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.
  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(gd, mtx, value));
  92         return (TRUE);
  93 }
  94
  95 #else
  96
  97 static __inline boolean_t
  98 spin_trylock_wr(struct spinlock *mtx)
  99 {
 100         globaldata_t gd = mycpu;
 101
 102         ++gd->gd_spinlocks_wr;
 103         return (TRUE);
 104 }
 105
 106 #endif
 107
 108 /*
 109  * Obtain an exclusive spinlock and return.  Shortcut the case where the only
 110  * cached read lock was from our own cpu (it can just be cleared).
 111  */
 112 static __inline void
 113 spin_lock_wr_quick(globaldata_t gd, struct spinlock *mtx)
 114 {
 115 #ifdef SMP
 116         int value;
 117 #endif
 118
 119         ++gd->gd_spinlocks_wr;
 120 #ifdef SMP
 121         if ((value = atomic_swap_int(&mtx->lock, SPINLOCK_EXCLUSIVE)) != 0) {
 122                 value &= ~gd->gd_cpumask;
 123                 if (value)
 124                         spin_lock_wr_contested(mtx, value);
 125         }
 126 #endif
 127 }
 128
 129 static __inline void
 130 spin_lock_wr(struct spinlock *mtx)
 131 {
 132         spin_lock_wr_quick(mycpu, mtx);
 133 }
 134
 135 #if 0
 136
 137 /*
 138  * Upgrade a shared spinlock to exclusive.  Return TRUE if we were
 139  * able to upgrade without another exclusive holder getting in before
 140  * us, FALSE otherwise.
 141  */
 142 static __inline int
 143 spin_lock_upgrade(struct spinlock *mtx)
 144 {
 145         globaldata_t gd = mycpu;
 146 #ifdef SMP
 147         int value;
 148 #endif
 149
 150         ++gd->gd_spinlocks_wr;
 151 #ifdef SMP
 152         value = atomic_swap_int(&mtx->lock, SPINLOCK_EXCLUSIVE);
 153         cpu_sfence();
 154 #endif
 155         gd->gd_spinlock_rd = NULL;
 156 #ifdef SMP
 157         value &= ~gd->gd_cpumask;
 158         if (value) {
 159                 spin_lock_wr_contested(mtx, value);
 160                 if (value & SPINLOCK_EXCLUSIVE)
 161                         return (FALSE);
 162                 XXX regain original shared lock?
 163         }
 164         return (TRUE);
 165 #endif
 166 }
 167
 168 #endif
 169
 170 /*
 171  * Obtain a shared spinlock and return.  This is a critical code path.
 172  *
 173  * The vast majority of the overhead is in the cpu_mfence() (5ns vs 1ns for
 174  * the entire rest of the procedure).  Unfortunately we have to ensure that
 175  * spinlock pointer is written out before we check the cpumask to interlock
 176  * against an exclusive spinlock that clears the cpumask and then checks
 177  * the spinlock pointer.
 178  *
 179  * But what is EXTREMELY important here is that we do not have to perform
 180  * a locked bus cycle on the spinlock itself if the shared bit for our cpu
 181  * is already found to be set.  We only need the mfence, and the mfence is
 182  * local to the cpu and never conflicts with other cpu's.
 183  *
 184  * This means that multiple parallel shared acessors (e.g. filedescriptor
 185  * table lookups, namecache lookups) run at full speed and incur NO cache
 186  * contention at all.  It is the difference between 10ns and 40-100ns.
 187  */
 188 static __inline void
 189 spin_lock_rd_quick(globaldata_t gd, struct spinlock *mtx)
 190 {
 191         gd->gd_spinlock_rd = mtx;
 192 #ifdef SMP
 193         cpu_mfence();
 194         if ((mtx->lock & gd->gd_cpumask) == 0)
 195                 spin_lock_rd_contested(mtx);
 196 #endif
 197 }
 198
 199 static __inline void
 200 spin_lock_rd(struct spinlock *mtx)
 201 {
 202         spin_lock_rd_quick(mycpu,mtx);
 203 }
 204
 205 /*
 206  * Release an exclusive spinlock.  We can just do this passively, only
 207  * ensuring that our spinlock count is left intact until the mutex is
 208  * cleared.
 209  */
 210 static __inline void
 211 spin_unlock_wr_quick(globaldata_t gd, struct spinlock *mtx)
 212 {
 213 #ifdef SMP
 214         mtx->lock = 0;
 215 #endif
 216         KKASSERT(gd->gd_spinlocks_wr > 0);
 217         --gd->gd_spinlocks_wr;
 218 }
 219
 220 static __inline void
 221 spin_unlock_wr(struct spinlock *mtx)
 222 {
 223         spin_unlock_wr_quick(mycpu, mtx);
 224 }
 225
 226 /*
 227  * Release a shared spinlock.  We leave the shared bit set in the spinlock
 228  * as a cache and simply clear the spinlock pointer for the cpu.  This
 229  * fast-paths another shared lock later at the cost of an exclusive lock
 230  * having to check per-cpu spinlock pointers to determine when there are no
 231  * shared holders remaining.
 232  */
 233 static __inline void
 234 spin_unlock_rd_quick(globaldata_t gd, struct spinlock *mtx)
 235 {
 236         KKASSERT(gd->gd_spinlock_rd == mtx);
 237         gd->gd_spinlock_rd = NULL;
 238 }
 239
 240 static __inline void
 241 spin_unlock_rd(struct spinlock *mtx)
 242 {
 243         spin_unlock_rd_quick(mycpu, mtx);
 244 }
 245
 246 static __inline void
 247 spin_init(struct spinlock *mtx)
 248 {
 249         mtx->lock = 0;
 250 }
 251
 252 static __inline void
 253 spin_uninit(struct spinlock *mtx)
 254 {
 255         /* unused */
 256 }
 257
 258 #endif  /* _KERNEL */
 259 #endif  /* _SYS_SPINLOCK2_H_ */
 260