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 #ifdef _KERNEL_VIRTUAL
73 struct spinlock pmap_spin
= SPINLOCK_INITIALIZER(pmap_spin
, "pmap_spin");
75 struct indefinite_info
{
84 #if !defined(KTR_SPIN_CONTENTION)
85 #define KTR_SPIN_CONTENTION KTR_ALL
87 #define SPIN_STRING "spin=%p type=%c"
88 #define SPIN_ARG_SIZE (sizeof(void *) + sizeof(int))
90 KTR_INFO_MASTER(spin
);
92 KTR_INFO(KTR_SPIN_CONTENTION
, spin
, beg
, 0, SPIN_STRING
, SPIN_ARG_SIZE
);
93 KTR_INFO(KTR_SPIN_CONTENTION
, spin
, end
, 1, SPIN_STRING
, SPIN_ARG_SIZE
);
96 #define logspin(name, spin, type) \
97 KTR_LOG(spin_ ## name, spin, type)
100 static int spin_lock_test_mode
;
103 #ifdef DEBUG_LOCKS_LATENCY
105 static long spinlocks_add_latency
;
106 SYSCTL_LONG(_debug
, OID_AUTO
, spinlocks_add_latency
, CTLFLAG_RW
,
107 &spinlocks_add_latency
, 0,
108 "Add spinlock latency");
114 * We need a fairly large pool to avoid contention on large SMP systems,
115 * particularly multi-chip systems.
117 /*#define SPINLOCK_NUM_POOL 8101*/
118 #define SPINLOCK_NUM_POOL 8192
119 #define SPINLOCK_NUM_POOL_MASK (SPINLOCK_NUM_POOL - 1)
121 static __cachealign
struct {
122 struct spinlock spin
;
123 char filler
[32 - sizeof(struct spinlock
)];
124 } pool_spinlocks
[SPINLOCK_NUM_POOL
];
126 static int spin_indefinite_check(struct spinlock
*spin
,
127 struct indefinite_info
*info
);
130 * We contested due to another exclusive lock holder. We lose.
132 * We have to unwind the attempt and may acquire the spinlock
133 * anyway while doing so.
136 spin_trylock_contested(struct spinlock
*spin
)
138 globaldata_t gd
= mycpu
;
141 * Handle degenerate case, else fail.
143 if (atomic_cmpset_int(&spin
->counta
, SPINLOCK_SHARED
|0, 1))
145 /*atomic_add_int(&spin->counta, -1);*/
147 --gd
->gd_curthread
->td_critcount
;
152 * The spin_lock() inline was unable to acquire the lock and calls this
153 * function with spin->counta already incremented.
155 * atomic_swap_int() is the absolute fastest spinlock instruction, at
156 * least on multi-socket systems. All instructions seem to be about
157 * the same on single-socket multi-core systems. However, atomic_swap_int()
158 * does not result in an even distribution of successful acquisitions.
160 * UNFORTUNATELY we cannot really use atomic_swap_int() when also implementing
161 * shared spin locks, so as we do a better job removing contention we've
162 * moved to atomic_cmpset_int() to be able handle multiple states.
164 * Another problem we have is that (at least on the 48-core opteron we test
165 * with) having all 48 cores contesting the same spin lock reduces
166 * performance to around 600,000 ops/sec, verses millions when fewer cores
167 * are going after the same lock.
169 * Backoff algorithms can create even worse starvation problems, and don't
170 * really improve performance when a lot of cores are contending.
172 * Our solution is to allow the data cache to lazy-update by reading it
173 * non-atomically and only attempting to acquire the lock if the lazy read
174 * looks good. This effectively limits cache bus bandwidth. A cpu_pause()
175 * (for intel/amd anyhow) is not strictly needed as cache bus resource use
176 * is governed by the lazy update.
178 * WARNING!!!! Performance matters here, by a huge margin.
180 * 48-core test with pre-read / -j 48 no-modules kernel compile
181 * with fanned-out inactive and active queues came in at 55 seconds.
183 * 48-core test with pre-read / -j 48 no-modules kernel compile
184 * came in at 75 seconds. Without pre-read it came in at 170 seconds.
186 * 4-core test with pre-read / -j 48 no-modules kernel compile
187 * came in at 83 seconds. Without pre-read it came in at 83 seconds
188 * as well (no difference).
191 _spin_lock_contested(struct spinlock
*spin
, const char *ident
, int value
)
193 struct indefinite_info info
= { 0, 0, ident
};
197 * Handle degenerate case.
199 if (value
== SPINLOCK_SHARED
) {
200 if (atomic_cmpset_int(&spin
->counta
, SPINLOCK_SHARED
|0, 1))
205 * Transfer our count to the high bits, then loop until we can
206 * acquire the low counter (== 1). No new shared lock can be
207 * acquired while we hold the EXCLWAIT bits.
209 * Force any existing shared locks to exclusive. The shared unlock
210 * understands that this may occur.
212 atomic_add_int(&spin
->counta
, SPINLOCK_EXCLWAIT
- 1);
213 if (value
& SPINLOCK_SHARED
)
214 atomic_clear_int(&spin
->counta
, SPINLOCK_SHARED
);
216 #ifdef DEBUG_LOCKS_LATENCY
218 for (j
= spinlocks_add_latency
; j
> 0; --j
)
223 /*logspin(beg, spin, 'w');*/
226 * If the low bits are zero, try to acquire the exclusive lock
227 * by transfering our high bit counter to the low bits.
229 * NOTE: Reading spin->counta prior to the swap is extremely
230 * important on multi-chip/many-core boxes. On 48-core
231 * this one change improves fully concurrent all-cores
232 * compiles by 100% or better.
234 * I can't emphasize enough how important the pre-read
235 * is in preventing hw cache bus armageddon on
236 * multi-chip systems. And on single-chip/multi-core
237 * systems it just doesn't hurt.
239 uint32_t ovalue
= spin
->counta
;
241 if ((ovalue
& (SPINLOCK_EXCLWAIT
- 1)) == 0 &&
242 atomic_cmpset_int(&spin
->counta
, ovalue
,
243 (ovalue
- SPINLOCK_EXCLWAIT
) | 1)) {
246 if ((++i
& 0x7F) == 0x7F) {
247 mycpu
->gd_cnt
.v_lock_name
[0] = 'X';
248 strncpy(mycpu
->gd_cnt
.v_lock_name
+ 1,
250 sizeof(mycpu
->gd_cnt
.v_lock_name
) - 2);
251 ++mycpu
->gd_cnt
.v_lock_colls
;
252 if (spin_indefinite_check(spin
, &info
))
255 #ifdef _KERNEL_VIRTUAL
259 /*logspin(end, spin, 'w');*/
263 * Shared spinlock attempt was contested.
265 * The caller has not modified counta.
268 _spin_lock_shared_contested(struct spinlock
*spin
, const char *ident
, int value
)
270 struct indefinite_info info
= { 0, 0, ident
};
273 #ifdef DEBUG_LOCKS_LATENCY
275 for (j
= spinlocks_add_latency
; j
> 0; --j
)
280 /*logspin(beg, spin, 'w');*/
283 * Loop until we can acquire the shared spinlock. Note that
284 * the low bits can be zero while the high EXCLWAIT bits are
285 * non-zero. In this situation exclusive requesters have
286 * priority (otherwise shared users on multiple cpus can hog
289 * NOTE: Reading spin->counta prior to the swap is extremely
290 * important on multi-chip/many-core boxes. On 48-core
291 * this one change improves fully concurrent all-cores
292 * compiles by 100% or better.
294 * I can't emphasize enough how important the pre-read
295 * is in preventing hw cache bus armageddon on
296 * multi-chip systems. And on single-chip/multi-core
297 * systems it just doesn't hurt.
299 uint32_t ovalue
= spin
->counta
;
303 if (atomic_cmpset_int(&spin
->counta
, 0,
304 SPINLOCK_SHARED
| 1))
306 } else if (ovalue
& SPINLOCK_SHARED
) {
307 if (atomic_cmpset_int(&spin
->counta
, ovalue
,
311 if ((++i
& 0x7F) == 0x7F) {
312 mycpu
->gd_cnt
.v_lock_name
[0] = 'S';
313 strncpy(mycpu
->gd_cnt
.v_lock_name
+ 1,
315 sizeof(mycpu
->gd_cnt
.v_lock_name
) - 2);
316 ++mycpu
->gd_cnt
.v_lock_colls
;
317 if (spin_indefinite_check(spin
, &info
))
320 #ifdef _KERNEL_VIRTUAL
324 /*logspin(end, spin, 'w');*/
328 * Pool functions (SHARED SPINLOCKS NOT SUPPORTED)
331 _spin_pool_hash(void *ptr
)
335 i
= ((int)(uintptr_t) ptr
>> 5) ^ ((int)(uintptr_t)ptr
>> 12);
336 i
&= SPINLOCK_NUM_POOL_MASK
;
341 _spin_pool_lock(void *chan
, const char *ident
)
345 sp
= &pool_spinlocks
[_spin_pool_hash(chan
)].spin
;
346 _spin_lock(sp
, ident
);
350 _spin_pool_unlock(void *chan
)
354 sp
= &pool_spinlocks
[_spin_pool_hash(chan
)].spin
;
361 spin_indefinite_check(struct spinlock
*spin
, struct indefinite_info
*info
)
365 cpu_spinlock_contested();
367 count
= sys_cputimer
->count();
368 if (info
->secs
== 0) {
371 } else if (count
- info
->base
> sys_cputimer
->freq
) {
372 kprintf("spin_lock: %s(%p), indefinite wait (%d secs)!\n",
373 info
->ident
, spin
, info
->secs
);
378 #if defined(INVARIANTS)
379 if (spin_lock_test_mode
) {
384 #if defined(INVARIANTS)
385 if (info
->secs
== 11)
388 if (info
->secs
== 60)
389 panic("spin_lock: %s(%p), indefinite wait!",
396 * If INVARIANTS is enabled various spinlock timing tests can be run
397 * by setting debug.spin_lock_test:
399 * 1 Test the indefinite wait code
400 * 2 Time the best-case exclusive lock overhead (spin_test_count)
401 * 3 Time the best-case shared lock overhead (spin_test_count)
406 static int spin_test_count
= 10000000;
407 SYSCTL_INT(_debug
, OID_AUTO
, spin_test_count
, CTLFLAG_RW
, &spin_test_count
, 0,
408 "Number of iterations to use for spinlock wait code test");
411 sysctl_spin_lock_test(SYSCTL_HANDLER_ARGS
)
413 struct spinlock spin
;
418 if ((error
= priv_check(curthread
, PRIV_ROOT
)) != 0)
420 if ((error
= SYSCTL_IN(req
, &value
, sizeof(value
))) != 0)
424 * Indefinite wait test
427 spin_init(&spin
, "sysctllock");
428 spin_lock(&spin
); /* force an indefinite wait */
429 spin_lock_test_mode
= 1;
431 spin_unlock(&spin
); /* Clean up the spinlock count */
433 spin_lock_test_mode
= 0;
437 * Time best-case exclusive spinlocks
440 globaldata_t gd
= mycpu
;
442 spin_init(&spin
, "sysctllocktest");
443 for (i
= spin_test_count
; i
> 0; --i
) {
444 _spin_lock_quick(gd
, &spin
, "test");
445 spin_unlock_quick(gd
, &spin
);
452 SYSCTL_PROC(_debug
, KERN_PROC_ALL
, spin_lock_test
, CTLFLAG_RW
|CTLTYPE_INT
,
453 0, 0, sysctl_spin_lock_test
, "I", "Test spinlock wait code");
455 #endif /* INVARIANTS */