xref: /freebsd/sys/kern/kern_timeout.c (revision 2c9a9dfc187d171de6b92654d71b977f067ed641)
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 static u_int __read_mostly callwheelsize;
134 static u_int __read_mostly callwheelmask;
135 
136 /*
137  * The callout cpu exec entities represent informations necessary for
138  * describing the state of callouts currently running on the CPU and the ones
139  * necessary for migrating callouts to the new callout cpu. In particular,
140  * the first entry of the array cc_exec_entity holds informations for callout
141  * running in SWI thread context, while the second one holds informations
142  * for callout running directly from hardware interrupt context.
143  * The cached informations are very important for deferring migration when
144  * the migrating callout is already running.
145  */
146 struct cc_exec {
147 	struct callout		*cc_curr;
148 	callout_func_t		*cc_drain;
149 	void			*cc_last_func;
150 	void			*cc_last_arg;
151 #ifdef SMP
152 	callout_func_t		*ce_migration_func;
153 	void			*ce_migration_arg;
154 	sbintime_t		ce_migration_time;
155 	sbintime_t		ce_migration_prec;
156 	int			ce_migration_cpu;
157 #endif
158 	bool			cc_cancel;
159 	bool			cc_waiting;
160 };
161 
162 /*
163  * There is one struct callout_cpu per cpu, holding all relevant
164  * state for the callout processing thread on the individual CPU.
165  */
166 struct callout_cpu {
167 	struct mtx_padalign	cc_lock;
168 	struct cc_exec 		cc_exec_entity[2];
169 	struct callout		*cc_next;
170 	struct callout		*cc_callout;
171 	struct callout_list	*cc_callwheel;
172 	struct callout_tailq	cc_expireq;
173 	struct callout_slist	cc_callfree;
174 	sbintime_t		cc_firstevent;
175 	sbintime_t		cc_lastscan;
176 	void			*cc_cookie;
177 	u_int			cc_bucket;
178 	u_int			cc_inited;
179 #ifdef KTR
180 	char			cc_ktr_event_name[20];
181 #endif
182 };
183 
184 #define	callout_migrating(c)	((c)->c_iflags & CALLOUT_DFRMIGRATION)
185 
186 #define	cc_exec_curr(cc, dir)		cc->cc_exec_entity[dir].cc_curr
187 #define	cc_exec_last_func(cc, dir)	cc->cc_exec_entity[dir].cc_last_func
188 #define	cc_exec_last_arg(cc, dir)	cc->cc_exec_entity[dir].cc_last_arg
189 #define	cc_exec_drain(cc, dir)		cc->cc_exec_entity[dir].cc_drain
190 #define	cc_exec_next(cc)		cc->cc_next
191 #define	cc_exec_cancel(cc, dir)		cc->cc_exec_entity[dir].cc_cancel
192 #define	cc_exec_waiting(cc, dir)	cc->cc_exec_entity[dir].cc_waiting
193 #ifdef SMP
194 #define	cc_migration_func(cc, dir)	cc->cc_exec_entity[dir].ce_migration_func
195 #define	cc_migration_arg(cc, dir)	cc->cc_exec_entity[dir].ce_migration_arg
196 #define	cc_migration_cpu(cc, dir)	cc->cc_exec_entity[dir].ce_migration_cpu
197 #define	cc_migration_time(cc, dir)	cc->cc_exec_entity[dir].ce_migration_time
198 #define	cc_migration_prec(cc, dir)	cc->cc_exec_entity[dir].ce_migration_prec
199 
200 struct callout_cpu cc_cpu[MAXCPU];
201 #define	CPUBLOCK	MAXCPU
202 #define	CC_CPU(cpu)	(&cc_cpu[(cpu)])
203 #define	CC_SELF()	CC_CPU(PCPU_GET(cpuid))
204 #else
205 struct callout_cpu cc_cpu;
206 #define	CC_CPU(cpu)	&cc_cpu
207 #define	CC_SELF()	&cc_cpu
208 #endif
209 #define	CC_LOCK(cc)	mtx_lock_spin(&(cc)->cc_lock)
210 #define	CC_UNLOCK(cc)	mtx_unlock_spin(&(cc)->cc_lock)
211 #define	CC_LOCK_ASSERT(cc)	mtx_assert(&(cc)->cc_lock, MA_OWNED)
212 
213 static int __read_mostly timeout_cpu;
214 
215 static void	callout_cpu_init(struct callout_cpu *cc, int cpu);
216 static void	softclock_call_cc(struct callout *c, struct callout_cpu *cc,
217 #ifdef CALLOUT_PROFILING
218 		    int *mpcalls, int *lockcalls, int *gcalls,
219 #endif
220 		    int direct);
221 
222 static MALLOC_DEFINE(M_CALLOUT, "callout", "Callout datastructures");
223 
224 /**
225  * Locked by cc_lock:
226  *   cc_curr         - If a callout is in progress, it is cc_curr.
227  *                     If cc_curr is non-NULL, threads waiting in
228  *                     callout_drain() will be woken up as soon as the
229  *                     relevant callout completes.
230  *   cc_cancel       - Changing to 1 with both callout_lock and cc_lock held
231  *                     guarantees that the current callout will not run.
232  *                     The softclock() function sets this to 0 before it
233  *                     drops callout_lock to acquire c_lock, and it calls
234  *                     the handler only if curr_cancelled is still 0 after
235  *                     cc_lock is successfully acquired.
236  *   cc_waiting      - If a thread is waiting in callout_drain(), then
237  *                     callout_wait is nonzero.  Set only when
238  *                     cc_curr is non-NULL.
239  */
240 
241 /*
242  * Resets the execution entity tied to a specific callout cpu.
243  */
244 static void
245 cc_cce_cleanup(struct callout_cpu *cc, int direct)
246 {
247 
248 	cc_exec_curr(cc, direct) = NULL;
249 	cc_exec_cancel(cc, direct) = false;
250 	cc_exec_waiting(cc, direct) = false;
251 #ifdef SMP
252 	cc_migration_cpu(cc, direct) = CPUBLOCK;
253 	cc_migration_time(cc, direct) = 0;
254 	cc_migration_prec(cc, direct) = 0;
255 	cc_migration_func(cc, direct) = NULL;
256 	cc_migration_arg(cc, direct) = NULL;
257 #endif
258 }
259 
260 /*
261  * Checks if migration is requested by a specific callout cpu.
262  */
263 static int
264 cc_cce_migrating(struct callout_cpu *cc, int direct)
265 {
266 
267 #ifdef SMP
268 	return (cc_migration_cpu(cc, direct) != CPUBLOCK);
269 #else
270 	return (0);
271 #endif
272 }
273 
274 /*
275  * Kernel low level callwheel initialization
276  * called on the BSP during kernel startup.
277  */
278 static void
279 callout_callwheel_init(void *dummy)
280 {
281 	struct callout_cpu *cc;
282 
283 	/*
284 	 * Calculate the size of the callout wheel and the preallocated
285 	 * timeout() structures.
286 	 * XXX: Clip callout to result of previous function of maxusers
287 	 * maximum 384.  This is still huge, but acceptable.
288 	 */
289 	memset(CC_CPU(curcpu), 0, sizeof(cc_cpu));
290 	ncallout = imin(16 + maxproc + maxfiles, 18508);
291 	TUNABLE_INT_FETCH("kern.ncallout", &ncallout);
292 
293 	/*
294 	 * Calculate callout wheel size, should be next power of two higher
295 	 * than 'ncallout'.
296 	 */
297 	callwheelsize = 1 << fls(ncallout);
298 	callwheelmask = callwheelsize - 1;
299 
300 	/*
301 	 * Fetch whether we're pinning the swi's or not.
302 	 */
303 	TUNABLE_INT_FETCH("kern.pin_default_swi", &pin_default_swi);
304 	TUNABLE_INT_FETCH("kern.pin_pcpu_swi", &pin_pcpu_swi);
305 
306 	/*
307 	 * Only BSP handles timeout(9) and receives a preallocation.
308 	 *
309 	 * XXX: Once all timeout(9) consumers are converted this can
310 	 * be removed.
311 	 */
312 	timeout_cpu = PCPU_GET(cpuid);
313 	cc = CC_CPU(timeout_cpu);
314 	cc->cc_callout = malloc(ncallout * sizeof(struct callout),
315 	    M_CALLOUT, M_WAITOK);
316 	callout_cpu_init(cc, timeout_cpu);
317 }
318 SYSINIT(callwheel_init, SI_SUB_CPU, SI_ORDER_ANY, callout_callwheel_init, NULL);
319 
320 /*
321  * Initialize the per-cpu callout structures.
322  */
323 static void
324 callout_cpu_init(struct callout_cpu *cc, int cpu)
325 {
326 	struct callout *c;
327 	int i;
328 
329 	mtx_init(&cc->cc_lock, "callout", NULL, MTX_SPIN | MTX_RECURSE);
330 	SLIST_INIT(&cc->cc_callfree);
331 	cc->cc_inited = 1;
332 	cc->cc_callwheel = malloc_domainset(sizeof(struct callout_list) *
333 	    callwheelsize, M_CALLOUT,
334 	    DOMAINSET_PREF(pcpu_find(cpu)->pc_domain), M_WAITOK);
335 	for (i = 0; i < callwheelsize; i++)
336 		LIST_INIT(&cc->cc_callwheel[i]);
337 	TAILQ_INIT(&cc->cc_expireq);
338 	cc->cc_firstevent = SBT_MAX;
339 	for (i = 0; i < 2; i++)
340 		cc_cce_cleanup(cc, i);
341 #ifdef KTR
342 	snprintf(cc->cc_ktr_event_name, sizeof(cc->cc_ktr_event_name),
343 	    "callwheel cpu %d", cpu);
344 #endif
345 	if (cc->cc_callout == NULL)	/* Only BSP handles timeout(9) */
346 		return;
347 	for (i = 0; i < ncallout; i++) {
348 		c = &cc->cc_callout[i];
349 		callout_init(c, 0);
350 		c->c_iflags = CALLOUT_LOCAL_ALLOC;
351 		SLIST_INSERT_HEAD(&cc->cc_callfree, c, c_links.sle);
352 	}
353 }
354 
355 #ifdef SMP
356 /*
357  * Switches the cpu tied to a specific callout.
358  * The function expects a locked incoming callout cpu and returns with
359  * locked outcoming callout cpu.
360  */
361 static struct callout_cpu *
362 callout_cpu_switch(struct callout *c, struct callout_cpu *cc, int new_cpu)
363 {
364 	struct callout_cpu *new_cc;
365 
366 	MPASS(c != NULL && cc != NULL);
367 	CC_LOCK_ASSERT(cc);
368 
369 	/*
370 	 * Avoid interrupts and preemption firing after the callout cpu
371 	 * is blocked in order to avoid deadlocks as the new thread
372 	 * may be willing to acquire the callout cpu lock.
373 	 */
374 	c->c_cpu = CPUBLOCK;
375 	spinlock_enter();
376 	CC_UNLOCK(cc);
377 	new_cc = CC_CPU(new_cpu);
378 	CC_LOCK(new_cc);
379 	spinlock_exit();
380 	c->c_cpu = new_cpu;
381 	return (new_cc);
382 }
383 #endif
384 
385 /*
386  * Start standard softclock thread.
387  */
388 static void
389 start_softclock(void *dummy)
390 {
391 	struct callout_cpu *cc;
392 	char name[MAXCOMLEN];
393 #ifdef SMP
394 	int cpu;
395 	struct intr_event *ie;
396 #endif
397 
398 	cc = CC_CPU(timeout_cpu);
399 	snprintf(name, sizeof(name), "clock (%d)", timeout_cpu);
400 	if (swi_add(&clk_intr_event, name, softclock, cc, SWI_CLOCK,
401 	    INTR_MPSAFE, &cc->cc_cookie))
402 		panic("died while creating standard software ithreads");
403 	if (pin_default_swi &&
404 	    (intr_event_bind(clk_intr_event, timeout_cpu) != 0)) {
405 		printf("%s: timeout clock couldn't be pinned to cpu %d\n",
406 		    __func__,
407 		    timeout_cpu);
408 	}
409 
410 #ifdef SMP
411 	CPU_FOREACH(cpu) {
412 		if (cpu == timeout_cpu)
413 			continue;
414 		cc = CC_CPU(cpu);
415 		cc->cc_callout = NULL;	/* Only BSP handles timeout(9). */
416 		callout_cpu_init(cc, cpu);
417 		snprintf(name, sizeof(name), "clock (%d)", cpu);
418 		ie = NULL;
419 		if (swi_add(&ie, name, softclock, cc, SWI_CLOCK,
420 		    INTR_MPSAFE, &cc->cc_cookie))
421 			panic("died while creating standard software ithreads");
422 		if (pin_pcpu_swi && (intr_event_bind(ie, cpu) != 0)) {
423 			printf("%s: per-cpu clock couldn't be pinned to "
424 			    "cpu %d\n",
425 			    __func__,
426 			    cpu);
427 		}
428 	}
429 #endif
430 }
431 SYSINIT(start_softclock, SI_SUB_SOFTINTR, SI_ORDER_FIRST, start_softclock, NULL);
432 
433 #define	CC_HASH_SHIFT	8
434 
435 static inline u_int
436 callout_hash(sbintime_t sbt)
437 {
438 
439 	return (sbt >> (32 - CC_HASH_SHIFT));
440 }
441 
442 static inline u_int
443 callout_get_bucket(sbintime_t sbt)
444 {
445 
446 	return (callout_hash(sbt) & callwheelmask);
447 }
448 
449 void
450 callout_process(sbintime_t now)
451 {
452 	struct callout *tmp, *tmpn;
453 	struct callout_cpu *cc;
454 	struct callout_list *sc;
455 	sbintime_t first, last, max, tmp_max;
456 	uint32_t lookahead;
457 	u_int firstb, lastb, nowb;
458 #ifdef CALLOUT_PROFILING
459 	int depth_dir = 0, mpcalls_dir = 0, lockcalls_dir = 0;
460 #endif
461 
462 	cc = CC_SELF();
463 	mtx_lock_spin_flags(&cc->cc_lock, MTX_QUIET);
464 
465 	/* Compute the buckets of the last scan and present times. */
466 	firstb = callout_hash(cc->cc_lastscan);
467 	cc->cc_lastscan = now;
468 	nowb = callout_hash(now);
469 
470 	/* Compute the last bucket and minimum time of the bucket after it. */
471 	if (nowb == firstb)
472 		lookahead = (SBT_1S / 16);
473 	else if (nowb - firstb == 1)
474 		lookahead = (SBT_1S / 8);
475 	else
476 		lookahead = (SBT_1S / 2);
477 	first = last = now;
478 	first += (lookahead / 2);
479 	last += lookahead;
480 	last &= (0xffffffffffffffffLLU << (32 - CC_HASH_SHIFT));
481 	lastb = callout_hash(last) - 1;
482 	max = last;
483 
484 	/*
485 	 * Check if we wrapped around the entire wheel from the last scan.
486 	 * In case, we need to scan entirely the wheel for pending callouts.
487 	 */
488 	if (lastb - firstb >= callwheelsize) {
489 		lastb = firstb + callwheelsize - 1;
490 		if (nowb - firstb >= callwheelsize)
491 			nowb = lastb;
492 	}
493 
494 	/* Iterate callwheel from firstb to nowb and then up to lastb. */
495 	do {
496 		sc = &cc->cc_callwheel[firstb & callwheelmask];
497 		tmp = LIST_FIRST(sc);
498 		while (tmp != NULL) {
499 			/* Run the callout if present time within allowed. */
500 			if (tmp->c_time <= now) {
501 				/*
502 				 * Consumer told us the callout may be run
503 				 * directly from hardware interrupt context.
504 				 */
505 				if (tmp->c_iflags & CALLOUT_DIRECT) {
506 #ifdef CALLOUT_PROFILING
507 					++depth_dir;
508 #endif
509 					cc_exec_next(cc) =
510 					    LIST_NEXT(tmp, c_links.le);
511 					cc->cc_bucket = firstb & callwheelmask;
512 					LIST_REMOVE(tmp, c_links.le);
513 					softclock_call_cc(tmp, cc,
514 #ifdef CALLOUT_PROFILING
515 					    &mpcalls_dir, &lockcalls_dir, NULL,
516 #endif
517 					    1);
518 					tmp = cc_exec_next(cc);
519 					cc_exec_next(cc) = NULL;
520 				} else {
521 					tmpn = LIST_NEXT(tmp, c_links.le);
522 					LIST_REMOVE(tmp, c_links.le);
523 					TAILQ_INSERT_TAIL(&cc->cc_expireq,
524 					    tmp, c_links.tqe);
525 					tmp->c_iflags |= CALLOUT_PROCESSED;
526 					tmp = tmpn;
527 				}
528 				continue;
529 			}
530 			/* Skip events from distant future. */
531 			if (tmp->c_time >= max)
532 				goto next;
533 			/*
534 			 * Event minimal time is bigger than present maximal
535 			 * time, so it cannot be aggregated.
536 			 */
537 			if (tmp->c_time > last) {
538 				lastb = nowb;
539 				goto next;
540 			}
541 			/* Update first and last time, respecting this event. */
542 			if (tmp->c_time < first)
543 				first = tmp->c_time;
544 			tmp_max = tmp->c_time + tmp->c_precision;
545 			if (tmp_max < last)
546 				last = tmp_max;
547 next:
548 			tmp = LIST_NEXT(tmp, c_links.le);
549 		}
550 		/* Proceed with the next bucket. */
551 		firstb++;
552 		/*
553 		 * Stop if we looked after present time and found
554 		 * some event we can't execute at now.
555 		 * Stop if we looked far enough into the future.
556 		 */
557 	} while (((int)(firstb - lastb)) <= 0);
558 	cc->cc_firstevent = last;
559 #ifndef NO_EVENTTIMERS
560 	cpu_new_callout(curcpu, last, first);
561 #endif
562 #ifdef CALLOUT_PROFILING
563 	avg_depth_dir += (depth_dir * 1000 - avg_depth_dir) >> 8;
564 	avg_mpcalls_dir += (mpcalls_dir * 1000 - avg_mpcalls_dir) >> 8;
565 	avg_lockcalls_dir += (lockcalls_dir * 1000 - avg_lockcalls_dir) >> 8;
566 #endif
567 	mtx_unlock_spin_flags(&cc->cc_lock, MTX_QUIET);
568 	/*
569 	 * swi_sched acquires the thread lock, so we don't want to call it
570 	 * with cc_lock held; incorrect locking order.
571 	 */
572 	if (!TAILQ_EMPTY(&cc->cc_expireq))
573 		swi_sched(cc->cc_cookie, 0);
574 }
575 
576 static struct callout_cpu *
577 callout_lock(struct callout *c)
578 {
579 	struct callout_cpu *cc;
580 	int cpu;
581 
582 	for (;;) {
583 		cpu = c->c_cpu;
584 #ifdef SMP
585 		if (cpu == CPUBLOCK) {
586 			while (c->c_cpu == CPUBLOCK)
587 				cpu_spinwait();
588 			continue;
589 		}
590 #endif
591 		cc = CC_CPU(cpu);
592 		CC_LOCK(cc);
593 		if (cpu == c->c_cpu)
594 			break;
595 		CC_UNLOCK(cc);
596 	}
597 	return (cc);
598 }
599 
600 static void
601 callout_cc_add(struct callout *c, struct callout_cpu *cc,
602     sbintime_t sbt, sbintime_t precision, void (*func)(void *),
603     void *arg, int cpu, int flags)
604 {
605 	int bucket;
606 
607 	CC_LOCK_ASSERT(cc);
608 	if (sbt < cc->cc_lastscan)
609 		sbt = cc->cc_lastscan;
610 	c->c_arg = arg;
611 	c->c_iflags |= CALLOUT_PENDING;
612 	c->c_iflags &= ~CALLOUT_PROCESSED;
613 	c->c_flags |= CALLOUT_ACTIVE;
614 	if (flags & C_DIRECT_EXEC)
615 		c->c_iflags |= CALLOUT_DIRECT;
616 	c->c_func = func;
617 	c->c_time = sbt;
618 	c->c_precision = precision;
619 	bucket = callout_get_bucket(c->c_time);
620 	CTR3(KTR_CALLOUT, "precision set for %p: %d.%08x",
621 	    c, (int)(c->c_precision >> 32),
622 	    (u_int)(c->c_precision & 0xffffffff));
623 	LIST_INSERT_HEAD(&cc->cc_callwheel[bucket], c, c_links.le);
624 	if (cc->cc_bucket == bucket)
625 		cc_exec_next(cc) = c;
626 #ifndef NO_EVENTTIMERS
627 	/*
628 	 * Inform the eventtimers(4) subsystem there's a new callout
629 	 * that has been inserted, but only if really required.
630 	 */
631 	if (SBT_MAX - c->c_time < c->c_precision)
632 		c->c_precision = SBT_MAX - c->c_time;
633 	sbt = c->c_time + c->c_precision;
634 	if (sbt < cc->cc_firstevent) {
635 		cc->cc_firstevent = sbt;
636 		cpu_new_callout(cpu, sbt, c->c_time);
637 	}
638 #endif
639 }
640 
641 static void
642 callout_cc_del(struct callout *c, struct callout_cpu *cc)
643 {
644 
645 	if ((c->c_iflags & CALLOUT_LOCAL_ALLOC) == 0)
646 		return;
647 	c->c_func = NULL;
648 	SLIST_INSERT_HEAD(&cc->cc_callfree, c, c_links.sle);
649 }
650 
651 static void
652 softclock_call_cc(struct callout *c, struct callout_cpu *cc,
653 #ifdef CALLOUT_PROFILING
654     int *mpcalls, int *lockcalls, int *gcalls,
655 #endif
656     int direct)
657 {
658 	struct rm_priotracker tracker;
659 	callout_func_t *c_func, *drain;
660 	void *c_arg;
661 	struct lock_class *class;
662 	struct lock_object *c_lock;
663 	uintptr_t lock_status;
664 	int c_iflags;
665 #ifdef SMP
666 	struct callout_cpu *new_cc;
667 	callout_func_t *new_func;
668 	void *new_arg;
669 	int flags, new_cpu;
670 	sbintime_t new_prec, new_time;
671 #endif
672 #if defined(DIAGNOSTIC) || defined(CALLOUT_PROFILING)
673 	sbintime_t sbt1, sbt2;
674 	struct timespec ts2;
675 	static sbintime_t maxdt = 2 * SBT_1MS;	/* 2 msec */
676 	static callout_func_t *lastfunc;
677 #endif
678 
679 	KASSERT((c->c_iflags & CALLOUT_PENDING) == CALLOUT_PENDING,
680 	    ("softclock_call_cc: pend %p %x", c, c->c_iflags));
681 	KASSERT((c->c_flags & CALLOUT_ACTIVE) == CALLOUT_ACTIVE,
682 	    ("softclock_call_cc: act %p %x", c, c->c_flags));
683 	class = (c->c_lock != NULL) ? LOCK_CLASS(c->c_lock) : NULL;
684 	lock_status = 0;
685 	if (c->c_flags & CALLOUT_SHAREDLOCK) {
686 		if (class == &lock_class_rm)
687 			lock_status = (uintptr_t)&tracker;
688 		else
689 			lock_status = 1;
690 	}
691 	c_lock = c->c_lock;
692 	c_func = c->c_func;
693 	c_arg = c->c_arg;
694 	c_iflags = c->c_iflags;
695 	if (c->c_iflags & CALLOUT_LOCAL_ALLOC)
696 		c->c_iflags = CALLOUT_LOCAL_ALLOC;
697 	else
698 		c->c_iflags &= ~CALLOUT_PENDING;
699 
700 	cc_exec_curr(cc, direct) = c;
701 	cc_exec_last_func(cc, direct) = c_func;
702 	cc_exec_last_arg(cc, direct) = c_arg;
703 	cc_exec_cancel(cc, direct) = false;
704 	cc_exec_drain(cc, direct) = NULL;
705 	CC_UNLOCK(cc);
706 	if (c_lock != NULL) {
707 		class->lc_lock(c_lock, lock_status);
708 		/*
709 		 * The callout may have been cancelled
710 		 * while we switched locks.
711 		 */
712 		if (cc_exec_cancel(cc, direct)) {
713 			class->lc_unlock(c_lock);
714 			goto skip;
715 		}
716 		/* The callout cannot be stopped now. */
717 		cc_exec_cancel(cc, direct) = true;
718 		if (c_lock == &Giant.lock_object) {
719 #ifdef CALLOUT_PROFILING
720 			(*gcalls)++;
721 #endif
722 			CTR3(KTR_CALLOUT, "callout giant %p func %p arg %p",
723 			    c, c_func, c_arg);
724 		} else {
725 #ifdef CALLOUT_PROFILING
726 			(*lockcalls)++;
727 #endif
728 			CTR3(KTR_CALLOUT, "callout lock %p func %p arg %p",
729 			    c, c_func, c_arg);
730 		}
731 	} else {
732 #ifdef CALLOUT_PROFILING
733 		(*mpcalls)++;
734 #endif
735 		CTR3(KTR_CALLOUT, "callout %p func %p arg %p",
736 		    c, c_func, c_arg);
737 	}
738 	KTR_STATE3(KTR_SCHED, "callout", cc->cc_ktr_event_name, "running",
739 	    "func:%p", c_func, "arg:%p", c_arg, "direct:%d", direct);
740 #if defined(DIAGNOSTIC) || defined(CALLOUT_PROFILING)
741 	sbt1 = sbinuptime();
742 #endif
743 	THREAD_NO_SLEEPING();
744 	SDT_PROBE1(callout_execute, , , callout__start, c);
745 	c_func(c_arg);
746 	SDT_PROBE1(callout_execute, , , callout__end, c);
747 	THREAD_SLEEPING_OK();
748 #if defined(DIAGNOSTIC) || defined(CALLOUT_PROFILING)
749 	sbt2 = sbinuptime();
750 	sbt2 -= sbt1;
751 	if (sbt2 > maxdt) {
752 		if (lastfunc != c_func || sbt2 > maxdt * 2) {
753 			ts2 = sbttots(sbt2);
754 			printf(
755 		"Expensive timeout(9) function: %p(%p) %jd.%09ld s\n",
756 			    c_func, c_arg, (intmax_t)ts2.tv_sec, ts2.tv_nsec);
757 		}
758 		maxdt = sbt2;
759 		lastfunc = c_func;
760 	}
761 #endif
762 	KTR_STATE0(KTR_SCHED, "callout", cc->cc_ktr_event_name, "idle");
763 	CTR1(KTR_CALLOUT, "callout %p finished", c);
764 	if ((c_iflags & CALLOUT_RETURNUNLOCKED) == 0)
765 		class->lc_unlock(c_lock);
766 skip:
767 	CC_LOCK(cc);
768 	KASSERT(cc_exec_curr(cc, direct) == c, ("mishandled cc_curr"));
769 	cc_exec_curr(cc, direct) = NULL;
770 	if (cc_exec_drain(cc, direct)) {
771 		drain = cc_exec_drain(cc, direct);
772 		cc_exec_drain(cc, direct) = NULL;
773 		CC_UNLOCK(cc);
774 		drain(c_arg);
775 		CC_LOCK(cc);
776 	}
777 	if (cc_exec_waiting(cc, direct)) {
778 		/*
779 		 * There is someone waiting for the
780 		 * callout to complete.
781 		 * If the callout was scheduled for
782 		 * migration just cancel it.
783 		 */
784 		if (cc_cce_migrating(cc, direct)) {
785 			cc_cce_cleanup(cc, direct);
786 
787 			/*
788 			 * It should be assert here that the callout is not
789 			 * destroyed but that is not easy.
790 			 */
791 			c->c_iflags &= ~CALLOUT_DFRMIGRATION;
792 		}
793 		cc_exec_waiting(cc, direct) = false;
794 		CC_UNLOCK(cc);
795 		wakeup(&cc_exec_waiting(cc, direct));
796 		CC_LOCK(cc);
797 	} else if (cc_cce_migrating(cc, direct)) {
798 		KASSERT((c_iflags & CALLOUT_LOCAL_ALLOC) == 0,
799 		    ("Migrating legacy callout %p", c));
800 #ifdef SMP
801 		/*
802 		 * If the callout was scheduled for
803 		 * migration just perform it now.
804 		 */
805 		new_cpu = cc_migration_cpu(cc, direct);
806 		new_time = cc_migration_time(cc, direct);
807 		new_prec = cc_migration_prec(cc, direct);
808 		new_func = cc_migration_func(cc, direct);
809 		new_arg = cc_migration_arg(cc, direct);
810 		cc_cce_cleanup(cc, direct);
811 
812 		/*
813 		 * It should be assert here that the callout is not destroyed
814 		 * but that is not easy.
815 		 *
816 		 * As first thing, handle deferred callout stops.
817 		 */
818 		if (!callout_migrating(c)) {
819 			CTR3(KTR_CALLOUT,
820 			     "deferred cancelled %p func %p arg %p",
821 			     c, new_func, new_arg);
822 			callout_cc_del(c, cc);
823 			return;
824 		}
825 		c->c_iflags &= ~CALLOUT_DFRMIGRATION;
826 
827 		new_cc = callout_cpu_switch(c, cc, new_cpu);
828 		flags = (direct) ? C_DIRECT_EXEC : 0;
829 		callout_cc_add(c, new_cc, new_time, new_prec, new_func,
830 		    new_arg, new_cpu, flags);
831 		CC_UNLOCK(new_cc);
832 		CC_LOCK(cc);
833 #else
834 		panic("migration should not happen");
835 #endif
836 	}
837 	/*
838 	 * If the current callout is locally allocated (from
839 	 * timeout(9)) then put it on the freelist.
840 	 *
841 	 * Note: we need to check the cached copy of c_iflags because
842 	 * if it was not local, then it's not safe to deref the
843 	 * callout pointer.
844 	 */
845 	KASSERT((c_iflags & CALLOUT_LOCAL_ALLOC) == 0 ||
846 	    c->c_iflags == CALLOUT_LOCAL_ALLOC,
847 	    ("corrupted callout"));
848 	if (c_iflags & CALLOUT_LOCAL_ALLOC)
849 		callout_cc_del(c, cc);
850 }
851 
852 /*
853  * The callout mechanism is based on the work of Adam M. Costello and
854  * George Varghese, published in a technical report entitled "Redesigning
855  * the BSD Callout and Timer Facilities" and modified slightly for inclusion
856  * in FreeBSD by Justin T. Gibbs.  The original work on the data structures
857  * used in this implementation was published by G. Varghese and T. Lauck in
858  * the paper "Hashed and Hierarchical Timing Wheels: Data Structures for
859  * the Efficient Implementation of a Timer Facility" in the Proceedings of
860  * the 11th ACM Annual Symposium on Operating Systems Principles,
861  * Austin, Texas Nov 1987.
862  */
863 
864 /*
865  * Software (low priority) clock interrupt.
866  * Run periodic events from timeout queue.
867  */
868 void
869 softclock(void *arg)
870 {
871 	struct callout_cpu *cc;
872 	struct callout *c;
873 #ifdef CALLOUT_PROFILING
874 	int depth = 0, gcalls = 0, lockcalls = 0, mpcalls = 0;
875 #endif
876 
877 	cc = (struct callout_cpu *)arg;
878 	CC_LOCK(cc);
879 	while ((c = TAILQ_FIRST(&cc->cc_expireq)) != NULL) {
880 		TAILQ_REMOVE(&cc->cc_expireq, c, c_links.tqe);
881 		softclock_call_cc(c, cc,
882 #ifdef CALLOUT_PROFILING
883 		    &mpcalls, &lockcalls, &gcalls,
884 #endif
885 		    0);
886 #ifdef CALLOUT_PROFILING
887 		++depth;
888 #endif
889 	}
890 #ifdef CALLOUT_PROFILING
891 	avg_depth += (depth * 1000 - avg_depth) >> 8;
892 	avg_mpcalls += (mpcalls * 1000 - avg_mpcalls) >> 8;
893 	avg_lockcalls += (lockcalls * 1000 - avg_lockcalls) >> 8;
894 	avg_gcalls += (gcalls * 1000 - avg_gcalls) >> 8;
895 #endif
896 	CC_UNLOCK(cc);
897 }
898 
899 /*
900  * timeout --
901  *	Execute a function after a specified length of time.
902  *
903  * untimeout --
904  *	Cancel previous timeout function call.
905  *
906  * callout_handle_init --
907  *	Initialize a handle so that using it with untimeout is benign.
908  *
909  *	See AT&T BCI Driver Reference Manual for specification.  This
910  *	implementation differs from that one in that although an
911  *	identification value is returned from timeout, the original
912  *	arguments to timeout as well as the identifier are used to
913  *	identify entries for untimeout.
914  */
915 struct callout_handle
916 timeout(timeout_t *ftn, void *arg, int to_ticks)
917 {
918 	struct callout_cpu *cc;
919 	struct callout *new;
920 	struct callout_handle handle;
921 
922 	cc = CC_CPU(timeout_cpu);
923 	CC_LOCK(cc);
924 	/* Fill in the next free callout structure. */
925 	new = SLIST_FIRST(&cc->cc_callfree);
926 	if (new == NULL)
927 		/* XXX Attempt to malloc first */
928 		panic("timeout table full");
929 	SLIST_REMOVE_HEAD(&cc->cc_callfree, c_links.sle);
930 	callout_reset(new, to_ticks, ftn, arg);
931 	handle.callout = new;
932 	CC_UNLOCK(cc);
933 
934 	return (handle);
935 }
936 
937 void
938 untimeout(timeout_t *ftn, void *arg, struct callout_handle handle)
939 {
940 	struct callout_cpu *cc;
941 
942 	/*
943 	 * Check for a handle that was initialized
944 	 * by callout_handle_init, but never used
945 	 * for a real timeout.
946 	 */
947 	if (handle.callout == NULL)
948 		return;
949 
950 	cc = callout_lock(handle.callout);
951 	if (handle.callout->c_func == ftn && handle.callout->c_arg == arg)
952 		callout_stop(handle.callout);
953 	CC_UNLOCK(cc);
954 }
955 
956 void
957 callout_handle_init(struct callout_handle *handle)
958 {
959 	handle->callout = NULL;
960 }
961 
962 void
963 callout_when(sbintime_t sbt, sbintime_t precision, int flags,
964     sbintime_t *res, sbintime_t *prec_res)
965 {
966 	sbintime_t to_sbt, to_pr;
967 
968 	if ((flags & (C_ABSOLUTE | C_PRECALC)) != 0) {
969 		*res = sbt;
970 		*prec_res = precision;
971 		return;
972 	}
973 	if ((flags & C_HARDCLOCK) != 0 && sbt < tick_sbt)
974 		sbt = tick_sbt;
975 	if ((flags & C_HARDCLOCK) != 0 ||
976 #ifdef NO_EVENTTIMERS
977 	    sbt >= sbt_timethreshold) {
978 		to_sbt = getsbinuptime();
979 
980 		/* Add safety belt for the case of hz > 1000. */
981 		to_sbt += tc_tick_sbt - tick_sbt;
982 #else
983 	    sbt >= sbt_tickthreshold) {
984 		/*
985 		 * Obtain the time of the last hardclock() call on
986 		 * this CPU directly from the kern_clocksource.c.
987 		 * This value is per-CPU, but it is equal for all
988 		 * active ones.
989 		 */
990 #ifdef __LP64__
991 		to_sbt = DPCPU_GET(hardclocktime);
992 #else
993 		spinlock_enter();
994 		to_sbt = DPCPU_GET(hardclocktime);
995 		spinlock_exit();
996 #endif
997 #endif
998 		if (cold && to_sbt == 0)
999 			to_sbt = sbinuptime();
1000 		if ((flags & C_HARDCLOCK) == 0)
1001 			to_sbt += tick_sbt;
1002 	} else
1003 		to_sbt = sbinuptime();
1004 	if (SBT_MAX - to_sbt < sbt)
1005 		to_sbt = SBT_MAX;
1006 	else
1007 		to_sbt += sbt;
1008 	*res = to_sbt;
1009 	to_pr = ((C_PRELGET(flags) < 0) ? sbt >> tc_precexp :
1010 	    sbt >> C_PRELGET(flags));
1011 	*prec_res = to_pr > precision ? to_pr : precision;
1012 }
1013 
1014 /*
1015  * New interface; clients allocate their own callout structures.
1016  *
1017  * callout_reset() - establish or change a timeout
1018  * callout_stop() - disestablish a timeout
1019  * callout_init() - initialize a callout structure so that it can
1020  *	safely be passed to callout_reset() and callout_stop()
1021  *
1022  * <sys/callout.h> defines three convenience macros:
1023  *
1024  * callout_active() - returns truth if callout has not been stopped,
1025  *	drained, or deactivated since the last time the callout was
1026  *	reset.
1027  * callout_pending() - returns truth if callout is still waiting for timeout
1028  * callout_deactivate() - marks the callout as having been serviced
1029  */
1030 int
1031 callout_reset_sbt_on(struct callout *c, sbintime_t sbt, sbintime_t prec,
1032     callout_func_t *ftn, void *arg, int cpu, int flags)
1033 {
1034 	sbintime_t to_sbt, precision;
1035 	struct callout_cpu *cc;
1036 	int cancelled, direct;
1037 	int ignore_cpu=0;
1038 
1039 	cancelled = 0;
1040 	if (cpu == -1) {
1041 		ignore_cpu = 1;
1042 	} else if ((cpu >= MAXCPU) ||
1043 		   ((CC_CPU(cpu))->cc_inited == 0)) {
1044 		/* Invalid CPU spec */
1045 		panic("Invalid CPU in callout %d", cpu);
1046 	}
1047 	callout_when(sbt, prec, flags, &to_sbt, &precision);
1048 
1049 	/*
1050 	 * This flag used to be added by callout_cc_add, but the
1051 	 * first time you call this we could end up with the
1052 	 * wrong direct flag if we don't do it before we add.
1053 	 */
1054 	if (flags & C_DIRECT_EXEC) {
1055 		direct = 1;
1056 	} else {
1057 		direct = 0;
1058 	}
1059 	KASSERT(!direct || c->c_lock == NULL,
1060 	    ("%s: direct callout %p has lock", __func__, c));
1061 	cc = callout_lock(c);
1062 	/*
1063 	 * Don't allow migration of pre-allocated callouts lest they
1064 	 * become unbalanced or handle the case where the user does
1065 	 * not care.
1066 	 */
1067 	if ((c->c_iflags & CALLOUT_LOCAL_ALLOC) ||
1068 	    ignore_cpu) {
1069 		cpu = c->c_cpu;
1070 	}
1071 
1072 	if (cc_exec_curr(cc, direct) == c) {
1073 		/*
1074 		 * We're being asked to reschedule a callout which is
1075 		 * currently in progress.  If there is a lock then we
1076 		 * can cancel the callout if it has not really started.
1077 		 */
1078 		if (c->c_lock != NULL && !cc_exec_cancel(cc, direct))
1079 			cancelled = cc_exec_cancel(cc, direct) = true;
1080 		if (cc_exec_waiting(cc, direct) || cc_exec_drain(cc, direct)) {
1081 			/*
1082 			 * Someone has called callout_drain to kill this
1083 			 * callout.  Don't reschedule.
1084 			 */
1085 			CTR4(KTR_CALLOUT, "%s %p func %p arg %p",
1086 			    cancelled ? "cancelled" : "failed to cancel",
1087 			    c, c->c_func, c->c_arg);
1088 			CC_UNLOCK(cc);
1089 			return (cancelled);
1090 		}
1091 #ifdef SMP
1092 		if (callout_migrating(c)) {
1093 			/*
1094 			 * This only occurs when a second callout_reset_sbt_on
1095 			 * is made after a previous one moved it into
1096 			 * deferred migration (below). Note we do *not* change
1097 			 * the prev_cpu even though the previous target may
1098 			 * be different.
1099 			 */
1100 			cc_migration_cpu(cc, direct) = cpu;
1101 			cc_migration_time(cc, direct) = to_sbt;
1102 			cc_migration_prec(cc, direct) = precision;
1103 			cc_migration_func(cc, direct) = ftn;
1104 			cc_migration_arg(cc, direct) = arg;
1105 			cancelled = 1;
1106 			CC_UNLOCK(cc);
1107 			return (cancelled);
1108 		}
1109 #endif
1110 	}
1111 	if (c->c_iflags & CALLOUT_PENDING) {
1112 		if ((c->c_iflags & CALLOUT_PROCESSED) == 0) {
1113 			if (cc_exec_next(cc) == c)
1114 				cc_exec_next(cc) = LIST_NEXT(c, c_links.le);
1115 			LIST_REMOVE(c, c_links.le);
1116 		} else {
1117 			TAILQ_REMOVE(&cc->cc_expireq, c, c_links.tqe);
1118 		}
1119 		cancelled = 1;
1120 		c->c_iflags &= ~ CALLOUT_PENDING;
1121 		c->c_flags &= ~ CALLOUT_ACTIVE;
1122 	}
1123 
1124 #ifdef SMP
1125 	/*
1126 	 * If the callout must migrate try to perform it immediately.
1127 	 * If the callout is currently running, just defer the migration
1128 	 * to a more appropriate moment.
1129 	 */
1130 	if (c->c_cpu != cpu) {
1131 		if (cc_exec_curr(cc, direct) == c) {
1132 			/*
1133 			 * Pending will have been removed since we are
1134 			 * actually executing the callout on another
1135 			 * CPU. That callout should be waiting on the
1136 			 * lock the caller holds. If we set both
1137 			 * active/and/pending after we return and the
1138 			 * lock on the executing callout proceeds, it
1139 			 * will then see pending is true and return.
1140 			 * At the return from the actual callout execution
1141 			 * the migration will occur in softclock_call_cc
1142 			 * and this new callout will be placed on the
1143 			 * new CPU via a call to callout_cpu_switch() which
1144 			 * will get the lock on the right CPU followed
1145 			 * by a call callout_cc_add() which will add it there.
1146 			 * (see above in softclock_call_cc()).
1147 			 */
1148 			cc_migration_cpu(cc, direct) = cpu;
1149 			cc_migration_time(cc, direct) = to_sbt;
1150 			cc_migration_prec(cc, direct) = precision;
1151 			cc_migration_func(cc, direct) = ftn;
1152 			cc_migration_arg(cc, direct) = arg;
1153 			c->c_iflags |= (CALLOUT_DFRMIGRATION | CALLOUT_PENDING);
1154 			c->c_flags |= CALLOUT_ACTIVE;
1155 			CTR6(KTR_CALLOUT,
1156 		    "migration of %p func %p arg %p in %d.%08x to %u deferred",
1157 			    c, c->c_func, c->c_arg, (int)(to_sbt >> 32),
1158 			    (u_int)(to_sbt & 0xffffffff), cpu);
1159 			CC_UNLOCK(cc);
1160 			return (cancelled);
1161 		}
1162 		cc = callout_cpu_switch(c, cc, cpu);
1163 	}
1164 #endif
1165 
1166 	callout_cc_add(c, cc, to_sbt, precision, ftn, arg, cpu, flags);
1167 	CTR6(KTR_CALLOUT, "%sscheduled %p func %p arg %p in %d.%08x",
1168 	    cancelled ? "re" : "", c, c->c_func, c->c_arg, (int)(to_sbt >> 32),
1169 	    (u_int)(to_sbt & 0xffffffff));
1170 	CC_UNLOCK(cc);
1171 
1172 	return (cancelled);
1173 }
1174 
1175 /*
1176  * Common idioms that can be optimized in the future.
1177  */
1178 int
1179 callout_schedule_on(struct callout *c, int to_ticks, int cpu)
1180 {
1181 	return callout_reset_on(c, to_ticks, c->c_func, c->c_arg, cpu);
1182 }
1183 
1184 int
1185 callout_schedule(struct callout *c, int to_ticks)
1186 {
1187 	return callout_reset_on(c, to_ticks, c->c_func, c->c_arg, c->c_cpu);
1188 }
1189 
1190 int
1191 _callout_stop_safe(struct callout *c, int flags, callout_func_t *drain)
1192 {
1193 	struct callout_cpu *cc, *old_cc;
1194 	struct lock_class *class;
1195 	int direct, sq_locked, use_lock;
1196 	int cancelled, not_on_a_list;
1197 
1198 	if ((flags & CS_DRAIN) != 0)
1199 		WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK, c->c_lock,
1200 		    "calling %s", __func__);
1201 
1202 	/*
1203 	 * Some old subsystems don't hold Giant while running a callout_stop(),
1204 	 * so just discard this check for the moment.
1205 	 */
1206 	if ((flags & CS_DRAIN) == 0 && c->c_lock != NULL) {
1207 		if (c->c_lock == &Giant.lock_object)
1208 			use_lock = mtx_owned(&Giant);
1209 		else {
1210 			use_lock = 1;
1211 			class = LOCK_CLASS(c->c_lock);
1212 			class->lc_assert(c->c_lock, LA_XLOCKED);
1213 		}
1214 	} else
1215 		use_lock = 0;
1216 	if (c->c_iflags & CALLOUT_DIRECT) {
1217 		direct = 1;
1218 	} else {
1219 		direct = 0;
1220 	}
1221 	sq_locked = 0;
1222 	old_cc = NULL;
1223 again:
1224 	cc = callout_lock(c);
1225 
1226 	if ((c->c_iflags & (CALLOUT_DFRMIGRATION | CALLOUT_PENDING)) ==
1227 	    (CALLOUT_DFRMIGRATION | CALLOUT_PENDING) &&
1228 	    ((c->c_flags & CALLOUT_ACTIVE) == CALLOUT_ACTIVE)) {
1229 		/*
1230 		 * Special case where this slipped in while we
1231 		 * were migrating *as* the callout is about to
1232 		 * execute. The caller probably holds the lock
1233 		 * the callout wants.
1234 		 *
1235 		 * Get rid of the migration first. Then set
1236 		 * the flag that tells this code *not* to
1237 		 * try to remove it from any lists (its not
1238 		 * on one yet). When the callout wheel runs,
1239 		 * it will ignore this callout.
1240 		 */
1241 		c->c_iflags &= ~CALLOUT_PENDING;
1242 		c->c_flags &= ~CALLOUT_ACTIVE;
1243 		not_on_a_list = 1;
1244 	} else {
1245 		not_on_a_list = 0;
1246 	}
1247 
1248 	/*
1249 	 * If the callout was migrating while the callout cpu lock was
1250 	 * dropped,  just drop the sleepqueue lock and check the states
1251 	 * again.
1252 	 */
1253 	if (sq_locked != 0 && cc != old_cc) {
1254 #ifdef SMP
1255 		CC_UNLOCK(cc);
1256 		sleepq_release(&cc_exec_waiting(old_cc, direct));
1257 		sq_locked = 0;
1258 		old_cc = NULL;
1259 		goto again;
1260 #else
1261 		panic("migration should not happen");
1262 #endif
1263 	}
1264 
1265 	/*
1266 	 * If the callout is running, try to stop it or drain it.
1267 	 */
1268 	if (cc_exec_curr(cc, direct) == c) {
1269 		/*
1270 		 * Succeed we to stop it or not, we must clear the
1271 		 * active flag - this is what API users expect.  If we're
1272 		 * draining and the callout is currently executing, first wait
1273 		 * until it finishes.
1274 		 */
1275 		if ((flags & CS_DRAIN) == 0)
1276 			c->c_flags &= ~CALLOUT_ACTIVE;
1277 
1278 		if ((flags & CS_DRAIN) != 0) {
1279 			/*
1280 			 * The current callout is running (or just
1281 			 * about to run) and blocking is allowed, so
1282 			 * just wait for the current invocation to
1283 			 * finish.
1284 			 */
1285 			while (cc_exec_curr(cc, direct) == c) {
1286 				/*
1287 				 * Use direct calls to sleepqueue interface
1288 				 * instead of cv/msleep in order to avoid
1289 				 * a LOR between cc_lock and sleepqueue
1290 				 * chain spinlocks.  This piece of code
1291 				 * emulates a msleep_spin() call actually.
1292 				 *
1293 				 * If we already have the sleepqueue chain
1294 				 * locked, then we can safely block.  If we
1295 				 * don't already have it locked, however,
1296 				 * we have to drop the cc_lock to lock
1297 				 * it.  This opens several races, so we
1298 				 * restart at the beginning once we have
1299 				 * both locks.  If nothing has changed, then
1300 				 * we will end up back here with sq_locked
1301 				 * set.
1302 				 */
1303 				if (!sq_locked) {
1304 					CC_UNLOCK(cc);
1305 					sleepq_lock(
1306 					    &cc_exec_waiting(cc, direct));
1307 					sq_locked = 1;
1308 					old_cc = cc;
1309 					goto again;
1310 				}
1311 
1312 				/*
1313 				 * Migration could be cancelled here, but
1314 				 * as long as it is still not sure when it
1315 				 * will be packed up, just let softclock()
1316 				 * take care of it.
1317 				 */
1318 				cc_exec_waiting(cc, direct) = true;
1319 				DROP_GIANT();
1320 				CC_UNLOCK(cc);
1321 				sleepq_add(
1322 				    &cc_exec_waiting(cc, direct),
1323 				    &cc->cc_lock.lock_object, "codrain",
1324 				    SLEEPQ_SLEEP, 0);
1325 				sleepq_wait(
1326 				    &cc_exec_waiting(cc, direct),
1327 					     0);
1328 				sq_locked = 0;
1329 				old_cc = NULL;
1330 
1331 				/* Reacquire locks previously released. */
1332 				PICKUP_GIANT();
1333 				CC_LOCK(cc);
1334 			}
1335 			c->c_flags &= ~CALLOUT_ACTIVE;
1336 		} else if (use_lock &&
1337 			   !cc_exec_cancel(cc, direct) && (drain == NULL)) {
1338 
1339 			/*
1340 			 * The current callout is waiting for its
1341 			 * lock which we hold.  Cancel the callout
1342 			 * and return.  After our caller drops the
1343 			 * lock, the callout will be skipped in
1344 			 * softclock(). This *only* works with a
1345 			 * callout_stop() *not* callout_drain() or
1346 			 * callout_async_drain().
1347 			 */
1348 			cc_exec_cancel(cc, direct) = true;
1349 			CTR3(KTR_CALLOUT, "cancelled %p func %p arg %p",
1350 			    c, c->c_func, c->c_arg);
1351 			KASSERT(!cc_cce_migrating(cc, direct),
1352 			    ("callout wrongly scheduled for migration"));
1353 			if (callout_migrating(c)) {
1354 				c->c_iflags &= ~CALLOUT_DFRMIGRATION;
1355 #ifdef SMP
1356 				cc_migration_cpu(cc, direct) = CPUBLOCK;
1357 				cc_migration_time(cc, direct) = 0;
1358 				cc_migration_prec(cc, direct) = 0;
1359 				cc_migration_func(cc, direct) = NULL;
1360 				cc_migration_arg(cc, direct) = NULL;
1361 #endif
1362 			}
1363 			CC_UNLOCK(cc);
1364 			KASSERT(!sq_locked, ("sleepqueue chain locked"));
1365 			return (1);
1366 		} else if (callout_migrating(c)) {
1367 			/*
1368 			 * The callout is currently being serviced
1369 			 * and the "next" callout is scheduled at
1370 			 * its completion with a migration. We remove
1371 			 * the migration flag so it *won't* get rescheduled,
1372 			 * but we can't stop the one thats running so
1373 			 * we return 0.
1374 			 */
1375 			c->c_iflags &= ~CALLOUT_DFRMIGRATION;
1376 #ifdef SMP
1377 			/*
1378 			 * We can't call cc_cce_cleanup here since
1379 			 * if we do it will remove .ce_curr and
1380 			 * its still running. This will prevent a
1381 			 * reschedule of the callout when the
1382 			 * execution completes.
1383 			 */
1384 			cc_migration_cpu(cc, direct) = CPUBLOCK;
1385 			cc_migration_time(cc, direct) = 0;
1386 			cc_migration_prec(cc, direct) = 0;
1387 			cc_migration_func(cc, direct) = NULL;
1388 			cc_migration_arg(cc, direct) = NULL;
1389 #endif
1390 			CTR3(KTR_CALLOUT, "postponing stop %p func %p arg %p",
1391 			    c, c->c_func, c->c_arg);
1392  			if (drain) {
1393 				cc_exec_drain(cc, direct) = drain;
1394 			}
1395 			CC_UNLOCK(cc);
1396 			return ((flags & CS_EXECUTING) != 0);
1397 		}
1398 		CTR3(KTR_CALLOUT, "failed to stop %p func %p arg %p",
1399 		    c, c->c_func, c->c_arg);
1400 		if (drain) {
1401 			cc_exec_drain(cc, direct) = drain;
1402 		}
1403 		KASSERT(!sq_locked, ("sleepqueue chain still locked"));
1404 		cancelled = ((flags & CS_EXECUTING) != 0);
1405 	} else
1406 		cancelled = 1;
1407 
1408 	if (sq_locked)
1409 		sleepq_release(&cc_exec_waiting(cc, direct));
1410 
1411 	if ((c->c_iflags & CALLOUT_PENDING) == 0) {
1412 		CTR3(KTR_CALLOUT, "failed to stop %p func %p arg %p",
1413 		    c, c->c_func, c->c_arg);
1414 		/*
1415 		 * For not scheduled and not executing callout return
1416 		 * negative value.
1417 		 */
1418 		if (cc_exec_curr(cc, direct) != c)
1419 			cancelled = -1;
1420 		CC_UNLOCK(cc);
1421 		return (cancelled);
1422 	}
1423 
1424 	c->c_iflags &= ~CALLOUT_PENDING;
1425 	c->c_flags &= ~CALLOUT_ACTIVE;
1426 
1427 	CTR3(KTR_CALLOUT, "cancelled %p func %p arg %p",
1428 	    c, c->c_func, c->c_arg);
1429 	if (not_on_a_list == 0) {
1430 		if ((c->c_iflags & CALLOUT_PROCESSED) == 0) {
1431 			if (cc_exec_next(cc) == c)
1432 				cc_exec_next(cc) = LIST_NEXT(c, c_links.le);
1433 			LIST_REMOVE(c, c_links.le);
1434 		} else {
1435 			TAILQ_REMOVE(&cc->cc_expireq, c, c_links.tqe);
1436 		}
1437 	}
1438 	callout_cc_del(c, cc);
1439 	CC_UNLOCK(cc);
1440 	return (cancelled);
1441 }
1442 
1443 void
1444 callout_init(struct callout *c, int mpsafe)
1445 {
1446 	bzero(c, sizeof *c);
1447 	if (mpsafe) {
1448 		c->c_lock = NULL;
1449 		c->c_iflags = CALLOUT_RETURNUNLOCKED;
1450 	} else {
1451 		c->c_lock = &Giant.lock_object;
1452 		c->c_iflags = 0;
1453 	}
1454 	c->c_cpu = timeout_cpu;
1455 }
1456 
1457 void
1458 _callout_init_lock(struct callout *c, struct lock_object *lock, int flags)
1459 {
1460 	bzero(c, sizeof *c);
1461 	c->c_lock = lock;
1462 	KASSERT((flags & ~(CALLOUT_RETURNUNLOCKED | CALLOUT_SHAREDLOCK)) == 0,
1463 	    ("callout_init_lock: bad flags %d", flags));
1464 	KASSERT(lock != NULL || (flags & CALLOUT_RETURNUNLOCKED) == 0,
1465 	    ("callout_init_lock: CALLOUT_RETURNUNLOCKED with no lock"));
1466 	KASSERT(lock == NULL || !(LOCK_CLASS(lock)->lc_flags &
1467 	    (LC_SPINLOCK | LC_SLEEPABLE)), ("%s: invalid lock class",
1468 	    __func__));
1469 	c->c_iflags = flags & (CALLOUT_RETURNUNLOCKED | CALLOUT_SHAREDLOCK);
1470 	c->c_cpu = timeout_cpu;
1471 }
1472 
1473 #ifdef APM_FIXUP_CALLTODO
1474 /*
1475  * Adjust the kernel calltodo timeout list.  This routine is used after
1476  * an APM resume to recalculate the calltodo timer list values with the
1477  * number of hz's we have been sleeping.  The next hardclock() will detect
1478  * that there are fired timers and run softclock() to execute them.
1479  *
1480  * Please note, I have not done an exhaustive analysis of what code this
1481  * might break.  I am motivated to have my select()'s and alarm()'s that
1482  * have expired during suspend firing upon resume so that the applications
1483  * which set the timer can do the maintanence the timer was for as close
1484  * as possible to the originally intended time.  Testing this code for a
1485  * week showed that resuming from a suspend resulted in 22 to 25 timers
1486  * firing, which seemed independent on whether the suspend was 2 hours or
1487  * 2 days.  Your milage may vary.   - Ken Key <key@cs.utk.edu>
1488  */
1489 void
1490 adjust_timeout_calltodo(struct timeval *time_change)
1491 {
1492 	struct callout *p;
1493 	unsigned long delta_ticks;
1494 
1495 	/*
1496 	 * How many ticks were we asleep?
1497 	 * (stolen from tvtohz()).
1498 	 */
1499 
1500 	/* Don't do anything */
1501 	if (time_change->tv_sec < 0)
1502 		return;
1503 	else if (time_change->tv_sec <= LONG_MAX / 1000000)
1504 		delta_ticks = howmany(time_change->tv_sec * 1000000 +
1505 		    time_change->tv_usec, tick) + 1;
1506 	else if (time_change->tv_sec <= LONG_MAX / hz)
1507 		delta_ticks = time_change->tv_sec * hz +
1508 		    howmany(time_change->tv_usec, tick) + 1;
1509 	else
1510 		delta_ticks = LONG_MAX;
1511 
1512 	if (delta_ticks > INT_MAX)
1513 		delta_ticks = INT_MAX;
1514 
1515 	/*
1516 	 * Now rip through the timer calltodo list looking for timers
1517 	 * to expire.
1518 	 */
1519 
1520 	/* don't collide with softclock() */
1521 	CC_LOCK(cc);
1522 	for (p = calltodo.c_next; p != NULL; p = p->c_next) {
1523 		p->c_time -= delta_ticks;
1524 
1525 		/* Break if the timer had more time on it than delta_ticks */
1526 		if (p->c_time > 0)
1527 			break;
1528 
1529 		/* take back the ticks the timer didn't use (p->c_time <= 0) */
1530 		delta_ticks = -p->c_time;
1531 	}
1532 	CC_UNLOCK(cc);
1533 
1534 	return;
1535 }
1536 #endif /* APM_FIXUP_CALLTODO */
1537 
1538 static int
1539 flssbt(sbintime_t sbt)
1540 {
1541 
1542 	sbt += (uint64_t)sbt >> 1;
1543 	if (sizeof(long) >= sizeof(sbintime_t))
1544 		return (flsl(sbt));
1545 	if (sbt >= SBT_1S)
1546 		return (flsl(((uint64_t)sbt) >> 32) + 32);
1547 	return (flsl(sbt));
1548 }
1549 
1550 /*
1551  * Dump immediate statistic snapshot of the scheduled callouts.
1552  */
1553 static int
1554 sysctl_kern_callout_stat(SYSCTL_HANDLER_ARGS)
1555 {
1556 	struct callout *tmp;
1557 	struct callout_cpu *cc;
1558 	struct callout_list *sc;
1559 	sbintime_t maxpr, maxt, medpr, medt, now, spr, st, t;
1560 	int ct[64], cpr[64], ccpbk[32];
1561 	int error, val, i, count, tcum, pcum, maxc, c, medc;
1562 #ifdef SMP
1563 	int cpu;
1564 #endif
1565 
1566 	val = 0;
1567 	error = sysctl_handle_int(oidp, &val, 0, req);
1568 	if (error != 0 || req->newptr == NULL)
1569 		return (error);
1570 	count = maxc = 0;
1571 	st = spr = maxt = maxpr = 0;
1572 	bzero(ccpbk, sizeof(ccpbk));
1573 	bzero(ct, sizeof(ct));
1574 	bzero(cpr, sizeof(cpr));
1575 	now = sbinuptime();
1576 #ifdef SMP
1577 	CPU_FOREACH(cpu) {
1578 		cc = CC_CPU(cpu);
1579 #else
1580 		cc = CC_CPU(timeout_cpu);
1581 #endif
1582 		CC_LOCK(cc);
1583 		for (i = 0; i < callwheelsize; i++) {
1584 			sc = &cc->cc_callwheel[i];
1585 			c = 0;
1586 			LIST_FOREACH(tmp, sc, c_links.le) {
1587 				c++;
1588 				t = tmp->c_time - now;
1589 				if (t < 0)
1590 					t = 0;
1591 				st += t / SBT_1US;
1592 				spr += tmp->c_precision / SBT_1US;
1593 				if (t > maxt)
1594 					maxt = t;
1595 				if (tmp->c_precision > maxpr)
1596 					maxpr = tmp->c_precision;
1597 				ct[flssbt(t)]++;
1598 				cpr[flssbt(tmp->c_precision)]++;
1599 			}
1600 			if (c > maxc)
1601 				maxc = c;
1602 			ccpbk[fls(c + c / 2)]++;
1603 			count += c;
1604 		}
1605 		CC_UNLOCK(cc);
1606 #ifdef SMP
1607 	}
1608 #endif
1609 
1610 	for (i = 0, tcum = 0; i < 64 && tcum < count / 2; i++)
1611 		tcum += ct[i];
1612 	medt = (i >= 2) ? (((sbintime_t)1) << (i - 2)) : 0;
1613 	for (i = 0, pcum = 0; i < 64 && pcum < count / 2; i++)
1614 		pcum += cpr[i];
1615 	medpr = (i >= 2) ? (((sbintime_t)1) << (i - 2)) : 0;
1616 	for (i = 0, c = 0; i < 32 && c < count / 2; i++)
1617 		c += ccpbk[i];
1618 	medc = (i >= 2) ? (1 << (i - 2)) : 0;
1619 
1620 	printf("Scheduled callouts statistic snapshot:\n");
1621 	printf("  Callouts: %6d  Buckets: %6d*%-3d  Bucket size: 0.%06ds\n",
1622 	    count, callwheelsize, mp_ncpus, 1000000 >> CC_HASH_SHIFT);
1623 	printf("  C/Bk: med %5d         avg %6d.%06jd  max %6d\n",
1624 	    medc,
1625 	    count / callwheelsize / mp_ncpus,
1626 	    (uint64_t)count * 1000000 / callwheelsize / mp_ncpus % 1000000,
1627 	    maxc);
1628 	printf("  Time: med %5jd.%06jds avg %6jd.%06jds max %6jd.%06jds\n",
1629 	    medt / SBT_1S, (medt & 0xffffffff) * 1000000 >> 32,
1630 	    (st / count) / 1000000, (st / count) % 1000000,
1631 	    maxt / SBT_1S, (maxt & 0xffffffff) * 1000000 >> 32);
1632 	printf("  Prec: med %5jd.%06jds avg %6jd.%06jds max %6jd.%06jds\n",
1633 	    medpr / SBT_1S, (medpr & 0xffffffff) * 1000000 >> 32,
1634 	    (spr / count) / 1000000, (spr / count) % 1000000,
1635 	    maxpr / SBT_1S, (maxpr & 0xffffffff) * 1000000 >> 32);
1636 	printf("  Distribution:       \tbuckets\t   time\t   tcum\t"
1637 	    "   prec\t   pcum\n");
1638 	for (i = 0, tcum = pcum = 0; i < 64; i++) {
1639 		if (ct[i] == 0 && cpr[i] == 0)
1640 			continue;
1641 		t = (i != 0) ? (((sbintime_t)1) << (i - 1)) : 0;
1642 		tcum += ct[i];
1643 		pcum += cpr[i];
1644 		printf("  %10jd.%06jds\t 2**%d\t%7d\t%7d\t%7d\t%7d\n",
1645 		    t / SBT_1S, (t & 0xffffffff) * 1000000 >> 32,
1646 		    i - 1 - (32 - CC_HASH_SHIFT),
1647 		    ct[i], tcum, cpr[i], pcum);
1648 	}
1649 	return (error);
1650 }
1651 SYSCTL_PROC(_kern, OID_AUTO, callout_stat,
1652     CTLTYPE_INT | CTLFLAG_RW | CTLFLAG_MPSAFE,
1653     0, 0, sysctl_kern_callout_stat, "I",
1654     "Dump immediate statistic snapshot of the scheduled callouts");
1655 
1656 #ifdef DDB
1657 static void
1658 _show_callout(struct callout *c)
1659 {
1660 
1661 	db_printf("callout %p\n", c);
1662 #define	C_DB_PRINTF(f, e)	db_printf("   %s = " f "\n", #e, c->e);
1663 	db_printf("   &c_links = %p\n", &(c->c_links));
1664 	C_DB_PRINTF("%" PRId64,	c_time);
1665 	C_DB_PRINTF("%" PRId64,	c_precision);
1666 	C_DB_PRINTF("%p",	c_arg);
1667 	C_DB_PRINTF("%p",	c_func);
1668 	C_DB_PRINTF("%p",	c_lock);
1669 	C_DB_PRINTF("%#x",	c_flags);
1670 	C_DB_PRINTF("%#x",	c_iflags);
1671 	C_DB_PRINTF("%d",	c_cpu);
1672 #undef	C_DB_PRINTF
1673 }
1674 
1675 DB_SHOW_COMMAND(callout, db_show_callout)
1676 {
1677 
1678 	if (!have_addr) {
1679 		db_printf("usage: show callout <struct callout *>\n");
1680 		return;
1681 	}
1682 
1683 	_show_callout((struct callout *)addr);
1684 }
1685 
1686 static void
1687 _show_last_callout(int cpu, int direct, const char *dirstr)
1688 {
1689 	struct callout_cpu *cc;
1690 	void *func, *arg;
1691 
1692 	cc = CC_CPU(cpu);
1693 	func = cc_exec_last_func(cc, direct);
1694 	arg = cc_exec_last_arg(cc, direct);
1695 	db_printf("cpu %d last%s callout function: %p ", cpu, dirstr, func);
1696 	db_printsym((db_expr_t)func, DB_STGY_ANY);
1697 	db_printf("\ncpu %d last%s callout argument: %p\n", cpu, dirstr, arg);
1698 }
1699 
1700 DB_SHOW_COMMAND(callout_last, db_show_callout_last)
1701 {
1702 	int cpu, last;
1703 
1704 	if (have_addr) {
1705 		if (addr < 0 || addr > mp_maxid || CPU_ABSENT(addr)) {
1706 			db_printf("no such cpu: %d\n", (int)addr);
1707 			return;
1708 		}
1709 		cpu = last = addr;
1710 	} else {
1711 		cpu = 0;
1712 		last = mp_maxid;
1713 	}
1714 
1715 	while (cpu <= last) {
1716 		if (!CPU_ABSENT(cpu)) {
1717 			_show_last_callout(cpu, 0, "");
1718 			_show_last_callout(cpu, 1, " direct");
1719 		}
1720 		cpu++;
1721 	}
1722 }
1723 #endif /* DDB */
1724