2 * NTP client/server, based on OpenNTPD 3.9p1
4 * Author: Adam Tkac <vonsch@gmail.com>
6 * Licensed under GPLv2, see file LICENSE in this tarball for details.
8 * Parts of OpenNTPD clock syncronization code is replaced by
9 * code which is based on ntp-4.2.6, whuch carries the following
12 ***********************************************************************
14 * Copyright (c) University of Delaware 1992-2009 *
16 * Permission to use, copy, modify, and distribute this software and *
17 * its documentation for any purpose with or without fee is hereby *
18 * granted, provided that the above copyright notice appears in all *
19 * copies and that both the copyright notice and this permission *
20 * notice appear in supporting documentation, and that the name *
21 * University of Delaware not be used in advertising or publicity *
22 * pertaining to distribution of the software without specific, *
23 * written prior permission. The University of Delaware makes no *
24 * representations about the suitability this software for any *
25 * purpose. It is provided "as is" without express or implied *
28 ***********************************************************************
32 #include <netinet/ip.h> /* For IPTOS_LOWDELAY definition */
33 #include <sys/timex.h>
34 #ifndef IPTOS_LOWDELAY
35 # define IPTOS_LOWDELAY 0x10
38 # error "Sorry, your kernel has to support IP_PKTINFO"
42 /* Verbosity control (max level of -dddd options accepted).
43 * max 5 is very talkative (and bloated). 2 is non-bloated,
44 * production level setting.
49 #define RETRY_INTERVAL 5 /* on error, retry in N secs */
50 #define RESPONSE_INTERVAL 15 /* wait for reply up to N secs */
52 #define FREQ_TOLERANCE 0.000015 /* % frequency tolerance (15 PPM) */
54 #define MINPOLL 4 /* % minimum poll interval (6: 64 s) */
55 #define MAXPOLL 12 /* % maximum poll interval (12: 1.1h, 17: 36.4h) (was 17) */
56 #define MINDISP 0.01 /* % minimum dispersion (s) */
57 #define MAXDISP 16 /* maximum dispersion (s) */
58 #define MAXSTRAT 16 /* maximum stratum (infinity metric) */
59 #define MAXDIST 1 /* % distance threshold (s) */
60 #define MIN_SELECTED 1 /* % minimum intersection survivors */
61 #define MIN_CLUSTERED 3 /* % minimum cluster survivors */
63 #define MAXDRIFT 0.000500 /* frequency drift we can correct (500 PPM) */
65 /* Clock discipline parameters and constants */
66 #define STEP_THRESHOLD 0.128 /* step threshold (s) */
67 #define WATCH_THRESHOLD 150 /* stepout threshold (s). std ntpd uses 900 (11 mins (!)) */
68 /* NB: set WATCH_THRESHOLD to ~60 when debugging to save time) */
69 #define PANIC_THRESHOLD 1000 /* panic threshold (s) */
71 /* Poll-adjust threshold.
72 * When we see that offset is small enough compared to discipline jitter,
73 * we grow a counter: += MINPOLL. When it goes over POLLADJ_LIMIT,
74 * we poll_exp++. If offset isn't small, counter -= poll_exp*2,
75 * and when it goes below -POLLADJ_LIMIT, we poll_exp--
77 #define POLLADJ_LIMIT 30
78 /* If offset < POLLADJ_GATE * discipline_jitter, then we can increase
79 * poll interval (we think we can't improve timekeeping
80 * by staying at smaller poll).
82 #define POLLADJ_GATE 4
83 /* Compromise Allan intercept (s). doc uses 1500, std ntpd uses 512 */
87 /* FLL loop gain [why it depends on MAXPOLL??] */
88 #define FLL (MAXPOLL + 1)
89 /* Parameter averaging constant */
98 NTP_MSGSIZE_NOAUTH = 48,
99 NTP_MSGSIZE = (NTP_MSGSIZE_NOAUTH + 4 + NTP_DIGESTSIZE),
102 MODE_MASK = (7 << 0),
103 VERSION_MASK = (7 << 3),
107 /* Leap Second Codes (high order two bits of m_status) */
108 LI_NOWARNING = (0 << 6), /* no warning */
109 LI_PLUSSEC = (1 << 6), /* add a second (61 seconds) */
110 LI_MINUSSEC = (2 << 6), /* minus a second (59 seconds) */
111 LI_ALARM = (3 << 6), /* alarm condition */
114 MODE_RES0 = 0, /* reserved */
115 MODE_SYM_ACT = 1, /* symmetric active */
116 MODE_SYM_PAS = 2, /* symmetric passive */
117 MODE_CLIENT = 3, /* client */
118 MODE_SERVER = 4, /* server */
119 MODE_BROADCAST = 5, /* broadcast */
120 MODE_RES1 = 6, /* reserved for NTP control message */
121 MODE_RES2 = 7, /* reserved for private use */
124 //TODO: better base selection
125 #define OFFSET_1900_1970 2208988800UL /* 1970 - 1900 in seconds */
127 #define NUM_DATAPOINTS 8
140 uint8_t m_status; /* status of local clock and leap info */
142 uint8_t m_ppoll; /* poll value */
143 int8_t m_precision_exp;
144 s_fixedpt_t m_rootdelay;
145 s_fixedpt_t m_rootdisp;
147 l_fixedpt_t m_reftime;
148 l_fixedpt_t m_orgtime;
149 l_fixedpt_t m_rectime;
150 l_fixedpt_t m_xmttime;
152 uint8_t m_digest[NTP_DIGESTSIZE];
162 len_and_sockaddr *p_lsa;
164 /* when to send new query (if p_fd == -1)
165 * or when receive times out (if p_fd >= 0): */
168 uint32_t lastpkt_refid;
169 uint8_t lastpkt_status;
170 uint8_t lastpkt_stratum;
171 uint8_t reachable_bits;
172 double next_action_time;
174 double lastpkt_recv_time;
175 double lastpkt_delay;
176 double lastpkt_rootdelay;
177 double lastpkt_rootdisp;
178 /* produced by filter algorithm: */
179 double filter_offset;
180 double filter_dispersion;
181 double filter_jitter;
182 datapoint_t filter_datapoint[NUM_DATAPOINTS];
183 /* last sent packet: */
193 /* Insert new options above this line. */
194 /* Non-compat options: */
197 OPT_l = (1 << 6) * ENABLE_FEATURE_NTPD_SERVER,
202 /* total round trip delay to currently selected reference clock */
204 /* reference timestamp: time when the system clock was last set or corrected */
206 /* total dispersion to currently selected reference clock */
209 #if ENABLE_FEATURE_NTPD_SERVER
214 /* refid: 32-bit code identifying the particular server or reference clock
215 * in stratum 0 packets this is a four-character ASCII string,
216 * called the kiss code, used for debugging and monitoring
217 * in stratum 1 packets this is a four-character ASCII string
218 * assigned to the reference clock by IANA. Example: "GPS "
219 * in stratum 2+ packets, it's IPv4 address or 4 first bytes of MD5 hash of IPv6
223 /* precision is defined as the larger of the resolution and time to
224 * read the clock, in log2 units. For instance, the precision of a
225 * mains-frequency clock incrementing at 60 Hz is 16 ms, even when the
226 * system clock hardware representation is to the nanosecond.
228 * Delays, jitters of various kinds are clamper down to precision.
230 * If precision_sec is too large, discipline_jitter gets clamped to it
231 * and if offset is much smaller than discipline_jitter, poll interval
232 * grows even though we really can benefit from staying at smaller one,
233 * collecting non-lagged datapoits and correcting the offset.
234 * (Lagged datapoits exist when poll_exp is large but we still have
235 * systematic offset error - the time distance between datapoints
236 * is significat and older datapoints have smaller offsets.
237 * This makes our offset estimation a bit smaller than reality)
238 * Due to this effect, setting G_precision_sec close to
239 * STEP_THRESHOLD isn't such a good idea - offsets may grow
240 * too big and we will step. I observed it with -6.
242 * OTOH, setting precision too small would result in futile attempts
243 * to syncronize to the unachievable precision.
245 * -6 is 1/64 sec, -7 is 1/128 sec and so on.
247 #define G_precision_exp -8
248 #define G_precision_sec (1.0 / (1 << (- G_precision_exp)))
250 /* Bool. After set to 1, never goes back to 0: */
251 smallint adjtimex_was_done;
252 smallint initial_poll_complete;
254 uint8_t discipline_state; // doc calls it c.state
255 uint8_t poll_exp; // s.poll
256 int polladj_count; // c.count
257 long kernel_freq_drift;
258 double last_update_offset; // c.last
259 double last_update_recv_time; // s.t
260 double discipline_jitter; // c.jitter
261 //TODO: add s.jitter - grep for it here and see clock_combine() in doc
262 #define USING_KERNEL_PLL_LOOP 1
263 #if !USING_KERNEL_PLL_LOOP
264 double discipline_freq_drift; // c.freq
265 //TODO: conditionally calculate wander? it's used only for logging
266 double discipline_wander; // c.wander
269 #define G (*ptr_to_globals)
271 static const int const_IPTOS_LOWDELAY = IPTOS_LOWDELAY;
274 #define VERB1 if (MAX_VERBOSE && G.verbose)
275 #define VERB2 if (MAX_VERBOSE >= 2 && G.verbose >= 2)
276 #define VERB3 if (MAX_VERBOSE >= 3 && G.verbose >= 3)
277 #define VERB4 if (MAX_VERBOSE >= 4 && G.verbose >= 4)
278 #define VERB5 if (MAX_VERBOSE >= 5 && G.verbose >= 5)
281 static double LOG2D(int a)
284 return 1.0 / (1UL << -a);
287 static ALWAYS_INLINE double SQUARE(double x)
291 static ALWAYS_INLINE double MAXD(double a, double b)
297 static ALWAYS_INLINE double MIND(double a, double b)
303 static NOINLINE double my_SQRT(double X)
310 double Xhalf = X * 0.5;
312 /* Fast and good approximation to 1/sqrt(X), black magic */
314 /*v.i = 0x5f3759df - (v.i >> 1);*/
315 v.i = 0x5f375a86 - (v.i >> 1); /* - this constant is slightly better */
316 invsqrt = v.f; /* better than 0.2% accuracy */
318 /* Refining it using Newton's method: x1 = x0 - f(x0)/f'(x0)
319 * f(x) = 1/(x*x) - X (f==0 when x = 1/sqrt(X))
321 * f(x)/f'(x) = (X - 1/(x*x)) / (2/(x*x*x)) = X*x*x*x/2 - x/2
322 * x1 = x0 - (X*x0*x0*x0/2 - x0/2) = 1.5*x0 - X*x0*x0*x0/2 = x0*(1.5 - (X/2)*x0*x0)
324 invsqrt = invsqrt * (1.5 - Xhalf * invsqrt * invsqrt); /* ~0.05% accuracy */
325 /* invsqrt = invsqrt * (1.5 - Xhalf * invsqrt * invsqrt); 2nd iter: ~0.0001% accuracy */
326 /* With 4 iterations, more than half results will be exact,
327 * at 6th iterations result stabilizes with about 72% results exact.
328 * We are well satisfied with 0.05% accuracy.
331 return X * invsqrt; /* X * 1/sqrt(X) ~= sqrt(X) */
333 static ALWAYS_INLINE double SQRT(double X)
335 /* If this arch doesn't use IEEE 754 floats, fall back to using libm */
336 if (sizeof(float) != 4)
339 /* This avoids needing libm, saves about 0.5k on x86-32 */
347 gettimeofday(&tv, NULL); /* never fails */
348 G.cur_time = tv.tv_sec + (1.0e-6 * tv.tv_usec) + OFFSET_1900_1970;
353 d_to_tv(double d, struct timeval *tv)
355 tv->tv_sec = (long)d;
356 tv->tv_usec = (d - tv->tv_sec) * 1000000;
360 lfp_to_d(l_fixedpt_t lfp)
363 lfp.int_partl = ntohl(lfp.int_partl);
364 lfp.fractionl = ntohl(lfp.fractionl);
365 ret = (double)lfp.int_partl + ((double)lfp.fractionl / UINT_MAX);
369 sfp_to_d(s_fixedpt_t sfp)
372 sfp.int_parts = ntohs(sfp.int_parts);
373 sfp.fractions = ntohs(sfp.fractions);
374 ret = (double)sfp.int_parts + ((double)sfp.fractions / USHRT_MAX);
377 #if ENABLE_FEATURE_NTPD_SERVER
382 lfp.int_partl = (uint32_t)d;
383 lfp.fractionl = (uint32_t)((d - lfp.int_partl) * UINT_MAX);
384 lfp.int_partl = htonl(lfp.int_partl);
385 lfp.fractionl = htonl(lfp.fractionl);
392 sfp.int_parts = (uint16_t)d;
393 sfp.fractions = (uint16_t)((d - sfp.int_parts) * USHRT_MAX);
394 sfp.int_parts = htons(sfp.int_parts);
395 sfp.fractions = htons(sfp.fractions);
401 dispersion(const datapoint_t *dp)
403 return dp->d_dispersion + FREQ_TOLERANCE * (G.cur_time - dp->d_recv_time);
407 root_distance(peer_t *p)
409 /* The root synchronization distance is the maximum error due to
410 * all causes of the local clock relative to the primary server.
411 * It is defined as half the total delay plus total dispersion
414 return MAXD(MINDISP, p->lastpkt_rootdelay + p->lastpkt_delay) / 2
415 + p->lastpkt_rootdisp
416 + p->filter_dispersion
417 + FREQ_TOLERANCE * (G.cur_time - p->lastpkt_recv_time)
422 set_next(peer_t *p, unsigned t)
424 p->next_action_time = G.cur_time + t;
428 * Peer clock filter and its helpers
431 filter_datapoints(peer_t *p)
435 double minoff, maxoff, wavg, sum, w;
436 double x = x; /* for compiler */
437 double oldest_off = oldest_off;
438 double oldest_age = oldest_age;
439 double newest_off = newest_off;
440 double newest_age = newest_age;
442 minoff = maxoff = p->filter_datapoint[0].d_offset;
443 for (i = 1; i < NUM_DATAPOINTS; i++) {
444 if (minoff > p->filter_datapoint[i].d_offset)
445 minoff = p->filter_datapoint[i].d_offset;
446 if (maxoff < p->filter_datapoint[i].d_offset)
447 maxoff = p->filter_datapoint[i].d_offset;
450 idx = p->datapoint_idx; /* most recent datapoint */
452 * Drop two outliers and take weighted average of the rest:
453 * most_recent/2 + older1/4 + older2/8 ... + older5/32 + older6/32
454 * we use older6/32, not older6/64 since sum of weights should be 1:
455 * 1/2 + 1/4 + 1/8 + 1/16 + 1/32 + 1/32 = 1
461 * filter_dispersion = \ -------------
468 for (i = 0; i < NUM_DATAPOINTS; i++) {
470 bb_error_msg("datapoint[%d]: off:%f disp:%f(%f) age:%f%s",
472 p->filter_datapoint[idx].d_offset,
473 p->filter_datapoint[idx].d_dispersion, dispersion(&p->filter_datapoint[idx]),
474 G.cur_time - p->filter_datapoint[idx].d_recv_time,
475 (minoff == p->filter_datapoint[idx].d_offset || maxoff == p->filter_datapoint[idx].d_offset)
476 ? " (outlier by offset)" : ""
480 sum += dispersion(&p->filter_datapoint[idx]) / (2 << i);
482 if (minoff == p->filter_datapoint[idx].d_offset) {
483 minoff -= 1; /* so that we don't match it ever again */
485 if (maxoff == p->filter_datapoint[idx].d_offset) {
488 oldest_off = p->filter_datapoint[idx].d_offset;
489 oldest_age = G.cur_time - p->filter_datapoint[idx].d_recv_time;
492 newest_off = oldest_off;
493 newest_age = oldest_age;
500 idx = (idx - 1) & (NUM_DATAPOINTS - 1);
502 p->filter_dispersion = sum;
503 wavg += x; /* add another older6/64 to form older6/32 */
504 /* Fix systematic underestimation with large poll intervals.
505 * Imagine that we still have a bit of uncorrected drift,
506 * and poll interval is big (say, 100 sec). Offsets form a progression:
507 * 0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 - 0.7 is most recent.
508 * The algorithm above drops 0.0 and 0.7 as outliers,
509 * and then we have this estimation, ~25% off from 0.7:
510 * 0.1/32 + 0.2/32 + 0.3/16 + 0.4/8 + 0.5/4 + 0.6/2 = 0.503125
512 x = oldest_age - newest_age;
514 x = newest_age / x; /* in above example, 100 / (600 - 100) */
515 if (x < 1) { /* paranoia check */
516 x = (newest_off - oldest_off) * x; /* 0.5 * 100/500 = 0.1 */
520 p->filter_offset = wavg;
522 /* +----- -----+ ^ 1/2
526 * filter_jitter = | --- * / (avg-offset_j) |
530 * where n is the number of valid datapoints in the filter (n > 1);
531 * if filter_jitter < precision then filter_jitter = precision
534 for (i = 0; i < NUM_DATAPOINTS; i++) {
535 sum += SQUARE(wavg - p->filter_datapoint[i].d_offset);
537 sum = SQRT(sum / NUM_DATAPOINTS);
538 p->filter_jitter = sum > G_precision_sec ? sum : G_precision_sec;
540 VERB3 bb_error_msg("filter offset:%f(corr:%e) disp:%f jitter:%f",
542 p->filter_dispersion,
548 reset_peer_stats(peer_t *p, double offset)
551 for (i = 0; i < NUM_DATAPOINTS; i++) {
552 if (offset < 16 * STEP_THRESHOLD) {
553 p->filter_datapoint[i].d_recv_time -= offset;
554 if (p->filter_datapoint[i].d_offset != 0) {
555 p->filter_datapoint[i].d_offset -= offset;
558 p->filter_datapoint[i].d_recv_time = G.cur_time;
559 p->filter_datapoint[i].d_offset = 0;
560 p->filter_datapoint[i].d_dispersion = MAXDISP;
563 if (offset < 16 * STEP_THRESHOLD) {
564 p->lastpkt_recv_time -= offset;
566 p->reachable_bits = 0;
567 p->lastpkt_recv_time = G.cur_time;
569 filter_datapoints(p); /* recalc p->filter_xxx */
570 p->next_action_time -= offset;
571 VERB5 bb_error_msg("%s->lastpkt_recv_time=%f", p->p_dotted, p->lastpkt_recv_time);
579 p = xzalloc(sizeof(*p));
580 p->p_lsa = xhost2sockaddr(s, 123);
581 p->p_dotted = xmalloc_sockaddr2dotted_noport(&p->p_lsa->u.sa);
583 p->p_xmt_msg.m_status = MODE_CLIENT | (NTP_VERSION << 3);
584 p->next_action_time = G.cur_time; /* = set_next(p, 0); */
585 reset_peer_stats(p, 16 * STEP_THRESHOLD);
587 llist_add_to(&G.ntp_peers, p);
593 const struct sockaddr *from, const struct sockaddr *to, socklen_t addrlen,
594 msg_t *msg, ssize_t len)
600 ret = sendto(fd, msg, len, MSG_DONTWAIT, to, addrlen);
602 ret = send_to_from(fd, msg, len, MSG_DONTWAIT, to, from, addrlen);
605 bb_perror_msg("send failed");
612 send_query_to_peer(peer_t *p)
614 /* Why do we need to bind()?
615 * See what happens when we don't bind:
617 * socket(PF_INET, SOCK_DGRAM, IPPROTO_IP) = 3
618 * setsockopt(3, SOL_IP, IP_TOS, [16], 4) = 0
619 * gettimeofday({1259071266, 327885}, NULL) = 0
620 * sendto(3, "xxx", 48, MSG_DONTWAIT, {sa_family=AF_INET, sin_port=htons(123), sin_addr=inet_addr("10.34.32.125")}, 16) = 48
621 * ^^^ we sent it from some source port picked by kernel.
622 * time(NULL) = 1259071266
623 * write(2, "ntpd: entering poll 15 secs\n", 28) = 28
624 * poll([{fd=3, events=POLLIN}], 1, 15000) = 1 ([{fd=3, revents=POLLIN}])
625 * recv(3, "yyy", 68, MSG_DONTWAIT) = 48
626 * ^^^ this recv will receive packets to any local port!
628 * Uncomment this and use strace to see it in action:
630 #define PROBE_LOCAL_ADDR /* { len_and_sockaddr lsa; lsa.len = LSA_SIZEOF_SA; getsockname(p->query.fd, &lsa.u.sa, &lsa.len); } */
634 len_and_sockaddr *local_lsa;
636 family = p->p_lsa->u.sa.sa_family;
637 p->p_fd = fd = xsocket_type(&local_lsa, family, SOCK_DGRAM);
638 /* local_lsa has "null" address and port 0 now.
639 * bind() ensures we have a *particular port* selected by kernel
640 * and remembered in p->p_fd, thus later recv(p->p_fd)
641 * receives only packets sent to this port.
644 xbind(fd, &local_lsa->u.sa, local_lsa->len);
646 #if ENABLE_FEATURE_IPV6
647 if (family == AF_INET)
649 setsockopt(fd, IPPROTO_IP, IP_TOS, &const_IPTOS_LOWDELAY, sizeof(const_IPTOS_LOWDELAY));
654 * Send out a random 64-bit number as our transmit time. The NTP
655 * server will copy said number into the originate field on the
656 * response that it sends us. This is totally legal per the SNTP spec.
658 * The impact of this is two fold: we no longer send out the current
659 * system time for the world to see (which may aid an attacker), and
660 * it gives us a (not very secure) way of knowing that we're not
661 * getting spoofed by an attacker that can't capture our traffic
662 * but can spoof packets from the NTP server we're communicating with.
664 * Save the real transmit timestamp locally.
666 p->p_xmt_msg.m_xmttime.int_partl = random();
667 p->p_xmt_msg.m_xmttime.fractionl = random();
668 p->p_xmttime = gettime1900d();
670 if (do_sendto(p->p_fd, /*from:*/ NULL, /*to:*/ &p->p_lsa->u.sa, /*addrlen:*/ p->p_lsa->len,
671 &p->p_xmt_msg, NTP_MSGSIZE_NOAUTH) == -1
675 set_next(p, RETRY_INTERVAL);
679 p->reachable_bits <<= 1;
680 VERB1 bb_error_msg("sent query to %s", p->p_dotted);
681 set_next(p, RESPONSE_INTERVAL);
686 step_time(double offset)
694 gettimeofday(&tv, NULL); /* never fails */
695 dtime = offset + tv.tv_sec;
696 dtime += 1.0e-6 * tv.tv_usec;
699 if (settimeofday(&tv, NULL) == -1)
700 bb_perror_msg_and_die("settimeofday");
703 strftime(buf, sizeof(buf), "%a %b %e %H:%M:%S %Z %Y", localtime(&tval));
705 bb_error_msg("setting clock to %s (offset %fs)", buf, offset);
707 /* Correct various fields which contain time-relative values: */
709 /* p->lastpkt_recv_time, p->next_action_time and such: */
710 for (item = G.ntp_peers; item != NULL; item = item->link) {
711 peer_t *pp = (peer_t *) item->data;
712 reset_peer_stats(pp, offset);
715 G.cur_time -= offset;
716 G.last_update_recv_time -= offset;
721 * Selection and clustering, and their helpers
729 compare_point_edge(const void *aa, const void *bb)
731 const point_t *a = aa;
732 const point_t *b = bb;
733 if (a->edge < b->edge) {
736 return (a->edge > b->edge);
743 compare_survivor_metric(const void *aa, const void *bb)
745 const survivor_t *a = aa;
746 const survivor_t *b = bb;
747 if (a->metric < b->metric) {
750 return (a->metric > b->metric);
753 fit(peer_t *p, double rd)
755 if ((p->reachable_bits & (p->reachable_bits-1)) == 0) {
756 /* One or zero bits in reachable_bits */
757 VERB3 bb_error_msg("peer %s unfit for selection: unreachable", p->p_dotted);
760 #if 0 /* we filter out such packets earlier */
761 if ((p->lastpkt_status & LI_ALARM) == LI_ALARM
762 || p->lastpkt_stratum >= MAXSTRAT
764 VERB3 bb_error_msg("peer %s unfit for selection: bad status/stratum", p->p_dotted);
768 /* rd is root_distance(p) */
769 if (rd > MAXDIST + FREQ_TOLERANCE * (1 << G.poll_exp)) {
770 VERB3 bb_error_msg("peer %s unfit for selection: root distance too high", p->p_dotted);
774 // /* Do we have a loop? */
775 // if (p->refid == p->dstaddr || p->refid == s.refid)
780 select_and_cluster(void)
784 int size = 3 * G.peer_cnt;
785 /* for selection algorithm */
787 unsigned num_points, num_candidates;
789 unsigned num_falsetickers;
790 /* for cluster algorithm */
791 survivor_t survivor[size];
792 unsigned num_survivors;
798 if (G.initial_poll_complete) while (item != NULL) {
799 peer_t *p = (peer_t *) item->data;
800 double rd = root_distance(p);
801 double offset = p->filter_offset;
808 VERB4 bb_error_msg("interval: [%f %f %f] %s",
814 point[num_points].p = p;
815 point[num_points].type = -1;
816 point[num_points].edge = offset - rd;
818 point[num_points].p = p;
819 point[num_points].type = 0;
820 point[num_points].edge = offset;
822 point[num_points].p = p;
823 point[num_points].type = 1;
824 point[num_points].edge = offset + rd;
828 num_candidates = num_points / 3;
829 if (num_candidates == 0) {
830 VERB3 bb_error_msg("no valid datapoints, no peer selected");
833 //TODO: sorting does not seem to be done in reference code
834 qsort(point, num_points, sizeof(point[0]), compare_point_edge);
836 /* Start with the assumption that there are no falsetickers.
837 * Attempt to find a nonempty intersection interval containing
838 * the midpoints of all truechimers.
839 * If a nonempty interval cannot be found, increase the number
840 * of assumed falsetickers by one and try again.
841 * If a nonempty interval is found and the number of falsetickers
842 * is less than the number of truechimers, a majority has been found
843 * and the midpoint of each truechimer represents
844 * the candidates available to the cluster algorithm.
846 num_falsetickers = 0;
849 unsigned num_midpoints = 0;
854 for (i = 0; i < num_points; i++) {
856 * if (point[i].type == -1) c++;
857 * if (point[i].type == 1) c--;
858 * and it's simpler to do it this way:
861 if (c >= num_candidates - num_falsetickers) {
862 /* If it was c++ and it got big enough... */
866 if (point[i].type == 0)
870 for (i = num_points-1; i >= 0; i--) {
872 if (c >= num_candidates - num_falsetickers) {
873 high = point[i].edge;
876 if (point[i].type == 0)
879 /* If the number of midpoints is greater than the number
880 * of allowed falsetickers, the intersection contains at
881 * least one truechimer with no midpoint - bad.
882 * Also, interval should be nonempty.
884 if (num_midpoints <= num_falsetickers && low < high)
887 if (num_falsetickers * 2 >= num_candidates) {
888 VERB3 bb_error_msg("too many falsetickers:%d (candidates:%d), no peer selected",
889 num_falsetickers, num_candidates);
893 VERB3 bb_error_msg("selected interval: [%f, %f]; candidates:%d falsetickers:%d",
894 low, high, num_candidates, num_falsetickers);
898 /* Construct a list of survivors (p, metric)
899 * from the chime list, where metric is dominated
900 * first by stratum and then by root distance.
901 * All other things being equal, this is the order of preference.
904 for (i = 0; i < num_points; i++) {
907 if (point[i].edge < low || point[i].edge > high)
910 survivor[num_survivors].p = p;
911 //TODO: save root_distance in point_t and reuse here?
912 survivor[num_survivors].metric = MAXDIST * p->lastpkt_stratum + root_distance(p);
913 VERB4 bb_error_msg("survivor[%d] metric:%f peer:%s",
914 num_survivors, survivor[num_survivors].metric, p->p_dotted);
917 /* There must be at least MIN_SELECTED survivors to satisfy the
918 * correctness assertions. Ordinarily, the Byzantine criteria
919 * require four survivors, but for the demonstration here, one
922 if (num_survivors < MIN_SELECTED) {
923 VERB3 bb_error_msg("num_survivors %d < %d, no peer selected",
924 num_survivors, MIN_SELECTED);
928 //looks like this is ONLY used by the fact that later we pick survivor[0].
929 //we can avoid sorting then, just find the minimum once!
930 qsort(survivor, num_survivors, sizeof(survivor[0]), compare_survivor_metric);
932 /* For each association p in turn, calculate the selection
933 * jitter p->sjitter as the square root of the sum of squares
934 * (p->offset - q->offset) over all q associations. The idea is
935 * to repeatedly discard the survivor with maximum selection
936 * jitter until a termination condition is met.
939 unsigned max_idx = max_idx;
940 double max_selection_jitter = max_selection_jitter;
941 double min_jitter = min_jitter;
943 if (num_survivors <= MIN_CLUSTERED) {
944 VERB3 bb_error_msg("num_survivors %d <= %d, not discarding more",
945 num_survivors, MIN_CLUSTERED);
949 /* To make sure a few survivors are left
950 * for the clustering algorithm to chew on,
951 * we stop if the number of survivors
952 * is less than or equal to MIN_CLUSTERED (3).
954 for (i = 0; i < num_survivors; i++) {
955 double selection_jitter_sq;
956 peer_t *p = survivor[i].p;
958 if (i == 0 || p->filter_jitter < min_jitter)
959 min_jitter = p->filter_jitter;
961 selection_jitter_sq = 0;
962 for (j = 0; j < num_survivors; j++) {
963 peer_t *q = survivor[j].p;
964 selection_jitter_sq += SQUARE(p->filter_offset - q->filter_offset);
966 if (i == 0 || selection_jitter_sq > max_selection_jitter) {
967 max_selection_jitter = selection_jitter_sq;
970 VERB5 bb_error_msg("survivor %d selection_jitter^2:%f",
971 i, selection_jitter_sq);
973 max_selection_jitter = SQRT(max_selection_jitter / num_survivors);
974 VERB4 bb_error_msg("max_selection_jitter (at %d):%f min_jitter:%f",
975 max_idx, max_selection_jitter, min_jitter);
977 /* If the maximum selection jitter is less than the
978 * minimum peer jitter, then tossing out more survivors
979 * will not lower the minimum peer jitter, so we might
982 if (max_selection_jitter < min_jitter) {
983 VERB3 bb_error_msg("max_selection_jitter:%f < min_jitter:%f, num_survivors:%d, not discarding more",
984 max_selection_jitter, min_jitter, num_survivors);
988 /* Delete survivor[max_idx] from the list
989 * and go around again.
991 VERB5 bb_error_msg("dropping survivor %d", max_idx);
993 while (max_idx < num_survivors) {
994 survivor[max_idx] = survivor[max_idx + 1];
999 /* Pick the best clock. If the old system peer is on the list
1000 * and at the same stratum as the first survivor on the list,
1001 * then don't do a clock hop. Otherwise, select the first
1002 * survivor on the list as the new system peer.
1004 //TODO - see clock_combine()
1005 VERB3 bb_error_msg("selected peer %s filter_offset:%f age:%f",
1006 survivor[0].p->p_dotted,
1007 survivor[0].p->filter_offset,
1008 G.cur_time - survivor[0].p->lastpkt_recv_time
1010 return survivor[0].p;
1015 * Local clock discipline and its helpers
1018 set_new_values(int disc_state, double offset, double recv_time)
1020 /* Enter new state and set state variables. Note we use the time
1021 * of the last clock filter sample, which must be earlier than
1024 VERB3 bb_error_msg("disc_state=%d last update offset=%f recv_time=%f",
1025 disc_state, offset, recv_time);
1026 G.discipline_state = disc_state;
1027 G.last_update_offset = offset;
1028 G.last_update_recv_time = recv_time;
1030 /* Clock state definitions */
1031 #define STATE_NSET 0 /* initial state, "nothing is set" */
1032 #define STATE_FSET 1 /* frequency set from file */
1033 #define STATE_SPIK 2 /* spike detected */
1034 #define STATE_FREQ 3 /* initial frequency */
1035 #define STATE_SYNC 4 /* clock synchronized (normal operation) */
1036 /* Return: -1: decrease poll interval, 0: leave as is, 1: increase */
1038 update_local_clock(peer_t *p)
1041 long old_tmx_offset;
1043 double offset = p->filter_offset;
1044 double recv_time = p->lastpkt_recv_time;
1046 #if !USING_KERNEL_PLL_LOOP
1049 double since_last_update;
1050 double etemp, dtemp;
1052 abs_offset = fabs(offset);
1054 /* If the offset is too large, give up and go home */
1055 if (abs_offset > PANIC_THRESHOLD) {
1056 bb_error_msg_and_die("offset %f far too big, exiting", offset);
1059 /* If this is an old update, for instance as the result
1060 * of a system peer change, avoid it. We never use
1061 * an old sample or the same sample twice.
1063 if (recv_time <= G.last_update_recv_time) {
1064 VERB3 bb_error_msg("same or older datapoint: %f >= %f, not using it",
1065 G.last_update_recv_time, recv_time);
1066 return 0; /* "leave poll interval as is" */
1069 /* Clock state machine transition function. This is where the
1070 * action is and defines how the system reacts to large time
1071 * and frequency errors.
1073 since_last_update = recv_time - G.reftime;
1074 #if !USING_KERNEL_PLL_LOOP
1077 if (G.discipline_state == STATE_FREQ) {
1078 /* Ignore updates until the stepout threshold */
1079 if (since_last_update < WATCH_THRESHOLD) {
1080 VERB3 bb_error_msg("measuring drift, datapoint ignored, %f sec remains",
1081 WATCH_THRESHOLD - since_last_update);
1082 return 0; /* "leave poll interval as is" */
1084 #if !USING_KERNEL_PLL_LOOP
1085 freq_drift = (offset - G.last_update_offset) / since_last_update;
1089 /* There are two main regimes: when the
1090 * offset exceeds the step threshold and when it does not.
1092 if (abs_offset > STEP_THRESHOLD) {
1093 switch (G.discipline_state) {
1095 /* The first outlyer: ignore it, switch to SPIK state */
1096 VERB3 bb_error_msg("offset:%f - spike detected", offset);
1097 G.discipline_state = STATE_SPIK;
1098 return -1; /* "decrease poll interval" */
1101 /* Ignore succeeding outlyers until either an inlyer
1102 * is found or the stepout threshold is exceeded.
1104 if (since_last_update < WATCH_THRESHOLD) {
1105 VERB3 bb_error_msg("spike detected, datapoint ignored, %f sec remains",
1106 WATCH_THRESHOLD - since_last_update);
1107 return -1; /* "decrease poll interval" */
1109 /* fall through: we need to step */
1112 /* Step the time and clamp down the poll interval.
1114 * In NSET state an initial frequency correction is
1115 * not available, usually because the frequency file has
1116 * not yet been written. Since the time is outside the
1117 * capture range, the clock is stepped. The frequency
1118 * will be set directly following the stepout interval.
1120 * In FSET state the initial frequency has been set
1121 * from the frequency file. Since the time is outside
1122 * the capture range, the clock is stepped immediately,
1123 * rather than after the stepout interval. Guys get
1124 * nervous if it takes 17 minutes to set the clock for
1127 * In SPIK state the stepout threshold has expired and
1128 * the phase is still above the step threshold. Note
1129 * that a single spike greater than the step threshold
1130 * is always suppressed, even at the longer poll
1133 VERB3 bb_error_msg("stepping time by %f; poll_exp=MINPOLL", offset);
1135 if (option_mask32 & OPT_q) {
1136 /* We were only asked to set time once. Done. */
1140 G.polladj_count = 0;
1141 G.poll_exp = MINPOLL;
1142 G.stratum = MAXSTRAT;
1143 if (G.discipline_state == STATE_NSET) {
1144 set_new_values(STATE_FREQ, /*offset:*/ 0, recv_time);
1145 return 1; /* "ok to increase poll interval" */
1147 set_new_values(STATE_SYNC, /*offset:*/ 0, recv_time);
1149 } else { /* abs_offset <= STEP_THRESHOLD */
1151 if (G.poll_exp < MINPOLL && G.initial_poll_complete) {
1152 VERB3 bb_error_msg("small offset:%f, disabling burst mode", offset);
1153 G.polladj_count = 0;
1154 G.poll_exp = MINPOLL;
1157 /* Compute the clock jitter as the RMS of exponentially
1158 * weighted offset differences. Used by the poll adjust code.
1160 etemp = SQUARE(G.discipline_jitter);
1161 dtemp = SQUARE(MAXD(fabs(offset - G.last_update_offset), G_precision_sec));
1162 G.discipline_jitter = SQRT(etemp + (dtemp - etemp) / AVG);
1163 VERB3 bb_error_msg("discipline jitter=%f", G.discipline_jitter);
1165 switch (G.discipline_state) {
1167 if (option_mask32 & OPT_q) {
1168 /* We were only asked to set time once.
1169 * The clock is precise enough, no need to step.
1173 /* This is the first update received and the frequency
1174 * has not been initialized. The first thing to do
1175 * is directly measure the oscillator frequency.
1177 set_new_values(STATE_FREQ, offset, recv_time);
1178 VERB3 bb_error_msg("transitioning to FREQ, datapoint ignored");
1179 return 0; /* "leave poll interval as is" */
1181 #if 0 /* this is dead code for now */
1183 /* This is the first update and the frequency
1184 * has been initialized. Adjust the phase, but
1185 * don't adjust the frequency until the next update.
1187 set_new_values(STATE_SYNC, offset, recv_time);
1188 /* freq_drift remains 0 */
1193 /* since_last_update >= WATCH_THRESHOLD, we waited enough.
1194 * Correct the phase and frequency and switch to SYNC state.
1195 * freq_drift was already estimated (see code above)
1197 set_new_values(STATE_SYNC, offset, recv_time);
1201 #if !USING_KERNEL_PLL_LOOP
1202 /* Compute freq_drift due to PLL and FLL contributions.
1204 * The FLL and PLL frequency gain constants
1205 * depend on the poll interval and Allan
1206 * intercept. The FLL is not used below one-half
1207 * the Allan intercept. Above that the loop gain
1208 * increases in steps to 1 / AVG.
1210 if ((1 << G.poll_exp) > ALLAN / 2) {
1211 etemp = FLL - G.poll_exp;
1214 freq_drift += (offset - G.last_update_offset) / (MAXD(since_last_update, ALLAN) * etemp);
1216 /* For the PLL the integration interval
1217 * (numerator) is the minimum of the update
1218 * interval and poll interval. This allows
1219 * oversampling, but not undersampling.
1221 etemp = MIND(since_last_update, (1 << G.poll_exp));
1222 dtemp = (4 * PLL) << G.poll_exp;
1223 freq_drift += offset * etemp / SQUARE(dtemp);
1225 set_new_values(STATE_SYNC, offset, recv_time);
1228 G.stratum = p->lastpkt_stratum + 1;
1231 G.reftime = G.cur_time;
1232 G.ntp_status = p->lastpkt_status;
1233 G.refid = p->lastpkt_refid;
1234 G.rootdelay = p->lastpkt_rootdelay + p->lastpkt_delay;
1235 dtemp = p->filter_jitter; // SQRT(SQUARE(p->filter_jitter) + SQUARE(s.jitter));
1236 dtemp += MAXD(p->filter_dispersion + FREQ_TOLERANCE * (G.cur_time - p->lastpkt_recv_time) + abs_offset, MINDISP);
1237 G.rootdisp = p->lastpkt_rootdisp + dtemp;
1238 VERB3 bb_error_msg("updating leap/refid/reftime/rootdisp from peer %s", p->p_dotted);
1240 /* We are in STATE_SYNC now, but did not do adjtimex yet.
1241 * (Any other state does not reach this, they all return earlier)
1242 * By this time, freq_drift and G.last_update_offset are set
1243 * to values suitable for adjtimex.
1245 #if !USING_KERNEL_PLL_LOOP
1246 /* Calculate the new frequency drift and frequency stability (wander).
1247 * Compute the clock wander as the RMS of exponentially weighted
1248 * frequency differences. This is not used directly, but can,
1249 * along with the jitter, be a highly useful monitoring and
1252 dtemp = G.discipline_freq_drift + freq_drift;
1253 G.discipline_freq_drift = MAXD(MIND(MAXDRIFT, dtemp), -MAXDRIFT);
1254 etemp = SQUARE(G.discipline_wander);
1255 dtemp = SQUARE(dtemp);
1256 G.discipline_wander = SQRT(etemp + (dtemp - etemp) / AVG);
1258 VERB3 bb_error_msg("discipline freq_drift=%.9f(int:%ld corr:%e) wander=%f",
1259 G.discipline_freq_drift,
1260 (long)(G.discipline_freq_drift * 65536e6),
1262 G.discipline_wander);
1265 memset(&tmx, 0, sizeof(tmx));
1266 if (adjtimex(&tmx) < 0)
1267 bb_perror_msg_and_die("adjtimex");
1268 VERB3 bb_error_msg("p adjtimex freq:%ld offset:%ld constant:%ld status:0x%x",
1269 tmx.freq, tmx.offset, tmx.constant, tmx.status);
1273 if (!G.adjtimex_was_done) {
1274 G.adjtimex_was_done = 1;
1275 /* When we use adjtimex for the very first time,
1276 * we need to ADD to pre-existing tmx.offset - it may be !0
1278 memset(&tmx, 0, sizeof(tmx));
1279 if (adjtimex(&tmx) < 0)
1280 bb_perror_msg_and_die("adjtimex");
1281 old_tmx_offset = tmx.offset;
1283 memset(&tmx, 0, sizeof(tmx));
1285 //doesn't work, offset remains 0 (!) in kernel:
1286 //ntpd: set adjtimex freq:1786097 tmx.offset:77487
1287 //ntpd: prev adjtimex freq:1786097 tmx.offset:0
1288 //ntpd: cur adjtimex freq:1786097 tmx.offset:0
1289 tmx.modes = ADJ_FREQUENCY | ADJ_OFFSET;
1290 /* 65536 is one ppm */
1291 tmx.freq = G.discipline_freq_drift * 65536e6;
1292 tmx.offset = G.last_update_offset * 1000000; /* usec */
1294 tmx.modes = ADJ_OFFSET | ADJ_STATUS | ADJ_TIMECONST;// | ADJ_MAXERROR | ADJ_ESTERROR;
1295 tmx.offset = (G.last_update_offset * 1000000) /* usec */
1296 /* + (G.last_update_offset < 0 ? -0.5 : 0.5) - too small to bother */
1297 + old_tmx_offset; /* almost always 0 */
1298 tmx.status = STA_PLL;
1299 if (G.ntp_status & LI_PLUSSEC)
1300 tmx.status |= STA_INS;
1301 if (G.ntp_status & LI_MINUSSEC)
1302 tmx.status |= STA_DEL;
1303 tmx.constant = G.poll_exp - 4;
1304 //tmx.esterror = (u_int32)(clock_jitter * 1e6);
1305 //tmx.maxerror = (u_int32)((sys_rootdelay / 2 + sys_rootdisp) * 1e6);
1306 rc = adjtimex(&tmx);
1308 bb_perror_msg_and_die("adjtimex");
1309 /* NB: here kernel returns constant == G.poll_exp, not == G.poll_exp - 4.
1310 * Not sure why. Perhaps it is normal.
1312 VERB3 bb_error_msg("adjtimex:%d freq:%ld offset:%ld constant:%ld status:0x%x",
1313 rc, tmx.freq, tmx.offset, tmx.constant, tmx.status);
1316 /* always gives the same output as above msg */
1317 memset(&tmx, 0, sizeof(tmx));
1318 if (adjtimex(&tmx) < 0)
1319 bb_perror_msg_and_die("adjtimex");
1320 VERB3 bb_error_msg("c adjtimex freq:%ld offset:%ld constant:%ld status:0x%x",
1321 tmx.freq, tmx.offset, tmx.constant, tmx.status);
1324 if (G.kernel_freq_drift != tmx.freq / 65536) {
1325 G.kernel_freq_drift = tmx.freq / 65536;
1326 VERB2 bb_error_msg("kernel clock drift: %ld ppm", G.kernel_freq_drift);
1329 return 1; /* "ok to increase poll interval" */
1334 * We've got a new reply packet from a peer, process it
1338 retry_interval(void)
1340 /* Local problem, want to retry soon */
1341 unsigned interval, r;
1342 interval = RETRY_INTERVAL;
1344 interval += r % (unsigned)(RETRY_INTERVAL / 4);
1345 VERB3 bb_error_msg("chose retry interval:%u", interval);
1349 poll_interval(int exponent)
1351 unsigned interval, r;
1352 exponent = G.poll_exp + exponent;
1355 interval = 1 << exponent;
1357 interval += ((r & (interval-1)) >> 4) + ((r >> 8) & 1); /* + 1/16 of interval, max */
1358 VERB3 bb_error_msg("chose poll interval:%u (poll_exp:%d exp:%d)", interval, G.poll_exp, exponent);
1361 static NOINLINE void
1362 recv_and_process_peer_pkt(peer_t *p)
1367 double T1, T2, T3, T4;
1369 datapoint_t *datapoint;
1372 /* We can recvfrom here and check from.IP, but some multihomed
1373 * ntp servers reply from their *other IP*.
1374 * TODO: maybe we should check at least what we can: from.port == 123?
1376 size = recv(p->p_fd, &msg, sizeof(msg), MSG_DONTWAIT);
1378 bb_perror_msg("recv(%s) error", p->p_dotted);
1379 if (errno == EHOSTUNREACH || errno == EHOSTDOWN
1380 || errno == ENETUNREACH || errno == ENETDOWN
1381 || errno == ECONNREFUSED || errno == EADDRNOTAVAIL
1384 //TODO: always do this?
1385 interval = retry_interval();
1386 goto set_next_and_close_sock;
1391 if (size != NTP_MSGSIZE_NOAUTH && size != NTP_MSGSIZE) {
1392 bb_error_msg("malformed packet received from %s", p->p_dotted);
1396 if (msg.m_orgtime.int_partl != p->p_xmt_msg.m_xmttime.int_partl
1397 || msg.m_orgtime.fractionl != p->p_xmt_msg.m_xmttime.fractionl
1402 if ((msg.m_status & LI_ALARM) == LI_ALARM
1403 || msg.m_stratum == 0
1404 || msg.m_stratum > NTP_MAXSTRATUM
1406 // TODO: stratum 0 responses may have commands in 32-bit m_refid field:
1407 // "DENY", "RSTR" - peer does not like us at all
1408 // "RATE" - peer is overloaded, reduce polling freq
1409 interval = poll_interval(0);
1410 bb_error_msg("reply from %s: not synced, next query in %us", p->p_dotted, interval);
1411 goto set_next_and_close_sock;
1414 // /* Verify valid root distance */
1415 // if (msg.m_rootdelay / 2 + msg.m_rootdisp >= MAXDISP || p->lastpkt_reftime > msg.m_xmt)
1416 // return; /* invalid header values */
1418 p->lastpkt_status = msg.m_status;
1419 p->lastpkt_stratum = msg.m_stratum;
1420 p->lastpkt_rootdelay = sfp_to_d(msg.m_rootdelay);
1421 p->lastpkt_rootdisp = sfp_to_d(msg.m_rootdisp);
1422 p->lastpkt_refid = msg.m_refid;
1425 * From RFC 2030 (with a correction to the delay math):
1427 * Timestamp Name ID When Generated
1428 * ------------------------------------------------------------
1429 * Originate Timestamp T1 time request sent by client
1430 * Receive Timestamp T2 time request received by server
1431 * Transmit Timestamp T3 time reply sent by server
1432 * Destination Timestamp T4 time reply received by client
1434 * The roundtrip delay and local clock offset are defined as
1436 * delay = (T4 - T1) - (T3 - T2); offset = ((T2 - T1) + (T3 - T4)) / 2
1439 T2 = lfp_to_d(msg.m_rectime);
1440 T3 = lfp_to_d(msg.m_xmttime);
1443 p->lastpkt_recv_time = T4;
1445 VERB5 bb_error_msg("%s->lastpkt_recv_time=%f", p->p_dotted, p->lastpkt_recv_time);
1446 p->datapoint_idx = p->reachable_bits ? (p->datapoint_idx + 1) % NUM_DATAPOINTS : 0;
1447 datapoint = &p->filter_datapoint[p->datapoint_idx];
1448 datapoint->d_recv_time = T4;
1449 datapoint->d_offset = ((T2 - T1) + (T3 - T4)) / 2;
1450 /* The delay calculation is a special case. In cases where the
1451 * server and client clocks are running at different rates and
1452 * with very fast networks, the delay can appear negative. In
1453 * order to avoid violating the Principle of Least Astonishment,
1454 * the delay is clamped not less than the system precision.
1456 p->lastpkt_delay = (T4 - T1) - (T3 - T2);
1457 if (p->lastpkt_delay < G_precision_sec)
1458 p->lastpkt_delay = G_precision_sec;
1459 datapoint->d_dispersion = LOG2D(msg.m_precision_exp) + G_precision_sec;
1460 if (!p->reachable_bits) {
1461 /* 1st datapoint ever - replicate offset in every element */
1463 for (i = 1; i < NUM_DATAPOINTS; i++) {
1464 p->filter_datapoint[i].d_offset = datapoint->d_offset;
1468 p->reachable_bits |= 1;
1470 bb_error_msg("reply from %s: reach 0x%02x offset %f delay %f status 0x%02x strat %d refid 0x%08x rootdelay %f",
1473 datapoint->d_offset,
1478 p->lastpkt_rootdelay
1479 /* not shown: m_ppoll, m_precision_exp, m_rootdisp,
1480 * m_reftime, m_orgtime, m_rectime, m_xmttime
1485 /* Muck with statictics and update the clock */
1486 filter_datapoints(p);
1487 q = select_and_cluster();
1491 if (!(option_mask32 & OPT_w))
1492 rc = update_local_clock(q);
1496 /* Adjust the poll interval by comparing the current offset
1497 * with the clock jitter. If the offset is less than
1498 * the clock jitter times a constant, then the averaging interval
1499 * is increased, otherwise it is decreased. A bit of hysteresis
1500 * helps calm the dance. Works best using burst mode.
1503 bb_error_msg("offset:%f POLLADJ_GATE*discipline_jitter:%f poll:%s",
1504 q->filter_offset, POLLADJ_GATE * G.discipline_jitter,
1505 fabs(q->filter_offset) < POLLADJ_GATE * G.discipline_jitter
1509 if (rc > 0 && fabs(q->filter_offset) < POLLADJ_GATE * G.discipline_jitter) {
1510 /* was += G.poll_exp but it is a bit
1511 * too optimistic for my taste at high poll_exp's */
1512 G.polladj_count += MINPOLL;
1513 if (G.polladj_count > POLLADJ_LIMIT) {
1514 G.polladj_count = 0;
1515 if (G.poll_exp < MAXPOLL) {
1517 VERB3 bb_error_msg("polladj: discipline_jitter:%f ++poll_exp=%d",
1518 G.discipline_jitter, G.poll_exp);
1521 VERB3 bb_error_msg("polladj: incr:%d", G.polladj_count);
1524 G.polladj_count -= G.poll_exp * 2;
1525 if (G.polladj_count < -POLLADJ_LIMIT) {
1526 G.polladj_count = 0;
1527 if (G.poll_exp > MINPOLL) {
1531 /* Correct p->next_action_time in each peer
1532 * which waits for sending, so that they send earlier.
1533 * Old pp->next_action_time are on the order
1534 * of t + (1 << old_poll_exp) + small_random,
1535 * we simply need to subtract ~half of that.
1537 for (item = G.ntp_peers; item != NULL; item = item->link) {
1538 peer_t *pp = (peer_t *) item->data;
1540 pp->next_action_time -= (1 << G.poll_exp);
1542 VERB3 bb_error_msg("polladj: discipline_jitter:%f --poll_exp=%d",
1543 G.discipline_jitter, G.poll_exp);
1546 VERB3 bb_error_msg("polladj: decr:%d", G.polladj_count);
1551 /* Decide when to send new query for this peer */
1552 interval = poll_interval(0);
1554 set_next_and_close_sock:
1555 set_next(p, interval);
1556 /* We do not expect any more packets from this peer for now.
1557 * Closing the socket informs kernel about it.
1558 * We open a new socket when we send a new query.
1566 #if ENABLE_FEATURE_NTPD_SERVER
1567 static NOINLINE void
1568 recv_and_process_client_pkt(void /*int fd*/)
1572 len_and_sockaddr *to;
1573 struct sockaddr *from;
1575 uint8_t query_status;
1576 l_fixedpt_t query_xmttime;
1578 to = get_sock_lsa(G.listen_fd);
1579 from = xzalloc(to->len);
1581 size = recv_from_to(G.listen_fd, &msg, sizeof(msg), MSG_DONTWAIT, from, &to->u.sa, to->len);
1582 if (size != NTP_MSGSIZE_NOAUTH && size != NTP_MSGSIZE) {
1585 if (errno == EAGAIN)
1587 bb_perror_msg_and_die("recv");
1589 addr = xmalloc_sockaddr2dotted_noport(from);
1590 bb_error_msg("malformed packet received from %s: size %u", addr, (int)size);
1595 query_status = msg.m_status;
1596 query_xmttime = msg.m_xmttime;
1598 /* Build a reply packet */
1599 memset(&msg, 0, sizeof(msg));
1600 msg.m_status = G.stratum < MAXSTRAT ? G.ntp_status : LI_ALARM;
1601 msg.m_status |= (query_status & VERSION_MASK);
1602 msg.m_status |= ((query_status & MODE_MASK) == MODE_CLIENT) ?
1603 MODE_SERVER : MODE_SYM_PAS;
1604 msg.m_stratum = G.stratum;
1605 msg.m_ppoll = G.poll_exp;
1606 msg.m_precision_exp = G_precision_exp;
1607 /* this time was obtained between poll() and recv() */
1608 msg.m_rectime = d_to_lfp(G.cur_time);
1609 msg.m_xmttime = d_to_lfp(gettime1900d()); /* this instant */
1610 msg.m_reftime = d_to_lfp(G.reftime);
1611 msg.m_orgtime = query_xmttime;
1612 msg.m_rootdelay = d_to_sfp(G.rootdelay);
1613 //simple code does not do this, fix simple code!
1614 msg.m_rootdisp = d_to_sfp(G.rootdisp);
1615 version = (query_status & VERSION_MASK); /* ... >> VERSION_SHIFT - done below instead */
1616 msg.m_refid = G.refid; // (version > (3 << VERSION_SHIFT)) ? G.refid : G.refid3;
1618 /* We reply from the local address packet was sent to,
1619 * this makes to/from look swapped here: */
1620 do_sendto(G.listen_fd,
1621 /*from:*/ &to->u.sa, /*to:*/ from, /*addrlen:*/ to->len,
1630 /* Upstream ntpd's options:
1632 * -4 Force DNS resolution of host names to the IPv4 namespace.
1633 * -6 Force DNS resolution of host names to the IPv6 namespace.
1634 * -a Require cryptographic authentication for broadcast client,
1635 * multicast client and symmetric passive associations.
1636 * This is the default.
1637 * -A Do not require cryptographic authentication for broadcast client,
1638 * multicast client and symmetric passive associations.
1639 * This is almost never a good idea.
1640 * -b Enable the client to synchronize to broadcast servers.
1642 * Specify the name and path of the configuration file,
1643 * default /etc/ntp.conf
1644 * -d Specify debugging mode. This option may occur more than once,
1645 * with each occurrence indicating greater detail of display.
1647 * Specify debugging level directly.
1649 * Specify the name and path of the frequency file.
1650 * This is the same operation as the "driftfile FILE"
1651 * configuration command.
1652 * -g Normally, ntpd exits with a message to the system log
1653 * if the offset exceeds the panic threshold, which is 1000 s
1654 * by default. This option allows the time to be set to any value
1655 * without restriction; however, this can happen only once.
1656 * If the threshold is exceeded after that, ntpd will exit
1657 * with a message to the system log. This option can be used
1658 * with the -q and -x options. See the tinker command for other options.
1660 * Chroot the server to the directory jaildir. This option also implies
1661 * that the server attempts to drop root privileges at startup
1662 * (otherwise, chroot gives very little additional security).
1663 * You may need to also specify a -u option.
1665 * Specify the name and path of the symmetric key file,
1666 * default /etc/ntp/keys. This is the same operation
1667 * as the "keys FILE" configuration command.
1669 * Specify the name and path of the log file. The default
1670 * is the system log file. This is the same operation as
1671 * the "logfile FILE" configuration command.
1672 * -L Do not listen to virtual IPs. The default is to listen.
1674 * -N To the extent permitted by the operating system,
1675 * run the ntpd at the highest priority.
1677 * Specify the name and path of the file used to record the ntpd
1678 * process ID. This is the same operation as the "pidfile FILE"
1679 * configuration command.
1681 * To the extent permitted by the operating system,
1682 * run the ntpd at the specified priority.
1683 * -q Exit the ntpd just after the first time the clock is set.
1684 * This behavior mimics that of the ntpdate program, which is
1685 * to be retired. The -g and -x options can be used with this option.
1686 * Note: The kernel time discipline is disabled with this option.
1688 * Specify the default propagation delay from the broadcast/multicast
1689 * server to this client. This is necessary only if the delay
1690 * cannot be computed automatically by the protocol.
1692 * Specify the directory path for files created by the statistics
1693 * facility. This is the same operation as the "statsdir DIR"
1694 * configuration command.
1696 * Add a key number to the trusted key list. This option can occur
1699 * Specify a user, and optionally a group, to switch to.
1702 * Add a system variable listed by default.
1703 * -x Normally, the time is slewed if the offset is less than the step
1704 * threshold, which is 128 ms by default, and stepped if above
1705 * the threshold. This option sets the threshold to 600 s, which is
1706 * well within the accuracy window to set the clock manually.
1707 * Note: since the slew rate of typical Unix kernels is limited
1708 * to 0.5 ms/s, each second of adjustment requires an amortization
1709 * interval of 2000 s. Thus, an adjustment as much as 600 s
1710 * will take almost 14 days to complete. This option can be used
1711 * with the -g and -q options. See the tinker command for other options.
1712 * Note: The kernel time discipline is disabled with this option.
1715 /* By doing init in a separate function we decrease stack usage
1718 static NOINLINE void ntp_init(char **argv)
1726 bb_error_msg_and_die(bb_msg_you_must_be_root);
1728 /* Set some globals */
1729 G.stratum = MAXSTRAT;
1731 G.poll_exp = BURSTPOLL; /* speeds up initial sync */
1732 G.reftime = G.last_update_recv_time = gettime1900d(); /* sets G.cur_time too */
1736 opt_complementary = "dd:p::"; /* d: counter, p: list */
1737 opts = getopt32(argv,
1739 "wp:"IF_FEATURE_NTPD_SERVER("l") /* NOT compat */
1741 "46aAbgL", /* compat, ignored */
1742 &peers, &G.verbose);
1743 if (!(opts & (OPT_p|OPT_l)))
1745 // if (opts & OPT_x) /* disable stepping, only slew is allowed */
1746 // G.time_was_stepped = 1;
1748 add_peers(llist_pop(&peers));
1749 if (!(opts & OPT_n)) {
1750 bb_daemonize_or_rexec(DAEMON_DEVNULL_STDIO, argv);
1751 logmode = LOGMODE_NONE;
1753 #if ENABLE_FEATURE_NTPD_SERVER
1756 G.listen_fd = create_and_bind_dgram_or_die(NULL, 123);
1757 socket_want_pktinfo(G.listen_fd);
1758 setsockopt(G.listen_fd, IPPROTO_IP, IP_TOS, &const_IPTOS_LOWDELAY, sizeof(const_IPTOS_LOWDELAY));
1761 /* I hesitate to set -20 prio. -15 should be high enough for timekeeping */
1763 setpriority(PRIO_PROCESS, 0, -15);
1765 bb_signals((1 << SIGTERM) | (1 << SIGINT), record_signo);
1766 bb_signals((1 << SIGPIPE) | (1 << SIGHUP), SIG_IGN);
1769 int ntpd_main(int argc UNUSED_PARAM, char **argv) MAIN_EXTERNALLY_VISIBLE;
1770 int ntpd_main(int argc UNUSED_PARAM, char **argv)
1778 memset(&G, 0, sizeof(G));
1779 SET_PTR_TO_GLOBALS(&G);
1783 /* If ENABLE_FEATURE_NTPD_SERVER, + 1 for listen_fd: */
1784 cnt = G.peer_cnt + ENABLE_FEATURE_NTPD_SERVER;
1785 idx2peer = xzalloc(sizeof(idx2peer[0]) * cnt);
1786 pfd = xzalloc(sizeof(pfd[0]) * cnt);
1788 /* Countdown: we never sync before we sent 5 packets to each peer
1789 * NB: if some peer is not responding, we may end up sending
1790 * fewer packets to it and more to other peers.
1791 * NB2: sync usually happens using 5-1=4 packets, since last reply
1792 * does not come back instantaneously.
1794 cnt = G.peer_cnt * 5;
1796 while (!bb_got_signal) {
1802 /* Nothing between here and poll() blocks for any significant time */
1804 nextaction = G.cur_time + 3600;
1807 #if ENABLE_FEATURE_NTPD_SERVER
1808 if (G.listen_fd != -1) {
1809 pfd[0].fd = G.listen_fd;
1810 pfd[0].events = POLLIN;
1814 /* Pass over peer list, send requests, time out on receives */
1815 for (item = G.ntp_peers; item != NULL; item = item->link) {
1816 peer_t *p = (peer_t *) item->data;
1818 if (p->next_action_time <= G.cur_time) {
1819 if (p->p_fd == -1) {
1820 /* Time to send new req */
1822 G.initial_poll_complete = 1;
1824 send_query_to_peer(p);
1826 /* Timed out waiting for reply */
1829 timeout = poll_interval(-2); /* -2: try a bit sooner */
1830 bb_error_msg("timed out waiting for %s, reach 0x%02x, next query in %us",
1831 p->p_dotted, p->reachable_bits, timeout);
1832 set_next(p, timeout);
1836 if (p->next_action_time < nextaction)
1837 nextaction = p->next_action_time;
1840 /* Wait for reply from this peer */
1841 pfd[i].fd = p->p_fd;
1842 pfd[i].events = POLLIN;
1848 timeout = nextaction - G.cur_time;
1851 timeout++; /* (nextaction - G.cur_time) rounds down, compensating */
1853 /* Here we may block */
1854 VERB2 bb_error_msg("poll %us, sockets:%u", timeout, i);
1855 nfds = poll(pfd, i, timeout * 1000);
1856 gettime1900d(); /* sets G.cur_time */
1860 /* Process any received packets */
1862 #if ENABLE_FEATURE_NTPD_SERVER
1863 if (G.listen_fd != -1) {
1864 if (pfd[0].revents /* & (POLLIN|POLLERR)*/) {
1866 recv_and_process_client_pkt(/*G.listen_fd*/);
1867 gettime1900d(); /* sets G.cur_time */
1872 for (; nfds != 0 && j < i; j++) {
1873 if (pfd[j].revents /* & (POLLIN|POLLERR)*/) {
1875 recv_and_process_peer_pkt(idx2peer[j]);
1876 gettime1900d(); /* sets G.cur_time */
1879 } /* while (!bb_got_signal) */
1881 kill_myself_with_sig(bb_got_signal);
1889 /*** openntpd-4.6 uses only adjtime, not adjtimex ***/
1891 /*** ntp-4.2.6/ntpd/ntp_loopfilter.c - adjtimex usage ***/
1895 direct_freq(double fp_offset)
1900 * If the kernel is enabled, we need the residual offset to
1901 * calculate the frequency correction.
1903 if (pll_control && kern_enable) {
1904 memset(&ntv, 0, sizeof(ntv));
1907 clock_offset = ntv.offset / 1e9;
1908 #else /* STA_NANO */
1909 clock_offset = ntv.offset / 1e6;
1910 #endif /* STA_NANO */
1911 drift_comp = FREQTOD(ntv.freq);
1913 #endif /* KERNEL_PLL */
1914 set_freq((fp_offset - clock_offset) / (current_time - clock_epoch) + drift_comp);
1920 set_freq(double freq) /* frequency update */
1928 * If the kernel is enabled, update the kernel frequency.
1930 if (pll_control && kern_enable) {
1931 memset(&ntv, 0, sizeof(ntv));
1932 ntv.modes = MOD_FREQUENCY;
1933 ntv.freq = DTOFREQ(drift_comp);
1935 snprintf(tbuf, sizeof(tbuf), "kernel %.3f PPM", drift_comp * 1e6);
1936 report_event(EVNT_FSET, NULL, tbuf);
1938 snprintf(tbuf, sizeof(tbuf), "ntpd %.3f PPM", drift_comp * 1e6);
1939 report_event(EVNT_FSET, NULL, tbuf);
1941 #else /* KERNEL_PLL */
1942 snprintf(tbuf, sizeof(tbuf), "ntpd %.3f PPM", drift_comp * 1e6);
1943 report_event(EVNT_FSET, NULL, tbuf);
1944 #endif /* KERNEL_PLL */
1953 * This code segment works when clock adjustments are made using
1954 * precision time kernel support and the ntp_adjtime() system
1955 * call. This support is available in Solaris 2.6 and later,
1956 * Digital Unix 4.0 and later, FreeBSD, Linux and specially
1957 * modified kernels for HP-UX 9 and Ultrix 4. In the case of the
1958 * DECstation 5000/240 and Alpha AXP, additional kernel
1959 * modifications provide a true microsecond clock and nanosecond
1960 * clock, respectively.
1962 * Important note: The kernel discipline is used only if the
1963 * step threshold is less than 0.5 s, as anything higher can
1964 * lead to overflow problems. This might occur if some misguided
1965 * lad set the step threshold to something ridiculous.
1967 if (pll_control && kern_enable) {
1969 #define MOD_BITS (MOD_OFFSET | MOD_MAXERROR | MOD_ESTERROR | MOD_STATUS | MOD_TIMECONST)
1972 * We initialize the structure for the ntp_adjtime()
1973 * system call. We have to convert everything to
1974 * microseconds or nanoseconds first. Do not update the
1975 * system variables if the ext_enable flag is set. In
1976 * this case, the external clock driver will update the
1977 * variables, which will be read later by the local
1978 * clock driver. Afterwards, remember the time and
1979 * frequency offsets for jitter and stability values and
1980 * to update the frequency file.
1982 memset(&ntv, 0, sizeof(ntv));
1984 ntv.modes = MOD_STATUS;
1987 ntv.modes = MOD_BITS | MOD_NANO;
1988 #else /* STA_NANO */
1989 ntv.modes = MOD_BITS;
1990 #endif /* STA_NANO */
1991 if (clock_offset < 0)
1996 ntv.offset = (int32)(clock_offset * 1e9 + dtemp);
1997 ntv.constant = sys_poll;
1998 #else /* STA_NANO */
1999 ntv.offset = (int32)(clock_offset * 1e6 + dtemp);
2000 ntv.constant = sys_poll - 4;
2001 #endif /* STA_NANO */
2002 ntv.esterror = (u_int32)(clock_jitter * 1e6);
2003 ntv.maxerror = (u_int32)((sys_rootdelay / 2 + sys_rootdisp) * 1e6);
2004 ntv.status = STA_PLL;
2007 * Enable/disable the PPS if requested.
2010 if (!(pll_status & STA_PPSTIME))
2011 report_event(EVNT_KERN,
2012 NULL, "PPS enabled");
2013 ntv.status |= STA_PPSTIME | STA_PPSFREQ;
2015 if (pll_status & STA_PPSTIME)
2016 report_event(EVNT_KERN,
2017 NULL, "PPS disabled");
2018 ntv.status &= ~(STA_PPSTIME |
2021 if (sys_leap == LEAP_ADDSECOND)
2022 ntv.status |= STA_INS;
2023 else if (sys_leap == LEAP_DELSECOND)
2024 ntv.status |= STA_DEL;
2028 * Pass the stuff to the kernel. If it squeals, turn off
2029 * the pps. In any case, fetch the kernel offset,
2030 * frequency and jitter.
2032 if (ntp_adjtime(&ntv) == TIME_ERROR) {
2033 if (!(ntv.status & STA_PPSSIGNAL))
2034 report_event(EVNT_KERN, NULL,
2037 pll_status = ntv.status;
2039 clock_offset = ntv.offset / 1e9;
2040 #else /* STA_NANO */
2041 clock_offset = ntv.offset / 1e6;
2042 #endif /* STA_NANO */
2043 clock_frequency = FREQTOD(ntv.freq);
2046 * If the kernel PPS is lit, monitor its performance.
2048 if (ntv.status & STA_PPSTIME) {
2050 clock_jitter = ntv.jitter / 1e9;
2051 #else /* STA_NANO */
2052 clock_jitter = ntv.jitter / 1e6;
2053 #endif /* STA_NANO */
2056 #if defined(STA_NANO) && NTP_API == 4
2058 * If the TAI changes, update the kernel TAI.
2060 if (loop_tai != sys_tai) {
2062 ntv.modes = MOD_TAI;
2063 ntv.constant = sys_tai;
2066 #endif /* STA_NANO */
2068 #endif /* KERNEL_PLL */