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 (may be repeated)"
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/timex.h>
98 #ifndef IPTOS_LOWDELAY
99 # define IPTOS_LOWDELAY 0x10
103 /* Verbosity control (max level of -dddd options accepted).
104 * max 6 is very talkative (and bloated). 3 is non-bloated,
105 * production level setting.
107 #define MAX_VERBOSE 3
110 /* High-level description of the algorithm:
112 * We start running with very small poll_exp, BURSTPOLL,
113 * in order to quickly accumulate INITIAL_SAMPLES datapoints
114 * for each peer. Then, time is stepped if the offset is larger
115 * than STEP_THRESHOLD, otherwise it isn't; anyway, we enlarge
116 * poll_exp to MINPOLL and enter frequency measurement step:
117 * we collect new datapoints but ignore them for WATCH_THRESHOLD
118 * seconds. After WATCH_THRESHOLD seconds we look at accumulated
119 * offset and estimate frequency drift.
121 * (frequency measurement step seems to not be strictly needed,
122 * it is conditionally disabled with USING_INITIAL_FREQ_ESTIMATION
125 * After this, we enter "steady state": we collect a datapoint,
126 * we select the best peer, if this datapoint is not a new one
127 * (IOW: if this datapoint isn't for selected peer), sleep
128 * and collect another one; otherwise, use its offset to update
129 * frequency drift, if offset is somewhat large, reduce poll_exp,
130 * otherwise increase poll_exp.
132 * If offset is larger than STEP_THRESHOLD, which shouldn't normally
133 * happen, we assume that something "bad" happened (computer
134 * was hibernated, someone set totally wrong date, etc),
135 * then the time is stepped, all datapoints are discarded,
136 * and we go back to steady state.
138 * Made some changes to speed up re-syncing after our clock goes bad
139 * (tested with suspending my laptop):
140 * - if largish offset (>= STEP_THRESHOLD == 1 sec) is seen
141 * from a peer, schedule next query for this peer soon
142 * without drastically lowering poll interval for everybody.
143 * This makes us collect enough data for step much faster:
144 * e.g. at poll = 10 (1024 secs), step was done within 5 minutes
145 * after first reply which indicated that our clock is 14 seconds off.
146 * - on step, do not discard d_dispersion data of the existing datapoints,
147 * do not clear reachable_bits. This prevents discarding first ~8
148 * datapoints after the step.
151 #define INITIAL_SAMPLES 4 /* how many samples do we want for init */
152 #define BAD_DELAY_GROWTH 4 /* drop packet if its delay grew by more than this */
154 #define RETRY_INTERVAL 32 /* on send/recv error, retry in N secs (need to be power of 2) */
155 #define NOREPLY_INTERVAL 512 /* sent, but got no reply: cap next query by this many seconds */
156 #define RESPONSE_INTERVAL 16 /* wait for reply up to N secs */
157 #define HOSTNAME_INTERVAL 4 /* hostname lookup failed. Wait N * peer->dns_errors secs for next try */
158 #define DNS_ERRORS_CAP 0x3f /* peer->dns_errors is in [0..63] */
160 /* Step threshold (sec). std ntpd uses 0.128.
162 #define STEP_THRESHOLD 1
163 /* Slew threshold (sec): adjtimex() won't accept offsets larger than this.
164 * Using exact power of 2 (1/8) results in smaller code
166 #define SLEW_THRESHOLD 0.125
167 //^^^^^^^^^^^^^^^^^^^^^^^^^^ TODO: man adjtimex about tmx.offset:
168 // "Since Linux 2.6.26, the supplied value is clamped to the range (-0.5s, +0.5s)"
169 // - can use this larger value instead?
171 /* Stepout threshold (sec). std ntpd uses 900 (11 mins (!)) */
172 //UNUSED: #define WATCH_THRESHOLD 128
173 /* NB: set WATCH_THRESHOLD to ~60 when debugging to save time) */
174 //UNUSED: #define PANIC_THRESHOLD 1000 /* panic threshold (sec) */
177 * If we got |offset| > BIGOFF from a peer, cap next query interval
178 * for this peer by this many seconds:
180 #define BIGOFF STEP_THRESHOLD
181 #define BIGOFF_INTERVAL (1 << 7) /* 128 s */
183 #define FREQ_TOLERANCE 0.000015 /* frequency tolerance (15 PPM) */
184 #define BURSTPOLL 0 /* initial poll */
185 #define MINPOLL 5 /* minimum poll interval. std ntpd uses 6 (6: 64 sec) */
187 * If offset > discipline_jitter * POLLADJ_GATE, and poll interval is > 2^BIGPOLL,
188 * then it is decreased _at once_. (If <= 2^BIGPOLL, it will be decreased _eventually_).
190 #define BIGPOLL 9 /* 2^9 sec ~= 8.5 min */
191 #define MAXPOLL 12 /* maximum poll interval (12: 1.1h, 17: 36.4h). std ntpd uses 17 */
193 * Actively lower poll when we see such big offsets.
194 * With SLEW_THRESHOLD = 0.125, it means we try to sync more aggressively
195 * if offset increases over ~0.04 sec
197 //#define POLLDOWN_OFFSET (SLEW_THRESHOLD / 3)
198 #define MINDISP 0.01 /* minimum dispersion (sec) */
199 #define MAXDISP 16 /* maximum dispersion (sec) */
200 #define MAXSTRAT 16 /* maximum stratum (infinity metric) */
201 #define MAXDIST 1 /* distance threshold (sec) */
202 #define MIN_SELECTED 1 /* minimum intersection survivors */
203 #define MIN_CLUSTERED 3 /* minimum cluster survivors */
205 #define MAXDRIFT 0.000500 /* frequency drift we can correct (500 PPM) */
207 /* Poll-adjust threshold.
208 * When we see that offset is small enough compared to discipline jitter,
209 * we grow a counter: += MINPOLL. When counter goes over POLLADJ_LIMIT,
210 * we poll_exp++. If offset isn't small, counter -= poll_exp*2,
211 * and when it goes below -POLLADJ_LIMIT, we poll_exp--.
212 * (Bumped from 30 to 40 since otherwise I often see poll_exp going *2* steps down)
214 #define POLLADJ_LIMIT 40
215 /* If offset < discipline_jitter * POLLADJ_GATE, then we decide to increase
216 * poll interval (we think we can't improve timekeeping
217 * by staying at smaller poll).
219 #define POLLADJ_GATE 4
220 #define TIMECONST_HACK_GATE 2
221 /* Compromise Allan intercept (sec). doc uses 1500, std ntpd uses 512 */
225 /* FLL loop gain [why it depends on MAXPOLL??] */
226 #define FLL (MAXPOLL + 1)
227 /* Parameter averaging constant */
236 NTP_MSGSIZE_NOAUTH = 48,
237 NTP_MSGSIZE = (NTP_MSGSIZE_NOAUTH + 4 + NTP_DIGESTSIZE),
240 MODE_MASK = (7 << 0),
241 VERSION_MASK = (7 << 3),
245 /* Leap Second Codes (high order two bits of m_status) */
246 LI_NOWARNING = (0 << 6), /* no warning */
247 LI_PLUSSEC = (1 << 6), /* add a second (61 seconds) */
248 LI_MINUSSEC = (2 << 6), /* minus a second (59 seconds) */
249 LI_ALARM = (3 << 6), /* alarm condition */
252 MODE_RES0 = 0, /* reserved */
253 MODE_SYM_ACT = 1, /* symmetric active */
254 MODE_SYM_PAS = 2, /* symmetric passive */
255 MODE_CLIENT = 3, /* client */
256 MODE_SERVER = 4, /* server */
257 MODE_BROADCAST = 5, /* broadcast */
258 MODE_RES1 = 6, /* reserved for NTP control message */
259 MODE_RES2 = 7, /* reserved for private use */
262 //TODO: better base selection
263 #define OFFSET_1900_1970 2208988800UL /* 1970 - 1900 in seconds */
265 #define NUM_DATAPOINTS 8
278 uint8_t m_status; /* status of local clock and leap info */
280 uint8_t m_ppoll; /* poll value */
281 int8_t m_precision_exp;
282 s_fixedpt_t m_rootdelay;
283 s_fixedpt_t m_rootdisp;
285 l_fixedpt_t m_reftime;
286 l_fixedpt_t m_orgtime;
287 l_fixedpt_t m_rectime;
288 l_fixedpt_t m_xmttime;
290 uint8_t m_digest[NTP_DIGESTSIZE];
300 len_and_sockaddr *p_lsa;
304 uint32_t lastpkt_refid;
305 uint8_t lastpkt_status;
306 uint8_t lastpkt_stratum;
307 uint8_t reachable_bits;
309 /* when to send new query (if p_fd == -1)
310 * or when receive times out (if p_fd >= 0): */
311 double next_action_time;
314 /* p_raw_delay is set even by "high delay" packets */
315 /* lastpkt_delay isn't */
316 double lastpkt_recv_time;
317 double lastpkt_delay;
318 double lastpkt_rootdelay;
319 double lastpkt_rootdisp;
320 /* produced by filter algorithm: */
321 double filter_offset;
322 double filter_dispersion;
323 double filter_jitter;
324 datapoint_t filter_datapoint[NUM_DATAPOINTS];
325 /* last sent packet: */
331 #define USING_KERNEL_PLL_LOOP 1
332 #define USING_INITIAL_FREQ_ESTIMATION 0
339 /* Insert new options above this line. */
340 /* Non-compat options: */
344 OPT_l = (1 << 7) * ENABLE_FEATURE_NTPD_SERVER,
345 OPT_I = (1 << 8) * ENABLE_FEATURE_NTPD_SERVER,
346 /* We hijack some bits for other purposes */
352 /* total round trip delay to currently selected reference clock */
354 /* reference timestamp: time when the system clock was last set or corrected */
356 /* total dispersion to currently selected reference clock */
359 double last_script_run;
362 #if ENABLE_FEATURE_NTPD_SERVER
365 # define G_listen_fd (G.listen_fd)
367 # define G_listen_fd (-1)
371 /* refid: 32-bit code identifying the particular server or reference clock
372 * in stratum 0 packets this is a four-character ASCII string,
373 * called the kiss code, used for debugging and monitoring
374 * in stratum 1 packets this is a four-character ASCII string
375 * assigned to the reference clock by IANA. Example: "GPS "
376 * in stratum 2+ packets, it's IPv4 address or 4 first bytes
377 * of MD5 hash of IPv6
381 /* precision is defined as the larger of the resolution and time to
382 * read the clock, in log2 units. For instance, the precision of a
383 * mains-frequency clock incrementing at 60 Hz is 16 ms, even when the
384 * system clock hardware representation is to the nanosecond.
386 * Delays, jitters of various kinds are clamped down to precision.
388 * If precision_sec is too large, discipline_jitter gets clamped to it
389 * and if offset is smaller than discipline_jitter * POLLADJ_GATE, poll
390 * interval grows even though we really can benefit from staying at
391 * smaller one, collecting non-lagged datapoits and correcting offset.
392 * (Lagged datapoits exist when poll_exp is large but we still have
393 * systematic offset error - the time distance between datapoints
394 * is significant and older datapoints have smaller offsets.
395 * This makes our offset estimation a bit smaller than reality)
396 * Due to this effect, setting G_precision_sec close to
397 * STEP_THRESHOLD isn't such a good idea - offsets may grow
398 * too big and we will step. I observed it with -6.
400 * OTOH, setting precision_sec far too small would result in futile
401 * attempts to synchronize to an unachievable precision.
403 * -6 is 1/64 sec, -7 is 1/128 sec and so on.
404 * -8 is 1/256 ~= 0.003906 (worked well for me --vda)
405 * -9 is 1/512 ~= 0.001953 (let's try this for some time)
407 #define G_precision_exp -9
409 * G_precision_exp is used only for construction outgoing packets.
410 * It's ok to set G_precision_sec to a slightly different value
411 * (One which is "nicer looking" in logs).
412 * Exact value would be (1.0 / (1 << (- G_precision_exp))):
414 #define G_precision_sec 0.002
417 #define STATE_NSET 0 /* initial state, "nothing is set" */
418 //#define STATE_FSET 1 /* frequency set from file */
419 //#define STATE_SPIK 2 /* spike detected */
420 //#define STATE_FREQ 3 /* initial frequency */
421 #define STATE_SYNC 4 /* clock synchronized (normal operation) */
422 uint8_t discipline_state; // doc calls it c.state
423 uint8_t poll_exp; // s.poll
424 int polladj_count; // c.count
426 long kernel_freq_drift;
427 peer_t *last_update_peer;
428 double last_update_offset; // c.last
429 double last_update_recv_time; // s.t
430 double discipline_jitter; // c.jitter
431 /* Since we only compare it with ints, can simplify code
432 * by not making this variable floating point:
434 unsigned offset_to_jitter_ratio;
435 //double cluster_offset; // s.offset
436 //double cluster_jitter; // s.jitter
437 #if !USING_KERNEL_PLL_LOOP
438 double discipline_freq_drift; // c.freq
439 /* Maybe conditionally calculate wander? it's used only for logging */
440 double discipline_wander; // c.wander
443 #define G (*ptr_to_globals)
446 #define VERB1 if (MAX_VERBOSE && G.verbose)
447 #define VERB2 if (MAX_VERBOSE >= 2 && G.verbose >= 2)
448 #define VERB3 if (MAX_VERBOSE >= 3 && G.verbose >= 3)
449 #define VERB4 if (MAX_VERBOSE >= 4 && G.verbose >= 4)
450 #define VERB5 if (MAX_VERBOSE >= 5 && G.verbose >= 5)
451 #define VERB6 if (MAX_VERBOSE >= 6 && G.verbose >= 6)
454 static double LOG2D(int a)
457 return 1.0 / (1UL << -a);
460 static ALWAYS_INLINE double SQUARE(double x)
464 static ALWAYS_INLINE double MAXD(double a, double b)
470 static ALWAYS_INLINE double MIND(double a, double b)
476 static NOINLINE double my_SQRT(double X)
483 double Xhalf = X * 0.5;
485 /* Fast and good approximation to 1/sqrt(X), black magic */
487 /*v.i = 0x5f3759df - (v.i >> 1);*/
488 v.i = 0x5f375a86 - (v.i >> 1); /* - this constant is slightly better */
489 invsqrt = v.f; /* better than 0.2% accuracy */
491 /* Refining it using Newton's method: x1 = x0 - f(x0)/f'(x0)
492 * f(x) = 1/(x*x) - X (f==0 when x = 1/sqrt(X))
494 * f(x)/f'(x) = (X - 1/(x*x)) / (2/(x*x*x)) = X*x*x*x/2 - x/2
495 * x1 = x0 - (X*x0*x0*x0/2 - x0/2) = 1.5*x0 - X*x0*x0*x0/2 = x0*(1.5 - (X/2)*x0*x0)
497 invsqrt = invsqrt * (1.5 - Xhalf * invsqrt * invsqrt); /* ~0.05% accuracy */
498 /* invsqrt = invsqrt * (1.5 - Xhalf * invsqrt * invsqrt); 2nd iter: ~0.0001% accuracy */
499 /* With 4 iterations, more than half results will be exact,
500 * at 6th iterations result stabilizes with about 72% results exact.
501 * We are well satisfied with 0.05% accuracy.
504 return X * invsqrt; /* X * 1/sqrt(X) ~= sqrt(X) */
506 static ALWAYS_INLINE double SQRT(double X)
508 /* If this arch doesn't use IEEE 754 floats, fall back to using libm */
509 if (sizeof(float) != 4)
512 /* This avoids needing libm, saves about 0.5k on x86-32 */
520 gettimeofday(&tv, NULL); /* never fails */
521 G.cur_time = tv.tv_sec + (1.0e-6 * tv.tv_usec) + OFFSET_1900_1970;
526 d_to_tv(double d, struct timeval *tv)
528 tv->tv_sec = (long)d;
529 tv->tv_usec = (d - tv->tv_sec) * 1000000;
533 lfp_to_d(l_fixedpt_t lfp)
536 lfp.int_partl = ntohl(lfp.int_partl);
537 lfp.fractionl = ntohl(lfp.fractionl);
538 ret = (double)lfp.int_partl + ((double)lfp.fractionl / UINT_MAX);
542 sfp_to_d(s_fixedpt_t sfp)
545 sfp.int_parts = ntohs(sfp.int_parts);
546 sfp.fractions = ntohs(sfp.fractions);
547 ret = (double)sfp.int_parts + ((double)sfp.fractions / USHRT_MAX);
550 #if ENABLE_FEATURE_NTPD_SERVER
555 lfp.int_partl = (uint32_t)d;
556 lfp.fractionl = (uint32_t)((d - lfp.int_partl) * UINT_MAX);
557 lfp.int_partl = htonl(lfp.int_partl);
558 lfp.fractionl = htonl(lfp.fractionl);
565 sfp.int_parts = (uint16_t)d;
566 sfp.fractions = (uint16_t)((d - sfp.int_parts) * USHRT_MAX);
567 sfp.int_parts = htons(sfp.int_parts);
568 sfp.fractions = htons(sfp.fractions);
574 dispersion(const datapoint_t *dp)
576 return dp->d_dispersion + FREQ_TOLERANCE * (G.cur_time - dp->d_recv_time);
580 root_distance(peer_t *p)
582 /* The root synchronization distance is the maximum error due to
583 * all causes of the local clock relative to the primary server.
584 * It is defined as half the total delay plus total dispersion
587 return MAXD(MINDISP, p->lastpkt_rootdelay + p->lastpkt_delay) / 2
588 + p->lastpkt_rootdisp
589 + p->filter_dispersion
590 + FREQ_TOLERANCE * (G.cur_time - p->lastpkt_recv_time)
595 set_next(peer_t *p, unsigned t)
597 p->next_action_time = G.cur_time + t;
601 * Peer clock filter and its helpers
604 filter_datapoints(peer_t *p)
611 /* Simulations have shown that use of *averaged* offset for p->filter_offset
612 * is in fact worse than simply using last received one: with large poll intervals
613 * (>= 2048) averaging code uses offset values which are outdated by hours,
614 * and time/frequency correction goes totally wrong when fed essentially bogus offsets.
617 double minoff, maxoff, w;
618 double x = x; /* for compiler */
619 double oldest_off = oldest_off;
620 double oldest_age = oldest_age;
621 double newest_off = newest_off;
622 double newest_age = newest_age;
624 fdp = p->filter_datapoint;
626 minoff = maxoff = fdp[0].d_offset;
627 for (i = 1; i < NUM_DATAPOINTS; i++) {
628 if (minoff > fdp[i].d_offset)
629 minoff = fdp[i].d_offset;
630 if (maxoff < fdp[i].d_offset)
631 maxoff = fdp[i].d_offset;
634 idx = p->datapoint_idx; /* most recent datapoint's index */
636 * Drop two outliers and take weighted average of the rest:
637 * most_recent/2 + older1/4 + older2/8 ... + older5/32 + older6/32
638 * we use older6/32, not older6/64 since sum of weights should be 1:
639 * 1/2 + 1/4 + 1/8 + 1/16 + 1/32 + 1/32 = 1
645 * filter_dispersion = \ -------------
652 for (i = 0; i < NUM_DATAPOINTS; i++) {
654 bb_error_msg("datapoint[%d]: off:%f disp:%f(%f) age:%f%s",
657 fdp[idx].d_dispersion, dispersion(&fdp[idx]),
658 G.cur_time - fdp[idx].d_recv_time,
659 (minoff == fdp[idx].d_offset || maxoff == fdp[idx].d_offset)
660 ? " (outlier by offset)" : ""
664 sum += dispersion(&fdp[idx]) / (2 << i);
666 if (minoff == fdp[idx].d_offset) {
667 minoff -= 1; /* so that we don't match it ever again */
669 if (maxoff == fdp[idx].d_offset) {
672 oldest_off = fdp[idx].d_offset;
673 oldest_age = G.cur_time - fdp[idx].d_recv_time;
676 newest_off = oldest_off;
677 newest_age = oldest_age;
684 idx = (idx - 1) & (NUM_DATAPOINTS - 1);
686 p->filter_dispersion = sum;
687 wavg += x; /* add another older6/64 to form older6/32 */
688 /* Fix systematic underestimation with large poll intervals.
689 * Imagine that we still have a bit of uncorrected drift,
690 * and poll interval is big (say, 100 sec). Offsets form a progression:
691 * 0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 - 0.7 is most recent.
692 * The algorithm above drops 0.0 and 0.7 as outliers,
693 * and then we have this estimation, ~25% off from 0.7:
694 * 0.1/32 + 0.2/32 + 0.3/16 + 0.4/8 + 0.5/4 + 0.6/2 = 0.503125
696 x = oldest_age - newest_age;
698 x = newest_age / x; /* in above example, 100 / (600 - 100) */
699 if (x < 1) { /* paranoia check */
700 x = (newest_off - oldest_off) * x; /* 0.5 * 100/500 = 0.1 */
704 p->filter_offset = wavg;
708 fdp = p->filter_datapoint;
709 idx = p->datapoint_idx; /* most recent datapoint's index */
711 /* filter_offset: simply use the most recent value */
712 p->filter_offset = fdp[idx].d_offset;
716 * filter_dispersion = \ -------------
723 for (i = 0; i < NUM_DATAPOINTS; i++) {
724 sum += dispersion(&fdp[idx]) / (2 << i);
725 wavg += fdp[idx].d_offset;
726 idx = (idx - 1) & (NUM_DATAPOINTS - 1);
728 wavg /= NUM_DATAPOINTS;
729 p->filter_dispersion = sum;
732 /* +----- -----+ ^ 1/2
736 * filter_jitter = | --- * / (avg-offset_j) |
740 * where n is the number of valid datapoints in the filter (n > 1);
741 * if filter_jitter < precision then filter_jitter = precision
744 for (i = 0; i < NUM_DATAPOINTS; i++) {
745 sum += SQUARE(wavg - fdp[i].d_offset);
747 sum = SQRT(sum / NUM_DATAPOINTS);
748 p->filter_jitter = sum > G_precision_sec ? sum : G_precision_sec;
750 VERB4 bb_error_msg("filter offset:%+f disp:%f jitter:%f",
752 p->filter_dispersion,
757 reset_peer_stats(peer_t *p, double offset)
760 bool small_ofs = fabs(offset) < STEP_THRESHOLD;
762 /* Used to set p->filter_datapoint[i].d_dispersion = MAXDISP
763 * and clear reachable bits, but this proved to be too aggressive:
764 * after step (tested with suspending laptop for ~30 secs),
765 * this caused all previous data to be considered invalid,
766 * making us needing to collect full ~8 datapoints per peer
767 * after step in order to start trusting them.
768 * In turn, this was making poll interval decrease even after
769 * step was done. (Poll interval decreases already before step
770 * in this scenario, because we see large offsets and end up with
771 * no good peer to select).
774 for (i = 0; i < NUM_DATAPOINTS; i++) {
776 p->filter_datapoint[i].d_recv_time += offset;
777 if (p->filter_datapoint[i].d_offset != 0) {
778 p->filter_datapoint[i].d_offset -= offset;
779 //bb_error_msg("p->filter_datapoint[%d].d_offset %f -> %f",
781 // p->filter_datapoint[i].d_offset + offset,
782 // p->filter_datapoint[i].d_offset);
785 p->filter_datapoint[i].d_recv_time = G.cur_time;
786 p->filter_datapoint[i].d_offset = 0;
787 /*p->filter_datapoint[i].d_dispersion = MAXDISP;*/
791 p->lastpkt_recv_time += offset;
793 /*p->reachable_bits = 0;*/
794 p->lastpkt_recv_time = G.cur_time;
796 filter_datapoints(p); /* recalc p->filter_xxx */
797 VERB6 bb_error_msg("%s->lastpkt_recv_time=%f", p->p_dotted, p->lastpkt_recv_time);
800 static len_and_sockaddr*
801 resolve_peer_hostname(peer_t *p)
803 len_and_sockaddr *lsa = host2sockaddr(p->p_hostname, 123);
808 p->p_dotted = xmalloc_sockaddr2dotted_noport(&lsa->u.sa);
809 VERB1 if (strcmp(p->p_hostname, p->p_dotted) != 0)
810 bb_error_msg("'%s' is %s", p->p_hostname, p->p_dotted);
814 p->dns_errors = ((p->dns_errors << 1) | 1) & DNS_ERRORS_CAP;
819 add_peers(const char *s)
824 p = xzalloc(sizeof(*p) + strlen(s));
825 strcpy(p->p_hostname, s);
827 p->p_xmt_msg.m_status = MODE_CLIENT | (NTP_VERSION << 3);
828 p->next_action_time = G.cur_time; /* = set_next(p, 0); */
829 reset_peer_stats(p, STEP_THRESHOLD);
831 /* Names like N.<country2chars>.pool.ntp.org are randomly resolved
832 * to a pool of machines. Sometimes different N's resolve to the same IP.
833 * It is not useful to have two peers with same IP. We skip duplicates.
835 if (resolve_peer_hostname(p)) {
836 for (item = G.ntp_peers; item != NULL; item = item->link) {
837 peer_t *pp = (peer_t *) item->data;
838 if (pp->p_dotted && strcmp(p->p_dotted, pp->p_dotted) == 0) {
839 bb_error_msg("duplicate peer %s (%s)", s, p->p_dotted);
848 llist_add_to(&G.ntp_peers, p);
854 const struct sockaddr *from, const struct sockaddr *to, socklen_t addrlen,
855 msg_t *msg, ssize_t len)
861 ret = sendto(fd, msg, len, MSG_DONTWAIT, to, addrlen);
863 ret = send_to_from(fd, msg, len, MSG_DONTWAIT, to, from, addrlen);
866 bb_perror_msg("send failed");
873 send_query_to_peer(peer_t *p)
878 /* Why do we need to bind()?
879 * See what happens when we don't bind:
881 * socket(PF_INET, SOCK_DGRAM, IPPROTO_IP) = 3
882 * setsockopt(3, SOL_IP, IP_TOS, [16], 4) = 0
883 * gettimeofday({1259071266, 327885}, NULL) = 0
884 * sendto(3, "xxx", 48, MSG_DONTWAIT, {sa_family=AF_INET, sin_port=htons(123), sin_addr=inet_addr("10.34.32.125")}, 16) = 48
885 * ^^^ we sent it from some source port picked by kernel.
886 * time(NULL) = 1259071266
887 * write(2, "ntpd: entering poll 15 secs\n", 28) = 28
888 * poll([{fd=3, events=POLLIN}], 1, 15000) = 1 ([{fd=3, revents=POLLIN}])
889 * recv(3, "yyy", 68, MSG_DONTWAIT) = 48
890 * ^^^ this recv will receive packets to any local port!
892 * Uncomment this and use strace to see it in action:
894 #define PROBE_LOCAL_ADDR /* { len_and_sockaddr lsa; lsa.len = LSA_SIZEOF_SA; getsockname(p->query.fd, &lsa.u.sa, &lsa.len); } */
898 len_and_sockaddr *local_lsa;
900 family = p->p_lsa->u.sa.sa_family;
901 p->p_fd = fd = xsocket_type(&local_lsa, family, SOCK_DGRAM);
902 /* local_lsa has "null" address and port 0 now.
903 * bind() ensures we have a *particular port* selected by kernel
904 * and remembered in p->p_fd, thus later recv(p->p_fd)
905 * receives only packets sent to this port.
908 xbind(fd, &local_lsa->u.sa, local_lsa->len);
910 #if ENABLE_FEATURE_IPV6
911 if (family == AF_INET)
913 setsockopt_int(fd, IPPROTO_IP, IP_TOS, IPTOS_LOWDELAY);
917 /* Emit message _before_ attempted send. Think of a very short
918 * roundtrip networks: we need to go back to recv loop ASAP,
919 * to reduce delay. Printing messages after send works against that.
921 VERB1 bb_error_msg("sending query to %s", p->p_dotted);
924 * Send out a random 64-bit number as our transmit time. The NTP
925 * server will copy said number into the originate field on the
926 * response that it sends us. This is totally legal per the SNTP spec.
928 * The impact of this is two fold: we no longer send out the current
929 * system time for the world to see (which may aid an attacker), and
930 * it gives us a (not very secure) way of knowing that we're not
931 * getting spoofed by an attacker that can't capture our traffic
932 * but can spoof packets from the NTP server we're communicating with.
934 * Save the real transmit timestamp locally.
936 p->p_xmt_msg.m_xmttime.int_partl = rand();
937 p->p_xmt_msg.m_xmttime.fractionl = rand();
938 p->p_xmttime = gettime1900d();
940 /* Were doing it only if sendto worked, but
941 * loss of sync detection needs reachable_bits updated
942 * even if sending fails *locally*:
943 * "network is unreachable" because cable was pulled?
944 * We still need to declare "unsync" if this condition persists.
946 p->reachable_bits <<= 1;
948 if (do_sendto(p->p_fd, /*from:*/ NULL, /*to:*/ &p->p_lsa->u.sa, /*addrlen:*/ p->p_lsa->len,
949 &p->p_xmt_msg, NTP_MSGSIZE_NOAUTH) == -1
954 * We know that we sent nothing.
955 * We can retry *soon* without fearing
956 * that we are flooding the peer.
958 set_next(p, RETRY_INTERVAL);
962 set_next(p, RESPONSE_INTERVAL);
966 /* Note that there is no provision to prevent several run_scripts
967 * to be started in quick succession. In fact, it happens rather often
968 * if initial syncronization results in a step.
969 * You will see "step" and then "stratum" script runs, sometimes
970 * as close as only 0.002 seconds apart.
971 * Script should be ready to deal with this.
973 static void run_script(const char *action, double offset)
976 char *env1, *env2, *env3, *env4;
978 G.last_script_run = G.cur_time;
983 argv[0] = (char*) G.script_name;
984 argv[1] = (char*) action;
987 VERB1 bb_error_msg("executing '%s %s'", G.script_name, action);
989 env1 = xasprintf("%s=%u", "stratum", G.stratum);
991 env2 = xasprintf("%s=%ld", "freq_drift_ppm", G.kernel_freq_drift);
993 env3 = xasprintf("%s=%u", "poll_interval", 1 << G.poll_exp);
995 env4 = xasprintf("%s=%f", "offset", offset);
997 /* Other items of potential interest: selected peer,
998 * rootdelay, reftime, rootdisp, refid, ntp_status,
999 * last_update_offset, last_update_recv_time, discipline_jitter,
1000 * how many peers have reachable_bits = 0?
1003 /* Don't want to wait: it may run hwclock --systohc, and that
1004 * may take some time (seconds): */
1005 /*spawn_and_wait(argv);*/
1008 unsetenv("stratum");
1009 unsetenv("freq_drift_ppm");
1010 unsetenv("poll_interval");
1018 static NOINLINE void
1019 step_time(double offset)
1023 struct timeval tvc, tvn;
1024 char buf[sizeof("yyyy-mm-dd hh:mm:ss") + /*paranoia:*/ 4];
1027 gettimeofday(&tvc, NULL); /* never fails */
1028 dtime = tvc.tv_sec + (1.0e-6 * tvc.tv_usec) + offset;
1029 d_to_tv(dtime, &tvn);
1030 if (settimeofday(&tvn, NULL) == -1)
1031 bb_perror_msg_and_die("settimeofday");
1035 strftime_YYYYMMDDHHMMSS(buf, sizeof(buf), &tval);
1036 bb_error_msg("current time is %s.%06u", buf, (unsigned)tvc.tv_usec);
1039 strftime_YYYYMMDDHHMMSS(buf, sizeof(buf), &tval);
1040 bb_error_msg("setting time to %s.%06u (offset %+fs)", buf, (unsigned)tvn.tv_usec, offset);
1041 //maybe? G.FREQHOLD_cnt = 0;
1043 /* Correct various fields which contain time-relative values: */
1046 G.cur_time += offset;
1047 G.last_update_recv_time += offset;
1048 G.last_script_run += offset;
1050 /* p->lastpkt_recv_time, p->next_action_time and such: */
1051 for (item = G.ntp_peers; item != NULL; item = item->link) {
1052 peer_t *pp = (peer_t *) item->data;
1053 reset_peer_stats(pp, offset);
1054 //bb_error_msg("offset:%+f pp->next_action_time:%f -> %f",
1055 // offset, pp->next_action_time, pp->next_action_time + offset);
1056 pp->next_action_time += offset;
1057 if (pp->p_fd >= 0) {
1058 /* We wait for reply from this peer too.
1059 * But due to step we are doing, reply's data is no longer
1060 * useful (in fact, it'll be bogus). Stop waiting for it.
1064 set_next(pp, RETRY_INTERVAL);
1069 static void clamp_pollexp_and_set_MAXSTRAT(void)
1071 if (G.poll_exp < MINPOLL)
1072 G.poll_exp = MINPOLL;
1073 if (G.poll_exp > BIGPOLL)
1074 G.poll_exp = BIGPOLL;
1075 G.polladj_count = 0;
1076 G.stratum = MAXSTRAT;
1081 * Selection and clustering, and their helpers
1087 double opt_rd; /* optimization */
1090 compare_point_edge(const void *aa, const void *bb)
1092 const point_t *a = aa;
1093 const point_t *b = bb;
1094 if (a->edge < b->edge) {
1097 return (a->edge > b->edge);
1104 compare_survivor_metric(const void *aa, const void *bb)
1106 const survivor_t *a = aa;
1107 const survivor_t *b = bb;
1108 if (a->metric < b->metric) {
1111 return (a->metric > b->metric);
1114 fit(peer_t *p, double rd)
1116 if ((p->reachable_bits & (p->reachable_bits-1)) == 0) {
1117 /* One or zero bits in reachable_bits */
1118 VERB4 bb_error_msg("peer %s unfit for selection: unreachable", p->p_dotted);
1121 #if 0 /* we filter out such packets earlier */
1122 if ((p->lastpkt_status & LI_ALARM) == LI_ALARM
1123 || p->lastpkt_stratum >= MAXSTRAT
1125 VERB4 bb_error_msg("peer %s unfit for selection: bad status/stratum", p->p_dotted);
1129 /* rd is root_distance(p) */
1130 if (rd > MAXDIST + FREQ_TOLERANCE * (1 << G.poll_exp)) {
1131 VERB4 bb_error_msg("peer %s unfit for selection: root distance too high", p->p_dotted);
1135 // /* Do we have a loop? */
1136 // if (p->refid == p->dstaddr || p->refid == s.refid)
1141 select_and_cluster(void)
1146 int size = 3 * G.peer_cnt;
1147 /* for selection algorithm */
1148 point_t point[size];
1149 unsigned num_points, num_candidates;
1151 unsigned num_falsetickers;
1152 /* for cluster algorithm */
1153 survivor_t survivor[size];
1154 unsigned num_survivors;
1160 while (item != NULL) {
1163 p = (peer_t *) item->data;
1164 rd = root_distance(p);
1165 offset = p->filter_offset;
1171 VERB5 bb_error_msg("interval: [%f %f %f] %s",
1177 point[num_points].p = p;
1178 point[num_points].type = -1;
1179 point[num_points].edge = offset - rd;
1180 point[num_points].opt_rd = rd;
1182 point[num_points].p = p;
1183 point[num_points].type = 0;
1184 point[num_points].edge = offset;
1185 point[num_points].opt_rd = rd;
1187 point[num_points].p = p;
1188 point[num_points].type = 1;
1189 point[num_points].edge = offset + rd;
1190 point[num_points].opt_rd = rd;
1194 num_candidates = num_points / 3;
1195 if (num_candidates == 0) {
1196 VERB3 bb_error_msg("no valid datapoints%s", ", no peer selected");
1199 //TODO: sorting does not seem to be done in reference code
1200 qsort(point, num_points, sizeof(point[0]), compare_point_edge);
1202 /* Start with the assumption that there are no falsetickers.
1203 * Attempt to find a nonempty intersection interval containing
1204 * the midpoints of all truechimers.
1205 * If a nonempty interval cannot be found, increase the number
1206 * of assumed falsetickers by one and try again.
1207 * If a nonempty interval is found and the number of falsetickers
1208 * is less than the number of truechimers, a majority has been found
1209 * and the midpoint of each truechimer represents
1210 * the candidates available to the cluster algorithm.
1212 num_falsetickers = 0;
1215 unsigned num_midpoints = 0;
1220 for (i = 0; i < num_points; i++) {
1222 * if (point[i].type == -1) c++;
1223 * if (point[i].type == 1) c--;
1224 * and it's simpler to do it this way:
1227 if (c >= num_candidates - num_falsetickers) {
1228 /* If it was c++ and it got big enough... */
1229 low = point[i].edge;
1232 if (point[i].type == 0)
1236 for (i = num_points-1; i >= 0; i--) {
1238 if (c >= num_candidates - num_falsetickers) {
1239 high = point[i].edge;
1242 if (point[i].type == 0)
1245 /* If the number of midpoints is greater than the number
1246 * of allowed falsetickers, the intersection contains at
1247 * least one truechimer with no midpoint - bad.
1248 * Also, interval should be nonempty.
1250 if (num_midpoints <= num_falsetickers && low < high)
1253 if (num_falsetickers * 2 >= num_candidates) {
1254 VERB3 bb_error_msg("falsetickers:%d, candidates:%d%s",
1255 num_falsetickers, num_candidates,
1256 ", no peer selected");
1260 VERB4 bb_error_msg("selected interval: [%f, %f]; candidates:%d falsetickers:%d",
1261 low, high, num_candidates, num_falsetickers);
1265 /* Construct a list of survivors (p, metric)
1266 * from the chime list, where metric is dominated
1267 * first by stratum and then by root distance.
1268 * All other things being equal, this is the order of preference.
1271 for (i = 0; i < num_points; i++) {
1272 if (point[i].edge < low || point[i].edge > high)
1275 survivor[num_survivors].p = p;
1276 /* x.opt_rd == root_distance(p); */
1277 survivor[num_survivors].metric = MAXDIST * p->lastpkt_stratum + point[i].opt_rd;
1278 VERB5 bb_error_msg("survivor[%d] metric:%f peer:%s",
1279 num_survivors, survivor[num_survivors].metric, p->p_dotted);
1282 /* There must be at least MIN_SELECTED survivors to satisfy the
1283 * correctness assertions. Ordinarily, the Byzantine criteria
1284 * require four survivors, but for the demonstration here, one
1287 if (num_survivors < MIN_SELECTED) {
1288 VERB3 bb_error_msg("survivors:%d%s",
1290 ", no peer selected");
1294 //looks like this is ONLY used by the fact that later we pick survivor[0].
1295 //we can avoid sorting then, just find the minimum once!
1296 qsort(survivor, num_survivors, sizeof(survivor[0]), compare_survivor_metric);
1298 /* For each association p in turn, calculate the selection
1299 * jitter p->sjitter as the square root of the sum of squares
1300 * (p->offset - q->offset) over all q associations. The idea is
1301 * to repeatedly discard the survivor with maximum selection
1302 * jitter until a termination condition is met.
1305 unsigned max_idx = max_idx;
1306 double max_selection_jitter = max_selection_jitter;
1307 double min_jitter = min_jitter;
1309 if (num_survivors <= MIN_CLUSTERED) {
1310 VERB4 bb_error_msg("num_survivors %d <= %d, not discarding more",
1311 num_survivors, MIN_CLUSTERED);
1315 /* To make sure a few survivors are left
1316 * for the clustering algorithm to chew on,
1317 * we stop if the number of survivors
1318 * is less than or equal to MIN_CLUSTERED (3).
1320 for (i = 0; i < num_survivors; i++) {
1321 double selection_jitter_sq;
1324 if (i == 0 || p->filter_jitter < min_jitter)
1325 min_jitter = p->filter_jitter;
1327 selection_jitter_sq = 0;
1328 for (j = 0; j < num_survivors; j++) {
1329 peer_t *q = survivor[j].p;
1330 selection_jitter_sq += SQUARE(p->filter_offset - q->filter_offset);
1332 if (i == 0 || selection_jitter_sq > max_selection_jitter) {
1333 max_selection_jitter = selection_jitter_sq;
1336 VERB6 bb_error_msg("survivor %d selection_jitter^2:%f",
1337 i, selection_jitter_sq);
1339 max_selection_jitter = SQRT(max_selection_jitter / num_survivors);
1340 VERB5 bb_error_msg("max_selection_jitter (at %d):%f min_jitter:%f",
1341 max_idx, max_selection_jitter, min_jitter);
1343 /* If the maximum selection jitter is less than the
1344 * minimum peer jitter, then tossing out more survivors
1345 * will not lower the minimum peer jitter, so we might
1348 if (max_selection_jitter < min_jitter) {
1349 VERB4 bb_error_msg("max_selection_jitter:%f < min_jitter:%f, num_survivors:%d, not discarding more",
1350 max_selection_jitter, min_jitter, num_survivors);
1354 /* Delete survivor[max_idx] from the list
1355 * and go around again.
1357 VERB6 bb_error_msg("dropping survivor %d", max_idx);
1359 while (max_idx < num_survivors) {
1360 survivor[max_idx] = survivor[max_idx + 1];
1366 /* Combine the offsets of the clustering algorithm survivors
1367 * using a weighted average with weight determined by the root
1368 * distance. Compute the selection jitter as the weighted RMS
1369 * difference between the first survivor and the remaining
1370 * survivors. In some cases the inherent clock jitter can be
1371 * reduced by not using this algorithm, especially when frequent
1372 * clockhopping is involved. bbox: thus we don't do it.
1376 for (i = 0; i < num_survivors; i++) {
1378 x = root_distance(p);
1380 z += p->filter_offset / x;
1381 w += SQUARE(p->filter_offset - survivor[0].p->filter_offset) / x;
1383 //G.cluster_offset = z / y;
1384 //G.cluster_jitter = SQRT(w / y);
1387 /* Pick the best clock. If the old system peer is on the list
1388 * and at the same stratum as the first survivor on the list,
1389 * then don't do a clock hop. Otherwise, select the first
1390 * survivor on the list as the new system peer.
1393 if (G.last_update_peer
1394 && G.last_update_peer->lastpkt_stratum <= p->lastpkt_stratum
1396 /* Starting from 1 is ok here */
1397 for (i = 1; i < num_survivors; i++) {
1398 if (G.last_update_peer == survivor[i].p) {
1399 VERB5 bb_error_msg("keeping old synced peer");
1400 p = G.last_update_peer;
1405 G.last_update_peer = p;
1407 VERB4 bb_error_msg("selected peer %s filter_offset:%+f age:%f",
1410 G.cur_time - p->lastpkt_recv_time
1417 * Local clock discipline and its helpers
1420 set_new_values(int disc_state, double offset, double recv_time)
1422 /* Enter new state and set state variables. Note we use the time
1423 * of the last clock filter sample, which must be earlier than
1426 VERB4 bb_error_msg("disc_state=%d last update offset=%f recv_time=%f",
1427 disc_state, offset, recv_time);
1428 G.discipline_state = disc_state;
1429 G.last_update_offset = offset;
1430 G.last_update_recv_time = recv_time;
1432 /* Return: -1: decrease poll interval, 0: leave as is, 1: increase */
1434 update_local_clock(peer_t *p)
1438 /* Note: can use G.cluster_offset instead: */
1439 double offset = p->filter_offset;
1440 double recv_time = p->lastpkt_recv_time;
1442 #if !USING_KERNEL_PLL_LOOP
1445 #if !USING_KERNEL_PLL_LOOP || USING_INITIAL_FREQ_ESTIMATION
1446 double since_last_update;
1448 double etemp, dtemp;
1450 abs_offset = fabs(offset);
1453 /* If needed, -S script can do it by looking at $offset
1454 * env var and killing parent */
1455 /* If the offset is too large, give up and go home */
1456 if (abs_offset > PANIC_THRESHOLD) {
1457 bb_error_msg_and_die("offset %f far too big, exiting", offset);
1461 /* If this is an old update, for instance as the result
1462 * of a system peer change, avoid it. We never use
1463 * an old sample or the same sample twice.
1465 if (recv_time <= G.last_update_recv_time) {
1466 VERB3 bb_error_msg("update from %s: same or older datapoint, not using it",
1468 return 0; /* "leave poll interval as is" */
1471 /* Clock state machine transition function. This is where the
1472 * action is and defines how the system reacts to large time
1473 * and frequency errors.
1475 #if !USING_KERNEL_PLL_LOOP || USING_INITIAL_FREQ_ESTIMATION
1476 since_last_update = recv_time - G.reftime;
1478 #if !USING_KERNEL_PLL_LOOP
1481 #if USING_INITIAL_FREQ_ESTIMATION
1482 if (G.discipline_state == STATE_FREQ) {
1483 /* Ignore updates until the stepout threshold */
1484 if (since_last_update < WATCH_THRESHOLD) {
1485 VERB4 bb_error_msg("measuring drift, datapoint ignored, %f sec remains",
1486 WATCH_THRESHOLD - since_last_update);
1487 return 0; /* "leave poll interval as is" */
1489 # if !USING_KERNEL_PLL_LOOP
1490 freq_drift = (offset - G.last_update_offset) / since_last_update;
1495 /* There are two main regimes: when the
1496 * offset exceeds the step threshold and when it does not.
1498 if (abs_offset > STEP_THRESHOLD) {
1502 // This "spike state" seems to be useless, peer selection already drops
1503 // occassional "bad" datapoints. If we are here, there were _many_
1504 // large offsets. When a few first large offsets are seen,
1505 // we end up in "no valid datapoints, no peer selected" state.
1506 // Only when enough of them are seen (which means it's not a fluke),
1507 // we end up here. Looks like _our_ clock is off.
1508 switch (G.discipline_state) {
1510 /* The first outlyer: ignore it, switch to SPIK state */
1511 VERB3 bb_error_msg("update from %s: offset:%+f, spike%s",
1512 p->p_dotted, offset,
1514 G.discipline_state = STATE_SPIK;
1515 return -1; /* "decrease poll interval" */
1518 /* Ignore succeeding outlyers until either an inlyer
1519 * is found or the stepout threshold is exceeded.
1521 remains = WATCH_THRESHOLD - since_last_update;
1523 VERB3 bb_error_msg("update from %s: offset:%+f, spike%s",
1524 p->p_dotted, offset,
1525 ", datapoint ignored");
1526 return -1; /* "decrease poll interval" */
1528 /* fall through: we need to step */
1532 /* Step the time and clamp down the poll interval.
1534 * In NSET state an initial frequency correction is
1535 * not available, usually because the frequency file has
1536 * not yet been written. Since the time is outside the
1537 * capture range, the clock is stepped. The frequency
1538 * will be set directly following the stepout interval.
1540 * In FSET state the initial frequency has been set
1541 * from the frequency file. Since the time is outside
1542 * the capture range, the clock is stepped immediately,
1543 * rather than after the stepout interval. Guys get
1544 * nervous if it takes 17 minutes to set the clock for
1547 * In SPIK state the stepout threshold has expired and
1548 * the phase is still above the step threshold. Note
1549 * that a single spike greater than the step threshold
1550 * is always suppressed, even at the longer poll
1553 VERB4 bb_error_msg("stepping time by %+f; poll_exp=MINPOLL", offset);
1555 if (option_mask32 & OPT_q) {
1556 /* We were only asked to set time once. Done. */
1560 clamp_pollexp_and_set_MAXSTRAT();
1562 run_script("step", offset);
1564 recv_time += offset;
1566 #if USING_INITIAL_FREQ_ESTIMATION
1567 if (G.discipline_state == STATE_NSET) {
1568 set_new_values(STATE_FREQ, /*offset:*/ 0, recv_time);
1569 return 1; /* "ok to increase poll interval" */
1572 abs_offset = offset = 0;
1573 set_new_values(STATE_SYNC, offset, recv_time);
1574 } else { /* abs_offset <= STEP_THRESHOLD */
1576 /* The ratio is calculated before jitter is updated to make
1577 * poll adjust code more sensitive to large offsets.
1579 G.offset_to_jitter_ratio = abs_offset / G.discipline_jitter;
1581 /* Compute the clock jitter as the RMS of exponentially
1582 * weighted offset differences. Used by the poll adjust code.
1584 etemp = SQUARE(G.discipline_jitter);
1585 dtemp = SQUARE(offset - G.last_update_offset);
1586 G.discipline_jitter = SQRT(etemp + (dtemp - etemp) / AVG);
1587 if (G.discipline_jitter < G_precision_sec)
1588 G.discipline_jitter = G_precision_sec;
1590 switch (G.discipline_state) {
1592 if (option_mask32 & OPT_q) {
1593 /* We were only asked to set time once.
1594 * The clock is precise enough, no need to step.
1598 #if USING_INITIAL_FREQ_ESTIMATION
1599 /* This is the first update received and the frequency
1600 * has not been initialized. The first thing to do
1601 * is directly measure the oscillator frequency.
1603 set_new_values(STATE_FREQ, offset, recv_time);
1605 set_new_values(STATE_SYNC, offset, recv_time);
1607 VERB4 bb_error_msg("transitioning to FREQ, datapoint ignored");
1608 return 0; /* "leave poll interval as is" */
1610 #if 0 /* this is dead code for now */
1612 /* This is the first update and the frequency
1613 * has been initialized. Adjust the phase, but
1614 * don't adjust the frequency until the next update.
1616 set_new_values(STATE_SYNC, offset, recv_time);
1617 /* freq_drift remains 0 */
1621 #if USING_INITIAL_FREQ_ESTIMATION
1623 /* since_last_update >= WATCH_THRESHOLD, we waited enough.
1624 * Correct the phase and frequency and switch to SYNC state.
1625 * freq_drift was already estimated (see code above)
1627 set_new_values(STATE_SYNC, offset, recv_time);
1632 #if !USING_KERNEL_PLL_LOOP
1633 /* Compute freq_drift due to PLL and FLL contributions.
1635 * The FLL and PLL frequency gain constants
1636 * depend on the poll interval and Allan
1637 * intercept. The FLL is not used below one-half
1638 * the Allan intercept. Above that the loop gain
1639 * increases in steps to 1 / AVG.
1641 if ((1 << G.poll_exp) > ALLAN / 2) {
1642 etemp = FLL - G.poll_exp;
1645 freq_drift += (offset - G.last_update_offset) / (MAXD(since_last_update, ALLAN) * etemp);
1647 /* For the PLL the integration interval
1648 * (numerator) is the minimum of the update
1649 * interval and poll interval. This allows
1650 * oversampling, but not undersampling.
1652 etemp = MIND(since_last_update, (1 << G.poll_exp));
1653 dtemp = (4 * PLL) << G.poll_exp;
1654 freq_drift += offset * etemp / SQUARE(dtemp);
1656 set_new_values(STATE_SYNC, offset, recv_time);
1659 if (G.stratum != p->lastpkt_stratum + 1) {
1660 G.stratum = p->lastpkt_stratum + 1;
1661 run_script("stratum", offset);
1665 G.reftime = G.cur_time;
1666 G.ntp_status = p->lastpkt_status;
1667 G.refid = p->lastpkt_refid;
1668 G.rootdelay = p->lastpkt_rootdelay + p->lastpkt_delay;
1669 dtemp = p->filter_jitter; // SQRT(SQUARE(p->filter_jitter) + SQUARE(G.cluster_jitter));
1670 dtemp += MAXD(p->filter_dispersion + FREQ_TOLERANCE * (G.cur_time - p->lastpkt_recv_time) + abs_offset, MINDISP);
1671 G.rootdisp = p->lastpkt_rootdisp + dtemp;
1672 VERB4 bb_error_msg("updating leap/refid/reftime/rootdisp from peer %s", p->p_dotted);
1674 /* We are in STATE_SYNC now, but did not do adjtimex yet.
1675 * (Any other state does not reach this, they all return earlier)
1676 * By this time, freq_drift and offset are set
1677 * to values suitable for adjtimex.
1679 #if !USING_KERNEL_PLL_LOOP
1680 /* Calculate the new frequency drift and frequency stability (wander).
1681 * Compute the clock wander as the RMS of exponentially weighted
1682 * frequency differences. This is not used directly, but can,
1683 * along with the jitter, be a highly useful monitoring and
1686 dtemp = G.discipline_freq_drift + freq_drift;
1687 G.discipline_freq_drift = MAXD(MIND(MAXDRIFT, dtemp), -MAXDRIFT);
1688 etemp = SQUARE(G.discipline_wander);
1689 dtemp = SQUARE(dtemp);
1690 G.discipline_wander = SQRT(etemp + (dtemp - etemp) / AVG);
1692 VERB4 bb_error_msg("discipline freq_drift=%.9f(int:%ld corr:%e) wander=%f",
1693 G.discipline_freq_drift,
1694 (long)(G.discipline_freq_drift * 65536e6),
1696 G.discipline_wander);
1699 memset(&tmx, 0, sizeof(tmx));
1700 if (adjtimex(&tmx) < 0)
1701 bb_perror_msg_and_die("adjtimex");
1702 bb_error_msg("p adjtimex freq:%ld offset:%+ld status:0x%x tc:%ld",
1703 tmx.freq, tmx.offset, tmx.status, tmx.constant);
1706 memset(&tmx, 0, sizeof(tmx));
1708 //doesn't work, offset remains 0 (!) in kernel:
1709 //ntpd: set adjtimex freq:1786097 tmx.offset:77487
1710 //ntpd: prev adjtimex freq:1786097 tmx.offset:0
1711 //ntpd: cur adjtimex freq:1786097 tmx.offset:0
1712 tmx.modes = ADJ_FREQUENCY | ADJ_OFFSET;
1713 /* 65536 is one ppm */
1714 tmx.freq = G.discipline_freq_drift * 65536e6;
1716 tmx.modes = ADJ_OFFSET | ADJ_STATUS | ADJ_TIMECONST;// | ADJ_MAXERROR | ADJ_ESTERROR;
1718 tmx.offset = (long)(offset * 1000000); /* usec */
1719 if (SLEW_THRESHOLD < STEP_THRESHOLD) {
1720 if (tmx.offset > (long)(SLEW_THRESHOLD * 1000000)) {
1721 tmx.offset = (long)(SLEW_THRESHOLD * 1000000);
1723 if (tmx.offset < -(long)(SLEW_THRESHOLD * 1000000)) {
1724 tmx.offset = -(long)(SLEW_THRESHOLD * 1000000);
1728 tmx.status = STA_PLL;
1729 if (G.FREQHOLD_cnt != 0) {
1730 /* man adjtimex on STA_FREQHOLD:
1731 * "Normally adjustments made via ADJ_OFFSET result in dampened
1732 * frequency adjustments also being made.
1733 * This flag prevents the small frequency adjustment from being
1734 * made when correcting for an ADJ_OFFSET value."
1736 * Use this flag for a few first adjustments at the beginning
1737 * of ntpd execution, otherwise even relatively small initial
1738 * offset tend to cause largish changes to in-kernel tmx.freq.
1739 * If ntpd was restarted due to e.g. switch to another network,
1740 * this destroys already well-established tmx.freq value.
1742 if (G.FREQHOLD_cnt < 0) {
1744 // Example: a laptop whose clock runs slower when hibernated,
1745 // after wake up it still has good tmx.freq, but accumulated ~0.5 sec offset:
1746 // Run with code where initial G.FREQHOLD_cnt was always 8:
1747 //15:17:52.947 no valid datapoints, no peer selected
1748 //15:17:56.515 update from:<IP> offset:+0.485133 delay:0.157762 jitter:0.209310 clock drift:-1.393ppm tc:4
1749 //15:17:57.719 update from:<IP> offset:+0.483825 delay:0.158070 jitter:0.181159 clock drift:-1.393ppm tc:4
1750 //15:17:59.925 update from:<IP> offset:+0.479504 delay:0.158147 jitter:0.156657 clock drift:-1.393ppm tc:4
1751 //15:18:33.322 update from:<IP> offset:+0.428119 delay:0.158317 jitter:0.138071 clock drift:-1.393ppm tc:4
1752 //15:19:06.718 update from:<IP> offset:+0.376932 delay:0.158276 jitter:0.122075 clock drift:-1.393ppm tc:4
1753 //15:19:39.114 update from:<IP> offset:+0.327022 delay:0.158384 jitter:0.108538 clock drift:-1.393ppm tc:4
1754 //15:20:12.715 update from:<IP> offset:+0.275596 delay:0.158297 jitter:0.097292 clock drift:-1.393ppm tc:4
1755 //15:20:45.111 update from:<IP> offset:+0.225715 delay:0.158271 jitter:0.087841 clock drift:-1.393ppm tc:4
1756 // If allwed to continue, it would start increasing tmx.freq now.
1757 // Instead, it was ^Ced, and started anew:
1758 //15:21:15.043 no valid datapoints, no peer selected
1759 //15:21:17.408 update from:<IP> offset:+0.175910 delay:0.158314 jitter:0.076683 clock drift:-1.393ppm tc:4
1760 //15:21:19.774 update from:<IP> offset:+0.171784 delay:0.158401 jitter:0.066436 clock drift:-1.393ppm tc:4
1761 //15:21:22.140 update from:<IP> offset:+0.171660 delay:0.158592 jitter:0.057536 clock drift:-1.393ppm tc:4
1762 //15:21:22.140 update from:<IP> offset:+0.167126 delay:0.158507 jitter:0.049792 clock drift:-1.393ppm tc:4
1763 //15:21:55.696 update from:<IP> offset:+0.115223 delay:0.158277 jitter:0.050240 clock drift:-1.393ppm tc:4
1764 //15:22:29.093 update from:<IP> offset:+0.068051 delay:0.158243 jitter:0.049405 clock drift:-1.393ppm tc:5
1765 //15:23:02.490 update from:<IP> offset:+0.051632 delay:0.158215 jitter:0.043545 clock drift:-1.393ppm tc:5
1766 //15:23:34.726 update from:<IP> offset:+0.039984 delay:0.158157 jitter:0.038106 clock drift:-1.393ppm tc:5
1767 // STA_FREQHOLD no longer set, started increasing tmx.freq now:
1768 //15:24:06.961 update from:<IP> offset:+0.030968 delay:0.158190 jitter:0.033306 clock drift:+2.387ppm tc:5
1769 //15:24:40.357 update from:<IP> offset:+0.023648 delay:0.158211 jitter:0.029072 clock drift:+5.454ppm tc:5
1770 //15:25:13.774 update from:<IP> offset:+0.018068 delay:0.157660 jitter:0.025288 clock drift:+7.728ppm tc:5
1771 //15:26:19.173 update from:<IP> offset:+0.010057 delay:0.157969 jitter:0.022255 clock drift:+8.361ppm tc:6
1772 //15:27:26.602 update from:<IP> offset:+0.006737 delay:0.158103 jitter:0.019316 clock drift:+8.792ppm tc:6
1773 //15:28:33.030 update from:<IP> offset:+0.004513 delay:0.158294 jitter:0.016765 clock drift:+9.080ppm tc:6
1774 //15:29:40.617 update from:<IP> offset:+0.002787 delay:0.157745 jitter:0.014543 clock drift:+9.258ppm tc:6
1775 //15:30:47.045 update from:<IP> offset:+0.001324 delay:0.157709 jitter:0.012594 clock drift:+9.342ppm tc:6
1776 //15:31:53.473 update from:<IP> offset:+0.000007 delay:0.158142 jitter:0.010922 clock drift:+9.343ppm tc:6
1777 //15:32:58.902 update from:<IP> offset:-0.000728 delay:0.158222 jitter:0.009454 clock drift:+9.298ppm tc:6
1779 * This expression would choose 15 in the above example.
1781 G.FREQHOLD_cnt = 8 + ((unsigned)(abs(tmx.offset)) >> 16);
1784 tmx.status |= STA_FREQHOLD;
1786 if (G.ntp_status & LI_PLUSSEC)
1787 tmx.status |= STA_INS;
1788 if (G.ntp_status & LI_MINUSSEC)
1789 tmx.status |= STA_DEL;
1791 tmx.constant = (int)G.poll_exp - 4;
1793 * The below if statement should be unnecessary, but...
1794 * It looks like Linux kernel's PLL is far too gentle in changing
1795 * tmx.freq in response to clock offset. Offset keeps growing
1796 * and eventually we fall back to smaller poll intervals.
1797 * We can make correction more aggressive (about x2) by supplying
1798 * PLL time constant which is one less than the real one.
1799 * To be on a safe side, let's do it only if offset is significantly
1800 * larger than jitter.
1802 if (G.offset_to_jitter_ratio >= TIMECONST_HACK_GATE)
1804 if (tmx.constant < 0)
1807 //tmx.esterror = (uint32_t)(clock_jitter * 1e6);
1808 //tmx.maxerror = (uint32_t)((sys_rootdelay / 2 + sys_rootdisp) * 1e6);
1809 rc = adjtimex(&tmx);
1811 bb_perror_msg_and_die("adjtimex");
1812 /* NB: here kernel returns constant == G.poll_exp, not == G.poll_exp - 4.
1813 * Not sure why. Perhaps it is normal.
1815 VERB4 bb_error_msg("adjtimex:%d freq:%ld offset:%+ld status:0x%x",
1816 rc, tmx.freq, tmx.offset, tmx.status);
1817 G.kernel_freq_drift = tmx.freq / 65536;
1818 VERB2 bb_error_msg("update from:%s offset:%+f delay:%f jitter:%f clock drift:%+.3fppm tc:%d",
1822 G.discipline_jitter,
1823 (double)tmx.freq / 65536,
1827 return 1; /* "ok to increase poll interval" */
1832 * We've got a new reply packet from a peer, process it
1836 poll_interval(int upper_bound)
1838 unsigned interval, r, mask;
1839 interval = 1 << G.poll_exp;
1840 if (interval > upper_bound)
1841 interval = upper_bound;
1842 mask = ((interval-1) >> 4) | 1;
1844 interval += r & mask; /* ~ random(0..1) * interval/16 */
1845 VERB4 bb_error_msg("chose poll interval:%u (poll_exp:%d)", interval, G.poll_exp);
1849 adjust_poll(int count)
1851 G.polladj_count += count;
1852 if (G.polladj_count > POLLADJ_LIMIT) {
1853 G.polladj_count = 0;
1854 if (G.poll_exp < MAXPOLL) {
1856 VERB4 bb_error_msg("polladj: discipline_jitter:%f ++poll_exp=%d",
1857 G.discipline_jitter, G.poll_exp);
1859 } else if (G.polladj_count < -POLLADJ_LIMIT || (count < 0 && G.poll_exp > BIGPOLL)) {
1860 G.polladj_count = 0;
1861 if (G.poll_exp > MINPOLL) {
1865 /* Correct p->next_action_time in each peer
1866 * which waits for sending, so that they send earlier.
1867 * Old pp->next_action_time are on the order
1868 * of t + (1 << old_poll_exp) + small_random,
1869 * we simply need to subtract ~half of that.
1871 for (item = G.ntp_peers; item != NULL; item = item->link) {
1872 peer_t *pp = (peer_t *) item->data;
1874 pp->next_action_time -= (1 << G.poll_exp);
1876 VERB4 bb_error_msg("polladj: discipline_jitter:%f --poll_exp=%d",
1877 G.discipline_jitter, G.poll_exp);
1880 VERB4 bb_error_msg("polladj: count:%d", G.polladj_count);
1883 static NOINLINE void
1884 recv_and_process_peer_pkt(peer_t *p)
1889 double T1, T2, T3, T4;
1891 double prev_delay, delay;
1893 datapoint_t *datapoint;
1898 /* We can recvfrom here and check from.IP, but some multihomed
1899 * ntp servers reply from their *other IP*.
1900 * TODO: maybe we should check at least what we can: from.port == 123?
1903 size = recv(p->p_fd, &msg, sizeof(msg), MSG_DONTWAIT);
1908 if (errno == EAGAIN)
1909 /* There was no packet after all
1910 * (poll() returning POLLIN for a fd
1911 * is not a ironclad guarantee that data is there)
1915 * If you need a different handling for a specific
1916 * errno, always explain it in comment.
1918 bb_perror_msg_and_die("recv(%s) error", p->p_dotted);
1921 if (size != NTP_MSGSIZE_NOAUTH && size != NTP_MSGSIZE) {
1922 bb_error_msg("malformed packet received from %s", p->p_dotted);
1926 if (msg.m_orgtime.int_partl != p->p_xmt_msg.m_xmttime.int_partl
1927 || msg.m_orgtime.fractionl != p->p_xmt_msg.m_xmttime.fractionl
1929 /* Somebody else's packet */
1933 /* We do not expect any more packets from this peer for now.
1934 * Closing the socket informs kernel about it.
1935 * We open a new socket when we send a new query.
1940 if ((msg.m_status & LI_ALARM) == LI_ALARM
1941 || msg.m_stratum == 0
1942 || msg.m_stratum > NTP_MAXSTRATUM
1944 bb_error_msg("reply from %s: peer is unsynced", p->p_dotted);
1946 * Stratum 0 responses may have commands in 32-bit m_refid field:
1947 * "DENY", "RSTR" - peer does not like us at all,
1948 * "RATE" - peer is overloaded, reduce polling freq.
1949 * If poll interval is small, increase it.
1951 if (G.poll_exp < BIGPOLL)
1952 goto increase_interval;
1953 goto pick_normal_interval;
1956 // /* Verify valid root distance */
1957 // if (msg.m_rootdelay / 2 + msg.m_rootdisp >= MAXDISP || p->lastpkt_reftime > msg.m_xmt)
1958 // return; /* invalid header values */
1961 * From RFC 2030 (with a correction to the delay math):
1963 * Timestamp Name ID When Generated
1964 * ------------------------------------------------------------
1965 * Originate Timestamp T1 time request sent by client
1966 * Receive Timestamp T2 time request received by server
1967 * Transmit Timestamp T3 time reply sent by server
1968 * Destination Timestamp T4 time reply received by client
1970 * The roundtrip delay and local clock offset are defined as
1972 * delay = (T4 - T1) - (T3 - T2); offset = ((T2 - T1) + (T3 - T4)) / 2
1975 T2 = lfp_to_d(msg.m_rectime);
1976 T3 = lfp_to_d(msg.m_xmttime);
1979 /* The delay calculation is a special case. In cases where the
1980 * server and client clocks are running at different rates and
1981 * with very fast networks, the delay can appear negative. In
1982 * order to avoid violating the Principle of Least Astonishment,
1983 * the delay is clamped not less than the system precision.
1985 delay = (T4 - T1) - (T3 - T2);
1986 if (delay < G_precision_sec)
1987 delay = G_precision_sec;
1989 * If this packet's delay is much bigger than the last one,
1990 * it's better to just ignore it than use its much less precise value.
1992 prev_delay = p->p_raw_delay;
1993 p->p_raw_delay = delay;
1994 if (p->reachable_bits && delay > prev_delay * BAD_DELAY_GROWTH) {
1995 bb_error_msg("reply from %s: delay %f is too high, ignoring", p->p_dotted, delay);
1996 goto pick_normal_interval;
1999 p->lastpkt_delay = delay;
2000 p->lastpkt_recv_time = T4;
2001 VERB6 bb_error_msg("%s->lastpkt_recv_time=%f", p->p_dotted, p->lastpkt_recv_time);
2002 p->lastpkt_status = msg.m_status;
2003 p->lastpkt_stratum = msg.m_stratum;
2004 p->lastpkt_rootdelay = sfp_to_d(msg.m_rootdelay);
2005 p->lastpkt_rootdisp = sfp_to_d(msg.m_rootdisp);
2006 p->lastpkt_refid = msg.m_refid;
2008 p->datapoint_idx = p->reachable_bits ? (p->datapoint_idx + 1) % NUM_DATAPOINTS : 0;
2009 datapoint = &p->filter_datapoint[p->datapoint_idx];
2010 datapoint->d_recv_time = T4;
2011 datapoint->d_offset = offset = ((T2 - T1) + (T3 - T4)) / 2;
2012 datapoint->d_dispersion = LOG2D(msg.m_precision_exp) + G_precision_sec;
2013 if (!p->reachable_bits) {
2014 /* 1st datapoint ever - replicate offset in every element */
2016 for (i = 0; i < NUM_DATAPOINTS; i++) {
2017 p->filter_datapoint[i].d_offset = offset;
2021 p->reachable_bits |= 1;
2022 if ((MAX_VERBOSE && G.verbose) || (option_mask32 & OPT_w)) {
2023 bb_error_msg("reply from %s: offset:%+f delay:%f status:0x%02x strat:%d refid:0x%08x rootdelay:%f reach:0x%02x",
2030 p->lastpkt_rootdelay,
2032 /* not shown: m_ppoll, m_precision_exp, m_rootdisp,
2033 * m_reftime, m_orgtime, m_rectime, m_xmttime
2038 /* Muck with statictics and update the clock */
2039 filter_datapoints(p);
2040 q = select_and_cluster();
2043 if (!(option_mask32 & OPT_w)) {
2044 rc = update_local_clock(q);
2046 //Disabled this because there is a case where largish offsets
2047 //are unavoidable: if network round-trip delay is, say, ~0.6s,
2048 //error in offset estimation would be ~delay/2 ~= 0.3s.
2049 //Thus, offsets will be usually in -0.3...0.3s range.
2050 //In this case, this code would keep poll interval small,
2051 //but it won't be helping.
2052 //BIGOFF check below deals with a case of seeing multi-second offsets.
2054 /* If drift is dangerously large, immediately
2055 * drop poll interval one step down.
2057 if (fabs(q->filter_offset) >= POLLDOWN_OFFSET) {
2058 VERB4 bb_error_msg("offset:%+f > POLLDOWN_OFFSET", q->filter_offset);
2059 adjust_poll(-POLLADJ_LIMIT * 3);
2065 /* No peer selected.
2066 * If poll interval is small, increase it.
2068 if (G.poll_exp < BIGPOLL)
2069 goto increase_interval;
2073 /* Adjust the poll interval by comparing the current offset
2074 * with the clock jitter. If the offset is less than
2075 * the clock jitter times a constant, then the averaging interval
2076 * is increased, otherwise it is decreased. A bit of hysteresis
2077 * helps calm the dance. Works best using burst mode.
2079 if (rc > 0 && G.offset_to_jitter_ratio <= POLLADJ_GATE) {
2080 /* was += G.poll_exp but it is a bit
2081 * too optimistic for my taste at high poll_exp's */
2083 adjust_poll(MINPOLL);
2086 bb_error_msg("want smaller interval: offset/jitter = %u",
2087 G.offset_to_jitter_ratio);
2088 adjust_poll(-G.poll_exp * 2);
2092 /* Decide when to send new query for this peer */
2093 pick_normal_interval:
2094 interval = poll_interval(INT_MAX);
2095 if (fabs(offset) >= BIGOFF && interval > BIGOFF_INTERVAL) {
2096 /* If we are synced, offsets are less than SLEW_THRESHOLD,
2097 * or at the very least not much larger than it.
2098 * Now we see a largish one.
2099 * Either this peer is feeling bad, or packet got corrupted,
2100 * or _our_ clock is wrong now and _all_ peers will show similar
2101 * largish offsets too.
2102 * I observed this with laptop suspend stopping clock.
2103 * In any case, it makes sense to make next request soonish:
2104 * cases 1 and 2: get a better datapoint,
2105 * case 3: allows to resync faster.
2107 interval = BIGOFF_INTERVAL;
2110 set_next(p, interval);
2113 #if ENABLE_FEATURE_NTPD_SERVER
2114 static NOINLINE void
2115 recv_and_process_client_pkt(void /*int fd*/)
2119 len_and_sockaddr *to;
2120 struct sockaddr *from;
2122 uint8_t query_status;
2123 l_fixedpt_t query_xmttime;
2125 to = get_sock_lsa(G_listen_fd);
2126 from = xzalloc(to->len);
2128 size = recv_from_to(G_listen_fd, &msg, sizeof(msg), MSG_DONTWAIT, from, &to->u.sa, to->len);
2129 if (size != NTP_MSGSIZE_NOAUTH && size != NTP_MSGSIZE) {
2132 if (errno == EAGAIN)
2134 bb_perror_msg_and_die("recv");
2136 addr = xmalloc_sockaddr2dotted_noport(from);
2137 bb_error_msg("malformed packet received from %s: size %u", addr, (int)size);
2142 /* Respond only to client and symmetric active packets */
2143 if ((msg.m_status & MODE_MASK) != MODE_CLIENT
2144 && (msg.m_status & MODE_MASK) != MODE_SYM_ACT
2149 query_status = msg.m_status;
2150 query_xmttime = msg.m_xmttime;
2152 /* Build a reply packet */
2153 memset(&msg, 0, sizeof(msg));
2154 msg.m_status = G.stratum < MAXSTRAT ? (G.ntp_status & LI_MASK) : LI_ALARM;
2155 msg.m_status |= (query_status & VERSION_MASK);
2156 msg.m_status |= ((query_status & MODE_MASK) == MODE_CLIENT) ?
2157 MODE_SERVER : MODE_SYM_PAS;
2158 msg.m_stratum = G.stratum;
2159 msg.m_ppoll = G.poll_exp;
2160 msg.m_precision_exp = G_precision_exp;
2161 /* this time was obtained between poll() and recv() */
2162 msg.m_rectime = d_to_lfp(G.cur_time);
2163 msg.m_xmttime = d_to_lfp(gettime1900d()); /* this instant */
2164 if (G.peer_cnt == 0) {
2165 /* we have no peers: "stratum 1 server" mode. reftime = our own time */
2166 G.reftime = G.cur_time;
2168 msg.m_reftime = d_to_lfp(G.reftime);
2169 msg.m_orgtime = query_xmttime;
2170 msg.m_rootdelay = d_to_sfp(G.rootdelay);
2171 //simple code does not do this, fix simple code!
2172 msg.m_rootdisp = d_to_sfp(G.rootdisp);
2173 //version = (query_status & VERSION_MASK); /* ... >> VERSION_SHIFT - done below instead */
2174 msg.m_refid = G.refid; // (version > (3 << VERSION_SHIFT)) ? G.refid : G.refid3;
2176 /* We reply from the local address packet was sent to,
2177 * this makes to/from look swapped here: */
2178 do_sendto(G_listen_fd,
2179 /*from:*/ &to->u.sa, /*to:*/ from, /*addrlen:*/ to->len,
2188 /* Upstream ntpd's options:
2190 * -4 Force DNS resolution of host names to the IPv4 namespace.
2191 * -6 Force DNS resolution of host names to the IPv6 namespace.
2192 * -a Require cryptographic authentication for broadcast client,
2193 * multicast client and symmetric passive associations.
2194 * This is the default.
2195 * -A Do not require cryptographic authentication for broadcast client,
2196 * multicast client and symmetric passive associations.
2197 * This is almost never a good idea.
2198 * -b Enable the client to synchronize to broadcast servers.
2200 * Specify the name and path of the configuration file,
2201 * default /etc/ntp.conf
2202 * -d Specify debugging mode. This option may occur more than once,
2203 * with each occurrence indicating greater detail of display.
2205 * Specify debugging level directly.
2207 * Specify the name and path of the frequency file.
2208 * This is the same operation as the "driftfile FILE"
2209 * configuration command.
2210 * -g Normally, ntpd exits with a message to the system log
2211 * if the offset exceeds the panic threshold, which is 1000 s
2212 * by default. This option allows the time to be set to any value
2213 * without restriction; however, this can happen only once.
2214 * If the threshold is exceeded after that, ntpd will exit
2215 * with a message to the system log. This option can be used
2216 * with the -q and -x options. See the tinker command for other options.
2218 * Chroot the server to the directory jaildir. This option also implies
2219 * that the server attempts to drop root privileges at startup
2220 * (otherwise, chroot gives very little additional security).
2221 * You may need to also specify a -u option.
2223 * Specify the name and path of the symmetric key file,
2224 * default /etc/ntp/keys. This is the same operation
2225 * as the "keys FILE" configuration command.
2227 * Specify the name and path of the log file. The default
2228 * is the system log file. This is the same operation as
2229 * the "logfile FILE" configuration command.
2230 * -L Do not listen to virtual IPs. The default is to listen.
2232 * -N To the extent permitted by the operating system,
2233 * run the ntpd at the highest priority.
2235 * Specify the name and path of the file used to record the ntpd
2236 * process ID. This is the same operation as the "pidfile FILE"
2237 * configuration command.
2239 * To the extent permitted by the operating system,
2240 * run the ntpd at the specified priority.
2241 * -q Exit the ntpd just after the first time the clock is set.
2242 * This behavior mimics that of the ntpdate program, which is
2243 * to be retired. The -g and -x options can be used with this option.
2244 * Note: The kernel time discipline is disabled with this option.
2246 * Specify the default propagation delay from the broadcast/multicast
2247 * server to this client. This is necessary only if the delay
2248 * cannot be computed automatically by the protocol.
2250 * Specify the directory path for files created by the statistics
2251 * facility. This is the same operation as the "statsdir DIR"
2252 * configuration command.
2254 * Add a key number to the trusted key list. This option can occur
2257 * Specify a user, and optionally a group, to switch to.
2260 * Add a system variable listed by default.
2261 * -x Normally, the time is slewed if the offset is less than the step
2262 * threshold, which is 128 ms by default, and stepped if above
2263 * the threshold. This option sets the threshold to 600 s, which is
2264 * well within the accuracy window to set the clock manually.
2265 * Note: since the slew rate of typical Unix kernels is limited
2266 * to 0.5 ms/s, each second of adjustment requires an amortization
2267 * interval of 2000 s. Thus, an adjustment as much as 600 s
2268 * will take almost 14 days to complete. This option can be used
2269 * with the -g and -q options. See the tinker command for other options.
2270 * Note: The kernel time discipline is disabled with this option.
2273 /* By doing init in a separate function we decrease stack usage
2276 static NOINLINE void ntp_init(char **argv)
2284 bb_error_msg_and_die(bb_msg_you_must_be_root);
2286 /* Set some globals */
2287 G.discipline_jitter = G_precision_sec;
2288 G.stratum = MAXSTRAT;
2290 G.poll_exp = BURSTPOLL; /* speeds up initial sync */
2291 G.last_script_run = G.reftime = G.last_update_recv_time = gettime1900d(); /* sets G.cur_time too */
2292 G.FREQHOLD_cnt = -1;
2296 opts = getopt32(argv, "^"
2298 "wp:*S:"IF_FEATURE_NTPD_SERVER("l") /* NOT compat */
2299 IF_FEATURE_NTPD_SERVER("I:") /* compat */
2301 "46aAbgL" /* compat, ignored */
2303 "dd:wn" /* -d: counter; -p: list; -w implies -n */
2304 IF_FEATURE_NTPD_SERVER(":Il") /* -I implies -l */
2305 , &peers, &G.script_name,
2306 #if ENABLE_FEATURE_NTPD_SERVER
2311 // if (opts & OPT_x) /* disable stepping, only slew is allowed */
2312 // G.time_was_stepped = 1;
2314 #if ENABLE_FEATURE_NTPD_SERVER
2317 G_listen_fd = create_and_bind_dgram_or_die(NULL, 123);
2319 if (setsockopt_bindtodevice(G_listen_fd, G.if_name))
2322 socket_want_pktinfo(G_listen_fd);
2323 setsockopt_int(G_listen_fd, IPPROTO_IP, IP_TOS, IPTOS_LOWDELAY);
2326 /* I hesitate to set -20 prio. -15 should be high enough for timekeeping */
2328 setpriority(PRIO_PROCESS, 0, -15);
2330 if (!(opts & OPT_n)) {
2331 bb_daemonize_or_rexec(DAEMON_DEVNULL_STDIO, argv);
2332 logmode = LOGMODE_NONE;
2337 add_peers(llist_pop(&peers));
2339 #if ENABLE_FEATURE_NTPD_CONF
2344 parser = config_open("/etc/ntp.conf");
2345 while (config_read(parser, token, 3, 1, "# \t", PARSE_NORMAL)) {
2346 if (strcmp(token[0], "server") == 0 && token[1]) {
2347 add_peers(token[1]);
2350 bb_error_msg("skipping %s:%u: unimplemented command '%s'",
2351 "/etc/ntp.conf", parser->lineno, token[0]
2354 config_close(parser);
2357 if (G.peer_cnt == 0) {
2358 if (!(opts & OPT_l))
2360 /* -l but no peers: "stratum 1 server" mode */
2363 /* If network is up, syncronization occurs in ~10 seconds.
2364 * We give "ntpd -q" 10 seconds to get first reply,
2365 * then another 50 seconds to finish syncing.
2367 * I tested ntpd 4.2.6p1 and apparently it never exits
2368 * (will try forever), but it does not feel right.
2369 * The goal of -q is to act like ntpdate: set time
2370 * after a reasonably small period of polling, or fail.
2373 option_mask32 |= OPT_qq;
2390 int ntpd_main(int argc UNUSED_PARAM, char **argv) MAIN_EXTERNALLY_VISIBLE;
2391 int ntpd_main(int argc UNUSED_PARAM, char **argv)
2399 memset(&G, 0, sizeof(G));
2400 SET_PTR_TO_GLOBALS(&G);
2404 /* If ENABLE_FEATURE_NTPD_SERVER, + 1 for listen_fd: */
2405 cnt = G.peer_cnt + ENABLE_FEATURE_NTPD_SERVER;
2406 idx2peer = xzalloc(sizeof(idx2peer[0]) * cnt);
2407 pfd = xzalloc(sizeof(pfd[0]) * cnt);
2409 /* Countdown: we never sync before we sent INITIAL_SAMPLES+1
2410 * packets to each peer.
2411 * NB: if some peer is not responding, we may end up sending
2412 * fewer packets to it and more to other peers.
2413 * NB2: sync usually happens using INITIAL_SAMPLES packets,
2414 * since last reply does not come back instantaneously.
2416 cnt = G.peer_cnt * (INITIAL_SAMPLES + 1);
2418 write_pidfile(CONFIG_PID_FILE_PATH "/ntpd.pid");
2420 while (!bb_got_signal) {
2426 /* Nothing between here and poll() blocks for any significant time */
2428 nextaction = G.last_script_run + (11*60);
2429 if (nextaction < G.cur_time + 1)
2430 nextaction = G.cur_time + 1;
2433 #if ENABLE_FEATURE_NTPD_SERVER
2434 if (G_listen_fd != -1) {
2435 pfd[0].fd = G_listen_fd;
2436 pfd[0].events = POLLIN;
2440 /* Pass over peer list, send requests, time out on receives */
2441 for (item = G.ntp_peers; item != NULL; item = item->link) {
2442 peer_t *p = (peer_t *) item->data;
2444 if (p->next_action_time <= G.cur_time) {
2445 if (p->p_fd == -1) {
2446 /* Time to send new req */
2448 VERB4 bb_error_msg("disabling burst mode");
2449 G.polladj_count = 0;
2450 G.poll_exp = MINPOLL;
2452 send_query_to_peer(p);
2454 /* Timed out waiting for reply */
2457 /* If poll interval is small, increase it */
2458 if (G.poll_exp < BIGPOLL)
2459 adjust_poll(MINPOLL);
2460 timeout = poll_interval(NOREPLY_INTERVAL);
2461 bb_error_msg("timed out waiting for %s, reach 0x%02x, next query in %us",
2462 p->p_dotted, p->reachable_bits, timeout);
2464 /* What if don't see it because it changed its IP? */
2465 if (p->reachable_bits == 0)
2466 resolve_peer_hostname(p);
2468 set_next(p, timeout);
2472 if (p->next_action_time < nextaction)
2473 nextaction = p->next_action_time;
2476 /* Wait for reply from this peer */
2477 pfd[i].fd = p->p_fd;
2478 pfd[i].events = POLLIN;
2484 timeout = nextaction - G.cur_time;
2487 timeout++; /* (nextaction - G.cur_time) rounds down, compensating */
2489 /* Here we may block */
2491 if (i > (ENABLE_FEATURE_NTPD_SERVER && G_listen_fd != -1)) {
2492 /* We wait for at least one reply.
2493 * Poll for it, without wasting time for message.
2494 * Since replies often come under 1 second, this also
2495 * reduces clutter in logs.
2497 nfds = poll(pfd, i, 1000);
2503 bb_error_msg("poll:%us sockets:%u interval:%us", timeout, i, 1 << G.poll_exp);
2505 nfds = poll(pfd, i, timeout * 1000);
2507 gettime1900d(); /* sets G.cur_time */
2513 break; /* poll was interrupted by a signal */
2515 if (G.cur_time - G.last_script_run > 11*60) {
2516 /* Useful for updating battery-backed RTC and such */
2517 run_script("periodic", G.last_update_offset);
2518 gettime1900d(); /* sets G.cur_time */
2521 /* Resolve peer names to IPs, if not resolved yet.
2522 * We do it only when poll timed out:
2523 * this way, we almost never overlap DNS resolution with
2524 * "request-reply" packet round trip.
2528 for (item = G.ntp_peers; item != NULL; item = item->link) {
2529 peer_t *p = (peer_t *) item->data;
2530 if (p->next_action_time <= ct && !p->p_lsa) {
2531 /* This can take up to ~10 sec per each DNS query */
2532 dns_error |= (!resolve_peer_hostname(p));
2537 /* Set next time for those which are still not resolved */
2538 gettime1900d(); /* sets G.cur_time (needed for set_next()) */
2539 for (item = G.ntp_peers; item != NULL; item = item->link) {
2540 peer_t *p = (peer_t *) item->data;
2541 if (p->next_action_time <= ct && !p->p_lsa) {
2542 set_next(p, HOSTNAME_INTERVAL * p->dns_errors);
2548 /* Process any received packets */
2550 #if ENABLE_FEATURE_NTPD_SERVER
2551 if (G.listen_fd != -1) {
2552 if (pfd[0].revents /* & (POLLIN|POLLERR)*/) {
2554 recv_and_process_client_pkt(/*G.listen_fd*/);
2555 gettime1900d(); /* sets G.cur_time */
2560 for (; nfds != 0 && j < i; j++) {
2561 if (pfd[j].revents /* & (POLLIN|POLLERR)*/) {
2563 * At init, alarm was set to 10 sec.
2564 * Now we did get a reply.
2565 * Increase timeout to 50 seconds to finish syncing.
2567 if (option_mask32 & OPT_qq) {
2568 option_mask32 &= ~OPT_qq;
2572 recv_and_process_peer_pkt(idx2peer[j]);
2573 gettime1900d(); /* sets G.cur_time */
2578 if (G.ntp_peers && G.stratum != MAXSTRAT) {
2579 for (item = G.ntp_peers; item != NULL; item = item->link) {
2580 peer_t *p = (peer_t *) item->data;
2581 if (p->reachable_bits)
2582 goto have_reachable_peer;
2584 /* No peer responded for last 8 packets, panic */
2585 clamp_pollexp_and_set_MAXSTRAT();
2586 run_script("unsync", 0.0);
2587 have_reachable_peer: ;
2589 } /* while (!bb_got_signal) */
2591 remove_pidfile(CONFIG_PID_FILE_PATH "/ntpd.pid");
2592 kill_myself_with_sig(bb_got_signal);
2600 /*** openntpd-4.6 uses only adjtime, not adjtimex ***/
2602 /*** ntp-4.2.6/ntpd/ntp_loopfilter.c - adjtimex usage ***/
2606 direct_freq(double fp_offset)
2610 * If the kernel is enabled, we need the residual offset to
2611 * calculate the frequency correction.
2613 if (pll_control && kern_enable) {
2614 memset(&ntv, 0, sizeof(ntv));
2617 clock_offset = ntv.offset / 1e9;
2618 #else /* STA_NANO */
2619 clock_offset = ntv.offset / 1e6;
2620 #endif /* STA_NANO */
2621 drift_comp = FREQTOD(ntv.freq);
2623 #endif /* KERNEL_PLL */
2624 set_freq((fp_offset - clock_offset) / (current_time - clock_epoch) + drift_comp);
2630 set_freq(double freq) /* frequency update */
2638 * If the kernel is enabled, update the kernel frequency.
2640 if (pll_control && kern_enable) {
2641 memset(&ntv, 0, sizeof(ntv));
2642 ntv.modes = MOD_FREQUENCY;
2643 ntv.freq = DTOFREQ(drift_comp);
2645 snprintf(tbuf, sizeof(tbuf), "kernel %.3f PPM", drift_comp * 1e6);
2646 report_event(EVNT_FSET, NULL, tbuf);
2648 snprintf(tbuf, sizeof(tbuf), "ntpd %.3f PPM", drift_comp * 1e6);
2649 report_event(EVNT_FSET, NULL, tbuf);
2651 #else /* KERNEL_PLL */
2652 snprintf(tbuf, sizeof(tbuf), "ntpd %.3f PPM", drift_comp * 1e6);
2653 report_event(EVNT_FSET, NULL, tbuf);
2654 #endif /* KERNEL_PLL */
2663 * This code segment works when clock adjustments are made using
2664 * precision time kernel support and the ntp_adjtime() system
2665 * call. This support is available in Solaris 2.6 and later,
2666 * Digital Unix 4.0 and later, FreeBSD, Linux and specially
2667 * modified kernels for HP-UX 9 and Ultrix 4. In the case of the
2668 * DECstation 5000/240 and Alpha AXP, additional kernel
2669 * modifications provide a true microsecond clock and nanosecond
2670 * clock, respectively.
2672 * Important note: The kernel discipline is used only if the
2673 * step threshold is less than 0.5 s, as anything higher can
2674 * lead to overflow problems. This might occur if some misguided
2675 * lad set the step threshold to something ridiculous.
2677 if (pll_control && kern_enable) {
2679 #define MOD_BITS (MOD_OFFSET | MOD_MAXERROR | MOD_ESTERROR | MOD_STATUS | MOD_TIMECONST)
2682 * We initialize the structure for the ntp_adjtime()
2683 * system call. We have to convert everything to
2684 * microseconds or nanoseconds first. Do not update the
2685 * system variables if the ext_enable flag is set. In
2686 * this case, the external clock driver will update the
2687 * variables, which will be read later by the local
2688 * clock driver. Afterwards, remember the time and
2689 * frequency offsets for jitter and stability values and
2690 * to update the frequency file.
2692 memset(&ntv, 0, sizeof(ntv));
2694 ntv.modes = MOD_STATUS;
2697 ntv.modes = MOD_BITS | MOD_NANO;
2698 #else /* STA_NANO */
2699 ntv.modes = MOD_BITS;
2700 #endif /* STA_NANO */
2701 if (clock_offset < 0)
2706 ntv.offset = (int32)(clock_offset * 1e9 + dtemp);
2707 ntv.constant = sys_poll;
2708 #else /* STA_NANO */
2709 ntv.offset = (int32)(clock_offset * 1e6 + dtemp);
2710 ntv.constant = sys_poll - 4;
2711 #endif /* STA_NANO */
2712 ntv.esterror = (u_int32)(clock_jitter * 1e6);
2713 ntv.maxerror = (u_int32)((sys_rootdelay / 2 + sys_rootdisp) * 1e6);
2714 ntv.status = STA_PLL;
2717 * Enable/disable the PPS if requested.
2720 if (!(pll_status & STA_PPSTIME))
2721 report_event(EVNT_KERN,
2722 NULL, "PPS enabled");
2723 ntv.status |= STA_PPSTIME | STA_PPSFREQ;
2725 if (pll_status & STA_PPSTIME)
2726 report_event(EVNT_KERN,
2727 NULL, "PPS disabled");
2728 ntv.status &= ~(STA_PPSTIME | STA_PPSFREQ);
2730 if (sys_leap == LEAP_ADDSECOND)
2731 ntv.status |= STA_INS;
2732 else if (sys_leap == LEAP_DELSECOND)
2733 ntv.status |= STA_DEL;
2737 * Pass the stuff to the kernel. If it squeals, turn off
2738 * the pps. In any case, fetch the kernel offset,
2739 * frequency and jitter.
2741 if (ntp_adjtime(&ntv) == TIME_ERROR) {
2742 if (!(ntv.status & STA_PPSSIGNAL))
2743 report_event(EVNT_KERN, NULL,
2746 pll_status = ntv.status;
2748 clock_offset = ntv.offset / 1e9;
2749 #else /* STA_NANO */
2750 clock_offset = ntv.offset / 1e6;
2751 #endif /* STA_NANO */
2752 clock_frequency = FREQTOD(ntv.freq);
2755 * If the kernel PPS is lit, monitor its performance.
2757 if (ntv.status & STA_PPSTIME) {
2759 clock_jitter = ntv.jitter / 1e9;
2760 #else /* STA_NANO */
2761 clock_jitter = ntv.jitter / 1e6;
2762 #endif /* STA_NANO */
2765 #if defined(STA_NANO) && NTP_API == 4
2767 * If the TAI changes, update the kernel TAI.
2769 if (loop_tai != sys_tai) {
2771 ntv.modes = MOD_TAI;
2772 ntv.constant = sys_tai;
2775 #endif /* STA_NANO */
2777 #endif /* KERNEL_PLL */