xref: /netbsd-src/sys/kern/kern_lock.c (revision 11a6dbe72840351315e0652b2fc6663628c84cad)
1 /*	$NetBSD: kern_lock.c,v 1.140 2008/04/28 20:24:03 martin 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  *
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 1.140 2008/04/28 20:24:03 martin Exp $");
35 
36 #include "opt_multiprocessor.h"
37 
38 #include <sys/param.h>
39 #include <sys/proc.h>
40 #include <sys/lock.h>
41 #include <sys/systm.h>
42 #include <sys/kernel.h>
43 #include <sys/lockdebug.h>
44 #include <sys/cpu.h>
45 #include <sys/syslog.h>
46 #include <sys/atomic.h>
47 
48 #include <machine/stdarg.h>
49 #include <machine/lock.h>
50 
51 #include <dev/lockstat.h>
52 
53 #define	RETURN_ADDRESS	(uintptr_t)__builtin_return_address(0)
54 
55 bool	kernel_lock_dodebug;
56 
57 __cpu_simple_lock_t kernel_lock[CACHE_LINE_SIZE / sizeof(__cpu_simple_lock_t)]
58     __aligned(CACHE_LINE_SIZE);
59 
60 #if defined(DEBUG) || defined(LKM)
61 void
62 assert_sleepable(void)
63 {
64 #if !defined(_RUMPKERNEL)
65 	const char *reason;
66 
67 	if (panicstr != NULL) {
68 		return;
69 	}
70 
71 	LOCKDEBUG_BARRIER(kernel_lock, 1);
72 
73 	reason = NULL;
74 	if (CURCPU_IDLE_P() && !cold) {
75 		reason = "idle";
76 	}
77 	if (cpu_intr_p()) {
78 		reason = "interrupt";
79 	}
80 	if ((curlwp->l_pflag & LP_INTR) != 0) {
81 		reason = "softint";
82 	}
83 
84 	if (reason) {
85 		panic("%s: %s caller=%p", __func__, reason,
86 		    (void *)RETURN_ADDRESS);
87 	}
88 #endif /* !defined(_RUMPKERNEL) */
89 }
90 #endif /* defined(DEBUG) || defined(LKM) */
91 
92 /*
93  * rump doesn't need the kernel lock so force it out.  We cannot
94  * currently easily include it for compilation because of
95  * a) SPINLOCK_* b) membar_producer().  They are defined in different
96  * places / way for each arch, so just simply do not bother to
97  * fight a lot for no gain (i.e. pain but still no gain).
98  */
99 #ifndef _RUMPKERNEL
100 /*
101  * Functions for manipulating the kernel_lock.  We put them here
102  * so that they show up in profiles.
103  */
104 
105 #define	_KERNEL_LOCK_ABORT(msg)						\
106     LOCKDEBUG_ABORT(kernel_lock, &_kernel_lock_ops, __func__, msg)
107 
108 #ifdef LOCKDEBUG
109 #define	_KERNEL_LOCK_ASSERT(cond)					\
110 do {									\
111 	if (!(cond))							\
112 		_KERNEL_LOCK_ABORT("assertion failed: " #cond);		\
113 } while (/* CONSTCOND */ 0)
114 #else
115 #define	_KERNEL_LOCK_ASSERT(cond)	/* nothing */
116 #endif
117 
118 void	_kernel_lock_dump(volatile void *);
119 
120 lockops_t _kernel_lock_ops = {
121 	"Kernel lock",
122 	0,
123 	_kernel_lock_dump
124 };
125 
126 /*
127  * Initialize the kernel lock.
128  */
129 void
130 kernel_lock_init(void)
131 {
132 
133 	KASSERT(CACHE_LINE_SIZE >= sizeof(__cpu_simple_lock_t));
134 	__cpu_simple_lock_init(kernel_lock);
135 	kernel_lock_dodebug = LOCKDEBUG_ALLOC(kernel_lock, &_kernel_lock_ops,
136 	    RETURN_ADDRESS);
137 }
138 
139 /*
140  * Print debugging information about the kernel lock.
141  */
142 void
143 _kernel_lock_dump(volatile void *junk)
144 {
145 	struct cpu_info *ci = curcpu();
146 
147 	(void)junk;
148 
149 	printf_nolog("curcpu holds : %18d wanted by: %#018lx\n",
150 	    ci->ci_biglock_count, (long)ci->ci_biglock_wanted);
151 }
152 
153 /*
154  * Acquire 'nlocks' holds on the kernel lock.  If 'l' is non-null, the
155  * acquisition is from process context.
156  */
157 void
158 _kernel_lock(int nlocks)
159 {
160 	struct cpu_info *ci;
161 	LOCKSTAT_TIMER(spintime);
162 	LOCKSTAT_FLAG(lsflag);
163 	struct lwp *owant;
164 	u_int spins;
165 	int s;
166 	struct lwp *l = curlwp;
167 
168 	_KERNEL_LOCK_ASSERT(nlocks > 0);
169 
170 	s = splvm();
171 	ci = curcpu();
172 	if (ci->ci_biglock_count != 0) {
173 		_KERNEL_LOCK_ASSERT(__SIMPLELOCK_LOCKED_P(kernel_lock));
174 		ci->ci_biglock_count += nlocks;
175 		l->l_blcnt += nlocks;
176 		splx(s);
177 		return;
178 	}
179 
180 	_KERNEL_LOCK_ASSERT(l->l_blcnt == 0);
181 	LOCKDEBUG_WANTLOCK(kernel_lock_dodebug, kernel_lock, RETURN_ADDRESS,
182 	    0);
183 
184 	if (__cpu_simple_lock_try(kernel_lock)) {
185 		ci->ci_biglock_count = nlocks;
186 		l->l_blcnt = nlocks;
187 		LOCKDEBUG_LOCKED(kernel_lock_dodebug, kernel_lock,
188 		    RETURN_ADDRESS, 0);
189 		splx(s);
190 		return;
191 	}
192 
193 	/*
194 	 * To remove the ordering constraint between adaptive mutexes
195 	 * and kernel_lock we must make it appear as if this thread is
196 	 * blocking.  For non-interlocked mutex release, a store fence
197 	 * is required to ensure that the result of any mutex_exit()
198 	 * by the current LWP becomes visible on the bus before the set
199 	 * of ci->ci_biglock_wanted becomes visible.
200 	 */
201 	membar_producer();
202 	owant = ci->ci_biglock_wanted;
203 	ci->ci_biglock_wanted = l;
204 
205 	/*
206 	 * Spin until we acquire the lock.  Once we have it, record the
207 	 * time spent with lockstat.
208 	 */
209 	LOCKSTAT_ENTER(lsflag);
210 	LOCKSTAT_START_TIMER(lsflag, spintime);
211 
212 	spins = 0;
213 	do {
214 		splx(s);
215 		while (__SIMPLELOCK_LOCKED_P(kernel_lock)) {
216 			if (SPINLOCK_SPINOUT(spins)) {
217 				extern volatile int start_init_exec;
218 				if (!start_init_exec)
219 					_KERNEL_LOCK_ABORT("spinout");
220 			}
221 			SPINLOCK_BACKOFF_HOOK;
222 			SPINLOCK_SPIN_HOOK;
223 		}
224 		s = splvm();
225 	} while (!__cpu_simple_lock_try(kernel_lock));
226 
227 	ci->ci_biglock_count = nlocks;
228 	l->l_blcnt = nlocks;
229 	LOCKSTAT_STOP_TIMER(lsflag, spintime);
230 	LOCKDEBUG_LOCKED(kernel_lock_dodebug, kernel_lock, RETURN_ADDRESS, 0);
231 	if (owant == NULL) {
232 		LOCKSTAT_EVENT_RA(lsflag, kernel_lock,
233 		    LB_KERNEL_LOCK | LB_SPIN, 1, spintime, RETURN_ADDRESS);
234 	}
235 	LOCKSTAT_EXIT(lsflag);
236 	splx(s);
237 
238 	/*
239 	 * Now that we have kernel_lock, reset ci_biglock_wanted.  This
240 	 * store must be unbuffered (immediately visible on the bus) in
241 	 * order for non-interlocked mutex release to work correctly.
242 	 * It must be visible before a mutex_exit() can execute on this
243 	 * processor.
244 	 *
245 	 * Note: only where CAS is available in hardware will this be
246 	 * an unbuffered write, but non-interlocked release cannot be
247 	 * done on CPUs without CAS in hardware.
248 	 */
249 	(void)atomic_swap_ptr(&ci->ci_biglock_wanted, owant);
250 
251 	/*
252 	 * Issue a memory barrier as we have acquired a lock.  This also
253 	 * prevents stores from a following mutex_exit() being reordered
254 	 * to occur before our store to ci_biglock_wanted above.
255 	 */
256 	membar_enter();
257 }
258 
259 /*
260  * Release 'nlocks' holds on the kernel lock.  If 'nlocks' is zero, release
261  * all holds.  If 'l' is non-null, the release is from process context.
262  */
263 void
264 _kernel_unlock(int nlocks, int *countp)
265 {
266 	struct cpu_info *ci;
267 	u_int olocks;
268 	int s;
269 	struct lwp *l = curlwp;
270 
271 	_KERNEL_LOCK_ASSERT(nlocks < 2);
272 
273 	olocks = l->l_blcnt;
274 
275 	if (olocks == 0) {
276 		_KERNEL_LOCK_ASSERT(nlocks <= 0);
277 		if (countp != NULL)
278 			*countp = 0;
279 		return;
280 	}
281 
282 	_KERNEL_LOCK_ASSERT(__SIMPLELOCK_LOCKED_P(kernel_lock));
283 
284 	if (nlocks == 0)
285 		nlocks = olocks;
286 	else if (nlocks == -1) {
287 		nlocks = 1;
288 		_KERNEL_LOCK_ASSERT(olocks == 1);
289 	}
290 	s = splvm();
291 	ci = curcpu();
292 	_KERNEL_LOCK_ASSERT(ci->ci_biglock_count >= l->l_blcnt);
293 	if (ci->ci_biglock_count == nlocks) {
294 		LOCKDEBUG_UNLOCKED(kernel_lock_dodebug, kernel_lock,
295 		    RETURN_ADDRESS, 0);
296 		ci->ci_biglock_count = 0;
297 		__cpu_simple_unlock(kernel_lock);
298 		l->l_blcnt -= nlocks;
299 		splx(s);
300 		if (l->l_dopreempt)
301 			kpreempt(0);
302 	} else {
303 		ci->ci_biglock_count -= nlocks;
304 		l->l_blcnt -= nlocks;
305 		splx(s);
306 	}
307 
308 	if (countp != NULL)
309 		*countp = olocks;
310 }
311 #endif /* !_RUMPKERNEL */
312