2 * NTP client/server, based on OpenNTPD 3.9p1
4 * Busybox port author: Adam Tkac (C) 2009 <vonsch@gmail.com>
6 * OpenNTPd 3.9p1 copyright holders:
7 * Copyright (c) 2003, 2004 Henning Brauer <henning@openbsd.org>
8 * Copyright (c) 2004 Alexander Guy <alexander.guy@andern.org>
10 * OpenNTPd code is licensed under ISC-style licence:
12 * Permission to use, copy, modify, and distribute this software for any
13 * purpose with or without fee is hereby granted, provided that the above
14 * copyright notice and this permission notice appear in all copies.
16 * THE SOFTWARE IS PROVIDED "AS IS" AND THE AUTHOR DISCLAIMS ALL WARRANTIES
17 * WITH REGARD TO THIS SOFTWARE INCLUDING ALL IMPLIED WARRANTIES OF
18 * MERCHANTABILITY AND FITNESS. IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR
19 * ANY SPECIAL, DIRECT, INDIRECT, OR CONSEQUENTIAL DAMAGES OR ANY DAMAGES
20 * WHATSOEVER RESULTING FROM LOSS OF MIND, USE, DATA OR PROFITS, WHETHER
21 * IN AN ACTION OF CONTRACT, NEGLIGENCE OR OTHER TORTIOUS ACTION, ARISING
22 * OUT OF OR IN CONNECTION WITH THE USE OR PERFORMANCE OF THIS SOFTWARE.
23 ***********************************************************************
25 * Parts of OpenNTPD clock syncronization code is replaced by
26 * code which is based on ntp-4.2.6, which carries the following
29 * Copyright (c) University of Delaware 1992-2009
31 * Permission to use, copy, modify, and distribute this software and
32 * its documentation for any purpose with or without fee is hereby
33 * granted, provided that the above copyright notice appears in all
34 * copies and that both the copyright notice and this permission
35 * notice appear in supporting documentation, and that the name
36 * University of Delaware not be used in advertising or publicity
37 * pertaining to distribution of the software without specific,
38 * written prior permission. The University of Delaware makes no
39 * representations about the suitability this software for any
40 * purpose. It is provided "as is" without express or implied warranty.
41 ***********************************************************************
44 //usage:#define ntpd_trivial_usage
45 //usage: "[-dnqNw"IF_FEATURE_NTPD_SERVER("l -I IFACE")"] [-S PROG] [-p PEER]..."
46 //usage:#define ntpd_full_usage "\n\n"
47 //usage: "NTP client/server\n"
48 //usage: "\n -d Verbose"
49 //usage: "\n -n Do not daemonize"
50 //usage: "\n -q Quit after clock is set"
51 //usage: "\n -N Run at high priority"
52 //usage: "\n -w Do not set time (only query peers), implies -n"
53 //usage: "\n -S PROG Run PROG after stepping time, stratum change, and every 11 mins"
54 //usage: "\n -p PEER Obtain time from PEER (may be repeated)"
55 //usage: IF_FEATURE_NTPD_CONF(
56 //usage: "\n If -p is not given, 'server HOST' lines"
57 //usage: "\n from /etc/ntp.conf are used"
59 //usage: IF_FEATURE_NTPD_SERVER(
60 //usage: "\n -l Also run as server on port 123"
61 //usage: "\n -I IFACE Bind server to IFACE, implies -l"
64 // -l and -p options are not compatible with "standard" ntpd:
65 // it has them as "-l logfile" and "-p pidfile".
66 // -S and -w are not compat either, "standard" ntpd has no such opts.
70 #include <netinet/ip.h> /* For IPTOS_LOWDELAY definition */
71 #include <sys/resource.h> /* setpriority */
72 #include <sys/timex.h>
73 #ifndef IPTOS_LOWDELAY
74 # define IPTOS_LOWDELAY 0x10
78 /* Verbosity control (max level of -dddd options accepted).
79 * max 6 is very talkative (and bloated). 3 is non-bloated,
80 * production level setting.
85 /* High-level description of the algorithm:
87 * We start running with very small poll_exp, BURSTPOLL,
88 * in order to quickly accumulate INITIAL_SAMPLES datapoints
89 * for each peer. Then, time is stepped if the offset is larger
90 * than STEP_THRESHOLD, otherwise it isn't; anyway, we enlarge
91 * poll_exp to MINPOLL and enter frequency measurement step:
92 * we collect new datapoints but ignore them for WATCH_THRESHOLD
93 * seconds. After WATCH_THRESHOLD seconds we look at accumulated
94 * offset and estimate frequency drift.
96 * (frequency measurement step seems to not be strictly needed,
97 * it is conditionally disabled with USING_INITIAL_FREQ_ESTIMATION
100 * After this, we enter "steady state": we collect a datapoint,
101 * we select the best peer, if this datapoint is not a new one
102 * (IOW: if this datapoint isn't for selected peer), sleep
103 * and collect another one; otherwise, use its offset to update
104 * frequency drift, if offset is somewhat large, reduce poll_exp,
105 * otherwise increase poll_exp.
107 * If offset is larger than STEP_THRESHOLD, which shouldn't normally
108 * happen, we assume that something "bad" happened (computer
109 * was hibernated, someone set totally wrong date, etc),
110 * then the time is stepped, all datapoints are discarded,
111 * and we go back to steady state.
113 * Made some changes to speed up re-syncing after our clock goes bad
114 * (tested with suspending my laptop):
115 * - if largish offset (>= STEP_THRESHOLD == 1 sec) is seen
116 * from a peer, schedule next query for this peer soon
117 * without drastically lowering poll interval for everybody.
118 * This makes us collect enough data for step much faster:
119 * e.g. at poll = 10 (1024 secs), step was done within 5 minutes
120 * after first reply which indicated that our clock is 14 seconds off.
121 * - on step, do not discard d_dispersion data of the existing datapoints,
122 * do not clear reachable_bits. This prevents discarding first ~8
123 * datapoints after the step.
126 #define INITIAL_SAMPLES 4 /* how many samples do we want for init */
127 #define BAD_DELAY_GROWTH 4 /* drop packet if its delay grew by more than this */
129 #define RETRY_INTERVAL 32 /* on send/recv error, retry in N secs (need to be power of 2) */
130 #define NOREPLY_INTERVAL 512 /* sent, but got no reply: cap next query by this many seconds */
131 #define RESPONSE_INTERVAL 16 /* wait for reply up to N secs */
133 /* Step threshold (sec). std ntpd uses 0.128.
135 #define STEP_THRESHOLD 1
136 /* Slew threshold (sec): adjtimex() won't accept offsets larger than this.
137 * Using exact power of 2 (1/8) results in smaller code
139 #define SLEW_THRESHOLD 0.125
140 /* Stepout threshold (sec). std ntpd uses 900 (11 mins (!)) */
141 #define WATCH_THRESHOLD 128
142 /* NB: set WATCH_THRESHOLD to ~60 when debugging to save time) */
143 //UNUSED: #define PANIC_THRESHOLD 1000 /* panic threshold (sec) */
146 * If we got |offset| > BIGOFF from a peer, cap next query interval
147 * for this peer by this many seconds:
149 #define BIGOFF STEP_THRESHOLD
150 #define BIGOFF_INTERVAL (1 << 7) /* 128 s */
152 #define FREQ_TOLERANCE 0.000015 /* frequency tolerance (15 PPM) */
153 #define BURSTPOLL 0 /* initial poll */
154 #define MINPOLL 5 /* minimum poll interval. std ntpd uses 6 (6: 64 sec) */
156 * If offset > discipline_jitter * POLLADJ_GATE, and poll interval is > 2^BIGPOLL,
157 * then it is decreased _at once_. (If <= 2^BIGPOLL, it will be decreased _eventually_).
159 #define BIGPOLL 9 /* 2^9 sec ~= 8.5 min */
160 #define MAXPOLL 12 /* maximum poll interval (12: 1.1h, 17: 36.4h). std ntpd uses 17 */
162 * Actively lower poll when we see such big offsets.
163 * With SLEW_THRESHOLD = 0.125, it means we try to sync more aggressively
164 * if offset increases over ~0.04 sec
166 //#define POLLDOWN_OFFSET (SLEW_THRESHOLD / 3)
167 #define MINDISP 0.01 /* minimum dispersion (sec) */
168 #define MAXDISP 16 /* maximum dispersion (sec) */
169 #define MAXSTRAT 16 /* maximum stratum (infinity metric) */
170 #define MAXDIST 1 /* distance threshold (sec) */
171 #define MIN_SELECTED 1 /* minimum intersection survivors */
172 #define MIN_CLUSTERED 3 /* minimum cluster survivors */
174 #define MAXDRIFT 0.000500 /* frequency drift we can correct (500 PPM) */
176 /* Poll-adjust threshold.
177 * When we see that offset is small enough compared to discipline jitter,
178 * we grow a counter: += MINPOLL. When counter goes over POLLADJ_LIMIT,
179 * we poll_exp++. If offset isn't small, counter -= poll_exp*2,
180 * and when it goes below -POLLADJ_LIMIT, we poll_exp--.
181 * (Bumped from 30 to 40 since otherwise I often see poll_exp going *2* steps down)
183 #define POLLADJ_LIMIT 40
184 /* If offset < discipline_jitter * POLLADJ_GATE, then we decide to increase
185 * poll interval (we think we can't improve timekeeping
186 * by staying at smaller poll).
188 #define POLLADJ_GATE 4
189 #define TIMECONST_HACK_GATE 2
190 /* Compromise Allan intercept (sec). doc uses 1500, std ntpd uses 512 */
194 /* FLL loop gain [why it depends on MAXPOLL??] */
195 #define FLL (MAXPOLL + 1)
196 /* Parameter averaging constant */
205 NTP_MSGSIZE_NOAUTH = 48,
206 NTP_MSGSIZE = (NTP_MSGSIZE_NOAUTH + 4 + NTP_DIGESTSIZE),
209 MODE_MASK = (7 << 0),
210 VERSION_MASK = (7 << 3),
214 /* Leap Second Codes (high order two bits of m_status) */
215 LI_NOWARNING = (0 << 6), /* no warning */
216 LI_PLUSSEC = (1 << 6), /* add a second (61 seconds) */
217 LI_MINUSSEC = (2 << 6), /* minus a second (59 seconds) */
218 LI_ALARM = (3 << 6), /* alarm condition */
221 MODE_RES0 = 0, /* reserved */
222 MODE_SYM_ACT = 1, /* symmetric active */
223 MODE_SYM_PAS = 2, /* symmetric passive */
224 MODE_CLIENT = 3, /* client */
225 MODE_SERVER = 4, /* server */
226 MODE_BROADCAST = 5, /* broadcast */
227 MODE_RES1 = 6, /* reserved for NTP control message */
228 MODE_RES2 = 7, /* reserved for private use */
231 //TODO: better base selection
232 #define OFFSET_1900_1970 2208988800UL /* 1970 - 1900 in seconds */
234 #define NUM_DATAPOINTS 8
247 uint8_t m_status; /* status of local clock and leap info */
249 uint8_t m_ppoll; /* poll value */
250 int8_t m_precision_exp;
251 s_fixedpt_t m_rootdelay;
252 s_fixedpt_t m_rootdisp;
254 l_fixedpt_t m_reftime;
255 l_fixedpt_t m_orgtime;
256 l_fixedpt_t m_rectime;
257 l_fixedpt_t m_xmttime;
259 uint8_t m_digest[NTP_DIGESTSIZE];
269 len_and_sockaddr *p_lsa;
273 uint32_t lastpkt_refid;
274 uint8_t lastpkt_status;
275 uint8_t lastpkt_stratum;
276 uint8_t reachable_bits;
277 /* when to send new query (if p_fd == -1)
278 * or when receive times out (if p_fd >= 0): */
279 double next_action_time;
282 /* p_raw_delay is set even by "high delay" packets */
283 /* lastpkt_delay isn't */
284 double lastpkt_recv_time;
285 double lastpkt_delay;
286 double lastpkt_rootdelay;
287 double lastpkt_rootdisp;
288 /* produced by filter algorithm: */
289 double filter_offset;
290 double filter_dispersion;
291 double filter_jitter;
292 datapoint_t filter_datapoint[NUM_DATAPOINTS];
293 /* last sent packet: */
298 #define USING_KERNEL_PLL_LOOP 1
299 #define USING_INITIAL_FREQ_ESTIMATION 0
306 /* Insert new options above this line. */
307 /* Non-compat options: */
311 OPT_l = (1 << 7) * ENABLE_FEATURE_NTPD_SERVER,
312 OPT_I = (1 << 8) * ENABLE_FEATURE_NTPD_SERVER,
313 /* We hijack some bits for other purposes */
319 /* total round trip delay to currently selected reference clock */
321 /* reference timestamp: time when the system clock was last set or corrected */
323 /* total dispersion to currently selected reference clock */
326 double last_script_run;
329 #if ENABLE_FEATURE_NTPD_SERVER
332 # define G_listen_fd (G.listen_fd)
334 # define G_listen_fd (-1)
338 /* refid: 32-bit code identifying the particular server or reference clock
339 * in stratum 0 packets this is a four-character ASCII string,
340 * called the kiss code, used for debugging and monitoring
341 * in stratum 1 packets this is a four-character ASCII string
342 * assigned to the reference clock by IANA. Example: "GPS "
343 * in stratum 2+ packets, it's IPv4 address or 4 first bytes
344 * of MD5 hash of IPv6
348 /* precision is defined as the larger of the resolution and time to
349 * read the clock, in log2 units. For instance, the precision of a
350 * mains-frequency clock incrementing at 60 Hz is 16 ms, even when the
351 * system clock hardware representation is to the nanosecond.
353 * Delays, jitters of various kinds are clamped down to precision.
355 * If precision_sec is too large, discipline_jitter gets clamped to it
356 * and if offset is smaller than discipline_jitter * POLLADJ_GATE, poll
357 * interval grows even though we really can benefit from staying at
358 * smaller one, collecting non-lagged datapoits and correcting offset.
359 * (Lagged datapoits exist when poll_exp is large but we still have
360 * systematic offset error - the time distance between datapoints
361 * is significant and older datapoints have smaller offsets.
362 * This makes our offset estimation a bit smaller than reality)
363 * Due to this effect, setting G_precision_sec close to
364 * STEP_THRESHOLD isn't such a good idea - offsets may grow
365 * too big and we will step. I observed it with -6.
367 * OTOH, setting precision_sec far too small would result in futile
368 * attempts to syncronize to an unachievable precision.
370 * -6 is 1/64 sec, -7 is 1/128 sec and so on.
371 * -8 is 1/256 ~= 0.003906 (worked well for me --vda)
372 * -9 is 1/512 ~= 0.001953 (let's try this for some time)
374 #define G_precision_exp -9
376 * G_precision_exp is used only for construction outgoing packets.
377 * It's ok to set G_precision_sec to a slightly different value
378 * (One which is "nicer looking" in logs).
379 * Exact value would be (1.0 / (1 << (- G_precision_exp))):
381 #define G_precision_sec 0.002
384 #define STATE_NSET 0 /* initial state, "nothing is set" */
385 //#define STATE_FSET 1 /* frequency set from file */
386 //#define STATE_SPIK 2 /* spike detected */
387 //#define STATE_FREQ 3 /* initial frequency */
388 #define STATE_SYNC 4 /* clock synchronized (normal operation) */
389 uint8_t discipline_state; // doc calls it c.state
390 uint8_t poll_exp; // s.poll
391 int polladj_count; // c.count
392 long kernel_freq_drift;
393 peer_t *last_update_peer;
394 double last_update_offset; // c.last
395 double last_update_recv_time; // s.t
396 double discipline_jitter; // c.jitter
397 /* Since we only compare it with ints, can simplify code
398 * by not making this variable floating point:
400 unsigned offset_to_jitter_ratio;
401 //double cluster_offset; // s.offset
402 //double cluster_jitter; // s.jitter
403 #if !USING_KERNEL_PLL_LOOP
404 double discipline_freq_drift; // c.freq
405 /* Maybe conditionally calculate wander? it's used only for logging */
406 double discipline_wander; // c.wander
409 #define G (*ptr_to_globals)
412 #define VERB1 if (MAX_VERBOSE && G.verbose)
413 #define VERB2 if (MAX_VERBOSE >= 2 && G.verbose >= 2)
414 #define VERB3 if (MAX_VERBOSE >= 3 && G.verbose >= 3)
415 #define VERB4 if (MAX_VERBOSE >= 4 && G.verbose >= 4)
416 #define VERB5 if (MAX_VERBOSE >= 5 && G.verbose >= 5)
417 #define VERB6 if (MAX_VERBOSE >= 6 && G.verbose >= 6)
420 static double LOG2D(int a)
423 return 1.0 / (1UL << -a);
426 static ALWAYS_INLINE double SQUARE(double x)
430 static ALWAYS_INLINE double MAXD(double a, double b)
436 static ALWAYS_INLINE double MIND(double a, double b)
442 static NOINLINE double my_SQRT(double X)
449 double Xhalf = X * 0.5;
451 /* Fast and good approximation to 1/sqrt(X), black magic */
453 /*v.i = 0x5f3759df - (v.i >> 1);*/
454 v.i = 0x5f375a86 - (v.i >> 1); /* - this constant is slightly better */
455 invsqrt = v.f; /* better than 0.2% accuracy */
457 /* Refining it using Newton's method: x1 = x0 - f(x0)/f'(x0)
458 * f(x) = 1/(x*x) - X (f==0 when x = 1/sqrt(X))
460 * f(x)/f'(x) = (X - 1/(x*x)) / (2/(x*x*x)) = X*x*x*x/2 - x/2
461 * x1 = x0 - (X*x0*x0*x0/2 - x0/2) = 1.5*x0 - X*x0*x0*x0/2 = x0*(1.5 - (X/2)*x0*x0)
463 invsqrt = invsqrt * (1.5 - Xhalf * invsqrt * invsqrt); /* ~0.05% accuracy */
464 /* invsqrt = invsqrt * (1.5 - Xhalf * invsqrt * invsqrt); 2nd iter: ~0.0001% accuracy */
465 /* With 4 iterations, more than half results will be exact,
466 * at 6th iterations result stabilizes with about 72% results exact.
467 * We are well satisfied with 0.05% accuracy.
470 return X * invsqrt; /* X * 1/sqrt(X) ~= sqrt(X) */
472 static ALWAYS_INLINE double SQRT(double X)
474 /* If this arch doesn't use IEEE 754 floats, fall back to using libm */
475 if (sizeof(float) != 4)
478 /* This avoids needing libm, saves about 0.5k on x86-32 */
486 gettimeofday(&tv, NULL); /* never fails */
487 G.cur_time = tv.tv_sec + (1.0e-6 * tv.tv_usec) + OFFSET_1900_1970;
492 d_to_tv(double d, struct timeval *tv)
494 tv->tv_sec = (long)d;
495 tv->tv_usec = (d - tv->tv_sec) * 1000000;
499 lfp_to_d(l_fixedpt_t lfp)
502 lfp.int_partl = ntohl(lfp.int_partl);
503 lfp.fractionl = ntohl(lfp.fractionl);
504 ret = (double)lfp.int_partl + ((double)lfp.fractionl / UINT_MAX);
508 sfp_to_d(s_fixedpt_t sfp)
511 sfp.int_parts = ntohs(sfp.int_parts);
512 sfp.fractions = ntohs(sfp.fractions);
513 ret = (double)sfp.int_parts + ((double)sfp.fractions / USHRT_MAX);
516 #if ENABLE_FEATURE_NTPD_SERVER
521 lfp.int_partl = (uint32_t)d;
522 lfp.fractionl = (uint32_t)((d - lfp.int_partl) * UINT_MAX);
523 lfp.int_partl = htonl(lfp.int_partl);
524 lfp.fractionl = htonl(lfp.fractionl);
531 sfp.int_parts = (uint16_t)d;
532 sfp.fractions = (uint16_t)((d - sfp.int_parts) * USHRT_MAX);
533 sfp.int_parts = htons(sfp.int_parts);
534 sfp.fractions = htons(sfp.fractions);
540 dispersion(const datapoint_t *dp)
542 return dp->d_dispersion + FREQ_TOLERANCE * (G.cur_time - dp->d_recv_time);
546 root_distance(peer_t *p)
548 /* The root synchronization distance is the maximum error due to
549 * all causes of the local clock relative to the primary server.
550 * It is defined as half the total delay plus total dispersion
553 return MAXD(MINDISP, p->lastpkt_rootdelay + p->lastpkt_delay) / 2
554 + p->lastpkt_rootdisp
555 + p->filter_dispersion
556 + FREQ_TOLERANCE * (G.cur_time - p->lastpkt_recv_time)
561 set_next(peer_t *p, unsigned t)
563 p->next_action_time = G.cur_time + t;
567 * Peer clock filter and its helpers
570 filter_datapoints(peer_t *p)
577 /* Simulations have shown that use of *averaged* offset for p->filter_offset
578 * is in fact worse than simply using last received one: with large poll intervals
579 * (>= 2048) averaging code uses offset values which are outdated by hours,
580 * and time/frequency correction goes totally wrong when fed essentially bogus offsets.
583 double minoff, maxoff, w;
584 double x = x; /* for compiler */
585 double oldest_off = oldest_off;
586 double oldest_age = oldest_age;
587 double newest_off = newest_off;
588 double newest_age = newest_age;
590 fdp = p->filter_datapoint;
592 minoff = maxoff = fdp[0].d_offset;
593 for (i = 1; i < NUM_DATAPOINTS; i++) {
594 if (minoff > fdp[i].d_offset)
595 minoff = fdp[i].d_offset;
596 if (maxoff < fdp[i].d_offset)
597 maxoff = fdp[i].d_offset;
600 idx = p->datapoint_idx; /* most recent datapoint's index */
602 * Drop two outliers and take weighted average of the rest:
603 * most_recent/2 + older1/4 + older2/8 ... + older5/32 + older6/32
604 * we use older6/32, not older6/64 since sum of weights should be 1:
605 * 1/2 + 1/4 + 1/8 + 1/16 + 1/32 + 1/32 = 1
611 * filter_dispersion = \ -------------
618 for (i = 0; i < NUM_DATAPOINTS; i++) {
620 bb_error_msg("datapoint[%d]: off:%f disp:%f(%f) age:%f%s",
623 fdp[idx].d_dispersion, dispersion(&fdp[idx]),
624 G.cur_time - fdp[idx].d_recv_time,
625 (minoff == fdp[idx].d_offset || maxoff == fdp[idx].d_offset)
626 ? " (outlier by offset)" : ""
630 sum += dispersion(&fdp[idx]) / (2 << i);
632 if (minoff == fdp[idx].d_offset) {
633 minoff -= 1; /* so that we don't match it ever again */
635 if (maxoff == fdp[idx].d_offset) {
638 oldest_off = fdp[idx].d_offset;
639 oldest_age = G.cur_time - fdp[idx].d_recv_time;
642 newest_off = oldest_off;
643 newest_age = oldest_age;
650 idx = (idx - 1) & (NUM_DATAPOINTS - 1);
652 p->filter_dispersion = sum;
653 wavg += x; /* add another older6/64 to form older6/32 */
654 /* Fix systematic underestimation with large poll intervals.
655 * Imagine that we still have a bit of uncorrected drift,
656 * and poll interval is big (say, 100 sec). Offsets form a progression:
657 * 0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 - 0.7 is most recent.
658 * The algorithm above drops 0.0 and 0.7 as outliers,
659 * and then we have this estimation, ~25% off from 0.7:
660 * 0.1/32 + 0.2/32 + 0.3/16 + 0.4/8 + 0.5/4 + 0.6/2 = 0.503125
662 x = oldest_age - newest_age;
664 x = newest_age / x; /* in above example, 100 / (600 - 100) */
665 if (x < 1) { /* paranoia check */
666 x = (newest_off - oldest_off) * x; /* 0.5 * 100/500 = 0.1 */
670 p->filter_offset = wavg;
674 fdp = p->filter_datapoint;
675 idx = p->datapoint_idx; /* most recent datapoint's index */
677 /* filter_offset: simply use the most recent value */
678 p->filter_offset = fdp[idx].d_offset;
682 * filter_dispersion = \ -------------
689 for (i = 0; i < NUM_DATAPOINTS; i++) {
690 sum += dispersion(&fdp[idx]) / (2 << i);
691 wavg += fdp[idx].d_offset;
692 idx = (idx - 1) & (NUM_DATAPOINTS - 1);
694 wavg /= NUM_DATAPOINTS;
695 p->filter_dispersion = sum;
698 /* +----- -----+ ^ 1/2
702 * filter_jitter = | --- * / (avg-offset_j) |
706 * where n is the number of valid datapoints in the filter (n > 1);
707 * if filter_jitter < precision then filter_jitter = precision
710 for (i = 0; i < NUM_DATAPOINTS; i++) {
711 sum += SQUARE(wavg - fdp[i].d_offset);
713 sum = SQRT(sum / NUM_DATAPOINTS);
714 p->filter_jitter = sum > G_precision_sec ? sum : G_precision_sec;
716 VERB4 bb_error_msg("filter offset:%+f disp:%f jitter:%f",
718 p->filter_dispersion,
723 reset_peer_stats(peer_t *p, double offset)
726 bool small_ofs = fabs(offset) < STEP_THRESHOLD;
728 /* Used to set p->filter_datapoint[i].d_dispersion = MAXDISP
729 * and clear reachable bits, but this proved to be too agressive:
730 * after step (tested with suspending laptop for ~30 secs),
731 * this caused all previous data to be considered invalid,
732 * making us needing to collect full ~8 datapoins per peer
733 * after step in order to start trusting them.
734 * In turn, this was making poll interval decrease even after
735 * step was done. (Poll interval decreases already before step
736 * in this scenario, because we see large offsets and end up with
737 * no good peer to select).
740 for (i = 0; i < NUM_DATAPOINTS; i++) {
742 p->filter_datapoint[i].d_recv_time += offset;
743 if (p->filter_datapoint[i].d_offset != 0) {
744 p->filter_datapoint[i].d_offset -= offset;
745 //bb_error_msg("p->filter_datapoint[%d].d_offset %f -> %f",
747 // p->filter_datapoint[i].d_offset + offset,
748 // p->filter_datapoint[i].d_offset);
751 p->filter_datapoint[i].d_recv_time = G.cur_time;
752 p->filter_datapoint[i].d_offset = 0;
753 /*p->filter_datapoint[i].d_dispersion = MAXDISP;*/
757 p->lastpkt_recv_time += offset;
759 /*p->reachable_bits = 0;*/
760 p->lastpkt_recv_time = G.cur_time;
762 filter_datapoints(p); /* recalc p->filter_xxx */
763 VERB6 bb_error_msg("%s->lastpkt_recv_time=%f", p->p_dotted, p->lastpkt_recv_time);
767 add_peers(const char *s)
772 p = xzalloc(sizeof(*p));
773 p->p_lsa = xhost2sockaddr(s, 123);
774 p->p_dotted = xmalloc_sockaddr2dotted_noport(&p->p_lsa->u.sa);
776 /* Names like N.<country2chars>.pool.ntp.org are randomly resolved
777 * to a pool of machines. Sometimes different N's resolve to the same IP.
778 * It is not useful to have two peers with same IP. We skip duplicates.
780 for (item = G.ntp_peers; item != NULL; item = item->link) {
781 peer_t *pp = (peer_t *) item->data;
782 if (strcmp(p->p_dotted, pp->p_dotted) == 0) {
783 bb_error_msg("duplicate peer %s (%s)", s, p->p_dotted);
790 p->p_xmt_msg.m_status = MODE_CLIENT | (NTP_VERSION << 3);
791 p->next_action_time = G.cur_time; /* = set_next(p, 0); */
792 reset_peer_stats(p, STEP_THRESHOLD);
794 llist_add_to(&G.ntp_peers, p);
800 const struct sockaddr *from, const struct sockaddr *to, socklen_t addrlen,
801 msg_t *msg, ssize_t len)
807 ret = sendto(fd, msg, len, MSG_DONTWAIT, to, addrlen);
809 ret = send_to_from(fd, msg, len, MSG_DONTWAIT, to, from, addrlen);
812 bb_perror_msg("send failed");
819 send_query_to_peer(peer_t *p)
821 /* Why do we need to bind()?
822 * See what happens when we don't bind:
824 * socket(PF_INET, SOCK_DGRAM, IPPROTO_IP) = 3
825 * setsockopt(3, SOL_IP, IP_TOS, [16], 4) = 0
826 * gettimeofday({1259071266, 327885}, NULL) = 0
827 * sendto(3, "xxx", 48, MSG_DONTWAIT, {sa_family=AF_INET, sin_port=htons(123), sin_addr=inet_addr("10.34.32.125")}, 16) = 48
828 * ^^^ we sent it from some source port picked by kernel.
829 * time(NULL) = 1259071266
830 * write(2, "ntpd: entering poll 15 secs\n", 28) = 28
831 * poll([{fd=3, events=POLLIN}], 1, 15000) = 1 ([{fd=3, revents=POLLIN}])
832 * recv(3, "yyy", 68, MSG_DONTWAIT) = 48
833 * ^^^ this recv will receive packets to any local port!
835 * Uncomment this and use strace to see it in action:
837 #define PROBE_LOCAL_ADDR /* { len_and_sockaddr lsa; lsa.len = LSA_SIZEOF_SA; getsockname(p->query.fd, &lsa.u.sa, &lsa.len); } */
841 len_and_sockaddr *local_lsa;
843 family = p->p_lsa->u.sa.sa_family;
844 p->p_fd = fd = xsocket_type(&local_lsa, family, SOCK_DGRAM);
845 /* local_lsa has "null" address and port 0 now.
846 * bind() ensures we have a *particular port* selected by kernel
847 * and remembered in p->p_fd, thus later recv(p->p_fd)
848 * receives only packets sent to this port.
851 xbind(fd, &local_lsa->u.sa, local_lsa->len);
853 #if ENABLE_FEATURE_IPV6
854 if (family == AF_INET)
856 setsockopt_int(fd, IPPROTO_IP, IP_TOS, IPTOS_LOWDELAY);
860 /* Emit message _before_ attempted send. Think of a very short
861 * roundtrip networks: we need to go back to recv loop ASAP,
862 * to reduce delay. Printing messages after send works against that.
864 VERB1 bb_error_msg("sending query to %s", p->p_dotted);
867 * Send out a random 64-bit number as our transmit time. The NTP
868 * server will copy said number into the originate field on the
869 * response that it sends us. This is totally legal per the SNTP spec.
871 * The impact of this is two fold: we no longer send out the current
872 * system time for the world to see (which may aid an attacker), and
873 * it gives us a (not very secure) way of knowing that we're not
874 * getting spoofed by an attacker that can't capture our traffic
875 * but can spoof packets from the NTP server we're communicating with.
877 * Save the real transmit timestamp locally.
879 p->p_xmt_msg.m_xmttime.int_partl = rand();
880 p->p_xmt_msg.m_xmttime.fractionl = rand();
881 p->p_xmttime = gettime1900d();
883 /* Were doing it only if sendto worked, but
884 * loss of sync detection needs reachable_bits updated
885 * even if sending fails *locally*:
886 * "network is unreachable" because cable was pulled?
887 * We still need to declare "unsync" if this condition persists.
889 p->reachable_bits <<= 1;
891 if (do_sendto(p->p_fd, /*from:*/ NULL, /*to:*/ &p->p_lsa->u.sa, /*addrlen:*/ p->p_lsa->len,
892 &p->p_xmt_msg, NTP_MSGSIZE_NOAUTH) == -1
897 * We know that we sent nothing.
898 * We can retry *soon* without fearing
899 * that we are flooding the peer.
901 set_next(p, RETRY_INTERVAL);
905 set_next(p, RESPONSE_INTERVAL);
909 /* Note that there is no provision to prevent several run_scripts
910 * to be started in quick succession. In fact, it happens rather often
911 * if initial syncronization results in a step.
912 * You will see "step" and then "stratum" script runs, sometimes
913 * as close as only 0.002 seconds apart.
914 * Script should be ready to deal with this.
916 static void run_script(const char *action, double offset)
919 char *env1, *env2, *env3, *env4;
921 G.last_script_run = G.cur_time;
926 argv[0] = (char*) G.script_name;
927 argv[1] = (char*) action;
930 VERB1 bb_error_msg("executing '%s %s'", G.script_name, action);
932 env1 = xasprintf("%s=%u", "stratum", G.stratum);
934 env2 = xasprintf("%s=%ld", "freq_drift_ppm", G.kernel_freq_drift);
936 env3 = xasprintf("%s=%u", "poll_interval", 1 << G.poll_exp);
938 env4 = xasprintf("%s=%f", "offset", offset);
940 /* Other items of potential interest: selected peer,
941 * rootdelay, reftime, rootdisp, refid, ntp_status,
942 * last_update_offset, last_update_recv_time, discipline_jitter,
943 * how many peers have reachable_bits = 0?
946 /* Don't want to wait: it may run hwclock --systohc, and that
947 * may take some time (seconds): */
948 /*spawn_and_wait(argv);*/
952 unsetenv("freq_drift_ppm");
953 unsetenv("poll_interval");
962 step_time(double offset)
966 struct timeval tvc, tvn;
967 char buf[sizeof("yyyy-mm-dd hh:mm:ss") + /*paranoia:*/ 4];
970 gettimeofday(&tvc, NULL); /* never fails */
971 dtime = tvc.tv_sec + (1.0e-6 * tvc.tv_usec) + offset;
972 d_to_tv(dtime, &tvn);
973 if (settimeofday(&tvn, NULL) == -1)
974 bb_perror_msg_and_die("settimeofday");
978 strftime_YYYYMMDDHHMMSS(buf, sizeof(buf), &tval);
979 bb_error_msg("current time is %s.%06u", buf, (unsigned)tvc.tv_usec);
982 strftime_YYYYMMDDHHMMSS(buf, sizeof(buf), &tval);
983 bb_error_msg("setting time to %s.%06u (offset %+fs)", buf, (unsigned)tvn.tv_usec, offset);
985 /* Correct various fields which contain time-relative values: */
988 G.cur_time += offset;
989 G.last_update_recv_time += offset;
990 G.last_script_run += offset;
992 /* p->lastpkt_recv_time, p->next_action_time and such: */
993 for (item = G.ntp_peers; item != NULL; item = item->link) {
994 peer_t *pp = (peer_t *) item->data;
995 reset_peer_stats(pp, offset);
996 //bb_error_msg("offset:%+f pp->next_action_time:%f -> %f",
997 // offset, pp->next_action_time, pp->next_action_time + offset);
998 pp->next_action_time += offset;
1000 /* We wait for reply from this peer too.
1001 * But due to step we are doing, reply's data is no longer
1002 * useful (in fact, it'll be bogus). Stop waiting for it.
1006 set_next(pp, RETRY_INTERVAL);
1011 static void clamp_pollexp_and_set_MAXSTRAT(void)
1013 if (G.poll_exp < MINPOLL)
1014 G.poll_exp = MINPOLL;
1015 if (G.poll_exp > BIGPOLL)
1016 G.poll_exp = BIGPOLL;
1017 G.polladj_count = 0;
1018 G.stratum = MAXSTRAT;
1023 * Selection and clustering, and their helpers
1029 double opt_rd; /* optimization */
1032 compare_point_edge(const void *aa, const void *bb)
1034 const point_t *a = aa;
1035 const point_t *b = bb;
1036 if (a->edge < b->edge) {
1039 return (a->edge > b->edge);
1046 compare_survivor_metric(const void *aa, const void *bb)
1048 const survivor_t *a = aa;
1049 const survivor_t *b = bb;
1050 if (a->metric < b->metric) {
1053 return (a->metric > b->metric);
1056 fit(peer_t *p, double rd)
1058 if ((p->reachable_bits & (p->reachable_bits-1)) == 0) {
1059 /* One or zero bits in reachable_bits */
1060 VERB4 bb_error_msg("peer %s unfit for selection: unreachable", p->p_dotted);
1063 #if 0 /* we filter out such packets earlier */
1064 if ((p->lastpkt_status & LI_ALARM) == LI_ALARM
1065 || p->lastpkt_stratum >= MAXSTRAT
1067 VERB4 bb_error_msg("peer %s unfit for selection: bad status/stratum", p->p_dotted);
1071 /* rd is root_distance(p) */
1072 if (rd > MAXDIST + FREQ_TOLERANCE * (1 << G.poll_exp)) {
1073 VERB4 bb_error_msg("peer %s unfit for selection: root distance too high", p->p_dotted);
1077 // /* Do we have a loop? */
1078 // if (p->refid == p->dstaddr || p->refid == s.refid)
1083 select_and_cluster(void)
1088 int size = 3 * G.peer_cnt;
1089 /* for selection algorithm */
1090 point_t point[size];
1091 unsigned num_points, num_candidates;
1093 unsigned num_falsetickers;
1094 /* for cluster algorithm */
1095 survivor_t survivor[size];
1096 unsigned num_survivors;
1102 while (item != NULL) {
1105 p = (peer_t *) item->data;
1106 rd = root_distance(p);
1107 offset = p->filter_offset;
1113 VERB5 bb_error_msg("interval: [%f %f %f] %s",
1119 point[num_points].p = p;
1120 point[num_points].type = -1;
1121 point[num_points].edge = offset - rd;
1122 point[num_points].opt_rd = rd;
1124 point[num_points].p = p;
1125 point[num_points].type = 0;
1126 point[num_points].edge = offset;
1127 point[num_points].opt_rd = rd;
1129 point[num_points].p = p;
1130 point[num_points].type = 1;
1131 point[num_points].edge = offset + rd;
1132 point[num_points].opt_rd = rd;
1136 num_candidates = num_points / 3;
1137 if (num_candidates == 0) {
1138 VERB3 bb_error_msg("no valid datapoints%s", ", no peer selected");
1141 //TODO: sorting does not seem to be done in reference code
1142 qsort(point, num_points, sizeof(point[0]), compare_point_edge);
1144 /* Start with the assumption that there are no falsetickers.
1145 * Attempt to find a nonempty intersection interval containing
1146 * the midpoints of all truechimers.
1147 * If a nonempty interval cannot be found, increase the number
1148 * of assumed falsetickers by one and try again.
1149 * If a nonempty interval is found and the number of falsetickers
1150 * is less than the number of truechimers, a majority has been found
1151 * and the midpoint of each truechimer represents
1152 * the candidates available to the cluster algorithm.
1154 num_falsetickers = 0;
1157 unsigned num_midpoints = 0;
1162 for (i = 0; i < num_points; i++) {
1164 * if (point[i].type == -1) c++;
1165 * if (point[i].type == 1) c--;
1166 * and it's simpler to do it this way:
1169 if (c >= num_candidates - num_falsetickers) {
1170 /* If it was c++ and it got big enough... */
1171 low = point[i].edge;
1174 if (point[i].type == 0)
1178 for (i = num_points-1; i >= 0; i--) {
1180 if (c >= num_candidates - num_falsetickers) {
1181 high = point[i].edge;
1184 if (point[i].type == 0)
1187 /* If the number of midpoints is greater than the number
1188 * of allowed falsetickers, the intersection contains at
1189 * least one truechimer with no midpoint - bad.
1190 * Also, interval should be nonempty.
1192 if (num_midpoints <= num_falsetickers && low < high)
1195 if (num_falsetickers * 2 >= num_candidates) {
1196 VERB3 bb_error_msg("falsetickers:%d, candidates:%d%s",
1197 num_falsetickers, num_candidates,
1198 ", no peer selected");
1202 VERB4 bb_error_msg("selected interval: [%f, %f]; candidates:%d falsetickers:%d",
1203 low, high, num_candidates, num_falsetickers);
1207 /* Construct a list of survivors (p, metric)
1208 * from the chime list, where metric is dominated
1209 * first by stratum and then by root distance.
1210 * All other things being equal, this is the order of preference.
1213 for (i = 0; i < num_points; i++) {
1214 if (point[i].edge < low || point[i].edge > high)
1217 survivor[num_survivors].p = p;
1218 /* x.opt_rd == root_distance(p); */
1219 survivor[num_survivors].metric = MAXDIST * p->lastpkt_stratum + point[i].opt_rd;
1220 VERB5 bb_error_msg("survivor[%d] metric:%f peer:%s",
1221 num_survivors, survivor[num_survivors].metric, p->p_dotted);
1224 /* There must be at least MIN_SELECTED survivors to satisfy the
1225 * correctness assertions. Ordinarily, the Byzantine criteria
1226 * require four survivors, but for the demonstration here, one
1229 if (num_survivors < MIN_SELECTED) {
1230 VERB3 bb_error_msg("survivors:%d%s",
1232 ", no peer selected");
1236 //looks like this is ONLY used by the fact that later we pick survivor[0].
1237 //we can avoid sorting then, just find the minimum once!
1238 qsort(survivor, num_survivors, sizeof(survivor[0]), compare_survivor_metric);
1240 /* For each association p in turn, calculate the selection
1241 * jitter p->sjitter as the square root of the sum of squares
1242 * (p->offset - q->offset) over all q associations. The idea is
1243 * to repeatedly discard the survivor with maximum selection
1244 * jitter until a termination condition is met.
1247 unsigned max_idx = max_idx;
1248 double max_selection_jitter = max_selection_jitter;
1249 double min_jitter = min_jitter;
1251 if (num_survivors <= MIN_CLUSTERED) {
1252 VERB4 bb_error_msg("num_survivors %d <= %d, not discarding more",
1253 num_survivors, MIN_CLUSTERED);
1257 /* To make sure a few survivors are left
1258 * for the clustering algorithm to chew on,
1259 * we stop if the number of survivors
1260 * is less than or equal to MIN_CLUSTERED (3).
1262 for (i = 0; i < num_survivors; i++) {
1263 double selection_jitter_sq;
1266 if (i == 0 || p->filter_jitter < min_jitter)
1267 min_jitter = p->filter_jitter;
1269 selection_jitter_sq = 0;
1270 for (j = 0; j < num_survivors; j++) {
1271 peer_t *q = survivor[j].p;
1272 selection_jitter_sq += SQUARE(p->filter_offset - q->filter_offset);
1274 if (i == 0 || selection_jitter_sq > max_selection_jitter) {
1275 max_selection_jitter = selection_jitter_sq;
1278 VERB6 bb_error_msg("survivor %d selection_jitter^2:%f",
1279 i, selection_jitter_sq);
1281 max_selection_jitter = SQRT(max_selection_jitter / num_survivors);
1282 VERB5 bb_error_msg("max_selection_jitter (at %d):%f min_jitter:%f",
1283 max_idx, max_selection_jitter, min_jitter);
1285 /* If the maximum selection jitter is less than the
1286 * minimum peer jitter, then tossing out more survivors
1287 * will not lower the minimum peer jitter, so we might
1290 if (max_selection_jitter < min_jitter) {
1291 VERB4 bb_error_msg("max_selection_jitter:%f < min_jitter:%f, num_survivors:%d, not discarding more",
1292 max_selection_jitter, min_jitter, num_survivors);
1296 /* Delete survivor[max_idx] from the list
1297 * and go around again.
1299 VERB6 bb_error_msg("dropping survivor %d", max_idx);
1301 while (max_idx < num_survivors) {
1302 survivor[max_idx] = survivor[max_idx + 1];
1308 /* Combine the offsets of the clustering algorithm survivors
1309 * using a weighted average with weight determined by the root
1310 * distance. Compute the selection jitter as the weighted RMS
1311 * difference between the first survivor and the remaining
1312 * survivors. In some cases the inherent clock jitter can be
1313 * reduced by not using this algorithm, especially when frequent
1314 * clockhopping is involved. bbox: thus we don't do it.
1318 for (i = 0; i < num_survivors; i++) {
1320 x = root_distance(p);
1322 z += p->filter_offset / x;
1323 w += SQUARE(p->filter_offset - survivor[0].p->filter_offset) / x;
1325 //G.cluster_offset = z / y;
1326 //G.cluster_jitter = SQRT(w / y);
1329 /* Pick the best clock. If the old system peer is on the list
1330 * and at the same stratum as the first survivor on the list,
1331 * then don't do a clock hop. Otherwise, select the first
1332 * survivor on the list as the new system peer.
1335 if (G.last_update_peer
1336 && G.last_update_peer->lastpkt_stratum <= p->lastpkt_stratum
1338 /* Starting from 1 is ok here */
1339 for (i = 1; i < num_survivors; i++) {
1340 if (G.last_update_peer == survivor[i].p) {
1341 VERB5 bb_error_msg("keeping old synced peer");
1342 p = G.last_update_peer;
1347 G.last_update_peer = p;
1349 VERB4 bb_error_msg("selected peer %s filter_offset:%+f age:%f",
1352 G.cur_time - p->lastpkt_recv_time
1359 * Local clock discipline and its helpers
1362 set_new_values(int disc_state, double offset, double recv_time)
1364 /* Enter new state and set state variables. Note we use the time
1365 * of the last clock filter sample, which must be earlier than
1368 VERB4 bb_error_msg("disc_state=%d last update offset=%f recv_time=%f",
1369 disc_state, offset, recv_time);
1370 G.discipline_state = disc_state;
1371 G.last_update_offset = offset;
1372 G.last_update_recv_time = recv_time;
1374 /* Return: -1: decrease poll interval, 0: leave as is, 1: increase */
1376 update_local_clock(peer_t *p)
1380 /* Note: can use G.cluster_offset instead: */
1381 double offset = p->filter_offset;
1382 double recv_time = p->lastpkt_recv_time;
1384 #if !USING_KERNEL_PLL_LOOP
1387 #if !USING_KERNEL_PLL_LOOP || USING_INITIAL_FREQ_ESTIMATION
1388 double since_last_update;
1390 double etemp, dtemp;
1392 abs_offset = fabs(offset);
1395 /* If needed, -S script can do it by looking at $offset
1396 * env var and killing parent */
1397 /* If the offset is too large, give up and go home */
1398 if (abs_offset > PANIC_THRESHOLD) {
1399 bb_error_msg_and_die("offset %f far too big, exiting", offset);
1403 /* If this is an old update, for instance as the result
1404 * of a system peer change, avoid it. We never use
1405 * an old sample or the same sample twice.
1407 if (recv_time <= G.last_update_recv_time) {
1408 VERB3 bb_error_msg("update from %s: same or older datapoint, not using it",
1410 return 0; /* "leave poll interval as is" */
1413 /* Clock state machine transition function. This is where the
1414 * action is and defines how the system reacts to large time
1415 * and frequency errors.
1417 #if !USING_KERNEL_PLL_LOOP || USING_INITIAL_FREQ_ESTIMATION
1418 since_last_update = recv_time - G.reftime;
1420 #if !USING_KERNEL_PLL_LOOP
1423 #if USING_INITIAL_FREQ_ESTIMATION
1424 if (G.discipline_state == STATE_FREQ) {
1425 /* Ignore updates until the stepout threshold */
1426 if (since_last_update < WATCH_THRESHOLD) {
1427 VERB4 bb_error_msg("measuring drift, datapoint ignored, %f sec remains",
1428 WATCH_THRESHOLD - since_last_update);
1429 return 0; /* "leave poll interval as is" */
1431 # if !USING_KERNEL_PLL_LOOP
1432 freq_drift = (offset - G.last_update_offset) / since_last_update;
1437 /* There are two main regimes: when the
1438 * offset exceeds the step threshold and when it does not.
1440 if (abs_offset > STEP_THRESHOLD) {
1444 // This "spike state" seems to be useless, peer selection already drops
1445 // occassional "bad" datapoints. If we are here, there were _many_
1446 // large offsets. When a few first large offsets are seen,
1447 // we end up in "no valid datapoints, no peer selected" state.
1448 // Only when enough of them are seen (which means it's not a fluke),
1449 // we end up here. Looks like _our_ clock is off.
1450 switch (G.discipline_state) {
1452 /* The first outlyer: ignore it, switch to SPIK state */
1453 VERB3 bb_error_msg("update from %s: offset:%+f, spike%s",
1454 p->p_dotted, offset,
1456 G.discipline_state = STATE_SPIK;
1457 return -1; /* "decrease poll interval" */
1460 /* Ignore succeeding outlyers until either an inlyer
1461 * is found or the stepout threshold is exceeded.
1463 remains = WATCH_THRESHOLD - since_last_update;
1465 VERB3 bb_error_msg("update from %s: offset:%+f, spike%s",
1466 p->p_dotted, offset,
1467 ", datapoint ignored");
1468 return -1; /* "decrease poll interval" */
1470 /* fall through: we need to step */
1474 /* Step the time and clamp down the poll interval.
1476 * In NSET state an initial frequency correction is
1477 * not available, usually because the frequency file has
1478 * not yet been written. Since the time is outside the
1479 * capture range, the clock is stepped. The frequency
1480 * will be set directly following the stepout interval.
1482 * In FSET state the initial frequency has been set
1483 * from the frequency file. Since the time is outside
1484 * the capture range, the clock is stepped immediately,
1485 * rather than after the stepout interval. Guys get
1486 * nervous if it takes 17 minutes to set the clock for
1489 * In SPIK state the stepout threshold has expired and
1490 * the phase is still above the step threshold. Note
1491 * that a single spike greater than the step threshold
1492 * is always suppressed, even at the longer poll
1495 VERB4 bb_error_msg("stepping time by %+f; poll_exp=MINPOLL", offset);
1497 if (option_mask32 & OPT_q) {
1498 /* We were only asked to set time once. Done. */
1502 clamp_pollexp_and_set_MAXSTRAT();
1504 run_script("step", offset);
1506 recv_time += offset;
1508 #if USING_INITIAL_FREQ_ESTIMATION
1509 if (G.discipline_state == STATE_NSET) {
1510 set_new_values(STATE_FREQ, /*offset:*/ 0, recv_time);
1511 return 1; /* "ok to increase poll interval" */
1514 abs_offset = offset = 0;
1515 set_new_values(STATE_SYNC, offset, recv_time);
1516 } else { /* abs_offset <= STEP_THRESHOLD */
1518 /* The ratio is calculated before jitter is updated to make
1519 * poll adjust code more sensitive to large offsets.
1521 G.offset_to_jitter_ratio = abs_offset / G.discipline_jitter;
1523 /* Compute the clock jitter as the RMS of exponentially
1524 * weighted offset differences. Used by the poll adjust code.
1526 etemp = SQUARE(G.discipline_jitter);
1527 dtemp = SQUARE(offset - G.last_update_offset);
1528 G.discipline_jitter = SQRT(etemp + (dtemp - etemp) / AVG);
1529 if (G.discipline_jitter < G_precision_sec)
1530 G.discipline_jitter = G_precision_sec;
1532 switch (G.discipline_state) {
1534 if (option_mask32 & OPT_q) {
1535 /* We were only asked to set time once.
1536 * The clock is precise enough, no need to step.
1540 #if USING_INITIAL_FREQ_ESTIMATION
1541 /* This is the first update received and the frequency
1542 * has not been initialized. The first thing to do
1543 * is directly measure the oscillator frequency.
1545 set_new_values(STATE_FREQ, offset, recv_time);
1547 set_new_values(STATE_SYNC, offset, recv_time);
1549 VERB4 bb_error_msg("transitioning to FREQ, datapoint ignored");
1550 return 0; /* "leave poll interval as is" */
1552 #if 0 /* this is dead code for now */
1554 /* This is the first update and the frequency
1555 * has been initialized. Adjust the phase, but
1556 * don't adjust the frequency until the next update.
1558 set_new_values(STATE_SYNC, offset, recv_time);
1559 /* freq_drift remains 0 */
1563 #if USING_INITIAL_FREQ_ESTIMATION
1565 /* since_last_update >= WATCH_THRESHOLD, we waited enough.
1566 * Correct the phase and frequency and switch to SYNC state.
1567 * freq_drift was already estimated (see code above)
1569 set_new_values(STATE_SYNC, offset, recv_time);
1574 #if !USING_KERNEL_PLL_LOOP
1575 /* Compute freq_drift due to PLL and FLL contributions.
1577 * The FLL and PLL frequency gain constants
1578 * depend on the poll interval and Allan
1579 * intercept. The FLL is not used below one-half
1580 * the Allan intercept. Above that the loop gain
1581 * increases in steps to 1 / AVG.
1583 if ((1 << G.poll_exp) > ALLAN / 2) {
1584 etemp = FLL - G.poll_exp;
1587 freq_drift += (offset - G.last_update_offset) / (MAXD(since_last_update, ALLAN) * etemp);
1589 /* For the PLL the integration interval
1590 * (numerator) is the minimum of the update
1591 * interval and poll interval. This allows
1592 * oversampling, but not undersampling.
1594 etemp = MIND(since_last_update, (1 << G.poll_exp));
1595 dtemp = (4 * PLL) << G.poll_exp;
1596 freq_drift += offset * etemp / SQUARE(dtemp);
1598 set_new_values(STATE_SYNC, offset, recv_time);
1601 if (G.stratum != p->lastpkt_stratum + 1) {
1602 G.stratum = p->lastpkt_stratum + 1;
1603 run_script("stratum", offset);
1607 G.reftime = G.cur_time;
1608 G.ntp_status = p->lastpkt_status;
1609 G.refid = p->lastpkt_refid;
1610 G.rootdelay = p->lastpkt_rootdelay + p->lastpkt_delay;
1611 dtemp = p->filter_jitter; // SQRT(SQUARE(p->filter_jitter) + SQUARE(G.cluster_jitter));
1612 dtemp += MAXD(p->filter_dispersion + FREQ_TOLERANCE * (G.cur_time - p->lastpkt_recv_time) + abs_offset, MINDISP);
1613 G.rootdisp = p->lastpkt_rootdisp + dtemp;
1614 VERB4 bb_error_msg("updating leap/refid/reftime/rootdisp from peer %s", p->p_dotted);
1616 /* We are in STATE_SYNC now, but did not do adjtimex yet.
1617 * (Any other state does not reach this, they all return earlier)
1618 * By this time, freq_drift and offset are set
1619 * to values suitable for adjtimex.
1621 #if !USING_KERNEL_PLL_LOOP
1622 /* Calculate the new frequency drift and frequency stability (wander).
1623 * Compute the clock wander as the RMS of exponentially weighted
1624 * frequency differences. This is not used directly, but can,
1625 * along with the jitter, be a highly useful monitoring and
1628 dtemp = G.discipline_freq_drift + freq_drift;
1629 G.discipline_freq_drift = MAXD(MIND(MAXDRIFT, dtemp), -MAXDRIFT);
1630 etemp = SQUARE(G.discipline_wander);
1631 dtemp = SQUARE(dtemp);
1632 G.discipline_wander = SQRT(etemp + (dtemp - etemp) / AVG);
1634 VERB4 bb_error_msg("discipline freq_drift=%.9f(int:%ld corr:%e) wander=%f",
1635 G.discipline_freq_drift,
1636 (long)(G.discipline_freq_drift * 65536e6),
1638 G.discipline_wander);
1641 memset(&tmx, 0, sizeof(tmx));
1642 if (adjtimex(&tmx) < 0)
1643 bb_perror_msg_and_die("adjtimex");
1644 bb_error_msg("p adjtimex freq:%ld offset:%+ld status:0x%x tc:%ld",
1645 tmx.freq, tmx.offset, tmx.status, tmx.constant);
1648 memset(&tmx, 0, sizeof(tmx));
1650 //doesn't work, offset remains 0 (!) in kernel:
1651 //ntpd: set adjtimex freq:1786097 tmx.offset:77487
1652 //ntpd: prev adjtimex freq:1786097 tmx.offset:0
1653 //ntpd: cur adjtimex freq:1786097 tmx.offset:0
1654 tmx.modes = ADJ_FREQUENCY | ADJ_OFFSET;
1655 /* 65536 is one ppm */
1656 tmx.freq = G.discipline_freq_drift * 65536e6;
1658 tmx.modes = ADJ_OFFSET | ADJ_STATUS | ADJ_TIMECONST;// | ADJ_MAXERROR | ADJ_ESTERROR;
1659 tmx.constant = (int)G.poll_exp - 4;
1661 * The below if statement should be unnecessary, but...
1662 * It looks like Linux kernel's PLL is far too gentle in changing
1663 * tmx.freq in response to clock offset. Offset keeps growing
1664 * and eventually we fall back to smaller poll intervals.
1665 * We can make correction more agressive (about x2) by supplying
1666 * PLL time constant which is one less than the real one.
1667 * To be on a safe side, let's do it only if offset is significantly
1668 * larger than jitter.
1670 if (G.offset_to_jitter_ratio >= TIMECONST_HACK_GATE)
1672 tmx.offset = (long)(offset * 1000000); /* usec */
1673 if (SLEW_THRESHOLD < STEP_THRESHOLD) {
1674 if (tmx.offset > (long)(SLEW_THRESHOLD * 1000000)) {
1675 tmx.offset = (long)(SLEW_THRESHOLD * 1000000);
1678 if (tmx.offset < -(long)(SLEW_THRESHOLD * 1000000)) {
1679 tmx.offset = -(long)(SLEW_THRESHOLD * 1000000);
1683 if (tmx.constant < 0)
1686 tmx.status = STA_PLL;
1687 if (G.ntp_status & LI_PLUSSEC)
1688 tmx.status |= STA_INS;
1689 if (G.ntp_status & LI_MINUSSEC)
1690 tmx.status |= STA_DEL;
1692 //tmx.esterror = (uint32_t)(clock_jitter * 1e6);
1693 //tmx.maxerror = (uint32_t)((sys_rootdelay / 2 + sys_rootdisp) * 1e6);
1694 rc = adjtimex(&tmx);
1696 bb_perror_msg_and_die("adjtimex");
1697 /* NB: here kernel returns constant == G.poll_exp, not == G.poll_exp - 4.
1698 * Not sure why. Perhaps it is normal.
1700 VERB4 bb_error_msg("adjtimex:%d freq:%ld offset:%+ld status:0x%x",
1701 rc, tmx.freq, tmx.offset, tmx.status);
1702 G.kernel_freq_drift = tmx.freq / 65536;
1703 VERB2 bb_error_msg("update from:%s offset:%+f delay:%f jitter:%f clock drift:%+.3fppm tc:%d",
1707 G.discipline_jitter,
1708 (double)tmx.freq / 65536,
1712 return 1; /* "ok to increase poll interval" */
1717 * We've got a new reply packet from a peer, process it
1721 poll_interval(int upper_bound)
1723 unsigned interval, r, mask;
1724 interval = 1 << G.poll_exp;
1725 if (interval > upper_bound)
1726 interval = upper_bound;
1727 mask = ((interval-1) >> 4) | 1;
1729 interval += r & mask; /* ~ random(0..1) * interval/16 */
1730 VERB4 bb_error_msg("chose poll interval:%u (poll_exp:%d)", interval, G.poll_exp);
1734 adjust_poll(int count)
1736 G.polladj_count += count;
1737 if (G.polladj_count > POLLADJ_LIMIT) {
1738 G.polladj_count = 0;
1739 if (G.poll_exp < MAXPOLL) {
1741 VERB4 bb_error_msg("polladj: discipline_jitter:%f ++poll_exp=%d",
1742 G.discipline_jitter, G.poll_exp);
1744 } else if (G.polladj_count < -POLLADJ_LIMIT || (count < 0 && G.poll_exp > BIGPOLL)) {
1745 G.polladj_count = 0;
1746 if (G.poll_exp > MINPOLL) {
1750 /* Correct p->next_action_time in each peer
1751 * which waits for sending, so that they send earlier.
1752 * Old pp->next_action_time are on the order
1753 * of t + (1 << old_poll_exp) + small_random,
1754 * we simply need to subtract ~half of that.
1756 for (item = G.ntp_peers; item != NULL; item = item->link) {
1757 peer_t *pp = (peer_t *) item->data;
1759 pp->next_action_time -= (1 << G.poll_exp);
1761 VERB4 bb_error_msg("polladj: discipline_jitter:%f --poll_exp=%d",
1762 G.discipline_jitter, G.poll_exp);
1765 VERB4 bb_error_msg("polladj: count:%d", G.polladj_count);
1768 static NOINLINE void
1769 recv_and_process_peer_pkt(peer_t *p)
1774 double T1, T2, T3, T4;
1776 double prev_delay, delay;
1778 datapoint_t *datapoint;
1783 /* We can recvfrom here and check from.IP, but some multihomed
1784 * ntp servers reply from their *other IP*.
1785 * TODO: maybe we should check at least what we can: from.port == 123?
1788 size = recv(p->p_fd, &msg, sizeof(msg), MSG_DONTWAIT);
1793 if (errno == EAGAIN)
1794 /* There was no packet after all
1795 * (poll() returning POLLIN for a fd
1796 * is not a ironclad guarantee that data is there)
1800 * If you need a different handling for a specific
1801 * errno, always explain it in comment.
1803 bb_perror_msg_and_die("recv(%s) error", p->p_dotted);
1806 if (size != NTP_MSGSIZE_NOAUTH && size != NTP_MSGSIZE) {
1807 bb_error_msg("malformed packet received from %s", p->p_dotted);
1811 if (msg.m_orgtime.int_partl != p->p_xmt_msg.m_xmttime.int_partl
1812 || msg.m_orgtime.fractionl != p->p_xmt_msg.m_xmttime.fractionl
1814 /* Somebody else's packet */
1818 /* We do not expect any more packets from this peer for now.
1819 * Closing the socket informs kernel about it.
1820 * We open a new socket when we send a new query.
1825 if ((msg.m_status & LI_ALARM) == LI_ALARM
1826 || msg.m_stratum == 0
1827 || msg.m_stratum > NTP_MAXSTRATUM
1829 bb_error_msg("reply from %s: peer is unsynced", p->p_dotted);
1831 * Stratum 0 responses may have commands in 32-bit m_refid field:
1832 * "DENY", "RSTR" - peer does not like us at all,
1833 * "RATE" - peer is overloaded, reduce polling freq.
1834 * If poll interval is small, increase it.
1836 if (G.poll_exp < BIGPOLL)
1837 goto increase_interval;
1838 goto pick_normal_interval;
1841 // /* Verify valid root distance */
1842 // if (msg.m_rootdelay / 2 + msg.m_rootdisp >= MAXDISP || p->lastpkt_reftime > msg.m_xmt)
1843 // return; /* invalid header values */
1846 * From RFC 2030 (with a correction to the delay math):
1848 * Timestamp Name ID When Generated
1849 * ------------------------------------------------------------
1850 * Originate Timestamp T1 time request sent by client
1851 * Receive Timestamp T2 time request received by server
1852 * Transmit Timestamp T3 time reply sent by server
1853 * Destination Timestamp T4 time reply received by client
1855 * The roundtrip delay and local clock offset are defined as
1857 * delay = (T4 - T1) - (T3 - T2); offset = ((T2 - T1) + (T3 - T4)) / 2
1860 T2 = lfp_to_d(msg.m_rectime);
1861 T3 = lfp_to_d(msg.m_xmttime);
1864 /* The delay calculation is a special case. In cases where the
1865 * server and client clocks are running at different rates and
1866 * with very fast networks, the delay can appear negative. In
1867 * order to avoid violating the Principle of Least Astonishment,
1868 * the delay is clamped not less than the system precision.
1870 delay = (T4 - T1) - (T3 - T2);
1871 if (delay < G_precision_sec)
1872 delay = G_precision_sec;
1874 * If this packet's delay is much bigger than the last one,
1875 * it's better to just ignore it than use its much less precise value.
1877 prev_delay = p->p_raw_delay;
1878 p->p_raw_delay = delay;
1879 if (p->reachable_bits && delay > prev_delay * BAD_DELAY_GROWTH) {
1880 bb_error_msg("reply from %s: delay %f is too high, ignoring", p->p_dotted, delay);
1881 goto pick_normal_interval;
1884 p->lastpkt_delay = delay;
1885 p->lastpkt_recv_time = T4;
1886 VERB6 bb_error_msg("%s->lastpkt_recv_time=%f", p->p_dotted, p->lastpkt_recv_time);
1887 p->lastpkt_status = msg.m_status;
1888 p->lastpkt_stratum = msg.m_stratum;
1889 p->lastpkt_rootdelay = sfp_to_d(msg.m_rootdelay);
1890 p->lastpkt_rootdisp = sfp_to_d(msg.m_rootdisp);
1891 p->lastpkt_refid = msg.m_refid;
1893 p->datapoint_idx = p->reachable_bits ? (p->datapoint_idx + 1) % NUM_DATAPOINTS : 0;
1894 datapoint = &p->filter_datapoint[p->datapoint_idx];
1895 datapoint->d_recv_time = T4;
1896 datapoint->d_offset = offset = ((T2 - T1) + (T3 - T4)) / 2;
1897 datapoint->d_dispersion = LOG2D(msg.m_precision_exp) + G_precision_sec;
1898 if (!p->reachable_bits) {
1899 /* 1st datapoint ever - replicate offset in every element */
1901 for (i = 0; i < NUM_DATAPOINTS; i++) {
1902 p->filter_datapoint[i].d_offset = offset;
1906 p->reachable_bits |= 1;
1907 if ((MAX_VERBOSE && G.verbose) || (option_mask32 & OPT_w)) {
1908 bb_error_msg("reply from %s: offset:%+f delay:%f status:0x%02x strat:%d refid:0x%08x rootdelay:%f reach:0x%02x",
1915 p->lastpkt_rootdelay,
1917 /* not shown: m_ppoll, m_precision_exp, m_rootdisp,
1918 * m_reftime, m_orgtime, m_rectime, m_xmttime
1923 /* Muck with statictics and update the clock */
1924 filter_datapoints(p);
1925 q = select_and_cluster();
1928 if (!(option_mask32 & OPT_w)) {
1929 rc = update_local_clock(q);
1931 //Disabled this because there is a case where largish offsets
1932 //are unavoidable: if network round-trip delay is, say, ~0.6s,
1933 //error in offset estimation would be ~delay/2 ~= 0.3s.
1934 //Thus, offsets will be usually in -0.3...0.3s range.
1935 //In this case, this code would keep poll interval small,
1936 //but it won't be helping.
1937 //BIGOFF check below deals with a case of seeing multi-second offsets.
1939 /* If drift is dangerously large, immediately
1940 * drop poll interval one step down.
1942 if (fabs(q->filter_offset) >= POLLDOWN_OFFSET) {
1943 VERB4 bb_error_msg("offset:%+f > POLLDOWN_OFFSET", q->filter_offset);
1944 adjust_poll(-POLLADJ_LIMIT * 3);
1950 /* No peer selected.
1951 * If poll interval is small, increase it.
1953 if (G.poll_exp < BIGPOLL)
1954 goto increase_interval;
1958 /* Adjust the poll interval by comparing the current offset
1959 * with the clock jitter. If the offset is less than
1960 * the clock jitter times a constant, then the averaging interval
1961 * is increased, otherwise it is decreased. A bit of hysteresis
1962 * helps calm the dance. Works best using burst mode.
1964 if (rc > 0 && G.offset_to_jitter_ratio <= POLLADJ_GATE) {
1965 /* was += G.poll_exp but it is a bit
1966 * too optimistic for my taste at high poll_exp's */
1968 adjust_poll(MINPOLL);
1971 bb_error_msg("want smaller interval: offset/jitter = %u",
1972 G.offset_to_jitter_ratio);
1973 adjust_poll(-G.poll_exp * 2);
1977 /* Decide when to send new query for this peer */
1978 pick_normal_interval:
1979 interval = poll_interval(INT_MAX);
1980 if (fabs(offset) >= BIGOFF && interval > BIGOFF_INTERVAL) {
1981 /* If we are synced, offsets are less than SLEW_THRESHOLD,
1982 * or at the very least not much larger than it.
1983 * Now we see a largish one.
1984 * Either this peer is feeling bad, or packet got corrupted,
1985 * or _our_ clock is wrong now and _all_ peers will show similar
1986 * largish offsets too.
1987 * I observed this with laptop suspend stopping clock.
1988 * In any case, it makes sense to make next request soonish:
1989 * cases 1 and 2: get a better datapoint,
1990 * case 3: allows to resync faster.
1992 interval = BIGOFF_INTERVAL;
1995 set_next(p, interval);
1998 #if ENABLE_FEATURE_NTPD_SERVER
1999 static NOINLINE void
2000 recv_and_process_client_pkt(void /*int fd*/)
2004 len_and_sockaddr *to;
2005 struct sockaddr *from;
2007 uint8_t query_status;
2008 l_fixedpt_t query_xmttime;
2010 to = get_sock_lsa(G_listen_fd);
2011 from = xzalloc(to->len);
2013 size = recv_from_to(G_listen_fd, &msg, sizeof(msg), MSG_DONTWAIT, from, &to->u.sa, to->len);
2014 if (size != NTP_MSGSIZE_NOAUTH && size != NTP_MSGSIZE) {
2017 if (errno == EAGAIN)
2019 bb_perror_msg_and_die("recv");
2021 addr = xmalloc_sockaddr2dotted_noport(from);
2022 bb_error_msg("malformed packet received from %s: size %u", addr, (int)size);
2027 query_status = msg.m_status;
2028 query_xmttime = msg.m_xmttime;
2030 /* Build a reply packet */
2031 memset(&msg, 0, sizeof(msg));
2032 msg.m_status = G.stratum < MAXSTRAT ? (G.ntp_status & LI_MASK) : LI_ALARM;
2033 msg.m_status |= (query_status & VERSION_MASK);
2034 msg.m_status |= ((query_status & MODE_MASK) == MODE_CLIENT) ?
2035 MODE_SERVER : MODE_SYM_PAS;
2036 msg.m_stratum = G.stratum;
2037 msg.m_ppoll = G.poll_exp;
2038 msg.m_precision_exp = G_precision_exp;
2039 /* this time was obtained between poll() and recv() */
2040 msg.m_rectime = d_to_lfp(G.cur_time);
2041 msg.m_xmttime = d_to_lfp(gettime1900d()); /* this instant */
2042 if (G.peer_cnt == 0) {
2043 /* we have no peers: "stratum 1 server" mode. reftime = our own time */
2044 G.reftime = G.cur_time;
2046 msg.m_reftime = d_to_lfp(G.reftime);
2047 msg.m_orgtime = query_xmttime;
2048 msg.m_rootdelay = d_to_sfp(G.rootdelay);
2049 //simple code does not do this, fix simple code!
2050 msg.m_rootdisp = d_to_sfp(G.rootdisp);
2051 //version = (query_status & VERSION_MASK); /* ... >> VERSION_SHIFT - done below instead */
2052 msg.m_refid = G.refid; // (version > (3 << VERSION_SHIFT)) ? G.refid : G.refid3;
2054 /* We reply from the local address packet was sent to,
2055 * this makes to/from look swapped here: */
2056 do_sendto(G_listen_fd,
2057 /*from:*/ &to->u.sa, /*to:*/ from, /*addrlen:*/ to->len,
2066 /* Upstream ntpd's options:
2068 * -4 Force DNS resolution of host names to the IPv4 namespace.
2069 * -6 Force DNS resolution of host names to the IPv6 namespace.
2070 * -a Require cryptographic authentication for broadcast client,
2071 * multicast client and symmetric passive associations.
2072 * This is the default.
2073 * -A Do not require cryptographic authentication for broadcast client,
2074 * multicast client and symmetric passive associations.
2075 * This is almost never a good idea.
2076 * -b Enable the client to synchronize to broadcast servers.
2078 * Specify the name and path of the configuration file,
2079 * default /etc/ntp.conf
2080 * -d Specify debugging mode. This option may occur more than once,
2081 * with each occurrence indicating greater detail of display.
2083 * Specify debugging level directly.
2085 * Specify the name and path of the frequency file.
2086 * This is the same operation as the "driftfile FILE"
2087 * configuration command.
2088 * -g Normally, ntpd exits with a message to the system log
2089 * if the offset exceeds the panic threshold, which is 1000 s
2090 * by default. This option allows the time to be set to any value
2091 * without restriction; however, this can happen only once.
2092 * If the threshold is exceeded after that, ntpd will exit
2093 * with a message to the system log. This option can be used
2094 * with the -q and -x options. See the tinker command for other options.
2096 * Chroot the server to the directory jaildir. This option also implies
2097 * that the server attempts to drop root privileges at startup
2098 * (otherwise, chroot gives very little additional security).
2099 * You may need to also specify a -u option.
2101 * Specify the name and path of the symmetric key file,
2102 * default /etc/ntp/keys. This is the same operation
2103 * as the "keys FILE" configuration command.
2105 * Specify the name and path of the log file. The default
2106 * is the system log file. This is the same operation as
2107 * the "logfile FILE" configuration command.
2108 * -L Do not listen to virtual IPs. The default is to listen.
2110 * -N To the extent permitted by the operating system,
2111 * run the ntpd at the highest priority.
2113 * Specify the name and path of the file used to record the ntpd
2114 * process ID. This is the same operation as the "pidfile FILE"
2115 * configuration command.
2117 * To the extent permitted by the operating system,
2118 * run the ntpd at the specified priority.
2119 * -q Exit the ntpd just after the first time the clock is set.
2120 * This behavior mimics that of the ntpdate program, which is
2121 * to be retired. The -g and -x options can be used with this option.
2122 * Note: The kernel time discipline is disabled with this option.
2124 * Specify the default propagation delay from the broadcast/multicast
2125 * server to this client. This is necessary only if the delay
2126 * cannot be computed automatically by the protocol.
2128 * Specify the directory path for files created by the statistics
2129 * facility. This is the same operation as the "statsdir DIR"
2130 * configuration command.
2132 * Add a key number to the trusted key list. This option can occur
2135 * Specify a user, and optionally a group, to switch to.
2138 * Add a system variable listed by default.
2139 * -x Normally, the time is slewed if the offset is less than the step
2140 * threshold, which is 128 ms by default, and stepped if above
2141 * the threshold. This option sets the threshold to 600 s, which is
2142 * well within the accuracy window to set the clock manually.
2143 * Note: since the slew rate of typical Unix kernels is limited
2144 * to 0.5 ms/s, each second of adjustment requires an amortization
2145 * interval of 2000 s. Thus, an adjustment as much as 600 s
2146 * will take almost 14 days to complete. This option can be used
2147 * with the -g and -q options. See the tinker command for other options.
2148 * Note: The kernel time discipline is disabled with this option.
2151 /* By doing init in a separate function we decrease stack usage
2154 static NOINLINE void ntp_init(char **argv)
2162 bb_error_msg_and_die(bb_msg_you_must_be_root);
2164 /* Set some globals */
2165 G.discipline_jitter = G_precision_sec;
2166 G.stratum = MAXSTRAT;
2168 G.poll_exp = BURSTPOLL; /* speeds up initial sync */
2169 G.last_script_run = G.reftime = G.last_update_recv_time = gettime1900d(); /* sets G.cur_time too */
2173 opt_complementary = "dd:p::wn" /* -d: counter; -p: list; -w implies -n */
2174 IF_FEATURE_NTPD_SERVER(":Il"); /* -I implies -l */
2175 opts = getopt32(argv,
2177 "wp:S:"IF_FEATURE_NTPD_SERVER("l") /* NOT compat */
2178 IF_FEATURE_NTPD_SERVER("I:") /* compat */
2180 "46aAbgL", /* compat, ignored */
2181 &peers,&G.script_name,
2182 #if ENABLE_FEATURE_NTPD_SERVER
2187 // if (opts & OPT_x) /* disable stepping, only slew is allowed */
2188 // G.time_was_stepped = 1;
2191 add_peers(llist_pop(&peers));
2193 #if ENABLE_FEATURE_NTPD_CONF
2198 parser = config_open("/etc/ntp.conf");
2199 while (config_read(parser, token, 3, 1, "# \t", PARSE_NORMAL)) {
2200 if (strcmp(token[0], "server") == 0 && token[1]) {
2201 add_peers(token[1]);
2204 bb_error_msg("skipping %s:%u: unimplemented command '%s'",
2205 "/etc/ntp.conf", parser->lineno, token[0]
2208 config_close(parser);
2211 if (G.peer_cnt == 0) {
2212 if (!(opts & OPT_l))
2214 /* -l but no peers: "stratum 1 server" mode */
2217 #if ENABLE_FEATURE_NTPD_SERVER
2220 G_listen_fd = create_and_bind_dgram_or_die(NULL, 123);
2222 if (setsockopt_bindtodevice(G_listen_fd, G.if_name))
2225 socket_want_pktinfo(G_listen_fd);
2226 setsockopt_int(G_listen_fd, IPPROTO_IP, IP_TOS, IPTOS_LOWDELAY);
2229 if (!(opts & OPT_n)) {
2230 bb_daemonize_or_rexec(DAEMON_DEVNULL_STDIO, argv);
2231 logmode = LOGMODE_NONE;
2233 /* I hesitate to set -20 prio. -15 should be high enough for timekeeping */
2235 setpriority(PRIO_PROCESS, 0, -15);
2237 /* If network is up, syncronization occurs in ~10 seconds.
2238 * We give "ntpd -q" 10 seconds to get first reply,
2239 * then another 50 seconds to finish syncing.
2241 * I tested ntpd 4.2.6p1 and apparently it never exits
2242 * (will try forever), but it does not feel right.
2243 * The goal of -q is to act like ntpdate: set time
2244 * after a reasonably small period of polling, or fail.
2247 option_mask32 |= OPT_qq;
2264 int ntpd_main(int argc UNUSED_PARAM, char **argv) MAIN_EXTERNALLY_VISIBLE;
2265 int ntpd_main(int argc UNUSED_PARAM, char **argv)
2273 memset(&G, 0, sizeof(G));
2274 SET_PTR_TO_GLOBALS(&G);
2278 /* If ENABLE_FEATURE_NTPD_SERVER, + 1 for listen_fd: */
2279 cnt = G.peer_cnt + ENABLE_FEATURE_NTPD_SERVER;
2280 idx2peer = xzalloc(sizeof(idx2peer[0]) * cnt);
2281 pfd = xzalloc(sizeof(pfd[0]) * cnt);
2283 /* Countdown: we never sync before we sent INITIAL_SAMPLES+1
2284 * packets to each peer.
2285 * NB: if some peer is not responding, we may end up sending
2286 * fewer packets to it and more to other peers.
2287 * NB2: sync usually happens using INITIAL_SAMPLES packets,
2288 * since last reply does not come back instantaneously.
2290 cnt = G.peer_cnt * (INITIAL_SAMPLES + 1);
2292 write_pidfile(CONFIG_PID_FILE_PATH "/ntpd.pid");
2294 while (!bb_got_signal) {
2300 /* Nothing between here and poll() blocks for any significant time */
2302 nextaction = G.cur_time + 3600;
2305 #if ENABLE_FEATURE_NTPD_SERVER
2306 if (G_listen_fd != -1) {
2307 pfd[0].fd = G_listen_fd;
2308 pfd[0].events = POLLIN;
2312 /* Pass over peer list, send requests, time out on receives */
2313 for (item = G.ntp_peers; item != NULL; item = item->link) {
2314 peer_t *p = (peer_t *) item->data;
2316 if (p->next_action_time <= G.cur_time) {
2317 if (p->p_fd == -1) {
2318 /* Time to send new req */
2320 VERB4 bb_error_msg("disabling burst mode");
2321 G.polladj_count = 0;
2322 G.poll_exp = MINPOLL;
2324 send_query_to_peer(p);
2326 /* Timed out waiting for reply */
2329 /* If poll interval is small, increase it */
2330 if (G.poll_exp < BIGPOLL)
2331 adjust_poll(MINPOLL);
2332 timeout = poll_interval(NOREPLY_INTERVAL);
2333 bb_error_msg("timed out waiting for %s, reach 0x%02x, next query in %us",
2334 p->p_dotted, p->reachable_bits, timeout);
2335 set_next(p, timeout);
2339 if (p->next_action_time < nextaction)
2340 nextaction = p->next_action_time;
2343 /* Wait for reply from this peer */
2344 pfd[i].fd = p->p_fd;
2345 pfd[i].events = POLLIN;
2351 timeout = nextaction - G.cur_time;
2354 timeout++; /* (nextaction - G.cur_time) rounds down, compensating */
2356 /* Here we may block */
2358 if (i > (ENABLE_FEATURE_NTPD_SERVER && G_listen_fd != -1)) {
2359 /* We wait for at least one reply.
2360 * Poll for it, without wasting time for message.
2361 * Since replies often come under 1 second, this also
2362 * reduces clutter in logs.
2364 nfds = poll(pfd, i, 1000);
2370 bb_error_msg("poll:%us sockets:%u interval:%us", timeout, i, 1 << G.poll_exp);
2372 nfds = poll(pfd, i, timeout * 1000);
2374 gettime1900d(); /* sets G.cur_time */
2376 if (!bb_got_signal /* poll wasn't interrupted by a signal */
2377 && G.cur_time - G.last_script_run > 11*60
2379 /* Useful for updating battery-backed RTC and such */
2380 run_script("periodic", G.last_update_offset);
2381 gettime1900d(); /* sets G.cur_time */
2386 /* Process any received packets */
2388 #if ENABLE_FEATURE_NTPD_SERVER
2389 if (G.listen_fd != -1) {
2390 if (pfd[0].revents /* & (POLLIN|POLLERR)*/) {
2392 recv_and_process_client_pkt(/*G.listen_fd*/);
2393 gettime1900d(); /* sets G.cur_time */
2398 for (; nfds != 0 && j < i; j++) {
2399 if (pfd[j].revents /* & (POLLIN|POLLERR)*/) {
2401 * At init, alarm was set to 10 sec.
2402 * Now we did get a reply.
2403 * Increase timeout to 50 seconds to finish syncing.
2405 if (option_mask32 & OPT_qq) {
2406 option_mask32 &= ~OPT_qq;
2410 recv_and_process_peer_pkt(idx2peer[j]);
2411 gettime1900d(); /* sets G.cur_time */
2416 if (G.ntp_peers && G.stratum != MAXSTRAT) {
2417 for (item = G.ntp_peers; item != NULL; item = item->link) {
2418 peer_t *p = (peer_t *) item->data;
2419 if (p->reachable_bits)
2420 goto have_reachable_peer;
2422 /* No peer responded for last 8 packets, panic */
2423 clamp_pollexp_and_set_MAXSTRAT();
2424 run_script("unsync", 0.0);
2425 have_reachable_peer: ;
2427 } /* while (!bb_got_signal) */
2429 remove_pidfile(CONFIG_PID_FILE_PATH "/ntpd.pid");
2430 kill_myself_with_sig(bb_got_signal);
2438 /*** openntpd-4.6 uses only adjtime, not adjtimex ***/
2440 /*** ntp-4.2.6/ntpd/ntp_loopfilter.c - adjtimex usage ***/
2444 direct_freq(double fp_offset)
2448 * If the kernel is enabled, we need the residual offset to
2449 * calculate the frequency correction.
2451 if (pll_control && kern_enable) {
2452 memset(&ntv, 0, sizeof(ntv));
2455 clock_offset = ntv.offset / 1e9;
2456 #else /* STA_NANO */
2457 clock_offset = ntv.offset / 1e6;
2458 #endif /* STA_NANO */
2459 drift_comp = FREQTOD(ntv.freq);
2461 #endif /* KERNEL_PLL */
2462 set_freq((fp_offset - clock_offset) / (current_time - clock_epoch) + drift_comp);
2468 set_freq(double freq) /* frequency update */
2476 * If the kernel is enabled, update the kernel frequency.
2478 if (pll_control && kern_enable) {
2479 memset(&ntv, 0, sizeof(ntv));
2480 ntv.modes = MOD_FREQUENCY;
2481 ntv.freq = DTOFREQ(drift_comp);
2483 snprintf(tbuf, sizeof(tbuf), "kernel %.3f PPM", drift_comp * 1e6);
2484 report_event(EVNT_FSET, NULL, tbuf);
2486 snprintf(tbuf, sizeof(tbuf), "ntpd %.3f PPM", drift_comp * 1e6);
2487 report_event(EVNT_FSET, NULL, tbuf);
2489 #else /* KERNEL_PLL */
2490 snprintf(tbuf, sizeof(tbuf), "ntpd %.3f PPM", drift_comp * 1e6);
2491 report_event(EVNT_FSET, NULL, tbuf);
2492 #endif /* KERNEL_PLL */
2501 * This code segment works when clock adjustments are made using
2502 * precision time kernel support and the ntp_adjtime() system
2503 * call. This support is available in Solaris 2.6 and later,
2504 * Digital Unix 4.0 and later, FreeBSD, Linux and specially
2505 * modified kernels for HP-UX 9 and Ultrix 4. In the case of the
2506 * DECstation 5000/240 and Alpha AXP, additional kernel
2507 * modifications provide a true microsecond clock and nanosecond
2508 * clock, respectively.
2510 * Important note: The kernel discipline is used only if the
2511 * step threshold is less than 0.5 s, as anything higher can
2512 * lead to overflow problems. This might occur if some misguided
2513 * lad set the step threshold to something ridiculous.
2515 if (pll_control && kern_enable) {
2517 #define MOD_BITS (MOD_OFFSET | MOD_MAXERROR | MOD_ESTERROR | MOD_STATUS | MOD_TIMECONST)
2520 * We initialize the structure for the ntp_adjtime()
2521 * system call. We have to convert everything to
2522 * microseconds or nanoseconds first. Do not update the
2523 * system variables if the ext_enable flag is set. In
2524 * this case, the external clock driver will update the
2525 * variables, which will be read later by the local
2526 * clock driver. Afterwards, remember the time and
2527 * frequency offsets for jitter and stability values and
2528 * to update the frequency file.
2530 memset(&ntv, 0, sizeof(ntv));
2532 ntv.modes = MOD_STATUS;
2535 ntv.modes = MOD_BITS | MOD_NANO;
2536 #else /* STA_NANO */
2537 ntv.modes = MOD_BITS;
2538 #endif /* STA_NANO */
2539 if (clock_offset < 0)
2544 ntv.offset = (int32)(clock_offset * 1e9 + dtemp);
2545 ntv.constant = sys_poll;
2546 #else /* STA_NANO */
2547 ntv.offset = (int32)(clock_offset * 1e6 + dtemp);
2548 ntv.constant = sys_poll - 4;
2549 #endif /* STA_NANO */
2550 ntv.esterror = (u_int32)(clock_jitter * 1e6);
2551 ntv.maxerror = (u_int32)((sys_rootdelay / 2 + sys_rootdisp) * 1e6);
2552 ntv.status = STA_PLL;
2555 * Enable/disable the PPS if requested.
2558 if (!(pll_status & STA_PPSTIME))
2559 report_event(EVNT_KERN,
2560 NULL, "PPS enabled");
2561 ntv.status |= STA_PPSTIME | STA_PPSFREQ;
2563 if (pll_status & STA_PPSTIME)
2564 report_event(EVNT_KERN,
2565 NULL, "PPS disabled");
2566 ntv.status &= ~(STA_PPSTIME | STA_PPSFREQ);
2568 if (sys_leap == LEAP_ADDSECOND)
2569 ntv.status |= STA_INS;
2570 else if (sys_leap == LEAP_DELSECOND)
2571 ntv.status |= STA_DEL;
2575 * Pass the stuff to the kernel. If it squeals, turn off
2576 * the pps. In any case, fetch the kernel offset,
2577 * frequency and jitter.
2579 if (ntp_adjtime(&ntv) == TIME_ERROR) {
2580 if (!(ntv.status & STA_PPSSIGNAL))
2581 report_event(EVNT_KERN, NULL,
2584 pll_status = ntv.status;
2586 clock_offset = ntv.offset / 1e9;
2587 #else /* STA_NANO */
2588 clock_offset = ntv.offset / 1e6;
2589 #endif /* STA_NANO */
2590 clock_frequency = FREQTOD(ntv.freq);
2593 * If the kernel PPS is lit, monitor its performance.
2595 if (ntv.status & STA_PPSTIME) {
2597 clock_jitter = ntv.jitter / 1e9;
2598 #else /* STA_NANO */
2599 clock_jitter = ntv.jitter / 1e6;
2600 #endif /* STA_NANO */
2603 #if defined(STA_NANO) && NTP_API == 4
2605 * If the TAI changes, update the kernel TAI.
2607 if (loop_tai != sys_tai) {
2609 ntv.modes = MOD_TAI;
2610 ntv.constant = sys_tai;
2613 #endif /* STA_NANO */
2615 #endif /* KERNEL_PLL */