path: root/networking/ntpd.c (plain)
blob: 1532669d3b45e723dd080ad44f0831fb383f0d5e

1	/*
2	* NTP client/server, based on OpenNTPD 3.9p1
3	*
4	* Busybox port author: Adam Tkac (C) 2009 <vonsch@gmail.com>
5	*
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>
9	*
10	* OpenNTPd code is licensed under ISC-style licence:
11	*
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.
15	*
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	***********************************************************************
24	*
25	* Parts of OpenNTPD clock syncronization code is replaced by
26	* code which is based on ntp-4.2.6, which carries the following
27	* copyright notice:
28	*
29	* Copyright (c) University of Delaware 1992-2009
30	*
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	***********************************************************************
42	*/
43	//config:config NTPD
44	//config: bool "ntpd"
45	//config: default y
46	//config: select PLATFORM_LINUX
47	//config: help
48	//config: The NTP client/server daemon.
49	//config:
50	//config:config FEATURE_NTPD_SERVER
51	//config: bool "Make ntpd usable as a NTP server"
52	//config: default y
53	//config: depends on NTPD
54	//config: help
55	//config: Make ntpd usable as a NTP server. If you disable this option
56	//config: ntpd will be usable only as a NTP client.
57	//config:
58	//config:config FEATURE_NTPD_CONF
59	//config: bool "Make ntpd understand /etc/ntp.conf"
60	//config: default y
61	//config: depends on NTPD
62	//config: help
63	//config: Make ntpd look in /etc/ntp.conf for peers. Only "server address"
64	//config: is supported.
65
66	//applet:IF_NTPD(APPLET(ntpd, BB_DIR_USR_SBIN, BB_SUID_DROP))
67
68	//kbuild:lib-$(CONFIG_NTPD) += ntpd.o
69
70	//usage:#define ntpd_trivial_usage
71	//usage: "[-dnqNw"IF_FEATURE_NTPD_SERVER("l -I IFACE")"] [-S PROG] [-p PEER]..."
72	//usage:#define ntpd_full_usage "\n\n"
73	//usage: "NTP client/server\n"
74	//usage: "\n -d Verbose"
75	//usage: "\n -n Do not daemonize"
76	//usage: "\n -q Quit after clock is set"
77	//usage: "\n -N Run at high priority"
78	//usage: "\n -w Do not set time (only query peers), implies -n"
79	//usage: "\n -S PROG Run PROG after stepping time, stratum change, and every 11 mins"
80	//usage: "\n -p PEER Obtain time from PEER (may be repeated)"
81	//usage: IF_FEATURE_NTPD_CONF(
82	//usage: "\n If -p is not given, 'server HOST' lines"
83	//usage: "\n from /etc/ntp.conf are used"
84	//usage: )
85	//usage: IF_FEATURE_NTPD_SERVER(
86	//usage: "\n -l Also run as server on port 123"
87	//usage: "\n -I IFACE Bind server to IFACE, implies -l"
88	//usage: )
89
90	// -l and -p options are not compatible with "standard" ntpd:
91	// it has them as "-l logfile" and "-p pidfile".
92	// -S and -w are not compat either, "standard" ntpd has no such opts.
93
94	#include "libbb.h"
95	#include <math.h>
96	#include <netinet/ip.h> /* For IPTOS_LOWDELAY definition */
97	#include <sys/resource.h> /* setpriority */
98
99	#ifdef __BIONIC__
100	#include <linux/timex.h>
101	extern int adjtimex (struct timex *);
102	#else
103	#include <sys/timex.h>
104	#endif
105	#ifndef IPTOS_LOWDELAY
106	# define IPTOS_LOWDELAY 0x10
107	#endif
108
109
110	/* Verbosity control (max level of -dddd options accepted).
111	* max 6 is very talkative (and bloated). 3 is non-bloated,
112	* production level setting.
113	*/
114	#define MAX_VERBOSE 3
115
116
117	/* High-level description of the algorithm:
118	*
119	* We start running with very small poll_exp, BURSTPOLL,
120	* in order to quickly accumulate INITIAL_SAMPLES datapoints
121	* for each peer. Then, time is stepped if the offset is larger
122	* than STEP_THRESHOLD, otherwise it isn't; anyway, we enlarge
123	* poll_exp to MINPOLL and enter frequency measurement step:
124	* we collect new datapoints but ignore them for WATCH_THRESHOLD
125	* seconds. After WATCH_THRESHOLD seconds we look at accumulated
126	* offset and estimate frequency drift.
127	*
128	* (frequency measurement step seems to not be strictly needed,
129	* it is conditionally disabled with USING_INITIAL_FREQ_ESTIMATION
130	* define set to 0)
131	*
132	* After this, we enter "steady state": we collect a datapoint,
133	* we select the best peer, if this datapoint is not a new one
134	* (IOW: if this datapoint isn't for selected peer), sleep
135	* and collect another one; otherwise, use its offset to update
136	* frequency drift, if offset is somewhat large, reduce poll_exp,
137	* otherwise increase poll_exp.
138	*
139	* If offset is larger than STEP_THRESHOLD, which shouldn't normally
140	* happen, we assume that something "bad" happened (computer
141	* was hibernated, someone set totally wrong date, etc),
142	* then the time is stepped, all datapoints are discarded,
143	* and we go back to steady state.
144	*
145	* Made some changes to speed up re-syncing after our clock goes bad
146	* (tested with suspending my laptop):
147	* - if largish offset (>= STEP_THRESHOLD == 1 sec) is seen
148	* from a peer, schedule next query for this peer soon
149	* without drastically lowering poll interval for everybody.
150	* This makes us collect enough data for step much faster:
151	* e.g. at poll = 10 (1024 secs), step was done within 5 minutes
152	* after first reply which indicated that our clock is 14 seconds off.
153	* - on step, do not discard d_dispersion data of the existing datapoints,
154	* do not clear reachable_bits. This prevents discarding first ~8
155	* datapoints after the step.
156	*/
157
158	#define INITIAL_SAMPLES 4 /* how many samples do we want for init */
159	#define BAD_DELAY_GROWTH 4 /* drop packet if its delay grew by more than this */
160
161	#define RETRY_INTERVAL 32 /* on send/recv error, retry in N secs (need to be power of 2) */
162	#define NOREPLY_INTERVAL 512 /* sent, but got no reply: cap next query by this many seconds */
163	#define RESPONSE_INTERVAL 16 /* wait for reply up to N secs */
164
165	/* Step threshold (sec). std ntpd uses 0.128.
166	*/
167	#define STEP_THRESHOLD 1
168	/* Slew threshold (sec): adjtimex() won't accept offsets larger than this.
169	* Using exact power of 2 (1/8) results in smaller code
170	*/
171	#define SLEW_THRESHOLD 0.125
172	/* Stepout threshold (sec). std ntpd uses 900 (11 mins (!)) */
173	#define WATCH_THRESHOLD 128
174	/* NB: set WATCH_THRESHOLD to ~60 when debugging to save time) */
175	//UNUSED: #define PANIC_THRESHOLD 1000 /* panic threshold (sec) */
176
177	/*
178	* If we got |offset| > BIGOFF from a peer, cap next query interval
179	* for this peer by this many seconds:
180	*/
181	#define BIGOFF STEP_THRESHOLD
182	#define BIGOFF_INTERVAL (1 << 7) /* 128 s */
183
184	#define FREQ_TOLERANCE 0.000015 /* frequency tolerance (15 PPM) */
185	#define BURSTPOLL 0 /* initial poll */
186	#define MINPOLL 5 /* minimum poll interval. std ntpd uses 6 (6: 64 sec) */
187	/*
188	* If offset > discipline_jitter * POLLADJ_GATE, and poll interval is > 2^BIGPOLL,
189	* then it is decreased _at once_. (If <= 2^BIGPOLL, it will be decreased _eventually_).
190	*/
191	#define BIGPOLL 9 /* 2^9 sec ~= 8.5 min */
192	#define MAXPOLL 12 /* maximum poll interval (12: 1.1h, 17: 36.4h). std ntpd uses 17 */
193	/*
194	* Actively lower poll when we see such big offsets.
195	* With SLEW_THRESHOLD = 0.125, it means we try to sync more aggressively
196	* if offset increases over ~0.04 sec
197	*/
198	//#define POLLDOWN_OFFSET (SLEW_THRESHOLD / 3)
199	#define MINDISP 0.01 /* minimum dispersion (sec) */
200	#define MAXDISP 16 /* maximum dispersion (sec) */
201	#define MAXSTRAT 16 /* maximum stratum (infinity metric) */
202	#define MAXDIST 1 /* distance threshold (sec) */
203	#define MIN_SELECTED 1 /* minimum intersection survivors */
204	#define MIN_CLUSTERED 3 /* minimum cluster survivors */
205
206	#define MAXDRIFT 0.000500 /* frequency drift we can correct (500 PPM) */
207
208	/* Poll-adjust threshold.
209	* When we see that offset is small enough compared to discipline jitter,
210	* we grow a counter: += MINPOLL. When counter goes over POLLADJ_LIMIT,
211	* we poll_exp++. If offset isn't small, counter -= poll_exp*2,
212	* and when it goes below -POLLADJ_LIMIT, we poll_exp--.
213	* (Bumped from 30 to 40 since otherwise I often see poll_exp going *2* steps down)
214	*/
215	#define POLLADJ_LIMIT 40
216	/* If offset < discipline_jitter * POLLADJ_GATE, then we decide to increase
217	* poll interval (we think we can't improve timekeeping
218	* by staying at smaller poll).
219	*/
220	#define POLLADJ_GATE 4
221	#define TIMECONST_HACK_GATE 2
222	/* Compromise Allan intercept (sec). doc uses 1500, std ntpd uses 512 */
223	#define ALLAN 512
224	/* PLL loop gain */
225	#define PLL 65536
226	/* FLL loop gain [why it depends on MAXPOLL??] */
227	#define FLL (MAXPOLL + 1)
228	/* Parameter averaging constant */
229	#define AVG 4
230
231
232	enum {
233	NTP_VERSION = 4,
234	NTP_MAXSTRATUM = 15,
235
236	NTP_DIGESTSIZE = 16,
237	NTP_MSGSIZE_NOAUTH = 48,
238	NTP_MSGSIZE = (NTP_MSGSIZE_NOAUTH + 4 + NTP_DIGESTSIZE),
239
240	/* Status Masks */
241	MODE_MASK = (7 << 0),
242	VERSION_MASK = (7 << 3),
243	VERSION_SHIFT = 3,
244	LI_MASK = (3 << 6),
245
246	/* Leap Second Codes (high order two bits of m_status) */
247	LI_NOWARNING = (0 << 6), /* no warning */
248	LI_PLUSSEC = (1 << 6), /* add a second (61 seconds) */
249	LI_MINUSSEC = (2 << 6), /* minus a second (59 seconds) */
250	LI_ALARM = (3 << 6), /* alarm condition */
251
252	/* Mode values */
253	MODE_RES0 = 0, /* reserved */
254	MODE_SYM_ACT = 1, /* symmetric active */
255	MODE_SYM_PAS = 2, /* symmetric passive */
256	MODE_CLIENT = 3, /* client */
257	MODE_SERVER = 4, /* server */
258	MODE_BROADCAST = 5, /* broadcast */
259	MODE_RES1 = 6, /* reserved for NTP control message */
260	MODE_RES2 = 7, /* reserved for private use */
261	};
262
263	//TODO: better base selection
264	#define OFFSET_1900_1970 2208988800UL /* 1970 - 1900 in seconds */
265
266	#define NUM_DATAPOINTS 8
267
268	typedef struct {
269	uint32_t int_partl;
270	uint32_t fractionl;
271	} l_fixedpt_t;
272
273	typedef struct {
274	uint16_t int_parts;
275	uint16_t fractions;
276	} s_fixedpt_t;
277
278	typedef struct {
279	uint8_t m_status; /* status of local clock and leap info */
280	uint8_t m_stratum;
281	uint8_t m_ppoll; /* poll value */
282	int8_t m_precision_exp;
283	s_fixedpt_t m_rootdelay;
284	s_fixedpt_t m_rootdisp;
285	uint32_t m_refid;
286	l_fixedpt_t m_reftime;
287	l_fixedpt_t m_orgtime;
288	l_fixedpt_t m_rectime;
289	l_fixedpt_t m_xmttime;
290	uint32_t m_keyid;
291	uint8_t m_digest[NTP_DIGESTSIZE];
292	} msg_t;
293
294	typedef struct {
295	double d_offset;
296	double d_recv_time;
297	double d_dispersion;
298	} datapoint_t;
299
300	typedef struct {
301	len_and_sockaddr *p_lsa;
302	char *p_dotted;
303	int p_fd;
304	int datapoint_idx;
305	uint32_t lastpkt_refid;
306	uint8_t lastpkt_status;
307	uint8_t lastpkt_stratum;
308	uint8_t reachable_bits;
309	/* when to send new query (if p_fd == -1)
310	* or when receive times out (if p_fd >= 0): */
311	double next_action_time;
312	double p_xmttime;
313	double p_raw_delay;
314	/* p_raw_delay is set even by "high delay" packets */
315	/* lastpkt_delay isn't */
316	double lastpkt_recv_time;
317	double lastpkt_delay;
318	double lastpkt_rootdelay;
319	double lastpkt_rootdisp;
320	/* produced by filter algorithm: */
321	double filter_offset;
322	double filter_dispersion;
323	double filter_jitter;
324	datapoint_t filter_datapoint[NUM_DATAPOINTS];
325	/* last sent packet: */
326	msg_t p_xmt_msg;
327	char p_hostname[1];
328	} peer_t;
329
330
331	#define USING_KERNEL_PLL_LOOP 1
332	#define USING_INITIAL_FREQ_ESTIMATION 0
333
334	enum {
335	OPT_n = (1 << 0),
336	OPT_q = (1 << 1),
337	OPT_N = (1 << 2),
338	OPT_x = (1 << 3),
339	/* Insert new options above this line. */
340	/* Non-compat options: */
341	OPT_w = (1 << 4),
342	OPT_p = (1 << 5),
343	OPT_S = (1 << 6),
344	OPT_l = (1 << 7) * ENABLE_FEATURE_NTPD_SERVER,
345	OPT_I = (1 << 8) * ENABLE_FEATURE_NTPD_SERVER,
346	/* We hijack some bits for other purposes */
347	OPT_qq = (1 << 31),
348	};
349
350	struct globals {
351	double cur_time;
352	/* total round trip delay to currently selected reference clock */
353	double rootdelay;
354	/* reference timestamp: time when the system clock was last set or corrected */
355	double reftime;
356	/* total dispersion to currently selected reference clock */
357	double rootdisp;
358
359	double last_script_run;
360	char *script_name;
361	llist_t *ntp_peers;
362	#if ENABLE_FEATURE_NTPD_SERVER
363	int listen_fd;
364	char *if_name;
365	# define G_listen_fd (G.listen_fd)
366	#else
367	# define G_listen_fd (-1)
368	#endif
369	unsigned verbose;
370	unsigned peer_cnt;
371	/* refid: 32-bit code identifying the particular server or reference clock
372	* in stratum 0 packets this is a four-character ASCII string,
373	* called the kiss code, used for debugging and monitoring
374	* in stratum 1 packets this is a four-character ASCII string
375	* assigned to the reference clock by IANA. Example: "GPS "
376	* in stratum 2+ packets, it's IPv4 address or 4 first bytes
377	* of MD5 hash of IPv6
378	*/
379	uint32_t refid;
380	uint8_t ntp_status;
381	/* precision is defined as the larger of the resolution and time to
382	* read the clock, in log2 units. For instance, the precision of a
383	* mains-frequency clock incrementing at 60 Hz is 16 ms, even when the
384	* system clock hardware representation is to the nanosecond.
385	*
386	* Delays, jitters of various kinds are clamped down to precision.
387	*
388	* If precision_sec is too large, discipline_jitter gets clamped to it
389	* and if offset is smaller than discipline_jitter * POLLADJ_GATE, poll
390	* interval grows even though we really can benefit from staying at
391	* smaller one, collecting non-lagged datapoits and correcting offset.
392	* (Lagged datapoits exist when poll_exp is large but we still have
393	* systematic offset error - the time distance between datapoints
394	* is significant and older datapoints have smaller offsets.
395	* This makes our offset estimation a bit smaller than reality)
396	* Due to this effect, setting G_precision_sec close to
397	* STEP_THRESHOLD isn't such a good idea - offsets may grow
398	* too big and we will step. I observed it with -6.
399	*
400	* OTOH, setting precision_sec far too small would result in futile
401	* attempts to syncronize to an unachievable precision.
402	*
403	* -6 is 1/64 sec, -7 is 1/128 sec and so on.
404	* -8 is 1/256 ~= 0.003906 (worked well for me --vda)
405	* -9 is 1/512 ~= 0.001953 (let's try this for some time)
406	*/
407	#define G_precision_exp -9
408	/*
409	* G_precision_exp is used only for construction outgoing packets.
410	* It's ok to set G_precision_sec to a slightly different value
411	* (One which is "nicer looking" in logs).
412	* Exact value would be (1.0 / (1 << (- G_precision_exp))):
413	*/
414	#define G_precision_sec 0.002
415	uint8_t stratum;
416
417	#define STATE_NSET 0 /* initial state, "nothing is set" */
418	//#define STATE_FSET 1 /* frequency set from file */
419	//#define STATE_SPIK 2 /* spike detected */
420	//#define STATE_FREQ 3 /* initial frequency */
421	#define STATE_SYNC 4 /* clock synchronized (normal operation) */
422	uint8_t discipline_state; // doc calls it c.state
423	uint8_t poll_exp; // s.poll
424	int polladj_count; // c.count
425	long kernel_freq_drift;
426	peer_t *last_update_peer;
427	double last_update_offset; // c.last
428	double last_update_recv_time; // s.t
429	double discipline_jitter; // c.jitter
430	/* Since we only compare it with ints, can simplify code
431	* by not making this variable floating point:
432	*/
433	unsigned offset_to_jitter_ratio;
434	//double cluster_offset; // s.offset
435	//double cluster_jitter; // s.jitter
436	#if !USING_KERNEL_PLL_LOOP
437	double discipline_freq_drift; // c.freq
438	/* Maybe conditionally calculate wander? it's used only for logging */
439	double discipline_wander; // c.wander
440	#endif
441	};
442	#define G (*ptr_to_globals)
443
444
445	#define VERB1 if (MAX_VERBOSE && G.verbose)
446	#define VERB2 if (MAX_VERBOSE >= 2 && G.verbose >= 2)
447	#define VERB3 if (MAX_VERBOSE >= 3 && G.verbose >= 3)
448	#define VERB4 if (MAX_VERBOSE >= 4 && G.verbose >= 4)
449	#define VERB5 if (MAX_VERBOSE >= 5 && G.verbose >= 5)
450	#define VERB6 if (MAX_VERBOSE >= 6 && G.verbose >= 6)
451
452
453	static double LOG2D(int a)
454	{
455	if (a < 0)
456	return 1.0 / (1UL << -a);
457	return 1UL << a;
458	}
459	static ALWAYS_INLINE double SQUARE(double x)
460	{
461	return x * x;
462	}
463	static ALWAYS_INLINE double MAXD(double a, double b)
464	{
465	if (a > b)
466	return a;
467	return b;
468	}
469	static ALWAYS_INLINE double MIND(double a, double b)
470	{
471	if (a < b)
472	return a;
473	return b;
474	}
475	static NOINLINE double my_SQRT(double X)
476	{
477	union {
478	float f;
479	int32_t i;
480	} v;
481	double invsqrt;
482	double Xhalf = X * 0.5;
483
484	/* Fast and good approximation to 1/sqrt(X), black magic */
485	v.f = X;
486	/*v.i = 0x5f3759df - (v.i >> 1);*/
487	v.i = 0x5f375a86 - (v.i >> 1); /* - this constant is slightly better */
488	invsqrt = v.f; /* better than 0.2% accuracy */
489
490	/* Refining it using Newton's method: x1 = x0 - f(x0)/f'(x0)
491	* f(x) = 1/(x*x) - X (f==0 when x = 1/sqrt(X))
492	* f'(x) = -2/(x*x*x)
493	* f(x)/f'(x) = (X - 1/(x*x)) / (2/(x*x*x)) = X*x*x*x/2 - x/2
494	* x1 = x0 - (X*x0*x0*x0/2 - x0/2) = 1.5*x0 - X*x0*x0*x0/2 = x0*(1.5 - (X/2)*x0*x0)
495	*/
496	invsqrt = invsqrt * (1.5 - Xhalf * invsqrt * invsqrt); /* ~0.05% accuracy */
497	/* invsqrt = invsqrt * (1.5 - Xhalf * invsqrt * invsqrt); 2nd iter: ~0.0001% accuracy */
498	/* With 4 iterations, more than half results will be exact,
499	* at 6th iterations result stabilizes with about 72% results exact.
500	* We are well satisfied with 0.05% accuracy.
501	*/
502
503	return X * invsqrt; /* X * 1/sqrt(X) ~= sqrt(X) */
504	}
505	static ALWAYS_INLINE double SQRT(double X)
506	{
507	/* If this arch doesn't use IEEE 754 floats, fall back to using libm */
508	if (sizeof(float) != 4)
509	return sqrt(X);
510
511	/* This avoids needing libm, saves about 0.5k on x86-32 */
512	return my_SQRT(X);
513	}
514
515	static double
516	gettime1900d(void)
517	{
518	struct timeval tv;
519	gettimeofday(&tv, NULL); /* never fails */
520	G.cur_time = tv.tv_sec + (1.0e-6 * tv.tv_usec) + OFFSET_1900_1970;
521	return G.cur_time;
522	}
523
524	static void
525	d_to_tv(double d, struct timeval *tv)
526	{
527	tv->tv_sec = (long)d;
528	tv->tv_usec = (d - tv->tv_sec) * 1000000;
529	}
530
531	static double
532	lfp_to_d(l_fixedpt_t lfp)
533	{
534	double ret;
535	lfp.int_partl = ntohl(lfp.int_partl);
536	lfp.fractionl = ntohl(lfp.fractionl);
537	ret = (double)lfp.int_partl + ((double)lfp.fractionl / UINT_MAX);
538	return ret;
539	}
540	static double
541	sfp_to_d(s_fixedpt_t sfp)
542	{
543	double ret;
544	sfp.int_parts = ntohs(sfp.int_parts);
545	sfp.fractions = ntohs(sfp.fractions);
546	ret = (double)sfp.int_parts + ((double)sfp.fractions / USHRT_MAX);
547	return ret;
548	}
549	#if ENABLE_FEATURE_NTPD_SERVER
550	static l_fixedpt_t
551	d_to_lfp(double d)
552	{
553	l_fixedpt_t lfp;
554	lfp.int_partl = (uint32_t)d;
555	lfp.fractionl = (uint32_t)((d - lfp.int_partl) * UINT_MAX);
556	lfp.int_partl = htonl(lfp.int_partl);
557	lfp.fractionl = htonl(lfp.fractionl);
558	return lfp;
559	}
560	static s_fixedpt_t
561	d_to_sfp(double d)
562	{
563	s_fixedpt_t sfp;
564	sfp.int_parts = (uint16_t)d;
565	sfp.fractions = (uint16_t)((d - sfp.int_parts) * USHRT_MAX);
566	sfp.int_parts = htons(sfp.int_parts);
567	sfp.fractions = htons(sfp.fractions);
568	return sfp;
569	}
570	#endif
571
572	static double
573	dispersion(const datapoint_t *dp)
574	{
575	return dp->d_dispersion + FREQ_TOLERANCE * (G.cur_time - dp->d_recv_time);
576	}
577
578	static double
579	root_distance(peer_t *p)
580	{
581	/* The root synchronization distance is the maximum error due to
582	* all causes of the local clock relative to the primary server.
583	* It is defined as half the total delay plus total dispersion
584	* plus peer jitter.
585	*/
586	return MAXD(MINDISP, p->lastpkt_rootdelay + p->lastpkt_delay) / 2
587	+ p->lastpkt_rootdisp
588	+ p->filter_dispersion
589	+ FREQ_TOLERANCE * (G.cur_time - p->lastpkt_recv_time)
590	+ p->filter_jitter;
591	}
592
593	static void
594	set_next(peer_t *p, unsigned t)
595	{
596	p->next_action_time = G.cur_time + t;
597	}
598
599	/*
600	* Peer clock filter and its helpers
601	*/
602	static void
603	filter_datapoints(peer_t *p)
604	{
605	int i, idx;
606	double sum, wavg;
607	datapoint_t *fdp;
608
609	#if 0
610	/* Simulations have shown that use of *averaged* offset for p->filter_offset
611	* is in fact worse than simply using last received one: with large poll intervals
612	* (>= 2048) averaging code uses offset values which are outdated by hours,
613	* and time/frequency correction goes totally wrong when fed essentially bogus offsets.
614	*/
615	int got_newest;
616	double minoff, maxoff, w;
617	double x = x; /* for compiler */
618	double oldest_off = oldest_off;
619	double oldest_age = oldest_age;
620	double newest_off = newest_off;
621	double newest_age = newest_age;
622
623	fdp = p->filter_datapoint;
624
625	minoff = maxoff = fdp[0].d_offset;
626	for (i = 1; i < NUM_DATAPOINTS; i++) {
627	if (minoff > fdp[i].d_offset)
628	minoff = fdp[i].d_offset;
629	if (maxoff < fdp[i].d_offset)
630	maxoff = fdp[i].d_offset;
631	}
632
633	idx = p->datapoint_idx; /* most recent datapoint's index */
634	/* Average offset:
635	* Drop two outliers and take weighted average of the rest:
636	* most_recent/2 + older1/4 + older2/8 ... + older5/32 + older6/32
637	* we use older6/32, not older6/64 since sum of weights should be 1:
638	* 1/2 + 1/4 + 1/8 + 1/16 + 1/32 + 1/32 = 1
639	*/
640	wavg = 0;
641	w = 0.5;
642	/* n-1
643	* --- dispersion(i)
644	* filter_dispersion = \ -------------
645	* / (i+1)
646	* --- 2
647	* i=0
648	*/
649	got_newest = 0;
650	sum = 0;
651	for (i = 0; i < NUM_DATAPOINTS; i++) {
652	VERB5 {
653	bb_error_msg("datapoint[%d]: off:%f disp:%f(%f) age:%f%s",
654	i,
655	fdp[idx].d_offset,
656	fdp[idx].d_dispersion, dispersion(&fdp[idx]),
657	G.cur_time - fdp[idx].d_recv_time,
658	(minoff == fdp[idx].d_offset || maxoff == fdp[idx].d_offset)
659	? " (outlier by offset)" : ""
660	);
661	}
662
663	sum += dispersion(&fdp[idx]) / (2 << i);
664
665	if (minoff == fdp[idx].d_offset) {
666	minoff -= 1; /* so that we don't match it ever again */
667	} else
668	if (maxoff == fdp[idx].d_offset) {
669	maxoff += 1;
670	} else {
671	oldest_off = fdp[idx].d_offset;
672	oldest_age = G.cur_time - fdp[idx].d_recv_time;
673	if (!got_newest) {
674	got_newest = 1;
675	newest_off = oldest_off;
676	newest_age = oldest_age;
677	}
678	x = oldest_off * w;
679	wavg += x;
680	w /= 2;
681	}
682
683	idx = (idx - 1) & (NUM_DATAPOINTS - 1);
684	}
685	p->filter_dispersion = sum;
686	wavg += x; /* add another older6/64 to form older6/32 */
687	/* Fix systematic underestimation with large poll intervals.
688	* Imagine that we still have a bit of uncorrected drift,
689	* and poll interval is big (say, 100 sec). Offsets form a progression:
690	* 0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 - 0.7 is most recent.
691	* The algorithm above drops 0.0 and 0.7 as outliers,
692	* and then we have this estimation, ~25% off from 0.7:
693	* 0.1/32 + 0.2/32 + 0.3/16 + 0.4/8 + 0.5/4 + 0.6/2 = 0.503125
694	*/
695	x = oldest_age - newest_age;
696	if (x != 0) {
697	x = newest_age / x; /* in above example, 100 / (600 - 100) */
698	if (x < 1) { /* paranoia check */
699	x = (newest_off - oldest_off) * x; /* 0.5 * 100/500 = 0.1 */
700	wavg += x;
701	}
702	}
703	p->filter_offset = wavg;
704
705	#else
706
707	fdp = p->filter_datapoint;
708	idx = p->datapoint_idx; /* most recent datapoint's index */
709
710	/* filter_offset: simply use the most recent value */
711	p->filter_offset = fdp[idx].d_offset;
712
713	/* n-1
714	* --- dispersion(i)
715	* filter_dispersion = \ -------------
716	* / (i+1)
717	* --- 2
718	* i=0
719	*/
720	wavg = 0;
721	sum = 0;
722	for (i = 0; i < NUM_DATAPOINTS; i++) {
723	sum += dispersion(&fdp[idx]) / (2 << i);
724	wavg += fdp[idx].d_offset;
725	idx = (idx - 1) & (NUM_DATAPOINTS - 1);
726	}
727	wavg /= NUM_DATAPOINTS;
728	p->filter_dispersion = sum;
729	#endif
730
731	/* +----- -----+ ^ 1/2
732	* | n-1 |
733	* | --- |
734	* | 1 \ 2 |
735	* filter_jitter = | --- * / (avg-offset_j) |
736	* | n --- |
737	* | j=0 |
738	* +----- -----+
739	* where n is the number of valid datapoints in the filter (n > 1);
740	* if filter_jitter < precision then filter_jitter = precision
741	*/
742	sum = 0;
743	for (i = 0; i < NUM_DATAPOINTS; i++) {
744	sum += SQUARE(wavg - fdp[i].d_offset);
745	}
746	sum = SQRT(sum / NUM_DATAPOINTS);
747	p->filter_jitter = sum > G_precision_sec ? sum : G_precision_sec;
748
749	VERB4 bb_error_msg("filter offset:%+f disp:%f jitter:%f",
750	p->filter_offset,
751	p->filter_dispersion,
752	p->filter_jitter);
753	}
754
755	static void
756	reset_peer_stats(peer_t *p, double offset)
757	{
758	int i;
759	bool small_ofs = fabs(offset) < STEP_THRESHOLD;
760
761	/* Used to set p->filter_datapoint[i].d_dispersion = MAXDISP
762	* and clear reachable bits, but this proved to be too agressive:
763	* after step (tested with suspending laptop for ~30 secs),
764	* this caused all previous data to be considered invalid,
765	* making us needing to collect full ~8 datapoins per peer
766	* after step in order to start trusting them.
767	* In turn, this was making poll interval decrease even after
768	* step was done. (Poll interval decreases already before step
769	* in this scenario, because we see large offsets and end up with
770	* no good peer to select).
771	*/
772
773	for (i = 0; i < NUM_DATAPOINTS; i++) {
774	if (small_ofs) {
775	p->filter_datapoint[i].d_recv_time += offset;
776	if (p->filter_datapoint[i].d_offset != 0) {
777	p->filter_datapoint[i].d_offset -= offset;
778	//bb_error_msg("p->filter_datapoint[%d].d_offset %f -> %f",
779	// i,
780	// p->filter_datapoint[i].d_offset + offset,
781	// p->filter_datapoint[i].d_offset);
782	}
783	} else {
784	p->filter_datapoint[i].d_recv_time = G.cur_time;
785	p->filter_datapoint[i].d_offset = 0;
786	/*p->filter_datapoint[i].d_dispersion = MAXDISP;*/
787	}
788	}
789	if (small_ofs) {
790	p->lastpkt_recv_time += offset;
791	} else {
792	/*p->reachable_bits = 0;*/
793	p->lastpkt_recv_time = G.cur_time;
794	}
795	filter_datapoints(p); /* recalc p->filter_xxx */
796	VERB6 bb_error_msg("%s->lastpkt_recv_time=%f", p->p_dotted, p->lastpkt_recv_time);
797	}
798
799	static void
800	resolve_peer_hostname(peer_t *p, int loop_on_fail)
801	{
802	len_and_sockaddr *lsa;
803
804	again:
805	lsa = host2sockaddr(p->p_hostname, 123);
806	if (!lsa) {
807	/* error message already emitted by host2sockaddr() */
808	if (!loop_on_fail)
809	return;
810	//FIXME: do this to avoid infinite looping on typo in a hostname?
811	//well... in which case, what is a good value for loop_on_fail?
812	//if (--loop_on_fail == 0)
813	// xfunc_die();
814	sleep(5);
815	goto again;
816	}
817	free(p->p_lsa);
818	free(p->p_dotted);
819	p->p_lsa = lsa;
820	p->p_dotted = xmalloc_sockaddr2dotted_noport(&lsa->u.sa);
821	}
822
823	static void
824	add_peers(const char *s)
825	{
826	llist_t *item;
827	peer_t *p;
828
829	p = xzalloc(sizeof(*p) + strlen(s));
830	strcpy(p->p_hostname, s);
831	resolve_peer_hostname(p, /*loop_on_fail=*/ 1);
832
833	/* Names like N.<country2chars>.pool.ntp.org are randomly resolved
834	* to a pool of machines. Sometimes different N's resolve to the same IP.
835	* It is not useful to have two peers with same IP. We skip duplicates.
836	*/
837	for (item = G.ntp_peers; item != NULL; item = item->link) {
838	peer_t *pp = (peer_t *) item->data;
839	if (strcmp(p->p_dotted, pp->p_dotted) == 0) {
840	bb_error_msg("duplicate peer %s (%s)", s, p->p_dotted);
841	free(p->p_lsa);
842	free(p->p_dotted);
843	free(p);
844	return;
845	}
846	}
847
848	p->p_fd = -1;
849	p->p_xmt_msg.m_status = MODE_CLIENT | (NTP_VERSION << 3);
850	p->next_action_time = G.cur_time; /* = set_next(p, 0); */
851	reset_peer_stats(p, STEP_THRESHOLD);
852
853	llist_add_to(&G.ntp_peers, p);
854	G.peer_cnt++;
855	}
856
857	static int
858	do_sendto(int fd,
859	const struct sockaddr *from, const struct sockaddr *to, socklen_t addrlen,
860	msg_t *msg, ssize_t len)
861	{
862	ssize_t ret;
863
864	errno = 0;
865	if (!from) {
866	ret = sendto(fd, msg, len, MSG_DONTWAIT, to, addrlen);
867	} else {
868	ret = send_to_from(fd, msg, len, MSG_DONTWAIT, to, from, addrlen);
869	}
870	if (ret != len) {
871	bb_perror_msg("send failed");
872	return -1;
873	}
874	return 0;
875	}
876
877	static void
878	send_query_to_peer(peer_t *p)
879	{
880	/* Why do we need to bind()?
881	* See what happens when we don't bind:
882	*
883	* socket(PF_INET, SOCK_DGRAM, IPPROTO_IP) = 3
884	* setsockopt(3, SOL_IP, IP_TOS, [16], 4) = 0
885	* gettimeofday({1259071266, 327885}, NULL) = 0
886	* sendto(3, "xxx", 48, MSG_DONTWAIT, {sa_family=AF_INET, sin_port=htons(123), sin_addr=inet_addr("10.34.32.125")}, 16) = 48
887	* ^^^ we sent it from some source port picked by kernel.
888	* time(NULL) = 1259071266
889	* write(2, "ntpd: entering poll 15 secs\n", 28) = 28
890	* poll([{fd=3, events=POLLIN}], 1, 15000) = 1 ([{fd=3, revents=POLLIN}])
891	* recv(3, "yyy", 68, MSG_DONTWAIT) = 48
892	* ^^^ this recv will receive packets to any local port!
893	*
894	* Uncomment this and use strace to see it in action:
895	*/
896	#define PROBE_LOCAL_ADDR /* { len_and_sockaddr lsa; lsa.len = LSA_SIZEOF_SA; getsockname(p->query.fd, &lsa.u.sa, &lsa.len); } */
897
898	if (p->p_fd == -1) {
899	int fd, family;
900	len_and_sockaddr *local_lsa;
901
902	family = p->p_lsa->u.sa.sa_family;
903	p->p_fd = fd = xsocket_type(&local_lsa, family, SOCK_DGRAM);
904	/* local_lsa has "null" address and port 0 now.
905	* bind() ensures we have a *particular port* selected by kernel
906	* and remembered in p->p_fd, thus later recv(p->p_fd)
907	* receives only packets sent to this port.
908	*/
909	PROBE_LOCAL_ADDR
910	xbind(fd, &local_lsa->u.sa, local_lsa->len);
911	PROBE_LOCAL_ADDR
912	#if ENABLE_FEATURE_IPV6
913	if (family == AF_INET)
914	#endif
915	setsockopt_int(fd, IPPROTO_IP, IP_TOS, IPTOS_LOWDELAY);
916	free(local_lsa);
917	}
918
919	/* Emit message _before_ attempted send. Think of a very short
920	* roundtrip networks: we need to go back to recv loop ASAP,
921	* to reduce delay. Printing messages after send works against that.
922	*/
923	VERB1 bb_error_msg("sending query to %s", p->p_dotted);
924
925	/*
926	* Send out a random 64-bit number as our transmit time. The NTP
927	* server will copy said number into the originate field on the
928	* response that it sends us. This is totally legal per the SNTP spec.
929	*
930	* The impact of this is two fold: we no longer send out the current
931	* system time for the world to see (which may aid an attacker), and
932	* it gives us a (not very secure) way of knowing that we're not
933	* getting spoofed by an attacker that can't capture our traffic
934	* but can spoof packets from the NTP server we're communicating with.
935	*
936	* Save the real transmit timestamp locally.
937	*/
938	p->p_xmt_msg.m_xmttime.int_partl = rand();
939	p->p_xmt_msg.m_xmttime.fractionl = rand();
940	p->p_xmttime = gettime1900d();
941
942	/* Were doing it only if sendto worked, but
943	* loss of sync detection needs reachable_bits updated
944	* even if sending fails *locally*:
945	* "network is unreachable" because cable was pulled?
946	* We still need to declare "unsync" if this condition persists.
947	*/
948	p->reachable_bits <<= 1;
949
950	if (do_sendto(p->p_fd, /*from:*/ NULL, /*to:*/ &p->p_lsa->u.sa, /*addrlen:*/ p->p_lsa->len,
951	&p->p_xmt_msg, NTP_MSGSIZE_NOAUTH) == -1
952	) {
953	close(p->p_fd);
954	p->p_fd = -1;
955	/*
956	* We know that we sent nothing.
957	* We can retry *soon* without fearing
958	* that we are flooding the peer.
959	*/
960	set_next(p, RETRY_INTERVAL);
961	return;
962	}
963
964	set_next(p, RESPONSE_INTERVAL);
965	}
966
967
968	/* Note that there is no provision to prevent several run_scripts
969	* to be started in quick succession. In fact, it happens rather often
970	* if initial syncronization results in a step.
971	* You will see "step" and then "stratum" script runs, sometimes
972	* as close as only 0.002 seconds apart.
973	* Script should be ready to deal with this.
974	*/
975	static void run_script(const char *action, double offset)
976	{
977	char *argv[3];
978	char *env1, *env2, *env3, *env4;
979
980	G.last_script_run = G.cur_time;
981
982	if (!G.script_name)
983	return;
984
985	argv[0] = (char*) G.script_name;
986	argv[1] = (char*) action;
987	argv[2] = NULL;
988
989	VERB1 bb_error_msg("executing '%s %s'", G.script_name, action);
990
991	env1 = xasprintf("%s=%u", "stratum", G.stratum);
992	putenv(env1);
993	env2 = xasprintf("%s=%ld", "freq_drift_ppm", G.kernel_freq_drift);
994	putenv(env2);
995	env3 = xasprintf("%s=%u", "poll_interval", 1 << G.poll_exp);
996	putenv(env3);
997	env4 = xasprintf("%s=%f", "offset", offset);
998	putenv(env4);
999	/* Other items of potential interest: selected peer,
1000	* rootdelay, reftime, rootdisp, refid, ntp_status,
1001	* last_update_offset, last_update_recv_time, discipline_jitter,
1002	* how many peers have reachable_bits = 0?
1003	*/
1004
1005	/* Don't want to wait: it may run hwclock --systohc, and that
1006	* may take some time (seconds): */
1007	/*spawn_and_wait(argv);*/
1008	spawn(argv);
1009
1010	unsetenv("stratum");
1011	unsetenv("freq_drift_ppm");
1012	unsetenv("poll_interval");
1013	unsetenv("offset");
1014	free(env1);
1015	free(env2);
1016	free(env3);
1017	free(env4);
1018	}
1019
1020	static NOINLINE void
1021	step_time(double offset)
1022	{
1023	llist_t *item;
1024	double dtime;
1025	struct timeval tvc, tvn;
1026	char buf[sizeof("yyyy-mm-dd hh:mm:ss") + /*paranoia:*/ 4];
1027	time_t tval;
1028
1029	gettimeofday(&tvc, NULL); /* never fails */
1030	dtime = tvc.tv_sec + (1.0e-6 * tvc.tv_usec) + offset;
1031	d_to_tv(dtime, &tvn);
1032	if (settimeofday(&tvn, NULL) == -1)
1033	bb_perror_msg_and_die("settimeofday");
1034
1035	VERB2 {
1036	tval = tvc.tv_sec;
1037	strftime_YYYYMMDDHHMMSS(buf, sizeof(buf), &tval);
1038	bb_error_msg("current time is %s.%06u", buf, (unsigned)tvc.tv_usec);
1039	}
1040	tval = tvn.tv_sec;
1041	strftime_YYYYMMDDHHMMSS(buf, sizeof(buf), &tval);
1042	bb_error_msg("setting time to %s.%06u (offset %+fs)", buf, (unsigned)tvn.tv_usec, offset);
1043
1044	/* Correct various fields which contain time-relative values: */
1045
1046	/* Globals: */
1047	G.cur_time += offset;
1048	G.last_update_recv_time += offset;
1049	G.last_script_run += offset;
1050
1051	/* p->lastpkt_recv_time, p->next_action_time and such: */
1052	for (item = G.ntp_peers; item != NULL; item = item->link) {
1053	peer_t *pp = (peer_t *) item->data;
1054	reset_peer_stats(pp, offset);
1055	//bb_error_msg("offset:%+f pp->next_action_time:%f -> %f",
1056	// offset, pp->next_action_time, pp->next_action_time + offset);
1057	pp->next_action_time += offset;
1058	if (pp->p_fd >= 0) {
1059	/* We wait for reply from this peer too.
1060	* But due to step we are doing, reply's data is no longer
1061	* useful (in fact, it'll be bogus). Stop waiting for it.
1062	*/
1063	close(pp->p_fd);
1064	pp->p_fd = -1;
1065	set_next(pp, RETRY_INTERVAL);
1066	}
1067	}
1068	}
1069
1070	static void clamp_pollexp_and_set_MAXSTRAT(void)
1071	{
1072	if (G.poll_exp < MINPOLL)
1073	G.poll_exp = MINPOLL;
1074	if (G.poll_exp > BIGPOLL)
1075	G.poll_exp = BIGPOLL;
1076	G.polladj_count = 0;
1077	G.stratum = MAXSTRAT;
1078	}
1079
1080
1081	/*
1082	* Selection and clustering, and their helpers
1083	*/
1084	typedef struct {
1085	peer_t *p;
1086	int type;
1087	double edge;
1088	double opt_rd; /* optimization */
1089	} point_t;
1090	static int
1091	compare_point_edge(const void *aa, const void *bb)
1092	{
1093	const point_t *a = aa;
1094	const point_t *b = bb;
1095	if (a->edge < b->edge) {
1096	return -1;
1097	}
1098	return (a->edge > b->edge);
1099	}
1100	typedef struct {
1101	peer_t *p;
1102	double metric;
1103	} survivor_t;
1104	static int
1105	compare_survivor_metric(const void *aa, const void *bb)
1106	{
1107	const survivor_t *a = aa;
1108	const survivor_t *b = bb;
1109	if (a->metric < b->metric) {
1110	return -1;
1111	}
1112	return (a->metric > b->metric);
1113	}
1114	static int
1115	fit(peer_t *p, double rd)
1116	{
1117	if ((p->reachable_bits & (p->reachable_bits-1)) == 0) {
1118	/* One or zero bits in reachable_bits */
1119	VERB4 bb_error_msg("peer %s unfit for selection: unreachable", p->p_dotted);
1120	return 0;
1121	}
1122	#if 0 /* we filter out such packets earlier */
1123	if ((p->lastpkt_status & LI_ALARM) == LI_ALARM
1124	|| p->lastpkt_stratum >= MAXSTRAT
1125	) {
1126	VERB4 bb_error_msg("peer %s unfit for selection: bad status/stratum", p->p_dotted);
1127	return 0;
1128	}
1129	#endif
1130	/* rd is root_distance(p) */
1131	if (rd > MAXDIST + FREQ_TOLERANCE * (1 << G.poll_exp)) {
1132	VERB4 bb_error_msg("peer %s unfit for selection: root distance too high", p->p_dotted);
1133	return 0;
1134	}
1135	//TODO
1136	// /* Do we have a loop? */
1137	// if (p->refid == p->dstaddr || p->refid == s.refid)
1138	// return 0;
1139	return 1;
1140	}
1141	static peer_t*
1142	select_and_cluster(void)
1143	{
1144	peer_t *p;
1145	llist_t *item;
1146	int i, j;
1147	int size = 3 * G.peer_cnt;
1148	/* for selection algorithm */
1149	point_t point[size];
1150	int num_points, num_candidates;
1151	double low, high;
1152	int num_falsetickers;
1153	/* for cluster algorithm */
1154	survivor_t survivor[size];
1155	int num_survivors;
1156
1157	/* Selection */
1158
1159	num_points = 0;
1160	item = G.ntp_peers;
1161	while (item != NULL) {
1162	double rd, offset;
1163
1164	p = (peer_t *) item->data;
1165	rd = root_distance(p);
1166	offset = p->filter_offset;
1167	if (!fit(p, rd)) {
1168	item = item->link;
1169	continue;
1170	}
1171
1172	VERB5 bb_error_msg("interval: [%f %f %f] %s",
1173	offset - rd,
1174	offset,
1175	offset + rd,
1176	p->p_dotted
1177	);
1178	point[num_points].p = p;
1179	point[num_points].type = -1;
1180	point[num_points].edge = offset - rd;
1181	point[num_points].opt_rd = rd;
1182	num_points++;
1183	point[num_points].p = p;
1184	point[num_points].type = 0;
1185	point[num_points].edge = offset;
1186	point[num_points].opt_rd = rd;
1187	num_points++;
1188	point[num_points].p = p;
1189	point[num_points].type = 1;
1190	point[num_points].edge = offset + rd;
1191	point[num_points].opt_rd = rd;
1192	num_points++;
1193	item = item->link;
1194	}
1195	num_candidates = num_points / 3;
1196	if (num_candidates == 0) {
1197	VERB3 bb_error_msg("no valid datapoints%s", ", no peer selected");
1198	return NULL;
1199	}
1200	//TODO: sorting does not seem to be done in reference code
1201	qsort(point, num_points, sizeof(point[0]), compare_point_edge);
1202
1203	/* Start with the assumption that there are no falsetickers.
1204	* Attempt to find a nonempty intersection interval containing
1205	* the midpoints of all truechimers.
1206	* If a nonempty interval cannot be found, increase the number
1207	* of assumed falsetickers by one and try again.
1208	* If a nonempty interval is found and the number of falsetickers
1209	* is less than the number of truechimers, a majority has been found
1210	* and the midpoint of each truechimer represents
1211	* the candidates available to the cluster algorithm.
1212	*/
1213	num_falsetickers = 0;
1214	while (1) {
1215	int c;
1216	int num_midpoints = 0;
1217
1218	low = 1 << 9;
1219	high = - (1 << 9);
1220	c = 0;
1221	for (i = 0; i < (int) num_points; i++) {
1222	/* We want to do:
1223	* if (point[i].type == -1) c++;
1224	* if (point[i].type == 1) c--;
1225	* and it's simpler to do it this way:
1226	*/
1227	c -= point[i].type;
1228	if (c >= num_candidates - num_falsetickers) {
1229	/* If it was c++ and it got big enough... */
1230	low = point[i].edge;
1231	break;
1232	}
1233	if (point[i].type == 0)
1234	num_midpoints++;
1235	}
1236	c = 0;
1237	for (i = num_points-1; i >= 0; i--) {
1238	c += point[i].type;
1239	if (c >= num_candidates - num_falsetickers) {
1240	high = point[i].edge;
1241	break;
1242	}
1243	if (point[i].type == 0)
1244	num_midpoints++;
1245	}
1246	/* If the number of midpoints is greater than the number
1247	* of allowed falsetickers, the intersection contains at
1248	* least one truechimer with no midpoint - bad.
1249	* Also, interval should be nonempty.
1250	*/
1251	if (num_midpoints <= num_falsetickers && low < high)
1252	break;
1253	num_falsetickers++;
1254	if (num_falsetickers * 2 >= num_candidates) {
1255	VERB3 bb_error_msg("falsetickers:%d, candidates:%d%s",
1256	num_falsetickers, num_candidates,
1257	", no peer selected");
1258	return NULL;
1259	}
1260	}
1261	VERB4 bb_error_msg("selected interval: [%f, %f]; candidates:%d falsetickers:%d",
1262	low, high, num_candidates, num_falsetickers);
1263
1264	/* Clustering */
1265
1266	/* Construct a list of survivors (p, metric)
1267	* from the chime list, where metric is dominated
1268	* first by stratum and then by root distance.
1269	* All other things being equal, this is the order of preference.
1270	*/
1271	num_survivors = 0;
1272	for (i = 0; i < num_points; i++) {
1273	if (point[i].edge < low || point[i].edge > high)
1274	continue;
1275	p = point[i].p;
1276	survivor[num_survivors].p = p;
1277	/* x.opt_rd == root_distance(p); */
1278	survivor[num_survivors].metric = MAXDIST * p->lastpkt_stratum + point[i].opt_rd;
1279	VERB5 bb_error_msg("survivor[%d] metric:%f peer:%s",
1280	num_survivors, survivor[num_survivors].metric, p->p_dotted);
1281	num_survivors++;
1282	}
1283	/* There must be at least MIN_SELECTED survivors to satisfy the
1284	* correctness assertions. Ordinarily, the Byzantine criteria
1285	* require four survivors, but for the demonstration here, one
1286	* is acceptable.
1287	*/
1288	if (num_survivors < MIN_SELECTED) {
1289	VERB3 bb_error_msg("survivors:%d%s",
1290	num_survivors,
1291	", no peer selected");
1292	return NULL;
1293	}
1294
1295	//looks like this is ONLY used by the fact that later we pick survivor[0].
1296	//we can avoid sorting then, just find the minimum once!
1297	qsort(survivor, num_survivors, sizeof(survivor[0]), compare_survivor_metric);
1298
1299	/* For each association p in turn, calculate the selection
1300	* jitter p->sjitter as the square root of the sum of squares
1301	* (p->offset - q->offset) over all q associations. The idea is
1302	* to repeatedly discard the survivor with maximum selection
1303	* jitter until a termination condition is met.
1304	*/
1305	while (1) {
1306	static int max_idx;
1307	double max_selection_jitter = max_selection_jitter;
1308	double min_jitter = min_jitter;
1309
1310	if (num_survivors <= MIN_CLUSTERED) {
1311	VERB4 bb_error_msg("num_survivors %d <= %d, not discarding more",
1312	num_survivors, MIN_CLUSTERED);
1313	break;
1314	}
1315
1316	/* To make sure a few survivors are left
1317	* for the clustering algorithm to chew on,
1318	* we stop if the number of survivors
1319	* is less than or equal to MIN_CLUSTERED (3).
1320	*/
1321	for (i = 0; i < num_survivors; i++) {
1322	double selection_jitter_sq;
1323
1324	p = survivor[i].p;
1325	if (i == 0 || p->filter_jitter < min_jitter)
1326	min_jitter = p->filter_jitter;
1327
1328	selection_jitter_sq = 0;
1329	for (j = 0; j < num_survivors; j++) {
1330	peer_t *q = survivor[j].p;
1331	selection_jitter_sq += SQUARE(p->filter_offset - q->filter_offset);
1332	}
1333	if (i == 0 || selection_jitter_sq > max_selection_jitter) {
1334	max_selection_jitter = selection_jitter_sq;
1335	max_idx = i;
1336	}
1337	VERB6 bb_error_msg("survivor %d selection_jitter^2:%f",
1338	i, selection_jitter_sq);
1339	}
1340	max_selection_jitter = SQRT(max_selection_jitter / num_survivors);
1341	VERB5 bb_error_msg("max_selection_jitter (at %d):%f min_jitter:%f",
1342	max_idx, max_selection_jitter, min_jitter);
1343
1344	/* If the maximum selection jitter is less than the
1345	* minimum peer jitter, then tossing out more survivors
1346	* will not lower the minimum peer jitter, so we might
1347	* as well stop.
1348	*/
1349	if (max_selection_jitter < min_jitter) {
1350	VERB4 bb_error_msg("max_selection_jitter:%f < min_jitter:%f, num_survivors:%d, not discarding more",
1351	max_selection_jitter, min_jitter, num_survivors);
1352	break;
1353	}
1354
1355	/* Delete survivor[max_idx] from the list
1356	* and go around again.
1357	*/
1358	VERB6 bb_error_msg("dropping survivor %d", max_idx);
1359	num_survivors--;
1360	while (max_idx < num_survivors) {
1361	survivor[max_idx] = survivor[max_idx + 1];
1362	max_idx++;
1363	}
1364	}
1365
1366	if (0) {
1367	/* Combine the offsets of the clustering algorithm survivors
1368	* using a weighted average with weight determined by the root
1369	* distance. Compute the selection jitter as the weighted RMS
1370	* difference between the first survivor and the remaining
1371	* survivors. In some cases the inherent clock jitter can be
1372	* reduced by not using this algorithm, especially when frequent
1373	* clockhopping is involved. bbox: thus we don't do it.
1374	*/
1375	double x, y, z, w;
1376	y = z = w = 0;
1377	for (i = 0; i < num_survivors; i++) {
1378	p = survivor[i].p;
1379	x = root_distance(p);
1380	y += 1 / x;
1381	z += p->filter_offset / x;
1382	w += SQUARE(p->filter_offset - survivor[0].p->filter_offset) / x;
1383	}
1384	//G.cluster_offset = z / y;
1385	//G.cluster_jitter = SQRT(w / y);
1386	}
1387
1388	/* Pick the best clock. If the old system peer is on the list
1389	* and at the same stratum as the first survivor on the list,
1390	* then don't do a clock hop. Otherwise, select the first
1391	* survivor on the list as the new system peer.
1392	*/
1393	p = survivor[0].p;
1394	if (G.last_update_peer
1395	&& G.last_update_peer->lastpkt_stratum <= p->lastpkt_stratum
1396	) {
1397	/* Starting from 1 is ok here */
1398	for (i = 1; i < num_survivors; i++) {
1399	if (G.last_update_peer == survivor[i].p) {
1400	VERB5 bb_error_msg("keeping old synced peer");
1401	p = G.last_update_peer;
1402	goto keep_old;
1403	}
1404	}
1405	}
1406	G.last_update_peer = p;
1407	keep_old:
1408	VERB4 bb_error_msg("selected peer %s filter_offset:%+f age:%f",
1409	p->p_dotted,
1410	p->filter_offset,
1411	G.cur_time - p->lastpkt_recv_time
1412	);
1413	return p;
1414	}
1415
1416
1417	/*
1418	* Local clock discipline and its helpers
1419	*/
1420	static void
1421	set_new_values(int disc_state, double offset, double recv_time)
1422	{
1423	/* Enter new state and set state variables. Note we use the time
1424	* of the last clock filter sample, which must be earlier than
1425	* the current time.
1426	*/
1427	VERB4 bb_error_msg("disc_state=%d last update offset=%f recv_time=%f",
1428	disc_state, offset, recv_time);
1429	G.discipline_state = disc_state;
1430	G.last_update_offset = offset;
1431	G.last_update_recv_time = recv_time;
1432	}
1433	/* Return: -1: decrease poll interval, 0: leave as is, 1: increase */
1434	static NOINLINE int
1435	update_local_clock(peer_t *p)
1436	{
1437	int rc;
1438	struct timex tmx;
1439	/* Note: can use G.cluster_offset instead: */
1440	double offset = p->filter_offset;
1441	double recv_time = p->lastpkt_recv_time;
1442	double abs_offset;
1443	#if !USING_KERNEL_PLL_LOOP
1444	double freq_drift;
1445	#endif
1446	#if !USING_KERNEL_PLL_LOOP || USING_INITIAL_FREQ_ESTIMATION
1447	double since_last_update;
1448	#endif
1449	double etemp, dtemp;
1450
1451	abs_offset = fabs(offset);
1452
1453	#if 0
1454	/* If needed, -S script can do it by looking at $offset
1455	* env var and killing parent */
1456	/* If the offset is too large, give up and go home */
1457	if (abs_offset > PANIC_THRESHOLD) {
1458	bb_error_msg_and_die("offset %f far too big, exiting", offset);
1459	}
1460	#endif
1461
1462	/* If this is an old update, for instance as the result
1463	* of a system peer change, avoid it. We never use
1464	* an old sample or the same sample twice.
1465	*/
1466	if (recv_time <= G.last_update_recv_time) {
1467	VERB3 bb_error_msg("update from %s: same or older datapoint, not using it",
1468	p->p_dotted);
1469	return 0; /* "leave poll interval as is" */
1470	}
1471
1472	/* Clock state machine transition function. This is where the
1473	* action is and defines how the system reacts to large time
1474	* and frequency errors.
1475	*/
1476	#if !USING_KERNEL_PLL_LOOP || USING_INITIAL_FREQ_ESTIMATION
1477	since_last_update = recv_time - G.reftime;
1478	#endif
1479	#if !USING_KERNEL_PLL_LOOP
1480	freq_drift = 0;
1481	#endif
1482	#if USING_INITIAL_FREQ_ESTIMATION
1483	if (G.discipline_state == STATE_FREQ) {
1484	/* Ignore updates until the stepout threshold */
1485	if (since_last_update < WATCH_THRESHOLD) {
1486	VERB4 bb_error_msg("measuring drift, datapoint ignored, %f sec remains",
1487	WATCH_THRESHOLD - since_last_update);
1488	return 0; /* "leave poll interval as is" */
1489	}
1490	# if !USING_KERNEL_PLL_LOOP
1491	freq_drift = (offset - G.last_update_offset) / since_last_update;
1492	# endif
1493	}
1494	#endif
1495
1496	/* There are two main regimes: when the
1497	* offset exceeds the step threshold and when it does not.
1498	*/
1499	if (abs_offset > STEP_THRESHOLD) {
1500	#if 0
1501	double remains;
1502
1503	// This "spike state" seems to be useless, peer selection already drops
1504	// occassional "bad" datapoints. If we are here, there were _many_
1505	// large offsets. When a few first large offsets are seen,
1506	// we end up in "no valid datapoints, no peer selected" state.
1507	// Only when enough of them are seen (which means it's not a fluke),
1508	// we end up here. Looks like _our_ clock is off.
1509	switch (G.discipline_state) {
1510	case STATE_SYNC:
1511	/* The first outlyer: ignore it, switch to SPIK state */
1512	VERB3 bb_error_msg("update from %s: offset:%+f, spike%s",
1513	p->p_dotted, offset,
1514	"");
1515	G.discipline_state = STATE_SPIK;
1516	return -1; /* "decrease poll interval" */
1517
1518	case STATE_SPIK:
1519	/* Ignore succeeding outlyers until either an inlyer
1520	* is found or the stepout threshold is exceeded.
1521	*/
1522	remains = WATCH_THRESHOLD - since_last_update;
1523	if (remains > 0) {
1524	VERB3 bb_error_msg("update from %s: offset:%+f, spike%s",
1525	p->p_dotted, offset,
1526	", datapoint ignored");
1527	return -1; /* "decrease poll interval" */
1528	}
1529	/* fall through: we need to step */
1530	} /* switch */
1531	#endif
1532
1533	/* Step the time and clamp down the poll interval.
1534	*
1535	* In NSET state an initial frequency correction is
1536	* not available, usually because the frequency file has
1537	* not yet been written. Since the time is outside the
1538	* capture range, the clock is stepped. The frequency
1539	* will be set directly following the stepout interval.
1540	*
1541	* In FSET state the initial frequency has been set
1542	* from the frequency file. Since the time is outside
1543	* the capture range, the clock is stepped immediately,
1544	* rather than after the stepout interval. Guys get
1545	* nervous if it takes 17 minutes to set the clock for
1546	* the first time.
1547	*
1548	* In SPIK state the stepout threshold has expired and
1549	* the phase is still above the step threshold. Note
1550	* that a single spike greater than the step threshold
1551	* is always suppressed, even at the longer poll
1552	* intervals.
1553	*/
1554	VERB4 bb_error_msg("stepping time by %+f; poll_exp=MINPOLL", offset);
1555	step_time(offset);
1556	if (option_mask32 & OPT_q) {
1557	/* We were only asked to set time once. Done. */
1558	exit(0);
1559	}
1560
1561	clamp_pollexp_and_set_MAXSTRAT();
1562
1563	run_script("step", offset);
1564
1565	recv_time += offset;
1566
1567	#if USING_INITIAL_FREQ_ESTIMATION
1568	if (G.discipline_state == STATE_NSET) {
1569	set_new_values(STATE_FREQ, /*offset:*/ 0, recv_time);
1570	return 1; /* "ok to increase poll interval" */
1571	}
1572	#endif
1573	abs_offset = offset = 0;
1574	set_new_values(STATE_SYNC, offset, recv_time);
1575	} else { /* abs_offset <= STEP_THRESHOLD */
1576
1577	/* The ratio is calculated before jitter is updated to make
1578	* poll adjust code more sensitive to large offsets.
1579	*/
1580	G.offset_to_jitter_ratio = abs_offset / G.discipline_jitter;
1581
1582	/* Compute the clock jitter as the RMS of exponentially
1583	* weighted offset differences. Used by the poll adjust code.
1584	*/
1585	etemp = SQUARE(G.discipline_jitter);
1586	dtemp = SQUARE(offset - G.last_update_offset);
1587	G.discipline_jitter = SQRT(etemp + (dtemp - etemp) / AVG);
1588	if (G.discipline_jitter < G_precision_sec)
1589	G.discipline_jitter = G_precision_sec;
1590
1591	switch (G.discipline_state) {
1592	case STATE_NSET:
1593	if (option_mask32 & OPT_q) {
1594	/* We were only asked to set time once.
1595	* The clock is precise enough, no need to step.
1596	*/
1597	exit(0);
1598	}
1599	#if USING_INITIAL_FREQ_ESTIMATION
1600	/* This is the first update received and the frequency
1601	* has not been initialized. The first thing to do
1602	* is directly measure the oscillator frequency.
1603	*/
1604	set_new_values(STATE_FREQ, offset, recv_time);
1605	#else
1606	set_new_values(STATE_SYNC, offset, recv_time);
1607	#endif
1608	VERB4 bb_error_msg("transitioning to FREQ, datapoint ignored");
1609	return 0; /* "leave poll interval as is" */
1610
1611	#if 0 /* this is dead code for now */
1612	case STATE_FSET:
1613	/* This is the first update and the frequency
1614	* has been initialized. Adjust the phase, but
1615	* don't adjust the frequency until the next update.
1616	*/
1617	set_new_values(STATE_SYNC, offset, recv_time);
1618	/* freq_drift remains 0 */
1619	break;
1620	#endif
1621
1622	#if USING_INITIAL_FREQ_ESTIMATION
1623	case STATE_FREQ:
1624	/* since_last_update >= WATCH_THRESHOLD, we waited enough.
1625	* Correct the phase and frequency and switch to SYNC state.
1626	* freq_drift was already estimated (see code above)
1627	*/
1628	set_new_values(STATE_SYNC, offset, recv_time);
1629	break;
1630	#endif
1631
1632	default:
1633	#if !USING_KERNEL_PLL_LOOP
1634	/* Compute freq_drift due to PLL and FLL contributions.
1635	*
1636	* The FLL and PLL frequency gain constants
1637	* depend on the poll interval and Allan
1638	* intercept. The FLL is not used below one-half
1639	* the Allan intercept. Above that the loop gain
1640	* increases in steps to 1 / AVG.
1641	*/
1642	if ((1 << G.poll_exp) > ALLAN / 2) {
1643	etemp = FLL - G.poll_exp;
1644	if (etemp < AVG)
1645	etemp = AVG;
1646	freq_drift += (offset - G.last_update_offset) / (MAXD(since_last_update, ALLAN) * etemp);
1647	}
1648	/* For the PLL the integration interval
1649	* (numerator) is the minimum of the update
1650	* interval and poll interval. This allows
1651	* oversampling, but not undersampling.
1652	*/
1653	etemp = MIND(since_last_update, (1 << G.poll_exp));
1654	dtemp = (4 * PLL) << G.poll_exp;
1655	freq_drift += offset * etemp / SQUARE(dtemp);
1656	#endif
1657	set_new_values(STATE_SYNC, offset, recv_time);
1658	break;
1659	}
1660	if (G.stratum != p->lastpkt_stratum + 1) {
1661	G.stratum = p->lastpkt_stratum + 1;
1662	run_script("stratum", offset);
1663	}
1664	}
1665
1666	G.reftime = G.cur_time;
1667	G.ntp_status = p->lastpkt_status;
1668	G.refid = p->lastpkt_refid;
1669	G.rootdelay = p->lastpkt_rootdelay + p->lastpkt_delay;
1670	dtemp = p->filter_jitter; // SQRT(SQUARE(p->filter_jitter) + SQUARE(G.cluster_jitter));
1671	dtemp += MAXD(p->filter_dispersion + FREQ_TOLERANCE * (G.cur_time - p->lastpkt_recv_time) + abs_offset, MINDISP);
1672	G.rootdisp = p->lastpkt_rootdisp + dtemp;
1673	VERB4 bb_error_msg("updating leap/refid/reftime/rootdisp from peer %s", p->p_dotted);
1674
1675	/* We are in STATE_SYNC now, but did not do adjtimex yet.
1676	* (Any other state does not reach this, they all return earlier)
1677	* By this time, freq_drift and offset are set
1678	* to values suitable for adjtimex.
1679	*/
1680	#if !USING_KERNEL_PLL_LOOP
1681	/* Calculate the new frequency drift and frequency stability (wander).
1682	* Compute the clock wander as the RMS of exponentially weighted
1683	* frequency differences. This is not used directly, but can,
1684	* along with the jitter, be a highly useful monitoring and
1685	* debugging tool.
1686	*/
1687	dtemp = G.discipline_freq_drift + freq_drift;
1688	G.discipline_freq_drift = MAXD(MIND(MAXDRIFT, dtemp), -MAXDRIFT);
1689	etemp = SQUARE(G.discipline_wander);
1690	dtemp = SQUARE(dtemp);
1691	G.discipline_wander = SQRT(etemp + (dtemp - etemp) / AVG);
1692
1693	VERB4 bb_error_msg("discipline freq_drift=%.9f(int:%ld corr:%e) wander=%f",
1694	G.discipline_freq_drift,
1695	(long)(G.discipline_freq_drift * 65536e6),
1696	freq_drift,
1697	G.discipline_wander);
1698	#endif
1699	VERB4 {
1700	memset(&tmx, 0, sizeof(tmx));
1701	if (adjtimex(&tmx) < 0)
1702	bb_perror_msg_and_die("adjtimex");
1703	bb_error_msg("p adjtimex freq:%ld offset:%+ld status:0x%x tc:%ld",
1704	tmx.freq, tmx.offset, tmx.status, tmx.constant);
1705	}
1706
1707	memset(&tmx, 0, sizeof(tmx));
1708	#if 0
1709	//doesn't work, offset remains 0 (!) in kernel:
1710	//ntpd: set adjtimex freq:1786097 tmx.offset:77487
1711	//ntpd: prev adjtimex freq:1786097 tmx.offset:0
1712	//ntpd: cur adjtimex freq:1786097 tmx.offset:0
1713	tmx.modes = ADJ_FREQUENCY | ADJ_OFFSET;
1714	/* 65536 is one ppm */
1715	tmx.freq = G.discipline_freq_drift * 65536e6;
1716	#endif
1717	tmx.modes = ADJ_OFFSET | ADJ_STATUS | ADJ_TIMECONST;// | ADJ_MAXERROR | ADJ_ESTERROR;
1718	tmx.constant = (int)G.poll_exp - 4;
1719	/* EXPERIMENTAL.
1720	* The below if statement should be unnecessary, but...
1721	* It looks like Linux kernel's PLL is far too gentle in changing
1722	* tmx.freq in response to clock offset. Offset keeps growing
1723	* and eventually we fall back to smaller poll intervals.
1724	* We can make correction more agressive (about x2) by supplying
1725	* PLL time constant which is one less than the real one.
1726	* To be on a safe side, let's do it only if offset is significantly
1727	* larger than jitter.
1728	*/
1729	if (G.offset_to_jitter_ratio >= TIMECONST_HACK_GATE)
1730	tmx.constant--;
1731	tmx.offset = (long)(offset * 1000000); /* usec */
1732	if (SLEW_THRESHOLD < STEP_THRESHOLD) {
1733	if (tmx.offset > (long)(SLEW_THRESHOLD * 1000000)) {
1734	tmx.offset = (long)(SLEW_THRESHOLD * 1000000);
1735	tmx.constant--;
1736	}
1737	if (tmx.offset < -(long)(SLEW_THRESHOLD * 1000000)) {
1738	tmx.offset = -(long)(SLEW_THRESHOLD * 1000000);
1739	tmx.constant--;
1740	}
1741	}
1742	if (tmx.constant < 0)
1743	tmx.constant = 0;
1744
1745	tmx.status = STA_PLL;
1746	if (G.ntp_status & LI_PLUSSEC)
1747	tmx.status |= STA_INS;
1748	if (G.ntp_status & LI_MINUSSEC)
1749	tmx.status |= STA_DEL;
1750
1751	//tmx.esterror = (uint32_t)(clock_jitter * 1e6);
1752	//tmx.maxerror = (uint32_t)((sys_rootdelay / 2 + sys_rootdisp) * 1e6);
1753	rc = adjtimex(&tmx);
1754	if (rc < 0)
1755	bb_perror_msg_and_die("adjtimex");
1756	/* NB: here kernel returns constant == G.poll_exp, not == G.poll_exp - 4.
1757	* Not sure why. Perhaps it is normal.
1758	*/
1759	VERB4 bb_error_msg("adjtimex:%d freq:%ld offset:%+ld status:0x%x",
1760	rc, tmx.freq, tmx.offset, tmx.status);
1761	G.kernel_freq_drift = tmx.freq / 65536;
1762	VERB2 bb_error_msg("update from:%s offset:%+f delay:%f jitter:%f clock drift:%+.3fppm tc:%d",
1763	p->p_dotted,
1764	offset,
1765	p->lastpkt_delay,
1766	G.discipline_jitter,
1767	(double)tmx.freq / 65536,
1768	(int)tmx.constant
1769	);
1770
1771	return 1; /* "ok to increase poll interval" */
1772	}
1773
1774
1775	/*
1776	* We've got a new reply packet from a peer, process it
1777	* (helpers first)
1778	*/
1779	static unsigned
1780	poll_interval(int upper_bound)
1781	{
1782	unsigned interval, r, mask;
1783	interval = 1 << G.poll_exp;
1784	if (interval > upper_bound)
1785	interval = upper_bound;
1786	mask = ((interval-1) >> 4) | 1;
1787	r = rand();
1788	interval += r & mask; /* ~ random(0..1) * interval/16 */
1789	VERB4 bb_error_msg("chose poll interval:%u (poll_exp:%d)", interval, G.poll_exp);
1790	return interval;
1791	}
1792	static void
1793	adjust_poll(int count)
1794	{
1795	G.polladj_count += count;
1796	if (G.polladj_count > POLLADJ_LIMIT) {
1797	G.polladj_count = 0;
1798	if (G.poll_exp < MAXPOLL) {
1799	G.poll_exp++;
1800	VERB4 bb_error_msg("polladj: discipline_jitter:%f ++poll_exp=%d",
1801	G.discipline_jitter, G.poll_exp);
1802	}
1803	} else if (G.polladj_count < -POLLADJ_LIMIT || (count < 0 && G.poll_exp > BIGPOLL)) {
1804	G.polladj_count = 0;
1805	if (G.poll_exp > MINPOLL) {
1806	llist_t *item;
1807
1808	G.poll_exp--;
1809	/* Correct p->next_action_time in each peer
1810	* which waits for sending, so that they send earlier.
1811	* Old pp->next_action_time are on the order
1812	* of t + (1 << old_poll_exp) + small_random,
1813	* we simply need to subtract ~half of that.
1814	*/
1815	for (item = G.ntp_peers; item != NULL; item = item->link) {
1816	peer_t *pp = (peer_t *) item->data;
1817	if (pp->p_fd < 0)
1818	pp->next_action_time -= (1 << G.poll_exp);
1819	}
1820	VERB4 bb_error_msg("polladj: discipline_jitter:%f --poll_exp=%d",
1821	G.discipline_jitter, G.poll_exp);
1822	}
1823	} else {
1824	VERB4 bb_error_msg("polladj: count:%d", G.polladj_count);
1825	}
1826	}
1827	static NOINLINE void
1828	recv_and_process_peer_pkt(peer_t *p)
1829	{
1830	int rc;
1831	ssize_t size;
1832	msg_t msg;
1833	double T1, T2, T3, T4;
1834	double offset;
1835	double prev_delay, delay;
1836	unsigned interval;
1837	datapoint_t *datapoint;
1838	peer_t *q;
1839
1840	offset = 0;
1841
1842	/* We can recvfrom here and check from.IP, but some multihomed
1843	* ntp servers reply from their *other IP*.
1844	* TODO: maybe we should check at least what we can: from.port == 123?
1845	*/
1846	recv_again:
1847	size = recv(p->p_fd, &msg, sizeof(msg), MSG_DONTWAIT);
1848	if (size < 0) {
1849	if (errno == EINTR)
1850	/* Signal caught */
1851	goto recv_again;
1852	if (errno == EAGAIN)
1853	/* There was no packet after all
1854	* (poll() returning POLLIN for a fd
1855	* is not a ironclad guarantee that data is there)
1856	*/
1857	return;
1858	/*
1859	* If you need a different handling for a specific
1860	* errno, always explain it in comment.
1861	*/
1862	bb_perror_msg_and_die("recv(%s) error", p->p_dotted);
1863	}
1864
1865	if (size != NTP_MSGSIZE_NOAUTH && size != NTP_MSGSIZE) {
1866	bb_error_msg("malformed packet received from %s", p->p_dotted);
1867	return;
1868	}
1869
1870	if (msg.m_orgtime.int_partl != p->p_xmt_msg.m_xmttime.int_partl
1871	|| msg.m_orgtime.fractionl != p->p_xmt_msg.m_xmttime.fractionl
1872	) {
1873	/* Somebody else's packet */
1874	return;
1875	}
1876
1877	/* We do not expect any more packets from this peer for now.
1878	* Closing the socket informs kernel about it.
1879	* We open a new socket when we send a new query.
1880	*/
1881	close(p->p_fd);
1882	p->p_fd = -1;
1883
1884	if ((msg.m_status & LI_ALARM) == LI_ALARM
1885	|| msg.m_stratum == 0
1886	|| msg.m_stratum > NTP_MAXSTRATUM
1887	) {
1888	bb_error_msg("reply from %s: peer is unsynced", p->p_dotted);
1889	/*
1890	* Stratum 0 responses may have commands in 32-bit m_refid field:
1891	* "DENY", "RSTR" - peer does not like us at all,
1892	* "RATE" - peer is overloaded, reduce polling freq.
1893	* If poll interval is small, increase it.
1894	*/
1895	if (G.poll_exp < BIGPOLL)
1896	goto increase_interval;
1897	goto pick_normal_interval;
1898	}
1899
1900	// /* Verify valid root distance */
1901	// if (msg.m_rootdelay / 2 + msg.m_rootdisp >= MAXDISP || p->lastpkt_reftime > msg.m_xmt)
1902	// return; /* invalid header values */
1903
1904	/*
1905	* From RFC 2030 (with a correction to the delay math):
1906	*
1907	* Timestamp Name ID When Generated
1908	* ------------------------------------------------------------
1909	* Originate Timestamp T1 time request sent by client
1910	* Receive Timestamp T2 time request received by server
1911	* Transmit Timestamp T3 time reply sent by server
1912	* Destination Timestamp T4 time reply received by client
1913	*
1914	* The roundtrip delay and local clock offset are defined as
1915	*
1916	* delay = (T4 - T1) - (T3 - T2); offset = ((T2 - T1) + (T3 - T4)) / 2
1917	*/
1918	T1 = p->p_xmttime;
1919	T2 = lfp_to_d(msg.m_rectime);
1920	T3 = lfp_to_d(msg.m_xmttime);
1921	T4 = G.cur_time;
1922
1923	/* The delay calculation is a special case. In cases where the
1924	* server and client clocks are running at different rates and
1925	* with very fast networks, the delay can appear negative. In
1926	* order to avoid violating the Principle of Least Astonishment,
1927	* the delay is clamped not less than the system precision.
1928	*/
1929	delay = (T4 - T1) - (T3 - T2);
1930	if (delay < G_precision_sec)
1931	delay = G_precision_sec;
1932	/*
1933	* If this packet's delay is much bigger than the last one,
1934	* it's better to just ignore it than use its much less precise value.
1935	*/
1936	prev_delay = p->p_raw_delay;
1937	p->p_raw_delay = delay;
1938	if (p->reachable_bits && delay > prev_delay * BAD_DELAY_GROWTH) {
1939	bb_error_msg("reply from %s: delay %f is too high, ignoring", p->p_dotted, delay);
1940	goto pick_normal_interval;
1941	}
1942
1943	p->lastpkt_delay = delay;
1944	p->lastpkt_recv_time = T4;
1945	VERB6 bb_error_msg("%s->lastpkt_recv_time=%f", p->p_dotted, p->lastpkt_recv_time);
1946	p->lastpkt_status = msg.m_status;
1947	p->lastpkt_stratum = msg.m_stratum;
1948	p->lastpkt_rootdelay = sfp_to_d(msg.m_rootdelay);
1949	p->lastpkt_rootdisp = sfp_to_d(msg.m_rootdisp);
1950	p->lastpkt_refid = msg.m_refid;
1951
1952	p->datapoint_idx = p->reachable_bits ? (p->datapoint_idx + 1) % NUM_DATAPOINTS : 0;
1953	datapoint = &p->filter_datapoint[p->datapoint_idx];
1954	datapoint->d_recv_time = T4;
1955	datapoint->d_offset = offset = ((T2 - T1) + (T3 - T4)) / 2;
1956	datapoint->d_dispersion = LOG2D(msg.m_precision_exp) + G_precision_sec;
1957	if (!p->reachable_bits) {
1958	/* 1st datapoint ever - replicate offset in every element */
1959	int i;
1960	for (i = 0; i < NUM_DATAPOINTS; i++) {
1961	p->filter_datapoint[i].d_offset = offset;
1962	}
1963	}
1964
1965	p->reachable_bits |= 1;
1966	if ((MAX_VERBOSE && G.verbose) || (option_mask32 & OPT_w)) {
1967	bb_error_msg("reply from %s: offset:%+f delay:%f status:0x%02x strat:%d refid:0x%08x rootdelay:%f reach:0x%02x",
1968	p->p_dotted,
1969	offset,
1970	p->lastpkt_delay,
1971	p->lastpkt_status,
1972	p->lastpkt_stratum,
1973	p->lastpkt_refid,
1974	p->lastpkt_rootdelay,
1975	p->reachable_bits
1976	/* not shown: m_ppoll, m_precision_exp, m_rootdisp,
1977	* m_reftime, m_orgtime, m_rectime, m_xmttime
1978	*/
1979	);
1980	}
1981
1982	/* Muck with statictics and update the clock */
1983	filter_datapoints(p);
1984	q = select_and_cluster();
1985	rc = 0;
1986	if (q) {
1987	if (!(option_mask32 & OPT_w)) {
1988	rc = update_local_clock(q);
1989	#if 0
1990	//Disabled this because there is a case where largish offsets
1991	//are unavoidable: if network round-trip delay is, say, ~0.6s,
1992	//error in offset estimation would be ~delay/2 ~= 0.3s.
1993	//Thus, offsets will be usually in -0.3...0.3s range.
1994	//In this case, this code would keep poll interval small,
1995	//but it won't be helping.
1996	//BIGOFF check below deals with a case of seeing multi-second offsets.
1997
1998	/* If drift is dangerously large, immediately
1999	* drop poll interval one step down.
2000	*/
2001	if (fabs(q->filter_offset) >= POLLDOWN_OFFSET) {
2002	VERB4 bb_error_msg("offset:%+f > POLLDOWN_OFFSET", q->filter_offset);
2003	adjust_poll(-POLLADJ_LIMIT * 3);
2004	rc = 0;
2005	}
2006	#endif
2007	}
2008	} else {
2009	/* No peer selected.
2010	* If poll interval is small, increase it.
2011	*/
2012	if (G.poll_exp < BIGPOLL)
2013	goto increase_interval;
2014	}
2015
2016	if (rc != 0) {
2017	/* Adjust the poll interval by comparing the current offset
2018	* with the clock jitter. If the offset is less than
2019	* the clock jitter times a constant, then the averaging interval
2020	* is increased, otherwise it is decreased. A bit of hysteresis
2021	* helps calm the dance. Works best using burst mode.
2022	*/
2023	if (rc > 0 && G.offset_to_jitter_ratio <= POLLADJ_GATE) {
2024	/* was += G.poll_exp but it is a bit
2025	* too optimistic for my taste at high poll_exp's */
2026	increase_interval:
2027	adjust_poll(MINPOLL);
2028	} else {
2029	VERB3 if (rc > 0)
2030	bb_error_msg("want smaller interval: offset/jitter = %u",
2031	G.offset_to_jitter_ratio);
2032	adjust_poll(-G.poll_exp * 2);
2033	}
2034	}
2035
2036	/* Decide when to send new query for this peer */
2037	pick_normal_interval:
2038	interval = poll_interval(INT_MAX);
2039	if (fabs(offset) >= BIGOFF && interval > BIGOFF_INTERVAL) {
2040	/* If we are synced, offsets are less than SLEW_THRESHOLD,
2041	* or at the very least not much larger than it.
2042	* Now we see a largish one.
2043	* Either this peer is feeling bad, or packet got corrupted,
2044	* or _our_ clock is wrong now and _all_ peers will show similar
2045	* largish offsets too.
2046	* I observed this with laptop suspend stopping clock.
2047	* In any case, it makes sense to make next request soonish:
2048	* cases 1 and 2: get a better datapoint,
2049	* case 3: allows to resync faster.
2050	*/
2051	interval = BIGOFF_INTERVAL;
2052	}
2053
2054	set_next(p, interval);
2055	}
2056
2057	#if ENABLE_FEATURE_NTPD_SERVER
2058	static NOINLINE void
2059	recv_and_process_client_pkt(void /*int fd*/)
2060	{
2061	ssize_t size;
2062	//uint8_t version;
2063	len_and_sockaddr *to;
2064	struct sockaddr *from;
2065	msg_t msg;
2066	uint8_t query_status;
2067	l_fixedpt_t query_xmttime;
2068
2069	to = get_sock_lsa(G_listen_fd);
2070	from = xzalloc(to->len);
2071
2072	size = recv_from_to(G_listen_fd, &msg, sizeof(msg), MSG_DONTWAIT, from, &to->u.sa, to->len);
2073	if (size != NTP_MSGSIZE_NOAUTH && size != NTP_MSGSIZE) {
2074	char *addr;
2075	if (size < 0) {
2076	if (errno == EAGAIN)
2077	goto bail;
2078	bb_perror_msg_and_die("recv");
2079	}
2080	addr = xmalloc_sockaddr2dotted_noport(from);
2081	bb_error_msg("malformed packet received from %s: size %u", addr, (int)size);
2082	free(addr);
2083	goto bail;
2084	}
2085
2086	/* Respond only to client and symmetric active packets */
2087	if ((msg.m_status & MODE_MASK) != MODE_CLIENT
2088	&& (msg.m_status & MODE_MASK) != MODE_SYM_ACT
2089	) {
2090	goto bail;
2091	}
2092
2093	query_status = msg.m_status;
2094	query_xmttime = msg.m_xmttime;
2095
2096	/* Build a reply packet */
2097	memset(&msg, 0, sizeof(msg));
2098	msg.m_status = G.stratum < MAXSTRAT ? (G.ntp_status & LI_MASK) : LI_ALARM;
2099	msg.m_status |= (query_status & VERSION_MASK);
2100	msg.m_status |= ((query_status & MODE_MASK) == MODE_CLIENT) ?
2101	MODE_SERVER : MODE_SYM_PAS;
2102	msg.m_stratum = G.stratum;
2103	msg.m_ppoll = G.poll_exp;
2104	msg.m_precision_exp = G_precision_exp;
2105	/* this time was obtained between poll() and recv() */
2106	msg.m_rectime = d_to_lfp(G.cur_time);
2107	msg.m_xmttime = d_to_lfp(gettime1900d()); /* this instant */
2108	if (G.peer_cnt == 0) {
2109	/* we have no peers: "stratum 1 server" mode. reftime = our own time */
2110	G.reftime = G.cur_time;
2111	}
2112	msg.m_reftime = d_to_lfp(G.reftime);
2113	msg.m_orgtime = query_xmttime;
2114	msg.m_rootdelay = d_to_sfp(G.rootdelay);
2115	//simple code does not do this, fix simple code!
2116	msg.m_rootdisp = d_to_sfp(G.rootdisp);
2117	//version = (query_status & VERSION_MASK); /* ... >> VERSION_SHIFT - done below instead */
2118	msg.m_refid = G.refid; // (version > (3 << VERSION_SHIFT)) ? G.refid : G.refid3;
2119
2120	/* We reply from the local address packet was sent to,
2121	* this makes to/from look swapped here: */
2122	do_sendto(G_listen_fd,
2123	/*from:*/ &to->u.sa, /*to:*/ from, /*addrlen:*/ to->len,
2124	&msg, size);
2125
2126	bail:
2127	free(to);
2128	free(from);
2129	}
2130	#endif
2131
2132	/* Upstream ntpd's options:
2133	*
2134	* -4 Force DNS resolution of host names to the IPv4 namespace.
2135	* -6 Force DNS resolution of host names to the IPv6 namespace.
2136	* -a Require cryptographic authentication for broadcast client,
2137	* multicast client and symmetric passive associations.
2138	* This is the default.
2139	* -A Do not require cryptographic authentication for broadcast client,
2140	* multicast client and symmetric passive associations.
2141	* This is almost never a good idea.
2142	* -b Enable the client to synchronize to broadcast servers.
2143	* -c conffile
2144	* Specify the name and path of the configuration file,
2145	* default /etc/ntp.conf
2146	* -d Specify debugging mode. This option may occur more than once,
2147	* with each occurrence indicating greater detail of display.
2148	* -D level
2149	* Specify debugging level directly.
2150	* -f driftfile
2151	* Specify the name and path of the frequency file.
2152	* This is the same operation as the "driftfile FILE"
2153	* configuration command.
2154	* -g Normally, ntpd exits with a message to the system log
2155	* if the offset exceeds the panic threshold, which is 1000 s
2156	* by default. This option allows the time to be set to any value
2157	* without restriction; however, this can happen only once.
2158	* If the threshold is exceeded after that, ntpd will exit
2159	* with a message to the system log. This option can be used
2160	* with the -q and -x options. See the tinker command for other options.
2161	* -i jaildir
2162	* Chroot the server to the directory jaildir. This option also implies
2163	* that the server attempts to drop root privileges at startup
2164	* (otherwise, chroot gives very little additional security).
2165	* You may need to also specify a -u option.
2166	* -k keyfile
2167	* Specify the name and path of the symmetric key file,
2168	* default /etc/ntp/keys. This is the same operation
2169	* as the "keys FILE" configuration command.
2170	* -l logfile
2171	* Specify the name and path of the log file. The default
2172	* is the system log file. This is the same operation as
2173	* the "logfile FILE" configuration command.
2174	* -L Do not listen to virtual IPs. The default is to listen.
2175	* -n Don't fork.
2176	* -N To the extent permitted by the operating system,
2177	* run the ntpd at the highest priority.
2178	* -p pidfile
2179	* Specify the name and path of the file used to record the ntpd
2180	* process ID. This is the same operation as the "pidfile FILE"
2181	* configuration command.
2182	* -P priority
2183	* To the extent permitted by the operating system,
2184	* run the ntpd at the specified priority.
2185	* -q Exit the ntpd just after the first time the clock is set.
2186	* This behavior mimics that of the ntpdate program, which is
2187	* to be retired. The -g and -x options can be used with this option.
2188	* Note: The kernel time discipline is disabled with this option.
2189	* -r broadcastdelay
2190	* Specify the default propagation delay from the broadcast/multicast
2191	* server to this client. This is necessary only if the delay
2192	* cannot be computed automatically by the protocol.
2193	* -s statsdir
2194	* Specify the directory path for files created by the statistics
2195	* facility. This is the same operation as the "statsdir DIR"
2196	* configuration command.
2197	* -t key
2198	* Add a key number to the trusted key list. This option can occur
2199	* more than once.
2200	* -u user[:group]
2201	* Specify a user, and optionally a group, to switch to.
2202	* -v variable
2203	* -V variable
2204	* Add a system variable listed by default.
2205	* -x Normally, the time is slewed if the offset is less than the step
2206	* threshold, which is 128 ms by default, and stepped if above
2207	* the threshold. This option sets the threshold to 600 s, which is
2208	* well within the accuracy window to set the clock manually.
2209	* Note: since the slew rate of typical Unix kernels is limited
2210	* to 0.5 ms/s, each second of adjustment requires an amortization
2211	* interval of 2000 s. Thus, an adjustment as much as 600 s
2212	* will take almost 14 days to complete. This option can be used
2213	* with the -g and -q options. See the tinker command for other options.
2214	* Note: The kernel time discipline is disabled with this option.
2215	*/
2216
2217	/* By doing init in a separate function we decrease stack usage
2218	* in main loop.
2219	*/
2220	static NOINLINE void ntp_init(char **argv)
2221	{
2222	unsigned opts;
2223	llist_t *peers;
2224
2225	srand(getpid());
2226
2227	if (getuid())
2228	bb_error_msg_and_die("%s", bb_msg_you_must_be_root);
2229
2230	/* Set some globals */
2231	G.discipline_jitter = G_precision_sec;
2232	G.stratum = MAXSTRAT;
2233	if (BURSTPOLL != 0)
2234	G.poll_exp = BURSTPOLL; /* speeds up initial sync */
2235	G.last_script_run = G.reftime = G.last_update_recv_time = gettime1900d(); /* sets G.cur_time too */
2236
2237	/* Parse options */
2238	peers = NULL;
2239	opt_complementary = "dd:wn" /* -d: counter; -p: list; -w implies -n */
2240	IF_FEATURE_NTPD_SERVER(":Il"); /* -I implies -l */
2241	opts = getopt32(argv,
2242	"nqNx" /* compat */
2243	"wp:*S:"IF_FEATURE_NTPD_SERVER("l") /* NOT compat */
2244	IF_FEATURE_NTPD_SERVER("I:") /* compat */
2245	"d" /* compat */
2246	"46aAbgL", /* compat, ignored */
2247	&peers,&G.script_name,
2248	#if ENABLE_FEATURE_NTPD_SERVER
2249	&G.if_name,
2250	#endif
2251	&G.verbose);
2252
2253	// if (opts & OPT_x) /* disable stepping, only slew is allowed */
2254	// G.time_was_stepped = 1;
2255
2256	#if ENABLE_FEATURE_NTPD_SERVER
2257	G_listen_fd = -1;
2258	if (opts & OPT_l) {
2259	G_listen_fd = create_and_bind_dgram_or_die(NULL, 123);
2260	if (G.if_name) {
2261	if (setsockopt_bindtodevice(G_listen_fd, G.if_name))
2262	xfunc_die();
2263	}
2264	socket_want_pktinfo(G_listen_fd);
2265	setsockopt_int(G_listen_fd, IPPROTO_IP, IP_TOS, IPTOS_LOWDELAY);
2266	}
2267	#endif
2268	/* I hesitate to set -20 prio. -15 should be high enough for timekeeping */
2269	if (opts & OPT_N)
2270	setpriority(PRIO_PROCESS, 0, -15);
2271
2272	/* add_peers() calls can retry DNS resolution (possibly forever).
2273	* Daemonize before them, or else boot can stall forever.
2274	*/
2275	if (!(opts & OPT_n)) {
2276	bb_daemonize_or_rexec(DAEMON_DEVNULL_STDIO, argv);
2277	logmode = LOGMODE_NONE;
2278	}
2279
2280	if (peers) {
2281	while (peers)
2282	add_peers(llist_pop(&peers));
2283	}
2284	#if ENABLE_FEATURE_NTPD_CONF
2285	else {
2286	parser_t *parser;
2287	char *token[3];
2288
2289	parser = config_open("/etc/ntp.conf");
2290	while (config_read(parser, token, 3, 1, "# \t", PARSE_NORMAL)) {
2291	if (strcmp(token[0], "server") == 0 && token[1]) {
2292	add_peers(token[1]);
2293	continue;
2294	}
2295	bb_error_msg("skipping %s:%u: unimplemented command '%s'",
2296	"/etc/ntp.conf", parser->lineno, token[0]
2297	);
2298	}
2299	config_close(parser);
2300	}
2301	#endif
2302	if (G.peer_cnt == 0) {
2303	if (!(opts & OPT_l))
2304	bb_show_usage();
2305	/* -l but no peers: "stratum 1 server" mode */
2306	G.stratum = 1;
2307	}
2308	/* If network is up, syncronization occurs in ~10 seconds.
2309	* We give "ntpd -q" 10 seconds to get first reply,
2310	* then another 50 seconds to finish syncing.
2311	*
2312	* I tested ntpd 4.2.6p1 and apparently it never exits
2313	* (will try forever), but it does not feel right.
2314	* The goal of -q is to act like ntpdate: set time
2315	* after a reasonably small period of polling, or fail.
2316	*/
2317	if (opts & OPT_q) {
2318	option_mask32 |= OPT_qq;
2319	alarm(10);
2320	}
2321
2322	bb_signals(0
2323	| (1 << SIGTERM)
2324	| (1 << SIGINT)
2325	| (1 << SIGALRM)
2326	, record_signo
2327	);
2328	bb_signals(0
2329	| (1 << SIGPIPE)
2330	| (1 << SIGCHLD)
2331	, SIG_IGN
2332	);
2333	}
2334
2335	int ntpd_main(int argc UNUSED_PARAM, char **argv) MAIN_EXTERNALLY_VISIBLE;
2336	int ntpd_main(int argc UNUSED_PARAM, char **argv)
2337	{
2338	#undef G
2339	struct globals G;
2340	struct pollfd *pfd;
2341	peer_t **idx2peer;
2342	unsigned cnt;
2343
2344	memset(&G, 0, sizeof(G));
2345	SET_PTR_TO_GLOBALS(&G);
2346
2347	ntp_init(argv);
2348
2349	/* If ENABLE_FEATURE_NTPD_SERVER, + 1 for listen_fd: */
2350	cnt = G.peer_cnt + ENABLE_FEATURE_NTPD_SERVER;
2351	idx2peer = xzalloc(sizeof(idx2peer[0]) * cnt);
2352	pfd = xzalloc(sizeof(pfd[0]) * cnt);
2353
2354	/* Countdown: we never sync before we sent INITIAL_SAMPLES+1
2355	* packets to each peer.
2356	* NB: if some peer is not responding, we may end up sending
2357	* fewer packets to it and more to other peers.
2358	* NB2: sync usually happens using INITIAL_SAMPLES packets,
2359	* since last reply does not come back instantaneously.
2360	*/
2361	cnt = G.peer_cnt * (INITIAL_SAMPLES + 1);
2362
2363	write_pidfile(CONFIG_PID_FILE_PATH "/ntpd.pid");
2364
2365	while (!bb_got_signal) {
2366	llist_t *item;
2367	unsigned i, j;
2368	int nfds, timeout;
2369	double nextaction;
2370
2371	/* Nothing between here and poll() blocks for any significant time */
2372
2373	nextaction = G.cur_time + 3600;
2374
2375	i = 0;
2376	#if ENABLE_FEATURE_NTPD_SERVER
2377	if (G_listen_fd != -1) {
2378	pfd[0].fd = G_listen_fd;
2379	pfd[0].events = POLLIN;
2380	i++;
2381	}
2382	#endif
2383	/* Pass over peer list, send requests, time out on receives */
2384	for (item = G.ntp_peers; item != NULL; item = item->link) {
2385	peer_t *p = (peer_t *) item->data;
2386
2387	if (p->next_action_time <= G.cur_time) {
2388	if (p->p_fd == -1) {
2389	/* Time to send new req */
2390	if (--cnt == 0) {
2391	VERB4 bb_error_msg("disabling burst mode");
2392	G.polladj_count = 0;
2393	G.poll_exp = MINPOLL;
2394	}
2395	send_query_to_peer(p);
2396	} else {
2397	/* Timed out waiting for reply */
2398	close(p->p_fd);
2399	p->p_fd = -1;
2400	/* If poll interval is small, increase it */
2401	if (G.poll_exp < BIGPOLL)
2402	adjust_poll(MINPOLL);
2403	timeout = poll_interval(NOREPLY_INTERVAL);
2404	bb_error_msg("timed out waiting for %s, reach 0x%02x, next query in %us",
2405	p->p_dotted, p->reachable_bits, timeout);
2406
2407	/* What if don't see it because it changed its IP? */
2408	if (p->reachable_bits == 0)
2409	resolve_peer_hostname(p, /*loop_on_fail=*/ 0);
2410
2411	set_next(p, timeout);
2412	}
2413	}
2414
2415	if (p->next_action_time < nextaction)
2416	nextaction = p->next_action_time;
2417
2418	if (p->p_fd >= 0) {
2419	/* Wait for reply from this peer */
2420	pfd[i].fd = p->p_fd;
2421	pfd[i].events = POLLIN;
2422	idx2peer[i] = p;
2423	i++;
2424	}
2425	}
2426
2427	timeout = nextaction - G.cur_time;
2428	if (timeout < 0)
2429	timeout = 0;
2430	timeout++; /* (nextaction - G.cur_time) rounds down, compensating */
2431
2432	/* Here we may block */
2433	VERB2 {
2434	if (i > (ENABLE_FEATURE_NTPD_SERVER && G_listen_fd != -1)) {
2435	/* We wait for at least one reply.
2436	* Poll for it, without wasting time for message.
2437	* Since replies often come under 1 second, this also
2438	* reduces clutter in logs.
2439	*/
2440	nfds = poll(pfd, i, 1000);
2441	if (nfds != 0)
2442	goto did_poll;
2443	if (--timeout <= 0)
2444	goto did_poll;
2445	}
2446	bb_error_msg("poll:%us sockets:%u interval:%us", timeout, i, 1 << G.poll_exp);
2447	}
2448	nfds = poll(pfd, i, timeout * 1000);
2449	did_poll:
2450	gettime1900d(); /* sets G.cur_time */
2451	if (nfds <= 0) {
2452	if (!bb_got_signal /* poll wasn't interrupted by a signal */
2453	&& G.cur_time - G.last_script_run > 11*60
2454	) {
2455	/* Useful for updating battery-backed RTC and such */
2456	run_script("periodic", G.last_update_offset);
2457	gettime1900d(); /* sets G.cur_time */
2458	}
2459	goto check_unsync;
2460	}
2461
2462	/* Process any received packets */
2463	j = 0;
2464	#if ENABLE_FEATURE_NTPD_SERVER
2465	if (G.listen_fd != -1) {
2466	if (pfd[0].revents /* & (POLLIN|POLLERR)*/) {
2467	nfds--;
2468	recv_and_process_client_pkt(/*G.listen_fd*/);
2469	gettime1900d(); /* sets G.cur_time */
2470	}
2471	j = 1;
2472	}
2473	#endif
2474	for (; nfds != 0 && j < i; j++) {
2475	if (pfd[j].revents /* & (POLLIN|POLLERR)*/) {
2476	/*
2477	* At init, alarm was set to 10 sec.
2478	* Now we did get a reply.
2479	* Increase timeout to 50 seconds to finish syncing.
2480	*/
2481	if (option_mask32 & OPT_qq) {
2482	option_mask32 &= ~OPT_qq;
2483	alarm(50);
2484	}
2485	nfds--;
2486	recv_and_process_peer_pkt(idx2peer[j]);
2487	gettime1900d(); /* sets G.cur_time */
2488	}
2489	}
2490
2491	check_unsync:
2492	if (G.ntp_peers && G.stratum != MAXSTRAT) {
2493	for (item = G.ntp_peers; item != NULL; item = item->link) {
2494	peer_t *p = (peer_t *) item->data;
2495	if (p->reachable_bits)
2496	goto have_reachable_peer;
2497	}
2498	/* No peer responded for last 8 packets, panic */
2499	clamp_pollexp_and_set_MAXSTRAT();
2500	run_script("unsync", 0.0);
2501	have_reachable_peer: ;
2502	}
2503	} /* while (!bb_got_signal) */
2504
2505	remove_pidfile(CONFIG_PID_FILE_PATH "/ntpd.pid");
2506	kill_myself_with_sig(bb_got_signal);
2507	}
2508
2509
2510
2511
2512
2513
2514	/*** openntpd-4.6 uses only adjtime, not adjtimex ***/
2515
2516	/*** ntp-4.2.6/ntpd/ntp_loopfilter.c - adjtimex usage ***/
2517
2518	#if 0
2519	static double
2520	direct_freq(double fp_offset)
2521	{
2522	#ifdef KERNEL_PLL
2523	/*
2524	* If the kernel is enabled, we need the residual offset to
2525	* calculate the frequency correction.
2526	*/
2527	if (pll_control && kern_enable) {
2528	memset(&ntv, 0, sizeof(ntv));
2529	ntp_adjtime(&ntv);
2530	#ifdef STA_NANO
2531	clock_offset = ntv.offset / 1e9;
2532	#else /* STA_NANO */
2533	clock_offset = ntv.offset / 1e6;
2534	#endif /* STA_NANO */
2535	drift_comp = FREQTOD(ntv.freq);
2536	}
2537	#endif /* KERNEL_PLL */
2538	set_freq((fp_offset - clock_offset) / (current_time - clock_epoch) + drift_comp);
2539	wander_resid = 0;
2540	return drift_comp;
2541	}
2542
2543	static void
2544	set_freq(double freq) /* frequency update */
2545	{
2546	char tbuf[80];
2547
2548	drift_comp = freq;
2549
2550	#ifdef KERNEL_PLL
2551	/*
2552	* If the kernel is enabled, update the kernel frequency.
2553	*/
2554	if (pll_control && kern_enable) {
2555	memset(&ntv, 0, sizeof(ntv));
2556	ntv.modes = MOD_FREQUENCY;
2557	ntv.freq = DTOFREQ(drift_comp);
2558	ntp_adjtime(&ntv);
2559	snprintf(tbuf, sizeof(tbuf), "kernel %.3f PPM", drift_comp * 1e6);
2560	report_event(EVNT_FSET, NULL, tbuf);
2561	} else {
2562	snprintf(tbuf, sizeof(tbuf), "ntpd %.3f PPM", drift_comp * 1e6);
2563	report_event(EVNT_FSET, NULL, tbuf);
2564	}
2565	#else /* KERNEL_PLL */
2566	snprintf(tbuf, sizeof(tbuf), "ntpd %.3f PPM", drift_comp * 1e6);
2567	report_event(EVNT_FSET, NULL, tbuf);
2568	#endif /* KERNEL_PLL */
2569	}
2570
2571	...
2572	...
2573	...
2574
2575	#ifdef KERNEL_PLL
2576	/*
2577	* This code segment works when clock adjustments are made using
2578	* precision time kernel support and the ntp_adjtime() system
2579	* call. This support is available in Solaris 2.6 and later,
2580	* Digital Unix 4.0 and later, FreeBSD, Linux and specially
2581	* modified kernels for HP-UX 9 and Ultrix 4. In the case of the
2582	* DECstation 5000/240 and Alpha AXP, additional kernel
2583	* modifications provide a true microsecond clock and nanosecond
2584	* clock, respectively.
2585	*
2586	* Important note: The kernel discipline is used only if the
2587	* step threshold is less than 0.5 s, as anything higher can
2588	* lead to overflow problems. This might occur if some misguided
2589	* lad set the step threshold to something ridiculous.
2590	*/
2591	if (pll_control && kern_enable) {
2592
2593	#define MOD_BITS (MOD_OFFSET | MOD_MAXERROR | MOD_ESTERROR | MOD_STATUS | MOD_TIMECONST)
2594
2595	/*
2596	* We initialize the structure for the ntp_adjtime()
2597	* system call. We have to convert everything to
2598	* microseconds or nanoseconds first. Do not update the
2599	* system variables if the ext_enable flag is set. In
2600	* this case, the external clock driver will update the
2601	* variables, which will be read later by the local
2602	* clock driver. Afterwards, remember the time and
2603	* frequency offsets for jitter and stability values and
2604	* to update the frequency file.
2605	*/
2606	memset(&ntv, 0, sizeof(ntv));
2607	if (ext_enable) {
2608	ntv.modes = MOD_STATUS;
2609	} else {
2610	#ifdef STA_NANO
2611	ntv.modes = MOD_BITS | MOD_NANO;
2612	#else /* STA_NANO */
2613	ntv.modes = MOD_BITS;
2614	#endif /* STA_NANO */
2615	if (clock_offset < 0)
2616	dtemp = -.5;
2617	else
2618	dtemp = .5;
2619	#ifdef STA_NANO
2620	ntv.offset = (int32)(clock_offset * 1e9 + dtemp);
2621	ntv.constant = sys_poll;
2622	#else /* STA_NANO */
2623	ntv.offset = (int32)(clock_offset * 1e6 + dtemp);
2624	ntv.constant = sys_poll - 4;
2625	#endif /* STA_NANO */
2626	ntv.esterror = (u_int32)(clock_jitter * 1e6);
2627	ntv.maxerror = (u_int32)((sys_rootdelay / 2 + sys_rootdisp) * 1e6);
2628	ntv.status = STA_PLL;
2629
2630	/*
2631	* Enable/disable the PPS if requested.
2632	*/
2633	if (pps_enable) {
2634	if (!(pll_status & STA_PPSTIME))
2635	report_event(EVNT_KERN,
2636	NULL, "PPS enabled");
2637	ntv.status |= STA_PPSTIME | STA_PPSFREQ;
2638	} else {
2639	if (pll_status & STA_PPSTIME)
2640	report_event(EVNT_KERN,
2641	NULL, "PPS disabled");
2642	ntv.status &= ~(STA_PPSTIME | STA_PPSFREQ);
2643	}
2644	if (sys_leap == LEAP_ADDSECOND)
2645	ntv.status |= STA_INS;
2646	else if (sys_leap == LEAP_DELSECOND)
2647	ntv.status |= STA_DEL;
2648	}
2649
2650	/*
2651	* Pass the stuff to the kernel. If it squeals, turn off
2652	* the pps. In any case, fetch the kernel offset,
2653	* frequency and jitter.
2654	*/
2655	if (ntp_adjtime(&ntv) == TIME_ERROR) {
2656	if (!(ntv.status & STA_PPSSIGNAL))
2657	report_event(EVNT_KERN, NULL,
2658	"PPS no signal");
2659	}
2660	pll_status = ntv.status;
2661	#ifdef STA_NANO
2662	clock_offset = ntv.offset / 1e9;
2663	#else /* STA_NANO */
2664	clock_offset = ntv.offset / 1e6;
2665	#endif /* STA_NANO */
2666	clock_frequency = FREQTOD(ntv.freq);
2667
2668	/*
2669	* If the kernel PPS is lit, monitor its performance.
2670	*/
2671	if (ntv.status & STA_PPSTIME) {
2672	#ifdef STA_NANO
2673	clock_jitter = ntv.jitter / 1e9;
2674	#else /* STA_NANO */
2675	clock_jitter = ntv.jitter / 1e6;
2676	#endif /* STA_NANO */
2677	}
2678
2679	#if defined(STA_NANO) && NTP_API == 4
2680	/*
2681	* If the TAI changes, update the kernel TAI.
2682	*/
2683	if (loop_tai != sys_tai) {
2684	loop_tai = sys_tai;
2685	ntv.modes = MOD_TAI;
2686	ntv.constant = sys_tai;
2687	ntp_adjtime(&ntv);
2688	}
2689	#endif /* STA_NANO */
2690	}
2691	#endif /* KERNEL_PLL */
2692	#endif
2693