xref: /netbsd-src/sys/kern/kern_lock.c (revision 627f7eb200a4419d89b531d55fccd2ee3ffdcde0)
1 /*	$NetBSD: kern_lock.c,v 1.172 2020/12/22 01:57:29 ad Exp $	*/
2 
3 /*-
4  * Copyright (c) 2002, 2006, 2007, 2008, 2009, 2020 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 1.172 2020/12/22 01:57:29 ad Exp $");
35 
36 #ifdef _KERNEL_OPT
37 #include "opt_lockdebug.h"
38 #endif
39 
40 #include <sys/param.h>
41 #include <sys/proc.h>
42 #include <sys/lock.h>
43 #include <sys/systm.h>
44 #include <sys/kernel.h>
45 #include <sys/lockdebug.h>
46 #include <sys/cpu.h>
47 #include <sys/syslog.h>
48 #include <sys/atomic.h>
49 #include <sys/lwp.h>
50 #include <sys/pserialize.h>
51 
52 #if defined(DIAGNOSTIC) && !defined(LOCKDEBUG)
53 #include <sys/ksyms.h>
54 #endif
55 
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     __cacheline_aligned;
66 
67 void
68 assert_sleepable(void)
69 {
70 	const char *reason;
71 	uint64_t pctr;
72 	bool idle;
73 
74 	if (panicstr != NULL) {
75 		return;
76 	}
77 
78 	LOCKDEBUG_BARRIER(kernel_lock, 1);
79 
80 	/*
81 	 * Avoid disabling/re-enabling preemption here since this
82 	 * routine may be called in delicate situations.
83 	 */
84 	do {
85 		pctr = lwp_pctr();
86 		__insn_barrier();
87 		idle = CURCPU_IDLE_P();
88 		__insn_barrier();
89 	} while (pctr != lwp_pctr());
90 
91 	reason = NULL;
92 	if (idle && !cold &&
93 	    kcpuset_isset(kcpuset_running, cpu_index(curcpu()))) {
94 		reason = "idle";
95 	}
96 	if (cpu_intr_p()) {
97 		reason = "interrupt";
98 	}
99 	if (cpu_softintr_p()) {
100 		reason = "softint";
101 	}
102 	if (!pserialize_not_in_read_section()) {
103 		reason = "pserialize";
104 	}
105 
106 	if (reason) {
107 		panic("%s: %s caller=%p", __func__, reason,
108 		    (void *)RETURN_ADDRESS);
109 	}
110 }
111 
112 /*
113  * Functions for manipulating the kernel_lock.  We put them here
114  * so that they show up in profiles.
115  */
116 
117 #define	_KERNEL_LOCK_ABORT(msg)						\
118     LOCKDEBUG_ABORT(__func__, __LINE__, kernel_lock, &_kernel_lock_ops, msg)
119 
120 #ifdef LOCKDEBUG
121 #define	_KERNEL_LOCK_ASSERT(cond)					\
122 do {									\
123 	if (!(cond))							\
124 		_KERNEL_LOCK_ABORT("assertion failed: " #cond);		\
125 } while (/* CONSTCOND */ 0)
126 #else
127 #define	_KERNEL_LOCK_ASSERT(cond)	/* nothing */
128 #endif
129 
130 static void	_kernel_lock_dump(const volatile void *, lockop_printer_t);
131 
132 lockops_t _kernel_lock_ops = {
133 	.lo_name = "Kernel lock",
134 	.lo_type = LOCKOPS_SPIN,
135 	.lo_dump = _kernel_lock_dump,
136 };
137 
138 /*
139  * Initialize the kernel lock.
140  */
141 void
142 kernel_lock_init(void)
143 {
144 
145 	__cpu_simple_lock_init(kernel_lock);
146 	kernel_lock_dodebug = LOCKDEBUG_ALLOC(kernel_lock, &_kernel_lock_ops,
147 	    RETURN_ADDRESS);
148 }
149 CTASSERT(CACHE_LINE_SIZE >= sizeof(__cpu_simple_lock_t));
150 
151 /*
152  * Print debugging information about the kernel lock.
153  */
154 static void
155 _kernel_lock_dump(const volatile void *junk, lockop_printer_t pr)
156 {
157 	struct cpu_info *ci = curcpu();
158 
159 	(void)junk;
160 
161 	pr("curcpu holds : %18d wanted by: %#018lx\n",
162 	    ci->ci_biglock_count, (long)ci->ci_biglock_wanted);
163 }
164 
165 /*
166  * Acquire 'nlocks' holds on the kernel lock.
167  *
168  * Although it may not look it, this is one of the most central, intricate
169  * routines in the kernel, and tons of code elsewhere depends on its exact
170  * behaviour.  If you change something in here, expect it to bite you in the
171  * rear.
172  */
173 void
174 _kernel_lock(int nlocks)
175 {
176 	struct cpu_info *ci;
177 	LOCKSTAT_TIMER(spintime);
178 	LOCKSTAT_FLAG(lsflag);
179 	struct lwp *owant;
180 #ifdef LOCKDEBUG
181 	u_int spins = 0;
182 #endif
183 	int s;
184 	struct lwp *l = curlwp;
185 
186 	_KERNEL_LOCK_ASSERT(nlocks > 0);
187 
188 	s = splvm();
189 	ci = curcpu();
190 	if (ci->ci_biglock_count != 0) {
191 		_KERNEL_LOCK_ASSERT(__SIMPLELOCK_LOCKED_P(kernel_lock));
192 		ci->ci_biglock_count += nlocks;
193 		l->l_blcnt += nlocks;
194 		splx(s);
195 		return;
196 	}
197 
198 	_KERNEL_LOCK_ASSERT(l->l_blcnt == 0);
199 	LOCKDEBUG_WANTLOCK(kernel_lock_dodebug, kernel_lock, RETURN_ADDRESS,
200 	    0);
201 
202 	if (__predict_true(__cpu_simple_lock_try(kernel_lock))) {
203 		ci->ci_biglock_count = nlocks;
204 		l->l_blcnt = nlocks;
205 		LOCKDEBUG_LOCKED(kernel_lock_dodebug, kernel_lock, NULL,
206 		    RETURN_ADDRESS, 0);
207 		splx(s);
208 		return;
209 	}
210 
211 	/*
212 	 * To remove the ordering constraint between adaptive mutexes
213 	 * and kernel_lock we must make it appear as if this thread is
214 	 * blocking.  For non-interlocked mutex release, a store fence
215 	 * is required to ensure that the result of any mutex_exit()
216 	 * by the current LWP becomes visible on the bus before the set
217 	 * of ci->ci_biglock_wanted becomes visible.
218 	 */
219 	membar_producer();
220 	owant = ci->ci_biglock_wanted;
221 	ci->ci_biglock_wanted = l;
222 #if defined(DIAGNOSTIC) && !defined(LOCKDEBUG)
223 	l->l_ld_wanted = __builtin_return_address(0);
224 #endif
225 
226 	/*
227 	 * Spin until we acquire the lock.  Once we have it, record the
228 	 * time spent with lockstat.
229 	 */
230 	LOCKSTAT_ENTER(lsflag);
231 	LOCKSTAT_START_TIMER(lsflag, spintime);
232 
233 	do {
234 		splx(s);
235 		while (__SIMPLELOCK_LOCKED_P(kernel_lock)) {
236 #ifdef LOCKDEBUG
237 			if (SPINLOCK_SPINOUT(spins)) {
238 				extern int start_init_exec;
239 				if (start_init_exec)
240 					_KERNEL_LOCK_ABORT("spinout");
241 			}
242 			SPINLOCK_BACKOFF_HOOK;
243 			SPINLOCK_SPIN_HOOK;
244 #endif
245 		}
246 		s = splvm();
247 	} while (!__cpu_simple_lock_try(kernel_lock));
248 
249 	ci->ci_biglock_count = nlocks;
250 	l->l_blcnt = nlocks;
251 	LOCKSTAT_STOP_TIMER(lsflag, spintime);
252 	LOCKDEBUG_LOCKED(kernel_lock_dodebug, kernel_lock, NULL,
253 	    RETURN_ADDRESS, 0);
254 	if (owant == NULL) {
255 		LOCKSTAT_EVENT_RA(lsflag, kernel_lock,
256 		    LB_KERNEL_LOCK | LB_SPIN, 1, spintime, RETURN_ADDRESS);
257 	}
258 	LOCKSTAT_EXIT(lsflag);
259 	splx(s);
260 
261 	/*
262 	 * Now that we have kernel_lock, reset ci_biglock_wanted.  This
263 	 * store must be unbuffered (immediately visible on the bus) in
264 	 * order for non-interlocked mutex release to work correctly.
265 	 * It must be visible before a mutex_exit() can execute on this
266 	 * processor.
267 	 *
268 	 * Note: only where CAS is available in hardware will this be
269 	 * an unbuffered write, but non-interlocked release cannot be
270 	 * done on CPUs without CAS in hardware.
271 	 */
272 	(void)atomic_swap_ptr(&ci->ci_biglock_wanted, owant);
273 
274 	/*
275 	 * Issue a memory barrier as we have acquired a lock.  This also
276 	 * prevents stores from a following mutex_exit() being reordered
277 	 * to occur before our store to ci_biglock_wanted above.
278 	 */
279 #ifndef __HAVE_ATOMIC_AS_MEMBAR
280 	membar_enter();
281 #endif
282 }
283 
284 /*
285  * Release 'nlocks' holds on the kernel lock.  If 'nlocks' is zero, release
286  * all holds.
287  */
288 void
289 _kernel_unlock(int nlocks, int *countp)
290 {
291 	struct cpu_info *ci;
292 	u_int olocks;
293 	int s;
294 	struct lwp *l = curlwp;
295 
296 	_KERNEL_LOCK_ASSERT(nlocks < 2);
297 
298 	olocks = l->l_blcnt;
299 
300 	if (olocks == 0) {
301 		_KERNEL_LOCK_ASSERT(nlocks <= 0);
302 		if (countp != NULL)
303 			*countp = 0;
304 		return;
305 	}
306 
307 	_KERNEL_LOCK_ASSERT(__SIMPLELOCK_LOCKED_P(kernel_lock));
308 
309 	if (nlocks == 0)
310 		nlocks = olocks;
311 	else if (nlocks == -1) {
312 		nlocks = 1;
313 		_KERNEL_LOCK_ASSERT(olocks == 1);
314 	}
315 	s = splvm();
316 	ci = curcpu();
317 	_KERNEL_LOCK_ASSERT(ci->ci_biglock_count >= l->l_blcnt);
318 	if (ci->ci_biglock_count == nlocks) {
319 		LOCKDEBUG_UNLOCKED(kernel_lock_dodebug, kernel_lock,
320 		    RETURN_ADDRESS, 0);
321 		ci->ci_biglock_count = 0;
322 		__cpu_simple_unlock(kernel_lock);
323 		l->l_blcnt -= nlocks;
324 		splx(s);
325 		if (l->l_dopreempt)
326 			kpreempt(0);
327 	} else {
328 		ci->ci_biglock_count -= nlocks;
329 		l->l_blcnt -= nlocks;
330 		splx(s);
331 	}
332 
333 	if (countp != NULL)
334 		*countp = olocks;
335 }
336 
337 bool
338 _kernel_locked_p(void)
339 {
340 	return __SIMPLELOCK_LOCKED_P(kernel_lock);
341 }
342