michael@0: /* -*- Mode: C++; tab-width: 4; indent-tabs-mode: nil; c-basic-offset: 2 -*- */
michael@0: /* This Source Code Form is subject to the terms of the Mozilla Public
michael@0:  * License, v. 2.0. If a copy of the MPL was not distributed with this
michael@0:  * file, You can obtain one at http://mozilla.org/MPL/2.0/. */
michael@0: 
michael@0: /*
michael@0: **     PR Atomic operations
michael@0: */
michael@0: 
michael@0: 
michael@0: #include "pratom.h"
michael@0: #include "primpl.h"
michael@0: 
michael@0: #include <string.h>
michael@0: 
michael@0: /*
michael@0:  * The following is a fallback implementation that emulates
michael@0:  * atomic operations for platforms without atomic operations.
michael@0:  * If a platform has atomic operations, it should define the
michael@0:  * macro _PR_HAVE_ATOMIC_OPS, and the following will not be
michael@0:  * compiled in.
michael@0:  */
michael@0: 
michael@0: #if !defined(_PR_HAVE_ATOMIC_OPS)
michael@0: 
michael@0: #if defined(_PR_PTHREADS) && !defined(_PR_DCETHREADS)
michael@0: /*
michael@0:  * PR_AtomicDecrement() is used in NSPR's thread-specific data
michael@0:  * destructor.  Because thread-specific data destructors may be
michael@0:  * invoked after a PR_Cleanup() call, we need an implementation
michael@0:  * of the atomic routines that doesn't need NSPR to be initialized.
michael@0:  */
michael@0: 
michael@0: /*
michael@0:  * We use a set of locks for all the emulated atomic operations.
michael@0:  * By hashing on the address of the integer to be locked the
michael@0:  * contention between multiple threads should be lessened.
michael@0:  *
michael@0:  * The number of atomic locks can be set by the environment variable
michael@0:  * NSPR_ATOMIC_HASH_LOCKS
michael@0:  */
michael@0: 
michael@0: /*
michael@0:  * lock counts should be a power of 2
michael@0:  */
michael@0: #define DEFAULT_ATOMIC_LOCKS	16	/* should be in sync with the number of initializers
michael@0: 										below */
michael@0: #define MAX_ATOMIC_LOCKS		(4 * 1024)
michael@0: 
michael@0: static pthread_mutex_t static_atomic_locks[DEFAULT_ATOMIC_LOCKS] = {
michael@0:         PTHREAD_MUTEX_INITIALIZER, PTHREAD_MUTEX_INITIALIZER,
michael@0:         PTHREAD_MUTEX_INITIALIZER, PTHREAD_MUTEX_INITIALIZER,
michael@0:         PTHREAD_MUTEX_INITIALIZER, PTHREAD_MUTEX_INITIALIZER,
michael@0:         PTHREAD_MUTEX_INITIALIZER, PTHREAD_MUTEX_INITIALIZER,
michael@0:         PTHREAD_MUTEX_INITIALIZER, PTHREAD_MUTEX_INITIALIZER,
michael@0:         PTHREAD_MUTEX_INITIALIZER, PTHREAD_MUTEX_INITIALIZER,
michael@0:         PTHREAD_MUTEX_INITIALIZER, PTHREAD_MUTEX_INITIALIZER,
michael@0:         PTHREAD_MUTEX_INITIALIZER, PTHREAD_MUTEX_INITIALIZER };
michael@0: 
michael@0: #ifdef DEBUG
michael@0: static PRInt32 static_hash_lock_counts[DEFAULT_ATOMIC_LOCKS];
michael@0: static PRInt32 *hash_lock_counts = static_hash_lock_counts;
michael@0: #endif
michael@0: 
michael@0: static PRUint32	num_atomic_locks = DEFAULT_ATOMIC_LOCKS;
michael@0: static pthread_mutex_t *atomic_locks = static_atomic_locks;
michael@0: static PRUint32 atomic_hash_mask = DEFAULT_ATOMIC_LOCKS - 1;
michael@0: 
michael@0: #define _PR_HASH_FOR_LOCK(ptr) 							\
michael@0: 			((PRUint32) (((PRUptrdiff) (ptr) >> 2)	^	\
michael@0: 						((PRUptrdiff) (ptr) >> 8)) &	\
michael@0: 						atomic_hash_mask)
michael@0: 
michael@0: void _PR_MD_INIT_ATOMIC()
michael@0: {
michael@0: char *eval;
michael@0: int index;
michael@0: 
michael@0: 
michael@0: 	PR_ASSERT(PR_FloorLog2(MAX_ATOMIC_LOCKS) ==
michael@0: 						PR_CeilingLog2(MAX_ATOMIC_LOCKS));
michael@0: 
michael@0: 	PR_ASSERT(PR_FloorLog2(DEFAULT_ATOMIC_LOCKS) ==
michael@0: 							PR_CeilingLog2(DEFAULT_ATOMIC_LOCKS));
michael@0: 
michael@0: 	if (((eval = getenv("NSPR_ATOMIC_HASH_LOCKS")) != NULL)  &&
michael@0: 		((num_atomic_locks = atoi(eval)) != DEFAULT_ATOMIC_LOCKS)) {
michael@0: 
michael@0: 		if (num_atomic_locks > MAX_ATOMIC_LOCKS)
michael@0: 			num_atomic_locks = MAX_ATOMIC_LOCKS;
michael@0: 		else if (num_atomic_locks < 1) 
michael@0: 			num_atomic_locks = 1;
michael@0: 		else {
michael@0: 			num_atomic_locks = PR_FloorLog2(num_atomic_locks);
michael@0: 			num_atomic_locks = 1L << num_atomic_locks;
michael@0: 		}
michael@0: 		atomic_locks = (pthread_mutex_t *) PR_Malloc(sizeof(pthread_mutex_t) *
michael@0: 						num_atomic_locks);
michael@0: 		if (atomic_locks) {
michael@0: 			for (index = 0; index < num_atomic_locks; index++) {
michael@0: 				if (pthread_mutex_init(&atomic_locks[index], NULL)) {
michael@0: 						PR_DELETE(atomic_locks);
michael@0: 						atomic_locks = NULL;
michael@0: 						break; 
michael@0: 				}
michael@0: 			}
michael@0: 		}
michael@0: #ifdef DEBUG
michael@0: 		if (atomic_locks) {
michael@0: 			hash_lock_counts = PR_CALLOC(num_atomic_locks * sizeof(PRInt32));
michael@0: 			if (hash_lock_counts == NULL) {
michael@0: 				PR_DELETE(atomic_locks);
michael@0: 				atomic_locks = NULL;
michael@0: 			}
michael@0: 		}
michael@0: #endif
michael@0: 		if (atomic_locks == NULL) {
michael@0: 			/*
michael@0: 			 *	Use statically allocated locks
michael@0: 			 */
michael@0: 			atomic_locks = static_atomic_locks;
michael@0: 			num_atomic_locks = DEFAULT_ATOMIC_LOCKS;
michael@0: 	#ifdef DEBUG
michael@0: 			hash_lock_counts = static_hash_lock_counts;
michael@0: 	#endif
michael@0: 		}
michael@0: 		atomic_hash_mask = num_atomic_locks - 1;
michael@0: 	}
michael@0: 	PR_ASSERT(PR_FloorLog2(num_atomic_locks) ==
michael@0: 								PR_CeilingLog2(num_atomic_locks));
michael@0: }
michael@0: 
michael@0: PRInt32
michael@0: _PR_MD_ATOMIC_INCREMENT(PRInt32 *val)
michael@0: {
michael@0:     PRInt32 rv;
michael@0:     PRInt32 idx = _PR_HASH_FOR_LOCK(val);
michael@0: 
michael@0:     pthread_mutex_lock(&atomic_locks[idx]);
michael@0:     rv = ++(*val);
michael@0: #ifdef DEBUG
michael@0:     hash_lock_counts[idx]++;
michael@0: #endif
michael@0:     pthread_mutex_unlock(&atomic_locks[idx]);
michael@0:     return rv;
michael@0: }
michael@0: 
michael@0: PRInt32
michael@0: _PR_MD_ATOMIC_ADD(PRInt32 *ptr, PRInt32 val)
michael@0: {
michael@0:     PRInt32 rv;
michael@0:     PRInt32 idx = _PR_HASH_FOR_LOCK(ptr);
michael@0: 
michael@0:     pthread_mutex_lock(&atomic_locks[idx]);
michael@0:     rv = ((*ptr) += val);
michael@0: #ifdef DEBUG
michael@0:     hash_lock_counts[idx]++;
michael@0: #endif
michael@0:     pthread_mutex_unlock(&atomic_locks[idx]);
michael@0:     return rv;
michael@0: }
michael@0: 
michael@0: PRInt32
michael@0: _PR_MD_ATOMIC_DECREMENT(PRInt32 *val)
michael@0: {
michael@0:     PRInt32 rv;
michael@0:     PRInt32 idx = _PR_HASH_FOR_LOCK(val);
michael@0: 
michael@0:     pthread_mutex_lock(&atomic_locks[idx]);
michael@0:     rv = --(*val);
michael@0: #ifdef DEBUG
michael@0:     hash_lock_counts[idx]++;
michael@0: #endif
michael@0:     pthread_mutex_unlock(&atomic_locks[idx]);
michael@0:     return rv;
michael@0: }
michael@0: 
michael@0: PRInt32
michael@0: _PR_MD_ATOMIC_SET(PRInt32 *val, PRInt32 newval)
michael@0: {
michael@0:     PRInt32 rv;
michael@0:     PRInt32 idx = _PR_HASH_FOR_LOCK(val);
michael@0: 
michael@0:     pthread_mutex_lock(&atomic_locks[idx]);
michael@0:     rv = *val;
michael@0:     *val = newval;
michael@0: #ifdef DEBUG
michael@0:     hash_lock_counts[idx]++;
michael@0: #endif
michael@0:     pthread_mutex_unlock(&atomic_locks[idx]);
michael@0:     return rv;
michael@0: }
michael@0: #else  /* _PR_PTHREADS && !_PR_DCETHREADS */
michael@0: /*
michael@0:  * We use a single lock for all the emulated atomic operations.
michael@0:  * The lock contention should be acceptable.
michael@0:  */
michael@0: static PRLock *atomic_lock = NULL;
michael@0: void _PR_MD_INIT_ATOMIC(void)
michael@0: {
michael@0:     if (atomic_lock == NULL) {
michael@0:         atomic_lock = PR_NewLock();
michael@0:     }
michael@0: }
michael@0: 
michael@0: PRInt32
michael@0: _PR_MD_ATOMIC_INCREMENT(PRInt32 *val)
michael@0: {
michael@0:     PRInt32 rv;
michael@0: 
michael@0:     if (!_pr_initialized) {
michael@0:         _PR_ImplicitInitialization();
michael@0:     }
michael@0:     PR_Lock(atomic_lock);
michael@0:     rv = ++(*val);
michael@0:     PR_Unlock(atomic_lock);
michael@0:     return rv;
michael@0: }
michael@0: 
michael@0: PRInt32
michael@0: _PR_MD_ATOMIC_ADD(PRInt32 *ptr, PRInt32 val)
michael@0: {
michael@0:     PRInt32 rv;
michael@0: 
michael@0:     if (!_pr_initialized) {
michael@0:         _PR_ImplicitInitialization();
michael@0:     }
michael@0:     PR_Lock(atomic_lock);
michael@0:     rv = ((*ptr) += val);
michael@0:     PR_Unlock(atomic_lock);
michael@0:     return rv;
michael@0: }
michael@0: 
michael@0: PRInt32
michael@0: _PR_MD_ATOMIC_DECREMENT(PRInt32 *val)
michael@0: {
michael@0:     PRInt32 rv;
michael@0: 
michael@0:     if (!_pr_initialized) {
michael@0:         _PR_ImplicitInitialization();
michael@0:     }
michael@0:     PR_Lock(atomic_lock);
michael@0:     rv = --(*val);
michael@0:     PR_Unlock(atomic_lock);
michael@0:     return rv;
michael@0: }
michael@0: 
michael@0: PRInt32
michael@0: _PR_MD_ATOMIC_SET(PRInt32 *val, PRInt32 newval)
michael@0: {
michael@0:     PRInt32 rv;
michael@0: 
michael@0:     if (!_pr_initialized) {
michael@0:         _PR_ImplicitInitialization();
michael@0:     }
michael@0:     PR_Lock(atomic_lock);
michael@0:     rv = *val;
michael@0:     *val = newval;
michael@0:     PR_Unlock(atomic_lock);
michael@0:     return rv;
michael@0: }
michael@0: #endif  /* _PR_PTHREADS && !_PR_DCETHREADS */
michael@0: 
michael@0: #endif  /* !_PR_HAVE_ATOMIC_OPS */
michael@0: 
michael@0: void _PR_InitAtomic(void)
michael@0: {
michael@0:     _PR_MD_INIT_ATOMIC();
michael@0: }
michael@0: 
michael@0: PR_IMPLEMENT(PRInt32)
michael@0: PR_AtomicIncrement(PRInt32 *val)
michael@0: {
michael@0:     return _PR_MD_ATOMIC_INCREMENT(val);
michael@0: }
michael@0: 
michael@0: PR_IMPLEMENT(PRInt32)
michael@0: PR_AtomicDecrement(PRInt32 *val)
michael@0: {
michael@0:     return _PR_MD_ATOMIC_DECREMENT(val);
michael@0: }
michael@0: 
michael@0: PR_IMPLEMENT(PRInt32)
michael@0: PR_AtomicSet(PRInt32 *val, PRInt32 newval)
michael@0: {
michael@0:     return _PR_MD_ATOMIC_SET(val, newval);
michael@0: }
michael@0: 
michael@0: PR_IMPLEMENT(PRInt32)
michael@0: PR_AtomicAdd(PRInt32 *ptr, PRInt32 val)
michael@0: {
michael@0:     return _PR_MD_ATOMIC_ADD(ptr, val);
michael@0: }
michael@0: /*
michael@0:  * For platforms, which don't support the CAS (compare-and-swap) instruction
michael@0:  * (or an equivalent), the stack operations are implemented by use of PRLock
michael@0:  */
michael@0: 
michael@0: PR_IMPLEMENT(PRStack *)
michael@0: PR_CreateStack(const char *stack_name)
michael@0: {
michael@0: PRStack *stack;
michael@0: 
michael@0:     if (!_pr_initialized) {
michael@0:         _PR_ImplicitInitialization();
michael@0:     }
michael@0: 
michael@0:     if ((stack = PR_NEW(PRStack)) == NULL) {
michael@0: 		return NULL;
michael@0: 	}
michael@0: 	if (stack_name) {
michael@0: 		stack->prstk_name = (char *) PR_Malloc(strlen(stack_name) + 1);
michael@0: 		if (stack->prstk_name == NULL) {
michael@0: 			PR_DELETE(stack);
michael@0: 			return NULL;
michael@0: 		}
michael@0: 		strcpy(stack->prstk_name, stack_name);
michael@0: 	} else
michael@0: 		stack->prstk_name = NULL;
michael@0: 
michael@0: #ifndef _PR_HAVE_ATOMIC_CAS
michael@0:     stack->prstk_lock = PR_NewLock();
michael@0: 	if (stack->prstk_lock == NULL) {
michael@0: 		PR_Free(stack->prstk_name);
michael@0: 		PR_DELETE(stack);
michael@0: 		return NULL;
michael@0: 	}
michael@0: #endif /* !_PR_HAVE_ATOMIC_CAS */
michael@0: 
michael@0: 	stack->prstk_head.prstk_elem_next = NULL;
michael@0: 	
michael@0:     return stack;
michael@0: }
michael@0: 
michael@0: PR_IMPLEMENT(PRStatus)
michael@0: PR_DestroyStack(PRStack *stack)
michael@0: {
michael@0: 	if (stack->prstk_head.prstk_elem_next != NULL) {
michael@0: 		PR_SetError(PR_INVALID_STATE_ERROR, 0);
michael@0: 		return PR_FAILURE;
michael@0: 	}
michael@0: 
michael@0: 	if (stack->prstk_name)
michael@0: 		PR_Free(stack->prstk_name);
michael@0: #ifndef _PR_HAVE_ATOMIC_CAS
michael@0: 	PR_DestroyLock(stack->prstk_lock);
michael@0: #endif /* !_PR_HAVE_ATOMIC_CAS */
michael@0: 	PR_DELETE(stack);
michael@0: 
michael@0: 	return PR_SUCCESS;
michael@0: }
michael@0: 
michael@0: #ifndef _PR_HAVE_ATOMIC_CAS
michael@0: 
michael@0: PR_IMPLEMENT(void)
michael@0: PR_StackPush(PRStack *stack, PRStackElem *stack_elem)
michael@0: {
michael@0:     PR_Lock(stack->prstk_lock);
michael@0: 	stack_elem->prstk_elem_next = stack->prstk_head.prstk_elem_next;
michael@0: 	stack->prstk_head.prstk_elem_next = stack_elem;
michael@0:     PR_Unlock(stack->prstk_lock);
michael@0:     return;
michael@0: }
michael@0: 
michael@0: PR_IMPLEMENT(PRStackElem *)
michael@0: PR_StackPop(PRStack *stack)
michael@0: {
michael@0: PRStackElem *element;
michael@0: 
michael@0:     PR_Lock(stack->prstk_lock);
michael@0: 	element = stack->prstk_head.prstk_elem_next;
michael@0: 	if (element != NULL) {
michael@0: 		stack->prstk_head.prstk_elem_next = element->prstk_elem_next;
michael@0: 		element->prstk_elem_next = NULL;	/* debugging aid */
michael@0: 	}
michael@0:     PR_Unlock(stack->prstk_lock);
michael@0:     return element;
michael@0: }
michael@0: #endif /* !_PR_HAVE_ATOMIC_CAS */