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)"
45 //usage: IF_FEATURE_NTPD_CONF(
46 //usage: "\n If -p is not given, read /etc/ntp.conf"
49 // -l and -p options are not compatible with "standard" ntpd:
50 // it has them as "-l logfile" and "-p pidfile".
51 // -S and -w are not compat either, "standard" ntpd has no such opts.
55 #include <netinet/ip.h> /* For IPTOS_LOWDELAY definition */
56 #include <sys/resource.h> /* setpriority */
57 #include <sys/timex.h>
58 #ifndef IPTOS_LOWDELAY
59 # define IPTOS_LOWDELAY 0x10
63 /* Verbosity control (max level of -dddd options accepted).
64 * max 6 is very talkative (and bloated). 3 is non-bloated,
65 * production level setting.
70 /* High-level description of the algorithm:
72 * We start running with very small poll_exp, BURSTPOLL,
73 * in order to quickly accumulate INITIAL_SAMPLES datapoints
74 * for each peer. Then, time is stepped if the offset is larger
75 * than STEP_THRESHOLD, otherwise it isn't; anyway, we enlarge
76 * poll_exp to MINPOLL and enter frequency measurement step:
77 * we collect new datapoints but ignore them for WATCH_THRESHOLD
78 * seconds. After WATCH_THRESHOLD seconds we look at accumulated
79 * offset and estimate frequency drift.
81 * (frequency measurement step seems to not be strictly needed,
82 * it is conditionally disabled with USING_INITIAL_FREQ_ESTIMATION
85 * After this, we enter "steady state": we collect a datapoint,
86 * we select the best peer, if this datapoint is not a new one
87 * (IOW: if this datapoint isn't for selected peer), sleep
88 * and collect another one; otherwise, use its offset to update
89 * frequency drift, if offset is somewhat large, reduce poll_exp,
90 * otherwise increase poll_exp.
92 * If offset is larger than STEP_THRESHOLD, which shouldn't normally
93 * happen, we assume that something "bad" happened (computer
94 * was hibernated, someone set totally wrong date, etc),
95 * then the time is stepped, all datapoints are discarded,
96 * and we go back to steady state.
98 * Made some changes to speed up re-syncing after our clock goes bad
99 * (tested with suspending my laptop):
100 * - if largish offset (>= STEP_THRESHOLD * 8 == 1 sec) is seen
101 * from a peer, schedule next query for this peer soon
102 * without drastically lowering poll interval for everybody.
103 * This makes us collect enough data for step much faster:
104 * e.g. at poll = 10 (1024 secs), step was done within 5 minutes
105 * after first reply which indicated that our clock is 14 seconds off.
106 * - on step, do not discard d_dispersion data of the existing datapoints,
107 * do not clear reachable_bits. This prevents discarding first ~8
108 * datapoints after the step.
111 #define RETRY_INTERVAL 5 /* on error, retry in N secs */
112 #define RESPONSE_INTERVAL 15 /* wait for reply up to N secs */
113 #define INITIAL_SAMPLES 4 /* how many samples do we want for init */
114 #define BAD_DELAY_GROWTH 4 /* drop packet if its delay grew by more than this */
116 /* Clock discipline parameters and constants */
118 /* Step threshold (sec). std ntpd uses 0.128.
119 * Using exact power of 2 (1/8) results in smaller code */
120 #define STEP_THRESHOLD 0.125
121 #define WATCH_THRESHOLD 128 /* stepout threshold (sec). std ntpd uses 900 (11 mins (!)) */
122 /* NB: set WATCH_THRESHOLD to ~60 when debugging to save time) */
123 //UNUSED: #define PANIC_THRESHOLD 1000 /* panic threshold (sec) */
125 #define FREQ_TOLERANCE 0.000015 /* frequency tolerance (15 PPM) */
126 #define BURSTPOLL 0 /* initial poll */
127 #define MINPOLL 5 /* minimum poll interval. std ntpd uses 6 (6: 64 sec) */
128 /* If we got largish offset from a peer, cap next query interval
129 * for this peer by this many seconds:
131 #define BIGOFF_INTERVAL (1 << 6)
132 /* If offset > discipline_jitter * POLLADJ_GATE, and poll interval is >= 2^BIGPOLL,
133 * then it is decreased _at once_. (If < 2^BIGPOLL, it will be decreased _eventually_).
135 #define BIGPOLL 10 /* 2^10 sec ~= 17 min */
136 #define MAXPOLL 12 /* maximum poll interval (12: 1.1h, 17: 36.4h). std ntpd uses 17 */
137 /* Actively lower poll when we see such big offsets.
138 * With STEP_THRESHOLD = 0.125, it means we try to sync more aggressively
139 * if offset increases over ~0.04 sec */
140 #define POLLDOWN_OFFSET (STEP_THRESHOLD / 3)
141 #define MINDISP 0.01 /* minimum dispersion (sec) */
142 #define MAXDISP 16 /* maximum dispersion (sec) */
143 #define MAXSTRAT 16 /* maximum stratum (infinity metric) */
144 #define MAXDIST 1 /* distance threshold (sec) */
145 #define MIN_SELECTED 1 /* minimum intersection survivors */
146 #define MIN_CLUSTERED 3 /* minimum cluster survivors */
148 #define MAXDRIFT 0.000500 /* frequency drift we can correct (500 PPM) */
150 /* Poll-adjust threshold.
151 * When we see that offset is small enough compared to discipline jitter,
152 * we grow a counter: += MINPOLL. When counter goes over POLLADJ_LIMIT,
153 * we poll_exp++. If offset isn't small, counter -= poll_exp*2,
154 * and when it goes below -POLLADJ_LIMIT, we poll_exp--.
155 * (Bumped from 30 to 40 since otherwise I often see poll_exp going *2* steps down)
157 #define POLLADJ_LIMIT 40
158 /* If offset < discipline_jitter * POLLADJ_GATE, then we decide to increase
159 * poll interval (we think we can't improve timekeeping
160 * by staying at smaller poll).
162 #define POLLADJ_GATE 4
163 #define TIMECONST_HACK_GATE 2
164 /* Compromise Allan intercept (sec). doc uses 1500, std ntpd uses 512 */
168 /* FLL loop gain [why it depends on MAXPOLL??] */
169 #define FLL (MAXPOLL + 1)
170 /* Parameter averaging constant */
179 NTP_MSGSIZE_NOAUTH = 48,
180 NTP_MSGSIZE = (NTP_MSGSIZE_NOAUTH + 4 + NTP_DIGESTSIZE),
183 MODE_MASK = (7 << 0),
184 VERSION_MASK = (7 << 3),
188 /* Leap Second Codes (high order two bits of m_status) */
189 LI_NOWARNING = (0 << 6), /* no warning */
190 LI_PLUSSEC = (1 << 6), /* add a second (61 seconds) */
191 LI_MINUSSEC = (2 << 6), /* minus a second (59 seconds) */
192 LI_ALARM = (3 << 6), /* alarm condition */
195 MODE_RES0 = 0, /* reserved */
196 MODE_SYM_ACT = 1, /* symmetric active */
197 MODE_SYM_PAS = 2, /* symmetric passive */
198 MODE_CLIENT = 3, /* client */
199 MODE_SERVER = 4, /* server */
200 MODE_BROADCAST = 5, /* broadcast */
201 MODE_RES1 = 6, /* reserved for NTP control message */
202 MODE_RES2 = 7, /* reserved for private use */
205 //TODO: better base selection
206 #define OFFSET_1900_1970 2208988800UL /* 1970 - 1900 in seconds */
208 #define NUM_DATAPOINTS 8
221 uint8_t m_status; /* status of local clock and leap info */
223 uint8_t m_ppoll; /* poll value */
224 int8_t m_precision_exp;
225 s_fixedpt_t m_rootdelay;
226 s_fixedpt_t m_rootdisp;
228 l_fixedpt_t m_reftime;
229 l_fixedpt_t m_orgtime;
230 l_fixedpt_t m_rectime;
231 l_fixedpt_t m_xmttime;
233 uint8_t m_digest[NTP_DIGESTSIZE];
243 len_and_sockaddr *p_lsa;
247 uint32_t lastpkt_refid;
248 uint8_t lastpkt_status;
249 uint8_t lastpkt_stratum;
250 uint8_t reachable_bits;
251 /* when to send new query (if p_fd == -1)
252 * or when receive times out (if p_fd >= 0): */
253 double next_action_time;
256 /* p_raw_delay is set even by "high delay" packets */
257 /* lastpkt_delay isn't */
258 double lastpkt_recv_time;
259 double lastpkt_delay;
260 double lastpkt_rootdelay;
261 double lastpkt_rootdisp;
262 /* produced by filter algorithm: */
263 double filter_offset;
264 double filter_dispersion;
265 double filter_jitter;
266 datapoint_t filter_datapoint[NUM_DATAPOINTS];
267 /* last sent packet: */
272 #define USING_KERNEL_PLL_LOOP 1
273 #define USING_INITIAL_FREQ_ESTIMATION 0
280 /* Insert new options above this line. */
281 /* Non-compat options: */
285 OPT_l = (1 << 7) * ENABLE_FEATURE_NTPD_SERVER,
286 /* We hijack some bits for other purposes */
292 /* total round trip delay to currently selected reference clock */
294 /* reference timestamp: time when the system clock was last set or corrected */
296 /* total dispersion to currently selected reference clock */
299 double last_script_run;
302 #if ENABLE_FEATURE_NTPD_SERVER
304 # define G_listen_fd (G.listen_fd)
306 # define G_listen_fd (-1)
310 /* refid: 32-bit code identifying the particular server or reference clock
311 * in stratum 0 packets this is a four-character ASCII string,
312 * called the kiss code, used for debugging and monitoring
313 * in stratum 1 packets this is a four-character ASCII string
314 * assigned to the reference clock by IANA. Example: "GPS "
315 * in stratum 2+ packets, it's IPv4 address or 4 first bytes
316 * of MD5 hash of IPv6
320 /* precision is defined as the larger of the resolution and time to
321 * read the clock, in log2 units. For instance, the precision of a
322 * mains-frequency clock incrementing at 60 Hz is 16 ms, even when the
323 * system clock hardware representation is to the nanosecond.
325 * Delays, jitters of various kinds are clamped down to precision.
327 * If precision_sec is too large, discipline_jitter gets clamped to it
328 * and if offset is smaller than discipline_jitter * POLLADJ_GATE, poll
329 * interval grows even though we really can benefit from staying at
330 * smaller one, collecting non-lagged datapoits and correcting offset.
331 * (Lagged datapoits exist when poll_exp is large but we still have
332 * systematic offset error - the time distance between datapoints
333 * is significant and older datapoints have smaller offsets.
334 * This makes our offset estimation a bit smaller than reality)
335 * Due to this effect, setting G_precision_sec close to
336 * STEP_THRESHOLD isn't such a good idea - offsets may grow
337 * too big and we will step. I observed it with -6.
339 * OTOH, setting precision_sec far too small would result in futile
340 * attempts to syncronize to an unachievable precision.
342 * -6 is 1/64 sec, -7 is 1/128 sec and so on.
343 * -8 is 1/256 ~= 0.003906 (worked well for me --vda)
344 * -9 is 1/512 ~= 0.001953 (let's try this for some time)
346 #define G_precision_exp -9
348 * G_precision_exp is used only for construction outgoing packets.
349 * It's ok to set G_precision_sec to a slightly different value
350 * (One which is "nicer looking" in logs).
351 * Exact value would be (1.0 / (1 << (- G_precision_exp))):
353 #define G_precision_sec 0.002
355 /* Bool. After set to 1, never goes back to 0: */
356 smallint initial_poll_complete;
358 #define STATE_NSET 0 /* initial state, "nothing is set" */
359 //#define STATE_FSET 1 /* frequency set from file */
360 //#define STATE_SPIK 2 /* spike detected */
361 //#define STATE_FREQ 3 /* initial frequency */
362 #define STATE_SYNC 4 /* clock synchronized (normal operation) */
363 uint8_t discipline_state; // doc calls it c.state
364 uint8_t poll_exp; // s.poll
365 int polladj_count; // c.count
366 long kernel_freq_drift;
367 peer_t *last_update_peer;
368 double last_update_offset; // c.last
369 double last_update_recv_time; // s.t
370 double discipline_jitter; // c.jitter
371 /* Since we only compare it with ints, can simplify code
372 * by not making this variable floating point:
374 unsigned offset_to_jitter_ratio;
375 //double cluster_offset; // s.offset
376 //double cluster_jitter; // s.jitter
377 #if !USING_KERNEL_PLL_LOOP
378 double discipline_freq_drift; // c.freq
379 /* Maybe conditionally calculate wander? it's used only for logging */
380 double discipline_wander; // c.wander
383 #define G (*ptr_to_globals)
385 static const int const_IPTOS_LOWDELAY = IPTOS_LOWDELAY;
388 #define VERB1 if (MAX_VERBOSE && G.verbose)
389 #define VERB2 if (MAX_VERBOSE >= 2 && G.verbose >= 2)
390 #define VERB3 if (MAX_VERBOSE >= 3 && G.verbose >= 3)
391 #define VERB4 if (MAX_VERBOSE >= 4 && G.verbose >= 4)
392 #define VERB5 if (MAX_VERBOSE >= 5 && G.verbose >= 5)
393 #define VERB6 if (MAX_VERBOSE >= 6 && G.verbose >= 6)
396 static double LOG2D(int a)
399 return 1.0 / (1UL << -a);
402 static ALWAYS_INLINE double SQUARE(double x)
406 static ALWAYS_INLINE double MAXD(double a, double b)
412 static ALWAYS_INLINE double MIND(double a, double b)
418 static NOINLINE double my_SQRT(double X)
425 double Xhalf = X * 0.5;
427 /* Fast and good approximation to 1/sqrt(X), black magic */
429 /*v.i = 0x5f3759df - (v.i >> 1);*/
430 v.i = 0x5f375a86 - (v.i >> 1); /* - this constant is slightly better */
431 invsqrt = v.f; /* better than 0.2% accuracy */
433 /* Refining it using Newton's method: x1 = x0 - f(x0)/f'(x0)
434 * f(x) = 1/(x*x) - X (f==0 when x = 1/sqrt(X))
436 * f(x)/f'(x) = (X - 1/(x*x)) / (2/(x*x*x)) = X*x*x*x/2 - x/2
437 * x1 = x0 - (X*x0*x0*x0/2 - x0/2) = 1.5*x0 - X*x0*x0*x0/2 = x0*(1.5 - (X/2)*x0*x0)
439 invsqrt = invsqrt * (1.5 - Xhalf * invsqrt * invsqrt); /* ~0.05% accuracy */
440 /* invsqrt = invsqrt * (1.5 - Xhalf * invsqrt * invsqrt); 2nd iter: ~0.0001% accuracy */
441 /* With 4 iterations, more than half results will be exact,
442 * at 6th iterations result stabilizes with about 72% results exact.
443 * We are well satisfied with 0.05% accuracy.
446 return X * invsqrt; /* X * 1/sqrt(X) ~= sqrt(X) */
448 static ALWAYS_INLINE double SQRT(double X)
450 /* If this arch doesn't use IEEE 754 floats, fall back to using libm */
451 if (sizeof(float) != 4)
454 /* This avoids needing libm, saves about 0.5k on x86-32 */
462 gettimeofday(&tv, NULL); /* never fails */
463 G.cur_time = tv.tv_sec + (1.0e-6 * tv.tv_usec) + OFFSET_1900_1970;
468 d_to_tv(double d, struct timeval *tv)
470 tv->tv_sec = (long)d;
471 tv->tv_usec = (d - tv->tv_sec) * 1000000;
475 lfp_to_d(l_fixedpt_t lfp)
478 lfp.int_partl = ntohl(lfp.int_partl);
479 lfp.fractionl = ntohl(lfp.fractionl);
480 ret = (double)lfp.int_partl + ((double)lfp.fractionl / UINT_MAX);
484 sfp_to_d(s_fixedpt_t sfp)
487 sfp.int_parts = ntohs(sfp.int_parts);
488 sfp.fractions = ntohs(sfp.fractions);
489 ret = (double)sfp.int_parts + ((double)sfp.fractions / USHRT_MAX);
492 #if ENABLE_FEATURE_NTPD_SERVER
497 lfp.int_partl = (uint32_t)d;
498 lfp.fractionl = (uint32_t)((d - lfp.int_partl) * UINT_MAX);
499 lfp.int_partl = htonl(lfp.int_partl);
500 lfp.fractionl = htonl(lfp.fractionl);
507 sfp.int_parts = (uint16_t)d;
508 sfp.fractions = (uint16_t)((d - sfp.int_parts) * USHRT_MAX);
509 sfp.int_parts = htons(sfp.int_parts);
510 sfp.fractions = htons(sfp.fractions);
516 dispersion(const datapoint_t *dp)
518 return dp->d_dispersion + FREQ_TOLERANCE * (G.cur_time - dp->d_recv_time);
522 root_distance(peer_t *p)
524 /* The root synchronization distance is the maximum error due to
525 * all causes of the local clock relative to the primary server.
526 * It is defined as half the total delay plus total dispersion
529 return MAXD(MINDISP, p->lastpkt_rootdelay + p->lastpkt_delay) / 2
530 + p->lastpkt_rootdisp
531 + p->filter_dispersion
532 + FREQ_TOLERANCE * (G.cur_time - p->lastpkt_recv_time)
537 set_next(peer_t *p, unsigned t)
539 p->next_action_time = G.cur_time + t;
543 * Peer clock filter and its helpers
546 filter_datapoints(peer_t *p)
553 /* Simulations have shown that use of *averaged* offset for p->filter_offset
554 * is in fact worse than simply using last received one: with large poll intervals
555 * (>= 2048) averaging code uses offset values which are outdated by hours,
556 * and time/frequency correction goes totally wrong when fed essentially bogus offsets.
559 double minoff, maxoff, w;
560 double x = x; /* for compiler */
561 double oldest_off = oldest_off;
562 double oldest_age = oldest_age;
563 double newest_off = newest_off;
564 double newest_age = newest_age;
566 fdp = p->filter_datapoint;
568 minoff = maxoff = fdp[0].d_offset;
569 for (i = 1; i < NUM_DATAPOINTS; i++) {
570 if (minoff > fdp[i].d_offset)
571 minoff = fdp[i].d_offset;
572 if (maxoff < fdp[i].d_offset)
573 maxoff = fdp[i].d_offset;
576 idx = p->datapoint_idx; /* most recent datapoint's index */
578 * Drop two outliers and take weighted average of the rest:
579 * most_recent/2 + older1/4 + older2/8 ... + older5/32 + older6/32
580 * we use older6/32, not older6/64 since sum of weights should be 1:
581 * 1/2 + 1/4 + 1/8 + 1/16 + 1/32 + 1/32 = 1
587 * filter_dispersion = \ -------------
594 for (i = 0; i < NUM_DATAPOINTS; i++) {
596 bb_error_msg("datapoint[%d]: off:%f disp:%f(%f) age:%f%s",
599 fdp[idx].d_dispersion, dispersion(&fdp[idx]),
600 G.cur_time - fdp[idx].d_recv_time,
601 (minoff == fdp[idx].d_offset || maxoff == fdp[idx].d_offset)
602 ? " (outlier by offset)" : ""
606 sum += dispersion(&fdp[idx]) / (2 << i);
608 if (minoff == fdp[idx].d_offset) {
609 minoff -= 1; /* so that we don't match it ever again */
611 if (maxoff == fdp[idx].d_offset) {
614 oldest_off = fdp[idx].d_offset;
615 oldest_age = G.cur_time - fdp[idx].d_recv_time;
618 newest_off = oldest_off;
619 newest_age = oldest_age;
626 idx = (idx - 1) & (NUM_DATAPOINTS - 1);
628 p->filter_dispersion = sum;
629 wavg += x; /* add another older6/64 to form older6/32 */
630 /* Fix systematic underestimation with large poll intervals.
631 * Imagine that we still have a bit of uncorrected drift,
632 * and poll interval is big (say, 100 sec). Offsets form a progression:
633 * 0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 - 0.7 is most recent.
634 * The algorithm above drops 0.0 and 0.7 as outliers,
635 * and then we have this estimation, ~25% off from 0.7:
636 * 0.1/32 + 0.2/32 + 0.3/16 + 0.4/8 + 0.5/4 + 0.6/2 = 0.503125
638 x = oldest_age - newest_age;
640 x = newest_age / x; /* in above example, 100 / (600 - 100) */
641 if (x < 1) { /* paranoia check */
642 x = (newest_off - oldest_off) * x; /* 0.5 * 100/500 = 0.1 */
646 p->filter_offset = wavg;
650 fdp = p->filter_datapoint;
651 idx = p->datapoint_idx; /* most recent datapoint's index */
653 /* filter_offset: simply use the most recent value */
654 p->filter_offset = fdp[idx].d_offset;
658 * filter_dispersion = \ -------------
665 for (i = 0; i < NUM_DATAPOINTS; i++) {
666 sum += dispersion(&fdp[idx]) / (2 << i);
667 wavg += fdp[idx].d_offset;
668 idx = (idx - 1) & (NUM_DATAPOINTS - 1);
670 wavg /= NUM_DATAPOINTS;
671 p->filter_dispersion = sum;
674 /* +----- -----+ ^ 1/2
678 * filter_jitter = | --- * / (avg-offset_j) |
682 * where n is the number of valid datapoints in the filter (n > 1);
683 * if filter_jitter < precision then filter_jitter = precision
686 for (i = 0; i < NUM_DATAPOINTS; i++) {
687 sum += SQUARE(wavg - fdp[i].d_offset);
689 sum = SQRT(sum / NUM_DATAPOINTS);
690 p->filter_jitter = sum > G_precision_sec ? sum : G_precision_sec;
692 VERB4 bb_error_msg("filter offset:%+f disp:%f jitter:%f",
694 p->filter_dispersion,
699 reset_peer_stats(peer_t *p, double offset)
702 bool small_ofs = fabs(offset) < 16 * STEP_THRESHOLD;
704 /* Used to set p->filter_datapoint[i].d_dispersion = MAXDISP
705 * and clear reachable bits, but this proved to be too agressive:
706 * after step (tested with suspinding laptop for ~30 secs),
707 * this caused all previous data to be considered invalid,
708 * making us needing to collect full ~8 datapoins per peer
709 * after step in order to start trusting them.
710 * In turn, this was making poll interval decrease even after
711 * step was done. (Poll interval decreases already before step
712 * in this scenario, because we see large offsets and end up with
713 * no good peer to select).
716 for (i = 0; i < NUM_DATAPOINTS; i++) {
718 p->filter_datapoint[i].d_recv_time += offset;
719 if (p->filter_datapoint[i].d_offset != 0) {
720 p->filter_datapoint[i].d_offset -= offset;
721 //bb_error_msg("p->filter_datapoint[%d].d_offset %f -> %f",
723 // p->filter_datapoint[i].d_offset + offset,
724 // p->filter_datapoint[i].d_offset);
727 p->filter_datapoint[i].d_recv_time = G.cur_time;
728 p->filter_datapoint[i].d_offset = 0;
729 /*p->filter_datapoint[i].d_dispersion = MAXDISP;*/
733 p->lastpkt_recv_time += offset;
735 /*p->reachable_bits = 0;*/
736 p->lastpkt_recv_time = G.cur_time;
738 filter_datapoints(p); /* recalc p->filter_xxx */
739 VERB6 bb_error_msg("%s->lastpkt_recv_time=%f", p->p_dotted, p->lastpkt_recv_time);
743 add_peers(const char *s)
747 p = xzalloc(sizeof(*p));
748 p->p_lsa = xhost2sockaddr(s, 123);
749 p->p_dotted = xmalloc_sockaddr2dotted_noport(&p->p_lsa->u.sa);
751 p->p_xmt_msg.m_status = MODE_CLIENT | (NTP_VERSION << 3);
752 p->next_action_time = G.cur_time; /* = set_next(p, 0); */
753 reset_peer_stats(p, 16 * STEP_THRESHOLD);
755 llist_add_to(&G.ntp_peers, p);
761 const struct sockaddr *from, const struct sockaddr *to, socklen_t addrlen,
762 msg_t *msg, ssize_t len)
768 ret = sendto(fd, msg, len, MSG_DONTWAIT, to, addrlen);
770 ret = send_to_from(fd, msg, len, MSG_DONTWAIT, to, from, addrlen);
773 bb_perror_msg("send failed");
780 send_query_to_peer(peer_t *p)
782 /* Why do we need to bind()?
783 * See what happens when we don't bind:
785 * socket(PF_INET, SOCK_DGRAM, IPPROTO_IP) = 3
786 * setsockopt(3, SOL_IP, IP_TOS, [16], 4) = 0
787 * gettimeofday({1259071266, 327885}, NULL) = 0
788 * sendto(3, "xxx", 48, MSG_DONTWAIT, {sa_family=AF_INET, sin_port=htons(123), sin_addr=inet_addr("10.34.32.125")}, 16) = 48
789 * ^^^ we sent it from some source port picked by kernel.
790 * time(NULL) = 1259071266
791 * write(2, "ntpd: entering poll 15 secs\n", 28) = 28
792 * poll([{fd=3, events=POLLIN}], 1, 15000) = 1 ([{fd=3, revents=POLLIN}])
793 * recv(3, "yyy", 68, MSG_DONTWAIT) = 48
794 * ^^^ this recv will receive packets to any local port!
796 * Uncomment this and use strace to see it in action:
798 #define PROBE_LOCAL_ADDR /* { len_and_sockaddr lsa; lsa.len = LSA_SIZEOF_SA; getsockname(p->query.fd, &lsa.u.sa, &lsa.len); } */
802 len_and_sockaddr *local_lsa;
804 family = p->p_lsa->u.sa.sa_family;
805 p->p_fd = fd = xsocket_type(&local_lsa, family, SOCK_DGRAM);
806 /* local_lsa has "null" address and port 0 now.
807 * bind() ensures we have a *particular port* selected by kernel
808 * and remembered in p->p_fd, thus later recv(p->p_fd)
809 * receives only packets sent to this port.
812 xbind(fd, &local_lsa->u.sa, local_lsa->len);
814 #if ENABLE_FEATURE_IPV6
815 if (family == AF_INET)
817 setsockopt(fd, IPPROTO_IP, IP_TOS, &const_IPTOS_LOWDELAY, sizeof(const_IPTOS_LOWDELAY));
821 /* Emit message _before_ attempted send. Think of a very short
822 * roundtrip networks: we need to go back to recv loop ASAP,
823 * to reduce delay. Printing messages after send works against that.
825 VERB1 bb_error_msg("sending query to %s", p->p_dotted);
828 * Send out a random 64-bit number as our transmit time. The NTP
829 * server will copy said number into the originate field on the
830 * response that it sends us. This is totally legal per the SNTP spec.
832 * The impact of this is two fold: we no longer send out the current
833 * system time for the world to see (which may aid an attacker), and
834 * it gives us a (not very secure) way of knowing that we're not
835 * getting spoofed by an attacker that can't capture our traffic
836 * but can spoof packets from the NTP server we're communicating with.
838 * Save the real transmit timestamp locally.
840 p->p_xmt_msg.m_xmttime.int_partl = rand();
841 p->p_xmt_msg.m_xmttime.fractionl = rand();
842 p->p_xmttime = gettime1900d();
844 /* Were doing it only if sendto worked, but
845 * loss of sync detection needs reachable_bits updated
846 * even if sending fails *locally*:
847 * "network is unreachable" because cable was pulled?
848 * We still need to declare "unsync" if this condition persists.
850 p->reachable_bits <<= 1;
852 if (do_sendto(p->p_fd, /*from:*/ NULL, /*to:*/ &p->p_lsa->u.sa, /*addrlen:*/ p->p_lsa->len,
853 &p->p_xmt_msg, NTP_MSGSIZE_NOAUTH) == -1
858 * We know that we sent nothing.
859 * We can retry *soon* without fearing
860 * that we are flooding the peer.
862 set_next(p, RETRY_INTERVAL);
866 set_next(p, RESPONSE_INTERVAL);
870 /* Note that there is no provision to prevent several run_scripts
871 * to be started in quick succession. In fact, it happens rather often
872 * if initial syncronization results in a step.
873 * You will see "step" and then "stratum" script runs, sometimes
874 * as close as only 0.002 seconds apart.
875 * Script should be ready to deal with this.
877 static void run_script(const char *action, double offset)
880 char *env1, *env2, *env3, *env4;
882 G.last_script_run = G.cur_time;
887 argv[0] = (char*) G.script_name;
888 argv[1] = (char*) action;
891 VERB1 bb_error_msg("executing '%s %s'", G.script_name, action);
893 env1 = xasprintf("%s=%u", "stratum", G.stratum);
895 env2 = xasprintf("%s=%ld", "freq_drift_ppm", G.kernel_freq_drift);
897 env3 = xasprintf("%s=%u", "poll_interval", 1 << G.poll_exp);
899 env4 = xasprintf("%s=%f", "offset", offset);
901 /* Other items of potential interest: selected peer,
902 * rootdelay, reftime, rootdisp, refid, ntp_status,
903 * last_update_offset, last_update_recv_time, discipline_jitter,
904 * how many peers have reachable_bits = 0?
907 /* Don't want to wait: it may run hwclock --systohc, and that
908 * may take some time (seconds): */
909 /*spawn_and_wait(argv);*/
913 unsetenv("freq_drift_ppm");
914 unsetenv("poll_interval");
923 step_time(double offset)
927 struct timeval tvc, tvn;
928 char buf[sizeof("yyyy-mm-dd hh:mm:ss") + /*paranoia:*/ 4];
931 gettimeofday(&tvc, NULL); /* never fails */
932 dtime = tvc.tv_sec + (1.0e-6 * tvc.tv_usec) + offset;
933 d_to_tv(dtime, &tvn);
934 if (settimeofday(&tvn, NULL) == -1)
935 bb_perror_msg_and_die("settimeofday");
939 strftime_YYYYMMDDHHMMSS(buf, sizeof(buf), &tval);
940 bb_error_msg("current time is %s.%06u", buf, (unsigned)tvc.tv_usec);
943 strftime_YYYYMMDDHHMMSS(buf, sizeof(buf), &tval);
944 bb_error_msg("setting time to %s.%06u (offset %+fs)", buf, (unsigned)tvn.tv_usec, offset);
946 /* Correct various fields which contain time-relative values: */
949 G.cur_time += offset;
950 G.last_update_recv_time += offset;
951 G.last_script_run += offset;
953 /* p->lastpkt_recv_time, p->next_action_time and such: */
954 for (item = G.ntp_peers; item != NULL; item = item->link) {
955 peer_t *pp = (peer_t *) item->data;
956 reset_peer_stats(pp, offset);
957 //bb_error_msg("offset:%+f pp->next_action_time:%f -> %f",
958 // offset, pp->next_action_time, pp->next_action_time + offset);
959 pp->next_action_time += offset;
961 /* We wait for reply from this peer too.
962 * But due to step we are doing, reply's data is no longer
963 * useful (in fact, it'll be bogus). Stop waiting for it.
967 set_next(pp, RETRY_INTERVAL);
974 * Selection and clustering, and their helpers
980 double opt_rd; /* optimization */
983 compare_point_edge(const void *aa, const void *bb)
985 const point_t *a = aa;
986 const point_t *b = bb;
987 if (a->edge < b->edge) {
990 return (a->edge > b->edge);
997 compare_survivor_metric(const void *aa, const void *bb)
999 const survivor_t *a = aa;
1000 const survivor_t *b = bb;
1001 if (a->metric < b->metric) {
1004 return (a->metric > b->metric);
1007 fit(peer_t *p, double rd)
1009 if ((p->reachable_bits & (p->reachable_bits-1)) == 0) {
1010 /* One or zero bits in reachable_bits */
1011 VERB4 bb_error_msg("peer %s unfit for selection: unreachable", p->p_dotted);
1014 #if 0 /* we filter out such packets earlier */
1015 if ((p->lastpkt_status & LI_ALARM) == LI_ALARM
1016 || p->lastpkt_stratum >= MAXSTRAT
1018 VERB4 bb_error_msg("peer %s unfit for selection: bad status/stratum", p->p_dotted);
1022 /* rd is root_distance(p) */
1023 if (rd > MAXDIST + FREQ_TOLERANCE * (1 << G.poll_exp)) {
1024 VERB4 bb_error_msg("peer %s unfit for selection: root distance too high", p->p_dotted);
1028 // /* Do we have a loop? */
1029 // if (p->refid == p->dstaddr || p->refid == s.refid)
1034 select_and_cluster(void)
1039 int size = 3 * G.peer_cnt;
1040 /* for selection algorithm */
1041 point_t point[size];
1042 unsigned num_points, num_candidates;
1044 unsigned num_falsetickers;
1045 /* for cluster algorithm */
1046 survivor_t survivor[size];
1047 unsigned num_survivors;
1053 if (G.initial_poll_complete) while (item != NULL) {
1056 p = (peer_t *) item->data;
1057 rd = root_distance(p);
1058 offset = p->filter_offset;
1064 VERB5 bb_error_msg("interval: [%f %f %f] %s",
1070 point[num_points].p = p;
1071 point[num_points].type = -1;
1072 point[num_points].edge = offset - rd;
1073 point[num_points].opt_rd = rd;
1075 point[num_points].p = p;
1076 point[num_points].type = 0;
1077 point[num_points].edge = offset;
1078 point[num_points].opt_rd = rd;
1080 point[num_points].p = p;
1081 point[num_points].type = 1;
1082 point[num_points].edge = offset + rd;
1083 point[num_points].opt_rd = rd;
1087 num_candidates = num_points / 3;
1088 if (num_candidates == 0) {
1089 VERB3 bb_error_msg("no valid datapoints%s", ", no peer selected");
1092 //TODO: sorting does not seem to be done in reference code
1093 qsort(point, num_points, sizeof(point[0]), compare_point_edge);
1095 /* Start with the assumption that there are no falsetickers.
1096 * Attempt to find a nonempty intersection interval containing
1097 * the midpoints of all truechimers.
1098 * If a nonempty interval cannot be found, increase the number
1099 * of assumed falsetickers by one and try again.
1100 * If a nonempty interval is found and the number of falsetickers
1101 * is less than the number of truechimers, a majority has been found
1102 * and the midpoint of each truechimer represents
1103 * the candidates available to the cluster algorithm.
1105 num_falsetickers = 0;
1108 unsigned num_midpoints = 0;
1113 for (i = 0; i < num_points; i++) {
1115 * if (point[i].type == -1) c++;
1116 * if (point[i].type == 1) c--;
1117 * and it's simpler to do it this way:
1120 if (c >= num_candidates - num_falsetickers) {
1121 /* If it was c++ and it got big enough... */
1122 low = point[i].edge;
1125 if (point[i].type == 0)
1129 for (i = num_points-1; i >= 0; i--) {
1131 if (c >= num_candidates - num_falsetickers) {
1132 high = point[i].edge;
1135 if (point[i].type == 0)
1138 /* If the number of midpoints is greater than the number
1139 * of allowed falsetickers, the intersection contains at
1140 * least one truechimer with no midpoint - bad.
1141 * Also, interval should be nonempty.
1143 if (num_midpoints <= num_falsetickers && low < high)
1146 if (num_falsetickers * 2 >= num_candidates) {
1147 VERB3 bb_error_msg("falsetickers:%d, candidates:%d%s",
1148 num_falsetickers, num_candidates,
1149 ", no peer selected");
1153 VERB4 bb_error_msg("selected interval: [%f, %f]; candidates:%d falsetickers:%d",
1154 low, high, num_candidates, num_falsetickers);
1158 /* Construct a list of survivors (p, metric)
1159 * from the chime list, where metric is dominated
1160 * first by stratum and then by root distance.
1161 * All other things being equal, this is the order of preference.
1164 for (i = 0; i < num_points; i++) {
1165 if (point[i].edge < low || point[i].edge > high)
1168 survivor[num_survivors].p = p;
1169 /* x.opt_rd == root_distance(p); */
1170 survivor[num_survivors].metric = MAXDIST * p->lastpkt_stratum + point[i].opt_rd;
1171 VERB5 bb_error_msg("survivor[%d] metric:%f peer:%s",
1172 num_survivors, survivor[num_survivors].metric, p->p_dotted);
1175 /* There must be at least MIN_SELECTED survivors to satisfy the
1176 * correctness assertions. Ordinarily, the Byzantine criteria
1177 * require four survivors, but for the demonstration here, one
1180 if (num_survivors < MIN_SELECTED) {
1181 VERB3 bb_error_msg("survivors:%d%s",
1183 ", no peer selected");
1187 //looks like this is ONLY used by the fact that later we pick survivor[0].
1188 //we can avoid sorting then, just find the minimum once!
1189 qsort(survivor, num_survivors, sizeof(survivor[0]), compare_survivor_metric);
1191 /* For each association p in turn, calculate the selection
1192 * jitter p->sjitter as the square root of the sum of squares
1193 * (p->offset - q->offset) over all q associations. The idea is
1194 * to repeatedly discard the survivor with maximum selection
1195 * jitter until a termination condition is met.
1198 unsigned max_idx = max_idx;
1199 double max_selection_jitter = max_selection_jitter;
1200 double min_jitter = min_jitter;
1202 if (num_survivors <= MIN_CLUSTERED) {
1203 VERB4 bb_error_msg("num_survivors %d <= %d, not discarding more",
1204 num_survivors, MIN_CLUSTERED);
1208 /* To make sure a few survivors are left
1209 * for the clustering algorithm to chew on,
1210 * we stop if the number of survivors
1211 * is less than or equal to MIN_CLUSTERED (3).
1213 for (i = 0; i < num_survivors; i++) {
1214 double selection_jitter_sq;
1217 if (i == 0 || p->filter_jitter < min_jitter)
1218 min_jitter = p->filter_jitter;
1220 selection_jitter_sq = 0;
1221 for (j = 0; j < num_survivors; j++) {
1222 peer_t *q = survivor[j].p;
1223 selection_jitter_sq += SQUARE(p->filter_offset - q->filter_offset);
1225 if (i == 0 || selection_jitter_sq > max_selection_jitter) {
1226 max_selection_jitter = selection_jitter_sq;
1229 VERB6 bb_error_msg("survivor %d selection_jitter^2:%f",
1230 i, selection_jitter_sq);
1232 max_selection_jitter = SQRT(max_selection_jitter / num_survivors);
1233 VERB5 bb_error_msg("max_selection_jitter (at %d):%f min_jitter:%f",
1234 max_idx, max_selection_jitter, min_jitter);
1236 /* If the maximum selection jitter is less than the
1237 * minimum peer jitter, then tossing out more survivors
1238 * will not lower the minimum peer jitter, so we might
1241 if (max_selection_jitter < min_jitter) {
1242 VERB4 bb_error_msg("max_selection_jitter:%f < min_jitter:%f, num_survivors:%d, not discarding more",
1243 max_selection_jitter, min_jitter, num_survivors);
1247 /* Delete survivor[max_idx] from the list
1248 * and go around again.
1250 VERB6 bb_error_msg("dropping survivor %d", max_idx);
1252 while (max_idx < num_survivors) {
1253 survivor[max_idx] = survivor[max_idx + 1];
1259 /* Combine the offsets of the clustering algorithm survivors
1260 * using a weighted average with weight determined by the root
1261 * distance. Compute the selection jitter as the weighted RMS
1262 * difference between the first survivor and the remaining
1263 * survivors. In some cases the inherent clock jitter can be
1264 * reduced by not using this algorithm, especially when frequent
1265 * clockhopping is involved. bbox: thus we don't do it.
1269 for (i = 0; i < num_survivors; i++) {
1271 x = root_distance(p);
1273 z += p->filter_offset / x;
1274 w += SQUARE(p->filter_offset - survivor[0].p->filter_offset) / x;
1276 //G.cluster_offset = z / y;
1277 //G.cluster_jitter = SQRT(w / y);
1280 /* Pick the best clock. If the old system peer is on the list
1281 * and at the same stratum as the first survivor on the list,
1282 * then don't do a clock hop. Otherwise, select the first
1283 * survivor on the list as the new system peer.
1286 if (G.last_update_peer
1287 && G.last_update_peer->lastpkt_stratum <= p->lastpkt_stratum
1289 /* Starting from 1 is ok here */
1290 for (i = 1; i < num_survivors; i++) {
1291 if (G.last_update_peer == survivor[i].p) {
1292 VERB5 bb_error_msg("keeping old synced peer");
1293 p = G.last_update_peer;
1298 G.last_update_peer = p;
1300 VERB4 bb_error_msg("selected peer %s filter_offset:%+f age:%f",
1303 G.cur_time - p->lastpkt_recv_time
1310 * Local clock discipline and its helpers
1313 set_new_values(int disc_state, double offset, double recv_time)
1315 /* Enter new state and set state variables. Note we use the time
1316 * of the last clock filter sample, which must be earlier than
1319 VERB4 bb_error_msg("disc_state=%d last update offset=%f recv_time=%f",
1320 disc_state, offset, recv_time);
1321 G.discipline_state = disc_state;
1322 G.last_update_offset = offset;
1323 G.last_update_recv_time = recv_time;
1325 /* Return: -1: decrease poll interval, 0: leave as is, 1: increase */
1327 update_local_clock(peer_t *p)
1331 /* Note: can use G.cluster_offset instead: */
1332 double offset = p->filter_offset;
1333 double recv_time = p->lastpkt_recv_time;
1335 #if !USING_KERNEL_PLL_LOOP
1338 #if !USING_KERNEL_PLL_LOOP || USING_INITIAL_FREQ_ESTIMATION
1339 double since_last_update;
1341 double etemp, dtemp;
1343 abs_offset = fabs(offset);
1346 /* If needed, -S script can do it by looking at $offset
1347 * env var and killing parent */
1348 /* If the offset is too large, give up and go home */
1349 if (abs_offset > PANIC_THRESHOLD) {
1350 bb_error_msg_and_die("offset %f far too big, exiting", offset);
1354 /* If this is an old update, for instance as the result
1355 * of a system peer change, avoid it. We never use
1356 * an old sample or the same sample twice.
1358 if (recv_time <= G.last_update_recv_time) {
1359 VERB3 bb_error_msg("update from %s: same or older datapoint, not using it",
1361 return 0; /* "leave poll interval as is" */
1364 /* Clock state machine transition function. This is where the
1365 * action is and defines how the system reacts to large time
1366 * and frequency errors.
1368 #if !USING_KERNEL_PLL_LOOP || USING_INITIAL_FREQ_ESTIMATION
1369 since_last_update = recv_time - G.reftime;
1371 #if !USING_KERNEL_PLL_LOOP
1374 #if USING_INITIAL_FREQ_ESTIMATION
1375 if (G.discipline_state == STATE_FREQ) {
1376 /* Ignore updates until the stepout threshold */
1377 if (since_last_update < WATCH_THRESHOLD) {
1378 VERB4 bb_error_msg("measuring drift, datapoint ignored, %f sec remains",
1379 WATCH_THRESHOLD - since_last_update);
1380 return 0; /* "leave poll interval as is" */
1382 # if !USING_KERNEL_PLL_LOOP
1383 freq_drift = (offset - G.last_update_offset) / since_last_update;
1388 /* There are two main regimes: when the
1389 * offset exceeds the step threshold and when it does not.
1391 if (abs_offset > STEP_THRESHOLD) {
1395 // This "spike state" seems to be useless, peer selection already drops
1396 // occassional "bad" datapoints. If we are here, there were _many_
1397 // large offsets. When a few first large offsets are seen,
1398 // we end up in "no valid datapoints, no peer selected" state.
1399 // Only when enough of them are seen (which means it's not a fluke),
1400 // we end up here. Looks like _our_ clock is off.
1401 switch (G.discipline_state) {
1403 /* The first outlyer: ignore it, switch to SPIK state */
1404 VERB3 bb_error_msg("update from %s: offset:%+f, spike%s",
1405 p->p_dotted, offset,
1407 G.discipline_state = STATE_SPIK;
1408 return -1; /* "decrease poll interval" */
1411 /* Ignore succeeding outlyers until either an inlyer
1412 * is found or the stepout threshold is exceeded.
1414 remains = WATCH_THRESHOLD - since_last_update;
1416 VERB3 bb_error_msg("update from %s: offset:%+f, spike%s",
1417 p->p_dotted, offset,
1418 ", datapoint ignored");
1419 return -1; /* "decrease poll interval" */
1421 /* fall through: we need to step */
1425 /* Step the time and clamp down the poll interval.
1427 * In NSET state an initial frequency correction is
1428 * not available, usually because the frequency file has
1429 * not yet been written. Since the time is outside the
1430 * capture range, the clock is stepped. The frequency
1431 * will be set directly following the stepout interval.
1433 * In FSET state the initial frequency has been set
1434 * from the frequency file. Since the time is outside
1435 * the capture range, the clock is stepped immediately,
1436 * rather than after the stepout interval. Guys get
1437 * nervous if it takes 17 minutes to set the clock for
1440 * In SPIK state the stepout threshold has expired and
1441 * the phase is still above the step threshold. Note
1442 * that a single spike greater than the step threshold
1443 * is always suppressed, even at the longer poll
1446 VERB4 bb_error_msg("stepping time by %+f; poll_exp=MINPOLL", offset);
1448 if (option_mask32 & OPT_q) {
1449 /* We were only asked to set time once. Done. */
1453 G.polladj_count = 0;
1454 G.poll_exp = MINPOLL;
1455 G.stratum = MAXSTRAT;
1457 run_script("step", offset);
1459 recv_time += offset;
1461 #if USING_INITIAL_FREQ_ESTIMATION
1462 if (G.discipline_state == STATE_NSET) {
1463 set_new_values(STATE_FREQ, /*offset:*/ 0, recv_time);
1464 return 1; /* "ok to increase poll interval" */
1467 abs_offset = offset = 0;
1468 set_new_values(STATE_SYNC, offset, recv_time);
1470 } else { /* abs_offset <= STEP_THRESHOLD */
1472 if (G.poll_exp < MINPOLL && G.initial_poll_complete) {
1473 VERB4 bb_error_msg("small offset:%+f, disabling burst mode", offset);
1474 G.polladj_count = 0;
1475 G.poll_exp = MINPOLL;
1478 /* Compute the clock jitter as the RMS of exponentially
1479 * weighted offset differences. Used by the poll adjust code.
1481 etemp = SQUARE(G.discipline_jitter);
1482 dtemp = SQUARE(offset - G.last_update_offset);
1483 G.discipline_jitter = SQRT(etemp + (dtemp - etemp) / AVG);
1485 switch (G.discipline_state) {
1487 if (option_mask32 & OPT_q) {
1488 /* We were only asked to set time once.
1489 * The clock is precise enough, no need to step.
1493 #if USING_INITIAL_FREQ_ESTIMATION
1494 /* This is the first update received and the frequency
1495 * has not been initialized. The first thing to do
1496 * is directly measure the oscillator frequency.
1498 set_new_values(STATE_FREQ, offset, recv_time);
1500 set_new_values(STATE_SYNC, offset, recv_time);
1502 VERB4 bb_error_msg("transitioning to FREQ, datapoint ignored");
1503 return 0; /* "leave poll interval as is" */
1505 #if 0 /* this is dead code for now */
1507 /* This is the first update and the frequency
1508 * has been initialized. Adjust the phase, but
1509 * don't adjust the frequency until the next update.
1511 set_new_values(STATE_SYNC, offset, recv_time);
1512 /* freq_drift remains 0 */
1516 #if USING_INITIAL_FREQ_ESTIMATION
1518 /* since_last_update >= WATCH_THRESHOLD, we waited enough.
1519 * Correct the phase and frequency and switch to SYNC state.
1520 * freq_drift was already estimated (see code above)
1522 set_new_values(STATE_SYNC, offset, recv_time);
1527 #if !USING_KERNEL_PLL_LOOP
1528 /* Compute freq_drift due to PLL and FLL contributions.
1530 * The FLL and PLL frequency gain constants
1531 * depend on the poll interval and Allan
1532 * intercept. The FLL is not used below one-half
1533 * the Allan intercept. Above that the loop gain
1534 * increases in steps to 1 / AVG.
1536 if ((1 << G.poll_exp) > ALLAN / 2) {
1537 etemp = FLL - G.poll_exp;
1540 freq_drift += (offset - G.last_update_offset) / (MAXD(since_last_update, ALLAN) * etemp);
1542 /* For the PLL the integration interval
1543 * (numerator) is the minimum of the update
1544 * interval and poll interval. This allows
1545 * oversampling, but not undersampling.
1547 etemp = MIND(since_last_update, (1 << G.poll_exp));
1548 dtemp = (4 * PLL) << G.poll_exp;
1549 freq_drift += offset * etemp / SQUARE(dtemp);
1551 set_new_values(STATE_SYNC, offset, recv_time);
1554 if (G.stratum != p->lastpkt_stratum + 1) {
1555 G.stratum = p->lastpkt_stratum + 1;
1556 run_script("stratum", offset);
1560 if (G.discipline_jitter < G_precision_sec)
1561 G.discipline_jitter = G_precision_sec;
1562 G.offset_to_jitter_ratio = abs_offset / G.discipline_jitter;
1564 G.reftime = G.cur_time;
1565 G.ntp_status = p->lastpkt_status;
1566 G.refid = p->lastpkt_refid;
1567 G.rootdelay = p->lastpkt_rootdelay + p->lastpkt_delay;
1568 dtemp = p->filter_jitter; // SQRT(SQUARE(p->filter_jitter) + SQUARE(G.cluster_jitter));
1569 dtemp += MAXD(p->filter_dispersion + FREQ_TOLERANCE * (G.cur_time - p->lastpkt_recv_time) + abs_offset, MINDISP);
1570 G.rootdisp = p->lastpkt_rootdisp + dtemp;
1571 VERB4 bb_error_msg("updating leap/refid/reftime/rootdisp from peer %s", p->p_dotted);
1573 /* We are in STATE_SYNC now, but did not do adjtimex yet.
1574 * (Any other state does not reach this, they all return earlier)
1575 * By this time, freq_drift and offset are set
1576 * to values suitable for adjtimex.
1578 #if !USING_KERNEL_PLL_LOOP
1579 /* Calculate the new frequency drift and frequency stability (wander).
1580 * Compute the clock wander as the RMS of exponentially weighted
1581 * frequency differences. This is not used directly, but can,
1582 * along with the jitter, be a highly useful monitoring and
1585 dtemp = G.discipline_freq_drift + freq_drift;
1586 G.discipline_freq_drift = MAXD(MIND(MAXDRIFT, dtemp), -MAXDRIFT);
1587 etemp = SQUARE(G.discipline_wander);
1588 dtemp = SQUARE(dtemp);
1589 G.discipline_wander = SQRT(etemp + (dtemp - etemp) / AVG);
1591 VERB4 bb_error_msg("discipline freq_drift=%.9f(int:%ld corr:%e) wander=%f",
1592 G.discipline_freq_drift,
1593 (long)(G.discipline_freq_drift * 65536e6),
1595 G.discipline_wander);
1598 memset(&tmx, 0, sizeof(tmx));
1599 if (adjtimex(&tmx) < 0)
1600 bb_perror_msg_and_die("adjtimex");
1601 bb_error_msg("p adjtimex freq:%ld offset:%+ld status:0x%x tc:%ld",
1602 tmx.freq, tmx.offset, tmx.status, tmx.constant);
1605 memset(&tmx, 0, sizeof(tmx));
1607 //doesn't work, offset remains 0 (!) in kernel:
1608 //ntpd: set adjtimex freq:1786097 tmx.offset:77487
1609 //ntpd: prev adjtimex freq:1786097 tmx.offset:0
1610 //ntpd: cur adjtimex freq:1786097 tmx.offset:0
1611 tmx.modes = ADJ_FREQUENCY | ADJ_OFFSET;
1612 /* 65536 is one ppm */
1613 tmx.freq = G.discipline_freq_drift * 65536e6;
1615 tmx.modes = ADJ_OFFSET | ADJ_STATUS | ADJ_TIMECONST;// | ADJ_MAXERROR | ADJ_ESTERROR;
1616 tmx.offset = (offset * 1000000); /* usec */
1617 tmx.status = STA_PLL;
1618 if (G.ntp_status & LI_PLUSSEC)
1619 tmx.status |= STA_INS;
1620 if (G.ntp_status & LI_MINUSSEC)
1621 tmx.status |= STA_DEL;
1623 tmx.constant = G.poll_exp - 4;
1625 * The below if statement should be unnecessary, but...
1626 * It looks like Linux kernel's PLL is far too gentle in changing
1627 * tmx.freq in response to clock offset. Offset keeps growing
1628 * and eventually we fall back to smaller poll intervals.
1629 * We can make correction more agressive (about x2) by supplying
1630 * PLL time constant which is one less than the real one.
1631 * To be on a safe side, let's do it only if offset is significantly
1632 * larger than jitter.
1634 if (tmx.constant > 0 && G.offset_to_jitter_ratio >= TIMECONST_HACK_GATE)
1637 //tmx.esterror = (uint32_t)(clock_jitter * 1e6);
1638 //tmx.maxerror = (uint32_t)((sys_rootdelay / 2 + sys_rootdisp) * 1e6);
1639 rc = adjtimex(&tmx);
1641 bb_perror_msg_and_die("adjtimex");
1642 /* NB: here kernel returns constant == G.poll_exp, not == G.poll_exp - 4.
1643 * Not sure why. Perhaps it is normal.
1645 VERB4 bb_error_msg("adjtimex:%d freq:%ld offset:%+ld status:0x%x",
1646 rc, tmx.freq, tmx.offset, tmx.status);
1647 G.kernel_freq_drift = tmx.freq / 65536;
1648 VERB2 bb_error_msg("update from:%s offset:%+f jitter:%f clock drift:%+.3fppm tc:%d",
1649 p->p_dotted, offset, G.discipline_jitter, (double)tmx.freq / 65536, (int)tmx.constant);
1651 return 1; /* "ok to increase poll interval" */
1656 * We've got a new reply packet from a peer, process it
1660 retry_interval(void)
1662 /* Local problem, want to retry soon */
1663 unsigned interval, r;
1664 interval = RETRY_INTERVAL;
1666 interval += r % (unsigned)(RETRY_INTERVAL / 4);
1667 VERB4 bb_error_msg("chose retry interval:%u", interval);
1671 poll_interval(int exponent)
1673 unsigned interval, r, mask;
1674 exponent = G.poll_exp + exponent;
1677 interval = 1 << exponent;
1678 mask = ((interval-1) >> 4) | 1;
1680 interval += r & mask; /* ~ random(0..1) * interval/16 */
1681 VERB4 bb_error_msg("chose poll interval:%u (poll_exp:%d exp:%d)", interval, G.poll_exp, exponent);
1684 static NOINLINE void
1685 recv_and_process_peer_pkt(peer_t *p)
1690 double T1, T2, T3, T4;
1692 double prev_delay, delay;
1694 datapoint_t *datapoint;
1699 /* We can recvfrom here and check from.IP, but some multihomed
1700 * ntp servers reply from their *other IP*.
1701 * TODO: maybe we should check at least what we can: from.port == 123?
1703 size = recv(p->p_fd, &msg, sizeof(msg), MSG_DONTWAIT);
1705 bb_perror_msg("recv(%s) error", p->p_dotted);
1706 if (errno == EHOSTUNREACH || errno == EHOSTDOWN
1707 || errno == ENETUNREACH || errno == ENETDOWN
1708 || errno == ECONNREFUSED || errno == EADDRNOTAVAIL
1711 //TODO: always do this?
1712 interval = retry_interval();
1713 goto set_next_and_ret;
1718 if (size != NTP_MSGSIZE_NOAUTH && size != NTP_MSGSIZE) {
1719 bb_error_msg("malformed packet received from %s", p->p_dotted);
1723 if (msg.m_orgtime.int_partl != p->p_xmt_msg.m_xmttime.int_partl
1724 || msg.m_orgtime.fractionl != p->p_xmt_msg.m_xmttime.fractionl
1726 /* Somebody else's packet */
1730 /* We do not expect any more packets from this peer for now.
1731 * Closing the socket informs kernel about it.
1732 * We open a new socket when we send a new query.
1737 if ((msg.m_status & LI_ALARM) == LI_ALARM
1738 || msg.m_stratum == 0
1739 || msg.m_stratum > NTP_MAXSTRATUM
1741 // TODO: stratum 0 responses may have commands in 32-bit m_refid field:
1742 // "DENY", "RSTR" - peer does not like us at all
1743 // "RATE" - peer is overloaded, reduce polling freq
1744 bb_error_msg("reply from %s: peer is unsynced", p->p_dotted);
1745 goto pick_normal_interval;
1748 // /* Verify valid root distance */
1749 // if (msg.m_rootdelay / 2 + msg.m_rootdisp >= MAXDISP || p->lastpkt_reftime > msg.m_xmt)
1750 // return; /* invalid header values */
1753 * From RFC 2030 (with a correction to the delay math):
1755 * Timestamp Name ID When Generated
1756 * ------------------------------------------------------------
1757 * Originate Timestamp T1 time request sent by client
1758 * Receive Timestamp T2 time request received by server
1759 * Transmit Timestamp T3 time reply sent by server
1760 * Destination Timestamp T4 time reply received by client
1762 * The roundtrip delay and local clock offset are defined as
1764 * delay = (T4 - T1) - (T3 - T2); offset = ((T2 - T1) + (T3 - T4)) / 2
1767 T2 = lfp_to_d(msg.m_rectime);
1768 T3 = lfp_to_d(msg.m_xmttime);
1771 /* The delay calculation is a special case. In cases where the
1772 * server and client clocks are running at different rates and
1773 * with very fast networks, the delay can appear negative. In
1774 * order to avoid violating the Principle of Least Astonishment,
1775 * the delay is clamped not less than the system precision.
1777 delay = (T4 - T1) - (T3 - T2);
1778 if (delay < G_precision_sec)
1779 delay = G_precision_sec;
1781 * If this packet's delay is much bigger than the last one,
1782 * it's better to just ignore it than use its much less precise value.
1784 prev_delay = p->p_raw_delay;
1785 p->p_raw_delay = delay;
1786 if (p->reachable_bits && delay > prev_delay * BAD_DELAY_GROWTH) {
1787 bb_error_msg("reply from %s: delay %f is too high, ignoring", p->p_dotted, p->lastpkt_delay);
1788 goto pick_normal_interval;
1791 p->lastpkt_delay = delay;
1792 p->lastpkt_recv_time = T4;
1793 VERB6 bb_error_msg("%s->lastpkt_recv_time=%f", p->p_dotted, p->lastpkt_recv_time);
1794 p->lastpkt_status = msg.m_status;
1795 p->lastpkt_stratum = msg.m_stratum;
1796 p->lastpkt_rootdelay = sfp_to_d(msg.m_rootdelay);
1797 p->lastpkt_rootdisp = sfp_to_d(msg.m_rootdisp);
1798 p->lastpkt_refid = msg.m_refid;
1800 p->datapoint_idx = p->reachable_bits ? (p->datapoint_idx + 1) % NUM_DATAPOINTS : 0;
1801 datapoint = &p->filter_datapoint[p->datapoint_idx];
1802 datapoint->d_recv_time = T4;
1803 datapoint->d_offset = offset = ((T2 - T1) + (T3 - T4)) / 2;
1804 datapoint->d_dispersion = LOG2D(msg.m_precision_exp) + G_precision_sec;
1805 if (!p->reachable_bits) {
1806 /* 1st datapoint ever - replicate offset in every element */
1808 for (i = 0; i < NUM_DATAPOINTS; i++) {
1809 p->filter_datapoint[i].d_offset = offset;
1813 p->reachable_bits |= 1;
1814 if ((MAX_VERBOSE && G.verbose) || (option_mask32 & OPT_w)) {
1815 bb_error_msg("reply from %s: offset:%+f delay:%f status:0x%02x strat:%d refid:0x%08x rootdelay:%f reach:0x%02x",
1822 p->lastpkt_rootdelay,
1824 /* not shown: m_ppoll, m_precision_exp, m_rootdisp,
1825 * m_reftime, m_orgtime, m_rectime, m_xmttime
1830 /* Muck with statictics and update the clock */
1831 filter_datapoints(p);
1832 q = select_and_cluster();
1836 if (!(option_mask32 & OPT_w)) {
1837 rc = update_local_clock(q);
1838 /* If drift is dangerously large, immediately
1839 * drop poll interval one step down.
1841 if (fabs(q->filter_offset) >= POLLDOWN_OFFSET) {
1842 VERB4 bb_error_msg("offset:%+f > POLLDOWN_OFFSET", q->filter_offset);
1847 /* else: no peer selected, rc = -1: we want to poll more often */
1850 /* Adjust the poll interval by comparing the current offset
1851 * with the clock jitter. If the offset is less than
1852 * the clock jitter times a constant, then the averaging interval
1853 * is increased, otherwise it is decreased. A bit of hysteresis
1854 * helps calm the dance. Works best using burst mode.
1856 if (rc > 0 && G.offset_to_jitter_ratio <= POLLADJ_GATE) {
1857 /* was += G.poll_exp but it is a bit
1858 * too optimistic for my taste at high poll_exp's */
1859 G.polladj_count += MINPOLL;
1860 if (G.polladj_count > POLLADJ_LIMIT) {
1861 G.polladj_count = 0;
1862 if (G.poll_exp < MAXPOLL) {
1864 VERB4 bb_error_msg("polladj: discipline_jitter:%f ++poll_exp=%d",
1865 G.discipline_jitter, G.poll_exp);
1868 VERB4 bb_error_msg("polladj: incr:%d", G.polladj_count);
1871 G.polladj_count -= G.poll_exp * 2;
1872 if (G.polladj_count < -POLLADJ_LIMIT || G.poll_exp >= BIGPOLL) {
1874 G.polladj_count = 0;
1875 if (G.poll_exp > MINPOLL) {
1879 /* Correct p->next_action_time in each peer
1880 * which waits for sending, so that they send earlier.
1881 * Old pp->next_action_time are on the order
1882 * of t + (1 << old_poll_exp) + small_random,
1883 * we simply need to subtract ~half of that.
1885 for (item = G.ntp_peers; item != NULL; item = item->link) {
1886 peer_t *pp = (peer_t *) item->data;
1888 pp->next_action_time -= (1 << G.poll_exp);
1890 VERB4 bb_error_msg("polladj: discipline_jitter:%f --poll_exp=%d",
1891 G.discipline_jitter, G.poll_exp);
1894 VERB4 bb_error_msg("polladj: decr:%d", G.polladj_count);
1899 /* Decide when to send new query for this peer */
1900 pick_normal_interval:
1901 interval = poll_interval(0);
1902 if (fabs(offset) >= STEP_THRESHOLD * 8 && interval > BIGOFF_INTERVAL) {
1903 /* If we are synced, offsets are less than STEP_THRESHOLD,
1904 * or at the very least not much larger than it.
1905 * Now we see a largish one.
1906 * Either this peer is feeling bad, or packet got corrupted,
1907 * or _our_ clock is wrong now and _all_ peers will show similar
1908 * largish offsets too.
1909 * I observed this with laptop suspend stopping clock.
1910 * In any case, it makes sense to make next request soonish:
1911 * cases 1 and 2: get a better datapoint,
1912 * case 3: allows to resync faster.
1914 interval = BIGOFF_INTERVAL;
1918 set_next(p, interval);
1921 #if ENABLE_FEATURE_NTPD_SERVER
1922 static NOINLINE void
1923 recv_and_process_client_pkt(void /*int fd*/)
1927 len_and_sockaddr *to;
1928 struct sockaddr *from;
1930 uint8_t query_status;
1931 l_fixedpt_t query_xmttime;
1933 to = get_sock_lsa(G_listen_fd);
1934 from = xzalloc(to->len);
1936 size = recv_from_to(G_listen_fd, &msg, sizeof(msg), MSG_DONTWAIT, from, &to->u.sa, to->len);
1937 if (size != NTP_MSGSIZE_NOAUTH && size != NTP_MSGSIZE) {
1940 if (errno == EAGAIN)
1942 bb_perror_msg_and_die("recv");
1944 addr = xmalloc_sockaddr2dotted_noport(from);
1945 bb_error_msg("malformed packet received from %s: size %u", addr, (int)size);
1950 query_status = msg.m_status;
1951 query_xmttime = msg.m_xmttime;
1953 /* Build a reply packet */
1954 memset(&msg, 0, sizeof(msg));
1955 msg.m_status = G.stratum < MAXSTRAT ? (G.ntp_status & LI_MASK) : LI_ALARM;
1956 msg.m_status |= (query_status & VERSION_MASK);
1957 msg.m_status |= ((query_status & MODE_MASK) == MODE_CLIENT) ?
1958 MODE_SERVER : MODE_SYM_PAS;
1959 msg.m_stratum = G.stratum;
1960 msg.m_ppoll = G.poll_exp;
1961 msg.m_precision_exp = G_precision_exp;
1962 /* this time was obtained between poll() and recv() */
1963 msg.m_rectime = d_to_lfp(G.cur_time);
1964 msg.m_xmttime = d_to_lfp(gettime1900d()); /* this instant */
1965 if (G.peer_cnt == 0) {
1966 /* we have no peers: "stratum 1 server" mode. reftime = our own time */
1967 G.reftime = G.cur_time;
1969 msg.m_reftime = d_to_lfp(G.reftime);
1970 msg.m_orgtime = query_xmttime;
1971 msg.m_rootdelay = d_to_sfp(G.rootdelay);
1972 //simple code does not do this, fix simple code!
1973 msg.m_rootdisp = d_to_sfp(G.rootdisp);
1974 //version = (query_status & VERSION_MASK); /* ... >> VERSION_SHIFT - done below instead */
1975 msg.m_refid = G.refid; // (version > (3 << VERSION_SHIFT)) ? G.refid : G.refid3;
1977 /* We reply from the local address packet was sent to,
1978 * this makes to/from look swapped here: */
1979 do_sendto(G_listen_fd,
1980 /*from:*/ &to->u.sa, /*to:*/ from, /*addrlen:*/ to->len,
1989 /* Upstream ntpd's options:
1991 * -4 Force DNS resolution of host names to the IPv4 namespace.
1992 * -6 Force DNS resolution of host names to the IPv6 namespace.
1993 * -a Require cryptographic authentication for broadcast client,
1994 * multicast client and symmetric passive associations.
1995 * This is the default.
1996 * -A Do not require cryptographic authentication for broadcast client,
1997 * multicast client and symmetric passive associations.
1998 * This is almost never a good idea.
1999 * -b Enable the client to synchronize to broadcast servers.
2001 * Specify the name and path of the configuration file,
2002 * default /etc/ntp.conf
2003 * -d Specify debugging mode. This option may occur more than once,
2004 * with each occurrence indicating greater detail of display.
2006 * Specify debugging level directly.
2008 * Specify the name and path of the frequency file.
2009 * This is the same operation as the "driftfile FILE"
2010 * configuration command.
2011 * -g Normally, ntpd exits with a message to the system log
2012 * if the offset exceeds the panic threshold, which is 1000 s
2013 * by default. This option allows the time to be set to any value
2014 * without restriction; however, this can happen only once.
2015 * If the threshold is exceeded after that, ntpd will exit
2016 * with a message to the system log. This option can be used
2017 * with the -q and -x options. See the tinker command for other options.
2019 * Chroot the server to the directory jaildir. This option also implies
2020 * that the server attempts to drop root privileges at startup
2021 * (otherwise, chroot gives very little additional security).
2022 * You may need to also specify a -u option.
2024 * Specify the name and path of the symmetric key file,
2025 * default /etc/ntp/keys. This is the same operation
2026 * as the "keys FILE" configuration command.
2028 * Specify the name and path of the log file. The default
2029 * is the system log file. This is the same operation as
2030 * the "logfile FILE" configuration command.
2031 * -L Do not listen to virtual IPs. The default is to listen.
2033 * -N To the extent permitted by the operating system,
2034 * run the ntpd at the highest priority.
2036 * Specify the name and path of the file used to record the ntpd
2037 * process ID. This is the same operation as the "pidfile FILE"
2038 * configuration command.
2040 * To the extent permitted by the operating system,
2041 * run the ntpd at the specified priority.
2042 * -q Exit the ntpd just after the first time the clock is set.
2043 * This behavior mimics that of the ntpdate program, which is
2044 * to be retired. The -g and -x options can be used with this option.
2045 * Note: The kernel time discipline is disabled with this option.
2047 * Specify the default propagation delay from the broadcast/multicast
2048 * server to this client. This is necessary only if the delay
2049 * cannot be computed automatically by the protocol.
2051 * Specify the directory path for files created by the statistics
2052 * facility. This is the same operation as the "statsdir DIR"
2053 * configuration command.
2055 * Add a key number to the trusted key list. This option can occur
2058 * Specify a user, and optionally a group, to switch to.
2061 * Add a system variable listed by default.
2062 * -x Normally, the time is slewed if the offset is less than the step
2063 * threshold, which is 128 ms by default, and stepped if above
2064 * the threshold. This option sets the threshold to 600 s, which is
2065 * well within the accuracy window to set the clock manually.
2066 * Note: since the slew rate of typical Unix kernels is limited
2067 * to 0.5 ms/s, each second of adjustment requires an amortization
2068 * interval of 2000 s. Thus, an adjustment as much as 600 s
2069 * will take almost 14 days to complete. This option can be used
2070 * with the -g and -q options. See the tinker command for other options.
2071 * Note: The kernel time discipline is disabled with this option.
2074 /* By doing init in a separate function we decrease stack usage
2077 static NOINLINE void ntp_init(char **argv)
2085 bb_error_msg_and_die(bb_msg_you_must_be_root);
2087 /* Set some globals */
2088 G.stratum = MAXSTRAT;
2090 G.poll_exp = BURSTPOLL; /* speeds up initial sync */
2091 G.last_script_run = G.reftime = G.last_update_recv_time = gettime1900d(); /* sets G.cur_time too */
2095 opt_complementary = "dd:p::wn"; /* d: counter; p: list; -w implies -n */
2096 opts = getopt32(argv,
2098 "wp:S:"IF_FEATURE_NTPD_SERVER("l") /* NOT compat */
2100 "46aAbgL", /* compat, ignored */
2101 &peers, &G.script_name, &G.verbose);
2103 // if (opts & OPT_x) /* disable stepping, only slew is allowed */
2104 // G.time_was_stepped = 1;
2107 add_peers(llist_pop(&peers));
2109 #if ENABLE_FEATURE_NTPD_CONF
2114 parser = config_open("/etc/ntp.conf");
2115 while (config_read(parser, token, 3, 1, "# \t", PARSE_NORMAL)) {
2116 if (strcmp(token[0], "server") == 0 && token[1]) {
2117 add_peers(token[1]);
2120 bb_error_msg("skipping %s:%u: unimplemented command '%s'",
2121 "/etc/ntp.conf", parser->lineno, token[0]
2124 config_close(parser);
2127 if (G.peer_cnt == 0) {
2128 if (!(opts & OPT_l))
2130 /* -l but no peers: "stratum 1 server" mode */
2133 if (!(opts & OPT_n)) {
2134 bb_daemonize_or_rexec(DAEMON_DEVNULL_STDIO, argv);
2135 logmode = LOGMODE_NONE;
2137 #if ENABLE_FEATURE_NTPD_SERVER
2140 G_listen_fd = create_and_bind_dgram_or_die(NULL, 123);
2141 socket_want_pktinfo(G_listen_fd);
2142 setsockopt(G_listen_fd, IPPROTO_IP, IP_TOS, &const_IPTOS_LOWDELAY, sizeof(const_IPTOS_LOWDELAY));
2145 /* I hesitate to set -20 prio. -15 should be high enough for timekeeping */
2147 setpriority(PRIO_PROCESS, 0, -15);
2149 /* If network is up, syncronization occurs in ~10 seconds.
2150 * We give "ntpd -q" 10 seconds to get first reply,
2151 * then another 50 seconds to finish syncing.
2153 * I tested ntpd 4.2.6p1 and apparently it never exits
2154 * (will try forever), but it does not feel right.
2155 * The goal of -q is to act like ntpdate: set time
2156 * after a reasonably small period of polling, or fail.
2159 option_mask32 |= OPT_qq;
2176 int ntpd_main(int argc UNUSED_PARAM, char **argv) MAIN_EXTERNALLY_VISIBLE;
2177 int ntpd_main(int argc UNUSED_PARAM, char **argv)
2185 memset(&G, 0, sizeof(G));
2186 SET_PTR_TO_GLOBALS(&G);
2190 /* If ENABLE_FEATURE_NTPD_SERVER, + 1 for listen_fd: */
2191 cnt = G.peer_cnt + ENABLE_FEATURE_NTPD_SERVER;
2192 idx2peer = xzalloc(sizeof(idx2peer[0]) * cnt);
2193 pfd = xzalloc(sizeof(pfd[0]) * cnt);
2195 /* Countdown: we never sync before we sent INITIAL_SAMPLES+1
2196 * packets to each peer.
2197 * NB: if some peer is not responding, we may end up sending
2198 * fewer packets to it and more to other peers.
2199 * NB2: sync usually happens using INITIAL_SAMPLES packets,
2200 * since last reply does not come back instantaneously.
2202 cnt = G.peer_cnt * (INITIAL_SAMPLES + 1);
2204 write_pidfile(CONFIG_PID_FILE_PATH "/ntpd.pid");
2206 while (!bb_got_signal) {
2212 /* Nothing between here and poll() blocks for any significant time */
2214 nextaction = G.cur_time + 3600;
2217 #if ENABLE_FEATURE_NTPD_SERVER
2218 if (G_listen_fd != -1) {
2219 pfd[0].fd = G_listen_fd;
2220 pfd[0].events = POLLIN;
2224 /* Pass over peer list, send requests, time out on receives */
2225 for (item = G.ntp_peers; item != NULL; item = item->link) {
2226 peer_t *p = (peer_t *) item->data;
2228 if (p->next_action_time <= G.cur_time) {
2229 if (p->p_fd == -1) {
2230 /* Time to send new req */
2232 G.initial_poll_complete = 1;
2234 send_query_to_peer(p);
2236 /* Timed out waiting for reply */
2239 timeout = poll_interval(-2); /* -2: try a bit sooner */
2240 bb_error_msg("timed out waiting for %s, reach 0x%02x, next query in %us",
2241 p->p_dotted, p->reachable_bits, timeout);
2242 set_next(p, timeout);
2246 if (p->next_action_time < nextaction)
2247 nextaction = p->next_action_time;
2250 /* Wait for reply from this peer */
2251 pfd[i].fd = p->p_fd;
2252 pfd[i].events = POLLIN;
2258 timeout = nextaction - G.cur_time;
2261 timeout++; /* (nextaction - G.cur_time) rounds down, compensating */
2263 /* Here we may block */
2265 if (i > (ENABLE_FEATURE_NTPD_SERVER && G_listen_fd != -1)) {
2266 /* We wait for at least one reply.
2267 * Poll for it, without wasting time for message.
2268 * Since replies often come under 1 second, this also
2269 * reduces clutter in logs.
2271 nfds = poll(pfd, i, 1000);
2277 bb_error_msg("poll:%us sockets:%u interval:%us", timeout, i, 1 << G.poll_exp);
2279 nfds = poll(pfd, i, timeout * 1000);
2281 gettime1900d(); /* sets G.cur_time */
2283 if (!bb_got_signal /* poll wasn't interrupted by a signal */
2284 && G.cur_time - G.last_script_run > 11*60
2286 /* Useful for updating battery-backed RTC and such */
2287 run_script("periodic", G.last_update_offset);
2288 gettime1900d(); /* sets G.cur_time */
2293 /* Process any received packets */
2295 #if ENABLE_FEATURE_NTPD_SERVER
2296 if (G.listen_fd != -1) {
2297 if (pfd[0].revents /* & (POLLIN|POLLERR)*/) {
2299 recv_and_process_client_pkt(/*G.listen_fd*/);
2300 gettime1900d(); /* sets G.cur_time */
2305 for (; nfds != 0 && j < i; j++) {
2306 if (pfd[j].revents /* & (POLLIN|POLLERR)*/) {
2308 * At init, alarm was set to 10 sec.
2309 * Now we did get a reply.
2310 * Increase timeout to 50 seconds to finish syncing.
2312 if (option_mask32 & OPT_qq) {
2313 option_mask32 &= ~OPT_qq;
2317 recv_and_process_peer_pkt(idx2peer[j]);
2318 gettime1900d(); /* sets G.cur_time */
2323 if (G.ntp_peers && G.stratum != MAXSTRAT) {
2324 for (item = G.ntp_peers; item != NULL; item = item->link) {
2325 peer_t *p = (peer_t *) item->data;
2326 if (p->reachable_bits)
2327 goto have_reachable_peer;
2329 /* No peer responded for last 8 packets, panic */
2330 G.polladj_count = 0;
2331 G.poll_exp = MINPOLL;
2332 G.stratum = MAXSTRAT;
2333 run_script("unsync", 0.0);
2334 have_reachable_peer: ;
2336 } /* while (!bb_got_signal) */
2338 remove_pidfile(CONFIG_PID_FILE_PATH "/ntpd.pid");
2339 kill_myself_with_sig(bb_got_signal);
2347 /*** openntpd-4.6 uses only adjtime, not adjtimex ***/
2349 /*** ntp-4.2.6/ntpd/ntp_loopfilter.c - adjtimex usage ***/
2353 direct_freq(double fp_offset)
2357 * If the kernel is enabled, we need the residual offset to
2358 * calculate the frequency correction.
2360 if (pll_control && kern_enable) {
2361 memset(&ntv, 0, sizeof(ntv));
2364 clock_offset = ntv.offset / 1e9;
2365 #else /* STA_NANO */
2366 clock_offset = ntv.offset / 1e6;
2367 #endif /* STA_NANO */
2368 drift_comp = FREQTOD(ntv.freq);
2370 #endif /* KERNEL_PLL */
2371 set_freq((fp_offset - clock_offset) / (current_time - clock_epoch) + drift_comp);
2377 set_freq(double freq) /* frequency update */
2385 * If the kernel is enabled, update the kernel frequency.
2387 if (pll_control && kern_enable) {
2388 memset(&ntv, 0, sizeof(ntv));
2389 ntv.modes = MOD_FREQUENCY;
2390 ntv.freq = DTOFREQ(drift_comp);
2392 snprintf(tbuf, sizeof(tbuf), "kernel %.3f PPM", drift_comp * 1e6);
2393 report_event(EVNT_FSET, NULL, tbuf);
2395 snprintf(tbuf, sizeof(tbuf), "ntpd %.3f PPM", drift_comp * 1e6);
2396 report_event(EVNT_FSET, NULL, tbuf);
2398 #else /* KERNEL_PLL */
2399 snprintf(tbuf, sizeof(tbuf), "ntpd %.3f PPM", drift_comp * 1e6);
2400 report_event(EVNT_FSET, NULL, tbuf);
2401 #endif /* KERNEL_PLL */
2410 * This code segment works when clock adjustments are made using
2411 * precision time kernel support and the ntp_adjtime() system
2412 * call. This support is available in Solaris 2.6 and later,
2413 * Digital Unix 4.0 and later, FreeBSD, Linux and specially
2414 * modified kernels for HP-UX 9 and Ultrix 4. In the case of the
2415 * DECstation 5000/240 and Alpha AXP, additional kernel
2416 * modifications provide a true microsecond clock and nanosecond
2417 * clock, respectively.
2419 * Important note: The kernel discipline is used only if the
2420 * step threshold is less than 0.5 s, as anything higher can
2421 * lead to overflow problems. This might occur if some misguided
2422 * lad set the step threshold to something ridiculous.
2424 if (pll_control && kern_enable) {
2426 #define MOD_BITS (MOD_OFFSET | MOD_MAXERROR | MOD_ESTERROR | MOD_STATUS | MOD_TIMECONST)
2429 * We initialize the structure for the ntp_adjtime()
2430 * system call. We have to convert everything to
2431 * microseconds or nanoseconds first. Do not update the
2432 * system variables if the ext_enable flag is set. In
2433 * this case, the external clock driver will update the
2434 * variables, which will be read later by the local
2435 * clock driver. Afterwards, remember the time and
2436 * frequency offsets for jitter and stability values and
2437 * to update the frequency file.
2439 memset(&ntv, 0, sizeof(ntv));
2441 ntv.modes = MOD_STATUS;
2444 ntv.modes = MOD_BITS | MOD_NANO;
2445 #else /* STA_NANO */
2446 ntv.modes = MOD_BITS;
2447 #endif /* STA_NANO */
2448 if (clock_offset < 0)
2453 ntv.offset = (int32)(clock_offset * 1e9 + dtemp);
2454 ntv.constant = sys_poll;
2455 #else /* STA_NANO */
2456 ntv.offset = (int32)(clock_offset * 1e6 + dtemp);
2457 ntv.constant = sys_poll - 4;
2458 #endif /* STA_NANO */
2459 ntv.esterror = (u_int32)(clock_jitter * 1e6);
2460 ntv.maxerror = (u_int32)((sys_rootdelay / 2 + sys_rootdisp) * 1e6);
2461 ntv.status = STA_PLL;
2464 * Enable/disable the PPS if requested.
2467 if (!(pll_status & STA_PPSTIME))
2468 report_event(EVNT_KERN,
2469 NULL, "PPS enabled");
2470 ntv.status |= STA_PPSTIME | STA_PPSFREQ;
2472 if (pll_status & STA_PPSTIME)
2473 report_event(EVNT_KERN,
2474 NULL, "PPS disabled");
2475 ntv.status &= ~(STA_PPSTIME | STA_PPSFREQ);
2477 if (sys_leap == LEAP_ADDSECOND)
2478 ntv.status |= STA_INS;
2479 else if (sys_leap == LEAP_DELSECOND)
2480 ntv.status |= STA_DEL;
2484 * Pass the stuff to the kernel. If it squeals, turn off
2485 * the pps. In any case, fetch the kernel offset,
2486 * frequency and jitter.
2488 if (ntp_adjtime(&ntv) == TIME_ERROR) {
2489 if (!(ntv.status & STA_PPSSIGNAL))
2490 report_event(EVNT_KERN, NULL,
2493 pll_status = ntv.status;
2495 clock_offset = ntv.offset / 1e9;
2496 #else /* STA_NANO */
2497 clock_offset = ntv.offset / 1e6;
2498 #endif /* STA_NANO */
2499 clock_frequency = FREQTOD(ntv.freq);
2502 * If the kernel PPS is lit, monitor its performance.
2504 if (ntv.status & STA_PPSTIME) {
2506 clock_jitter = ntv.jitter / 1e9;
2507 #else /* STA_NANO */
2508 clock_jitter = ntv.jitter / 1e6;
2509 #endif /* STA_NANO */
2512 #if defined(STA_NANO) && NTP_API == 4
2514 * If the TAI changes, update the kernel TAI.
2516 if (loop_tai != sys_tai) {
2518 ntv.modes = MOD_TAI;
2519 ntv.constant = sys_tai;
2522 #endif /* STA_NANO */
2524 #endif /* KERNEL_PLL */