xref: /netbsd-src/sys/kern/kern_lock.c (revision 8ac07aec990b9d2e483062509d0a9fa5b4f57cf2)
1 /*	$NetBSD: kern_lock.c,v 1.137 2008/04/01 19:49:31 drochner 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 1.137 2008/04/01 19:49:31 drochner 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(DEBUG) || defined(LKM)
68 void
69 assert_sleepable(void)
70 {
71 #if !defined(_RUMPKERNEL)
72 	const char *reason;
73 
74 	if (panicstr != NULL) {
75 		return;
76 	}
77 
78 	LOCKDEBUG_BARRIER(kernel_lock, 1);
79 
80 	reason = NULL;
81 	if (CURCPU_IDLE_P() && !cold) {
82 		reason = "idle";
83 	}
84 	if (cpu_intr_p()) {
85 		reason = "interrupt";
86 	}
87 	if ((curlwp->l_pflag & LP_INTR) != 0) {
88 		reason = "softint";
89 	}
90 
91 	if (reason) {
92 		panic("%s: %s caller=%p", __func__, reason,
93 		    (void *)RETURN_ADDRESS);
94 	}
95 #endif /* !defined(_RUMPKERNEL) */
96 }
97 #endif /* defined(DEBUG) || defined(LKM) */
98 
99 /*
100  * rump doesn't need the kernel lock so force it out.  We cannot
101  * currently easily include it for compilation because of
102  * a) SPINLOCK_* b) membar_producer().  They are defined in different
103  * places / way for each arch, so just simply do not bother to
104  * fight a lot for no gain (i.e. pain but still no gain).
105  */
106 #ifndef _RUMPKERNEL
107 /*
108  * Functions for manipulating the kernel_lock.  We put them here
109  * so that they show up in profiles.
110  */
111 
112 #define	_KERNEL_LOCK_ABORT(msg)						\
113     LOCKDEBUG_ABORT(kernel_lock, &_kernel_lock_ops, __func__, msg)
114 
115 #ifdef LOCKDEBUG
116 #define	_KERNEL_LOCK_ASSERT(cond)					\
117 do {									\
118 	if (!(cond))							\
119 		_KERNEL_LOCK_ABORT("assertion failed: " #cond);		\
120 } while (/* CONSTCOND */ 0)
121 #else
122 #define	_KERNEL_LOCK_ASSERT(cond)	/* nothing */
123 #endif
124 
125 void	_kernel_lock_dump(volatile void *);
126 
127 lockops_t _kernel_lock_ops = {
128 	"Kernel lock",
129 	0,
130 	_kernel_lock_dump
131 };
132 
133 /*
134  * Initialize the kernel lock.
135  */
136 void
137 kernel_lock_init(void)
138 {
139 
140 	KASSERT(CACHE_LINE_SIZE >= sizeof(__cpu_simple_lock_t));
141 	__cpu_simple_lock_init(kernel_lock);
142 	kernel_lock_dodebug = LOCKDEBUG_ALLOC(kernel_lock, &_kernel_lock_ops,
143 	    RETURN_ADDRESS);
144 }
145 
146 /*
147  * Print debugging information about the kernel lock.
148  */
149 void
150 _kernel_lock_dump(volatile void *junk)
151 {
152 	struct cpu_info *ci = curcpu();
153 
154 	(void)junk;
155 
156 	printf_nolog("curcpu holds : %18d wanted by: %#018lx\n",
157 	    ci->ci_biglock_count, (long)ci->ci_biglock_wanted);
158 }
159 
160 /*
161  * Acquire 'nlocks' holds on the kernel lock.  If 'l' is non-null, the
162  * acquisition is from process context.
163  */
164 void
165 _kernel_lock(int nlocks)
166 {
167 	struct cpu_info *ci = curcpu();
168 	LOCKSTAT_TIMER(spintime);
169 	LOCKSTAT_FLAG(lsflag);
170 	struct lwp *owant;
171 	u_int spins;
172 	int s;
173 	struct lwp *l = curlwp;
174 
175 	_KERNEL_LOCK_ASSERT(nlocks > 0);
176 
177 	if (ci->ci_biglock_count != 0) {
178 		_KERNEL_LOCK_ASSERT(__SIMPLELOCK_LOCKED_P(kernel_lock));
179 		ci->ci_biglock_count += nlocks;
180 		l->l_blcnt += nlocks;
181 		return;
182 	}
183 
184 	_KERNEL_LOCK_ASSERT(l->l_blcnt == 0);
185 	LOCKDEBUG_WANTLOCK(kernel_lock_dodebug, kernel_lock, RETURN_ADDRESS,
186 	    0);
187 
188 	s = splvm();
189 	if (__cpu_simple_lock_try(kernel_lock)) {
190 		ci->ci_biglock_count = nlocks;
191 		l->l_blcnt = nlocks;
192 		LOCKDEBUG_LOCKED(kernel_lock_dodebug, kernel_lock,
193 		    RETURN_ADDRESS, 0);
194 		splx(s);
195 		return;
196 	}
197 
198 	/*
199 	 * To remove the ordering constraint between adaptive mutexes
200 	 * and kernel_lock we must make it appear as if this thread is
201 	 * blocking.  For non-interlocked mutex release, a store fence
202 	 * is required to ensure that the result of any mutex_exit()
203 	 * by the current LWP becomes visible on the bus before the set
204 	 * of ci->ci_biglock_wanted becomes visible.
205 	 */
206 	membar_producer();
207 	owant = ci->ci_biglock_wanted;
208 	ci->ci_biglock_wanted = l;
209 
210 	/*
211 	 * Spin until we acquire the lock.  Once we have it, record the
212 	 * time spent with lockstat.
213 	 */
214 	LOCKSTAT_ENTER(lsflag);
215 	LOCKSTAT_START_TIMER(lsflag, spintime);
216 
217 	spins = 0;
218 	do {
219 		splx(s);
220 		while (__SIMPLELOCK_LOCKED_P(kernel_lock)) {
221 			if (SPINLOCK_SPINOUT(spins)) {
222 				extern volatile int start_init_exec;
223 				if (!start_init_exec)
224 					_KERNEL_LOCK_ABORT("spinout");
225 			}
226 			SPINLOCK_BACKOFF_HOOK;
227 			SPINLOCK_SPIN_HOOK;
228 		}
229 		s = splvm();
230 	} while (!__cpu_simple_lock_try(kernel_lock));
231 
232 	ci->ci_biglock_count = nlocks;
233 	l->l_blcnt = nlocks;
234 	LOCKSTAT_STOP_TIMER(lsflag, spintime);
235 	LOCKDEBUG_LOCKED(kernel_lock_dodebug, kernel_lock, RETURN_ADDRESS, 0);
236 	if (owant == NULL) {
237 		LOCKSTAT_EVENT_RA(lsflag, kernel_lock,
238 		    LB_KERNEL_LOCK | LB_SPIN, 1, spintime, RETURN_ADDRESS);
239 	}
240 	LOCKSTAT_EXIT(lsflag);
241 	splx(s);
242 
243 	/*
244 	 * Now that we have kernel_lock, reset ci_biglock_wanted.  This
245 	 * store must be unbuffered (immediately visible on the bus) in
246 	 * order for non-interlocked mutex release to work correctly.
247 	 * It must be visible before a mutex_exit() can execute on this
248 	 * processor.
249 	 *
250 	 * Note: only where CAS is available in hardware will this be
251 	 * an unbuffered write, but non-interlocked release cannot be
252 	 * done on CPUs without CAS in hardware.
253 	 */
254 	(void)atomic_swap_ptr(&ci->ci_biglock_wanted, owant);
255 
256 	/*
257 	 * Issue a memory barrier as we have acquired a lock.  This also
258 	 * prevents stores from a following mutex_exit() being reordered
259 	 * to occur before our store to ci_biglock_wanted above.
260 	 */
261 	membar_enter();
262 }
263 
264 /*
265  * Release 'nlocks' holds on the kernel lock.  If 'nlocks' is zero, release
266  * all holds.  If 'l' is non-null, the release is from process context.
267  */
268 void
269 _kernel_unlock(int nlocks, int *countp)
270 {
271 	struct cpu_info *ci = curcpu();
272 	u_int olocks;
273 	int s;
274 	struct lwp *l = curlwp;
275 
276 	_KERNEL_LOCK_ASSERT(nlocks < 2);
277 
278 	olocks = l->l_blcnt;
279 
280 	if (olocks == 0) {
281 		_KERNEL_LOCK_ASSERT(nlocks <= 0);
282 		if (countp != NULL)
283 			*countp = 0;
284 		return;
285 	}
286 
287 	_KERNEL_LOCK_ASSERT(__SIMPLELOCK_LOCKED_P(kernel_lock));
288 
289 	if (nlocks == 0)
290 		nlocks = olocks;
291 	else if (nlocks == -1) {
292 		nlocks = 1;
293 		_KERNEL_LOCK_ASSERT(olocks == 1);
294 	}
295 
296 	_KERNEL_LOCK_ASSERT(ci->ci_biglock_count >= l->l_blcnt);
297 
298 	l->l_blcnt -= nlocks;
299 	if (ci->ci_biglock_count == nlocks) {
300 		s = splvm();
301 		LOCKDEBUG_UNLOCKED(kernel_lock_dodebug, kernel_lock,
302 		    RETURN_ADDRESS, 0);
303 		ci->ci_biglock_count = 0;
304 		__cpu_simple_unlock(kernel_lock);
305 		splx(s);
306 	} else
307 		ci->ci_biglock_count -= nlocks;
308 
309 	if (countp != NULL)
310 		*countp = olocks;
311 }
312 #endif /* !_RUMPKERNEL */
313