xref: /netbsd-src/sys/kern/kern_lock.c (revision 53b02e147d4ed531c0d2a5ca9b3e8026ba3e99b5)
1 /*	$NetBSD: kern_lock.c,v 1.173 2021/10/31 16:26:26 skrll 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 1.173 2021/10/31 16:26:26 skrll 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 		reason = "idle";
94 	}
95 	if (cpu_intr_p()) {
96 		reason = "interrupt";
97 	}
98 	if (cpu_softintr_p()) {
99 		reason = "softint";
100 	}
101 	if (!pserialize_not_in_read_section()) {
102 		reason = "pserialize";
103 	}
104 
105 	if (reason) {
106 		panic("%s: %s caller=%p", __func__, reason,
107 		    (void *)RETURN_ADDRESS);
108 	}
109 }
110 
111 /*
112  * Functions for manipulating the kernel_lock.  We put them here
113  * so that they show up in profiles.
114  */
115 
116 #define	_KERNEL_LOCK_ABORT(msg)						\
117     LOCKDEBUG_ABORT(__func__, __LINE__, kernel_lock, &_kernel_lock_ops, msg)
118 
119 #ifdef LOCKDEBUG
120 #define	_KERNEL_LOCK_ASSERT(cond)					\
121 do {									\
122 	if (!(cond))							\
123 		_KERNEL_LOCK_ABORT("assertion failed: " #cond);		\
124 } while (/* CONSTCOND */ 0)
125 #else
126 #define	_KERNEL_LOCK_ASSERT(cond)	/* nothing */
127 #endif
128 
129 static void	_kernel_lock_dump(const volatile void *, lockop_printer_t);
130 
131 lockops_t _kernel_lock_ops = {
132 	.lo_name = "Kernel lock",
133 	.lo_type = LOCKOPS_SPIN,
134 	.lo_dump = _kernel_lock_dump,
135 };
136 
137 /*
138  * Initialize the kernel lock.
139  */
140 void
141 kernel_lock_init(void)
142 {
143 
144 	__cpu_simple_lock_init(kernel_lock);
145 	kernel_lock_dodebug = LOCKDEBUG_ALLOC(kernel_lock, &_kernel_lock_ops,
146 	    RETURN_ADDRESS);
147 }
148 CTASSERT(CACHE_LINE_SIZE >= sizeof(__cpu_simple_lock_t));
149 
150 /*
151  * Print debugging information about the kernel lock.
152  */
153 static void
154 _kernel_lock_dump(const volatile void *junk, lockop_printer_t pr)
155 {
156 	struct cpu_info *ci = curcpu();
157 
158 	(void)junk;
159 
160 	pr("curcpu holds : %18d wanted by: %#018lx\n",
161 	    ci->ci_biglock_count, (long)ci->ci_biglock_wanted);
162 }
163 
164 /*
165  * Acquire 'nlocks' holds on the kernel lock.
166  *
167  * Although it may not look it, this is one of the most central, intricate
168  * routines in the kernel, and tons of code elsewhere depends on its exact
169  * behaviour.  If you change something in here, expect it to bite you in the
170  * rear.
171  */
172 void
173 _kernel_lock(int nlocks)
174 {
175 	struct cpu_info *ci;
176 	LOCKSTAT_TIMER(spintime);
177 	LOCKSTAT_FLAG(lsflag);
178 	struct lwp *owant;
179 #ifdef LOCKDEBUG
180 	u_int spins = 0;
181 #endif
182 	int s;
183 	struct lwp *l = curlwp;
184 
185 	_KERNEL_LOCK_ASSERT(nlocks > 0);
186 
187 	s = splvm();
188 	ci = curcpu();
189 	if (ci->ci_biglock_count != 0) {
190 		_KERNEL_LOCK_ASSERT(__SIMPLELOCK_LOCKED_P(kernel_lock));
191 		ci->ci_biglock_count += nlocks;
192 		l->l_blcnt += nlocks;
193 		splx(s);
194 		return;
195 	}
196 
197 	_KERNEL_LOCK_ASSERT(l->l_blcnt == 0);
198 	LOCKDEBUG_WANTLOCK(kernel_lock_dodebug, kernel_lock, RETURN_ADDRESS,
199 	    0);
200 
201 	if (__predict_true(__cpu_simple_lock_try(kernel_lock))) {
202 		ci->ci_biglock_count = nlocks;
203 		l->l_blcnt = nlocks;
204 		LOCKDEBUG_LOCKED(kernel_lock_dodebug, kernel_lock, NULL,
205 		    RETURN_ADDRESS, 0);
206 		splx(s);
207 		return;
208 	}
209 
210 	/*
211 	 * To remove the ordering constraint between adaptive mutexes
212 	 * and kernel_lock we must make it appear as if this thread is
213 	 * blocking.  For non-interlocked mutex release, a store fence
214 	 * is required to ensure that the result of any mutex_exit()
215 	 * by the current LWP becomes visible on the bus before the set
216 	 * of ci->ci_biglock_wanted becomes visible.
217 	 */
218 	membar_producer();
219 	owant = ci->ci_biglock_wanted;
220 	ci->ci_biglock_wanted = l;
221 #if defined(DIAGNOSTIC) && !defined(LOCKDEBUG)
222 	l->l_ld_wanted = __builtin_return_address(0);
223 #endif
224 
225 	/*
226 	 * Spin until we acquire the lock.  Once we have it, record the
227 	 * time spent with lockstat.
228 	 */
229 	LOCKSTAT_ENTER(lsflag);
230 	LOCKSTAT_START_TIMER(lsflag, spintime);
231 
232 	do {
233 		splx(s);
234 		while (__SIMPLELOCK_LOCKED_P(kernel_lock)) {
235 #ifdef LOCKDEBUG
236 			if (SPINLOCK_SPINOUT(spins)) {
237 				extern int start_init_exec;
238 				if (start_init_exec)
239 					_KERNEL_LOCK_ABORT("spinout");
240 			}
241 			SPINLOCK_BACKOFF_HOOK;
242 			SPINLOCK_SPIN_HOOK;
243 #endif
244 		}
245 		s = splvm();
246 	} while (!__cpu_simple_lock_try(kernel_lock));
247 
248 	ci->ci_biglock_count = nlocks;
249 	l->l_blcnt = nlocks;
250 	LOCKSTAT_STOP_TIMER(lsflag, spintime);
251 	LOCKDEBUG_LOCKED(kernel_lock_dodebug, kernel_lock, NULL,
252 	    RETURN_ADDRESS, 0);
253 	if (owant == NULL) {
254 		LOCKSTAT_EVENT_RA(lsflag, kernel_lock,
255 		    LB_KERNEL_LOCK | LB_SPIN, 1, spintime, RETURN_ADDRESS);
256 	}
257 	LOCKSTAT_EXIT(lsflag);
258 	splx(s);
259 
260 	/*
261 	 * Now that we have kernel_lock, reset ci_biglock_wanted.  This
262 	 * store must be unbuffered (immediately visible on the bus) in
263 	 * order for non-interlocked mutex release to work correctly.
264 	 * It must be visible before a mutex_exit() can execute on this
265 	 * processor.
266 	 *
267 	 * Note: only where CAS is available in hardware will this be
268 	 * an unbuffered write, but non-interlocked release cannot be
269 	 * done on CPUs without CAS in hardware.
270 	 */
271 	(void)atomic_swap_ptr(&ci->ci_biglock_wanted, owant);
272 
273 	/*
274 	 * Issue a memory barrier as we have acquired a lock.  This also
275 	 * prevents stores from a following mutex_exit() being reordered
276 	 * to occur before our store to ci_biglock_wanted above.
277 	 */
278 #ifndef __HAVE_ATOMIC_AS_MEMBAR
279 	membar_enter();
280 #endif
281 }
282 
283 /*
284  * Release 'nlocks' holds on the kernel lock.  If 'nlocks' is zero, release
285  * all holds.
286  */
287 void
288 _kernel_unlock(int nlocks, int *countp)
289 {
290 	struct cpu_info *ci;
291 	u_int olocks;
292 	int s;
293 	struct lwp *l = curlwp;
294 
295 	_KERNEL_LOCK_ASSERT(nlocks < 2);
296 
297 	olocks = l->l_blcnt;
298 
299 	if (olocks == 0) {
300 		_KERNEL_LOCK_ASSERT(nlocks <= 0);
301 		if (countp != NULL)
302 			*countp = 0;
303 		return;
304 	}
305 
306 	_KERNEL_LOCK_ASSERT(__SIMPLELOCK_LOCKED_P(kernel_lock));
307 
308 	if (nlocks == 0)
309 		nlocks = olocks;
310 	else if (nlocks == -1) {
311 		nlocks = 1;
312 		_KERNEL_LOCK_ASSERT(olocks == 1);
313 	}
314 	s = splvm();
315 	ci = curcpu();
316 	_KERNEL_LOCK_ASSERT(ci->ci_biglock_count >= l->l_blcnt);
317 	if (ci->ci_biglock_count == nlocks) {
318 		LOCKDEBUG_UNLOCKED(kernel_lock_dodebug, kernel_lock,
319 		    RETURN_ADDRESS, 0);
320 		ci->ci_biglock_count = 0;
321 		__cpu_simple_unlock(kernel_lock);
322 		l->l_blcnt -= nlocks;
323 		splx(s);
324 		if (l->l_dopreempt)
325 			kpreempt(0);
326 	} else {
327 		ci->ci_biglock_count -= nlocks;
328 		l->l_blcnt -= nlocks;
329 		splx(s);
330 	}
331 
332 	if (countp != NULL)
333 		*countp = olocks;
334 }
335 
336 bool
337 _kernel_locked_p(void)
338 {
339 	return __SIMPLELOCK_LOCKED_P(kernel_lock);
340 }
341