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_DSCP_AF21 definition */
97 #include <sys/timex.h>
98 #ifndef IPTOS_DSCP_AF21
99 # define IPTOS_DSCP_AF21 0x48
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 MIN_FREQHOLD 12 /* adjust offset, but not freq in this many first adjustments */
153 #define BAD_DELAY_GROWTH 4 /* drop packet if its delay grew by more than this factor */
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 //^^^^^^^^^^^^^^^^^^^^^^^^^^ TODO: man adjtimex about tmx.offset:
169 // "Since Linux 2.6.26, the supplied value is clamped to the range (-0.5s, +0.5s)"
170 // - can use this larger value instead?
172 /* Stepout threshold (sec). std ntpd uses 900 (11 mins (!)) */
173 //UNUSED: #define WATCH_THRESHOLD 128
174 /* NB: set WATCH_THRESHOLD to ~60 when debugging to save time) */
175 //UNUSED: #define PANIC_THRESHOLD 1000 /* panic threshold (sec) */
178 * If we got |offset| > BIGOFF from a peer, cap next query interval
179 * for this peer by this many seconds:
181 #define BIGOFF STEP_THRESHOLD
182 #define BIGOFF_INTERVAL (1 << 7) /* 128 s */
184 #define FREQ_TOLERANCE 0.000015 /* frequency tolerance (15 PPM) */
185 #define BURSTPOLL 0 /* initial poll */
186 #define MINPOLL 5 /* minimum poll interval. std ntpd uses 6 (6: 64 sec) */
188 * If offset > discipline_jitter * POLLADJ_GATE, and poll interval is > 2^BIGPOLL,
189 * then it is decreased _at once_. (If <= 2^BIGPOLL, it will be decreased _eventually_).
191 #define BIGPOLL 9 /* 2^9 sec ~= 8.5 min */
192 #define MAXPOLL 12 /* maximum poll interval (12: 1.1h, 17: 36.4h). std ntpd uses 17 */
194 * Actively lower poll when we see such big offsets.
195 * With SLEW_THRESHOLD = 0.125, it means we try to sync more aggressively
196 * if offset increases over ~0.04 sec
198 //#define POLLDOWN_OFFSET (SLEW_THRESHOLD / 3)
199 #define MINDISP 0.01 /* minimum dispersion (sec) */
200 #define MAXDISP 16 /* maximum dispersion (sec) */
201 #define MAXSTRAT 16 /* maximum stratum (infinity metric) */
202 #define MAXDIST 1 /* distance threshold (sec) */
203 #define MIN_SELECTED 1 /* minimum intersection survivors */
204 #define MIN_CLUSTERED 3 /* minimum cluster survivors */
206 #define MAXDRIFT 0.000500 /* frequency drift we can correct (500 PPM) */
208 /* Poll-adjust threshold.
209 * When we see that offset is small enough compared to discipline jitter,
210 * we grow a counter: += MINPOLL. When counter goes over POLLADJ_LIMIT,
211 * we poll_exp++. If offset isn't small, counter -= poll_exp*2,
212 * and when it goes below -POLLADJ_LIMIT, we poll_exp--.
213 * (Bumped from 30 to 40 since otherwise I often see poll_exp going *2* steps down)
215 #define POLLADJ_LIMIT 40
216 /* If offset < discipline_jitter * POLLADJ_GATE, then we decide to increase
217 * poll interval (we think we can't improve timekeeping
218 * by staying at smaller poll).
220 #define POLLADJ_GATE 4
221 #define TIMECONST_HACK_GATE 2
222 /* Compromise Allan intercept (sec). doc uses 1500, std ntpd uses 512 */
226 /* FLL loop gain [why it depends on MAXPOLL??] */
227 #define FLL (MAXPOLL + 1)
228 /* Parameter averaging constant */
237 NTP_MSGSIZE_NOAUTH = 48,
238 NTP_MSGSIZE = (NTP_MSGSIZE_NOAUTH + 4 + NTP_DIGESTSIZE),
241 MODE_MASK = (7 << 0),
242 VERSION_MASK = (7 << 3),
246 /* Leap Second Codes (high order two bits of m_status) */
247 LI_NOWARNING = (0 << 6), /* no warning */
248 LI_PLUSSEC = (1 << 6), /* add a second (61 seconds) */
249 LI_MINUSSEC = (2 << 6), /* minus a second (59 seconds) */
250 LI_ALARM = (3 << 6), /* alarm condition */
253 MODE_RES0 = 0, /* reserved */
254 MODE_SYM_ACT = 1, /* symmetric active */
255 MODE_SYM_PAS = 2, /* symmetric passive */
256 MODE_CLIENT = 3, /* client */
257 MODE_SERVER = 4, /* server */
258 MODE_BROADCAST = 5, /* broadcast */
259 MODE_RES1 = 6, /* reserved for NTP control message */
260 MODE_RES2 = 7, /* reserved for private use */
263 //TODO: better base selection
264 #define OFFSET_1900_1970 2208988800UL /* 1970 - 1900 in seconds */
266 #define NUM_DATAPOINTS 8
279 uint8_t m_status; /* status of local clock and leap info */
281 uint8_t m_ppoll; /* poll value */
282 int8_t m_precision_exp;
283 s_fixedpt_t m_rootdelay;
284 s_fixedpt_t m_rootdisp;
286 l_fixedpt_t m_reftime;
287 l_fixedpt_t m_orgtime;
288 l_fixedpt_t m_rectime;
289 l_fixedpt_t m_xmttime;
291 uint8_t m_digest[NTP_DIGESTSIZE];
301 len_and_sockaddr *p_lsa;
305 uint32_t lastpkt_refid;
306 uint8_t lastpkt_status;
307 uint8_t lastpkt_stratum;
308 uint8_t reachable_bits;
310 /* when to send new query (if p_fd == -1)
311 * or when receive times out (if p_fd >= 0): */
312 double next_action_time;
315 /* p_raw_delay is set even by "high delay" packets */
316 /* lastpkt_delay isn't */
317 double lastpkt_recv_time;
318 double lastpkt_delay;
319 double lastpkt_rootdelay;
320 double lastpkt_rootdisp;
321 /* produced by filter algorithm: */
322 double filter_offset;
323 double filter_dispersion;
324 double filter_jitter;
325 datapoint_t filter_datapoint[NUM_DATAPOINTS];
326 /* last sent packet: */
332 #define USING_KERNEL_PLL_LOOP 1
333 #define USING_INITIAL_FREQ_ESTIMATION 0
340 /* Insert new options above this line. */
341 /* Non-compat options: */
345 OPT_l = (1 << 7) * ENABLE_FEATURE_NTPD_SERVER,
346 OPT_I = (1 << 8) * ENABLE_FEATURE_NTPD_SERVER,
347 /* We hijack some bits for other purposes */
353 /* total round trip delay to currently selected reference clock */
355 /* reference timestamp: time when the system clock was last set or corrected */
357 /* total dispersion to currently selected reference clock */
360 double last_script_run;
363 #if ENABLE_FEATURE_NTPD_SERVER
366 # define G_listen_fd (G.listen_fd)
368 # define G_listen_fd (-1)
372 /* refid: 32-bit code identifying the particular server or reference clock
373 * in stratum 0 packets this is a four-character ASCII string,
374 * called the kiss code, used for debugging and monitoring
375 * in stratum 1 packets this is a four-character ASCII string
376 * assigned to the reference clock by IANA. Example: "GPS "
377 * in stratum 2+ packets, it's IPv4 address or 4 first bytes
378 * of MD5 hash of IPv6
382 /* precision is defined as the larger of the resolution and time to
383 * read the clock, in log2 units. For instance, the precision of a
384 * mains-frequency clock incrementing at 60 Hz is 16 ms, even when the
385 * system clock hardware representation is to the nanosecond.
387 * Delays, jitters of various kinds are clamped down to precision.
389 * If precision_sec is too large, discipline_jitter gets clamped to it
390 * and if offset is smaller than discipline_jitter * POLLADJ_GATE, poll
391 * interval grows even though we really can benefit from staying at
392 * smaller one, collecting non-lagged datapoits and correcting offset.
393 * (Lagged datapoits exist when poll_exp is large but we still have
394 * systematic offset error - the time distance between datapoints
395 * is significant and older datapoints have smaller offsets.
396 * This makes our offset estimation a bit smaller than reality)
397 * Due to this effect, setting G_precision_sec close to
398 * STEP_THRESHOLD isn't such a good idea - offsets may grow
399 * too big and we will step. I observed it with -6.
401 * OTOH, setting precision_sec far too small would result in futile
402 * attempts to synchronize to an unachievable precision.
404 * -6 is 1/64 sec, -7 is 1/128 sec and so on.
405 * -8 is 1/256 ~= 0.003906 (worked well for me --vda)
406 * -9 is 1/512 ~= 0.001953 (let's try this for some time)
408 #define G_precision_exp -9
410 * G_precision_exp is used only for construction outgoing packets.
411 * It's ok to set G_precision_sec to a slightly different value
412 * (One which is "nicer looking" in logs).
413 * Exact value would be (1.0 / (1 << (- G_precision_exp))):
415 #define G_precision_sec 0.002
418 #define STATE_NSET 0 /* initial state, "nothing is set" */
419 //#define STATE_FSET 1 /* frequency set from file */
420 //#define STATE_SPIK 2 /* spike detected */
421 //#define STATE_FREQ 3 /* initial frequency */
422 #define STATE_SYNC 4 /* clock synchronized (normal operation) */
423 uint8_t discipline_state; // doc calls it c.state
424 uint8_t poll_exp; // s.poll
425 int polladj_count; // c.count
427 long kernel_freq_drift;
428 peer_t *last_update_peer;
429 double last_update_offset; // c.last
430 double last_update_recv_time; // s.t
431 double discipline_jitter; // c.jitter
432 /* Since we only compare it with ints, can simplify code
433 * by not making this variable floating point:
435 unsigned offset_to_jitter_ratio;
436 //double cluster_offset; // s.offset
437 //double cluster_jitter; // s.jitter
438 #if !USING_KERNEL_PLL_LOOP
439 double discipline_freq_drift; // c.freq
440 /* Maybe conditionally calculate wander? it's used only for logging */
441 double discipline_wander; // c.wander
444 #define G (*ptr_to_globals)
447 #define VERB1 if (MAX_VERBOSE && G.verbose)
448 #define VERB2 if (MAX_VERBOSE >= 2 && G.verbose >= 2)
449 #define VERB3 if (MAX_VERBOSE >= 3 && G.verbose >= 3)
450 #define VERB4 if (MAX_VERBOSE >= 4 && G.verbose >= 4)
451 #define VERB5 if (MAX_VERBOSE >= 5 && G.verbose >= 5)
452 #define VERB6 if (MAX_VERBOSE >= 6 && G.verbose >= 6)
455 static double LOG2D(int a)
458 return 1.0 / (1UL << -a);
461 static ALWAYS_INLINE double SQUARE(double x)
465 static ALWAYS_INLINE double MAXD(double a, double b)
471 static ALWAYS_INLINE double MIND(double a, double b)
477 static NOINLINE double my_SQRT(double X)
484 double Xhalf = X * 0.5;
486 /* Fast and good approximation to 1/sqrt(X), black magic */
488 /*v.i = 0x5f3759df - (v.i >> 1);*/
489 v.i = 0x5f375a86 - (v.i >> 1); /* - this constant is slightly better */
490 invsqrt = v.f; /* better than 0.2% accuracy */
492 /* Refining it using Newton's method: x1 = x0 - f(x0)/f'(x0)
493 * f(x) = 1/(x*x) - X (f==0 when x = 1/sqrt(X))
495 * f(x)/f'(x) = (X - 1/(x*x)) / (2/(x*x*x)) = X*x*x*x/2 - x/2
496 * x1 = x0 - (X*x0*x0*x0/2 - x0/2) = 1.5*x0 - X*x0*x0*x0/2 = x0*(1.5 - (X/2)*x0*x0)
498 invsqrt = invsqrt * (1.5 - Xhalf * invsqrt * invsqrt); /* ~0.05% accuracy */
499 /* invsqrt = invsqrt * (1.5 - Xhalf * invsqrt * invsqrt); 2nd iter: ~0.0001% accuracy */
500 /* With 4 iterations, more than half results will be exact,
501 * at 6th iterations result stabilizes with about 72% results exact.
502 * We are well satisfied with 0.05% accuracy.
505 return X * invsqrt; /* X * 1/sqrt(X) ~= sqrt(X) */
507 static ALWAYS_INLINE double SQRT(double X)
509 /* If this arch doesn't use IEEE 754 floats, fall back to using libm */
510 if (sizeof(float) != 4)
513 /* This avoids needing libm, saves about 0.5k on x86-32 */
521 gettimeofday(&tv, NULL); /* never fails */
522 G.cur_time = tv.tv_sec + (1.0e-6 * tv.tv_usec) + OFFSET_1900_1970;
527 d_to_tv(double d, struct timeval *tv)
529 tv->tv_sec = (long)d;
530 tv->tv_usec = (d - tv->tv_sec) * 1000000;
534 lfp_to_d(l_fixedpt_t lfp)
537 lfp.int_partl = ntohl(lfp.int_partl);
538 lfp.fractionl = ntohl(lfp.fractionl);
539 ret = (double)lfp.int_partl + ((double)lfp.fractionl / UINT_MAX);
543 sfp_to_d(s_fixedpt_t sfp)
546 sfp.int_parts = ntohs(sfp.int_parts);
547 sfp.fractions = ntohs(sfp.fractions);
548 ret = (double)sfp.int_parts + ((double)sfp.fractions / USHRT_MAX);
551 #if ENABLE_FEATURE_NTPD_SERVER
556 lfp.int_partl = (uint32_t)d;
557 lfp.fractionl = (uint32_t)((d - lfp.int_partl) * UINT_MAX);
558 lfp.int_partl = htonl(lfp.int_partl);
559 lfp.fractionl = htonl(lfp.fractionl);
566 sfp.int_parts = (uint16_t)d;
567 sfp.fractions = (uint16_t)((d - sfp.int_parts) * USHRT_MAX);
568 sfp.int_parts = htons(sfp.int_parts);
569 sfp.fractions = htons(sfp.fractions);
575 dispersion(const datapoint_t *dp)
577 return dp->d_dispersion + FREQ_TOLERANCE * (G.cur_time - dp->d_recv_time);
581 root_distance(peer_t *p)
583 /* The root synchronization distance is the maximum error due to
584 * all causes of the local clock relative to the primary server.
585 * It is defined as half the total delay plus total dispersion
588 return MAXD(MINDISP, p->lastpkt_rootdelay + p->lastpkt_delay) / 2
589 + p->lastpkt_rootdisp
590 + p->filter_dispersion
591 + FREQ_TOLERANCE * (G.cur_time - p->lastpkt_recv_time)
596 set_next(peer_t *p, unsigned t)
598 p->next_action_time = G.cur_time + t;
602 * Peer clock filter and its helpers
605 filter_datapoints(peer_t *p)
612 /* Simulations have shown that use of *averaged* offset for p->filter_offset
613 * is in fact worse than simply using last received one: with large poll intervals
614 * (>= 2048) averaging code uses offset values which are outdated by hours,
615 * and time/frequency correction goes totally wrong when fed essentially bogus offsets.
618 double minoff, maxoff, w;
619 double x = x; /* for compiler */
620 double oldest_off = oldest_off;
621 double oldest_age = oldest_age;
622 double newest_off = newest_off;
623 double newest_age = newest_age;
625 fdp = p->filter_datapoint;
627 minoff = maxoff = fdp[0].d_offset;
628 for (i = 1; i < NUM_DATAPOINTS; i++) {
629 if (minoff > fdp[i].d_offset)
630 minoff = fdp[i].d_offset;
631 if (maxoff < fdp[i].d_offset)
632 maxoff = fdp[i].d_offset;
635 idx = p->datapoint_idx; /* most recent datapoint's index */
637 * Drop two outliers and take weighted average of the rest:
638 * most_recent/2 + older1/4 + older2/8 ... + older5/32 + older6/32
639 * we use older6/32, not older6/64 since sum of weights should be 1:
640 * 1/2 + 1/4 + 1/8 + 1/16 + 1/32 + 1/32 = 1
646 * filter_dispersion = \ -------------
653 for (i = 0; i < NUM_DATAPOINTS; i++) {
655 bb_error_msg("datapoint[%d]: off:%f disp:%f(%f) age:%f%s",
658 fdp[idx].d_dispersion, dispersion(&fdp[idx]),
659 G.cur_time - fdp[idx].d_recv_time,
660 (minoff == fdp[idx].d_offset || maxoff == fdp[idx].d_offset)
661 ? " (outlier by offset)" : ""
665 sum += dispersion(&fdp[idx]) / (2 << i);
667 if (minoff == fdp[idx].d_offset) {
668 minoff -= 1; /* so that we don't match it ever again */
670 if (maxoff == fdp[idx].d_offset) {
673 oldest_off = fdp[idx].d_offset;
674 oldest_age = G.cur_time - fdp[idx].d_recv_time;
677 newest_off = oldest_off;
678 newest_age = oldest_age;
685 idx = (idx - 1) & (NUM_DATAPOINTS - 1);
687 p->filter_dispersion = sum;
688 wavg += x; /* add another older6/64 to form older6/32 */
689 /* Fix systematic underestimation with large poll intervals.
690 * Imagine that we still have a bit of uncorrected drift,
691 * and poll interval is big (say, 100 sec). Offsets form a progression:
692 * 0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 - 0.7 is most recent.
693 * The algorithm above drops 0.0 and 0.7 as outliers,
694 * and then we have this estimation, ~25% off from 0.7:
695 * 0.1/32 + 0.2/32 + 0.3/16 + 0.4/8 + 0.5/4 + 0.6/2 = 0.503125
697 x = oldest_age - newest_age;
699 x = newest_age / x; /* in above example, 100 / (600 - 100) */
700 if (x < 1) { /* paranoia check */
701 x = (newest_off - oldest_off) * x; /* 0.5 * 100/500 = 0.1 */
705 p->filter_offset = wavg;
709 fdp = p->filter_datapoint;
710 idx = p->datapoint_idx; /* most recent datapoint's index */
712 /* filter_offset: simply use the most recent value */
713 p->filter_offset = fdp[idx].d_offset;
717 * filter_dispersion = \ -------------
724 for (i = 0; i < NUM_DATAPOINTS; i++) {
725 sum += dispersion(&fdp[idx]) / (2 << i);
726 wavg += fdp[idx].d_offset;
727 idx = (idx - 1) & (NUM_DATAPOINTS - 1);
729 wavg /= NUM_DATAPOINTS;
730 p->filter_dispersion = sum;
733 /* +----- -----+ ^ 1/2
737 * filter_jitter = | --- * / (avg-offset_j) |
741 * where n is the number of valid datapoints in the filter (n > 1);
742 * if filter_jitter < precision then filter_jitter = precision
745 for (i = 0; i < NUM_DATAPOINTS; i++) {
746 sum += SQUARE(wavg - fdp[i].d_offset);
748 sum = SQRT(sum / NUM_DATAPOINTS);
749 p->filter_jitter = sum > G_precision_sec ? sum : G_precision_sec;
751 VERB4 bb_error_msg("filter offset:%+f disp:%f jitter:%f",
753 p->filter_dispersion,
758 reset_peer_stats(peer_t *p, double offset)
761 bool small_ofs = fabs(offset) < STEP_THRESHOLD;
763 /* Used to set p->filter_datapoint[i].d_dispersion = MAXDISP
764 * and clear reachable bits, but this proved to be too aggressive:
765 * after step (tested with suspending laptop for ~30 secs),
766 * this caused all previous data to be considered invalid,
767 * making us needing to collect full ~8 datapoints per peer
768 * after step in order to start trusting them.
769 * In turn, this was making poll interval decrease even after
770 * step was done. (Poll interval decreases already before step
771 * in this scenario, because we see large offsets and end up with
772 * no good peer to select).
775 for (i = 0; i < NUM_DATAPOINTS; i++) {
777 p->filter_datapoint[i].d_recv_time += offset;
778 if (p->filter_datapoint[i].d_offset != 0) {
779 p->filter_datapoint[i].d_offset -= offset;
780 //bb_error_msg("p->filter_datapoint[%d].d_offset %f -> %f",
782 // p->filter_datapoint[i].d_offset + offset,
783 // p->filter_datapoint[i].d_offset);
786 p->filter_datapoint[i].d_recv_time = G.cur_time;
787 p->filter_datapoint[i].d_offset = 0;
788 /*p->filter_datapoint[i].d_dispersion = MAXDISP;*/
792 p->lastpkt_recv_time += offset;
794 /*p->reachable_bits = 0;*/
795 p->lastpkt_recv_time = G.cur_time;
797 filter_datapoints(p); /* recalc p->filter_xxx */
798 VERB6 bb_error_msg("%s->lastpkt_recv_time=%f", p->p_dotted, p->lastpkt_recv_time);
801 static len_and_sockaddr*
802 resolve_peer_hostname(peer_t *p)
804 len_and_sockaddr *lsa = host2sockaddr(p->p_hostname, 123);
809 p->p_dotted = xmalloc_sockaddr2dotted_noport(&lsa->u.sa);
810 VERB1 if (strcmp(p->p_hostname, p->p_dotted) != 0)
811 bb_error_msg("'%s' is %s", p->p_hostname, p->p_dotted);
815 p->dns_errors = ((p->dns_errors << 1) | 1) & DNS_ERRORS_CAP;
820 add_peers(const char *s)
825 p = xzalloc(sizeof(*p) + strlen(s));
826 strcpy(p->p_hostname, s);
828 p->p_xmt_msg.m_status = MODE_CLIENT | (NTP_VERSION << 3);
829 p->next_action_time = G.cur_time; /* = set_next(p, 0); */
830 reset_peer_stats(p, STEP_THRESHOLD);
832 /* Names like N.<country2chars>.pool.ntp.org are randomly resolved
833 * to a pool of machines. Sometimes different N's resolve to the same IP.
834 * It is not useful to have two peers with same IP. We skip duplicates.
836 if (resolve_peer_hostname(p)) {
837 for (item = G.ntp_peers; item != NULL; item = item->link) {
838 peer_t *pp = (peer_t *) item->data;
839 if (pp->p_dotted && strcmp(p->p_dotted, pp->p_dotted) == 0) {
840 bb_error_msg("duplicate peer %s (%s)", s, p->p_dotted);
849 llist_add_to(&G.ntp_peers, p);
855 const struct sockaddr *from, const struct sockaddr *to, socklen_t addrlen,
856 msg_t *msg, ssize_t len)
862 ret = sendto(fd, msg, len, MSG_DONTWAIT, to, addrlen);
864 ret = send_to_from(fd, msg, len, MSG_DONTWAIT, to, from, addrlen);
867 bb_perror_msg("send failed");
874 send_query_to_peer(peer_t *p)
879 /* Why do we need to bind()?
880 * See what happens when we don't bind:
882 * socket(PF_INET, SOCK_DGRAM, IPPROTO_IP) = 3
883 * setsockopt(3, SOL_IP, IP_TOS, [16], 4) = 0
884 * gettimeofday({1259071266, 327885}, NULL) = 0
885 * sendto(3, "xxx", 48, MSG_DONTWAIT, {sa_family=AF_INET, sin_port=htons(123), sin_addr=inet_addr("10.34.32.125")}, 16) = 48
886 * ^^^ we sent it from some source port picked by kernel.
887 * time(NULL) = 1259071266
888 * write(2, "ntpd: entering poll 15 secs\n", 28) = 28
889 * poll([{fd=3, events=POLLIN}], 1, 15000) = 1 ([{fd=3, revents=POLLIN}])
890 * recv(3, "yyy", 68, MSG_DONTWAIT) = 48
891 * ^^^ this recv will receive packets to any local port!
893 * Uncomment this and use strace to see it in action:
895 #define PROBE_LOCAL_ADDR /* { len_and_sockaddr lsa; lsa.len = LSA_SIZEOF_SA; getsockname(p->query.fd, &lsa.u.sa, &lsa.len); } */
899 len_and_sockaddr *local_lsa;
901 family = p->p_lsa->u.sa.sa_family;
902 p->p_fd = fd = xsocket_type(&local_lsa, family, SOCK_DGRAM);
903 /* local_lsa has "null" address and port 0 now.
904 * bind() ensures we have a *particular port* selected by kernel
905 * and remembered in p->p_fd, thus later recv(p->p_fd)
906 * receives only packets sent to this port.
909 xbind(fd, &local_lsa->u.sa, local_lsa->len);
911 #if ENABLE_FEATURE_IPV6
912 if (family == AF_INET)
914 setsockopt_int(fd, IPPROTO_IP, IP_TOS, IPTOS_DSCP_AF21);
918 /* Emit message _before_ attempted send. Think of a very short
919 * roundtrip networks: we need to go back to recv loop ASAP,
920 * to reduce delay. Printing messages after send works against that.
922 VERB1 bb_error_msg("sending query to %s", p->p_dotted);
925 * Send out a random 64-bit number as our transmit time. The NTP
926 * server will copy said number into the originate field on the
927 * response that it sends us. This is totally legal per the SNTP spec.
929 * The impact of this is two fold: we no longer send out the current
930 * system time for the world to see (which may aid an attacker), and
931 * it gives us a (not very secure) way of knowing that we're not
932 * getting spoofed by an attacker that can't capture our traffic
933 * but can spoof packets from the NTP server we're communicating with.
935 * Save the real transmit timestamp locally.
937 p->p_xmt_msg.m_xmttime.int_partl = rand();
938 p->p_xmt_msg.m_xmttime.fractionl = rand();
939 p->p_xmttime = gettime1900d();
941 /* Were doing it only if sendto worked, but
942 * loss of sync detection needs reachable_bits updated
943 * even if sending fails *locally*:
944 * "network is unreachable" because cable was pulled?
945 * We still need to declare "unsync" if this condition persists.
947 p->reachable_bits <<= 1;
949 if (do_sendto(p->p_fd, /*from:*/ NULL, /*to:*/ &p->p_lsa->u.sa, /*addrlen:*/ p->p_lsa->len,
950 &p->p_xmt_msg, NTP_MSGSIZE_NOAUTH) == -1
955 * We know that we sent nothing.
956 * We can retry *soon* without fearing
957 * that we are flooding the peer.
959 set_next(p, RETRY_INTERVAL);
963 set_next(p, RESPONSE_INTERVAL);
967 /* Note that there is no provision to prevent several run_scripts
968 * to be started in quick succession. In fact, it happens rather often
969 * if initial syncronization results in a step.
970 * You will see "step" and then "stratum" script runs, sometimes
971 * as close as only 0.002 seconds apart.
972 * Script should be ready to deal with this.
974 static void run_script(const char *action, double offset)
977 char *env1, *env2, *env3, *env4;
979 G.last_script_run = G.cur_time;
984 argv[0] = (char*) G.script_name;
985 argv[1] = (char*) action;
988 VERB1 bb_error_msg("executing '%s %s'", G.script_name, action);
990 env1 = xasprintf("%s=%u", "stratum", G.stratum);
992 env2 = xasprintf("%s=%ld", "freq_drift_ppm", G.kernel_freq_drift);
994 env3 = xasprintf("%s=%u", "poll_interval", 1 << G.poll_exp);
996 env4 = xasprintf("%s=%f", "offset", offset);
998 /* Other items of potential interest: selected peer,
999 * rootdelay, reftime, rootdisp, refid, ntp_status,
1000 * last_update_offset, last_update_recv_time, discipline_jitter,
1001 * how many peers have reachable_bits = 0?
1004 /* Don't want to wait: it may run hwclock --systohc, and that
1005 * may take some time (seconds): */
1006 /*spawn_and_wait(argv);*/
1009 unsetenv("stratum");
1010 unsetenv("freq_drift_ppm");
1011 unsetenv("poll_interval");
1019 static NOINLINE void
1020 step_time(double offset)
1024 struct timeval tvc, tvn;
1025 char buf[sizeof("yyyy-mm-dd hh:mm:ss") + /*paranoia:*/ 4];
1028 gettimeofday(&tvc, NULL); /* never fails */
1029 dtime = tvc.tv_sec + (1.0e-6 * tvc.tv_usec) + offset;
1030 d_to_tv(dtime, &tvn);
1031 if (settimeofday(&tvn, NULL) == -1)
1032 bb_perror_msg_and_die("settimeofday");
1036 strftime_YYYYMMDDHHMMSS(buf, sizeof(buf), &tval);
1037 bb_error_msg("current time is %s.%06u", buf, (unsigned)tvc.tv_usec);
1040 strftime_YYYYMMDDHHMMSS(buf, sizeof(buf), &tval);
1041 bb_error_msg("setting time to %s.%06u (offset %+fs)", buf, (unsigned)tvn.tv_usec, offset);
1042 //maybe? G.FREQHOLD_cnt = 0;
1044 /* Correct various fields which contain time-relative values: */
1047 G.cur_time += offset;
1048 G.last_update_recv_time += offset;
1049 G.last_script_run += offset;
1051 /* p->lastpkt_recv_time, p->next_action_time and such: */
1052 for (item = G.ntp_peers; item != NULL; item = item->link) {
1053 peer_t *pp = (peer_t *) item->data;
1054 reset_peer_stats(pp, offset);
1055 //bb_error_msg("offset:%+f pp->next_action_time:%f -> %f",
1056 // offset, pp->next_action_time, pp->next_action_time + offset);
1057 pp->next_action_time += offset;
1058 if (pp->p_fd >= 0) {
1059 /* We wait for reply from this peer too.
1060 * But due to step we are doing, reply's data is no longer
1061 * useful (in fact, it'll be bogus). Stop waiting for it.
1065 set_next(pp, RETRY_INTERVAL);
1070 static void clamp_pollexp_and_set_MAXSTRAT(void)
1072 if (G.poll_exp < MINPOLL)
1073 G.poll_exp = MINPOLL;
1074 if (G.poll_exp > BIGPOLL)
1075 G.poll_exp = BIGPOLL;
1076 G.polladj_count = 0;
1077 G.stratum = MAXSTRAT;
1082 * Selection and clustering, and their helpers
1088 double opt_rd; /* optimization */
1091 compare_point_edge(const void *aa, const void *bb)
1093 const point_t *a = aa;
1094 const point_t *b = bb;
1095 if (a->edge < b->edge) {
1098 return (a->edge > b->edge);
1105 compare_survivor_metric(const void *aa, const void *bb)
1107 const survivor_t *a = aa;
1108 const survivor_t *b = bb;
1109 if (a->metric < b->metric) {
1112 return (a->metric > b->metric);
1115 fit(peer_t *p, double rd)
1117 if ((p->reachable_bits & (p->reachable_bits-1)) == 0) {
1118 /* One or zero bits in reachable_bits */
1119 VERB4 bb_error_msg("peer %s unfit for selection: unreachable", p->p_dotted);
1122 #if 0 /* we filter out such packets earlier */
1123 if ((p->lastpkt_status & LI_ALARM) == LI_ALARM
1124 || p->lastpkt_stratum >= MAXSTRAT
1126 VERB4 bb_error_msg("peer %s unfit for selection: bad status/stratum", p->p_dotted);
1130 /* rd is root_distance(p) */
1131 if (rd > MAXDIST + FREQ_TOLERANCE * (1 << G.poll_exp)) {
1132 VERB4 bb_error_msg("peer %s unfit for selection: root distance too high", p->p_dotted);
1136 // /* Do we have a loop? */
1137 // if (p->refid == p->dstaddr || p->refid == s.refid)
1142 select_and_cluster(void)
1147 int size = 3 * G.peer_cnt;
1148 /* for selection algorithm */
1149 point_t point[size];
1150 unsigned num_points, num_candidates;
1152 unsigned num_falsetickers;
1153 /* for cluster algorithm */
1154 survivor_t survivor[size];
1155 unsigned num_survivors;
1161 while (item != NULL) {
1164 p = (peer_t *) item->data;
1165 rd = root_distance(p);
1166 offset = p->filter_offset;
1172 VERB5 bb_error_msg("interval: [%f %f %f] %s",
1178 point[num_points].p = p;
1179 point[num_points].type = -1;
1180 point[num_points].edge = offset - rd;
1181 point[num_points].opt_rd = rd;
1183 point[num_points].p = p;
1184 point[num_points].type = 0;
1185 point[num_points].edge = offset;
1186 point[num_points].opt_rd = rd;
1188 point[num_points].p = p;
1189 point[num_points].type = 1;
1190 point[num_points].edge = offset + rd;
1191 point[num_points].opt_rd = rd;
1195 num_candidates = num_points / 3;
1196 if (num_candidates == 0) {
1197 VERB3 bb_error_msg("no valid datapoints%s", ", no peer selected");
1200 //TODO: sorting does not seem to be done in reference code
1201 qsort(point, num_points, sizeof(point[0]), compare_point_edge);
1203 /* Start with the assumption that there are no falsetickers.
1204 * Attempt to find a nonempty intersection interval containing
1205 * the midpoints of all truechimers.
1206 * If a nonempty interval cannot be found, increase the number
1207 * of assumed falsetickers by one and try again.
1208 * If a nonempty interval is found and the number of falsetickers
1209 * is less than the number of truechimers, a majority has been found
1210 * and the midpoint of each truechimer represents
1211 * the candidates available to the cluster algorithm.
1213 num_falsetickers = 0;
1216 unsigned num_midpoints = 0;
1221 for (i = 0; i < num_points; i++) {
1223 * if (point[i].type == -1) c++;
1224 * if (point[i].type == 1) c--;
1225 * and it's simpler to do it this way:
1228 if (c >= num_candidates - num_falsetickers) {
1229 /* If it was c++ and it got big enough... */
1230 low = point[i].edge;
1233 if (point[i].type == 0)
1237 for (i = num_points-1; i >= 0; i--) {
1239 if (c >= num_candidates - num_falsetickers) {
1240 high = point[i].edge;
1243 if (point[i].type == 0)
1246 /* If the number of midpoints is greater than the number
1247 * of allowed falsetickers, the intersection contains at
1248 * least one truechimer with no midpoint - bad.
1249 * Also, interval should be nonempty.
1251 if (num_midpoints <= num_falsetickers && low < high)
1254 if (num_falsetickers * 2 >= num_candidates) {
1255 VERB3 bb_error_msg("falsetickers:%d, candidates:%d%s",
1256 num_falsetickers, num_candidates,
1257 ", no peer selected");
1261 VERB4 bb_error_msg("selected interval: [%f, %f]; candidates:%d falsetickers:%d",
1262 low, high, num_candidates, num_falsetickers);
1266 /* Construct a list of survivors (p, metric)
1267 * from the chime list, where metric is dominated
1268 * first by stratum and then by root distance.
1269 * All other things being equal, this is the order of preference.
1272 for (i = 0; i < num_points; i++) {
1273 if (point[i].edge < low || point[i].edge > high)
1276 survivor[num_survivors].p = p;
1277 /* x.opt_rd == root_distance(p); */
1278 survivor[num_survivors].metric = MAXDIST * p->lastpkt_stratum + point[i].opt_rd;
1279 VERB5 bb_error_msg("survivor[%d] metric:%f peer:%s",
1280 num_survivors, survivor[num_survivors].metric, p->p_dotted);
1283 /* There must be at least MIN_SELECTED survivors to satisfy the
1284 * correctness assertions. Ordinarily, the Byzantine criteria
1285 * require four survivors, but for the demonstration here, one
1288 if (num_survivors < MIN_SELECTED) {
1289 VERB3 bb_error_msg("survivors:%d%s",
1291 ", no peer selected");
1295 //looks like this is ONLY used by the fact that later we pick survivor[0].
1296 //we can avoid sorting then, just find the minimum once!
1297 qsort(survivor, num_survivors, sizeof(survivor[0]), compare_survivor_metric);
1299 /* For each association p in turn, calculate the selection
1300 * jitter p->sjitter as the square root of the sum of squares
1301 * (p->offset - q->offset) over all q associations. The idea is
1302 * to repeatedly discard the survivor with maximum selection
1303 * jitter until a termination condition is met.
1306 unsigned max_idx = max_idx;
1307 double max_selection_jitter = max_selection_jitter;
1308 double min_jitter = min_jitter;
1310 if (num_survivors <= MIN_CLUSTERED) {
1311 VERB4 bb_error_msg("num_survivors %d <= %d, not discarding more",
1312 num_survivors, MIN_CLUSTERED);
1316 /* To make sure a few survivors are left
1317 * for the clustering algorithm to chew on,
1318 * we stop if the number of survivors
1319 * is less than or equal to MIN_CLUSTERED (3).
1321 for (i = 0; i < num_survivors; i++) {
1322 double selection_jitter_sq;
1325 if (i == 0 || p->filter_jitter < min_jitter)
1326 min_jitter = p->filter_jitter;
1328 selection_jitter_sq = 0;
1329 for (j = 0; j < num_survivors; j++) {
1330 peer_t *q = survivor[j].p;
1331 selection_jitter_sq += SQUARE(p->filter_offset - q->filter_offset);
1333 if (i == 0 || selection_jitter_sq > max_selection_jitter) {
1334 max_selection_jitter = selection_jitter_sq;
1337 VERB6 bb_error_msg("survivor %d selection_jitter^2:%f",
1338 i, selection_jitter_sq);
1340 max_selection_jitter = SQRT(max_selection_jitter / num_survivors);
1341 VERB5 bb_error_msg("max_selection_jitter (at %d):%f min_jitter:%f",
1342 max_idx, max_selection_jitter, min_jitter);
1344 /* If the maximum selection jitter is less than the
1345 * minimum peer jitter, then tossing out more survivors
1346 * will not lower the minimum peer jitter, so we might
1349 if (max_selection_jitter < min_jitter) {
1350 VERB4 bb_error_msg("max_selection_jitter:%f < min_jitter:%f, num_survivors:%d, not discarding more",
1351 max_selection_jitter, min_jitter, num_survivors);
1355 /* Delete survivor[max_idx] from the list
1356 * and go around again.
1358 VERB6 bb_error_msg("dropping survivor %d", max_idx);
1360 while (max_idx < num_survivors) {
1361 survivor[max_idx] = survivor[max_idx + 1];
1367 /* Combine the offsets of the clustering algorithm survivors
1368 * using a weighted average with weight determined by the root
1369 * distance. Compute the selection jitter as the weighted RMS
1370 * difference between the first survivor and the remaining
1371 * survivors. In some cases the inherent clock jitter can be
1372 * reduced by not using this algorithm, especially when frequent
1373 * clockhopping is involved. bbox: thus we don't do it.
1377 for (i = 0; i < num_survivors; i++) {
1379 x = root_distance(p);
1381 z += p->filter_offset / x;
1382 w += SQUARE(p->filter_offset - survivor[0].p->filter_offset) / x;
1384 //G.cluster_offset = z / y;
1385 //G.cluster_jitter = SQRT(w / y);
1388 /* Pick the best clock. If the old system peer is on the list
1389 * and at the same stratum as the first survivor on the list,
1390 * then don't do a clock hop. Otherwise, select the first
1391 * survivor on the list as the new system peer.
1394 if (G.last_update_peer
1395 && G.last_update_peer->lastpkt_stratum <= p->lastpkt_stratum
1397 /* Starting from 1 is ok here */
1398 for (i = 1; i < num_survivors; i++) {
1399 if (G.last_update_peer == survivor[i].p) {
1400 VERB5 bb_error_msg("keeping old synced peer");
1401 p = G.last_update_peer;
1406 G.last_update_peer = p;
1408 VERB4 bb_error_msg("selected peer %s filter_offset:%+f age:%f",
1411 G.cur_time - p->lastpkt_recv_time
1418 * Local clock discipline and its helpers
1421 set_new_values(int disc_state, double offset, double recv_time)
1423 /* Enter new state and set state variables. Note we use the time
1424 * of the last clock filter sample, which must be earlier than
1427 VERB4 bb_error_msg("disc_state=%d last update offset=%f recv_time=%f",
1428 disc_state, offset, recv_time);
1429 G.discipline_state = disc_state;
1430 G.last_update_offset = offset;
1431 G.last_update_recv_time = recv_time;
1433 /* Return: -1: decrease poll interval, 0: leave as is, 1: increase */
1435 update_local_clock(peer_t *p)
1439 /* Note: can use G.cluster_offset instead: */
1440 double offset = p->filter_offset;
1441 double recv_time = p->lastpkt_recv_time;
1443 #if !USING_KERNEL_PLL_LOOP
1446 #if !USING_KERNEL_PLL_LOOP || USING_INITIAL_FREQ_ESTIMATION
1447 double since_last_update;
1449 double etemp, dtemp;
1451 abs_offset = fabs(offset);
1454 /* If needed, -S script can do it by looking at $offset
1455 * env var and killing parent */
1456 /* If the offset is too large, give up and go home */
1457 if (abs_offset > PANIC_THRESHOLD) {
1458 bb_error_msg_and_die("offset %f far too big, exiting", offset);
1462 /* If this is an old update, for instance as the result
1463 * of a system peer change, avoid it. We never use
1464 * an old sample or the same sample twice.
1466 if (recv_time <= G.last_update_recv_time) {
1467 VERB3 bb_error_msg("update from %s: same or older datapoint, not using it",
1469 return 0; /* "leave poll interval as is" */
1472 /* Clock state machine transition function. This is where the
1473 * action is and defines how the system reacts to large time
1474 * and frequency errors.
1476 #if !USING_KERNEL_PLL_LOOP || USING_INITIAL_FREQ_ESTIMATION
1477 since_last_update = recv_time - G.reftime;
1479 #if !USING_KERNEL_PLL_LOOP
1482 #if USING_INITIAL_FREQ_ESTIMATION
1483 if (G.discipline_state == STATE_FREQ) {
1484 /* Ignore updates until the stepout threshold */
1485 if (since_last_update < WATCH_THRESHOLD) {
1486 VERB4 bb_error_msg("measuring drift, datapoint ignored, %f sec remains",
1487 WATCH_THRESHOLD - since_last_update);
1488 return 0; /* "leave poll interval as is" */
1490 # if !USING_KERNEL_PLL_LOOP
1491 freq_drift = (offset - G.last_update_offset) / since_last_update;
1496 /* There are two main regimes: when the
1497 * offset exceeds the step threshold and when it does not.
1499 if (abs_offset > STEP_THRESHOLD) {
1503 // This "spike state" seems to be useless, peer selection already drops
1504 // occassional "bad" datapoints. If we are here, there were _many_
1505 // large offsets. When a few first large offsets are seen,
1506 // we end up in "no valid datapoints, no peer selected" state.
1507 // Only when enough of them are seen (which means it's not a fluke),
1508 // we end up here. Looks like _our_ clock is off.
1509 switch (G.discipline_state) {
1511 /* The first outlyer: ignore it, switch to SPIK state */
1512 VERB3 bb_error_msg("update from %s: offset:%+f, spike%s",
1513 p->p_dotted, offset,
1515 G.discipline_state = STATE_SPIK;
1516 return -1; /* "decrease poll interval" */
1519 /* Ignore succeeding outlyers until either an inlyer
1520 * is found or the stepout threshold is exceeded.
1522 remains = WATCH_THRESHOLD - since_last_update;
1524 VERB3 bb_error_msg("update from %s: offset:%+f, spike%s",
1525 p->p_dotted, offset,
1526 ", datapoint ignored");
1527 return -1; /* "decrease poll interval" */
1529 /* fall through: we need to step */
1533 /* Step the time and clamp down the poll interval.
1535 * In NSET state an initial frequency correction is
1536 * not available, usually because the frequency file has
1537 * not yet been written. Since the time is outside the
1538 * capture range, the clock is stepped. The frequency
1539 * will be set directly following the stepout interval.
1541 * In FSET state the initial frequency has been set
1542 * from the frequency file. Since the time is outside
1543 * the capture range, the clock is stepped immediately,
1544 * rather than after the stepout interval. Guys get
1545 * nervous if it takes 17 minutes to set the clock for
1548 * In SPIK state the stepout threshold has expired and
1549 * the phase is still above the step threshold. Note
1550 * that a single spike greater than the step threshold
1551 * is always suppressed, even at the longer poll
1554 VERB4 bb_error_msg("stepping time by %+f; poll_exp=MINPOLL", offset);
1556 if (option_mask32 & OPT_q) {
1557 /* We were only asked to set time once. Done. */
1561 clamp_pollexp_and_set_MAXSTRAT();
1563 run_script("step", offset);
1565 recv_time += offset;
1567 #if USING_INITIAL_FREQ_ESTIMATION
1568 if (G.discipline_state == STATE_NSET) {
1569 set_new_values(STATE_FREQ, /*offset:*/ 0, recv_time);
1570 return 1; /* "ok to increase poll interval" */
1573 abs_offset = offset = 0;
1574 set_new_values(STATE_SYNC, offset, recv_time);
1575 } else { /* abs_offset <= STEP_THRESHOLD */
1577 /* The ratio is calculated before jitter is updated to make
1578 * poll adjust code more sensitive to large offsets.
1580 G.offset_to_jitter_ratio = abs_offset / G.discipline_jitter;
1582 /* Compute the clock jitter as the RMS of exponentially
1583 * weighted offset differences. Used by the poll adjust code.
1585 etemp = SQUARE(G.discipline_jitter);
1586 dtemp = SQUARE(offset - G.last_update_offset);
1587 G.discipline_jitter = SQRT(etemp + (dtemp - etemp) / AVG);
1588 if (G.discipline_jitter < G_precision_sec)
1589 G.discipline_jitter = G_precision_sec;
1591 switch (G.discipline_state) {
1593 if (option_mask32 & OPT_q) {
1594 /* We were only asked to set time once.
1595 * The clock is precise enough, no need to step.
1599 #if USING_INITIAL_FREQ_ESTIMATION
1600 /* This is the first update received and the frequency
1601 * has not been initialized. The first thing to do
1602 * is directly measure the oscillator frequency.
1604 set_new_values(STATE_FREQ, offset, recv_time);
1606 set_new_values(STATE_SYNC, offset, recv_time);
1608 VERB4 bb_error_msg("transitioning to FREQ, datapoint ignored");
1609 return 0; /* "leave poll interval as is" */
1611 #if 0 /* this is dead code for now */
1613 /* This is the first update and the frequency
1614 * has been initialized. Adjust the phase, but
1615 * don't adjust the frequency until the next update.
1617 set_new_values(STATE_SYNC, offset, recv_time);
1618 /* freq_drift remains 0 */
1622 #if USING_INITIAL_FREQ_ESTIMATION
1624 /* since_last_update >= WATCH_THRESHOLD, we waited enough.
1625 * Correct the phase and frequency and switch to SYNC state.
1626 * freq_drift was already estimated (see code above)
1628 set_new_values(STATE_SYNC, offset, recv_time);
1633 #if !USING_KERNEL_PLL_LOOP
1634 /* Compute freq_drift due to PLL and FLL contributions.
1636 * The FLL and PLL frequency gain constants
1637 * depend on the poll interval and Allan
1638 * intercept. The FLL is not used below one-half
1639 * the Allan intercept. Above that the loop gain
1640 * increases in steps to 1 / AVG.
1642 if ((1 << G.poll_exp) > ALLAN / 2) {
1643 etemp = FLL - G.poll_exp;
1646 freq_drift += (offset - G.last_update_offset) / (MAXD(since_last_update, ALLAN) * etemp);
1648 /* For the PLL the integration interval
1649 * (numerator) is the minimum of the update
1650 * interval and poll interval. This allows
1651 * oversampling, but not undersampling.
1653 etemp = MIND(since_last_update, (1 << G.poll_exp));
1654 dtemp = (4 * PLL) << G.poll_exp;
1655 freq_drift += offset * etemp / SQUARE(dtemp);
1657 set_new_values(STATE_SYNC, offset, recv_time);
1660 if (G.stratum != p->lastpkt_stratum + 1) {
1661 G.stratum = p->lastpkt_stratum + 1;
1662 run_script("stratum", offset);
1666 G.reftime = G.cur_time;
1667 G.ntp_status = p->lastpkt_status;
1668 G.refid = p->lastpkt_refid;
1669 G.rootdelay = p->lastpkt_rootdelay + p->lastpkt_delay;
1670 dtemp = p->filter_jitter; // SQRT(SQUARE(p->filter_jitter) + SQUARE(G.cluster_jitter));
1671 dtemp += MAXD(p->filter_dispersion + FREQ_TOLERANCE * (G.cur_time - p->lastpkt_recv_time) + abs_offset, MINDISP);
1672 G.rootdisp = p->lastpkt_rootdisp + dtemp;
1673 VERB4 bb_error_msg("updating leap/refid/reftime/rootdisp from peer %s", p->p_dotted);
1675 /* We are in STATE_SYNC now, but did not do adjtimex yet.
1676 * (Any other state does not reach this, they all return earlier)
1677 * By this time, freq_drift and offset are set
1678 * to values suitable for adjtimex.
1680 #if !USING_KERNEL_PLL_LOOP
1681 /* Calculate the new frequency drift and frequency stability (wander).
1682 * Compute the clock wander as the RMS of exponentially weighted
1683 * frequency differences. This is not used directly, but can,
1684 * along with the jitter, be a highly useful monitoring and
1687 dtemp = G.discipline_freq_drift + freq_drift;
1688 G.discipline_freq_drift = MAXD(MIND(MAXDRIFT, dtemp), -MAXDRIFT);
1689 etemp = SQUARE(G.discipline_wander);
1690 dtemp = SQUARE(dtemp);
1691 G.discipline_wander = SQRT(etemp + (dtemp - etemp) / AVG);
1693 VERB4 bb_error_msg("discipline freq_drift=%.9f(int:%ld corr:%e) wander=%f",
1694 G.discipline_freq_drift,
1695 (long)(G.discipline_freq_drift * 65536e6),
1697 G.discipline_wander);
1700 memset(&tmx, 0, sizeof(tmx));
1701 if (adjtimex(&tmx) < 0)
1702 bb_perror_msg_and_die("adjtimex");
1703 bb_error_msg("p adjtimex freq:%ld offset:%+ld status:0x%x tc:%ld",
1704 tmx.freq, tmx.offset, tmx.status, tmx.constant);
1707 memset(&tmx, 0, sizeof(tmx));
1709 //doesn't work, offset remains 0 (!) in kernel:
1710 //ntpd: set adjtimex freq:1786097 tmx.offset:77487
1711 //ntpd: prev adjtimex freq:1786097 tmx.offset:0
1712 //ntpd: cur adjtimex freq:1786097 tmx.offset:0
1713 tmx.modes = ADJ_FREQUENCY | ADJ_OFFSET;
1714 /* 65536 is one ppm */
1715 tmx.freq = G.discipline_freq_drift * 65536e6;
1717 tmx.modes = ADJ_OFFSET | ADJ_STATUS | ADJ_TIMECONST;// | ADJ_MAXERROR | ADJ_ESTERROR;
1719 tmx.offset = (long)(offset * 1000000); /* usec */
1720 if (SLEW_THRESHOLD < STEP_THRESHOLD) {
1721 if (tmx.offset > (long)(SLEW_THRESHOLD * 1000000)) {
1722 tmx.offset = (long)(SLEW_THRESHOLD * 1000000);
1724 if (tmx.offset < -(long)(SLEW_THRESHOLD * 1000000)) {
1725 tmx.offset = -(long)(SLEW_THRESHOLD * 1000000);
1729 tmx.status = STA_PLL;
1730 if (G.FREQHOLD_cnt != 0) {
1731 /* man adjtimex on STA_FREQHOLD:
1732 * "Normally adjustments made via ADJ_OFFSET result in dampened
1733 * frequency adjustments also being made.
1734 * This flag prevents the small frequency adjustment from being
1735 * made when correcting for an ADJ_OFFSET value."
1737 * Use this flag for a few first adjustments at the beginning
1738 * of ntpd execution, otherwise even relatively small initial
1739 * offset tend to cause largish changes to in-kernel tmx.freq.
1740 * If ntpd was restarted due to e.g. switch to another network,
1741 * this destroys already well-established tmx.freq value.
1743 if (G.FREQHOLD_cnt < 0) {
1745 // Example: a laptop whose clock runs slower when hibernated,
1746 // after wake up it still has good tmx.freq, but accumulated ~0.5 sec offset:
1747 // Run with code where initial G.FREQHOLD_cnt was always 8:
1748 //15:17:52.947 no valid datapoints, no peer selected
1749 //15:17:56.515 update from:<IP> offset:+0.485133 delay:0.157762 jitter:0.209310 clock drift:-1.393ppm tc:4
1750 //15:17:57.719 update from:<IP> offset:+0.483825 delay:0.158070 jitter:0.181159 clock drift:-1.393ppm tc:4
1751 //15:17:59.925 update from:<IP> offset:+0.479504 delay:0.158147 jitter:0.156657 clock drift:-1.393ppm tc:4
1752 //15:18:33.322 update from:<IP> offset:+0.428119 delay:0.158317 jitter:0.138071 clock drift:-1.393ppm tc:4
1753 //15:19:06.718 update from:<IP> offset:+0.376932 delay:0.158276 jitter:0.122075 clock drift:-1.393ppm tc:4
1754 //15:19:39.114 update from:<IP> offset:+0.327022 delay:0.158384 jitter:0.108538 clock drift:-1.393ppm tc:4
1755 //15:20:12.715 update from:<IP> offset:+0.275596 delay:0.158297 jitter:0.097292 clock drift:-1.393ppm tc:4
1756 //15:20:45.111 update from:<IP> offset:+0.225715 delay:0.158271 jitter:0.087841 clock drift:-1.393ppm tc:4
1757 // If allowed to continue, it would start increasing tmx.freq now.
1758 // Instead, it was ^Ced, and started anew:
1759 //15:21:15.043 no valid datapoints, no peer selected
1760 //15:21:17.408 update from:<IP> offset:+0.175910 delay:0.158314 jitter:0.076683 clock drift:-1.393ppm tc:4
1761 //15:21:19.774 update from:<IP> offset:+0.171784 delay:0.158401 jitter:0.066436 clock drift:-1.393ppm tc:4
1762 //15:21:22.140 update from:<IP> offset:+0.171660 delay:0.158592 jitter:0.057536 clock drift:-1.393ppm tc:4
1763 //15:21:22.140 update from:<IP> offset:+0.167126 delay:0.158507 jitter:0.049792 clock drift:-1.393ppm tc:4
1764 //15:21:55.696 update from:<IP> offset:+0.115223 delay:0.158277 jitter:0.050240 clock drift:-1.393ppm tc:4
1765 //15:22:29.093 update from:<IP> offset:+0.068051 delay:0.158243 jitter:0.049405 clock drift:-1.393ppm tc:5
1766 //15:23:02.490 update from:<IP> offset:+0.051632 delay:0.158215 jitter:0.043545 clock drift:-1.393ppm tc:5
1767 //15:23:34.726 update from:<IP> offset:+0.039984 delay:0.158157 jitter:0.038106 clock drift:-1.393ppm tc:5
1768 // STA_FREQHOLD no longer set, started increasing tmx.freq now:
1769 //15:24:06.961 update from:<IP> offset:+0.030968 delay:0.158190 jitter:0.033306 clock drift:+2.387ppm tc:5
1770 //15:24:40.357 update from:<IP> offset:+0.023648 delay:0.158211 jitter:0.029072 clock drift:+5.454ppm tc:5
1771 //15:25:13.774 update from:<IP> offset:+0.018068 delay:0.157660 jitter:0.025288 clock drift:+7.728ppm tc:5
1772 //15:26:19.173 update from:<IP> offset:+0.010057 delay:0.157969 jitter:0.022255 clock drift:+8.361ppm tc:6
1773 //15:27:26.602 update from:<IP> offset:+0.006737 delay:0.158103 jitter:0.019316 clock drift:+8.792ppm tc:6
1774 //15:28:33.030 update from:<IP> offset:+0.004513 delay:0.158294 jitter:0.016765 clock drift:+9.080ppm tc:6
1775 //15:29:40.617 update from:<IP> offset:+0.002787 delay:0.157745 jitter:0.014543 clock drift:+9.258ppm tc:6
1776 //15:30:47.045 update from:<IP> offset:+0.001324 delay:0.157709 jitter:0.012594 clock drift:+9.342ppm tc:6
1777 //15:31:53.473 update from:<IP> offset:+0.000007 delay:0.158142 jitter:0.010922 clock drift:+9.343ppm tc:6
1778 //15:32:58.902 update from:<IP> offset:-0.000728 delay:0.158222 jitter:0.009454 clock drift:+9.298ppm tc:6
1780 * This expression would choose MIN_FREQHOLD + 8 in the above example.
1782 G.FREQHOLD_cnt = 1 + MIN_FREQHOLD + ((unsigned)(abs(tmx.offset)) >> 16);
1785 tmx.status |= STA_FREQHOLD;
1787 if (G.ntp_status & LI_PLUSSEC)
1788 tmx.status |= STA_INS;
1789 if (G.ntp_status & LI_MINUSSEC)
1790 tmx.status |= STA_DEL;
1792 tmx.constant = (int)G.poll_exp - 4;
1794 * The below if statement should be unnecessary, but...
1795 * It looks like Linux kernel's PLL is far too gentle in changing
1796 * tmx.freq in response to clock offset. Offset keeps growing
1797 * and eventually we fall back to smaller poll intervals.
1798 * We can make correction more aggressive (about x2) by supplying
1799 * PLL time constant which is one less than the real one.
1800 * To be on a safe side, let's do it only if offset is significantly
1801 * larger than jitter.
1803 if (G.offset_to_jitter_ratio >= TIMECONST_HACK_GATE)
1805 if (tmx.constant < 0)
1808 //tmx.esterror = (uint32_t)(clock_jitter * 1e6);
1809 //tmx.maxerror = (uint32_t)((sys_rootdelay / 2 + sys_rootdisp) * 1e6);
1810 rc = adjtimex(&tmx);
1812 bb_perror_msg_and_die("adjtimex");
1813 /* NB: here kernel returns constant == G.poll_exp, not == G.poll_exp - 4.
1814 * Not sure why. Perhaps it is normal.
1816 VERB4 bb_error_msg("adjtimex:%d freq:%ld offset:%+ld status:0x%x",
1817 rc, tmx.freq, tmx.offset, tmx.status);
1818 G.kernel_freq_drift = tmx.freq / 65536;
1819 VERB2 bb_error_msg("update from:%s offset:%+f delay:%f jitter:%f clock drift:%+.3fppm tc:%d",
1823 G.discipline_jitter,
1824 (double)tmx.freq / 65536,
1828 return 1; /* "ok to increase poll interval" */
1833 * We've got a new reply packet from a peer, process it
1837 poll_interval(int upper_bound)
1839 unsigned interval, r, mask;
1840 interval = 1 << G.poll_exp;
1841 if (interval > upper_bound)
1842 interval = upper_bound;
1843 mask = ((interval-1) >> 4) | 1;
1845 interval += r & mask; /* ~ random(0..1) * interval/16 */
1846 VERB4 bb_error_msg("chose poll interval:%u (poll_exp:%d)", interval, G.poll_exp);
1850 adjust_poll(int count)
1852 G.polladj_count += count;
1853 if (G.polladj_count > POLLADJ_LIMIT) {
1854 G.polladj_count = 0;
1855 if (G.poll_exp < MAXPOLL) {
1857 VERB4 bb_error_msg("polladj: discipline_jitter:%f ++poll_exp=%d",
1858 G.discipline_jitter, G.poll_exp);
1860 } else if (G.polladj_count < -POLLADJ_LIMIT || (count < 0 && G.poll_exp > BIGPOLL)) {
1861 G.polladj_count = 0;
1862 if (G.poll_exp > MINPOLL) {
1866 /* Correct p->next_action_time in each peer
1867 * which waits for sending, so that they send earlier.
1868 * Old pp->next_action_time are on the order
1869 * of t + (1 << old_poll_exp) + small_random,
1870 * we simply need to subtract ~half of that.
1872 for (item = G.ntp_peers; item != NULL; item = item->link) {
1873 peer_t *pp = (peer_t *) item->data;
1875 pp->next_action_time -= (1 << G.poll_exp);
1877 VERB4 bb_error_msg("polladj: discipline_jitter:%f --poll_exp=%d",
1878 G.discipline_jitter, G.poll_exp);
1881 VERB4 bb_error_msg("polladj: count:%d", G.polladj_count);
1884 static NOINLINE void
1885 recv_and_process_peer_pkt(peer_t *p)
1890 double T1, T2, T3, T4;
1892 double prev_delay, delay;
1894 datapoint_t *datapoint;
1899 /* We can recvfrom here and check from.IP, but some multihomed
1900 * ntp servers reply from their *other IP*.
1901 * TODO: maybe we should check at least what we can: from.port == 123?
1904 size = recv(p->p_fd, &msg, sizeof(msg), MSG_DONTWAIT);
1909 if (errno == EAGAIN)
1910 /* There was no packet after all
1911 * (poll() returning POLLIN for a fd
1912 * is not a ironclad guarantee that data is there)
1916 * If you need a different handling for a specific
1917 * errno, always explain it in comment.
1919 bb_perror_msg_and_die("recv(%s) error", p->p_dotted);
1922 if (size != NTP_MSGSIZE_NOAUTH && size != NTP_MSGSIZE) {
1923 bb_error_msg("malformed packet received from %s", p->p_dotted);
1927 if (msg.m_orgtime.int_partl != p->p_xmt_msg.m_xmttime.int_partl
1928 || msg.m_orgtime.fractionl != p->p_xmt_msg.m_xmttime.fractionl
1930 /* Somebody else's packet */
1934 /* We do not expect any more packets from this peer for now.
1935 * Closing the socket informs kernel about it.
1936 * We open a new socket when we send a new query.
1941 if ((msg.m_status & LI_ALARM) == LI_ALARM
1942 || msg.m_stratum == 0
1943 || msg.m_stratum > NTP_MAXSTRATUM
1945 bb_error_msg("reply from %s: peer is unsynced", p->p_dotted);
1947 * Stratum 0 responses may have commands in 32-bit m_refid field:
1948 * "DENY", "RSTR" - peer does not like us at all,
1949 * "RATE" - peer is overloaded, reduce polling freq.
1950 * If poll interval is small, increase it.
1952 if (G.poll_exp < BIGPOLL)
1953 goto increase_interval;
1954 goto pick_normal_interval;
1957 // /* Verify valid root distance */
1958 // if (msg.m_rootdelay / 2 + msg.m_rootdisp >= MAXDISP || p->lastpkt_reftime > msg.m_xmt)
1959 // return; /* invalid header values */
1962 * From RFC 2030 (with a correction to the delay math):
1964 * Timestamp Name ID When Generated
1965 * ------------------------------------------------------------
1966 * Originate Timestamp T1 time request sent by client
1967 * Receive Timestamp T2 time request received by server
1968 * Transmit Timestamp T3 time reply sent by server
1969 * Destination Timestamp T4 time reply received by client
1971 * The roundtrip delay and local clock offset are defined as
1973 * delay = (T4 - T1) - (T3 - T2); offset = ((T2 - T1) + (T3 - T4)) / 2
1976 T2 = lfp_to_d(msg.m_rectime);
1977 T3 = lfp_to_d(msg.m_xmttime);
1979 delay = (T4 - T1) - (T3 - T2);
1982 * If this packet's delay is much bigger than the last one,
1983 * it's better to just ignore it than use its much less precise value.
1985 prev_delay = p->p_raw_delay;
1986 p->p_raw_delay = (delay < 0 ? 0.0 : delay);
1987 if (p->reachable_bits
1988 && delay > prev_delay * BAD_DELAY_GROWTH
1989 && delay > 1.0 / (8 * 1024) /* larger than ~0.000122 */
1991 bb_error_msg("reply from %s: delay %f is too high, ignoring", p->p_dotted, delay);
1992 goto pick_normal_interval;
1995 /* The delay calculation is a special case. In cases where the
1996 * server and client clocks are running at different rates and
1997 * with very fast networks, the delay can appear negative. In
1998 * order to avoid violating the Principle of Least Astonishment,
1999 * the delay is clamped not less than the system precision.
2001 if (delay < G_precision_sec)
2002 delay = G_precision_sec;
2003 p->lastpkt_delay = delay;
2004 p->lastpkt_recv_time = T4;
2005 VERB6 bb_error_msg("%s->lastpkt_recv_time=%f", p->p_dotted, p->lastpkt_recv_time);
2006 p->lastpkt_status = msg.m_status;
2007 p->lastpkt_stratum = msg.m_stratum;
2008 p->lastpkt_rootdelay = sfp_to_d(msg.m_rootdelay);
2009 p->lastpkt_rootdisp = sfp_to_d(msg.m_rootdisp);
2010 p->lastpkt_refid = msg.m_refid;
2012 p->datapoint_idx = p->reachable_bits ? (p->datapoint_idx + 1) % NUM_DATAPOINTS : 0;
2013 datapoint = &p->filter_datapoint[p->datapoint_idx];
2014 datapoint->d_recv_time = T4;
2015 datapoint->d_offset = offset = ((T2 - T1) + (T3 - T4)) / 2;
2016 datapoint->d_dispersion = LOG2D(msg.m_precision_exp) + G_precision_sec;
2017 if (!p->reachable_bits) {
2018 /* 1st datapoint ever - replicate offset in every element */
2020 for (i = 0; i < NUM_DATAPOINTS; i++) {
2021 p->filter_datapoint[i].d_offset = offset;
2025 p->reachable_bits |= 1;
2026 if ((MAX_VERBOSE && G.verbose) || (option_mask32 & OPT_w)) {
2027 bb_error_msg("reply from %s: offset:%+f delay:%f status:0x%02x strat:%d refid:0x%08x rootdelay:%f reach:0x%02x",
2034 p->lastpkt_rootdelay,
2036 /* not shown: m_ppoll, m_precision_exp, m_rootdisp,
2037 * m_reftime, m_orgtime, m_rectime, m_xmttime
2042 /* Muck with statictics and update the clock */
2043 filter_datapoints(p);
2044 q = select_and_cluster();
2047 if (!(option_mask32 & OPT_w)) {
2048 rc = update_local_clock(q);
2050 //Disabled this because there is a case where largish offsets
2051 //are unavoidable: if network round-trip delay is, say, ~0.6s,
2052 //error in offset estimation would be ~delay/2 ~= 0.3s.
2053 //Thus, offsets will be usually in -0.3...0.3s range.
2054 //In this case, this code would keep poll interval small,
2055 //but it won't be helping.
2056 //BIGOFF check below deals with a case of seeing multi-second offsets.
2058 /* If drift is dangerously large, immediately
2059 * drop poll interval one step down.
2061 if (fabs(q->filter_offset) >= POLLDOWN_OFFSET) {
2062 VERB4 bb_error_msg("offset:%+f > POLLDOWN_OFFSET", q->filter_offset);
2063 adjust_poll(-POLLADJ_LIMIT * 3);
2069 /* No peer selected.
2070 * If poll interval is small, increase it.
2072 if (G.poll_exp < BIGPOLL)
2073 goto increase_interval;
2077 /* Adjust the poll interval by comparing the current offset
2078 * with the clock jitter. If the offset is less than
2079 * the clock jitter times a constant, then the averaging interval
2080 * is increased, otherwise it is decreased. A bit of hysteresis
2081 * helps calm the dance. Works best using burst mode.
2083 if (rc > 0 && G.offset_to_jitter_ratio <= POLLADJ_GATE) {
2084 /* was += G.poll_exp but it is a bit
2085 * too optimistic for my taste at high poll_exp's */
2087 adjust_poll(MINPOLL);
2090 bb_error_msg("want smaller interval: offset/jitter = %u",
2091 G.offset_to_jitter_ratio);
2092 adjust_poll(-G.poll_exp * 2);
2096 /* Decide when to send new query for this peer */
2097 pick_normal_interval:
2098 interval = poll_interval(INT_MAX);
2099 if (fabs(offset) >= BIGOFF && interval > BIGOFF_INTERVAL) {
2100 /* If we are synced, offsets are less than SLEW_THRESHOLD,
2101 * or at the very least not much larger than it.
2102 * Now we see a largish one.
2103 * Either this peer is feeling bad, or packet got corrupted,
2104 * or _our_ clock is wrong now and _all_ peers will show similar
2105 * largish offsets too.
2106 * I observed this with laptop suspend stopping clock.
2107 * In any case, it makes sense to make next request soonish:
2108 * cases 1 and 2: get a better datapoint,
2109 * case 3: allows to resync faster.
2111 interval = BIGOFF_INTERVAL;
2114 set_next(p, interval);
2117 #if ENABLE_FEATURE_NTPD_SERVER
2118 static NOINLINE void
2119 recv_and_process_client_pkt(void /*int fd*/)
2123 len_and_sockaddr *to;
2124 struct sockaddr *from;
2126 uint8_t query_status;
2127 l_fixedpt_t query_xmttime;
2129 to = get_sock_lsa(G_listen_fd);
2130 from = xzalloc(to->len);
2132 size = recv_from_to(G_listen_fd, &msg, sizeof(msg), MSG_DONTWAIT, from, &to->u.sa, to->len);
2133 if (size != NTP_MSGSIZE_NOAUTH && size != NTP_MSGSIZE) {
2136 if (errno == EAGAIN)
2138 bb_perror_msg_and_die("recv");
2140 addr = xmalloc_sockaddr2dotted_noport(from);
2141 bb_error_msg("malformed packet received from %s: size %u", addr, (int)size);
2146 /* Respond only to client and symmetric active packets */
2147 if ((msg.m_status & MODE_MASK) != MODE_CLIENT
2148 && (msg.m_status & MODE_MASK) != MODE_SYM_ACT
2153 query_status = msg.m_status;
2154 query_xmttime = msg.m_xmttime;
2156 /* Build a reply packet */
2157 memset(&msg, 0, sizeof(msg));
2158 msg.m_status = G.stratum < MAXSTRAT ? (G.ntp_status & LI_MASK) : LI_ALARM;
2159 msg.m_status |= (query_status & VERSION_MASK);
2160 msg.m_status |= ((query_status & MODE_MASK) == MODE_CLIENT) ?
2161 MODE_SERVER : MODE_SYM_PAS;
2162 msg.m_stratum = G.stratum;
2163 msg.m_ppoll = G.poll_exp;
2164 msg.m_precision_exp = G_precision_exp;
2165 /* this time was obtained between poll() and recv() */
2166 msg.m_rectime = d_to_lfp(G.cur_time);
2167 msg.m_xmttime = d_to_lfp(gettime1900d()); /* this instant */
2168 if (G.peer_cnt == 0) {
2169 /* we have no peers: "stratum 1 server" mode. reftime = our own time */
2170 G.reftime = G.cur_time;
2172 msg.m_reftime = d_to_lfp(G.reftime);
2173 msg.m_orgtime = query_xmttime;
2174 msg.m_rootdelay = d_to_sfp(G.rootdelay);
2175 //simple code does not do this, fix simple code!
2176 msg.m_rootdisp = d_to_sfp(G.rootdisp);
2177 //version = (query_status & VERSION_MASK); /* ... >> VERSION_SHIFT - done below instead */
2178 msg.m_refid = G.refid; // (version > (3 << VERSION_SHIFT)) ? G.refid : G.refid3;
2180 /* We reply from the local address packet was sent to,
2181 * this makes to/from look swapped here: */
2182 do_sendto(G_listen_fd,
2183 /*from:*/ &to->u.sa, /*to:*/ from, /*addrlen:*/ to->len,
2192 /* Upstream ntpd's options:
2194 * -4 Force DNS resolution of host names to the IPv4 namespace.
2195 * -6 Force DNS resolution of host names to the IPv6 namespace.
2196 * -a Require cryptographic authentication for broadcast client,
2197 * multicast client and symmetric passive associations.
2198 * This is the default.
2199 * -A Do not require cryptographic authentication for broadcast client,
2200 * multicast client and symmetric passive associations.
2201 * This is almost never a good idea.
2202 * -b Enable the client to synchronize to broadcast servers.
2204 * Specify the name and path of the configuration file,
2205 * default /etc/ntp.conf
2206 * -d Specify debugging mode. This option may occur more than once,
2207 * with each occurrence indicating greater detail of display.
2209 * Specify debugging level directly.
2211 * Specify the name and path of the frequency file.
2212 * This is the same operation as the "driftfile FILE"
2213 * configuration command.
2214 * -g Normally, ntpd exits with a message to the system log
2215 * if the offset exceeds the panic threshold, which is 1000 s
2216 * by default. This option allows the time to be set to any value
2217 * without restriction; however, this can happen only once.
2218 * If the threshold is exceeded after that, ntpd will exit
2219 * with a message to the system log. This option can be used
2220 * with the -q and -x options. See the tinker command for other options.
2222 * Chroot the server to the directory jaildir. This option also implies
2223 * that the server attempts to drop root privileges at startup
2224 * (otherwise, chroot gives very little additional security).
2225 * You may need to also specify a -u option.
2227 * Specify the name and path of the symmetric key file,
2228 * default /etc/ntp/keys. This is the same operation
2229 * as the "keys FILE" configuration command.
2231 * Specify the name and path of the log file. The default
2232 * is the system log file. This is the same operation as
2233 * the "logfile FILE" configuration command.
2234 * -L Do not listen to virtual IPs. The default is to listen.
2236 * -N To the extent permitted by the operating system,
2237 * run the ntpd at the highest priority.
2239 * Specify the name and path of the file used to record the ntpd
2240 * process ID. This is the same operation as the "pidfile FILE"
2241 * configuration command.
2243 * To the extent permitted by the operating system,
2244 * run the ntpd at the specified priority.
2245 * -q Exit the ntpd just after the first time the clock is set.
2246 * This behavior mimics that of the ntpdate program, which is
2247 * to be retired. The -g and -x options can be used with this option.
2248 * Note: The kernel time discipline is disabled with this option.
2250 * Specify the default propagation delay from the broadcast/multicast
2251 * server to this client. This is necessary only if the delay
2252 * cannot be computed automatically by the protocol.
2254 * Specify the directory path for files created by the statistics
2255 * facility. This is the same operation as the "statsdir DIR"
2256 * configuration command.
2258 * Add a key number to the trusted key list. This option can occur
2261 * Specify a user, and optionally a group, to switch to.
2264 * Add a system variable listed by default.
2265 * -x Normally, the time is slewed if the offset is less than the step
2266 * threshold, which is 128 ms by default, and stepped if above
2267 * the threshold. This option sets the threshold to 600 s, which is
2268 * well within the accuracy window to set the clock manually.
2269 * Note: since the slew rate of typical Unix kernels is limited
2270 * to 0.5 ms/s, each second of adjustment requires an amortization
2271 * interval of 2000 s. Thus, an adjustment as much as 600 s
2272 * will take almost 14 days to complete. This option can be used
2273 * with the -g and -q options. See the tinker command for other options.
2274 * Note: The kernel time discipline is disabled with this option.
2277 /* By doing init in a separate function we decrease stack usage
2280 static NOINLINE void ntp_init(char **argv)
2288 bb_error_msg_and_die(bb_msg_you_must_be_root);
2290 /* Set some globals */
2291 G.discipline_jitter = G_precision_sec;
2292 G.stratum = MAXSTRAT;
2294 G.poll_exp = BURSTPOLL; /* speeds up initial sync */
2295 G.last_script_run = G.reftime = G.last_update_recv_time = gettime1900d(); /* sets G.cur_time too */
2296 G.FREQHOLD_cnt = -1;
2300 opts = getopt32(argv, "^"
2302 "wp:*S:"IF_FEATURE_NTPD_SERVER("l") /* NOT compat */
2303 IF_FEATURE_NTPD_SERVER("I:") /* compat */
2305 "46aAbgL" /* compat, ignored */
2307 "dd:wn" /* -d: counter; -p: list; -w implies -n */
2308 IF_FEATURE_NTPD_SERVER(":Il") /* -I implies -l */
2309 , &peers, &G.script_name,
2310 #if ENABLE_FEATURE_NTPD_SERVER
2315 // if (opts & OPT_x) /* disable stepping, only slew is allowed */
2316 // G.time_was_stepped = 1;
2318 #if ENABLE_FEATURE_NTPD_SERVER
2321 G_listen_fd = create_and_bind_dgram_or_die(NULL, 123);
2323 if (setsockopt_bindtodevice(G_listen_fd, G.if_name))
2326 socket_want_pktinfo(G_listen_fd);
2327 setsockopt_int(G_listen_fd, IPPROTO_IP, IP_TOS, IPTOS_DSCP_AF21);
2330 /* I hesitate to set -20 prio. -15 should be high enough for timekeeping */
2332 setpriority(PRIO_PROCESS, 0, -15);
2334 if (!(opts & OPT_n)) {
2335 bb_daemonize_or_rexec(DAEMON_DEVNULL_STDIO, argv);
2336 logmode = LOGMODE_NONE;
2341 add_peers(llist_pop(&peers));
2343 #if ENABLE_FEATURE_NTPD_CONF
2348 parser = config_open("/etc/ntp.conf");
2349 while (config_read(parser, token, 3, 1, "# \t", PARSE_NORMAL)) {
2350 if (strcmp(token[0], "server") == 0 && token[1]) {
2351 add_peers(token[1]);
2354 bb_error_msg("skipping %s:%u: unimplemented command '%s'",
2355 "/etc/ntp.conf", parser->lineno, token[0]
2358 config_close(parser);
2361 if (G.peer_cnt == 0) {
2362 if (!(opts & OPT_l))
2364 /* -l but no peers: "stratum 1 server" mode */
2367 /* If network is up, syncronization occurs in ~10 seconds.
2368 * We give "ntpd -q" 10 seconds to get first reply,
2369 * then another 50 seconds to finish syncing.
2371 * I tested ntpd 4.2.6p1 and apparently it never exits
2372 * (will try forever), but it does not feel right.
2373 * The goal of -q is to act like ntpdate: set time
2374 * after a reasonably small period of polling, or fail.
2377 option_mask32 |= OPT_qq;
2394 int ntpd_main(int argc UNUSED_PARAM, char **argv) MAIN_EXTERNALLY_VISIBLE;
2395 int ntpd_main(int argc UNUSED_PARAM, char **argv)
2403 memset(&G, 0, sizeof(G));
2404 SET_PTR_TO_GLOBALS(&G);
2408 /* If ENABLE_FEATURE_NTPD_SERVER, + 1 for listen_fd: */
2409 cnt = G.peer_cnt + ENABLE_FEATURE_NTPD_SERVER;
2410 idx2peer = xzalloc(sizeof(idx2peer[0]) * cnt);
2411 pfd = xzalloc(sizeof(pfd[0]) * cnt);
2413 /* Countdown: we never sync before we sent INITIAL_SAMPLES+1
2414 * packets to each peer.
2415 * NB: if some peer is not responding, we may end up sending
2416 * fewer packets to it and more to other peers.
2417 * NB2: sync usually happens using INITIAL_SAMPLES packets,
2418 * since last reply does not come back instantaneously.
2420 cnt = G.peer_cnt * (INITIAL_SAMPLES + 1);
2422 write_pidfile(CONFIG_PID_FILE_PATH "/ntpd.pid");
2424 while (!bb_got_signal) {
2430 /* Nothing between here and poll() blocks for any significant time */
2432 nextaction = G.last_script_run + (11*60);
2433 if (nextaction < G.cur_time + 1)
2434 nextaction = G.cur_time + 1;
2437 #if ENABLE_FEATURE_NTPD_SERVER
2438 if (G_listen_fd != -1) {
2439 pfd[0].fd = G_listen_fd;
2440 pfd[0].events = POLLIN;
2444 /* Pass over peer list, send requests, time out on receives */
2445 for (item = G.ntp_peers; item != NULL; item = item->link) {
2446 peer_t *p = (peer_t *) item->data;
2448 if (p->next_action_time <= G.cur_time) {
2449 if (p->p_fd == -1) {
2450 /* Time to send new req */
2452 VERB4 bb_error_msg("disabling burst mode");
2453 G.polladj_count = 0;
2454 G.poll_exp = MINPOLL;
2456 send_query_to_peer(p);
2458 /* Timed out waiting for reply */
2461 /* If poll interval is small, increase it */
2462 if (G.poll_exp < BIGPOLL)
2463 adjust_poll(MINPOLL);
2464 timeout = poll_interval(NOREPLY_INTERVAL);
2465 bb_error_msg("timed out waiting for %s, reach 0x%02x, next query in %us",
2466 p->p_dotted, p->reachable_bits, timeout);
2468 /* What if don't see it because it changed its IP? */
2469 if (p->reachable_bits == 0)
2470 resolve_peer_hostname(p);
2472 set_next(p, timeout);
2476 if (p->next_action_time < nextaction)
2477 nextaction = p->next_action_time;
2480 /* Wait for reply from this peer */
2481 pfd[i].fd = p->p_fd;
2482 pfd[i].events = POLLIN;
2488 timeout = nextaction - G.cur_time;
2491 timeout++; /* (nextaction - G.cur_time) rounds down, compensating */
2493 /* Here we may block */
2495 if (i > (ENABLE_FEATURE_NTPD_SERVER && G_listen_fd != -1)) {
2496 /* We wait for at least one reply.
2497 * Poll for it, without wasting time for message.
2498 * Since replies often come under 1 second, this also
2499 * reduces clutter in logs.
2501 nfds = poll(pfd, i, 1000);
2507 bb_error_msg("poll:%us sockets:%u interval:%us", timeout, i, 1 << G.poll_exp);
2509 nfds = poll(pfd, i, timeout * 1000);
2511 gettime1900d(); /* sets G.cur_time */
2517 break; /* poll was interrupted by a signal */
2519 if (G.cur_time - G.last_script_run > 11*60) {
2520 /* Useful for updating battery-backed RTC and such */
2521 run_script("periodic", G.last_update_offset);
2522 gettime1900d(); /* sets G.cur_time */
2525 /* Resolve peer names to IPs, if not resolved yet.
2526 * We do it only when poll timed out:
2527 * this way, we almost never overlap DNS resolution with
2528 * "request-reply" packet round trip.
2532 for (item = G.ntp_peers; item != NULL; item = item->link) {
2533 peer_t *p = (peer_t *) item->data;
2534 if (p->next_action_time <= ct && !p->p_lsa) {
2535 /* This can take up to ~10 sec per each DNS query */
2536 dns_error |= (!resolve_peer_hostname(p));
2541 /* Set next time for those which are still not resolved */
2542 gettime1900d(); /* sets G.cur_time (needed for set_next()) */
2543 for (item = G.ntp_peers; item != NULL; item = item->link) {
2544 peer_t *p = (peer_t *) item->data;
2545 if (p->next_action_time <= ct && !p->p_lsa) {
2546 set_next(p, HOSTNAME_INTERVAL * p->dns_errors);
2552 /* Process any received packets */
2554 #if ENABLE_FEATURE_NTPD_SERVER
2555 if (G.listen_fd != -1) {
2556 if (pfd[0].revents /* & (POLLIN|POLLERR)*/) {
2558 recv_and_process_client_pkt(/*G.listen_fd*/);
2559 gettime1900d(); /* sets G.cur_time */
2564 for (; nfds != 0 && j < i; j++) {
2565 if (pfd[j].revents /* & (POLLIN|POLLERR)*/) {
2567 * At init, alarm was set to 10 sec.
2568 * Now we did get a reply.
2569 * Increase timeout to 50 seconds to finish syncing.
2571 if (option_mask32 & OPT_qq) {
2572 option_mask32 &= ~OPT_qq;
2576 recv_and_process_peer_pkt(idx2peer[j]);
2577 gettime1900d(); /* sets G.cur_time */
2582 if (G.ntp_peers && G.stratum != MAXSTRAT) {
2583 for (item = G.ntp_peers; item != NULL; item = item->link) {
2584 peer_t *p = (peer_t *) item->data;
2585 if (p->reachable_bits)
2586 goto have_reachable_peer;
2588 /* No peer responded for last 8 packets, panic */
2589 clamp_pollexp_and_set_MAXSTRAT();
2590 run_script("unsync", 0.0);
2591 have_reachable_peer: ;
2593 } /* while (!bb_got_signal) */
2595 remove_pidfile(CONFIG_PID_FILE_PATH "/ntpd.pid");
2596 kill_myself_with_sig(bb_got_signal);
2604 /*** openntpd-4.6 uses only adjtime, not adjtimex ***/
2606 /*** ntp-4.2.6/ntpd/ntp_loopfilter.c - adjtimex usage ***/
2610 direct_freq(double fp_offset)
2614 * If the kernel is enabled, we need the residual offset to
2615 * calculate the frequency correction.
2617 if (pll_control && kern_enable) {
2618 memset(&ntv, 0, sizeof(ntv));
2621 clock_offset = ntv.offset / 1e9;
2622 #else /* STA_NANO */
2623 clock_offset = ntv.offset / 1e6;
2624 #endif /* STA_NANO */
2625 drift_comp = FREQTOD(ntv.freq);
2627 #endif /* KERNEL_PLL */
2628 set_freq((fp_offset - clock_offset) / (current_time - clock_epoch) + drift_comp);
2634 set_freq(double freq) /* frequency update */
2642 * If the kernel is enabled, update the kernel frequency.
2644 if (pll_control && kern_enable) {
2645 memset(&ntv, 0, sizeof(ntv));
2646 ntv.modes = MOD_FREQUENCY;
2647 ntv.freq = DTOFREQ(drift_comp);
2649 snprintf(tbuf, sizeof(tbuf), "kernel %.3f PPM", drift_comp * 1e6);
2650 report_event(EVNT_FSET, NULL, tbuf);
2652 snprintf(tbuf, sizeof(tbuf), "ntpd %.3f PPM", drift_comp * 1e6);
2653 report_event(EVNT_FSET, NULL, tbuf);
2655 #else /* KERNEL_PLL */
2656 snprintf(tbuf, sizeof(tbuf), "ntpd %.3f PPM", drift_comp * 1e6);
2657 report_event(EVNT_FSET, NULL, tbuf);
2658 #endif /* KERNEL_PLL */
2667 * This code segment works when clock adjustments are made using
2668 * precision time kernel support and the ntp_adjtime() system
2669 * call. This support is available in Solaris 2.6 and later,
2670 * Digital Unix 4.0 and later, FreeBSD, Linux and specially
2671 * modified kernels for HP-UX 9 and Ultrix 4. In the case of the
2672 * DECstation 5000/240 and Alpha AXP, additional kernel
2673 * modifications provide a true microsecond clock and nanosecond
2674 * clock, respectively.
2676 * Important note: The kernel discipline is used only if the
2677 * step threshold is less than 0.5 s, as anything higher can
2678 * lead to overflow problems. This might occur if some misguided
2679 * lad set the step threshold to something ridiculous.
2681 if (pll_control && kern_enable) {
2683 #define MOD_BITS (MOD_OFFSET | MOD_MAXERROR | MOD_ESTERROR | MOD_STATUS | MOD_TIMECONST)
2686 * We initialize the structure for the ntp_adjtime()
2687 * system call. We have to convert everything to
2688 * microseconds or nanoseconds first. Do not update the
2689 * system variables if the ext_enable flag is set. In
2690 * this case, the external clock driver will update the
2691 * variables, which will be read later by the local
2692 * clock driver. Afterwards, remember the time and
2693 * frequency offsets for jitter and stability values and
2694 * to update the frequency file.
2696 memset(&ntv, 0, sizeof(ntv));
2698 ntv.modes = MOD_STATUS;
2701 ntv.modes = MOD_BITS | MOD_NANO;
2702 #else /* STA_NANO */
2703 ntv.modes = MOD_BITS;
2704 #endif /* STA_NANO */
2705 if (clock_offset < 0)
2710 ntv.offset = (int32)(clock_offset * 1e9 + dtemp);
2711 ntv.constant = sys_poll;
2712 #else /* STA_NANO */
2713 ntv.offset = (int32)(clock_offset * 1e6 + dtemp);
2714 ntv.constant = sys_poll - 4;
2715 #endif /* STA_NANO */
2716 ntv.esterror = (u_int32)(clock_jitter * 1e6);
2717 ntv.maxerror = (u_int32)((sys_rootdelay / 2 + sys_rootdisp) * 1e6);
2718 ntv.status = STA_PLL;
2721 * Enable/disable the PPS if requested.
2724 if (!(pll_status & STA_PPSTIME))
2725 report_event(EVNT_KERN,
2726 NULL, "PPS enabled");
2727 ntv.status |= STA_PPSTIME | STA_PPSFREQ;
2729 if (pll_status & STA_PPSTIME)
2730 report_event(EVNT_KERN,
2731 NULL, "PPS disabled");
2732 ntv.status &= ~(STA_PPSTIME | STA_PPSFREQ);
2734 if (sys_leap == LEAP_ADDSECOND)
2735 ntv.status |= STA_INS;
2736 else if (sys_leap == LEAP_DELSECOND)
2737 ntv.status |= STA_DEL;
2741 * Pass the stuff to the kernel. If it squeals, turn off
2742 * the pps. In any case, fetch the kernel offset,
2743 * frequency and jitter.
2745 if (ntp_adjtime(&ntv) == TIME_ERROR) {
2746 if (!(ntv.status & STA_PPSSIGNAL))
2747 report_event(EVNT_KERN, NULL,
2750 pll_status = ntv.status;
2752 clock_offset = ntv.offset / 1e9;
2753 #else /* STA_NANO */
2754 clock_offset = ntv.offset / 1e6;
2755 #endif /* STA_NANO */
2756 clock_frequency = FREQTOD(ntv.freq);
2759 * If the kernel PPS is lit, monitor its performance.
2761 if (ntv.status & STA_PPSTIME) {
2763 clock_jitter = ntv.jitter / 1e9;
2764 #else /* STA_NANO */
2765 clock_jitter = ntv.jitter / 1e6;
2766 #endif /* STA_NANO */
2769 #if defined(STA_NANO) && NTP_API == 4
2771 * If the TAI changes, update the kernel TAI.
2773 if (loop_tai != sys_tai) {
2775 ntv.modes = MOD_TAI;
2776 ntv.constant = sys_tai;
2779 #endif /* STA_NANO */
2781 #endif /* KERNEL_PLL */