2 * SPDX-License-Identifier: Beerware
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 * ----------------------------------------------------------------------------
11 * Copyright (c) 2011, 2015, 2016 The FreeBSD Foundation
12 * All rights reserved.
14 * Portions of this software were developed by Julien Ridoux at the University
15 * of Melbourne under sponsorship from the FreeBSD Foundation.
17 * Portions of this software were developed by Konstantin Belousov
18 * under sponsorship from the FreeBSD Foundation.
21 #include <sys/cdefs.h>
22 __FBSDID("$FreeBSD$");
25 #include "opt_ffclock.h"
27 #include <sys/param.h>
28 #include <sys/kernel.h>
29 #include <sys/limits.h>
31 #include <sys/mutex.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>
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.
50 #define LARGE_STEP 200
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
59 dummy_get_timecount(struct timecounter *tc)
66 static struct timecounter dummy_timecounter = {
67 dummy_get_timecount, 0, ~0u, 1000000, "dummy", -1000000
71 /* These fields must be initialized by the driver. */
72 struct timecounter *th_counter;
73 int64_t th_adjustment;
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. */
83 struct timehands *th_next;
86 static struct timehands ths[16] = {
88 .th_counter = &dummy_timecounter,
89 .th_scale = (uint64_t)-1 / 1000000,
90 .th_offset = { .sec = 1 },
95 static struct timehands *volatile timehands = &ths[0];
96 struct timecounter *timecounter = &dummy_timecounter;
97 static struct timecounter *timecounters = &dummy_timecounter;
99 int tc_min_ticktock_freq = 1;
101 volatile time_t time_second = 1;
102 volatile time_t time_uptime = 1;
104 static int sysctl_kern_boottime(SYSCTL_HANDLER_ARGS);
105 SYSCTL_PROC(_kern, KERN_BOOTTIME, boottime, CTLTYPE_STRUCT|CTLFLAG_RD,
106 NULL, 0, sysctl_kern_boottime, "S,timeval", "System boottime");
108 SYSCTL_NODE(_kern, OID_AUTO, timecounter, CTLFLAG_RW, 0, "");
109 static SYSCTL_NODE(_kern_timecounter, OID_AUTO, tc, CTLFLAG_RW, 0, "");
111 static int timestepwarnings;
112 SYSCTL_INT(_kern_timecounter, OID_AUTO, stepwarnings, CTLFLAG_RW,
113 ×tepwarnings, 0, "Log time steps");
115 static int timehands_count = 2;
116 SYSCTL_INT(_kern_timecounter, OID_AUTO, timehands_count, CTLFLAG_RDTUN,
117 &timehands_count, 0, "Count of timehands in rotation");
119 struct bintime bt_timethreshold;
120 struct bintime bt_tickthreshold;
121 sbintime_t sbt_timethreshold;
122 sbintime_t sbt_tickthreshold;
123 struct bintime tc_tick_bt;
124 sbintime_t tc_tick_sbt;
126 int tc_timepercentage = TC_DEFAULTPERC;
127 static int sysctl_kern_timecounter_adjprecision(SYSCTL_HANDLER_ARGS);
128 SYSCTL_PROC(_kern_timecounter, OID_AUTO, alloweddeviation,
129 CTLTYPE_INT | CTLFLAG_RWTUN | CTLFLAG_MPSAFE, 0, 0,
130 sysctl_kern_timecounter_adjprecision, "I",
131 "Allowed time interval deviation in percents");
133 volatile int rtc_generation = 1;
135 static int tc_chosen; /* Non-zero if a specific tc was chosen via sysctl. */
137 static void tc_windup(struct bintime *new_boottimebin);
138 static void cpu_tick_calibrate(int);
140 void dtrace_getnanotime(struct timespec *tsp);
143 sysctl_kern_boottime(SYSCTL_HANDLER_ARGS)
145 struct timeval boottime;
147 getboottime(&boottime);
149 /* i386 is the only arch which uses a 32bits time_t */
154 if (req->flags & SCTL_MASK32) {
155 tv[0] = boottime.tv_sec;
156 tv[1] = boottime.tv_usec;
157 return (SYSCTL_OUT(req, tv, sizeof(tv)));
161 return (SYSCTL_OUT(req, &boottime, sizeof(boottime)));
165 sysctl_kern_timecounter_get(SYSCTL_HANDLER_ARGS)
168 struct timecounter *tc = arg1;
170 ncount = tc->tc_get_timecount(tc);
171 return (sysctl_handle_int(oidp, &ncount, 0, req));
175 sysctl_kern_timecounter_freq(SYSCTL_HANDLER_ARGS)
178 struct timecounter *tc = arg1;
180 freq = tc->tc_frequency;
181 return (sysctl_handle_64(oidp, &freq, 0, req));
185 * Return the difference between the timehands' counter value now and what
186 * was when we copied it to the timehands' offset_count.
188 static __inline u_int
189 tc_delta(struct timehands *th)
191 struct timecounter *tc;
194 return ((tc->tc_get_timecount(tc) - th->th_offset_count) &
195 tc->tc_counter_mask);
199 * Functions for reading the time. We have to loop until we are sure that
200 * the timehands that we operated on was not updated under our feet. See
201 * the comment in <sys/time.h> for a description of these 12 functions.
206 fbclock_binuptime(struct bintime *bt)
208 struct timehands *th;
213 gen = atomic_load_acq_int(&th->th_generation);
215 bintime_addx(bt, th->th_scale * tc_delta(th));
216 atomic_thread_fence_acq();
217 } while (gen == 0 || gen != th->th_generation);
221 fbclock_nanouptime(struct timespec *tsp)
225 fbclock_binuptime(&bt);
226 bintime2timespec(&bt, tsp);
230 fbclock_microuptime(struct timeval *tvp)
234 fbclock_binuptime(&bt);
235 bintime2timeval(&bt, tvp);
239 fbclock_bintime(struct bintime *bt)
241 struct timehands *th;
246 gen = atomic_load_acq_int(&th->th_generation);
247 *bt = th->th_bintime;
248 bintime_addx(bt, th->th_scale * tc_delta(th));
249 atomic_thread_fence_acq();
250 } while (gen == 0 || gen != th->th_generation);
254 fbclock_nanotime(struct timespec *tsp)
258 fbclock_bintime(&bt);
259 bintime2timespec(&bt, tsp);
263 fbclock_microtime(struct timeval *tvp)
267 fbclock_bintime(&bt);
268 bintime2timeval(&bt, tvp);
272 fbclock_getbinuptime(struct bintime *bt)
274 struct timehands *th;
279 gen = atomic_load_acq_int(&th->th_generation);
281 atomic_thread_fence_acq();
282 } while (gen == 0 || gen != th->th_generation);
286 fbclock_getnanouptime(struct timespec *tsp)
288 struct timehands *th;
293 gen = atomic_load_acq_int(&th->th_generation);
294 bintime2timespec(&th->th_offset, tsp);
295 atomic_thread_fence_acq();
296 } while (gen == 0 || gen != th->th_generation);
300 fbclock_getmicrouptime(struct timeval *tvp)
302 struct timehands *th;
307 gen = atomic_load_acq_int(&th->th_generation);
308 bintime2timeval(&th->th_offset, tvp);
309 atomic_thread_fence_acq();
310 } while (gen == 0 || gen != th->th_generation);
314 fbclock_getbintime(struct bintime *bt)
316 struct timehands *th;
321 gen = atomic_load_acq_int(&th->th_generation);
322 *bt = th->th_bintime;
323 atomic_thread_fence_acq();
324 } while (gen == 0 || gen != th->th_generation);
328 fbclock_getnanotime(struct timespec *tsp)
330 struct timehands *th;
335 gen = atomic_load_acq_int(&th->th_generation);
336 *tsp = th->th_nanotime;
337 atomic_thread_fence_acq();
338 } while (gen == 0 || gen != th->th_generation);
342 fbclock_getmicrotime(struct timeval *tvp)
344 struct timehands *th;
349 gen = atomic_load_acq_int(&th->th_generation);
350 *tvp = th->th_microtime;
351 atomic_thread_fence_acq();
352 } while (gen == 0 || gen != th->th_generation);
356 binuptime(struct bintime *bt)
358 struct timehands *th;
363 gen = atomic_load_acq_int(&th->th_generation);
365 bintime_addx(bt, th->th_scale * tc_delta(th));
366 atomic_thread_fence_acq();
367 } while (gen == 0 || gen != th->th_generation);
371 nanouptime(struct timespec *tsp)
376 bintime2timespec(&bt, tsp);
380 microuptime(struct timeval *tvp)
385 bintime2timeval(&bt, tvp);
389 bintime(struct bintime *bt)
391 struct timehands *th;
396 gen = atomic_load_acq_int(&th->th_generation);
397 *bt = th->th_bintime;
398 bintime_addx(bt, th->th_scale * tc_delta(th));
399 atomic_thread_fence_acq();
400 } while (gen == 0 || gen != th->th_generation);
404 nanotime(struct timespec *tsp)
409 bintime2timespec(&bt, tsp);
413 microtime(struct timeval *tvp)
418 bintime2timeval(&bt, tvp);
422 getbinuptime(struct bintime *bt)
424 struct timehands *th;
429 gen = atomic_load_acq_int(&th->th_generation);
431 atomic_thread_fence_acq();
432 } while (gen == 0 || gen != th->th_generation);
436 getnanouptime(struct timespec *tsp)
438 struct timehands *th;
443 gen = atomic_load_acq_int(&th->th_generation);
444 bintime2timespec(&th->th_offset, tsp);
445 atomic_thread_fence_acq();
446 } while (gen == 0 || gen != th->th_generation);
450 getmicrouptime(struct timeval *tvp)
452 struct timehands *th;
457 gen = atomic_load_acq_int(&th->th_generation);
458 bintime2timeval(&th->th_offset, tvp);
459 atomic_thread_fence_acq();
460 } while (gen == 0 || gen != th->th_generation);
464 getbintime(struct bintime *bt)
466 struct timehands *th;
471 gen = atomic_load_acq_int(&th->th_generation);
472 *bt = th->th_bintime;
473 atomic_thread_fence_acq();
474 } while (gen == 0 || gen != th->th_generation);
478 getnanotime(struct timespec *tsp)
480 struct timehands *th;
485 gen = atomic_load_acq_int(&th->th_generation);
486 *tsp = th->th_nanotime;
487 atomic_thread_fence_acq();
488 } while (gen == 0 || gen != th->th_generation);
492 getmicrotime(struct timeval *tvp)
494 struct timehands *th;
499 gen = atomic_load_acq_int(&th->th_generation);
500 *tvp = th->th_microtime;
501 atomic_thread_fence_acq();
502 } while (gen == 0 || gen != th->th_generation);
507 getboottime(struct timeval *boottime)
509 struct bintime boottimebin;
511 getboottimebin(&boottimebin);
512 bintime2timeval(&boottimebin, boottime);
516 getboottimebin(struct bintime *boottimebin)
518 struct timehands *th;
523 gen = atomic_load_acq_int(&th->th_generation);
524 *boottimebin = th->th_boottime;
525 atomic_thread_fence_acq();
526 } while (gen == 0 || gen != th->th_generation);
531 * Support for feed-forward synchronization algorithms. This is heavily inspired
532 * by the timehands mechanism but kept independent from it. *_windup() functions
533 * have some connection to avoid accessing the timecounter hardware more than
537 /* Feed-forward clock estimates kept updated by the synchronization daemon. */
538 struct ffclock_estimate ffclock_estimate;
539 struct bintime ffclock_boottime; /* Feed-forward boot time estimate. */
540 uint32_t ffclock_status; /* Feed-forward clock status. */
541 int8_t ffclock_updated; /* New estimates are available. */
542 struct mtx ffclock_mtx; /* Mutex on ffclock_estimate. */
545 struct ffclock_estimate cest;
546 struct bintime tick_time;
547 struct bintime tick_time_lerp;
548 ffcounter tick_ffcount;
549 uint64_t period_lerp;
550 volatile uint8_t gen;
551 struct fftimehands *next;
554 #define NUM_ELEMENTS(x) (sizeof(x) / sizeof(*x))
556 static struct fftimehands ffth[10];
557 static struct fftimehands *volatile fftimehands = ffth;
562 struct fftimehands *cur;
563 struct fftimehands *last;
565 memset(ffth, 0, sizeof(ffth));
567 last = ffth + NUM_ELEMENTS(ffth) - 1;
568 for (cur = ffth; cur < last; cur++)
573 ffclock_status = FFCLOCK_STA_UNSYNC;
574 mtx_init(&ffclock_mtx, "ffclock lock", NULL, MTX_DEF);
578 * Reset the feed-forward clock estimates. Called from inittodr() to get things
579 * kick started and uses the timecounter nominal frequency as a first period
580 * estimate. Note: this function may be called several time just after boot.
581 * Note: this is the only function that sets the value of boot time for the
582 * monotonic (i.e. uptime) version of the feed-forward clock.
585 ffclock_reset_clock(struct timespec *ts)
587 struct timecounter *tc;
588 struct ffclock_estimate cest;
590 tc = timehands->th_counter;
591 memset(&cest, 0, sizeof(struct ffclock_estimate));
593 timespec2bintime(ts, &ffclock_boottime);
594 timespec2bintime(ts, &(cest.update_time));
595 ffclock_read_counter(&cest.update_ffcount);
596 cest.leapsec_next = 0;
597 cest.period = ((1ULL << 63) / tc->tc_frequency) << 1;
600 cest.status = FFCLOCK_STA_UNSYNC;
601 cest.leapsec_total = 0;
604 mtx_lock(&ffclock_mtx);
605 bcopy(&cest, &ffclock_estimate, sizeof(struct ffclock_estimate));
606 ffclock_updated = INT8_MAX;
607 mtx_unlock(&ffclock_mtx);
609 printf("ffclock reset: %s (%llu Hz), time = %ld.%09lu\n", tc->tc_name,
610 (unsigned long long)tc->tc_frequency, (long)ts->tv_sec,
611 (unsigned long)ts->tv_nsec);
615 * Sub-routine to convert a time interval measured in RAW counter units to time
616 * in seconds stored in bintime format.
617 * NOTE: bintime_mul requires u_int, but the value of the ffcounter may be
618 * larger than the max value of u_int (on 32 bit architecture). Loop to consume
622 ffclock_convert_delta(ffcounter ffdelta, uint64_t period, struct bintime *bt)
625 ffcounter delta, delta_max;
627 delta_max = (1ULL << (8 * sizeof(unsigned int))) - 1;
630 if (ffdelta > delta_max)
636 bintime_mul(&bt2, (unsigned int)delta);
637 bintime_add(bt, &bt2);
639 } while (ffdelta > 0);
643 * Update the fftimehands.
644 * Push the tick ffcount and time(s) forward based on current clock estimate.
645 * The conversion from ffcounter to bintime relies on the difference clock
646 * principle, whose accuracy relies on computing small time intervals. If a new
647 * clock estimate has been passed by the synchronisation daemon, make it
648 * current, and compute the linear interpolation for monotonic time if needed.
651 ffclock_windup(unsigned int delta)
653 struct ffclock_estimate *cest;
654 struct fftimehands *ffth;
655 struct bintime bt, gap_lerp;
658 unsigned int polling;
659 uint8_t forward_jump, ogen;
662 * Pick the next timehand, copy current ffclock estimates and move tick
663 * times and counter forward.
666 ffth = fftimehands->next;
670 bcopy(&fftimehands->cest, cest, sizeof(struct ffclock_estimate));
671 ffdelta = (ffcounter)delta;
672 ffth->period_lerp = fftimehands->period_lerp;
674 ffth->tick_time = fftimehands->tick_time;
675 ffclock_convert_delta(ffdelta, cest->period, &bt);
676 bintime_add(&ffth->tick_time, &bt);
678 ffth->tick_time_lerp = fftimehands->tick_time_lerp;
679 ffclock_convert_delta(ffdelta, ffth->period_lerp, &bt);
680 bintime_add(&ffth->tick_time_lerp, &bt);
682 ffth->tick_ffcount = fftimehands->tick_ffcount + ffdelta;
685 * Assess the status of the clock, if the last update is too old, it is
686 * likely the synchronisation daemon is dead and the clock is free
689 if (ffclock_updated == 0) {
690 ffdelta = ffth->tick_ffcount - cest->update_ffcount;
691 ffclock_convert_delta(ffdelta, cest->period, &bt);
692 if (bt.sec > 2 * FFCLOCK_SKM_SCALE)
693 ffclock_status |= FFCLOCK_STA_UNSYNC;
697 * If available, grab updated clock estimates and make them current.
698 * Recompute time at this tick using the updated estimates. The clock
699 * estimates passed the feed-forward synchronisation daemon may result
700 * in time conversion that is not monotonically increasing (just after
701 * the update). time_lerp is a particular linear interpolation over the
702 * synchronisation algo polling period that ensures monotonicity for the
703 * clock ids requesting it.
705 if (ffclock_updated > 0) {
706 bcopy(&ffclock_estimate, cest, sizeof(struct ffclock_estimate));
707 ffdelta = ffth->tick_ffcount - cest->update_ffcount;
708 ffth->tick_time = cest->update_time;
709 ffclock_convert_delta(ffdelta, cest->period, &bt);
710 bintime_add(&ffth->tick_time, &bt);
712 /* ffclock_reset sets ffclock_updated to INT8_MAX */
713 if (ffclock_updated == INT8_MAX)
714 ffth->tick_time_lerp = ffth->tick_time;
716 if (bintime_cmp(&ffth->tick_time, &ffth->tick_time_lerp, >))
721 bintime_clear(&gap_lerp);
723 gap_lerp = ffth->tick_time;
724 bintime_sub(&gap_lerp, &ffth->tick_time_lerp);
726 gap_lerp = ffth->tick_time_lerp;
727 bintime_sub(&gap_lerp, &ffth->tick_time);
731 * The reset from the RTC clock may be far from accurate, and
732 * reducing the gap between real time and interpolated time
733 * could take a very long time if the interpolated clock insists
734 * on strict monotonicity. The clock is reset under very strict
735 * conditions (kernel time is known to be wrong and
736 * synchronization daemon has been restarted recently.
737 * ffclock_boottime absorbs the jump to ensure boot time is
738 * correct and uptime functions stay consistent.
740 if (((ffclock_status & FFCLOCK_STA_UNSYNC) == FFCLOCK_STA_UNSYNC) &&
741 ((cest->status & FFCLOCK_STA_UNSYNC) == 0) &&
742 ((cest->status & FFCLOCK_STA_WARMUP) == FFCLOCK_STA_WARMUP)) {
744 bintime_add(&ffclock_boottime, &gap_lerp);
746 bintime_sub(&ffclock_boottime, &gap_lerp);
747 ffth->tick_time_lerp = ffth->tick_time;
748 bintime_clear(&gap_lerp);
751 ffclock_status = cest->status;
752 ffth->period_lerp = cest->period;
755 * Compute corrected period used for the linear interpolation of
756 * time. The rate of linear interpolation is capped to 5000PPM
759 if (bintime_isset(&gap_lerp)) {
760 ffdelta = cest->update_ffcount;
761 ffdelta -= fftimehands->cest.update_ffcount;
762 ffclock_convert_delta(ffdelta, cest->period, &bt);
765 bt.frac = 5000000 * (uint64_t)18446744073LL;
766 bintime_mul(&bt, polling);
767 if (bintime_cmp(&gap_lerp, &bt, >))
770 /* Approximate 1 sec by 1-(1/2^64) to ease arithmetic */
772 if (gap_lerp.sec > 0) {
774 frac /= ffdelta / gap_lerp.sec;
776 frac += gap_lerp.frac / ffdelta;
779 ffth->period_lerp += frac;
781 ffth->period_lerp -= frac;
793 * Adjust the fftimehands when the timecounter is changed. Stating the obvious,
794 * the old and new hardware counter cannot be read simultaneously. tc_windup()
795 * does read the two counters 'back to back', but a few cycles are effectively
796 * lost, and not accumulated in tick_ffcount. This is a fairly radical
797 * operation for a feed-forward synchronization daemon, and it is its job to not
798 * pushing irrelevant data to the kernel. Because there is no locking here,
799 * simply force to ignore pending or next update to give daemon a chance to
800 * realize the counter has changed.
803 ffclock_change_tc(struct timehands *th)
805 struct fftimehands *ffth;
806 struct ffclock_estimate *cest;
807 struct timecounter *tc;
811 ffth = fftimehands->next;
816 bcopy(&(fftimehands->cest), cest, sizeof(struct ffclock_estimate));
817 cest->period = ((1ULL << 63) / tc->tc_frequency ) << 1;
820 cest->status |= FFCLOCK_STA_UNSYNC;
822 ffth->tick_ffcount = fftimehands->tick_ffcount;
823 ffth->tick_time_lerp = fftimehands->tick_time_lerp;
824 ffth->tick_time = fftimehands->tick_time;
825 ffth->period_lerp = cest->period;
827 /* Do not lock but ignore next update from synchronization daemon. */
837 * Retrieve feed-forward counter and time of last kernel tick.
840 ffclock_last_tick(ffcounter *ffcount, struct bintime *bt, uint32_t flags)
842 struct fftimehands *ffth;
846 * No locking but check generation has not changed. Also need to make
847 * sure ffdelta is positive, i.e. ffcount > tick_ffcount.
852 if ((flags & FFCLOCK_LERP) == FFCLOCK_LERP)
853 *bt = ffth->tick_time_lerp;
855 *bt = ffth->tick_time;
856 *ffcount = ffth->tick_ffcount;
857 } while (gen == 0 || gen != ffth->gen);
861 * Absolute clock conversion. Low level function to convert ffcounter to
862 * bintime. The ffcounter is converted using the current ffclock period estimate
863 * or the "interpolated period" to ensure monotonicity.
864 * NOTE: this conversion may have been deferred, and the clock updated since the
865 * hardware counter has been read.
868 ffclock_convert_abs(ffcounter ffcount, struct bintime *bt, uint32_t flags)
870 struct fftimehands *ffth;
876 * No locking but check generation has not changed. Also need to make
877 * sure ffdelta is positive, i.e. ffcount > tick_ffcount.
882 if (ffcount > ffth->tick_ffcount)
883 ffdelta = ffcount - ffth->tick_ffcount;
885 ffdelta = ffth->tick_ffcount - ffcount;
887 if ((flags & FFCLOCK_LERP) == FFCLOCK_LERP) {
888 *bt = ffth->tick_time_lerp;
889 ffclock_convert_delta(ffdelta, ffth->period_lerp, &bt2);
891 *bt = ffth->tick_time;
892 ffclock_convert_delta(ffdelta, ffth->cest.period, &bt2);
895 if (ffcount > ffth->tick_ffcount)
896 bintime_add(bt, &bt2);
898 bintime_sub(bt, &bt2);
899 } while (gen == 0 || gen != ffth->gen);
903 * Difference clock conversion.
904 * Low level function to Convert a time interval measured in RAW counter units
905 * into bintime. The difference clock allows measuring small intervals much more
906 * reliably than the absolute clock.
909 ffclock_convert_diff(ffcounter ffdelta, struct bintime *bt)
911 struct fftimehands *ffth;
914 /* No locking but check generation has not changed. */
918 ffclock_convert_delta(ffdelta, ffth->cest.period, bt);
919 } while (gen == 0 || gen != ffth->gen);
923 * Access to current ffcounter value.
926 ffclock_read_counter(ffcounter *ffcount)
928 struct timehands *th;
929 struct fftimehands *ffth;
930 unsigned int gen, delta;
933 * ffclock_windup() called from tc_windup(), safe to rely on
934 * th->th_generation only, for correct delta and ffcounter.
938 gen = atomic_load_acq_int(&th->th_generation);
940 delta = tc_delta(th);
941 *ffcount = ffth->tick_ffcount;
942 atomic_thread_fence_acq();
943 } while (gen == 0 || gen != th->th_generation);
949 binuptime(struct bintime *bt)
952 binuptime_fromclock(bt, sysclock_active);
956 nanouptime(struct timespec *tsp)
959 nanouptime_fromclock(tsp, sysclock_active);
963 microuptime(struct timeval *tvp)
966 microuptime_fromclock(tvp, sysclock_active);
970 bintime(struct bintime *bt)
973 bintime_fromclock(bt, sysclock_active);
977 nanotime(struct timespec *tsp)
980 nanotime_fromclock(tsp, sysclock_active);
984 microtime(struct timeval *tvp)
987 microtime_fromclock(tvp, sysclock_active);
991 getbinuptime(struct bintime *bt)
994 getbinuptime_fromclock(bt, sysclock_active);
998 getnanouptime(struct timespec *tsp)
1001 getnanouptime_fromclock(tsp, sysclock_active);
1005 getmicrouptime(struct timeval *tvp)
1008 getmicrouptime_fromclock(tvp, sysclock_active);
1012 getbintime(struct bintime *bt)
1015 getbintime_fromclock(bt, sysclock_active);
1019 getnanotime(struct timespec *tsp)
1022 getnanotime_fromclock(tsp, sysclock_active);
1026 getmicrotime(struct timeval *tvp)
1029 getmicrouptime_fromclock(tvp, sysclock_active);
1032 #endif /* FFCLOCK */
1035 * This is a clone of getnanotime and used for walltimestamps.
1036 * The dtrace_ prefix prevents fbt from creating probes for
1037 * it so walltimestamp can be safely used in all fbt probes.
1040 dtrace_getnanotime(struct timespec *tsp)
1042 struct timehands *th;
1047 gen = atomic_load_acq_int(&th->th_generation);
1048 *tsp = th->th_nanotime;
1049 atomic_thread_fence_acq();
1050 } while (gen == 0 || gen != th->th_generation);
1054 * System clock currently providing time to the system. Modifiable via sysctl
1055 * when the FFCLOCK option is defined.
1057 int sysclock_active = SYSCLOCK_FBCK;
1059 /* Internal NTP status and error estimates. */
1060 extern int time_status;
1061 extern long time_esterror;
1064 * Take a snapshot of sysclock data which can be used to compare system clocks
1065 * and generate timestamps after the fact.
1068 sysclock_getsnapshot(struct sysclock_snap *clock_snap, int fast)
1070 struct fbclock_info *fbi;
1071 struct timehands *th;
1073 unsigned int delta, gen;
1076 struct fftimehands *ffth;
1077 struct ffclock_info *ffi;
1078 struct ffclock_estimate cest;
1080 ffi = &clock_snap->ff_info;
1083 fbi = &clock_snap->fb_info;
1088 gen = atomic_load_acq_int(&th->th_generation);
1089 fbi->th_scale = th->th_scale;
1090 fbi->tick_time = th->th_offset;
1093 ffi->tick_time = ffth->tick_time_lerp;
1094 ffi->tick_time_lerp = ffth->tick_time_lerp;
1095 ffi->period = ffth->cest.period;
1096 ffi->period_lerp = ffth->period_lerp;
1097 clock_snap->ffcount = ffth->tick_ffcount;
1101 delta = tc_delta(th);
1102 atomic_thread_fence_acq();
1103 } while (gen == 0 || gen != th->th_generation);
1105 clock_snap->delta = delta;
1106 clock_snap->sysclock_active = sysclock_active;
1108 /* Record feedback clock status and error. */
1109 clock_snap->fb_info.status = time_status;
1110 /* XXX: Very crude estimate of feedback clock error. */
1111 bt.sec = time_esterror / 1000000;
1112 bt.frac = ((time_esterror - bt.sec) * 1000000) *
1113 (uint64_t)18446744073709ULL;
1114 clock_snap->fb_info.error = bt;
1118 clock_snap->ffcount += delta;
1120 /* Record feed-forward clock leap second adjustment. */
1121 ffi->leapsec_adjustment = cest.leapsec_total;
1122 if (clock_snap->ffcount > cest.leapsec_next)
1123 ffi->leapsec_adjustment -= cest.leapsec;
1125 /* Record feed-forward clock status and error. */
1126 clock_snap->ff_info.status = cest.status;
1127 ffcount = clock_snap->ffcount - cest.update_ffcount;
1128 ffclock_convert_delta(ffcount, cest.period, &bt);
1129 /* 18446744073709 = int(2^64/1e12), err_bound_rate in [ps/s]. */
1130 bintime_mul(&bt, cest.errb_rate * (uint64_t)18446744073709ULL);
1131 /* 18446744073 = int(2^64 / 1e9), since err_abs in [ns]. */
1132 bintime_addx(&bt, cest.errb_abs * (uint64_t)18446744073ULL);
1133 clock_snap->ff_info.error = bt;
1138 * Convert a sysclock snapshot into a struct bintime based on the specified
1139 * clock source and flags.
1142 sysclock_snap2bintime(struct sysclock_snap *cs, struct bintime *bt,
1143 int whichclock, uint32_t flags)
1145 struct bintime boottimebin;
1151 switch (whichclock) {
1153 *bt = cs->fb_info.tick_time;
1155 /* If snapshot was created with !fast, delta will be >0. */
1157 bintime_addx(bt, cs->fb_info.th_scale * cs->delta);
1159 if ((flags & FBCLOCK_UPTIME) == 0) {
1160 getboottimebin(&boottimebin);
1161 bintime_add(bt, &boottimebin);
1166 if (flags & FFCLOCK_LERP) {
1167 *bt = cs->ff_info.tick_time_lerp;
1168 period = cs->ff_info.period_lerp;
1170 *bt = cs->ff_info.tick_time;
1171 period = cs->ff_info.period;
1174 /* If snapshot was created with !fast, delta will be >0. */
1175 if (cs->delta > 0) {
1176 ffclock_convert_delta(cs->delta, period, &bt2);
1177 bintime_add(bt, &bt2);
1180 /* Leap second adjustment. */
1181 if (flags & FFCLOCK_LEAPSEC)
1182 bt->sec -= cs->ff_info.leapsec_adjustment;
1184 /* Boot time adjustment, for uptime/monotonic clocks. */
1185 if (flags & FFCLOCK_UPTIME)
1186 bintime_sub(bt, &ffclock_boottime);
1198 * Initialize a new timecounter and possibly use it.
1201 tc_init(struct timecounter *tc)
1204 struct sysctl_oid *tc_root;
1206 u = tc->tc_frequency / tc->tc_counter_mask;
1207 /* XXX: We need some margin here, 10% is a guess */
1210 if (u > hz && tc->tc_quality >= 0) {
1211 tc->tc_quality = -2000;
1213 printf("Timecounter \"%s\" frequency %ju Hz",
1214 tc->tc_name, (uintmax_t)tc->tc_frequency);
1215 printf(" -- Insufficient hz, needs at least %u\n", u);
1217 } else if (tc->tc_quality >= 0 || bootverbose) {
1218 printf("Timecounter \"%s\" frequency %ju Hz quality %d\n",
1219 tc->tc_name, (uintmax_t)tc->tc_frequency,
1223 tc->tc_next = timecounters;
1226 * Set up sysctl tree for this counter.
1228 tc_root = SYSCTL_ADD_NODE_WITH_LABEL(NULL,
1229 SYSCTL_STATIC_CHILDREN(_kern_timecounter_tc), OID_AUTO, tc->tc_name,
1230 CTLFLAG_RW, 0, "timecounter description", "timecounter");
1231 SYSCTL_ADD_UINT(NULL, SYSCTL_CHILDREN(tc_root), OID_AUTO,
1232 "mask", CTLFLAG_RD, &(tc->tc_counter_mask), 0,
1233 "mask for implemented bits");
1234 SYSCTL_ADD_PROC(NULL, SYSCTL_CHILDREN(tc_root), OID_AUTO,
1235 "counter", CTLTYPE_UINT | CTLFLAG_RD, tc, sizeof(*tc),
1236 sysctl_kern_timecounter_get, "IU", "current timecounter value");
1237 SYSCTL_ADD_PROC(NULL, SYSCTL_CHILDREN(tc_root), OID_AUTO,
1238 "frequency", CTLTYPE_U64 | CTLFLAG_RD, tc, sizeof(*tc),
1239 sysctl_kern_timecounter_freq, "QU", "timecounter frequency");
1240 SYSCTL_ADD_INT(NULL, SYSCTL_CHILDREN(tc_root), OID_AUTO,
1241 "quality", CTLFLAG_RD, &(tc->tc_quality), 0,
1242 "goodness of time counter");
1244 * Do not automatically switch if the current tc was specifically
1245 * chosen. Never automatically use a timecounter with negative quality.
1246 * Even though we run on the dummy counter, switching here may be
1247 * worse since this timecounter may not be monotonic.
1251 if (tc->tc_quality < 0)
1253 if (tc->tc_quality < timecounter->tc_quality)
1255 if (tc->tc_quality == timecounter->tc_quality &&
1256 tc->tc_frequency < timecounter->tc_frequency)
1258 (void)tc->tc_get_timecount(tc);
1259 (void)tc->tc_get_timecount(tc);
1263 /* Report the frequency of the current timecounter. */
1265 tc_getfrequency(void)
1268 return (timehands->th_counter->tc_frequency);
1272 sleeping_on_old_rtc(struct thread *td)
1276 * td_rtcgen is modified by curthread when it is running,
1277 * and by other threads in this function. By finding the thread
1278 * on a sleepqueue and holding the lock on the sleepqueue
1279 * chain, we guarantee that the thread is not running and that
1280 * modifying td_rtcgen is safe. Setting td_rtcgen to zero informs
1281 * the thread that it was woken due to a real-time clock adjustment.
1282 * (The declaration of td_rtcgen refers to this comment.)
1284 if (td->td_rtcgen != 0 && td->td_rtcgen != rtc_generation) {
1291 static struct mtx tc_setclock_mtx;
1292 MTX_SYSINIT(tc_setclock_init, &tc_setclock_mtx, "tcsetc", MTX_SPIN);
1295 * Step our concept of UTC. This is done by modifying our estimate of
1299 tc_setclock(struct timespec *ts)
1301 struct timespec tbef, taft;
1302 struct bintime bt, bt2;
1304 timespec2bintime(ts, &bt);
1306 mtx_lock_spin(&tc_setclock_mtx);
1307 cpu_tick_calibrate(1);
1309 bintime_sub(&bt, &bt2);
1311 /* XXX fiddle all the little crinkly bits around the fiords... */
1313 mtx_unlock_spin(&tc_setclock_mtx);
1315 /* Avoid rtc_generation == 0, since td_rtcgen == 0 is special. */
1316 atomic_add_rel_int(&rtc_generation, 2);
1317 sleepq_chains_remove_matching(sleeping_on_old_rtc);
1318 if (timestepwarnings) {
1321 "Time stepped from %jd.%09ld to %jd.%09ld (%jd.%09ld)\n",
1322 (intmax_t)tbef.tv_sec, tbef.tv_nsec,
1323 (intmax_t)taft.tv_sec, taft.tv_nsec,
1324 (intmax_t)ts->tv_sec, ts->tv_nsec);
1329 * Initialize the next struct timehands in the ring and make
1330 * it the active timehands. Along the way we might switch to a different
1331 * timecounter and/or do seconds processing in NTP. Slightly magic.
1334 tc_windup(struct bintime *new_boottimebin)
1337 struct timehands *th, *tho;
1339 u_int delta, ncount, ogen;
1344 * Make the next timehands a copy of the current one, but do
1345 * not overwrite the generation or next pointer. While we
1346 * update the contents, the generation must be zero. We need
1347 * to ensure that the zero generation is visible before the
1348 * data updates become visible, which requires release fence.
1349 * For similar reasons, re-reading of the generation after the
1350 * data is read should use acquire fence.
1354 ogen = th->th_generation;
1355 th->th_generation = 0;
1356 atomic_thread_fence_rel();
1357 memcpy(th, tho, offsetof(struct timehands, th_generation));
1358 if (new_boottimebin != NULL)
1359 th->th_boottime = *new_boottimebin;
1362 * Capture a timecounter delta on the current timecounter and if
1363 * changing timecounters, a counter value from the new timecounter.
1364 * Update the offset fields accordingly.
1366 delta = tc_delta(th);
1367 if (th->th_counter != timecounter)
1368 ncount = timecounter->tc_get_timecount(timecounter);
1372 ffclock_windup(delta);
1374 th->th_offset_count += delta;
1375 th->th_offset_count &= th->th_counter->tc_counter_mask;
1376 while (delta > th->th_counter->tc_frequency) {
1377 /* Eat complete unadjusted seconds. */
1378 delta -= th->th_counter->tc_frequency;
1379 th->th_offset.sec++;
1381 if ((delta > th->th_counter->tc_frequency / 2) &&
1382 (th->th_scale * delta < ((uint64_t)1 << 63))) {
1383 /* The product th_scale * delta just barely overflows. */
1384 th->th_offset.sec++;
1386 bintime_addx(&th->th_offset, th->th_scale * delta);
1389 * Hardware latching timecounters may not generate interrupts on
1390 * PPS events, so instead we poll them. There is a finite risk that
1391 * the hardware might capture a count which is later than the one we
1392 * got above, and therefore possibly in the next NTP second which might
1393 * have a different rate than the current NTP second. It doesn't
1394 * matter in practice.
1396 if (tho->th_counter->tc_poll_pps)
1397 tho->th_counter->tc_poll_pps(tho->th_counter);
1400 * Deal with NTP second processing. The for loop normally
1401 * iterates at most once, but in extreme situations it might
1402 * keep NTP sane if timeouts are not run for several seconds.
1403 * At boot, the time step can be large when the TOD hardware
1404 * has been read, so on really large steps, we call
1405 * ntp_update_second only twice. We need to call it twice in
1406 * case we missed a leap second.
1409 bintime_add(&bt, &th->th_boottime);
1410 i = bt.sec - tho->th_microtime.tv_sec;
1413 for (; i > 0; i--) {
1415 ntp_update_second(&th->th_adjustment, &bt.sec);
1417 th->th_boottime.sec += bt.sec - t;
1419 /* Update the UTC timestamps used by the get*() functions. */
1420 th->th_bintime = bt;
1421 bintime2timeval(&bt, &th->th_microtime);
1422 bintime2timespec(&bt, &th->th_nanotime);
1424 /* Now is a good time to change timecounters. */
1425 if (th->th_counter != timecounter) {
1427 if ((timecounter->tc_flags & TC_FLAGS_C2STOP) != 0)
1428 cpu_disable_c2_sleep++;
1429 if ((th->th_counter->tc_flags & TC_FLAGS_C2STOP) != 0)
1430 cpu_disable_c2_sleep--;
1432 th->th_counter = timecounter;
1433 th->th_offset_count = ncount;
1434 tc_min_ticktock_freq = max(1, timecounter->tc_frequency /
1435 (((uint64_t)timecounter->tc_counter_mask + 1) / 3));
1437 ffclock_change_tc(th);
1442 * Recalculate the scaling factor. We want the number of 1/2^64
1443 * fractions of a second per period of the hardware counter, taking
1444 * into account the th_adjustment factor which the NTP PLL/adjtime(2)
1445 * processing provides us with.
1447 * The th_adjustment is nanoseconds per second with 32 bit binary
1448 * fraction and we want 64 bit binary fraction of second:
1450 * x = a * 2^32 / 10^9 = a * 4.294967296
1452 * The range of th_adjustment is +/- 5000PPM so inside a 64bit int
1453 * we can only multiply by about 850 without overflowing, that
1454 * leaves no suitably precise fractions for multiply before divide.
1456 * Divide before multiply with a fraction of 2199/512 results in a
1457 * systematic undercompensation of 10PPM of th_adjustment. On a
1458 * 5000PPM adjustment this is a 0.05PPM error. This is acceptable.
1460 * We happily sacrifice the lowest of the 64 bits of our result
1461 * to the goddess of code clarity.
1464 scale = (uint64_t)1 << 63;
1465 scale += (th->th_adjustment / 1024) * 2199;
1466 scale /= th->th_counter->tc_frequency;
1467 th->th_scale = scale * 2;
1470 * Now that the struct timehands is again consistent, set the new
1471 * generation number, making sure to not make it zero.
1475 atomic_store_rel_int(&th->th_generation, ogen);
1477 /* Go live with the new struct timehands. */
1479 switch (sysclock_active) {
1482 time_second = th->th_microtime.tv_sec;
1483 time_uptime = th->th_offset.sec;
1487 time_second = fftimehands->tick_time_lerp.sec;
1488 time_uptime = fftimehands->tick_time_lerp.sec - ffclock_boottime.sec;
1494 timekeep_push_vdso();
1497 /* Report or change the active timecounter hardware. */
1499 sysctl_kern_timecounter_hardware(SYSCTL_HANDLER_ARGS)
1502 struct timecounter *newtc, *tc;
1506 strlcpy(newname, tc->tc_name, sizeof(newname));
1508 error = sysctl_handle_string(oidp, &newname[0], sizeof(newname), req);
1509 if (error != 0 || req->newptr == NULL)
1511 /* Record that the tc in use now was specifically chosen. */
1513 if (strcmp(newname, tc->tc_name) == 0)
1515 for (newtc = timecounters; newtc != NULL; newtc = newtc->tc_next) {
1516 if (strcmp(newname, newtc->tc_name) != 0)
1519 /* Warm up new timecounter. */
1520 (void)newtc->tc_get_timecount(newtc);
1521 (void)newtc->tc_get_timecount(newtc);
1523 timecounter = newtc;
1526 * The vdso timehands update is deferred until the next
1529 * This is prudent given that 'timekeep_push_vdso()' does not
1530 * use any locking and that it can be called in hard interrupt
1531 * context via 'tc_windup()'.
1538 SYSCTL_PROC(_kern_timecounter, OID_AUTO, hardware, CTLTYPE_STRING | CTLFLAG_RW,
1539 0, 0, sysctl_kern_timecounter_hardware, "A",
1540 "Timecounter hardware selected");
1543 /* Report the available timecounter hardware. */
1545 sysctl_kern_timecounter_choice(SYSCTL_HANDLER_ARGS)
1548 struct timecounter *tc;
1551 sbuf_new_for_sysctl(&sb, NULL, 0, req);
1552 for (tc = timecounters; tc != NULL; tc = tc->tc_next) {
1553 if (tc != timecounters)
1554 sbuf_putc(&sb, ' ');
1555 sbuf_printf(&sb, "%s(%d)", tc->tc_name, tc->tc_quality);
1557 error = sbuf_finish(&sb);
1562 SYSCTL_PROC(_kern_timecounter, OID_AUTO, choice, CTLTYPE_STRING | CTLFLAG_RD,
1563 0, 0, sysctl_kern_timecounter_choice, "A", "Timecounter hardware detected");
1566 * RFC 2783 PPS-API implementation.
1570 * Return true if the driver is aware of the abi version extensions in the
1571 * pps_state structure, and it supports at least the given abi version number.
1574 abi_aware(struct pps_state *pps, int vers)
1577 return ((pps->kcmode & KCMODE_ABIFLAG) && pps->driver_abi >= vers);
1581 pps_fetch(struct pps_fetch_args *fapi, struct pps_state *pps)
1584 pps_seq_t aseq, cseq;
1587 if (fapi->tsformat && fapi->tsformat != PPS_TSFMT_TSPEC)
1591 * If no timeout is requested, immediately return whatever values were
1592 * most recently captured. If timeout seconds is -1, that's a request
1593 * to block without a timeout. WITNESS won't let us sleep forever
1594 * without a lock (we really don't need a lock), so just repeatedly
1595 * sleep a long time.
1597 if (fapi->timeout.tv_sec || fapi->timeout.tv_nsec) {
1598 if (fapi->timeout.tv_sec == -1)
1601 tv.tv_sec = fapi->timeout.tv_sec;
1602 tv.tv_usec = fapi->timeout.tv_nsec / 1000;
1605 aseq = atomic_load_int(&pps->ppsinfo.assert_sequence);
1606 cseq = atomic_load_int(&pps->ppsinfo.clear_sequence);
1607 while (aseq == atomic_load_int(&pps->ppsinfo.assert_sequence) &&
1608 cseq == atomic_load_int(&pps->ppsinfo.clear_sequence)) {
1609 if (abi_aware(pps, 1) && pps->driver_mtx != NULL) {
1610 if (pps->flags & PPSFLAG_MTX_SPIN) {
1611 err = msleep_spin(pps, pps->driver_mtx,
1614 err = msleep(pps, pps->driver_mtx, PCATCH,
1618 err = tsleep(pps, PCATCH, "ppsfch", timo);
1620 if (err == EWOULDBLOCK) {
1621 if (fapi->timeout.tv_sec == -1) {
1626 } else if (err != 0) {
1632 pps->ppsinfo.current_mode = pps->ppsparam.mode;
1633 fapi->pps_info_buf = pps->ppsinfo;
1639 pps_ioctl(u_long cmd, caddr_t data, struct pps_state *pps)
1642 struct pps_fetch_args *fapi;
1644 struct pps_fetch_ffc_args *fapi_ffc;
1647 struct pps_kcbind_args *kapi;
1650 KASSERT(pps != NULL, ("NULL pps pointer in pps_ioctl"));
1652 case PPS_IOC_CREATE:
1654 case PPS_IOC_DESTROY:
1656 case PPS_IOC_SETPARAMS:
1657 app = (pps_params_t *)data;
1658 if (app->mode & ~pps->ppscap)
1661 /* Ensure only a single clock is selected for ffc timestamp. */
1662 if ((app->mode & PPS_TSCLK_MASK) == PPS_TSCLK_MASK)
1665 pps->ppsparam = *app;
1667 case PPS_IOC_GETPARAMS:
1668 app = (pps_params_t *)data;
1669 *app = pps->ppsparam;
1670 app->api_version = PPS_API_VERS_1;
1672 case PPS_IOC_GETCAP:
1673 *(int*)data = pps->ppscap;
1676 fapi = (struct pps_fetch_args *)data;
1677 return (pps_fetch(fapi, pps));
1679 case PPS_IOC_FETCH_FFCOUNTER:
1680 fapi_ffc = (struct pps_fetch_ffc_args *)data;
1681 if (fapi_ffc->tsformat && fapi_ffc->tsformat !=
1684 if (fapi_ffc->timeout.tv_sec || fapi_ffc->timeout.tv_nsec)
1685 return (EOPNOTSUPP);
1686 pps->ppsinfo_ffc.current_mode = pps->ppsparam.mode;
1687 fapi_ffc->pps_info_buf_ffc = pps->ppsinfo_ffc;
1688 /* Overwrite timestamps if feedback clock selected. */
1689 switch (pps->ppsparam.mode & PPS_TSCLK_MASK) {
1690 case PPS_TSCLK_FBCK:
1691 fapi_ffc->pps_info_buf_ffc.assert_timestamp =
1692 pps->ppsinfo.assert_timestamp;
1693 fapi_ffc->pps_info_buf_ffc.clear_timestamp =
1694 pps->ppsinfo.clear_timestamp;
1696 case PPS_TSCLK_FFWD:
1702 #endif /* FFCLOCK */
1703 case PPS_IOC_KCBIND:
1705 kapi = (struct pps_kcbind_args *)data;
1706 /* XXX Only root should be able to do this */
1707 if (kapi->tsformat && kapi->tsformat != PPS_TSFMT_TSPEC)
1709 if (kapi->kernel_consumer != PPS_KC_HARDPPS)
1711 if (kapi->edge & ~pps->ppscap)
1713 pps->kcmode = (kapi->edge & KCMODE_EDGEMASK) |
1714 (pps->kcmode & KCMODE_ABIFLAG);
1717 return (EOPNOTSUPP);
1725 pps_init(struct pps_state *pps)
1727 pps->ppscap |= PPS_TSFMT_TSPEC | PPS_CANWAIT;
1728 if (pps->ppscap & PPS_CAPTUREASSERT)
1729 pps->ppscap |= PPS_OFFSETASSERT;
1730 if (pps->ppscap & PPS_CAPTURECLEAR)
1731 pps->ppscap |= PPS_OFFSETCLEAR;
1733 pps->ppscap |= PPS_TSCLK_MASK;
1735 pps->kcmode &= ~KCMODE_ABIFLAG;
1739 pps_init_abi(struct pps_state *pps)
1743 if (pps->driver_abi > 0) {
1744 pps->kcmode |= KCMODE_ABIFLAG;
1745 pps->kernel_abi = PPS_ABI_VERSION;
1750 pps_capture(struct pps_state *pps)
1752 struct timehands *th;
1754 KASSERT(pps != NULL, ("NULL pps pointer in pps_capture"));
1756 pps->capgen = atomic_load_acq_int(&th->th_generation);
1759 pps->capffth = fftimehands;
1761 pps->capcount = th->th_counter->tc_get_timecount(th->th_counter);
1762 atomic_thread_fence_acq();
1763 if (pps->capgen != th->th_generation)
1768 pps_event(struct pps_state *pps, int event)
1771 struct timespec ts, *tsp, *osp;
1772 u_int tcount, *pcount;
1776 struct timespec *tsp_ffc;
1777 pps_seq_t *pseq_ffc;
1784 KASSERT(pps != NULL, ("NULL pps pointer in pps_event"));
1785 /* Nothing to do if not currently set to capture this event type. */
1786 if ((event & pps->ppsparam.mode) == 0)
1788 /* If the timecounter was wound up underneath us, bail out. */
1789 if (pps->capgen == 0 || pps->capgen !=
1790 atomic_load_acq_int(&pps->capth->th_generation))
1793 /* Things would be easier with arrays. */
1794 if (event == PPS_CAPTUREASSERT) {
1795 tsp = &pps->ppsinfo.assert_timestamp;
1796 osp = &pps->ppsparam.assert_offset;
1797 foff = pps->ppsparam.mode & PPS_OFFSETASSERT;
1799 fhard = pps->kcmode & PPS_CAPTUREASSERT;
1801 pcount = &pps->ppscount[0];
1802 pseq = &pps->ppsinfo.assert_sequence;
1804 ffcount = &pps->ppsinfo_ffc.assert_ffcount;
1805 tsp_ffc = &pps->ppsinfo_ffc.assert_timestamp;
1806 pseq_ffc = &pps->ppsinfo_ffc.assert_sequence;
1809 tsp = &pps->ppsinfo.clear_timestamp;
1810 osp = &pps->ppsparam.clear_offset;
1811 foff = pps->ppsparam.mode & PPS_OFFSETCLEAR;
1813 fhard = pps->kcmode & PPS_CAPTURECLEAR;
1815 pcount = &pps->ppscount[1];
1816 pseq = &pps->ppsinfo.clear_sequence;
1818 ffcount = &pps->ppsinfo_ffc.clear_ffcount;
1819 tsp_ffc = &pps->ppsinfo_ffc.clear_timestamp;
1820 pseq_ffc = &pps->ppsinfo_ffc.clear_sequence;
1825 * If the timecounter changed, we cannot compare the count values, so
1826 * we have to drop the rest of the PPS-stuff until the next event.
1828 if (pps->ppstc != pps->capth->th_counter) {
1829 pps->ppstc = pps->capth->th_counter;
1830 *pcount = pps->capcount;
1831 pps->ppscount[2] = pps->capcount;
1835 /* Convert the count to a timespec. */
1836 tcount = pps->capcount - pps->capth->th_offset_count;
1837 tcount &= pps->capth->th_counter->tc_counter_mask;
1838 bt = pps->capth->th_bintime;
1839 bintime_addx(&bt, pps->capth->th_scale * tcount);
1840 bintime2timespec(&bt, &ts);
1842 /* If the timecounter was wound up underneath us, bail out. */
1843 atomic_thread_fence_acq();
1844 if (pps->capgen != pps->capth->th_generation)
1847 *pcount = pps->capcount;
1852 timespecadd(tsp, osp, tsp);
1853 if (tsp->tv_nsec < 0) {
1854 tsp->tv_nsec += 1000000000;
1860 *ffcount = pps->capffth->tick_ffcount + tcount;
1861 bt = pps->capffth->tick_time;
1862 ffclock_convert_delta(tcount, pps->capffth->cest.period, &bt);
1863 bintime_add(&bt, &pps->capffth->tick_time);
1864 bintime2timespec(&bt, &ts);
1874 * Feed the NTP PLL/FLL.
1875 * The FLL wants to know how many (hardware) nanoseconds
1876 * elapsed since the previous event.
1878 tcount = pps->capcount - pps->ppscount[2];
1879 pps->ppscount[2] = pps->capcount;
1880 tcount &= pps->capth->th_counter->tc_counter_mask;
1881 scale = (uint64_t)1 << 63;
1882 scale /= pps->capth->th_counter->tc_frequency;
1886 bintime_addx(&bt, scale * tcount);
1887 bintime2timespec(&bt, &ts);
1888 hardpps(tsp, ts.tv_nsec + 1000000000 * ts.tv_sec);
1892 /* Wakeup anyone sleeping in pps_fetch(). */
1897 * Timecounters need to be updated every so often to prevent the hardware
1898 * counter from overflowing. Updating also recalculates the cached values
1899 * used by the get*() family of functions, so their precision depends on
1900 * the update frequency.
1904 SYSCTL_INT(_kern_timecounter, OID_AUTO, tick, CTLFLAG_RD, &tc_tick, 0,
1905 "Approximate number of hardclock ticks in a millisecond");
1908 tc_ticktock(int cnt)
1912 if (mtx_trylock_spin(&tc_setclock_mtx)) {
1914 if (count >= tc_tick) {
1918 mtx_unlock_spin(&tc_setclock_mtx);
1922 static void __inline
1923 tc_adjprecision(void)
1927 if (tc_timepercentage > 0) {
1928 t = (99 + tc_timepercentage) / tc_timepercentage;
1929 tc_precexp = fls(t + (t >> 1)) - 1;
1930 FREQ2BT(hz / tc_tick, &bt_timethreshold);
1931 FREQ2BT(hz, &bt_tickthreshold);
1932 bintime_shift(&bt_timethreshold, tc_precexp);
1933 bintime_shift(&bt_tickthreshold, tc_precexp);
1936 bt_timethreshold.sec = INT_MAX;
1937 bt_timethreshold.frac = ~(uint64_t)0;
1938 bt_tickthreshold = bt_timethreshold;
1940 sbt_timethreshold = bttosbt(bt_timethreshold);
1941 sbt_tickthreshold = bttosbt(bt_tickthreshold);
1945 sysctl_kern_timecounter_adjprecision(SYSCTL_HANDLER_ARGS)
1949 val = tc_timepercentage;
1950 error = sysctl_handle_int(oidp, &val, 0, req);
1951 if (error != 0 || req->newptr == NULL)
1953 tc_timepercentage = val;
1962 inittimecounter(void *dummy)
1964 struct timehands *thp;
1969 * Set the initial timeout to
1970 * max(1, <approx. number of hardclock ticks in a millisecond>).
1971 * People should probably not use the sysctl to set the timeout
1972 * to smaller than its initial value, since that value is the
1973 * smallest reasonable one. If they want better timestamps they
1974 * should use the non-"get"* functions.
1977 tc_tick = (hz + 500) / 1000;
1981 FREQ2BT(hz, &tick_bt);
1982 tick_sbt = bttosbt(tick_bt);
1983 tick_rate = hz / tc_tick;
1984 FREQ2BT(tick_rate, &tc_tick_bt);
1985 tc_tick_sbt = bttosbt(tc_tick_bt);
1986 p = (tc_tick * 1000000) / hz;
1987 printf("Timecounters tick every %d.%03u msec\n", p / 1000, p % 1000);
1993 /* Set up the requested number of timehands. */
1994 if (timehands_count < 1)
1995 timehands_count = 1;
1996 if (timehands_count > nitems(ths))
1997 timehands_count = nitems(ths);
1998 for (i = 1, thp = &ths[0]; i < timehands_count; thp = &ths[i++])
1999 thp->th_next = &ths[i];
2000 thp->th_next = &ths[0];
2002 /* warm up new timecounter (again) and get rolling. */
2003 (void)timecounter->tc_get_timecount(timecounter);
2004 (void)timecounter->tc_get_timecount(timecounter);
2005 mtx_lock_spin(&tc_setclock_mtx);
2007 mtx_unlock_spin(&tc_setclock_mtx);
2010 SYSINIT(timecounter, SI_SUB_CLOCKS, SI_ORDER_SECOND, inittimecounter, NULL);
2012 /* Cpu tick handling -------------------------------------------------*/
2014 static int cpu_tick_variable;
2015 static uint64_t cpu_tick_frequency;
2017 DPCPU_DEFINE_STATIC(uint64_t, tc_cpu_ticks_base);
2018 DPCPU_DEFINE_STATIC(unsigned, tc_cpu_ticks_last);
2023 struct timecounter *tc;
2024 uint64_t res, *base;
2028 base = DPCPU_PTR(tc_cpu_ticks_base);
2029 last = DPCPU_PTR(tc_cpu_ticks_last);
2030 tc = timehands->th_counter;
2031 u = tc->tc_get_timecount(tc) & tc->tc_counter_mask;
2033 *base += (uint64_t)tc->tc_counter_mask + 1;
2041 cpu_tick_calibration(void)
2043 static time_t last_calib;
2045 if (time_uptime != last_calib && !(time_uptime & 0xf)) {
2046 cpu_tick_calibrate(0);
2047 last_calib = time_uptime;
2052 * This function gets called every 16 seconds on only one designated
2053 * CPU in the system from hardclock() via cpu_tick_calibration()().
2055 * Whenever the real time clock is stepped we get called with reset=1
2056 * to make sure we handle suspend/resume and similar events correctly.
2060 cpu_tick_calibrate(int reset)
2062 static uint64_t c_last;
2063 uint64_t c_this, c_delta;
2064 static struct bintime t_last;
2065 struct bintime t_this, t_delta;
2069 /* The clock was stepped, abort & reset */
2074 /* we don't calibrate fixed rate cputicks */
2075 if (!cpu_tick_variable)
2078 getbinuptime(&t_this);
2079 c_this = cpu_ticks();
2080 if (t_last.sec != 0) {
2081 c_delta = c_this - c_last;
2083 bintime_sub(&t_delta, &t_last);
2086 * 2^(64-20) / 16[s] =
2088 * 17.592.186.044.416 / 16 =
2089 * 1.099.511.627.776 [Hz]
2091 divi = t_delta.sec << 20;
2092 divi |= t_delta.frac >> (64 - 20);
2095 if (c_delta > cpu_tick_frequency) {
2096 if (0 && bootverbose)
2097 printf("cpu_tick increased to %ju Hz\n",
2099 cpu_tick_frequency = c_delta;
2107 set_cputicker(cpu_tick_f *func, uint64_t freq, unsigned var)
2111 cpu_ticks = tc_cpu_ticks;
2113 cpu_tick_frequency = freq;
2114 cpu_tick_variable = var;
2123 if (cpu_ticks == tc_cpu_ticks)
2124 return (tc_getfrequency());
2125 return (cpu_tick_frequency);
2129 * We need to be slightly careful converting cputicks to microseconds.
2130 * There is plenty of margin in 64 bits of microseconds (half a million
2131 * years) and in 64 bits at 4 GHz (146 years), but if we do a multiply
2132 * before divide conversion (to retain precision) we find that the
2133 * margin shrinks to 1.5 hours (one millionth of 146y).
2134 * With a three prong approach we never lose significant bits, no
2135 * matter what the cputick rate and length of timeinterval is.
2139 cputick2usec(uint64_t tick)
2142 if (tick > 18446744073709551LL) /* floor(2^64 / 1000) */
2143 return (tick / (cpu_tickrate() / 1000000LL));
2144 else if (tick > 18446744073709LL) /* floor(2^64 / 1000000) */
2145 return ((tick * 1000LL) / (cpu_tickrate() / 1000LL));
2147 return ((tick * 1000000LL) / cpu_tickrate());
2150 cpu_tick_f *cpu_ticks = tc_cpu_ticks;
2152 static int vdso_th_enable = 1;
2154 sysctl_fast_gettime(SYSCTL_HANDLER_ARGS)
2156 int old_vdso_th_enable, error;
2158 old_vdso_th_enable = vdso_th_enable;
2159 error = sysctl_handle_int(oidp, &old_vdso_th_enable, 0, req);
2162 vdso_th_enable = old_vdso_th_enable;
2165 SYSCTL_PROC(_kern_timecounter, OID_AUTO, fast_gettime,
2166 CTLTYPE_INT | CTLFLAG_RW | CTLFLAG_MPSAFE,
2167 NULL, 0, sysctl_fast_gettime, "I", "Enable fast time of day");
2170 tc_fill_vdso_timehands(struct vdso_timehands *vdso_th)
2172 struct timehands *th;
2176 vdso_th->th_scale = th->th_scale;
2177 vdso_th->th_offset_count = th->th_offset_count;
2178 vdso_th->th_counter_mask = th->th_counter->tc_counter_mask;
2179 vdso_th->th_offset = th->th_offset;
2180 vdso_th->th_boottime = th->th_boottime;
2181 if (th->th_counter->tc_fill_vdso_timehands != NULL) {
2182 enabled = th->th_counter->tc_fill_vdso_timehands(vdso_th,
2186 if (!vdso_th_enable)
2191 #ifdef COMPAT_FREEBSD32
2193 tc_fill_vdso_timehands32(struct vdso_timehands32 *vdso_th32)
2195 struct timehands *th;
2199 *(uint64_t *)&vdso_th32->th_scale[0] = th->th_scale;
2200 vdso_th32->th_offset_count = th->th_offset_count;
2201 vdso_th32->th_counter_mask = th->th_counter->tc_counter_mask;
2202 vdso_th32->th_offset.sec = th->th_offset.sec;
2203 *(uint64_t *)&vdso_th32->th_offset.frac[0] = th->th_offset.frac;
2204 vdso_th32->th_boottime.sec = th->th_boottime.sec;
2205 *(uint64_t *)&vdso_th32->th_boottime.frac[0] = th->th_boottime.frac;
2206 if (th->th_counter->tc_fill_vdso_timehands32 != NULL) {
2207 enabled = th->th_counter->tc_fill_vdso_timehands32(vdso_th32,
2211 if (!vdso_th_enable)