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 ***********************************************************************
46 //config: select PLATFORM_LINUX
48 //config: The NTP client/server daemon.
50 //config:config FEATURE_NTPD_SERVER
51 //config: bool "Make ntpd usable as a NTP server"
53 //config: depends on NTPD
55 //config: Make ntpd usable as a NTP server. If you disable this option
56 //config: ntpd will be usable only as a NTP client.
58 //config:config FEATURE_NTPD_CONF
59 //config: bool "Make ntpd understand /etc/ntp.conf"
61 //config: depends on NTPD
63 //config: Make ntpd look in /etc/ntp.conf for peers. Only "server address"
64 //config: is supported.
66 //applet:IF_NTPD(APPLET(ntpd, BB_DIR_USR_SBIN, BB_SUID_DROP))
68 //kbuild:lib-$(CONFIG_NTPD) += ntpd.o
70 //usage:#define ntpd_trivial_usage
71 //usage: "[-dnqNw"IF_FEATURE_NTPD_SERVER("l -I IFACE")"] [-S PROG] [-p PEER]..."
72 //usage:#define ntpd_full_usage "\n\n"
73 //usage: "NTP client/server\n"
74 //usage: "\n -d Verbose"
75 //usage: "\n -n Do not daemonize"
76 //usage: "\n -q Quit after clock is set"
77 //usage: "\n -N Run at high priority"
78 //usage: "\n -w Do not set time (only query peers), implies -n"
79 //usage: "\n -S PROG Run PROG after stepping time, stratum change, and every 11 mins"
80 //usage: "\n -p PEER Obtain time from PEER (may be repeated)"
81 //usage: IF_FEATURE_NTPD_CONF(
82 //usage: "\n If -p is not given, 'server HOST' lines"
83 //usage: "\n from /etc/ntp.conf are used"
85 //usage: IF_FEATURE_NTPD_SERVER(
86 //usage: "\n -l Also run as server on port 123"
87 //usage: "\n -I IFACE Bind server to IFACE, implies -l"
90 // -l and -p options are not compatible with "standard" ntpd:
91 // it has them as "-l logfile" and "-p pidfile".
92 // -S and -w are not compat either, "standard" ntpd has no such opts.
96 #include <netinet/ip.h> /* For IPTOS_LOWDELAY definition */
97 #include <sys/resource.h> /* setpriority */
98 #include <sys/timex.h>
99 #ifndef IPTOS_LOWDELAY
100 # define IPTOS_LOWDELAY 0x10
104 /* Verbosity control (max level of -dddd options accepted).
105 * max 6 is very talkative (and bloated). 3 is non-bloated,
106 * production level setting.
108 #define MAX_VERBOSE 3
111 /* High-level description of the algorithm:
113 * We start running with very small poll_exp, BURSTPOLL,
114 * in order to quickly accumulate INITIAL_SAMPLES datapoints
115 * for each peer. Then, time is stepped if the offset is larger
116 * than STEP_THRESHOLD, otherwise it isn't; anyway, we enlarge
117 * poll_exp to MINPOLL and enter frequency measurement step:
118 * we collect new datapoints but ignore them for WATCH_THRESHOLD
119 * seconds. After WATCH_THRESHOLD seconds we look at accumulated
120 * offset and estimate frequency drift.
122 * (frequency measurement step seems to not be strictly needed,
123 * it is conditionally disabled with USING_INITIAL_FREQ_ESTIMATION
126 * After this, we enter "steady state": we collect a datapoint,
127 * we select the best peer, if this datapoint is not a new one
128 * (IOW: if this datapoint isn't for selected peer), sleep
129 * and collect another one; otherwise, use its offset to update
130 * frequency drift, if offset is somewhat large, reduce poll_exp,
131 * otherwise increase poll_exp.
133 * If offset is larger than STEP_THRESHOLD, which shouldn't normally
134 * happen, we assume that something "bad" happened (computer
135 * was hibernated, someone set totally wrong date, etc),
136 * then the time is stepped, all datapoints are discarded,
137 * and we go back to steady state.
139 * Made some changes to speed up re-syncing after our clock goes bad
140 * (tested with suspending my laptop):
141 * - if largish offset (>= STEP_THRESHOLD == 1 sec) is seen
142 * from a peer, schedule next query for this peer soon
143 * without drastically lowering poll interval for everybody.
144 * This makes us collect enough data for step much faster:
145 * e.g. at poll = 10 (1024 secs), step was done within 5 minutes
146 * after first reply which indicated that our clock is 14 seconds off.
147 * - on step, do not discard d_dispersion data of the existing datapoints,
148 * do not clear reachable_bits. This prevents discarding first ~8
149 * datapoints after the step.
152 #define INITIAL_SAMPLES 4 /* how many samples do we want for init */
153 #define BAD_DELAY_GROWTH 4 /* drop packet if its delay grew by more than this */
155 #define RETRY_INTERVAL 32 /* on send/recv error, retry in N secs (need to be power of 2) */
156 #define NOREPLY_INTERVAL 512 /* sent, but got no reply: cap next query by this many seconds */
157 #define RESPONSE_INTERVAL 16 /* wait for reply up to N secs */
159 /* Step threshold (sec). std ntpd uses 0.128.
161 #define STEP_THRESHOLD 1
162 /* Slew threshold (sec): adjtimex() won't accept offsets larger than this.
163 * Using exact power of 2 (1/8) results in smaller code
165 #define SLEW_THRESHOLD 0.125
166 /* Stepout threshold (sec). std ntpd uses 900 (11 mins (!)) */
167 #define WATCH_THRESHOLD 128
168 /* NB: set WATCH_THRESHOLD to ~60 when debugging to save time) */
169 //UNUSED: #define PANIC_THRESHOLD 1000 /* panic threshold (sec) */
172 * If we got |offset| > BIGOFF from a peer, cap next query interval
173 * for this peer by this many seconds:
175 #define BIGOFF STEP_THRESHOLD
176 #define BIGOFF_INTERVAL (1 << 7) /* 128 s */
178 #define FREQ_TOLERANCE 0.000015 /* frequency tolerance (15 PPM) */
179 #define BURSTPOLL 0 /* initial poll */
180 #define MINPOLL 5 /* minimum poll interval. std ntpd uses 6 (6: 64 sec) */
182 * If offset > discipline_jitter * POLLADJ_GATE, and poll interval is > 2^BIGPOLL,
183 * then it is decreased _at once_. (If <= 2^BIGPOLL, it will be decreased _eventually_).
185 #define BIGPOLL 9 /* 2^9 sec ~= 8.5 min */
186 #define MAXPOLL 12 /* maximum poll interval (12: 1.1h, 17: 36.4h). std ntpd uses 17 */
188 * Actively lower poll when we see such big offsets.
189 * With SLEW_THRESHOLD = 0.125, it means we try to sync more aggressively
190 * if offset increases over ~0.04 sec
192 //#define POLLDOWN_OFFSET (SLEW_THRESHOLD / 3)
193 #define MINDISP 0.01 /* minimum dispersion (sec) */
194 #define MAXDISP 16 /* maximum dispersion (sec) */
195 #define MAXSTRAT 16 /* maximum stratum (infinity metric) */
196 #define MAXDIST 1 /* distance threshold (sec) */
197 #define MIN_SELECTED 1 /* minimum intersection survivors */
198 #define MIN_CLUSTERED 3 /* minimum cluster survivors */
200 #define MAXDRIFT 0.000500 /* frequency drift we can correct (500 PPM) */
202 /* Poll-adjust threshold.
203 * When we see that offset is small enough compared to discipline jitter,
204 * we grow a counter: += MINPOLL. When counter goes over POLLADJ_LIMIT,
205 * we poll_exp++. If offset isn't small, counter -= poll_exp*2,
206 * and when it goes below -POLLADJ_LIMIT, we poll_exp--.
207 * (Bumped from 30 to 40 since otherwise I often see poll_exp going *2* steps down)
209 #define POLLADJ_LIMIT 40
210 /* If offset < discipline_jitter * POLLADJ_GATE, then we decide to increase
211 * poll interval (we think we can't improve timekeeping
212 * by staying at smaller poll).
214 #define POLLADJ_GATE 4
215 #define TIMECONST_HACK_GATE 2
216 /* Compromise Allan intercept (sec). doc uses 1500, std ntpd uses 512 */
220 /* FLL loop gain [why it depends on MAXPOLL??] */
221 #define FLL (MAXPOLL + 1)
222 /* Parameter averaging constant */
231 NTP_MSGSIZE_NOAUTH = 48,
232 NTP_MSGSIZE = (NTP_MSGSIZE_NOAUTH + 4 + NTP_DIGESTSIZE),
235 MODE_MASK = (7 << 0),
236 VERSION_MASK = (7 << 3),
240 /* Leap Second Codes (high order two bits of m_status) */
241 LI_NOWARNING = (0 << 6), /* no warning */
242 LI_PLUSSEC = (1 << 6), /* add a second (61 seconds) */
243 LI_MINUSSEC = (2 << 6), /* minus a second (59 seconds) */
244 LI_ALARM = (3 << 6), /* alarm condition */
247 MODE_RES0 = 0, /* reserved */
248 MODE_SYM_ACT = 1, /* symmetric active */
249 MODE_SYM_PAS = 2, /* symmetric passive */
250 MODE_CLIENT = 3, /* client */
251 MODE_SERVER = 4, /* server */
252 MODE_BROADCAST = 5, /* broadcast */
253 MODE_RES1 = 6, /* reserved for NTP control message */
254 MODE_RES2 = 7, /* reserved for private use */
257 //TODO: better base selection
258 #define OFFSET_1900_1970 2208988800UL /* 1970 - 1900 in seconds */
260 #define NUM_DATAPOINTS 8
273 uint8_t m_status; /* status of local clock and leap info */
275 uint8_t m_ppoll; /* poll value */
276 int8_t m_precision_exp;
277 s_fixedpt_t m_rootdelay;
278 s_fixedpt_t m_rootdisp;
280 l_fixedpt_t m_reftime;
281 l_fixedpt_t m_orgtime;
282 l_fixedpt_t m_rectime;
283 l_fixedpt_t m_xmttime;
285 uint8_t m_digest[NTP_DIGESTSIZE];
295 len_and_sockaddr *p_lsa;
299 uint32_t lastpkt_refid;
300 uint8_t lastpkt_status;
301 uint8_t lastpkt_stratum;
302 uint8_t reachable_bits;
303 /* when to send new query (if p_fd == -1)
304 * or when receive times out (if p_fd >= 0): */
305 double next_action_time;
308 /* p_raw_delay is set even by "high delay" packets */
309 /* lastpkt_delay isn't */
310 double lastpkt_recv_time;
311 double lastpkt_delay;
312 double lastpkt_rootdelay;
313 double lastpkt_rootdisp;
314 /* produced by filter algorithm: */
315 double filter_offset;
316 double filter_dispersion;
317 double filter_jitter;
318 datapoint_t filter_datapoint[NUM_DATAPOINTS];
319 /* last sent packet: */
325 #define USING_KERNEL_PLL_LOOP 1
326 #define USING_INITIAL_FREQ_ESTIMATION 0
333 /* Insert new options above this line. */
334 /* Non-compat options: */
338 OPT_l = (1 << 7) * ENABLE_FEATURE_NTPD_SERVER,
339 OPT_I = (1 << 8) * ENABLE_FEATURE_NTPD_SERVER,
340 /* We hijack some bits for other purposes */
346 /* total round trip delay to currently selected reference clock */
348 /* reference timestamp: time when the system clock was last set or corrected */
350 /* total dispersion to currently selected reference clock */
353 double last_script_run;
356 #if ENABLE_FEATURE_NTPD_SERVER
359 # define G_listen_fd (G.listen_fd)
361 # define G_listen_fd (-1)
365 /* refid: 32-bit code identifying the particular server or reference clock
366 * in stratum 0 packets this is a four-character ASCII string,
367 * called the kiss code, used for debugging and monitoring
368 * in stratum 1 packets this is a four-character ASCII string
369 * assigned to the reference clock by IANA. Example: "GPS "
370 * in stratum 2+ packets, it's IPv4 address or 4 first bytes
371 * of MD5 hash of IPv6
375 /* precision is defined as the larger of the resolution and time to
376 * read the clock, in log2 units. For instance, the precision of a
377 * mains-frequency clock incrementing at 60 Hz is 16 ms, even when the
378 * system clock hardware representation is to the nanosecond.
380 * Delays, jitters of various kinds are clamped down to precision.
382 * If precision_sec is too large, discipline_jitter gets clamped to it
383 * and if offset is smaller than discipline_jitter * POLLADJ_GATE, poll
384 * interval grows even though we really can benefit from staying at
385 * smaller one, collecting non-lagged datapoits and correcting offset.
386 * (Lagged datapoits exist when poll_exp is large but we still have
387 * systematic offset error - the time distance between datapoints
388 * is significant and older datapoints have smaller offsets.
389 * This makes our offset estimation a bit smaller than reality)
390 * Due to this effect, setting G_precision_sec close to
391 * STEP_THRESHOLD isn't such a good idea - offsets may grow
392 * too big and we will step. I observed it with -6.
394 * OTOH, setting precision_sec far too small would result in futile
395 * attempts to syncronize to an unachievable precision.
397 * -6 is 1/64 sec, -7 is 1/128 sec and so on.
398 * -8 is 1/256 ~= 0.003906 (worked well for me --vda)
399 * -9 is 1/512 ~= 0.001953 (let's try this for some time)
401 #define G_precision_exp -9
403 * G_precision_exp is used only for construction outgoing packets.
404 * It's ok to set G_precision_sec to a slightly different value
405 * (One which is "nicer looking" in logs).
406 * Exact value would be (1.0 / (1 << (- G_precision_exp))):
408 #define G_precision_sec 0.002
411 #define STATE_NSET 0 /* initial state, "nothing is set" */
412 //#define STATE_FSET 1 /* frequency set from file */
413 //#define STATE_SPIK 2 /* spike detected */
414 //#define STATE_FREQ 3 /* initial frequency */
415 #define STATE_SYNC 4 /* clock synchronized (normal operation) */
416 uint8_t discipline_state; // doc calls it c.state
417 uint8_t poll_exp; // s.poll
418 int polladj_count; // c.count
419 long kernel_freq_drift;
420 peer_t *last_update_peer;
421 double last_update_offset; // c.last
422 double last_update_recv_time; // s.t
423 double discipline_jitter; // c.jitter
424 /* Since we only compare it with ints, can simplify code
425 * by not making this variable floating point:
427 unsigned offset_to_jitter_ratio;
428 //double cluster_offset; // s.offset
429 //double cluster_jitter; // s.jitter
430 #if !USING_KERNEL_PLL_LOOP
431 double discipline_freq_drift; // c.freq
432 /* Maybe conditionally calculate wander? it's used only for logging */
433 double discipline_wander; // c.wander
436 #define G (*ptr_to_globals)
439 #define VERB1 if (MAX_VERBOSE && G.verbose)
440 #define VERB2 if (MAX_VERBOSE >= 2 && G.verbose >= 2)
441 #define VERB3 if (MAX_VERBOSE >= 3 && G.verbose >= 3)
442 #define VERB4 if (MAX_VERBOSE >= 4 && G.verbose >= 4)
443 #define VERB5 if (MAX_VERBOSE >= 5 && G.verbose >= 5)
444 #define VERB6 if (MAX_VERBOSE >= 6 && G.verbose >= 6)
447 static double LOG2D(int a)
450 return 1.0 / (1UL << -a);
453 static ALWAYS_INLINE double SQUARE(double x)
457 static ALWAYS_INLINE double MAXD(double a, double b)
463 static ALWAYS_INLINE double MIND(double a, double b)
469 static NOINLINE double my_SQRT(double X)
476 double Xhalf = X * 0.5;
478 /* Fast and good approximation to 1/sqrt(X), black magic */
480 /*v.i = 0x5f3759df - (v.i >> 1);*/
481 v.i = 0x5f375a86 - (v.i >> 1); /* - this constant is slightly better */
482 invsqrt = v.f; /* better than 0.2% accuracy */
484 /* Refining it using Newton's method: x1 = x0 - f(x0)/f'(x0)
485 * f(x) = 1/(x*x) - X (f==0 when x = 1/sqrt(X))
487 * f(x)/f'(x) = (X - 1/(x*x)) / (2/(x*x*x)) = X*x*x*x/2 - x/2
488 * x1 = x0 - (X*x0*x0*x0/2 - x0/2) = 1.5*x0 - X*x0*x0*x0/2 = x0*(1.5 - (X/2)*x0*x0)
490 invsqrt = invsqrt * (1.5 - Xhalf * invsqrt * invsqrt); /* ~0.05% accuracy */
491 /* invsqrt = invsqrt * (1.5 - Xhalf * invsqrt * invsqrt); 2nd iter: ~0.0001% accuracy */
492 /* With 4 iterations, more than half results will be exact,
493 * at 6th iterations result stabilizes with about 72% results exact.
494 * We are well satisfied with 0.05% accuracy.
497 return X * invsqrt; /* X * 1/sqrt(X) ~= sqrt(X) */
499 static ALWAYS_INLINE double SQRT(double X)
501 /* If this arch doesn't use IEEE 754 floats, fall back to using libm */
502 if (sizeof(float) != 4)
505 /* This avoids needing libm, saves about 0.5k on x86-32 */
513 gettimeofday(&tv, NULL); /* never fails */
514 G.cur_time = tv.tv_sec + (1.0e-6 * tv.tv_usec) + OFFSET_1900_1970;
519 d_to_tv(double d, struct timeval *tv)
521 tv->tv_sec = (long)d;
522 tv->tv_usec = (d - tv->tv_sec) * 1000000;
526 lfp_to_d(l_fixedpt_t lfp)
529 lfp.int_partl = ntohl(lfp.int_partl);
530 lfp.fractionl = ntohl(lfp.fractionl);
531 ret = (double)lfp.int_partl + ((double)lfp.fractionl / UINT_MAX);
535 sfp_to_d(s_fixedpt_t sfp)
538 sfp.int_parts = ntohs(sfp.int_parts);
539 sfp.fractions = ntohs(sfp.fractions);
540 ret = (double)sfp.int_parts + ((double)sfp.fractions / USHRT_MAX);
543 #if ENABLE_FEATURE_NTPD_SERVER
548 lfp.int_partl = (uint32_t)d;
549 lfp.fractionl = (uint32_t)((d - lfp.int_partl) * UINT_MAX);
550 lfp.int_partl = htonl(lfp.int_partl);
551 lfp.fractionl = htonl(lfp.fractionl);
558 sfp.int_parts = (uint16_t)d;
559 sfp.fractions = (uint16_t)((d - sfp.int_parts) * USHRT_MAX);
560 sfp.int_parts = htons(sfp.int_parts);
561 sfp.fractions = htons(sfp.fractions);
567 dispersion(const datapoint_t *dp)
569 return dp->d_dispersion + FREQ_TOLERANCE * (G.cur_time - dp->d_recv_time);
573 root_distance(peer_t *p)
575 /* The root synchronization distance is the maximum error due to
576 * all causes of the local clock relative to the primary server.
577 * It is defined as half the total delay plus total dispersion
580 return MAXD(MINDISP, p->lastpkt_rootdelay + p->lastpkt_delay) / 2
581 + p->lastpkt_rootdisp
582 + p->filter_dispersion
583 + FREQ_TOLERANCE * (G.cur_time - p->lastpkt_recv_time)
588 set_next(peer_t *p, unsigned t)
590 p->next_action_time = G.cur_time + t;
594 * Peer clock filter and its helpers
597 filter_datapoints(peer_t *p)
604 /* Simulations have shown that use of *averaged* offset for p->filter_offset
605 * is in fact worse than simply using last received one: with large poll intervals
606 * (>= 2048) averaging code uses offset values which are outdated by hours,
607 * and time/frequency correction goes totally wrong when fed essentially bogus offsets.
610 double minoff, maxoff, w;
611 double x = x; /* for compiler */
612 double oldest_off = oldest_off;
613 double oldest_age = oldest_age;
614 double newest_off = newest_off;
615 double newest_age = newest_age;
617 fdp = p->filter_datapoint;
619 minoff = maxoff = fdp[0].d_offset;
620 for (i = 1; i < NUM_DATAPOINTS; i++) {
621 if (minoff > fdp[i].d_offset)
622 minoff = fdp[i].d_offset;
623 if (maxoff < fdp[i].d_offset)
624 maxoff = fdp[i].d_offset;
627 idx = p->datapoint_idx; /* most recent datapoint's index */
629 * Drop two outliers and take weighted average of the rest:
630 * most_recent/2 + older1/4 + older2/8 ... + older5/32 + older6/32
631 * we use older6/32, not older6/64 since sum of weights should be 1:
632 * 1/2 + 1/4 + 1/8 + 1/16 + 1/32 + 1/32 = 1
638 * filter_dispersion = \ -------------
645 for (i = 0; i < NUM_DATAPOINTS; i++) {
647 bb_error_msg("datapoint[%d]: off:%f disp:%f(%f) age:%f%s",
650 fdp[idx].d_dispersion, dispersion(&fdp[idx]),
651 G.cur_time - fdp[idx].d_recv_time,
652 (minoff == fdp[idx].d_offset || maxoff == fdp[idx].d_offset)
653 ? " (outlier by offset)" : ""
657 sum += dispersion(&fdp[idx]) / (2 << i);
659 if (minoff == fdp[idx].d_offset) {
660 minoff -= 1; /* so that we don't match it ever again */
662 if (maxoff == fdp[idx].d_offset) {
665 oldest_off = fdp[idx].d_offset;
666 oldest_age = G.cur_time - fdp[idx].d_recv_time;
669 newest_off = oldest_off;
670 newest_age = oldest_age;
677 idx = (idx - 1) & (NUM_DATAPOINTS - 1);
679 p->filter_dispersion = sum;
680 wavg += x; /* add another older6/64 to form older6/32 */
681 /* Fix systematic underestimation with large poll intervals.
682 * Imagine that we still have a bit of uncorrected drift,
683 * and poll interval is big (say, 100 sec). Offsets form a progression:
684 * 0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 - 0.7 is most recent.
685 * The algorithm above drops 0.0 and 0.7 as outliers,
686 * and then we have this estimation, ~25% off from 0.7:
687 * 0.1/32 + 0.2/32 + 0.3/16 + 0.4/8 + 0.5/4 + 0.6/2 = 0.503125
689 x = oldest_age - newest_age;
691 x = newest_age / x; /* in above example, 100 / (600 - 100) */
692 if (x < 1) { /* paranoia check */
693 x = (newest_off - oldest_off) * x; /* 0.5 * 100/500 = 0.1 */
697 p->filter_offset = wavg;
701 fdp = p->filter_datapoint;
702 idx = p->datapoint_idx; /* most recent datapoint's index */
704 /* filter_offset: simply use the most recent value */
705 p->filter_offset = fdp[idx].d_offset;
709 * filter_dispersion = \ -------------
716 for (i = 0; i < NUM_DATAPOINTS; i++) {
717 sum += dispersion(&fdp[idx]) / (2 << i);
718 wavg += fdp[idx].d_offset;
719 idx = (idx - 1) & (NUM_DATAPOINTS - 1);
721 wavg /= NUM_DATAPOINTS;
722 p->filter_dispersion = sum;
725 /* +----- -----+ ^ 1/2
729 * filter_jitter = | --- * / (avg-offset_j) |
733 * where n is the number of valid datapoints in the filter (n > 1);
734 * if filter_jitter < precision then filter_jitter = precision
737 for (i = 0; i < NUM_DATAPOINTS; i++) {
738 sum += SQUARE(wavg - fdp[i].d_offset);
740 sum = SQRT(sum / NUM_DATAPOINTS);
741 p->filter_jitter = sum > G_precision_sec ? sum : G_precision_sec;
743 VERB4 bb_error_msg("filter offset:%+f disp:%f jitter:%f",
745 p->filter_dispersion,
750 reset_peer_stats(peer_t *p, double offset)
753 bool small_ofs = fabs(offset) < STEP_THRESHOLD;
755 /* Used to set p->filter_datapoint[i].d_dispersion = MAXDISP
756 * and clear reachable bits, but this proved to be too agressive:
757 * after step (tested with suspending laptop for ~30 secs),
758 * this caused all previous data to be considered invalid,
759 * making us needing to collect full ~8 datapoins per peer
760 * after step in order to start trusting them.
761 * In turn, this was making poll interval decrease even after
762 * step was done. (Poll interval decreases already before step
763 * in this scenario, because we see large offsets and end up with
764 * no good peer to select).
767 for (i = 0; i < NUM_DATAPOINTS; i++) {
769 p->filter_datapoint[i].d_recv_time += offset;
770 if (p->filter_datapoint[i].d_offset != 0) {
771 p->filter_datapoint[i].d_offset -= offset;
772 //bb_error_msg("p->filter_datapoint[%d].d_offset %f -> %f",
774 // p->filter_datapoint[i].d_offset + offset,
775 // p->filter_datapoint[i].d_offset);
778 p->filter_datapoint[i].d_recv_time = G.cur_time;
779 p->filter_datapoint[i].d_offset = 0;
780 /*p->filter_datapoint[i].d_dispersion = MAXDISP;*/
784 p->lastpkt_recv_time += offset;
786 /*p->reachable_bits = 0;*/
787 p->lastpkt_recv_time = G.cur_time;
789 filter_datapoints(p); /* recalc p->filter_xxx */
790 VERB6 bb_error_msg("%s->lastpkt_recv_time=%f", p->p_dotted, p->lastpkt_recv_time);
794 resolve_peer_hostname(peer_t *p, int loop_on_fail)
796 len_and_sockaddr *lsa;
799 lsa = host2sockaddr(p->p_hostname, 123);
801 /* error message already emitted by host2sockaddr() */
804 //FIXME: do this to avoid infinite looping on typo in a hostname?
805 //well... in which case, what is a good value for loop_on_fail?
806 //if (--loop_on_fail == 0)
814 p->p_dotted = xmalloc_sockaddr2dotted_noport(&lsa->u.sa);
818 add_peers(const char *s)
823 p = xzalloc(sizeof(*p) + strlen(s));
824 strcpy(p->p_hostname, s);
825 resolve_peer_hostname(p, /*loop_on_fail=*/ 1);
827 /* Names like N.<country2chars>.pool.ntp.org are randomly resolved
828 * to a pool of machines. Sometimes different N's resolve to the same IP.
829 * It is not useful to have two peers with same IP. We skip duplicates.
831 for (item = G.ntp_peers; item != NULL; item = item->link) {
832 peer_t *pp = (peer_t *) item->data;
833 if (strcmp(p->p_dotted, pp->p_dotted) == 0) {
834 bb_error_msg("duplicate peer %s (%s)", s, p->p_dotted);
843 p->p_xmt_msg.m_status = MODE_CLIENT | (NTP_VERSION << 3);
844 p->next_action_time = G.cur_time; /* = set_next(p, 0); */
845 reset_peer_stats(p, STEP_THRESHOLD);
847 llist_add_to(&G.ntp_peers, p);
853 const struct sockaddr *from, const struct sockaddr *to, socklen_t addrlen,
854 msg_t *msg, ssize_t len)
860 ret = sendto(fd, msg, len, MSG_DONTWAIT, to, addrlen);
862 ret = send_to_from(fd, msg, len, MSG_DONTWAIT, to, from, addrlen);
865 bb_perror_msg("send failed");
872 send_query_to_peer(peer_t *p)
874 /* Why do we need to bind()?
875 * See what happens when we don't bind:
877 * socket(PF_INET, SOCK_DGRAM, IPPROTO_IP) = 3
878 * setsockopt(3, SOL_IP, IP_TOS, [16], 4) = 0
879 * gettimeofday({1259071266, 327885}, NULL) = 0
880 * sendto(3, "xxx", 48, MSG_DONTWAIT, {sa_family=AF_INET, sin_port=htons(123), sin_addr=inet_addr("10.34.32.125")}, 16) = 48
881 * ^^^ we sent it from some source port picked by kernel.
882 * time(NULL) = 1259071266
883 * write(2, "ntpd: entering poll 15 secs\n", 28) = 28
884 * poll([{fd=3, events=POLLIN}], 1, 15000) = 1 ([{fd=3, revents=POLLIN}])
885 * recv(3, "yyy", 68, MSG_DONTWAIT) = 48
886 * ^^^ this recv will receive packets to any local port!
888 * Uncomment this and use strace to see it in action:
890 #define PROBE_LOCAL_ADDR /* { len_and_sockaddr lsa; lsa.len = LSA_SIZEOF_SA; getsockname(p->query.fd, &lsa.u.sa, &lsa.len); } */
894 len_and_sockaddr *local_lsa;
896 family = p->p_lsa->u.sa.sa_family;
897 p->p_fd = fd = xsocket_type(&local_lsa, family, SOCK_DGRAM);
898 /* local_lsa has "null" address and port 0 now.
899 * bind() ensures we have a *particular port* selected by kernel
900 * and remembered in p->p_fd, thus later recv(p->p_fd)
901 * receives only packets sent to this port.
904 xbind(fd, &local_lsa->u.sa, local_lsa->len);
906 #if ENABLE_FEATURE_IPV6
907 if (family == AF_INET)
909 setsockopt_int(fd, IPPROTO_IP, IP_TOS, IPTOS_LOWDELAY);
913 /* Emit message _before_ attempted send. Think of a very short
914 * roundtrip networks: we need to go back to recv loop ASAP,
915 * to reduce delay. Printing messages after send works against that.
917 VERB1 bb_error_msg("sending query to %s", p->p_dotted);
920 * Send out a random 64-bit number as our transmit time. The NTP
921 * server will copy said number into the originate field on the
922 * response that it sends us. This is totally legal per the SNTP spec.
924 * The impact of this is two fold: we no longer send out the current
925 * system time for the world to see (which may aid an attacker), and
926 * it gives us a (not very secure) way of knowing that we're not
927 * getting spoofed by an attacker that can't capture our traffic
928 * but can spoof packets from the NTP server we're communicating with.
930 * Save the real transmit timestamp locally.
932 p->p_xmt_msg.m_xmttime.int_partl = rand();
933 p->p_xmt_msg.m_xmttime.fractionl = rand();
934 p->p_xmttime = gettime1900d();
936 /* Were doing it only if sendto worked, but
937 * loss of sync detection needs reachable_bits updated
938 * even if sending fails *locally*:
939 * "network is unreachable" because cable was pulled?
940 * We still need to declare "unsync" if this condition persists.
942 p->reachable_bits <<= 1;
944 if (do_sendto(p->p_fd, /*from:*/ NULL, /*to:*/ &p->p_lsa->u.sa, /*addrlen:*/ p->p_lsa->len,
945 &p->p_xmt_msg, NTP_MSGSIZE_NOAUTH) == -1
950 * We know that we sent nothing.
951 * We can retry *soon* without fearing
952 * that we are flooding the peer.
954 set_next(p, RETRY_INTERVAL);
958 set_next(p, RESPONSE_INTERVAL);
962 /* Note that there is no provision to prevent several run_scripts
963 * to be started in quick succession. In fact, it happens rather often
964 * if initial syncronization results in a step.
965 * You will see "step" and then "stratum" script runs, sometimes
966 * as close as only 0.002 seconds apart.
967 * Script should be ready to deal with this.
969 static void run_script(const char *action, double offset)
972 char *env1, *env2, *env3, *env4;
974 G.last_script_run = G.cur_time;
979 argv[0] = (char*) G.script_name;
980 argv[1] = (char*) action;
983 VERB1 bb_error_msg("executing '%s %s'", G.script_name, action);
985 env1 = xasprintf("%s=%u", "stratum", G.stratum);
987 env2 = xasprintf("%s=%ld", "freq_drift_ppm", G.kernel_freq_drift);
989 env3 = xasprintf("%s=%u", "poll_interval", 1 << G.poll_exp);
991 env4 = xasprintf("%s=%f", "offset", offset);
993 /* Other items of potential interest: selected peer,
994 * rootdelay, reftime, rootdisp, refid, ntp_status,
995 * last_update_offset, last_update_recv_time, discipline_jitter,
996 * how many peers have reachable_bits = 0?
999 /* Don't want to wait: it may run hwclock --systohc, and that
1000 * may take some time (seconds): */
1001 /*spawn_and_wait(argv);*/
1004 unsetenv("stratum");
1005 unsetenv("freq_drift_ppm");
1006 unsetenv("poll_interval");
1014 static NOINLINE void
1015 step_time(double offset)
1019 struct timeval tvc, tvn;
1020 char buf[sizeof("yyyy-mm-dd hh:mm:ss") + /*paranoia:*/ 4];
1023 gettimeofday(&tvc, NULL); /* never fails */
1024 dtime = tvc.tv_sec + (1.0e-6 * tvc.tv_usec) + offset;
1025 d_to_tv(dtime, &tvn);
1026 if (settimeofday(&tvn, NULL) == -1)
1027 bb_perror_msg_and_die("settimeofday");
1031 strftime_YYYYMMDDHHMMSS(buf, sizeof(buf), &tval);
1032 bb_error_msg("current time is %s.%06u", buf, (unsigned)tvc.tv_usec);
1035 strftime_YYYYMMDDHHMMSS(buf, sizeof(buf), &tval);
1036 bb_error_msg("setting time to %s.%06u (offset %+fs)", buf, (unsigned)tvn.tv_usec, offset);
1038 /* Correct various fields which contain time-relative values: */
1041 G.cur_time += offset;
1042 G.last_update_recv_time += offset;
1043 G.last_script_run += offset;
1045 /* p->lastpkt_recv_time, p->next_action_time and such: */
1046 for (item = G.ntp_peers; item != NULL; item = item->link) {
1047 peer_t *pp = (peer_t *) item->data;
1048 reset_peer_stats(pp, offset);
1049 //bb_error_msg("offset:%+f pp->next_action_time:%f -> %f",
1050 // offset, pp->next_action_time, pp->next_action_time + offset);
1051 pp->next_action_time += offset;
1052 if (pp->p_fd >= 0) {
1053 /* We wait for reply from this peer too.
1054 * But due to step we are doing, reply's data is no longer
1055 * useful (in fact, it'll be bogus). Stop waiting for it.
1059 set_next(pp, RETRY_INTERVAL);
1064 static void clamp_pollexp_and_set_MAXSTRAT(void)
1066 if (G.poll_exp < MINPOLL)
1067 G.poll_exp = MINPOLL;
1068 if (G.poll_exp > BIGPOLL)
1069 G.poll_exp = BIGPOLL;
1070 G.polladj_count = 0;
1071 G.stratum = MAXSTRAT;
1076 * Selection and clustering, and their helpers
1082 double opt_rd; /* optimization */
1085 compare_point_edge(const void *aa, const void *bb)
1087 const point_t *a = aa;
1088 const point_t *b = bb;
1089 if (a->edge < b->edge) {
1092 return (a->edge > b->edge);
1099 compare_survivor_metric(const void *aa, const void *bb)
1101 const survivor_t *a = aa;
1102 const survivor_t *b = bb;
1103 if (a->metric < b->metric) {
1106 return (a->metric > b->metric);
1109 fit(peer_t *p, double rd)
1111 if ((p->reachable_bits & (p->reachable_bits-1)) == 0) {
1112 /* One or zero bits in reachable_bits */
1113 VERB4 bb_error_msg("peer %s unfit for selection: unreachable", p->p_dotted);
1116 #if 0 /* we filter out such packets earlier */
1117 if ((p->lastpkt_status & LI_ALARM) == LI_ALARM
1118 || p->lastpkt_stratum >= MAXSTRAT
1120 VERB4 bb_error_msg("peer %s unfit for selection: bad status/stratum", p->p_dotted);
1124 /* rd is root_distance(p) */
1125 if (rd > MAXDIST + FREQ_TOLERANCE * (1 << G.poll_exp)) {
1126 VERB4 bb_error_msg("peer %s unfit for selection: root distance too high", p->p_dotted);
1130 // /* Do we have a loop? */
1131 // if (p->refid == p->dstaddr || p->refid == s.refid)
1136 select_and_cluster(void)
1141 int size = 3 * G.peer_cnt;
1142 /* for selection algorithm */
1143 point_t point[size];
1144 unsigned num_points, num_candidates;
1146 unsigned num_falsetickers;
1147 /* for cluster algorithm */
1148 survivor_t survivor[size];
1149 unsigned num_survivors;
1155 while (item != NULL) {
1158 p = (peer_t *) item->data;
1159 rd = root_distance(p);
1160 offset = p->filter_offset;
1166 VERB5 bb_error_msg("interval: [%f %f %f] %s",
1172 point[num_points].p = p;
1173 point[num_points].type = -1;
1174 point[num_points].edge = offset - rd;
1175 point[num_points].opt_rd = rd;
1177 point[num_points].p = p;
1178 point[num_points].type = 0;
1179 point[num_points].edge = offset;
1180 point[num_points].opt_rd = rd;
1182 point[num_points].p = p;
1183 point[num_points].type = 1;
1184 point[num_points].edge = offset + rd;
1185 point[num_points].opt_rd = rd;
1189 num_candidates = num_points / 3;
1190 if (num_candidates == 0) {
1191 VERB3 bb_error_msg("no valid datapoints%s", ", no peer selected");
1194 //TODO: sorting does not seem to be done in reference code
1195 qsort(point, num_points, sizeof(point[0]), compare_point_edge);
1197 /* Start with the assumption that there are no falsetickers.
1198 * Attempt to find a nonempty intersection interval containing
1199 * the midpoints of all truechimers.
1200 * If a nonempty interval cannot be found, increase the number
1201 * of assumed falsetickers by one and try again.
1202 * If a nonempty interval is found and the number of falsetickers
1203 * is less than the number of truechimers, a majority has been found
1204 * and the midpoint of each truechimer represents
1205 * the candidates available to the cluster algorithm.
1207 num_falsetickers = 0;
1210 unsigned num_midpoints = 0;
1215 for (i = 0; i < num_points; i++) {
1217 * if (point[i].type == -1) c++;
1218 * if (point[i].type == 1) c--;
1219 * and it's simpler to do it this way:
1222 if (c >= num_candidates - num_falsetickers) {
1223 /* If it was c++ and it got big enough... */
1224 low = point[i].edge;
1227 if (point[i].type == 0)
1231 for (i = num_points-1; i >= 0; i--) {
1233 if (c >= num_candidates - num_falsetickers) {
1234 high = point[i].edge;
1237 if (point[i].type == 0)
1240 /* If the number of midpoints is greater than the number
1241 * of allowed falsetickers, the intersection contains at
1242 * least one truechimer with no midpoint - bad.
1243 * Also, interval should be nonempty.
1245 if (num_midpoints <= num_falsetickers && low < high)
1248 if (num_falsetickers * 2 >= num_candidates) {
1249 VERB3 bb_error_msg("falsetickers:%d, candidates:%d%s",
1250 num_falsetickers, num_candidates,
1251 ", no peer selected");
1255 VERB4 bb_error_msg("selected interval: [%f, %f]; candidates:%d falsetickers:%d",
1256 low, high, num_candidates, num_falsetickers);
1260 /* Construct a list of survivors (p, metric)
1261 * from the chime list, where metric is dominated
1262 * first by stratum and then by root distance.
1263 * All other things being equal, this is the order of preference.
1266 for (i = 0; i < num_points; i++) {
1267 if (point[i].edge < low || point[i].edge > high)
1270 survivor[num_survivors].p = p;
1271 /* x.opt_rd == root_distance(p); */
1272 survivor[num_survivors].metric = MAXDIST * p->lastpkt_stratum + point[i].opt_rd;
1273 VERB5 bb_error_msg("survivor[%d] metric:%f peer:%s",
1274 num_survivors, survivor[num_survivors].metric, p->p_dotted);
1277 /* There must be at least MIN_SELECTED survivors to satisfy the
1278 * correctness assertions. Ordinarily, the Byzantine criteria
1279 * require four survivors, but for the demonstration here, one
1282 if (num_survivors < MIN_SELECTED) {
1283 VERB3 bb_error_msg("survivors:%d%s",
1285 ", no peer selected");
1289 //looks like this is ONLY used by the fact that later we pick survivor[0].
1290 //we can avoid sorting then, just find the minimum once!
1291 qsort(survivor, num_survivors, sizeof(survivor[0]), compare_survivor_metric);
1293 /* For each association p in turn, calculate the selection
1294 * jitter p->sjitter as the square root of the sum of squares
1295 * (p->offset - q->offset) over all q associations. The idea is
1296 * to repeatedly discard the survivor with maximum selection
1297 * jitter until a termination condition is met.
1300 unsigned max_idx = max_idx;
1301 double max_selection_jitter = max_selection_jitter;
1302 double min_jitter = min_jitter;
1304 if (num_survivors <= MIN_CLUSTERED) {
1305 VERB4 bb_error_msg("num_survivors %d <= %d, not discarding more",
1306 num_survivors, MIN_CLUSTERED);
1310 /* To make sure a few survivors are left
1311 * for the clustering algorithm to chew on,
1312 * we stop if the number of survivors
1313 * is less than or equal to MIN_CLUSTERED (3).
1315 for (i = 0; i < num_survivors; i++) {
1316 double selection_jitter_sq;
1319 if (i == 0 || p->filter_jitter < min_jitter)
1320 min_jitter = p->filter_jitter;
1322 selection_jitter_sq = 0;
1323 for (j = 0; j < num_survivors; j++) {
1324 peer_t *q = survivor[j].p;
1325 selection_jitter_sq += SQUARE(p->filter_offset - q->filter_offset);
1327 if (i == 0 || selection_jitter_sq > max_selection_jitter) {
1328 max_selection_jitter = selection_jitter_sq;
1331 VERB6 bb_error_msg("survivor %d selection_jitter^2:%f",
1332 i, selection_jitter_sq);
1334 max_selection_jitter = SQRT(max_selection_jitter / num_survivors);
1335 VERB5 bb_error_msg("max_selection_jitter (at %d):%f min_jitter:%f",
1336 max_idx, max_selection_jitter, min_jitter);
1338 /* If the maximum selection jitter is less than the
1339 * minimum peer jitter, then tossing out more survivors
1340 * will not lower the minimum peer jitter, so we might
1343 if (max_selection_jitter < min_jitter) {
1344 VERB4 bb_error_msg("max_selection_jitter:%f < min_jitter:%f, num_survivors:%d, not discarding more",
1345 max_selection_jitter, min_jitter, num_survivors);
1349 /* Delete survivor[max_idx] from the list
1350 * and go around again.
1352 VERB6 bb_error_msg("dropping survivor %d", max_idx);
1354 while (max_idx < num_survivors) {
1355 survivor[max_idx] = survivor[max_idx + 1];
1361 /* Combine the offsets of the clustering algorithm survivors
1362 * using a weighted average with weight determined by the root
1363 * distance. Compute the selection jitter as the weighted RMS
1364 * difference between the first survivor and the remaining
1365 * survivors. In some cases the inherent clock jitter can be
1366 * reduced by not using this algorithm, especially when frequent
1367 * clockhopping is involved. bbox: thus we don't do it.
1371 for (i = 0; i < num_survivors; i++) {
1373 x = root_distance(p);
1375 z += p->filter_offset / x;
1376 w += SQUARE(p->filter_offset - survivor[0].p->filter_offset) / x;
1378 //G.cluster_offset = z / y;
1379 //G.cluster_jitter = SQRT(w / y);
1382 /* Pick the best clock. If the old system peer is on the list
1383 * and at the same stratum as the first survivor on the list,
1384 * then don't do a clock hop. Otherwise, select the first
1385 * survivor on the list as the new system peer.
1388 if (G.last_update_peer
1389 && G.last_update_peer->lastpkt_stratum <= p->lastpkt_stratum
1391 /* Starting from 1 is ok here */
1392 for (i = 1; i < num_survivors; i++) {
1393 if (G.last_update_peer == survivor[i].p) {
1394 VERB5 bb_error_msg("keeping old synced peer");
1395 p = G.last_update_peer;
1400 G.last_update_peer = p;
1402 VERB4 bb_error_msg("selected peer %s filter_offset:%+f age:%f",
1405 G.cur_time - p->lastpkt_recv_time
1412 * Local clock discipline and its helpers
1415 set_new_values(int disc_state, double offset, double recv_time)
1417 /* Enter new state and set state variables. Note we use the time
1418 * of the last clock filter sample, which must be earlier than
1421 VERB4 bb_error_msg("disc_state=%d last update offset=%f recv_time=%f",
1422 disc_state, offset, recv_time);
1423 G.discipline_state = disc_state;
1424 G.last_update_offset = offset;
1425 G.last_update_recv_time = recv_time;
1427 /* Return: -1: decrease poll interval, 0: leave as is, 1: increase */
1429 update_local_clock(peer_t *p)
1433 /* Note: can use G.cluster_offset instead: */
1434 double offset = p->filter_offset;
1435 double recv_time = p->lastpkt_recv_time;
1437 #if !USING_KERNEL_PLL_LOOP
1440 #if !USING_KERNEL_PLL_LOOP || USING_INITIAL_FREQ_ESTIMATION
1441 double since_last_update;
1443 double etemp, dtemp;
1445 abs_offset = fabs(offset);
1448 /* If needed, -S script can do it by looking at $offset
1449 * env var and killing parent */
1450 /* If the offset is too large, give up and go home */
1451 if (abs_offset > PANIC_THRESHOLD) {
1452 bb_error_msg_and_die("offset %f far too big, exiting", offset);
1456 /* If this is an old update, for instance as the result
1457 * of a system peer change, avoid it. We never use
1458 * an old sample or the same sample twice.
1460 if (recv_time <= G.last_update_recv_time) {
1461 VERB3 bb_error_msg("update from %s: same or older datapoint, not using it",
1463 return 0; /* "leave poll interval as is" */
1466 /* Clock state machine transition function. This is where the
1467 * action is and defines how the system reacts to large time
1468 * and frequency errors.
1470 #if !USING_KERNEL_PLL_LOOP || USING_INITIAL_FREQ_ESTIMATION
1471 since_last_update = recv_time - G.reftime;
1473 #if !USING_KERNEL_PLL_LOOP
1476 #if USING_INITIAL_FREQ_ESTIMATION
1477 if (G.discipline_state == STATE_FREQ) {
1478 /* Ignore updates until the stepout threshold */
1479 if (since_last_update < WATCH_THRESHOLD) {
1480 VERB4 bb_error_msg("measuring drift, datapoint ignored, %f sec remains",
1481 WATCH_THRESHOLD - since_last_update);
1482 return 0; /* "leave poll interval as is" */
1484 # if !USING_KERNEL_PLL_LOOP
1485 freq_drift = (offset - G.last_update_offset) / since_last_update;
1490 /* There are two main regimes: when the
1491 * offset exceeds the step threshold and when it does not.
1493 if (abs_offset > STEP_THRESHOLD) {
1497 // This "spike state" seems to be useless, peer selection already drops
1498 // occassional "bad" datapoints. If we are here, there were _many_
1499 // large offsets. When a few first large offsets are seen,
1500 // we end up in "no valid datapoints, no peer selected" state.
1501 // Only when enough of them are seen (which means it's not a fluke),
1502 // we end up here. Looks like _our_ clock is off.
1503 switch (G.discipline_state) {
1505 /* The first outlyer: ignore it, switch to SPIK state */
1506 VERB3 bb_error_msg("update from %s: offset:%+f, spike%s",
1507 p->p_dotted, offset,
1509 G.discipline_state = STATE_SPIK;
1510 return -1; /* "decrease poll interval" */
1513 /* Ignore succeeding outlyers until either an inlyer
1514 * is found or the stepout threshold is exceeded.
1516 remains = WATCH_THRESHOLD - since_last_update;
1518 VERB3 bb_error_msg("update from %s: offset:%+f, spike%s",
1519 p->p_dotted, offset,
1520 ", datapoint ignored");
1521 return -1; /* "decrease poll interval" */
1523 /* fall through: we need to step */
1527 /* Step the time and clamp down the poll interval.
1529 * In NSET state an initial frequency correction is
1530 * not available, usually because the frequency file has
1531 * not yet been written. Since the time is outside the
1532 * capture range, the clock is stepped. The frequency
1533 * will be set directly following the stepout interval.
1535 * In FSET state the initial frequency has been set
1536 * from the frequency file. Since the time is outside
1537 * the capture range, the clock is stepped immediately,
1538 * rather than after the stepout interval. Guys get
1539 * nervous if it takes 17 minutes to set the clock for
1542 * In SPIK state the stepout threshold has expired and
1543 * the phase is still above the step threshold. Note
1544 * that a single spike greater than the step threshold
1545 * is always suppressed, even at the longer poll
1548 VERB4 bb_error_msg("stepping time by %+f; poll_exp=MINPOLL", offset);
1550 if (option_mask32 & OPT_q) {
1551 /* We were only asked to set time once. Done. */
1555 clamp_pollexp_and_set_MAXSTRAT();
1557 run_script("step", offset);
1559 recv_time += offset;
1561 #if USING_INITIAL_FREQ_ESTIMATION
1562 if (G.discipline_state == STATE_NSET) {
1563 set_new_values(STATE_FREQ, /*offset:*/ 0, recv_time);
1564 return 1; /* "ok to increase poll interval" */
1567 abs_offset = offset = 0;
1568 set_new_values(STATE_SYNC, offset, recv_time);
1569 } else { /* abs_offset <= STEP_THRESHOLD */
1571 /* The ratio is calculated before jitter is updated to make
1572 * poll adjust code more sensitive to large offsets.
1574 G.offset_to_jitter_ratio = abs_offset / G.discipline_jitter;
1576 /* Compute the clock jitter as the RMS of exponentially
1577 * weighted offset differences. Used by the poll adjust code.
1579 etemp = SQUARE(G.discipline_jitter);
1580 dtemp = SQUARE(offset - G.last_update_offset);
1581 G.discipline_jitter = SQRT(etemp + (dtemp - etemp) / AVG);
1582 if (G.discipline_jitter < G_precision_sec)
1583 G.discipline_jitter = G_precision_sec;
1585 switch (G.discipline_state) {
1587 if (option_mask32 & OPT_q) {
1588 /* We were only asked to set time once.
1589 * The clock is precise enough, no need to step.
1593 #if USING_INITIAL_FREQ_ESTIMATION
1594 /* This is the first update received and the frequency
1595 * has not been initialized. The first thing to do
1596 * is directly measure the oscillator frequency.
1598 set_new_values(STATE_FREQ, offset, recv_time);
1600 set_new_values(STATE_SYNC, offset, recv_time);
1602 VERB4 bb_error_msg("transitioning to FREQ, datapoint ignored");
1603 return 0; /* "leave poll interval as is" */
1605 #if 0 /* this is dead code for now */
1607 /* This is the first update and the frequency
1608 * has been initialized. Adjust the phase, but
1609 * don't adjust the frequency until the next update.
1611 set_new_values(STATE_SYNC, offset, recv_time);
1612 /* freq_drift remains 0 */
1616 #if USING_INITIAL_FREQ_ESTIMATION
1618 /* since_last_update >= WATCH_THRESHOLD, we waited enough.
1619 * Correct the phase and frequency and switch to SYNC state.
1620 * freq_drift was already estimated (see code above)
1622 set_new_values(STATE_SYNC, offset, recv_time);
1627 #if !USING_KERNEL_PLL_LOOP
1628 /* Compute freq_drift due to PLL and FLL contributions.
1630 * The FLL and PLL frequency gain constants
1631 * depend on the poll interval and Allan
1632 * intercept. The FLL is not used below one-half
1633 * the Allan intercept. Above that the loop gain
1634 * increases in steps to 1 / AVG.
1636 if ((1 << G.poll_exp) > ALLAN / 2) {
1637 etemp = FLL - G.poll_exp;
1640 freq_drift += (offset - G.last_update_offset) / (MAXD(since_last_update, ALLAN) * etemp);
1642 /* For the PLL the integration interval
1643 * (numerator) is the minimum of the update
1644 * interval and poll interval. This allows
1645 * oversampling, but not undersampling.
1647 etemp = MIND(since_last_update, (1 << G.poll_exp));
1648 dtemp = (4 * PLL) << G.poll_exp;
1649 freq_drift += offset * etemp / SQUARE(dtemp);
1651 set_new_values(STATE_SYNC, offset, recv_time);
1654 if (G.stratum != p->lastpkt_stratum + 1) {
1655 G.stratum = p->lastpkt_stratum + 1;
1656 run_script("stratum", offset);
1660 G.reftime = G.cur_time;
1661 G.ntp_status = p->lastpkt_status;
1662 G.refid = p->lastpkt_refid;
1663 G.rootdelay = p->lastpkt_rootdelay + p->lastpkt_delay;
1664 dtemp = p->filter_jitter; // SQRT(SQUARE(p->filter_jitter) + SQUARE(G.cluster_jitter));
1665 dtemp += MAXD(p->filter_dispersion + FREQ_TOLERANCE * (G.cur_time - p->lastpkt_recv_time) + abs_offset, MINDISP);
1666 G.rootdisp = p->lastpkt_rootdisp + dtemp;
1667 VERB4 bb_error_msg("updating leap/refid/reftime/rootdisp from peer %s", p->p_dotted);
1669 /* We are in STATE_SYNC now, but did not do adjtimex yet.
1670 * (Any other state does not reach this, they all return earlier)
1671 * By this time, freq_drift and offset are set
1672 * to values suitable for adjtimex.
1674 #if !USING_KERNEL_PLL_LOOP
1675 /* Calculate the new frequency drift and frequency stability (wander).
1676 * Compute the clock wander as the RMS of exponentially weighted
1677 * frequency differences. This is not used directly, but can,
1678 * along with the jitter, be a highly useful monitoring and
1681 dtemp = G.discipline_freq_drift + freq_drift;
1682 G.discipline_freq_drift = MAXD(MIND(MAXDRIFT, dtemp), -MAXDRIFT);
1683 etemp = SQUARE(G.discipline_wander);
1684 dtemp = SQUARE(dtemp);
1685 G.discipline_wander = SQRT(etemp + (dtemp - etemp) / AVG);
1687 VERB4 bb_error_msg("discipline freq_drift=%.9f(int:%ld corr:%e) wander=%f",
1688 G.discipline_freq_drift,
1689 (long)(G.discipline_freq_drift * 65536e6),
1691 G.discipline_wander);
1694 memset(&tmx, 0, sizeof(tmx));
1695 if (adjtimex(&tmx) < 0)
1696 bb_perror_msg_and_die("adjtimex");
1697 bb_error_msg("p adjtimex freq:%ld offset:%+ld status:0x%x tc:%ld",
1698 tmx.freq, tmx.offset, tmx.status, tmx.constant);
1701 memset(&tmx, 0, sizeof(tmx));
1703 //doesn't work, offset remains 0 (!) in kernel:
1704 //ntpd: set adjtimex freq:1786097 tmx.offset:77487
1705 //ntpd: prev adjtimex freq:1786097 tmx.offset:0
1706 //ntpd: cur adjtimex freq:1786097 tmx.offset:0
1707 tmx.modes = ADJ_FREQUENCY | ADJ_OFFSET;
1708 /* 65536 is one ppm */
1709 tmx.freq = G.discipline_freq_drift * 65536e6;
1711 tmx.modes = ADJ_OFFSET | ADJ_STATUS | ADJ_TIMECONST;// | ADJ_MAXERROR | ADJ_ESTERROR;
1712 tmx.constant = (int)G.poll_exp - 4;
1714 * The below if statement should be unnecessary, but...
1715 * It looks like Linux kernel's PLL is far too gentle in changing
1716 * tmx.freq in response to clock offset. Offset keeps growing
1717 * and eventually we fall back to smaller poll intervals.
1718 * We can make correction more agressive (about x2) by supplying
1719 * PLL time constant which is one less than the real one.
1720 * To be on a safe side, let's do it only if offset is significantly
1721 * larger than jitter.
1723 if (G.offset_to_jitter_ratio >= TIMECONST_HACK_GATE)
1725 tmx.offset = (long)(offset * 1000000); /* usec */
1726 if (SLEW_THRESHOLD < STEP_THRESHOLD) {
1727 if (tmx.offset > (long)(SLEW_THRESHOLD * 1000000)) {
1728 tmx.offset = (long)(SLEW_THRESHOLD * 1000000);
1731 if (tmx.offset < -(long)(SLEW_THRESHOLD * 1000000)) {
1732 tmx.offset = -(long)(SLEW_THRESHOLD * 1000000);
1736 if (tmx.constant < 0)
1739 tmx.status = STA_PLL;
1740 if (G.ntp_status & LI_PLUSSEC)
1741 tmx.status |= STA_INS;
1742 if (G.ntp_status & LI_MINUSSEC)
1743 tmx.status |= STA_DEL;
1745 //tmx.esterror = (uint32_t)(clock_jitter * 1e6);
1746 //tmx.maxerror = (uint32_t)((sys_rootdelay / 2 + sys_rootdisp) * 1e6);
1747 rc = adjtimex(&tmx);
1749 bb_perror_msg_and_die("adjtimex");
1750 /* NB: here kernel returns constant == G.poll_exp, not == G.poll_exp - 4.
1751 * Not sure why. Perhaps it is normal.
1753 VERB4 bb_error_msg("adjtimex:%d freq:%ld offset:%+ld status:0x%x",
1754 rc, tmx.freq, tmx.offset, tmx.status);
1755 G.kernel_freq_drift = tmx.freq / 65536;
1756 VERB2 bb_error_msg("update from:%s offset:%+f delay:%f jitter:%f clock drift:%+.3fppm tc:%d",
1760 G.discipline_jitter,
1761 (double)tmx.freq / 65536,
1765 return 1; /* "ok to increase poll interval" */
1770 * We've got a new reply packet from a peer, process it
1774 poll_interval(int upper_bound)
1776 unsigned interval, r, mask;
1777 interval = 1 << G.poll_exp;
1778 if (interval > upper_bound)
1779 interval = upper_bound;
1780 mask = ((interval-1) >> 4) | 1;
1782 interval += r & mask; /* ~ random(0..1) * interval/16 */
1783 VERB4 bb_error_msg("chose poll interval:%u (poll_exp:%d)", interval, G.poll_exp);
1787 adjust_poll(int count)
1789 G.polladj_count += count;
1790 if (G.polladj_count > POLLADJ_LIMIT) {
1791 G.polladj_count = 0;
1792 if (G.poll_exp < MAXPOLL) {
1794 VERB4 bb_error_msg("polladj: discipline_jitter:%f ++poll_exp=%d",
1795 G.discipline_jitter, G.poll_exp);
1797 } else if (G.polladj_count < -POLLADJ_LIMIT || (count < 0 && G.poll_exp > BIGPOLL)) {
1798 G.polladj_count = 0;
1799 if (G.poll_exp > MINPOLL) {
1803 /* Correct p->next_action_time in each peer
1804 * which waits for sending, so that they send earlier.
1805 * Old pp->next_action_time are on the order
1806 * of t + (1 << old_poll_exp) + small_random,
1807 * we simply need to subtract ~half of that.
1809 for (item = G.ntp_peers; item != NULL; item = item->link) {
1810 peer_t *pp = (peer_t *) item->data;
1812 pp->next_action_time -= (1 << G.poll_exp);
1814 VERB4 bb_error_msg("polladj: discipline_jitter:%f --poll_exp=%d",
1815 G.discipline_jitter, G.poll_exp);
1818 VERB4 bb_error_msg("polladj: count:%d", G.polladj_count);
1821 static NOINLINE void
1822 recv_and_process_peer_pkt(peer_t *p)
1827 double T1, T2, T3, T4;
1829 double prev_delay, delay;
1831 datapoint_t *datapoint;
1836 /* We can recvfrom here and check from.IP, but some multihomed
1837 * ntp servers reply from their *other IP*.
1838 * TODO: maybe we should check at least what we can: from.port == 123?
1841 size = recv(p->p_fd, &msg, sizeof(msg), MSG_DONTWAIT);
1846 if (errno == EAGAIN)
1847 /* There was no packet after all
1848 * (poll() returning POLLIN for a fd
1849 * is not a ironclad guarantee that data is there)
1853 * If you need a different handling for a specific
1854 * errno, always explain it in comment.
1856 bb_perror_msg_and_die("recv(%s) error", p->p_dotted);
1859 if (size != NTP_MSGSIZE_NOAUTH && size != NTP_MSGSIZE) {
1860 bb_error_msg("malformed packet received from %s", p->p_dotted);
1864 if (msg.m_orgtime.int_partl != p->p_xmt_msg.m_xmttime.int_partl
1865 || msg.m_orgtime.fractionl != p->p_xmt_msg.m_xmttime.fractionl
1867 /* Somebody else's packet */
1871 /* We do not expect any more packets from this peer for now.
1872 * Closing the socket informs kernel about it.
1873 * We open a new socket when we send a new query.
1878 if ((msg.m_status & LI_ALARM) == LI_ALARM
1879 || msg.m_stratum == 0
1880 || msg.m_stratum > NTP_MAXSTRATUM
1882 bb_error_msg("reply from %s: peer is unsynced", p->p_dotted);
1884 * Stratum 0 responses may have commands in 32-bit m_refid field:
1885 * "DENY", "RSTR" - peer does not like us at all,
1886 * "RATE" - peer is overloaded, reduce polling freq.
1887 * If poll interval is small, increase it.
1889 if (G.poll_exp < BIGPOLL)
1890 goto increase_interval;
1891 goto pick_normal_interval;
1894 // /* Verify valid root distance */
1895 // if (msg.m_rootdelay / 2 + msg.m_rootdisp >= MAXDISP || p->lastpkt_reftime > msg.m_xmt)
1896 // return; /* invalid header values */
1899 * From RFC 2030 (with a correction to the delay math):
1901 * Timestamp Name ID When Generated
1902 * ------------------------------------------------------------
1903 * Originate Timestamp T1 time request sent by client
1904 * Receive Timestamp T2 time request received by server
1905 * Transmit Timestamp T3 time reply sent by server
1906 * Destination Timestamp T4 time reply received by client
1908 * The roundtrip delay and local clock offset are defined as
1910 * delay = (T4 - T1) - (T3 - T2); offset = ((T2 - T1) + (T3 - T4)) / 2
1913 T2 = lfp_to_d(msg.m_rectime);
1914 T3 = lfp_to_d(msg.m_xmttime);
1917 /* The delay calculation is a special case. In cases where the
1918 * server and client clocks are running at different rates and
1919 * with very fast networks, the delay can appear negative. In
1920 * order to avoid violating the Principle of Least Astonishment,
1921 * the delay is clamped not less than the system precision.
1923 delay = (T4 - T1) - (T3 - T2);
1924 if (delay < G_precision_sec)
1925 delay = G_precision_sec;
1927 * If this packet's delay is much bigger than the last one,
1928 * it's better to just ignore it than use its much less precise value.
1930 prev_delay = p->p_raw_delay;
1931 p->p_raw_delay = delay;
1932 if (p->reachable_bits && delay > prev_delay * BAD_DELAY_GROWTH) {
1933 bb_error_msg("reply from %s: delay %f is too high, ignoring", p->p_dotted, delay);
1934 goto pick_normal_interval;
1937 p->lastpkt_delay = delay;
1938 p->lastpkt_recv_time = T4;
1939 VERB6 bb_error_msg("%s->lastpkt_recv_time=%f", p->p_dotted, p->lastpkt_recv_time);
1940 p->lastpkt_status = msg.m_status;
1941 p->lastpkt_stratum = msg.m_stratum;
1942 p->lastpkt_rootdelay = sfp_to_d(msg.m_rootdelay);
1943 p->lastpkt_rootdisp = sfp_to_d(msg.m_rootdisp);
1944 p->lastpkt_refid = msg.m_refid;
1946 p->datapoint_idx = p->reachable_bits ? (p->datapoint_idx + 1) % NUM_DATAPOINTS : 0;
1947 datapoint = &p->filter_datapoint[p->datapoint_idx];
1948 datapoint->d_recv_time = T4;
1949 datapoint->d_offset = offset = ((T2 - T1) + (T3 - T4)) / 2;
1950 datapoint->d_dispersion = LOG2D(msg.m_precision_exp) + G_precision_sec;
1951 if (!p->reachable_bits) {
1952 /* 1st datapoint ever - replicate offset in every element */
1954 for (i = 0; i < NUM_DATAPOINTS; i++) {
1955 p->filter_datapoint[i].d_offset = offset;
1959 p->reachable_bits |= 1;
1960 if ((MAX_VERBOSE && G.verbose) || (option_mask32 & OPT_w)) {
1961 bb_error_msg("reply from %s: offset:%+f delay:%f status:0x%02x strat:%d refid:0x%08x rootdelay:%f reach:0x%02x",
1968 p->lastpkt_rootdelay,
1970 /* not shown: m_ppoll, m_precision_exp, m_rootdisp,
1971 * m_reftime, m_orgtime, m_rectime, m_xmttime
1976 /* Muck with statictics and update the clock */
1977 filter_datapoints(p);
1978 q = select_and_cluster();
1981 if (!(option_mask32 & OPT_w)) {
1982 rc = update_local_clock(q);
1984 //Disabled this because there is a case where largish offsets
1985 //are unavoidable: if network round-trip delay is, say, ~0.6s,
1986 //error in offset estimation would be ~delay/2 ~= 0.3s.
1987 //Thus, offsets will be usually in -0.3...0.3s range.
1988 //In this case, this code would keep poll interval small,
1989 //but it won't be helping.
1990 //BIGOFF check below deals with a case of seeing multi-second offsets.
1992 /* If drift is dangerously large, immediately
1993 * drop poll interval one step down.
1995 if (fabs(q->filter_offset) >= POLLDOWN_OFFSET) {
1996 VERB4 bb_error_msg("offset:%+f > POLLDOWN_OFFSET", q->filter_offset);
1997 adjust_poll(-POLLADJ_LIMIT * 3);
2003 /* No peer selected.
2004 * If poll interval is small, increase it.
2006 if (G.poll_exp < BIGPOLL)
2007 goto increase_interval;
2011 /* Adjust the poll interval by comparing the current offset
2012 * with the clock jitter. If the offset is less than
2013 * the clock jitter times a constant, then the averaging interval
2014 * is increased, otherwise it is decreased. A bit of hysteresis
2015 * helps calm the dance. Works best using burst mode.
2017 if (rc > 0 && G.offset_to_jitter_ratio <= POLLADJ_GATE) {
2018 /* was += G.poll_exp but it is a bit
2019 * too optimistic for my taste at high poll_exp's */
2021 adjust_poll(MINPOLL);
2024 bb_error_msg("want smaller interval: offset/jitter = %u",
2025 G.offset_to_jitter_ratio);
2026 adjust_poll(-G.poll_exp * 2);
2030 /* Decide when to send new query for this peer */
2031 pick_normal_interval:
2032 interval = poll_interval(INT_MAX);
2033 if (fabs(offset) >= BIGOFF && interval > BIGOFF_INTERVAL) {
2034 /* If we are synced, offsets are less than SLEW_THRESHOLD,
2035 * or at the very least not much larger than it.
2036 * Now we see a largish one.
2037 * Either this peer is feeling bad, or packet got corrupted,
2038 * or _our_ clock is wrong now and _all_ peers will show similar
2039 * largish offsets too.
2040 * I observed this with laptop suspend stopping clock.
2041 * In any case, it makes sense to make next request soonish:
2042 * cases 1 and 2: get a better datapoint,
2043 * case 3: allows to resync faster.
2045 interval = BIGOFF_INTERVAL;
2048 set_next(p, interval);
2051 #if ENABLE_FEATURE_NTPD_SERVER
2052 static NOINLINE void
2053 recv_and_process_client_pkt(void /*int fd*/)
2057 len_and_sockaddr *to;
2058 struct sockaddr *from;
2060 uint8_t query_status;
2061 l_fixedpt_t query_xmttime;
2063 to = get_sock_lsa(G_listen_fd);
2064 from = xzalloc(to->len);
2066 size = recv_from_to(G_listen_fd, &msg, sizeof(msg), MSG_DONTWAIT, from, &to->u.sa, to->len);
2067 if (size != NTP_MSGSIZE_NOAUTH && size != NTP_MSGSIZE) {
2070 if (errno == EAGAIN)
2072 bb_perror_msg_and_die("recv");
2074 addr = xmalloc_sockaddr2dotted_noport(from);
2075 bb_error_msg("malformed packet received from %s: size %u", addr, (int)size);
2080 /* Respond only to client and symmetric active packets */
2081 if ((msg.m_status & MODE_MASK) != MODE_CLIENT
2082 && (msg.m_status & MODE_MASK) != MODE_SYM_ACT
2087 query_status = msg.m_status;
2088 query_xmttime = msg.m_xmttime;
2090 /* Build a reply packet */
2091 memset(&msg, 0, sizeof(msg));
2092 msg.m_status = G.stratum < MAXSTRAT ? (G.ntp_status & LI_MASK) : LI_ALARM;
2093 msg.m_status |= (query_status & VERSION_MASK);
2094 msg.m_status |= ((query_status & MODE_MASK) == MODE_CLIENT) ?
2095 MODE_SERVER : MODE_SYM_PAS;
2096 msg.m_stratum = G.stratum;
2097 msg.m_ppoll = G.poll_exp;
2098 msg.m_precision_exp = G_precision_exp;
2099 /* this time was obtained between poll() and recv() */
2100 msg.m_rectime = d_to_lfp(G.cur_time);
2101 msg.m_xmttime = d_to_lfp(gettime1900d()); /* this instant */
2102 if (G.peer_cnt == 0) {
2103 /* we have no peers: "stratum 1 server" mode. reftime = our own time */
2104 G.reftime = G.cur_time;
2106 msg.m_reftime = d_to_lfp(G.reftime);
2107 msg.m_orgtime = query_xmttime;
2108 msg.m_rootdelay = d_to_sfp(G.rootdelay);
2109 //simple code does not do this, fix simple code!
2110 msg.m_rootdisp = d_to_sfp(G.rootdisp);
2111 //version = (query_status & VERSION_MASK); /* ... >> VERSION_SHIFT - done below instead */
2112 msg.m_refid = G.refid; // (version > (3 << VERSION_SHIFT)) ? G.refid : G.refid3;
2114 /* We reply from the local address packet was sent to,
2115 * this makes to/from look swapped here: */
2116 do_sendto(G_listen_fd,
2117 /*from:*/ &to->u.sa, /*to:*/ from, /*addrlen:*/ to->len,
2126 /* Upstream ntpd's options:
2128 * -4 Force DNS resolution of host names to the IPv4 namespace.
2129 * -6 Force DNS resolution of host names to the IPv6 namespace.
2130 * -a Require cryptographic authentication for broadcast client,
2131 * multicast client and symmetric passive associations.
2132 * This is the default.
2133 * -A Do not require cryptographic authentication for broadcast client,
2134 * multicast client and symmetric passive associations.
2135 * This is almost never a good idea.
2136 * -b Enable the client to synchronize to broadcast servers.
2138 * Specify the name and path of the configuration file,
2139 * default /etc/ntp.conf
2140 * -d Specify debugging mode. This option may occur more than once,
2141 * with each occurrence indicating greater detail of display.
2143 * Specify debugging level directly.
2145 * Specify the name and path of the frequency file.
2146 * This is the same operation as the "driftfile FILE"
2147 * configuration command.
2148 * -g Normally, ntpd exits with a message to the system log
2149 * if the offset exceeds the panic threshold, which is 1000 s
2150 * by default. This option allows the time to be set to any value
2151 * without restriction; however, this can happen only once.
2152 * If the threshold is exceeded after that, ntpd will exit
2153 * with a message to the system log. This option can be used
2154 * with the -q and -x options. See the tinker command for other options.
2156 * Chroot the server to the directory jaildir. This option also implies
2157 * that the server attempts to drop root privileges at startup
2158 * (otherwise, chroot gives very little additional security).
2159 * You may need to also specify a -u option.
2161 * Specify the name and path of the symmetric key file,
2162 * default /etc/ntp/keys. This is the same operation
2163 * as the "keys FILE" configuration command.
2165 * Specify the name and path of the log file. The default
2166 * is the system log file. This is the same operation as
2167 * the "logfile FILE" configuration command.
2168 * -L Do not listen to virtual IPs. The default is to listen.
2170 * -N To the extent permitted by the operating system,
2171 * run the ntpd at the highest priority.
2173 * Specify the name and path of the file used to record the ntpd
2174 * process ID. This is the same operation as the "pidfile FILE"
2175 * configuration command.
2177 * To the extent permitted by the operating system,
2178 * run the ntpd at the specified priority.
2179 * -q Exit the ntpd just after the first time the clock is set.
2180 * This behavior mimics that of the ntpdate program, which is
2181 * to be retired. The -g and -x options can be used with this option.
2182 * Note: The kernel time discipline is disabled with this option.
2184 * Specify the default propagation delay from the broadcast/multicast
2185 * server to this client. This is necessary only if the delay
2186 * cannot be computed automatically by the protocol.
2188 * Specify the directory path for files created by the statistics
2189 * facility. This is the same operation as the "statsdir DIR"
2190 * configuration command.
2192 * Add a key number to the trusted key list. This option can occur
2195 * Specify a user, and optionally a group, to switch to.
2198 * Add a system variable listed by default.
2199 * -x Normally, the time is slewed if the offset is less than the step
2200 * threshold, which is 128 ms by default, and stepped if above
2201 * the threshold. This option sets the threshold to 600 s, which is
2202 * well within the accuracy window to set the clock manually.
2203 * Note: since the slew rate of typical Unix kernels is limited
2204 * to 0.5 ms/s, each second of adjustment requires an amortization
2205 * interval of 2000 s. Thus, an adjustment as much as 600 s
2206 * will take almost 14 days to complete. This option can be used
2207 * with the -g and -q options. See the tinker command for other options.
2208 * Note: The kernel time discipline is disabled with this option.
2211 /* By doing init in a separate function we decrease stack usage
2214 static NOINLINE void ntp_init(char **argv)
2222 bb_error_msg_and_die(bb_msg_you_must_be_root);
2224 /* Set some globals */
2225 G.discipline_jitter = G_precision_sec;
2226 G.stratum = MAXSTRAT;
2228 G.poll_exp = BURSTPOLL; /* speeds up initial sync */
2229 G.last_script_run = G.reftime = G.last_update_recv_time = gettime1900d(); /* sets G.cur_time too */
2233 opt_complementary = "dd:wn" /* -d: counter; -p: list; -w implies -n */
2234 IF_FEATURE_NTPD_SERVER(":Il"); /* -I implies -l */
2235 opts = getopt32(argv,
2237 "wp:*S:"IF_FEATURE_NTPD_SERVER("l") /* NOT compat */
2238 IF_FEATURE_NTPD_SERVER("I:") /* compat */
2240 "46aAbgL", /* compat, ignored */
2241 &peers,&G.script_name,
2242 #if ENABLE_FEATURE_NTPD_SERVER
2247 // if (opts & OPT_x) /* disable stepping, only slew is allowed */
2248 // G.time_was_stepped = 1;
2250 #if ENABLE_FEATURE_NTPD_SERVER
2253 G_listen_fd = create_and_bind_dgram_or_die(NULL, 123);
2255 if (setsockopt_bindtodevice(G_listen_fd, G.if_name))
2258 socket_want_pktinfo(G_listen_fd);
2259 setsockopt_int(G_listen_fd, IPPROTO_IP, IP_TOS, IPTOS_LOWDELAY);
2262 /* I hesitate to set -20 prio. -15 should be high enough for timekeeping */
2264 setpriority(PRIO_PROCESS, 0, -15);
2266 /* add_peers() calls can retry DNS resolution (possibly forever).
2267 * Daemonize before them, or else boot can stall forever.
2269 if (!(opts & OPT_n)) {
2270 bb_daemonize_or_rexec(DAEMON_DEVNULL_STDIO, argv);
2271 logmode = LOGMODE_NONE;
2276 add_peers(llist_pop(&peers));
2278 #if ENABLE_FEATURE_NTPD_CONF
2283 parser = config_open("/etc/ntp.conf");
2284 while (config_read(parser, token, 3, 1, "# \t", PARSE_NORMAL)) {
2285 if (strcmp(token[0], "server") == 0 && token[1]) {
2286 add_peers(token[1]);
2289 bb_error_msg("skipping %s:%u: unimplemented command '%s'",
2290 "/etc/ntp.conf", parser->lineno, token[0]
2293 config_close(parser);
2296 if (G.peer_cnt == 0) {
2297 if (!(opts & OPT_l))
2299 /* -l but no peers: "stratum 1 server" mode */
2302 /* If network is up, syncronization occurs in ~10 seconds.
2303 * We give "ntpd -q" 10 seconds to get first reply,
2304 * then another 50 seconds to finish syncing.
2306 * I tested ntpd 4.2.6p1 and apparently it never exits
2307 * (will try forever), but it does not feel right.
2308 * The goal of -q is to act like ntpdate: set time
2309 * after a reasonably small period of polling, or fail.
2312 option_mask32 |= OPT_qq;
2329 int ntpd_main(int argc UNUSED_PARAM, char **argv) MAIN_EXTERNALLY_VISIBLE;
2330 int ntpd_main(int argc UNUSED_PARAM, char **argv)
2338 memset(&G, 0, sizeof(G));
2339 SET_PTR_TO_GLOBALS(&G);
2343 /* If ENABLE_FEATURE_NTPD_SERVER, + 1 for listen_fd: */
2344 cnt = G.peer_cnt + ENABLE_FEATURE_NTPD_SERVER;
2345 idx2peer = xzalloc(sizeof(idx2peer[0]) * cnt);
2346 pfd = xzalloc(sizeof(pfd[0]) * cnt);
2348 /* Countdown: we never sync before we sent INITIAL_SAMPLES+1
2349 * packets to each peer.
2350 * NB: if some peer is not responding, we may end up sending
2351 * fewer packets to it and more to other peers.
2352 * NB2: sync usually happens using INITIAL_SAMPLES packets,
2353 * since last reply does not come back instantaneously.
2355 cnt = G.peer_cnt * (INITIAL_SAMPLES + 1);
2357 write_pidfile(CONFIG_PID_FILE_PATH "/ntpd.pid");
2359 while (!bb_got_signal) {
2365 /* Nothing between here and poll() blocks for any significant time */
2367 nextaction = G.cur_time + 3600;
2370 #if ENABLE_FEATURE_NTPD_SERVER
2371 if (G_listen_fd != -1) {
2372 pfd[0].fd = G_listen_fd;
2373 pfd[0].events = POLLIN;
2377 /* Pass over peer list, send requests, time out on receives */
2378 for (item = G.ntp_peers; item != NULL; item = item->link) {
2379 peer_t *p = (peer_t *) item->data;
2381 if (p->next_action_time <= G.cur_time) {
2382 if (p->p_fd == -1) {
2383 /* Time to send new req */
2385 VERB4 bb_error_msg("disabling burst mode");
2386 G.polladj_count = 0;
2387 G.poll_exp = MINPOLL;
2389 send_query_to_peer(p);
2391 /* Timed out waiting for reply */
2394 /* If poll interval is small, increase it */
2395 if (G.poll_exp < BIGPOLL)
2396 adjust_poll(MINPOLL);
2397 timeout = poll_interval(NOREPLY_INTERVAL);
2398 bb_error_msg("timed out waiting for %s, reach 0x%02x, next query in %us",
2399 p->p_dotted, p->reachable_bits, timeout);
2401 /* What if don't see it because it changed its IP? */
2402 if (p->reachable_bits == 0)
2403 resolve_peer_hostname(p, /*loop_on_fail=*/ 0);
2405 set_next(p, timeout);
2409 if (p->next_action_time < nextaction)
2410 nextaction = p->next_action_time;
2413 /* Wait for reply from this peer */
2414 pfd[i].fd = p->p_fd;
2415 pfd[i].events = POLLIN;
2421 timeout = nextaction - G.cur_time;
2424 timeout++; /* (nextaction - G.cur_time) rounds down, compensating */
2426 /* Here we may block */
2428 if (i > (ENABLE_FEATURE_NTPD_SERVER && G_listen_fd != -1)) {
2429 /* We wait for at least one reply.
2430 * Poll for it, without wasting time for message.
2431 * Since replies often come under 1 second, this also
2432 * reduces clutter in logs.
2434 nfds = poll(pfd, i, 1000);
2440 bb_error_msg("poll:%us sockets:%u interval:%us", timeout, i, 1 << G.poll_exp);
2442 nfds = poll(pfd, i, timeout * 1000);
2444 gettime1900d(); /* sets G.cur_time */
2446 if (!bb_got_signal /* poll wasn't interrupted by a signal */
2447 && G.cur_time - G.last_script_run > 11*60
2449 /* Useful for updating battery-backed RTC and such */
2450 run_script("periodic", G.last_update_offset);
2451 gettime1900d(); /* sets G.cur_time */
2456 /* Process any received packets */
2458 #if ENABLE_FEATURE_NTPD_SERVER
2459 if (G.listen_fd != -1) {
2460 if (pfd[0].revents /* & (POLLIN|POLLERR)*/) {
2462 recv_and_process_client_pkt(/*G.listen_fd*/);
2463 gettime1900d(); /* sets G.cur_time */
2468 for (; nfds != 0 && j < i; j++) {
2469 if (pfd[j].revents /* & (POLLIN|POLLERR)*/) {
2471 * At init, alarm was set to 10 sec.
2472 * Now we did get a reply.
2473 * Increase timeout to 50 seconds to finish syncing.
2475 if (option_mask32 & OPT_qq) {
2476 option_mask32 &= ~OPT_qq;
2480 recv_and_process_peer_pkt(idx2peer[j]);
2481 gettime1900d(); /* sets G.cur_time */
2486 if (G.ntp_peers && G.stratum != MAXSTRAT) {
2487 for (item = G.ntp_peers; item != NULL; item = item->link) {
2488 peer_t *p = (peer_t *) item->data;
2489 if (p->reachable_bits)
2490 goto have_reachable_peer;
2492 /* No peer responded for last 8 packets, panic */
2493 clamp_pollexp_and_set_MAXSTRAT();
2494 run_script("unsync", 0.0);
2495 have_reachable_peer: ;
2497 } /* while (!bb_got_signal) */
2499 remove_pidfile(CONFIG_PID_FILE_PATH "/ntpd.pid");
2500 kill_myself_with_sig(bb_got_signal);
2508 /*** openntpd-4.6 uses only adjtime, not adjtimex ***/
2510 /*** ntp-4.2.6/ntpd/ntp_loopfilter.c - adjtimex usage ***/
2514 direct_freq(double fp_offset)
2518 * If the kernel is enabled, we need the residual offset to
2519 * calculate the frequency correction.
2521 if (pll_control && kern_enable) {
2522 memset(&ntv, 0, sizeof(ntv));
2525 clock_offset = ntv.offset / 1e9;
2526 #else /* STA_NANO */
2527 clock_offset = ntv.offset / 1e6;
2528 #endif /* STA_NANO */
2529 drift_comp = FREQTOD(ntv.freq);
2531 #endif /* KERNEL_PLL */
2532 set_freq((fp_offset - clock_offset) / (current_time - clock_epoch) + drift_comp);
2538 set_freq(double freq) /* frequency update */
2546 * If the kernel is enabled, update the kernel frequency.
2548 if (pll_control && kern_enable) {
2549 memset(&ntv, 0, sizeof(ntv));
2550 ntv.modes = MOD_FREQUENCY;
2551 ntv.freq = DTOFREQ(drift_comp);
2553 snprintf(tbuf, sizeof(tbuf), "kernel %.3f PPM", drift_comp * 1e6);
2554 report_event(EVNT_FSET, NULL, tbuf);
2556 snprintf(tbuf, sizeof(tbuf), "ntpd %.3f PPM", drift_comp * 1e6);
2557 report_event(EVNT_FSET, NULL, tbuf);
2559 #else /* KERNEL_PLL */
2560 snprintf(tbuf, sizeof(tbuf), "ntpd %.3f PPM", drift_comp * 1e6);
2561 report_event(EVNT_FSET, NULL, tbuf);
2562 #endif /* KERNEL_PLL */
2571 * This code segment works when clock adjustments are made using
2572 * precision time kernel support and the ntp_adjtime() system
2573 * call. This support is available in Solaris 2.6 and later,
2574 * Digital Unix 4.0 and later, FreeBSD, Linux and specially
2575 * modified kernels for HP-UX 9 and Ultrix 4. In the case of the
2576 * DECstation 5000/240 and Alpha AXP, additional kernel
2577 * modifications provide a true microsecond clock and nanosecond
2578 * clock, respectively.
2580 * Important note: The kernel discipline is used only if the
2581 * step threshold is less than 0.5 s, as anything higher can
2582 * lead to overflow problems. This might occur if some misguided
2583 * lad set the step threshold to something ridiculous.
2585 if (pll_control && kern_enable) {
2587 #define MOD_BITS (MOD_OFFSET | MOD_MAXERROR | MOD_ESTERROR | MOD_STATUS | MOD_TIMECONST)
2590 * We initialize the structure for the ntp_adjtime()
2591 * system call. We have to convert everything to
2592 * microseconds or nanoseconds first. Do not update the
2593 * system variables if the ext_enable flag is set. In
2594 * this case, the external clock driver will update the
2595 * variables, which will be read later by the local
2596 * clock driver. Afterwards, remember the time and
2597 * frequency offsets for jitter and stability values and
2598 * to update the frequency file.
2600 memset(&ntv, 0, sizeof(ntv));
2602 ntv.modes = MOD_STATUS;
2605 ntv.modes = MOD_BITS | MOD_NANO;
2606 #else /* STA_NANO */
2607 ntv.modes = MOD_BITS;
2608 #endif /* STA_NANO */
2609 if (clock_offset < 0)
2614 ntv.offset = (int32)(clock_offset * 1e9 + dtemp);
2615 ntv.constant = sys_poll;
2616 #else /* STA_NANO */
2617 ntv.offset = (int32)(clock_offset * 1e6 + dtemp);
2618 ntv.constant = sys_poll - 4;
2619 #endif /* STA_NANO */
2620 ntv.esterror = (u_int32)(clock_jitter * 1e6);
2621 ntv.maxerror = (u_int32)((sys_rootdelay / 2 + sys_rootdisp) * 1e6);
2622 ntv.status = STA_PLL;
2625 * Enable/disable the PPS if requested.
2628 if (!(pll_status & STA_PPSTIME))
2629 report_event(EVNT_KERN,
2630 NULL, "PPS enabled");
2631 ntv.status |= STA_PPSTIME | STA_PPSFREQ;
2633 if (pll_status & STA_PPSTIME)
2634 report_event(EVNT_KERN,
2635 NULL, "PPS disabled");
2636 ntv.status &= ~(STA_PPSTIME | STA_PPSFREQ);
2638 if (sys_leap == LEAP_ADDSECOND)
2639 ntv.status |= STA_INS;
2640 else if (sys_leap == LEAP_DELSECOND)
2641 ntv.status |= STA_DEL;
2645 * Pass the stuff to the kernel. If it squeals, turn off
2646 * the pps. In any case, fetch the kernel offset,
2647 * frequency and jitter.
2649 if (ntp_adjtime(&ntv) == TIME_ERROR) {
2650 if (!(ntv.status & STA_PPSSIGNAL))
2651 report_event(EVNT_KERN, NULL,
2654 pll_status = ntv.status;
2656 clock_offset = ntv.offset / 1e9;
2657 #else /* STA_NANO */
2658 clock_offset = ntv.offset / 1e6;
2659 #endif /* STA_NANO */
2660 clock_frequency = FREQTOD(ntv.freq);
2663 * If the kernel PPS is lit, monitor its performance.
2665 if (ntv.status & STA_PPSTIME) {
2667 clock_jitter = ntv.jitter / 1e9;
2668 #else /* STA_NANO */
2669 clock_jitter = ntv.jitter / 1e6;
2670 #endif /* STA_NANO */
2673 #if defined(STA_NANO) && NTP_API == 4
2675 * If the TAI changes, update the kernel TAI.
2677 if (loop_tai != sys_tai) {
2679 ntv.modes = MOD_TAI;
2680 ntv.constant = sys_tai;
2683 #endif /* STA_NANO */
2685 #endif /* KERNEL_PLL */