FreeBSD/Linux Kernel Cross Reference
sys/kern/kern_tc.c
1 /*-
2 * ----------------------------------------------------------------------------
3 * "THE BEER-WARE LICENSE" (Revision 42):
4 * <phk@FreeBSD.ORG> wrote this file. As long as you retain this notice you
5 * can do whatever you want with this stuff. If we meet some day, and you think
6 * this stuff is worth it, you can buy me a beer in return. Poul-Henning Kamp
7 * ----------------------------------------------------------------------------
8 *
9 * $FreeBSD: releng/5.1/sys/kern/kern_tc.c 112367 2003-03-18 08:45:25Z phk $
10 */
11
12 #include "opt_ntp.h"
13
14 #include <sys/param.h>
15 #include <sys/kernel.h>
16 #include <sys/sysctl.h>
17 #include <sys/systm.h>
18 #include <sys/timepps.h>
19 #include <sys/timetc.h>
20 #include <sys/timex.h>
21
22 /*
23 * Implement a dummy timecounter which we can use until we get a real one
24 * in the air. This allows the console and other early stuff to use
25 * time services.
26 */
27
28 static u_int
29 dummy_get_timecount(struct timecounter *tc)
30 {
31 static u_int now;
32
33 return (++now);
34 }
35
36 static struct timecounter dummy_timecounter = {
37 dummy_get_timecount, 0, ~0u, 1000000, "dummy",
38 };
39
40 struct timehands {
41 /* These fields must be initialized by the driver. */
42 struct timecounter *th_counter;
43 int64_t th_adjustment;
44 u_int64_t th_scale;
45 u_int th_offset_count;
46 struct bintime th_offset;
47 struct timeval th_microtime;
48 struct timespec th_nanotime;
49 /* Fields not to be copied in tc_windup start with th_generation. */
50 volatile u_int th_generation;
51 struct timehands *th_next;
52 };
53
54 extern struct timehands th0;
55 static struct timehands th9 = { NULL, 0, 0, 0, {0, 0}, {0, 0}, {0, 0}, 0, &th0};
56 static struct timehands th8 = { NULL, 0, 0, 0, {0, 0}, {0, 0}, {0, 0}, 0, &th9};
57 static struct timehands th7 = { NULL, 0, 0, 0, {0, 0}, {0, 0}, {0, 0}, 0, &th8};
58 static struct timehands th6 = { NULL, 0, 0, 0, {0, 0}, {0, 0}, {0, 0}, 0, &th7};
59 static struct timehands th5 = { NULL, 0, 0, 0, {0, 0}, {0, 0}, {0, 0}, 0, &th6};
60 static struct timehands th4 = { NULL, 0, 0, 0, {0, 0}, {0, 0}, {0, 0}, 0, &th5};
61 static struct timehands th3 = { NULL, 0, 0, 0, {0, 0}, {0, 0}, {0, 0}, 0, &th4};
62 static struct timehands th2 = { NULL, 0, 0, 0, {0, 0}, {0, 0}, {0, 0}, 0, &th3};
63 static struct timehands th1 = { NULL, 0, 0, 0, {0, 0}, {0, 0}, {0, 0}, 0, &th2};
64 static struct timehands th0 = {
65 &dummy_timecounter,
66 0,
67 (uint64_t)-1 / 1000000,
68 0,
69 {1, 0},
70 {0, 0},
71 {0, 0},
72 1,
73 &th1
74 };
75
76 static struct timehands *volatile timehands = &th0;
77 struct timecounter *timecounter = &dummy_timecounter;
78 static struct timecounter *timecounters = &dummy_timecounter;
79
80 time_t time_second = 1;
81 time_t time_uptime = 0;
82
83 static struct bintime boottimebin;
84 struct timeval boottime;
85 SYSCTL_STRUCT(_kern, KERN_BOOTTIME, boottime, CTLFLAG_RD,
86 &boottime, timeval, "System boottime");
87
88 SYSCTL_NODE(_kern, OID_AUTO, timecounter, CTLFLAG_RW, 0, "");
89
90 #define TC_STATS(foo) \
91 static u_int foo; \
92 SYSCTL_UINT(_kern_timecounter, OID_AUTO, foo, CTLFLAG_RD, &foo, 0, "");\
93 struct __hack
94
95 TC_STATS(nbinuptime); TC_STATS(nnanouptime); TC_STATS(nmicrouptime);
96 TC_STATS(nbintime); TC_STATS(nnanotime); TC_STATS(nmicrotime);
97 TC_STATS(ngetbinuptime); TC_STATS(ngetnanouptime); TC_STATS(ngetmicrouptime);
98 TC_STATS(ngetbintime); TC_STATS(ngetnanotime); TC_STATS(ngetmicrotime);
99 TC_STATS(nsetclock);
100
101 #undef TC_STATS
102
103 static void tc_windup(void);
104
105 /*
106 * Return the difference between the timehands' counter value now and what
107 * was when we copied it to the timehands' offset_count.
108 */
109 static __inline u_int
110 tc_delta(struct timehands *th)
111 {
112 struct timecounter *tc;
113
114 tc = th->th_counter;
115 return ((tc->tc_get_timecount(tc) - th->th_offset_count) &
116 tc->tc_counter_mask);
117 }
118
119 /*
120 * Functions for reading the time. We have to loop until we are sure that
121 * the timehands that we operated on was not updated under our feet. See
122 * the comment in <sys/time.h> for a description of these 12 functions.
123 */
124
125 void
126 binuptime(struct bintime *bt)
127 {
128 struct timehands *th;
129 u_int gen;
130
131 nbinuptime++;
132 do {
133 th = timehands;
134 gen = th->th_generation;
135 *bt = th->th_offset;
136 bintime_addx(bt, th->th_scale * tc_delta(th));
137 } while (gen == 0 || gen != th->th_generation);
138 }
139
140 void
141 nanouptime(struct timespec *tsp)
142 {
143 struct bintime bt;
144
145 nnanouptime++;
146 binuptime(&bt);
147 bintime2timespec(&bt, tsp);
148 }
149
150 void
151 microuptime(struct timeval *tvp)
152 {
153 struct bintime bt;
154
155 nmicrouptime++;
156 binuptime(&bt);
157 bintime2timeval(&bt, tvp);
158 }
159
160 void
161 bintime(struct bintime *bt)
162 {
163
164 nbintime++;
165 binuptime(bt);
166 bintime_add(bt, &boottimebin);
167 }
168
169 void
170 nanotime(struct timespec *tsp)
171 {
172 struct bintime bt;
173
174 nnanotime++;
175 bintime(&bt);
176 bintime2timespec(&bt, tsp);
177 }
178
179 void
180 microtime(struct timeval *tvp)
181 {
182 struct bintime bt;
183
184 nmicrotime++;
185 bintime(&bt);
186 bintime2timeval(&bt, tvp);
187 }
188
189 void
190 getbinuptime(struct bintime *bt)
191 {
192 struct timehands *th;
193 u_int gen;
194
195 ngetbinuptime++;
196 do {
197 th = timehands;
198 gen = th->th_generation;
199 *bt = th->th_offset;
200 } while (gen == 0 || gen != th->th_generation);
201 }
202
203 void
204 getnanouptime(struct timespec *tsp)
205 {
206 struct timehands *th;
207 u_int gen;
208
209 ngetnanouptime++;
210 do {
211 th = timehands;
212 gen = th->th_generation;
213 bintime2timespec(&th->th_offset, tsp);
214 } while (gen == 0 || gen != th->th_generation);
215 }
216
217 void
218 getmicrouptime(struct timeval *tvp)
219 {
220 struct timehands *th;
221 u_int gen;
222
223 ngetmicrouptime++;
224 do {
225 th = timehands;
226 gen = th->th_generation;
227 bintime2timeval(&th->th_offset, tvp);
228 } while (gen == 0 || gen != th->th_generation);
229 }
230
231 void
232 getbintime(struct bintime *bt)
233 {
234 struct timehands *th;
235 u_int gen;
236
237 ngetbintime++;
238 do {
239 th = timehands;
240 gen = th->th_generation;
241 *bt = th->th_offset;
242 } while (gen == 0 || gen != th->th_generation);
243 bintime_add(bt, &boottimebin);
244 }
245
246 void
247 getnanotime(struct timespec *tsp)
248 {
249 struct timehands *th;
250 u_int gen;
251
252 ngetnanotime++;
253 do {
254 th = timehands;
255 gen = th->th_generation;
256 *tsp = th->th_nanotime;
257 } while (gen == 0 || gen != th->th_generation);
258 }
259
260 void
261 getmicrotime(struct timeval *tvp)
262 {
263 struct timehands *th;
264 u_int gen;
265
266 ngetmicrotime++;
267 do {
268 th = timehands;
269 gen = th->th_generation;
270 *tvp = th->th_microtime;
271 } while (gen == 0 || gen != th->th_generation);
272 }
273
274 /*
275 * Initialize a new timecounter.
276 * We should really try to rank the timecounters and intelligently determine
277 * if the new timecounter is better than the current one. This is subject
278 * to further study. For now always use the new timecounter.
279 */
280 void
281 tc_init(struct timecounter *tc)
282 {
283 unsigned u;
284
285 printf("Timecounter \"%s\" frequency %ju Hz",
286 tc->tc_name, (intmax_t)tc->tc_frequency);
287
288 u = tc->tc_frequency / tc->tc_counter_mask;
289 if (u > hz) {
290 printf(" -- Insufficient hz, needs at least %u\n", u);
291 return;
292 }
293 tc->tc_next = timecounters;
294 timecounters = tc;
295 printf("\n");
296 (void)tc->tc_get_timecount(tc);
297 (void)tc->tc_get_timecount(tc);
298 timecounter = tc;
299 }
300
301 /* Report the frequency of the current timecounter. */
302 u_int64_t
303 tc_getfrequency(void)
304 {
305
306 return (timehands->th_counter->tc_frequency);
307 }
308
309 /*
310 * Step our concept of GMT. This is done by modifying our estimate of
311 * when we booted. XXX: needs futher work.
312 */
313 void
314 tc_setclock(struct timespec *ts)
315 {
316 struct timespec ts2;
317
318 nsetclock++;
319 nanouptime(&ts2);
320 boottime.tv_sec = ts->tv_sec - ts2.tv_sec;
321 /* XXX boottime should probably be a timespec. */
322 boottime.tv_usec = (ts->tv_nsec - ts2.tv_nsec) / 1000;
323 if (boottime.tv_usec < 0) {
324 boottime.tv_usec += 1000000;
325 boottime.tv_sec--;
326 }
327 timeval2bintime(&boottime, &boottimebin);
328
329 /* XXX fiddle all the little crinkly bits around the fiords... */
330 tc_windup();
331 }
332
333 /*
334 * Initialize the next struct timehands in the ring and make
335 * it the active timehands. Along the way we might switch to a different
336 * timecounter and/or do seconds processing in NTP. Slightly magic.
337 */
338 static void
339 tc_windup(void)
340 {
341 struct bintime bt;
342 struct timehands *th, *tho;
343 u_int64_t scale;
344 u_int delta, ncount, ogen;
345 int i;
346
347 /*
348 * Make the next timehands a copy of the current one, but do not
349 * overwrite the generation or next pointer. While we update
350 * the contents, the generation must be zero.
351 */
352 tho = timehands;
353 th = tho->th_next;
354 ogen = th->th_generation;
355 th->th_generation = 0;
356 bcopy(tho, th, offsetof(struct timehands, th_generation));
357
358 /*
359 * Capture a timecounter delta on the current timecounter and if
360 * changing timecounters, a counter value from the new timecounter.
361 * Update the offset fields accordingly.
362 */
363 delta = tc_delta(th);
364 if (th->th_counter != timecounter)
365 ncount = timecounter->tc_get_timecount(timecounter);
366 else
367 ncount = 0;
368 th->th_offset_count += delta;
369 th->th_offset_count &= th->th_counter->tc_counter_mask;
370 bintime_addx(&th->th_offset, th->th_scale * delta);
371
372 /*
373 * Hardware latching timecounters may not generate interrupts on
374 * PPS events, so instead we poll them. There is a finite risk that
375 * the hardware might capture a count which is later than the one we
376 * got above, and therefore possibly in the next NTP second which might
377 * have a different rate than the current NTP second. It doesn't
378 * matter in practice.
379 */
380 if (tho->th_counter->tc_poll_pps)
381 tho->th_counter->tc_poll_pps(tho->th_counter);
382
383 /*
384 * Deal with NTP second processing. The for loop normally only
385 * iterates once, but in extreme situations it might keep NTP sane
386 * if timeouts are not run for several seconds.
387 */
388 for (i = th->th_offset.sec - tho->th_offset.sec; i > 0; i--)
389 ntp_update_second(&th->th_adjustment, &th->th_offset.sec);
390
391 /* Now is a good time to change timecounters. */
392 if (th->th_counter != timecounter) {
393 th->th_counter = timecounter;
394 th->th_offset_count = ncount;
395 }
396
397 /*-
398 * Recalculate the scaling factor. We want the number of 1/2^64
399 * fractions of a second per period of the hardware counter, taking
400 * into account the th_adjustment factor which the NTP PLL/adjtime(2)
401 * processing provides us with.
402 *
403 * The th_adjustment is nanoseconds per second with 32 bit binary
404 * fraction and want 64 bit binary fraction of second:
405 *
406 * x = a * 2^32 / 10^9 = a * 4.294967296
407 *
408 * The range of th_adjustment is +/- 5000PPM so inside a 64bit int
409 * we can only multiply by about 850 without overflowing, but that
410 * leaves suitably precise fractions for multiply before divide.
411 *
412 * Divide before multiply with a fraction of 2199/512 results in a
413 * systematic undercompensation of 10PPM of th_adjustment. On a
414 * 5000PPM adjustment this is a 0.05PPM error. This is acceptable.
415 *
416 * We happily sacrifice the lowest of the 64 bits of our result
417 * to the goddess of code clarity.
418 *
419 */
420 scale = (u_int64_t)1 << 63;
421 scale += (th->th_adjustment / 1024) * 2199;
422 scale /= th->th_counter->tc_frequency;
423 th->th_scale = scale * 2;
424
425 /* Update the GMT timestamps used for the get*() functions. */
426 bt = th->th_offset;
427 bintime_add(&bt, &boottimebin);
428 bintime2timeval(&bt, &th->th_microtime);
429 bintime2timespec(&bt, &th->th_nanotime);
430
431 /*
432 * Now that the struct timehands is again consistent, set the new
433 * generation number, making sure to not make it zero.
434 */
435 if (++ogen == 0)
436 ogen = 1;
437 th->th_generation = ogen;
438
439 /* Go live with the new struct timehands. */
440 time_second = th->th_microtime.tv_sec;
441 time_uptime = th->th_offset.sec;
442 timehands = th;
443 }
444
445 /* Report or change the active timecounter hardware. */
446 static int
447 sysctl_kern_timecounter_hardware(SYSCTL_HANDLER_ARGS)
448 {
449 char newname[32];
450 struct timecounter *newtc, *tc;
451 int error;
452
453 tc = timecounter;
454 strlcpy(newname, tc->tc_name, sizeof(newname));
455
456 error = sysctl_handle_string(oidp, &newname[0], sizeof(newname), req);
457 if (error != 0 || req->newptr == NULL ||
458 strcmp(newname, tc->tc_name) == 0)
459 return (error);
460 for (newtc = timecounters; newtc != NULL; newtc = newtc->tc_next) {
461 if (strcmp(newname, newtc->tc_name) != 0)
462 continue;
463
464 /* Warm up new timecounter. */
465 (void)newtc->tc_get_timecount(newtc);
466 (void)newtc->tc_get_timecount(newtc);
467
468 timecounter = newtc;
469 return (0);
470 }
471 return (EINVAL);
472 }
473
474 SYSCTL_PROC(_kern_timecounter, OID_AUTO, hardware, CTLTYPE_STRING | CTLFLAG_RW,
475 0, 0, sysctl_kern_timecounter_hardware, "A", "");
476
477 /*
478 * RFC 2783 PPS-API implementation.
479 */
480
481 int
482 pps_ioctl(u_long cmd, caddr_t data, struct pps_state *pps)
483 {
484 pps_params_t *app;
485 struct pps_fetch_args *fapi;
486 #ifdef PPS_SYNC
487 struct pps_kcbind_args *kapi;
488 #endif
489
490 switch (cmd) {
491 case PPS_IOC_CREATE:
492 return (0);
493 case PPS_IOC_DESTROY:
494 return (0);
495 case PPS_IOC_SETPARAMS:
496 app = (pps_params_t *)data;
497 if (app->mode & ~pps->ppscap)
498 return (EINVAL);
499 pps->ppsparam = *app;
500 return (0);
501 case PPS_IOC_GETPARAMS:
502 app = (pps_params_t *)data;
503 *app = pps->ppsparam;
504 app->api_version = PPS_API_VERS_1;
505 return (0);
506 case PPS_IOC_GETCAP:
507 *(int*)data = pps->ppscap;
508 return (0);
509 case PPS_IOC_FETCH:
510 fapi = (struct pps_fetch_args *)data;
511 if (fapi->tsformat && fapi->tsformat != PPS_TSFMT_TSPEC)
512 return (EINVAL);
513 if (fapi->timeout.tv_sec || fapi->timeout.tv_nsec)
514 return (EOPNOTSUPP);
515 pps->ppsinfo.current_mode = pps->ppsparam.mode;
516 fapi->pps_info_buf = pps->ppsinfo;
517 return (0);
518 case PPS_IOC_KCBIND:
519 #ifdef PPS_SYNC
520 kapi = (struct pps_kcbind_args *)data;
521 /* XXX Only root should be able to do this */
522 if (kapi->tsformat && kapi->tsformat != PPS_TSFMT_TSPEC)
523 return (EINVAL);
524 if (kapi->kernel_consumer != PPS_KC_HARDPPS)
525 return (EINVAL);
526 if (kapi->edge & ~pps->ppscap)
527 return (EINVAL);
528 pps->kcmode = kapi->edge;
529 return (0);
530 #else
531 return (EOPNOTSUPP);
532 #endif
533 default:
534 return (ENOTTY);
535 }
536 }
537
538 void
539 pps_init(struct pps_state *pps)
540 {
541 pps->ppscap |= PPS_TSFMT_TSPEC;
542 if (pps->ppscap & PPS_CAPTUREASSERT)
543 pps->ppscap |= PPS_OFFSETASSERT;
544 if (pps->ppscap & PPS_CAPTURECLEAR)
545 pps->ppscap |= PPS_OFFSETCLEAR;
546 }
547
548 void
549 pps_capture(struct pps_state *pps)
550 {
551 struct timehands *th;
552
553 th = timehands;
554 pps->capgen = th->th_generation;
555 pps->capth = th;
556 pps->capcount = th->th_counter->tc_get_timecount(th->th_counter);
557 if (pps->capgen != th->th_generation)
558 pps->capgen = 0;
559 }
560
561 void
562 pps_event(struct pps_state *pps, int event)
563 {
564 struct bintime bt;
565 struct timespec ts, *tsp, *osp;
566 u_int tcount, *pcount;
567 int foff, fhard;
568 pps_seq_t *pseq;
569
570 /* If the timecounter was wound up underneath us, bail out. */
571 if (pps->capgen == 0 || pps->capgen != pps->capth->th_generation)
572 return;
573
574 /* Things would be easier with arrays. */
575 if (event == PPS_CAPTUREASSERT) {
576 tsp = &pps->ppsinfo.assert_timestamp;
577 osp = &pps->ppsparam.assert_offset;
578 foff = pps->ppsparam.mode & PPS_OFFSETASSERT;
579 fhard = pps->kcmode & PPS_CAPTUREASSERT;
580 pcount = &pps->ppscount[0];
581 pseq = &pps->ppsinfo.assert_sequence;
582 } else {
583 tsp = &pps->ppsinfo.clear_timestamp;
584 osp = &pps->ppsparam.clear_offset;
585 foff = pps->ppsparam.mode & PPS_OFFSETCLEAR;
586 fhard = pps->kcmode & PPS_CAPTURECLEAR;
587 pcount = &pps->ppscount[1];
588 pseq = &pps->ppsinfo.clear_sequence;
589 }
590
591 /*
592 * If the timecounter changed, we cannot compare the count values, so
593 * we have to drop the rest of the PPS-stuff until the next event.
594 */
595 if (pps->ppstc != pps->capth->th_counter) {
596 pps->ppstc = pps->capth->th_counter;
597 *pcount = pps->capcount;
598 pps->ppscount[2] = pps->capcount;
599 return;
600 }
601
602 /* Return if nothing really happened. */
603 if (*pcount == pps->capcount)
604 return;
605
606 /* Convert the count to a timespec. */
607 tcount = pps->capcount - pps->capth->th_offset_count;
608 tcount &= pps->capth->th_counter->tc_counter_mask;
609 bt = pps->capth->th_offset;
610 bintime_addx(&bt, pps->capth->th_scale * tcount);
611 bintime_add(&bt, &boottimebin);
612 bintime2timespec(&bt, &ts);
613
614 /* If the timecounter was wound up underneath us, bail out. */
615 if (pps->capgen != pps->capth->th_generation)
616 return;
617
618 *pcount = pps->capcount;
619 (*pseq)++;
620 *tsp = ts;
621
622 if (foff) {
623 timespecadd(tsp, osp);
624 if (tsp->tv_nsec < 0) {
625 tsp->tv_nsec += 1000000000;
626 tsp->tv_sec -= 1;
627 }
628 }
629 #ifdef PPS_SYNC
630 if (fhard) {
631 u_int64_t scale;
632
633 /*
634 * Feed the NTP PLL/FLL.
635 * The FLL wants to know how many (hardware) nanoseconds
636 * elapsed since the previous event.
637 */
638 tcount = pps->capcount - pps->ppscount[2];
639 pps->ppscount[2] = pps->capcount;
640 tcount &= pps->capth->th_counter->tc_counter_mask;
641 scale = (u_int64_t)1 << 63;
642 scale /= pps->capth->th_counter->tc_frequency;
643 scale *= 2;
644 bt.sec = 0;
645 bt.frac = 0;
646 bintime_addx(&bt, scale * tcount);
647 bintime2timespec(&bt, &ts);
648 hardpps(tsp, ts.tv_nsec + 1000000000 * ts.tv_sec);
649 }
650 #endif
651 }
652
653 /*
654 * Timecounters need to be updated every so often to prevent the hardware
655 * counter from overflowing. Updating also recalculates the cached values
656 * used by the get*() family of functions, so their precision depends on
657 * the update frequency.
658 */
659
660 static int tc_tick;
661 SYSCTL_INT(_kern_timecounter, OID_AUTO, tick, CTLFLAG_RD, &tc_tick, 0, "");
662
663 void
664 tc_ticktock(void)
665 {
666 static int count;
667
668 if (++count < tc_tick)
669 return;
670 count = 0;
671 tc_windup();
672 }
673
674 static void
675 inittimecounter(void *dummy)
676 {
677 u_int p;
678
679 /*
680 * Set the initial timeout to
681 * max(1, <approx. number of hardclock ticks in a millisecond>).
682 * People should probably not use the sysctl to set the timeout
683 * to smaller than its inital value, since that value is the
684 * smallest reasonable one. If they want better timestamps they
685 * should use the non-"get"* functions.
686 */
687 if (hz > 1000)
688 tc_tick = (hz + 500) / 1000;
689 else
690 tc_tick = 1;
691 p = (tc_tick * 1000000) / hz;
692 printf("Timecounters tick every %d.%03u msec\n", p / 1000, p % 1000);
693
694 /* warm up new timecounter (again) and get rolling. */
695 (void)timecounter->tc_get_timecount(timecounter);
696 (void)timecounter->tc_get_timecount(timecounter);
697 }
698
699 SYSINIT(timecounter, SI_SUB_CLOCKS, SI_ORDER_SECOND, inittimecounter, NULL)
Cache object: c3a76a90065119f6e06739f6b8e03931
|