source: branches/3.2/mindi-busybox/networking/ntpd.c @ 3232

Last change on this file since 3232 was 3232, checked in by bruno, 6 years ago
  • Update mindi-busybox to 1.21.1
  • Property svn:eol-style set to native
File size: 74.0 KB
Line 
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 source tree.
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
31//usage:#define ntpd_trivial_usage
32//usage:    "[-dnqNw"IF_FEATURE_NTPD_SERVER("l")"] [-S PROG] [-p PEER]..."
33//usage:#define ntpd_full_usage "\n\n"
34//usage:       "NTP client/server\n"
35//usage:     "\n    -d  Verbose"
36//usage:     "\n    -n  Do not daemonize"
37//usage:     "\n    -q  Quit after clock is set"
38//usage:     "\n    -N  Run at high priority"
39//usage:     "\n    -w  Do not set time (only query peers), implies -n"
40//usage:    IF_FEATURE_NTPD_SERVER(
41//usage:     "\n    -l  Run as server on port 123"
42//usage:    )
43//usage:     "\n    -S PROG Run PROG after stepping time, stratum change, and every 11 mins"
44//usage:     "\n    -p PEER Obtain time from PEER (may be repeated)"
45
46#include "libbb.h"
47#include <math.h>
48#include <netinet/ip.h> /* For IPTOS_LOWDELAY definition */
49#include <sys/resource.h> /* setpriority */
50#include <sys/timex.h>
51#ifndef IPTOS_LOWDELAY
52# define IPTOS_LOWDELAY 0x10
53#endif
54#ifndef IP_PKTINFO
55# error "Sorry, your kernel has to support IP_PKTINFO"
56#endif
57
58
59/* Verbosity control (max level of -dddd options accepted).
60 * max 5 is very talkative (and bloated). 2 is non-bloated,
61 * production level setting.
62 */
63#define MAX_VERBOSE     2
64
65
66/* High-level description of the algorithm:
67 *
68 * We start running with very small poll_exp, BURSTPOLL,
69 * in order to quickly accumulate INITIAL_SAMPLES datapoints
70 * for each peer. Then, time is stepped if the offset is larger
71 * than STEP_THRESHOLD, otherwise it isn't; anyway, we enlarge
72 * poll_exp to MINPOLL and enter frequency measurement step:
73 * we collect new datapoints but ignore them for WATCH_THRESHOLD
74 * seconds. After WATCH_THRESHOLD seconds we look at accumulated
75 * offset and estimate frequency drift.
76 *
77 * (frequency measurement step seems to not be strictly needed,
78 * it is conditionally disabled with USING_INITIAL_FREQ_ESTIMATION
79 * define set to 0)
80 *
81 * After this, we enter "steady state": we collect a datapoint,
82 * we select the best peer, if this datapoint is not a new one
83 * (IOW: if this datapoint isn't for selected peer), sleep
84 * and collect another one; otherwise, use its offset to update
85 * frequency drift, if offset is somewhat large, reduce poll_exp,
86 * otherwise increase poll_exp.
87 *
88 * If offset is larger than STEP_THRESHOLD, which shouldn't normally
89 * happen, we assume that something "bad" happened (computer
90 * was hibernated, someone set totally wrong date, etc),
91 * then the time is stepped, all datapoints are discarded,
92 * and we go back to steady state.
93 */
94
95#define RETRY_INTERVAL  5       /* on error, retry in N secs */
96#define RESPONSE_INTERVAL 15    /* wait for reply up to N secs */
97#define INITIAL_SAMPLES 4       /* how many samples do we want for init */
98
99/* Clock discipline parameters and constants */
100
101/* Step threshold (sec). std ntpd uses 0.128.
102 * Using exact power of 2 (1/8) results in smaller code */
103#define STEP_THRESHOLD  0.125
104#define WATCH_THRESHOLD 128     /* stepout threshold (sec). std ntpd uses 900 (11 mins (!)) */
105/* NB: set WATCH_THRESHOLD to ~60 when debugging to save time) */
106//UNUSED: #define PANIC_THRESHOLD 1000    /* panic threshold (sec) */
107
108#define FREQ_TOLERANCE  0.000015 /* frequency tolerance (15 PPM) */
109#define BURSTPOLL       0       /* initial poll */
110#define MINPOLL         5       /* minimum poll interval. std ntpd uses 6 (6: 64 sec) */
111/* If offset > discipline_jitter * POLLADJ_GATE, and poll interval is >= 2^BIGPOLL,
112 * then it is decreased _at once_. (If < 2^BIGPOLL, it will be decreased _eventually_).
113 */
114#define BIGPOLL         10      /* 2^10 sec ~= 17 min */
115#define MAXPOLL         12      /* maximum poll interval (12: 1.1h, 17: 36.4h). std ntpd uses 17 */
116/* Actively lower poll when we see such big offsets.
117 * With STEP_THRESHOLD = 0.125, it means we try to sync more aggressively
118 * if offset increases over ~0.04 sec */
119#define POLLDOWN_OFFSET (STEP_THRESHOLD / 3)
120#define MINDISP         0.01    /* minimum dispersion (sec) */
121#define MAXDISP         16      /* maximum dispersion (sec) */
122#define MAXSTRAT        16      /* maximum stratum (infinity metric) */
123#define MAXDIST         1       /* distance threshold (sec) */
124#define MIN_SELECTED    1       /* minimum intersection survivors */
125#define MIN_CLUSTERED   3       /* minimum cluster survivors */
126
127#define MAXDRIFT        0.000500 /* frequency drift we can correct (500 PPM) */
128
129/* Poll-adjust threshold.
130 * When we see that offset is small enough compared to discipline jitter,
131 * we grow a counter: += MINPOLL. When counter goes over POLLADJ_LIMIT,
132 * we poll_exp++. If offset isn't small, counter -= poll_exp*2,
133 * and when it goes below -POLLADJ_LIMIT, we poll_exp--.
134 * (Bumped from 30 to 40 since otherwise I often see poll_exp going *2* steps down)
135 */
136#define POLLADJ_LIMIT   40
137/* If offset < discipline_jitter * POLLADJ_GATE, then we decide to increase
138 * poll interval (we think we can't improve timekeeping
139 * by staying at smaller poll).
140 */
141#define POLLADJ_GATE    4
142#define TIMECONST_HACK_GATE 2
143/* Compromise Allan intercept (sec). doc uses 1500, std ntpd uses 512 */
144#define ALLAN           512
145/* PLL loop gain */
146#define PLL             65536
147/* FLL loop gain [why it depends on MAXPOLL??] */
148#define FLL             (MAXPOLL + 1)
149/* Parameter averaging constant */
150#define AVG             4
151
152
153enum {
154    NTP_VERSION     = 4,
155    NTP_MAXSTRATUM  = 15,
156
157    NTP_DIGESTSIZE     = 16,
158    NTP_MSGSIZE_NOAUTH = 48,
159    NTP_MSGSIZE        = (NTP_MSGSIZE_NOAUTH + 4 + NTP_DIGESTSIZE),
160
161    /* Status Masks */
162    MODE_MASK       = (7 << 0),
163    VERSION_MASK    = (7 << 3),
164    VERSION_SHIFT   = 3,
165    LI_MASK         = (3 << 6),
166
167    /* Leap Second Codes (high order two bits of m_status) */
168    LI_NOWARNING    = (0 << 6),    /* no warning */
169    LI_PLUSSEC      = (1 << 6),    /* add a second (61 seconds) */
170    LI_MINUSSEC     = (2 << 6),    /* minus a second (59 seconds) */
171    LI_ALARM        = (3 << 6),    /* alarm condition */
172
173    /* Mode values */
174    MODE_RES0       = 0,    /* reserved */
175    MODE_SYM_ACT    = 1,    /* symmetric active */
176    MODE_SYM_PAS    = 2,    /* symmetric passive */
177    MODE_CLIENT     = 3,    /* client */
178    MODE_SERVER     = 4,    /* server */
179    MODE_BROADCAST  = 5,    /* broadcast */
180    MODE_RES1       = 6,    /* reserved for NTP control message */
181    MODE_RES2       = 7,    /* reserved for private use */
182};
183
184//TODO: better base selection
185#define OFFSET_1900_1970 2208988800UL  /* 1970 - 1900 in seconds */
186
187#define NUM_DATAPOINTS  8
188
189typedef struct {
190    uint32_t int_partl;
191    uint32_t fractionl;
192} l_fixedpt_t;
193
194typedef struct {
195    uint16_t int_parts;
196    uint16_t fractions;
197} s_fixedpt_t;
198
199typedef struct {
200    uint8_t     m_status;     /* status of local clock and leap info */
201    uint8_t     m_stratum;
202    uint8_t     m_ppoll;      /* poll value */
203    int8_t      m_precision_exp;
204    s_fixedpt_t m_rootdelay;
205    s_fixedpt_t m_rootdisp;
206    uint32_t    m_refid;
207    l_fixedpt_t m_reftime;
208    l_fixedpt_t m_orgtime;
209    l_fixedpt_t m_rectime;
210    l_fixedpt_t m_xmttime;
211    uint32_t    m_keyid;
212    uint8_t     m_digest[NTP_DIGESTSIZE];
213} msg_t;
214
215typedef struct {
216    double d_offset;
217    double d_recv_time;
218    double d_dispersion;
219} datapoint_t;
220
221typedef struct {
222    len_and_sockaddr *p_lsa;
223    char             *p_dotted;
224    int              p_fd;
225    int              datapoint_idx;
226    uint32_t         lastpkt_refid;
227    uint8_t          lastpkt_status;
228    uint8_t          lastpkt_stratum;
229    uint8_t          reachable_bits;
230        /* when to send new query (if p_fd == -1)
231         * or when receive times out (if p_fd >= 0): */
232    double           next_action_time;
233    double           p_xmttime;
234    double           lastpkt_recv_time;
235    double           lastpkt_delay;
236    double           lastpkt_rootdelay;
237    double           lastpkt_rootdisp;
238    /* produced by filter algorithm: */
239    double           filter_offset;
240    double           filter_dispersion;
241    double           filter_jitter;
242    datapoint_t      filter_datapoint[NUM_DATAPOINTS];
243    /* last sent packet: */
244    msg_t            p_xmt_msg;
245} peer_t;
246
247
248#define USING_KERNEL_PLL_LOOP          1
249#define USING_INITIAL_FREQ_ESTIMATION  0
250
251enum {
252    OPT_n = (1 << 0),
253    OPT_q = (1 << 1),
254    OPT_N = (1 << 2),
255    OPT_x = (1 << 3),
256    /* Insert new options above this line. */
257    /* Non-compat options: */
258    OPT_w = (1 << 4),
259    OPT_p = (1 << 5),
260    OPT_S = (1 << 6),
261    OPT_l = (1 << 7) * ENABLE_FEATURE_NTPD_SERVER,
262    /* We hijack some bits for other purposes */
263    OPT_qq = (1 << 31),
264};
265
266struct globals {
267    double   cur_time;
268    /* total round trip delay to currently selected reference clock */
269    double   rootdelay;
270    /* reference timestamp: time when the system clock was last set or corrected */
271    double   reftime;
272    /* total dispersion to currently selected reference clock */
273    double   rootdisp;
274
275    double   last_script_run;
276    char     *script_name;
277    llist_t  *ntp_peers;
278#if ENABLE_FEATURE_NTPD_SERVER
279    int      listen_fd;
280# define G_listen_fd (G.listen_fd)
281#else
282# define G_listen_fd (-1)
283#endif
284    unsigned verbose;
285    unsigned peer_cnt;
286    /* refid: 32-bit code identifying the particular server or reference clock
287     * in stratum 0 packets this is a four-character ASCII string,
288     * called the kiss code, used for debugging and monitoring
289     * in stratum 1 packets this is a four-character ASCII string
290     * assigned to the reference clock by IANA. Example: "GPS "
291     * in stratum 2+ packets, it's IPv4 address or 4 first bytes
292     * of MD5 hash of IPv6
293     */
294    uint32_t refid;
295    uint8_t  ntp_status;
296    /* precision is defined as the larger of the resolution and time to
297     * read the clock, in log2 units.  For instance, the precision of a
298     * mains-frequency clock incrementing at 60 Hz is 16 ms, even when the
299     * system clock hardware representation is to the nanosecond.
300     *
301     * Delays, jitters of various kinds are clamped down to precision.
302     *
303     * If precision_sec is too large, discipline_jitter gets clamped to it
304     * and if offset is smaller than discipline_jitter * POLLADJ_GATE, poll
305     * interval grows even though we really can benefit from staying at
306     * smaller one, collecting non-lagged datapoits and correcting offset.
307     * (Lagged datapoits exist when poll_exp is large but we still have
308     * systematic offset error - the time distance between datapoints
309     * is significant and older datapoints have smaller offsets.
310     * This makes our offset estimation a bit smaller than reality)
311     * Due to this effect, setting G_precision_sec close to
312     * STEP_THRESHOLD isn't such a good idea - offsets may grow
313     * too big and we will step. I observed it with -6.
314     *
315     * OTOH, setting precision_sec far too small would result in futile
316     * attempts to syncronize to an unachievable precision.
317     *
318     * -6 is 1/64 sec, -7 is 1/128 sec and so on.
319     * -8 is 1/256 ~= 0.003906 (worked well for me --vda)
320     * -9 is 1/512 ~= 0.001953 (let's try this for some time)
321     */
322#define G_precision_exp  -9
323    /*
324     * G_precision_exp is used only for construction outgoing packets.
325     * It's ok to set G_precision_sec to a slightly different value
326     * (One which is "nicer looking" in logs).
327     * Exact value would be (1.0 / (1 << (- G_precision_exp))):
328     */
329#define G_precision_sec  0.002
330    uint8_t  stratum;
331    /* Bool. After set to 1, never goes back to 0: */
332    smallint initial_poll_complete;
333
334#define STATE_NSET      0       /* initial state, "nothing is set" */
335//#define STATE_FSET    1       /* frequency set from file */
336#define STATE_SPIK      2       /* spike detected */
337//#define STATE_FREQ    3       /* initial frequency */
338#define STATE_SYNC      4       /* clock synchronized (normal operation) */
339    uint8_t  discipline_state;      // doc calls it c.state
340    uint8_t  poll_exp;              // s.poll
341    int      polladj_count;         // c.count
342    long     kernel_freq_drift;
343    peer_t   *last_update_peer;
344    double   last_update_offset;    // c.last
345    double   last_update_recv_time; // s.t
346    double   discipline_jitter;     // c.jitter
347    /* Since we only compare it with ints, can simplify code
348     * by not making this variable floating point:
349     */
350    unsigned offset_to_jitter_ratio;
351    //double   cluster_offset;        // s.offset
352    //double   cluster_jitter;        // s.jitter
353#if !USING_KERNEL_PLL_LOOP
354    double   discipline_freq_drift; // c.freq
355    /* Maybe conditionally calculate wander? it's used only for logging */
356    double   discipline_wander;     // c.wander
357#endif
358};
359#define G (*ptr_to_globals)
360
361static const int const_IPTOS_LOWDELAY = IPTOS_LOWDELAY;
362
363
364#define VERB1 if (MAX_VERBOSE && G.verbose)
365#define VERB2 if (MAX_VERBOSE >= 2 && G.verbose >= 2)
366#define VERB3 if (MAX_VERBOSE >= 3 && G.verbose >= 3)
367#define VERB4 if (MAX_VERBOSE >= 4 && G.verbose >= 4)
368#define VERB5 if (MAX_VERBOSE >= 5 && G.verbose >= 5)
369
370
371static double LOG2D(int a)
372{
373    if (a < 0)
374        return 1.0 / (1UL << -a);
375    return 1UL << a;
376}
377static ALWAYS_INLINE double SQUARE(double x)
378{
379    return x * x;
380}
381static ALWAYS_INLINE double MAXD(double a, double b)
382{
383    if (a > b)
384        return a;
385    return b;
386}
387static ALWAYS_INLINE double MIND(double a, double b)
388{
389    if (a < b)
390        return a;
391    return b;
392}
393static NOINLINE double my_SQRT(double X)
394{
395    union {
396        float   f;
397        int32_t i;
398    } v;
399    double invsqrt;
400    double Xhalf = X * 0.5;
401
402    /* Fast and good approximation to 1/sqrt(X), black magic */
403    v.f = X;
404    /*v.i = 0x5f3759df - (v.i >> 1);*/
405    v.i = 0x5f375a86 - (v.i >> 1); /* - this constant is slightly better */
406    invsqrt = v.f; /* better than 0.2% accuracy */
407
408    /* Refining it using Newton's method: x1 = x0 - f(x0)/f'(x0)
409     * f(x) = 1/(x*x) - X  (f==0 when x = 1/sqrt(X))
410     * f'(x) = -2/(x*x*x)
411     * f(x)/f'(x) = (X - 1/(x*x)) / (2/(x*x*x)) = X*x*x*x/2 - x/2
412     * x1 = x0 - (X*x0*x0*x0/2 - x0/2) = 1.5*x0 - X*x0*x0*x0/2 = x0*(1.5 - (X/2)*x0*x0)
413     */
414    invsqrt = invsqrt * (1.5 - Xhalf * invsqrt * invsqrt); /* ~0.05% accuracy */
415    /* invsqrt = invsqrt * (1.5 - Xhalf * invsqrt * invsqrt); 2nd iter: ~0.0001% accuracy */
416    /* With 4 iterations, more than half results will be exact,
417     * at 6th iterations result stabilizes with about 72% results exact.
418     * We are well satisfied with 0.05% accuracy.
419     */
420
421    return X * invsqrt; /* X * 1/sqrt(X) ~= sqrt(X) */
422}
423static ALWAYS_INLINE double SQRT(double X)
424{
425    /* If this arch doesn't use IEEE 754 floats, fall back to using libm */
426    if (sizeof(float) != 4)
427        return sqrt(X);
428
429    /* This avoids needing libm, saves about 0.5k on x86-32 */
430    return my_SQRT(X);
431}
432
433static double
434gettime1900d(void)
435{
436    struct timeval tv;
437    gettimeofday(&tv, NULL); /* never fails */
438    G.cur_time = tv.tv_sec + (1.0e-6 * tv.tv_usec) + OFFSET_1900_1970;
439    return G.cur_time;
440}
441
442static void
443d_to_tv(double d, struct timeval *tv)
444{
445    tv->tv_sec = (long)d;
446    tv->tv_usec = (d - tv->tv_sec) * 1000000;
447}
448
449static double
450lfp_to_d(l_fixedpt_t lfp)
451{
452    double ret;
453    lfp.int_partl = ntohl(lfp.int_partl);
454    lfp.fractionl = ntohl(lfp.fractionl);
455    ret = (double)lfp.int_partl + ((double)lfp.fractionl / UINT_MAX);
456    return ret;
457}
458static double
459sfp_to_d(s_fixedpt_t sfp)
460{
461    double ret;
462    sfp.int_parts = ntohs(sfp.int_parts);
463    sfp.fractions = ntohs(sfp.fractions);
464    ret = (double)sfp.int_parts + ((double)sfp.fractions / USHRT_MAX);
465    return ret;
466}
467#if ENABLE_FEATURE_NTPD_SERVER
468static l_fixedpt_t
469d_to_lfp(double d)
470{
471    l_fixedpt_t lfp;
472    lfp.int_partl = (uint32_t)d;
473    lfp.fractionl = (uint32_t)((d - lfp.int_partl) * UINT_MAX);
474    lfp.int_partl = htonl(lfp.int_partl);
475    lfp.fractionl = htonl(lfp.fractionl);
476    return lfp;
477}
478static s_fixedpt_t
479d_to_sfp(double d)
480{
481    s_fixedpt_t sfp;
482    sfp.int_parts = (uint16_t)d;
483    sfp.fractions = (uint16_t)((d - sfp.int_parts) * USHRT_MAX);
484    sfp.int_parts = htons(sfp.int_parts);
485    sfp.fractions = htons(sfp.fractions);
486    return sfp;
487}
488#endif
489
490static double
491dispersion(const datapoint_t *dp)
492{
493    return dp->d_dispersion + FREQ_TOLERANCE * (G.cur_time - dp->d_recv_time);
494}
495
496static double
497root_distance(peer_t *p)
498{
499    /* The root synchronization distance is the maximum error due to
500     * all causes of the local clock relative to the primary server.
501     * It is defined as half the total delay plus total dispersion
502     * plus peer jitter.
503     */
504    return MAXD(MINDISP, p->lastpkt_rootdelay + p->lastpkt_delay) / 2
505        + p->lastpkt_rootdisp
506        + p->filter_dispersion
507        + FREQ_TOLERANCE * (G.cur_time - p->lastpkt_recv_time)
508        + p->filter_jitter;
509}
510
511static void
512set_next(peer_t *p, unsigned t)
513{
514    p->next_action_time = G.cur_time + t;
515}
516
517/*
518 * Peer clock filter and its helpers
519 */
520static void
521filter_datapoints(peer_t *p)
522{
523    int i, idx;
524    double sum, wavg;
525    datapoint_t *fdp;
526
527#if 0
528/* Simulations have shown that use of *averaged* offset for p->filter_offset
529 * is in fact worse than simply using last received one: with large poll intervals
530 * (>= 2048) averaging code uses offset values which are outdated by hours,
531 * and time/frequency correction goes totally wrong when fed essentially bogus offsets.
532 */
533    int got_newest;
534    double minoff, maxoff, w;
535    double x = x; /* for compiler */
536    double oldest_off = oldest_off;
537    double oldest_age = oldest_age;
538    double newest_off = newest_off;
539    double newest_age = newest_age;
540
541    fdp = p->filter_datapoint;
542
543    minoff = maxoff = fdp[0].d_offset;
544    for (i = 1; i < NUM_DATAPOINTS; i++) {
545        if (minoff > fdp[i].d_offset)
546            minoff = fdp[i].d_offset;
547        if (maxoff < fdp[i].d_offset)
548            maxoff = fdp[i].d_offset;
549    }
550
551    idx = p->datapoint_idx; /* most recent datapoint's index */
552    /* Average offset:
553     * Drop two outliers and take weighted average of the rest:
554     * most_recent/2 + older1/4 + older2/8 ... + older5/32 + older6/32
555     * we use older6/32, not older6/64 since sum of weights should be 1:
556     * 1/2 + 1/4 + 1/8 + 1/16 + 1/32 + 1/32 = 1
557     */
558    wavg = 0;
559    w = 0.5;
560    /*                     n-1
561     *                     ---    dispersion(i)
562     * filter_dispersion =  \     -------------
563     *                      /       (i+1)
564     *                     ---     2
565     *                     i=0
566     */
567    got_newest = 0;
568    sum = 0;
569    for (i = 0; i < NUM_DATAPOINTS; i++) {
570        VERB4 {
571            bb_error_msg("datapoint[%d]: off:%f disp:%f(%f) age:%f%s",
572                i,
573                fdp[idx].d_offset,
574                fdp[idx].d_dispersion, dispersion(&fdp[idx]),
575                G.cur_time - fdp[idx].d_recv_time,
576                (minoff == fdp[idx].d_offset || maxoff == fdp[idx].d_offset)
577                    ? " (outlier by offset)" : ""
578            );
579        }
580
581        sum += dispersion(&fdp[idx]) / (2 << i);
582
583        if (minoff == fdp[idx].d_offset) {
584            minoff -= 1; /* so that we don't match it ever again */
585        } else
586        if (maxoff == fdp[idx].d_offset) {
587            maxoff += 1;
588        } else {
589            oldest_off = fdp[idx].d_offset;
590            oldest_age = G.cur_time - fdp[idx].d_recv_time;
591            if (!got_newest) {
592                got_newest = 1;
593                newest_off = oldest_off;
594                newest_age = oldest_age;
595            }
596            x = oldest_off * w;
597            wavg += x;
598            w /= 2;
599        }
600
601        idx = (idx - 1) & (NUM_DATAPOINTS - 1);
602    }
603    p->filter_dispersion = sum;
604    wavg += x; /* add another older6/64 to form older6/32 */
605    /* Fix systematic underestimation with large poll intervals.
606     * Imagine that we still have a bit of uncorrected drift,
607     * and poll interval is big (say, 100 sec). Offsets form a progression:
608     * 0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 - 0.7 is most recent.
609     * The algorithm above drops 0.0 and 0.7 as outliers,
610     * and then we have this estimation, ~25% off from 0.7:
611     * 0.1/32 + 0.2/32 + 0.3/16 + 0.4/8 + 0.5/4 + 0.6/2 = 0.503125
612     */
613    x = oldest_age - newest_age;
614    if (x != 0) {
615        x = newest_age / x; /* in above example, 100 / (600 - 100) */
616        if (x < 1) { /* paranoia check */
617            x = (newest_off - oldest_off) * x; /* 0.5 * 100/500 = 0.1 */
618            wavg += x;
619        }
620    }
621    p->filter_offset = wavg;
622
623#else
624
625    fdp = p->filter_datapoint;
626    idx = p->datapoint_idx; /* most recent datapoint's index */
627
628    /* filter_offset: simply use the most recent value */
629    p->filter_offset = fdp[idx].d_offset;
630
631    /*                     n-1
632     *                     ---    dispersion(i)
633     * filter_dispersion =  \     -------------
634     *                      /       (i+1)
635     *                     ---     2
636     *                     i=0
637     */
638    wavg = 0;
639    sum = 0;
640    for (i = 0; i < NUM_DATAPOINTS; i++) {
641        sum += dispersion(&fdp[idx]) / (2 << i);
642        wavg += fdp[idx].d_offset;
643        idx = (idx - 1) & (NUM_DATAPOINTS - 1);
644    }
645    wavg /= NUM_DATAPOINTS;
646    p->filter_dispersion = sum;
647#endif
648
649    /*                  +-----                 -----+ ^ 1/2
650     *                  |       n-1                 |
651     *                  |       ---                 |
652     *                  |  1    \                2  |
653     * filter_jitter =  | --- * /  (avg-offset_j)   |
654     *                  |  n    ---                 |
655     *                  |       j=0                 |
656     *                  +-----                 -----+
657     * where n is the number of valid datapoints in the filter (n > 1);
658     * if filter_jitter < precision then filter_jitter = precision
659     */
660    sum = 0;
661    for (i = 0; i < NUM_DATAPOINTS; i++) {
662        sum += SQUARE(wavg - fdp[i].d_offset);
663    }
664    sum = SQRT(sum / NUM_DATAPOINTS);
665    p->filter_jitter = sum > G_precision_sec ? sum : G_precision_sec;
666
667    VERB3 bb_error_msg("filter offset:%+f disp:%f jitter:%f",
668            p->filter_offset,
669            p->filter_dispersion,
670            p->filter_jitter);
671}
672
673static void
674reset_peer_stats(peer_t *p, double offset)
675{
676    int i;
677    bool small_ofs = fabs(offset) < 16 * STEP_THRESHOLD;
678
679    for (i = 0; i < NUM_DATAPOINTS; i++) {
680        if (small_ofs) {
681            p->filter_datapoint[i].d_recv_time += offset;
682            if (p->filter_datapoint[i].d_offset != 0) {
683                p->filter_datapoint[i].d_offset -= offset;
684                //bb_error_msg("p->filter_datapoint[%d].d_offset %f -> %f",
685                //  i,
686                //  p->filter_datapoint[i].d_offset + offset,
687                //  p->filter_datapoint[i].d_offset);
688            }
689        } else {
690            p->filter_datapoint[i].d_recv_time  = G.cur_time;
691            p->filter_datapoint[i].d_offset     = 0;
692            p->filter_datapoint[i].d_dispersion = MAXDISP;
693        }
694    }
695    if (small_ofs) {
696        p->lastpkt_recv_time += offset;
697    } else {
698        p->reachable_bits = 0;
699        p->lastpkt_recv_time = G.cur_time;
700    }
701    filter_datapoints(p); /* recalc p->filter_xxx */
702    VERB5 bb_error_msg("%s->lastpkt_recv_time=%f", p->p_dotted, p->lastpkt_recv_time);
703}
704
705static void
706add_peers(char *s)
707{
708    peer_t *p;
709
710    p = xzalloc(sizeof(*p));
711    p->p_lsa = xhost2sockaddr(s, 123);
712    p->p_dotted = xmalloc_sockaddr2dotted_noport(&p->p_lsa->u.sa);
713    p->p_fd = -1;
714    p->p_xmt_msg.m_status = MODE_CLIENT | (NTP_VERSION << 3);
715    p->next_action_time = G.cur_time; /* = set_next(p, 0); */
716    reset_peer_stats(p, 16 * STEP_THRESHOLD);
717
718    llist_add_to(&G.ntp_peers, p);
719    G.peer_cnt++;
720}
721
722static int
723do_sendto(int fd,
724        const struct sockaddr *from, const struct sockaddr *to, socklen_t addrlen,
725        msg_t *msg, ssize_t len)
726{
727    ssize_t ret;
728
729    errno = 0;
730    if (!from) {
731        ret = sendto(fd, msg, len, MSG_DONTWAIT, to, addrlen);
732    } else {
733        ret = send_to_from(fd, msg, len, MSG_DONTWAIT, to, from, addrlen);
734    }
735    if (ret != len) {
736        bb_perror_msg("send failed");
737        return -1;
738    }
739    return 0;
740}
741
742static void
743send_query_to_peer(peer_t *p)
744{
745    /* Why do we need to bind()?
746     * See what happens when we don't bind:
747     *
748     * socket(PF_INET, SOCK_DGRAM, IPPROTO_IP) = 3
749     * setsockopt(3, SOL_IP, IP_TOS, [16], 4) = 0
750     * gettimeofday({1259071266, 327885}, NULL) = 0
751     * sendto(3, "xxx", 48, MSG_DONTWAIT, {sa_family=AF_INET, sin_port=htons(123), sin_addr=inet_addr("10.34.32.125")}, 16) = 48
752     * ^^^ we sent it from some source port picked by kernel.
753     * time(NULL)              = 1259071266
754     * write(2, "ntpd: entering poll 15 secs\n", 28) = 28
755     * poll([{fd=3, events=POLLIN}], 1, 15000) = 1 ([{fd=3, revents=POLLIN}])
756     * recv(3, "yyy", 68, MSG_DONTWAIT) = 48
757     * ^^^ this recv will receive packets to any local port!
758     *
759     * Uncomment this and use strace to see it in action:
760     */
761#define PROBE_LOCAL_ADDR /* { len_and_sockaddr lsa; lsa.len = LSA_SIZEOF_SA; getsockname(p->query.fd, &lsa.u.sa, &lsa.len); } */
762
763    if (p->p_fd == -1) {
764        int fd, family;
765        len_and_sockaddr *local_lsa;
766
767        family = p->p_lsa->u.sa.sa_family;
768        p->p_fd = fd = xsocket_type(&local_lsa, family, SOCK_DGRAM);
769        /* local_lsa has "null" address and port 0 now.
770         * bind() ensures we have a *particular port* selected by kernel
771         * and remembered in p->p_fd, thus later recv(p->p_fd)
772         * receives only packets sent to this port.
773         */
774        PROBE_LOCAL_ADDR
775        xbind(fd, &local_lsa->u.sa, local_lsa->len);
776        PROBE_LOCAL_ADDR
777#if ENABLE_FEATURE_IPV6
778        if (family == AF_INET)
779#endif
780            setsockopt(fd, IPPROTO_IP, IP_TOS, &const_IPTOS_LOWDELAY, sizeof(const_IPTOS_LOWDELAY));
781        free(local_lsa);
782    }
783
784    /* Emit message _before_ attempted send. Think of a very short
785     * roundtrip networks: we need to go back to recv loop ASAP,
786     * to reduce delay. Printing messages after send works against that.
787     */
788    VERB1 bb_error_msg("sending query to %s", p->p_dotted);
789
790    /*
791     * Send out a random 64-bit number as our transmit time.  The NTP
792     * server will copy said number into the originate field on the
793     * response that it sends us.  This is totally legal per the SNTP spec.
794     *
795     * The impact of this is two fold: we no longer send out the current
796     * system time for the world to see (which may aid an attacker), and
797     * it gives us a (not very secure) way of knowing that we're not
798     * getting spoofed by an attacker that can't capture our traffic
799     * but can spoof packets from the NTP server we're communicating with.
800     *
801     * Save the real transmit timestamp locally.
802     */
803    p->p_xmt_msg.m_xmttime.int_partl = random();
804    p->p_xmt_msg.m_xmttime.fractionl = random();
805    p->p_xmttime = gettime1900d();
806
807    if (do_sendto(p->p_fd, /*from:*/ NULL, /*to:*/ &p->p_lsa->u.sa, /*addrlen:*/ p->p_lsa->len,
808            &p->p_xmt_msg, NTP_MSGSIZE_NOAUTH) == -1
809    ) {
810        close(p->p_fd);
811        p->p_fd = -1;
812        set_next(p, RETRY_INTERVAL);
813        return;
814    }
815
816    p->reachable_bits <<= 1;
817    set_next(p, RESPONSE_INTERVAL);
818}
819
820
821/* Note that there is no provision to prevent several run_scripts
822 * to be done in quick succession. In fact, it happens rather often
823 * if initial syncronization results in a step.
824 * You will see "step" and then "stratum" script runs, sometimes
825 * as close as only 0.002 seconds apart.
826 * Script should be ready to deal with this.
827 */
828static void run_script(const char *action, double offset)
829{
830    char *argv[3];
831    char *env1, *env2, *env3, *env4;
832
833    if (!G.script_name)
834        return;
835
836    argv[0] = (char*) G.script_name;
837    argv[1] = (char*) action;
838    argv[2] = NULL;
839
840    VERB1 bb_error_msg("executing '%s %s'", G.script_name, action);
841
842    env1 = xasprintf("%s=%u", "stratum", G.stratum);
843    putenv(env1);
844    env2 = xasprintf("%s=%ld", "freq_drift_ppm", G.kernel_freq_drift);
845    putenv(env2);
846    env3 = xasprintf("%s=%u", "poll_interval", 1 << G.poll_exp);
847    putenv(env3);
848    env4 = xasprintf("%s=%f", "offset", offset);
849    putenv(env4);
850    /* Other items of potential interest: selected peer,
851     * rootdelay, reftime, rootdisp, refid, ntp_status,
852     * last_update_offset, last_update_recv_time, discipline_jitter,
853     * how many peers have reachable_bits = 0?
854     */
855
856    /* Don't want to wait: it may run hwclock --systohc, and that
857     * may take some time (seconds): */
858    /*spawn_and_wait(argv);*/
859    spawn(argv);
860
861    unsetenv("stratum");
862    unsetenv("freq_drift_ppm");
863    unsetenv("poll_interval");
864    unsetenv("offset");
865    free(env1);
866    free(env2);
867    free(env3);
868    free(env4);
869
870    G.last_script_run = G.cur_time;
871}
872
873static NOINLINE void
874step_time(double offset)
875{
876    llist_t *item;
877    double dtime;
878    struct timeval tvc, tvn;
879    char buf[sizeof("yyyy-mm-dd hh:mm:ss") + /*paranoia:*/ 4];
880    time_t tval;
881
882    gettimeofday(&tvc, NULL); /* never fails */
883    dtime = tvc.tv_sec + (1.0e-6 * tvc.tv_usec) + offset;
884    d_to_tv(dtime, &tvn);
885    if (settimeofday(&tvn, NULL) == -1)
886        bb_perror_msg_and_die("settimeofday");
887
888    VERB2 {
889        tval = tvc.tv_sec;
890        strftime(buf, sizeof(buf), "%Y-%m-%d %H:%M:%S", localtime(&tval));
891        bb_error_msg("current time is %s.%06u", buf, (unsigned)tvc.tv_usec);
892    }
893    tval = tvn.tv_sec;
894    strftime(buf, sizeof(buf), "%Y-%m-%d %H:%M:%S", localtime(&tval));
895    bb_error_msg("setting time to %s.%06u (offset %+fs)", buf, (unsigned)tvn.tv_usec, offset);
896
897    /* Correct various fields which contain time-relative values: */
898
899    /* Globals: */
900    G.cur_time += offset;
901    G.last_update_recv_time += offset;
902    G.last_script_run += offset;
903
904    /* p->lastpkt_recv_time, p->next_action_time and such: */
905    for (item = G.ntp_peers; item != NULL; item = item->link) {
906        peer_t *pp = (peer_t *) item->data;
907        reset_peer_stats(pp, offset);
908        //bb_error_msg("offset:%+f pp->next_action_time:%f -> %f",
909        //  offset, pp->next_action_time, pp->next_action_time + offset);
910        pp->next_action_time += offset;
911        if (pp->p_fd >= 0) {
912            /* We wait for reply from this peer too.
913             * But due to step we are doing, reply's data is no longer
914             * useful (in fact, it'll be bogus). Stop waiting for it.
915             */
916            close(pp->p_fd);
917            pp->p_fd = -1;
918            set_next(pp, RETRY_INTERVAL);
919        }
920    }
921}
922
923
924/*
925 * Selection and clustering, and their helpers
926 */
927typedef struct {
928    peer_t *p;
929    int    type;
930    double edge;
931    double opt_rd; /* optimization */
932} point_t;
933static int
934compare_point_edge(const void *aa, const void *bb)
935{
936    const point_t *a = aa;
937    const point_t *b = bb;
938    if (a->edge < b->edge) {
939        return -1;
940    }
941    return (a->edge > b->edge);
942}
943typedef struct {
944    peer_t *p;
945    double metric;
946} survivor_t;
947static int
948compare_survivor_metric(const void *aa, const void *bb)
949{
950    const survivor_t *a = aa;
951    const survivor_t *b = bb;
952    if (a->metric < b->metric) {
953        return -1;
954    }
955    return (a->metric > b->metric);
956}
957static int
958fit(peer_t *p, double rd)
959{
960    if ((p->reachable_bits & (p->reachable_bits-1)) == 0) {
961        /* One or zero bits in reachable_bits */
962        VERB3 bb_error_msg("peer %s unfit for selection: unreachable", p->p_dotted);
963        return 0;
964    }
965#if 0 /* we filter out such packets earlier */
966    if ((p->lastpkt_status & LI_ALARM) == LI_ALARM
967     || p->lastpkt_stratum >= MAXSTRAT
968    ) {
969        VERB3 bb_error_msg("peer %s unfit for selection: bad status/stratum", p->p_dotted);
970        return 0;
971    }
972#endif
973    /* rd is root_distance(p) */
974    if (rd > MAXDIST + FREQ_TOLERANCE * (1 << G.poll_exp)) {
975        VERB3 bb_error_msg("peer %s unfit for selection: root distance too high", p->p_dotted);
976        return 0;
977    }
978//TODO
979//  /* Do we have a loop? */
980//  if (p->refid == p->dstaddr || p->refid == s.refid)
981//      return 0;
982    return 1;
983}
984static peer_t*
985select_and_cluster(void)
986{
987    peer_t     *p;
988    llist_t    *item;
989    int        i, j;
990    int        size = 3 * G.peer_cnt;
991    /* for selection algorithm */
992    point_t    point[size];
993    unsigned   num_points, num_candidates;
994    double     low, high;
995    unsigned   num_falsetickers;
996    /* for cluster algorithm */
997    survivor_t survivor[size];
998    unsigned   num_survivors;
999
1000    /* Selection */
1001
1002    num_points = 0;
1003    item = G.ntp_peers;
1004    if (G.initial_poll_complete) while (item != NULL) {
1005        double rd, offset;
1006
1007        p = (peer_t *) item->data;
1008        rd = root_distance(p);
1009        offset = p->filter_offset;
1010        if (!fit(p, rd)) {
1011            item = item->link;
1012            continue;
1013        }
1014
1015        VERB4 bb_error_msg("interval: [%f %f %f] %s",
1016                offset - rd,
1017                offset,
1018                offset + rd,
1019                p->p_dotted
1020        );
1021        point[num_points].p = p;
1022        point[num_points].type = -1;
1023        point[num_points].edge = offset - rd;
1024        point[num_points].opt_rd = rd;
1025        num_points++;
1026        point[num_points].p = p;
1027        point[num_points].type = 0;
1028        point[num_points].edge = offset;
1029        point[num_points].opt_rd = rd;
1030        num_points++;
1031        point[num_points].p = p;
1032        point[num_points].type = 1;
1033        point[num_points].edge = offset + rd;
1034        point[num_points].opt_rd = rd;
1035        num_points++;
1036        item = item->link;
1037    }
1038    num_candidates = num_points / 3;
1039    if (num_candidates == 0) {
1040        VERB3 bb_error_msg("no valid datapoints, no peer selected");
1041        return NULL;
1042    }
1043//TODO: sorting does not seem to be done in reference code
1044    qsort(point, num_points, sizeof(point[0]), compare_point_edge);
1045
1046    /* Start with the assumption that there are no falsetickers.
1047     * Attempt to find a nonempty intersection interval containing
1048     * the midpoints of all truechimers.
1049     * If a nonempty interval cannot be found, increase the number
1050     * of assumed falsetickers by one and try again.
1051     * If a nonempty interval is found and the number of falsetickers
1052     * is less than the number of truechimers, a majority has been found
1053     * and the midpoint of each truechimer represents
1054     * the candidates available to the cluster algorithm.
1055     */
1056    num_falsetickers = 0;
1057    while (1) {
1058        int c;
1059        unsigned num_midpoints = 0;
1060
1061        low = 1 << 9;
1062        high = - (1 << 9);
1063        c = 0;
1064        for (i = 0; i < num_points; i++) {
1065            /* We want to do:
1066             * if (point[i].type == -1) c++;
1067             * if (point[i].type == 1) c--;
1068             * and it's simpler to do it this way:
1069             */
1070            c -= point[i].type;
1071            if (c >= num_candidates - num_falsetickers) {
1072                /* If it was c++ and it got big enough... */
1073                low = point[i].edge;
1074                break;
1075            }
1076            if (point[i].type == 0)
1077                num_midpoints++;
1078        }
1079        c = 0;
1080        for (i = num_points-1; i >= 0; i--) {
1081            c += point[i].type;
1082            if (c >= num_candidates - num_falsetickers) {
1083                high = point[i].edge;
1084                break;
1085            }
1086            if (point[i].type == 0)
1087                num_midpoints++;
1088        }
1089        /* If the number of midpoints is greater than the number
1090         * of allowed falsetickers, the intersection contains at
1091         * least one truechimer with no midpoint - bad.
1092         * Also, interval should be nonempty.
1093         */
1094        if (num_midpoints <= num_falsetickers && low < high)
1095            break;
1096        num_falsetickers++;
1097        if (num_falsetickers * 2 >= num_candidates) {
1098            VERB3 bb_error_msg("too many falsetickers:%d (candidates:%d), no peer selected",
1099                    num_falsetickers, num_candidates);
1100            return NULL;
1101        }
1102    }
1103    VERB3 bb_error_msg("selected interval: [%f, %f]; candidates:%d falsetickers:%d",
1104            low, high, num_candidates, num_falsetickers);
1105
1106    /* Clustering */
1107
1108    /* Construct a list of survivors (p, metric)
1109     * from the chime list, where metric is dominated
1110     * first by stratum and then by root distance.
1111     * All other things being equal, this is the order of preference.
1112     */
1113    num_survivors = 0;
1114    for (i = 0; i < num_points; i++) {
1115        if (point[i].edge < low || point[i].edge > high)
1116            continue;
1117        p = point[i].p;
1118        survivor[num_survivors].p = p;
1119        /* x.opt_rd == root_distance(p); */
1120        survivor[num_survivors].metric = MAXDIST * p->lastpkt_stratum + point[i].opt_rd;
1121        VERB4 bb_error_msg("survivor[%d] metric:%f peer:%s",
1122            num_survivors, survivor[num_survivors].metric, p->p_dotted);
1123        num_survivors++;
1124    }
1125    /* There must be at least MIN_SELECTED survivors to satisfy the
1126     * correctness assertions. Ordinarily, the Byzantine criteria
1127     * require four survivors, but for the demonstration here, one
1128     * is acceptable.
1129     */
1130    if (num_survivors < MIN_SELECTED) {
1131        VERB3 bb_error_msg("num_survivors %d < %d, no peer selected",
1132                num_survivors, MIN_SELECTED);
1133        return NULL;
1134    }
1135
1136//looks like this is ONLY used by the fact that later we pick survivor[0].
1137//we can avoid sorting then, just find the minimum once!
1138    qsort(survivor, num_survivors, sizeof(survivor[0]), compare_survivor_metric);
1139
1140    /* For each association p in turn, calculate the selection
1141     * jitter p->sjitter as the square root of the sum of squares
1142     * (p->offset - q->offset) over all q associations. The idea is
1143     * to repeatedly discard the survivor with maximum selection
1144     * jitter until a termination condition is met.
1145     */
1146    while (1) {
1147        unsigned max_idx = max_idx;
1148        double max_selection_jitter = max_selection_jitter;
1149        double min_jitter = min_jitter;
1150
1151        if (num_survivors <= MIN_CLUSTERED) {
1152            VERB3 bb_error_msg("num_survivors %d <= %d, not discarding more",
1153                    num_survivors, MIN_CLUSTERED);
1154            break;
1155        }
1156
1157        /* To make sure a few survivors are left
1158         * for the clustering algorithm to chew on,
1159         * we stop if the number of survivors
1160         * is less than or equal to MIN_CLUSTERED (3).
1161         */
1162        for (i = 0; i < num_survivors; i++) {
1163            double selection_jitter_sq;
1164
1165            p = survivor[i].p;
1166            if (i == 0 || p->filter_jitter < min_jitter)
1167                min_jitter = p->filter_jitter;
1168
1169            selection_jitter_sq = 0;
1170            for (j = 0; j < num_survivors; j++) {
1171                peer_t *q = survivor[j].p;
1172                selection_jitter_sq += SQUARE(p->filter_offset - q->filter_offset);
1173            }
1174            if (i == 0 || selection_jitter_sq > max_selection_jitter) {
1175                max_selection_jitter = selection_jitter_sq;
1176                max_idx = i;
1177            }
1178            VERB5 bb_error_msg("survivor %d selection_jitter^2:%f",
1179                    i, selection_jitter_sq);
1180        }
1181        max_selection_jitter = SQRT(max_selection_jitter / num_survivors);
1182        VERB4 bb_error_msg("max_selection_jitter (at %d):%f min_jitter:%f",
1183                max_idx, max_selection_jitter, min_jitter);
1184
1185        /* If the maximum selection jitter is less than the
1186         * minimum peer jitter, then tossing out more survivors
1187         * will not lower the minimum peer jitter, so we might
1188         * as well stop.
1189         */
1190        if (max_selection_jitter < min_jitter) {
1191            VERB3 bb_error_msg("max_selection_jitter:%f < min_jitter:%f, num_survivors:%d, not discarding more",
1192                    max_selection_jitter, min_jitter, num_survivors);
1193            break;
1194        }
1195
1196        /* Delete survivor[max_idx] from the list
1197         * and go around again.
1198         */
1199        VERB5 bb_error_msg("dropping survivor %d", max_idx);
1200        num_survivors--;
1201        while (max_idx < num_survivors) {
1202            survivor[max_idx] = survivor[max_idx + 1];
1203            max_idx++;
1204        }
1205    }
1206
1207    if (0) {
1208        /* Combine the offsets of the clustering algorithm survivors
1209         * using a weighted average with weight determined by the root
1210         * distance. Compute the selection jitter as the weighted RMS
1211         * difference between the first survivor and the remaining
1212         * survivors. In some cases the inherent clock jitter can be
1213         * reduced by not using this algorithm, especially when frequent
1214         * clockhopping is involved. bbox: thus we don't do it.
1215         */
1216        double x, y, z, w;
1217        y = z = w = 0;
1218        for (i = 0; i < num_survivors; i++) {
1219            p = survivor[i].p;
1220            x = root_distance(p);
1221            y += 1 / x;
1222            z += p->filter_offset / x;
1223            w += SQUARE(p->filter_offset - survivor[0].p->filter_offset) / x;
1224        }
1225        //G.cluster_offset = z / y;
1226        //G.cluster_jitter = SQRT(w / y);
1227    }
1228
1229    /* Pick the best clock. If the old system peer is on the list
1230     * and at the same stratum as the first survivor on the list,
1231     * then don't do a clock hop. Otherwise, select the first
1232     * survivor on the list as the new system peer.
1233     */
1234    p = survivor[0].p;
1235    if (G.last_update_peer
1236     && G.last_update_peer->lastpkt_stratum <= p->lastpkt_stratum
1237    ) {
1238        /* Starting from 1 is ok here */
1239        for (i = 1; i < num_survivors; i++) {
1240            if (G.last_update_peer == survivor[i].p) {
1241                VERB4 bb_error_msg("keeping old synced peer");
1242                p = G.last_update_peer;
1243                goto keep_old;
1244            }
1245        }
1246    }
1247    G.last_update_peer = p;
1248 keep_old:
1249    VERB3 bb_error_msg("selected peer %s filter_offset:%+f age:%f",
1250            p->p_dotted,
1251            p->filter_offset,
1252            G.cur_time - p->lastpkt_recv_time
1253    );
1254    return p;
1255}
1256
1257
1258/*
1259 * Local clock discipline and its helpers
1260 */
1261static void
1262set_new_values(int disc_state, double offset, double recv_time)
1263{
1264    /* Enter new state and set state variables. Note we use the time
1265     * of the last clock filter sample, which must be earlier than
1266     * the current time.
1267     */
1268    VERB3 bb_error_msg("disc_state=%d last update offset=%f recv_time=%f",
1269            disc_state, offset, recv_time);
1270    G.discipline_state = disc_state;
1271    G.last_update_offset = offset;
1272    G.last_update_recv_time = recv_time;
1273}
1274/* Return: -1: decrease poll interval, 0: leave as is, 1: increase */
1275static NOINLINE int
1276update_local_clock(peer_t *p)
1277{
1278    int rc;
1279    struct timex tmx;
1280    /* Note: can use G.cluster_offset instead: */
1281    double offset = p->filter_offset;
1282    double recv_time = p->lastpkt_recv_time;
1283    double abs_offset;
1284#if !USING_KERNEL_PLL_LOOP
1285    double freq_drift;
1286#endif
1287    double since_last_update;
1288    double etemp, dtemp;
1289
1290    abs_offset = fabs(offset);
1291
1292#if 0
1293    /* If needed, -S script can do it by looking at $offset
1294     * env var and killing parent */
1295    /* If the offset is too large, give up and go home */
1296    if (abs_offset > PANIC_THRESHOLD) {
1297        bb_error_msg_and_die("offset %f far too big, exiting", offset);
1298    }
1299#endif
1300
1301    /* If this is an old update, for instance as the result
1302     * of a system peer change, avoid it. We never use
1303     * an old sample or the same sample twice.
1304     */
1305    if (recv_time <= G.last_update_recv_time) {
1306        VERB3 bb_error_msg("same or older datapoint: %f >= %f, not using it",
1307                G.last_update_recv_time, recv_time);
1308        return 0; /* "leave poll interval as is" */
1309    }
1310
1311    /* Clock state machine transition function. This is where the
1312     * action is and defines how the system reacts to large time
1313     * and frequency errors.
1314     */
1315    since_last_update = recv_time - G.reftime;
1316#if !USING_KERNEL_PLL_LOOP
1317    freq_drift = 0;
1318#endif
1319#if USING_INITIAL_FREQ_ESTIMATION
1320    if (G.discipline_state == STATE_FREQ) {
1321        /* Ignore updates until the stepout threshold */
1322        if (since_last_update < WATCH_THRESHOLD) {
1323            VERB3 bb_error_msg("measuring drift, datapoint ignored, %f sec remains",
1324                    WATCH_THRESHOLD - since_last_update);
1325            return 0; /* "leave poll interval as is" */
1326        }
1327# if !USING_KERNEL_PLL_LOOP
1328        freq_drift = (offset - G.last_update_offset) / since_last_update;
1329# endif
1330    }
1331#endif
1332
1333    /* There are two main regimes: when the
1334     * offset exceeds the step threshold and when it does not.
1335     */
1336    if (abs_offset > STEP_THRESHOLD) {
1337        switch (G.discipline_state) {
1338        case STATE_SYNC:
1339            /* The first outlyer: ignore it, switch to SPIK state */
1340            VERB3 bb_error_msg("offset:%+f - spike detected", offset);
1341            G.discipline_state = STATE_SPIK;
1342            return -1; /* "decrease poll interval" */
1343
1344        case STATE_SPIK:
1345            /* Ignore succeeding outlyers until either an inlyer
1346             * is found or the stepout threshold is exceeded.
1347             */
1348            if (since_last_update < WATCH_THRESHOLD) {
1349                VERB3 bb_error_msg("spike detected, datapoint ignored, %f sec remains",
1350                        WATCH_THRESHOLD - since_last_update);
1351                return -1; /* "decrease poll interval" */
1352            }
1353            /* fall through: we need to step */
1354        } /* switch */
1355
1356        /* Step the time and clamp down the poll interval.
1357         *
1358         * In NSET state an initial frequency correction is
1359         * not available, usually because the frequency file has
1360         * not yet been written. Since the time is outside the
1361         * capture range, the clock is stepped. The frequency
1362         * will be set directly following the stepout interval.
1363         *
1364         * In FSET state the initial frequency has been set
1365         * from the frequency file. Since the time is outside
1366         * the capture range, the clock is stepped immediately,
1367         * rather than after the stepout interval. Guys get
1368         * nervous if it takes 17 minutes to set the clock for
1369         * the first time.
1370         *
1371         * In SPIK state the stepout threshold has expired and
1372         * the phase is still above the step threshold. Note
1373         * that a single spike greater than the step threshold
1374         * is always suppressed, even at the longer poll
1375         * intervals.
1376         */
1377        VERB3 bb_error_msg("stepping time by %+f; poll_exp=MINPOLL", offset);
1378        step_time(offset);
1379        if (option_mask32 & OPT_q) {
1380            /* We were only asked to set time once. Done. */
1381            exit(0);
1382        }
1383
1384        G.polladj_count = 0;
1385        G.poll_exp = MINPOLL;
1386        G.stratum = MAXSTRAT;
1387
1388        run_script("step", offset);
1389
1390#if USING_INITIAL_FREQ_ESTIMATION
1391        if (G.discipline_state == STATE_NSET) {
1392            set_new_values(STATE_FREQ, /*offset:*/ 0, recv_time);
1393            return 1; /* "ok to increase poll interval" */
1394        }
1395#endif
1396        abs_offset = offset = 0;
1397        set_new_values(STATE_SYNC, offset, recv_time);
1398
1399    } else { /* abs_offset <= STEP_THRESHOLD */
1400
1401        if (G.poll_exp < MINPOLL && G.initial_poll_complete) {
1402            VERB3 bb_error_msg("small offset:%+f, disabling burst mode", offset);
1403            G.polladj_count = 0;
1404            G.poll_exp = MINPOLL;
1405        }
1406
1407        /* Compute the clock jitter as the RMS of exponentially
1408         * weighted offset differences. Used by the poll adjust code.
1409         */
1410        etemp = SQUARE(G.discipline_jitter);
1411        dtemp = SQUARE(offset - G.last_update_offset);
1412        G.discipline_jitter = SQRT(etemp + (dtemp - etemp) / AVG);
1413
1414        switch (G.discipline_state) {
1415        case STATE_NSET:
1416            if (option_mask32 & OPT_q) {
1417                /* We were only asked to set time once.
1418                 * The clock is precise enough, no need to step.
1419                 */
1420                exit(0);
1421            }
1422#if USING_INITIAL_FREQ_ESTIMATION
1423            /* This is the first update received and the frequency
1424             * has not been initialized. The first thing to do
1425             * is directly measure the oscillator frequency.
1426             */
1427            set_new_values(STATE_FREQ, offset, recv_time);
1428#else
1429            set_new_values(STATE_SYNC, offset, recv_time);
1430#endif
1431            VERB3 bb_error_msg("transitioning to FREQ, datapoint ignored");
1432            return 0; /* "leave poll interval as is" */
1433
1434#if 0 /* this is dead code for now */
1435        case STATE_FSET:
1436            /* This is the first update and the frequency
1437             * has been initialized. Adjust the phase, but
1438             * don't adjust the frequency until the next update.
1439             */
1440            set_new_values(STATE_SYNC, offset, recv_time);
1441            /* freq_drift remains 0 */
1442            break;
1443#endif
1444
1445#if USING_INITIAL_FREQ_ESTIMATION
1446        case STATE_FREQ:
1447            /* since_last_update >= WATCH_THRESHOLD, we waited enough.
1448             * Correct the phase and frequency and switch to SYNC state.
1449             * freq_drift was already estimated (see code above)
1450             */
1451            set_new_values(STATE_SYNC, offset, recv_time);
1452            break;
1453#endif
1454
1455        default:
1456#if !USING_KERNEL_PLL_LOOP
1457            /* Compute freq_drift due to PLL and FLL contributions.
1458             *
1459             * The FLL and PLL frequency gain constants
1460             * depend on the poll interval and Allan
1461             * intercept. The FLL is not used below one-half
1462             * the Allan intercept. Above that the loop gain
1463             * increases in steps to 1 / AVG.
1464             */
1465            if ((1 << G.poll_exp) > ALLAN / 2) {
1466                etemp = FLL - G.poll_exp;
1467                if (etemp < AVG)
1468                    etemp = AVG;
1469                freq_drift += (offset - G.last_update_offset) / (MAXD(since_last_update, ALLAN) * etemp);
1470            }
1471            /* For the PLL the integration interval
1472             * (numerator) is the minimum of the update
1473             * interval and poll interval. This allows
1474             * oversampling, but not undersampling.
1475             */
1476            etemp = MIND(since_last_update, (1 << G.poll_exp));
1477            dtemp = (4 * PLL) << G.poll_exp;
1478            freq_drift += offset * etemp / SQUARE(dtemp);
1479#endif
1480            set_new_values(STATE_SYNC, offset, recv_time);
1481            break;
1482        }
1483        if (G.stratum != p->lastpkt_stratum + 1) {
1484            G.stratum = p->lastpkt_stratum + 1;
1485            run_script("stratum", offset);
1486        }
1487    }
1488
1489    if (G.discipline_jitter < G_precision_sec)
1490        G.discipline_jitter = G_precision_sec;
1491    G.offset_to_jitter_ratio = abs_offset / G.discipline_jitter;
1492
1493    G.reftime = G.cur_time;
1494    G.ntp_status = p->lastpkt_status;
1495    G.refid = p->lastpkt_refid;
1496    G.rootdelay = p->lastpkt_rootdelay + p->lastpkt_delay;
1497    dtemp = p->filter_jitter; // SQRT(SQUARE(p->filter_jitter) + SQUARE(G.cluster_jitter));
1498    dtemp += MAXD(p->filter_dispersion + FREQ_TOLERANCE * (G.cur_time - p->lastpkt_recv_time) + abs_offset, MINDISP);
1499    G.rootdisp = p->lastpkt_rootdisp + dtemp;
1500    VERB3 bb_error_msg("updating leap/refid/reftime/rootdisp from peer %s", p->p_dotted);
1501
1502    /* We are in STATE_SYNC now, but did not do adjtimex yet.
1503     * (Any other state does not reach this, they all return earlier)
1504     * By this time, freq_drift and offset are set
1505     * to values suitable for adjtimex.
1506     */
1507#if !USING_KERNEL_PLL_LOOP
1508    /* Calculate the new frequency drift and frequency stability (wander).
1509     * Compute the clock wander as the RMS of exponentially weighted
1510     * frequency differences. This is not used directly, but can,
1511     * along with the jitter, be a highly useful monitoring and
1512     * debugging tool.
1513     */
1514    dtemp = G.discipline_freq_drift + freq_drift;
1515    G.discipline_freq_drift = MAXD(MIND(MAXDRIFT, dtemp), -MAXDRIFT);
1516    etemp = SQUARE(G.discipline_wander);
1517    dtemp = SQUARE(dtemp);
1518    G.discipline_wander = SQRT(etemp + (dtemp - etemp) / AVG);
1519
1520    VERB3 bb_error_msg("discipline freq_drift=%.9f(int:%ld corr:%e) wander=%f",
1521            G.discipline_freq_drift,
1522            (long)(G.discipline_freq_drift * 65536e6),
1523            freq_drift,
1524            G.discipline_wander);
1525#endif
1526    VERB3 {
1527        memset(&tmx, 0, sizeof(tmx));
1528        if (adjtimex(&tmx) < 0)
1529            bb_perror_msg_and_die("adjtimex");
1530        bb_error_msg("p adjtimex freq:%ld offset:%+ld status:0x%x tc:%ld",
1531                tmx.freq, tmx.offset, tmx.status, tmx.constant);
1532    }
1533
1534    memset(&tmx, 0, sizeof(tmx));
1535#if 0
1536//doesn't work, offset remains 0 (!) in kernel:
1537//ntpd:  set adjtimex freq:1786097 tmx.offset:77487
1538//ntpd: prev adjtimex freq:1786097 tmx.offset:0
1539//ntpd:  cur adjtimex freq:1786097 tmx.offset:0
1540    tmx.modes = ADJ_FREQUENCY | ADJ_OFFSET;
1541    /* 65536 is one ppm */
1542    tmx.freq = G.discipline_freq_drift * 65536e6;
1543#endif
1544    tmx.modes = ADJ_OFFSET | ADJ_STATUS | ADJ_TIMECONST;// | ADJ_MAXERROR | ADJ_ESTERROR;
1545    tmx.offset = (offset * 1000000); /* usec */
1546    tmx.status = STA_PLL;
1547    if (G.ntp_status & LI_PLUSSEC)
1548        tmx.status |= STA_INS;
1549    if (G.ntp_status & LI_MINUSSEC)
1550        tmx.status |= STA_DEL;
1551
1552    tmx.constant = G.poll_exp - 4;
1553    /* EXPERIMENTAL.
1554     * The below if statement should be unnecessary, but...
1555     * It looks like Linux kernel's PLL is far too gentle in changing
1556     * tmx.freq in response to clock offset. Offset keeps growing
1557     * and eventually we fall back to smaller poll intervals.
1558     * We can make correction more agressive (about x2) by supplying
1559     * PLL time constant which is one less than the real one.
1560     * To be on a safe side, let's do it only if offset is significantly
1561     * larger than jitter.
1562     */
1563    if (tmx.constant > 0 && G.offset_to_jitter_ratio >= TIMECONST_HACK_GATE)
1564        tmx.constant--;
1565
1566    //tmx.esterror = (uint32_t)(clock_jitter * 1e6);
1567    //tmx.maxerror = (uint32_t)((sys_rootdelay / 2 + sys_rootdisp) * 1e6);
1568    rc = adjtimex(&tmx);
1569    if (rc < 0)
1570        bb_perror_msg_and_die("adjtimex");
1571    /* NB: here kernel returns constant == G.poll_exp, not == G.poll_exp - 4.
1572     * Not sure why. Perhaps it is normal.
1573     */
1574    VERB3 bb_error_msg("adjtimex:%d freq:%ld offset:%+ld status:0x%x",
1575                rc, tmx.freq, tmx.offset, tmx.status);
1576    G.kernel_freq_drift = tmx.freq / 65536;
1577    VERB2 bb_error_msg("update from:%s offset:%+f jitter:%f clock drift:%+.3fppm tc:%d",
1578            p->p_dotted, offset, G.discipline_jitter, (double)tmx.freq / 65536, (int)tmx.constant);
1579
1580    return 1; /* "ok to increase poll interval" */
1581}
1582
1583
1584/*
1585 * We've got a new reply packet from a peer, process it
1586 * (helpers first)
1587 */
1588static unsigned
1589retry_interval(void)
1590{
1591    /* Local problem, want to retry soon */
1592    unsigned interval, r;
1593    interval = RETRY_INTERVAL;
1594    r = random();
1595    interval += r % (unsigned)(RETRY_INTERVAL / 4);
1596    VERB3 bb_error_msg("chose retry interval:%u", interval);
1597    return interval;
1598}
1599static unsigned
1600poll_interval(int exponent)
1601{
1602    unsigned interval, r;
1603    exponent = G.poll_exp + exponent;
1604    if (exponent < 0)
1605        exponent = 0;
1606    interval = 1 << exponent;
1607    r = random();
1608    interval += ((r & (interval-1)) >> 4) + ((r >> 8) & 1); /* + 1/16 of interval, max */
1609    VERB3 bb_error_msg("chose poll interval:%u (poll_exp:%d exp:%d)", interval, G.poll_exp, exponent);
1610    return interval;
1611}
1612static NOINLINE void
1613recv_and_process_peer_pkt(peer_t *p)
1614{
1615    int         rc;
1616    ssize_t     size;
1617    msg_t       msg;
1618    double      T1, T2, T3, T4;
1619    unsigned    interval;
1620    datapoint_t *datapoint;
1621    peer_t      *q;
1622
1623    /* We can recvfrom here and check from.IP, but some multihomed
1624     * ntp servers reply from their *other IP*.
1625     * TODO: maybe we should check at least what we can: from.port == 123?
1626     */
1627    size = recv(p->p_fd, &msg, sizeof(msg), MSG_DONTWAIT);
1628    if (size == -1) {
1629        bb_perror_msg("recv(%s) error", p->p_dotted);
1630        if (errno == EHOSTUNREACH || errno == EHOSTDOWN
1631         || errno == ENETUNREACH || errno == ENETDOWN
1632         || errno == ECONNREFUSED || errno == EADDRNOTAVAIL
1633         || errno == EAGAIN
1634        ) {
1635//TODO: always do this?
1636            interval = retry_interval();
1637            goto set_next_and_ret;
1638        }
1639        xfunc_die();
1640    }
1641
1642    if (size != NTP_MSGSIZE_NOAUTH && size != NTP_MSGSIZE) {
1643        bb_error_msg("malformed packet received from %s", p->p_dotted);
1644        return;
1645    }
1646
1647    if (msg.m_orgtime.int_partl != p->p_xmt_msg.m_xmttime.int_partl
1648     || msg.m_orgtime.fractionl != p->p_xmt_msg.m_xmttime.fractionl
1649    ) {
1650        /* Somebody else's packet */
1651        return;
1652    }
1653
1654    /* We do not expect any more packets from this peer for now.
1655     * Closing the socket informs kernel about it.
1656     * We open a new socket when we send a new query.
1657     */
1658    close(p->p_fd);
1659    p->p_fd = -1;
1660
1661    if ((msg.m_status & LI_ALARM) == LI_ALARM
1662     || msg.m_stratum == 0
1663     || msg.m_stratum > NTP_MAXSTRATUM
1664    ) {
1665// TODO: stratum 0 responses may have commands in 32-bit m_refid field:
1666// "DENY", "RSTR" - peer does not like us at all
1667// "RATE" - peer is overloaded, reduce polling freq
1668        interval = poll_interval(0);
1669        bb_error_msg("reply from %s: peer is unsynced, next query in %us", p->p_dotted, interval);
1670        goto set_next_and_ret;
1671    }
1672
1673//  /* Verify valid root distance */
1674//  if (msg.m_rootdelay / 2 + msg.m_rootdisp >= MAXDISP || p->lastpkt_reftime > msg.m_xmt)
1675//      return;                 /* invalid header values */
1676
1677    p->lastpkt_status = msg.m_status;
1678    p->lastpkt_stratum = msg.m_stratum;
1679    p->lastpkt_rootdelay = sfp_to_d(msg.m_rootdelay);
1680    p->lastpkt_rootdisp = sfp_to_d(msg.m_rootdisp);
1681    p->lastpkt_refid = msg.m_refid;
1682
1683    /*
1684     * From RFC 2030 (with a correction to the delay math):
1685     *
1686     * Timestamp Name          ID   When Generated
1687     * ------------------------------------------------------------
1688     * Originate Timestamp     T1   time request sent by client
1689     * Receive Timestamp       T2   time request received by server
1690     * Transmit Timestamp      T3   time reply sent by server
1691     * Destination Timestamp   T4   time reply received by client
1692     *
1693     * The roundtrip delay and local clock offset are defined as
1694     *
1695     * delay = (T4 - T1) - (T3 - T2); offset = ((T2 - T1) + (T3 - T4)) / 2
1696     */
1697    T1 = p->p_xmttime;
1698    T2 = lfp_to_d(msg.m_rectime);
1699    T3 = lfp_to_d(msg.m_xmttime);
1700    T4 = G.cur_time;
1701
1702    p->lastpkt_recv_time = T4;
1703
1704    VERB5 bb_error_msg("%s->lastpkt_recv_time=%f", p->p_dotted, p->lastpkt_recv_time);
1705    p->datapoint_idx = p->reachable_bits ? (p->datapoint_idx + 1) % NUM_DATAPOINTS : 0;
1706    datapoint = &p->filter_datapoint[p->datapoint_idx];
1707    datapoint->d_recv_time = T4;
1708    datapoint->d_offset    = ((T2 - T1) + (T3 - T4)) / 2;
1709    /* The delay calculation is a special case. In cases where the
1710     * server and client clocks are running at different rates and
1711     * with very fast networks, the delay can appear negative. In
1712     * order to avoid violating the Principle of Least Astonishment,
1713     * the delay is clamped not less than the system precision.
1714     */
1715    p->lastpkt_delay = (T4 - T1) - (T3 - T2);
1716    if (p->lastpkt_delay < G_precision_sec)
1717        p->lastpkt_delay = G_precision_sec;
1718    datapoint->d_dispersion = LOG2D(msg.m_precision_exp) + G_precision_sec;
1719    if (!p->reachable_bits) {
1720        /* 1st datapoint ever - replicate offset in every element */
1721        int i;
1722        for (i = 0; i < NUM_DATAPOINTS; i++) {
1723            p->filter_datapoint[i].d_offset = datapoint->d_offset;
1724        }
1725    }
1726
1727    p->reachable_bits |= 1;
1728    if ((MAX_VERBOSE && G.verbose) || (option_mask32 & OPT_w)) {
1729        bb_error_msg("reply from %s: offset:%+f delay:%f status:0x%02x strat:%d refid:0x%08x rootdelay:%f reach:0x%02x",
1730            p->p_dotted,
1731            datapoint->d_offset,
1732            p->lastpkt_delay,
1733            p->lastpkt_status,
1734            p->lastpkt_stratum,
1735            p->lastpkt_refid,
1736            p->lastpkt_rootdelay,
1737            p->reachable_bits
1738            /* not shown: m_ppoll, m_precision_exp, m_rootdisp,
1739             * m_reftime, m_orgtime, m_rectime, m_xmttime
1740             */
1741        );
1742    }
1743
1744    /* Muck with statictics and update the clock */
1745    filter_datapoints(p);
1746    q = select_and_cluster();
1747    rc = -1;
1748    if (q) {
1749        rc = 0;
1750        if (!(option_mask32 & OPT_w)) {
1751            rc = update_local_clock(q);
1752            /* If drift is dangerously large, immediately
1753             * drop poll interval one step down.
1754             */
1755            if (fabs(q->filter_offset) >= POLLDOWN_OFFSET) {
1756                VERB3 bb_error_msg("offset:%+f > POLLDOWN_OFFSET", q->filter_offset);
1757                goto poll_down;
1758            }
1759        }
1760    }
1761    /* else: no peer selected, rc = -1: we want to poll more often */
1762
1763    if (rc != 0) {
1764        /* Adjust the poll interval by comparing the current offset
1765         * with the clock jitter. If the offset is less than
1766         * the clock jitter times a constant, then the averaging interval
1767         * is increased, otherwise it is decreased. A bit of hysteresis
1768         * helps calm the dance. Works best using burst mode.
1769         */
1770        if (rc > 0 && G.offset_to_jitter_ratio <= POLLADJ_GATE) {
1771            /* was += G.poll_exp but it is a bit
1772             * too optimistic for my taste at high poll_exp's */
1773            G.polladj_count += MINPOLL;
1774            if (G.polladj_count > POLLADJ_LIMIT) {
1775                G.polladj_count = 0;
1776                if (G.poll_exp < MAXPOLL) {
1777                    G.poll_exp++;
1778                    VERB3 bb_error_msg("polladj: discipline_jitter:%f ++poll_exp=%d",
1779                            G.discipline_jitter, G.poll_exp);
1780                }
1781            } else {
1782                VERB3 bb_error_msg("polladj: incr:%d", G.polladj_count);
1783            }
1784        } else {
1785            G.polladj_count -= G.poll_exp * 2;
1786            if (G.polladj_count < -POLLADJ_LIMIT || G.poll_exp >= BIGPOLL) {
1787 poll_down:
1788                G.polladj_count = 0;
1789                if (G.poll_exp > MINPOLL) {
1790                    llist_t *item;
1791
1792                    G.poll_exp--;
1793                    /* Correct p->next_action_time in each peer
1794                     * which waits for sending, so that they send earlier.
1795                     * Old pp->next_action_time are on the order
1796                     * of t + (1 << old_poll_exp) + small_random,
1797                     * we simply need to subtract ~half of that.
1798                     */
1799                    for (item = G.ntp_peers; item != NULL; item = item->link) {
1800                        peer_t *pp = (peer_t *) item->data;
1801                        if (pp->p_fd < 0)
1802                            pp->next_action_time -= (1 << G.poll_exp);
1803                    }
1804                    VERB3 bb_error_msg("polladj: discipline_jitter:%f --poll_exp=%d",
1805                            G.discipline_jitter, G.poll_exp);
1806                }
1807            } else {
1808                VERB3 bb_error_msg("polladj: decr:%d", G.polladj_count);
1809            }
1810        }
1811    }
1812
1813    /* Decide when to send new query for this peer */
1814    interval = poll_interval(0);
1815
1816 set_next_and_ret:
1817    set_next(p, interval);
1818}
1819
1820#if ENABLE_FEATURE_NTPD_SERVER
1821static NOINLINE void
1822recv_and_process_client_pkt(void /*int fd*/)
1823{
1824    ssize_t          size;
1825    //uint8_t          version;
1826    len_and_sockaddr *to;
1827    struct sockaddr  *from;
1828    msg_t            msg;
1829    uint8_t          query_status;
1830    l_fixedpt_t      query_xmttime;
1831
1832    to = get_sock_lsa(G_listen_fd);
1833    from = xzalloc(to->len);
1834
1835    size = recv_from_to(G_listen_fd, &msg, sizeof(msg), MSG_DONTWAIT, from, &to->u.sa, to->len);
1836    if (size != NTP_MSGSIZE_NOAUTH && size != NTP_MSGSIZE) {
1837        char *addr;
1838        if (size < 0) {
1839            if (errno == EAGAIN)
1840                goto bail;
1841            bb_perror_msg_and_die("recv");
1842        }
1843        addr = xmalloc_sockaddr2dotted_noport(from);
1844        bb_error_msg("malformed packet received from %s: size %u", addr, (int)size);
1845        free(addr);
1846        goto bail;
1847    }
1848
1849    query_status = msg.m_status;
1850    query_xmttime = msg.m_xmttime;
1851
1852    /* Build a reply packet */
1853    memset(&msg, 0, sizeof(msg));
1854    msg.m_status = G.stratum < MAXSTRAT ? (G.ntp_status & LI_MASK) : LI_ALARM;
1855    msg.m_status |= (query_status & VERSION_MASK);
1856    msg.m_status |= ((query_status & MODE_MASK) == MODE_CLIENT) ?
1857            MODE_SERVER : MODE_SYM_PAS;
1858    msg.m_stratum = G.stratum;
1859    msg.m_ppoll = G.poll_exp;
1860    msg.m_precision_exp = G_precision_exp;
1861    /* this time was obtained between poll() and recv() */
1862    msg.m_rectime = d_to_lfp(G.cur_time);
1863    msg.m_xmttime = d_to_lfp(gettime1900d()); /* this instant */
1864    if (G.peer_cnt == 0) {
1865        /* we have no peers: "stratum 1 server" mode. reftime = our own time */
1866        G.reftime = G.cur_time;
1867    }
1868    msg.m_reftime = d_to_lfp(G.reftime);
1869    msg.m_orgtime = query_xmttime;
1870    msg.m_rootdelay = d_to_sfp(G.rootdelay);
1871//simple code does not do this, fix simple code!
1872    msg.m_rootdisp = d_to_sfp(G.rootdisp);
1873    //version = (query_status & VERSION_MASK); /* ... >> VERSION_SHIFT - done below instead */
1874    msg.m_refid = G.refid; // (version > (3 << VERSION_SHIFT)) ? G.refid : G.refid3;
1875
1876    /* We reply from the local address packet was sent to,
1877     * this makes to/from look swapped here: */
1878    do_sendto(G_listen_fd,
1879        /*from:*/ &to->u.sa, /*to:*/ from, /*addrlen:*/ to->len,
1880        &msg, size);
1881
1882 bail:
1883    free(to);
1884    free(from);
1885}
1886#endif
1887
1888/* Upstream ntpd's options:
1889 *
1890 * -4   Force DNS resolution of host names to the IPv4 namespace.
1891 * -6   Force DNS resolution of host names to the IPv6 namespace.
1892 * -a   Require cryptographic authentication for broadcast client,
1893 *      multicast client and symmetric passive associations.
1894 *      This is the default.
1895 * -A   Do not require cryptographic authentication for broadcast client,
1896 *      multicast client and symmetric passive associations.
1897 *      This is almost never a good idea.
1898 * -b   Enable the client to synchronize to broadcast servers.
1899 * -c conffile
1900 *      Specify the name and path of the configuration file,
1901 *      default /etc/ntp.conf
1902 * -d   Specify debugging mode. This option may occur more than once,
1903 *      with each occurrence indicating greater detail of display.
1904 * -D level
1905 *      Specify debugging level directly.
1906 * -f driftfile
1907 *      Specify the name and path of the frequency file.
1908 *      This is the same operation as the "driftfile FILE"
1909 *      configuration command.
1910 * -g   Normally, ntpd exits with a message to the system log
1911 *      if the offset exceeds the panic threshold, which is 1000 s
1912 *      by default. This option allows the time to be set to any value
1913 *      without restriction; however, this can happen only once.
1914 *      If the threshold is exceeded after that, ntpd will exit
1915 *      with a message to the system log. This option can be used
1916 *      with the -q and -x options. See the tinker command for other options.
1917 * -i jaildir
1918 *      Chroot the server to the directory jaildir. This option also implies
1919 *      that the server attempts to drop root privileges at startup
1920 *      (otherwise, chroot gives very little additional security).
1921 *      You may need to also specify a -u option.
1922 * -k keyfile
1923 *      Specify the name and path of the symmetric key file,
1924 *      default /etc/ntp/keys. This is the same operation
1925 *      as the "keys FILE" configuration command.
1926 * -l logfile
1927 *      Specify the name and path of the log file. The default
1928 *      is the system log file. This is the same operation as
1929 *      the "logfile FILE" configuration command.
1930 * -L   Do not listen to virtual IPs. The default is to listen.
1931 * -n   Don't fork.
1932 * -N   To the extent permitted by the operating system,
1933 *      run the ntpd at the highest priority.
1934 * -p pidfile
1935 *      Specify the name and path of the file used to record the ntpd
1936 *      process ID. This is the same operation as the "pidfile FILE"
1937 *      configuration command.
1938 * -P priority
1939 *      To the extent permitted by the operating system,
1940 *      run the ntpd at the specified priority.
1941 * -q   Exit the ntpd just after the first time the clock is set.
1942 *      This behavior mimics that of the ntpdate program, which is
1943 *      to be retired. The -g and -x options can be used with this option.
1944 *      Note: The kernel time discipline is disabled with this option.
1945 * -r broadcastdelay
1946 *      Specify the default propagation delay from the broadcast/multicast
1947 *      server to this client. This is necessary only if the delay
1948 *      cannot be computed automatically by the protocol.
1949 * -s statsdir
1950 *      Specify the directory path for files created by the statistics
1951 *      facility. This is the same operation as the "statsdir DIR"
1952 *      configuration command.
1953 * -t key
1954 *      Add a key number to the trusted key list. This option can occur
1955 *      more than once.
1956 * -u user[:group]
1957 *      Specify a user, and optionally a group, to switch to.
1958 * -v variable
1959 * -V variable
1960 *      Add a system variable listed by default.
1961 * -x   Normally, the time is slewed if the offset is less than the step
1962 *      threshold, which is 128 ms by default, and stepped if above
1963 *      the threshold. This option sets the threshold to 600 s, which is
1964 *      well within the accuracy window to set the clock manually.
1965 *      Note: since the slew rate of typical Unix kernels is limited
1966 *      to 0.5 ms/s, each second of adjustment requires an amortization
1967 *      interval of 2000 s. Thus, an adjustment as much as 600 s
1968 *      will take almost 14 days to complete. This option can be used
1969 *      with the -g and -q options. See the tinker command for other options.
1970 *      Note: The kernel time discipline is disabled with this option.
1971 */
1972
1973/* By doing init in a separate function we decrease stack usage
1974 * in main loop.
1975 */
1976static NOINLINE void ntp_init(char **argv)
1977{
1978    unsigned opts;
1979    llist_t *peers;
1980
1981    srandom(getpid());
1982
1983    if (getuid())
1984        bb_error_msg_and_die(bb_msg_you_must_be_root);
1985
1986    /* Set some globals */
1987    G.stratum = MAXSTRAT;
1988    if (BURSTPOLL != 0)
1989        G.poll_exp = BURSTPOLL; /* speeds up initial sync */
1990    G.last_script_run = G.reftime = G.last_update_recv_time = gettime1900d(); /* sets G.cur_time too */
1991
1992    /* Parse options */
1993    peers = NULL;
1994    opt_complementary = "dd:p::wn"; /* d: counter; p: list; -w implies -n */
1995    opts = getopt32(argv,
1996            "nqNx" /* compat */
1997            "wp:S:"IF_FEATURE_NTPD_SERVER("l") /* NOT compat */
1998            "d" /* compat */
1999            "46aAbgL", /* compat, ignored */
2000            &peers, &G.script_name, &G.verbose);
2001    if (!(opts & (OPT_p|OPT_l)))
2002        bb_show_usage();
2003//  if (opts & OPT_x) /* disable stepping, only slew is allowed */
2004//      G.time_was_stepped = 1;
2005    if (peers) {
2006        while (peers)
2007            add_peers(llist_pop(&peers));
2008    } else {
2009        /* -l but no peers: "stratum 1 server" mode */
2010        G.stratum = 1;
2011    }
2012    if (!(opts & OPT_n)) {
2013        bb_daemonize_or_rexec(DAEMON_DEVNULL_STDIO, argv);
2014        logmode = LOGMODE_NONE;
2015    }
2016#if ENABLE_FEATURE_NTPD_SERVER
2017    G_listen_fd = -1;
2018    if (opts & OPT_l) {
2019        G_listen_fd = create_and_bind_dgram_or_die(NULL, 123);
2020        socket_want_pktinfo(G_listen_fd);
2021        setsockopt(G_listen_fd, IPPROTO_IP, IP_TOS, &const_IPTOS_LOWDELAY, sizeof(const_IPTOS_LOWDELAY));
2022    }
2023#endif
2024    /* I hesitate to set -20 prio. -15 should be high enough for timekeeping */
2025    if (opts & OPT_N)
2026        setpriority(PRIO_PROCESS, 0, -15);
2027
2028    /* If network is up, syncronization occurs in ~10 seconds.
2029     * We give "ntpd -q" 10 seconds to get first reply,
2030     * then another 50 seconds to finish syncing.
2031     *
2032     * I tested ntpd 4.2.6p1 and apparently it never exits
2033     * (will try forever), but it does not feel right.
2034     * The goal of -q is to act like ntpdate: set time
2035     * after a reasonably small period of polling, or fail.
2036     */
2037    if (opts & OPT_q) {
2038        option_mask32 |= OPT_qq;
2039        alarm(10);
2040    }
2041
2042    bb_signals(0
2043        | (1 << SIGTERM)
2044        | (1 << SIGINT)
2045        | (1 << SIGALRM)
2046        , record_signo
2047    );
2048    bb_signals(0
2049        | (1 << SIGPIPE)
2050        | (1 << SIGCHLD)
2051        , SIG_IGN
2052    );
2053}
2054
2055int ntpd_main(int argc UNUSED_PARAM, char **argv) MAIN_EXTERNALLY_VISIBLE;
2056int ntpd_main(int argc UNUSED_PARAM, char **argv)
2057{
2058#undef G
2059    struct globals G;
2060    struct pollfd *pfd;
2061    peer_t **idx2peer;
2062    unsigned cnt;
2063
2064    memset(&G, 0, sizeof(G));
2065    SET_PTR_TO_GLOBALS(&G);
2066
2067    ntp_init(argv);
2068
2069    /* If ENABLE_FEATURE_NTPD_SERVER, + 1 for listen_fd: */
2070    cnt = G.peer_cnt + ENABLE_FEATURE_NTPD_SERVER;
2071    idx2peer = xzalloc(sizeof(idx2peer[0]) * cnt);
2072    pfd = xzalloc(sizeof(pfd[0]) * cnt);
2073
2074    /* Countdown: we never sync before we sent INITIAL_SAMPLES+1
2075     * packets to each peer.
2076     * NB: if some peer is not responding, we may end up sending
2077     * fewer packets to it and more to other peers.
2078     * NB2: sync usually happens using INITIAL_SAMPLES packets,
2079     * since last reply does not come back instantaneously.
2080     */
2081    cnt = G.peer_cnt * (INITIAL_SAMPLES + 1);
2082
2083    write_pidfile(CONFIG_PID_FILE_PATH "/ntpd.pid");
2084
2085    while (!bb_got_signal) {
2086        llist_t *item;
2087        unsigned i, j;
2088        int nfds, timeout;
2089        double nextaction;
2090
2091        /* Nothing between here and poll() blocks for any significant time */
2092
2093        nextaction = G.cur_time + 3600;
2094
2095        i = 0;
2096#if ENABLE_FEATURE_NTPD_SERVER
2097        if (G_listen_fd != -1) {
2098            pfd[0].fd = G_listen_fd;
2099            pfd[0].events = POLLIN;
2100            i++;
2101        }
2102#endif
2103        /* Pass over peer list, send requests, time out on receives */
2104        for (item = G.ntp_peers; item != NULL; item = item->link) {
2105            peer_t *p = (peer_t *) item->data;
2106
2107            if (p->next_action_time <= G.cur_time) {
2108                if (p->p_fd == -1) {
2109                    /* Time to send new req */
2110                    if (--cnt == 0) {
2111                        G.initial_poll_complete = 1;
2112                    }
2113                    send_query_to_peer(p);
2114                } else {
2115                    /* Timed out waiting for reply */
2116                    close(p->p_fd);
2117                    p->p_fd = -1;
2118                    timeout = poll_interval(-2); /* -2: try a bit sooner */
2119                    bb_error_msg("timed out waiting for %s, reach 0x%02x, next query in %us",
2120                            p->p_dotted, p->reachable_bits, timeout);
2121                    set_next(p, timeout);
2122                }
2123            }
2124
2125            if (p->next_action_time < nextaction)
2126                nextaction = p->next_action_time;
2127
2128            if (p->p_fd >= 0) {
2129                /* Wait for reply from this peer */
2130                pfd[i].fd = p->p_fd;
2131                pfd[i].events = POLLIN;
2132                idx2peer[i] = p;
2133                i++;
2134            }
2135        }
2136
2137        timeout = nextaction - G.cur_time;
2138        if (timeout < 0)
2139            timeout = 0;
2140        timeout++; /* (nextaction - G.cur_time) rounds down, compensating */
2141
2142        /* Here we may block */
2143        VERB2 {
2144            if (i > (ENABLE_FEATURE_NTPD_SERVER && G_listen_fd != -1)) {
2145                /* We wait for at least one reply.
2146                 * Poll for it, without wasting time for message.
2147                 * Since replies often come under 1 second, this also
2148                 * reduces clutter in logs.
2149                 */
2150                nfds = poll(pfd, i, 1000);
2151                if (nfds != 0)
2152                    goto did_poll;
2153                if (--timeout <= 0)
2154                    goto did_poll;
2155            }
2156            bb_error_msg("poll:%us sockets:%u interval:%us", timeout, i, 1 << G.poll_exp);
2157        }
2158        nfds = poll(pfd, i, timeout * 1000);
2159 did_poll:
2160        gettime1900d(); /* sets G.cur_time */
2161        if (nfds <= 0) {
2162            if (G.script_name && G.cur_time - G.last_script_run > 11*60) {
2163                /* Useful for updating battery-backed RTC and such */
2164                run_script("periodic", G.last_update_offset);
2165                gettime1900d(); /* sets G.cur_time */
2166            }
2167            continue;
2168        }
2169
2170        /* Process any received packets */
2171        j = 0;
2172#if ENABLE_FEATURE_NTPD_SERVER
2173        if (G.listen_fd != -1) {
2174            if (pfd[0].revents /* & (POLLIN|POLLERR)*/) {
2175                nfds--;
2176                recv_and_process_client_pkt(/*G.listen_fd*/);
2177                gettime1900d(); /* sets G.cur_time */
2178            }
2179            j = 1;
2180        }
2181#endif
2182        for (; nfds != 0 && j < i; j++) {
2183            if (pfd[j].revents /* & (POLLIN|POLLERR)*/) {
2184                /*
2185                 * At init, alarm was set to 10 sec.
2186                 * Now we did get a reply.
2187                 * Increase timeout to 50 seconds to finish syncing.
2188                 */
2189                if (option_mask32 & OPT_qq) {
2190                    option_mask32 &= ~OPT_qq;
2191                    alarm(50);
2192                }
2193                nfds--;
2194                recv_and_process_peer_pkt(idx2peer[j]);
2195                gettime1900d(); /* sets G.cur_time */
2196            }
2197        }
2198    } /* while (!bb_got_signal) */
2199
2200    remove_pidfile(CONFIG_PID_FILE_PATH "/ntpd.pid");
2201    kill_myself_with_sig(bb_got_signal);
2202}
2203
2204
2205
2206
2207
2208
2209/*** openntpd-4.6 uses only adjtime, not adjtimex ***/
2210
2211/*** ntp-4.2.6/ntpd/ntp_loopfilter.c - adjtimex usage ***/
2212
2213#if 0
2214static double
2215direct_freq(double fp_offset)
2216{
2217#ifdef KERNEL_PLL
2218    /*
2219     * If the kernel is enabled, we need the residual offset to
2220     * calculate the frequency correction.
2221     */
2222    if (pll_control && kern_enable) {
2223        memset(&ntv, 0, sizeof(ntv));
2224        ntp_adjtime(&ntv);
2225#ifdef STA_NANO
2226        clock_offset = ntv.offset / 1e9;
2227#else /* STA_NANO */
2228        clock_offset = ntv.offset / 1e6;
2229#endif /* STA_NANO */
2230        drift_comp = FREQTOD(ntv.freq);
2231    }
2232#endif /* KERNEL_PLL */
2233    set_freq((fp_offset - clock_offset) / (current_time - clock_epoch) + drift_comp);
2234    wander_resid = 0;
2235    return drift_comp;
2236}
2237
2238static void
2239set_freq(double freq) /* frequency update */
2240{
2241    char tbuf[80];
2242
2243    drift_comp = freq;
2244
2245#ifdef KERNEL_PLL
2246    /*
2247     * If the kernel is enabled, update the kernel frequency.
2248     */
2249    if (pll_control && kern_enable) {
2250        memset(&ntv, 0, sizeof(ntv));
2251        ntv.modes = MOD_FREQUENCY;
2252        ntv.freq = DTOFREQ(drift_comp);
2253        ntp_adjtime(&ntv);
2254        snprintf(tbuf, sizeof(tbuf), "kernel %.3f PPM", drift_comp * 1e6);
2255        report_event(EVNT_FSET, NULL, tbuf);
2256    } else {
2257        snprintf(tbuf, sizeof(tbuf), "ntpd %.3f PPM", drift_comp * 1e6);
2258        report_event(EVNT_FSET, NULL, tbuf);
2259    }
2260#else /* KERNEL_PLL */
2261    snprintf(tbuf, sizeof(tbuf), "ntpd %.3f PPM", drift_comp * 1e6);
2262    report_event(EVNT_FSET, NULL, tbuf);
2263#endif /* KERNEL_PLL */
2264}
2265
2266...
2267...
2268...
2269
2270#ifdef KERNEL_PLL
2271    /*
2272     * This code segment works when clock adjustments are made using
2273     * precision time kernel support and the ntp_adjtime() system
2274     * call. This support is available in Solaris 2.6 and later,
2275     * Digital Unix 4.0 and later, FreeBSD, Linux and specially
2276     * modified kernels for HP-UX 9 and Ultrix 4. In the case of the
2277     * DECstation 5000/240 and Alpha AXP, additional kernel
2278     * modifications provide a true microsecond clock and nanosecond
2279     * clock, respectively.
2280     *
2281     * Important note: The kernel discipline is used only if the
2282     * step threshold is less than 0.5 s, as anything higher can
2283     * lead to overflow problems. This might occur if some misguided
2284     * lad set the step threshold to something ridiculous.
2285     */
2286    if (pll_control && kern_enable) {
2287
2288#define MOD_BITS (MOD_OFFSET | MOD_MAXERROR | MOD_ESTERROR | MOD_STATUS | MOD_TIMECONST)
2289
2290        /*
2291         * We initialize the structure for the ntp_adjtime()
2292         * system call. We have to convert everything to
2293         * microseconds or nanoseconds first. Do not update the
2294         * system variables if the ext_enable flag is set. In
2295         * this case, the external clock driver will update the
2296         * variables, which will be read later by the local
2297         * clock driver. Afterwards, remember the time and
2298         * frequency offsets for jitter and stability values and
2299         * to update the frequency file.
2300         */
2301        memset(&ntv,  0, sizeof(ntv));
2302        if (ext_enable) {
2303            ntv.modes = MOD_STATUS;
2304        } else {
2305#ifdef STA_NANO
2306            ntv.modes = MOD_BITS | MOD_NANO;
2307#else /* STA_NANO */
2308            ntv.modes = MOD_BITS;
2309#endif /* STA_NANO */
2310            if (clock_offset < 0)
2311                dtemp = -.5;
2312            else
2313                dtemp = .5;
2314#ifdef STA_NANO
2315            ntv.offset = (int32)(clock_offset * 1e9 + dtemp);
2316            ntv.constant = sys_poll;
2317#else /* STA_NANO */
2318            ntv.offset = (int32)(clock_offset * 1e6 + dtemp);
2319            ntv.constant = sys_poll - 4;
2320#endif /* STA_NANO */
2321            ntv.esterror = (u_int32)(clock_jitter * 1e6);
2322            ntv.maxerror = (u_int32)((sys_rootdelay / 2 + sys_rootdisp) * 1e6);
2323            ntv.status = STA_PLL;
2324
2325            /*
2326             * Enable/disable the PPS if requested.
2327             */
2328            if (pps_enable) {
2329                if (!(pll_status & STA_PPSTIME))
2330                    report_event(EVNT_KERN,
2331                        NULL, "PPS enabled");
2332                ntv.status |= STA_PPSTIME | STA_PPSFREQ;
2333            } else {
2334                if (pll_status & STA_PPSTIME)
2335                    report_event(EVNT_KERN,
2336                        NULL, "PPS disabled");
2337                ntv.status &= ~(STA_PPSTIME | STA_PPSFREQ);
2338            }
2339            if (sys_leap == LEAP_ADDSECOND)
2340                ntv.status |= STA_INS;
2341            else if (sys_leap == LEAP_DELSECOND)
2342                ntv.status |= STA_DEL;
2343        }
2344
2345        /*
2346         * Pass the stuff to the kernel. If it squeals, turn off
2347         * the pps. In any case, fetch the kernel offset,
2348         * frequency and jitter.
2349         */
2350        if (ntp_adjtime(&ntv) == TIME_ERROR) {
2351            if (!(ntv.status & STA_PPSSIGNAL))
2352                report_event(EVNT_KERN, NULL,
2353                        "PPS no signal");
2354        }
2355        pll_status = ntv.status;
2356#ifdef STA_NANO
2357        clock_offset = ntv.offset / 1e9;
2358#else /* STA_NANO */
2359        clock_offset = ntv.offset / 1e6;
2360#endif /* STA_NANO */
2361        clock_frequency = FREQTOD(ntv.freq);
2362
2363        /*
2364         * If the kernel PPS is lit, monitor its performance.
2365         */
2366        if (ntv.status & STA_PPSTIME) {
2367#ifdef STA_NANO
2368            clock_jitter = ntv.jitter / 1e9;
2369#else /* STA_NANO */
2370            clock_jitter = ntv.jitter / 1e6;
2371#endif /* STA_NANO */
2372        }
2373
2374#if defined(STA_NANO) && NTP_API == 4
2375        /*
2376         * If the TAI changes, update the kernel TAI.
2377         */
2378        if (loop_tai != sys_tai) {
2379            loop_tai = sys_tai;
2380            ntv.modes = MOD_TAI;
2381            ntv.constant = sys_tai;
2382            ntp_adjtime(&ntv);
2383        }
2384#endif /* STA_NANO */
2385    }
2386#endif /* KERNEL_PLL */
2387#endif
Note: See TracBrowser for help on using the repository browser.