2 * SPDX-License-Identifier: BSD-3-Clause
4 * Copyright (c) 1982, 1986, 1991, 1993
5 * The Regents of the University of California. All rights reserved.
6 * (c) UNIX System Laboratories, Inc.
7 * All or some portions of this file are derived from material licensed
8 * to the University of California by American Telephone and Telegraph
9 * Co. or Unix System Laboratories, Inc. and are reproduced herein with
10 * the permission of UNIX System Laboratories, Inc.
12 * Redistribution and use in source and binary forms, with or without
13 * modification, are permitted provided that the following conditions
15 * 1. Redistributions of source code must retain the above copyright
16 * notice, this list of conditions and the following disclaimer.
17 * 2. Redistributions in binary form must reproduce the above copyright
18 * notice, this list of conditions and the following disclaimer in the
19 * documentation and/or other materials provided with the distribution.
20 * 3. Neither the name of the University nor the names of its contributors
21 * may be used to endorse or promote products derived from this software
22 * without specific prior written permission.
24 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
25 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
26 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
27 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
28 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
29 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
30 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
31 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
32 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
33 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
36 * From: @(#)kern_clock.c 8.5 (Berkeley) 1/21/94
39 #include <sys/cdefs.h>
40 __FBSDID("$FreeBSD$");
42 #include "opt_callout_profiling.h"
45 #include "opt_timer.h"
49 #include <sys/param.h>
50 #include <sys/systm.h>
52 #include <sys/callout.h>
53 #include <sys/domainset.h>
55 #include <sys/interrupt.h>
56 #include <sys/kernel.h>
59 #include <sys/malloc.h>
60 #include <sys/mutex.h>
63 #include <sys/sleepqueue.h>
64 #include <sys/sysctl.h>
69 #include <ddb/db_sym.h>
70 #include <machine/_inttypes.h>
74 #include <machine/cpu.h>
77 #ifndef NO_EVENTTIMERS
78 DPCPU_DECLARE(sbintime_t, hardclocktime);
81 SDT_PROVIDER_DEFINE(callout_execute);
82 SDT_PROBE_DEFINE1(callout_execute, , , callout__start, "struct callout *");
83 SDT_PROBE_DEFINE1(callout_execute, , , callout__end, "struct callout *");
85 #ifdef CALLOUT_PROFILING
87 SYSCTL_INT(_debug, OID_AUTO, to_avg_depth, CTLFLAG_RD, &avg_depth, 0,
88 "Average number of items examined per softclock call. Units = 1/1000");
89 static int avg_gcalls;
90 SYSCTL_INT(_debug, OID_AUTO, to_avg_gcalls, CTLFLAG_RD, &avg_gcalls, 0,
91 "Average number of Giant callouts made per softclock call. Units = 1/1000");
92 static int avg_lockcalls;
93 SYSCTL_INT(_debug, OID_AUTO, to_avg_lockcalls, CTLFLAG_RD, &avg_lockcalls, 0,
94 "Average number of lock callouts made per softclock call. Units = 1/1000");
95 static int avg_mpcalls;
96 SYSCTL_INT(_debug, OID_AUTO, to_avg_mpcalls, CTLFLAG_RD, &avg_mpcalls, 0,
97 "Average number of MP callouts made per softclock call. Units = 1/1000");
98 static int avg_depth_dir;
99 SYSCTL_INT(_debug, OID_AUTO, to_avg_depth_dir, CTLFLAG_RD, &avg_depth_dir, 0,
100 "Average number of direct callouts examined per callout_process call. "
102 static int avg_lockcalls_dir;
103 SYSCTL_INT(_debug, OID_AUTO, to_avg_lockcalls_dir, CTLFLAG_RD,
104 &avg_lockcalls_dir, 0, "Average number of lock direct callouts made per "
105 "callout_process call. Units = 1/1000");
106 static int avg_mpcalls_dir;
107 SYSCTL_INT(_debug, OID_AUTO, to_avg_mpcalls_dir, CTLFLAG_RD, &avg_mpcalls_dir,
108 0, "Average number of MP direct callouts made per callout_process call. "
113 SYSCTL_INT(_kern, OID_AUTO, ncallout, CTLFLAG_RDTUN | CTLFLAG_NOFETCH, &ncallout, 0,
114 "Number of entries in callwheel and size of timeout() preallocation");
117 static int pin_default_swi = 1;
118 static int pin_pcpu_swi = 1;
120 static int pin_default_swi = 0;
121 static int pin_pcpu_swi = 0;
124 SYSCTL_INT(_kern, OID_AUTO, pin_default_swi, CTLFLAG_RDTUN | CTLFLAG_NOFETCH, &pin_default_swi,
125 0, "Pin the default (non-per-cpu) swi (shared with PCPU 0 swi)");
126 SYSCTL_INT(_kern, OID_AUTO, pin_pcpu_swi, CTLFLAG_RDTUN | CTLFLAG_NOFETCH, &pin_pcpu_swi,
127 0, "Pin the per-CPU swis (except PCPU 0, which is also default");
131 * allocate more timeout table slots when table overflows.
133 static u_int __read_mostly callwheelsize;
134 static u_int __read_mostly callwheelmask;
137 * The callout cpu exec entities represent informations necessary for
138 * describing the state of callouts currently running on the CPU and the ones
139 * necessary for migrating callouts to the new callout cpu. In particular,
140 * the first entry of the array cc_exec_entity holds informations for callout
141 * running in SWI thread context, while the second one holds informations
142 * for callout running directly from hardware interrupt context.
143 * The cached informations are very important for deferring migration when
144 * the migrating callout is already running.
147 struct callout *cc_curr;
148 callout_func_t *cc_drain;
152 callout_func_t *ce_migration_func;
153 void *ce_migration_arg;
154 sbintime_t ce_migration_time;
155 sbintime_t ce_migration_prec;
156 int ce_migration_cpu;
163 * There is one struct callout_cpu per cpu, holding all relevant
164 * state for the callout processing thread on the individual CPU.
167 struct mtx_padalign cc_lock;
168 struct cc_exec cc_exec_entity[2];
169 struct callout *cc_next;
170 struct callout *cc_callout;
171 struct callout_list *cc_callwheel;
172 struct callout_tailq cc_expireq;
173 struct callout_slist cc_callfree;
174 sbintime_t cc_firstevent;
175 sbintime_t cc_lastscan;
180 char cc_ktr_event_name[20];
184 #define callout_migrating(c) ((c)->c_iflags & CALLOUT_DFRMIGRATION)
186 #define cc_exec_curr(cc, dir) cc->cc_exec_entity[dir].cc_curr
187 #define cc_exec_last_func(cc, dir) cc->cc_exec_entity[dir].cc_last_func
188 #define cc_exec_last_arg(cc, dir) cc->cc_exec_entity[dir].cc_last_arg
189 #define cc_exec_drain(cc, dir) cc->cc_exec_entity[dir].cc_drain
190 #define cc_exec_next(cc) cc->cc_next
191 #define cc_exec_cancel(cc, dir) cc->cc_exec_entity[dir].cc_cancel
192 #define cc_exec_waiting(cc, dir) cc->cc_exec_entity[dir].cc_waiting
194 #define cc_migration_func(cc, dir) cc->cc_exec_entity[dir].ce_migration_func
195 #define cc_migration_arg(cc, dir) cc->cc_exec_entity[dir].ce_migration_arg
196 #define cc_migration_cpu(cc, dir) cc->cc_exec_entity[dir].ce_migration_cpu
197 #define cc_migration_time(cc, dir) cc->cc_exec_entity[dir].ce_migration_time
198 #define cc_migration_prec(cc, dir) cc->cc_exec_entity[dir].ce_migration_prec
200 struct callout_cpu cc_cpu[MAXCPU];
201 #define CPUBLOCK MAXCPU
202 #define CC_CPU(cpu) (&cc_cpu[(cpu)])
203 #define CC_SELF() CC_CPU(PCPU_GET(cpuid))
205 struct callout_cpu cc_cpu;
206 #define CC_CPU(cpu) &cc_cpu
207 #define CC_SELF() &cc_cpu
209 #define CC_LOCK(cc) mtx_lock_spin(&(cc)->cc_lock)
210 #define CC_UNLOCK(cc) mtx_unlock_spin(&(cc)->cc_lock)
211 #define CC_LOCK_ASSERT(cc) mtx_assert(&(cc)->cc_lock, MA_OWNED)
213 static int __read_mostly timeout_cpu;
215 static void callout_cpu_init(struct callout_cpu *cc, int cpu);
216 static void softclock_call_cc(struct callout *c, struct callout_cpu *cc,
217 #ifdef CALLOUT_PROFILING
218 int *mpcalls, int *lockcalls, int *gcalls,
222 static MALLOC_DEFINE(M_CALLOUT, "callout", "Callout datastructures");
226 * cc_curr - If a callout is in progress, it is cc_curr.
227 * If cc_curr is non-NULL, threads waiting in
228 * callout_drain() will be woken up as soon as the
229 * relevant callout completes.
230 * cc_cancel - Changing to 1 with both callout_lock and cc_lock held
231 * guarantees that the current callout will not run.
232 * The softclock() function sets this to 0 before it
233 * drops callout_lock to acquire c_lock, and it calls
234 * the handler only if curr_cancelled is still 0 after
235 * cc_lock is successfully acquired.
236 * cc_waiting - If a thread is waiting in callout_drain(), then
237 * callout_wait is nonzero. Set only when
238 * cc_curr is non-NULL.
242 * Resets the execution entity tied to a specific callout cpu.
245 cc_cce_cleanup(struct callout_cpu *cc, int direct)
248 cc_exec_curr(cc, direct) = NULL;
249 cc_exec_cancel(cc, direct) = false;
250 cc_exec_waiting(cc, direct) = false;
252 cc_migration_cpu(cc, direct) = CPUBLOCK;
253 cc_migration_time(cc, direct) = 0;
254 cc_migration_prec(cc, direct) = 0;
255 cc_migration_func(cc, direct) = NULL;
256 cc_migration_arg(cc, direct) = NULL;
261 * Checks if migration is requested by a specific callout cpu.
264 cc_cce_migrating(struct callout_cpu *cc, int direct)
268 return (cc_migration_cpu(cc, direct) != CPUBLOCK);
275 * Kernel low level callwheel initialization
276 * called on the BSP during kernel startup.
279 callout_callwheel_init(void *dummy)
281 struct callout_cpu *cc;
284 * Calculate the size of the callout wheel and the preallocated
285 * timeout() structures.
286 * XXX: Clip callout to result of previous function of maxusers
287 * maximum 384. This is still huge, but acceptable.
289 memset(CC_CPU(curcpu), 0, sizeof(cc_cpu));
290 ncallout = imin(16 + maxproc + maxfiles, 18508);
291 TUNABLE_INT_FETCH("kern.ncallout", &ncallout);
294 * Calculate callout wheel size, should be next power of two higher
297 callwheelsize = 1 << fls(ncallout);
298 callwheelmask = callwheelsize - 1;
301 * Fetch whether we're pinning the swi's or not.
303 TUNABLE_INT_FETCH("kern.pin_default_swi", &pin_default_swi);
304 TUNABLE_INT_FETCH("kern.pin_pcpu_swi", &pin_pcpu_swi);
307 * Only BSP handles timeout(9) and receives a preallocation.
309 * XXX: Once all timeout(9) consumers are converted this can
312 timeout_cpu = PCPU_GET(cpuid);
313 cc = CC_CPU(timeout_cpu);
314 cc->cc_callout = malloc(ncallout * sizeof(struct callout),
315 M_CALLOUT, M_WAITOK);
316 callout_cpu_init(cc, timeout_cpu);
318 SYSINIT(callwheel_init, SI_SUB_CPU, SI_ORDER_ANY, callout_callwheel_init, NULL);
321 * Initialize the per-cpu callout structures.
324 callout_cpu_init(struct callout_cpu *cc, int cpu)
329 mtx_init(&cc->cc_lock, "callout", NULL, MTX_SPIN | MTX_RECURSE);
330 SLIST_INIT(&cc->cc_callfree);
332 cc->cc_callwheel = malloc_domainset(sizeof(struct callout_list) *
333 callwheelsize, M_CALLOUT,
334 DOMAINSET_PREF(pcpu_find(cpu)->pc_domain), M_WAITOK);
335 for (i = 0; i < callwheelsize; i++)
336 LIST_INIT(&cc->cc_callwheel[i]);
337 TAILQ_INIT(&cc->cc_expireq);
338 cc->cc_firstevent = SBT_MAX;
339 for (i = 0; i < 2; i++)
340 cc_cce_cleanup(cc, i);
342 snprintf(cc->cc_ktr_event_name, sizeof(cc->cc_ktr_event_name),
343 "callwheel cpu %d", cpu);
345 if (cc->cc_callout == NULL) /* Only BSP handles timeout(9) */
347 for (i = 0; i < ncallout; i++) {
348 c = &cc->cc_callout[i];
350 c->c_iflags = CALLOUT_LOCAL_ALLOC;
351 SLIST_INSERT_HEAD(&cc->cc_callfree, c, c_links.sle);
357 * Switches the cpu tied to a specific callout.
358 * The function expects a locked incoming callout cpu and returns with
359 * locked outcoming callout cpu.
361 static struct callout_cpu *
362 callout_cpu_switch(struct callout *c, struct callout_cpu *cc, int new_cpu)
364 struct callout_cpu *new_cc;
366 MPASS(c != NULL && cc != NULL);
370 * Avoid interrupts and preemption firing after the callout cpu
371 * is blocked in order to avoid deadlocks as the new thread
372 * may be willing to acquire the callout cpu lock.
377 new_cc = CC_CPU(new_cpu);
386 * Start standard softclock thread.
389 start_softclock(void *dummy)
391 struct callout_cpu *cc;
392 char name[MAXCOMLEN];
395 struct intr_event *ie;
398 cc = CC_CPU(timeout_cpu);
399 snprintf(name, sizeof(name), "clock (%d)", timeout_cpu);
400 if (swi_add(&clk_intr_event, name, softclock, cc, SWI_CLOCK,
401 INTR_MPSAFE, &cc->cc_cookie))
402 panic("died while creating standard software ithreads");
403 if (pin_default_swi &&
404 (intr_event_bind(clk_intr_event, timeout_cpu) != 0)) {
405 printf("%s: timeout clock couldn't be pinned to cpu %d\n",
412 if (cpu == timeout_cpu)
415 cc->cc_callout = NULL; /* Only BSP handles timeout(9). */
416 callout_cpu_init(cc, cpu);
417 snprintf(name, sizeof(name), "clock (%d)", cpu);
419 if (swi_add(&ie, name, softclock, cc, SWI_CLOCK,
420 INTR_MPSAFE, &cc->cc_cookie))
421 panic("died while creating standard software ithreads");
422 if (pin_pcpu_swi && (intr_event_bind(ie, cpu) != 0)) {
423 printf("%s: per-cpu clock couldn't be pinned to "
431 SYSINIT(start_softclock, SI_SUB_SOFTINTR, SI_ORDER_FIRST, start_softclock, NULL);
433 #define CC_HASH_SHIFT 8
436 callout_hash(sbintime_t sbt)
439 return (sbt >> (32 - CC_HASH_SHIFT));
443 callout_get_bucket(sbintime_t sbt)
446 return (callout_hash(sbt) & callwheelmask);
450 callout_process(sbintime_t now)
452 struct callout *tmp, *tmpn;
453 struct callout_cpu *cc;
454 struct callout_list *sc;
455 sbintime_t first, last, max, tmp_max;
457 u_int firstb, lastb, nowb;
458 #ifdef CALLOUT_PROFILING
459 int depth_dir = 0, mpcalls_dir = 0, lockcalls_dir = 0;
463 mtx_lock_spin_flags(&cc->cc_lock, MTX_QUIET);
465 /* Compute the buckets of the last scan and present times. */
466 firstb = callout_hash(cc->cc_lastscan);
467 cc->cc_lastscan = now;
468 nowb = callout_hash(now);
470 /* Compute the last bucket and minimum time of the bucket after it. */
472 lookahead = (SBT_1S / 16);
473 else if (nowb - firstb == 1)
474 lookahead = (SBT_1S / 8);
476 lookahead = (SBT_1S / 2);
478 first += (lookahead / 2);
480 last &= (0xffffffffffffffffLLU << (32 - CC_HASH_SHIFT));
481 lastb = callout_hash(last) - 1;
485 * Check if we wrapped around the entire wheel from the last scan.
486 * In case, we need to scan entirely the wheel for pending callouts.
488 if (lastb - firstb >= callwheelsize) {
489 lastb = firstb + callwheelsize - 1;
490 if (nowb - firstb >= callwheelsize)
494 /* Iterate callwheel from firstb to nowb and then up to lastb. */
496 sc = &cc->cc_callwheel[firstb & callwheelmask];
497 tmp = LIST_FIRST(sc);
498 while (tmp != NULL) {
499 /* Run the callout if present time within allowed. */
500 if (tmp->c_time <= now) {
502 * Consumer told us the callout may be run
503 * directly from hardware interrupt context.
505 if (tmp->c_iflags & CALLOUT_DIRECT) {
506 #ifdef CALLOUT_PROFILING
510 LIST_NEXT(tmp, c_links.le);
511 cc->cc_bucket = firstb & callwheelmask;
512 LIST_REMOVE(tmp, c_links.le);
513 softclock_call_cc(tmp, cc,
514 #ifdef CALLOUT_PROFILING
515 &mpcalls_dir, &lockcalls_dir, NULL,
518 tmp = cc_exec_next(cc);
519 cc_exec_next(cc) = NULL;
521 tmpn = LIST_NEXT(tmp, c_links.le);
522 LIST_REMOVE(tmp, c_links.le);
523 TAILQ_INSERT_TAIL(&cc->cc_expireq,
525 tmp->c_iflags |= CALLOUT_PROCESSED;
530 /* Skip events from distant future. */
531 if (tmp->c_time >= max)
534 * Event minimal time is bigger than present maximal
535 * time, so it cannot be aggregated.
537 if (tmp->c_time > last) {
541 /* Update first and last time, respecting this event. */
542 if (tmp->c_time < first)
544 tmp_max = tmp->c_time + tmp->c_precision;
548 tmp = LIST_NEXT(tmp, c_links.le);
550 /* Proceed with the next bucket. */
553 * Stop if we looked after present time and found
554 * some event we can't execute at now.
555 * Stop if we looked far enough into the future.
557 } while (((int)(firstb - lastb)) <= 0);
558 cc->cc_firstevent = last;
559 #ifndef NO_EVENTTIMERS
560 cpu_new_callout(curcpu, last, first);
562 #ifdef CALLOUT_PROFILING
563 avg_depth_dir += (depth_dir * 1000 - avg_depth_dir) >> 8;
564 avg_mpcalls_dir += (mpcalls_dir * 1000 - avg_mpcalls_dir) >> 8;
565 avg_lockcalls_dir += (lockcalls_dir * 1000 - avg_lockcalls_dir) >> 8;
567 mtx_unlock_spin_flags(&cc->cc_lock, MTX_QUIET);
569 * swi_sched acquires the thread lock, so we don't want to call it
570 * with cc_lock held; incorrect locking order.
572 if (!TAILQ_EMPTY(&cc->cc_expireq))
573 swi_sched(cc->cc_cookie, 0);
576 static struct callout_cpu *
577 callout_lock(struct callout *c)
579 struct callout_cpu *cc;
585 if (cpu == CPUBLOCK) {
586 while (c->c_cpu == CPUBLOCK)
601 callout_cc_add(struct callout *c, struct callout_cpu *cc,
602 sbintime_t sbt, sbintime_t precision, void (*func)(void *),
603 void *arg, int cpu, int flags)
608 if (sbt < cc->cc_lastscan)
609 sbt = cc->cc_lastscan;
611 c->c_iflags |= CALLOUT_PENDING;
612 c->c_iflags &= ~CALLOUT_PROCESSED;
613 c->c_flags |= CALLOUT_ACTIVE;
614 if (flags & C_DIRECT_EXEC)
615 c->c_iflags |= CALLOUT_DIRECT;
618 c->c_precision = precision;
619 bucket = callout_get_bucket(c->c_time);
620 CTR3(KTR_CALLOUT, "precision set for %p: %d.%08x",
621 c, (int)(c->c_precision >> 32),
622 (u_int)(c->c_precision & 0xffffffff));
623 LIST_INSERT_HEAD(&cc->cc_callwheel[bucket], c, c_links.le);
624 if (cc->cc_bucket == bucket)
625 cc_exec_next(cc) = c;
626 #ifndef NO_EVENTTIMERS
628 * Inform the eventtimers(4) subsystem there's a new callout
629 * that has been inserted, but only if really required.
631 if (SBT_MAX - c->c_time < c->c_precision)
632 c->c_precision = SBT_MAX - c->c_time;
633 sbt = c->c_time + c->c_precision;
634 if (sbt < cc->cc_firstevent) {
635 cc->cc_firstevent = sbt;
636 cpu_new_callout(cpu, sbt, c->c_time);
642 callout_cc_del(struct callout *c, struct callout_cpu *cc)
645 if ((c->c_iflags & CALLOUT_LOCAL_ALLOC) == 0)
648 SLIST_INSERT_HEAD(&cc->cc_callfree, c, c_links.sle);
652 softclock_call_cc(struct callout *c, struct callout_cpu *cc,
653 #ifdef CALLOUT_PROFILING
654 int *mpcalls, int *lockcalls, int *gcalls,
658 struct rm_priotracker tracker;
659 callout_func_t *c_func, *drain;
661 struct lock_class *class;
662 struct lock_object *c_lock;
663 uintptr_t lock_status;
666 struct callout_cpu *new_cc;
667 callout_func_t *new_func;
670 sbintime_t new_prec, new_time;
672 #if defined(DIAGNOSTIC) || defined(CALLOUT_PROFILING)
673 sbintime_t sbt1, sbt2;
675 static sbintime_t maxdt = 2 * SBT_1MS; /* 2 msec */
676 static callout_func_t *lastfunc;
679 KASSERT((c->c_iflags & CALLOUT_PENDING) == CALLOUT_PENDING,
680 ("softclock_call_cc: pend %p %x", c, c->c_iflags));
681 KASSERT((c->c_flags & CALLOUT_ACTIVE) == CALLOUT_ACTIVE,
682 ("softclock_call_cc: act %p %x", c, c->c_flags));
683 class = (c->c_lock != NULL) ? LOCK_CLASS(c->c_lock) : NULL;
685 if (c->c_flags & CALLOUT_SHAREDLOCK) {
686 if (class == &lock_class_rm)
687 lock_status = (uintptr_t)&tracker;
694 c_iflags = c->c_iflags;
695 if (c->c_iflags & CALLOUT_LOCAL_ALLOC)
696 c->c_iflags = CALLOUT_LOCAL_ALLOC;
698 c->c_iflags &= ~CALLOUT_PENDING;
700 cc_exec_curr(cc, direct) = c;
701 cc_exec_last_func(cc, direct) = c_func;
702 cc_exec_last_arg(cc, direct) = c_arg;
703 cc_exec_cancel(cc, direct) = false;
704 cc_exec_drain(cc, direct) = NULL;
706 if (c_lock != NULL) {
707 class->lc_lock(c_lock, lock_status);
709 * The callout may have been cancelled
710 * while we switched locks.
712 if (cc_exec_cancel(cc, direct)) {
713 class->lc_unlock(c_lock);
716 /* The callout cannot be stopped now. */
717 cc_exec_cancel(cc, direct) = true;
718 if (c_lock == &Giant.lock_object) {
719 #ifdef CALLOUT_PROFILING
722 CTR3(KTR_CALLOUT, "callout giant %p func %p arg %p",
725 #ifdef CALLOUT_PROFILING
728 CTR3(KTR_CALLOUT, "callout lock %p func %p arg %p",
732 #ifdef CALLOUT_PROFILING
735 CTR3(KTR_CALLOUT, "callout %p func %p arg %p",
738 KTR_STATE3(KTR_SCHED, "callout", cc->cc_ktr_event_name, "running",
739 "func:%p", c_func, "arg:%p", c_arg, "direct:%d", direct);
740 #if defined(DIAGNOSTIC) || defined(CALLOUT_PROFILING)
743 THREAD_NO_SLEEPING();
744 SDT_PROBE1(callout_execute, , , callout__start, c);
746 SDT_PROBE1(callout_execute, , , callout__end, c);
747 THREAD_SLEEPING_OK();
748 #if defined(DIAGNOSTIC) || defined(CALLOUT_PROFILING)
752 if (lastfunc != c_func || sbt2 > maxdt * 2) {
755 "Expensive timeout(9) function: %p(%p) %jd.%09ld s\n",
756 c_func, c_arg, (intmax_t)ts2.tv_sec, ts2.tv_nsec);
762 KTR_STATE0(KTR_SCHED, "callout", cc->cc_ktr_event_name, "idle");
763 CTR1(KTR_CALLOUT, "callout %p finished", c);
764 if ((c_iflags & CALLOUT_RETURNUNLOCKED) == 0)
765 class->lc_unlock(c_lock);
768 KASSERT(cc_exec_curr(cc, direct) == c, ("mishandled cc_curr"));
769 cc_exec_curr(cc, direct) = NULL;
770 if (cc_exec_drain(cc, direct)) {
771 drain = cc_exec_drain(cc, direct);
772 cc_exec_drain(cc, direct) = NULL;
777 if (cc_exec_waiting(cc, direct)) {
779 * There is someone waiting for the
780 * callout to complete.
781 * If the callout was scheduled for
782 * migration just cancel it.
784 if (cc_cce_migrating(cc, direct)) {
785 cc_cce_cleanup(cc, direct);
788 * It should be assert here that the callout is not
789 * destroyed but that is not easy.
791 c->c_iflags &= ~CALLOUT_DFRMIGRATION;
793 cc_exec_waiting(cc, direct) = false;
795 wakeup(&cc_exec_waiting(cc, direct));
797 } else if (cc_cce_migrating(cc, direct)) {
798 KASSERT((c_iflags & CALLOUT_LOCAL_ALLOC) == 0,
799 ("Migrating legacy callout %p", c));
802 * If the callout was scheduled for
803 * migration just perform it now.
805 new_cpu = cc_migration_cpu(cc, direct);
806 new_time = cc_migration_time(cc, direct);
807 new_prec = cc_migration_prec(cc, direct);
808 new_func = cc_migration_func(cc, direct);
809 new_arg = cc_migration_arg(cc, direct);
810 cc_cce_cleanup(cc, direct);
813 * It should be assert here that the callout is not destroyed
814 * but that is not easy.
816 * As first thing, handle deferred callout stops.
818 if (!callout_migrating(c)) {
820 "deferred cancelled %p func %p arg %p",
821 c, new_func, new_arg);
822 callout_cc_del(c, cc);
825 c->c_iflags &= ~CALLOUT_DFRMIGRATION;
827 new_cc = callout_cpu_switch(c, cc, new_cpu);
828 flags = (direct) ? C_DIRECT_EXEC : 0;
829 callout_cc_add(c, new_cc, new_time, new_prec, new_func,
830 new_arg, new_cpu, flags);
834 panic("migration should not happen");
838 * If the current callout is locally allocated (from
839 * timeout(9)) then put it on the freelist.
841 * Note: we need to check the cached copy of c_iflags because
842 * if it was not local, then it's not safe to deref the
845 KASSERT((c_iflags & CALLOUT_LOCAL_ALLOC) == 0 ||
846 c->c_iflags == CALLOUT_LOCAL_ALLOC,
847 ("corrupted callout"));
848 if (c_iflags & CALLOUT_LOCAL_ALLOC)
849 callout_cc_del(c, cc);
853 * The callout mechanism is based on the work of Adam M. Costello and
854 * George Varghese, published in a technical report entitled "Redesigning
855 * the BSD Callout and Timer Facilities" and modified slightly for inclusion
856 * in FreeBSD by Justin T. Gibbs. The original work on the data structures
857 * used in this implementation was published by G. Varghese and T. Lauck in
858 * the paper "Hashed and Hierarchical Timing Wheels: Data Structures for
859 * the Efficient Implementation of a Timer Facility" in the Proceedings of
860 * the 11th ACM Annual Symposium on Operating Systems Principles,
861 * Austin, Texas Nov 1987.
865 * Software (low priority) clock interrupt.
866 * Run periodic events from timeout queue.
871 struct callout_cpu *cc;
873 #ifdef CALLOUT_PROFILING
874 int depth = 0, gcalls = 0, lockcalls = 0, mpcalls = 0;
877 cc = (struct callout_cpu *)arg;
879 while ((c = TAILQ_FIRST(&cc->cc_expireq)) != NULL) {
880 TAILQ_REMOVE(&cc->cc_expireq, c, c_links.tqe);
881 softclock_call_cc(c, cc,
882 #ifdef CALLOUT_PROFILING
883 &mpcalls, &lockcalls, &gcalls,
886 #ifdef CALLOUT_PROFILING
890 #ifdef CALLOUT_PROFILING
891 avg_depth += (depth * 1000 - avg_depth) >> 8;
892 avg_mpcalls += (mpcalls * 1000 - avg_mpcalls) >> 8;
893 avg_lockcalls += (lockcalls * 1000 - avg_lockcalls) >> 8;
894 avg_gcalls += (gcalls * 1000 - avg_gcalls) >> 8;
901 * Execute a function after a specified length of time.
904 * Cancel previous timeout function call.
906 * callout_handle_init --
907 * Initialize a handle so that using it with untimeout is benign.
909 * See AT&T BCI Driver Reference Manual for specification. This
910 * implementation differs from that one in that although an
911 * identification value is returned from timeout, the original
912 * arguments to timeout as well as the identifier are used to
913 * identify entries for untimeout.
915 struct callout_handle
916 timeout(timeout_t *ftn, void *arg, int to_ticks)
918 struct callout_cpu *cc;
920 struct callout_handle handle;
922 cc = CC_CPU(timeout_cpu);
924 /* Fill in the next free callout structure. */
925 new = SLIST_FIRST(&cc->cc_callfree);
927 /* XXX Attempt to malloc first */
928 panic("timeout table full");
929 SLIST_REMOVE_HEAD(&cc->cc_callfree, c_links.sle);
930 callout_reset(new, to_ticks, ftn, arg);
931 handle.callout = new;
938 untimeout(timeout_t *ftn, void *arg, struct callout_handle handle)
940 struct callout_cpu *cc;
943 * Check for a handle that was initialized
944 * by callout_handle_init, but never used
945 * for a real timeout.
947 if (handle.callout == NULL)
950 cc = callout_lock(handle.callout);
951 if (handle.callout->c_func == ftn && handle.callout->c_arg == arg)
952 callout_stop(handle.callout);
957 callout_handle_init(struct callout_handle *handle)
959 handle->callout = NULL;
963 callout_when(sbintime_t sbt, sbintime_t precision, int flags,
964 sbintime_t *res, sbintime_t *prec_res)
966 sbintime_t to_sbt, to_pr;
968 if ((flags & (C_ABSOLUTE | C_PRECALC)) != 0) {
970 *prec_res = precision;
973 if ((flags & C_HARDCLOCK) != 0 && sbt < tick_sbt)
975 if ((flags & C_HARDCLOCK) != 0 ||
976 #ifdef NO_EVENTTIMERS
977 sbt >= sbt_timethreshold) {
978 to_sbt = getsbinuptime();
980 /* Add safety belt for the case of hz > 1000. */
981 to_sbt += tc_tick_sbt - tick_sbt;
983 sbt >= sbt_tickthreshold) {
985 * Obtain the time of the last hardclock() call on
986 * this CPU directly from the kern_clocksource.c.
987 * This value is per-CPU, but it is equal for all
991 to_sbt = DPCPU_GET(hardclocktime);
994 to_sbt = DPCPU_GET(hardclocktime);
998 if (cold && to_sbt == 0)
999 to_sbt = sbinuptime();
1000 if ((flags & C_HARDCLOCK) == 0)
1003 to_sbt = sbinuptime();
1004 if (SBT_MAX - to_sbt < sbt)
1009 to_pr = ((C_PRELGET(flags) < 0) ? sbt >> tc_precexp :
1010 sbt >> C_PRELGET(flags));
1011 *prec_res = to_pr > precision ? to_pr : precision;
1015 * New interface; clients allocate their own callout structures.
1017 * callout_reset() - establish or change a timeout
1018 * callout_stop() - disestablish a timeout
1019 * callout_init() - initialize a callout structure so that it can
1020 * safely be passed to callout_reset() and callout_stop()
1022 * <sys/callout.h> defines three convenience macros:
1024 * callout_active() - returns truth if callout has not been stopped,
1025 * drained, or deactivated since the last time the callout was
1027 * callout_pending() - returns truth if callout is still waiting for timeout
1028 * callout_deactivate() - marks the callout as having been serviced
1031 callout_reset_sbt_on(struct callout *c, sbintime_t sbt, sbintime_t prec,
1032 callout_func_t *ftn, void *arg, int cpu, int flags)
1034 sbintime_t to_sbt, precision;
1035 struct callout_cpu *cc;
1036 int cancelled, direct;
1042 } else if ((cpu >= MAXCPU) ||
1043 ((CC_CPU(cpu))->cc_inited == 0)) {
1044 /* Invalid CPU spec */
1045 panic("Invalid CPU in callout %d", cpu);
1047 callout_when(sbt, prec, flags, &to_sbt, &precision);
1050 * This flag used to be added by callout_cc_add, but the
1051 * first time you call this we could end up with the
1052 * wrong direct flag if we don't do it before we add.
1054 if (flags & C_DIRECT_EXEC) {
1059 KASSERT(!direct || c->c_lock == NULL,
1060 ("%s: direct callout %p has lock", __func__, c));
1061 cc = callout_lock(c);
1063 * Don't allow migration of pre-allocated callouts lest they
1064 * become unbalanced or handle the case where the user does
1067 if ((c->c_iflags & CALLOUT_LOCAL_ALLOC) ||
1072 if (cc_exec_curr(cc, direct) == c) {
1074 * We're being asked to reschedule a callout which is
1075 * currently in progress. If there is a lock then we
1076 * can cancel the callout if it has not really started.
1078 if (c->c_lock != NULL && !cc_exec_cancel(cc, direct))
1079 cancelled = cc_exec_cancel(cc, direct) = true;
1080 if (cc_exec_waiting(cc, direct) || cc_exec_drain(cc, direct)) {
1082 * Someone has called callout_drain to kill this
1083 * callout. Don't reschedule.
1085 CTR4(KTR_CALLOUT, "%s %p func %p arg %p",
1086 cancelled ? "cancelled" : "failed to cancel",
1087 c, c->c_func, c->c_arg);
1092 if (callout_migrating(c)) {
1094 * This only occurs when a second callout_reset_sbt_on
1095 * is made after a previous one moved it into
1096 * deferred migration (below). Note we do *not* change
1097 * the prev_cpu even though the previous target may
1100 cc_migration_cpu(cc, direct) = cpu;
1101 cc_migration_time(cc, direct) = to_sbt;
1102 cc_migration_prec(cc, direct) = precision;
1103 cc_migration_func(cc, direct) = ftn;
1104 cc_migration_arg(cc, direct) = arg;
1111 if (c->c_iflags & CALLOUT_PENDING) {
1112 if ((c->c_iflags & CALLOUT_PROCESSED) == 0) {
1113 if (cc_exec_next(cc) == c)
1114 cc_exec_next(cc) = LIST_NEXT(c, c_links.le);
1115 LIST_REMOVE(c, c_links.le);
1117 TAILQ_REMOVE(&cc->cc_expireq, c, c_links.tqe);
1120 c->c_iflags &= ~ CALLOUT_PENDING;
1121 c->c_flags &= ~ CALLOUT_ACTIVE;
1126 * If the callout must migrate try to perform it immediately.
1127 * If the callout is currently running, just defer the migration
1128 * to a more appropriate moment.
1130 if (c->c_cpu != cpu) {
1131 if (cc_exec_curr(cc, direct) == c) {
1133 * Pending will have been removed since we are
1134 * actually executing the callout on another
1135 * CPU. That callout should be waiting on the
1136 * lock the caller holds. If we set both
1137 * active/and/pending after we return and the
1138 * lock on the executing callout proceeds, it
1139 * will then see pending is true and return.
1140 * At the return from the actual callout execution
1141 * the migration will occur in softclock_call_cc
1142 * and this new callout will be placed on the
1143 * new CPU via a call to callout_cpu_switch() which
1144 * will get the lock on the right CPU followed
1145 * by a call callout_cc_add() which will add it there.
1146 * (see above in softclock_call_cc()).
1148 cc_migration_cpu(cc, direct) = cpu;
1149 cc_migration_time(cc, direct) = to_sbt;
1150 cc_migration_prec(cc, direct) = precision;
1151 cc_migration_func(cc, direct) = ftn;
1152 cc_migration_arg(cc, direct) = arg;
1153 c->c_iflags |= (CALLOUT_DFRMIGRATION | CALLOUT_PENDING);
1154 c->c_flags |= CALLOUT_ACTIVE;
1156 "migration of %p func %p arg %p in %d.%08x to %u deferred",
1157 c, c->c_func, c->c_arg, (int)(to_sbt >> 32),
1158 (u_int)(to_sbt & 0xffffffff), cpu);
1162 cc = callout_cpu_switch(c, cc, cpu);
1166 callout_cc_add(c, cc, to_sbt, precision, ftn, arg, cpu, flags);
1167 CTR6(KTR_CALLOUT, "%sscheduled %p func %p arg %p in %d.%08x",
1168 cancelled ? "re" : "", c, c->c_func, c->c_arg, (int)(to_sbt >> 32),
1169 (u_int)(to_sbt & 0xffffffff));
1176 * Common idioms that can be optimized in the future.
1179 callout_schedule_on(struct callout *c, int to_ticks, int cpu)
1181 return callout_reset_on(c, to_ticks, c->c_func, c->c_arg, cpu);
1185 callout_schedule(struct callout *c, int to_ticks)
1187 return callout_reset_on(c, to_ticks, c->c_func, c->c_arg, c->c_cpu);
1191 _callout_stop_safe(struct callout *c, int flags, callout_func_t *drain)
1193 struct callout_cpu *cc, *old_cc;
1194 struct lock_class *class;
1195 int direct, sq_locked, use_lock;
1196 int cancelled, not_on_a_list;
1198 if ((flags & CS_DRAIN) != 0)
1199 WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK, c->c_lock,
1200 "calling %s", __func__);
1203 * Some old subsystems don't hold Giant while running a callout_stop(),
1204 * so just discard this check for the moment.
1206 if ((flags & CS_DRAIN) == 0 && c->c_lock != NULL) {
1207 if (c->c_lock == &Giant.lock_object)
1208 use_lock = mtx_owned(&Giant);
1211 class = LOCK_CLASS(c->c_lock);
1212 class->lc_assert(c->c_lock, LA_XLOCKED);
1216 if (c->c_iflags & CALLOUT_DIRECT) {
1224 cc = callout_lock(c);
1226 if ((c->c_iflags & (CALLOUT_DFRMIGRATION | CALLOUT_PENDING)) ==
1227 (CALLOUT_DFRMIGRATION | CALLOUT_PENDING) &&
1228 ((c->c_flags & CALLOUT_ACTIVE) == CALLOUT_ACTIVE)) {
1230 * Special case where this slipped in while we
1231 * were migrating *as* the callout is about to
1232 * execute. The caller probably holds the lock
1233 * the callout wants.
1235 * Get rid of the migration first. Then set
1236 * the flag that tells this code *not* to
1237 * try to remove it from any lists (its not
1238 * on one yet). When the callout wheel runs,
1239 * it will ignore this callout.
1241 c->c_iflags &= ~CALLOUT_PENDING;
1242 c->c_flags &= ~CALLOUT_ACTIVE;
1249 * If the callout was migrating while the callout cpu lock was
1250 * dropped, just drop the sleepqueue lock and check the states
1253 if (sq_locked != 0 && cc != old_cc) {
1256 sleepq_release(&cc_exec_waiting(old_cc, direct));
1261 panic("migration should not happen");
1266 * If the callout is running, try to stop it or drain it.
1268 if (cc_exec_curr(cc, direct) == c) {
1270 * Succeed we to stop it or not, we must clear the
1271 * active flag - this is what API users expect. If we're
1272 * draining and the callout is currently executing, first wait
1273 * until it finishes.
1275 if ((flags & CS_DRAIN) == 0)
1276 c->c_flags &= ~CALLOUT_ACTIVE;
1278 if ((flags & CS_DRAIN) != 0) {
1280 * The current callout is running (or just
1281 * about to run) and blocking is allowed, so
1282 * just wait for the current invocation to
1285 while (cc_exec_curr(cc, direct) == c) {
1287 * Use direct calls to sleepqueue interface
1288 * instead of cv/msleep in order to avoid
1289 * a LOR between cc_lock and sleepqueue
1290 * chain spinlocks. This piece of code
1291 * emulates a msleep_spin() call actually.
1293 * If we already have the sleepqueue chain
1294 * locked, then we can safely block. If we
1295 * don't already have it locked, however,
1296 * we have to drop the cc_lock to lock
1297 * it. This opens several races, so we
1298 * restart at the beginning once we have
1299 * both locks. If nothing has changed, then
1300 * we will end up back here with sq_locked
1306 &cc_exec_waiting(cc, direct));
1313 * Migration could be cancelled here, but
1314 * as long as it is still not sure when it
1315 * will be packed up, just let softclock()
1318 cc_exec_waiting(cc, direct) = true;
1322 &cc_exec_waiting(cc, direct),
1323 &cc->cc_lock.lock_object, "codrain",
1326 &cc_exec_waiting(cc, direct),
1331 /* Reacquire locks previously released. */
1335 c->c_flags &= ~CALLOUT_ACTIVE;
1336 } else if (use_lock &&
1337 !cc_exec_cancel(cc, direct) && (drain == NULL)) {
1340 * The current callout is waiting for its
1341 * lock which we hold. Cancel the callout
1342 * and return. After our caller drops the
1343 * lock, the callout will be skipped in
1344 * softclock(). This *only* works with a
1345 * callout_stop() *not* callout_drain() or
1346 * callout_async_drain().
1348 cc_exec_cancel(cc, direct) = true;
1349 CTR3(KTR_CALLOUT, "cancelled %p func %p arg %p",
1350 c, c->c_func, c->c_arg);
1351 KASSERT(!cc_cce_migrating(cc, direct),
1352 ("callout wrongly scheduled for migration"));
1353 if (callout_migrating(c)) {
1354 c->c_iflags &= ~CALLOUT_DFRMIGRATION;
1356 cc_migration_cpu(cc, direct) = CPUBLOCK;
1357 cc_migration_time(cc, direct) = 0;
1358 cc_migration_prec(cc, direct) = 0;
1359 cc_migration_func(cc, direct) = NULL;
1360 cc_migration_arg(cc, direct) = NULL;
1364 KASSERT(!sq_locked, ("sleepqueue chain locked"));
1366 } else if (callout_migrating(c)) {
1368 * The callout is currently being serviced
1369 * and the "next" callout is scheduled at
1370 * its completion with a migration. We remove
1371 * the migration flag so it *won't* get rescheduled,
1372 * but we can't stop the one thats running so
1375 c->c_iflags &= ~CALLOUT_DFRMIGRATION;
1378 * We can't call cc_cce_cleanup here since
1379 * if we do it will remove .ce_curr and
1380 * its still running. This will prevent a
1381 * reschedule of the callout when the
1382 * execution completes.
1384 cc_migration_cpu(cc, direct) = CPUBLOCK;
1385 cc_migration_time(cc, direct) = 0;
1386 cc_migration_prec(cc, direct) = 0;
1387 cc_migration_func(cc, direct) = NULL;
1388 cc_migration_arg(cc, direct) = NULL;
1390 CTR3(KTR_CALLOUT, "postponing stop %p func %p arg %p",
1391 c, c->c_func, c->c_arg);
1393 cc_exec_drain(cc, direct) = drain;
1396 return ((flags & CS_EXECUTING) != 0);
1398 CTR3(KTR_CALLOUT, "failed to stop %p func %p arg %p",
1399 c, c->c_func, c->c_arg);
1401 cc_exec_drain(cc, direct) = drain;
1403 KASSERT(!sq_locked, ("sleepqueue chain still locked"));
1404 cancelled = ((flags & CS_EXECUTING) != 0);
1409 sleepq_release(&cc_exec_waiting(cc, direct));
1411 if ((c->c_iflags & CALLOUT_PENDING) == 0) {
1412 CTR3(KTR_CALLOUT, "failed to stop %p func %p arg %p",
1413 c, c->c_func, c->c_arg);
1415 * For not scheduled and not executing callout return
1418 if (cc_exec_curr(cc, direct) != c)
1424 c->c_iflags &= ~CALLOUT_PENDING;
1425 c->c_flags &= ~CALLOUT_ACTIVE;
1427 CTR3(KTR_CALLOUT, "cancelled %p func %p arg %p",
1428 c, c->c_func, c->c_arg);
1429 if (not_on_a_list == 0) {
1430 if ((c->c_iflags & CALLOUT_PROCESSED) == 0) {
1431 if (cc_exec_next(cc) == c)
1432 cc_exec_next(cc) = LIST_NEXT(c, c_links.le);
1433 LIST_REMOVE(c, c_links.le);
1435 TAILQ_REMOVE(&cc->cc_expireq, c, c_links.tqe);
1438 callout_cc_del(c, cc);
1444 callout_init(struct callout *c, int mpsafe)
1446 bzero(c, sizeof *c);
1449 c->c_iflags = CALLOUT_RETURNUNLOCKED;
1451 c->c_lock = &Giant.lock_object;
1454 c->c_cpu = timeout_cpu;
1458 _callout_init_lock(struct callout *c, struct lock_object *lock, int flags)
1460 bzero(c, sizeof *c);
1462 KASSERT((flags & ~(CALLOUT_RETURNUNLOCKED | CALLOUT_SHAREDLOCK)) == 0,
1463 ("callout_init_lock: bad flags %d", flags));
1464 KASSERT(lock != NULL || (flags & CALLOUT_RETURNUNLOCKED) == 0,
1465 ("callout_init_lock: CALLOUT_RETURNUNLOCKED with no lock"));
1466 KASSERT(lock == NULL || !(LOCK_CLASS(lock)->lc_flags &
1467 (LC_SPINLOCK | LC_SLEEPABLE)), ("%s: invalid lock class",
1469 c->c_iflags = flags & (CALLOUT_RETURNUNLOCKED | CALLOUT_SHAREDLOCK);
1470 c->c_cpu = timeout_cpu;
1473 #ifdef APM_FIXUP_CALLTODO
1475 * Adjust the kernel calltodo timeout list. This routine is used after
1476 * an APM resume to recalculate the calltodo timer list values with the
1477 * number of hz's we have been sleeping. The next hardclock() will detect
1478 * that there are fired timers and run softclock() to execute them.
1480 * Please note, I have not done an exhaustive analysis of what code this
1481 * might break. I am motivated to have my select()'s and alarm()'s that
1482 * have expired during suspend firing upon resume so that the applications
1483 * which set the timer can do the maintanence the timer was for as close
1484 * as possible to the originally intended time. Testing this code for a
1485 * week showed that resuming from a suspend resulted in 22 to 25 timers
1486 * firing, which seemed independent on whether the suspend was 2 hours or
1487 * 2 days. Your milage may vary. - Ken Key <key@cs.utk.edu>
1490 adjust_timeout_calltodo(struct timeval *time_change)
1493 unsigned long delta_ticks;
1496 * How many ticks were we asleep?
1497 * (stolen from tvtohz()).
1500 /* Don't do anything */
1501 if (time_change->tv_sec < 0)
1503 else if (time_change->tv_sec <= LONG_MAX / 1000000)
1504 delta_ticks = howmany(time_change->tv_sec * 1000000 +
1505 time_change->tv_usec, tick) + 1;
1506 else if (time_change->tv_sec <= LONG_MAX / hz)
1507 delta_ticks = time_change->tv_sec * hz +
1508 howmany(time_change->tv_usec, tick) + 1;
1510 delta_ticks = LONG_MAX;
1512 if (delta_ticks > INT_MAX)
1513 delta_ticks = INT_MAX;
1516 * Now rip through the timer calltodo list looking for timers
1520 /* don't collide with softclock() */
1522 for (p = calltodo.c_next; p != NULL; p = p->c_next) {
1523 p->c_time -= delta_ticks;
1525 /* Break if the timer had more time on it than delta_ticks */
1529 /* take back the ticks the timer didn't use (p->c_time <= 0) */
1530 delta_ticks = -p->c_time;
1536 #endif /* APM_FIXUP_CALLTODO */
1539 flssbt(sbintime_t sbt)
1542 sbt += (uint64_t)sbt >> 1;
1543 if (sizeof(long) >= sizeof(sbintime_t))
1546 return (flsl(((uint64_t)sbt) >> 32) + 32);
1551 * Dump immediate statistic snapshot of the scheduled callouts.
1554 sysctl_kern_callout_stat(SYSCTL_HANDLER_ARGS)
1556 struct callout *tmp;
1557 struct callout_cpu *cc;
1558 struct callout_list *sc;
1559 sbintime_t maxpr, maxt, medpr, medt, now, spr, st, t;
1560 int ct[64], cpr[64], ccpbk[32];
1561 int error, val, i, count, tcum, pcum, maxc, c, medc;
1567 error = sysctl_handle_int(oidp, &val, 0, req);
1568 if (error != 0 || req->newptr == NULL)
1571 st = spr = maxt = maxpr = 0;
1572 bzero(ccpbk, sizeof(ccpbk));
1573 bzero(ct, sizeof(ct));
1574 bzero(cpr, sizeof(cpr));
1580 cc = CC_CPU(timeout_cpu);
1583 for (i = 0; i < callwheelsize; i++) {
1584 sc = &cc->cc_callwheel[i];
1586 LIST_FOREACH(tmp, sc, c_links.le) {
1588 t = tmp->c_time - now;
1592 spr += tmp->c_precision / SBT_1US;
1595 if (tmp->c_precision > maxpr)
1596 maxpr = tmp->c_precision;
1598 cpr[flssbt(tmp->c_precision)]++;
1602 ccpbk[fls(c + c / 2)]++;
1610 for (i = 0, tcum = 0; i < 64 && tcum < count / 2; i++)
1612 medt = (i >= 2) ? (((sbintime_t)1) << (i - 2)) : 0;
1613 for (i = 0, pcum = 0; i < 64 && pcum < count / 2; i++)
1615 medpr = (i >= 2) ? (((sbintime_t)1) << (i - 2)) : 0;
1616 for (i = 0, c = 0; i < 32 && c < count / 2; i++)
1618 medc = (i >= 2) ? (1 << (i - 2)) : 0;
1620 printf("Scheduled callouts statistic snapshot:\n");
1621 printf(" Callouts: %6d Buckets: %6d*%-3d Bucket size: 0.%06ds\n",
1622 count, callwheelsize, mp_ncpus, 1000000 >> CC_HASH_SHIFT);
1623 printf(" C/Bk: med %5d avg %6d.%06jd max %6d\n",
1625 count / callwheelsize / mp_ncpus,
1626 (uint64_t)count * 1000000 / callwheelsize / mp_ncpus % 1000000,
1628 printf(" Time: med %5jd.%06jds avg %6jd.%06jds max %6jd.%06jds\n",
1629 medt / SBT_1S, (medt & 0xffffffff) * 1000000 >> 32,
1630 (st / count) / 1000000, (st / count) % 1000000,
1631 maxt / SBT_1S, (maxt & 0xffffffff) * 1000000 >> 32);
1632 printf(" Prec: med %5jd.%06jds avg %6jd.%06jds max %6jd.%06jds\n",
1633 medpr / SBT_1S, (medpr & 0xffffffff) * 1000000 >> 32,
1634 (spr / count) / 1000000, (spr / count) % 1000000,
1635 maxpr / SBT_1S, (maxpr & 0xffffffff) * 1000000 >> 32);
1636 printf(" Distribution: \tbuckets\t time\t tcum\t"
1638 for (i = 0, tcum = pcum = 0; i < 64; i++) {
1639 if (ct[i] == 0 && cpr[i] == 0)
1641 t = (i != 0) ? (((sbintime_t)1) << (i - 1)) : 0;
1644 printf(" %10jd.%06jds\t 2**%d\t%7d\t%7d\t%7d\t%7d\n",
1645 t / SBT_1S, (t & 0xffffffff) * 1000000 >> 32,
1646 i - 1 - (32 - CC_HASH_SHIFT),
1647 ct[i], tcum, cpr[i], pcum);
1651 SYSCTL_PROC(_kern, OID_AUTO, callout_stat,
1652 CTLTYPE_INT | CTLFLAG_RW | CTLFLAG_MPSAFE,
1653 0, 0, sysctl_kern_callout_stat, "I",
1654 "Dump immediate statistic snapshot of the scheduled callouts");
1658 _show_callout(struct callout *c)
1661 db_printf("callout %p\n", c);
1662 #define C_DB_PRINTF(f, e) db_printf(" %s = " f "\n", #e, c->e);
1663 db_printf(" &c_links = %p\n", &(c->c_links));
1664 C_DB_PRINTF("%" PRId64, c_time);
1665 C_DB_PRINTF("%" PRId64, c_precision);
1666 C_DB_PRINTF("%p", c_arg);
1667 C_DB_PRINTF("%p", c_func);
1668 C_DB_PRINTF("%p", c_lock);
1669 C_DB_PRINTF("%#x", c_flags);
1670 C_DB_PRINTF("%#x", c_iflags);
1671 C_DB_PRINTF("%d", c_cpu);
1675 DB_SHOW_COMMAND(callout, db_show_callout)
1679 db_printf("usage: show callout <struct callout *>\n");
1683 _show_callout((struct callout *)addr);
1687 _show_last_callout(int cpu, int direct, const char *dirstr)
1689 struct callout_cpu *cc;
1693 func = cc_exec_last_func(cc, direct);
1694 arg = cc_exec_last_arg(cc, direct);
1695 db_printf("cpu %d last%s callout function: %p ", cpu, dirstr, func);
1696 db_printsym((db_expr_t)func, DB_STGY_ANY);
1697 db_printf("\ncpu %d last%s callout argument: %p\n", cpu, dirstr, arg);
1700 DB_SHOW_COMMAND(callout_last, db_show_callout_last)
1705 if (addr < 0 || addr > mp_maxid || CPU_ABSENT(addr)) {
1706 db_printf("no such cpu: %d\n", (int)addr);
1715 while (cpu <= last) {
1716 if (!CPU_ABSENT(cpu)) {
1717 _show_last_callout(cpu, 0, "");
1718 _show_last_callout(cpu, 1, " direct");