1/*-
2 * Copyright (c) 1982, 1986, 1991, 1993
3 * The Regents of the University of California. All rights reserved.
4 * (c) UNIX System Laboratories, Inc.
5 * All or some portions of this file are derived from material licensed
6 * to the University of California by American Telephone and Telegraph
7 * Co. or Unix System Laboratories, Inc. and are reproduced herein with
8 * the permission of UNIX System Laboratories, Inc.
9 *
10 * Redistribution and use in source and binary forms, with or without
11 * modification, are permitted provided that the following conditions
12 * are met:
13 * 1. Redistributions of source code must retain the above copyright
14 * notice, this list of conditions and the following disclaimer.
15 * 2. Redistributions in binary form must reproduce the above copyright
16 * notice, this list of conditions and the following disclaimer in the
17 * documentation and/or other materials provided with the distribution.
18 * 3. All advertising materials mentioning features or use of this software
19 * must display the following acknowledgement:
20 * This product includes software developed by the University of
21 * California, Berkeley and its contributors.
22 * 4. Neither the name of the University nor the names of its contributors
23 * may be used to endorse or promote products derived from this software
24 * without specific prior written permission.
25 *
26 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
27 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
28 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
29 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
30 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
31 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
32 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
33 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
34 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
35 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
36 * SUCH DAMAGE.
37 *
38 * @(#)kern_clock.c 8.5 (Berkeley) 1/21/94
39 * $Id: kern_clock.c,v 1.11 1994/12/12 11:58:46 bde Exp $
40 */
41
42/* Portions of this software are covered by the following: */
43/******************************************************************************
44 * *
45 * Copyright (c) David L. Mills 1993, 1994 *
46 * *
47 * Permission to use, copy, modify, and distribute this software and its *
48 * documentation for any purpose and without fee is hereby granted, provided *
49 * that the above copyright notice appears in all copies and that both the *
50 * copyright notice and this permission notice appear in supporting *
51 * documentation, and that the name University of Delaware not be used in *
52 * advertising or publicity pertaining to distribution of the software *
53 * without specific, written prior permission. The University of Delaware *
54 * makes no representations about the suitability this software for any *
55 * purpose. It is provided "as is" without express or implied warranty. *
56 * *
57 *****************************************************************************/
58
59#include <sys/param.h>
60#include <sys/systm.h>
61#include <sys/dkstat.h>
62#include <sys/callout.h>
63#include <sys/kernel.h>
64#include <sys/proc.h>
65#include <sys/resourcevar.h>
66#include <sys/signalvar.h>
67#include <sys/timex.h>
68#include <vm/vm.h>
69#include <sys/sysctl.h>
70
71#include <machine/cpu.h>
72#include <machine/clock.h>
73
74#ifdef GPROF
75#include <sys/gmon.h>
76#endif
77
78/* Does anybody else really care about these? */
79struct callout *callfree, *callout, calltodo;
80int ncallout;
81
82/* Some of these don't belong here, but it's easiest to concentrate them. */
83long cp_time[CPUSTATES];
84long dk_seek[DK_NDRIVE];
85long dk_time[DK_NDRIVE];
86long dk_wds[DK_NDRIVE];
87long dk_wpms[DK_NDRIVE];
88long dk_xfer[DK_NDRIVE];
89
90int dk_busy;
91int dk_ndrive = 0;
92char dk_names[DK_NDRIVE][DK_NAMELEN];
93
94long tk_cancc;
95long tk_nin;
96long tk_nout;
97long tk_rawcc;
98
99/*
100 * Clock handling routines.
101 *
102 * This code is written to operate with two timers that run independently of
103 * each other. The main clock, running hz times per second, is used to keep
104 * track of real time. The second timer handles kernel and user profiling,
105 * and does resource use estimation. If the second timer is programmable,
106 * it is randomized to avoid aliasing between the two clocks. For example,
107 * the randomization prevents an adversary from always giving up the cpu
108 * just before its quantum expires. Otherwise, it would never accumulate
109 * cpu ticks. The mean frequency of the second timer is stathz.
110 *
111 * If no second timer exists, stathz will be zero; in this case we drive
112 * profiling and statistics off the main clock. This WILL NOT be accurate;
113 * do not do it unless absolutely necessary.
114 *
115 * The statistics clock may (or may not) be run at a higher rate while
116 * profiling. This profile clock runs at profhz. We require that profhz
117 * be an integral multiple of stathz.
118 *
119 * If the statistics clock is running fast, it must be divided by the ratio
120 * profhz/stathz for statistics. (For profiling, every tick counts.)
121 */
122
123/*
124 * TODO:
125 * allocate more timeout table slots when table overflows.
126 */
127
128/*
129 * Bump a timeval by a small number of usec's.
130 */
131#define BUMPTIME(t, usec) { \
132 register volatile struct timeval *tp = (t); \
133 register long us; \
134 \
135 tp->tv_usec = us = tp->tv_usec + (usec); \
136 if (us >= 1000000) { \
137 tp->tv_usec = us - 1000000; \
138 tp->tv_sec++; \
139 } \
140}
141
142int stathz;
143int profhz;
144int profprocs;
145int ticks;
146static int psdiv, pscnt; /* prof => stat divider */
147int psratio; /* ratio: prof / stat */
148
149volatile struct timeval time;
150volatile struct timeval mono_time;
151
152/*
153 * Phase-lock loop (PLL) definitions
154 *
155 * The following variables are read and set by the ntp_adjtime() system
156 * call.
157 *
158 * time_state shows the state of the system clock, with values defined
159 * in the timex.h header file.
160 *
161 * time_status shows the status of the system clock, with bits defined
162 * in the timex.h header file.
163 *
164 * time_offset is used by the PLL to adjust the system time in small
165 * increments.
166 *
167 * time_constant determines the bandwidth or "stiffness" of the PLL.
168 *
169 * time_tolerance determines maximum frequency error or tolerance of the
170 * CPU clock oscillator and is a property of the architecture; however,
171 * in principle it could change as result of the presence of external
172 * discipline signals, for instance.
173 *
174 * time_precision is usually equal to the kernel tick variable; however,
175 * in cases where a precision clock counter or external clock is
176 * available, the resolution can be much less than this and depend on
177 * whether the external clock is working or not.
178 *
179 * time_maxerror is initialized by a ntp_adjtime() call and increased by
180 * the kernel once each second to reflect the maximum error
181 * bound growth.
182 *
183 * time_esterror is set and read by the ntp_adjtime() call, but
184 * otherwise not used by the kernel.
185 */
186int time_status = STA_UNSYNC; /* clock status bits */
187int time_state = TIME_OK; /* clock state */
188long time_offset = 0; /* time offset (us) */
189long time_constant = 0; /* pll time constant */
190long time_tolerance = MAXFREQ; /* frequency tolerance (scaled ppm) */
191long time_precision = 1; /* clock precision (us) */
192long time_maxerror = MAXPHASE; /* maximum error (us) */
193long time_esterror = MAXPHASE; /* estimated error (us) */
194
195/*
196 * The following variables establish the state of the PLL and the
197 * residual time and frequency offset of the local clock. The scale
198 * factors are defined in the timex.h header file.
199 *
200 * time_phase and time_freq are the phase increment and the frequency
201 * increment, respectively, of the kernel time variable at each tick of
202 * the clock.
203 *
204 * time_freq is set via ntp_adjtime() from a value stored in a file when
205 * the synchronization daemon is first started. Its value is retrieved
206 * via ntp_adjtime() and written to the file about once per hour by the
207 * daemon.
208 *
209 * time_adj is the adjustment added to the value of tick at each timer
210 * interrupt and is recomputed at each timer interrupt.
211 *
212 * time_reftime is the second's portion of the system time on the last
213 * call to ntp_adjtime(). It is used to adjust the time_freq variable
214 * and to increase the time_maxerror as the time since last update
215 * increases.
216 */
217long time_phase = 0; /* phase offset (scaled us) */
218long time_freq = 0; /* frequency offset (scaled ppm) */
219long time_adj = 0; /* tick adjust (scaled 1 / hz) */
220long time_reftime = 0; /* time at last adjustment (s) */
221
222#ifdef PPS_SYNC
223/*
224 * The following variables are used only if the if the kernel PPS
225 * discipline code is configured (PPS_SYNC). The scale factors are
226 * defined in the timex.h header file.
227 *
228 * pps_time contains the time at each calibration interval, as read by
229 * microtime().
230 *
231 * pps_offset is the time offset produced by the time median filter
232 * pps_tf[], while pps_jitter is the dispersion measured by this
233 * filter.
234 *
235 * pps_freq is the frequency offset produced by the frequency median
236 * filter pps_ff[], while pps_stabil is the dispersion measured by
237 * this filter.
238 *
239 * pps_usec is latched from a high resolution counter or external clock
240 * at pps_time. Here we want the hardware counter contents only, not the
241 * contents plus the time_tv.usec as usual.
242 *
243 * pps_valid counts the number of seconds since the last PPS update. It
244 * is used as a watchdog timer to disable the PPS discipline should the
245 * PPS signal be lost.
246 *
247 * pps_glitch counts the number of seconds since the beginning of an
248 * offset burst more than tick/2 from current nominal offset. It is used
249 * mainly to suppress error bursts due to priority conflicts between the
250 * PPS interrupt and timer interrupt.
251 *
252 * pps_count counts the seconds of the calibration interval, the
253 * duration of which is pps_shift in powers of two.
254 *
255 * pps_intcnt counts the calibration intervals for use in the interval-
256 * adaptation algorithm. It's just too complicated for words.
257 */
258struct timeval pps_time; /* kernel time at last interval */
259long pps_offset = 0; /* pps time offset (us) */
260long pps_jitter = MAXTIME; /* pps time dispersion (jitter) (us) */
261long pps_tf[] = {0, 0, 0}; /* pps time offset median filter (us) */
262long pps_freq = 0; /* frequency offset (scaled ppm) */
263long pps_stabil = MAXFREQ; /* frequency dispersion (scaled ppm) */
264long pps_ff[] = {0, 0, 0}; /* frequency offset median filter */
265long pps_usec = 0; /* microsec counter at last interval */
266long pps_valid = PPS_VALID; /* pps signal watchdog counter */
267int pps_glitch = 0; /* pps signal glitch counter */
268int pps_count = 0; /* calibration interval counter (s) */
269int pps_shift = PPS_SHIFT; /* interval duration (s) (shift) */
270int pps_intcnt = 0; /* intervals at current duration */
271
272/*
273 * PPS signal quality monitors
274 *
275 * pps_jitcnt counts the seconds that have been discarded because the
276 * jitter measured by the time median filter exceeds the limit MAXTIME
277 * (100 us).
278 *
279 * pps_calcnt counts the frequency calibration intervals, which are
280 * variable from 4 s to 256 s.
281 *
282 * pps_errcnt counts the calibration intervals which have been discarded
283 * because the wander exceeds the limit MAXFREQ (100 ppm) or where the
284 * calibration interval jitter exceeds two ticks.
285 *
286 * pps_stbcnt counts the calibration intervals that have been discarded
287 * because the frequency wander exceeds the limit MAXFREQ / 4 (25 us).
288 */
289long pps_jitcnt = 0; /* jitter limit exceeded */
290long pps_calcnt = 0; /* calibration intervals */
291long pps_errcnt = 0; /* calibration errors */
292long pps_stbcnt = 0; /* stability limit exceeded */
293#endif /* PPS_SYNC */
294
295/* XXX none of this stuff works under FreeBSD */
296#ifdef EXT_CLOCK
297/*
298 * External clock definitions
299 *
300 * The following definitions and declarations are used only if an
301 * external clock (HIGHBALL or TPRO) is configured on the system.
302 */
303#define CLOCK_INTERVAL 30 /* CPU clock update interval (s) */
304
305/*
306 * The clock_count variable is set to CLOCK_INTERVAL at each PPS
307 * interrupt and decremented once each second.
308 */
309int clock_count = 0; /* CPU clock counter */
310
311#ifdef HIGHBALL
312/*
313 * The clock_offset and clock_cpu variables are used by the HIGHBALL
314 * interface. The clock_offset variable defines the offset between
315 * system time and the HIGBALL counters. The clock_cpu variable contains
316 * the offset between the system clock and the HIGHBALL clock for use in
317 * disciplining the kernel time variable.
318 */
319extern struct timeval clock_offset; /* Highball clock offset */
320long clock_cpu = 0; /* CPU clock adjust */
321#endif /* HIGHBALL */
322#endif /* EXT_CLOCK */
323
324/*
325 * hardupdate() - local clock update
326 *
327 * This routine is called by ntp_adjtime() to update the local clock
328 * phase and frequency. This is used to implement an adaptive-parameter,
329 * first-order, type-II phase-lock loop. The code computes new time and
330 * frequency offsets each time it is called. The hardclock() routine
331 * amortizes these offsets at each tick interrupt. If the kernel PPS
332 * discipline code is configured (PPS_SYNC), the PPS signal itself
333 * determines the new time offset, instead of the calling argument.
334 * Presumably, calls to ntp_adjtime() occur only when the caller
335 * believes the local clock is valid within some bound (+-128 ms with
336 * NTP). If the caller's time is far different than the PPS time, an
337 * argument will ensue, and it's not clear who will lose.
338 *
339 * For default SHIFT_UPDATE = 12, the offset is limited to +-512 ms, the
340 * maximum interval between updates is 4096 s and the maximum frequency
341 * offset is +-31.25 ms/s.
342 *
343 * Note: splclock() is in effect.
344 */
345void
346hardupdate(offset)
347 long offset;
348{
349 long ltemp, mtemp;
350
351 if (!(time_status & STA_PLL) && !(time_status & STA_PPSTIME))
352 return;
353 ltemp = offset;
354#ifdef PPS_SYNC
355 if (time_status & STA_PPSTIME && time_status & STA_PPSSIGNAL)
356 ltemp = pps_offset;
357#endif /* PPS_SYNC */
358 if (ltemp > MAXPHASE)
359 time_offset = MAXPHASE << SHIFT_UPDATE;
360 else if (ltemp < -MAXPHASE)
361 time_offset = -(MAXPHASE << SHIFT_UPDATE);
362 else
363 time_offset = ltemp << SHIFT_UPDATE;
364 mtemp = time.tv_sec - time_reftime;
365 time_reftime = time.tv_sec;
366 if (mtemp > MAXSEC)
367 mtemp = 0;
368
369 /* ugly multiply should be replaced */
370 if (ltemp < 0)
371 time_freq -= (-ltemp * mtemp) >> (time_constant +
372 time_constant + SHIFT_KF - SHIFT_USEC);
373 else
374 time_freq += (ltemp * mtemp) >> (time_constant +
375 time_constant + SHIFT_KF - SHIFT_USEC);
376 if (time_freq > time_tolerance)
377 time_freq = time_tolerance;
378 else if (time_freq < -time_tolerance)
379 time_freq = -time_tolerance;
380}
381
382
383
384/*
385 * Initialize clock frequencies and start both clocks running.
386 */
387void
388initclocks()
389{
390 register int i;
391
392 /*
393 * Set divisors to 1 (normal case) and let the machine-specific
394 * code do its bit.
395 */
396 psdiv = pscnt = 1;
397 cpu_initclocks();
398
399 /*
400 * Compute profhz/stathz, and fix profhz if needed.
401 */
402 i = stathz ? stathz : hz;
403 if (profhz == 0)
404 profhz = i;
405 psratio = profhz / i;
406}
407
408/*
409 * The real-time timer, interrupting hz times per second.
410 */
411void
412hardclock(frame)
413 register struct clockframe *frame;
414{
415 register struct callout *p1;
416 register struct proc *p;
417 register int needsoft;
418
419 /*
420 * Update real-time timeout queue.
421 * At front of queue are some number of events which are ``due''.
422 * The time to these is <= 0 and if negative represents the
423 * number of ticks which have passed since it was supposed to happen.
424 * The rest of the q elements (times > 0) are events yet to happen,
425 * where the time for each is given as a delta from the previous.
426 * Decrementing just the first of these serves to decrement the time
427 * to all events.
428 */
429 needsoft = 0;
430 for (p1 = calltodo.c_next; p1 != NULL; p1 = p1->c_next) {
431 if (--p1->c_time > 0)
432 break;
433 needsoft = 1;
434 if (p1->c_time == 0)
435 break;
436 }
437
438 p = curproc;
439 if (p) {
440 register struct pstats *pstats;
441
442 /*
443 * Run current process's virtual and profile time, as needed.
444 */
445 pstats = p->p_stats;
446 if (CLKF_USERMODE(frame) &&
447 timerisset(&pstats->p_timer[ITIMER_VIRTUAL].it_value) &&
448 itimerdecr(&pstats->p_timer[ITIMER_VIRTUAL], tick) == 0)
449 psignal(p, SIGVTALRM);
450 if (timerisset(&pstats->p_timer[ITIMER_PROF].it_value) &&
451 itimerdecr(&pstats->p_timer[ITIMER_PROF], tick) == 0)
452 psignal(p, SIGPROF);
453 }
454
455 /*
456 * If no separate statistics clock is available, run it from here.
457 */
458 if (stathz == 0)
459 statclock(frame);
460
461 /*
462 * Increment the time-of-day.
463 */
464 ticks++;
465 {
466 int time_update;
467 struct timeval newtime = time;
468 long ltemp;
469
470 if (timedelta == 0) {
471 time_update = tick;
472 } else {
473 time_update = tick + tickdelta;
474 timedelta -= tickdelta;
475 }
476 BUMPTIME(&mono_time, time_update);
477
478 /*
479 * Compute the phase adjustment. If the low-order bits
480 * (time_phase) of the update overflow, bump the high-order bits
481 * (time_update).
482 */
483 time_phase += time_adj;
484 if (time_phase <= -FINEUSEC) {
485 ltemp = -time_phase >> SHIFT_SCALE;
486 time_phase += ltemp << SHIFT_SCALE;
487 time_update -= ltemp;
488 }
489 else if (time_phase >= FINEUSEC) {
490 ltemp = time_phase >> SHIFT_SCALE;
491 time_phase -= ltemp << SHIFT_SCALE;
492 time_update += ltemp;
493 }
494
495 newtime.tv_usec += time_update;
496 /*
497 * On rollover of the second the phase adjustment to be used for
498 * the next second is calculated. Also, the maximum error is
499 * increased by the tolerance. If the PPS frequency discipline
500 * code is present, the phase is increased to compensate for the
501 * CPU clock oscillator frequency error.
502 *
503 * With SHIFT_SCALE = 23, the maximum frequency adjustment is
504 * +-256 us per tick, or 25.6 ms/s at a clock frequency of 100
505 * Hz. The time contribution is shifted right a minimum of two
506 * bits, while the frequency contribution is a right shift.
507 * Thus, overflow is prevented if the frequency contribution is
508 * limited to half the maximum or 15.625 ms/s.
509 */
510 if (newtime.tv_usec >= 1000000) {
511 newtime.tv_usec -= 1000000;
512 newtime.tv_sec++;
513 time_maxerror += time_tolerance >> SHIFT_USEC;
514 if (time_offset < 0) {
515 ltemp = -time_offset >>
516 (SHIFT_KG + time_constant);
517 time_offset += ltemp;
518 time_adj = -ltemp <<
519 (SHIFT_SCALE - SHIFT_HZ - SHIFT_UPDATE);
520 } else {
521 ltemp = time_offset >>
522 (SHIFT_KG + time_constant);
523 time_offset -= ltemp;
524 time_adj = ltemp <<
525 (SHIFT_SCALE - SHIFT_HZ - SHIFT_UPDATE);
526 }
527#ifdef PPS_SYNC
528 /*
529 * Gnaw on the watchdog counter and update the frequency
530 * computed by the pll and the PPS signal.
531 */
532 pps_valid++;
533 if (pps_valid == PPS_VALID) {
534 pps_jitter = MAXTIME;
535 pps_stabil = MAXFREQ;
536 time_status &= ~(STA_PPSSIGNAL | STA_PPSJITTER |
537 STA_PPSWANDER | STA_PPSERROR);
538 }
539 ltemp = time_freq + pps_freq;
540#else
541 ltemp = time_freq;
542#endif /* PPS_SYNC */
543 if (ltemp < 0)
544 time_adj -= -ltemp >>
545 (SHIFT_USEC + SHIFT_HZ - SHIFT_SCALE);
546 else
547 time_adj += ltemp >>
548 (SHIFT_USEC + SHIFT_HZ - SHIFT_SCALE);
549
550 /*
551 * When the CPU clock oscillator frequency is not a
552 * power of two in Hz, the SHIFT_HZ is only an
553 * approximate scale factor. In the SunOS kernel, this
554 * results in a PLL gain factor of 1/1.28 = 0.78 what it
555 * should be. In the following code the overall gain is
556 * increased by a factor of 1.25, which results in a
557 * residual error less than 3 percent.
558 */
559 /* Same thing applies for FreeBSD --GAW */
560 if (hz == 100) {
561 if (time_adj < 0)
562 time_adj -= -time_adj >> 2;
563 else
564 time_adj += time_adj >> 2;
565 }
566
567 /* XXX - this is really bogus, but can't be fixed until
568 xntpd's idea of the system clock is fixed to know how
569 the user wants leap seconds handled; in the mean time,
570 we assume that users of NTP are running without proper
571 leap second support (this is now the default anyway) */
572 /*
573 * Leap second processing. If in leap-insert state at
574 * the end of the day, the system clock is set back one
575 * second; if in leap-delete state, the system clock is
576 * set ahead one second. The microtime() routine or
577 * external clock driver will insure that reported time
578 * is always monotonic. The ugly divides should be
579 * replaced.
580 */
581 switch (time_state) {
582
583 case TIME_OK:
584 if (time_status & STA_INS)
585 time_state = TIME_INS;
586 else if (time_status & STA_DEL)
587 time_state = TIME_DEL;
588 break;
589
590 case TIME_INS:
591 if (newtime.tv_sec % 86400 == 0) {
592 newtime.tv_sec--;
593 time_state = TIME_OOP;
594 }
595 break;
596
597 case TIME_DEL:
598 if ((newtime.tv_sec + 1) % 86400 == 0) {
599 newtime.tv_sec++;
600 time_state = TIME_WAIT;
601 }
602 break;
603
604 case TIME_OOP:
605 time_state = TIME_WAIT;
606 break;
607
608 case TIME_WAIT:
609 if (!(time_status & (STA_INS | STA_DEL)))
610 time_state = TIME_OK;
611 }
612 }
613 CPU_CLOCKUPDATE(&time, &newtime);
614 }
615
616 /*
617 * Process callouts at a very low cpu priority, so we don't keep the
618 * relatively high clock interrupt priority any longer than necessary.
619 */
620 if (needsoft) {
621 if (CLKF_BASEPRI(frame)) {
622 /*
623 * Save the overhead of a software interrupt;
624 * it will happen as soon as we return, so do it now.
625 */
626 (void)splsoftclock();
627 softclock();
628 } else
629 setsoftclock();
630 }
631}
632
633/*
634 * Software (low priority) clock interrupt.
635 * Run periodic events from timeout queue.
636 */
637/*ARGSUSED*/
638void
639softclock()
640{
641 register struct callout *c;
642 register void *arg;
643 register void (*func) __P((void *));
644 register int s;
645
646 s = splhigh();
647 while ((c = calltodo.c_next) != NULL && c->c_time <= 0) {
648 func = c->c_func;
649 arg = c->c_arg;
650 calltodo.c_next = c->c_next;
651 c->c_next = callfree;
652 callfree = c;
653 splx(s);
654 (*func)(arg);
655 (void) splhigh();
656 }
657 splx(s);
658}
659
660/*
661 * timeout --
662 * Execute a function after a specified length of time.
663 *
664 * untimeout --
665 * Cancel previous timeout function call.
666 *
667 * See AT&T BCI Driver Reference Manual for specification. This
668 * implementation differs from that one in that no identification
669 * value is returned from timeout, rather, the original arguments
670 * to timeout are used to identify entries for untimeout.
671 */
672void
673timeout(ftn, arg, ticks)
674 timeout_t ftn;
675 void *arg;
676 register int ticks;
677{
678 register struct callout *new, *p, *t;
679 register int s;
680
681 if (ticks <= 0)
682 ticks = 1;
683
684 /* Lock out the clock. */
685 s = splhigh();
686
687 /* Fill in the next free callout structure. */
688 if (callfree == NULL)
689 panic("timeout table full");
690 new = callfree;
691 callfree = new->c_next;
692 new->c_arg = arg;
693 new->c_func = ftn;
694
695 /*
696 * The time for each event is stored as a difference from the time
697 * of the previous event on the queue. Walk the queue, correcting
698 * the ticks argument for queue entries passed. Correct the ticks
699 * value for the queue entry immediately after the insertion point
700 * as well. Watch out for negative c_time values; these represent
701 * overdue events.
702 */
703 for (p = &calltodo;
704 (t = p->c_next) != NULL && ticks > t->c_time; p = t)
705 if (t->c_time > 0)
706 ticks -= t->c_time;
707 new->c_time = ticks;
708 if (t != NULL)
709 t->c_time -= ticks;
710
711 /* Insert the new entry into the queue. */
712 p->c_next = new;
713 new->c_next = t;
714 splx(s);
715}
716
717void
718untimeout(ftn, arg)
719 timeout_t ftn;
720 void *arg;
721{
722 register struct callout *p, *t;
723 register int s;
724
725 s = splhigh();
726 for (p = &calltodo; (t = p->c_next) != NULL; p = t)
727 if (t->c_func == ftn && t->c_arg == arg) {
728 /* Increment next entry's tick count. */
729 if (t->c_next && t->c_time > 0)
730 t->c_next->c_time += t->c_time;
731
732 /* Move entry from callout queue to callfree queue. */
733 p->c_next = t->c_next;
734 t->c_next = callfree;
735 callfree = t;
736 break;
737 }
738 splx(s);
739}
740
741/*
742 * Compute number of hz until specified time. Used to
743 * compute third argument to timeout() from an absolute time.
744 */
745int
746hzto(tv)
747 struct timeval *tv;
748{
749 register unsigned long ticks;
750 register long sec, usec;
751 int s;
752
753 /*
754 * If the number of usecs in the whole seconds part of the time
755 * difference fits in a long, then the total number of usecs will
756 * fit in an unsigned long. Compute the total and convert it to
757 * ticks, rounding up and adding 1 to allow for the current tick
758 * to expire. Rounding also depends on unsigned long arithmetic
759 * to avoid overflow.
760 *
761 * Otherwise, if the number of ticks in the whole seconds part of
762 * the time difference fits in a long, then convert the parts to
763 * ticks separately and add, using similar rounding methods and
764 * overflow avoidance. This method would work in the previous
765 * case but it is slightly slower and assumes that hz is integral.
766 *
767 * Otherwise, round the time difference down to the maximum
768 * representable value.
769 *
770 * If ints have 32 bits, then the maximum value for any timeout in
771 * 10ms ticks is 248 days.
772 */
773 s = splclock();
774 sec = tv->tv_sec - time.tv_sec;
775 usec = tv->tv_usec - time.tv_usec;
776 splx(s);
777 if (usec < 0) {
778 sec--;
779 usec += 1000000;
780 }
781 if (sec < 0) {
782#ifdef DIAGNOSTIC
783 printf("hzto: negative time difference %ld sec %ld usec\n",
784 sec, usec);
785#endif
786 ticks = 1;
787 } else if (sec <= LONG_MAX / 1000000)
788 ticks = (sec * 1000000 + (unsigned long)usec + (tick - 1))
789 / tick + 1;
790 else if (sec <= LONG_MAX / hz)
791 ticks = sec * hz
792 + ((unsigned long)usec + (tick - 1)) / tick + 1;
793 else
794 ticks = LONG_MAX;
795 if (ticks > INT_MAX)
796 ticks = INT_MAX;
797 return (ticks);
798}
799
800/*
801 * Start profiling on a process.
802 *
803 * Kernel profiling passes proc0 which never exits and hence
804 * keeps the profile clock running constantly.
805 */
806void
807startprofclock(p)
808 register struct proc *p;
809{
810 int s;
811
812 if ((p->p_flag & P_PROFIL) == 0) {
813 p->p_flag |= P_PROFIL;
814 if (++profprocs == 1 && stathz != 0) {
815 s = splstatclock();
816 psdiv = pscnt = psratio;
817 setstatclockrate(profhz);
818 splx(s);
819 }
820 }
821}
822
823/*
824 * Stop profiling on a process.
825 */
826void
827stopprofclock(p)
828 register struct proc *p;
829{
830 int s;
831
832 if (p->p_flag & P_PROFIL) {
833 p->p_flag &= ~P_PROFIL;
834 if (--profprocs == 0 && stathz != 0) {
835 s = splstatclock();
836 psdiv = pscnt = 1;
837 setstatclockrate(stathz);
838 splx(s);
839 }
840 }
841}
842
843/*
844 * Statistics clock. Grab profile sample, and if divider reaches 0,
845 * do process and kernel statistics.
846 */
847void
848statclock(frame)
849 register struct clockframe *frame;
850{
851#ifdef GPROF
852 register struct gmonparam *g;
853#endif
854 register struct proc *p = curproc;
855 register int i;
856
857 if (p) {
858 struct pstats *pstats;
859 struct rusage *ru;
860 struct vmspace *vm;
861
862 /* bump the resource usage of integral space use */
863 if ((pstats = p->p_stats) && (ru = &pstats->p_ru) && (vm = p->p_vmspace)) {
864 ru->ru_ixrss += vm->vm_tsize * PAGE_SIZE / 1024;
865 ru->ru_idrss += vm->vm_dsize * PAGE_SIZE / 1024;
866 ru->ru_isrss += vm->vm_ssize * PAGE_SIZE / 1024;
867 if ((vm->vm_pmap.pm_stats.resident_count * PAGE_SIZE / 1024) >
868 ru->ru_maxrss) {
869 ru->ru_maxrss =
870 vm->vm_pmap.pm_stats.resident_count * PAGE_SIZE / 1024;
871 }
872 }
873 }
874
875 if (CLKF_USERMODE(frame)) {
876 if (p->p_flag & P_PROFIL)
877 addupc_intr(p, CLKF_PC(frame), 1);
878 if (--pscnt > 0)
879 return;
880 /*
881 * Came from user mode; CPU was in user state.
882 * If this process is being profiled record the tick.
883 */
884 p->p_uticks++;
885 if (p->p_nice > NZERO)
886 cp_time[CP_NICE]++;
887 else
888 cp_time[CP_USER]++;
889 } else {
890#ifdef GPROF
891 /*
892 * Kernel statistics are just like addupc_intr, only easier.
893 */
894 g = &_gmonparam;
895 if (g->state == GMON_PROF_ON) {
896 i = CLKF_PC(frame) - g->lowpc;
897 if (i < g->textsize) {
898 i /= HISTFRACTION * sizeof(*g->kcount);
899 g->kcount[i]++;
900 }
901 }
902#endif
903 if (--pscnt > 0)
904 return;
905 /*
906 * Came from kernel mode, so we were:
907 * - handling an interrupt,
908 * - doing syscall or trap work on behalf of the current
909 * user process, or
910 * - spinning in the idle loop.
911 * Whichever it is, charge the time as appropriate.
912 * Note that we charge interrupts to the current process,
913 * regardless of whether they are ``for'' that process,
914 * so that we know how much of its real time was spent
915 * in ``non-process'' (i.e., interrupt) work.
916 */
917 if (CLKF_INTR(frame)) {
918 if (p != NULL)
919 p->p_iticks++;
920 cp_time[CP_INTR]++;
921 } else if (p != NULL) {
922 p->p_sticks++;
923 cp_time[CP_SYS]++;
924 } else
925 cp_time[CP_IDLE]++;
926 }
927 pscnt = psdiv;
928
929 /*
930 * We maintain statistics shown by user-level statistics
931 * programs: the amount of time in each cpu state, and
932 * the amount of time each of DK_NDRIVE ``drives'' is busy.
933 *
934 * XXX should either run linked list of drives, or (better)
935 * grab timestamps in the start & done code.
936 */
937 for (i = 0; i < DK_NDRIVE; i++)
938 if (dk_busy & (1 << i))
939 dk_time[i]++;
940
941 /*
942 * We adjust the priority of the current process. The priority of
943 * a process gets worse as it accumulates CPU time. The cpu usage
944 * estimator (p_estcpu) is increased here. The formula for computing
945 * priorities (in kern_synch.c) will compute a different value each
946 * time p_estcpu increases by 4. The cpu usage estimator ramps up
947 * quite quickly when the process is running (linearly), and decays
948 * away exponentially, at a rate which is proportionally slower when
949 * the system is busy. The basic principal is that the system will
950 * 90% forget that the process used a lot of CPU time in 5 * loadav
951 * seconds. This causes the system to favor processes which haven't
952 * run much recently, and to round-robin among other processes.
953 */
954 if (p != NULL) {
955 p->p_cpticks++;
956 if (++p->p_estcpu == 0)
957 p->p_estcpu--;
958 if ((p->p_estcpu & 3) == 0) {
959 resetpriority(p);
960 if (p->p_priority >= PUSER)
961 p->p_priority = p->p_usrpri;
962 }
963 }
964}
965
966/*
967 * Return information about system clocks.
968 */
969int
970sysctl_clockrate(where, sizep)
971 register char *where;
972 size_t *sizep;
973{
974 struct clockinfo clkinfo;
975
976 /*
977 * Construct clockinfo structure.
978 */
979 clkinfo.hz = hz;
980 clkinfo.tick = tick;
981 clkinfo.profhz = profhz;
982 clkinfo.stathz = stathz ? stathz : hz;
983 return (sysctl_rdstruct(where, sizep, NULL, &clkinfo, sizeof(clkinfo)));
984}
985
986/*#ifdef PPS_SYNC*/
987#if 0
988/* This code is completely bogus; if anybody ever wants to use it, get
989 * the current version from Dave Mills. */
990
991/*
992 * hardpps() - discipline CPU clock oscillator to external pps signal
993 *
994 * This routine is called at each PPS interrupt in order to discipline
995 * the CPU clock oscillator to the PPS signal. It integrates successive
996 * phase differences between the two oscillators and calculates the
997 * frequency offset. This is used in hardclock() to discipline the CPU
998 * clock oscillator so that intrinsic frequency error is cancelled out.
999 * The code requires the caller to capture the time and hardware
1000 * counter value at the designated PPS signal transition.
1001 */
1002void
1003hardpps(tvp, usec)
1004 struct timeval *tvp; /* time at PPS */
1005 long usec; /* hardware counter at PPS */
1006{
1007 long u_usec, v_usec, bigtick;
1008 long cal_sec, cal_usec;
1009
1010 /*
1011 * During the calibration interval adjust the starting time when
1012 * the tick overflows. At the end of the interval compute the
1013 * duration of the interval and the difference of the hardware
1014 * counters at the beginning and end of the interval. This code
1015 * is deliciously complicated by the fact valid differences may
1016 * exceed the value of tick when using long calibration
1017 * intervals and small ticks. Note that the counter can be
1018 * greater than tick if caught at just the wrong instant, but
1019 * the values returned and used here are correct.
1020 */
1021 bigtick = (long)tick << SHIFT_USEC;
1022 pps_usec -= ntp_pll.ybar;
1023 if (pps_usec >= bigtick)
1024 pps_usec -= bigtick;
1025 if (pps_usec < 0)
1026 pps_usec += bigtick;
1027 pps_time.tv_sec++;
1028 pps_count++;
1029 if (pps_count < (1 << pps_shift))
1030 return;
1031 pps_count = 0;
1032 ntp_pll.calcnt++;
1033 u_usec = usec << SHIFT_USEC;
1034 v_usec = pps_usec - u_usec;
1035 if (v_usec >= bigtick >> 1)
1036 v_usec -= bigtick;
1037 if (v_usec < -(bigtick >> 1))
1038 v_usec += bigtick;
1039 if (v_usec < 0)
1040 v_usec = -(-v_usec >> ntp_pll.shift);
1041 else
1042 v_usec = v_usec >> ntp_pll.shift;
1043 pps_usec = u_usec;
1044 cal_sec = tvp->tv_sec;
1045 cal_usec = tvp->tv_usec;
1046 cal_sec -= pps_time.tv_sec;
1047 cal_usec -= pps_time.tv_usec;
1048 if (cal_usec < 0) {
1049 cal_usec += 1000000;
1050 cal_sec--;
1051 }
1052 pps_time = *tvp;
1053
1054 /*
1055 * Check for lost interrupts, noise, excessive jitter and
1056 * excessive frequency error. The number of timer ticks during
1057 * the interval may vary +-1 tick. Add to this a margin of one
1058 * tick for the PPS signal jitter and maximum frequency
1059 * deviation. If the limits are exceeded, the calibration
1060 * interval is reset to the minimum and we start over.
1061 */
1062 u_usec = (long)tick << 1;
1063 if (!((cal_sec == -1 && cal_usec > (1000000 - u_usec))
1064 || (cal_sec == 0 && cal_usec < u_usec))
1065 || v_usec > ntp_pll.tolerance || v_usec < -ntp_pll.tolerance) {
1066 ntp_pll.jitcnt++;
1067 ntp_pll.shift = NTP_PLL.SHIFT;
1068 pps_dispinc = PPS_DISPINC;
1069 ntp_pll.intcnt = 0;
1070 return;
1071 }
1072
1073 /*
1074 * A three-stage median filter is used to help deglitch the pps
1075 * signal. The median sample becomes the offset estimate; the
1076 * difference between the other two samples becomes the
1077 * dispersion estimate.
1078 */
1079 pps_mf[2] = pps_mf[1];
1080 pps_mf[1] = pps_mf[0];
1081 pps_mf[0] = v_usec;
1082 if (pps_mf[0] > pps_mf[1]) {
1083 if (pps_mf[1] > pps_mf[2]) {
1084 u_usec = pps_mf[1]; /* 0 1 2 */
1085 v_usec = pps_mf[0] - pps_mf[2];
1086 } else if (pps_mf[2] > pps_mf[0]) {
1087 u_usec = pps_mf[0]; /* 2 0 1 */
1088 v_usec = pps_mf[2] - pps_mf[1];
1089 } else {
1090 u_usec = pps_mf[2]; /* 0 2 1 */
1091 v_usec = pps_mf[0] - pps_mf[1];
1092 }
1093 } else {
1094 if (pps_mf[1] < pps_mf[2]) {
1095 u_usec = pps_mf[1]; /* 2 1 0 */
1096 v_usec = pps_mf[2] - pps_mf[0];
1097 } else if (pps_mf[2] < pps_mf[0]) {
1098 u_usec = pps_mf[0]; /* 1 0 2 */
1099 v_usec = pps_mf[1] - pps_mf[2];
1100 } else {
1101 u_usec = pps_mf[2]; /* 1 2 0 */
1102 v_usec = pps_mf[1] - pps_mf[0];
1103 }
1104 }
1105
1106 /*
1107 * Here the dispersion average is updated. If it is less than
1108 * the threshold pps_dispmax, the frequency average is updated
1109 * as well, but clamped to the tolerance.
1110 */
1111 v_usec = (v_usec >> 1) - ntp_pll.disp;
1112 if (v_usec < 0)
1113 ntp_pll.disp -= -v_usec >> PPS_AVG;
1114 else
1115 ntp_pll.disp += v_usec >> PPS_AVG;
1116 if (ntp_pll.disp > pps_dispmax) {
1117 ntp_pll.discnt++;
1118 return;
1119 }
1120 if (u_usec < 0) {
1121 ntp_pll.ybar -= -u_usec >> PPS_AVG;
1122 if (ntp_pll.ybar < -ntp_pll.tolerance)
1123 ntp_pll.ybar = -ntp_pll.tolerance;
1124 u_usec = -u_usec;
1125 } else {
1126 ntp_pll.ybar += u_usec >> PPS_AVG;
1127 if (ntp_pll.ybar > ntp_pll.tolerance)
1128 ntp_pll.ybar = ntp_pll.tolerance;
1129 }
1130
1131 /*
1132 * Here the calibration interval is adjusted. If the maximum
1133 * time difference is greater than tick/4, reduce the interval
1134 * by half. If this is not the case for four consecutive
1135 * intervals, double the interval.
1136 */
1137 if (u_usec << ntp_pll.shift > bigtick >> 2) {
1138 ntp_pll.intcnt = 0;
1139 if (ntp_pll.shift > NTP_PLL.SHIFT) {
1140 ntp_pll.shift--;
1141 pps_dispinc <<= 1;
1142 }
1143 } else if (ntp_pll.intcnt >= 4) {
1144 ntp_pll.intcnt = 0;
1145 if (ntp_pll.shift < NTP_PLL.SHIFTMAX) {
1146 ntp_pll.shift++;
1147 pps_dispinc >>= 1;
1148 }
1149 } else
1150 ntp_pll.intcnt++;
1151}
1152#endif /* PPS_SYNC */