1 /* $NetBSD: lock.h,v 1.23 2008/01/09 00:23:18 yamt Exp $ */ 2 3 /*- 4 * Copyright (c) 2000, 2006 The NetBSD Foundation, Inc. 5 * All rights reserved. 6 * 7 * This code is derived from software contributed to The NetBSD Foundation 8 * by Jason R. Thorpe and Andrew Doran. 9 * 10 * Redistribution and use in source and binary forms, with or without 11 * modification, are permitted provided that the following conditions 12 * are met: 13 * 1. Redistributions of source code must retain the above copyright 14 * notice, this list of conditions and the following disclaimer. 15 * 2. Redistributions in binary form must reproduce the above copyright 16 * notice, this list of conditions and the following disclaimer in the 17 * documentation and/or other materials provided with the distribution. 18 * 3. All advertising materials mentioning features or use of this software 19 * must display the following acknowledgement: 20 * This product includes software developed by the NetBSD 21 * Foundation, Inc. and its contributors. 22 * 4. Neither the name of The NetBSD Foundation nor the names of its 23 * contributors may be used to endorse or promote products derived 24 * from this software without specific prior written permission. 25 * 26 * THIS SOFTWARE IS PROVIDED BY THE NETBSD FOUNDATION, INC. AND CONTRIBUTORS 27 * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED 28 * TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR 29 * PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE FOUNDATION OR CONTRIBUTORS 30 * BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR 31 * CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF 32 * SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS 33 * INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN 34 * CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) 35 * ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE 36 * POSSIBILITY OF SUCH DAMAGE. 37 */ 38 39 /* 40 * Machine-dependent spin lock operations. 41 */ 42 43 #ifndef _X86_LOCK_H_ 44 #define _X86_LOCK_H_ 45 46 static __inline int 47 __SIMPLELOCK_LOCKED_P(__cpu_simple_lock_t *__ptr) 48 { 49 return *__ptr == __SIMPLELOCK_LOCKED; 50 } 51 52 static __inline int 53 __SIMPLELOCK_UNLOCKED_P(__cpu_simple_lock_t *__ptr) 54 { 55 return *__ptr == __SIMPLELOCK_UNLOCKED; 56 } 57 58 static __inline void 59 __cpu_simple_lock_set(__cpu_simple_lock_t *__ptr) 60 { 61 62 *__ptr = __SIMPLELOCK_LOCKED; 63 } 64 65 static __inline void 66 __cpu_simple_lock_clear(__cpu_simple_lock_t *__ptr) 67 { 68 69 *__ptr = __SIMPLELOCK_UNLOCKED; 70 } 71 72 #ifdef _KERNEL 73 74 #include <machine/cpufunc.h> 75 76 void __cpu_simple_lock_init(__cpu_simple_lock_t *); 77 void __cpu_simple_lock(__cpu_simple_lock_t *); 78 int __cpu_simple_lock_try(__cpu_simple_lock_t *); 79 void __cpu_simple_unlock(__cpu_simple_lock_t *); 80 81 #define SPINLOCK_SPIN_HOOK /* nothing */ 82 83 #ifdef SPINLOCK_BACKOFF_HOOK 84 #undef SPINLOCK_BACKOFF_HOOK 85 #endif 86 #define SPINLOCK_BACKOFF_HOOK x86_pause() 87 88 #else 89 90 static __inline void __cpu_simple_lock_init(__cpu_simple_lock_t *) 91 __unused; 92 static __inline void __cpu_simple_lock(__cpu_simple_lock_t *) 93 __unused; 94 static __inline int __cpu_simple_lock_try(__cpu_simple_lock_t *) 95 __unused; 96 static __inline void __cpu_simple_unlock(__cpu_simple_lock_t *) 97 __unused; 98 99 static __inline void 100 __cpu_simple_lock_init(__cpu_simple_lock_t *lockp) 101 { 102 103 *lockp = __SIMPLELOCK_UNLOCKED; 104 __insn_barrier(); 105 } 106 107 static __inline int 108 __cpu_simple_lock_try(__cpu_simple_lock_t *lockp) 109 { 110 uint8_t val; 111 112 val = __SIMPLELOCK_LOCKED; 113 __asm volatile ("xchgb %0,(%2)" : 114 "=r" (val) 115 :"0" (val), "r" (lockp)); 116 __insn_barrier(); 117 return val == __SIMPLELOCK_UNLOCKED; 118 } 119 120 static __inline void 121 __cpu_simple_lock(__cpu_simple_lock_t *lockp) 122 { 123 124 while (!__cpu_simple_lock_try(lockp)) 125 /* nothing */; 126 __insn_barrier(); 127 } 128 129 /* 130 * Note on x86 memory ordering 131 * 132 * When releasing a lock we must ensure that no stores or loads from within 133 * the critical section are re-ordered by the CPU to occur outside of it: 134 * they must have completed and be visible to other processors once the lock 135 * has been released. 136 * 137 * NetBSD usually runs with the kernel mapped (via MTRR) in a WB (write 138 * back) memory region. In that case, memory ordering on x86 platforms 139 * looks like this: 140 * 141 * i386 All loads/stores occur in instruction sequence. 142 * 143 * i486 All loads/stores occur in instruction sequence. In 144 * Pentium exceptional circumstances, loads can be re-ordered around 145 * stores, but for the purposes of releasing a lock it does 146 * not matter. Stores may not be immediately visible to other 147 * processors as they can be buffered. However, since the 148 * stores are buffered in order the lock release will always be 149 * the last operation in the critical section that becomes 150 * visible to other CPUs. 151 * 152 * Pentium Pro The "Intel 64 and IA-32 Architectures Software Developer's 153 * onwards Manual" volume 3A (order number 248966) says that (1) "Reads 154 * can be carried out speculatively and in any order" and (2) 155 * "Reads can pass buffered stores, but the processor is 156 * self-consistent.". This would be a problem for the below, 157 * and would mandate a locked instruction cycle or load fence 158 * before releasing the simple lock. 159 * 160 * The "Intel Pentium 4 Processor Optimization" guide (order 161 * number 253668-022US) says: "Loads can be moved before stores 162 * that occurred earlier in the program if they are not 163 * predicted to load from the same linear address.". This is 164 * not a problem since the only loads that can be re-ordered 165 * take place once the lock has been released via a store. 166 * 167 * The above two documents seem to contradict each other, 168 * however with the exception of early steppings of the Pentium 169 * Pro, the second document is closer to the truth: a store 170 * will always act as a load fence for all loads that precede 171 * the store in instruction order. 172 * 173 * Again, note that stores can be buffered and will not always 174 * become immediately visible to other CPUs: they are however 175 * buffered in order. 176 * 177 * AMD64 Stores occur in order and are buffered. Loads can be 178 * reordered, however stores act as load fences, meaning that 179 * loads can not be reordered around stores. 180 */ 181 static __inline void 182 __cpu_simple_unlock(__cpu_simple_lock_t *lockp) 183 { 184 185 __insn_barrier(); 186 *lockp = __SIMPLELOCK_UNLOCKED; 187 } 188 189 #endif /* _KERNEL */ 190 191 #endif /* _X86_LOCK_H_ */ 192