Changeset 3232 in MondoRescue for branches/3.2/mindi-busybox/networking/ntpd.c

Timestamp:

Jan 1, 2014, 12:47:38 AM (10 years ago)

Author:

Bruno Cornec

Message:

• Update mindi-busybox to 1.21.1

File:

• branches/3.2/mindi-busybox/networking/ntpd.c (modified) (57 diffs)

branches/3.2/mindi-busybox/networking/ntpd.c

@@ -28 +28 @@
  ***********************************************************************
  */
+
+//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
@@ -92 +109 @@
 #define BURSTPOLL       0       /* initial poll */
 #define MINPOLL         5       /* minimum poll interval. std ntpd uses 6 (6: 64 sec) */
-#define BIGPOLL         10      /* drop to lower poll at any trouble (10: 17 min) */
+/* 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 0.03 sec */
-#define POLLDOWN_OFFSET (STEP_THRESHOLD / 4)
+ * 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) */
@@ -109 +129 @@
 /* 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--
- * (bumped from 30 to 36 since otherwise I often see poll_exp going *2* steps down)
+ * 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   36
-/* 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
+#define TIMECONST_HACK_GATE 2
 /* Compromise Allan intercept (sec). doc uses 1500, std ntpd uses 512 */
 #define ALLAN           512
@@ -193 +214 @@
 
 typedef struct {
+    double d_offset;
     double d_recv_time;
-    double d_offset;
     double d_dispersion;
 } datapoint_t;
@@ -201 +222 @@
     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;
@@ -209 +228 @@
     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;
@@ -239 +260 @@
     OPT_S = (1 << 6),
     OPT_l = (1 << 7) * ENABLE_FEATURE_NTPD_SERVER,
+    /* We hijack some bits for other purposes */
+    OPT_qq = (1 << 31),
 };
 
@@ -255 +278 @@
 #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;
@@ -272 +299 @@
      * 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
@@ -286 +313 @@
      * 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.
-     */
-#define G_precision_exp  -8
-#define G_precision_sec  (1.0 / (1 << (- G_precision_exp)))
+     * -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_sec  0.002
     uint8_t  stratum;
     /* Bool. After set to 1, never goes back to 0: */
@@ -310 +345 @@
     double   last_update_recv_time; // s.t
     double   discipline_jitter;     // c.jitter
+    /* 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
@@ -483 +522 @@
 {
     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;
@@ -491 +539 @@
     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;
-    }
-
-    idx = p->datapoint_idx; /* most recent datapoint */
+        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's index */
     /* Average offset:
      * Drop two outliers and take weighted average of the rest:
@@ -521 +571 @@
             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);
-
-        if (minoff == p->filter_datapoint[idx].d_offset) {
+        sum += dispersion(&fdp[idx]) / (2 << i);
+
+        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;
@@ -570 +620 @@
     }
     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
@@ -584 +660 @@
     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,
+    VERB3 bb_error_msg("filter offset:%+f disp:%f jitter:%f",
+            p->filter_offset,
             p->filter_dispersion,
             p->filter_jitter);
@@ -605 +681 @@
             p->filter_datapoint[i].d_recv_time += offset;
             if (p->filter_datapoint[i].d_offset != 0) {
-                p->filter_datapoint[i].d_offset += offset;
+                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 {
@@ -702 +782 @@
     }
 
+    /* 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
@@ -729 +815 @@
 
     p->reachable_bits <<= 1;
-    VERB1 bb_error_msg("sent query to %s", p->p_dotted);
     set_next(p, RESPONSE_INTERVAL);
 }
@@ -791 +876 @@
     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(buf, sizeof(buf), "%Y-%m-%d %H:%M:%S", localtime(&tval));
+        bb_error_msg("current time is %s.%06u", buf, (unsigned)tvc.tv_usec);
+    }
+    tval = tvn.tv_sec;
+    strftime(buf, sizeof(buf), "%Y-%m-%d %H:%M:%S", localtime(&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: */
@@ -814 +906 @@
         peer_t *pp = (peer_t *) item->data;
         reset_peer_stats(pp, offset);
-        //bb_error_msg("offset:%f pp->next_action_time:%f -> %f",
+        //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;
-    }
-    /* Globals: */
-    G.cur_time += offset;
-    G.last_update_recv_time += offset;
-    G.last_script_run += 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);
+        }
+    }
 }
 
@@ -883 +980 @@
 //  if (p->refid == p->dstaddr || p->refid == s.refid)
 //      return 0;
-        return 1;
+    return 1;
 }
 static peer_t*
@@ -1150 +1247 @@
     G.last_update_peer = p;
  keep_old:
-    VERB3 bb_error_msg("selected peer %s filter_offset:%f age:%f",
+    VERB3 bb_error_msg("selected peer %s filter_offset:%+f age:%f",
             p->p_dotted,
             p->filter_offset,
@@ -1241 +1338 @@
         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("offset:%+f - spike detected", offset);
             G.discipline_state = STATE_SPIK;
             return -1; /* "decrease poll interval" */
@@ -1278 +1375 @@
          * intervals.
          */
-        VERB3 bb_error_msg("stepping time by %f; poll_exp=MINPOLL", offset);
+        VERB3 bb_error_msg("stepping time by %+f; poll_exp=MINPOLL", offset);
         step_time(offset);
         if (option_mask32 & OPT_q) {
@@ -1297 +1394 @@
         }
 #endif
-        set_new_values(STATE_SYNC, /*offset:*/ 0, recv_time);
+        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);
+            VERB3 bb_error_msg("small offset:%+f, disabling burst mode", offset);
             G.polladj_count = 0;
             G.poll_exp = MINPOLL;
@@ -1311 +1409 @@
          */
         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) {
@@ -1390 +1487 @@
     }
 
+    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;
@@ -1401 +1502 @@
     /* 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.
      */
@@ -1427 +1528 @@
         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);
     }
 
@@ -1440 +1541 @@
     /* 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 */
+    tmx.offset = (offset * 1000000); /* usec */
     tmx.status = STA_PLL;
     if (G.ntp_status & LI_PLUSSEC)
@@ -1450 +1549 @@
     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)
@@ -1459 +1572 @@
      * 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
+    VERB3 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 peer:%s, offset:%f, clock drift:%ld ppm",
-            p->p_dotted, G.last_update_offset, G.kernel_freq_drift);
+    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" */
@@ -1532 +1635 @@
 //TODO: always do this?
             interval = retry_interval();
-            goto set_next_and_close_sock;
+            goto set_next_and_ret;
         }
         xfunc_die();
@@ -1539 +1642 @@
     if (size != NTP_MSGSIZE_NOAUTH && size != NTP_MSGSIZE) {
         bb_error_msg("malformed packet received from %s", p->p_dotted);
-        goto bail;
+        return;
     }
 
@@ -1545 +1648 @@
      || 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
@@ -1556 +1667 @@
 // "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, next query in %us", p->p_dotted, interval);
+        goto set_next_and_ret;
     }
 
@@ -1609 +1720 @@
         /* 1st datapoint ever - replicate offset in every element */
         int i;
-        for (i = 1; i < NUM_DATAPOINTS; i++) {
+        for (i = 0; i < NUM_DATAPOINTS; i++) {
             p->filter_datapoint[i].d_offset = datapoint->d_offset;
         }
@@ -1616 +1727 @@
     p->reachable_bits |= 1;
     if ((MAX_VERBOSE && G.verbose) || (option_mask32 & OPT_w)) {
-        bb_error_msg("reply from %s: reach 0x%02x offset %f delay %f status 0x%02x strat %d refid 0x%08x rootdelay %f",
+        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,
             p->lastpkt_delay,
@@ -1624 +1734 @@
             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
@@ -1643 +1754 @@
              */
             if (fabs(q->filter_offset) >= POLLDOWN_OFFSET) {
-                VERB3 bb_error_msg("offset:%f > POLLDOWN_OFFSET", q->filter_offset);
+                VERB3 bb_error_msg("offset:%+f > POLLDOWN_OFFSET", q->filter_offset);
                 goto poll_down;
             }
@@ -1657 +1768 @@
          * 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 */
@@ -1710 +1814 @@
     interval = poll_interval(0);
 
- 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;
 }
 
@@ -1727 +1823 @@
 {
     ssize_t          size;
-    uint8_t          version;
+    //uint8_t          version;
     len_and_sockaddr *to;
     struct sockaddr  *from;
@@ -1734 +1830 @@
     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;
@@ -1756 +1852 @@
     /* 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;
@@ -1775 +1871 @@
 //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);
@@ -1919 +2015 @@
     }
 #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
@@ -1931 +2027 @@
 
     /* If network is up, syncronization occurs in ~10 seconds.
-     * We give "ntpd -q" a full minute to finish, then we exit.
+     * 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
@@ -1938 +2035 @@
      * after a reasonably small period of polling, or fail.
      */
-    if (opts & OPT_q)
-        alarm(60);
+    if (opts & OPT_q) {
+        option_mask32 |= OPT_qq;
+        alarm(10);
+    }
 
     bb_signals(0
@@ -1982 +2081 @@
     cnt = G.peer_cnt * (INITIAL_SAMPLES + 1);
 
+    write_pidfile(CONFIG_PID_FILE_PATH "/ntpd.pid");
+
     while (!bb_got_signal) {
         llist_t *item;
@@ -1994 +2095 @@
         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++;
@@ -2040 +2141 @@
 
         /* Here we may block */
-        VERB2 bb_error_msg("poll %us, sockets:%u, poll interval:%us", timeout, i, 1 << G.poll_exp);
+        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) {
@@ -2066 +2182 @@
         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]);
@@ -2073 +2198 @@
     } /* while (!bb_got_signal) */
 
+    remove_pidfile(CONFIG_PID_FILE_PATH "/ntpd.pid");
     kill_myself_with_sig(bb_got_signal);
 }
@@ -2203 +2329 @@
                 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)
@@ -2226 +2351 @@
             if (!(ntv.status & STA_PPSSIGNAL))
                 report_event(EVNT_KERN, NULL,
-                    "PPS no signal");
+                        "PPS no signal");
         }
         pll_status = ntv.status;