Home | History | Annotate | Line # | Download | only in kern
kern_timeout.c revision 1.17.20.2
      1 /*	$NetBSD: kern_timeout.c,v 1.17.20.2 2006/10/20 19:24:29 ad Exp $	*/
      2 
      3 /*-
      4  * Copyright (c) 2003, 2006 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.
      9  *
     10  * Redistribution and use in source and binary forms, with or without
     11  * modification, are permitted provided that the following conditions
     12  * are met:
     13  * 1. Redistributions of source code must retain the above copyright
     14  *    notice, this list of conditions and the following disclaimer.
     15  * 2. Redistributions in binary form must reproduce the above copyright
     16  *    notice, this list of conditions and the following disclaimer in the
     17  *    documentation and/or other materials provided with the distribution.
     18  * 3. All advertising materials mentioning features or use of this software
     19  *    must display the following acknowledgement:
     20  *	This product includes software developed by the NetBSD
     21  *	Foundation, Inc. and its contributors.
     22  * 4. Neither the name of The NetBSD Foundation nor the names of its
     23  *    contributors may be used to endorse or promote products derived
     24  *    from this software without specific prior written permission.
     25  *
     26  * THIS SOFTWARE IS PROVIDED BY THE NETBSD FOUNDATION, INC. AND CONTRIBUTORS
     27  * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED
     28  * TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
     29  * PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL THE FOUNDATION OR CONTRIBUTORS
     30  * BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
     31  * CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
     32  * SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
     33  * INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
     34  * CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
     35  * ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
     36  * POSSIBILITY OF SUCH DAMAGE.
     37  */
     38 
     39 /*
     40  * Copyright (c) 2001 Thomas Nordin <nordin (at) openbsd.org>
     41  * Copyright (c) 2000-2001 Artur Grabowski <art (at) openbsd.org>
     42  * All rights reserved.
     43  *
     44  * Redistribution and use in source and binary forms, with or without
     45  * modification, are permitted provided that the following conditions
     46  * are met:
     47  *
     48  * 1. Redistributions of source code must retain the above copyright
     49  *    notice, this list of conditions and the following disclaimer.
     50  * 2. Redistributions in binary form must reproduce the above copyright
     51  *    notice, this list of conditions and the following disclaimer in the
     52  *    documentation and/or other materials provided with the distribution.
     53  * 3. The name of the author may not be used to endorse or promote products
     54  *    derived from this software without specific prior written permission.
     55  *
     56  * THIS SOFTWARE IS PROVIDED ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES,
     57  * INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY
     58  * AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL
     59  * THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
     60  * EXEMPLARY, OR CONSEQUENTIAL  DAMAGES (INCLUDING, BUT NOT LIMITED TO,
     61  * PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS;
     62  * OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY,
     63  * WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR
     64  * OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF
     65  * ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
     66  */
     67 
     68 #include <sys/cdefs.h>
     69 __KERNEL_RCSID(0, "$NetBSD: kern_timeout.c,v 1.17.20.2 2006/10/20 19:24:29 ad Exp $");
     70 
     71 /*
     72  * Adapted from OpenBSD: kern_timeout.c,v 1.15 2002/12/08 04:21:07 art Exp,
     73  * modified to match NetBSD's pre-existing callout API.
     74  */
     75 
     76 #include <sys/param.h>
     77 #include <sys/systm.h>
     78 #include <sys/kernel.h>
     79 #include <sys/lock.h>
     80 #include <sys/callout.h>
     81 #include <sys/mutex.h>
     82 
     83 #ifdef DDB
     84 #include <machine/db_machdep.h>
     85 #include <ddb/db_interface.h>
     86 #include <ddb/db_access.h>
     87 #include <ddb/db_sym.h>
     88 #include <ddb/db_output.h>
     89 #endif
     90 
     91 /*
     92  * Timeouts are kept in a hierarchical timing wheel. The c_time is the value
     93  * of the global variable "hardclock_ticks" when the timeout should be called.
     94  * There are four levels with 256 buckets each. See 'Scheme 7' in
     95  * "Hashed and Hierarchical Timing Wheels: Efficient Data Structures for
     96  * Implementing a Timer Facility" by George Varghese and Tony Lauck.
     97  */
     98 #define BUCKETS 1024
     99 #define WHEELSIZE 256
    100 #define WHEELMASK 255
    101 #define WHEELBITS 8
    102 
    103 static struct callout_circq timeout_wheel[BUCKETS];	/* Queues of timeouts */
    104 static struct callout_circq timeout_todo;		/* Worklist */
    105 
    106 #define MASKWHEEL(wheel, time) (((time) >> ((wheel)*WHEELBITS)) & WHEELMASK)
    107 
    108 #define BUCKET(rel, abs)						\
    109     (((rel) <= (1 << (2*WHEELBITS)))					\
    110     	? ((rel) <= (1 << WHEELBITS))					\
    111             ? &timeout_wheel[MASKWHEEL(0, (abs))]			\
    112             : &timeout_wheel[MASKWHEEL(1, (abs)) + WHEELSIZE]		\
    113         : ((rel) <= (1 << (3*WHEELBITS)))				\
    114             ? &timeout_wheel[MASKWHEEL(2, (abs)) + 2*WHEELSIZE]		\
    115             : &timeout_wheel[MASKWHEEL(3, (abs)) + 3*WHEELSIZE])
    116 
    117 #define MOVEBUCKET(wheel, time)						\
    118     CIRCQ_APPEND(&timeout_todo,						\
    119         &timeout_wheel[MASKWHEEL((wheel), (time)) + (wheel)*WHEELSIZE])
    120 
    121 /*
    122  * All wheels are locked with the same lock (which must also block out all
    123  * interrupts).
    124  */
    125 kmutex_t callout_mutex;
    126 
    127 /*
    128  * Circular queue definitions.
    129  */
    130 
    131 #define CIRCQ_INIT(list)						\
    132 do {									\
    133         (list)->cq_next_l = (list);					\
    134         (list)->cq_prev_l = (list);					\
    135 } while (/*CONSTCOND*/0)
    136 
    137 #define CIRCQ_INSERT(elem, list)					\
    138 do {									\
    139         (elem)->cq_prev_e = (list)->cq_prev_e;				\
    140         (elem)->cq_next_l = (list);					\
    141         (list)->cq_prev_l->cq_next_l = (elem);				\
    142         (list)->cq_prev_l = (elem);					\
    143 } while (/*CONSTCOND*/0)
    144 
    145 #define CIRCQ_APPEND(fst, snd)						\
    146 do {									\
    147         if (!CIRCQ_EMPTY(snd)) {					\
    148                 (fst)->cq_prev_l->cq_next_l = (snd)->cq_next_l;		\
    149                 (snd)->cq_next_l->cq_prev_l = (fst)->cq_prev_l;		\
    150                 (snd)->cq_prev_l->cq_next_l = (fst);			\
    151                 (fst)->cq_prev_l = (snd)->cq_prev_l;			\
    152                 CIRCQ_INIT(snd);					\
    153         }								\
    154 } while (/*CONSTCOND*/0)
    155 
    156 #define CIRCQ_REMOVE(elem)						\
    157 do {									\
    158         (elem)->cq_next_l->cq_prev_e = (elem)->cq_prev_e;		\
    159         (elem)->cq_prev_l->cq_next_e = (elem)->cq_next_e;		\
    160 } while (/*CONSTCOND*/0)
    161 
    162 #define CIRCQ_FIRST(list)	((list)->cq_next_e)
    163 #define CIRCQ_NEXT(elem)	((elem)->cq_next_e)
    164 #define CIRCQ_LAST(elem,list)	((elem)->cq_next_l == (list))
    165 #define CIRCQ_EMPTY(list)	((list)->cq_next_l == (list))
    166 
    167 /*
    168  * Some of the "math" in here is a bit tricky.
    169  *
    170  * We have to beware of wrapping ints.
    171  * We use the fact that any element added to the queue must be added with a
    172  * positive time. That means that any element `to' on the queue cannot be
    173  * scheduled to timeout further in time than INT_MAX, but c->c_time can
    174  * be positive or negative so comparing it with anything is dangerous.
    175  * The only way we can use the c->c_time value in any predictable way
    176  * is when we calculate how far in the future `to' will timeout -
    177  * "c->c_time - hardclock_ticks". The result will always be positive for
    178  * future timeouts and 0 or negative for due timeouts.
    179  */
    180 
    181 #ifdef CALLOUT_EVENT_COUNTERS
    182 static struct evcnt callout_ev_late;
    183 #endif
    184 
    185 /*
    186  * callout_barrier:
    187  *
    188  *	If the callout is running on another CPU, busy wait until it
    189  *	completes.
    190  */
    191 static inline void
    192 callout_barrier(struct callout *c)
    193 {
    194 #ifdef MULTIPROCESSOR
    195 	void *oncpu;
    196 
    197 	LOCK_ASSERT(mutex_owned(&callout_mutex));
    198 
    199 	/*
    200 	 * The callout may have already been dispatched to run on the
    201 	 * current CPU.  It's possible for us to arrive here before it
    202 	 * actually runs because the SPL is dropped from IPL_SCHED in
    203 	 * softclock(), and IPL_SOFTCLOCK is low priority. We can't deal
    204 	 * with that race easily, so for now the caller must deal with
    205 	 * it.
    206 	 */
    207 	oncpu = c->c_oncpu;
    208 	if (oncpu != NULL && oncpu != curcpu()) {
    209 		mutex_exit(&callout_mutex);
    210 		while (c->c_oncpu != NULL)
    211 			;
    212 		mutex_enter(&callout_mutex);
    213 	}
    214 #endif
    215 }
    216 
    217 /*
    218  * callout_startup:
    219  *
    220  *	Initialize the callout facility, called at system startup time.
    221  */
    222 void
    223 callout_startup(void)
    224 {
    225 	int b;
    226 
    227 	CIRCQ_INIT(&timeout_todo);
    228 	for (b = 0; b < BUCKETS; b++)
    229 		CIRCQ_INIT(&timeout_wheel[b]);
    230 	mutex_init(&callout_mutex, MUTEX_SPIN, IPL_SCHED);
    231 
    232 #ifdef CALLOUT_EVENT_COUNTERS
    233 	evcnt_attach_dynamic(&callout_ev_late, EVCNT_TYPE_MISC,
    234 	    NULL, "callout", "late");
    235 #endif
    236 }
    237 
    238 /*
    239  * callout_init:
    240  *
    241  *	Initialize a callout structure.
    242  */
    243 void
    244 callout_init(struct callout *c)
    245 {
    246 
    247 	memset(c, 0, sizeof(*c));
    248 }
    249 
    250 /*
    251  * callout_reset:
    252  *
    253  *	Reset a callout structure with a new function and argument, and
    254  *	schedule it to run.
    255  */
    256 void
    257 callout_reset(struct callout *c, int to_ticks, void (*func)(void *), void *arg)
    258 {
    259 	int old_time;
    260 
    261 	KASSERT(to_ticks >= 0);
    262 
    263 	mutex_enter(&callout_mutex);
    264 
    265 	callout_barrier(c);
    266 
    267 	/* Initialize the time here, it won't change. */
    268 	old_time = c->c_time;
    269 	c->c_time = to_ticks + hardclock_ticks;
    270 	c->c_flags &= ~(CALLOUT_FIRED|CALLOUT_INVOKING);
    271 
    272 	c->c_func = func;
    273 	c->c_arg = arg;
    274 
    275 	/*
    276 	 * If this timeout is already scheduled and now is moved
    277 	 * earlier, reschedule it now. Otherwise leave it in place
    278 	 * and let it be rescheduled later.
    279 	 */
    280 	if (callout_pending(c)) {
    281 		if (c->c_time - old_time < 0) {
    282 			CIRCQ_REMOVE(&c->c_list);
    283 			CIRCQ_INSERT(&c->c_list, &timeout_todo);
    284 		}
    285 	} else {
    286 		c->c_flags |= CALLOUT_PENDING;
    287 		CIRCQ_INSERT(&c->c_list, &timeout_todo);
    288 	}
    289 
    290 	mutex_exit(&callout_mutex);
    291 }
    292 
    293 /*
    294  * callout_schedule:
    295  *
    296  *	Schedule a callout to run.  The function and argument must
    297  *	already be set in the callout structure.
    298  */
    299 void
    300 callout_schedule(struct callout *c, int to_ticks)
    301 {
    302 	int old_time;
    303 
    304 	KASSERT(to_ticks >= 0);
    305 
    306 	mutex_enter(&callout_mutex);
    307 
    308 	callout_barrier(c);
    309 
    310 	/* Initialize the time here, it won't change. */
    311 	old_time = c->c_time;
    312 	c->c_time = to_ticks + hardclock_ticks;
    313 	c->c_flags &= ~(CALLOUT_FIRED|CALLOUT_INVOKING);
    314 
    315 	/*
    316 	 * If this timeout is already scheduled and now is moved
    317 	 * earlier, reschedule it now. Otherwise leave it in place
    318 	 * and let it be rescheduled later.
    319 	 */
    320 	if (callout_pending(c)) {
    321 		if (c->c_time - old_time < 0) {
    322 			CIRCQ_REMOVE(&c->c_list);
    323 			CIRCQ_INSERT(&c->c_list, &timeout_todo);
    324 		}
    325 	} else {
    326 		c->c_flags |= CALLOUT_PENDING;
    327 		CIRCQ_INSERT(&c->c_list, &timeout_todo);
    328 	}
    329 
    330 	mutex_exit(&callout_mutex);
    331 }
    332 
    333 /*
    334  * callout_stop:
    335  *
    336  *	Cancel a pending callout.
    337  */
    338 void
    339 callout_stop(struct callout *c)
    340 {
    341 
    342 	mutex_enter(&callout_mutex);
    343 
    344 	callout_barrier(c);
    345 
    346 	if (callout_pending(c))
    347 		CIRCQ_REMOVE(&c->c_list);
    348 
    349 	c->c_flags &= ~(CALLOUT_PENDING|CALLOUT_FIRED);
    350 
    351 	mutex_exit(&callout_mutex);
    352 }
    353 
    354 /*
    355  * This is called from hardclock() once every tick.
    356  * We return !0 if we need to schedule a softclock.
    357  */
    358 int
    359 callout_hardclock(void)
    360 {
    361 	int needsoftclock;
    362 
    363 	mutex_enter(&callout_mutex);
    364 
    365 	MOVEBUCKET(0, hardclock_ticks);
    366 	if (MASKWHEEL(0, hardclock_ticks) == 0) {
    367 		MOVEBUCKET(1, hardclock_ticks);
    368 		if (MASKWHEEL(1, hardclock_ticks) == 0) {
    369 			MOVEBUCKET(2, hardclock_ticks);
    370 			if (MASKWHEEL(2, hardclock_ticks) == 0)
    371 				MOVEBUCKET(3, hardclock_ticks);
    372 		}
    373 	}
    374 
    375 	needsoftclock = !CIRCQ_EMPTY(&timeout_todo);
    376 	mutex_exit(&callout_mutex);
    377 
    378 	return needsoftclock;
    379 }
    380 
    381 /* ARGSUSED */
    382 void
    383 softclock(void *v)
    384 {
    385 	struct cpu_info *ci;
    386 	struct callout *c;
    387 	void (*func)(void *);
    388 	void *arg;
    389 
    390 	ci = curcpu();
    391 
    392 	mutex_enter(&callout_mutex);
    393 
    394 	while (!CIRCQ_EMPTY(&timeout_todo)) {
    395 		c = CIRCQ_FIRST(&timeout_todo);
    396 		CIRCQ_REMOVE(&c->c_list);
    397 
    398 		/* If due run it, otherwise insert it into the right bucket. */
    399 		if (c->c_time - hardclock_ticks > 0) {
    400 			CIRCQ_INSERT(&c->c_list,
    401 			    BUCKET((c->c_time - hardclock_ticks), c->c_time));
    402 		} else {
    403 #ifdef CALLOUT_EVENT_COUNTERS
    404 			if (c->c_time - hardclock_ticks < 0)
    405 				callout_ev_late.ev_count++;
    406 #endif
    407 			c->c_flags = (c->c_flags & ~CALLOUT_PENDING) |
    408 			    (CALLOUT_FIRED|CALLOUT_INVOKING);
    409 
    410 			func = c->c_func;
    411 			arg = c->c_arg;
    412 
    413 #ifdef MULTIPROCESSOR
    414 			c->c_oncpu = ci;
    415 #endif
    416 			mutex_exit(&callout_mutex);
    417 			(*func)(arg);
    418 			mutex_enter(&callout_mutex);
    419 #ifdef MULTIPROCESSOR
    420 			c->c_oncpu = NULL;
    421 #endif
    422 		}
    423 	}
    424 
    425 	mutex_exit(&callout_mutex);
    426 }
    427 
    428 #ifdef DDB
    429 static void
    430 db_show_callout_bucket(struct callout_circq *bucket)
    431 {
    432 	struct callout *c;
    433 	db_expr_t offset;
    434 	const char *name;
    435 	static char question[] = "?";
    436 
    437 	if (CIRCQ_EMPTY(bucket))
    438 		return;
    439 
    440 	for (c = CIRCQ_FIRST(bucket); /*nothing*/; c = CIRCQ_NEXT(&c->c_list)) {
    441 		db_find_sym_and_offset((db_addr_t)(intptr_t)c->c_func, &name,
    442 		    &offset);
    443 		name = name ? name : question;
    444 #ifdef _LP64
    445 #define	POINTER_WIDTH	"%16lx"
    446 #else
    447 #define	POINTER_WIDTH	"%8lx"
    448 #endif
    449 		db_printf("%9d %2d/%-4d " POINTER_WIDTH "  %s\n",
    450 		    c->c_time - hardclock_ticks,
    451 		    (int)((bucket - timeout_wheel) / WHEELSIZE),
    452 		    (int)(bucket - timeout_wheel), (u_long) c->c_arg, name);
    453 
    454 		if (CIRCQ_LAST(&c->c_list, bucket))
    455 			break;
    456 	}
    457 }
    458 
    459 void
    460 db_show_callout(db_expr_t addr, int haddr, db_expr_t count, const char *modif)
    461 {
    462 	int b;
    463 
    464 	db_printf("hardclock_ticks now: %d\n", hardclock_ticks);
    465 #ifdef _LP64
    466 	db_printf("    ticks  wheel               arg  func\n");
    467 #else
    468 	db_printf("    ticks  wheel       arg  func\n");
    469 #endif
    470 
    471 	/*
    472 	 * Don't lock the callwheel; all the other CPUs are paused
    473 	 * anyhow, and we might be called in a circumstance where
    474 	 * some other CPU was paused while holding the lock.
    475 	 */
    476 
    477 	db_show_callout_bucket(&timeout_todo);
    478 	for (b = 0; b < BUCKETS; b++)
    479 		db_show_callout_bucket(&timeout_wheel[b]);
    480 }
    481 #endif /* DDB */
    482