2 * NTP client/server, based on OpenNTPD 3.9p1
4 * Busybox port author: Adam Tkac (C) 2009 <vonsch@gmail.com>
6 * OpenNTPd 3.9p1 copyright holders:
7 * Copyright (c) 2003, 2004 Henning Brauer <henning@openbsd.org>
8 * Copyright (c) 2004 Alexander Guy <alexander.guy@andern.org>
10 * OpenNTPd code is licensed under ISC-style licence:
12 * Permission to use, copy, modify, and distribute this software for any
13 * purpose with or without fee is hereby granted, provided that the above
14 * copyright notice and this permission notice appear in all copies.
16 * THE SOFTWARE IS PROVIDED "AS IS" AND THE AUTHOR DISCLAIMS ALL WARRANTIES
17 * WITH REGARD TO THIS SOFTWARE INCLUDING ALL IMPLIED WARRANTIES OF
18 * MERCHANTABILITY AND FITNESS. IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR
19 * ANY SPECIAL, DIRECT, INDIRECT, OR CONSEQUENTIAL DAMAGES OR ANY DAMAGES
20 * WHATSOEVER RESULTING FROM LOSS OF MIND, USE, DATA OR PROFITS, WHETHER
21 * IN AN ACTION OF CONTRACT, NEGLIGENCE OR OTHER TORTIOUS ACTION, ARISING
22 * OUT OF OR IN CONNECTION WITH THE USE OR PERFORMANCE OF THIS SOFTWARE.
23 ***********************************************************************
25 * Parts of OpenNTPD clock syncronization code is replaced by
26 * code which is based on ntp-4.2.6, which carries the following
29 * Copyright (c) University of Delaware 1992-2009
31 * Permission to use, copy, modify, and distribute this software and
32 * its documentation for any purpose with or without fee is hereby
33 * granted, provided that the above copyright notice appears in all
34 * copies and that both the copyright notice and this permission
35 * notice appear in supporting documentation, and that the name
36 * University of Delaware not be used in advertising or publicity
37 * pertaining to distribution of the software without specific,
38 * written prior permission. The University of Delaware makes no
39 * representations about the suitability this software for any
40 * purpose. It is provided "as is" without express or implied warranty.
41 ***********************************************************************
44 //usage:#define ntpd_trivial_usage
45 //usage: "[-dnqNw"IF_FEATURE_NTPD_SERVER("l -I IFACE")"] [-S PROG] [-p PEER]..."
46 //usage:#define ntpd_full_usage "\n\n"
47 //usage: "NTP client/server\n"
48 //usage: "\n -d Verbose"
49 //usage: "\n -n Do not daemonize"
50 //usage: "\n -q Quit after clock is set"
51 //usage: "\n -N Run at high priority"
52 //usage: "\n -w Do not set time (only query peers), implies -n"
53 //usage: "\n -S PROG Run PROG after stepping time, stratum change, and every 11 mins"
54 //usage: "\n -p PEER Obtain time from PEER (may be repeated)"
55 //usage: IF_FEATURE_NTPD_CONF(
56 //usage: "\n If -p is not given, read /etc/ntp.conf"
58 //usage: IF_FEATURE_NTPD_SERVER(
59 //usage: "\n -l Also run as server on port 123"
60 //usage: "\n -I IFACE Bind server to IFACE, implies -l"
63 // -l and -p options are not compatible with "standard" ntpd:
64 // it has them as "-l logfile" and "-p pidfile".
65 // -S and -w are not compat either, "standard" ntpd has no such opts.
69 #include <netinet/ip.h> /* For IPTOS_LOWDELAY definition */
70 #include <sys/resource.h> /* setpriority */
71 #include <sys/timex.h>
72 #ifndef IPTOS_LOWDELAY
73 # define IPTOS_LOWDELAY 0x10
77 /* Verbosity control (max level of -dddd options accepted).
78 * max 6 is very talkative (and bloated). 3 is non-bloated,
79 * production level setting.
84 /* High-level description of the algorithm:
86 * We start running with very small poll_exp, BURSTPOLL,
87 * in order to quickly accumulate INITIAL_SAMPLES datapoints
88 * for each peer. Then, time is stepped if the offset is larger
89 * than STEP_THRESHOLD, otherwise it isn't; anyway, we enlarge
90 * poll_exp to MINPOLL and enter frequency measurement step:
91 * we collect new datapoints but ignore them for WATCH_THRESHOLD
92 * seconds. After WATCH_THRESHOLD seconds we look at accumulated
93 * offset and estimate frequency drift.
95 * (frequency measurement step seems to not be strictly needed,
96 * it is conditionally disabled with USING_INITIAL_FREQ_ESTIMATION
99 * After this, we enter "steady state": we collect a datapoint,
100 * we select the best peer, if this datapoint is not a new one
101 * (IOW: if this datapoint isn't for selected peer), sleep
102 * and collect another one; otherwise, use its offset to update
103 * frequency drift, if offset is somewhat large, reduce poll_exp,
104 * otherwise increase poll_exp.
106 * If offset is larger than STEP_THRESHOLD, which shouldn't normally
107 * happen, we assume that something "bad" happened (computer
108 * was hibernated, someone set totally wrong date, etc),
109 * then the time is stepped, all datapoints are discarded,
110 * and we go back to steady state.
112 * Made some changes to speed up re-syncing after our clock goes bad
113 * (tested with suspending my laptop):
114 * - if largish offset (>= STEP_THRESHOLD * 8 == 1 sec) is seen
115 * from a peer, schedule next query for this peer soon
116 * without drastically lowering poll interval for everybody.
117 * This makes us collect enough data for step much faster:
118 * e.g. at poll = 10 (1024 secs), step was done within 5 minutes
119 * after first reply which indicated that our clock is 14 seconds off.
120 * - on step, do not discard d_dispersion data of the existing datapoints,
121 * do not clear reachable_bits. This prevents discarding first ~8
122 * datapoints after the step.
125 #define INITIAL_SAMPLES 4 /* how many samples do we want for init */
126 #define BAD_DELAY_GROWTH 4 /* drop packet if its delay grew by more than this */
128 #define RETRY_INTERVAL 32 /* on send/recv error, retry in N secs (need to be power of 2) */
129 #define NOREPLY_INTERVAL 512 /* sent, but got no reply: cap next query by this many seconds */
130 #define RESPONSE_INTERVAL 16 /* wait for reply up to N secs */
132 /* Step threshold (sec). std ntpd uses 0.128.
133 * Using exact power of 2 (1/8) results in smaller code
135 #define STEP_THRESHOLD 0.125
136 /* Stepout threshold (sec). std ntpd uses 900 (11 mins (!)) */
137 #define WATCH_THRESHOLD 128
138 /* NB: set WATCH_THRESHOLD to ~60 when debugging to save time) */
139 //UNUSED: #define PANIC_THRESHOLD 1000 /* panic threshold (sec) */
142 * If we got |offset| > BIGOFF from a peer, cap next query interval
143 * for this peer by this many seconds:
145 #define BIGOFF (STEP_THRESHOLD * 8)
146 #define BIGOFF_INTERVAL (1 << 7) /* 128 s */
148 #define FREQ_TOLERANCE 0.000015 /* frequency tolerance (15 PPM) */
149 #define BURSTPOLL 0 /* initial poll */
150 #define MINPOLL 5 /* minimum poll interval. std ntpd uses 6 (6: 64 sec) */
152 * If offset > discipline_jitter * POLLADJ_GATE, and poll interval is > 2^BIGPOLL,
153 * then it is decreased _at once_. (If <= 2^BIGPOLL, it will be decreased _eventually_).
155 #define BIGPOLL 9 /* 2^9 sec ~= 8.5 min */
156 #define MAXPOLL 12 /* maximum poll interval (12: 1.1h, 17: 36.4h). std ntpd uses 17 */
158 * Actively lower poll when we see such big offsets.
159 * With STEP_THRESHOLD = 0.125, it means we try to sync more aggressively
160 * if offset increases over ~0.04 sec
162 //#define POLLDOWN_OFFSET (STEP_THRESHOLD / 3)
163 #define MINDISP 0.01 /* minimum dispersion (sec) */
164 #define MAXDISP 16 /* maximum dispersion (sec) */
165 #define MAXSTRAT 16 /* maximum stratum (infinity metric) */
166 #define MAXDIST 1 /* distance threshold (sec) */
167 #define MIN_SELECTED 1 /* minimum intersection survivors */
168 #define MIN_CLUSTERED 3 /* minimum cluster survivors */
170 #define MAXDRIFT 0.000500 /* frequency drift we can correct (500 PPM) */
172 /* Poll-adjust threshold.
173 * When we see that offset is small enough compared to discipline jitter,
174 * we grow a counter: += MINPOLL. When counter goes over POLLADJ_LIMIT,
175 * we poll_exp++. If offset isn't small, counter -= poll_exp*2,
176 * and when it goes below -POLLADJ_LIMIT, we poll_exp--.
177 * (Bumped from 30 to 40 since otherwise I often see poll_exp going *2* steps down)
179 #define POLLADJ_LIMIT 40
180 /* If offset < discipline_jitter * POLLADJ_GATE, then we decide to increase
181 * poll interval (we think we can't improve timekeeping
182 * by staying at smaller poll).
184 #define POLLADJ_GATE 4
185 #define TIMECONST_HACK_GATE 2
186 /* Compromise Allan intercept (sec). doc uses 1500, std ntpd uses 512 */
190 /* FLL loop gain [why it depends on MAXPOLL??] */
191 #define FLL (MAXPOLL + 1)
192 /* Parameter averaging constant */
201 NTP_MSGSIZE_NOAUTH = 48,
202 NTP_MSGSIZE = (NTP_MSGSIZE_NOAUTH + 4 + NTP_DIGESTSIZE),
205 MODE_MASK = (7 << 0),
206 VERSION_MASK = (7 << 3),
210 /* Leap Second Codes (high order two bits of m_status) */
211 LI_NOWARNING = (0 << 6), /* no warning */
212 LI_PLUSSEC = (1 << 6), /* add a second (61 seconds) */
213 LI_MINUSSEC = (2 << 6), /* minus a second (59 seconds) */
214 LI_ALARM = (3 << 6), /* alarm condition */
217 MODE_RES0 = 0, /* reserved */
218 MODE_SYM_ACT = 1, /* symmetric active */
219 MODE_SYM_PAS = 2, /* symmetric passive */
220 MODE_CLIENT = 3, /* client */
221 MODE_SERVER = 4, /* server */
222 MODE_BROADCAST = 5, /* broadcast */
223 MODE_RES1 = 6, /* reserved for NTP control message */
224 MODE_RES2 = 7, /* reserved for private use */
227 //TODO: better base selection
228 #define OFFSET_1900_1970 2208988800UL /* 1970 - 1900 in seconds */
230 #define NUM_DATAPOINTS 8
243 uint8_t m_status; /* status of local clock and leap info */
245 uint8_t m_ppoll; /* poll value */
246 int8_t m_precision_exp;
247 s_fixedpt_t m_rootdelay;
248 s_fixedpt_t m_rootdisp;
250 l_fixedpt_t m_reftime;
251 l_fixedpt_t m_orgtime;
252 l_fixedpt_t m_rectime;
253 l_fixedpt_t m_xmttime;
255 uint8_t m_digest[NTP_DIGESTSIZE];
265 len_and_sockaddr *p_lsa;
269 uint32_t lastpkt_refid;
270 uint8_t lastpkt_status;
271 uint8_t lastpkt_stratum;
272 uint8_t reachable_bits;
273 /* when to send new query (if p_fd == -1)
274 * or when receive times out (if p_fd >= 0): */
275 double next_action_time;
278 /* p_raw_delay is set even by "high delay" packets */
279 /* lastpkt_delay isn't */
280 double lastpkt_recv_time;
281 double lastpkt_delay;
282 double lastpkt_rootdelay;
283 double lastpkt_rootdisp;
284 /* produced by filter algorithm: */
285 double filter_offset;
286 double filter_dispersion;
287 double filter_jitter;
288 datapoint_t filter_datapoint[NUM_DATAPOINTS];
289 /* last sent packet: */
294 #define USING_KERNEL_PLL_LOOP 1
295 #define USING_INITIAL_FREQ_ESTIMATION 0
302 /* Insert new options above this line. */
303 /* Non-compat options: */
307 OPT_l = (1 << 7) * ENABLE_FEATURE_NTPD_SERVER,
308 OPT_I = (1 << 8) * ENABLE_FEATURE_NTPD_SERVER,
309 /* We hijack some bits for other purposes */
315 /* total round trip delay to currently selected reference clock */
317 /* reference timestamp: time when the system clock was last set or corrected */
319 /* total dispersion to currently selected reference clock */
322 double last_script_run;
325 #if ENABLE_FEATURE_NTPD_SERVER
328 # define G_listen_fd (G.listen_fd)
330 # define G_listen_fd (-1)
334 /* refid: 32-bit code identifying the particular server or reference clock
335 * in stratum 0 packets this is a four-character ASCII string,
336 * called the kiss code, used for debugging and monitoring
337 * in stratum 1 packets this is a four-character ASCII string
338 * assigned to the reference clock by IANA. Example: "GPS "
339 * in stratum 2+ packets, it's IPv4 address or 4 first bytes
340 * of MD5 hash of IPv6
344 /* precision is defined as the larger of the resolution and time to
345 * read the clock, in log2 units. For instance, the precision of a
346 * mains-frequency clock incrementing at 60 Hz is 16 ms, even when the
347 * system clock hardware representation is to the nanosecond.
349 * Delays, jitters of various kinds are clamped down to precision.
351 * If precision_sec is too large, discipline_jitter gets clamped to it
352 * and if offset is smaller than discipline_jitter * POLLADJ_GATE, poll
353 * interval grows even though we really can benefit from staying at
354 * smaller one, collecting non-lagged datapoits and correcting offset.
355 * (Lagged datapoits exist when poll_exp is large but we still have
356 * systematic offset error - the time distance between datapoints
357 * is significant and older datapoints have smaller offsets.
358 * This makes our offset estimation a bit smaller than reality)
359 * Due to this effect, setting G_precision_sec close to
360 * STEP_THRESHOLD isn't such a good idea - offsets may grow
361 * too big and we will step. I observed it with -6.
363 * OTOH, setting precision_sec far too small would result in futile
364 * attempts to syncronize to an unachievable precision.
366 * -6 is 1/64 sec, -7 is 1/128 sec and so on.
367 * -8 is 1/256 ~= 0.003906 (worked well for me --vda)
368 * -9 is 1/512 ~= 0.001953 (let's try this for some time)
370 #define G_precision_exp -9
372 * G_precision_exp is used only for construction outgoing packets.
373 * It's ok to set G_precision_sec to a slightly different value
374 * (One which is "nicer looking" in logs).
375 * Exact value would be (1.0 / (1 << (- G_precision_exp))):
377 #define G_precision_sec 0.002
380 #define STATE_NSET 0 /* initial state, "nothing is set" */
381 //#define STATE_FSET 1 /* frequency set from file */
382 //#define STATE_SPIK 2 /* spike detected */
383 //#define STATE_FREQ 3 /* initial frequency */
384 #define STATE_SYNC 4 /* clock synchronized (normal operation) */
385 uint8_t discipline_state; // doc calls it c.state
386 uint8_t poll_exp; // s.poll
387 int polladj_count; // c.count
388 long kernel_freq_drift;
389 peer_t *last_update_peer;
390 double last_update_offset; // c.last
391 double last_update_recv_time; // s.t
392 double discipline_jitter; // c.jitter
393 /* Since we only compare it with ints, can simplify code
394 * by not making this variable floating point:
396 unsigned offset_to_jitter_ratio;
397 //double cluster_offset; // s.offset
398 //double cluster_jitter; // s.jitter
399 #if !USING_KERNEL_PLL_LOOP
400 double discipline_freq_drift; // c.freq
401 /* Maybe conditionally calculate wander? it's used only for logging */
402 double discipline_wander; // c.wander
405 #define G (*ptr_to_globals)
407 static const int const_IPTOS_LOWDELAY = IPTOS_LOWDELAY;
410 #define VERB1 if (MAX_VERBOSE && G.verbose)
411 #define VERB2 if (MAX_VERBOSE >= 2 && G.verbose >= 2)
412 #define VERB3 if (MAX_VERBOSE >= 3 && G.verbose >= 3)
413 #define VERB4 if (MAX_VERBOSE >= 4 && G.verbose >= 4)
414 #define VERB5 if (MAX_VERBOSE >= 5 && G.verbose >= 5)
415 #define VERB6 if (MAX_VERBOSE >= 6 && G.verbose >= 6)
418 static double LOG2D(int a)
421 return 1.0 / (1UL << -a);
424 static ALWAYS_INLINE double SQUARE(double x)
428 static ALWAYS_INLINE double MAXD(double a, double b)
434 static ALWAYS_INLINE double MIND(double a, double b)
440 static NOINLINE double my_SQRT(double X)
447 double Xhalf = X * 0.5;
449 /* Fast and good approximation to 1/sqrt(X), black magic */
451 /*v.i = 0x5f3759df - (v.i >> 1);*/
452 v.i = 0x5f375a86 - (v.i >> 1); /* - this constant is slightly better */
453 invsqrt = v.f; /* better than 0.2% accuracy */
455 /* Refining it using Newton's method: x1 = x0 - f(x0)/f'(x0)
456 * f(x) = 1/(x*x) - X (f==0 when x = 1/sqrt(X))
458 * f(x)/f'(x) = (X - 1/(x*x)) / (2/(x*x*x)) = X*x*x*x/2 - x/2
459 * x1 = x0 - (X*x0*x0*x0/2 - x0/2) = 1.5*x0 - X*x0*x0*x0/2 = x0*(1.5 - (X/2)*x0*x0)
461 invsqrt = invsqrt * (1.5 - Xhalf * invsqrt * invsqrt); /* ~0.05% accuracy */
462 /* invsqrt = invsqrt * (1.5 - Xhalf * invsqrt * invsqrt); 2nd iter: ~0.0001% accuracy */
463 /* With 4 iterations, more than half results will be exact,
464 * at 6th iterations result stabilizes with about 72% results exact.
465 * We are well satisfied with 0.05% accuracy.
468 return X * invsqrt; /* X * 1/sqrt(X) ~= sqrt(X) */
470 static ALWAYS_INLINE double SQRT(double X)
472 /* If this arch doesn't use IEEE 754 floats, fall back to using libm */
473 if (sizeof(float) != 4)
476 /* This avoids needing libm, saves about 0.5k on x86-32 */
484 gettimeofday(&tv, NULL); /* never fails */
485 G.cur_time = tv.tv_sec + (1.0e-6 * tv.tv_usec) + OFFSET_1900_1970;
490 d_to_tv(double d, struct timeval *tv)
492 tv->tv_sec = (long)d;
493 tv->tv_usec = (d - tv->tv_sec) * 1000000;
497 lfp_to_d(l_fixedpt_t lfp)
500 lfp.int_partl = ntohl(lfp.int_partl);
501 lfp.fractionl = ntohl(lfp.fractionl);
502 ret = (double)lfp.int_partl + ((double)lfp.fractionl / UINT_MAX);
506 sfp_to_d(s_fixedpt_t sfp)
509 sfp.int_parts = ntohs(sfp.int_parts);
510 sfp.fractions = ntohs(sfp.fractions);
511 ret = (double)sfp.int_parts + ((double)sfp.fractions / USHRT_MAX);
514 #if ENABLE_FEATURE_NTPD_SERVER
519 lfp.int_partl = (uint32_t)d;
520 lfp.fractionl = (uint32_t)((d - lfp.int_partl) * UINT_MAX);
521 lfp.int_partl = htonl(lfp.int_partl);
522 lfp.fractionl = htonl(lfp.fractionl);
529 sfp.int_parts = (uint16_t)d;
530 sfp.fractions = (uint16_t)((d - sfp.int_parts) * USHRT_MAX);
531 sfp.int_parts = htons(sfp.int_parts);
532 sfp.fractions = htons(sfp.fractions);
538 dispersion(const datapoint_t *dp)
540 return dp->d_dispersion + FREQ_TOLERANCE * (G.cur_time - dp->d_recv_time);
544 root_distance(peer_t *p)
546 /* The root synchronization distance is the maximum error due to
547 * all causes of the local clock relative to the primary server.
548 * It is defined as half the total delay plus total dispersion
551 return MAXD(MINDISP, p->lastpkt_rootdelay + p->lastpkt_delay) / 2
552 + p->lastpkt_rootdisp
553 + p->filter_dispersion
554 + FREQ_TOLERANCE * (G.cur_time - p->lastpkt_recv_time)
559 set_next(peer_t *p, unsigned t)
561 p->next_action_time = G.cur_time + t;
565 * Peer clock filter and its helpers
568 filter_datapoints(peer_t *p)
575 /* Simulations have shown that use of *averaged* offset for p->filter_offset
576 * is in fact worse than simply using last received one: with large poll intervals
577 * (>= 2048) averaging code uses offset values which are outdated by hours,
578 * and time/frequency correction goes totally wrong when fed essentially bogus offsets.
581 double minoff, maxoff, w;
582 double x = x; /* for compiler */
583 double oldest_off = oldest_off;
584 double oldest_age = oldest_age;
585 double newest_off = newest_off;
586 double newest_age = newest_age;
588 fdp = p->filter_datapoint;
590 minoff = maxoff = fdp[0].d_offset;
591 for (i = 1; i < NUM_DATAPOINTS; i++) {
592 if (minoff > fdp[i].d_offset)
593 minoff = fdp[i].d_offset;
594 if (maxoff < fdp[i].d_offset)
595 maxoff = fdp[i].d_offset;
598 idx = p->datapoint_idx; /* most recent datapoint's index */
600 * Drop two outliers and take weighted average of the rest:
601 * most_recent/2 + older1/4 + older2/8 ... + older5/32 + older6/32
602 * we use older6/32, not older6/64 since sum of weights should be 1:
603 * 1/2 + 1/4 + 1/8 + 1/16 + 1/32 + 1/32 = 1
609 * filter_dispersion = \ -------------
616 for (i = 0; i < NUM_DATAPOINTS; i++) {
618 bb_error_msg("datapoint[%d]: off:%f disp:%f(%f) age:%f%s",
621 fdp[idx].d_dispersion, dispersion(&fdp[idx]),
622 G.cur_time - fdp[idx].d_recv_time,
623 (minoff == fdp[idx].d_offset || maxoff == fdp[idx].d_offset)
624 ? " (outlier by offset)" : ""
628 sum += dispersion(&fdp[idx]) / (2 << i);
630 if (minoff == fdp[idx].d_offset) {
631 minoff -= 1; /* so that we don't match it ever again */
633 if (maxoff == fdp[idx].d_offset) {
636 oldest_off = fdp[idx].d_offset;
637 oldest_age = G.cur_time - fdp[idx].d_recv_time;
640 newest_off = oldest_off;
641 newest_age = oldest_age;
648 idx = (idx - 1) & (NUM_DATAPOINTS - 1);
650 p->filter_dispersion = sum;
651 wavg += x; /* add another older6/64 to form older6/32 */
652 /* Fix systematic underestimation with large poll intervals.
653 * Imagine that we still have a bit of uncorrected drift,
654 * and poll interval is big (say, 100 sec). Offsets form a progression:
655 * 0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 - 0.7 is most recent.
656 * The algorithm above drops 0.0 and 0.7 as outliers,
657 * and then we have this estimation, ~25% off from 0.7:
658 * 0.1/32 + 0.2/32 + 0.3/16 + 0.4/8 + 0.5/4 + 0.6/2 = 0.503125
660 x = oldest_age - newest_age;
662 x = newest_age / x; /* in above example, 100 / (600 - 100) */
663 if (x < 1) { /* paranoia check */
664 x = (newest_off - oldest_off) * x; /* 0.5 * 100/500 = 0.1 */
668 p->filter_offset = wavg;
672 fdp = p->filter_datapoint;
673 idx = p->datapoint_idx; /* most recent datapoint's index */
675 /* filter_offset: simply use the most recent value */
676 p->filter_offset = fdp[idx].d_offset;
680 * filter_dispersion = \ -------------
687 for (i = 0; i < NUM_DATAPOINTS; i++) {
688 sum += dispersion(&fdp[idx]) / (2 << i);
689 wavg += fdp[idx].d_offset;
690 idx = (idx - 1) & (NUM_DATAPOINTS - 1);
692 wavg /= NUM_DATAPOINTS;
693 p->filter_dispersion = sum;
696 /* +----- -----+ ^ 1/2
700 * filter_jitter = | --- * / (avg-offset_j) |
704 * where n is the number of valid datapoints in the filter (n > 1);
705 * if filter_jitter < precision then filter_jitter = precision
708 for (i = 0; i < NUM_DATAPOINTS; i++) {
709 sum += SQUARE(wavg - fdp[i].d_offset);
711 sum = SQRT(sum / NUM_DATAPOINTS);
712 p->filter_jitter = sum > G_precision_sec ? sum : G_precision_sec;
714 VERB4 bb_error_msg("filter offset:%+f disp:%f jitter:%f",
716 p->filter_dispersion,
721 reset_peer_stats(peer_t *p, double offset)
724 bool small_ofs = fabs(offset) < 16 * STEP_THRESHOLD;
726 /* Used to set p->filter_datapoint[i].d_dispersion = MAXDISP
727 * and clear reachable bits, but this proved to be too agressive:
728 * after step (tested with suspinding laptop for ~30 secs),
729 * this caused all previous data to be considered invalid,
730 * making us needing to collect full ~8 datapoins per peer
731 * after step in order to start trusting them.
732 * In turn, this was making poll interval decrease even after
733 * step was done. (Poll interval decreases already before step
734 * in this scenario, because we see large offsets and end up with
735 * no good peer to select).
738 for (i = 0; i < NUM_DATAPOINTS; i++) {
740 p->filter_datapoint[i].d_recv_time += offset;
741 if (p->filter_datapoint[i].d_offset != 0) {
742 p->filter_datapoint[i].d_offset -= offset;
743 //bb_error_msg("p->filter_datapoint[%d].d_offset %f -> %f",
745 // p->filter_datapoint[i].d_offset + offset,
746 // p->filter_datapoint[i].d_offset);
749 p->filter_datapoint[i].d_recv_time = G.cur_time;
750 p->filter_datapoint[i].d_offset = 0;
751 /*p->filter_datapoint[i].d_dispersion = MAXDISP;*/
755 p->lastpkt_recv_time += offset;
757 /*p->reachable_bits = 0;*/
758 p->lastpkt_recv_time = G.cur_time;
760 filter_datapoints(p); /* recalc p->filter_xxx */
761 VERB6 bb_error_msg("%s->lastpkt_recv_time=%f", p->p_dotted, p->lastpkt_recv_time);
765 add_peers(const char *s)
769 p = xzalloc(sizeof(*p));
770 p->p_lsa = xhost2sockaddr(s, 123);
771 p->p_dotted = xmalloc_sockaddr2dotted_noport(&p->p_lsa->u.sa);
773 p->p_xmt_msg.m_status = MODE_CLIENT | (NTP_VERSION << 3);
774 p->next_action_time = G.cur_time; /* = set_next(p, 0); */
775 reset_peer_stats(p, 16 * STEP_THRESHOLD);
777 llist_add_to(&G.ntp_peers, p);
783 const struct sockaddr *from, const struct sockaddr *to, socklen_t addrlen,
784 msg_t *msg, ssize_t len)
790 ret = sendto(fd, msg, len, MSG_DONTWAIT, to, addrlen);
792 ret = send_to_from(fd, msg, len, MSG_DONTWAIT, to, from, addrlen);
795 bb_perror_msg("send failed");
802 send_query_to_peer(peer_t *p)
804 /* Why do we need to bind()?
805 * See what happens when we don't bind:
807 * socket(PF_INET, SOCK_DGRAM, IPPROTO_IP) = 3
808 * setsockopt(3, SOL_IP, IP_TOS, [16], 4) = 0
809 * gettimeofday({1259071266, 327885}, NULL) = 0
810 * sendto(3, "xxx", 48, MSG_DONTWAIT, {sa_family=AF_INET, sin_port=htons(123), sin_addr=inet_addr("10.34.32.125")}, 16) = 48
811 * ^^^ we sent it from some source port picked by kernel.
812 * time(NULL) = 1259071266
813 * write(2, "ntpd: entering poll 15 secs\n", 28) = 28
814 * poll([{fd=3, events=POLLIN}], 1, 15000) = 1 ([{fd=3, revents=POLLIN}])
815 * recv(3, "yyy", 68, MSG_DONTWAIT) = 48
816 * ^^^ this recv will receive packets to any local port!
818 * Uncomment this and use strace to see it in action:
820 #define PROBE_LOCAL_ADDR /* { len_and_sockaddr lsa; lsa.len = LSA_SIZEOF_SA; getsockname(p->query.fd, &lsa.u.sa, &lsa.len); } */
824 len_and_sockaddr *local_lsa;
826 family = p->p_lsa->u.sa.sa_family;
827 p->p_fd = fd = xsocket_type(&local_lsa, family, SOCK_DGRAM);
828 /* local_lsa has "null" address and port 0 now.
829 * bind() ensures we have a *particular port* selected by kernel
830 * and remembered in p->p_fd, thus later recv(p->p_fd)
831 * receives only packets sent to this port.
834 xbind(fd, &local_lsa->u.sa, local_lsa->len);
836 #if ENABLE_FEATURE_IPV6
837 if (family == AF_INET)
839 setsockopt(fd, IPPROTO_IP, IP_TOS, &const_IPTOS_LOWDELAY, sizeof(const_IPTOS_LOWDELAY));
843 /* Emit message _before_ attempted send. Think of a very short
844 * roundtrip networks: we need to go back to recv loop ASAP,
845 * to reduce delay. Printing messages after send works against that.
847 VERB1 bb_error_msg("sending query to %s", p->p_dotted);
850 * Send out a random 64-bit number as our transmit time. The NTP
851 * server will copy said number into the originate field on the
852 * response that it sends us. This is totally legal per the SNTP spec.
854 * The impact of this is two fold: we no longer send out the current
855 * system time for the world to see (which may aid an attacker), and
856 * it gives us a (not very secure) way of knowing that we're not
857 * getting spoofed by an attacker that can't capture our traffic
858 * but can spoof packets from the NTP server we're communicating with.
860 * Save the real transmit timestamp locally.
862 p->p_xmt_msg.m_xmttime.int_partl = rand();
863 p->p_xmt_msg.m_xmttime.fractionl = rand();
864 p->p_xmttime = gettime1900d();
866 /* Were doing it only if sendto worked, but
867 * loss of sync detection needs reachable_bits updated
868 * even if sending fails *locally*:
869 * "network is unreachable" because cable was pulled?
870 * We still need to declare "unsync" if this condition persists.
872 p->reachable_bits <<= 1;
874 if (do_sendto(p->p_fd, /*from:*/ NULL, /*to:*/ &p->p_lsa->u.sa, /*addrlen:*/ p->p_lsa->len,
875 &p->p_xmt_msg, NTP_MSGSIZE_NOAUTH) == -1
880 * We know that we sent nothing.
881 * We can retry *soon* without fearing
882 * that we are flooding the peer.
884 set_next(p, RETRY_INTERVAL);
888 set_next(p, RESPONSE_INTERVAL);
892 /* Note that there is no provision to prevent several run_scripts
893 * to be started in quick succession. In fact, it happens rather often
894 * if initial syncronization results in a step.
895 * You will see "step" and then "stratum" script runs, sometimes
896 * as close as only 0.002 seconds apart.
897 * Script should be ready to deal with this.
899 static void run_script(const char *action, double offset)
902 char *env1, *env2, *env3, *env4;
904 G.last_script_run = G.cur_time;
909 argv[0] = (char*) G.script_name;
910 argv[1] = (char*) action;
913 VERB1 bb_error_msg("executing '%s %s'", G.script_name, action);
915 env1 = xasprintf("%s=%u", "stratum", G.stratum);
917 env2 = xasprintf("%s=%ld", "freq_drift_ppm", G.kernel_freq_drift);
919 env3 = xasprintf("%s=%u", "poll_interval", 1 << G.poll_exp);
921 env4 = xasprintf("%s=%f", "offset", offset);
923 /* Other items of potential interest: selected peer,
924 * rootdelay, reftime, rootdisp, refid, ntp_status,
925 * last_update_offset, last_update_recv_time, discipline_jitter,
926 * how many peers have reachable_bits = 0?
929 /* Don't want to wait: it may run hwclock --systohc, and that
930 * may take some time (seconds): */
931 /*spawn_and_wait(argv);*/
935 unsetenv("freq_drift_ppm");
936 unsetenv("poll_interval");
945 step_time(double offset)
949 struct timeval tvc, tvn;
950 char buf[sizeof("yyyy-mm-dd hh:mm:ss") + /*paranoia:*/ 4];
953 gettimeofday(&tvc, NULL); /* never fails */
954 dtime = tvc.tv_sec + (1.0e-6 * tvc.tv_usec) + offset;
955 d_to_tv(dtime, &tvn);
956 if (settimeofday(&tvn, NULL) == -1)
957 bb_perror_msg_and_die("settimeofday");
961 strftime_YYYYMMDDHHMMSS(buf, sizeof(buf), &tval);
962 bb_error_msg("current time is %s.%06u", buf, (unsigned)tvc.tv_usec);
965 strftime_YYYYMMDDHHMMSS(buf, sizeof(buf), &tval);
966 bb_error_msg("setting time to %s.%06u (offset %+fs)", buf, (unsigned)tvn.tv_usec, offset);
968 /* Correct various fields which contain time-relative values: */
971 G.cur_time += offset;
972 G.last_update_recv_time += offset;
973 G.last_script_run += offset;
975 /* p->lastpkt_recv_time, p->next_action_time and such: */
976 for (item = G.ntp_peers; item != NULL; item = item->link) {
977 peer_t *pp = (peer_t *) item->data;
978 reset_peer_stats(pp, offset);
979 //bb_error_msg("offset:%+f pp->next_action_time:%f -> %f",
980 // offset, pp->next_action_time, pp->next_action_time + offset);
981 pp->next_action_time += offset;
983 /* We wait for reply from this peer too.
984 * But due to step we are doing, reply's data is no longer
985 * useful (in fact, it'll be bogus). Stop waiting for it.
989 set_next(pp, RETRY_INTERVAL);
994 static void clamp_pollexp_and_set_MAXSTRAT(void)
996 if (G.poll_exp < MINPOLL)
997 G.poll_exp = MINPOLL;
998 if (G.poll_exp > BIGPOLL)
999 G.poll_exp = BIGPOLL;
1000 G.polladj_count = 0;
1001 G.stratum = MAXSTRAT;
1006 * Selection and clustering, and their helpers
1012 double opt_rd; /* optimization */
1015 compare_point_edge(const void *aa, const void *bb)
1017 const point_t *a = aa;
1018 const point_t *b = bb;
1019 if (a->edge < b->edge) {
1022 return (a->edge > b->edge);
1029 compare_survivor_metric(const void *aa, const void *bb)
1031 const survivor_t *a = aa;
1032 const survivor_t *b = bb;
1033 if (a->metric < b->metric) {
1036 return (a->metric > b->metric);
1039 fit(peer_t *p, double rd)
1041 if ((p->reachable_bits & (p->reachable_bits-1)) == 0) {
1042 /* One or zero bits in reachable_bits */
1043 VERB4 bb_error_msg("peer %s unfit for selection: unreachable", p->p_dotted);
1046 #if 0 /* we filter out such packets earlier */
1047 if ((p->lastpkt_status & LI_ALARM) == LI_ALARM
1048 || p->lastpkt_stratum >= MAXSTRAT
1050 VERB4 bb_error_msg("peer %s unfit for selection: bad status/stratum", p->p_dotted);
1054 /* rd is root_distance(p) */
1055 if (rd > MAXDIST + FREQ_TOLERANCE * (1 << G.poll_exp)) {
1056 VERB4 bb_error_msg("peer %s unfit for selection: root distance too high", p->p_dotted);
1060 // /* Do we have a loop? */
1061 // if (p->refid == p->dstaddr || p->refid == s.refid)
1066 select_and_cluster(void)
1071 int size = 3 * G.peer_cnt;
1072 /* for selection algorithm */
1073 point_t point[size];
1074 unsigned num_points, num_candidates;
1076 unsigned num_falsetickers;
1077 /* for cluster algorithm */
1078 survivor_t survivor[size];
1079 unsigned num_survivors;
1085 while (item != NULL) {
1088 p = (peer_t *) item->data;
1089 rd = root_distance(p);
1090 offset = p->filter_offset;
1096 VERB5 bb_error_msg("interval: [%f %f %f] %s",
1102 point[num_points].p = p;
1103 point[num_points].type = -1;
1104 point[num_points].edge = offset - rd;
1105 point[num_points].opt_rd = rd;
1107 point[num_points].p = p;
1108 point[num_points].type = 0;
1109 point[num_points].edge = offset;
1110 point[num_points].opt_rd = rd;
1112 point[num_points].p = p;
1113 point[num_points].type = 1;
1114 point[num_points].edge = offset + rd;
1115 point[num_points].opt_rd = rd;
1119 num_candidates = num_points / 3;
1120 if (num_candidates == 0) {
1121 VERB3 bb_error_msg("no valid datapoints%s", ", no peer selected");
1124 //TODO: sorting does not seem to be done in reference code
1125 qsort(point, num_points, sizeof(point[0]), compare_point_edge);
1127 /* Start with the assumption that there are no falsetickers.
1128 * Attempt to find a nonempty intersection interval containing
1129 * the midpoints of all truechimers.
1130 * If a nonempty interval cannot be found, increase the number
1131 * of assumed falsetickers by one and try again.
1132 * If a nonempty interval is found and the number of falsetickers
1133 * is less than the number of truechimers, a majority has been found
1134 * and the midpoint of each truechimer represents
1135 * the candidates available to the cluster algorithm.
1137 num_falsetickers = 0;
1140 unsigned num_midpoints = 0;
1145 for (i = 0; i < num_points; i++) {
1147 * if (point[i].type == -1) c++;
1148 * if (point[i].type == 1) c--;
1149 * and it's simpler to do it this way:
1152 if (c >= num_candidates - num_falsetickers) {
1153 /* If it was c++ and it got big enough... */
1154 low = point[i].edge;
1157 if (point[i].type == 0)
1161 for (i = num_points-1; i >= 0; i--) {
1163 if (c >= num_candidates - num_falsetickers) {
1164 high = point[i].edge;
1167 if (point[i].type == 0)
1170 /* If the number of midpoints is greater than the number
1171 * of allowed falsetickers, the intersection contains at
1172 * least one truechimer with no midpoint - bad.
1173 * Also, interval should be nonempty.
1175 if (num_midpoints <= num_falsetickers && low < high)
1178 if (num_falsetickers * 2 >= num_candidates) {
1179 VERB3 bb_error_msg("falsetickers:%d, candidates:%d%s",
1180 num_falsetickers, num_candidates,
1181 ", no peer selected");
1185 VERB4 bb_error_msg("selected interval: [%f, %f]; candidates:%d falsetickers:%d",
1186 low, high, num_candidates, num_falsetickers);
1190 /* Construct a list of survivors (p, metric)
1191 * from the chime list, where metric is dominated
1192 * first by stratum and then by root distance.
1193 * All other things being equal, this is the order of preference.
1196 for (i = 0; i < num_points; i++) {
1197 if (point[i].edge < low || point[i].edge > high)
1200 survivor[num_survivors].p = p;
1201 /* x.opt_rd == root_distance(p); */
1202 survivor[num_survivors].metric = MAXDIST * p->lastpkt_stratum + point[i].opt_rd;
1203 VERB5 bb_error_msg("survivor[%d] metric:%f peer:%s",
1204 num_survivors, survivor[num_survivors].metric, p->p_dotted);
1207 /* There must be at least MIN_SELECTED survivors to satisfy the
1208 * correctness assertions. Ordinarily, the Byzantine criteria
1209 * require four survivors, but for the demonstration here, one
1212 if (num_survivors < MIN_SELECTED) {
1213 VERB3 bb_error_msg("survivors:%d%s",
1215 ", no peer selected");
1219 //looks like this is ONLY used by the fact that later we pick survivor[0].
1220 //we can avoid sorting then, just find the minimum once!
1221 qsort(survivor, num_survivors, sizeof(survivor[0]), compare_survivor_metric);
1223 /* For each association p in turn, calculate the selection
1224 * jitter p->sjitter as the square root of the sum of squares
1225 * (p->offset - q->offset) over all q associations. The idea is
1226 * to repeatedly discard the survivor with maximum selection
1227 * jitter until a termination condition is met.
1230 unsigned max_idx = max_idx;
1231 double max_selection_jitter = max_selection_jitter;
1232 double min_jitter = min_jitter;
1234 if (num_survivors <= MIN_CLUSTERED) {
1235 VERB4 bb_error_msg("num_survivors %d <= %d, not discarding more",
1236 num_survivors, MIN_CLUSTERED);
1240 /* To make sure a few survivors are left
1241 * for the clustering algorithm to chew on,
1242 * we stop if the number of survivors
1243 * is less than or equal to MIN_CLUSTERED (3).
1245 for (i = 0; i < num_survivors; i++) {
1246 double selection_jitter_sq;
1249 if (i == 0 || p->filter_jitter < min_jitter)
1250 min_jitter = p->filter_jitter;
1252 selection_jitter_sq = 0;
1253 for (j = 0; j < num_survivors; j++) {
1254 peer_t *q = survivor[j].p;
1255 selection_jitter_sq += SQUARE(p->filter_offset - q->filter_offset);
1257 if (i == 0 || selection_jitter_sq > max_selection_jitter) {
1258 max_selection_jitter = selection_jitter_sq;
1261 VERB6 bb_error_msg("survivor %d selection_jitter^2:%f",
1262 i, selection_jitter_sq);
1264 max_selection_jitter = SQRT(max_selection_jitter / num_survivors);
1265 VERB5 bb_error_msg("max_selection_jitter (at %d):%f min_jitter:%f",
1266 max_idx, max_selection_jitter, min_jitter);
1268 /* If the maximum selection jitter is less than the
1269 * minimum peer jitter, then tossing out more survivors
1270 * will not lower the minimum peer jitter, so we might
1273 if (max_selection_jitter < min_jitter) {
1274 VERB4 bb_error_msg("max_selection_jitter:%f < min_jitter:%f, num_survivors:%d, not discarding more",
1275 max_selection_jitter, min_jitter, num_survivors);
1279 /* Delete survivor[max_idx] from the list
1280 * and go around again.
1282 VERB6 bb_error_msg("dropping survivor %d", max_idx);
1284 while (max_idx < num_survivors) {
1285 survivor[max_idx] = survivor[max_idx + 1];
1291 /* Combine the offsets of the clustering algorithm survivors
1292 * using a weighted average with weight determined by the root
1293 * distance. Compute the selection jitter as the weighted RMS
1294 * difference between the first survivor and the remaining
1295 * survivors. In some cases the inherent clock jitter can be
1296 * reduced by not using this algorithm, especially when frequent
1297 * clockhopping is involved. bbox: thus we don't do it.
1301 for (i = 0; i < num_survivors; i++) {
1303 x = root_distance(p);
1305 z += p->filter_offset / x;
1306 w += SQUARE(p->filter_offset - survivor[0].p->filter_offset) / x;
1308 //G.cluster_offset = z / y;
1309 //G.cluster_jitter = SQRT(w / y);
1312 /* Pick the best clock. If the old system peer is on the list
1313 * and at the same stratum as the first survivor on the list,
1314 * then don't do a clock hop. Otherwise, select the first
1315 * survivor on the list as the new system peer.
1318 if (G.last_update_peer
1319 && G.last_update_peer->lastpkt_stratum <= p->lastpkt_stratum
1321 /* Starting from 1 is ok here */
1322 for (i = 1; i < num_survivors; i++) {
1323 if (G.last_update_peer == survivor[i].p) {
1324 VERB5 bb_error_msg("keeping old synced peer");
1325 p = G.last_update_peer;
1330 G.last_update_peer = p;
1332 VERB4 bb_error_msg("selected peer %s filter_offset:%+f age:%f",
1335 G.cur_time - p->lastpkt_recv_time
1342 * Local clock discipline and its helpers
1345 set_new_values(int disc_state, double offset, double recv_time)
1347 /* Enter new state and set state variables. Note we use the time
1348 * of the last clock filter sample, which must be earlier than
1351 VERB4 bb_error_msg("disc_state=%d last update offset=%f recv_time=%f",
1352 disc_state, offset, recv_time);
1353 G.discipline_state = disc_state;
1354 G.last_update_offset = offset;
1355 G.last_update_recv_time = recv_time;
1357 /* Return: -1: decrease poll interval, 0: leave as is, 1: increase */
1359 update_local_clock(peer_t *p)
1363 /* Note: can use G.cluster_offset instead: */
1364 double offset = p->filter_offset;
1365 double recv_time = p->lastpkt_recv_time;
1367 #if !USING_KERNEL_PLL_LOOP
1370 #if !USING_KERNEL_PLL_LOOP || USING_INITIAL_FREQ_ESTIMATION
1371 double since_last_update;
1373 double etemp, dtemp;
1375 abs_offset = fabs(offset);
1378 /* If needed, -S script can do it by looking at $offset
1379 * env var and killing parent */
1380 /* If the offset is too large, give up and go home */
1381 if (abs_offset > PANIC_THRESHOLD) {
1382 bb_error_msg_and_die("offset %f far too big, exiting", offset);
1386 /* If this is an old update, for instance as the result
1387 * of a system peer change, avoid it. We never use
1388 * an old sample or the same sample twice.
1390 if (recv_time <= G.last_update_recv_time) {
1391 VERB3 bb_error_msg("update from %s: same or older datapoint, not using it",
1393 return 0; /* "leave poll interval as is" */
1396 /* Clock state machine transition function. This is where the
1397 * action is and defines how the system reacts to large time
1398 * and frequency errors.
1400 #if !USING_KERNEL_PLL_LOOP || USING_INITIAL_FREQ_ESTIMATION
1401 since_last_update = recv_time - G.reftime;
1403 #if !USING_KERNEL_PLL_LOOP
1406 #if USING_INITIAL_FREQ_ESTIMATION
1407 if (G.discipline_state == STATE_FREQ) {
1408 /* Ignore updates until the stepout threshold */
1409 if (since_last_update < WATCH_THRESHOLD) {
1410 VERB4 bb_error_msg("measuring drift, datapoint ignored, %f sec remains",
1411 WATCH_THRESHOLD - since_last_update);
1412 return 0; /* "leave poll interval as is" */
1414 # if !USING_KERNEL_PLL_LOOP
1415 freq_drift = (offset - G.last_update_offset) / since_last_update;
1420 /* There are two main regimes: when the
1421 * offset exceeds the step threshold and when it does not.
1423 if (abs_offset > STEP_THRESHOLD) {
1427 // This "spike state" seems to be useless, peer selection already drops
1428 // occassional "bad" datapoints. If we are here, there were _many_
1429 // large offsets. When a few first large offsets are seen,
1430 // we end up in "no valid datapoints, no peer selected" state.
1431 // Only when enough of them are seen (which means it's not a fluke),
1432 // we end up here. Looks like _our_ clock is off.
1433 switch (G.discipline_state) {
1435 /* The first outlyer: ignore it, switch to SPIK state */
1436 VERB3 bb_error_msg("update from %s: offset:%+f, spike%s",
1437 p->p_dotted, offset,
1439 G.discipline_state = STATE_SPIK;
1440 return -1; /* "decrease poll interval" */
1443 /* Ignore succeeding outlyers until either an inlyer
1444 * is found or the stepout threshold is exceeded.
1446 remains = WATCH_THRESHOLD - since_last_update;
1448 VERB3 bb_error_msg("update from %s: offset:%+f, spike%s",
1449 p->p_dotted, offset,
1450 ", datapoint ignored");
1451 return -1; /* "decrease poll interval" */
1453 /* fall through: we need to step */
1457 /* Step the time and clamp down the poll interval.
1459 * In NSET state an initial frequency correction is
1460 * not available, usually because the frequency file has
1461 * not yet been written. Since the time is outside the
1462 * capture range, the clock is stepped. The frequency
1463 * will be set directly following the stepout interval.
1465 * In FSET state the initial frequency has been set
1466 * from the frequency file. Since the time is outside
1467 * the capture range, the clock is stepped immediately,
1468 * rather than after the stepout interval. Guys get
1469 * nervous if it takes 17 minutes to set the clock for
1472 * In SPIK state the stepout threshold has expired and
1473 * the phase is still above the step threshold. Note
1474 * that a single spike greater than the step threshold
1475 * is always suppressed, even at the longer poll
1478 VERB4 bb_error_msg("stepping time by %+f; poll_exp=MINPOLL", offset);
1480 if (option_mask32 & OPT_q) {
1481 /* We were only asked to set time once. Done. */
1485 clamp_pollexp_and_set_MAXSTRAT();
1487 run_script("step", offset);
1489 recv_time += offset;
1491 #if USING_INITIAL_FREQ_ESTIMATION
1492 if (G.discipline_state == STATE_NSET) {
1493 set_new_values(STATE_FREQ, /*offset:*/ 0, recv_time);
1494 return 1; /* "ok to increase poll interval" */
1497 abs_offset = offset = 0;
1498 set_new_values(STATE_SYNC, offset, recv_time);
1500 } else { /* abs_offset <= STEP_THRESHOLD */
1502 /* The ratio is calculated before jitter is updated to make
1503 * poll adjust code more sensitive to large offsets.
1505 G.offset_to_jitter_ratio = abs_offset / G.discipline_jitter;
1507 /* Compute the clock jitter as the RMS of exponentially
1508 * weighted offset differences. Used by the poll adjust code.
1510 etemp = SQUARE(G.discipline_jitter);
1511 dtemp = SQUARE(offset - G.last_update_offset);
1512 G.discipline_jitter = SQRT(etemp + (dtemp - etemp) / AVG);
1513 if (G.discipline_jitter < G_precision_sec)
1514 G.discipline_jitter = G_precision_sec;
1516 switch (G.discipline_state) {
1518 if (option_mask32 & OPT_q) {
1519 /* We were only asked to set time once.
1520 * The clock is precise enough, no need to step.
1524 #if USING_INITIAL_FREQ_ESTIMATION
1525 /* This is the first update received and the frequency
1526 * has not been initialized. The first thing to do
1527 * is directly measure the oscillator frequency.
1529 set_new_values(STATE_FREQ, offset, recv_time);
1531 set_new_values(STATE_SYNC, offset, recv_time);
1533 VERB4 bb_error_msg("transitioning to FREQ, datapoint ignored");
1534 return 0; /* "leave poll interval as is" */
1536 #if 0 /* this is dead code for now */
1538 /* This is the first update and the frequency
1539 * has been initialized. Adjust the phase, but
1540 * don't adjust the frequency until the next update.
1542 set_new_values(STATE_SYNC, offset, recv_time);
1543 /* freq_drift remains 0 */
1547 #if USING_INITIAL_FREQ_ESTIMATION
1549 /* since_last_update >= WATCH_THRESHOLD, we waited enough.
1550 * Correct the phase and frequency and switch to SYNC state.
1551 * freq_drift was already estimated (see code above)
1553 set_new_values(STATE_SYNC, offset, recv_time);
1558 #if !USING_KERNEL_PLL_LOOP
1559 /* Compute freq_drift due to PLL and FLL contributions.
1561 * The FLL and PLL frequency gain constants
1562 * depend on the poll interval and Allan
1563 * intercept. The FLL is not used below one-half
1564 * the Allan intercept. Above that the loop gain
1565 * increases in steps to 1 / AVG.
1567 if ((1 << G.poll_exp) > ALLAN / 2) {
1568 etemp = FLL - G.poll_exp;
1571 freq_drift += (offset - G.last_update_offset) / (MAXD(since_last_update, ALLAN) * etemp);
1573 /* For the PLL the integration interval
1574 * (numerator) is the minimum of the update
1575 * interval and poll interval. This allows
1576 * oversampling, but not undersampling.
1578 etemp = MIND(since_last_update, (1 << G.poll_exp));
1579 dtemp = (4 * PLL) << G.poll_exp;
1580 freq_drift += offset * etemp / SQUARE(dtemp);
1582 set_new_values(STATE_SYNC, offset, recv_time);
1585 if (G.stratum != p->lastpkt_stratum + 1) {
1586 G.stratum = p->lastpkt_stratum + 1;
1587 run_script("stratum", offset);
1591 G.reftime = G.cur_time;
1592 G.ntp_status = p->lastpkt_status;
1593 G.refid = p->lastpkt_refid;
1594 G.rootdelay = p->lastpkt_rootdelay + p->lastpkt_delay;
1595 dtemp = p->filter_jitter; // SQRT(SQUARE(p->filter_jitter) + SQUARE(G.cluster_jitter));
1596 dtemp += MAXD(p->filter_dispersion + FREQ_TOLERANCE * (G.cur_time - p->lastpkt_recv_time) + abs_offset, MINDISP);
1597 G.rootdisp = p->lastpkt_rootdisp + dtemp;
1598 VERB4 bb_error_msg("updating leap/refid/reftime/rootdisp from peer %s", p->p_dotted);
1600 /* We are in STATE_SYNC now, but did not do adjtimex yet.
1601 * (Any other state does not reach this, they all return earlier)
1602 * By this time, freq_drift and offset are set
1603 * to values suitable for adjtimex.
1605 #if !USING_KERNEL_PLL_LOOP
1606 /* Calculate the new frequency drift and frequency stability (wander).
1607 * Compute the clock wander as the RMS of exponentially weighted
1608 * frequency differences. This is not used directly, but can,
1609 * along with the jitter, be a highly useful monitoring and
1612 dtemp = G.discipline_freq_drift + freq_drift;
1613 G.discipline_freq_drift = MAXD(MIND(MAXDRIFT, dtemp), -MAXDRIFT);
1614 etemp = SQUARE(G.discipline_wander);
1615 dtemp = SQUARE(dtemp);
1616 G.discipline_wander = SQRT(etemp + (dtemp - etemp) / AVG);
1618 VERB4 bb_error_msg("discipline freq_drift=%.9f(int:%ld corr:%e) wander=%f",
1619 G.discipline_freq_drift,
1620 (long)(G.discipline_freq_drift * 65536e6),
1622 G.discipline_wander);
1625 memset(&tmx, 0, sizeof(tmx));
1626 if (adjtimex(&tmx) < 0)
1627 bb_perror_msg_and_die("adjtimex");
1628 bb_error_msg("p adjtimex freq:%ld offset:%+ld status:0x%x tc:%ld",
1629 tmx.freq, tmx.offset, tmx.status, tmx.constant);
1632 memset(&tmx, 0, sizeof(tmx));
1634 //doesn't work, offset remains 0 (!) in kernel:
1635 //ntpd: set adjtimex freq:1786097 tmx.offset:77487
1636 //ntpd: prev adjtimex freq:1786097 tmx.offset:0
1637 //ntpd: cur adjtimex freq:1786097 tmx.offset:0
1638 tmx.modes = ADJ_FREQUENCY | ADJ_OFFSET;
1639 /* 65536 is one ppm */
1640 tmx.freq = G.discipline_freq_drift * 65536e6;
1642 tmx.modes = ADJ_OFFSET | ADJ_STATUS | ADJ_TIMECONST;// | ADJ_MAXERROR | ADJ_ESTERROR;
1643 tmx.offset = (offset * 1000000); /* usec */
1644 tmx.status = STA_PLL;
1645 if (G.ntp_status & LI_PLUSSEC)
1646 tmx.status |= STA_INS;
1647 if (G.ntp_status & LI_MINUSSEC)
1648 tmx.status |= STA_DEL;
1650 tmx.constant = (int)G.poll_exp - 4 > 0 ? (int)G.poll_exp - 4 : 0;
1652 * The below if statement should be unnecessary, but...
1653 * It looks like Linux kernel's PLL is far too gentle in changing
1654 * tmx.freq in response to clock offset. Offset keeps growing
1655 * and eventually we fall back to smaller poll intervals.
1656 * We can make correction more agressive (about x2) by supplying
1657 * PLL time constant which is one less than the real one.
1658 * To be on a safe side, let's do it only if offset is significantly
1659 * larger than jitter.
1661 if (tmx.constant > 0 && G.offset_to_jitter_ratio >= TIMECONST_HACK_GATE)
1664 //tmx.esterror = (uint32_t)(clock_jitter * 1e6);
1665 //tmx.maxerror = (uint32_t)((sys_rootdelay / 2 + sys_rootdisp) * 1e6);
1666 rc = adjtimex(&tmx);
1668 bb_perror_msg_and_die("adjtimex");
1669 /* NB: here kernel returns constant == G.poll_exp, not == G.poll_exp - 4.
1670 * Not sure why. Perhaps it is normal.
1672 VERB4 bb_error_msg("adjtimex:%d freq:%ld offset:%+ld status:0x%x",
1673 rc, tmx.freq, tmx.offset, tmx.status);
1674 G.kernel_freq_drift = tmx.freq / 65536;
1675 VERB2 bb_error_msg("update from:%s offset:%+f jitter:%f clock drift:%+.3fppm tc:%d",
1676 p->p_dotted, offset, G.discipline_jitter, (double)tmx.freq / 65536, (int)tmx.constant);
1678 return 1; /* "ok to increase poll interval" */
1683 * We've got a new reply packet from a peer, process it
1687 poll_interval(int upper_bound)
1689 unsigned interval, r, mask;
1690 interval = 1 << G.poll_exp;
1691 if (interval > upper_bound)
1692 interval = upper_bound;
1693 mask = ((interval-1) >> 4) | 1;
1695 interval += r & mask; /* ~ random(0..1) * interval/16 */
1696 VERB4 bb_error_msg("chose poll interval:%u (poll_exp:%d)", interval, G.poll_exp);
1700 adjust_poll(int count)
1702 G.polladj_count += count;
1703 if (G.polladj_count > POLLADJ_LIMIT) {
1704 G.polladj_count = 0;
1705 if (G.poll_exp < MAXPOLL) {
1707 VERB4 bb_error_msg("polladj: discipline_jitter:%f ++poll_exp=%d",
1708 G.discipline_jitter, G.poll_exp);
1710 } else if (G.polladj_count < -POLLADJ_LIMIT || (count < 0 && G.poll_exp > BIGPOLL)) {
1711 G.polladj_count = 0;
1712 if (G.poll_exp > MINPOLL) {
1716 /* Correct p->next_action_time in each peer
1717 * which waits for sending, so that they send earlier.
1718 * Old pp->next_action_time are on the order
1719 * of t + (1 << old_poll_exp) + small_random,
1720 * we simply need to subtract ~half of that.
1722 for (item = G.ntp_peers; item != NULL; item = item->link) {
1723 peer_t *pp = (peer_t *) item->data;
1725 pp->next_action_time -= (1 << G.poll_exp);
1727 VERB4 bb_error_msg("polladj: discipline_jitter:%f --poll_exp=%d",
1728 G.discipline_jitter, G.poll_exp);
1731 VERB4 bb_error_msg("polladj: count:%d", G.polladj_count);
1734 static NOINLINE void
1735 recv_and_process_peer_pkt(peer_t *p)
1740 double T1, T2, T3, T4;
1742 double prev_delay, delay;
1744 datapoint_t *datapoint;
1749 /* We can recvfrom here and check from.IP, but some multihomed
1750 * ntp servers reply from their *other IP*.
1751 * TODO: maybe we should check at least what we can: from.port == 123?
1754 size = recv(p->p_fd, &msg, sizeof(msg), MSG_DONTWAIT);
1759 if (errno == EAGAIN)
1760 /* There was no packet after all
1761 * (poll() returning POLLIN for a fd
1762 * is not a ironclad guarantee that data is there)
1766 * If you need a different handling for a specific
1767 * errno, always explain it in comment.
1769 bb_perror_msg_and_die("recv(%s) error", p->p_dotted);
1772 if (size != NTP_MSGSIZE_NOAUTH && size != NTP_MSGSIZE) {
1773 bb_error_msg("malformed packet received from %s", p->p_dotted);
1777 if (msg.m_orgtime.int_partl != p->p_xmt_msg.m_xmttime.int_partl
1778 || msg.m_orgtime.fractionl != p->p_xmt_msg.m_xmttime.fractionl
1780 /* Somebody else's packet */
1784 /* We do not expect any more packets from this peer for now.
1785 * Closing the socket informs kernel about it.
1786 * We open a new socket when we send a new query.
1791 if ((msg.m_status & LI_ALARM) == LI_ALARM
1792 || msg.m_stratum == 0
1793 || msg.m_stratum > NTP_MAXSTRATUM
1795 bb_error_msg("reply from %s: peer is unsynced", p->p_dotted);
1797 * Stratum 0 responses may have commands in 32-bit m_refid field:
1798 * "DENY", "RSTR" - peer does not like us at all,
1799 * "RATE" - peer is overloaded, reduce polling freq.
1800 * If poll interval is small, increase it.
1802 if (G.poll_exp < BIGPOLL)
1803 goto increase_interval;
1804 goto pick_normal_interval;
1807 // /* Verify valid root distance */
1808 // if (msg.m_rootdelay / 2 + msg.m_rootdisp >= MAXDISP || p->lastpkt_reftime > msg.m_xmt)
1809 // return; /* invalid header values */
1812 * From RFC 2030 (with a correction to the delay math):
1814 * Timestamp Name ID When Generated
1815 * ------------------------------------------------------------
1816 * Originate Timestamp T1 time request sent by client
1817 * Receive Timestamp T2 time request received by server
1818 * Transmit Timestamp T3 time reply sent by server
1819 * Destination Timestamp T4 time reply received by client
1821 * The roundtrip delay and local clock offset are defined as
1823 * delay = (T4 - T1) - (T3 - T2); offset = ((T2 - T1) + (T3 - T4)) / 2
1826 T2 = lfp_to_d(msg.m_rectime);
1827 T3 = lfp_to_d(msg.m_xmttime);
1830 /* The delay calculation is a special case. In cases where the
1831 * server and client clocks are running at different rates and
1832 * with very fast networks, the delay can appear negative. In
1833 * order to avoid violating the Principle of Least Astonishment,
1834 * the delay is clamped not less than the system precision.
1836 delay = (T4 - T1) - (T3 - T2);
1837 if (delay < G_precision_sec)
1838 delay = G_precision_sec;
1840 * If this packet's delay is much bigger than the last one,
1841 * it's better to just ignore it than use its much less precise value.
1843 prev_delay = p->p_raw_delay;
1844 p->p_raw_delay = delay;
1845 if (p->reachable_bits && delay > prev_delay * BAD_DELAY_GROWTH) {
1846 bb_error_msg("reply from %s: delay %f is too high, ignoring", p->p_dotted, delay);
1847 goto pick_normal_interval;
1850 p->lastpkt_delay = delay;
1851 p->lastpkt_recv_time = T4;
1852 VERB6 bb_error_msg("%s->lastpkt_recv_time=%f", p->p_dotted, p->lastpkt_recv_time);
1853 p->lastpkt_status = msg.m_status;
1854 p->lastpkt_stratum = msg.m_stratum;
1855 p->lastpkt_rootdelay = sfp_to_d(msg.m_rootdelay);
1856 p->lastpkt_rootdisp = sfp_to_d(msg.m_rootdisp);
1857 p->lastpkt_refid = msg.m_refid;
1859 p->datapoint_idx = p->reachable_bits ? (p->datapoint_idx + 1) % NUM_DATAPOINTS : 0;
1860 datapoint = &p->filter_datapoint[p->datapoint_idx];
1861 datapoint->d_recv_time = T4;
1862 datapoint->d_offset = offset = ((T2 - T1) + (T3 - T4)) / 2;
1863 datapoint->d_dispersion = LOG2D(msg.m_precision_exp) + G_precision_sec;
1864 if (!p->reachable_bits) {
1865 /* 1st datapoint ever - replicate offset in every element */
1867 for (i = 0; i < NUM_DATAPOINTS; i++) {
1868 p->filter_datapoint[i].d_offset = offset;
1872 p->reachable_bits |= 1;
1873 if ((MAX_VERBOSE && G.verbose) || (option_mask32 & OPT_w)) {
1874 bb_error_msg("reply from %s: offset:%+f delay:%f status:0x%02x strat:%d refid:0x%08x rootdelay:%f reach:0x%02x",
1881 p->lastpkt_rootdelay,
1883 /* not shown: m_ppoll, m_precision_exp, m_rootdisp,
1884 * m_reftime, m_orgtime, m_rectime, m_xmttime
1889 /* Muck with statictics and update the clock */
1890 filter_datapoints(p);
1891 q = select_and_cluster();
1894 if (!(option_mask32 & OPT_w)) {
1895 rc = update_local_clock(q);
1897 //Disabled this because there is a case where largish offsets
1898 //are unavoidable: if network round-trip delay is, say, ~0.6s,
1899 //error in offset estimation would be ~delay/2 ~= 0.3s.
1900 //Thus, offsets will be usually in -0.3...0.3s range.
1901 //In this case, this code would keep poll interval small,
1902 //but it won't be helping.
1903 //BIGOFF check below deals with a case of seeing multi-second offsets.
1905 /* If drift is dangerously large, immediately
1906 * drop poll interval one step down.
1908 if (fabs(q->filter_offset) >= POLLDOWN_OFFSET) {
1909 VERB4 bb_error_msg("offset:%+f > POLLDOWN_OFFSET", q->filter_offset);
1910 adjust_poll(-POLLADJ_LIMIT * 3);
1916 /* No peer selected.
1917 * If poll interval is small, increase it.
1919 if (G.poll_exp < BIGPOLL)
1920 goto increase_interval;
1924 /* Adjust the poll interval by comparing the current offset
1925 * with the clock jitter. If the offset is less than
1926 * the clock jitter times a constant, then the averaging interval
1927 * is increased, otherwise it is decreased. A bit of hysteresis
1928 * helps calm the dance. Works best using burst mode.
1930 if (rc > 0 && G.offset_to_jitter_ratio <= POLLADJ_GATE) {
1931 /* was += G.poll_exp but it is a bit
1932 * too optimistic for my taste at high poll_exp's */
1934 adjust_poll(MINPOLL);
1936 adjust_poll(-G.poll_exp * 2);
1940 /* Decide when to send new query for this peer */
1941 pick_normal_interval:
1942 interval = poll_interval(INT_MAX);
1943 if (fabs(offset) >= BIGOFF && interval > BIGOFF_INTERVAL) {
1944 /* If we are synced, offsets are less than STEP_THRESHOLD,
1945 * or at the very least not much larger than it.
1946 * Now we see a largish one.
1947 * Either this peer is feeling bad, or packet got corrupted,
1948 * or _our_ clock is wrong now and _all_ peers will show similar
1949 * largish offsets too.
1950 * I observed this with laptop suspend stopping clock.
1951 * In any case, it makes sense to make next request soonish:
1952 * cases 1 and 2: get a better datapoint,
1953 * case 3: allows to resync faster.
1955 interval = BIGOFF_INTERVAL;
1958 set_next(p, interval);
1961 #if ENABLE_FEATURE_NTPD_SERVER
1962 static NOINLINE void
1963 recv_and_process_client_pkt(void /*int fd*/)
1967 len_and_sockaddr *to;
1968 struct sockaddr *from;
1970 uint8_t query_status;
1971 l_fixedpt_t query_xmttime;
1973 to = get_sock_lsa(G_listen_fd);
1974 from = xzalloc(to->len);
1976 size = recv_from_to(G_listen_fd, &msg, sizeof(msg), MSG_DONTWAIT, from, &to->u.sa, to->len);
1977 if (size != NTP_MSGSIZE_NOAUTH && size != NTP_MSGSIZE) {
1980 if (errno == EAGAIN)
1982 bb_perror_msg_and_die("recv");
1984 addr = xmalloc_sockaddr2dotted_noport(from);
1985 bb_error_msg("malformed packet received from %s: size %u", addr, (int)size);
1990 query_status = msg.m_status;
1991 query_xmttime = msg.m_xmttime;
1993 /* Build a reply packet */
1994 memset(&msg, 0, sizeof(msg));
1995 msg.m_status = G.stratum < MAXSTRAT ? (G.ntp_status & LI_MASK) : LI_ALARM;
1996 msg.m_status |= (query_status & VERSION_MASK);
1997 msg.m_status |= ((query_status & MODE_MASK) == MODE_CLIENT) ?
1998 MODE_SERVER : MODE_SYM_PAS;
1999 msg.m_stratum = G.stratum;
2000 msg.m_ppoll = G.poll_exp;
2001 msg.m_precision_exp = G_precision_exp;
2002 /* this time was obtained between poll() and recv() */
2003 msg.m_rectime = d_to_lfp(G.cur_time);
2004 msg.m_xmttime = d_to_lfp(gettime1900d()); /* this instant */
2005 if (G.peer_cnt == 0) {
2006 /* we have no peers: "stratum 1 server" mode. reftime = our own time */
2007 G.reftime = G.cur_time;
2009 msg.m_reftime = d_to_lfp(G.reftime);
2010 msg.m_orgtime = query_xmttime;
2011 msg.m_rootdelay = d_to_sfp(G.rootdelay);
2012 //simple code does not do this, fix simple code!
2013 msg.m_rootdisp = d_to_sfp(G.rootdisp);
2014 //version = (query_status & VERSION_MASK); /* ... >> VERSION_SHIFT - done below instead */
2015 msg.m_refid = G.refid; // (version > (3 << VERSION_SHIFT)) ? G.refid : G.refid3;
2017 /* We reply from the local address packet was sent to,
2018 * this makes to/from look swapped here: */
2019 do_sendto(G_listen_fd,
2020 /*from:*/ &to->u.sa, /*to:*/ from, /*addrlen:*/ to->len,
2029 /* Upstream ntpd's options:
2031 * -4 Force DNS resolution of host names to the IPv4 namespace.
2032 * -6 Force DNS resolution of host names to the IPv6 namespace.
2033 * -a Require cryptographic authentication for broadcast client,
2034 * multicast client and symmetric passive associations.
2035 * This is the default.
2036 * -A Do not require cryptographic authentication for broadcast client,
2037 * multicast client and symmetric passive associations.
2038 * This is almost never a good idea.
2039 * -b Enable the client to synchronize to broadcast servers.
2041 * Specify the name and path of the configuration file,
2042 * default /etc/ntp.conf
2043 * -d Specify debugging mode. This option may occur more than once,
2044 * with each occurrence indicating greater detail of display.
2046 * Specify debugging level directly.
2048 * Specify the name and path of the frequency file.
2049 * This is the same operation as the "driftfile FILE"
2050 * configuration command.
2051 * -g Normally, ntpd exits with a message to the system log
2052 * if the offset exceeds the panic threshold, which is 1000 s
2053 * by default. This option allows the time to be set to any value
2054 * without restriction; however, this can happen only once.
2055 * If the threshold is exceeded after that, ntpd will exit
2056 * with a message to the system log. This option can be used
2057 * with the -q and -x options. See the tinker command for other options.
2059 * Chroot the server to the directory jaildir. This option also implies
2060 * that the server attempts to drop root privileges at startup
2061 * (otherwise, chroot gives very little additional security).
2062 * You may need to also specify a -u option.
2064 * Specify the name and path of the symmetric key file,
2065 * default /etc/ntp/keys. This is the same operation
2066 * as the "keys FILE" configuration command.
2068 * Specify the name and path of the log file. The default
2069 * is the system log file. This is the same operation as
2070 * the "logfile FILE" configuration command.
2071 * -L Do not listen to virtual IPs. The default is to listen.
2073 * -N To the extent permitted by the operating system,
2074 * run the ntpd at the highest priority.
2076 * Specify the name and path of the file used to record the ntpd
2077 * process ID. This is the same operation as the "pidfile FILE"
2078 * configuration command.
2080 * To the extent permitted by the operating system,
2081 * run the ntpd at the specified priority.
2082 * -q Exit the ntpd just after the first time the clock is set.
2083 * This behavior mimics that of the ntpdate program, which is
2084 * to be retired. The -g and -x options can be used with this option.
2085 * Note: The kernel time discipline is disabled with this option.
2087 * Specify the default propagation delay from the broadcast/multicast
2088 * server to this client. This is necessary only if the delay
2089 * cannot be computed automatically by the protocol.
2091 * Specify the directory path for files created by the statistics
2092 * facility. This is the same operation as the "statsdir DIR"
2093 * configuration command.
2095 * Add a key number to the trusted key list. This option can occur
2098 * Specify a user, and optionally a group, to switch to.
2101 * Add a system variable listed by default.
2102 * -x Normally, the time is slewed if the offset is less than the step
2103 * threshold, which is 128 ms by default, and stepped if above
2104 * the threshold. This option sets the threshold to 600 s, which is
2105 * well within the accuracy window to set the clock manually.
2106 * Note: since the slew rate of typical Unix kernels is limited
2107 * to 0.5 ms/s, each second of adjustment requires an amortization
2108 * interval of 2000 s. Thus, an adjustment as much as 600 s
2109 * will take almost 14 days to complete. This option can be used
2110 * with the -g and -q options. See the tinker command for other options.
2111 * Note: The kernel time discipline is disabled with this option.
2114 /* By doing init in a separate function we decrease stack usage
2117 static NOINLINE void ntp_init(char **argv)
2125 bb_error_msg_and_die(bb_msg_you_must_be_root);
2127 /* Set some globals */
2128 G.discipline_jitter = G_precision_sec;
2129 G.stratum = MAXSTRAT;
2131 G.poll_exp = BURSTPOLL; /* speeds up initial sync */
2132 G.last_script_run = G.reftime = G.last_update_recv_time = gettime1900d(); /* sets G.cur_time too */
2136 opt_complementary = "dd:p::wn" /* -d: counter; -p: list; -w implies -n */
2137 IF_FEATURE_NTPD_SERVER(":Il"); /* -I implies -l */
2138 opts = getopt32(argv,
2140 "wp:S:"IF_FEATURE_NTPD_SERVER("l") /* NOT compat */
2141 IF_FEATURE_NTPD_SERVER("I:") /* compat */
2143 "46aAbgL", /* compat, ignored */
2144 &peers,&G.script_name,
2145 #if ENABLE_FEATURE_NTPD_SERVER
2150 // if (opts & OPT_x) /* disable stepping, only slew is allowed */
2151 // G.time_was_stepped = 1;
2154 add_peers(llist_pop(&peers));
2156 #if ENABLE_FEATURE_NTPD_CONF
2161 parser = config_open("/etc/ntp.conf");
2162 while (config_read(parser, token, 3, 1, "# \t", PARSE_NORMAL)) {
2163 if (strcmp(token[0], "server") == 0 && token[1]) {
2164 add_peers(token[1]);
2167 bb_error_msg("skipping %s:%u: unimplemented command '%s'",
2168 "/etc/ntp.conf", parser->lineno, token[0]
2171 config_close(parser);
2174 if (G.peer_cnt == 0) {
2175 if (!(opts & OPT_l))
2177 /* -l but no peers: "stratum 1 server" mode */
2180 #if ENABLE_FEATURE_NTPD_SERVER
2183 G_listen_fd = create_and_bind_dgram_or_die(NULL, 123);
2185 if (setsockopt_bindtodevice(G_listen_fd, G.if_name))
2188 socket_want_pktinfo(G_listen_fd);
2189 setsockopt(G_listen_fd, IPPROTO_IP, IP_TOS, &const_IPTOS_LOWDELAY, sizeof(const_IPTOS_LOWDELAY));
2192 if (!(opts & OPT_n)) {
2193 bb_daemonize_or_rexec(DAEMON_DEVNULL_STDIO, argv);
2194 logmode = LOGMODE_NONE;
2196 /* I hesitate to set -20 prio. -15 should be high enough for timekeeping */
2198 setpriority(PRIO_PROCESS, 0, -15);
2200 /* If network is up, syncronization occurs in ~10 seconds.
2201 * We give "ntpd -q" 10 seconds to get first reply,
2202 * then another 50 seconds to finish syncing.
2204 * I tested ntpd 4.2.6p1 and apparently it never exits
2205 * (will try forever), but it does not feel right.
2206 * The goal of -q is to act like ntpdate: set time
2207 * after a reasonably small period of polling, or fail.
2210 option_mask32 |= OPT_qq;
2227 int ntpd_main(int argc UNUSED_PARAM, char **argv) MAIN_EXTERNALLY_VISIBLE;
2228 int ntpd_main(int argc UNUSED_PARAM, char **argv)
2236 memset(&G, 0, sizeof(G));
2237 SET_PTR_TO_GLOBALS(&G);
2241 /* If ENABLE_FEATURE_NTPD_SERVER, + 1 for listen_fd: */
2242 cnt = G.peer_cnt + ENABLE_FEATURE_NTPD_SERVER;
2243 idx2peer = xzalloc(sizeof(idx2peer[0]) * cnt);
2244 pfd = xzalloc(sizeof(pfd[0]) * cnt);
2246 /* Countdown: we never sync before we sent INITIAL_SAMPLES+1
2247 * packets to each peer.
2248 * NB: if some peer is not responding, we may end up sending
2249 * fewer packets to it and more to other peers.
2250 * NB2: sync usually happens using INITIAL_SAMPLES packets,
2251 * since last reply does not come back instantaneously.
2253 cnt = G.peer_cnt * (INITIAL_SAMPLES + 1);
2255 write_pidfile(CONFIG_PID_FILE_PATH "/ntpd.pid");
2257 while (!bb_got_signal) {
2263 /* Nothing between here and poll() blocks for any significant time */
2265 nextaction = G.cur_time + 3600;
2268 #if ENABLE_FEATURE_NTPD_SERVER
2269 if (G_listen_fd != -1) {
2270 pfd[0].fd = G_listen_fd;
2271 pfd[0].events = POLLIN;
2275 /* Pass over peer list, send requests, time out on receives */
2276 for (item = G.ntp_peers; item != NULL; item = item->link) {
2277 peer_t *p = (peer_t *) item->data;
2279 if (p->next_action_time <= G.cur_time) {
2280 if (p->p_fd == -1) {
2281 /* Time to send new req */
2283 VERB4 bb_error_msg("disabling burst mode");
2284 G.polladj_count = 0;
2285 G.poll_exp = MINPOLL;
2287 send_query_to_peer(p);
2289 /* Timed out waiting for reply */
2292 /* If poll interval is small, increase it */
2293 if (G.poll_exp < BIGPOLL)
2294 adjust_poll(MINPOLL);
2295 timeout = poll_interval(NOREPLY_INTERVAL);
2296 bb_error_msg("timed out waiting for %s, reach 0x%02x, next query in %us",
2297 p->p_dotted, p->reachable_bits, timeout);
2298 set_next(p, timeout);
2302 if (p->next_action_time < nextaction)
2303 nextaction = p->next_action_time;
2306 /* Wait for reply from this peer */
2307 pfd[i].fd = p->p_fd;
2308 pfd[i].events = POLLIN;
2314 timeout = nextaction - G.cur_time;
2317 timeout++; /* (nextaction - G.cur_time) rounds down, compensating */
2319 /* Here we may block */
2321 if (i > (ENABLE_FEATURE_NTPD_SERVER && G_listen_fd != -1)) {
2322 /* We wait for at least one reply.
2323 * Poll for it, without wasting time for message.
2324 * Since replies often come under 1 second, this also
2325 * reduces clutter in logs.
2327 nfds = poll(pfd, i, 1000);
2333 bb_error_msg("poll:%us sockets:%u interval:%us", timeout, i, 1 << G.poll_exp);
2335 nfds = poll(pfd, i, timeout * 1000);
2337 gettime1900d(); /* sets G.cur_time */
2339 if (!bb_got_signal /* poll wasn't interrupted by a signal */
2340 && G.cur_time - G.last_script_run > 11*60
2342 /* Useful for updating battery-backed RTC and such */
2343 run_script("periodic", G.last_update_offset);
2344 gettime1900d(); /* sets G.cur_time */
2349 /* Process any received packets */
2351 #if ENABLE_FEATURE_NTPD_SERVER
2352 if (G.listen_fd != -1) {
2353 if (pfd[0].revents /* & (POLLIN|POLLERR)*/) {
2355 recv_and_process_client_pkt(/*G.listen_fd*/);
2356 gettime1900d(); /* sets G.cur_time */
2361 for (; nfds != 0 && j < i; j++) {
2362 if (pfd[j].revents /* & (POLLIN|POLLERR)*/) {
2364 * At init, alarm was set to 10 sec.
2365 * Now we did get a reply.
2366 * Increase timeout to 50 seconds to finish syncing.
2368 if (option_mask32 & OPT_qq) {
2369 option_mask32 &= ~OPT_qq;
2373 recv_and_process_peer_pkt(idx2peer[j]);
2374 gettime1900d(); /* sets G.cur_time */
2379 if (G.ntp_peers && G.stratum != MAXSTRAT) {
2380 for (item = G.ntp_peers; item != NULL; item = item->link) {
2381 peer_t *p = (peer_t *) item->data;
2382 if (p->reachable_bits)
2383 goto have_reachable_peer;
2385 /* No peer responded for last 8 packets, panic */
2386 clamp_pollexp_and_set_MAXSTRAT();
2387 run_script("unsync", 0.0);
2388 have_reachable_peer: ;
2390 } /* while (!bb_got_signal) */
2392 remove_pidfile(CONFIG_PID_FILE_PATH "/ntpd.pid");
2393 kill_myself_with_sig(bb_got_signal);
2401 /*** openntpd-4.6 uses only adjtime, not adjtimex ***/
2403 /*** ntp-4.2.6/ntpd/ntp_loopfilter.c - adjtimex usage ***/
2407 direct_freq(double fp_offset)
2411 * If the kernel is enabled, we need the residual offset to
2412 * calculate the frequency correction.
2414 if (pll_control && kern_enable) {
2415 memset(&ntv, 0, sizeof(ntv));
2418 clock_offset = ntv.offset / 1e9;
2419 #else /* STA_NANO */
2420 clock_offset = ntv.offset / 1e6;
2421 #endif /* STA_NANO */
2422 drift_comp = FREQTOD(ntv.freq);
2424 #endif /* KERNEL_PLL */
2425 set_freq((fp_offset - clock_offset) / (current_time - clock_epoch) + drift_comp);
2431 set_freq(double freq) /* frequency update */
2439 * If the kernel is enabled, update the kernel frequency.
2441 if (pll_control && kern_enable) {
2442 memset(&ntv, 0, sizeof(ntv));
2443 ntv.modes = MOD_FREQUENCY;
2444 ntv.freq = DTOFREQ(drift_comp);
2446 snprintf(tbuf, sizeof(tbuf), "kernel %.3f PPM", drift_comp * 1e6);
2447 report_event(EVNT_FSET, NULL, tbuf);
2449 snprintf(tbuf, sizeof(tbuf), "ntpd %.3f PPM", drift_comp * 1e6);
2450 report_event(EVNT_FSET, NULL, tbuf);
2452 #else /* KERNEL_PLL */
2453 snprintf(tbuf, sizeof(tbuf), "ntpd %.3f PPM", drift_comp * 1e6);
2454 report_event(EVNT_FSET, NULL, tbuf);
2455 #endif /* KERNEL_PLL */
2464 * This code segment works when clock adjustments are made using
2465 * precision time kernel support and the ntp_adjtime() system
2466 * call. This support is available in Solaris 2.6 and later,
2467 * Digital Unix 4.0 and later, FreeBSD, Linux and specially
2468 * modified kernels for HP-UX 9 and Ultrix 4. In the case of the
2469 * DECstation 5000/240 and Alpha AXP, additional kernel
2470 * modifications provide a true microsecond clock and nanosecond
2471 * clock, respectively.
2473 * Important note: The kernel discipline is used only if the
2474 * step threshold is less than 0.5 s, as anything higher can
2475 * lead to overflow problems. This might occur if some misguided
2476 * lad set the step threshold to something ridiculous.
2478 if (pll_control && kern_enable) {
2480 #define MOD_BITS (MOD_OFFSET | MOD_MAXERROR | MOD_ESTERROR | MOD_STATUS | MOD_TIMECONST)
2483 * We initialize the structure for the ntp_adjtime()
2484 * system call. We have to convert everything to
2485 * microseconds or nanoseconds first. Do not update the
2486 * system variables if the ext_enable flag is set. In
2487 * this case, the external clock driver will update the
2488 * variables, which will be read later by the local
2489 * clock driver. Afterwards, remember the time and
2490 * frequency offsets for jitter and stability values and
2491 * to update the frequency file.
2493 memset(&ntv, 0, sizeof(ntv));
2495 ntv.modes = MOD_STATUS;
2498 ntv.modes = MOD_BITS | MOD_NANO;
2499 #else /* STA_NANO */
2500 ntv.modes = MOD_BITS;
2501 #endif /* STA_NANO */
2502 if (clock_offset < 0)
2507 ntv.offset = (int32)(clock_offset * 1e9 + dtemp);
2508 ntv.constant = sys_poll;
2509 #else /* STA_NANO */
2510 ntv.offset = (int32)(clock_offset * 1e6 + dtemp);
2511 ntv.constant = sys_poll - 4;
2512 #endif /* STA_NANO */
2513 ntv.esterror = (u_int32)(clock_jitter * 1e6);
2514 ntv.maxerror = (u_int32)((sys_rootdelay / 2 + sys_rootdisp) * 1e6);
2515 ntv.status = STA_PLL;
2518 * Enable/disable the PPS if requested.
2521 if (!(pll_status & STA_PPSTIME))
2522 report_event(EVNT_KERN,
2523 NULL, "PPS enabled");
2524 ntv.status |= STA_PPSTIME | STA_PPSFREQ;
2526 if (pll_status & STA_PPSTIME)
2527 report_event(EVNT_KERN,
2528 NULL, "PPS disabled");
2529 ntv.status &= ~(STA_PPSTIME | STA_PPSFREQ);
2531 if (sys_leap == LEAP_ADDSECOND)
2532 ntv.status |= STA_INS;
2533 else if (sys_leap == LEAP_DELSECOND)
2534 ntv.status |= STA_DEL;
2538 * Pass the stuff to the kernel. If it squeals, turn off
2539 * the pps. In any case, fetch the kernel offset,
2540 * frequency and jitter.
2542 if (ntp_adjtime(&ntv) == TIME_ERROR) {
2543 if (!(ntv.status & STA_PPSSIGNAL))
2544 report_event(EVNT_KERN, NULL,
2547 pll_status = ntv.status;
2549 clock_offset = ntv.offset / 1e9;
2550 #else /* STA_NANO */
2551 clock_offset = ntv.offset / 1e6;
2552 #endif /* STA_NANO */
2553 clock_frequency = FREQTOD(ntv.freq);
2556 * If the kernel PPS is lit, monitor its performance.
2558 if (ntv.status & STA_PPSTIME) {
2560 clock_jitter = ntv.jitter / 1e9;
2561 #else /* STA_NANO */
2562 clock_jitter = ntv.jitter / 1e6;
2563 #endif /* STA_NANO */
2566 #if defined(STA_NANO) && NTP_API == 4
2568 * If the TAI changes, update the kernel TAI.
2570 if (loop_tai != sys_tai) {
2572 ntv.modes = MOD_TAI;
2573 ntv.constant = sys_tai;
2576 #endif /* STA_NANO */
2578 #endif /* KERNEL_PLL */