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
13 * Portions of this software were developed by Julien Ridoux at the University
14 * of Melbourne under sponsorship from the FreeBSD Foundation.
16 * Portions of this software were developed by Konstantin Belousov
17 * under sponsorship from the FreeBSD Foundation.
20 #include <sys/cdefs.h>
22 #include "opt_ffclock.h"
24 #include <sys/param.h>
25 #include <sys/kernel.h>
26 #include <sys/limits.h>
28 #include <sys/mutex.h>
31 #include <sys/sleepqueue.h>
32 #include <sys/sysctl.h>
33 #include <sys/syslog.h>
34 #include <sys/systm.h>
35 #include <sys/timeffc.h>
36 #include <sys/timepps.h>
37 #include <sys/timetc.h>
38 #include <sys/timex.h>
42 * A large step happens on boot. This constant detects such steps.
43 * It is relatively small so that ntp_update_second gets called enough
44 * in the typical 'missed a couple of seconds' case, but doesn't loop
45 * forever when the time step is large.
47 #define LARGE_STEP 200
50 * Implement a dummy timecounter which we can use until we get a real one
51 * in the air. This allows the console and other early stuff to use
56 dummy_get_timecount(struct timecounter *tc)
63 static struct timecounter dummy_timecounter = {
64 dummy_get_timecount, 0, ~0u, 1000000, "dummy", -1000000
68 /* These fields must be initialized by the driver. */
69 struct timecounter *th_counter;
70 int64_t th_adjustment;
73 u_int th_offset_count;
74 struct bintime th_offset;
75 struct bintime th_bintime;
76 struct timeval th_microtime;
77 struct timespec th_nanotime;
78 struct bintime th_boottime;
79 /* Fields not to be copied in tc_windup start with th_generation. */
81 struct timehands *th_next;
84 static struct timehands ths[16] = {
86 .th_counter = &dummy_timecounter,
87 .th_scale = (uint64_t)-1 / 1000000,
88 .th_large_delta = 1000000,
89 .th_offset = { .sec = 1 },
94 static struct timehands *volatile timehands = &ths[0];
95 struct timecounter *timecounter = &dummy_timecounter;
96 static struct timecounter *timecounters = &dummy_timecounter;
98 /* Mutex to protect the timecounter list. */
99 static struct mtx tc_lock;
101 int tc_min_ticktock_freq = 1;
103 volatile time_t time_second = 1;
104 volatile time_t time_uptime = 1;
106 static int sysctl_kern_boottime(SYSCTL_HANDLER_ARGS);
107 SYSCTL_PROC(_kern, KERN_BOOTTIME, boottime,
108 CTLTYPE_STRUCT | CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, 0,
109 sysctl_kern_boottime, "S,timeval",
112 SYSCTL_NODE(_kern, OID_AUTO, timecounter, CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
114 static SYSCTL_NODE(_kern_timecounter, OID_AUTO, tc,
115 CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
118 static int timestepwarnings;
119 SYSCTL_INT(_kern_timecounter, OID_AUTO, stepwarnings, CTLFLAG_RWTUN,
120 ×tepwarnings, 0, "Log time steps");
122 static int timehands_count = 2;
123 SYSCTL_INT(_kern_timecounter, OID_AUTO, timehands_count,
124 CTLFLAG_RDTUN | CTLFLAG_NOFETCH,
125 &timehands_count, 0, "Count of timehands in rotation");
127 struct bintime bt_timethreshold;
128 struct bintime bt_tickthreshold;
129 sbintime_t sbt_timethreshold;
130 sbintime_t sbt_tickthreshold;
131 struct bintime tc_tick_bt;
132 sbintime_t tc_tick_sbt;
134 int tc_timepercentage = TC_DEFAULTPERC;
135 static int sysctl_kern_timecounter_adjprecision(SYSCTL_HANDLER_ARGS);
136 SYSCTL_PROC(_kern_timecounter, OID_AUTO, alloweddeviation,
137 CTLTYPE_INT | CTLFLAG_RWTUN | CTLFLAG_MPSAFE, 0, 0,
138 sysctl_kern_timecounter_adjprecision, "I",
139 "Allowed time interval deviation in percents");
141 volatile int rtc_generation = 1;
143 static int tc_chosen; /* Non-zero if a specific tc was chosen via sysctl. */
144 static char tc_from_tunable[16];
146 static void tc_windup(struct bintime *new_boottimebin);
147 static void cpu_tick_calibrate(int);
149 void dtrace_getnanotime(struct timespec *tsp);
150 void dtrace_getnanouptime(struct timespec *tsp);
153 sysctl_kern_boottime(SYSCTL_HANDLER_ARGS)
155 struct timeval boottime;
157 getboottime(&boottime);
159 /* i386 is the only arch which uses a 32bits time_t */
164 if (req->flags & SCTL_MASK32) {
165 tv[0] = boottime.tv_sec;
166 tv[1] = boottime.tv_usec;
167 return (SYSCTL_OUT(req, tv, sizeof(tv)));
171 return (SYSCTL_OUT(req, &boottime, sizeof(boottime)));
175 sysctl_kern_timecounter_get(SYSCTL_HANDLER_ARGS)
178 struct timecounter *tc = arg1;
180 ncount = tc->tc_get_timecount(tc);
181 return (sysctl_handle_int(oidp, &ncount, 0, req));
185 sysctl_kern_timecounter_freq(SYSCTL_HANDLER_ARGS)
188 struct timecounter *tc = arg1;
190 freq = tc->tc_frequency;
191 return (sysctl_handle_64(oidp, &freq, 0, req));
195 * Return the difference between the timehands' counter value now and what
196 * was when we copied it to the timehands' offset_count.
198 static __inline u_int
199 tc_delta(struct timehands *th)
201 struct timecounter *tc;
204 return ((tc->tc_get_timecount(tc) - th->th_offset_count) &
205 tc->tc_counter_mask);
209 bintime_add_tc_delta(struct bintime *bt, uint64_t scale,
210 uint64_t large_delta, uint64_t delta)
214 if (__predict_false(delta >= large_delta)) {
215 /* Avoid overflow for scale * delta. */
216 x = (scale >> 32) * delta;
218 bintime_addx(bt, x << 32);
219 bintime_addx(bt, (scale & 0xffffffff) * delta);
221 bintime_addx(bt, scale * delta);
226 * Functions for reading the time. We have to loop until we are sure that
227 * the timehands that we operated on was not updated under our feet. See
228 * the comment in <sys/time.h> for a description of these 12 functions.
232 bintime_off(struct bintime *bt, u_int off)
234 struct timehands *th;
237 u_int delta, gen, large_delta;
241 gen = atomic_load_acq_int(&th->th_generation);
242 btp = (struct bintime *)((vm_offset_t)th + off);
244 scale = th->th_scale;
245 delta = tc_delta(th);
246 large_delta = th->th_large_delta;
247 atomic_thread_fence_acq();
248 } while (gen == 0 || gen != th->th_generation);
250 bintime_add_tc_delta(bt, scale, large_delta, delta);
252 #define GETTHBINTIME(dst, member) \
254 _Static_assert(_Generic(((struct timehands *)NULL)->member, \
255 struct bintime: 1, default: 0) == 1, \
256 "struct timehands member is not of struct bintime type"); \
257 bintime_off(dst, __offsetof(struct timehands, member)); \
261 getthmember(void *out, size_t out_size, u_int off)
263 struct timehands *th;
268 gen = atomic_load_acq_int(&th->th_generation);
269 memcpy(out, (char *)th + off, out_size);
270 atomic_thread_fence_acq();
271 } while (gen == 0 || gen != th->th_generation);
273 #define GETTHMEMBER(dst, member) \
275 _Static_assert(_Generic(*dst, \
276 __typeof(((struct timehands *)NULL)->member): 1, \
278 "*dst and struct timehands member have different types"); \
279 getthmember(dst, sizeof(*dst), __offsetof(struct timehands, \
285 fbclock_binuptime(struct bintime *bt)
288 GETTHBINTIME(bt, th_offset);
292 fbclock_nanouptime(struct timespec *tsp)
296 fbclock_binuptime(&bt);
297 bintime2timespec(&bt, tsp);
301 fbclock_microuptime(struct timeval *tvp)
305 fbclock_binuptime(&bt);
306 bintime2timeval(&bt, tvp);
310 fbclock_bintime(struct bintime *bt)
313 GETTHBINTIME(bt, th_bintime);
317 fbclock_nanotime(struct timespec *tsp)
321 fbclock_bintime(&bt);
322 bintime2timespec(&bt, tsp);
326 fbclock_microtime(struct timeval *tvp)
330 fbclock_bintime(&bt);
331 bintime2timeval(&bt, tvp);
335 fbclock_getbinuptime(struct bintime *bt)
338 GETTHMEMBER(bt, th_offset);
342 fbclock_getnanouptime(struct timespec *tsp)
346 GETTHMEMBER(&bt, th_offset);
347 bintime2timespec(&bt, tsp);
351 fbclock_getmicrouptime(struct timeval *tvp)
355 GETTHMEMBER(&bt, th_offset);
356 bintime2timeval(&bt, tvp);
360 fbclock_getbintime(struct bintime *bt)
363 GETTHMEMBER(bt, th_bintime);
367 fbclock_getnanotime(struct timespec *tsp)
370 GETTHMEMBER(tsp, th_nanotime);
374 fbclock_getmicrotime(struct timeval *tvp)
377 GETTHMEMBER(tvp, th_microtime);
382 binuptime(struct bintime *bt)
385 GETTHBINTIME(bt, th_offset);
389 nanouptime(struct timespec *tsp)
394 bintime2timespec(&bt, tsp);
398 microuptime(struct timeval *tvp)
403 bintime2timeval(&bt, tvp);
407 bintime(struct bintime *bt)
410 GETTHBINTIME(bt, th_bintime);
414 nanotime(struct timespec *tsp)
419 bintime2timespec(&bt, tsp);
423 microtime(struct timeval *tvp)
428 bintime2timeval(&bt, tvp);
432 getbinuptime(struct bintime *bt)
435 GETTHMEMBER(bt, th_offset);
439 getnanouptime(struct timespec *tsp)
443 GETTHMEMBER(&bt, th_offset);
444 bintime2timespec(&bt, tsp);
448 getmicrouptime(struct timeval *tvp)
452 GETTHMEMBER(&bt, th_offset);
453 bintime2timeval(&bt, tvp);
457 getbintime(struct bintime *bt)
460 GETTHMEMBER(bt, th_bintime);
464 getnanotime(struct timespec *tsp)
467 GETTHMEMBER(tsp, th_nanotime);
471 getmicrotime(struct timeval *tvp)
474 GETTHMEMBER(tvp, th_microtime);
479 getboottime(struct timeval *boottime)
481 struct bintime boottimebin;
483 getboottimebin(&boottimebin);
484 bintime2timeval(&boottimebin, boottime);
488 getboottimebin(struct bintime *boottimebin)
491 GETTHMEMBER(boottimebin, th_boottime);
496 * Support for feed-forward synchronization algorithms. This is heavily inspired
497 * by the timehands mechanism but kept independent from it. *_windup() functions
498 * have some connection to avoid accessing the timecounter hardware more than
502 /* Feed-forward clock estimates kept updated by the synchronization daemon. */
503 struct ffclock_estimate ffclock_estimate;
504 struct bintime ffclock_boottime; /* Feed-forward boot time estimate. */
505 uint32_t ffclock_status; /* Feed-forward clock status. */
506 int8_t ffclock_updated; /* New estimates are available. */
507 struct mtx ffclock_mtx; /* Mutex on ffclock_estimate. */
510 struct ffclock_estimate cest;
511 struct bintime tick_time;
512 struct bintime tick_time_lerp;
513 ffcounter tick_ffcount;
514 uint64_t period_lerp;
515 volatile uint8_t gen;
516 struct fftimehands *next;
519 #define NUM_ELEMENTS(x) (sizeof(x) / sizeof(*x))
521 static struct fftimehands ffth[10];
522 static struct fftimehands *volatile fftimehands = ffth;
527 struct fftimehands *cur;
528 struct fftimehands *last;
530 memset(ffth, 0, sizeof(ffth));
532 last = ffth + NUM_ELEMENTS(ffth) - 1;
533 for (cur = ffth; cur < last; cur++)
538 ffclock_status = FFCLOCK_STA_UNSYNC;
539 mtx_init(&ffclock_mtx, "ffclock lock", NULL, MTX_DEF);
543 * Reset the feed-forward clock estimates. Called from inittodr() to get things
544 * kick started and uses the timecounter nominal frequency as a first period
545 * estimate. Note: this function may be called several time just after boot.
546 * Note: this is the only function that sets the value of boot time for the
547 * monotonic (i.e. uptime) version of the feed-forward clock.
550 ffclock_reset_clock(struct timespec *ts)
552 struct timecounter *tc;
553 struct ffclock_estimate cest;
555 tc = timehands->th_counter;
556 memset(&cest, 0, sizeof(struct ffclock_estimate));
558 timespec2bintime(ts, &ffclock_boottime);
559 timespec2bintime(ts, &(cest.update_time));
560 ffclock_read_counter(&cest.update_ffcount);
561 cest.leapsec_next = 0;
562 cest.period = ((1ULL << 63) / tc->tc_frequency) << 1;
565 cest.status = FFCLOCK_STA_UNSYNC;
566 cest.leapsec_total = 0;
569 mtx_lock(&ffclock_mtx);
570 bcopy(&cest, &ffclock_estimate, sizeof(struct ffclock_estimate));
571 ffclock_updated = INT8_MAX;
572 mtx_unlock(&ffclock_mtx);
574 printf("ffclock reset: %s (%llu Hz), time = %ld.%09lu\n", tc->tc_name,
575 (unsigned long long)tc->tc_frequency, (long)ts->tv_sec,
576 (unsigned long)ts->tv_nsec);
580 * Sub-routine to convert a time interval measured in RAW counter units to time
581 * in seconds stored in bintime format.
582 * NOTE: bintime_mul requires u_int, but the value of the ffcounter may be
583 * larger than the max value of u_int (on 32 bit architecture). Loop to consume
587 ffclock_convert_delta(ffcounter ffdelta, uint64_t period, struct bintime *bt)
590 ffcounter delta, delta_max;
592 delta_max = (1ULL << (8 * sizeof(unsigned int))) - 1;
595 if (ffdelta > delta_max)
601 bintime_mul(&bt2, (unsigned int)delta);
602 bintime_add(bt, &bt2);
604 } while (ffdelta > 0);
608 * Update the fftimehands.
609 * Push the tick ffcount and time(s) forward based on current clock estimate.
610 * The conversion from ffcounter to bintime relies on the difference clock
611 * principle, whose accuracy relies on computing small time intervals. If a new
612 * clock estimate has been passed by the synchronisation daemon, make it
613 * current, and compute the linear interpolation for monotonic time if needed.
616 ffclock_windup(unsigned int delta)
618 struct ffclock_estimate *cest;
619 struct fftimehands *ffth;
620 struct bintime bt, gap_lerp;
623 unsigned int polling;
624 uint8_t forward_jump, ogen;
627 * Pick the next timehand, copy current ffclock estimates and move tick
628 * times and counter forward.
631 ffth = fftimehands->next;
635 bcopy(&fftimehands->cest, cest, sizeof(struct ffclock_estimate));
636 ffdelta = (ffcounter)delta;
637 ffth->period_lerp = fftimehands->period_lerp;
639 ffth->tick_time = fftimehands->tick_time;
640 ffclock_convert_delta(ffdelta, cest->period, &bt);
641 bintime_add(&ffth->tick_time, &bt);
643 ffth->tick_time_lerp = fftimehands->tick_time_lerp;
644 ffclock_convert_delta(ffdelta, ffth->period_lerp, &bt);
645 bintime_add(&ffth->tick_time_lerp, &bt);
647 ffth->tick_ffcount = fftimehands->tick_ffcount + ffdelta;
650 * Assess the status of the clock, if the last update is too old, it is
651 * likely the synchronisation daemon is dead and the clock is free
654 if (ffclock_updated == 0) {
655 ffdelta = ffth->tick_ffcount - cest->update_ffcount;
656 ffclock_convert_delta(ffdelta, cest->period, &bt);
657 if (bt.sec > 2 * FFCLOCK_SKM_SCALE)
658 ffclock_status |= FFCLOCK_STA_UNSYNC;
662 * If available, grab updated clock estimates and make them current.
663 * Recompute time at this tick using the updated estimates. The clock
664 * estimates passed the feed-forward synchronisation daemon may result
665 * in time conversion that is not monotonically increasing (just after
666 * the update). time_lerp is a particular linear interpolation over the
667 * synchronisation algo polling period that ensures monotonicity for the
668 * clock ids requesting it.
670 if (ffclock_updated > 0) {
671 bcopy(&ffclock_estimate, cest, sizeof(struct ffclock_estimate));
672 ffdelta = ffth->tick_ffcount - cest->update_ffcount;
673 ffth->tick_time = cest->update_time;
674 ffclock_convert_delta(ffdelta, cest->period, &bt);
675 bintime_add(&ffth->tick_time, &bt);
677 /* ffclock_reset sets ffclock_updated to INT8_MAX */
678 if (ffclock_updated == INT8_MAX)
679 ffth->tick_time_lerp = ffth->tick_time;
681 if (bintime_cmp(&ffth->tick_time, &ffth->tick_time_lerp, >))
686 bintime_clear(&gap_lerp);
688 gap_lerp = ffth->tick_time;
689 bintime_sub(&gap_lerp, &ffth->tick_time_lerp);
691 gap_lerp = ffth->tick_time_lerp;
692 bintime_sub(&gap_lerp, &ffth->tick_time);
696 * The reset from the RTC clock may be far from accurate, and
697 * reducing the gap between real time and interpolated time
698 * could take a very long time if the interpolated clock insists
699 * on strict monotonicity. The clock is reset under very strict
700 * conditions (kernel time is known to be wrong and
701 * synchronization daemon has been restarted recently.
702 * ffclock_boottime absorbs the jump to ensure boot time is
703 * correct and uptime functions stay consistent.
705 if (((ffclock_status & FFCLOCK_STA_UNSYNC) == FFCLOCK_STA_UNSYNC) &&
706 ((cest->status & FFCLOCK_STA_UNSYNC) == 0) &&
707 ((cest->status & FFCLOCK_STA_WARMUP) == FFCLOCK_STA_WARMUP)) {
709 bintime_add(&ffclock_boottime, &gap_lerp);
711 bintime_sub(&ffclock_boottime, &gap_lerp);
712 ffth->tick_time_lerp = ffth->tick_time;
713 bintime_clear(&gap_lerp);
716 ffclock_status = cest->status;
717 ffth->period_lerp = cest->period;
720 * Compute corrected period used for the linear interpolation of
721 * time. The rate of linear interpolation is capped to 5000PPM
724 if (bintime_isset(&gap_lerp)) {
725 ffdelta = cest->update_ffcount;
726 ffdelta -= fftimehands->cest.update_ffcount;
727 ffclock_convert_delta(ffdelta, cest->period, &bt);
730 bt.frac = 5000000 * (uint64_t)18446744073LL;
731 bintime_mul(&bt, polling);
732 if (bintime_cmp(&gap_lerp, &bt, >))
735 /* Approximate 1 sec by 1-(1/2^64) to ease arithmetic */
737 if (gap_lerp.sec > 0) {
739 frac /= ffdelta / gap_lerp.sec;
741 frac += gap_lerp.frac / ffdelta;
744 ffth->period_lerp += frac;
746 ffth->period_lerp -= frac;
758 * Adjust the fftimehands when the timecounter is changed. Stating the obvious,
759 * the old and new hardware counter cannot be read simultaneously. tc_windup()
760 * does read the two counters 'back to back', but a few cycles are effectively
761 * lost, and not accumulated in tick_ffcount. This is a fairly radical
762 * operation for a feed-forward synchronization daemon, and it is its job to not
763 * pushing irrelevant data to the kernel. Because there is no locking here,
764 * simply force to ignore pending or next update to give daemon a chance to
765 * realize the counter has changed.
768 ffclock_change_tc(struct timehands *th)
770 struct fftimehands *ffth;
771 struct ffclock_estimate *cest;
772 struct timecounter *tc;
776 ffth = fftimehands->next;
781 bcopy(&(fftimehands->cest), cest, sizeof(struct ffclock_estimate));
782 cest->period = ((1ULL << 63) / tc->tc_frequency ) << 1;
785 cest->status |= FFCLOCK_STA_UNSYNC;
787 ffth->tick_ffcount = fftimehands->tick_ffcount;
788 ffth->tick_time_lerp = fftimehands->tick_time_lerp;
789 ffth->tick_time = fftimehands->tick_time;
790 ffth->period_lerp = cest->period;
792 /* Do not lock but ignore next update from synchronization daemon. */
802 * Retrieve feed-forward counter and time of last kernel tick.
805 ffclock_last_tick(ffcounter *ffcount, struct bintime *bt, uint32_t flags)
807 struct fftimehands *ffth;
811 * No locking but check generation has not changed. Also need to make
812 * sure ffdelta is positive, i.e. ffcount > tick_ffcount.
817 if ((flags & FFCLOCK_LERP) == FFCLOCK_LERP)
818 *bt = ffth->tick_time_lerp;
820 *bt = ffth->tick_time;
821 *ffcount = ffth->tick_ffcount;
822 } while (gen == 0 || gen != ffth->gen);
826 * Absolute clock conversion. Low level function to convert ffcounter to
827 * bintime. The ffcounter is converted using the current ffclock period estimate
828 * or the "interpolated period" to ensure monotonicity.
829 * NOTE: this conversion may have been deferred, and the clock updated since the
830 * hardware counter has been read.
833 ffclock_convert_abs(ffcounter ffcount, struct bintime *bt, uint32_t flags)
835 struct fftimehands *ffth;
841 * No locking but check generation has not changed. Also need to make
842 * sure ffdelta is positive, i.e. ffcount > tick_ffcount.
847 if (ffcount > ffth->tick_ffcount)
848 ffdelta = ffcount - ffth->tick_ffcount;
850 ffdelta = ffth->tick_ffcount - ffcount;
852 if ((flags & FFCLOCK_LERP) == FFCLOCK_LERP) {
853 *bt = ffth->tick_time_lerp;
854 ffclock_convert_delta(ffdelta, ffth->period_lerp, &bt2);
856 *bt = ffth->tick_time;
857 ffclock_convert_delta(ffdelta, ffth->cest.period, &bt2);
860 if (ffcount > ffth->tick_ffcount)
861 bintime_add(bt, &bt2);
863 bintime_sub(bt, &bt2);
864 } while (gen == 0 || gen != ffth->gen);
868 * Difference clock conversion.
869 * Low level function to Convert a time interval measured in RAW counter units
870 * into bintime. The difference clock allows measuring small intervals much more
871 * reliably than the absolute clock.
874 ffclock_convert_diff(ffcounter ffdelta, struct bintime *bt)
876 struct fftimehands *ffth;
879 /* No locking but check generation has not changed. */
883 ffclock_convert_delta(ffdelta, ffth->cest.period, bt);
884 } while (gen == 0 || gen != ffth->gen);
888 * Access to current ffcounter value.
891 ffclock_read_counter(ffcounter *ffcount)
893 struct timehands *th;
894 struct fftimehands *ffth;
895 unsigned int gen, delta;
898 * ffclock_windup() called from tc_windup(), safe to rely on
899 * th->th_generation only, for correct delta and ffcounter.
903 gen = atomic_load_acq_int(&th->th_generation);
905 delta = tc_delta(th);
906 *ffcount = ffth->tick_ffcount;
907 atomic_thread_fence_acq();
908 } while (gen == 0 || gen != th->th_generation);
914 binuptime(struct bintime *bt)
917 binuptime_fromclock(bt, sysclock_active);
921 nanouptime(struct timespec *tsp)
924 nanouptime_fromclock(tsp, sysclock_active);
928 microuptime(struct timeval *tvp)
931 microuptime_fromclock(tvp, sysclock_active);
935 bintime(struct bintime *bt)
938 bintime_fromclock(bt, sysclock_active);
942 nanotime(struct timespec *tsp)
945 nanotime_fromclock(tsp, sysclock_active);
949 microtime(struct timeval *tvp)
952 microtime_fromclock(tvp, sysclock_active);
956 getbinuptime(struct bintime *bt)
959 getbinuptime_fromclock(bt, sysclock_active);
963 getnanouptime(struct timespec *tsp)
966 getnanouptime_fromclock(tsp, sysclock_active);
970 getmicrouptime(struct timeval *tvp)
973 getmicrouptime_fromclock(tvp, sysclock_active);
977 getbintime(struct bintime *bt)
980 getbintime_fromclock(bt, sysclock_active);
984 getnanotime(struct timespec *tsp)
987 getnanotime_fromclock(tsp, sysclock_active);
991 getmicrotime(struct timeval *tvp)
994 getmicrouptime_fromclock(tvp, sysclock_active);
1000 * This is a clone of getnanotime and used for walltimestamps.
1001 * The dtrace_ prefix prevents fbt from creating probes for
1002 * it so walltimestamp can be safely used in all fbt probes.
1005 dtrace_getnanotime(struct timespec *tsp)
1008 GETTHMEMBER(tsp, th_nanotime);
1012 * This is a clone of getnanouptime used for time since boot.
1013 * The dtrace_ prefix prevents fbt from creating probes for
1014 * it so an uptime that can be safely used in all fbt probes.
1017 dtrace_getnanouptime(struct timespec *tsp)
1021 GETTHMEMBER(&bt, th_offset);
1022 bintime2timespec(&bt, tsp);
1026 * System clock currently providing time to the system. Modifiable via sysctl
1027 * when the FFCLOCK option is defined.
1029 int sysclock_active = SYSCLOCK_FBCK;
1031 /* Internal NTP status and error estimates. */
1032 extern int time_status;
1033 extern long time_esterror;
1036 * Take a snapshot of sysclock data which can be used to compare system clocks
1037 * and generate timestamps after the fact.
1040 sysclock_getsnapshot(struct sysclock_snap *clock_snap, int fast)
1042 struct fbclock_info *fbi;
1043 struct timehands *th;
1045 unsigned int delta, gen;
1048 struct fftimehands *ffth;
1049 struct ffclock_info *ffi;
1050 struct ffclock_estimate cest;
1052 ffi = &clock_snap->ff_info;
1055 fbi = &clock_snap->fb_info;
1060 gen = atomic_load_acq_int(&th->th_generation);
1061 fbi->th_scale = th->th_scale;
1062 fbi->tick_time = th->th_offset;
1065 ffi->tick_time = ffth->tick_time_lerp;
1066 ffi->tick_time_lerp = ffth->tick_time_lerp;
1067 ffi->period = ffth->cest.period;
1068 ffi->period_lerp = ffth->period_lerp;
1069 clock_snap->ffcount = ffth->tick_ffcount;
1073 delta = tc_delta(th);
1074 atomic_thread_fence_acq();
1075 } while (gen == 0 || gen != th->th_generation);
1077 clock_snap->delta = delta;
1078 clock_snap->sysclock_active = sysclock_active;
1080 /* Record feedback clock status and error. */
1081 clock_snap->fb_info.status = time_status;
1082 /* XXX: Very crude estimate of feedback clock error. */
1083 bt.sec = time_esterror / 1000000;
1084 bt.frac = ((time_esterror - bt.sec) * 1000000) *
1085 (uint64_t)18446744073709ULL;
1086 clock_snap->fb_info.error = bt;
1090 clock_snap->ffcount += delta;
1092 /* Record feed-forward clock leap second adjustment. */
1093 ffi->leapsec_adjustment = cest.leapsec_total;
1094 if (clock_snap->ffcount > cest.leapsec_next)
1095 ffi->leapsec_adjustment -= cest.leapsec;
1097 /* Record feed-forward clock status and error. */
1098 clock_snap->ff_info.status = cest.status;
1099 ffcount = clock_snap->ffcount - cest.update_ffcount;
1100 ffclock_convert_delta(ffcount, cest.period, &bt);
1101 /* 18446744073709 = int(2^64/1e12), err_bound_rate in [ps/s]. */
1102 bintime_mul(&bt, cest.errb_rate * (uint64_t)18446744073709ULL);
1103 /* 18446744073 = int(2^64 / 1e9), since err_abs in [ns]. */
1104 bintime_addx(&bt, cest.errb_abs * (uint64_t)18446744073ULL);
1105 clock_snap->ff_info.error = bt;
1110 * Convert a sysclock snapshot into a struct bintime based on the specified
1111 * clock source and flags.
1114 sysclock_snap2bintime(struct sysclock_snap *cs, struct bintime *bt,
1115 int whichclock, uint32_t flags)
1117 struct bintime boottimebin;
1123 switch (whichclock) {
1125 *bt = cs->fb_info.tick_time;
1127 /* If snapshot was created with !fast, delta will be >0. */
1129 bintime_addx(bt, cs->fb_info.th_scale * cs->delta);
1131 if ((flags & FBCLOCK_UPTIME) == 0) {
1132 getboottimebin(&boottimebin);
1133 bintime_add(bt, &boottimebin);
1138 if (flags & FFCLOCK_LERP) {
1139 *bt = cs->ff_info.tick_time_lerp;
1140 period = cs->ff_info.period_lerp;
1142 *bt = cs->ff_info.tick_time;
1143 period = cs->ff_info.period;
1146 /* If snapshot was created with !fast, delta will be >0. */
1147 if (cs->delta > 0) {
1148 ffclock_convert_delta(cs->delta, period, &bt2);
1149 bintime_add(bt, &bt2);
1152 /* Leap second adjustment. */
1153 if (flags & FFCLOCK_LEAPSEC)
1154 bt->sec -= cs->ff_info.leapsec_adjustment;
1156 /* Boot time adjustment, for uptime/monotonic clocks. */
1157 if (flags & FFCLOCK_UPTIME)
1158 bintime_sub(bt, &ffclock_boottime);
1170 * Initialize a new timecounter and possibly use it.
1173 tc_init(struct timecounter *tc)
1176 struct sysctl_oid *tc_root;
1178 u = tc->tc_frequency / tc->tc_counter_mask;
1179 /* XXX: We need some margin here, 10% is a guess */
1182 if (u > hz && tc->tc_quality >= 0) {
1183 tc->tc_quality = -2000;
1185 printf("Timecounter \"%s\" frequency %ju Hz",
1186 tc->tc_name, (uintmax_t)tc->tc_frequency);
1187 printf(" -- Insufficient hz, needs at least %u\n", u);
1189 } else if (tc->tc_quality >= 0 || bootverbose) {
1190 printf("Timecounter \"%s\" frequency %ju Hz quality %d\n",
1191 tc->tc_name, (uintmax_t)tc->tc_frequency,
1196 * Set up sysctl tree for this counter.
1198 tc_root = SYSCTL_ADD_NODE_WITH_LABEL(NULL,
1199 SYSCTL_STATIC_CHILDREN(_kern_timecounter_tc), OID_AUTO, tc->tc_name,
1200 CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
1201 "timecounter description", "timecounter");
1202 SYSCTL_ADD_UINT(NULL, SYSCTL_CHILDREN(tc_root), OID_AUTO,
1203 "mask", CTLFLAG_RD, &(tc->tc_counter_mask), 0,
1204 "mask for implemented bits");
1205 SYSCTL_ADD_PROC(NULL, SYSCTL_CHILDREN(tc_root), OID_AUTO,
1206 "counter", CTLTYPE_UINT | CTLFLAG_RD | CTLFLAG_MPSAFE, tc,
1207 sizeof(*tc), sysctl_kern_timecounter_get, "IU",
1208 "current timecounter value");
1209 SYSCTL_ADD_PROC(NULL, SYSCTL_CHILDREN(tc_root), OID_AUTO,
1210 "frequency", CTLTYPE_U64 | CTLFLAG_RD | CTLFLAG_MPSAFE, tc,
1211 sizeof(*tc), sysctl_kern_timecounter_freq, "QU",
1212 "timecounter frequency");
1213 SYSCTL_ADD_INT(NULL, SYSCTL_CHILDREN(tc_root), OID_AUTO,
1214 "quality", CTLFLAG_RD, &(tc->tc_quality), 0,
1215 "goodness of time counter");
1218 tc->tc_next = timecounters;
1222 * Do not automatically switch if the current tc was specifically
1223 * chosen. Never automatically use a timecounter with negative quality.
1224 * Even though we run on the dummy counter, switching here may be
1225 * worse since this timecounter may not be monotonic.
1229 if (tc->tc_quality < 0)
1231 if (tc_from_tunable[0] != '\0' &&
1232 strcmp(tc->tc_name, tc_from_tunable) == 0) {
1234 tc_from_tunable[0] = '\0';
1236 if (tc->tc_quality < timecounter->tc_quality)
1238 if (tc->tc_quality == timecounter->tc_quality &&
1239 tc->tc_frequency < timecounter->tc_frequency)
1242 (void)tc->tc_get_timecount(tc);
1245 mtx_unlock(&tc_lock);
1248 /* Report the frequency of the current timecounter. */
1250 tc_getfrequency(void)
1253 return (timehands->th_counter->tc_frequency);
1257 sleeping_on_old_rtc(struct thread *td)
1261 * td_rtcgen is modified by curthread when it is running,
1262 * and by other threads in this function. By finding the thread
1263 * on a sleepqueue and holding the lock on the sleepqueue
1264 * chain, we guarantee that the thread is not running and that
1265 * modifying td_rtcgen is safe. Setting td_rtcgen to zero informs
1266 * the thread that it was woken due to a real-time clock adjustment.
1267 * (The declaration of td_rtcgen refers to this comment.)
1269 if (td->td_rtcgen != 0 && td->td_rtcgen != rtc_generation) {
1276 static struct mtx tc_setclock_mtx;
1277 MTX_SYSINIT(tc_setclock_init, &tc_setclock_mtx, "tcsetc", MTX_SPIN);
1280 * Step our concept of UTC. This is done by modifying our estimate of
1284 tc_setclock(struct timespec *ts)
1286 struct timespec tbef, taft;
1287 struct bintime bt, bt2;
1289 timespec2bintime(ts, &bt);
1291 mtx_lock_spin(&tc_setclock_mtx);
1292 cpu_tick_calibrate(1);
1294 bintime_sub(&bt, &bt2);
1296 /* XXX fiddle all the little crinkly bits around the fiords... */
1298 mtx_unlock_spin(&tc_setclock_mtx);
1300 /* Avoid rtc_generation == 0, since td_rtcgen == 0 is special. */
1301 atomic_add_rel_int(&rtc_generation, 2);
1302 sleepq_chains_remove_matching(sleeping_on_old_rtc);
1303 if (timestepwarnings) {
1306 "Time stepped from %jd.%09ld to %jd.%09ld (%jd.%09ld)\n",
1307 (intmax_t)tbef.tv_sec, tbef.tv_nsec,
1308 (intmax_t)taft.tv_sec, taft.tv_nsec,
1309 (intmax_t)ts->tv_sec, ts->tv_nsec);
1314 * Recalculate the scaling factor. We want the number of 1/2^64
1315 * fractions of a second per period of the hardware counter, taking
1316 * into account the th_adjustment factor which the NTP PLL/adjtime(2)
1317 * processing provides us with.
1319 * The th_adjustment is nanoseconds per second with 32 bit binary
1320 * fraction and we want 64 bit binary fraction of second:
1322 * x = a * 2^32 / 10^9 = a * 4.294967296
1324 * The range of th_adjustment is +/- 5000PPM so inside a 64bit int
1325 * we can only multiply by about 850 without overflowing, that
1326 * leaves no suitably precise fractions for multiply before divide.
1328 * Divide before multiply with a fraction of 2199/512 results in a
1329 * systematic undercompensation of 10PPM of th_adjustment. On a
1330 * 5000PPM adjustment this is a 0.05PPM error. This is acceptable.
1332 * We happily sacrifice the lowest of the 64 bits of our result
1333 * to the goddess of code clarity.
1336 recalculate_scaling_factor_and_large_delta(struct timehands *th)
1340 scale = (uint64_t)1 << 63;
1341 scale += (th->th_adjustment / 1024) * 2199;
1342 scale /= th->th_counter->tc_frequency;
1343 th->th_scale = scale * 2;
1344 th->th_large_delta = MIN(((uint64_t)1 << 63) / scale, UINT_MAX);
1348 * Initialize the next struct timehands in the ring and make
1349 * it the active timehands. Along the way we might switch to a different
1350 * timecounter and/or do seconds processing in NTP. Slightly magic.
1353 tc_windup(struct bintime *new_boottimebin)
1356 struct timecounter *tc;
1357 struct timehands *th, *tho;
1358 u_int delta, ncount, ogen;
1363 * Make the next timehands a copy of the current one, but do
1364 * not overwrite the generation or next pointer. While we
1365 * update the contents, the generation must be zero. We need
1366 * to ensure that the zero generation is visible before the
1367 * data updates become visible, which requires release fence.
1368 * For similar reasons, re-reading of the generation after the
1369 * data is read should use acquire fence.
1373 ogen = th->th_generation;
1374 th->th_generation = 0;
1375 atomic_thread_fence_rel();
1376 memcpy(th, tho, offsetof(struct timehands, th_generation));
1377 if (new_boottimebin != NULL)
1378 th->th_boottime = *new_boottimebin;
1381 * Capture a timecounter delta on the current timecounter and if
1382 * changing timecounters, a counter value from the new timecounter.
1383 * Update the offset fields accordingly.
1385 tc = atomic_load_ptr(&timecounter);
1386 delta = tc_delta(th);
1387 if (th->th_counter != tc)
1388 ncount = tc->tc_get_timecount(tc);
1392 ffclock_windup(delta);
1394 th->th_offset_count += delta;
1395 th->th_offset_count &= th->th_counter->tc_counter_mask;
1396 bintime_add_tc_delta(&th->th_offset, th->th_scale,
1397 th->th_large_delta, delta);
1400 * Hardware latching timecounters may not generate interrupts on
1401 * PPS events, so instead we poll them. There is a finite risk that
1402 * the hardware might capture a count which is later than the one we
1403 * got above, and therefore possibly in the next NTP second which might
1404 * have a different rate than the current NTP second. It doesn't
1405 * matter in practice.
1407 if (tho->th_counter->tc_poll_pps)
1408 tho->th_counter->tc_poll_pps(tho->th_counter);
1411 * Deal with NTP second processing. The loop normally
1412 * iterates at most once, but in extreme situations it might
1413 * keep NTP sane if timeouts are not run for several seconds.
1414 * At boot, the time step can be large when the TOD hardware
1415 * has been read, so on really large steps, we call
1416 * ntp_update_second only twice. We need to call it twice in
1417 * case we missed a leap second.
1420 bintime_add(&bt, &th->th_boottime);
1421 i = bt.sec - tho->th_microtime.tv_sec;
1428 ntp_update_second(&th->th_adjustment, &bt.sec);
1430 th->th_boottime.sec += bt.sec - t;
1434 recalculate_scaling_factor_and_large_delta(th);
1437 /* Update the UTC timestamps used by the get*() functions. */
1438 th->th_bintime = bt;
1439 bintime2timeval(&bt, &th->th_microtime);
1440 bintime2timespec(&bt, &th->th_nanotime);
1442 /* Now is a good time to change timecounters. */
1443 if (th->th_counter != tc) {
1445 if ((tc->tc_flags & TC_FLAGS_C2STOP) != 0)
1446 cpu_disable_c2_sleep++;
1447 if ((th->th_counter->tc_flags & TC_FLAGS_C2STOP) != 0)
1448 cpu_disable_c2_sleep--;
1450 th->th_counter = tc;
1451 th->th_offset_count = ncount;
1452 tc_min_ticktock_freq = max(1, tc->tc_frequency /
1453 (((uint64_t)tc->tc_counter_mask + 1) / 3));
1454 recalculate_scaling_factor_and_large_delta(th);
1456 ffclock_change_tc(th);
1461 * Now that the struct timehands is again consistent, set the new
1462 * generation number, making sure to not make it zero.
1466 atomic_store_rel_int(&th->th_generation, ogen);
1468 /* Go live with the new struct timehands. */
1470 switch (sysclock_active) {
1473 time_second = th->th_microtime.tv_sec;
1474 time_uptime = th->th_offset.sec;
1478 time_second = fftimehands->tick_time_lerp.sec;
1479 time_uptime = fftimehands->tick_time_lerp.sec - ffclock_boottime.sec;
1485 timekeep_push_vdso();
1488 /* Report or change the active timecounter hardware. */
1490 sysctl_kern_timecounter_hardware(SYSCTL_HANDLER_ARGS)
1493 struct timecounter *newtc, *tc;
1498 strlcpy(newname, tc->tc_name, sizeof(newname));
1499 mtx_unlock(&tc_lock);
1501 error = sysctl_handle_string(oidp, &newname[0], sizeof(newname), req);
1502 if (error != 0 || req->newptr == NULL)
1506 /* Record that the tc in use now was specifically chosen. */
1508 if (strcmp(newname, tc->tc_name) == 0) {
1509 mtx_unlock(&tc_lock);
1512 for (newtc = timecounters; newtc != NULL; newtc = newtc->tc_next) {
1513 if (strcmp(newname, newtc->tc_name) != 0)
1516 /* Warm up new timecounter. */
1517 (void)newtc->tc_get_timecount(newtc);
1519 timecounter = newtc;
1522 * The vdso timehands update is deferred until the next
1525 * This is prudent given that 'timekeep_push_vdso()' does not
1526 * use any locking and that it can be called in hard interrupt
1527 * context via 'tc_windup()'.
1531 mtx_unlock(&tc_lock);
1532 return (newtc != NULL ? 0 : EINVAL);
1534 SYSCTL_PROC(_kern_timecounter, OID_AUTO, hardware,
1535 CTLTYPE_STRING | CTLFLAG_RWTUN | CTLFLAG_NOFETCH | CTLFLAG_MPSAFE, 0, 0,
1536 sysctl_kern_timecounter_hardware, "A",
1537 "Timecounter hardware selected");
1539 /* Report the available timecounter hardware. */
1541 sysctl_kern_timecounter_choice(SYSCTL_HANDLER_ARGS)
1544 struct timecounter *tc;
1547 error = sysctl_wire_old_buffer(req, 0);
1550 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 mtx_unlock(&tc_lock);
1558 error = sbuf_finish(&sb);
1563 SYSCTL_PROC(_kern_timecounter, OID_AUTO, choice,
1564 CTLTYPE_STRING | CTLFLAG_RD | CTLFLAG_MPSAFE, 0, 0,
1565 sysctl_kern_timecounter_choice, "A",
1566 "Timecounter hardware detected");
1569 * RFC 2783 PPS-API implementation.
1573 * Return true if the driver is aware of the abi version extensions in the
1574 * pps_state structure, and it supports at least the given abi version number.
1577 abi_aware(struct pps_state *pps, int vers)
1580 return ((pps->kcmode & KCMODE_ABIFLAG) && pps->driver_abi >= vers);
1584 pps_fetch(struct pps_fetch_args *fapi, struct pps_state *pps)
1587 pps_seq_t aseq, cseq;
1590 if (fapi->tsformat && fapi->tsformat != PPS_TSFMT_TSPEC)
1594 * If no timeout is requested, immediately return whatever values were
1595 * most recently captured. If timeout seconds is -1, that's a request
1596 * to block without a timeout. WITNESS won't let us sleep forever
1597 * without a lock (we really don't need a lock), so just repeatedly
1598 * sleep a long time.
1600 if (fapi->timeout.tv_sec || fapi->timeout.tv_nsec) {
1601 if (fapi->timeout.tv_sec == -1)
1604 tv.tv_sec = fapi->timeout.tv_sec;
1605 tv.tv_usec = fapi->timeout.tv_nsec / 1000;
1608 aseq = atomic_load_int(&pps->ppsinfo.assert_sequence);
1609 cseq = atomic_load_int(&pps->ppsinfo.clear_sequence);
1610 while (aseq == atomic_load_int(&pps->ppsinfo.assert_sequence) &&
1611 cseq == atomic_load_int(&pps->ppsinfo.clear_sequence)) {
1612 if (abi_aware(pps, 1) && pps->driver_mtx != NULL) {
1613 if (pps->flags & PPSFLAG_MTX_SPIN) {
1614 err = msleep_spin(pps, pps->driver_mtx,
1617 err = msleep(pps, pps->driver_mtx, PCATCH,
1621 err = tsleep(pps, PCATCH, "ppsfch", timo);
1623 if (err == EWOULDBLOCK) {
1624 if (fapi->timeout.tv_sec == -1) {
1629 } else if (err != 0) {
1635 pps->ppsinfo.current_mode = pps->ppsparam.mode;
1636 fapi->pps_info_buf = pps->ppsinfo;
1642 pps_ioctl(u_long cmd, caddr_t data, struct pps_state *pps)
1645 struct pps_fetch_args *fapi;
1647 struct pps_fetch_ffc_args *fapi_ffc;
1650 struct pps_kcbind_args *kapi;
1653 KASSERT(pps != NULL, ("NULL pps pointer in pps_ioctl"));
1655 case PPS_IOC_CREATE:
1657 case PPS_IOC_DESTROY:
1659 case PPS_IOC_SETPARAMS:
1660 app = (pps_params_t *)data;
1661 if (app->mode & ~pps->ppscap)
1664 /* Ensure only a single clock is selected for ffc timestamp. */
1665 if ((app->mode & PPS_TSCLK_MASK) == PPS_TSCLK_MASK)
1668 pps->ppsparam = *app;
1670 case PPS_IOC_GETPARAMS:
1671 app = (pps_params_t *)data;
1672 *app = pps->ppsparam;
1673 app->api_version = PPS_API_VERS_1;
1675 case PPS_IOC_GETCAP:
1676 *(int*)data = pps->ppscap;
1679 fapi = (struct pps_fetch_args *)data;
1680 return (pps_fetch(fapi, pps));
1682 case PPS_IOC_FETCH_FFCOUNTER:
1683 fapi_ffc = (struct pps_fetch_ffc_args *)data;
1684 if (fapi_ffc->tsformat && fapi_ffc->tsformat !=
1687 if (fapi_ffc->timeout.tv_sec || fapi_ffc->timeout.tv_nsec)
1688 return (EOPNOTSUPP);
1689 pps->ppsinfo_ffc.current_mode = pps->ppsparam.mode;
1690 fapi_ffc->pps_info_buf_ffc = pps->ppsinfo_ffc;
1691 /* Overwrite timestamps if feedback clock selected. */
1692 switch (pps->ppsparam.mode & PPS_TSCLK_MASK) {
1693 case PPS_TSCLK_FBCK:
1694 fapi_ffc->pps_info_buf_ffc.assert_timestamp =
1695 pps->ppsinfo.assert_timestamp;
1696 fapi_ffc->pps_info_buf_ffc.clear_timestamp =
1697 pps->ppsinfo.clear_timestamp;
1699 case PPS_TSCLK_FFWD:
1705 #endif /* FFCLOCK */
1706 case PPS_IOC_KCBIND:
1708 kapi = (struct pps_kcbind_args *)data;
1709 /* XXX Only root should be able to do this */
1710 if (kapi->tsformat && kapi->tsformat != PPS_TSFMT_TSPEC)
1712 if (kapi->kernel_consumer != PPS_KC_HARDPPS)
1714 if (kapi->edge & ~pps->ppscap)
1716 pps->kcmode = (kapi->edge & KCMODE_EDGEMASK) |
1717 (pps->kcmode & KCMODE_ABIFLAG);
1720 return (EOPNOTSUPP);
1728 pps_init(struct pps_state *pps)
1730 pps->ppscap |= PPS_TSFMT_TSPEC | PPS_CANWAIT;
1731 if (pps->ppscap & PPS_CAPTUREASSERT)
1732 pps->ppscap |= PPS_OFFSETASSERT;
1733 if (pps->ppscap & PPS_CAPTURECLEAR)
1734 pps->ppscap |= PPS_OFFSETCLEAR;
1736 pps->ppscap |= PPS_TSCLK_MASK;
1738 pps->kcmode &= ~KCMODE_ABIFLAG;
1742 pps_init_abi(struct pps_state *pps)
1746 if (pps->driver_abi > 0) {
1747 pps->kcmode |= KCMODE_ABIFLAG;
1748 pps->kernel_abi = PPS_ABI_VERSION;
1753 pps_capture(struct pps_state *pps)
1755 struct timehands *th;
1757 KASSERT(pps != NULL, ("NULL pps pointer in pps_capture"));
1759 pps->capgen = atomic_load_acq_int(&th->th_generation);
1762 pps->capffth = fftimehands;
1764 pps->capcount = th->th_counter->tc_get_timecount(th->th_counter);
1765 atomic_thread_fence_acq();
1766 if (pps->capgen != th->th_generation)
1771 pps_event(struct pps_state *pps, int event)
1774 struct timespec ts, *tsp, *osp;
1775 u_int tcount, *pcount;
1779 struct timespec *tsp_ffc;
1780 pps_seq_t *pseq_ffc;
1787 KASSERT(pps != NULL, ("NULL pps pointer in pps_event"));
1788 /* Nothing to do if not currently set to capture this event type. */
1789 if ((event & pps->ppsparam.mode) == 0)
1791 /* If the timecounter was wound up underneath us, bail out. */
1792 if (pps->capgen == 0 || pps->capgen !=
1793 atomic_load_acq_int(&pps->capth->th_generation))
1796 /* Things would be easier with arrays. */
1797 if (event == PPS_CAPTUREASSERT) {
1798 tsp = &pps->ppsinfo.assert_timestamp;
1799 osp = &pps->ppsparam.assert_offset;
1800 foff = pps->ppsparam.mode & PPS_OFFSETASSERT;
1802 fhard = pps->kcmode & PPS_CAPTUREASSERT;
1804 pcount = &pps->ppscount[0];
1805 pseq = &pps->ppsinfo.assert_sequence;
1807 ffcount = &pps->ppsinfo_ffc.assert_ffcount;
1808 tsp_ffc = &pps->ppsinfo_ffc.assert_timestamp;
1809 pseq_ffc = &pps->ppsinfo_ffc.assert_sequence;
1812 tsp = &pps->ppsinfo.clear_timestamp;
1813 osp = &pps->ppsparam.clear_offset;
1814 foff = pps->ppsparam.mode & PPS_OFFSETCLEAR;
1816 fhard = pps->kcmode & PPS_CAPTURECLEAR;
1818 pcount = &pps->ppscount[1];
1819 pseq = &pps->ppsinfo.clear_sequence;
1821 ffcount = &pps->ppsinfo_ffc.clear_ffcount;
1822 tsp_ffc = &pps->ppsinfo_ffc.clear_timestamp;
1823 pseq_ffc = &pps->ppsinfo_ffc.clear_sequence;
1828 * If the timecounter changed, we cannot compare the count values, so
1829 * we have to drop the rest of the PPS-stuff until the next event.
1831 if (pps->ppstc != pps->capth->th_counter) {
1832 pps->ppstc = pps->capth->th_counter;
1833 *pcount = pps->capcount;
1834 pps->ppscount[2] = pps->capcount;
1838 /* Convert the count to a timespec. */
1839 tcount = pps->capcount - pps->capth->th_offset_count;
1840 tcount &= pps->capth->th_counter->tc_counter_mask;
1841 bt = pps->capth->th_bintime;
1842 bintime_addx(&bt, pps->capth->th_scale * tcount);
1843 bintime2timespec(&bt, &ts);
1845 /* If the timecounter was wound up underneath us, bail out. */
1846 atomic_thread_fence_acq();
1847 if (pps->capgen != pps->capth->th_generation)
1850 *pcount = pps->capcount;
1855 timespecadd(tsp, osp, tsp);
1856 if (tsp->tv_nsec < 0) {
1857 tsp->tv_nsec += 1000000000;
1863 *ffcount = pps->capffth->tick_ffcount + tcount;
1864 bt = pps->capffth->tick_time;
1865 ffclock_convert_delta(tcount, pps->capffth->cest.period, &bt);
1866 bintime_add(&bt, &pps->capffth->tick_time);
1867 bintime2timespec(&bt, &ts);
1877 * Feed the NTP PLL/FLL.
1878 * The FLL wants to know how many (hardware) nanoseconds
1879 * elapsed since the previous event.
1881 tcount = pps->capcount - pps->ppscount[2];
1882 pps->ppscount[2] = pps->capcount;
1883 tcount &= pps->capth->th_counter->tc_counter_mask;
1884 scale = (uint64_t)1 << 63;
1885 scale /= pps->capth->th_counter->tc_frequency;
1889 bintime_addx(&bt, scale * tcount);
1890 bintime2timespec(&bt, &ts);
1891 hardpps(tsp, ts.tv_nsec + 1000000000 * ts.tv_sec);
1895 /* Wakeup anyone sleeping in pps_fetch(). */
1900 * Timecounters need to be updated every so often to prevent the hardware
1901 * counter from overflowing. Updating also recalculates the cached values
1902 * used by the get*() family of functions, so their precision depends on
1903 * the update frequency.
1907 SYSCTL_INT(_kern_timecounter, OID_AUTO, tick, CTLFLAG_RD, &tc_tick, 0,
1908 "Approximate number of hardclock ticks in a millisecond");
1911 tc_ticktock(int cnt)
1915 if (mtx_trylock_spin(&tc_setclock_mtx)) {
1917 if (count >= tc_tick) {
1921 mtx_unlock_spin(&tc_setclock_mtx);
1925 static void __inline
1926 tc_adjprecision(void)
1930 if (tc_timepercentage > 0) {
1931 t = (99 + tc_timepercentage) / tc_timepercentage;
1932 tc_precexp = fls(t + (t >> 1)) - 1;
1933 FREQ2BT(hz / tc_tick, &bt_timethreshold);
1934 FREQ2BT(hz, &bt_tickthreshold);
1935 bintime_shift(&bt_timethreshold, tc_precexp);
1936 bintime_shift(&bt_tickthreshold, tc_precexp);
1939 bt_timethreshold.sec = INT_MAX;
1940 bt_timethreshold.frac = ~(uint64_t)0;
1941 bt_tickthreshold = bt_timethreshold;
1943 sbt_timethreshold = bttosbt(bt_timethreshold);
1944 sbt_tickthreshold = bttosbt(bt_tickthreshold);
1948 sysctl_kern_timecounter_adjprecision(SYSCTL_HANDLER_ARGS)
1952 val = tc_timepercentage;
1953 error = sysctl_handle_int(oidp, &val, 0, req);
1954 if (error != 0 || req->newptr == NULL)
1956 tc_timepercentage = val;
1964 /* Set up the requested number of timehands. */
1966 inittimehands(void *dummy)
1968 struct timehands *thp;
1971 TUNABLE_INT_FETCH("kern.timecounter.timehands_count",
1973 if (timehands_count < 1)
1974 timehands_count = 1;
1975 if (timehands_count > nitems(ths))
1976 timehands_count = nitems(ths);
1977 for (i = 1, thp = &ths[0]; i < timehands_count; thp = &ths[i++])
1978 thp->th_next = &ths[i];
1979 thp->th_next = &ths[0];
1981 TUNABLE_STR_FETCH("kern.timecounter.hardware", tc_from_tunable,
1982 sizeof(tc_from_tunable));
1984 mtx_init(&tc_lock, "tc", NULL, MTX_DEF);
1986 SYSINIT(timehands, SI_SUB_TUNABLES, SI_ORDER_ANY, inittimehands, NULL);
1989 inittimecounter(void *dummy)
1995 * Set the initial timeout to
1996 * max(1, <approx. number of hardclock ticks in a millisecond>).
1997 * People should probably not use the sysctl to set the timeout
1998 * to smaller than its initial value, since that value is the
1999 * smallest reasonable one. If they want better timestamps they
2000 * should use the non-"get"* functions.
2003 tc_tick = (hz + 500) / 1000;
2007 FREQ2BT(hz, &tick_bt);
2008 tick_sbt = bttosbt(tick_bt);
2009 tick_rate = hz / tc_tick;
2010 FREQ2BT(tick_rate, &tc_tick_bt);
2011 tc_tick_sbt = bttosbt(tc_tick_bt);
2012 p = (tc_tick * 1000000) / hz;
2013 printf("Timecounters tick every %d.%03u msec\n", p / 1000, p % 1000);
2019 /* warm up new timecounter (again) and get rolling. */
2020 (void)timecounter->tc_get_timecount(timecounter);
2021 mtx_lock_spin(&tc_setclock_mtx);
2023 mtx_unlock_spin(&tc_setclock_mtx);
2026 SYSINIT(timecounter, SI_SUB_CLOCKS, SI_ORDER_SECOND, inittimecounter, NULL);
2028 /* Cpu tick handling -------------------------------------------------*/
2030 static bool cpu_tick_variable;
2031 static uint64_t cpu_tick_frequency;
2033 DPCPU_DEFINE_STATIC(uint64_t, tc_cpu_ticks_base);
2034 DPCPU_DEFINE_STATIC(unsigned, tc_cpu_ticks_last);
2039 struct timecounter *tc;
2040 uint64_t res, *base;
2044 base = DPCPU_PTR(tc_cpu_ticks_base);
2045 last = DPCPU_PTR(tc_cpu_ticks_last);
2046 tc = timehands->th_counter;
2047 u = tc->tc_get_timecount(tc) & tc->tc_counter_mask;
2049 *base += (uint64_t)tc->tc_counter_mask + 1;
2057 cpu_tick_calibration(void)
2059 static time_t last_calib;
2061 if (time_uptime != last_calib && !(time_uptime & 0xf)) {
2062 cpu_tick_calibrate(0);
2063 last_calib = time_uptime;
2068 * This function gets called every 16 seconds on only one designated
2069 * CPU in the system from hardclock() via cpu_tick_calibration()().
2071 * Whenever the real time clock is stepped we get called with reset=1
2072 * to make sure we handle suspend/resume and similar events correctly.
2076 cpu_tick_calibrate(int reset)
2078 static uint64_t c_last;
2079 uint64_t c_this, c_delta;
2080 static struct bintime t_last;
2081 struct bintime t_this, t_delta;
2085 /* The clock was stepped, abort & reset */
2090 /* we don't calibrate fixed rate cputicks */
2091 if (!cpu_tick_variable)
2094 getbinuptime(&t_this);
2095 c_this = cpu_ticks();
2096 if (t_last.sec != 0) {
2097 c_delta = c_this - c_last;
2099 bintime_sub(&t_delta, &t_last);
2102 * 2^(64-20) / 16[s] =
2104 * 17.592.186.044.416 / 16 =
2105 * 1.099.511.627.776 [Hz]
2107 divi = t_delta.sec << 20;
2108 divi |= t_delta.frac >> (64 - 20);
2111 if (c_delta > cpu_tick_frequency) {
2112 if (0 && bootverbose)
2113 printf("cpu_tick increased to %ju Hz\n",
2115 cpu_tick_frequency = c_delta;
2123 set_cputicker(cpu_tick_f *func, uint64_t freq, bool isvariable)
2127 cpu_ticks = tc_cpu_ticks;
2129 cpu_tick_frequency = freq;
2130 cpu_tick_variable = isvariable;
2139 if (cpu_ticks == tc_cpu_ticks)
2140 return (tc_getfrequency());
2141 return (cpu_tick_frequency);
2145 * We need to be slightly careful converting cputicks to microseconds.
2146 * There is plenty of margin in 64 bits of microseconds (half a million
2147 * years) and in 64 bits at 4 GHz (146 years), but if we do a multiply
2148 * before divide conversion (to retain precision) we find that the
2149 * margin shrinks to 1.5 hours (one millionth of 146y).
2150 * With a three prong approach we never lose significant bits, no
2151 * matter what the cputick rate and length of timeinterval is.
2155 cputick2usec(uint64_t tick)
2158 if (tick > 18446744073709551LL) /* floor(2^64 / 1000) */
2159 return (tick / (cpu_tickrate() / 1000000LL));
2160 else if (tick > 18446744073709LL) /* floor(2^64 / 1000000) */
2161 return ((tick * 1000LL) / (cpu_tickrate() / 1000LL));
2163 return ((tick * 1000000LL) / cpu_tickrate());
2166 cpu_tick_f *cpu_ticks = tc_cpu_ticks;
2168 static int vdso_th_enable = 1;
2170 sysctl_fast_gettime(SYSCTL_HANDLER_ARGS)
2172 int old_vdso_th_enable, error;
2174 old_vdso_th_enable = vdso_th_enable;
2175 error = sysctl_handle_int(oidp, &old_vdso_th_enable, 0, req);
2178 vdso_th_enable = old_vdso_th_enable;
2181 SYSCTL_PROC(_kern_timecounter, OID_AUTO, fast_gettime,
2182 CTLTYPE_INT | CTLFLAG_RW | CTLFLAG_MPSAFE,
2183 NULL, 0, sysctl_fast_gettime, "I", "Enable fast time of day");
2186 tc_fill_vdso_timehands(struct vdso_timehands *vdso_th)
2188 struct timehands *th;
2192 vdso_th->th_scale = th->th_scale;
2193 vdso_th->th_offset_count = th->th_offset_count;
2194 vdso_th->th_counter_mask = th->th_counter->tc_counter_mask;
2195 vdso_th->th_offset = th->th_offset;
2196 vdso_th->th_boottime = th->th_boottime;
2197 if (th->th_counter->tc_fill_vdso_timehands != NULL) {
2198 enabled = th->th_counter->tc_fill_vdso_timehands(vdso_th,
2202 if (!vdso_th_enable)
2207 #ifdef COMPAT_FREEBSD32
2209 tc_fill_vdso_timehands32(struct vdso_timehands32 *vdso_th32)
2211 struct timehands *th;
2215 *(uint64_t *)&vdso_th32->th_scale[0] = th->th_scale;
2216 vdso_th32->th_offset_count = th->th_offset_count;
2217 vdso_th32->th_counter_mask = th->th_counter->tc_counter_mask;
2218 vdso_th32->th_offset.sec = th->th_offset.sec;
2219 *(uint64_t *)&vdso_th32->th_offset.frac[0] = th->th_offset.frac;
2220 vdso_th32->th_boottime.sec = th->th_boottime.sec;
2221 *(uint64_t *)&vdso_th32->th_boottime.frac[0] = th->th_boottime.frac;
2222 if (th->th_counter->tc_fill_vdso_timehands32 != NULL) {
2223 enabled = th->th_counter->tc_fill_vdso_timehands32(vdso_th32,
2227 if (!vdso_th_enable)
2233 #include "opt_ddb.h"
2235 #include <ddb/ddb.h>
2237 DB_SHOW_COMMAND(timecounter, db_show_timecounter)
2239 struct timehands *th;
2240 struct timecounter *tc;
2244 tc = th->th_counter;
2245 val1 = tc->tc_get_timecount(tc);
2246 __compiler_membar();
2247 val2 = tc->tc_get_timecount(tc);
2249 db_printf("timecounter %p %s\n", tc, tc->tc_name);
2250 db_printf(" mask %#x freq %ju qual %d flags %#x priv %p\n",
2251 tc->tc_counter_mask, (uintmax_t)tc->tc_frequency, tc->tc_quality,
2252 tc->tc_flags, tc->tc_priv);
2253 db_printf(" val %#x %#x\n", val1, val2);
2254 db_printf("timehands adj %#jx scale %#jx ldelta %d off_cnt %d gen %d\n",
2255 (uintmax_t)th->th_adjustment, (uintmax_t)th->th_scale,
2256 th->th_large_delta, th->th_offset_count, th->th_generation);
2257 db_printf(" offset %jd %jd boottime %jd %jd\n",
2258 (intmax_t)th->th_offset.sec, (uintmax_t)th->th_offset.frac,
2259 (intmax_t)th->th_boottime.sec, (uintmax_t)th->th_boottime.frac);