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. It 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 #define RETRY_INTERVAL 5 /* on error, retry in N secs */
43 #define QUERYTIME_MAX 15 /* wait for reply up to N secs */
45 #define FREQ_TOLERANCE 15e-6 /* % frequency tolerance (15 PPM) */
46 #define MINPOLL 4 /* % minimum poll interval (6: 64 s) */
47 #define MAXPOLL 12 /* % maximum poll interval (12: 1.1h, 17: 36.4h) (was 17) */
48 #define MINDISP 0.01 /* % minimum dispersion (s) */
49 #define MAXDISP 16 /* maximum dispersion (s) */
50 #define MAXSTRAT 16 /* maximum stratum (infinity metric) */
51 #define MAXDIST 1 /* % distance threshold (s) */
52 #define MIN_SELECTED 1 /* % minimum intersection survivors */
53 #define MIN_CLUSTERED 3 /* % minimum cluster survivors */
55 #define MAXFREQ 0.000500 /* frequency tolerance (500 PPM) */
57 /* Clock discipline parameters and constants */
58 #define STEP_THRESHOLD 0.128 /* step threshold (s) */
59 #define WATCH_THRESHOLD 150 /* stepout threshold (s). std ntpd uses 900 (11 mins (!)) */
60 /* NB: set WATCH_THRESHOLD to ~60 when debugging to save time) */
61 #define PANIC_THRESHOLD 1000 /* panic threshold (s) */
63 /* Poll-adjust threshold.
64 * When we see that offset is small enough compared to discipline jitter,
65 * we grow a counter: += poll_ext. When it goes over POLLADJ_LIMIT,
66 * we poll_ext++. If offset isn't small, counter -= poll_ext*2,
67 * and when it goes below -POLLADJ_LIMIT, we poll_ext--
69 #define POLLADJ_LIMIT 30
70 /* If offset < POLLADJ_GATE * discipline_jitter, then we can increase
71 * poll interval (we think we can't improve timekeeping
72 * by staying at smaller poll).
74 #define POLLADJ_GATE 4
75 /* Compromise Allan intercept (s). doc uses 1500, std ntpd uses 512 */
79 /* FLL loop gain [why it depends on MAXPOLL??] */
80 #define FLL (MAXPOLL + 1)
81 /* Parameter averaging constant */
84 /* Verbosity control (max level of -dddd options accepted).
85 * max 5 is very talkative (and bloated). 2 is non-bloated,
86 * production level setting.
90 #define VERB1 if (MAX_VERBOSE && G.verbose)
91 #define VERB2 if (MAX_VERBOSE >= 2 && G.verbose >= 2)
92 #define VERB3 if (MAX_VERBOSE >= 3 && G.verbose >= 3)
93 #define VERB4 if (MAX_VERBOSE >= 4 && G.verbose >= 4)
94 #define VERB5 if (MAX_VERBOSE >= 5 && G.verbose >= 5)
102 NTP_MSGSIZE_NOAUTH = 48,
103 NTP_MSGSIZE = (NTP_MSGSIZE_NOAUTH + 4 + NTP_DIGESTSIZE),
106 MODE_MASK = (7 << 0),
107 VERSION_MASK = (7 << 3),
111 /* Leap Second Codes (high order two bits of m_status) */
112 LI_NOWARNING = (0 << 6), /* no warning */
113 LI_PLUSSEC = (1 << 6), /* add a second (61 seconds) */
114 LI_MINUSSEC = (2 << 6), /* minus a second (59 seconds) */
115 LI_ALARM = (3 << 6), /* alarm condition */
118 MODE_RES0 = 0, /* reserved */
119 MODE_SYM_ACT = 1, /* symmetric active */
120 MODE_SYM_PAS = 2, /* symmetric passive */
121 MODE_CLIENT = 3, /* client */
122 MODE_SERVER = 4, /* server */
123 MODE_BROADCAST = 5, /* broadcast */
124 MODE_RES1 = 6, /* reserved for NTP control message */
125 MODE_RES2 = 7, /* reserved for private use */
128 //TODO: better base selection
129 #define OFFSET_1900_1970 2208988800UL /* 1970 - 1900 in seconds */
131 #define NUM_DATAPOINTS 8
144 uint8_t m_status; /* status of local clock and leap info */
146 uint8_t m_ppoll; /* poll value */
147 int8_t m_precision_exp;
148 s_fixedpt_t m_rootdelay;
149 s_fixedpt_t m_rootdisp;
151 l_fixedpt_t m_reftime;
152 l_fixedpt_t m_orgtime;
153 l_fixedpt_t m_rectime;
154 l_fixedpt_t m_xmttime;
156 uint8_t m_digest[NTP_DIGESTSIZE];
166 len_and_sockaddr *p_lsa;
168 /* when to send new query (if p_fd == -1)
169 * or when receive times out (if p_fd >= 0): */
170 time_t next_action_time;
173 uint32_t lastpkt_refid;
174 uint8_t lastpkt_leap;
175 uint8_t lastpkt_stratum;
176 uint8_t p_reachable_bits;
178 double lastpkt_recv_time;
179 double lastpkt_delay;
180 double lastpkt_rootdelay;
181 double lastpkt_rootdisp;
182 /* produced by filter algorithm: */
183 double filter_offset;
184 double filter_dispersion;
185 double filter_jitter;
186 datapoint_t filter_datapoint[NUM_DATAPOINTS];
187 /* last sent packet: */
197 /* Insert new options above this line. */
198 /* Non-compat options: */
200 OPT_l = (1 << 5) * ENABLE_FEATURE_NTPD_SERVER,
204 /* total round trip delay to currently selected reference clock */
206 /* reference timestamp: time when the system clock was last set or corrected */
208 /* total dispersion to currently selected reference clock */
211 #if ENABLE_FEATURE_NTPD_SERVER
216 /* refid: 32-bit code identifying the particular server or reference clock
217 * in stratum 0 packets this is a four-character ASCII string,
218 * called the kiss code, used for debugging and monitoring
219 * in stratum 1 packets this is a four-character ASCII string
220 * assigned to the reference clock by IANA. Example: "GPS "
221 * in stratum 2+ packets, it's IPv4 address or 4 first bytes of MD5 hash of IPv6
225 /* precision is defined as the larger of the resolution and time to
226 * read the clock, in log2 units. For instance, the precision of a
227 * mains-frequency clock incrementing at 60 Hz is 16 ms, even when the
228 * system clock hardware representation is to the nanosecond.
230 * Delays, jitters of various kinds are clamper down to precision.
232 * If precision_sec is too large, discipline_jitter gets clamped to it
233 * and if offset is much smaller than discipline_jitter, poll interval
234 * grows even though we really can benefit from staying at smaller one,
235 * collecting non-lagged datapoits and correcting the offset.
236 * (Lagged datapoits exist when poll_exp is large but we still have
237 * systematic offset error - the time distance between datapoints
238 * is significat and older datapoints have smaller offsets.
239 * This makes our offset estimation a bit smaller than reality)
240 * Due to this effect, setting G_precision_sec close to
241 * STEP_THRESHOLD isn't such a good idea - offsets may grow
242 * too big and we will step. I observed it with -6.
244 * OTOH, setting precision too small would result in futile attempts
245 * to syncronize to the unachievable precision.
247 * -6 is 1/64 sec, -7 is 1/128 sec and so on.
249 #define G_precision_exp -8
250 #define G_precision_sec (1.0 / (1 << (- G_precision_exp)))
252 /* Bool. After set to 1, never goes back to 0: */
254 // uint8_t time_was_stepped;
255 uint8_t adjtimex_was_done;
257 uint8_t discipline_state; // doc calls it c.state
258 uint8_t poll_exp; // s.poll
259 int polladj_count; // c.count
260 double discipline_jitter; // c.jitter
261 double last_update_offset; // c.last
262 double discipline_freq_drift; // c.freq
263 //TODO: conditionally calculate wander? it's used only for logging
264 double discipline_wander; // c.wander
265 double last_update_recv_time; // s.t
266 //TODO: add s.jitter - grep for it here and see clock_combine() in doc
268 #define G (*ptr_to_globals)
270 static const int const_IPTOS_LOWDELAY = IPTOS_LOWDELAY;
273 static double LOG2D(int a)
276 return 1.0 / (1UL << -a);
279 static ALWAYS_INLINE double SQUARE(double x)
283 static ALWAYS_INLINE double MAXD(double a, double b)
289 static ALWAYS_INLINE double MIND(double a, double b)
295 #define SQRT(x) (sqrt(x))
301 gettimeofday(&tv, NULL); /* never fails */
302 return (tv.tv_sec + 1.0e-6 * tv.tv_usec + OFFSET_1900_1970);
306 d_to_tv(double d, struct timeval *tv)
308 tv->tv_sec = (long)d;
309 tv->tv_usec = (d - tv->tv_sec) * 1000000;
313 lfp_to_d(l_fixedpt_t lfp)
316 lfp.int_partl = ntohl(lfp.int_partl);
317 lfp.fractionl = ntohl(lfp.fractionl);
318 ret = (double)lfp.int_partl + ((double)lfp.fractionl / UINT_MAX);
322 sfp_to_d(s_fixedpt_t sfp)
325 sfp.int_parts = ntohs(sfp.int_parts);
326 sfp.fractions = ntohs(sfp.fractions);
327 ret = (double)sfp.int_parts + ((double)sfp.fractions / USHRT_MAX);
330 #if ENABLE_FEATURE_NTPD_SERVER
335 lfp.int_partl = (uint32_t)d;
336 lfp.fractionl = (uint32_t)((d - lfp.int_partl) * UINT_MAX);
337 lfp.int_partl = htonl(lfp.int_partl);
338 lfp.fractionl = htonl(lfp.fractionl);
345 sfp.int_parts = (uint16_t)d;
346 sfp.fractions = (uint16_t)((d - sfp.int_parts) * USHRT_MAX);
347 sfp.int_parts = htons(sfp.int_parts);
348 sfp.fractions = htons(sfp.fractions);
354 dispersion(const datapoint_t *dp, double t)
356 return dp->d_dispersion + FREQ_TOLERANCE * (t - dp->d_recv_time);
360 root_distance(peer_t *p, double t)
362 /* The root synchronization distance is the maximum error due to
363 * all causes of the local clock relative to the primary server.
364 * It is defined as half the total delay plus total dispersion
367 return MAXD(MINDISP, p->lastpkt_rootdelay + p->lastpkt_delay) / 2
368 + p->lastpkt_rootdisp
369 + p->filter_dispersion
370 + FREQ_TOLERANCE * (t - p->lastpkt_recv_time)
375 set_next(peer_t *p, unsigned t)
377 p->next_action_time = time(NULL) + t;
381 * Peer clock filter and its helpers
384 filter_datapoints(peer_t *p, double t)
387 double minoff, maxoff, wavg, sum, w;
390 minoff = maxoff = p->filter_datapoint[0].d_offset;
391 for (i = 1; i < NUM_DATAPOINTS; i++) {
392 if (minoff > p->filter_datapoint[i].d_offset)
393 minoff = p->filter_datapoint[i].d_offset;
394 if (maxoff < p->filter_datapoint[i].d_offset)
395 maxoff = p->filter_datapoint[i].d_offset;
398 idx = p->datapoint_idx; /* most recent datapoint */
400 * Drop two outliers and take weighted average of the rest:
401 * most_recent/2 + older1/4 + older2/8 ... + older5/32 + older6/32
402 * we use older6/32, not older6/64 since sum of weights should be 1:
403 * 1/2 + 1/4 + 1/8 + 1/16 + 1/32 + 1/32 = 1
409 // filter_dispersion = \ -------------
414 for (i = 0; i < NUM_DATAPOINTS; i++) {
416 bb_error_msg("datapoint[%d]: off:%f disp:%f(%f) age:%f%s",
418 p->filter_datapoint[idx].d_offset,
419 p->filter_datapoint[idx].d_dispersion, dispersion(&p->filter_datapoint[idx], t),
420 t - p->filter_datapoint[idx].d_recv_time,
421 (minoff == p->filter_datapoint[idx].d_offset || maxoff == p->filter_datapoint[idx].d_offset)
422 ? " (outlier by offset)" : ""
426 sum += dispersion(&p->filter_datapoint[idx], t) / (2 << i);
428 if (minoff == p->filter_datapoint[idx].d_offset) {
431 if (maxoff == p->filter_datapoint[idx].d_offset) {
434 x = p->filter_datapoint[idx].d_offset * w;
439 idx = (idx - 1) & (NUM_DATAPOINTS - 1);
441 wavg += x; /* add another older6/64 to form older6/32 */
442 p->filter_offset = wavg;
443 p->filter_dispersion = sum;
445 // +----- -----+ ^ 1/2
449 // filter_jitter = --- * | / (avg-offset_j) |
453 // where n is the number of valid datapoints in the filter (n > 1);
454 // if filter_jitter < precision then filter_jitter = precision
456 for (i = 0; i < NUM_DATAPOINTS; i++) {
457 sum += SQUARE(wavg - p->filter_datapoint[i].d_offset);
459 sum = SQRT(sum) / NUM_DATAPOINTS;
460 p->filter_jitter = sum > G_precision_sec ? sum : G_precision_sec;
462 VERB3 bb_error_msg("filter offset:%f disp:%f jitter:%f",
463 p->filter_offset, p->filter_dispersion, p->filter_jitter);
468 reset_peer_stats(peer_t *p, double t, double offset)
471 for (i = 0; i < NUM_DATAPOINTS; i++) {
472 if (offset < 16 * STEP_THRESHOLD) {
473 p->filter_datapoint[i].d_recv_time -= offset;
474 if (p->filter_datapoint[i].d_offset != 0) {
475 p->filter_datapoint[i].d_offset -= offset;
478 p->filter_datapoint[i].d_recv_time = t;
479 p->filter_datapoint[i].d_offset = 0;
480 p->filter_datapoint[i].d_dispersion = MAXDISP;
483 if (offset < 16 * STEP_THRESHOLD) {
484 p->lastpkt_recv_time -= offset;
486 p->p_reachable_bits = 0;
487 p->lastpkt_recv_time = t;
489 filter_datapoints(p, t); /* recalc p->filter_xxx */
490 p->next_action_time -= (time_t)offset;
491 VERB5 bb_error_msg("%s->lastpkt_recv_time=%f", p->p_dotted, p->lastpkt_recv_time);
499 p = xzalloc(sizeof(*p));
500 p->p_lsa = xhost2sockaddr(s, 123);
501 p->p_dotted = xmalloc_sockaddr2dotted_noport(&p->p_lsa->u.sa);
503 p->p_xmt_msg.m_status = MODE_CLIENT | (NTP_VERSION << 3);
504 p->next_action_time = time(NULL); /* = set_next(p, 0); */
505 reset_peer_stats(p, gettime1900d(), 16 * STEP_THRESHOLD);
506 /* Speed up initial sync: with small offsets from peers,
507 * 3 samples will sync
509 p->filter_datapoint[6].d_dispersion = 0;
510 p->filter_datapoint[7].d_dispersion = 0;
512 llist_add_to(&G.ntp_peers, p);
518 const struct sockaddr *from, const struct sockaddr *to, socklen_t addrlen,
519 msg_t *msg, ssize_t len)
525 ret = sendto(fd, msg, len, MSG_DONTWAIT, to, addrlen);
527 ret = send_to_from(fd, msg, len, MSG_DONTWAIT, to, from, addrlen);
530 bb_perror_msg("send failed");
537 send_query_to_peer(peer_t *p)
539 // Why do we need to bind()?
540 // See what happens when we don't bind:
542 // socket(PF_INET, SOCK_DGRAM, IPPROTO_IP) = 3
543 // setsockopt(3, SOL_IP, IP_TOS, [16], 4) = 0
544 // gettimeofday({1259071266, 327885}, NULL) = 0
545 // sendto(3, "xxx", 48, MSG_DONTWAIT, {sa_family=AF_INET, sin_port=htons(123), sin_addr=inet_addr("10.34.32.125")}, 16) = 48
546 // ^^^ we sent it from some source port picked by kernel.
547 // time(NULL) = 1259071266
548 // write(2, "ntpd: entering poll 15 secs\n", 28) = 28
549 // poll([{fd=3, events=POLLIN}], 1, 15000) = 1 ([{fd=3, revents=POLLIN}])
550 // recv(3, "yyy", 68, MSG_DONTWAIT) = 48
551 // ^^^ this recv will receive packets to any local port!
553 // Uncomment this and use strace to see it in action:
554 #define PROBE_LOCAL_ADDR // { len_and_sockaddr lsa; lsa.len = LSA_SIZEOF_SA; getsockname(p->query.fd, &lsa.u.sa, &lsa.len); }
558 len_and_sockaddr *local_lsa;
560 family = p->p_lsa->u.sa.sa_family;
561 p->p_fd = fd = xsocket_type(&local_lsa, family, SOCK_DGRAM);
562 /* local_lsa has "null" address and port 0 now.
563 * bind() ensures we have a *particular port* selected by kernel
564 * and remembered in p->p_fd, thus later recv(p->p_fd)
565 * receives only packets sent to this port.
568 xbind(fd, &local_lsa->u.sa, local_lsa->len);
570 #if ENABLE_FEATURE_IPV6
571 if (family == AF_INET)
573 setsockopt(fd, IPPROTO_IP, IP_TOS, &const_IPTOS_LOWDELAY, sizeof(const_IPTOS_LOWDELAY));
578 * Send out a random 64-bit number as our transmit time. The NTP
579 * server will copy said number into the originate field on the
580 * response that it sends us. This is totally legal per the SNTP spec.
582 * The impact of this is two fold: we no longer send out the current
583 * system time for the world to see (which may aid an attacker), and
584 * it gives us a (not very secure) way of knowing that we're not
585 * getting spoofed by an attacker that can't capture our traffic
586 * but can spoof packets from the NTP server we're communicating with.
588 * Save the real transmit timestamp locally.
590 p->p_xmt_msg.m_xmttime.int_partl = random();
591 p->p_xmt_msg.m_xmttime.fractionl = random();
592 p->p_xmttime = gettime1900d();
594 if (do_sendto(p->p_fd, /*from:*/ NULL, /*to:*/ &p->p_lsa->u.sa, /*addrlen:*/ p->p_lsa->len,
595 &p->p_xmt_msg, NTP_MSGSIZE_NOAUTH) == -1
599 set_next(p, RETRY_INTERVAL);
603 p->p_reachable_bits <<= 1;
604 VERB1 bb_error_msg("sent query to %s", p->p_dotted);
605 set_next(p, QUERYTIME_MAX);
612 step_time(double offset)
619 gettimeofday(&tv, NULL); /* never fails */
620 dtime = offset + tv.tv_sec;
621 dtime += 1.0e-6 * tv.tv_usec;
624 if (settimeofday(&tv, NULL) == -1)
625 bb_perror_msg_and_die("settimeofday");
628 strftime(buf, sizeof(buf), "%a %b %e %H:%M:%S %Z %Y", localtime(&tval));
630 bb_error_msg("setting clock to %s (offset %fs)", buf, offset);
632 // G.time_was_stepped = 1;
637 * Selection and clustering, and their helpers
645 compare_point_edge(const void *aa, const void *bb)
647 const point_t *a = aa;
648 const point_t *b = bb;
649 if (a->edge < b->edge) {
652 return (a->edge > b->edge);
659 compare_survivor_metric(const void *aa, const void *bb)
661 const survivor_t *a = aa;
662 const survivor_t *b = bb;
663 if (a->metric < b->metric)
665 return (a->metric > b->metric);
668 fit(peer_t *p, double rd)
670 if (p->p_reachable_bits == 0) {
671 VERB3 bb_error_msg("peer %s unfit for selection: unreachable", p->p_dotted);
674 //TODO: we never accept such packets anyway, right?
675 if ((p->lastpkt_leap & LI_ALARM) == LI_ALARM
676 || p->lastpkt_stratum >= MAXSTRAT
678 VERB3 bb_error_msg("peer %s unfit for selection: bad status/stratum", p->p_dotted);
681 /* rd is root_distance(p, t) */
682 if (rd > MAXDIST + FREQ_TOLERANCE * (1 << G.poll_exp)) {
683 VERB3 bb_error_msg("peer %s unfit for selection: root distance too high", p->p_dotted);
687 // /* Do we have a loop? */
688 // if (p->refid == p->dstaddr || p->refid == s.refid)
693 select_and_cluster(double t)
697 int size = 3 * G.peer_cnt;
698 /* for selection algorithm */
700 unsigned num_points, num_candidates;
702 unsigned num_falsetickers;
703 /* for cluster algorithm */
704 survivor_t survivor[size];
705 unsigned num_survivors;
711 while (item != NULL) {
712 peer_t *p = (peer_t *) item->data;
713 double rd = root_distance(p, t);
714 double offset = p->filter_offset;
721 VERB4 bb_error_msg("interval: [%f %f %f] %s",
727 point[num_points].p = p;
728 point[num_points].type = -1;
729 point[num_points].edge = offset - rd;
731 point[num_points].p = p;
732 point[num_points].type = 0;
733 point[num_points].edge = offset;
735 point[num_points].p = p;
736 point[num_points].type = 1;
737 point[num_points].edge = offset + rd;
741 num_candidates = num_points / 3;
742 if (num_candidates == 0) {
743 VERB3 bb_error_msg("no valid datapoints, no peer selected");
744 return NULL; /* never happers? */
746 //TODO: sorting does not seem to be done in reference code
747 qsort(point, num_points, sizeof(point[0]), compare_point_edge);
749 /* Start with the assumption that there are no falsetickers.
750 * Attempt to find a nonempty intersection interval containing
751 * the midpoints of all truechimers.
752 * If a nonempty interval cannot be found, increase the number
753 * of assumed falsetickers by one and try again.
754 * If a nonempty interval is found and the number of falsetickers
755 * is less than the number of truechimers, a majority has been found
756 * and the midpoint of each truechimer represents
757 * the candidates available to the cluster algorithm.
759 num_falsetickers = 0;
762 unsigned num_midpoints = 0;
767 for (i = 0; i < num_points; i++) {
769 * if (point[i].type == -1) c++;
770 * if (point[i].type == 1) c--;
771 * and it's simpler to do it this way:
774 if (c >= num_candidates - num_falsetickers) {
775 /* If it was c++ and it got big enough... */
779 if (point[i].type == 0)
783 for (i = num_points-1; i >= 0; i--) {
785 if (c >= num_candidates - num_falsetickers) {
786 high = point[i].edge;
789 if (point[i].type == 0)
792 /* If the number of midpoints is greater than the number
793 * of allowed falsetickers, the intersection contains at
794 * least one truechimer with no midpoint - bad.
795 * Also, interval should be nonempty.
797 if (num_midpoints <= num_falsetickers && low < high)
800 if (num_falsetickers * 2 >= num_candidates) {
801 VERB3 bb_error_msg("too many falsetickers:%d (candidates:%d), no peer selected",
802 num_falsetickers, num_candidates);
806 VERB3 bb_error_msg("selected interval: [%f, %f]; candidates:%d falsetickers:%d",
807 low, high, num_candidates, num_falsetickers);
811 /* Construct a list of survivors (p, metric)
812 * from the chime list, where metric is dominated
813 * first by stratum and then by root distance.
814 * All other things being equal, this is the order of preference.
817 for (i = 0; i < num_points; i++) {
820 if (point[i].edge < low || point[i].edge > high)
823 survivor[num_survivors].p = p;
824 //TODO: save root_distance in point_t and reuse here?
825 survivor[num_survivors].metric = MAXDIST * p->lastpkt_stratum + root_distance(p, t);
826 VERB4 bb_error_msg("survivor[%d] metric:%f peer:%s",
827 num_survivors, survivor[num_survivors].metric, p->p_dotted);
830 /* There must be at least MIN_SELECTED survivors to satisfy the
831 * correctness assertions. Ordinarily, the Byzantine criteria
832 * require four survivors, but for the demonstration here, one
835 if (num_survivors < MIN_SELECTED) {
836 VERB3 bb_error_msg("num_survivors %d < %d, no peer selected",
837 num_survivors, MIN_SELECTED);
841 //looks like this is ONLY used by the fact that later we pick survivor[0].
842 //we can avoid sorting then, just find the minimum once!
843 qsort(survivor, num_survivors, sizeof(survivor[0]), compare_survivor_metric);
845 /* For each association p in turn, calculate the selection
846 * jitter p->sjitter as the square root of the sum of squares
847 * (p->offset - q->offset) over all q associations. The idea is
848 * to repeatedly discard the survivor with maximum selection
849 * jitter until a termination condition is met.
852 unsigned max_idx = max_idx;
853 double max_selection_jitter = max_selection_jitter;
854 double min_jitter = min_jitter;
856 if (num_survivors <= MIN_CLUSTERED) {
857 bb_error_msg("num_survivors %d <= %d, not discarding more",
858 num_survivors, MIN_CLUSTERED);
862 /* To make sure a few survivors are left
863 * for the clustering algorithm to chew on,
864 * we stop if the number of survivors
865 * is less than or equal to MIN_CLUSTERED (3).
867 for (i = 0; i < num_survivors; i++) {
868 double selection_jitter_sq;
869 peer_t *p = survivor[i].p;
871 if (i == 0 || p->filter_jitter < min_jitter)
872 min_jitter = p->filter_jitter;
874 selection_jitter_sq = 0;
875 for (j = 0; j < num_survivors; j++) {
876 peer_t *q = survivor[j].p;
877 //TODO: where is 1/(n-1) * ... multiplier?
878 selection_jitter_sq += SQUARE(p->filter_offset - q->filter_offset);
880 if (i == 0 || selection_jitter_sq > max_selection_jitter) {
881 max_selection_jitter = selection_jitter_sq;
884 VERB5 bb_error_msg("survivor %d selection_jitter^2:%f",
885 i, selection_jitter_sq);
887 max_selection_jitter = SQRT(max_selection_jitter);
888 VERB4 bb_error_msg("max_selection_jitter (at %d):%f min_jitter:%f",
889 max_idx, max_selection_jitter, min_jitter);
891 /* If the maximum selection jitter is less than the
892 * minimum peer jitter, then tossing out more survivors
893 * will not lower the minimum peer jitter, so we might
896 if (max_selection_jitter < min_jitter) {
897 VERB3 bb_error_msg("max_selection_jitter:%f < min_jitter:%f, num_survivors:%d, not discarding more",
898 max_selection_jitter, min_jitter, num_survivors);
902 /* Delete survivor[max_idx] from the list
903 * and go around again.
905 VERB5 bb_error_msg("dropping survivor %d", max_idx);
907 while (max_idx < num_survivors) {
908 survivor[max_idx] = survivor[max_idx + 1];
913 /* Pick the best clock. If the old system peer is on the list
914 * and at the same stratum as the first survivor on the list,
915 * then don't do a clock hop. Otherwise, select the first
916 * survivor on the list as the new system peer.
918 //TODO - see clock_combine()
919 VERB3 bb_error_msg("selected peer %s filter_offset:%f age:%f",
920 survivor[0].p->p_dotted,
921 survivor[0].p->filter_offset,
922 t - survivor[0].p->lastpkt_recv_time
924 return survivor[0].p;
929 * Local clock discipline and its helpers
932 set_new_values(int disc_state, double offset, double recv_time)
934 /* Enter new state and set state variables. Note we use the time
935 * of the last clock filter sample, which must be earlier than
938 VERB3 bb_error_msg("disc_state=%d last_update_offset=%f last_update_recv_time=%f",
939 disc_state, offset, recv_time);
940 G.discipline_state = disc_state;
941 G.last_update_offset = offset;
942 G.last_update_recv_time = recv_time;
944 /* Clock state definitions */
945 #define STATE_NSET 0 /* initial state, "nothing is set" */
946 #define STATE_FSET 1 /* frequency set from file */
947 #define STATE_SPIK 2 /* spike detected */
948 #define STATE_FREQ 3 /* initial frequency */
949 #define STATE_SYNC 4 /* clock synchronized (normal operation) */
950 /* Return: -1: decrease poll interval, 0: leave as is, 1: increase */
952 update_local_clock(peer_t *p, double t)
957 double offset = p->filter_offset;
958 double recv_time = p->lastpkt_recv_time;
961 double since_last_update;
964 abs_offset = fabs(offset);
966 /* If the offset is too large, give up and go home */
967 if (abs_offset > PANIC_THRESHOLD) {
968 bb_error_msg_and_die("offset %f far too big, exiting", offset);
971 /* If this is an old update, for instance as the result
972 * of a system peer change, avoid it. We never use
973 * an old sample or the same sample twice.
975 if (recv_time <= G.last_update_recv_time) {
976 VERB3 bb_error_msg("same or older datapoint: %f >= %f, not using it",
977 G.last_update_recv_time, recv_time);
978 return 0; /* "leave poll interval as is" */
981 /* Clock state machine transition function. This is where the
982 * action is and defines how the system reacts to large time
983 * and frequency errors.
985 since_last_update = recv_time - G.reftime;
987 if (G.discipline_state == STATE_FREQ) {
988 /* Ignore updates until the stepout threshold */
989 if (since_last_update < WATCH_THRESHOLD) {
990 VERB3 bb_error_msg("measuring drift, datapoint ignored, %f sec remains",
991 WATCH_THRESHOLD - since_last_update);
992 return 0; /* "leave poll interval as is" */
994 freq_drift = (offset - G.last_update_offset) / since_last_update;
997 /* There are two main regimes: when the
998 * offset exceeds the step threshold and when it does not.
1000 if (abs_offset > STEP_THRESHOLD) {
1003 switch (G.discipline_state) {
1005 /* The first outlyer: ignore it, switch to SPIK state */
1006 VERB3 bb_error_msg("offset:%f - spike detected", offset);
1007 G.discipline_state = STATE_SPIK;
1008 return -1; /* "decrease poll interval" */
1011 /* Ignore succeeding outlyers until either an inlyer
1012 * is found or the stepout threshold is exceeded.
1014 if (since_last_update < WATCH_THRESHOLD) {
1015 VERB3 bb_error_msg("spike detected, datapoint ignored, %f sec remains",
1016 WATCH_THRESHOLD - since_last_update);
1017 return -1; /* "decrease poll interval" */
1019 /* fall through: we need to step */
1022 /* Step the time and clamp down the poll interval.
1024 * In NSET state an initial frequency correction is
1025 * not available, usually because the frequency file has
1026 * not yet been written. Since the time is outside the
1027 * capture range, the clock is stepped. The frequency
1028 * will be set directly following the stepout interval.
1030 * In FSET state the initial frequency has been set
1031 * from the frequency file. Since the time is outside
1032 * the capture range, the clock is stepped immediately,
1033 * rather than after the stepout interval. Guys get
1034 * nervous if it takes 17 minutes to set the clock for
1037 * In SPIK state the stepout threshold has expired and
1038 * the phase is still above the step threshold. Note
1039 * that a single spike greater than the step threshold
1040 * is always suppressed, even at the longer poll
1043 VERB3 bb_error_msg("stepping time by %f; poll_exp=MINPOLL", offset);
1045 if (option_mask32 & OPT_q) {
1046 /* We were only asked to set time once. Done. */
1050 G.polladj_count = 0;
1051 G.poll_exp = MINPOLL;
1052 G.stratum = MAXSTRAT;
1053 for (item = G.ntp_peers; item != NULL; item = item->link) {
1054 peer_t *pp = (peer_t *) item->data;
1055 reset_peer_stats(pp, t, offset);
1057 if (G.discipline_state == STATE_NSET) {
1058 set_new_values(STATE_FREQ, /*offset:*/ 0, recv_time);
1059 return 1; /* "ok to increase poll interval" */
1061 set_new_values(STATE_SYNC, /*offset:*/ 0, recv_time);
1063 } else { /* abs_offset <= STEP_THRESHOLD */
1065 if (G.poll_exp < MINPOLL) {
1066 VERB3 bb_error_msg("saw small offset %f, disabling burst mode", offset);
1067 G.poll_exp = MINPOLL;
1070 /* Compute the clock jitter as the RMS of exponentially
1071 * weighted offset differences. Used by the poll adjust code.
1073 etemp = SQUARE(G.discipline_jitter);
1074 dtemp = SQUARE(MAXD(fabs(offset - G.last_update_offset), G_precision_sec));
1075 G.discipline_jitter = SQRT(etemp + (dtemp - etemp) / AVG);
1076 VERB3 bb_error_msg("discipline jitter=%f", G.discipline_jitter);
1078 switch (G.discipline_state) {
1080 if (option_mask32 & OPT_q) {
1081 /* We were only asked to set time once.
1082 * The clock is precise enough, no need to step.
1086 /* This is the first update received and the frequency
1087 * has not been initialized. The first thing to do
1088 * is directly measure the oscillator frequency.
1090 set_new_values(STATE_FREQ, offset, recv_time);
1091 VERB3 bb_error_msg("transitioning to FREQ, datapoint ignored");
1092 return -1; /* "decrease poll interval" */
1094 #if 0 /* this is dead code for now */
1096 /* This is the first update and the frequency
1097 * has been initialized. Adjust the phase, but
1098 * don't adjust the frequency until the next update.
1100 set_new_values(STATE_SYNC, offset, recv_time);
1101 /* freq_drift remains 0 */
1106 /* since_last_update >= WATCH_THRESHOLD, we waited enough.
1107 * Correct the phase and frequency and switch to SYNC state.
1108 * freq_drift was already estimated (see code above)
1110 set_new_values(STATE_SYNC, offset, recv_time);
1114 /* Compute freq_drift due to PLL and FLL contributions.
1116 * The FLL and PLL frequency gain constants
1117 * depend on the poll interval and Allan
1118 * intercept. The FLL is not used below one-half
1119 * the Allan intercept. Above that the loop gain
1120 * increases in steps to 1 / AVG.
1122 if ((1 << G.poll_exp) > ALLAN / 2) {
1123 etemp = FLL - G.poll_exp;
1126 freq_drift += (offset - G.last_update_offset) / (MAXD(since_last_update, ALLAN) * etemp);
1128 /* For the PLL the integration interval
1129 * (numerator) is the minimum of the update
1130 * interval and poll interval. This allows
1131 * oversampling, but not undersampling.
1133 etemp = MIND(since_last_update, (1 << G.poll_exp));
1134 dtemp = (4 * PLL) << G.poll_exp;
1135 freq_drift += offset * etemp / SQUARE(dtemp);
1136 set_new_values(STATE_SYNC, offset, recv_time);
1139 G.stratum = p->lastpkt_stratum + 1;
1143 G.leap = p->lastpkt_leap;
1144 G.refid = p->lastpkt_refid;
1145 G.rootdelay = p->lastpkt_rootdelay + p->lastpkt_delay;
1146 dtemp = p->filter_jitter; // SQRT(SQUARE(p->filter_jitter) + SQUARE(s.jitter));
1147 dtemp += MAXD(p->filter_dispersion + FREQ_TOLERANCE * (t - p->lastpkt_recv_time) + abs_offset, MINDISP);
1148 G.rootdisp = p->lastpkt_rootdisp + dtemp;
1149 VERB3 bb_error_msg("updating leap/refid/reftime/rootdisp from peer %s", p->p_dotted);
1151 /* We are in STATE_SYNC now, but did not do adjtimex yet.
1152 * (Any other state does not reach this, they all return earlier)
1153 * By this time, freq_drift and G.last_update_offset are set
1154 * to values suitable for adjtimex.
1156 * Calculate the new frequency drift and frequency stability (wander).
1157 * Compute the clock wander as the RMS of exponentially weighted
1158 * frequency differences. This is not used directly, but can,
1159 * along with the jitter, be a highly useful monitoring and
1162 dtemp = G.discipline_freq_drift + freq_drift;
1163 G.discipline_freq_drift = MAXD(MIND(MAXFREQ, dtemp), -MAXFREQ);
1164 etemp = SQUARE(G.discipline_wander);
1165 dtemp = SQUARE(dtemp);
1166 G.discipline_wander = SQRT(etemp + (dtemp - etemp) / AVG);
1169 bb_error_msg("discipline freq_drift=%.9f(int:%ld corr:%e) wander=%f",
1170 G.discipline_freq_drift,
1171 (long)(G.discipline_freq_drift * 65536e6),
1173 G.discipline_wander);
1174 memset(&tmx, 0, sizeof(tmx));
1175 if (adjtimex(&tmx) < 0)
1176 bb_perror_msg_and_die("adjtimex");
1177 VERB3 bb_error_msg("p adjtimex freq:%ld offset:%ld constant:%ld status:0x%x",
1178 tmx.freq, tmx.offset, tmx.constant, tmx.status);
1182 if (!G.adjtimex_was_done) {
1183 G.adjtimex_was_done = 1;
1184 /* When we use adjtimex for the very first time,
1185 * we need to ADD to pre-existing tmx.offset - it may be !0
1187 memset(&tmx, 0, sizeof(tmx));
1188 if (adjtimex(&tmx) < 0)
1189 bb_perror_msg_and_die("adjtimex");
1190 old_tmx_offset = tmx.offset;
1192 memset(&tmx, 0, sizeof(tmx));
1194 //doesn't work, offset remains 0 (!):
1195 //ntpd: set adjtimex freq:1786097 tmx.offset:77487
1196 //ntpd: prev adjtimex freq:1786097 tmx.offset:0
1197 //ntpd: cur adjtimex freq:1786097 tmx.offset:0
1198 tmx.modes = ADJ_FREQUENCY | ADJ_OFFSET;
1199 /* 65536 is one ppm */
1200 tmx.freq = G.discipline_freq_drift * 65536e6;
1201 tmx.offset = G.last_update_offset * 1000000; /* usec */
1203 tmx.modes = ADJ_OFFSET | ADJ_STATUS | ADJ_TIMECONST;// | ADJ_MAXERROR | ADJ_ESTERROR;
1204 tmx.offset = (G.last_update_offset * 1000000) /* usec */
1205 /* + (G.last_update_offset < 0 ? -0.5 : 0.5) - too small to bother */
1206 + old_tmx_offset; /* almost always 0 */
1207 tmx.status = STA_PLL;
1208 //if (sys_leap == LEAP_ADDSECOND)
1209 // tmx.status |= STA_INS;
1210 //else if (sys_leap == LEAP_DELSECOND)
1211 // tmx.status |= STA_DEL;
1212 tmx.constant = G.poll_exp - 4;
1213 //tmx.esterror = (u_int32)(clock_jitter * 1e6);
1214 //tmx.maxerror = (u_int32)((sys_rootdelay / 2 + sys_rootdisp) * 1e6);
1215 VERB3 bb_error_msg("b adjtimex freq:%ld offset:%ld constant:%ld status:0x%x",
1216 tmx.freq, tmx.offset, tmx.constant, tmx.status);
1217 rc = adjtimex(&tmx);
1219 bb_perror_msg_and_die("adjtimex");
1221 bb_error_msg("adjtimex:%d freq:%ld offset:%ld constant:%ld status:0x%x",
1222 rc, tmx.freq, tmx.offset, tmx.constant, tmx.status);
1223 memset(&tmx, 0, sizeof(tmx));
1224 if (adjtimex(&tmx) < 0)
1225 bb_perror_msg_and_die("adjtimex");
1226 VERB3 bb_error_msg("c adjtimex freq:%ld offset:%ld constant:%ld status:0x%x",
1227 tmx.freq, tmx.offset, tmx.constant, tmx.status);
1229 // #define STA_MODE 0x4000 /* mode (0 = PLL, 1 = FLL) (ro) */ - ?
1230 // it appeared after a while:
1231 //ntpd: p adjtimex freq:-14545653 offset:-5396 constant:10 status:0x41
1232 //ntpd: c adjtimex freq:-14547835 offset:-8307 constant:10 status:0x1
1233 //ntpd: p adjtimex freq:-14547835 offset:-6398 constant:10 status:0x41
1234 //ntpd: c adjtimex freq:-14550486 offset:-10158 constant:10 status:0x1
1235 //ntpd: p adjtimex freq:-14550486 offset:-6132 constant:10 status:0x41
1236 //ntpd: c adjtimex freq:-14636129 offset:-10158 constant:10 status:0x4001
1237 //ntpd: p adjtimex freq:-14636129 offset:-10002 constant:10 status:0x4041
1238 //ntpd: c adjtimex freq:-14636245 offset:-7497 constant:10 status:0x1
1239 //ntpd: p adjtimex freq:-14636245 offset:-4573 constant:10 status:0x41
1240 //ntpd: c adjtimex freq:-14642034 offset:-11715 constant:10 status:0x1
1241 //ntpd: p adjtimex freq:-14642034 offset:-4098 constant:10 status:0x41
1242 //ntpd: c adjtimex freq:-14699112 offset:-11746 constant:10 status:0x4001
1243 //ntpd: p adjtimex freq:-14699112 offset:-4239 constant:10 status:0x4041
1244 //ntpd: c adjtimex freq:-14762330 offset:-12786 constant:10 status:0x4001
1245 //ntpd: p adjtimex freq:-14762330 offset:-4434 constant:10 status:0x4041
1246 //ntpd: b adjtimex freq:0 offset:-9669 constant:8 status:0x1
1247 //ntpd: adjtimex:0 freq:-14809095 offset:-9669 constant:10 status:0x4001
1248 //ntpd: c adjtimex freq:-14809095 offset:-9669 constant:10 status:0x4001
1250 return 1; /* "ok to increase poll interval" */
1255 * We've got a new reply packet from a peer, process it
1259 retry_interval(void)
1261 /* Local problem, want to retry soon */
1262 unsigned interval, r;
1263 interval = RETRY_INTERVAL;
1265 interval += r % (unsigned)(RETRY_INTERVAL / 4);
1266 VERB3 bb_error_msg("chose retry interval:%u", interval);
1270 poll_interval(int exponent) /* exp is always -1 or 0 */
1272 /* Want to send next packet at (1 << G.poll_exp) + small random value */
1273 unsigned interval, r;
1274 exponent += G.poll_exp; /* G.poll_exp is always > 0 */
1275 /* never true: if (exp < 0) exp = 0; */
1276 interval = 1 << exponent;
1278 interval += ((r & (interval-1)) >> 4) + ((r >> 8) & 1); /* + 1/16 of interval, max */
1279 VERB3 bb_error_msg("chose poll interval:%u (poll_exp:%d exp:%d)", interval, G.poll_exp, exponent);
1283 recv_and_process_peer_pkt(peer_t *p)
1288 double T1, T2, T3, T4;
1290 datapoint_t *datapoint;
1293 /* We can recvfrom here and check from.IP, but some multihomed
1294 * ntp servers reply from their *other IP*.
1295 * TODO: maybe we should check at least what we can: from.port == 123?
1297 size = recv(p->p_fd, &msg, sizeof(msg), MSG_DONTWAIT);
1299 bb_perror_msg("recv(%s) error", p->p_dotted);
1300 if (errno == EHOSTUNREACH || errno == EHOSTDOWN
1301 || errno == ENETUNREACH || errno == ENETDOWN
1302 || errno == ECONNREFUSED || errno == EADDRNOTAVAIL
1305 //TODO: always do this?
1306 set_next(p, retry_interval());
1312 if (size != NTP_MSGSIZE_NOAUTH && size != NTP_MSGSIZE) {
1313 bb_error_msg("malformed packet received from %s", p->p_dotted);
1317 if (msg.m_orgtime.int_partl != p->p_xmt_msg.m_xmttime.int_partl
1318 || msg.m_orgtime.fractionl != p->p_xmt_msg.m_xmttime.fractionl
1323 if ((msg.m_status & LI_ALARM) == LI_ALARM
1324 || msg.m_stratum == 0
1325 || msg.m_stratum > NTP_MAXSTRATUM
1327 // TODO: stratum 0 responses may have commands in 32-bit m_refid field:
1328 // "DENY", "RSTR" - peer does not like us at all
1329 // "RATE" - peer is overloaded, reduce polling freq
1330 interval = poll_interval(0);
1331 bb_error_msg("reply from %s: not synced, next query in %us", p->p_dotted, interval);
1336 // * Verify the server is synchronized with valid stratum and
1337 // * reference time not later than the transmit time.
1339 // if (p->lastpkt_leap == NOSYNC || p->lastpkt_stratum >= MAXSTRAT)
1340 // return; /* unsynchronized */
1342 // /* Verify valid root distance */
1343 // if (msg.m_rootdelay / 2 + msg.m_rootdisp >= MAXDISP || p->lastpkt_reftime > msg.m_xmt)
1344 // return; /* invalid header values */
1346 p->lastpkt_leap = msg.m_status;
1347 p->lastpkt_rootdelay = sfp_to_d(msg.m_rootdelay);
1348 p->lastpkt_rootdisp = sfp_to_d(msg.m_rootdisp);
1349 p->lastpkt_refid = msg.m_refid;
1352 * From RFC 2030 (with a correction to the delay math):
1354 * Timestamp Name ID When Generated
1355 * ------------------------------------------------------------
1356 * Originate Timestamp T1 time request sent by client
1357 * Receive Timestamp T2 time request received by server
1358 * Transmit Timestamp T3 time reply sent by server
1359 * Destination Timestamp T4 time reply received by client
1361 * The roundtrip delay and local clock offset are defined as
1363 * delay = (T4 - T1) - (T3 - T2); offset = ((T2 - T1) + (T3 - T4)) / 2
1366 T2 = lfp_to_d(msg.m_rectime);
1367 T3 = lfp_to_d(msg.m_xmttime);
1368 T4 = gettime1900d();
1370 p->lastpkt_recv_time = T4;
1372 VERB5 bb_error_msg("%s->lastpkt_recv_time=%f", p->p_dotted, p->lastpkt_recv_time);
1373 p->datapoint_idx = p->p_reachable_bits ? (p->datapoint_idx + 1) % NUM_DATAPOINTS : 0;
1374 datapoint = &p->filter_datapoint[p->datapoint_idx];
1375 datapoint->d_recv_time = T4;
1376 datapoint->d_offset = ((T2 - T1) + (T3 - T4)) / 2;
1377 /* The delay calculation is a special case. In cases where the
1378 * server and client clocks are running at different rates and
1379 * with very fast networks, the delay can appear negative. In
1380 * order to avoid violating the Principle of Least Astonishment,
1381 * the delay is clamped not less than the system precision.
1383 p->lastpkt_delay = (T4 - T1) - (T3 - T2);
1384 datapoint->d_dispersion = LOG2D(msg.m_precision_exp) + G_precision_sec;
1385 if (!p->p_reachable_bits) {
1386 /* 1st datapoint ever - replicate offset in every element */
1388 for (i = 1; i < NUM_DATAPOINTS; i++) {
1389 p->filter_datapoint[i].d_offset = datapoint->d_offset;
1393 p->p_reachable_bits |= 1;
1395 bb_error_msg("reply from %s: reach 0x%02x offset %f delay %f",
1397 p->p_reachable_bits,
1398 datapoint->d_offset, p->lastpkt_delay);
1401 /* Muck with statictics and update the clock */
1402 filter_datapoints(p, T4);
1403 q = select_and_cluster(T4);
1406 rc = update_local_clock(q, T4);
1409 /* Adjust the poll interval by comparing the current offset
1410 * with the clock jitter. If the offset is less than
1411 * the clock jitter times a constant, then the averaging interval
1412 * is increased, otherwise it is decreased. A bit of hysteresis
1413 * helps calm the dance. Works best using burst mode.
1416 bb_error_msg("offset:%f POLLADJ_GATE*discipline_jitter:%f poll:%s",
1417 q->filter_offset, POLLADJ_GATE * G.discipline_jitter,
1418 fabs(q->filter_offset) < POLLADJ_GATE * G.discipline_jitter
1422 if (rc > 0 && fabs(q->filter_offset) < POLLADJ_GATE * G.discipline_jitter) {
1423 G.polladj_count += G.poll_exp;
1424 if (G.polladj_count > POLLADJ_LIMIT) {
1425 G.polladj_count = 0;
1426 if (G.poll_exp < MAXPOLL) {
1428 VERB3 bb_error_msg("polladj: discipline_jitter:%f ++poll_exp=%d",
1429 G.discipline_jitter, G.poll_exp);
1432 VERB3 bb_error_msg("polladj: incr:%d", G.polladj_count);
1435 G.polladj_count -= G.poll_exp * 2;
1436 if (G.polladj_count < -POLLADJ_LIMIT) {
1437 G.polladj_count = 0;
1438 if (G.poll_exp > MINPOLL) {
1440 VERB3 bb_error_msg("polladj: discipline_jitter:%f --poll_exp=%d",
1441 G.discipline_jitter, G.poll_exp);
1444 VERB3 bb_error_msg("polladj: decr:%d", G.polladj_count);
1449 /* Decide when to send new query for this peer */
1450 interval = poll_interval(0);
1451 set_next(p, interval);
1454 /* We do not expect any more packets from this peer for now.
1455 * Closing the socket informs kernel about it.
1456 * We open a new socket when we send a new query.
1464 #if ENABLE_FEATURE_NTPD_SERVER
1466 recv_and_process_client_pkt(void /*int fd*/)
1471 len_and_sockaddr *to;
1472 struct sockaddr *from;
1474 uint8_t query_status;
1475 l_fixedpt_t query_xmttime;
1477 to = get_sock_lsa(G.listen_fd);
1478 from = xzalloc(to->len);
1480 size = recv_from_to(G.listen_fd, &msg, sizeof(msg), MSG_DONTWAIT, from, &to->u.sa, to->len);
1481 if (size != NTP_MSGSIZE_NOAUTH && size != NTP_MSGSIZE) {
1484 if (errno == EAGAIN)
1486 bb_perror_msg_and_die("recv");
1488 addr = xmalloc_sockaddr2dotted_noport(from);
1489 bb_error_msg("malformed packet received from %s: size %u", addr, (int)size);
1494 query_status = msg.m_status;
1495 query_xmttime = msg.m_xmttime;
1497 /* Build a reply packet */
1498 memset(&msg, 0, sizeof(msg));
1499 msg.m_status = G.stratum < MAXSTRAT ? G.leap : LI_ALARM;
1500 msg.m_status |= (query_status & VERSION_MASK);
1501 msg.m_status |= ((query_status & MODE_MASK) == MODE_CLIENT) ?
1502 MODE_SERVER : MODE_SYM_PAS;
1503 msg.m_stratum = G.stratum;
1504 msg.m_ppoll = G.poll_exp;
1505 msg.m_precision_exp = G_precision_exp;
1506 rectime = gettime1900d();
1507 msg.m_xmttime = msg.m_rectime = d_to_lfp(rectime);
1508 msg.m_reftime = d_to_lfp(G.reftime);
1509 msg.m_orgtime = query_xmttime;
1510 msg.m_rootdelay = d_to_sfp(G.rootdelay);
1511 //simple code does not do this, fix simple code!
1512 msg.m_rootdisp = d_to_sfp(G.rootdisp);
1513 version = (query_status & VERSION_MASK); /* ... >> VERSION_SHIFT - done below instead */
1514 msg.m_refid = G.refid; // (version > (3 << VERSION_SHIFT)) ? G.refid : G.refid3;
1516 /* We reply from the local address packet was sent to,
1517 * this makes to/from look swapped here: */
1518 do_sendto(G.listen_fd,
1519 /*from:*/ &to->u.sa, /*to:*/ from, /*addrlen:*/ to->len,
1528 /* Upstream ntpd's options:
1530 * -4 Force DNS resolution of host names to the IPv4 namespace.
1531 * -6 Force DNS resolution of host names to the IPv6 namespace.
1532 * -a Require cryptographic authentication for broadcast client,
1533 * multicast client and symmetric passive associations.
1534 * This is the default.
1535 * -A Do not require cryptographic authentication for broadcast client,
1536 * multicast client and symmetric passive associations.
1537 * This is almost never a good idea.
1538 * -b Enable the client to synchronize to broadcast servers.
1540 * Specify the name and path of the configuration file,
1541 * default /etc/ntp.conf
1542 * -d Specify debugging mode. This option may occur more than once,
1543 * with each occurrence indicating greater detail of display.
1545 * Specify debugging level directly.
1547 * Specify the name and path of the frequency file.
1548 * This is the same operation as the "driftfile FILE"
1549 * configuration command.
1550 * -g Normally, ntpd exits with a message to the system log
1551 * if the offset exceeds the panic threshold, which is 1000 s
1552 * by default. This option allows the time to be set to any value
1553 * without restriction; however, this can happen only once.
1554 * If the threshold is exceeded after that, ntpd will exit
1555 * with a message to the system log. This option can be used
1556 * with the -q and -x options. See the tinker command for other options.
1558 * Chroot the server to the directory jaildir. This option also implies
1559 * that the server attempts to drop root privileges at startup
1560 * (otherwise, chroot gives very little additional security).
1561 * You may need to also specify a -u option.
1563 * Specify the name and path of the symmetric key file,
1564 * default /etc/ntp/keys. This is the same operation
1565 * as the "keys FILE" configuration command.
1567 * Specify the name and path of the log file. The default
1568 * is the system log file. This is the same operation as
1569 * the "logfile FILE" configuration command.
1570 * -L Do not listen to virtual IPs. The default is to listen.
1572 * -N To the extent permitted by the operating system,
1573 * run the ntpd at the highest priority.
1575 * Specify the name and path of the file used to record the ntpd
1576 * process ID. This is the same operation as the "pidfile FILE"
1577 * configuration command.
1579 * To the extent permitted by the operating system,
1580 * run the ntpd at the specified priority.
1581 * -q Exit the ntpd just after the first time the clock is set.
1582 * This behavior mimics that of the ntpdate program, which is
1583 * to be retired. The -g and -x options can be used with this option.
1584 * Note: The kernel time discipline is disabled with this option.
1586 * Specify the default propagation delay from the broadcast/multicast
1587 * server to this client. This is necessary only if the delay
1588 * cannot be computed automatically by the protocol.
1590 * Specify the directory path for files created by the statistics
1591 * facility. This is the same operation as the "statsdir DIR"
1592 * configuration command.
1594 * Add a key number to the trusted key list. This option can occur
1597 * Specify a user, and optionally a group, to switch to.
1600 * Add a system variable listed by default.
1601 * -x Normally, the time is slewed if the offset is less than the step
1602 * threshold, which is 128 ms by default, and stepped if above
1603 * the threshold. This option sets the threshold to 600 s, which is
1604 * well within the accuracy window to set the clock manually.
1605 * Note: since the slew rate of typical Unix kernels is limited
1606 * to 0.5 ms/s, each second of adjustment requires an amortization
1607 * interval of 2000 s. Thus, an adjustment as much as 600 s
1608 * will take almost 14 days to complete. This option can be used
1609 * with the -g and -q options. See the tinker command for other options.
1610 * Note: The kernel time discipline is disabled with this option.
1613 /* By doing init in a separate function we decrease stack usage
1616 static NOINLINE void ntp_init(char **argv)
1624 bb_error_msg_and_die(bb_msg_you_must_be_root);
1626 /* Set some globals */
1628 /* With constant b = 100, G.precision_exp is also constant -6.
1629 * Uncomment this to verify.
1636 /* We can use sys_clock_getres but assuming 10ms tick should be fine */
1637 clock_getres(CLOCK_REALTIME, &tp);
1639 tp.tv_nsec = 10000000;
1640 b = 1000000000 / tp.tv_nsec; /* convert to Hz */
1642 b = 100; /* b = 1000000000/10000000 = 100 */
1646 /*G.precision_exp = prec;*/
1647 /*G.precision_sec = (1.0 / (1 << (- prec)));*/
1648 bb_error_msg("G.precision_exp:%d sec:%f", prec, G_precision_sec); /* -6 */
1651 G.stratum = MAXSTRAT;
1652 G.poll_exp = 1; /* should use MINPOLL, but 1 speeds up initial sync */
1653 G.reftime = G.last_update_recv_time = gettime1900d();
1657 opt_complementary = "dd:p::"; /* d: counter, p: list */
1658 opts = getopt32(argv,
1660 "p:"IF_FEATURE_NTPD_SERVER("l") /* NOT compat */
1662 "46aAbgL", /* compat, ignored */
1663 &peers, &G.verbose);
1664 if (!(opts & (OPT_p|OPT_l)))
1666 // if (opts & OPT_x) /* disable stepping, only slew is allowed */
1667 // G.time_was_stepped = 1;
1669 add_peers(llist_pop(&peers));
1670 if (!(opts & OPT_n)) {
1671 bb_daemonize_or_rexec(DAEMON_DEVNULL_STDIO, argv);
1672 logmode = LOGMODE_NONE;
1674 #if ENABLE_FEATURE_NTPD_SERVER
1677 G.listen_fd = create_and_bind_dgram_or_die(NULL, 123);
1678 socket_want_pktinfo(G.listen_fd);
1679 setsockopt(G.listen_fd, IPPROTO_IP, IP_TOS, &const_IPTOS_LOWDELAY, sizeof(const_IPTOS_LOWDELAY));
1682 /* I hesitate to set -20 prio. -15 should be high enough for timekeeping */
1684 setpriority(PRIO_PROCESS, 0, -15);
1686 bb_signals((1 << SIGTERM) | (1 << SIGINT), record_signo);
1687 bb_signals((1 << SIGPIPE) | (1 << SIGHUP), SIG_IGN);
1690 int ntpd_main(int argc UNUSED_PARAM, char **argv) MAIN_EXTERNALLY_VISIBLE;
1691 int ntpd_main(int argc UNUSED_PARAM, char **argv)
1697 memset(&g, 0, sizeof(g));
1698 SET_PTR_TO_GLOBALS(&g);
1703 /* if ENABLE_FEATURE_NTPD_SERVER, + 1 for listen_fd: */
1704 unsigned cnt = g.peer_cnt + ENABLE_FEATURE_NTPD_SERVER;
1705 idx2peer = xzalloc(sizeof(idx2peer[0]) * cnt);
1706 pfd = xzalloc(sizeof(pfd[0]) * cnt);
1709 while (!bb_got_signal) {
1712 unsigned sent_cnt, trial_cnt;
1714 time_t cur_time, nextaction;
1716 /* Nothing between here and poll() blocks for any significant time */
1718 cur_time = time(NULL);
1719 nextaction = cur_time + 3600;
1722 #if ENABLE_FEATURE_NTPD_SERVER
1723 if (g.listen_fd != -1) {
1724 pfd[0].fd = g.listen_fd;
1725 pfd[0].events = POLLIN;
1729 /* Pass over peer list, send requests, time out on receives */
1730 sent_cnt = trial_cnt = 0;
1731 for (item = g.ntp_peers; item != NULL; item = item->link) {
1732 peer_t *p = (peer_t *) item->data;
1734 /* Overflow-safe "if (p->next_action_time <= cur_time) ..." */
1735 if ((int)(cur_time - p->next_action_time) >= 0) {
1736 if (p->p_fd == -1) {
1737 /* Time to send new req */
1739 if (send_query_to_peer(p) == 0)
1742 /* Timed out waiting for reply */
1745 timeout = poll_interval(-1); /* try a bit faster */
1746 bb_error_msg("timed out waiting for %s, reach 0x%02x, next query in %us",
1747 p->p_dotted, p->p_reachable_bits, timeout);
1748 set_next(p, timeout);
1752 if (p->next_action_time < nextaction)
1753 nextaction = p->next_action_time;
1756 /* Wait for reply from this peer */
1757 pfd[i].fd = p->p_fd;
1758 pfd[i].events = POLLIN;
1764 // if ((trial_cnt > 0 && sent_cnt == 0) || g.peer_cnt == 0) {
1765 // G.time_was_stepped = 1;
1768 timeout = nextaction - cur_time;
1772 /* Here we may block */
1773 VERB2 bb_error_msg("poll %us, sockets:%u", timeout, i);
1774 nfds = poll(pfd, i, timeout * 1000);
1778 /* Process any received packets */
1780 #if ENABLE_FEATURE_NTPD_SERVER
1781 if (g.listen_fd != -1) {
1782 if (pfd[0].revents /* & (POLLIN|POLLERR)*/) {
1784 recv_and_process_client_pkt(/*g.listen_fd*/);
1789 for (; nfds != 0 && j < i; j++) {
1790 if (pfd[j].revents /* & (POLLIN|POLLERR)*/) {
1792 recv_and_process_peer_pkt(idx2peer[j]);
1795 } /* while (!bb_got_signal) */
1797 kill_myself_with_sig(bb_got_signal);
1805 /*** openntpd-4.6 uses only adjtime, not adjtimex ***/
1807 /*** ntp-4.2.6/ntpd/ntp_loopfilter.c - adjtimex usage ***/
1811 direct_freq(double fp_offset)
1816 * If the kernel is enabled, we need the residual offset to
1817 * calculate the frequency correction.
1819 if (pll_control && kern_enable) {
1820 memset(&ntv, 0, sizeof(ntv));
1823 clock_offset = ntv.offset / 1e9;
1824 #else /* STA_NANO */
1825 clock_offset = ntv.offset / 1e6;
1826 #endif /* STA_NANO */
1827 drift_comp = FREQTOD(ntv.freq);
1829 #endif /* KERNEL_PLL */
1830 set_freq((fp_offset - clock_offset) / (current_time - clock_epoch) + drift_comp);
1836 set_freq(double freq) /* frequency update */
1844 * If the kernel is enabled, update the kernel frequency.
1846 if (pll_control && kern_enable) {
1847 memset(&ntv, 0, sizeof(ntv));
1848 ntv.modes = MOD_FREQUENCY;
1849 ntv.freq = DTOFREQ(drift_comp);
1851 snprintf(tbuf, sizeof(tbuf), "kernel %.3f PPM", drift_comp * 1e6);
1852 report_event(EVNT_FSET, NULL, tbuf);
1854 snprintf(tbuf, sizeof(tbuf), "ntpd %.3f PPM", drift_comp * 1e6);
1855 report_event(EVNT_FSET, NULL, tbuf);
1857 #else /* KERNEL_PLL */
1858 snprintf(tbuf, sizeof(tbuf), "ntpd %.3f PPM", drift_comp * 1e6);
1859 report_event(EVNT_FSET, NULL, tbuf);
1860 #endif /* KERNEL_PLL */
1869 * This code segment works when clock adjustments are made using
1870 * precision time kernel support and the ntp_adjtime() system
1871 * call. This support is available in Solaris 2.6 and later,
1872 * Digital Unix 4.0 and later, FreeBSD, Linux and specially
1873 * modified kernels for HP-UX 9 and Ultrix 4. In the case of the
1874 * DECstation 5000/240 and Alpha AXP, additional kernel
1875 * modifications provide a true microsecond clock and nanosecond
1876 * clock, respectively.
1878 * Important note: The kernel discipline is used only if the
1879 * step threshold is less than 0.5 s, as anything higher can
1880 * lead to overflow problems. This might occur if some misguided
1881 * lad set the step threshold to something ridiculous.
1883 if (pll_control && kern_enable) {
1885 #define MOD_BITS (MOD_OFFSET | MOD_MAXERROR | MOD_ESTERROR | MOD_STATUS | MOD_TIMECONST)
1888 * We initialize the structure for the ntp_adjtime()
1889 * system call. We have to convert everything to
1890 * microseconds or nanoseconds first. Do not update the
1891 * system variables if the ext_enable flag is set. In
1892 * this case, the external clock driver will update the
1893 * variables, which will be read later by the local
1894 * clock driver. Afterwards, remember the time and
1895 * frequency offsets for jitter and stability values and
1896 * to update the frequency file.
1898 memset(&ntv, 0, sizeof(ntv));
1900 ntv.modes = MOD_STATUS;
1903 ntv.modes = MOD_BITS | MOD_NANO;
1904 #else /* STA_NANO */
1905 ntv.modes = MOD_BITS;
1906 #endif /* STA_NANO */
1907 if (clock_offset < 0)
1912 ntv.offset = (int32)(clock_offset * 1e9 + dtemp);
1913 ntv.constant = sys_poll;
1914 #else /* STA_NANO */
1915 ntv.offset = (int32)(clock_offset * 1e6 + dtemp);
1916 ntv.constant = sys_poll - 4;
1917 #endif /* STA_NANO */
1918 ntv.esterror = (u_int32)(clock_jitter * 1e6);
1919 ntv.maxerror = (u_int32)((sys_rootdelay / 2 + sys_rootdisp) * 1e6);
1920 ntv.status = STA_PLL;
1923 * Enable/disable the PPS if requested.
1926 if (!(pll_status & STA_PPSTIME))
1927 report_event(EVNT_KERN,
1928 NULL, "PPS enabled");
1929 ntv.status |= STA_PPSTIME | STA_PPSFREQ;
1931 if (pll_status & STA_PPSTIME)
1932 report_event(EVNT_KERN,
1933 NULL, "PPS disabled");
1934 ntv.status &= ~(STA_PPSTIME |
1937 if (sys_leap == LEAP_ADDSECOND)
1938 ntv.status |= STA_INS;
1939 else if (sys_leap == LEAP_DELSECOND)
1940 ntv.status |= STA_DEL;
1944 * Pass the stuff to the kernel. If it squeals, turn off
1945 * the pps. In any case, fetch the kernel offset,
1946 * frequency and jitter.
1948 if (ntp_adjtime(&ntv) == TIME_ERROR) {
1949 if (!(ntv.status & STA_PPSSIGNAL))
1950 report_event(EVNT_KERN, NULL,
1953 pll_status = ntv.status;
1955 clock_offset = ntv.offset / 1e9;
1956 #else /* STA_NANO */
1957 clock_offset = ntv.offset / 1e6;
1958 #endif /* STA_NANO */
1959 clock_frequency = FREQTOD(ntv.freq);
1962 * If the kernel PPS is lit, monitor its performance.
1964 if (ntv.status & STA_PPSTIME) {
1966 clock_jitter = ntv.jitter / 1e9;
1967 #else /* STA_NANO */
1968 clock_jitter = ntv.jitter / 1e6;
1969 #endif /* STA_NANO */
1972 #if defined(STA_NANO) && NTP_API == 4
1974 * If the TAI changes, update the kernel TAI.
1976 if (loop_tai != sys_tai) {
1978 ntv.modes = MOD_TAI;
1979 ntv.constant = sys_tai;
1982 #endif /* STA_NANO */
1984 #endif /* KERNEL_PLL */