xref: /dflybsd-src/sys/kern/kern_spinlock.c (revision 5688255a2465ea75743dbf7967ff76164c91aefe)
1 /*
2  * Copyright (c) 2005 Jeffrey M. Hsu.  All rights reserved.
3  *
4  * This code is derived from software contributed to The DragonFly Project
5  * by Jeffrey M. Hsu. and Matthew Dillon
6  *
7  * Redistribution and use in source and binary forms, with or without
8  * modification, are permitted provided that the following conditions
9  * are met:
10  * 1. Redistributions of source code must retain the above copyright
11  *    notice, this list of conditions and the following disclaimer.
12  * 2. Redistributions in binary form must reproduce the above copyright
13  *    notice, this list of conditions and the following disclaimer in the
14  *    documentation and/or other materials provided with the distribution.
15  * 3. Neither the name of The DragonFly Project nor the names of its
16  *    contributors may be used to endorse or promote products derived
17  *    from this software without specific, prior written permission.
18  *
19  * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
20  * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
21  * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS
22  * FOR A PARTICULAR PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL THE
23  * COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
24  * INCIDENTAL, SPECIAL, EXEMPLARY OR CONSEQUENTIAL DAMAGES (INCLUDING,
25  * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
26  * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED
27  * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
28  * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT
29  * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
30  * SUCH DAMAGE.
31  *
32  * $DragonFly: src/sys/kern/kern_spinlock.c,v 1.16 2008/09/11 01:11:42 y0netan1 Exp $
33  */
34 
35 #include <sys/param.h>
36 #include <sys/systm.h>
37 #include <sys/types.h>
38 #include <sys/kernel.h>
39 #include <sys/sysctl.h>
40 #ifdef INVARIANTS
41 #include <sys/proc.h>
42 #endif
43 #include <sys/priv.h>
44 #include <machine/atomic.h>
45 #include <machine/cpufunc.h>
46 #include <machine/specialreg.h>
47 #include <machine/clock.h>
48 #include <sys/spinlock.h>
49 #include <sys/spinlock2.h>
50 #include <sys/ktr.h>
51 
52 #define	BACKOFF_INITIAL	1
53 #define	BACKOFF_LIMIT	256
54 
55 #ifdef SMP
56 
57 /*
58  * Kernal Trace
59  */
60 #if !defined(KTR_SPIN_CONTENTION)
61 #define KTR_SPIN_CONTENTION	KTR_ALL
62 #endif
63 #define SPIN_STRING	"spin=%p type=%c"
64 #define SPIN_ARG_SIZE	(sizeof(void *) + sizeof(int))
65 
66 KTR_INFO_MASTER(spin);
67 KTR_INFO(KTR_SPIN_CONTENTION, spin, beg, 0, SPIN_STRING, SPIN_ARG_SIZE);
68 KTR_INFO(KTR_SPIN_CONTENTION, spin, end, 1, SPIN_STRING, SPIN_ARG_SIZE);
69 KTR_INFO(KTR_SPIN_CONTENTION, spin, backoff, 2,
70 	 "spin=%p bo1=%d thr=%p bo=%d",
71 	 ((2 * sizeof(void *)) + (2 * sizeof(int))));
72 KTR_INFO(KTR_SPIN_CONTENTION, spin, bofail, 3, SPIN_STRING, SPIN_ARG_SIZE);
73 
74 #define logspin(name, mtx, type)			\
75 	KTR_LOG(spin_ ## name, mtx, type)
76 
77 #define logspin_backoff(mtx, bo1, thr, bo)		\
78 	KTR_LOG(spin_backoff, mtx, bo1, thr, bo)
79 
80 #ifdef INVARIANTS
81 static int spin_lock_test_mode;
82 #endif
83 
84 static int64_t spinlocks_contested1;
85 SYSCTL_QUAD(_debug, OID_AUTO, spinlocks_contested1, CTLFLAG_RD,
86 	    &spinlocks_contested1, 0, "");
87 
88 static int64_t spinlocks_contested2;
89 SYSCTL_QUAD(_debug, OID_AUTO, spinlocks_contested2, CTLFLAG_RD,
90 	    &spinlocks_contested2, 0, "");
91 
92 static int spinlocks_backoff_limit = BACKOFF_LIMIT;
93 SYSCTL_INT(_debug, OID_AUTO, spinlocks_bolim, CTLFLAG_RW,
94 	   &spinlocks_backoff_limit, 0, "");
95 
96 struct exponential_backoff {
97 	int backoff;
98 	int nsec;
99 	struct spinlock *mtx;
100 	sysclock_t base;
101 };
102 static int exponential_backoff(struct exponential_backoff *bo);
103 
104 static __inline
105 void
106 exponential_init(struct exponential_backoff *bo, struct spinlock *mtx)
107 {
108 	bo->backoff = BACKOFF_INITIAL;
109 	bo->nsec = 0;
110 	bo->mtx = mtx;
111 }
112 
113 /*
114  * We were either contested due to another exclusive lock holder,
115  * or due to the presence of shared locks.  We have to undo the mess
116  * we created by returning the shared locks.
117  *
118  * If there was another exclusive lock holder only the exclusive bit
119  * in value will be the only bit set.  We don't have to do anything since
120  * restoration does not involve any work.
121  *
122  * Otherwise we successfully obtained the exclusive bit.  Attempt to
123  * clear the shared bits.  If we are able to clear the shared bits
124  * we win.  Otherwise we lose and we have to restore the shared bits
125  * we couldn't clear (and also clear our exclusive bit).
126  */
127 int
128 spin_trylock_wr_contested(globaldata_t gd, struct spinlock *mtx, int value)
129 {
130 	int bit;
131 
132 	++spinlocks_contested1;
133 	if ((value & SPINLOCK_EXCLUSIVE) == 0) {
134 		while (value) {
135 			bit = bsfl(value);
136 			if (globaldata_find(bit)->gd_spinlock_rd == mtx) {
137 				atomic_swap_int(&mtx->lock, value);
138 				--gd->gd_spinlocks_wr;
139 				--gd->gd_curthread->td_critcount;
140 				return (FALSE);
141 			}
142 			value &= ~(1 << bit);
143 		}
144 		return (TRUE);
145 	}
146 	--gd->gd_spinlocks_wr;
147 	--gd->gd_curthread->td_critcount;
148 	return (FALSE);
149 }
150 
151 /*
152  * We were either contested due to another exclusive lock holder,
153  * or due to the presence of shared locks
154  *
155  * NOTE: If value indicates an exclusively held mutex, no shared bits
156  * would have been set and we can throw away value.
157  */
158 void
159 spin_lock_wr_contested(struct spinlock *mtx, int value)
160 {
161 	struct exponential_backoff backoff;
162 	globaldata_t gd = mycpu;
163 	int bit;
164 	int mask;
165 
166 	/*
167 	 * Wait until we can gain exclusive access vs another exclusive
168 	 * holder.
169 	 */
170 	exponential_init(&backoff, mtx);
171 	++spinlocks_contested1;
172 	logspin(beg, mtx, 'w');
173 
174 	while (value & SPINLOCK_EXCLUSIVE) {
175 		value = atomic_swap_int(&mtx->lock, SPINLOCK_EXCLUSIVE);
176 		if (exponential_backoff(&backoff)) {
177 			value &= ~SPINLOCK_EXCLUSIVE;
178 			break;
179 		}
180 	}
181 
182 	/*
183 	 * Kill the cached shared bit for our own cpu.  This is the most
184 	 * common case and there's no sense wasting cpu on it.  Since
185 	 * spinlocks aren't recursive, we can't own a shared ref on the
186 	 * spinlock while trying to get an exclusive one.
187 	 *
188 	 * If multiple bits are set do not stall on any single cpu.  Check
189 	 * all cpus that have the cache bit set, then loop and check again,
190 	 * until we've cleaned all the bits.
191 	 */
192 	value &= ~gd->gd_cpumask;
193 
194 	while ((mask = value) != 0) {
195 		while (mask) {
196 			bit = bsfl(value);
197 			if (globaldata_find(bit)->gd_spinlock_rd != mtx) {
198 				value &= ~(1 << bit);
199 			} else if (exponential_backoff(&backoff)) {
200 				value = 0;
201 				break;
202 			}
203 			mask &= ~(1 << bit);
204 		}
205 	}
206 	logspin(end, mtx, 'w');
207 }
208 
209 /*
210  * The cache bit wasn't set for our cpu.  Loop until we can set the bit.
211  * As with the spin_lock_rd() inline we need a memory fence after setting
212  * gd_spinlock_rd to interlock against exclusive spinlocks waiting for
213  * that field to clear.
214  */
215 void
216 spin_lock_rd_contested(struct spinlock *mtx)
217 {
218 	struct exponential_backoff backoff;
219 	globaldata_t gd = mycpu;
220 	int value = mtx->lock;
221 
222 	/*
223 	 * Shortcut the op if we can just set the cache bit.  This case
224 	 * occurs when the last lock was an exclusive lock.
225 	 */
226 	while ((value & SPINLOCK_EXCLUSIVE) == 0) {
227 		if (atomic_cmpset_int(&mtx->lock, value, value|gd->gd_cpumask))
228 			return;
229 		value = mtx->lock;
230 	}
231 
232 	exponential_init(&backoff, mtx);
233 	++spinlocks_contested1;
234 
235 	logspin(beg, mtx, 'r');
236 
237 	while ((value & gd->gd_cpumask) == 0) {
238 		if (value & SPINLOCK_EXCLUSIVE) {
239 			gd->gd_spinlock_rd = NULL;
240 			if (exponential_backoff(&backoff)) {
241 				gd->gd_spinlock_rd = mtx;
242 				break;
243 			}
244 			gd->gd_spinlock_rd = mtx;
245 			cpu_mfence();
246 		} else {
247 			if (atomic_cmpset_int(&mtx->lock, value, value|gd->gd_cpumask))
248 				break;
249 		}
250 		value = mtx->lock;
251 	}
252 	logspin(end, mtx, 'r');
253 }
254 
255 /*
256  * Handle exponential backoff and indefinite waits.
257  *
258  * If the system is handling a panic we hand the spinlock over to the caller
259  * after 1 second.  After 10 seconds we attempt to print a debugger
260  * backtrace.  We also run pending interrupts in order to allow a console
261  * break into DDB.
262  */
263 static
264 int
265 exponential_backoff(struct exponential_backoff *bo)
266 {
267 	sysclock_t count;
268 	int backoff;
269 
270 #ifdef _RDTSC_SUPPORTED_
271 	if (cpu_feature & CPUID_TSC) {
272 		backoff =
273 		(((u_long)rdtsc() ^ (((u_long)curthread) >> 5)) &
274 		 (bo->backoff - 1)) + BACKOFF_INITIAL;
275 	} else
276 #endif
277 		backoff = bo->backoff;
278 	logspin_backoff(bo->mtx, bo->backoff, curthread, backoff);
279 
280 	/*
281 	 * Quick backoff
282 	 */
283 	for (; backoff; --backoff)
284 		cpu_pause();
285 	if (bo->backoff < spinlocks_backoff_limit) {
286 		bo->backoff <<= 1;
287 		return (FALSE);
288 	} else {
289 		bo->backoff = BACKOFF_INITIAL;
290 	}
291 
292 	logspin(bofail, bo->mtx, 'u');
293 
294 	/*
295 	 * Indefinite
296 	 */
297 	++spinlocks_contested2;
298 	cpu_spinlock_contested();
299 	if (bo->nsec == 0) {
300 		bo->base = sys_cputimer->count();
301 		bo->nsec = 1;
302 	}
303 
304 	count = sys_cputimer->count();
305 	if (count - bo->base > sys_cputimer->freq) {
306 		kprintf("spin_lock: %p, indefinite wait!\n", bo->mtx);
307 		if (panicstr)
308 			return (TRUE);
309 #if defined(INVARIANTS)
310 		if (spin_lock_test_mode) {
311 			print_backtrace(-1);
312 			return (TRUE);
313 		}
314 #endif
315 		++bo->nsec;
316 #if defined(INVARIANTS)
317 		if (bo->nsec == 11)
318 			print_backtrace(-1);
319 #endif
320 		if (bo->nsec == 60)
321 			panic("spin_lock: %p, indefinite wait!\n", bo->mtx);
322 		bo->base = count;
323 	}
324 	return (FALSE);
325 }
326 
327 /*
328  * If INVARIANTS is enabled various spinlock timing tests can be run
329  * by setting debug.spin_lock_test:
330  *
331  *	1	Test the indefinite wait code
332  *	2	Time the best-case exclusive lock overhead (spin_test_count)
333  *	3	Time the best-case shared lock overhead (spin_test_count)
334  */
335 
336 #ifdef INVARIANTS
337 
338 static int spin_test_count = 10000000;
339 SYSCTL_INT(_debug, OID_AUTO, spin_test_count, CTLFLAG_RW, &spin_test_count, 0, "");
340 
341 static int
342 sysctl_spin_lock_test(SYSCTL_HANDLER_ARGS)
343 {
344         struct spinlock mtx;
345 	int error;
346 	int value = 0;
347 	int i;
348 
349 	if ((error = priv_check(curthread, PRIV_ROOT)) != 0)
350 		return (error);
351 	if ((error = SYSCTL_IN(req, &value, sizeof(value))) != 0)
352 		return (error);
353 
354 	/*
355 	 * Indefinite wait test
356 	 */
357 	if (value == 1) {
358 		spin_init(&mtx);
359 		spin_lock_wr(&mtx);	/* force an indefinite wait */
360 		spin_lock_test_mode = 1;
361 		spin_lock_wr(&mtx);
362 		spin_unlock_wr(&mtx);	/* Clean up the spinlock count */
363 		spin_unlock_wr(&mtx);
364 		spin_lock_test_mode = 0;
365 	}
366 
367 	/*
368 	 * Time best-case exclusive spinlocks
369 	 */
370 	if (value == 2) {
371 		globaldata_t gd = mycpu;
372 
373 		spin_init(&mtx);
374 		for (i = spin_test_count; i > 0; --i) {
375 		    spin_lock_wr_quick(gd, &mtx);
376 		    spin_unlock_wr_quick(gd, &mtx);
377 		}
378 	}
379 
380 	/*
381 	 * Time best-case shared spinlocks
382 	 */
383 	if (value == 3) {
384 		globaldata_t gd = mycpu;
385 
386 		spin_init(&mtx);
387 		for (i = spin_test_count; i > 0; --i) {
388 		    spin_lock_rd_quick(gd, &mtx);
389 		    spin_unlock_rd_quick(gd, &mtx);
390 		}
391 	}
392         return (0);
393 }
394 
395 SYSCTL_PROC(_debug, KERN_PROC_ALL, spin_lock_test, CTLFLAG_RW|CTLTYPE_INT,
396         0, 0, sysctl_spin_lock_test, "I", "Test spinlock wait code");
397 
398 #endif	/* INVARIANTS */
399 #endif	/* SMP */
400