xref: /freebsd/sys/kern/kern_timeout.c (revision a3cf0ef5a295c885c895fabfd56470c0d1db322d)
1 /*-
2  * Copyright (c) 1982, 1986, 1991, 1993
3  *	The Regents of the University of California.  All rights reserved.
4  * (c) UNIX System Laboratories, Inc.
5  * All or some portions of this file are derived from material licensed
6  * to the University of California by American Telephone and Telegraph
7  * Co. or Unix System Laboratories, Inc. and are reproduced herein with
8  * the permission of UNIX System Laboratories, Inc.
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  * 4. Neither the name of the University nor the names of its contributors
19  *    may be used to endorse or promote products derived from this software
20  *    without specific prior written permission.
21  *
22  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
23  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
24  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
25  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
26  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
27  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
28  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
29  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
30  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
31  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
32  * SUCH DAMAGE.
33  *
34  *	From: @(#)kern_clock.c	8.5 (Berkeley) 1/21/94
35  */
36 
37 #include <sys/cdefs.h>
38 __FBSDID("$FreeBSD$");
39 
40 #include "opt_kdtrace.h"
41 
42 #include <sys/param.h>
43 #include <sys/systm.h>
44 #include <sys/bus.h>
45 #include <sys/callout.h>
46 #include <sys/condvar.h>
47 #include <sys/interrupt.h>
48 #include <sys/kernel.h>
49 #include <sys/ktr.h>
50 #include <sys/lock.h>
51 #include <sys/malloc.h>
52 #include <sys/mutex.h>
53 #include <sys/proc.h>
54 #include <sys/sdt.h>
55 #include <sys/sleepqueue.h>
56 #include <sys/sysctl.h>
57 #include <sys/smp.h>
58 
59 SDT_PROVIDER_DEFINE(callout_execute);
60 SDT_PROBE_DEFINE(callout_execute, kernel, , callout_start, callout-start);
61 SDT_PROBE_ARGTYPE(callout_execute, kernel, , callout_start, 0,
62     "struct callout *");
63 SDT_PROBE_DEFINE(callout_execute, kernel, , callout_end, callout-end);
64 SDT_PROBE_ARGTYPE(callout_execute, kernel, , callout_end, 0,
65     "struct callout *");
66 
67 static int avg_depth;
68 SYSCTL_INT(_debug, OID_AUTO, to_avg_depth, CTLFLAG_RD, &avg_depth, 0,
69     "Average number of items examined per softclock call. Units = 1/1000");
70 static int avg_gcalls;
71 SYSCTL_INT(_debug, OID_AUTO, to_avg_gcalls, CTLFLAG_RD, &avg_gcalls, 0,
72     "Average number of Giant callouts made per softclock call. Units = 1/1000");
73 static int avg_lockcalls;
74 SYSCTL_INT(_debug, OID_AUTO, to_avg_lockcalls, CTLFLAG_RD, &avg_lockcalls, 0,
75     "Average number of lock callouts made per softclock call. Units = 1/1000");
76 static int avg_mpcalls;
77 SYSCTL_INT(_debug, OID_AUTO, to_avg_mpcalls, CTLFLAG_RD, &avg_mpcalls, 0,
78     "Average number of MP callouts made per softclock call. Units = 1/1000");
79 /*
80  * TODO:
81  *	allocate more timeout table slots when table overflows.
82  */
83 int callwheelsize, callwheelbits, callwheelmask;
84 
85 /*
86  * There is one struct callout_cpu per cpu, holding all relevant
87  * state for the callout processing thread on the individual CPU.
88  * In particular:
89  *	cc_ticks is incremented once per tick in callout_cpu().
90  *	It tracks the global 'ticks' but in a way that the individual
91  *	threads should not worry about races in the order in which
92  *	hardclock() and hardclock_cpu() run on the various CPUs.
93  *	cc_softclock is advanced in callout_cpu() to point to the
94  *	first entry in cc_callwheel that may need handling. In turn,
95  *	a softclock() is scheduled so it can serve the various entries i
96  *	such that cc_softclock <= i <= cc_ticks .
97  *	XXX maybe cc_softclock and cc_ticks should be volatile ?
98  *
99  *	cc_ticks is also used in callout_reset_cpu() to determine
100  *	when the callout should be served.
101  */
102 struct callout_cpu {
103 	struct mtx		cc_lock;
104 	struct callout		*cc_callout;
105 	struct callout_tailq	*cc_callwheel;
106 	struct callout_list	cc_callfree;
107 	struct callout		*cc_next;
108 	struct callout		*cc_curr;
109 	void			*cc_cookie;
110 	int 			cc_ticks;
111 	int 			cc_softticks;
112 	int			cc_cancel;
113 	int			cc_waiting;
114 	int 			cc_firsttick;
115 };
116 
117 #ifdef SMP
118 struct callout_cpu cc_cpu[MAXCPU];
119 #define	CC_CPU(cpu)	(&cc_cpu[(cpu)])
120 #define	CC_SELF()	CC_CPU(PCPU_GET(cpuid))
121 #else
122 struct callout_cpu cc_cpu;
123 #define	CC_CPU(cpu)	&cc_cpu
124 #define	CC_SELF()	&cc_cpu
125 #endif
126 #define	CC_LOCK(cc)	mtx_lock_spin(&(cc)->cc_lock)
127 #define	CC_UNLOCK(cc)	mtx_unlock_spin(&(cc)->cc_lock)
128 
129 static int timeout_cpu;
130 void (*callout_new_inserted)(int cpu, int ticks) = NULL;
131 
132 MALLOC_DEFINE(M_CALLOUT, "callout", "Callout datastructures");
133 
134 /**
135  * Locked by cc_lock:
136  *   cc_curr         - If a callout is in progress, it is curr_callout.
137  *                     If curr_callout is non-NULL, threads waiting in
138  *                     callout_drain() will be woken up as soon as the
139  *                     relevant callout completes.
140  *   cc_cancel       - Changing to 1 with both callout_lock and c_lock held
141  *                     guarantees that the current callout will not run.
142  *                     The softclock() function sets this to 0 before it
143  *                     drops callout_lock to acquire c_lock, and it calls
144  *                     the handler only if curr_cancelled is still 0 after
145  *                     c_lock is successfully acquired.
146  *   cc_waiting      - If a thread is waiting in callout_drain(), then
147  *                     callout_wait is nonzero.  Set only when
148  *                     curr_callout is non-NULL.
149  */
150 
151 /*
152  * kern_timeout_callwheel_alloc() - kernel low level callwheel initialization
153  *
154  *	This code is called very early in the kernel initialization sequence,
155  *	and may be called more then once.
156  */
157 caddr_t
158 kern_timeout_callwheel_alloc(caddr_t v)
159 {
160 	struct callout_cpu *cc;
161 
162 	timeout_cpu = PCPU_GET(cpuid);
163 	cc = CC_CPU(timeout_cpu);
164 	/*
165 	 * Calculate callout wheel size
166 	 */
167 	for (callwheelsize = 1, callwheelbits = 0;
168 	     callwheelsize < ncallout;
169 	     callwheelsize <<= 1, ++callwheelbits)
170 		;
171 	callwheelmask = callwheelsize - 1;
172 
173 	cc->cc_callout = (struct callout *)v;
174 	v = (caddr_t)(cc->cc_callout + ncallout);
175 	cc->cc_callwheel = (struct callout_tailq *)v;
176 	v = (caddr_t)(cc->cc_callwheel + callwheelsize);
177 	return(v);
178 }
179 
180 static void
181 callout_cpu_init(struct callout_cpu *cc)
182 {
183 	struct callout *c;
184 	int i;
185 
186 	mtx_init(&cc->cc_lock, "callout", NULL, MTX_SPIN | MTX_RECURSE);
187 	SLIST_INIT(&cc->cc_callfree);
188 	for (i = 0; i < callwheelsize; i++) {
189 		TAILQ_INIT(&cc->cc_callwheel[i]);
190 	}
191 	if (cc->cc_callout == NULL)
192 		return;
193 	for (i = 0; i < ncallout; i++) {
194 		c = &cc->cc_callout[i];
195 		callout_init(c, 0);
196 		c->c_flags = CALLOUT_LOCAL_ALLOC;
197 		SLIST_INSERT_HEAD(&cc->cc_callfree, c, c_links.sle);
198 	}
199 }
200 
201 /*
202  * kern_timeout_callwheel_init() - initialize previously reserved callwheel
203  *				   space.
204  *
205  *	This code is called just once, after the space reserved for the
206  *	callout wheel has been finalized.
207  */
208 void
209 kern_timeout_callwheel_init(void)
210 {
211 	callout_cpu_init(CC_CPU(timeout_cpu));
212 }
213 
214 /*
215  * Start standard softclock thread.
216  */
217 static void
218 start_softclock(void *dummy)
219 {
220 	struct callout_cpu *cc;
221 #ifdef SMP
222 	int cpu;
223 #endif
224 
225 	cc = CC_CPU(timeout_cpu);
226 	if (swi_add(&clk_intr_event, "clock", softclock, cc, SWI_CLOCK,
227 	    INTR_MPSAFE, &cc->cc_cookie))
228 		panic("died while creating standard software ithreads");
229 #ifdef SMP
230 	CPU_FOREACH(cpu) {
231 		if (cpu == timeout_cpu)
232 			continue;
233 		cc = CC_CPU(cpu);
234 		if (swi_add(NULL, "clock", softclock, cc, SWI_CLOCK,
235 		    INTR_MPSAFE, &cc->cc_cookie))
236 			panic("died while creating standard software ithreads");
237 		cc->cc_callout = NULL;	/* Only cpu0 handles timeout(). */
238 		cc->cc_callwheel = malloc(
239 		    sizeof(struct callout_tailq) * callwheelsize, M_CALLOUT,
240 		    M_WAITOK);
241 		callout_cpu_init(cc);
242 	}
243 #endif
244 }
245 
246 SYSINIT(start_softclock, SI_SUB_SOFTINTR, SI_ORDER_FIRST, start_softclock, NULL);
247 
248 void
249 callout_tick(void)
250 {
251 	struct callout_cpu *cc;
252 	int need_softclock;
253 	int bucket;
254 
255 	/*
256 	 * Process callouts at a very low cpu priority, so we don't keep the
257 	 * relatively high clock interrupt priority any longer than necessary.
258 	 */
259 	need_softclock = 0;
260 	cc = CC_SELF();
261 	mtx_lock_spin_flags(&cc->cc_lock, MTX_QUIET);
262 	cc->cc_firsttick = cc->cc_ticks = ticks;
263 	for (; (cc->cc_softticks - cc->cc_ticks) <= 0; cc->cc_softticks++) {
264 		bucket = cc->cc_softticks & callwheelmask;
265 		if (!TAILQ_EMPTY(&cc->cc_callwheel[bucket])) {
266 			need_softclock = 1;
267 			break;
268 		}
269 	}
270 	mtx_unlock_spin_flags(&cc->cc_lock, MTX_QUIET);
271 	/*
272 	 * swi_sched acquires the thread lock, so we don't want to call it
273 	 * with cc_lock held; incorrect locking order.
274 	 */
275 	if (need_softclock)
276 		swi_sched(cc->cc_cookie, 0);
277 }
278 
279 int
280 callout_tickstofirst(int limit)
281 {
282 	struct callout_cpu *cc;
283 	struct callout *c;
284 	struct callout_tailq *sc;
285 	int curticks;
286 	int skip = 1;
287 
288 	cc = CC_SELF();
289 	mtx_lock_spin_flags(&cc->cc_lock, MTX_QUIET);
290 	curticks = cc->cc_ticks;
291 	while( skip < ncallout && skip < limit ) {
292 		sc = &cc->cc_callwheel[ (curticks+skip) & callwheelmask ];
293 		/* search scanning ticks */
294 		TAILQ_FOREACH( c, sc, c_links.tqe ){
295 			if (c->c_time - curticks <= ncallout)
296 				goto out;
297 		}
298 		skip++;
299 	}
300 out:
301 	cc->cc_firsttick = curticks + skip;
302 	mtx_unlock_spin_flags(&cc->cc_lock, MTX_QUIET);
303 	return (skip);
304 }
305 
306 static struct callout_cpu *
307 callout_lock(struct callout *c)
308 {
309 	struct callout_cpu *cc;
310 	int cpu;
311 
312 	for (;;) {
313 		cpu = c->c_cpu;
314 		cc = CC_CPU(cpu);
315 		CC_LOCK(cc);
316 		if (cpu == c->c_cpu)
317 			break;
318 		CC_UNLOCK(cc);
319 	}
320 	return (cc);
321 }
322 
323 /*
324  * The callout mechanism is based on the work of Adam M. Costello and
325  * George Varghese, published in a technical report entitled "Redesigning
326  * the BSD Callout and Timer Facilities" and modified slightly for inclusion
327  * in FreeBSD by Justin T. Gibbs.  The original work on the data structures
328  * used in this implementation was published by G. Varghese and T. Lauck in
329  * the paper "Hashed and Hierarchical Timing Wheels: Data Structures for
330  * the Efficient Implementation of a Timer Facility" in the Proceedings of
331  * the 11th ACM Annual Symposium on Operating Systems Principles,
332  * Austin, Texas Nov 1987.
333  */
334 
335 /*
336  * Software (low priority) clock interrupt.
337  * Run periodic events from timeout queue.
338  */
339 void
340 softclock(void *arg)
341 {
342 	struct callout_cpu *cc;
343 	struct callout *c;
344 	struct callout_tailq *bucket;
345 	int curticks;
346 	int steps;	/* #steps since we last allowed interrupts */
347 	int depth;
348 	int mpcalls;
349 	int lockcalls;
350 	int gcalls;
351 #ifdef DIAGNOSTIC
352 	struct bintime bt1, bt2;
353 	struct timespec ts2;
354 	static uint64_t maxdt = 36893488147419102LL;	/* 2 msec */
355 	static timeout_t *lastfunc;
356 #endif
357 
358 #ifndef MAX_SOFTCLOCK_STEPS
359 #define MAX_SOFTCLOCK_STEPS 100 /* Maximum allowed value of steps. */
360 #endif /* MAX_SOFTCLOCK_STEPS */
361 
362 	mpcalls = 0;
363 	lockcalls = 0;
364 	gcalls = 0;
365 	depth = 0;
366 	steps = 0;
367 	cc = (struct callout_cpu *)arg;
368 	CC_LOCK(cc);
369 	while (cc->cc_softticks - 1 != cc->cc_ticks) {
370 		/*
371 		 * cc_softticks may be modified by hard clock, so cache
372 		 * it while we work on a given bucket.
373 		 */
374 		curticks = cc->cc_softticks;
375 		cc->cc_softticks++;
376 		bucket = &cc->cc_callwheel[curticks & callwheelmask];
377 		c = TAILQ_FIRST(bucket);
378 		while (c) {
379 			depth++;
380 			if (c->c_time != curticks) {
381 				c = TAILQ_NEXT(c, c_links.tqe);
382 				++steps;
383 				if (steps >= MAX_SOFTCLOCK_STEPS) {
384 					cc->cc_next = c;
385 					/* Give interrupts a chance. */
386 					CC_UNLOCK(cc);
387 					;	/* nothing */
388 					CC_LOCK(cc);
389 					c = cc->cc_next;
390 					steps = 0;
391 				}
392 			} else {
393 				void (*c_func)(void *);
394 				void *c_arg;
395 				struct lock_class *class;
396 				struct lock_object *c_lock;
397 				int c_flags, sharedlock;
398 
399 				cc->cc_next = TAILQ_NEXT(c, c_links.tqe);
400 				TAILQ_REMOVE(bucket, c, c_links.tqe);
401 				class = (c->c_lock != NULL) ?
402 				    LOCK_CLASS(c->c_lock) : NULL;
403 				sharedlock = (c->c_flags & CALLOUT_SHAREDLOCK) ?
404 				    0 : 1;
405 				c_lock = c->c_lock;
406 				c_func = c->c_func;
407 				c_arg = c->c_arg;
408 				c_flags = c->c_flags;
409 				if (c->c_flags & CALLOUT_LOCAL_ALLOC) {
410 					c->c_flags = CALLOUT_LOCAL_ALLOC;
411 				} else {
412 					c->c_flags =
413 					    (c->c_flags & ~CALLOUT_PENDING);
414 				}
415 				cc->cc_curr = c;
416 				cc->cc_cancel = 0;
417 				CC_UNLOCK(cc);
418 				if (c_lock != NULL) {
419 					class->lc_lock(c_lock, sharedlock);
420 					/*
421 					 * The callout may have been cancelled
422 					 * while we switched locks.
423 					 */
424 					if (cc->cc_cancel) {
425 						class->lc_unlock(c_lock);
426 						goto skip;
427 					}
428 					/* The callout cannot be stopped now. */
429 					cc->cc_cancel = 1;
430 
431 					if (c_lock == &Giant.lock_object) {
432 						gcalls++;
433 						CTR3(KTR_CALLOUT,
434 						    "callout %p func %p arg %p",
435 						    c, c_func, c_arg);
436 					} else {
437 						lockcalls++;
438 						CTR3(KTR_CALLOUT, "callout lock"
439 						    " %p func %p arg %p",
440 						    c, c_func, c_arg);
441 					}
442 				} else {
443 					mpcalls++;
444 					CTR3(KTR_CALLOUT,
445 					    "callout mpsafe %p func %p arg %p",
446 					    c, c_func, c_arg);
447 				}
448 #ifdef DIAGNOSTIC
449 				binuptime(&bt1);
450 #endif
451 				THREAD_NO_SLEEPING();
452 				SDT_PROBE(callout_execute, kernel, ,
453 				    callout_start, c, 0, 0, 0, 0);
454 				c_func(c_arg);
455 				SDT_PROBE(callout_execute, kernel, ,
456 				    callout_end, c, 0, 0, 0, 0);
457 				THREAD_SLEEPING_OK();
458 #ifdef DIAGNOSTIC
459 				binuptime(&bt2);
460 				bintime_sub(&bt2, &bt1);
461 				if (bt2.frac > maxdt) {
462 					if (lastfunc != c_func ||
463 					    bt2.frac > maxdt * 2) {
464 						bintime2timespec(&bt2, &ts2);
465 						printf(
466 			"Expensive timeout(9) function: %p(%p) %jd.%09ld s\n",
467 						    c_func, c_arg,
468 						    (intmax_t)ts2.tv_sec,
469 						    ts2.tv_nsec);
470 					}
471 					maxdt = bt2.frac;
472 					lastfunc = c_func;
473 				}
474 #endif
475 				CTR1(KTR_CALLOUT, "callout %p finished", c);
476 				if ((c_flags & CALLOUT_RETURNUNLOCKED) == 0)
477 					class->lc_unlock(c_lock);
478 			skip:
479 				CC_LOCK(cc);
480 				/*
481 				 * If the current callout is locally
482 				 * allocated (from timeout(9))
483 				 * then put it on the freelist.
484 				 *
485 				 * Note: we need to check the cached
486 				 * copy of c_flags because if it was not
487 				 * local, then it's not safe to deref the
488 				 * callout pointer.
489 				 */
490 				if (c_flags & CALLOUT_LOCAL_ALLOC) {
491 					KASSERT(c->c_flags ==
492 					    CALLOUT_LOCAL_ALLOC,
493 					    ("corrupted callout"));
494 					c->c_func = NULL;
495 					SLIST_INSERT_HEAD(&cc->cc_callfree, c,
496 					    c_links.sle);
497 				}
498 				cc->cc_curr = NULL;
499 				if (cc->cc_waiting) {
500 					/*
501 					 * There is someone waiting
502 					 * for the callout to complete.
503 					 */
504 					cc->cc_waiting = 0;
505 					CC_UNLOCK(cc);
506 					wakeup(&cc->cc_waiting);
507 					CC_LOCK(cc);
508 				}
509 				steps = 0;
510 				c = cc->cc_next;
511 			}
512 		}
513 	}
514 	avg_depth += (depth * 1000 - avg_depth) >> 8;
515 	avg_mpcalls += (mpcalls * 1000 - avg_mpcalls) >> 8;
516 	avg_lockcalls += (lockcalls * 1000 - avg_lockcalls) >> 8;
517 	avg_gcalls += (gcalls * 1000 - avg_gcalls) >> 8;
518 	cc->cc_next = NULL;
519 	CC_UNLOCK(cc);
520 }
521 
522 /*
523  * timeout --
524  *	Execute a function after a specified length of time.
525  *
526  * untimeout --
527  *	Cancel previous timeout function call.
528  *
529  * callout_handle_init --
530  *	Initialize a handle so that using it with untimeout is benign.
531  *
532  *	See AT&T BCI Driver Reference Manual for specification.  This
533  *	implementation differs from that one in that although an
534  *	identification value is returned from timeout, the original
535  *	arguments to timeout as well as the identifier are used to
536  *	identify entries for untimeout.
537  */
538 struct callout_handle
539 timeout(ftn, arg, to_ticks)
540 	timeout_t *ftn;
541 	void *arg;
542 	int to_ticks;
543 {
544 	struct callout_cpu *cc;
545 	struct callout *new;
546 	struct callout_handle handle;
547 
548 	cc = CC_CPU(timeout_cpu);
549 	CC_LOCK(cc);
550 	/* Fill in the next free callout structure. */
551 	new = SLIST_FIRST(&cc->cc_callfree);
552 	if (new == NULL)
553 		/* XXX Attempt to malloc first */
554 		panic("timeout table full");
555 	SLIST_REMOVE_HEAD(&cc->cc_callfree, c_links.sle);
556 	callout_reset(new, to_ticks, ftn, arg);
557 	handle.callout = new;
558 	CC_UNLOCK(cc);
559 
560 	return (handle);
561 }
562 
563 void
564 untimeout(ftn, arg, handle)
565 	timeout_t *ftn;
566 	void *arg;
567 	struct callout_handle handle;
568 {
569 	struct callout_cpu *cc;
570 
571 	/*
572 	 * Check for a handle that was initialized
573 	 * by callout_handle_init, but never used
574 	 * for a real timeout.
575 	 */
576 	if (handle.callout == NULL)
577 		return;
578 
579 	cc = callout_lock(handle.callout);
580 	if (handle.callout->c_func == ftn && handle.callout->c_arg == arg)
581 		callout_stop(handle.callout);
582 	CC_UNLOCK(cc);
583 }
584 
585 void
586 callout_handle_init(struct callout_handle *handle)
587 {
588 	handle->callout = NULL;
589 }
590 
591 /*
592  * New interface; clients allocate their own callout structures.
593  *
594  * callout_reset() - establish or change a timeout
595  * callout_stop() - disestablish a timeout
596  * callout_init() - initialize a callout structure so that it can
597  *	safely be passed to callout_reset() and callout_stop()
598  *
599  * <sys/callout.h> defines three convenience macros:
600  *
601  * callout_active() - returns truth if callout has not been stopped,
602  *	drained, or deactivated since the last time the callout was
603  *	reset.
604  * callout_pending() - returns truth if callout is still waiting for timeout
605  * callout_deactivate() - marks the callout as having been serviced
606  */
607 int
608 callout_reset_on(struct callout *c, int to_ticks, void (*ftn)(void *),
609     void *arg, int cpu)
610 {
611 	struct callout_cpu *cc;
612 	int cancelled = 0;
613 
614 	/*
615 	 * Don't allow migration of pre-allocated callouts lest they
616 	 * become unbalanced.
617 	 */
618 	if (c->c_flags & CALLOUT_LOCAL_ALLOC)
619 		cpu = c->c_cpu;
620 retry:
621 	cc = callout_lock(c);
622 	if (cc->cc_curr == c) {
623 		/*
624 		 * We're being asked to reschedule a callout which is
625 		 * currently in progress.  If there is a lock then we
626 		 * can cancel the callout if it has not really started.
627 		 */
628 		if (c->c_lock != NULL && !cc->cc_cancel)
629 			cancelled = cc->cc_cancel = 1;
630 		if (cc->cc_waiting) {
631 			/*
632 			 * Someone has called callout_drain to kill this
633 			 * callout.  Don't reschedule.
634 			 */
635 			CTR4(KTR_CALLOUT, "%s %p func %p arg %p",
636 			    cancelled ? "cancelled" : "failed to cancel",
637 			    c, c->c_func, c->c_arg);
638 			CC_UNLOCK(cc);
639 			return (cancelled);
640 		}
641 	}
642 	if (c->c_flags & CALLOUT_PENDING) {
643 		if (cc->cc_next == c) {
644 			cc->cc_next = TAILQ_NEXT(c, c_links.tqe);
645 		}
646 		TAILQ_REMOVE(&cc->cc_callwheel[c->c_time & callwheelmask], c,
647 		    c_links.tqe);
648 
649 		cancelled = 1;
650 		c->c_flags &= ~(CALLOUT_ACTIVE | CALLOUT_PENDING);
651 	}
652 	/*
653 	 * If the lock must migrate we have to check the state again as
654 	 * we can't hold both the new and old locks simultaneously.
655 	 */
656 	if (c->c_cpu != cpu) {
657 		c->c_cpu = cpu;
658 		CC_UNLOCK(cc);
659 		goto retry;
660 	}
661 
662 	if (to_ticks <= 0)
663 		to_ticks = 1;
664 
665 	c->c_arg = arg;
666 	c->c_flags |= (CALLOUT_ACTIVE | CALLOUT_PENDING);
667 	c->c_func = ftn;
668 	c->c_time = ticks + to_ticks;
669 	TAILQ_INSERT_TAIL(&cc->cc_callwheel[c->c_time & callwheelmask],
670 			  c, c_links.tqe);
671 	if ((c->c_time - cc->cc_firsttick) < 0 &&
672 	    callout_new_inserted != NULL) {
673 		cc->cc_firsttick = c->c_time;
674 		(*callout_new_inserted)(cpu,
675 		    to_ticks + (ticks - cc->cc_ticks));
676 	}
677 	CTR5(KTR_CALLOUT, "%sscheduled %p func %p arg %p in %d",
678 	    cancelled ? "re" : "", c, c->c_func, c->c_arg, to_ticks);
679 	CC_UNLOCK(cc);
680 
681 	return (cancelled);
682 }
683 
684 /*
685  * Common idioms that can be optimized in the future.
686  */
687 int
688 callout_schedule_on(struct callout *c, int to_ticks, int cpu)
689 {
690 	return callout_reset_on(c, to_ticks, c->c_func, c->c_arg, cpu);
691 }
692 
693 int
694 callout_schedule(struct callout *c, int to_ticks)
695 {
696 	return callout_reset_on(c, to_ticks, c->c_func, c->c_arg, c->c_cpu);
697 }
698 
699 int
700 _callout_stop_safe(c, safe)
701 	struct	callout *c;
702 	int	safe;
703 {
704 	struct callout_cpu *cc;
705 	struct lock_class *class;
706 	int use_lock, sq_locked;
707 
708 	/*
709 	 * Some old subsystems don't hold Giant while running a callout_stop(),
710 	 * so just discard this check for the moment.
711 	 */
712 	if (!safe && c->c_lock != NULL) {
713 		if (c->c_lock == &Giant.lock_object)
714 			use_lock = mtx_owned(&Giant);
715 		else {
716 			use_lock = 1;
717 			class = LOCK_CLASS(c->c_lock);
718 			class->lc_assert(c->c_lock, LA_XLOCKED);
719 		}
720 	} else
721 		use_lock = 0;
722 
723 	sq_locked = 0;
724 again:
725 	cc = callout_lock(c);
726 	/*
727 	 * If the callout isn't pending, it's not on the queue, so
728 	 * don't attempt to remove it from the queue.  We can try to
729 	 * stop it by other means however.
730 	 */
731 	if (!(c->c_flags & CALLOUT_PENDING)) {
732 		c->c_flags &= ~CALLOUT_ACTIVE;
733 
734 		/*
735 		 * If it wasn't on the queue and it isn't the current
736 		 * callout, then we can't stop it, so just bail.
737 		 */
738 		if (cc->cc_curr != c) {
739 			CTR3(KTR_CALLOUT, "failed to stop %p func %p arg %p",
740 			    c, c->c_func, c->c_arg);
741 			CC_UNLOCK(cc);
742 			if (sq_locked)
743 				sleepq_release(&cc->cc_waiting);
744 			return (0);
745 		}
746 
747 		if (safe) {
748 			/*
749 			 * The current callout is running (or just
750 			 * about to run) and blocking is allowed, so
751 			 * just wait for the current invocation to
752 			 * finish.
753 			 */
754 			while (cc->cc_curr == c) {
755 
756 				/*
757 				 * Use direct calls to sleepqueue interface
758 				 * instead of cv/msleep in order to avoid
759 				 * a LOR between cc_lock and sleepqueue
760 				 * chain spinlocks.  This piece of code
761 				 * emulates a msleep_spin() call actually.
762 				 *
763 				 * If we already have the sleepqueue chain
764 				 * locked, then we can safely block.  If we
765 				 * don't already have it locked, however,
766 				 * we have to drop the cc_lock to lock
767 				 * it.  This opens several races, so we
768 				 * restart at the beginning once we have
769 				 * both locks.  If nothing has changed, then
770 				 * we will end up back here with sq_locked
771 				 * set.
772 				 */
773 				if (!sq_locked) {
774 					CC_UNLOCK(cc);
775 					sleepq_lock(&cc->cc_waiting);
776 					sq_locked = 1;
777 					goto again;
778 				}
779 				cc->cc_waiting = 1;
780 				DROP_GIANT();
781 				CC_UNLOCK(cc);
782 				sleepq_add(&cc->cc_waiting,
783 				    &cc->cc_lock.lock_object, "codrain",
784 				    SLEEPQ_SLEEP, 0);
785 				sleepq_wait(&cc->cc_waiting, 0);
786 				sq_locked = 0;
787 
788 				/* Reacquire locks previously released. */
789 				PICKUP_GIANT();
790 				CC_LOCK(cc);
791 			}
792 		} else if (use_lock && !cc->cc_cancel) {
793 			/*
794 			 * The current callout is waiting for its
795 			 * lock which we hold.  Cancel the callout
796 			 * and return.  After our caller drops the
797 			 * lock, the callout will be skipped in
798 			 * softclock().
799 			 */
800 			cc->cc_cancel = 1;
801 			CTR3(KTR_CALLOUT, "cancelled %p func %p arg %p",
802 			    c, c->c_func, c->c_arg);
803 			CC_UNLOCK(cc);
804 			KASSERT(!sq_locked, ("sleepqueue chain locked"));
805 			return (1);
806 		}
807 		CTR3(KTR_CALLOUT, "failed to stop %p func %p arg %p",
808 		    c, c->c_func, c->c_arg);
809 		CC_UNLOCK(cc);
810 		KASSERT(!sq_locked, ("sleepqueue chain still locked"));
811 		return (0);
812 	}
813 	if (sq_locked)
814 		sleepq_release(&cc->cc_waiting);
815 
816 	c->c_flags &= ~(CALLOUT_ACTIVE | CALLOUT_PENDING);
817 
818 	if (cc->cc_next == c) {
819 		cc->cc_next = TAILQ_NEXT(c, c_links.tqe);
820 	}
821 	TAILQ_REMOVE(&cc->cc_callwheel[c->c_time & callwheelmask], c,
822 	    c_links.tqe);
823 
824 	CTR3(KTR_CALLOUT, "cancelled %p func %p arg %p",
825 	    c, c->c_func, c->c_arg);
826 
827 	if (c->c_flags & CALLOUT_LOCAL_ALLOC) {
828 		c->c_func = NULL;
829 		SLIST_INSERT_HEAD(&cc->cc_callfree, c, c_links.sle);
830 	}
831 	CC_UNLOCK(cc);
832 	return (1);
833 }
834 
835 void
836 callout_init(c, mpsafe)
837 	struct	callout *c;
838 	int mpsafe;
839 {
840 	bzero(c, sizeof *c);
841 	if (mpsafe) {
842 		c->c_lock = NULL;
843 		c->c_flags = CALLOUT_RETURNUNLOCKED;
844 	} else {
845 		c->c_lock = &Giant.lock_object;
846 		c->c_flags = 0;
847 	}
848 	c->c_cpu = timeout_cpu;
849 }
850 
851 void
852 _callout_init_lock(c, lock, flags)
853 	struct	callout *c;
854 	struct	lock_object *lock;
855 	int flags;
856 {
857 	bzero(c, sizeof *c);
858 	c->c_lock = lock;
859 	KASSERT((flags & ~(CALLOUT_RETURNUNLOCKED | CALLOUT_SHAREDLOCK)) == 0,
860 	    ("callout_init_lock: bad flags %d", flags));
861 	KASSERT(lock != NULL || (flags & CALLOUT_RETURNUNLOCKED) == 0,
862 	    ("callout_init_lock: CALLOUT_RETURNUNLOCKED with no lock"));
863 	KASSERT(lock == NULL || !(LOCK_CLASS(lock)->lc_flags &
864 	    (LC_SPINLOCK | LC_SLEEPABLE)), ("%s: invalid lock class",
865 	    __func__));
866 	c->c_flags = flags & (CALLOUT_RETURNUNLOCKED | CALLOUT_SHAREDLOCK);
867 	c->c_cpu = timeout_cpu;
868 }
869 
870 #ifdef APM_FIXUP_CALLTODO
871 /*
872  * Adjust the kernel calltodo timeout list.  This routine is used after
873  * an APM resume to recalculate the calltodo timer list values with the
874  * number of hz's we have been sleeping.  The next hardclock() will detect
875  * that there are fired timers and run softclock() to execute them.
876  *
877  * Please note, I have not done an exhaustive analysis of what code this
878  * might break.  I am motivated to have my select()'s and alarm()'s that
879  * have expired during suspend firing upon resume so that the applications
880  * which set the timer can do the maintanence the timer was for as close
881  * as possible to the originally intended time.  Testing this code for a
882  * week showed that resuming from a suspend resulted in 22 to 25 timers
883  * firing, which seemed independant on whether the suspend was 2 hours or
884  * 2 days.  Your milage may vary.   - Ken Key <key@cs.utk.edu>
885  */
886 void
887 adjust_timeout_calltodo(time_change)
888     struct timeval *time_change;
889 {
890 	register struct callout *p;
891 	unsigned long delta_ticks;
892 
893 	/*
894 	 * How many ticks were we asleep?
895 	 * (stolen from tvtohz()).
896 	 */
897 
898 	/* Don't do anything */
899 	if (time_change->tv_sec < 0)
900 		return;
901 	else if (time_change->tv_sec <= LONG_MAX / 1000000)
902 		delta_ticks = (time_change->tv_sec * 1000000 +
903 			       time_change->tv_usec + (tick - 1)) / tick + 1;
904 	else if (time_change->tv_sec <= LONG_MAX / hz)
905 		delta_ticks = time_change->tv_sec * hz +
906 			      (time_change->tv_usec + (tick - 1)) / tick + 1;
907 	else
908 		delta_ticks = LONG_MAX;
909 
910 	if (delta_ticks > INT_MAX)
911 		delta_ticks = INT_MAX;
912 
913 	/*
914 	 * Now rip through the timer calltodo list looking for timers
915 	 * to expire.
916 	 */
917 
918 	/* don't collide with softclock() */
919 	CC_LOCK(cc);
920 	for (p = calltodo.c_next; p != NULL; p = p->c_next) {
921 		p->c_time -= delta_ticks;
922 
923 		/* Break if the timer had more time on it than delta_ticks */
924 		if (p->c_time > 0)
925 			break;
926 
927 		/* take back the ticks the timer didn't use (p->c_time <= 0) */
928 		delta_ticks = -p->c_time;
929 	}
930 	CC_UNLOCK(cc);
931 
932 	return;
933 }
934 #endif /* APM_FIXUP_CALLTODO */
935