Home | History | Annotate | Line # | Download | only in kern
kern_tc.c revision 1.1.1.1.6.4
      1 /* $NetBSD: kern_tc.c,v 1.1.1.1.6.4 2006/09/14 12:31:48 yamt Exp $ */
      2 
      3 /*-
      4  * ----------------------------------------------------------------------------
      5  * "THE BEER-WARE LICENSE" (Revision 42):
      6  * <phk (at) FreeBSD.ORG> wrote this file.  As long as you retain this notice you
      7  * can do whatever you want with this stuff. If we meet some day, and you think
      8  * this stuff is worth it, you can buy me a beer in return.   Poul-Henning Kamp
      9  * ---------------------------------------------------------------------------
     10  */
     11 
     12 #include <sys/cdefs.h>
     13 /* __FBSDID("$FreeBSD: src/sys/kern/kern_tc.c,v 1.166 2005/09/19 22:16:31 andre Exp $"); */
     14 __KERNEL_RCSID(0, "$NetBSD: kern_tc.c,v 1.1.1.1.6.4 2006/09/14 12:31:48 yamt Exp $");
     15 
     16 #include "opt_ntp.h"
     17 
     18 #include <sys/param.h>
     19 #ifdef __HAVE_TIMECOUNTER	/* XXX */
     20 #include <sys/kernel.h>
     21 #include <sys/reboot.h>	/* XXX just to get AB_VERBOSE */
     22 #include <sys/sysctl.h>
     23 #include <sys/syslog.h>
     24 #include <sys/systm.h>
     25 #include <sys/timepps.h>
     26 #include <sys/timetc.h>
     27 #include <sys/timex.h>
     28 #include <sys/evcnt.h>
     29 #include <sys/kauth.h>
     30 
     31 /*
     32  * A large step happens on boot.  This constant detects such steps.
     33  * It is relatively small so that ntp_update_second gets called enough
     34  * in the typical 'missed a couple of seconds' case, but doesn't loop
     35  * forever when the time step is large.
     36  */
     37 #define LARGE_STEP	200
     38 
     39 /*
     40  * Implement a dummy timecounter which we can use until we get a real one
     41  * in the air.  This allows the console and other early stuff to use
     42  * time services.
     43  */
     44 
     45 static u_int
     46 dummy_get_timecount(struct timecounter *tc)
     47 {
     48 	static u_int now;
     49 
     50 	return (++now);
     51 }
     52 
     53 static struct timecounter dummy_timecounter = {
     54 	dummy_get_timecount, 0, ~0u, 1000000, "dummy", -1000000, NULL, NULL,
     55 };
     56 
     57 struct timehands {
     58 	/* These fields must be initialized by the driver. */
     59 	struct timecounter	*th_counter;
     60 	int64_t			th_adjustment;
     61 	u_int64_t		th_scale;
     62 	u_int	 		th_offset_count;
     63 	struct bintime		th_offset;
     64 	struct timeval		th_microtime;
     65 	struct timespec		th_nanotime;
     66 	/* Fields not to be copied in tc_windup start with th_generation. */
     67 	volatile u_int		th_generation;
     68 	struct timehands	*th_next;
     69 };
     70 
     71 static struct timehands th0;
     72 static struct timehands th9 = { .th_next = &th0, };
     73 static struct timehands th8 = { .th_next = &th9, };
     74 static struct timehands th7 = { .th_next = &th8, };
     75 static struct timehands th6 = { .th_next = &th7, };
     76 static struct timehands th5 = { .th_next = &th6, };
     77 static struct timehands th4 = { .th_next = &th5, };
     78 static struct timehands th3 = { .th_next = &th4, };
     79 static struct timehands th2 = { .th_next = &th3, };
     80 static struct timehands th1 = { .th_next = &th2, };
     81 static struct timehands th0 = {
     82 	.th_counter = &dummy_timecounter,
     83 	.th_scale = (uint64_t)-1 / 1000000,
     84 	.th_offset = { .sec = 1, .frac = 0 },
     85 	.th_generation = 1,
     86 	.th_next = &th1,
     87 };
     88 
     89 static struct timehands *volatile timehands = &th0;
     90 struct timecounter *timecounter = &dummy_timecounter;
     91 static struct timecounter *timecounters = &dummy_timecounter;
     92 
     93 time_t time_second = 1;
     94 time_t time_uptime = 1;
     95 
     96 static struct bintime timebasebin;
     97 
     98 static int timestepwarnings;
     99 
    100 #ifdef __FreeBSD__
    101 SYSCTL_INT(_kern_timecounter, OID_AUTO, stepwarnings, CTLFLAG_RW,
    102     &timestepwarnings, 0, "");
    103 #endif /* __FreeBSD__ */
    104 
    105 /*
    106  * sysctl helper routine for kern.timercounter.current
    107  */
    108 static int
    109 sysctl_kern_timecounter_hardware(SYSCTLFN_ARGS)
    110 {
    111 	struct sysctlnode node;
    112 	int error;
    113 	char newname[MAX_TCNAMELEN];
    114 	struct timecounter *newtc, *tc;
    115 
    116 	tc = timecounter;
    117 
    118 	strlcpy(newname, tc->tc_name, sizeof(newname));
    119 
    120 	node = *rnode;
    121 	node.sysctl_data = newname;
    122 	node.sysctl_size = sizeof(newname);
    123 
    124 	error = sysctl_lookup(SYSCTLFN_CALL(&node));
    125 
    126 	if (error ||
    127 	    newp == NULL ||
    128 	    strncmp(newname, tc->tc_name, sizeof(newname)) == 0)
    129 		return error;
    130 
    131 	if (l != NULL && (error = kauth_authorize_generic(l->l_cred,
    132 	    KAUTH_GENERIC_ISSUSER, &l->l_acflag)) != 0)
    133 		return (error);
    134 
    135 	/* XXX locking */
    136 
    137 	for (newtc = timecounters; newtc != NULL; newtc = newtc->tc_next) {
    138 		if (strcmp(newname, newtc->tc_name) != 0)
    139 			continue;
    140 
    141 		/* Warm up new timecounter. */
    142 		(void)newtc->tc_get_timecount(newtc);
    143 		(void)newtc->tc_get_timecount(newtc);
    144 
    145 		timecounter = newtc;
    146 
    147 		/* XXX unlock */
    148 
    149 		return (0);
    150 	}
    151 
    152 	/* XXX unlock */
    153 
    154 	return (EINVAL);
    155 }
    156 
    157 static int
    158 sysctl_kern_timecounter_choice(SYSCTLFN_ARGS)
    159 {
    160 	char buf[MAX_TCNAMELEN+48];
    161 	char *where = oldp;
    162 	const char *spc;
    163 	struct timecounter *tc;
    164 	size_t needed, left, slen;
    165 	int error;
    166 
    167 	if (newp != NULL)
    168 		return (EPERM);
    169 	if (namelen != 0)
    170 		return (EINVAL);
    171 
    172 	spc = "";
    173 	error = 0;
    174 	needed = 0;
    175 	left = *oldlenp;
    176 
    177 	/* XXX locking */
    178 
    179 	for (tc = timecounters; error == 0 && tc != NULL; tc = tc->tc_next) {
    180 		if (where == NULL) {
    181 			needed += sizeof(buf);  /* be conservative */
    182 		} else {
    183 			slen = snprintf(buf, sizeof(buf), "%s%s(q=%d, f=%" PRId64
    184 					" Hz)", spc, tc->tc_name, tc->tc_quality,
    185 					tc->tc_frequency);
    186 			if (left < slen + 1)
    187 				break;
    188 			/* XXX use sysctl_copyout? (from sysctl_hw_disknames) */
    189 			error = copyout(buf, where, slen + 1);
    190 			spc = " ";
    191 			where += slen;
    192 			needed += slen;
    193 			left -= slen;
    194 		}
    195 	}
    196 
    197 	/* XXX unlock */
    198 
    199 	*oldlenp = needed;
    200 	return (error);
    201 }
    202 
    203 SYSCTL_SETUP(sysctl_timecounter_setup, "sysctl timecounter setup")
    204 {
    205 	const struct sysctlnode *node;
    206 
    207 	sysctl_createv(clog, 0, NULL, &node,
    208 		       CTLFLAG_PERMANENT,
    209 		       CTLTYPE_NODE, "timecounter",
    210 		       SYSCTL_DESCR("time counter information"),
    211 		       NULL, 0, NULL, 0,
    212 		       CTL_KERN, CTL_CREATE, CTL_EOL);
    213 
    214 	if (node != NULL) {
    215 		sysctl_createv(clog, 0, NULL, NULL,
    216 			       CTLFLAG_PERMANENT,
    217 			       CTLTYPE_STRING, "choice",
    218 			       SYSCTL_DESCR("available counters"),
    219 			       sysctl_kern_timecounter_choice, 0, NULL, 0,
    220 			       CTL_KERN, node->sysctl_num, CTL_CREATE, CTL_EOL);
    221 
    222 		sysctl_createv(clog, 0, NULL, NULL,
    223 			       CTLFLAG_PERMANENT|CTLFLAG_READWRITE,
    224 			       CTLTYPE_STRING, "hardware",
    225 			       SYSCTL_DESCR("currently active time counter"),
    226 			       sysctl_kern_timecounter_hardware, 0, NULL, MAX_TCNAMELEN,
    227 			       CTL_KERN, node->sysctl_num, CTL_CREATE, CTL_EOL);
    228 
    229 		sysctl_createv(clog, 0, NULL, NULL,
    230 			       CTLFLAG_PERMANENT|CTLFLAG_READWRITE,
    231 			       CTLTYPE_INT, "timestepwarnings",
    232 			       SYSCTL_DESCR("log time steps"),
    233 			       NULL, 0, &timestepwarnings, 0,
    234 			       CTL_KERN, node->sysctl_num, CTL_CREATE, CTL_EOL);
    235 	}
    236 }
    237 
    238 #define	TC_STATS(name)							\
    239 static struct evcnt n##name =						\
    240     EVCNT_INITIALIZER(EVCNT_TYPE_MISC, NULL, "timecounter", #name);	\
    241 EVCNT_ATTACH_STATIC(n##name)
    242 
    243 TC_STATS(binuptime);    TC_STATS(nanouptime);    TC_STATS(microuptime);
    244 TC_STATS(bintime);      TC_STATS(nanotime);      TC_STATS(microtime);
    245 TC_STATS(getbinuptime); TC_STATS(getnanouptime); TC_STATS(getmicrouptime);
    246 TC_STATS(getbintime);   TC_STATS(getnanotime);   TC_STATS(getmicrotime);
    247 TC_STATS(setclock);
    248 
    249 #undef TC_STATS
    250 
    251 static void tc_windup(void);
    252 
    253 /*
    254  * Return the difference between the timehands' counter value now and what
    255  * was when we copied it to the timehands' offset_count.
    256  */
    257 static __inline u_int
    258 tc_delta(struct timehands *th)
    259 {
    260 	struct timecounter *tc;
    261 
    262 	tc = th->th_counter;
    263 	return ((tc->tc_get_timecount(tc) -
    264 		 th->th_offset_count) & tc->tc_counter_mask);
    265 }
    266 
    267 /*
    268  * Functions for reading the time.  We have to loop until we are sure that
    269  * the timehands that we operated on was not updated under our feet.  See
    270  * the comment in <sys/time.h> for a description of these 12 functions.
    271  */
    272 
    273 void
    274 binuptime(struct bintime *bt)
    275 {
    276 	struct timehands *th;
    277 	u_int gen;
    278 
    279 	nbinuptime.ev_count++;
    280 	do {
    281 		th = timehands;
    282 		gen = th->th_generation;
    283 		*bt = th->th_offset;
    284 		bintime_addx(bt, th->th_scale * tc_delta(th));
    285 	} while (gen == 0 || gen != th->th_generation);
    286 }
    287 
    288 void
    289 nanouptime(struct timespec *tsp)
    290 {
    291 	struct bintime bt;
    292 
    293 	nnanouptime.ev_count++;
    294 	binuptime(&bt);
    295 	bintime2timespec(&bt, tsp);
    296 }
    297 
    298 void
    299 microuptime(struct timeval *tvp)
    300 {
    301 	struct bintime bt;
    302 
    303 	nmicrouptime.ev_count++;
    304 	binuptime(&bt);
    305 	bintime2timeval(&bt, tvp);
    306 }
    307 
    308 void
    309 bintime(struct bintime *bt)
    310 {
    311 
    312 	nbintime.ev_count++;
    313 	binuptime(bt);
    314 	bintime_add(bt, &timebasebin);
    315 }
    316 
    317 void
    318 nanotime(struct timespec *tsp)
    319 {
    320 	struct bintime bt;
    321 
    322 	nnanotime.ev_count++;
    323 	bintime(&bt);
    324 	bintime2timespec(&bt, tsp);
    325 }
    326 
    327 void
    328 microtime(struct timeval *tvp)
    329 {
    330 	struct bintime bt;
    331 
    332 	nmicrotime.ev_count++;
    333 	bintime(&bt);
    334 	bintime2timeval(&bt, tvp);
    335 }
    336 
    337 void
    338 getbinuptime(struct bintime *bt)
    339 {
    340 	struct timehands *th;
    341 	u_int gen;
    342 
    343 	ngetbinuptime.ev_count++;
    344 	do {
    345 		th = timehands;
    346 		gen = th->th_generation;
    347 		*bt = th->th_offset;
    348 	} while (gen == 0 || gen != th->th_generation);
    349 }
    350 
    351 void
    352 getnanouptime(struct timespec *tsp)
    353 {
    354 	struct timehands *th;
    355 	u_int gen;
    356 
    357 	ngetnanouptime.ev_count++;
    358 	do {
    359 		th = timehands;
    360 		gen = th->th_generation;
    361 		bintime2timespec(&th->th_offset, tsp);
    362 	} while (gen == 0 || gen != th->th_generation);
    363 }
    364 
    365 void
    366 getmicrouptime(struct timeval *tvp)
    367 {
    368 	struct timehands *th;
    369 	u_int gen;
    370 
    371 	ngetmicrouptime.ev_count++;
    372 	do {
    373 		th = timehands;
    374 		gen = th->th_generation;
    375 		bintime2timeval(&th->th_offset, tvp);
    376 	} while (gen == 0 || gen != th->th_generation);
    377 }
    378 
    379 void
    380 getbintime(struct bintime *bt)
    381 {
    382 	struct timehands *th;
    383 	u_int gen;
    384 
    385 	ngetbintime.ev_count++;
    386 	do {
    387 		th = timehands;
    388 		gen = th->th_generation;
    389 		*bt = th->th_offset;
    390 	} while (gen == 0 || gen != th->th_generation);
    391 	bintime_add(bt, &timebasebin);
    392 }
    393 
    394 void
    395 getnanotime(struct timespec *tsp)
    396 {
    397 	struct timehands *th;
    398 	u_int gen;
    399 
    400 	ngetnanotime.ev_count++;
    401 	do {
    402 		th = timehands;
    403 		gen = th->th_generation;
    404 		*tsp = th->th_nanotime;
    405 	} while (gen == 0 || gen != th->th_generation);
    406 }
    407 
    408 void
    409 getmicrotime(struct timeval *tvp)
    410 {
    411 	struct timehands *th;
    412 	u_int gen;
    413 
    414 	ngetmicrotime.ev_count++;
    415 	do {
    416 		th = timehands;
    417 		gen = th->th_generation;
    418 		*tvp = th->th_microtime;
    419 	} while (gen == 0 || gen != th->th_generation);
    420 }
    421 
    422 /*
    423  * Initialize a new timecounter and possibly use it.
    424  */
    425 void
    426 tc_init(struct timecounter *tc)
    427 {
    428 	u_int u;
    429 	int s;
    430 
    431 	u = tc->tc_frequency / tc->tc_counter_mask;
    432 	/* XXX: We need some margin here, 10% is a guess */
    433 	u *= 11;
    434 	u /= 10;
    435 	if (u > hz && tc->tc_quality >= 0) {
    436 		tc->tc_quality = -2000;
    437 		if (bootverbose) {
    438 			printf("timecounter: Timecounter \"%s\" frequency %ju Hz",
    439 			    tc->tc_name, (uintmax_t)tc->tc_frequency);
    440 			printf(" -- Insufficient hz, needs at least %u\n", u);
    441 		}
    442 	} else if (tc->tc_quality >= 0 || bootverbose) {
    443 		printf("timecounter: Timecounter \"%s\" frequency %ju Hz quality %d\n",
    444 		    tc->tc_name, (uintmax_t)tc->tc_frequency,
    445 		    tc->tc_quality);
    446 	}
    447 
    448 	s = splclock();
    449 
    450 	tc->tc_next = timecounters;
    451 	timecounters = tc;
    452 	/*
    453 	 * Never automatically use a timecounter with negative quality.
    454 	 * Even though we run on the dummy counter, switching here may be
    455 	 * worse since this timecounter may not be monotonous.
    456 	 */
    457 	if (tc->tc_quality < 0)
    458 		goto out;
    459 	if (tc->tc_quality < timecounter->tc_quality)
    460 		goto out;
    461 	if (tc->tc_quality == timecounter->tc_quality &&
    462 	    tc->tc_frequency < timecounter->tc_frequency)
    463 		goto out;
    464 	(void)tc->tc_get_timecount(tc);
    465 	(void)tc->tc_get_timecount(tc);
    466 	timecounter = tc;
    467 	tc_windup();
    468 
    469  out:
    470 	splx(s);
    471 }
    472 
    473 /* Report the frequency of the current timecounter. */
    474 u_int64_t
    475 tc_getfrequency(void)
    476 {
    477 
    478 	return (timehands->th_counter->tc_frequency);
    479 }
    480 
    481 /*
    482  * Step our concept of UTC.  This is done by modifying our estimate of
    483  * when we booted.
    484  * XXX: not locked.
    485  */
    486 void
    487 tc_setclock(struct timespec *ts)
    488 {
    489 	struct timespec ts2;
    490 	struct bintime bt, bt2;
    491 
    492 	nsetclock.ev_count++;
    493 	binuptime(&bt2);
    494 	timespec2bintime(ts, &bt);
    495 	bintime_sub(&bt, &bt2);
    496 	bintime_add(&bt2, &timebasebin);
    497 	timebasebin = bt;
    498 
    499 	/* XXX fiddle all the little crinkly bits around the fiords... */
    500 	tc_windup();
    501 	if (timestepwarnings) {
    502 		bintime2timespec(&bt2, &ts2);
    503 		log(LOG_INFO, "Time stepped from %jd.%09ld to %jd.%09ld\n",
    504 		    (intmax_t)ts2.tv_sec, ts2.tv_nsec,
    505 		    (intmax_t)ts->tv_sec, ts->tv_nsec);
    506 	}
    507 }
    508 
    509 /*
    510  * Initialize the next struct timehands in the ring and make
    511  * it the active timehands.  Along the way we might switch to a different
    512  * timecounter and/or do seconds processing in NTP.  Slightly magic.
    513  */
    514 static void
    515 tc_windup(void)
    516 {
    517 	struct bintime bt;
    518 	struct timehands *th, *tho;
    519 	u_int64_t scale;
    520 	u_int delta, ncount, ogen;
    521 	int i;
    522 	time_t t;
    523 
    524 	/*
    525 	 * Make the next timehands a copy of the current one, but do not
    526 	 * overwrite the generation or next pointer.  While we update
    527 	 * the contents, the generation must be zero.
    528 	 */
    529 	tho = timehands;
    530 	th = tho->th_next;
    531 	ogen = th->th_generation;
    532 	th->th_generation = 0;
    533 	bcopy(tho, th, offsetof(struct timehands, th_generation));
    534 
    535 	/*
    536 	 * Capture a timecounter delta on the current timecounter and if
    537 	 * changing timecounters, a counter value from the new timecounter.
    538 	 * Update the offset fields accordingly.
    539 	 */
    540 	delta = tc_delta(th);
    541 	if (th->th_counter != timecounter)
    542 		ncount = timecounter->tc_get_timecount(timecounter);
    543 	else
    544 		ncount = 0;
    545 	th->th_offset_count += delta;
    546 	th->th_offset_count &= th->th_counter->tc_counter_mask;
    547 	bintime_addx(&th->th_offset, th->th_scale * delta);
    548 
    549 	/*
    550 	 * Hardware latching timecounters may not generate interrupts on
    551 	 * PPS events, so instead we poll them.  There is a finite risk that
    552 	 * the hardware might capture a count which is later than the one we
    553 	 * got above, and therefore possibly in the next NTP second which might
    554 	 * have a different rate than the current NTP second.  It doesn't
    555 	 * matter in practice.
    556 	 */
    557 	if (tho->th_counter->tc_poll_pps)
    558 		tho->th_counter->tc_poll_pps(tho->th_counter);
    559 
    560 	/*
    561 	 * Deal with NTP second processing.  The for loop normally
    562 	 * iterates at most once, but in extreme situations it might
    563 	 * keep NTP sane if timeouts are not run for several seconds.
    564 	 * At boot, the time step can be large when the TOD hardware
    565 	 * has been read, so on really large steps, we call
    566 	 * ntp_update_second only twice.  We need to call it twice in
    567 	 * case we missed a leap second.
    568 	 * If NTP is not compiled in ntp_update_second still calculates
    569 	 * the adjustment resulting from adjtime() calls.
    570 	 */
    571 	bt = th->th_offset;
    572 	bintime_add(&bt, &timebasebin);
    573 	i = bt.sec - tho->th_microtime.tv_sec;
    574 	if (i > LARGE_STEP)
    575 		i = 2;
    576 	for (; i > 0; i--) {
    577 		t = bt.sec;
    578 		ntp_update_second(&th->th_adjustment, &bt.sec);
    579 		if (bt.sec != t)
    580 			timebasebin.sec += bt.sec - t;
    581 	}
    582 
    583 	/* Update the UTC timestamps used by the get*() functions. */
    584 	/* XXX shouldn't do this here.  Should force non-`get' versions. */
    585 	bintime2timeval(&bt, &th->th_microtime);
    586 	bintime2timespec(&bt, &th->th_nanotime);
    587 
    588 	/* Now is a good time to change timecounters. */
    589 	if (th->th_counter != timecounter) {
    590 		th->th_counter = timecounter;
    591 		th->th_offset_count = ncount;
    592 
    593 		printf("timecounter: selected timecounter \"%s\" frequency %ju Hz quality %d\n",
    594 		    timecounter->tc_name, (uintmax_t)timecounter->tc_frequency,
    595 		    timecounter->tc_quality);
    596 	}
    597 
    598 	/*-
    599 	 * Recalculate the scaling factor.  We want the number of 1/2^64
    600 	 * fractions of a second per period of the hardware counter, taking
    601 	 * into account the th_adjustment factor which the NTP PLL/adjtime(2)
    602 	 * processing provides us with.
    603 	 *
    604 	 * The th_adjustment is nanoseconds per second with 32 bit binary
    605 	 * fraction and we want 64 bit binary fraction of second:
    606 	 *
    607 	 *	 x = a * 2^32 / 10^9 = a * 4.294967296
    608 	 *
    609 	 * The range of th_adjustment is +/- 5000PPM so inside a 64bit int
    610 	 * we can only multiply by about 850 without overflowing, but that
    611 	 * leaves suitably precise fractions for multiply before divide.
    612 	 *
    613 	 * Divide before multiply with a fraction of 2199/512 results in a
    614 	 * systematic undercompensation of 10PPM of th_adjustment.  On a
    615 	 * 5000PPM adjustment this is a 0.05PPM error.  This is acceptable.
    616  	 *
    617 	 * We happily sacrifice the lowest of the 64 bits of our result
    618 	 * to the goddess of code clarity.
    619 	 *
    620 	 */
    621 	scale = (u_int64_t)1 << 63;
    622 	scale += (th->th_adjustment / 1024) * 2199;
    623 	scale /= th->th_counter->tc_frequency;
    624 	th->th_scale = scale * 2;
    625 
    626 	/*
    627 	 * Now that the struct timehands is again consistent, set the new
    628 	 * generation number, making sure to not make it zero.
    629 	 */
    630 	if (++ogen == 0)
    631 		ogen = 1;
    632 	th->th_generation = ogen;
    633 
    634 	/* Go live with the new struct timehands. */
    635 	time_second = th->th_microtime.tv_sec;
    636 	time_uptime = th->th_offset.sec;
    637 	timehands = th;
    638 }
    639 
    640 #ifdef __FreeBSD__
    641 /* Report or change the active timecounter hardware. */
    642 static int
    643 sysctl_kern_timecounter_hardware(SYSCTL_HANDLER_ARGS)
    644 {
    645 	char newname[32];
    646 	struct timecounter *newtc, *tc;
    647 	int error;
    648 
    649 	tc = timecounter;
    650 	strlcpy(newname, tc->tc_name, sizeof(newname));
    651 
    652 	error = sysctl_handle_string(oidp, &newname[0], sizeof(newname), req);
    653 	if (error != 0 || req->newptr == NULL ||
    654 	    strcmp(newname, tc->tc_name) == 0)
    655 		return (error);
    656 
    657 	for (newtc = timecounters; newtc != NULL; newtc = newtc->tc_next) {
    658 		if (strcmp(newname, newtc->tc_name) != 0)
    659 			continue;
    660 
    661 		/* Warm up new timecounter. */
    662 		(void)newtc->tc_get_timecount(newtc);
    663 		(void)newtc->tc_get_timecount(newtc);
    664 
    665 		timecounter = newtc;
    666 		return (0);
    667 	}
    668 	return (EINVAL);
    669 }
    670 
    671 SYSCTL_PROC(_kern_timecounter, OID_AUTO, hardware, CTLTYPE_STRING | CTLFLAG_RW,
    672     0, 0, sysctl_kern_timecounter_hardware, "A", "");
    673 
    674 
    675 /* Report or change the active timecounter hardware. */
    676 static int
    677 sysctl_kern_timecounter_choice(SYSCTL_HANDLER_ARGS)
    678 {
    679 	char buf[32], *spc;
    680 	struct timecounter *tc;
    681 	int error;
    682 
    683 	spc = "";
    684 	error = 0;
    685 	for (tc = timecounters; error == 0 && tc != NULL; tc = tc->tc_next) {
    686 		sprintf(buf, "%s%s(%d)",
    687 		    spc, tc->tc_name, tc->tc_quality);
    688 		error = SYSCTL_OUT(req, buf, strlen(buf));
    689 		spc = " ";
    690 	}
    691 	return (error);
    692 }
    693 
    694 SYSCTL_PROC(_kern_timecounter, OID_AUTO, choice, CTLTYPE_STRING | CTLFLAG_RD,
    695     0, 0, sysctl_kern_timecounter_choice, "A", "");
    696 #endif /* __FreeBSD__ */
    697 
    698 /*
    699  * RFC 2783 PPS-API implementation.
    700  */
    701 
    702 int
    703 pps_ioctl(u_long cmd, caddr_t data, struct pps_state *pps)
    704 {
    705 	pps_params_t *app;
    706 	pps_info_t *pipi;
    707 #ifdef PPS_SYNC
    708 	int *epi;
    709 #endif
    710 
    711 	KASSERT(pps != NULL); /* XXX ("NULL pps pointer in pps_ioctl") */
    712 	switch (cmd) {
    713 	case PPS_IOC_CREATE:
    714 		return (0);
    715 	case PPS_IOC_DESTROY:
    716 		return (0);
    717 	case PPS_IOC_SETPARAMS:
    718 		app = (pps_params_t *)data;
    719 		if (app->mode & ~pps->ppscap)
    720 			return (EINVAL);
    721 		pps->ppsparam = *app;
    722 		return (0);
    723 	case PPS_IOC_GETPARAMS:
    724 		app = (pps_params_t *)data;
    725 		*app = pps->ppsparam;
    726 		app->api_version = PPS_API_VERS_1;
    727 		return (0);
    728 	case PPS_IOC_GETCAP:
    729 		*(int*)data = pps->ppscap;
    730 		return (0);
    731 	case PPS_IOC_FETCH:
    732 		pipi = (pps_info_t *)data;
    733 		pps->ppsinfo.current_mode = pps->ppsparam.mode;
    734 		*pipi = pps->ppsinfo;
    735 		return (0);
    736 	case PPS_IOC_KCBIND:
    737 #ifdef PPS_SYNC
    738 		epi = (int *)data;
    739 		/* XXX Only root should be able to do this */
    740 		if (*epi & ~pps->ppscap)
    741 			return (EINVAL);
    742 		pps->kcmode = *epi;
    743 		return (0);
    744 #else
    745 		return (EOPNOTSUPP);
    746 #endif
    747 	default:
    748 		return (EPASSTHROUGH);
    749 	}
    750 }
    751 
    752 void
    753 pps_init(struct pps_state *pps)
    754 {
    755 	pps->ppscap |= PPS_TSFMT_TSPEC;
    756 	if (pps->ppscap & PPS_CAPTUREASSERT)
    757 		pps->ppscap |= PPS_OFFSETASSERT;
    758 	if (pps->ppscap & PPS_CAPTURECLEAR)
    759 		pps->ppscap |= PPS_OFFSETCLEAR;
    760 }
    761 
    762 void
    763 pps_capture(struct pps_state *pps)
    764 {
    765 	struct timehands *th;
    766 
    767 	KASSERT(pps != NULL); /* XXX ("NULL pps pointer in pps_capture") */
    768 	th = timehands;
    769 	pps->capgen = th->th_generation;
    770 	pps->capth = th;
    771 	pps->capcount = th->th_counter->tc_get_timecount(th->th_counter);
    772 	if (pps->capgen != th->th_generation)
    773 		pps->capgen = 0;
    774 }
    775 
    776 void
    777 pps_event(struct pps_state *pps, int event)
    778 {
    779 	struct bintime bt;
    780 	struct timespec ts, *tsp, *osp;
    781 	u_int tcount, *pcount;
    782 	int foff, fhard;
    783 	pps_seq_t *pseq;
    784 
    785 	KASSERT(pps != NULL); /* XXX ("NULL pps pointer in pps_event") */
    786 	/* If the timecounter was wound up underneath us, bail out. */
    787 	if (pps->capgen == 0 || pps->capgen != pps->capth->th_generation)
    788 		return;
    789 
    790 	/* Things would be easier with arrays. */
    791 	if (event == PPS_CAPTUREASSERT) {
    792 		tsp = &pps->ppsinfo.assert_timestamp;
    793 		osp = &pps->ppsparam.assert_offset;
    794 		foff = pps->ppsparam.mode & PPS_OFFSETASSERT;
    795 		fhard = pps->kcmode & PPS_CAPTUREASSERT;
    796 		pcount = &pps->ppscount[0];
    797 		pseq = &pps->ppsinfo.assert_sequence;
    798 	} else {
    799 		tsp = &pps->ppsinfo.clear_timestamp;
    800 		osp = &pps->ppsparam.clear_offset;
    801 		foff = pps->ppsparam.mode & PPS_OFFSETCLEAR;
    802 		fhard = pps->kcmode & PPS_CAPTURECLEAR;
    803 		pcount = &pps->ppscount[1];
    804 		pseq = &pps->ppsinfo.clear_sequence;
    805 	}
    806 
    807 	/*
    808 	 * If the timecounter changed, we cannot compare the count values, so
    809 	 * we have to drop the rest of the PPS-stuff until the next event.
    810 	 */
    811 	if (pps->ppstc != pps->capth->th_counter) {
    812 		pps->ppstc = pps->capth->th_counter;
    813 		*pcount = pps->capcount;
    814 		pps->ppscount[2] = pps->capcount;
    815 		return;
    816 	}
    817 
    818 	/* Convert the count to a timespec. */
    819 	tcount = pps->capcount - pps->capth->th_offset_count;
    820 	tcount &= pps->capth->th_counter->tc_counter_mask;
    821 	bt = pps->capth->th_offset;
    822 	bintime_addx(&bt, pps->capth->th_scale * tcount);
    823 	bintime_add(&bt, &timebasebin);
    824 	bintime2timespec(&bt, &ts);
    825 
    826 	/* If the timecounter was wound up underneath us, bail out. */
    827 	if (pps->capgen != pps->capth->th_generation)
    828 		return;
    829 
    830 	*pcount = pps->capcount;
    831 	(*pseq)++;
    832 	*tsp = ts;
    833 
    834 	if (foff) {
    835 		timespecadd(tsp, osp, tsp);
    836 		if (tsp->tv_nsec < 0) {
    837 			tsp->tv_nsec += 1000000000;
    838 			tsp->tv_sec -= 1;
    839 		}
    840 	}
    841 #ifdef PPS_SYNC
    842 	if (fhard) {
    843 		u_int64_t scale;
    844 
    845 		/*
    846 		 * Feed the NTP PLL/FLL.
    847 		 * The FLL wants to know how many (hardware) nanoseconds
    848 		 * elapsed since the previous event.
    849 		 */
    850 		tcount = pps->capcount - pps->ppscount[2];
    851 		pps->ppscount[2] = pps->capcount;
    852 		tcount &= pps->capth->th_counter->tc_counter_mask;
    853 		scale = (u_int64_t)1 << 63;
    854 		scale /= pps->capth->th_counter->tc_frequency;
    855 		scale *= 2;
    856 		bt.sec = 0;
    857 		bt.frac = 0;
    858 		bintime_addx(&bt, scale * tcount);
    859 		bintime2timespec(&bt, &ts);
    860 		hardpps(tsp, ts.tv_nsec + 1000000000 * ts.tv_sec);
    861 	}
    862 #endif
    863 }
    864 
    865 /*
    866  * Timecounters need to be updated every so often to prevent the hardware
    867  * counter from overflowing.  Updating also recalculates the cached values
    868  * used by the get*() family of functions, so their precision depends on
    869  * the update frequency.
    870  */
    871 
    872 static int tc_tick;
    873 #ifdef __FreeBSD__
    874 SYSCTL_INT(_kern_timecounter, OID_AUTO, tick, CTLFLAG_RD, &tc_tick, 0, "");
    875 #endif /* __FreeBSD__ */
    876 
    877 void
    878 tc_ticktock(void)
    879 {
    880 	static int count;
    881 
    882 	if (++count < tc_tick)
    883 		return;
    884 	count = 0;
    885 	tc_windup();
    886 }
    887 
    888 void
    889 inittimecounter(void)
    890 {
    891 	u_int p;
    892 
    893 	/*
    894 	 * Set the initial timeout to
    895 	 * max(1, <approx. number of hardclock ticks in a millisecond>).
    896 	 * People should probably not use the sysctl to set the timeout
    897 	 * to smaller than its inital value, since that value is the
    898 	 * smallest reasonable one.  If they want better timestamps they
    899 	 * should use the non-"get"* functions.
    900 	 */
    901 	if (hz > 1000)
    902 		tc_tick = (hz + 500) / 1000;
    903 	else
    904 		tc_tick = 1;
    905 	p = (tc_tick * 1000000) / hz;
    906 	printf("timecounter: Timecounters tick every %d.%03u msec\n", p / 1000, p % 1000);
    907 
    908 	/* warm up new timecounter (again) and get rolling. */
    909 	(void)timecounter->tc_get_timecount(timecounter);
    910 	(void)timecounter->tc_get_timecount(timecounter);
    911 }
    912 
    913 #ifdef __FreeBSD__
    914 SYSINIT(timecounter, SI_SUB_CLOCKS, SI_ORDER_SECOND, inittimecounter, NULL)
    915 #endif /* __FreeBSD__ */
    916 #endif /* __HAVE_TIMECOUNTER */
    917