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 static NOINLINE double my_SQRT(double X)
309 double Xhalf = X * 0.5;
311 /* Fast and good approximation to 1/sqrt(X), black magic */
313 /*v.i = 0x5f3759df - (v.i >> 1);*/
314 v.i = 0x5f375a86 - (v.i >> 1); /* - this constant is slightly better */
315 invsqrt = v.f; /* better than 0.2% accuracy */
317 /* Refining it using Newton's method: x1 = x0 - f(x0)/f'(x0)
318 * f(x) = 1/(x*x) - X (f==0 when x = 1/sqrt(X))
320 * f(x)/f'(x) = (X - 1/(x*x)) / (2/(x*x*x)) = X*x*x*x/2 - x/2
321 * x1 = x0 - (X*x0*x0*x0/2 - x0/2) = 1.5*x0 - X*x0*x0*x0/2 = x0*(1.5 - (X/2)*x0*x0)
323 invsqrt = invsqrt * (1.5 - Xhalf * invsqrt * invsqrt); /* ~0.05% accuracy */
324 /* invsqrt = invsqrt * (1.5 - Xhalf * invsqrt * invsqrt); 2nd iter: ~0.0001% accuracy */
325 /* With 4 iterations, more than half results will be exact,
326 * at 6th iterations result stabilizes with about 72% results exact.
327 * We are well satisfied with 0.05% accuracy.
330 return X * invsqrt; /* X * 1/sqrt(X) ~= sqrt(X) */
332 static ALWAYS_INLINE double SQRT(double X)
334 /* If this arch doesn't use IEEE 754 floats, fall back to using libm */
335 if (sizeof(float) != 4)
338 /* This avoids needing libm, saves about 1.2k on x86-32 */
346 gettimeofday(&tv, NULL); /* never fails */
347 G.cur_time = tv.tv_sec + (1.0e-6 * tv.tv_usec) + OFFSET_1900_1970;
352 d_to_tv(double d, struct timeval *tv)
354 tv->tv_sec = (long)d;
355 tv->tv_usec = (d - tv->tv_sec) * 1000000;
359 lfp_to_d(l_fixedpt_t lfp)
362 lfp.int_partl = ntohl(lfp.int_partl);
363 lfp.fractionl = ntohl(lfp.fractionl);
364 ret = (double)lfp.int_partl + ((double)lfp.fractionl / UINT_MAX);
368 sfp_to_d(s_fixedpt_t sfp)
371 sfp.int_parts = ntohs(sfp.int_parts);
372 sfp.fractions = ntohs(sfp.fractions);
373 ret = (double)sfp.int_parts + ((double)sfp.fractions / USHRT_MAX);
376 #if ENABLE_FEATURE_NTPD_SERVER
381 lfp.int_partl = (uint32_t)d;
382 lfp.fractionl = (uint32_t)((d - lfp.int_partl) * UINT_MAX);
383 lfp.int_partl = htonl(lfp.int_partl);
384 lfp.fractionl = htonl(lfp.fractionl);
391 sfp.int_parts = (uint16_t)d;
392 sfp.fractions = (uint16_t)((d - sfp.int_parts) * USHRT_MAX);
393 sfp.int_parts = htons(sfp.int_parts);
394 sfp.fractions = htons(sfp.fractions);
400 dispersion(const datapoint_t *dp)
402 return dp->d_dispersion + FREQ_TOLERANCE * (G.cur_time - dp->d_recv_time);
406 root_distance(peer_t *p)
408 /* The root synchronization distance is the maximum error due to
409 * all causes of the local clock relative to the primary server.
410 * It is defined as half the total delay plus total dispersion
413 return MAXD(MINDISP, p->lastpkt_rootdelay + p->lastpkt_delay) / 2
414 + p->lastpkt_rootdisp
415 + p->filter_dispersion
416 + FREQ_TOLERANCE * (G.cur_time - p->lastpkt_recv_time)
421 set_next(peer_t *p, unsigned t)
423 p->next_action_time = G.cur_time + t;
427 * Peer clock filter and its helpers
430 filter_datapoints(peer_t *p)
434 double minoff, maxoff, wavg, sum, w;
435 double x = x; /* for compiler */
436 double oldest_off = oldest_off;
437 double oldest_age = oldest_age;
438 double newest_off = newest_off;
439 double newest_age = newest_age;
441 minoff = maxoff = p->filter_datapoint[0].d_offset;
442 for (i = 1; i < NUM_DATAPOINTS; i++) {
443 if (minoff > p->filter_datapoint[i].d_offset)
444 minoff = p->filter_datapoint[i].d_offset;
445 if (maxoff < p->filter_datapoint[i].d_offset)
446 maxoff = p->filter_datapoint[i].d_offset;
449 idx = p->datapoint_idx; /* most recent datapoint */
451 * Drop two outliers and take weighted average of the rest:
452 * most_recent/2 + older1/4 + older2/8 ... + older5/32 + older6/32
453 * we use older6/32, not older6/64 since sum of weights should be 1:
454 * 1/2 + 1/4 + 1/8 + 1/16 + 1/32 + 1/32 = 1
460 * filter_dispersion = \ -------------
467 for (i = 0; i < NUM_DATAPOINTS; i++) {
469 bb_error_msg("datapoint[%d]: off:%f disp:%f(%f) age:%f%s",
471 p->filter_datapoint[idx].d_offset,
472 p->filter_datapoint[idx].d_dispersion, dispersion(&p->filter_datapoint[idx]),
473 G.cur_time - p->filter_datapoint[idx].d_recv_time,
474 (minoff == p->filter_datapoint[idx].d_offset || maxoff == p->filter_datapoint[idx].d_offset)
475 ? " (outlier by offset)" : ""
479 sum += dispersion(&p->filter_datapoint[idx]) / (2 << i);
481 if (minoff == p->filter_datapoint[idx].d_offset) {
482 minoff -= 1; /* so that we don't match it ever again */
484 if (maxoff == p->filter_datapoint[idx].d_offset) {
487 oldest_off = p->filter_datapoint[idx].d_offset;
488 oldest_age = G.cur_time - p->filter_datapoint[idx].d_recv_time;
491 newest_off = oldest_off;
492 newest_age = oldest_age;
499 idx = (idx - 1) & (NUM_DATAPOINTS - 1);
501 p->filter_dispersion = sum;
502 wavg += x; /* add another older6/64 to form older6/32 */
503 /* Fix systematic underestimation with large poll intervals.
504 * Imagine that we still have a bit of uncorrected drift,
505 * and poll interval is big (say, 100 sec). Offsets form a progression:
506 * 0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 - 0.7 is most recent.
507 * The algorithm above drops 0.0 and 0.7 as outliers,
508 * and then we have this estimation, ~25% off from 0.7:
509 * 0.1/32 + 0.2/32 + 0.3/16 + 0.4/8 + 0.5/4 + 0.6/2 = 0.503125
511 x = oldest_age - newest_age;
513 x = newest_age / x; /* in above example, 100 / (600 - 100) */
514 if (x < 1) { /* paranoia check */
515 x = (newest_off - oldest_off) * x; /* 0.5 * 100/500 = 0.1 */
519 p->filter_offset = wavg;
521 /* +----- -----+ ^ 1/2
525 * filter_jitter = | --- * / (avg-offset_j) |
529 * where n is the number of valid datapoints in the filter (n > 1);
530 * if filter_jitter < precision then filter_jitter = precision
533 for (i = 0; i < NUM_DATAPOINTS; i++) {
534 sum += SQUARE(wavg - p->filter_datapoint[i].d_offset);
536 sum = SQRT(sum / NUM_DATAPOINTS);
537 p->filter_jitter = sum > G_precision_sec ? sum : G_precision_sec;
539 VERB3 bb_error_msg("filter offset:%f(corr:%e) disp:%f jitter:%f",
541 p->filter_dispersion,
547 reset_peer_stats(peer_t *p, double offset)
550 for (i = 0; i < NUM_DATAPOINTS; i++) {
551 if (offset < 16 * STEP_THRESHOLD) {
552 p->filter_datapoint[i].d_recv_time -= offset;
553 if (p->filter_datapoint[i].d_offset != 0) {
554 p->filter_datapoint[i].d_offset -= offset;
557 p->filter_datapoint[i].d_recv_time = G.cur_time;
558 p->filter_datapoint[i].d_offset = 0;
559 p->filter_datapoint[i].d_dispersion = MAXDISP;
562 if (offset < 16 * STEP_THRESHOLD) {
563 p->lastpkt_recv_time -= offset;
565 p->reachable_bits = 0;
566 p->lastpkt_recv_time = G.cur_time;
568 filter_datapoints(p); /* recalc p->filter_xxx */
569 p->next_action_time -= offset;
570 VERB5 bb_error_msg("%s->lastpkt_recv_time=%f", p->p_dotted, p->lastpkt_recv_time);
578 p = xzalloc(sizeof(*p));
579 p->p_lsa = xhost2sockaddr(s, 123);
580 p->p_dotted = xmalloc_sockaddr2dotted_noport(&p->p_lsa->u.sa);
582 p->p_xmt_msg.m_status = MODE_CLIENT | (NTP_VERSION << 3);
583 p->next_action_time = G.cur_time; /* = set_next(p, 0); */
584 reset_peer_stats(p, 16 * STEP_THRESHOLD);
586 llist_add_to(&G.ntp_peers, p);
592 const struct sockaddr *from, const struct sockaddr *to, socklen_t addrlen,
593 msg_t *msg, ssize_t len)
599 ret = sendto(fd, msg, len, MSG_DONTWAIT, to, addrlen);
601 ret = send_to_from(fd, msg, len, MSG_DONTWAIT, to, from, addrlen);
604 bb_perror_msg("send failed");
611 send_query_to_peer(peer_t *p)
613 /* Why do we need to bind()?
614 * See what happens when we don't bind:
616 * socket(PF_INET, SOCK_DGRAM, IPPROTO_IP) = 3
617 * setsockopt(3, SOL_IP, IP_TOS, [16], 4) = 0
618 * gettimeofday({1259071266, 327885}, NULL) = 0
619 * sendto(3, "xxx", 48, MSG_DONTWAIT, {sa_family=AF_INET, sin_port=htons(123), sin_addr=inet_addr("10.34.32.125")}, 16) = 48
620 * ^^^ we sent it from some source port picked by kernel.
621 * time(NULL) = 1259071266
622 * write(2, "ntpd: entering poll 15 secs\n", 28) = 28
623 * poll([{fd=3, events=POLLIN}], 1, 15000) = 1 ([{fd=3, revents=POLLIN}])
624 * recv(3, "yyy", 68, MSG_DONTWAIT) = 48
625 * ^^^ this recv will receive packets to any local port!
627 * Uncomment this and use strace to see it in action:
629 #define PROBE_LOCAL_ADDR /* { len_and_sockaddr lsa; lsa.len = LSA_SIZEOF_SA; getsockname(p->query.fd, &lsa.u.sa, &lsa.len); } */
633 len_and_sockaddr *local_lsa;
635 family = p->p_lsa->u.sa.sa_family;
636 p->p_fd = fd = xsocket_type(&local_lsa, family, SOCK_DGRAM);
637 /* local_lsa has "null" address and port 0 now.
638 * bind() ensures we have a *particular port* selected by kernel
639 * and remembered in p->p_fd, thus later recv(p->p_fd)
640 * receives only packets sent to this port.
643 xbind(fd, &local_lsa->u.sa, local_lsa->len);
645 #if ENABLE_FEATURE_IPV6
646 if (family == AF_INET)
648 setsockopt(fd, IPPROTO_IP, IP_TOS, &const_IPTOS_LOWDELAY, sizeof(const_IPTOS_LOWDELAY));
653 * Send out a random 64-bit number as our transmit time. The NTP
654 * server will copy said number into the originate field on the
655 * response that it sends us. This is totally legal per the SNTP spec.
657 * The impact of this is two fold: we no longer send out the current
658 * system time for the world to see (which may aid an attacker), and
659 * it gives us a (not very secure) way of knowing that we're not
660 * getting spoofed by an attacker that can't capture our traffic
661 * but can spoof packets from the NTP server we're communicating with.
663 * Save the real transmit timestamp locally.
665 p->p_xmt_msg.m_xmttime.int_partl = random();
666 p->p_xmt_msg.m_xmttime.fractionl = random();
667 p->p_xmttime = gettime1900d();
669 if (do_sendto(p->p_fd, /*from:*/ NULL, /*to:*/ &p->p_lsa->u.sa, /*addrlen:*/ p->p_lsa->len,
670 &p->p_xmt_msg, NTP_MSGSIZE_NOAUTH) == -1
674 set_next(p, RETRY_INTERVAL);
678 p->reachable_bits <<= 1;
679 VERB1 bb_error_msg("sent query to %s", p->p_dotted);
680 set_next(p, RESPONSE_INTERVAL);
685 step_time(double offset)
693 gettimeofday(&tv, NULL); /* never fails */
694 dtime = offset + tv.tv_sec;
695 dtime += 1.0e-6 * tv.tv_usec;
698 if (settimeofday(&tv, NULL) == -1)
699 bb_perror_msg_and_die("settimeofday");
702 strftime(buf, sizeof(buf), "%a %b %e %H:%M:%S %Z %Y", localtime(&tval));
704 bb_error_msg("setting clock to %s (offset %fs)", buf, offset);
706 /* Correct various fields which contain time-relative values: */
708 /* p->lastpkt_recv_time, p->next_action_time and such: */
709 for (item = G.ntp_peers; item != NULL; item = item->link) {
710 peer_t *pp = (peer_t *) item->data;
711 reset_peer_stats(pp, offset);
714 G.cur_time -= offset;
715 G.last_update_recv_time -= offset;
720 * Selection and clustering, and their helpers
728 compare_point_edge(const void *aa, const void *bb)
730 const point_t *a = aa;
731 const point_t *b = bb;
732 if (a->edge < b->edge) {
735 return (a->edge > b->edge);
742 compare_survivor_metric(const void *aa, const void *bb)
744 const survivor_t *a = aa;
745 const survivor_t *b = bb;
746 if (a->metric < b->metric) {
749 return (a->metric > b->metric);
752 fit(peer_t *p, double rd)
754 if ((p->reachable_bits & (p->reachable_bits-1)) == 0) {
755 /* One or zero bits in reachable_bits */
756 VERB3 bb_error_msg("peer %s unfit for selection: unreachable", p->p_dotted);
759 #if 0 /* we filter out such packets earlier */
760 if ((p->lastpkt_status & LI_ALARM) == LI_ALARM
761 || p->lastpkt_stratum >= MAXSTRAT
763 VERB3 bb_error_msg("peer %s unfit for selection: bad status/stratum", p->p_dotted);
767 /* rd is root_distance(p) */
768 if (rd > MAXDIST + FREQ_TOLERANCE * (1 << G.poll_exp)) {
769 VERB3 bb_error_msg("peer %s unfit for selection: root distance too high", p->p_dotted);
773 // /* Do we have a loop? */
774 // if (p->refid == p->dstaddr || p->refid == s.refid)
779 select_and_cluster(void)
783 int size = 3 * G.peer_cnt;
784 /* for selection algorithm */
786 unsigned num_points, num_candidates;
788 unsigned num_falsetickers;
789 /* for cluster algorithm */
790 survivor_t survivor[size];
791 unsigned num_survivors;
797 if (G.initial_poll_complete) while (item != NULL) {
798 peer_t *p = (peer_t *) item->data;
799 double rd = root_distance(p);
800 double offset = p->filter_offset;
807 VERB4 bb_error_msg("interval: [%f %f %f] %s",
813 point[num_points].p = p;
814 point[num_points].type = -1;
815 point[num_points].edge = offset - rd;
817 point[num_points].p = p;
818 point[num_points].type = 0;
819 point[num_points].edge = offset;
821 point[num_points].p = p;
822 point[num_points].type = 1;
823 point[num_points].edge = offset + rd;
827 num_candidates = num_points / 3;
828 if (num_candidates == 0) {
829 VERB3 bb_error_msg("no valid datapoints, no peer selected");
832 //TODO: sorting does not seem to be done in reference code
833 qsort(point, num_points, sizeof(point[0]), compare_point_edge);
835 /* Start with the assumption that there are no falsetickers.
836 * Attempt to find a nonempty intersection interval containing
837 * the midpoints of all truechimers.
838 * If a nonempty interval cannot be found, increase the number
839 * of assumed falsetickers by one and try again.
840 * If a nonempty interval is found and the number of falsetickers
841 * is less than the number of truechimers, a majority has been found
842 * and the midpoint of each truechimer represents
843 * the candidates available to the cluster algorithm.
845 num_falsetickers = 0;
848 unsigned num_midpoints = 0;
853 for (i = 0; i < num_points; i++) {
855 * if (point[i].type == -1) c++;
856 * if (point[i].type == 1) c--;
857 * and it's simpler to do it this way:
860 if (c >= num_candidates - num_falsetickers) {
861 /* If it was c++ and it got big enough... */
865 if (point[i].type == 0)
869 for (i = num_points-1; i >= 0; i--) {
871 if (c >= num_candidates - num_falsetickers) {
872 high = point[i].edge;
875 if (point[i].type == 0)
878 /* If the number of midpoints is greater than the number
879 * of allowed falsetickers, the intersection contains at
880 * least one truechimer with no midpoint - bad.
881 * Also, interval should be nonempty.
883 if (num_midpoints <= num_falsetickers && low < high)
886 if (num_falsetickers * 2 >= num_candidates) {
887 VERB3 bb_error_msg("too many falsetickers:%d (candidates:%d), no peer selected",
888 num_falsetickers, num_candidates);
892 VERB3 bb_error_msg("selected interval: [%f, %f]; candidates:%d falsetickers:%d",
893 low, high, num_candidates, num_falsetickers);
897 /* Construct a list of survivors (p, metric)
898 * from the chime list, where metric is dominated
899 * first by stratum and then by root distance.
900 * All other things being equal, this is the order of preference.
903 for (i = 0; i < num_points; i++) {
906 if (point[i].edge < low || point[i].edge > high)
909 survivor[num_survivors].p = p;
910 //TODO: save root_distance in point_t and reuse here?
911 survivor[num_survivors].metric = MAXDIST * p->lastpkt_stratum + root_distance(p);
912 VERB4 bb_error_msg("survivor[%d] metric:%f peer:%s",
913 num_survivors, survivor[num_survivors].metric, p->p_dotted);
916 /* There must be at least MIN_SELECTED survivors to satisfy the
917 * correctness assertions. Ordinarily, the Byzantine criteria
918 * require four survivors, but for the demonstration here, one
921 if (num_survivors < MIN_SELECTED) {
922 VERB3 bb_error_msg("num_survivors %d < %d, no peer selected",
923 num_survivors, MIN_SELECTED);
927 //looks like this is ONLY used by the fact that later we pick survivor[0].
928 //we can avoid sorting then, just find the minimum once!
929 qsort(survivor, num_survivors, sizeof(survivor[0]), compare_survivor_metric);
931 /* For each association p in turn, calculate the selection
932 * jitter p->sjitter as the square root of the sum of squares
933 * (p->offset - q->offset) over all q associations. The idea is
934 * to repeatedly discard the survivor with maximum selection
935 * jitter until a termination condition is met.
938 unsigned max_idx = max_idx;
939 double max_selection_jitter = max_selection_jitter;
940 double min_jitter = min_jitter;
942 if (num_survivors <= MIN_CLUSTERED) {
943 bb_error_msg("num_survivors %d <= %d, not discarding more",
944 num_survivors, MIN_CLUSTERED);
948 /* To make sure a few survivors are left
949 * for the clustering algorithm to chew on,
950 * we stop if the number of survivors
951 * is less than or equal to MIN_CLUSTERED (3).
953 for (i = 0; i < num_survivors; i++) {
954 double selection_jitter_sq;
955 peer_t *p = survivor[i].p;
957 if (i == 0 || p->filter_jitter < min_jitter)
958 min_jitter = p->filter_jitter;
960 selection_jitter_sq = 0;
961 for (j = 0; j < num_survivors; j++) {
962 peer_t *q = survivor[j].p;
963 selection_jitter_sq += SQUARE(p->filter_offset - q->filter_offset);
965 if (i == 0 || selection_jitter_sq > max_selection_jitter) {
966 max_selection_jitter = selection_jitter_sq;
969 VERB5 bb_error_msg("survivor %d selection_jitter^2:%f",
970 i, selection_jitter_sq);
972 max_selection_jitter = SQRT(max_selection_jitter / num_survivors);
973 VERB4 bb_error_msg("max_selection_jitter (at %d):%f min_jitter:%f",
974 max_idx, max_selection_jitter, min_jitter);
976 /* If the maximum selection jitter is less than the
977 * minimum peer jitter, then tossing out more survivors
978 * will not lower the minimum peer jitter, so we might
981 if (max_selection_jitter < min_jitter) {
982 VERB3 bb_error_msg("max_selection_jitter:%f < min_jitter:%f, num_survivors:%d, not discarding more",
983 max_selection_jitter, min_jitter, num_survivors);
987 /* Delete survivor[max_idx] from the list
988 * and go around again.
990 VERB5 bb_error_msg("dropping survivor %d", max_idx);
992 while (max_idx < num_survivors) {
993 survivor[max_idx] = survivor[max_idx + 1];
998 /* Pick the best clock. If the old system peer is on the list
999 * and at the same stratum as the first survivor on the list,
1000 * then don't do a clock hop. Otherwise, select the first
1001 * survivor on the list as the new system peer.
1003 //TODO - see clock_combine()
1004 VERB3 bb_error_msg("selected peer %s filter_offset:%f age:%f",
1005 survivor[0].p->p_dotted,
1006 survivor[0].p->filter_offset,
1007 G.cur_time - survivor[0].p->lastpkt_recv_time
1009 return survivor[0].p;
1014 * Local clock discipline and its helpers
1017 set_new_values(int disc_state, double offset, double recv_time)
1019 /* Enter new state and set state variables. Note we use the time
1020 * of the last clock filter sample, which must be earlier than
1023 VERB3 bb_error_msg("disc_state=%d last update offset=%f recv_time=%f",
1024 disc_state, offset, recv_time);
1025 G.discipline_state = disc_state;
1026 G.last_update_offset = offset;
1027 G.last_update_recv_time = recv_time;
1029 /* Clock state definitions */
1030 #define STATE_NSET 0 /* initial state, "nothing is set" */
1031 #define STATE_FSET 1 /* frequency set from file */
1032 #define STATE_SPIK 2 /* spike detected */
1033 #define STATE_FREQ 3 /* initial frequency */
1034 #define STATE_SYNC 4 /* clock synchronized (normal operation) */
1035 /* Return: -1: decrease poll interval, 0: leave as is, 1: increase */
1037 update_local_clock(peer_t *p)
1040 long old_tmx_offset;
1042 double offset = p->filter_offset;
1043 double recv_time = p->lastpkt_recv_time;
1045 #if !USING_KERNEL_PLL_LOOP
1048 double since_last_update;
1049 double etemp, dtemp;
1051 abs_offset = fabs(offset);
1053 /* If the offset is too large, give up and go home */
1054 if (abs_offset > PANIC_THRESHOLD) {
1055 bb_error_msg_and_die("offset %f far too big, exiting", offset);
1058 /* If this is an old update, for instance as the result
1059 * of a system peer change, avoid it. We never use
1060 * an old sample or the same sample twice.
1062 if (recv_time <= G.last_update_recv_time) {
1063 VERB3 bb_error_msg("same or older datapoint: %f >= %f, not using it",
1064 G.last_update_recv_time, recv_time);
1065 return 0; /* "leave poll interval as is" */
1068 /* Clock state machine transition function. This is where the
1069 * action is and defines how the system reacts to large time
1070 * and frequency errors.
1072 since_last_update = recv_time - G.reftime;
1073 #if !USING_KERNEL_PLL_LOOP
1076 if (G.discipline_state == STATE_FREQ) {
1077 /* Ignore updates until the stepout threshold */
1078 if (since_last_update < WATCH_THRESHOLD) {
1079 VERB3 bb_error_msg("measuring drift, datapoint ignored, %f sec remains",
1080 WATCH_THRESHOLD - since_last_update);
1081 return 0; /* "leave poll interval as is" */
1083 #if !USING_KERNEL_PLL_LOOP
1084 freq_drift = (offset - G.last_update_offset) / since_last_update;
1088 /* There are two main regimes: when the
1089 * offset exceeds the step threshold and when it does not.
1091 if (abs_offset > STEP_THRESHOLD) {
1092 switch (G.discipline_state) {
1094 /* The first outlyer: ignore it, switch to SPIK state */
1095 VERB3 bb_error_msg("offset:%f - spike detected", offset);
1096 G.discipline_state = STATE_SPIK;
1097 return -1; /* "decrease poll interval" */
1100 /* Ignore succeeding outlyers until either an inlyer
1101 * is found or the stepout threshold is exceeded.
1103 if (since_last_update < WATCH_THRESHOLD) {
1104 VERB3 bb_error_msg("spike detected, datapoint ignored, %f sec remains",
1105 WATCH_THRESHOLD - since_last_update);
1106 return -1; /* "decrease poll interval" */
1108 /* fall through: we need to step */
1111 /* Step the time and clamp down the poll interval.
1113 * In NSET state an initial frequency correction is
1114 * not available, usually because the frequency file has
1115 * not yet been written. Since the time is outside the
1116 * capture range, the clock is stepped. The frequency
1117 * will be set directly following the stepout interval.
1119 * In FSET state the initial frequency has been set
1120 * from the frequency file. Since the time is outside
1121 * the capture range, the clock is stepped immediately,
1122 * rather than after the stepout interval. Guys get
1123 * nervous if it takes 17 minutes to set the clock for
1126 * In SPIK state the stepout threshold has expired and
1127 * the phase is still above the step threshold. Note
1128 * that a single spike greater than the step threshold
1129 * is always suppressed, even at the longer poll
1132 VERB3 bb_error_msg("stepping time by %f; poll_exp=MINPOLL", offset);
1134 if (option_mask32 & OPT_q) {
1135 /* We were only asked to set time once. Done. */
1139 G.polladj_count = 0;
1140 G.poll_exp = MINPOLL;
1141 G.stratum = MAXSTRAT;
1142 if (G.discipline_state == STATE_NSET) {
1143 set_new_values(STATE_FREQ, /*offset:*/ 0, recv_time);
1144 return 1; /* "ok to increase poll interval" */
1146 set_new_values(STATE_SYNC, /*offset:*/ 0, recv_time);
1148 } else { /* abs_offset <= STEP_THRESHOLD */
1150 if (G.poll_exp < MINPOLL && G.initial_poll_complete) {
1151 VERB3 bb_error_msg("small offset:%f, disabling burst mode", offset);
1152 G.polladj_count = 0;
1153 G.poll_exp = MINPOLL;
1156 /* Compute the clock jitter as the RMS of exponentially
1157 * weighted offset differences. Used by the poll adjust code.
1159 etemp = SQUARE(G.discipline_jitter);
1160 dtemp = SQUARE(MAXD(fabs(offset - G.last_update_offset), G_precision_sec));
1161 G.discipline_jitter = SQRT(etemp + (dtemp - etemp) / AVG);
1162 VERB3 bb_error_msg("discipline jitter=%f", G.discipline_jitter);
1164 switch (G.discipline_state) {
1166 if (option_mask32 & OPT_q) {
1167 /* We were only asked to set time once.
1168 * The clock is precise enough, no need to step.
1172 /* This is the first update received and the frequency
1173 * has not been initialized. The first thing to do
1174 * is directly measure the oscillator frequency.
1176 set_new_values(STATE_FREQ, offset, recv_time);
1177 VERB3 bb_error_msg("transitioning to FREQ, datapoint ignored");
1178 return 0; /* "leave poll interval as is" */
1180 #if 0 /* this is dead code for now */
1182 /* This is the first update and the frequency
1183 * has been initialized. Adjust the phase, but
1184 * don't adjust the frequency until the next update.
1186 set_new_values(STATE_SYNC, offset, recv_time);
1187 /* freq_drift remains 0 */
1192 /* since_last_update >= WATCH_THRESHOLD, we waited enough.
1193 * Correct the phase and frequency and switch to SYNC state.
1194 * freq_drift was already estimated (see code above)
1196 set_new_values(STATE_SYNC, offset, recv_time);
1200 #if !USING_KERNEL_PLL_LOOP
1201 /* Compute freq_drift due to PLL and FLL contributions.
1203 * The FLL and PLL frequency gain constants
1204 * depend on the poll interval and Allan
1205 * intercept. The FLL is not used below one-half
1206 * the Allan intercept. Above that the loop gain
1207 * increases in steps to 1 / AVG.
1209 if ((1 << G.poll_exp) > ALLAN / 2) {
1210 etemp = FLL - G.poll_exp;
1213 freq_drift += (offset - G.last_update_offset) / (MAXD(since_last_update, ALLAN) * etemp);
1215 /* For the PLL the integration interval
1216 * (numerator) is the minimum of the update
1217 * interval and poll interval. This allows
1218 * oversampling, but not undersampling.
1220 etemp = MIND(since_last_update, (1 << G.poll_exp));
1221 dtemp = (4 * PLL) << G.poll_exp;
1222 freq_drift += offset * etemp / SQUARE(dtemp);
1224 set_new_values(STATE_SYNC, offset, recv_time);
1227 G.stratum = p->lastpkt_stratum + 1;
1230 G.reftime = G.cur_time;
1231 G.ntp_status = p->lastpkt_status;
1232 G.refid = p->lastpkt_refid;
1233 G.rootdelay = p->lastpkt_rootdelay + p->lastpkt_delay;
1234 dtemp = p->filter_jitter; // SQRT(SQUARE(p->filter_jitter) + SQUARE(s.jitter));
1235 dtemp += MAXD(p->filter_dispersion + FREQ_TOLERANCE * (G.cur_time - p->lastpkt_recv_time) + abs_offset, MINDISP);
1236 G.rootdisp = p->lastpkt_rootdisp + dtemp;
1237 VERB3 bb_error_msg("updating leap/refid/reftime/rootdisp from peer %s", p->p_dotted);
1239 /* We are in STATE_SYNC now, but did not do adjtimex yet.
1240 * (Any other state does not reach this, they all return earlier)
1241 * By this time, freq_drift and G.last_update_offset are set
1242 * to values suitable for adjtimex.
1244 #if !USING_KERNEL_PLL_LOOP
1245 /* Calculate the new frequency drift and frequency stability (wander).
1246 * Compute the clock wander as the RMS of exponentially weighted
1247 * frequency differences. This is not used directly, but can,
1248 * along with the jitter, be a highly useful monitoring and
1251 dtemp = G.discipline_freq_drift + freq_drift;
1252 G.discipline_freq_drift = MAXD(MIND(MAXDRIFT, dtemp), -MAXDRIFT);
1253 etemp = SQUARE(G.discipline_wander);
1254 dtemp = SQUARE(dtemp);
1255 G.discipline_wander = SQRT(etemp + (dtemp - etemp) / AVG);
1257 VERB3 bb_error_msg("discipline freq_drift=%.9f(int:%ld corr:%e) wander=%f",
1258 G.discipline_freq_drift,
1259 (long)(G.discipline_freq_drift * 65536e6),
1261 G.discipline_wander);
1264 memset(&tmx, 0, sizeof(tmx));
1265 if (adjtimex(&tmx) < 0)
1266 bb_perror_msg_and_die("adjtimex");
1267 VERB3 bb_error_msg("p adjtimex freq:%ld offset:%ld constant:%ld status:0x%x",
1268 tmx.freq, tmx.offset, tmx.constant, tmx.status);
1272 if (!G.adjtimex_was_done) {
1273 G.adjtimex_was_done = 1;
1274 /* When we use adjtimex for the very first time,
1275 * we need to ADD to pre-existing tmx.offset - it may be !0
1277 memset(&tmx, 0, sizeof(tmx));
1278 if (adjtimex(&tmx) < 0)
1279 bb_perror_msg_and_die("adjtimex");
1280 old_tmx_offset = tmx.offset;
1282 memset(&tmx, 0, sizeof(tmx));
1284 //doesn't work, offset remains 0 (!) in kernel:
1285 //ntpd: set adjtimex freq:1786097 tmx.offset:77487
1286 //ntpd: prev adjtimex freq:1786097 tmx.offset:0
1287 //ntpd: cur adjtimex freq:1786097 tmx.offset:0
1288 tmx.modes = ADJ_FREQUENCY | ADJ_OFFSET;
1289 /* 65536 is one ppm */
1290 tmx.freq = G.discipline_freq_drift * 65536e6;
1291 tmx.offset = G.last_update_offset * 1000000; /* usec */
1293 tmx.modes = ADJ_OFFSET | ADJ_STATUS | ADJ_TIMECONST;// | ADJ_MAXERROR | ADJ_ESTERROR;
1294 tmx.offset = (G.last_update_offset * 1000000) /* usec */
1295 /* + (G.last_update_offset < 0 ? -0.5 : 0.5) - too small to bother */
1296 + old_tmx_offset; /* almost always 0 */
1297 tmx.status = STA_PLL;
1298 if (G.ntp_status & LI_PLUSSEC)
1299 tmx.status |= STA_INS;
1300 if (G.ntp_status & LI_MINUSSEC)
1301 tmx.status |= STA_DEL;
1302 tmx.constant = G.poll_exp - 4;
1303 //tmx.esterror = (u_int32)(clock_jitter * 1e6);
1304 //tmx.maxerror = (u_int32)((sys_rootdelay / 2 + sys_rootdisp) * 1e6);
1305 rc = adjtimex(&tmx);
1307 bb_perror_msg_and_die("adjtimex");
1308 /* NB: here kernel returns constant == G.poll_exp, not == G.poll_exp - 4.
1309 * Not sure why. Perhaps it is normal.
1311 VERB3 bb_error_msg("adjtimex:%d freq:%ld offset:%ld constant:%ld status:0x%x",
1312 rc, tmx.freq, tmx.offset, tmx.constant, tmx.status);
1315 /* always gives the same output as above msg */
1316 memset(&tmx, 0, sizeof(tmx));
1317 if (adjtimex(&tmx) < 0)
1318 bb_perror_msg_and_die("adjtimex");
1319 VERB3 bb_error_msg("c adjtimex freq:%ld offset:%ld constant:%ld status:0x%x",
1320 tmx.freq, tmx.offset, tmx.constant, tmx.status);
1323 if (G.kernel_freq_drift != tmx.freq / 65536) {
1324 G.kernel_freq_drift = tmx.freq / 65536;
1325 VERB2 bb_error_msg("kernel clock drift: %ld ppm", G.kernel_freq_drift);
1328 return 1; /* "ok to increase poll interval" */
1333 * We've got a new reply packet from a peer, process it
1337 retry_interval(void)
1339 /* Local problem, want to retry soon */
1340 unsigned interval, r;
1341 interval = RETRY_INTERVAL;
1343 interval += r % (unsigned)(RETRY_INTERVAL / 4);
1344 VERB3 bb_error_msg("chose retry interval:%u", interval);
1348 poll_interval(int exponent)
1350 unsigned interval, r;
1351 exponent = G.poll_exp + exponent;
1354 interval = 1 << exponent;
1356 interval += ((r & (interval-1)) >> 4) + ((r >> 8) & 1); /* + 1/16 of interval, max */
1357 VERB3 bb_error_msg("chose poll interval:%u (poll_exp:%d exp:%d)", interval, G.poll_exp, exponent);
1360 static NOINLINE void
1361 recv_and_process_peer_pkt(peer_t *p)
1366 double T1, T2, T3, T4;
1368 datapoint_t *datapoint;
1371 /* We can recvfrom here and check from.IP, but some multihomed
1372 * ntp servers reply from their *other IP*.
1373 * TODO: maybe we should check at least what we can: from.port == 123?
1375 size = recv(p->p_fd, &msg, sizeof(msg), MSG_DONTWAIT);
1377 bb_perror_msg("recv(%s) error", p->p_dotted);
1378 if (errno == EHOSTUNREACH || errno == EHOSTDOWN
1379 || errno == ENETUNREACH || errno == ENETDOWN
1380 || errno == ECONNREFUSED || errno == EADDRNOTAVAIL
1383 //TODO: always do this?
1384 set_next(p, retry_interval());
1390 if (size != NTP_MSGSIZE_NOAUTH && size != NTP_MSGSIZE) {
1391 bb_error_msg("malformed packet received from %s", p->p_dotted);
1395 if (msg.m_orgtime.int_partl != p->p_xmt_msg.m_xmttime.int_partl
1396 || msg.m_orgtime.fractionl != p->p_xmt_msg.m_xmttime.fractionl
1401 if ((msg.m_status & LI_ALARM) == LI_ALARM
1402 || msg.m_stratum == 0
1403 || msg.m_stratum > NTP_MAXSTRATUM
1405 // TODO: stratum 0 responses may have commands in 32-bit m_refid field:
1406 // "DENY", "RSTR" - peer does not like us at all
1407 // "RATE" - peer is overloaded, reduce polling freq
1408 interval = poll_interval(0);
1409 bb_error_msg("reply from %s: not synced, next query in %us", p->p_dotted, interval);
1413 // /* Verify valid root distance */
1414 // if (msg.m_rootdelay / 2 + msg.m_rootdisp >= MAXDISP || p->lastpkt_reftime > msg.m_xmt)
1415 // return; /* invalid header values */
1417 p->lastpkt_status = msg.m_status;
1418 p->lastpkt_stratum = msg.m_stratum;
1419 p->lastpkt_rootdelay = sfp_to_d(msg.m_rootdelay);
1420 p->lastpkt_rootdisp = sfp_to_d(msg.m_rootdisp);
1421 p->lastpkt_refid = msg.m_refid;
1424 * From RFC 2030 (with a correction to the delay math):
1426 * Timestamp Name ID When Generated
1427 * ------------------------------------------------------------
1428 * Originate Timestamp T1 time request sent by client
1429 * Receive Timestamp T2 time request received by server
1430 * Transmit Timestamp T3 time reply sent by server
1431 * Destination Timestamp T4 time reply received by client
1433 * The roundtrip delay and local clock offset are defined as
1435 * delay = (T4 - T1) - (T3 - T2); offset = ((T2 - T1) + (T3 - T4)) / 2
1438 T2 = lfp_to_d(msg.m_rectime);
1439 T3 = lfp_to_d(msg.m_xmttime);
1442 p->lastpkt_recv_time = T4;
1444 VERB5 bb_error_msg("%s->lastpkt_recv_time=%f", p->p_dotted, p->lastpkt_recv_time);
1445 p->datapoint_idx = p->reachable_bits ? (p->datapoint_idx + 1) % NUM_DATAPOINTS : 0;
1446 datapoint = &p->filter_datapoint[p->datapoint_idx];
1447 datapoint->d_recv_time = T4;
1448 datapoint->d_offset = ((T2 - T1) + (T3 - T4)) / 2;
1449 /* The delay calculation is a special case. In cases where the
1450 * server and client clocks are running at different rates and
1451 * with very fast networks, the delay can appear negative. In
1452 * order to avoid violating the Principle of Least Astonishment,
1453 * the delay is clamped not less than the system precision.
1455 p->lastpkt_delay = (T4 - T1) - (T3 - T2);
1456 if (p->lastpkt_delay < G_precision_sec)
1457 p->lastpkt_delay = G_precision_sec;
1458 datapoint->d_dispersion = LOG2D(msg.m_precision_exp) + G_precision_sec;
1459 if (!p->reachable_bits) {
1460 /* 1st datapoint ever - replicate offset in every element */
1462 for (i = 1; i < NUM_DATAPOINTS; i++) {
1463 p->filter_datapoint[i].d_offset = datapoint->d_offset;
1467 p->reachable_bits |= 1;
1469 bb_error_msg("reply from %s: reach 0x%02x offset %f delay %f",
1472 datapoint->d_offset, p->lastpkt_delay);
1475 /* Muck with statictics and update the clock */
1476 filter_datapoints(p);
1477 q = select_and_cluster();
1480 rc = update_local_clock(q);
1483 /* Adjust the poll interval by comparing the current offset
1484 * with the clock jitter. If the offset is less than
1485 * the clock jitter times a constant, then the averaging interval
1486 * is increased, otherwise it is decreased. A bit of hysteresis
1487 * helps calm the dance. Works best using burst mode.
1490 bb_error_msg("offset:%f POLLADJ_GATE*discipline_jitter:%f poll:%s",
1491 q->filter_offset, POLLADJ_GATE * G.discipline_jitter,
1492 fabs(q->filter_offset) < POLLADJ_GATE * G.discipline_jitter
1496 if (rc > 0 && fabs(q->filter_offset) < POLLADJ_GATE * G.discipline_jitter) {
1497 /* was += G.poll_exp but it is a bit
1498 * too optimistic for my taste at high poll_exp's */
1499 G.polladj_count += MINPOLL;
1500 if (G.polladj_count > POLLADJ_LIMIT) {
1501 G.polladj_count = 0;
1502 if (G.poll_exp < MAXPOLL) {
1504 VERB3 bb_error_msg("polladj: discipline_jitter:%f ++poll_exp=%d",
1505 G.discipline_jitter, G.poll_exp);
1508 VERB3 bb_error_msg("polladj: incr:%d", G.polladj_count);
1511 G.polladj_count -= G.poll_exp * 2;
1512 if (G.polladj_count < -POLLADJ_LIMIT) {
1513 G.polladj_count = 0;
1514 if (G.poll_exp > MINPOLL) {
1518 /* Correct p->next_action_time in each peer
1519 * which waits for sending, so that they send earlier.
1520 * Old pp->next_action_time are on the order
1521 * of t + (1 << old_poll_exp) + small_random,
1522 * we simply need to subtract ~half of that.
1524 for (item = G.ntp_peers; item != NULL; item = item->link) {
1525 peer_t *pp = (peer_t *) item->data;
1527 pp->next_action_time -= (1 << G.poll_exp);
1529 VERB3 bb_error_msg("polladj: discipline_jitter:%f --poll_exp=%d",
1530 G.discipline_jitter, G.poll_exp);
1533 VERB3 bb_error_msg("polladj: decr:%d", G.polladj_count);
1538 /* Decide when to send new query for this peer */
1539 interval = poll_interval(0);
1540 set_next(p, interval);
1543 /* We do not expect any more packets from this peer for now.
1544 * Closing the socket informs kernel about it.
1545 * We open a new socket when we send a new query.
1553 #if ENABLE_FEATURE_NTPD_SERVER
1554 static NOINLINE void
1555 recv_and_process_client_pkt(void /*int fd*/)
1559 len_and_sockaddr *to;
1560 struct sockaddr *from;
1562 uint8_t query_status;
1563 l_fixedpt_t query_xmttime;
1565 to = get_sock_lsa(G.listen_fd);
1566 from = xzalloc(to->len);
1568 size = recv_from_to(G.listen_fd, &msg, sizeof(msg), MSG_DONTWAIT, from, &to->u.sa, to->len);
1569 if (size != NTP_MSGSIZE_NOAUTH && size != NTP_MSGSIZE) {
1572 if (errno == EAGAIN)
1574 bb_perror_msg_and_die("recv");
1576 addr = xmalloc_sockaddr2dotted_noport(from);
1577 bb_error_msg("malformed packet received from %s: size %u", addr, (int)size);
1582 query_status = msg.m_status;
1583 query_xmttime = msg.m_xmttime;
1585 /* Build a reply packet */
1586 memset(&msg, 0, sizeof(msg));
1587 msg.m_status = G.stratum < MAXSTRAT ? G.ntp_status : LI_ALARM;
1588 msg.m_status |= (query_status & VERSION_MASK);
1589 msg.m_status |= ((query_status & MODE_MASK) == MODE_CLIENT) ?
1590 MODE_SERVER : MODE_SYM_PAS;
1591 msg.m_stratum = G.stratum;
1592 msg.m_ppoll = G.poll_exp;
1593 msg.m_precision_exp = G_precision_exp;
1594 /* this time was obtained between poll() and recv() */
1595 msg.m_rectime = d_to_lfp(G.cur_time);
1596 msg.m_xmttime = d_to_lfp(gettime1900d()); /* this instant */
1597 msg.m_reftime = d_to_lfp(G.reftime);
1598 msg.m_orgtime = query_xmttime;
1599 msg.m_rootdelay = d_to_sfp(G.rootdelay);
1600 //simple code does not do this, fix simple code!
1601 msg.m_rootdisp = d_to_sfp(G.rootdisp);
1602 version = (query_status & VERSION_MASK); /* ... >> VERSION_SHIFT - done below instead */
1603 msg.m_refid = G.refid; // (version > (3 << VERSION_SHIFT)) ? G.refid : G.refid3;
1605 /* We reply from the local address packet was sent to,
1606 * this makes to/from look swapped here: */
1607 do_sendto(G.listen_fd,
1608 /*from:*/ &to->u.sa, /*to:*/ from, /*addrlen:*/ to->len,
1617 /* Upstream ntpd's options:
1619 * -4 Force DNS resolution of host names to the IPv4 namespace.
1620 * -6 Force DNS resolution of host names to the IPv6 namespace.
1621 * -a Require cryptographic authentication for broadcast client,
1622 * multicast client and symmetric passive associations.
1623 * This is the default.
1624 * -A Do not require cryptographic authentication for broadcast client,
1625 * multicast client and symmetric passive associations.
1626 * This is almost never a good idea.
1627 * -b Enable the client to synchronize to broadcast servers.
1629 * Specify the name and path of the configuration file,
1630 * default /etc/ntp.conf
1631 * -d Specify debugging mode. This option may occur more than once,
1632 * with each occurrence indicating greater detail of display.
1634 * Specify debugging level directly.
1636 * Specify the name and path of the frequency file.
1637 * This is the same operation as the "driftfile FILE"
1638 * configuration command.
1639 * -g Normally, ntpd exits with a message to the system log
1640 * if the offset exceeds the panic threshold, which is 1000 s
1641 * by default. This option allows the time to be set to any value
1642 * without restriction; however, this can happen only once.
1643 * If the threshold is exceeded after that, ntpd will exit
1644 * with a message to the system log. This option can be used
1645 * with the -q and -x options. See the tinker command for other options.
1647 * Chroot the server to the directory jaildir. This option also implies
1648 * that the server attempts to drop root privileges at startup
1649 * (otherwise, chroot gives very little additional security).
1650 * You may need to also specify a -u option.
1652 * Specify the name and path of the symmetric key file,
1653 * default /etc/ntp/keys. This is the same operation
1654 * as the "keys FILE" configuration command.
1656 * Specify the name and path of the log file. The default
1657 * is the system log file. This is the same operation as
1658 * the "logfile FILE" configuration command.
1659 * -L Do not listen to virtual IPs. The default is to listen.
1661 * -N To the extent permitted by the operating system,
1662 * run the ntpd at the highest priority.
1664 * Specify the name and path of the file used to record the ntpd
1665 * process ID. This is the same operation as the "pidfile FILE"
1666 * configuration command.
1668 * To the extent permitted by the operating system,
1669 * run the ntpd at the specified priority.
1670 * -q Exit the ntpd just after the first time the clock is set.
1671 * This behavior mimics that of the ntpdate program, which is
1672 * to be retired. The -g and -x options can be used with this option.
1673 * Note: The kernel time discipline is disabled with this option.
1675 * Specify the default propagation delay from the broadcast/multicast
1676 * server to this client. This is necessary only if the delay
1677 * cannot be computed automatically by the protocol.
1679 * Specify the directory path for files created by the statistics
1680 * facility. This is the same operation as the "statsdir DIR"
1681 * configuration command.
1683 * Add a key number to the trusted key list. This option can occur
1686 * Specify a user, and optionally a group, to switch to.
1689 * Add a system variable listed by default.
1690 * -x Normally, the time is slewed if the offset is less than the step
1691 * threshold, which is 128 ms by default, and stepped if above
1692 * the threshold. This option sets the threshold to 600 s, which is
1693 * well within the accuracy window to set the clock manually.
1694 * Note: since the slew rate of typical Unix kernels is limited
1695 * to 0.5 ms/s, each second of adjustment requires an amortization
1696 * interval of 2000 s. Thus, an adjustment as much as 600 s
1697 * will take almost 14 days to complete. This option can be used
1698 * with the -g and -q options. See the tinker command for other options.
1699 * Note: The kernel time discipline is disabled with this option.
1702 /* By doing init in a separate function we decrease stack usage
1705 static NOINLINE void ntp_init(char **argv)
1713 bb_error_msg_and_die(bb_msg_you_must_be_root);
1715 /* Set some globals */
1716 G.stratum = MAXSTRAT;
1718 G.poll_exp = BURSTPOLL; /* speeds up initial sync */
1719 G.reftime = G.last_update_recv_time = gettime1900d(); /* sets G.cur_time too */
1723 opt_complementary = "dd:p::"; /* d: counter, p: list */
1724 opts = getopt32(argv,
1726 "p:"IF_FEATURE_NTPD_SERVER("l") /* NOT compat */
1728 "46aAbgL", /* compat, ignored */
1729 &peers, &G.verbose);
1730 if (!(opts & (OPT_p|OPT_l)))
1732 // if (opts & OPT_x) /* disable stepping, only slew is allowed */
1733 // G.time_was_stepped = 1;
1735 add_peers(llist_pop(&peers));
1736 if (!(opts & OPT_n)) {
1737 bb_daemonize_or_rexec(DAEMON_DEVNULL_STDIO, argv);
1738 logmode = LOGMODE_NONE;
1740 #if ENABLE_FEATURE_NTPD_SERVER
1743 G.listen_fd = create_and_bind_dgram_or_die(NULL, 123);
1744 socket_want_pktinfo(G.listen_fd);
1745 setsockopt(G.listen_fd, IPPROTO_IP, IP_TOS, &const_IPTOS_LOWDELAY, sizeof(const_IPTOS_LOWDELAY));
1748 /* I hesitate to set -20 prio. -15 should be high enough for timekeeping */
1750 setpriority(PRIO_PROCESS, 0, -15);
1752 bb_signals((1 << SIGTERM) | (1 << SIGINT), record_signo);
1753 bb_signals((1 << SIGPIPE) | (1 << SIGHUP), SIG_IGN);
1756 int ntpd_main(int argc UNUSED_PARAM, char **argv) MAIN_EXTERNALLY_VISIBLE;
1757 int ntpd_main(int argc UNUSED_PARAM, char **argv)
1765 memset(&G, 0, sizeof(G));
1766 SET_PTR_TO_GLOBALS(&G);
1770 /* If ENABLE_FEATURE_NTPD_SERVER, + 1 for listen_fd: */
1771 cnt = G.peer_cnt + ENABLE_FEATURE_NTPD_SERVER;
1772 idx2peer = xzalloc(sizeof(idx2peer[0]) * cnt);
1773 pfd = xzalloc(sizeof(pfd[0]) * cnt);
1775 /* Countdown: we never sync before we sent 5 packets to each peer
1776 * NB: if some peer is not responding, we may end up sending
1777 * fewer packets to it and more to other peers.
1778 * NB2: sync usually happens using 5-1=4 packets, since last reply
1779 * does not come back instantaneously.
1781 cnt = G.peer_cnt * 5;
1783 while (!bb_got_signal) {
1789 /* Nothing between here and poll() blocks for any significant time */
1791 nextaction = G.cur_time + 3600;
1794 #if ENABLE_FEATURE_NTPD_SERVER
1795 if (G.listen_fd != -1) {
1796 pfd[0].fd = G.listen_fd;
1797 pfd[0].events = POLLIN;
1801 /* Pass over peer list, send requests, time out on receives */
1802 for (item = G.ntp_peers; item != NULL; item = item->link) {
1803 peer_t *p = (peer_t *) item->data;
1805 if (p->next_action_time <= G.cur_time) {
1806 if (p->p_fd == -1) {
1807 /* Time to send new req */
1809 G.initial_poll_complete = 1;
1811 send_query_to_peer(p);
1813 /* Timed out waiting for reply */
1816 timeout = poll_interval(-2); /* -2: try a bit sooner */
1817 bb_error_msg("timed out waiting for %s, reach 0x%02x, next query in %us",
1818 p->p_dotted, p->reachable_bits, timeout);
1819 set_next(p, timeout);
1823 if (p->next_action_time < nextaction)
1824 nextaction = p->next_action_time;
1827 /* Wait for reply from this peer */
1828 pfd[i].fd = p->p_fd;
1829 pfd[i].events = POLLIN;
1835 timeout = nextaction - G.cur_time;
1838 timeout++; /* (nextaction - G.cur_time) rounds down, compensating */
1840 /* Here we may block */
1841 VERB2 bb_error_msg("poll %us, sockets:%u", timeout, i);
1842 nfds = poll(pfd, i, timeout * 1000);
1843 gettime1900d(); /* sets G.cur_time */
1847 /* Process any received packets */
1849 #if ENABLE_FEATURE_NTPD_SERVER
1850 if (G.listen_fd != -1) {
1851 if (pfd[0].revents /* & (POLLIN|POLLERR)*/) {
1853 recv_and_process_client_pkt(/*G.listen_fd*/);
1854 gettime1900d(); /* sets G.cur_time */
1859 for (; nfds != 0 && j < i; j++) {
1860 if (pfd[j].revents /* & (POLLIN|POLLERR)*/) {
1862 recv_and_process_peer_pkt(idx2peer[j]);
1863 gettime1900d(); /* sets G.cur_time */
1866 } /* while (!bb_got_signal) */
1868 kill_myself_with_sig(bb_got_signal);
1876 /*** openntpd-4.6 uses only adjtime, not adjtimex ***/
1878 /*** ntp-4.2.6/ntpd/ntp_loopfilter.c - adjtimex usage ***/
1882 direct_freq(double fp_offset)
1887 * If the kernel is enabled, we need the residual offset to
1888 * calculate the frequency correction.
1890 if (pll_control && kern_enable) {
1891 memset(&ntv, 0, sizeof(ntv));
1894 clock_offset = ntv.offset / 1e9;
1895 #else /* STA_NANO */
1896 clock_offset = ntv.offset / 1e6;
1897 #endif /* STA_NANO */
1898 drift_comp = FREQTOD(ntv.freq);
1900 #endif /* KERNEL_PLL */
1901 set_freq((fp_offset - clock_offset) / (current_time - clock_epoch) + drift_comp);
1907 set_freq(double freq) /* frequency update */
1915 * If the kernel is enabled, update the kernel frequency.
1917 if (pll_control && kern_enable) {
1918 memset(&ntv, 0, sizeof(ntv));
1919 ntv.modes = MOD_FREQUENCY;
1920 ntv.freq = DTOFREQ(drift_comp);
1922 snprintf(tbuf, sizeof(tbuf), "kernel %.3f PPM", drift_comp * 1e6);
1923 report_event(EVNT_FSET, NULL, tbuf);
1925 snprintf(tbuf, sizeof(tbuf), "ntpd %.3f PPM", drift_comp * 1e6);
1926 report_event(EVNT_FSET, NULL, tbuf);
1928 #else /* KERNEL_PLL */
1929 snprintf(tbuf, sizeof(tbuf), "ntpd %.3f PPM", drift_comp * 1e6);
1930 report_event(EVNT_FSET, NULL, tbuf);
1931 #endif /* KERNEL_PLL */
1940 * This code segment works when clock adjustments are made using
1941 * precision time kernel support and the ntp_adjtime() system
1942 * call. This support is available in Solaris 2.6 and later,
1943 * Digital Unix 4.0 and later, FreeBSD, Linux and specially
1944 * modified kernels for HP-UX 9 and Ultrix 4. In the case of the
1945 * DECstation 5000/240 and Alpha AXP, additional kernel
1946 * modifications provide a true microsecond clock and nanosecond
1947 * clock, respectively.
1949 * Important note: The kernel discipline is used only if the
1950 * step threshold is less than 0.5 s, as anything higher can
1951 * lead to overflow problems. This might occur if some misguided
1952 * lad set the step threshold to something ridiculous.
1954 if (pll_control && kern_enable) {
1956 #define MOD_BITS (MOD_OFFSET | MOD_MAXERROR | MOD_ESTERROR | MOD_STATUS | MOD_TIMECONST)
1959 * We initialize the structure for the ntp_adjtime()
1960 * system call. We have to convert everything to
1961 * microseconds or nanoseconds first. Do not update the
1962 * system variables if the ext_enable flag is set. In
1963 * this case, the external clock driver will update the
1964 * variables, which will be read later by the local
1965 * clock driver. Afterwards, remember the time and
1966 * frequency offsets for jitter and stability values and
1967 * to update the frequency file.
1969 memset(&ntv, 0, sizeof(ntv));
1971 ntv.modes = MOD_STATUS;
1974 ntv.modes = MOD_BITS | MOD_NANO;
1975 #else /* STA_NANO */
1976 ntv.modes = MOD_BITS;
1977 #endif /* STA_NANO */
1978 if (clock_offset < 0)
1983 ntv.offset = (int32)(clock_offset * 1e9 + dtemp);
1984 ntv.constant = sys_poll;
1985 #else /* STA_NANO */
1986 ntv.offset = (int32)(clock_offset * 1e6 + dtemp);
1987 ntv.constant = sys_poll - 4;
1988 #endif /* STA_NANO */
1989 ntv.esterror = (u_int32)(clock_jitter * 1e6);
1990 ntv.maxerror = (u_int32)((sys_rootdelay / 2 + sys_rootdisp) * 1e6);
1991 ntv.status = STA_PLL;
1994 * Enable/disable the PPS if requested.
1997 if (!(pll_status & STA_PPSTIME))
1998 report_event(EVNT_KERN,
1999 NULL, "PPS enabled");
2000 ntv.status |= STA_PPSTIME | STA_PPSFREQ;
2002 if (pll_status & STA_PPSTIME)
2003 report_event(EVNT_KERN,
2004 NULL, "PPS disabled");
2005 ntv.status &= ~(STA_PPSTIME |
2008 if (sys_leap == LEAP_ADDSECOND)
2009 ntv.status |= STA_INS;
2010 else if (sys_leap == LEAP_DELSECOND)
2011 ntv.status |= STA_DEL;
2015 * Pass the stuff to the kernel. If it squeals, turn off
2016 * the pps. In any case, fetch the kernel offset,
2017 * frequency and jitter.
2019 if (ntp_adjtime(&ntv) == TIME_ERROR) {
2020 if (!(ntv.status & STA_PPSSIGNAL))
2021 report_event(EVNT_KERN, NULL,
2024 pll_status = ntv.status;
2026 clock_offset = ntv.offset / 1e9;
2027 #else /* STA_NANO */
2028 clock_offset = ntv.offset / 1e6;
2029 #endif /* STA_NANO */
2030 clock_frequency = FREQTOD(ntv.freq);
2033 * If the kernel PPS is lit, monitor its performance.
2035 if (ntv.status & STA_PPSTIME) {
2037 clock_jitter = ntv.jitter / 1e9;
2038 #else /* STA_NANO */
2039 clock_jitter = ntv.jitter / 1e6;
2040 #endif /* STA_NANO */
2043 #if defined(STA_NANO) && NTP_API == 4
2045 * If the TAI changes, update the kernel TAI.
2047 if (loop_tai != sys_tai) {
2049 ntv.modes = MOD_TAI;
2050 ntv.constant = sys_tai;
2053 #endif /* STA_NANO */
2055 #endif /* KERNEL_PLL */