xref: /freebsd/sys/kern/kern_timeout.c (revision 65b71c48d226f3ef3eb1c4452dccc678a7d2ac96)
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_callout_profiling.h"
41 #if defined(__arm__)
42 #include "opt_timer.h"
43 #endif
44 #include "opt_rss.h"
45 
46 #include <sys/param.h>
47 #include <sys/systm.h>
48 #include <sys/bus.h>
49 #include <sys/callout.h>
50 #include <sys/file.h>
51 #include <sys/interrupt.h>
52 #include <sys/kernel.h>
53 #include <sys/ktr.h>
54 #include <sys/lock.h>
55 #include <sys/malloc.h>
56 #include <sys/mutex.h>
57 #include <sys/proc.h>
58 #include <sys/sdt.h>
59 #include <sys/sleepqueue.h>
60 #include <sys/sysctl.h>
61 #include <sys/smp.h>
62 
63 #ifdef SMP
64 #include <machine/cpu.h>
65 #endif
66 
67 #ifndef NO_EVENTTIMERS
68 DPCPU_DECLARE(sbintime_t, hardclocktime);
69 #endif
70 
71 SDT_PROVIDER_DEFINE(callout_execute);
72 SDT_PROBE_DEFINE1(callout_execute, kernel, , callout__start,
73     "struct callout *");
74 SDT_PROBE_DEFINE1(callout_execute, kernel, , callout__end,
75     "struct callout *");
76 
77 #ifdef CALLOUT_PROFILING
78 static int avg_depth;
79 SYSCTL_INT(_debug, OID_AUTO, to_avg_depth, CTLFLAG_RD, &avg_depth, 0,
80     "Average number of items examined per softclock call. Units = 1/1000");
81 static int avg_gcalls;
82 SYSCTL_INT(_debug, OID_AUTO, to_avg_gcalls, CTLFLAG_RD, &avg_gcalls, 0,
83     "Average number of Giant callouts made per softclock call. Units = 1/1000");
84 static int avg_lockcalls;
85 SYSCTL_INT(_debug, OID_AUTO, to_avg_lockcalls, CTLFLAG_RD, &avg_lockcalls, 0,
86     "Average number of lock callouts made per softclock call. Units = 1/1000");
87 static int avg_mpcalls;
88 SYSCTL_INT(_debug, OID_AUTO, to_avg_mpcalls, CTLFLAG_RD, &avg_mpcalls, 0,
89     "Average number of MP callouts made per softclock call. Units = 1/1000");
90 static int avg_depth_dir;
91 SYSCTL_INT(_debug, OID_AUTO, to_avg_depth_dir, CTLFLAG_RD, &avg_depth_dir, 0,
92     "Average number of direct callouts examined per callout_process call. "
93     "Units = 1/1000");
94 static int avg_lockcalls_dir;
95 SYSCTL_INT(_debug, OID_AUTO, to_avg_lockcalls_dir, CTLFLAG_RD,
96     &avg_lockcalls_dir, 0, "Average number of lock direct callouts made per "
97     "callout_process call. Units = 1/1000");
98 static int avg_mpcalls_dir;
99 SYSCTL_INT(_debug, OID_AUTO, to_avg_mpcalls_dir, CTLFLAG_RD, &avg_mpcalls_dir,
100     0, "Average number of MP direct callouts made per callout_process call. "
101     "Units = 1/1000");
102 #endif
103 
104 static int ncallout;
105 SYSCTL_INT(_kern, OID_AUTO, ncallout, CTLFLAG_RDTUN | CTLFLAG_NOFETCH, &ncallout, 0,
106     "Number of entries in callwheel and size of timeout() preallocation");
107 
108 #ifdef	RSS
109 static int pin_default_swi = 1;
110 static int pin_pcpu_swi = 1;
111 #else
112 static int pin_default_swi = 0;
113 static int pin_pcpu_swi = 0;
114 #endif
115 
116 SYSCTL_INT(_kern, OID_AUTO, pin_default_swi, CTLFLAG_RDTUN | CTLFLAG_NOFETCH, &pin_default_swi,
117     0, "Pin the default (non-per-cpu) swi (shared with PCPU 0 swi)");
118 SYSCTL_INT(_kern, OID_AUTO, pin_pcpu_swi, CTLFLAG_RDTUN | CTLFLAG_NOFETCH, &pin_pcpu_swi,
119     0, "Pin the per-CPU swis (except PCPU 0, which is also default");
120 
121 /*
122  * TODO:
123  *	allocate more timeout table slots when table overflows.
124  */
125 u_int callwheelsize, callwheelmask;
126 
127 /*
128  * The callout cpu exec entities represent informations necessary for
129  * describing the state of callouts currently running on the CPU and the ones
130  * necessary for migrating callouts to the new callout cpu. In particular,
131  * the first entry of the array cc_exec_entity holds informations for callout
132  * running in SWI thread context, while the second one holds informations
133  * for callout running directly from hardware interrupt context.
134  * The cached informations are very important for deferring migration when
135  * the migrating callout is already running.
136  */
137 struct cc_exec {
138 	struct callout		*cc_curr;
139 #ifdef SMP
140 	void			(*ce_migration_func)(void *);
141 	void			*ce_migration_arg;
142 	int			ce_migration_cpu;
143 	sbintime_t		ce_migration_time;
144 	sbintime_t		ce_migration_prec;
145 #endif
146 	bool			cc_cancel;
147 	bool			cc_waiting;
148 };
149 
150 /*
151  * There is one struct callout_cpu per cpu, holding all relevant
152  * state for the callout processing thread on the individual CPU.
153  */
154 struct callout_cpu {
155 	struct mtx_padalign	cc_lock;
156 	struct cc_exec 		cc_exec_entity[2];
157 	struct callout		*cc_next;
158 	struct callout		*cc_callout;
159 	struct callout_list	*cc_callwheel;
160 	struct callout_tailq	cc_expireq;
161 	struct callout_slist	cc_callfree;
162 	sbintime_t		cc_firstevent;
163 	sbintime_t		cc_lastscan;
164 	void			*cc_cookie;
165 	u_int			cc_bucket;
166 	u_int			cc_inited;
167 	char			cc_ktr_event_name[20];
168 };
169 
170 #define	callout_migrating(c)	((c)->c_iflags & CALLOUT_DFRMIGRATION)
171 
172 #define	cc_exec_curr(cc, dir)		cc->cc_exec_entity[dir].cc_curr
173 #define	cc_exec_next(cc)		cc->cc_next
174 #define	cc_exec_cancel(cc, dir)		cc->cc_exec_entity[dir].cc_cancel
175 #define	cc_exec_waiting(cc, dir)	cc->cc_exec_entity[dir].cc_waiting
176 #ifdef SMP
177 #define	cc_migration_func(cc, dir)	cc->cc_exec_entity[dir].ce_migration_func
178 #define	cc_migration_arg(cc, dir)	cc->cc_exec_entity[dir].ce_migration_arg
179 #define	cc_migration_cpu(cc, dir)	cc->cc_exec_entity[dir].ce_migration_cpu
180 #define	cc_migration_time(cc, dir)	cc->cc_exec_entity[dir].ce_migration_time
181 #define	cc_migration_prec(cc, dir)	cc->cc_exec_entity[dir].ce_migration_prec
182 
183 struct callout_cpu cc_cpu[MAXCPU];
184 #define	CPUBLOCK	MAXCPU
185 #define	CC_CPU(cpu)	(&cc_cpu[(cpu)])
186 #define	CC_SELF()	CC_CPU(PCPU_GET(cpuid))
187 #else
188 struct callout_cpu cc_cpu;
189 #define	CC_CPU(cpu)	&cc_cpu
190 #define	CC_SELF()	&cc_cpu
191 #endif
192 #define	CC_LOCK(cc)	mtx_lock_spin(&(cc)->cc_lock)
193 #define	CC_UNLOCK(cc)	mtx_unlock_spin(&(cc)->cc_lock)
194 #define	CC_LOCK_ASSERT(cc)	mtx_assert(&(cc)->cc_lock, MA_OWNED)
195 
196 static int timeout_cpu;
197 
198 static void	callout_cpu_init(struct callout_cpu *cc, int cpu);
199 static void	softclock_call_cc(struct callout *c, struct callout_cpu *cc,
200 #ifdef CALLOUT_PROFILING
201 		    int *mpcalls, int *lockcalls, int *gcalls,
202 #endif
203 		    int direct);
204 
205 static MALLOC_DEFINE(M_CALLOUT, "callout", "Callout datastructures");
206 
207 /**
208  * Locked by cc_lock:
209  *   cc_curr         - If a callout is in progress, it is cc_curr.
210  *                     If cc_curr is non-NULL, threads waiting in
211  *                     callout_drain() will be woken up as soon as the
212  *                     relevant callout completes.
213  *   cc_cancel       - Changing to 1 with both callout_lock and cc_lock held
214  *                     guarantees that the current callout will not run.
215  *                     The softclock() function sets this to 0 before it
216  *                     drops callout_lock to acquire c_lock, and it calls
217  *                     the handler only if curr_cancelled is still 0 after
218  *                     cc_lock is successfully acquired.
219  *   cc_waiting      - If a thread is waiting in callout_drain(), then
220  *                     callout_wait is nonzero.  Set only when
221  *                     cc_curr is non-NULL.
222  */
223 
224 /*
225  * Resets the execution entity tied to a specific callout cpu.
226  */
227 static void
228 cc_cce_cleanup(struct callout_cpu *cc, int direct)
229 {
230 
231 	cc_exec_curr(cc, direct) = NULL;
232 	cc_exec_cancel(cc, direct) = false;
233 	cc_exec_waiting(cc, direct) = false;
234 #ifdef SMP
235 	cc_migration_cpu(cc, direct) = CPUBLOCK;
236 	cc_migration_time(cc, direct) = 0;
237 	cc_migration_prec(cc, direct) = 0;
238 	cc_migration_func(cc, direct) = NULL;
239 	cc_migration_arg(cc, direct) = NULL;
240 #endif
241 }
242 
243 /*
244  * Checks if migration is requested by a specific callout cpu.
245  */
246 static int
247 cc_cce_migrating(struct callout_cpu *cc, int direct)
248 {
249 
250 #ifdef SMP
251 	return (cc_migration_cpu(cc, direct) != CPUBLOCK);
252 #else
253 	return (0);
254 #endif
255 }
256 
257 /*
258  * Kernel low level callwheel initialization
259  * called on cpu0 during kernel startup.
260  */
261 static void
262 callout_callwheel_init(void *dummy)
263 {
264 	struct callout_cpu *cc;
265 
266 	/*
267 	 * Calculate the size of the callout wheel and the preallocated
268 	 * timeout() structures.
269 	 * XXX: Clip callout to result of previous function of maxusers
270 	 * maximum 384.  This is still huge, but acceptable.
271 	 */
272 	memset(CC_CPU(0), 0, sizeof(cc_cpu));
273 	ncallout = imin(16 + maxproc + maxfiles, 18508);
274 	TUNABLE_INT_FETCH("kern.ncallout", &ncallout);
275 
276 	/*
277 	 * Calculate callout wheel size, should be next power of two higher
278 	 * than 'ncallout'.
279 	 */
280 	callwheelsize = 1 << fls(ncallout);
281 	callwheelmask = callwheelsize - 1;
282 
283 	/*
284 	 * Fetch whether we're pinning the swi's or not.
285 	 */
286 	TUNABLE_INT_FETCH("kern.pin_default_swi", &pin_default_swi);
287 	TUNABLE_INT_FETCH("kern.pin_pcpu_swi", &pin_pcpu_swi);
288 
289 	/*
290 	 * Only cpu0 handles timeout(9) and receives a preallocation.
291 	 *
292 	 * XXX: Once all timeout(9) consumers are converted this can
293 	 * be removed.
294 	 */
295 	timeout_cpu = PCPU_GET(cpuid);
296 	cc = CC_CPU(timeout_cpu);
297 	cc->cc_callout = malloc(ncallout * sizeof(struct callout),
298 	    M_CALLOUT, M_WAITOK);
299 	callout_cpu_init(cc, timeout_cpu);
300 }
301 SYSINIT(callwheel_init, SI_SUB_CPU, SI_ORDER_ANY, callout_callwheel_init, NULL);
302 
303 /*
304  * Initialize the per-cpu callout structures.
305  */
306 static void
307 callout_cpu_init(struct callout_cpu *cc, int cpu)
308 {
309 	struct callout *c;
310 	int i;
311 
312 	mtx_init(&cc->cc_lock, "callout", NULL, MTX_SPIN | MTX_RECURSE);
313 	SLIST_INIT(&cc->cc_callfree);
314 	cc->cc_inited = 1;
315 	cc->cc_callwheel = malloc(sizeof(struct callout_list) * callwheelsize,
316 	    M_CALLOUT, M_WAITOK);
317 	for (i = 0; i < callwheelsize; i++)
318 		LIST_INIT(&cc->cc_callwheel[i]);
319 	TAILQ_INIT(&cc->cc_expireq);
320 	cc->cc_firstevent = SBT_MAX;
321 	for (i = 0; i < 2; i++)
322 		cc_cce_cleanup(cc, i);
323 	snprintf(cc->cc_ktr_event_name, sizeof(cc->cc_ktr_event_name),
324 	    "callwheel cpu %d", cpu);
325 	if (cc->cc_callout == NULL)	/* Only cpu0 handles timeout(9) */
326 		return;
327 	for (i = 0; i < ncallout; i++) {
328 		c = &cc->cc_callout[i];
329 		callout_init(c, 0);
330 		c->c_iflags = CALLOUT_LOCAL_ALLOC;
331 		SLIST_INSERT_HEAD(&cc->cc_callfree, c, c_links.sle);
332 	}
333 }
334 
335 #ifdef SMP
336 /*
337  * Switches the cpu tied to a specific callout.
338  * The function expects a locked incoming callout cpu and returns with
339  * locked outcoming callout cpu.
340  */
341 static struct callout_cpu *
342 callout_cpu_switch(struct callout *c, struct callout_cpu *cc, int new_cpu)
343 {
344 	struct callout_cpu *new_cc;
345 
346 	MPASS(c != NULL && cc != NULL);
347 	CC_LOCK_ASSERT(cc);
348 
349 	/*
350 	 * Avoid interrupts and preemption firing after the callout cpu
351 	 * is blocked in order to avoid deadlocks as the new thread
352 	 * may be willing to acquire the callout cpu lock.
353 	 */
354 	c->c_cpu = CPUBLOCK;
355 	spinlock_enter();
356 	CC_UNLOCK(cc);
357 	new_cc = CC_CPU(new_cpu);
358 	CC_LOCK(new_cc);
359 	spinlock_exit();
360 	c->c_cpu = new_cpu;
361 	return (new_cc);
362 }
363 #endif
364 
365 /*
366  * Start standard softclock thread.
367  */
368 static void
369 start_softclock(void *dummy)
370 {
371 	struct callout_cpu *cc;
372 	char name[MAXCOMLEN];
373 #ifdef SMP
374 	int cpu;
375 	struct intr_event *ie;
376 #endif
377 
378 	cc = CC_CPU(timeout_cpu);
379 	snprintf(name, sizeof(name), "clock (%d)", timeout_cpu);
380 	if (swi_add(&clk_intr_event, name, softclock, cc, SWI_CLOCK,
381 	    INTR_MPSAFE, &cc->cc_cookie))
382 		panic("died while creating standard software ithreads");
383 	if (pin_default_swi &&
384 	    (intr_event_bind(clk_intr_event, timeout_cpu) != 0)) {
385 		printf("%s: timeout clock couldn't be pinned to cpu %d\n",
386 		    __func__,
387 		    timeout_cpu);
388 	}
389 
390 #ifdef SMP
391 	CPU_FOREACH(cpu) {
392 		if (cpu == timeout_cpu)
393 			continue;
394 		cc = CC_CPU(cpu);
395 		cc->cc_callout = NULL;	/* Only cpu0 handles timeout(9). */
396 		callout_cpu_init(cc, cpu);
397 		snprintf(name, sizeof(name), "clock (%d)", cpu);
398 		ie = NULL;
399 		if (swi_add(&ie, name, softclock, cc, SWI_CLOCK,
400 		    INTR_MPSAFE, &cc->cc_cookie))
401 			panic("died while creating standard software ithreads");
402 		if (pin_pcpu_swi && (intr_event_bind(ie, cpu) != 0)) {
403 			printf("%s: per-cpu clock couldn't be pinned to "
404 			    "cpu %d\n",
405 			    __func__,
406 			    cpu);
407 		}
408 	}
409 #endif
410 }
411 SYSINIT(start_softclock, SI_SUB_SOFTINTR, SI_ORDER_FIRST, start_softclock, NULL);
412 
413 #define	CC_HASH_SHIFT	8
414 
415 static inline u_int
416 callout_hash(sbintime_t sbt)
417 {
418 
419 	return (sbt >> (32 - CC_HASH_SHIFT));
420 }
421 
422 static inline u_int
423 callout_get_bucket(sbintime_t sbt)
424 {
425 
426 	return (callout_hash(sbt) & callwheelmask);
427 }
428 
429 void
430 callout_process(sbintime_t now)
431 {
432 	struct callout *tmp, *tmpn;
433 	struct callout_cpu *cc;
434 	struct callout_list *sc;
435 	sbintime_t first, last, max, tmp_max;
436 	uint32_t lookahead;
437 	u_int firstb, lastb, nowb;
438 #ifdef CALLOUT_PROFILING
439 	int depth_dir = 0, mpcalls_dir = 0, lockcalls_dir = 0;
440 #endif
441 
442 	cc = CC_SELF();
443 	mtx_lock_spin_flags(&cc->cc_lock, MTX_QUIET);
444 
445 	/* Compute the buckets of the last scan and present times. */
446 	firstb = callout_hash(cc->cc_lastscan);
447 	cc->cc_lastscan = now;
448 	nowb = callout_hash(now);
449 
450 	/* Compute the last bucket and minimum time of the bucket after it. */
451 	if (nowb == firstb)
452 		lookahead = (SBT_1S / 16);
453 	else if (nowb - firstb == 1)
454 		lookahead = (SBT_1S / 8);
455 	else
456 		lookahead = (SBT_1S / 2);
457 	first = last = now;
458 	first += (lookahead / 2);
459 	last += lookahead;
460 	last &= (0xffffffffffffffffLLU << (32 - CC_HASH_SHIFT));
461 	lastb = callout_hash(last) - 1;
462 	max = last;
463 
464 	/*
465 	 * Check if we wrapped around the entire wheel from the last scan.
466 	 * In case, we need to scan entirely the wheel for pending callouts.
467 	 */
468 	if (lastb - firstb >= callwheelsize) {
469 		lastb = firstb + callwheelsize - 1;
470 		if (nowb - firstb >= callwheelsize)
471 			nowb = lastb;
472 	}
473 
474 	/* Iterate callwheel from firstb to nowb and then up to lastb. */
475 	do {
476 		sc = &cc->cc_callwheel[firstb & callwheelmask];
477 		tmp = LIST_FIRST(sc);
478 		while (tmp != NULL) {
479 			/* Run the callout if present time within allowed. */
480 			if (tmp->c_time <= now) {
481 				/*
482 				 * Consumer told us the callout may be run
483 				 * directly from hardware interrupt context.
484 				 */
485 				if (tmp->c_iflags & CALLOUT_DIRECT) {
486 #ifdef CALLOUT_PROFILING
487 					++depth_dir;
488 #endif
489 					cc_exec_next(cc) =
490 					    LIST_NEXT(tmp, c_links.le);
491 					cc->cc_bucket = firstb & callwheelmask;
492 					LIST_REMOVE(tmp, c_links.le);
493 					softclock_call_cc(tmp, cc,
494 #ifdef CALLOUT_PROFILING
495 					    &mpcalls_dir, &lockcalls_dir, NULL,
496 #endif
497 					    1);
498 					tmp = cc_exec_next(cc);
499 					cc_exec_next(cc) = NULL;
500 				} else {
501 					tmpn = LIST_NEXT(tmp, c_links.le);
502 					LIST_REMOVE(tmp, c_links.le);
503 					TAILQ_INSERT_TAIL(&cc->cc_expireq,
504 					    tmp, c_links.tqe);
505 					tmp->c_iflags |= CALLOUT_PROCESSED;
506 					tmp = tmpn;
507 				}
508 				continue;
509 			}
510 			/* Skip events from distant future. */
511 			if (tmp->c_time >= max)
512 				goto next;
513 			/*
514 			 * Event minimal time is bigger than present maximal
515 			 * time, so it cannot be aggregated.
516 			 */
517 			if (tmp->c_time > last) {
518 				lastb = nowb;
519 				goto next;
520 			}
521 			/* Update first and last time, respecting this event. */
522 			if (tmp->c_time < first)
523 				first = tmp->c_time;
524 			tmp_max = tmp->c_time + tmp->c_precision;
525 			if (tmp_max < last)
526 				last = tmp_max;
527 next:
528 			tmp = LIST_NEXT(tmp, c_links.le);
529 		}
530 		/* Proceed with the next bucket. */
531 		firstb++;
532 		/*
533 		 * Stop if we looked after present time and found
534 		 * some event we can't execute at now.
535 		 * Stop if we looked far enough into the future.
536 		 */
537 	} while (((int)(firstb - lastb)) <= 0);
538 	cc->cc_firstevent = last;
539 #ifndef NO_EVENTTIMERS
540 	cpu_new_callout(curcpu, last, first);
541 #endif
542 #ifdef CALLOUT_PROFILING
543 	avg_depth_dir += (depth_dir * 1000 - avg_depth_dir) >> 8;
544 	avg_mpcalls_dir += (mpcalls_dir * 1000 - avg_mpcalls_dir) >> 8;
545 	avg_lockcalls_dir += (lockcalls_dir * 1000 - avg_lockcalls_dir) >> 8;
546 #endif
547 	mtx_unlock_spin_flags(&cc->cc_lock, MTX_QUIET);
548 	/*
549 	 * swi_sched acquires the thread lock, so we don't want to call it
550 	 * with cc_lock held; incorrect locking order.
551 	 */
552 	if (!TAILQ_EMPTY(&cc->cc_expireq))
553 		swi_sched(cc->cc_cookie, 0);
554 }
555 
556 static struct callout_cpu *
557 callout_lock(struct callout *c)
558 {
559 	struct callout_cpu *cc;
560 	int cpu;
561 
562 	for (;;) {
563 		cpu = c->c_cpu;
564 #ifdef SMP
565 		if (cpu == CPUBLOCK) {
566 			while (c->c_cpu == CPUBLOCK)
567 				cpu_spinwait();
568 			continue;
569 		}
570 #endif
571 		cc = CC_CPU(cpu);
572 		CC_LOCK(cc);
573 		if (cpu == c->c_cpu)
574 			break;
575 		CC_UNLOCK(cc);
576 	}
577 	return (cc);
578 }
579 
580 static void
581 callout_cc_add(struct callout *c, struct callout_cpu *cc,
582     sbintime_t sbt, sbintime_t precision, void (*func)(void *),
583     void *arg, int cpu, int flags)
584 {
585 	int bucket;
586 
587 	CC_LOCK_ASSERT(cc);
588 	if (sbt < cc->cc_lastscan)
589 		sbt = cc->cc_lastscan;
590 	c->c_arg = arg;
591 	c->c_iflags |= CALLOUT_PENDING;
592 	c->c_iflags &= ~CALLOUT_PROCESSED;
593 	c->c_flags |= CALLOUT_ACTIVE;
594 	if (flags & C_DIRECT_EXEC)
595 		c->c_iflags |= CALLOUT_DIRECT;
596 	c->c_func = func;
597 	c->c_time = sbt;
598 	c->c_precision = precision;
599 	bucket = callout_get_bucket(c->c_time);
600 	CTR3(KTR_CALLOUT, "precision set for %p: %d.%08x",
601 	    c, (int)(c->c_precision >> 32),
602 	    (u_int)(c->c_precision & 0xffffffff));
603 	LIST_INSERT_HEAD(&cc->cc_callwheel[bucket], c, c_links.le);
604 	if (cc->cc_bucket == bucket)
605 		cc_exec_next(cc) = c;
606 #ifndef NO_EVENTTIMERS
607 	/*
608 	 * Inform the eventtimers(4) subsystem there's a new callout
609 	 * that has been inserted, but only if really required.
610 	 */
611 	if (SBT_MAX - c->c_time < c->c_precision)
612 		c->c_precision = SBT_MAX - c->c_time;
613 	sbt = c->c_time + c->c_precision;
614 	if (sbt < cc->cc_firstevent) {
615 		cc->cc_firstevent = sbt;
616 		cpu_new_callout(cpu, sbt, c->c_time);
617 	}
618 #endif
619 }
620 
621 static void
622 callout_cc_del(struct callout *c, struct callout_cpu *cc)
623 {
624 
625 	if ((c->c_iflags & CALLOUT_LOCAL_ALLOC) == 0)
626 		return;
627 	c->c_func = NULL;
628 	SLIST_INSERT_HEAD(&cc->cc_callfree, c, c_links.sle);
629 }
630 
631 static void
632 softclock_call_cc(struct callout *c, struct callout_cpu *cc,
633 #ifdef CALLOUT_PROFILING
634     int *mpcalls, int *lockcalls, int *gcalls,
635 #endif
636     int direct)
637 {
638 	struct rm_priotracker tracker;
639 	void (*c_func)(void *);
640 	void *c_arg;
641 	struct lock_class *class;
642 	struct lock_object *c_lock;
643 	uintptr_t lock_status;
644 	int c_iflags;
645 #ifdef SMP
646 	struct callout_cpu *new_cc;
647 	void (*new_func)(void *);
648 	void *new_arg;
649 	int flags, new_cpu;
650 	sbintime_t new_prec, new_time;
651 #endif
652 #if defined(DIAGNOSTIC) || defined(CALLOUT_PROFILING)
653 	sbintime_t sbt1, sbt2;
654 	struct timespec ts2;
655 	static sbintime_t maxdt = 2 * SBT_1MS;	/* 2 msec */
656 	static timeout_t *lastfunc;
657 #endif
658 
659 	KASSERT((c->c_iflags & CALLOUT_PENDING) == CALLOUT_PENDING,
660 	    ("softclock_call_cc: pend %p %x", c, c->c_iflags));
661 	KASSERT((c->c_flags & CALLOUT_ACTIVE) == CALLOUT_ACTIVE,
662 	    ("softclock_call_cc: act %p %x", c, c->c_flags));
663 	class = (c->c_lock != NULL) ? LOCK_CLASS(c->c_lock) : NULL;
664 	lock_status = 0;
665 	if (c->c_flags & CALLOUT_SHAREDLOCK) {
666 		if (class == &lock_class_rm)
667 			lock_status = (uintptr_t)&tracker;
668 		else
669 			lock_status = 1;
670 	}
671 	c_lock = c->c_lock;
672 	c_func = c->c_func;
673 	c_arg = c->c_arg;
674 	c_iflags = c->c_iflags;
675 	if (c->c_iflags & CALLOUT_LOCAL_ALLOC)
676 		c->c_iflags = CALLOUT_LOCAL_ALLOC;
677 	else
678 		c->c_iflags &= ~CALLOUT_PENDING;
679 
680 	cc_exec_curr(cc, direct) = c;
681 	cc_exec_cancel(cc, direct) = false;
682 	CC_UNLOCK(cc);
683 	if (c_lock != NULL) {
684 		class->lc_lock(c_lock, lock_status);
685 		/*
686 		 * The callout may have been cancelled
687 		 * while we switched locks.
688 		 */
689 		if (cc_exec_cancel(cc, direct)) {
690 			class->lc_unlock(c_lock);
691 			goto skip;
692 		}
693 		/* The callout cannot be stopped now. */
694 		cc_exec_cancel(cc, direct) = true;
695 		if (c_lock == &Giant.lock_object) {
696 #ifdef CALLOUT_PROFILING
697 			(*gcalls)++;
698 #endif
699 			CTR3(KTR_CALLOUT, "callout giant %p func %p arg %p",
700 			    c, c_func, c_arg);
701 		} else {
702 #ifdef CALLOUT_PROFILING
703 			(*lockcalls)++;
704 #endif
705 			CTR3(KTR_CALLOUT, "callout lock %p func %p arg %p",
706 			    c, c_func, c_arg);
707 		}
708 	} else {
709 #ifdef CALLOUT_PROFILING
710 		(*mpcalls)++;
711 #endif
712 		CTR3(KTR_CALLOUT, "callout %p func %p arg %p",
713 		    c, c_func, c_arg);
714 	}
715 	KTR_STATE3(KTR_SCHED, "callout", cc->cc_ktr_event_name, "running",
716 	    "func:%p", c_func, "arg:%p", c_arg, "direct:%d", direct);
717 #if defined(DIAGNOSTIC) || defined(CALLOUT_PROFILING)
718 	sbt1 = sbinuptime();
719 #endif
720 	THREAD_NO_SLEEPING();
721 	SDT_PROBE1(callout_execute, kernel, , callout__start, c);
722 	c_func(c_arg);
723 	SDT_PROBE1(callout_execute, kernel, , callout__end, c);
724 	THREAD_SLEEPING_OK();
725 #if defined(DIAGNOSTIC) || defined(CALLOUT_PROFILING)
726 	sbt2 = sbinuptime();
727 	sbt2 -= sbt1;
728 	if (sbt2 > maxdt) {
729 		if (lastfunc != c_func || sbt2 > maxdt * 2) {
730 			ts2 = sbttots(sbt2);
731 			printf(
732 		"Expensive timeout(9) function: %p(%p) %jd.%09ld s\n",
733 			    c_func, c_arg, (intmax_t)ts2.tv_sec, ts2.tv_nsec);
734 		}
735 		maxdt = sbt2;
736 		lastfunc = c_func;
737 	}
738 #endif
739 	KTR_STATE0(KTR_SCHED, "callout", cc->cc_ktr_event_name, "idle");
740 	CTR1(KTR_CALLOUT, "callout %p finished", c);
741 	if ((c_iflags & CALLOUT_RETURNUNLOCKED) == 0)
742 		class->lc_unlock(c_lock);
743 skip:
744 	CC_LOCK(cc);
745 	KASSERT(cc_exec_curr(cc, direct) == c, ("mishandled cc_curr"));
746 	cc_exec_curr(cc, direct) = NULL;
747 	if (cc_exec_waiting(cc, direct)) {
748 		/*
749 		 * There is someone waiting for the
750 		 * callout to complete.
751 		 * If the callout was scheduled for
752 		 * migration just cancel it.
753 		 */
754 		if (cc_cce_migrating(cc, direct)) {
755 			cc_cce_cleanup(cc, direct);
756 
757 			/*
758 			 * It should be assert here that the callout is not
759 			 * destroyed but that is not easy.
760 			 */
761 			c->c_iflags &= ~CALLOUT_DFRMIGRATION;
762 		}
763 		cc_exec_waiting(cc, direct) = false;
764 		CC_UNLOCK(cc);
765 		wakeup(&cc_exec_waiting(cc, direct));
766 		CC_LOCK(cc);
767 	} else if (cc_cce_migrating(cc, direct)) {
768 		KASSERT((c_iflags & CALLOUT_LOCAL_ALLOC) == 0,
769 		    ("Migrating legacy callout %p", c));
770 #ifdef SMP
771 		/*
772 		 * If the callout was scheduled for
773 		 * migration just perform it now.
774 		 */
775 		new_cpu = cc_migration_cpu(cc, direct);
776 		new_time = cc_migration_time(cc, direct);
777 		new_prec = cc_migration_prec(cc, direct);
778 		new_func = cc_migration_func(cc, direct);
779 		new_arg = cc_migration_arg(cc, direct);
780 		cc_cce_cleanup(cc, direct);
781 
782 		/*
783 		 * It should be assert here that the callout is not destroyed
784 		 * but that is not easy.
785 		 *
786 		 * As first thing, handle deferred callout stops.
787 		 */
788 		if (!callout_migrating(c)) {
789 			CTR3(KTR_CALLOUT,
790 			     "deferred cancelled %p func %p arg %p",
791 			     c, new_func, new_arg);
792 			callout_cc_del(c, cc);
793 			return;
794 		}
795 		c->c_iflags &= ~CALLOUT_DFRMIGRATION;
796 
797 		new_cc = callout_cpu_switch(c, cc, new_cpu);
798 		flags = (direct) ? C_DIRECT_EXEC : 0;
799 		callout_cc_add(c, new_cc, new_time, new_prec, new_func,
800 		    new_arg, new_cpu, flags);
801 		CC_UNLOCK(new_cc);
802 		CC_LOCK(cc);
803 #else
804 		panic("migration should not happen");
805 #endif
806 	}
807 	/*
808 	 * If the current callout is locally allocated (from
809 	 * timeout(9)) then put it on the freelist.
810 	 *
811 	 * Note: we need to check the cached copy of c_iflags because
812 	 * if it was not local, then it's not safe to deref the
813 	 * callout pointer.
814 	 */
815 	KASSERT((c_iflags & CALLOUT_LOCAL_ALLOC) == 0 ||
816 	    c->c_iflags == CALLOUT_LOCAL_ALLOC,
817 	    ("corrupted callout"));
818 	if (c_iflags & CALLOUT_LOCAL_ALLOC)
819 		callout_cc_del(c, cc);
820 }
821 
822 /*
823  * The callout mechanism is based on the work of Adam M. Costello and
824  * George Varghese, published in a technical report entitled "Redesigning
825  * the BSD Callout and Timer Facilities" and modified slightly for inclusion
826  * in FreeBSD by Justin T. Gibbs.  The original work on the data structures
827  * used in this implementation was published by G. Varghese and T. Lauck in
828  * the paper "Hashed and Hierarchical Timing Wheels: Data Structures for
829  * the Efficient Implementation of a Timer Facility" in the Proceedings of
830  * the 11th ACM Annual Symposium on Operating Systems Principles,
831  * Austin, Texas Nov 1987.
832  */
833 
834 /*
835  * Software (low priority) clock interrupt.
836  * Run periodic events from timeout queue.
837  */
838 void
839 softclock(void *arg)
840 {
841 	struct callout_cpu *cc;
842 	struct callout *c;
843 #ifdef CALLOUT_PROFILING
844 	int depth = 0, gcalls = 0, lockcalls = 0, mpcalls = 0;
845 #endif
846 
847 	cc = (struct callout_cpu *)arg;
848 	CC_LOCK(cc);
849 	while ((c = TAILQ_FIRST(&cc->cc_expireq)) != NULL) {
850 		TAILQ_REMOVE(&cc->cc_expireq, c, c_links.tqe);
851 		softclock_call_cc(c, cc,
852 #ifdef CALLOUT_PROFILING
853 		    &mpcalls, &lockcalls, &gcalls,
854 #endif
855 		    0);
856 #ifdef CALLOUT_PROFILING
857 		++depth;
858 #endif
859 	}
860 #ifdef CALLOUT_PROFILING
861 	avg_depth += (depth * 1000 - avg_depth) >> 8;
862 	avg_mpcalls += (mpcalls * 1000 - avg_mpcalls) >> 8;
863 	avg_lockcalls += (lockcalls * 1000 - avg_lockcalls) >> 8;
864 	avg_gcalls += (gcalls * 1000 - avg_gcalls) >> 8;
865 #endif
866 	CC_UNLOCK(cc);
867 }
868 
869 /*
870  * timeout --
871  *	Execute a function after a specified length of time.
872  *
873  * untimeout --
874  *	Cancel previous timeout function call.
875  *
876  * callout_handle_init --
877  *	Initialize a handle so that using it with untimeout is benign.
878  *
879  *	See AT&T BCI Driver Reference Manual for specification.  This
880  *	implementation differs from that one in that although an
881  *	identification value is returned from timeout, the original
882  *	arguments to timeout as well as the identifier are used to
883  *	identify entries for untimeout.
884  */
885 struct callout_handle
886 timeout(timeout_t *ftn, void *arg, int to_ticks)
887 {
888 	struct callout_cpu *cc;
889 	struct callout *new;
890 	struct callout_handle handle;
891 
892 	cc = CC_CPU(timeout_cpu);
893 	CC_LOCK(cc);
894 	/* Fill in the next free callout structure. */
895 	new = SLIST_FIRST(&cc->cc_callfree);
896 	if (new == NULL)
897 		/* XXX Attempt to malloc first */
898 		panic("timeout table full");
899 	SLIST_REMOVE_HEAD(&cc->cc_callfree, c_links.sle);
900 	callout_reset(new, to_ticks, ftn, arg);
901 	handle.callout = new;
902 	CC_UNLOCK(cc);
903 
904 	return (handle);
905 }
906 
907 void
908 untimeout(timeout_t *ftn, void *arg, struct callout_handle handle)
909 {
910 	struct callout_cpu *cc;
911 
912 	/*
913 	 * Check for a handle that was initialized
914 	 * by callout_handle_init, but never used
915 	 * for a real timeout.
916 	 */
917 	if (handle.callout == NULL)
918 		return;
919 
920 	cc = callout_lock(handle.callout);
921 	if (handle.callout->c_func == ftn && handle.callout->c_arg == arg)
922 		callout_stop(handle.callout);
923 	CC_UNLOCK(cc);
924 }
925 
926 void
927 callout_handle_init(struct callout_handle *handle)
928 {
929 	handle->callout = NULL;
930 }
931 
932 /*
933  * New interface; clients allocate their own callout structures.
934  *
935  * callout_reset() - establish or change a timeout
936  * callout_stop() - disestablish a timeout
937  * callout_init() - initialize a callout structure so that it can
938  *	safely be passed to callout_reset() and callout_stop()
939  *
940  * <sys/callout.h> defines three convenience macros:
941  *
942  * callout_active() - returns truth if callout has not been stopped,
943  *	drained, or deactivated since the last time the callout was
944  *	reset.
945  * callout_pending() - returns truth if callout is still waiting for timeout
946  * callout_deactivate() - marks the callout as having been serviced
947  */
948 int
949 callout_reset_sbt_on(struct callout *c, sbintime_t sbt, sbintime_t precision,
950     void (*ftn)(void *), void *arg, int cpu, int flags)
951 {
952 	sbintime_t to_sbt, pr;
953 	struct callout_cpu *cc;
954 	int cancelled, direct;
955 	int ignore_cpu=0;
956 
957 	cancelled = 0;
958 	if (cpu == -1) {
959 		ignore_cpu = 1;
960 	} else if ((cpu >= MAXCPU) ||
961 		   ((CC_CPU(cpu))->cc_inited == 0)) {
962 		/* Invalid CPU spec */
963 		panic("Invalid CPU in callout %d", cpu);
964 	}
965 	if (flags & C_ABSOLUTE) {
966 		to_sbt = sbt;
967 	} else {
968 		if ((flags & C_HARDCLOCK) && (sbt < tick_sbt))
969 			sbt = tick_sbt;
970 		if ((flags & C_HARDCLOCK) ||
971 #ifdef NO_EVENTTIMERS
972 		    sbt >= sbt_timethreshold) {
973 			to_sbt = getsbinuptime();
974 
975 			/* Add safety belt for the case of hz > 1000. */
976 			to_sbt += tc_tick_sbt - tick_sbt;
977 #else
978 		    sbt >= sbt_tickthreshold) {
979 			/*
980 			 * Obtain the time of the last hardclock() call on
981 			 * this CPU directly from the kern_clocksource.c.
982 			 * This value is per-CPU, but it is equal for all
983 			 * active ones.
984 			 */
985 #ifdef __LP64__
986 			to_sbt = DPCPU_GET(hardclocktime);
987 #else
988 			spinlock_enter();
989 			to_sbt = DPCPU_GET(hardclocktime);
990 			spinlock_exit();
991 #endif
992 #endif
993 			if ((flags & C_HARDCLOCK) == 0)
994 				to_sbt += tick_sbt;
995 		} else
996 			to_sbt = sbinuptime();
997 		if (SBT_MAX - to_sbt < sbt)
998 			to_sbt = SBT_MAX;
999 		else
1000 			to_sbt += sbt;
1001 		pr = ((C_PRELGET(flags) < 0) ? sbt >> tc_precexp :
1002 		    sbt >> C_PRELGET(flags));
1003 		if (pr > precision)
1004 			precision = pr;
1005 	}
1006 	/*
1007 	 * This flag used to be added by callout_cc_add, but the
1008 	 * first time you call this we could end up with the
1009 	 * wrong direct flag if we don't do it before we add.
1010 	 */
1011 	if (flags & C_DIRECT_EXEC) {
1012 		direct = 1;
1013 	} else {
1014 		direct = 0;
1015 	}
1016 	KASSERT(!direct || c->c_lock == NULL,
1017 	    ("%s: direct callout %p has lock", __func__, c));
1018 	cc = callout_lock(c);
1019 	/*
1020 	 * Don't allow migration of pre-allocated callouts lest they
1021 	 * become unbalanced or handle the case where the user does
1022 	 * not care.
1023 	 */
1024 	if ((c->c_iflags & CALLOUT_LOCAL_ALLOC) ||
1025 	    ignore_cpu) {
1026 		cpu = c->c_cpu;
1027 	}
1028 
1029 	if (cc_exec_curr(cc, direct) == c) {
1030 		/*
1031 		 * We're being asked to reschedule a callout which is
1032 		 * currently in progress.  If there is a lock then we
1033 		 * can cancel the callout if it has not really started.
1034 		 */
1035 		if (c->c_lock != NULL && !cc_exec_cancel(cc, direct))
1036 			cancelled = cc_exec_cancel(cc, direct) = true;
1037 		if (cc_exec_waiting(cc, direct)) {
1038 			/*
1039 			 * Someone has called callout_drain to kill this
1040 			 * callout.  Don't reschedule.
1041 			 */
1042 			CTR4(KTR_CALLOUT, "%s %p func %p arg %p",
1043 			    cancelled ? "cancelled" : "failed to cancel",
1044 			    c, c->c_func, c->c_arg);
1045 			CC_UNLOCK(cc);
1046 			return (cancelled);
1047 		}
1048 #ifdef SMP
1049 		if (callout_migrating(c)) {
1050 			/*
1051 			 * This only occurs when a second callout_reset_sbt_on
1052 			 * is made after a previous one moved it into
1053 			 * deferred migration (below). Note we do *not* change
1054 			 * the prev_cpu even though the previous target may
1055 			 * be different.
1056 			 */
1057 			cc_migration_cpu(cc, direct) = cpu;
1058 			cc_migration_time(cc, direct) = to_sbt;
1059 			cc_migration_prec(cc, direct) = precision;
1060 			cc_migration_func(cc, direct) = ftn;
1061 			cc_migration_arg(cc, direct) = arg;
1062 			cancelled = 1;
1063 			CC_UNLOCK(cc);
1064 			return (cancelled);
1065 		}
1066 #endif
1067 	}
1068 	if (c->c_iflags & CALLOUT_PENDING) {
1069 		if ((c->c_iflags & CALLOUT_PROCESSED) == 0) {
1070 			if (cc_exec_next(cc) == c)
1071 				cc_exec_next(cc) = LIST_NEXT(c, c_links.le);
1072 			LIST_REMOVE(c, c_links.le);
1073 		} else {
1074 			TAILQ_REMOVE(&cc->cc_expireq, c, c_links.tqe);
1075 		}
1076 		cancelled = 1;
1077 		c->c_iflags &= ~ CALLOUT_PENDING;
1078 		c->c_flags &= ~ CALLOUT_ACTIVE;
1079 	}
1080 
1081 #ifdef SMP
1082 	/*
1083 	 * If the callout must migrate try to perform it immediately.
1084 	 * If the callout is currently running, just defer the migration
1085 	 * to a more appropriate moment.
1086 	 */
1087 	if (c->c_cpu != cpu) {
1088 		if (cc_exec_curr(cc, direct) == c) {
1089 			/*
1090 			 * Pending will have been removed since we are
1091 			 * actually executing the callout on another
1092 			 * CPU. That callout should be waiting on the
1093 			 * lock the caller holds. If we set both
1094 			 * active/and/pending after we return and the
1095 			 * lock on the executing callout proceeds, it
1096 			 * will then see pending is true and return.
1097 			 * At the return from the actual callout execution
1098 			 * the migration will occur in softclock_call_cc
1099 			 * and this new callout will be placed on the
1100 			 * new CPU via a call to callout_cpu_switch() which
1101 			 * will get the lock on the right CPU followed
1102 			 * by a call callout_cc_add() which will add it there.
1103 			 * (see above in softclock_call_cc()).
1104 			 */
1105 			cc_migration_cpu(cc, direct) = cpu;
1106 			cc_migration_time(cc, direct) = to_sbt;
1107 			cc_migration_prec(cc, direct) = precision;
1108 			cc_migration_func(cc, direct) = ftn;
1109 			cc_migration_arg(cc, direct) = arg;
1110 			c->c_iflags |= (CALLOUT_DFRMIGRATION | CALLOUT_PENDING);
1111 			c->c_flags |= CALLOUT_ACTIVE;
1112 			CTR6(KTR_CALLOUT,
1113 		    "migration of %p func %p arg %p in %d.%08x to %u deferred",
1114 			    c, c->c_func, c->c_arg, (int)(to_sbt >> 32),
1115 			    (u_int)(to_sbt & 0xffffffff), cpu);
1116 			CC_UNLOCK(cc);
1117 			return (cancelled);
1118 		}
1119 		cc = callout_cpu_switch(c, cc, cpu);
1120 	}
1121 #endif
1122 
1123 	callout_cc_add(c, cc, to_sbt, precision, ftn, arg, cpu, flags);
1124 	CTR6(KTR_CALLOUT, "%sscheduled %p func %p arg %p in %d.%08x",
1125 	    cancelled ? "re" : "", c, c->c_func, c->c_arg, (int)(to_sbt >> 32),
1126 	    (u_int)(to_sbt & 0xffffffff));
1127 	CC_UNLOCK(cc);
1128 
1129 	return (cancelled);
1130 }
1131 
1132 /*
1133  * Common idioms that can be optimized in the future.
1134  */
1135 int
1136 callout_schedule_on(struct callout *c, int to_ticks, int cpu)
1137 {
1138 	return callout_reset_on(c, to_ticks, c->c_func, c->c_arg, cpu);
1139 }
1140 
1141 int
1142 callout_schedule(struct callout *c, int to_ticks)
1143 {
1144 	return callout_reset_on(c, to_ticks, c->c_func, c->c_arg, c->c_cpu);
1145 }
1146 
1147 int
1148 _callout_stop_safe(struct callout *c, int safe)
1149 {
1150 	struct callout_cpu *cc, *old_cc;
1151 	struct lock_class *class;
1152 	int direct, sq_locked, use_lock;
1153 	int not_on_a_list;
1154 
1155 	if (safe)
1156 		WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK, c->c_lock,
1157 		    "calling %s", __func__);
1158 
1159 	/*
1160 	 * Some old subsystems don't hold Giant while running a callout_stop(),
1161 	 * so just discard this check for the moment.
1162 	 */
1163 	if (!safe && c->c_lock != NULL) {
1164 		if (c->c_lock == &Giant.lock_object)
1165 			use_lock = mtx_owned(&Giant);
1166 		else {
1167 			use_lock = 1;
1168 			class = LOCK_CLASS(c->c_lock);
1169 			class->lc_assert(c->c_lock, LA_XLOCKED);
1170 		}
1171 	} else
1172 		use_lock = 0;
1173 	if (c->c_iflags & CALLOUT_DIRECT) {
1174 		direct = 1;
1175 	} else {
1176 		direct = 0;
1177 	}
1178 	sq_locked = 0;
1179 	old_cc = NULL;
1180 again:
1181 	cc = callout_lock(c);
1182 
1183 	if ((c->c_iflags & (CALLOUT_DFRMIGRATION | CALLOUT_PENDING)) ==
1184 	    (CALLOUT_DFRMIGRATION | CALLOUT_PENDING) &&
1185 	    ((c->c_flags & CALLOUT_ACTIVE) == CALLOUT_ACTIVE)) {
1186 		/*
1187 		 * Special case where this slipped in while we
1188 		 * were migrating *as* the callout is about to
1189 		 * execute. The caller probably holds the lock
1190 		 * the callout wants.
1191 		 *
1192 		 * Get rid of the migration first. Then set
1193 		 * the flag that tells this code *not* to
1194 		 * try to remove it from any lists (its not
1195 		 * on one yet). When the callout wheel runs,
1196 		 * it will ignore this callout.
1197 		 */
1198 		c->c_iflags &= ~CALLOUT_PENDING;
1199 		c->c_flags &= ~CALLOUT_ACTIVE;
1200 		not_on_a_list = 1;
1201 	} else {
1202 		not_on_a_list = 0;
1203 	}
1204 
1205 	/*
1206 	 * If the callout was migrating while the callout cpu lock was
1207 	 * dropped,  just drop the sleepqueue lock and check the states
1208 	 * again.
1209 	 */
1210 	if (sq_locked != 0 && cc != old_cc) {
1211 #ifdef SMP
1212 		CC_UNLOCK(cc);
1213 		sleepq_release(&cc_exec_waiting(old_cc, direct));
1214 		sq_locked = 0;
1215 		old_cc = NULL;
1216 		goto again;
1217 #else
1218 		panic("migration should not happen");
1219 #endif
1220 	}
1221 
1222 	/*
1223 	 * If the callout isn't pending, it's not on the queue, so
1224 	 * don't attempt to remove it from the queue.  We can try to
1225 	 * stop it by other means however.
1226 	 */
1227 	if (!(c->c_iflags & CALLOUT_PENDING)) {
1228 		c->c_flags &= ~CALLOUT_ACTIVE;
1229 
1230 		/*
1231 		 * If it wasn't on the queue and it isn't the current
1232 		 * callout, then we can't stop it, so just bail.
1233 		 */
1234 		if (cc_exec_curr(cc, direct) != c) {
1235 			CTR3(KTR_CALLOUT, "failed to stop %p func %p arg %p",
1236 			    c, c->c_func, c->c_arg);
1237 			CC_UNLOCK(cc);
1238 			if (sq_locked)
1239 				sleepq_release(&cc_exec_waiting(cc, direct));
1240 			return (0);
1241 		}
1242 
1243 		if (safe) {
1244 			/*
1245 			 * The current callout is running (or just
1246 			 * about to run) and blocking is allowed, so
1247 			 * just wait for the current invocation to
1248 			 * finish.
1249 			 */
1250 			while (cc_exec_curr(cc, direct) == c) {
1251 				/*
1252 				 * Use direct calls to sleepqueue interface
1253 				 * instead of cv/msleep in order to avoid
1254 				 * a LOR between cc_lock and sleepqueue
1255 				 * chain spinlocks.  This piece of code
1256 				 * emulates a msleep_spin() call actually.
1257 				 *
1258 				 * If we already have the sleepqueue chain
1259 				 * locked, then we can safely block.  If we
1260 				 * don't already have it locked, however,
1261 				 * we have to drop the cc_lock to lock
1262 				 * it.  This opens several races, so we
1263 				 * restart at the beginning once we have
1264 				 * both locks.  If nothing has changed, then
1265 				 * we will end up back here with sq_locked
1266 				 * set.
1267 				 */
1268 				if (!sq_locked) {
1269 					CC_UNLOCK(cc);
1270 					sleepq_lock(
1271 					    &cc_exec_waiting(cc, direct));
1272 					sq_locked = 1;
1273 					old_cc = cc;
1274 					goto again;
1275 				}
1276 
1277 				/*
1278 				 * Migration could be cancelled here, but
1279 				 * as long as it is still not sure when it
1280 				 * will be packed up, just let softclock()
1281 				 * take care of it.
1282 				 */
1283 				cc_exec_waiting(cc, direct) = true;
1284 				DROP_GIANT();
1285 				CC_UNLOCK(cc);
1286 				sleepq_add(
1287 				    &cc_exec_waiting(cc, direct),
1288 				    &cc->cc_lock.lock_object, "codrain",
1289 				    SLEEPQ_SLEEP, 0);
1290 				sleepq_wait(
1291 				    &cc_exec_waiting(cc, direct),
1292 					     0);
1293 				sq_locked = 0;
1294 				old_cc = NULL;
1295 
1296 				/* Reacquire locks previously released. */
1297 				PICKUP_GIANT();
1298 				CC_LOCK(cc);
1299 			}
1300 		} else if (use_lock &&
1301 			   !cc_exec_cancel(cc, direct)) {
1302 
1303 			/*
1304 			 * The current callout is waiting for its
1305 			 * lock which we hold.  Cancel the callout
1306 			 * and return.  After our caller drops the
1307 			 * lock, the callout will be skipped in
1308 			 * softclock().
1309 			 */
1310 			cc_exec_cancel(cc, direct) = true;
1311 			CTR3(KTR_CALLOUT, "cancelled %p func %p arg %p",
1312 			    c, c->c_func, c->c_arg);
1313 			KASSERT(!cc_cce_migrating(cc, direct),
1314 			    ("callout wrongly scheduled for migration"));
1315 			if (callout_migrating(c)) {
1316 				c->c_iflags &= ~CALLOUT_DFRMIGRATION;
1317 #ifdef SMP
1318 				cc_migration_cpu(cc, direct) = CPUBLOCK;
1319 				cc_migration_time(cc, direct) = 0;
1320 				cc_migration_prec(cc, direct) = 0;
1321 				cc_migration_func(cc, direct) = NULL;
1322 				cc_migration_arg(cc, direct) = NULL;
1323 #endif
1324 			}
1325 			CC_UNLOCK(cc);
1326 			KASSERT(!sq_locked, ("sleepqueue chain locked"));
1327 			return (1);
1328 		} else if (callout_migrating(c)) {
1329 			/*
1330 			 * The callout is currently being serviced
1331 			 * and the "next" callout is scheduled at
1332 			 * its completion with a migration. We remove
1333 			 * the migration flag so it *won't* get rescheduled,
1334 			 * but we can't stop the one thats running so
1335 			 * we return 0.
1336 			 */
1337 			c->c_iflags &= ~CALLOUT_DFRMIGRATION;
1338 #ifdef SMP
1339 			/*
1340 			 * We can't call cc_cce_cleanup here since
1341 			 * if we do it will remove .ce_curr and
1342 			 * its still running. This will prevent a
1343 			 * reschedule of the callout when the
1344 			 * execution completes.
1345 			 */
1346 			cc_migration_cpu(cc, direct) = CPUBLOCK;
1347 			cc_migration_time(cc, direct) = 0;
1348 			cc_migration_prec(cc, direct) = 0;
1349 			cc_migration_func(cc, direct) = NULL;
1350 			cc_migration_arg(cc, direct) = NULL;
1351 #endif
1352 			CTR3(KTR_CALLOUT, "postponing stop %p func %p arg %p",
1353 			    c, c->c_func, c->c_arg);
1354 			CC_UNLOCK(cc);
1355 			return (0);
1356 		}
1357 		CTR3(KTR_CALLOUT, "failed to stop %p func %p arg %p",
1358 		    c, c->c_func, c->c_arg);
1359 		CC_UNLOCK(cc);
1360 		KASSERT(!sq_locked, ("sleepqueue chain still locked"));
1361 		return (0);
1362 	}
1363 	if (sq_locked)
1364 		sleepq_release(&cc_exec_waiting(cc, direct));
1365 
1366 	c->c_iflags &= ~CALLOUT_PENDING;
1367 	c->c_flags &= ~CALLOUT_ACTIVE;
1368 
1369 	CTR3(KTR_CALLOUT, "cancelled %p func %p arg %p",
1370 	    c, c->c_func, c->c_arg);
1371 	if (not_on_a_list == 0) {
1372 		if ((c->c_iflags & CALLOUT_PROCESSED) == 0) {
1373 			if (cc_exec_next(cc) == c)
1374 				cc_exec_next(cc) = LIST_NEXT(c, c_links.le);
1375 			LIST_REMOVE(c, c_links.le);
1376 		} else {
1377 			TAILQ_REMOVE(&cc->cc_expireq, c, c_links.tqe);
1378 		}
1379 	}
1380 	callout_cc_del(c, cc);
1381 	CC_UNLOCK(cc);
1382 	return (1);
1383 }
1384 
1385 void
1386 callout_init(struct callout *c, int mpsafe)
1387 {
1388 	bzero(c, sizeof *c);
1389 	if (mpsafe) {
1390 		c->c_lock = NULL;
1391 		c->c_iflags = CALLOUT_RETURNUNLOCKED;
1392 	} else {
1393 		c->c_lock = &Giant.lock_object;
1394 		c->c_iflags = 0;
1395 	}
1396 	c->c_cpu = timeout_cpu;
1397 }
1398 
1399 void
1400 _callout_init_lock(struct callout *c, struct lock_object *lock, int flags)
1401 {
1402 	bzero(c, sizeof *c);
1403 	c->c_lock = lock;
1404 	KASSERT((flags & ~(CALLOUT_RETURNUNLOCKED | CALLOUT_SHAREDLOCK)) == 0,
1405 	    ("callout_init_lock: bad flags %d", flags));
1406 	KASSERT(lock != NULL || (flags & CALLOUT_RETURNUNLOCKED) == 0,
1407 	    ("callout_init_lock: CALLOUT_RETURNUNLOCKED with no lock"));
1408 	KASSERT(lock == NULL || !(LOCK_CLASS(lock)->lc_flags &
1409 	    (LC_SPINLOCK | LC_SLEEPABLE)), ("%s: invalid lock class",
1410 	    __func__));
1411 	c->c_iflags = flags & (CALLOUT_RETURNUNLOCKED | CALLOUT_SHAREDLOCK);
1412 	c->c_cpu = timeout_cpu;
1413 }
1414 
1415 #ifdef APM_FIXUP_CALLTODO
1416 /*
1417  * Adjust the kernel calltodo timeout list.  This routine is used after
1418  * an APM resume to recalculate the calltodo timer list values with the
1419  * number of hz's we have been sleeping.  The next hardclock() will detect
1420  * that there are fired timers and run softclock() to execute them.
1421  *
1422  * Please note, I have not done an exhaustive analysis of what code this
1423  * might break.  I am motivated to have my select()'s and alarm()'s that
1424  * have expired during suspend firing upon resume so that the applications
1425  * which set the timer can do the maintanence the timer was for as close
1426  * as possible to the originally intended time.  Testing this code for a
1427  * week showed that resuming from a suspend resulted in 22 to 25 timers
1428  * firing, which seemed independant on whether the suspend was 2 hours or
1429  * 2 days.  Your milage may vary.   - Ken Key <key@cs.utk.edu>
1430  */
1431 void
1432 adjust_timeout_calltodo(struct timeval *time_change)
1433 {
1434 	register struct callout *p;
1435 	unsigned long delta_ticks;
1436 
1437 	/*
1438 	 * How many ticks were we asleep?
1439 	 * (stolen from tvtohz()).
1440 	 */
1441 
1442 	/* Don't do anything */
1443 	if (time_change->tv_sec < 0)
1444 		return;
1445 	else if (time_change->tv_sec <= LONG_MAX / 1000000)
1446 		delta_ticks = (time_change->tv_sec * 1000000 +
1447 			       time_change->tv_usec + (tick - 1)) / tick + 1;
1448 	else if (time_change->tv_sec <= LONG_MAX / hz)
1449 		delta_ticks = time_change->tv_sec * hz +
1450 			      (time_change->tv_usec + (tick - 1)) / tick + 1;
1451 	else
1452 		delta_ticks = LONG_MAX;
1453 
1454 	if (delta_ticks > INT_MAX)
1455 		delta_ticks = INT_MAX;
1456 
1457 	/*
1458 	 * Now rip through the timer calltodo list looking for timers
1459 	 * to expire.
1460 	 */
1461 
1462 	/* don't collide with softclock() */
1463 	CC_LOCK(cc);
1464 	for (p = calltodo.c_next; p != NULL; p = p->c_next) {
1465 		p->c_time -= delta_ticks;
1466 
1467 		/* Break if the timer had more time on it than delta_ticks */
1468 		if (p->c_time > 0)
1469 			break;
1470 
1471 		/* take back the ticks the timer didn't use (p->c_time <= 0) */
1472 		delta_ticks = -p->c_time;
1473 	}
1474 	CC_UNLOCK(cc);
1475 
1476 	return;
1477 }
1478 #endif /* APM_FIXUP_CALLTODO */
1479 
1480 static int
1481 flssbt(sbintime_t sbt)
1482 {
1483 
1484 	sbt += (uint64_t)sbt >> 1;
1485 	if (sizeof(long) >= sizeof(sbintime_t))
1486 		return (flsl(sbt));
1487 	if (sbt >= SBT_1S)
1488 		return (flsl(((uint64_t)sbt) >> 32) + 32);
1489 	return (flsl(sbt));
1490 }
1491 
1492 /*
1493  * Dump immediate statistic snapshot of the scheduled callouts.
1494  */
1495 static int
1496 sysctl_kern_callout_stat(SYSCTL_HANDLER_ARGS)
1497 {
1498 	struct callout *tmp;
1499 	struct callout_cpu *cc;
1500 	struct callout_list *sc;
1501 	sbintime_t maxpr, maxt, medpr, medt, now, spr, st, t;
1502 	int ct[64], cpr[64], ccpbk[32];
1503 	int error, val, i, count, tcum, pcum, maxc, c, medc;
1504 #ifdef SMP
1505 	int cpu;
1506 #endif
1507 
1508 	val = 0;
1509 	error = sysctl_handle_int(oidp, &val, 0, req);
1510 	if (error != 0 || req->newptr == NULL)
1511 		return (error);
1512 	count = maxc = 0;
1513 	st = spr = maxt = maxpr = 0;
1514 	bzero(ccpbk, sizeof(ccpbk));
1515 	bzero(ct, sizeof(ct));
1516 	bzero(cpr, sizeof(cpr));
1517 	now = sbinuptime();
1518 #ifdef SMP
1519 	CPU_FOREACH(cpu) {
1520 		cc = CC_CPU(cpu);
1521 #else
1522 		cc = CC_CPU(timeout_cpu);
1523 #endif
1524 		CC_LOCK(cc);
1525 		for (i = 0; i < callwheelsize; i++) {
1526 			sc = &cc->cc_callwheel[i];
1527 			c = 0;
1528 			LIST_FOREACH(tmp, sc, c_links.le) {
1529 				c++;
1530 				t = tmp->c_time - now;
1531 				if (t < 0)
1532 					t = 0;
1533 				st += t / SBT_1US;
1534 				spr += tmp->c_precision / SBT_1US;
1535 				if (t > maxt)
1536 					maxt = t;
1537 				if (tmp->c_precision > maxpr)
1538 					maxpr = tmp->c_precision;
1539 				ct[flssbt(t)]++;
1540 				cpr[flssbt(tmp->c_precision)]++;
1541 			}
1542 			if (c > maxc)
1543 				maxc = c;
1544 			ccpbk[fls(c + c / 2)]++;
1545 			count += c;
1546 		}
1547 		CC_UNLOCK(cc);
1548 #ifdef SMP
1549 	}
1550 #endif
1551 
1552 	for (i = 0, tcum = 0; i < 64 && tcum < count / 2; i++)
1553 		tcum += ct[i];
1554 	medt = (i >= 2) ? (((sbintime_t)1) << (i - 2)) : 0;
1555 	for (i = 0, pcum = 0; i < 64 && pcum < count / 2; i++)
1556 		pcum += cpr[i];
1557 	medpr = (i >= 2) ? (((sbintime_t)1) << (i - 2)) : 0;
1558 	for (i = 0, c = 0; i < 32 && c < count / 2; i++)
1559 		c += ccpbk[i];
1560 	medc = (i >= 2) ? (1 << (i - 2)) : 0;
1561 
1562 	printf("Scheduled callouts statistic snapshot:\n");
1563 	printf("  Callouts: %6d  Buckets: %6d*%-3d  Bucket size: 0.%06ds\n",
1564 	    count, callwheelsize, mp_ncpus, 1000000 >> CC_HASH_SHIFT);
1565 	printf("  C/Bk: med %5d         avg %6d.%06jd  max %6d\n",
1566 	    medc,
1567 	    count / callwheelsize / mp_ncpus,
1568 	    (uint64_t)count * 1000000 / callwheelsize / mp_ncpus % 1000000,
1569 	    maxc);
1570 	printf("  Time: med %5jd.%06jds avg %6jd.%06jds max %6jd.%06jds\n",
1571 	    medt / SBT_1S, (medt & 0xffffffff) * 1000000 >> 32,
1572 	    (st / count) / 1000000, (st / count) % 1000000,
1573 	    maxt / SBT_1S, (maxt & 0xffffffff) * 1000000 >> 32);
1574 	printf("  Prec: med %5jd.%06jds avg %6jd.%06jds max %6jd.%06jds\n",
1575 	    medpr / SBT_1S, (medpr & 0xffffffff) * 1000000 >> 32,
1576 	    (spr / count) / 1000000, (spr / count) % 1000000,
1577 	    maxpr / SBT_1S, (maxpr & 0xffffffff) * 1000000 >> 32);
1578 	printf("  Distribution:       \tbuckets\t   time\t   tcum\t"
1579 	    "   prec\t   pcum\n");
1580 	for (i = 0, tcum = pcum = 0; i < 64; i++) {
1581 		if (ct[i] == 0 && cpr[i] == 0)
1582 			continue;
1583 		t = (i != 0) ? (((sbintime_t)1) << (i - 1)) : 0;
1584 		tcum += ct[i];
1585 		pcum += cpr[i];
1586 		printf("  %10jd.%06jds\t 2**%d\t%7d\t%7d\t%7d\t%7d\n",
1587 		    t / SBT_1S, (t & 0xffffffff) * 1000000 >> 32,
1588 		    i - 1 - (32 - CC_HASH_SHIFT),
1589 		    ct[i], tcum, cpr[i], pcum);
1590 	}
1591 	return (error);
1592 }
1593 SYSCTL_PROC(_kern, OID_AUTO, callout_stat,
1594     CTLTYPE_INT | CTLFLAG_RW | CTLFLAG_MPSAFE,
1595     0, 0, sysctl_kern_callout_stat, "I",
1596     "Dump immediate statistic snapshot of the scheduled callouts");
1597