ntpd: add anti-clock-hopping code
authorDenys Vlasenko <vda.linux@googlemail.com>
Sun, 17 Jan 2010 01:51:33 +0000 (02:51 +0100)
committerDenys Vlasenko <vda.linux@googlemail.com>
Sun, 17 Jan 2010 01:51:33 +0000 (02:51 +0100)
function                                             old     new   delta
select_and_cluster                                   837     950    +113
update_local_clock                                   759     767      +8
root_distance                                         61       -     -61

Signed-off-by: Denys Vlasenko <vda.linux@googlemail.com>
networking/ntpd.c

index b2cd0a3..aca79c9 100644 (file)
@@ -306,13 +306,15 @@ struct globals {
        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
+       //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
 };
@@ -821,6 +823,7 @@ typedef struct {
        peer_t *p;
        int    type;
        double edge;
+       double opt_rd; /* optimization */
 } point_t;
 static int
 compare_point_edge(const void *aa, const void *bb)
@@ -876,6 +879,7 @@ fit(peer_t *p, double rd)
 static peer_t*
 select_and_cluster(void)
 {
+       peer_t     *p;
        llist_t    *item;
        int        i, j;
        int        size = 3 * G.peer_cnt;
@@ -893,10 +897,11 @@ select_and_cluster(void)
        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;
@@ -911,14 +916,17 @@ select_and_cluster(void)
                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;
        }
@@ -999,14 +1007,12 @@ select_and_cluster(void)
         */
        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);
+               /* x.opt_rd == root_distance(p); */
+               survivor[num_survivors].metric = MAXDIST * p->lastpkt_stratum + point[i].opt_rd;
                VERB4 bb_error_msg("survivor[%d] metric:%f peer:%s",
                        num_survivors, survivor[num_survivors].metric, p->p_dotted);
                num_survivors++;
@@ -1050,8 +1056,8 @@ select_and_cluster(void)
                 */
                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;
 
@@ -1093,18 +1099,54 @@ select_and_cluster(void)
                }
        }
 
+       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()
+       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) {
+                               VERB4 bb_error_msg("keeping old synced peer");
+                               p = G.last_update_peer;
+                               goto keep_old;
+                       }
+               }
+       }
+       G.last_update_peer = p;
+ keep_old:
        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->p_dotted,
+                       p->filter_offset,
+                       G.cur_time - p->lastpkt_recv_time
        );
-       return survivor[0].p;
+       return p;
 }
 
 
@@ -1131,6 +1173,7 @@ 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;
@@ -1343,7 +1386,7 @@ update_local_clock(peer_t *p)
        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);
@@ -1433,7 +1476,8 @@ update_local_clock(peer_t *p)
        }
 #endif
        G.kernel_freq_drift = tmx.freq / 65536;
-       VERB2 bb_error_msg("update offset:%f, clock drift:%ld ppm", G.last_update_offset, G.kernel_freq_drift);
+       VERB2 bb_error_msg("update peer:%s, offset:%f, clock drift:%ld ppm",
+                       p->p_dotted, G.last_update_offset, G.kernel_freq_drift);
 
        return 1; /* "ok to increase poll interval" */
 }