xref: /freebsd/sys/kern/kern_timeout.c (revision 9af6c78cd43b18e169f10802142c61638bd62bed)
1 /*-
2  * SPDX-License-Identifier: BSD-3-Clause
3  *
4  * Copyright (c) 1982, 1986, 1991, 1993
5  *	The Regents of the University of California.  All rights reserved.
6  * (c) UNIX System Laboratories, Inc.
7  * All or some portions of this file are derived from material licensed
8  * to the University of California by American Telephone and Telegraph
9  * Co. or Unix System Laboratories, Inc. and are reproduced herein with
10  * the permission of UNIX System Laboratories, Inc.
11  *
12  * Redistribution and use in source and binary forms, with or without
13  * modification, are permitted provided that the following conditions
14  * are met:
15  * 1. Redistributions of source code must retain the above copyright
16  *    notice, this list of conditions and the following disclaimer.
17  * 2. Redistributions in binary form must reproduce the above copyright
18  *    notice, this list of conditions and the following disclaimer in the
19  *    documentation and/or other materials provided with the distribution.
20  * 3. Neither the name of the University nor the names of its contributors
21  *    may be used to endorse or promote products derived from this software
22  *    without specific prior written permission.
23  *
24  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
25  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
26  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
27  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
28  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
29  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
30  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
31  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
32  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
33  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
34  * SUCH DAMAGE.
35  *
36  *	From: @(#)kern_clock.c	8.5 (Berkeley) 1/21/94
37  */
38 
39 #include <sys/cdefs.h>
40 __FBSDID("$FreeBSD$");
41 
42 #include "opt_callout_profiling.h"
43 #include "opt_ddb.h"
44 #if defined(__arm__)
45 #include "opt_timer.h"
46 #endif
47 #include "opt_rss.h"
48 
49 #include <sys/param.h>
50 #include <sys/systm.h>
51 #include <sys/bus.h>
52 #include <sys/callout.h>
53 #include <sys/domainset.h>
54 #include <sys/file.h>
55 #include <sys/interrupt.h>
56 #include <sys/kernel.h>
57 #include <sys/ktr.h>
58 #include <sys/lock.h>
59 #include <sys/malloc.h>
60 #include <sys/mutex.h>
61 #include <sys/proc.h>
62 #include <sys/sdt.h>
63 #include <sys/sleepqueue.h>
64 #include <sys/sysctl.h>
65 #include <sys/smp.h>
66 
67 #ifdef DDB
68 #include <ddb/ddb.h>
69 #include <ddb/db_sym.h>
70 #include <machine/_inttypes.h>
71 #endif
72 
73 #ifdef SMP
74 #include <machine/cpu.h>
75 #endif
76 
77 #ifndef NO_EVENTTIMERS
78 DPCPU_DECLARE(sbintime_t, hardclocktime);
79 #endif
80 
81 SDT_PROVIDER_DEFINE(callout_execute);
82 SDT_PROBE_DEFINE1(callout_execute, , , callout__start, "struct callout *");
83 SDT_PROBE_DEFINE1(callout_execute, , , callout__end, "struct callout *");
84 
85 #ifdef CALLOUT_PROFILING
86 static int avg_depth;
87 SYSCTL_INT(_debug, OID_AUTO, to_avg_depth, CTLFLAG_RD, &avg_depth, 0,
88     "Average number of items examined per softclock call. Units = 1/1000");
89 static int avg_gcalls;
90 SYSCTL_INT(_debug, OID_AUTO, to_avg_gcalls, CTLFLAG_RD, &avg_gcalls, 0,
91     "Average number of Giant callouts made per softclock call. Units = 1/1000");
92 static int avg_lockcalls;
93 SYSCTL_INT(_debug, OID_AUTO, to_avg_lockcalls, CTLFLAG_RD, &avg_lockcalls, 0,
94     "Average number of lock callouts made per softclock call. Units = 1/1000");
95 static int avg_mpcalls;
96 SYSCTL_INT(_debug, OID_AUTO, to_avg_mpcalls, CTLFLAG_RD, &avg_mpcalls, 0,
97     "Average number of MP callouts made per softclock call. Units = 1/1000");
98 static int avg_depth_dir;
99 SYSCTL_INT(_debug, OID_AUTO, to_avg_depth_dir, CTLFLAG_RD, &avg_depth_dir, 0,
100     "Average number of direct callouts examined per callout_process call. "
101     "Units = 1/1000");
102 static int avg_lockcalls_dir;
103 SYSCTL_INT(_debug, OID_AUTO, to_avg_lockcalls_dir, CTLFLAG_RD,
104     &avg_lockcalls_dir, 0, "Average number of lock direct callouts made per "
105     "callout_process call. Units = 1/1000");
106 static int avg_mpcalls_dir;
107 SYSCTL_INT(_debug, OID_AUTO, to_avg_mpcalls_dir, CTLFLAG_RD, &avg_mpcalls_dir,
108     0, "Average number of MP direct callouts made per callout_process call. "
109     "Units = 1/1000");
110 #endif
111 
112 static int ncallout;
113 SYSCTL_INT(_kern, OID_AUTO, ncallout, CTLFLAG_RDTUN | CTLFLAG_NOFETCH, &ncallout, 0,
114     "Number of entries in callwheel and size of timeout() preallocation");
115 
116 #ifdef	RSS
117 static int pin_default_swi = 1;
118 static int pin_pcpu_swi = 1;
119 #else
120 static int pin_default_swi = 0;
121 static int pin_pcpu_swi = 0;
122 #endif
123 
124 SYSCTL_INT(_kern, OID_AUTO, pin_default_swi, CTLFLAG_RDTUN | CTLFLAG_NOFETCH, &pin_default_swi,
125     0, "Pin the default (non-per-cpu) swi (shared with PCPU 0 swi)");
126 SYSCTL_INT(_kern, OID_AUTO, pin_pcpu_swi, CTLFLAG_RDTUN | CTLFLAG_NOFETCH, &pin_pcpu_swi,
127     0, "Pin the per-CPU swis (except PCPU 0, which is also default");
128 
129 /*
130  * TODO:
131  *	allocate more timeout table slots when table overflows.
132  */
133 u_int callwheelsize, callwheelmask;
134 
135 /*
136  * The callout cpu exec entities represent informations necessary for
137  * describing the state of callouts currently running on the CPU and the ones
138  * necessary for migrating callouts to the new callout cpu. In particular,
139  * the first entry of the array cc_exec_entity holds informations for callout
140  * running in SWI thread context, while the second one holds informations
141  * for callout running directly from hardware interrupt context.
142  * The cached informations are very important for deferring migration when
143  * the migrating callout is already running.
144  */
145 struct cc_exec {
146 	struct callout		*cc_curr;
147 	void			(*cc_drain)(void *);
148 	void			*cc_last_func;
149 	void			*cc_last_arg;
150 #ifdef SMP
151 	void			(*ce_migration_func)(void *);
152 	void			*ce_migration_arg;
153 	sbintime_t		ce_migration_time;
154 	sbintime_t		ce_migration_prec;
155 	int			ce_migration_cpu;
156 #endif
157 	bool			cc_cancel;
158 	bool			cc_waiting;
159 };
160 
161 /*
162  * There is one struct callout_cpu per cpu, holding all relevant
163  * state for the callout processing thread on the individual CPU.
164  */
165 struct callout_cpu {
166 	struct mtx_padalign	cc_lock;
167 	struct cc_exec 		cc_exec_entity[2];
168 	struct callout		*cc_next;
169 	struct callout		*cc_callout;
170 	struct callout_list	*cc_callwheel;
171 	struct callout_tailq	cc_expireq;
172 	struct callout_slist	cc_callfree;
173 	sbintime_t		cc_firstevent;
174 	sbintime_t		cc_lastscan;
175 	void			*cc_cookie;
176 	u_int			cc_bucket;
177 	u_int			cc_inited;
178 	char			cc_ktr_event_name[20];
179 };
180 
181 #define	callout_migrating(c)	((c)->c_iflags & CALLOUT_DFRMIGRATION)
182 
183 #define	cc_exec_curr(cc, dir)		cc->cc_exec_entity[dir].cc_curr
184 #define	cc_exec_last_func(cc, dir)	cc->cc_exec_entity[dir].cc_last_func
185 #define	cc_exec_last_arg(cc, dir)	cc->cc_exec_entity[dir].cc_last_arg
186 #define	cc_exec_drain(cc, dir)		cc->cc_exec_entity[dir].cc_drain
187 #define	cc_exec_next(cc)		cc->cc_next
188 #define	cc_exec_cancel(cc, dir)		cc->cc_exec_entity[dir].cc_cancel
189 #define	cc_exec_waiting(cc, dir)	cc->cc_exec_entity[dir].cc_waiting
190 #ifdef SMP
191 #define	cc_migration_func(cc, dir)	cc->cc_exec_entity[dir].ce_migration_func
192 #define	cc_migration_arg(cc, dir)	cc->cc_exec_entity[dir].ce_migration_arg
193 #define	cc_migration_cpu(cc, dir)	cc->cc_exec_entity[dir].ce_migration_cpu
194 #define	cc_migration_time(cc, dir)	cc->cc_exec_entity[dir].ce_migration_time
195 #define	cc_migration_prec(cc, dir)	cc->cc_exec_entity[dir].ce_migration_prec
196 
197 struct callout_cpu cc_cpu[MAXCPU];
198 #define	CPUBLOCK	MAXCPU
199 #define	CC_CPU(cpu)	(&cc_cpu[(cpu)])
200 #define	CC_SELF()	CC_CPU(PCPU_GET(cpuid))
201 #else
202 struct callout_cpu cc_cpu;
203 #define	CC_CPU(cpu)	&cc_cpu
204 #define	CC_SELF()	&cc_cpu
205 #endif
206 #define	CC_LOCK(cc)	mtx_lock_spin(&(cc)->cc_lock)
207 #define	CC_UNLOCK(cc)	mtx_unlock_spin(&(cc)->cc_lock)
208 #define	CC_LOCK_ASSERT(cc)	mtx_assert(&(cc)->cc_lock, MA_OWNED)
209 
210 static int timeout_cpu;
211 
212 static void	callout_cpu_init(struct callout_cpu *cc, int cpu);
213 static void	softclock_call_cc(struct callout *c, struct callout_cpu *cc,
214 #ifdef CALLOUT_PROFILING
215 		    int *mpcalls, int *lockcalls, int *gcalls,
216 #endif
217 		    int direct);
218 
219 static MALLOC_DEFINE(M_CALLOUT, "callout", "Callout datastructures");
220 
221 /**
222  * Locked by cc_lock:
223  *   cc_curr         - If a callout is in progress, it is cc_curr.
224  *                     If cc_curr is non-NULL, threads waiting in
225  *                     callout_drain() will be woken up as soon as the
226  *                     relevant callout completes.
227  *   cc_cancel       - Changing to 1 with both callout_lock and cc_lock held
228  *                     guarantees that the current callout will not run.
229  *                     The softclock() function sets this to 0 before it
230  *                     drops callout_lock to acquire c_lock, and it calls
231  *                     the handler only if curr_cancelled is still 0 after
232  *                     cc_lock is successfully acquired.
233  *   cc_waiting      - If a thread is waiting in callout_drain(), then
234  *                     callout_wait is nonzero.  Set only when
235  *                     cc_curr is non-NULL.
236  */
237 
238 /*
239  * Resets the execution entity tied to a specific callout cpu.
240  */
241 static void
242 cc_cce_cleanup(struct callout_cpu *cc, int direct)
243 {
244 
245 	cc_exec_curr(cc, direct) = NULL;
246 	cc_exec_cancel(cc, direct) = false;
247 	cc_exec_waiting(cc, direct) = false;
248 #ifdef SMP
249 	cc_migration_cpu(cc, direct) = CPUBLOCK;
250 	cc_migration_time(cc, direct) = 0;
251 	cc_migration_prec(cc, direct) = 0;
252 	cc_migration_func(cc, direct) = NULL;
253 	cc_migration_arg(cc, direct) = NULL;
254 #endif
255 }
256 
257 /*
258  * Checks if migration is requested by a specific callout cpu.
259  */
260 static int
261 cc_cce_migrating(struct callout_cpu *cc, int direct)
262 {
263 
264 #ifdef SMP
265 	return (cc_migration_cpu(cc, direct) != CPUBLOCK);
266 #else
267 	return (0);
268 #endif
269 }
270 
271 /*
272  * Kernel low level callwheel initialization
273  * called on the BSP during kernel startup.
274  */
275 static void
276 callout_callwheel_init(void *dummy)
277 {
278 	struct callout_cpu *cc;
279 
280 	/*
281 	 * Calculate the size of the callout wheel and the preallocated
282 	 * timeout() structures.
283 	 * XXX: Clip callout to result of previous function of maxusers
284 	 * maximum 384.  This is still huge, but acceptable.
285 	 */
286 	memset(CC_CPU(curcpu), 0, sizeof(cc_cpu));
287 	ncallout = imin(16 + maxproc + maxfiles, 18508);
288 	TUNABLE_INT_FETCH("kern.ncallout", &ncallout);
289 
290 	/*
291 	 * Calculate callout wheel size, should be next power of two higher
292 	 * than 'ncallout'.
293 	 */
294 	callwheelsize = 1 << fls(ncallout);
295 	callwheelmask = callwheelsize - 1;
296 
297 	/*
298 	 * Fetch whether we're pinning the swi's or not.
299 	 */
300 	TUNABLE_INT_FETCH("kern.pin_default_swi", &pin_default_swi);
301 	TUNABLE_INT_FETCH("kern.pin_pcpu_swi", &pin_pcpu_swi);
302 
303 	/*
304 	 * Only BSP handles timeout(9) and receives a preallocation.
305 	 *
306 	 * XXX: Once all timeout(9) consumers are converted this can
307 	 * be removed.
308 	 */
309 	timeout_cpu = PCPU_GET(cpuid);
310 	cc = CC_CPU(timeout_cpu);
311 	cc->cc_callout = malloc(ncallout * sizeof(struct callout),
312 	    M_CALLOUT, M_WAITOK);
313 	callout_cpu_init(cc, timeout_cpu);
314 }
315 SYSINIT(callwheel_init, SI_SUB_CPU, SI_ORDER_ANY, callout_callwheel_init, NULL);
316 
317 /*
318  * Initialize the per-cpu callout structures.
319  */
320 static void
321 callout_cpu_init(struct callout_cpu *cc, int cpu)
322 {
323 	struct callout *c;
324 	int i;
325 
326 	mtx_init(&cc->cc_lock, "callout", NULL, MTX_SPIN | MTX_RECURSE);
327 	SLIST_INIT(&cc->cc_callfree);
328 	cc->cc_inited = 1;
329 	cc->cc_callwheel = malloc_domainset(sizeof(struct callout_list) *
330 	    callwheelsize, M_CALLOUT,
331 	    DOMAINSET_PREF(pcpu_find(cpu)->pc_domain), M_WAITOK);
332 	for (i = 0; i < callwheelsize; i++)
333 		LIST_INIT(&cc->cc_callwheel[i]);
334 	TAILQ_INIT(&cc->cc_expireq);
335 	cc->cc_firstevent = SBT_MAX;
336 	for (i = 0; i < 2; i++)
337 		cc_cce_cleanup(cc, i);
338 	snprintf(cc->cc_ktr_event_name, sizeof(cc->cc_ktr_event_name),
339 	    "callwheel cpu %d", cpu);
340 	if (cc->cc_callout == NULL)	/* Only BSP handles timeout(9) */
341 		return;
342 	for (i = 0; i < ncallout; i++) {
343 		c = &cc->cc_callout[i];
344 		callout_init(c, 0);
345 		c->c_iflags = CALLOUT_LOCAL_ALLOC;
346 		SLIST_INSERT_HEAD(&cc->cc_callfree, c, c_links.sle);
347 	}
348 }
349 
350 #ifdef SMP
351 /*
352  * Switches the cpu tied to a specific callout.
353  * The function expects a locked incoming callout cpu and returns with
354  * locked outcoming callout cpu.
355  */
356 static struct callout_cpu *
357 callout_cpu_switch(struct callout *c, struct callout_cpu *cc, int new_cpu)
358 {
359 	struct callout_cpu *new_cc;
360 
361 	MPASS(c != NULL && cc != NULL);
362 	CC_LOCK_ASSERT(cc);
363 
364 	/*
365 	 * Avoid interrupts and preemption firing after the callout cpu
366 	 * is blocked in order to avoid deadlocks as the new thread
367 	 * may be willing to acquire the callout cpu lock.
368 	 */
369 	c->c_cpu = CPUBLOCK;
370 	spinlock_enter();
371 	CC_UNLOCK(cc);
372 	new_cc = CC_CPU(new_cpu);
373 	CC_LOCK(new_cc);
374 	spinlock_exit();
375 	c->c_cpu = new_cpu;
376 	return (new_cc);
377 }
378 #endif
379 
380 /*
381  * Start standard softclock thread.
382  */
383 static void
384 start_softclock(void *dummy)
385 {
386 	struct callout_cpu *cc;
387 	char name[MAXCOMLEN];
388 #ifdef SMP
389 	int cpu;
390 	struct intr_event *ie;
391 #endif
392 
393 	cc = CC_CPU(timeout_cpu);
394 	snprintf(name, sizeof(name), "clock (%d)", timeout_cpu);
395 	if (swi_add(&clk_intr_event, name, softclock, cc, SWI_CLOCK,
396 	    INTR_MPSAFE, &cc->cc_cookie))
397 		panic("died while creating standard software ithreads");
398 	if (pin_default_swi &&
399 	    (intr_event_bind(clk_intr_event, timeout_cpu) != 0)) {
400 		printf("%s: timeout clock couldn't be pinned to cpu %d\n",
401 		    __func__,
402 		    timeout_cpu);
403 	}
404 
405 #ifdef SMP
406 	CPU_FOREACH(cpu) {
407 		if (cpu == timeout_cpu)
408 			continue;
409 		cc = CC_CPU(cpu);
410 		cc->cc_callout = NULL;	/* Only BSP handles timeout(9). */
411 		callout_cpu_init(cc, cpu);
412 		snprintf(name, sizeof(name), "clock (%d)", cpu);
413 		ie = NULL;
414 		if (swi_add(&ie, name, softclock, cc, SWI_CLOCK,
415 		    INTR_MPSAFE, &cc->cc_cookie))
416 			panic("died while creating standard software ithreads");
417 		if (pin_pcpu_swi && (intr_event_bind(ie, cpu) != 0)) {
418 			printf("%s: per-cpu clock couldn't be pinned to "
419 			    "cpu %d\n",
420 			    __func__,
421 			    cpu);
422 		}
423 	}
424 #endif
425 }
426 SYSINIT(start_softclock, SI_SUB_SOFTINTR, SI_ORDER_FIRST, start_softclock, NULL);
427 
428 #define	CC_HASH_SHIFT	8
429 
430 static inline u_int
431 callout_hash(sbintime_t sbt)
432 {
433 
434 	return (sbt >> (32 - CC_HASH_SHIFT));
435 }
436 
437 static inline u_int
438 callout_get_bucket(sbintime_t sbt)
439 {
440 
441 	return (callout_hash(sbt) & callwheelmask);
442 }
443 
444 void
445 callout_process(sbintime_t now)
446 {
447 	struct callout *tmp, *tmpn;
448 	struct callout_cpu *cc;
449 	struct callout_list *sc;
450 	sbintime_t first, last, max, tmp_max;
451 	uint32_t lookahead;
452 	u_int firstb, lastb, nowb;
453 #ifdef CALLOUT_PROFILING
454 	int depth_dir = 0, mpcalls_dir = 0, lockcalls_dir = 0;
455 #endif
456 
457 	cc = CC_SELF();
458 	mtx_lock_spin_flags(&cc->cc_lock, MTX_QUIET);
459 
460 	/* Compute the buckets of the last scan and present times. */
461 	firstb = callout_hash(cc->cc_lastscan);
462 	cc->cc_lastscan = now;
463 	nowb = callout_hash(now);
464 
465 	/* Compute the last bucket and minimum time of the bucket after it. */
466 	if (nowb == firstb)
467 		lookahead = (SBT_1S / 16);
468 	else if (nowb - firstb == 1)
469 		lookahead = (SBT_1S / 8);
470 	else
471 		lookahead = (SBT_1S / 2);
472 	first = last = now;
473 	first += (lookahead / 2);
474 	last += lookahead;
475 	last &= (0xffffffffffffffffLLU << (32 - CC_HASH_SHIFT));
476 	lastb = callout_hash(last) - 1;
477 	max = last;
478 
479 	/*
480 	 * Check if we wrapped around the entire wheel from the last scan.
481 	 * In case, we need to scan entirely the wheel for pending callouts.
482 	 */
483 	if (lastb - firstb >= callwheelsize) {
484 		lastb = firstb + callwheelsize - 1;
485 		if (nowb - firstb >= callwheelsize)
486 			nowb = lastb;
487 	}
488 
489 	/* Iterate callwheel from firstb to nowb and then up to lastb. */
490 	do {
491 		sc = &cc->cc_callwheel[firstb & callwheelmask];
492 		tmp = LIST_FIRST(sc);
493 		while (tmp != NULL) {
494 			/* Run the callout if present time within allowed. */
495 			if (tmp->c_time <= now) {
496 				/*
497 				 * Consumer told us the callout may be run
498 				 * directly from hardware interrupt context.
499 				 */
500 				if (tmp->c_iflags & CALLOUT_DIRECT) {
501 #ifdef CALLOUT_PROFILING
502 					++depth_dir;
503 #endif
504 					cc_exec_next(cc) =
505 					    LIST_NEXT(tmp, c_links.le);
506 					cc->cc_bucket = firstb & callwheelmask;
507 					LIST_REMOVE(tmp, c_links.le);
508 					softclock_call_cc(tmp, cc,
509 #ifdef CALLOUT_PROFILING
510 					    &mpcalls_dir, &lockcalls_dir, NULL,
511 #endif
512 					    1);
513 					tmp = cc_exec_next(cc);
514 					cc_exec_next(cc) = NULL;
515 				} else {
516 					tmpn = LIST_NEXT(tmp, c_links.le);
517 					LIST_REMOVE(tmp, c_links.le);
518 					TAILQ_INSERT_TAIL(&cc->cc_expireq,
519 					    tmp, c_links.tqe);
520 					tmp->c_iflags |= CALLOUT_PROCESSED;
521 					tmp = tmpn;
522 				}
523 				continue;
524 			}
525 			/* Skip events from distant future. */
526 			if (tmp->c_time >= max)
527 				goto next;
528 			/*
529 			 * Event minimal time is bigger than present maximal
530 			 * time, so it cannot be aggregated.
531 			 */
532 			if (tmp->c_time > last) {
533 				lastb = nowb;
534 				goto next;
535 			}
536 			/* Update first and last time, respecting this event. */
537 			if (tmp->c_time < first)
538 				first = tmp->c_time;
539 			tmp_max = tmp->c_time + tmp->c_precision;
540 			if (tmp_max < last)
541 				last = tmp_max;
542 next:
543 			tmp = LIST_NEXT(tmp, c_links.le);
544 		}
545 		/* Proceed with the next bucket. */
546 		firstb++;
547 		/*
548 		 * Stop if we looked after present time and found
549 		 * some event we can't execute at now.
550 		 * Stop if we looked far enough into the future.
551 		 */
552 	} while (((int)(firstb - lastb)) <= 0);
553 	cc->cc_firstevent = last;
554 #ifndef NO_EVENTTIMERS
555 	cpu_new_callout(curcpu, last, first);
556 #endif
557 #ifdef CALLOUT_PROFILING
558 	avg_depth_dir += (depth_dir * 1000 - avg_depth_dir) >> 8;
559 	avg_mpcalls_dir += (mpcalls_dir * 1000 - avg_mpcalls_dir) >> 8;
560 	avg_lockcalls_dir += (lockcalls_dir * 1000 - avg_lockcalls_dir) >> 8;
561 #endif
562 	mtx_unlock_spin_flags(&cc->cc_lock, MTX_QUIET);
563 	/*
564 	 * swi_sched acquires the thread lock, so we don't want to call it
565 	 * with cc_lock held; incorrect locking order.
566 	 */
567 	if (!TAILQ_EMPTY(&cc->cc_expireq))
568 		swi_sched(cc->cc_cookie, 0);
569 }
570 
571 static struct callout_cpu *
572 callout_lock(struct callout *c)
573 {
574 	struct callout_cpu *cc;
575 	int cpu;
576 
577 	for (;;) {
578 		cpu = c->c_cpu;
579 #ifdef SMP
580 		if (cpu == CPUBLOCK) {
581 			while (c->c_cpu == CPUBLOCK)
582 				cpu_spinwait();
583 			continue;
584 		}
585 #endif
586 		cc = CC_CPU(cpu);
587 		CC_LOCK(cc);
588 		if (cpu == c->c_cpu)
589 			break;
590 		CC_UNLOCK(cc);
591 	}
592 	return (cc);
593 }
594 
595 static void
596 callout_cc_add(struct callout *c, struct callout_cpu *cc,
597     sbintime_t sbt, sbintime_t precision, void (*func)(void *),
598     void *arg, int cpu, int flags)
599 {
600 	int bucket;
601 
602 	CC_LOCK_ASSERT(cc);
603 	if (sbt < cc->cc_lastscan)
604 		sbt = cc->cc_lastscan;
605 	c->c_arg = arg;
606 	c->c_iflags |= CALLOUT_PENDING;
607 	c->c_iflags &= ~CALLOUT_PROCESSED;
608 	c->c_flags |= CALLOUT_ACTIVE;
609 	if (flags & C_DIRECT_EXEC)
610 		c->c_iflags |= CALLOUT_DIRECT;
611 	c->c_func = func;
612 	c->c_time = sbt;
613 	c->c_precision = precision;
614 	bucket = callout_get_bucket(c->c_time);
615 	CTR3(KTR_CALLOUT, "precision set for %p: %d.%08x",
616 	    c, (int)(c->c_precision >> 32),
617 	    (u_int)(c->c_precision & 0xffffffff));
618 	LIST_INSERT_HEAD(&cc->cc_callwheel[bucket], c, c_links.le);
619 	if (cc->cc_bucket == bucket)
620 		cc_exec_next(cc) = c;
621 #ifndef NO_EVENTTIMERS
622 	/*
623 	 * Inform the eventtimers(4) subsystem there's a new callout
624 	 * that has been inserted, but only if really required.
625 	 */
626 	if (SBT_MAX - c->c_time < c->c_precision)
627 		c->c_precision = SBT_MAX - c->c_time;
628 	sbt = c->c_time + c->c_precision;
629 	if (sbt < cc->cc_firstevent) {
630 		cc->cc_firstevent = sbt;
631 		cpu_new_callout(cpu, sbt, c->c_time);
632 	}
633 #endif
634 }
635 
636 static void
637 callout_cc_del(struct callout *c, struct callout_cpu *cc)
638 {
639 
640 	if ((c->c_iflags & CALLOUT_LOCAL_ALLOC) == 0)
641 		return;
642 	c->c_func = NULL;
643 	SLIST_INSERT_HEAD(&cc->cc_callfree, c, c_links.sle);
644 }
645 
646 static void
647 softclock_call_cc(struct callout *c, struct callout_cpu *cc,
648 #ifdef CALLOUT_PROFILING
649     int *mpcalls, int *lockcalls, int *gcalls,
650 #endif
651     int direct)
652 {
653 	struct rm_priotracker tracker;
654 	void (*c_func)(void *);
655 	void *c_arg;
656 	struct lock_class *class;
657 	struct lock_object *c_lock;
658 	uintptr_t lock_status;
659 	int c_iflags;
660 #ifdef SMP
661 	struct callout_cpu *new_cc;
662 	void (*new_func)(void *);
663 	void *new_arg;
664 	int flags, new_cpu;
665 	sbintime_t new_prec, new_time;
666 #endif
667 #if defined(DIAGNOSTIC) || defined(CALLOUT_PROFILING)
668 	sbintime_t sbt1, sbt2;
669 	struct timespec ts2;
670 	static sbintime_t maxdt = 2 * SBT_1MS;	/* 2 msec */
671 	static timeout_t *lastfunc;
672 #endif
673 
674 	KASSERT((c->c_iflags & CALLOUT_PENDING) == CALLOUT_PENDING,
675 	    ("softclock_call_cc: pend %p %x", c, c->c_iflags));
676 	KASSERT((c->c_flags & CALLOUT_ACTIVE) == CALLOUT_ACTIVE,
677 	    ("softclock_call_cc: act %p %x", c, c->c_flags));
678 	class = (c->c_lock != NULL) ? LOCK_CLASS(c->c_lock) : NULL;
679 	lock_status = 0;
680 	if (c->c_flags & CALLOUT_SHAREDLOCK) {
681 		if (class == &lock_class_rm)
682 			lock_status = (uintptr_t)&tracker;
683 		else
684 			lock_status = 1;
685 	}
686 	c_lock = c->c_lock;
687 	c_func = c->c_func;
688 	c_arg = c->c_arg;
689 	c_iflags = c->c_iflags;
690 	if (c->c_iflags & CALLOUT_LOCAL_ALLOC)
691 		c->c_iflags = CALLOUT_LOCAL_ALLOC;
692 	else
693 		c->c_iflags &= ~CALLOUT_PENDING;
694 
695 	cc_exec_curr(cc, direct) = c;
696 	cc_exec_last_func(cc, direct) = c_func;
697 	cc_exec_last_arg(cc, direct) = c_arg;
698 	cc_exec_cancel(cc, direct) = false;
699 	cc_exec_drain(cc, direct) = NULL;
700 	CC_UNLOCK(cc);
701 	if (c_lock != NULL) {
702 		class->lc_lock(c_lock, lock_status);
703 		/*
704 		 * The callout may have been cancelled
705 		 * while we switched locks.
706 		 */
707 		if (cc_exec_cancel(cc, direct)) {
708 			class->lc_unlock(c_lock);
709 			goto skip;
710 		}
711 		/* The callout cannot be stopped now. */
712 		cc_exec_cancel(cc, direct) = true;
713 		if (c_lock == &Giant.lock_object) {
714 #ifdef CALLOUT_PROFILING
715 			(*gcalls)++;
716 #endif
717 			CTR3(KTR_CALLOUT, "callout giant %p func %p arg %p",
718 			    c, c_func, c_arg);
719 		} else {
720 #ifdef CALLOUT_PROFILING
721 			(*lockcalls)++;
722 #endif
723 			CTR3(KTR_CALLOUT, "callout lock %p func %p arg %p",
724 			    c, c_func, c_arg);
725 		}
726 	} else {
727 #ifdef CALLOUT_PROFILING
728 		(*mpcalls)++;
729 #endif
730 		CTR3(KTR_CALLOUT, "callout %p func %p arg %p",
731 		    c, c_func, c_arg);
732 	}
733 	KTR_STATE3(KTR_SCHED, "callout", cc->cc_ktr_event_name, "running",
734 	    "func:%p", c_func, "arg:%p", c_arg, "direct:%d", direct);
735 #if defined(DIAGNOSTIC) || defined(CALLOUT_PROFILING)
736 	sbt1 = sbinuptime();
737 #endif
738 	THREAD_NO_SLEEPING();
739 	SDT_PROBE1(callout_execute, , , callout__start, c);
740 	c_func(c_arg);
741 	SDT_PROBE1(callout_execute, , , callout__end, c);
742 	THREAD_SLEEPING_OK();
743 #if defined(DIAGNOSTIC) || defined(CALLOUT_PROFILING)
744 	sbt2 = sbinuptime();
745 	sbt2 -= sbt1;
746 	if (sbt2 > maxdt) {
747 		if (lastfunc != c_func || sbt2 > maxdt * 2) {
748 			ts2 = sbttots(sbt2);
749 			printf(
750 		"Expensive timeout(9) function: %p(%p) %jd.%09ld s\n",
751 			    c_func, c_arg, (intmax_t)ts2.tv_sec, ts2.tv_nsec);
752 		}
753 		maxdt = sbt2;
754 		lastfunc = c_func;
755 	}
756 #endif
757 	KTR_STATE0(KTR_SCHED, "callout", cc->cc_ktr_event_name, "idle");
758 	CTR1(KTR_CALLOUT, "callout %p finished", c);
759 	if ((c_iflags & CALLOUT_RETURNUNLOCKED) == 0)
760 		class->lc_unlock(c_lock);
761 skip:
762 	CC_LOCK(cc);
763 	KASSERT(cc_exec_curr(cc, direct) == c, ("mishandled cc_curr"));
764 	cc_exec_curr(cc, direct) = NULL;
765 	if (cc_exec_drain(cc, direct)) {
766 		void (*drain)(void *);
767 
768 		drain = cc_exec_drain(cc, direct);
769 		cc_exec_drain(cc, direct) = NULL;
770 		CC_UNLOCK(cc);
771 		drain(c_arg);
772 		CC_LOCK(cc);
773 	}
774 	if (cc_exec_waiting(cc, direct)) {
775 		/*
776 		 * There is someone waiting for the
777 		 * callout to complete.
778 		 * If the callout was scheduled for
779 		 * migration just cancel it.
780 		 */
781 		if (cc_cce_migrating(cc, direct)) {
782 			cc_cce_cleanup(cc, direct);
783 
784 			/*
785 			 * It should be assert here that the callout is not
786 			 * destroyed but that is not easy.
787 			 */
788 			c->c_iflags &= ~CALLOUT_DFRMIGRATION;
789 		}
790 		cc_exec_waiting(cc, direct) = false;
791 		CC_UNLOCK(cc);
792 		wakeup(&cc_exec_waiting(cc, direct));
793 		CC_LOCK(cc);
794 	} else if (cc_cce_migrating(cc, direct)) {
795 		KASSERT((c_iflags & CALLOUT_LOCAL_ALLOC) == 0,
796 		    ("Migrating legacy callout %p", c));
797 #ifdef SMP
798 		/*
799 		 * If the callout was scheduled for
800 		 * migration just perform it now.
801 		 */
802 		new_cpu = cc_migration_cpu(cc, direct);
803 		new_time = cc_migration_time(cc, direct);
804 		new_prec = cc_migration_prec(cc, direct);
805 		new_func = cc_migration_func(cc, direct);
806 		new_arg = cc_migration_arg(cc, direct);
807 		cc_cce_cleanup(cc, direct);
808 
809 		/*
810 		 * It should be assert here that the callout is not destroyed
811 		 * but that is not easy.
812 		 *
813 		 * As first thing, handle deferred callout stops.
814 		 */
815 		if (!callout_migrating(c)) {
816 			CTR3(KTR_CALLOUT,
817 			     "deferred cancelled %p func %p arg %p",
818 			     c, new_func, new_arg);
819 			callout_cc_del(c, cc);
820 			return;
821 		}
822 		c->c_iflags &= ~CALLOUT_DFRMIGRATION;
823 
824 		new_cc = callout_cpu_switch(c, cc, new_cpu);
825 		flags = (direct) ? C_DIRECT_EXEC : 0;
826 		callout_cc_add(c, new_cc, new_time, new_prec, new_func,
827 		    new_arg, new_cpu, flags);
828 		CC_UNLOCK(new_cc);
829 		CC_LOCK(cc);
830 #else
831 		panic("migration should not happen");
832 #endif
833 	}
834 	/*
835 	 * If the current callout is locally allocated (from
836 	 * timeout(9)) then put it on the freelist.
837 	 *
838 	 * Note: we need to check the cached copy of c_iflags because
839 	 * if it was not local, then it's not safe to deref the
840 	 * callout pointer.
841 	 */
842 	KASSERT((c_iflags & CALLOUT_LOCAL_ALLOC) == 0 ||
843 	    c->c_iflags == CALLOUT_LOCAL_ALLOC,
844 	    ("corrupted callout"));
845 	if (c_iflags & CALLOUT_LOCAL_ALLOC)
846 		callout_cc_del(c, cc);
847 }
848 
849 /*
850  * The callout mechanism is based on the work of Adam M. Costello and
851  * George Varghese, published in a technical report entitled "Redesigning
852  * the BSD Callout and Timer Facilities" and modified slightly for inclusion
853  * in FreeBSD by Justin T. Gibbs.  The original work on the data structures
854  * used in this implementation was published by G. Varghese and T. Lauck in
855  * the paper "Hashed and Hierarchical Timing Wheels: Data Structures for
856  * the Efficient Implementation of a Timer Facility" in the Proceedings of
857  * the 11th ACM Annual Symposium on Operating Systems Principles,
858  * Austin, Texas Nov 1987.
859  */
860 
861 /*
862  * Software (low priority) clock interrupt.
863  * Run periodic events from timeout queue.
864  */
865 void
866 softclock(void *arg)
867 {
868 	struct callout_cpu *cc;
869 	struct callout *c;
870 #ifdef CALLOUT_PROFILING
871 	int depth = 0, gcalls = 0, lockcalls = 0, mpcalls = 0;
872 #endif
873 
874 	cc = (struct callout_cpu *)arg;
875 	CC_LOCK(cc);
876 	while ((c = TAILQ_FIRST(&cc->cc_expireq)) != NULL) {
877 		TAILQ_REMOVE(&cc->cc_expireq, c, c_links.tqe);
878 		softclock_call_cc(c, cc,
879 #ifdef CALLOUT_PROFILING
880 		    &mpcalls, &lockcalls, &gcalls,
881 #endif
882 		    0);
883 #ifdef CALLOUT_PROFILING
884 		++depth;
885 #endif
886 	}
887 #ifdef CALLOUT_PROFILING
888 	avg_depth += (depth * 1000 - avg_depth) >> 8;
889 	avg_mpcalls += (mpcalls * 1000 - avg_mpcalls) >> 8;
890 	avg_lockcalls += (lockcalls * 1000 - avg_lockcalls) >> 8;
891 	avg_gcalls += (gcalls * 1000 - avg_gcalls) >> 8;
892 #endif
893 	CC_UNLOCK(cc);
894 }
895 
896 /*
897  * timeout --
898  *	Execute a function after a specified length of time.
899  *
900  * untimeout --
901  *	Cancel previous timeout function call.
902  *
903  * callout_handle_init --
904  *	Initialize a handle so that using it with untimeout is benign.
905  *
906  *	See AT&T BCI Driver Reference Manual for specification.  This
907  *	implementation differs from that one in that although an
908  *	identification value is returned from timeout, the original
909  *	arguments to timeout as well as the identifier are used to
910  *	identify entries for untimeout.
911  */
912 struct callout_handle
913 timeout(timeout_t *ftn, void *arg, int to_ticks)
914 {
915 	struct callout_cpu *cc;
916 	struct callout *new;
917 	struct callout_handle handle;
918 
919 	cc = CC_CPU(timeout_cpu);
920 	CC_LOCK(cc);
921 	/* Fill in the next free callout structure. */
922 	new = SLIST_FIRST(&cc->cc_callfree);
923 	if (new == NULL)
924 		/* XXX Attempt to malloc first */
925 		panic("timeout table full");
926 	SLIST_REMOVE_HEAD(&cc->cc_callfree, c_links.sle);
927 	callout_reset(new, to_ticks, ftn, arg);
928 	handle.callout = new;
929 	CC_UNLOCK(cc);
930 
931 	return (handle);
932 }
933 
934 void
935 untimeout(timeout_t *ftn, void *arg, struct callout_handle handle)
936 {
937 	struct callout_cpu *cc;
938 
939 	/*
940 	 * Check for a handle that was initialized
941 	 * by callout_handle_init, but never used
942 	 * for a real timeout.
943 	 */
944 	if (handle.callout == NULL)
945 		return;
946 
947 	cc = callout_lock(handle.callout);
948 	if (handle.callout->c_func == ftn && handle.callout->c_arg == arg)
949 		callout_stop(handle.callout);
950 	CC_UNLOCK(cc);
951 }
952 
953 void
954 callout_handle_init(struct callout_handle *handle)
955 {
956 	handle->callout = NULL;
957 }
958 
959 void
960 callout_when(sbintime_t sbt, sbintime_t precision, int flags,
961     sbintime_t *res, sbintime_t *prec_res)
962 {
963 	sbintime_t to_sbt, to_pr;
964 
965 	if ((flags & (C_ABSOLUTE | C_PRECALC)) != 0) {
966 		*res = sbt;
967 		*prec_res = precision;
968 		return;
969 	}
970 	if ((flags & C_HARDCLOCK) != 0 && sbt < tick_sbt)
971 		sbt = tick_sbt;
972 	if ((flags & C_HARDCLOCK) != 0 ||
973 #ifdef NO_EVENTTIMERS
974 	    sbt >= sbt_timethreshold) {
975 		to_sbt = getsbinuptime();
976 
977 		/* Add safety belt for the case of hz > 1000. */
978 		to_sbt += tc_tick_sbt - tick_sbt;
979 #else
980 	    sbt >= sbt_tickthreshold) {
981 		/*
982 		 * Obtain the time of the last hardclock() call on
983 		 * this CPU directly from the kern_clocksource.c.
984 		 * This value is per-CPU, but it is equal for all
985 		 * active ones.
986 		 */
987 #ifdef __LP64__
988 		to_sbt = DPCPU_GET(hardclocktime);
989 #else
990 		spinlock_enter();
991 		to_sbt = DPCPU_GET(hardclocktime);
992 		spinlock_exit();
993 #endif
994 #endif
995 		if (cold && to_sbt == 0)
996 			to_sbt = sbinuptime();
997 		if ((flags & C_HARDCLOCK) == 0)
998 			to_sbt += tick_sbt;
999 	} else
1000 		to_sbt = sbinuptime();
1001 	if (SBT_MAX - to_sbt < sbt)
1002 		to_sbt = SBT_MAX;
1003 	else
1004 		to_sbt += sbt;
1005 	*res = to_sbt;
1006 	to_pr = ((C_PRELGET(flags) < 0) ? sbt >> tc_precexp :
1007 	    sbt >> C_PRELGET(flags));
1008 	*prec_res = to_pr > precision ? to_pr : precision;
1009 }
1010 
1011 /*
1012  * New interface; clients allocate their own callout structures.
1013  *
1014  * callout_reset() - establish or change a timeout
1015  * callout_stop() - disestablish a timeout
1016  * callout_init() - initialize a callout structure so that it can
1017  *	safely be passed to callout_reset() and callout_stop()
1018  *
1019  * <sys/callout.h> defines three convenience macros:
1020  *
1021  * callout_active() - returns truth if callout has not been stopped,
1022  *	drained, or deactivated since the last time the callout was
1023  *	reset.
1024  * callout_pending() - returns truth if callout is still waiting for timeout
1025  * callout_deactivate() - marks the callout as having been serviced
1026  */
1027 int
1028 callout_reset_sbt_on(struct callout *c, sbintime_t sbt, sbintime_t prec,
1029     void (*ftn)(void *), void *arg, int cpu, int flags)
1030 {
1031 	sbintime_t to_sbt, precision;
1032 	struct callout_cpu *cc;
1033 	int cancelled, direct;
1034 	int ignore_cpu=0;
1035 
1036 	cancelled = 0;
1037 	if (cpu == -1) {
1038 		ignore_cpu = 1;
1039 	} else if ((cpu >= MAXCPU) ||
1040 		   ((CC_CPU(cpu))->cc_inited == 0)) {
1041 		/* Invalid CPU spec */
1042 		panic("Invalid CPU in callout %d", cpu);
1043 	}
1044 	callout_when(sbt, prec, flags, &to_sbt, &precision);
1045 
1046 	/*
1047 	 * This flag used to be added by callout_cc_add, but the
1048 	 * first time you call this we could end up with the
1049 	 * wrong direct flag if we don't do it before we add.
1050 	 */
1051 	if (flags & C_DIRECT_EXEC) {
1052 		direct = 1;
1053 	} else {
1054 		direct = 0;
1055 	}
1056 	KASSERT(!direct || c->c_lock == NULL,
1057 	    ("%s: direct callout %p has lock", __func__, c));
1058 	cc = callout_lock(c);
1059 	/*
1060 	 * Don't allow migration of pre-allocated callouts lest they
1061 	 * become unbalanced or handle the case where the user does
1062 	 * not care.
1063 	 */
1064 	if ((c->c_iflags & CALLOUT_LOCAL_ALLOC) ||
1065 	    ignore_cpu) {
1066 		cpu = c->c_cpu;
1067 	}
1068 
1069 	if (cc_exec_curr(cc, direct) == c) {
1070 		/*
1071 		 * We're being asked to reschedule a callout which is
1072 		 * currently in progress.  If there is a lock then we
1073 		 * can cancel the callout if it has not really started.
1074 		 */
1075 		if (c->c_lock != NULL && !cc_exec_cancel(cc, direct))
1076 			cancelled = cc_exec_cancel(cc, direct) = true;
1077 		if (cc_exec_waiting(cc, direct) || cc_exec_drain(cc, direct)) {
1078 			/*
1079 			 * Someone has called callout_drain to kill this
1080 			 * callout.  Don't reschedule.
1081 			 */
1082 			CTR4(KTR_CALLOUT, "%s %p func %p arg %p",
1083 			    cancelled ? "cancelled" : "failed to cancel",
1084 			    c, c->c_func, c->c_arg);
1085 			CC_UNLOCK(cc);
1086 			return (cancelled);
1087 		}
1088 #ifdef SMP
1089 		if (callout_migrating(c)) {
1090 			/*
1091 			 * This only occurs when a second callout_reset_sbt_on
1092 			 * is made after a previous one moved it into
1093 			 * deferred migration (below). Note we do *not* change
1094 			 * the prev_cpu even though the previous target may
1095 			 * be different.
1096 			 */
1097 			cc_migration_cpu(cc, direct) = cpu;
1098 			cc_migration_time(cc, direct) = to_sbt;
1099 			cc_migration_prec(cc, direct) = precision;
1100 			cc_migration_func(cc, direct) = ftn;
1101 			cc_migration_arg(cc, direct) = arg;
1102 			cancelled = 1;
1103 			CC_UNLOCK(cc);
1104 			return (cancelled);
1105 		}
1106 #endif
1107 	}
1108 	if (c->c_iflags & CALLOUT_PENDING) {
1109 		if ((c->c_iflags & CALLOUT_PROCESSED) == 0) {
1110 			if (cc_exec_next(cc) == c)
1111 				cc_exec_next(cc) = LIST_NEXT(c, c_links.le);
1112 			LIST_REMOVE(c, c_links.le);
1113 		} else {
1114 			TAILQ_REMOVE(&cc->cc_expireq, c, c_links.tqe);
1115 		}
1116 		cancelled = 1;
1117 		c->c_iflags &= ~ CALLOUT_PENDING;
1118 		c->c_flags &= ~ CALLOUT_ACTIVE;
1119 	}
1120 
1121 #ifdef SMP
1122 	/*
1123 	 * If the callout must migrate try to perform it immediately.
1124 	 * If the callout is currently running, just defer the migration
1125 	 * to a more appropriate moment.
1126 	 */
1127 	if (c->c_cpu != cpu) {
1128 		if (cc_exec_curr(cc, direct) == c) {
1129 			/*
1130 			 * Pending will have been removed since we are
1131 			 * actually executing the callout on another
1132 			 * CPU. That callout should be waiting on the
1133 			 * lock the caller holds. If we set both
1134 			 * active/and/pending after we return and the
1135 			 * lock on the executing callout proceeds, it
1136 			 * will then see pending is true and return.
1137 			 * At the return from the actual callout execution
1138 			 * the migration will occur in softclock_call_cc
1139 			 * and this new callout will be placed on the
1140 			 * new CPU via a call to callout_cpu_switch() which
1141 			 * will get the lock on the right CPU followed
1142 			 * by a call callout_cc_add() which will add it there.
1143 			 * (see above in softclock_call_cc()).
1144 			 */
1145 			cc_migration_cpu(cc, direct) = cpu;
1146 			cc_migration_time(cc, direct) = to_sbt;
1147 			cc_migration_prec(cc, direct) = precision;
1148 			cc_migration_func(cc, direct) = ftn;
1149 			cc_migration_arg(cc, direct) = arg;
1150 			c->c_iflags |= (CALLOUT_DFRMIGRATION | CALLOUT_PENDING);
1151 			c->c_flags |= CALLOUT_ACTIVE;
1152 			CTR6(KTR_CALLOUT,
1153 		    "migration of %p func %p arg %p in %d.%08x to %u deferred",
1154 			    c, c->c_func, c->c_arg, (int)(to_sbt >> 32),
1155 			    (u_int)(to_sbt & 0xffffffff), cpu);
1156 			CC_UNLOCK(cc);
1157 			return (cancelled);
1158 		}
1159 		cc = callout_cpu_switch(c, cc, cpu);
1160 	}
1161 #endif
1162 
1163 	callout_cc_add(c, cc, to_sbt, precision, ftn, arg, cpu, flags);
1164 	CTR6(KTR_CALLOUT, "%sscheduled %p func %p arg %p in %d.%08x",
1165 	    cancelled ? "re" : "", c, c->c_func, c->c_arg, (int)(to_sbt >> 32),
1166 	    (u_int)(to_sbt & 0xffffffff));
1167 	CC_UNLOCK(cc);
1168 
1169 	return (cancelled);
1170 }
1171 
1172 /*
1173  * Common idioms that can be optimized in the future.
1174  */
1175 int
1176 callout_schedule_on(struct callout *c, int to_ticks, int cpu)
1177 {
1178 	return callout_reset_on(c, to_ticks, c->c_func, c->c_arg, cpu);
1179 }
1180 
1181 int
1182 callout_schedule(struct callout *c, int to_ticks)
1183 {
1184 	return callout_reset_on(c, to_ticks, c->c_func, c->c_arg, c->c_cpu);
1185 }
1186 
1187 int
1188 _callout_stop_safe(struct callout *c, int flags, void (*drain)(void *))
1189 {
1190 	struct callout_cpu *cc, *old_cc;
1191 	struct lock_class *class;
1192 	int direct, sq_locked, use_lock;
1193 	int cancelled, not_on_a_list;
1194 
1195 	if ((flags & CS_DRAIN) != 0)
1196 		WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK, c->c_lock,
1197 		    "calling %s", __func__);
1198 
1199 	/*
1200 	 * Some old subsystems don't hold Giant while running a callout_stop(),
1201 	 * so just discard this check for the moment.
1202 	 */
1203 	if ((flags & CS_DRAIN) == 0 && c->c_lock != NULL) {
1204 		if (c->c_lock == &Giant.lock_object)
1205 			use_lock = mtx_owned(&Giant);
1206 		else {
1207 			use_lock = 1;
1208 			class = LOCK_CLASS(c->c_lock);
1209 			class->lc_assert(c->c_lock, LA_XLOCKED);
1210 		}
1211 	} else
1212 		use_lock = 0;
1213 	if (c->c_iflags & CALLOUT_DIRECT) {
1214 		direct = 1;
1215 	} else {
1216 		direct = 0;
1217 	}
1218 	sq_locked = 0;
1219 	old_cc = NULL;
1220 again:
1221 	cc = callout_lock(c);
1222 
1223 	if ((c->c_iflags & (CALLOUT_DFRMIGRATION | CALLOUT_PENDING)) ==
1224 	    (CALLOUT_DFRMIGRATION | CALLOUT_PENDING) &&
1225 	    ((c->c_flags & CALLOUT_ACTIVE) == CALLOUT_ACTIVE)) {
1226 		/*
1227 		 * Special case where this slipped in while we
1228 		 * were migrating *as* the callout is about to
1229 		 * execute. The caller probably holds the lock
1230 		 * the callout wants.
1231 		 *
1232 		 * Get rid of the migration first. Then set
1233 		 * the flag that tells this code *not* to
1234 		 * try to remove it from any lists (its not
1235 		 * on one yet). When the callout wheel runs,
1236 		 * it will ignore this callout.
1237 		 */
1238 		c->c_iflags &= ~CALLOUT_PENDING;
1239 		c->c_flags &= ~CALLOUT_ACTIVE;
1240 		not_on_a_list = 1;
1241 	} else {
1242 		not_on_a_list = 0;
1243 	}
1244 
1245 	/*
1246 	 * If the callout was migrating while the callout cpu lock was
1247 	 * dropped,  just drop the sleepqueue lock and check the states
1248 	 * again.
1249 	 */
1250 	if (sq_locked != 0 && cc != old_cc) {
1251 #ifdef SMP
1252 		CC_UNLOCK(cc);
1253 		sleepq_release(&cc_exec_waiting(old_cc, direct));
1254 		sq_locked = 0;
1255 		old_cc = NULL;
1256 		goto again;
1257 #else
1258 		panic("migration should not happen");
1259 #endif
1260 	}
1261 
1262 	/*
1263 	 * If the callout is running, try to stop it or drain it.
1264 	 */
1265 	if (cc_exec_curr(cc, direct) == c) {
1266 		/*
1267 		 * Succeed we to stop it or not, we must clear the
1268 		 * active flag - this is what API users expect.  If we're
1269 		 * draining and the callout is currently executing, first wait
1270 		 * until it finishes.
1271 		 */
1272 		if ((flags & CS_DRAIN) == 0)
1273 			c->c_flags &= ~CALLOUT_ACTIVE;
1274 
1275 		if ((flags & CS_DRAIN) != 0) {
1276 			/*
1277 			 * The current callout is running (or just
1278 			 * about to run) and blocking is allowed, so
1279 			 * just wait for the current invocation to
1280 			 * finish.
1281 			 */
1282 			while (cc_exec_curr(cc, direct) == c) {
1283 				/*
1284 				 * Use direct calls to sleepqueue interface
1285 				 * instead of cv/msleep in order to avoid
1286 				 * a LOR between cc_lock and sleepqueue
1287 				 * chain spinlocks.  This piece of code
1288 				 * emulates a msleep_spin() call actually.
1289 				 *
1290 				 * If we already have the sleepqueue chain
1291 				 * locked, then we can safely block.  If we
1292 				 * don't already have it locked, however,
1293 				 * we have to drop the cc_lock to lock
1294 				 * it.  This opens several races, so we
1295 				 * restart at the beginning once we have
1296 				 * both locks.  If nothing has changed, then
1297 				 * we will end up back here with sq_locked
1298 				 * set.
1299 				 */
1300 				if (!sq_locked) {
1301 					CC_UNLOCK(cc);
1302 					sleepq_lock(
1303 					    &cc_exec_waiting(cc, direct));
1304 					sq_locked = 1;
1305 					old_cc = cc;
1306 					goto again;
1307 				}
1308 
1309 				/*
1310 				 * Migration could be cancelled here, but
1311 				 * as long as it is still not sure when it
1312 				 * will be packed up, just let softclock()
1313 				 * take care of it.
1314 				 */
1315 				cc_exec_waiting(cc, direct) = true;
1316 				DROP_GIANT();
1317 				CC_UNLOCK(cc);
1318 				sleepq_add(
1319 				    &cc_exec_waiting(cc, direct),
1320 				    &cc->cc_lock.lock_object, "codrain",
1321 				    SLEEPQ_SLEEP, 0);
1322 				sleepq_wait(
1323 				    &cc_exec_waiting(cc, direct),
1324 					     0);
1325 				sq_locked = 0;
1326 				old_cc = NULL;
1327 
1328 				/* Reacquire locks previously released. */
1329 				PICKUP_GIANT();
1330 				CC_LOCK(cc);
1331 			}
1332 			c->c_flags &= ~CALLOUT_ACTIVE;
1333 		} else if (use_lock &&
1334 			   !cc_exec_cancel(cc, direct) && (drain == NULL)) {
1335 
1336 			/*
1337 			 * The current callout is waiting for its
1338 			 * lock which we hold.  Cancel the callout
1339 			 * and return.  After our caller drops the
1340 			 * lock, the callout will be skipped in
1341 			 * softclock(). This *only* works with a
1342 			 * callout_stop() *not* callout_drain() or
1343 			 * callout_async_drain().
1344 			 */
1345 			cc_exec_cancel(cc, direct) = true;
1346 			CTR3(KTR_CALLOUT, "cancelled %p func %p arg %p",
1347 			    c, c->c_func, c->c_arg);
1348 			KASSERT(!cc_cce_migrating(cc, direct),
1349 			    ("callout wrongly scheduled for migration"));
1350 			if (callout_migrating(c)) {
1351 				c->c_iflags &= ~CALLOUT_DFRMIGRATION;
1352 #ifdef SMP
1353 				cc_migration_cpu(cc, direct) = CPUBLOCK;
1354 				cc_migration_time(cc, direct) = 0;
1355 				cc_migration_prec(cc, direct) = 0;
1356 				cc_migration_func(cc, direct) = NULL;
1357 				cc_migration_arg(cc, direct) = NULL;
1358 #endif
1359 			}
1360 			CC_UNLOCK(cc);
1361 			KASSERT(!sq_locked, ("sleepqueue chain locked"));
1362 			return (1);
1363 		} else if (callout_migrating(c)) {
1364 			/*
1365 			 * The callout is currently being serviced
1366 			 * and the "next" callout is scheduled at
1367 			 * its completion with a migration. We remove
1368 			 * the migration flag so it *won't* get rescheduled,
1369 			 * but we can't stop the one thats running so
1370 			 * we return 0.
1371 			 */
1372 			c->c_iflags &= ~CALLOUT_DFRMIGRATION;
1373 #ifdef SMP
1374 			/*
1375 			 * We can't call cc_cce_cleanup here since
1376 			 * if we do it will remove .ce_curr and
1377 			 * its still running. This will prevent a
1378 			 * reschedule of the callout when the
1379 			 * execution completes.
1380 			 */
1381 			cc_migration_cpu(cc, direct) = CPUBLOCK;
1382 			cc_migration_time(cc, direct) = 0;
1383 			cc_migration_prec(cc, direct) = 0;
1384 			cc_migration_func(cc, direct) = NULL;
1385 			cc_migration_arg(cc, direct) = NULL;
1386 #endif
1387 			CTR3(KTR_CALLOUT, "postponing stop %p func %p arg %p",
1388 			    c, c->c_func, c->c_arg);
1389  			if (drain) {
1390 				cc_exec_drain(cc, direct) = drain;
1391 			}
1392 			CC_UNLOCK(cc);
1393 			return ((flags & CS_EXECUTING) != 0);
1394 		}
1395 		CTR3(KTR_CALLOUT, "failed to stop %p func %p arg %p",
1396 		    c, c->c_func, c->c_arg);
1397 		if (drain) {
1398 			cc_exec_drain(cc, direct) = drain;
1399 		}
1400 		KASSERT(!sq_locked, ("sleepqueue chain still locked"));
1401 		cancelled = ((flags & CS_EXECUTING) != 0);
1402 	} else
1403 		cancelled = 1;
1404 
1405 	if (sq_locked)
1406 		sleepq_release(&cc_exec_waiting(cc, direct));
1407 
1408 	if ((c->c_iflags & CALLOUT_PENDING) == 0) {
1409 		CTR3(KTR_CALLOUT, "failed to stop %p func %p arg %p",
1410 		    c, c->c_func, c->c_arg);
1411 		/*
1412 		 * For not scheduled and not executing callout return
1413 		 * negative value.
1414 		 */
1415 		if (cc_exec_curr(cc, direct) != c)
1416 			cancelled = -1;
1417 		CC_UNLOCK(cc);
1418 		return (cancelled);
1419 	}
1420 
1421 	c->c_iflags &= ~CALLOUT_PENDING;
1422 	c->c_flags &= ~CALLOUT_ACTIVE;
1423 
1424 	CTR3(KTR_CALLOUT, "cancelled %p func %p arg %p",
1425 	    c, c->c_func, c->c_arg);
1426 	if (not_on_a_list == 0) {
1427 		if ((c->c_iflags & CALLOUT_PROCESSED) == 0) {
1428 			if (cc_exec_next(cc) == c)
1429 				cc_exec_next(cc) = LIST_NEXT(c, c_links.le);
1430 			LIST_REMOVE(c, c_links.le);
1431 		} else {
1432 			TAILQ_REMOVE(&cc->cc_expireq, c, c_links.tqe);
1433 		}
1434 	}
1435 	callout_cc_del(c, cc);
1436 	CC_UNLOCK(cc);
1437 	return (cancelled);
1438 }
1439 
1440 void
1441 callout_init(struct callout *c, int mpsafe)
1442 {
1443 	bzero(c, sizeof *c);
1444 	if (mpsafe) {
1445 		c->c_lock = NULL;
1446 		c->c_iflags = CALLOUT_RETURNUNLOCKED;
1447 	} else {
1448 		c->c_lock = &Giant.lock_object;
1449 		c->c_iflags = 0;
1450 	}
1451 	c->c_cpu = timeout_cpu;
1452 }
1453 
1454 void
1455 _callout_init_lock(struct callout *c, struct lock_object *lock, int flags)
1456 {
1457 	bzero(c, sizeof *c);
1458 	c->c_lock = lock;
1459 	KASSERT((flags & ~(CALLOUT_RETURNUNLOCKED | CALLOUT_SHAREDLOCK)) == 0,
1460 	    ("callout_init_lock: bad flags %d", flags));
1461 	KASSERT(lock != NULL || (flags & CALLOUT_RETURNUNLOCKED) == 0,
1462 	    ("callout_init_lock: CALLOUT_RETURNUNLOCKED with no lock"));
1463 	KASSERT(lock == NULL || !(LOCK_CLASS(lock)->lc_flags &
1464 	    (LC_SPINLOCK | LC_SLEEPABLE)), ("%s: invalid lock class",
1465 	    __func__));
1466 	c->c_iflags = flags & (CALLOUT_RETURNUNLOCKED | CALLOUT_SHAREDLOCK);
1467 	c->c_cpu = timeout_cpu;
1468 }
1469 
1470 #ifdef APM_FIXUP_CALLTODO
1471 /*
1472  * Adjust the kernel calltodo timeout list.  This routine is used after
1473  * an APM resume to recalculate the calltodo timer list values with the
1474  * number of hz's we have been sleeping.  The next hardclock() will detect
1475  * that there are fired timers and run softclock() to execute them.
1476  *
1477  * Please note, I have not done an exhaustive analysis of what code this
1478  * might break.  I am motivated to have my select()'s and alarm()'s that
1479  * have expired during suspend firing upon resume so that the applications
1480  * which set the timer can do the maintanence the timer was for as close
1481  * as possible to the originally intended time.  Testing this code for a
1482  * week showed that resuming from a suspend resulted in 22 to 25 timers
1483  * firing, which seemed independent on whether the suspend was 2 hours or
1484  * 2 days.  Your milage may vary.   - Ken Key <key@cs.utk.edu>
1485  */
1486 void
1487 adjust_timeout_calltodo(struct timeval *time_change)
1488 {
1489 	struct callout *p;
1490 	unsigned long delta_ticks;
1491 
1492 	/*
1493 	 * How many ticks were we asleep?
1494 	 * (stolen from tvtohz()).
1495 	 */
1496 
1497 	/* Don't do anything */
1498 	if (time_change->tv_sec < 0)
1499 		return;
1500 	else if (time_change->tv_sec <= LONG_MAX / 1000000)
1501 		delta_ticks = howmany(time_change->tv_sec * 1000000 +
1502 		    time_change->tv_usec, tick) + 1;
1503 	else if (time_change->tv_sec <= LONG_MAX / hz)
1504 		delta_ticks = time_change->tv_sec * hz +
1505 		    howmany(time_change->tv_usec, tick) + 1;
1506 	else
1507 		delta_ticks = LONG_MAX;
1508 
1509 	if (delta_ticks > INT_MAX)
1510 		delta_ticks = INT_MAX;
1511 
1512 	/*
1513 	 * Now rip through the timer calltodo list looking for timers
1514 	 * to expire.
1515 	 */
1516 
1517 	/* don't collide with softclock() */
1518 	CC_LOCK(cc);
1519 	for (p = calltodo.c_next; p != NULL; p = p->c_next) {
1520 		p->c_time -= delta_ticks;
1521 
1522 		/* Break if the timer had more time on it than delta_ticks */
1523 		if (p->c_time > 0)
1524 			break;
1525 
1526 		/* take back the ticks the timer didn't use (p->c_time <= 0) */
1527 		delta_ticks = -p->c_time;
1528 	}
1529 	CC_UNLOCK(cc);
1530 
1531 	return;
1532 }
1533 #endif /* APM_FIXUP_CALLTODO */
1534 
1535 static int
1536 flssbt(sbintime_t sbt)
1537 {
1538 
1539 	sbt += (uint64_t)sbt >> 1;
1540 	if (sizeof(long) >= sizeof(sbintime_t))
1541 		return (flsl(sbt));
1542 	if (sbt >= SBT_1S)
1543 		return (flsl(((uint64_t)sbt) >> 32) + 32);
1544 	return (flsl(sbt));
1545 }
1546 
1547 /*
1548  * Dump immediate statistic snapshot of the scheduled callouts.
1549  */
1550 static int
1551 sysctl_kern_callout_stat(SYSCTL_HANDLER_ARGS)
1552 {
1553 	struct callout *tmp;
1554 	struct callout_cpu *cc;
1555 	struct callout_list *sc;
1556 	sbintime_t maxpr, maxt, medpr, medt, now, spr, st, t;
1557 	int ct[64], cpr[64], ccpbk[32];
1558 	int error, val, i, count, tcum, pcum, maxc, c, medc;
1559 #ifdef SMP
1560 	int cpu;
1561 #endif
1562 
1563 	val = 0;
1564 	error = sysctl_handle_int(oidp, &val, 0, req);
1565 	if (error != 0 || req->newptr == NULL)
1566 		return (error);
1567 	count = maxc = 0;
1568 	st = spr = maxt = maxpr = 0;
1569 	bzero(ccpbk, sizeof(ccpbk));
1570 	bzero(ct, sizeof(ct));
1571 	bzero(cpr, sizeof(cpr));
1572 	now = sbinuptime();
1573 #ifdef SMP
1574 	CPU_FOREACH(cpu) {
1575 		cc = CC_CPU(cpu);
1576 #else
1577 		cc = CC_CPU(timeout_cpu);
1578 #endif
1579 		CC_LOCK(cc);
1580 		for (i = 0; i < callwheelsize; i++) {
1581 			sc = &cc->cc_callwheel[i];
1582 			c = 0;
1583 			LIST_FOREACH(tmp, sc, c_links.le) {
1584 				c++;
1585 				t = tmp->c_time - now;
1586 				if (t < 0)
1587 					t = 0;
1588 				st += t / SBT_1US;
1589 				spr += tmp->c_precision / SBT_1US;
1590 				if (t > maxt)
1591 					maxt = t;
1592 				if (tmp->c_precision > maxpr)
1593 					maxpr = tmp->c_precision;
1594 				ct[flssbt(t)]++;
1595 				cpr[flssbt(tmp->c_precision)]++;
1596 			}
1597 			if (c > maxc)
1598 				maxc = c;
1599 			ccpbk[fls(c + c / 2)]++;
1600 			count += c;
1601 		}
1602 		CC_UNLOCK(cc);
1603 #ifdef SMP
1604 	}
1605 #endif
1606 
1607 	for (i = 0, tcum = 0; i < 64 && tcum < count / 2; i++)
1608 		tcum += ct[i];
1609 	medt = (i >= 2) ? (((sbintime_t)1) << (i - 2)) : 0;
1610 	for (i = 0, pcum = 0; i < 64 && pcum < count / 2; i++)
1611 		pcum += cpr[i];
1612 	medpr = (i >= 2) ? (((sbintime_t)1) << (i - 2)) : 0;
1613 	for (i = 0, c = 0; i < 32 && c < count / 2; i++)
1614 		c += ccpbk[i];
1615 	medc = (i >= 2) ? (1 << (i - 2)) : 0;
1616 
1617 	printf("Scheduled callouts statistic snapshot:\n");
1618 	printf("  Callouts: %6d  Buckets: %6d*%-3d  Bucket size: 0.%06ds\n",
1619 	    count, callwheelsize, mp_ncpus, 1000000 >> CC_HASH_SHIFT);
1620 	printf("  C/Bk: med %5d         avg %6d.%06jd  max %6d\n",
1621 	    medc,
1622 	    count / callwheelsize / mp_ncpus,
1623 	    (uint64_t)count * 1000000 / callwheelsize / mp_ncpus % 1000000,
1624 	    maxc);
1625 	printf("  Time: med %5jd.%06jds avg %6jd.%06jds max %6jd.%06jds\n",
1626 	    medt / SBT_1S, (medt & 0xffffffff) * 1000000 >> 32,
1627 	    (st / count) / 1000000, (st / count) % 1000000,
1628 	    maxt / SBT_1S, (maxt & 0xffffffff) * 1000000 >> 32);
1629 	printf("  Prec: med %5jd.%06jds avg %6jd.%06jds max %6jd.%06jds\n",
1630 	    medpr / SBT_1S, (medpr & 0xffffffff) * 1000000 >> 32,
1631 	    (spr / count) / 1000000, (spr / count) % 1000000,
1632 	    maxpr / SBT_1S, (maxpr & 0xffffffff) * 1000000 >> 32);
1633 	printf("  Distribution:       \tbuckets\t   time\t   tcum\t"
1634 	    "   prec\t   pcum\n");
1635 	for (i = 0, tcum = pcum = 0; i < 64; i++) {
1636 		if (ct[i] == 0 && cpr[i] == 0)
1637 			continue;
1638 		t = (i != 0) ? (((sbintime_t)1) << (i - 1)) : 0;
1639 		tcum += ct[i];
1640 		pcum += cpr[i];
1641 		printf("  %10jd.%06jds\t 2**%d\t%7d\t%7d\t%7d\t%7d\n",
1642 		    t / SBT_1S, (t & 0xffffffff) * 1000000 >> 32,
1643 		    i - 1 - (32 - CC_HASH_SHIFT),
1644 		    ct[i], tcum, cpr[i], pcum);
1645 	}
1646 	return (error);
1647 }
1648 SYSCTL_PROC(_kern, OID_AUTO, callout_stat,
1649     CTLTYPE_INT | CTLFLAG_RW | CTLFLAG_MPSAFE,
1650     0, 0, sysctl_kern_callout_stat, "I",
1651     "Dump immediate statistic snapshot of the scheduled callouts");
1652 
1653 #ifdef DDB
1654 static void
1655 _show_callout(struct callout *c)
1656 {
1657 
1658 	db_printf("callout %p\n", c);
1659 #define	C_DB_PRINTF(f, e)	db_printf("   %s = " f "\n", #e, c->e);
1660 	db_printf("   &c_links = %p\n", &(c->c_links));
1661 	C_DB_PRINTF("%" PRId64,	c_time);
1662 	C_DB_PRINTF("%" PRId64,	c_precision);
1663 	C_DB_PRINTF("%p",	c_arg);
1664 	C_DB_PRINTF("%p",	c_func);
1665 	C_DB_PRINTF("%p",	c_lock);
1666 	C_DB_PRINTF("%#x",	c_flags);
1667 	C_DB_PRINTF("%#x",	c_iflags);
1668 	C_DB_PRINTF("%d",	c_cpu);
1669 #undef	C_DB_PRINTF
1670 }
1671 
1672 DB_SHOW_COMMAND(callout, db_show_callout)
1673 {
1674 
1675 	if (!have_addr) {
1676 		db_printf("usage: show callout <struct callout *>\n");
1677 		return;
1678 	}
1679 
1680 	_show_callout((struct callout *)addr);
1681 }
1682 
1683 static void
1684 _show_last_callout(int cpu, int direct, const char *dirstr)
1685 {
1686 	struct callout_cpu *cc;
1687 	void *func, *arg;
1688 
1689 	cc = CC_CPU(cpu);
1690 	func = cc_exec_last_func(cc, direct);
1691 	arg = cc_exec_last_arg(cc, direct);
1692 	db_printf("cpu %d last%s callout function: %p ", cpu, dirstr, func);
1693 	db_printsym((db_expr_t)func, DB_STGY_ANY);
1694 	db_printf("\ncpu %d last%s callout argument: %p\n", cpu, dirstr, arg);
1695 }
1696 
1697 DB_SHOW_COMMAND(callout_last, db_show_callout_last)
1698 {
1699 	int cpu, last;
1700 
1701 	if (have_addr) {
1702 		if (addr < 0 || addr > mp_maxid || CPU_ABSENT(addr)) {
1703 			db_printf("no such cpu: %d\n", (int)addr);
1704 			return;
1705 		}
1706 		cpu = last = addr;
1707 	} else {
1708 		cpu = 0;
1709 		last = mp_maxid;
1710 	}
1711 
1712 	while (cpu <= last) {
1713 		if (!CPU_ABSENT(cpu)) {
1714 			_show_last_callout(cpu, 0, "");
1715 			_show_last_callout(cpu, 1, " direct");
1716 		}
1717 		cpu++;
1718 	}
1719 }
1720 #endif /* DDB */
1721