*
* Author: Adam Tkac <vonsch@gmail.com>
*
- * Licensed under GPLv2, see file LICENSE in this tarball for details.
+ * Licensed under GPLv2, see file LICENSE in this source tree.
*
* Parts of OpenNTPD clock syncronization code is replaced by
* code which is based on ntp-4.2.6, whuch carries the following
* *
***********************************************************************
*/
+
+//usage:#define ntpd_trivial_usage
+//usage: "[-dnqNw"IF_FEATURE_NTPD_SERVER("l")"] [-S PROG] [-p PEER]..."
+//usage:#define ntpd_full_usage "\n\n"
+//usage: "NTP client/server\n"
+//usage: "\n -d Verbose"
+//usage: "\n -n Do not daemonize"
+//usage: "\n -q Quit after clock is set"
+//usage: "\n -N Run at high priority"
+//usage: "\n -w Do not set time (only query peers), implies -n"
+//usage: IF_FEATURE_NTPD_SERVER(
+//usage: "\n -l Run as server on port 123"
+//usage: )
+//usage: "\n -S PROG Run PROG after stepping time, stratum change, and every 11 mins"
+//usage: "\n -p PEER Obtain time from PEER (may be repeated)"
+
#include "libbb.h"
#include <math.h>
#include <netinet/ip.h> /* For IPTOS_LOWDELAY definition */
+#include <sys/resource.h> /* setpriority */
#include <sys/timex.h>
#ifndef IPTOS_LOWDELAY
# define IPTOS_LOWDELAY 0x10
/* Verbosity control (max level of -dddd options accepted).
- * max 5 is very talkative (and bloated). 2 is non-bloated,
+ * max 6 is very talkative (and bloated). 3 is non-bloated,
* production level setting.
*/
-#define MAX_VERBOSE 2
+#define MAX_VERBOSE 3
+
+/* High-level description of the algorithm:
+ *
+ * We start running with very small poll_exp, BURSTPOLL,
+ * in order to quickly accumulate INITIAL_SAMPLES datapoints
+ * for each peer. Then, time is stepped if the offset is larger
+ * than STEP_THRESHOLD, otherwise it isn't; anyway, we enlarge
+ * poll_exp to MINPOLL and enter frequency measurement step:
+ * we collect new datapoints but ignore them for WATCH_THRESHOLD
+ * seconds. After WATCH_THRESHOLD seconds we look at accumulated
+ * offset and estimate frequency drift.
+ *
+ * (frequency measurement step seems to not be strictly needed,
+ * it is conditionally disabled with USING_INITIAL_FREQ_ESTIMATION
+ * define set to 0)
+ *
+ * After this, we enter "steady state": we collect a datapoint,
+ * we select the best peer, if this datapoint is not a new one
+ * (IOW: if this datapoint isn't for selected peer), sleep
+ * and collect another one; otherwise, use its offset to update
+ * frequency drift, if offset is somewhat large, reduce poll_exp,
+ * otherwise increase poll_exp.
+ *
+ * If offset is larger than STEP_THRESHOLD, which shouldn't normally
+ * happen, we assume that something "bad" happened (computer
+ * was hibernated, someone set totally wrong date, etc),
+ * then the time is stepped, all datapoints are discarded,
+ * and we go back to steady state.
+ *
+ * Made some changes to speed up re-syncing after our clock goes bad
+ * (tested with suspending my laptop):
+ * - if largish offset (>= STEP_THRESHOLD * 8 == 1 sec) is seen
+ * from a peer, schedule next query for this peer soon
+ * without drastically lowering poll interval for everybody.
+ * This makes us collect enough data for step much faster:
+ * e.g. at poll = 10 (1024 secs), step was done within 5 minutes
+ * after first reply which indicated that our clock is 14 seconds off.
+ * - on step, do not discard d_dispersion data of the existing datapoints,
+ * do not clear reachable_bits. This prevents discarding first ~8
+ * datapoints after the step.
+ */
-#define RETRY_INTERVAL 5 /* on error, retry in N secs */
+#define RETRY_INTERVAL 5 /* on error, retry in N secs */
#define RESPONSE_INTERVAL 15 /* wait for reply up to N secs */
+#define INITIAL_SAMPLES 4 /* how many samples do we want for init */
+#define BAD_DELAY_GROWTH 4 /* drop packet if its delay grew by more than this */
+
+/* Clock discipline parameters and constants */
+
+/* Step threshold (sec). std ntpd uses 0.128.
+ * Using exact power of 2 (1/8) results in smaller code */
+#define STEP_THRESHOLD 0.125
+#define WATCH_THRESHOLD 128 /* stepout threshold (sec). std ntpd uses 900 (11 mins (!)) */
+/* NB: set WATCH_THRESHOLD to ~60 when debugging to save time) */
+//UNUSED: #define PANIC_THRESHOLD 1000 /* panic threshold (sec) */
-#define FREQ_TOLERANCE 0.000015 /* % frequency tolerance (15 PPM) */
-#define BURSTPOLL 0
-#define MINPOLL 4 /* % minimum poll interval (6: 64 s) */
-#define MAXPOLL 12 /* % maximum poll interval (12: 1.1h, 17: 36.4h) (was 17) */
-#define MINDISP 0.01 /* % minimum dispersion (s) */
-#define MAXDISP 16 /* maximum dispersion (s) */
+#define FREQ_TOLERANCE 0.000015 /* frequency tolerance (15 PPM) */
+#define BURSTPOLL 0 /* initial poll */
+#define MINPOLL 5 /* minimum poll interval. std ntpd uses 6 (6: 64 sec) */
+/* If we got largish offset from a peer, cap next query interval
+ * for this peer by this many seconds:
+ */
+#define BIGOFF_INTERVAL (1 << 6)
+/* If offset > discipline_jitter * POLLADJ_GATE, and poll interval is >= 2^BIGPOLL,
+ * then it is decreased _at once_. (If < 2^BIGPOLL, it will be decreased _eventually_).
+ */
+#define BIGPOLL 10 /* 2^10 sec ~= 17 min */
+#define MAXPOLL 12 /* maximum poll interval (12: 1.1h, 17: 36.4h). std ntpd uses 17 */
+/* Actively lower poll when we see such big offsets.
+ * With STEP_THRESHOLD = 0.125, it means we try to sync more aggressively
+ * if offset increases over ~0.04 sec */
+#define POLLDOWN_OFFSET (STEP_THRESHOLD / 3)
+#define MINDISP 0.01 /* minimum dispersion (sec) */
+#define MAXDISP 16 /* maximum dispersion (sec) */
#define MAXSTRAT 16 /* maximum stratum (infinity metric) */
-#define MAXDIST 1 /* % distance threshold (s) */
-#define MIN_SELECTED 1 /* % minimum intersection survivors */
-#define MIN_CLUSTERED 3 /* % minimum cluster survivors */
+#define MAXDIST 1 /* distance threshold (sec) */
+#define MIN_SELECTED 1 /* minimum intersection survivors */
+#define MIN_CLUSTERED 3 /* minimum cluster survivors */
#define MAXDRIFT 0.000500 /* frequency drift we can correct (500 PPM) */
-/* Clock discipline parameters and constants */
-#define STEP_THRESHOLD 0.128 /* step threshold (s) */
-#define WATCH_THRESHOLD 150 /* stepout threshold (s). std ntpd uses 900 (11 mins (!)) */
-/* NB: set WATCH_THRESHOLD to ~60 when debugging to save time) */
-#define PANIC_THRESHOLD 1000 /* panic threshold (s) */
-
/* Poll-adjust threshold.
* When we see that offset is small enough compared to discipline jitter,
- * we grow a counter: += MINPOLL. When it goes over POLLADJ_LIMIT,
+ * we grow a counter: += MINPOLL. When counter goes over POLLADJ_LIMIT,
* we poll_exp++. If offset isn't small, counter -= poll_exp*2,
- * and when it goes below -POLLADJ_LIMIT, we poll_exp--
+ * and when it goes below -POLLADJ_LIMIT, we poll_exp--.
+ * (Bumped from 30 to 40 since otherwise I often see poll_exp going *2* steps down)
*/
-#define POLLADJ_LIMIT 30
-/* If offset < POLLADJ_GATE * discipline_jitter, then we can increase
+#define POLLADJ_LIMIT 40
+/* If offset < discipline_jitter * POLLADJ_GATE, then we decide to increase
* poll interval (we think we can't improve timekeeping
* by staying at smaller poll).
*/
#define POLLADJ_GATE 4
-/* Compromise Allan intercept (s). doc uses 1500, std ntpd uses 512 */
+#define TIMECONST_HACK_GATE 2
+/* Compromise Allan intercept (sec). doc uses 1500, std ntpd uses 512 */
#define ALLAN 512
/* PLL loop gain */
#define PLL 65536
} msg_t;
typedef struct {
- double d_recv_time;
double d_offset;
+ double d_recv_time;
double d_dispersion;
} datapoint_t;
typedef struct {
len_and_sockaddr *p_lsa;
char *p_dotted;
- /* when to send new query (if p_fd == -1)
- * or when receive times out (if p_fd >= 0): */
int p_fd;
int datapoint_idx;
uint32_t lastpkt_refid;
uint8_t lastpkt_status;
uint8_t lastpkt_stratum;
uint8_t reachable_bits;
+ /* when to send new query (if p_fd == -1)
+ * or when receive times out (if p_fd >= 0): */
double next_action_time;
double p_xmttime;
double lastpkt_recv_time;
} peer_t;
+#define USING_KERNEL_PLL_LOOP 1
+#define USING_INITIAL_FREQ_ESTIMATION 0
+
enum {
OPT_n = (1 << 0),
OPT_q = (1 << 1),
/* Non-compat options: */
OPT_w = (1 << 4),
OPT_p = (1 << 5),
- OPT_l = (1 << 6) * ENABLE_FEATURE_NTPD_SERVER,
+ OPT_S = (1 << 6),
+ OPT_l = (1 << 7) * ENABLE_FEATURE_NTPD_SERVER,
+ /* We hijack some bits for other purposes */
+ OPT_qq = (1 << 31),
};
struct globals {
double reftime;
/* total dispersion to currently selected reference clock */
double rootdisp;
+
+ double last_script_run;
+ char *script_name;
llist_t *ntp_peers;
#if ENABLE_FEATURE_NTPD_SERVER
int listen_fd;
+# define G_listen_fd (G.listen_fd)
+#else
+# define G_listen_fd (-1)
#endif
unsigned verbose;
unsigned peer_cnt;
/* refid: 32-bit code identifying the particular server or reference clock
- * in stratum 0 packets this is a four-character ASCII string,
- * called the kiss code, used for debugging and monitoring
- * in stratum 1 packets this is a four-character ASCII string
- * assigned to the reference clock by IANA. Example: "GPS "
- * in stratum 2+ packets, it's IPv4 address or 4 first bytes of MD5 hash of IPv6
+ * in stratum 0 packets this is a four-character ASCII string,
+ * called the kiss code, used for debugging and monitoring
+ * in stratum 1 packets this is a four-character ASCII string
+ * assigned to the reference clock by IANA. Example: "GPS "
+ * in stratum 2+ packets, it's IPv4 address or 4 first bytes
+ * of MD5 hash of IPv6
*/
uint32_t refid;
uint8_t ntp_status;
* mains-frequency clock incrementing at 60 Hz is 16 ms, even when the
* system clock hardware representation is to the nanosecond.
*
- * Delays, jitters of various kinds are clamper down to precision.
+ * Delays, jitters of various kinds are clamped down to precision.
*
* If precision_sec is too large, discipline_jitter gets clamped to it
- * and if offset is much smaller than discipline_jitter, poll interval
- * grows even though we really can benefit from staying at smaller one,
- * collecting non-lagged datapoits and correcting the offset.
+ * and if offset is smaller than discipline_jitter * POLLADJ_GATE, poll
+ * interval grows even though we really can benefit from staying at
+ * smaller one, collecting non-lagged datapoits and correcting offset.
* (Lagged datapoits exist when poll_exp is large but we still have
* systematic offset error - the time distance between datapoints
- * is significat and older datapoints have smaller offsets.
+ * is significant and older datapoints have smaller offsets.
* This makes our offset estimation a bit smaller than reality)
* Due to this effect, setting G_precision_sec close to
* STEP_THRESHOLD isn't such a good idea - offsets may grow
* too big and we will step. I observed it with -6.
*
- * OTOH, setting precision too small would result in futile attempts
- * to syncronize to the unachievable precision.
+ * OTOH, setting precision_sec far too small would result in futile
+ * attempts to syncronize to an unachievable precision.
*
* -6 is 1/64 sec, -7 is 1/128 sec and so on.
+ * -8 is 1/256 ~= 0.003906 (worked well for me --vda)
+ * -9 is 1/512 ~= 0.001953 (let's try this for some time)
+ */
+#define G_precision_exp -9
+ /*
+ * G_precision_exp is used only for construction outgoing packets.
+ * It's ok to set G_precision_sec to a slightly different value
+ * (One which is "nicer looking" in logs).
+ * Exact value would be (1.0 / (1 << (- G_precision_exp))):
*/
-#define G_precision_exp -8
-#define G_precision_sec (1.0 / (1 << (- G_precision_exp)))
+#define G_precision_sec 0.002
uint8_t stratum;
/* Bool. After set to 1, never goes back to 0: */
- smallint adjtimex_was_done;
smallint initial_poll_complete;
+#define STATE_NSET 0 /* initial state, "nothing is set" */
+//#define STATE_FSET 1 /* frequency set from file */
+//#define STATE_SPIK 2 /* spike detected */
+//#define STATE_FREQ 3 /* initial frequency */
+#define STATE_SYNC 4 /* clock synchronized (normal operation) */
uint8_t discipline_state; // doc calls it c.state
uint8_t poll_exp; // s.poll
int polladj_count; // c.count
long kernel_freq_drift;
+ peer_t *last_update_peer;
double last_update_offset; // c.last
double last_update_recv_time; // s.t
double discipline_jitter; // c.jitter
-//TODO: add s.jitter - grep for it here and see clock_combine() in doc
-#define USING_KERNEL_PLL_LOOP 1
+ /* Since we only compare it with ints, can simplify code
+ * by not making this variable floating point:
+ */
+ unsigned offset_to_jitter_ratio;
+ //double cluster_offset; // s.offset
+ //double cluster_jitter; // s.jitter
#if !USING_KERNEL_PLL_LOOP
double discipline_freq_drift; // c.freq
-//TODO: conditionally calculate wander? it's used only for logging
+ /* Maybe conditionally calculate wander? it's used only for logging */
double discipline_wander; // c.wander
#endif
};
#define VERB3 if (MAX_VERBOSE >= 3 && G.verbose >= 3)
#define VERB4 if (MAX_VERBOSE >= 4 && G.verbose >= 4)
#define VERB5 if (MAX_VERBOSE >= 5 && G.verbose >= 5)
+#define VERB6 if (MAX_VERBOSE >= 6 && G.verbose >= 6)
static double LOG2D(int a)
filter_datapoints(peer_t *p)
{
int i, idx;
+ double sum, wavg;
+ datapoint_t *fdp;
+
+#if 0
+/* Simulations have shown that use of *averaged* offset for p->filter_offset
+ * is in fact worse than simply using last received one: with large poll intervals
+ * (>= 2048) averaging code uses offset values which are outdated by hours,
+ * and time/frequency correction goes totally wrong when fed essentially bogus offsets.
+ */
int got_newest;
- double minoff, maxoff, wavg, sum, w;
+ double minoff, maxoff, w;
double x = x; /* for compiler */
double oldest_off = oldest_off;
double oldest_age = oldest_age;
double newest_off = newest_off;
double newest_age = newest_age;
- minoff = maxoff = p->filter_datapoint[0].d_offset;
+ fdp = p->filter_datapoint;
+
+ minoff = maxoff = fdp[0].d_offset;
for (i = 1; i < NUM_DATAPOINTS; i++) {
- if (minoff > p->filter_datapoint[i].d_offset)
- minoff = p->filter_datapoint[i].d_offset;
- if (maxoff < p->filter_datapoint[i].d_offset)
- maxoff = p->filter_datapoint[i].d_offset;
+ if (minoff > fdp[i].d_offset)
+ minoff = fdp[i].d_offset;
+ if (maxoff < fdp[i].d_offset)
+ maxoff = fdp[i].d_offset;
}
- idx = p->datapoint_idx; /* most recent datapoint */
+ idx = p->datapoint_idx; /* most recent datapoint's index */
/* Average offset:
* Drop two outliers and take weighted average of the rest:
* most_recent/2 + older1/4 + older2/8 ... + older5/32 + older6/32
got_newest = 0;
sum = 0;
for (i = 0; i < NUM_DATAPOINTS; i++) {
- VERB4 {
+ VERB5 {
bb_error_msg("datapoint[%d]: off:%f disp:%f(%f) age:%f%s",
i,
- p->filter_datapoint[idx].d_offset,
- p->filter_datapoint[idx].d_dispersion, dispersion(&p->filter_datapoint[idx]),
- G.cur_time - p->filter_datapoint[idx].d_recv_time,
- (minoff == p->filter_datapoint[idx].d_offset || maxoff == p->filter_datapoint[idx].d_offset)
+ fdp[idx].d_offset,
+ fdp[idx].d_dispersion, dispersion(&fdp[idx]),
+ G.cur_time - fdp[idx].d_recv_time,
+ (minoff == fdp[idx].d_offset || maxoff == fdp[idx].d_offset)
? " (outlier by offset)" : ""
);
}
- sum += dispersion(&p->filter_datapoint[idx]) / (2 << i);
+ sum += dispersion(&fdp[idx]) / (2 << i);
- if (minoff == p->filter_datapoint[idx].d_offset) {
+ if (minoff == fdp[idx].d_offset) {
minoff -= 1; /* so that we don't match it ever again */
} else
- if (maxoff == p->filter_datapoint[idx].d_offset) {
+ if (maxoff == fdp[idx].d_offset) {
maxoff += 1;
} else {
- oldest_off = p->filter_datapoint[idx].d_offset;
- oldest_age = G.cur_time - p->filter_datapoint[idx].d_recv_time;
+ oldest_off = fdp[idx].d_offset;
+ oldest_age = G.cur_time - fdp[idx].d_recv_time;
if (!got_newest) {
got_newest = 1;
newest_off = oldest_off;
}
p->filter_offset = wavg;
+#else
+
+ fdp = p->filter_datapoint;
+ idx = p->datapoint_idx; /* most recent datapoint's index */
+
+ /* filter_offset: simply use the most recent value */
+ p->filter_offset = fdp[idx].d_offset;
+
+ /* n-1
+ * --- dispersion(i)
+ * filter_dispersion = \ -------------
+ * / (i+1)
+ * --- 2
+ * i=0
+ */
+ wavg = 0;
+ sum = 0;
+ for (i = 0; i < NUM_DATAPOINTS; i++) {
+ sum += dispersion(&fdp[idx]) / (2 << i);
+ wavg += fdp[idx].d_offset;
+ idx = (idx - 1) & (NUM_DATAPOINTS - 1);
+ }
+ wavg /= NUM_DATAPOINTS;
+ p->filter_dispersion = sum;
+#endif
+
/* +----- -----+ ^ 1/2
* | n-1 |
* | --- |
*/
sum = 0;
for (i = 0; i < NUM_DATAPOINTS; i++) {
- sum += SQUARE(wavg - p->filter_datapoint[i].d_offset);
+ sum += SQUARE(wavg - fdp[i].d_offset);
}
sum = SQRT(sum / NUM_DATAPOINTS);
p->filter_jitter = sum > G_precision_sec ? sum : G_precision_sec;
- VERB3 bb_error_msg("filter offset:%f(corr:%e) disp:%f jitter:%f",
- p->filter_offset, x,
+ VERB4 bb_error_msg("filter offset:%+f disp:%f jitter:%f",
+ p->filter_offset,
p->filter_dispersion,
p->filter_jitter);
-
}
static void
reset_peer_stats(peer_t *p, double offset)
{
int i;
+ bool small_ofs = fabs(offset) < 16 * STEP_THRESHOLD;
+
+ /* Used to set p->filter_datapoint[i].d_dispersion = MAXDISP
+ * and clear reachable bits, but this proved to be too agressive:
+ * after step (tested with suspinding laptop for ~30 secs),
+ * this caused all previous data to be considered invalid,
+ * making us needing to collect full ~8 datapoins per peer
+ * after step in order to start trusting them.
+ * In turn, this was making poll interval decrease even after
+ * step was done. (Poll interval decreases already before step
+ * in this scenario, because we see large offsets and end up with
+ * no good peer to select).
+ */
+
for (i = 0; i < NUM_DATAPOINTS; i++) {
- if (offset < 16 * STEP_THRESHOLD) {
- p->filter_datapoint[i].d_recv_time -= offset;
+ if (small_ofs) {
+ p->filter_datapoint[i].d_recv_time += offset;
if (p->filter_datapoint[i].d_offset != 0) {
p->filter_datapoint[i].d_offset -= offset;
+ //bb_error_msg("p->filter_datapoint[%d].d_offset %f -> %f",
+ // i,
+ // p->filter_datapoint[i].d_offset + offset,
+ // p->filter_datapoint[i].d_offset);
}
} else {
p->filter_datapoint[i].d_recv_time = G.cur_time;
p->filter_datapoint[i].d_offset = 0;
- p->filter_datapoint[i].d_dispersion = MAXDISP;
+ /*p->filter_datapoint[i].d_dispersion = MAXDISP;*/
}
}
- if (offset < 16 * STEP_THRESHOLD) {
- p->lastpkt_recv_time -= offset;
+ if (small_ofs) {
+ p->lastpkt_recv_time += offset;
} else {
- p->reachable_bits = 0;
+ /*p->reachable_bits = 0;*/
p->lastpkt_recv_time = G.cur_time;
}
filter_datapoints(p); /* recalc p->filter_xxx */
- p->next_action_time -= offset;
- VERB5 bb_error_msg("%s->lastpkt_recv_time=%f", p->p_dotted, p->lastpkt_recv_time);
+ VERB6 bb_error_msg("%s->lastpkt_recv_time=%f", p->p_dotted, p->lastpkt_recv_time);
}
static void
free(local_lsa);
}
+ /* Emit message _before_ attempted send. Think of a very short
+ * roundtrip networks: we need to go back to recv loop ASAP,
+ * to reduce delay. Printing messages after send works against that.
+ */
+ VERB1 bb_error_msg("sending query to %s", p->p_dotted);
+
/*
* Send out a random 64-bit number as our transmit time. The NTP
* server will copy said number into the originate field on the
p->p_xmt_msg.m_xmttime.fractionl = random();
p->p_xmttime = gettime1900d();
+ /* Were doing it only if sendto worked, but
+ * loss of sync detection needs reachable_bits updated
+ * even if sending fails *locally*:
+ * "network is unreachable" because cable was pulled?
+ * We still need to declare "unsync" if this condition persists.
+ */
+ p->reachable_bits <<= 1;
+
if (do_sendto(p->p_fd, /*from:*/ NULL, /*to:*/ &p->p_lsa->u.sa, /*addrlen:*/ p->p_lsa->len,
&p->p_xmt_msg, NTP_MSGSIZE_NOAUTH) == -1
) {
close(p->p_fd);
p->p_fd = -1;
+ /*
+ * We know that we sent nothing.
+ * We can retry *soon* without fearing
+ * that we are flooding the peer.
+ */
set_next(p, RETRY_INTERVAL);
return;
}
- p->reachable_bits <<= 1;
- VERB1 bb_error_msg("sent query to %s", p->p_dotted);
set_next(p, RESPONSE_INTERVAL);
}
+/* Note that there is no provision to prevent several run_scripts
+ * to be started in quick succession. In fact, it happens rather often
+ * if initial syncronization results in a step.
+ * You will see "step" and then "stratum" script runs, sometimes
+ * as close as only 0.002 seconds apart.
+ * Script should be ready to deal with this.
+ */
+static void run_script(const char *action, double offset)
+{
+ char *argv[3];
+ char *env1, *env2, *env3, *env4;
+
+ G.last_script_run = G.cur_time;
+
+ if (!G.script_name)
+ return;
+
+ argv[0] = (char*) G.script_name;
+ argv[1] = (char*) action;
+ argv[2] = NULL;
+
+ VERB1 bb_error_msg("executing '%s %s'", G.script_name, action);
+
+ env1 = xasprintf("%s=%u", "stratum", G.stratum);
+ putenv(env1);
+ env2 = xasprintf("%s=%ld", "freq_drift_ppm", G.kernel_freq_drift);
+ putenv(env2);
+ env3 = xasprintf("%s=%u", "poll_interval", 1 << G.poll_exp);
+ putenv(env3);
+ env4 = xasprintf("%s=%f", "offset", offset);
+ putenv(env4);
+ /* Other items of potential interest: selected peer,
+ * rootdelay, reftime, rootdisp, refid, ntp_status,
+ * last_update_offset, last_update_recv_time, discipline_jitter,
+ * how many peers have reachable_bits = 0?
+ */
+
+ /* Don't want to wait: it may run hwclock --systohc, and that
+ * may take some time (seconds): */
+ /*spawn_and_wait(argv);*/
+ spawn(argv);
+
+ unsetenv("stratum");
+ unsetenv("freq_drift_ppm");
+ unsetenv("poll_interval");
+ unsetenv("offset");
+ free(env1);
+ free(env2);
+ free(env3);
+ free(env4);
+}
+
static NOINLINE void
step_time(double offset)
{
llist_t *item;
double dtime;
- struct timeval tv;
- char buf[80];
+ struct timeval tvc, tvn;
+ char buf[sizeof("yyyy-mm-dd hh:mm:ss") + /*paranoia:*/ 4];
time_t tval;
- gettimeofday(&tv, NULL); /* never fails */
- dtime = offset + tv.tv_sec;
- dtime += 1.0e-6 * tv.tv_usec;
- d_to_tv(dtime, &tv);
-
- if (settimeofday(&tv, NULL) == -1)
+ gettimeofday(&tvc, NULL); /* never fails */
+ dtime = tvc.tv_sec + (1.0e-6 * tvc.tv_usec) + offset;
+ d_to_tv(dtime, &tvn);
+ if (settimeofday(&tvn, NULL) == -1)
bb_perror_msg_and_die("settimeofday");
- tval = tv.tv_sec;
- strftime(buf, sizeof(buf), "%a %b %e %H:%M:%S %Z %Y", localtime(&tval));
-
- bb_error_msg("setting clock to %s (offset %fs)", buf, offset);
+ VERB2 {
+ tval = tvc.tv_sec;
+ strftime_YYYYMMDDHHMMSS(buf, sizeof(buf), &tval);
+ bb_error_msg("current time is %s.%06u", buf, (unsigned)tvc.tv_usec);
+ }
+ tval = tvn.tv_sec;
+ strftime_YYYYMMDDHHMMSS(buf, sizeof(buf), &tval);
+ bb_error_msg("setting time to %s.%06u (offset %+fs)", buf, (unsigned)tvn.tv_usec, offset);
/* Correct various fields which contain time-relative values: */
+ /* Globals: */
+ G.cur_time += offset;
+ G.last_update_recv_time += offset;
+ G.last_script_run += offset;
+
/* p->lastpkt_recv_time, p->next_action_time and such: */
for (item = G.ntp_peers; item != NULL; item = item->link) {
peer_t *pp = (peer_t *) item->data;
reset_peer_stats(pp, offset);
+ //bb_error_msg("offset:%+f pp->next_action_time:%f -> %f",
+ // offset, pp->next_action_time, pp->next_action_time + offset);
+ pp->next_action_time += offset;
+ if (pp->p_fd >= 0) {
+ /* We wait for reply from this peer too.
+ * But due to step we are doing, reply's data is no longer
+ * useful (in fact, it'll be bogus). Stop waiting for it.
+ */
+ close(pp->p_fd);
+ pp->p_fd = -1;
+ set_next(pp, RETRY_INTERVAL);
+ }
}
- /* Globals: */
- G.cur_time -= offset;
- G.last_update_recv_time -= offset;
}
peer_t *p;
int type;
double edge;
+ double opt_rd; /* optimization */
} point_t;
static int
compare_point_edge(const void *aa, const void *bb)
{
if ((p->reachable_bits & (p->reachable_bits-1)) == 0) {
/* One or zero bits in reachable_bits */
- VERB3 bb_error_msg("peer %s unfit for selection: unreachable", p->p_dotted);
+ VERB4 bb_error_msg("peer %s unfit for selection: unreachable", p->p_dotted);
return 0;
}
-#if 0 /* we filter out such packets earlier */
+#if 0 /* we filter out such packets earlier */
if ((p->lastpkt_status & LI_ALARM) == LI_ALARM
|| p->lastpkt_stratum >= MAXSTRAT
) {
- VERB3 bb_error_msg("peer %s unfit for selection: bad status/stratum", p->p_dotted);
+ VERB4 bb_error_msg("peer %s unfit for selection: bad status/stratum", p->p_dotted);
return 0;
}
#endif
/* rd is root_distance(p) */
if (rd > MAXDIST + FREQ_TOLERANCE * (1 << G.poll_exp)) {
- VERB3 bb_error_msg("peer %s unfit for selection: root distance too high", p->p_dotted);
+ VERB4 bb_error_msg("peer %s unfit for selection: root distance too high", p->p_dotted);
return 0;
}
//TODO
// /* Do we have a loop? */
// if (p->refid == p->dstaddr || p->refid == s.refid)
// return 0;
- return 1;
+ return 1;
}
static peer_t*
select_and_cluster(void)
{
+ peer_t *p;
llist_t *item;
int i, j;
int size = 3 * G.peer_cnt;
num_points = 0;
item = G.ntp_peers;
if (G.initial_poll_complete) while (item != NULL) {
- peer_t *p = (peer_t *) item->data;
- double rd = root_distance(p);
- double offset = p->filter_offset;
+ double rd, offset;
+ p = (peer_t *) item->data;
+ rd = root_distance(p);
+ offset = p->filter_offset;
if (!fit(p, rd)) {
item = item->link;
continue;
}
- VERB4 bb_error_msg("interval: [%f %f %f] %s",
+ VERB5 bb_error_msg("interval: [%f %f %f] %s",
offset - rd,
offset,
offset + rd,
point[num_points].p = p;
point[num_points].type = -1;
point[num_points].edge = offset - rd;
+ point[num_points].opt_rd = rd;
num_points++;
point[num_points].p = p;
point[num_points].type = 0;
point[num_points].edge = offset;
+ point[num_points].opt_rd = rd;
num_points++;
point[num_points].p = p;
point[num_points].type = 1;
point[num_points].edge = offset + rd;
+ point[num_points].opt_rd = rd;
num_points++;
item = item->link;
}
num_candidates = num_points / 3;
if (num_candidates == 0) {
- VERB3 bb_error_msg("no valid datapoints, no peer selected");
+ VERB3 bb_error_msg("no valid datapoints%s", ", no peer selected");
return NULL;
}
//TODO: sorting does not seem to be done in reference code
break;
num_falsetickers++;
if (num_falsetickers * 2 >= num_candidates) {
- VERB3 bb_error_msg("too many falsetickers:%d (candidates:%d), no peer selected",
- num_falsetickers, num_candidates);
+ VERB3 bb_error_msg("falsetickers:%d, candidates:%d%s",
+ num_falsetickers, num_candidates,
+ ", no peer selected");
return NULL;
}
}
- VERB3 bb_error_msg("selected interval: [%f, %f]; candidates:%d falsetickers:%d",
+ VERB4 bb_error_msg("selected interval: [%f, %f]; candidates:%d falsetickers:%d",
low, high, num_candidates, num_falsetickers);
/* Clustering */
*/
num_survivors = 0;
for (i = 0; i < num_points; i++) {
- peer_t *p;
-
if (point[i].edge < low || point[i].edge > high)
continue;
p = point[i].p;
survivor[num_survivors].p = p;
-//TODO: save root_distance in point_t and reuse here?
- survivor[num_survivors].metric = MAXDIST * p->lastpkt_stratum + root_distance(p);
- VERB4 bb_error_msg("survivor[%d] metric:%f peer:%s",
+ /* x.opt_rd == root_distance(p); */
+ survivor[num_survivors].metric = MAXDIST * p->lastpkt_stratum + point[i].opt_rd;
+ VERB5 bb_error_msg("survivor[%d] metric:%f peer:%s",
num_survivors, survivor[num_survivors].metric, p->p_dotted);
num_survivors++;
}
* is acceptable.
*/
if (num_survivors < MIN_SELECTED) {
- VERB3 bb_error_msg("num_survivors %d < %d, no peer selected",
- num_survivors, MIN_SELECTED);
+ VERB3 bb_error_msg("survivors:%d%s",
+ num_survivors,
+ ", no peer selected");
return NULL;
}
double min_jitter = min_jitter;
if (num_survivors <= MIN_CLUSTERED) {
- VERB3 bb_error_msg("num_survivors %d <= %d, not discarding more",
+ VERB4 bb_error_msg("num_survivors %d <= %d, not discarding more",
num_survivors, MIN_CLUSTERED);
break;
}
*/
for (i = 0; i < num_survivors; i++) {
double selection_jitter_sq;
- peer_t *p = survivor[i].p;
+ p = survivor[i].p;
if (i == 0 || p->filter_jitter < min_jitter)
min_jitter = p->filter_jitter;
max_selection_jitter = selection_jitter_sq;
max_idx = i;
}
- VERB5 bb_error_msg("survivor %d selection_jitter^2:%f",
+ VERB6 bb_error_msg("survivor %d selection_jitter^2:%f",
i, selection_jitter_sq);
}
max_selection_jitter = SQRT(max_selection_jitter / num_survivors);
- VERB4 bb_error_msg("max_selection_jitter (at %d):%f min_jitter:%f",
+ VERB5 bb_error_msg("max_selection_jitter (at %d):%f min_jitter:%f",
max_idx, max_selection_jitter, min_jitter);
/* If the maximum selection jitter is less than the
* as well stop.
*/
if (max_selection_jitter < min_jitter) {
- VERB3 bb_error_msg("max_selection_jitter:%f < min_jitter:%f, num_survivors:%d, not discarding more",
+ VERB4 bb_error_msg("max_selection_jitter:%f < min_jitter:%f, num_survivors:%d, not discarding more",
max_selection_jitter, min_jitter, num_survivors);
break;
}
/* Delete survivor[max_idx] from the list
* and go around again.
*/
- VERB5 bb_error_msg("dropping survivor %d", max_idx);
+ VERB6 bb_error_msg("dropping survivor %d", max_idx);
num_survivors--;
while (max_idx < num_survivors) {
survivor[max_idx] = survivor[max_idx + 1];
}
}
+ if (0) {
+ /* Combine the offsets of the clustering algorithm survivors
+ * using a weighted average with weight determined by the root
+ * distance. Compute the selection jitter as the weighted RMS
+ * difference between the first survivor and the remaining
+ * survivors. In some cases the inherent clock jitter can be
+ * reduced by not using this algorithm, especially when frequent
+ * clockhopping is involved. bbox: thus we don't do it.
+ */
+ double x, y, z, w;
+ y = z = w = 0;
+ for (i = 0; i < num_survivors; i++) {
+ p = survivor[i].p;
+ x = root_distance(p);
+ y += 1 / x;
+ z += p->filter_offset / x;
+ w += SQUARE(p->filter_offset - survivor[0].p->filter_offset) / x;
+ }
+ //G.cluster_offset = z / y;
+ //G.cluster_jitter = SQRT(w / y);
+ }
+
/* Pick the best clock. If the old system peer is on the list
* and at the same stratum as the first survivor on the list,
* then don't do a clock hop. Otherwise, select the first
* survivor on the list as the new system peer.
*/
-//TODO - see clock_combine()
- VERB3 bb_error_msg("selected peer %s filter_offset:%f age:%f",
- survivor[0].p->p_dotted,
- survivor[0].p->filter_offset,
- G.cur_time - survivor[0].p->lastpkt_recv_time
+ p = survivor[0].p;
+ if (G.last_update_peer
+ && G.last_update_peer->lastpkt_stratum <= p->lastpkt_stratum
+ ) {
+ /* Starting from 1 is ok here */
+ for (i = 1; i < num_survivors; i++) {
+ if (G.last_update_peer == survivor[i].p) {
+ VERB5 bb_error_msg("keeping old synced peer");
+ p = G.last_update_peer;
+ goto keep_old;
+ }
+ }
+ }
+ G.last_update_peer = p;
+ keep_old:
+ VERB4 bb_error_msg("selected peer %s filter_offset:%+f age:%f",
+ p->p_dotted,
+ p->filter_offset,
+ G.cur_time - p->lastpkt_recv_time
);
- return survivor[0].p;
+ return p;
}
* of the last clock filter sample, which must be earlier than
* the current time.
*/
- VERB3 bb_error_msg("disc_state=%d last update offset=%f recv_time=%f",
+ VERB4 bb_error_msg("disc_state=%d last update offset=%f recv_time=%f",
disc_state, offset, recv_time);
G.discipline_state = disc_state;
G.last_update_offset = offset;
G.last_update_recv_time = recv_time;
}
-/* Clock state definitions */
-#define STATE_NSET 0 /* initial state, "nothing is set" */
-#define STATE_FSET 1 /* frequency set from file */
-#define STATE_SPIK 2 /* spike detected */
-#define STATE_FREQ 3 /* initial frequency */
-#define STATE_SYNC 4 /* clock synchronized (normal operation) */
/* Return: -1: decrease poll interval, 0: leave as is, 1: increase */
static NOINLINE int
update_local_clock(peer_t *p)
{
int rc;
- long old_tmx_offset;
struct timex tmx;
+ /* Note: can use G.cluster_offset instead: */
double offset = p->filter_offset;
double recv_time = p->lastpkt_recv_time;
double abs_offset;
abs_offset = fabs(offset);
+#if 0
+ /* If needed, -S script can do it by looking at $offset
+ * env var and killing parent */
/* If the offset is too large, give up and go home */
if (abs_offset > PANIC_THRESHOLD) {
bb_error_msg_and_die("offset %f far too big, exiting", offset);
}
+#endif
/* If this is an old update, for instance as the result
* of a system peer change, avoid it. We never use
* an old sample or the same sample twice.
*/
if (recv_time <= G.last_update_recv_time) {
- VERB3 bb_error_msg("same or older datapoint: %f >= %f, not using it",
- G.last_update_recv_time, recv_time);
+ VERB3 bb_error_msg("update from %s: same or older datapoint, not using it",
+ p->p_dotted);
return 0; /* "leave poll interval as is" */
}
#if !USING_KERNEL_PLL_LOOP
freq_drift = 0;
#endif
+#if USING_INITIAL_FREQ_ESTIMATION
if (G.discipline_state == STATE_FREQ) {
/* Ignore updates until the stepout threshold */
if (since_last_update < WATCH_THRESHOLD) {
- VERB3 bb_error_msg("measuring drift, datapoint ignored, %f sec remains",
+ VERB4 bb_error_msg("measuring drift, datapoint ignored, %f sec remains",
WATCH_THRESHOLD - since_last_update);
return 0; /* "leave poll interval as is" */
}
-#if !USING_KERNEL_PLL_LOOP
+# if !USING_KERNEL_PLL_LOOP
freq_drift = (offset - G.last_update_offset) / since_last_update;
-#endif
+# endif
}
+#endif
/* There are two main regimes: when the
* offset exceeds the step threshold and when it does not.
*/
if (abs_offset > STEP_THRESHOLD) {
+#if 0
+ double remains;
+
+// This "spike state" seems to be useless, peer selection already drops
+// occassional "bad" datapoints. If we are here, there were _many_
+// large offsets. When a few first large offsets are seen,
+// we end up in "no valid datapoints, no peer selected" state.
+// Only when enough of them are seen (which means it's not a fluke),
+// we end up here. Looks like _our_ clock is off.
switch (G.discipline_state) {
case STATE_SYNC:
/* The first outlyer: ignore it, switch to SPIK state */
- VERB3 bb_error_msg("offset:%f - spike detected", offset);
+ VERB3 bb_error_msg("update from %s: offset:%+f, spike%s",
+ p->p_dotted, offset,
+ "");
G.discipline_state = STATE_SPIK;
return -1; /* "decrease poll interval" */
/* Ignore succeeding outlyers until either an inlyer
* is found or the stepout threshold is exceeded.
*/
- if (since_last_update < WATCH_THRESHOLD) {
- VERB3 bb_error_msg("spike detected, datapoint ignored, %f sec remains",
- WATCH_THRESHOLD - since_last_update);
+ remains = WATCH_THRESHOLD - since_last_update;
+ if (remains > 0) {
+ VERB3 bb_error_msg("update from %s: offset:%+f, spike%s",
+ p->p_dotted, offset,
+ ", datapoint ignored");
return -1; /* "decrease poll interval" */
}
/* fall through: we need to step */
} /* switch */
+#endif
/* Step the time and clamp down the poll interval.
*
* is always suppressed, even at the longer poll
* intervals.
*/
- VERB3 bb_error_msg("stepping time by %f; poll_exp=MINPOLL", offset);
+ VERB4 bb_error_msg("stepping time by %+f; poll_exp=MINPOLL", offset);
step_time(offset);
if (option_mask32 & OPT_q) {
/* We were only asked to set time once. Done. */
G.polladj_count = 0;
G.poll_exp = MINPOLL;
G.stratum = MAXSTRAT;
+
+ run_script("step", offset);
+
+ recv_time += offset;
+
+#if USING_INITIAL_FREQ_ESTIMATION
if (G.discipline_state == STATE_NSET) {
set_new_values(STATE_FREQ, /*offset:*/ 0, recv_time);
return 1; /* "ok to increase poll interval" */
}
- set_new_values(STATE_SYNC, /*offset:*/ 0, recv_time);
+#endif
+ abs_offset = offset = 0;
+ set_new_values(STATE_SYNC, offset, recv_time);
} else { /* abs_offset <= STEP_THRESHOLD */
if (G.poll_exp < MINPOLL && G.initial_poll_complete) {
- VERB3 bb_error_msg("small offset:%f, disabling burst mode", offset);
+ VERB4 bb_error_msg("small offset:%+f, disabling burst mode", offset);
G.polladj_count = 0;
G.poll_exp = MINPOLL;
}
* weighted offset differences. Used by the poll adjust code.
*/
etemp = SQUARE(G.discipline_jitter);
- dtemp = SQUARE(MAXD(fabs(offset - G.last_update_offset), G_precision_sec));
+ dtemp = SQUARE(offset - G.last_update_offset);
G.discipline_jitter = SQRT(etemp + (dtemp - etemp) / AVG);
- VERB3 bb_error_msg("discipline jitter=%f", G.discipline_jitter);
switch (G.discipline_state) {
case STATE_NSET:
*/
exit(0);
}
+#if USING_INITIAL_FREQ_ESTIMATION
/* This is the first update received and the frequency
* has not been initialized. The first thing to do
* is directly measure the oscillator frequency.
*/
set_new_values(STATE_FREQ, offset, recv_time);
- VERB3 bb_error_msg("transitioning to FREQ, datapoint ignored");
+#else
+ set_new_values(STATE_SYNC, offset, recv_time);
+#endif
+ VERB4 bb_error_msg("transitioning to FREQ, datapoint ignored");
return 0; /* "leave poll interval as is" */
#if 0 /* this is dead code for now */
break;
#endif
+#if USING_INITIAL_FREQ_ESTIMATION
case STATE_FREQ:
/* since_last_update >= WATCH_THRESHOLD, we waited enough.
* Correct the phase and frequency and switch to SYNC state.
*/
set_new_values(STATE_SYNC, offset, recv_time);
break;
+#endif
default:
#if !USING_KERNEL_PLL_LOOP
set_new_values(STATE_SYNC, offset, recv_time);
break;
}
- G.stratum = p->lastpkt_stratum + 1;
+ if (G.stratum != p->lastpkt_stratum + 1) {
+ G.stratum = p->lastpkt_stratum + 1;
+ run_script("stratum", offset);
+ }
}
+ if (G.discipline_jitter < G_precision_sec)
+ G.discipline_jitter = G_precision_sec;
+ G.offset_to_jitter_ratio = abs_offset / G.discipline_jitter;
+
G.reftime = G.cur_time;
G.ntp_status = p->lastpkt_status;
G.refid = p->lastpkt_refid;
G.rootdelay = p->lastpkt_rootdelay + p->lastpkt_delay;
- dtemp = p->filter_jitter; // SQRT(SQUARE(p->filter_jitter) + SQUARE(s.jitter));
+ dtemp = p->filter_jitter; // SQRT(SQUARE(p->filter_jitter) + SQUARE(G.cluster_jitter));
dtemp += MAXD(p->filter_dispersion + FREQ_TOLERANCE * (G.cur_time - p->lastpkt_recv_time) + abs_offset, MINDISP);
G.rootdisp = p->lastpkt_rootdisp + dtemp;
- VERB3 bb_error_msg("updating leap/refid/reftime/rootdisp from peer %s", p->p_dotted);
+ VERB4 bb_error_msg("updating leap/refid/reftime/rootdisp from peer %s", p->p_dotted);
/* We are in STATE_SYNC now, but did not do adjtimex yet.
* (Any other state does not reach this, they all return earlier)
- * By this time, freq_drift and G.last_update_offset are set
+ * By this time, freq_drift and offset are set
* to values suitable for adjtimex.
*/
#if !USING_KERNEL_PLL_LOOP
dtemp = SQUARE(dtemp);
G.discipline_wander = SQRT(etemp + (dtemp - etemp) / AVG);
- VERB3 bb_error_msg("discipline freq_drift=%.9f(int:%ld corr:%e) wander=%f",
+ VERB4 bb_error_msg("discipline freq_drift=%.9f(int:%ld corr:%e) wander=%f",
G.discipline_freq_drift,
(long)(G.discipline_freq_drift * 65536e6),
freq_drift,
G.discipline_wander);
#endif
- VERB3 {
+ VERB4 {
memset(&tmx, 0, sizeof(tmx));
if (adjtimex(&tmx) < 0)
bb_perror_msg_and_die("adjtimex");
- VERB3 bb_error_msg("p adjtimex freq:%ld offset:%ld constant:%ld status:0x%x",
- tmx.freq, tmx.offset, tmx.constant, tmx.status);
+ bb_error_msg("p adjtimex freq:%ld offset:%+ld status:0x%x tc:%ld",
+ tmx.freq, tmx.offset, tmx.status, tmx.constant);
}
- old_tmx_offset = 0;
- if (!G.adjtimex_was_done) {
- G.adjtimex_was_done = 1;
- /* When we use adjtimex for the very first time,
- * we need to ADD to pre-existing tmx.offset - it may be !0
- */
- memset(&tmx, 0, sizeof(tmx));
- if (adjtimex(&tmx) < 0)
- bb_perror_msg_and_die("adjtimex");
- old_tmx_offset = tmx.offset;
- }
memset(&tmx, 0, sizeof(tmx));
#if 0
//doesn't work, offset remains 0 (!) in kernel:
tmx.modes = ADJ_FREQUENCY | ADJ_OFFSET;
/* 65536 is one ppm */
tmx.freq = G.discipline_freq_drift * 65536e6;
- tmx.offset = G.last_update_offset * 1000000; /* usec */
#endif
tmx.modes = ADJ_OFFSET | ADJ_STATUS | ADJ_TIMECONST;// | ADJ_MAXERROR | ADJ_ESTERROR;
- tmx.offset = (G.last_update_offset * 1000000) /* usec */
- /* + (G.last_update_offset < 0 ? -0.5 : 0.5) - too small to bother */
- + old_tmx_offset; /* almost always 0 */
+ tmx.offset = (offset * 1000000); /* usec */
tmx.status = STA_PLL;
if (G.ntp_status & LI_PLUSSEC)
tmx.status |= STA_INS;
if (G.ntp_status & LI_MINUSSEC)
tmx.status |= STA_DEL;
+
tmx.constant = G.poll_exp - 4;
- //tmx.esterror = (u_int32)(clock_jitter * 1e6);
- //tmx.maxerror = (u_int32)((sys_rootdelay / 2 + sys_rootdisp) * 1e6);
+ /* EXPERIMENTAL.
+ * The below if statement should be unnecessary, but...
+ * It looks like Linux kernel's PLL is far too gentle in changing
+ * tmx.freq in response to clock offset. Offset keeps growing
+ * and eventually we fall back to smaller poll intervals.
+ * We can make correction more agressive (about x2) by supplying
+ * PLL time constant which is one less than the real one.
+ * To be on a safe side, let's do it only if offset is significantly
+ * larger than jitter.
+ */
+ if (tmx.constant > 0 && G.offset_to_jitter_ratio >= TIMECONST_HACK_GATE)
+ tmx.constant--;
+
+ //tmx.esterror = (uint32_t)(clock_jitter * 1e6);
+ //tmx.maxerror = (uint32_t)((sys_rootdelay / 2 + sys_rootdisp) * 1e6);
rc = adjtimex(&tmx);
if (rc < 0)
bb_perror_msg_and_die("adjtimex");
/* NB: here kernel returns constant == G.poll_exp, not == G.poll_exp - 4.
* Not sure why. Perhaps it is normal.
*/
- VERB3 bb_error_msg("adjtimex:%d freq:%ld offset:%ld constant:%ld status:0x%x",
- rc, tmx.freq, tmx.offset, tmx.constant, tmx.status);
-#if 0
- VERB3 {
- /* always gives the same output as above msg */
- memset(&tmx, 0, sizeof(tmx));
- if (adjtimex(&tmx) < 0)
- bb_perror_msg_and_die("adjtimex");
- VERB3 bb_error_msg("c adjtimex freq:%ld offset:%ld constant:%ld status:0x%x",
- tmx.freq, tmx.offset, tmx.constant, tmx.status);
- }
-#endif
- if (G.kernel_freq_drift != tmx.freq / 65536) {
- G.kernel_freq_drift = tmx.freq / 65536;
- VERB2 bb_error_msg("kernel clock drift: %ld ppm", G.kernel_freq_drift);
- }
+ VERB4 bb_error_msg("adjtimex:%d freq:%ld offset:%+ld status:0x%x",
+ rc, tmx.freq, tmx.offset, tmx.status);
+ G.kernel_freq_drift = tmx.freq / 65536;
+ VERB2 bb_error_msg("update from:%s offset:%+f jitter:%f clock drift:%+.3fppm tc:%d",
+ p->p_dotted, offset, G.discipline_jitter, (double)tmx.freq / 65536, (int)tmx.constant);
return 1; /* "ok to increase poll interval" */
}
interval = RETRY_INTERVAL;
r = random();
interval += r % (unsigned)(RETRY_INTERVAL / 4);
- VERB3 bb_error_msg("chose retry interval:%u", interval);
+ VERB4 bb_error_msg("chose retry interval:%u", interval);
return interval;
}
static unsigned
interval = 1 << exponent;
r = random();
interval += ((r & (interval-1)) >> 4) + ((r >> 8) & 1); /* + 1/16 of interval, max */
- VERB3 bb_error_msg("chose poll interval:%u (poll_exp:%d exp:%d)", interval, G.poll_exp, exponent);
+ VERB4 bb_error_msg("chose poll interval:%u (poll_exp:%d exp:%d)", interval, G.poll_exp, exponent);
return interval;
}
static NOINLINE void
ssize_t size;
msg_t msg;
double T1, T2, T3, T4;
+ double dv, offset;
unsigned interval;
datapoint_t *datapoint;
peer_t *q;
+ offset = 0;
+
/* We can recvfrom here and check from.IP, but some multihomed
* ntp servers reply from their *other IP*.
* TODO: maybe we should check at least what we can: from.port == 123?
) {
//TODO: always do this?
interval = retry_interval();
- goto set_next_and_close_sock;
+ goto set_next_and_ret;
}
xfunc_die();
}
if (size != NTP_MSGSIZE_NOAUTH && size != NTP_MSGSIZE) {
bb_error_msg("malformed packet received from %s", p->p_dotted);
- goto bail;
+ return;
}
if (msg.m_orgtime.int_partl != p->p_xmt_msg.m_xmttime.int_partl
|| msg.m_orgtime.fractionl != p->p_xmt_msg.m_xmttime.fractionl
) {
- goto bail;
+ /* Somebody else's packet */
+ return;
}
+ /* We do not expect any more packets from this peer for now.
+ * Closing the socket informs kernel about it.
+ * We open a new socket when we send a new query.
+ */
+ close(p->p_fd);
+ p->p_fd = -1;
+
if ((msg.m_status & LI_ALARM) == LI_ALARM
|| msg.m_stratum == 0
|| msg.m_stratum > NTP_MAXSTRATUM
// TODO: stratum 0 responses may have commands in 32-bit m_refid field:
// "DENY", "RSTR" - peer does not like us at all
// "RATE" - peer is overloaded, reduce polling freq
- interval = poll_interval(0);
- bb_error_msg("reply from %s: not synced, next query in %us", p->p_dotted, interval);
- goto set_next_and_close_sock;
+ bb_error_msg("reply from %s: peer is unsynced", p->p_dotted);
+ goto pick_normal_interval;
}
// /* Verify valid root distance */
T4 = G.cur_time;
p->lastpkt_recv_time = T4;
+ VERB6 bb_error_msg("%s->lastpkt_recv_time=%f", p->p_dotted, p->lastpkt_recv_time);
- VERB5 bb_error_msg("%s->lastpkt_recv_time=%f", p->p_dotted, p->lastpkt_recv_time);
- p->datapoint_idx = p->reachable_bits ? (p->datapoint_idx + 1) % NUM_DATAPOINTS : 0;
- datapoint = &p->filter_datapoint[p->datapoint_idx];
- datapoint->d_recv_time = T4;
- datapoint->d_offset = ((T2 - T1) + (T3 - T4)) / 2;
/* The delay calculation is a special case. In cases where the
* server and client clocks are running at different rates and
* with very fast networks, the delay can appear negative. In
* order to avoid violating the Principle of Least Astonishment,
* the delay is clamped not less than the system precision.
*/
+ dv = p->lastpkt_delay;
p->lastpkt_delay = (T4 - T1) - (T3 - T2);
if (p->lastpkt_delay < G_precision_sec)
p->lastpkt_delay = G_precision_sec;
+ /*
+ * If this packet's delay is much bigger than the last one,
+ * it's better to just ignore it than use its much less precise value.
+ */
+ if (p->reachable_bits && p->lastpkt_delay > dv * BAD_DELAY_GROWTH) {
+ bb_error_msg("reply from %s: delay %f is too high, ignoring", p->p_dotted, p->lastpkt_delay);
+ goto pick_normal_interval;
+ }
+
+ p->datapoint_idx = p->reachable_bits ? (p->datapoint_idx + 1) % NUM_DATAPOINTS : 0;
+ datapoint = &p->filter_datapoint[p->datapoint_idx];
+ datapoint->d_recv_time = T4;
+ datapoint->d_offset = offset = ((T2 - T1) + (T3 - T4)) / 2;
datapoint->d_dispersion = LOG2D(msg.m_precision_exp) + G_precision_sec;
if (!p->reachable_bits) {
/* 1st datapoint ever - replicate offset in every element */
int i;
- for (i = 1; i < NUM_DATAPOINTS; i++) {
- p->filter_datapoint[i].d_offset = datapoint->d_offset;
+ for (i = 0; i < NUM_DATAPOINTS; i++) {
+ p->filter_datapoint[i].d_offset = offset;
}
}
p->reachable_bits |= 1;
- VERB1 {
- bb_error_msg("reply from %s: reach 0x%02x offset %f delay %f status 0x%02x strat %d refid 0x%08x rootdelay %f",
+ if ((MAX_VERBOSE && G.verbose) || (option_mask32 & OPT_w)) {
+ bb_error_msg("reply from %s: offset:%+f delay:%f status:0x%02x strat:%d refid:0x%08x rootdelay:%f reach:0x%02x",
p->p_dotted,
- p->reachable_bits,
- datapoint->d_offset,
+ offset,
p->lastpkt_delay,
p->lastpkt_status,
p->lastpkt_stratum,
p->lastpkt_refid,
- p->lastpkt_rootdelay
+ p->lastpkt_rootdelay,
+ p->reachable_bits
/* not shown: m_ppoll, m_precision_exp, m_rootdisp,
* m_reftime, m_orgtime, m_rectime, m_xmttime
*/
rc = -1;
if (q) {
rc = 0;
- if (!(option_mask32 & OPT_w))
+ if (!(option_mask32 & OPT_w)) {
rc = update_local_clock(q);
+ /* If drift is dangerously large, immediately
+ * drop poll interval one step down.
+ */
+ if (fabs(q->filter_offset) >= POLLDOWN_OFFSET) {
+ VERB4 bb_error_msg("offset:%+f > POLLDOWN_OFFSET", q->filter_offset);
+ goto poll_down;
+ }
+ }
}
+ /* else: no peer selected, rc = -1: we want to poll more often */
if (rc != 0) {
/* Adjust the poll interval by comparing the current offset
* is increased, otherwise it is decreased. A bit of hysteresis
* helps calm the dance. Works best using burst mode.
*/
- VERB4 if (rc > 0) {
- bb_error_msg("offset:%f POLLADJ_GATE*discipline_jitter:%f poll:%s",
- q->filter_offset, POLLADJ_GATE * G.discipline_jitter,
- fabs(q->filter_offset) < POLLADJ_GATE * G.discipline_jitter
- ? "grows" : "falls"
- );
- }
- if (rc > 0 && fabs(q->filter_offset) < POLLADJ_GATE * G.discipline_jitter) {
+ if (rc > 0 && G.offset_to_jitter_ratio <= POLLADJ_GATE) {
/* was += G.poll_exp but it is a bit
* too optimistic for my taste at high poll_exp's */
G.polladj_count += MINPOLL;
G.polladj_count = 0;
if (G.poll_exp < MAXPOLL) {
G.poll_exp++;
- VERB3 bb_error_msg("polladj: discipline_jitter:%f ++poll_exp=%d",
+ VERB4 bb_error_msg("polladj: discipline_jitter:%f ++poll_exp=%d",
G.discipline_jitter, G.poll_exp);
}
} else {
- VERB3 bb_error_msg("polladj: incr:%d", G.polladj_count);
+ VERB4 bb_error_msg("polladj: incr:%d", G.polladj_count);
}
} else {
G.polladj_count -= G.poll_exp * 2;
- if (G.polladj_count < -POLLADJ_LIMIT) {
+ if (G.polladj_count < -POLLADJ_LIMIT || G.poll_exp >= BIGPOLL) {
+ poll_down:
G.polladj_count = 0;
if (G.poll_exp > MINPOLL) {
llist_t *item;
if (pp->p_fd < 0)
pp->next_action_time -= (1 << G.poll_exp);
}
- VERB3 bb_error_msg("polladj: discipline_jitter:%f --poll_exp=%d",
+ VERB4 bb_error_msg("polladj: discipline_jitter:%f --poll_exp=%d",
G.discipline_jitter, G.poll_exp);
}
} else {
- VERB3 bb_error_msg("polladj: decr:%d", G.polladj_count);
+ VERB4 bb_error_msg("polladj: decr:%d", G.polladj_count);
}
}
}
/* Decide when to send new query for this peer */
+ pick_normal_interval:
interval = poll_interval(0);
+ if (fabs(offset) >= STEP_THRESHOLD * 8 && interval > BIGOFF_INTERVAL) {
+ /* If we are synced, offsets are less than STEP_THRESHOLD,
+ * or at the very least not much larger than it.
+ * Now we see a largish one.
+ * Either this peer is feeling bad, or packet got corrupted,
+ * or _our_ clock is wrong now and _all_ peers will show similar
+ * largish offsets too.
+ * I observed this with laptop suspend stopping clock.
+ * In any case, it makes sense to make next request soonish:
+ * cases 1 and 2: get a better datapoint,
+ * case 3: allows to resync faster.
+ */
+ interval = BIGOFF_INTERVAL;
+ }
- set_next_and_close_sock:
+ set_next_and_ret:
set_next(p, interval);
- /* We do not expect any more packets from this peer for now.
- * Closing the socket informs kernel about it.
- * We open a new socket when we send a new query.
- */
- close(p->p_fd);
- p->p_fd = -1;
- bail:
- return;
}
#if ENABLE_FEATURE_NTPD_SERVER
recv_and_process_client_pkt(void /*int fd*/)
{
ssize_t size;
- uint8_t version;
+ //uint8_t version;
len_and_sockaddr *to;
struct sockaddr *from;
msg_t msg;
uint8_t query_status;
l_fixedpt_t query_xmttime;
- to = get_sock_lsa(G.listen_fd);
+ to = get_sock_lsa(G_listen_fd);
from = xzalloc(to->len);
- size = recv_from_to(G.listen_fd, &msg, sizeof(msg), MSG_DONTWAIT, from, &to->u.sa, to->len);
+ size = recv_from_to(G_listen_fd, &msg, sizeof(msg), MSG_DONTWAIT, from, &to->u.sa, to->len);
if (size != NTP_MSGSIZE_NOAUTH && size != NTP_MSGSIZE) {
char *addr;
if (size < 0) {
/* Build a reply packet */
memset(&msg, 0, sizeof(msg));
- msg.m_status = G.stratum < MAXSTRAT ? G.ntp_status : LI_ALARM;
+ msg.m_status = G.stratum < MAXSTRAT ? (G.ntp_status & LI_MASK) : LI_ALARM;
msg.m_status |= (query_status & VERSION_MASK);
msg.m_status |= ((query_status & MODE_MASK) == MODE_CLIENT) ?
- MODE_SERVER : MODE_SYM_PAS;
+ MODE_SERVER : MODE_SYM_PAS;
msg.m_stratum = G.stratum;
msg.m_ppoll = G.poll_exp;
msg.m_precision_exp = G_precision_exp;
/* this time was obtained between poll() and recv() */
msg.m_rectime = d_to_lfp(G.cur_time);
msg.m_xmttime = d_to_lfp(gettime1900d()); /* this instant */
+ if (G.peer_cnt == 0) {
+ /* we have no peers: "stratum 1 server" mode. reftime = our own time */
+ G.reftime = G.cur_time;
+ }
msg.m_reftime = d_to_lfp(G.reftime);
msg.m_orgtime = query_xmttime;
msg.m_rootdelay = d_to_sfp(G.rootdelay);
//simple code does not do this, fix simple code!
msg.m_rootdisp = d_to_sfp(G.rootdisp);
- version = (query_status & VERSION_MASK); /* ... >> VERSION_SHIFT - done below instead */
+ //version = (query_status & VERSION_MASK); /* ... >> VERSION_SHIFT - done below instead */
msg.m_refid = G.refid; // (version > (3 << VERSION_SHIFT)) ? G.refid : G.refid3;
/* We reply from the local address packet was sent to,
* this makes to/from look swapped here: */
- do_sendto(G.listen_fd,
+ do_sendto(G_listen_fd,
/*from:*/ &to->u.sa, /*to:*/ from, /*addrlen:*/ to->len,
&msg, size);
G.stratum = MAXSTRAT;
if (BURSTPOLL != 0)
G.poll_exp = BURSTPOLL; /* speeds up initial sync */
- G.reftime = G.last_update_recv_time = gettime1900d(); /* sets G.cur_time too */
+ G.last_script_run = G.reftime = G.last_update_recv_time = gettime1900d(); /* sets G.cur_time too */
/* Parse options */
peers = NULL;
- opt_complementary = "dd:p::"; /* d: counter, p: list */
+ opt_complementary = "dd:p::wn"; /* d: counter; p: list; -w implies -n */
opts = getopt32(argv,
"nqNx" /* compat */
- "wp:"IF_FEATURE_NTPD_SERVER("l") /* NOT compat */
+ "wp:S:"IF_FEATURE_NTPD_SERVER("l") /* NOT compat */
"d" /* compat */
"46aAbgL", /* compat, ignored */
- &peers, &G.verbose);
+ &peers, &G.script_name, &G.verbose);
if (!(opts & (OPT_p|OPT_l)))
bb_show_usage();
// if (opts & OPT_x) /* disable stepping, only slew is allowed */
// G.time_was_stepped = 1;
- while (peers)
- add_peers(llist_pop(&peers));
+ if (peers) {
+ while (peers)
+ add_peers(llist_pop(&peers));
+ } else {
+ /* -l but no peers: "stratum 1 server" mode */
+ G.stratum = 1;
+ }
if (!(opts & OPT_n)) {
bb_daemonize_or_rexec(DAEMON_DEVNULL_STDIO, argv);
logmode = LOGMODE_NONE;
}
#if ENABLE_FEATURE_NTPD_SERVER
- G.listen_fd = -1;
+ G_listen_fd = -1;
if (opts & OPT_l) {
- G.listen_fd = create_and_bind_dgram_or_die(NULL, 123);
- socket_want_pktinfo(G.listen_fd);
- setsockopt(G.listen_fd, IPPROTO_IP, IP_TOS, &const_IPTOS_LOWDELAY, sizeof(const_IPTOS_LOWDELAY));
+ G_listen_fd = create_and_bind_dgram_or_die(NULL, 123);
+ socket_want_pktinfo(G_listen_fd);
+ setsockopt(G_listen_fd, IPPROTO_IP, IP_TOS, &const_IPTOS_LOWDELAY, sizeof(const_IPTOS_LOWDELAY));
}
#endif
/* I hesitate to set -20 prio. -15 should be high enough for timekeeping */
if (opts & OPT_N)
setpriority(PRIO_PROCESS, 0, -15);
- bb_signals((1 << SIGTERM) | (1 << SIGINT), record_signo);
- bb_signals((1 << SIGPIPE) | (1 << SIGHUP), SIG_IGN);
+ /* If network is up, syncronization occurs in ~10 seconds.
+ * We give "ntpd -q" 10 seconds to get first reply,
+ * then another 50 seconds to finish syncing.
+ *
+ * I tested ntpd 4.2.6p1 and apparently it never exits
+ * (will try forever), but it does not feel right.
+ * The goal of -q is to act like ntpdate: set time
+ * after a reasonably small period of polling, or fail.
+ */
+ if (opts & OPT_q) {
+ option_mask32 |= OPT_qq;
+ alarm(10);
+ }
+
+ bb_signals(0
+ | (1 << SIGTERM)
+ | (1 << SIGINT)
+ | (1 << SIGALRM)
+ , record_signo
+ );
+ bb_signals(0
+ | (1 << SIGPIPE)
+ | (1 << SIGCHLD)
+ , SIG_IGN
+ );
}
int ntpd_main(int argc UNUSED_PARAM, char **argv) MAIN_EXTERNALLY_VISIBLE;
idx2peer = xzalloc(sizeof(idx2peer[0]) * cnt);
pfd = xzalloc(sizeof(pfd[0]) * cnt);
- /* Countdown: we never sync before we sent 5 packets to each peer
+ /* Countdown: we never sync before we sent INITIAL_SAMPLES+1
+ * packets to each peer.
* NB: if some peer is not responding, we may end up sending
* fewer packets to it and more to other peers.
- * NB2: sync usually happens using 5-1=4 packets, since last reply
- * does not come back instantaneously.
+ * NB2: sync usually happens using INITIAL_SAMPLES packets,
+ * since last reply does not come back instantaneously.
*/
- cnt = G.peer_cnt * 5;
+ cnt = G.peer_cnt * (INITIAL_SAMPLES + 1);
+
+ write_pidfile(CONFIG_PID_FILE_PATH "/ntpd.pid");
while (!bb_got_signal) {
llist_t *item;
i = 0;
#if ENABLE_FEATURE_NTPD_SERVER
- if (G.listen_fd != -1) {
- pfd[0].fd = G.listen_fd;
+ if (G_listen_fd != -1) {
+ pfd[0].fd = G_listen_fd;
pfd[0].events = POLLIN;
i++;
}
timeout++; /* (nextaction - G.cur_time) rounds down, compensating */
/* Here we may block */
- VERB2 bb_error_msg("poll %us, sockets:%u", timeout, i);
+ VERB2 {
+ if (i > (ENABLE_FEATURE_NTPD_SERVER && G_listen_fd != -1)) {
+ /* We wait for at least one reply.
+ * Poll for it, without wasting time for message.
+ * Since replies often come under 1 second, this also
+ * reduces clutter in logs.
+ */
+ nfds = poll(pfd, i, 1000);
+ if (nfds != 0)
+ goto did_poll;
+ if (--timeout <= 0)
+ goto did_poll;
+ }
+ bb_error_msg("poll:%us sockets:%u interval:%us", timeout, i, 1 << G.poll_exp);
+ }
nfds = poll(pfd, i, timeout * 1000);
+ did_poll:
gettime1900d(); /* sets G.cur_time */
- if (nfds <= 0)
- continue;
+ if (nfds <= 0) {
+ if (!bb_got_signal /* poll wasn't interrupted by a signal */
+ && G.cur_time - G.last_script_run > 11*60
+ ) {
+ /* Useful for updating battery-backed RTC and such */
+ run_script("periodic", G.last_update_offset);
+ gettime1900d(); /* sets G.cur_time */
+ }
+ goto check_unsync;
+ }
/* Process any received packets */
j = 0;
#endif
for (; nfds != 0 && j < i; j++) {
if (pfd[j].revents /* & (POLLIN|POLLERR)*/) {
+ /*
+ * At init, alarm was set to 10 sec.
+ * Now we did get a reply.
+ * Increase timeout to 50 seconds to finish syncing.
+ */
+ if (option_mask32 & OPT_qq) {
+ option_mask32 &= ~OPT_qq;
+ alarm(50);
+ }
nfds--;
recv_and_process_peer_pkt(idx2peer[j]);
gettime1900d(); /* sets G.cur_time */
}
}
+
+ check_unsync:
+ if (G.ntp_peers && G.stratum != MAXSTRAT) {
+ for (item = G.ntp_peers; item != NULL; item = item->link) {
+ peer_t *p = (peer_t *) item->data;
+ if (p->reachable_bits)
+ goto have_reachable_peer;
+ }
+ /* No peer responded for last 8 packets, panic */
+ G.polladj_count = 0;
+ G.poll_exp = MINPOLL;
+ G.stratum = MAXSTRAT;
+ run_script("unsync", 0.0);
+ have_reachable_peer: ;
+ }
} /* while (!bb_got_signal) */
+ remove_pidfile(CONFIG_PID_FILE_PATH "/ntpd.pid");
kill_myself_with_sig(bb_got_signal);
}
static double
direct_freq(double fp_offset)
{
-
#ifdef KERNEL_PLL
/*
* If the kernel is enabled, we need the residual offset to
}
static void
-set_freq(double freq) /* frequency update */
+set_freq(double freq) /* frequency update */
{
char tbuf[80];
if (pps_enable) {
if (!(pll_status & STA_PPSTIME))
report_event(EVNT_KERN,
- NULL, "PPS enabled");
+ NULL, "PPS enabled");
ntv.status |= STA_PPSTIME | STA_PPSFREQ;
} else {
if (pll_status & STA_PPSTIME)
report_event(EVNT_KERN,
- NULL, "PPS disabled");
- ntv.status &= ~(STA_PPSTIME |
- STA_PPSFREQ);
+ NULL, "PPS disabled");
+ ntv.status &= ~(STA_PPSTIME | STA_PPSFREQ);
}
if (sys_leap == LEAP_ADDSECOND)
ntv.status |= STA_INS;
if (ntp_adjtime(&ntv) == TIME_ERROR) {
if (!(ntv.status & STA_PPSSIGNAL))
report_event(EVNT_KERN, NULL,
- "PPS no signal");
+ "PPS no signal");
}
pll_status = ntv.status;
#ifdef STA_NANO
projects / platform / upstream / busybox.git / blobdiff