sys/sys/spinlock2.h

   1 /*
   2  * Copyright (c) 2005 Jeffrey M. Hsu.  All rights reserved.
   3  *
   4  * This code is derived from software contributed to The DragonFly Project
   5  * by Jeffrey M. Hsu.
   6  *
   7  * Redistribution and use in source and binary forms, with or without
   8  * modification, are permitted provided that the following conditions
   9  * are met:
  10  * 1. Redistributions of source code must retain the above copyright
  11  *    notice, this list of conditions and the following disclaimer.
  12  * 2. Redistributions in binary form must reproduce the above copyright
  13  *    notice, this list of conditions and the following disclaimer in the
  14  *    documentation and/or other materials provided with the distribution.
  15  * 3. Neither the name of The DragonFly Project nor the names of its
  16  *    contributors may be used to endorse or promote products derived
  17  *    from this software without specific, prior written permission.
  18  *
  19  * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
  20  * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
  21  * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS
  22  * FOR A PARTICULAR PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL THE
  23  * COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
  24  * INCIDENTAL, SPECIAL, EXEMPLARY OR CONSEQUENTIAL DAMAGES (INCLUDING,
  25  * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
  26  * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED
  27  * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
  28  * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT
  29  * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
  30  * SUCH DAMAGE.
  31  *
  32  * $DragonFly: src/sys/sys/spinlock2.h,v 1.11 2006/06/01 21:02:39 dillon Exp $
  33  */
  34
  35 #ifndef _SYS_SPINLOCK2_H_
  36 #define _SYS_SPINLOCK2_H_
  37
  38 #ifndef _KERNEL
  39
  40 #error "This file should not be included by userland programs."
  41
  42 #else
  43
  44 #ifndef _SYS_SYSTM_H_
  45 #include <sys/systm.h>
  46 #endif
  47 #ifndef _SYS_THREAD2_H_
  48 #include <sys/thread2.h>
  49 #endif
  50 #ifndef _SYS_GLOBALDATA_H_
  51 #include <sys/globaldata.h>
  52 #endif
  53 #ifndef _MACHINE_ATOMIC_H_
  54 #include <machine/atomic.h>
  55 #endif
  56 #ifndef _MACHINE_CPUFUNC_H_
  57 #include <machine/cpufunc.h>
  58 #endif
  59
  60 /*
  61  * SPECIAL NOTE!  Obtaining a spinlock does not enter a critical section
  62  * or protect against FAST interrupts but it will prevent thread preemption.
  63  * Because the spinlock code path is ultra critical, we do not check for
  64  * LWKT reschedule requests (due to an interrupt thread not being able to
  65  * preempt).
  66  */
  67
  68 #ifdef SMP
  69
  70 extern int spin_trylock_wr_contested(struct spinlock *mtx, int value);
  71 extern void spin_lock_wr_contested(struct spinlock *mtx, int value);
  72 extern void spin_lock_rd_contested(struct spinlock *mtx);
  73
  74 #endif
  75
  76 #ifdef SMP
  77
  78 /*
  79  * Attempt to obtain an exclusive spinlock.  Returns FALSE on failure,
  80  * TRUE on success.  Since the caller assumes that spinlocks must actually
  81  * work when using this function, it is only made available to SMP builds.
  82  */
  83 static __inline boolean_t
  84 spin_trylock_wr(struct spinlock *mtx)
  85 {
  86         globaldata_t gd = mycpu;
  87         int value;
  88
  89         ++gd->gd_spinlocks_wr;
  90         if ((value = atomic_swap_int(&mtx->lock, SPINLOCK_EXCLUSIVE)) != 0)
  91                 return (spin_trylock_wr_contested(mtx, value));
  92         return (TRUE);
  93 }
  94
  95 #endif
  96
  97 /*
  98  * Obtain an exclusive spinlock and return.  Shortcut the case where the only
  99  * cached read lock was from our own cpu (it can just be cleared).
 100  */
 101 static __inline void
 102 spin_lock_wr_quick(globaldata_t gd, struct spinlock *mtx)
 103 {
 104 #ifdef SMP
 105         int value;
 106 #endif
 107
 108         ++gd->gd_spinlocks_wr;
 109 #ifdef SMP
 110         if ((value = atomic_swap_int(&mtx->lock, SPINLOCK_EXCLUSIVE)) != 0) {
 111                 value &= ~gd->gd_cpumask;
 112                 if (value)
 113                         spin_lock_wr_contested(mtx, value);
 114         }
 115 #endif
 116 }
 117
 118 static __inline void
 119 spin_lock_wr(struct spinlock *mtx)
 120 {
 121         spin_lock_wr_quick(mycpu, mtx);
 122 }
 123
 124 #if 0
 125
 126 /*
 127  * Upgrade a shared spinlock to exclusive.  Return TRUE if we were
 128  * able to upgrade without another exclusive holder getting in before
 129  * us, FALSE otherwise.
 130  */
 131 static __inline int
 132 spin_lock_upgrade(struct spinlock *mtx)
 133 {
 134         globaldata_t gd = mycpu;
 135 #ifdef SMP
 136         int value;
 137 #endif
 138
 139         ++gd->gd_spinlocks_wr;
 140 #ifdef SMP
 141         value = atomic_swap_int(&mtx->lock, SPINLOCK_EXCLUSIVE);
 142         cpu_sfence();
 143 #endif
 144         gd->gd_spinlock_rd = NULL;
 145 #ifdef SMP
 146         value &= ~gd->gd_cpumask;
 147         if (value) {
 148                 spin_lock_wr_contested(mtx, value);
 149                 if (value & SPINLOCK_EXCLUSIVE)
 150                         return (FALSE);
 151                 XXX regain original shared lock?
 152         }
 153         return (TRUE);
 154 #endif
 155 }
 156
 157 #endif
 158
 159 /*
 160  * Obtain a shared spinlock and return.  This is a critical code path.
 161  *
 162  * The vast majority of the overhead is in the cpu_mfence() (5ns vs 1ns for
 163  * the entire rest of the procedure).  Unfortunately we have to ensure that
 164  * spinlock pointer is written out before we check the cpumask to interlock
 165  * against an exclusive spinlock that clears the cpumask and then checks
 166  * the spinlock pointer.
 167  *
 168  * But what is EXTREMELY important here is that we do not have to perform
 169  * a locked bus cycle on the spinlock itself if the shared bit for our cpu
 170  * is already found to be set.  We only need the mfence, and the mfence is
 171  * local to the cpu and never conflicts with other cpu's.
 172  *
 173  * This means that multiple parallel shared acessors (e.g. filedescriptor
 174  * table lookups, namecache lookups) run at full speed and incur NO cache
 175  * contention at all.  It is the difference between 10ns and 40-100ns.
 176  */
 177 static __inline void
 178 spin_lock_rd_quick(globaldata_t gd, struct spinlock *mtx)
 179 {
 180         gd->gd_spinlock_rd = mtx;
 181 #ifdef SMP
 182         cpu_mfence();
 183         if ((mtx->lock & gd->gd_cpumask) == 0)
 184                 spin_lock_rd_contested(mtx);
 185 #endif
 186 }
 187
 188 static __inline void
 189 spin_lock_rd(struct spinlock *mtx)
 190 {
 191         spin_lock_rd_quick(mycpu,mtx);
 192 }
 193
 194 /*
 195  * Release an exclusive spinlock.  We can just do this passively, only
 196  * ensuring that our spinlock count is left intact until the mutex is
 197  * cleared.
 198  */
 199 static __inline void
 200 spin_unlock_wr_quick(globaldata_t gd, struct spinlock *mtx)
 201 {
 202 #ifdef SMP
 203         mtx->lock = 0;
 204 #endif
 205         --gd->gd_spinlocks_wr;
 206 }
 207
 208 static __inline void
 209 spin_unlock_wr(struct spinlock *mtx)
 210 {
 211         spin_unlock_wr_quick(mycpu, mtx);
 212 }
 213
 214 /*
 215  * Release a shared spinlock.  We leave the shared bit set in the spinlock
 216  * as a cache and simply clear the spinlock pointer for the cpu.  This
 217  * fast-paths another shared lock later at the cost of an exclusive lock
 218  * having to check per-cpu spinlock pointers to determine when there are no
 219  * shared holders remaining.
 220  */
 221 static __inline void
 222 spin_unlock_rd_quick(globaldata_t gd, struct spinlock *mtx)
 223 {
 224         KKASSERT(gd->gd_spinlock_rd == mtx);
 225         gd->gd_spinlock_rd = NULL;
 226 }
 227
 228 static __inline void
 229 spin_unlock_rd(struct spinlock *mtx)
 230 {
 231         spin_unlock_rd_quick(mycpu, mtx);
 232 }
 233
 234 static __inline void
 235 spin_init(struct spinlock *mtx)
 236 {
 237         mtx->lock = 0;
 238 }
 239
 240 static __inline void
 241 spin_uninit(struct spinlock *mtx)
 242 {
 243         /* unused */
 244 }
 245
 246 #endif  /* _KERNEL */
 247 #endif  /* _SYS_SPINLOCK2_H_ */
 248