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.
10 * Redistribution and use in source and binary forms, with or without
11 * modification, are permitted provided that the following conditions
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 * 4. 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.
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
34 * From: @(#)kern_clock.c 8.5 (Berkeley) 1/21/94
37 #include <sys/cdefs.h>
38 __FBSDID("$FreeBSD$");
40 #include "opt_callout_profiling.h"
41 #include "opt_kdtrace.h"
43 #include "opt_timer.h"
46 #include <sys/param.h>
47 #include <sys/systm.h>
49 #include <sys/callout.h>
51 #include <sys/interrupt.h>
52 #include <sys/kernel.h>
55 #include <sys/malloc.h>
56 #include <sys/mutex.h>
59 #include <sys/sleepqueue.h>
60 #include <sys/sysctl.h>
64 #include <machine/cpu.h>
67 #ifndef NO_EVENTTIMERS
68 DPCPU_DECLARE(sbintime_t, hardclocktime);
71 SDT_PROVIDER_DEFINE(callout_execute);
72 SDT_PROBE_DEFINE1(callout_execute, kernel, , callout_start, callout-start,
74 SDT_PROBE_DEFINE1(callout_execute, kernel, , callout_end, callout-end,
77 #ifdef CALLOUT_PROFILING
79 SYSCTL_INT(_debug, OID_AUTO, to_avg_depth, CTLFLAG_RD, &avg_depth, 0,
80 "Average number of items examined per softclock call. Units = 1/1000");
81 static int avg_gcalls;
82 SYSCTL_INT(_debug, OID_AUTO, to_avg_gcalls, CTLFLAG_RD, &avg_gcalls, 0,
83 "Average number of Giant callouts made per softclock call. Units = 1/1000");
84 static int avg_lockcalls;
85 SYSCTL_INT(_debug, OID_AUTO, to_avg_lockcalls, CTLFLAG_RD, &avg_lockcalls, 0,
86 "Average number of lock callouts made per softclock call. Units = 1/1000");
87 static int avg_mpcalls;
88 SYSCTL_INT(_debug, OID_AUTO, to_avg_mpcalls, CTLFLAG_RD, &avg_mpcalls, 0,
89 "Average number of MP callouts made per softclock call. Units = 1/1000");
90 static int avg_depth_dir;
91 SYSCTL_INT(_debug, OID_AUTO, to_avg_depth_dir, CTLFLAG_RD, &avg_depth_dir, 0,
92 "Average number of direct callouts examined per callout_process call. "
94 static int avg_lockcalls_dir;
95 SYSCTL_INT(_debug, OID_AUTO, to_avg_lockcalls_dir, CTLFLAG_RD,
96 &avg_lockcalls_dir, 0, "Average number of lock direct callouts made per "
97 "callout_process call. Units = 1/1000");
98 static int avg_mpcalls_dir;
99 SYSCTL_INT(_debug, OID_AUTO, to_avg_mpcalls_dir, CTLFLAG_RD, &avg_mpcalls_dir,
100 0, "Average number of MP direct callouts made per callout_process call. "
105 SYSCTL_INT(_kern, OID_AUTO, ncallout, CTLFLAG_RDTUN, &ncallout, 0,
106 "Number of entries in callwheel and size of timeout() preallocation");
110 * allocate more timeout table slots when table overflows.
112 u_int callwheelsize, callwheelmask;
115 * The callout cpu exec entities represent informations necessary for
116 * describing the state of callouts currently running on the CPU and the ones
117 * necessary for migrating callouts to the new callout cpu. In particular,
118 * the first entry of the array cc_exec_entity holds informations for callout
119 * running in SWI thread context, while the second one holds informations
120 * for callout running directly from hardware interrupt context.
121 * The cached informations are very important for deferring migration when
122 * the migrating callout is already running.
125 struct callout *cc_next;
126 struct callout *cc_curr;
128 void (*ce_migration_func)(void *);
129 void *ce_migration_arg;
130 int ce_migration_cpu;
131 sbintime_t ce_migration_time;
132 sbintime_t ce_migration_prec;
139 * There is one struct callout_cpu per cpu, holding all relevant
140 * state for the callout processing thread on the individual CPU.
143 struct mtx_padalign cc_lock;
144 struct cc_exec cc_exec_entity[2];
145 struct callout *cc_callout;
146 struct callout_list *cc_callwheel;
147 struct callout_tailq cc_expireq;
148 struct callout_slist cc_callfree;
149 sbintime_t cc_firstevent;
150 sbintime_t cc_lastscan;
155 #define cc_exec_curr cc_exec_entity[0].cc_curr
156 #define cc_exec_next cc_exec_entity[0].cc_next
157 #define cc_exec_cancel cc_exec_entity[0].cc_cancel
158 #define cc_exec_waiting cc_exec_entity[0].cc_waiting
159 #define cc_exec_curr_dir cc_exec_entity[1].cc_curr
160 #define cc_exec_next_dir cc_exec_entity[1].cc_next
161 #define cc_exec_cancel_dir cc_exec_entity[1].cc_cancel
162 #define cc_exec_waiting_dir cc_exec_entity[1].cc_waiting
165 #define cc_migration_func cc_exec_entity[0].ce_migration_func
166 #define cc_migration_arg cc_exec_entity[0].ce_migration_arg
167 #define cc_migration_cpu cc_exec_entity[0].ce_migration_cpu
168 #define cc_migration_time cc_exec_entity[0].ce_migration_time
169 #define cc_migration_prec cc_exec_entity[0].ce_migration_prec
170 #define cc_migration_func_dir cc_exec_entity[1].ce_migration_func
171 #define cc_migration_arg_dir cc_exec_entity[1].ce_migration_arg
172 #define cc_migration_cpu_dir cc_exec_entity[1].ce_migration_cpu
173 #define cc_migration_time_dir cc_exec_entity[1].ce_migration_time
174 #define cc_migration_prec_dir cc_exec_entity[1].ce_migration_prec
176 struct callout_cpu cc_cpu[MAXCPU];
177 #define CPUBLOCK MAXCPU
178 #define CC_CPU(cpu) (&cc_cpu[(cpu)])
179 #define CC_SELF() CC_CPU(PCPU_GET(cpuid))
181 struct callout_cpu cc_cpu;
182 #define CC_CPU(cpu) &cc_cpu
183 #define CC_SELF() &cc_cpu
185 #define CC_LOCK(cc) mtx_lock_spin(&(cc)->cc_lock)
186 #define CC_UNLOCK(cc) mtx_unlock_spin(&(cc)->cc_lock)
187 #define CC_LOCK_ASSERT(cc) mtx_assert(&(cc)->cc_lock, MA_OWNED)
189 static int timeout_cpu;
191 static void callout_cpu_init(struct callout_cpu *cc);
192 static void softclock_call_cc(struct callout *c, struct callout_cpu *cc,
193 #ifdef CALLOUT_PROFILING
194 int *mpcalls, int *lockcalls, int *gcalls,
198 static MALLOC_DEFINE(M_CALLOUT, "callout", "Callout datastructures");
202 * cc_curr - If a callout is in progress, it is cc_curr.
203 * If cc_curr is non-NULL, threads waiting in
204 * callout_drain() will be woken up as soon as the
205 * relevant callout completes.
206 * cc_cancel - Changing to 1 with both callout_lock and cc_lock held
207 * guarantees that the current callout will not run.
208 * The softclock() function sets this to 0 before it
209 * drops callout_lock to acquire c_lock, and it calls
210 * the handler only if curr_cancelled is still 0 after
211 * cc_lock is successfully acquired.
212 * cc_waiting - If a thread is waiting in callout_drain(), then
213 * callout_wait is nonzero. Set only when
214 * cc_curr is non-NULL.
218 * Resets the execution entity tied to a specific callout cpu.
221 cc_cce_cleanup(struct callout_cpu *cc, int direct)
224 cc->cc_exec_entity[direct].cc_curr = NULL;
225 cc->cc_exec_entity[direct].cc_next = NULL;
226 cc->cc_exec_entity[direct].cc_cancel = false;
227 cc->cc_exec_entity[direct].cc_waiting = false;
229 cc->cc_exec_entity[direct].ce_migration_cpu = CPUBLOCK;
230 cc->cc_exec_entity[direct].ce_migration_time = 0;
231 cc->cc_exec_entity[direct].ce_migration_prec = 0;
232 cc->cc_exec_entity[direct].ce_migration_func = NULL;
233 cc->cc_exec_entity[direct].ce_migration_arg = NULL;
238 * Checks if migration is requested by a specific callout cpu.
241 cc_cce_migrating(struct callout_cpu *cc, int direct)
245 return (cc->cc_exec_entity[direct].ce_migration_cpu != CPUBLOCK);
252 * Kernel low level callwheel initialization
253 * called on cpu0 during kernel startup.
256 callout_callwheel_init(void *dummy)
258 struct callout_cpu *cc;
261 * Calculate the size of the callout wheel and the preallocated
262 * timeout() structures.
263 * XXX: Clip callout to result of previous function of maxusers
264 * maximum 384. This is still huge, but acceptable.
266 ncallout = imin(16 + maxproc + maxfiles, 18508);
267 TUNABLE_INT_FETCH("kern.ncallout", &ncallout);
270 * Calculate callout wheel size, should be next power of two higher
273 callwheelsize = 1 << fls(ncallout);
274 callwheelmask = callwheelsize - 1;
277 * Only cpu0 handles timeout(9) and receives a preallocation.
279 * XXX: Once all timeout(9) consumers are converted this can
282 timeout_cpu = PCPU_GET(cpuid);
283 cc = CC_CPU(timeout_cpu);
284 cc->cc_callout = malloc(ncallout * sizeof(struct callout),
285 M_CALLOUT, M_WAITOK);
286 callout_cpu_init(cc);
288 SYSINIT(callwheel_init, SI_SUB_CPU, SI_ORDER_ANY, callout_callwheel_init, NULL);
291 * Initialize the per-cpu callout structures.
294 callout_cpu_init(struct callout_cpu *cc)
299 mtx_init(&cc->cc_lock, "callout", NULL, MTX_SPIN | MTX_RECURSE);
300 SLIST_INIT(&cc->cc_callfree);
301 cc->cc_callwheel = malloc(sizeof(struct callout_list) * callwheelsize,
302 M_CALLOUT, M_WAITOK);
303 for (i = 0; i < callwheelsize; i++)
304 LIST_INIT(&cc->cc_callwheel[i]);
305 TAILQ_INIT(&cc->cc_expireq);
306 cc->cc_firstevent = INT64_MAX;
307 for (i = 0; i < 2; i++)
308 cc_cce_cleanup(cc, i);
309 if (cc->cc_callout == NULL) /* Only cpu0 handles timeout(9) */
311 for (i = 0; i < ncallout; i++) {
312 c = &cc->cc_callout[i];
314 c->c_flags = CALLOUT_LOCAL_ALLOC;
315 SLIST_INSERT_HEAD(&cc->cc_callfree, c, c_links.sle);
321 * Switches the cpu tied to a specific callout.
322 * The function expects a locked incoming callout cpu and returns with
323 * locked outcoming callout cpu.
325 static struct callout_cpu *
326 callout_cpu_switch(struct callout *c, struct callout_cpu *cc, int new_cpu)
328 struct callout_cpu *new_cc;
330 MPASS(c != NULL && cc != NULL);
334 * Avoid interrupts and preemption firing after the callout cpu
335 * is blocked in order to avoid deadlocks as the new thread
336 * may be willing to acquire the callout cpu lock.
341 new_cc = CC_CPU(new_cpu);
350 * Start standard softclock thread.
353 start_softclock(void *dummy)
355 struct callout_cpu *cc;
360 cc = CC_CPU(timeout_cpu);
361 if (swi_add(&clk_intr_event, "clock", softclock, cc, SWI_CLOCK,
362 INTR_MPSAFE, &cc->cc_cookie))
363 panic("died while creating standard software ithreads");
366 if (cpu == timeout_cpu)
369 cc->cc_callout = NULL; /* Only cpu0 handles timeout(9). */
370 callout_cpu_init(cc);
371 if (swi_add(NULL, "clock", softclock, cc, SWI_CLOCK,
372 INTR_MPSAFE, &cc->cc_cookie))
373 panic("died while creating standard software ithreads");
377 SYSINIT(start_softclock, SI_SUB_SOFTINTR, SI_ORDER_FIRST, start_softclock, NULL);
379 #define CC_HASH_SHIFT 8
382 callout_hash(sbintime_t sbt)
385 return (sbt >> (32 - CC_HASH_SHIFT));
389 callout_get_bucket(sbintime_t sbt)
392 return (callout_hash(sbt) & callwheelmask);
396 callout_process(sbintime_t now)
398 struct callout *tmp, *tmpn;
399 struct callout_cpu *cc;
400 struct callout_list *sc;
401 sbintime_t first, last, max, tmp_max;
403 u_int firstb, lastb, nowb;
404 #ifdef CALLOUT_PROFILING
405 int depth_dir = 0, mpcalls_dir = 0, lockcalls_dir = 0;
409 mtx_lock_spin_flags(&cc->cc_lock, MTX_QUIET);
411 /* Compute the buckets of the last scan and present times. */
412 firstb = callout_hash(cc->cc_lastscan);
413 cc->cc_lastscan = now;
414 nowb = callout_hash(now);
416 /* Compute the last bucket and minimum time of the bucket after it. */
418 lookahead = (SBT_1S / 16);
419 else if (nowb - firstb == 1)
420 lookahead = (SBT_1S / 8);
422 lookahead = (SBT_1S / 2);
424 first += (lookahead / 2);
426 last &= (0xffffffffffffffffLLU << (32 - CC_HASH_SHIFT));
427 lastb = callout_hash(last) - 1;
431 * Check if we wrapped around the entire wheel from the last scan.
432 * In case, we need to scan entirely the wheel for pending callouts.
434 if (lastb - firstb >= callwheelsize) {
435 lastb = firstb + callwheelsize - 1;
436 if (nowb - firstb >= callwheelsize)
440 /* Iterate callwheel from firstb to nowb and then up to lastb. */
442 sc = &cc->cc_callwheel[firstb & callwheelmask];
443 tmp = LIST_FIRST(sc);
444 while (tmp != NULL) {
445 /* Run the callout if present time within allowed. */
446 if (tmp->c_time <= now) {
448 * Consumer told us the callout may be run
449 * directly from hardware interrupt context.
451 if (tmp->c_flags & CALLOUT_DIRECT) {
452 #ifdef CALLOUT_PROFILING
455 cc->cc_exec_next_dir =
456 LIST_NEXT(tmp, c_links.le);
457 cc->cc_bucket = firstb & callwheelmask;
458 LIST_REMOVE(tmp, c_links.le);
459 softclock_call_cc(tmp, cc,
460 #ifdef CALLOUT_PROFILING
461 &mpcalls_dir, &lockcalls_dir, NULL,
464 tmp = cc->cc_exec_next_dir;
466 tmpn = LIST_NEXT(tmp, c_links.le);
467 LIST_REMOVE(tmp, c_links.le);
468 TAILQ_INSERT_TAIL(&cc->cc_expireq,
470 tmp->c_flags |= CALLOUT_PROCESSED;
475 /* Skip events from distant future. */
476 if (tmp->c_time >= max)
479 * Event minimal time is bigger than present maximal
480 * time, so it cannot be aggregated.
482 if (tmp->c_time > last) {
486 /* Update first and last time, respecting this event. */
487 if (tmp->c_time < first)
489 tmp_max = tmp->c_time + tmp->c_precision;
493 tmp = LIST_NEXT(tmp, c_links.le);
495 /* Proceed with the next bucket. */
498 * Stop if we looked after present time and found
499 * some event we can't execute at now.
500 * Stop if we looked far enough into the future.
502 } while (((int)(firstb - lastb)) <= 0);
503 cc->cc_firstevent = last;
504 #ifndef NO_EVENTTIMERS
505 cpu_new_callout(curcpu, last, first);
507 #ifdef CALLOUT_PROFILING
508 avg_depth_dir += (depth_dir * 1000 - avg_depth_dir) >> 8;
509 avg_mpcalls_dir += (mpcalls_dir * 1000 - avg_mpcalls_dir) >> 8;
510 avg_lockcalls_dir += (lockcalls_dir * 1000 - avg_lockcalls_dir) >> 8;
512 mtx_unlock_spin_flags(&cc->cc_lock, MTX_QUIET);
514 * swi_sched acquires the thread lock, so we don't want to call it
515 * with cc_lock held; incorrect locking order.
517 if (!TAILQ_EMPTY(&cc->cc_expireq))
518 swi_sched(cc->cc_cookie, 0);
521 static struct callout_cpu *
522 callout_lock(struct callout *c)
524 struct callout_cpu *cc;
530 if (cpu == CPUBLOCK) {
531 while (c->c_cpu == CPUBLOCK)
546 callout_cc_add(struct callout *c, struct callout_cpu *cc,
547 sbintime_t sbt, sbintime_t precision, void (*func)(void *),
548 void *arg, int cpu, int flags)
553 if (sbt < cc->cc_lastscan)
554 sbt = cc->cc_lastscan;
556 c->c_flags |= (CALLOUT_ACTIVE | CALLOUT_PENDING);
557 if (flags & C_DIRECT_EXEC)
558 c->c_flags |= CALLOUT_DIRECT;
559 c->c_flags &= ~CALLOUT_PROCESSED;
562 c->c_precision = precision;
563 bucket = callout_get_bucket(c->c_time);
564 CTR3(KTR_CALLOUT, "precision set for %p: %d.%08x",
565 c, (int)(c->c_precision >> 32),
566 (u_int)(c->c_precision & 0xffffffff));
567 LIST_INSERT_HEAD(&cc->cc_callwheel[bucket], c, c_links.le);
568 if (cc->cc_bucket == bucket)
569 cc->cc_exec_next_dir = c;
570 #ifndef NO_EVENTTIMERS
572 * Inform the eventtimers(4) subsystem there's a new callout
573 * that has been inserted, but only if really required.
575 sbt = c->c_time + c->c_precision;
576 if (sbt < cc->cc_firstevent) {
577 cc->cc_firstevent = sbt;
578 cpu_new_callout(cpu, sbt, c->c_time);
584 callout_cc_del(struct callout *c, struct callout_cpu *cc)
587 if ((c->c_flags & CALLOUT_LOCAL_ALLOC) == 0)
590 SLIST_INSERT_HEAD(&cc->cc_callfree, c, c_links.sle);
594 softclock_call_cc(struct callout *c, struct callout_cpu *cc,
595 #ifdef CALLOUT_PROFILING
596 int *mpcalls, int *lockcalls, int *gcalls,
600 struct rm_priotracker tracker;
601 void (*c_func)(void *);
603 struct lock_class *class;
604 struct lock_object *c_lock;
605 uintptr_t lock_status;
608 struct callout_cpu *new_cc;
609 void (*new_func)(void *);
612 sbintime_t new_prec, new_time;
614 #if defined(DIAGNOSTIC) || defined(CALLOUT_PROFILING)
615 sbintime_t sbt1, sbt2;
617 static sbintime_t maxdt = 2 * SBT_1MS; /* 2 msec */
618 static timeout_t *lastfunc;
621 KASSERT((c->c_flags & (CALLOUT_PENDING | CALLOUT_ACTIVE)) ==
622 (CALLOUT_PENDING | CALLOUT_ACTIVE),
623 ("softclock_call_cc: pend|act %p %x", c, c->c_flags));
624 class = (c->c_lock != NULL) ? LOCK_CLASS(c->c_lock) : NULL;
626 if (c->c_flags & CALLOUT_SHAREDLOCK) {
627 if (class == &lock_class_rm)
628 lock_status = (uintptr_t)&tracker;
635 c_flags = c->c_flags;
636 if (c->c_flags & CALLOUT_LOCAL_ALLOC)
637 c->c_flags = CALLOUT_LOCAL_ALLOC;
639 c->c_flags &= ~CALLOUT_PENDING;
640 cc->cc_exec_entity[direct].cc_curr = c;
641 cc->cc_exec_entity[direct].cc_cancel = false;
643 if (c_lock != NULL) {
644 class->lc_lock(c_lock, lock_status);
646 * The callout may have been cancelled
647 * while we switched locks.
649 if (cc->cc_exec_entity[direct].cc_cancel) {
650 class->lc_unlock(c_lock);
653 /* The callout cannot be stopped now. */
654 cc->cc_exec_entity[direct].cc_cancel = true;
655 if (c_lock == &Giant.lock_object) {
656 #ifdef CALLOUT_PROFILING
659 CTR3(KTR_CALLOUT, "callout giant %p func %p arg %p",
662 #ifdef CALLOUT_PROFILING
665 CTR3(KTR_CALLOUT, "callout lock %p func %p arg %p",
669 #ifdef CALLOUT_PROFILING
672 CTR3(KTR_CALLOUT, "callout %p func %p arg %p",
675 #if defined(DIAGNOSTIC) || defined(CALLOUT_PROFILING)
678 THREAD_NO_SLEEPING();
679 SDT_PROBE(callout_execute, kernel, , callout_start, c, 0, 0, 0, 0);
681 SDT_PROBE(callout_execute, kernel, , callout_end, c, 0, 0, 0, 0);
682 THREAD_SLEEPING_OK();
683 #if defined(DIAGNOSTIC) || defined(CALLOUT_PROFILING)
687 if (lastfunc != c_func || sbt2 > maxdt * 2) {
690 "Expensive timeout(9) function: %p(%p) %jd.%09ld s\n",
691 c_func, c_arg, (intmax_t)ts2.tv_sec, ts2.tv_nsec);
697 CTR1(KTR_CALLOUT, "callout %p finished", c);
698 if ((c_flags & CALLOUT_RETURNUNLOCKED) == 0)
699 class->lc_unlock(c_lock);
702 KASSERT(cc->cc_exec_entity[direct].cc_curr == c, ("mishandled cc_curr"));
703 cc->cc_exec_entity[direct].cc_curr = NULL;
704 if (cc->cc_exec_entity[direct].cc_waiting) {
706 * There is someone waiting for the
707 * callout to complete.
708 * If the callout was scheduled for
709 * migration just cancel it.
711 if (cc_cce_migrating(cc, direct)) {
712 cc_cce_cleanup(cc, direct);
715 * It should be assert here that the callout is not
716 * destroyed but that is not easy.
718 c->c_flags &= ~CALLOUT_DFRMIGRATION;
720 cc->cc_exec_entity[direct].cc_waiting = false;
722 wakeup(&cc->cc_exec_entity[direct].cc_waiting);
724 } else if (cc_cce_migrating(cc, direct)) {
725 KASSERT((c_flags & CALLOUT_LOCAL_ALLOC) == 0,
726 ("Migrating legacy callout %p", c));
729 * If the callout was scheduled for
730 * migration just perform it now.
732 new_cpu = cc->cc_exec_entity[direct].ce_migration_cpu;
733 new_time = cc->cc_exec_entity[direct].ce_migration_time;
734 new_prec = cc->cc_exec_entity[direct].ce_migration_prec;
735 new_func = cc->cc_exec_entity[direct].ce_migration_func;
736 new_arg = cc->cc_exec_entity[direct].ce_migration_arg;
737 cc_cce_cleanup(cc, direct);
740 * It should be assert here that the callout is not destroyed
741 * but that is not easy.
743 * As first thing, handle deferred callout stops.
745 if ((c->c_flags & CALLOUT_DFRMIGRATION) == 0) {
747 "deferred cancelled %p func %p arg %p",
748 c, new_func, new_arg);
749 callout_cc_del(c, cc);
752 c->c_flags &= ~CALLOUT_DFRMIGRATION;
754 new_cc = callout_cpu_switch(c, cc, new_cpu);
755 flags = (direct) ? C_DIRECT_EXEC : 0;
756 callout_cc_add(c, new_cc, new_time, new_prec, new_func,
757 new_arg, new_cpu, flags);
761 panic("migration should not happen");
765 * If the current callout is locally allocated (from
766 * timeout(9)) then put it on the freelist.
768 * Note: we need to check the cached copy of c_flags because
769 * if it was not local, then it's not safe to deref the
772 KASSERT((c_flags & CALLOUT_LOCAL_ALLOC) == 0 ||
773 c->c_flags == CALLOUT_LOCAL_ALLOC,
774 ("corrupted callout"));
775 if (c_flags & CALLOUT_LOCAL_ALLOC)
776 callout_cc_del(c, cc);
780 * The callout mechanism is based on the work of Adam M. Costello and
781 * George Varghese, published in a technical report entitled "Redesigning
782 * the BSD Callout and Timer Facilities" and modified slightly for inclusion
783 * in FreeBSD by Justin T. Gibbs. The original work on the data structures
784 * used in this implementation was published by G. Varghese and T. Lauck in
785 * the paper "Hashed and Hierarchical Timing Wheels: Data Structures for
786 * the Efficient Implementation of a Timer Facility" in the Proceedings of
787 * the 11th ACM Annual Symposium on Operating Systems Principles,
788 * Austin, Texas Nov 1987.
792 * Software (low priority) clock interrupt.
793 * Run periodic events from timeout queue.
798 struct callout_cpu *cc;
800 #ifdef CALLOUT_PROFILING
801 int depth = 0, gcalls = 0, lockcalls = 0, mpcalls = 0;
804 cc = (struct callout_cpu *)arg;
806 while ((c = TAILQ_FIRST(&cc->cc_expireq)) != NULL) {
807 TAILQ_REMOVE(&cc->cc_expireq, c, c_links.tqe);
808 softclock_call_cc(c, cc,
809 #ifdef CALLOUT_PROFILING
810 &mpcalls, &lockcalls, &gcalls,
813 #ifdef CALLOUT_PROFILING
817 #ifdef CALLOUT_PROFILING
818 avg_depth += (depth * 1000 - avg_depth) >> 8;
819 avg_mpcalls += (mpcalls * 1000 - avg_mpcalls) >> 8;
820 avg_lockcalls += (lockcalls * 1000 - avg_lockcalls) >> 8;
821 avg_gcalls += (gcalls * 1000 - avg_gcalls) >> 8;
828 * Execute a function after a specified length of time.
831 * Cancel previous timeout function call.
833 * callout_handle_init --
834 * Initialize a handle so that using it with untimeout is benign.
836 * See AT&T BCI Driver Reference Manual for specification. This
837 * implementation differs from that one in that although an
838 * identification value is returned from timeout, the original
839 * arguments to timeout as well as the identifier are used to
840 * identify entries for untimeout.
842 struct callout_handle
843 timeout(ftn, arg, to_ticks)
848 struct callout_cpu *cc;
850 struct callout_handle handle;
852 cc = CC_CPU(timeout_cpu);
854 /* Fill in the next free callout structure. */
855 new = SLIST_FIRST(&cc->cc_callfree);
857 /* XXX Attempt to malloc first */
858 panic("timeout table full");
859 SLIST_REMOVE_HEAD(&cc->cc_callfree, c_links.sle);
860 callout_reset(new, to_ticks, ftn, arg);
861 handle.callout = new;
868 untimeout(ftn, arg, handle)
871 struct callout_handle handle;
873 struct callout_cpu *cc;
876 * Check for a handle that was initialized
877 * by callout_handle_init, but never used
878 * for a real timeout.
880 if (handle.callout == NULL)
883 cc = callout_lock(handle.callout);
884 if (handle.callout->c_func == ftn && handle.callout->c_arg == arg)
885 callout_stop(handle.callout);
890 callout_handle_init(struct callout_handle *handle)
892 handle->callout = NULL;
896 * New interface; clients allocate their own callout structures.
898 * callout_reset() - establish or change a timeout
899 * callout_stop() - disestablish a timeout
900 * callout_init() - initialize a callout structure so that it can
901 * safely be passed to callout_reset() and callout_stop()
903 * <sys/callout.h> defines three convenience macros:
905 * callout_active() - returns truth if callout has not been stopped,
906 * drained, or deactivated since the last time the callout was
908 * callout_pending() - returns truth if callout is still waiting for timeout
909 * callout_deactivate() - marks the callout as having been serviced
912 callout_reset_sbt_on(struct callout *c, sbintime_t sbt, sbintime_t precision,
913 void (*ftn)(void *), void *arg, int cpu, int flags)
915 sbintime_t to_sbt, pr;
916 struct callout_cpu *cc;
917 int cancelled, direct;
920 if (flags & C_ABSOLUTE) {
923 if ((flags & C_HARDCLOCK) && (sbt < tick_sbt))
925 if ((flags & C_HARDCLOCK) ||
926 #ifdef NO_EVENTTIMERS
927 sbt >= sbt_timethreshold) {
928 to_sbt = getsbinuptime();
930 /* Add safety belt for the case of hz > 1000. */
931 to_sbt += tc_tick_sbt - tick_sbt;
933 sbt >= sbt_tickthreshold) {
935 * Obtain the time of the last hardclock() call on
936 * this CPU directly from the kern_clocksource.c.
937 * This value is per-CPU, but it is equal for all
941 to_sbt = DPCPU_GET(hardclocktime);
944 to_sbt = DPCPU_GET(hardclocktime);
948 if ((flags & C_HARDCLOCK) == 0)
951 to_sbt = sbinuptime();
953 pr = ((C_PRELGET(flags) < 0) ? sbt >> tc_precexp :
954 sbt >> C_PRELGET(flags));
959 * Don't allow migration of pre-allocated callouts lest they
962 if (c->c_flags & CALLOUT_LOCAL_ALLOC)
964 direct = (c->c_flags & CALLOUT_DIRECT) != 0;
965 KASSERT(!direct || c->c_lock == NULL,
966 ("%s: direct callout %p has lock", __func__, c));
967 cc = callout_lock(c);
968 if (cc->cc_exec_entity[direct].cc_curr == c) {
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.
974 if (c->c_lock != NULL && !cc->cc_exec_entity[direct].cc_cancel)
975 cancelled = cc->cc_exec_entity[direct].cc_cancel = true;
976 if (cc->cc_exec_entity[direct].cc_waiting) {
978 * Someone has called callout_drain to kill this
979 * callout. Don't reschedule.
981 CTR4(KTR_CALLOUT, "%s %p func %p arg %p",
982 cancelled ? "cancelled" : "failed to cancel",
983 c, c->c_func, c->c_arg);
988 if (c->c_flags & CALLOUT_PENDING) {
989 if ((c->c_flags & CALLOUT_PROCESSED) == 0) {
990 if (cc->cc_exec_next_dir == c)
991 cc->cc_exec_next_dir = LIST_NEXT(c, c_links.le);
992 LIST_REMOVE(c, c_links.le);
994 TAILQ_REMOVE(&cc->cc_expireq, c, c_links.tqe);
996 c->c_flags &= ~(CALLOUT_ACTIVE | CALLOUT_PENDING);
1001 * If the callout must migrate try to perform it immediately.
1002 * If the callout is currently running, just defer the migration
1003 * to a more appropriate moment.
1005 if (c->c_cpu != cpu) {
1006 if (cc->cc_exec_entity[direct].cc_curr == c) {
1007 cc->cc_exec_entity[direct].ce_migration_cpu = cpu;
1008 cc->cc_exec_entity[direct].ce_migration_time
1010 cc->cc_exec_entity[direct].ce_migration_prec
1012 cc->cc_exec_entity[direct].ce_migration_func = ftn;
1013 cc->cc_exec_entity[direct].ce_migration_arg = arg;
1014 c->c_flags |= CALLOUT_DFRMIGRATION;
1016 "migration of %p func %p arg %p in %d.%08x to %u deferred",
1017 c, c->c_func, c->c_arg, (int)(to_sbt >> 32),
1018 (u_int)(to_sbt & 0xffffffff), cpu);
1022 cc = callout_cpu_switch(c, cc, cpu);
1026 callout_cc_add(c, cc, to_sbt, precision, ftn, arg, cpu, flags);
1027 CTR6(KTR_CALLOUT, "%sscheduled %p func %p arg %p in %d.%08x",
1028 cancelled ? "re" : "", c, c->c_func, c->c_arg, (int)(to_sbt >> 32),
1029 (u_int)(to_sbt & 0xffffffff));
1036 * Common idioms that can be optimized in the future.
1039 callout_schedule_on(struct callout *c, int to_ticks, int cpu)
1041 return callout_reset_on(c, to_ticks, c->c_func, c->c_arg, cpu);
1045 callout_schedule(struct callout *c, int to_ticks)
1047 return callout_reset_on(c, to_ticks, c->c_func, c->c_arg, c->c_cpu);
1051 _callout_stop_safe(c, safe)
1055 struct callout_cpu *cc, *old_cc;
1056 struct lock_class *class;
1057 int direct, sq_locked, use_lock;
1060 * Some old subsystems don't hold Giant while running a callout_stop(),
1061 * so just discard this check for the moment.
1063 if (!safe && c->c_lock != NULL) {
1064 if (c->c_lock == &Giant.lock_object)
1065 use_lock = mtx_owned(&Giant);
1068 class = LOCK_CLASS(c->c_lock);
1069 class->lc_assert(c->c_lock, LA_XLOCKED);
1073 direct = (c->c_flags & CALLOUT_DIRECT) != 0;
1077 cc = callout_lock(c);
1080 * If the callout was migrating while the callout cpu lock was
1081 * dropped, just drop the sleepqueue lock and check the states
1084 if (sq_locked != 0 && cc != old_cc) {
1087 sleepq_release(&old_cc->cc_exec_entity[direct].cc_waiting);
1092 panic("migration should not happen");
1097 * If the callout isn't pending, it's not on the queue, so
1098 * don't attempt to remove it from the queue. We can try to
1099 * stop it by other means however.
1101 if (!(c->c_flags & CALLOUT_PENDING)) {
1102 c->c_flags &= ~CALLOUT_ACTIVE;
1105 * If it wasn't on the queue and it isn't the current
1106 * callout, then we can't stop it, so just bail.
1108 if (cc->cc_exec_entity[direct].cc_curr != c) {
1109 CTR3(KTR_CALLOUT, "failed to stop %p func %p arg %p",
1110 c, c->c_func, c->c_arg);
1114 &cc->cc_exec_entity[direct].cc_waiting);
1120 * The current callout is running (or just
1121 * about to run) and blocking is allowed, so
1122 * just wait for the current invocation to
1125 while (cc->cc_exec_entity[direct].cc_curr == c) {
1127 * Use direct calls to sleepqueue interface
1128 * instead of cv/msleep in order to avoid
1129 * a LOR between cc_lock and sleepqueue
1130 * chain spinlocks. This piece of code
1131 * emulates a msleep_spin() call actually.
1133 * If we already have the sleepqueue chain
1134 * locked, then we can safely block. If we
1135 * don't already have it locked, however,
1136 * we have to drop the cc_lock to lock
1137 * it. This opens several races, so we
1138 * restart at the beginning once we have
1139 * both locks. If nothing has changed, then
1140 * we will end up back here with sq_locked
1146 &cc->cc_exec_entity[direct].cc_waiting);
1153 * Migration could be cancelled here, but
1154 * as long as it is still not sure when it
1155 * will be packed up, just let softclock()
1158 cc->cc_exec_entity[direct].cc_waiting = true;
1162 &cc->cc_exec_entity[direct].cc_waiting,
1163 &cc->cc_lock.lock_object, "codrain",
1166 &cc->cc_exec_entity[direct].cc_waiting,
1171 /* Reacquire locks previously released. */
1175 } else if (use_lock &&
1176 !cc->cc_exec_entity[direct].cc_cancel) {
1178 * The current callout is waiting for its
1179 * lock which we hold. Cancel the callout
1180 * and return. After our caller drops the
1181 * lock, the callout will be skipped in
1184 cc->cc_exec_entity[direct].cc_cancel = true;
1185 CTR3(KTR_CALLOUT, "cancelled %p func %p arg %p",
1186 c, c->c_func, c->c_arg);
1187 KASSERT(!cc_cce_migrating(cc, direct),
1188 ("callout wrongly scheduled for migration"));
1190 KASSERT(!sq_locked, ("sleepqueue chain locked"));
1192 } else if ((c->c_flags & CALLOUT_DFRMIGRATION) != 0) {
1193 c->c_flags &= ~CALLOUT_DFRMIGRATION;
1194 CTR3(KTR_CALLOUT, "postponing stop %p func %p arg %p",
1195 c, c->c_func, c->c_arg);
1199 CTR3(KTR_CALLOUT, "failed to stop %p func %p arg %p",
1200 c, c->c_func, c->c_arg);
1202 KASSERT(!sq_locked, ("sleepqueue chain still locked"));
1206 sleepq_release(&cc->cc_exec_entity[direct].cc_waiting);
1208 c->c_flags &= ~(CALLOUT_ACTIVE | CALLOUT_PENDING);
1210 CTR3(KTR_CALLOUT, "cancelled %p func %p arg %p",
1211 c, c->c_func, c->c_arg);
1212 if ((c->c_flags & CALLOUT_PROCESSED) == 0) {
1213 if (cc->cc_exec_next_dir == c)
1214 cc->cc_exec_next_dir = LIST_NEXT(c, c_links.le);
1215 LIST_REMOVE(c, c_links.le);
1217 TAILQ_REMOVE(&cc->cc_expireq, c, c_links.tqe);
1218 callout_cc_del(c, cc);
1225 callout_init(c, mpsafe)
1229 bzero(c, sizeof *c);
1232 c->c_flags = CALLOUT_RETURNUNLOCKED;
1234 c->c_lock = &Giant.lock_object;
1237 c->c_cpu = timeout_cpu;
1241 _callout_init_lock(c, lock, flags)
1243 struct lock_object *lock;
1246 bzero(c, sizeof *c);
1248 KASSERT((flags & ~(CALLOUT_RETURNUNLOCKED | CALLOUT_SHAREDLOCK)) == 0,
1249 ("callout_init_lock: bad flags %d", flags));
1250 KASSERT(lock != NULL || (flags & CALLOUT_RETURNUNLOCKED) == 0,
1251 ("callout_init_lock: CALLOUT_RETURNUNLOCKED with no lock"));
1252 KASSERT(lock == NULL || !(LOCK_CLASS(lock)->lc_flags &
1253 (LC_SPINLOCK | LC_SLEEPABLE)), ("%s: invalid lock class",
1255 c->c_flags = flags & (CALLOUT_RETURNUNLOCKED | CALLOUT_SHAREDLOCK);
1256 c->c_cpu = timeout_cpu;
1259 #ifdef APM_FIXUP_CALLTODO
1261 * Adjust the kernel calltodo timeout list. This routine is used after
1262 * an APM resume to recalculate the calltodo timer list values with the
1263 * number of hz's we have been sleeping. The next hardclock() will detect
1264 * that there are fired timers and run softclock() to execute them.
1266 * Please note, I have not done an exhaustive analysis of what code this
1267 * might break. I am motivated to have my select()'s and alarm()'s that
1268 * have expired during suspend firing upon resume so that the applications
1269 * which set the timer can do the maintanence the timer was for as close
1270 * as possible to the originally intended time. Testing this code for a
1271 * week showed that resuming from a suspend resulted in 22 to 25 timers
1272 * firing, which seemed independant on whether the suspend was 2 hours or
1273 * 2 days. Your milage may vary. - Ken Key <key@cs.utk.edu>
1276 adjust_timeout_calltodo(time_change)
1277 struct timeval *time_change;
1279 register struct callout *p;
1280 unsigned long delta_ticks;
1283 * How many ticks were we asleep?
1284 * (stolen from tvtohz()).
1287 /* Don't do anything */
1288 if (time_change->tv_sec < 0)
1290 else if (time_change->tv_sec <= LONG_MAX / 1000000)
1291 delta_ticks = (time_change->tv_sec * 1000000 +
1292 time_change->tv_usec + (tick - 1)) / tick + 1;
1293 else if (time_change->tv_sec <= LONG_MAX / hz)
1294 delta_ticks = time_change->tv_sec * hz +
1295 (time_change->tv_usec + (tick - 1)) / tick + 1;
1297 delta_ticks = LONG_MAX;
1299 if (delta_ticks > INT_MAX)
1300 delta_ticks = INT_MAX;
1303 * Now rip through the timer calltodo list looking for timers
1307 /* don't collide with softclock() */
1309 for (p = calltodo.c_next; p != NULL; p = p->c_next) {
1310 p->c_time -= delta_ticks;
1312 /* Break if the timer had more time on it than delta_ticks */
1316 /* take back the ticks the timer didn't use (p->c_time <= 0) */
1317 delta_ticks = -p->c_time;
1323 #endif /* APM_FIXUP_CALLTODO */
1326 flssbt(sbintime_t sbt)
1329 sbt += (uint64_t)sbt >> 1;
1330 if (sizeof(long) >= sizeof(sbintime_t))
1333 return (flsl(((uint64_t)sbt) >> 32) + 32);
1338 * Dump immediate statistic snapshot of the scheduled callouts.
1341 sysctl_kern_callout_stat(SYSCTL_HANDLER_ARGS)
1343 struct callout *tmp;
1344 struct callout_cpu *cc;
1345 struct callout_list *sc;
1346 sbintime_t maxpr, maxt, medpr, medt, now, spr, st, t;
1347 int ct[64], cpr[64], ccpbk[32];
1348 int error, val, i, count, tcum, pcum, maxc, c, medc;
1354 error = sysctl_handle_int(oidp, &val, 0, req);
1355 if (error != 0 || req->newptr == NULL)
1358 st = spr = maxt = maxpr = 0;
1359 bzero(ccpbk, sizeof(ccpbk));
1360 bzero(ct, sizeof(ct));
1361 bzero(cpr, sizeof(cpr));
1367 cc = CC_CPU(timeout_cpu);
1370 for (i = 0; i < callwheelsize; i++) {
1371 sc = &cc->cc_callwheel[i];
1373 LIST_FOREACH(tmp, sc, c_links.le) {
1375 t = tmp->c_time - now;
1379 spr += tmp->c_precision / SBT_1US;
1382 if (tmp->c_precision > maxpr)
1383 maxpr = tmp->c_precision;
1385 cpr[flssbt(tmp->c_precision)]++;
1389 ccpbk[fls(c + c / 2)]++;
1397 for (i = 0, tcum = 0; i < 64 && tcum < count / 2; i++)
1399 medt = (i >= 2) ? (((sbintime_t)1) << (i - 2)) : 0;
1400 for (i = 0, pcum = 0; i < 64 && pcum < count / 2; i++)
1402 medpr = (i >= 2) ? (((sbintime_t)1) << (i - 2)) : 0;
1403 for (i = 0, c = 0; i < 32 && c < count / 2; i++)
1405 medc = (i >= 2) ? (1 << (i - 2)) : 0;
1407 printf("Scheduled callouts statistic snapshot:\n");
1408 printf(" Callouts: %6d Buckets: %6d*%-3d Bucket size: 0.%06ds\n",
1409 count, callwheelsize, mp_ncpus, 1000000 >> CC_HASH_SHIFT);
1410 printf(" C/Bk: med %5d avg %6d.%06jd max %6d\n",
1412 count / callwheelsize / mp_ncpus,
1413 (uint64_t)count * 1000000 / callwheelsize / mp_ncpus % 1000000,
1415 printf(" Time: med %5jd.%06jds avg %6jd.%06jds max %6jd.%06jds\n",
1416 medt / SBT_1S, (medt & 0xffffffff) * 1000000 >> 32,
1417 (st / count) / 1000000, (st / count) % 1000000,
1418 maxt / SBT_1S, (maxt & 0xffffffff) * 1000000 >> 32);
1419 printf(" Prec: med %5jd.%06jds avg %6jd.%06jds max %6jd.%06jds\n",
1420 medpr / SBT_1S, (medpr & 0xffffffff) * 1000000 >> 32,
1421 (spr / count) / 1000000, (spr / count) % 1000000,
1422 maxpr / SBT_1S, (maxpr & 0xffffffff) * 1000000 >> 32);
1423 printf(" Distribution: \tbuckets\t time\t tcum\t"
1425 for (i = 0, tcum = pcum = 0; i < 64; i++) {
1426 if (ct[i] == 0 && cpr[i] == 0)
1428 t = (i != 0) ? (((sbintime_t)1) << (i - 1)) : 0;
1431 printf(" %10jd.%06jds\t 2**%d\t%7d\t%7d\t%7d\t%7d\n",
1432 t / SBT_1S, (t & 0xffffffff) * 1000000 >> 32,
1433 i - 1 - (32 - CC_HASH_SHIFT),
1434 ct[i], tcum, cpr[i], pcum);
1438 SYSCTL_PROC(_kern, OID_AUTO, callout_stat,
1439 CTLTYPE_INT | CTLFLAG_RW | CTLFLAG_MPSAFE,
1440 0, 0, sysctl_kern_callout_stat, "I",
1441 "Dump immediate statistic snapshot of the scheduled callouts");