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 source tree.
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 ***********************************************************************
31 //usage:#define ntpd_trivial_usage
32 //usage: "[-dnqNw"IF_FEATURE_NTPD_SERVER("l")"] [-S PROG] [-p PEER]..."
33 //usage:#define ntpd_full_usage "\n\n"
34 //usage: "NTP client/server\n"
35 //usage: "\n -d Verbose"
36 //usage: "\n -n Do not daemonize"
37 //usage: "\n -q Quit after clock is set"
38 //usage: "\n -N Run at high priority"
39 //usage: "\n -w Do not set time (only query peers), implies -n"
40 //usage: IF_FEATURE_NTPD_SERVER(
41 //usage: "\n -l Run as server on port 123"
43 //usage: "\n -S PROG Run PROG after stepping time, stratum change, and every 11 mins"
44 //usage: "\n -p PEER Obtain time from PEER (may be repeated)"
48 #include <netinet/ip.h> /* For IPTOS_LOWDELAY definition */
49 #include <sys/resource.h> /* setpriority */
50 #include <sys/timex.h>
51 #ifndef IPTOS_LOWDELAY
52 # define IPTOS_LOWDELAY 0x10
55 # error "Sorry, your kernel has to support IP_PKTINFO"
59 /* Verbosity control (max level of -dddd options accepted).
60 * max 6 is very talkative (and bloated). 3 is non-bloated,
61 * production level setting.
66 /* High-level description of the algorithm:
68 * We start running with very small poll_exp, BURSTPOLL,
69 * in order to quickly accumulate INITIAL_SAMPLES datapoints
70 * for each peer. Then, time is stepped if the offset is larger
71 * than STEP_THRESHOLD, otherwise it isn't; anyway, we enlarge
72 * poll_exp to MINPOLL and enter frequency measurement step:
73 * we collect new datapoints but ignore them for WATCH_THRESHOLD
74 * seconds. After WATCH_THRESHOLD seconds we look at accumulated
75 * offset and estimate frequency drift.
77 * (frequency measurement step seems to not be strictly needed,
78 * it is conditionally disabled with USING_INITIAL_FREQ_ESTIMATION
81 * After this, we enter "steady state": we collect a datapoint,
82 * we select the best peer, if this datapoint is not a new one
83 * (IOW: if this datapoint isn't for selected peer), sleep
84 * and collect another one; otherwise, use its offset to update
85 * frequency drift, if offset is somewhat large, reduce poll_exp,
86 * otherwise increase poll_exp.
88 * If offset is larger than STEP_THRESHOLD, which shouldn't normally
89 * happen, we assume that something "bad" happened (computer
90 * was hibernated, someone set totally wrong date, etc),
91 * then the time is stepped, all datapoints are discarded,
92 * and we go back to steady state.
94 * Made some changes to speed up re-syncing after our clock goes bad
95 * (tested with suspending my laptop):
96 * - if largish offset (>= STEP_THRESHOLD * 8 == 1 sec) is seen
97 * from a peer, schedule next query for this peer soon
98 * without drastically lowering poll interval for everybody.
99 * This makes us collect enough data for step much faster:
100 * e.g. at poll = 10 (1024 secs), step was done within 5 minutes
101 * after first reply which indicated that our clock is 14 seconds off.
102 * - on step, do not discard d_dispersion data of the existing datapoints,
103 * do not clear reachable_bits. This prevents discarding first ~8
104 * datapoints after the step.
107 #define RETRY_INTERVAL 5 /* on error, retry in N secs */
108 #define RESPONSE_INTERVAL 15 /* wait for reply up to N secs */
109 #define INITIAL_SAMPLES 4 /* how many samples do we want for init */
110 #define BAD_DELAY_GROWTH 4 /* drop packet if its delay grew by more than this */
112 /* Clock discipline parameters and constants */
114 /* Step threshold (sec). std ntpd uses 0.128.
115 * Using exact power of 2 (1/8) results in smaller code */
116 #define STEP_THRESHOLD 0.125
117 #define WATCH_THRESHOLD 128 /* stepout threshold (sec). std ntpd uses 900 (11 mins (!)) */
118 /* NB: set WATCH_THRESHOLD to ~60 when debugging to save time) */
119 //UNUSED: #define PANIC_THRESHOLD 1000 /* panic threshold (sec) */
121 #define FREQ_TOLERANCE 0.000015 /* frequency tolerance (15 PPM) */
122 #define BURSTPOLL 0 /* initial poll */
123 #define MINPOLL 5 /* minimum poll interval. std ntpd uses 6 (6: 64 sec) */
124 /* If we got largish offset from a peer, cap next query interval
125 * for this peer by this many seconds:
127 #define BIGOFF_INTERVAL (1 << 6)
128 /* If offset > discipline_jitter * POLLADJ_GATE, and poll interval is >= 2^BIGPOLL,
129 * then it is decreased _at once_. (If < 2^BIGPOLL, it will be decreased _eventually_).
131 #define BIGPOLL 10 /* 2^10 sec ~= 17 min */
132 #define MAXPOLL 12 /* maximum poll interval (12: 1.1h, 17: 36.4h). std ntpd uses 17 */
133 /* Actively lower poll when we see such big offsets.
134 * With STEP_THRESHOLD = 0.125, it means we try to sync more aggressively
135 * if offset increases over ~0.04 sec */
136 #define POLLDOWN_OFFSET (STEP_THRESHOLD / 3)
137 #define MINDISP 0.01 /* minimum dispersion (sec) */
138 #define MAXDISP 16 /* maximum dispersion (sec) */
139 #define MAXSTRAT 16 /* maximum stratum (infinity metric) */
140 #define MAXDIST 1 /* distance threshold (sec) */
141 #define MIN_SELECTED 1 /* minimum intersection survivors */
142 #define MIN_CLUSTERED 3 /* minimum cluster survivors */
144 #define MAXDRIFT 0.000500 /* frequency drift we can correct (500 PPM) */
146 /* Poll-adjust threshold.
147 * When we see that offset is small enough compared to discipline jitter,
148 * we grow a counter: += MINPOLL. When counter goes over POLLADJ_LIMIT,
149 * we poll_exp++. If offset isn't small, counter -= poll_exp*2,
150 * and when it goes below -POLLADJ_LIMIT, we poll_exp--.
151 * (Bumped from 30 to 40 since otherwise I often see poll_exp going *2* steps down)
153 #define POLLADJ_LIMIT 40
154 /* If offset < discipline_jitter * POLLADJ_GATE, then we decide to increase
155 * poll interval (we think we can't improve timekeeping
156 * by staying at smaller poll).
158 #define POLLADJ_GATE 4
159 #define TIMECONST_HACK_GATE 2
160 /* Compromise Allan intercept (sec). doc uses 1500, std ntpd uses 512 */
164 /* FLL loop gain [why it depends on MAXPOLL??] */
165 #define FLL (MAXPOLL + 1)
166 /* Parameter averaging constant */
175 NTP_MSGSIZE_NOAUTH = 48,
176 NTP_MSGSIZE = (NTP_MSGSIZE_NOAUTH + 4 + NTP_DIGESTSIZE),
179 MODE_MASK = (7 << 0),
180 VERSION_MASK = (7 << 3),
184 /* Leap Second Codes (high order two bits of m_status) */
185 LI_NOWARNING = (0 << 6), /* no warning */
186 LI_PLUSSEC = (1 << 6), /* add a second (61 seconds) */
187 LI_MINUSSEC = (2 << 6), /* minus a second (59 seconds) */
188 LI_ALARM = (3 << 6), /* alarm condition */
191 MODE_RES0 = 0, /* reserved */
192 MODE_SYM_ACT = 1, /* symmetric active */
193 MODE_SYM_PAS = 2, /* symmetric passive */
194 MODE_CLIENT = 3, /* client */
195 MODE_SERVER = 4, /* server */
196 MODE_BROADCAST = 5, /* broadcast */
197 MODE_RES1 = 6, /* reserved for NTP control message */
198 MODE_RES2 = 7, /* reserved for private use */
201 //TODO: better base selection
202 #define OFFSET_1900_1970 2208988800UL /* 1970 - 1900 in seconds */
204 #define NUM_DATAPOINTS 8
217 uint8_t m_status; /* status of local clock and leap info */
219 uint8_t m_ppoll; /* poll value */
220 int8_t m_precision_exp;
221 s_fixedpt_t m_rootdelay;
222 s_fixedpt_t m_rootdisp;
224 l_fixedpt_t m_reftime;
225 l_fixedpt_t m_orgtime;
226 l_fixedpt_t m_rectime;
227 l_fixedpt_t m_xmttime;
229 uint8_t m_digest[NTP_DIGESTSIZE];
239 len_and_sockaddr *p_lsa;
243 uint32_t lastpkt_refid;
244 uint8_t lastpkt_status;
245 uint8_t lastpkt_stratum;
246 uint8_t reachable_bits;
247 /* when to send new query (if p_fd == -1)
248 * or when receive times out (if p_fd >= 0): */
249 double next_action_time;
251 double lastpkt_recv_time;
252 double lastpkt_delay;
253 double lastpkt_rootdelay;
254 double lastpkt_rootdisp;
255 /* produced by filter algorithm: */
256 double filter_offset;
257 double filter_dispersion;
258 double filter_jitter;
259 datapoint_t filter_datapoint[NUM_DATAPOINTS];
260 /* last sent packet: */
265 #define USING_KERNEL_PLL_LOOP 1
266 #define USING_INITIAL_FREQ_ESTIMATION 0
273 /* Insert new options above this line. */
274 /* Non-compat options: */
278 OPT_l = (1 << 7) * ENABLE_FEATURE_NTPD_SERVER,
279 /* We hijack some bits for other purposes */
285 /* total round trip delay to currently selected reference clock */
287 /* reference timestamp: time when the system clock was last set or corrected */
289 /* total dispersion to currently selected reference clock */
292 double last_script_run;
295 #if ENABLE_FEATURE_NTPD_SERVER
297 # define G_listen_fd (G.listen_fd)
299 # define G_listen_fd (-1)
303 /* refid: 32-bit code identifying the particular server or reference clock
304 * in stratum 0 packets this is a four-character ASCII string,
305 * called the kiss code, used for debugging and monitoring
306 * in stratum 1 packets this is a four-character ASCII string
307 * assigned to the reference clock by IANA. Example: "GPS "
308 * in stratum 2+ packets, it's IPv4 address or 4 first bytes
309 * of MD5 hash of IPv6
313 /* precision is defined as the larger of the resolution and time to
314 * read the clock, in log2 units. For instance, the precision of a
315 * mains-frequency clock incrementing at 60 Hz is 16 ms, even when the
316 * system clock hardware representation is to the nanosecond.
318 * Delays, jitters of various kinds are clamped down to precision.
320 * If precision_sec is too large, discipline_jitter gets clamped to it
321 * and if offset is smaller than discipline_jitter * POLLADJ_GATE, poll
322 * interval grows even though we really can benefit from staying at
323 * smaller one, collecting non-lagged datapoits and correcting offset.
324 * (Lagged datapoits exist when poll_exp is large but we still have
325 * systematic offset error - the time distance between datapoints
326 * is significant and older datapoints have smaller offsets.
327 * This makes our offset estimation a bit smaller than reality)
328 * Due to this effect, setting G_precision_sec close to
329 * STEP_THRESHOLD isn't such a good idea - offsets may grow
330 * too big and we will step. I observed it with -6.
332 * OTOH, setting precision_sec far too small would result in futile
333 * attempts to syncronize to an unachievable precision.
335 * -6 is 1/64 sec, -7 is 1/128 sec and so on.
336 * -8 is 1/256 ~= 0.003906 (worked well for me --vda)
337 * -9 is 1/512 ~= 0.001953 (let's try this for some time)
339 #define G_precision_exp -9
341 * G_precision_exp is used only for construction outgoing packets.
342 * It's ok to set G_precision_sec to a slightly different value
343 * (One which is "nicer looking" in logs).
344 * Exact value would be (1.0 / (1 << (- G_precision_exp))):
346 #define G_precision_sec 0.002
348 /* Bool. After set to 1, never goes back to 0: */
349 smallint initial_poll_complete;
351 #define STATE_NSET 0 /* initial state, "nothing is set" */
352 //#define STATE_FSET 1 /* frequency set from file */
353 //#define STATE_SPIK 2 /* spike detected */
354 //#define STATE_FREQ 3 /* initial frequency */
355 #define STATE_SYNC 4 /* clock synchronized (normal operation) */
356 uint8_t discipline_state; // doc calls it c.state
357 uint8_t poll_exp; // s.poll
358 int polladj_count; // c.count
359 long kernel_freq_drift;
360 peer_t *last_update_peer;
361 double last_update_offset; // c.last
362 double last_update_recv_time; // s.t
363 double discipline_jitter; // c.jitter
364 /* Since we only compare it with ints, can simplify code
365 * by not making this variable floating point:
367 unsigned offset_to_jitter_ratio;
368 //double cluster_offset; // s.offset
369 //double cluster_jitter; // s.jitter
370 #if !USING_KERNEL_PLL_LOOP
371 double discipline_freq_drift; // c.freq
372 /* Maybe conditionally calculate wander? it's used only for logging */
373 double discipline_wander; // c.wander
376 #define G (*ptr_to_globals)
378 static const int const_IPTOS_LOWDELAY = IPTOS_LOWDELAY;
381 #define VERB1 if (MAX_VERBOSE && G.verbose)
382 #define VERB2 if (MAX_VERBOSE >= 2 && G.verbose >= 2)
383 #define VERB3 if (MAX_VERBOSE >= 3 && G.verbose >= 3)
384 #define VERB4 if (MAX_VERBOSE >= 4 && G.verbose >= 4)
385 #define VERB5 if (MAX_VERBOSE >= 5 && G.verbose >= 5)
386 #define VERB6 if (MAX_VERBOSE >= 6 && G.verbose >= 6)
389 static double LOG2D(int a)
392 return 1.0 / (1UL << -a);
395 static ALWAYS_INLINE double SQUARE(double x)
399 static ALWAYS_INLINE double MAXD(double a, double b)
405 static ALWAYS_INLINE double MIND(double a, double b)
411 static NOINLINE double my_SQRT(double X)
418 double Xhalf = X * 0.5;
420 /* Fast and good approximation to 1/sqrt(X), black magic */
422 /*v.i = 0x5f3759df - (v.i >> 1);*/
423 v.i = 0x5f375a86 - (v.i >> 1); /* - this constant is slightly better */
424 invsqrt = v.f; /* better than 0.2% accuracy */
426 /* Refining it using Newton's method: x1 = x0 - f(x0)/f'(x0)
427 * f(x) = 1/(x*x) - X (f==0 when x = 1/sqrt(X))
429 * f(x)/f'(x) = (X - 1/(x*x)) / (2/(x*x*x)) = X*x*x*x/2 - x/2
430 * x1 = x0 - (X*x0*x0*x0/2 - x0/2) = 1.5*x0 - X*x0*x0*x0/2 = x0*(1.5 - (X/2)*x0*x0)
432 invsqrt = invsqrt * (1.5 - Xhalf * invsqrt * invsqrt); /* ~0.05% accuracy */
433 /* invsqrt = invsqrt * (1.5 - Xhalf * invsqrt * invsqrt); 2nd iter: ~0.0001% accuracy */
434 /* With 4 iterations, more than half results will be exact,
435 * at 6th iterations result stabilizes with about 72% results exact.
436 * We are well satisfied with 0.05% accuracy.
439 return X * invsqrt; /* X * 1/sqrt(X) ~= sqrt(X) */
441 static ALWAYS_INLINE double SQRT(double X)
443 /* If this arch doesn't use IEEE 754 floats, fall back to using libm */
444 if (sizeof(float) != 4)
447 /* This avoids needing libm, saves about 0.5k on x86-32 */
455 gettimeofday(&tv, NULL); /* never fails */
456 G.cur_time = tv.tv_sec + (1.0e-6 * tv.tv_usec) + OFFSET_1900_1970;
461 d_to_tv(double d, struct timeval *tv)
463 tv->tv_sec = (long)d;
464 tv->tv_usec = (d - tv->tv_sec) * 1000000;
468 lfp_to_d(l_fixedpt_t lfp)
471 lfp.int_partl = ntohl(lfp.int_partl);
472 lfp.fractionl = ntohl(lfp.fractionl);
473 ret = (double)lfp.int_partl + ((double)lfp.fractionl / UINT_MAX);
477 sfp_to_d(s_fixedpt_t sfp)
480 sfp.int_parts = ntohs(sfp.int_parts);
481 sfp.fractions = ntohs(sfp.fractions);
482 ret = (double)sfp.int_parts + ((double)sfp.fractions / USHRT_MAX);
485 #if ENABLE_FEATURE_NTPD_SERVER
490 lfp.int_partl = (uint32_t)d;
491 lfp.fractionl = (uint32_t)((d - lfp.int_partl) * UINT_MAX);
492 lfp.int_partl = htonl(lfp.int_partl);
493 lfp.fractionl = htonl(lfp.fractionl);
500 sfp.int_parts = (uint16_t)d;
501 sfp.fractions = (uint16_t)((d - sfp.int_parts) * USHRT_MAX);
502 sfp.int_parts = htons(sfp.int_parts);
503 sfp.fractions = htons(sfp.fractions);
509 dispersion(const datapoint_t *dp)
511 return dp->d_dispersion + FREQ_TOLERANCE * (G.cur_time - dp->d_recv_time);
515 root_distance(peer_t *p)
517 /* The root synchronization distance is the maximum error due to
518 * all causes of the local clock relative to the primary server.
519 * It is defined as half the total delay plus total dispersion
522 return MAXD(MINDISP, p->lastpkt_rootdelay + p->lastpkt_delay) / 2
523 + p->lastpkt_rootdisp
524 + p->filter_dispersion
525 + FREQ_TOLERANCE * (G.cur_time - p->lastpkt_recv_time)
530 set_next(peer_t *p, unsigned t)
532 p->next_action_time = G.cur_time + t;
536 * Peer clock filter and its helpers
539 filter_datapoints(peer_t *p)
546 /* Simulations have shown that use of *averaged* offset for p->filter_offset
547 * is in fact worse than simply using last received one: with large poll intervals
548 * (>= 2048) averaging code uses offset values which are outdated by hours,
549 * and time/frequency correction goes totally wrong when fed essentially bogus offsets.
552 double minoff, maxoff, w;
553 double x = x; /* for compiler */
554 double oldest_off = oldest_off;
555 double oldest_age = oldest_age;
556 double newest_off = newest_off;
557 double newest_age = newest_age;
559 fdp = p->filter_datapoint;
561 minoff = maxoff = fdp[0].d_offset;
562 for (i = 1; i < NUM_DATAPOINTS; i++) {
563 if (minoff > fdp[i].d_offset)
564 minoff = fdp[i].d_offset;
565 if (maxoff < fdp[i].d_offset)
566 maxoff = fdp[i].d_offset;
569 idx = p->datapoint_idx; /* most recent datapoint's index */
571 * Drop two outliers and take weighted average of the rest:
572 * most_recent/2 + older1/4 + older2/8 ... + older5/32 + older6/32
573 * we use older6/32, not older6/64 since sum of weights should be 1:
574 * 1/2 + 1/4 + 1/8 + 1/16 + 1/32 + 1/32 = 1
580 * filter_dispersion = \ -------------
587 for (i = 0; i < NUM_DATAPOINTS; i++) {
589 bb_error_msg("datapoint[%d]: off:%f disp:%f(%f) age:%f%s",
592 fdp[idx].d_dispersion, dispersion(&fdp[idx]),
593 G.cur_time - fdp[idx].d_recv_time,
594 (minoff == fdp[idx].d_offset || maxoff == fdp[idx].d_offset)
595 ? " (outlier by offset)" : ""
599 sum += dispersion(&fdp[idx]) / (2 << i);
601 if (minoff == fdp[idx].d_offset) {
602 minoff -= 1; /* so that we don't match it ever again */
604 if (maxoff == fdp[idx].d_offset) {
607 oldest_off = fdp[idx].d_offset;
608 oldest_age = G.cur_time - fdp[idx].d_recv_time;
611 newest_off = oldest_off;
612 newest_age = oldest_age;
619 idx = (idx - 1) & (NUM_DATAPOINTS - 1);
621 p->filter_dispersion = sum;
622 wavg += x; /* add another older6/64 to form older6/32 */
623 /* Fix systematic underestimation with large poll intervals.
624 * Imagine that we still have a bit of uncorrected drift,
625 * and poll interval is big (say, 100 sec). Offsets form a progression:
626 * 0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 - 0.7 is most recent.
627 * The algorithm above drops 0.0 and 0.7 as outliers,
628 * and then we have this estimation, ~25% off from 0.7:
629 * 0.1/32 + 0.2/32 + 0.3/16 + 0.4/8 + 0.5/4 + 0.6/2 = 0.503125
631 x = oldest_age - newest_age;
633 x = newest_age / x; /* in above example, 100 / (600 - 100) */
634 if (x < 1) { /* paranoia check */
635 x = (newest_off - oldest_off) * x; /* 0.5 * 100/500 = 0.1 */
639 p->filter_offset = wavg;
643 fdp = p->filter_datapoint;
644 idx = p->datapoint_idx; /* most recent datapoint's index */
646 /* filter_offset: simply use the most recent value */
647 p->filter_offset = fdp[idx].d_offset;
651 * filter_dispersion = \ -------------
658 for (i = 0; i < NUM_DATAPOINTS; i++) {
659 sum += dispersion(&fdp[idx]) / (2 << i);
660 wavg += fdp[idx].d_offset;
661 idx = (idx - 1) & (NUM_DATAPOINTS - 1);
663 wavg /= NUM_DATAPOINTS;
664 p->filter_dispersion = sum;
667 /* +----- -----+ ^ 1/2
671 * filter_jitter = | --- * / (avg-offset_j) |
675 * where n is the number of valid datapoints in the filter (n > 1);
676 * if filter_jitter < precision then filter_jitter = precision
679 for (i = 0; i < NUM_DATAPOINTS; i++) {
680 sum += SQUARE(wavg - fdp[i].d_offset);
682 sum = SQRT(sum / NUM_DATAPOINTS);
683 p->filter_jitter = sum > G_precision_sec ? sum : G_precision_sec;
685 VERB4 bb_error_msg("filter offset:%+f disp:%f jitter:%f",
687 p->filter_dispersion,
692 reset_peer_stats(peer_t *p, double offset)
695 bool small_ofs = fabs(offset) < 16 * STEP_THRESHOLD;
697 /* Used to set p->filter_datapoint[i].d_dispersion = MAXDISP
698 * and clear reachable bits, but this proved to be too agressive:
699 * after step (tested with suspinding laptop for ~30 secs),
700 * this caused all previous data to be considered invalid,
701 * making us needing to collect full ~8 datapoins per peer
702 * after step in order to start trusting them.
703 * In turn, this was making poll interval decrease even after
704 * step was done. (Poll interval decreases already before step
705 * in this scenario, because we see large offsets and end up with
706 * no good peer to select).
709 for (i = 0; i < NUM_DATAPOINTS; i++) {
711 p->filter_datapoint[i].d_recv_time += offset;
712 if (p->filter_datapoint[i].d_offset != 0) {
713 p->filter_datapoint[i].d_offset -= offset;
714 //bb_error_msg("p->filter_datapoint[%d].d_offset %f -> %f",
716 // p->filter_datapoint[i].d_offset + offset,
717 // p->filter_datapoint[i].d_offset);
720 p->filter_datapoint[i].d_recv_time = G.cur_time;
721 p->filter_datapoint[i].d_offset = 0;
722 /*p->filter_datapoint[i].d_dispersion = MAXDISP;*/
726 p->lastpkt_recv_time += offset;
728 /*p->reachable_bits = 0;*/
729 p->lastpkt_recv_time = G.cur_time;
731 filter_datapoints(p); /* recalc p->filter_xxx */
732 VERB6 bb_error_msg("%s->lastpkt_recv_time=%f", p->p_dotted, p->lastpkt_recv_time);
740 p = xzalloc(sizeof(*p));
741 p->p_lsa = xhost2sockaddr(s, 123);
742 p->p_dotted = xmalloc_sockaddr2dotted_noport(&p->p_lsa->u.sa);
744 p->p_xmt_msg.m_status = MODE_CLIENT | (NTP_VERSION << 3);
745 p->next_action_time = G.cur_time; /* = set_next(p, 0); */
746 reset_peer_stats(p, 16 * STEP_THRESHOLD);
748 llist_add_to(&G.ntp_peers, p);
754 const struct sockaddr *from, const struct sockaddr *to, socklen_t addrlen,
755 msg_t *msg, ssize_t len)
761 ret = sendto(fd, msg, len, MSG_DONTWAIT, to, addrlen);
763 ret = send_to_from(fd, msg, len, MSG_DONTWAIT, to, from, addrlen);
766 bb_perror_msg("send failed");
773 send_query_to_peer(peer_t *p)
775 /* Why do we need to bind()?
776 * See what happens when we don't bind:
778 * socket(PF_INET, SOCK_DGRAM, IPPROTO_IP) = 3
779 * setsockopt(3, SOL_IP, IP_TOS, [16], 4) = 0
780 * gettimeofday({1259071266, 327885}, NULL) = 0
781 * sendto(3, "xxx", 48, MSG_DONTWAIT, {sa_family=AF_INET, sin_port=htons(123), sin_addr=inet_addr("10.34.32.125")}, 16) = 48
782 * ^^^ we sent it from some source port picked by kernel.
783 * time(NULL) = 1259071266
784 * write(2, "ntpd: entering poll 15 secs\n", 28) = 28
785 * poll([{fd=3, events=POLLIN}], 1, 15000) = 1 ([{fd=3, revents=POLLIN}])
786 * recv(3, "yyy", 68, MSG_DONTWAIT) = 48
787 * ^^^ this recv will receive packets to any local port!
789 * Uncomment this and use strace to see it in action:
791 #define PROBE_LOCAL_ADDR /* { len_and_sockaddr lsa; lsa.len = LSA_SIZEOF_SA; getsockname(p->query.fd, &lsa.u.sa, &lsa.len); } */
795 len_and_sockaddr *local_lsa;
797 family = p->p_lsa->u.sa.sa_family;
798 p->p_fd = fd = xsocket_type(&local_lsa, family, SOCK_DGRAM);
799 /* local_lsa has "null" address and port 0 now.
800 * bind() ensures we have a *particular port* selected by kernel
801 * and remembered in p->p_fd, thus later recv(p->p_fd)
802 * receives only packets sent to this port.
805 xbind(fd, &local_lsa->u.sa, local_lsa->len);
807 #if ENABLE_FEATURE_IPV6
808 if (family == AF_INET)
810 setsockopt(fd, IPPROTO_IP, IP_TOS, &const_IPTOS_LOWDELAY, sizeof(const_IPTOS_LOWDELAY));
814 /* Emit message _before_ attempted send. Think of a very short
815 * roundtrip networks: we need to go back to recv loop ASAP,
816 * to reduce delay. Printing messages after send works against that.
818 VERB1 bb_error_msg("sending query to %s", p->p_dotted);
821 * Send out a random 64-bit number as our transmit time. The NTP
822 * server will copy said number into the originate field on the
823 * response that it sends us. This is totally legal per the SNTP spec.
825 * The impact of this is two fold: we no longer send out the current
826 * system time for the world to see (which may aid an attacker), and
827 * it gives us a (not very secure) way of knowing that we're not
828 * getting spoofed by an attacker that can't capture our traffic
829 * but can spoof packets from the NTP server we're communicating with.
831 * Save the real transmit timestamp locally.
833 p->p_xmt_msg.m_xmttime.int_partl = random();
834 p->p_xmt_msg.m_xmttime.fractionl = random();
835 p->p_xmttime = gettime1900d();
837 /* Were doing it only if sendto worked, but
838 * loss of sync detection needs reachable_bits updated
839 * even if sending fails *locally*:
840 * "network is unreachable" because cable was pulled?
841 * We still need to declare "unsync" if this condition persists.
843 p->reachable_bits <<= 1;
845 if (do_sendto(p->p_fd, /*from:*/ NULL, /*to:*/ &p->p_lsa->u.sa, /*addrlen:*/ p->p_lsa->len,
846 &p->p_xmt_msg, NTP_MSGSIZE_NOAUTH) == -1
851 * We know that we sent nothing.
852 * We can retry *soon* without fearing
853 * that we are flooding the peer.
855 set_next(p, RETRY_INTERVAL);
859 set_next(p, RESPONSE_INTERVAL);
863 /* Note that there is no provision to prevent several run_scripts
864 * to be started in quick succession. In fact, it happens rather often
865 * if initial syncronization results in a step.
866 * You will see "step" and then "stratum" script runs, sometimes
867 * as close as only 0.002 seconds apart.
868 * Script should be ready to deal with this.
870 static void run_script(const char *action, double offset)
873 char *env1, *env2, *env3, *env4;
875 G.last_script_run = G.cur_time;
880 argv[0] = (char*) G.script_name;
881 argv[1] = (char*) action;
884 VERB1 bb_error_msg("executing '%s %s'", G.script_name, action);
886 env1 = xasprintf("%s=%u", "stratum", G.stratum);
888 env2 = xasprintf("%s=%ld", "freq_drift_ppm", G.kernel_freq_drift);
890 env3 = xasprintf("%s=%u", "poll_interval", 1 << G.poll_exp);
892 env4 = xasprintf("%s=%f", "offset", offset);
894 /* Other items of potential interest: selected peer,
895 * rootdelay, reftime, rootdisp, refid, ntp_status,
896 * last_update_offset, last_update_recv_time, discipline_jitter,
897 * how many peers have reachable_bits = 0?
900 /* Don't want to wait: it may run hwclock --systohc, and that
901 * may take some time (seconds): */
902 /*spawn_and_wait(argv);*/
906 unsetenv("freq_drift_ppm");
907 unsetenv("poll_interval");
916 step_time(double offset)
920 struct timeval tvc, tvn;
921 char buf[sizeof("yyyy-mm-dd hh:mm:ss") + /*paranoia:*/ 4];
924 gettimeofday(&tvc, NULL); /* never fails */
925 dtime = tvc.tv_sec + (1.0e-6 * tvc.tv_usec) + offset;
926 d_to_tv(dtime, &tvn);
927 if (settimeofday(&tvn, NULL) == -1)
928 bb_perror_msg_and_die("settimeofday");
932 strftime_YYYYMMDDHHMMSS(buf, sizeof(buf), &tval);
933 bb_error_msg("current time is %s.%06u", buf, (unsigned)tvc.tv_usec);
936 strftime_YYYYMMDDHHMMSS(buf, sizeof(buf), &tval);
937 bb_error_msg("setting time to %s.%06u (offset %+fs)", buf, (unsigned)tvn.tv_usec, offset);
939 /* Correct various fields which contain time-relative values: */
942 G.cur_time += offset;
943 G.last_update_recv_time += offset;
944 G.last_script_run += offset;
946 /* p->lastpkt_recv_time, p->next_action_time and such: */
947 for (item = G.ntp_peers; item != NULL; item = item->link) {
948 peer_t *pp = (peer_t *) item->data;
949 reset_peer_stats(pp, offset);
950 //bb_error_msg("offset:%+f pp->next_action_time:%f -> %f",
951 // offset, pp->next_action_time, pp->next_action_time + offset);
952 pp->next_action_time += offset;
954 /* We wait for reply from this peer too.
955 * But due to step we are doing, reply's data is no longer
956 * useful (in fact, it'll be bogus). Stop waiting for it.
960 set_next(pp, RETRY_INTERVAL);
967 * Selection and clustering, and their helpers
973 double opt_rd; /* optimization */
976 compare_point_edge(const void *aa, const void *bb)
978 const point_t *a = aa;
979 const point_t *b = bb;
980 if (a->edge < b->edge) {
983 return (a->edge > b->edge);
990 compare_survivor_metric(const void *aa, const void *bb)
992 const survivor_t *a = aa;
993 const survivor_t *b = bb;
994 if (a->metric < b->metric) {
997 return (a->metric > b->metric);
1000 fit(peer_t *p, double rd)
1002 if ((p->reachable_bits & (p->reachable_bits-1)) == 0) {
1003 /* One or zero bits in reachable_bits */
1004 VERB4 bb_error_msg("peer %s unfit for selection: unreachable", p->p_dotted);
1007 #if 0 /* we filter out such packets earlier */
1008 if ((p->lastpkt_status & LI_ALARM) == LI_ALARM
1009 || p->lastpkt_stratum >= MAXSTRAT
1011 VERB4 bb_error_msg("peer %s unfit for selection: bad status/stratum", p->p_dotted);
1015 /* rd is root_distance(p) */
1016 if (rd > MAXDIST + FREQ_TOLERANCE * (1 << G.poll_exp)) {
1017 VERB4 bb_error_msg("peer %s unfit for selection: root distance too high", p->p_dotted);
1021 // /* Do we have a loop? */
1022 // if (p->refid == p->dstaddr || p->refid == s.refid)
1027 select_and_cluster(void)
1032 int size = 3 * G.peer_cnt;
1033 /* for selection algorithm */
1034 point_t point[size];
1035 unsigned num_points, num_candidates;
1037 unsigned num_falsetickers;
1038 /* for cluster algorithm */
1039 survivor_t survivor[size];
1040 unsigned num_survivors;
1046 if (G.initial_poll_complete) while (item != NULL) {
1049 p = (peer_t *) item->data;
1050 rd = root_distance(p);
1051 offset = p->filter_offset;
1057 VERB5 bb_error_msg("interval: [%f %f %f] %s",
1063 point[num_points].p = p;
1064 point[num_points].type = -1;
1065 point[num_points].edge = offset - rd;
1066 point[num_points].opt_rd = rd;
1068 point[num_points].p = p;
1069 point[num_points].type = 0;
1070 point[num_points].edge = offset;
1071 point[num_points].opt_rd = rd;
1073 point[num_points].p = p;
1074 point[num_points].type = 1;
1075 point[num_points].edge = offset + rd;
1076 point[num_points].opt_rd = rd;
1080 num_candidates = num_points / 3;
1081 if (num_candidates == 0) {
1082 VERB3 bb_error_msg("no valid datapoints%s", ", no peer selected");
1085 //TODO: sorting does not seem to be done in reference code
1086 qsort(point, num_points, sizeof(point[0]), compare_point_edge);
1088 /* Start with the assumption that there are no falsetickers.
1089 * Attempt to find a nonempty intersection interval containing
1090 * the midpoints of all truechimers.
1091 * If a nonempty interval cannot be found, increase the number
1092 * of assumed falsetickers by one and try again.
1093 * If a nonempty interval is found and the number of falsetickers
1094 * is less than the number of truechimers, a majority has been found
1095 * and the midpoint of each truechimer represents
1096 * the candidates available to the cluster algorithm.
1098 num_falsetickers = 0;
1101 unsigned num_midpoints = 0;
1106 for (i = 0; i < num_points; i++) {
1108 * if (point[i].type == -1) c++;
1109 * if (point[i].type == 1) c--;
1110 * and it's simpler to do it this way:
1113 if (c >= num_candidates - num_falsetickers) {
1114 /* If it was c++ and it got big enough... */
1115 low = point[i].edge;
1118 if (point[i].type == 0)
1122 for (i = num_points-1; i >= 0; i--) {
1124 if (c >= num_candidates - num_falsetickers) {
1125 high = point[i].edge;
1128 if (point[i].type == 0)
1131 /* If the number of midpoints is greater than the number
1132 * of allowed falsetickers, the intersection contains at
1133 * least one truechimer with no midpoint - bad.
1134 * Also, interval should be nonempty.
1136 if (num_midpoints <= num_falsetickers && low < high)
1139 if (num_falsetickers * 2 >= num_candidates) {
1140 VERB3 bb_error_msg("falsetickers:%d, candidates:%d%s",
1141 num_falsetickers, num_candidates,
1142 ", no peer selected");
1146 VERB4 bb_error_msg("selected interval: [%f, %f]; candidates:%d falsetickers:%d",
1147 low, high, num_candidates, num_falsetickers);
1151 /* Construct a list of survivors (p, metric)
1152 * from the chime list, where metric is dominated
1153 * first by stratum and then by root distance.
1154 * All other things being equal, this is the order of preference.
1157 for (i = 0; i < num_points; i++) {
1158 if (point[i].edge < low || point[i].edge > high)
1161 survivor[num_survivors].p = p;
1162 /* x.opt_rd == root_distance(p); */
1163 survivor[num_survivors].metric = MAXDIST * p->lastpkt_stratum + point[i].opt_rd;
1164 VERB5 bb_error_msg("survivor[%d] metric:%f peer:%s",
1165 num_survivors, survivor[num_survivors].metric, p->p_dotted);
1168 /* There must be at least MIN_SELECTED survivors to satisfy the
1169 * correctness assertions. Ordinarily, the Byzantine criteria
1170 * require four survivors, but for the demonstration here, one
1173 if (num_survivors < MIN_SELECTED) {
1174 VERB3 bb_error_msg("survivors:%d%s",
1176 ", no peer selected");
1180 //looks like this is ONLY used by the fact that later we pick survivor[0].
1181 //we can avoid sorting then, just find the minimum once!
1182 qsort(survivor, num_survivors, sizeof(survivor[0]), compare_survivor_metric);
1184 /* For each association p in turn, calculate the selection
1185 * jitter p->sjitter as the square root of the sum of squares
1186 * (p->offset - q->offset) over all q associations. The idea is
1187 * to repeatedly discard the survivor with maximum selection
1188 * jitter until a termination condition is met.
1191 unsigned max_idx = max_idx;
1192 double max_selection_jitter = max_selection_jitter;
1193 double min_jitter = min_jitter;
1195 if (num_survivors <= MIN_CLUSTERED) {
1196 VERB4 bb_error_msg("num_survivors %d <= %d, not discarding more",
1197 num_survivors, MIN_CLUSTERED);
1201 /* To make sure a few survivors are left
1202 * for the clustering algorithm to chew on,
1203 * we stop if the number of survivors
1204 * is less than or equal to MIN_CLUSTERED (3).
1206 for (i = 0; i < num_survivors; i++) {
1207 double selection_jitter_sq;
1210 if (i == 0 || p->filter_jitter < min_jitter)
1211 min_jitter = p->filter_jitter;
1213 selection_jitter_sq = 0;
1214 for (j = 0; j < num_survivors; j++) {
1215 peer_t *q = survivor[j].p;
1216 selection_jitter_sq += SQUARE(p->filter_offset - q->filter_offset);
1218 if (i == 0 || selection_jitter_sq > max_selection_jitter) {
1219 max_selection_jitter = selection_jitter_sq;
1222 VERB6 bb_error_msg("survivor %d selection_jitter^2:%f",
1223 i, selection_jitter_sq);
1225 max_selection_jitter = SQRT(max_selection_jitter / num_survivors);
1226 VERB5 bb_error_msg("max_selection_jitter (at %d):%f min_jitter:%f",
1227 max_idx, max_selection_jitter, min_jitter);
1229 /* If the maximum selection jitter is less than the
1230 * minimum peer jitter, then tossing out more survivors
1231 * will not lower the minimum peer jitter, so we might
1234 if (max_selection_jitter < min_jitter) {
1235 VERB4 bb_error_msg("max_selection_jitter:%f < min_jitter:%f, num_survivors:%d, not discarding more",
1236 max_selection_jitter, min_jitter, num_survivors);
1240 /* Delete survivor[max_idx] from the list
1241 * and go around again.
1243 VERB6 bb_error_msg("dropping survivor %d", max_idx);
1245 while (max_idx < num_survivors) {
1246 survivor[max_idx] = survivor[max_idx + 1];
1252 /* Combine the offsets of the clustering algorithm survivors
1253 * using a weighted average with weight determined by the root
1254 * distance. Compute the selection jitter as the weighted RMS
1255 * difference between the first survivor and the remaining
1256 * survivors. In some cases the inherent clock jitter can be
1257 * reduced by not using this algorithm, especially when frequent
1258 * clockhopping is involved. bbox: thus we don't do it.
1262 for (i = 0; i < num_survivors; i++) {
1264 x = root_distance(p);
1266 z += p->filter_offset / x;
1267 w += SQUARE(p->filter_offset - survivor[0].p->filter_offset) / x;
1269 //G.cluster_offset = z / y;
1270 //G.cluster_jitter = SQRT(w / y);
1273 /* Pick the best clock. If the old system peer is on the list
1274 * and at the same stratum as the first survivor on the list,
1275 * then don't do a clock hop. Otherwise, select the first
1276 * survivor on the list as the new system peer.
1279 if (G.last_update_peer
1280 && G.last_update_peer->lastpkt_stratum <= p->lastpkt_stratum
1282 /* Starting from 1 is ok here */
1283 for (i = 1; i < num_survivors; i++) {
1284 if (G.last_update_peer == survivor[i].p) {
1285 VERB5 bb_error_msg("keeping old synced peer");
1286 p = G.last_update_peer;
1291 G.last_update_peer = p;
1293 VERB4 bb_error_msg("selected peer %s filter_offset:%+f age:%f",
1296 G.cur_time - p->lastpkt_recv_time
1303 * Local clock discipline and its helpers
1306 set_new_values(int disc_state, double offset, double recv_time)
1308 /* Enter new state and set state variables. Note we use the time
1309 * of the last clock filter sample, which must be earlier than
1312 VERB4 bb_error_msg("disc_state=%d last update offset=%f recv_time=%f",
1313 disc_state, offset, recv_time);
1314 G.discipline_state = disc_state;
1315 G.last_update_offset = offset;
1316 G.last_update_recv_time = recv_time;
1318 /* Return: -1: decrease poll interval, 0: leave as is, 1: increase */
1320 update_local_clock(peer_t *p)
1324 /* Note: can use G.cluster_offset instead: */
1325 double offset = p->filter_offset;
1326 double recv_time = p->lastpkt_recv_time;
1328 #if !USING_KERNEL_PLL_LOOP
1331 double since_last_update;
1332 double etemp, dtemp;
1334 abs_offset = fabs(offset);
1337 /* If needed, -S script can do it by looking at $offset
1338 * env var and killing parent */
1339 /* If the offset is too large, give up and go home */
1340 if (abs_offset > PANIC_THRESHOLD) {
1341 bb_error_msg_and_die("offset %f far too big, exiting", offset);
1345 /* If this is an old update, for instance as the result
1346 * of a system peer change, avoid it. We never use
1347 * an old sample or the same sample twice.
1349 if (recv_time <= G.last_update_recv_time) {
1350 VERB3 bb_error_msg("update from %s: same or older datapoint, not using it",
1352 return 0; /* "leave poll interval as is" */
1355 /* Clock state machine transition function. This is where the
1356 * action is and defines how the system reacts to large time
1357 * and frequency errors.
1359 since_last_update = recv_time - G.reftime;
1360 #if !USING_KERNEL_PLL_LOOP
1363 #if USING_INITIAL_FREQ_ESTIMATION
1364 if (G.discipline_state == STATE_FREQ) {
1365 /* Ignore updates until the stepout threshold */
1366 if (since_last_update < WATCH_THRESHOLD) {
1367 VERB4 bb_error_msg("measuring drift, datapoint ignored, %f sec remains",
1368 WATCH_THRESHOLD - since_last_update);
1369 return 0; /* "leave poll interval as is" */
1371 # if !USING_KERNEL_PLL_LOOP
1372 freq_drift = (offset - G.last_update_offset) / since_last_update;
1377 /* There are two main regimes: when the
1378 * offset exceeds the step threshold and when it does not.
1380 if (abs_offset > STEP_THRESHOLD) {
1384 // This "spike state" seems to be useless, peer selection already drops
1385 // occassional "bad" datapoints. If we are here, there were _many_
1386 // large offsets. When a few first large offsets are seen,
1387 // we end up in "no valid datapoints, no peer selected" state.
1388 // Only when enough of them are seen (which means it's not a fluke),
1389 // we end up here. Looks like _our_ clock is off.
1390 switch (G.discipline_state) {
1392 /* The first outlyer: ignore it, switch to SPIK state */
1393 VERB3 bb_error_msg("update from %s: offset:%+f, spike%s",
1394 p->p_dotted, offset,
1396 G.discipline_state = STATE_SPIK;
1397 return -1; /* "decrease poll interval" */
1400 /* Ignore succeeding outlyers until either an inlyer
1401 * is found or the stepout threshold is exceeded.
1403 remains = WATCH_THRESHOLD - since_last_update;
1405 VERB3 bb_error_msg("update from %s: offset:%+f, spike%s",
1406 p->p_dotted, offset,
1407 ", datapoint ignored");
1408 return -1; /* "decrease poll interval" */
1410 /* fall through: we need to step */
1414 /* Step the time and clamp down the poll interval.
1416 * In NSET state an initial frequency correction is
1417 * not available, usually because the frequency file has
1418 * not yet been written. Since the time is outside the
1419 * capture range, the clock is stepped. The frequency
1420 * will be set directly following the stepout interval.
1422 * In FSET state the initial frequency has been set
1423 * from the frequency file. Since the time is outside
1424 * the capture range, the clock is stepped immediately,
1425 * rather than after the stepout interval. Guys get
1426 * nervous if it takes 17 minutes to set the clock for
1429 * In SPIK state the stepout threshold has expired and
1430 * the phase is still above the step threshold. Note
1431 * that a single spike greater than the step threshold
1432 * is always suppressed, even at the longer poll
1435 VERB4 bb_error_msg("stepping time by %+f; poll_exp=MINPOLL", offset);
1437 if (option_mask32 & OPT_q) {
1438 /* We were only asked to set time once. Done. */
1442 G.polladj_count = 0;
1443 G.poll_exp = MINPOLL;
1444 G.stratum = MAXSTRAT;
1446 run_script("step", offset);
1448 recv_time += offset;
1450 #if USING_INITIAL_FREQ_ESTIMATION
1451 if (G.discipline_state == STATE_NSET) {
1452 set_new_values(STATE_FREQ, /*offset:*/ 0, recv_time);
1453 return 1; /* "ok to increase poll interval" */
1456 abs_offset = offset = 0;
1457 set_new_values(STATE_SYNC, offset, recv_time);
1459 } else { /* abs_offset <= STEP_THRESHOLD */
1461 if (G.poll_exp < MINPOLL && G.initial_poll_complete) {
1462 VERB4 bb_error_msg("small offset:%+f, disabling burst mode", offset);
1463 G.polladj_count = 0;
1464 G.poll_exp = MINPOLL;
1467 /* Compute the clock jitter as the RMS of exponentially
1468 * weighted offset differences. Used by the poll adjust code.
1470 etemp = SQUARE(G.discipline_jitter);
1471 dtemp = SQUARE(offset - G.last_update_offset);
1472 G.discipline_jitter = SQRT(etemp + (dtemp - etemp) / AVG);
1474 switch (G.discipline_state) {
1476 if (option_mask32 & OPT_q) {
1477 /* We were only asked to set time once.
1478 * The clock is precise enough, no need to step.
1482 #if USING_INITIAL_FREQ_ESTIMATION
1483 /* This is the first update received and the frequency
1484 * has not been initialized. The first thing to do
1485 * is directly measure the oscillator frequency.
1487 set_new_values(STATE_FREQ, offset, recv_time);
1489 set_new_values(STATE_SYNC, offset, recv_time);
1491 VERB4 bb_error_msg("transitioning to FREQ, datapoint ignored");
1492 return 0; /* "leave poll interval as is" */
1494 #if 0 /* this is dead code for now */
1496 /* This is the first update and the frequency
1497 * has been initialized. Adjust the phase, but
1498 * don't adjust the frequency until the next update.
1500 set_new_values(STATE_SYNC, offset, recv_time);
1501 /* freq_drift remains 0 */
1505 #if USING_INITIAL_FREQ_ESTIMATION
1507 /* since_last_update >= WATCH_THRESHOLD, we waited enough.
1508 * Correct the phase and frequency and switch to SYNC state.
1509 * freq_drift was already estimated (see code above)
1511 set_new_values(STATE_SYNC, offset, recv_time);
1516 #if !USING_KERNEL_PLL_LOOP
1517 /* Compute freq_drift due to PLL and FLL contributions.
1519 * The FLL and PLL frequency gain constants
1520 * depend on the poll interval and Allan
1521 * intercept. The FLL is not used below one-half
1522 * the Allan intercept. Above that the loop gain
1523 * increases in steps to 1 / AVG.
1525 if ((1 << G.poll_exp) > ALLAN / 2) {
1526 etemp = FLL - G.poll_exp;
1529 freq_drift += (offset - G.last_update_offset) / (MAXD(since_last_update, ALLAN) * etemp);
1531 /* For the PLL the integration interval
1532 * (numerator) is the minimum of the update
1533 * interval and poll interval. This allows
1534 * oversampling, but not undersampling.
1536 etemp = MIND(since_last_update, (1 << G.poll_exp));
1537 dtemp = (4 * PLL) << G.poll_exp;
1538 freq_drift += offset * etemp / SQUARE(dtemp);
1540 set_new_values(STATE_SYNC, offset, recv_time);
1543 if (G.stratum != p->lastpkt_stratum + 1) {
1544 G.stratum = p->lastpkt_stratum + 1;
1545 run_script("stratum", offset);
1549 if (G.discipline_jitter < G_precision_sec)
1550 G.discipline_jitter = G_precision_sec;
1551 G.offset_to_jitter_ratio = abs_offset / G.discipline_jitter;
1553 G.reftime = G.cur_time;
1554 G.ntp_status = p->lastpkt_status;
1555 G.refid = p->lastpkt_refid;
1556 G.rootdelay = p->lastpkt_rootdelay + p->lastpkt_delay;
1557 dtemp = p->filter_jitter; // SQRT(SQUARE(p->filter_jitter) + SQUARE(G.cluster_jitter));
1558 dtemp += MAXD(p->filter_dispersion + FREQ_TOLERANCE * (G.cur_time - p->lastpkt_recv_time) + abs_offset, MINDISP);
1559 G.rootdisp = p->lastpkt_rootdisp + dtemp;
1560 VERB4 bb_error_msg("updating leap/refid/reftime/rootdisp from peer %s", p->p_dotted);
1562 /* We are in STATE_SYNC now, but did not do adjtimex yet.
1563 * (Any other state does not reach this, they all return earlier)
1564 * By this time, freq_drift and offset are set
1565 * to values suitable for adjtimex.
1567 #if !USING_KERNEL_PLL_LOOP
1568 /* Calculate the new frequency drift and frequency stability (wander).
1569 * Compute the clock wander as the RMS of exponentially weighted
1570 * frequency differences. This is not used directly, but can,
1571 * along with the jitter, be a highly useful monitoring and
1574 dtemp = G.discipline_freq_drift + freq_drift;
1575 G.discipline_freq_drift = MAXD(MIND(MAXDRIFT, dtemp), -MAXDRIFT);
1576 etemp = SQUARE(G.discipline_wander);
1577 dtemp = SQUARE(dtemp);
1578 G.discipline_wander = SQRT(etemp + (dtemp - etemp) / AVG);
1580 VERB4 bb_error_msg("discipline freq_drift=%.9f(int:%ld corr:%e) wander=%f",
1581 G.discipline_freq_drift,
1582 (long)(G.discipline_freq_drift * 65536e6),
1584 G.discipline_wander);
1587 memset(&tmx, 0, sizeof(tmx));
1588 if (adjtimex(&tmx) < 0)
1589 bb_perror_msg_and_die("adjtimex");
1590 bb_error_msg("p adjtimex freq:%ld offset:%+ld status:0x%x tc:%ld",
1591 tmx.freq, tmx.offset, tmx.status, tmx.constant);
1594 memset(&tmx, 0, sizeof(tmx));
1596 //doesn't work, offset remains 0 (!) in kernel:
1597 //ntpd: set adjtimex freq:1786097 tmx.offset:77487
1598 //ntpd: prev adjtimex freq:1786097 tmx.offset:0
1599 //ntpd: cur adjtimex freq:1786097 tmx.offset:0
1600 tmx.modes = ADJ_FREQUENCY | ADJ_OFFSET;
1601 /* 65536 is one ppm */
1602 tmx.freq = G.discipline_freq_drift * 65536e6;
1604 tmx.modes = ADJ_OFFSET | ADJ_STATUS | ADJ_TIMECONST;// | ADJ_MAXERROR | ADJ_ESTERROR;
1605 tmx.offset = (offset * 1000000); /* usec */
1606 tmx.status = STA_PLL;
1607 if (G.ntp_status & LI_PLUSSEC)
1608 tmx.status |= STA_INS;
1609 if (G.ntp_status & LI_MINUSSEC)
1610 tmx.status |= STA_DEL;
1612 tmx.constant = G.poll_exp - 4;
1614 * The below if statement should be unnecessary, but...
1615 * It looks like Linux kernel's PLL is far too gentle in changing
1616 * tmx.freq in response to clock offset. Offset keeps growing
1617 * and eventually we fall back to smaller poll intervals.
1618 * We can make correction more agressive (about x2) by supplying
1619 * PLL time constant which is one less than the real one.
1620 * To be on a safe side, let's do it only if offset is significantly
1621 * larger than jitter.
1623 if (tmx.constant > 0 && G.offset_to_jitter_ratio >= TIMECONST_HACK_GATE)
1626 //tmx.esterror = (uint32_t)(clock_jitter * 1e6);
1627 //tmx.maxerror = (uint32_t)((sys_rootdelay / 2 + sys_rootdisp) * 1e6);
1628 rc = adjtimex(&tmx);
1630 bb_perror_msg_and_die("adjtimex");
1631 /* NB: here kernel returns constant == G.poll_exp, not == G.poll_exp - 4.
1632 * Not sure why. Perhaps it is normal.
1634 VERB4 bb_error_msg("adjtimex:%d freq:%ld offset:%+ld status:0x%x",
1635 rc, tmx.freq, tmx.offset, tmx.status);
1636 G.kernel_freq_drift = tmx.freq / 65536;
1637 VERB2 bb_error_msg("update from:%s offset:%+f jitter:%f clock drift:%+.3fppm tc:%d",
1638 p->p_dotted, offset, G.discipline_jitter, (double)tmx.freq / 65536, (int)tmx.constant);
1640 return 1; /* "ok to increase poll interval" */
1645 * We've got a new reply packet from a peer, process it
1649 retry_interval(void)
1651 /* Local problem, want to retry soon */
1652 unsigned interval, r;
1653 interval = RETRY_INTERVAL;
1655 interval += r % (unsigned)(RETRY_INTERVAL / 4);
1656 VERB4 bb_error_msg("chose retry interval:%u", interval);
1660 poll_interval(int exponent)
1662 unsigned interval, r;
1663 exponent = G.poll_exp + exponent;
1666 interval = 1 << exponent;
1668 interval += ((r & (interval-1)) >> 4) + ((r >> 8) & 1); /* + 1/16 of interval, max */
1669 VERB4 bb_error_msg("chose poll interval:%u (poll_exp:%d exp:%d)", interval, G.poll_exp, exponent);
1672 static NOINLINE void
1673 recv_and_process_peer_pkt(peer_t *p)
1678 double T1, T2, T3, T4;
1681 datapoint_t *datapoint;
1686 /* We can recvfrom here and check from.IP, but some multihomed
1687 * ntp servers reply from their *other IP*.
1688 * TODO: maybe we should check at least what we can: from.port == 123?
1690 size = recv(p->p_fd, &msg, sizeof(msg), MSG_DONTWAIT);
1692 bb_perror_msg("recv(%s) error", p->p_dotted);
1693 if (errno == EHOSTUNREACH || errno == EHOSTDOWN
1694 || errno == ENETUNREACH || errno == ENETDOWN
1695 || errno == ECONNREFUSED || errno == EADDRNOTAVAIL
1698 //TODO: always do this?
1699 interval = retry_interval();
1700 goto set_next_and_ret;
1705 if (size != NTP_MSGSIZE_NOAUTH && size != NTP_MSGSIZE) {
1706 bb_error_msg("malformed packet received from %s", p->p_dotted);
1710 if (msg.m_orgtime.int_partl != p->p_xmt_msg.m_xmttime.int_partl
1711 || msg.m_orgtime.fractionl != p->p_xmt_msg.m_xmttime.fractionl
1713 /* Somebody else's packet */
1717 /* We do not expect any more packets from this peer for now.
1718 * Closing the socket informs kernel about it.
1719 * We open a new socket when we send a new query.
1724 if ((msg.m_status & LI_ALARM) == LI_ALARM
1725 || msg.m_stratum == 0
1726 || msg.m_stratum > NTP_MAXSTRATUM
1728 // TODO: stratum 0 responses may have commands in 32-bit m_refid field:
1729 // "DENY", "RSTR" - peer does not like us at all
1730 // "RATE" - peer is overloaded, reduce polling freq
1731 bb_error_msg("reply from %s: peer is unsynced", p->p_dotted);
1732 goto pick_normal_interval;
1735 // /* Verify valid root distance */
1736 // if (msg.m_rootdelay / 2 + msg.m_rootdisp >= MAXDISP || p->lastpkt_reftime > msg.m_xmt)
1737 // return; /* invalid header values */
1739 p->lastpkt_status = msg.m_status;
1740 p->lastpkt_stratum = msg.m_stratum;
1741 p->lastpkt_rootdelay = sfp_to_d(msg.m_rootdelay);
1742 p->lastpkt_rootdisp = sfp_to_d(msg.m_rootdisp);
1743 p->lastpkt_refid = msg.m_refid;
1746 * From RFC 2030 (with a correction to the delay math):
1748 * Timestamp Name ID When Generated
1749 * ------------------------------------------------------------
1750 * Originate Timestamp T1 time request sent by client
1751 * Receive Timestamp T2 time request received by server
1752 * Transmit Timestamp T3 time reply sent by server
1753 * Destination Timestamp T4 time reply received by client
1755 * The roundtrip delay and local clock offset are defined as
1757 * delay = (T4 - T1) - (T3 - T2); offset = ((T2 - T1) + (T3 - T4)) / 2
1760 T2 = lfp_to_d(msg.m_rectime);
1761 T3 = lfp_to_d(msg.m_xmttime);
1764 p->lastpkt_recv_time = T4;
1765 VERB6 bb_error_msg("%s->lastpkt_recv_time=%f", p->p_dotted, p->lastpkt_recv_time);
1767 /* The delay calculation is a special case. In cases where the
1768 * server and client clocks are running at different rates and
1769 * with very fast networks, the delay can appear negative. In
1770 * order to avoid violating the Principle of Least Astonishment,
1771 * the delay is clamped not less than the system precision.
1773 dv = p->lastpkt_delay;
1774 p->lastpkt_delay = (T4 - T1) - (T3 - T2);
1775 if (p->lastpkt_delay < G_precision_sec)
1776 p->lastpkt_delay = G_precision_sec;
1778 * If this packet's delay is much bigger than the last one,
1779 * it's better to just ignore it than use its much less precise value.
1781 if (p->reachable_bits && p->lastpkt_delay > dv * BAD_DELAY_GROWTH) {
1782 bb_error_msg("reply from %s: delay %f is too high, ignoring", p->p_dotted, p->lastpkt_delay);
1783 goto pick_normal_interval;
1786 p->datapoint_idx = p->reachable_bits ? (p->datapoint_idx + 1) % NUM_DATAPOINTS : 0;
1787 datapoint = &p->filter_datapoint[p->datapoint_idx];
1788 datapoint->d_recv_time = T4;
1789 datapoint->d_offset = offset = ((T2 - T1) + (T3 - T4)) / 2;
1790 datapoint->d_dispersion = LOG2D(msg.m_precision_exp) + G_precision_sec;
1791 if (!p->reachable_bits) {
1792 /* 1st datapoint ever - replicate offset in every element */
1794 for (i = 0; i < NUM_DATAPOINTS; i++) {
1795 p->filter_datapoint[i].d_offset = offset;
1799 p->reachable_bits |= 1;
1800 if ((MAX_VERBOSE && G.verbose) || (option_mask32 & OPT_w)) {
1801 bb_error_msg("reply from %s: offset:%+f delay:%f status:0x%02x strat:%d refid:0x%08x rootdelay:%f reach:0x%02x",
1808 p->lastpkt_rootdelay,
1810 /* not shown: m_ppoll, m_precision_exp, m_rootdisp,
1811 * m_reftime, m_orgtime, m_rectime, m_xmttime
1816 /* Muck with statictics and update the clock */
1817 filter_datapoints(p);
1818 q = select_and_cluster();
1822 if (!(option_mask32 & OPT_w)) {
1823 rc = update_local_clock(q);
1824 /* If drift is dangerously large, immediately
1825 * drop poll interval one step down.
1827 if (fabs(q->filter_offset) >= POLLDOWN_OFFSET) {
1828 VERB4 bb_error_msg("offset:%+f > POLLDOWN_OFFSET", q->filter_offset);
1833 /* else: no peer selected, rc = -1: we want to poll more often */
1836 /* Adjust the poll interval by comparing the current offset
1837 * with the clock jitter. If the offset is less than
1838 * the clock jitter times a constant, then the averaging interval
1839 * is increased, otherwise it is decreased. A bit of hysteresis
1840 * helps calm the dance. Works best using burst mode.
1842 if (rc > 0 && G.offset_to_jitter_ratio <= POLLADJ_GATE) {
1843 /* was += G.poll_exp but it is a bit
1844 * too optimistic for my taste at high poll_exp's */
1845 G.polladj_count += MINPOLL;
1846 if (G.polladj_count > POLLADJ_LIMIT) {
1847 G.polladj_count = 0;
1848 if (G.poll_exp < MAXPOLL) {
1850 VERB4 bb_error_msg("polladj: discipline_jitter:%f ++poll_exp=%d",
1851 G.discipline_jitter, G.poll_exp);
1854 VERB4 bb_error_msg("polladj: incr:%d", G.polladj_count);
1857 G.polladj_count -= G.poll_exp * 2;
1858 if (G.polladj_count < -POLLADJ_LIMIT || G.poll_exp >= BIGPOLL) {
1860 G.polladj_count = 0;
1861 if (G.poll_exp > MINPOLL) {
1865 /* Correct p->next_action_time in each peer
1866 * which waits for sending, so that they send earlier.
1867 * Old pp->next_action_time are on the order
1868 * of t + (1 << old_poll_exp) + small_random,
1869 * we simply need to subtract ~half of that.
1871 for (item = G.ntp_peers; item != NULL; item = item->link) {
1872 peer_t *pp = (peer_t *) item->data;
1874 pp->next_action_time -= (1 << G.poll_exp);
1876 VERB4 bb_error_msg("polladj: discipline_jitter:%f --poll_exp=%d",
1877 G.discipline_jitter, G.poll_exp);
1880 VERB4 bb_error_msg("polladj: decr:%d", G.polladj_count);
1885 /* Decide when to send new query for this peer */
1886 pick_normal_interval:
1887 interval = poll_interval(0);
1888 if (fabs(offset) >= STEP_THRESHOLD * 8 && interval > BIGOFF_INTERVAL) {
1889 /* If we are synced, offsets are less than STEP_THRESHOLD,
1890 * or at the very least not much larger than it.
1891 * Now we see a largish one.
1892 * Either this peer is feeling bad, or packet got corrupted,
1893 * or _our_ clock is wrong now and _all_ peers will show similar
1894 * largish offsets too.
1895 * I observed this with laptop suspend stopping clock.
1896 * In any case, it makes sense to make next request soonish:
1897 * cases 1 and 2: get a better datapoint,
1898 * case 3: allows to resync faster.
1900 interval = BIGOFF_INTERVAL;
1904 set_next(p, interval);
1907 #if ENABLE_FEATURE_NTPD_SERVER
1908 static NOINLINE void
1909 recv_and_process_client_pkt(void /*int fd*/)
1913 len_and_sockaddr *to;
1914 struct sockaddr *from;
1916 uint8_t query_status;
1917 l_fixedpt_t query_xmttime;
1919 to = get_sock_lsa(G_listen_fd);
1920 from = xzalloc(to->len);
1922 size = recv_from_to(G_listen_fd, &msg, sizeof(msg), MSG_DONTWAIT, from, &to->u.sa, to->len);
1923 if (size != NTP_MSGSIZE_NOAUTH && size != NTP_MSGSIZE) {
1926 if (errno == EAGAIN)
1928 bb_perror_msg_and_die("recv");
1930 addr = xmalloc_sockaddr2dotted_noport(from);
1931 bb_error_msg("malformed packet received from %s: size %u", addr, (int)size);
1936 query_status = msg.m_status;
1937 query_xmttime = msg.m_xmttime;
1939 /* Build a reply packet */
1940 memset(&msg, 0, sizeof(msg));
1941 msg.m_status = G.stratum < MAXSTRAT ? (G.ntp_status & LI_MASK) : LI_ALARM;
1942 msg.m_status |= (query_status & VERSION_MASK);
1943 msg.m_status |= ((query_status & MODE_MASK) == MODE_CLIENT) ?
1944 MODE_SERVER : MODE_SYM_PAS;
1945 msg.m_stratum = G.stratum;
1946 msg.m_ppoll = G.poll_exp;
1947 msg.m_precision_exp = G_precision_exp;
1948 /* this time was obtained between poll() and recv() */
1949 msg.m_rectime = d_to_lfp(G.cur_time);
1950 msg.m_xmttime = d_to_lfp(gettime1900d()); /* this instant */
1951 if (G.peer_cnt == 0) {
1952 /* we have no peers: "stratum 1 server" mode. reftime = our own time */
1953 G.reftime = G.cur_time;
1955 msg.m_reftime = d_to_lfp(G.reftime);
1956 msg.m_orgtime = query_xmttime;
1957 msg.m_rootdelay = d_to_sfp(G.rootdelay);
1958 //simple code does not do this, fix simple code!
1959 msg.m_rootdisp = d_to_sfp(G.rootdisp);
1960 //version = (query_status & VERSION_MASK); /* ... >> VERSION_SHIFT - done below instead */
1961 msg.m_refid = G.refid; // (version > (3 << VERSION_SHIFT)) ? G.refid : G.refid3;
1963 /* We reply from the local address packet was sent to,
1964 * this makes to/from look swapped here: */
1965 do_sendto(G_listen_fd,
1966 /*from:*/ &to->u.sa, /*to:*/ from, /*addrlen:*/ to->len,
1975 /* Upstream ntpd's options:
1977 * -4 Force DNS resolution of host names to the IPv4 namespace.
1978 * -6 Force DNS resolution of host names to the IPv6 namespace.
1979 * -a Require cryptographic authentication for broadcast client,
1980 * multicast client and symmetric passive associations.
1981 * This is the default.
1982 * -A Do not require cryptographic authentication for broadcast client,
1983 * multicast client and symmetric passive associations.
1984 * This is almost never a good idea.
1985 * -b Enable the client to synchronize to broadcast servers.
1987 * Specify the name and path of the configuration file,
1988 * default /etc/ntp.conf
1989 * -d Specify debugging mode. This option may occur more than once,
1990 * with each occurrence indicating greater detail of display.
1992 * Specify debugging level directly.
1994 * Specify the name and path of the frequency file.
1995 * This is the same operation as the "driftfile FILE"
1996 * configuration command.
1997 * -g Normally, ntpd exits with a message to the system log
1998 * if the offset exceeds the panic threshold, which is 1000 s
1999 * by default. This option allows the time to be set to any value
2000 * without restriction; however, this can happen only once.
2001 * If the threshold is exceeded after that, ntpd will exit
2002 * with a message to the system log. This option can be used
2003 * with the -q and -x options. See the tinker command for other options.
2005 * Chroot the server to the directory jaildir. This option also implies
2006 * that the server attempts to drop root privileges at startup
2007 * (otherwise, chroot gives very little additional security).
2008 * You may need to also specify a -u option.
2010 * Specify the name and path of the symmetric key file,
2011 * default /etc/ntp/keys. This is the same operation
2012 * as the "keys FILE" configuration command.
2014 * Specify the name and path of the log file. The default
2015 * is the system log file. This is the same operation as
2016 * the "logfile FILE" configuration command.
2017 * -L Do not listen to virtual IPs. The default is to listen.
2019 * -N To the extent permitted by the operating system,
2020 * run the ntpd at the highest priority.
2022 * Specify the name and path of the file used to record the ntpd
2023 * process ID. This is the same operation as the "pidfile FILE"
2024 * configuration command.
2026 * To the extent permitted by the operating system,
2027 * run the ntpd at the specified priority.
2028 * -q Exit the ntpd just after the first time the clock is set.
2029 * This behavior mimics that of the ntpdate program, which is
2030 * to be retired. The -g and -x options can be used with this option.
2031 * Note: The kernel time discipline is disabled with this option.
2033 * Specify the default propagation delay from the broadcast/multicast
2034 * server to this client. This is necessary only if the delay
2035 * cannot be computed automatically by the protocol.
2037 * Specify the directory path for files created by the statistics
2038 * facility. This is the same operation as the "statsdir DIR"
2039 * configuration command.
2041 * Add a key number to the trusted key list. This option can occur
2044 * Specify a user, and optionally a group, to switch to.
2047 * Add a system variable listed by default.
2048 * -x Normally, the time is slewed if the offset is less than the step
2049 * threshold, which is 128 ms by default, and stepped if above
2050 * the threshold. This option sets the threshold to 600 s, which is
2051 * well within the accuracy window to set the clock manually.
2052 * Note: since the slew rate of typical Unix kernels is limited
2053 * to 0.5 ms/s, each second of adjustment requires an amortization
2054 * interval of 2000 s. Thus, an adjustment as much as 600 s
2055 * will take almost 14 days to complete. This option can be used
2056 * with the -g and -q options. See the tinker command for other options.
2057 * Note: The kernel time discipline is disabled with this option.
2060 /* By doing init in a separate function we decrease stack usage
2063 static NOINLINE void ntp_init(char **argv)
2071 bb_error_msg_and_die(bb_msg_you_must_be_root);
2073 /* Set some globals */
2074 G.stratum = MAXSTRAT;
2076 G.poll_exp = BURSTPOLL; /* speeds up initial sync */
2077 G.last_script_run = G.reftime = G.last_update_recv_time = gettime1900d(); /* sets G.cur_time too */
2081 opt_complementary = "dd:p::wn"; /* d: counter; p: list; -w implies -n */
2082 opts = getopt32(argv,
2084 "wp:S:"IF_FEATURE_NTPD_SERVER("l") /* NOT compat */
2086 "46aAbgL", /* compat, ignored */
2087 &peers, &G.script_name, &G.verbose);
2088 if (!(opts & (OPT_p|OPT_l)))
2090 // if (opts & OPT_x) /* disable stepping, only slew is allowed */
2091 // G.time_was_stepped = 1;
2094 add_peers(llist_pop(&peers));
2096 /* -l but no peers: "stratum 1 server" mode */
2099 if (!(opts & OPT_n)) {
2100 bb_daemonize_or_rexec(DAEMON_DEVNULL_STDIO, argv);
2101 logmode = LOGMODE_NONE;
2103 #if ENABLE_FEATURE_NTPD_SERVER
2106 G_listen_fd = create_and_bind_dgram_or_die(NULL, 123);
2107 socket_want_pktinfo(G_listen_fd);
2108 setsockopt(G_listen_fd, IPPROTO_IP, IP_TOS, &const_IPTOS_LOWDELAY, sizeof(const_IPTOS_LOWDELAY));
2111 /* I hesitate to set -20 prio. -15 should be high enough for timekeeping */
2113 setpriority(PRIO_PROCESS, 0, -15);
2115 /* If network is up, syncronization occurs in ~10 seconds.
2116 * We give "ntpd -q" 10 seconds to get first reply,
2117 * then another 50 seconds to finish syncing.
2119 * I tested ntpd 4.2.6p1 and apparently it never exits
2120 * (will try forever), but it does not feel right.
2121 * The goal of -q is to act like ntpdate: set time
2122 * after a reasonably small period of polling, or fail.
2125 option_mask32 |= OPT_qq;
2142 int ntpd_main(int argc UNUSED_PARAM, char **argv) MAIN_EXTERNALLY_VISIBLE;
2143 int ntpd_main(int argc UNUSED_PARAM, char **argv)
2151 memset(&G, 0, sizeof(G));
2152 SET_PTR_TO_GLOBALS(&G);
2156 /* If ENABLE_FEATURE_NTPD_SERVER, + 1 for listen_fd: */
2157 cnt = G.peer_cnt + ENABLE_FEATURE_NTPD_SERVER;
2158 idx2peer = xzalloc(sizeof(idx2peer[0]) * cnt);
2159 pfd = xzalloc(sizeof(pfd[0]) * cnt);
2161 /* Countdown: we never sync before we sent INITIAL_SAMPLES+1
2162 * packets to each peer.
2163 * NB: if some peer is not responding, we may end up sending
2164 * fewer packets to it and more to other peers.
2165 * NB2: sync usually happens using INITIAL_SAMPLES packets,
2166 * since last reply does not come back instantaneously.
2168 cnt = G.peer_cnt * (INITIAL_SAMPLES + 1);
2170 write_pidfile(CONFIG_PID_FILE_PATH "/ntpd.pid");
2172 while (!bb_got_signal) {
2178 /* Nothing between here and poll() blocks for any significant time */
2180 nextaction = G.cur_time + 3600;
2183 #if ENABLE_FEATURE_NTPD_SERVER
2184 if (G_listen_fd != -1) {
2185 pfd[0].fd = G_listen_fd;
2186 pfd[0].events = POLLIN;
2190 /* Pass over peer list, send requests, time out on receives */
2191 for (item = G.ntp_peers; item != NULL; item = item->link) {
2192 peer_t *p = (peer_t *) item->data;
2194 if (p->next_action_time <= G.cur_time) {
2195 if (p->p_fd == -1) {
2196 /* Time to send new req */
2198 G.initial_poll_complete = 1;
2200 send_query_to_peer(p);
2202 /* Timed out waiting for reply */
2205 timeout = poll_interval(-2); /* -2: try a bit sooner */
2206 bb_error_msg("timed out waiting for %s, reach 0x%02x, next query in %us",
2207 p->p_dotted, p->reachable_bits, timeout);
2208 set_next(p, timeout);
2212 if (p->next_action_time < nextaction)
2213 nextaction = p->next_action_time;
2216 /* Wait for reply from this peer */
2217 pfd[i].fd = p->p_fd;
2218 pfd[i].events = POLLIN;
2224 timeout = nextaction - G.cur_time;
2227 timeout++; /* (nextaction - G.cur_time) rounds down, compensating */
2229 /* Here we may block */
2231 if (i > (ENABLE_FEATURE_NTPD_SERVER && G_listen_fd != -1)) {
2232 /* We wait for at least one reply.
2233 * Poll for it, without wasting time for message.
2234 * Since replies often come under 1 second, this also
2235 * reduces clutter in logs.
2237 nfds = poll(pfd, i, 1000);
2243 bb_error_msg("poll:%us sockets:%u interval:%us", timeout, i, 1 << G.poll_exp);
2245 nfds = poll(pfd, i, timeout * 1000);
2247 gettime1900d(); /* sets G.cur_time */
2249 if (!bb_got_signal /* poll wasn't interrupted by a signal */
2250 && G.cur_time - G.last_script_run > 11*60
2252 /* Useful for updating battery-backed RTC and such */
2253 run_script("periodic", G.last_update_offset);
2254 gettime1900d(); /* sets G.cur_time */
2259 /* Process any received packets */
2261 #if ENABLE_FEATURE_NTPD_SERVER
2262 if (G.listen_fd != -1) {
2263 if (pfd[0].revents /* & (POLLIN|POLLERR)*/) {
2265 recv_and_process_client_pkt(/*G.listen_fd*/);
2266 gettime1900d(); /* sets G.cur_time */
2271 for (; nfds != 0 && j < i; j++) {
2272 if (pfd[j].revents /* & (POLLIN|POLLERR)*/) {
2274 * At init, alarm was set to 10 sec.
2275 * Now we did get a reply.
2276 * Increase timeout to 50 seconds to finish syncing.
2278 if (option_mask32 & OPT_qq) {
2279 option_mask32 &= ~OPT_qq;
2283 recv_and_process_peer_pkt(idx2peer[j]);
2284 gettime1900d(); /* sets G.cur_time */
2289 if (G.ntp_peers && G.stratum != MAXSTRAT) {
2290 for (item = G.ntp_peers; item != NULL; item = item->link) {
2291 peer_t *p = (peer_t *) item->data;
2292 if (p->reachable_bits)
2293 goto have_reachable_peer;
2295 /* No peer responded for last 8 packets, panic */
2296 G.polladj_count = 0;
2297 G.poll_exp = MINPOLL;
2298 G.stratum = MAXSTRAT;
2299 run_script("unsync", 0.0);
2300 have_reachable_peer: ;
2302 } /* while (!bb_got_signal) */
2304 remove_pidfile(CONFIG_PID_FILE_PATH "/ntpd.pid");
2305 kill_myself_with_sig(bb_got_signal);
2313 /*** openntpd-4.6 uses only adjtime, not adjtimex ***/
2315 /*** ntp-4.2.6/ntpd/ntp_loopfilter.c - adjtimex usage ***/
2319 direct_freq(double fp_offset)
2323 * If the kernel is enabled, we need the residual offset to
2324 * calculate the frequency correction.
2326 if (pll_control && kern_enable) {
2327 memset(&ntv, 0, sizeof(ntv));
2330 clock_offset = ntv.offset / 1e9;
2331 #else /* STA_NANO */
2332 clock_offset = ntv.offset / 1e6;
2333 #endif /* STA_NANO */
2334 drift_comp = FREQTOD(ntv.freq);
2336 #endif /* KERNEL_PLL */
2337 set_freq((fp_offset - clock_offset) / (current_time - clock_epoch) + drift_comp);
2343 set_freq(double freq) /* frequency update */
2351 * If the kernel is enabled, update the kernel frequency.
2353 if (pll_control && kern_enable) {
2354 memset(&ntv, 0, sizeof(ntv));
2355 ntv.modes = MOD_FREQUENCY;
2356 ntv.freq = DTOFREQ(drift_comp);
2358 snprintf(tbuf, sizeof(tbuf), "kernel %.3f PPM", drift_comp * 1e6);
2359 report_event(EVNT_FSET, NULL, tbuf);
2361 snprintf(tbuf, sizeof(tbuf), "ntpd %.3f PPM", drift_comp * 1e6);
2362 report_event(EVNT_FSET, NULL, tbuf);
2364 #else /* KERNEL_PLL */
2365 snprintf(tbuf, sizeof(tbuf), "ntpd %.3f PPM", drift_comp * 1e6);
2366 report_event(EVNT_FSET, NULL, tbuf);
2367 #endif /* KERNEL_PLL */
2376 * This code segment works when clock adjustments are made using
2377 * precision time kernel support and the ntp_adjtime() system
2378 * call. This support is available in Solaris 2.6 and later,
2379 * Digital Unix 4.0 and later, FreeBSD, Linux and specially
2380 * modified kernels for HP-UX 9 and Ultrix 4. In the case of the
2381 * DECstation 5000/240 and Alpha AXP, additional kernel
2382 * modifications provide a true microsecond clock and nanosecond
2383 * clock, respectively.
2385 * Important note: The kernel discipline is used only if the
2386 * step threshold is less than 0.5 s, as anything higher can
2387 * lead to overflow problems. This might occur if some misguided
2388 * lad set the step threshold to something ridiculous.
2390 if (pll_control && kern_enable) {
2392 #define MOD_BITS (MOD_OFFSET | MOD_MAXERROR | MOD_ESTERROR | MOD_STATUS | MOD_TIMECONST)
2395 * We initialize the structure for the ntp_adjtime()
2396 * system call. We have to convert everything to
2397 * microseconds or nanoseconds first. Do not update the
2398 * system variables if the ext_enable flag is set. In
2399 * this case, the external clock driver will update the
2400 * variables, which will be read later by the local
2401 * clock driver. Afterwards, remember the time and
2402 * frequency offsets for jitter and stability values and
2403 * to update the frequency file.
2405 memset(&ntv, 0, sizeof(ntv));
2407 ntv.modes = MOD_STATUS;
2410 ntv.modes = MOD_BITS | MOD_NANO;
2411 #else /* STA_NANO */
2412 ntv.modes = MOD_BITS;
2413 #endif /* STA_NANO */
2414 if (clock_offset < 0)
2419 ntv.offset = (int32)(clock_offset * 1e9 + dtemp);
2420 ntv.constant = sys_poll;
2421 #else /* STA_NANO */
2422 ntv.offset = (int32)(clock_offset * 1e6 + dtemp);
2423 ntv.constant = sys_poll - 4;
2424 #endif /* STA_NANO */
2425 ntv.esterror = (u_int32)(clock_jitter * 1e6);
2426 ntv.maxerror = (u_int32)((sys_rootdelay / 2 + sys_rootdisp) * 1e6);
2427 ntv.status = STA_PLL;
2430 * Enable/disable the PPS if requested.
2433 if (!(pll_status & STA_PPSTIME))
2434 report_event(EVNT_KERN,
2435 NULL, "PPS enabled");
2436 ntv.status |= STA_PPSTIME | STA_PPSFREQ;
2438 if (pll_status & STA_PPSTIME)
2439 report_event(EVNT_KERN,
2440 NULL, "PPS disabled");
2441 ntv.status &= ~(STA_PPSTIME | STA_PPSFREQ);
2443 if (sys_leap == LEAP_ADDSECOND)
2444 ntv.status |= STA_INS;
2445 else if (sys_leap == LEAP_DELSECOND)
2446 ntv.status |= STA_DEL;
2450 * Pass the stuff to the kernel. If it squeals, turn off
2451 * the pps. In any case, fetch the kernel offset,
2452 * frequency and jitter.
2454 if (ntp_adjtime(&ntv) == TIME_ERROR) {
2455 if (!(ntv.status & STA_PPSSIGNAL))
2456 report_event(EVNT_KERN, NULL,
2459 pll_status = ntv.status;
2461 clock_offset = ntv.offset / 1e9;
2462 #else /* STA_NANO */
2463 clock_offset = ntv.offset / 1e6;
2464 #endif /* STA_NANO */
2465 clock_frequency = FREQTOD(ntv.freq);
2468 * If the kernel PPS is lit, monitor its performance.
2470 if (ntv.status & STA_PPSTIME) {
2472 clock_jitter = ntv.jitter / 1e9;
2473 #else /* STA_NANO */
2474 clock_jitter = ntv.jitter / 1e6;
2475 #endif /* STA_NANO */
2478 #if defined(STA_NANO) && NTP_API == 4
2480 * If the TAI changes, update the kernel TAI.
2482 if (loop_tai != sys_tai) {
2484 ntv.modes = MOD_TAI;
2485 ntv.constant = sys_tai;
2488 #endif /* STA_NANO */
2490 #endif /* KERNEL_PLL */