X-Git-Url: https://git.librecmc.org/?a=blobdiff_plain;f=networking%2Fntpd.c;h=6d9183a4bfa04264d665d12f1ad79f10e16c8dd2;hb=765b0eed3ef29a80115708c3249d3a541509cd24;hp=038f2bded87a731ddb7689007976f05939ad6fc4;hpb=6959f6bc23abb9057c3ed4330dad1bc8f4a1dd7a;p=oweals%2Fbusybox.git diff --git a/networking/ntpd.c b/networking/ntpd.c index 038f2bded..6d9183a4b 100644 --- a/networking/ntpd.c +++ b/networking/ntpd.c @@ -46,41 +46,80 @@ #define MAX_VERBOSE 2 +/* High-level description of the algorithm: + * + * We start running with very small poll_exp, BURSTPOLL, + * in order to quickly accumulate INITIAL_SAMLPES 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. + */ + #define RETRY_INTERVAL 5 /* on error, retry in N secs */ #define RESPONSE_INTERVAL 15 /* wait for reply up to N secs */ +#define INITIAL_SAMLPES 4 /* how many samples do we want for init */ + +/* Clock discipline parameters and constants */ -#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) */ +/* 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 /* 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) */ +#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) +#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 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) */ -#define POLLADJ_LIMIT 30 +#define POLLADJ_LIMIT 36 /* If offset < POLLADJ_GATE * discipline_jitter, then we can 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 */ +/* Compromise Allan intercept (sec). doc uses 1500, std ntpd uses 512 */ #define ALLAN 512 /* PLL loop gain */ #define PLL 65536 @@ -185,6 +224,9 @@ typedef struct { } peer_t; +#define USING_KERNEL_PLL_LOOP 1 +#define USING_INITIAL_FREQ_ESTIMATION 0 + enum { OPT_n = (1 << 0), OPT_q = (1 << 1), @@ -252,21 +294,26 @@ struct globals { #define G_precision_sec (1.0 / (1 << (- G_precision_exp))) 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 + //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 }; @@ -552,8 +599,10 @@ static void reset_peer_stats(peer_t *p, double offset) { int i; + bool small_ofs = fabs(offset) < 16 * STEP_THRESHOLD; + for (i = 0; i < NUM_DATAPOINTS; i++) { - if (offset < 16 * STEP_THRESHOLD) { + 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; @@ -564,7 +613,7 @@ reset_peer_stats(peer_t *p, double offset) p->filter_datapoint[i].d_dispersion = MAXDISP; } } - if (offset < 16 * STEP_THRESHOLD) { + if (small_ofs) { p->lastpkt_recv_time -= offset; } else { p->reachable_bits = 0; @@ -686,10 +735,17 @@ send_query_to_peer(peer_t *p) } -static void run_script(const char *action) +/* Note that there is no provision to prevent several run_scripts + * to be done 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; + char *env1, *env2, *env3, *env4; if (!G.script_name) return; @@ -700,25 +756,35 @@ static void run_script(const char *action) VERB1 bb_error_msg("executing '%s %s'", G.script_name, action); - env1 = xasprintf("stratum=%u", G.stratum); + env1 = xasprintf("%s=%u", "stratum", G.stratum); putenv(env1); - env2 = xasprintf("freq_drift_ppm=%ld", G.kernel_freq_drift); + 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, poll_exp, - * last_update_offset, last_update_recv_time, discipline_jitter + * 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): */ - /*wait4pid(spawn(argv));*/ + /*spawn_and_wait(argv);*/ spawn(argv); - G.last_script_run = G.cur_time; unsetenv("stratum"); unsetenv("freq_drift_ppm"); + unsetenv("poll_interval"); + unsetenv("offset"); free(env1); free(env2); + free(env3); + free(env4); + + G.last_script_run = G.cur_time; } static NOINLINE void @@ -753,6 +819,7 @@ step_time(double offset) /* Globals: */ G.cur_time -= offset; G.last_update_recv_time -= offset; + G.last_script_run -= offset; } @@ -763,6 +830,7 @@ 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) @@ -818,6 +886,7 @@ 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; @@ -835,10 +904,11 @@ 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; @@ -853,14 +923,17 @@ 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; } @@ -941,14 +1014,12 @@ 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++; @@ -992,8 +1063,8 @@ 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; @@ -1035,18 +1106,54 @@ 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; } @@ -1066,19 +1173,13 @@ set_new_values(int disc_state, double offset, double recv_time) 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; @@ -1090,10 +1191,14 @@ update_local_clock(peer_t *p) 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 @@ -1113,6 +1218,7 @@ update_local_clock(peer_t *p) #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) { @@ -1120,10 +1226,11 @@ update_local_clock(peer_t *p) 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. @@ -1180,12 +1287,14 @@ update_local_clock(peer_t *p) G.poll_exp = MINPOLL; G.stratum = MAXSTRAT; - run_script("step"); + run_script("step", 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" */ } +#endif set_new_values(STATE_SYNC, /*offset:*/ 0, recv_time); } else { /* abs_offset <= STEP_THRESHOLD */ @@ -1212,11 +1321,15 @@ update_local_clock(peer_t *p) */ 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); +#else + set_new_values(STATE_SYNC, offset, recv_time); +#endif VERB3 bb_error_msg("transitioning to FREQ, datapoint ignored"); return 0; /* "leave poll interval as is" */ @@ -1231,6 +1344,7 @@ update_local_clock(peer_t *p) 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. @@ -1238,6 +1352,7 @@ update_local_clock(peer_t *p) */ set_new_values(STATE_SYNC, offset, recv_time); break; +#endif default: #if !USING_KERNEL_PLL_LOOP @@ -1269,7 +1384,7 @@ update_local_clock(peer_t *p) } if (G.stratum != p->lastpkt_stratum + 1) { G.stratum = p->lastpkt_stratum + 1; - run_script("stratum"); + run_script("stratum", offset); } } @@ -1277,7 +1392,7 @@ 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); @@ -1314,17 +1429,6 @@ update_local_clock(peer_t *p) tmx.freq, tmx.offset, tmx.constant, tmx.status); } - 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: @@ -1337,9 +1441,8 @@ update_local_clock(peer_t *p) 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 */ + 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.status = STA_PLL; if (G.ntp_status & LI_PLUSSEC) tmx.status |= STA_INS; @@ -1366,10 +1469,9 @@ update_local_clock(peer_t *p) 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); - } + 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); return 1; /* "ok to increase poll interval" */ } @@ -1533,9 +1635,18 @@ recv_and_process_peer_pkt(peer_t *p) 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) { + VERB3 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 @@ -1567,7 +1678,8 @@ recv_and_process_peer_pkt(peer_t *p) } } 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; @@ -1831,13 +1943,14 @@ int ntpd_main(int argc UNUSED_PARAM, char **argv) 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_SAMLPES+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_SAMLPES packets, + * since last reply does not come back instantaneously. */ - cnt = G.peer_cnt * 5; + cnt = G.peer_cnt * (INITIAL_SAMLPES + 1); while (!bb_got_signal) { llist_t *item; @@ -1897,15 +2010,13 @@ int ntpd_main(int argc UNUSED_PARAM, char **argv) timeout++; /* (nextaction - G.cur_time) rounds down, compensating */ /* Here we may block */ - VERB2 bb_error_msg("poll %us, sockets:%u", timeout, i); + VERB2 bb_error_msg("poll %us, sockets:%u, poll interval:%us", timeout, i, 1 << G.poll_exp); nfds = poll(pfd, i, timeout * 1000); gettime1900d(); /* sets G.cur_time */ if (nfds <= 0) { - if (G.adjtimex_was_done - && G.cur_time - G.last_script_run > 11*60 - ) { + if (G.script_name && G.cur_time - G.last_script_run > 11*60) { /* Useful for updating battery-backed RTC and such */ - run_script("periodic"); + run_script("periodic", G.last_update_offset); gettime1900d(); /* sets G.cur_time */ } continue;