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 * 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.
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: releng/9.1/sys/kern/kern_timeout.c 235221 2012-05-10 10:16:21Z kib $");
39
40 #include "opt_kdtrace.h"
41
42 #include <sys/param.h>
43 #include <sys/systm.h>
44 #include <sys/bus.h>
45 #include <sys/callout.h>
46 #include <sys/condvar.h>
47 #include <sys/interrupt.h>
48 #include <sys/kernel.h>
49 #include <sys/ktr.h>
50 #include <sys/lock.h>
51 #include <sys/malloc.h>
52 #include <sys/mutex.h>
53 #include <sys/proc.h>
54 #include <sys/sdt.h>
55 #include <sys/sleepqueue.h>
56 #include <sys/sysctl.h>
57 #include <sys/smp.h>
58
59 #ifdef SMP
60 #include <machine/cpu.h>
61 #endif
62
63 SDT_PROVIDER_DEFINE(callout_execute);
64 SDT_PROBE_DEFINE(callout_execute, kernel, , callout_start, callout-start);
65 SDT_PROBE_ARGTYPE(callout_execute, kernel, , callout_start, 0,
66 "struct callout *");
67 SDT_PROBE_DEFINE(callout_execute, kernel, , callout_end, callout-end);
68 SDT_PROBE_ARGTYPE(callout_execute, kernel, , callout_end, 0,
69 "struct callout *");
70
71 static int avg_depth;
72 SYSCTL_INT(_debug, OID_AUTO, to_avg_depth, CTLFLAG_RD, &avg_depth, 0,
73 "Average number of items examined per softclock call. Units = 1/1000");
74 static int avg_gcalls;
75 SYSCTL_INT(_debug, OID_AUTO, to_avg_gcalls, CTLFLAG_RD, &avg_gcalls, 0,
76 "Average number of Giant callouts made per softclock call. Units = 1/1000");
77 static int avg_lockcalls;
78 SYSCTL_INT(_debug, OID_AUTO, to_avg_lockcalls, CTLFLAG_RD, &avg_lockcalls, 0,
79 "Average number of lock callouts made per softclock call. Units = 1/1000");
80 static int avg_mpcalls;
81 SYSCTL_INT(_debug, OID_AUTO, to_avg_mpcalls, CTLFLAG_RD, &avg_mpcalls, 0,
82 "Average number of MP callouts made per softclock call. Units = 1/1000");
83 /*
84 * TODO:
85 * allocate more timeout table slots when table overflows.
86 */
87 int callwheelsize, callwheelbits, callwheelmask;
88
89 /*
90 * The callout cpu migration entity represents informations necessary for
91 * describing the migrating callout to the new callout cpu.
92 * The cached informations are very important for deferring migration when
93 * the migrating callout is already running.
94 */
95 struct cc_mig_ent {
96 #ifdef SMP
97 void (*ce_migration_func)(void *);
98 void *ce_migration_arg;
99 int ce_migration_cpu;
100 int ce_migration_ticks;
101 #endif
102 };
103
104 /*
105 * There is one struct callout_cpu per cpu, holding all relevant
106 * state for the callout processing thread on the individual CPU.
107 * In particular:
108 * cc_ticks is incremented once per tick in callout_cpu().
109 * It tracks the global 'ticks' but in a way that the individual
110 * threads should not worry about races in the order in which
111 * hardclock() and hardclock_cpu() run on the various CPUs.
112 * cc_softclock is advanced in callout_cpu() to point to the
113 * first entry in cc_callwheel that may need handling. In turn,
114 * a softclock() is scheduled so it can serve the various entries i
115 * such that cc_softclock <= i <= cc_ticks .
116 * XXX maybe cc_softclock and cc_ticks should be volatile ?
117 *
118 * cc_ticks is also used in callout_reset_cpu() to determine
119 * when the callout should be served.
120 */
121 struct callout_cpu {
122 struct cc_mig_ent cc_migrating_entity;
123 struct mtx cc_lock;
124 struct callout *cc_callout;
125 struct callout_tailq *cc_callwheel;
126 struct callout_list cc_callfree;
127 struct callout *cc_next;
128 struct callout *cc_curr;
129 void *cc_cookie;
130 int cc_ticks;
131 int cc_softticks;
132 int cc_cancel;
133 int cc_waiting;
134 int cc_firsttick;
135 };
136
137 #ifdef SMP
138 #define cc_migration_func cc_migrating_entity.ce_migration_func
139 #define cc_migration_arg cc_migrating_entity.ce_migration_arg
140 #define cc_migration_cpu cc_migrating_entity.ce_migration_cpu
141 #define cc_migration_ticks cc_migrating_entity.ce_migration_ticks
142
143 struct callout_cpu cc_cpu[MAXCPU];
144 #define CPUBLOCK MAXCPU
145 #define CC_CPU(cpu) (&cc_cpu[(cpu)])
146 #define CC_SELF() CC_CPU(PCPU_GET(cpuid))
147 #else
148 struct callout_cpu cc_cpu;
149 #define CC_CPU(cpu) &cc_cpu
150 #define CC_SELF() &cc_cpu
151 #endif
152 #define CC_LOCK(cc) mtx_lock_spin(&(cc)->cc_lock)
153 #define CC_UNLOCK(cc) mtx_unlock_spin(&(cc)->cc_lock)
154 #define CC_LOCK_ASSERT(cc) mtx_assert(&(cc)->cc_lock, MA_OWNED)
155
156 static int timeout_cpu;
157 void (*callout_new_inserted)(int cpu, int ticks) = NULL;
158
159 MALLOC_DEFINE(M_CALLOUT, "callout", "Callout datastructures");
160
161 /**
162 * Locked by cc_lock:
163 * cc_curr - If a callout is in progress, it is curr_callout.
164 * If curr_callout is non-NULL, threads waiting in
165 * callout_drain() will be woken up as soon as the
166 * relevant callout completes.
167 * cc_cancel - Changing to 1 with both callout_lock and c_lock held
168 * guarantees that the current callout will not run.
169 * The softclock() function sets this to 0 before it
170 * drops callout_lock to acquire c_lock, and it calls
171 * the handler only if curr_cancelled is still 0 after
172 * c_lock is successfully acquired.
173 * cc_waiting - If a thread is waiting in callout_drain(), then
174 * callout_wait is nonzero. Set only when
175 * curr_callout is non-NULL.
176 */
177
178 /*
179 * Resets the migration entity tied to a specific callout cpu.
180 */
181 static void
182 cc_cme_cleanup(struct callout_cpu *cc)
183 {
184
185 #ifdef SMP
186 cc->cc_migration_cpu = CPUBLOCK;
187 cc->cc_migration_ticks = 0;
188 cc->cc_migration_func = NULL;
189 cc->cc_migration_arg = NULL;
190 #endif
191 }
192
193 /*
194 * Checks if migration is requested by a specific callout cpu.
195 */
196 static int
197 cc_cme_migrating(struct callout_cpu *cc)
198 {
199
200 #ifdef SMP
201 return (cc->cc_migration_cpu != CPUBLOCK);
202 #else
203 return (0);
204 #endif
205 }
206
207 /*
208 * kern_timeout_callwheel_alloc() - kernel low level callwheel initialization
209 *
210 * This code is called very early in the kernel initialization sequence,
211 * and may be called more then once.
212 */
213 caddr_t
214 kern_timeout_callwheel_alloc(caddr_t v)
215 {
216 struct callout_cpu *cc;
217
218 timeout_cpu = PCPU_GET(cpuid);
219 cc = CC_CPU(timeout_cpu);
220 /*
221 * Calculate callout wheel size
222 */
223 for (callwheelsize = 1, callwheelbits = 0;
224 callwheelsize < ncallout;
225 callwheelsize <<= 1, ++callwheelbits)
226 ;
227 callwheelmask = callwheelsize - 1;
228
229 cc->cc_callout = (struct callout *)v;
230 v = (caddr_t)(cc->cc_callout + ncallout);
231 cc->cc_callwheel = (struct callout_tailq *)v;
232 v = (caddr_t)(cc->cc_callwheel + callwheelsize);
233 return(v);
234 }
235
236 static void
237 callout_cpu_init(struct callout_cpu *cc)
238 {
239 struct callout *c;
240 int i;
241
242 mtx_init(&cc->cc_lock, "callout", NULL, MTX_SPIN | MTX_RECURSE);
243 SLIST_INIT(&cc->cc_callfree);
244 for (i = 0; i < callwheelsize; i++) {
245 TAILQ_INIT(&cc->cc_callwheel[i]);
246 }
247 cc_cme_cleanup(cc);
248 if (cc->cc_callout == NULL)
249 return;
250 for (i = 0; i < ncallout; i++) {
251 c = &cc->cc_callout[i];
252 callout_init(c, 0);
253 c->c_flags = CALLOUT_LOCAL_ALLOC;
254 SLIST_INSERT_HEAD(&cc->cc_callfree, c, c_links.sle);
255 }
256 }
257
258 #ifdef SMP
259 /*
260 * Switches the cpu tied to a specific callout.
261 * The function expects a locked incoming callout cpu and returns with
262 * locked outcoming callout cpu.
263 */
264 static struct callout_cpu *
265 callout_cpu_switch(struct callout *c, struct callout_cpu *cc, int new_cpu)
266 {
267 struct callout_cpu *new_cc;
268
269 MPASS(c != NULL && cc != NULL);
270 CC_LOCK_ASSERT(cc);
271
272 /*
273 * Avoid interrupts and preemption firing after the callout cpu
274 * is blocked in order to avoid deadlocks as the new thread
275 * may be willing to acquire the callout cpu lock.
276 */
277 c->c_cpu = CPUBLOCK;
278 spinlock_enter();
279 CC_UNLOCK(cc);
280 new_cc = CC_CPU(new_cpu);
281 CC_LOCK(new_cc);
282 spinlock_exit();
283 c->c_cpu = new_cpu;
284 return (new_cc);
285 }
286 #endif
287
288 /*
289 * kern_timeout_callwheel_init() - initialize previously reserved callwheel
290 * space.
291 *
292 * This code is called just once, after the space reserved for the
293 * callout wheel has been finalized.
294 */
295 void
296 kern_timeout_callwheel_init(void)
297 {
298 callout_cpu_init(CC_CPU(timeout_cpu));
299 }
300
301 /*
302 * Start standard softclock thread.
303 */
304 static void
305 start_softclock(void *dummy)
306 {
307 struct callout_cpu *cc;
308 #ifdef SMP
309 int cpu;
310 #endif
311
312 cc = CC_CPU(timeout_cpu);
313 if (swi_add(&clk_intr_event, "clock", softclock, cc, SWI_CLOCK,
314 INTR_MPSAFE, &cc->cc_cookie))
315 panic("died while creating standard software ithreads");
316 #ifdef SMP
317 CPU_FOREACH(cpu) {
318 if (cpu == timeout_cpu)
319 continue;
320 cc = CC_CPU(cpu);
321 if (swi_add(NULL, "clock", softclock, cc, SWI_CLOCK,
322 INTR_MPSAFE, &cc->cc_cookie))
323 panic("died while creating standard software ithreads");
324 cc->cc_callout = NULL; /* Only cpu0 handles timeout(). */
325 cc->cc_callwheel = malloc(
326 sizeof(struct callout_tailq) * callwheelsize, M_CALLOUT,
327 M_WAITOK);
328 callout_cpu_init(cc);
329 }
330 #endif
331 }
332
333 SYSINIT(start_softclock, SI_SUB_SOFTINTR, SI_ORDER_FIRST, start_softclock, NULL);
334
335 void
336 callout_tick(void)
337 {
338 struct callout_cpu *cc;
339 int need_softclock;
340 int bucket;
341
342 /*
343 * Process callouts at a very low cpu priority, so we don't keep the
344 * relatively high clock interrupt priority any longer than necessary.
345 */
346 need_softclock = 0;
347 cc = CC_SELF();
348 mtx_lock_spin_flags(&cc->cc_lock, MTX_QUIET);
349 cc->cc_firsttick = cc->cc_ticks = ticks;
350 for (; (cc->cc_softticks - cc->cc_ticks) <= 0; cc->cc_softticks++) {
351 bucket = cc->cc_softticks & callwheelmask;
352 if (!TAILQ_EMPTY(&cc->cc_callwheel[bucket])) {
353 need_softclock = 1;
354 break;
355 }
356 }
357 mtx_unlock_spin_flags(&cc->cc_lock, MTX_QUIET);
358 /*
359 * swi_sched acquires the thread lock, so we don't want to call it
360 * with cc_lock held; incorrect locking order.
361 */
362 if (need_softclock)
363 swi_sched(cc->cc_cookie, 0);
364 }
365
366 int
367 callout_tickstofirst(int limit)
368 {
369 struct callout_cpu *cc;
370 struct callout *c;
371 struct callout_tailq *sc;
372 int curticks;
373 int skip = 1;
374
375 cc = CC_SELF();
376 mtx_lock_spin_flags(&cc->cc_lock, MTX_QUIET);
377 curticks = cc->cc_ticks;
378 while( skip < ncallout && skip < limit ) {
379 sc = &cc->cc_callwheel[ (curticks+skip) & callwheelmask ];
380 /* search scanning ticks */
381 TAILQ_FOREACH( c, sc, c_links.tqe ){
382 if (c->c_time - curticks <= ncallout)
383 goto out;
384 }
385 skip++;
386 }
387 out:
388 cc->cc_firsttick = curticks + skip;
389 mtx_unlock_spin_flags(&cc->cc_lock, MTX_QUIET);
390 return (skip);
391 }
392
393 static struct callout_cpu *
394 callout_lock(struct callout *c)
395 {
396 struct callout_cpu *cc;
397 int cpu;
398
399 for (;;) {
400 cpu = c->c_cpu;
401 #ifdef SMP
402 if (cpu == CPUBLOCK) {
403 while (c->c_cpu == CPUBLOCK)
404 cpu_spinwait();
405 continue;
406 }
407 #endif
408 cc = CC_CPU(cpu);
409 CC_LOCK(cc);
410 if (cpu == c->c_cpu)
411 break;
412 CC_UNLOCK(cc);
413 }
414 return (cc);
415 }
416
417 static void
418 callout_cc_add(struct callout *c, struct callout_cpu *cc, int to_ticks,
419 void (*func)(void *), void *arg, int cpu)
420 {
421
422 CC_LOCK_ASSERT(cc);
423
424 if (to_ticks <= 0)
425 to_ticks = 1;
426 c->c_arg = arg;
427 c->c_flags |= (CALLOUT_ACTIVE | CALLOUT_PENDING);
428 c->c_func = func;
429 c->c_time = ticks + to_ticks;
430 TAILQ_INSERT_TAIL(&cc->cc_callwheel[c->c_time & callwheelmask],
431 c, c_links.tqe);
432 if ((c->c_time - cc->cc_firsttick) < 0 &&
433 callout_new_inserted != NULL) {
434 cc->cc_firsttick = c->c_time;
435 (*callout_new_inserted)(cpu,
436 to_ticks + (ticks - cc->cc_ticks));
437 }
438 }
439
440 static void
441 callout_cc_del(struct callout *c, struct callout_cpu *cc)
442 {
443
444 if (cc->cc_next == c)
445 cc->cc_next = TAILQ_NEXT(c, c_links.tqe);
446 if (c->c_flags & CALLOUT_LOCAL_ALLOC) {
447 c->c_func = NULL;
448 SLIST_INSERT_HEAD(&cc->cc_callfree, c, c_links.sle);
449 }
450 }
451
452 static struct callout *
453 softclock_call_cc(struct callout *c, struct callout_cpu *cc, int *mpcalls,
454 int *lockcalls, int *gcalls)
455 {
456 void (*c_func)(void *);
457 void *c_arg;
458 struct lock_class *class;
459 struct lock_object *c_lock;
460 int c_flags, sharedlock;
461 #ifdef SMP
462 struct callout_cpu *new_cc;
463 void (*new_func)(void *);
464 void *new_arg;
465 int new_cpu, new_ticks;
466 #endif
467 #ifdef DIAGNOSTIC
468 struct bintime bt1, bt2;
469 struct timespec ts2;
470 static uint64_t maxdt = 36893488147419102LL; /* 2 msec */
471 static timeout_t *lastfunc;
472 #endif
473
474 cc->cc_next = TAILQ_NEXT(c, c_links.tqe);
475 class = (c->c_lock != NULL) ? LOCK_CLASS(c->c_lock) : NULL;
476 sharedlock = (c->c_flags & CALLOUT_SHAREDLOCK) ? 0 : 1;
477 c_lock = c->c_lock;
478 c_func = c->c_func;
479 c_arg = c->c_arg;
480 c_flags = c->c_flags;
481 if (c->c_flags & CALLOUT_LOCAL_ALLOC)
482 c->c_flags = CALLOUT_LOCAL_ALLOC;
483 else
484 c->c_flags &= ~CALLOUT_PENDING;
485 cc->cc_curr = c;
486 cc->cc_cancel = 0;
487 CC_UNLOCK(cc);
488 if (c_lock != NULL) {
489 class->lc_lock(c_lock, sharedlock);
490 /*
491 * The callout may have been cancelled
492 * while we switched locks.
493 */
494 if (cc->cc_cancel) {
495 class->lc_unlock(c_lock);
496 goto skip;
497 }
498 /* The callout cannot be stopped now. */
499 cc->cc_cancel = 1;
500
501 if (c_lock == &Giant.lock_object) {
502 (*gcalls)++;
503 CTR3(KTR_CALLOUT, "callout %p func %p arg %p",
504 c, c_func, c_arg);
505 } else {
506 (*lockcalls)++;
507 CTR3(KTR_CALLOUT, "callout lock %p func %p arg %p",
508 c, c_func, c_arg);
509 }
510 } else {
511 (*mpcalls)++;
512 CTR3(KTR_CALLOUT, "callout mpsafe %p func %p arg %p",
513 c, c_func, c_arg);
514 }
515 #ifdef DIAGNOSTIC
516 binuptime(&bt1);
517 #endif
518 THREAD_NO_SLEEPING();
519 SDT_PROBE(callout_execute, kernel, , callout_start, c, 0, 0, 0, 0);
520 c_func(c_arg);
521 SDT_PROBE(callout_execute, kernel, , callout_end, c, 0, 0, 0, 0);
522 THREAD_SLEEPING_OK();
523 #ifdef DIAGNOSTIC
524 binuptime(&bt2);
525 bintime_sub(&bt2, &bt1);
526 if (bt2.frac > maxdt) {
527 if (lastfunc != c_func || bt2.frac > maxdt * 2) {
528 bintime2timespec(&bt2, &ts2);
529 printf(
530 "Expensive timeout(9) function: %p(%p) %jd.%09ld s\n",
531 c_func, c_arg, (intmax_t)ts2.tv_sec, ts2.tv_nsec);
532 }
533 maxdt = bt2.frac;
534 lastfunc = c_func;
535 }
536 #endif
537 CTR1(KTR_CALLOUT, "callout %p finished", c);
538 if ((c_flags & CALLOUT_RETURNUNLOCKED) == 0)
539 class->lc_unlock(c_lock);
540 skip:
541 CC_LOCK(cc);
542 /*
543 * If the current callout is locally allocated (from
544 * timeout(9)) then put it on the freelist.
545 *
546 * Note: we need to check the cached copy of c_flags because
547 * if it was not local, then it's not safe to deref the
548 * callout pointer.
549 */
550 if (c_flags & CALLOUT_LOCAL_ALLOC) {
551 KASSERT(c->c_flags == CALLOUT_LOCAL_ALLOC,
552 ("corrupted callout"));
553 c->c_func = NULL;
554 SLIST_INSERT_HEAD(&cc->cc_callfree, c, c_links.sle);
555 }
556 cc->cc_curr = NULL;
557 if (cc->cc_waiting) {
558 /*
559 * There is someone waiting for the
560 * callout to complete.
561 * If the callout was scheduled for
562 * migration just cancel it.
563 */
564 if (cc_cme_migrating(cc))
565 cc_cme_cleanup(cc);
566 cc->cc_waiting = 0;
567 CC_UNLOCK(cc);
568 wakeup(&cc->cc_waiting);
569 CC_LOCK(cc);
570 } else if (cc_cme_migrating(cc)) {
571 #ifdef SMP
572 /*
573 * If the callout was scheduled for
574 * migration just perform it now.
575 */
576 new_cpu = cc->cc_migration_cpu;
577 new_ticks = cc->cc_migration_ticks;
578 new_func = cc->cc_migration_func;
579 new_arg = cc->cc_migration_arg;
580 cc_cme_cleanup(cc);
581
582 /*
583 * Handle deferred callout stops
584 */
585 if ((c->c_flags & CALLOUT_DFRMIGRATION) == 0) {
586 CTR3(KTR_CALLOUT,
587 "deferred cancelled %p func %p arg %p",
588 c, new_func, new_arg);
589 callout_cc_del(c, cc);
590 goto nextc;
591 }
592
593 c->c_flags &= ~CALLOUT_DFRMIGRATION;
594
595 /*
596 * It should be assert here that the
597 * callout is not destroyed but that
598 * is not easy.
599 */
600 new_cc = callout_cpu_switch(c, cc, new_cpu);
601 callout_cc_add(c, new_cc, new_ticks, new_func, new_arg,
602 new_cpu);
603 CC_UNLOCK(new_cc);
604 CC_LOCK(cc);
605 #else
606 panic("migration should not happen");
607 #endif
608 }
609 #ifdef SMP
610 nextc:
611 #endif
612 return (cc->cc_next);
613 }
614
615 /*
616 * The callout mechanism is based on the work of Adam M. Costello and
617 * George Varghese, published in a technical report entitled "Redesigning
618 * the BSD Callout and Timer Facilities" and modified slightly for inclusion
619 * in FreeBSD by Justin T. Gibbs. The original work on the data structures
620 * used in this implementation was published by G. Varghese and T. Lauck in
621 * the paper "Hashed and Hierarchical Timing Wheels: Data Structures for
622 * the Efficient Implementation of a Timer Facility" in the Proceedings of
623 * the 11th ACM Annual Symposium on Operating Systems Principles,
624 * Austin, Texas Nov 1987.
625 */
626
627 /*
628 * Software (low priority) clock interrupt.
629 * Run periodic events from timeout queue.
630 */
631 void
632 softclock(void *arg)
633 {
634 struct callout_cpu *cc;
635 struct callout *c;
636 struct callout_tailq *bucket;
637 int curticks;
638 int steps; /* #steps since we last allowed interrupts */
639 int depth;
640 int mpcalls;
641 int lockcalls;
642 int gcalls;
643
644 #ifndef MAX_SOFTCLOCK_STEPS
645 #define MAX_SOFTCLOCK_STEPS 100 /* Maximum allowed value of steps. */
646 #endif /* MAX_SOFTCLOCK_STEPS */
647
648 mpcalls = 0;
649 lockcalls = 0;
650 gcalls = 0;
651 depth = 0;
652 steps = 0;
653 cc = (struct callout_cpu *)arg;
654 CC_LOCK(cc);
655 while (cc->cc_softticks - 1 != cc->cc_ticks) {
656 /*
657 * cc_softticks may be modified by hard clock, so cache
658 * it while we work on a given bucket.
659 */
660 curticks = cc->cc_softticks;
661 cc->cc_softticks++;
662 bucket = &cc->cc_callwheel[curticks & callwheelmask];
663 c = TAILQ_FIRST(bucket);
664 while (c != NULL) {
665 depth++;
666 if (c->c_time != curticks) {
667 c = TAILQ_NEXT(c, c_links.tqe);
668 ++steps;
669 if (steps >= MAX_SOFTCLOCK_STEPS) {
670 cc->cc_next = c;
671 /* Give interrupts a chance. */
672 CC_UNLOCK(cc);
673 ; /* nothing */
674 CC_LOCK(cc);
675 c = cc->cc_next;
676 steps = 0;
677 }
678 } else {
679 TAILQ_REMOVE(bucket, c, c_links.tqe);
680 c = softclock_call_cc(c, cc, &mpcalls,
681 &lockcalls, &gcalls);
682 steps = 0;
683 }
684 }
685 }
686 avg_depth += (depth * 1000 - avg_depth) >> 8;
687 avg_mpcalls += (mpcalls * 1000 - avg_mpcalls) >> 8;
688 avg_lockcalls += (lockcalls * 1000 - avg_lockcalls) >> 8;
689 avg_gcalls += (gcalls * 1000 - avg_gcalls) >> 8;
690 cc->cc_next = NULL;
691 CC_UNLOCK(cc);
692 }
693
694 /*
695 * timeout --
696 * Execute a function after a specified length of time.
697 *
698 * untimeout --
699 * Cancel previous timeout function call.
700 *
701 * callout_handle_init --
702 * Initialize a handle so that using it with untimeout is benign.
703 *
704 * See AT&T BCI Driver Reference Manual for specification. This
705 * implementation differs from that one in that although an
706 * identification value is returned from timeout, the original
707 * arguments to timeout as well as the identifier are used to
708 * identify entries for untimeout.
709 */
710 struct callout_handle
711 timeout(ftn, arg, to_ticks)
712 timeout_t *ftn;
713 void *arg;
714 int to_ticks;
715 {
716 struct callout_cpu *cc;
717 struct callout *new;
718 struct callout_handle handle;
719
720 cc = CC_CPU(timeout_cpu);
721 CC_LOCK(cc);
722 /* Fill in the next free callout structure. */
723 new = SLIST_FIRST(&cc->cc_callfree);
724 if (new == NULL)
725 /* XXX Attempt to malloc first */
726 panic("timeout table full");
727 SLIST_REMOVE_HEAD(&cc->cc_callfree, c_links.sle);
728 callout_reset(new, to_ticks, ftn, arg);
729 handle.callout = new;
730 CC_UNLOCK(cc);
731
732 return (handle);
733 }
734
735 void
736 untimeout(ftn, arg, handle)
737 timeout_t *ftn;
738 void *arg;
739 struct callout_handle handle;
740 {
741 struct callout_cpu *cc;
742
743 /*
744 * Check for a handle that was initialized
745 * by callout_handle_init, but never used
746 * for a real timeout.
747 */
748 if (handle.callout == NULL)
749 return;
750
751 cc = callout_lock(handle.callout);
752 if (handle.callout->c_func == ftn && handle.callout->c_arg == arg)
753 callout_stop(handle.callout);
754 CC_UNLOCK(cc);
755 }
756
757 void
758 callout_handle_init(struct callout_handle *handle)
759 {
760 handle->callout = NULL;
761 }
762
763 /*
764 * New interface; clients allocate their own callout structures.
765 *
766 * callout_reset() - establish or change a timeout
767 * callout_stop() - disestablish a timeout
768 * callout_init() - initialize a callout structure so that it can
769 * safely be passed to callout_reset() and callout_stop()
770 *
771 * <sys/callout.h> defines three convenience macros:
772 *
773 * callout_active() - returns truth if callout has not been stopped,
774 * drained, or deactivated since the last time the callout was
775 * reset.
776 * callout_pending() - returns truth if callout is still waiting for timeout
777 * callout_deactivate() - marks the callout as having been serviced
778 */
779 int
780 callout_reset_on(struct callout *c, int to_ticks, void (*ftn)(void *),
781 void *arg, int cpu)
782 {
783 struct callout_cpu *cc;
784 int cancelled = 0;
785
786 /*
787 * Don't allow migration of pre-allocated callouts lest they
788 * become unbalanced.
789 */
790 if (c->c_flags & CALLOUT_LOCAL_ALLOC)
791 cpu = c->c_cpu;
792 cc = callout_lock(c);
793 if (cc->cc_curr == c) {
794 /*
795 * We're being asked to reschedule a callout which is
796 * currently in progress. If there is a lock then we
797 * can cancel the callout if it has not really started.
798 */
799 if (c->c_lock != NULL && !cc->cc_cancel)
800 cancelled = cc->cc_cancel = 1;
801 if (cc->cc_waiting) {
802 /*
803 * Someone has called callout_drain to kill this
804 * callout. Don't reschedule.
805 */
806 CTR4(KTR_CALLOUT, "%s %p func %p arg %p",
807 cancelled ? "cancelled" : "failed to cancel",
808 c, c->c_func, c->c_arg);
809 CC_UNLOCK(cc);
810 return (cancelled);
811 }
812 }
813 if (c->c_flags & CALLOUT_PENDING) {
814 if (cc->cc_next == c) {
815 cc->cc_next = TAILQ_NEXT(c, c_links.tqe);
816 }
817 TAILQ_REMOVE(&cc->cc_callwheel[c->c_time & callwheelmask], c,
818 c_links.tqe);
819
820 cancelled = 1;
821 c->c_flags &= ~(CALLOUT_ACTIVE | CALLOUT_PENDING);
822 }
823
824 #ifdef SMP
825 /*
826 * If the callout must migrate try to perform it immediately.
827 * If the callout is currently running, just defer the migration
828 * to a more appropriate moment.
829 */
830 if (c->c_cpu != cpu) {
831 if (cc->cc_curr == c) {
832 cc->cc_migration_cpu = cpu;
833 cc->cc_migration_ticks = to_ticks;
834 cc->cc_migration_func = ftn;
835 cc->cc_migration_arg = arg;
836 c->c_flags |= CALLOUT_DFRMIGRATION;
837 CTR5(KTR_CALLOUT,
838 "migration of %p func %p arg %p in %d to %u deferred",
839 c, c->c_func, c->c_arg, to_ticks, cpu);
840 CC_UNLOCK(cc);
841 return (cancelled);
842 }
843 cc = callout_cpu_switch(c, cc, cpu);
844 }
845 #endif
846
847 callout_cc_add(c, cc, to_ticks, ftn, arg, cpu);
848 CTR5(KTR_CALLOUT, "%sscheduled %p func %p arg %p in %d",
849 cancelled ? "re" : "", c, c->c_func, c->c_arg, to_ticks);
850 CC_UNLOCK(cc);
851
852 return (cancelled);
853 }
854
855 /*
856 * Common idioms that can be optimized in the future.
857 */
858 int
859 callout_schedule_on(struct callout *c, int to_ticks, int cpu)
860 {
861 return callout_reset_on(c, to_ticks, c->c_func, c->c_arg, cpu);
862 }
863
864 int
865 callout_schedule(struct callout *c, int to_ticks)
866 {
867 return callout_reset_on(c, to_ticks, c->c_func, c->c_arg, c->c_cpu);
868 }
869
870 int
871 _callout_stop_safe(c, safe)
872 struct callout *c;
873 int safe;
874 {
875 struct callout_cpu *cc, *old_cc;
876 struct lock_class *class;
877 int use_lock, sq_locked;
878
879 /*
880 * Some old subsystems don't hold Giant while running a callout_stop(),
881 * so just discard this check for the moment.
882 */
883 if (!safe && c->c_lock != NULL) {
884 if (c->c_lock == &Giant.lock_object)
885 use_lock = mtx_owned(&Giant);
886 else {
887 use_lock = 1;
888 class = LOCK_CLASS(c->c_lock);
889 class->lc_assert(c->c_lock, LA_XLOCKED);
890 }
891 } else
892 use_lock = 0;
893
894 sq_locked = 0;
895 old_cc = NULL;
896 again:
897 cc = callout_lock(c);
898
899 /*
900 * If the callout was migrating while the callout cpu lock was
901 * dropped, just drop the sleepqueue lock and check the states
902 * again.
903 */
904 if (sq_locked != 0 && cc != old_cc) {
905 #ifdef SMP
906 CC_UNLOCK(cc);
907 sleepq_release(&old_cc->cc_waiting);
908 sq_locked = 0;
909 old_cc = NULL;
910 goto again;
911 #else
912 panic("migration should not happen");
913 #endif
914 }
915
916 /*
917 * If the callout isn't pending, it's not on the queue, so
918 * don't attempt to remove it from the queue. We can try to
919 * stop it by other means however.
920 */
921 if (!(c->c_flags & CALLOUT_PENDING)) {
922 c->c_flags &= ~CALLOUT_ACTIVE;
923
924 /*
925 * If it wasn't on the queue and it isn't the current
926 * callout, then we can't stop it, so just bail.
927 */
928 if (cc->cc_curr != c) {
929 CTR3(KTR_CALLOUT, "failed to stop %p func %p arg %p",
930 c, c->c_func, c->c_arg);
931 CC_UNLOCK(cc);
932 if (sq_locked)
933 sleepq_release(&cc->cc_waiting);
934 return (0);
935 }
936
937 if (safe) {
938 /*
939 * The current callout is running (or just
940 * about to run) and blocking is allowed, so
941 * just wait for the current invocation to
942 * finish.
943 */
944 while (cc->cc_curr == c) {
945
946 /*
947 * Use direct calls to sleepqueue interface
948 * instead of cv/msleep in order to avoid
949 * a LOR between cc_lock and sleepqueue
950 * chain spinlocks. This piece of code
951 * emulates a msleep_spin() call actually.
952 *
953 * If we already have the sleepqueue chain
954 * locked, then we can safely block. If we
955 * don't already have it locked, however,
956 * we have to drop the cc_lock to lock
957 * it. This opens several races, so we
958 * restart at the beginning once we have
959 * both locks. If nothing has changed, then
960 * we will end up back here with sq_locked
961 * set.
962 */
963 if (!sq_locked) {
964 CC_UNLOCK(cc);
965 sleepq_lock(&cc->cc_waiting);
966 sq_locked = 1;
967 old_cc = cc;
968 goto again;
969 }
970
971 /*
972 * Migration could be cancelled here, but
973 * as long as it is still not sure when it
974 * will be packed up, just let softclock()
975 * take care of it.
976 */
977 cc->cc_waiting = 1;
978 DROP_GIANT();
979 CC_UNLOCK(cc);
980 sleepq_add(&cc->cc_waiting,
981 &cc->cc_lock.lock_object, "codrain",
982 SLEEPQ_SLEEP, 0);
983 sleepq_wait(&cc->cc_waiting, 0);
984 sq_locked = 0;
985 old_cc = NULL;
986
987 /* Reacquire locks previously released. */
988 PICKUP_GIANT();
989 CC_LOCK(cc);
990 }
991 } else if (use_lock && !cc->cc_cancel) {
992 /*
993 * The current callout is waiting for its
994 * lock which we hold. Cancel the callout
995 * and return. After our caller drops the
996 * lock, the callout will be skipped in
997 * softclock().
998 */
999 cc->cc_cancel = 1;
1000 CTR3(KTR_CALLOUT, "cancelled %p func %p arg %p",
1001 c, c->c_func, c->c_arg);
1002 KASSERT(!cc_cme_migrating(cc),
1003 ("callout wrongly scheduled for migration"));
1004 CC_UNLOCK(cc);
1005 KASSERT(!sq_locked, ("sleepqueue chain locked"));
1006 return (1);
1007 } else if ((c->c_flags & CALLOUT_DFRMIGRATION) != 0) {
1008 c->c_flags &= ~CALLOUT_DFRMIGRATION;
1009 CTR3(KTR_CALLOUT, "postponing stop %p func %p arg %p",
1010 c, c->c_func, c->c_arg);
1011 CC_UNLOCK(cc);
1012 return (1);
1013 }
1014 CTR3(KTR_CALLOUT, "failed to stop %p func %p arg %p",
1015 c, c->c_func, c->c_arg);
1016 CC_UNLOCK(cc);
1017 KASSERT(!sq_locked, ("sleepqueue chain still locked"));
1018 return (0);
1019 }
1020 if (sq_locked)
1021 sleepq_release(&cc->cc_waiting);
1022
1023 c->c_flags &= ~(CALLOUT_ACTIVE | CALLOUT_PENDING);
1024
1025 CTR3(KTR_CALLOUT, "cancelled %p func %p arg %p",
1026 c, c->c_func, c->c_arg);
1027 TAILQ_REMOVE(&cc->cc_callwheel[c->c_time & callwheelmask], c,
1028 c_links.tqe);
1029 callout_cc_del(c, cc);
1030
1031 CC_UNLOCK(cc);
1032 return (1);
1033 }
1034
1035 void
1036 callout_init(c, mpsafe)
1037 struct callout *c;
1038 int mpsafe;
1039 {
1040 bzero(c, sizeof *c);
1041 if (mpsafe) {
1042 c->c_lock = NULL;
1043 c->c_flags = CALLOUT_RETURNUNLOCKED;
1044 } else {
1045 c->c_lock = &Giant.lock_object;
1046 c->c_flags = 0;
1047 }
1048 c->c_cpu = timeout_cpu;
1049 }
1050
1051 void
1052 _callout_init_lock(c, lock, flags)
1053 struct callout *c;
1054 struct lock_object *lock;
1055 int flags;
1056 {
1057 bzero(c, sizeof *c);
1058 c->c_lock = lock;
1059 KASSERT((flags & ~(CALLOUT_RETURNUNLOCKED | CALLOUT_SHAREDLOCK)) == 0,
1060 ("callout_init_lock: bad flags %d", flags));
1061 KASSERT(lock != NULL || (flags & CALLOUT_RETURNUNLOCKED) == 0,
1062 ("callout_init_lock: CALLOUT_RETURNUNLOCKED with no lock"));
1063 KASSERT(lock == NULL || !(LOCK_CLASS(lock)->lc_flags &
1064 (LC_SPINLOCK | LC_SLEEPABLE)), ("%s: invalid lock class",
1065 __func__));
1066 c->c_flags = flags & (CALLOUT_RETURNUNLOCKED | CALLOUT_SHAREDLOCK);
1067 c->c_cpu = timeout_cpu;
1068 }
1069
1070 #ifdef APM_FIXUP_CALLTODO
1071 /*
1072 * Adjust the kernel calltodo timeout list. This routine is used after
1073 * an APM resume to recalculate the calltodo timer list values with the
1074 * number of hz's we have been sleeping. The next hardclock() will detect
1075 * that there are fired timers and run softclock() to execute them.
1076 *
1077 * Please note, I have not done an exhaustive analysis of what code this
1078 * might break. I am motivated to have my select()'s and alarm()'s that
1079 * have expired during suspend firing upon resume so that the applications
1080 * which set the timer can do the maintanence the timer was for as close
1081 * as possible to the originally intended time. Testing this code for a
1082 * week showed that resuming from a suspend resulted in 22 to 25 timers
1083 * firing, which seemed independant on whether the suspend was 2 hours or
1084 * 2 days. Your milage may vary. - Ken Key <key@cs.utk.edu>
1085 */
1086 void
1087 adjust_timeout_calltodo(time_change)
1088 struct timeval *time_change;
1089 {
1090 register struct callout *p;
1091 unsigned long delta_ticks;
1092
1093 /*
1094 * How many ticks were we asleep?
1095 * (stolen from tvtohz()).
1096 */
1097
1098 /* Don't do anything */
1099 if (time_change->tv_sec < 0)
1100 return;
1101 else if (time_change->tv_sec <= LONG_MAX / 1000000)
1102 delta_ticks = (time_change->tv_sec * 1000000 +
1103 time_change->tv_usec + (tick - 1)) / tick + 1;
1104 else if (time_change->tv_sec <= LONG_MAX / hz)
1105 delta_ticks = time_change->tv_sec * hz +
1106 (time_change->tv_usec + (tick - 1)) / tick + 1;
1107 else
1108 delta_ticks = LONG_MAX;
1109
1110 if (delta_ticks > INT_MAX)
1111 delta_ticks = INT_MAX;
1112
1113 /*
1114 * Now rip through the timer calltodo list looking for timers
1115 * to expire.
1116 */
1117
1118 /* don't collide with softclock() */
1119 CC_LOCK(cc);
1120 for (p = calltodo.c_next; p != NULL; p = p->c_next) {
1121 p->c_time -= delta_ticks;
1122
1123 /* Break if the timer had more time on it than delta_ticks */
1124 if (p->c_time > 0)
1125 break;
1126
1127 /* take back the ticks the timer didn't use (p->c_time <= 0) */
1128 delta_ticks = -p->c_time;
1129 }
1130 CC_UNLOCK(cc);
1131
1132 return;
1133 }
1134 #endif /* APM_FIXUP_CALLTODO */
Cache object: db6cb871d2063610b3f2b895e5c69f6d
|