2 * ----------------------------------------------------------------------------
3 * "THE BEER-WARE LICENSE" (Revision 42):
4 * <phk@FreeBSD.ORG> wrote this file. As long as you retain this notice you
5 * can do whatever you want with this stuff. If we meet some day, and you think
6 * this stuff is worth it, you can buy me a beer in return. Poul-Henning Kamp
7 * ----------------------------------------------------------------------------
9 * Copyright (c) 2011, 2015, 2016 The FreeBSD Foundation
10 * All rights reserved.
12 * Portions of this software were developed by Julien Ridoux at the University
13 * of Melbourne under sponsorship from the FreeBSD Foundation.
15 * Portions of this software were developed by Konstantin Belousov
16 * under sponsorship from the FreeBSD Foundation.
19 #include <sys/cdefs.h>
20 __FBSDID("$FreeBSD$");
22 #include "opt_compat.h"
24 #include "opt_ffclock.h"
26 #include <sys/param.h>
27 #include <sys/kernel.h>
28 #include <sys/limits.h>
30 #include <sys/mutex.h>
33 #include <sys/sleepqueue.h>
34 #include <sys/sysctl.h>
35 #include <sys/syslog.h>
36 #include <sys/systm.h>
37 #include <sys/timeffc.h>
38 #include <sys/timepps.h>
39 #include <sys/timetc.h>
40 #include <sys/timex.h>
44 * A large step happens on boot. This constant detects such steps.
45 * It is relatively small so that ntp_update_second gets called enough
46 * in the typical 'missed a couple of seconds' case, but doesn't loop
47 * forever when the time step is large.
49 #define LARGE_STEP 200
52 * Implement a dummy timecounter which we can use until we get a real one
53 * in the air. This allows the console and other early stuff to use
58 dummy_get_timecount(struct timecounter *tc)
65 static struct timecounter dummy_timecounter = {
66 dummy_get_timecount, 0, ~0u, 1000000, "dummy", -1000000
70 /* These fields must be initialized by the driver. */
71 struct timecounter *th_counter;
72 int64_t th_adjustment;
74 u_int th_offset_count;
75 struct bintime th_offset;
76 struct bintime th_bintime;
77 struct timeval th_microtime;
78 struct timespec th_nanotime;
79 struct bintime th_boottime;
80 /* Fields not to be copied in tc_windup start with th_generation. */
82 struct timehands *th_next;
85 static struct timehands th0;
86 static struct timehands th1 = {
89 static struct timehands th0 = {
90 .th_counter = &dummy_timecounter,
91 .th_scale = (uint64_t)-1 / 1000000,
92 .th_offset = { .sec = 1 },
97 static struct timehands *volatile timehands = &th0;
98 struct timecounter *timecounter = &dummy_timecounter;
99 static struct timecounter *timecounters = &dummy_timecounter;
101 int tc_min_ticktock_freq = 1;
103 volatile time_t time_second = 1;
104 volatile time_t time_uptime = 1;
106 struct bintime boottimebin;
107 struct timeval boottime;
108 static int sysctl_kern_boottime(SYSCTL_HANDLER_ARGS);
109 SYSCTL_PROC(_kern, KERN_BOOTTIME, boottime, CTLTYPE_STRUCT|CTLFLAG_RD,
110 NULL, 0, sysctl_kern_boottime, "S,timeval", "System boottime");
112 SYSCTL_NODE(_kern, OID_AUTO, timecounter, CTLFLAG_RW, 0, "");
113 static SYSCTL_NODE(_kern_timecounter, OID_AUTO, tc, CTLFLAG_RW, 0, "");
115 static int timestepwarnings;
116 SYSCTL_INT(_kern_timecounter, OID_AUTO, stepwarnings, CTLFLAG_RW,
117 ×tepwarnings, 0, "Log time steps");
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_x;
147 getboottime(&boottime_x);
153 if (req->flags & SCTL_MASK32) {
154 tv[0] = boottime_x.tv_sec;
155 tv[1] = boottime_x.tv_usec;
156 return (SYSCTL_OUT(req, tv, sizeof(tv)));
160 return (SYSCTL_OUT(req, &boottime_x, sizeof(boottime_x)));
164 sysctl_kern_timecounter_get(SYSCTL_HANDLER_ARGS)
167 struct timecounter *tc = arg1;
169 ncount = tc->tc_get_timecount(tc);
170 return (sysctl_handle_int(oidp, &ncount, 0, req));
174 sysctl_kern_timecounter_freq(SYSCTL_HANDLER_ARGS)
177 struct timecounter *tc = arg1;
179 freq = tc->tc_frequency;
180 return (sysctl_handle_64(oidp, &freq, 0, req));
184 * Return the difference between the timehands' counter value now and what
185 * was when we copied it to the timehands' offset_count.
187 static __inline u_int
188 tc_delta(struct timehands *th)
190 struct timecounter *tc;
193 return ((tc->tc_get_timecount(tc) - th->th_offset_count) &
194 tc->tc_counter_mask);
198 * Functions for reading the time. We have to loop until we are sure that
199 * the timehands that we operated on was not updated under our feet. See
200 * the comment in <sys/time.h> for a description of these 12 functions.
205 fbclock_binuptime(struct bintime *bt)
207 struct timehands *th;
212 gen = atomic_load_acq_int(&th->th_generation);
214 bintime_addx(bt, th->th_scale * tc_delta(th));
215 atomic_thread_fence_acq();
216 } while (gen == 0 || gen != th->th_generation);
220 fbclock_nanouptime(struct timespec *tsp)
224 fbclock_binuptime(&bt);
225 bintime2timespec(&bt, tsp);
229 fbclock_microuptime(struct timeval *tvp)
233 fbclock_binuptime(&bt);
234 bintime2timeval(&bt, tvp);
238 fbclock_bintime(struct bintime *bt)
240 struct timehands *th;
245 gen = atomic_load_acq_int(&th->th_generation);
246 *bt = th->th_bintime;
247 bintime_addx(bt, th->th_scale * tc_delta(th));
248 atomic_thread_fence_acq();
249 } while (gen == 0 || gen != th->th_generation);
253 fbclock_nanotime(struct timespec *tsp)
257 fbclock_bintime(&bt);
258 bintime2timespec(&bt, tsp);
262 fbclock_microtime(struct timeval *tvp)
266 fbclock_bintime(&bt);
267 bintime2timeval(&bt, tvp);
271 fbclock_getbinuptime(struct bintime *bt)
273 struct timehands *th;
278 gen = atomic_load_acq_int(&th->th_generation);
280 atomic_thread_fence_acq();
281 } while (gen == 0 || gen != th->th_generation);
285 fbclock_getnanouptime(struct timespec *tsp)
287 struct timehands *th;
292 gen = atomic_load_acq_int(&th->th_generation);
293 bintime2timespec(&th->th_offset, tsp);
294 atomic_thread_fence_acq();
295 } while (gen == 0 || gen != th->th_generation);
299 fbclock_getmicrouptime(struct timeval *tvp)
301 struct timehands *th;
306 gen = atomic_load_acq_int(&th->th_generation);
307 bintime2timeval(&th->th_offset, tvp);
308 atomic_thread_fence_acq();
309 } while (gen == 0 || gen != th->th_generation);
313 fbclock_getbintime(struct bintime *bt)
315 struct timehands *th;
320 gen = atomic_load_acq_int(&th->th_generation);
321 *bt = th->th_bintime;
322 atomic_thread_fence_acq();
323 } while (gen == 0 || gen != th->th_generation);
327 fbclock_getnanotime(struct timespec *tsp)
329 struct timehands *th;
334 gen = atomic_load_acq_int(&th->th_generation);
335 *tsp = th->th_nanotime;
336 atomic_thread_fence_acq();
337 } while (gen == 0 || gen != th->th_generation);
341 fbclock_getmicrotime(struct timeval *tvp)
343 struct timehands *th;
348 gen = atomic_load_acq_int(&th->th_generation);
349 *tvp = th->th_microtime;
350 atomic_thread_fence_acq();
351 } while (gen == 0 || gen != th->th_generation);
355 binuptime(struct bintime *bt)
357 struct timehands *th;
362 gen = atomic_load_acq_int(&th->th_generation);
364 bintime_addx(bt, th->th_scale * tc_delta(th));
365 atomic_thread_fence_acq();
366 } while (gen == 0 || gen != th->th_generation);
370 nanouptime(struct timespec *tsp)
375 bintime2timespec(&bt, tsp);
379 microuptime(struct timeval *tvp)
384 bintime2timeval(&bt, tvp);
388 bintime(struct bintime *bt)
390 struct timehands *th;
395 gen = atomic_load_acq_int(&th->th_generation);
396 *bt = th->th_bintime;
397 bintime_addx(bt, th->th_scale * tc_delta(th));
398 atomic_thread_fence_acq();
399 } while (gen == 0 || gen != th->th_generation);
403 nanotime(struct timespec *tsp)
408 bintime2timespec(&bt, tsp);
412 microtime(struct timeval *tvp)
417 bintime2timeval(&bt, tvp);
421 getbinuptime(struct bintime *bt)
423 struct timehands *th;
428 gen = atomic_load_acq_int(&th->th_generation);
430 atomic_thread_fence_acq();
431 } while (gen == 0 || gen != th->th_generation);
435 getnanouptime(struct timespec *tsp)
437 struct timehands *th;
442 gen = atomic_load_acq_int(&th->th_generation);
443 bintime2timespec(&th->th_offset, tsp);
444 atomic_thread_fence_acq();
445 } while (gen == 0 || gen != th->th_generation);
449 getmicrouptime(struct timeval *tvp)
451 struct timehands *th;
456 gen = atomic_load_acq_int(&th->th_generation);
457 bintime2timeval(&th->th_offset, tvp);
458 atomic_thread_fence_acq();
459 } while (gen == 0 || gen != th->th_generation);
463 getbintime(struct bintime *bt)
465 struct timehands *th;
470 gen = atomic_load_acq_int(&th->th_generation);
471 *bt = th->th_bintime;
472 atomic_thread_fence_acq();
473 } while (gen == 0 || gen != th->th_generation);
477 getnanotime(struct timespec *tsp)
479 struct timehands *th;
484 gen = atomic_load_acq_int(&th->th_generation);
485 *tsp = th->th_nanotime;
486 atomic_thread_fence_acq();
487 } while (gen == 0 || gen != th->th_generation);
491 getmicrotime(struct timeval *tvp)
493 struct timehands *th;
498 gen = atomic_load_acq_int(&th->th_generation);
499 *tvp = th->th_microtime;
500 atomic_thread_fence_acq();
501 } while (gen == 0 || gen != th->th_generation);
506 getboottime(struct timeval *boottime_x)
508 struct bintime boottimebin_x;
510 getboottimebin(&boottimebin_x);
511 bintime2timeval(&boottimebin_x, boottime_x);
515 getboottimebin(struct bintime *boottimebin_x)
517 struct timehands *th;
522 gen = atomic_load_acq_int(&th->th_generation);
523 *boottimebin_x = th->th_boottime;
524 atomic_thread_fence_acq();
525 } while (gen == 0 || gen != th->th_generation);
530 * Support for feed-forward synchronization algorithms. This is heavily inspired
531 * by the timehands mechanism but kept independent from it. *_windup() functions
532 * have some connection to avoid accessing the timecounter hardware more than
536 /* Feed-forward clock estimates kept updated by the synchronization daemon. */
537 struct ffclock_estimate ffclock_estimate;
538 struct bintime ffclock_boottime; /* Feed-forward boot time estimate. */
539 uint32_t ffclock_status; /* Feed-forward clock status. */
540 int8_t ffclock_updated; /* New estimates are available. */
541 struct mtx ffclock_mtx; /* Mutex on ffclock_estimate. */
544 struct ffclock_estimate cest;
545 struct bintime tick_time;
546 struct bintime tick_time_lerp;
547 ffcounter tick_ffcount;
548 uint64_t period_lerp;
549 volatile uint8_t gen;
550 struct fftimehands *next;
553 #define NUM_ELEMENTS(x) (sizeof(x) / sizeof(*x))
555 static struct fftimehands ffth[10];
556 static struct fftimehands *volatile fftimehands = ffth;
561 struct fftimehands *cur;
562 struct fftimehands *last;
564 memset(ffth, 0, sizeof(ffth));
566 last = ffth + NUM_ELEMENTS(ffth) - 1;
567 for (cur = ffth; cur < last; cur++)
572 ffclock_status = FFCLOCK_STA_UNSYNC;
573 mtx_init(&ffclock_mtx, "ffclock lock", NULL, MTX_DEF);
577 * Reset the feed-forward clock estimates. Called from inittodr() to get things
578 * kick started and uses the timecounter nominal frequency as a first period
579 * estimate. Note: this function may be called several time just after boot.
580 * Note: this is the only function that sets the value of boot time for the
581 * monotonic (i.e. uptime) version of the feed-forward clock.
584 ffclock_reset_clock(struct timespec *ts)
586 struct timecounter *tc;
587 struct ffclock_estimate cest;
589 tc = timehands->th_counter;
590 memset(&cest, 0, sizeof(struct ffclock_estimate));
592 timespec2bintime(ts, &ffclock_boottime);
593 timespec2bintime(ts, &(cest.update_time));
594 ffclock_read_counter(&cest.update_ffcount);
595 cest.leapsec_next = 0;
596 cest.period = ((1ULL << 63) / tc->tc_frequency) << 1;
599 cest.status = FFCLOCK_STA_UNSYNC;
600 cest.leapsec_total = 0;
603 mtx_lock(&ffclock_mtx);
604 bcopy(&cest, &ffclock_estimate, sizeof(struct ffclock_estimate));
605 ffclock_updated = INT8_MAX;
606 mtx_unlock(&ffclock_mtx);
608 printf("ffclock reset: %s (%llu Hz), time = %ld.%09lu\n", tc->tc_name,
609 (unsigned long long)tc->tc_frequency, (long)ts->tv_sec,
610 (unsigned long)ts->tv_nsec);
614 * Sub-routine to convert a time interval measured in RAW counter units to time
615 * in seconds stored in bintime format.
616 * NOTE: bintime_mul requires u_int, but the value of the ffcounter may be
617 * larger than the max value of u_int (on 32 bit architecture). Loop to consume
621 ffclock_convert_delta(ffcounter ffdelta, uint64_t period, struct bintime *bt)
624 ffcounter delta, delta_max;
626 delta_max = (1ULL << (8 * sizeof(unsigned int))) - 1;
629 if (ffdelta > delta_max)
635 bintime_mul(&bt2, (unsigned int)delta);
636 bintime_add(bt, &bt2);
638 } while (ffdelta > 0);
642 * Update the fftimehands.
643 * Push the tick ffcount and time(s) forward based on current clock estimate.
644 * The conversion from ffcounter to bintime relies on the difference clock
645 * principle, whose accuracy relies on computing small time intervals. If a new
646 * clock estimate has been passed by the synchronisation daemon, make it
647 * current, and compute the linear interpolation for monotonic time if needed.
650 ffclock_windup(unsigned int delta)
652 struct ffclock_estimate *cest;
653 struct fftimehands *ffth;
654 struct bintime bt, gap_lerp;
657 unsigned int polling;
658 uint8_t forward_jump, ogen;
661 * Pick the next timehand, copy current ffclock estimates and move tick
662 * times and counter forward.
665 ffth = fftimehands->next;
669 bcopy(&fftimehands->cest, cest, sizeof(struct ffclock_estimate));
670 ffdelta = (ffcounter)delta;
671 ffth->period_lerp = fftimehands->period_lerp;
673 ffth->tick_time = fftimehands->tick_time;
674 ffclock_convert_delta(ffdelta, cest->period, &bt);
675 bintime_add(&ffth->tick_time, &bt);
677 ffth->tick_time_lerp = fftimehands->tick_time_lerp;
678 ffclock_convert_delta(ffdelta, ffth->period_lerp, &bt);
679 bintime_add(&ffth->tick_time_lerp, &bt);
681 ffth->tick_ffcount = fftimehands->tick_ffcount + ffdelta;
684 * Assess the status of the clock, if the last update is too old, it is
685 * likely the synchronisation daemon is dead and the clock is free
688 if (ffclock_updated == 0) {
689 ffdelta = ffth->tick_ffcount - cest->update_ffcount;
690 ffclock_convert_delta(ffdelta, cest->period, &bt);
691 if (bt.sec > 2 * FFCLOCK_SKM_SCALE)
692 ffclock_status |= FFCLOCK_STA_UNSYNC;
696 * If available, grab updated clock estimates and make them current.
697 * Recompute time at this tick using the updated estimates. The clock
698 * estimates passed the feed-forward synchronisation daemon may result
699 * in time conversion that is not monotonically increasing (just after
700 * the update). time_lerp is a particular linear interpolation over the
701 * synchronisation algo polling period that ensures monotonicity for the
702 * clock ids requesting it.
704 if (ffclock_updated > 0) {
705 bcopy(&ffclock_estimate, cest, sizeof(struct ffclock_estimate));
706 ffdelta = ffth->tick_ffcount - cest->update_ffcount;
707 ffth->tick_time = cest->update_time;
708 ffclock_convert_delta(ffdelta, cest->period, &bt);
709 bintime_add(&ffth->tick_time, &bt);
711 /* ffclock_reset sets ffclock_updated to INT8_MAX */
712 if (ffclock_updated == INT8_MAX)
713 ffth->tick_time_lerp = ffth->tick_time;
715 if (bintime_cmp(&ffth->tick_time, &ffth->tick_time_lerp, >))
720 bintime_clear(&gap_lerp);
722 gap_lerp = ffth->tick_time;
723 bintime_sub(&gap_lerp, &ffth->tick_time_lerp);
725 gap_lerp = ffth->tick_time_lerp;
726 bintime_sub(&gap_lerp, &ffth->tick_time);
730 * The reset from the RTC clock may be far from accurate, and
731 * reducing the gap between real time and interpolated time
732 * could take a very long time if the interpolated clock insists
733 * on strict monotonicity. The clock is reset under very strict
734 * conditions (kernel time is known to be wrong and
735 * synchronization daemon has been restarted recently.
736 * ffclock_boottime absorbs the jump to ensure boot time is
737 * correct and uptime functions stay consistent.
739 if (((ffclock_status & FFCLOCK_STA_UNSYNC) == FFCLOCK_STA_UNSYNC) &&
740 ((cest->status & FFCLOCK_STA_UNSYNC) == 0) &&
741 ((cest->status & FFCLOCK_STA_WARMUP) == FFCLOCK_STA_WARMUP)) {
743 bintime_add(&ffclock_boottime, &gap_lerp);
745 bintime_sub(&ffclock_boottime, &gap_lerp);
746 ffth->tick_time_lerp = ffth->tick_time;
747 bintime_clear(&gap_lerp);
750 ffclock_status = cest->status;
751 ffth->period_lerp = cest->period;
754 * Compute corrected period used for the linear interpolation of
755 * time. The rate of linear interpolation is capped to 5000PPM
758 if (bintime_isset(&gap_lerp)) {
759 ffdelta = cest->update_ffcount;
760 ffdelta -= fftimehands->cest.update_ffcount;
761 ffclock_convert_delta(ffdelta, cest->period, &bt);
764 bt.frac = 5000000 * (uint64_t)18446744073LL;
765 bintime_mul(&bt, polling);
766 if (bintime_cmp(&gap_lerp, &bt, >))
769 /* Approximate 1 sec by 1-(1/2^64) to ease arithmetic */
771 if (gap_lerp.sec > 0) {
773 frac /= ffdelta / gap_lerp.sec;
775 frac += gap_lerp.frac / ffdelta;
778 ffth->period_lerp += frac;
780 ffth->period_lerp -= frac;
792 * Adjust the fftimehands when the timecounter is changed. Stating the obvious,
793 * the old and new hardware counter cannot be read simultaneously. tc_windup()
794 * does read the two counters 'back to back', but a few cycles are effectively
795 * lost, and not accumulated in tick_ffcount. This is a fairly radical
796 * operation for a feed-forward synchronization daemon, and it is its job to not
797 * pushing irrelevant data to the kernel. Because there is no locking here,
798 * simply force to ignore pending or next update to give daemon a chance to
799 * realize the counter has changed.
802 ffclock_change_tc(struct timehands *th)
804 struct fftimehands *ffth;
805 struct ffclock_estimate *cest;
806 struct timecounter *tc;
810 ffth = fftimehands->next;
815 bcopy(&(fftimehands->cest), cest, sizeof(struct ffclock_estimate));
816 cest->period = ((1ULL << 63) / tc->tc_frequency ) << 1;
819 cest->status |= FFCLOCK_STA_UNSYNC;
821 ffth->tick_ffcount = fftimehands->tick_ffcount;
822 ffth->tick_time_lerp = fftimehands->tick_time_lerp;
823 ffth->tick_time = fftimehands->tick_time;
824 ffth->period_lerp = cest->period;
826 /* Do not lock but ignore next update from synchronization daemon. */
836 * Retrieve feed-forward counter and time of last kernel tick.
839 ffclock_last_tick(ffcounter *ffcount, struct bintime *bt, uint32_t flags)
841 struct fftimehands *ffth;
845 * No locking but check generation has not changed. Also need to make
846 * sure ffdelta is positive, i.e. ffcount > tick_ffcount.
851 if ((flags & FFCLOCK_LERP) == FFCLOCK_LERP)
852 *bt = ffth->tick_time_lerp;
854 *bt = ffth->tick_time;
855 *ffcount = ffth->tick_ffcount;
856 } while (gen == 0 || gen != ffth->gen);
860 * Absolute clock conversion. Low level function to convert ffcounter to
861 * bintime. The ffcounter is converted using the current ffclock period estimate
862 * or the "interpolated period" to ensure monotonicity.
863 * NOTE: this conversion may have been deferred, and the clock updated since the
864 * hardware counter has been read.
867 ffclock_convert_abs(ffcounter ffcount, struct bintime *bt, uint32_t flags)
869 struct fftimehands *ffth;
875 * No locking but check generation has not changed. Also need to make
876 * sure ffdelta is positive, i.e. ffcount > tick_ffcount.
881 if (ffcount > ffth->tick_ffcount)
882 ffdelta = ffcount - ffth->tick_ffcount;
884 ffdelta = ffth->tick_ffcount - ffcount;
886 if ((flags & FFCLOCK_LERP) == FFCLOCK_LERP) {
887 *bt = ffth->tick_time_lerp;
888 ffclock_convert_delta(ffdelta, ffth->period_lerp, &bt2);
890 *bt = ffth->tick_time;
891 ffclock_convert_delta(ffdelta, ffth->cest.period, &bt2);
894 if (ffcount > ffth->tick_ffcount)
895 bintime_add(bt, &bt2);
897 bintime_sub(bt, &bt2);
898 } while (gen == 0 || gen != ffth->gen);
902 * Difference clock conversion.
903 * Low level function to Convert a time interval measured in RAW counter units
904 * into bintime. The difference clock allows measuring small intervals much more
905 * reliably than the absolute clock.
908 ffclock_convert_diff(ffcounter ffdelta, struct bintime *bt)
910 struct fftimehands *ffth;
913 /* No locking but check generation has not changed. */
917 ffclock_convert_delta(ffdelta, ffth->cest.period, bt);
918 } while (gen == 0 || gen != ffth->gen);
922 * Access to current ffcounter value.
925 ffclock_read_counter(ffcounter *ffcount)
927 struct timehands *th;
928 struct fftimehands *ffth;
929 unsigned int gen, delta;
932 * ffclock_windup() called from tc_windup(), safe to rely on
933 * th->th_generation only, for correct delta and ffcounter.
937 gen = atomic_load_acq_int(&th->th_generation);
939 delta = tc_delta(th);
940 *ffcount = ffth->tick_ffcount;
941 atomic_thread_fence_acq();
942 } while (gen == 0 || gen != th->th_generation);
948 binuptime(struct bintime *bt)
951 binuptime_fromclock(bt, sysclock_active);
955 nanouptime(struct timespec *tsp)
958 nanouptime_fromclock(tsp, sysclock_active);
962 microuptime(struct timeval *tvp)
965 microuptime_fromclock(tvp, sysclock_active);
969 bintime(struct bintime *bt)
972 bintime_fromclock(bt, sysclock_active);
976 nanotime(struct timespec *tsp)
979 nanotime_fromclock(tsp, sysclock_active);
983 microtime(struct timeval *tvp)
986 microtime_fromclock(tvp, sysclock_active);
990 getbinuptime(struct bintime *bt)
993 getbinuptime_fromclock(bt, sysclock_active);
997 getnanouptime(struct timespec *tsp)
1000 getnanouptime_fromclock(tsp, sysclock_active);
1004 getmicrouptime(struct timeval *tvp)
1007 getmicrouptime_fromclock(tvp, sysclock_active);
1011 getbintime(struct bintime *bt)
1014 getbintime_fromclock(bt, sysclock_active);
1018 getnanotime(struct timespec *tsp)
1021 getnanotime_fromclock(tsp, sysclock_active);
1025 getmicrotime(struct timeval *tvp)
1028 getmicrouptime_fromclock(tvp, sysclock_active);
1031 #endif /* FFCLOCK */
1034 * This is a clone of getnanotime and used for walltimestamps.
1035 * The dtrace_ prefix prevents fbt from creating probes for
1036 * it so walltimestamp can be safely used in all fbt probes.
1039 dtrace_getnanotime(struct timespec *tsp)
1041 struct timehands *th;
1046 gen = atomic_load_acq_int(&th->th_generation);
1047 *tsp = th->th_nanotime;
1048 atomic_thread_fence_acq();
1049 } while (gen == 0 || gen != th->th_generation);
1053 * System clock currently providing time to the system. Modifiable via sysctl
1054 * when the FFCLOCK option is defined.
1056 int sysclock_active = SYSCLOCK_FBCK;
1058 /* Internal NTP status and error estimates. */
1059 extern int time_status;
1060 extern long time_esterror;
1063 * Take a snapshot of sysclock data which can be used to compare system clocks
1064 * and generate timestamps after the fact.
1067 sysclock_getsnapshot(struct sysclock_snap *clock_snap, int fast)
1069 struct fbclock_info *fbi;
1070 struct timehands *th;
1072 unsigned int delta, gen;
1075 struct fftimehands *ffth;
1076 struct ffclock_info *ffi;
1077 struct ffclock_estimate cest;
1079 ffi = &clock_snap->ff_info;
1082 fbi = &clock_snap->fb_info;
1087 gen = atomic_load_acq_int(&th->th_generation);
1088 fbi->th_scale = th->th_scale;
1089 fbi->tick_time = th->th_offset;
1092 ffi->tick_time = ffth->tick_time_lerp;
1093 ffi->tick_time_lerp = ffth->tick_time_lerp;
1094 ffi->period = ffth->cest.period;
1095 ffi->period_lerp = ffth->period_lerp;
1096 clock_snap->ffcount = ffth->tick_ffcount;
1100 delta = tc_delta(th);
1101 atomic_thread_fence_acq();
1102 } while (gen == 0 || gen != th->th_generation);
1104 clock_snap->delta = delta;
1105 clock_snap->sysclock_active = sysclock_active;
1107 /* Record feedback clock status and error. */
1108 clock_snap->fb_info.status = time_status;
1109 /* XXX: Very crude estimate of feedback clock error. */
1110 bt.sec = time_esterror / 1000000;
1111 bt.frac = ((time_esterror - bt.sec) * 1000000) *
1112 (uint64_t)18446744073709ULL;
1113 clock_snap->fb_info.error = bt;
1117 clock_snap->ffcount += delta;
1119 /* Record feed-forward clock leap second adjustment. */
1120 ffi->leapsec_adjustment = cest.leapsec_total;
1121 if (clock_snap->ffcount > cest.leapsec_next)
1122 ffi->leapsec_adjustment -= cest.leapsec;
1124 /* Record feed-forward clock status and error. */
1125 clock_snap->ff_info.status = cest.status;
1126 ffcount = clock_snap->ffcount - cest.update_ffcount;
1127 ffclock_convert_delta(ffcount, cest.period, &bt);
1128 /* 18446744073709 = int(2^64/1e12), err_bound_rate in [ps/s]. */
1129 bintime_mul(&bt, cest.errb_rate * (uint64_t)18446744073709ULL);
1130 /* 18446744073 = int(2^64 / 1e9), since err_abs in [ns]. */
1131 bintime_addx(&bt, cest.errb_abs * (uint64_t)18446744073ULL);
1132 clock_snap->ff_info.error = bt;
1137 * Convert a sysclock snapshot into a struct bintime based on the specified
1138 * clock source and flags.
1141 sysclock_snap2bintime(struct sysclock_snap *cs, struct bintime *bt,
1142 int whichclock, uint32_t flags)
1144 struct bintime boottimebin_x;
1150 switch (whichclock) {
1152 *bt = cs->fb_info.tick_time;
1154 /* If snapshot was created with !fast, delta will be >0. */
1156 bintime_addx(bt, cs->fb_info.th_scale * cs->delta);
1158 if ((flags & FBCLOCK_UPTIME) == 0) {
1159 getboottimebin(&boottimebin_x);
1160 bintime_add(bt, &boottimebin_x);
1165 if (flags & FFCLOCK_LERP) {
1166 *bt = cs->ff_info.tick_time_lerp;
1167 period = cs->ff_info.period_lerp;
1169 *bt = cs->ff_info.tick_time;
1170 period = cs->ff_info.period;
1173 /* If snapshot was created with !fast, delta will be >0. */
1174 if (cs->delta > 0) {
1175 ffclock_convert_delta(cs->delta, period, &bt2);
1176 bintime_add(bt, &bt2);
1179 /* Leap second adjustment. */
1180 if (flags & FFCLOCK_LEAPSEC)
1181 bt->sec -= cs->ff_info.leapsec_adjustment;
1183 /* Boot time adjustment, for uptime/monotonic clocks. */
1184 if (flags & FFCLOCK_UPTIME)
1185 bintime_sub(bt, &ffclock_boottime);
1197 * Initialize a new timecounter and possibly use it.
1200 tc_init(struct timecounter *tc)
1203 struct sysctl_oid *tc_root;
1205 u = tc->tc_frequency / tc->tc_counter_mask;
1206 /* XXX: We need some margin here, 10% is a guess */
1209 if (u > hz && tc->tc_quality >= 0) {
1210 tc->tc_quality = -2000;
1212 printf("Timecounter \"%s\" frequency %ju Hz",
1213 tc->tc_name, (uintmax_t)tc->tc_frequency);
1214 printf(" -- Insufficient hz, needs at least %u\n", u);
1216 } else if (tc->tc_quality >= 0 || bootverbose) {
1217 printf("Timecounter \"%s\" frequency %ju Hz quality %d\n",
1218 tc->tc_name, (uintmax_t)tc->tc_frequency,
1222 tc->tc_next = timecounters;
1225 * Set up sysctl tree for this counter.
1227 tc_root = SYSCTL_ADD_NODE(NULL,
1228 SYSCTL_STATIC_CHILDREN(_kern_timecounter_tc), OID_AUTO, tc->tc_name,
1229 CTLFLAG_RW, 0, "timecounter description");
1230 SYSCTL_ADD_UINT(NULL, SYSCTL_CHILDREN(tc_root), OID_AUTO,
1231 "mask", CTLFLAG_RD, &(tc->tc_counter_mask), 0,
1232 "mask for implemented bits");
1233 SYSCTL_ADD_PROC(NULL, SYSCTL_CHILDREN(tc_root), OID_AUTO,
1234 "counter", CTLTYPE_UINT | CTLFLAG_RD, tc, sizeof(*tc),
1235 sysctl_kern_timecounter_get, "IU", "current timecounter value");
1236 SYSCTL_ADD_PROC(NULL, SYSCTL_CHILDREN(tc_root), OID_AUTO,
1237 "frequency", CTLTYPE_U64 | CTLFLAG_RD, tc, sizeof(*tc),
1238 sysctl_kern_timecounter_freq, "QU", "timecounter frequency");
1239 SYSCTL_ADD_INT(NULL, SYSCTL_CHILDREN(tc_root), OID_AUTO,
1240 "quality", CTLFLAG_RD, &(tc->tc_quality), 0,
1241 "goodness of time counter");
1243 * Do not automatically switch if the current tc was specifically
1244 * chosen. Never automatically use a timecounter with negative quality.
1245 * Even though we run on the dummy counter, switching here may be
1246 * worse since this timecounter may not be monotonic.
1250 if (tc->tc_quality < 0)
1252 if (tc->tc_quality < timecounter->tc_quality)
1254 if (tc->tc_quality == timecounter->tc_quality &&
1255 tc->tc_frequency < timecounter->tc_frequency)
1257 (void)tc->tc_get_timecount(tc);
1258 (void)tc->tc_get_timecount(tc);
1262 /* Report the frequency of the current timecounter. */
1264 tc_getfrequency(void)
1267 return (timehands->th_counter->tc_frequency);
1271 sleeping_on_old_rtc(struct thread *td)
1275 * td_rtcgen is modified by curthread when it is running,
1276 * and by other threads in this function. By finding the thread
1277 * on a sleepqueue and holding the lock on the sleepqueue
1278 * chain, we guarantee that the thread is not running and that
1279 * modifying td_rtcgen is safe. Setting td_rtcgen to zero informs
1280 * the thread that it was woken due to a real-time clock adjustment.
1281 * (The declaration of td_rtcgen refers to this comment.)
1283 if (td->td_rtcgen != 0 && td->td_rtcgen != rtc_generation) {
1290 static struct mtx tc_setclock_mtx;
1291 MTX_SYSINIT(tc_setclock_init, &tc_setclock_mtx, "tcsetc", MTX_SPIN);
1294 * Step our concept of UTC. This is done by modifying our estimate of
1298 tc_setclock(struct timespec *ts)
1300 struct timespec tbef, taft;
1301 struct bintime bt, bt2;
1303 timespec2bintime(ts, &bt);
1305 mtx_lock_spin(&tc_setclock_mtx);
1306 cpu_tick_calibrate(1);
1308 bintime_sub(&bt, &bt2);
1310 /* XXX fiddle all the little crinkly bits around the fiords... */
1312 mtx_unlock_spin(&tc_setclock_mtx);
1313 getboottimebin(&boottimebin);
1314 bintime2timeval(&boottimebin, &boottime);
1316 /* Avoid rtc_generation == 0, since td_rtcgen == 0 is special. */
1317 atomic_add_rel_int(&rtc_generation, 2);
1318 sleepq_chains_remove_matching(sleeping_on_old_rtc);
1319 if (timestepwarnings) {
1322 "Time stepped from %jd.%09ld to %jd.%09ld (%jd.%09ld)\n",
1323 (intmax_t)tbef.tv_sec, tbef.tv_nsec,
1324 (intmax_t)taft.tv_sec, taft.tv_nsec,
1325 (intmax_t)ts->tv_sec, ts->tv_nsec);
1330 * Initialize the next struct timehands in the ring and make
1331 * it the active timehands. Along the way we might switch to a different
1332 * timecounter and/or do seconds processing in NTP. Slightly magic.
1335 tc_windup(struct bintime *new_boottimebin)
1338 struct timehands *th, *tho;
1340 u_int delta, ncount, ogen;
1345 * Make the next timehands a copy of the current one, but do
1346 * not overwrite the generation or next pointer. While we
1347 * update the contents, the generation must be zero. We need
1348 * to ensure that the zero generation is visible before the
1349 * data updates become visible, which requires release fence.
1350 * For similar reasons, re-reading of the generation after the
1351 * data is read should use acquire fence.
1355 ogen = th->th_generation;
1356 th->th_generation = 0;
1357 atomic_thread_fence_rel();
1358 bcopy(tho, th, offsetof(struct timehands, th_generation));
1359 if (new_boottimebin != NULL)
1360 th->th_boottime = *new_boottimebin;
1363 * Capture a timecounter delta on the current timecounter and if
1364 * changing timecounters, a counter value from the new timecounter.
1365 * Update the offset fields accordingly.
1367 delta = tc_delta(th);
1368 if (th->th_counter != timecounter)
1369 ncount = timecounter->tc_get_timecount(timecounter);
1373 ffclock_windup(delta);
1375 th->th_offset_count += delta;
1376 th->th_offset_count &= th->th_counter->tc_counter_mask;
1377 while (delta > th->th_counter->tc_frequency) {
1378 /* Eat complete unadjusted seconds. */
1379 delta -= th->th_counter->tc_frequency;
1380 th->th_offset.sec++;
1382 if ((delta > th->th_counter->tc_frequency / 2) &&
1383 (th->th_scale * delta < ((uint64_t)1 << 63))) {
1384 /* The product th_scale * delta just barely overflows. */
1385 th->th_offset.sec++;
1387 bintime_addx(&th->th_offset, th->th_scale * delta);
1390 * Hardware latching timecounters may not generate interrupts on
1391 * PPS events, so instead we poll them. There is a finite risk that
1392 * the hardware might capture a count which is later than the one we
1393 * got above, and therefore possibly in the next NTP second which might
1394 * have a different rate than the current NTP second. It doesn't
1395 * matter in practice.
1397 if (tho->th_counter->tc_poll_pps)
1398 tho->th_counter->tc_poll_pps(tho->th_counter);
1401 * Deal with NTP second processing. The for loop normally
1402 * iterates at most once, but in extreme situations it might
1403 * keep NTP sane if timeouts are not run for several seconds.
1404 * At boot, the time step can be large when the TOD hardware
1405 * has been read, so on really large steps, we call
1406 * ntp_update_second only twice. We need to call it twice in
1407 * case we missed a leap second.
1410 bintime_add(&bt, &th->th_boottime);
1411 i = bt.sec - tho->th_microtime.tv_sec;
1414 for (; i > 0; i--) {
1416 ntp_update_second(&th->th_adjustment, &bt.sec);
1418 th->th_boottime.sec += bt.sec - t;
1420 /* Update the UTC timestamps used by the get*() functions. */
1421 th->th_bintime = bt;
1422 bintime2timeval(&bt, &th->th_microtime);
1423 bintime2timespec(&bt, &th->th_nanotime);
1425 /* Now is a good time to change timecounters. */
1426 if (th->th_counter != timecounter) {
1428 if ((timecounter->tc_flags & TC_FLAGS_C2STOP) != 0)
1429 cpu_disable_c2_sleep++;
1430 if ((th->th_counter->tc_flags & TC_FLAGS_C2STOP) != 0)
1431 cpu_disable_c2_sleep--;
1433 th->th_counter = timecounter;
1434 th->th_offset_count = ncount;
1435 tc_min_ticktock_freq = max(1, timecounter->tc_frequency /
1436 (((uint64_t)timecounter->tc_counter_mask + 1) / 3));
1438 ffclock_change_tc(th);
1443 * Recalculate the scaling factor. We want the number of 1/2^64
1444 * fractions of a second per period of the hardware counter, taking
1445 * into account the th_adjustment factor which the NTP PLL/adjtime(2)
1446 * processing provides us with.
1448 * The th_adjustment is nanoseconds per second with 32 bit binary
1449 * fraction and we want 64 bit binary fraction of second:
1451 * x = a * 2^32 / 10^9 = a * 4.294967296
1453 * The range of th_adjustment is +/- 5000PPM so inside a 64bit int
1454 * we can only multiply by about 850 without overflowing, that
1455 * leaves no suitably precise fractions for multiply before divide.
1457 * Divide before multiply with a fraction of 2199/512 results in a
1458 * systematic undercompensation of 10PPM of th_adjustment. On a
1459 * 5000PPM adjustment this is a 0.05PPM error. This is acceptable.
1461 * We happily sacrifice the lowest of the 64 bits of our result
1462 * to the goddess of code clarity.
1465 scale = (uint64_t)1 << 63;
1466 scale += (th->th_adjustment / 1024) * 2199;
1467 scale /= th->th_counter->tc_frequency;
1468 th->th_scale = scale * 2;
1471 * Now that the struct timehands is again consistent, set the new
1472 * generation number, making sure to not make it zero.
1476 atomic_store_rel_int(&th->th_generation, ogen);
1478 /* Go live with the new struct timehands. */
1480 switch (sysclock_active) {
1483 time_second = th->th_microtime.tv_sec;
1484 time_uptime = th->th_offset.sec;
1488 time_second = fftimehands->tick_time_lerp.sec;
1489 time_uptime = fftimehands->tick_time_lerp.sec - ffclock_boottime.sec;
1495 timekeep_push_vdso();
1498 /* Report or change the active timecounter hardware. */
1500 sysctl_kern_timecounter_hardware(SYSCTL_HANDLER_ARGS)
1503 struct timecounter *newtc, *tc;
1507 strlcpy(newname, tc->tc_name, sizeof(newname));
1509 error = sysctl_handle_string(oidp, &newname[0], sizeof(newname), req);
1510 if (error != 0 || req->newptr == NULL)
1512 /* Record that the tc in use now was specifically chosen. */
1514 if (strcmp(newname, tc->tc_name) == 0)
1516 for (newtc = timecounters; newtc != NULL; newtc = newtc->tc_next) {
1517 if (strcmp(newname, newtc->tc_name) != 0)
1520 /* Warm up new timecounter. */
1521 (void)newtc->tc_get_timecount(newtc);
1522 (void)newtc->tc_get_timecount(newtc);
1524 timecounter = newtc;
1527 * The vdso timehands update is deferred until the next
1530 * This is prudent given that 'timekeep_push_vdso()' does not
1531 * use any locking and that it can be called in hard interrupt
1532 * context via 'tc_windup()'.
1539 SYSCTL_PROC(_kern_timecounter, OID_AUTO, hardware, CTLTYPE_STRING | CTLFLAG_RW,
1540 0, 0, sysctl_kern_timecounter_hardware, "A",
1541 "Timecounter hardware selected");
1544 /* Report the available timecounter hardware. */
1546 sysctl_kern_timecounter_choice(SYSCTL_HANDLER_ARGS)
1549 struct timecounter *tc;
1552 sbuf_new_for_sysctl(&sb, NULL, 0, req);
1553 for (tc = timecounters; tc != NULL; tc = tc->tc_next) {
1554 if (tc != timecounters)
1555 sbuf_putc(&sb, ' ');
1556 sbuf_printf(&sb, "%s(%d)", tc->tc_name, tc->tc_quality);
1558 error = sbuf_finish(&sb);
1563 SYSCTL_PROC(_kern_timecounter, OID_AUTO, choice, CTLTYPE_STRING | CTLFLAG_RD,
1564 0, 0, sysctl_kern_timecounter_choice, "A", "Timecounter hardware detected");
1567 * RFC 2783 PPS-API implementation.
1571 * Return true if the driver is aware of the abi version extensions in the
1572 * pps_state structure, and it supports at least the given abi version number.
1575 abi_aware(struct pps_state *pps, int vers)
1578 return ((pps->kcmode & KCMODE_ABIFLAG) && pps->driver_abi >= vers);
1582 pps_fetch(struct pps_fetch_args *fapi, struct pps_state *pps)
1585 pps_seq_t aseq, cseq;
1588 if (fapi->tsformat && fapi->tsformat != PPS_TSFMT_TSPEC)
1592 * If no timeout is requested, immediately return whatever values were
1593 * most recently captured. If timeout seconds is -1, that's a request
1594 * to block without a timeout. WITNESS won't let us sleep forever
1595 * without a lock (we really don't need a lock), so just repeatedly
1596 * sleep a long time.
1598 if (fapi->timeout.tv_sec || fapi->timeout.tv_nsec) {
1599 if (fapi->timeout.tv_sec == -1)
1602 tv.tv_sec = fapi->timeout.tv_sec;
1603 tv.tv_usec = fapi->timeout.tv_nsec / 1000;
1606 aseq = atomic_load_int(&pps->ppsinfo.assert_sequence);
1607 cseq = atomic_load_int(&pps->ppsinfo.clear_sequence);
1608 while (aseq == atomic_load_int(&pps->ppsinfo.assert_sequence) &&
1609 cseq == atomic_load_int(&pps->ppsinfo.clear_sequence)) {
1610 if (abi_aware(pps, 1) && pps->driver_mtx != NULL) {
1611 if (pps->flags & PPSFLAG_MTX_SPIN) {
1612 err = msleep_spin(pps, pps->driver_mtx,
1615 err = msleep(pps, pps->driver_mtx, PCATCH,
1619 err = tsleep(pps, PCATCH, "ppsfch", timo);
1621 if (err == EWOULDBLOCK) {
1622 if (fapi->timeout.tv_sec == -1) {
1627 } else if (err != 0) {
1633 pps->ppsinfo.current_mode = pps->ppsparam.mode;
1634 fapi->pps_info_buf = pps->ppsinfo;
1640 pps_ioctl(u_long cmd, caddr_t data, struct pps_state *pps)
1643 struct pps_fetch_args *fapi;
1645 struct pps_fetch_ffc_args *fapi_ffc;
1648 struct pps_kcbind_args *kapi;
1651 KASSERT(pps != NULL, ("NULL pps pointer in pps_ioctl"));
1653 case PPS_IOC_CREATE:
1655 case PPS_IOC_DESTROY:
1657 case PPS_IOC_SETPARAMS:
1658 app = (pps_params_t *)data;
1659 if (app->mode & ~pps->ppscap)
1662 /* Ensure only a single clock is selected for ffc timestamp. */
1663 if ((app->mode & PPS_TSCLK_MASK) == PPS_TSCLK_MASK)
1666 pps->ppsparam = *app;
1668 case PPS_IOC_GETPARAMS:
1669 app = (pps_params_t *)data;
1670 *app = pps->ppsparam;
1671 app->api_version = PPS_API_VERS_1;
1673 case PPS_IOC_GETCAP:
1674 *(int*)data = pps->ppscap;
1677 fapi = (struct pps_fetch_args *)data;
1678 return (pps_fetch(fapi, pps));
1680 case PPS_IOC_FETCH_FFCOUNTER:
1681 fapi_ffc = (struct pps_fetch_ffc_args *)data;
1682 if (fapi_ffc->tsformat && fapi_ffc->tsformat !=
1685 if (fapi_ffc->timeout.tv_sec || fapi_ffc->timeout.tv_nsec)
1686 return (EOPNOTSUPP);
1687 pps->ppsinfo_ffc.current_mode = pps->ppsparam.mode;
1688 fapi_ffc->pps_info_buf_ffc = pps->ppsinfo_ffc;
1689 /* Overwrite timestamps if feedback clock selected. */
1690 switch (pps->ppsparam.mode & PPS_TSCLK_MASK) {
1691 case PPS_TSCLK_FBCK:
1692 fapi_ffc->pps_info_buf_ffc.assert_timestamp =
1693 pps->ppsinfo.assert_timestamp;
1694 fapi_ffc->pps_info_buf_ffc.clear_timestamp =
1695 pps->ppsinfo.clear_timestamp;
1697 case PPS_TSCLK_FFWD:
1703 #endif /* FFCLOCK */
1704 case PPS_IOC_KCBIND:
1706 kapi = (struct pps_kcbind_args *)data;
1707 /* XXX Only root should be able to do this */
1708 if (kapi->tsformat && kapi->tsformat != PPS_TSFMT_TSPEC)
1710 if (kapi->kernel_consumer != PPS_KC_HARDPPS)
1712 if (kapi->edge & ~pps->ppscap)
1714 pps->kcmode = (kapi->edge & KCMODE_EDGEMASK) |
1715 (pps->kcmode & KCMODE_ABIFLAG);
1718 return (EOPNOTSUPP);
1726 pps_init(struct pps_state *pps)
1728 pps->ppscap |= PPS_TSFMT_TSPEC | PPS_CANWAIT;
1729 if (pps->ppscap & PPS_CAPTUREASSERT)
1730 pps->ppscap |= PPS_OFFSETASSERT;
1731 if (pps->ppscap & PPS_CAPTURECLEAR)
1732 pps->ppscap |= PPS_OFFSETCLEAR;
1734 pps->ppscap |= PPS_TSCLK_MASK;
1736 pps->kcmode &= ~KCMODE_ABIFLAG;
1740 pps_init_abi(struct pps_state *pps)
1744 if (pps->driver_abi > 0) {
1745 pps->kcmode |= KCMODE_ABIFLAG;
1746 pps->kernel_abi = PPS_ABI_VERSION;
1751 pps_capture(struct pps_state *pps)
1753 struct timehands *th;
1755 KASSERT(pps != NULL, ("NULL pps pointer in pps_capture"));
1757 pps->capgen = atomic_load_acq_int(&th->th_generation);
1760 pps->capffth = fftimehands;
1762 pps->capcount = th->th_counter->tc_get_timecount(th->th_counter);
1763 atomic_thread_fence_acq();
1764 if (pps->capgen != th->th_generation)
1769 pps_event(struct pps_state *pps, int event)
1772 struct timespec ts, *tsp, *osp;
1773 u_int tcount, *pcount;
1777 struct timespec *tsp_ffc;
1778 pps_seq_t *pseq_ffc;
1785 KASSERT(pps != NULL, ("NULL pps pointer in pps_event"));
1786 /* Nothing to do if not currently set to capture this event type. */
1787 if ((event & pps->ppsparam.mode) == 0)
1789 /* If the timecounter was wound up underneath us, bail out. */
1790 if (pps->capgen == 0 || pps->capgen !=
1791 atomic_load_acq_int(&pps->capth->th_generation))
1794 /* Things would be easier with arrays. */
1795 if (event == PPS_CAPTUREASSERT) {
1796 tsp = &pps->ppsinfo.assert_timestamp;
1797 osp = &pps->ppsparam.assert_offset;
1798 foff = pps->ppsparam.mode & PPS_OFFSETASSERT;
1800 fhard = pps->kcmode & PPS_CAPTUREASSERT;
1802 pcount = &pps->ppscount[0];
1803 pseq = &pps->ppsinfo.assert_sequence;
1805 ffcount = &pps->ppsinfo_ffc.assert_ffcount;
1806 tsp_ffc = &pps->ppsinfo_ffc.assert_timestamp;
1807 pseq_ffc = &pps->ppsinfo_ffc.assert_sequence;
1810 tsp = &pps->ppsinfo.clear_timestamp;
1811 osp = &pps->ppsparam.clear_offset;
1812 foff = pps->ppsparam.mode & PPS_OFFSETCLEAR;
1814 fhard = pps->kcmode & PPS_CAPTURECLEAR;
1816 pcount = &pps->ppscount[1];
1817 pseq = &pps->ppsinfo.clear_sequence;
1819 ffcount = &pps->ppsinfo_ffc.clear_ffcount;
1820 tsp_ffc = &pps->ppsinfo_ffc.clear_timestamp;
1821 pseq_ffc = &pps->ppsinfo_ffc.clear_sequence;
1826 * If the timecounter changed, we cannot compare the count values, so
1827 * we have to drop the rest of the PPS-stuff until the next event.
1829 if (pps->ppstc != pps->capth->th_counter) {
1830 pps->ppstc = pps->capth->th_counter;
1831 *pcount = pps->capcount;
1832 pps->ppscount[2] = pps->capcount;
1836 /* Convert the count to a timespec. */
1837 tcount = pps->capcount - pps->capth->th_offset_count;
1838 tcount &= pps->capth->th_counter->tc_counter_mask;
1839 bt = pps->capth->th_bintime;
1840 bintime_addx(&bt, pps->capth->th_scale * tcount);
1841 bintime2timespec(&bt, &ts);
1843 /* If the timecounter was wound up underneath us, bail out. */
1844 atomic_thread_fence_acq();
1845 if (pps->capgen != pps->capth->th_generation)
1848 *pcount = pps->capcount;
1853 timespecadd(tsp, osp);
1854 if (tsp->tv_nsec < 0) {
1855 tsp->tv_nsec += 1000000000;
1861 *ffcount = pps->capffth->tick_ffcount + tcount;
1862 bt = pps->capffth->tick_time;
1863 ffclock_convert_delta(tcount, pps->capffth->cest.period, &bt);
1864 bintime_add(&bt, &pps->capffth->tick_time);
1865 bintime2timespec(&bt, &ts);
1875 * Feed the NTP PLL/FLL.
1876 * The FLL wants to know how many (hardware) nanoseconds
1877 * elapsed since the previous event.
1879 tcount = pps->capcount - pps->ppscount[2];
1880 pps->ppscount[2] = pps->capcount;
1881 tcount &= pps->capth->th_counter->tc_counter_mask;
1882 scale = (uint64_t)1 << 63;
1883 scale /= pps->capth->th_counter->tc_frequency;
1887 bintime_addx(&bt, scale * tcount);
1888 bintime2timespec(&bt, &ts);
1889 hardpps(tsp, ts.tv_nsec + 1000000000 * ts.tv_sec);
1893 /* Wakeup anyone sleeping in pps_fetch(). */
1898 * Timecounters need to be updated every so often to prevent the hardware
1899 * counter from overflowing. Updating also recalculates the cached values
1900 * used by the get*() family of functions, so their precision depends on
1901 * the update frequency.
1905 SYSCTL_INT(_kern_timecounter, OID_AUTO, tick, CTLFLAG_RD, &tc_tick, 0,
1906 "Approximate number of hardclock ticks in a millisecond");
1909 tc_ticktock(int cnt)
1913 if (mtx_trylock_spin(&tc_setclock_mtx)) {
1915 if (count >= tc_tick) {
1919 mtx_unlock_spin(&tc_setclock_mtx);
1923 static void __inline
1924 tc_adjprecision(void)
1928 if (tc_timepercentage > 0) {
1929 t = (99 + tc_timepercentage) / tc_timepercentage;
1930 tc_precexp = fls(t + (t >> 1)) - 1;
1931 FREQ2BT(hz / tc_tick, &bt_timethreshold);
1932 FREQ2BT(hz, &bt_tickthreshold);
1933 bintime_shift(&bt_timethreshold, tc_precexp);
1934 bintime_shift(&bt_tickthreshold, tc_precexp);
1937 bt_timethreshold.sec = INT_MAX;
1938 bt_timethreshold.frac = ~(uint64_t)0;
1939 bt_tickthreshold = bt_timethreshold;
1941 sbt_timethreshold = bttosbt(bt_timethreshold);
1942 sbt_tickthreshold = bttosbt(bt_tickthreshold);
1946 sysctl_kern_timecounter_adjprecision(SYSCTL_HANDLER_ARGS)
1950 val = tc_timepercentage;
1951 error = sysctl_handle_int(oidp, &val, 0, req);
1952 if (error != 0 || req->newptr == NULL)
1954 tc_timepercentage = val;
1963 inittimecounter(void *dummy)
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);
1992 /* warm up new timecounter (again) and get rolling. */
1993 (void)timecounter->tc_get_timecount(timecounter);
1994 (void)timecounter->tc_get_timecount(timecounter);
1995 mtx_lock_spin(&tc_setclock_mtx);
1997 mtx_unlock_spin(&tc_setclock_mtx);
2000 SYSINIT(timecounter, SI_SUB_CLOCKS, SI_ORDER_SECOND, inittimecounter, NULL);
2002 /* Cpu tick handling -------------------------------------------------*/
2004 static int cpu_tick_variable;
2005 static uint64_t cpu_tick_frequency;
2007 static DPCPU_DEFINE(uint64_t, tc_cpu_ticks_base);
2008 static DPCPU_DEFINE(unsigned, tc_cpu_ticks_last);
2013 struct timecounter *tc;
2014 uint64_t res, *base;
2018 base = DPCPU_PTR(tc_cpu_ticks_base);
2019 last = DPCPU_PTR(tc_cpu_ticks_last);
2020 tc = timehands->th_counter;
2021 u = tc->tc_get_timecount(tc) & tc->tc_counter_mask;
2023 *base += (uint64_t)tc->tc_counter_mask + 1;
2031 cpu_tick_calibration(void)
2033 static time_t last_calib;
2035 if (time_uptime != last_calib && !(time_uptime & 0xf)) {
2036 cpu_tick_calibrate(0);
2037 last_calib = time_uptime;
2042 * This function gets called every 16 seconds on only one designated
2043 * CPU in the system from hardclock() via cpu_tick_calibration()().
2045 * Whenever the real time clock is stepped we get called with reset=1
2046 * to make sure we handle suspend/resume and similar events correctly.
2050 cpu_tick_calibrate(int reset)
2052 static uint64_t c_last;
2053 uint64_t c_this, c_delta;
2054 static struct bintime t_last;
2055 struct bintime t_this, t_delta;
2059 /* The clock was stepped, abort & reset */
2064 /* we don't calibrate fixed rate cputicks */
2065 if (!cpu_tick_variable)
2068 getbinuptime(&t_this);
2069 c_this = cpu_ticks();
2070 if (t_last.sec != 0) {
2071 c_delta = c_this - c_last;
2073 bintime_sub(&t_delta, &t_last);
2076 * 2^(64-20) / 16[s] =
2078 * 17.592.186.044.416 / 16 =
2079 * 1.099.511.627.776 [Hz]
2081 divi = t_delta.sec << 20;
2082 divi |= t_delta.frac >> (64 - 20);
2085 if (c_delta > cpu_tick_frequency) {
2086 if (0 && bootverbose)
2087 printf("cpu_tick increased to %ju Hz\n",
2089 cpu_tick_frequency = c_delta;
2097 set_cputicker(cpu_tick_f *func, uint64_t freq, unsigned var)
2101 cpu_ticks = tc_cpu_ticks;
2103 cpu_tick_frequency = freq;
2104 cpu_tick_variable = var;
2113 if (cpu_ticks == tc_cpu_ticks)
2114 return (tc_getfrequency());
2115 return (cpu_tick_frequency);
2119 * We need to be slightly careful converting cputicks to microseconds.
2120 * There is plenty of margin in 64 bits of microseconds (half a million
2121 * years) and in 64 bits at 4 GHz (146 years), but if we do a multiply
2122 * before divide conversion (to retain precision) we find that the
2123 * margin shrinks to 1.5 hours (one millionth of 146y).
2124 * With a three prong approach we never lose significant bits, no
2125 * matter what the cputick rate and length of timeinterval is.
2129 cputick2usec(uint64_t tick)
2132 if (tick > 18446744073709551LL) /* floor(2^64 / 1000) */
2133 return (tick / (cpu_tickrate() / 1000000LL));
2134 else if (tick > 18446744073709LL) /* floor(2^64 / 1000000) */
2135 return ((tick * 1000LL) / (cpu_tickrate() / 1000LL));
2137 return ((tick * 1000000LL) / cpu_tickrate());
2140 cpu_tick_f *cpu_ticks = tc_cpu_ticks;
2142 static int vdso_th_enable = 1;
2144 sysctl_fast_gettime(SYSCTL_HANDLER_ARGS)
2146 int old_vdso_th_enable, error;
2148 old_vdso_th_enable = vdso_th_enable;
2149 error = sysctl_handle_int(oidp, &old_vdso_th_enable, 0, req);
2152 vdso_th_enable = old_vdso_th_enable;
2155 SYSCTL_PROC(_kern_timecounter, OID_AUTO, fast_gettime,
2156 CTLTYPE_INT | CTLFLAG_RW | CTLFLAG_MPSAFE,
2157 NULL, 0, sysctl_fast_gettime, "I", "Enable fast time of day");
2160 tc_fill_vdso_timehands(struct vdso_timehands *vdso_th)
2162 struct timehands *th;
2166 vdso_th->th_scale = th->th_scale;
2167 vdso_th->th_offset_count = th->th_offset_count;
2168 vdso_th->th_counter_mask = th->th_counter->tc_counter_mask;
2169 vdso_th->th_offset = th->th_offset;
2170 vdso_th->th_boottime = th->th_boottime;
2171 if (th->th_counter->tc_fill_vdso_timehands != NULL) {
2172 enabled = th->th_counter->tc_fill_vdso_timehands(vdso_th,
2176 if (!vdso_th_enable)
2181 #ifdef COMPAT_FREEBSD32
2183 tc_fill_vdso_timehands32(struct vdso_timehands32 *vdso_th32)
2185 struct timehands *th;
2189 *(uint64_t *)&vdso_th32->th_scale[0] = th->th_scale;
2190 vdso_th32->th_offset_count = th->th_offset_count;
2191 vdso_th32->th_counter_mask = th->th_counter->tc_counter_mask;
2192 vdso_th32->th_offset.sec = th->th_offset.sec;
2193 *(uint64_t *)&vdso_th32->th_offset.frac[0] = th->th_offset.frac;
2194 vdso_th32->th_boottime.sec = th->th_boottime.sec;
2195 *(uint64_t *)&vdso_th32->th_boottime.frac[0] = th->th_boottime.frac;
2196 if (th->th_counter->tc_fill_vdso_timehands32 != NULL) {
2197 enabled = th->th_counter->tc_fill_vdso_timehands32(vdso_th32,
2201 if (!vdso_th_enable)