xref: /netbsd-src/sys/kern/kern_lock.c (revision 267197ec1eebfcb9810ea27a89625b6ddf68e3e7)
1 /*	$NetBSD: kern_lock.c,v 1.134 2008/01/30 14:54:26 ad Exp $	*/
2 
3 /*-
4  * Copyright (c) 2002, 2006, 2007, 2008 The NetBSD Foundation, Inc.
5  * All rights reserved.
6  *
7  * This code is derived from software contributed to The NetBSD Foundation
8  * by Jason R. Thorpe of the Numerical Aerospace Simulation Facility,
9  * NASA Ames Research Center, and by Andrew Doran.
10  *
11  * Redistribution and use in source and binary forms, with or without
12  * modification, are permitted provided that the following conditions
13  * are met:
14  * 1. Redistributions of source code must retain the above copyright
15  *    notice, this list of conditions and the following disclaimer.
16  * 2. Redistributions in binary form must reproduce the above copyright
17  *    notice, this list of conditions and the following disclaimer in the
18  *    documentation and/or other materials provided with the distribution.
19  * 3. All advertising materials mentioning features or use of this software
20  *    must display the following acknowledgement:
21  *	This product includes software developed by the NetBSD
22  *	Foundation, Inc. and its contributors.
23  * 4. Neither the name of The NetBSD Foundation nor the names of its
24  *    contributors may be used to endorse or promote products derived
25  *    from this software without specific prior written permission.
26  *
27  * THIS SOFTWARE IS PROVIDED BY THE NETBSD FOUNDATION, INC. AND CONTRIBUTORS
28  * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED
29  * TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
30  * PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL THE FOUNDATION OR CONTRIBUTORS
31  * BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
32  * CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
33  * SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
34  * INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
35  * CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
36  * ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
37  * POSSIBILITY OF SUCH DAMAGE.
38  */
39 
40 #include <sys/cdefs.h>
41 __KERNEL_RCSID(0, "$NetBSD: kern_lock.c,v 1.134 2008/01/30 14:54:26 ad Exp $");
42 
43 #include "opt_multiprocessor.h"
44 
45 #include <sys/param.h>
46 #include <sys/proc.h>
47 #include <sys/lock.h>
48 #include <sys/systm.h>
49 #include <sys/kernel.h>
50 #include <sys/lockdebug.h>
51 #include <sys/cpu.h>
52 #include <sys/syslog.h>
53 #include <sys/atomic.h>
54 
55 #include <machine/stdarg.h>
56 #include <machine/lock.h>
57 
58 #include <dev/lockstat.h>
59 
60 #define	RETURN_ADDRESS	(uintptr_t)__builtin_return_address(0)
61 
62 bool	kernel_lock_dodebug;
63 
64 __cpu_simple_lock_t kernel_lock[CACHE_LINE_SIZE / sizeof(__cpu_simple_lock_t)]
65     __aligned(CACHE_LINE_SIZE);
66 
67 #if defined(LOCKDEBUG)
68 void
69 assert_sleepable(struct simplelock *interlock, const char *msg)
70 {
71 
72 	if (panicstr != NULL)
73 		return;
74 	LOCKDEBUG_BARRIER(kernel_lock, 1);
75 	if (CURCPU_IDLE_P() && !cold) {
76 		panic("assert_sleepable: idle");
77 	}
78 }
79 #endif
80 
81 /*
82  * rump doesn't need the kernel lock so force it out.  We cannot
83  * currently easily include it for compilation because of
84  * a) SPINLOCK_* b) membar_producer().  They are defined in different
85  * places / way for each arch, so just simply do not bother to
86  * fight a lot for no gain (i.e. pain but still no gain).
87  */
88 #ifndef _RUMPKERNEL
89 /*
90  * Functions for manipulating the kernel_lock.  We put them here
91  * so that they show up in profiles.
92  */
93 
94 #define	_KERNEL_LOCK_ABORT(msg)						\
95     LOCKDEBUG_ABORT(kernel_lock, &_kernel_lock_ops, __func__, msg)
96 
97 #ifdef LOCKDEBUG
98 #define	_KERNEL_LOCK_ASSERT(cond)					\
99 do {									\
100 	if (!(cond))							\
101 		_KERNEL_LOCK_ABORT("assertion failed: " #cond);		\
102 } while (/* CONSTCOND */ 0)
103 #else
104 #define	_KERNEL_LOCK_ASSERT(cond)	/* nothing */
105 #endif
106 
107 void	_kernel_lock_dump(volatile void *);
108 
109 lockops_t _kernel_lock_ops = {
110 	"Kernel lock",
111 	0,
112 	_kernel_lock_dump
113 };
114 
115 /*
116  * Initialize the kernel lock.
117  */
118 void
119 kernel_lock_init(void)
120 {
121 
122 	KASSERT(CACHE_LINE_SIZE >= sizeof(__cpu_simple_lock_t));
123 	__cpu_simple_lock_init(kernel_lock);
124 	kernel_lock_dodebug = LOCKDEBUG_ALLOC(kernel_lock, &_kernel_lock_ops,
125 	    RETURN_ADDRESS);
126 }
127 
128 /*
129  * Print debugging information about the kernel lock.
130  */
131 void
132 _kernel_lock_dump(volatile void *junk)
133 {
134 	struct cpu_info *ci = curcpu();
135 
136 	(void)junk;
137 
138 	printf_nolog("curcpu holds : %18d wanted by: %#018lx\n",
139 	    ci->ci_biglock_count, (long)ci->ci_biglock_wanted);
140 }
141 
142 /*
143  * Acquire 'nlocks' holds on the kernel lock.  If 'l' is non-null, the
144  * acquisition is from process context.
145  */
146 void
147 _kernel_lock(int nlocks, struct lwp *l)
148 {
149 	struct cpu_info *ci = curcpu();
150 	LOCKSTAT_TIMER(spintime);
151 	LOCKSTAT_FLAG(lsflag);
152 	struct lwp *owant;
153 	u_int spins;
154 	int s;
155 
156 	if (nlocks == 0)
157 		return;
158 	_KERNEL_LOCK_ASSERT(nlocks > 0);
159 
160 	l = curlwp;
161 
162 	if (ci->ci_biglock_count != 0) {
163 		_KERNEL_LOCK_ASSERT(__SIMPLELOCK_LOCKED_P(kernel_lock));
164 		ci->ci_biglock_count += nlocks;
165 		l->l_blcnt += nlocks;
166 		return;
167 	}
168 
169 	_KERNEL_LOCK_ASSERT(l->l_blcnt == 0);
170 	LOCKDEBUG_WANTLOCK(kernel_lock_dodebug, kernel_lock, RETURN_ADDRESS,
171 	    0);
172 
173 	s = splvm();
174 	if (__cpu_simple_lock_try(kernel_lock)) {
175 		ci->ci_biglock_count = nlocks;
176 		l->l_blcnt = nlocks;
177 		LOCKDEBUG_LOCKED(kernel_lock_dodebug, kernel_lock,
178 		    RETURN_ADDRESS, 0);
179 		splx(s);
180 		return;
181 	}
182 
183 	/*
184 	 * To remove the ordering constraint between adaptive mutexes
185 	 * and kernel_lock we must make it appear as if this thread is
186 	 * blocking.  For non-interlocked mutex release, a store fence
187 	 * is required to ensure that the result of any mutex_exit()
188 	 * by the current LWP becomes visible on the bus before the set
189 	 * of ci->ci_biglock_wanted becomes visible.
190 	 */
191 	membar_producer();
192 	owant = ci->ci_biglock_wanted;
193 	ci->ci_biglock_wanted = l;
194 
195 	/*
196 	 * Spin until we acquire the lock.  Once we have it, record the
197 	 * time spent with lockstat.
198 	 */
199 	LOCKSTAT_ENTER(lsflag);
200 	LOCKSTAT_START_TIMER(lsflag, spintime);
201 
202 	spins = 0;
203 	do {
204 		splx(s);
205 		while (__SIMPLELOCK_LOCKED_P(kernel_lock)) {
206 			if (SPINLOCK_SPINOUT(spins)) {
207 				_KERNEL_LOCK_ABORT("spinout");
208 			}
209 			SPINLOCK_BACKOFF_HOOK;
210 			SPINLOCK_SPIN_HOOK;
211 		}
212 		s = splvm();
213 	} while (!__cpu_simple_lock_try(kernel_lock));
214 
215 	ci->ci_biglock_count = nlocks;
216 	l->l_blcnt = nlocks;
217 	LOCKSTAT_STOP_TIMER(lsflag, spintime);
218 	LOCKDEBUG_LOCKED(kernel_lock_dodebug, kernel_lock, RETURN_ADDRESS, 0);
219 	if (owant == NULL) {
220 		LOCKSTAT_EVENT_RA(lsflag, kernel_lock,
221 		    LB_KERNEL_LOCK | LB_SPIN, 1, spintime, RETURN_ADDRESS);
222 	}
223 	LOCKSTAT_EXIT(lsflag);
224 	splx(s);
225 
226 	/*
227 	 * Now that we have kernel_lock, reset ci_biglock_wanted.  This
228 	 * store must be unbuffered (immediately visible on the bus) in
229 	 * order for non-interlocked mutex release to work correctly.
230 	 * It must be visible before a mutex_exit() can execute on this
231 	 * processor.
232 	 *
233 	 * Note: only where CAS is available in hardware will this be
234 	 * an unbuffered write, but non-interlocked release cannot be
235 	 * done on CPUs without CAS in hardware.
236 	 */
237 	(void)atomic_swap_ptr(&ci->ci_biglock_wanted, owant);
238 
239 	/*
240 	 * Issue a memory barrier as we have acquired a lock.  This also
241 	 * prevents stores from a following mutex_exit() being reordered
242 	 * to occur before our store to ci_biglock_wanted above.
243 	 */
244 	membar_enter();
245 }
246 
247 /*
248  * Release 'nlocks' holds on the kernel lock.  If 'nlocks' is zero, release
249  * all holds.  If 'l' is non-null, the release is from process context.
250  */
251 void
252 _kernel_unlock(int nlocks, struct lwp *l, int *countp)
253 {
254 	struct cpu_info *ci = curcpu();
255 	u_int olocks;
256 	int s;
257 
258 	l = curlwp;
259 
260 	_KERNEL_LOCK_ASSERT(nlocks < 2);
261 
262 	olocks = l->l_blcnt;
263 
264 	if (olocks == 0) {
265 		_KERNEL_LOCK_ASSERT(nlocks <= 0);
266 		if (countp != NULL)
267 			*countp = 0;
268 		return;
269 	}
270 
271 	_KERNEL_LOCK_ASSERT(__SIMPLELOCK_LOCKED_P(kernel_lock));
272 
273 	if (nlocks == 0)
274 		nlocks = olocks;
275 	else if (nlocks == -1) {
276 		nlocks = 1;
277 		_KERNEL_LOCK_ASSERT(olocks == 1);
278 	}
279 
280 	_KERNEL_LOCK_ASSERT(ci->ci_biglock_count >= l->l_blcnt);
281 
282 	l->l_blcnt -= nlocks;
283 	if (ci->ci_biglock_count == nlocks) {
284 		s = splvm();
285 		LOCKDEBUG_UNLOCKED(kernel_lock_dodebug, kernel_lock,
286 		    RETURN_ADDRESS, 0);
287 		ci->ci_biglock_count = 0;
288 		__cpu_simple_unlock(kernel_lock);
289 		splx(s);
290 	} else
291 		ci->ci_biglock_count -= nlocks;
292 
293 	if (countp != NULL)
294 		*countp = olocks;
295 }
296 #endif /* !_RUMPKERNEL */
297