xref: /netbsd-src/sys/kern/kern_lock.c (revision ba65fde2d7fefa7d39838fa5fa855e62bd606b5e)
1 /*	$NetBSD: kern_lock.c,v 1.153 2012/08/30 02:23:14 matt Exp $	*/
2 
3 /*-
4  * Copyright (c) 2002, 2006, 2007, 2008, 2009 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  *
20  * THIS SOFTWARE IS PROVIDED BY THE NETBSD FOUNDATION, INC. AND CONTRIBUTORS
21  * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED
22  * TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
23  * PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL THE FOUNDATION OR CONTRIBUTORS
24  * BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
25  * CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
26  * SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
27  * INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
28  * CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
29  * ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
30  * POSSIBILITY OF SUCH DAMAGE.
31  */
32 
33 #include <sys/cdefs.h>
34 __KERNEL_RCSID(0, "$NetBSD: kern_lock.c,v 1.153 2012/08/30 02:23:14 matt Exp $");
35 
36 #include <sys/param.h>
37 #include <sys/proc.h>
38 #include <sys/lock.h>
39 #include <sys/systm.h>
40 #include <sys/kernel.h>
41 #include <sys/lockdebug.h>
42 #include <sys/cpu.h>
43 #include <sys/syslog.h>
44 #include <sys/atomic.h>
45 #include <sys/lwp.h>
46 
47 #include <machine/lock.h>
48 
49 #include <dev/lockstat.h>
50 
51 #define	RETURN_ADDRESS	(uintptr_t)__builtin_return_address(0)
52 
53 bool	kernel_lock_dodebug;
54 
55 __cpu_simple_lock_t kernel_lock[CACHE_LINE_SIZE / sizeof(__cpu_simple_lock_t)]
56     __cacheline_aligned;
57 
58 void
59 assert_sleepable(void)
60 {
61 	const char *reason;
62 	uint64_t pctr;
63 	bool idle;
64 
65 	if (panicstr != NULL) {
66 		return;
67 	}
68 
69 	LOCKDEBUG_BARRIER(kernel_lock, 1);
70 
71 	/*
72 	 * Avoid disabling/re-enabling preemption here since this
73 	 * routine may be called in delicate situations.
74 	 */
75 	do {
76 		pctr = lwp_pctr();
77 		idle = CURCPU_IDLE_P();
78 	} while (pctr != lwp_pctr());
79 
80 	reason = NULL;
81 	if (idle && !cold) {
82 		reason = "idle";
83 	}
84 	if (cpu_intr_p()) {
85 		reason = "interrupt";
86 	}
87 	if (cpu_softintr_p()) {
88 		reason = "softint";
89 	}
90 
91 	if (reason) {
92 		panic("%s: %s caller=%p", __func__, reason,
93 		    (void *)RETURN_ADDRESS);
94 	}
95 }
96 
97 /*
98  * Functions for manipulating the kernel_lock.  We put them here
99  * so that they show up in profiles.
100  */
101 
102 #define	_KERNEL_LOCK_ABORT(msg)						\
103     LOCKDEBUG_ABORT(kernel_lock, &_kernel_lock_ops, __func__, msg)
104 
105 #ifdef LOCKDEBUG
106 #define	_KERNEL_LOCK_ASSERT(cond)					\
107 do {									\
108 	if (!(cond))							\
109 		_KERNEL_LOCK_ABORT("assertion failed: " #cond);		\
110 } while (/* CONSTCOND */ 0)
111 #else
112 #define	_KERNEL_LOCK_ASSERT(cond)	/* nothing */
113 #endif
114 
115 void	_kernel_lock_dump(volatile void *);
116 
117 lockops_t _kernel_lock_ops = {
118 	"Kernel lock",
119 	LOCKOPS_SPIN,
120 	_kernel_lock_dump
121 };
122 
123 /*
124  * Initialize the kernel lock.
125  */
126 void
127 kernel_lock_init(void)
128 {
129 
130 	CTASSERT(CACHE_LINE_SIZE >= sizeof(__cpu_simple_lock_t));
131 	__cpu_simple_lock_init(kernel_lock);
132 	kernel_lock_dodebug = LOCKDEBUG_ALLOC(kernel_lock, &_kernel_lock_ops,
133 	    RETURN_ADDRESS);
134 }
135 
136 /*
137  * Print debugging information about the kernel lock.
138  */
139 void
140 _kernel_lock_dump(volatile void *junk)
141 {
142 	struct cpu_info *ci = curcpu();
143 
144 	(void)junk;
145 
146 	printf_nolog("curcpu holds : %18d wanted by: %#018lx\n",
147 	    ci->ci_biglock_count, (long)ci->ci_biglock_wanted);
148 }
149 
150 /*
151  * Acquire 'nlocks' holds on the kernel lock.
152  */
153 void
154 _kernel_lock(int nlocks)
155 {
156 	struct cpu_info *ci;
157 	LOCKSTAT_TIMER(spintime);
158 	LOCKSTAT_FLAG(lsflag);
159 	struct lwp *owant;
160 	u_int spins;
161 	int s;
162 	struct lwp *l = curlwp;
163 
164 	_KERNEL_LOCK_ASSERT(nlocks > 0);
165 
166 	s = splvm();
167 	ci = curcpu();
168 	if (ci->ci_biglock_count != 0) {
169 		_KERNEL_LOCK_ASSERT(__SIMPLELOCK_LOCKED_P(kernel_lock));
170 		ci->ci_biglock_count += nlocks;
171 		l->l_blcnt += nlocks;
172 		splx(s);
173 		return;
174 	}
175 
176 	_KERNEL_LOCK_ASSERT(l->l_blcnt == 0);
177 	LOCKDEBUG_WANTLOCK(kernel_lock_dodebug, kernel_lock, RETURN_ADDRESS,
178 	    false, false);
179 
180 	if (__cpu_simple_lock_try(kernel_lock)) {
181 		ci->ci_biglock_count = nlocks;
182 		l->l_blcnt = nlocks;
183 		LOCKDEBUG_LOCKED(kernel_lock_dodebug, kernel_lock, NULL,
184 		    RETURN_ADDRESS, 0);
185 		splx(s);
186 		return;
187 	}
188 
189 	/*
190 	 * To remove the ordering constraint between adaptive mutexes
191 	 * and kernel_lock we must make it appear as if this thread is
192 	 * blocking.  For non-interlocked mutex release, a store fence
193 	 * is required to ensure that the result of any mutex_exit()
194 	 * by the current LWP becomes visible on the bus before the set
195 	 * of ci->ci_biglock_wanted becomes visible.
196 	 */
197 	membar_producer();
198 	owant = ci->ci_biglock_wanted;
199 	ci->ci_biglock_wanted = l;
200 
201 	/*
202 	 * Spin until we acquire the lock.  Once we have it, record the
203 	 * time spent with lockstat.
204 	 */
205 	LOCKSTAT_ENTER(lsflag);
206 	LOCKSTAT_START_TIMER(lsflag, spintime);
207 
208 	spins = 0;
209 	do {
210 		splx(s);
211 		while (__SIMPLELOCK_LOCKED_P(kernel_lock)) {
212 			if (SPINLOCK_SPINOUT(spins)) {
213 				extern int start_init_exec;
214 				if (!start_init_exec)
215 					_KERNEL_LOCK_ABORT("spinout");
216 			}
217 			SPINLOCK_BACKOFF_HOOK;
218 			SPINLOCK_SPIN_HOOK;
219 		}
220 		s = splvm();
221 	} while (!__cpu_simple_lock_try(kernel_lock));
222 
223 	ci->ci_biglock_count = nlocks;
224 	l->l_blcnt = nlocks;
225 	LOCKSTAT_STOP_TIMER(lsflag, spintime);
226 	LOCKDEBUG_LOCKED(kernel_lock_dodebug, kernel_lock, NULL,
227 	    RETURN_ADDRESS, 0);
228 	if (owant == NULL) {
229 		LOCKSTAT_EVENT_RA(lsflag, kernel_lock,
230 		    LB_KERNEL_LOCK | LB_SPIN, 1, spintime, RETURN_ADDRESS);
231 	}
232 	LOCKSTAT_EXIT(lsflag);
233 	splx(s);
234 
235 	/*
236 	 * Now that we have kernel_lock, reset ci_biglock_wanted.  This
237 	 * store must be unbuffered (immediately visible on the bus) in
238 	 * order for non-interlocked mutex release to work correctly.
239 	 * It must be visible before a mutex_exit() can execute on this
240 	 * processor.
241 	 *
242 	 * Note: only where CAS is available in hardware will this be
243 	 * an unbuffered write, but non-interlocked release cannot be
244 	 * done on CPUs without CAS in hardware.
245 	 */
246 	(void)atomic_swap_ptr(&ci->ci_biglock_wanted, owant);
247 
248 	/*
249 	 * Issue a memory barrier as we have acquired a lock.  This also
250 	 * prevents stores from a following mutex_exit() being reordered
251 	 * to occur before our store to ci_biglock_wanted above.
252 	 */
253 	membar_enter();
254 }
255 
256 /*
257  * Release 'nlocks' holds on the kernel lock.  If 'nlocks' is zero, release
258  * all holds.
259  */
260 void
261 _kernel_unlock(int nlocks, int *countp)
262 {
263 	struct cpu_info *ci;
264 	u_int olocks;
265 	int s;
266 	struct lwp *l = curlwp;
267 
268 	_KERNEL_LOCK_ASSERT(nlocks < 2);
269 
270 	olocks = l->l_blcnt;
271 
272 	if (olocks == 0) {
273 		_KERNEL_LOCK_ASSERT(nlocks <= 0);
274 		if (countp != NULL)
275 			*countp = 0;
276 		return;
277 	}
278 
279 	_KERNEL_LOCK_ASSERT(__SIMPLELOCK_LOCKED_P(kernel_lock));
280 
281 	if (nlocks == 0)
282 		nlocks = olocks;
283 	else if (nlocks == -1) {
284 		nlocks = 1;
285 		_KERNEL_LOCK_ASSERT(olocks == 1);
286 	}
287 	s = splvm();
288 	ci = curcpu();
289 	_KERNEL_LOCK_ASSERT(ci->ci_biglock_count >= l->l_blcnt);
290 	if (ci->ci_biglock_count == nlocks) {
291 		LOCKDEBUG_UNLOCKED(kernel_lock_dodebug, kernel_lock,
292 		    RETURN_ADDRESS, 0);
293 		ci->ci_biglock_count = 0;
294 		__cpu_simple_unlock(kernel_lock);
295 		l->l_blcnt -= nlocks;
296 		splx(s);
297 		if (l->l_dopreempt)
298 			kpreempt(0);
299 	} else {
300 		ci->ci_biglock_count -= nlocks;
301 		l->l_blcnt -= nlocks;
302 		splx(s);
303 	}
304 
305 	if (countp != NULL)
306 		*countp = olocks;
307 }
308 
309 bool
310 _kernel_locked_p(void)
311 {
312 	return __SIMPLELOCK_LOCKED_P(kernel_lock);
313 }
314