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: */
196 OPT_l = (1 << 5) * ENABLE_FEATURE_NTPD_SERVER,
201 /* total round trip delay to currently selected reference clock */
203 /* reference timestamp: time when the system clock was last set or corrected */
205 /* total dispersion to currently selected reference clock */
208 #if ENABLE_FEATURE_NTPD_SERVER
213 /* refid: 32-bit code identifying the particular server or reference clock
214 * in stratum 0 packets this is a four-character ASCII string,
215 * called the kiss code, used for debugging and monitoring
216 * in stratum 1 packets this is a four-character ASCII string
217 * assigned to the reference clock by IANA. Example: "GPS "
218 * in stratum 2+ packets, it's IPv4 address or 4 first bytes of MD5 hash of IPv6
222 /* precision is defined as the larger of the resolution and time to
223 * read the clock, in log2 units. For instance, the precision of a
224 * mains-frequency clock incrementing at 60 Hz is 16 ms, even when the
225 * system clock hardware representation is to the nanosecond.
227 * Delays, jitters of various kinds are clamper down to precision.
229 * If precision_sec is too large, discipline_jitter gets clamped to it
230 * and if offset is much smaller than discipline_jitter, poll interval
231 * grows even though we really can benefit from staying at smaller one,
232 * collecting non-lagged datapoits and correcting the offset.
233 * (Lagged datapoits exist when poll_exp is large but we still have
234 * systematic offset error - the time distance between datapoints
235 * is significat and older datapoints have smaller offsets.
236 * This makes our offset estimation a bit smaller than reality)
237 * Due to this effect, setting G_precision_sec close to
238 * STEP_THRESHOLD isn't such a good idea - offsets may grow
239 * too big and we will step. I observed it with -6.
241 * OTOH, setting precision too small would result in futile attempts
242 * to syncronize to the unachievable precision.
244 * -6 is 1/64 sec, -7 is 1/128 sec and so on.
246 #define G_precision_exp -8
247 #define G_precision_sec (1.0 / (1 << (- G_precision_exp)))
249 /* Bool. After set to 1, never goes back to 0: */
250 smallint adjtimex_was_done;
251 smallint initial_poll_complete;
253 uint8_t discipline_state; // doc calls it c.state
254 uint8_t poll_exp; // s.poll
255 int polladj_count; // c.count
256 long kernel_freq_drift;
257 double last_update_offset; // c.last
258 double last_update_recv_time; // s.t
259 double discipline_jitter; // c.jitter
260 //TODO: add s.jitter - grep for it here and see clock_combine() in doc
261 #define USING_KERNEL_PLL_LOOP 1
262 #if !USING_KERNEL_PLL_LOOP
263 double discipline_freq_drift; // c.freq
264 //TODO: conditionally calculate wander? it's used only for logging
265 double discipline_wander; // c.wander
268 #define G (*ptr_to_globals)
270 static const int const_IPTOS_LOWDELAY = IPTOS_LOWDELAY;
273 #define VERB1 if (MAX_VERBOSE && G.verbose)
274 #define VERB2 if (MAX_VERBOSE >= 2 && G.verbose >= 2)
275 #define VERB3 if (MAX_VERBOSE >= 3 && G.verbose >= 3)
276 #define VERB4 if (MAX_VERBOSE >= 4 && G.verbose >= 4)
277 #define VERB5 if (MAX_VERBOSE >= 5 && G.verbose >= 5)
280 static double LOG2D(int a)
283 return 1.0 / (1UL << -a);
286 static ALWAYS_INLINE double SQUARE(double x)
290 static ALWAYS_INLINE double MAXD(double a, double b)
296 static ALWAYS_INLINE double MIND(double a, double b)
302 #define SQRT(x) (sqrt(x))
308 gettimeofday(&tv, NULL); /* never fails */
309 G.cur_time = tv.tv_sec + (1.0e-6 * tv.tv_usec) + OFFSET_1900_1970;
314 d_to_tv(double d, struct timeval *tv)
316 tv->tv_sec = (long)d;
317 tv->tv_usec = (d - tv->tv_sec) * 1000000;
321 lfp_to_d(l_fixedpt_t lfp)
324 lfp.int_partl = ntohl(lfp.int_partl);
325 lfp.fractionl = ntohl(lfp.fractionl);
326 ret = (double)lfp.int_partl + ((double)lfp.fractionl / UINT_MAX);
330 sfp_to_d(s_fixedpt_t sfp)
333 sfp.int_parts = ntohs(sfp.int_parts);
334 sfp.fractions = ntohs(sfp.fractions);
335 ret = (double)sfp.int_parts + ((double)sfp.fractions / USHRT_MAX);
338 #if ENABLE_FEATURE_NTPD_SERVER
343 lfp.int_partl = (uint32_t)d;
344 lfp.fractionl = (uint32_t)((d - lfp.int_partl) * UINT_MAX);
345 lfp.int_partl = htonl(lfp.int_partl);
346 lfp.fractionl = htonl(lfp.fractionl);
353 sfp.int_parts = (uint16_t)d;
354 sfp.fractions = (uint16_t)((d - sfp.int_parts) * USHRT_MAX);
355 sfp.int_parts = htons(sfp.int_parts);
356 sfp.fractions = htons(sfp.fractions);
362 dispersion(const datapoint_t *dp)
364 return dp->d_dispersion + FREQ_TOLERANCE * (G.cur_time - dp->d_recv_time);
368 root_distance(peer_t *p)
370 /* The root synchronization distance is the maximum error due to
371 * all causes of the local clock relative to the primary server.
372 * It is defined as half the total delay plus total dispersion
375 return MAXD(MINDISP, p->lastpkt_rootdelay + p->lastpkt_delay) / 2
376 + p->lastpkt_rootdisp
377 + p->filter_dispersion
378 + FREQ_TOLERANCE * (G.cur_time - p->lastpkt_recv_time)
383 set_next(peer_t *p, unsigned t)
385 p->next_action_time = G.cur_time + t;
389 * Peer clock filter and its helpers
392 filter_datapoints(peer_t *p)
396 double minoff, maxoff, wavg, sum, w;
397 double x = x; /* for compiler */
398 double oldest_off = oldest_off;
399 double oldest_age = oldest_age;
400 double newest_off = newest_off;
401 double newest_age = newest_age;
403 minoff = maxoff = p->filter_datapoint[0].d_offset;
404 for (i = 1; i < NUM_DATAPOINTS; i++) {
405 if (minoff > p->filter_datapoint[i].d_offset)
406 minoff = p->filter_datapoint[i].d_offset;
407 if (maxoff < p->filter_datapoint[i].d_offset)
408 maxoff = p->filter_datapoint[i].d_offset;
411 idx = p->datapoint_idx; /* most recent datapoint */
413 * Drop two outliers and take weighted average of the rest:
414 * most_recent/2 + older1/4 + older2/8 ... + older5/32 + older6/32
415 * we use older6/32, not older6/64 since sum of weights should be 1:
416 * 1/2 + 1/4 + 1/8 + 1/16 + 1/32 + 1/32 = 1
422 * filter_dispersion = \ -------------
429 for (i = 0; i < NUM_DATAPOINTS; i++) {
431 bb_error_msg("datapoint[%d]: off:%f disp:%f(%f) age:%f%s",
433 p->filter_datapoint[idx].d_offset,
434 p->filter_datapoint[idx].d_dispersion, dispersion(&p->filter_datapoint[idx]),
435 G.cur_time - p->filter_datapoint[idx].d_recv_time,
436 (minoff == p->filter_datapoint[idx].d_offset || maxoff == p->filter_datapoint[idx].d_offset)
437 ? " (outlier by offset)" : ""
441 sum += dispersion(&p->filter_datapoint[idx]) / (2 << i);
443 if (minoff == p->filter_datapoint[idx].d_offset) {
444 minoff -= 1; /* so that we don't match it ever again */
446 if (maxoff == p->filter_datapoint[idx].d_offset) {
449 oldest_off = p->filter_datapoint[idx].d_offset;
450 oldest_age = G.cur_time - p->filter_datapoint[idx].d_recv_time;
453 newest_off = oldest_off;
454 newest_age = oldest_age;
461 idx = (idx - 1) & (NUM_DATAPOINTS - 1);
463 p->filter_dispersion = sum;
464 wavg += x; /* add another older6/64 to form older6/32 */
465 /* Fix systematic underestimation with large poll intervals.
466 * Imagine that we still have a bit of uncorrected drift,
467 * and poll interval is big (say, 100 sec). Offsets form a progression:
468 * 0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 - 0.7 is most recent.
469 * The algorithm above drops 0.0 and 0.7 as outliers,
470 * and then we have this estimation, ~25% off from 0.7:
471 * 0.1/32 + 0.2/32 + 0.3/16 + 0.4/8 + 0.5/4 + 0.6/2 = 0.503125
473 x = oldest_age - newest_age;
475 x = newest_age / x; /* in above example, 100 / (600 - 100) */
476 if (x < 1) { /* paranoia check */
477 x = (newest_off - oldest_off) * x; /* 0.5 * 100/500 = 0.1 */
481 p->filter_offset = wavg;
483 /* +----- -----+ ^ 1/2
487 * filter_jitter = | --- * / (avg-offset_j) |
491 * where n is the number of valid datapoints in the filter (n > 1);
492 * if filter_jitter < precision then filter_jitter = precision
495 for (i = 0; i < NUM_DATAPOINTS; i++) {
496 sum += SQUARE(wavg - p->filter_datapoint[i].d_offset);
498 sum = SQRT(sum / NUM_DATAPOINTS);
499 p->filter_jitter = sum > G_precision_sec ? sum : G_precision_sec;
501 VERB3 bb_error_msg("filter offset:%f(corr:%e) disp:%f jitter:%f",
503 p->filter_dispersion,
509 reset_peer_stats(peer_t *p, double offset)
512 for (i = 0; i < NUM_DATAPOINTS; i++) {
513 if (offset < 16 * STEP_THRESHOLD) {
514 p->filter_datapoint[i].d_recv_time -= offset;
515 if (p->filter_datapoint[i].d_offset != 0) {
516 p->filter_datapoint[i].d_offset -= offset;
519 p->filter_datapoint[i].d_recv_time = G.cur_time;
520 p->filter_datapoint[i].d_offset = 0;
521 p->filter_datapoint[i].d_dispersion = MAXDISP;
524 if (offset < 16 * STEP_THRESHOLD) {
525 p->lastpkt_recv_time -= offset;
527 p->reachable_bits = 0;
528 p->lastpkt_recv_time = G.cur_time;
530 filter_datapoints(p); /* recalc p->filter_xxx */
531 p->next_action_time -= offset;
532 VERB5 bb_error_msg("%s->lastpkt_recv_time=%f", p->p_dotted, p->lastpkt_recv_time);
540 p = xzalloc(sizeof(*p));
541 p->p_lsa = xhost2sockaddr(s, 123);
542 p->p_dotted = xmalloc_sockaddr2dotted_noport(&p->p_lsa->u.sa);
544 p->p_xmt_msg.m_status = MODE_CLIENT | (NTP_VERSION << 3);
545 p->next_action_time = G.cur_time; /* = set_next(p, 0); */
546 reset_peer_stats(p, 16 * STEP_THRESHOLD);
548 llist_add_to(&G.ntp_peers, p);
554 const struct sockaddr *from, const struct sockaddr *to, socklen_t addrlen,
555 msg_t *msg, ssize_t len)
561 ret = sendto(fd, msg, len, MSG_DONTWAIT, to, addrlen);
563 ret = send_to_from(fd, msg, len, MSG_DONTWAIT, to, from, addrlen);
566 bb_perror_msg("send failed");
573 send_query_to_peer(peer_t *p)
575 /* Why do we need to bind()?
576 * See what happens when we don't bind:
578 * socket(PF_INET, SOCK_DGRAM, IPPROTO_IP) = 3
579 * setsockopt(3, SOL_IP, IP_TOS, [16], 4) = 0
580 * gettimeofday({1259071266, 327885}, NULL) = 0
581 * sendto(3, "xxx", 48, MSG_DONTWAIT, {sa_family=AF_INET, sin_port=htons(123), sin_addr=inet_addr("10.34.32.125")}, 16) = 48
582 * ^^^ we sent it from some source port picked by kernel.
583 * time(NULL) = 1259071266
584 * write(2, "ntpd: entering poll 15 secs\n", 28) = 28
585 * poll([{fd=3, events=POLLIN}], 1, 15000) = 1 ([{fd=3, revents=POLLIN}])
586 * recv(3, "yyy", 68, MSG_DONTWAIT) = 48
587 * ^^^ this recv will receive packets to any local port!
589 * Uncomment this and use strace to see it in action:
591 #define PROBE_LOCAL_ADDR /* { len_and_sockaddr lsa; lsa.len = LSA_SIZEOF_SA; getsockname(p->query.fd, &lsa.u.sa, &lsa.len); } */
595 len_and_sockaddr *local_lsa;
597 family = p->p_lsa->u.sa.sa_family;
598 p->p_fd = fd = xsocket_type(&local_lsa, family, SOCK_DGRAM);
599 /* local_lsa has "null" address and port 0 now.
600 * bind() ensures we have a *particular port* selected by kernel
601 * and remembered in p->p_fd, thus later recv(p->p_fd)
602 * receives only packets sent to this port.
605 xbind(fd, &local_lsa->u.sa, local_lsa->len);
607 #if ENABLE_FEATURE_IPV6
608 if (family == AF_INET)
610 setsockopt(fd, IPPROTO_IP, IP_TOS, &const_IPTOS_LOWDELAY, sizeof(const_IPTOS_LOWDELAY));
615 * Send out a random 64-bit number as our transmit time. The NTP
616 * server will copy said number into the originate field on the
617 * response that it sends us. This is totally legal per the SNTP spec.
619 * The impact of this is two fold: we no longer send out the current
620 * system time for the world to see (which may aid an attacker), and
621 * it gives us a (not very secure) way of knowing that we're not
622 * getting spoofed by an attacker that can't capture our traffic
623 * but can spoof packets from the NTP server we're communicating with.
625 * Save the real transmit timestamp locally.
627 p->p_xmt_msg.m_xmttime.int_partl = random();
628 p->p_xmt_msg.m_xmttime.fractionl = random();
629 p->p_xmttime = gettime1900d();
631 if (do_sendto(p->p_fd, /*from:*/ NULL, /*to:*/ &p->p_lsa->u.sa, /*addrlen:*/ p->p_lsa->len,
632 &p->p_xmt_msg, NTP_MSGSIZE_NOAUTH) == -1
636 set_next(p, RETRY_INTERVAL);
640 p->reachable_bits <<= 1;
641 VERB1 bb_error_msg("sent query to %s", p->p_dotted);
642 set_next(p, RESPONSE_INTERVAL);
647 step_time(double offset)
655 gettimeofday(&tv, NULL); /* never fails */
656 dtime = offset + tv.tv_sec;
657 dtime += 1.0e-6 * tv.tv_usec;
660 if (settimeofday(&tv, NULL) == -1)
661 bb_perror_msg_and_die("settimeofday");
664 strftime(buf, sizeof(buf), "%a %b %e %H:%M:%S %Z %Y", localtime(&tval));
666 bb_error_msg("setting clock to %s (offset %fs)", buf, offset);
668 /* Correct various fields which contain time-relative values: */
670 /* p->lastpkt_recv_time, p->next_action_time and such: */
671 for (item = G.ntp_peers; item != NULL; item = item->link) {
672 peer_t *pp = (peer_t *) item->data;
673 reset_peer_stats(pp, offset);
676 G.cur_time -= offset;
677 G.last_update_recv_time -= offset;
682 * Selection and clustering, and their helpers
690 compare_point_edge(const void *aa, const void *bb)
692 const point_t *a = aa;
693 const point_t *b = bb;
694 if (a->edge < b->edge) {
697 return (a->edge > b->edge);
704 compare_survivor_metric(const void *aa, const void *bb)
706 const survivor_t *a = aa;
707 const survivor_t *b = bb;
708 if (a->metric < b->metric) {
711 return (a->metric > b->metric);
714 fit(peer_t *p, double rd)
716 if ((p->reachable_bits & (p->reachable_bits-1)) == 0) {
717 /* One or zero bits in reachable_bits */
718 VERB3 bb_error_msg("peer %s unfit for selection: unreachable", p->p_dotted);
721 #if 0 /* we filter out such packets earlier */
722 if ((p->lastpkt_status & LI_ALARM) == LI_ALARM
723 || p->lastpkt_stratum >= MAXSTRAT
725 VERB3 bb_error_msg("peer %s unfit for selection: bad status/stratum", p->p_dotted);
729 /* rd is root_distance(p) */
730 if (rd > MAXDIST + FREQ_TOLERANCE * (1 << G.poll_exp)) {
731 VERB3 bb_error_msg("peer %s unfit for selection: root distance too high", p->p_dotted);
735 // /* Do we have a loop? */
736 // if (p->refid == p->dstaddr || p->refid == s.refid)
741 select_and_cluster(void)
745 int size = 3 * G.peer_cnt;
746 /* for selection algorithm */
748 unsigned num_points, num_candidates;
750 unsigned num_falsetickers;
751 /* for cluster algorithm */
752 survivor_t survivor[size];
753 unsigned num_survivors;
759 if (G.initial_poll_complete) while (item != NULL) {
760 peer_t *p = (peer_t *) item->data;
761 double rd = root_distance(p);
762 double offset = p->filter_offset;
769 VERB4 bb_error_msg("interval: [%f %f %f] %s",
775 point[num_points].p = p;
776 point[num_points].type = -1;
777 point[num_points].edge = offset - rd;
779 point[num_points].p = p;
780 point[num_points].type = 0;
781 point[num_points].edge = offset;
783 point[num_points].p = p;
784 point[num_points].type = 1;
785 point[num_points].edge = offset + rd;
789 num_candidates = num_points / 3;
790 if (num_candidates == 0) {
791 VERB3 bb_error_msg("no valid datapoints, no peer selected");
794 //TODO: sorting does not seem to be done in reference code
795 qsort(point, num_points, sizeof(point[0]), compare_point_edge);
797 /* Start with the assumption that there are no falsetickers.
798 * Attempt to find a nonempty intersection interval containing
799 * the midpoints of all truechimers.
800 * If a nonempty interval cannot be found, increase the number
801 * of assumed falsetickers by one and try again.
802 * If a nonempty interval is found and the number of falsetickers
803 * is less than the number of truechimers, a majority has been found
804 * and the midpoint of each truechimer represents
805 * the candidates available to the cluster algorithm.
807 num_falsetickers = 0;
810 unsigned num_midpoints = 0;
815 for (i = 0; i < num_points; i++) {
817 * if (point[i].type == -1) c++;
818 * if (point[i].type == 1) c--;
819 * and it's simpler to do it this way:
822 if (c >= num_candidates - num_falsetickers) {
823 /* If it was c++ and it got big enough... */
827 if (point[i].type == 0)
831 for (i = num_points-1; i >= 0; i--) {
833 if (c >= num_candidates - num_falsetickers) {
834 high = point[i].edge;
837 if (point[i].type == 0)
840 /* If the number of midpoints is greater than the number
841 * of allowed falsetickers, the intersection contains at
842 * least one truechimer with no midpoint - bad.
843 * Also, interval should be nonempty.
845 if (num_midpoints <= num_falsetickers && low < high)
848 if (num_falsetickers * 2 >= num_candidates) {
849 VERB3 bb_error_msg("too many falsetickers:%d (candidates:%d), no peer selected",
850 num_falsetickers, num_candidates);
854 VERB3 bb_error_msg("selected interval: [%f, %f]; candidates:%d falsetickers:%d",
855 low, high, num_candidates, num_falsetickers);
859 /* Construct a list of survivors (p, metric)
860 * from the chime list, where metric is dominated
861 * first by stratum and then by root distance.
862 * All other things being equal, this is the order of preference.
865 for (i = 0; i < num_points; i++) {
868 if (point[i].edge < low || point[i].edge > high)
871 survivor[num_survivors].p = p;
872 //TODO: save root_distance in point_t and reuse here?
873 survivor[num_survivors].metric = MAXDIST * p->lastpkt_stratum + root_distance(p);
874 VERB4 bb_error_msg("survivor[%d] metric:%f peer:%s",
875 num_survivors, survivor[num_survivors].metric, p->p_dotted);
878 /* There must be at least MIN_SELECTED survivors to satisfy the
879 * correctness assertions. Ordinarily, the Byzantine criteria
880 * require four survivors, but for the demonstration here, one
883 if (num_survivors < MIN_SELECTED) {
884 VERB3 bb_error_msg("num_survivors %d < %d, no peer selected",
885 num_survivors, MIN_SELECTED);
889 //looks like this is ONLY used by the fact that later we pick survivor[0].
890 //we can avoid sorting then, just find the minimum once!
891 qsort(survivor, num_survivors, sizeof(survivor[0]), compare_survivor_metric);
893 /* For each association p in turn, calculate the selection
894 * jitter p->sjitter as the square root of the sum of squares
895 * (p->offset - q->offset) over all q associations. The idea is
896 * to repeatedly discard the survivor with maximum selection
897 * jitter until a termination condition is met.
900 unsigned max_idx = max_idx;
901 double max_selection_jitter = max_selection_jitter;
902 double min_jitter = min_jitter;
904 if (num_survivors <= MIN_CLUSTERED) {
905 bb_error_msg("num_survivors %d <= %d, not discarding more",
906 num_survivors, MIN_CLUSTERED);
910 /* To make sure a few survivors are left
911 * for the clustering algorithm to chew on,
912 * we stop if the number of survivors
913 * is less than or equal to MIN_CLUSTERED (3).
915 for (i = 0; i < num_survivors; i++) {
916 double selection_jitter_sq;
917 peer_t *p = survivor[i].p;
919 if (i == 0 || p->filter_jitter < min_jitter)
920 min_jitter = p->filter_jitter;
922 selection_jitter_sq = 0;
923 for (j = 0; j < num_survivors; j++) {
924 peer_t *q = survivor[j].p;
925 selection_jitter_sq += SQUARE(p->filter_offset - q->filter_offset);
927 if (i == 0 || selection_jitter_sq > max_selection_jitter) {
928 max_selection_jitter = selection_jitter_sq;
931 VERB5 bb_error_msg("survivor %d selection_jitter^2:%f",
932 i, selection_jitter_sq);
934 max_selection_jitter = SQRT(max_selection_jitter / num_survivors);
935 VERB4 bb_error_msg("max_selection_jitter (at %d):%f min_jitter:%f",
936 max_idx, max_selection_jitter, min_jitter);
938 /* If the maximum selection jitter is less than the
939 * minimum peer jitter, then tossing out more survivors
940 * will not lower the minimum peer jitter, so we might
943 if (max_selection_jitter < min_jitter) {
944 VERB3 bb_error_msg("max_selection_jitter:%f < min_jitter:%f, num_survivors:%d, not discarding more",
945 max_selection_jitter, min_jitter, num_survivors);
949 /* Delete survivor[max_idx] from the list
950 * and go around again.
952 VERB5 bb_error_msg("dropping survivor %d", max_idx);
954 while (max_idx < num_survivors) {
955 survivor[max_idx] = survivor[max_idx + 1];
960 /* Pick the best clock. If the old system peer is on the list
961 * and at the same stratum as the first survivor on the list,
962 * then don't do a clock hop. Otherwise, select the first
963 * survivor on the list as the new system peer.
965 //TODO - see clock_combine()
966 VERB3 bb_error_msg("selected peer %s filter_offset:%f age:%f",
967 survivor[0].p->p_dotted,
968 survivor[0].p->filter_offset,
969 G.cur_time - survivor[0].p->lastpkt_recv_time
971 return survivor[0].p;
976 * Local clock discipline and its helpers
979 set_new_values(int disc_state, double offset, double recv_time)
981 /* Enter new state and set state variables. Note we use the time
982 * of the last clock filter sample, which must be earlier than
985 VERB3 bb_error_msg("disc_state=%d last update offset=%f recv_time=%f",
986 disc_state, offset, recv_time);
987 G.discipline_state = disc_state;
988 G.last_update_offset = offset;
989 G.last_update_recv_time = recv_time;
991 /* Clock state definitions */
992 #define STATE_NSET 0 /* initial state, "nothing is set" */
993 #define STATE_FSET 1 /* frequency set from file */
994 #define STATE_SPIK 2 /* spike detected */
995 #define STATE_FREQ 3 /* initial frequency */
996 #define STATE_SYNC 4 /* clock synchronized (normal operation) */
997 /* Return: -1: decrease poll interval, 0: leave as is, 1: increase */
999 update_local_clock(peer_t *p)
1002 long old_tmx_offset;
1004 double offset = p->filter_offset;
1005 double recv_time = p->lastpkt_recv_time;
1007 #if !USING_KERNEL_PLL_LOOP
1010 double since_last_update;
1011 double etemp, dtemp;
1013 abs_offset = fabs(offset);
1015 /* If the offset is too large, give up and go home */
1016 if (abs_offset > PANIC_THRESHOLD) {
1017 bb_error_msg_and_die("offset %f far too big, exiting", offset);
1020 /* If this is an old update, for instance as the result
1021 * of a system peer change, avoid it. We never use
1022 * an old sample or the same sample twice.
1024 if (recv_time <= G.last_update_recv_time) {
1025 VERB3 bb_error_msg("same or older datapoint: %f >= %f, not using it",
1026 G.last_update_recv_time, recv_time);
1027 return 0; /* "leave poll interval as is" */
1030 /* Clock state machine transition function. This is where the
1031 * action is and defines how the system reacts to large time
1032 * and frequency errors.
1034 since_last_update = recv_time - G.reftime;
1035 #if !USING_KERNEL_PLL_LOOP
1038 if (G.discipline_state == STATE_FREQ) {
1039 /* Ignore updates until the stepout threshold */
1040 if (since_last_update < WATCH_THRESHOLD) {
1041 VERB3 bb_error_msg("measuring drift, datapoint ignored, %f sec remains",
1042 WATCH_THRESHOLD - since_last_update);
1043 return 0; /* "leave poll interval as is" */
1045 #if !USING_KERNEL_PLL_LOOP
1046 freq_drift = (offset - G.last_update_offset) / since_last_update;
1050 /* There are two main regimes: when the
1051 * offset exceeds the step threshold and when it does not.
1053 if (abs_offset > STEP_THRESHOLD) {
1054 switch (G.discipline_state) {
1056 /* The first outlyer: ignore it, switch to SPIK state */
1057 VERB3 bb_error_msg("offset:%f - spike detected", offset);
1058 G.discipline_state = STATE_SPIK;
1059 return -1; /* "decrease poll interval" */
1062 /* Ignore succeeding outlyers until either an inlyer
1063 * is found or the stepout threshold is exceeded.
1065 if (since_last_update < WATCH_THRESHOLD) {
1066 VERB3 bb_error_msg("spike detected, datapoint ignored, %f sec remains",
1067 WATCH_THRESHOLD - since_last_update);
1068 return -1; /* "decrease poll interval" */
1070 /* fall through: we need to step */
1073 /* Step the time and clamp down the poll interval.
1075 * In NSET state an initial frequency correction is
1076 * not available, usually because the frequency file has
1077 * not yet been written. Since the time is outside the
1078 * capture range, the clock is stepped. The frequency
1079 * will be set directly following the stepout interval.
1081 * In FSET state the initial frequency has been set
1082 * from the frequency file. Since the time is outside
1083 * the capture range, the clock is stepped immediately,
1084 * rather than after the stepout interval. Guys get
1085 * nervous if it takes 17 minutes to set the clock for
1088 * In SPIK state the stepout threshold has expired and
1089 * the phase is still above the step threshold. Note
1090 * that a single spike greater than the step threshold
1091 * is always suppressed, even at the longer poll
1094 VERB3 bb_error_msg("stepping time by %f; poll_exp=MINPOLL", offset);
1096 if (option_mask32 & OPT_q) {
1097 /* We were only asked to set time once. Done. */
1101 G.polladj_count = 0;
1102 G.poll_exp = MINPOLL;
1103 G.stratum = MAXSTRAT;
1104 if (G.discipline_state == STATE_NSET) {
1105 set_new_values(STATE_FREQ, /*offset:*/ 0, recv_time);
1106 return 1; /* "ok to increase poll interval" */
1108 set_new_values(STATE_SYNC, /*offset:*/ 0, recv_time);
1110 } else { /* abs_offset <= STEP_THRESHOLD */
1112 if (G.poll_exp < MINPOLL && G.initial_poll_complete) {
1113 VERB3 bb_error_msg("small offset:%f, disabling burst mode", offset);
1114 G.polladj_count = 0;
1115 G.poll_exp = MINPOLL;
1118 /* Compute the clock jitter as the RMS of exponentially
1119 * weighted offset differences. Used by the poll adjust code.
1121 etemp = SQUARE(G.discipline_jitter);
1122 dtemp = SQUARE(MAXD(fabs(offset - G.last_update_offset), G_precision_sec));
1123 G.discipline_jitter = SQRT(etemp + (dtemp - etemp) / AVG);
1124 VERB3 bb_error_msg("discipline jitter=%f", G.discipline_jitter);
1126 switch (G.discipline_state) {
1128 if (option_mask32 & OPT_q) {
1129 /* We were only asked to set time once.
1130 * The clock is precise enough, no need to step.
1134 /* This is the first update received and the frequency
1135 * has not been initialized. The first thing to do
1136 * is directly measure the oscillator frequency.
1138 set_new_values(STATE_FREQ, offset, recv_time);
1139 VERB3 bb_error_msg("transitioning to FREQ, datapoint ignored");
1140 return 0; /* "leave poll interval as is" */
1142 #if 0 /* this is dead code for now */
1144 /* This is the first update and the frequency
1145 * has been initialized. Adjust the phase, but
1146 * don't adjust the frequency until the next update.
1148 set_new_values(STATE_SYNC, offset, recv_time);
1149 /* freq_drift remains 0 */
1154 /* since_last_update >= WATCH_THRESHOLD, we waited enough.
1155 * Correct the phase and frequency and switch to SYNC state.
1156 * freq_drift was already estimated (see code above)
1158 set_new_values(STATE_SYNC, offset, recv_time);
1162 #if !USING_KERNEL_PLL_LOOP
1163 /* Compute freq_drift due to PLL and FLL contributions.
1165 * The FLL and PLL frequency gain constants
1166 * depend on the poll interval and Allan
1167 * intercept. The FLL is not used below one-half
1168 * the Allan intercept. Above that the loop gain
1169 * increases in steps to 1 / AVG.
1171 if ((1 << G.poll_exp) > ALLAN / 2) {
1172 etemp = FLL - G.poll_exp;
1175 freq_drift += (offset - G.last_update_offset) / (MAXD(since_last_update, ALLAN) * etemp);
1177 /* For the PLL the integration interval
1178 * (numerator) is the minimum of the update
1179 * interval and poll interval. This allows
1180 * oversampling, but not undersampling.
1182 etemp = MIND(since_last_update, (1 << G.poll_exp));
1183 dtemp = (4 * PLL) << G.poll_exp;
1184 freq_drift += offset * etemp / SQUARE(dtemp);
1186 set_new_values(STATE_SYNC, offset, recv_time);
1189 G.stratum = p->lastpkt_stratum + 1;
1192 G.reftime = G.cur_time;
1193 G.ntp_status = p->lastpkt_status;
1194 G.refid = p->lastpkt_refid;
1195 G.rootdelay = p->lastpkt_rootdelay + p->lastpkt_delay;
1196 dtemp = p->filter_jitter; // SQRT(SQUARE(p->filter_jitter) + SQUARE(s.jitter));
1197 dtemp += MAXD(p->filter_dispersion + FREQ_TOLERANCE * (G.cur_time - p->lastpkt_recv_time) + abs_offset, MINDISP);
1198 G.rootdisp = p->lastpkt_rootdisp + dtemp;
1199 VERB3 bb_error_msg("updating leap/refid/reftime/rootdisp from peer %s", p->p_dotted);
1201 /* We are in STATE_SYNC now, but did not do adjtimex yet.
1202 * (Any other state does not reach this, they all return earlier)
1203 * By this time, freq_drift and G.last_update_offset are set
1204 * to values suitable for adjtimex.
1206 #if !USING_KERNEL_PLL_LOOP
1207 /* Calculate the new frequency drift and frequency stability (wander).
1208 * Compute the clock wander as the RMS of exponentially weighted
1209 * frequency differences. This is not used directly, but can,
1210 * along with the jitter, be a highly useful monitoring and
1213 dtemp = G.discipline_freq_drift + freq_drift;
1214 G.discipline_freq_drift = MAXD(MIND(MAXDRIFT, dtemp), -MAXDRIFT);
1215 etemp = SQUARE(G.discipline_wander);
1216 dtemp = SQUARE(dtemp);
1217 G.discipline_wander = SQRT(etemp + (dtemp - etemp) / AVG);
1219 VERB3 bb_error_msg("discipline freq_drift=%.9f(int:%ld corr:%e) wander=%f",
1220 G.discipline_freq_drift,
1221 (long)(G.discipline_freq_drift * 65536e6),
1223 G.discipline_wander);
1226 memset(&tmx, 0, sizeof(tmx));
1227 if (adjtimex(&tmx) < 0)
1228 bb_perror_msg_and_die("adjtimex");
1229 VERB3 bb_error_msg("p adjtimex freq:%ld offset:%ld constant:%ld status:0x%x",
1230 tmx.freq, tmx.offset, tmx.constant, tmx.status);
1234 if (!G.adjtimex_was_done) {
1235 G.adjtimex_was_done = 1;
1236 /* When we use adjtimex for the very first time,
1237 * we need to ADD to pre-existing tmx.offset - it may be !0
1239 memset(&tmx, 0, sizeof(tmx));
1240 if (adjtimex(&tmx) < 0)
1241 bb_perror_msg_and_die("adjtimex");
1242 old_tmx_offset = tmx.offset;
1244 memset(&tmx, 0, sizeof(tmx));
1246 //doesn't work, offset remains 0 (!) in kernel:
1247 //ntpd: set adjtimex freq:1786097 tmx.offset:77487
1248 //ntpd: prev adjtimex freq:1786097 tmx.offset:0
1249 //ntpd: cur adjtimex freq:1786097 tmx.offset:0
1250 tmx.modes = ADJ_FREQUENCY | ADJ_OFFSET;
1251 /* 65536 is one ppm */
1252 tmx.freq = G.discipline_freq_drift * 65536e6;
1253 tmx.offset = G.last_update_offset * 1000000; /* usec */
1255 tmx.modes = ADJ_OFFSET | ADJ_STATUS | ADJ_TIMECONST;// | ADJ_MAXERROR | ADJ_ESTERROR;
1256 tmx.offset = (G.last_update_offset * 1000000) /* usec */
1257 /* + (G.last_update_offset < 0 ? -0.5 : 0.5) - too small to bother */
1258 + old_tmx_offset; /* almost always 0 */
1259 tmx.status = STA_PLL;
1260 if (G.ntp_status & LI_PLUSSEC)
1261 tmx.status |= STA_INS;
1262 if (G.ntp_status & LI_MINUSSEC)
1263 tmx.status |= STA_DEL;
1264 tmx.constant = G.poll_exp - 4;
1265 //tmx.esterror = (u_int32)(clock_jitter * 1e6);
1266 //tmx.maxerror = (u_int32)((sys_rootdelay / 2 + sys_rootdisp) * 1e6);
1267 rc = adjtimex(&tmx);
1269 bb_perror_msg_and_die("adjtimex");
1270 /* NB: here kernel returns constant == G.poll_exp, not == G.poll_exp - 4.
1271 * Not sure why. Perhaps it is normal.
1273 VERB3 bb_error_msg("adjtimex:%d freq:%ld offset:%ld constant:%ld status:0x%x",
1274 rc, tmx.freq, tmx.offset, tmx.constant, tmx.status);
1277 /* always gives the same output as above msg */
1278 memset(&tmx, 0, sizeof(tmx));
1279 if (adjtimex(&tmx) < 0)
1280 bb_perror_msg_and_die("adjtimex");
1281 VERB3 bb_error_msg("c adjtimex freq:%ld offset:%ld constant:%ld status:0x%x",
1282 tmx.freq, tmx.offset, tmx.constant, tmx.status);
1285 if (G.kernel_freq_drift != tmx.freq / 65536) {
1286 G.kernel_freq_drift = tmx.freq / 65536;
1287 VERB2 bb_error_msg("kernel clock drift: %ld ppm", G.kernel_freq_drift);
1290 return 1; /* "ok to increase poll interval" */
1295 * We've got a new reply packet from a peer, process it
1299 retry_interval(void)
1301 /* Local problem, want to retry soon */
1302 unsigned interval, r;
1303 interval = RETRY_INTERVAL;
1305 interval += r % (unsigned)(RETRY_INTERVAL / 4);
1306 VERB3 bb_error_msg("chose retry interval:%u", interval);
1310 poll_interval(int exponent)
1312 unsigned interval, r;
1313 exponent = G.poll_exp + exponent;
1316 interval = 1 << exponent;
1318 interval += ((r & (interval-1)) >> 4) + ((r >> 8) & 1); /* + 1/16 of interval, max */
1319 VERB3 bb_error_msg("chose poll interval:%u (poll_exp:%d exp:%d)", interval, G.poll_exp, exponent);
1322 static NOINLINE void
1323 recv_and_process_peer_pkt(peer_t *p)
1328 double T1, T2, T3, T4;
1330 datapoint_t *datapoint;
1333 /* We can recvfrom here and check from.IP, but some multihomed
1334 * ntp servers reply from their *other IP*.
1335 * TODO: maybe we should check at least what we can: from.port == 123?
1337 size = recv(p->p_fd, &msg, sizeof(msg), MSG_DONTWAIT);
1339 bb_perror_msg("recv(%s) error", p->p_dotted);
1340 if (errno == EHOSTUNREACH || errno == EHOSTDOWN
1341 || errno == ENETUNREACH || errno == ENETDOWN
1342 || errno == ECONNREFUSED || errno == EADDRNOTAVAIL
1345 //TODO: always do this?
1346 set_next(p, retry_interval());
1352 if (size != NTP_MSGSIZE_NOAUTH && size != NTP_MSGSIZE) {
1353 bb_error_msg("malformed packet received from %s", p->p_dotted);
1357 if (msg.m_orgtime.int_partl != p->p_xmt_msg.m_xmttime.int_partl
1358 || msg.m_orgtime.fractionl != p->p_xmt_msg.m_xmttime.fractionl
1363 if ((msg.m_status & LI_ALARM) == LI_ALARM
1364 || msg.m_stratum == 0
1365 || msg.m_stratum > NTP_MAXSTRATUM
1367 // TODO: stratum 0 responses may have commands in 32-bit m_refid field:
1368 // "DENY", "RSTR" - peer does not like us at all
1369 // "RATE" - peer is overloaded, reduce polling freq
1370 interval = poll_interval(0);
1371 bb_error_msg("reply from %s: not synced, next query in %us", p->p_dotted, interval);
1375 // /* Verify valid root distance */
1376 // if (msg.m_rootdelay / 2 + msg.m_rootdisp >= MAXDISP || p->lastpkt_reftime > msg.m_xmt)
1377 // return; /* invalid header values */
1379 p->lastpkt_status = msg.m_status;
1380 p->lastpkt_stratum = msg.m_stratum;
1381 p->lastpkt_rootdelay = sfp_to_d(msg.m_rootdelay);
1382 p->lastpkt_rootdisp = sfp_to_d(msg.m_rootdisp);
1383 p->lastpkt_refid = msg.m_refid;
1386 * From RFC 2030 (with a correction to the delay math):
1388 * Timestamp Name ID When Generated
1389 * ------------------------------------------------------------
1390 * Originate Timestamp T1 time request sent by client
1391 * Receive Timestamp T2 time request received by server
1392 * Transmit Timestamp T3 time reply sent by server
1393 * Destination Timestamp T4 time reply received by client
1395 * The roundtrip delay and local clock offset are defined as
1397 * delay = (T4 - T1) - (T3 - T2); offset = ((T2 - T1) + (T3 - T4)) / 2
1400 T2 = lfp_to_d(msg.m_rectime);
1401 T3 = lfp_to_d(msg.m_xmttime);
1404 p->lastpkt_recv_time = T4;
1406 VERB5 bb_error_msg("%s->lastpkt_recv_time=%f", p->p_dotted, p->lastpkt_recv_time);
1407 p->datapoint_idx = p->reachable_bits ? (p->datapoint_idx + 1) % NUM_DATAPOINTS : 0;
1408 datapoint = &p->filter_datapoint[p->datapoint_idx];
1409 datapoint->d_recv_time = T4;
1410 datapoint->d_offset = ((T2 - T1) + (T3 - T4)) / 2;
1411 /* The delay calculation is a special case. In cases where the
1412 * server and client clocks are running at different rates and
1413 * with very fast networks, the delay can appear negative. In
1414 * order to avoid violating the Principle of Least Astonishment,
1415 * the delay is clamped not less than the system precision.
1417 p->lastpkt_delay = (T4 - T1) - (T3 - T2);
1418 if (p->lastpkt_delay < G_precision_sec)
1419 p->lastpkt_delay = G_precision_sec;
1420 datapoint->d_dispersion = LOG2D(msg.m_precision_exp) + G_precision_sec;
1421 if (!p->reachable_bits) {
1422 /* 1st datapoint ever - replicate offset in every element */
1424 for (i = 1; i < NUM_DATAPOINTS; i++) {
1425 p->filter_datapoint[i].d_offset = datapoint->d_offset;
1429 p->reachable_bits |= 1;
1431 bb_error_msg("reply from %s: reach 0x%02x offset %f delay %f",
1434 datapoint->d_offset, p->lastpkt_delay);
1437 /* Muck with statictics and update the clock */
1438 filter_datapoints(p);
1439 q = select_and_cluster();
1442 rc = update_local_clock(q);
1445 /* Adjust the poll interval by comparing the current offset
1446 * with the clock jitter. If the offset is less than
1447 * the clock jitter times a constant, then the averaging interval
1448 * is increased, otherwise it is decreased. A bit of hysteresis
1449 * helps calm the dance. Works best using burst mode.
1452 bb_error_msg("offset:%f POLLADJ_GATE*discipline_jitter:%f poll:%s",
1453 q->filter_offset, POLLADJ_GATE * G.discipline_jitter,
1454 fabs(q->filter_offset) < POLLADJ_GATE * G.discipline_jitter
1458 if (rc > 0 && fabs(q->filter_offset) < POLLADJ_GATE * G.discipline_jitter) {
1459 /* was += G.poll_exp but it is a bit
1460 * too optimistic for my taste at high poll_exp's */
1461 G.polladj_count += MINPOLL;
1462 if (G.polladj_count > POLLADJ_LIMIT) {
1463 G.polladj_count = 0;
1464 if (G.poll_exp < MAXPOLL) {
1466 VERB3 bb_error_msg("polladj: discipline_jitter:%f ++poll_exp=%d",
1467 G.discipline_jitter, G.poll_exp);
1470 VERB3 bb_error_msg("polladj: incr:%d", G.polladj_count);
1473 G.polladj_count -= G.poll_exp * 2;
1474 if (G.polladj_count < -POLLADJ_LIMIT) {
1475 G.polladj_count = 0;
1476 if (G.poll_exp > MINPOLL) {
1480 /* Correct p->next_action_time in each peer
1481 * which waits for sending, so that they send earlier.
1482 * Old pp->next_action_time are on the order
1483 * of t + (1 << old_poll_exp) + small_random,
1484 * we simply need to subtract ~half of that.
1486 for (item = G.ntp_peers; item != NULL; item = item->link) {
1487 peer_t *pp = (peer_t *) item->data;
1489 pp->next_action_time -= (1 << G.poll_exp);
1491 VERB3 bb_error_msg("polladj: discipline_jitter:%f --poll_exp=%d",
1492 G.discipline_jitter, G.poll_exp);
1495 VERB3 bb_error_msg("polladj: decr:%d", G.polladj_count);
1500 /* Decide when to send new query for this peer */
1501 interval = poll_interval(0);
1502 set_next(p, interval);
1505 /* We do not expect any more packets from this peer for now.
1506 * Closing the socket informs kernel about it.
1507 * We open a new socket when we send a new query.
1515 #if ENABLE_FEATURE_NTPD_SERVER
1516 static NOINLINE void
1517 recv_and_process_client_pkt(void /*int fd*/)
1521 len_and_sockaddr *to;
1522 struct sockaddr *from;
1524 uint8_t query_status;
1525 l_fixedpt_t query_xmttime;
1527 to = get_sock_lsa(G.listen_fd);
1528 from = xzalloc(to->len);
1530 size = recv_from_to(G.listen_fd, &msg, sizeof(msg), MSG_DONTWAIT, from, &to->u.sa, to->len);
1531 if (size != NTP_MSGSIZE_NOAUTH && size != NTP_MSGSIZE) {
1534 if (errno == EAGAIN)
1536 bb_perror_msg_and_die("recv");
1538 addr = xmalloc_sockaddr2dotted_noport(from);
1539 bb_error_msg("malformed packet received from %s: size %u", addr, (int)size);
1544 query_status = msg.m_status;
1545 query_xmttime = msg.m_xmttime;
1547 /* Build a reply packet */
1548 memset(&msg, 0, sizeof(msg));
1549 msg.m_status = G.stratum < MAXSTRAT ? G.ntp_status : LI_ALARM;
1550 msg.m_status |= (query_status & VERSION_MASK);
1551 msg.m_status |= ((query_status & MODE_MASK) == MODE_CLIENT) ?
1552 MODE_SERVER : MODE_SYM_PAS;
1553 msg.m_stratum = G.stratum;
1554 msg.m_ppoll = G.poll_exp;
1555 msg.m_precision_exp = G_precision_exp;
1556 /* this time was obtained between poll() and recv() */
1557 msg.m_rectime = d_to_lfp(G.cur_time);
1558 msg.m_xmttime = d_to_lfp(gettime1900d()); /* this instant */
1559 msg.m_reftime = d_to_lfp(G.reftime);
1560 msg.m_orgtime = query_xmttime;
1561 msg.m_rootdelay = d_to_sfp(G.rootdelay);
1562 //simple code does not do this, fix simple code!
1563 msg.m_rootdisp = d_to_sfp(G.rootdisp);
1564 version = (query_status & VERSION_MASK); /* ... >> VERSION_SHIFT - done below instead */
1565 msg.m_refid = G.refid; // (version > (3 << VERSION_SHIFT)) ? G.refid : G.refid3;
1567 /* We reply from the local address packet was sent to,
1568 * this makes to/from look swapped here: */
1569 do_sendto(G.listen_fd,
1570 /*from:*/ &to->u.sa, /*to:*/ from, /*addrlen:*/ to->len,
1579 /* Upstream ntpd's options:
1581 * -4 Force DNS resolution of host names to the IPv4 namespace.
1582 * -6 Force DNS resolution of host names to the IPv6 namespace.
1583 * -a Require cryptographic authentication for broadcast client,
1584 * multicast client and symmetric passive associations.
1585 * This is the default.
1586 * -A Do not require cryptographic authentication for broadcast client,
1587 * multicast client and symmetric passive associations.
1588 * This is almost never a good idea.
1589 * -b Enable the client to synchronize to broadcast servers.
1591 * Specify the name and path of the configuration file,
1592 * default /etc/ntp.conf
1593 * -d Specify debugging mode. This option may occur more than once,
1594 * with each occurrence indicating greater detail of display.
1596 * Specify debugging level directly.
1598 * Specify the name and path of the frequency file.
1599 * This is the same operation as the "driftfile FILE"
1600 * configuration command.
1601 * -g Normally, ntpd exits with a message to the system log
1602 * if the offset exceeds the panic threshold, which is 1000 s
1603 * by default. This option allows the time to be set to any value
1604 * without restriction; however, this can happen only once.
1605 * If the threshold is exceeded after that, ntpd will exit
1606 * with a message to the system log. This option can be used
1607 * with the -q and -x options. See the tinker command for other options.
1609 * Chroot the server to the directory jaildir. This option also implies
1610 * that the server attempts to drop root privileges at startup
1611 * (otherwise, chroot gives very little additional security).
1612 * You may need to also specify a -u option.
1614 * Specify the name and path of the symmetric key file,
1615 * default /etc/ntp/keys. This is the same operation
1616 * as the "keys FILE" configuration command.
1618 * Specify the name and path of the log file. The default
1619 * is the system log file. This is the same operation as
1620 * the "logfile FILE" configuration command.
1621 * -L Do not listen to virtual IPs. The default is to listen.
1623 * -N To the extent permitted by the operating system,
1624 * run the ntpd at the highest priority.
1626 * Specify the name and path of the file used to record the ntpd
1627 * process ID. This is the same operation as the "pidfile FILE"
1628 * configuration command.
1630 * To the extent permitted by the operating system,
1631 * run the ntpd at the specified priority.
1632 * -q Exit the ntpd just after the first time the clock is set.
1633 * This behavior mimics that of the ntpdate program, which is
1634 * to be retired. The -g and -x options can be used with this option.
1635 * Note: The kernel time discipline is disabled with this option.
1637 * Specify the default propagation delay from the broadcast/multicast
1638 * server to this client. This is necessary only if the delay
1639 * cannot be computed automatically by the protocol.
1641 * Specify the directory path for files created by the statistics
1642 * facility. This is the same operation as the "statsdir DIR"
1643 * configuration command.
1645 * Add a key number to the trusted key list. This option can occur
1648 * Specify a user, and optionally a group, to switch to.
1651 * Add a system variable listed by default.
1652 * -x Normally, the time is slewed if the offset is less than the step
1653 * threshold, which is 128 ms by default, and stepped if above
1654 * the threshold. This option sets the threshold to 600 s, which is
1655 * well within the accuracy window to set the clock manually.
1656 * Note: since the slew rate of typical Unix kernels is limited
1657 * to 0.5 ms/s, each second of adjustment requires an amortization
1658 * interval of 2000 s. Thus, an adjustment as much as 600 s
1659 * will take almost 14 days to complete. This option can be used
1660 * with the -g and -q options. See the tinker command for other options.
1661 * Note: The kernel time discipline is disabled with this option.
1664 /* By doing init in a separate function we decrease stack usage
1667 static NOINLINE void ntp_init(char **argv)
1675 bb_error_msg_and_die(bb_msg_you_must_be_root);
1677 /* Set some globals */
1678 G.stratum = MAXSTRAT;
1680 G.poll_exp = BURSTPOLL; /* speeds up initial sync */
1681 G.reftime = G.last_update_recv_time = gettime1900d(); /* sets G.cur_time too */
1685 opt_complementary = "dd:p::"; /* d: counter, p: list */
1686 opts = getopt32(argv,
1688 "p:"IF_FEATURE_NTPD_SERVER("l") /* NOT compat */
1690 "46aAbgL", /* compat, ignored */
1691 &peers, &G.verbose);
1692 if (!(opts & (OPT_p|OPT_l)))
1694 // if (opts & OPT_x) /* disable stepping, only slew is allowed */
1695 // G.time_was_stepped = 1;
1697 add_peers(llist_pop(&peers));
1698 if (!(opts & OPT_n)) {
1699 bb_daemonize_or_rexec(DAEMON_DEVNULL_STDIO, argv);
1700 logmode = LOGMODE_NONE;
1702 #if ENABLE_FEATURE_NTPD_SERVER
1705 G.listen_fd = create_and_bind_dgram_or_die(NULL, 123);
1706 socket_want_pktinfo(G.listen_fd);
1707 setsockopt(G.listen_fd, IPPROTO_IP, IP_TOS, &const_IPTOS_LOWDELAY, sizeof(const_IPTOS_LOWDELAY));
1710 /* I hesitate to set -20 prio. -15 should be high enough for timekeeping */
1712 setpriority(PRIO_PROCESS, 0, -15);
1714 bb_signals((1 << SIGTERM) | (1 << SIGINT), record_signo);
1715 bb_signals((1 << SIGPIPE) | (1 << SIGHUP), SIG_IGN);
1718 int ntpd_main(int argc UNUSED_PARAM, char **argv) MAIN_EXTERNALLY_VISIBLE;
1719 int ntpd_main(int argc UNUSED_PARAM, char **argv)
1727 memset(&G, 0, sizeof(G));
1728 SET_PTR_TO_GLOBALS(&G);
1732 /* If ENABLE_FEATURE_NTPD_SERVER, + 1 for listen_fd: */
1733 cnt = G.peer_cnt + ENABLE_FEATURE_NTPD_SERVER;
1734 idx2peer = xzalloc(sizeof(idx2peer[0]) * cnt);
1735 pfd = xzalloc(sizeof(pfd[0]) * cnt);
1737 /* Countdown: we never sync before we sent 5 packets to each peer
1738 * NB: if some peer is not responding, we may end up sending
1739 * fewer packets to it and more to other peers.
1740 * NB2: sync usually happens using 5-1=4 packets, since last reply
1741 * does not come back instantaneously.
1743 cnt = G.peer_cnt * 5;
1745 while (!bb_got_signal) {
1751 /* Nothing between here and poll() blocks for any significant time */
1753 nextaction = G.cur_time + 3600;
1756 #if ENABLE_FEATURE_NTPD_SERVER
1757 if (G.listen_fd != -1) {
1758 pfd[0].fd = G.listen_fd;
1759 pfd[0].events = POLLIN;
1763 /* Pass over peer list, send requests, time out on receives */
1764 for (item = G.ntp_peers; item != NULL; item = item->link) {
1765 peer_t *p = (peer_t *) item->data;
1767 if (p->next_action_time <= G.cur_time) {
1768 if (p->p_fd == -1) {
1769 /* Time to send new req */
1771 G.initial_poll_complete = 1;
1773 send_query_to_peer(p);
1775 /* Timed out waiting for reply */
1778 timeout = poll_interval(-2); /* -2: try a bit sooner */
1779 bb_error_msg("timed out waiting for %s, reach 0x%02x, next query in %us",
1780 p->p_dotted, p->reachable_bits, timeout);
1781 set_next(p, timeout);
1785 if (p->next_action_time < nextaction)
1786 nextaction = p->next_action_time;
1789 /* Wait for reply from this peer */
1790 pfd[i].fd = p->p_fd;
1791 pfd[i].events = POLLIN;
1797 timeout = nextaction - G.cur_time;
1800 timeout++; /* (nextaction - G.cur_time) rounds down, compensating */
1802 /* Here we may block */
1803 VERB2 bb_error_msg("poll %us, sockets:%u", timeout, i);
1804 nfds = poll(pfd, i, timeout * 1000);
1805 gettime1900d(); /* sets G.cur_time */
1809 /* Process any received packets */
1811 #if ENABLE_FEATURE_NTPD_SERVER
1812 if (G.listen_fd != -1) {
1813 if (pfd[0].revents /* & (POLLIN|POLLERR)*/) {
1815 recv_and_process_client_pkt(/*G.listen_fd*/);
1816 gettime1900d(); /* sets G.cur_time */
1821 for (; nfds != 0 && j < i; j++) {
1822 if (pfd[j].revents /* & (POLLIN|POLLERR)*/) {
1824 recv_and_process_peer_pkt(idx2peer[j]);
1825 gettime1900d(); /* sets G.cur_time */
1828 } /* while (!bb_got_signal) */
1830 kill_myself_with_sig(bb_got_signal);
1838 /*** openntpd-4.6 uses only adjtime, not adjtimex ***/
1840 /*** ntp-4.2.6/ntpd/ntp_loopfilter.c - adjtimex usage ***/
1844 direct_freq(double fp_offset)
1849 * If the kernel is enabled, we need the residual offset to
1850 * calculate the frequency correction.
1852 if (pll_control && kern_enable) {
1853 memset(&ntv, 0, sizeof(ntv));
1856 clock_offset = ntv.offset / 1e9;
1857 #else /* STA_NANO */
1858 clock_offset = ntv.offset / 1e6;
1859 #endif /* STA_NANO */
1860 drift_comp = FREQTOD(ntv.freq);
1862 #endif /* KERNEL_PLL */
1863 set_freq((fp_offset - clock_offset) / (current_time - clock_epoch) + drift_comp);
1869 set_freq(double freq) /* frequency update */
1877 * If the kernel is enabled, update the kernel frequency.
1879 if (pll_control && kern_enable) {
1880 memset(&ntv, 0, sizeof(ntv));
1881 ntv.modes = MOD_FREQUENCY;
1882 ntv.freq = DTOFREQ(drift_comp);
1884 snprintf(tbuf, sizeof(tbuf), "kernel %.3f PPM", drift_comp * 1e6);
1885 report_event(EVNT_FSET, NULL, tbuf);
1887 snprintf(tbuf, sizeof(tbuf), "ntpd %.3f PPM", drift_comp * 1e6);
1888 report_event(EVNT_FSET, NULL, tbuf);
1890 #else /* KERNEL_PLL */
1891 snprintf(tbuf, sizeof(tbuf), "ntpd %.3f PPM", drift_comp * 1e6);
1892 report_event(EVNT_FSET, NULL, tbuf);
1893 #endif /* KERNEL_PLL */
1902 * This code segment works when clock adjustments are made using
1903 * precision time kernel support and the ntp_adjtime() system
1904 * call. This support is available in Solaris 2.6 and later,
1905 * Digital Unix 4.0 and later, FreeBSD, Linux and specially
1906 * modified kernels for HP-UX 9 and Ultrix 4. In the case of the
1907 * DECstation 5000/240 and Alpha AXP, additional kernel
1908 * modifications provide a true microsecond clock and nanosecond
1909 * clock, respectively.
1911 * Important note: The kernel discipline is used only if the
1912 * step threshold is less than 0.5 s, as anything higher can
1913 * lead to overflow problems. This might occur if some misguided
1914 * lad set the step threshold to something ridiculous.
1916 if (pll_control && kern_enable) {
1918 #define MOD_BITS (MOD_OFFSET | MOD_MAXERROR | MOD_ESTERROR | MOD_STATUS | MOD_TIMECONST)
1921 * We initialize the structure for the ntp_adjtime()
1922 * system call. We have to convert everything to
1923 * microseconds or nanoseconds first. Do not update the
1924 * system variables if the ext_enable flag is set. In
1925 * this case, the external clock driver will update the
1926 * variables, which will be read later by the local
1927 * clock driver. Afterwards, remember the time and
1928 * frequency offsets for jitter and stability values and
1929 * to update the frequency file.
1931 memset(&ntv, 0, sizeof(ntv));
1933 ntv.modes = MOD_STATUS;
1936 ntv.modes = MOD_BITS | MOD_NANO;
1937 #else /* STA_NANO */
1938 ntv.modes = MOD_BITS;
1939 #endif /* STA_NANO */
1940 if (clock_offset < 0)
1945 ntv.offset = (int32)(clock_offset * 1e9 + dtemp);
1946 ntv.constant = sys_poll;
1947 #else /* STA_NANO */
1948 ntv.offset = (int32)(clock_offset * 1e6 + dtemp);
1949 ntv.constant = sys_poll - 4;
1950 #endif /* STA_NANO */
1951 ntv.esterror = (u_int32)(clock_jitter * 1e6);
1952 ntv.maxerror = (u_int32)((sys_rootdelay / 2 + sys_rootdisp) * 1e6);
1953 ntv.status = STA_PLL;
1956 * Enable/disable the PPS if requested.
1959 if (!(pll_status & STA_PPSTIME))
1960 report_event(EVNT_KERN,
1961 NULL, "PPS enabled");
1962 ntv.status |= STA_PPSTIME | STA_PPSFREQ;
1964 if (pll_status & STA_PPSTIME)
1965 report_event(EVNT_KERN,
1966 NULL, "PPS disabled");
1967 ntv.status &= ~(STA_PPSTIME |
1970 if (sys_leap == LEAP_ADDSECOND)
1971 ntv.status |= STA_INS;
1972 else if (sys_leap == LEAP_DELSECOND)
1973 ntv.status |= STA_DEL;
1977 * Pass the stuff to the kernel. If it squeals, turn off
1978 * the pps. In any case, fetch the kernel offset,
1979 * frequency and jitter.
1981 if (ntp_adjtime(&ntv) == TIME_ERROR) {
1982 if (!(ntv.status & STA_PPSSIGNAL))
1983 report_event(EVNT_KERN, NULL,
1986 pll_status = ntv.status;
1988 clock_offset = ntv.offset / 1e9;
1989 #else /* STA_NANO */
1990 clock_offset = ntv.offset / 1e6;
1991 #endif /* STA_NANO */
1992 clock_frequency = FREQTOD(ntv.freq);
1995 * If the kernel PPS is lit, monitor its performance.
1997 if (ntv.status & STA_PPSTIME) {
1999 clock_jitter = ntv.jitter / 1e9;
2000 #else /* STA_NANO */
2001 clock_jitter = ntv.jitter / 1e6;
2002 #endif /* STA_NANO */
2005 #if defined(STA_NANO) && NTP_API == 4
2007 * If the TAI changes, update the kernel TAI.
2009 if (loop_tai != sys_tai) {
2011 ntv.modes = MOD_TAI;
2012 ntv.constant = sys_tai;
2015 #endif /* STA_NANO */
2017 #endif /* KERNEL_PLL */