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 -I IFACE")"] [-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"
42 //usage: "\n -I IFACE Bind server to IFACE, implies -l"
44 //usage: "\n -S PROG Run PROG after stepping time, stratum change, and every 11 mins"
45 //usage: "\n -p PEER Obtain time from PEER (may be repeated)"
46 //usage: IF_FEATURE_NTPD_CONF(
47 //usage: "\n If -p is not given, read /etc/ntp.conf"
50 // -l and -p options are not compatible with "standard" ntpd:
51 // it has them as "-l logfile" and "-p pidfile".
52 // -S and -w are not compat either, "standard" ntpd has no such opts.
56 #include <netinet/ip.h> /* For IPTOS_LOWDELAY definition */
57 #include <sys/resource.h> /* setpriority */
58 #include <sys/timex.h>
59 #ifndef IPTOS_LOWDELAY
60 # define IPTOS_LOWDELAY 0x10
64 /* Verbosity control (max level of -dddd options accepted).
65 * max 6 is very talkative (and bloated). 3 is non-bloated,
66 * production level setting.
71 /* High-level description of the algorithm:
73 * We start running with very small poll_exp, BURSTPOLL,
74 * in order to quickly accumulate INITIAL_SAMPLES datapoints
75 * for each peer. Then, time is stepped if the offset is larger
76 * than STEP_THRESHOLD, otherwise it isn't; anyway, we enlarge
77 * poll_exp to MINPOLL and enter frequency measurement step:
78 * we collect new datapoints but ignore them for WATCH_THRESHOLD
79 * seconds. After WATCH_THRESHOLD seconds we look at accumulated
80 * offset and estimate frequency drift.
82 * (frequency measurement step seems to not be strictly needed,
83 * it is conditionally disabled with USING_INITIAL_FREQ_ESTIMATION
86 * After this, we enter "steady state": we collect a datapoint,
87 * we select the best peer, if this datapoint is not a new one
88 * (IOW: if this datapoint isn't for selected peer), sleep
89 * and collect another one; otherwise, use its offset to update
90 * frequency drift, if offset is somewhat large, reduce poll_exp,
91 * otherwise increase poll_exp.
93 * If offset is larger than STEP_THRESHOLD, which shouldn't normally
94 * happen, we assume that something "bad" happened (computer
95 * was hibernated, someone set totally wrong date, etc),
96 * then the time is stepped, all datapoints are discarded,
97 * and we go back to steady state.
99 * Made some changes to speed up re-syncing after our clock goes bad
100 * (tested with suspending my laptop):
101 * - if largish offset (>= STEP_THRESHOLD * 8 == 1 sec) is seen
102 * from a peer, schedule next query for this peer soon
103 * without drastically lowering poll interval for everybody.
104 * This makes us collect enough data for step much faster:
105 * e.g. at poll = 10 (1024 secs), step was done within 5 minutes
106 * after first reply which indicated that our clock is 14 seconds off.
107 * - on step, do not discard d_dispersion data of the existing datapoints,
108 * do not clear reachable_bits. This prevents discarding first ~8
109 * datapoints after the step.
112 #define INITIAL_SAMPLES 4 /* how many samples do we want for init */
113 #define BAD_DELAY_GROWTH 4 /* drop packet if its delay grew by more than this */
115 #define RETRY_INTERVAL 32 /* on send/recv error, retry in N secs (need to be power of 2) */
116 #define NOREPLY_INTERVAL 512 /* sent, but got no reply: cap next query by this many seconds */
117 #define RESPONSE_INTERVAL 16 /* wait for reply up to N secs */
119 /* Step threshold (sec). std ntpd uses 0.128.
120 * Using exact power of 2 (1/8) results in smaller code
122 #define STEP_THRESHOLD 0.125
123 /* Stepout threshold (sec). std ntpd uses 900 (11 mins (!)) */
124 #define WATCH_THRESHOLD 128
125 /* NB: set WATCH_THRESHOLD to ~60 when debugging to save time) */
126 //UNUSED: #define PANIC_THRESHOLD 1000 /* panic threshold (sec) */
129 * If we got |offset| > BIGOFF from a peer, cap next query interval
130 * for this peer by this many seconds:
132 #define BIGOFF (STEP_THRESHOLD * 8)
133 #define BIGOFF_INTERVAL (1 << 7) /* 128 s */
135 #define FREQ_TOLERANCE 0.000015 /* frequency tolerance (15 PPM) */
136 #define BURSTPOLL 0 /* initial poll */
137 #define MINPOLL 5 /* minimum poll interval. std ntpd uses 6 (6: 64 sec) */
139 * If offset > discipline_jitter * POLLADJ_GATE, and poll interval is > 2^BIGPOLL,
140 * then it is decreased _at once_. (If <= 2^BIGPOLL, it will be decreased _eventually_).
142 #define BIGPOLL 9 /* 2^9 sec ~= 8.5 min */
143 #define MAXPOLL 12 /* maximum poll interval (12: 1.1h, 17: 36.4h). std ntpd uses 17 */
145 * Actively lower poll when we see such big offsets.
146 * With STEP_THRESHOLD = 0.125, it means we try to sync more aggressively
147 * if offset increases over ~0.04 sec
149 //#define POLLDOWN_OFFSET (STEP_THRESHOLD / 3)
150 #define MINDISP 0.01 /* minimum dispersion (sec) */
151 #define MAXDISP 16 /* maximum dispersion (sec) */
152 #define MAXSTRAT 16 /* maximum stratum (infinity metric) */
153 #define MAXDIST 1 /* distance threshold (sec) */
154 #define MIN_SELECTED 1 /* minimum intersection survivors */
155 #define MIN_CLUSTERED 3 /* minimum cluster survivors */
157 #define MAXDRIFT 0.000500 /* frequency drift we can correct (500 PPM) */
159 /* Poll-adjust threshold.
160 * When we see that offset is small enough compared to discipline jitter,
161 * we grow a counter: += MINPOLL. When counter goes over POLLADJ_LIMIT,
162 * we poll_exp++. If offset isn't small, counter -= poll_exp*2,
163 * and when it goes below -POLLADJ_LIMIT, we poll_exp--.
164 * (Bumped from 30 to 40 since otherwise I often see poll_exp going *2* steps down)
166 #define POLLADJ_LIMIT 40
167 /* If offset < discipline_jitter * POLLADJ_GATE, then we decide to increase
168 * poll interval (we think we can't improve timekeeping
169 * by staying at smaller poll).
171 #define POLLADJ_GATE 4
172 #define TIMECONST_HACK_GATE 2
173 /* Compromise Allan intercept (sec). doc uses 1500, std ntpd uses 512 */
177 /* FLL loop gain [why it depends on MAXPOLL??] */
178 #define FLL (MAXPOLL + 1)
179 /* Parameter averaging constant */
188 NTP_MSGSIZE_NOAUTH = 48,
189 NTP_MSGSIZE = (NTP_MSGSIZE_NOAUTH + 4 + NTP_DIGESTSIZE),
192 MODE_MASK = (7 << 0),
193 VERSION_MASK = (7 << 3),
197 /* Leap Second Codes (high order two bits of m_status) */
198 LI_NOWARNING = (0 << 6), /* no warning */
199 LI_PLUSSEC = (1 << 6), /* add a second (61 seconds) */
200 LI_MINUSSEC = (2 << 6), /* minus a second (59 seconds) */
201 LI_ALARM = (3 << 6), /* alarm condition */
204 MODE_RES0 = 0, /* reserved */
205 MODE_SYM_ACT = 1, /* symmetric active */
206 MODE_SYM_PAS = 2, /* symmetric passive */
207 MODE_CLIENT = 3, /* client */
208 MODE_SERVER = 4, /* server */
209 MODE_BROADCAST = 5, /* broadcast */
210 MODE_RES1 = 6, /* reserved for NTP control message */
211 MODE_RES2 = 7, /* reserved for private use */
214 //TODO: better base selection
215 #define OFFSET_1900_1970 2208988800UL /* 1970 - 1900 in seconds */
217 #define NUM_DATAPOINTS 8
230 uint8_t m_status; /* status of local clock and leap info */
232 uint8_t m_ppoll; /* poll value */
233 int8_t m_precision_exp;
234 s_fixedpt_t m_rootdelay;
235 s_fixedpt_t m_rootdisp;
237 l_fixedpt_t m_reftime;
238 l_fixedpt_t m_orgtime;
239 l_fixedpt_t m_rectime;
240 l_fixedpt_t m_xmttime;
242 uint8_t m_digest[NTP_DIGESTSIZE];
252 len_and_sockaddr *p_lsa;
256 uint32_t lastpkt_refid;
257 uint8_t lastpkt_status;
258 uint8_t lastpkt_stratum;
259 uint8_t reachable_bits;
260 /* when to send new query (if p_fd == -1)
261 * or when receive times out (if p_fd >= 0): */
262 double next_action_time;
265 /* p_raw_delay is set even by "high delay" packets */
266 /* lastpkt_delay isn't */
267 double lastpkt_recv_time;
268 double lastpkt_delay;
269 double lastpkt_rootdelay;
270 double lastpkt_rootdisp;
271 /* produced by filter algorithm: */
272 double filter_offset;
273 double filter_dispersion;
274 double filter_jitter;
275 datapoint_t filter_datapoint[NUM_DATAPOINTS];
276 /* last sent packet: */
281 #define USING_KERNEL_PLL_LOOP 1
282 #define USING_INITIAL_FREQ_ESTIMATION 0
289 /* Insert new options above this line. */
290 /* Non-compat options: */
294 OPT_l = (1 << 7) * ENABLE_FEATURE_NTPD_SERVER,
295 OPT_I = (1 << 8) * ENABLE_FEATURE_NTPD_SERVER,
296 /* We hijack some bits for other purposes */
302 /* total round trip delay to currently selected reference clock */
304 /* reference timestamp: time when the system clock was last set or corrected */
306 /* total dispersion to currently selected reference clock */
309 double last_script_run;
312 #if ENABLE_FEATURE_NTPD_SERVER
315 # define G_listen_fd (G.listen_fd)
317 # define G_listen_fd (-1)
321 /* refid: 32-bit code identifying the particular server or reference clock
322 * in stratum 0 packets this is a four-character ASCII string,
323 * called the kiss code, used for debugging and monitoring
324 * in stratum 1 packets this is a four-character ASCII string
325 * assigned to the reference clock by IANA. Example: "GPS "
326 * in stratum 2+ packets, it's IPv4 address or 4 first bytes
327 * of MD5 hash of IPv6
331 /* precision is defined as the larger of the resolution and time to
332 * read the clock, in log2 units. For instance, the precision of a
333 * mains-frequency clock incrementing at 60 Hz is 16 ms, even when the
334 * system clock hardware representation is to the nanosecond.
336 * Delays, jitters of various kinds are clamped down to precision.
338 * If precision_sec is too large, discipline_jitter gets clamped to it
339 * and if offset is smaller than discipline_jitter * POLLADJ_GATE, poll
340 * interval grows even though we really can benefit from staying at
341 * smaller one, collecting non-lagged datapoits and correcting offset.
342 * (Lagged datapoits exist when poll_exp is large but we still have
343 * systematic offset error - the time distance between datapoints
344 * is significant and older datapoints have smaller offsets.
345 * This makes our offset estimation a bit smaller than reality)
346 * Due to this effect, setting G_precision_sec close to
347 * STEP_THRESHOLD isn't such a good idea - offsets may grow
348 * too big and we will step. I observed it with -6.
350 * OTOH, setting precision_sec far too small would result in futile
351 * attempts to syncronize to an unachievable precision.
353 * -6 is 1/64 sec, -7 is 1/128 sec and so on.
354 * -8 is 1/256 ~= 0.003906 (worked well for me --vda)
355 * -9 is 1/512 ~= 0.001953 (let's try this for some time)
357 #define G_precision_exp -9
359 * G_precision_exp is used only for construction outgoing packets.
360 * It's ok to set G_precision_sec to a slightly different value
361 * (One which is "nicer looking" in logs).
362 * Exact value would be (1.0 / (1 << (- G_precision_exp))):
364 #define G_precision_sec 0.002
366 /* Bool. After set to 1, never goes back to 0: */
367 smallint initial_poll_complete;
369 #define STATE_NSET 0 /* initial state, "nothing is set" */
370 //#define STATE_FSET 1 /* frequency set from file */
371 //#define STATE_SPIK 2 /* spike detected */
372 //#define STATE_FREQ 3 /* initial frequency */
373 #define STATE_SYNC 4 /* clock synchronized (normal operation) */
374 uint8_t discipline_state; // doc calls it c.state
375 uint8_t poll_exp; // s.poll
376 int polladj_count; // c.count
377 long kernel_freq_drift;
378 peer_t *last_update_peer;
379 double last_update_offset; // c.last
380 double last_update_recv_time; // s.t
381 double discipline_jitter; // c.jitter
382 /* Since we only compare it with ints, can simplify code
383 * by not making this variable floating point:
385 unsigned offset_to_jitter_ratio;
386 //double cluster_offset; // s.offset
387 //double cluster_jitter; // s.jitter
388 #if !USING_KERNEL_PLL_LOOP
389 double discipline_freq_drift; // c.freq
390 /* Maybe conditionally calculate wander? it's used only for logging */
391 double discipline_wander; // c.wander
394 #define G (*ptr_to_globals)
396 static const int const_IPTOS_LOWDELAY = IPTOS_LOWDELAY;
399 #define VERB1 if (MAX_VERBOSE && G.verbose)
400 #define VERB2 if (MAX_VERBOSE >= 2 && G.verbose >= 2)
401 #define VERB3 if (MAX_VERBOSE >= 3 && G.verbose >= 3)
402 #define VERB4 if (MAX_VERBOSE >= 4 && G.verbose >= 4)
403 #define VERB5 if (MAX_VERBOSE >= 5 && G.verbose >= 5)
404 #define VERB6 if (MAX_VERBOSE >= 6 && G.verbose >= 6)
407 static double LOG2D(int a)
410 return 1.0 / (1UL << -a);
413 static ALWAYS_INLINE double SQUARE(double x)
417 static ALWAYS_INLINE double MAXD(double a, double b)
423 static ALWAYS_INLINE double MIND(double a, double b)
429 static NOINLINE double my_SQRT(double X)
436 double Xhalf = X * 0.5;
438 /* Fast and good approximation to 1/sqrt(X), black magic */
440 /*v.i = 0x5f3759df - (v.i >> 1);*/
441 v.i = 0x5f375a86 - (v.i >> 1); /* - this constant is slightly better */
442 invsqrt = v.f; /* better than 0.2% accuracy */
444 /* Refining it using Newton's method: x1 = x0 - f(x0)/f'(x0)
445 * f(x) = 1/(x*x) - X (f==0 when x = 1/sqrt(X))
447 * f(x)/f'(x) = (X - 1/(x*x)) / (2/(x*x*x)) = X*x*x*x/2 - x/2
448 * x1 = x0 - (X*x0*x0*x0/2 - x0/2) = 1.5*x0 - X*x0*x0*x0/2 = x0*(1.5 - (X/2)*x0*x0)
450 invsqrt = invsqrt * (1.5 - Xhalf * invsqrt * invsqrt); /* ~0.05% accuracy */
451 /* invsqrt = invsqrt * (1.5 - Xhalf * invsqrt * invsqrt); 2nd iter: ~0.0001% accuracy */
452 /* With 4 iterations, more than half results will be exact,
453 * at 6th iterations result stabilizes with about 72% results exact.
454 * We are well satisfied with 0.05% accuracy.
457 return X * invsqrt; /* X * 1/sqrt(X) ~= sqrt(X) */
459 static ALWAYS_INLINE double SQRT(double X)
461 /* If this arch doesn't use IEEE 754 floats, fall back to using libm */
462 if (sizeof(float) != 4)
465 /* This avoids needing libm, saves about 0.5k on x86-32 */
473 gettimeofday(&tv, NULL); /* never fails */
474 G.cur_time = tv.tv_sec + (1.0e-6 * tv.tv_usec) + OFFSET_1900_1970;
479 d_to_tv(double d, struct timeval *tv)
481 tv->tv_sec = (long)d;
482 tv->tv_usec = (d - tv->tv_sec) * 1000000;
486 lfp_to_d(l_fixedpt_t lfp)
489 lfp.int_partl = ntohl(lfp.int_partl);
490 lfp.fractionl = ntohl(lfp.fractionl);
491 ret = (double)lfp.int_partl + ((double)lfp.fractionl / UINT_MAX);
495 sfp_to_d(s_fixedpt_t sfp)
498 sfp.int_parts = ntohs(sfp.int_parts);
499 sfp.fractions = ntohs(sfp.fractions);
500 ret = (double)sfp.int_parts + ((double)sfp.fractions / USHRT_MAX);
503 #if ENABLE_FEATURE_NTPD_SERVER
508 lfp.int_partl = (uint32_t)d;
509 lfp.fractionl = (uint32_t)((d - lfp.int_partl) * UINT_MAX);
510 lfp.int_partl = htonl(lfp.int_partl);
511 lfp.fractionl = htonl(lfp.fractionl);
518 sfp.int_parts = (uint16_t)d;
519 sfp.fractions = (uint16_t)((d - sfp.int_parts) * USHRT_MAX);
520 sfp.int_parts = htons(sfp.int_parts);
521 sfp.fractions = htons(sfp.fractions);
527 dispersion(const datapoint_t *dp)
529 return dp->d_dispersion + FREQ_TOLERANCE * (G.cur_time - dp->d_recv_time);
533 root_distance(peer_t *p)
535 /* The root synchronization distance is the maximum error due to
536 * all causes of the local clock relative to the primary server.
537 * It is defined as half the total delay plus total dispersion
540 return MAXD(MINDISP, p->lastpkt_rootdelay + p->lastpkt_delay) / 2
541 + p->lastpkt_rootdisp
542 + p->filter_dispersion
543 + FREQ_TOLERANCE * (G.cur_time - p->lastpkt_recv_time)
548 set_next(peer_t *p, unsigned t)
550 p->next_action_time = G.cur_time + t;
554 * Peer clock filter and its helpers
557 filter_datapoints(peer_t *p)
564 /* Simulations have shown that use of *averaged* offset for p->filter_offset
565 * is in fact worse than simply using last received one: with large poll intervals
566 * (>= 2048) averaging code uses offset values which are outdated by hours,
567 * and time/frequency correction goes totally wrong when fed essentially bogus offsets.
570 double minoff, maxoff, w;
571 double x = x; /* for compiler */
572 double oldest_off = oldest_off;
573 double oldest_age = oldest_age;
574 double newest_off = newest_off;
575 double newest_age = newest_age;
577 fdp = p->filter_datapoint;
579 minoff = maxoff = fdp[0].d_offset;
580 for (i = 1; i < NUM_DATAPOINTS; i++) {
581 if (minoff > fdp[i].d_offset)
582 minoff = fdp[i].d_offset;
583 if (maxoff < fdp[i].d_offset)
584 maxoff = fdp[i].d_offset;
587 idx = p->datapoint_idx; /* most recent datapoint's index */
589 * Drop two outliers and take weighted average of the rest:
590 * most_recent/2 + older1/4 + older2/8 ... + older5/32 + older6/32
591 * we use older6/32, not older6/64 since sum of weights should be 1:
592 * 1/2 + 1/4 + 1/8 + 1/16 + 1/32 + 1/32 = 1
598 * filter_dispersion = \ -------------
605 for (i = 0; i < NUM_DATAPOINTS; i++) {
607 bb_error_msg("datapoint[%d]: off:%f disp:%f(%f) age:%f%s",
610 fdp[idx].d_dispersion, dispersion(&fdp[idx]),
611 G.cur_time - fdp[idx].d_recv_time,
612 (minoff == fdp[idx].d_offset || maxoff == fdp[idx].d_offset)
613 ? " (outlier by offset)" : ""
617 sum += dispersion(&fdp[idx]) / (2 << i);
619 if (minoff == fdp[idx].d_offset) {
620 minoff -= 1; /* so that we don't match it ever again */
622 if (maxoff == fdp[idx].d_offset) {
625 oldest_off = fdp[idx].d_offset;
626 oldest_age = G.cur_time - fdp[idx].d_recv_time;
629 newest_off = oldest_off;
630 newest_age = oldest_age;
637 idx = (idx - 1) & (NUM_DATAPOINTS - 1);
639 p->filter_dispersion = sum;
640 wavg += x; /* add another older6/64 to form older6/32 */
641 /* Fix systematic underestimation with large poll intervals.
642 * Imagine that we still have a bit of uncorrected drift,
643 * and poll interval is big (say, 100 sec). Offsets form a progression:
644 * 0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 - 0.7 is most recent.
645 * The algorithm above drops 0.0 and 0.7 as outliers,
646 * and then we have this estimation, ~25% off from 0.7:
647 * 0.1/32 + 0.2/32 + 0.3/16 + 0.4/8 + 0.5/4 + 0.6/2 = 0.503125
649 x = oldest_age - newest_age;
651 x = newest_age / x; /* in above example, 100 / (600 - 100) */
652 if (x < 1) { /* paranoia check */
653 x = (newest_off - oldest_off) * x; /* 0.5 * 100/500 = 0.1 */
657 p->filter_offset = wavg;
661 fdp = p->filter_datapoint;
662 idx = p->datapoint_idx; /* most recent datapoint's index */
664 /* filter_offset: simply use the most recent value */
665 p->filter_offset = fdp[idx].d_offset;
669 * filter_dispersion = \ -------------
676 for (i = 0; i < NUM_DATAPOINTS; i++) {
677 sum += dispersion(&fdp[idx]) / (2 << i);
678 wavg += fdp[idx].d_offset;
679 idx = (idx - 1) & (NUM_DATAPOINTS - 1);
681 wavg /= NUM_DATAPOINTS;
682 p->filter_dispersion = sum;
685 /* +----- -----+ ^ 1/2
689 * filter_jitter = | --- * / (avg-offset_j) |
693 * where n is the number of valid datapoints in the filter (n > 1);
694 * if filter_jitter < precision then filter_jitter = precision
697 for (i = 0; i < NUM_DATAPOINTS; i++) {
698 sum += SQUARE(wavg - fdp[i].d_offset);
700 sum = SQRT(sum / NUM_DATAPOINTS);
701 p->filter_jitter = sum > G_precision_sec ? sum : G_precision_sec;
703 VERB4 bb_error_msg("filter offset:%+f disp:%f jitter:%f",
705 p->filter_dispersion,
710 reset_peer_stats(peer_t *p, double offset)
713 bool small_ofs = fabs(offset) < 16 * STEP_THRESHOLD;
715 /* Used to set p->filter_datapoint[i].d_dispersion = MAXDISP
716 * and clear reachable bits, but this proved to be too agressive:
717 * after step (tested with suspinding laptop for ~30 secs),
718 * this caused all previous data to be considered invalid,
719 * making us needing to collect full ~8 datapoins per peer
720 * after step in order to start trusting them.
721 * In turn, this was making poll interval decrease even after
722 * step was done. (Poll interval decreases already before step
723 * in this scenario, because we see large offsets and end up with
724 * no good peer to select).
727 for (i = 0; i < NUM_DATAPOINTS; i++) {
729 p->filter_datapoint[i].d_recv_time += offset;
730 if (p->filter_datapoint[i].d_offset != 0) {
731 p->filter_datapoint[i].d_offset -= offset;
732 //bb_error_msg("p->filter_datapoint[%d].d_offset %f -> %f",
734 // p->filter_datapoint[i].d_offset + offset,
735 // p->filter_datapoint[i].d_offset);
738 p->filter_datapoint[i].d_recv_time = G.cur_time;
739 p->filter_datapoint[i].d_offset = 0;
740 /*p->filter_datapoint[i].d_dispersion = MAXDISP;*/
744 p->lastpkt_recv_time += offset;
746 /*p->reachable_bits = 0;*/
747 p->lastpkt_recv_time = G.cur_time;
749 filter_datapoints(p); /* recalc p->filter_xxx */
750 VERB6 bb_error_msg("%s->lastpkt_recv_time=%f", p->p_dotted, p->lastpkt_recv_time);
754 add_peers(const char *s)
758 p = xzalloc(sizeof(*p));
759 p->p_lsa = xhost2sockaddr(s, 123);
760 p->p_dotted = xmalloc_sockaddr2dotted_noport(&p->p_lsa->u.sa);
762 p->p_xmt_msg.m_status = MODE_CLIENT | (NTP_VERSION << 3);
763 p->next_action_time = G.cur_time; /* = set_next(p, 0); */
764 reset_peer_stats(p, 16 * STEP_THRESHOLD);
766 llist_add_to(&G.ntp_peers, p);
772 const struct sockaddr *from, const struct sockaddr *to, socklen_t addrlen,
773 msg_t *msg, ssize_t len)
779 ret = sendto(fd, msg, len, MSG_DONTWAIT, to, addrlen);
781 ret = send_to_from(fd, msg, len, MSG_DONTWAIT, to, from, addrlen);
784 bb_perror_msg("send failed");
791 send_query_to_peer(peer_t *p)
793 /* Why do we need to bind()?
794 * See what happens when we don't bind:
796 * socket(PF_INET, SOCK_DGRAM, IPPROTO_IP) = 3
797 * setsockopt(3, SOL_IP, IP_TOS, [16], 4) = 0
798 * gettimeofday({1259071266, 327885}, NULL) = 0
799 * sendto(3, "xxx", 48, MSG_DONTWAIT, {sa_family=AF_INET, sin_port=htons(123), sin_addr=inet_addr("10.34.32.125")}, 16) = 48
800 * ^^^ we sent it from some source port picked by kernel.
801 * time(NULL) = 1259071266
802 * write(2, "ntpd: entering poll 15 secs\n", 28) = 28
803 * poll([{fd=3, events=POLLIN}], 1, 15000) = 1 ([{fd=3, revents=POLLIN}])
804 * recv(3, "yyy", 68, MSG_DONTWAIT) = 48
805 * ^^^ this recv will receive packets to any local port!
807 * Uncomment this and use strace to see it in action:
809 #define PROBE_LOCAL_ADDR /* { len_and_sockaddr lsa; lsa.len = LSA_SIZEOF_SA; getsockname(p->query.fd, &lsa.u.sa, &lsa.len); } */
813 len_and_sockaddr *local_lsa;
815 family = p->p_lsa->u.sa.sa_family;
816 p->p_fd = fd = xsocket_type(&local_lsa, family, SOCK_DGRAM);
817 /* local_lsa has "null" address and port 0 now.
818 * bind() ensures we have a *particular port* selected by kernel
819 * and remembered in p->p_fd, thus later recv(p->p_fd)
820 * receives only packets sent to this port.
823 xbind(fd, &local_lsa->u.sa, local_lsa->len);
825 #if ENABLE_FEATURE_IPV6
826 if (family == AF_INET)
828 setsockopt(fd, IPPROTO_IP, IP_TOS, &const_IPTOS_LOWDELAY, sizeof(const_IPTOS_LOWDELAY));
832 /* Emit message _before_ attempted send. Think of a very short
833 * roundtrip networks: we need to go back to recv loop ASAP,
834 * to reduce delay. Printing messages after send works against that.
836 VERB1 bb_error_msg("sending query to %s", p->p_dotted);
839 * Send out a random 64-bit number as our transmit time. The NTP
840 * server will copy said number into the originate field on the
841 * response that it sends us. This is totally legal per the SNTP spec.
843 * The impact of this is two fold: we no longer send out the current
844 * system time for the world to see (which may aid an attacker), and
845 * it gives us a (not very secure) way of knowing that we're not
846 * getting spoofed by an attacker that can't capture our traffic
847 * but can spoof packets from the NTP server we're communicating with.
849 * Save the real transmit timestamp locally.
851 p->p_xmt_msg.m_xmttime.int_partl = rand();
852 p->p_xmt_msg.m_xmttime.fractionl = rand();
853 p->p_xmttime = gettime1900d();
855 /* Were doing it only if sendto worked, but
856 * loss of sync detection needs reachable_bits updated
857 * even if sending fails *locally*:
858 * "network is unreachable" because cable was pulled?
859 * We still need to declare "unsync" if this condition persists.
861 p->reachable_bits <<= 1;
863 if (do_sendto(p->p_fd, /*from:*/ NULL, /*to:*/ &p->p_lsa->u.sa, /*addrlen:*/ p->p_lsa->len,
864 &p->p_xmt_msg, NTP_MSGSIZE_NOAUTH) == -1
869 * We know that we sent nothing.
870 * We can retry *soon* without fearing
871 * that we are flooding the peer.
873 set_next(p, RETRY_INTERVAL);
877 set_next(p, RESPONSE_INTERVAL);
881 /* Note that there is no provision to prevent several run_scripts
882 * to be started in quick succession. In fact, it happens rather often
883 * if initial syncronization results in a step.
884 * You will see "step" and then "stratum" script runs, sometimes
885 * as close as only 0.002 seconds apart.
886 * Script should be ready to deal with this.
888 static void run_script(const char *action, double offset)
891 char *env1, *env2, *env3, *env4;
893 G.last_script_run = G.cur_time;
898 argv[0] = (char*) G.script_name;
899 argv[1] = (char*) action;
902 VERB1 bb_error_msg("executing '%s %s'", G.script_name, action);
904 env1 = xasprintf("%s=%u", "stratum", G.stratum);
906 env2 = xasprintf("%s=%ld", "freq_drift_ppm", G.kernel_freq_drift);
908 env3 = xasprintf("%s=%u", "poll_interval", 1 << G.poll_exp);
910 env4 = xasprintf("%s=%f", "offset", offset);
912 /* Other items of potential interest: selected peer,
913 * rootdelay, reftime, rootdisp, refid, ntp_status,
914 * last_update_offset, last_update_recv_time, discipline_jitter,
915 * how many peers have reachable_bits = 0?
918 /* Don't want to wait: it may run hwclock --systohc, and that
919 * may take some time (seconds): */
920 /*spawn_and_wait(argv);*/
924 unsetenv("freq_drift_ppm");
925 unsetenv("poll_interval");
934 step_time(double offset)
938 struct timeval tvc, tvn;
939 char buf[sizeof("yyyy-mm-dd hh:mm:ss") + /*paranoia:*/ 4];
942 gettimeofday(&tvc, NULL); /* never fails */
943 dtime = tvc.tv_sec + (1.0e-6 * tvc.tv_usec) + offset;
944 d_to_tv(dtime, &tvn);
945 if (settimeofday(&tvn, NULL) == -1)
946 bb_perror_msg_and_die("settimeofday");
950 strftime_YYYYMMDDHHMMSS(buf, sizeof(buf), &tval);
951 bb_error_msg("current time is %s.%06u", buf, (unsigned)tvc.tv_usec);
954 strftime_YYYYMMDDHHMMSS(buf, sizeof(buf), &tval);
955 bb_error_msg("setting time to %s.%06u (offset %+fs)", buf, (unsigned)tvn.tv_usec, offset);
957 /* Correct various fields which contain time-relative values: */
960 G.cur_time += offset;
961 G.last_update_recv_time += offset;
962 G.last_script_run += offset;
964 /* p->lastpkt_recv_time, p->next_action_time and such: */
965 for (item = G.ntp_peers; item != NULL; item = item->link) {
966 peer_t *pp = (peer_t *) item->data;
967 reset_peer_stats(pp, offset);
968 //bb_error_msg("offset:%+f pp->next_action_time:%f -> %f",
969 // offset, pp->next_action_time, pp->next_action_time + offset);
970 pp->next_action_time += offset;
972 /* We wait for reply from this peer too.
973 * But due to step we are doing, reply's data is no longer
974 * useful (in fact, it'll be bogus). Stop waiting for it.
978 set_next(pp, RETRY_INTERVAL);
983 static void clamp_pollexp_and_set_MAXSTRAT(void)
985 if (G.poll_exp < MINPOLL)
986 G.poll_exp = MINPOLL;
987 if (G.poll_exp > BIGPOLL)
988 G.poll_exp = BIGPOLL;
990 G.stratum = MAXSTRAT;
995 * Selection and clustering, and their helpers
1001 double opt_rd; /* optimization */
1004 compare_point_edge(const void *aa, const void *bb)
1006 const point_t *a = aa;
1007 const point_t *b = bb;
1008 if (a->edge < b->edge) {
1011 return (a->edge > b->edge);
1018 compare_survivor_metric(const void *aa, const void *bb)
1020 const survivor_t *a = aa;
1021 const survivor_t *b = bb;
1022 if (a->metric < b->metric) {
1025 return (a->metric > b->metric);
1028 fit(peer_t *p, double rd)
1030 if ((p->reachable_bits & (p->reachable_bits-1)) == 0) {
1031 /* One or zero bits in reachable_bits */
1032 VERB4 bb_error_msg("peer %s unfit for selection: unreachable", p->p_dotted);
1035 #if 0 /* we filter out such packets earlier */
1036 if ((p->lastpkt_status & LI_ALARM) == LI_ALARM
1037 || p->lastpkt_stratum >= MAXSTRAT
1039 VERB4 bb_error_msg("peer %s unfit for selection: bad status/stratum", p->p_dotted);
1043 /* rd is root_distance(p) */
1044 if (rd > MAXDIST + FREQ_TOLERANCE * (1 << G.poll_exp)) {
1045 VERB4 bb_error_msg("peer %s unfit for selection: root distance too high", p->p_dotted);
1049 // /* Do we have a loop? */
1050 // if (p->refid == p->dstaddr || p->refid == s.refid)
1055 select_and_cluster(void)
1060 int size = 3 * G.peer_cnt;
1061 /* for selection algorithm */
1062 point_t point[size];
1063 unsigned num_points, num_candidates;
1065 unsigned num_falsetickers;
1066 /* for cluster algorithm */
1067 survivor_t survivor[size];
1068 unsigned num_survivors;
1074 if (G.initial_poll_complete) while (item != NULL) {
1077 p = (peer_t *) item->data;
1078 rd = root_distance(p);
1079 offset = p->filter_offset;
1085 VERB5 bb_error_msg("interval: [%f %f %f] %s",
1091 point[num_points].p = p;
1092 point[num_points].type = -1;
1093 point[num_points].edge = offset - rd;
1094 point[num_points].opt_rd = rd;
1096 point[num_points].p = p;
1097 point[num_points].type = 0;
1098 point[num_points].edge = offset;
1099 point[num_points].opt_rd = rd;
1101 point[num_points].p = p;
1102 point[num_points].type = 1;
1103 point[num_points].edge = offset + rd;
1104 point[num_points].opt_rd = rd;
1108 num_candidates = num_points / 3;
1109 if (num_candidates == 0) {
1110 VERB3 bb_error_msg("no valid datapoints%s", ", no peer selected");
1113 //TODO: sorting does not seem to be done in reference code
1114 qsort(point, num_points, sizeof(point[0]), compare_point_edge);
1116 /* Start with the assumption that there are no falsetickers.
1117 * Attempt to find a nonempty intersection interval containing
1118 * the midpoints of all truechimers.
1119 * If a nonempty interval cannot be found, increase the number
1120 * of assumed falsetickers by one and try again.
1121 * If a nonempty interval is found and the number of falsetickers
1122 * is less than the number of truechimers, a majority has been found
1123 * and the midpoint of each truechimer represents
1124 * the candidates available to the cluster algorithm.
1126 num_falsetickers = 0;
1129 unsigned num_midpoints = 0;
1134 for (i = 0; i < num_points; i++) {
1136 * if (point[i].type == -1) c++;
1137 * if (point[i].type == 1) c--;
1138 * and it's simpler to do it this way:
1141 if (c >= num_candidates - num_falsetickers) {
1142 /* If it was c++ and it got big enough... */
1143 low = point[i].edge;
1146 if (point[i].type == 0)
1150 for (i = num_points-1; i >= 0; i--) {
1152 if (c >= num_candidates - num_falsetickers) {
1153 high = point[i].edge;
1156 if (point[i].type == 0)
1159 /* If the number of midpoints is greater than the number
1160 * of allowed falsetickers, the intersection contains at
1161 * least one truechimer with no midpoint - bad.
1162 * Also, interval should be nonempty.
1164 if (num_midpoints <= num_falsetickers && low < high)
1167 if (num_falsetickers * 2 >= num_candidates) {
1168 VERB3 bb_error_msg("falsetickers:%d, candidates:%d%s",
1169 num_falsetickers, num_candidates,
1170 ", no peer selected");
1174 VERB4 bb_error_msg("selected interval: [%f, %f]; candidates:%d falsetickers:%d",
1175 low, high, num_candidates, num_falsetickers);
1179 /* Construct a list of survivors (p, metric)
1180 * from the chime list, where metric is dominated
1181 * first by stratum and then by root distance.
1182 * All other things being equal, this is the order of preference.
1185 for (i = 0; i < num_points; i++) {
1186 if (point[i].edge < low || point[i].edge > high)
1189 survivor[num_survivors].p = p;
1190 /* x.opt_rd == root_distance(p); */
1191 survivor[num_survivors].metric = MAXDIST * p->lastpkt_stratum + point[i].opt_rd;
1192 VERB5 bb_error_msg("survivor[%d] metric:%f peer:%s",
1193 num_survivors, survivor[num_survivors].metric, p->p_dotted);
1196 /* There must be at least MIN_SELECTED survivors to satisfy the
1197 * correctness assertions. Ordinarily, the Byzantine criteria
1198 * require four survivors, but for the demonstration here, one
1201 if (num_survivors < MIN_SELECTED) {
1202 VERB3 bb_error_msg("survivors:%d%s",
1204 ", no peer selected");
1208 //looks like this is ONLY used by the fact that later we pick survivor[0].
1209 //we can avoid sorting then, just find the minimum once!
1210 qsort(survivor, num_survivors, sizeof(survivor[0]), compare_survivor_metric);
1212 /* For each association p in turn, calculate the selection
1213 * jitter p->sjitter as the square root of the sum of squares
1214 * (p->offset - q->offset) over all q associations. The idea is
1215 * to repeatedly discard the survivor with maximum selection
1216 * jitter until a termination condition is met.
1219 unsigned max_idx = max_idx;
1220 double max_selection_jitter = max_selection_jitter;
1221 double min_jitter = min_jitter;
1223 if (num_survivors <= MIN_CLUSTERED) {
1224 VERB4 bb_error_msg("num_survivors %d <= %d, not discarding more",
1225 num_survivors, MIN_CLUSTERED);
1229 /* To make sure a few survivors are left
1230 * for the clustering algorithm to chew on,
1231 * we stop if the number of survivors
1232 * is less than or equal to MIN_CLUSTERED (3).
1234 for (i = 0; i < num_survivors; i++) {
1235 double selection_jitter_sq;
1238 if (i == 0 || p->filter_jitter < min_jitter)
1239 min_jitter = p->filter_jitter;
1241 selection_jitter_sq = 0;
1242 for (j = 0; j < num_survivors; j++) {
1243 peer_t *q = survivor[j].p;
1244 selection_jitter_sq += SQUARE(p->filter_offset - q->filter_offset);
1246 if (i == 0 || selection_jitter_sq > max_selection_jitter) {
1247 max_selection_jitter = selection_jitter_sq;
1250 VERB6 bb_error_msg("survivor %d selection_jitter^2:%f",
1251 i, selection_jitter_sq);
1253 max_selection_jitter = SQRT(max_selection_jitter / num_survivors);
1254 VERB5 bb_error_msg("max_selection_jitter (at %d):%f min_jitter:%f",
1255 max_idx, max_selection_jitter, min_jitter);
1257 /* If the maximum selection jitter is less than the
1258 * minimum peer jitter, then tossing out more survivors
1259 * will not lower the minimum peer jitter, so we might
1262 if (max_selection_jitter < min_jitter) {
1263 VERB4 bb_error_msg("max_selection_jitter:%f < min_jitter:%f, num_survivors:%d, not discarding more",
1264 max_selection_jitter, min_jitter, num_survivors);
1268 /* Delete survivor[max_idx] from the list
1269 * and go around again.
1271 VERB6 bb_error_msg("dropping survivor %d", max_idx);
1273 while (max_idx < num_survivors) {
1274 survivor[max_idx] = survivor[max_idx + 1];
1280 /* Combine the offsets of the clustering algorithm survivors
1281 * using a weighted average with weight determined by the root
1282 * distance. Compute the selection jitter as the weighted RMS
1283 * difference between the first survivor and the remaining
1284 * survivors. In some cases the inherent clock jitter can be
1285 * reduced by not using this algorithm, especially when frequent
1286 * clockhopping is involved. bbox: thus we don't do it.
1290 for (i = 0; i < num_survivors; i++) {
1292 x = root_distance(p);
1294 z += p->filter_offset / x;
1295 w += SQUARE(p->filter_offset - survivor[0].p->filter_offset) / x;
1297 //G.cluster_offset = z / y;
1298 //G.cluster_jitter = SQRT(w / y);
1301 /* Pick the best clock. If the old system peer is on the list
1302 * and at the same stratum as the first survivor on the list,
1303 * then don't do a clock hop. Otherwise, select the first
1304 * survivor on the list as the new system peer.
1307 if (G.last_update_peer
1308 && G.last_update_peer->lastpkt_stratum <= p->lastpkt_stratum
1310 /* Starting from 1 is ok here */
1311 for (i = 1; i < num_survivors; i++) {
1312 if (G.last_update_peer == survivor[i].p) {
1313 VERB5 bb_error_msg("keeping old synced peer");
1314 p = G.last_update_peer;
1319 G.last_update_peer = p;
1321 VERB4 bb_error_msg("selected peer %s filter_offset:%+f age:%f",
1324 G.cur_time - p->lastpkt_recv_time
1331 * Local clock discipline and its helpers
1334 set_new_values(int disc_state, double offset, double recv_time)
1336 /* Enter new state and set state variables. Note we use the time
1337 * of the last clock filter sample, which must be earlier than
1340 VERB4 bb_error_msg("disc_state=%d last update offset=%f recv_time=%f",
1341 disc_state, offset, recv_time);
1342 G.discipline_state = disc_state;
1343 G.last_update_offset = offset;
1344 G.last_update_recv_time = recv_time;
1346 /* Return: -1: decrease poll interval, 0: leave as is, 1: increase */
1348 update_local_clock(peer_t *p)
1352 /* Note: can use G.cluster_offset instead: */
1353 double offset = p->filter_offset;
1354 double recv_time = p->lastpkt_recv_time;
1356 #if !USING_KERNEL_PLL_LOOP
1359 #if !USING_KERNEL_PLL_LOOP || USING_INITIAL_FREQ_ESTIMATION
1360 double since_last_update;
1362 double etemp, dtemp;
1364 abs_offset = fabs(offset);
1367 /* If needed, -S script can do it by looking at $offset
1368 * env var and killing parent */
1369 /* If the offset is too large, give up and go home */
1370 if (abs_offset > PANIC_THRESHOLD) {
1371 bb_error_msg_and_die("offset %f far too big, exiting", offset);
1375 /* If this is an old update, for instance as the result
1376 * of a system peer change, avoid it. We never use
1377 * an old sample or the same sample twice.
1379 if (recv_time <= G.last_update_recv_time) {
1380 VERB3 bb_error_msg("update from %s: same or older datapoint, not using it",
1382 return 0; /* "leave poll interval as is" */
1385 /* Clock state machine transition function. This is where the
1386 * action is and defines how the system reacts to large time
1387 * and frequency errors.
1389 #if !USING_KERNEL_PLL_LOOP || USING_INITIAL_FREQ_ESTIMATION
1390 since_last_update = recv_time - G.reftime;
1392 #if !USING_KERNEL_PLL_LOOP
1395 #if USING_INITIAL_FREQ_ESTIMATION
1396 if (G.discipline_state == STATE_FREQ) {
1397 /* Ignore updates until the stepout threshold */
1398 if (since_last_update < WATCH_THRESHOLD) {
1399 VERB4 bb_error_msg("measuring drift, datapoint ignored, %f sec remains",
1400 WATCH_THRESHOLD - since_last_update);
1401 return 0; /* "leave poll interval as is" */
1403 # if !USING_KERNEL_PLL_LOOP
1404 freq_drift = (offset - G.last_update_offset) / since_last_update;
1409 /* There are two main regimes: when the
1410 * offset exceeds the step threshold and when it does not.
1412 if (abs_offset > STEP_THRESHOLD) {
1416 // This "spike state" seems to be useless, peer selection already drops
1417 // occassional "bad" datapoints. If we are here, there were _many_
1418 // large offsets. When a few first large offsets are seen,
1419 // we end up in "no valid datapoints, no peer selected" state.
1420 // Only when enough of them are seen (which means it's not a fluke),
1421 // we end up here. Looks like _our_ clock is off.
1422 switch (G.discipline_state) {
1424 /* The first outlyer: ignore it, switch to SPIK state */
1425 VERB3 bb_error_msg("update from %s: offset:%+f, spike%s",
1426 p->p_dotted, offset,
1428 G.discipline_state = STATE_SPIK;
1429 return -1; /* "decrease poll interval" */
1432 /* Ignore succeeding outlyers until either an inlyer
1433 * is found or the stepout threshold is exceeded.
1435 remains = WATCH_THRESHOLD - since_last_update;
1437 VERB3 bb_error_msg("update from %s: offset:%+f, spike%s",
1438 p->p_dotted, offset,
1439 ", datapoint ignored");
1440 return -1; /* "decrease poll interval" */
1442 /* fall through: we need to step */
1446 /* Step the time and clamp down the poll interval.
1448 * In NSET state an initial frequency correction is
1449 * not available, usually because the frequency file has
1450 * not yet been written. Since the time is outside the
1451 * capture range, the clock is stepped. The frequency
1452 * will be set directly following the stepout interval.
1454 * In FSET state the initial frequency has been set
1455 * from the frequency file. Since the time is outside
1456 * the capture range, the clock is stepped immediately,
1457 * rather than after the stepout interval. Guys get
1458 * nervous if it takes 17 minutes to set the clock for
1461 * In SPIK state the stepout threshold has expired and
1462 * the phase is still above the step threshold. Note
1463 * that a single spike greater than the step threshold
1464 * is always suppressed, even at the longer poll
1467 VERB4 bb_error_msg("stepping time by %+f; poll_exp=MINPOLL", offset);
1469 if (option_mask32 & OPT_q) {
1470 /* We were only asked to set time once. Done. */
1474 clamp_pollexp_and_set_MAXSTRAT();
1476 run_script("step", offset);
1478 recv_time += offset;
1480 #if USING_INITIAL_FREQ_ESTIMATION
1481 if (G.discipline_state == STATE_NSET) {
1482 set_new_values(STATE_FREQ, /*offset:*/ 0, recv_time);
1483 return 1; /* "ok to increase poll interval" */
1486 abs_offset = offset = 0;
1487 set_new_values(STATE_SYNC, offset, recv_time);
1489 } else { /* abs_offset <= STEP_THRESHOLD */
1491 /* The ratio is calculated before jitter is updated to make
1492 * poll adjust code more sensitive to large offsets.
1494 G.offset_to_jitter_ratio = abs_offset / G.discipline_jitter;
1496 /* Compute the clock jitter as the RMS of exponentially
1497 * weighted offset differences. Used by the poll adjust code.
1499 etemp = SQUARE(G.discipline_jitter);
1500 dtemp = SQUARE(offset - G.last_update_offset);
1501 G.discipline_jitter = SQRT(etemp + (dtemp - etemp) / AVG);
1502 if (G.discipline_jitter < G_precision_sec)
1503 G.discipline_jitter = G_precision_sec;
1505 switch (G.discipline_state) {
1507 if (option_mask32 & OPT_q) {
1508 /* We were only asked to set time once.
1509 * The clock is precise enough, no need to step.
1513 #if USING_INITIAL_FREQ_ESTIMATION
1514 /* This is the first update received and the frequency
1515 * has not been initialized. The first thing to do
1516 * is directly measure the oscillator frequency.
1518 set_new_values(STATE_FREQ, offset, recv_time);
1520 set_new_values(STATE_SYNC, offset, recv_time);
1522 VERB4 bb_error_msg("transitioning to FREQ, datapoint ignored");
1523 return 0; /* "leave poll interval as is" */
1525 #if 0 /* this is dead code for now */
1527 /* This is the first update and the frequency
1528 * has been initialized. Adjust the phase, but
1529 * don't adjust the frequency until the next update.
1531 set_new_values(STATE_SYNC, offset, recv_time);
1532 /* freq_drift remains 0 */
1536 #if USING_INITIAL_FREQ_ESTIMATION
1538 /* since_last_update >= WATCH_THRESHOLD, we waited enough.
1539 * Correct the phase and frequency and switch to SYNC state.
1540 * freq_drift was already estimated (see code above)
1542 set_new_values(STATE_SYNC, offset, recv_time);
1547 #if !USING_KERNEL_PLL_LOOP
1548 /* Compute freq_drift due to PLL and FLL contributions.
1550 * The FLL and PLL frequency gain constants
1551 * depend on the poll interval and Allan
1552 * intercept. The FLL is not used below one-half
1553 * the Allan intercept. Above that the loop gain
1554 * increases in steps to 1 / AVG.
1556 if ((1 << G.poll_exp) > ALLAN / 2) {
1557 etemp = FLL - G.poll_exp;
1560 freq_drift += (offset - G.last_update_offset) / (MAXD(since_last_update, ALLAN) * etemp);
1562 /* For the PLL the integration interval
1563 * (numerator) is the minimum of the update
1564 * interval and poll interval. This allows
1565 * oversampling, but not undersampling.
1567 etemp = MIND(since_last_update, (1 << G.poll_exp));
1568 dtemp = (4 * PLL) << G.poll_exp;
1569 freq_drift += offset * etemp / SQUARE(dtemp);
1571 set_new_values(STATE_SYNC, offset, recv_time);
1574 if (G.stratum != p->lastpkt_stratum + 1) {
1575 G.stratum = p->lastpkt_stratum + 1;
1576 run_script("stratum", offset);
1580 G.reftime = G.cur_time;
1581 G.ntp_status = p->lastpkt_status;
1582 G.refid = p->lastpkt_refid;
1583 G.rootdelay = p->lastpkt_rootdelay + p->lastpkt_delay;
1584 dtemp = p->filter_jitter; // SQRT(SQUARE(p->filter_jitter) + SQUARE(G.cluster_jitter));
1585 dtemp += MAXD(p->filter_dispersion + FREQ_TOLERANCE * (G.cur_time - p->lastpkt_recv_time) + abs_offset, MINDISP);
1586 G.rootdisp = p->lastpkt_rootdisp + dtemp;
1587 VERB4 bb_error_msg("updating leap/refid/reftime/rootdisp from peer %s", p->p_dotted);
1589 /* We are in STATE_SYNC now, but did not do adjtimex yet.
1590 * (Any other state does not reach this, they all return earlier)
1591 * By this time, freq_drift and offset are set
1592 * to values suitable for adjtimex.
1594 #if !USING_KERNEL_PLL_LOOP
1595 /* Calculate the new frequency drift and frequency stability (wander).
1596 * Compute the clock wander as the RMS of exponentially weighted
1597 * frequency differences. This is not used directly, but can,
1598 * along with the jitter, be a highly useful monitoring and
1601 dtemp = G.discipline_freq_drift + freq_drift;
1602 G.discipline_freq_drift = MAXD(MIND(MAXDRIFT, dtemp), -MAXDRIFT);
1603 etemp = SQUARE(G.discipline_wander);
1604 dtemp = SQUARE(dtemp);
1605 G.discipline_wander = SQRT(etemp + (dtemp - etemp) / AVG);
1607 VERB4 bb_error_msg("discipline freq_drift=%.9f(int:%ld corr:%e) wander=%f",
1608 G.discipline_freq_drift,
1609 (long)(G.discipline_freq_drift * 65536e6),
1611 G.discipline_wander);
1614 memset(&tmx, 0, sizeof(tmx));
1615 if (adjtimex(&tmx) < 0)
1616 bb_perror_msg_and_die("adjtimex");
1617 bb_error_msg("p adjtimex freq:%ld offset:%+ld status:0x%x tc:%ld",
1618 tmx.freq, tmx.offset, tmx.status, tmx.constant);
1621 memset(&tmx, 0, sizeof(tmx));
1623 //doesn't work, offset remains 0 (!) in kernel:
1624 //ntpd: set adjtimex freq:1786097 tmx.offset:77487
1625 //ntpd: prev adjtimex freq:1786097 tmx.offset:0
1626 //ntpd: cur adjtimex freq:1786097 tmx.offset:0
1627 tmx.modes = ADJ_FREQUENCY | ADJ_OFFSET;
1628 /* 65536 is one ppm */
1629 tmx.freq = G.discipline_freq_drift * 65536e6;
1631 tmx.modes = ADJ_OFFSET | ADJ_STATUS | ADJ_TIMECONST;// | ADJ_MAXERROR | ADJ_ESTERROR;
1632 tmx.offset = (offset * 1000000); /* usec */
1633 tmx.status = STA_PLL;
1634 if (G.ntp_status & LI_PLUSSEC)
1635 tmx.status |= STA_INS;
1636 if (G.ntp_status & LI_MINUSSEC)
1637 tmx.status |= STA_DEL;
1639 tmx.constant = G.poll_exp - 4;
1641 * The below if statement should be unnecessary, but...
1642 * It looks like Linux kernel's PLL is far too gentle in changing
1643 * tmx.freq in response to clock offset. Offset keeps growing
1644 * and eventually we fall back to smaller poll intervals.
1645 * We can make correction more agressive (about x2) by supplying
1646 * PLL time constant which is one less than the real one.
1647 * To be on a safe side, let's do it only if offset is significantly
1648 * larger than jitter.
1650 if (tmx.constant > 0 && G.offset_to_jitter_ratio >= TIMECONST_HACK_GATE)
1653 //tmx.esterror = (uint32_t)(clock_jitter * 1e6);
1654 //tmx.maxerror = (uint32_t)((sys_rootdelay / 2 + sys_rootdisp) * 1e6);
1655 rc = adjtimex(&tmx);
1657 bb_perror_msg_and_die("adjtimex");
1658 /* NB: here kernel returns constant == G.poll_exp, not == G.poll_exp - 4.
1659 * Not sure why. Perhaps it is normal.
1661 VERB4 bb_error_msg("adjtimex:%d freq:%ld offset:%+ld status:0x%x",
1662 rc, tmx.freq, tmx.offset, tmx.status);
1663 G.kernel_freq_drift = tmx.freq / 65536;
1664 VERB2 bb_error_msg("update from:%s offset:%+f jitter:%f clock drift:%+.3fppm tc:%d",
1665 p->p_dotted, offset, G.discipline_jitter, (double)tmx.freq / 65536, (int)tmx.constant);
1667 return 1; /* "ok to increase poll interval" */
1672 * We've got a new reply packet from a peer, process it
1676 poll_interval(int upper_bound)
1678 unsigned interval, r, mask;
1679 interval = 1 << G.poll_exp;
1680 if (interval > upper_bound)
1681 interval = upper_bound;
1682 mask = ((interval-1) >> 4) | 1;
1684 interval += r & mask; /* ~ random(0..1) * interval/16 */
1685 VERB4 bb_error_msg("chose poll interval:%u (poll_exp:%d)", interval, G.poll_exp);
1689 adjust_poll(int count)
1691 G.polladj_count += count;
1692 if (G.polladj_count > POLLADJ_LIMIT) {
1693 G.polladj_count = 0;
1694 if (G.poll_exp < MAXPOLL) {
1696 VERB4 bb_error_msg("polladj: discipline_jitter:%f ++poll_exp=%d",
1697 G.discipline_jitter, G.poll_exp);
1699 } else if (G.polladj_count < -POLLADJ_LIMIT || (count < 0 && G.poll_exp > BIGPOLL)) {
1700 G.polladj_count = 0;
1701 if (G.poll_exp > MINPOLL) {
1705 /* Correct p->next_action_time in each peer
1706 * which waits for sending, so that they send earlier.
1707 * Old pp->next_action_time are on the order
1708 * of t + (1 << old_poll_exp) + small_random,
1709 * we simply need to subtract ~half of that.
1711 for (item = G.ntp_peers; item != NULL; item = item->link) {
1712 peer_t *pp = (peer_t *) item->data;
1714 pp->next_action_time -= (1 << G.poll_exp);
1716 VERB4 bb_error_msg("polladj: discipline_jitter:%f --poll_exp=%d",
1717 G.discipline_jitter, G.poll_exp);
1720 VERB4 bb_error_msg("polladj: count:%d", G.polladj_count);
1723 static NOINLINE void
1724 recv_and_process_peer_pkt(peer_t *p)
1729 double T1, T2, T3, T4;
1731 double prev_delay, delay;
1733 datapoint_t *datapoint;
1738 /* We can recvfrom here and check from.IP, but some multihomed
1739 * ntp servers reply from their *other IP*.
1740 * TODO: maybe we should check at least what we can: from.port == 123?
1743 size = recv(p->p_fd, &msg, sizeof(msg), MSG_DONTWAIT);
1748 if (errno == EAGAIN)
1749 /* There was no packet after all
1750 * (poll() returning POLLIN for a fd
1751 * is not a ironclad guarantee that data is there)
1755 * If you need a different handling for a specific
1756 * errno, always explain it in comment.
1758 bb_perror_msg_and_die("recv(%s) error", p->p_dotted);
1761 if (size != NTP_MSGSIZE_NOAUTH && size != NTP_MSGSIZE) {
1762 bb_error_msg("malformed packet received from %s", p->p_dotted);
1766 if (msg.m_orgtime.int_partl != p->p_xmt_msg.m_xmttime.int_partl
1767 || msg.m_orgtime.fractionl != p->p_xmt_msg.m_xmttime.fractionl
1769 /* Somebody else's packet */
1773 /* We do not expect any more packets from this peer for now.
1774 * Closing the socket informs kernel about it.
1775 * We open a new socket when we send a new query.
1780 if ((msg.m_status & LI_ALARM) == LI_ALARM
1781 || msg.m_stratum == 0
1782 || msg.m_stratum > NTP_MAXSTRATUM
1784 bb_error_msg("reply from %s: peer is unsynced", p->p_dotted);
1786 * Stratum 0 responses may have commands in 32-bit m_refid field:
1787 * "DENY", "RSTR" - peer does not like us at all,
1788 * "RATE" - peer is overloaded, reduce polling freq.
1789 * If poll interval is small, increase it.
1791 if (G.poll_exp < BIGPOLL)
1792 goto increase_interval;
1793 goto pick_normal_interval;
1796 // /* Verify valid root distance */
1797 // if (msg.m_rootdelay / 2 + msg.m_rootdisp >= MAXDISP || p->lastpkt_reftime > msg.m_xmt)
1798 // return; /* invalid header values */
1801 * From RFC 2030 (with a correction to the delay math):
1803 * Timestamp Name ID When Generated
1804 * ------------------------------------------------------------
1805 * Originate Timestamp T1 time request sent by client
1806 * Receive Timestamp T2 time request received by server
1807 * Transmit Timestamp T3 time reply sent by server
1808 * Destination Timestamp T4 time reply received by client
1810 * The roundtrip delay and local clock offset are defined as
1812 * delay = (T4 - T1) - (T3 - T2); offset = ((T2 - T1) + (T3 - T4)) / 2
1815 T2 = lfp_to_d(msg.m_rectime);
1816 T3 = lfp_to_d(msg.m_xmttime);
1819 /* The delay calculation is a special case. In cases where the
1820 * server and client clocks are running at different rates and
1821 * with very fast networks, the delay can appear negative. In
1822 * order to avoid violating the Principle of Least Astonishment,
1823 * the delay is clamped not less than the system precision.
1825 delay = (T4 - T1) - (T3 - T2);
1826 if (delay < G_precision_sec)
1827 delay = G_precision_sec;
1829 * If this packet's delay is much bigger than the last one,
1830 * it's better to just ignore it than use its much less precise value.
1832 prev_delay = p->p_raw_delay;
1833 p->p_raw_delay = delay;
1834 if (p->reachable_bits && delay > prev_delay * BAD_DELAY_GROWTH) {
1835 bb_error_msg("reply from %s: delay %f is too high, ignoring", p->p_dotted, delay);
1836 goto pick_normal_interval;
1839 p->lastpkt_delay = delay;
1840 p->lastpkt_recv_time = T4;
1841 VERB6 bb_error_msg("%s->lastpkt_recv_time=%f", p->p_dotted, p->lastpkt_recv_time);
1842 p->lastpkt_status = msg.m_status;
1843 p->lastpkt_stratum = msg.m_stratum;
1844 p->lastpkt_rootdelay = sfp_to_d(msg.m_rootdelay);
1845 p->lastpkt_rootdisp = sfp_to_d(msg.m_rootdisp);
1846 p->lastpkt_refid = msg.m_refid;
1848 p->datapoint_idx = p->reachable_bits ? (p->datapoint_idx + 1) % NUM_DATAPOINTS : 0;
1849 datapoint = &p->filter_datapoint[p->datapoint_idx];
1850 datapoint->d_recv_time = T4;
1851 datapoint->d_offset = offset = ((T2 - T1) + (T3 - T4)) / 2;
1852 datapoint->d_dispersion = LOG2D(msg.m_precision_exp) + G_precision_sec;
1853 if (!p->reachable_bits) {
1854 /* 1st datapoint ever - replicate offset in every element */
1856 for (i = 0; i < NUM_DATAPOINTS; i++) {
1857 p->filter_datapoint[i].d_offset = offset;
1861 p->reachable_bits |= 1;
1862 if ((MAX_VERBOSE && G.verbose) || (option_mask32 & OPT_w)) {
1863 bb_error_msg("reply from %s: offset:%+f delay:%f status:0x%02x strat:%d refid:0x%08x rootdelay:%f reach:0x%02x",
1870 p->lastpkt_rootdelay,
1872 /* not shown: m_ppoll, m_precision_exp, m_rootdisp,
1873 * m_reftime, m_orgtime, m_rectime, m_xmttime
1878 /* Muck with statictics and update the clock */
1879 filter_datapoints(p);
1880 q = select_and_cluster();
1883 if (!(option_mask32 & OPT_w)) {
1884 rc = update_local_clock(q);
1886 //Disabled this because there is a case where largish offsets
1887 //are unavoidable: if network round-trip delay is, say, ~0.6s,
1888 //error in offset estimation would be ~delay/2 ~= 0.3s.
1889 //Thus, offsets will be usually in -0.3...0.3s range.
1890 //In this case, this code would keep poll interval small,
1891 //but it won't be helping.
1892 //BIGOFF check below deals with a case of seeing multi-second offsets.
1894 /* If drift is dangerously large, immediately
1895 * drop poll interval one step down.
1897 if (fabs(q->filter_offset) >= POLLDOWN_OFFSET) {
1898 VERB4 bb_error_msg("offset:%+f > POLLDOWN_OFFSET", q->filter_offset);
1899 adjust_poll(-POLLADJ_LIMIT * 3);
1905 /* No peer selected.
1906 * If poll interval is small, increase it.
1908 if (G.poll_exp < BIGPOLL)
1909 goto increase_interval;
1913 /* Adjust the poll interval by comparing the current offset
1914 * with the clock jitter. If the offset is less than
1915 * the clock jitter times a constant, then the averaging interval
1916 * is increased, otherwise it is decreased. A bit of hysteresis
1917 * helps calm the dance. Works best using burst mode.
1919 if (rc > 0 && G.offset_to_jitter_ratio <= POLLADJ_GATE) {
1920 /* was += G.poll_exp but it is a bit
1921 * too optimistic for my taste at high poll_exp's */
1923 adjust_poll(MINPOLL);
1925 adjust_poll(-G.poll_exp * 2);
1929 /* Decide when to send new query for this peer */
1930 pick_normal_interval:
1931 interval = poll_interval(INT_MAX);
1932 if (fabs(offset) >= BIGOFF && interval > BIGOFF_INTERVAL) {
1933 /* If we are synced, offsets are less than STEP_THRESHOLD,
1934 * or at the very least not much larger than it.
1935 * Now we see a largish one.
1936 * Either this peer is feeling bad, or packet got corrupted,
1937 * or _our_ clock is wrong now and _all_ peers will show similar
1938 * largish offsets too.
1939 * I observed this with laptop suspend stopping clock.
1940 * In any case, it makes sense to make next request soonish:
1941 * cases 1 and 2: get a better datapoint,
1942 * case 3: allows to resync faster.
1944 interval = BIGOFF_INTERVAL;
1947 set_next(p, interval);
1950 #if ENABLE_FEATURE_NTPD_SERVER
1951 static NOINLINE void
1952 recv_and_process_client_pkt(void /*int fd*/)
1956 len_and_sockaddr *to;
1957 struct sockaddr *from;
1959 uint8_t query_status;
1960 l_fixedpt_t query_xmttime;
1962 to = get_sock_lsa(G_listen_fd);
1963 from = xzalloc(to->len);
1965 size = recv_from_to(G_listen_fd, &msg, sizeof(msg), MSG_DONTWAIT, from, &to->u.sa, to->len);
1966 if (size != NTP_MSGSIZE_NOAUTH && size != NTP_MSGSIZE) {
1969 if (errno == EAGAIN)
1971 bb_perror_msg_and_die("recv");
1973 addr = xmalloc_sockaddr2dotted_noport(from);
1974 bb_error_msg("malformed packet received from %s: size %u", addr, (int)size);
1979 query_status = msg.m_status;
1980 query_xmttime = msg.m_xmttime;
1982 /* Build a reply packet */
1983 memset(&msg, 0, sizeof(msg));
1984 msg.m_status = G.stratum < MAXSTRAT ? (G.ntp_status & LI_MASK) : LI_ALARM;
1985 msg.m_status |= (query_status & VERSION_MASK);
1986 msg.m_status |= ((query_status & MODE_MASK) == MODE_CLIENT) ?
1987 MODE_SERVER : MODE_SYM_PAS;
1988 msg.m_stratum = G.stratum;
1989 msg.m_ppoll = G.poll_exp;
1990 msg.m_precision_exp = G_precision_exp;
1991 /* this time was obtained between poll() and recv() */
1992 msg.m_rectime = d_to_lfp(G.cur_time);
1993 msg.m_xmttime = d_to_lfp(gettime1900d()); /* this instant */
1994 if (G.peer_cnt == 0) {
1995 /* we have no peers: "stratum 1 server" mode. reftime = our own time */
1996 G.reftime = G.cur_time;
1998 msg.m_reftime = d_to_lfp(G.reftime);
1999 msg.m_orgtime = query_xmttime;
2000 msg.m_rootdelay = d_to_sfp(G.rootdelay);
2001 //simple code does not do this, fix simple code!
2002 msg.m_rootdisp = d_to_sfp(G.rootdisp);
2003 //version = (query_status & VERSION_MASK); /* ... >> VERSION_SHIFT - done below instead */
2004 msg.m_refid = G.refid; // (version > (3 << VERSION_SHIFT)) ? G.refid : G.refid3;
2006 /* We reply from the local address packet was sent to,
2007 * this makes to/from look swapped here: */
2008 do_sendto(G_listen_fd,
2009 /*from:*/ &to->u.sa, /*to:*/ from, /*addrlen:*/ to->len,
2018 /* Upstream ntpd's options:
2020 * -4 Force DNS resolution of host names to the IPv4 namespace.
2021 * -6 Force DNS resolution of host names to the IPv6 namespace.
2022 * -a Require cryptographic authentication for broadcast client,
2023 * multicast client and symmetric passive associations.
2024 * This is the default.
2025 * -A Do not require cryptographic authentication for broadcast client,
2026 * multicast client and symmetric passive associations.
2027 * This is almost never a good idea.
2028 * -b Enable the client to synchronize to broadcast servers.
2030 * Specify the name and path of the configuration file,
2031 * default /etc/ntp.conf
2032 * -d Specify debugging mode. This option may occur more than once,
2033 * with each occurrence indicating greater detail of display.
2035 * Specify debugging level directly.
2037 * Specify the name and path of the frequency file.
2038 * This is the same operation as the "driftfile FILE"
2039 * configuration command.
2040 * -g Normally, ntpd exits with a message to the system log
2041 * if the offset exceeds the panic threshold, which is 1000 s
2042 * by default. This option allows the time to be set to any value
2043 * without restriction; however, this can happen only once.
2044 * If the threshold is exceeded after that, ntpd will exit
2045 * with a message to the system log. This option can be used
2046 * with the -q and -x options. See the tinker command for other options.
2048 * Chroot the server to the directory jaildir. This option also implies
2049 * that the server attempts to drop root privileges at startup
2050 * (otherwise, chroot gives very little additional security).
2051 * You may need to also specify a -u option.
2053 * Specify the name and path of the symmetric key file,
2054 * default /etc/ntp/keys. This is the same operation
2055 * as the "keys FILE" configuration command.
2057 * Specify the name and path of the log file. The default
2058 * is the system log file. This is the same operation as
2059 * the "logfile FILE" configuration command.
2060 * -L Do not listen to virtual IPs. The default is to listen.
2062 * -N To the extent permitted by the operating system,
2063 * run the ntpd at the highest priority.
2065 * Specify the name and path of the file used to record the ntpd
2066 * process ID. This is the same operation as the "pidfile FILE"
2067 * configuration command.
2069 * To the extent permitted by the operating system,
2070 * run the ntpd at the specified priority.
2071 * -q Exit the ntpd just after the first time the clock is set.
2072 * This behavior mimics that of the ntpdate program, which is
2073 * to be retired. The -g and -x options can be used with this option.
2074 * Note: The kernel time discipline is disabled with this option.
2076 * Specify the default propagation delay from the broadcast/multicast
2077 * server to this client. This is necessary only if the delay
2078 * cannot be computed automatically by the protocol.
2080 * Specify the directory path for files created by the statistics
2081 * facility. This is the same operation as the "statsdir DIR"
2082 * configuration command.
2084 * Add a key number to the trusted key list. This option can occur
2087 * Specify a user, and optionally a group, to switch to.
2090 * Add a system variable listed by default.
2091 * -x Normally, the time is slewed if the offset is less than the step
2092 * threshold, which is 128 ms by default, and stepped if above
2093 * the threshold. This option sets the threshold to 600 s, which is
2094 * well within the accuracy window to set the clock manually.
2095 * Note: since the slew rate of typical Unix kernels is limited
2096 * to 0.5 ms/s, each second of adjustment requires an amortization
2097 * interval of 2000 s. Thus, an adjustment as much as 600 s
2098 * will take almost 14 days to complete. This option can be used
2099 * with the -g and -q options. See the tinker command for other options.
2100 * Note: The kernel time discipline is disabled with this option.
2103 /* By doing init in a separate function we decrease stack usage
2106 static NOINLINE void ntp_init(char **argv)
2114 bb_error_msg_and_die(bb_msg_you_must_be_root);
2116 /* Set some globals */
2117 G.discipline_jitter = G_precision_sec;
2118 G.stratum = MAXSTRAT;
2120 G.poll_exp = BURSTPOLL; /* speeds up initial sync */
2121 G.last_script_run = G.reftime = G.last_update_recv_time = gettime1900d(); /* sets G.cur_time too */
2125 opt_complementary = "dd:p::wn" /* -d: counter; -p: list; -w implies -n */
2126 IF_FEATURE_NTPD_SERVER(":Il"); /* -I implies -l */
2127 opts = getopt32(argv,
2129 "wp:S:"IF_FEATURE_NTPD_SERVER("l") /* NOT compat */
2130 IF_FEATURE_NTPD_SERVER("I:") /* compat */
2132 "46aAbgL", /* compat, ignored */
2133 &peers,&G.script_name,
2134 #if ENABLE_FEATURE_NTPD_SERVER
2139 // if (opts & OPT_x) /* disable stepping, only slew is allowed */
2140 // G.time_was_stepped = 1;
2143 add_peers(llist_pop(&peers));
2145 #if ENABLE_FEATURE_NTPD_CONF
2150 parser = config_open("/etc/ntp.conf");
2151 while (config_read(parser, token, 3, 1, "# \t", PARSE_NORMAL)) {
2152 if (strcmp(token[0], "server") == 0 && token[1]) {
2153 add_peers(token[1]);
2156 bb_error_msg("skipping %s:%u: unimplemented command '%s'",
2157 "/etc/ntp.conf", parser->lineno, token[0]
2160 config_close(parser);
2163 if (G.peer_cnt == 0) {
2164 if (!(opts & OPT_l))
2166 /* -l but no peers: "stratum 1 server" mode */
2169 #if ENABLE_FEATURE_NTPD_SERVER
2172 G_listen_fd = create_and_bind_dgram_or_die(NULL, 123);
2174 if (setsockopt_bindtodevice(G_listen_fd, G.if_name))
2177 socket_want_pktinfo(G_listen_fd);
2178 setsockopt(G_listen_fd, IPPROTO_IP, IP_TOS, &const_IPTOS_LOWDELAY, sizeof(const_IPTOS_LOWDELAY));
2181 if (!(opts & OPT_n)) {
2182 bb_daemonize_or_rexec(DAEMON_DEVNULL_STDIO, argv);
2183 logmode = LOGMODE_NONE;
2185 /* I hesitate to set -20 prio. -15 should be high enough for timekeeping */
2187 setpriority(PRIO_PROCESS, 0, -15);
2189 /* If network is up, syncronization occurs in ~10 seconds.
2190 * We give "ntpd -q" 10 seconds to get first reply,
2191 * then another 50 seconds to finish syncing.
2193 * I tested ntpd 4.2.6p1 and apparently it never exits
2194 * (will try forever), but it does not feel right.
2195 * The goal of -q is to act like ntpdate: set time
2196 * after a reasonably small period of polling, or fail.
2199 option_mask32 |= OPT_qq;
2216 int ntpd_main(int argc UNUSED_PARAM, char **argv) MAIN_EXTERNALLY_VISIBLE;
2217 int ntpd_main(int argc UNUSED_PARAM, char **argv)
2225 memset(&G, 0, sizeof(G));
2226 SET_PTR_TO_GLOBALS(&G);
2230 /* If ENABLE_FEATURE_NTPD_SERVER, + 1 for listen_fd: */
2231 cnt = G.peer_cnt + ENABLE_FEATURE_NTPD_SERVER;
2232 idx2peer = xzalloc(sizeof(idx2peer[0]) * cnt);
2233 pfd = xzalloc(sizeof(pfd[0]) * cnt);
2235 /* Countdown: we never sync before we sent INITIAL_SAMPLES+1
2236 * packets to each peer.
2237 * NB: if some peer is not responding, we may end up sending
2238 * fewer packets to it and more to other peers.
2239 * NB2: sync usually happens using INITIAL_SAMPLES packets,
2240 * since last reply does not come back instantaneously.
2242 cnt = G.peer_cnt * (INITIAL_SAMPLES + 1);
2244 write_pidfile(CONFIG_PID_FILE_PATH "/ntpd.pid");
2246 while (!bb_got_signal) {
2252 /* Nothing between here and poll() blocks for any significant time */
2254 nextaction = G.cur_time + 3600;
2257 #if ENABLE_FEATURE_NTPD_SERVER
2258 if (G_listen_fd != -1) {
2259 pfd[0].fd = G_listen_fd;
2260 pfd[0].events = POLLIN;
2264 /* Pass over peer list, send requests, time out on receives */
2265 for (item = G.ntp_peers; item != NULL; item = item->link) {
2266 peer_t *p = (peer_t *) item->data;
2268 if (p->next_action_time <= G.cur_time) {
2269 if (p->p_fd == -1) {
2270 /* Time to send new req */
2272 VERB4 bb_error_msg("disabling burst mode");
2273 G.polladj_count = 0;
2274 G.poll_exp = MINPOLL;
2275 G.initial_poll_complete = 1;
2277 send_query_to_peer(p);
2279 /* Timed out waiting for reply */
2282 /* If poll interval is small, increase it */
2283 if (G.poll_exp < BIGPOLL)
2284 adjust_poll(MINPOLL);
2285 timeout = poll_interval(NOREPLY_INTERVAL);
2286 bb_error_msg("timed out waiting for %s, reach 0x%02x, next query in %us",
2287 p->p_dotted, p->reachable_bits, timeout);
2288 set_next(p, timeout);
2292 if (p->next_action_time < nextaction)
2293 nextaction = p->next_action_time;
2296 /* Wait for reply from this peer */
2297 pfd[i].fd = p->p_fd;
2298 pfd[i].events = POLLIN;
2304 timeout = nextaction - G.cur_time;
2307 timeout++; /* (nextaction - G.cur_time) rounds down, compensating */
2309 /* Here we may block */
2311 if (i > (ENABLE_FEATURE_NTPD_SERVER && G_listen_fd != -1)) {
2312 /* We wait for at least one reply.
2313 * Poll for it, without wasting time for message.
2314 * Since replies often come under 1 second, this also
2315 * reduces clutter in logs.
2317 nfds = poll(pfd, i, 1000);
2323 bb_error_msg("poll:%us sockets:%u interval:%us", timeout, i, 1 << G.poll_exp);
2325 nfds = poll(pfd, i, timeout * 1000);
2327 gettime1900d(); /* sets G.cur_time */
2329 if (!bb_got_signal /* poll wasn't interrupted by a signal */
2330 && G.cur_time - G.last_script_run > 11*60
2332 /* Useful for updating battery-backed RTC and such */
2333 run_script("periodic", G.last_update_offset);
2334 gettime1900d(); /* sets G.cur_time */
2339 /* Process any received packets */
2341 #if ENABLE_FEATURE_NTPD_SERVER
2342 if (G.listen_fd != -1) {
2343 if (pfd[0].revents /* & (POLLIN|POLLERR)*/) {
2345 recv_and_process_client_pkt(/*G.listen_fd*/);
2346 gettime1900d(); /* sets G.cur_time */
2351 for (; nfds != 0 && j < i; j++) {
2352 if (pfd[j].revents /* & (POLLIN|POLLERR)*/) {
2354 * At init, alarm was set to 10 sec.
2355 * Now we did get a reply.
2356 * Increase timeout to 50 seconds to finish syncing.
2358 if (option_mask32 & OPT_qq) {
2359 option_mask32 &= ~OPT_qq;
2363 recv_and_process_peer_pkt(idx2peer[j]);
2364 gettime1900d(); /* sets G.cur_time */
2369 if (G.ntp_peers && G.stratum != MAXSTRAT) {
2370 for (item = G.ntp_peers; item != NULL; item = item->link) {
2371 peer_t *p = (peer_t *) item->data;
2372 if (p->reachable_bits)
2373 goto have_reachable_peer;
2375 /* No peer responded for last 8 packets, panic */
2376 clamp_pollexp_and_set_MAXSTRAT();
2377 run_script("unsync", 0.0);
2378 have_reachable_peer: ;
2380 } /* while (!bb_got_signal) */
2382 remove_pidfile(CONFIG_PID_FILE_PATH "/ntpd.pid");
2383 kill_myself_with_sig(bb_got_signal);
2391 /*** openntpd-4.6 uses only adjtime, not adjtimex ***/
2393 /*** ntp-4.2.6/ntpd/ntp_loopfilter.c - adjtimex usage ***/
2397 direct_freq(double fp_offset)
2401 * If the kernel is enabled, we need the residual offset to
2402 * calculate the frequency correction.
2404 if (pll_control && kern_enable) {
2405 memset(&ntv, 0, sizeof(ntv));
2408 clock_offset = ntv.offset / 1e9;
2409 #else /* STA_NANO */
2410 clock_offset = ntv.offset / 1e6;
2411 #endif /* STA_NANO */
2412 drift_comp = FREQTOD(ntv.freq);
2414 #endif /* KERNEL_PLL */
2415 set_freq((fp_offset - clock_offset) / (current_time - clock_epoch) + drift_comp);
2421 set_freq(double freq) /* frequency update */
2429 * If the kernel is enabled, update the kernel frequency.
2431 if (pll_control && kern_enable) {
2432 memset(&ntv, 0, sizeof(ntv));
2433 ntv.modes = MOD_FREQUENCY;
2434 ntv.freq = DTOFREQ(drift_comp);
2436 snprintf(tbuf, sizeof(tbuf), "kernel %.3f PPM", drift_comp * 1e6);
2437 report_event(EVNT_FSET, NULL, tbuf);
2439 snprintf(tbuf, sizeof(tbuf), "ntpd %.3f PPM", drift_comp * 1e6);
2440 report_event(EVNT_FSET, NULL, tbuf);
2442 #else /* KERNEL_PLL */
2443 snprintf(tbuf, sizeof(tbuf), "ntpd %.3f PPM", drift_comp * 1e6);
2444 report_event(EVNT_FSET, NULL, tbuf);
2445 #endif /* KERNEL_PLL */
2454 * This code segment works when clock adjustments are made using
2455 * precision time kernel support and the ntp_adjtime() system
2456 * call. This support is available in Solaris 2.6 and later,
2457 * Digital Unix 4.0 and later, FreeBSD, Linux and specially
2458 * modified kernels for HP-UX 9 and Ultrix 4. In the case of the
2459 * DECstation 5000/240 and Alpha AXP, additional kernel
2460 * modifications provide a true microsecond clock and nanosecond
2461 * clock, respectively.
2463 * Important note: The kernel discipline is used only if the
2464 * step threshold is less than 0.5 s, as anything higher can
2465 * lead to overflow problems. This might occur if some misguided
2466 * lad set the step threshold to something ridiculous.
2468 if (pll_control && kern_enable) {
2470 #define MOD_BITS (MOD_OFFSET | MOD_MAXERROR | MOD_ESTERROR | MOD_STATUS | MOD_TIMECONST)
2473 * We initialize the structure for the ntp_adjtime()
2474 * system call. We have to convert everything to
2475 * microseconds or nanoseconds first. Do not update the
2476 * system variables if the ext_enable flag is set. In
2477 * this case, the external clock driver will update the
2478 * variables, which will be read later by the local
2479 * clock driver. Afterwards, remember the time and
2480 * frequency offsets for jitter and stability values and
2481 * to update the frequency file.
2483 memset(&ntv, 0, sizeof(ntv));
2485 ntv.modes = MOD_STATUS;
2488 ntv.modes = MOD_BITS | MOD_NANO;
2489 #else /* STA_NANO */
2490 ntv.modes = MOD_BITS;
2491 #endif /* STA_NANO */
2492 if (clock_offset < 0)
2497 ntv.offset = (int32)(clock_offset * 1e9 + dtemp);
2498 ntv.constant = sys_poll;
2499 #else /* STA_NANO */
2500 ntv.offset = (int32)(clock_offset * 1e6 + dtemp);
2501 ntv.constant = sys_poll - 4;
2502 #endif /* STA_NANO */
2503 ntv.esterror = (u_int32)(clock_jitter * 1e6);
2504 ntv.maxerror = (u_int32)((sys_rootdelay / 2 + sys_rootdisp) * 1e6);
2505 ntv.status = STA_PLL;
2508 * Enable/disable the PPS if requested.
2511 if (!(pll_status & STA_PPSTIME))
2512 report_event(EVNT_KERN,
2513 NULL, "PPS enabled");
2514 ntv.status |= STA_PPSTIME | STA_PPSFREQ;
2516 if (pll_status & STA_PPSTIME)
2517 report_event(EVNT_KERN,
2518 NULL, "PPS disabled");
2519 ntv.status &= ~(STA_PPSTIME | STA_PPSFREQ);
2521 if (sys_leap == LEAP_ADDSECOND)
2522 ntv.status |= STA_INS;
2523 else if (sys_leap == LEAP_DELSECOND)
2524 ntv.status |= STA_DEL;
2528 * Pass the stuff to the kernel. If it squeals, turn off
2529 * the pps. In any case, fetch the kernel offset,
2530 * frequency and jitter.
2532 if (ntp_adjtime(&ntv) == TIME_ERROR) {
2533 if (!(ntv.status & STA_PPSSIGNAL))
2534 report_event(EVNT_KERN, NULL,
2537 pll_status = ntv.status;
2539 clock_offset = ntv.offset / 1e9;
2540 #else /* STA_NANO */
2541 clock_offset = ntv.offset / 1e6;
2542 #endif /* STA_NANO */
2543 clock_frequency = FREQTOD(ntv.freq);
2546 * If the kernel PPS is lit, monitor its performance.
2548 if (ntv.status & STA_PPSTIME) {
2550 clock_jitter = ntv.jitter / 1e9;
2551 #else /* STA_NANO */
2552 clock_jitter = ntv.jitter / 1e6;
2553 #endif /* STA_NANO */
2556 #if defined(STA_NANO) && NTP_API == 4
2558 * If the TAI changes, update the kernel TAI.
2560 if (loop_tai != sys_tai) {
2562 ntv.modes = MOD_TAI;
2563 ntv.constant = sys_tai;
2566 #endif /* STA_NANO */
2568 #endif /* KERNEL_PLL */