xref: /dragonfly/sys/kern/kern_timeout.c (revision 2038fb68)
1 /*
2  * Copyright (c) 2004 The DragonFly Project.  All rights reserved.
3  *
4  * This code is derived from software contributed to The DragonFly Project
5  * by Matthew Dillon <dillon@backplane.com>
6  *
7  * Redistribution and use in source and binary forms, with or without
8  * modification, are permitted provided that the following conditions
9  * are met:
10  *
11  * 1. Redistributions of source code must retain the above copyright
12  *    notice, this list of conditions and the following disclaimer.
13  * 2. Redistributions in binary form must reproduce the above copyright
14  *    notice, this list of conditions and the following disclaimer in
15  *    the documentation and/or other materials provided with the
16  *    distribution.
17  * 3. Neither the name of The DragonFly Project nor the names of its
18  *    contributors may be used to endorse or promote products derived
19  *    from this software without specific, prior written permission.
20  *
21  * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
22  * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
23  * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS
24  * FOR A PARTICULAR PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL THE
25  * COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
26  * INCIDENTAL, SPECIAL, EXEMPLARY OR CONSEQUENTIAL DAMAGES (INCLUDING,
27  * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
28  * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED
29  * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
30  * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT
31  * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
32  * SUCH DAMAGE.
33  */
34 /*
35  * Copyright (c) 1982, 1986, 1991, 1993
36  *	The Regents of the University of California.  All rights reserved.
37  * (c) UNIX System Laboratories, Inc.
38  * All or some portions of this file are derived from material licensed
39  * to the University of California by American Telephone and Telegraph
40  * Co. or Unix System Laboratories, Inc. and are reproduced herein with
41  * the permission of UNIX System Laboratories, Inc.
42  *
43  * Redistribution and use in source and binary forms, with or without
44  * modification, are permitted provided that the following conditions
45  * are met:
46  * 1. Redistributions of source code must retain the above copyright
47  *    notice, this list of conditions and the following disclaimer.
48  * 2. Redistributions in binary form must reproduce the above copyright
49  *    notice, this list of conditions and the following disclaimer in the
50  *    documentation and/or other materials provided with the distribution.
51  * 3. All advertising materials mentioning features or use of this software
52  *    must display the following acknowledgement:
53  *	This product includes software developed by the University of
54  *	California, Berkeley and its contributors.
55  * 4. Neither the name of the University nor the names of its contributors
56  *    may be used to endorse or promote products derived from this software
57  *    without specific prior written permission.
58  *
59  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
60  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
61  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
62  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
63  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
64  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
65  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
66  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
67  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
68  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
69  * SUCH DAMAGE.
70  *
71  *	From: @(#)kern_clock.c	8.5 (Berkeley) 1/21/94
72  * $FreeBSD: src/sys/kern/kern_timeout.c,v 1.59.2.1 2001/11/13 18:24:52 archie Exp $
73  * $DragonFly: src/sys/kern/kern_timeout.c,v 1.27 2007/11/14 18:27:52 swildner Exp $
74  */
75 /*
76  * DRAGONFLY BGL STATUS
77  *
78  *	All the API functions should be MP safe.
79  *
80  *	The callback functions will be flagged as being MP safe if the
81  *	timeout structure is initialized with callout_init_mp() instead of
82  *	callout_init().
83  *
84  *	The helper threads cannot be made preempt-capable until after we
85  *	clean up all the uses of splsoftclock() and related interlocks (which
86  *	require the related functions to be MP safe as well).
87  */
88 /*
89  * The callout mechanism is based on the work of Adam M. Costello and
90  * George Varghese, published in a technical report entitled "Redesigning
91  * the BSD Callout and Timer Facilities" and modified slightly for inclusion
92  * in FreeBSD by Justin T. Gibbs.  The original work on the data structures
93  * used in this implementation was published by G. Varghese and T. Lauck in
94  * the paper "Hashed and Hierarchical Timing Wheels: Data Structures for
95  * the Efficient Implementation of a Timer Facility" in the Proceedings of
96  * the 11th ACM Annual Symposium on Operating Systems Principles,
97  * Austin, Texas Nov 1987.
98  *
99  * The per-cpu augmentation was done by Matthew Dillon.
100  */
101 
102 #include <sys/param.h>
103 #include <sys/systm.h>
104 #include <sys/callout.h>
105 #include <sys/kernel.h>
106 #include <sys/interrupt.h>
107 #include <sys/thread.h>
108 #include <sys/thread2.h>
109 
110 #ifndef MAX_SOFTCLOCK_STEPS
111 #define MAX_SOFTCLOCK_STEPS 100 /* Maximum allowed value of steps. */
112 #endif
113 
114 
115 struct softclock_pcpu {
116 	struct callout_tailq *callwheel;
117 	struct callout * volatile next;
118 	int softticks;		/* softticks index */
119 	int curticks;		/* per-cpu ticks counter */
120 	int isrunning;
121 	struct thread thread;
122 
123 };
124 
125 typedef struct softclock_pcpu *softclock_pcpu_t;
126 
127 /*
128  * TODO:
129  *	allocate more timeout table slots when table overflows.
130  */
131 static MALLOC_DEFINE(M_CALLOUT, "callout", "callout structures");
132 static int callwheelsize;
133 static int callwheelbits;
134 static int callwheelmask;
135 static struct softclock_pcpu softclock_pcpu_ary[MAXCPU];
136 
137 static void softclock_handler(void *arg);
138 
139 static void
140 swi_softclock_setup(void *arg)
141 {
142 	int cpu;
143 	int i;
144 
145 	/*
146 	 * Figure out how large a callwheel we need.  It must be a power of 2.
147 	 */
148 	callwheelsize = 1;
149 	callwheelbits = 0;
150 	while (callwheelsize < ncallout) {
151 		callwheelsize <<= 1;
152 		++callwheelbits;
153 	}
154 	callwheelmask = callwheelsize - 1;
155 
156 	/*
157 	 * Initialize per-cpu data structures.
158 	 */
159 	for (cpu = 0; cpu < ncpus; ++cpu) {
160 		softclock_pcpu_t sc;
161 
162 		sc = &softclock_pcpu_ary[cpu];
163 
164 		sc->callwheel = kmalloc(sizeof(*sc->callwheel) * callwheelsize,
165 					M_CALLOUT, M_WAITOK|M_ZERO);
166 		for (i = 0; i < callwheelsize; ++i)
167 			TAILQ_INIT(&sc->callwheel[i]);
168 
169 		/*
170 		 * Create a preemption-capable thread for each cpu to handle
171 		 * softclock timeouts on that cpu.  The preemption can only
172 		 * be blocked by a critical section.  The thread can itself
173 		 * be preempted by normal interrupts.
174 		 */
175 		lwkt_create(softclock_handler, sc, NULL,
176 			    &sc->thread, TDF_STOPREQ|TDF_INTTHREAD, cpu,
177 			    "softclock %d", cpu);
178 #if 0
179 		/*
180 		 * Do not make the thread preemptable until we clean up all
181 		 * the splsoftclock() calls in the system.  Since the threads
182 		 * are no longer operated as a software interrupt, the
183 		 * splsoftclock() calls will not have any effect on them.
184 		 */
185 		sc->thread.td_preemptable = lwkt_preempt;
186 #endif
187 	}
188 }
189 
190 /*
191  * Must occur after ncpus has been initialized.
192  */
193 SYSINIT(softclock_setup, SI_BOOT2_SOFTCLOCK, SI_ORDER_SECOND,
194 	swi_softclock_setup, NULL);
195 
196 /*
197  * This routine is called from the hardclock() (basically a FASTint/IPI) on
198  * each cpu in the system.  sc->curticks is this cpu's notion of the timebase.
199  * It IS NOT NECESSARILY SYNCHRONIZED WITH 'ticks'!  sc->softticks is where
200  * the callwheel is currently indexed.
201  *
202  * WARNING!  The MP lock is not necessarily held on call, nor can it be
203  * safely obtained.
204  *
205  * sc->softticks is adjusted by either this routine or our helper thread
206  * depending on whether the helper thread is running or not.
207  */
208 void
209 hardclock_softtick(globaldata_t gd)
210 {
211 	softclock_pcpu_t sc;
212 
213 	sc = &softclock_pcpu_ary[gd->gd_cpuid];
214 	++sc->curticks;
215 	if (sc->isrunning)
216 		return;
217 	if (sc->softticks == sc->curticks) {
218 		/*
219 		 * in sync, only wakeup the thread if there is something to
220 		 * do.
221 		 */
222 		if (TAILQ_FIRST(&sc->callwheel[sc->softticks & callwheelmask]))
223 		{
224 			sc->isrunning = 1;
225 			lwkt_schedule(&sc->thread);
226 		} else {
227 			++sc->softticks;
228 		}
229 	} else {
230 		/*
231 		 * out of sync, wakeup the thread unconditionally so it can
232 		 * catch up.
233 		 */
234 		sc->isrunning = 1;
235 		lwkt_schedule(&sc->thread);
236 	}
237 }
238 
239 /*
240  * This procedure is the main loop of our per-cpu helper thread.  The
241  * sc->isrunning flag prevents us from racing hardclock_softtick() and
242  * a critical section is sufficient to interlock sc->curticks and protect
243  * us from remote IPI's / list removal.
244  *
245  * The thread starts with the MP lock held and not in a critical section.
246  * The loop itself is MP safe while individual callbacks may or may not
247  * be, so we obtain or release the MP lock as appropriate.
248  */
249 static void
250 softclock_handler(void *arg)
251 {
252 	softclock_pcpu_t sc;
253 	struct callout *c;
254 	struct callout_tailq *bucket;
255 	void (*c_func)(void *);
256 	void *c_arg;
257 #ifdef SMP
258 	int mpsafe = 0;
259 #endif
260 
261 	lwkt_setpri_self(TDPRI_SOFT_NORM);
262 
263 	sc = arg;
264 	crit_enter();
265 loop:
266 	while (sc->softticks != (int)(sc->curticks + 1)) {
267 		bucket = &sc->callwheel[sc->softticks & callwheelmask];
268 
269 		for (c = TAILQ_FIRST(bucket); c; c = sc->next) {
270 			if (c->c_time != sc->softticks) {
271 				sc->next = TAILQ_NEXT(c, c_links.tqe);
272 				continue;
273 			}
274 #ifdef SMP
275 			if (c->c_flags & CALLOUT_MPSAFE) {
276 				if (mpsafe == 0) {
277 					mpsafe = 1;
278 					rel_mplock();
279 				}
280 			} else {
281 				/*
282 				 * The request might be removed while we
283 				 * are waiting to get the MP lock.  If it
284 				 * was removed sc->next will point to the
285 				 * next valid request or NULL, loop up.
286 				 */
287 				if (mpsafe) {
288 					mpsafe = 0;
289 					sc->next = c;
290 					get_mplock();
291 					if (c != sc->next)
292 						continue;
293 				}
294 			}
295 #endif
296 			sc->next = TAILQ_NEXT(c, c_links.tqe);
297 			TAILQ_REMOVE(bucket, c, c_links.tqe);
298 
299 			c_func = c->c_func;
300 			c_arg = c->c_arg;
301 			c->c_func = NULL;
302 			KKASSERT(c->c_flags & CALLOUT_DID_INIT);
303 			c->c_flags &= ~CALLOUT_PENDING;
304 			crit_exit();
305 			c_func(c_arg);
306 			crit_enter();
307 			/* NOTE: list may have changed */
308 		}
309 		++sc->softticks;
310 	}
311 	sc->isrunning = 0;
312 	lwkt_deschedule_self(&sc->thread);	/* == curthread */
313 	lwkt_switch();
314 	goto loop;
315 	/* NOT REACHED */
316 }
317 
318 /*
319  * New interface; clients allocate their own callout structures.
320  *
321  * callout_reset() - establish or change a timeout
322  * callout_stop() - disestablish a timeout
323  * callout_init() - initialize a callout structure so that it can
324  *			safely be passed to callout_reset() and callout_stop()
325  * callout_init_mp() - same but any installed functions must be MP safe.
326  *
327  * <sys/callout.h> defines three convenience macros:
328  *
329  * callout_active() - returns truth if callout has not been serviced
330  * callout_pending() - returns truth if callout is still waiting for timeout
331  * callout_deactivate() - marks the callout as having been serviced
332  */
333 
334 /*
335  * Start or restart a timeout.  Install the callout structure in the
336  * callwheel.  Callers may legally pass any value, even if 0 or negative,
337  * but since the sc->curticks index may have already been processed a
338  * minimum timeout of 1 tick will be enforced.
339  *
340  * The callout is installed on and will be processed on the current cpu's
341  * callout wheel.
342  *
343  * WARNING! This function may be called from any cpu but the caller must
344  * serialize callout_stop() and callout_reset() calls on the passed
345  * structure regardless of cpu.
346  */
347 void
348 callout_reset(struct callout *c, int to_ticks, void (*ftn)(void *),
349 		void *arg)
350 {
351 	softclock_pcpu_t sc;
352 	globaldata_t gd;
353 
354 #ifdef INVARIANTS
355         if ((c->c_flags & CALLOUT_DID_INIT) == 0) {
356 		callout_init(c);
357 		kprintf(
358 		    "callout_reset(%p) from %p: callout was not initialized\n",
359 		    c, ((int **)&c)[-1]);
360 		print_backtrace();
361 	}
362 #endif
363 	gd = mycpu;
364 	sc = &softclock_pcpu_ary[gd->gd_cpuid];
365 	crit_enter_gd(gd);
366 
367 	if (c->c_flags & CALLOUT_PENDING)
368 		callout_stop(c);
369 
370 	if (to_ticks <= 0)
371 		to_ticks = 1;
372 
373 	c->c_arg = arg;
374 	c->c_flags |= (CALLOUT_ACTIVE | CALLOUT_PENDING);
375 	c->c_func = ftn;
376 	c->c_time = sc->curticks + to_ticks;
377 #ifdef SMP
378 	c->c_gd = gd;
379 #endif
380 
381 	TAILQ_INSERT_TAIL(&sc->callwheel[c->c_time & callwheelmask],
382 			  c, c_links.tqe);
383 	crit_exit_gd(gd);
384 }
385 
386 /*
387  * Stop a running timer.  WARNING!  If called on a cpu other then the one
388  * the callout was started on this function will liveloop on its IPI to
389  * the target cpu to process the request.  It is possible for the callout
390  * to execute in that case.
391  *
392  * WARNING! This function may be called from any cpu but the caller must
393  * serialize callout_stop() and callout_reset() calls on the passed
394  * structure regardless of cpu.
395  *
396  * WARNING! This routine may be called from an IPI
397  */
398 int
399 callout_stop(struct callout *c)
400 {
401 	globaldata_t gd = mycpu;
402 #ifdef SMP
403 	globaldata_t tgd;
404 #endif
405 	softclock_pcpu_t sc;
406 
407 #ifdef INVARIANTS
408         if ((c->c_flags & CALLOUT_DID_INIT) == 0) {
409 		callout_init(c);
410 		kprintf(
411 		    "callout_stop(%p) from %p: callout was not initialized\n",
412 		    c, ((int **)&c)[-1]);
413 		print_backtrace();
414 	}
415 #endif
416 	crit_enter_gd(gd);
417 
418 	/*
419 	 * Don't attempt to delete a callout that's not on the queue.  The
420 	 * callout may not have a cpu assigned to it.  Callers do not have
421 	 * to be on the issuing cpu but must still serialize access to the
422 	 * callout structure.
423 	 *
424 	 * We are not cpu-localized here and cannot safely modify the
425 	 * flags field in the callout structure.  Note that most of the
426 	 * time CALLOUT_ACTIVE will be 0 if CALLOUT_PENDING is also 0.
427 	 *
428 	 * If we race another cpu's dispatch of this callout it is possible
429 	 * for CALLOUT_ACTIVE to be set with CALLOUT_PENDING unset.  This
430 	 * will cause us to fall through and synchronize with the other
431 	 * cpu.
432 	 */
433 	if ((c->c_flags & CALLOUT_PENDING) == 0) {
434 #ifdef SMP
435 		if ((c->c_flags & CALLOUT_ACTIVE) == 0) {
436 			crit_exit_gd(gd);
437 			return (0);
438 		}
439 		if (c->c_gd == NULL || c->c_gd == gd) {
440 			c->c_flags &= ~CALLOUT_ACTIVE;
441 			crit_exit_gd(gd);
442 			return (0);
443 		}
444 		/* fall-through to the cpu-localization code. */
445 #else
446 		c->c_flags &= ~CALLOUT_ACTIVE;
447 		crit_exit_gd(gd);
448 		return (0);
449 #endif
450 	}
451 #ifdef SMP
452 	if ((tgd = c->c_gd) != gd) {
453 		/*
454 		 * If the callout is owned by a different CPU we have to
455 		 * execute the function synchronously on the target cpu.
456 		 */
457 		int seq;
458 
459 		cpu_ccfence();	/* don't let tgd alias c_gd */
460 		seq = lwkt_send_ipiq(tgd, (void *)callout_stop, c);
461 		lwkt_wait_ipiq(tgd, seq);
462 	} else
463 #endif
464 	{
465 		/*
466 		 * If the callout is owned by the same CPU we can
467 		 * process it directly, but if we are racing our helper
468 		 * thread (sc->next), we have to adjust sc->next.  The
469 		 * race is interlocked by a critical section.
470 		 */
471 		sc = &softclock_pcpu_ary[gd->gd_cpuid];
472 
473 		c->c_flags &= ~(CALLOUT_ACTIVE | CALLOUT_PENDING);
474 		if (sc->next == c)
475 			sc->next = TAILQ_NEXT(c, c_links.tqe);
476 
477 		TAILQ_REMOVE(&sc->callwheel[c->c_time & callwheelmask],
478 				c, c_links.tqe);
479 		c->c_func = NULL;
480 	}
481 	crit_exit_gd(gd);
482 	return (1);
483 }
484 
485 /*
486  * Prepare a callout structure for use by callout_reset() and/or
487  * callout_stop().  The MP version of this routine requires that the callback
488  * function installed by callout_reset() be MP safe.
489  */
490 void
491 callout_init(struct callout *c)
492 {
493 	bzero(c, sizeof *c);
494 	c->c_flags = CALLOUT_DID_INIT;
495 }
496 
497 void
498 callout_init_mp(struct callout *c)
499 {
500 	callout_init(c);
501 	c->c_flags |= CALLOUT_MPSAFE;
502 }
503 
504 /* What, are you joking?  This is nuts! -Matt */
505 #if 0
506 #ifdef APM_FIXUP_CALLTODO
507 /*
508  * Adjust the kernel calltodo timeout list.  This routine is used after
509  * an APM resume to recalculate the calltodo timer list values with the
510  * number of hz's we have been sleeping.  The next hardclock() will detect
511  * that there are fired timers and run softclock() to execute them.
512  *
513  * Please note, I have not done an exhaustive analysis of what code this
514  * might break.  I am motivated to have my select()'s and alarm()'s that
515  * have expired during suspend firing upon resume so that the applications
516  * which set the timer can do the maintanence the timer was for as close
517  * as possible to the originally intended time.  Testing this code for a
518  * week showed that resuming from a suspend resulted in 22 to 25 timers
519  * firing, which seemed independant on whether the suspend was 2 hours or
520  * 2 days.  Your milage may vary.   - Ken Key <key@cs.utk.edu>
521  */
522 void
523 adjust_timeout_calltodo(struct timeval *time_change)
524 {
525 	struct callout *p;
526 	unsigned long delta_ticks;
527 
528 	/*
529 	 * How many ticks were we asleep?
530 	 * (stolen from tvtohz()).
531 	 */
532 
533 	/* Don't do anything */
534 	if (time_change->tv_sec < 0)
535 		return;
536 	else if (time_change->tv_sec <= LONG_MAX / 1000000)
537 		delta_ticks = (time_change->tv_sec * 1000000 +
538 			       time_change->tv_usec + (tick - 1)) / tick + 1;
539 	else if (time_change->tv_sec <= LONG_MAX / hz)
540 		delta_ticks = time_change->tv_sec * hz +
541 			      (time_change->tv_usec + (tick - 1)) / tick + 1;
542 	else
543 		delta_ticks = LONG_MAX;
544 
545 	if (delta_ticks > INT_MAX)
546 		delta_ticks = INT_MAX;
547 
548 	/*
549 	 * Now rip through the timer calltodo list looking for timers
550 	 * to expire.
551 	 */
552 
553 	/* don't collide with softclock() */
554 	crit_enter();
555 	for (p = calltodo.c_next; p != NULL; p = p->c_next) {
556 		p->c_time -= delta_ticks;
557 
558 		/* Break if the timer had more time on it than delta_ticks */
559 		if (p->c_time > 0)
560 			break;
561 
562 		/* take back the ticks the timer didn't use (p->c_time <= 0) */
563 		delta_ticks = -p->c_time;
564 	}
565 	crit_exit();
566 
567 	return;
568 }
569 #endif /* APM_FIXUP_CALLTODO */
570 #endif
571 
572