2 * Copyright (c) 2005 Jeffrey M. Hsu. All rights reserved.
4 * This code is derived from software contributed to The DragonFly Project
5 * by Jeffrey M. Hsu. and Matthew Dillon
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/kern/kern_spinlock.c,v 1.4 2006/05/21 20:23:25 dillon Exp $
35 #include <sys/param.h>
36 #include <sys/systm.h>
37 #include <sys/types.h>
38 #include <sys/kernel.h>
39 #include <sys/sysctl.h>
44 #include <machine/atomic.h>
45 #include <machine/cpufunc.h>
46 #include <machine/clock.h>
47 #include <sys/spinlock.h>
48 #include <sys/spinlock2.h>
50 #define BACKOFF_INITIAL 1
51 #define BACKOFF_LIMIT 256
56 static int spin_lock_test_mode;
59 static int64_t spinlocks_contested1;
60 SYSCTL_QUAD(_debug, OID_AUTO, spinlocks_contested1, CTLFLAG_RD, &spinlocks_contested1, 0, "");
61 static int64_t spinlocks_contested2;
62 SYSCTL_QUAD(_debug, OID_AUTO, spinlocks_contested2, CTLFLAG_RD, &spinlocks_contested2, 0, "");
64 struct exponential_backoff {
70 static int exponential_backoff(struct exponential_backoff *bo);
74 exponential_init(struct exponential_backoff *bo, struct spinlock *mtx)
76 bo->backoff = BACKOFF_INITIAL;
82 * We were either contested due to another exclusive lock holder,
83 * or due to the presence of shared locks. We have to undo the mess
84 * we created by returning the shared locks.
86 * If there was another exclusive lock holder only the exclusive bit
87 * in value will be the only bit set. We don't have to do anything since
88 * restoration does not involve any work.
90 * Otherwise we successfully obtained the exclusive bit. Attempt to
91 * clear the shared bits. If we are able to clear the shared bits
92 * we win. Otherwise we lose and we have to restore the shared bits
93 * we couldn't clear (and also clear our exclusive bit).
96 spin_trylock_wr_contested(struct spinlock *mtx, int value)
100 ++spinlocks_contested1;
101 if ((value & SPINLOCK_EXCLUSIVE) == 0) {
104 if (globaldata_find(bit)->gd_spinlocks_rd != 0) {
105 atomic_swap_int(&mtx->lock, value);
108 value &= ~(1 << bit);
116 * We were either contested due to another exclusive lock holder,
117 * or due to the presence of shared locks
119 * NOTE: If value indicates an exclusively held mutex, no shared bits
120 * would have been set and we can throw away value.
123 spin_lock_wr_contested(struct spinlock *mtx, int value)
125 struct exponential_backoff backoff;
126 globaldata_t gd = mycpu;
131 * Wait until we can gain exclusive access vs another exclusive
134 exponential_init(&backoff, mtx);
135 ++spinlocks_contested1;
137 while (value & SPINLOCK_EXCLUSIVE) {
138 value = atomic_swap_int(&mtx->lock, 0x80000000);
139 if (exponential_backoff(&backoff)) {
140 value &= ~SPINLOCK_EXCLUSIVE;
146 * Kill the cached shared bit for our own cpu. This is the most
147 * common case and there's no sense wasting cpu on it. Since
148 * spinlocks aren't recursive, we can't own a shared ref on the
149 * spinlock while trying to get an exclusive one.
151 * If multiple bits are set do not stall on any single cpu. Check
152 * all cpus that have the cache bit set, then loop and check again,
153 * until we've cleaned all the bits.
155 value &= ~gd->gd_cpumask;
157 while ((mask = value) != 0) {
160 if (globaldata_find(bit)->gd_spinlocks_rd == 0) {
161 value &= ~(1 << bit);
162 } else if (exponential_backoff(&backoff)) {
172 * The cache bit wasn't set for our cpu. Loop until we can set the bit.
173 * As with the spin_lock_rd() inline we need a memory fence after incrementing
174 * gd_spinlocks_rd to interlock against exclusive spinlocks waiting for
175 * that field to clear.
178 spin_lock_rd_contested(struct spinlock *mtx)
180 struct exponential_backoff backoff;
181 globaldata_t gd = mycpu;
182 int value = mtx->lock;
184 exponential_init(&backoff, mtx);
185 ++spinlocks_contested1;
187 while ((value & gd->gd_cpumask) == 0) {
188 if (value & SPINLOCK_EXCLUSIVE) {
189 --gd->gd_spinlocks_rd;
190 if (exponential_backoff(&backoff)) {
191 ++gd->gd_spinlocks_rd;
194 ++gd->gd_spinlocks_rd;
197 if (atomic_cmpset_int(&mtx->lock, value, value|gd->gd_cpumask))
205 * Handle exponential backoff and indefinite waits.
207 * If the system is handling a panic we hand the spinlock over to the caller
208 * after 1 second. After 10 seconds we attempt to print a debugger
209 * backtrace. We also run pending interrupts in order to allow a console
214 exponential_backoff(struct exponential_backoff *bo)
222 for (i = 0; i < bo->backoff; ++i)
224 if (bo->backoff < BACKOFF_LIMIT) {
232 ++spinlocks_contested2;
234 bo->base = sys_cputimer->count();
238 count = sys_cputimer->count();
239 if (count - bo->base > sys_cputimer->freq) {
240 printf("spin_lock: %p, indefinite wait!\n", bo->mtx);
244 if (spin_lock_test_mode) {
245 db_print_backtrace();
249 if (++bo->nsec == 11)
250 db_print_backtrace();
252 panic("spin_lock: %p, indefinite wait!\n", bo->mtx);
260 * If INVARIANTS is enabled various spinlock timing tests can be run
261 * by setting debug.spin_lock_test:
263 * 1 Test the indefinite wait code
264 * 2 Time the best-case exclusive lock overhead (spin_test_count)
265 * 3 Time the best-case shared lock overhead (spin_test_count)
270 static int spin_test_count = 10000000;
271 SYSCTL_INT(_debug, OID_AUTO, spin_test_count, CTLFLAG_RW, &spin_test_count, 0, "");
274 sysctl_spin_lock_test(SYSCTL_HANDLER_ARGS)
281 if ((error = suser(curthread)) != 0)
283 if ((error = SYSCTL_IN(req, &value, sizeof(value))) != 0)
287 * Indefinite wait test
291 spin_lock_wr(&mtx); /* force an indefinite wait */
292 spin_lock_test_mode = 1;
294 spin_unlock_wr(&mtx); /* Clean up the spinlock count */
295 spin_unlock_wr(&mtx);
296 spin_lock_test_mode = 0;
300 * Time best-case exclusive spinlocks
303 globaldata_t gd = mycpu;
306 for (i = spin_test_count; i > 0; --i) {
307 spin_lock_wr_quick(gd, &mtx);
308 spin_unlock_wr_quick(gd, &mtx);
313 * Time best-case shared spinlocks
316 globaldata_t gd = mycpu;
319 for (i = spin_test_count; i > 0; --i) {
320 spin_lock_rd_quick(gd, &mtx);
321 spin_unlock_rd_quick(gd, &mtx);
327 SYSCTL_PROC(_debug, KERN_PROC_ALL, spin_lock_test, CTLFLAG_RW|CTLTYPE_INT,
328 0, 0, sysctl_spin_lock_test, "I", "Test spinlock wait code");
330 #endif /* INVARIANTS */