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("$FreeBSD$");
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_DEFINE1(callout_execute, kernel, , callout__start,
65 "struct callout *");
66 SDT_PROBE_DEFINE1(callout_execute, kernel, , callout__end,
67 "struct callout *");
68
69 static int avg_depth;
70 SYSCTL_INT(_debug, OID_AUTO, to_avg_depth, CTLFLAG_RD, &avg_depth, 0,
71 "Average number of items examined per softclock call. Units = 1/1000");
72 static int avg_gcalls;
73 SYSCTL_INT(_debug, OID_AUTO, to_avg_gcalls, CTLFLAG_RD, &avg_gcalls, 0,
74 "Average number of Giant callouts made per softclock call. Units = 1/1000");
75 static int avg_lockcalls;
76 SYSCTL_INT(_debug, OID_AUTO, to_avg_lockcalls, CTLFLAG_RD, &avg_lockcalls, 0,
77 "Average number of lock callouts made per softclock call. Units = 1/1000");
78 static int avg_mpcalls;
79 SYSCTL_INT(_debug, OID_AUTO, to_avg_mpcalls, CTLFLAG_RD, &avg_mpcalls, 0,
80 "Average number of MP callouts made per softclock call. Units = 1/1000");
81 /*
82 * TODO:
83 * allocate more timeout table slots when table overflows.
84 */
85 int callwheelsize, callwheelbits, callwheelmask;
86
87 /*
88 * The callout cpu migration entity represents informations necessary for
89 * describing the migrating callout to the new callout cpu.
90 * The cached informations are very important for deferring migration when
91 * the migrating callout is already running.
92 */
93 struct cc_mig_ent {
94 #ifdef SMP
95 void (*ce_migration_func)(void *);
96 void *ce_migration_arg;
97 int ce_migration_cpu;
98 int ce_migration_ticks;
99 #endif
100 };
101
102 /*
103 * There is one struct callout_cpu per cpu, holding all relevant
104 * state for the callout processing thread on the individual CPU.
105 * In particular:
106 * cc_ticks is incremented once per tick in callout_cpu().
107 * It tracks the global 'ticks' but in a way that the individual
108 * threads should not worry about races in the order in which
109 * hardclock() and hardclock_cpu() run on the various CPUs.
110 * cc_softclock is advanced in callout_cpu() to point to the
111 * first entry in cc_callwheel that may need handling. In turn,
112 * a softclock() is scheduled so it can serve the various entries i
113 * such that cc_softclock <= i <= cc_ticks .
114 * XXX maybe cc_softclock and cc_ticks should be volatile ?
115 *
116 * cc_ticks is also used in callout_reset_cpu() to determine
117 * when the callout should be served.
118 */
119 struct callout_cpu {
120 struct cc_mig_ent cc_migrating_entity;
121 struct mtx cc_lock;
122 struct callout *cc_callout;
123 struct callout_tailq *cc_callwheel;
124 struct callout_list cc_callfree;
125 struct callout *cc_next;
126 struct callout *cc_curr;
127 void *cc_cookie;
128 int cc_ticks;
129 int cc_softticks;
130 int cc_cancel;
131 int cc_waiting;
132 int cc_firsttick;
133 };
134
135 #ifdef SMP
136 #define cc_migration_func cc_migrating_entity.ce_migration_func
137 #define cc_migration_arg cc_migrating_entity.ce_migration_arg
138 #define cc_migration_cpu cc_migrating_entity.ce_migration_cpu
139 #define cc_migration_ticks cc_migrating_entity.ce_migration_ticks
140
141 struct callout_cpu cc_cpu[MAXCPU];
142 #define CPUBLOCK MAXCPU
143 #define CC_CPU(cpu) (&cc_cpu[(cpu)])
144 #define CC_SELF() CC_CPU(PCPU_GET(cpuid))
145 #else
146 struct callout_cpu cc_cpu;
147 #define CC_CPU(cpu) &cc_cpu
148 #define CC_SELF() &cc_cpu
149 #endif
150 #define CC_LOCK(cc) mtx_lock_spin(&(cc)->cc_lock)
151 #define CC_UNLOCK(cc) mtx_unlock_spin(&(cc)->cc_lock)
152 #define CC_LOCK_ASSERT(cc) mtx_assert(&(cc)->cc_lock, MA_OWNED)
153
154 static int timeout_cpu;
155 void (*callout_new_inserted)(int cpu, int ticks) = NULL;
156
157 static MALLOC_DEFINE(M_CALLOUT, "callout", "Callout datastructures");
158
159 /**
160 * Locked by cc_lock:
161 * cc_curr - If a callout is in progress, it is curr_callout.
162 * If curr_callout is non-NULL, threads waiting in
163 * callout_drain() will be woken up as soon as the
164 * relevant callout completes.
165 * cc_cancel - Changing to 1 with both callout_lock and c_lock held
166 * guarantees that the current callout will not run.
167 * The softclock() function sets this to 0 before it
168 * drops callout_lock to acquire c_lock, and it calls
169 * the handler only if curr_cancelled is still 0 after
170 * c_lock is successfully acquired.
171 * cc_waiting - If a thread is waiting in callout_drain(), then
172 * callout_wait is nonzero. Set only when
173 * curr_callout is non-NULL.
174 */
175
176 /*
177 * Resets the migration entity tied to a specific callout cpu.
178 */
179 static void
180 cc_cme_cleanup(struct callout_cpu *cc)
181 {
182
183 #ifdef SMP
184 cc->cc_migration_cpu = CPUBLOCK;
185 cc->cc_migration_ticks = 0;
186 cc->cc_migration_func = NULL;
187 cc->cc_migration_arg = NULL;
188 #endif
189 }
190
191 /*
192 * Checks if migration is requested by a specific callout cpu.
193 */
194 static int
195 cc_cme_migrating(struct callout_cpu *cc)
196 {
197
198 #ifdef SMP
199 return (cc->cc_migration_cpu != CPUBLOCK);
200 #else
201 return (0);
202 #endif
203 }
204
205 /*
206 * kern_timeout_callwheel_alloc() - kernel low level callwheel initialization
207 *
208 * This code is called very early in the kernel initialization sequence,
209 * and may be called more then once.
210 */
211 caddr_t
212 kern_timeout_callwheel_alloc(caddr_t v)
213 {
214 struct callout_cpu *cc;
215
216 timeout_cpu = PCPU_GET(cpuid);
217 cc = CC_CPU(timeout_cpu);
218 /*
219 * Calculate callout wheel size
220 */
221 for (callwheelsize = 1, callwheelbits = 0;
222 callwheelsize < ncallout;
223 callwheelsize <<= 1, ++callwheelbits)
224 ;
225 callwheelmask = callwheelsize - 1;
226
227 cc->cc_callout = (struct callout *)v;
228 v = (caddr_t)(cc->cc_callout + ncallout);
229 cc->cc_callwheel = (struct callout_tailq *)v;
230 v = (caddr_t)(cc->cc_callwheel + callwheelsize);
231 return(v);
232 }
233
234 static void
235 callout_cpu_init(struct callout_cpu *cc)
236 {
237 struct callout *c;
238 int i;
239
240 mtx_init(&cc->cc_lock, "callout", NULL, MTX_SPIN | MTX_RECURSE);
241 SLIST_INIT(&cc->cc_callfree);
242 for (i = 0; i < callwheelsize; i++) {
243 TAILQ_INIT(&cc->cc_callwheel[i]);
244 }
245 cc_cme_cleanup(cc);
246 if (cc->cc_callout == NULL)
247 return;
248 for (i = 0; i < ncallout; i++) {
249 c = &cc->cc_callout[i];
250 callout_init(c, 0);
251 c->c_flags = CALLOUT_LOCAL_ALLOC;
252 SLIST_INSERT_HEAD(&cc->cc_callfree, c, c_links.sle);
253 }
254 }
255
256 #ifdef SMP
257 /*
258 * Switches the cpu tied to a specific callout.
259 * The function expects a locked incoming callout cpu and returns with
260 * locked outcoming callout cpu.
261 */
262 static struct callout_cpu *
263 callout_cpu_switch(struct callout *c, struct callout_cpu *cc, int new_cpu)
264 {
265 struct callout_cpu *new_cc;
266
267 MPASS(c != NULL && cc != NULL);
268 CC_LOCK_ASSERT(cc);
269
270 /*
271 * Avoid interrupts and preemption firing after the callout cpu
272 * is blocked in order to avoid deadlocks as the new thread
273 * may be willing to acquire the callout cpu lock.
274 */
275 c->c_cpu = CPUBLOCK;
276 spinlock_enter();
277 CC_UNLOCK(cc);
278 new_cc = CC_CPU(new_cpu);
279 CC_LOCK(new_cc);
280 spinlock_exit();
281 c->c_cpu = new_cpu;
282 return (new_cc);
283 }
284 #endif
285
286 /*
287 * kern_timeout_callwheel_init() - initialize previously reserved callwheel
288 * space.
289 *
290 * This code is called just once, after the space reserved for the
291 * callout wheel has been finalized.
292 */
293 void
294 kern_timeout_callwheel_init(void)
295 {
296 callout_cpu_init(CC_CPU(timeout_cpu));
297 }
298
299 /*
300 * Start standard softclock thread.
301 */
302 static void
303 start_softclock(void *dummy)
304 {
305 struct callout_cpu *cc;
306 #ifdef SMP
307 int cpu;
308 #endif
309
310 cc = CC_CPU(timeout_cpu);
311 if (swi_add(&clk_intr_event, "clock", softclock, cc, SWI_CLOCK,
312 INTR_MPSAFE, &cc->cc_cookie))
313 panic("died while creating standard software ithreads");
314 #ifdef SMP
315 CPU_FOREACH(cpu) {
316 if (cpu == timeout_cpu)
317 continue;
318 cc = CC_CPU(cpu);
319 if (swi_add(NULL, "clock", softclock, cc, SWI_CLOCK,
320 INTR_MPSAFE, &cc->cc_cookie))
321 panic("died while creating standard software ithreads");
322 cc->cc_callout = NULL; /* Only cpu0 handles timeout(). */
323 cc->cc_callwheel = malloc(
324 sizeof(struct callout_tailq) * callwheelsize, M_CALLOUT,
325 M_WAITOK);
326 callout_cpu_init(cc);
327 }
328 #endif
329 }
330
331 SYSINIT(start_softclock, SI_SUB_SOFTINTR, SI_ORDER_FIRST, start_softclock, NULL);
332
333 void
334 callout_tick(void)
335 {
336 struct callout_cpu *cc;
337 int need_softclock;
338 int bucket;
339
340 /*
341 * Process callouts at a very low cpu priority, so we don't keep the
342 * relatively high clock interrupt priority any longer than necessary.
343 */
344 need_softclock = 0;
345 cc = CC_SELF();
346 mtx_lock_spin_flags(&cc->cc_lock, MTX_QUIET);
347 cc->cc_firsttick = cc->cc_ticks = ticks;
348 for (; (cc->cc_softticks - cc->cc_ticks) <= 0; cc->cc_softticks++) {
349 bucket = cc->cc_softticks & callwheelmask;
350 if (!TAILQ_EMPTY(&cc->cc_callwheel[bucket])) {
351 need_softclock = 1;
352 break;
353 }
354 }
355 mtx_unlock_spin_flags(&cc->cc_lock, MTX_QUIET);
356 /*
357 * swi_sched acquires the thread lock, so we don't want to call it
358 * with cc_lock held; incorrect locking order.
359 */
360 if (need_softclock)
361 swi_sched(cc->cc_cookie, 0);
362 }
363
364 int
365 callout_tickstofirst(int limit)
366 {
367 struct callout_cpu *cc;
368 struct callout *c;
369 struct callout_tailq *sc;
370 int curticks;
371 int skip = 1;
372
373 cc = CC_SELF();
374 mtx_lock_spin_flags(&cc->cc_lock, MTX_QUIET);
375 curticks = cc->cc_ticks;
376 while( skip < ncallout && skip < limit ) {
377 sc = &cc->cc_callwheel[ (curticks+skip) & callwheelmask ];
378 /* search scanning ticks */
379 TAILQ_FOREACH( c, sc, c_links.tqe ){
380 if (c->c_time - curticks <= ncallout)
381 goto out;
382 }
383 skip++;
384 }
385 out:
386 cc->cc_firsttick = curticks + skip;
387 mtx_unlock_spin_flags(&cc->cc_lock, MTX_QUIET);
388 return (skip);
389 }
390
391 static struct callout_cpu *
392 callout_lock(struct callout *c)
393 {
394 struct callout_cpu *cc;
395 int cpu;
396
397 for (;;) {
398 cpu = c->c_cpu;
399 #ifdef SMP
400 if (cpu == CPUBLOCK) {
401 while (c->c_cpu == CPUBLOCK)
402 cpu_spinwait();
403 continue;
404 }
405 #endif
406 cc = CC_CPU(cpu);
407 CC_LOCK(cc);
408 if (cpu == c->c_cpu)
409 break;
410 CC_UNLOCK(cc);
411 }
412 return (cc);
413 }
414
415 static void
416 callout_cc_add(struct callout *c, struct callout_cpu *cc, int to_ticks,
417 void (*func)(void *), void *arg, int cpu)
418 {
419
420 CC_LOCK_ASSERT(cc);
421
422 if (to_ticks <= 0)
423 to_ticks = 1;
424 c->c_arg = arg;
425 c->c_flags |= (CALLOUT_ACTIVE | CALLOUT_PENDING);
426 c->c_func = func;
427 c->c_time = ticks + to_ticks;
428 TAILQ_INSERT_TAIL(&cc->cc_callwheel[c->c_time & callwheelmask],
429 c, c_links.tqe);
430 if ((c->c_time - cc->cc_firsttick) < 0 &&
431 callout_new_inserted != NULL) {
432 cc->cc_firsttick = c->c_time;
433 (*callout_new_inserted)(cpu,
434 to_ticks + (ticks - cc->cc_ticks));
435 }
436 }
437
438 static void
439 callout_cc_del(struct callout *c, struct callout_cpu *cc)
440 {
441
442 if ((c->c_flags & CALLOUT_LOCAL_ALLOC) == 0)
443 return;
444 c->c_func = NULL;
445 SLIST_INSERT_HEAD(&cc->cc_callfree, c, c_links.sle);
446 }
447
448 static void
449 softclock_call_cc(struct callout *c, struct callout_cpu *cc, int *mpcalls,
450 int *lockcalls, int *gcalls)
451 {
452 void (*c_func)(void *);
453 void *c_arg;
454 struct lock_class *class;
455 struct lock_object *c_lock;
456 int c_flags, sharedlock;
457 #ifdef SMP
458 struct callout_cpu *new_cc;
459 void (*new_func)(void *);
460 void *new_arg;
461 int new_cpu, new_ticks;
462 #endif
463 #ifdef DIAGNOSTIC
464 struct bintime bt1, bt2;
465 struct timespec ts2;
466 static uint64_t maxdt = 36893488147419102LL; /* 2 msec */
467 static timeout_t *lastfunc;
468 #endif
469
470 KASSERT((c->c_flags & (CALLOUT_PENDING | CALLOUT_ACTIVE)) ==
471 (CALLOUT_PENDING | CALLOUT_ACTIVE),
472 ("softclock_call_cc: pend|act %p %x", c, c->c_flags));
473 class = (c->c_lock != NULL) ? LOCK_CLASS(c->c_lock) : NULL;
474 sharedlock = (c->c_flags & CALLOUT_SHAREDLOCK) ? 0 : 1;
475 c_lock = c->c_lock;
476 c_func = c->c_func;
477 c_arg = c->c_arg;
478 c_flags = c->c_flags;
479 if (c->c_flags & CALLOUT_LOCAL_ALLOC)
480 c->c_flags = CALLOUT_LOCAL_ALLOC;
481 else
482 c->c_flags &= ~CALLOUT_PENDING;
483 cc->cc_curr = c;
484 cc->cc_cancel = 0;
485 CC_UNLOCK(cc);
486 if (c_lock != NULL) {
487 class->lc_lock(c_lock, sharedlock);
488 /*
489 * The callout may have been cancelled
490 * while we switched locks.
491 */
492 if (cc->cc_cancel) {
493 class->lc_unlock(c_lock);
494 goto skip;
495 }
496 /* The callout cannot be stopped now. */
497 cc->cc_cancel = 1;
498
499 if (c_lock == &Giant.lock_object) {
500 (*gcalls)++;
501 CTR3(KTR_CALLOUT, "callout %p func %p arg %p",
502 c, c_func, c_arg);
503 } else {
504 (*lockcalls)++;
505 CTR3(KTR_CALLOUT, "callout lock %p func %p arg %p",
506 c, c_func, c_arg);
507 }
508 } else {
509 (*mpcalls)++;
510 CTR3(KTR_CALLOUT, "callout mpsafe %p func %p arg %p",
511 c, c_func, c_arg);
512 }
513 #ifdef DIAGNOSTIC
514 binuptime(&bt1);
515 #endif
516 THREAD_NO_SLEEPING();
517 SDT_PROBE1(callout_execute, kernel, , callout__start, c);
518 c_func(c_arg);
519 SDT_PROBE1(callout_execute, kernel, , callout__end, c);
520 THREAD_SLEEPING_OK();
521 #ifdef DIAGNOSTIC
522 binuptime(&bt2);
523 bintime_sub(&bt2, &bt1);
524 if (bt2.frac > maxdt) {
525 if (lastfunc != c_func || bt2.frac > maxdt * 2) {
526 bintime2timespec(&bt2, &ts2);
527 printf(
528 "Expensive timeout(9) function: %p(%p) %jd.%09ld s\n",
529 c_func, c_arg, (intmax_t)ts2.tv_sec, ts2.tv_nsec);
530 }
531 maxdt = bt2.frac;
532 lastfunc = c_func;
533 }
534 #endif
535 CTR1(KTR_CALLOUT, "callout %p finished", c);
536 if ((c_flags & CALLOUT_RETURNUNLOCKED) == 0)
537 class->lc_unlock(c_lock);
538 skip:
539 CC_LOCK(cc);
540 KASSERT(cc->cc_curr == c, ("mishandled cc_curr"));
541 cc->cc_curr = NULL;
542 if (cc->cc_waiting) {
543 /*
544 * There is someone waiting for the
545 * callout to complete.
546 * If the callout was scheduled for
547 * migration just cancel it.
548 */
549 if (cc_cme_migrating(cc)) {
550 cc_cme_cleanup(cc);
551
552 /*
553 * It should be assert here that the callout is not
554 * destroyed but that is not easy.
555 */
556 c->c_flags &= ~CALLOUT_DFRMIGRATION;
557 }
558 cc->cc_waiting = 0;
559 CC_UNLOCK(cc);
560 wakeup(&cc->cc_waiting);
561 CC_LOCK(cc);
562 } else if (cc_cme_migrating(cc)) {
563 KASSERT((c_flags & CALLOUT_LOCAL_ALLOC) == 0,
564 ("Migrating legacy callout %p", c));
565 #ifdef SMP
566 /*
567 * If the callout was scheduled for
568 * migration just perform it now.
569 */
570 new_cpu = cc->cc_migration_cpu;
571 new_ticks = cc->cc_migration_ticks;
572 new_func = cc->cc_migration_func;
573 new_arg = cc->cc_migration_arg;
574 cc_cme_cleanup(cc);
575
576 /*
577 * It should be assert here that the callout is not destroyed
578 * but that is not easy.
579 *
580 * As first thing, handle deferred callout stops.
581 */
582 if ((c->c_flags & CALLOUT_DFRMIGRATION) == 0) {
583 CTR3(KTR_CALLOUT,
584 "deferred cancelled %p func %p arg %p",
585 c, new_func, new_arg);
586 callout_cc_del(c, cc);
587 return;
588 }
589 c->c_flags &= ~CALLOUT_DFRMIGRATION;
590
591 new_cc = callout_cpu_switch(c, cc, new_cpu);
592 callout_cc_add(c, new_cc, new_ticks, new_func, new_arg,
593 new_cpu);
594 CC_UNLOCK(new_cc);
595 CC_LOCK(cc);
596 #else
597 panic("migration should not happen");
598 #endif
599 }
600 /*
601 * If the current callout is locally allocated (from
602 * timeout(9)) then put it on the freelist.
603 *
604 * Note: we need to check the cached copy of c_flags because
605 * if it was not local, then it's not safe to deref the
606 * callout pointer.
607 */
608 KASSERT((c_flags & CALLOUT_LOCAL_ALLOC) == 0 ||
609 c->c_flags == CALLOUT_LOCAL_ALLOC,
610 ("corrupted callout"));
611 if (c_flags & CALLOUT_LOCAL_ALLOC)
612 callout_cc_del(c, cc);
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 cc->cc_next = TAILQ_NEXT(c, c_links.tqe);
680 TAILQ_REMOVE(bucket, c, c_links.tqe);
681 softclock_call_cc(c, cc, &mpcalls,
682 &lockcalls, &gcalls);
683 steps = 0;
684 c = cc->cc_next;
685 }
686 }
687 }
688 avg_depth += (depth * 1000 - avg_depth) >> 8;
689 avg_mpcalls += (mpcalls * 1000 - avg_mpcalls) >> 8;
690 avg_lockcalls += (lockcalls * 1000 - avg_lockcalls) >> 8;
691 avg_gcalls += (gcalls * 1000 - avg_gcalls) >> 8;
692 cc->cc_next = NULL;
693 CC_UNLOCK(cc);
694 }
695
696 /*
697 * timeout --
698 * Execute a function after a specified length of time.
699 *
700 * untimeout --
701 * Cancel previous timeout function call.
702 *
703 * callout_handle_init --
704 * Initialize a handle so that using it with untimeout is benign.
705 *
706 * See AT&T BCI Driver Reference Manual for specification. This
707 * implementation differs from that one in that although an
708 * identification value is returned from timeout, the original
709 * arguments to timeout as well as the identifier are used to
710 * identify entries for untimeout.
711 */
712 struct callout_handle
713 timeout(ftn, arg, to_ticks)
714 timeout_t *ftn;
715 void *arg;
716 int to_ticks;
717 {
718 struct callout_cpu *cc;
719 struct callout *new;
720 struct callout_handle handle;
721
722 cc = CC_CPU(timeout_cpu);
723 CC_LOCK(cc);
724 /* Fill in the next free callout structure. */
725 new = SLIST_FIRST(&cc->cc_callfree);
726 if (new == NULL)
727 /* XXX Attempt to malloc first */
728 panic("timeout table full");
729 SLIST_REMOVE_HEAD(&cc->cc_callfree, c_links.sle);
730 callout_reset(new, to_ticks, ftn, arg);
731 handle.callout = new;
732 CC_UNLOCK(cc);
733
734 return (handle);
735 }
736
737 void
738 untimeout(ftn, arg, handle)
739 timeout_t *ftn;
740 void *arg;
741 struct callout_handle handle;
742 {
743 struct callout_cpu *cc;
744
745 /*
746 * Check for a handle that was initialized
747 * by callout_handle_init, but never used
748 * for a real timeout.
749 */
750 if (handle.callout == NULL)
751 return;
752
753 cc = callout_lock(handle.callout);
754 if (handle.callout->c_func == ftn && handle.callout->c_arg == arg)
755 callout_stop(handle.callout);
756 CC_UNLOCK(cc);
757 }
758
759 void
760 callout_handle_init(struct callout_handle *handle)
761 {
762 handle->callout = NULL;
763 }
764
765 /*
766 * New interface; clients allocate their own callout structures.
767 *
768 * callout_reset() - establish or change a timeout
769 * callout_stop() - disestablish a timeout
770 * callout_init() - initialize a callout structure so that it can
771 * safely be passed to callout_reset() and callout_stop()
772 *
773 * <sys/callout.h> defines three convenience macros:
774 *
775 * callout_active() - returns truth if callout has not been stopped,
776 * drained, or deactivated since the last time the callout was
777 * reset.
778 * callout_pending() - returns truth if callout is still waiting for timeout
779 * callout_deactivate() - marks the callout as having been serviced
780 */
781 int
782 callout_reset_on(struct callout *c, int to_ticks, void (*ftn)(void *),
783 void *arg, int cpu)
784 {
785 struct callout_cpu *cc;
786 int cancelled = 0;
787
788 /*
789 * Don't allow migration of pre-allocated callouts lest they
790 * become unbalanced.
791 */
792 if (c->c_flags & CALLOUT_LOCAL_ALLOC)
793 cpu = c->c_cpu;
794 cc = callout_lock(c);
795 if (cc->cc_curr == c) {
796 /*
797 * We're being asked to reschedule a callout which is
798 * currently in progress. If there is a lock then we
799 * can cancel the callout if it has not really started.
800 */
801 if (c->c_lock != NULL && !cc->cc_cancel)
802 cancelled = cc->cc_cancel = 1;
803 if (cc->cc_waiting) {
804 /*
805 * Someone has called callout_drain to kill this
806 * callout. Don't reschedule.
807 */
808 CTR4(KTR_CALLOUT, "%s %p func %p arg %p",
809 cancelled ? "cancelled" : "failed to cancel",
810 c, c->c_func, c->c_arg);
811 CC_UNLOCK(cc);
812 return (cancelled);
813 }
814 }
815 if (c->c_flags & CALLOUT_PENDING) {
816 if (cc->cc_next == c) {
817 cc->cc_next = TAILQ_NEXT(c, c_links.tqe);
818 }
819 TAILQ_REMOVE(&cc->cc_callwheel[c->c_time & callwheelmask], c,
820 c_links.tqe);
821
822 cancelled = 1;
823 c->c_flags &= ~(CALLOUT_ACTIVE | CALLOUT_PENDING);
824 }
825
826 #ifdef SMP
827 /*
828 * If the callout must migrate try to perform it immediately.
829 * If the callout is currently running, just defer the migration
830 * to a more appropriate moment.
831 */
832 if (c->c_cpu != cpu) {
833 if (cc->cc_curr == c) {
834 cc->cc_migration_cpu = cpu;
835 cc->cc_migration_ticks = to_ticks;
836 cc->cc_migration_func = ftn;
837 cc->cc_migration_arg = arg;
838 c->c_flags |= CALLOUT_DFRMIGRATION;
839 CTR5(KTR_CALLOUT,
840 "migration of %p func %p arg %p in %d to %u deferred",
841 c, c->c_func, c->c_arg, to_ticks, cpu);
842 CC_UNLOCK(cc);
843 return (cancelled);
844 }
845 cc = callout_cpu_switch(c, cc, cpu);
846 }
847 #endif
848
849 callout_cc_add(c, cc, to_ticks, ftn, arg, cpu);
850 CTR5(KTR_CALLOUT, "%sscheduled %p func %p arg %p in %d",
851 cancelled ? "re" : "", c, c->c_func, c->c_arg, to_ticks);
852 CC_UNLOCK(cc);
853
854 return (cancelled);
855 }
856
857 /*
858 * Common idioms that can be optimized in the future.
859 */
860 int
861 callout_schedule_on(struct callout *c, int to_ticks, int cpu)
862 {
863 return callout_reset_on(c, to_ticks, c->c_func, c->c_arg, cpu);
864 }
865
866 int
867 callout_schedule(struct callout *c, int to_ticks)
868 {
869 return callout_reset_on(c, to_ticks, c->c_func, c->c_arg, c->c_cpu);
870 }
871
872 int
873 _callout_stop_safe(c, safe)
874 struct callout *c;
875 int safe;
876 {
877 struct callout_cpu *cc, *old_cc;
878 struct lock_class *class;
879 int use_lock, sq_locked;
880
881 /*
882 * Some old subsystems don't hold Giant while running a callout_stop(),
883 * so just discard this check for the moment.
884 */
885 if (!safe && c->c_lock != NULL) {
886 if (c->c_lock == &Giant.lock_object)
887 use_lock = mtx_owned(&Giant);
888 else {
889 use_lock = 1;
890 class = LOCK_CLASS(c->c_lock);
891 class->lc_assert(c->c_lock, LA_XLOCKED);
892 }
893 } else
894 use_lock = 0;
895
896 sq_locked = 0;
897 old_cc = NULL;
898 again:
899 cc = callout_lock(c);
900
901 /*
902 * If the callout was migrating while the callout cpu lock was
903 * dropped, just drop the sleepqueue lock and check the states
904 * again.
905 */
906 if (sq_locked != 0 && cc != old_cc) {
907 #ifdef SMP
908 CC_UNLOCK(cc);
909 sleepq_release(&old_cc->cc_waiting);
910 sq_locked = 0;
911 old_cc = NULL;
912 goto again;
913 #else
914 panic("migration should not happen");
915 #endif
916 }
917
918 /*
919 * If the callout isn't pending, it's not on the queue, so
920 * don't attempt to remove it from the queue. We can try to
921 * stop it by other means however.
922 */
923 if (!(c->c_flags & CALLOUT_PENDING)) {
924 c->c_flags &= ~CALLOUT_ACTIVE;
925
926 /*
927 * If it wasn't on the queue and it isn't the current
928 * callout, then we can't stop it, so just bail.
929 */
930 if (cc->cc_curr != c) {
931 CTR3(KTR_CALLOUT, "failed to stop %p func %p arg %p",
932 c, c->c_func, c->c_arg);
933 CC_UNLOCK(cc);
934 if (sq_locked)
935 sleepq_release(&cc->cc_waiting);
936 return (0);
937 }
938
939 if (safe) {
940 /*
941 * The current callout is running (or just
942 * about to run) and blocking is allowed, so
943 * just wait for the current invocation to
944 * finish.
945 */
946 while (cc->cc_curr == c) {
947
948 /*
949 * Use direct calls to sleepqueue interface
950 * instead of cv/msleep in order to avoid
951 * a LOR between cc_lock and sleepqueue
952 * chain spinlocks. This piece of code
953 * emulates a msleep_spin() call actually.
954 *
955 * If we already have the sleepqueue chain
956 * locked, then we can safely block. If we
957 * don't already have it locked, however,
958 * we have to drop the cc_lock to lock
959 * it. This opens several races, so we
960 * restart at the beginning once we have
961 * both locks. If nothing has changed, then
962 * we will end up back here with sq_locked
963 * set.
964 */
965 if (!sq_locked) {
966 CC_UNLOCK(cc);
967 sleepq_lock(&cc->cc_waiting);
968 sq_locked = 1;
969 old_cc = cc;
970 goto again;
971 }
972
973 /*
974 * Migration could be cancelled here, but
975 * as long as it is still not sure when it
976 * will be packed up, just let softclock()
977 * take care of it.
978 */
979 cc->cc_waiting = 1;
980 DROP_GIANT();
981 CC_UNLOCK(cc);
982 sleepq_add(&cc->cc_waiting,
983 &cc->cc_lock.lock_object, "codrain",
984 SLEEPQ_SLEEP, 0);
985 sleepq_wait(&cc->cc_waiting, 0);
986 sq_locked = 0;
987 old_cc = NULL;
988
989 /* Reacquire locks previously released. */
990 PICKUP_GIANT();
991 CC_LOCK(cc);
992 }
993 } else if (use_lock && !cc->cc_cancel) {
994 /*
995 * The current callout is waiting for its
996 * lock which we hold. Cancel the callout
997 * and return. After our caller drops the
998 * lock, the callout will be skipped in
999 * softclock().
1000 */
1001 cc->cc_cancel = 1;
1002 CTR3(KTR_CALLOUT, "cancelled %p func %p arg %p",
1003 c, c->c_func, c->c_arg);
1004 KASSERT(!cc_cme_migrating(cc),
1005 ("callout wrongly scheduled for migration"));
1006 CC_UNLOCK(cc);
1007 KASSERT(!sq_locked, ("sleepqueue chain locked"));
1008 return (1);
1009 } else if ((c->c_flags & CALLOUT_DFRMIGRATION) != 0) {
1010 c->c_flags &= ~CALLOUT_DFRMIGRATION;
1011 CTR3(KTR_CALLOUT, "postponing stop %p func %p arg %p",
1012 c, c->c_func, c->c_arg);
1013 CC_UNLOCK(cc);
1014 return (1);
1015 }
1016 CTR3(KTR_CALLOUT, "failed to stop %p func %p arg %p",
1017 c, c->c_func, c->c_arg);
1018 CC_UNLOCK(cc);
1019 KASSERT(!sq_locked, ("sleepqueue chain still locked"));
1020 return (0);
1021 }
1022 if (sq_locked)
1023 sleepq_release(&cc->cc_waiting);
1024
1025 c->c_flags &= ~(CALLOUT_ACTIVE | CALLOUT_PENDING);
1026
1027 CTR3(KTR_CALLOUT, "cancelled %p func %p arg %p",
1028 c, c->c_func, c->c_arg);
1029 if (cc->cc_next == c)
1030 cc->cc_next = TAILQ_NEXT(c, c_links.tqe);
1031 TAILQ_REMOVE(&cc->cc_callwheel[c->c_time & callwheelmask], c,
1032 c_links.tqe);
1033 callout_cc_del(c, cc);
1034
1035 CC_UNLOCK(cc);
1036 return (1);
1037 }
1038
1039 void
1040 callout_init(c, mpsafe)
1041 struct callout *c;
1042 int mpsafe;
1043 {
1044 bzero(c, sizeof *c);
1045 if (mpsafe) {
1046 c->c_lock = NULL;
1047 c->c_flags = CALLOUT_RETURNUNLOCKED;
1048 } else {
1049 c->c_lock = &Giant.lock_object;
1050 c->c_flags = 0;
1051 }
1052 c->c_cpu = timeout_cpu;
1053 }
1054
1055 void
1056 _callout_init_lock(c, lock, flags)
1057 struct callout *c;
1058 struct lock_object *lock;
1059 int flags;
1060 {
1061 bzero(c, sizeof *c);
1062 c->c_lock = lock;
1063 KASSERT((flags & ~(CALLOUT_RETURNUNLOCKED | CALLOUT_SHAREDLOCK)) == 0,
1064 ("callout_init_lock: bad flags %d", flags));
1065 KASSERT(lock != NULL || (flags & CALLOUT_RETURNUNLOCKED) == 0,
1066 ("callout_init_lock: CALLOUT_RETURNUNLOCKED with no lock"));
1067 KASSERT(lock == NULL || !(LOCK_CLASS(lock)->lc_flags &
1068 (LC_SPINLOCK | LC_SLEEPABLE)), ("%s: invalid lock class",
1069 __func__));
1070 c->c_flags = flags & (CALLOUT_RETURNUNLOCKED | CALLOUT_SHAREDLOCK);
1071 c->c_cpu = timeout_cpu;
1072 }
1073
1074 #ifdef APM_FIXUP_CALLTODO
1075 /*
1076 * Adjust the kernel calltodo timeout list. This routine is used after
1077 * an APM resume to recalculate the calltodo timer list values with the
1078 * number of hz's we have been sleeping. The next hardclock() will detect
1079 * that there are fired timers and run softclock() to execute them.
1080 *
1081 * Please note, I have not done an exhaustive analysis of what code this
1082 * might break. I am motivated to have my select()'s and alarm()'s that
1083 * have expired during suspend firing upon resume so that the applications
1084 * which set the timer can do the maintanence the timer was for as close
1085 * as possible to the originally intended time. Testing this code for a
1086 * week showed that resuming from a suspend resulted in 22 to 25 timers
1087 * firing, which seemed independant on whether the suspend was 2 hours or
1088 * 2 days. Your milage may vary. - Ken Key <key@cs.utk.edu>
1089 */
1090 void
1091 adjust_timeout_calltodo(time_change)
1092 struct timeval *time_change;
1093 {
1094 register struct callout *p;
1095 unsigned long delta_ticks;
1096
1097 /*
1098 * How many ticks were we asleep?
1099 * (stolen from tvtohz()).
1100 */
1101
1102 /* Don't do anything */
1103 if (time_change->tv_sec < 0)
1104 return;
1105 else if (time_change->tv_sec <= LONG_MAX / 1000000)
1106 delta_ticks = (time_change->tv_sec * 1000000 +
1107 time_change->tv_usec + (tick - 1)) / tick + 1;
1108 else if (time_change->tv_sec <= LONG_MAX / hz)
1109 delta_ticks = time_change->tv_sec * hz +
1110 (time_change->tv_usec + (tick - 1)) / tick + 1;
1111 else
1112 delta_ticks = LONG_MAX;
1113
1114 if (delta_ticks > INT_MAX)
1115 delta_ticks = INT_MAX;
1116
1117 /*
1118 * Now rip through the timer calltodo list looking for timers
1119 * to expire.
1120 */
1121
1122 /* don't collide with softclock() */
1123 CC_LOCK(cc);
1124 for (p = calltodo.c_next; p != NULL; p = p->c_next) {
1125 p->c_time -= delta_ticks;
1126
1127 /* Break if the timer had more time on it than delta_ticks */
1128 if (p->c_time > 0)
1129 break;
1130
1131 /* take back the ticks the timer didn't use (p->c_time <= 0) */
1132 delta_ticks = -p->c_time;
1133 }
1134 CC_UNLOCK(cc);
1135
1136 return;
1137 }
1138 #endif /* APM_FIXUP_CALLTODO */
Cache object: 95c13515083671113fb76782b7f99af4
|