* 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 */
+#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 */
#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 STATE_NSET 0 /* initial state, "nothing is set" */
//#define STATE_FSET 1 /* frequency set from file */
-#define STATE_SPIK 2 /* spike detected */
+//#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
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 (small_ofs) {
p->filter_datapoint[i].d_recv_time += 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 (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 */
* offset exceeds the step threshold and when it does not.
*/
if (abs_offset > STEP_THRESHOLD) {
+#if 0
double remains;
-// TODO: this "spike state" seems to be useless, peer selection already drops
-// occassional "bad" datapoints. If we are here, there were _many_ large offsets -
-// looks like _our_ clock is off.
+// 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 */
}
/* fall through: we need to step */
} /* switch */
+#endif
/* Step the time and clamp down the poll interval.
*
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);
ssize_t size;
msg_t msg;
double T1, T2, T3, T4;
- double dv;
+ 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?
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;
+ 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 = 0; i < NUM_DATAPOINTS; i++) {
- p->filter_datapoint[i].d_offset = datapoint->d_offset;
+ p->filter_datapoint[i].d_offset = offset;
}
}
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,
- datapoint->d_offset,
+ offset,
p->lastpkt_delay,
p->lastpkt_status,
p->lastpkt_stratum,
/* 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_ret:
set_next(p, interval);
projects / platform / upstream / busybox.git / blobdiff