2 * Copyright (c) 2005 Jeffrey M. Hsu. All rights reserved.
4 * This code is derived from software contributed to The DragonFly Project
7 * Redistribution and use in source and binary forms, with or without
8 * modification, are permitted provided that the following conditions
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.
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
32 * $DragonFly: src/sys/sys/spinlock2.h,v 1.12 2008/06/04 04:34:54 nth Exp $
35 #ifndef _SYS_SPINLOCK2_H_
36 #define _SYS_SPINLOCK2_H_
40 #error "This file should not be included by userland programs."
45 #include <sys/systm.h>
47 #ifndef _SYS_THREAD2_H_
48 #include <sys/thread2.h>
50 #ifndef _SYS_GLOBALDATA_H_
51 #include <sys/globaldata.h>
53 #ifndef _MACHINE_ATOMIC_H_
54 #include <machine/atomic.h>
56 #ifndef _MACHINE_CPUFUNC_H_
57 #include <machine/cpufunc.h>
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
70 extern int spin_trylock_wr_contested(globaldata_t gd, struct spinlock *mtx,
72 extern void spin_lock_wr_contested(struct spinlock *mtx, int value);
73 extern void spin_lock_rd_contested(struct spinlock *mtx);
80 * Attempt to obtain an exclusive spinlock. Returns FALSE on failure,
83 static __inline boolean_t
84 spin_trylock_wr(struct spinlock *mtx)
86 globaldata_t gd = mycpu;
89 ++gd->gd_curthread->td_critcount;
91 ++gd->gd_spinlocks_wr;
92 if ((value = atomic_swap_int(&mtx->lock, SPINLOCK_EXCLUSIVE)) != 0)
93 return (spin_trylock_wr_contested(gd, mtx, value));
99 static __inline boolean_t
100 spin_trylock_wr(struct spinlock *mtx)
102 globaldata_t gd = mycpu;
104 ++gd->gd_curthread->td_critcount;
106 ++gd->gd_spinlocks_wr;
113 * Obtain an exclusive spinlock and return. Shortcut the case where the only
114 * cached read lock was from our own cpu (it can just be cleared).
117 spin_lock_wr_quick(globaldata_t gd, struct spinlock *mtx)
123 ++gd->gd_curthread->td_critcount;
125 ++gd->gd_spinlocks_wr;
127 if ((value = atomic_swap_int(&mtx->lock, SPINLOCK_EXCLUSIVE)) != 0) {
128 value &= ~gd->gd_cpumask;
130 spin_lock_wr_contested(mtx, value);
136 spin_lock_wr(struct spinlock *mtx)
138 spin_lock_wr_quick(mycpu, mtx);
142 * Obtain a shared spinlock and return. This is a critical code path.
144 * The vast majority of the overhead is in the cpu_mfence() (5ns vs 1ns for
145 * the entire rest of the procedure). Unfortunately we have to ensure that
146 * spinlock pointer is written out before we check the cpumask to interlock
147 * against an exclusive spinlock that clears the cpumask and then checks
148 * the spinlock pointer.
150 * But what is EXTREMELY important here is that we do not have to perform
151 * a locked bus cycle on the spinlock itself if the shared bit for our cpu
152 * is already found to be set. We only need the mfence, and the mfence is
153 * local to the cpu and never conflicts with other cpu's.
155 * This means that multiple parallel shared acessors (e.g. filedescriptor
156 * table lookups, namecache lookups) run at full speed and incur NO cache
157 * contention at all. It is the difference between 10ns and 40-100ns.
160 spin_lock_rd_quick(globaldata_t gd, struct spinlock *mtx)
162 ++gd->gd_curthread->td_critcount;
164 gd->gd_spinlock_rd = mtx;
167 if ((mtx->lock & gd->gd_cpumask) == 0)
168 spin_lock_rd_contested(mtx);
173 spin_lock_rd(struct spinlock *mtx)
175 spin_lock_rd_quick(mycpu,mtx);
179 * Release an exclusive spinlock. We can just do this passively, only
180 * ensuring that our spinlock count is left intact until the mutex is
184 spin_unlock_wr_quick(globaldata_t gd, struct spinlock *mtx)
189 KKASSERT(gd->gd_spinlocks_wr > 0);
190 --gd->gd_spinlocks_wr;
192 --gd->gd_curthread->td_critcount;
196 spin_unlock_wr(struct spinlock *mtx)
198 spin_unlock_wr_quick(mycpu, mtx);
202 * Release a shared spinlock. We leave the shared bit set in the spinlock
203 * as a cache and simply clear the spinlock pointer for the cpu. This
204 * fast-paths another shared lock later at the cost of an exclusive lock
205 * having to check per-cpu spinlock pointers to determine when there are no
206 * shared holders remaining.
209 spin_unlock_rd_quick(globaldata_t gd, struct spinlock *mtx)
211 KKASSERT(gd->gd_spinlock_rd == mtx);
212 gd->gd_spinlock_rd = NULL;
214 --gd->gd_curthread->td_critcount;
218 spin_unlock_rd(struct spinlock *mtx)
220 spin_unlock_rd_quick(mycpu, mtx);
224 spin_init(struct spinlock *mtx)
230 spin_uninit(struct spinlock *mtx)
236 #endif /* _SYS_SPINLOCK2_H_ */