xref: /freebsd/sys/kern/kern_timeout.c (revision d0b2dbfa0ecf2bbc9709efc5e20baf8e4b44bbbf)
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 #include "opt_callout_profiling.h"
41 #include "opt_ddb.h"
42 #include "opt_rss.h"
43 
44 #include <sys/param.h>
45 #include <sys/systm.h>
46 #include <sys/bus.h>
47 #include <sys/callout.h>
48 #include <sys/domainset.h>
49 #include <sys/file.h>
50 #include <sys/interrupt.h>
51 #include <sys/kernel.h>
52 #include <sys/ktr.h>
53 #include <sys/kthread.h>
54 #include <sys/lock.h>
55 #include <sys/malloc.h>
56 #include <sys/mutex.h>
57 #include <sys/proc.h>
58 #include <sys/random.h>
59 #include <sys/sched.h>
60 #include <sys/sdt.h>
61 #include <sys/sleepqueue.h>
62 #include <sys/sysctl.h>
63 #include <sys/smp.h>
64 #include <sys/unistd.h>
65 
66 #ifdef DDB
67 #include <ddb/ddb.h>
68 #include <ddb/db_sym.h>
69 #include <machine/_inttypes.h>
70 #endif
71 
72 #ifdef SMP
73 #include <machine/cpu.h>
74 #endif
75 
76 DPCPU_DECLARE(sbintime_t, hardclocktime);
77 
78 SDT_PROVIDER_DEFINE(callout_execute);
79 SDT_PROBE_DEFINE1(callout_execute, , , callout__start, "struct callout *");
80 SDT_PROBE_DEFINE1(callout_execute, , , callout__end, "struct callout *");
81 
82 static void	softclock_thread(void *arg);
83 
84 #ifdef CALLOUT_PROFILING
85 static int avg_depth;
86 SYSCTL_INT(_debug, OID_AUTO, to_avg_depth, CTLFLAG_RD, &avg_depth, 0,
87     "Average number of items examined per softclock call. Units = 1/1000");
88 static int avg_gcalls;
89 SYSCTL_INT(_debug, OID_AUTO, to_avg_gcalls, CTLFLAG_RD, &avg_gcalls, 0,
90     "Average number of Giant callouts made per softclock call. Units = 1/1000");
91 static int avg_lockcalls;
92 SYSCTL_INT(_debug, OID_AUTO, to_avg_lockcalls, CTLFLAG_RD, &avg_lockcalls, 0,
93     "Average number of lock callouts made per softclock call. Units = 1/1000");
94 static int avg_mpcalls;
95 SYSCTL_INT(_debug, OID_AUTO, to_avg_mpcalls, CTLFLAG_RD, &avg_mpcalls, 0,
96     "Average number of MP callouts made per softclock call. Units = 1/1000");
97 static int avg_depth_dir;
98 SYSCTL_INT(_debug, OID_AUTO, to_avg_depth_dir, CTLFLAG_RD, &avg_depth_dir, 0,
99     "Average number of direct callouts examined per callout_process call. "
100     "Units = 1/1000");
101 static int avg_lockcalls_dir;
102 SYSCTL_INT(_debug, OID_AUTO, to_avg_lockcalls_dir, CTLFLAG_RD,
103     &avg_lockcalls_dir, 0, "Average number of lock direct callouts made per "
104     "callout_process call. Units = 1/1000");
105 static int avg_mpcalls_dir;
106 SYSCTL_INT(_debug, OID_AUTO, to_avg_mpcalls_dir, CTLFLAG_RD, &avg_mpcalls_dir,
107     0, "Average number of MP direct callouts made per callout_process call. "
108     "Units = 1/1000");
109 #endif
110 
111 static int ncallout;
112 SYSCTL_INT(_kern, OID_AUTO, ncallout, CTLFLAG_RDTUN | CTLFLAG_NOFETCH, &ncallout, 0,
113     "Number of entries in callwheel and size of timeout() preallocation");
114 
115 #ifdef	RSS
116 static int pin_default_swi = 1;
117 static int pin_pcpu_swi = 1;
118 #else
119 static int pin_default_swi = 0;
120 static int pin_pcpu_swi = 0;
121 #endif
122 
123 SYSCTL_INT(_kern, OID_AUTO, pin_default_swi, CTLFLAG_RDTUN | CTLFLAG_NOFETCH, &pin_default_swi,
124     0, "Pin the default (non-per-cpu) swi (shared with PCPU 0 swi)");
125 SYSCTL_INT(_kern, OID_AUTO, pin_pcpu_swi, CTLFLAG_RDTUN | CTLFLAG_NOFETCH, &pin_pcpu_swi,
126     0, "Pin the per-CPU swis (except PCPU 0, which is also default)");
127 
128 /*
129  * TODO:
130  *	allocate more timeout table slots when table overflows.
131  */
132 static u_int __read_mostly callwheelsize;
133 static u_int __read_mostly 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 	callout_func_t		*cc_drain;
148 	void			*cc_last_func;
149 	void			*cc_last_arg;
150 #ifdef SMP
151 	callout_func_t		*ce_migration_func;
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_list	*cc_callwheel;
170 	struct callout_tailq	cc_expireq;
171 	sbintime_t		cc_firstevent;
172 	sbintime_t		cc_lastscan;
173 	struct thread		*cc_thread;
174 	u_int			cc_bucket;
175 #ifdef KTR
176 	char			cc_ktr_event_name[20];
177 #endif
178 };
179 
180 #define	callout_migrating(c)	((c)->c_iflags & CALLOUT_DFRMIGRATION)
181 
182 #define	cc_exec_curr(cc, dir)		cc->cc_exec_entity[dir].cc_curr
183 #define	cc_exec_last_func(cc, dir)	cc->cc_exec_entity[dir].cc_last_func
184 #define	cc_exec_last_arg(cc, dir)	cc->cc_exec_entity[dir].cc_last_arg
185 #define	cc_exec_drain(cc, dir)		cc->cc_exec_entity[dir].cc_drain
186 #define	cc_exec_next(cc)		cc->cc_next
187 #define	cc_exec_cancel(cc, dir)		cc->cc_exec_entity[dir].cc_cancel
188 #define	cc_exec_waiting(cc, dir)	cc->cc_exec_entity[dir].cc_waiting
189 #ifdef SMP
190 #define	cc_migration_func(cc, dir)	cc->cc_exec_entity[dir].ce_migration_func
191 #define	cc_migration_arg(cc, dir)	cc->cc_exec_entity[dir].ce_migration_arg
192 #define	cc_migration_cpu(cc, dir)	cc->cc_exec_entity[dir].ce_migration_cpu
193 #define	cc_migration_time(cc, dir)	cc->cc_exec_entity[dir].ce_migration_time
194 #define	cc_migration_prec(cc, dir)	cc->cc_exec_entity[dir].ce_migration_prec
195 
196 DPCPU_DEFINE_STATIC(struct callout_cpu, cc_cpu);
197 #define	CPUBLOCK	MAXCPU
198 #define	CC_CPU(cpu)	DPCPU_ID_PTR(cpu, cc_cpu)
199 #define	CC_SELF()	CC_CPU(PCPU_GET(cpuid))
200 #else
201 static struct callout_cpu cc_cpu;
202 #define	CC_CPU(cpu)	(&cc_cpu)
203 #define	CC_SELF()	(&cc_cpu)
204 #endif
205 #define	CC_LOCK(cc)	mtx_lock_spin(&(cc)->cc_lock)
206 #define	CC_UNLOCK(cc)	mtx_unlock_spin(&(cc)->cc_lock)
207 #define	CC_LOCK_ASSERT(cc)	mtx_assert(&(cc)->cc_lock, MA_OWNED)
208 
209 static int __read_mostly cc_default_cpu;
210 
211 static void	callout_cpu_init(struct callout_cpu *cc, int cpu);
212 static void	softclock_call_cc(struct callout *c, struct callout_cpu *cc,
213 #ifdef CALLOUT_PROFILING
214 		    int *mpcalls, int *lockcalls, int *gcalls,
215 #endif
216 		    int direct);
217 
218 static MALLOC_DEFINE(M_CALLOUT, "callout", "Callout datastructures");
219 
220 /**
221  * Locked by cc_lock:
222  *   cc_curr         - If a callout is in progress, it is cc_curr.
223  *                     If cc_curr is non-NULL, threads waiting in
224  *                     callout_drain() will be woken up as soon as the
225  *                     relevant callout completes.
226  *   cc_cancel       - Changing to 1 with both callout_lock and cc_lock held
227  *                     guarantees that the current callout will not run.
228  *                     The softclock_call_cc() function sets this to 0 before it
229  *                     drops callout_lock to acquire c_lock, and it calls
230  *                     the handler only if curr_cancelled is still 0 after
231  *                     cc_lock is successfully acquired.
232  *   cc_waiting      - If a thread is waiting in callout_drain(), then
233  *                     callout_wait is nonzero.  Set only when
234  *                     cc_curr is non-NULL.
235  */
236 
237 /*
238  * Resets the execution entity tied to a specific callout cpu.
239  */
240 static void
241 cc_cce_cleanup(struct callout_cpu *cc, int direct)
242 {
243 
244 	cc_exec_curr(cc, direct) = NULL;
245 	cc_exec_cancel(cc, direct) = false;
246 	cc_exec_waiting(cc, direct) = false;
247 #ifdef SMP
248 	cc_migration_cpu(cc, direct) = CPUBLOCK;
249 	cc_migration_time(cc, direct) = 0;
250 	cc_migration_prec(cc, direct) = 0;
251 	cc_migration_func(cc, direct) = NULL;
252 	cc_migration_arg(cc, direct) = NULL;
253 #endif
254 }
255 
256 /*
257  * Checks if migration is requested by a specific callout cpu.
258  */
259 static int
260 cc_cce_migrating(struct callout_cpu *cc, int direct)
261 {
262 
263 #ifdef SMP
264 	return (cc_migration_cpu(cc, direct) != CPUBLOCK);
265 #else
266 	return (0);
267 #endif
268 }
269 
270 /*
271  * Kernel low level callwheel initialization
272  * called on the BSP during kernel startup.
273  */
274 static void
275 callout_callwheel_init(void *dummy)
276 {
277 	struct callout_cpu *cc;
278 	int cpu;
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 	ncallout = imin(16 + maxproc + maxfiles, 18508);
287 	TUNABLE_INT_FETCH("kern.ncallout", &ncallout);
288 
289 	/*
290 	 * Calculate callout wheel size, should be next power of two higher
291 	 * than 'ncallout'.
292 	 */
293 	callwheelsize = 1 << fls(ncallout);
294 	callwheelmask = callwheelsize - 1;
295 
296 	/*
297 	 * Fetch whether we're pinning the swi's or not.
298 	 */
299 	TUNABLE_INT_FETCH("kern.pin_default_swi", &pin_default_swi);
300 	TUNABLE_INT_FETCH("kern.pin_pcpu_swi", &pin_pcpu_swi);
301 
302 	/*
303 	 * Initialize callout wheels.  The software interrupt threads
304 	 * are created later.
305 	 */
306 	cc_default_cpu = PCPU_GET(cpuid);
307 	CPU_FOREACH(cpu) {
308 		cc = CC_CPU(cpu);
309 		callout_cpu_init(cc, cpu);
310 	}
311 }
312 SYSINIT(callwheel_init, SI_SUB_CPU, SI_ORDER_ANY, callout_callwheel_init, NULL);
313 
314 /*
315  * Initialize the per-cpu callout structures.
316  */
317 static void
318 callout_cpu_init(struct callout_cpu *cc, int cpu)
319 {
320 	int i;
321 
322 	mtx_init(&cc->cc_lock, "callout", NULL, MTX_SPIN);
323 	cc->cc_callwheel = malloc_domainset(sizeof(struct callout_list) *
324 	    callwheelsize, M_CALLOUT,
325 	    DOMAINSET_PREF(pcpu_find(cpu)->pc_domain), M_WAITOK);
326 	for (i = 0; i < callwheelsize; i++)
327 		LIST_INIT(&cc->cc_callwheel[i]);
328 	TAILQ_INIT(&cc->cc_expireq);
329 	cc->cc_firstevent = SBT_MAX;
330 	for (i = 0; i < 2; i++)
331 		cc_cce_cleanup(cc, i);
332 #ifdef KTR
333 	snprintf(cc->cc_ktr_event_name, sizeof(cc->cc_ktr_event_name),
334 	    "callwheel cpu %d", cpu);
335 #endif
336 }
337 
338 #ifdef SMP
339 /*
340  * Switches the cpu tied to a specific callout.
341  * The function expects a locked incoming callout cpu and returns with
342  * locked outcoming callout cpu.
343  */
344 static struct callout_cpu *
345 callout_cpu_switch(struct callout *c, struct callout_cpu *cc, int new_cpu)
346 {
347 	struct callout_cpu *new_cc;
348 
349 	MPASS(c != NULL && cc != NULL);
350 	CC_LOCK_ASSERT(cc);
351 
352 	/*
353 	 * Avoid interrupts and preemption firing after the callout cpu
354 	 * is blocked in order to avoid deadlocks as the new thread
355 	 * may be willing to acquire the callout cpu lock.
356 	 */
357 	c->c_cpu = CPUBLOCK;
358 	spinlock_enter();
359 	CC_UNLOCK(cc);
360 	new_cc = CC_CPU(new_cpu);
361 	CC_LOCK(new_cc);
362 	spinlock_exit();
363 	c->c_cpu = new_cpu;
364 	return (new_cc);
365 }
366 #endif
367 
368 /*
369  * Start softclock threads.
370  */
371 static void
372 start_softclock(void *dummy)
373 {
374 	struct proc *p;
375 	struct thread *td;
376 	struct callout_cpu *cc;
377 	int cpu, error;
378 	bool pin_swi;
379 
380 	p = NULL;
381 	CPU_FOREACH(cpu) {
382 		cc = CC_CPU(cpu);
383 		error = kproc_kthread_add(softclock_thread, cc, &p, &td,
384 		    RFSTOPPED, 0, "clock", "clock (%d)", cpu);
385 		if (error != 0)
386 			panic("failed to create softclock thread for cpu %d: %d",
387 			    cpu, error);
388 		CC_LOCK(cc);
389 		cc->cc_thread = td;
390 		thread_lock(td);
391 		sched_class(td, PRI_ITHD);
392 		sched_ithread_prio(td, PI_SOFTCLOCK);
393 		TD_SET_IWAIT(td);
394 		thread_lock_set(td, (struct mtx *)&cc->cc_lock);
395 		thread_unlock(td);
396 		if (cpu == cc_default_cpu)
397 			pin_swi = pin_default_swi;
398 		else
399 			pin_swi = pin_pcpu_swi;
400 		if (pin_swi) {
401 			error = cpuset_setithread(td->td_tid, cpu);
402 			if (error != 0)
403 				printf("%s: %s clock couldn't be pinned to cpu %d: %d\n",
404 				    __func__, cpu == cc_default_cpu ?
405 				    "default" : "per-cpu", cpu, error);
406 		}
407 	}
408 }
409 SYSINIT(start_softclock, SI_SUB_SOFTINTR, SI_ORDER_FIRST, start_softclock, NULL);
410 
411 #define	CC_HASH_SHIFT	8
412 
413 static inline u_int
414 callout_hash(sbintime_t sbt)
415 {
416 
417 	return (sbt >> (32 - CC_HASH_SHIFT));
418 }
419 
420 static inline u_int
421 callout_get_bucket(sbintime_t sbt)
422 {
423 
424 	return (callout_hash(sbt) & callwheelmask);
425 }
426 
427 void
428 callout_process(sbintime_t now)
429 {
430 	struct callout_entropy {
431 		struct callout_cpu *cc;
432 		struct thread *td;
433 		sbintime_t now;
434 	} entropy;
435 	struct callout *c, *next;
436 	struct callout_cpu *cc;
437 	struct callout_list *sc;
438 	struct thread *td;
439 	sbintime_t first, last, lookahead, max, tmp_max;
440 	u_int firstb, lastb, nowb;
441 #ifdef CALLOUT_PROFILING
442 	int depth_dir = 0, mpcalls_dir = 0, lockcalls_dir = 0;
443 #endif
444 
445 	cc = CC_SELF();
446 	mtx_lock_spin_flags(&cc->cc_lock, MTX_QUIET);
447 
448 	/* Compute the buckets of the last scan and present times. */
449 	firstb = callout_hash(cc->cc_lastscan);
450 	cc->cc_lastscan = now;
451 	nowb = callout_hash(now);
452 
453 	/* Compute the last bucket and minimum time of the bucket after it. */
454 	if (nowb == firstb)
455 		lookahead = (SBT_1S / 16);
456 	else if (nowb - firstb == 1)
457 		lookahead = (SBT_1S / 8);
458 	else
459 		lookahead = SBT_1S;
460 	first = last = now;
461 	first += (lookahead / 2);
462 	last += lookahead;
463 	last &= (0xffffffffffffffffLLU << (32 - CC_HASH_SHIFT));
464 	lastb = callout_hash(last) - 1;
465 	max = last;
466 
467 	/*
468 	 * Check if we wrapped around the entire wheel from the last scan.
469 	 * In case, we need to scan entirely the wheel for pending callouts.
470 	 */
471 	if (lastb - firstb >= callwheelsize) {
472 		lastb = firstb + callwheelsize - 1;
473 		if (nowb - firstb >= callwheelsize)
474 			nowb = lastb;
475 	}
476 
477 	/* Iterate callwheel from firstb to nowb and then up to lastb. */
478 	do {
479 		sc = &cc->cc_callwheel[firstb & callwheelmask];
480 		LIST_FOREACH_SAFE(c, sc, c_links.le, next) {
481 			/* Run the callout if present time within allowed. */
482 			if (c->c_time <= now) {
483 				/*
484 				 * Consumer told us the callout may be run
485 				 * directly from hardware interrupt context.
486 				 */
487 				if (c->c_iflags & CALLOUT_DIRECT) {
488 #ifdef CALLOUT_PROFILING
489 					++depth_dir;
490 #endif
491 					cc_exec_next(cc) = next;
492 					cc->cc_bucket = firstb & callwheelmask;
493 					LIST_REMOVE(c, c_links.le);
494 					softclock_call_cc(c, cc,
495 #ifdef CALLOUT_PROFILING
496 					    &mpcalls_dir, &lockcalls_dir, NULL,
497 #endif
498 					    1);
499 					next = cc_exec_next(cc);
500 					cc_exec_next(cc) = NULL;
501 				} else {
502 					LIST_REMOVE(c, c_links.le);
503 					TAILQ_INSERT_TAIL(&cc->cc_expireq,
504 					    c, c_links.tqe);
505 					c->c_iflags |= CALLOUT_PROCESSED;
506 				}
507 			} else if (c->c_time >= max) {
508 				/*
509 				 * Skip events in the distant future.
510 				 */
511 				;
512 			} else if (c->c_time > last) {
513 				/*
514 				 * Event minimal time is bigger than present
515 				 * maximal time, so it cannot be aggregated.
516 				 */
517 				lastb = nowb;
518 			} else {
519 				/*
520 				 * Update first and last time, respecting this
521 				 * event.
522 				 */
523 				if (c->c_time < first)
524 					first = c->c_time;
525 				tmp_max = c->c_time + c->c_precision;
526 				if (tmp_max < last)
527 					last = tmp_max;
528 			}
529 		}
530 		/* Proceed with the next bucket. */
531 		firstb++;
532 		/*
533 		 * Stop if we looked after present time and found
534 		 * some event we can't execute at now.
535 		 * Stop if we looked far enough into the future.
536 		 */
537 	} while (((int)(firstb - lastb)) <= 0);
538 	cc->cc_firstevent = last;
539 	cpu_new_callout(curcpu, last, first);
540 
541 #ifdef CALLOUT_PROFILING
542 	avg_depth_dir += (depth_dir * 1000 - avg_depth_dir) >> 8;
543 	avg_mpcalls_dir += (mpcalls_dir * 1000 - avg_mpcalls_dir) >> 8;
544 	avg_lockcalls_dir += (lockcalls_dir * 1000 - avg_lockcalls_dir) >> 8;
545 #endif
546 	if (!TAILQ_EMPTY(&cc->cc_expireq)) {
547 		entropy.cc = cc;
548 		entropy.td = curthread;
549 		entropy.now = now;
550 		random_harvest_queue(&entropy, sizeof(entropy), RANDOM_CALLOUT);
551 
552 		td = cc->cc_thread;
553 		if (TD_AWAITING_INTR(td)) {
554 			thread_lock_block_wait(td);
555 			THREAD_LOCK_ASSERT(td, MA_OWNED);
556 			TD_CLR_IWAIT(td);
557 			sched_wakeup(td, SRQ_INTR);
558 		} else
559 			mtx_unlock_spin_flags(&cc->cc_lock, MTX_QUIET);
560 	} else
561 		mtx_unlock_spin_flags(&cc->cc_lock, MTX_QUIET);
562 }
563 
564 static struct callout_cpu *
565 callout_lock(struct callout *c)
566 {
567 	struct callout_cpu *cc;
568 	int cpu;
569 
570 	for (;;) {
571 		cpu = c->c_cpu;
572 #ifdef SMP
573 		if (cpu == CPUBLOCK) {
574 			while (c->c_cpu == CPUBLOCK)
575 				cpu_spinwait();
576 			continue;
577 		}
578 #endif
579 		cc = CC_CPU(cpu);
580 		CC_LOCK(cc);
581 		if (cpu == c->c_cpu)
582 			break;
583 		CC_UNLOCK(cc);
584 	}
585 	return (cc);
586 }
587 
588 static void
589 callout_cc_add(struct callout *c, struct callout_cpu *cc,
590     sbintime_t sbt, sbintime_t precision, void (*func)(void *),
591     void *arg, int flags)
592 {
593 	int bucket;
594 
595 	CC_LOCK_ASSERT(cc);
596 	if (sbt < cc->cc_lastscan)
597 		sbt = cc->cc_lastscan;
598 	c->c_arg = arg;
599 	c->c_iflags |= CALLOUT_PENDING;
600 	c->c_iflags &= ~CALLOUT_PROCESSED;
601 	c->c_flags |= CALLOUT_ACTIVE;
602 	if (flags & C_DIRECT_EXEC)
603 		c->c_iflags |= CALLOUT_DIRECT;
604 	c->c_func = func;
605 	c->c_time = sbt;
606 	c->c_precision = precision;
607 	bucket = callout_get_bucket(c->c_time);
608 	CTR3(KTR_CALLOUT, "precision set for %p: %d.%08x",
609 	    c, (int)(c->c_precision >> 32),
610 	    (u_int)(c->c_precision & 0xffffffff));
611 	LIST_INSERT_HEAD(&cc->cc_callwheel[bucket], c, c_links.le);
612 	if (cc->cc_bucket == bucket)
613 		cc_exec_next(cc) = c;
614 
615 	/*
616 	 * Inform the eventtimers(4) subsystem there's a new callout
617 	 * that has been inserted, but only if really required.
618 	 */
619 	if (SBT_MAX - c->c_time < c->c_precision)
620 		c->c_precision = SBT_MAX - c->c_time;
621 	sbt = c->c_time + c->c_precision;
622 	if (sbt < cc->cc_firstevent) {
623 		cc->cc_firstevent = sbt;
624 		cpu_new_callout(c->c_cpu, sbt, c->c_time);
625 	}
626 }
627 
628 static void
629 softclock_call_cc(struct callout *c, struct callout_cpu *cc,
630 #ifdef CALLOUT_PROFILING
631     int *mpcalls, int *lockcalls, int *gcalls,
632 #endif
633     int direct)
634 {
635 	struct rm_priotracker tracker;
636 	callout_func_t *c_func, *drain;
637 	void *c_arg;
638 	struct lock_class *class;
639 	struct lock_object *c_lock;
640 	uintptr_t lock_status;
641 	int c_iflags;
642 #ifdef SMP
643 	struct callout_cpu *new_cc;
644 	callout_func_t *new_func;
645 	void *new_arg;
646 	int flags, new_cpu;
647 	sbintime_t new_prec, new_time;
648 #endif
649 #if defined(DIAGNOSTIC) || defined(CALLOUT_PROFILING)
650 	sbintime_t sbt1, sbt2;
651 	struct timespec ts2;
652 	static sbintime_t maxdt = 2 * SBT_1MS;	/* 2 msec */
653 	static callout_func_t *lastfunc;
654 #endif
655 
656 	KASSERT((c->c_iflags & CALLOUT_PENDING) == CALLOUT_PENDING,
657 	    ("softclock_call_cc: pend %p %x", c, c->c_iflags));
658 	KASSERT((c->c_flags & CALLOUT_ACTIVE) == CALLOUT_ACTIVE,
659 	    ("softclock_call_cc: act %p %x", c, c->c_flags));
660 	class = (c->c_lock != NULL) ? LOCK_CLASS(c->c_lock) : NULL;
661 	lock_status = 0;
662 	if (c->c_iflags & CALLOUT_SHAREDLOCK) {
663 		if (class == &lock_class_rm)
664 			lock_status = (uintptr_t)&tracker;
665 		else
666 			lock_status = 1;
667 	}
668 	c_lock = c->c_lock;
669 	c_func = c->c_func;
670 	c_arg = c->c_arg;
671 	c_iflags = c->c_iflags;
672 	c->c_iflags &= ~CALLOUT_PENDING;
673 
674 	cc_exec_curr(cc, direct) = c;
675 	cc_exec_last_func(cc, direct) = c_func;
676 	cc_exec_last_arg(cc, direct) = c_arg;
677 	cc_exec_cancel(cc, direct) = false;
678 	cc_exec_drain(cc, direct) = NULL;
679 	CC_UNLOCK(cc);
680 	if (c_lock != NULL) {
681 		class->lc_lock(c_lock, lock_status);
682 		/*
683 		 * The callout may have been cancelled
684 		 * while we switched locks.
685 		 */
686 		if (cc_exec_cancel(cc, direct)) {
687 			class->lc_unlock(c_lock);
688 			goto skip;
689 		}
690 		/* The callout cannot be stopped now. */
691 		cc_exec_cancel(cc, direct) = true;
692 		if (c_lock == &Giant.lock_object) {
693 #ifdef CALLOUT_PROFILING
694 			(*gcalls)++;
695 #endif
696 			CTR3(KTR_CALLOUT, "callout giant %p func %p arg %p",
697 			    c, c_func, c_arg);
698 		} else {
699 #ifdef CALLOUT_PROFILING
700 			(*lockcalls)++;
701 #endif
702 			CTR3(KTR_CALLOUT, "callout lock %p func %p arg %p",
703 			    c, c_func, c_arg);
704 		}
705 	} else {
706 #ifdef CALLOUT_PROFILING
707 		(*mpcalls)++;
708 #endif
709 		CTR3(KTR_CALLOUT, "callout %p func %p arg %p",
710 		    c, c_func, c_arg);
711 	}
712 	KTR_STATE3(KTR_SCHED, "callout", cc->cc_ktr_event_name, "running",
713 	    "func:%p", c_func, "arg:%p", c_arg, "direct:%d", direct);
714 #if defined(DIAGNOSTIC) || defined(CALLOUT_PROFILING)
715 	sbt1 = sbinuptime();
716 #endif
717 	THREAD_NO_SLEEPING();
718 	SDT_PROBE1(callout_execute, , , callout__start, c);
719 	c_func(c_arg);
720 	SDT_PROBE1(callout_execute, , , callout__end, c);
721 	THREAD_SLEEPING_OK();
722 #if defined(DIAGNOSTIC) || defined(CALLOUT_PROFILING)
723 	sbt2 = sbinuptime();
724 	sbt2 -= sbt1;
725 	if (sbt2 > maxdt) {
726 		if (lastfunc != c_func || sbt2 > maxdt * 2) {
727 			ts2 = sbttots(sbt2);
728 			printf(
729 		"Expensive callout(9) function: %p(%p) %jd.%09ld s\n",
730 			    c_func, c_arg, (intmax_t)ts2.tv_sec, ts2.tv_nsec);
731 		}
732 		maxdt = sbt2;
733 		lastfunc = c_func;
734 	}
735 #endif
736 	KTR_STATE0(KTR_SCHED, "callout", cc->cc_ktr_event_name, "idle");
737 	CTR1(KTR_CALLOUT, "callout %p finished", c);
738 	if ((c_iflags & CALLOUT_RETURNUNLOCKED) == 0)
739 		class->lc_unlock(c_lock);
740 skip:
741 	CC_LOCK(cc);
742 	KASSERT(cc_exec_curr(cc, direct) == c, ("mishandled cc_curr"));
743 	cc_exec_curr(cc, direct) = NULL;
744 	if (cc_exec_drain(cc, direct)) {
745 		drain = cc_exec_drain(cc, direct);
746 		cc_exec_drain(cc, direct) = NULL;
747 		CC_UNLOCK(cc);
748 		drain(c_arg);
749 		CC_LOCK(cc);
750 	}
751 	if (cc_exec_waiting(cc, direct)) {
752 		/*
753 		 * There is someone waiting for the
754 		 * callout to complete.
755 		 * If the callout was scheduled for
756 		 * migration just cancel it.
757 		 */
758 		if (cc_cce_migrating(cc, direct)) {
759 			cc_cce_cleanup(cc, direct);
760 
761 			/*
762 			 * It should be assert here that the callout is not
763 			 * destroyed but that is not easy.
764 			 */
765 			c->c_iflags &= ~CALLOUT_DFRMIGRATION;
766 		}
767 		cc_exec_waiting(cc, direct) = false;
768 		CC_UNLOCK(cc);
769 		wakeup(&cc_exec_waiting(cc, direct));
770 		CC_LOCK(cc);
771 	} else if (cc_cce_migrating(cc, direct)) {
772 #ifdef SMP
773 		/*
774 		 * If the callout was scheduled for
775 		 * migration just perform it now.
776 		 */
777 		new_cpu = cc_migration_cpu(cc, direct);
778 		new_time = cc_migration_time(cc, direct);
779 		new_prec = cc_migration_prec(cc, direct);
780 		new_func = cc_migration_func(cc, direct);
781 		new_arg = cc_migration_arg(cc, direct);
782 		cc_cce_cleanup(cc, direct);
783 
784 		/*
785 		 * It should be assert here that the callout is not destroyed
786 		 * but that is not easy.
787 		 *
788 		 * As first thing, handle deferred callout stops.
789 		 */
790 		if (!callout_migrating(c)) {
791 			CTR3(KTR_CALLOUT,
792 			     "deferred cancelled %p func %p arg %p",
793 			     c, new_func, new_arg);
794 			return;
795 		}
796 		c->c_iflags &= ~CALLOUT_DFRMIGRATION;
797 
798 		new_cc = callout_cpu_switch(c, cc, new_cpu);
799 		flags = (direct) ? C_DIRECT_EXEC : 0;
800 		callout_cc_add(c, new_cc, new_time, new_prec, new_func,
801 		    new_arg, flags);
802 		CC_UNLOCK(new_cc);
803 		CC_LOCK(cc);
804 #else
805 		panic("migration should not happen");
806 #endif
807 	}
808 }
809 
810 /*
811  * The callout mechanism is based on the work of Adam M. Costello and
812  * George Varghese, published in a technical report entitled "Redesigning
813  * the BSD Callout and Timer Facilities" and modified slightly for inclusion
814  * in FreeBSD by Justin T. Gibbs.  The original work on the data structures
815  * used in this implementation was published by G. Varghese and T. Lauck in
816  * the paper "Hashed and Hierarchical Timing Wheels: Data Structures for
817  * the Efficient Implementation of a Timer Facility" in the Proceedings of
818  * the 11th ACM Annual Symposium on Operating Systems Principles,
819  * Austin, Texas Nov 1987.
820  */
821 
822 /*
823  * Software (low priority) clock interrupt thread handler.
824  * Run periodic events from timeout queue.
825  */
826 static void
827 softclock_thread(void *arg)
828 {
829 	struct thread *td = curthread;
830 	struct callout_cpu *cc;
831 	struct callout *c;
832 #ifdef CALLOUT_PROFILING
833 	int depth, gcalls, lockcalls, mpcalls;
834 #endif
835 
836 	cc = (struct callout_cpu *)arg;
837 	CC_LOCK(cc);
838 	for (;;) {
839 		while (TAILQ_EMPTY(&cc->cc_expireq)) {
840 			/*
841 			 * Use CC_LOCK(cc) as the thread_lock while
842 			 * idle.
843 			 */
844 			thread_lock(td);
845 			thread_lock_set(td, (struct mtx *)&cc->cc_lock);
846 			TD_SET_IWAIT(td);
847 			mi_switch(SW_VOL | SWT_IWAIT);
848 
849 			/* mi_switch() drops thread_lock(). */
850 			CC_LOCK(cc);
851 		}
852 
853 #ifdef CALLOUT_PROFILING
854 		depth = gcalls = lockcalls = mpcalls = 0;
855 #endif
856 		while ((c = TAILQ_FIRST(&cc->cc_expireq)) != NULL) {
857 			TAILQ_REMOVE(&cc->cc_expireq, c, c_links.tqe);
858 			softclock_call_cc(c, cc,
859 #ifdef CALLOUT_PROFILING
860 			    &mpcalls, &lockcalls, &gcalls,
861 #endif
862 			    0);
863 #ifdef CALLOUT_PROFILING
864 			++depth;
865 #endif
866 		}
867 #ifdef CALLOUT_PROFILING
868 		avg_depth += (depth * 1000 - avg_depth) >> 8;
869 		avg_mpcalls += (mpcalls * 1000 - avg_mpcalls) >> 8;
870 		avg_lockcalls += (lockcalls * 1000 - avg_lockcalls) >> 8;
871 		avg_gcalls += (gcalls * 1000 - avg_gcalls) >> 8;
872 #endif
873 	}
874 }
875 
876 void
877 callout_when(sbintime_t sbt, sbintime_t precision, int flags,
878     sbintime_t *res, sbintime_t *prec_res)
879 {
880 	sbintime_t to_sbt, to_pr;
881 
882 	if ((flags & (C_ABSOLUTE | C_PRECALC)) != 0) {
883 		*res = sbt;
884 		*prec_res = precision;
885 		return;
886 	}
887 	if ((flags & C_HARDCLOCK) != 0 && sbt < tick_sbt)
888 		sbt = tick_sbt;
889 	if ((flags & C_HARDCLOCK) != 0 || sbt >= sbt_tickthreshold) {
890 		/*
891 		 * Obtain the time of the last hardclock() call on
892 		 * this CPU directly from the kern_clocksource.c.
893 		 * This value is per-CPU, but it is equal for all
894 		 * active ones.
895 		 */
896 #ifdef __LP64__
897 		to_sbt = DPCPU_GET(hardclocktime);
898 #else
899 		spinlock_enter();
900 		to_sbt = DPCPU_GET(hardclocktime);
901 		spinlock_exit();
902 #endif
903 		if (cold && to_sbt == 0)
904 			to_sbt = sbinuptime();
905 		if ((flags & C_HARDCLOCK) == 0)
906 			to_sbt += tick_sbt;
907 	} else
908 		to_sbt = sbinuptime();
909 	if (SBT_MAX - to_sbt < sbt)
910 		to_sbt = SBT_MAX;
911 	else
912 		to_sbt += sbt;
913 	*res = to_sbt;
914 	to_pr = ((C_PRELGET(flags) < 0) ? sbt >> tc_precexp :
915 	    sbt >> C_PRELGET(flags));
916 	*prec_res = to_pr > precision ? to_pr : precision;
917 }
918 
919 /*
920  * New interface; clients allocate their own callout structures.
921  *
922  * callout_reset() - establish or change a timeout
923  * callout_stop() - disestablish a timeout
924  * callout_init() - initialize a callout structure so that it can
925  *	safely be passed to callout_reset() and callout_stop()
926  *
927  * <sys/callout.h> defines three convenience macros:
928  *
929  * callout_active() - returns truth if callout has not been stopped,
930  *	drained, or deactivated since the last time the callout was
931  *	reset.
932  * callout_pending() - returns truth if callout is still waiting for timeout
933  * callout_deactivate() - marks the callout as having been serviced
934  */
935 int
936 callout_reset_sbt_on(struct callout *c, sbintime_t sbt, sbintime_t prec,
937     callout_func_t *ftn, void *arg, int cpu, int flags)
938 {
939 	sbintime_t to_sbt, precision;
940 	struct callout_cpu *cc;
941 	int cancelled, direct;
942 
943 	cancelled = 0;
944 	callout_when(sbt, prec, flags, &to_sbt, &precision);
945 
946 	/*
947 	 * This flag used to be added by callout_cc_add, but the
948 	 * first time you call this we could end up with the
949 	 * wrong direct flag if we don't do it before we add.
950 	 */
951 	if (flags & C_DIRECT_EXEC) {
952 		direct = 1;
953 	} else {
954 		direct = 0;
955 	}
956 	KASSERT(!direct || c->c_lock == NULL ||
957 	    (LOCK_CLASS(c->c_lock)->lc_flags & LC_SPINLOCK),
958 	    ("%s: direct callout %p has non-spin lock", __func__, c));
959 
960 	cc = callout_lock(c);
961 	if (cpu == -1)
962 		cpu = c->c_cpu;
963 	KASSERT(cpu >= 0 && cpu <= mp_maxid && !CPU_ABSENT(cpu),
964 	    ("%s: invalid cpu %d", __func__, cpu));
965 
966 	if (cc_exec_curr(cc, direct) == c) {
967 		/*
968 		 * We're being asked to reschedule a callout which is
969 		 * currently in progress.  If there is a lock then we
970 		 * can cancel the callout if it has not really started.
971 		 */
972 		if (c->c_lock != NULL && !cc_exec_cancel(cc, direct))
973 			cancelled = cc_exec_cancel(cc, direct) = true;
974 		if (cc_exec_waiting(cc, direct) || cc_exec_drain(cc, direct)) {
975 			/*
976 			 * Someone has called callout_drain to kill this
977 			 * callout.  Don't reschedule.
978 			 */
979 			CTR4(KTR_CALLOUT, "%s %p func %p arg %p",
980 			    cancelled ? "cancelled" : "failed to cancel",
981 			    c, c->c_func, c->c_arg);
982 			CC_UNLOCK(cc);
983 			return (cancelled);
984 		}
985 #ifdef SMP
986 		if (callout_migrating(c)) {
987 			/*
988 			 * This only occurs when a second callout_reset_sbt_on
989 			 * is made after a previous one moved it into
990 			 * deferred migration (below). Note we do *not* change
991 			 * the prev_cpu even though the previous target may
992 			 * be different.
993 			 */
994 			cc_migration_cpu(cc, direct) = cpu;
995 			cc_migration_time(cc, direct) = to_sbt;
996 			cc_migration_prec(cc, direct) = precision;
997 			cc_migration_func(cc, direct) = ftn;
998 			cc_migration_arg(cc, direct) = arg;
999 			cancelled = 1;
1000 			CC_UNLOCK(cc);
1001 			return (cancelled);
1002 		}
1003 #endif
1004 	}
1005 	if (c->c_iflags & CALLOUT_PENDING) {
1006 		if ((c->c_iflags & CALLOUT_PROCESSED) == 0) {
1007 			if (cc_exec_next(cc) == c)
1008 				cc_exec_next(cc) = LIST_NEXT(c, c_links.le);
1009 			LIST_REMOVE(c, c_links.le);
1010 		} else {
1011 			TAILQ_REMOVE(&cc->cc_expireq, c, c_links.tqe);
1012 		}
1013 		cancelled = 1;
1014 		c->c_iflags &= ~ CALLOUT_PENDING;
1015 		c->c_flags &= ~ CALLOUT_ACTIVE;
1016 	}
1017 
1018 #ifdef SMP
1019 	/*
1020 	 * If the callout must migrate try to perform it immediately.
1021 	 * If the callout is currently running, just defer the migration
1022 	 * to a more appropriate moment.
1023 	 */
1024 	if (c->c_cpu != cpu) {
1025 		if (cc_exec_curr(cc, direct) == c) {
1026 			/*
1027 			 * Pending will have been removed since we are
1028 			 * actually executing the callout on another
1029 			 * CPU. That callout should be waiting on the
1030 			 * lock the caller holds. If we set both
1031 			 * active/and/pending after we return and the
1032 			 * lock on the executing callout proceeds, it
1033 			 * will then see pending is true and return.
1034 			 * At the return from the actual callout execution
1035 			 * the migration will occur in softclock_call_cc
1036 			 * and this new callout will be placed on the
1037 			 * new CPU via a call to callout_cpu_switch() which
1038 			 * will get the lock on the right CPU followed
1039 			 * by a call callout_cc_add() which will add it there.
1040 			 * (see above in softclock_call_cc()).
1041 			 */
1042 			cc_migration_cpu(cc, direct) = cpu;
1043 			cc_migration_time(cc, direct) = to_sbt;
1044 			cc_migration_prec(cc, direct) = precision;
1045 			cc_migration_func(cc, direct) = ftn;
1046 			cc_migration_arg(cc, direct) = arg;
1047 			c->c_iflags |= (CALLOUT_DFRMIGRATION | CALLOUT_PENDING);
1048 			c->c_flags |= CALLOUT_ACTIVE;
1049 			CTR6(KTR_CALLOUT,
1050 		    "migration of %p func %p arg %p in %d.%08x to %u deferred",
1051 			    c, c->c_func, c->c_arg, (int)(to_sbt >> 32),
1052 			    (u_int)(to_sbt & 0xffffffff), cpu);
1053 			CC_UNLOCK(cc);
1054 			return (cancelled);
1055 		}
1056 		cc = callout_cpu_switch(c, cc, cpu);
1057 	}
1058 #endif
1059 
1060 	callout_cc_add(c, cc, to_sbt, precision, ftn, arg, flags);
1061 	CTR6(KTR_CALLOUT, "%sscheduled %p func %p arg %p in %d.%08x",
1062 	    cancelled ? "re" : "", c, c->c_func, c->c_arg, (int)(to_sbt >> 32),
1063 	    (u_int)(to_sbt & 0xffffffff));
1064 	CC_UNLOCK(cc);
1065 
1066 	return (cancelled);
1067 }
1068 
1069 /*
1070  * Common idioms that can be optimized in the future.
1071  */
1072 int
1073 callout_schedule_on(struct callout *c, int to_ticks, int cpu)
1074 {
1075 	return callout_reset_on(c, to_ticks, c->c_func, c->c_arg, cpu);
1076 }
1077 
1078 int
1079 callout_schedule(struct callout *c, int to_ticks)
1080 {
1081 	return callout_reset_on(c, to_ticks, c->c_func, c->c_arg, c->c_cpu);
1082 }
1083 
1084 int
1085 _callout_stop_safe(struct callout *c, int flags, callout_func_t *drain)
1086 {
1087 	struct callout_cpu *cc, *old_cc;
1088 	struct lock_class *class;
1089 	int direct, sq_locked, use_lock;
1090 	int cancelled, not_on_a_list;
1091 
1092 	if ((flags & CS_DRAIN) != 0)
1093 		WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK, c->c_lock,
1094 		    "calling %s", __func__);
1095 
1096 	KASSERT((flags & CS_DRAIN) == 0 || drain == NULL,
1097 	    ("Cannot set drain callback and CS_DRAIN flag at the same time"));
1098 
1099 	/*
1100 	 * Some old subsystems don't hold Giant while running a callout_stop(),
1101 	 * so just discard this check for the moment.
1102 	 */
1103 	if ((flags & CS_DRAIN) == 0 && c->c_lock != NULL) {
1104 		if (c->c_lock == &Giant.lock_object)
1105 			use_lock = mtx_owned(&Giant);
1106 		else {
1107 			use_lock = 1;
1108 			class = LOCK_CLASS(c->c_lock);
1109 			class->lc_assert(c->c_lock, LA_XLOCKED);
1110 		}
1111 	} else
1112 		use_lock = 0;
1113 	if (c->c_iflags & CALLOUT_DIRECT) {
1114 		direct = 1;
1115 	} else {
1116 		direct = 0;
1117 	}
1118 	sq_locked = 0;
1119 	old_cc = NULL;
1120 again:
1121 	cc = callout_lock(c);
1122 
1123 	if ((c->c_iflags & (CALLOUT_DFRMIGRATION | CALLOUT_PENDING)) ==
1124 	    (CALLOUT_DFRMIGRATION | CALLOUT_PENDING) &&
1125 	    ((c->c_flags & CALLOUT_ACTIVE) == CALLOUT_ACTIVE)) {
1126 		/*
1127 		 * Special case where this slipped in while we
1128 		 * were migrating *as* the callout is about to
1129 		 * execute. The caller probably holds the lock
1130 		 * the callout wants.
1131 		 *
1132 		 * Get rid of the migration first. Then set
1133 		 * the flag that tells this code *not* to
1134 		 * try to remove it from any lists (its not
1135 		 * on one yet). When the callout wheel runs,
1136 		 * it will ignore this callout.
1137 		 */
1138 		c->c_iflags &= ~CALLOUT_PENDING;
1139 		c->c_flags &= ~CALLOUT_ACTIVE;
1140 		not_on_a_list = 1;
1141 	} else {
1142 		not_on_a_list = 0;
1143 	}
1144 
1145 	/*
1146 	 * If the callout was migrating while the callout cpu lock was
1147 	 * dropped,  just drop the sleepqueue lock and check the states
1148 	 * again.
1149 	 */
1150 	if (sq_locked != 0 && cc != old_cc) {
1151 #ifdef SMP
1152 		CC_UNLOCK(cc);
1153 		sleepq_release(&cc_exec_waiting(old_cc, direct));
1154 		sq_locked = 0;
1155 		old_cc = NULL;
1156 		goto again;
1157 #else
1158 		panic("migration should not happen");
1159 #endif
1160 	}
1161 
1162 	/*
1163 	 * If the callout is running, try to stop it or drain it.
1164 	 */
1165 	if (cc_exec_curr(cc, direct) == c) {
1166 		/*
1167 		 * Succeed we to stop it or not, we must clear the
1168 		 * active flag - this is what API users expect.  If we're
1169 		 * draining and the callout is currently executing, first wait
1170 		 * until it finishes.
1171 		 */
1172 		if ((flags & CS_DRAIN) == 0)
1173 			c->c_flags &= ~CALLOUT_ACTIVE;
1174 
1175 		if ((flags & CS_DRAIN) != 0) {
1176 			/*
1177 			 * The current callout is running (or just
1178 			 * about to run) and blocking is allowed, so
1179 			 * just wait for the current invocation to
1180 			 * finish.
1181 			 */
1182 			if (cc_exec_curr(cc, direct) == c) {
1183 				/*
1184 				 * Use direct calls to sleepqueue interface
1185 				 * instead of cv/msleep in order to avoid
1186 				 * a LOR between cc_lock and sleepqueue
1187 				 * chain spinlocks.  This piece of code
1188 				 * emulates a msleep_spin() call actually.
1189 				 *
1190 				 * If we already have the sleepqueue chain
1191 				 * locked, then we can safely block.  If we
1192 				 * don't already have it locked, however,
1193 				 * we have to drop the cc_lock to lock
1194 				 * it.  This opens several races, so we
1195 				 * restart at the beginning once we have
1196 				 * both locks.  If nothing has changed, then
1197 				 * we will end up back here with sq_locked
1198 				 * set.
1199 				 */
1200 				if (!sq_locked) {
1201 					CC_UNLOCK(cc);
1202 					sleepq_lock(
1203 					    &cc_exec_waiting(cc, direct));
1204 					sq_locked = 1;
1205 					old_cc = cc;
1206 					goto again;
1207 				}
1208 
1209 				/*
1210 				 * Migration could be cancelled here, but
1211 				 * as long as it is still not sure when it
1212 				 * will be packed up, just let softclock()
1213 				 * take care of it.
1214 				 */
1215 				cc_exec_waiting(cc, direct) = true;
1216 				DROP_GIANT();
1217 				CC_UNLOCK(cc);
1218 				sleepq_add(
1219 				    &cc_exec_waiting(cc, direct),
1220 				    &cc->cc_lock.lock_object, "codrain",
1221 				    SLEEPQ_SLEEP, 0);
1222 				sleepq_wait(
1223 				    &cc_exec_waiting(cc, direct),
1224 					     0);
1225 				sq_locked = 0;
1226 				old_cc = NULL;
1227 
1228 				/* Reacquire locks previously released. */
1229 				PICKUP_GIANT();
1230 				goto again;
1231 			}
1232 			c->c_flags &= ~CALLOUT_ACTIVE;
1233 		} else if (use_lock &&
1234 			   !cc_exec_cancel(cc, direct) && (drain == NULL)) {
1235 
1236 			/*
1237 			 * The current callout is waiting for its
1238 			 * lock which we hold.  Cancel the callout
1239 			 * and return.  After our caller drops the
1240 			 * lock, the callout will be skipped in
1241 			 * softclock(). This *only* works with a
1242 			 * callout_stop() *not* callout_drain() or
1243 			 * callout_async_drain().
1244 			 */
1245 			cc_exec_cancel(cc, direct) = true;
1246 			CTR3(KTR_CALLOUT, "cancelled %p func %p arg %p",
1247 			    c, c->c_func, c->c_arg);
1248 			KASSERT(!cc_cce_migrating(cc, direct),
1249 			    ("callout wrongly scheduled for migration"));
1250 			if (callout_migrating(c)) {
1251 				c->c_iflags &= ~CALLOUT_DFRMIGRATION;
1252 #ifdef SMP
1253 				cc_migration_cpu(cc, direct) = CPUBLOCK;
1254 				cc_migration_time(cc, direct) = 0;
1255 				cc_migration_prec(cc, direct) = 0;
1256 				cc_migration_func(cc, direct) = NULL;
1257 				cc_migration_arg(cc, direct) = NULL;
1258 #endif
1259 			}
1260 			CC_UNLOCK(cc);
1261 			KASSERT(!sq_locked, ("sleepqueue chain locked"));
1262 			return (1);
1263 		} else if (callout_migrating(c)) {
1264 			/*
1265 			 * The callout is currently being serviced
1266 			 * and the "next" callout is scheduled at
1267 			 * its completion with a migration. We remove
1268 			 * the migration flag so it *won't* get rescheduled,
1269 			 * but we can't stop the one thats running so
1270 			 * we return 0.
1271 			 */
1272 			c->c_iflags &= ~CALLOUT_DFRMIGRATION;
1273 #ifdef SMP
1274 			/*
1275 			 * We can't call cc_cce_cleanup here since
1276 			 * if we do it will remove .ce_curr and
1277 			 * its still running. This will prevent a
1278 			 * reschedule of the callout when the
1279 			 * execution completes.
1280 			 */
1281 			cc_migration_cpu(cc, direct) = CPUBLOCK;
1282 			cc_migration_time(cc, direct) = 0;
1283 			cc_migration_prec(cc, direct) = 0;
1284 			cc_migration_func(cc, direct) = NULL;
1285 			cc_migration_arg(cc, direct) = NULL;
1286 #endif
1287 			CTR3(KTR_CALLOUT, "postponing stop %p func %p arg %p",
1288 			    c, c->c_func, c->c_arg);
1289  			if (drain) {
1290 				KASSERT(cc_exec_drain(cc, direct) == NULL,
1291 				    ("callout drain function already set to %p",
1292 				    cc_exec_drain(cc, direct)));
1293 				cc_exec_drain(cc, direct) = drain;
1294 			}
1295 			CC_UNLOCK(cc);
1296 			return (0);
1297 		} else {
1298 			CTR3(KTR_CALLOUT, "failed to stop %p func %p arg %p",
1299 			    c, c->c_func, c->c_arg);
1300 			if (drain) {
1301 				KASSERT(cc_exec_drain(cc, direct) == NULL,
1302 				    ("callout drain function already set to %p",
1303 				    cc_exec_drain(cc, direct)));
1304 				cc_exec_drain(cc, direct) = drain;
1305 			}
1306 		}
1307 		KASSERT(!sq_locked, ("sleepqueue chain still locked"));
1308 		cancelled = 0;
1309 	} else
1310 		cancelled = 1;
1311 
1312 	if (sq_locked)
1313 		sleepq_release(&cc_exec_waiting(cc, direct));
1314 
1315 	if ((c->c_iflags & CALLOUT_PENDING) == 0) {
1316 		CTR3(KTR_CALLOUT, "failed to stop %p func %p arg %p",
1317 		    c, c->c_func, c->c_arg);
1318 		/*
1319 		 * For not scheduled and not executing callout return
1320 		 * negative value.
1321 		 */
1322 		if (cc_exec_curr(cc, direct) != c)
1323 			cancelled = -1;
1324 		CC_UNLOCK(cc);
1325 		return (cancelled);
1326 	}
1327 
1328 	c->c_iflags &= ~CALLOUT_PENDING;
1329 	c->c_flags &= ~CALLOUT_ACTIVE;
1330 
1331 	CTR3(KTR_CALLOUT, "cancelled %p func %p arg %p",
1332 	    c, c->c_func, c->c_arg);
1333 	if (not_on_a_list == 0) {
1334 		if ((c->c_iflags & CALLOUT_PROCESSED) == 0) {
1335 			if (cc_exec_next(cc) == c)
1336 				cc_exec_next(cc) = LIST_NEXT(c, c_links.le);
1337 			LIST_REMOVE(c, c_links.le);
1338 		} else {
1339 			TAILQ_REMOVE(&cc->cc_expireq, c, c_links.tqe);
1340 		}
1341 	}
1342 	CC_UNLOCK(cc);
1343 	return (cancelled);
1344 }
1345 
1346 void
1347 callout_init(struct callout *c, int mpsafe)
1348 {
1349 	bzero(c, sizeof *c);
1350 	if (mpsafe) {
1351 		c->c_lock = NULL;
1352 		c->c_iflags = CALLOUT_RETURNUNLOCKED;
1353 	} else {
1354 		c->c_lock = &Giant.lock_object;
1355 		c->c_iflags = 0;
1356 	}
1357 	c->c_cpu = cc_default_cpu;
1358 }
1359 
1360 void
1361 _callout_init_lock(struct callout *c, struct lock_object *lock, int flags)
1362 {
1363 	bzero(c, sizeof *c);
1364 	c->c_lock = lock;
1365 	KASSERT((flags & ~(CALLOUT_RETURNUNLOCKED | CALLOUT_SHAREDLOCK)) == 0,
1366 	    ("callout_init_lock: bad flags %d", flags));
1367 	KASSERT(lock != NULL || (flags & CALLOUT_RETURNUNLOCKED) == 0,
1368 	    ("callout_init_lock: CALLOUT_RETURNUNLOCKED with no lock"));
1369 	KASSERT(lock == NULL || !(LOCK_CLASS(lock)->lc_flags & LC_SLEEPABLE),
1370 	    ("%s: callout %p has sleepable lock", __func__, c));
1371 	c->c_iflags = flags & (CALLOUT_RETURNUNLOCKED | CALLOUT_SHAREDLOCK);
1372 	c->c_cpu = cc_default_cpu;
1373 }
1374 
1375 static int
1376 flssbt(sbintime_t sbt)
1377 {
1378 
1379 	sbt += (uint64_t)sbt >> 1;
1380 	if (sizeof(long) >= sizeof(sbintime_t))
1381 		return (flsl(sbt));
1382 	if (sbt >= SBT_1S)
1383 		return (flsl(((uint64_t)sbt) >> 32) + 32);
1384 	return (flsl(sbt));
1385 }
1386 
1387 /*
1388  * Dump immediate statistic snapshot of the scheduled callouts.
1389  */
1390 static int
1391 sysctl_kern_callout_stat(SYSCTL_HANDLER_ARGS)
1392 {
1393 	struct callout *tmp;
1394 	struct callout_cpu *cc;
1395 	struct callout_list *sc;
1396 	sbintime_t maxpr, maxt, medpr, medt, now, spr, st, t;
1397 	int ct[64], cpr[64], ccpbk[32];
1398 	int error, val, i, count, tcum, pcum, maxc, c, medc;
1399 	int cpu;
1400 
1401 	val = 0;
1402 	error = sysctl_handle_int(oidp, &val, 0, req);
1403 	if (error != 0 || req->newptr == NULL)
1404 		return (error);
1405 	count = maxc = 0;
1406 	st = spr = maxt = maxpr = 0;
1407 	bzero(ccpbk, sizeof(ccpbk));
1408 	bzero(ct, sizeof(ct));
1409 	bzero(cpr, sizeof(cpr));
1410 	now = sbinuptime();
1411 	CPU_FOREACH(cpu) {
1412 		cc = CC_CPU(cpu);
1413 		CC_LOCK(cc);
1414 		for (i = 0; i < callwheelsize; i++) {
1415 			sc = &cc->cc_callwheel[i];
1416 			c = 0;
1417 			LIST_FOREACH(tmp, sc, c_links.le) {
1418 				c++;
1419 				t = tmp->c_time - now;
1420 				if (t < 0)
1421 					t = 0;
1422 				st += t / SBT_1US;
1423 				spr += tmp->c_precision / SBT_1US;
1424 				if (t > maxt)
1425 					maxt = t;
1426 				if (tmp->c_precision > maxpr)
1427 					maxpr = tmp->c_precision;
1428 				ct[flssbt(t)]++;
1429 				cpr[flssbt(tmp->c_precision)]++;
1430 			}
1431 			if (c > maxc)
1432 				maxc = c;
1433 			ccpbk[fls(c + c / 2)]++;
1434 			count += c;
1435 		}
1436 		CC_UNLOCK(cc);
1437 	}
1438 
1439 	for (i = 0, tcum = 0; i < 64 && tcum < count / 2; i++)
1440 		tcum += ct[i];
1441 	medt = (i >= 2) ? (((sbintime_t)1) << (i - 2)) : 0;
1442 	for (i = 0, pcum = 0; i < 64 && pcum < count / 2; i++)
1443 		pcum += cpr[i];
1444 	medpr = (i >= 2) ? (((sbintime_t)1) << (i - 2)) : 0;
1445 	for (i = 0, c = 0; i < 32 && c < count / 2; i++)
1446 		c += ccpbk[i];
1447 	medc = (i >= 2) ? (1 << (i - 2)) : 0;
1448 
1449 	printf("Scheduled callouts statistic snapshot:\n");
1450 	printf("  Callouts: %6d  Buckets: %6d*%-3d  Bucket size: 0.%06ds\n",
1451 	    count, callwheelsize, mp_ncpus, 1000000 >> CC_HASH_SHIFT);
1452 	printf("  C/Bk: med %5d         avg %6d.%06jd  max %6d\n",
1453 	    medc,
1454 	    count / callwheelsize / mp_ncpus,
1455 	    (uint64_t)count * 1000000 / callwheelsize / mp_ncpus % 1000000,
1456 	    maxc);
1457 	printf("  Time: med %5jd.%06jds avg %6jd.%06jds max %6jd.%06jds\n",
1458 	    medt / SBT_1S, (medt & 0xffffffff) * 1000000 >> 32,
1459 	    (st / count) / 1000000, (st / count) % 1000000,
1460 	    maxt / SBT_1S, (maxt & 0xffffffff) * 1000000 >> 32);
1461 	printf("  Prec: med %5jd.%06jds avg %6jd.%06jds max %6jd.%06jds\n",
1462 	    medpr / SBT_1S, (medpr & 0xffffffff) * 1000000 >> 32,
1463 	    (spr / count) / 1000000, (spr / count) % 1000000,
1464 	    maxpr / SBT_1S, (maxpr & 0xffffffff) * 1000000 >> 32);
1465 	printf("  Distribution:       \tbuckets\t   time\t   tcum\t"
1466 	    "   prec\t   pcum\n");
1467 	for (i = 0, tcum = pcum = 0; i < 64; i++) {
1468 		if (ct[i] == 0 && cpr[i] == 0)
1469 			continue;
1470 		t = (i != 0) ? (((sbintime_t)1) << (i - 1)) : 0;
1471 		tcum += ct[i];
1472 		pcum += cpr[i];
1473 		printf("  %10jd.%06jds\t 2**%d\t%7d\t%7d\t%7d\t%7d\n",
1474 		    t / SBT_1S, (t & 0xffffffff) * 1000000 >> 32,
1475 		    i - 1 - (32 - CC_HASH_SHIFT),
1476 		    ct[i], tcum, cpr[i], pcum);
1477 	}
1478 	return (error);
1479 }
1480 SYSCTL_PROC(_kern, OID_AUTO, callout_stat,
1481     CTLTYPE_INT | CTLFLAG_RW | CTLFLAG_MPSAFE,
1482     0, 0, sysctl_kern_callout_stat, "I",
1483     "Dump immediate statistic snapshot of the scheduled callouts");
1484 
1485 #ifdef DDB
1486 static void
1487 _show_callout(struct callout *c)
1488 {
1489 
1490 	db_printf("callout %p\n", c);
1491 #define	C_DB_PRINTF(f, e)	db_printf("   %s = " f "\n", #e, c->e);
1492 	db_printf("   &c_links = %p\n", &(c->c_links));
1493 	C_DB_PRINTF("%" PRId64,	c_time);
1494 	C_DB_PRINTF("%" PRId64,	c_precision);
1495 	C_DB_PRINTF("%p",	c_arg);
1496 	C_DB_PRINTF("%p",	c_func);
1497 	C_DB_PRINTF("%p",	c_lock);
1498 	C_DB_PRINTF("%#x",	c_flags);
1499 	C_DB_PRINTF("%#x",	c_iflags);
1500 	C_DB_PRINTF("%d",	c_cpu);
1501 #undef	C_DB_PRINTF
1502 }
1503 
1504 DB_SHOW_COMMAND(callout, db_show_callout)
1505 {
1506 
1507 	if (!have_addr) {
1508 		db_printf("usage: show callout <struct callout *>\n");
1509 		return;
1510 	}
1511 
1512 	_show_callout((struct callout *)addr);
1513 }
1514 
1515 static void
1516 _show_last_callout(int cpu, int direct, const char *dirstr)
1517 {
1518 	struct callout_cpu *cc;
1519 	void *func, *arg;
1520 
1521 	cc = CC_CPU(cpu);
1522 	func = cc_exec_last_func(cc, direct);
1523 	arg = cc_exec_last_arg(cc, direct);
1524 	db_printf("cpu %d last%s callout function: %p ", cpu, dirstr, func);
1525 	db_printsym((db_expr_t)func, DB_STGY_ANY);
1526 	db_printf("\ncpu %d last%s callout argument: %p\n", cpu, dirstr, arg);
1527 }
1528 
1529 DB_SHOW_COMMAND_FLAGS(callout_last, db_show_callout_last, DB_CMD_MEMSAFE)
1530 {
1531 	int cpu, last;
1532 
1533 	if (have_addr) {
1534 		if (addr < 0 || addr > mp_maxid || CPU_ABSENT(addr)) {
1535 			db_printf("no such cpu: %d\n", (int)addr);
1536 			return;
1537 		}
1538 		cpu = last = addr;
1539 	} else {
1540 		cpu = 0;
1541 		last = mp_maxid;
1542 	}
1543 
1544 	while (cpu <= last) {
1545 		if (!CPU_ABSENT(cpu)) {
1546 			_show_last_callout(cpu, 0, "");
1547 			_show_last_callout(cpu, 1, " direct");
1548 		}
1549 		cpu++;
1550 	}
1551 }
1552 #endif /* DDB */
1553