FreeBSD/Linux Kernel Cross Reference
sys/kern/kern_tc.c
1 /*-
2 * SPDX-License-Identifier: Beerware
3 *
4 * ----------------------------------------------------------------------------
5 * "THE BEER-WARE LICENSE" (Revision 42):
6 * <phk@FreeBSD.ORG> wrote this file. As long as you retain this notice you
7 * can do whatever you want with this stuff. If we meet some day, and you think
8 * this stuff is worth it, you can buy me a beer in return. Poul-Henning Kamp
9 * ----------------------------------------------------------------------------
10 *
11 * Copyright (c) 2011, 2015, 2016 The FreeBSD Foundation
12 * All rights reserved.
13 *
14 * Portions of this software were developed by Julien Ridoux at the University
15 * of Melbourne under sponsorship from the FreeBSD Foundation.
16 *
17 * Portions of this software were developed by Konstantin Belousov
18 * under sponsorship from the FreeBSD Foundation.
19 */
20
21 #include <sys/cdefs.h>
22 __FBSDID("$FreeBSD: releng/12.0/sys/kern/kern_tc.c 336914 2018-07-30 15:46:40Z asomers $");
23
24 #include "opt_ntp.h"
25 #include "opt_ffclock.h"
26
27 #include <sys/param.h>
28 #include <sys/kernel.h>
29 #include <sys/limits.h>
30 #include <sys/lock.h>
31 #include <sys/mutex.h>
32 #include <sys/proc.h>
33 #include <sys/sbuf.h>
34 #include <sys/sleepqueue.h>
35 #include <sys/sysctl.h>
36 #include <sys/syslog.h>
37 #include <sys/systm.h>
38 #include <sys/timeffc.h>
39 #include <sys/timepps.h>
40 #include <sys/timetc.h>
41 #include <sys/timex.h>
42 #include <sys/vdso.h>
43
44 /*
45 * A large step happens on boot. This constant detects such steps.
46 * It is relatively small so that ntp_update_second gets called enough
47 * in the typical 'missed a couple of seconds' case, but doesn't loop
48 * forever when the time step is large.
49 */
50 #define LARGE_STEP 200
51
52 /*
53 * Implement a dummy timecounter which we can use until we get a real one
54 * in the air. This allows the console and other early stuff to use
55 * time services.
56 */
57
58 static u_int
59 dummy_get_timecount(struct timecounter *tc)
60 {
61 static u_int now;
62
63 return (++now);
64 }
65
66 static struct timecounter dummy_timecounter = {
67 dummy_get_timecount, 0, ~0u, 1000000, "dummy", -1000000
68 };
69
70 struct timehands {
71 /* These fields must be initialized by the driver. */
72 struct timecounter *th_counter;
73 int64_t th_adjustment;
74 uint64_t th_scale;
75 u_int th_offset_count;
76 struct bintime th_offset;
77 struct bintime th_bintime;
78 struct timeval th_microtime;
79 struct timespec th_nanotime;
80 struct bintime th_boottime;
81 /* Fields not to be copied in tc_windup start with th_generation. */
82 u_int th_generation;
83 struct timehands *th_next;
84 };
85
86 static struct timehands th0;
87 static struct timehands th1 = {
88 .th_next = &th0
89 };
90 static struct timehands th0 = {
91 .th_counter = &dummy_timecounter,
92 .th_scale = (uint64_t)-1 / 1000000,
93 .th_offset = { .sec = 1 },
94 .th_generation = 1,
95 .th_next = &th1
96 };
97
98 static struct timehands *volatile timehands = &th0;
99 struct timecounter *timecounter = &dummy_timecounter;
100 static struct timecounter *timecounters = &dummy_timecounter;
101
102 int tc_min_ticktock_freq = 1;
103
104 volatile time_t time_second = 1;
105 volatile time_t time_uptime = 1;
106
107 static int sysctl_kern_boottime(SYSCTL_HANDLER_ARGS);
108 SYSCTL_PROC(_kern, KERN_BOOTTIME, boottime, CTLTYPE_STRUCT|CTLFLAG_RD,
109 NULL, 0, sysctl_kern_boottime, "S,timeval", "System boottime");
110
111 SYSCTL_NODE(_kern, OID_AUTO, timecounter, CTLFLAG_RW, 0, "");
112 static SYSCTL_NODE(_kern_timecounter, OID_AUTO, tc, CTLFLAG_RW, 0, "");
113
114 static int timestepwarnings;
115 SYSCTL_INT(_kern_timecounter, OID_AUTO, stepwarnings, CTLFLAG_RW,
116 ×tepwarnings, 0, "Log time steps");
117
118 struct bintime bt_timethreshold;
119 struct bintime bt_tickthreshold;
120 sbintime_t sbt_timethreshold;
121 sbintime_t sbt_tickthreshold;
122 struct bintime tc_tick_bt;
123 sbintime_t tc_tick_sbt;
124 int tc_precexp;
125 int tc_timepercentage = TC_DEFAULTPERC;
126 static int sysctl_kern_timecounter_adjprecision(SYSCTL_HANDLER_ARGS);
127 SYSCTL_PROC(_kern_timecounter, OID_AUTO, alloweddeviation,
128 CTLTYPE_INT | CTLFLAG_RWTUN | CTLFLAG_MPSAFE, 0, 0,
129 sysctl_kern_timecounter_adjprecision, "I",
130 "Allowed time interval deviation in percents");
131
132 volatile int rtc_generation = 1;
133
134 static int tc_chosen; /* Non-zero if a specific tc was chosen via sysctl. */
135
136 static void tc_windup(struct bintime *new_boottimebin);
137 static void cpu_tick_calibrate(int);
138
139 void dtrace_getnanotime(struct timespec *tsp);
140
141 static int
142 sysctl_kern_boottime(SYSCTL_HANDLER_ARGS)
143 {
144 struct timeval boottime;
145
146 getboottime(&boottime);
147
148 #ifndef __mips__
149 #ifdef SCTL_MASK32
150 int tv[2];
151
152 if (req->flags & SCTL_MASK32) {
153 tv[0] = boottime.tv_sec;
154 tv[1] = boottime.tv_usec;
155 return (SYSCTL_OUT(req, tv, sizeof(tv)));
156 }
157 #endif
158 #endif
159 return (SYSCTL_OUT(req, &boottime, sizeof(boottime)));
160 }
161
162 static int
163 sysctl_kern_timecounter_get(SYSCTL_HANDLER_ARGS)
164 {
165 u_int ncount;
166 struct timecounter *tc = arg1;
167
168 ncount = tc->tc_get_timecount(tc);
169 return (sysctl_handle_int(oidp, &ncount, 0, req));
170 }
171
172 static int
173 sysctl_kern_timecounter_freq(SYSCTL_HANDLER_ARGS)
174 {
175 uint64_t freq;
176 struct timecounter *tc = arg1;
177
178 freq = tc->tc_frequency;
179 return (sysctl_handle_64(oidp, &freq, 0, req));
180 }
181
182 /*
183 * Return the difference between the timehands' counter value now and what
184 * was when we copied it to the timehands' offset_count.
185 */
186 static __inline u_int
187 tc_delta(struct timehands *th)
188 {
189 struct timecounter *tc;
190
191 tc = th->th_counter;
192 return ((tc->tc_get_timecount(tc) - th->th_offset_count) &
193 tc->tc_counter_mask);
194 }
195
196 /*
197 * Functions for reading the time. We have to loop until we are sure that
198 * the timehands that we operated on was not updated under our feet. See
199 * the comment in <sys/time.h> for a description of these 12 functions.
200 */
201
202 #ifdef FFCLOCK
203 void
204 fbclock_binuptime(struct bintime *bt)
205 {
206 struct timehands *th;
207 unsigned int gen;
208
209 do {
210 th = timehands;
211 gen = atomic_load_acq_int(&th->th_generation);
212 *bt = th->th_offset;
213 bintime_addx(bt, th->th_scale * tc_delta(th));
214 atomic_thread_fence_acq();
215 } while (gen == 0 || gen != th->th_generation);
216 }
217
218 void
219 fbclock_nanouptime(struct timespec *tsp)
220 {
221 struct bintime bt;
222
223 fbclock_binuptime(&bt);
224 bintime2timespec(&bt, tsp);
225 }
226
227 void
228 fbclock_microuptime(struct timeval *tvp)
229 {
230 struct bintime bt;
231
232 fbclock_binuptime(&bt);
233 bintime2timeval(&bt, tvp);
234 }
235
236 void
237 fbclock_bintime(struct bintime *bt)
238 {
239 struct timehands *th;
240 unsigned int gen;
241
242 do {
243 th = timehands;
244 gen = atomic_load_acq_int(&th->th_generation);
245 *bt = th->th_bintime;
246 bintime_addx(bt, th->th_scale * tc_delta(th));
247 atomic_thread_fence_acq();
248 } while (gen == 0 || gen != th->th_generation);
249 }
250
251 void
252 fbclock_nanotime(struct timespec *tsp)
253 {
254 struct bintime bt;
255
256 fbclock_bintime(&bt);
257 bintime2timespec(&bt, tsp);
258 }
259
260 void
261 fbclock_microtime(struct timeval *tvp)
262 {
263 struct bintime bt;
264
265 fbclock_bintime(&bt);
266 bintime2timeval(&bt, tvp);
267 }
268
269 void
270 fbclock_getbinuptime(struct bintime *bt)
271 {
272 struct timehands *th;
273 unsigned int gen;
274
275 do {
276 th = timehands;
277 gen = atomic_load_acq_int(&th->th_generation);
278 *bt = th->th_offset;
279 atomic_thread_fence_acq();
280 } while (gen == 0 || gen != th->th_generation);
281 }
282
283 void
284 fbclock_getnanouptime(struct timespec *tsp)
285 {
286 struct timehands *th;
287 unsigned int gen;
288
289 do {
290 th = timehands;
291 gen = atomic_load_acq_int(&th->th_generation);
292 bintime2timespec(&th->th_offset, tsp);
293 atomic_thread_fence_acq();
294 } while (gen == 0 || gen != th->th_generation);
295 }
296
297 void
298 fbclock_getmicrouptime(struct timeval *tvp)
299 {
300 struct timehands *th;
301 unsigned int gen;
302
303 do {
304 th = timehands;
305 gen = atomic_load_acq_int(&th->th_generation);
306 bintime2timeval(&th->th_offset, tvp);
307 atomic_thread_fence_acq();
308 } while (gen == 0 || gen != th->th_generation);
309 }
310
311 void
312 fbclock_getbintime(struct bintime *bt)
313 {
314 struct timehands *th;
315 unsigned int gen;
316
317 do {
318 th = timehands;
319 gen = atomic_load_acq_int(&th->th_generation);
320 *bt = th->th_bintime;
321 atomic_thread_fence_acq();
322 } while (gen == 0 || gen != th->th_generation);
323 }
324
325 void
326 fbclock_getnanotime(struct timespec *tsp)
327 {
328 struct timehands *th;
329 unsigned int gen;
330
331 do {
332 th = timehands;
333 gen = atomic_load_acq_int(&th->th_generation);
334 *tsp = th->th_nanotime;
335 atomic_thread_fence_acq();
336 } while (gen == 0 || gen != th->th_generation);
337 }
338
339 void
340 fbclock_getmicrotime(struct timeval *tvp)
341 {
342 struct timehands *th;
343 unsigned int gen;
344
345 do {
346 th = timehands;
347 gen = atomic_load_acq_int(&th->th_generation);
348 *tvp = th->th_microtime;
349 atomic_thread_fence_acq();
350 } while (gen == 0 || gen != th->th_generation);
351 }
352 #else /* !FFCLOCK */
353 void
354 binuptime(struct bintime *bt)
355 {
356 struct timehands *th;
357 u_int gen;
358
359 do {
360 th = timehands;
361 gen = atomic_load_acq_int(&th->th_generation);
362 *bt = th->th_offset;
363 bintime_addx(bt, th->th_scale * tc_delta(th));
364 atomic_thread_fence_acq();
365 } while (gen == 0 || gen != th->th_generation);
366 }
367
368 void
369 nanouptime(struct timespec *tsp)
370 {
371 struct bintime bt;
372
373 binuptime(&bt);
374 bintime2timespec(&bt, tsp);
375 }
376
377 void
378 microuptime(struct timeval *tvp)
379 {
380 struct bintime bt;
381
382 binuptime(&bt);
383 bintime2timeval(&bt, tvp);
384 }
385
386 void
387 bintime(struct bintime *bt)
388 {
389 struct timehands *th;
390 u_int gen;
391
392 do {
393 th = timehands;
394 gen = atomic_load_acq_int(&th->th_generation);
395 *bt = th->th_bintime;
396 bintime_addx(bt, th->th_scale * tc_delta(th));
397 atomic_thread_fence_acq();
398 } while (gen == 0 || gen != th->th_generation);
399 }
400
401 void
402 nanotime(struct timespec *tsp)
403 {
404 struct bintime bt;
405
406 bintime(&bt);
407 bintime2timespec(&bt, tsp);
408 }
409
410 void
411 microtime(struct timeval *tvp)
412 {
413 struct bintime bt;
414
415 bintime(&bt);
416 bintime2timeval(&bt, tvp);
417 }
418
419 void
420 getbinuptime(struct bintime *bt)
421 {
422 struct timehands *th;
423 u_int gen;
424
425 do {
426 th = timehands;
427 gen = atomic_load_acq_int(&th->th_generation);
428 *bt = th->th_offset;
429 atomic_thread_fence_acq();
430 } while (gen == 0 || gen != th->th_generation);
431 }
432
433 void
434 getnanouptime(struct timespec *tsp)
435 {
436 struct timehands *th;
437 u_int gen;
438
439 do {
440 th = timehands;
441 gen = atomic_load_acq_int(&th->th_generation);
442 bintime2timespec(&th->th_offset, tsp);
443 atomic_thread_fence_acq();
444 } while (gen == 0 || gen != th->th_generation);
445 }
446
447 void
448 getmicrouptime(struct timeval *tvp)
449 {
450 struct timehands *th;
451 u_int gen;
452
453 do {
454 th = timehands;
455 gen = atomic_load_acq_int(&th->th_generation);
456 bintime2timeval(&th->th_offset, tvp);
457 atomic_thread_fence_acq();
458 } while (gen == 0 || gen != th->th_generation);
459 }
460
461 void
462 getbintime(struct bintime *bt)
463 {
464 struct timehands *th;
465 u_int gen;
466
467 do {
468 th = timehands;
469 gen = atomic_load_acq_int(&th->th_generation);
470 *bt = th->th_bintime;
471 atomic_thread_fence_acq();
472 } while (gen == 0 || gen != th->th_generation);
473 }
474
475 void
476 getnanotime(struct timespec *tsp)
477 {
478 struct timehands *th;
479 u_int gen;
480
481 do {
482 th = timehands;
483 gen = atomic_load_acq_int(&th->th_generation);
484 *tsp = th->th_nanotime;
485 atomic_thread_fence_acq();
486 } while (gen == 0 || gen != th->th_generation);
487 }
488
489 void
490 getmicrotime(struct timeval *tvp)
491 {
492 struct timehands *th;
493 u_int gen;
494
495 do {
496 th = timehands;
497 gen = atomic_load_acq_int(&th->th_generation);
498 *tvp = th->th_microtime;
499 atomic_thread_fence_acq();
500 } while (gen == 0 || gen != th->th_generation);
501 }
502 #endif /* FFCLOCK */
503
504 void
505 getboottime(struct timeval *boottime)
506 {
507 struct bintime boottimebin;
508
509 getboottimebin(&boottimebin);
510 bintime2timeval(&boottimebin, boottime);
511 }
512
513 void
514 getboottimebin(struct bintime *boottimebin)
515 {
516 struct timehands *th;
517 u_int gen;
518
519 do {
520 th = timehands;
521 gen = atomic_load_acq_int(&th->th_generation);
522 *boottimebin = th->th_boottime;
523 atomic_thread_fence_acq();
524 } while (gen == 0 || gen != th->th_generation);
525 }
526
527 #ifdef FFCLOCK
528 /*
529 * Support for feed-forward synchronization algorithms. This is heavily inspired
530 * by the timehands mechanism but kept independent from it. *_windup() functions
531 * have some connection to avoid accessing the timecounter hardware more than
532 * necessary.
533 */
534
535 /* Feed-forward clock estimates kept updated by the synchronization daemon. */
536 struct ffclock_estimate ffclock_estimate;
537 struct bintime ffclock_boottime; /* Feed-forward boot time estimate. */
538 uint32_t ffclock_status; /* Feed-forward clock status. */
539 int8_t ffclock_updated; /* New estimates are available. */
540 struct mtx ffclock_mtx; /* Mutex on ffclock_estimate. */
541
542 struct fftimehands {
543 struct ffclock_estimate cest;
544 struct bintime tick_time;
545 struct bintime tick_time_lerp;
546 ffcounter tick_ffcount;
547 uint64_t period_lerp;
548 volatile uint8_t gen;
549 struct fftimehands *next;
550 };
551
552 #define NUM_ELEMENTS(x) (sizeof(x) / sizeof(*x))
553
554 static struct fftimehands ffth[10];
555 static struct fftimehands *volatile fftimehands = ffth;
556
557 static void
558 ffclock_init(void)
559 {
560 struct fftimehands *cur;
561 struct fftimehands *last;
562
563 memset(ffth, 0, sizeof(ffth));
564
565 last = ffth + NUM_ELEMENTS(ffth) - 1;
566 for (cur = ffth; cur < last; cur++)
567 cur->next = cur + 1;
568 last->next = ffth;
569
570 ffclock_updated = 0;
571 ffclock_status = FFCLOCK_STA_UNSYNC;
572 mtx_init(&ffclock_mtx, "ffclock lock", NULL, MTX_DEF);
573 }
574
575 /*
576 * Reset the feed-forward clock estimates. Called from inittodr() to get things
577 * kick started and uses the timecounter nominal frequency as a first period
578 * estimate. Note: this function may be called several time just after boot.
579 * Note: this is the only function that sets the value of boot time for the
580 * monotonic (i.e. uptime) version of the feed-forward clock.
581 */
582 void
583 ffclock_reset_clock(struct timespec *ts)
584 {
585 struct timecounter *tc;
586 struct ffclock_estimate cest;
587
588 tc = timehands->th_counter;
589 memset(&cest, 0, sizeof(struct ffclock_estimate));
590
591 timespec2bintime(ts, &ffclock_boottime);
592 timespec2bintime(ts, &(cest.update_time));
593 ffclock_read_counter(&cest.update_ffcount);
594 cest.leapsec_next = 0;
595 cest.period = ((1ULL << 63) / tc->tc_frequency) << 1;
596 cest.errb_abs = 0;
597 cest.errb_rate = 0;
598 cest.status = FFCLOCK_STA_UNSYNC;
599 cest.leapsec_total = 0;
600 cest.leapsec = 0;
601
602 mtx_lock(&ffclock_mtx);
603 bcopy(&cest, &ffclock_estimate, sizeof(struct ffclock_estimate));
604 ffclock_updated = INT8_MAX;
605 mtx_unlock(&ffclock_mtx);
606
607 printf("ffclock reset: %s (%llu Hz), time = %ld.%09lu\n", tc->tc_name,
608 (unsigned long long)tc->tc_frequency, (long)ts->tv_sec,
609 (unsigned long)ts->tv_nsec);
610 }
611
612 /*
613 * Sub-routine to convert a time interval measured in RAW counter units to time
614 * in seconds stored in bintime format.
615 * NOTE: bintime_mul requires u_int, but the value of the ffcounter may be
616 * larger than the max value of u_int (on 32 bit architecture). Loop to consume
617 * extra cycles.
618 */
619 static void
620 ffclock_convert_delta(ffcounter ffdelta, uint64_t period, struct bintime *bt)
621 {
622 struct bintime bt2;
623 ffcounter delta, delta_max;
624
625 delta_max = (1ULL << (8 * sizeof(unsigned int))) - 1;
626 bintime_clear(bt);
627 do {
628 if (ffdelta > delta_max)
629 delta = delta_max;
630 else
631 delta = ffdelta;
632 bt2.sec = 0;
633 bt2.frac = period;
634 bintime_mul(&bt2, (unsigned int)delta);
635 bintime_add(bt, &bt2);
636 ffdelta -= delta;
637 } while (ffdelta > 0);
638 }
639
640 /*
641 * Update the fftimehands.
642 * Push the tick ffcount and time(s) forward based on current clock estimate.
643 * The conversion from ffcounter to bintime relies on the difference clock
644 * principle, whose accuracy relies on computing small time intervals. If a new
645 * clock estimate has been passed by the synchronisation daemon, make it
646 * current, and compute the linear interpolation for monotonic time if needed.
647 */
648 static void
649 ffclock_windup(unsigned int delta)
650 {
651 struct ffclock_estimate *cest;
652 struct fftimehands *ffth;
653 struct bintime bt, gap_lerp;
654 ffcounter ffdelta;
655 uint64_t frac;
656 unsigned int polling;
657 uint8_t forward_jump, ogen;
658
659 /*
660 * Pick the next timehand, copy current ffclock estimates and move tick
661 * times and counter forward.
662 */
663 forward_jump = 0;
664 ffth = fftimehands->next;
665 ogen = ffth->gen;
666 ffth->gen = 0;
667 cest = &ffth->cest;
668 bcopy(&fftimehands->cest, cest, sizeof(struct ffclock_estimate));
669 ffdelta = (ffcounter)delta;
670 ffth->period_lerp = fftimehands->period_lerp;
671
672 ffth->tick_time = fftimehands->tick_time;
673 ffclock_convert_delta(ffdelta, cest->period, &bt);
674 bintime_add(&ffth->tick_time, &bt);
675
676 ffth->tick_time_lerp = fftimehands->tick_time_lerp;
677 ffclock_convert_delta(ffdelta, ffth->period_lerp, &bt);
678 bintime_add(&ffth->tick_time_lerp, &bt);
679
680 ffth->tick_ffcount = fftimehands->tick_ffcount + ffdelta;
681
682 /*
683 * Assess the status of the clock, if the last update is too old, it is
684 * likely the synchronisation daemon is dead and the clock is free
685 * running.
686 */
687 if (ffclock_updated == 0) {
688 ffdelta = ffth->tick_ffcount - cest->update_ffcount;
689 ffclock_convert_delta(ffdelta, cest->period, &bt);
690 if (bt.sec > 2 * FFCLOCK_SKM_SCALE)
691 ffclock_status |= FFCLOCK_STA_UNSYNC;
692 }
693
694 /*
695 * If available, grab updated clock estimates and make them current.
696 * Recompute time at this tick using the updated estimates. The clock
697 * estimates passed the feed-forward synchronisation daemon may result
698 * in time conversion that is not monotonically increasing (just after
699 * the update). time_lerp is a particular linear interpolation over the
700 * synchronisation algo polling period that ensures monotonicity for the
701 * clock ids requesting it.
702 */
703 if (ffclock_updated > 0) {
704 bcopy(&ffclock_estimate, cest, sizeof(struct ffclock_estimate));
705 ffdelta = ffth->tick_ffcount - cest->update_ffcount;
706 ffth->tick_time = cest->update_time;
707 ffclock_convert_delta(ffdelta, cest->period, &bt);
708 bintime_add(&ffth->tick_time, &bt);
709
710 /* ffclock_reset sets ffclock_updated to INT8_MAX */
711 if (ffclock_updated == INT8_MAX)
712 ffth->tick_time_lerp = ffth->tick_time;
713
714 if (bintime_cmp(&ffth->tick_time, &ffth->tick_time_lerp, >))
715 forward_jump = 1;
716 else
717 forward_jump = 0;
718
719 bintime_clear(&gap_lerp);
720 if (forward_jump) {
721 gap_lerp = ffth->tick_time;
722 bintime_sub(&gap_lerp, &ffth->tick_time_lerp);
723 } else {
724 gap_lerp = ffth->tick_time_lerp;
725 bintime_sub(&gap_lerp, &ffth->tick_time);
726 }
727
728 /*
729 * The reset from the RTC clock may be far from accurate, and
730 * reducing the gap between real time and interpolated time
731 * could take a very long time if the interpolated clock insists
732 * on strict monotonicity. The clock is reset under very strict
733 * conditions (kernel time is known to be wrong and
734 * synchronization daemon has been restarted recently.
735 * ffclock_boottime absorbs the jump to ensure boot time is
736 * correct and uptime functions stay consistent.
737 */
738 if (((ffclock_status & FFCLOCK_STA_UNSYNC) == FFCLOCK_STA_UNSYNC) &&
739 ((cest->status & FFCLOCK_STA_UNSYNC) == 0) &&
740 ((cest->status & FFCLOCK_STA_WARMUP) == FFCLOCK_STA_WARMUP)) {
741 if (forward_jump)
742 bintime_add(&ffclock_boottime, &gap_lerp);
743 else
744 bintime_sub(&ffclock_boottime, &gap_lerp);
745 ffth->tick_time_lerp = ffth->tick_time;
746 bintime_clear(&gap_lerp);
747 }
748
749 ffclock_status = cest->status;
750 ffth->period_lerp = cest->period;
751
752 /*
753 * Compute corrected period used for the linear interpolation of
754 * time. The rate of linear interpolation is capped to 5000PPM
755 * (5ms/s).
756 */
757 if (bintime_isset(&gap_lerp)) {
758 ffdelta = cest->update_ffcount;
759 ffdelta -= fftimehands->cest.update_ffcount;
760 ffclock_convert_delta(ffdelta, cest->period, &bt);
761 polling = bt.sec;
762 bt.sec = 0;
763 bt.frac = 5000000 * (uint64_t)18446744073LL;
764 bintime_mul(&bt, polling);
765 if (bintime_cmp(&gap_lerp, &bt, >))
766 gap_lerp = bt;
767
768 /* Approximate 1 sec by 1-(1/2^64) to ease arithmetic */
769 frac = 0;
770 if (gap_lerp.sec > 0) {
771 frac -= 1;
772 frac /= ffdelta / gap_lerp.sec;
773 }
774 frac += gap_lerp.frac / ffdelta;
775
776 if (forward_jump)
777 ffth->period_lerp += frac;
778 else
779 ffth->period_lerp -= frac;
780 }
781
782 ffclock_updated = 0;
783 }
784 if (++ogen == 0)
785 ogen = 1;
786 ffth->gen = ogen;
787 fftimehands = ffth;
788 }
789
790 /*
791 * Adjust the fftimehands when the timecounter is changed. Stating the obvious,
792 * the old and new hardware counter cannot be read simultaneously. tc_windup()
793 * does read the two counters 'back to back', but a few cycles are effectively
794 * lost, and not accumulated in tick_ffcount. This is a fairly radical
795 * operation for a feed-forward synchronization daemon, and it is its job to not
796 * pushing irrelevant data to the kernel. Because there is no locking here,
797 * simply force to ignore pending or next update to give daemon a chance to
798 * realize the counter has changed.
799 */
800 static void
801 ffclock_change_tc(struct timehands *th)
802 {
803 struct fftimehands *ffth;
804 struct ffclock_estimate *cest;
805 struct timecounter *tc;
806 uint8_t ogen;
807
808 tc = th->th_counter;
809 ffth = fftimehands->next;
810 ogen = ffth->gen;
811 ffth->gen = 0;
812
813 cest = &ffth->cest;
814 bcopy(&(fftimehands->cest), cest, sizeof(struct ffclock_estimate));
815 cest->period = ((1ULL << 63) / tc->tc_frequency ) << 1;
816 cest->errb_abs = 0;
817 cest->errb_rate = 0;
818 cest->status |= FFCLOCK_STA_UNSYNC;
819
820 ffth->tick_ffcount = fftimehands->tick_ffcount;
821 ffth->tick_time_lerp = fftimehands->tick_time_lerp;
822 ffth->tick_time = fftimehands->tick_time;
823 ffth->period_lerp = cest->period;
824
825 /* Do not lock but ignore next update from synchronization daemon. */
826 ffclock_updated--;
827
828 if (++ogen == 0)
829 ogen = 1;
830 ffth->gen = ogen;
831 fftimehands = ffth;
832 }
833
834 /*
835 * Retrieve feed-forward counter and time of last kernel tick.
836 */
837 void
838 ffclock_last_tick(ffcounter *ffcount, struct bintime *bt, uint32_t flags)
839 {
840 struct fftimehands *ffth;
841 uint8_t gen;
842
843 /*
844 * No locking but check generation has not changed. Also need to make
845 * sure ffdelta is positive, i.e. ffcount > tick_ffcount.
846 */
847 do {
848 ffth = fftimehands;
849 gen = ffth->gen;
850 if ((flags & FFCLOCK_LERP) == FFCLOCK_LERP)
851 *bt = ffth->tick_time_lerp;
852 else
853 *bt = ffth->tick_time;
854 *ffcount = ffth->tick_ffcount;
855 } while (gen == 0 || gen != ffth->gen);
856 }
857
858 /*
859 * Absolute clock conversion. Low level function to convert ffcounter to
860 * bintime. The ffcounter is converted using the current ffclock period estimate
861 * or the "interpolated period" to ensure monotonicity.
862 * NOTE: this conversion may have been deferred, and the clock updated since the
863 * hardware counter has been read.
864 */
865 void
866 ffclock_convert_abs(ffcounter ffcount, struct bintime *bt, uint32_t flags)
867 {
868 struct fftimehands *ffth;
869 struct bintime bt2;
870 ffcounter ffdelta;
871 uint8_t gen;
872
873 /*
874 * No locking but check generation has not changed. Also need to make
875 * sure ffdelta is positive, i.e. ffcount > tick_ffcount.
876 */
877 do {
878 ffth = fftimehands;
879 gen = ffth->gen;
880 if (ffcount > ffth->tick_ffcount)
881 ffdelta = ffcount - ffth->tick_ffcount;
882 else
883 ffdelta = ffth->tick_ffcount - ffcount;
884
885 if ((flags & FFCLOCK_LERP) == FFCLOCK_LERP) {
886 *bt = ffth->tick_time_lerp;
887 ffclock_convert_delta(ffdelta, ffth->period_lerp, &bt2);
888 } else {
889 *bt = ffth->tick_time;
890 ffclock_convert_delta(ffdelta, ffth->cest.period, &bt2);
891 }
892
893 if (ffcount > ffth->tick_ffcount)
894 bintime_add(bt, &bt2);
895 else
896 bintime_sub(bt, &bt2);
897 } while (gen == 0 || gen != ffth->gen);
898 }
899
900 /*
901 * Difference clock conversion.
902 * Low level function to Convert a time interval measured in RAW counter units
903 * into bintime. The difference clock allows measuring small intervals much more
904 * reliably than the absolute clock.
905 */
906 void
907 ffclock_convert_diff(ffcounter ffdelta, struct bintime *bt)
908 {
909 struct fftimehands *ffth;
910 uint8_t gen;
911
912 /* No locking but check generation has not changed. */
913 do {
914 ffth = fftimehands;
915 gen = ffth->gen;
916 ffclock_convert_delta(ffdelta, ffth->cest.period, bt);
917 } while (gen == 0 || gen != ffth->gen);
918 }
919
920 /*
921 * Access to current ffcounter value.
922 */
923 void
924 ffclock_read_counter(ffcounter *ffcount)
925 {
926 struct timehands *th;
927 struct fftimehands *ffth;
928 unsigned int gen, delta;
929
930 /*
931 * ffclock_windup() called from tc_windup(), safe to rely on
932 * th->th_generation only, for correct delta and ffcounter.
933 */
934 do {
935 th = timehands;
936 gen = atomic_load_acq_int(&th->th_generation);
937 ffth = fftimehands;
938 delta = tc_delta(th);
939 *ffcount = ffth->tick_ffcount;
940 atomic_thread_fence_acq();
941 } while (gen == 0 || gen != th->th_generation);
942
943 *ffcount += delta;
944 }
945
946 void
947 binuptime(struct bintime *bt)
948 {
949
950 binuptime_fromclock(bt, sysclock_active);
951 }
952
953 void
954 nanouptime(struct timespec *tsp)
955 {
956
957 nanouptime_fromclock(tsp, sysclock_active);
958 }
959
960 void
961 microuptime(struct timeval *tvp)
962 {
963
964 microuptime_fromclock(tvp, sysclock_active);
965 }
966
967 void
968 bintime(struct bintime *bt)
969 {
970
971 bintime_fromclock(bt, sysclock_active);
972 }
973
974 void
975 nanotime(struct timespec *tsp)
976 {
977
978 nanotime_fromclock(tsp, sysclock_active);
979 }
980
981 void
982 microtime(struct timeval *tvp)
983 {
984
985 microtime_fromclock(tvp, sysclock_active);
986 }
987
988 void
989 getbinuptime(struct bintime *bt)
990 {
991
992 getbinuptime_fromclock(bt, sysclock_active);
993 }
994
995 void
996 getnanouptime(struct timespec *tsp)
997 {
998
999 getnanouptime_fromclock(tsp, sysclock_active);
1000 }
1001
1002 void
1003 getmicrouptime(struct timeval *tvp)
1004 {
1005
1006 getmicrouptime_fromclock(tvp, sysclock_active);
1007 }
1008
1009 void
1010 getbintime(struct bintime *bt)
1011 {
1012
1013 getbintime_fromclock(bt, sysclock_active);
1014 }
1015
1016 void
1017 getnanotime(struct timespec *tsp)
1018 {
1019
1020 getnanotime_fromclock(tsp, sysclock_active);
1021 }
1022
1023 void
1024 getmicrotime(struct timeval *tvp)
1025 {
1026
1027 getmicrouptime_fromclock(tvp, sysclock_active);
1028 }
1029
1030 #endif /* FFCLOCK */
1031
1032 /*
1033 * This is a clone of getnanotime and used for walltimestamps.
1034 * The dtrace_ prefix prevents fbt from creating probes for
1035 * it so walltimestamp can be safely used in all fbt probes.
1036 */
1037 void
1038 dtrace_getnanotime(struct timespec *tsp)
1039 {
1040 struct timehands *th;
1041 u_int gen;
1042
1043 do {
1044 th = timehands;
1045 gen = atomic_load_acq_int(&th->th_generation);
1046 *tsp = th->th_nanotime;
1047 atomic_thread_fence_acq();
1048 } while (gen == 0 || gen != th->th_generation);
1049 }
1050
1051 /*
1052 * System clock currently providing time to the system. Modifiable via sysctl
1053 * when the FFCLOCK option is defined.
1054 */
1055 int sysclock_active = SYSCLOCK_FBCK;
1056
1057 /* Internal NTP status and error estimates. */
1058 extern int time_status;
1059 extern long time_esterror;
1060
1061 /*
1062 * Take a snapshot of sysclock data which can be used to compare system clocks
1063 * and generate timestamps after the fact.
1064 */
1065 void
1066 sysclock_getsnapshot(struct sysclock_snap *clock_snap, int fast)
1067 {
1068 struct fbclock_info *fbi;
1069 struct timehands *th;
1070 struct bintime bt;
1071 unsigned int delta, gen;
1072 #ifdef FFCLOCK
1073 ffcounter ffcount;
1074 struct fftimehands *ffth;
1075 struct ffclock_info *ffi;
1076 struct ffclock_estimate cest;
1077
1078 ffi = &clock_snap->ff_info;
1079 #endif
1080
1081 fbi = &clock_snap->fb_info;
1082 delta = 0;
1083
1084 do {
1085 th = timehands;
1086 gen = atomic_load_acq_int(&th->th_generation);
1087 fbi->th_scale = th->th_scale;
1088 fbi->tick_time = th->th_offset;
1089 #ifdef FFCLOCK
1090 ffth = fftimehands;
1091 ffi->tick_time = ffth->tick_time_lerp;
1092 ffi->tick_time_lerp = ffth->tick_time_lerp;
1093 ffi->period = ffth->cest.period;
1094 ffi->period_lerp = ffth->period_lerp;
1095 clock_snap->ffcount = ffth->tick_ffcount;
1096 cest = ffth->cest;
1097 #endif
1098 if (!fast)
1099 delta = tc_delta(th);
1100 atomic_thread_fence_acq();
1101 } while (gen == 0 || gen != th->th_generation);
1102
1103 clock_snap->delta = delta;
1104 clock_snap->sysclock_active = sysclock_active;
1105
1106 /* Record feedback clock status and error. */
1107 clock_snap->fb_info.status = time_status;
1108 /* XXX: Very crude estimate of feedback clock error. */
1109 bt.sec = time_esterror / 1000000;
1110 bt.frac = ((time_esterror - bt.sec) * 1000000) *
1111 (uint64_t)18446744073709ULL;
1112 clock_snap->fb_info.error = bt;
1113
1114 #ifdef FFCLOCK
1115 if (!fast)
1116 clock_snap->ffcount += delta;
1117
1118 /* Record feed-forward clock leap second adjustment. */
1119 ffi->leapsec_adjustment = cest.leapsec_total;
1120 if (clock_snap->ffcount > cest.leapsec_next)
1121 ffi->leapsec_adjustment -= cest.leapsec;
1122
1123 /* Record feed-forward clock status and error. */
1124 clock_snap->ff_info.status = cest.status;
1125 ffcount = clock_snap->ffcount - cest.update_ffcount;
1126 ffclock_convert_delta(ffcount, cest.period, &bt);
1127 /* 18446744073709 = int(2^64/1e12), err_bound_rate in [ps/s]. */
1128 bintime_mul(&bt, cest.errb_rate * (uint64_t)18446744073709ULL);
1129 /* 18446744073 = int(2^64 / 1e9), since err_abs in [ns]. */
1130 bintime_addx(&bt, cest.errb_abs * (uint64_t)18446744073ULL);
1131 clock_snap->ff_info.error = bt;
1132 #endif
1133 }
1134
1135 /*
1136 * Convert a sysclock snapshot into a struct bintime based on the specified
1137 * clock source and flags.
1138 */
1139 int
1140 sysclock_snap2bintime(struct sysclock_snap *cs, struct bintime *bt,
1141 int whichclock, uint32_t flags)
1142 {
1143 struct bintime boottimebin;
1144 #ifdef FFCLOCK
1145 struct bintime bt2;
1146 uint64_t period;
1147 #endif
1148
1149 switch (whichclock) {
1150 case SYSCLOCK_FBCK:
1151 *bt = cs->fb_info.tick_time;
1152
1153 /* If snapshot was created with !fast, delta will be >0. */
1154 if (cs->delta > 0)
1155 bintime_addx(bt, cs->fb_info.th_scale * cs->delta);
1156
1157 if ((flags & FBCLOCK_UPTIME) == 0) {
1158 getboottimebin(&boottimebin);
1159 bintime_add(bt, &boottimebin);
1160 }
1161 break;
1162 #ifdef FFCLOCK
1163 case SYSCLOCK_FFWD:
1164 if (flags & FFCLOCK_LERP) {
1165 *bt = cs->ff_info.tick_time_lerp;
1166 period = cs->ff_info.period_lerp;
1167 } else {
1168 *bt = cs->ff_info.tick_time;
1169 period = cs->ff_info.period;
1170 }
1171
1172 /* If snapshot was created with !fast, delta will be >0. */
1173 if (cs->delta > 0) {
1174 ffclock_convert_delta(cs->delta, period, &bt2);
1175 bintime_add(bt, &bt2);
1176 }
1177
1178 /* Leap second adjustment. */
1179 if (flags & FFCLOCK_LEAPSEC)
1180 bt->sec -= cs->ff_info.leapsec_adjustment;
1181
1182 /* Boot time adjustment, for uptime/monotonic clocks. */
1183 if (flags & FFCLOCK_UPTIME)
1184 bintime_sub(bt, &ffclock_boottime);
1185 break;
1186 #endif
1187 default:
1188 return (EINVAL);
1189 break;
1190 }
1191
1192 return (0);
1193 }
1194
1195 /*
1196 * Initialize a new timecounter and possibly use it.
1197 */
1198 void
1199 tc_init(struct timecounter *tc)
1200 {
1201 u_int u;
1202 struct sysctl_oid *tc_root;
1203
1204 u = tc->tc_frequency / tc->tc_counter_mask;
1205 /* XXX: We need some margin here, 10% is a guess */
1206 u *= 11;
1207 u /= 10;
1208 if (u > hz && tc->tc_quality >= 0) {
1209 tc->tc_quality = -2000;
1210 if (bootverbose) {
1211 printf("Timecounter \"%s\" frequency %ju Hz",
1212 tc->tc_name, (uintmax_t)tc->tc_frequency);
1213 printf(" -- Insufficient hz, needs at least %u\n", u);
1214 }
1215 } else if (tc->tc_quality >= 0 || bootverbose) {
1216 printf("Timecounter \"%s\" frequency %ju Hz quality %d\n",
1217 tc->tc_name, (uintmax_t)tc->tc_frequency,
1218 tc->tc_quality);
1219 }
1220
1221 tc->tc_next = timecounters;
1222 timecounters = tc;
1223 /*
1224 * Set up sysctl tree for this counter.
1225 */
1226 tc_root = SYSCTL_ADD_NODE_WITH_LABEL(NULL,
1227 SYSCTL_STATIC_CHILDREN(_kern_timecounter_tc), OID_AUTO, tc->tc_name,
1228 CTLFLAG_RW, 0, "timecounter description", "timecounter");
1229 SYSCTL_ADD_UINT(NULL, SYSCTL_CHILDREN(tc_root), OID_AUTO,
1230 "mask", CTLFLAG_RD, &(tc->tc_counter_mask), 0,
1231 "mask for implemented bits");
1232 SYSCTL_ADD_PROC(NULL, SYSCTL_CHILDREN(tc_root), OID_AUTO,
1233 "counter", CTLTYPE_UINT | CTLFLAG_RD, tc, sizeof(*tc),
1234 sysctl_kern_timecounter_get, "IU", "current timecounter value");
1235 SYSCTL_ADD_PROC(NULL, SYSCTL_CHILDREN(tc_root), OID_AUTO,
1236 "frequency", CTLTYPE_U64 | CTLFLAG_RD, tc, sizeof(*tc),
1237 sysctl_kern_timecounter_freq, "QU", "timecounter frequency");
1238 SYSCTL_ADD_INT(NULL, SYSCTL_CHILDREN(tc_root), OID_AUTO,
1239 "quality", CTLFLAG_RD, &(tc->tc_quality), 0,
1240 "goodness of time counter");
1241 /*
1242 * Do not automatically switch if the current tc was specifically
1243 * chosen. Never automatically use a timecounter with negative quality.
1244 * Even though we run on the dummy counter, switching here may be
1245 * worse since this timecounter may not be monotonic.
1246 */
1247 if (tc_chosen)
1248 return;
1249 if (tc->tc_quality < 0)
1250 return;
1251 if (tc->tc_quality < timecounter->tc_quality)
1252 return;
1253 if (tc->tc_quality == timecounter->tc_quality &&
1254 tc->tc_frequency < timecounter->tc_frequency)
1255 return;
1256 (void)tc->tc_get_timecount(tc);
1257 (void)tc->tc_get_timecount(tc);
1258 timecounter = tc;
1259 }
1260
1261 /* Report the frequency of the current timecounter. */
1262 uint64_t
1263 tc_getfrequency(void)
1264 {
1265
1266 return (timehands->th_counter->tc_frequency);
1267 }
1268
1269 static bool
1270 sleeping_on_old_rtc(struct thread *td)
1271 {
1272
1273 /*
1274 * td_rtcgen is modified by curthread when it is running,
1275 * and by other threads in this function. By finding the thread
1276 * on a sleepqueue and holding the lock on the sleepqueue
1277 * chain, we guarantee that the thread is not running and that
1278 * modifying td_rtcgen is safe. Setting td_rtcgen to zero informs
1279 * the thread that it was woken due to a real-time clock adjustment.
1280 * (The declaration of td_rtcgen refers to this comment.)
1281 */
1282 if (td->td_rtcgen != 0 && td->td_rtcgen != rtc_generation) {
1283 td->td_rtcgen = 0;
1284 return (true);
1285 }
1286 return (false);
1287 }
1288
1289 static struct mtx tc_setclock_mtx;
1290 MTX_SYSINIT(tc_setclock_init, &tc_setclock_mtx, "tcsetc", MTX_SPIN);
1291
1292 /*
1293 * Step our concept of UTC. This is done by modifying our estimate of
1294 * when we booted.
1295 */
1296 void
1297 tc_setclock(struct timespec *ts)
1298 {
1299 struct timespec tbef, taft;
1300 struct bintime bt, bt2;
1301
1302 timespec2bintime(ts, &bt);
1303 nanotime(&tbef);
1304 mtx_lock_spin(&tc_setclock_mtx);
1305 cpu_tick_calibrate(1);
1306 binuptime(&bt2);
1307 bintime_sub(&bt, &bt2);
1308
1309 /* XXX fiddle all the little crinkly bits around the fiords... */
1310 tc_windup(&bt);
1311 mtx_unlock_spin(&tc_setclock_mtx);
1312
1313 /* Avoid rtc_generation == 0, since td_rtcgen == 0 is special. */
1314 atomic_add_rel_int(&rtc_generation, 2);
1315 sleepq_chains_remove_matching(sleeping_on_old_rtc);
1316 if (timestepwarnings) {
1317 nanotime(&taft);
1318 log(LOG_INFO,
1319 "Time stepped from %jd.%09ld to %jd.%09ld (%jd.%09ld)\n",
1320 (intmax_t)tbef.tv_sec, tbef.tv_nsec,
1321 (intmax_t)taft.tv_sec, taft.tv_nsec,
1322 (intmax_t)ts->tv_sec, ts->tv_nsec);
1323 }
1324 }
1325
1326 /*
1327 * Initialize the next struct timehands in the ring and make
1328 * it the active timehands. Along the way we might switch to a different
1329 * timecounter and/or do seconds processing in NTP. Slightly magic.
1330 */
1331 static void
1332 tc_windup(struct bintime *new_boottimebin)
1333 {
1334 struct bintime bt;
1335 struct timehands *th, *tho;
1336 uint64_t scale;
1337 u_int delta, ncount, ogen;
1338 int i;
1339 time_t t;
1340
1341 /*
1342 * Make the next timehands a copy of the current one, but do
1343 * not overwrite the generation or next pointer. While we
1344 * update the contents, the generation must be zero. We need
1345 * to ensure that the zero generation is visible before the
1346 * data updates become visible, which requires release fence.
1347 * For similar reasons, re-reading of the generation after the
1348 * data is read should use acquire fence.
1349 */
1350 tho = timehands;
1351 th = tho->th_next;
1352 ogen = th->th_generation;
1353 th->th_generation = 0;
1354 atomic_thread_fence_rel();
1355 memcpy(th, tho, offsetof(struct timehands, th_generation));
1356 if (new_boottimebin != NULL)
1357 th->th_boottime = *new_boottimebin;
1358
1359 /*
1360 * Capture a timecounter delta on the current timecounter and if
1361 * changing timecounters, a counter value from the new timecounter.
1362 * Update the offset fields accordingly.
1363 */
1364 delta = tc_delta(th);
1365 if (th->th_counter != timecounter)
1366 ncount = timecounter->tc_get_timecount(timecounter);
1367 else
1368 ncount = 0;
1369 #ifdef FFCLOCK
1370 ffclock_windup(delta);
1371 #endif
1372 th->th_offset_count += delta;
1373 th->th_offset_count &= th->th_counter->tc_counter_mask;
1374 while (delta > th->th_counter->tc_frequency) {
1375 /* Eat complete unadjusted seconds. */
1376 delta -= th->th_counter->tc_frequency;
1377 th->th_offset.sec++;
1378 }
1379 if ((delta > th->th_counter->tc_frequency / 2) &&
1380 (th->th_scale * delta < ((uint64_t)1 << 63))) {
1381 /* The product th_scale * delta just barely overflows. */
1382 th->th_offset.sec++;
1383 }
1384 bintime_addx(&th->th_offset, th->th_scale * delta);
1385
1386 /*
1387 * Hardware latching timecounters may not generate interrupts on
1388 * PPS events, so instead we poll them. There is a finite risk that
1389 * the hardware might capture a count which is later than the one we
1390 * got above, and therefore possibly in the next NTP second which might
1391 * have a different rate than the current NTP second. It doesn't
1392 * matter in practice.
1393 */
1394 if (tho->th_counter->tc_poll_pps)
1395 tho->th_counter->tc_poll_pps(tho->th_counter);
1396
1397 /*
1398 * Deal with NTP second processing. The for loop normally
1399 * iterates at most once, but in extreme situations it might
1400 * keep NTP sane if timeouts are not run for several seconds.
1401 * At boot, the time step can be large when the TOD hardware
1402 * has been read, so on really large steps, we call
1403 * ntp_update_second only twice. We need to call it twice in
1404 * case we missed a leap second.
1405 */
1406 bt = th->th_offset;
1407 bintime_add(&bt, &th->th_boottime);
1408 i = bt.sec - tho->th_microtime.tv_sec;
1409 if (i > LARGE_STEP)
1410 i = 2;
1411 for (; i > 0; i--) {
1412 t = bt.sec;
1413 ntp_update_second(&th->th_adjustment, &bt.sec);
1414 if (bt.sec != t)
1415 th->th_boottime.sec += bt.sec - t;
1416 }
1417 /* Update the UTC timestamps used by the get*() functions. */
1418 th->th_bintime = bt;
1419 bintime2timeval(&bt, &th->th_microtime);
1420 bintime2timespec(&bt, &th->th_nanotime);
1421
1422 /* Now is a good time to change timecounters. */
1423 if (th->th_counter != timecounter) {
1424 #ifndef __arm__
1425 if ((timecounter->tc_flags & TC_FLAGS_C2STOP) != 0)
1426 cpu_disable_c2_sleep++;
1427 if ((th->th_counter->tc_flags & TC_FLAGS_C2STOP) != 0)
1428 cpu_disable_c2_sleep--;
1429 #endif
1430 th->th_counter = timecounter;
1431 th->th_offset_count = ncount;
1432 tc_min_ticktock_freq = max(1, timecounter->tc_frequency /
1433 (((uint64_t)timecounter->tc_counter_mask + 1) / 3));
1434 #ifdef FFCLOCK
1435 ffclock_change_tc(th);
1436 #endif
1437 }
1438
1439 /*-
1440 * Recalculate the scaling factor. We want the number of 1/2^64
1441 * fractions of a second per period of the hardware counter, taking
1442 * into account the th_adjustment factor which the NTP PLL/adjtime(2)
1443 * processing provides us with.
1444 *
1445 * The th_adjustment is nanoseconds per second with 32 bit binary
1446 * fraction and we want 64 bit binary fraction of second:
1447 *
1448 * x = a * 2^32 / 10^9 = a * 4.294967296
1449 *
1450 * The range of th_adjustment is +/- 5000PPM so inside a 64bit int
1451 * we can only multiply by about 850 without overflowing, that
1452 * leaves no suitably precise fractions for multiply before divide.
1453 *
1454 * Divide before multiply with a fraction of 2199/512 results in a
1455 * systematic undercompensation of 10PPM of th_adjustment. On a
1456 * 5000PPM adjustment this is a 0.05PPM error. This is acceptable.
1457 *
1458 * We happily sacrifice the lowest of the 64 bits of our result
1459 * to the goddess of code clarity.
1460 *
1461 */
1462 scale = (uint64_t)1 << 63;
1463 scale += (th->th_adjustment / 1024) * 2199;
1464 scale /= th->th_counter->tc_frequency;
1465 th->th_scale = scale * 2;
1466
1467 /*
1468 * Now that the struct timehands is again consistent, set the new
1469 * generation number, making sure to not make it zero.
1470 */
1471 if (++ogen == 0)
1472 ogen = 1;
1473 atomic_store_rel_int(&th->th_generation, ogen);
1474
1475 /* Go live with the new struct timehands. */
1476 #ifdef FFCLOCK
1477 switch (sysclock_active) {
1478 case SYSCLOCK_FBCK:
1479 #endif
1480 time_second = th->th_microtime.tv_sec;
1481 time_uptime = th->th_offset.sec;
1482 #ifdef FFCLOCK
1483 break;
1484 case SYSCLOCK_FFWD:
1485 time_second = fftimehands->tick_time_lerp.sec;
1486 time_uptime = fftimehands->tick_time_lerp.sec - ffclock_boottime.sec;
1487 break;
1488 }
1489 #endif
1490
1491 timehands = th;
1492 timekeep_push_vdso();
1493 }
1494
1495 /* Report or change the active timecounter hardware. */
1496 static int
1497 sysctl_kern_timecounter_hardware(SYSCTL_HANDLER_ARGS)
1498 {
1499 char newname[32];
1500 struct timecounter *newtc, *tc;
1501 int error;
1502
1503 tc = timecounter;
1504 strlcpy(newname, tc->tc_name, sizeof(newname));
1505
1506 error = sysctl_handle_string(oidp, &newname[0], sizeof(newname), req);
1507 if (error != 0 || req->newptr == NULL)
1508 return (error);
1509 /* Record that the tc in use now was specifically chosen. */
1510 tc_chosen = 1;
1511 if (strcmp(newname, tc->tc_name) == 0)
1512 return (0);
1513 for (newtc = timecounters; newtc != NULL; newtc = newtc->tc_next) {
1514 if (strcmp(newname, newtc->tc_name) != 0)
1515 continue;
1516
1517 /* Warm up new timecounter. */
1518 (void)newtc->tc_get_timecount(newtc);
1519 (void)newtc->tc_get_timecount(newtc);
1520
1521 timecounter = newtc;
1522
1523 /*
1524 * The vdso timehands update is deferred until the next
1525 * 'tc_windup()'.
1526 *
1527 * This is prudent given that 'timekeep_push_vdso()' does not
1528 * use any locking and that it can be called in hard interrupt
1529 * context via 'tc_windup()'.
1530 */
1531 return (0);
1532 }
1533 return (EINVAL);
1534 }
1535
1536 SYSCTL_PROC(_kern_timecounter, OID_AUTO, hardware, CTLTYPE_STRING | CTLFLAG_RW,
1537 0, 0, sysctl_kern_timecounter_hardware, "A",
1538 "Timecounter hardware selected");
1539
1540
1541 /* Report the available timecounter hardware. */
1542 static int
1543 sysctl_kern_timecounter_choice(SYSCTL_HANDLER_ARGS)
1544 {
1545 struct sbuf sb;
1546 struct timecounter *tc;
1547 int error;
1548
1549 sbuf_new_for_sysctl(&sb, NULL, 0, req);
1550 for (tc = timecounters; tc != NULL; tc = tc->tc_next) {
1551 if (tc != timecounters)
1552 sbuf_putc(&sb, ' ');
1553 sbuf_printf(&sb, "%s(%d)", tc->tc_name, tc->tc_quality);
1554 }
1555 error = sbuf_finish(&sb);
1556 sbuf_delete(&sb);
1557 return (error);
1558 }
1559
1560 SYSCTL_PROC(_kern_timecounter, OID_AUTO, choice, CTLTYPE_STRING | CTLFLAG_RD,
1561 0, 0, sysctl_kern_timecounter_choice, "A", "Timecounter hardware detected");
1562
1563 /*
1564 * RFC 2783 PPS-API implementation.
1565 */
1566
1567 /*
1568 * Return true if the driver is aware of the abi version extensions in the
1569 * pps_state structure, and it supports at least the given abi version number.
1570 */
1571 static inline int
1572 abi_aware(struct pps_state *pps, int vers)
1573 {
1574
1575 return ((pps->kcmode & KCMODE_ABIFLAG) && pps->driver_abi >= vers);
1576 }
1577
1578 static int
1579 pps_fetch(struct pps_fetch_args *fapi, struct pps_state *pps)
1580 {
1581 int err, timo;
1582 pps_seq_t aseq, cseq;
1583 struct timeval tv;
1584
1585 if (fapi->tsformat && fapi->tsformat != PPS_TSFMT_TSPEC)
1586 return (EINVAL);
1587
1588 /*
1589 * If no timeout is requested, immediately return whatever values were
1590 * most recently captured. If timeout seconds is -1, that's a request
1591 * to block without a timeout. WITNESS won't let us sleep forever
1592 * without a lock (we really don't need a lock), so just repeatedly
1593 * sleep a long time.
1594 */
1595 if (fapi->timeout.tv_sec || fapi->timeout.tv_nsec) {
1596 if (fapi->timeout.tv_sec == -1)
1597 timo = 0x7fffffff;
1598 else {
1599 tv.tv_sec = fapi->timeout.tv_sec;
1600 tv.tv_usec = fapi->timeout.tv_nsec / 1000;
1601 timo = tvtohz(&tv);
1602 }
1603 aseq = atomic_load_int(&pps->ppsinfo.assert_sequence);
1604 cseq = atomic_load_int(&pps->ppsinfo.clear_sequence);
1605 while (aseq == atomic_load_int(&pps->ppsinfo.assert_sequence) &&
1606 cseq == atomic_load_int(&pps->ppsinfo.clear_sequence)) {
1607 if (abi_aware(pps, 1) && pps->driver_mtx != NULL) {
1608 if (pps->flags & PPSFLAG_MTX_SPIN) {
1609 err = msleep_spin(pps, pps->driver_mtx,
1610 "ppsfch", timo);
1611 } else {
1612 err = msleep(pps, pps->driver_mtx, PCATCH,
1613 "ppsfch", timo);
1614 }
1615 } else {
1616 err = tsleep(pps, PCATCH, "ppsfch", timo);
1617 }
1618 if (err == EWOULDBLOCK) {
1619 if (fapi->timeout.tv_sec == -1) {
1620 continue;
1621 } else {
1622 return (ETIMEDOUT);
1623 }
1624 } else if (err != 0) {
1625 return (err);
1626 }
1627 }
1628 }
1629
1630 pps->ppsinfo.current_mode = pps->ppsparam.mode;
1631 fapi->pps_info_buf = pps->ppsinfo;
1632
1633 return (0);
1634 }
1635
1636 int
1637 pps_ioctl(u_long cmd, caddr_t data, struct pps_state *pps)
1638 {
1639 pps_params_t *app;
1640 struct pps_fetch_args *fapi;
1641 #ifdef FFCLOCK
1642 struct pps_fetch_ffc_args *fapi_ffc;
1643 #endif
1644 #ifdef PPS_SYNC
1645 struct pps_kcbind_args *kapi;
1646 #endif
1647
1648 KASSERT(pps != NULL, ("NULL pps pointer in pps_ioctl"));
1649 switch (cmd) {
1650 case PPS_IOC_CREATE:
1651 return (0);
1652 case PPS_IOC_DESTROY:
1653 return (0);
1654 case PPS_IOC_SETPARAMS:
1655 app = (pps_params_t *)data;
1656 if (app->mode & ~pps->ppscap)
1657 return (EINVAL);
1658 #ifdef FFCLOCK
1659 /* Ensure only a single clock is selected for ffc timestamp. */
1660 if ((app->mode & PPS_TSCLK_MASK) == PPS_TSCLK_MASK)
1661 return (EINVAL);
1662 #endif
1663 pps->ppsparam = *app;
1664 return (0);
1665 case PPS_IOC_GETPARAMS:
1666 app = (pps_params_t *)data;
1667 *app = pps->ppsparam;
1668 app->api_version = PPS_API_VERS_1;
1669 return (0);
1670 case PPS_IOC_GETCAP:
1671 *(int*)data = pps->ppscap;
1672 return (0);
1673 case PPS_IOC_FETCH:
1674 fapi = (struct pps_fetch_args *)data;
1675 return (pps_fetch(fapi, pps));
1676 #ifdef FFCLOCK
1677 case PPS_IOC_FETCH_FFCOUNTER:
1678 fapi_ffc = (struct pps_fetch_ffc_args *)data;
1679 if (fapi_ffc->tsformat && fapi_ffc->tsformat !=
1680 PPS_TSFMT_TSPEC)
1681 return (EINVAL);
1682 if (fapi_ffc->timeout.tv_sec || fapi_ffc->timeout.tv_nsec)
1683 return (EOPNOTSUPP);
1684 pps->ppsinfo_ffc.current_mode = pps->ppsparam.mode;
1685 fapi_ffc->pps_info_buf_ffc = pps->ppsinfo_ffc;
1686 /* Overwrite timestamps if feedback clock selected. */
1687 switch (pps->ppsparam.mode & PPS_TSCLK_MASK) {
1688 case PPS_TSCLK_FBCK:
1689 fapi_ffc->pps_info_buf_ffc.assert_timestamp =
1690 pps->ppsinfo.assert_timestamp;
1691 fapi_ffc->pps_info_buf_ffc.clear_timestamp =
1692 pps->ppsinfo.clear_timestamp;
1693 break;
1694 case PPS_TSCLK_FFWD:
1695 break;
1696 default:
1697 break;
1698 }
1699 return (0);
1700 #endif /* FFCLOCK */
1701 case PPS_IOC_KCBIND:
1702 #ifdef PPS_SYNC
1703 kapi = (struct pps_kcbind_args *)data;
1704 /* XXX Only root should be able to do this */
1705 if (kapi->tsformat && kapi->tsformat != PPS_TSFMT_TSPEC)
1706 return (EINVAL);
1707 if (kapi->kernel_consumer != PPS_KC_HARDPPS)
1708 return (EINVAL);
1709 if (kapi->edge & ~pps->ppscap)
1710 return (EINVAL);
1711 pps->kcmode = (kapi->edge & KCMODE_EDGEMASK) |
1712 (pps->kcmode & KCMODE_ABIFLAG);
1713 return (0);
1714 #else
1715 return (EOPNOTSUPP);
1716 #endif
1717 default:
1718 return (ENOIOCTL);
1719 }
1720 }
1721
1722 void
1723 pps_init(struct pps_state *pps)
1724 {
1725 pps->ppscap |= PPS_TSFMT_TSPEC | PPS_CANWAIT;
1726 if (pps->ppscap & PPS_CAPTUREASSERT)
1727 pps->ppscap |= PPS_OFFSETASSERT;
1728 if (pps->ppscap & PPS_CAPTURECLEAR)
1729 pps->ppscap |= PPS_OFFSETCLEAR;
1730 #ifdef FFCLOCK
1731 pps->ppscap |= PPS_TSCLK_MASK;
1732 #endif
1733 pps->kcmode &= ~KCMODE_ABIFLAG;
1734 }
1735
1736 void
1737 pps_init_abi(struct pps_state *pps)
1738 {
1739
1740 pps_init(pps);
1741 if (pps->driver_abi > 0) {
1742 pps->kcmode |= KCMODE_ABIFLAG;
1743 pps->kernel_abi = PPS_ABI_VERSION;
1744 }
1745 }
1746
1747 void
1748 pps_capture(struct pps_state *pps)
1749 {
1750 struct timehands *th;
1751
1752 KASSERT(pps != NULL, ("NULL pps pointer in pps_capture"));
1753 th = timehands;
1754 pps->capgen = atomic_load_acq_int(&th->th_generation);
1755 pps->capth = th;
1756 #ifdef FFCLOCK
1757 pps->capffth = fftimehands;
1758 #endif
1759 pps->capcount = th->th_counter->tc_get_timecount(th->th_counter);
1760 atomic_thread_fence_acq();
1761 if (pps->capgen != th->th_generation)
1762 pps->capgen = 0;
1763 }
1764
1765 void
1766 pps_event(struct pps_state *pps, int event)
1767 {
1768 struct bintime bt;
1769 struct timespec ts, *tsp, *osp;
1770 u_int tcount, *pcount;
1771 int foff;
1772 pps_seq_t *pseq;
1773 #ifdef FFCLOCK
1774 struct timespec *tsp_ffc;
1775 pps_seq_t *pseq_ffc;
1776 ffcounter *ffcount;
1777 #endif
1778 #ifdef PPS_SYNC
1779 int fhard;
1780 #endif
1781
1782 KASSERT(pps != NULL, ("NULL pps pointer in pps_event"));
1783 /* Nothing to do if not currently set to capture this event type. */
1784 if ((event & pps->ppsparam.mode) == 0)
1785 return;
1786 /* If the timecounter was wound up underneath us, bail out. */
1787 if (pps->capgen == 0 || pps->capgen !=
1788 atomic_load_acq_int(&pps->capth->th_generation))
1789 return;
1790
1791 /* Things would be easier with arrays. */
1792 if (event == PPS_CAPTUREASSERT) {
1793 tsp = &pps->ppsinfo.assert_timestamp;
1794 osp = &pps->ppsparam.assert_offset;
1795 foff = pps->ppsparam.mode & PPS_OFFSETASSERT;
1796 #ifdef PPS_SYNC
1797 fhard = pps->kcmode & PPS_CAPTUREASSERT;
1798 #endif
1799 pcount = &pps->ppscount[0];
1800 pseq = &pps->ppsinfo.assert_sequence;
1801 #ifdef FFCLOCK
1802 ffcount = &pps->ppsinfo_ffc.assert_ffcount;
1803 tsp_ffc = &pps->ppsinfo_ffc.assert_timestamp;
1804 pseq_ffc = &pps->ppsinfo_ffc.assert_sequence;
1805 #endif
1806 } else {
1807 tsp = &pps->ppsinfo.clear_timestamp;
1808 osp = &pps->ppsparam.clear_offset;
1809 foff = pps->ppsparam.mode & PPS_OFFSETCLEAR;
1810 #ifdef PPS_SYNC
1811 fhard = pps->kcmode & PPS_CAPTURECLEAR;
1812 #endif
1813 pcount = &pps->ppscount[1];
1814 pseq = &pps->ppsinfo.clear_sequence;
1815 #ifdef FFCLOCK
1816 ffcount = &pps->ppsinfo_ffc.clear_ffcount;
1817 tsp_ffc = &pps->ppsinfo_ffc.clear_timestamp;
1818 pseq_ffc = &pps->ppsinfo_ffc.clear_sequence;
1819 #endif
1820 }
1821
1822 /*
1823 * If the timecounter changed, we cannot compare the count values, so
1824 * we have to drop the rest of the PPS-stuff until the next event.
1825 */
1826 if (pps->ppstc != pps->capth->th_counter) {
1827 pps->ppstc = pps->capth->th_counter;
1828 *pcount = pps->capcount;
1829 pps->ppscount[2] = pps->capcount;
1830 return;
1831 }
1832
1833 /* Convert the count to a timespec. */
1834 tcount = pps->capcount - pps->capth->th_offset_count;
1835 tcount &= pps->capth->th_counter->tc_counter_mask;
1836 bt = pps->capth->th_bintime;
1837 bintime_addx(&bt, pps->capth->th_scale * tcount);
1838 bintime2timespec(&bt, &ts);
1839
1840 /* If the timecounter was wound up underneath us, bail out. */
1841 atomic_thread_fence_acq();
1842 if (pps->capgen != pps->capth->th_generation)
1843 return;
1844
1845 *pcount = pps->capcount;
1846 (*pseq)++;
1847 *tsp = ts;
1848
1849 if (foff) {
1850 timespecadd(tsp, osp, tsp);
1851 if (tsp->tv_nsec < 0) {
1852 tsp->tv_nsec += 1000000000;
1853 tsp->tv_sec -= 1;
1854 }
1855 }
1856
1857 #ifdef FFCLOCK
1858 *ffcount = pps->capffth->tick_ffcount + tcount;
1859 bt = pps->capffth->tick_time;
1860 ffclock_convert_delta(tcount, pps->capffth->cest.period, &bt);
1861 bintime_add(&bt, &pps->capffth->tick_time);
1862 bintime2timespec(&bt, &ts);
1863 (*pseq_ffc)++;
1864 *tsp_ffc = ts;
1865 #endif
1866
1867 #ifdef PPS_SYNC
1868 if (fhard) {
1869 uint64_t scale;
1870
1871 /*
1872 * Feed the NTP PLL/FLL.
1873 * The FLL wants to know how many (hardware) nanoseconds
1874 * elapsed since the previous event.
1875 */
1876 tcount = pps->capcount - pps->ppscount[2];
1877 pps->ppscount[2] = pps->capcount;
1878 tcount &= pps->capth->th_counter->tc_counter_mask;
1879 scale = (uint64_t)1 << 63;
1880 scale /= pps->capth->th_counter->tc_frequency;
1881 scale *= 2;
1882 bt.sec = 0;
1883 bt.frac = 0;
1884 bintime_addx(&bt, scale * tcount);
1885 bintime2timespec(&bt, &ts);
1886 hardpps(tsp, ts.tv_nsec + 1000000000 * ts.tv_sec);
1887 }
1888 #endif
1889
1890 /* Wakeup anyone sleeping in pps_fetch(). */
1891 wakeup(pps);
1892 }
1893
1894 /*
1895 * Timecounters need to be updated every so often to prevent the hardware
1896 * counter from overflowing. Updating also recalculates the cached values
1897 * used by the get*() family of functions, so their precision depends on
1898 * the update frequency.
1899 */
1900
1901 static int tc_tick;
1902 SYSCTL_INT(_kern_timecounter, OID_AUTO, tick, CTLFLAG_RD, &tc_tick, 0,
1903 "Approximate number of hardclock ticks in a millisecond");
1904
1905 void
1906 tc_ticktock(int cnt)
1907 {
1908 static int count;
1909
1910 if (mtx_trylock_spin(&tc_setclock_mtx)) {
1911 count += cnt;
1912 if (count >= tc_tick) {
1913 count = 0;
1914 tc_windup(NULL);
1915 }
1916 mtx_unlock_spin(&tc_setclock_mtx);
1917 }
1918 }
1919
1920 static void __inline
1921 tc_adjprecision(void)
1922 {
1923 int t;
1924
1925 if (tc_timepercentage > 0) {
1926 t = (99 + tc_timepercentage) / tc_timepercentage;
1927 tc_precexp = fls(t + (t >> 1)) - 1;
1928 FREQ2BT(hz / tc_tick, &bt_timethreshold);
1929 FREQ2BT(hz, &bt_tickthreshold);
1930 bintime_shift(&bt_timethreshold, tc_precexp);
1931 bintime_shift(&bt_tickthreshold, tc_precexp);
1932 } else {
1933 tc_precexp = 31;
1934 bt_timethreshold.sec = INT_MAX;
1935 bt_timethreshold.frac = ~(uint64_t)0;
1936 bt_tickthreshold = bt_timethreshold;
1937 }
1938 sbt_timethreshold = bttosbt(bt_timethreshold);
1939 sbt_tickthreshold = bttosbt(bt_tickthreshold);
1940 }
1941
1942 static int
1943 sysctl_kern_timecounter_adjprecision(SYSCTL_HANDLER_ARGS)
1944 {
1945 int error, val;
1946
1947 val = tc_timepercentage;
1948 error = sysctl_handle_int(oidp, &val, 0, req);
1949 if (error != 0 || req->newptr == NULL)
1950 return (error);
1951 tc_timepercentage = val;
1952 if (cold)
1953 goto done;
1954 tc_adjprecision();
1955 done:
1956 return (0);
1957 }
1958
1959 static void
1960 inittimecounter(void *dummy)
1961 {
1962 u_int p;
1963 int tick_rate;
1964
1965 /*
1966 * Set the initial timeout to
1967 * max(1, <approx. number of hardclock ticks in a millisecond>).
1968 * People should probably not use the sysctl to set the timeout
1969 * to smaller than its initial value, since that value is the
1970 * smallest reasonable one. If they want better timestamps they
1971 * should use the non-"get"* functions.
1972 */
1973 if (hz > 1000)
1974 tc_tick = (hz + 500) / 1000;
1975 else
1976 tc_tick = 1;
1977 tc_adjprecision();
1978 FREQ2BT(hz, &tick_bt);
1979 tick_sbt = bttosbt(tick_bt);
1980 tick_rate = hz / tc_tick;
1981 FREQ2BT(tick_rate, &tc_tick_bt);
1982 tc_tick_sbt = bttosbt(tc_tick_bt);
1983 p = (tc_tick * 1000000) / hz;
1984 printf("Timecounters tick every %d.%03u msec\n", p / 1000, p % 1000);
1985
1986 #ifdef FFCLOCK
1987 ffclock_init();
1988 #endif
1989 /* warm up new timecounter (again) and get rolling. */
1990 (void)timecounter->tc_get_timecount(timecounter);
1991 (void)timecounter->tc_get_timecount(timecounter);
1992 mtx_lock_spin(&tc_setclock_mtx);
1993 tc_windup(NULL);
1994 mtx_unlock_spin(&tc_setclock_mtx);
1995 }
1996
1997 SYSINIT(timecounter, SI_SUB_CLOCKS, SI_ORDER_SECOND, inittimecounter, NULL);
1998
1999 /* Cpu tick handling -------------------------------------------------*/
2000
2001 static int cpu_tick_variable;
2002 static uint64_t cpu_tick_frequency;
2003
2004 DPCPU_DEFINE_STATIC(uint64_t, tc_cpu_ticks_base);
2005 DPCPU_DEFINE_STATIC(unsigned, tc_cpu_ticks_last);
2006
2007 static uint64_t
2008 tc_cpu_ticks(void)
2009 {
2010 struct timecounter *tc;
2011 uint64_t res, *base;
2012 unsigned u, *last;
2013
2014 critical_enter();
2015 base = DPCPU_PTR(tc_cpu_ticks_base);
2016 last = DPCPU_PTR(tc_cpu_ticks_last);
2017 tc = timehands->th_counter;
2018 u = tc->tc_get_timecount(tc) & tc->tc_counter_mask;
2019 if (u < *last)
2020 *base += (uint64_t)tc->tc_counter_mask + 1;
2021 *last = u;
2022 res = u + *base;
2023 critical_exit();
2024 return (res);
2025 }
2026
2027 void
2028 cpu_tick_calibration(void)
2029 {
2030 static time_t last_calib;
2031
2032 if (time_uptime != last_calib && !(time_uptime & 0xf)) {
2033 cpu_tick_calibrate(0);
2034 last_calib = time_uptime;
2035 }
2036 }
2037
2038 /*
2039 * This function gets called every 16 seconds on only one designated
2040 * CPU in the system from hardclock() via cpu_tick_calibration()().
2041 *
2042 * Whenever the real time clock is stepped we get called with reset=1
2043 * to make sure we handle suspend/resume and similar events correctly.
2044 */
2045
2046 static void
2047 cpu_tick_calibrate(int reset)
2048 {
2049 static uint64_t c_last;
2050 uint64_t c_this, c_delta;
2051 static struct bintime t_last;
2052 struct bintime t_this, t_delta;
2053 uint32_t divi;
2054
2055 if (reset) {
2056 /* The clock was stepped, abort & reset */
2057 t_last.sec = 0;
2058 return;
2059 }
2060
2061 /* we don't calibrate fixed rate cputicks */
2062 if (!cpu_tick_variable)
2063 return;
2064
2065 getbinuptime(&t_this);
2066 c_this = cpu_ticks();
2067 if (t_last.sec != 0) {
2068 c_delta = c_this - c_last;
2069 t_delta = t_this;
2070 bintime_sub(&t_delta, &t_last);
2071 /*
2072 * Headroom:
2073 * 2^(64-20) / 16[s] =
2074 * 2^(44) / 16[s] =
2075 * 17.592.186.044.416 / 16 =
2076 * 1.099.511.627.776 [Hz]
2077 */
2078 divi = t_delta.sec << 20;
2079 divi |= t_delta.frac >> (64 - 20);
2080 c_delta <<= 20;
2081 c_delta /= divi;
2082 if (c_delta > cpu_tick_frequency) {
2083 if (0 && bootverbose)
2084 printf("cpu_tick increased to %ju Hz\n",
2085 c_delta);
2086 cpu_tick_frequency = c_delta;
2087 }
2088 }
2089 c_last = c_this;
2090 t_last = t_this;
2091 }
2092
2093 void
2094 set_cputicker(cpu_tick_f *func, uint64_t freq, unsigned var)
2095 {
2096
2097 if (func == NULL) {
2098 cpu_ticks = tc_cpu_ticks;
2099 } else {
2100 cpu_tick_frequency = freq;
2101 cpu_tick_variable = var;
2102 cpu_ticks = func;
2103 }
2104 }
2105
2106 uint64_t
2107 cpu_tickrate(void)
2108 {
2109
2110 if (cpu_ticks == tc_cpu_ticks)
2111 return (tc_getfrequency());
2112 return (cpu_tick_frequency);
2113 }
2114
2115 /*
2116 * We need to be slightly careful converting cputicks to microseconds.
2117 * There is plenty of margin in 64 bits of microseconds (half a million
2118 * years) and in 64 bits at 4 GHz (146 years), but if we do a multiply
2119 * before divide conversion (to retain precision) we find that the
2120 * margin shrinks to 1.5 hours (one millionth of 146y).
2121 * With a three prong approach we never lose significant bits, no
2122 * matter what the cputick rate and length of timeinterval is.
2123 */
2124
2125 uint64_t
2126 cputick2usec(uint64_t tick)
2127 {
2128
2129 if (tick > 18446744073709551LL) /* floor(2^64 / 1000) */
2130 return (tick / (cpu_tickrate() / 1000000LL));
2131 else if (tick > 18446744073709LL) /* floor(2^64 / 1000000) */
2132 return ((tick * 1000LL) / (cpu_tickrate() / 1000LL));
2133 else
2134 return ((tick * 1000000LL) / cpu_tickrate());
2135 }
2136
2137 cpu_tick_f *cpu_ticks = tc_cpu_ticks;
2138
2139 static int vdso_th_enable = 1;
2140 static int
2141 sysctl_fast_gettime(SYSCTL_HANDLER_ARGS)
2142 {
2143 int old_vdso_th_enable, error;
2144
2145 old_vdso_th_enable = vdso_th_enable;
2146 error = sysctl_handle_int(oidp, &old_vdso_th_enable, 0, req);
2147 if (error != 0)
2148 return (error);
2149 vdso_th_enable = old_vdso_th_enable;
2150 return (0);
2151 }
2152 SYSCTL_PROC(_kern_timecounter, OID_AUTO, fast_gettime,
2153 CTLTYPE_INT | CTLFLAG_RW | CTLFLAG_MPSAFE,
2154 NULL, 0, sysctl_fast_gettime, "I", "Enable fast time of day");
2155
2156 uint32_t
2157 tc_fill_vdso_timehands(struct vdso_timehands *vdso_th)
2158 {
2159 struct timehands *th;
2160 uint32_t enabled;
2161
2162 th = timehands;
2163 vdso_th->th_scale = th->th_scale;
2164 vdso_th->th_offset_count = th->th_offset_count;
2165 vdso_th->th_counter_mask = th->th_counter->tc_counter_mask;
2166 vdso_th->th_offset = th->th_offset;
2167 vdso_th->th_boottime = th->th_boottime;
2168 if (th->th_counter->tc_fill_vdso_timehands != NULL) {
2169 enabled = th->th_counter->tc_fill_vdso_timehands(vdso_th,
2170 th->th_counter);
2171 } else
2172 enabled = 0;
2173 if (!vdso_th_enable)
2174 enabled = 0;
2175 return (enabled);
2176 }
2177
2178 #ifdef COMPAT_FREEBSD32
2179 uint32_t
2180 tc_fill_vdso_timehands32(struct vdso_timehands32 *vdso_th32)
2181 {
2182 struct timehands *th;
2183 uint32_t enabled;
2184
2185 th = timehands;
2186 *(uint64_t *)&vdso_th32->th_scale[0] = th->th_scale;
2187 vdso_th32->th_offset_count = th->th_offset_count;
2188 vdso_th32->th_counter_mask = th->th_counter->tc_counter_mask;
2189 vdso_th32->th_offset.sec = th->th_offset.sec;
2190 *(uint64_t *)&vdso_th32->th_offset.frac[0] = th->th_offset.frac;
2191 vdso_th32->th_boottime.sec = th->th_boottime.sec;
2192 *(uint64_t *)&vdso_th32->th_boottime.frac[0] = th->th_boottime.frac;
2193 if (th->th_counter->tc_fill_vdso_timehands32 != NULL) {
2194 enabled = th->th_counter->tc_fill_vdso_timehands32(vdso_th32,
2195 th->th_counter);
2196 } else
2197 enabled = 0;
2198 if (!vdso_th_enable)
2199 enabled = 0;
2200 return (enabled);
2201 }
2202 #endif
Cache object: e6bcbbf8d27f6d2c63bfa8394e6dc728
|