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 //config: bool "ntpd (17 kb)"
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 */
158 #define HOSTNAME_INTERVAL 4 /* hostname lookup failed. Wait N * peer->dns_errors secs for next try */
159 #define DNS_ERRORS_CAP 0x3f /* peer->dns_errors is in [0..63] */
161 /* Step threshold (sec). std ntpd uses 0.128.
163 #define STEP_THRESHOLD 1
164 /* Slew threshold (sec): adjtimex() won't accept offsets larger than this.
165 * Using exact power of 2 (1/8) results in smaller code
167 #define SLEW_THRESHOLD 0.125
168 /* Stepout threshold (sec). std ntpd uses 900 (11 mins (!)) */
169 #define WATCH_THRESHOLD 128
170 /* NB: set WATCH_THRESHOLD to ~60 when debugging to save time) */
171 //UNUSED: #define PANIC_THRESHOLD 1000 /* panic threshold (sec) */
174 * If we got |offset| > BIGOFF from a peer, cap next query interval
175 * for this peer by this many seconds:
177 #define BIGOFF STEP_THRESHOLD
178 #define BIGOFF_INTERVAL (1 << 7) /* 128 s */
180 #define FREQ_TOLERANCE 0.000015 /* frequency tolerance (15 PPM) */
181 #define BURSTPOLL 0 /* initial poll */
182 #define MINPOLL 5 /* minimum poll interval. std ntpd uses 6 (6: 64 sec) */
184 * If offset > discipline_jitter * POLLADJ_GATE, and poll interval is > 2^BIGPOLL,
185 * then it is decreased _at once_. (If <= 2^BIGPOLL, it will be decreased _eventually_).
187 #define BIGPOLL 9 /* 2^9 sec ~= 8.5 min */
188 #define MAXPOLL 12 /* maximum poll interval (12: 1.1h, 17: 36.4h). std ntpd uses 17 */
190 * Actively lower poll when we see such big offsets.
191 * With SLEW_THRESHOLD = 0.125, it means we try to sync more aggressively
192 * if offset increases over ~0.04 sec
194 //#define POLLDOWN_OFFSET (SLEW_THRESHOLD / 3)
195 #define MINDISP 0.01 /* minimum dispersion (sec) */
196 #define MAXDISP 16 /* maximum dispersion (sec) */
197 #define MAXSTRAT 16 /* maximum stratum (infinity metric) */
198 #define MAXDIST 1 /* distance threshold (sec) */
199 #define MIN_SELECTED 1 /* minimum intersection survivors */
200 #define MIN_CLUSTERED 3 /* minimum cluster survivors */
202 #define MAXDRIFT 0.000500 /* frequency drift we can correct (500 PPM) */
204 /* Poll-adjust threshold.
205 * When we see that offset is small enough compared to discipline jitter,
206 * we grow a counter: += MINPOLL. When counter goes over POLLADJ_LIMIT,
207 * we poll_exp++. If offset isn't small, counter -= poll_exp*2,
208 * and when it goes below -POLLADJ_LIMIT, we poll_exp--.
209 * (Bumped from 30 to 40 since otherwise I often see poll_exp going *2* steps down)
211 #define POLLADJ_LIMIT 40
212 /* If offset < discipline_jitter * POLLADJ_GATE, then we decide to increase
213 * poll interval (we think we can't improve timekeeping
214 * by staying at smaller poll).
216 #define POLLADJ_GATE 4
217 #define TIMECONST_HACK_GATE 2
218 /* Compromise Allan intercept (sec). doc uses 1500, std ntpd uses 512 */
222 /* FLL loop gain [why it depends on MAXPOLL??] */
223 #define FLL (MAXPOLL + 1)
224 /* Parameter averaging constant */
233 NTP_MSGSIZE_NOAUTH = 48,
234 NTP_MSGSIZE = (NTP_MSGSIZE_NOAUTH + 4 + NTP_DIGESTSIZE),
237 MODE_MASK = (7 << 0),
238 VERSION_MASK = (7 << 3),
242 /* Leap Second Codes (high order two bits of m_status) */
243 LI_NOWARNING = (0 << 6), /* no warning */
244 LI_PLUSSEC = (1 << 6), /* add a second (61 seconds) */
245 LI_MINUSSEC = (2 << 6), /* minus a second (59 seconds) */
246 LI_ALARM = (3 << 6), /* alarm condition */
249 MODE_RES0 = 0, /* reserved */
250 MODE_SYM_ACT = 1, /* symmetric active */
251 MODE_SYM_PAS = 2, /* symmetric passive */
252 MODE_CLIENT = 3, /* client */
253 MODE_SERVER = 4, /* server */
254 MODE_BROADCAST = 5, /* broadcast */
255 MODE_RES1 = 6, /* reserved for NTP control message */
256 MODE_RES2 = 7, /* reserved for private use */
259 //TODO: better base selection
260 #define OFFSET_1900_1970 2208988800UL /* 1970 - 1900 in seconds */
262 #define NUM_DATAPOINTS 8
275 uint8_t m_status; /* status of local clock and leap info */
277 uint8_t m_ppoll; /* poll value */
278 int8_t m_precision_exp;
279 s_fixedpt_t m_rootdelay;
280 s_fixedpt_t m_rootdisp;
282 l_fixedpt_t m_reftime;
283 l_fixedpt_t m_orgtime;
284 l_fixedpt_t m_rectime;
285 l_fixedpt_t m_xmttime;
287 uint8_t m_digest[NTP_DIGESTSIZE];
297 len_and_sockaddr *p_lsa;
301 uint32_t lastpkt_refid;
302 uint8_t lastpkt_status;
303 uint8_t lastpkt_stratum;
304 uint8_t reachable_bits;
306 /* when to send new query (if p_fd == -1)
307 * or when receive times out (if p_fd >= 0): */
308 double next_action_time;
311 /* p_raw_delay is set even by "high delay" packets */
312 /* lastpkt_delay isn't */
313 double lastpkt_recv_time;
314 double lastpkt_delay;
315 double lastpkt_rootdelay;
316 double lastpkt_rootdisp;
317 /* produced by filter algorithm: */
318 double filter_offset;
319 double filter_dispersion;
320 double filter_jitter;
321 datapoint_t filter_datapoint[NUM_DATAPOINTS];
322 /* last sent packet: */
328 #define USING_KERNEL_PLL_LOOP 1
329 #define USING_INITIAL_FREQ_ESTIMATION 0
336 /* Insert new options above this line. */
337 /* Non-compat options: */
341 OPT_l = (1 << 7) * ENABLE_FEATURE_NTPD_SERVER,
342 OPT_I = (1 << 8) * ENABLE_FEATURE_NTPD_SERVER,
343 /* We hijack some bits for other purposes */
349 /* total round trip delay to currently selected reference clock */
351 /* reference timestamp: time when the system clock was last set or corrected */
353 /* total dispersion to currently selected reference clock */
356 double last_script_run;
359 #if ENABLE_FEATURE_NTPD_SERVER
362 # define G_listen_fd (G.listen_fd)
364 # define G_listen_fd (-1)
368 /* refid: 32-bit code identifying the particular server or reference clock
369 * in stratum 0 packets this is a four-character ASCII string,
370 * called the kiss code, used for debugging and monitoring
371 * in stratum 1 packets this is a four-character ASCII string
372 * assigned to the reference clock by IANA. Example: "GPS "
373 * in stratum 2+ packets, it's IPv4 address or 4 first bytes
374 * of MD5 hash of IPv6
378 /* precision is defined as the larger of the resolution and time to
379 * read the clock, in log2 units. For instance, the precision of a
380 * mains-frequency clock incrementing at 60 Hz is 16 ms, even when the
381 * system clock hardware representation is to the nanosecond.
383 * Delays, jitters of various kinds are clamped down to precision.
385 * If precision_sec is too large, discipline_jitter gets clamped to it
386 * and if offset is smaller than discipline_jitter * POLLADJ_GATE, poll
387 * interval grows even though we really can benefit from staying at
388 * smaller one, collecting non-lagged datapoits and correcting offset.
389 * (Lagged datapoits exist when poll_exp is large but we still have
390 * systematic offset error - the time distance between datapoints
391 * is significant and older datapoints have smaller offsets.
392 * This makes our offset estimation a bit smaller than reality)
393 * Due to this effect, setting G_precision_sec close to
394 * STEP_THRESHOLD isn't such a good idea - offsets may grow
395 * too big and we will step. I observed it with -6.
397 * OTOH, setting precision_sec far too small would result in futile
398 * attempts to synchronize to an unachievable precision.
400 * -6 is 1/64 sec, -7 is 1/128 sec and so on.
401 * -8 is 1/256 ~= 0.003906 (worked well for me --vda)
402 * -9 is 1/512 ~= 0.001953 (let's try this for some time)
404 #define G_precision_exp -9
406 * G_precision_exp is used only for construction outgoing packets.
407 * It's ok to set G_precision_sec to a slightly different value
408 * (One which is "nicer looking" in logs).
409 * Exact value would be (1.0 / (1 << (- G_precision_exp))):
411 #define G_precision_sec 0.002
414 #define STATE_NSET 0 /* initial state, "nothing is set" */
415 //#define STATE_FSET 1 /* frequency set from file */
416 //#define STATE_SPIK 2 /* spike detected */
417 //#define STATE_FREQ 3 /* initial frequency */
418 #define STATE_SYNC 4 /* clock synchronized (normal operation) */
419 uint8_t discipline_state; // doc calls it c.state
420 uint8_t poll_exp; // s.poll
421 int polladj_count; // c.count
422 long kernel_freq_drift;
423 peer_t *last_update_peer;
424 double last_update_offset; // c.last
425 double last_update_recv_time; // s.t
426 double discipline_jitter; // c.jitter
427 /* Since we only compare it with ints, can simplify code
428 * by not making this variable floating point:
430 unsigned offset_to_jitter_ratio;
431 //double cluster_offset; // s.offset
432 //double cluster_jitter; // s.jitter
433 #if !USING_KERNEL_PLL_LOOP
434 double discipline_freq_drift; // c.freq
435 /* Maybe conditionally calculate wander? it's used only for logging */
436 double discipline_wander; // c.wander
439 #define G (*ptr_to_globals)
442 #define VERB1 if (MAX_VERBOSE && G.verbose)
443 #define VERB2 if (MAX_VERBOSE >= 2 && G.verbose >= 2)
444 #define VERB3 if (MAX_VERBOSE >= 3 && G.verbose >= 3)
445 #define VERB4 if (MAX_VERBOSE >= 4 && G.verbose >= 4)
446 #define VERB5 if (MAX_VERBOSE >= 5 && G.verbose >= 5)
447 #define VERB6 if (MAX_VERBOSE >= 6 && G.verbose >= 6)
450 static double LOG2D(int a)
453 return 1.0 / (1UL << -a);
456 static ALWAYS_INLINE double SQUARE(double x)
460 static ALWAYS_INLINE double MAXD(double a, double b)
466 static ALWAYS_INLINE double MIND(double a, double b)
472 static NOINLINE double my_SQRT(double X)
479 double Xhalf = X * 0.5;
481 /* Fast and good approximation to 1/sqrt(X), black magic */
483 /*v.i = 0x5f3759df - (v.i >> 1);*/
484 v.i = 0x5f375a86 - (v.i >> 1); /* - this constant is slightly better */
485 invsqrt = v.f; /* better than 0.2% accuracy */
487 /* Refining it using Newton's method: x1 = x0 - f(x0)/f'(x0)
488 * f(x) = 1/(x*x) - X (f==0 when x = 1/sqrt(X))
490 * f(x)/f'(x) = (X - 1/(x*x)) / (2/(x*x*x)) = X*x*x*x/2 - x/2
491 * x1 = x0 - (X*x0*x0*x0/2 - x0/2) = 1.5*x0 - X*x0*x0*x0/2 = x0*(1.5 - (X/2)*x0*x0)
493 invsqrt = invsqrt * (1.5 - Xhalf * invsqrt * invsqrt); /* ~0.05% accuracy */
494 /* invsqrt = invsqrt * (1.5 - Xhalf * invsqrt * invsqrt); 2nd iter: ~0.0001% accuracy */
495 /* With 4 iterations, more than half results will be exact,
496 * at 6th iterations result stabilizes with about 72% results exact.
497 * We are well satisfied with 0.05% accuracy.
500 return X * invsqrt; /* X * 1/sqrt(X) ~= sqrt(X) */
502 static ALWAYS_INLINE double SQRT(double X)
504 /* If this arch doesn't use IEEE 754 floats, fall back to using libm */
505 if (sizeof(float) != 4)
508 /* This avoids needing libm, saves about 0.5k on x86-32 */
516 gettimeofday(&tv, NULL); /* never fails */
517 G.cur_time = tv.tv_sec + (1.0e-6 * tv.tv_usec) + OFFSET_1900_1970;
522 d_to_tv(double d, struct timeval *tv)
524 tv->tv_sec = (long)d;
525 tv->tv_usec = (d - tv->tv_sec) * 1000000;
529 lfp_to_d(l_fixedpt_t lfp)
532 lfp.int_partl = ntohl(lfp.int_partl);
533 lfp.fractionl = ntohl(lfp.fractionl);
534 ret = (double)lfp.int_partl + ((double)lfp.fractionl / UINT_MAX);
538 sfp_to_d(s_fixedpt_t sfp)
541 sfp.int_parts = ntohs(sfp.int_parts);
542 sfp.fractions = ntohs(sfp.fractions);
543 ret = (double)sfp.int_parts + ((double)sfp.fractions / USHRT_MAX);
546 #if ENABLE_FEATURE_NTPD_SERVER
551 lfp.int_partl = (uint32_t)d;
552 lfp.fractionl = (uint32_t)((d - lfp.int_partl) * UINT_MAX);
553 lfp.int_partl = htonl(lfp.int_partl);
554 lfp.fractionl = htonl(lfp.fractionl);
561 sfp.int_parts = (uint16_t)d;
562 sfp.fractions = (uint16_t)((d - sfp.int_parts) * USHRT_MAX);
563 sfp.int_parts = htons(sfp.int_parts);
564 sfp.fractions = htons(sfp.fractions);
570 dispersion(const datapoint_t *dp)
572 return dp->d_dispersion + FREQ_TOLERANCE * (G.cur_time - dp->d_recv_time);
576 root_distance(peer_t *p)
578 /* The root synchronization distance is the maximum error due to
579 * all causes of the local clock relative to the primary server.
580 * It is defined as half the total delay plus total dispersion
583 return MAXD(MINDISP, p->lastpkt_rootdelay + p->lastpkt_delay) / 2
584 + p->lastpkt_rootdisp
585 + p->filter_dispersion
586 + FREQ_TOLERANCE * (G.cur_time - p->lastpkt_recv_time)
591 set_next(peer_t *p, unsigned t)
593 p->next_action_time = G.cur_time + t;
597 * Peer clock filter and its helpers
600 filter_datapoints(peer_t *p)
607 /* Simulations have shown that use of *averaged* offset for p->filter_offset
608 * is in fact worse than simply using last received one: with large poll intervals
609 * (>= 2048) averaging code uses offset values which are outdated by hours,
610 * and time/frequency correction goes totally wrong when fed essentially bogus offsets.
613 double minoff, maxoff, w;
614 double x = x; /* for compiler */
615 double oldest_off = oldest_off;
616 double oldest_age = oldest_age;
617 double newest_off = newest_off;
618 double newest_age = newest_age;
620 fdp = p->filter_datapoint;
622 minoff = maxoff = fdp[0].d_offset;
623 for (i = 1; i < NUM_DATAPOINTS; i++) {
624 if (minoff > fdp[i].d_offset)
625 minoff = fdp[i].d_offset;
626 if (maxoff < fdp[i].d_offset)
627 maxoff = fdp[i].d_offset;
630 idx = p->datapoint_idx; /* most recent datapoint's index */
632 * Drop two outliers and take weighted average of the rest:
633 * most_recent/2 + older1/4 + older2/8 ... + older5/32 + older6/32
634 * we use older6/32, not older6/64 since sum of weights should be 1:
635 * 1/2 + 1/4 + 1/8 + 1/16 + 1/32 + 1/32 = 1
641 * filter_dispersion = \ -------------
648 for (i = 0; i < NUM_DATAPOINTS; i++) {
650 bb_error_msg("datapoint[%d]: off:%f disp:%f(%f) age:%f%s",
653 fdp[idx].d_dispersion, dispersion(&fdp[idx]),
654 G.cur_time - fdp[idx].d_recv_time,
655 (minoff == fdp[idx].d_offset || maxoff == fdp[idx].d_offset)
656 ? " (outlier by offset)" : ""
660 sum += dispersion(&fdp[idx]) / (2 << i);
662 if (minoff == fdp[idx].d_offset) {
663 minoff -= 1; /* so that we don't match it ever again */
665 if (maxoff == fdp[idx].d_offset) {
668 oldest_off = fdp[idx].d_offset;
669 oldest_age = G.cur_time - fdp[idx].d_recv_time;
672 newest_off = oldest_off;
673 newest_age = oldest_age;
680 idx = (idx - 1) & (NUM_DATAPOINTS - 1);
682 p->filter_dispersion = sum;
683 wavg += x; /* add another older6/64 to form older6/32 */
684 /* Fix systematic underestimation with large poll intervals.
685 * Imagine that we still have a bit of uncorrected drift,
686 * and poll interval is big (say, 100 sec). Offsets form a progression:
687 * 0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 - 0.7 is most recent.
688 * The algorithm above drops 0.0 and 0.7 as outliers,
689 * and then we have this estimation, ~25% off from 0.7:
690 * 0.1/32 + 0.2/32 + 0.3/16 + 0.4/8 + 0.5/4 + 0.6/2 = 0.503125
692 x = oldest_age - newest_age;
694 x = newest_age / x; /* in above example, 100 / (600 - 100) */
695 if (x < 1) { /* paranoia check */
696 x = (newest_off - oldest_off) * x; /* 0.5 * 100/500 = 0.1 */
700 p->filter_offset = wavg;
704 fdp = p->filter_datapoint;
705 idx = p->datapoint_idx; /* most recent datapoint's index */
707 /* filter_offset: simply use the most recent value */
708 p->filter_offset = fdp[idx].d_offset;
712 * filter_dispersion = \ -------------
719 for (i = 0; i < NUM_DATAPOINTS; i++) {
720 sum += dispersion(&fdp[idx]) / (2 << i);
721 wavg += fdp[idx].d_offset;
722 idx = (idx - 1) & (NUM_DATAPOINTS - 1);
724 wavg /= NUM_DATAPOINTS;
725 p->filter_dispersion = sum;
728 /* +----- -----+ ^ 1/2
732 * filter_jitter = | --- * / (avg-offset_j) |
736 * where n is the number of valid datapoints in the filter (n > 1);
737 * if filter_jitter < precision then filter_jitter = precision
740 for (i = 0; i < NUM_DATAPOINTS; i++) {
741 sum += SQUARE(wavg - fdp[i].d_offset);
743 sum = SQRT(sum / NUM_DATAPOINTS);
744 p->filter_jitter = sum > G_precision_sec ? sum : G_precision_sec;
746 VERB4 bb_error_msg("filter offset:%+f disp:%f jitter:%f",
748 p->filter_dispersion,
753 reset_peer_stats(peer_t *p, double offset)
756 bool small_ofs = fabs(offset) < STEP_THRESHOLD;
758 /* Used to set p->filter_datapoint[i].d_dispersion = MAXDISP
759 * and clear reachable bits, but this proved to be too aggressive:
760 * after step (tested with suspending laptop for ~30 secs),
761 * this caused all previous data to be considered invalid,
762 * making us needing to collect full ~8 datapoints per peer
763 * after step in order to start trusting them.
764 * In turn, this was making poll interval decrease even after
765 * step was done. (Poll interval decreases already before step
766 * in this scenario, because we see large offsets and end up with
767 * no good peer to select).
770 for (i = 0; i < NUM_DATAPOINTS; i++) {
772 p->filter_datapoint[i].d_recv_time += offset;
773 if (p->filter_datapoint[i].d_offset != 0) {
774 p->filter_datapoint[i].d_offset -= offset;
775 //bb_error_msg("p->filter_datapoint[%d].d_offset %f -> %f",
777 // p->filter_datapoint[i].d_offset + offset,
778 // p->filter_datapoint[i].d_offset);
781 p->filter_datapoint[i].d_recv_time = G.cur_time;
782 p->filter_datapoint[i].d_offset = 0;
783 /*p->filter_datapoint[i].d_dispersion = MAXDISP;*/
787 p->lastpkt_recv_time += offset;
789 /*p->reachable_bits = 0;*/
790 p->lastpkt_recv_time = G.cur_time;
792 filter_datapoints(p); /* recalc p->filter_xxx */
793 VERB6 bb_error_msg("%s->lastpkt_recv_time=%f", p->p_dotted, p->lastpkt_recv_time);
796 static len_and_sockaddr*
797 resolve_peer_hostname(peer_t *p)
799 len_and_sockaddr *lsa = host2sockaddr(p->p_hostname, 123);
804 p->p_dotted = xmalloc_sockaddr2dotted_noport(&lsa->u.sa);
805 VERB1 if (strcmp(p->p_hostname, p->p_dotted) != 0)
806 bb_error_msg("'%s' is %s", p->p_hostname, p->p_dotted);
809 p->dns_errors = ((p->dns_errors << 1) | 1) & DNS_ERRORS_CAP;
814 add_peers(const char *s)
819 p = xzalloc(sizeof(*p) + strlen(s));
820 strcpy(p->p_hostname, s);
822 p->p_xmt_msg.m_status = MODE_CLIENT | (NTP_VERSION << 3);
823 p->next_action_time = G.cur_time; /* = set_next(p, 0); */
824 reset_peer_stats(p, STEP_THRESHOLD);
826 /* Names like N.<country2chars>.pool.ntp.org are randomly resolved
827 * to a pool of machines. Sometimes different N's resolve to the same IP.
828 * It is not useful to have two peers with same IP. We skip duplicates.
830 if (resolve_peer_hostname(p)) {
831 for (item = G.ntp_peers; item != NULL; item = item->link) {
832 peer_t *pp = (peer_t *) item->data;
833 if (pp->p_dotted && strcmp(p->p_dotted, pp->p_dotted) == 0) {
834 bb_error_msg("duplicate peer %s (%s)", s, p->p_dotted);
843 llist_add_to(&G.ntp_peers, p);
849 const struct sockaddr *from, const struct sockaddr *to, socklen_t addrlen,
850 msg_t *msg, ssize_t len)
856 ret = sendto(fd, msg, len, MSG_DONTWAIT, to, addrlen);
858 ret = send_to_from(fd, msg, len, MSG_DONTWAIT, to, from, addrlen);
861 bb_perror_msg("send failed");
868 send_query_to_peer(peer_t *p)
873 /* Why do we need to bind()?
874 * See what happens when we don't bind:
876 * socket(PF_INET, SOCK_DGRAM, IPPROTO_IP) = 3
877 * setsockopt(3, SOL_IP, IP_TOS, [16], 4) = 0
878 * gettimeofday({1259071266, 327885}, NULL) = 0
879 * sendto(3, "xxx", 48, MSG_DONTWAIT, {sa_family=AF_INET, sin_port=htons(123), sin_addr=inet_addr("10.34.32.125")}, 16) = 48
880 * ^^^ we sent it from some source port picked by kernel.
881 * time(NULL) = 1259071266
882 * write(2, "ntpd: entering poll 15 secs\n", 28) = 28
883 * poll([{fd=3, events=POLLIN}], 1, 15000) = 1 ([{fd=3, revents=POLLIN}])
884 * recv(3, "yyy", 68, MSG_DONTWAIT) = 48
885 * ^^^ this recv will receive packets to any local port!
887 * Uncomment this and use strace to see it in action:
889 #define PROBE_LOCAL_ADDR /* { len_and_sockaddr lsa; lsa.len = LSA_SIZEOF_SA; getsockname(p->query.fd, &lsa.u.sa, &lsa.len); } */
893 len_and_sockaddr *local_lsa;
895 family = p->p_lsa->u.sa.sa_family;
896 p->p_fd = fd = xsocket_type(&local_lsa, family, SOCK_DGRAM);
897 /* local_lsa has "null" address and port 0 now.
898 * bind() ensures we have a *particular port* selected by kernel
899 * and remembered in p->p_fd, thus later recv(p->p_fd)
900 * receives only packets sent to this port.
903 xbind(fd, &local_lsa->u.sa, local_lsa->len);
905 #if ENABLE_FEATURE_IPV6
906 if (family == AF_INET)
908 setsockopt_int(fd, IPPROTO_IP, IP_TOS, IPTOS_LOWDELAY);
912 /* Emit message _before_ attempted send. Think of a very short
913 * roundtrip networks: we need to go back to recv loop ASAP,
914 * to reduce delay. Printing messages after send works against that.
916 VERB1 bb_error_msg("sending query to %s", p->p_dotted);
919 * Send out a random 64-bit number as our transmit time. The NTP
920 * server will copy said number into the originate field on the
921 * response that it sends us. This is totally legal per the SNTP spec.
923 * The impact of this is two fold: we no longer send out the current
924 * system time for the world to see (which may aid an attacker), and
925 * it gives us a (not very secure) way of knowing that we're not
926 * getting spoofed by an attacker that can't capture our traffic
927 * but can spoof packets from the NTP server we're communicating with.
929 * Save the real transmit timestamp locally.
931 p->p_xmt_msg.m_xmttime.int_partl = rand();
932 p->p_xmt_msg.m_xmttime.fractionl = rand();
933 p->p_xmttime = gettime1900d();
935 /* Were doing it only if sendto worked, but
936 * loss of sync detection needs reachable_bits updated
937 * even if sending fails *locally*:
938 * "network is unreachable" because cable was pulled?
939 * We still need to declare "unsync" if this condition persists.
941 p->reachable_bits <<= 1;
943 if (do_sendto(p->p_fd, /*from:*/ NULL, /*to:*/ &p->p_lsa->u.sa, /*addrlen:*/ p->p_lsa->len,
944 &p->p_xmt_msg, NTP_MSGSIZE_NOAUTH) == -1
949 * We know that we sent nothing.
950 * We can retry *soon* without fearing
951 * that we are flooding the peer.
953 set_next(p, RETRY_INTERVAL);
957 set_next(p, RESPONSE_INTERVAL);
961 /* Note that there is no provision to prevent several run_scripts
962 * to be started in quick succession. In fact, it happens rather often
963 * if initial syncronization results in a step.
964 * You will see "step" and then "stratum" script runs, sometimes
965 * as close as only 0.002 seconds apart.
966 * Script should be ready to deal with this.
968 static void run_script(const char *action, double offset)
971 char *env1, *env2, *env3, *env4;
973 G.last_script_run = G.cur_time;
978 argv[0] = (char*) G.script_name;
979 argv[1] = (char*) action;
982 VERB1 bb_error_msg("executing '%s %s'", G.script_name, action);
984 env1 = xasprintf("%s=%u", "stratum", G.stratum);
986 env2 = xasprintf("%s=%ld", "freq_drift_ppm", G.kernel_freq_drift);
988 env3 = xasprintf("%s=%u", "poll_interval", 1 << G.poll_exp);
990 env4 = xasprintf("%s=%f", "offset", offset);
992 /* Other items of potential interest: selected peer,
993 * rootdelay, reftime, rootdisp, refid, ntp_status,
994 * last_update_offset, last_update_recv_time, discipline_jitter,
995 * how many peers have reachable_bits = 0?
998 /* Don't want to wait: it may run hwclock --systohc, and that
999 * may take some time (seconds): */
1000 /*spawn_and_wait(argv);*/
1003 unsetenv("stratum");
1004 unsetenv("freq_drift_ppm");
1005 unsetenv("poll_interval");
1013 static NOINLINE void
1014 step_time(double offset)
1018 struct timeval tvc, tvn;
1019 char buf[sizeof("yyyy-mm-dd hh:mm:ss") + /*paranoia:*/ 4];
1022 gettimeofday(&tvc, NULL); /* never fails */
1023 dtime = tvc.tv_sec + (1.0e-6 * tvc.tv_usec) + offset;
1024 d_to_tv(dtime, &tvn);
1025 if (settimeofday(&tvn, NULL) == -1)
1026 bb_perror_msg_and_die("settimeofday");
1030 strftime_YYYYMMDDHHMMSS(buf, sizeof(buf), &tval);
1031 bb_error_msg("current time is %s.%06u", buf, (unsigned)tvc.tv_usec);
1034 strftime_YYYYMMDDHHMMSS(buf, sizeof(buf), &tval);
1035 bb_error_msg("setting time to %s.%06u (offset %+fs)", buf, (unsigned)tvn.tv_usec, offset);
1037 /* Correct various fields which contain time-relative values: */
1040 G.cur_time += offset;
1041 G.last_update_recv_time += offset;
1042 G.last_script_run += offset;
1044 /* p->lastpkt_recv_time, p->next_action_time and such: */
1045 for (item = G.ntp_peers; item != NULL; item = item->link) {
1046 peer_t *pp = (peer_t *) item->data;
1047 reset_peer_stats(pp, offset);
1048 //bb_error_msg("offset:%+f pp->next_action_time:%f -> %f",
1049 // offset, pp->next_action_time, pp->next_action_time + offset);
1050 pp->next_action_time += offset;
1051 if (pp->p_fd >= 0) {
1052 /* We wait for reply from this peer too.
1053 * But due to step we are doing, reply's data is no longer
1054 * useful (in fact, it'll be bogus). Stop waiting for it.
1058 set_next(pp, RETRY_INTERVAL);
1063 static void clamp_pollexp_and_set_MAXSTRAT(void)
1065 if (G.poll_exp < MINPOLL)
1066 G.poll_exp = MINPOLL;
1067 if (G.poll_exp > BIGPOLL)
1068 G.poll_exp = BIGPOLL;
1069 G.polladj_count = 0;
1070 G.stratum = MAXSTRAT;
1075 * Selection and clustering, and their helpers
1081 double opt_rd; /* optimization */
1084 compare_point_edge(const void *aa, const void *bb)
1086 const point_t *a = aa;
1087 const point_t *b = bb;
1088 if (a->edge < b->edge) {
1091 return (a->edge > b->edge);
1098 compare_survivor_metric(const void *aa, const void *bb)
1100 const survivor_t *a = aa;
1101 const survivor_t *b = bb;
1102 if (a->metric < b->metric) {
1105 return (a->metric > b->metric);
1108 fit(peer_t *p, double rd)
1110 if ((p->reachable_bits & (p->reachable_bits-1)) == 0) {
1111 /* One or zero bits in reachable_bits */
1112 VERB4 bb_error_msg("peer %s unfit for selection: unreachable", p->p_dotted);
1115 #if 0 /* we filter out such packets earlier */
1116 if ((p->lastpkt_status & LI_ALARM) == LI_ALARM
1117 || p->lastpkt_stratum >= MAXSTRAT
1119 VERB4 bb_error_msg("peer %s unfit for selection: bad status/stratum", p->p_dotted);
1123 /* rd is root_distance(p) */
1124 if (rd > MAXDIST + FREQ_TOLERANCE * (1 << G.poll_exp)) {
1125 VERB4 bb_error_msg("peer %s unfit for selection: root distance too high", p->p_dotted);
1129 // /* Do we have a loop? */
1130 // if (p->refid == p->dstaddr || p->refid == s.refid)
1135 select_and_cluster(void)
1140 int size = 3 * G.peer_cnt;
1141 /* for selection algorithm */
1142 point_t point[size];
1143 unsigned num_points, num_candidates;
1145 unsigned num_falsetickers;
1146 /* for cluster algorithm */
1147 survivor_t survivor[size];
1148 unsigned num_survivors;
1154 while (item != NULL) {
1157 p = (peer_t *) item->data;
1158 rd = root_distance(p);
1159 offset = p->filter_offset;
1165 VERB5 bb_error_msg("interval: [%f %f %f] %s",
1171 point[num_points].p = p;
1172 point[num_points].type = -1;
1173 point[num_points].edge = offset - rd;
1174 point[num_points].opt_rd = rd;
1176 point[num_points].p = p;
1177 point[num_points].type = 0;
1178 point[num_points].edge = offset;
1179 point[num_points].opt_rd = rd;
1181 point[num_points].p = p;
1182 point[num_points].type = 1;
1183 point[num_points].edge = offset + rd;
1184 point[num_points].opt_rd = rd;
1188 num_candidates = num_points / 3;
1189 if (num_candidates == 0) {
1190 VERB3 bb_error_msg("no valid datapoints%s", ", no peer selected");
1193 //TODO: sorting does not seem to be done in reference code
1194 qsort(point, num_points, sizeof(point[0]), compare_point_edge);
1196 /* Start with the assumption that there are no falsetickers.
1197 * Attempt to find a nonempty intersection interval containing
1198 * the midpoints of all truechimers.
1199 * If a nonempty interval cannot be found, increase the number
1200 * of assumed falsetickers by one and try again.
1201 * If a nonempty interval is found and the number of falsetickers
1202 * is less than the number of truechimers, a majority has been found
1203 * and the midpoint of each truechimer represents
1204 * the candidates available to the cluster algorithm.
1206 num_falsetickers = 0;
1209 unsigned num_midpoints = 0;
1214 for (i = 0; i < num_points; i++) {
1216 * if (point[i].type == -1) c++;
1217 * if (point[i].type == 1) c--;
1218 * and it's simpler to do it this way:
1221 if (c >= num_candidates - num_falsetickers) {
1222 /* If it was c++ and it got big enough... */
1223 low = point[i].edge;
1226 if (point[i].type == 0)
1230 for (i = num_points-1; i >= 0; i--) {
1232 if (c >= num_candidates - num_falsetickers) {
1233 high = point[i].edge;
1236 if (point[i].type == 0)
1239 /* If the number of midpoints is greater than the number
1240 * of allowed falsetickers, the intersection contains at
1241 * least one truechimer with no midpoint - bad.
1242 * Also, interval should be nonempty.
1244 if (num_midpoints <= num_falsetickers && low < high)
1247 if (num_falsetickers * 2 >= num_candidates) {
1248 VERB3 bb_error_msg("falsetickers:%d, candidates:%d%s",
1249 num_falsetickers, num_candidates,
1250 ", no peer selected");
1254 VERB4 bb_error_msg("selected interval: [%f, %f]; candidates:%d falsetickers:%d",
1255 low, high, num_candidates, num_falsetickers);
1259 /* Construct a list of survivors (p, metric)
1260 * from the chime list, where metric is dominated
1261 * first by stratum and then by root distance.
1262 * All other things being equal, this is the order of preference.
1265 for (i = 0; i < num_points; i++) {
1266 if (point[i].edge < low || point[i].edge > high)
1269 survivor[num_survivors].p = p;
1270 /* x.opt_rd == root_distance(p); */
1271 survivor[num_survivors].metric = MAXDIST * p->lastpkt_stratum + point[i].opt_rd;
1272 VERB5 bb_error_msg("survivor[%d] metric:%f peer:%s",
1273 num_survivors, survivor[num_survivors].metric, p->p_dotted);
1276 /* There must be at least MIN_SELECTED survivors to satisfy the
1277 * correctness assertions. Ordinarily, the Byzantine criteria
1278 * require four survivors, but for the demonstration here, one
1281 if (num_survivors < MIN_SELECTED) {
1282 VERB3 bb_error_msg("survivors:%d%s",
1284 ", no peer selected");
1288 //looks like this is ONLY used by the fact that later we pick survivor[0].
1289 //we can avoid sorting then, just find the minimum once!
1290 qsort(survivor, num_survivors, sizeof(survivor[0]), compare_survivor_metric);
1292 /* For each association p in turn, calculate the selection
1293 * jitter p->sjitter as the square root of the sum of squares
1294 * (p->offset - q->offset) over all q associations. The idea is
1295 * to repeatedly discard the survivor with maximum selection
1296 * jitter until a termination condition is met.
1299 unsigned max_idx = max_idx;
1300 double max_selection_jitter = max_selection_jitter;
1301 double min_jitter = min_jitter;
1303 if (num_survivors <= MIN_CLUSTERED) {
1304 VERB4 bb_error_msg("num_survivors %d <= %d, not discarding more",
1305 num_survivors, MIN_CLUSTERED);
1309 /* To make sure a few survivors are left
1310 * for the clustering algorithm to chew on,
1311 * we stop if the number of survivors
1312 * is less than or equal to MIN_CLUSTERED (3).
1314 for (i = 0; i < num_survivors; i++) {
1315 double selection_jitter_sq;
1318 if (i == 0 || p->filter_jitter < min_jitter)
1319 min_jitter = p->filter_jitter;
1321 selection_jitter_sq = 0;
1322 for (j = 0; j < num_survivors; j++) {
1323 peer_t *q = survivor[j].p;
1324 selection_jitter_sq += SQUARE(p->filter_offset - q->filter_offset);
1326 if (i == 0 || selection_jitter_sq > max_selection_jitter) {
1327 max_selection_jitter = selection_jitter_sq;
1330 VERB6 bb_error_msg("survivor %d selection_jitter^2:%f",
1331 i, selection_jitter_sq);
1333 max_selection_jitter = SQRT(max_selection_jitter / num_survivors);
1334 VERB5 bb_error_msg("max_selection_jitter (at %d):%f min_jitter:%f",
1335 max_idx, max_selection_jitter, min_jitter);
1337 /* If the maximum selection jitter is less than the
1338 * minimum peer jitter, then tossing out more survivors
1339 * will not lower the minimum peer jitter, so we might
1342 if (max_selection_jitter < min_jitter) {
1343 VERB4 bb_error_msg("max_selection_jitter:%f < min_jitter:%f, num_survivors:%d, not discarding more",
1344 max_selection_jitter, min_jitter, num_survivors);
1348 /* Delete survivor[max_idx] from the list
1349 * and go around again.
1351 VERB6 bb_error_msg("dropping survivor %d", max_idx);
1353 while (max_idx < num_survivors) {
1354 survivor[max_idx] = survivor[max_idx + 1];
1360 /* Combine the offsets of the clustering algorithm survivors
1361 * using a weighted average with weight determined by the root
1362 * distance. Compute the selection jitter as the weighted RMS
1363 * difference between the first survivor and the remaining
1364 * survivors. In some cases the inherent clock jitter can be
1365 * reduced by not using this algorithm, especially when frequent
1366 * clockhopping is involved. bbox: thus we don't do it.
1370 for (i = 0; i < num_survivors; i++) {
1372 x = root_distance(p);
1374 z += p->filter_offset / x;
1375 w += SQUARE(p->filter_offset - survivor[0].p->filter_offset) / x;
1377 //G.cluster_offset = z / y;
1378 //G.cluster_jitter = SQRT(w / y);
1381 /* Pick the best clock. If the old system peer is on the list
1382 * and at the same stratum as the first survivor on the list,
1383 * then don't do a clock hop. Otherwise, select the first
1384 * survivor on the list as the new system peer.
1387 if (G.last_update_peer
1388 && G.last_update_peer->lastpkt_stratum <= p->lastpkt_stratum
1390 /* Starting from 1 is ok here */
1391 for (i = 1; i < num_survivors; i++) {
1392 if (G.last_update_peer == survivor[i].p) {
1393 VERB5 bb_error_msg("keeping old synced peer");
1394 p = G.last_update_peer;
1399 G.last_update_peer = p;
1401 VERB4 bb_error_msg("selected peer %s filter_offset:%+f age:%f",
1404 G.cur_time - p->lastpkt_recv_time
1411 * Local clock discipline and its helpers
1414 set_new_values(int disc_state, double offset, double recv_time)
1416 /* Enter new state and set state variables. Note we use the time
1417 * of the last clock filter sample, which must be earlier than
1420 VERB4 bb_error_msg("disc_state=%d last update offset=%f recv_time=%f",
1421 disc_state, offset, recv_time);
1422 G.discipline_state = disc_state;
1423 G.last_update_offset = offset;
1424 G.last_update_recv_time = recv_time;
1426 /* Return: -1: decrease poll interval, 0: leave as is, 1: increase */
1428 update_local_clock(peer_t *p)
1432 /* Note: can use G.cluster_offset instead: */
1433 double offset = p->filter_offset;
1434 double recv_time = p->lastpkt_recv_time;
1436 #if !USING_KERNEL_PLL_LOOP
1439 #if !USING_KERNEL_PLL_LOOP || USING_INITIAL_FREQ_ESTIMATION
1440 double since_last_update;
1442 double etemp, dtemp;
1444 abs_offset = fabs(offset);
1447 /* If needed, -S script can do it by looking at $offset
1448 * env var and killing parent */
1449 /* If the offset is too large, give up and go home */
1450 if (abs_offset > PANIC_THRESHOLD) {
1451 bb_error_msg_and_die("offset %f far too big, exiting", offset);
1455 /* If this is an old update, for instance as the result
1456 * of a system peer change, avoid it. We never use
1457 * an old sample or the same sample twice.
1459 if (recv_time <= G.last_update_recv_time) {
1460 VERB3 bb_error_msg("update from %s: same or older datapoint, not using it",
1462 return 0; /* "leave poll interval as is" */
1465 /* Clock state machine transition function. This is where the
1466 * action is and defines how the system reacts to large time
1467 * and frequency errors.
1469 #if !USING_KERNEL_PLL_LOOP || USING_INITIAL_FREQ_ESTIMATION
1470 since_last_update = recv_time - G.reftime;
1472 #if !USING_KERNEL_PLL_LOOP
1475 #if USING_INITIAL_FREQ_ESTIMATION
1476 if (G.discipline_state == STATE_FREQ) {
1477 /* Ignore updates until the stepout threshold */
1478 if (since_last_update < WATCH_THRESHOLD) {
1479 VERB4 bb_error_msg("measuring drift, datapoint ignored, %f sec remains",
1480 WATCH_THRESHOLD - since_last_update);
1481 return 0; /* "leave poll interval as is" */
1483 # if !USING_KERNEL_PLL_LOOP
1484 freq_drift = (offset - G.last_update_offset) / since_last_update;
1489 /* There are two main regimes: when the
1490 * offset exceeds the step threshold and when it does not.
1492 if (abs_offset > STEP_THRESHOLD) {
1496 // This "spike state" seems to be useless, peer selection already drops
1497 // occassional "bad" datapoints. If we are here, there were _many_
1498 // large offsets. When a few first large offsets are seen,
1499 // we end up in "no valid datapoints, no peer selected" state.
1500 // Only when enough of them are seen (which means it's not a fluke),
1501 // we end up here. Looks like _our_ clock is off.
1502 switch (G.discipline_state) {
1504 /* The first outlyer: ignore it, switch to SPIK state */
1505 VERB3 bb_error_msg("update from %s: offset:%+f, spike%s",
1506 p->p_dotted, offset,
1508 G.discipline_state = STATE_SPIK;
1509 return -1; /* "decrease poll interval" */
1512 /* Ignore succeeding outlyers until either an inlyer
1513 * is found or the stepout threshold is exceeded.
1515 remains = WATCH_THRESHOLD - since_last_update;
1517 VERB3 bb_error_msg("update from %s: offset:%+f, spike%s",
1518 p->p_dotted, offset,
1519 ", datapoint ignored");
1520 return -1; /* "decrease poll interval" */
1522 /* fall through: we need to step */
1526 /* Step the time and clamp down the poll interval.
1528 * In NSET state an initial frequency correction is
1529 * not available, usually because the frequency file has
1530 * not yet been written. Since the time is outside the
1531 * capture range, the clock is stepped. The frequency
1532 * will be set directly following the stepout interval.
1534 * In FSET state the initial frequency has been set
1535 * from the frequency file. Since the time is outside
1536 * the capture range, the clock is stepped immediately,
1537 * rather than after the stepout interval. Guys get
1538 * nervous if it takes 17 minutes to set the clock for
1541 * In SPIK state the stepout threshold has expired and
1542 * the phase is still above the step threshold. Note
1543 * that a single spike greater than the step threshold
1544 * is always suppressed, even at the longer poll
1547 VERB4 bb_error_msg("stepping time by %+f; poll_exp=MINPOLL", offset);
1549 if (option_mask32 & OPT_q) {
1550 /* We were only asked to set time once. Done. */
1554 clamp_pollexp_and_set_MAXSTRAT();
1556 run_script("step", offset);
1558 recv_time += offset;
1560 #if USING_INITIAL_FREQ_ESTIMATION
1561 if (G.discipline_state == STATE_NSET) {
1562 set_new_values(STATE_FREQ, /*offset:*/ 0, recv_time);
1563 return 1; /* "ok to increase poll interval" */
1566 abs_offset = offset = 0;
1567 set_new_values(STATE_SYNC, offset, recv_time);
1568 } else { /* abs_offset <= STEP_THRESHOLD */
1570 /* The ratio is calculated before jitter is updated to make
1571 * poll adjust code more sensitive to large offsets.
1573 G.offset_to_jitter_ratio = abs_offset / G.discipline_jitter;
1575 /* Compute the clock jitter as the RMS of exponentially
1576 * weighted offset differences. Used by the poll adjust code.
1578 etemp = SQUARE(G.discipline_jitter);
1579 dtemp = SQUARE(offset - G.last_update_offset);
1580 G.discipline_jitter = SQRT(etemp + (dtemp - etemp) / AVG);
1581 if (G.discipline_jitter < G_precision_sec)
1582 G.discipline_jitter = G_precision_sec;
1584 switch (G.discipline_state) {
1586 if (option_mask32 & OPT_q) {
1587 /* We were only asked to set time once.
1588 * The clock is precise enough, no need to step.
1592 #if USING_INITIAL_FREQ_ESTIMATION
1593 /* This is the first update received and the frequency
1594 * has not been initialized. The first thing to do
1595 * is directly measure the oscillator frequency.
1597 set_new_values(STATE_FREQ, offset, recv_time);
1599 set_new_values(STATE_SYNC, offset, recv_time);
1601 VERB4 bb_error_msg("transitioning to FREQ, datapoint ignored");
1602 return 0; /* "leave poll interval as is" */
1604 #if 0 /* this is dead code for now */
1606 /* This is the first update and the frequency
1607 * has been initialized. Adjust the phase, but
1608 * don't adjust the frequency until the next update.
1610 set_new_values(STATE_SYNC, offset, recv_time);
1611 /* freq_drift remains 0 */
1615 #if USING_INITIAL_FREQ_ESTIMATION
1617 /* since_last_update >= WATCH_THRESHOLD, we waited enough.
1618 * Correct the phase and frequency and switch to SYNC state.
1619 * freq_drift was already estimated (see code above)
1621 set_new_values(STATE_SYNC, offset, recv_time);
1626 #if !USING_KERNEL_PLL_LOOP
1627 /* Compute freq_drift due to PLL and FLL contributions.
1629 * The FLL and PLL frequency gain constants
1630 * depend on the poll interval and Allan
1631 * intercept. The FLL is not used below one-half
1632 * the Allan intercept. Above that the loop gain
1633 * increases in steps to 1 / AVG.
1635 if ((1 << G.poll_exp) > ALLAN / 2) {
1636 etemp = FLL - G.poll_exp;
1639 freq_drift += (offset - G.last_update_offset) / (MAXD(since_last_update, ALLAN) * etemp);
1641 /* For the PLL the integration interval
1642 * (numerator) is the minimum of the update
1643 * interval and poll interval. This allows
1644 * oversampling, but not undersampling.
1646 etemp = MIND(since_last_update, (1 << G.poll_exp));
1647 dtemp = (4 * PLL) << G.poll_exp;
1648 freq_drift += offset * etemp / SQUARE(dtemp);
1650 set_new_values(STATE_SYNC, offset, recv_time);
1653 if (G.stratum != p->lastpkt_stratum + 1) {
1654 G.stratum = p->lastpkt_stratum + 1;
1655 run_script("stratum", offset);
1659 G.reftime = G.cur_time;
1660 G.ntp_status = p->lastpkt_status;
1661 G.refid = p->lastpkt_refid;
1662 G.rootdelay = p->lastpkt_rootdelay + p->lastpkt_delay;
1663 dtemp = p->filter_jitter; // SQRT(SQUARE(p->filter_jitter) + SQUARE(G.cluster_jitter));
1664 dtemp += MAXD(p->filter_dispersion + FREQ_TOLERANCE * (G.cur_time - p->lastpkt_recv_time) + abs_offset, MINDISP);
1665 G.rootdisp = p->lastpkt_rootdisp + dtemp;
1666 VERB4 bb_error_msg("updating leap/refid/reftime/rootdisp from peer %s", p->p_dotted);
1668 /* We are in STATE_SYNC now, but did not do adjtimex yet.
1669 * (Any other state does not reach this, they all return earlier)
1670 * By this time, freq_drift and offset are set
1671 * to values suitable for adjtimex.
1673 #if !USING_KERNEL_PLL_LOOP
1674 /* Calculate the new frequency drift and frequency stability (wander).
1675 * Compute the clock wander as the RMS of exponentially weighted
1676 * frequency differences. This is not used directly, but can,
1677 * along with the jitter, be a highly useful monitoring and
1680 dtemp = G.discipline_freq_drift + freq_drift;
1681 G.discipline_freq_drift = MAXD(MIND(MAXDRIFT, dtemp), -MAXDRIFT);
1682 etemp = SQUARE(G.discipline_wander);
1683 dtemp = SQUARE(dtemp);
1684 G.discipline_wander = SQRT(etemp + (dtemp - etemp) / AVG);
1686 VERB4 bb_error_msg("discipline freq_drift=%.9f(int:%ld corr:%e) wander=%f",
1687 G.discipline_freq_drift,
1688 (long)(G.discipline_freq_drift * 65536e6),
1690 G.discipline_wander);
1693 memset(&tmx, 0, sizeof(tmx));
1694 if (adjtimex(&tmx) < 0)
1695 bb_perror_msg_and_die("adjtimex");
1696 bb_error_msg("p adjtimex freq:%ld offset:%+ld status:0x%x tc:%ld",
1697 tmx.freq, tmx.offset, tmx.status, tmx.constant);
1700 memset(&tmx, 0, sizeof(tmx));
1702 //doesn't work, offset remains 0 (!) in kernel:
1703 //ntpd: set adjtimex freq:1786097 tmx.offset:77487
1704 //ntpd: prev adjtimex freq:1786097 tmx.offset:0
1705 //ntpd: cur adjtimex freq:1786097 tmx.offset:0
1706 tmx.modes = ADJ_FREQUENCY | ADJ_OFFSET;
1707 /* 65536 is one ppm */
1708 tmx.freq = G.discipline_freq_drift * 65536e6;
1710 tmx.modes = ADJ_OFFSET | ADJ_STATUS | ADJ_TIMECONST;// | ADJ_MAXERROR | ADJ_ESTERROR;
1711 tmx.constant = (int)G.poll_exp - 4;
1713 * The below if statement should be unnecessary, but...
1714 * It looks like Linux kernel's PLL is far too gentle in changing
1715 * tmx.freq in response to clock offset. Offset keeps growing
1716 * and eventually we fall back to smaller poll intervals.
1717 * We can make correction more aggressive (about x2) by supplying
1718 * PLL time constant which is one less than the real one.
1719 * To be on a safe side, let's do it only if offset is significantly
1720 * larger than jitter.
1722 if (G.offset_to_jitter_ratio >= TIMECONST_HACK_GATE)
1724 tmx.offset = (long)(offset * 1000000); /* usec */
1725 if (SLEW_THRESHOLD < STEP_THRESHOLD) {
1726 if (tmx.offset > (long)(SLEW_THRESHOLD * 1000000)) {
1727 tmx.offset = (long)(SLEW_THRESHOLD * 1000000);
1730 if (tmx.offset < -(long)(SLEW_THRESHOLD * 1000000)) {
1731 tmx.offset = -(long)(SLEW_THRESHOLD * 1000000);
1735 if (tmx.constant < 0)
1738 tmx.status = STA_PLL;
1739 if (G.ntp_status & LI_PLUSSEC)
1740 tmx.status |= STA_INS;
1741 if (G.ntp_status & LI_MINUSSEC)
1742 tmx.status |= STA_DEL;
1744 //tmx.esterror = (uint32_t)(clock_jitter * 1e6);
1745 //tmx.maxerror = (uint32_t)((sys_rootdelay / 2 + sys_rootdisp) * 1e6);
1746 rc = adjtimex(&tmx);
1748 bb_perror_msg_and_die("adjtimex");
1749 /* NB: here kernel returns constant == G.poll_exp, not == G.poll_exp - 4.
1750 * Not sure why. Perhaps it is normal.
1752 VERB4 bb_error_msg("adjtimex:%d freq:%ld offset:%+ld status:0x%x",
1753 rc, tmx.freq, tmx.offset, tmx.status);
1754 G.kernel_freq_drift = tmx.freq / 65536;
1755 VERB2 bb_error_msg("update from:%s offset:%+f delay:%f jitter:%f clock drift:%+.3fppm tc:%d",
1759 G.discipline_jitter,
1760 (double)tmx.freq / 65536,
1764 return 1; /* "ok to increase poll interval" */
1769 * We've got a new reply packet from a peer, process it
1773 poll_interval(int upper_bound)
1775 unsigned interval, r, mask;
1776 interval = 1 << G.poll_exp;
1777 if (interval > upper_bound)
1778 interval = upper_bound;
1779 mask = ((interval-1) >> 4) | 1;
1781 interval += r & mask; /* ~ random(0..1) * interval/16 */
1782 VERB4 bb_error_msg("chose poll interval:%u (poll_exp:%d)", interval, G.poll_exp);
1786 adjust_poll(int count)
1788 G.polladj_count += count;
1789 if (G.polladj_count > POLLADJ_LIMIT) {
1790 G.polladj_count = 0;
1791 if (G.poll_exp < MAXPOLL) {
1793 VERB4 bb_error_msg("polladj: discipline_jitter:%f ++poll_exp=%d",
1794 G.discipline_jitter, G.poll_exp);
1796 } else if (G.polladj_count < -POLLADJ_LIMIT || (count < 0 && G.poll_exp > BIGPOLL)) {
1797 G.polladj_count = 0;
1798 if (G.poll_exp > MINPOLL) {
1802 /* Correct p->next_action_time in each peer
1803 * which waits for sending, so that they send earlier.
1804 * Old pp->next_action_time are on the order
1805 * of t + (1 << old_poll_exp) + small_random,
1806 * we simply need to subtract ~half of that.
1808 for (item = G.ntp_peers; item != NULL; item = item->link) {
1809 peer_t *pp = (peer_t *) item->data;
1811 pp->next_action_time -= (1 << G.poll_exp);
1813 VERB4 bb_error_msg("polladj: discipline_jitter:%f --poll_exp=%d",
1814 G.discipline_jitter, G.poll_exp);
1817 VERB4 bb_error_msg("polladj: count:%d", G.polladj_count);
1820 static NOINLINE void
1821 recv_and_process_peer_pkt(peer_t *p)
1826 double T1, T2, T3, T4;
1828 double prev_delay, delay;
1830 datapoint_t *datapoint;
1835 /* We can recvfrom here and check from.IP, but some multihomed
1836 * ntp servers reply from their *other IP*.
1837 * TODO: maybe we should check at least what we can: from.port == 123?
1840 size = recv(p->p_fd, &msg, sizeof(msg), MSG_DONTWAIT);
1845 if (errno == EAGAIN)
1846 /* There was no packet after all
1847 * (poll() returning POLLIN for a fd
1848 * is not a ironclad guarantee that data is there)
1852 * If you need a different handling for a specific
1853 * errno, always explain it in comment.
1855 bb_perror_msg_and_die("recv(%s) error", p->p_dotted);
1858 if (size != NTP_MSGSIZE_NOAUTH && size != NTP_MSGSIZE) {
1859 bb_error_msg("malformed packet received from %s", p->p_dotted);
1863 if (msg.m_orgtime.int_partl != p->p_xmt_msg.m_xmttime.int_partl
1864 || msg.m_orgtime.fractionl != p->p_xmt_msg.m_xmttime.fractionl
1866 /* Somebody else's packet */
1870 /* We do not expect any more packets from this peer for now.
1871 * Closing the socket informs kernel about it.
1872 * We open a new socket when we send a new query.
1877 if ((msg.m_status & LI_ALARM) == LI_ALARM
1878 || msg.m_stratum == 0
1879 || msg.m_stratum > NTP_MAXSTRATUM
1881 bb_error_msg("reply from %s: peer is unsynced", p->p_dotted);
1883 * Stratum 0 responses may have commands in 32-bit m_refid field:
1884 * "DENY", "RSTR" - peer does not like us at all,
1885 * "RATE" - peer is overloaded, reduce polling freq.
1886 * If poll interval is small, increase it.
1888 if (G.poll_exp < BIGPOLL)
1889 goto increase_interval;
1890 goto pick_normal_interval;
1893 // /* Verify valid root distance */
1894 // if (msg.m_rootdelay / 2 + msg.m_rootdisp >= MAXDISP || p->lastpkt_reftime > msg.m_xmt)
1895 // return; /* invalid header values */
1898 * From RFC 2030 (with a correction to the delay math):
1900 * Timestamp Name ID When Generated
1901 * ------------------------------------------------------------
1902 * Originate Timestamp T1 time request sent by client
1903 * Receive Timestamp T2 time request received by server
1904 * Transmit Timestamp T3 time reply sent by server
1905 * Destination Timestamp T4 time reply received by client
1907 * The roundtrip delay and local clock offset are defined as
1909 * delay = (T4 - T1) - (T3 - T2); offset = ((T2 - T1) + (T3 - T4)) / 2
1912 T2 = lfp_to_d(msg.m_rectime);
1913 T3 = lfp_to_d(msg.m_xmttime);
1916 /* The delay calculation is a special case. In cases where the
1917 * server and client clocks are running at different rates and
1918 * with very fast networks, the delay can appear negative. In
1919 * order to avoid violating the Principle of Least Astonishment,
1920 * the delay is clamped not less than the system precision.
1922 delay = (T4 - T1) - (T3 - T2);
1923 if (delay < G_precision_sec)
1924 delay = G_precision_sec;
1926 * If this packet's delay is much bigger than the last one,
1927 * it's better to just ignore it than use its much less precise value.
1929 prev_delay = p->p_raw_delay;
1930 p->p_raw_delay = delay;
1931 if (p->reachable_bits && delay > prev_delay * BAD_DELAY_GROWTH) {
1932 bb_error_msg("reply from %s: delay %f is too high, ignoring", p->p_dotted, delay);
1933 goto pick_normal_interval;
1936 p->lastpkt_delay = delay;
1937 p->lastpkt_recv_time = T4;
1938 VERB6 bb_error_msg("%s->lastpkt_recv_time=%f", p->p_dotted, p->lastpkt_recv_time);
1939 p->lastpkt_status = msg.m_status;
1940 p->lastpkt_stratum = msg.m_stratum;
1941 p->lastpkt_rootdelay = sfp_to_d(msg.m_rootdelay);
1942 p->lastpkt_rootdisp = sfp_to_d(msg.m_rootdisp);
1943 p->lastpkt_refid = msg.m_refid;
1945 p->datapoint_idx = p->reachable_bits ? (p->datapoint_idx + 1) % NUM_DATAPOINTS : 0;
1946 datapoint = &p->filter_datapoint[p->datapoint_idx];
1947 datapoint->d_recv_time = T4;
1948 datapoint->d_offset = offset = ((T2 - T1) + (T3 - T4)) / 2;
1949 datapoint->d_dispersion = LOG2D(msg.m_precision_exp) + G_precision_sec;
1950 if (!p->reachable_bits) {
1951 /* 1st datapoint ever - replicate offset in every element */
1953 for (i = 0; i < NUM_DATAPOINTS; i++) {
1954 p->filter_datapoint[i].d_offset = offset;
1958 p->reachable_bits |= 1;
1959 if ((MAX_VERBOSE && G.verbose) || (option_mask32 & OPT_w)) {
1960 bb_error_msg("reply from %s: offset:%+f delay:%f status:0x%02x strat:%d refid:0x%08x rootdelay:%f reach:0x%02x",
1967 p->lastpkt_rootdelay,
1969 /* not shown: m_ppoll, m_precision_exp, m_rootdisp,
1970 * m_reftime, m_orgtime, m_rectime, m_xmttime
1975 /* Muck with statictics and update the clock */
1976 filter_datapoints(p);
1977 q = select_and_cluster();
1980 if (!(option_mask32 & OPT_w)) {
1981 rc = update_local_clock(q);
1983 //Disabled this because there is a case where largish offsets
1984 //are unavoidable: if network round-trip delay is, say, ~0.6s,
1985 //error in offset estimation would be ~delay/2 ~= 0.3s.
1986 //Thus, offsets will be usually in -0.3...0.3s range.
1987 //In this case, this code would keep poll interval small,
1988 //but it won't be helping.
1989 //BIGOFF check below deals with a case of seeing multi-second offsets.
1991 /* If drift is dangerously large, immediately
1992 * drop poll interval one step down.
1994 if (fabs(q->filter_offset) >= POLLDOWN_OFFSET) {
1995 VERB4 bb_error_msg("offset:%+f > POLLDOWN_OFFSET", q->filter_offset);
1996 adjust_poll(-POLLADJ_LIMIT * 3);
2002 /* No peer selected.
2003 * If poll interval is small, increase it.
2005 if (G.poll_exp < BIGPOLL)
2006 goto increase_interval;
2010 /* Adjust the poll interval by comparing the current offset
2011 * with the clock jitter. If the offset is less than
2012 * the clock jitter times a constant, then the averaging interval
2013 * is increased, otherwise it is decreased. A bit of hysteresis
2014 * helps calm the dance. Works best using burst mode.
2016 if (rc > 0 && G.offset_to_jitter_ratio <= POLLADJ_GATE) {
2017 /* was += G.poll_exp but it is a bit
2018 * too optimistic for my taste at high poll_exp's */
2020 adjust_poll(MINPOLL);
2023 bb_error_msg("want smaller interval: offset/jitter = %u",
2024 G.offset_to_jitter_ratio);
2025 adjust_poll(-G.poll_exp * 2);
2029 /* Decide when to send new query for this peer */
2030 pick_normal_interval:
2031 interval = poll_interval(INT_MAX);
2032 if (fabs(offset) >= BIGOFF && interval > BIGOFF_INTERVAL) {
2033 /* If we are synced, offsets are less than SLEW_THRESHOLD,
2034 * or at the very least not much larger than it.
2035 * Now we see a largish one.
2036 * Either this peer is feeling bad, or packet got corrupted,
2037 * or _our_ clock is wrong now and _all_ peers will show similar
2038 * largish offsets too.
2039 * I observed this with laptop suspend stopping clock.
2040 * In any case, it makes sense to make next request soonish:
2041 * cases 1 and 2: get a better datapoint,
2042 * case 3: allows to resync faster.
2044 interval = BIGOFF_INTERVAL;
2047 set_next(p, interval);
2050 #if ENABLE_FEATURE_NTPD_SERVER
2051 static NOINLINE void
2052 recv_and_process_client_pkt(void /*int fd*/)
2056 len_and_sockaddr *to;
2057 struct sockaddr *from;
2059 uint8_t query_status;
2060 l_fixedpt_t query_xmttime;
2062 to = get_sock_lsa(G_listen_fd);
2063 from = xzalloc(to->len);
2065 size = recv_from_to(G_listen_fd, &msg, sizeof(msg), MSG_DONTWAIT, from, &to->u.sa, to->len);
2066 if (size != NTP_MSGSIZE_NOAUTH && size != NTP_MSGSIZE) {
2069 if (errno == EAGAIN)
2071 bb_perror_msg_and_die("recv");
2073 addr = xmalloc_sockaddr2dotted_noport(from);
2074 bb_error_msg("malformed packet received from %s: size %u", addr, (int)size);
2079 /* Respond only to client and symmetric active packets */
2080 if ((msg.m_status & MODE_MASK) != MODE_CLIENT
2081 && (msg.m_status & MODE_MASK) != MODE_SYM_ACT
2086 query_status = msg.m_status;
2087 query_xmttime = msg.m_xmttime;
2089 /* Build a reply packet */
2090 memset(&msg, 0, sizeof(msg));
2091 msg.m_status = G.stratum < MAXSTRAT ? (G.ntp_status & LI_MASK) : LI_ALARM;
2092 msg.m_status |= (query_status & VERSION_MASK);
2093 msg.m_status |= ((query_status & MODE_MASK) == MODE_CLIENT) ?
2094 MODE_SERVER : MODE_SYM_PAS;
2095 msg.m_stratum = G.stratum;
2096 msg.m_ppoll = G.poll_exp;
2097 msg.m_precision_exp = G_precision_exp;
2098 /* this time was obtained between poll() and recv() */
2099 msg.m_rectime = d_to_lfp(G.cur_time);
2100 msg.m_xmttime = d_to_lfp(gettime1900d()); /* this instant */
2101 if (G.peer_cnt == 0) {
2102 /* we have no peers: "stratum 1 server" mode. reftime = our own time */
2103 G.reftime = G.cur_time;
2105 msg.m_reftime = d_to_lfp(G.reftime);
2106 msg.m_orgtime = query_xmttime;
2107 msg.m_rootdelay = d_to_sfp(G.rootdelay);
2108 //simple code does not do this, fix simple code!
2109 msg.m_rootdisp = d_to_sfp(G.rootdisp);
2110 //version = (query_status & VERSION_MASK); /* ... >> VERSION_SHIFT - done below instead */
2111 msg.m_refid = G.refid; // (version > (3 << VERSION_SHIFT)) ? G.refid : G.refid3;
2113 /* We reply from the local address packet was sent to,
2114 * this makes to/from look swapped here: */
2115 do_sendto(G_listen_fd,
2116 /*from:*/ &to->u.sa, /*to:*/ from, /*addrlen:*/ to->len,
2125 /* Upstream ntpd's options:
2127 * -4 Force DNS resolution of host names to the IPv4 namespace.
2128 * -6 Force DNS resolution of host names to the IPv6 namespace.
2129 * -a Require cryptographic authentication for broadcast client,
2130 * multicast client and symmetric passive associations.
2131 * This is the default.
2132 * -A Do not require cryptographic authentication for broadcast client,
2133 * multicast client and symmetric passive associations.
2134 * This is almost never a good idea.
2135 * -b Enable the client to synchronize to broadcast servers.
2137 * Specify the name and path of the configuration file,
2138 * default /etc/ntp.conf
2139 * -d Specify debugging mode. This option may occur more than once,
2140 * with each occurrence indicating greater detail of display.
2142 * Specify debugging level directly.
2144 * Specify the name and path of the frequency file.
2145 * This is the same operation as the "driftfile FILE"
2146 * configuration command.
2147 * -g Normally, ntpd exits with a message to the system log
2148 * if the offset exceeds the panic threshold, which is 1000 s
2149 * by default. This option allows the time to be set to any value
2150 * without restriction; however, this can happen only once.
2151 * If the threshold is exceeded after that, ntpd will exit
2152 * with a message to the system log. This option can be used
2153 * with the -q and -x options. See the tinker command for other options.
2155 * Chroot the server to the directory jaildir. This option also implies
2156 * that the server attempts to drop root privileges at startup
2157 * (otherwise, chroot gives very little additional security).
2158 * You may need to also specify a -u option.
2160 * Specify the name and path of the symmetric key file,
2161 * default /etc/ntp/keys. This is the same operation
2162 * as the "keys FILE" configuration command.
2164 * Specify the name and path of the log file. The default
2165 * is the system log file. This is the same operation as
2166 * the "logfile FILE" configuration command.
2167 * -L Do not listen to virtual IPs. The default is to listen.
2169 * -N To the extent permitted by the operating system,
2170 * run the ntpd at the highest priority.
2172 * Specify the name and path of the file used to record the ntpd
2173 * process ID. This is the same operation as the "pidfile FILE"
2174 * configuration command.
2176 * To the extent permitted by the operating system,
2177 * run the ntpd at the specified priority.
2178 * -q Exit the ntpd just after the first time the clock is set.
2179 * This behavior mimics that of the ntpdate program, which is
2180 * to be retired. The -g and -x options can be used with this option.
2181 * Note: The kernel time discipline is disabled with this option.
2183 * Specify the default propagation delay from the broadcast/multicast
2184 * server to this client. This is necessary only if the delay
2185 * cannot be computed automatically by the protocol.
2187 * Specify the directory path for files created by the statistics
2188 * facility. This is the same operation as the "statsdir DIR"
2189 * configuration command.
2191 * Add a key number to the trusted key list. This option can occur
2194 * Specify a user, and optionally a group, to switch to.
2197 * Add a system variable listed by default.
2198 * -x Normally, the time is slewed if the offset is less than the step
2199 * threshold, which is 128 ms by default, and stepped if above
2200 * the threshold. This option sets the threshold to 600 s, which is
2201 * well within the accuracy window to set the clock manually.
2202 * Note: since the slew rate of typical Unix kernels is limited
2203 * to 0.5 ms/s, each second of adjustment requires an amortization
2204 * interval of 2000 s. Thus, an adjustment as much as 600 s
2205 * will take almost 14 days to complete. This option can be used
2206 * with the -g and -q options. See the tinker command for other options.
2207 * Note: The kernel time discipline is disabled with this option.
2210 /* By doing init in a separate function we decrease stack usage
2213 static NOINLINE void ntp_init(char **argv)
2221 bb_error_msg_and_die(bb_msg_you_must_be_root);
2223 /* Set some globals */
2224 G.discipline_jitter = G_precision_sec;
2225 G.stratum = MAXSTRAT;
2227 G.poll_exp = BURSTPOLL; /* speeds up initial sync */
2228 G.last_script_run = G.reftime = G.last_update_recv_time = gettime1900d(); /* sets G.cur_time too */
2232 opts = getopt32(argv, "^"
2234 "wp:*S:"IF_FEATURE_NTPD_SERVER("l") /* NOT compat */
2235 IF_FEATURE_NTPD_SERVER("I:") /* compat */
2237 "46aAbgL" /* compat, ignored */
2239 "dd:wn" /* -d: counter; -p: list; -w implies -n */
2240 IF_FEATURE_NTPD_SERVER(":Il") /* -I implies -l */
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 if (!(opts & OPT_n)) {
2267 bb_daemonize_or_rexec(DAEMON_DEVNULL_STDIO, argv);
2268 logmode = LOGMODE_NONE;
2273 add_peers(llist_pop(&peers));
2275 #if ENABLE_FEATURE_NTPD_CONF
2280 parser = config_open("/etc/ntp.conf");
2281 while (config_read(parser, token, 3, 1, "# \t", PARSE_NORMAL)) {
2282 if (strcmp(token[0], "server") == 0 && token[1]) {
2283 add_peers(token[1]);
2286 bb_error_msg("skipping %s:%u: unimplemented command '%s'",
2287 "/etc/ntp.conf", parser->lineno, token[0]
2290 config_close(parser);
2293 if (G.peer_cnt == 0) {
2294 if (!(opts & OPT_l))
2296 /* -l but no peers: "stratum 1 server" mode */
2299 /* If network is up, syncronization occurs in ~10 seconds.
2300 * We give "ntpd -q" 10 seconds to get first reply,
2301 * then another 50 seconds to finish syncing.
2303 * I tested ntpd 4.2.6p1 and apparently it never exits
2304 * (will try forever), but it does not feel right.
2305 * The goal of -q is to act like ntpdate: set time
2306 * after a reasonably small period of polling, or fail.
2309 option_mask32 |= OPT_qq;
2326 int ntpd_main(int argc UNUSED_PARAM, char **argv) MAIN_EXTERNALLY_VISIBLE;
2327 int ntpd_main(int argc UNUSED_PARAM, char **argv)
2335 memset(&G, 0, sizeof(G));
2336 SET_PTR_TO_GLOBALS(&G);
2340 /* If ENABLE_FEATURE_NTPD_SERVER, + 1 for listen_fd: */
2341 cnt = G.peer_cnt + ENABLE_FEATURE_NTPD_SERVER;
2342 idx2peer = xzalloc(sizeof(idx2peer[0]) * cnt);
2343 pfd = xzalloc(sizeof(pfd[0]) * cnt);
2345 /* Countdown: we never sync before we sent INITIAL_SAMPLES+1
2346 * packets to each peer.
2347 * NB: if some peer is not responding, we may end up sending
2348 * fewer packets to it and more to other peers.
2349 * NB2: sync usually happens using INITIAL_SAMPLES packets,
2350 * since last reply does not come back instantaneously.
2352 cnt = G.peer_cnt * (INITIAL_SAMPLES + 1);
2354 write_pidfile(CONFIG_PID_FILE_PATH "/ntpd.pid");
2356 while (!bb_got_signal) {
2362 /* Nothing between here and poll() blocks for any significant time */
2364 nextaction = G.cur_time + 3600;
2367 #if ENABLE_FEATURE_NTPD_SERVER
2368 if (G_listen_fd != -1) {
2369 pfd[0].fd = G_listen_fd;
2370 pfd[0].events = POLLIN;
2374 /* Pass over peer list, send requests, time out on receives */
2375 for (item = G.ntp_peers; item != NULL; item = item->link) {
2376 peer_t *p = (peer_t *) item->data;
2378 if (p->next_action_time <= G.cur_time) {
2379 if (p->p_fd == -1) {
2380 /* Time to send new req */
2382 VERB4 bb_error_msg("disabling burst mode");
2383 G.polladj_count = 0;
2384 G.poll_exp = MINPOLL;
2386 send_query_to_peer(p);
2388 /* Timed out waiting for reply */
2391 /* If poll interval is small, increase it */
2392 if (G.poll_exp < BIGPOLL)
2393 adjust_poll(MINPOLL);
2394 timeout = poll_interval(NOREPLY_INTERVAL);
2395 bb_error_msg("timed out waiting for %s, reach 0x%02x, next query in %us",
2396 p->p_dotted, p->reachable_bits, timeout);
2398 /* What if don't see it because it changed its IP? */
2399 if (p->reachable_bits == 0)
2400 resolve_peer_hostname(p);
2402 set_next(p, timeout);
2406 if (p->next_action_time < nextaction)
2407 nextaction = p->next_action_time;
2410 /* Wait for reply from this peer */
2411 pfd[i].fd = p->p_fd;
2412 pfd[i].events = POLLIN;
2418 timeout = nextaction - G.cur_time;
2421 timeout++; /* (nextaction - G.cur_time) rounds down, compensating */
2423 /* Here we may block */
2425 if (i > (ENABLE_FEATURE_NTPD_SERVER && G_listen_fd != -1)) {
2426 /* We wait for at least one reply.
2427 * Poll for it, without wasting time for message.
2428 * Since replies often come under 1 second, this also
2429 * reduces clutter in logs.
2431 nfds = poll(pfd, i, 1000);
2437 bb_error_msg("poll:%us sockets:%u interval:%us", timeout, i, 1 << G.poll_exp);
2439 nfds = poll(pfd, i, timeout * 1000);
2441 gettime1900d(); /* sets G.cur_time */
2446 break; /* poll was interrupted by a signal */
2448 if (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 */
2454 /* Resolve peer names to IPs, if not resolved yet.
2455 * We do it only when poll timed out:
2456 * this way, we almost never overlap DNS resolution with
2457 * "request-reply" packet round trip.
2460 for (item = G.ntp_peers; item != NULL; item = item->link) {
2461 peer_t *p = (peer_t *) item->data;
2462 if (p->next_action_time <= ct && !p->p_lsa) {
2463 /* This can take up to ~10 sec per each DNS query */
2464 resolve_peer_hostname(p);
2467 gettime1900d(); /* sets G.cur_time (needed for set_next()) */
2468 /* Set next time for those which are still not resolved */
2469 for (item = G.ntp_peers; item != NULL; item = item->link) {
2470 peer_t *p = (peer_t *) item->data;
2471 if (p->next_action_time <= ct && !p->p_lsa) {
2472 set_next(p, HOSTNAME_INTERVAL * p->dns_errors);
2478 /* Process any received packets */
2480 #if ENABLE_FEATURE_NTPD_SERVER
2481 if (G.listen_fd != -1) {
2482 if (pfd[0].revents /* & (POLLIN|POLLERR)*/) {
2484 recv_and_process_client_pkt(/*G.listen_fd*/);
2485 gettime1900d(); /* sets G.cur_time */
2490 for (; nfds != 0 && j < i; j++) {
2491 if (pfd[j].revents /* & (POLLIN|POLLERR)*/) {
2493 * At init, alarm was set to 10 sec.
2494 * Now we did get a reply.
2495 * Increase timeout to 50 seconds to finish syncing.
2497 if (option_mask32 & OPT_qq) {
2498 option_mask32 &= ~OPT_qq;
2502 recv_and_process_peer_pkt(idx2peer[j]);
2503 gettime1900d(); /* sets G.cur_time */
2508 if (G.ntp_peers && G.stratum != MAXSTRAT) {
2509 for (item = G.ntp_peers; item != NULL; item = item->link) {
2510 peer_t *p = (peer_t *) item->data;
2511 if (p->reachable_bits)
2512 goto have_reachable_peer;
2514 /* No peer responded for last 8 packets, panic */
2515 clamp_pollexp_and_set_MAXSTRAT();
2516 run_script("unsync", 0.0);
2517 have_reachable_peer: ;
2519 } /* while (!bb_got_signal) */
2521 remove_pidfile(CONFIG_PID_FILE_PATH "/ntpd.pid");
2522 kill_myself_with_sig(bb_got_signal);
2530 /*** openntpd-4.6 uses only adjtime, not adjtimex ***/
2532 /*** ntp-4.2.6/ntpd/ntp_loopfilter.c - adjtimex usage ***/
2536 direct_freq(double fp_offset)
2540 * If the kernel is enabled, we need the residual offset to
2541 * calculate the frequency correction.
2543 if (pll_control && kern_enable) {
2544 memset(&ntv, 0, sizeof(ntv));
2547 clock_offset = ntv.offset / 1e9;
2548 #else /* STA_NANO */
2549 clock_offset = ntv.offset / 1e6;
2550 #endif /* STA_NANO */
2551 drift_comp = FREQTOD(ntv.freq);
2553 #endif /* KERNEL_PLL */
2554 set_freq((fp_offset - clock_offset) / (current_time - clock_epoch) + drift_comp);
2560 set_freq(double freq) /* frequency update */
2568 * If the kernel is enabled, update the kernel frequency.
2570 if (pll_control && kern_enable) {
2571 memset(&ntv, 0, sizeof(ntv));
2572 ntv.modes = MOD_FREQUENCY;
2573 ntv.freq = DTOFREQ(drift_comp);
2575 snprintf(tbuf, sizeof(tbuf), "kernel %.3f PPM", drift_comp * 1e6);
2576 report_event(EVNT_FSET, NULL, tbuf);
2578 snprintf(tbuf, sizeof(tbuf), "ntpd %.3f PPM", drift_comp * 1e6);
2579 report_event(EVNT_FSET, NULL, tbuf);
2581 #else /* KERNEL_PLL */
2582 snprintf(tbuf, sizeof(tbuf), "ntpd %.3f PPM", drift_comp * 1e6);
2583 report_event(EVNT_FSET, NULL, tbuf);
2584 #endif /* KERNEL_PLL */
2593 * This code segment works when clock adjustments are made using
2594 * precision time kernel support and the ntp_adjtime() system
2595 * call. This support is available in Solaris 2.6 and later,
2596 * Digital Unix 4.0 and later, FreeBSD, Linux and specially
2597 * modified kernels for HP-UX 9 and Ultrix 4. In the case of the
2598 * DECstation 5000/240 and Alpha AXP, additional kernel
2599 * modifications provide a true microsecond clock and nanosecond
2600 * clock, respectively.
2602 * Important note: The kernel discipline is used only if the
2603 * step threshold is less than 0.5 s, as anything higher can
2604 * lead to overflow problems. This might occur if some misguided
2605 * lad set the step threshold to something ridiculous.
2607 if (pll_control && kern_enable) {
2609 #define MOD_BITS (MOD_OFFSET | MOD_MAXERROR | MOD_ESTERROR | MOD_STATUS | MOD_TIMECONST)
2612 * We initialize the structure for the ntp_adjtime()
2613 * system call. We have to convert everything to
2614 * microseconds or nanoseconds first. Do not update the
2615 * system variables if the ext_enable flag is set. In
2616 * this case, the external clock driver will update the
2617 * variables, which will be read later by the local
2618 * clock driver. Afterwards, remember the time and
2619 * frequency offsets for jitter and stability values and
2620 * to update the frequency file.
2622 memset(&ntv, 0, sizeof(ntv));
2624 ntv.modes = MOD_STATUS;
2627 ntv.modes = MOD_BITS | MOD_NANO;
2628 #else /* STA_NANO */
2629 ntv.modes = MOD_BITS;
2630 #endif /* STA_NANO */
2631 if (clock_offset < 0)
2636 ntv.offset = (int32)(clock_offset * 1e9 + dtemp);
2637 ntv.constant = sys_poll;
2638 #else /* STA_NANO */
2639 ntv.offset = (int32)(clock_offset * 1e6 + dtemp);
2640 ntv.constant = sys_poll - 4;
2641 #endif /* STA_NANO */
2642 ntv.esterror = (u_int32)(clock_jitter * 1e6);
2643 ntv.maxerror = (u_int32)((sys_rootdelay / 2 + sys_rootdisp) * 1e6);
2644 ntv.status = STA_PLL;
2647 * Enable/disable the PPS if requested.
2650 if (!(pll_status & STA_PPSTIME))
2651 report_event(EVNT_KERN,
2652 NULL, "PPS enabled");
2653 ntv.status |= STA_PPSTIME | STA_PPSFREQ;
2655 if (pll_status & STA_PPSTIME)
2656 report_event(EVNT_KERN,
2657 NULL, "PPS disabled");
2658 ntv.status &= ~(STA_PPSTIME | STA_PPSFREQ);
2660 if (sys_leap == LEAP_ADDSECOND)
2661 ntv.status |= STA_INS;
2662 else if (sys_leap == LEAP_DELSECOND)
2663 ntv.status |= STA_DEL;
2667 * Pass the stuff to the kernel. If it squeals, turn off
2668 * the pps. In any case, fetch the kernel offset,
2669 * frequency and jitter.
2671 if (ntp_adjtime(&ntv) == TIME_ERROR) {
2672 if (!(ntv.status & STA_PPSSIGNAL))
2673 report_event(EVNT_KERN, NULL,
2676 pll_status = ntv.status;
2678 clock_offset = ntv.offset / 1e9;
2679 #else /* STA_NANO */
2680 clock_offset = ntv.offset / 1e6;
2681 #endif /* STA_NANO */
2682 clock_frequency = FREQTOD(ntv.freq);
2685 * If the kernel PPS is lit, monitor its performance.
2687 if (ntv.status & STA_PPSTIME) {
2689 clock_jitter = ntv.jitter / 1e9;
2690 #else /* STA_NANO */
2691 clock_jitter = ntv.jitter / 1e6;
2692 #endif /* STA_NANO */
2695 #if defined(STA_NANO) && NTP_API == 4
2697 * If the TAI changes, update the kernel TAI.
2699 if (loop_tai != sys_tai) {
2701 ntv.modes = MOD_TAI;
2702 ntv.constant = sys_tai;
2705 #endif /* STA_NANO */
2707 #endif /* KERNEL_PLL */