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 /* High-level description of the algorithm:
51 * We start running with very small poll_exp, BURSTPOLL,
52 * in order to quickly accumulate INITIAL_SAMLPES datapoints
53 * for each peer. Then, time is stepped if the offset is larger
54 * than STEP_THRESHOLD, otherwise it isn't; anyway, we enlarge
55 * poll_exp to MINPOLL and enter frequency measurement step:
56 * we collect new datapoints but ignore them for WATCH_THRESHOLD
57 * seconds. After WATCH_THRESHOLD seconds we look at accumulated
58 * offset and estimate frequency drift.
60 * After this, we enter "steady state": we collect a datapoint,
61 * we select the best peer, if this datapoint is not a new one
62 * (IOW: if this datapoint isn't for selected peer), sleep
63 * and collect another one; otherwise, use its offset to update
64 * frequency drift, if offset is somewhat large, reduce poll_exp,
65 * otherwise increase poll_exp.
67 * If offset is larger than STEP_THRESHOLD, which shouldn't normally
68 * happen, we assume that something "bad" happened (computer
69 * was hibernated, someone set totally wrong date, etc),
70 * then the time is stepped, all datapoints are discarded,
71 * and we go back to steady state.
74 #define RETRY_INTERVAL 5 /* on error, retry in N secs */
75 #define RESPONSE_INTERVAL 15 /* wait for reply up to N secs */
76 #define INITIAL_SAMLPES 4 /* how many samples do we want for init */
78 /* Clock discipline parameters and constants */
79 #define STEP_THRESHOLD 0.128 /* step threshold (s) */
80 #define WATCH_THRESHOLD 150 /* stepout threshold (s). std ntpd uses 900 (11 mins (!)) */
81 /* NB: set WATCH_THRESHOLD to ~60 when debugging to save time) */
82 //UNUSED: #define PANIC_THRESHOLD 1000 /* panic threshold (s) */
84 #define FREQ_TOLERANCE 0.000015 /* frequency tolerance (15 PPM) */
85 #define BURSTPOLL 0 /* initial poll */
86 #define MINPOLL 4 /* minimum poll interval (6: 64 s) */
87 #define BIGPOLL 10 /* drop to lower poll at any trouble (10: 17 min) */
88 #define MAXPOLL 12 /* maximum poll interval (12: 1.1h, 17: 36.4h) (was 17) */
89 #define POLLDOWN_OFFSET (STEP_THRESHOLD / 3) /* actively lower poll when we see such big offsets */
90 #define MINDISP 0.01 /* minimum dispersion (s) */
91 #define MAXDISP 16 /* maximum dispersion (s) */
92 #define MAXSTRAT 16 /* maximum stratum (infinity metric) */
93 #define MAXDIST 1 /* distance threshold (s) */
94 #define MIN_SELECTED 1 /* minimum intersection survivors */
95 #define MIN_CLUSTERED 3 /* minimum cluster survivors */
97 #define MAXDRIFT 0.000500 /* frequency drift we can correct (500 PPM) */
99 /* Poll-adjust threshold.
100 * When we see that offset is small enough compared to discipline jitter,
101 * we grow a counter: += MINPOLL. When it goes over POLLADJ_LIMIT,
102 * we poll_exp++. If offset isn't small, counter -= poll_exp*2,
103 * and when it goes below -POLLADJ_LIMIT, we poll_exp--
104 * (bumped from 30 to 36 since otherwise I often see poll_exp going *2* steps down)
106 #define POLLADJ_LIMIT 36
107 /* If offset < POLLADJ_GATE * discipline_jitter, then we can increase
108 * poll interval (we think we can't improve timekeeping
109 * by staying at smaller poll).
111 #define POLLADJ_GATE 4
112 /* Compromise Allan intercept (s). doc uses 1500, std ntpd uses 512 */
116 /* FLL loop gain [why it depends on MAXPOLL??] */
117 #define FLL (MAXPOLL + 1)
118 /* Parameter averaging constant */
127 NTP_MSGSIZE_NOAUTH = 48,
128 NTP_MSGSIZE = (NTP_MSGSIZE_NOAUTH + 4 + NTP_DIGESTSIZE),
131 MODE_MASK = (7 << 0),
132 VERSION_MASK = (7 << 3),
136 /* Leap Second Codes (high order two bits of m_status) */
137 LI_NOWARNING = (0 << 6), /* no warning */
138 LI_PLUSSEC = (1 << 6), /* add a second (61 seconds) */
139 LI_MINUSSEC = (2 << 6), /* minus a second (59 seconds) */
140 LI_ALARM = (3 << 6), /* alarm condition */
143 MODE_RES0 = 0, /* reserved */
144 MODE_SYM_ACT = 1, /* symmetric active */
145 MODE_SYM_PAS = 2, /* symmetric passive */
146 MODE_CLIENT = 3, /* client */
147 MODE_SERVER = 4, /* server */
148 MODE_BROADCAST = 5, /* broadcast */
149 MODE_RES1 = 6, /* reserved for NTP control message */
150 MODE_RES2 = 7, /* reserved for private use */
153 //TODO: better base selection
154 #define OFFSET_1900_1970 2208988800UL /* 1970 - 1900 in seconds */
156 #define NUM_DATAPOINTS 8
169 uint8_t m_status; /* status of local clock and leap info */
171 uint8_t m_ppoll; /* poll value */
172 int8_t m_precision_exp;
173 s_fixedpt_t m_rootdelay;
174 s_fixedpt_t m_rootdisp;
176 l_fixedpt_t m_reftime;
177 l_fixedpt_t m_orgtime;
178 l_fixedpt_t m_rectime;
179 l_fixedpt_t m_xmttime;
181 uint8_t m_digest[NTP_DIGESTSIZE];
191 len_and_sockaddr *p_lsa;
193 /* when to send new query (if p_fd == -1)
194 * or when receive times out (if p_fd >= 0): */
197 uint32_t lastpkt_refid;
198 uint8_t lastpkt_status;
199 uint8_t lastpkt_stratum;
200 uint8_t reachable_bits;
201 double next_action_time;
203 double lastpkt_recv_time;
204 double lastpkt_delay;
205 double lastpkt_rootdelay;
206 double lastpkt_rootdisp;
207 /* produced by filter algorithm: */
208 double filter_offset;
209 double filter_dispersion;
210 double filter_jitter;
211 datapoint_t filter_datapoint[NUM_DATAPOINTS];
212 /* last sent packet: */
222 /* Insert new options above this line. */
223 /* Non-compat options: */
227 OPT_l = (1 << 7) * ENABLE_FEATURE_NTPD_SERVER,
232 /* total round trip delay to currently selected reference clock */
234 /* reference timestamp: time when the system clock was last set or corrected */
236 /* total dispersion to currently selected reference clock */
239 double last_script_run;
242 #if ENABLE_FEATURE_NTPD_SERVER
247 /* refid: 32-bit code identifying the particular server or reference clock
248 * in stratum 0 packets this is a four-character ASCII string,
249 * called the kiss code, used for debugging and monitoring
250 * in stratum 1 packets this is a four-character ASCII string
251 * assigned to the reference clock by IANA. Example: "GPS "
252 * in stratum 2+ packets, it's IPv4 address or 4 first bytes of MD5 hash of IPv6
256 /* precision is defined as the larger of the resolution and time to
257 * read the clock, in log2 units. For instance, the precision of a
258 * mains-frequency clock incrementing at 60 Hz is 16 ms, even when the
259 * system clock hardware representation is to the nanosecond.
261 * Delays, jitters of various kinds are clamper down to precision.
263 * If precision_sec is too large, discipline_jitter gets clamped to it
264 * and if offset is much smaller than discipline_jitter, poll interval
265 * grows even though we really can benefit from staying at smaller one,
266 * collecting non-lagged datapoits and correcting the offset.
267 * (Lagged datapoits exist when poll_exp is large but we still have
268 * systematic offset error - the time distance between datapoints
269 * is significat and older datapoints have smaller offsets.
270 * This makes our offset estimation a bit smaller than reality)
271 * Due to this effect, setting G_precision_sec close to
272 * STEP_THRESHOLD isn't such a good idea - offsets may grow
273 * too big and we will step. I observed it with -6.
275 * OTOH, setting precision too small would result in futile attempts
276 * to syncronize to the unachievable precision.
278 * -6 is 1/64 sec, -7 is 1/128 sec and so on.
280 #define G_precision_exp -8
281 #define G_precision_sec (1.0 / (1 << (- G_precision_exp)))
283 /* Bool. After set to 1, never goes back to 0: */
284 smallint adjtimex_was_done;
285 smallint initial_poll_complete;
287 uint8_t discipline_state; // doc calls it c.state
288 uint8_t poll_exp; // s.poll
289 int polladj_count; // c.count
290 long kernel_freq_drift;
291 double last_update_offset; // c.last
292 double last_update_recv_time; // s.t
293 double discipline_jitter; // c.jitter
294 //TODO: add s.jitter - grep for it here and see clock_combine() in doc
295 #define USING_KERNEL_PLL_LOOP 1
296 #if !USING_KERNEL_PLL_LOOP
297 double discipline_freq_drift; // c.freq
298 //TODO: conditionally calculate wander? it's used only for logging
299 double discipline_wander; // c.wander
302 #define G (*ptr_to_globals)
304 static const int const_IPTOS_LOWDELAY = IPTOS_LOWDELAY;
307 #define VERB1 if (MAX_VERBOSE && G.verbose)
308 #define VERB2 if (MAX_VERBOSE >= 2 && G.verbose >= 2)
309 #define VERB3 if (MAX_VERBOSE >= 3 && G.verbose >= 3)
310 #define VERB4 if (MAX_VERBOSE >= 4 && G.verbose >= 4)
311 #define VERB5 if (MAX_VERBOSE >= 5 && G.verbose >= 5)
314 static double LOG2D(int a)
317 return 1.0 / (1UL << -a);
320 static ALWAYS_INLINE double SQUARE(double x)
324 static ALWAYS_INLINE double MAXD(double a, double b)
330 static ALWAYS_INLINE double MIND(double a, double b)
336 static NOINLINE double my_SQRT(double X)
343 double Xhalf = X * 0.5;
345 /* Fast and good approximation to 1/sqrt(X), black magic */
347 /*v.i = 0x5f3759df - (v.i >> 1);*/
348 v.i = 0x5f375a86 - (v.i >> 1); /* - this constant is slightly better */
349 invsqrt = v.f; /* better than 0.2% accuracy */
351 /* Refining it using Newton's method: x1 = x0 - f(x0)/f'(x0)
352 * f(x) = 1/(x*x) - X (f==0 when x = 1/sqrt(X))
354 * f(x)/f'(x) = (X - 1/(x*x)) / (2/(x*x*x)) = X*x*x*x/2 - x/2
355 * x1 = x0 - (X*x0*x0*x0/2 - x0/2) = 1.5*x0 - X*x0*x0*x0/2 = x0*(1.5 - (X/2)*x0*x0)
357 invsqrt = invsqrt * (1.5 - Xhalf * invsqrt * invsqrt); /* ~0.05% accuracy */
358 /* invsqrt = invsqrt * (1.5 - Xhalf * invsqrt * invsqrt); 2nd iter: ~0.0001% accuracy */
359 /* With 4 iterations, more than half results will be exact,
360 * at 6th iterations result stabilizes with about 72% results exact.
361 * We are well satisfied with 0.05% accuracy.
364 return X * invsqrt; /* X * 1/sqrt(X) ~= sqrt(X) */
366 static ALWAYS_INLINE double SQRT(double X)
368 /* If this arch doesn't use IEEE 754 floats, fall back to using libm */
369 if (sizeof(float) != 4)
372 /* This avoids needing libm, saves about 0.5k on x86-32 */
380 gettimeofday(&tv, NULL); /* never fails */
381 G.cur_time = tv.tv_sec + (1.0e-6 * tv.tv_usec) + OFFSET_1900_1970;
386 d_to_tv(double d, struct timeval *tv)
388 tv->tv_sec = (long)d;
389 tv->tv_usec = (d - tv->tv_sec) * 1000000;
393 lfp_to_d(l_fixedpt_t lfp)
396 lfp.int_partl = ntohl(lfp.int_partl);
397 lfp.fractionl = ntohl(lfp.fractionl);
398 ret = (double)lfp.int_partl + ((double)lfp.fractionl / UINT_MAX);
402 sfp_to_d(s_fixedpt_t sfp)
405 sfp.int_parts = ntohs(sfp.int_parts);
406 sfp.fractions = ntohs(sfp.fractions);
407 ret = (double)sfp.int_parts + ((double)sfp.fractions / USHRT_MAX);
410 #if ENABLE_FEATURE_NTPD_SERVER
415 lfp.int_partl = (uint32_t)d;
416 lfp.fractionl = (uint32_t)((d - lfp.int_partl) * UINT_MAX);
417 lfp.int_partl = htonl(lfp.int_partl);
418 lfp.fractionl = htonl(lfp.fractionl);
425 sfp.int_parts = (uint16_t)d;
426 sfp.fractions = (uint16_t)((d - sfp.int_parts) * USHRT_MAX);
427 sfp.int_parts = htons(sfp.int_parts);
428 sfp.fractions = htons(sfp.fractions);
434 dispersion(const datapoint_t *dp)
436 return dp->d_dispersion + FREQ_TOLERANCE * (G.cur_time - dp->d_recv_time);
440 root_distance(peer_t *p)
442 /* The root synchronization distance is the maximum error due to
443 * all causes of the local clock relative to the primary server.
444 * It is defined as half the total delay plus total dispersion
447 return MAXD(MINDISP, p->lastpkt_rootdelay + p->lastpkt_delay) / 2
448 + p->lastpkt_rootdisp
449 + p->filter_dispersion
450 + FREQ_TOLERANCE * (G.cur_time - p->lastpkt_recv_time)
455 set_next(peer_t *p, unsigned t)
457 p->next_action_time = G.cur_time + t;
461 * Peer clock filter and its helpers
464 filter_datapoints(peer_t *p)
468 double minoff, maxoff, wavg, sum, w;
469 double x = x; /* for compiler */
470 double oldest_off = oldest_off;
471 double oldest_age = oldest_age;
472 double newest_off = newest_off;
473 double newest_age = newest_age;
475 minoff = maxoff = p->filter_datapoint[0].d_offset;
476 for (i = 1; i < NUM_DATAPOINTS; i++) {
477 if (minoff > p->filter_datapoint[i].d_offset)
478 minoff = p->filter_datapoint[i].d_offset;
479 if (maxoff < p->filter_datapoint[i].d_offset)
480 maxoff = p->filter_datapoint[i].d_offset;
483 idx = p->datapoint_idx; /* most recent datapoint */
485 * Drop two outliers and take weighted average of the rest:
486 * most_recent/2 + older1/4 + older2/8 ... + older5/32 + older6/32
487 * we use older6/32, not older6/64 since sum of weights should be 1:
488 * 1/2 + 1/4 + 1/8 + 1/16 + 1/32 + 1/32 = 1
494 * filter_dispersion = \ -------------
501 for (i = 0; i < NUM_DATAPOINTS; i++) {
503 bb_error_msg("datapoint[%d]: off:%f disp:%f(%f) age:%f%s",
505 p->filter_datapoint[idx].d_offset,
506 p->filter_datapoint[idx].d_dispersion, dispersion(&p->filter_datapoint[idx]),
507 G.cur_time - p->filter_datapoint[idx].d_recv_time,
508 (minoff == p->filter_datapoint[idx].d_offset || maxoff == p->filter_datapoint[idx].d_offset)
509 ? " (outlier by offset)" : ""
513 sum += dispersion(&p->filter_datapoint[idx]) / (2 << i);
515 if (minoff == p->filter_datapoint[idx].d_offset) {
516 minoff -= 1; /* so that we don't match it ever again */
518 if (maxoff == p->filter_datapoint[idx].d_offset) {
521 oldest_off = p->filter_datapoint[idx].d_offset;
522 oldest_age = G.cur_time - p->filter_datapoint[idx].d_recv_time;
525 newest_off = oldest_off;
526 newest_age = oldest_age;
533 idx = (idx - 1) & (NUM_DATAPOINTS - 1);
535 p->filter_dispersion = sum;
536 wavg += x; /* add another older6/64 to form older6/32 */
537 /* Fix systematic underestimation with large poll intervals.
538 * Imagine that we still have a bit of uncorrected drift,
539 * and poll interval is big (say, 100 sec). Offsets form a progression:
540 * 0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 - 0.7 is most recent.
541 * The algorithm above drops 0.0 and 0.7 as outliers,
542 * and then we have this estimation, ~25% off from 0.7:
543 * 0.1/32 + 0.2/32 + 0.3/16 + 0.4/8 + 0.5/4 + 0.6/2 = 0.503125
545 x = oldest_age - newest_age;
547 x = newest_age / x; /* in above example, 100 / (600 - 100) */
548 if (x < 1) { /* paranoia check */
549 x = (newest_off - oldest_off) * x; /* 0.5 * 100/500 = 0.1 */
553 p->filter_offset = wavg;
555 /* +----- -----+ ^ 1/2
559 * filter_jitter = | --- * / (avg-offset_j) |
563 * where n is the number of valid datapoints in the filter (n > 1);
564 * if filter_jitter < precision then filter_jitter = precision
567 for (i = 0; i < NUM_DATAPOINTS; i++) {
568 sum += SQUARE(wavg - p->filter_datapoint[i].d_offset);
570 sum = SQRT(sum / NUM_DATAPOINTS);
571 p->filter_jitter = sum > G_precision_sec ? sum : G_precision_sec;
573 VERB3 bb_error_msg("filter offset:%f(corr:%e) disp:%f jitter:%f",
575 p->filter_dispersion,
581 reset_peer_stats(peer_t *p, double offset)
584 for (i = 0; i < NUM_DATAPOINTS; i++) {
585 if (offset < 16 * STEP_THRESHOLD) {
586 p->filter_datapoint[i].d_recv_time -= offset;
587 if (p->filter_datapoint[i].d_offset != 0) {
588 p->filter_datapoint[i].d_offset -= offset;
591 p->filter_datapoint[i].d_recv_time = G.cur_time;
592 p->filter_datapoint[i].d_offset = 0;
593 p->filter_datapoint[i].d_dispersion = MAXDISP;
596 if (offset < 16 * STEP_THRESHOLD) {
597 p->lastpkt_recv_time -= offset;
599 p->reachable_bits = 0;
600 p->lastpkt_recv_time = G.cur_time;
602 filter_datapoints(p); /* recalc p->filter_xxx */
603 p->next_action_time -= offset;
604 VERB5 bb_error_msg("%s->lastpkt_recv_time=%f", p->p_dotted, p->lastpkt_recv_time);
612 p = xzalloc(sizeof(*p));
613 p->p_lsa = xhost2sockaddr(s, 123);
614 p->p_dotted = xmalloc_sockaddr2dotted_noport(&p->p_lsa->u.sa);
616 p->p_xmt_msg.m_status = MODE_CLIENT | (NTP_VERSION << 3);
617 p->next_action_time = G.cur_time; /* = set_next(p, 0); */
618 reset_peer_stats(p, 16 * STEP_THRESHOLD);
620 llist_add_to(&G.ntp_peers, p);
626 const struct sockaddr *from, const struct sockaddr *to, socklen_t addrlen,
627 msg_t *msg, ssize_t len)
633 ret = sendto(fd, msg, len, MSG_DONTWAIT, to, addrlen);
635 ret = send_to_from(fd, msg, len, MSG_DONTWAIT, to, from, addrlen);
638 bb_perror_msg("send failed");
645 send_query_to_peer(peer_t *p)
647 /* Why do we need to bind()?
648 * See what happens when we don't bind:
650 * socket(PF_INET, SOCK_DGRAM, IPPROTO_IP) = 3
651 * setsockopt(3, SOL_IP, IP_TOS, [16], 4) = 0
652 * gettimeofday({1259071266, 327885}, NULL) = 0
653 * sendto(3, "xxx", 48, MSG_DONTWAIT, {sa_family=AF_INET, sin_port=htons(123), sin_addr=inet_addr("10.34.32.125")}, 16) = 48
654 * ^^^ we sent it from some source port picked by kernel.
655 * time(NULL) = 1259071266
656 * write(2, "ntpd: entering poll 15 secs\n", 28) = 28
657 * poll([{fd=3, events=POLLIN}], 1, 15000) = 1 ([{fd=3, revents=POLLIN}])
658 * recv(3, "yyy", 68, MSG_DONTWAIT) = 48
659 * ^^^ this recv will receive packets to any local port!
661 * Uncomment this and use strace to see it in action:
663 #define PROBE_LOCAL_ADDR /* { len_and_sockaddr lsa; lsa.len = LSA_SIZEOF_SA; getsockname(p->query.fd, &lsa.u.sa, &lsa.len); } */
667 len_and_sockaddr *local_lsa;
669 family = p->p_lsa->u.sa.sa_family;
670 p->p_fd = fd = xsocket_type(&local_lsa, family, SOCK_DGRAM);
671 /* local_lsa has "null" address and port 0 now.
672 * bind() ensures we have a *particular port* selected by kernel
673 * and remembered in p->p_fd, thus later recv(p->p_fd)
674 * receives only packets sent to this port.
677 xbind(fd, &local_lsa->u.sa, local_lsa->len);
679 #if ENABLE_FEATURE_IPV6
680 if (family == AF_INET)
682 setsockopt(fd, IPPROTO_IP, IP_TOS, &const_IPTOS_LOWDELAY, sizeof(const_IPTOS_LOWDELAY));
687 * Send out a random 64-bit number as our transmit time. The NTP
688 * server will copy said number into the originate field on the
689 * response that it sends us. This is totally legal per the SNTP spec.
691 * The impact of this is two fold: we no longer send out the current
692 * system time for the world to see (which may aid an attacker), and
693 * it gives us a (not very secure) way of knowing that we're not
694 * getting spoofed by an attacker that can't capture our traffic
695 * but can spoof packets from the NTP server we're communicating with.
697 * Save the real transmit timestamp locally.
699 p->p_xmt_msg.m_xmttime.int_partl = random();
700 p->p_xmt_msg.m_xmttime.fractionl = random();
701 p->p_xmttime = gettime1900d();
703 if (do_sendto(p->p_fd, /*from:*/ NULL, /*to:*/ &p->p_lsa->u.sa, /*addrlen:*/ p->p_lsa->len,
704 &p->p_xmt_msg, NTP_MSGSIZE_NOAUTH) == -1
708 set_next(p, RETRY_INTERVAL);
712 p->reachable_bits <<= 1;
713 VERB1 bb_error_msg("sent query to %s", p->p_dotted);
714 set_next(p, RESPONSE_INTERVAL);
718 static void run_script(const char *action, double offset)
721 char *env1, *env2, *env3, *env4;
726 argv[0] = (char*) G.script_name;
727 argv[1] = (char*) action;
730 VERB1 bb_error_msg("executing '%s %s'", G.script_name, action);
732 env1 = xasprintf("%s=%u", "stratum", G.stratum);
734 env2 = xasprintf("%s=%ld", "freq_drift_ppm", G.kernel_freq_drift);
736 env3 = xasprintf("%s=%u", "poll_interval", 1 << G.poll_exp);
738 env4 = xasprintf("%s=%f", "offset", offset);
740 /* Other items of potential interest: selected peer,
741 * rootdelay, reftime, rootdisp, refid, ntp_status,
742 * last_update_offset, last_update_recv_time, discipline_jitter,
743 * how many peers have reachable_bits = 0?
746 /* Don't want to wait: it may run hwclock --systohc, and that
747 * may take some time (seconds): */
748 /*wait4pid(spawn(argv));*/
752 unsetenv("freq_drift_ppm");
753 unsetenv("poll_interval");
760 G.last_script_run = G.cur_time;
764 step_time(double offset)
772 gettimeofday(&tv, NULL); /* never fails */
773 dtime = offset + tv.tv_sec;
774 dtime += 1.0e-6 * tv.tv_usec;
777 if (settimeofday(&tv, NULL) == -1)
778 bb_perror_msg_and_die("settimeofday");
781 strftime(buf, sizeof(buf), "%a %b %e %H:%M:%S %Z %Y", localtime(&tval));
783 bb_error_msg("setting clock to %s (offset %fs)", buf, offset);
785 /* Correct various fields which contain time-relative values: */
787 /* p->lastpkt_recv_time, p->next_action_time and such: */
788 for (item = G.ntp_peers; item != NULL; item = item->link) {
789 peer_t *pp = (peer_t *) item->data;
790 reset_peer_stats(pp, offset);
793 G.cur_time -= offset;
794 G.last_update_recv_time -= offset;
799 * Selection and clustering, and their helpers
807 compare_point_edge(const void *aa, const void *bb)
809 const point_t *a = aa;
810 const point_t *b = bb;
811 if (a->edge < b->edge) {
814 return (a->edge > b->edge);
821 compare_survivor_metric(const void *aa, const void *bb)
823 const survivor_t *a = aa;
824 const survivor_t *b = bb;
825 if (a->metric < b->metric) {
828 return (a->metric > b->metric);
831 fit(peer_t *p, double rd)
833 if ((p->reachable_bits & (p->reachable_bits-1)) == 0) {
834 /* One or zero bits in reachable_bits */
835 VERB3 bb_error_msg("peer %s unfit for selection: unreachable", p->p_dotted);
838 #if 0 /* we filter out such packets earlier */
839 if ((p->lastpkt_status & LI_ALARM) == LI_ALARM
840 || p->lastpkt_stratum >= MAXSTRAT
842 VERB3 bb_error_msg("peer %s unfit for selection: bad status/stratum", p->p_dotted);
846 /* rd is root_distance(p) */
847 if (rd > MAXDIST + FREQ_TOLERANCE * (1 << G.poll_exp)) {
848 VERB3 bb_error_msg("peer %s unfit for selection: root distance too high", p->p_dotted);
852 // /* Do we have a loop? */
853 // if (p->refid == p->dstaddr || p->refid == s.refid)
858 select_and_cluster(void)
862 int size = 3 * G.peer_cnt;
863 /* for selection algorithm */
865 unsigned num_points, num_candidates;
867 unsigned num_falsetickers;
868 /* for cluster algorithm */
869 survivor_t survivor[size];
870 unsigned num_survivors;
876 if (G.initial_poll_complete) while (item != NULL) {
877 peer_t *p = (peer_t *) item->data;
878 double rd = root_distance(p);
879 double offset = p->filter_offset;
886 VERB4 bb_error_msg("interval: [%f %f %f] %s",
892 point[num_points].p = p;
893 point[num_points].type = -1;
894 point[num_points].edge = offset - rd;
896 point[num_points].p = p;
897 point[num_points].type = 0;
898 point[num_points].edge = offset;
900 point[num_points].p = p;
901 point[num_points].type = 1;
902 point[num_points].edge = offset + rd;
906 num_candidates = num_points / 3;
907 if (num_candidates == 0) {
908 VERB3 bb_error_msg("no valid datapoints, no peer selected");
911 //TODO: sorting does not seem to be done in reference code
912 qsort(point, num_points, sizeof(point[0]), compare_point_edge);
914 /* Start with the assumption that there are no falsetickers.
915 * Attempt to find a nonempty intersection interval containing
916 * the midpoints of all truechimers.
917 * If a nonempty interval cannot be found, increase the number
918 * of assumed falsetickers by one and try again.
919 * If a nonempty interval is found and the number of falsetickers
920 * is less than the number of truechimers, a majority has been found
921 * and the midpoint of each truechimer represents
922 * the candidates available to the cluster algorithm.
924 num_falsetickers = 0;
927 unsigned num_midpoints = 0;
932 for (i = 0; i < num_points; i++) {
934 * if (point[i].type == -1) c++;
935 * if (point[i].type == 1) c--;
936 * and it's simpler to do it this way:
939 if (c >= num_candidates - num_falsetickers) {
940 /* If it was c++ and it got big enough... */
944 if (point[i].type == 0)
948 for (i = num_points-1; i >= 0; i--) {
950 if (c >= num_candidates - num_falsetickers) {
951 high = point[i].edge;
954 if (point[i].type == 0)
957 /* If the number of midpoints is greater than the number
958 * of allowed falsetickers, the intersection contains at
959 * least one truechimer with no midpoint - bad.
960 * Also, interval should be nonempty.
962 if (num_midpoints <= num_falsetickers && low < high)
965 if (num_falsetickers * 2 >= num_candidates) {
966 VERB3 bb_error_msg("too many falsetickers:%d (candidates:%d), no peer selected",
967 num_falsetickers, num_candidates);
971 VERB3 bb_error_msg("selected interval: [%f, %f]; candidates:%d falsetickers:%d",
972 low, high, num_candidates, num_falsetickers);
976 /* Construct a list of survivors (p, metric)
977 * from the chime list, where metric is dominated
978 * first by stratum and then by root distance.
979 * All other things being equal, this is the order of preference.
982 for (i = 0; i < num_points; i++) {
985 if (point[i].edge < low || point[i].edge > high)
988 survivor[num_survivors].p = p;
989 //TODO: save root_distance in point_t and reuse here?
990 survivor[num_survivors].metric = MAXDIST * p->lastpkt_stratum + root_distance(p);
991 VERB4 bb_error_msg("survivor[%d] metric:%f peer:%s",
992 num_survivors, survivor[num_survivors].metric, p->p_dotted);
995 /* There must be at least MIN_SELECTED survivors to satisfy the
996 * correctness assertions. Ordinarily, the Byzantine criteria
997 * require four survivors, but for the demonstration here, one
1000 if (num_survivors < MIN_SELECTED) {
1001 VERB3 bb_error_msg("num_survivors %d < %d, no peer selected",
1002 num_survivors, MIN_SELECTED);
1006 //looks like this is ONLY used by the fact that later we pick survivor[0].
1007 //we can avoid sorting then, just find the minimum once!
1008 qsort(survivor, num_survivors, sizeof(survivor[0]), compare_survivor_metric);
1010 /* For each association p in turn, calculate the selection
1011 * jitter p->sjitter as the square root of the sum of squares
1012 * (p->offset - q->offset) over all q associations. The idea is
1013 * to repeatedly discard the survivor with maximum selection
1014 * jitter until a termination condition is met.
1017 unsigned max_idx = max_idx;
1018 double max_selection_jitter = max_selection_jitter;
1019 double min_jitter = min_jitter;
1021 if (num_survivors <= MIN_CLUSTERED) {
1022 VERB3 bb_error_msg("num_survivors %d <= %d, not discarding more",
1023 num_survivors, MIN_CLUSTERED);
1027 /* To make sure a few survivors are left
1028 * for the clustering algorithm to chew on,
1029 * we stop if the number of survivors
1030 * is less than or equal to MIN_CLUSTERED (3).
1032 for (i = 0; i < num_survivors; i++) {
1033 double selection_jitter_sq;
1034 peer_t *p = survivor[i].p;
1036 if (i == 0 || p->filter_jitter < min_jitter)
1037 min_jitter = p->filter_jitter;
1039 selection_jitter_sq = 0;
1040 for (j = 0; j < num_survivors; j++) {
1041 peer_t *q = survivor[j].p;
1042 selection_jitter_sq += SQUARE(p->filter_offset - q->filter_offset);
1044 if (i == 0 || selection_jitter_sq > max_selection_jitter) {
1045 max_selection_jitter = selection_jitter_sq;
1048 VERB5 bb_error_msg("survivor %d selection_jitter^2:%f",
1049 i, selection_jitter_sq);
1051 max_selection_jitter = SQRT(max_selection_jitter / num_survivors);
1052 VERB4 bb_error_msg("max_selection_jitter (at %d):%f min_jitter:%f",
1053 max_idx, max_selection_jitter, min_jitter);
1055 /* If the maximum selection jitter is less than the
1056 * minimum peer jitter, then tossing out more survivors
1057 * will not lower the minimum peer jitter, so we might
1060 if (max_selection_jitter < min_jitter) {
1061 VERB3 bb_error_msg("max_selection_jitter:%f < min_jitter:%f, num_survivors:%d, not discarding more",
1062 max_selection_jitter, min_jitter, num_survivors);
1066 /* Delete survivor[max_idx] from the list
1067 * and go around again.
1069 VERB5 bb_error_msg("dropping survivor %d", max_idx);
1071 while (max_idx < num_survivors) {
1072 survivor[max_idx] = survivor[max_idx + 1];
1077 /* Pick the best clock. If the old system peer is on the list
1078 * and at the same stratum as the first survivor on the list,
1079 * then don't do a clock hop. Otherwise, select the first
1080 * survivor on the list as the new system peer.
1082 //TODO - see clock_combine()
1083 VERB3 bb_error_msg("selected peer %s filter_offset:%f age:%f",
1084 survivor[0].p->p_dotted,
1085 survivor[0].p->filter_offset,
1086 G.cur_time - survivor[0].p->lastpkt_recv_time
1088 return survivor[0].p;
1093 * Local clock discipline and its helpers
1096 set_new_values(int disc_state, double offset, double recv_time)
1098 /* Enter new state and set state variables. Note we use the time
1099 * of the last clock filter sample, which must be earlier than
1102 VERB3 bb_error_msg("disc_state=%d last update offset=%f recv_time=%f",
1103 disc_state, offset, recv_time);
1104 G.discipline_state = disc_state;
1105 G.last_update_offset = offset;
1106 G.last_update_recv_time = recv_time;
1108 /* Clock state definitions */
1109 #define STATE_NSET 0 /* initial state, "nothing is set" */
1110 #define STATE_FSET 1 /* frequency set from file */
1111 #define STATE_SPIK 2 /* spike detected */
1112 #define STATE_FREQ 3 /* initial frequency */
1113 #define STATE_SYNC 4 /* clock synchronized (normal operation) */
1114 /* Return: -1: decrease poll interval, 0: leave as is, 1: increase */
1116 update_local_clock(peer_t *p)
1119 long old_tmx_offset;
1121 double offset = p->filter_offset;
1122 double recv_time = p->lastpkt_recv_time;
1124 #if !USING_KERNEL_PLL_LOOP
1127 double since_last_update;
1128 double etemp, dtemp;
1130 abs_offset = fabs(offset);
1133 /* If needed, -S script can detect this by looking at $offset
1134 * env var and kill parent */
1135 /* If the offset is too large, give up and go home */
1136 if (abs_offset > PANIC_THRESHOLD) {
1137 bb_error_msg_and_die("offset %f far too big, exiting", offset);
1141 /* If this is an old update, for instance as the result
1142 * of a system peer change, avoid it. We never use
1143 * an old sample or the same sample twice.
1145 if (recv_time <= G.last_update_recv_time) {
1146 VERB3 bb_error_msg("same or older datapoint: %f >= %f, not using it",
1147 G.last_update_recv_time, recv_time);
1148 return 0; /* "leave poll interval as is" */
1151 /* Clock state machine transition function. This is where the
1152 * action is and defines how the system reacts to large time
1153 * and frequency errors.
1155 since_last_update = recv_time - G.reftime;
1156 #if !USING_KERNEL_PLL_LOOP
1159 if (G.discipline_state == STATE_FREQ) {
1160 /* Ignore updates until the stepout threshold */
1161 if (since_last_update < WATCH_THRESHOLD) {
1162 VERB3 bb_error_msg("measuring drift, datapoint ignored, %f sec remains",
1163 WATCH_THRESHOLD - since_last_update);
1164 return 0; /* "leave poll interval as is" */
1166 #if !USING_KERNEL_PLL_LOOP
1167 freq_drift = (offset - G.last_update_offset) / since_last_update;
1171 /* There are two main regimes: when the
1172 * offset exceeds the step threshold and when it does not.
1174 if (abs_offset > STEP_THRESHOLD) {
1175 switch (G.discipline_state) {
1177 /* The first outlyer: ignore it, switch to SPIK state */
1178 VERB3 bb_error_msg("offset:%f - spike detected", offset);
1179 G.discipline_state = STATE_SPIK;
1180 return -1; /* "decrease poll interval" */
1183 /* Ignore succeeding outlyers until either an inlyer
1184 * is found or the stepout threshold is exceeded.
1186 if (since_last_update < WATCH_THRESHOLD) {
1187 VERB3 bb_error_msg("spike detected, datapoint ignored, %f sec remains",
1188 WATCH_THRESHOLD - since_last_update);
1189 return -1; /* "decrease poll interval" */
1191 /* fall through: we need to step */
1194 /* Step the time and clamp down the poll interval.
1196 * In NSET state an initial frequency correction is
1197 * not available, usually because the frequency file has
1198 * not yet been written. Since the time is outside the
1199 * capture range, the clock is stepped. The frequency
1200 * will be set directly following the stepout interval.
1202 * In FSET state the initial frequency has been set
1203 * from the frequency file. Since the time is outside
1204 * the capture range, the clock is stepped immediately,
1205 * rather than after the stepout interval. Guys get
1206 * nervous if it takes 17 minutes to set the clock for
1209 * In SPIK state the stepout threshold has expired and
1210 * the phase is still above the step threshold. Note
1211 * that a single spike greater than the step threshold
1212 * is always suppressed, even at the longer poll
1215 VERB3 bb_error_msg("stepping time by %f; poll_exp=MINPOLL", offset);
1217 if (option_mask32 & OPT_q) {
1218 /* We were only asked to set time once. Done. */
1222 G.polladj_count = 0;
1223 G.poll_exp = MINPOLL;
1224 G.stratum = MAXSTRAT;
1226 run_script("step", offset);
1228 if (G.discipline_state == STATE_NSET) {
1229 set_new_values(STATE_FREQ, /*offset:*/ 0, recv_time);
1230 return 1; /* "ok to increase poll interval" */
1232 set_new_values(STATE_SYNC, /*offset:*/ 0, recv_time);
1234 } else { /* abs_offset <= STEP_THRESHOLD */
1236 if (G.poll_exp < MINPOLL && G.initial_poll_complete) {
1237 VERB3 bb_error_msg("small offset:%f, disabling burst mode", offset);
1238 G.polladj_count = 0;
1239 G.poll_exp = MINPOLL;
1242 /* Compute the clock jitter as the RMS of exponentially
1243 * weighted offset differences. Used by the poll adjust code.
1245 etemp = SQUARE(G.discipline_jitter);
1246 dtemp = SQUARE(MAXD(fabs(offset - G.last_update_offset), G_precision_sec));
1247 G.discipline_jitter = SQRT(etemp + (dtemp - etemp) / AVG);
1248 VERB3 bb_error_msg("discipline jitter=%f", G.discipline_jitter);
1250 switch (G.discipline_state) {
1252 if (option_mask32 & OPT_q) {
1253 /* We were only asked to set time once.
1254 * The clock is precise enough, no need to step.
1258 /* This is the first update received and the frequency
1259 * has not been initialized. The first thing to do
1260 * is directly measure the oscillator frequency.
1262 set_new_values(STATE_FREQ, offset, recv_time);
1263 VERB3 bb_error_msg("transitioning to FREQ, datapoint ignored");
1264 return 0; /* "leave poll interval as is" */
1266 #if 0 /* this is dead code for now */
1268 /* This is the first update and the frequency
1269 * has been initialized. Adjust the phase, but
1270 * don't adjust the frequency until the next update.
1272 set_new_values(STATE_SYNC, offset, recv_time);
1273 /* freq_drift remains 0 */
1278 /* since_last_update >= WATCH_THRESHOLD, we waited enough.
1279 * Correct the phase and frequency and switch to SYNC state.
1280 * freq_drift was already estimated (see code above)
1282 set_new_values(STATE_SYNC, offset, recv_time);
1286 #if !USING_KERNEL_PLL_LOOP
1287 /* Compute freq_drift due to PLL and FLL contributions.
1289 * The FLL and PLL frequency gain constants
1290 * depend on the poll interval and Allan
1291 * intercept. The FLL is not used below one-half
1292 * the Allan intercept. Above that the loop gain
1293 * increases in steps to 1 / AVG.
1295 if ((1 << G.poll_exp) > ALLAN / 2) {
1296 etemp = FLL - G.poll_exp;
1299 freq_drift += (offset - G.last_update_offset) / (MAXD(since_last_update, ALLAN) * etemp);
1301 /* For the PLL the integration interval
1302 * (numerator) is the minimum of the update
1303 * interval and poll interval. This allows
1304 * oversampling, but not undersampling.
1306 etemp = MIND(since_last_update, (1 << G.poll_exp));
1307 dtemp = (4 * PLL) << G.poll_exp;
1308 freq_drift += offset * etemp / SQUARE(dtemp);
1310 set_new_values(STATE_SYNC, offset, recv_time);
1313 if (G.stratum != p->lastpkt_stratum + 1) {
1314 G.stratum = p->lastpkt_stratum + 1;
1315 run_script("stratum", offset);
1319 G.reftime = G.cur_time;
1320 G.ntp_status = p->lastpkt_status;
1321 G.refid = p->lastpkt_refid;
1322 G.rootdelay = p->lastpkt_rootdelay + p->lastpkt_delay;
1323 dtemp = p->filter_jitter; // SQRT(SQUARE(p->filter_jitter) + SQUARE(s.jitter));
1324 dtemp += MAXD(p->filter_dispersion + FREQ_TOLERANCE * (G.cur_time - p->lastpkt_recv_time) + abs_offset, MINDISP);
1325 G.rootdisp = p->lastpkt_rootdisp + dtemp;
1326 VERB3 bb_error_msg("updating leap/refid/reftime/rootdisp from peer %s", p->p_dotted);
1328 /* We are in STATE_SYNC now, but did not do adjtimex yet.
1329 * (Any other state does not reach this, they all return earlier)
1330 * By this time, freq_drift and G.last_update_offset are set
1331 * to values suitable for adjtimex.
1333 #if !USING_KERNEL_PLL_LOOP
1334 /* Calculate the new frequency drift and frequency stability (wander).
1335 * Compute the clock wander as the RMS of exponentially weighted
1336 * frequency differences. This is not used directly, but can,
1337 * along with the jitter, be a highly useful monitoring and
1340 dtemp = G.discipline_freq_drift + freq_drift;
1341 G.discipline_freq_drift = MAXD(MIND(MAXDRIFT, dtemp), -MAXDRIFT);
1342 etemp = SQUARE(G.discipline_wander);
1343 dtemp = SQUARE(dtemp);
1344 G.discipline_wander = SQRT(etemp + (dtemp - etemp) / AVG);
1346 VERB3 bb_error_msg("discipline freq_drift=%.9f(int:%ld corr:%e) wander=%f",
1347 G.discipline_freq_drift,
1348 (long)(G.discipline_freq_drift * 65536e6),
1350 G.discipline_wander);
1353 memset(&tmx, 0, sizeof(tmx));
1354 if (adjtimex(&tmx) < 0)
1355 bb_perror_msg_and_die("adjtimex");
1356 VERB3 bb_error_msg("p adjtimex freq:%ld offset:%ld constant:%ld status:0x%x",
1357 tmx.freq, tmx.offset, tmx.constant, tmx.status);
1361 if (!G.adjtimex_was_done) {
1362 G.adjtimex_was_done = 1;
1363 /* When we use adjtimex for the very first time,
1364 * we need to ADD to pre-existing tmx.offset - it may be !0
1366 memset(&tmx, 0, sizeof(tmx));
1367 if (adjtimex(&tmx) < 0)
1368 bb_perror_msg_and_die("adjtimex");
1369 old_tmx_offset = tmx.offset;
1371 memset(&tmx, 0, sizeof(tmx));
1373 //doesn't work, offset remains 0 (!) in kernel:
1374 //ntpd: set adjtimex freq:1786097 tmx.offset:77487
1375 //ntpd: prev adjtimex freq:1786097 tmx.offset:0
1376 //ntpd: cur adjtimex freq:1786097 tmx.offset:0
1377 tmx.modes = ADJ_FREQUENCY | ADJ_OFFSET;
1378 /* 65536 is one ppm */
1379 tmx.freq = G.discipline_freq_drift * 65536e6;
1380 tmx.offset = G.last_update_offset * 1000000; /* usec */
1382 tmx.modes = ADJ_OFFSET | ADJ_STATUS | ADJ_TIMECONST;// | ADJ_MAXERROR | ADJ_ESTERROR;
1383 tmx.offset = (G.last_update_offset * 1000000) /* usec */
1384 /* + (G.last_update_offset < 0 ? -0.5 : 0.5) - too small to bother */
1385 + old_tmx_offset; /* almost always 0 */
1386 tmx.status = STA_PLL;
1387 if (G.ntp_status & LI_PLUSSEC)
1388 tmx.status |= STA_INS;
1389 if (G.ntp_status & LI_MINUSSEC)
1390 tmx.status |= STA_DEL;
1391 tmx.constant = G.poll_exp - 4;
1392 //tmx.esterror = (u_int32)(clock_jitter * 1e6);
1393 //tmx.maxerror = (u_int32)((sys_rootdelay / 2 + sys_rootdisp) * 1e6);
1394 rc = adjtimex(&tmx);
1396 bb_perror_msg_and_die("adjtimex");
1397 /* NB: here kernel returns constant == G.poll_exp, not == G.poll_exp - 4.
1398 * Not sure why. Perhaps it is normal.
1400 VERB3 bb_error_msg("adjtimex:%d freq:%ld offset:%ld constant:%ld status:0x%x",
1401 rc, tmx.freq, tmx.offset, tmx.constant, tmx.status);
1404 /* always gives the same output as above msg */
1405 memset(&tmx, 0, sizeof(tmx));
1406 if (adjtimex(&tmx) < 0)
1407 bb_perror_msg_and_die("adjtimex");
1408 VERB3 bb_error_msg("c adjtimex freq:%ld offset:%ld constant:%ld status:0x%x",
1409 tmx.freq, tmx.offset, tmx.constant, tmx.status);
1412 G.kernel_freq_drift = tmx.freq / 65536;
1413 VERB2 bb_error_msg("update offset:%f, clock drift:%ld ppm", G.last_update_offset, G.kernel_freq_drift);
1415 return 1; /* "ok to increase poll interval" */
1420 * We've got a new reply packet from a peer, process it
1424 retry_interval(void)
1426 /* Local problem, want to retry soon */
1427 unsigned interval, r;
1428 interval = RETRY_INTERVAL;
1430 interval += r % (unsigned)(RETRY_INTERVAL / 4);
1431 VERB3 bb_error_msg("chose retry interval:%u", interval);
1435 poll_interval(int exponent)
1437 unsigned interval, r;
1438 exponent = G.poll_exp + exponent;
1441 interval = 1 << exponent;
1443 interval += ((r & (interval-1)) >> 4) + ((r >> 8) & 1); /* + 1/16 of interval, max */
1444 VERB3 bb_error_msg("chose poll interval:%u (poll_exp:%d exp:%d)", interval, G.poll_exp, exponent);
1447 static NOINLINE void
1448 recv_and_process_peer_pkt(peer_t *p)
1453 double T1, T2, T3, T4;
1455 datapoint_t *datapoint;
1458 /* We can recvfrom here and check from.IP, but some multihomed
1459 * ntp servers reply from their *other IP*.
1460 * TODO: maybe we should check at least what we can: from.port == 123?
1462 size = recv(p->p_fd, &msg, sizeof(msg), MSG_DONTWAIT);
1464 bb_perror_msg("recv(%s) error", p->p_dotted);
1465 if (errno == EHOSTUNREACH || errno == EHOSTDOWN
1466 || errno == ENETUNREACH || errno == ENETDOWN
1467 || errno == ECONNREFUSED || errno == EADDRNOTAVAIL
1470 //TODO: always do this?
1471 interval = retry_interval();
1472 goto set_next_and_close_sock;
1477 if (size != NTP_MSGSIZE_NOAUTH && size != NTP_MSGSIZE) {
1478 bb_error_msg("malformed packet received from %s", p->p_dotted);
1482 if (msg.m_orgtime.int_partl != p->p_xmt_msg.m_xmttime.int_partl
1483 || msg.m_orgtime.fractionl != p->p_xmt_msg.m_xmttime.fractionl
1488 if ((msg.m_status & LI_ALARM) == LI_ALARM
1489 || msg.m_stratum == 0
1490 || msg.m_stratum > NTP_MAXSTRATUM
1492 // TODO: stratum 0 responses may have commands in 32-bit m_refid field:
1493 // "DENY", "RSTR" - peer does not like us at all
1494 // "RATE" - peer is overloaded, reduce polling freq
1495 interval = poll_interval(0);
1496 bb_error_msg("reply from %s: not synced, next query in %us", p->p_dotted, interval);
1497 goto set_next_and_close_sock;
1500 // /* Verify valid root distance */
1501 // if (msg.m_rootdelay / 2 + msg.m_rootdisp >= MAXDISP || p->lastpkt_reftime > msg.m_xmt)
1502 // return; /* invalid header values */
1504 p->lastpkt_status = msg.m_status;
1505 p->lastpkt_stratum = msg.m_stratum;
1506 p->lastpkt_rootdelay = sfp_to_d(msg.m_rootdelay);
1507 p->lastpkt_rootdisp = sfp_to_d(msg.m_rootdisp);
1508 p->lastpkt_refid = msg.m_refid;
1511 * From RFC 2030 (with a correction to the delay math):
1513 * Timestamp Name ID When Generated
1514 * ------------------------------------------------------------
1515 * Originate Timestamp T1 time request sent by client
1516 * Receive Timestamp T2 time request received by server
1517 * Transmit Timestamp T3 time reply sent by server
1518 * Destination Timestamp T4 time reply received by client
1520 * The roundtrip delay and local clock offset are defined as
1522 * delay = (T4 - T1) - (T3 - T2); offset = ((T2 - T1) + (T3 - T4)) / 2
1525 T2 = lfp_to_d(msg.m_rectime);
1526 T3 = lfp_to_d(msg.m_xmttime);
1529 p->lastpkt_recv_time = T4;
1531 VERB5 bb_error_msg("%s->lastpkt_recv_time=%f", p->p_dotted, p->lastpkt_recv_time);
1532 p->datapoint_idx = p->reachable_bits ? (p->datapoint_idx + 1) % NUM_DATAPOINTS : 0;
1533 datapoint = &p->filter_datapoint[p->datapoint_idx];
1534 datapoint->d_recv_time = T4;
1535 datapoint->d_offset = ((T2 - T1) + (T3 - T4)) / 2;
1536 /* The delay calculation is a special case. In cases where the
1537 * server and client clocks are running at different rates and
1538 * with very fast networks, the delay can appear negative. In
1539 * order to avoid violating the Principle of Least Astonishment,
1540 * the delay is clamped not less than the system precision.
1542 p->lastpkt_delay = (T4 - T1) - (T3 - T2);
1543 if (p->lastpkt_delay < G_precision_sec)
1544 p->lastpkt_delay = G_precision_sec;
1545 datapoint->d_dispersion = LOG2D(msg.m_precision_exp) + G_precision_sec;
1546 if (!p->reachable_bits) {
1547 /* 1st datapoint ever - replicate offset in every element */
1549 for (i = 1; i < NUM_DATAPOINTS; i++) {
1550 p->filter_datapoint[i].d_offset = datapoint->d_offset;
1554 p->reachable_bits |= 1;
1555 if ((MAX_VERBOSE && G.verbose) || (option_mask32 & OPT_w)) {
1556 bb_error_msg("reply from %s: reach 0x%02x offset %f delay %f status 0x%02x strat %d refid 0x%08x rootdelay %f",
1559 datapoint->d_offset,
1564 p->lastpkt_rootdelay
1565 /* not shown: m_ppoll, m_precision_exp, m_rootdisp,
1566 * m_reftime, m_orgtime, m_rectime, m_xmttime
1571 /* Muck with statictics and update the clock */
1572 filter_datapoints(p);
1573 q = select_and_cluster();
1577 if (!(option_mask32 & OPT_w)) {
1578 rc = update_local_clock(q);
1579 /* If drift is dangerously large, immediately
1580 * drop poll interval one step down.
1582 if (q->filter_offset < -POLLDOWN_OFFSET
1583 || q->filter_offset > POLLDOWN_OFFSET
1585 VERB3 bb_error_msg("offset:%f > POLLDOWN_OFFSET", q->filter_offset);
1590 /* else: no peer selected, rc = -1: we want to poll more often */
1593 /* Adjust the poll interval by comparing the current offset
1594 * with the clock jitter. If the offset is less than
1595 * the clock jitter times a constant, then the averaging interval
1596 * is increased, otherwise it is decreased. A bit of hysteresis
1597 * helps calm the dance. Works best using burst mode.
1600 bb_error_msg("offset:%f POLLADJ_GATE*discipline_jitter:%f poll:%s",
1601 q->filter_offset, POLLADJ_GATE * G.discipline_jitter,
1602 fabs(q->filter_offset) < POLLADJ_GATE * G.discipline_jitter
1606 if (rc > 0 && fabs(q->filter_offset) < POLLADJ_GATE * G.discipline_jitter) {
1607 /* was += G.poll_exp but it is a bit
1608 * too optimistic for my taste at high poll_exp's */
1609 G.polladj_count += MINPOLL;
1610 if (G.polladj_count > POLLADJ_LIMIT) {
1611 G.polladj_count = 0;
1612 if (G.poll_exp < MAXPOLL) {
1614 VERB3 bb_error_msg("polladj: discipline_jitter:%f ++poll_exp=%d",
1615 G.discipline_jitter, G.poll_exp);
1618 VERB3 bb_error_msg("polladj: incr:%d", G.polladj_count);
1621 G.polladj_count -= G.poll_exp * 2;
1622 if (G.polladj_count < -POLLADJ_LIMIT || G.poll_exp >= BIGPOLL) {
1624 G.polladj_count = 0;
1625 if (G.poll_exp > MINPOLL) {
1629 /* Correct p->next_action_time in each peer
1630 * which waits for sending, so that they send earlier.
1631 * Old pp->next_action_time are on the order
1632 * of t + (1 << old_poll_exp) + small_random,
1633 * we simply need to subtract ~half of that.
1635 for (item = G.ntp_peers; item != NULL; item = item->link) {
1636 peer_t *pp = (peer_t *) item->data;
1638 pp->next_action_time -= (1 << G.poll_exp);
1640 VERB3 bb_error_msg("polladj: discipline_jitter:%f --poll_exp=%d",
1641 G.discipline_jitter, G.poll_exp);
1644 VERB3 bb_error_msg("polladj: decr:%d", G.polladj_count);
1649 /* Decide when to send new query for this peer */
1650 interval = poll_interval(0);
1652 set_next_and_close_sock:
1653 set_next(p, interval);
1654 /* We do not expect any more packets from this peer for now.
1655 * Closing the socket informs kernel about it.
1656 * We open a new socket when we send a new query.
1664 #if ENABLE_FEATURE_NTPD_SERVER
1665 static NOINLINE void
1666 recv_and_process_client_pkt(void /*int fd*/)
1670 len_and_sockaddr *to;
1671 struct sockaddr *from;
1673 uint8_t query_status;
1674 l_fixedpt_t query_xmttime;
1676 to = get_sock_lsa(G.listen_fd);
1677 from = xzalloc(to->len);
1679 size = recv_from_to(G.listen_fd, &msg, sizeof(msg), MSG_DONTWAIT, from, &to->u.sa, to->len);
1680 if (size != NTP_MSGSIZE_NOAUTH && size != NTP_MSGSIZE) {
1683 if (errno == EAGAIN)
1685 bb_perror_msg_and_die("recv");
1687 addr = xmalloc_sockaddr2dotted_noport(from);
1688 bb_error_msg("malformed packet received from %s: size %u", addr, (int)size);
1693 query_status = msg.m_status;
1694 query_xmttime = msg.m_xmttime;
1696 /* Build a reply packet */
1697 memset(&msg, 0, sizeof(msg));
1698 msg.m_status = G.stratum < MAXSTRAT ? G.ntp_status : LI_ALARM;
1699 msg.m_status |= (query_status & VERSION_MASK);
1700 msg.m_status |= ((query_status & MODE_MASK) == MODE_CLIENT) ?
1701 MODE_SERVER : MODE_SYM_PAS;
1702 msg.m_stratum = G.stratum;
1703 msg.m_ppoll = G.poll_exp;
1704 msg.m_precision_exp = G_precision_exp;
1705 /* this time was obtained between poll() and recv() */
1706 msg.m_rectime = d_to_lfp(G.cur_time);
1707 msg.m_xmttime = d_to_lfp(gettime1900d()); /* this instant */
1708 msg.m_reftime = d_to_lfp(G.reftime);
1709 msg.m_orgtime = query_xmttime;
1710 msg.m_rootdelay = d_to_sfp(G.rootdelay);
1711 //simple code does not do this, fix simple code!
1712 msg.m_rootdisp = d_to_sfp(G.rootdisp);
1713 version = (query_status & VERSION_MASK); /* ... >> VERSION_SHIFT - done below instead */
1714 msg.m_refid = G.refid; // (version > (3 << VERSION_SHIFT)) ? G.refid : G.refid3;
1716 /* We reply from the local address packet was sent to,
1717 * this makes to/from look swapped here: */
1718 do_sendto(G.listen_fd,
1719 /*from:*/ &to->u.sa, /*to:*/ from, /*addrlen:*/ to->len,
1728 /* Upstream ntpd's options:
1730 * -4 Force DNS resolution of host names to the IPv4 namespace.
1731 * -6 Force DNS resolution of host names to the IPv6 namespace.
1732 * -a Require cryptographic authentication for broadcast client,
1733 * multicast client and symmetric passive associations.
1734 * This is the default.
1735 * -A Do not require cryptographic authentication for broadcast client,
1736 * multicast client and symmetric passive associations.
1737 * This is almost never a good idea.
1738 * -b Enable the client to synchronize to broadcast servers.
1740 * Specify the name and path of the configuration file,
1741 * default /etc/ntp.conf
1742 * -d Specify debugging mode. This option may occur more than once,
1743 * with each occurrence indicating greater detail of display.
1745 * Specify debugging level directly.
1747 * Specify the name and path of the frequency file.
1748 * This is the same operation as the "driftfile FILE"
1749 * configuration command.
1750 * -g Normally, ntpd exits with a message to the system log
1751 * if the offset exceeds the panic threshold, which is 1000 s
1752 * by default. This option allows the time to be set to any value
1753 * without restriction; however, this can happen only once.
1754 * If the threshold is exceeded after that, ntpd will exit
1755 * with a message to the system log. This option can be used
1756 * with the -q and -x options. See the tinker command for other options.
1758 * Chroot the server to the directory jaildir. This option also implies
1759 * that the server attempts to drop root privileges at startup
1760 * (otherwise, chroot gives very little additional security).
1761 * You may need to also specify a -u option.
1763 * Specify the name and path of the symmetric key file,
1764 * default /etc/ntp/keys. This is the same operation
1765 * as the "keys FILE" configuration command.
1767 * Specify the name and path of the log file. The default
1768 * is the system log file. This is the same operation as
1769 * the "logfile FILE" configuration command.
1770 * -L Do not listen to virtual IPs. The default is to listen.
1772 * -N To the extent permitted by the operating system,
1773 * run the ntpd at the highest priority.
1775 * Specify the name and path of the file used to record the ntpd
1776 * process ID. This is the same operation as the "pidfile FILE"
1777 * configuration command.
1779 * To the extent permitted by the operating system,
1780 * run the ntpd at the specified priority.
1781 * -q Exit the ntpd just after the first time the clock is set.
1782 * This behavior mimics that of the ntpdate program, which is
1783 * to be retired. The -g and -x options can be used with this option.
1784 * Note: The kernel time discipline is disabled with this option.
1786 * Specify the default propagation delay from the broadcast/multicast
1787 * server to this client. This is necessary only if the delay
1788 * cannot be computed automatically by the protocol.
1790 * Specify the directory path for files created by the statistics
1791 * facility. This is the same operation as the "statsdir DIR"
1792 * configuration command.
1794 * Add a key number to the trusted key list. This option can occur
1797 * Specify a user, and optionally a group, to switch to.
1800 * Add a system variable listed by default.
1801 * -x Normally, the time is slewed if the offset is less than the step
1802 * threshold, which is 128 ms by default, and stepped if above
1803 * the threshold. This option sets the threshold to 600 s, which is
1804 * well within the accuracy window to set the clock manually.
1805 * Note: since the slew rate of typical Unix kernels is limited
1806 * to 0.5 ms/s, each second of adjustment requires an amortization
1807 * interval of 2000 s. Thus, an adjustment as much as 600 s
1808 * will take almost 14 days to complete. This option can be used
1809 * with the -g and -q options. See the tinker command for other options.
1810 * Note: The kernel time discipline is disabled with this option.
1813 /* By doing init in a separate function we decrease stack usage
1816 static NOINLINE void ntp_init(char **argv)
1824 bb_error_msg_and_die(bb_msg_you_must_be_root);
1826 /* Set some globals */
1827 G.stratum = MAXSTRAT;
1829 G.poll_exp = BURSTPOLL; /* speeds up initial sync */
1830 G.last_script_run = G.reftime = G.last_update_recv_time = gettime1900d(); /* sets G.cur_time too */
1834 opt_complementary = "dd:p::wn"; /* d: counter; p: list; -w implies -n */
1835 opts = getopt32(argv,
1837 "wp:S:"IF_FEATURE_NTPD_SERVER("l") /* NOT compat */
1839 "46aAbgL", /* compat, ignored */
1840 &peers, &G.script_name, &G.verbose);
1841 if (!(opts & (OPT_p|OPT_l)))
1843 // if (opts & OPT_x) /* disable stepping, only slew is allowed */
1844 // G.time_was_stepped = 1;
1846 add_peers(llist_pop(&peers));
1847 if (!(opts & OPT_n)) {
1848 bb_daemonize_or_rexec(DAEMON_DEVNULL_STDIO, argv);
1849 logmode = LOGMODE_NONE;
1851 #if ENABLE_FEATURE_NTPD_SERVER
1854 G.listen_fd = create_and_bind_dgram_or_die(NULL, 123);
1855 socket_want_pktinfo(G.listen_fd);
1856 setsockopt(G.listen_fd, IPPROTO_IP, IP_TOS, &const_IPTOS_LOWDELAY, sizeof(const_IPTOS_LOWDELAY));
1859 /* I hesitate to set -20 prio. -15 should be high enough for timekeeping */
1861 setpriority(PRIO_PROCESS, 0, -15);
1863 bb_signals((1 << SIGTERM) | (1 << SIGINT), record_signo);
1864 /* Removed SIGHUP here: */
1865 bb_signals((1 << SIGPIPE) | (1 << SIGCHLD), SIG_IGN);
1868 int ntpd_main(int argc UNUSED_PARAM, char **argv) MAIN_EXTERNALLY_VISIBLE;
1869 int ntpd_main(int argc UNUSED_PARAM, char **argv)
1877 memset(&G, 0, sizeof(G));
1878 SET_PTR_TO_GLOBALS(&G);
1882 /* If ENABLE_FEATURE_NTPD_SERVER, + 1 for listen_fd: */
1883 cnt = G.peer_cnt + ENABLE_FEATURE_NTPD_SERVER;
1884 idx2peer = xzalloc(sizeof(idx2peer[0]) * cnt);
1885 pfd = xzalloc(sizeof(pfd[0]) * cnt);
1887 /* Countdown: we never sync before we sent INITIAL_SAMLPES+1
1888 * packets to each peer.
1889 * NB: if some peer is not responding, we may end up sending
1890 * fewer packets to it and more to other peers.
1891 * NB2: sync usually happens using INITIAL_SAMLPES packets,
1892 * since last reply does not come back instantaneously.
1894 cnt = G.peer_cnt * (INITIAL_SAMLPES + 1);
1896 while (!bb_got_signal) {
1902 /* Nothing between here and poll() blocks for any significant time */
1904 nextaction = G.cur_time + 3600;
1907 #if ENABLE_FEATURE_NTPD_SERVER
1908 if (G.listen_fd != -1) {
1909 pfd[0].fd = G.listen_fd;
1910 pfd[0].events = POLLIN;
1914 /* Pass over peer list, send requests, time out on receives */
1915 for (item = G.ntp_peers; item != NULL; item = item->link) {
1916 peer_t *p = (peer_t *) item->data;
1918 if (p->next_action_time <= G.cur_time) {
1919 if (p->p_fd == -1) {
1920 /* Time to send new req */
1922 G.initial_poll_complete = 1;
1924 send_query_to_peer(p);
1926 /* Timed out waiting for reply */
1929 timeout = poll_interval(-2); /* -2: try a bit sooner */
1930 bb_error_msg("timed out waiting for %s, reach 0x%02x, next query in %us",
1931 p->p_dotted, p->reachable_bits, timeout);
1932 set_next(p, timeout);
1936 if (p->next_action_time < nextaction)
1937 nextaction = p->next_action_time;
1940 /* Wait for reply from this peer */
1941 pfd[i].fd = p->p_fd;
1942 pfd[i].events = POLLIN;
1948 timeout = nextaction - G.cur_time;
1951 timeout++; /* (nextaction - G.cur_time) rounds down, compensating */
1953 /* Here we may block */
1954 VERB2 bb_error_msg("poll %us, sockets:%u, poll interval:%us", timeout, i, 1 << G.poll_exp);
1955 nfds = poll(pfd, i, timeout * 1000);
1956 gettime1900d(); /* sets G.cur_time */
1958 if (G.adjtimex_was_done
1959 && G.cur_time - G.last_script_run > 11*60
1961 /* Useful for updating battery-backed RTC and such */
1962 run_script("periodic", G.last_update_offset);
1963 gettime1900d(); /* sets G.cur_time */
1968 /* Process any received packets */
1970 #if ENABLE_FEATURE_NTPD_SERVER
1971 if (G.listen_fd != -1) {
1972 if (pfd[0].revents /* & (POLLIN|POLLERR)*/) {
1974 recv_and_process_client_pkt(/*G.listen_fd*/);
1975 gettime1900d(); /* sets G.cur_time */
1980 for (; nfds != 0 && j < i; j++) {
1981 if (pfd[j].revents /* & (POLLIN|POLLERR)*/) {
1983 recv_and_process_peer_pkt(idx2peer[j]);
1984 gettime1900d(); /* sets G.cur_time */
1987 } /* while (!bb_got_signal) */
1989 kill_myself_with_sig(bb_got_signal);
1997 /*** openntpd-4.6 uses only adjtime, not adjtimex ***/
1999 /*** ntp-4.2.6/ntpd/ntp_loopfilter.c - adjtimex usage ***/
2003 direct_freq(double fp_offset)
2008 * If the kernel is enabled, we need the residual offset to
2009 * calculate the frequency correction.
2011 if (pll_control && kern_enable) {
2012 memset(&ntv, 0, sizeof(ntv));
2015 clock_offset = ntv.offset / 1e9;
2016 #else /* STA_NANO */
2017 clock_offset = ntv.offset / 1e6;
2018 #endif /* STA_NANO */
2019 drift_comp = FREQTOD(ntv.freq);
2021 #endif /* KERNEL_PLL */
2022 set_freq((fp_offset - clock_offset) / (current_time - clock_epoch) + drift_comp);
2028 set_freq(double freq) /* frequency update */
2036 * If the kernel is enabled, update the kernel frequency.
2038 if (pll_control && kern_enable) {
2039 memset(&ntv, 0, sizeof(ntv));
2040 ntv.modes = MOD_FREQUENCY;
2041 ntv.freq = DTOFREQ(drift_comp);
2043 snprintf(tbuf, sizeof(tbuf), "kernel %.3f PPM", drift_comp * 1e6);
2044 report_event(EVNT_FSET, NULL, tbuf);
2046 snprintf(tbuf, sizeof(tbuf), "ntpd %.3f PPM", drift_comp * 1e6);
2047 report_event(EVNT_FSET, NULL, tbuf);
2049 #else /* KERNEL_PLL */
2050 snprintf(tbuf, sizeof(tbuf), "ntpd %.3f PPM", drift_comp * 1e6);
2051 report_event(EVNT_FSET, NULL, tbuf);
2052 #endif /* KERNEL_PLL */
2061 * This code segment works when clock adjustments are made using
2062 * precision time kernel support and the ntp_adjtime() system
2063 * call. This support is available in Solaris 2.6 and later,
2064 * Digital Unix 4.0 and later, FreeBSD, Linux and specially
2065 * modified kernels for HP-UX 9 and Ultrix 4. In the case of the
2066 * DECstation 5000/240 and Alpha AXP, additional kernel
2067 * modifications provide a true microsecond clock and nanosecond
2068 * clock, respectively.
2070 * Important note: The kernel discipline is used only if the
2071 * step threshold is less than 0.5 s, as anything higher can
2072 * lead to overflow problems. This might occur if some misguided
2073 * lad set the step threshold to something ridiculous.
2075 if (pll_control && kern_enable) {
2077 #define MOD_BITS (MOD_OFFSET | MOD_MAXERROR | MOD_ESTERROR | MOD_STATUS | MOD_TIMECONST)
2080 * We initialize the structure for the ntp_adjtime()
2081 * system call. We have to convert everything to
2082 * microseconds or nanoseconds first. Do not update the
2083 * system variables if the ext_enable flag is set. In
2084 * this case, the external clock driver will update the
2085 * variables, which will be read later by the local
2086 * clock driver. Afterwards, remember the time and
2087 * frequency offsets for jitter and stability values and
2088 * to update the frequency file.
2090 memset(&ntv, 0, sizeof(ntv));
2092 ntv.modes = MOD_STATUS;
2095 ntv.modes = MOD_BITS | MOD_NANO;
2096 #else /* STA_NANO */
2097 ntv.modes = MOD_BITS;
2098 #endif /* STA_NANO */
2099 if (clock_offset < 0)
2104 ntv.offset = (int32)(clock_offset * 1e9 + dtemp);
2105 ntv.constant = sys_poll;
2106 #else /* STA_NANO */
2107 ntv.offset = (int32)(clock_offset * 1e6 + dtemp);
2108 ntv.constant = sys_poll - 4;
2109 #endif /* STA_NANO */
2110 ntv.esterror = (u_int32)(clock_jitter * 1e6);
2111 ntv.maxerror = (u_int32)((sys_rootdelay / 2 + sys_rootdisp) * 1e6);
2112 ntv.status = STA_PLL;
2115 * Enable/disable the PPS if requested.
2118 if (!(pll_status & STA_PPSTIME))
2119 report_event(EVNT_KERN,
2120 NULL, "PPS enabled");
2121 ntv.status |= STA_PPSTIME | STA_PPSFREQ;
2123 if (pll_status & STA_PPSTIME)
2124 report_event(EVNT_KERN,
2125 NULL, "PPS disabled");
2126 ntv.status &= ~(STA_PPSTIME |
2129 if (sys_leap == LEAP_ADDSECOND)
2130 ntv.status |= STA_INS;
2131 else if (sys_leap == LEAP_DELSECOND)
2132 ntv.status |= STA_DEL;
2136 * Pass the stuff to the kernel. If it squeals, turn off
2137 * the pps. In any case, fetch the kernel offset,
2138 * frequency and jitter.
2140 if (ntp_adjtime(&ntv) == TIME_ERROR) {
2141 if (!(ntv.status & STA_PPSSIGNAL))
2142 report_event(EVNT_KERN, NULL,
2145 pll_status = ntv.status;
2147 clock_offset = ntv.offset / 1e9;
2148 #else /* STA_NANO */
2149 clock_offset = ntv.offset / 1e6;
2150 #endif /* STA_NANO */
2151 clock_frequency = FREQTOD(ntv.freq);
2154 * If the kernel PPS is lit, monitor its performance.
2156 if (ntv.status & STA_PPSTIME) {
2158 clock_jitter = ntv.jitter / 1e9;
2159 #else /* STA_NANO */
2160 clock_jitter = ntv.jitter / 1e6;
2161 #endif /* STA_NANO */
2164 #if defined(STA_NANO) && NTP_API == 4
2166 * If the TAI changes, update the kernel TAI.
2168 if (loop_tai != sys_tai) {
2170 ntv.modes = MOD_TAI;
2171 ntv.constant = sys_tai;
2174 #endif /* STA_NANO */
2176 #endif /* KERNEL_PLL */