FreeBSD/Linux Kernel Cross Reference
sys/kern/kern_clock.c
1 /*-
2 * Copyright (c) 1997, 1998 Poul-Henning Kamp <phk@FreeBSD.org>
3 * Copyright (c) 1982, 1986, 1991, 1993
4 * The Regents of the University of California. All rights reserved.
5 * (c) UNIX System Laboratories, Inc.
6 * All or some portions of this file are derived from material licensed
7 * to the University of California by American Telephone and Telegraph
8 * Co. or Unix System Laboratories, Inc. and are reproduced herein with
9 * the permission of UNIX System Laboratories, Inc.
10 *
11 * Redistribution and use in source and binary forms, with or without
12 * modification, are permitted provided that the following conditions
13 * are met:
14 * 1. Redistributions of source code must retain the above copyright
15 * notice, this list of conditions and the following disclaimer.
16 * 2. Redistributions in binary form must reproduce the above copyright
17 * notice, this list of conditions and the following disclaimer in the
18 * documentation and/or other materials provided with the distribution.
19 * 3. All advertising materials mentioning features or use of this software
20 * must display the following acknowledgement:
21 * This product includes software developed by the University of
22 * California, Berkeley and its contributors.
23 * 4. Neither the name of the University nor the names of its contributors
24 * may be used to endorse or promote products derived from this software
25 * without specific prior written permission.
26 *
27 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
28 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
29 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
30 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
31 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
32 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
33 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
34 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
35 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
36 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
37 * SUCH DAMAGE.
38 *
39 * @(#)kern_clock.c 8.5 (Berkeley) 1/21/94
40 * $FreeBSD$
41 */
42
43 #include "opt_ntp.h"
44
45 #include <sys/param.h>
46 #include <sys/systm.h>
47 #include <sys/dkstat.h>
48 #include <sys/callout.h>
49 #include <sys/kernel.h>
50 #include <sys/proc.h>
51 #include <sys/malloc.h>
52 #include <sys/resourcevar.h>
53 #include <sys/signalvar.h>
54 #include <sys/timex.h>
55 #include <sys/timepps.h>
56 #include <vm/vm.h>
57 #include <sys/lock.h>
58 #include <vm/pmap.h>
59 #include <vm/vm_map.h>
60 #include <sys/sysctl.h>
61
62 #include <machine/cpu.h>
63 #include <machine/limits.h>
64
65 #ifdef GPROF
66 #include <sys/gmon.h>
67 #endif
68
69 #if defined(SMP) && defined(BETTER_CLOCK)
70 #include <machine/smp.h>
71 #endif
72
73 /*
74 * Number of timecounters used to implement stable storage
75 */
76 #ifndef NTIMECOUNTER
77 #define NTIMECOUNTER 5
78 #endif
79
80 static MALLOC_DEFINE(M_TIMECOUNTER, "timecounter",
81 "Timecounter stable storage");
82
83 static void initclocks __P((void *dummy));
84 SYSINIT(clocks, SI_SUB_CLOCKS, SI_ORDER_FIRST, initclocks, NULL)
85
86 static void tco_forward __P((int force));
87 static void tco_setscales __P((struct timecounter *tc));
88 static __inline unsigned tco_delta __P((struct timecounter *tc));
89
90 /* Some of these don't belong here, but it's easiest to concentrate them. */
91 #if defined(SMP) && defined(BETTER_CLOCK)
92 long cp_time[CPUSTATES];
93 #else
94 static long cp_time[CPUSTATES];
95 #endif
96
97 long tk_cancc;
98 long tk_nin;
99 long tk_nout;
100 long tk_rawcc;
101
102 time_t time_second;
103
104 /*
105 * Which update policy to use.
106 * 0 - every tick, bad hardware may fail with "calcru negative..."
107 * 1 - more resistent to the above hardware, but less efficient.
108 */
109 static int tco_method;
110
111 /*
112 * Implement a dummy timecounter which we can use until we get a real one
113 * in the air. This allows the console and other early stuff to use
114 * timeservices.
115 */
116
117 static unsigned
118 dummy_get_timecount(struct timecounter *tc)
119 {
120 static unsigned now;
121 return (++now);
122 }
123
124 static struct timecounter dummy_timecounter = {
125 dummy_get_timecount,
126 0,
127 ~0u,
128 1000000,
129 "dummy"
130 };
131
132 struct timecounter *timecounter = &dummy_timecounter;
133
134 /*
135 * Clock handling routines.
136 *
137 * This code is written to operate with two timers that run independently of
138 * each other.
139 *
140 * The main timer, running hz times per second, is used to trigger interval
141 * timers, timeouts and rescheduling as needed.
142 *
143 * The second timer handles kernel and user profiling,
144 * and does resource use estimation. If the second timer is programmable,
145 * it is randomized to avoid aliasing between the two clocks. For example,
146 * the randomization prevents an adversary from always giving up the cpu
147 * just before its quantum expires. Otherwise, it would never accumulate
148 * cpu ticks. The mean frequency of the second timer is stathz.
149 *
150 * If no second timer exists, stathz will be zero; in this case we drive
151 * profiling and statistics off the main clock. This WILL NOT be accurate;
152 * do not do it unless absolutely necessary.
153 *
154 * The statistics clock may (or may not) be run at a higher rate while
155 * profiling. This profile clock runs at profhz. We require that profhz
156 * be an integral multiple of stathz.
157 *
158 * If the statistics clock is running fast, it must be divided by the ratio
159 * profhz/stathz for statistics. (For profiling, every tick counts.)
160 *
161 * Time-of-day is maintained using a "timecounter", which may or may
162 * not be related to the hardware generating the above mentioned
163 * interrupts.
164 */
165
166 int stathz;
167 int profhz;
168 static int profprocs;
169 int ticks;
170 static int psdiv, pscnt; /* prof => stat divider */
171 int psratio; /* ratio: prof / stat */
172
173 /*
174 * Initialize clock frequencies and start both clocks running.
175 */
176 /* ARGSUSED*/
177 static void
178 initclocks(dummy)
179 void *dummy;
180 {
181 register int i;
182
183 /*
184 * Set divisors to 1 (normal case) and let the machine-specific
185 * code do its bit.
186 */
187 psdiv = pscnt = 1;
188 cpu_initclocks();
189
190 /*
191 * Compute profhz/stathz, and fix profhz if needed.
192 */
193 i = stathz ? stathz : hz;
194 if (profhz == 0)
195 profhz = i;
196 psratio = profhz / i;
197 }
198
199 /*
200 * The real-time timer, interrupting hz times per second.
201 */
202 void
203 hardclock(frame)
204 register struct clockframe *frame;
205 {
206 register struct proc *p;
207
208 p = curproc;
209 if (p) {
210 register struct pstats *pstats;
211
212 /*
213 * Run current process's virtual and profile time, as needed.
214 */
215 pstats = p->p_stats;
216 if (CLKF_USERMODE(frame) &&
217 timevalisset(&pstats->p_timer[ITIMER_VIRTUAL].it_value) &&
218 itimerdecr(&pstats->p_timer[ITIMER_VIRTUAL], tick) == 0)
219 psignal(p, SIGVTALRM);
220 if (timevalisset(&pstats->p_timer[ITIMER_PROF].it_value) &&
221 itimerdecr(&pstats->p_timer[ITIMER_PROF], tick) == 0)
222 psignal(p, SIGPROF);
223 }
224
225 #if defined(SMP) && defined(BETTER_CLOCK)
226 forward_hardclock(pscnt);
227 #endif
228
229 /*
230 * If no separate statistics clock is available, run it from here.
231 */
232 if (stathz == 0)
233 statclock(frame);
234
235 tco_forward(0);
236 ticks++;
237
238 /*
239 * Process callouts at a very low cpu priority, so we don't keep the
240 * relatively high clock interrupt priority any longer than necessary.
241 */
242 if (TAILQ_FIRST(&callwheel[ticks & callwheelmask]) != NULL) {
243 if (CLKF_BASEPRI(frame)) {
244 /*
245 * Save the overhead of a software interrupt;
246 * it will happen as soon as we return, so do it now.
247 */
248 (void)splsoftclock();
249 softclock();
250 } else
251 setsoftclock();
252 } else if (softticks + 1 == ticks)
253 ++softticks;
254 }
255
256 /*
257 * Compute number of ticks in the specified amount of time.
258 */
259 int
260 tvtohz(tv)
261 struct timeval *tv;
262 {
263 register unsigned long ticks;
264 register long sec, usec;
265
266 /*
267 * If the number of usecs in the whole seconds part of the time
268 * difference fits in a long, then the total number of usecs will
269 * fit in an unsigned long. Compute the total and convert it to
270 * ticks, rounding up and adding 1 to allow for the current tick
271 * to expire. Rounding also depends on unsigned long arithmetic
272 * to avoid overflow.
273 *
274 * Otherwise, if the number of ticks in the whole seconds part of
275 * the time difference fits in a long, then convert the parts to
276 * ticks separately and add, using similar rounding methods and
277 * overflow avoidance. This method would work in the previous
278 * case but it is slightly slower and assumes that hz is integral.
279 *
280 * Otherwise, round the time difference down to the maximum
281 * representable value.
282 *
283 * If ints have 32 bits, then the maximum value for any timeout in
284 * 10ms ticks is 248 days.
285 */
286 sec = tv->tv_sec;
287 usec = tv->tv_usec;
288 if (usec < 0) {
289 sec--;
290 usec += 1000000;
291 }
292 if (sec < 0) {
293 #ifdef DIAGNOSTIC
294 if (usec > 0) {
295 sec++;
296 usec -= 1000000;
297 }
298 printf("tvotohz: negative time difference %ld sec %ld usec\n",
299 sec, usec);
300 #endif
301 ticks = 1;
302 } else if (sec <= LONG_MAX / 1000000)
303 ticks = (sec * 1000000 + (unsigned long)usec + (tick - 1))
304 / tick + 1;
305 else if (sec <= LONG_MAX / hz)
306 ticks = sec * hz
307 + ((unsigned long)usec + (tick - 1)) / tick + 1;
308 else
309 ticks = LONG_MAX;
310 if (ticks > INT_MAX)
311 ticks = INT_MAX;
312 return ((int)ticks);
313 }
314
315 /*
316 * Start profiling on a process.
317 *
318 * Kernel profiling passes proc0 which never exits and hence
319 * keeps the profile clock running constantly.
320 */
321 void
322 startprofclock(p)
323 register struct proc *p;
324 {
325 int s;
326
327 if ((p->p_flag & P_PROFIL) == 0) {
328 p->p_flag |= P_PROFIL;
329 if (++profprocs == 1 && stathz != 0) {
330 s = splstatclock();
331 psdiv = pscnt = psratio;
332 setstatclockrate(profhz);
333 splx(s);
334 }
335 }
336 }
337
338 /*
339 * Stop profiling on a process.
340 */
341 void
342 stopprofclock(p)
343 register struct proc *p;
344 {
345 int s;
346
347 if (p->p_flag & P_PROFIL) {
348 p->p_flag &= ~P_PROFIL;
349 if (--profprocs == 0 && stathz != 0) {
350 s = splstatclock();
351 psdiv = pscnt = 1;
352 setstatclockrate(stathz);
353 splx(s);
354 }
355 }
356 }
357
358 /*
359 * Statistics clock. Grab profile sample, and if divider reaches 0,
360 * do process and kernel statistics.
361 */
362 void
363 statclock(frame)
364 register struct clockframe *frame;
365 {
366 #ifdef GPROF
367 register struct gmonparam *g;
368 int i;
369 #endif
370 register struct proc *p;
371 struct pstats *pstats;
372 long rss;
373 struct rusage *ru;
374 struct vmspace *vm;
375
376 if (curproc != NULL && CLKF_USERMODE(frame)) {
377 p = curproc;
378 if (p->p_flag & P_PROFIL)
379 addupc_intr(p, CLKF_PC(frame), 1);
380 #if defined(SMP) && defined(BETTER_CLOCK)
381 if (stathz != 0)
382 forward_statclock(pscnt);
383 #endif
384 if (--pscnt > 0)
385 return;
386 /*
387 * Came from user mode; CPU was in user state.
388 * If this process is being profiled record the tick.
389 */
390 p->p_uticks++;
391 if (p->p_nice > NZERO)
392 cp_time[CP_NICE]++;
393 else
394 cp_time[CP_USER]++;
395 } else {
396 #ifdef GPROF
397 /*
398 * Kernel statistics are just like addupc_intr, only easier.
399 */
400 g = &_gmonparam;
401 if (g->state == GMON_PROF_ON) {
402 i = CLKF_PC(frame) - g->lowpc;
403 if (i < g->textsize) {
404 i /= HISTFRACTION * sizeof(*g->kcount);
405 g->kcount[i]++;
406 }
407 }
408 #endif
409 #if defined(SMP) && defined(BETTER_CLOCK)
410 if (stathz != 0)
411 forward_statclock(pscnt);
412 #endif
413 if (--pscnt > 0)
414 return;
415 /*
416 * Came from kernel mode, so we were:
417 * - handling an interrupt,
418 * - doing syscall or trap work on behalf of the current
419 * user process, or
420 * - spinning in the idle loop.
421 * Whichever it is, charge the time as appropriate.
422 * Note that we charge interrupts to the current process,
423 * regardless of whether they are ``for'' that process,
424 * so that we know how much of its real time was spent
425 * in ``non-process'' (i.e., interrupt) work.
426 */
427 p = curproc;
428 if (CLKF_INTR(frame)) {
429 if (p != NULL)
430 p->p_iticks++;
431 cp_time[CP_INTR]++;
432 } else if (p != NULL) {
433 p->p_sticks++;
434 cp_time[CP_SYS]++;
435 } else
436 cp_time[CP_IDLE]++;
437 }
438 pscnt = psdiv;
439
440 /*
441 * We maintain statistics shown by user-level statistics
442 * programs: the amount of time in each cpu state.
443 */
444
445 /*
446 * We adjust the priority of the current process. The priority of
447 * a process gets worse as it accumulates CPU time. The cpu usage
448 * estimator (p_estcpu) is increased here. The formula for computing
449 * priorities (in kern_synch.c) will compute a different value each
450 * time p_estcpu increases by 4. The cpu usage estimator ramps up
451 * quite quickly when the process is running (linearly), and decays
452 * away exponentially, at a rate which is proportionally slower when
453 * the system is busy. The basic principal is that the system will
454 * 90% forget that the process used a lot of CPU time in 5 * loadav
455 * seconds. This causes the system to favor processes which haven't
456 * run much recently, and to round-robin among other processes.
457 */
458 if (p != NULL) {
459 p->p_cpticks++;
460 if (++p->p_estcpu == 0)
461 p->p_estcpu--;
462 if ((p->p_estcpu & 3) == 0) {
463 resetpriority(p);
464 if (p->p_priority >= PUSER)
465 p->p_priority = p->p_usrpri;
466 }
467
468 /* Update resource usage integrals and maximums. */
469 if ((pstats = p->p_stats) != NULL &&
470 (ru = &pstats->p_ru) != NULL &&
471 (vm = p->p_vmspace) != NULL) {
472 ru->ru_ixrss += vm->vm_tsize * PAGE_SIZE / 1024;
473 ru->ru_idrss += vm->vm_dsize * PAGE_SIZE / 1024;
474 ru->ru_isrss += vm->vm_ssize * PAGE_SIZE / 1024;
475 rss = vm->vm_pmap.pm_stats.resident_count *
476 PAGE_SIZE / 1024;
477 if (ru->ru_maxrss < rss)
478 ru->ru_maxrss = rss;
479 }
480 }
481 }
482
483 /*
484 * Return information about system clocks.
485 */
486 static int
487 sysctl_kern_clockrate SYSCTL_HANDLER_ARGS
488 {
489 struct clockinfo clkinfo;
490 /*
491 * Construct clockinfo structure.
492 */
493 clkinfo.hz = hz;
494 clkinfo.tick = tick;
495 clkinfo.tickadj = tickadj;
496 clkinfo.profhz = profhz;
497 clkinfo.stathz = stathz ? stathz : hz;
498 return (sysctl_handle_opaque(oidp, &clkinfo, sizeof clkinfo, req));
499 }
500
501 SYSCTL_PROC(_kern, KERN_CLOCKRATE, clockrate, CTLTYPE_STRUCT|CTLFLAG_RD,
502 0, 0, sysctl_kern_clockrate, "S,clockinfo","");
503
504 static __inline unsigned
505 tco_delta(struct timecounter *tc)
506 {
507
508 return ((tc->tc_get_timecount(tc) - tc->tc_offset_count) &
509 tc->tc_counter_mask);
510 }
511
512 /*
513 * We have eight functions for looking at the clock, four for
514 * microseconds and four for nanoseconds. For each there is fast
515 * but less precise version "get{nano|micro}[up]time" which will
516 * return a time which is up to 1/HZ previous to the call, whereas
517 * the raw version "{nano|micro}[up]time" will return a timestamp
518 * which is as precise as possible. The "up" variants return the
519 * time relative to system boot, these are well suited for time
520 * interval measurements.
521 */
522
523 void
524 getmicrotime(struct timeval *tvp)
525 {
526 struct timecounter *tc;
527
528 if (!tco_method) {
529 tc = timecounter;
530 *tvp = tc->tc_microtime;
531 } else {
532 microtime(tvp);
533 }
534 }
535
536 void
537 getnanotime(struct timespec *tsp)
538 {
539 struct timecounter *tc;
540
541 if (!tco_method) {
542 tc = timecounter;
543 *tsp = tc->tc_nanotime;
544 } else {
545 nanotime(tsp);
546 }
547 }
548
549 void
550 microtime(struct timeval *tv)
551 {
552 struct timecounter *tc;
553
554 tc = (struct timecounter *)timecounter;
555 tv->tv_sec = tc->tc_offset_sec;
556 tv->tv_usec = tc->tc_offset_micro;
557 tv->tv_usec += ((u_int64_t)tco_delta(tc) * tc->tc_scale_micro) >> 32;
558 tv->tv_usec += boottime.tv_usec;
559 tv->tv_sec += boottime.tv_sec;
560 while (tv->tv_usec >= 1000000) {
561 tv->tv_usec -= 1000000;
562 tv->tv_sec++;
563 }
564 }
565
566 void
567 nanotime(struct timespec *ts)
568 {
569 unsigned count;
570 u_int64_t delta;
571 struct timecounter *tc;
572
573 tc = (struct timecounter *)timecounter;
574 ts->tv_sec = tc->tc_offset_sec;
575 count = tco_delta(tc);
576 delta = tc->tc_offset_nano;
577 delta += ((u_int64_t)count * tc->tc_scale_nano_f);
578 delta >>= 32;
579 delta += ((u_int64_t)count * tc->tc_scale_nano_i);
580 delta += boottime.tv_usec * 1000;
581 ts->tv_sec += boottime.tv_sec;
582 while (delta >= 1000000000) {
583 delta -= 1000000000;
584 ts->tv_sec++;
585 }
586 ts->tv_nsec = delta;
587 }
588
589 void
590 getmicrouptime(struct timeval *tvp)
591 {
592 struct timecounter *tc;
593
594 if (!tco_method) {
595 tc = timecounter;
596 tvp->tv_sec = tc->tc_offset_sec;
597 tvp->tv_usec = tc->tc_offset_micro;
598 } else {
599 microuptime(tvp);
600 }
601 }
602
603 void
604 getnanouptime(struct timespec *tsp)
605 {
606 struct timecounter *tc;
607
608 if (!tco_method) {
609 tc = timecounter;
610 tsp->tv_sec = tc->tc_offset_sec;
611 tsp->tv_nsec = tc->tc_offset_nano >> 32;
612 } else {
613 nanouptime(tsp);
614 }
615 }
616
617 void
618 microuptime(struct timeval *tv)
619 {
620 struct timecounter *tc;
621
622 tc = (struct timecounter *)timecounter;
623 tv->tv_sec = tc->tc_offset_sec;
624 tv->tv_usec = tc->tc_offset_micro;
625 tv->tv_usec += ((u_int64_t)tco_delta(tc) * tc->tc_scale_micro) >> 32;
626 if (tv->tv_usec >= 1000000) {
627 tv->tv_usec -= 1000000;
628 tv->tv_sec++;
629 }
630 }
631
632 void
633 nanouptime(struct timespec *ts)
634 {
635 unsigned count;
636 u_int64_t delta;
637 struct timecounter *tc;
638
639 tc = (struct timecounter *)timecounter;
640 ts->tv_sec = tc->tc_offset_sec;
641 count = tco_delta(tc);
642 delta = tc->tc_offset_nano;
643 delta += ((u_int64_t)count * tc->tc_scale_nano_f);
644 delta >>= 32;
645 delta += ((u_int64_t)count * tc->tc_scale_nano_i);
646 if (delta >= 1000000000) {
647 delta -= 1000000000;
648 ts->tv_sec++;
649 }
650 ts->tv_nsec = delta;
651 }
652
653 static void
654 tco_setscales(struct timecounter *tc)
655 {
656 u_int64_t scale;
657
658 scale = 1000000000LL << 32;
659 scale += tc->tc_adjustment;
660 scale /= tc->tc_tweak->tc_frequency;
661 tc->tc_scale_micro = scale / 1000;
662 tc->tc_scale_nano_f = scale & 0xffffffff;
663 tc->tc_scale_nano_i = scale >> 32;
664 }
665
666 void
667 update_timecounter(struct timecounter *tc)
668 {
669 tco_setscales(tc);
670 }
671
672 void
673 init_timecounter(struct timecounter *tc)
674 {
675 struct timespec ts1;
676 struct timecounter *t1, *t2, *t3;
677 int i;
678
679 tc->tc_adjustment = 0;
680 tc->tc_tweak = tc;
681 tco_setscales(tc);
682 tc->tc_offset_count = tc->tc_get_timecount(tc);
683 MALLOC(t1, struct timecounter *, sizeof *t1, M_TIMECOUNTER, M_WAITOK);
684 *t1 = *tc;
685 t2 = t1;
686 for (i = 1; i < NTIMECOUNTER; i++) {
687 MALLOC(t3, struct timecounter *, sizeof *t3,
688 M_TIMECOUNTER, M_WAITOK);
689 *t3 = *tc;
690 t3->tc_other = t2;
691 t2 = t3;
692 }
693 t1->tc_other = t3;
694 tc = t1;
695
696 printf("Timecounter \"%s\" frequency %lu Hz\n",
697 tc->tc_name, (u_long)tc->tc_frequency);
698
699 /* XXX: For now always start using the counter. */
700 tc->tc_offset_count = tc->tc_get_timecount(tc);
701 nanouptime(&ts1);
702 tc->tc_offset_nano = (u_int64_t)ts1.tv_nsec << 32;
703 tc->tc_offset_micro = ts1.tv_nsec / 1000;
704 tc->tc_offset_sec = ts1.tv_sec;
705 timecounter = tc;
706 }
707
708 void
709 set_timecounter(struct timespec *ts)
710 {
711 struct timespec ts2;
712
713 nanouptime(&ts2);
714 boottime.tv_sec = ts->tv_sec - ts2.tv_sec;
715 boottime.tv_usec = (ts->tv_nsec - ts2.tv_nsec) / 1000;
716 if (boottime.tv_usec < 0) {
717 boottime.tv_usec += 1000000;
718 boottime.tv_sec--;
719 }
720 /* fiddle all the little crinkly bits around the fiords... */
721 tco_forward(1);
722 }
723
724
725 #if 0 /* Currently unused */
726 void
727 switch_timecounter(struct timecounter *newtc)
728 {
729 int s;
730 struct timecounter *tc;
731 struct timespec ts;
732
733 s = splclock();
734 tc = timecounter;
735 if (newtc == tc || newtc == tc->tc_other) {
736 splx(s);
737 return;
738 }
739 nanouptime(&ts);
740 newtc->tc_offset_sec = ts.tv_sec;
741 newtc->tc_offset_nano = (u_int64_t)ts.tv_nsec << 32;
742 newtc->tc_offset_micro = ts.tv_nsec / 1000;
743 newtc->tc_offset_count = newtc->tc_get_timecount(newtc);
744 timecounter = newtc;
745 splx(s);
746 }
747 #endif
748
749 static struct timecounter *
750 sync_other_counter(void)
751 {
752 struct timecounter *tc, *tcn, *tco;
753 unsigned delta;
754
755 tco = timecounter;
756 tc = tco->tc_other;
757 tcn = tc->tc_other;
758 *tc = *tco;
759 tc->tc_other = tcn;
760 delta = tco_delta(tc);
761 tc->tc_offset_count += delta;
762 tc->tc_offset_count &= tc->tc_counter_mask;
763 tc->tc_offset_nano += (u_int64_t)delta * tc->tc_scale_nano_f;
764 tc->tc_offset_nano += (u_int64_t)delta * tc->tc_scale_nano_i << 32;
765 return (tc);
766 }
767
768 static void
769 tco_forward(int force)
770 {
771 struct timecounter *tc, *tco;
772
773 tco = timecounter;
774 tc = sync_other_counter();
775 /*
776 * We may be inducing a tiny error here, the tc_poll_pps() may
777 * process a latched count which happens after the tco_delta()
778 * in sync_other_counter(), which would extend the previous
779 * counters parameters into the domain of this new one.
780 * Since the timewindow is very small for this, the error is
781 * going to be only a few weenieseconds (as Dave Mills would
782 * say), so lets just not talk more about it, OK ?
783 */
784 if (tco->tc_poll_pps)
785 tco->tc_poll_pps(tco);
786 if (timedelta != 0) {
787 tc->tc_offset_nano += (u_int64_t)(tickdelta * 1000) << 32;
788 timedelta -= tickdelta;
789 force++;
790 }
791
792 while (tc->tc_offset_nano >= 1000000000ULL << 32) {
793 tc->tc_offset_nano -= 1000000000ULL << 32;
794 tc->tc_offset_sec++;
795 ntp_update_second(tc); /* XXX only needed if xntpd runs */
796 tco_setscales(tc);
797 force++;
798 }
799
800 if (tco_method && !force)
801 return;
802
803 tc->tc_offset_micro = (tc->tc_offset_nano / 1000) >> 32;
804
805 /* Figure out the wall-clock time */
806 tc->tc_nanotime.tv_sec = tc->tc_offset_sec + boottime.tv_sec;
807 tc->tc_nanotime.tv_nsec =
808 (tc->tc_offset_nano >> 32) + boottime.tv_usec * 1000;
809 tc->tc_microtime.tv_usec = tc->tc_offset_micro + boottime.tv_usec;
810 if (tc->tc_nanotime.tv_nsec >= 1000000000) {
811 tc->tc_nanotime.tv_nsec -= 1000000000;
812 tc->tc_microtime.tv_usec -= 1000000;
813 tc->tc_nanotime.tv_sec++;
814 }
815 time_second = tc->tc_microtime.tv_sec = tc->tc_nanotime.tv_sec;
816
817 timecounter = tc;
818 }
819
820 SYSCTL_NODE(_kern, OID_AUTO, timecounter, CTLFLAG_RW, 0, "");
821
822 SYSCTL_INT(_kern_timecounter, OID_AUTO, method, CTLFLAG_RW, &tco_method, 0,
823 "This variable determines the method used for updating timecounters. "
824 "If the default algorithm (0) fails with \"calcru negative...\" messages "
825 "try the alternate algorithm (1) which handles bad hardware better."
826
827 );
828
829
830 int
831 pps_ioctl(u_long cmd, caddr_t data, struct pps_state *pps)
832 {
833 pps_params_t *app;
834 struct pps_fetch_args *fapi;
835 #ifdef PPS_SYNC
836 struct pps_kcbind_args *kapi;
837 #endif
838
839 switch (cmd) {
840 case PPS_IOC_CREATE:
841 return (0);
842 case PPS_IOC_DESTROY:
843 return (0);
844 case PPS_IOC_SETPARAMS:
845 app = (pps_params_t *)data;
846 if (app->mode & ~pps->ppscap)
847 return (EINVAL);
848 pps->ppsparam = *app;
849 return (0);
850 case PPS_IOC_GETPARAMS:
851 app = (pps_params_t *)data;
852 *app = pps->ppsparam;
853 app->api_version = PPS_API_VERS_1;
854 return (0);
855 case PPS_IOC_GETCAP:
856 *(int*)data = pps->ppscap;
857 return (0);
858 case PPS_IOC_FETCH:
859 fapi = (struct pps_fetch_args *)data;
860 if (fapi->tsformat && fapi->tsformat != PPS_TSFMT_TSPEC)
861 return (EINVAL);
862 if (fapi->timeout.tv_sec || fapi->timeout.tv_nsec)
863 return (EOPNOTSUPP);
864 pps->ppsinfo.current_mode = pps->ppsparam.mode;
865 fapi->pps_info_buf = pps->ppsinfo;
866 return (0);
867 case PPS_IOC_KCBIND:
868 #ifdef PPS_SYNC
869 kapi = (struct pps_kcbind_args *)data;
870 /* XXX Only root should be able to do this */
871 if (kapi->tsformat && kapi->tsformat != PPS_TSFMT_TSPEC)
872 return (EINVAL);
873 if (kapi->kernel_consumer != PPS_KC_HARDPPS)
874 return (EINVAL);
875 if (kapi->edge & ~pps->ppscap)
876 return (EINVAL);
877 pps->kcmode = kapi->edge;
878 return (0);
879 #else
880 return (EOPNOTSUPP);
881 #endif
882 default:
883 return (ENOTTY);
884 }
885 }
886
887 void
888 pps_init(struct pps_state *pps)
889 {
890 pps->ppscap |= PPS_TSFMT_TSPEC;
891 if (pps->ppscap & PPS_CAPTUREASSERT)
892 pps->ppscap |= PPS_OFFSETASSERT;
893 if (pps->ppscap & PPS_CAPTURECLEAR)
894 pps->ppscap |= PPS_OFFSETCLEAR;
895 }
896
897 void
898 pps_event(struct pps_state *pps, struct timecounter *tc, unsigned count, int event)
899 {
900 struct timespec ts, *tsp, *osp;
901 u_int64_t delta;
902 unsigned tcount, *pcount;
903 int foff, fhard;
904 pps_seq_t *pseq;
905
906 /* Things would be easier with arrays... */
907 if (event == PPS_CAPTUREASSERT) {
908 tsp = &pps->ppsinfo.assert_timestamp;
909 osp = &pps->ppsparam.assert_offset;
910 foff = pps->ppsparam.mode & PPS_OFFSETASSERT;
911 fhard = pps->kcmode & PPS_CAPTUREASSERT;
912 pcount = &pps->ppscount[0];
913 pseq = &pps->ppsinfo.assert_sequence;
914 } else {
915 tsp = &pps->ppsinfo.clear_timestamp;
916 osp = &pps->ppsparam.clear_offset;
917 foff = pps->ppsparam.mode & PPS_OFFSETCLEAR;
918 fhard = pps->kcmode & PPS_CAPTURECLEAR;
919 pcount = &pps->ppscount[1];
920 pseq = &pps->ppsinfo.clear_sequence;
921 }
922
923 /* The timecounter changed: bail */
924 if (!pps->ppstc ||
925 pps->ppstc->tc_name != tc->tc_name ||
926 tc->tc_name != timecounter->tc_name) {
927 pps->ppstc = tc;
928 *pcount = count;
929 return;
930 }
931
932 /* Nothing really happened */
933 if (*pcount == count)
934 return;
935
936 *pcount = count;
937
938 /* Convert the count to timespec */
939 ts.tv_sec = tc->tc_offset_sec;
940 tcount = count - tc->tc_offset_count;
941 tcount &= tc->tc_counter_mask;
942 delta = tc->tc_offset_nano;
943 delta += ((u_int64_t)tcount * tc->tc_scale_nano_f);
944 delta >>= 32;
945 delta += ((u_int64_t)tcount * tc->tc_scale_nano_i);
946 delta += boottime.tv_usec * 1000;
947 ts.tv_sec += boottime.tv_sec;
948 while (delta >= 1000000000) {
949 delta -= 1000000000;
950 ts.tv_sec++;
951 }
952 ts.tv_nsec = delta;
953
954 (*pseq)++;
955 *tsp = ts;
956
957 if (foff) {
958 timespecadd(tsp, osp);
959 if (tsp->tv_nsec < 0) {
960 tsp->tv_nsec += 1000000000;
961 tsp->tv_sec -= 1;
962 }
963 }
964 #ifdef PPS_SYNC
965 if (fhard) {
966 /* magic, at its best... */
967 tcount = count - pps->ppscount[2];
968 pps->ppscount[2] = count;
969 tcount &= tc->tc_counter_mask;
970 delta = ((u_int64_t)tcount * tc->tc_tweak->tc_scale_nano_f);
971 delta >>= 32;
972 delta += ((u_int64_t)tcount * tc->tc_tweak->tc_scale_nano_i);
973 hardpps(tsp, delta);
974 }
975 #endif
976 }
Cache object: 972247d99a4b10dfdf14d77bca827d05
|