xref: /freebsd/sys/kern/kern_timeout.c (revision 41059135ce931c0f1014a999ffabc6bc470ce856)
1 /*-
2  * Copyright (c) 1982, 1986, 1991, 1993
3  *	The Regents of the University of California.  All rights reserved.
4  * (c) UNIX System Laboratories, Inc.
5  * All or some portions of this file are derived from material licensed
6  * to the University of California by American Telephone and Telegraph
7  * Co. or Unix System Laboratories, Inc. and are reproduced herein with
8  * the permission of UNIX System Laboratories, Inc.
9  *
10  * Redistribution and use in source and binary forms, with or without
11  * modification, are permitted provided that the following conditions
12  * are met:
13  * 1. Redistributions of source code must retain the above copyright
14  *    notice, this list of conditions and the following disclaimer.
15  * 2. Redistributions in binary form must reproduce the above copyright
16  *    notice, this list of conditions and the following disclaimer in the
17  *    documentation and/or other materials provided with the distribution.
18  * 3. Neither the name of the University nor the names of its contributors
19  *    may be used to endorse or promote products derived from this software
20  *    without specific prior written permission.
21  *
22  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
23  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
24  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
25  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
26  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
27  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
28  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
29  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
30  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
31  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
32  * SUCH DAMAGE.
33  *
34  *	From: @(#)kern_clock.c	8.5 (Berkeley) 1/21/94
35  */
36 
37 #include <sys/cdefs.h>
38 __FBSDID("$FreeBSD$");
39 
40 #include "opt_callout_profiling.h"
41 #include "opt_ddb.h"
42 #if defined(__arm__)
43 #include "opt_timer.h"
44 #endif
45 #include "opt_rss.h"
46 
47 #include <sys/param.h>
48 #include <sys/systm.h>
49 #include <sys/bus.h>
50 #include <sys/callout.h>
51 #include <sys/file.h>
52 #include <sys/interrupt.h>
53 #include <sys/kernel.h>
54 #include <sys/ktr.h>
55 #include <sys/lock.h>
56 #include <sys/malloc.h>
57 #include <sys/mutex.h>
58 #include <sys/proc.h>
59 #include <sys/sdt.h>
60 #include <sys/sleepqueue.h>
61 #include <sys/sysctl.h>
62 #include <sys/smp.h>
63 
64 #ifdef DDB
65 #include <ddb/ddb.h>
66 #include <machine/_inttypes.h>
67 #endif
68 
69 #ifdef SMP
70 #include <machine/cpu.h>
71 #endif
72 
73 #ifndef NO_EVENTTIMERS
74 DPCPU_DECLARE(sbintime_t, hardclocktime);
75 #endif
76 
77 SDT_PROVIDER_DEFINE(callout_execute);
78 SDT_PROBE_DEFINE1(callout_execute, , , callout__start, "struct callout *");
79 SDT_PROBE_DEFINE1(callout_execute, , , callout__end, "struct callout *");
80 
81 #ifdef CALLOUT_PROFILING
82 static int avg_depth;
83 SYSCTL_INT(_debug, OID_AUTO, to_avg_depth, CTLFLAG_RD, &avg_depth, 0,
84     "Average number of items examined per softclock call. Units = 1/1000");
85 static int avg_gcalls;
86 SYSCTL_INT(_debug, OID_AUTO, to_avg_gcalls, CTLFLAG_RD, &avg_gcalls, 0,
87     "Average number of Giant callouts made per softclock call. Units = 1/1000");
88 static int avg_lockcalls;
89 SYSCTL_INT(_debug, OID_AUTO, to_avg_lockcalls, CTLFLAG_RD, &avg_lockcalls, 0,
90     "Average number of lock callouts made per softclock call. Units = 1/1000");
91 static int avg_mpcalls;
92 SYSCTL_INT(_debug, OID_AUTO, to_avg_mpcalls, CTLFLAG_RD, &avg_mpcalls, 0,
93     "Average number of MP callouts made per softclock call. Units = 1/1000");
94 static int avg_depth_dir;
95 SYSCTL_INT(_debug, OID_AUTO, to_avg_depth_dir, CTLFLAG_RD, &avg_depth_dir, 0,
96     "Average number of direct callouts examined per callout_process call. "
97     "Units = 1/1000");
98 static int avg_lockcalls_dir;
99 SYSCTL_INT(_debug, OID_AUTO, to_avg_lockcalls_dir, CTLFLAG_RD,
100     &avg_lockcalls_dir, 0, "Average number of lock direct callouts made per "
101     "callout_process call. Units = 1/1000");
102 static int avg_mpcalls_dir;
103 SYSCTL_INT(_debug, OID_AUTO, to_avg_mpcalls_dir, CTLFLAG_RD, &avg_mpcalls_dir,
104     0, "Average number of MP direct callouts made per callout_process call. "
105     "Units = 1/1000");
106 #endif
107 
108 static int ncallout;
109 SYSCTL_INT(_kern, OID_AUTO, ncallout, CTLFLAG_RDTUN | CTLFLAG_NOFETCH, &ncallout, 0,
110     "Number of entries in callwheel and size of timeout() preallocation");
111 
112 #ifdef	RSS
113 static int pin_default_swi = 1;
114 static int pin_pcpu_swi = 1;
115 #else
116 static int pin_default_swi = 0;
117 static int pin_pcpu_swi = 0;
118 #endif
119 
120 SYSCTL_INT(_kern, OID_AUTO, pin_default_swi, CTLFLAG_RDTUN | CTLFLAG_NOFETCH, &pin_default_swi,
121     0, "Pin the default (non-per-cpu) swi (shared with PCPU 0 swi)");
122 SYSCTL_INT(_kern, OID_AUTO, pin_pcpu_swi, CTLFLAG_RDTUN | CTLFLAG_NOFETCH, &pin_pcpu_swi,
123     0, "Pin the per-CPU swis (except PCPU 0, which is also default");
124 
125 /*
126  * TODO:
127  *	allocate more timeout table slots when table overflows.
128  */
129 u_int callwheelsize, callwheelmask;
130 
131 /*
132  * The callout cpu exec entities represent informations necessary for
133  * describing the state of callouts currently running on the CPU and the ones
134  * necessary for migrating callouts to the new callout cpu. In particular,
135  * the first entry of the array cc_exec_entity holds informations for callout
136  * running in SWI thread context, while the second one holds informations
137  * for callout running directly from hardware interrupt context.
138  * The cached informations are very important for deferring migration when
139  * the migrating callout is already running.
140  */
141 struct cc_exec {
142 	struct callout		*cc_curr;
143 	void			(*cc_drain)(void *);
144 #ifdef SMP
145 	void			(*ce_migration_func)(void *);
146 	void			*ce_migration_arg;
147 	int			ce_migration_cpu;
148 	sbintime_t		ce_migration_time;
149 	sbintime_t		ce_migration_prec;
150 #endif
151 	bool			cc_cancel;
152 	bool			cc_waiting;
153 };
154 
155 /*
156  * There is one struct callout_cpu per cpu, holding all relevant
157  * state for the callout processing thread on the individual CPU.
158  */
159 struct callout_cpu {
160 	struct mtx_padalign	cc_lock;
161 	struct cc_exec 		cc_exec_entity[2];
162 	struct callout		*cc_next;
163 	struct callout		*cc_callout;
164 	struct callout_list	*cc_callwheel;
165 	struct callout_tailq	cc_expireq;
166 	struct callout_slist	cc_callfree;
167 	sbintime_t		cc_firstevent;
168 	sbintime_t		cc_lastscan;
169 	void			*cc_cookie;
170 	u_int			cc_bucket;
171 	u_int			cc_inited;
172 	char			cc_ktr_event_name[20];
173 };
174 
175 #define	callout_migrating(c)	((c)->c_iflags & CALLOUT_DFRMIGRATION)
176 
177 #define	cc_exec_curr(cc, dir)		cc->cc_exec_entity[dir].cc_curr
178 #define	cc_exec_drain(cc, dir)		cc->cc_exec_entity[dir].cc_drain
179 #define	cc_exec_next(cc)		cc->cc_next
180 #define	cc_exec_cancel(cc, dir)		cc->cc_exec_entity[dir].cc_cancel
181 #define	cc_exec_waiting(cc, dir)	cc->cc_exec_entity[dir].cc_waiting
182 #ifdef SMP
183 #define	cc_migration_func(cc, dir)	cc->cc_exec_entity[dir].ce_migration_func
184 #define	cc_migration_arg(cc, dir)	cc->cc_exec_entity[dir].ce_migration_arg
185 #define	cc_migration_cpu(cc, dir)	cc->cc_exec_entity[dir].ce_migration_cpu
186 #define	cc_migration_time(cc, dir)	cc->cc_exec_entity[dir].ce_migration_time
187 #define	cc_migration_prec(cc, dir)	cc->cc_exec_entity[dir].ce_migration_prec
188 
189 struct callout_cpu cc_cpu[MAXCPU];
190 #define	CPUBLOCK	MAXCPU
191 #define	CC_CPU(cpu)	(&cc_cpu[(cpu)])
192 #define	CC_SELF()	CC_CPU(PCPU_GET(cpuid))
193 #else
194 struct callout_cpu cc_cpu;
195 #define	CC_CPU(cpu)	&cc_cpu
196 #define	CC_SELF()	&cc_cpu
197 #endif
198 #define	CC_LOCK(cc)	mtx_lock_spin(&(cc)->cc_lock)
199 #define	CC_UNLOCK(cc)	mtx_unlock_spin(&(cc)->cc_lock)
200 #define	CC_LOCK_ASSERT(cc)	mtx_assert(&(cc)->cc_lock, MA_OWNED)
201 
202 static int timeout_cpu;
203 
204 static void	callout_cpu_init(struct callout_cpu *cc, int cpu);
205 static void	softclock_call_cc(struct callout *c, struct callout_cpu *cc,
206 #ifdef CALLOUT_PROFILING
207 		    int *mpcalls, int *lockcalls, int *gcalls,
208 #endif
209 		    int direct);
210 
211 static MALLOC_DEFINE(M_CALLOUT, "callout", "Callout datastructures");
212 
213 /**
214  * Locked by cc_lock:
215  *   cc_curr         - If a callout is in progress, it is cc_curr.
216  *                     If cc_curr is non-NULL, threads waiting in
217  *                     callout_drain() will be woken up as soon as the
218  *                     relevant callout completes.
219  *   cc_cancel       - Changing to 1 with both callout_lock and cc_lock held
220  *                     guarantees that the current callout will not run.
221  *                     The softclock() function sets this to 0 before it
222  *                     drops callout_lock to acquire c_lock, and it calls
223  *                     the handler only if curr_cancelled is still 0 after
224  *                     cc_lock is successfully acquired.
225  *   cc_waiting      - If a thread is waiting in callout_drain(), then
226  *                     callout_wait is nonzero.  Set only when
227  *                     cc_curr is non-NULL.
228  */
229 
230 /*
231  * Resets the execution entity tied to a specific callout cpu.
232  */
233 static void
234 cc_cce_cleanup(struct callout_cpu *cc, int direct)
235 {
236 
237 	cc_exec_curr(cc, direct) = NULL;
238 	cc_exec_cancel(cc, direct) = false;
239 	cc_exec_waiting(cc, direct) = false;
240 #ifdef SMP
241 	cc_migration_cpu(cc, direct) = CPUBLOCK;
242 	cc_migration_time(cc, direct) = 0;
243 	cc_migration_prec(cc, direct) = 0;
244 	cc_migration_func(cc, direct) = NULL;
245 	cc_migration_arg(cc, direct) = NULL;
246 #endif
247 }
248 
249 /*
250  * Checks if migration is requested by a specific callout cpu.
251  */
252 static int
253 cc_cce_migrating(struct callout_cpu *cc, int direct)
254 {
255 
256 #ifdef SMP
257 	return (cc_migration_cpu(cc, direct) != CPUBLOCK);
258 #else
259 	return (0);
260 #endif
261 }
262 
263 /*
264  * Kernel low level callwheel initialization
265  * called on cpu0 during kernel startup.
266  */
267 static void
268 callout_callwheel_init(void *dummy)
269 {
270 	struct callout_cpu *cc;
271 
272 	/*
273 	 * Calculate the size of the callout wheel and the preallocated
274 	 * timeout() structures.
275 	 * XXX: Clip callout to result of previous function of maxusers
276 	 * maximum 384.  This is still huge, but acceptable.
277 	 */
278 	memset(CC_CPU(0), 0, sizeof(cc_cpu));
279 	ncallout = imin(16 + maxproc + maxfiles, 18508);
280 	TUNABLE_INT_FETCH("kern.ncallout", &ncallout);
281 
282 	/*
283 	 * Calculate callout wheel size, should be next power of two higher
284 	 * than 'ncallout'.
285 	 */
286 	callwheelsize = 1 << fls(ncallout);
287 	callwheelmask = callwheelsize - 1;
288 
289 	/*
290 	 * Fetch whether we're pinning the swi's or not.
291 	 */
292 	TUNABLE_INT_FETCH("kern.pin_default_swi", &pin_default_swi);
293 	TUNABLE_INT_FETCH("kern.pin_pcpu_swi", &pin_pcpu_swi);
294 
295 	/*
296 	 * Only cpu0 handles timeout(9) and receives a preallocation.
297 	 *
298 	 * XXX: Once all timeout(9) consumers are converted this can
299 	 * be removed.
300 	 */
301 	timeout_cpu = PCPU_GET(cpuid);
302 	cc = CC_CPU(timeout_cpu);
303 	cc->cc_callout = malloc(ncallout * sizeof(struct callout),
304 	    M_CALLOUT, M_WAITOK);
305 	callout_cpu_init(cc, timeout_cpu);
306 }
307 SYSINIT(callwheel_init, SI_SUB_CPU, SI_ORDER_ANY, callout_callwheel_init, NULL);
308 
309 /*
310  * Initialize the per-cpu callout structures.
311  */
312 static void
313 callout_cpu_init(struct callout_cpu *cc, int cpu)
314 {
315 	struct callout *c;
316 	int i;
317 
318 	mtx_init(&cc->cc_lock, "callout", NULL, MTX_SPIN | MTX_RECURSE);
319 	SLIST_INIT(&cc->cc_callfree);
320 	cc->cc_inited = 1;
321 	cc->cc_callwheel = malloc(sizeof(struct callout_list) * callwheelsize,
322 	    M_CALLOUT, M_WAITOK);
323 	for (i = 0; i < callwheelsize; i++)
324 		LIST_INIT(&cc->cc_callwheel[i]);
325 	TAILQ_INIT(&cc->cc_expireq);
326 	cc->cc_firstevent = SBT_MAX;
327 	for (i = 0; i < 2; i++)
328 		cc_cce_cleanup(cc, i);
329 	snprintf(cc->cc_ktr_event_name, sizeof(cc->cc_ktr_event_name),
330 	    "callwheel cpu %d", cpu);
331 	if (cc->cc_callout == NULL)	/* Only cpu0 handles timeout(9) */
332 		return;
333 	for (i = 0; i < ncallout; i++) {
334 		c = &cc->cc_callout[i];
335 		callout_init(c, 0);
336 		c->c_iflags = CALLOUT_LOCAL_ALLOC;
337 		SLIST_INSERT_HEAD(&cc->cc_callfree, c, c_links.sle);
338 	}
339 }
340 
341 #ifdef SMP
342 /*
343  * Switches the cpu tied to a specific callout.
344  * The function expects a locked incoming callout cpu and returns with
345  * locked outcoming callout cpu.
346  */
347 static struct callout_cpu *
348 callout_cpu_switch(struct callout *c, struct callout_cpu *cc, int new_cpu)
349 {
350 	struct callout_cpu *new_cc;
351 
352 	MPASS(c != NULL && cc != NULL);
353 	CC_LOCK_ASSERT(cc);
354 
355 	/*
356 	 * Avoid interrupts and preemption firing after the callout cpu
357 	 * is blocked in order to avoid deadlocks as the new thread
358 	 * may be willing to acquire the callout cpu lock.
359 	 */
360 	c->c_cpu = CPUBLOCK;
361 	spinlock_enter();
362 	CC_UNLOCK(cc);
363 	new_cc = CC_CPU(new_cpu);
364 	CC_LOCK(new_cc);
365 	spinlock_exit();
366 	c->c_cpu = new_cpu;
367 	return (new_cc);
368 }
369 #endif
370 
371 /*
372  * Start standard softclock thread.
373  */
374 static void
375 start_softclock(void *dummy)
376 {
377 	struct callout_cpu *cc;
378 	char name[MAXCOMLEN];
379 #ifdef SMP
380 	int cpu;
381 	struct intr_event *ie;
382 #endif
383 
384 	cc = CC_CPU(timeout_cpu);
385 	snprintf(name, sizeof(name), "clock (%d)", timeout_cpu);
386 	if (swi_add(&clk_intr_event, name, softclock, cc, SWI_CLOCK,
387 	    INTR_MPSAFE, &cc->cc_cookie))
388 		panic("died while creating standard software ithreads");
389 	if (pin_default_swi &&
390 	    (intr_event_bind(clk_intr_event, timeout_cpu) != 0)) {
391 		printf("%s: timeout clock couldn't be pinned to cpu %d\n",
392 		    __func__,
393 		    timeout_cpu);
394 	}
395 
396 #ifdef SMP
397 	CPU_FOREACH(cpu) {
398 		if (cpu == timeout_cpu)
399 			continue;
400 		cc = CC_CPU(cpu);
401 		cc->cc_callout = NULL;	/* Only cpu0 handles timeout(9). */
402 		callout_cpu_init(cc, cpu);
403 		snprintf(name, sizeof(name), "clock (%d)", cpu);
404 		ie = NULL;
405 		if (swi_add(&ie, name, softclock, cc, SWI_CLOCK,
406 		    INTR_MPSAFE, &cc->cc_cookie))
407 			panic("died while creating standard software ithreads");
408 		if (pin_pcpu_swi && (intr_event_bind(ie, cpu) != 0)) {
409 			printf("%s: per-cpu clock couldn't be pinned to "
410 			    "cpu %d\n",
411 			    __func__,
412 			    cpu);
413 		}
414 	}
415 #endif
416 }
417 SYSINIT(start_softclock, SI_SUB_SOFTINTR, SI_ORDER_FIRST, start_softclock, NULL);
418 
419 #define	CC_HASH_SHIFT	8
420 
421 static inline u_int
422 callout_hash(sbintime_t sbt)
423 {
424 
425 	return (sbt >> (32 - CC_HASH_SHIFT));
426 }
427 
428 static inline u_int
429 callout_get_bucket(sbintime_t sbt)
430 {
431 
432 	return (callout_hash(sbt) & callwheelmask);
433 }
434 
435 void
436 callout_process(sbintime_t now)
437 {
438 	struct callout *tmp, *tmpn;
439 	struct callout_cpu *cc;
440 	struct callout_list *sc;
441 	sbintime_t first, last, max, tmp_max;
442 	uint32_t lookahead;
443 	u_int firstb, lastb, nowb;
444 #ifdef CALLOUT_PROFILING
445 	int depth_dir = 0, mpcalls_dir = 0, lockcalls_dir = 0;
446 #endif
447 
448 	cc = CC_SELF();
449 	mtx_lock_spin_flags(&cc->cc_lock, MTX_QUIET);
450 
451 	/* Compute the buckets of the last scan and present times. */
452 	firstb = callout_hash(cc->cc_lastscan);
453 	cc->cc_lastscan = now;
454 	nowb = callout_hash(now);
455 
456 	/* Compute the last bucket and minimum time of the bucket after it. */
457 	if (nowb == firstb)
458 		lookahead = (SBT_1S / 16);
459 	else if (nowb - firstb == 1)
460 		lookahead = (SBT_1S / 8);
461 	else
462 		lookahead = (SBT_1S / 2);
463 	first = last = now;
464 	first += (lookahead / 2);
465 	last += lookahead;
466 	last &= (0xffffffffffffffffLLU << (32 - CC_HASH_SHIFT));
467 	lastb = callout_hash(last) - 1;
468 	max = last;
469 
470 	/*
471 	 * Check if we wrapped around the entire wheel from the last scan.
472 	 * In case, we need to scan entirely the wheel for pending callouts.
473 	 */
474 	if (lastb - firstb >= callwheelsize) {
475 		lastb = firstb + callwheelsize - 1;
476 		if (nowb - firstb >= callwheelsize)
477 			nowb = lastb;
478 	}
479 
480 	/* Iterate callwheel from firstb to nowb and then up to lastb. */
481 	do {
482 		sc = &cc->cc_callwheel[firstb & callwheelmask];
483 		tmp = LIST_FIRST(sc);
484 		while (tmp != NULL) {
485 			/* Run the callout if present time within allowed. */
486 			if (tmp->c_time <= now) {
487 				/*
488 				 * Consumer told us the callout may be run
489 				 * directly from hardware interrupt context.
490 				 */
491 				if (tmp->c_iflags & CALLOUT_DIRECT) {
492 #ifdef CALLOUT_PROFILING
493 					++depth_dir;
494 #endif
495 					cc_exec_next(cc) =
496 					    LIST_NEXT(tmp, c_links.le);
497 					cc->cc_bucket = firstb & callwheelmask;
498 					LIST_REMOVE(tmp, c_links.le);
499 					softclock_call_cc(tmp, cc,
500 #ifdef CALLOUT_PROFILING
501 					    &mpcalls_dir, &lockcalls_dir, NULL,
502 #endif
503 					    1);
504 					tmp = cc_exec_next(cc);
505 					cc_exec_next(cc) = NULL;
506 				} else {
507 					tmpn = LIST_NEXT(tmp, c_links.le);
508 					LIST_REMOVE(tmp, c_links.le);
509 					TAILQ_INSERT_TAIL(&cc->cc_expireq,
510 					    tmp, c_links.tqe);
511 					tmp->c_iflags |= CALLOUT_PROCESSED;
512 					tmp = tmpn;
513 				}
514 				continue;
515 			}
516 			/* Skip events from distant future. */
517 			if (tmp->c_time >= max)
518 				goto next;
519 			/*
520 			 * Event minimal time is bigger than present maximal
521 			 * time, so it cannot be aggregated.
522 			 */
523 			if (tmp->c_time > last) {
524 				lastb = nowb;
525 				goto next;
526 			}
527 			/* Update first and last time, respecting this event. */
528 			if (tmp->c_time < first)
529 				first = tmp->c_time;
530 			tmp_max = tmp->c_time + tmp->c_precision;
531 			if (tmp_max < last)
532 				last = tmp_max;
533 next:
534 			tmp = LIST_NEXT(tmp, c_links.le);
535 		}
536 		/* Proceed with the next bucket. */
537 		firstb++;
538 		/*
539 		 * Stop if we looked after present time and found
540 		 * some event we can't execute at now.
541 		 * Stop if we looked far enough into the future.
542 		 */
543 	} while (((int)(firstb - lastb)) <= 0);
544 	cc->cc_firstevent = last;
545 #ifndef NO_EVENTTIMERS
546 	cpu_new_callout(curcpu, last, first);
547 #endif
548 #ifdef CALLOUT_PROFILING
549 	avg_depth_dir += (depth_dir * 1000 - avg_depth_dir) >> 8;
550 	avg_mpcalls_dir += (mpcalls_dir * 1000 - avg_mpcalls_dir) >> 8;
551 	avg_lockcalls_dir += (lockcalls_dir * 1000 - avg_lockcalls_dir) >> 8;
552 #endif
553 	mtx_unlock_spin_flags(&cc->cc_lock, MTX_QUIET);
554 	/*
555 	 * swi_sched acquires the thread lock, so we don't want to call it
556 	 * with cc_lock held; incorrect locking order.
557 	 */
558 	if (!TAILQ_EMPTY(&cc->cc_expireq))
559 		swi_sched(cc->cc_cookie, 0);
560 }
561 
562 static struct callout_cpu *
563 callout_lock(struct callout *c)
564 {
565 	struct callout_cpu *cc;
566 	int cpu;
567 
568 	for (;;) {
569 		cpu = c->c_cpu;
570 #ifdef SMP
571 		if (cpu == CPUBLOCK) {
572 			while (c->c_cpu == CPUBLOCK)
573 				cpu_spinwait();
574 			continue;
575 		}
576 #endif
577 		cc = CC_CPU(cpu);
578 		CC_LOCK(cc);
579 		if (cpu == c->c_cpu)
580 			break;
581 		CC_UNLOCK(cc);
582 	}
583 	return (cc);
584 }
585 
586 static void
587 callout_cc_add(struct callout *c, struct callout_cpu *cc,
588     sbintime_t sbt, sbintime_t precision, void (*func)(void *),
589     void *arg, int cpu, int flags)
590 {
591 	int bucket;
592 
593 	CC_LOCK_ASSERT(cc);
594 	if (sbt < cc->cc_lastscan)
595 		sbt = cc->cc_lastscan;
596 	c->c_arg = arg;
597 	c->c_iflags |= CALLOUT_PENDING;
598 	c->c_iflags &= ~CALLOUT_PROCESSED;
599 	c->c_flags |= CALLOUT_ACTIVE;
600 	if (flags & C_DIRECT_EXEC)
601 		c->c_iflags |= CALLOUT_DIRECT;
602 	c->c_func = func;
603 	c->c_time = sbt;
604 	c->c_precision = precision;
605 	bucket = callout_get_bucket(c->c_time);
606 	CTR3(KTR_CALLOUT, "precision set for %p: %d.%08x",
607 	    c, (int)(c->c_precision >> 32),
608 	    (u_int)(c->c_precision & 0xffffffff));
609 	LIST_INSERT_HEAD(&cc->cc_callwheel[bucket], c, c_links.le);
610 	if (cc->cc_bucket == bucket)
611 		cc_exec_next(cc) = c;
612 #ifndef NO_EVENTTIMERS
613 	/*
614 	 * Inform the eventtimers(4) subsystem there's a new callout
615 	 * that has been inserted, but only if really required.
616 	 */
617 	if (SBT_MAX - c->c_time < c->c_precision)
618 		c->c_precision = SBT_MAX - c->c_time;
619 	sbt = c->c_time + c->c_precision;
620 	if (sbt < cc->cc_firstevent) {
621 		cc->cc_firstevent = sbt;
622 		cpu_new_callout(cpu, sbt, c->c_time);
623 	}
624 #endif
625 }
626 
627 static void
628 callout_cc_del(struct callout *c, struct callout_cpu *cc)
629 {
630 
631 	if ((c->c_iflags & CALLOUT_LOCAL_ALLOC) == 0)
632 		return;
633 	c->c_func = NULL;
634 	SLIST_INSERT_HEAD(&cc->cc_callfree, c, c_links.sle);
635 }
636 
637 static void
638 softclock_call_cc(struct callout *c, struct callout_cpu *cc,
639 #ifdef CALLOUT_PROFILING
640     int *mpcalls, int *lockcalls, int *gcalls,
641 #endif
642     int direct)
643 {
644 	struct rm_priotracker tracker;
645 	void (*c_func)(void *);
646 	void *c_arg;
647 	struct lock_class *class;
648 	struct lock_object *c_lock;
649 	uintptr_t lock_status;
650 	int c_iflags;
651 #ifdef SMP
652 	struct callout_cpu *new_cc;
653 	void (*new_func)(void *);
654 	void *new_arg;
655 	int flags, new_cpu;
656 	sbintime_t new_prec, new_time;
657 #endif
658 #if defined(DIAGNOSTIC) || defined(CALLOUT_PROFILING)
659 	sbintime_t sbt1, sbt2;
660 	struct timespec ts2;
661 	static sbintime_t maxdt = 2 * SBT_1MS;	/* 2 msec */
662 	static timeout_t *lastfunc;
663 #endif
664 
665 	KASSERT((c->c_iflags & CALLOUT_PENDING) == CALLOUT_PENDING,
666 	    ("softclock_call_cc: pend %p %x", c, c->c_iflags));
667 	KASSERT((c->c_flags & CALLOUT_ACTIVE) == CALLOUT_ACTIVE,
668 	    ("softclock_call_cc: act %p %x", c, c->c_flags));
669 	class = (c->c_lock != NULL) ? LOCK_CLASS(c->c_lock) : NULL;
670 	lock_status = 0;
671 	if (c->c_flags & CALLOUT_SHAREDLOCK) {
672 		if (class == &lock_class_rm)
673 			lock_status = (uintptr_t)&tracker;
674 		else
675 			lock_status = 1;
676 	}
677 	c_lock = c->c_lock;
678 	c_func = c->c_func;
679 	c_arg = c->c_arg;
680 	c_iflags = c->c_iflags;
681 	if (c->c_iflags & CALLOUT_LOCAL_ALLOC)
682 		c->c_iflags = CALLOUT_LOCAL_ALLOC;
683 	else
684 		c->c_iflags &= ~CALLOUT_PENDING;
685 
686 	cc_exec_curr(cc, direct) = c;
687 	cc_exec_cancel(cc, direct) = false;
688 	cc_exec_drain(cc, direct) = NULL;
689 	CC_UNLOCK(cc);
690 	if (c_lock != NULL) {
691 		class->lc_lock(c_lock, lock_status);
692 		/*
693 		 * The callout may have been cancelled
694 		 * while we switched locks.
695 		 */
696 		if (cc_exec_cancel(cc, direct)) {
697 			class->lc_unlock(c_lock);
698 			goto skip;
699 		}
700 		/* The callout cannot be stopped now. */
701 		cc_exec_cancel(cc, direct) = true;
702 		if (c_lock == &Giant.lock_object) {
703 #ifdef CALLOUT_PROFILING
704 			(*gcalls)++;
705 #endif
706 			CTR3(KTR_CALLOUT, "callout giant %p func %p arg %p",
707 			    c, c_func, c_arg);
708 		} else {
709 #ifdef CALLOUT_PROFILING
710 			(*lockcalls)++;
711 #endif
712 			CTR3(KTR_CALLOUT, "callout lock %p func %p arg %p",
713 			    c, c_func, c_arg);
714 		}
715 	} else {
716 #ifdef CALLOUT_PROFILING
717 		(*mpcalls)++;
718 #endif
719 		CTR3(KTR_CALLOUT, "callout %p func %p arg %p",
720 		    c, c_func, c_arg);
721 	}
722 	KTR_STATE3(KTR_SCHED, "callout", cc->cc_ktr_event_name, "running",
723 	    "func:%p", c_func, "arg:%p", c_arg, "direct:%d", direct);
724 #if defined(DIAGNOSTIC) || defined(CALLOUT_PROFILING)
725 	sbt1 = sbinuptime();
726 #endif
727 	THREAD_NO_SLEEPING();
728 	SDT_PROBE1(callout_execute, , , callout__start, c);
729 	c_func(c_arg);
730 	SDT_PROBE1(callout_execute, , , callout__end, c);
731 	THREAD_SLEEPING_OK();
732 #if defined(DIAGNOSTIC) || defined(CALLOUT_PROFILING)
733 	sbt2 = sbinuptime();
734 	sbt2 -= sbt1;
735 	if (sbt2 > maxdt) {
736 		if (lastfunc != c_func || sbt2 > maxdt * 2) {
737 			ts2 = sbttots(sbt2);
738 			printf(
739 		"Expensive timeout(9) function: %p(%p) %jd.%09ld s\n",
740 			    c_func, c_arg, (intmax_t)ts2.tv_sec, ts2.tv_nsec);
741 		}
742 		maxdt = sbt2;
743 		lastfunc = c_func;
744 	}
745 #endif
746 	KTR_STATE0(KTR_SCHED, "callout", cc->cc_ktr_event_name, "idle");
747 	CTR1(KTR_CALLOUT, "callout %p finished", c);
748 	if ((c_iflags & CALLOUT_RETURNUNLOCKED) == 0)
749 		class->lc_unlock(c_lock);
750 skip:
751 	CC_LOCK(cc);
752 	KASSERT(cc_exec_curr(cc, direct) == c, ("mishandled cc_curr"));
753 	cc_exec_curr(cc, direct) = NULL;
754 	if (cc_exec_drain(cc, direct)) {
755 		void (*drain)(void *);
756 
757 		drain = cc_exec_drain(cc, direct);
758 		cc_exec_drain(cc, direct) = NULL;
759 		CC_UNLOCK(cc);
760 		drain(c_arg);
761 		CC_LOCK(cc);
762 	}
763 	if (cc_exec_waiting(cc, direct)) {
764 		/*
765 		 * There is someone waiting for the
766 		 * callout to complete.
767 		 * If the callout was scheduled for
768 		 * migration just cancel it.
769 		 */
770 		if (cc_cce_migrating(cc, direct)) {
771 			cc_cce_cleanup(cc, direct);
772 
773 			/*
774 			 * It should be assert here that the callout is not
775 			 * destroyed but that is not easy.
776 			 */
777 			c->c_iflags &= ~CALLOUT_DFRMIGRATION;
778 		}
779 		cc_exec_waiting(cc, direct) = false;
780 		CC_UNLOCK(cc);
781 		wakeup(&cc_exec_waiting(cc, direct));
782 		CC_LOCK(cc);
783 	} else if (cc_cce_migrating(cc, direct)) {
784 		KASSERT((c_iflags & CALLOUT_LOCAL_ALLOC) == 0,
785 		    ("Migrating legacy callout %p", c));
786 #ifdef SMP
787 		/*
788 		 * If the callout was scheduled for
789 		 * migration just perform it now.
790 		 */
791 		new_cpu = cc_migration_cpu(cc, direct);
792 		new_time = cc_migration_time(cc, direct);
793 		new_prec = cc_migration_prec(cc, direct);
794 		new_func = cc_migration_func(cc, direct);
795 		new_arg = cc_migration_arg(cc, direct);
796 		cc_cce_cleanup(cc, direct);
797 
798 		/*
799 		 * It should be assert here that the callout is not destroyed
800 		 * but that is not easy.
801 		 *
802 		 * As first thing, handle deferred callout stops.
803 		 */
804 		if (!callout_migrating(c)) {
805 			CTR3(KTR_CALLOUT,
806 			     "deferred cancelled %p func %p arg %p",
807 			     c, new_func, new_arg);
808 			callout_cc_del(c, cc);
809 			return;
810 		}
811 		c->c_iflags &= ~CALLOUT_DFRMIGRATION;
812 
813 		new_cc = callout_cpu_switch(c, cc, new_cpu);
814 		flags = (direct) ? C_DIRECT_EXEC : 0;
815 		callout_cc_add(c, new_cc, new_time, new_prec, new_func,
816 		    new_arg, new_cpu, flags);
817 		CC_UNLOCK(new_cc);
818 		CC_LOCK(cc);
819 #else
820 		panic("migration should not happen");
821 #endif
822 	}
823 	/*
824 	 * If the current callout is locally allocated (from
825 	 * timeout(9)) then put it on the freelist.
826 	 *
827 	 * Note: we need to check the cached copy of c_iflags because
828 	 * if it was not local, then it's not safe to deref the
829 	 * callout pointer.
830 	 */
831 	KASSERT((c_iflags & CALLOUT_LOCAL_ALLOC) == 0 ||
832 	    c->c_iflags == CALLOUT_LOCAL_ALLOC,
833 	    ("corrupted callout"));
834 	if (c_iflags & CALLOUT_LOCAL_ALLOC)
835 		callout_cc_del(c, cc);
836 }
837 
838 /*
839  * The callout mechanism is based on the work of Adam M. Costello and
840  * George Varghese, published in a technical report entitled "Redesigning
841  * the BSD Callout and Timer Facilities" and modified slightly for inclusion
842  * in FreeBSD by Justin T. Gibbs.  The original work on the data structures
843  * used in this implementation was published by G. Varghese and T. Lauck in
844  * the paper "Hashed and Hierarchical Timing Wheels: Data Structures for
845  * the Efficient Implementation of a Timer Facility" in the Proceedings of
846  * the 11th ACM Annual Symposium on Operating Systems Principles,
847  * Austin, Texas Nov 1987.
848  */
849 
850 /*
851  * Software (low priority) clock interrupt.
852  * Run periodic events from timeout queue.
853  */
854 void
855 softclock(void *arg)
856 {
857 	struct callout_cpu *cc;
858 	struct callout *c;
859 #ifdef CALLOUT_PROFILING
860 	int depth = 0, gcalls = 0, lockcalls = 0, mpcalls = 0;
861 #endif
862 
863 	cc = (struct callout_cpu *)arg;
864 	CC_LOCK(cc);
865 	while ((c = TAILQ_FIRST(&cc->cc_expireq)) != NULL) {
866 		TAILQ_REMOVE(&cc->cc_expireq, c, c_links.tqe);
867 		softclock_call_cc(c, cc,
868 #ifdef CALLOUT_PROFILING
869 		    &mpcalls, &lockcalls, &gcalls,
870 #endif
871 		    0);
872 #ifdef CALLOUT_PROFILING
873 		++depth;
874 #endif
875 	}
876 #ifdef CALLOUT_PROFILING
877 	avg_depth += (depth * 1000 - avg_depth) >> 8;
878 	avg_mpcalls += (mpcalls * 1000 - avg_mpcalls) >> 8;
879 	avg_lockcalls += (lockcalls * 1000 - avg_lockcalls) >> 8;
880 	avg_gcalls += (gcalls * 1000 - avg_gcalls) >> 8;
881 #endif
882 	CC_UNLOCK(cc);
883 }
884 
885 /*
886  * timeout --
887  *	Execute a function after a specified length of time.
888  *
889  * untimeout --
890  *	Cancel previous timeout function call.
891  *
892  * callout_handle_init --
893  *	Initialize a handle so that using it with untimeout is benign.
894  *
895  *	See AT&T BCI Driver Reference Manual for specification.  This
896  *	implementation differs from that one in that although an
897  *	identification value is returned from timeout, the original
898  *	arguments to timeout as well as the identifier are used to
899  *	identify entries for untimeout.
900  */
901 struct callout_handle
902 timeout(timeout_t *ftn, void *arg, int to_ticks)
903 {
904 	struct callout_cpu *cc;
905 	struct callout *new;
906 	struct callout_handle handle;
907 
908 	cc = CC_CPU(timeout_cpu);
909 	CC_LOCK(cc);
910 	/* Fill in the next free callout structure. */
911 	new = SLIST_FIRST(&cc->cc_callfree);
912 	if (new == NULL)
913 		/* XXX Attempt to malloc first */
914 		panic("timeout table full");
915 	SLIST_REMOVE_HEAD(&cc->cc_callfree, c_links.sle);
916 	callout_reset(new, to_ticks, ftn, arg);
917 	handle.callout = new;
918 	CC_UNLOCK(cc);
919 
920 	return (handle);
921 }
922 
923 void
924 untimeout(timeout_t *ftn, void *arg, struct callout_handle handle)
925 {
926 	struct callout_cpu *cc;
927 
928 	/*
929 	 * Check for a handle that was initialized
930 	 * by callout_handle_init, but never used
931 	 * for a real timeout.
932 	 */
933 	if (handle.callout == NULL)
934 		return;
935 
936 	cc = callout_lock(handle.callout);
937 	if (handle.callout->c_func == ftn && handle.callout->c_arg == arg)
938 		callout_stop(handle.callout);
939 	CC_UNLOCK(cc);
940 }
941 
942 void
943 callout_handle_init(struct callout_handle *handle)
944 {
945 	handle->callout = NULL;
946 }
947 
948 void
949 callout_when(sbintime_t sbt, sbintime_t precision, int flags,
950     sbintime_t *res, sbintime_t *prec_res)
951 {
952 	sbintime_t to_sbt, to_pr;
953 
954 	if ((flags & (C_ABSOLUTE | C_PRECALC)) != 0) {
955 		*res = sbt;
956 		*prec_res = precision;
957 		return;
958 	}
959 	if ((flags & C_HARDCLOCK) != 0 && sbt < tick_sbt)
960 		sbt = tick_sbt;
961 	if ((flags & C_HARDCLOCK) != 0 ||
962 #ifdef NO_EVENTTIMERS
963 	    sbt >= sbt_timethreshold) {
964 		to_sbt = getsbinuptime();
965 
966 		/* Add safety belt for the case of hz > 1000. */
967 		to_sbt += tc_tick_sbt - tick_sbt;
968 #else
969 	    sbt >= sbt_tickthreshold) {
970 		/*
971 		 * Obtain the time of the last hardclock() call on
972 		 * this CPU directly from the kern_clocksource.c.
973 		 * This value is per-CPU, but it is equal for all
974 		 * active ones.
975 		 */
976 #ifdef __LP64__
977 		to_sbt = DPCPU_GET(hardclocktime);
978 #else
979 		spinlock_enter();
980 		to_sbt = DPCPU_GET(hardclocktime);
981 		spinlock_exit();
982 #endif
983 #endif
984 		if (cold && to_sbt == 0)
985 			to_sbt = sbinuptime();
986 		if ((flags & C_HARDCLOCK) == 0)
987 			to_sbt += tick_sbt;
988 	} else
989 		to_sbt = sbinuptime();
990 	if (SBT_MAX - to_sbt < sbt)
991 		to_sbt = SBT_MAX;
992 	else
993 		to_sbt += sbt;
994 	*res = to_sbt;
995 	to_pr = ((C_PRELGET(flags) < 0) ? sbt >> tc_precexp :
996 	    sbt >> C_PRELGET(flags));
997 	*prec_res = to_pr > precision ? to_pr : precision;
998 }
999 
1000 /*
1001  * New interface; clients allocate their own callout structures.
1002  *
1003  * callout_reset() - establish or change a timeout
1004  * callout_stop() - disestablish a timeout
1005  * callout_init() - initialize a callout structure so that it can
1006  *	safely be passed to callout_reset() and callout_stop()
1007  *
1008  * <sys/callout.h> defines three convenience macros:
1009  *
1010  * callout_active() - returns truth if callout has not been stopped,
1011  *	drained, or deactivated since the last time the callout was
1012  *	reset.
1013  * callout_pending() - returns truth if callout is still waiting for timeout
1014  * callout_deactivate() - marks the callout as having been serviced
1015  */
1016 int
1017 callout_reset_sbt_on(struct callout *c, sbintime_t sbt, sbintime_t prec,
1018     void (*ftn)(void *), void *arg, int cpu, int flags)
1019 {
1020 	sbintime_t to_sbt, precision;
1021 	struct callout_cpu *cc;
1022 	int cancelled, direct;
1023 	int ignore_cpu=0;
1024 
1025 	cancelled = 0;
1026 	if (cpu == -1) {
1027 		ignore_cpu = 1;
1028 	} else if ((cpu >= MAXCPU) ||
1029 		   ((CC_CPU(cpu))->cc_inited == 0)) {
1030 		/* Invalid CPU spec */
1031 		panic("Invalid CPU in callout %d", cpu);
1032 	}
1033 	callout_when(sbt, prec, flags, &to_sbt, &precision);
1034 
1035 	/*
1036 	 * This flag used to be added by callout_cc_add, but the
1037 	 * first time you call this we could end up with the
1038 	 * wrong direct flag if we don't do it before we add.
1039 	 */
1040 	if (flags & C_DIRECT_EXEC) {
1041 		direct = 1;
1042 	} else {
1043 		direct = 0;
1044 	}
1045 	KASSERT(!direct || c->c_lock == NULL,
1046 	    ("%s: direct callout %p has lock", __func__, c));
1047 	cc = callout_lock(c);
1048 	/*
1049 	 * Don't allow migration of pre-allocated callouts lest they
1050 	 * become unbalanced or handle the case where the user does
1051 	 * not care.
1052 	 */
1053 	if ((c->c_iflags & CALLOUT_LOCAL_ALLOC) ||
1054 	    ignore_cpu) {
1055 		cpu = c->c_cpu;
1056 	}
1057 
1058 	if (cc_exec_curr(cc, direct) == c) {
1059 		/*
1060 		 * We're being asked to reschedule a callout which is
1061 		 * currently in progress.  If there is a lock then we
1062 		 * can cancel the callout if it has not really started.
1063 		 */
1064 		if (c->c_lock != NULL && !cc_exec_cancel(cc, direct))
1065 			cancelled = cc_exec_cancel(cc, direct) = true;
1066 		if (cc_exec_waiting(cc, direct) || cc_exec_drain(cc, direct)) {
1067 			/*
1068 			 * Someone has called callout_drain to kill this
1069 			 * callout.  Don't reschedule.
1070 			 */
1071 			CTR4(KTR_CALLOUT, "%s %p func %p arg %p",
1072 			    cancelled ? "cancelled" : "failed to cancel",
1073 			    c, c->c_func, c->c_arg);
1074 			CC_UNLOCK(cc);
1075 			return (cancelled);
1076 		}
1077 #ifdef SMP
1078 		if (callout_migrating(c)) {
1079 			/*
1080 			 * This only occurs when a second callout_reset_sbt_on
1081 			 * is made after a previous one moved it into
1082 			 * deferred migration (below). Note we do *not* change
1083 			 * the prev_cpu even though the previous target may
1084 			 * be different.
1085 			 */
1086 			cc_migration_cpu(cc, direct) = cpu;
1087 			cc_migration_time(cc, direct) = to_sbt;
1088 			cc_migration_prec(cc, direct) = precision;
1089 			cc_migration_func(cc, direct) = ftn;
1090 			cc_migration_arg(cc, direct) = arg;
1091 			cancelled = 1;
1092 			CC_UNLOCK(cc);
1093 			return (cancelled);
1094 		}
1095 #endif
1096 	}
1097 	if (c->c_iflags & CALLOUT_PENDING) {
1098 		if ((c->c_iflags & CALLOUT_PROCESSED) == 0) {
1099 			if (cc_exec_next(cc) == c)
1100 				cc_exec_next(cc) = LIST_NEXT(c, c_links.le);
1101 			LIST_REMOVE(c, c_links.le);
1102 		} else {
1103 			TAILQ_REMOVE(&cc->cc_expireq, c, c_links.tqe);
1104 		}
1105 		cancelled = 1;
1106 		c->c_iflags &= ~ CALLOUT_PENDING;
1107 		c->c_flags &= ~ CALLOUT_ACTIVE;
1108 	}
1109 
1110 #ifdef SMP
1111 	/*
1112 	 * If the callout must migrate try to perform it immediately.
1113 	 * If the callout is currently running, just defer the migration
1114 	 * to a more appropriate moment.
1115 	 */
1116 	if (c->c_cpu != cpu) {
1117 		if (cc_exec_curr(cc, direct) == c) {
1118 			/*
1119 			 * Pending will have been removed since we are
1120 			 * actually executing the callout on another
1121 			 * CPU. That callout should be waiting on the
1122 			 * lock the caller holds. If we set both
1123 			 * active/and/pending after we return and the
1124 			 * lock on the executing callout proceeds, it
1125 			 * will then see pending is true and return.
1126 			 * At the return from the actual callout execution
1127 			 * the migration will occur in softclock_call_cc
1128 			 * and this new callout will be placed on the
1129 			 * new CPU via a call to callout_cpu_switch() which
1130 			 * will get the lock on the right CPU followed
1131 			 * by a call callout_cc_add() which will add it there.
1132 			 * (see above in softclock_call_cc()).
1133 			 */
1134 			cc_migration_cpu(cc, direct) = cpu;
1135 			cc_migration_time(cc, direct) = to_sbt;
1136 			cc_migration_prec(cc, direct) = precision;
1137 			cc_migration_func(cc, direct) = ftn;
1138 			cc_migration_arg(cc, direct) = arg;
1139 			c->c_iflags |= (CALLOUT_DFRMIGRATION | CALLOUT_PENDING);
1140 			c->c_flags |= CALLOUT_ACTIVE;
1141 			CTR6(KTR_CALLOUT,
1142 		    "migration of %p func %p arg %p in %d.%08x to %u deferred",
1143 			    c, c->c_func, c->c_arg, (int)(to_sbt >> 32),
1144 			    (u_int)(to_sbt & 0xffffffff), cpu);
1145 			CC_UNLOCK(cc);
1146 			return (cancelled);
1147 		}
1148 		cc = callout_cpu_switch(c, cc, cpu);
1149 	}
1150 #endif
1151 
1152 	callout_cc_add(c, cc, to_sbt, precision, ftn, arg, cpu, flags);
1153 	CTR6(KTR_CALLOUT, "%sscheduled %p func %p arg %p in %d.%08x",
1154 	    cancelled ? "re" : "", c, c->c_func, c->c_arg, (int)(to_sbt >> 32),
1155 	    (u_int)(to_sbt & 0xffffffff));
1156 	CC_UNLOCK(cc);
1157 
1158 	return (cancelled);
1159 }
1160 
1161 /*
1162  * Common idioms that can be optimized in the future.
1163  */
1164 int
1165 callout_schedule_on(struct callout *c, int to_ticks, int cpu)
1166 {
1167 	return callout_reset_on(c, to_ticks, c->c_func, c->c_arg, cpu);
1168 }
1169 
1170 int
1171 callout_schedule(struct callout *c, int to_ticks)
1172 {
1173 	return callout_reset_on(c, to_ticks, c->c_func, c->c_arg, c->c_cpu);
1174 }
1175 
1176 int
1177 _callout_stop_safe(struct callout *c, int flags, void (*drain)(void *))
1178 {
1179 	struct callout_cpu *cc, *old_cc;
1180 	struct lock_class *class;
1181 	int direct, sq_locked, use_lock;
1182 	int cancelled, not_on_a_list;
1183 
1184 	if ((flags & CS_DRAIN) != 0)
1185 		WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK, c->c_lock,
1186 		    "calling %s", __func__);
1187 
1188 	/*
1189 	 * Some old subsystems don't hold Giant while running a callout_stop(),
1190 	 * so just discard this check for the moment.
1191 	 */
1192 	if ((flags & CS_DRAIN) == 0 && c->c_lock != NULL) {
1193 		if (c->c_lock == &Giant.lock_object)
1194 			use_lock = mtx_owned(&Giant);
1195 		else {
1196 			use_lock = 1;
1197 			class = LOCK_CLASS(c->c_lock);
1198 			class->lc_assert(c->c_lock, LA_XLOCKED);
1199 		}
1200 	} else
1201 		use_lock = 0;
1202 	if (c->c_iflags & CALLOUT_DIRECT) {
1203 		direct = 1;
1204 	} else {
1205 		direct = 0;
1206 	}
1207 	sq_locked = 0;
1208 	old_cc = NULL;
1209 again:
1210 	cc = callout_lock(c);
1211 
1212 	if ((c->c_iflags & (CALLOUT_DFRMIGRATION | CALLOUT_PENDING)) ==
1213 	    (CALLOUT_DFRMIGRATION | CALLOUT_PENDING) &&
1214 	    ((c->c_flags & CALLOUT_ACTIVE) == CALLOUT_ACTIVE)) {
1215 		/*
1216 		 * Special case where this slipped in while we
1217 		 * were migrating *as* the callout is about to
1218 		 * execute. The caller probably holds the lock
1219 		 * the callout wants.
1220 		 *
1221 		 * Get rid of the migration first. Then set
1222 		 * the flag that tells this code *not* to
1223 		 * try to remove it from any lists (its not
1224 		 * on one yet). When the callout wheel runs,
1225 		 * it will ignore this callout.
1226 		 */
1227 		c->c_iflags &= ~CALLOUT_PENDING;
1228 		c->c_flags &= ~CALLOUT_ACTIVE;
1229 		not_on_a_list = 1;
1230 	} else {
1231 		not_on_a_list = 0;
1232 	}
1233 
1234 	/*
1235 	 * If the callout was migrating while the callout cpu lock was
1236 	 * dropped,  just drop the sleepqueue lock and check the states
1237 	 * again.
1238 	 */
1239 	if (sq_locked != 0 && cc != old_cc) {
1240 #ifdef SMP
1241 		CC_UNLOCK(cc);
1242 		sleepq_release(&cc_exec_waiting(old_cc, direct));
1243 		sq_locked = 0;
1244 		old_cc = NULL;
1245 		goto again;
1246 #else
1247 		panic("migration should not happen");
1248 #endif
1249 	}
1250 
1251 	/*
1252 	 * If the callout is running, try to stop it or drain it.
1253 	 */
1254 	if (cc_exec_curr(cc, direct) == c) {
1255 		/*
1256 		 * Succeed we to stop it or not, we must clear the
1257 		 * active flag - this is what API users expect.  If we're
1258 		 * draining and the callout is currently executing, first wait
1259 		 * until it finishes.
1260 		 */
1261 		if ((flags & CS_DRAIN) == 0)
1262 			c->c_flags &= ~CALLOUT_ACTIVE;
1263 
1264 		if ((flags & CS_DRAIN) != 0) {
1265 			/*
1266 			 * The current callout is running (or just
1267 			 * about to run) and blocking is allowed, so
1268 			 * just wait for the current invocation to
1269 			 * finish.
1270 			 */
1271 			while (cc_exec_curr(cc, direct) == c) {
1272 				/*
1273 				 * Use direct calls to sleepqueue interface
1274 				 * instead of cv/msleep in order to avoid
1275 				 * a LOR between cc_lock and sleepqueue
1276 				 * chain spinlocks.  This piece of code
1277 				 * emulates a msleep_spin() call actually.
1278 				 *
1279 				 * If we already have the sleepqueue chain
1280 				 * locked, then we can safely block.  If we
1281 				 * don't already have it locked, however,
1282 				 * we have to drop the cc_lock to lock
1283 				 * it.  This opens several races, so we
1284 				 * restart at the beginning once we have
1285 				 * both locks.  If nothing has changed, then
1286 				 * we will end up back here with sq_locked
1287 				 * set.
1288 				 */
1289 				if (!sq_locked) {
1290 					CC_UNLOCK(cc);
1291 					sleepq_lock(
1292 					    &cc_exec_waiting(cc, direct));
1293 					sq_locked = 1;
1294 					old_cc = cc;
1295 					goto again;
1296 				}
1297 
1298 				/*
1299 				 * Migration could be cancelled here, but
1300 				 * as long as it is still not sure when it
1301 				 * will be packed up, just let softclock()
1302 				 * take care of it.
1303 				 */
1304 				cc_exec_waiting(cc, direct) = true;
1305 				DROP_GIANT();
1306 				CC_UNLOCK(cc);
1307 				sleepq_add(
1308 				    &cc_exec_waiting(cc, direct),
1309 				    &cc->cc_lock.lock_object, "codrain",
1310 				    SLEEPQ_SLEEP, 0);
1311 				sleepq_wait(
1312 				    &cc_exec_waiting(cc, direct),
1313 					     0);
1314 				sq_locked = 0;
1315 				old_cc = NULL;
1316 
1317 				/* Reacquire locks previously released. */
1318 				PICKUP_GIANT();
1319 				CC_LOCK(cc);
1320 			}
1321 			c->c_flags &= ~CALLOUT_ACTIVE;
1322 		} else if (use_lock &&
1323 			   !cc_exec_cancel(cc, direct) && (drain == NULL)) {
1324 
1325 			/*
1326 			 * The current callout is waiting for its
1327 			 * lock which we hold.  Cancel the callout
1328 			 * and return.  After our caller drops the
1329 			 * lock, the callout will be skipped in
1330 			 * softclock(). This *only* works with a
1331 			 * callout_stop() *not* callout_drain() or
1332 			 * callout_async_drain().
1333 			 */
1334 			cc_exec_cancel(cc, direct) = true;
1335 			CTR3(KTR_CALLOUT, "cancelled %p func %p arg %p",
1336 			    c, c->c_func, c->c_arg);
1337 			KASSERT(!cc_cce_migrating(cc, direct),
1338 			    ("callout wrongly scheduled for migration"));
1339 			if (callout_migrating(c)) {
1340 				c->c_iflags &= ~CALLOUT_DFRMIGRATION;
1341 #ifdef SMP
1342 				cc_migration_cpu(cc, direct) = CPUBLOCK;
1343 				cc_migration_time(cc, direct) = 0;
1344 				cc_migration_prec(cc, direct) = 0;
1345 				cc_migration_func(cc, direct) = NULL;
1346 				cc_migration_arg(cc, direct) = NULL;
1347 #endif
1348 			}
1349 			CC_UNLOCK(cc);
1350 			KASSERT(!sq_locked, ("sleepqueue chain locked"));
1351 			return (1);
1352 		} else if (callout_migrating(c)) {
1353 			/*
1354 			 * The callout is currently being serviced
1355 			 * and the "next" callout is scheduled at
1356 			 * its completion with a migration. We remove
1357 			 * the migration flag so it *won't* get rescheduled,
1358 			 * but we can't stop the one thats running so
1359 			 * we return 0.
1360 			 */
1361 			c->c_iflags &= ~CALLOUT_DFRMIGRATION;
1362 #ifdef SMP
1363 			/*
1364 			 * We can't call cc_cce_cleanup here since
1365 			 * if we do it will remove .ce_curr and
1366 			 * its still running. This will prevent a
1367 			 * reschedule of the callout when the
1368 			 * execution completes.
1369 			 */
1370 			cc_migration_cpu(cc, direct) = CPUBLOCK;
1371 			cc_migration_time(cc, direct) = 0;
1372 			cc_migration_prec(cc, direct) = 0;
1373 			cc_migration_func(cc, direct) = NULL;
1374 			cc_migration_arg(cc, direct) = NULL;
1375 #endif
1376 			CTR3(KTR_CALLOUT, "postponing stop %p func %p arg %p",
1377 			    c, c->c_func, c->c_arg);
1378  			if (drain) {
1379 				cc_exec_drain(cc, direct) = drain;
1380 			}
1381 			CC_UNLOCK(cc);
1382 			return ((flags & CS_EXECUTING) != 0);
1383 		}
1384 		CTR3(KTR_CALLOUT, "failed to stop %p func %p arg %p",
1385 		    c, c->c_func, c->c_arg);
1386 		if (drain) {
1387 			cc_exec_drain(cc, direct) = drain;
1388 		}
1389 		KASSERT(!sq_locked, ("sleepqueue chain still locked"));
1390 		cancelled = ((flags & CS_EXECUTING) != 0);
1391 	} else
1392 		cancelled = 1;
1393 
1394 	if (sq_locked)
1395 		sleepq_release(&cc_exec_waiting(cc, direct));
1396 
1397 	if ((c->c_iflags & CALLOUT_PENDING) == 0) {
1398 		CTR3(KTR_CALLOUT, "failed to stop %p func %p arg %p",
1399 		    c, c->c_func, c->c_arg);
1400 		/*
1401 		 * For not scheduled and not executing callout return
1402 		 * negative value.
1403 		 */
1404 		if (cc_exec_curr(cc, direct) != c)
1405 			cancelled = -1;
1406 		CC_UNLOCK(cc);
1407 		return (cancelled);
1408 	}
1409 
1410 	c->c_iflags &= ~CALLOUT_PENDING;
1411 	c->c_flags &= ~CALLOUT_ACTIVE;
1412 
1413 	CTR3(KTR_CALLOUT, "cancelled %p func %p arg %p",
1414 	    c, c->c_func, c->c_arg);
1415 	if (not_on_a_list == 0) {
1416 		if ((c->c_iflags & CALLOUT_PROCESSED) == 0) {
1417 			if (cc_exec_next(cc) == c)
1418 				cc_exec_next(cc) = LIST_NEXT(c, c_links.le);
1419 			LIST_REMOVE(c, c_links.le);
1420 		} else {
1421 			TAILQ_REMOVE(&cc->cc_expireq, c, c_links.tqe);
1422 		}
1423 	}
1424 	callout_cc_del(c, cc);
1425 	CC_UNLOCK(cc);
1426 	return (cancelled);
1427 }
1428 
1429 void
1430 callout_init(struct callout *c, int mpsafe)
1431 {
1432 	bzero(c, sizeof *c);
1433 	if (mpsafe) {
1434 		c->c_lock = NULL;
1435 		c->c_iflags = CALLOUT_RETURNUNLOCKED;
1436 	} else {
1437 		c->c_lock = &Giant.lock_object;
1438 		c->c_iflags = 0;
1439 	}
1440 	c->c_cpu = timeout_cpu;
1441 }
1442 
1443 void
1444 _callout_init_lock(struct callout *c, struct lock_object *lock, int flags)
1445 {
1446 	bzero(c, sizeof *c);
1447 	c->c_lock = lock;
1448 	KASSERT((flags & ~(CALLOUT_RETURNUNLOCKED | CALLOUT_SHAREDLOCK)) == 0,
1449 	    ("callout_init_lock: bad flags %d", flags));
1450 	KASSERT(lock != NULL || (flags & CALLOUT_RETURNUNLOCKED) == 0,
1451 	    ("callout_init_lock: CALLOUT_RETURNUNLOCKED with no lock"));
1452 	KASSERT(lock == NULL || !(LOCK_CLASS(lock)->lc_flags &
1453 	    (LC_SPINLOCK | LC_SLEEPABLE)), ("%s: invalid lock class",
1454 	    __func__));
1455 	c->c_iflags = flags & (CALLOUT_RETURNUNLOCKED | CALLOUT_SHAREDLOCK);
1456 	c->c_cpu = timeout_cpu;
1457 }
1458 
1459 #ifdef APM_FIXUP_CALLTODO
1460 /*
1461  * Adjust the kernel calltodo timeout list.  This routine is used after
1462  * an APM resume to recalculate the calltodo timer list values with the
1463  * number of hz's we have been sleeping.  The next hardclock() will detect
1464  * that there are fired timers and run softclock() to execute them.
1465  *
1466  * Please note, I have not done an exhaustive analysis of what code this
1467  * might break.  I am motivated to have my select()'s and alarm()'s that
1468  * have expired during suspend firing upon resume so that the applications
1469  * which set the timer can do the maintanence the timer was for as close
1470  * as possible to the originally intended time.  Testing this code for a
1471  * week showed that resuming from a suspend resulted in 22 to 25 timers
1472  * firing, which seemed independent on whether the suspend was 2 hours or
1473  * 2 days.  Your milage may vary.   - Ken Key <key@cs.utk.edu>
1474  */
1475 void
1476 adjust_timeout_calltodo(struct timeval *time_change)
1477 {
1478 	struct callout *p;
1479 	unsigned long delta_ticks;
1480 
1481 	/*
1482 	 * How many ticks were we asleep?
1483 	 * (stolen from tvtohz()).
1484 	 */
1485 
1486 	/* Don't do anything */
1487 	if (time_change->tv_sec < 0)
1488 		return;
1489 	else if (time_change->tv_sec <= LONG_MAX / 1000000)
1490 		delta_ticks = howmany(time_change->tv_sec * 1000000 +
1491 		    time_change->tv_usec, tick) + 1;
1492 	else if (time_change->tv_sec <= LONG_MAX / hz)
1493 		delta_ticks = time_change->tv_sec * hz +
1494 		    howmany(time_change->tv_usec, tick) + 1;
1495 	else
1496 		delta_ticks = LONG_MAX;
1497 
1498 	if (delta_ticks > INT_MAX)
1499 		delta_ticks = INT_MAX;
1500 
1501 	/*
1502 	 * Now rip through the timer calltodo list looking for timers
1503 	 * to expire.
1504 	 */
1505 
1506 	/* don't collide with softclock() */
1507 	CC_LOCK(cc);
1508 	for (p = calltodo.c_next; p != NULL; p = p->c_next) {
1509 		p->c_time -= delta_ticks;
1510 
1511 		/* Break if the timer had more time on it than delta_ticks */
1512 		if (p->c_time > 0)
1513 			break;
1514 
1515 		/* take back the ticks the timer didn't use (p->c_time <= 0) */
1516 		delta_ticks = -p->c_time;
1517 	}
1518 	CC_UNLOCK(cc);
1519 
1520 	return;
1521 }
1522 #endif /* APM_FIXUP_CALLTODO */
1523 
1524 static int
1525 flssbt(sbintime_t sbt)
1526 {
1527 
1528 	sbt += (uint64_t)sbt >> 1;
1529 	if (sizeof(long) >= sizeof(sbintime_t))
1530 		return (flsl(sbt));
1531 	if (sbt >= SBT_1S)
1532 		return (flsl(((uint64_t)sbt) >> 32) + 32);
1533 	return (flsl(sbt));
1534 }
1535 
1536 /*
1537  * Dump immediate statistic snapshot of the scheduled callouts.
1538  */
1539 static int
1540 sysctl_kern_callout_stat(SYSCTL_HANDLER_ARGS)
1541 {
1542 	struct callout *tmp;
1543 	struct callout_cpu *cc;
1544 	struct callout_list *sc;
1545 	sbintime_t maxpr, maxt, medpr, medt, now, spr, st, t;
1546 	int ct[64], cpr[64], ccpbk[32];
1547 	int error, val, i, count, tcum, pcum, maxc, c, medc;
1548 #ifdef SMP
1549 	int cpu;
1550 #endif
1551 
1552 	val = 0;
1553 	error = sysctl_handle_int(oidp, &val, 0, req);
1554 	if (error != 0 || req->newptr == NULL)
1555 		return (error);
1556 	count = maxc = 0;
1557 	st = spr = maxt = maxpr = 0;
1558 	bzero(ccpbk, sizeof(ccpbk));
1559 	bzero(ct, sizeof(ct));
1560 	bzero(cpr, sizeof(cpr));
1561 	now = sbinuptime();
1562 #ifdef SMP
1563 	CPU_FOREACH(cpu) {
1564 		cc = CC_CPU(cpu);
1565 #else
1566 		cc = CC_CPU(timeout_cpu);
1567 #endif
1568 		CC_LOCK(cc);
1569 		for (i = 0; i < callwheelsize; i++) {
1570 			sc = &cc->cc_callwheel[i];
1571 			c = 0;
1572 			LIST_FOREACH(tmp, sc, c_links.le) {
1573 				c++;
1574 				t = tmp->c_time - now;
1575 				if (t < 0)
1576 					t = 0;
1577 				st += t / SBT_1US;
1578 				spr += tmp->c_precision / SBT_1US;
1579 				if (t > maxt)
1580 					maxt = t;
1581 				if (tmp->c_precision > maxpr)
1582 					maxpr = tmp->c_precision;
1583 				ct[flssbt(t)]++;
1584 				cpr[flssbt(tmp->c_precision)]++;
1585 			}
1586 			if (c > maxc)
1587 				maxc = c;
1588 			ccpbk[fls(c + c / 2)]++;
1589 			count += c;
1590 		}
1591 		CC_UNLOCK(cc);
1592 #ifdef SMP
1593 	}
1594 #endif
1595 
1596 	for (i = 0, tcum = 0; i < 64 && tcum < count / 2; i++)
1597 		tcum += ct[i];
1598 	medt = (i >= 2) ? (((sbintime_t)1) << (i - 2)) : 0;
1599 	for (i = 0, pcum = 0; i < 64 && pcum < count / 2; i++)
1600 		pcum += cpr[i];
1601 	medpr = (i >= 2) ? (((sbintime_t)1) << (i - 2)) : 0;
1602 	for (i = 0, c = 0; i < 32 && c < count / 2; i++)
1603 		c += ccpbk[i];
1604 	medc = (i >= 2) ? (1 << (i - 2)) : 0;
1605 
1606 	printf("Scheduled callouts statistic snapshot:\n");
1607 	printf("  Callouts: %6d  Buckets: %6d*%-3d  Bucket size: 0.%06ds\n",
1608 	    count, callwheelsize, mp_ncpus, 1000000 >> CC_HASH_SHIFT);
1609 	printf("  C/Bk: med %5d         avg %6d.%06jd  max %6d\n",
1610 	    medc,
1611 	    count / callwheelsize / mp_ncpus,
1612 	    (uint64_t)count * 1000000 / callwheelsize / mp_ncpus % 1000000,
1613 	    maxc);
1614 	printf("  Time: med %5jd.%06jds avg %6jd.%06jds max %6jd.%06jds\n",
1615 	    medt / SBT_1S, (medt & 0xffffffff) * 1000000 >> 32,
1616 	    (st / count) / 1000000, (st / count) % 1000000,
1617 	    maxt / SBT_1S, (maxt & 0xffffffff) * 1000000 >> 32);
1618 	printf("  Prec: med %5jd.%06jds avg %6jd.%06jds max %6jd.%06jds\n",
1619 	    medpr / SBT_1S, (medpr & 0xffffffff) * 1000000 >> 32,
1620 	    (spr / count) / 1000000, (spr / count) % 1000000,
1621 	    maxpr / SBT_1S, (maxpr & 0xffffffff) * 1000000 >> 32);
1622 	printf("  Distribution:       \tbuckets\t   time\t   tcum\t"
1623 	    "   prec\t   pcum\n");
1624 	for (i = 0, tcum = pcum = 0; i < 64; i++) {
1625 		if (ct[i] == 0 && cpr[i] == 0)
1626 			continue;
1627 		t = (i != 0) ? (((sbintime_t)1) << (i - 1)) : 0;
1628 		tcum += ct[i];
1629 		pcum += cpr[i];
1630 		printf("  %10jd.%06jds\t 2**%d\t%7d\t%7d\t%7d\t%7d\n",
1631 		    t / SBT_1S, (t & 0xffffffff) * 1000000 >> 32,
1632 		    i - 1 - (32 - CC_HASH_SHIFT),
1633 		    ct[i], tcum, cpr[i], pcum);
1634 	}
1635 	return (error);
1636 }
1637 SYSCTL_PROC(_kern, OID_AUTO, callout_stat,
1638     CTLTYPE_INT | CTLFLAG_RW | CTLFLAG_MPSAFE,
1639     0, 0, sysctl_kern_callout_stat, "I",
1640     "Dump immediate statistic snapshot of the scheduled callouts");
1641 
1642 #ifdef DDB
1643 static void
1644 _show_callout(struct callout *c)
1645 {
1646 
1647 	db_printf("callout %p\n", c);
1648 #define	C_DB_PRINTF(f, e)	db_printf("   %s = " f "\n", #e, c->e);
1649 	db_printf("   &c_links = %p\n", &(c->c_links));
1650 	C_DB_PRINTF("%" PRId64,	c_time);
1651 	C_DB_PRINTF("%" PRId64,	c_precision);
1652 	C_DB_PRINTF("%p",	c_arg);
1653 	C_DB_PRINTF("%p",	c_func);
1654 	C_DB_PRINTF("%p",	c_lock);
1655 	C_DB_PRINTF("%#x",	c_flags);
1656 	C_DB_PRINTF("%#x",	c_iflags);
1657 	C_DB_PRINTF("%d",	c_cpu);
1658 #undef	C_DB_PRINTF
1659 }
1660 
1661 DB_SHOW_COMMAND(callout, db_show_callout)
1662 {
1663 
1664 	if (!have_addr) {
1665 		db_printf("usage: show callout <struct callout *>\n");
1666 		return;
1667 	}
1668 
1669 	_show_callout((struct callout *)addr);
1670 }
1671 #endif /* DDB */
1672