the program and return ** its address. */ #ifdef HAVE_DLL #if defined(USE_DLFCN) && !defined(NO_DLOPEN_NULL) #include static void * pr_FindSymbolInProg(const char *name) { void *h; void *sym; h = dlopen(0, RTLD_LAZY); if (h == NULL) { return NULL; } sym = dlsym(h, name); (void)dlclose(h); return sym; } #elif defined(USE_HPSHL) #include static void * pr_FindSymbolInProg(const char *name) { shl_t h = NULL; void *sym; if (shl_findsym(&h, name, TYPE_DATA, &sym) == -1) { return NULL; } return sym; } #elif defined(USE_MACH_DYLD) || defined(NO_DLOPEN_NULL) static void * pr_FindSymbolInProg(const char *name) { /* FIXME: not implemented */ return NULL; } #else #error "The zone allocator is not supported on this platform" #endif #else /* !defined(HAVE_DLL) */ static void * pr_FindSymbolInProg(const char *name) { /* can't be implemented */ return NULL; } #endif /* HAVE_DLL */ void _PR_InitZones(void) { int i, j; char *envp; PRBool *sym; if ((sym = (PRBool *)pr_FindSymbolInProg("nspr_use_zone_allocator")) != NULL) { use_zone_allocator = *sym; } else if ((envp = getenv("NSPR_USE_ZONE_ALLOCATOR")) != NULL) { use_zone_allocator = (atoi(envp) == 1); } if (!use_zone_allocator) { return; } for (j = 0; j < THREAD_POOLS; j++) { for (i = 0; i < MEM_ZONES; i++) { MemoryZone *mz = &zones[i][j]; int rv = pthread_mutex_init(&mz->lock, NULL); PR_ASSERT(0 == rv); if (rv != 0) { goto loser; } mz->blockSize = 16 << ( 2 * i); } } return; loser: _PR_DestroyZones(); return; } PR_IMPLEMENT(void) PR_FPrintZoneStats(PRFileDesc *debug_out) { int i, j; for (j = 0; j < THREAD_POOLS; j++) { for (i = 0; i < MEM_ZONES; i++) { MemoryZone *mz = &zones[i][j]; MemoryZone zone = *mz; if (zone.elements || zone.misses || zone.hits) { PR_fprintf(debug_out, "pool: %d, zone: %d, size: %d, free: %d, hit: %d, miss: %d, contend: %d\n", j, i, zone.blockSize, zone.elements, zone.hits, zone.misses, zone.contention); } } } } static void * pr_ZoneMalloc(PRUint32 size) { void *rv; unsigned int zone; size_t blockSize; MemBlockHdr *mb, *mt; MemoryZone *mz; /* Always allocate a non-zero amount of bytes */ if (size < 1) { size = 1; } for (zone = 0, blockSize = 16; zone < MEM_ZONES; ++zone, blockSize <<= 2) { if (size <= blockSize) { break; } } if (zone < MEM_ZONES) { pthread_t me = pthread_self(); unsigned int pool = (PRUptrdiff)me % THREAD_POOLS; PRUint32 wasLocked; mz = &zones[zone][pool]; wasLocked = mz->locked; pthread_mutex_lock(&mz->lock); mz->locked = 1; if (wasLocked) { mz->contention++; } if (mz->head) { mb = mz->head; PR_ASSERT(mb->s.magic == ZONE_MAGIC); PR_ASSERT(mb->s.zone == mz); PR_ASSERT(mb->s.blockSize == blockSize); PR_ASSERT(mz->blockSize == blockSize); mt = (MemBlockHdr *)(((char *)(mb + 1)) + blockSize); PR_ASSERT(mt->s.magic == ZONE_MAGIC); PR_ASSERT(mt->s.zone == mz); PR_ASSERT(mt->s.blockSize == blockSize); mz->hits++; mz->elements--; mz->head = mb->s.next; /* take off free list */ mz->locked = 0; pthread_mutex_unlock(&mz->lock); mt->s.next = mb->s.next = NULL; mt->s.requestedSize = mb->s.requestedSize = size; rv = (void *)(mb + 1); return rv; } mz->misses++; mz->locked = 0; pthread_mutex_unlock(&mz->lock); mb = (MemBlockHdr *)malloc(blockSize + 2 * (sizeof *mb)); if (!mb) { PR_SetError(PR_OUT_OF_MEMORY_ERROR, 0); return NULL; } mb->s.next = NULL; mb->s.zone = mz; mb->s.magic = ZONE_MAGIC; mb->s.blockSize = blockSize; mb->s.requestedSize = size; mt = (MemBlockHdr *)(((char *)(mb + 1)) + blockSize); memcpy(mt, mb, sizeof *mb); rv = (void *)(mb + 1); return rv; } /* size was too big. Create a block with no zone */ blockSize = (size & 15) ? size + 16 - (size & 15) : size; mb = (MemBlockHdr *)malloc(blockSize + 2 * (sizeof *mb)); if (!mb) { PR_SetError(PR_OUT_OF_MEMORY_ERROR, 0); return NULL; } mb->s.next = NULL; mb->s.zone = NULL; mb->s.magic = ZONE_MAGIC; mb->s.blockSize = blockSize; mb->s.requestedSize = size; mt = (MemBlockHdr *)(((char *)(mb + 1)) + blockSize); memcpy(mt, mb, sizeof *mb); rv = (void *)(mb + 1); return rv; } static void * pr_ZoneCalloc(PRUint32 nelem, PRUint32 elsize) { PRUint32 size = nelem * elsize; void *p = pr_ZoneMalloc(size); if (p) { memset(p, 0, size); } return p; } static void * pr_ZoneRealloc(void *oldptr, PRUint32 bytes) { void *rv; MemBlockHdr *mb; int ours; MemBlockHdr phony; if (!oldptr) { return pr_ZoneMalloc(bytes); } mb = (MemBlockHdr *)((char *)oldptr - (sizeof *mb)); if (mb->s.magic != ZONE_MAGIC) { /* Maybe this just came from ordinary malloc */ #ifdef DEBUG fprintf(stderr, "Warning: reallocing memory block %p from ordinary malloc\n", oldptr); #endif /* * We are going to realloc oldptr. If realloc succeeds, the * original value of oldptr will point to freed memory. So this * function must not fail after a successfull realloc call. We * must perform any operation that may fail before the realloc * call. */ rv = pr_ZoneMalloc(bytes); /* this may fail */ if (!rv) { return rv; } /* We don't know how big it is. But we can fix that. */ oldptr = realloc(oldptr, bytes); /* * If realloc returns NULL, this function loses the original * value of oldptr. This isn't a leak because the caller of * this function still has the original value of oldptr. */ if (!oldptr) { if (bytes) { PR_SetError(PR_OUT_OF_MEMORY_ERROR, 0); pr_ZoneFree(rv); return oldptr; } } phony.s.requestedSize = bytes; mb = &phony; ours = 0; } else { size_t blockSize = mb->s.blockSize; MemBlockHdr *mt = (MemBlockHdr *)(((char *)(mb + 1)) + blockSize); PR_ASSERT(mt->s.magic == ZONE_MAGIC); PR_ASSERT(mt->s.zone == mb->s.zone); PR_ASSERT(mt->s.blockSize == blockSize); if (bytes <= blockSize) { /* The block is already big enough. */ mt->s.requestedSize = mb->s.requestedSize = bytes; return oldptr; } ours = 1; rv = pr_ZoneMalloc(bytes); if (!rv) { return rv; } } if (oldptr && mb->s.requestedSize) { memcpy(rv, oldptr, mb->s.requestedSize); } if (ours) { pr_ZoneFree(oldptr); } else if (oldptr) { free(oldptr); } return rv; } static void pr_ZoneFree(void *ptr) { MemBlockHdr *mb, *mt; MemoryZone *mz; size_t blockSize; PRUint32 wasLocked; if (!ptr) { return; } mb = (MemBlockHdr *)((char *)ptr - (sizeof *mb)); if (mb->s.magic != ZONE_MAGIC) { /* maybe this came from ordinary malloc */ #ifdef DEBUG fprintf(stderr, "Warning: freeing memory block %p from ordinary malloc\n", ptr); #endif free(ptr); return; } blockSize = mb->s.blockSize; mz = mb->s.zone; mt = (MemBlockHdr *)(((char *)(mb + 1)) + blockSize); PR_ASSERT(mt->s.magic == ZONE_MAGIC); PR_ASSERT(mt->s.zone == mz); PR_ASSERT(mt->s.blockSize == blockSize); if (!mz) { PR_ASSERT(blockSize > 65536); /* This block was not in any zone. Just free it. */ free(mb); return; } PR_ASSERT(mz->blockSize == blockSize); wasLocked = mz->locked; pthread_mutex_lock(&mz->lock); mz->locked = 1; if (wasLocked) { mz->contention++; } mt->s.next = mb->s.next = mz->head; /* put on head of list */ mz->head = mb; mz->elements++; mz->locked = 0; pthread_mutex_unlock(&mz->lock); } PR_IMPLEMENT(void *) PR_Malloc(PRUint32 size) { if (!_pr_initialized) { _PR_ImplicitInitialization(); } return use_zone_allocator ? pr_ZoneMalloc(size) : malloc(size); } PR_IMPLEMENT(void *) PR_Calloc(PRUint32 nelem, PRUint32 elsize) { if (!_pr_initialized) { _PR_ImplicitInitialization(); } return use_zone_allocator ? pr_ZoneCalloc(nelem, elsize) : calloc(nelem, elsize); } PR_IMPLEMENT(void *) PR_Realloc(void *ptr, PRUint32 size) { if (!_pr_initialized) { _PR_ImplicitInitialization(); } return use_zone_allocator ? pr_ZoneRealloc(ptr, size) : realloc(ptr, size); } PR_IMPLEMENT(void) PR_Free(void *ptr) { if (use_zone_allocator) { pr_ZoneFree(ptr); } else { free(ptr); } } #else /* !defined(_PR_ZONE_ALLOCATOR) */ /* ** The PR_Malloc, PR_Calloc, PR_Realloc, and PR_Free functions simply ** call their libc equivalents now. This may seem redundant, but it ** ensures that we are calling into the same runtime library. On ** Win32, it is possible to have multiple runtime libraries (e.g., ** objects compiled with /MD and /MDd) in the same process, and ** they maintain separate heaps, which cannot be mixed. */ PR_IMPLEMENT(void *) PR_Malloc(PRUint32 size) { #if defined (WIN16) return PR_MD_malloc( (size_t) size); #else return malloc(size); #endif } PR_IMPLEMENT(void *) PR_Calloc(PRUint32 nelem, PRUint32 elsize) { #if defined (WIN16) return PR_MD_calloc( (size_t)nelem, (size_t)elsize ); #else return calloc(nelem, elsize); #endif } PR_IMPLEMENT(void *) PR_Realloc(void *ptr, PRUint32 size) { #if defined (WIN16) return PR_MD_realloc( ptr, (size_t) size); #else return realloc(ptr, size); #endif } PR_IMPLEMENT(void) PR_Free(void *ptr) { #if defined (WIN16) PR_MD_free( ptr ); #else free(ptr); #endif } #endif /* _PR_ZONE_ALLOCATOR */ /* ** Complexity alert! ** ** If malloc/calloc/free (etc.) were implemented to use pr lock's then ** the entry points could block when called if some other thread had the ** lock. ** ** Most of the time this isn't a problem. However, in the case that we ** are using the thread safe malloc code after PR_Init but before ** PR_AttachThread has been called (on a native thread that nspr has yet ** to be told about) we could get royally screwed if the lock was busy ** and we tried to context switch the thread away. In this scenario ** PR_CURRENT_THREAD() == NULL ** ** To avoid this unfortunate case, we use the low level locking ** facilities for malloc protection instead of the slightly higher level ** locking. This makes malloc somewhat faster so maybe it's a good thing ** anyway. */ #ifdef _PR_OVERRIDE_MALLOC /* Imports */ extern void *_PR_UnlockedMalloc(size_t size); extern void *_PR_UnlockedMemalign(size_t alignment, size_t size); extern void _PR_UnlockedFree(void *ptr); extern void *_PR_UnlockedRealloc(void *ptr, size_t size); extern void *_PR_UnlockedCalloc(size_t n, size_t elsize); static PRBool _PR_malloc_initialised = PR_FALSE; #ifdef _PR_PTHREADS static pthread_mutex_t _PR_MD_malloc_crustylock; #define _PR_Lock_Malloc() { \ if(PR_TRUE == _PR_malloc_initialised) { \ PRStatus rv; \ rv = pthread_mutex_lock(&_PR_MD_malloc_crustylock); \ PR_ASSERT(0 == rv); \ } #define _PR_Unlock_Malloc() if(PR_TRUE == _PR_malloc_initialised) { \ PRStatus rv; \ rv = pthread_mutex_unlock(&_PR_MD_malloc_crustylock); \ PR_ASSERT(0 == rv); \ } \ } #else /* _PR_PTHREADS */ static _MDLock _PR_MD_malloc_crustylock; #define _PR_Lock_Malloc() { \ PRIntn _is; \ if(PR_TRUE == _PR_malloc_initialised) { \ if (_PR_MD_CURRENT_THREAD() && \ !_PR_IS_NATIVE_THREAD( \ _PR_MD_CURRENT_THREAD())) \ _PR_INTSOFF(_is); \ _PR_MD_LOCK(&_PR_MD_malloc_crustylock); \ } #define _PR_Unlock_Malloc() if(PR_TRUE == _PR_malloc_initialised) { \ _PR_MD_UNLOCK(&_PR_MD_malloc_crustylock); \ if (_PR_MD_CURRENT_THREAD() && \ !_PR_IS_NATIVE_THREAD( \ _PR_MD_CURRENT_THREAD())) \ _PR_INTSON(_is); \ } \ } #endif /* _PR_PTHREADS */ PR_IMPLEMENT(PRStatus) _PR_MallocInit(void) { PRStatus rv = PR_SUCCESS; if( PR_TRUE == _PR_malloc_initialised ) { return PR_SUCCESS; } #ifdef _PR_PTHREADS { int status; pthread_mutexattr_t mattr; status = _PT_PTHREAD_MUTEXATTR_INIT(&mattr); PR_ASSERT(0 == status); status = _PT_PTHREAD_MUTEX_INIT(_PR_MD_malloc_crustylock, mattr); PR_ASSERT(0 == status); status = _PT_PTHREAD_MUTEXATTR_DESTROY(&mattr); PR_ASSERT(0 == status); } #else /* _PR_PTHREADS */ _MD_NEW_LOCK(&_PR_MD_malloc_crustylock); #endif /* _PR_PTHREADS */ if( PR_SUCCESS == rv ) { _PR_malloc_initialised = PR_TRUE; } return rv; } void *malloc(size_t size) { void *p; _PR_Lock_Malloc(); p = _PR_UnlockedMalloc(size); _PR_Unlock_Malloc(); return p; } void free(void *ptr) { _PR_Lock_Malloc(); _PR_UnlockedFree(ptr); _PR_Unlock_Malloc(); } void *realloc(void *ptr, size_t size) { void *p; _PR_Lock_Malloc(); p = _PR_UnlockedRealloc(ptr, size); _PR_Unlock_Malloc(); return p; } void *calloc(size_t n, size_t elsize) { void *p; _PR_Lock_Malloc(); p = _PR_UnlockedCalloc(n, elsize); _PR_Unlock_Malloc(); return p; } void cfree(void *p) { _PR_Lock_Malloc(); _PR_UnlockedFree(p); _PR_Unlock_Malloc(); } void _PR_InitMem(void) { PRStatus rv; rv = _PR_MallocInit(); PR_ASSERT(PR_SUCCESS == rv); } #endif /* _PR_OVERRIDE_MALLOC */