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
34 * The implementation is designed to avoid looping when compatible operations
37 * To acquire a spinlock we first increment counta. Then we check if counta
38 * meets our requirements. For an exclusive spinlock it must be 1, of a
39 * shared spinlock it must either be 1 or the SHARED_SPINLOCK bit must be set.
41 * Shared spinlock failure case: Decrement the count, loop until we can
42 * transition from 0 to SHARED_SPINLOCK|1, or until we find SHARED_SPINLOCK
43 * is set and increment the count.
45 * Exclusive spinlock failure case: While maintaining the count, clear the
46 * SHARED_SPINLOCK flag unconditionally. Then use an atomic add to transfer
47 * the count from the low bits to the high bits of counta. Then loop until
48 * all low bits are 0. Once the low bits drop to 0 we can transfer the
49 * count back with an atomic_cmpset_int(), atomically, and return.
51 #include <sys/param.h>
52 #include <sys/systm.h>
53 #include <sys/types.h>
54 #include <sys/kernel.h>
55 #include <sys/sysctl.h>
60 #include <machine/atomic.h>
61 #include <machine/cpu.h>
62 #include <machine/cpufunc.h>
63 #include <machine/specialreg.h>
64 #include <machine/clock.h>
65 #include <sys/spinlock.h>
66 #include <sys/spinlock2.h>
69 struct spinlock pmap_spin = SPINLOCK_INITIALIZER(pmap_spin);
73 struct indefinite_info {
81 #if !defined(KTR_SPIN_CONTENTION)
82 #define KTR_SPIN_CONTENTION KTR_ALL
84 #define SPIN_STRING "spin=%p type=%c"
85 #define SPIN_ARG_SIZE (sizeof(void *) + sizeof(int))
87 KTR_INFO_MASTER(spin);
89 KTR_INFO(KTR_SPIN_CONTENTION, spin, beg, 0, SPIN_STRING, SPIN_ARG_SIZE);
90 KTR_INFO(KTR_SPIN_CONTENTION, spin, end, 1, SPIN_STRING, SPIN_ARG_SIZE);
93 #define logspin(name, spin, type) \
94 KTR_LOG(spin_ ## name, spin, type)
97 static int spin_lock_test_mode;
100 static int64_t spinlocks_contested1;
101 SYSCTL_QUAD(_debug, OID_AUTO, spinlocks_contested1, CTLFLAG_RD,
102 &spinlocks_contested1, 0,
103 "Spinlock contention count due to collisions with exclusive lock holders");
105 static int64_t spinlocks_contested2;
106 SYSCTL_QUAD(_debug, OID_AUTO, spinlocks_contested2, CTLFLAG_RD,
107 &spinlocks_contested2, 0,
108 "Serious spinlock contention count");
110 #ifdef DEBUG_LOCKS_LATENCY
112 static long spinlocks_add_latency;
113 SYSCTL_LONG(_debug, OID_AUTO, spinlocks_add_latency, CTLFLAG_RW,
114 &spinlocks_add_latency, 0,
115 "Add spinlock latency");
121 * We need a fairly large pool to avoid contention on large SMP systems,
122 * particularly multi-chip systems.
124 /*#define SPINLOCK_NUM_POOL 8101*/
125 #define SPINLOCK_NUM_POOL 8192
126 #define SPINLOCK_NUM_POOL_MASK (SPINLOCK_NUM_POOL - 1)
128 static __cachealign struct {
129 struct spinlock spin;
130 char filler[32 - sizeof(struct spinlock)];
131 } pool_spinlocks[SPINLOCK_NUM_POOL];
133 static int spin_indefinite_check(struct spinlock *spin,
134 struct indefinite_info *info);
137 * We contested due to another exclusive lock holder. We lose.
139 * We have to unwind the attempt and may acquire the spinlock
140 * anyway while doing so. countb was incremented on our behalf.
143 spin_trylock_contested(struct spinlock *spin)
145 globaldata_t gd = mycpu;
147 /*++spinlocks_contested1;*/
148 /*atomic_add_int(&spin->counta, -1);*/
150 --gd->gd_curthread->td_critcount;
155 * The spin_lock() inline was unable to acquire the lock.
157 * atomic_swap_int() is the absolute fastest spinlock instruction, at
158 * least on multi-socket systems. All instructions seem to be about
159 * the same on single-socket multi-core systems. However, atomic_swap_int()
160 * does not result in an even distribution of successful acquisitions.
162 * UNFORTUNATELY we cannot really use atomic_swap_int() when also implementing
163 * shared spin locks, so as we do a better job removing contention we've
164 * moved to atomic_cmpset_int() to be able handle multiple states.
166 * Another problem we have is that (at least on the 48-core opteron we test
167 * with) having all 48 cores contesting the same spin lock reduces
168 * performance to around 600,000 ops/sec, verses millions when fewer cores
169 * are going after the same lock.
171 * Backoff algorithms can create even worse starvation problems, and don't
172 * really improve performance when a lot of cores are contending.
174 * Our solution is to allow the data cache to lazy-update by reading it
175 * non-atomically and only attempting to acquire the lock if the lazy read
176 * looks good. This effectively limits cache bus bandwidth. A cpu_pause()
177 * (for intel/amd anyhow) is not strictly needed as cache bus resource use
178 * is governed by the lazy update.
180 * WARNING!!!! Performance matters here, by a huge margin.
182 * 48-core test with pre-read / -j 48 no-modules kernel compile
183 * with fanned-out inactive and active queues came in at 55 seconds.
185 * 48-core test with pre-read / -j 48 no-modules kernel compile
186 * came in at 75 seconds. Without pre-read it came in at 170 seconds.
188 * 4-core test with pre-read / -j 48 no-modules kernel compile
189 * came in at 83 seconds. Without pre-read it came in at 83 seconds
190 * as well (no difference).
193 spin_lock_contested(struct spinlock *spin)
195 struct indefinite_info info = { 0, 0 };
199 * Force any existing shared locks to exclusive so no new shared
200 * locks can occur. Transfer our count to the high bits, then
201 * loop until we can acquire the low counter (== 1).
203 atomic_clear_int(&spin->counta, SPINLOCK_SHARED);
204 atomic_add_int(&spin->counta, SPINLOCK_EXCLWAIT - 1);
206 #ifdef DEBUG_LOCKS_LATENCY
208 for (j = spinlocks_add_latency; j > 0; --j)
211 if (spin_lock_test_mode > 10 &&
212 spin->countb > spin_lock_test_mode &&
213 (spin_lock_test_mode & 0xFF) == mycpu->gd_cpuid) {
221 /*logspin(beg, spin, 'w');*/
224 * If the low bits are zero, try to acquire the exclusive lock
225 * by transfering our high bit counter to the low bits.
227 * NOTE: Reading spin->counta prior to the swap is extremely
228 * important on multi-chip/many-core boxes. On 48-core
229 * this one change improves fully concurrent all-cores
230 * compiles by 100% or better.
232 * I can't emphasize enough how important the pre-read
233 * is in preventing hw cache bus armageddon on
234 * multi-chip systems. And on single-chip/multi-core
235 * systems it just doesn't hurt.
237 uint32_t ovalue = spin->counta;
239 if ((ovalue & (SPINLOCK_EXCLWAIT - 1)) == 0 &&
240 atomic_cmpset_int(&spin->counta, ovalue,
241 (ovalue - SPINLOCK_EXCLWAIT) | 1)) {
244 if ((++i & 0x7F) == 0x7F) {
246 if (spin_indefinite_check(spin, &info))
250 /*logspin(end, spin, 'w');*/
257 spin_lock_shared_contested(struct spinlock *spin)
259 struct indefinite_info info = { 0, 0 };
262 atomic_add_int(&spin->counta, -1);
263 #ifdef DEBUG_LOCKS_LATENCY
265 for (j = spinlocks_add_latency; j > 0; --j)
268 if (spin_lock_test_mode > 10 &&
269 spin->countb > spin_lock_test_mode &&
270 (spin_lock_test_mode & 0xFF) == mycpu->gd_cpuid) {
278 /*logspin(beg, spin, 'w');*/
281 * NOTE: Reading spin->counta prior to the swap is extremely
282 * important on multi-chip/many-core boxes. On 48-core
283 * this one change improves fully concurrent all-cores
284 * compiles by 100% or better.
286 * I can't emphasize enough how important the pre-read
287 * is in preventing hw cache bus armageddon on
288 * multi-chip systems. And on single-chip/multi-core
289 * systems it just doesn't hurt.
291 uint32_t ovalue = spin->counta;
295 if (atomic_cmpset_int(&spin->counta, 0,
296 SPINLOCK_SHARED | 1))
298 } else if (ovalue & SPINLOCK_SHARED) {
299 if (atomic_cmpset_int(&spin->counta, ovalue,
303 if ((++i & 0x7F) == 0x7F) {
305 if (spin_indefinite_check(spin, &info))
309 /*logspin(end, spin, 'w');*/
313 * Pool functions (SHARED SPINLOCKS NOT SUPPORTED)
316 _spin_pool_hash(void *ptr)
320 i = ((int)(uintptr_t) ptr >> 5) ^ ((int)(uintptr_t)ptr >> 12);
321 i &= SPINLOCK_NUM_POOL_MASK;
326 _spin_pool_lock(void *chan)
330 sp = &pool_spinlocks[_spin_pool_hash(chan)].spin;
335 _spin_pool_unlock(void *chan)
339 sp = &pool_spinlocks[_spin_pool_hash(chan)].spin;
346 spin_indefinite_check(struct spinlock *spin, struct indefinite_info *info)
350 cpu_spinlock_contested();
352 count = sys_cputimer->count();
353 if (info->secs == 0) {
356 } else if (count - info->base > sys_cputimer->freq) {
357 kprintf("spin_lock: %p, indefinite wait (%d secs)!\n",
363 #if defined(INVARIANTS)
364 if (spin_lock_test_mode) {
369 #if defined(INVARIANTS)
370 if (info->secs == 11)
373 if (info->secs == 60)
374 panic("spin_lock: %p, indefinite wait!", spin);
380 * If INVARIANTS is enabled various spinlock timing tests can be run
381 * by setting debug.spin_lock_test:
383 * 1 Test the indefinite wait code
384 * 2 Time the best-case exclusive lock overhead (spin_test_count)
385 * 3 Time the best-case shared lock overhead (spin_test_count)
390 static int spin_test_count = 10000000;
391 SYSCTL_INT(_debug, OID_AUTO, spin_test_count, CTLFLAG_RW, &spin_test_count, 0,
392 "Number of iterations to use for spinlock wait code test");
395 sysctl_spin_lock_test(SYSCTL_HANDLER_ARGS)
397 struct spinlock spin;
402 if ((error = priv_check(curthread, PRIV_ROOT)) != 0)
404 if ((error = SYSCTL_IN(req, &value, sizeof(value))) != 0)
408 * Indefinite wait test
412 spin_lock(&spin); /* force an indefinite wait */
413 spin_lock_test_mode = 1;
415 spin_unlock(&spin); /* Clean up the spinlock count */
417 spin_lock_test_mode = 0;
421 * Time best-case exclusive spinlocks
424 globaldata_t gd = mycpu;
427 for (i = spin_test_count; i > 0; --i) {
428 spin_lock_quick(gd, &spin);
429 spin_unlock_quick(gd, &spin);
436 SYSCTL_PROC(_debug, KERN_PROC_ALL, spin_lock_test, CTLFLAG_RW|CTLTYPE_INT,
437 0, 0, sysctl_spin_lock_test, "I", "Test spinlock wait code");
439 #endif /* INVARIANTS */