kern_clock.c revision 1.106.2.2 1 /* $NetBSD: kern_clock.c,v 1.106.2.2 2007/02/20 21:48:44 rmind Exp $ */
2
3 /*-
4 * Copyright (c) 2000, 2004, 2006, 2007 The NetBSD Foundation, Inc.
5 * All rights reserved.
6 *
7 * This code is derived from software contributed to The NetBSD Foundation
8 * by Jason R. Thorpe of the Numerical Aerospace Simulation Facility,
9 * NASA Ames Research Center.
10 * This code is derived from software contributed to The NetBSD Foundation
11 * by Charles M. Hannum.
12 *
13 * Redistribution and use in source and binary forms, with or without
14 * modification, are permitted provided that the following conditions
15 * are met:
16 * 1. Redistributions of source code must retain the above copyright
17 * notice, this list of conditions and the following disclaimer.
18 * 2. Redistributions in binary form must reproduce the above copyright
19 * notice, this list of conditions and the following disclaimer in the
20 * documentation and/or other materials provided with the distribution.
21 * 3. All advertising materials mentioning features or use of this software
22 * must display the following acknowledgement:
23 * This product includes software developed by the NetBSD
24 * Foundation, Inc. and its contributors.
25 * 4. Neither the name of The NetBSD Foundation nor the names of its
26 * contributors may be used to endorse or promote products derived
27 * from this software without specific prior written permission.
28 *
29 * THIS SOFTWARE IS PROVIDED BY THE NETBSD FOUNDATION, INC. AND CONTRIBUTORS
30 * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED
31 * TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
32 * PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE FOUNDATION OR CONTRIBUTORS
33 * BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
34 * CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
35 * SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
36 * INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
37 * CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
38 * ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
39 * POSSIBILITY OF SUCH DAMAGE.
40 */
41
42 /*-
43 * Copyright (c) 1982, 1986, 1991, 1993
44 * The Regents of the University of California. All rights reserved.
45 * (c) UNIX System Laboratories, Inc.
46 * All or some portions of this file are derived from material licensed
47 * to the University of California by American Telephone and Telegraph
48 * Co. or Unix System Laboratories, Inc. and are reproduced herein with
49 * the permission of UNIX System Laboratories, Inc.
50 *
51 * Redistribution and use in source and binary forms, with or without
52 * modification, are permitted provided that the following conditions
53 * are met:
54 * 1. Redistributions of source code must retain the above copyright
55 * notice, this list of conditions and the following disclaimer.
56 * 2. Redistributions in binary form must reproduce the above copyright
57 * notice, this list of conditions and the following disclaimer in the
58 * documentation and/or other materials provided with the distribution.
59 * 3. Neither the name of the University nor the names of its contributors
60 * may be used to endorse or promote products derived from this software
61 * without specific prior written permission.
62 *
63 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
64 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
65 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
66 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
67 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
68 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
69 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
70 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
71 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
72 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
73 * SUCH DAMAGE.
74 *
75 * @(#)kern_clock.c 8.5 (Berkeley) 1/21/94
76 */
77
78 #include <sys/cdefs.h>
79 __KERNEL_RCSID(0, "$NetBSD: kern_clock.c,v 1.106.2.2 2007/02/20 21:48:44 rmind Exp $");
80
81 #include "opt_ntp.h"
82 #include "opt_multiprocessor.h"
83 #include "opt_perfctrs.h"
84
85 #include <sys/param.h>
86 #include <sys/systm.h>
87 #include <sys/callout.h>
88 #include <sys/kernel.h>
89 #include <sys/proc.h>
90 #include <sys/resourcevar.h>
91 #include <sys/signalvar.h>
92 #include <sys/sysctl.h>
93 #include <sys/timex.h>
94 #include <sys/sched.h>
95 #include <sys/time.h>
96 #ifdef __HAVE_TIMECOUNTER
97 #include <sys/timetc.h>
98 #endif
99
100 #include <machine/cpu.h>
101 #include <machine/intr.h>
102
103 #ifdef GPROF
104 #include <sys/gmon.h>
105 #endif
106
107 /*
108 * Clock handling routines.
109 *
110 * This code is written to operate with two timers that run independently of
111 * each other. The main clock, running hz times per second, is used to keep
112 * track of real time. The second timer handles kernel and user profiling,
113 * and does resource use estimation. If the second timer is programmable,
114 * it is randomized to avoid aliasing between the two clocks. For example,
115 * the randomization prevents an adversary from always giving up the CPU
116 * just before its quantum expires. Otherwise, it would never accumulate
117 * CPU ticks. The mean frequency of the second timer is stathz.
118 *
119 * If no second timer exists, stathz will be zero; in this case we drive
120 * profiling and statistics off the main clock. This WILL NOT be accurate;
121 * do not do it unless absolutely necessary.
122 *
123 * The statistics clock may (or may not) be run at a higher rate while
124 * profiling. This profile clock runs at profhz. We require that profhz
125 * be an integral multiple of stathz.
126 *
127 * If the statistics clock is running fast, it must be divided by the ratio
128 * profhz/stathz for statistics. (For profiling, every tick counts.)
129 */
130
131 #ifndef __HAVE_TIMECOUNTER
132 #ifdef NTP /* NTP phase-locked loop in kernel */
133 /*
134 * Phase/frequency-lock loop (PLL/FLL) definitions
135 *
136 * The following variables are read and set by the ntp_adjtime() system
137 * call.
138 *
139 * time_state shows the state of the system clock, with values defined
140 * in the timex.h header file.
141 *
142 * time_status shows the status of the system clock, with bits defined
143 * in the timex.h header file.
144 *
145 * time_offset is used by the PLL/FLL to adjust the system time in small
146 * increments.
147 *
148 * time_constant determines the bandwidth or "stiffness" of the PLL.
149 *
150 * time_tolerance determines maximum frequency error or tolerance of the
151 * CPU clock oscillator and is a property of the architecture; however,
152 * in principle it could change as result of the presence of external
153 * discipline signals, for instance.
154 *
155 * time_precision is usually equal to the kernel tick variable; however,
156 * in cases where a precision clock counter or external clock is
157 * available, the resolution can be much less than this and depend on
158 * whether the external clock is working or not.
159 *
160 * time_maxerror is initialized by a ntp_adjtime() call and increased by
161 * the kernel once each second to reflect the maximum error bound
162 * growth.
163 *
164 * time_esterror is set and read by the ntp_adjtime() call, but
165 * otherwise not used by the kernel.
166 */
167 int time_state = TIME_OK; /* clock state */
168 int time_status = STA_UNSYNC; /* clock status bits */
169 long time_offset = 0; /* time offset (us) */
170 long time_constant = 0; /* pll time constant */
171 long time_tolerance = MAXFREQ; /* frequency tolerance (scaled ppm) */
172 long time_precision = 1; /* clock precision (us) */
173 long time_maxerror = MAXPHASE; /* maximum error (us) */
174 long time_esterror = MAXPHASE; /* estimated error (us) */
175
176 /*
177 * The following variables establish the state of the PLL/FLL and the
178 * residual time and frequency offset of the local clock. The scale
179 * factors are defined in the timex.h header file.
180 *
181 * time_phase and time_freq are the phase increment and the frequency
182 * increment, respectively, of the kernel time variable.
183 *
184 * time_freq is set via ntp_adjtime() from a value stored in a file when
185 * the synchronization daemon is first started. Its value is retrieved
186 * via ntp_adjtime() and written to the file about once per hour by the
187 * daemon.
188 *
189 * time_adj is the adjustment added to the value of tick at each timer
190 * interrupt and is recomputed from time_phase and time_freq at each
191 * seconds rollover.
192 *
193 * time_reftime is the second's portion of the system time at the last
194 * call to ntp_adjtime(). It is used to adjust the time_freq variable
195 * and to increase the time_maxerror as the time since last update
196 * increases.
197 */
198 long time_phase = 0; /* phase offset (scaled us) */
199 long time_freq = 0; /* frequency offset (scaled ppm) */
200 long time_adj = 0; /* tick adjust (scaled 1 / hz) */
201 long time_reftime = 0; /* time at last adjustment (s) */
202
203 #ifdef PPS_SYNC
204 /*
205 * The following variables are used only if the kernel PPS discipline
206 * code is configured (PPS_SYNC). The scale factors are defined in the
207 * timex.h header file.
208 *
209 * pps_time contains the time at each calibration interval, as read by
210 * microtime(). pps_count counts the seconds of the calibration
211 * interval, the duration of which is nominally pps_shift in powers of
212 * two.
213 *
214 * pps_offset is the time offset produced by the time median filter
215 * pps_tf[], while pps_jitter is the dispersion (jitter) measured by
216 * this filter.
217 *
218 * pps_freq is the frequency offset produced by the frequency median
219 * filter pps_ff[], while pps_stabil is the dispersion (wander) measured
220 * by this filter.
221 *
222 * pps_usec is latched from a high resolution counter or external clock
223 * at pps_time. Here we want the hardware counter contents only, not the
224 * contents plus the time_tv.usec as usual.
225 *
226 * pps_valid counts the number of seconds since the last PPS update. It
227 * is used as a watchdog timer to disable the PPS discipline should the
228 * PPS signal be lost.
229 *
230 * pps_glitch counts the number of seconds since the beginning of an
231 * offset burst more than tick/2 from current nominal offset. It is used
232 * mainly to suppress error bursts due to priority conflicts between the
233 * PPS interrupt and timer interrupt.
234 *
235 * pps_intcnt counts the calibration intervals for use in the interval-
236 * adaptation algorithm. It's just too complicated for words.
237 *
238 * pps_kc_hardpps_source contains an arbitrary value that uniquely
239 * identifies the currently bound source of the PPS signal, or NULL
240 * if no source is bound.
241 *
242 * pps_kc_hardpps_mode indicates which transitions, if any, of the PPS
243 * signal should be reported.
244 */
245 struct timeval pps_time; /* kernel time at last interval */
246 long pps_tf[] = {0, 0, 0}; /* pps time offset median filter (us) */
247 long pps_offset = 0; /* pps time offset (us) */
248 long pps_jitter = MAXTIME; /* time dispersion (jitter) (us) */
249 long pps_ff[] = {0, 0, 0}; /* pps frequency offset median filter */
250 long pps_freq = 0; /* frequency offset (scaled ppm) */
251 long pps_stabil = MAXFREQ; /* frequency dispersion (scaled ppm) */
252 long pps_usec = 0; /* microsec counter at last interval */
253 long pps_valid = PPS_VALID; /* pps signal watchdog counter */
254 int pps_glitch = 0; /* pps signal glitch counter */
255 int pps_count = 0; /* calibration interval counter (s) */
256 int pps_shift = PPS_SHIFT; /* interval duration (s) (shift) */
257 int pps_intcnt = 0; /* intervals at current duration */
258 void *pps_kc_hardpps_source = NULL; /* current PPS supplier's identifier */
259 int pps_kc_hardpps_mode = 0; /* interesting edges of PPS signal */
260
261 /*
262 * PPS signal quality monitors
263 *
264 * pps_jitcnt counts the seconds that have been discarded because the
265 * jitter measured by the time median filter exceeds the limit MAXTIME
266 * (100 us).
267 *
268 * pps_calcnt counts the frequency calibration intervals, which are
269 * variable from 4 s to 256 s.
270 *
271 * pps_errcnt counts the calibration intervals which have been discarded
272 * because the wander exceeds the limit MAXFREQ (100 ppm) or where the
273 * calibration interval jitter exceeds two ticks.
274 *
275 * pps_stbcnt counts the calibration intervals that have been discarded
276 * because the frequency wander exceeds the limit MAXFREQ / 4 (25 us).
277 */
278 long pps_jitcnt = 0; /* jitter limit exceeded */
279 long pps_calcnt = 0; /* calibration intervals */
280 long pps_errcnt = 0; /* calibration errors */
281 long pps_stbcnt = 0; /* stability limit exceeded */
282 #endif /* PPS_SYNC */
283
284 #ifdef EXT_CLOCK
285 /*
286 * External clock definitions
287 *
288 * The following definitions and declarations are used only if an
289 * external clock is configured on the system.
290 */
291 #define CLOCK_INTERVAL 30 /* CPU clock update interval (s) */
292
293 /*
294 * The clock_count variable is set to CLOCK_INTERVAL at each PPS
295 * interrupt and decremented once each second.
296 */
297 int clock_count = 0; /* CPU clock counter */
298
299 #ifdef HIGHBALL
300 /*
301 * The clock_offset and clock_cpu variables are used by the HIGHBALL
302 * interface. The clock_offset variable defines the offset between
303 * system time and the HIGBALL counters. The clock_cpu variable contains
304 * the offset between the system clock and the HIGHBALL clock for use in
305 * disciplining the kernel time variable.
306 */
307 extern struct timeval clock_offset; /* Highball clock offset */
308 long clock_cpu = 0; /* CPU clock adjust */
309 #endif /* HIGHBALL */
310 #endif /* EXT_CLOCK */
311 #endif /* NTP */
312
313 /*
314 * Bump a timeval by a small number of usec's.
315 */
316 #define BUMPTIME(t, usec) { \
317 volatile struct timeval *tp = (t); \
318 long us; \
319 \
320 tp->tv_usec = us = tp->tv_usec + (usec); \
321 if (us >= 1000000) { \
322 tp->tv_usec = us - 1000000; \
323 tp->tv_sec++; \
324 } \
325 }
326 #endif /* !__HAVE_TIMECOUNTER */
327
328 int stathz;
329 int profhz;
330 int profsrc;
331 int schedhz;
332 int profprocs;
333 int hardclock_ticks;
334 #ifdef SCHED_4BSD
335 static int statscheddiv; /* stat => sched divider (used if schedhz == 0) */
336 #endif
337 static int psdiv; /* prof => stat divider */
338 int psratio; /* ratio: prof / stat */
339 #ifndef __HAVE_TIMECOUNTER
340 int tickfix, tickfixinterval; /* used if tick not really integral */
341 #ifndef NTP
342 static int tickfixcnt; /* accumulated fractional error */
343 #else
344 int fixtick; /* used by NTP for same */
345 int shifthz;
346 #endif
347
348 /*
349 * We might want ldd to load the both words from time at once.
350 * To succeed we need to be quadword aligned.
351 * The sparc already does that, and that it has worked so far is a fluke.
352 */
353 volatile struct timeval time __attribute__((__aligned__(__alignof__(quad_t))));
354 volatile struct timeval mono_time;
355 #endif /* !__HAVE_TIMECOUNTER */
356
357 void *softclock_si;
358
359 #ifdef __HAVE_TIMECOUNTER
360 static u_int get_intr_timecount(struct timecounter *);
361
362 static struct timecounter intr_timecounter = {
363 get_intr_timecount, /* get_timecount */
364 0, /* no poll_pps */
365 ~0u, /* counter_mask */
366 0, /* frequency */
367 "clockinterrupt", /* name */
368 0, /* quality - minimum implementation level for a clock */
369 NULL, /* prev */
370 NULL, /* next */
371 };
372
373 static u_int
374 get_intr_timecount(struct timecounter *tc)
375 {
376
377 return (u_int)hardclock_ticks;
378 }
379 #endif
380
381 /*
382 * Initialize clock frequencies and start both clocks running.
383 */
384 void
385 initclocks(void)
386 {
387 int i;
388
389 softclock_si = softintr_establish(IPL_SOFTCLOCK, softclock, NULL);
390 if (softclock_si == NULL)
391 panic("initclocks: unable to register softclock intr");
392
393 /*
394 * Set divisors to 1 (normal case) and let the machine-specific
395 * code do its bit.
396 */
397 psdiv = 1;
398 #ifdef __HAVE_TIMECOUNTER
399 /*
400 * provide minimum default time counter
401 * will only run at interrupt resolution
402 */
403 intr_timecounter.tc_frequency = hz;
404 tc_init(&intr_timecounter);
405 #endif
406 cpu_initclocks();
407
408 /*
409 * Compute profhz and stathz, fix profhz if needed.
410 */
411 i = stathz ? stathz : hz;
412 if (profhz == 0)
413 profhz = i;
414 psratio = profhz / i;
415 #ifdef SCHED_4BSD
416 if (schedhz == 0) {
417 /* 16Hz is best */
418 statscheddiv = i / 16;
419 if (statscheddiv <= 0)
420 panic("statscheddiv");
421 }
422 #endif /* SCHED_4BSD */
423
424 #ifndef __HAVE_TIMECOUNTER
425 #ifdef NTP
426 switch (hz) {
427 case 1:
428 shifthz = SHIFT_SCALE - 0;
429 break;
430 case 2:
431 shifthz = SHIFT_SCALE - 1;
432 break;
433 case 4:
434 shifthz = SHIFT_SCALE - 2;
435 break;
436 case 8:
437 shifthz = SHIFT_SCALE - 3;
438 break;
439 case 16:
440 shifthz = SHIFT_SCALE - 4;
441 break;
442 case 32:
443 shifthz = SHIFT_SCALE - 5;
444 break;
445 case 50:
446 case 60:
447 case 64:
448 shifthz = SHIFT_SCALE - 6;
449 break;
450 case 96:
451 case 100:
452 case 128:
453 shifthz = SHIFT_SCALE - 7;
454 break;
455 case 256:
456 shifthz = SHIFT_SCALE - 8;
457 break;
458 case 512:
459 shifthz = SHIFT_SCALE - 9;
460 break;
461 case 1000:
462 case 1024:
463 shifthz = SHIFT_SCALE - 10;
464 break;
465 case 1200:
466 case 2048:
467 shifthz = SHIFT_SCALE - 11;
468 break;
469 case 4096:
470 shifthz = SHIFT_SCALE - 12;
471 break;
472 case 8192:
473 shifthz = SHIFT_SCALE - 13;
474 break;
475 case 16384:
476 shifthz = SHIFT_SCALE - 14;
477 break;
478 case 32768:
479 shifthz = SHIFT_SCALE - 15;
480 break;
481 case 65536:
482 shifthz = SHIFT_SCALE - 16;
483 break;
484 default:
485 panic("weird hz");
486 }
487 if (fixtick == 0) {
488 /*
489 * Give MD code a chance to set this to a better
490 * value; but, if it doesn't, we should.
491 */
492 fixtick = (1000000 - (hz*tick));
493 }
494 #endif /* NTP */
495 #endif /* !__HAVE_TIMECOUNTER */
496 }
497
498 /*
499 * The real-time timer, interrupting hz times per second.
500 */
501 void
502 hardclock(struct clockframe *frame)
503 {
504 struct lwp *l;
505 struct proc *p;
506 struct cpu_info *ci = curcpu();
507 struct ptimer *pt;
508 #ifndef __HAVE_TIMECOUNTER
509 int delta;
510 extern int tickdelta;
511 extern long timedelta;
512 #ifdef NTP
513 int time_update;
514 int ltemp;
515 #endif /* NTP */
516 #endif /* __HAVE_TIMECOUNTER */
517
518 l = curlwp;
519 if (!CURCPU_IDLE_P()) {
520 p = l->l_proc;
521 /*
522 * Run current process's virtual and profile time, as needed.
523 */
524 if (CLKF_USERMODE(frame) && p->p_timers &&
525 (pt = LIST_FIRST(&p->p_timers->pts_virtual)) != NULL)
526 if (itimerdecr(pt, tick) == 0)
527 itimerfire(pt);
528 if (p->p_timers &&
529 (pt = LIST_FIRST(&p->p_timers->pts_prof)) != NULL)
530 if (itimerdecr(pt, tick) == 0)
531 itimerfire(pt);
532 }
533
534 /*
535 * If no separate statistics clock is available, run it from here.
536 */
537 if (stathz == 0)
538 statclock(frame);
539 if ((--ci->ci_schedstate.spc_ticks) <= 0)
540 sched_tick(ci);
541
542 #if defined(MULTIPROCESSOR)
543 /*
544 * If we are not the primary CPU, we're not allowed to do
545 * any more work.
546 */
547 if (CPU_IS_PRIMARY(ci) == 0)
548 return;
549 #endif
550
551 hardclock_ticks++;
552
553 #ifdef __HAVE_TIMECOUNTER
554 tc_ticktock();
555 #else /* __HAVE_TIMECOUNTER */
556 /*
557 * Increment the time-of-day. The increment is normally just
558 * ``tick''. If the machine is one which has a clock frequency
559 * such that ``hz'' would not divide the second evenly into
560 * milliseconds, a periodic adjustment must be applied. Finally,
561 * if we are still adjusting the time (see adjtime()),
562 * ``tickdelta'' may also be added in.
563 */
564 delta = tick;
565
566 #ifndef NTP
567 if (tickfix) {
568 tickfixcnt += tickfix;
569 if (tickfixcnt >= tickfixinterval) {
570 delta++;
571 tickfixcnt -= tickfixinterval;
572 }
573 }
574 #endif /* !NTP */
575 /* Imprecise 4bsd adjtime() handling */
576 if (timedelta != 0) {
577 delta += tickdelta;
578 timedelta -= tickdelta;
579 }
580
581 #ifdef notyet
582 microset();
583 #endif
584
585 #ifndef NTP
586 BUMPTIME(&time, delta); /* XXX Now done using NTP code below */
587 #endif
588 BUMPTIME(&mono_time, delta);
589
590 #ifdef NTP
591 time_update = delta;
592
593 /*
594 * Compute the phase adjustment. If the low-order bits
595 * (time_phase) of the update overflow, bump the high-order bits
596 * (time_update).
597 */
598 time_phase += time_adj;
599 if (time_phase <= -FINEUSEC) {
600 ltemp = -time_phase >> SHIFT_SCALE;
601 time_phase += ltemp << SHIFT_SCALE;
602 time_update -= ltemp;
603 } else if (time_phase >= FINEUSEC) {
604 ltemp = time_phase >> SHIFT_SCALE;
605 time_phase -= ltemp << SHIFT_SCALE;
606 time_update += ltemp;
607 }
608
609 #ifdef HIGHBALL
610 /*
611 * If the HIGHBALL board is installed, we need to adjust the
612 * external clock offset in order to close the hardware feedback
613 * loop. This will adjust the external clock phase and frequency
614 * in small amounts. The additional phase noise and frequency
615 * wander this causes should be minimal. We also need to
616 * discipline the kernel time variable, since the PLL is used to
617 * discipline the external clock. If the Highball board is not
618 * present, we discipline kernel time with the PLL as usual. We
619 * assume that the external clock phase adjustment (time_update)
620 * and kernel phase adjustment (clock_cpu) are less than the
621 * value of tick.
622 */
623 clock_offset.tv_usec += time_update;
624 if (clock_offset.tv_usec >= 1000000) {
625 clock_offset.tv_sec++;
626 clock_offset.tv_usec -= 1000000;
627 }
628 if (clock_offset.tv_usec < 0) {
629 clock_offset.tv_sec--;
630 clock_offset.tv_usec += 1000000;
631 }
632 time.tv_usec += clock_cpu;
633 clock_cpu = 0;
634 #else
635 time.tv_usec += time_update;
636 #endif /* HIGHBALL */
637
638 /*
639 * On rollover of the second the phase adjustment to be used for
640 * the next second is calculated. Also, the maximum error is
641 * increased by the tolerance. If the PPS frequency discipline
642 * code is present, the phase is increased to compensate for the
643 * CPU clock oscillator frequency error.
644 *
645 * On a 32-bit machine and given parameters in the timex.h
646 * header file, the maximum phase adjustment is +-512 ms and
647 * maximum frequency offset is a tad less than) +-512 ppm. On a
648 * 64-bit machine, you shouldn't need to ask.
649 */
650 if (time.tv_usec >= 1000000) {
651 time.tv_usec -= 1000000;
652 time.tv_sec++;
653 time_maxerror += time_tolerance >> SHIFT_USEC;
654
655 /*
656 * Leap second processing. If in leap-insert state at
657 * the end of the day, the system clock is set back one
658 * second; if in leap-delete state, the system clock is
659 * set ahead one second. The microtime() routine or
660 * external clock driver will insure that reported time
661 * is always monotonic. The ugly divides should be
662 * replaced.
663 */
664 switch (time_state) {
665 case TIME_OK:
666 if (time_status & STA_INS)
667 time_state = TIME_INS;
668 else if (time_status & STA_DEL)
669 time_state = TIME_DEL;
670 break;
671
672 case TIME_INS:
673 if (time.tv_sec % 86400 == 0) {
674 time.tv_sec--;
675 time_state = TIME_OOP;
676 }
677 break;
678
679 case TIME_DEL:
680 if ((time.tv_sec + 1) % 86400 == 0) {
681 time.tv_sec++;
682 time_state = TIME_WAIT;
683 }
684 break;
685
686 case TIME_OOP:
687 time_state = TIME_WAIT;
688 break;
689
690 case TIME_WAIT:
691 if (!(time_status & (STA_INS | STA_DEL)))
692 time_state = TIME_OK;
693 break;
694 }
695
696 /*
697 * Compute the phase adjustment for the next second. In
698 * PLL mode, the offset is reduced by a fixed factor
699 * times the time constant. In FLL mode the offset is
700 * used directly. In either mode, the maximum phase
701 * adjustment for each second is clamped so as to spread
702 * the adjustment over not more than the number of
703 * seconds between updates.
704 */
705 if (time_offset < 0) {
706 ltemp = -time_offset;
707 if (!(time_status & STA_FLL))
708 ltemp >>= SHIFT_KG + time_constant;
709 if (ltemp > (MAXPHASE / MINSEC) << SHIFT_UPDATE)
710 ltemp = (MAXPHASE / MINSEC) <<
711 SHIFT_UPDATE;
712 time_offset += ltemp;
713 time_adj = -ltemp << (shifthz - SHIFT_UPDATE);
714 } else if (time_offset > 0) {
715 ltemp = time_offset;
716 if (!(time_status & STA_FLL))
717 ltemp >>= SHIFT_KG + time_constant;
718 if (ltemp > (MAXPHASE / MINSEC) << SHIFT_UPDATE)
719 ltemp = (MAXPHASE / MINSEC) <<
720 SHIFT_UPDATE;
721 time_offset -= ltemp;
722 time_adj = ltemp << (shifthz - SHIFT_UPDATE);
723 } else
724 time_adj = 0;
725
726 /*
727 * Compute the frequency estimate and additional phase
728 * adjustment due to frequency error for the next
729 * second. When the PPS signal is engaged, gnaw on the
730 * watchdog counter and update the frequency computed by
731 * the pll and the PPS signal.
732 */
733 #ifdef PPS_SYNC
734 pps_valid++;
735 if (pps_valid == PPS_VALID) {
736 pps_jitter = MAXTIME;
737 pps_stabil = MAXFREQ;
738 time_status &= ~(STA_PPSSIGNAL | STA_PPSJITTER |
739 STA_PPSWANDER | STA_PPSERROR);
740 }
741 ltemp = time_freq + pps_freq;
742 #else
743 ltemp = time_freq;
744 #endif /* PPS_SYNC */
745
746 if (ltemp < 0)
747 time_adj -= -ltemp >> (SHIFT_USEC - shifthz);
748 else
749 time_adj += ltemp >> (SHIFT_USEC - shifthz);
750 time_adj += (long)fixtick << shifthz;
751
752 /*
753 * When the CPU clock oscillator frequency is not a
754 * power of 2 in Hz, shifthz is only an approximate
755 * scale factor.
756 *
757 * To determine the adjustment, you can do the following:
758 * bc -q
759 * scale=24
760 * obase=2
761 * idealhz/realhz
762 * where `idealhz' is the next higher power of 2, and `realhz'
763 * is the actual value. You may need to factor this result
764 * into a sequence of 2 multipliers to get better precision.
765 *
766 * Likewise, the error can be calculated with (e.g. for 100Hz):
767 * bc -q
768 * scale=24
769 * ((1+2^-2+2^-5)*(1-2^-10)*realhz-idealhz)/idealhz
770 * (and then multiply by 1000000 to get ppm).
771 */
772 switch (hz) {
773 case 60:
774 /* A factor of 1.000100010001 gives about 15ppm
775 error. */
776 if (time_adj < 0) {
777 time_adj -= (-time_adj >> 4);
778 time_adj -= (-time_adj >> 8);
779 } else {
780 time_adj += (time_adj >> 4);
781 time_adj += (time_adj >> 8);
782 }
783 break;
784
785 case 96:
786 /* A factor of 1.0101010101 gives about 244ppm error. */
787 if (time_adj < 0) {
788 time_adj -= (-time_adj >> 2);
789 time_adj -= (-time_adj >> 4) + (-time_adj >> 8);
790 } else {
791 time_adj += (time_adj >> 2);
792 time_adj += (time_adj >> 4) + (time_adj >> 8);
793 }
794 break;
795
796 case 50:
797 case 100:
798 /* A factor of 1.010001111010111 gives about 1ppm
799 error. */
800 if (time_adj < 0) {
801 time_adj -= (-time_adj >> 2) + (-time_adj >> 5);
802 time_adj += (-time_adj >> 10);
803 } else {
804 time_adj += (time_adj >> 2) + (time_adj >> 5);
805 time_adj -= (time_adj >> 10);
806 }
807 break;
808
809 case 1000:
810 /* A factor of 1.000001100010100001 gives about 50ppm
811 error. */
812 if (time_adj < 0) {
813 time_adj -= (-time_adj >> 6) + (-time_adj >> 11);
814 time_adj -= (-time_adj >> 7);
815 } else {
816 time_adj += (time_adj >> 6) + (time_adj >> 11);
817 time_adj += (time_adj >> 7);
818 }
819 break;
820
821 case 1200:
822 /* A factor of 1.1011010011100001 gives about 64ppm
823 error. */
824 if (time_adj < 0) {
825 time_adj -= (-time_adj >> 1) + (-time_adj >> 6);
826 time_adj -= (-time_adj >> 3) + (-time_adj >> 10);
827 } else {
828 time_adj += (time_adj >> 1) + (time_adj >> 6);
829 time_adj += (time_adj >> 3) + (time_adj >> 10);
830 }
831 break;
832 }
833
834 #ifdef EXT_CLOCK
835 /*
836 * If an external clock is present, it is necessary to
837 * discipline the kernel time variable anyway, since not
838 * all system components use the microtime() interface.
839 * Here, the time offset between the external clock and
840 * kernel time variable is computed every so often.
841 */
842 clock_count++;
843 if (clock_count > CLOCK_INTERVAL) {
844 clock_count = 0;
845 microtime(&clock_ext);
846 delta.tv_sec = clock_ext.tv_sec - time.tv_sec;
847 delta.tv_usec = clock_ext.tv_usec -
848 time.tv_usec;
849 if (delta.tv_usec < 0)
850 delta.tv_sec--;
851 if (delta.tv_usec >= 500000) {
852 delta.tv_usec -= 1000000;
853 delta.tv_sec++;
854 }
855 if (delta.tv_usec < -500000) {
856 delta.tv_usec += 1000000;
857 delta.tv_sec--;
858 }
859 if (delta.tv_sec > 0 || (delta.tv_sec == 0 &&
860 delta.tv_usec > MAXPHASE) ||
861 delta.tv_sec < -1 || (delta.tv_sec == -1 &&
862 delta.tv_usec < -MAXPHASE)) {
863 time = clock_ext;
864 delta.tv_sec = 0;
865 delta.tv_usec = 0;
866 }
867 #ifdef HIGHBALL
868 clock_cpu = delta.tv_usec;
869 #else /* HIGHBALL */
870 hardupdate(delta.tv_usec);
871 #endif /* HIGHBALL */
872 }
873 #endif /* EXT_CLOCK */
874 }
875
876 #endif /* NTP */
877 #endif /* !__HAVE_TIMECOUNTER */
878
879 /*
880 * Update real-time timeout queue. Callouts are processed at a
881 * very low CPU priority, so we don't keep the relatively high
882 * clock interrupt priority any longer than necessary.
883 */
884 if (callout_hardclock())
885 softintr_schedule(softclock_si);
886 }
887
888 #ifdef __HAVE_TIMECOUNTER
889 /*
890 * Compute number of hz until specified time. Used to compute second
891 * argument to callout_reset() from an absolute time.
892 */
893 int
894 hzto(struct timeval *tvp)
895 {
896 struct timeval now, tv;
897
898 tv = *tvp; /* Don't modify original tvp. */
899 getmicrotime(&now);
900 timersub(&tv, &now, &tv);
901 return tvtohz(&tv);
902 }
903 #endif /* __HAVE_TIMECOUNTER */
904
905 /*
906 * Compute number of ticks in the specified amount of time.
907 */
908 int
909 tvtohz(struct timeval *tv)
910 {
911 unsigned long ticks;
912 long sec, usec;
913
914 /*
915 * If the number of usecs in the whole seconds part of the time
916 * difference fits in a long, then the total number of usecs will
917 * fit in an unsigned long. Compute the total and convert it to
918 * ticks, rounding up and adding 1 to allow for the current tick
919 * to expire. Rounding also depends on unsigned long arithmetic
920 * to avoid overflow.
921 *
922 * Otherwise, if the number of ticks in the whole seconds part of
923 * the time difference fits in a long, then convert the parts to
924 * ticks separately and add, using similar rounding methods and
925 * overflow avoidance. This method would work in the previous
926 * case, but it is slightly slower and assumes that hz is integral.
927 *
928 * Otherwise, round the time difference down to the maximum
929 * representable value.
930 *
931 * If ints are 32-bit, then the maximum value for any timeout in
932 * 10ms ticks is 248 days.
933 */
934 sec = tv->tv_sec;
935 usec = tv->tv_usec;
936
937 if (usec < 0) {
938 sec--;
939 usec += 1000000;
940 }
941
942 if (sec < 0 || (sec == 0 && usec <= 0)) {
943 /*
944 * Would expire now or in the past. Return 0 ticks.
945 * This is different from the legacy hzto() interface,
946 * and callers need to check for it.
947 */
948 ticks = 0;
949 } else if (sec <= (LONG_MAX / 1000000))
950 ticks = (((sec * 1000000) + (unsigned long)usec + (tick - 1))
951 / tick) + 1;
952 else if (sec <= (LONG_MAX / hz))
953 ticks = (sec * hz) +
954 (((unsigned long)usec + (tick - 1)) / tick) + 1;
955 else
956 ticks = LONG_MAX;
957
958 if (ticks > INT_MAX)
959 ticks = INT_MAX;
960
961 return ((int)ticks);
962 }
963
964 #ifndef __HAVE_TIMECOUNTER
965 /*
966 * Compute number of hz until specified time. Used to compute second
967 * argument to callout_reset() from an absolute time.
968 */
969 int
970 hzto(struct timeval *tv)
971 {
972 unsigned long ticks;
973 long sec, usec;
974 int s;
975
976 /*
977 * If the number of usecs in the whole seconds part of the time
978 * difference fits in a long, then the total number of usecs will
979 * fit in an unsigned long. Compute the total and convert it to
980 * ticks, rounding up and adding 1 to allow for the current tick
981 * to expire. Rounding also depends on unsigned long arithmetic
982 * to avoid overflow.
983 *
984 * Otherwise, if the number of ticks in the whole seconds part of
985 * the time difference fits in a long, then convert the parts to
986 * ticks separately and add, using similar rounding methods and
987 * overflow avoidance. This method would work in the previous
988 * case, but it is slightly slower and assume that hz is integral.
989 *
990 * Otherwise, round the time difference down to the maximum
991 * representable value.
992 *
993 * If ints are 32-bit, then the maximum value for any timeout in
994 * 10ms ticks is 248 days.
995 */
996 s = splclock();
997 sec = tv->tv_sec - time.tv_sec;
998 usec = tv->tv_usec - time.tv_usec;
999 splx(s);
1000
1001 if (usec < 0) {
1002 sec--;
1003 usec += 1000000;
1004 }
1005
1006 if (sec < 0 || (sec == 0 && usec <= 0)) {
1007 /*
1008 * Would expire now or in the past. Return 0 ticks.
1009 * This is different from the legacy hzto() interface,
1010 * and callers need to check for it.
1011 */
1012 ticks = 0;
1013 } else if (sec <= (LONG_MAX / 1000000))
1014 ticks = (((sec * 1000000) + (unsigned long)usec + (tick - 1))
1015 / tick) + 1;
1016 else if (sec <= (LONG_MAX / hz))
1017 ticks = (sec * hz) +
1018 (((unsigned long)usec + (tick - 1)) / tick) + 1;
1019 else
1020 ticks = LONG_MAX;
1021
1022 if (ticks > INT_MAX)
1023 ticks = INT_MAX;
1024
1025 return ((int)ticks);
1026 }
1027 #endif /* !__HAVE_TIMECOUNTER */
1028
1029 /*
1030 * Compute number of ticks in the specified amount of time.
1031 */
1032 int
1033 tstohz(struct timespec *ts)
1034 {
1035 struct timeval tv;
1036
1037 /*
1038 * usec has great enough resolution for hz, so convert to a
1039 * timeval and use tvtohz() above.
1040 */
1041 TIMESPEC_TO_TIMEVAL(&tv, ts);
1042 return tvtohz(&tv);
1043 }
1044
1045 /*
1046 * Start profiling on a process.
1047 *
1048 * Kernel profiling passes proc0 which never exits and hence
1049 * keeps the profile clock running constantly.
1050 */
1051 void
1052 startprofclock(struct proc *p)
1053 {
1054
1055 LOCK_ASSERT(mutex_owned(&p->p_stmutex));
1056
1057 if ((p->p_stflag & PST_PROFIL) == 0) {
1058 p->p_stflag |= PST_PROFIL;
1059 /*
1060 * This is only necessary if using the clock as the
1061 * profiling source.
1062 */
1063 if (++profprocs == 1 && stathz != 0)
1064 psdiv = psratio;
1065 }
1066 }
1067
1068 /*
1069 * Stop profiling on a process.
1070 */
1071 void
1072 stopprofclock(struct proc *p)
1073 {
1074
1075 LOCK_ASSERT(mutex_owned(&p->p_stmutex));
1076
1077 if (p->p_stflag & PST_PROFIL) {
1078 p->p_stflag &= ~PST_PROFIL;
1079 /*
1080 * This is only necessary if using the clock as the
1081 * profiling source.
1082 */
1083 if (--profprocs == 0 && stathz != 0)
1084 psdiv = 1;
1085 }
1086 }
1087
1088 #if defined(PERFCTRS)
1089 /*
1090 * Independent profiling "tick" in case we're using a separate
1091 * clock or profiling event source. Currently, that's just
1092 * performance counters--hence the wrapper.
1093 */
1094 void
1095 proftick(struct clockframe *frame)
1096 {
1097 #ifdef GPROF
1098 struct gmonparam *g;
1099 intptr_t i;
1100 #endif
1101 struct lwp *l;
1102 struct proc *p;
1103
1104 l = curlwp;
1105 p = (l ? l->l_proc : NULL);
1106 if (CLKF_USERMODE(frame)) {
1107 mutex_spin_enter(&p->p_stmutex);
1108 if (p->p_stflag & PST_PROFIL)
1109 addupc_intr(l, CLKF_PC(frame));
1110 mutex_spin_exit(&p->p_stmutex);
1111 } else {
1112 #ifdef GPROF
1113 g = &_gmonparam;
1114 if (g->state == GMON_PROF_ON) {
1115 i = CLKF_PC(frame) - g->lowpc;
1116 if (i < g->textsize) {
1117 i /= HISTFRACTION * sizeof(*g->kcount);
1118 g->kcount[i]++;
1119 }
1120 }
1121 #endif
1122 #ifdef PROC_PC
1123 if (p != NULL) {
1124 mutex_spin_enter(&p->p_stmutex);
1125 if (p->p_stflag & PST_PROFIL))
1126 addupc_intr(l, PROC_PC(p));
1127 mutex_spin_exit(&p->p_stmutex);
1128 }
1129 #endif
1130 }
1131 }
1132 #endif
1133
1134 /*
1135 * Statistics clock. Grab profile sample, and if divider reaches 0,
1136 * do process and kernel statistics.
1137 */
1138 void
1139 statclock(struct clockframe *frame)
1140 {
1141 #ifdef GPROF
1142 struct gmonparam *g;
1143 intptr_t i;
1144 #endif
1145 struct cpu_info *ci = curcpu();
1146 struct schedstate_percpu *spc = &ci->ci_schedstate;
1147 struct proc *p;
1148 struct lwp *l;
1149
1150 /*
1151 * Notice changes in divisor frequency, and adjust clock
1152 * frequency accordingly.
1153 */
1154 if (spc->spc_psdiv != psdiv) {
1155 spc->spc_psdiv = psdiv;
1156 spc->spc_pscnt = psdiv;
1157 if (psdiv == 1) {
1158 setstatclockrate(stathz);
1159 } else {
1160 setstatclockrate(profhz);
1161 }
1162 }
1163 l = curlwp;
1164 if ((l->l_flag & L_IDLE) != 0) {
1165 /*
1166 * don't account idle lwps as swapper.
1167 */
1168 p = NULL;
1169 } else {
1170 p = l->l_proc;
1171 mutex_spin_enter(&p->p_stmutex);
1172 }
1173
1174 if (CLKF_USERMODE(frame)) {
1175 if ((p->p_stflag & PST_PROFIL) && profsrc == PROFSRC_CLOCK)
1176 addupc_intr(l, CLKF_PC(frame));
1177 if (--spc->spc_pscnt > 0) {
1178 mutex_spin_exit(&p->p_stmutex);
1179 return;
1180 }
1181
1182 /*
1183 * Came from user mode; CPU was in user state.
1184 * If this process is being profiled record the tick.
1185 */
1186 p->p_uticks++;
1187 if (p->p_nice > NZERO)
1188 spc->spc_cp_time[CP_NICE]++;
1189 else
1190 spc->spc_cp_time[CP_USER]++;
1191 } else {
1192 #ifdef GPROF
1193 /*
1194 * Kernel statistics are just like addupc_intr, only easier.
1195 */
1196 g = &_gmonparam;
1197 if (profsrc == PROFSRC_CLOCK && g->state == GMON_PROF_ON) {
1198 i = CLKF_PC(frame) - g->lowpc;
1199 if (i < g->textsize) {
1200 i /= HISTFRACTION * sizeof(*g->kcount);
1201 g->kcount[i]++;
1202 }
1203 }
1204 #endif
1205 #ifdef LWP_PC
1206 if (p != NULL && profsrc == PROFSRC_CLOCK &&
1207 (p->p_stflag & PST_PROFIL)) {
1208 addupc_intr(l, LWP_PC(l));
1209 }
1210 #endif
1211 if (--spc->spc_pscnt > 0) {
1212 if (p != NULL)
1213 mutex_spin_exit(&p->p_stmutex);
1214 return;
1215 }
1216 /*
1217 * Came from kernel mode, so we were:
1218 * - handling an interrupt,
1219 * - doing syscall or trap work on behalf of the current
1220 * user process, or
1221 * - spinning in the idle loop.
1222 * Whichever it is, charge the time as appropriate.
1223 * Note that we charge interrupts to the current process,
1224 * regardless of whether they are ``for'' that process,
1225 * so that we know how much of its real time was spent
1226 * in ``non-process'' (i.e., interrupt) work.
1227 */
1228 if (CLKF_INTR(frame)) {
1229 if (p != NULL) {
1230 p->p_iticks++;
1231 }
1232 spc->spc_cp_time[CP_INTR]++;
1233 } else if (p != NULL) {
1234 p->p_sticks++;
1235 spc->spc_cp_time[CP_SYS]++;
1236 } else {
1237 spc->spc_cp_time[CP_IDLE]++;
1238 }
1239 }
1240 spc->spc_pscnt = psdiv;
1241
1242 if (p == NULL) {
1243 return;
1244 }
1245
1246 ++p->p_cpticks;
1247 mutex_spin_exit(&p->p_stmutex);
1248
1249 #ifdef SCHED_4BSD
1250 /*
1251 * If no separate schedclock is provided, call it here
1252 * at about 16 Hz.
1253 */
1254 if (schedhz == 0) {
1255 if ((int)(--ci->ci_schedstate.spc_schedticks) <= 0) {
1256 schedclock(l);
1257 ci->ci_schedstate.spc_schedticks = statscheddiv;
1258 }
1259 }
1260 #endif /* SCHED_4BSD */
1261 }
1262
1263 #ifndef __HAVE_TIMECOUNTER
1264 #ifdef NTP /* NTP phase-locked loop in kernel */
1265 /*
1266 * hardupdate() - local clock update
1267 *
1268 * This routine is called by ntp_adjtime() to update the local clock
1269 * phase and frequency. The implementation is of an adaptive-parameter,
1270 * hybrid phase/frequency-lock loop (PLL/FLL). The routine computes new
1271 * time and frequency offset estimates for each call. If the kernel PPS
1272 * discipline code is configured (PPS_SYNC), the PPS signal itself
1273 * determines the new time offset, instead of the calling argument.
1274 * Presumably, calls to ntp_adjtime() occur only when the caller
1275 * believes the local clock is valid within some bound (+-128 ms with
1276 * NTP). If the caller's time is far different than the PPS time, an
1277 * argument will ensue, and it's not clear who will lose.
1278 *
1279 * For uncompensated quartz crystal oscillatores and nominal update
1280 * intervals less than 1024 s, operation should be in phase-lock mode
1281 * (STA_FLL = 0), where the loop is disciplined to phase. For update
1282 * intervals greater than thiss, operation should be in frequency-lock
1283 * mode (STA_FLL = 1), where the loop is disciplined to frequency.
1284 *
1285 * Note: splclock() is in effect.
1286 */
1287 void
1288 hardupdate(long offset)
1289 {
1290 long ltemp, mtemp;
1291
1292 if (!(time_status & STA_PLL) && !(time_status & STA_PPSTIME))
1293 return;
1294 ltemp = offset;
1295 #ifdef PPS_SYNC
1296 if (time_status & STA_PPSTIME && time_status & STA_PPSSIGNAL)
1297 ltemp = pps_offset;
1298 #endif /* PPS_SYNC */
1299
1300 /*
1301 * Scale the phase adjustment and clamp to the operating range.
1302 */
1303 if (ltemp > MAXPHASE)
1304 time_offset = MAXPHASE << SHIFT_UPDATE;
1305 else if (ltemp < -MAXPHASE)
1306 time_offset = -(MAXPHASE << SHIFT_UPDATE);
1307 else
1308 time_offset = ltemp << SHIFT_UPDATE;
1309
1310 /*
1311 * Select whether the frequency is to be controlled and in which
1312 * mode (PLL or FLL). Clamp to the operating range. Ugly
1313 * multiply/divide should be replaced someday.
1314 */
1315 if (time_status & STA_FREQHOLD || time_reftime == 0)
1316 time_reftime = time.tv_sec;
1317 mtemp = time.tv_sec - time_reftime;
1318 time_reftime = time.tv_sec;
1319 if (time_status & STA_FLL) {
1320 if (mtemp >= MINSEC) {
1321 ltemp = ((time_offset / mtemp) << (SHIFT_USEC -
1322 SHIFT_UPDATE));
1323 if (ltemp < 0)
1324 time_freq -= -ltemp >> SHIFT_KH;
1325 else
1326 time_freq += ltemp >> SHIFT_KH;
1327 }
1328 } else {
1329 if (mtemp < MAXSEC) {
1330 ltemp *= mtemp;
1331 if (ltemp < 0)
1332 time_freq -= -ltemp >> (time_constant +
1333 time_constant + SHIFT_KF -
1334 SHIFT_USEC);
1335 else
1336 time_freq += ltemp >> (time_constant +
1337 time_constant + SHIFT_KF -
1338 SHIFT_USEC);
1339 }
1340 }
1341 if (time_freq > time_tolerance)
1342 time_freq = time_tolerance;
1343 else if (time_freq < -time_tolerance)
1344 time_freq = -time_tolerance;
1345 }
1346
1347 #ifdef PPS_SYNC
1348 /*
1349 * hardpps() - discipline CPU clock oscillator to external PPS signal
1350 *
1351 * This routine is called at each PPS interrupt in order to discipline
1352 * the CPU clock oscillator to the PPS signal. It measures the PPS phase
1353 * and leaves it in a handy spot for the hardclock() routine. It
1354 * integrates successive PPS phase differences and calculates the
1355 * frequency offset. This is used in hardclock() to discipline the CPU
1356 * clock oscillator so that intrinsic frequency error is cancelled out.
1357 * The code requires the caller to capture the time and hardware counter
1358 * value at the on-time PPS signal transition.
1359 *
1360 * Note that, on some Unix systems, this routine runs at an interrupt
1361 * priority level higher than the timer interrupt routine hardclock().
1362 * Therefore, the variables used are distinct from the hardclock()
1363 * variables, except for certain exceptions: The PPS frequency pps_freq
1364 * and phase pps_offset variables are determined by this routine and
1365 * updated atomically. The time_tolerance variable can be considered a
1366 * constant, since it is infrequently changed, and then only when the
1367 * PPS signal is disabled. The watchdog counter pps_valid is updated
1368 * once per second by hardclock() and is atomically cleared in this
1369 * routine.
1370 */
1371 void
1372 hardpps(struct timeval *tvp, /* time at PPS */
1373 long usec /* hardware counter at PPS */)
1374 {
1375 long u_usec, v_usec, bigtick;
1376 long cal_sec, cal_usec;
1377
1378 /*
1379 * An occasional glitch can be produced when the PPS interrupt
1380 * occurs in the hardclock() routine before the time variable is
1381 * updated. Here the offset is discarded when the difference
1382 * between it and the last one is greater than tick/2, but not
1383 * if the interval since the first discard exceeds 30 s.
1384 */
1385 time_status |= STA_PPSSIGNAL;
1386 time_status &= ~(STA_PPSJITTER | STA_PPSWANDER | STA_PPSERROR);
1387 pps_valid = 0;
1388 u_usec = -tvp->tv_usec;
1389 if (u_usec < -500000)
1390 u_usec += 1000000;
1391 v_usec = pps_offset - u_usec;
1392 if (v_usec < 0)
1393 v_usec = -v_usec;
1394 if (v_usec > (tick >> 1)) {
1395 if (pps_glitch > MAXGLITCH) {
1396 pps_glitch = 0;
1397 pps_tf[2] = u_usec;
1398 pps_tf[1] = u_usec;
1399 } else {
1400 pps_glitch++;
1401 u_usec = pps_offset;
1402 }
1403 } else
1404 pps_glitch = 0;
1405
1406 /*
1407 * A three-stage median filter is used to help deglitch the pps
1408 * time. The median sample becomes the time offset estimate; the
1409 * difference between the other two samples becomes the time
1410 * dispersion (jitter) estimate.
1411 */
1412 pps_tf[2] = pps_tf[1];
1413 pps_tf[1] = pps_tf[0];
1414 pps_tf[0] = u_usec;
1415 if (pps_tf[0] > pps_tf[1]) {
1416 if (pps_tf[1] > pps_tf[2]) {
1417 pps_offset = pps_tf[1]; /* 0 1 2 */
1418 v_usec = pps_tf[0] - pps_tf[2];
1419 } else if (pps_tf[2] > pps_tf[0]) {
1420 pps_offset = pps_tf[0]; /* 2 0 1 */
1421 v_usec = pps_tf[2] - pps_tf[1];
1422 } else {
1423 pps_offset = pps_tf[2]; /* 0 2 1 */
1424 v_usec = pps_tf[0] - pps_tf[1];
1425 }
1426 } else {
1427 if (pps_tf[1] < pps_tf[2]) {
1428 pps_offset = pps_tf[1]; /* 2 1 0 */
1429 v_usec = pps_tf[2] - pps_tf[0];
1430 } else if (pps_tf[2] < pps_tf[0]) {
1431 pps_offset = pps_tf[0]; /* 1 0 2 */
1432 v_usec = pps_tf[1] - pps_tf[2];
1433 } else {
1434 pps_offset = pps_tf[2]; /* 1 2 0 */
1435 v_usec = pps_tf[1] - pps_tf[0];
1436 }
1437 }
1438 if (v_usec > MAXTIME)
1439 pps_jitcnt++;
1440 v_usec = (v_usec << PPS_AVG) - pps_jitter;
1441 if (v_usec < 0)
1442 pps_jitter -= -v_usec >> PPS_AVG;
1443 else
1444 pps_jitter += v_usec >> PPS_AVG;
1445 if (pps_jitter > (MAXTIME >> 1))
1446 time_status |= STA_PPSJITTER;
1447
1448 /*
1449 * During the calibration interval adjust the starting time when
1450 * the tick overflows. At the end of the interval compute the
1451 * duration of the interval and the difference of the hardware
1452 * counters at the beginning and end of the interval. This code
1453 * is deliciously complicated by the fact valid differences may
1454 * exceed the value of tick when using long calibration
1455 * intervals and small ticks. Note that the counter can be
1456 * greater than tick if caught at just the wrong instant, but
1457 * the values returned and used here are correct.
1458 */
1459 bigtick = (long)tick << SHIFT_USEC;
1460 pps_usec -= pps_freq;
1461 if (pps_usec >= bigtick)
1462 pps_usec -= bigtick;
1463 if (pps_usec < 0)
1464 pps_usec += bigtick;
1465 pps_time.tv_sec++;
1466 pps_count++;
1467 if (pps_count < (1 << pps_shift))
1468 return;
1469 pps_count = 0;
1470 pps_calcnt++;
1471 u_usec = usec << SHIFT_USEC;
1472 v_usec = pps_usec - u_usec;
1473 if (v_usec >= bigtick >> 1)
1474 v_usec -= bigtick;
1475 if (v_usec < -(bigtick >> 1))
1476 v_usec += bigtick;
1477 if (v_usec < 0)
1478 v_usec = -(-v_usec >> pps_shift);
1479 else
1480 v_usec = v_usec >> pps_shift;
1481 pps_usec = u_usec;
1482 cal_sec = tvp->tv_sec;
1483 cal_usec = tvp->tv_usec;
1484 cal_sec -= pps_time.tv_sec;
1485 cal_usec -= pps_time.tv_usec;
1486 if (cal_usec < 0) {
1487 cal_usec += 1000000;
1488 cal_sec--;
1489 }
1490 pps_time = *tvp;
1491
1492 /*
1493 * Check for lost interrupts, noise, excessive jitter and
1494 * excessive frequency error. The number of timer ticks during
1495 * the interval may vary +-1 tick. Add to this a margin of one
1496 * tick for the PPS signal jitter and maximum frequency
1497 * deviation. If the limits are exceeded, the calibration
1498 * interval is reset to the minimum and we start over.
1499 */
1500 u_usec = (long)tick << 1;
1501 if (!((cal_sec == -1 && cal_usec > (1000000 - u_usec))
1502 || (cal_sec == 0 && cal_usec < u_usec))
1503 || v_usec > time_tolerance || v_usec < -time_tolerance) {
1504 pps_errcnt++;
1505 pps_shift = PPS_SHIFT;
1506 pps_intcnt = 0;
1507 time_status |= STA_PPSERROR;
1508 return;
1509 }
1510
1511 /*
1512 * A three-stage median filter is used to help deglitch the pps
1513 * frequency. The median sample becomes the frequency offset
1514 * estimate; the difference between the other two samples
1515 * becomes the frequency dispersion (stability) estimate.
1516 */
1517 pps_ff[2] = pps_ff[1];
1518 pps_ff[1] = pps_ff[0];
1519 pps_ff[0] = v_usec;
1520 if (pps_ff[0] > pps_ff[1]) {
1521 if (pps_ff[1] > pps_ff[2]) {
1522 u_usec = pps_ff[1]; /* 0 1 2 */
1523 v_usec = pps_ff[0] - pps_ff[2];
1524 } else if (pps_ff[2] > pps_ff[0]) {
1525 u_usec = pps_ff[0]; /* 2 0 1 */
1526 v_usec = pps_ff[2] - pps_ff[1];
1527 } else {
1528 u_usec = pps_ff[2]; /* 0 2 1 */
1529 v_usec = pps_ff[0] - pps_ff[1];
1530 }
1531 } else {
1532 if (pps_ff[1] < pps_ff[2]) {
1533 u_usec = pps_ff[1]; /* 2 1 0 */
1534 v_usec = pps_ff[2] - pps_ff[0];
1535 } else if (pps_ff[2] < pps_ff[0]) {
1536 u_usec = pps_ff[0]; /* 1 0 2 */
1537 v_usec = pps_ff[1] - pps_ff[2];
1538 } else {
1539 u_usec = pps_ff[2]; /* 1 2 0 */
1540 v_usec = pps_ff[1] - pps_ff[0];
1541 }
1542 }
1543
1544 /*
1545 * Here the frequency dispersion (stability) is updated. If it
1546 * is less than one-fourth the maximum (MAXFREQ), the frequency
1547 * offset is updated as well, but clamped to the tolerance. It
1548 * will be processed later by the hardclock() routine.
1549 */
1550 v_usec = (v_usec >> 1) - pps_stabil;
1551 if (v_usec < 0)
1552 pps_stabil -= -v_usec >> PPS_AVG;
1553 else
1554 pps_stabil += v_usec >> PPS_AVG;
1555 if (pps_stabil > MAXFREQ >> 2) {
1556 pps_stbcnt++;
1557 time_status |= STA_PPSWANDER;
1558 return;
1559 }
1560 if (time_status & STA_PPSFREQ) {
1561 if (u_usec < 0) {
1562 pps_freq -= -u_usec >> PPS_AVG;
1563 if (pps_freq < -time_tolerance)
1564 pps_freq = -time_tolerance;
1565 u_usec = -u_usec;
1566 } else {
1567 pps_freq += u_usec >> PPS_AVG;
1568 if (pps_freq > time_tolerance)
1569 pps_freq = time_tolerance;
1570 }
1571 }
1572
1573 /*
1574 * Here the calibration interval is adjusted. If the maximum
1575 * time difference is greater than tick / 4, reduce the interval
1576 * by half. If this is not the case for four consecutive
1577 * intervals, double the interval.
1578 */
1579 if (u_usec << pps_shift > bigtick >> 2) {
1580 pps_intcnt = 0;
1581 if (pps_shift > PPS_SHIFT)
1582 pps_shift--;
1583 } else if (pps_intcnt >= 4) {
1584 pps_intcnt = 0;
1585 if (pps_shift < PPS_SHIFTMAX)
1586 pps_shift++;
1587 } else
1588 pps_intcnt++;
1589 }
1590 #endif /* PPS_SYNC */
1591 #endif /* NTP */
1592
1593 /* timecounter compat functions */
1594 void
1595 nanotime(struct timespec *ts)
1596 {
1597 struct timeval tv;
1598
1599 microtime(&tv);
1600 TIMEVAL_TO_TIMESPEC(&tv, ts);
1601 }
1602
1603 void
1604 getbinuptime(struct bintime *bt)
1605 {
1606 struct timeval tv;
1607
1608 microtime(&tv);
1609 timeval2bintime(&tv, bt);
1610 }
1611
1612 void
1613 nanouptime(struct timespec *tsp)
1614 {
1615 int s;
1616
1617 s = splclock();
1618 TIMEVAL_TO_TIMESPEC(&mono_time, tsp);
1619 splx(s);
1620 }
1621
1622 void
1623 getnanouptime(struct timespec *tsp)
1624 {
1625 int s;
1626
1627 s = splclock();
1628 TIMEVAL_TO_TIMESPEC(&mono_time, tsp);
1629 splx(s);
1630 }
1631
1632 void
1633 getmicrouptime(struct timeval *tvp)
1634 {
1635 int s;
1636
1637 s = splclock();
1638 *tvp = mono_time;
1639 splx(s);
1640 }
1641
1642 void
1643 getnanotime(struct timespec *tsp)
1644 {
1645 int s;
1646
1647 s = splclock();
1648 TIMEVAL_TO_TIMESPEC(&time, tsp);
1649 splx(s);
1650 }
1651
1652 void
1653 getmicrotime(struct timeval *tvp)
1654 {
1655 int s;
1656
1657 s = splclock();
1658 *tvp = time;
1659 splx(s);
1660 }
1661 #endif /* !__HAVE_TIMECOUNTER */
1662