xref: /freebsd/sys/kern/kern_timeout.c (revision 8ddb146abcdf061be9f2c0db7e391697dafad85c)
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 #include "opt_rss.h"
45 
46 #include <sys/param.h>
47 #include <sys/systm.h>
48 #include <sys/bus.h>
49 #include <sys/callout.h>
50 #include <sys/domainset.h>
51 #include <sys/file.h>
52 #include <sys/interrupt.h>
53 #include <sys/kernel.h>
54 #include <sys/ktr.h>
55 #include <sys/kthread.h>
56 #include <sys/lock.h>
57 #include <sys/malloc.h>
58 #include <sys/mutex.h>
59 #include <sys/proc.h>
60 #include <sys/random.h>
61 #include <sys/sched.h>
62 #include <sys/sdt.h>
63 #include <sys/sleepqueue.h>
64 #include <sys/sysctl.h>
65 #include <sys/smp.h>
66 #include <sys/unistd.h>
67 
68 #ifdef DDB
69 #include <ddb/ddb.h>
70 #include <ddb/db_sym.h>
71 #include <machine/_inttypes.h>
72 #endif
73 
74 #ifdef SMP
75 #include <machine/cpu.h>
76 #endif
77 
78 DPCPU_DECLARE(sbintime_t, hardclocktime);
79 
80 SDT_PROVIDER_DEFINE(callout_execute);
81 SDT_PROBE_DEFINE1(callout_execute, , , callout__start, "struct callout *");
82 SDT_PROBE_DEFINE1(callout_execute, , , callout__end, "struct callout *");
83 
84 static void	softclock_thread(void *arg);
85 
86 #ifdef CALLOUT_PROFILING
87 static int avg_depth;
88 SYSCTL_INT(_debug, OID_AUTO, to_avg_depth, CTLFLAG_RD, &avg_depth, 0,
89     "Average number of items examined per softclock call. Units = 1/1000");
90 static int avg_gcalls;
91 SYSCTL_INT(_debug, OID_AUTO, to_avg_gcalls, CTLFLAG_RD, &avg_gcalls, 0,
92     "Average number of Giant callouts made per softclock call. Units = 1/1000");
93 static int avg_lockcalls;
94 SYSCTL_INT(_debug, OID_AUTO, to_avg_lockcalls, CTLFLAG_RD, &avg_lockcalls, 0,
95     "Average number of lock callouts made per softclock call. Units = 1/1000");
96 static int avg_mpcalls;
97 SYSCTL_INT(_debug, OID_AUTO, to_avg_mpcalls, CTLFLAG_RD, &avg_mpcalls, 0,
98     "Average number of MP callouts made per softclock call. Units = 1/1000");
99 static int avg_depth_dir;
100 SYSCTL_INT(_debug, OID_AUTO, to_avg_depth_dir, CTLFLAG_RD, &avg_depth_dir, 0,
101     "Average number of direct callouts examined per callout_process call. "
102     "Units = 1/1000");
103 static int avg_lockcalls_dir;
104 SYSCTL_INT(_debug, OID_AUTO, to_avg_lockcalls_dir, CTLFLAG_RD,
105     &avg_lockcalls_dir, 0, "Average number of lock direct callouts made per "
106     "callout_process call. Units = 1/1000");
107 static int avg_mpcalls_dir;
108 SYSCTL_INT(_debug, OID_AUTO, to_avg_mpcalls_dir, CTLFLAG_RD, &avg_mpcalls_dir,
109     0, "Average number of MP direct callouts made per callout_process call. "
110     "Units = 1/1000");
111 #endif
112 
113 static int ncallout;
114 SYSCTL_INT(_kern, OID_AUTO, ncallout, CTLFLAG_RDTUN | CTLFLAG_NOFETCH, &ncallout, 0,
115     "Number of entries in callwheel and size of timeout() preallocation");
116 
117 #ifdef	RSS
118 static int pin_default_swi = 1;
119 static int pin_pcpu_swi = 1;
120 #else
121 static int pin_default_swi = 0;
122 static int pin_pcpu_swi = 0;
123 #endif
124 
125 SYSCTL_INT(_kern, OID_AUTO, pin_default_swi, CTLFLAG_RDTUN | CTLFLAG_NOFETCH, &pin_default_swi,
126     0, "Pin the default (non-per-cpu) swi (shared with PCPU 0 swi)");
127 SYSCTL_INT(_kern, OID_AUTO, pin_pcpu_swi, CTLFLAG_RDTUN | CTLFLAG_NOFETCH, &pin_pcpu_swi,
128     0, "Pin the per-CPU swis (except PCPU 0, which is also default)");
129 
130 /*
131  * TODO:
132  *	allocate more timeout table slots when table overflows.
133  */
134 static u_int __read_mostly callwheelsize;
135 static u_int __read_mostly callwheelmask;
136 
137 /*
138  * The callout cpu exec entities represent informations necessary for
139  * describing the state of callouts currently running on the CPU and the ones
140  * necessary for migrating callouts to the new callout cpu. In particular,
141  * the first entry of the array cc_exec_entity holds informations for callout
142  * running in SWI thread context, while the second one holds informations
143  * for callout running directly from hardware interrupt context.
144  * The cached informations are very important for deferring migration when
145  * the migrating callout is already running.
146  */
147 struct cc_exec {
148 	struct callout		*cc_curr;
149 	callout_func_t		*cc_drain;
150 	void			*cc_last_func;
151 	void			*cc_last_arg;
152 #ifdef SMP
153 	callout_func_t		*ce_migration_func;
154 	void			*ce_migration_arg;
155 	sbintime_t		ce_migration_time;
156 	sbintime_t		ce_migration_prec;
157 	int			ce_migration_cpu;
158 #endif
159 	bool			cc_cancel;
160 	bool			cc_waiting;
161 };
162 
163 /*
164  * There is one struct callout_cpu per cpu, holding all relevant
165  * state for the callout processing thread on the individual CPU.
166  */
167 struct callout_cpu {
168 	struct mtx_padalign	cc_lock;
169 	struct cc_exec 		cc_exec_entity[2];
170 	struct callout		*cc_next;
171 	struct callout_list	*cc_callwheel;
172 	struct callout_tailq	cc_expireq;
173 	sbintime_t		cc_firstevent;
174 	sbintime_t		cc_lastscan;
175 	struct thread		*cc_thread;
176 	u_int			cc_bucket;
177 #ifdef KTR
178 	char			cc_ktr_event_name[20];
179 #endif
180 };
181 
182 #define	callout_migrating(c)	((c)->c_iflags & CALLOUT_DFRMIGRATION)
183 
184 #define	cc_exec_curr(cc, dir)		cc->cc_exec_entity[dir].cc_curr
185 #define	cc_exec_last_func(cc, dir)	cc->cc_exec_entity[dir].cc_last_func
186 #define	cc_exec_last_arg(cc, dir)	cc->cc_exec_entity[dir].cc_last_arg
187 #define	cc_exec_drain(cc, dir)		cc->cc_exec_entity[dir].cc_drain
188 #define	cc_exec_next(cc)		cc->cc_next
189 #define	cc_exec_cancel(cc, dir)		cc->cc_exec_entity[dir].cc_cancel
190 #define	cc_exec_waiting(cc, dir)	cc->cc_exec_entity[dir].cc_waiting
191 #ifdef SMP
192 #define	cc_migration_func(cc, dir)	cc->cc_exec_entity[dir].ce_migration_func
193 #define	cc_migration_arg(cc, dir)	cc->cc_exec_entity[dir].ce_migration_arg
194 #define	cc_migration_cpu(cc, dir)	cc->cc_exec_entity[dir].ce_migration_cpu
195 #define	cc_migration_time(cc, dir)	cc->cc_exec_entity[dir].ce_migration_time
196 #define	cc_migration_prec(cc, dir)	cc->cc_exec_entity[dir].ce_migration_prec
197 
198 static struct callout_cpu cc_cpu[MAXCPU];
199 #define	CPUBLOCK	MAXCPU
200 #define	CC_CPU(cpu)	(&cc_cpu[(cpu)])
201 #define	CC_SELF()	CC_CPU(PCPU_GET(cpuid))
202 #else
203 static struct callout_cpu cc_cpu;
204 #define	CC_CPU(cpu)	(&cc_cpu)
205 #define	CC_SELF()	(&cc_cpu)
206 #endif
207 #define	CC_LOCK(cc)	mtx_lock_spin(&(cc)->cc_lock)
208 #define	CC_UNLOCK(cc)	mtx_unlock_spin(&(cc)->cc_lock)
209 #define	CC_LOCK_ASSERT(cc)	mtx_assert(&(cc)->cc_lock, MA_OWNED)
210 
211 static int __read_mostly cc_default_cpu;
212 
213 static void	callout_cpu_init(struct callout_cpu *cc, int cpu);
214 static void	softclock_call_cc(struct callout *c, struct callout_cpu *cc,
215 #ifdef CALLOUT_PROFILING
216 		    int *mpcalls, int *lockcalls, int *gcalls,
217 #endif
218 		    int direct);
219 
220 static MALLOC_DEFINE(M_CALLOUT, "callout", "Callout datastructures");
221 
222 /**
223  * Locked by cc_lock:
224  *   cc_curr         - If a callout is in progress, it is cc_curr.
225  *                     If cc_curr is non-NULL, threads waiting in
226  *                     callout_drain() will be woken up as soon as the
227  *                     relevant callout completes.
228  *   cc_cancel       - Changing to 1 with both callout_lock and cc_lock held
229  *                     guarantees that the current callout will not run.
230  *                     The softclock_call_cc() function sets this to 0 before it
231  *                     drops callout_lock to acquire c_lock, and it calls
232  *                     the handler only if curr_cancelled is still 0 after
233  *                     cc_lock is successfully acquired.
234  *   cc_waiting      - If a thread is waiting in callout_drain(), then
235  *                     callout_wait is nonzero.  Set only when
236  *                     cc_curr is non-NULL.
237  */
238 
239 /*
240  * Resets the execution entity tied to a specific callout cpu.
241  */
242 static void
243 cc_cce_cleanup(struct callout_cpu *cc, int direct)
244 {
245 
246 	cc_exec_curr(cc, direct) = NULL;
247 	cc_exec_cancel(cc, direct) = false;
248 	cc_exec_waiting(cc, direct) = false;
249 #ifdef SMP
250 	cc_migration_cpu(cc, direct) = CPUBLOCK;
251 	cc_migration_time(cc, direct) = 0;
252 	cc_migration_prec(cc, direct) = 0;
253 	cc_migration_func(cc, direct) = NULL;
254 	cc_migration_arg(cc, direct) = NULL;
255 #endif
256 }
257 
258 /*
259  * Checks if migration is requested by a specific callout cpu.
260  */
261 static int
262 cc_cce_migrating(struct callout_cpu *cc, int direct)
263 {
264 
265 #ifdef SMP
266 	return (cc_migration_cpu(cc, direct) != CPUBLOCK);
267 #else
268 	return (0);
269 #endif
270 }
271 
272 /*
273  * Kernel low level callwheel initialization
274  * called on the BSP during kernel startup.
275  */
276 static void
277 callout_callwheel_init(void *dummy)
278 {
279 	struct callout_cpu *cc;
280 	int cpu;
281 
282 	/*
283 	 * Calculate the size of the callout wheel and the preallocated
284 	 * timeout() structures.
285 	 * XXX: Clip callout to result of previous function of maxusers
286 	 * maximum 384.  This is still huge, but acceptable.
287 	 */
288 	ncallout = imin(16 + maxproc + maxfiles, 18508);
289 	TUNABLE_INT_FETCH("kern.ncallout", &ncallout);
290 
291 	/*
292 	 * Calculate callout wheel size, should be next power of two higher
293 	 * than 'ncallout'.
294 	 */
295 	callwheelsize = 1 << fls(ncallout);
296 	callwheelmask = callwheelsize - 1;
297 
298 	/*
299 	 * Fetch whether we're pinning the swi's or not.
300 	 */
301 	TUNABLE_INT_FETCH("kern.pin_default_swi", &pin_default_swi);
302 	TUNABLE_INT_FETCH("kern.pin_pcpu_swi", &pin_pcpu_swi);
303 
304 	/*
305 	 * Initialize callout wheels.  The software interrupt threads
306 	 * are created later.
307 	 */
308 	cc_default_cpu = PCPU_GET(cpuid);
309 	CPU_FOREACH(cpu) {
310 		cc = CC_CPU(cpu);
311 		callout_cpu_init(cc, cpu);
312 	}
313 }
314 SYSINIT(callwheel_init, SI_SUB_CPU, SI_ORDER_ANY, callout_callwheel_init, NULL);
315 
316 /*
317  * Initialize the per-cpu callout structures.
318  */
319 static void
320 callout_cpu_init(struct callout_cpu *cc, int cpu)
321 {
322 	int i;
323 
324 	mtx_init(&cc->cc_lock, "callout", NULL, MTX_SPIN);
325 	cc->cc_callwheel = malloc_domainset(sizeof(struct callout_list) *
326 	    callwheelsize, M_CALLOUT,
327 	    DOMAINSET_PREF(pcpu_find(cpu)->pc_domain), M_WAITOK);
328 	for (i = 0; i < callwheelsize; i++)
329 		LIST_INIT(&cc->cc_callwheel[i]);
330 	TAILQ_INIT(&cc->cc_expireq);
331 	cc->cc_firstevent = SBT_MAX;
332 	for (i = 0; i < 2; i++)
333 		cc_cce_cleanup(cc, i);
334 #ifdef KTR
335 	snprintf(cc->cc_ktr_event_name, sizeof(cc->cc_ktr_event_name),
336 	    "callwheel cpu %d", cpu);
337 #endif
338 }
339 
340 #ifdef SMP
341 /*
342  * Switches the cpu tied to a specific callout.
343  * The function expects a locked incoming callout cpu and returns with
344  * locked outcoming callout cpu.
345  */
346 static struct callout_cpu *
347 callout_cpu_switch(struct callout *c, struct callout_cpu *cc, int new_cpu)
348 {
349 	struct callout_cpu *new_cc;
350 
351 	MPASS(c != NULL && cc != NULL);
352 	CC_LOCK_ASSERT(cc);
353 
354 	/*
355 	 * Avoid interrupts and preemption firing after the callout cpu
356 	 * is blocked in order to avoid deadlocks as the new thread
357 	 * may be willing to acquire the callout cpu lock.
358 	 */
359 	c->c_cpu = CPUBLOCK;
360 	spinlock_enter();
361 	CC_UNLOCK(cc);
362 	new_cc = CC_CPU(new_cpu);
363 	CC_LOCK(new_cc);
364 	spinlock_exit();
365 	c->c_cpu = new_cpu;
366 	return (new_cc);
367 }
368 #endif
369 
370 /*
371  * Start softclock threads.
372  */
373 static void
374 start_softclock(void *dummy)
375 {
376 	struct proc *p;
377 	struct thread *td;
378 	struct callout_cpu *cc;
379 	int cpu, error;
380 	bool pin_swi;
381 
382 	p = NULL;
383 	CPU_FOREACH(cpu) {
384 		cc = CC_CPU(cpu);
385 		error = kproc_kthread_add(softclock_thread, cc, &p, &td,
386 		    RFSTOPPED, 0, "clock", "clock (%d)", cpu);
387 		if (error != 0)
388 			panic("failed to create softclock thread for cpu %d: %d",
389 			    cpu, error);
390 		CC_LOCK(cc);
391 		cc->cc_thread = td;
392 		thread_lock(td);
393 		sched_class(td, PRI_ITHD);
394 		sched_ithread_prio(td, PI_SOFTCLOCK);
395 		TD_SET_IWAIT(td);
396 		thread_lock_set(td, (struct mtx *)&cc->cc_lock);
397 		thread_unlock(td);
398 		if (cpu == cc_default_cpu)
399 			pin_swi = pin_default_swi;
400 		else
401 			pin_swi = pin_pcpu_swi;
402 		if (pin_swi) {
403 			error = cpuset_setithread(td->td_tid, cpu);
404 			if (error != 0)
405 				printf("%s: %s clock couldn't be pinned to cpu %d: %d\n",
406 				    __func__, cpu == cc_default_cpu ?
407 				    "default" : "per-cpu", cpu, error);
408 		}
409 	}
410 }
411 SYSINIT(start_softclock, SI_SUB_SOFTINTR, SI_ORDER_FIRST, start_softclock, NULL);
412 
413 #define	CC_HASH_SHIFT	8
414 
415 static inline u_int
416 callout_hash(sbintime_t sbt)
417 {
418 
419 	return (sbt >> (32 - CC_HASH_SHIFT));
420 }
421 
422 static inline u_int
423 callout_get_bucket(sbintime_t sbt)
424 {
425 
426 	return (callout_hash(sbt) & callwheelmask);
427 }
428 
429 void
430 callout_process(sbintime_t now)
431 {
432 	struct callout_entropy {
433 		struct callout_cpu *cc;
434 		struct thread *td;
435 		sbintime_t now;
436 	} entropy;
437 	struct callout *c, *next;
438 	struct callout_cpu *cc;
439 	struct callout_list *sc;
440 	struct thread *td;
441 	sbintime_t first, last, lookahead, max, tmp_max;
442 	u_int firstb, lastb, nowb;
443 #ifdef CALLOUT_PROFILING
444 	int depth_dir = 0, mpcalls_dir = 0, lockcalls_dir = 0;
445 #endif
446 
447 	cc = CC_SELF();
448 	mtx_lock_spin_flags(&cc->cc_lock, MTX_QUIET);
449 
450 	/* Compute the buckets of the last scan and present times. */
451 	firstb = callout_hash(cc->cc_lastscan);
452 	cc->cc_lastscan = now;
453 	nowb = callout_hash(now);
454 
455 	/* Compute the last bucket and minimum time of the bucket after it. */
456 	if (nowb == firstb)
457 		lookahead = (SBT_1S / 16);
458 	else if (nowb - firstb == 1)
459 		lookahead = (SBT_1S / 8);
460 	else
461 		lookahead = SBT_1S;
462 	first = last = now;
463 	first += (lookahead / 2);
464 	last += lookahead;
465 	last &= (0xffffffffffffffffLLU << (32 - CC_HASH_SHIFT));
466 	lastb = callout_hash(last) - 1;
467 	max = last;
468 
469 	/*
470 	 * Check if we wrapped around the entire wheel from the last scan.
471 	 * In case, we need to scan entirely the wheel for pending callouts.
472 	 */
473 	if (lastb - firstb >= callwheelsize) {
474 		lastb = firstb + callwheelsize - 1;
475 		if (nowb - firstb >= callwheelsize)
476 			nowb = lastb;
477 	}
478 
479 	/* Iterate callwheel from firstb to nowb and then up to lastb. */
480 	do {
481 		sc = &cc->cc_callwheel[firstb & callwheelmask];
482 		LIST_FOREACH_SAFE(c, sc, c_links.le, next) {
483 			/* Run the callout if present time within allowed. */
484 			if (c->c_time <= now) {
485 				/*
486 				 * Consumer told us the callout may be run
487 				 * directly from hardware interrupt context.
488 				 */
489 				if (c->c_iflags & CALLOUT_DIRECT) {
490 #ifdef CALLOUT_PROFILING
491 					++depth_dir;
492 #endif
493 					cc_exec_next(cc) = next;
494 					cc->cc_bucket = firstb & callwheelmask;
495 					LIST_REMOVE(c, c_links.le);
496 					softclock_call_cc(c, cc,
497 #ifdef CALLOUT_PROFILING
498 					    &mpcalls_dir, &lockcalls_dir, NULL,
499 #endif
500 					    1);
501 					next = cc_exec_next(cc);
502 					cc_exec_next(cc) = NULL;
503 				} else {
504 					LIST_REMOVE(c, c_links.le);
505 					TAILQ_INSERT_TAIL(&cc->cc_expireq,
506 					    c, c_links.tqe);
507 					c->c_iflags |= CALLOUT_PROCESSED;
508 				}
509 			} else if (c->c_time >= max) {
510 				/*
511 				 * Skip events in the distant future.
512 				 */
513 				;
514 			} else if (c->c_time > last) {
515 				/*
516 				 * Event minimal time is bigger than present
517 				 * maximal time, so it cannot be aggregated.
518 				 */
519 				lastb = nowb;
520 			} else {
521 				/*
522 				 * Update first and last time, respecting this
523 				 * event.
524 				 */
525 				if (c->c_time < first)
526 					first = c->c_time;
527 				tmp_max = c->c_time + c->c_precision;
528 				if (tmp_max < last)
529 					last = tmp_max;
530 			}
531 		}
532 		/* Proceed with the next bucket. */
533 		firstb++;
534 		/*
535 		 * Stop if we looked after present time and found
536 		 * some event we can't execute at now.
537 		 * Stop if we looked far enough into the future.
538 		 */
539 	} while (((int)(firstb - lastb)) <= 0);
540 	cc->cc_firstevent = last;
541 	cpu_new_callout(curcpu, last, first);
542 
543 #ifdef CALLOUT_PROFILING
544 	avg_depth_dir += (depth_dir * 1000 - avg_depth_dir) >> 8;
545 	avg_mpcalls_dir += (mpcalls_dir * 1000 - avg_mpcalls_dir) >> 8;
546 	avg_lockcalls_dir += (lockcalls_dir * 1000 - avg_lockcalls_dir) >> 8;
547 #endif
548 	if (!TAILQ_EMPTY(&cc->cc_expireq)) {
549 		entropy.cc = cc;
550 		entropy.td = curthread;
551 		entropy.now = now;
552 		random_harvest_queue(&entropy, sizeof(entropy), RANDOM_CALLOUT);
553 
554 		td = cc->cc_thread;
555 		if (TD_AWAITING_INTR(td)) {
556 			thread_lock_block_wait(td);
557 			THREAD_LOCK_ASSERT(td, MA_OWNED);
558 			TD_CLR_IWAIT(td);
559 			sched_wakeup(td, SRQ_INTR);
560 		} else
561 			mtx_unlock_spin_flags(&cc->cc_lock, MTX_QUIET);
562 	} else
563 		mtx_unlock_spin_flags(&cc->cc_lock, MTX_QUIET);
564 }
565 
566 static struct callout_cpu *
567 callout_lock(struct callout *c)
568 {
569 	struct callout_cpu *cc;
570 	int cpu;
571 
572 	for (;;) {
573 		cpu = c->c_cpu;
574 #ifdef SMP
575 		if (cpu == CPUBLOCK) {
576 			while (c->c_cpu == CPUBLOCK)
577 				cpu_spinwait();
578 			continue;
579 		}
580 #endif
581 		cc = CC_CPU(cpu);
582 		CC_LOCK(cc);
583 		if (cpu == c->c_cpu)
584 			break;
585 		CC_UNLOCK(cc);
586 	}
587 	return (cc);
588 }
589 
590 static void
591 callout_cc_add(struct callout *c, struct callout_cpu *cc,
592     sbintime_t sbt, sbintime_t precision, void (*func)(void *),
593     void *arg, int flags)
594 {
595 	int bucket;
596 
597 	CC_LOCK_ASSERT(cc);
598 	if (sbt < cc->cc_lastscan)
599 		sbt = cc->cc_lastscan;
600 	c->c_arg = arg;
601 	c->c_iflags |= CALLOUT_PENDING;
602 	c->c_iflags &= ~CALLOUT_PROCESSED;
603 	c->c_flags |= CALLOUT_ACTIVE;
604 	if (flags & C_DIRECT_EXEC)
605 		c->c_iflags |= CALLOUT_DIRECT;
606 	c->c_func = func;
607 	c->c_time = sbt;
608 	c->c_precision = precision;
609 	bucket = callout_get_bucket(c->c_time);
610 	CTR3(KTR_CALLOUT, "precision set for %p: %d.%08x",
611 	    c, (int)(c->c_precision >> 32),
612 	    (u_int)(c->c_precision & 0xffffffff));
613 	LIST_INSERT_HEAD(&cc->cc_callwheel[bucket], c, c_links.le);
614 	if (cc->cc_bucket == bucket)
615 		cc_exec_next(cc) = c;
616 
617 	/*
618 	 * Inform the eventtimers(4) subsystem there's a new callout
619 	 * that has been inserted, but only if really required.
620 	 */
621 	if (SBT_MAX - c->c_time < c->c_precision)
622 		c->c_precision = SBT_MAX - c->c_time;
623 	sbt = c->c_time + c->c_precision;
624 	if (sbt < cc->cc_firstevent) {
625 		cc->cc_firstevent = sbt;
626 		cpu_new_callout(c->c_cpu, sbt, c->c_time);
627 	}
628 }
629 
630 static void
631 softclock_call_cc(struct callout *c, struct callout_cpu *cc,
632 #ifdef CALLOUT_PROFILING
633     int *mpcalls, int *lockcalls, int *gcalls,
634 #endif
635     int direct)
636 {
637 	struct rm_priotracker tracker;
638 	callout_func_t *c_func, *drain;
639 	void *c_arg;
640 	struct lock_class *class;
641 	struct lock_object *c_lock;
642 	uintptr_t lock_status;
643 	int c_iflags;
644 #ifdef SMP
645 	struct callout_cpu *new_cc;
646 	callout_func_t *new_func;
647 	void *new_arg;
648 	int flags, new_cpu;
649 	sbintime_t new_prec, new_time;
650 #endif
651 #if defined(DIAGNOSTIC) || defined(CALLOUT_PROFILING)
652 	sbintime_t sbt1, sbt2;
653 	struct timespec ts2;
654 	static sbintime_t maxdt = 2 * SBT_1MS;	/* 2 msec */
655 	static callout_func_t *lastfunc;
656 #endif
657 
658 	KASSERT((c->c_iflags & CALLOUT_PENDING) == CALLOUT_PENDING,
659 	    ("softclock_call_cc: pend %p %x", c, c->c_iflags));
660 	KASSERT((c->c_flags & CALLOUT_ACTIVE) == CALLOUT_ACTIVE,
661 	    ("softclock_call_cc: act %p %x", c, c->c_flags));
662 	class = (c->c_lock != NULL) ? LOCK_CLASS(c->c_lock) : NULL;
663 	lock_status = 0;
664 	if (c->c_iflags & CALLOUT_SHAREDLOCK) {
665 		if (class == &lock_class_rm)
666 			lock_status = (uintptr_t)&tracker;
667 		else
668 			lock_status = 1;
669 	}
670 	c_lock = c->c_lock;
671 	c_func = c->c_func;
672 	c_arg = c->c_arg;
673 	c_iflags = c->c_iflags;
674 	c->c_iflags &= ~CALLOUT_PENDING;
675 
676 	cc_exec_curr(cc, direct) = c;
677 	cc_exec_last_func(cc, direct) = c_func;
678 	cc_exec_last_arg(cc, direct) = c_arg;
679 	cc_exec_cancel(cc, direct) = false;
680 	cc_exec_drain(cc, direct) = NULL;
681 	CC_UNLOCK(cc);
682 	if (c_lock != NULL) {
683 		class->lc_lock(c_lock, lock_status);
684 		/*
685 		 * The callout may have been cancelled
686 		 * while we switched locks.
687 		 */
688 		if (cc_exec_cancel(cc, direct)) {
689 			class->lc_unlock(c_lock);
690 			goto skip;
691 		}
692 		/* The callout cannot be stopped now. */
693 		cc_exec_cancel(cc, direct) = true;
694 		if (c_lock == &Giant.lock_object) {
695 #ifdef CALLOUT_PROFILING
696 			(*gcalls)++;
697 #endif
698 			CTR3(KTR_CALLOUT, "callout giant %p func %p arg %p",
699 			    c, c_func, c_arg);
700 		} else {
701 #ifdef CALLOUT_PROFILING
702 			(*lockcalls)++;
703 #endif
704 			CTR3(KTR_CALLOUT, "callout lock %p func %p arg %p",
705 			    c, c_func, c_arg);
706 		}
707 	} else {
708 #ifdef CALLOUT_PROFILING
709 		(*mpcalls)++;
710 #endif
711 		CTR3(KTR_CALLOUT, "callout %p func %p arg %p",
712 		    c, c_func, c_arg);
713 	}
714 	KTR_STATE3(KTR_SCHED, "callout", cc->cc_ktr_event_name, "running",
715 	    "func:%p", c_func, "arg:%p", c_arg, "direct:%d", direct);
716 #if defined(DIAGNOSTIC) || defined(CALLOUT_PROFILING)
717 	sbt1 = sbinuptime();
718 #endif
719 	THREAD_NO_SLEEPING();
720 	SDT_PROBE1(callout_execute, , , callout__start, c);
721 	c_func(c_arg);
722 	SDT_PROBE1(callout_execute, , , callout__end, c);
723 	THREAD_SLEEPING_OK();
724 #if defined(DIAGNOSTIC) || defined(CALLOUT_PROFILING)
725 	sbt2 = sbinuptime();
726 	sbt2 -= sbt1;
727 	if (sbt2 > maxdt) {
728 		if (lastfunc != c_func || sbt2 > maxdt * 2) {
729 			ts2 = sbttots(sbt2);
730 			printf(
731 		"Expensive timeout(9) function: %p(%p) %jd.%09ld s\n",
732 			    c_func, c_arg, (intmax_t)ts2.tv_sec, ts2.tv_nsec);
733 		}
734 		maxdt = sbt2;
735 		lastfunc = c_func;
736 	}
737 #endif
738 	KTR_STATE0(KTR_SCHED, "callout", cc->cc_ktr_event_name, "idle");
739 	CTR1(KTR_CALLOUT, "callout %p finished", c);
740 	if ((c_iflags & CALLOUT_RETURNUNLOCKED) == 0)
741 		class->lc_unlock(c_lock);
742 skip:
743 	CC_LOCK(cc);
744 	KASSERT(cc_exec_curr(cc, direct) == c, ("mishandled cc_curr"));
745 	cc_exec_curr(cc, direct) = NULL;
746 	if (cc_exec_drain(cc, direct)) {
747 		drain = cc_exec_drain(cc, direct);
748 		cc_exec_drain(cc, direct) = NULL;
749 		CC_UNLOCK(cc);
750 		drain(c_arg);
751 		CC_LOCK(cc);
752 	}
753 	if (cc_exec_waiting(cc, direct)) {
754 		/*
755 		 * There is someone waiting for the
756 		 * callout to complete.
757 		 * If the callout was scheduled for
758 		 * migration just cancel it.
759 		 */
760 		if (cc_cce_migrating(cc, direct)) {
761 			cc_cce_cleanup(cc, direct);
762 
763 			/*
764 			 * It should be assert here that the callout is not
765 			 * destroyed but that is not easy.
766 			 */
767 			c->c_iflags &= ~CALLOUT_DFRMIGRATION;
768 		}
769 		cc_exec_waiting(cc, direct) = false;
770 		CC_UNLOCK(cc);
771 		wakeup(&cc_exec_waiting(cc, direct));
772 		CC_LOCK(cc);
773 	} else if (cc_cce_migrating(cc, direct)) {
774 #ifdef SMP
775 		/*
776 		 * If the callout was scheduled for
777 		 * migration just perform it now.
778 		 */
779 		new_cpu = cc_migration_cpu(cc, direct);
780 		new_time = cc_migration_time(cc, direct);
781 		new_prec = cc_migration_prec(cc, direct);
782 		new_func = cc_migration_func(cc, direct);
783 		new_arg = cc_migration_arg(cc, direct);
784 		cc_cce_cleanup(cc, direct);
785 
786 		/*
787 		 * It should be assert here that the callout is not destroyed
788 		 * but that is not easy.
789 		 *
790 		 * As first thing, handle deferred callout stops.
791 		 */
792 		if (!callout_migrating(c)) {
793 			CTR3(KTR_CALLOUT,
794 			     "deferred cancelled %p func %p arg %p",
795 			     c, new_func, new_arg);
796 			return;
797 		}
798 		c->c_iflags &= ~CALLOUT_DFRMIGRATION;
799 
800 		new_cc = callout_cpu_switch(c, cc, new_cpu);
801 		flags = (direct) ? C_DIRECT_EXEC : 0;
802 		callout_cc_add(c, new_cc, new_time, new_prec, new_func,
803 		    new_arg, flags);
804 		CC_UNLOCK(new_cc);
805 		CC_LOCK(cc);
806 #else
807 		panic("migration should not happen");
808 #endif
809 	}
810 }
811 
812 /*
813  * The callout mechanism is based on the work of Adam M. Costello and
814  * George Varghese, published in a technical report entitled "Redesigning
815  * the BSD Callout and Timer Facilities" and modified slightly for inclusion
816  * in FreeBSD by Justin T. Gibbs.  The original work on the data structures
817  * used in this implementation was published by G. Varghese and T. Lauck in
818  * the paper "Hashed and Hierarchical Timing Wheels: Data Structures for
819  * the Efficient Implementation of a Timer Facility" in the Proceedings of
820  * the 11th ACM Annual Symposium on Operating Systems Principles,
821  * Austin, Texas Nov 1987.
822  */
823 
824 /*
825  * Software (low priority) clock interrupt thread handler.
826  * Run periodic events from timeout queue.
827  */
828 static void
829 softclock_thread(void *arg)
830 {
831 	struct thread *td = curthread;
832 	struct callout_cpu *cc;
833 	struct callout *c;
834 #ifdef CALLOUT_PROFILING
835 	int depth, gcalls, lockcalls, mpcalls;
836 #endif
837 
838 	cc = (struct callout_cpu *)arg;
839 	CC_LOCK(cc);
840 	for (;;) {
841 		while (TAILQ_EMPTY(&cc->cc_expireq)) {
842 			/*
843 			 * Use CC_LOCK(cc) as the thread_lock while
844 			 * idle.
845 			 */
846 			thread_lock(td);
847 			thread_lock_set(td, (struct mtx *)&cc->cc_lock);
848 			TD_SET_IWAIT(td);
849 			mi_switch(SW_VOL | SWT_IWAIT);
850 
851 			/* mi_switch() drops thread_lock(). */
852 			CC_LOCK(cc);
853 		}
854 
855 #ifdef CALLOUT_PROFILING
856 		depth = gcalls = lockcalls = mpcalls = 0;
857 #endif
858 		while ((c = TAILQ_FIRST(&cc->cc_expireq)) != NULL) {
859 			TAILQ_REMOVE(&cc->cc_expireq, c, c_links.tqe);
860 			softclock_call_cc(c, cc,
861 #ifdef CALLOUT_PROFILING
862 			    &mpcalls, &lockcalls, &gcalls,
863 #endif
864 			    0);
865 #ifdef CALLOUT_PROFILING
866 			++depth;
867 #endif
868 		}
869 #ifdef CALLOUT_PROFILING
870 		avg_depth += (depth * 1000 - avg_depth) >> 8;
871 		avg_mpcalls += (mpcalls * 1000 - avg_mpcalls) >> 8;
872 		avg_lockcalls += (lockcalls * 1000 - avg_lockcalls) >> 8;
873 		avg_gcalls += (gcalls * 1000 - avg_gcalls) >> 8;
874 #endif
875 	}
876 }
877 
878 void
879 callout_when(sbintime_t sbt, sbintime_t precision, int flags,
880     sbintime_t *res, sbintime_t *prec_res)
881 {
882 	sbintime_t to_sbt, to_pr;
883 
884 	if ((flags & (C_ABSOLUTE | C_PRECALC)) != 0) {
885 		*res = sbt;
886 		*prec_res = precision;
887 		return;
888 	}
889 	if ((flags & C_HARDCLOCK) != 0 && sbt < tick_sbt)
890 		sbt = tick_sbt;
891 	if ((flags & C_HARDCLOCK) != 0 || sbt >= sbt_tickthreshold) {
892 		/*
893 		 * Obtain the time of the last hardclock() call on
894 		 * this CPU directly from the kern_clocksource.c.
895 		 * This value is per-CPU, but it is equal for all
896 		 * active ones.
897 		 */
898 #ifdef __LP64__
899 		to_sbt = DPCPU_GET(hardclocktime);
900 #else
901 		spinlock_enter();
902 		to_sbt = DPCPU_GET(hardclocktime);
903 		spinlock_exit();
904 #endif
905 		if (cold && to_sbt == 0)
906 			to_sbt = sbinuptime();
907 		if ((flags & C_HARDCLOCK) == 0)
908 			to_sbt += tick_sbt;
909 	} else
910 		to_sbt = sbinuptime();
911 	if (SBT_MAX - to_sbt < sbt)
912 		to_sbt = SBT_MAX;
913 	else
914 		to_sbt += sbt;
915 	*res = to_sbt;
916 	to_pr = ((C_PRELGET(flags) < 0) ? sbt >> tc_precexp :
917 	    sbt >> C_PRELGET(flags));
918 	*prec_res = to_pr > precision ? to_pr : precision;
919 }
920 
921 /*
922  * New interface; clients allocate their own callout structures.
923  *
924  * callout_reset() - establish or change a timeout
925  * callout_stop() - disestablish a timeout
926  * callout_init() - initialize a callout structure so that it can
927  *	safely be passed to callout_reset() and callout_stop()
928  *
929  * <sys/callout.h> defines three convenience macros:
930  *
931  * callout_active() - returns truth if callout has not been stopped,
932  *	drained, or deactivated since the last time the callout was
933  *	reset.
934  * callout_pending() - returns truth if callout is still waiting for timeout
935  * callout_deactivate() - marks the callout as having been serviced
936  */
937 int
938 callout_reset_sbt_on(struct callout *c, sbintime_t sbt, sbintime_t prec,
939     callout_func_t *ftn, void *arg, int cpu, int flags)
940 {
941 	sbintime_t to_sbt, precision;
942 	struct callout_cpu *cc;
943 	int cancelled, direct;
944 
945 	cancelled = 0;
946 	callout_when(sbt, prec, flags, &to_sbt, &precision);
947 
948 	/*
949 	 * This flag used to be added by callout_cc_add, but the
950 	 * first time you call this we could end up with the
951 	 * wrong direct flag if we don't do it before we add.
952 	 */
953 	if (flags & C_DIRECT_EXEC) {
954 		direct = 1;
955 	} else {
956 		direct = 0;
957 	}
958 	KASSERT(!direct || c->c_lock == NULL ||
959 	    (LOCK_CLASS(c->c_lock)->lc_flags & LC_SPINLOCK),
960 	    ("%s: direct callout %p has non-spin lock", __func__, c));
961 
962 	cc = callout_lock(c);
963 	if (cpu == -1)
964 		cpu = c->c_cpu;
965 	KASSERT(cpu >= 0 && cpu <= mp_maxid && !CPU_ABSENT(cpu),
966 	    ("%s: invalid cpu %d", __func__, cpu));
967 
968 	if (cc_exec_curr(cc, direct) == c) {
969 		/*
970 		 * We're being asked to reschedule a callout which is
971 		 * currently in progress.  If there is a lock then we
972 		 * can cancel the callout if it has not really started.
973 		 */
974 		if (c->c_lock != NULL && !cc_exec_cancel(cc, direct))
975 			cancelled = cc_exec_cancel(cc, direct) = true;
976 		if (cc_exec_waiting(cc, direct) || cc_exec_drain(cc, direct)) {
977 			/*
978 			 * Someone has called callout_drain to kill this
979 			 * callout.  Don't reschedule.
980 			 */
981 			CTR4(KTR_CALLOUT, "%s %p func %p arg %p",
982 			    cancelled ? "cancelled" : "failed to cancel",
983 			    c, c->c_func, c->c_arg);
984 			CC_UNLOCK(cc);
985 			return (cancelled);
986 		}
987 #ifdef SMP
988 		if (callout_migrating(c)) {
989 			/*
990 			 * This only occurs when a second callout_reset_sbt_on
991 			 * is made after a previous one moved it into
992 			 * deferred migration (below). Note we do *not* change
993 			 * the prev_cpu even though the previous target may
994 			 * be different.
995 			 */
996 			cc_migration_cpu(cc, direct) = cpu;
997 			cc_migration_time(cc, direct) = to_sbt;
998 			cc_migration_prec(cc, direct) = precision;
999 			cc_migration_func(cc, direct) = ftn;
1000 			cc_migration_arg(cc, direct) = arg;
1001 			cancelled = 1;
1002 			CC_UNLOCK(cc);
1003 			return (cancelled);
1004 		}
1005 #endif
1006 	}
1007 	if (c->c_iflags & CALLOUT_PENDING) {
1008 		if ((c->c_iflags & CALLOUT_PROCESSED) == 0) {
1009 			if (cc_exec_next(cc) == c)
1010 				cc_exec_next(cc) = LIST_NEXT(c, c_links.le);
1011 			LIST_REMOVE(c, c_links.le);
1012 		} else {
1013 			TAILQ_REMOVE(&cc->cc_expireq, c, c_links.tqe);
1014 		}
1015 		cancelled = 1;
1016 		c->c_iflags &= ~ CALLOUT_PENDING;
1017 		c->c_flags &= ~ CALLOUT_ACTIVE;
1018 	}
1019 
1020 #ifdef SMP
1021 	/*
1022 	 * If the callout must migrate try to perform it immediately.
1023 	 * If the callout is currently running, just defer the migration
1024 	 * to a more appropriate moment.
1025 	 */
1026 	if (c->c_cpu != cpu) {
1027 		if (cc_exec_curr(cc, direct) == c) {
1028 			/*
1029 			 * Pending will have been removed since we are
1030 			 * actually executing the callout on another
1031 			 * CPU. That callout should be waiting on the
1032 			 * lock the caller holds. If we set both
1033 			 * active/and/pending after we return and the
1034 			 * lock on the executing callout proceeds, it
1035 			 * will then see pending is true and return.
1036 			 * At the return from the actual callout execution
1037 			 * the migration will occur in softclock_call_cc
1038 			 * and this new callout will be placed on the
1039 			 * new CPU via a call to callout_cpu_switch() which
1040 			 * will get the lock on the right CPU followed
1041 			 * by a call callout_cc_add() which will add it there.
1042 			 * (see above in softclock_call_cc()).
1043 			 */
1044 			cc_migration_cpu(cc, direct) = cpu;
1045 			cc_migration_time(cc, direct) = to_sbt;
1046 			cc_migration_prec(cc, direct) = precision;
1047 			cc_migration_func(cc, direct) = ftn;
1048 			cc_migration_arg(cc, direct) = arg;
1049 			c->c_iflags |= (CALLOUT_DFRMIGRATION | CALLOUT_PENDING);
1050 			c->c_flags |= CALLOUT_ACTIVE;
1051 			CTR6(KTR_CALLOUT,
1052 		    "migration of %p func %p arg %p in %d.%08x to %u deferred",
1053 			    c, c->c_func, c->c_arg, (int)(to_sbt >> 32),
1054 			    (u_int)(to_sbt & 0xffffffff), cpu);
1055 			CC_UNLOCK(cc);
1056 			return (cancelled);
1057 		}
1058 		cc = callout_cpu_switch(c, cc, cpu);
1059 	}
1060 #endif
1061 
1062 	callout_cc_add(c, cc, to_sbt, precision, ftn, arg, flags);
1063 	CTR6(KTR_CALLOUT, "%sscheduled %p func %p arg %p in %d.%08x",
1064 	    cancelled ? "re" : "", c, c->c_func, c->c_arg, (int)(to_sbt >> 32),
1065 	    (u_int)(to_sbt & 0xffffffff));
1066 	CC_UNLOCK(cc);
1067 
1068 	return (cancelled);
1069 }
1070 
1071 /*
1072  * Common idioms that can be optimized in the future.
1073  */
1074 int
1075 callout_schedule_on(struct callout *c, int to_ticks, int cpu)
1076 {
1077 	return callout_reset_on(c, to_ticks, c->c_func, c->c_arg, cpu);
1078 }
1079 
1080 int
1081 callout_schedule(struct callout *c, int to_ticks)
1082 {
1083 	return callout_reset_on(c, to_ticks, c->c_func, c->c_arg, c->c_cpu);
1084 }
1085 
1086 int
1087 _callout_stop_safe(struct callout *c, int flags, callout_func_t *drain)
1088 {
1089 	struct callout_cpu *cc, *old_cc;
1090 	struct lock_class *class;
1091 	int direct, sq_locked, use_lock;
1092 	int cancelled, not_on_a_list;
1093 
1094 	if ((flags & CS_DRAIN) != 0)
1095 		WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK, c->c_lock,
1096 		    "calling %s", __func__);
1097 
1098 	KASSERT((flags & CS_DRAIN) == 0 || drain == NULL,
1099 	    ("Cannot set drain callback and CS_DRAIN flag at the same time"));
1100 
1101 	/*
1102 	 * Some old subsystems don't hold Giant while running a callout_stop(),
1103 	 * so just discard this check for the moment.
1104 	 */
1105 	if ((flags & CS_DRAIN) == 0 && c->c_lock != NULL) {
1106 		if (c->c_lock == &Giant.lock_object)
1107 			use_lock = mtx_owned(&Giant);
1108 		else {
1109 			use_lock = 1;
1110 			class = LOCK_CLASS(c->c_lock);
1111 			class->lc_assert(c->c_lock, LA_XLOCKED);
1112 		}
1113 	} else
1114 		use_lock = 0;
1115 	if (c->c_iflags & CALLOUT_DIRECT) {
1116 		direct = 1;
1117 	} else {
1118 		direct = 0;
1119 	}
1120 	sq_locked = 0;
1121 	old_cc = NULL;
1122 again:
1123 	cc = callout_lock(c);
1124 
1125 	if ((c->c_iflags & (CALLOUT_DFRMIGRATION | CALLOUT_PENDING)) ==
1126 	    (CALLOUT_DFRMIGRATION | CALLOUT_PENDING) &&
1127 	    ((c->c_flags & CALLOUT_ACTIVE) == CALLOUT_ACTIVE)) {
1128 		/*
1129 		 * Special case where this slipped in while we
1130 		 * were migrating *as* the callout is about to
1131 		 * execute. The caller probably holds the lock
1132 		 * the callout wants.
1133 		 *
1134 		 * Get rid of the migration first. Then set
1135 		 * the flag that tells this code *not* to
1136 		 * try to remove it from any lists (its not
1137 		 * on one yet). When the callout wheel runs,
1138 		 * it will ignore this callout.
1139 		 */
1140 		c->c_iflags &= ~CALLOUT_PENDING;
1141 		c->c_flags &= ~CALLOUT_ACTIVE;
1142 		not_on_a_list = 1;
1143 	} else {
1144 		not_on_a_list = 0;
1145 	}
1146 
1147 	/*
1148 	 * If the callout was migrating while the callout cpu lock was
1149 	 * dropped,  just drop the sleepqueue lock and check the states
1150 	 * again.
1151 	 */
1152 	if (sq_locked != 0 && cc != old_cc) {
1153 #ifdef SMP
1154 		CC_UNLOCK(cc);
1155 		sleepq_release(&cc_exec_waiting(old_cc, direct));
1156 		sq_locked = 0;
1157 		old_cc = NULL;
1158 		goto again;
1159 #else
1160 		panic("migration should not happen");
1161 #endif
1162 	}
1163 
1164 	/*
1165 	 * If the callout is running, try to stop it or drain it.
1166 	 */
1167 	if (cc_exec_curr(cc, direct) == c) {
1168 		/*
1169 		 * Succeed we to stop it or not, we must clear the
1170 		 * active flag - this is what API users expect.  If we're
1171 		 * draining and the callout is currently executing, first wait
1172 		 * until it finishes.
1173 		 */
1174 		if ((flags & CS_DRAIN) == 0)
1175 			c->c_flags &= ~CALLOUT_ACTIVE;
1176 
1177 		if ((flags & CS_DRAIN) != 0) {
1178 			/*
1179 			 * The current callout is running (or just
1180 			 * about to run) and blocking is allowed, so
1181 			 * just wait for the current invocation to
1182 			 * finish.
1183 			 */
1184 			if (cc_exec_curr(cc, direct) == c) {
1185 				/*
1186 				 * Use direct calls to sleepqueue interface
1187 				 * instead of cv/msleep in order to avoid
1188 				 * a LOR between cc_lock and sleepqueue
1189 				 * chain spinlocks.  This piece of code
1190 				 * emulates a msleep_spin() call actually.
1191 				 *
1192 				 * If we already have the sleepqueue chain
1193 				 * locked, then we can safely block.  If we
1194 				 * don't already have it locked, however,
1195 				 * we have to drop the cc_lock to lock
1196 				 * it.  This opens several races, so we
1197 				 * restart at the beginning once we have
1198 				 * both locks.  If nothing has changed, then
1199 				 * we will end up back here with sq_locked
1200 				 * set.
1201 				 */
1202 				if (!sq_locked) {
1203 					CC_UNLOCK(cc);
1204 					sleepq_lock(
1205 					    &cc_exec_waiting(cc, direct));
1206 					sq_locked = 1;
1207 					old_cc = cc;
1208 					goto again;
1209 				}
1210 
1211 				/*
1212 				 * Migration could be cancelled here, but
1213 				 * as long as it is still not sure when it
1214 				 * will be packed up, just let softclock()
1215 				 * take care of it.
1216 				 */
1217 				cc_exec_waiting(cc, direct) = true;
1218 				DROP_GIANT();
1219 				CC_UNLOCK(cc);
1220 				sleepq_add(
1221 				    &cc_exec_waiting(cc, direct),
1222 				    &cc->cc_lock.lock_object, "codrain",
1223 				    SLEEPQ_SLEEP, 0);
1224 				sleepq_wait(
1225 				    &cc_exec_waiting(cc, direct),
1226 					     0);
1227 				sq_locked = 0;
1228 				old_cc = NULL;
1229 
1230 				/* Reacquire locks previously released. */
1231 				PICKUP_GIANT();
1232 				goto again;
1233 			}
1234 			c->c_flags &= ~CALLOUT_ACTIVE;
1235 		} else if (use_lock &&
1236 			   !cc_exec_cancel(cc, direct) && (drain == NULL)) {
1237 
1238 			/*
1239 			 * The current callout is waiting for its
1240 			 * lock which we hold.  Cancel the callout
1241 			 * and return.  After our caller drops the
1242 			 * lock, the callout will be skipped in
1243 			 * softclock(). This *only* works with a
1244 			 * callout_stop() *not* callout_drain() or
1245 			 * callout_async_drain().
1246 			 */
1247 			cc_exec_cancel(cc, direct) = true;
1248 			CTR3(KTR_CALLOUT, "cancelled %p func %p arg %p",
1249 			    c, c->c_func, c->c_arg);
1250 			KASSERT(!cc_cce_migrating(cc, direct),
1251 			    ("callout wrongly scheduled for migration"));
1252 			if (callout_migrating(c)) {
1253 				c->c_iflags &= ~CALLOUT_DFRMIGRATION;
1254 #ifdef SMP
1255 				cc_migration_cpu(cc, direct) = CPUBLOCK;
1256 				cc_migration_time(cc, direct) = 0;
1257 				cc_migration_prec(cc, direct) = 0;
1258 				cc_migration_func(cc, direct) = NULL;
1259 				cc_migration_arg(cc, direct) = NULL;
1260 #endif
1261 			}
1262 			CC_UNLOCK(cc);
1263 			KASSERT(!sq_locked, ("sleepqueue chain locked"));
1264 			return (1);
1265 		} else if (callout_migrating(c)) {
1266 			/*
1267 			 * The callout is currently being serviced
1268 			 * and the "next" callout is scheduled at
1269 			 * its completion with a migration. We remove
1270 			 * the migration flag so it *won't* get rescheduled,
1271 			 * but we can't stop the one thats running so
1272 			 * we return 0.
1273 			 */
1274 			c->c_iflags &= ~CALLOUT_DFRMIGRATION;
1275 #ifdef SMP
1276 			/*
1277 			 * We can't call cc_cce_cleanup here since
1278 			 * if we do it will remove .ce_curr and
1279 			 * its still running. This will prevent a
1280 			 * reschedule of the callout when the
1281 			 * execution completes.
1282 			 */
1283 			cc_migration_cpu(cc, direct) = CPUBLOCK;
1284 			cc_migration_time(cc, direct) = 0;
1285 			cc_migration_prec(cc, direct) = 0;
1286 			cc_migration_func(cc, direct) = NULL;
1287 			cc_migration_arg(cc, direct) = NULL;
1288 #endif
1289 			CTR3(KTR_CALLOUT, "postponing stop %p func %p arg %p",
1290 			    c, c->c_func, c->c_arg);
1291  			if (drain) {
1292 				KASSERT(cc_exec_drain(cc, direct) == NULL,
1293 				    ("callout drain function already set to %p",
1294 				    cc_exec_drain(cc, direct)));
1295 				cc_exec_drain(cc, direct) = drain;
1296 			}
1297 			CC_UNLOCK(cc);
1298 			return (0);
1299 		} else {
1300 			CTR3(KTR_CALLOUT, "failed to stop %p func %p arg %p",
1301 			    c, c->c_func, c->c_arg);
1302 			if (drain) {
1303 				KASSERT(cc_exec_drain(cc, direct) == NULL,
1304 				    ("callout drain function already set to %p",
1305 				    cc_exec_drain(cc, direct)));
1306 				cc_exec_drain(cc, direct) = drain;
1307 			}
1308 		}
1309 		KASSERT(!sq_locked, ("sleepqueue chain still locked"));
1310 		cancelled = 0;
1311 	} else
1312 		cancelled = 1;
1313 
1314 	if (sq_locked)
1315 		sleepq_release(&cc_exec_waiting(cc, direct));
1316 
1317 	if ((c->c_iflags & CALLOUT_PENDING) == 0) {
1318 		CTR3(KTR_CALLOUT, "failed to stop %p func %p arg %p",
1319 		    c, c->c_func, c->c_arg);
1320 		/*
1321 		 * For not scheduled and not executing callout return
1322 		 * negative value.
1323 		 */
1324 		if (cc_exec_curr(cc, direct) != c)
1325 			cancelled = -1;
1326 		CC_UNLOCK(cc);
1327 		return (cancelled);
1328 	}
1329 
1330 	c->c_iflags &= ~CALLOUT_PENDING;
1331 	c->c_flags &= ~CALLOUT_ACTIVE;
1332 
1333 	CTR3(KTR_CALLOUT, "cancelled %p func %p arg %p",
1334 	    c, c->c_func, c->c_arg);
1335 	if (not_on_a_list == 0) {
1336 		if ((c->c_iflags & CALLOUT_PROCESSED) == 0) {
1337 			if (cc_exec_next(cc) == c)
1338 				cc_exec_next(cc) = LIST_NEXT(c, c_links.le);
1339 			LIST_REMOVE(c, c_links.le);
1340 		} else {
1341 			TAILQ_REMOVE(&cc->cc_expireq, c, c_links.tqe);
1342 		}
1343 	}
1344 	CC_UNLOCK(cc);
1345 	return (cancelled);
1346 }
1347 
1348 void
1349 callout_init(struct callout *c, int mpsafe)
1350 {
1351 	bzero(c, sizeof *c);
1352 	if (mpsafe) {
1353 		c->c_lock = NULL;
1354 		c->c_iflags = CALLOUT_RETURNUNLOCKED;
1355 	} else {
1356 		c->c_lock = &Giant.lock_object;
1357 		c->c_iflags = 0;
1358 	}
1359 	c->c_cpu = cc_default_cpu;
1360 }
1361 
1362 void
1363 _callout_init_lock(struct callout *c, struct lock_object *lock, int flags)
1364 {
1365 	bzero(c, sizeof *c);
1366 	c->c_lock = lock;
1367 	KASSERT((flags & ~(CALLOUT_RETURNUNLOCKED | CALLOUT_SHAREDLOCK)) == 0,
1368 	    ("callout_init_lock: bad flags %d", flags));
1369 	KASSERT(lock != NULL || (flags & CALLOUT_RETURNUNLOCKED) == 0,
1370 	    ("callout_init_lock: CALLOUT_RETURNUNLOCKED with no lock"));
1371 	KASSERT(lock == NULL || !(LOCK_CLASS(lock)->lc_flags & LC_SLEEPABLE),
1372 	    ("%s: callout %p has sleepable lock", __func__, c));
1373 	c->c_iflags = flags & (CALLOUT_RETURNUNLOCKED | CALLOUT_SHAREDLOCK);
1374 	c->c_cpu = cc_default_cpu;
1375 }
1376 
1377 static int
1378 flssbt(sbintime_t sbt)
1379 {
1380 
1381 	sbt += (uint64_t)sbt >> 1;
1382 	if (sizeof(long) >= sizeof(sbintime_t))
1383 		return (flsl(sbt));
1384 	if (sbt >= SBT_1S)
1385 		return (flsl(((uint64_t)sbt) >> 32) + 32);
1386 	return (flsl(sbt));
1387 }
1388 
1389 /*
1390  * Dump immediate statistic snapshot of the scheduled callouts.
1391  */
1392 static int
1393 sysctl_kern_callout_stat(SYSCTL_HANDLER_ARGS)
1394 {
1395 	struct callout *tmp;
1396 	struct callout_cpu *cc;
1397 	struct callout_list *sc;
1398 	sbintime_t maxpr, maxt, medpr, medt, now, spr, st, t;
1399 	int ct[64], cpr[64], ccpbk[32];
1400 	int error, val, i, count, tcum, pcum, maxc, c, medc;
1401 	int cpu;
1402 
1403 	val = 0;
1404 	error = sysctl_handle_int(oidp, &val, 0, req);
1405 	if (error != 0 || req->newptr == NULL)
1406 		return (error);
1407 	count = maxc = 0;
1408 	st = spr = maxt = maxpr = 0;
1409 	bzero(ccpbk, sizeof(ccpbk));
1410 	bzero(ct, sizeof(ct));
1411 	bzero(cpr, sizeof(cpr));
1412 	now = sbinuptime();
1413 	CPU_FOREACH(cpu) {
1414 		cc = CC_CPU(cpu);
1415 		CC_LOCK(cc);
1416 		for (i = 0; i < callwheelsize; i++) {
1417 			sc = &cc->cc_callwheel[i];
1418 			c = 0;
1419 			LIST_FOREACH(tmp, sc, c_links.le) {
1420 				c++;
1421 				t = tmp->c_time - now;
1422 				if (t < 0)
1423 					t = 0;
1424 				st += t / SBT_1US;
1425 				spr += tmp->c_precision / SBT_1US;
1426 				if (t > maxt)
1427 					maxt = t;
1428 				if (tmp->c_precision > maxpr)
1429 					maxpr = tmp->c_precision;
1430 				ct[flssbt(t)]++;
1431 				cpr[flssbt(tmp->c_precision)]++;
1432 			}
1433 			if (c > maxc)
1434 				maxc = c;
1435 			ccpbk[fls(c + c / 2)]++;
1436 			count += c;
1437 		}
1438 		CC_UNLOCK(cc);
1439 	}
1440 
1441 	for (i = 0, tcum = 0; i < 64 && tcum < count / 2; i++)
1442 		tcum += ct[i];
1443 	medt = (i >= 2) ? (((sbintime_t)1) << (i - 2)) : 0;
1444 	for (i = 0, pcum = 0; i < 64 && pcum < count / 2; i++)
1445 		pcum += cpr[i];
1446 	medpr = (i >= 2) ? (((sbintime_t)1) << (i - 2)) : 0;
1447 	for (i = 0, c = 0; i < 32 && c < count / 2; i++)
1448 		c += ccpbk[i];
1449 	medc = (i >= 2) ? (1 << (i - 2)) : 0;
1450 
1451 	printf("Scheduled callouts statistic snapshot:\n");
1452 	printf("  Callouts: %6d  Buckets: %6d*%-3d  Bucket size: 0.%06ds\n",
1453 	    count, callwheelsize, mp_ncpus, 1000000 >> CC_HASH_SHIFT);
1454 	printf("  C/Bk: med %5d         avg %6d.%06jd  max %6d\n",
1455 	    medc,
1456 	    count / callwheelsize / mp_ncpus,
1457 	    (uint64_t)count * 1000000 / callwheelsize / mp_ncpus % 1000000,
1458 	    maxc);
1459 	printf("  Time: med %5jd.%06jds avg %6jd.%06jds max %6jd.%06jds\n",
1460 	    medt / SBT_1S, (medt & 0xffffffff) * 1000000 >> 32,
1461 	    (st / count) / 1000000, (st / count) % 1000000,
1462 	    maxt / SBT_1S, (maxt & 0xffffffff) * 1000000 >> 32);
1463 	printf("  Prec: med %5jd.%06jds avg %6jd.%06jds max %6jd.%06jds\n",
1464 	    medpr / SBT_1S, (medpr & 0xffffffff) * 1000000 >> 32,
1465 	    (spr / count) / 1000000, (spr / count) % 1000000,
1466 	    maxpr / SBT_1S, (maxpr & 0xffffffff) * 1000000 >> 32);
1467 	printf("  Distribution:       \tbuckets\t   time\t   tcum\t"
1468 	    "   prec\t   pcum\n");
1469 	for (i = 0, tcum = pcum = 0; i < 64; i++) {
1470 		if (ct[i] == 0 && cpr[i] == 0)
1471 			continue;
1472 		t = (i != 0) ? (((sbintime_t)1) << (i - 1)) : 0;
1473 		tcum += ct[i];
1474 		pcum += cpr[i];
1475 		printf("  %10jd.%06jds\t 2**%d\t%7d\t%7d\t%7d\t%7d\n",
1476 		    t / SBT_1S, (t & 0xffffffff) * 1000000 >> 32,
1477 		    i - 1 - (32 - CC_HASH_SHIFT),
1478 		    ct[i], tcum, cpr[i], pcum);
1479 	}
1480 	return (error);
1481 }
1482 SYSCTL_PROC(_kern, OID_AUTO, callout_stat,
1483     CTLTYPE_INT | CTLFLAG_RW | CTLFLAG_MPSAFE,
1484     0, 0, sysctl_kern_callout_stat, "I",
1485     "Dump immediate statistic snapshot of the scheduled callouts");
1486 
1487 #ifdef DDB
1488 static void
1489 _show_callout(struct callout *c)
1490 {
1491 
1492 	db_printf("callout %p\n", c);
1493 #define	C_DB_PRINTF(f, e)	db_printf("   %s = " f "\n", #e, c->e);
1494 	db_printf("   &c_links = %p\n", &(c->c_links));
1495 	C_DB_PRINTF("%" PRId64,	c_time);
1496 	C_DB_PRINTF("%" PRId64,	c_precision);
1497 	C_DB_PRINTF("%p",	c_arg);
1498 	C_DB_PRINTF("%p",	c_func);
1499 	C_DB_PRINTF("%p",	c_lock);
1500 	C_DB_PRINTF("%#x",	c_flags);
1501 	C_DB_PRINTF("%#x",	c_iflags);
1502 	C_DB_PRINTF("%d",	c_cpu);
1503 #undef	C_DB_PRINTF
1504 }
1505 
1506 DB_SHOW_COMMAND(callout, db_show_callout)
1507 {
1508 
1509 	if (!have_addr) {
1510 		db_printf("usage: show callout <struct callout *>\n");
1511 		return;
1512 	}
1513 
1514 	_show_callout((struct callout *)addr);
1515 }
1516 
1517 static void
1518 _show_last_callout(int cpu, int direct, const char *dirstr)
1519 {
1520 	struct callout_cpu *cc;
1521 	void *func, *arg;
1522 
1523 	cc = CC_CPU(cpu);
1524 	func = cc_exec_last_func(cc, direct);
1525 	arg = cc_exec_last_arg(cc, direct);
1526 	db_printf("cpu %d last%s callout function: %p ", cpu, dirstr, func);
1527 	db_printsym((db_expr_t)func, DB_STGY_ANY);
1528 	db_printf("\ncpu %d last%s callout argument: %p\n", cpu, dirstr, arg);
1529 }
1530 
1531 DB_SHOW_COMMAND_FLAGS(callout_last, db_show_callout_last, DB_CMD_MEMSAFE)
1532 {
1533 	int cpu, last;
1534 
1535 	if (have_addr) {
1536 		if (addr < 0 || addr > mp_maxid || CPU_ABSENT(addr)) {
1537 			db_printf("no such cpu: %d\n", (int)addr);
1538 			return;
1539 		}
1540 		cpu = last = addr;
1541 	} else {
1542 		cpu = 0;
1543 		last = mp_maxid;
1544 	}
1545 
1546 	while (cpu <= last) {
1547 		if (!CPU_ABSENT(cpu)) {
1548 			_show_last_callout(cpu, 0, "");
1549 			_show_last_callout(cpu, 1, " direct");
1550 		}
1551 		cpu++;
1552 	}
1553 }
1554 #endif /* DDB */
1555