Contents of /tags/mkinitrd-6_2_2/busybox/networking/ntpd.c
Parent Directory | Revision Log
Revision 1005 -
(show annotations)
(download)
Sun May 30 12:33:03 2010 UTC (14 years, 4 months ago) by niro
File MIME type: text/plain
File size: 71111 byte(s)
Sun May 30 12:33:03 2010 UTC (14 years, 4 months ago) by niro
File MIME type: text/plain
File size: 71111 byte(s)
tagged 'mkinitrd-6_2_2'
1 | /* |
2 | * NTP client/server, based on OpenNTPD 3.9p1 |
3 | * |
4 | * Author: Adam Tkac <vonsch@gmail.com> |
5 | * |
6 | * Licensed under GPLv2, see file LICENSE in this tarball for details. |
7 | * |
8 | * Parts of OpenNTPD clock syncronization code is replaced by |
9 | * code which is based on ntp-4.2.6, whuch carries the following |
10 | * copyright notice: |
11 | * |
12 | *********************************************************************** |
13 | * * |
14 | * Copyright (c) University of Delaware 1992-2009 * |
15 | * * |
16 | * Permission to use, copy, modify, and distribute this software and * |
17 | * its documentation for any purpose with or without fee is hereby * |
18 | * granted, provided that the above copyright notice appears in all * |
19 | * copies and that both the copyright notice and this permission * |
20 | * notice appear in supporting documentation, and that the name * |
21 | * University of Delaware not be used in advertising or publicity * |
22 | * pertaining to distribution of the software without specific, * |
23 | * written prior permission. The University of Delaware makes no * |
24 | * representations about the suitability this software for any * |
25 | * purpose. It is provided "as is" without express or implied * |
26 | * warranty. * |
27 | * * |
28 | *********************************************************************** |
29 | */ |
30 | #include "libbb.h" |
31 | #include <math.h> |
32 | #include <netinet/ip.h> /* For IPTOS_LOWDELAY definition */ |
33 | #include <sys/timex.h> |
34 | #ifndef IPTOS_LOWDELAY |
35 | # define IPTOS_LOWDELAY 0x10 |
36 | #endif |
37 | #ifndef IP_PKTINFO |
38 | # error "Sorry, your kernel has to support IP_PKTINFO" |
39 | #endif |
40 | |
41 | |
42 | /* Verbosity control (max level of -dddd options accepted). |
43 | * max 5 is very talkative (and bloated). 2 is non-bloated, |
44 | * production level setting. |
45 | */ |
46 | #define MAX_VERBOSE 2 |
47 | |
48 | |
49 | /* High-level description of the algorithm: |
50 | * |
51 | * We start running with very small poll_exp, BURSTPOLL, |
52 | * in order to quickly accumulate INITIAL_SAMLPES datapoints |
53 | * for each peer. Then, time is stepped if the offset is larger |
54 | * than STEP_THRESHOLD, otherwise it isn't; anyway, we enlarge |
55 | * poll_exp to MINPOLL and enter frequency measurement step: |
56 | * we collect new datapoints but ignore them for WATCH_THRESHOLD |
57 | * seconds. After WATCH_THRESHOLD seconds we look at accumulated |
58 | * offset and estimate frequency drift. |
59 | * |
60 | * (frequency measurement step seems to not be strictly needed, |
61 | * it is conditionally disabled with USING_INITIAL_FREQ_ESTIMATION |
62 | * define set to 0) |
63 | * |
64 | * After this, we enter "steady state": we collect a datapoint, |
65 | * we select the best peer, if this datapoint is not a new one |
66 | * (IOW: if this datapoint isn't for selected peer), sleep |
67 | * and collect another one; otherwise, use its offset to update |
68 | * frequency drift, if offset is somewhat large, reduce poll_exp, |
69 | * otherwise increase poll_exp. |
70 | * |
71 | * If offset is larger than STEP_THRESHOLD, which shouldn't normally |
72 | * happen, we assume that something "bad" happened (computer |
73 | * was hibernated, someone set totally wrong date, etc), |
74 | * then the time is stepped, all datapoints are discarded, |
75 | * and we go back to steady state. |
76 | */ |
77 | |
78 | #define RETRY_INTERVAL 5 /* on error, retry in N secs */ |
79 | #define RESPONSE_INTERVAL 15 /* wait for reply up to N secs */ |
80 | #define INITIAL_SAMLPES 4 /* how many samples do we want for init */ |
81 | |
82 | /* Clock discipline parameters and constants */ |
83 | |
84 | /* Step threshold (sec). std ntpd uses 0.128. |
85 | * Using exact power of 2 (1/8) results in smaller code */ |
86 | #define STEP_THRESHOLD 0.125 |
87 | #define WATCH_THRESHOLD 128 /* stepout threshold (sec). std ntpd uses 900 (11 mins (!)) */ |
88 | /* NB: set WATCH_THRESHOLD to ~60 when debugging to save time) */ |
89 | //UNUSED: #define PANIC_THRESHOLD 1000 /* panic threshold (sec) */ |
90 | |
91 | #define FREQ_TOLERANCE 0.000015 /* frequency tolerance (15 PPM) */ |
92 | #define BURSTPOLL 0 /* initial poll */ |
93 | #define MINPOLL 5 /* minimum poll interval. std ntpd uses 6 (6: 64 sec) */ |
94 | #define BIGPOLL 10 /* drop to lower poll at any trouble (10: 17 min) */ |
95 | #define MAXPOLL 12 /* maximum poll interval (12: 1.1h, 17: 36.4h). std ntpd uses 17 */ |
96 | /* Actively lower poll when we see such big offsets. |
97 | * With STEP_THRESHOLD = 0.125, it means we try to sync more aggressively |
98 | * if offset increases over 0.03 sec */ |
99 | #define POLLDOWN_OFFSET (STEP_THRESHOLD / 4) |
100 | #define MINDISP 0.01 /* minimum dispersion (sec) */ |
101 | #define MAXDISP 16 /* maximum dispersion (sec) */ |
102 | #define MAXSTRAT 16 /* maximum stratum (infinity metric) */ |
103 | #define MAXDIST 1 /* distance threshold (sec) */ |
104 | #define MIN_SELECTED 1 /* minimum intersection survivors */ |
105 | #define MIN_CLUSTERED 3 /* minimum cluster survivors */ |
106 | |
107 | #define MAXDRIFT 0.000500 /* frequency drift we can correct (500 PPM) */ |
108 | |
109 | /* Poll-adjust threshold. |
110 | * When we see that offset is small enough compared to discipline jitter, |
111 | * we grow a counter: += MINPOLL. When it goes over POLLADJ_LIMIT, |
112 | * we poll_exp++. If offset isn't small, counter -= poll_exp*2, |
113 | * and when it goes below -POLLADJ_LIMIT, we poll_exp-- |
114 | * (bumped from 30 to 36 since otherwise I often see poll_exp going *2* steps down) |
115 | */ |
116 | #define POLLADJ_LIMIT 36 |
117 | /* If offset < POLLADJ_GATE * discipline_jitter, then we can increase |
118 | * poll interval (we think we can't improve timekeeping |
119 | * by staying at smaller poll). |
120 | */ |
121 | #define POLLADJ_GATE 4 |
122 | /* Compromise Allan intercept (sec). doc uses 1500, std ntpd uses 512 */ |
123 | #define ALLAN 512 |
124 | /* PLL loop gain */ |
125 | #define PLL 65536 |
126 | /* FLL loop gain [why it depends on MAXPOLL??] */ |
127 | #define FLL (MAXPOLL + 1) |
128 | /* Parameter averaging constant */ |
129 | #define AVG 4 |
130 | |
131 | |
132 | enum { |
133 | NTP_VERSION = 4, |
134 | NTP_MAXSTRATUM = 15, |
135 | |
136 | NTP_DIGESTSIZE = 16, |
137 | NTP_MSGSIZE_NOAUTH = 48, |
138 | NTP_MSGSIZE = (NTP_MSGSIZE_NOAUTH + 4 + NTP_DIGESTSIZE), |
139 | |
140 | /* Status Masks */ |
141 | MODE_MASK = (7 << 0), |
142 | VERSION_MASK = (7 << 3), |
143 | VERSION_SHIFT = 3, |
144 | LI_MASK = (3 << 6), |
145 | |
146 | /* Leap Second Codes (high order two bits of m_status) */ |
147 | LI_NOWARNING = (0 << 6), /* no warning */ |
148 | LI_PLUSSEC = (1 << 6), /* add a second (61 seconds) */ |
149 | LI_MINUSSEC = (2 << 6), /* minus a second (59 seconds) */ |
150 | LI_ALARM = (3 << 6), /* alarm condition */ |
151 | |
152 | /* Mode values */ |
153 | MODE_RES0 = 0, /* reserved */ |
154 | MODE_SYM_ACT = 1, /* symmetric active */ |
155 | MODE_SYM_PAS = 2, /* symmetric passive */ |
156 | MODE_CLIENT = 3, /* client */ |
157 | MODE_SERVER = 4, /* server */ |
158 | MODE_BROADCAST = 5, /* broadcast */ |
159 | MODE_RES1 = 6, /* reserved for NTP control message */ |
160 | MODE_RES2 = 7, /* reserved for private use */ |
161 | }; |
162 | |
163 | //TODO: better base selection |
164 | #define OFFSET_1900_1970 2208988800UL /* 1970 - 1900 in seconds */ |
165 | |
166 | #define NUM_DATAPOINTS 8 |
167 | |
168 | typedef struct { |
169 | uint32_t int_partl; |
170 | uint32_t fractionl; |
171 | } l_fixedpt_t; |
172 | |
173 | typedef struct { |
174 | uint16_t int_parts; |
175 | uint16_t fractions; |
176 | } s_fixedpt_t; |
177 | |
178 | typedef struct { |
179 | uint8_t m_status; /* status of local clock and leap info */ |
180 | uint8_t m_stratum; |
181 | uint8_t m_ppoll; /* poll value */ |
182 | int8_t m_precision_exp; |
183 | s_fixedpt_t m_rootdelay; |
184 | s_fixedpt_t m_rootdisp; |
185 | uint32_t m_refid; |
186 | l_fixedpt_t m_reftime; |
187 | l_fixedpt_t m_orgtime; |
188 | l_fixedpt_t m_rectime; |
189 | l_fixedpt_t m_xmttime; |
190 | uint32_t m_keyid; |
191 | uint8_t m_digest[NTP_DIGESTSIZE]; |
192 | } msg_t; |
193 | |
194 | typedef struct { |
195 | double d_recv_time; |
196 | double d_offset; |
197 | double d_dispersion; |
198 | } datapoint_t; |
199 | |
200 | typedef struct { |
201 | len_and_sockaddr *p_lsa; |
202 | char *p_dotted; |
203 | /* when to send new query (if p_fd == -1) |
204 | * or when receive times out (if p_fd >= 0): */ |
205 | int p_fd; |
206 | int datapoint_idx; |
207 | uint32_t lastpkt_refid; |
208 | uint8_t lastpkt_status; |
209 | uint8_t lastpkt_stratum; |
210 | uint8_t reachable_bits; |
211 | double next_action_time; |
212 | double p_xmttime; |
213 | double lastpkt_recv_time; |
214 | double lastpkt_delay; |
215 | double lastpkt_rootdelay; |
216 | double lastpkt_rootdisp; |
217 | /* produced by filter algorithm: */ |
218 | double filter_offset; |
219 | double filter_dispersion; |
220 | double filter_jitter; |
221 | datapoint_t filter_datapoint[NUM_DATAPOINTS]; |
222 | /* last sent packet: */ |
223 | msg_t p_xmt_msg; |
224 | } peer_t; |
225 | |
226 | |
227 | #define USING_KERNEL_PLL_LOOP 1 |
228 | #define USING_INITIAL_FREQ_ESTIMATION 0 |
229 | |
230 | enum { |
231 | OPT_n = (1 << 0), |
232 | OPT_q = (1 << 1), |
233 | OPT_N = (1 << 2), |
234 | OPT_x = (1 << 3), |
235 | /* Insert new options above this line. */ |
236 | /* Non-compat options: */ |
237 | OPT_w = (1 << 4), |
238 | OPT_p = (1 << 5), |
239 | OPT_S = (1 << 6), |
240 | OPT_l = (1 << 7) * ENABLE_FEATURE_NTPD_SERVER, |
241 | }; |
242 | |
243 | struct globals { |
244 | double cur_time; |
245 | /* total round trip delay to currently selected reference clock */ |
246 | double rootdelay; |
247 | /* reference timestamp: time when the system clock was last set or corrected */ |
248 | double reftime; |
249 | /* total dispersion to currently selected reference clock */ |
250 | double rootdisp; |
251 | |
252 | double last_script_run; |
253 | char *script_name; |
254 | llist_t *ntp_peers; |
255 | #if ENABLE_FEATURE_NTPD_SERVER |
256 | int listen_fd; |
257 | #endif |
258 | unsigned verbose; |
259 | unsigned peer_cnt; |
260 | /* refid: 32-bit code identifying the particular server or reference clock |
261 | * in stratum 0 packets this is a four-character ASCII string, |
262 | * called the kiss code, used for debugging and monitoring |
263 | * in stratum 1 packets this is a four-character ASCII string |
264 | * assigned to the reference clock by IANA. Example: "GPS " |
265 | * in stratum 2+ packets, it's IPv4 address or 4 first bytes of MD5 hash of IPv6 |
266 | */ |
267 | uint32_t refid; |
268 | uint8_t ntp_status; |
269 | /* precision is defined as the larger of the resolution and time to |
270 | * read the clock, in log2 units. For instance, the precision of a |
271 | * mains-frequency clock incrementing at 60 Hz is 16 ms, even when the |
272 | * system clock hardware representation is to the nanosecond. |
273 | * |
274 | * Delays, jitters of various kinds are clamper down to precision. |
275 | * |
276 | * If precision_sec is too large, discipline_jitter gets clamped to it |
277 | * and if offset is much smaller than discipline_jitter, poll interval |
278 | * grows even though we really can benefit from staying at smaller one, |
279 | * collecting non-lagged datapoits and correcting the offset. |
280 | * (Lagged datapoits exist when poll_exp is large but we still have |
281 | * systematic offset error - the time distance between datapoints |
282 | * is significat and older datapoints have smaller offsets. |
283 | * This makes our offset estimation a bit smaller than reality) |
284 | * Due to this effect, setting G_precision_sec close to |
285 | * STEP_THRESHOLD isn't such a good idea - offsets may grow |
286 | * too big and we will step. I observed it with -6. |
287 | * |
288 | * OTOH, setting precision too small would result in futile attempts |
289 | * to syncronize to the unachievable precision. |
290 | * |
291 | * -6 is 1/64 sec, -7 is 1/128 sec and so on. |
292 | */ |
293 | #define G_precision_exp -8 |
294 | #define G_precision_sec (1.0 / (1 << (- G_precision_exp))) |
295 | uint8_t stratum; |
296 | /* Bool. After set to 1, never goes back to 0: */ |
297 | smallint initial_poll_complete; |
298 | |
299 | #define STATE_NSET 0 /* initial state, "nothing is set" */ |
300 | //#define STATE_FSET 1 /* frequency set from file */ |
301 | #define STATE_SPIK 2 /* spike detected */ |
302 | //#define STATE_FREQ 3 /* initial frequency */ |
303 | #define STATE_SYNC 4 /* clock synchronized (normal operation) */ |
304 | uint8_t discipline_state; // doc calls it c.state |
305 | uint8_t poll_exp; // s.poll |
306 | int polladj_count; // c.count |
307 | long kernel_freq_drift; |
308 | peer_t *last_update_peer; |
309 | double last_update_offset; // c.last |
310 | double last_update_recv_time; // s.t |
311 | double discipline_jitter; // c.jitter |
312 | //double cluster_offset; // s.offset |
313 | //double cluster_jitter; // s.jitter |
314 | #if !USING_KERNEL_PLL_LOOP |
315 | double discipline_freq_drift; // c.freq |
316 | /* Maybe conditionally calculate wander? it's used only for logging */ |
317 | double discipline_wander; // c.wander |
318 | #endif |
319 | }; |
320 | #define G (*ptr_to_globals) |
321 | |
322 | static const int const_IPTOS_LOWDELAY = IPTOS_LOWDELAY; |
323 | |
324 | |
325 | #define VERB1 if (MAX_VERBOSE && G.verbose) |
326 | #define VERB2 if (MAX_VERBOSE >= 2 && G.verbose >= 2) |
327 | #define VERB3 if (MAX_VERBOSE >= 3 && G.verbose >= 3) |
328 | #define VERB4 if (MAX_VERBOSE >= 4 && G.verbose >= 4) |
329 | #define VERB5 if (MAX_VERBOSE >= 5 && G.verbose >= 5) |
330 | |
331 | |
332 | static double LOG2D(int a) |
333 | { |
334 | if (a < 0) |
335 | return 1.0 / (1UL << -a); |
336 | return 1UL << a; |
337 | } |
338 | static ALWAYS_INLINE double SQUARE(double x) |
339 | { |
340 | return x * x; |
341 | } |
342 | static ALWAYS_INLINE double MAXD(double a, double b) |
343 | { |
344 | if (a > b) |
345 | return a; |
346 | return b; |
347 | } |
348 | static ALWAYS_INLINE double MIND(double a, double b) |
349 | { |
350 | if (a < b) |
351 | return a; |
352 | return b; |
353 | } |
354 | static NOINLINE double my_SQRT(double X) |
355 | { |
356 | union { |
357 | float f; |
358 | int32_t i; |
359 | } v; |
360 | double invsqrt; |
361 | double Xhalf = X * 0.5; |
362 | |
363 | /* Fast and good approximation to 1/sqrt(X), black magic */ |
364 | v.f = X; |
365 | /*v.i = 0x5f3759df - (v.i >> 1);*/ |
366 | v.i = 0x5f375a86 - (v.i >> 1); /* - this constant is slightly better */ |
367 | invsqrt = v.f; /* better than 0.2% accuracy */ |
368 | |
369 | /* Refining it using Newton's method: x1 = x0 - f(x0)/f'(x0) |
370 | * f(x) = 1/(x*x) - X (f==0 when x = 1/sqrt(X)) |
371 | * f'(x) = -2/(x*x*x) |
372 | * f(x)/f'(x) = (X - 1/(x*x)) / (2/(x*x*x)) = X*x*x*x/2 - x/2 |
373 | * x1 = x0 - (X*x0*x0*x0/2 - x0/2) = 1.5*x0 - X*x0*x0*x0/2 = x0*(1.5 - (X/2)*x0*x0) |
374 | */ |
375 | invsqrt = invsqrt * (1.5 - Xhalf * invsqrt * invsqrt); /* ~0.05% accuracy */ |
376 | /* invsqrt = invsqrt * (1.5 - Xhalf * invsqrt * invsqrt); 2nd iter: ~0.0001% accuracy */ |
377 | /* With 4 iterations, more than half results will be exact, |
378 | * at 6th iterations result stabilizes with about 72% results exact. |
379 | * We are well satisfied with 0.05% accuracy. |
380 | */ |
381 | |
382 | return X * invsqrt; /* X * 1/sqrt(X) ~= sqrt(X) */ |
383 | } |
384 | static ALWAYS_INLINE double SQRT(double X) |
385 | { |
386 | /* If this arch doesn't use IEEE 754 floats, fall back to using libm */ |
387 | if (sizeof(float) != 4) |
388 | return sqrt(X); |
389 | |
390 | /* This avoids needing libm, saves about 0.5k on x86-32 */ |
391 | return my_SQRT(X); |
392 | } |
393 | |
394 | static double |
395 | gettime1900d(void) |
396 | { |
397 | struct timeval tv; |
398 | gettimeofday(&tv, NULL); /* never fails */ |
399 | G.cur_time = tv.tv_sec + (1.0e-6 * tv.tv_usec) + OFFSET_1900_1970; |
400 | return G.cur_time; |
401 | } |
402 | |
403 | static void |
404 | d_to_tv(double d, struct timeval *tv) |
405 | { |
406 | tv->tv_sec = (long)d; |
407 | tv->tv_usec = (d - tv->tv_sec) * 1000000; |
408 | } |
409 | |
410 | static double |
411 | lfp_to_d(l_fixedpt_t lfp) |
412 | { |
413 | double ret; |
414 | lfp.int_partl = ntohl(lfp.int_partl); |
415 | lfp.fractionl = ntohl(lfp.fractionl); |
416 | ret = (double)lfp.int_partl + ((double)lfp.fractionl / UINT_MAX); |
417 | return ret; |
418 | } |
419 | static double |
420 | sfp_to_d(s_fixedpt_t sfp) |
421 | { |
422 | double ret; |
423 | sfp.int_parts = ntohs(sfp.int_parts); |
424 | sfp.fractions = ntohs(sfp.fractions); |
425 | ret = (double)sfp.int_parts + ((double)sfp.fractions / USHRT_MAX); |
426 | return ret; |
427 | } |
428 | #if ENABLE_FEATURE_NTPD_SERVER |
429 | static l_fixedpt_t |
430 | d_to_lfp(double d) |
431 | { |
432 | l_fixedpt_t lfp; |
433 | lfp.int_partl = (uint32_t)d; |
434 | lfp.fractionl = (uint32_t)((d - lfp.int_partl) * UINT_MAX); |
435 | lfp.int_partl = htonl(lfp.int_partl); |
436 | lfp.fractionl = htonl(lfp.fractionl); |
437 | return lfp; |
438 | } |
439 | static s_fixedpt_t |
440 | d_to_sfp(double d) |
441 | { |
442 | s_fixedpt_t sfp; |
443 | sfp.int_parts = (uint16_t)d; |
444 | sfp.fractions = (uint16_t)((d - sfp.int_parts) * USHRT_MAX); |
445 | sfp.int_parts = htons(sfp.int_parts); |
446 | sfp.fractions = htons(sfp.fractions); |
447 | return sfp; |
448 | } |
449 | #endif |
450 | |
451 | static double |
452 | dispersion(const datapoint_t *dp) |
453 | { |
454 | return dp->d_dispersion + FREQ_TOLERANCE * (G.cur_time - dp->d_recv_time); |
455 | } |
456 | |
457 | static double |
458 | root_distance(peer_t *p) |
459 | { |
460 | /* The root synchronization distance is the maximum error due to |
461 | * all causes of the local clock relative to the primary server. |
462 | * It is defined as half the total delay plus total dispersion |
463 | * plus peer jitter. |
464 | */ |
465 | return MAXD(MINDISP, p->lastpkt_rootdelay + p->lastpkt_delay) / 2 |
466 | + p->lastpkt_rootdisp |
467 | + p->filter_dispersion |
468 | + FREQ_TOLERANCE * (G.cur_time - p->lastpkt_recv_time) |
469 | + p->filter_jitter; |
470 | } |
471 | |
472 | static void |
473 | set_next(peer_t *p, unsigned t) |
474 | { |
475 | p->next_action_time = G.cur_time + t; |
476 | } |
477 | |
478 | /* |
479 | * Peer clock filter and its helpers |
480 | */ |
481 | static void |
482 | filter_datapoints(peer_t *p) |
483 | { |
484 | int i, idx; |
485 | int got_newest; |
486 | double minoff, maxoff, wavg, sum, w; |
487 | double x = x; /* for compiler */ |
488 | double oldest_off = oldest_off; |
489 | double oldest_age = oldest_age; |
490 | double newest_off = newest_off; |
491 | double newest_age = newest_age; |
492 | |
493 | minoff = maxoff = p->filter_datapoint[0].d_offset; |
494 | for (i = 1; i < NUM_DATAPOINTS; i++) { |
495 | if (minoff > p->filter_datapoint[i].d_offset) |
496 | minoff = p->filter_datapoint[i].d_offset; |
497 | if (maxoff < p->filter_datapoint[i].d_offset) |
498 | maxoff = p->filter_datapoint[i].d_offset; |
499 | } |
500 | |
501 | idx = p->datapoint_idx; /* most recent datapoint */ |
502 | /* Average offset: |
503 | * Drop two outliers and take weighted average of the rest: |
504 | * most_recent/2 + older1/4 + older2/8 ... + older5/32 + older6/32 |
505 | * we use older6/32, not older6/64 since sum of weights should be 1: |
506 | * 1/2 + 1/4 + 1/8 + 1/16 + 1/32 + 1/32 = 1 |
507 | */ |
508 | wavg = 0; |
509 | w = 0.5; |
510 | /* n-1 |
511 | * --- dispersion(i) |
512 | * filter_dispersion = \ ------------- |
513 | * / (i+1) |
514 | * --- 2 |
515 | * i=0 |
516 | */ |
517 | got_newest = 0; |
518 | sum = 0; |
519 | for (i = 0; i < NUM_DATAPOINTS; i++) { |
520 | VERB4 { |
521 | bb_error_msg("datapoint[%d]: off:%f disp:%f(%f) age:%f%s", |
522 | i, |
523 | p->filter_datapoint[idx].d_offset, |
524 | p->filter_datapoint[idx].d_dispersion, dispersion(&p->filter_datapoint[idx]), |
525 | G.cur_time - p->filter_datapoint[idx].d_recv_time, |
526 | (minoff == p->filter_datapoint[idx].d_offset || maxoff == p->filter_datapoint[idx].d_offset) |
527 | ? " (outlier by offset)" : "" |
528 | ); |
529 | } |
530 | |
531 | sum += dispersion(&p->filter_datapoint[idx]) / (2 << i); |
532 | |
533 | if (minoff == p->filter_datapoint[idx].d_offset) { |
534 | minoff -= 1; /* so that we don't match it ever again */ |
535 | } else |
536 | if (maxoff == p->filter_datapoint[idx].d_offset) { |
537 | maxoff += 1; |
538 | } else { |
539 | oldest_off = p->filter_datapoint[idx].d_offset; |
540 | oldest_age = G.cur_time - p->filter_datapoint[idx].d_recv_time; |
541 | if (!got_newest) { |
542 | got_newest = 1; |
543 | newest_off = oldest_off; |
544 | newest_age = oldest_age; |
545 | } |
546 | x = oldest_off * w; |
547 | wavg += x; |
548 | w /= 2; |
549 | } |
550 | |
551 | idx = (idx - 1) & (NUM_DATAPOINTS - 1); |
552 | } |
553 | p->filter_dispersion = sum; |
554 | wavg += x; /* add another older6/64 to form older6/32 */ |
555 | /* Fix systematic underestimation with large poll intervals. |
556 | * Imagine that we still have a bit of uncorrected drift, |
557 | * and poll interval is big (say, 100 sec). Offsets form a progression: |
558 | * 0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 - 0.7 is most recent. |
559 | * The algorithm above drops 0.0 and 0.7 as outliers, |
560 | * and then we have this estimation, ~25% off from 0.7: |
561 | * 0.1/32 + 0.2/32 + 0.3/16 + 0.4/8 + 0.5/4 + 0.6/2 = 0.503125 |
562 | */ |
563 | x = oldest_age - newest_age; |
564 | if (x != 0) { |
565 | x = newest_age / x; /* in above example, 100 / (600 - 100) */ |
566 | if (x < 1) { /* paranoia check */ |
567 | x = (newest_off - oldest_off) * x; /* 0.5 * 100/500 = 0.1 */ |
568 | wavg += x; |
569 | } |
570 | } |
571 | p->filter_offset = wavg; |
572 | |
573 | /* +----- -----+ ^ 1/2 |
574 | * | n-1 | |
575 | * | --- | |
576 | * | 1 \ 2 | |
577 | * filter_jitter = | --- * / (avg-offset_j) | |
578 | * | n --- | |
579 | * | j=0 | |
580 | * +----- -----+ |
581 | * where n is the number of valid datapoints in the filter (n > 1); |
582 | * if filter_jitter < precision then filter_jitter = precision |
583 | */ |
584 | sum = 0; |
585 | for (i = 0; i < NUM_DATAPOINTS; i++) { |
586 | sum += SQUARE(wavg - p->filter_datapoint[i].d_offset); |
587 | } |
588 | sum = SQRT(sum / NUM_DATAPOINTS); |
589 | p->filter_jitter = sum > G_precision_sec ? sum : G_precision_sec; |
590 | |
591 | VERB3 bb_error_msg("filter offset:%f(corr:%e) disp:%f jitter:%f", |
592 | p->filter_offset, x, |
593 | p->filter_dispersion, |
594 | p->filter_jitter); |
595 | |
596 | } |
597 | |
598 | static void |
599 | reset_peer_stats(peer_t *p, double offset) |
600 | { |
601 | int i; |
602 | bool small_ofs = fabs(offset) < 16 * STEP_THRESHOLD; |
603 | |
604 | for (i = 0; i < NUM_DATAPOINTS; i++) { |
605 | if (small_ofs) { |
606 | p->filter_datapoint[i].d_recv_time -= offset; |
607 | if (p->filter_datapoint[i].d_offset != 0) { |
608 | p->filter_datapoint[i].d_offset -= offset; |
609 | } |
610 | } else { |
611 | p->filter_datapoint[i].d_recv_time = G.cur_time; |
612 | p->filter_datapoint[i].d_offset = 0; |
613 | p->filter_datapoint[i].d_dispersion = MAXDISP; |
614 | } |
615 | } |
616 | if (small_ofs) { |
617 | p->lastpkt_recv_time -= offset; |
618 | } else { |
619 | p->reachable_bits = 0; |
620 | p->lastpkt_recv_time = G.cur_time; |
621 | } |
622 | filter_datapoints(p); /* recalc p->filter_xxx */ |
623 | p->next_action_time -= offset; |
624 | VERB5 bb_error_msg("%s->lastpkt_recv_time=%f", p->p_dotted, p->lastpkt_recv_time); |
625 | } |
626 | |
627 | static void |
628 | add_peers(char *s) |
629 | { |
630 | peer_t *p; |
631 | |
632 | p = xzalloc(sizeof(*p)); |
633 | p->p_lsa = xhost2sockaddr(s, 123); |
634 | p->p_dotted = xmalloc_sockaddr2dotted_noport(&p->p_lsa->u.sa); |
635 | p->p_fd = -1; |
636 | p->p_xmt_msg.m_status = MODE_CLIENT | (NTP_VERSION << 3); |
637 | p->next_action_time = G.cur_time; /* = set_next(p, 0); */ |
638 | reset_peer_stats(p, 16 * STEP_THRESHOLD); |
639 | |
640 | llist_add_to(&G.ntp_peers, p); |
641 | G.peer_cnt++; |
642 | } |
643 | |
644 | static int |
645 | do_sendto(int fd, |
646 | const struct sockaddr *from, const struct sockaddr *to, socklen_t addrlen, |
647 | msg_t *msg, ssize_t len) |
648 | { |
649 | ssize_t ret; |
650 | |
651 | errno = 0; |
652 | if (!from) { |
653 | ret = sendto(fd, msg, len, MSG_DONTWAIT, to, addrlen); |
654 | } else { |
655 | ret = send_to_from(fd, msg, len, MSG_DONTWAIT, to, from, addrlen); |
656 | } |
657 | if (ret != len) { |
658 | bb_perror_msg("send failed"); |
659 | return -1; |
660 | } |
661 | return 0; |
662 | } |
663 | |
664 | static void |
665 | send_query_to_peer(peer_t *p) |
666 | { |
667 | /* Why do we need to bind()? |
668 | * See what happens when we don't bind: |
669 | * |
670 | * socket(PF_INET, SOCK_DGRAM, IPPROTO_IP) = 3 |
671 | * setsockopt(3, SOL_IP, IP_TOS, [16], 4) = 0 |
672 | * gettimeofday({1259071266, 327885}, NULL) = 0 |
673 | * sendto(3, "xxx", 48, MSG_DONTWAIT, {sa_family=AF_INET, sin_port=htons(123), sin_addr=inet_addr("10.34.32.125")}, 16) = 48 |
674 | * ^^^ we sent it from some source port picked by kernel. |
675 | * time(NULL) = 1259071266 |
676 | * write(2, "ntpd: entering poll 15 secs\n", 28) = 28 |
677 | * poll([{fd=3, events=POLLIN}], 1, 15000) = 1 ([{fd=3, revents=POLLIN}]) |
678 | * recv(3, "yyy", 68, MSG_DONTWAIT) = 48 |
679 | * ^^^ this recv will receive packets to any local port! |
680 | * |
681 | * Uncomment this and use strace to see it in action: |
682 | */ |
683 | #define PROBE_LOCAL_ADDR /* { len_and_sockaddr lsa; lsa.len = LSA_SIZEOF_SA; getsockname(p->query.fd, &lsa.u.sa, &lsa.len); } */ |
684 | |
685 | if (p->p_fd == -1) { |
686 | int fd, family; |
687 | len_and_sockaddr *local_lsa; |
688 | |
689 | family = p->p_lsa->u.sa.sa_family; |
690 | p->p_fd = fd = xsocket_type(&local_lsa, family, SOCK_DGRAM); |
691 | /* local_lsa has "null" address and port 0 now. |
692 | * bind() ensures we have a *particular port* selected by kernel |
693 | * and remembered in p->p_fd, thus later recv(p->p_fd) |
694 | * receives only packets sent to this port. |
695 | */ |
696 | PROBE_LOCAL_ADDR |
697 | xbind(fd, &local_lsa->u.sa, local_lsa->len); |
698 | PROBE_LOCAL_ADDR |
699 | #if ENABLE_FEATURE_IPV6 |
700 | if (family == AF_INET) |
701 | #endif |
702 | setsockopt(fd, IPPROTO_IP, IP_TOS, &const_IPTOS_LOWDELAY, sizeof(const_IPTOS_LOWDELAY)); |
703 | free(local_lsa); |
704 | } |
705 | |
706 | /* |
707 | * Send out a random 64-bit number as our transmit time. The NTP |
708 | * server will copy said number into the originate field on the |
709 | * response that it sends us. This is totally legal per the SNTP spec. |
710 | * |
711 | * The impact of this is two fold: we no longer send out the current |
712 | * system time for the world to see (which may aid an attacker), and |
713 | * it gives us a (not very secure) way of knowing that we're not |
714 | * getting spoofed by an attacker that can't capture our traffic |
715 | * but can spoof packets from the NTP server we're communicating with. |
716 | * |
717 | * Save the real transmit timestamp locally. |
718 | */ |
719 | p->p_xmt_msg.m_xmttime.int_partl = random(); |
720 | p->p_xmt_msg.m_xmttime.fractionl = random(); |
721 | p->p_xmttime = gettime1900d(); |
722 | |
723 | if (do_sendto(p->p_fd, /*from:*/ NULL, /*to:*/ &p->p_lsa->u.sa, /*addrlen:*/ p->p_lsa->len, |
724 | &p->p_xmt_msg, NTP_MSGSIZE_NOAUTH) == -1 |
725 | ) { |
726 | close(p->p_fd); |
727 | p->p_fd = -1; |
728 | set_next(p, RETRY_INTERVAL); |
729 | return; |
730 | } |
731 | |
732 | p->reachable_bits <<= 1; |
733 | VERB1 bb_error_msg("sent query to %s", p->p_dotted); |
734 | set_next(p, RESPONSE_INTERVAL); |
735 | } |
736 | |
737 | |
738 | /* Note that there is no provision to prevent several run_scripts |
739 | * to be done in quick succession. In fact, it happens rather often |
740 | * if initial syncronization results in a step. |
741 | * You will see "step" and then "stratum" script runs, sometimes |
742 | * as close as only 0.002 seconds apart. |
743 | * Script should be ready to deal with this. |
744 | */ |
745 | static void run_script(const char *action, double offset) |
746 | { |
747 | char *argv[3]; |
748 | char *env1, *env2, *env3, *env4; |
749 | |
750 | if (!G.script_name) |
751 | return; |
752 | |
753 | argv[0] = (char*) G.script_name; |
754 | argv[1] = (char*) action; |
755 | argv[2] = NULL; |
756 | |
757 | VERB1 bb_error_msg("executing '%s %s'", G.script_name, action); |
758 | |
759 | env1 = xasprintf("%s=%u", "stratum", G.stratum); |
760 | putenv(env1); |
761 | env2 = xasprintf("%s=%ld", "freq_drift_ppm", G.kernel_freq_drift); |
762 | putenv(env2); |
763 | env3 = xasprintf("%s=%u", "poll_interval", 1 << G.poll_exp); |
764 | putenv(env3); |
765 | env4 = xasprintf("%s=%f", "offset", offset); |
766 | putenv(env4); |
767 | /* Other items of potential interest: selected peer, |
768 | * rootdelay, reftime, rootdisp, refid, ntp_status, |
769 | * last_update_offset, last_update_recv_time, discipline_jitter, |
770 | * how many peers have reachable_bits = 0? |
771 | */ |
772 | |
773 | /* Don't want to wait: it may run hwclock --systohc, and that |
774 | * may take some time (seconds): */ |
775 | /*wait4pid(spawn(argv));*/ |
776 | spawn(argv); |
777 | |
778 | unsetenv("stratum"); |
779 | unsetenv("freq_drift_ppm"); |
780 | unsetenv("poll_interval"); |
781 | unsetenv("offset"); |
782 | free(env1); |
783 | free(env2); |
784 | free(env3); |
785 | free(env4); |
786 | |
787 | G.last_script_run = G.cur_time; |
788 | } |
789 | |
790 | static NOINLINE void |
791 | step_time(double offset) |
792 | { |
793 | llist_t *item; |
794 | double dtime; |
795 | struct timeval tv; |
796 | char buf[80]; |
797 | time_t tval; |
798 | |
799 | gettimeofday(&tv, NULL); /* never fails */ |
800 | dtime = offset + tv.tv_sec; |
801 | dtime += 1.0e-6 * tv.tv_usec; |
802 | d_to_tv(dtime, &tv); |
803 | |
804 | if (settimeofday(&tv, NULL) == -1) |
805 | bb_perror_msg_and_die("settimeofday"); |
806 | |
807 | tval = tv.tv_sec; |
808 | strftime(buf, sizeof(buf), "%a %b %e %H:%M:%S %Z %Y", localtime(&tval)); |
809 | |
810 | bb_error_msg("setting clock to %s (offset %fs)", buf, offset); |
811 | |
812 | /* Correct various fields which contain time-relative values: */ |
813 | |
814 | /* p->lastpkt_recv_time, p->next_action_time and such: */ |
815 | for (item = G.ntp_peers; item != NULL; item = item->link) { |
816 | peer_t *pp = (peer_t *) item->data; |
817 | reset_peer_stats(pp, offset); |
818 | } |
819 | /* Globals: */ |
820 | G.cur_time -= offset; |
821 | G.last_update_recv_time -= offset; |
822 | G.last_script_run -= offset; |
823 | } |
824 | |
825 | |
826 | /* |
827 | * Selection and clustering, and their helpers |
828 | */ |
829 | typedef struct { |
830 | peer_t *p; |
831 | int type; |
832 | double edge; |
833 | double opt_rd; /* optimization */ |
834 | } point_t; |
835 | static int |
836 | compare_point_edge(const void *aa, const void *bb) |
837 | { |
838 | const point_t *a = aa; |
839 | const point_t *b = bb; |
840 | if (a->edge < b->edge) { |
841 | return -1; |
842 | } |
843 | return (a->edge > b->edge); |
844 | } |
845 | typedef struct { |
846 | peer_t *p; |
847 | double metric; |
848 | } survivor_t; |
849 | static int |
850 | compare_survivor_metric(const void *aa, const void *bb) |
851 | { |
852 | const survivor_t *a = aa; |
853 | const survivor_t *b = bb; |
854 | if (a->metric < b->metric) { |
855 | return -1; |
856 | } |
857 | return (a->metric > b->metric); |
858 | } |
859 | static int |
860 | fit(peer_t *p, double rd) |
861 | { |
862 | if ((p->reachable_bits & (p->reachable_bits-1)) == 0) { |
863 | /* One or zero bits in reachable_bits */ |
864 | VERB3 bb_error_msg("peer %s unfit for selection: unreachable", p->p_dotted); |
865 | return 0; |
866 | } |
867 | #if 0 /* we filter out such packets earlier */ |
868 | if ((p->lastpkt_status & LI_ALARM) == LI_ALARM |
869 | || p->lastpkt_stratum >= MAXSTRAT |
870 | ) { |
871 | VERB3 bb_error_msg("peer %s unfit for selection: bad status/stratum", p->p_dotted); |
872 | return 0; |
873 | } |
874 | #endif |
875 | /* rd is root_distance(p) */ |
876 | if (rd > MAXDIST + FREQ_TOLERANCE * (1 << G.poll_exp)) { |
877 | VERB3 bb_error_msg("peer %s unfit for selection: root distance too high", p->p_dotted); |
878 | return 0; |
879 | } |
880 | //TODO |
881 | // /* Do we have a loop? */ |
882 | // if (p->refid == p->dstaddr || p->refid == s.refid) |
883 | // return 0; |
884 | return 1; |
885 | } |
886 | static peer_t* |
887 | select_and_cluster(void) |
888 | { |
889 | peer_t *p; |
890 | llist_t *item; |
891 | int i, j; |
892 | int size = 3 * G.peer_cnt; |
893 | /* for selection algorithm */ |
894 | point_t point[size]; |
895 | unsigned num_points, num_candidates; |
896 | double low, high; |
897 | unsigned num_falsetickers; |
898 | /* for cluster algorithm */ |
899 | survivor_t survivor[size]; |
900 | unsigned num_survivors; |
901 | |
902 | /* Selection */ |
903 | |
904 | num_points = 0; |
905 | item = G.ntp_peers; |
906 | if (G.initial_poll_complete) while (item != NULL) { |
907 | double rd, offset; |
908 | |
909 | p = (peer_t *) item->data; |
910 | rd = root_distance(p); |
911 | offset = p->filter_offset; |
912 | if (!fit(p, rd)) { |
913 | item = item->link; |
914 | continue; |
915 | } |
916 | |
917 | VERB4 bb_error_msg("interval: [%f %f %f] %s", |
918 | offset - rd, |
919 | offset, |
920 | offset + rd, |
921 | p->p_dotted |
922 | ); |
923 | point[num_points].p = p; |
924 | point[num_points].type = -1; |
925 | point[num_points].edge = offset - rd; |
926 | point[num_points].opt_rd = rd; |
927 | num_points++; |
928 | point[num_points].p = p; |
929 | point[num_points].type = 0; |
930 | point[num_points].edge = offset; |
931 | point[num_points].opt_rd = rd; |
932 | num_points++; |
933 | point[num_points].p = p; |
934 | point[num_points].type = 1; |
935 | point[num_points].edge = offset + rd; |
936 | point[num_points].opt_rd = rd; |
937 | num_points++; |
938 | item = item->link; |
939 | } |
940 | num_candidates = num_points / 3; |
941 | if (num_candidates == 0) { |
942 | VERB3 bb_error_msg("no valid datapoints, no peer selected"); |
943 | return NULL; |
944 | } |
945 | //TODO: sorting does not seem to be done in reference code |
946 | qsort(point, num_points, sizeof(point[0]), compare_point_edge); |
947 | |
948 | /* Start with the assumption that there are no falsetickers. |
949 | * Attempt to find a nonempty intersection interval containing |
950 | * the midpoints of all truechimers. |
951 | * If a nonempty interval cannot be found, increase the number |
952 | * of assumed falsetickers by one and try again. |
953 | * If a nonempty interval is found and the number of falsetickers |
954 | * is less than the number of truechimers, a majority has been found |
955 | * and the midpoint of each truechimer represents |
956 | * the candidates available to the cluster algorithm. |
957 | */ |
958 | num_falsetickers = 0; |
959 | while (1) { |
960 | int c; |
961 | unsigned num_midpoints = 0; |
962 | |
963 | low = 1 << 9; |
964 | high = - (1 << 9); |
965 | c = 0; |
966 | for (i = 0; i < num_points; i++) { |
967 | /* We want to do: |
968 | * if (point[i].type == -1) c++; |
969 | * if (point[i].type == 1) c--; |
970 | * and it's simpler to do it this way: |
971 | */ |
972 | c -= point[i].type; |
973 | if (c >= num_candidates - num_falsetickers) { |
974 | /* If it was c++ and it got big enough... */ |
975 | low = point[i].edge; |
976 | break; |
977 | } |
978 | if (point[i].type == 0) |
979 | num_midpoints++; |
980 | } |
981 | c = 0; |
982 | for (i = num_points-1; i >= 0; i--) { |
983 | c += point[i].type; |
984 | if (c >= num_candidates - num_falsetickers) { |
985 | high = point[i].edge; |
986 | break; |
987 | } |
988 | if (point[i].type == 0) |
989 | num_midpoints++; |
990 | } |
991 | /* If the number of midpoints is greater than the number |
992 | * of allowed falsetickers, the intersection contains at |
993 | * least one truechimer with no midpoint - bad. |
994 | * Also, interval should be nonempty. |
995 | */ |
996 | if (num_midpoints <= num_falsetickers && low < high) |
997 | break; |
998 | num_falsetickers++; |
999 | if (num_falsetickers * 2 >= num_candidates) { |
1000 | VERB3 bb_error_msg("too many falsetickers:%d (candidates:%d), no peer selected", |
1001 | num_falsetickers, num_candidates); |
1002 | return NULL; |
1003 | } |
1004 | } |
1005 | VERB3 bb_error_msg("selected interval: [%f, %f]; candidates:%d falsetickers:%d", |
1006 | low, high, num_candidates, num_falsetickers); |
1007 | |
1008 | /* Clustering */ |
1009 | |
1010 | /* Construct a list of survivors (p, metric) |
1011 | * from the chime list, where metric is dominated |
1012 | * first by stratum and then by root distance. |
1013 | * All other things being equal, this is the order of preference. |
1014 | */ |
1015 | num_survivors = 0; |
1016 | for (i = 0; i < num_points; i++) { |
1017 | if (point[i].edge < low || point[i].edge > high) |
1018 | continue; |
1019 | p = point[i].p; |
1020 | survivor[num_survivors].p = p; |
1021 | /* x.opt_rd == root_distance(p); */ |
1022 | survivor[num_survivors].metric = MAXDIST * p->lastpkt_stratum + point[i].opt_rd; |
1023 | VERB4 bb_error_msg("survivor[%d] metric:%f peer:%s", |
1024 | num_survivors, survivor[num_survivors].metric, p->p_dotted); |
1025 | num_survivors++; |
1026 | } |
1027 | /* There must be at least MIN_SELECTED survivors to satisfy the |
1028 | * correctness assertions. Ordinarily, the Byzantine criteria |
1029 | * require four survivors, but for the demonstration here, one |
1030 | * is acceptable. |
1031 | */ |
1032 | if (num_survivors < MIN_SELECTED) { |
1033 | VERB3 bb_error_msg("num_survivors %d < %d, no peer selected", |
1034 | num_survivors, MIN_SELECTED); |
1035 | return NULL; |
1036 | } |
1037 | |
1038 | //looks like this is ONLY used by the fact that later we pick survivor[0]. |
1039 | //we can avoid sorting then, just find the minimum once! |
1040 | qsort(survivor, num_survivors, sizeof(survivor[0]), compare_survivor_metric); |
1041 | |
1042 | /* For each association p in turn, calculate the selection |
1043 | * jitter p->sjitter as the square root of the sum of squares |
1044 | * (p->offset - q->offset) over all q associations. The idea is |
1045 | * to repeatedly discard the survivor with maximum selection |
1046 | * jitter until a termination condition is met. |
1047 | */ |
1048 | while (1) { |
1049 | unsigned max_idx = max_idx; |
1050 | double max_selection_jitter = max_selection_jitter; |
1051 | double min_jitter = min_jitter; |
1052 | |
1053 | if (num_survivors <= MIN_CLUSTERED) { |
1054 | VERB3 bb_error_msg("num_survivors %d <= %d, not discarding more", |
1055 | num_survivors, MIN_CLUSTERED); |
1056 | break; |
1057 | } |
1058 | |
1059 | /* To make sure a few survivors are left |
1060 | * for the clustering algorithm to chew on, |
1061 | * we stop if the number of survivors |
1062 | * is less than or equal to MIN_CLUSTERED (3). |
1063 | */ |
1064 | for (i = 0; i < num_survivors; i++) { |
1065 | double selection_jitter_sq; |
1066 | |
1067 | p = survivor[i].p; |
1068 | if (i == 0 || p->filter_jitter < min_jitter) |
1069 | min_jitter = p->filter_jitter; |
1070 | |
1071 | selection_jitter_sq = 0; |
1072 | for (j = 0; j < num_survivors; j++) { |
1073 | peer_t *q = survivor[j].p; |
1074 | selection_jitter_sq += SQUARE(p->filter_offset - q->filter_offset); |
1075 | } |
1076 | if (i == 0 || selection_jitter_sq > max_selection_jitter) { |
1077 | max_selection_jitter = selection_jitter_sq; |
1078 | max_idx = i; |
1079 | } |
1080 | VERB5 bb_error_msg("survivor %d selection_jitter^2:%f", |
1081 | i, selection_jitter_sq); |
1082 | } |
1083 | max_selection_jitter = SQRT(max_selection_jitter / num_survivors); |
1084 | VERB4 bb_error_msg("max_selection_jitter (at %d):%f min_jitter:%f", |
1085 | max_idx, max_selection_jitter, min_jitter); |
1086 | |
1087 | /* If the maximum selection jitter is less than the |
1088 | * minimum peer jitter, then tossing out more survivors |
1089 | * will not lower the minimum peer jitter, so we might |
1090 | * as well stop. |
1091 | */ |
1092 | if (max_selection_jitter < min_jitter) { |
1093 | VERB3 bb_error_msg("max_selection_jitter:%f < min_jitter:%f, num_survivors:%d, not discarding more", |
1094 | max_selection_jitter, min_jitter, num_survivors); |
1095 | break; |
1096 | } |
1097 | |
1098 | /* Delete survivor[max_idx] from the list |
1099 | * and go around again. |
1100 | */ |
1101 | VERB5 bb_error_msg("dropping survivor %d", max_idx); |
1102 | num_survivors--; |
1103 | while (max_idx < num_survivors) { |
1104 | survivor[max_idx] = survivor[max_idx + 1]; |
1105 | max_idx++; |
1106 | } |
1107 | } |
1108 | |
1109 | if (0) { |
1110 | /* Combine the offsets of the clustering algorithm survivors |
1111 | * using a weighted average with weight determined by the root |
1112 | * distance. Compute the selection jitter as the weighted RMS |
1113 | * difference between the first survivor and the remaining |
1114 | * survivors. In some cases the inherent clock jitter can be |
1115 | * reduced by not using this algorithm, especially when frequent |
1116 | * clockhopping is involved. bbox: thus we don't do it. |
1117 | */ |
1118 | double x, y, z, w; |
1119 | y = z = w = 0; |
1120 | for (i = 0; i < num_survivors; i++) { |
1121 | p = survivor[i].p; |
1122 | x = root_distance(p); |
1123 | y += 1 / x; |
1124 | z += p->filter_offset / x; |
1125 | w += SQUARE(p->filter_offset - survivor[0].p->filter_offset) / x; |
1126 | } |
1127 | //G.cluster_offset = z / y; |
1128 | //G.cluster_jitter = SQRT(w / y); |
1129 | } |
1130 | |
1131 | /* Pick the best clock. If the old system peer is on the list |
1132 | * and at the same stratum as the first survivor on the list, |
1133 | * then don't do a clock hop. Otherwise, select the first |
1134 | * survivor on the list as the new system peer. |
1135 | */ |
1136 | p = survivor[0].p; |
1137 | if (G.last_update_peer |
1138 | && G.last_update_peer->lastpkt_stratum <= p->lastpkt_stratum |
1139 | ) { |
1140 | /* Starting from 1 is ok here */ |
1141 | for (i = 1; i < num_survivors; i++) { |
1142 | if (G.last_update_peer == survivor[i].p) { |
1143 | VERB4 bb_error_msg("keeping old synced peer"); |
1144 | p = G.last_update_peer; |
1145 | goto keep_old; |
1146 | } |
1147 | } |
1148 | } |
1149 | G.last_update_peer = p; |
1150 | keep_old: |
1151 | VERB3 bb_error_msg("selected peer %s filter_offset:%f age:%f", |
1152 | p->p_dotted, |
1153 | p->filter_offset, |
1154 | G.cur_time - p->lastpkt_recv_time |
1155 | ); |
1156 | return p; |
1157 | } |
1158 | |
1159 | |
1160 | /* |
1161 | * Local clock discipline and its helpers |
1162 | */ |
1163 | static void |
1164 | set_new_values(int disc_state, double offset, double recv_time) |
1165 | { |
1166 | /* Enter new state and set state variables. Note we use the time |
1167 | * of the last clock filter sample, which must be earlier than |
1168 | * the current time. |
1169 | */ |
1170 | VERB3 bb_error_msg("disc_state=%d last update offset=%f recv_time=%f", |
1171 | disc_state, offset, recv_time); |
1172 | G.discipline_state = disc_state; |
1173 | G.last_update_offset = offset; |
1174 | G.last_update_recv_time = recv_time; |
1175 | } |
1176 | /* Return: -1: decrease poll interval, 0: leave as is, 1: increase */ |
1177 | static NOINLINE int |
1178 | update_local_clock(peer_t *p) |
1179 | { |
1180 | int rc; |
1181 | struct timex tmx; |
1182 | /* Note: can use G.cluster_offset instead: */ |
1183 | double offset = p->filter_offset; |
1184 | double recv_time = p->lastpkt_recv_time; |
1185 | double abs_offset; |
1186 | #if !USING_KERNEL_PLL_LOOP |
1187 | double freq_drift; |
1188 | #endif |
1189 | double since_last_update; |
1190 | double etemp, dtemp; |
1191 | |
1192 | abs_offset = fabs(offset); |
1193 | |
1194 | #if 0 |
1195 | /* If needed, -S script can do it by looking at $offset |
1196 | * env var and killing parent */ |
1197 | /* If the offset is too large, give up and go home */ |
1198 | if (abs_offset > PANIC_THRESHOLD) { |
1199 | bb_error_msg_and_die("offset %f far too big, exiting", offset); |
1200 | } |
1201 | #endif |
1202 | |
1203 | /* If this is an old update, for instance as the result |
1204 | * of a system peer change, avoid it. We never use |
1205 | * an old sample or the same sample twice. |
1206 | */ |
1207 | if (recv_time <= G.last_update_recv_time) { |
1208 | VERB3 bb_error_msg("same or older datapoint: %f >= %f, not using it", |
1209 | G.last_update_recv_time, recv_time); |
1210 | return 0; /* "leave poll interval as is" */ |
1211 | } |
1212 | |
1213 | /* Clock state machine transition function. This is where the |
1214 | * action is and defines how the system reacts to large time |
1215 | * and frequency errors. |
1216 | */ |
1217 | since_last_update = recv_time - G.reftime; |
1218 | #if !USING_KERNEL_PLL_LOOP |
1219 | freq_drift = 0; |
1220 | #endif |
1221 | #if USING_INITIAL_FREQ_ESTIMATION |
1222 | if (G.discipline_state == STATE_FREQ) { |
1223 | /* Ignore updates until the stepout threshold */ |
1224 | if (since_last_update < WATCH_THRESHOLD) { |
1225 | VERB3 bb_error_msg("measuring drift, datapoint ignored, %f sec remains", |
1226 | WATCH_THRESHOLD - since_last_update); |
1227 | return 0; /* "leave poll interval as is" */ |
1228 | } |
1229 | # if !USING_KERNEL_PLL_LOOP |
1230 | freq_drift = (offset - G.last_update_offset) / since_last_update; |
1231 | # endif |
1232 | } |
1233 | #endif |
1234 | |
1235 | /* There are two main regimes: when the |
1236 | * offset exceeds the step threshold and when it does not. |
1237 | */ |
1238 | if (abs_offset > STEP_THRESHOLD) { |
1239 | switch (G.discipline_state) { |
1240 | case STATE_SYNC: |
1241 | /* The first outlyer: ignore it, switch to SPIK state */ |
1242 | VERB3 bb_error_msg("offset:%f - spike detected", offset); |
1243 | G.discipline_state = STATE_SPIK; |
1244 | return -1; /* "decrease poll interval" */ |
1245 | |
1246 | case STATE_SPIK: |
1247 | /* Ignore succeeding outlyers until either an inlyer |
1248 | * is found or the stepout threshold is exceeded. |
1249 | */ |
1250 | if (since_last_update < WATCH_THRESHOLD) { |
1251 | VERB3 bb_error_msg("spike detected, datapoint ignored, %f sec remains", |
1252 | WATCH_THRESHOLD - since_last_update); |
1253 | return -1; /* "decrease poll interval" */ |
1254 | } |
1255 | /* fall through: we need to step */ |
1256 | } /* switch */ |
1257 | |
1258 | /* Step the time and clamp down the poll interval. |
1259 | * |
1260 | * In NSET state an initial frequency correction is |
1261 | * not available, usually because the frequency file has |
1262 | * not yet been written. Since the time is outside the |
1263 | * capture range, the clock is stepped. The frequency |
1264 | * will be set directly following the stepout interval. |
1265 | * |
1266 | * In FSET state the initial frequency has been set |
1267 | * from the frequency file. Since the time is outside |
1268 | * the capture range, the clock is stepped immediately, |
1269 | * rather than after the stepout interval. Guys get |
1270 | * nervous if it takes 17 minutes to set the clock for |
1271 | * the first time. |
1272 | * |
1273 | * In SPIK state the stepout threshold has expired and |
1274 | * the phase is still above the step threshold. Note |
1275 | * that a single spike greater than the step threshold |
1276 | * is always suppressed, even at the longer poll |
1277 | * intervals. |
1278 | */ |
1279 | VERB3 bb_error_msg("stepping time by %f; poll_exp=MINPOLL", offset); |
1280 | step_time(offset); |
1281 | if (option_mask32 & OPT_q) { |
1282 | /* We were only asked to set time once. Done. */ |
1283 | exit(0); |
1284 | } |
1285 | |
1286 | G.polladj_count = 0; |
1287 | G.poll_exp = MINPOLL; |
1288 | G.stratum = MAXSTRAT; |
1289 | |
1290 | run_script("step", offset); |
1291 | |
1292 | #if USING_INITIAL_FREQ_ESTIMATION |
1293 | if (G.discipline_state == STATE_NSET) { |
1294 | set_new_values(STATE_FREQ, /*offset:*/ 0, recv_time); |
1295 | return 1; /* "ok to increase poll interval" */ |
1296 | } |
1297 | #endif |
1298 | set_new_values(STATE_SYNC, /*offset:*/ 0, recv_time); |
1299 | |
1300 | } else { /* abs_offset <= STEP_THRESHOLD */ |
1301 | |
1302 | if (G.poll_exp < MINPOLL && G.initial_poll_complete) { |
1303 | VERB3 bb_error_msg("small offset:%f, disabling burst mode", offset); |
1304 | G.polladj_count = 0; |
1305 | G.poll_exp = MINPOLL; |
1306 | } |
1307 | |
1308 | /* Compute the clock jitter as the RMS of exponentially |
1309 | * weighted offset differences. Used by the poll adjust code. |
1310 | */ |
1311 | etemp = SQUARE(G.discipline_jitter); |
1312 | dtemp = SQUARE(MAXD(fabs(offset - G.last_update_offset), G_precision_sec)); |
1313 | G.discipline_jitter = SQRT(etemp + (dtemp - etemp) / AVG); |
1314 | VERB3 bb_error_msg("discipline jitter=%f", G.discipline_jitter); |
1315 | |
1316 | switch (G.discipline_state) { |
1317 | case STATE_NSET: |
1318 | if (option_mask32 & OPT_q) { |
1319 | /* We were only asked to set time once. |
1320 | * The clock is precise enough, no need to step. |
1321 | */ |
1322 | exit(0); |
1323 | } |
1324 | #if USING_INITIAL_FREQ_ESTIMATION |
1325 | /* This is the first update received and the frequency |
1326 | * has not been initialized. The first thing to do |
1327 | * is directly measure the oscillator frequency. |
1328 | */ |
1329 | set_new_values(STATE_FREQ, offset, recv_time); |
1330 | #else |
1331 | set_new_values(STATE_SYNC, offset, recv_time); |
1332 | #endif |
1333 | VERB3 bb_error_msg("transitioning to FREQ, datapoint ignored"); |
1334 | return 0; /* "leave poll interval as is" */ |
1335 | |
1336 | #if 0 /* this is dead code for now */ |
1337 | case STATE_FSET: |
1338 | /* This is the first update and the frequency |
1339 | * has been initialized. Adjust the phase, but |
1340 | * don't adjust the frequency until the next update. |
1341 | */ |
1342 | set_new_values(STATE_SYNC, offset, recv_time); |
1343 | /* freq_drift remains 0 */ |
1344 | break; |
1345 | #endif |
1346 | |
1347 | #if USING_INITIAL_FREQ_ESTIMATION |
1348 | case STATE_FREQ: |
1349 | /* since_last_update >= WATCH_THRESHOLD, we waited enough. |
1350 | * Correct the phase and frequency and switch to SYNC state. |
1351 | * freq_drift was already estimated (see code above) |
1352 | */ |
1353 | set_new_values(STATE_SYNC, offset, recv_time); |
1354 | break; |
1355 | #endif |
1356 | |
1357 | default: |
1358 | #if !USING_KERNEL_PLL_LOOP |
1359 | /* Compute freq_drift due to PLL and FLL contributions. |
1360 | * |
1361 | * The FLL and PLL frequency gain constants |
1362 | * depend on the poll interval and Allan |
1363 | * intercept. The FLL is not used below one-half |
1364 | * the Allan intercept. Above that the loop gain |
1365 | * increases in steps to 1 / AVG. |
1366 | */ |
1367 | if ((1 << G.poll_exp) > ALLAN / 2) { |
1368 | etemp = FLL - G.poll_exp; |
1369 | if (etemp < AVG) |
1370 | etemp = AVG; |
1371 | freq_drift += (offset - G.last_update_offset) / (MAXD(since_last_update, ALLAN) * etemp); |
1372 | } |
1373 | /* For the PLL the integration interval |
1374 | * (numerator) is the minimum of the update |
1375 | * interval and poll interval. This allows |
1376 | * oversampling, but not undersampling. |
1377 | */ |
1378 | etemp = MIND(since_last_update, (1 << G.poll_exp)); |
1379 | dtemp = (4 * PLL) << G.poll_exp; |
1380 | freq_drift += offset * etemp / SQUARE(dtemp); |
1381 | #endif |
1382 | set_new_values(STATE_SYNC, offset, recv_time); |
1383 | break; |
1384 | } |
1385 | if (G.stratum != p->lastpkt_stratum + 1) { |
1386 | G.stratum = p->lastpkt_stratum + 1; |
1387 | run_script("stratum", offset); |
1388 | } |
1389 | } |
1390 | |
1391 | G.reftime = G.cur_time; |
1392 | G.ntp_status = p->lastpkt_status; |
1393 | G.refid = p->lastpkt_refid; |
1394 | G.rootdelay = p->lastpkt_rootdelay + p->lastpkt_delay; |
1395 | dtemp = p->filter_jitter; // SQRT(SQUARE(p->filter_jitter) + SQUARE(G.cluster_jitter)); |
1396 | dtemp += MAXD(p->filter_dispersion + FREQ_TOLERANCE * (G.cur_time - p->lastpkt_recv_time) + abs_offset, MINDISP); |
1397 | G.rootdisp = p->lastpkt_rootdisp + dtemp; |
1398 | VERB3 bb_error_msg("updating leap/refid/reftime/rootdisp from peer %s", p->p_dotted); |
1399 | |
1400 | /* We are in STATE_SYNC now, but did not do adjtimex yet. |
1401 | * (Any other state does not reach this, they all return earlier) |
1402 | * By this time, freq_drift and G.last_update_offset are set |
1403 | * to values suitable for adjtimex. |
1404 | */ |
1405 | #if !USING_KERNEL_PLL_LOOP |
1406 | /* Calculate the new frequency drift and frequency stability (wander). |
1407 | * Compute the clock wander as the RMS of exponentially weighted |
1408 | * frequency differences. This is not used directly, but can, |
1409 | * along with the jitter, be a highly useful monitoring and |
1410 | * debugging tool. |
1411 | */ |
1412 | dtemp = G.discipline_freq_drift + freq_drift; |
1413 | G.discipline_freq_drift = MAXD(MIND(MAXDRIFT, dtemp), -MAXDRIFT); |
1414 | etemp = SQUARE(G.discipline_wander); |
1415 | dtemp = SQUARE(dtemp); |
1416 | G.discipline_wander = SQRT(etemp + (dtemp - etemp) / AVG); |
1417 | |
1418 | VERB3 bb_error_msg("discipline freq_drift=%.9f(int:%ld corr:%e) wander=%f", |
1419 | G.discipline_freq_drift, |
1420 | (long)(G.discipline_freq_drift * 65536e6), |
1421 | freq_drift, |
1422 | G.discipline_wander); |
1423 | #endif |
1424 | VERB3 { |
1425 | memset(&tmx, 0, sizeof(tmx)); |
1426 | if (adjtimex(&tmx) < 0) |
1427 | bb_perror_msg_and_die("adjtimex"); |
1428 | VERB3 bb_error_msg("p adjtimex freq:%ld offset:%ld constant:%ld status:0x%x", |
1429 | tmx.freq, tmx.offset, tmx.constant, tmx.status); |
1430 | } |
1431 | |
1432 | memset(&tmx, 0, sizeof(tmx)); |
1433 | #if 0 |
1434 | //doesn't work, offset remains 0 (!) in kernel: |
1435 | //ntpd: set adjtimex freq:1786097 tmx.offset:77487 |
1436 | //ntpd: prev adjtimex freq:1786097 tmx.offset:0 |
1437 | //ntpd: cur adjtimex freq:1786097 tmx.offset:0 |
1438 | tmx.modes = ADJ_FREQUENCY | ADJ_OFFSET; |
1439 | /* 65536 is one ppm */ |
1440 | tmx.freq = G.discipline_freq_drift * 65536e6; |
1441 | tmx.offset = G.last_update_offset * 1000000; /* usec */ |
1442 | #endif |
1443 | tmx.modes = ADJ_OFFSET | ADJ_STATUS | ADJ_TIMECONST;// | ADJ_MAXERROR | ADJ_ESTERROR; |
1444 | tmx.offset = (G.last_update_offset * 1000000); /* usec */ |
1445 | /* + (G.last_update_offset < 0 ? -0.5 : 0.5) - too small to bother */ |
1446 | tmx.status = STA_PLL; |
1447 | if (G.ntp_status & LI_PLUSSEC) |
1448 | tmx.status |= STA_INS; |
1449 | if (G.ntp_status & LI_MINUSSEC) |
1450 | tmx.status |= STA_DEL; |
1451 | tmx.constant = G.poll_exp - 4; |
1452 | //tmx.esterror = (u_int32)(clock_jitter * 1e6); |
1453 | //tmx.maxerror = (u_int32)((sys_rootdelay / 2 + sys_rootdisp) * 1e6); |
1454 | rc = adjtimex(&tmx); |
1455 | if (rc < 0) |
1456 | bb_perror_msg_and_die("adjtimex"); |
1457 | /* NB: here kernel returns constant == G.poll_exp, not == G.poll_exp - 4. |
1458 | * Not sure why. Perhaps it is normal. |
1459 | */ |
1460 | VERB3 bb_error_msg("adjtimex:%d freq:%ld offset:%ld constant:%ld status:0x%x", |
1461 | rc, tmx.freq, tmx.offset, tmx.constant, tmx.status); |
1462 | #if 0 |
1463 | VERB3 { |
1464 | /* always gives the same output as above msg */ |
1465 | memset(&tmx, 0, sizeof(tmx)); |
1466 | if (adjtimex(&tmx) < 0) |
1467 | bb_perror_msg_and_die("adjtimex"); |
1468 | VERB3 bb_error_msg("c adjtimex freq:%ld offset:%ld constant:%ld status:0x%x", |
1469 | tmx.freq, tmx.offset, tmx.constant, tmx.status); |
1470 | } |
1471 | #endif |
1472 | G.kernel_freq_drift = tmx.freq / 65536; |
1473 | VERB2 bb_error_msg("update peer:%s, offset:%f, clock drift:%ld ppm", |
1474 | p->p_dotted, G.last_update_offset, G.kernel_freq_drift); |
1475 | |
1476 | return 1; /* "ok to increase poll interval" */ |
1477 | } |
1478 | |
1479 | |
1480 | /* |
1481 | * We've got a new reply packet from a peer, process it |
1482 | * (helpers first) |
1483 | */ |
1484 | static unsigned |
1485 | retry_interval(void) |
1486 | { |
1487 | /* Local problem, want to retry soon */ |
1488 | unsigned interval, r; |
1489 | interval = RETRY_INTERVAL; |
1490 | r = random(); |
1491 | interval += r % (unsigned)(RETRY_INTERVAL / 4); |
1492 | VERB3 bb_error_msg("chose retry interval:%u", interval); |
1493 | return interval; |
1494 | } |
1495 | static unsigned |
1496 | poll_interval(int exponent) |
1497 | { |
1498 | unsigned interval, r; |
1499 | exponent = G.poll_exp + exponent; |
1500 | if (exponent < 0) |
1501 | exponent = 0; |
1502 | interval = 1 << exponent; |
1503 | r = random(); |
1504 | interval += ((r & (interval-1)) >> 4) + ((r >> 8) & 1); /* + 1/16 of interval, max */ |
1505 | VERB3 bb_error_msg("chose poll interval:%u (poll_exp:%d exp:%d)", interval, G.poll_exp, exponent); |
1506 | return interval; |
1507 | } |
1508 | static NOINLINE void |
1509 | recv_and_process_peer_pkt(peer_t *p) |
1510 | { |
1511 | int rc; |
1512 | ssize_t size; |
1513 | msg_t msg; |
1514 | double T1, T2, T3, T4; |
1515 | unsigned interval; |
1516 | datapoint_t *datapoint; |
1517 | peer_t *q; |
1518 | |
1519 | /* We can recvfrom here and check from.IP, but some multihomed |
1520 | * ntp servers reply from their *other IP*. |
1521 | * TODO: maybe we should check at least what we can: from.port == 123? |
1522 | */ |
1523 | size = recv(p->p_fd, &msg, sizeof(msg), MSG_DONTWAIT); |
1524 | if (size == -1) { |
1525 | bb_perror_msg("recv(%s) error", p->p_dotted); |
1526 | if (errno == EHOSTUNREACH || errno == EHOSTDOWN |
1527 | || errno == ENETUNREACH || errno == ENETDOWN |
1528 | || errno == ECONNREFUSED || errno == EADDRNOTAVAIL |
1529 | || errno == EAGAIN |
1530 | ) { |
1531 | //TODO: always do this? |
1532 | interval = retry_interval(); |
1533 | goto set_next_and_close_sock; |
1534 | } |
1535 | xfunc_die(); |
1536 | } |
1537 | |
1538 | if (size != NTP_MSGSIZE_NOAUTH && size != NTP_MSGSIZE) { |
1539 | bb_error_msg("malformed packet received from %s", p->p_dotted); |
1540 | goto bail; |
1541 | } |
1542 | |
1543 | if (msg.m_orgtime.int_partl != p->p_xmt_msg.m_xmttime.int_partl |
1544 | || msg.m_orgtime.fractionl != p->p_xmt_msg.m_xmttime.fractionl |
1545 | ) { |
1546 | goto bail; |
1547 | } |
1548 | |
1549 | if ((msg.m_status & LI_ALARM) == LI_ALARM |
1550 | || msg.m_stratum == 0 |
1551 | || msg.m_stratum > NTP_MAXSTRATUM |
1552 | ) { |
1553 | // TODO: stratum 0 responses may have commands in 32-bit m_refid field: |
1554 | // "DENY", "RSTR" - peer does not like us at all |
1555 | // "RATE" - peer is overloaded, reduce polling freq |
1556 | interval = poll_interval(0); |
1557 | bb_error_msg("reply from %s: not synced, next query in %us", p->p_dotted, interval); |
1558 | goto set_next_and_close_sock; |
1559 | } |
1560 | |
1561 | // /* Verify valid root distance */ |
1562 | // if (msg.m_rootdelay / 2 + msg.m_rootdisp >= MAXDISP || p->lastpkt_reftime > msg.m_xmt) |
1563 | // return; /* invalid header values */ |
1564 | |
1565 | p->lastpkt_status = msg.m_status; |
1566 | p->lastpkt_stratum = msg.m_stratum; |
1567 | p->lastpkt_rootdelay = sfp_to_d(msg.m_rootdelay); |
1568 | p->lastpkt_rootdisp = sfp_to_d(msg.m_rootdisp); |
1569 | p->lastpkt_refid = msg.m_refid; |
1570 | |
1571 | /* |
1572 | * From RFC 2030 (with a correction to the delay math): |
1573 | * |
1574 | * Timestamp Name ID When Generated |
1575 | * ------------------------------------------------------------ |
1576 | * Originate Timestamp T1 time request sent by client |
1577 | * Receive Timestamp T2 time request received by server |
1578 | * Transmit Timestamp T3 time reply sent by server |
1579 | * Destination Timestamp T4 time reply received by client |
1580 | * |
1581 | * The roundtrip delay and local clock offset are defined as |
1582 | * |
1583 | * delay = (T4 - T1) - (T3 - T2); offset = ((T2 - T1) + (T3 - T4)) / 2 |
1584 | */ |
1585 | T1 = p->p_xmttime; |
1586 | T2 = lfp_to_d(msg.m_rectime); |
1587 | T3 = lfp_to_d(msg.m_xmttime); |
1588 | T4 = G.cur_time; |
1589 | |
1590 | p->lastpkt_recv_time = T4; |
1591 | |
1592 | VERB5 bb_error_msg("%s->lastpkt_recv_time=%f", p->p_dotted, p->lastpkt_recv_time); |
1593 | p->datapoint_idx = p->reachable_bits ? (p->datapoint_idx + 1) % NUM_DATAPOINTS : 0; |
1594 | datapoint = &p->filter_datapoint[p->datapoint_idx]; |
1595 | datapoint->d_recv_time = T4; |
1596 | datapoint->d_offset = ((T2 - T1) + (T3 - T4)) / 2; |
1597 | /* The delay calculation is a special case. In cases where the |
1598 | * server and client clocks are running at different rates and |
1599 | * with very fast networks, the delay can appear negative. In |
1600 | * order to avoid violating the Principle of Least Astonishment, |
1601 | * the delay is clamped not less than the system precision. |
1602 | */ |
1603 | p->lastpkt_delay = (T4 - T1) - (T3 - T2); |
1604 | if (p->lastpkt_delay < G_precision_sec) |
1605 | p->lastpkt_delay = G_precision_sec; |
1606 | datapoint->d_dispersion = LOG2D(msg.m_precision_exp) + G_precision_sec; |
1607 | if (!p->reachable_bits) { |
1608 | /* 1st datapoint ever - replicate offset in every element */ |
1609 | int i; |
1610 | for (i = 1; i < NUM_DATAPOINTS; i++) { |
1611 | p->filter_datapoint[i].d_offset = datapoint->d_offset; |
1612 | } |
1613 | } |
1614 | |
1615 | p->reachable_bits |= 1; |
1616 | if ((MAX_VERBOSE && G.verbose) || (option_mask32 & OPT_w)) { |
1617 | bb_error_msg("reply from %s: reach 0x%02x offset %f delay %f status 0x%02x strat %d refid 0x%08x rootdelay %f", |
1618 | p->p_dotted, |
1619 | p->reachable_bits, |
1620 | datapoint->d_offset, |
1621 | p->lastpkt_delay, |
1622 | p->lastpkt_status, |
1623 | p->lastpkt_stratum, |
1624 | p->lastpkt_refid, |
1625 | p->lastpkt_rootdelay |
1626 | /* not shown: m_ppoll, m_precision_exp, m_rootdisp, |
1627 | * m_reftime, m_orgtime, m_rectime, m_xmttime |
1628 | */ |
1629 | ); |
1630 | } |
1631 | |
1632 | /* Muck with statictics and update the clock */ |
1633 | filter_datapoints(p); |
1634 | q = select_and_cluster(); |
1635 | rc = -1; |
1636 | if (q) { |
1637 | rc = 0; |
1638 | if (!(option_mask32 & OPT_w)) { |
1639 | rc = update_local_clock(q); |
1640 | /* If drift is dangerously large, immediately |
1641 | * drop poll interval one step down. |
1642 | */ |
1643 | if (fabs(q->filter_offset) >= POLLDOWN_OFFSET) { |
1644 | VERB3 bb_error_msg("offset:%f > POLLDOWN_OFFSET", q->filter_offset); |
1645 | goto poll_down; |
1646 | } |
1647 | } |
1648 | } |
1649 | /* else: no peer selected, rc = -1: we want to poll more often */ |
1650 | |
1651 | if (rc != 0) { |
1652 | /* Adjust the poll interval by comparing the current offset |
1653 | * with the clock jitter. If the offset is less than |
1654 | * the clock jitter times a constant, then the averaging interval |
1655 | * is increased, otherwise it is decreased. A bit of hysteresis |
1656 | * helps calm the dance. Works best using burst mode. |
1657 | */ |
1658 | VERB4 if (rc > 0) { |
1659 | bb_error_msg("offset:%f POLLADJ_GATE*discipline_jitter:%f poll:%s", |
1660 | q->filter_offset, POLLADJ_GATE * G.discipline_jitter, |
1661 | fabs(q->filter_offset) < POLLADJ_GATE * G.discipline_jitter |
1662 | ? "grows" : "falls" |
1663 | ); |
1664 | } |
1665 | if (rc > 0 && fabs(q->filter_offset) < POLLADJ_GATE * G.discipline_jitter) { |
1666 | /* was += G.poll_exp but it is a bit |
1667 | * too optimistic for my taste at high poll_exp's */ |
1668 | G.polladj_count += MINPOLL; |
1669 | if (G.polladj_count > POLLADJ_LIMIT) { |
1670 | G.polladj_count = 0; |
1671 | if (G.poll_exp < MAXPOLL) { |
1672 | G.poll_exp++; |
1673 | VERB3 bb_error_msg("polladj: discipline_jitter:%f ++poll_exp=%d", |
1674 | G.discipline_jitter, G.poll_exp); |
1675 | } |
1676 | } else { |
1677 | VERB3 bb_error_msg("polladj: incr:%d", G.polladj_count); |
1678 | } |
1679 | } else { |
1680 | G.polladj_count -= G.poll_exp * 2; |
1681 | if (G.polladj_count < -POLLADJ_LIMIT || G.poll_exp >= BIGPOLL) { |
1682 | poll_down: |
1683 | G.polladj_count = 0; |
1684 | if (G.poll_exp > MINPOLL) { |
1685 | llist_t *item; |
1686 | |
1687 | G.poll_exp--; |
1688 | /* Correct p->next_action_time in each peer |
1689 | * which waits for sending, so that they send earlier. |
1690 | * Old pp->next_action_time are on the order |
1691 | * of t + (1 << old_poll_exp) + small_random, |
1692 | * we simply need to subtract ~half of that. |
1693 | */ |
1694 | for (item = G.ntp_peers; item != NULL; item = item->link) { |
1695 | peer_t *pp = (peer_t *) item->data; |
1696 | if (pp->p_fd < 0) |
1697 | pp->next_action_time -= (1 << G.poll_exp); |
1698 | } |
1699 | VERB3 bb_error_msg("polladj: discipline_jitter:%f --poll_exp=%d", |
1700 | G.discipline_jitter, G.poll_exp); |
1701 | } |
1702 | } else { |
1703 | VERB3 bb_error_msg("polladj: decr:%d", G.polladj_count); |
1704 | } |
1705 | } |
1706 | } |
1707 | |
1708 | /* Decide when to send new query for this peer */ |
1709 | interval = poll_interval(0); |
1710 | |
1711 | set_next_and_close_sock: |
1712 | set_next(p, interval); |
1713 | /* We do not expect any more packets from this peer for now. |
1714 | * Closing the socket informs kernel about it. |
1715 | * We open a new socket when we send a new query. |
1716 | */ |
1717 | close(p->p_fd); |
1718 | p->p_fd = -1; |
1719 | bail: |
1720 | return; |
1721 | } |
1722 | |
1723 | #if ENABLE_FEATURE_NTPD_SERVER |
1724 | static NOINLINE void |
1725 | recv_and_process_client_pkt(void /*int fd*/) |
1726 | { |
1727 | ssize_t size; |
1728 | uint8_t version; |
1729 | len_and_sockaddr *to; |
1730 | struct sockaddr *from; |
1731 | msg_t msg; |
1732 | uint8_t query_status; |
1733 | l_fixedpt_t query_xmttime; |
1734 | |
1735 | to = get_sock_lsa(G.listen_fd); |
1736 | from = xzalloc(to->len); |
1737 | |
1738 | size = recv_from_to(G.listen_fd, &msg, sizeof(msg), MSG_DONTWAIT, from, &to->u.sa, to->len); |
1739 | if (size != NTP_MSGSIZE_NOAUTH && size != NTP_MSGSIZE) { |
1740 | char *addr; |
1741 | if (size < 0) { |
1742 | if (errno == EAGAIN) |
1743 | goto bail; |
1744 | bb_perror_msg_and_die("recv"); |
1745 | } |
1746 | addr = xmalloc_sockaddr2dotted_noport(from); |
1747 | bb_error_msg("malformed packet received from %s: size %u", addr, (int)size); |
1748 | free(addr); |
1749 | goto bail; |
1750 | } |
1751 | |
1752 | query_status = msg.m_status; |
1753 | query_xmttime = msg.m_xmttime; |
1754 | |
1755 | /* Build a reply packet */ |
1756 | memset(&msg, 0, sizeof(msg)); |
1757 | msg.m_status = G.stratum < MAXSTRAT ? G.ntp_status : LI_ALARM; |
1758 | msg.m_status |= (query_status & VERSION_MASK); |
1759 | msg.m_status |= ((query_status & MODE_MASK) == MODE_CLIENT) ? |
1760 | MODE_SERVER : MODE_SYM_PAS; |
1761 | msg.m_stratum = G.stratum; |
1762 | msg.m_ppoll = G.poll_exp; |
1763 | msg.m_precision_exp = G_precision_exp; |
1764 | /* this time was obtained between poll() and recv() */ |
1765 | msg.m_rectime = d_to_lfp(G.cur_time); |
1766 | msg.m_xmttime = d_to_lfp(gettime1900d()); /* this instant */ |
1767 | msg.m_reftime = d_to_lfp(G.reftime); |
1768 | msg.m_orgtime = query_xmttime; |
1769 | msg.m_rootdelay = d_to_sfp(G.rootdelay); |
1770 | //simple code does not do this, fix simple code! |
1771 | msg.m_rootdisp = d_to_sfp(G.rootdisp); |
1772 | version = (query_status & VERSION_MASK); /* ... >> VERSION_SHIFT - done below instead */ |
1773 | msg.m_refid = G.refid; // (version > (3 << VERSION_SHIFT)) ? G.refid : G.refid3; |
1774 | |
1775 | /* We reply from the local address packet was sent to, |
1776 | * this makes to/from look swapped here: */ |
1777 | do_sendto(G.listen_fd, |
1778 | /*from:*/ &to->u.sa, /*to:*/ from, /*addrlen:*/ to->len, |
1779 | &msg, size); |
1780 | |
1781 | bail: |
1782 | free(to); |
1783 | free(from); |
1784 | } |
1785 | #endif |
1786 | |
1787 | /* Upstream ntpd's options: |
1788 | * |
1789 | * -4 Force DNS resolution of host names to the IPv4 namespace. |
1790 | * -6 Force DNS resolution of host names to the IPv6 namespace. |
1791 | * -a Require cryptographic authentication for broadcast client, |
1792 | * multicast client and symmetric passive associations. |
1793 | * This is the default. |
1794 | * -A Do not require cryptographic authentication for broadcast client, |
1795 | * multicast client and symmetric passive associations. |
1796 | * This is almost never a good idea. |
1797 | * -b Enable the client to synchronize to broadcast servers. |
1798 | * -c conffile |
1799 | * Specify the name and path of the configuration file, |
1800 | * default /etc/ntp.conf |
1801 | * -d Specify debugging mode. This option may occur more than once, |
1802 | * with each occurrence indicating greater detail of display. |
1803 | * -D level |
1804 | * Specify debugging level directly. |
1805 | * -f driftfile |
1806 | * Specify the name and path of the frequency file. |
1807 | * This is the same operation as the "driftfile FILE" |
1808 | * configuration command. |
1809 | * -g Normally, ntpd exits with a message to the system log |
1810 | * if the offset exceeds the panic threshold, which is 1000 s |
1811 | * by default. This option allows the time to be set to any value |
1812 | * without restriction; however, this can happen only once. |
1813 | * If the threshold is exceeded after that, ntpd will exit |
1814 | * with a message to the system log. This option can be used |
1815 | * with the -q and -x options. See the tinker command for other options. |
1816 | * -i jaildir |
1817 | * Chroot the server to the directory jaildir. This option also implies |
1818 | * that the server attempts to drop root privileges at startup |
1819 | * (otherwise, chroot gives very little additional security). |
1820 | * You may need to also specify a -u option. |
1821 | * -k keyfile |
1822 | * Specify the name and path of the symmetric key file, |
1823 | * default /etc/ntp/keys. This is the same operation |
1824 | * as the "keys FILE" configuration command. |
1825 | * -l logfile |
1826 | * Specify the name and path of the log file. The default |
1827 | * is the system log file. This is the same operation as |
1828 | * the "logfile FILE" configuration command. |
1829 | * -L Do not listen to virtual IPs. The default is to listen. |
1830 | * -n Don't fork. |
1831 | * -N To the extent permitted by the operating system, |
1832 | * run the ntpd at the highest priority. |
1833 | * -p pidfile |
1834 | * Specify the name and path of the file used to record the ntpd |
1835 | * process ID. This is the same operation as the "pidfile FILE" |
1836 | * configuration command. |
1837 | * -P priority |
1838 | * To the extent permitted by the operating system, |
1839 | * run the ntpd at the specified priority. |
1840 | * -q Exit the ntpd just after the first time the clock is set. |
1841 | * This behavior mimics that of the ntpdate program, which is |
1842 | * to be retired. The -g and -x options can be used with this option. |
1843 | * Note: The kernel time discipline is disabled with this option. |
1844 | * -r broadcastdelay |
1845 | * Specify the default propagation delay from the broadcast/multicast |
1846 | * server to this client. This is necessary only if the delay |
1847 | * cannot be computed automatically by the protocol. |
1848 | * -s statsdir |
1849 | * Specify the directory path for files created by the statistics |
1850 | * facility. This is the same operation as the "statsdir DIR" |
1851 | * configuration command. |
1852 | * -t key |
1853 | * Add a key number to the trusted key list. This option can occur |
1854 | * more than once. |
1855 | * -u user[:group] |
1856 | * Specify a user, and optionally a group, to switch to. |
1857 | * -v variable |
1858 | * -V variable |
1859 | * Add a system variable listed by default. |
1860 | * -x Normally, the time is slewed if the offset is less than the step |
1861 | * threshold, which is 128 ms by default, and stepped if above |
1862 | * the threshold. This option sets the threshold to 600 s, which is |
1863 | * well within the accuracy window to set the clock manually. |
1864 | * Note: since the slew rate of typical Unix kernels is limited |
1865 | * to 0.5 ms/s, each second of adjustment requires an amortization |
1866 | * interval of 2000 s. Thus, an adjustment as much as 600 s |
1867 | * will take almost 14 days to complete. This option can be used |
1868 | * with the -g and -q options. See the tinker command for other options. |
1869 | * Note: The kernel time discipline is disabled with this option. |
1870 | */ |
1871 | |
1872 | /* By doing init in a separate function we decrease stack usage |
1873 | * in main loop. |
1874 | */ |
1875 | static NOINLINE void ntp_init(char **argv) |
1876 | { |
1877 | unsigned opts; |
1878 | llist_t *peers; |
1879 | |
1880 | srandom(getpid()); |
1881 | |
1882 | if (getuid()) |
1883 | bb_error_msg_and_die(bb_msg_you_must_be_root); |
1884 | |
1885 | /* Set some globals */ |
1886 | G.stratum = MAXSTRAT; |
1887 | if (BURSTPOLL != 0) |
1888 | G.poll_exp = BURSTPOLL; /* speeds up initial sync */ |
1889 | G.last_script_run = G.reftime = G.last_update_recv_time = gettime1900d(); /* sets G.cur_time too */ |
1890 | |
1891 | /* Parse options */ |
1892 | peers = NULL; |
1893 | opt_complementary = "dd:p::wn"; /* d: counter; p: list; -w implies -n */ |
1894 | opts = getopt32(argv, |
1895 | "nqNx" /* compat */ |
1896 | "wp:S:"IF_FEATURE_NTPD_SERVER("l") /* NOT compat */ |
1897 | "d" /* compat */ |
1898 | "46aAbgL", /* compat, ignored */ |
1899 | &peers, &G.script_name, &G.verbose); |
1900 | if (!(opts & (OPT_p|OPT_l))) |
1901 | bb_show_usage(); |
1902 | // if (opts & OPT_x) /* disable stepping, only slew is allowed */ |
1903 | // G.time_was_stepped = 1; |
1904 | while (peers) |
1905 | add_peers(llist_pop(&peers)); |
1906 | if (!(opts & OPT_n)) { |
1907 | bb_daemonize_or_rexec(DAEMON_DEVNULL_STDIO, argv); |
1908 | logmode = LOGMODE_NONE; |
1909 | } |
1910 | #if ENABLE_FEATURE_NTPD_SERVER |
1911 | G.listen_fd = -1; |
1912 | if (opts & OPT_l) { |
1913 | G.listen_fd = create_and_bind_dgram_or_die(NULL, 123); |
1914 | socket_want_pktinfo(G.listen_fd); |
1915 | setsockopt(G.listen_fd, IPPROTO_IP, IP_TOS, &const_IPTOS_LOWDELAY, sizeof(const_IPTOS_LOWDELAY)); |
1916 | } |
1917 | #endif |
1918 | /* I hesitate to set -20 prio. -15 should be high enough for timekeeping */ |
1919 | if (opts & OPT_N) |
1920 | setpriority(PRIO_PROCESS, 0, -15); |
1921 | |
1922 | bb_signals((1 << SIGTERM) | (1 << SIGINT), record_signo); |
1923 | /* Removed SIGHUP here: */ |
1924 | bb_signals((1 << SIGPIPE) | (1 << SIGCHLD), SIG_IGN); |
1925 | } |
1926 | |
1927 | int ntpd_main(int argc UNUSED_PARAM, char **argv) MAIN_EXTERNALLY_VISIBLE; |
1928 | int ntpd_main(int argc UNUSED_PARAM, char **argv) |
1929 | { |
1930 | #undef G |
1931 | struct globals G; |
1932 | struct pollfd *pfd; |
1933 | peer_t **idx2peer; |
1934 | unsigned cnt; |
1935 | |
1936 | memset(&G, 0, sizeof(G)); |
1937 | SET_PTR_TO_GLOBALS(&G); |
1938 | |
1939 | ntp_init(argv); |
1940 | |
1941 | /* If ENABLE_FEATURE_NTPD_SERVER, + 1 for listen_fd: */ |
1942 | cnt = G.peer_cnt + ENABLE_FEATURE_NTPD_SERVER; |
1943 | idx2peer = xzalloc(sizeof(idx2peer[0]) * cnt); |
1944 | pfd = xzalloc(sizeof(pfd[0]) * cnt); |
1945 | |
1946 | /* Countdown: we never sync before we sent INITIAL_SAMLPES+1 |
1947 | * packets to each peer. |
1948 | * NB: if some peer is not responding, we may end up sending |
1949 | * fewer packets to it and more to other peers. |
1950 | * NB2: sync usually happens using INITIAL_SAMLPES packets, |
1951 | * since last reply does not come back instantaneously. |
1952 | */ |
1953 | cnt = G.peer_cnt * (INITIAL_SAMLPES + 1); |
1954 | |
1955 | while (!bb_got_signal) { |
1956 | llist_t *item; |
1957 | unsigned i, j; |
1958 | int nfds, timeout; |
1959 | double nextaction; |
1960 | |
1961 | /* Nothing between here and poll() blocks for any significant time */ |
1962 | |
1963 | nextaction = G.cur_time + 3600; |
1964 | |
1965 | i = 0; |
1966 | #if ENABLE_FEATURE_NTPD_SERVER |
1967 | if (G.listen_fd != -1) { |
1968 | pfd[0].fd = G.listen_fd; |
1969 | pfd[0].events = POLLIN; |
1970 | i++; |
1971 | } |
1972 | #endif |
1973 | /* Pass over peer list, send requests, time out on receives */ |
1974 | for (item = G.ntp_peers; item != NULL; item = item->link) { |
1975 | peer_t *p = (peer_t *) item->data; |
1976 | |
1977 | if (p->next_action_time <= G.cur_time) { |
1978 | if (p->p_fd == -1) { |
1979 | /* Time to send new req */ |
1980 | if (--cnt == 0) { |
1981 | G.initial_poll_complete = 1; |
1982 | } |
1983 | send_query_to_peer(p); |
1984 | } else { |
1985 | /* Timed out waiting for reply */ |
1986 | close(p->p_fd); |
1987 | p->p_fd = -1; |
1988 | timeout = poll_interval(-2); /* -2: try a bit sooner */ |
1989 | bb_error_msg("timed out waiting for %s, reach 0x%02x, next query in %us", |
1990 | p->p_dotted, p->reachable_bits, timeout); |
1991 | set_next(p, timeout); |
1992 | } |
1993 | } |
1994 | |
1995 | if (p->next_action_time < nextaction) |
1996 | nextaction = p->next_action_time; |
1997 | |
1998 | if (p->p_fd >= 0) { |
1999 | /* Wait for reply from this peer */ |
2000 | pfd[i].fd = p->p_fd; |
2001 | pfd[i].events = POLLIN; |
2002 | idx2peer[i] = p; |
2003 | i++; |
2004 | } |
2005 | } |
2006 | |
2007 | timeout = nextaction - G.cur_time; |
2008 | if (timeout < 0) |
2009 | timeout = 0; |
2010 | timeout++; /* (nextaction - G.cur_time) rounds down, compensating */ |
2011 | |
2012 | /* Here we may block */ |
2013 | VERB2 bb_error_msg("poll %us, sockets:%u, poll interval:%us", timeout, i, 1 << G.poll_exp); |
2014 | nfds = poll(pfd, i, timeout * 1000); |
2015 | gettime1900d(); /* sets G.cur_time */ |
2016 | if (nfds <= 0) { |
2017 | if (G.script_name && G.cur_time - G.last_script_run > 11*60) { |
2018 | /* Useful for updating battery-backed RTC and such */ |
2019 | run_script("periodic", G.last_update_offset); |
2020 | gettime1900d(); /* sets G.cur_time */ |
2021 | } |
2022 | continue; |
2023 | } |
2024 | |
2025 | /* Process any received packets */ |
2026 | j = 0; |
2027 | #if ENABLE_FEATURE_NTPD_SERVER |
2028 | if (G.listen_fd != -1) { |
2029 | if (pfd[0].revents /* & (POLLIN|POLLERR)*/) { |
2030 | nfds--; |
2031 | recv_and_process_client_pkt(/*G.listen_fd*/); |
2032 | gettime1900d(); /* sets G.cur_time */ |
2033 | } |
2034 | j = 1; |
2035 | } |
2036 | #endif |
2037 | for (; nfds != 0 && j < i; j++) { |
2038 | if (pfd[j].revents /* & (POLLIN|POLLERR)*/) { |
2039 | nfds--; |
2040 | recv_and_process_peer_pkt(idx2peer[j]); |
2041 | gettime1900d(); /* sets G.cur_time */ |
2042 | } |
2043 | } |
2044 | } /* while (!bb_got_signal) */ |
2045 | |
2046 | kill_myself_with_sig(bb_got_signal); |
2047 | } |
2048 | |
2049 | |
2050 | |
2051 | |
2052 | |
2053 | |
2054 | /*** openntpd-4.6 uses only adjtime, not adjtimex ***/ |
2055 | |
2056 | /*** ntp-4.2.6/ntpd/ntp_loopfilter.c - adjtimex usage ***/ |
2057 | |
2058 | #if 0 |
2059 | static double |
2060 | direct_freq(double fp_offset) |
2061 | { |
2062 | |
2063 | #ifdef KERNEL_PLL |
2064 | /* |
2065 | * If the kernel is enabled, we need the residual offset to |
2066 | * calculate the frequency correction. |
2067 | */ |
2068 | if (pll_control && kern_enable) { |
2069 | memset(&ntv, 0, sizeof(ntv)); |
2070 | ntp_adjtime(&ntv); |
2071 | #ifdef STA_NANO |
2072 | clock_offset = ntv.offset / 1e9; |
2073 | #else /* STA_NANO */ |
2074 | clock_offset = ntv.offset / 1e6; |
2075 | #endif /* STA_NANO */ |
2076 | drift_comp = FREQTOD(ntv.freq); |
2077 | } |
2078 | #endif /* KERNEL_PLL */ |
2079 | set_freq((fp_offset - clock_offset) / (current_time - clock_epoch) + drift_comp); |
2080 | wander_resid = 0; |
2081 | return drift_comp; |
2082 | } |
2083 | |
2084 | static void |
2085 | set_freq(double freq) /* frequency update */ |
2086 | { |
2087 | char tbuf[80]; |
2088 | |
2089 | drift_comp = freq; |
2090 | |
2091 | #ifdef KERNEL_PLL |
2092 | /* |
2093 | * If the kernel is enabled, update the kernel frequency. |
2094 | */ |
2095 | if (pll_control && kern_enable) { |
2096 | memset(&ntv, 0, sizeof(ntv)); |
2097 | ntv.modes = MOD_FREQUENCY; |
2098 | ntv.freq = DTOFREQ(drift_comp); |
2099 | ntp_adjtime(&ntv); |
2100 | snprintf(tbuf, sizeof(tbuf), "kernel %.3f PPM", drift_comp * 1e6); |
2101 | report_event(EVNT_FSET, NULL, tbuf); |
2102 | } else { |
2103 | snprintf(tbuf, sizeof(tbuf), "ntpd %.3f PPM", drift_comp * 1e6); |
2104 | report_event(EVNT_FSET, NULL, tbuf); |
2105 | } |
2106 | #else /* KERNEL_PLL */ |
2107 | snprintf(tbuf, sizeof(tbuf), "ntpd %.3f PPM", drift_comp * 1e6); |
2108 | report_event(EVNT_FSET, NULL, tbuf); |
2109 | #endif /* KERNEL_PLL */ |
2110 | } |
2111 | |
2112 | ... |
2113 | ... |
2114 | ... |
2115 | |
2116 | #ifdef KERNEL_PLL |
2117 | /* |
2118 | * This code segment works when clock adjustments are made using |
2119 | * precision time kernel support and the ntp_adjtime() system |
2120 | * call. This support is available in Solaris 2.6 and later, |
2121 | * Digital Unix 4.0 and later, FreeBSD, Linux and specially |
2122 | * modified kernels for HP-UX 9 and Ultrix 4. In the case of the |
2123 | * DECstation 5000/240 and Alpha AXP, additional kernel |
2124 | * modifications provide a true microsecond clock and nanosecond |
2125 | * clock, respectively. |
2126 | * |
2127 | * Important note: The kernel discipline is used only if the |
2128 | * step threshold is less than 0.5 s, as anything higher can |
2129 | * lead to overflow problems. This might occur if some misguided |
2130 | * lad set the step threshold to something ridiculous. |
2131 | */ |
2132 | if (pll_control && kern_enable) { |
2133 | |
2134 | #define MOD_BITS (MOD_OFFSET | MOD_MAXERROR | MOD_ESTERROR | MOD_STATUS | MOD_TIMECONST) |
2135 | |
2136 | /* |
2137 | * We initialize the structure for the ntp_adjtime() |
2138 | * system call. We have to convert everything to |
2139 | * microseconds or nanoseconds first. Do not update the |
2140 | * system variables if the ext_enable flag is set. In |
2141 | * this case, the external clock driver will update the |
2142 | * variables, which will be read later by the local |
2143 | * clock driver. Afterwards, remember the time and |
2144 | * frequency offsets for jitter and stability values and |
2145 | * to update the frequency file. |
2146 | */ |
2147 | memset(&ntv, 0, sizeof(ntv)); |
2148 | if (ext_enable) { |
2149 | ntv.modes = MOD_STATUS; |
2150 | } else { |
2151 | #ifdef STA_NANO |
2152 | ntv.modes = MOD_BITS | MOD_NANO; |
2153 | #else /* STA_NANO */ |
2154 | ntv.modes = MOD_BITS; |
2155 | #endif /* STA_NANO */ |
2156 | if (clock_offset < 0) |
2157 | dtemp = -.5; |
2158 | else |
2159 | dtemp = .5; |
2160 | #ifdef STA_NANO |
2161 | ntv.offset = (int32)(clock_offset * 1e9 + dtemp); |
2162 | ntv.constant = sys_poll; |
2163 | #else /* STA_NANO */ |
2164 | ntv.offset = (int32)(clock_offset * 1e6 + dtemp); |
2165 | ntv.constant = sys_poll - 4; |
2166 | #endif /* STA_NANO */ |
2167 | ntv.esterror = (u_int32)(clock_jitter * 1e6); |
2168 | ntv.maxerror = (u_int32)((sys_rootdelay / 2 + sys_rootdisp) * 1e6); |
2169 | ntv.status = STA_PLL; |
2170 | |
2171 | /* |
2172 | * Enable/disable the PPS if requested. |
2173 | */ |
2174 | if (pps_enable) { |
2175 | if (!(pll_status & STA_PPSTIME)) |
2176 | report_event(EVNT_KERN, |
2177 | NULL, "PPS enabled"); |
2178 | ntv.status |= STA_PPSTIME | STA_PPSFREQ; |
2179 | } else { |
2180 | if (pll_status & STA_PPSTIME) |
2181 | report_event(EVNT_KERN, |
2182 | NULL, "PPS disabled"); |
2183 | ntv.status &= ~(STA_PPSTIME | |
2184 | STA_PPSFREQ); |
2185 | } |
2186 | if (sys_leap == LEAP_ADDSECOND) |
2187 | ntv.status |= STA_INS; |
2188 | else if (sys_leap == LEAP_DELSECOND) |
2189 | ntv.status |= STA_DEL; |
2190 | } |
2191 | |
2192 | /* |
2193 | * Pass the stuff to the kernel. If it squeals, turn off |
2194 | * the pps. In any case, fetch the kernel offset, |
2195 | * frequency and jitter. |
2196 | */ |
2197 | if (ntp_adjtime(&ntv) == TIME_ERROR) { |
2198 | if (!(ntv.status & STA_PPSSIGNAL)) |
2199 | report_event(EVNT_KERN, NULL, |
2200 | "PPS no signal"); |
2201 | } |
2202 | pll_status = ntv.status; |
2203 | #ifdef STA_NANO |
2204 | clock_offset = ntv.offset / 1e9; |
2205 | #else /* STA_NANO */ |
2206 | clock_offset = ntv.offset / 1e6; |
2207 | #endif /* STA_NANO */ |
2208 | clock_frequency = FREQTOD(ntv.freq); |
2209 | |
2210 | /* |
2211 | * If the kernel PPS is lit, monitor its performance. |
2212 | */ |
2213 | if (ntv.status & STA_PPSTIME) { |
2214 | #ifdef STA_NANO |
2215 | clock_jitter = ntv.jitter / 1e9; |
2216 | #else /* STA_NANO */ |
2217 | clock_jitter = ntv.jitter / 1e6; |
2218 | #endif /* STA_NANO */ |
2219 | } |
2220 | |
2221 | #if defined(STA_NANO) && NTP_API == 4 |
2222 | /* |
2223 | * If the TAI changes, update the kernel TAI. |
2224 | */ |
2225 | if (loop_tai != sys_tai) { |
2226 | loop_tai = sys_tai; |
2227 | ntv.modes = MOD_TAI; |
2228 | ntv.constant = sys_tai; |
2229 | ntp_adjtime(&ntv); |
2230 | } |
2231 | #endif /* STA_NANO */ |
2232 | } |
2233 | #endif /* KERNEL_PLL */ |
2234 | #endif |