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 1.12 2008/06/04 04:34:54 nth 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(globaldata_t gd, struct spinlock *mtx,
  71     int value);
  72 extern void spin_lock_wr_contested(struct spinlock *mtx, int value);
  73 extern void spin_lock_rd_contested(struct spinlock *mtx);
  74
  75 #endif
  76
  77 #ifdef SMP
  78
  79 /*
  80  * Attempt to obtain an exclusive spinlock.  Returns FALSE on failure,
  81  * TRUE on success.  Since the caller assumes that spinlocks must actually
  82  * work when using this function, it is only made available to SMP builds.
  83  */
  84 static __inline boolean_t
  85 spin_trylock_wr(struct spinlock *mtx)
  86 {
  87         globaldata_t gd = mycpu;
  88         int value;
  89
  90         ++gd->gd_spinlocks_wr;
  91         if ((value = atomic_swap_int(&mtx->lock, SPINLOCK_EXCLUSIVE)) != 0)
  92                 return (spin_trylock_wr_contested(gd, mtx, value));
  93         return (TRUE);
  94 }
  95
  96 #endif
  97
  98 /*
  99  * Obtain an exclusive spinlock and return.  Shortcut the case where the only
 100  * cached read lock was from our own cpu (it can just be cleared).
 101  */
 102 static __inline void
 103 spin_lock_wr_quick(globaldata_t gd, struct spinlock *mtx)
 104 {
 105 #ifdef SMP
 106         int value;
 107 #endif
 108
 109         ++gd->gd_spinlocks_wr;
 110 #ifdef SMP
 111         if ((value = atomic_swap_int(&mtx->lock, SPINLOCK_EXCLUSIVE)) != 0) {
 112                 value &= ~gd->gd_cpumask;
 113                 if (value)
 114                         spin_lock_wr_contested(mtx, value);
 115         }
 116 #endif
 117 }
 118
 119 static __inline void
 120 spin_lock_wr(struct spinlock *mtx)
 121 {
 122         spin_lock_wr_quick(mycpu, mtx);
 123 }
 124
 125 #if 0
 126
 127 /*
 128  * Upgrade a shared spinlock to exclusive.  Return TRUE if we were
 129  * able to upgrade without another exclusive holder getting in before
 130  * us, FALSE otherwise.
 131  */
 132 static __inline int
 133 spin_lock_upgrade(struct spinlock *mtx)
 134 {
 135         globaldata_t gd = mycpu;
 136 #ifdef SMP
 137         int value;
 138 #endif
 139
 140         ++gd->gd_spinlocks_wr;
 141 #ifdef SMP
 142         value = atomic_swap_int(&mtx->lock, SPINLOCK_EXCLUSIVE);
 143         cpu_sfence();
 144 #endif
 145         gd->gd_spinlock_rd = NULL;
 146 #ifdef SMP
 147         value &= ~gd->gd_cpumask;
 148         if (value) {
 149                 spin_lock_wr_contested(mtx, value);
 150                 if (value & SPINLOCK_EXCLUSIVE)
 151                         return (FALSE);
 152                 XXX regain original shared lock?
 153         }
 154         return (TRUE);
 155 #endif
 156 }
 157
 158 #endif
 159
 160 /*
 161  * Obtain a shared spinlock and return.  This is a critical code path.
 162  *
 163  * The vast majority of the overhead is in the cpu_mfence() (5ns vs 1ns for
 164  * the entire rest of the procedure).  Unfortunately we have to ensure that
 165  * spinlock pointer is written out before we check the cpumask to interlock
 166  * against an exclusive spinlock that clears the cpumask and then checks
 167  * the spinlock pointer.
 168  *
 169  * But what is EXTREMELY important here is that we do not have to perform
 170  * a locked bus cycle on the spinlock itself if the shared bit for our cpu
 171  * is already found to be set.  We only need the mfence, and the mfence is
 172  * local to the cpu and never conflicts with other cpu's.
 173  *
 174  * This means that multiple parallel shared acessors (e.g. filedescriptor
 175  * table lookups, namecache lookups) run at full speed and incur NO cache
 176  * contention at all.  It is the difference between 10ns and 40-100ns.
 177  */
 178 static __inline void
 179 spin_lock_rd_quick(globaldata_t gd, struct spinlock *mtx)
 180 {
 181         gd->gd_spinlock_rd = mtx;
 182 #ifdef SMP
 183         cpu_mfence();
 184         if ((mtx->lock & gd->gd_cpumask) == 0)
 185                 spin_lock_rd_contested(mtx);
 186 #endif
 187 }
 188
 189 static __inline void
 190 spin_lock_rd(struct spinlock *mtx)
 191 {
 192         spin_lock_rd_quick(mycpu,mtx);
 193 }
 194
 195 /*
 196  * Release an exclusive spinlock.  We can just do this passively, only
 197  * ensuring that our spinlock count is left intact until the mutex is
 198  * cleared.
 199  */
 200 static __inline void
 201 spin_unlock_wr_quick(globaldata_t gd, struct spinlock *mtx)
 202 {
 203 #ifdef SMP
 204         mtx->lock = 0;
 205 #endif
 206         --gd->gd_spinlocks_wr;
 207 }
 208
 209 static __inline void
 210 spin_unlock_wr(struct spinlock *mtx)
 211 {
 212         spin_unlock_wr_quick(mycpu, mtx);
 213 }
 214
 215 /*
 216  * Release a shared spinlock.  We leave the shared bit set in the spinlock
 217  * as a cache and simply clear the spinlock pointer for the cpu.  This
 218  * fast-paths another shared lock later at the cost of an exclusive lock
 219  * having to check per-cpu spinlock pointers to determine when there are no
 220  * shared holders remaining.
 221  */
 222 static __inline void
 223 spin_unlock_rd_quick(globaldata_t gd, struct spinlock *mtx)
 224 {
 225         KKASSERT(gd->gd_spinlock_rd == mtx);
 226         gd->gd_spinlock_rd = NULL;
 227 }
 228
 229 static __inline void
 230 spin_unlock_rd(struct spinlock *mtx)
 231 {
 232         spin_unlock_rd_quick(mycpu, mtx);
 233 }
 234
 235 static __inline void
 236 spin_init(struct spinlock *mtx)
 237 {
 238         mtx->lock = 0;
 239 }
 240
 241 static __inline void
 242 spin_uninit(struct spinlock *mtx)
 243 {
 244         /* unused */
 245 }
 246
 247 #endif  /* _KERNEL */
 248 #endif  /* _SYS_SPINLOCK2_H_ */
 249