/* $NetBSD: refclock_chu.c,v 1.3 1998/03/06 18:17:23 christos Exp $ */ /* * refclock_chu - clock driver for the CHU time code */ #ifdef HAVE_CONFIG_H #include #endif #if defined(REFCLOCK) && defined(CHUCLK) #include #include #include #include "ntpd.h" #include "ntp_io.h" #include "ntp_refclock.h" #include "ntp_unixtime.h" #include #include "ntp_stdlib.h" /* * The CHU time signal includes a time code which is modulated at the * standard Bell 103 frequencies (i.e. mark=2225Hz, space=2025Hz). * and formatted into 8 bit characters with one start bit and two * stop bits. The time code is composed of 10 8-bit characters. * The second 5 bytes of the timecode are a redundancy check, and * are a copy of the first 5 bytes. * * It is assumed that you have built or modified a Bell 103 standard * modem, attached the input to the output of a radio and cabled the * output to a serial port on your computer, i.e. what you are receiving * is essentially the output of your radio. It is also assumed you have * installed a special CHU line discipline to condition the output from * the terminal driver and take accurate time stamps. * * There are two types of timecodes. One is sent in the 32nd * through 39th second of the minute. * * 6dddhhmmss6dddhhmmss * * where ddd is the day of the year, hh is the hour (in UTC), mm is * the minute and ss the second. The 6 is a constant. Note that * the code is sent twice. * * The second sort of timecode is sent only during the 31st second * past the minute. * * xdyyyyttabXDYYYYTTAB * * In this case, the second part of the code is the one's complement * of the code. This differentiates it from the other timecode * format. * * d is the absolute value of DUT (in tenths of a second). yyyy * is the year. tt is the difference between UTC and TAI. a is * a canadian daylight time flag and b is a serial number. * x is a bitwise field. The least significant bit of x is * one if DUT is negative. The 2nd bit is set if a leap second * will be added at the next opportunity. The 3rd bit is set if * a leap second will be deleted at the next opportunity. * The 4th bit is an even parity bit for the other three bits * in this nibble. * * The start bit in each character has a precise relationship to * the on-time second. Most often UART's synchronize themselves to the * start bit and will post an interrupt at the center of the first stop * bit. Thus each character's interrupt should occur at a fixed offset * from the on-time second. This means that a timestamp taken at the * arrival of each character in the code will provide an independent * estimate of the offset. Since there are 10 characters in the time * code and the code is sent 9 times per minute, this means you * potentially get 90 offset samples per minute. Much of the code in * here is dedicated to producing a single offset estimate from these * samples. * * A note about the line discipline. It is possible to receive the * CHU time code in raw mode, but this has disadvantages. In particular, * this puts a lot of code between the interrupt and the time you freeze * a time stamp, decreasing precision. It is also expensive in terms of * context switches, and made even more expensive by the way I do I/O. * Worse, since you are listening directly to the output of your radio, * CHU is noisy and will make you spend a lot of time receiving noise. * * The line discipline fixes a lot of this. It knows that the CHU time * code consists of 10 bytes which arrive with an intercharacter * spacing of about 37 ms, and that the data is BCD, and filters on this * basis. It delivers block of ten characters plus their associated time * stamps all at once. The time stamps are hence about as accurate as * a Unix machine can get them, and much of the noise disappears in the * kernel with no context switching cost. * * The kernel module also will insure that the packets that are * delivered have the correct redundancy bytes, and will return * a flag in chutype to differentiate one sort of packet from * the other. */ /* * CHU definitions */ #define DEVICE "/dev/chu%d" /* device name and unit */ #define SPEED232 B300 /* uart speed (300 baud) */ #define PRECISION (-9) /* what the heck */ #define REFID "CHU\0" /* reference ID */ #define DESCRIPTION "Scratchbuilt CHU Receiver" /* WRU */ #define NCHUCODES 8 /* expect 8 CHU codes per minute */ #ifndef CHULDISC #define CHULDISC 10 /* XXX temp CHU line discipline */ #endif /* * To compute a quality for the estimate (a pseudo dispersion) we add a * fixed 10 ms for each missing code in the minute and add to this * the sum of the differences between the remaining offsets and the * estimated sample offset. */ #define CHUDELAYPENALTY 0x0000028f /* * Default fudge factors */ #define DEFPROPDELAY 0x00624dd3 /* 0.0015 seconds, 1.5 ms */ #define DEFFILTFUDGE 0x000d1b71 /* 0.0002 seconds, 200 us */ /* * Hacks to avoid excercising the multiplier. I have no pride. */ #define MULBY10(x) (((x)<<3) + ((x)<<1)) #define MULBY60(x) (((x)<<6) - ((x)<<2)) /* watch overflow */ #define MULBY24(x) (((x)<<4) + ((x)<<3)) /* * Constants for use when multiplying by 0.1. ZEROPTONE is 0.1 * as an l_fp fraction, NZPOBITS is the number of significant bits * in ZEROPTONE. */ #define ZEROPTONE 0x1999999a #define NZPOBITS 29 static char hexstring[]="0123456789abcdef"; /* * Unit control structure. */ struct chuunit { struct peer *peer; /* peer structure pointer */ struct event chutimer; /* timeout timer structure */ l_fp offsets[NCHUCODES]; /* offsets computed from each code */ l_fp rectimes[NCHUCODES]; /* times we received this stuff */ u_long reftimes[NCHUCODES]; /* time of last code received */ u_char lastcode[NCHUCHARS * 4]; /* last code we received */ u_char expect; /* the next offset expected */ u_short haveoffset; /* flag word indicating valid offsets */ u_short flags; /* operational flags */ u_long responses; /* number of responses */ int pollcnt; /* poll message counter */ }; #define CHUTIMERSET 0x1 /* timer is set to fire */ /* * The CHU table. This gives the expected time of arrival of each * character after the on-time second and is computed as follows: * The CHU time code is sent at 300 bps. Your average UART will * synchronize at the edge of the start bit and will consider the * character complete at the middle of the first stop bit, i.e. * 0.031667 ms later (some UARTS may complete the character at the * end of the stop bit instead of the middle, but you can fudge this). * Thus the expected time of each interrupt is the start bit time plus * 0.031667 seconds. These times are in chutable[]. */ #define CHARDELAY 0x081b4e82 static u_long chutable[NCHUCHARS] = { 0x22222222 + CHARDELAY, /* 0.1333333333 */ 0x2b851eb8 + CHARDELAY, /* 0.170 (exactly) */ 0x34e81b4e + CHARDELAY, /* 0.2066666667 */ 0x3f92c5f9 + CHARDELAY, /* 0.2483333333 */ 0x47ae147b + CHARDELAY, /* 0.280 (exactly) */ 0x51111111 + CHARDELAY, /* 0.3166666667 */ 0x5a740da7 + CHARDELAY, /* 0.3533333333 */ 0x63d70a3d + CHARDELAY, /* 0.390 (exactly) */ 0x6d3a06d4 + CHARDELAY, /* 0.4266666667 */ 0x769d0370 + CHARDELAY, /* 0.4633333333 */ }; /* * Imported from the timer module */ extern u_long current_time; extern struct event timerqueue[]; /* * Imported from ntpd module */ extern int debug; /* global debug flag */ /* * Function prototypes */ static int chu_start P((int, struct peer *)); static void chu_shutdown P((int, struct peer *)); static void chu_receive P((struct recvbuf *)); static void chu_process P((struct chuunit *)); static void chu_poll P((int, struct peer *)); static void chu_timeout P((struct peer *)); /* * Transfer vector */ struct refclock refclock_chu = { chu_start, /* start up driver */ chu_shutdown, /* shut down driver */ chu_poll, /* transmit poll message */ noentry, /* not used (old chu_control) */ noentry, /* initialize driver (not used) */ noentry, /* not used (old chu_buginfo) */ NOFLAGS /* not used */ }; /* * chu_start - open the CHU device and initialize data for processing */ static int chu_start(unit, peer) int unit; struct peer *peer; { register struct chuunit *up; struct refclockproc *pp; int fd; char device[20]; /* * Open serial port and set CHU line discipline */ (void) sprintf(device, DEVICE, unit); if (!(fd = refclock_open(device, SPEED232, LDISC_CHU))) return (0); /* * Allocate and initialize unit structure */ if (!(up = (struct chuunit *) emalloc(sizeof(struct chuunit)))) { (void) close(fd); return (0); } memset((char *)up, 0, sizeof(struct chuunit)); up->chutimer.peer = (struct peer *)up; up->chutimer.event_handler = chu_timeout; up->peer = peer; pp = peer->procptr; pp->io.clock_recv = chu_receive; pp->io.srcclock = (caddr_t)peer; pp->io.datalen = 0; pp->io.fd = fd; if (!io_addclock(&pp->io)) { (void) close(fd); free(up); return (0); } pp->unitptr = (caddr_t)up; /* * Initialize miscellaneous variables */ peer->precision = PRECISION; pp->clockdesc = DESCRIPTION; memcpy((char *)&pp->refid, REFID, 4); up->pollcnt = 2; return (1); } /* * chu_shutdown - shut down the clock */ static void chu_shutdown(unit, peer) int unit; struct peer *peer; { register struct chuunit *up; struct refclockproc *pp; pp = peer->procptr; up = (struct chuunit *)pp->unitptr; io_closeclock(&pp->io); free(up); } /* * chu_receive - receive data from a CHU clock, do format checks and compute * an estimate from the sample data */ static void chu_receive(rbufp) struct recvbuf *rbufp; { register struct chuunit *up; struct refclockproc *pp; struct peer *peer; int i; u_long date_ui; u_long tmp; u_char *code; struct chucode *chuc; int isneg; u_long reftime; l_fp off[NCHUCHARS]; int day, hour, minute, second; /* * Do a length check on the data. Should be what we asked for. */ if (rbufp->recv_length != sizeof(struct chucode)) { msyslog(LOG_ERR, "chu_receive: received %d bytes, expected %d", rbufp->recv_length, sizeof(struct chucode)); return; } /* * Get the clock this applies to and a pointer to the data */ peer = (struct peer *)rbufp->recv_srcclock; pp = peer->procptr; up = (struct chuunit *)pp->unitptr; chuc = (struct chucode *)&rbufp->recv_space; up->responses++; /* * Just for fun, we can debug the whole frame if * we want. */ for (i = 0; i < NCHUCHARS; i++) { pp->a_lastcode[2 * i] = hexstring[chuc->codechars[i] & 0xf]; pp->a_lastcode[2 * i + 1] = hexstring[chuc->codechars[i] >> 4]; } pp->lencode = 2 * i; pp->a_lastcode[pp->lencode] = '\0'; #ifdef DEBUG if (debug > 3) { printf("chu: %s packet\n", (chuc->chutype == CHU_YEAR)? "year":"time"); for (i = 0; i < NCHUCHARS; i++) { char c[64]; sprintf(c,"%c%c %s", hexstring[chuc->codechars[i] & 0xf], hexstring[chuc->codechars[i] >> 4], ctime(&(chuc->codetimes[i].tv_sec))); c[strlen(c) - 1] = 0; /* ctime() adds \n */ printf("chu: %s .%06ld\n", c, chuc->codetimes[i].tv_usec); } } #endif /* * At this point we're assured that both halves of the * data match because of what the kernel has done. * But there's more than one data format. We need to * check chutype to see what to do now. If it's a * year packet, then we fiddle with it specially. */ if (chuc->chutype == CHU_YEAR) { u_char leapbits,parity; /* * Break out the code into the BCD nibbles. * Put it in the half of lastcode. */ code = up->a_lastcode; code += 2*NCHUCHARS; for (i = 0; i < NCHUCHARS; i++) { *code++ = chuc->codechars[i] & 0xf; *code++ = (chuc->codechars[i] >> 4) & 0xf; } leapbits = chuc->codechars[0]&0xf; /* * Now make sure that the leap nibble * is even parity. */ parity = (leapbits ^ (leapbits >> 2))&0x3; parity = (parity ^ (parity>>1))&0x1; if (parity) { refclock_report(peer, CEVNT_BADREPLY); return; } /* * This just happens to work. :-) */ pp->leap = (leapbits >> 1) & 0x3; return; } if (chuc->chutype != CHU_TIME) { refclock_report(peer, CEVNT_BADREPLY); return; } /* * Break out the code into the BCD nibbles. Only need to fiddle * with the first half since both are identical. Note the first * BCD character is the low order nibble, the second the high order. */ code = up->a_lastcode; for (i = 0; i < NCHUCHARS; i++) { *code++ = chuc->codechars[i] & 0xf; *code++ = (chuc->codechars[i] >> 4) & 0xf; } /* * Format check. Make sure the two halves match. * There's really no need for this, but it can't hurt. */ for (i = 0; i < NCHUCHARS/2; i++) if (chuc->codechars[i] != chuc->codechars[i+(NCHUCHARS/2)]) { refclock_report(peer, CEVNT_BADREPLY); return; } /* * If the first nibble isn't a 6, we're up the creek */ code = up->a_lastcode; if (*code++ != 6) { refclock_report(peer, CEVNT_BADREPLY); return; } /* * Collect the day, the hour, the minute and the second. */ day = *code++; day = MULBY10(day) + *code++; day = MULBY10(day) + *code++; hour = *code++; hour = MULBY10(hour) + *code++; minute = *code++; minute = MULBY10(minute) + *code++; second = *code++; second = MULBY10(second) + *code++; /* * Sanity check the day and time. Note that this * only occurs on the 32st through the 39th second * of the minute. */ if (day < 1 || day > 366 || hour > 23 || minute > 59 || second < 32 || second > 39) { pp->baddata++; if (day < 1 || day > 366) { refclock_report(peer, CEVNT_BADDATE); } else { refclock_report(peer, CEVNT_BADTIME); } return; } /* * Compute the NTP date from the input data and the * receive timestamp. If this doesn't work, mark the * date as bad and forget it. */ if (!clocktime(day, hour, minute, second, GMT, rbufp->recv_time.l_ui, &pp->yearstart, (u_int32 *)&reftime)) { refclock_report(peer, CEVNT_BADDATE); return; } date_ui = reftime;; /* * We've now got the integral seconds part of the time code (we hope). * The fractional part comes from the table. We next compute * the offsets for each character. */ for (i = 0; i < NCHUCHARS; i++) { register u_long tmp2; off[i].l_ui = date_ui; off[i].l_uf = chutable[i]; tmp = chuc->codetimes[i].tv_sec + JAN_1970; TVUTOTSF(chuc->codetimes[i].tv_usec, tmp2); M_SUB(off[i].l_ui, off[i].l_uf, tmp, tmp2); } if (!pp->sloppyclockflag) { u_short ord[NCHUCHARS]; /* * In here we assume the clock has adequate bits * to take timestamps with reasonable accuracy. * Note that the time stamps may contain errors * for a couple of reasons. Timing is actually * referenced to the start bit in each character * in the time code. If this is obscured by static * you can still get a valid character but have the * timestamp offset by +-1.5 ms. Also, we may suffer * from interrupt delays if the interrupt is being * held off when the character arrives. Note the * latter error is always in the form of a delay. * * After fiddling I arrived at the following scheme. * We sort the times into order by offset. We then * drop the most positive 2 offset values (which may * correspond to a character arriving early due to * static) and the most negative 4 (which may correspond * to delayed characters, either from static or from * interrupt latency). We then take the mean of the * remaining 4 offsets as our estimate. */ /* * Set up the order array. */ for (i = 0; i < NCHUCHARS; i++) ord[i] = (u_short)i; /* * Sort them into order. Reuse variables with abandon. */ for (tmp = 0; tmp < (NCHUCHARS-1); tmp++) { for (i = (int)tmp+1; i < NCHUCHARS; i++) { if (!L_ISGEQ(&off[ord[i]], &off[ord[tmp]])) { date_ui = (u_long)ord[i]; ord[i] = ord[tmp]; ord[tmp] = (u_short)date_ui; } } } /* * Done the sort. We drop 0, 1, 2 and 3 at the negative * end, and 8 and 9 at the positive. Take the sum of * 4, 5, 6 and 7. */ date_ui = off[ord[4]].l_ui; tmp = off[ord[4]].l_uf; for (i = 5; i <= 7; i++) M_ADD(date_ui, tmp, off[ord[i]].l_ui, off[ord[i]].l_uf); /* * Round properly, then right shift two bits for the * divide by four. */ if (tmp & 0x2) M_ADDUF(date_ui, tmp, 0x4); M_RSHIFT(date_ui, tmp); M_RSHIFT(date_ui, tmp); } else { /* * Here is a *big* problem. On a machine where the * low order bit in the clock is on the order of half * a millisecond or more we don't really have enough * precision to make intelligent choices about which * samples might be in error and which aren't. More * than this, in the case of error free data we can * pick up a few bits of precision by taking the mean * of the whole bunch. This is what we do. The problem * comes when it comes time to divide the 64 bit sum of * the 10 samples by 10, a procedure which really sucks. * Oh, well, grin and bear it. Compute the sum first. */ date_ui = 0; tmp = 0; for (i = 0; i < NCHUCHARS; i++) M_ADD(date_ui, tmp, off[i].l_ui, off[i].l_uf); if (M_ISNEG(date_ui, tmp)) isneg = 1; else isneg = 0; /* * Here is a multiply-by-0.1 optimization that should apply * just about everywhere. If the magnitude of the sum * is less than 9 we don't have to worry about overflow * out of a 64 bit product, even after rounding. */ if (date_ui < 9 || date_ui > 0xfffffff7) { register u_long prod_ui; register u_long prod_uf; prod_ui = prod_uf = 0; /* * This code knows the low order bit in 0.1 is zero */ for (i = 1; i < NZPOBITS; i++) { M_LSHIFT(date_ui, tmp); if (ZEROPTONE & (1<expect) { /* * This shouldn't actually happen, but might if a single * bit error occurred in the code which fooled us. * Throw away all previous data. */ up->expect = 0; up->haveoffset = 0; if (up->flags & CHUTIMERSET) { TIMER_DEQUEUE(&up->chutimer); up->flags &= ~CHUTIMERSET; } } up->offsets[i].l_ui = date_ui; up->offsets[i].l_uf = tmp; up->rectimes[i] = rbufp->recv_time; up->reftimes[i] = reftime; up->expect = i + 1; up->haveoffset |= (1 << i); if (up->expect >= NCHUCODES) { /* * Got a full second's worth. Dequeue timer and * process this. */ if (up->flags & CHUTIMERSET) { TIMER_DEQUEUE(&up->chutimer); up->flags &= ~CHUTIMERSET; } chu_process(up); } else if (!(up->flags & CHUTIMERSET)) { /* * Try to take an interrupt sometime after the * 42 second mark (leaves an extra 2 seconds for * slop). Round it up to an even multiple of * 4 seconds. */ up->chutimer.event_time = current_time + (u_long)(10 - i) + (1<chutimer.event_time &= ~((1<chutimer); up->flags |= CHUTIMERSET; } } /* * chu_timeout - process a timeout event */ static void chu_timeout(fakepeer) struct peer *fakepeer; { /* * If we got here it means we received some time codes * but didn't get the one which should have arrived on * the 39th second. Process what we have. */ ((struct chuunit *)fakepeer)->flags &= ~CHUTIMERSET; chu_process((struct chuunit *)fakepeer); } /* * chu_process - process the raw offset estimates we have and pass * the results on to the NTP clock filters. */ static void chu_process(up) register struct chuunit *up; { struct peer *peer; struct refclockproc *pp; int i; s_fp bestoff; s_fp tmpoff; u_fp dispersion; int imax; /* * The most positive offset. */ peer = up->peer; pp = peer->procptr; imax = NCHUCODES; for (i = 0; i < NCHUCODES; i++) if (up->haveoffset & (1<offsets[i], &up->offsets[imax])) imax = i; /* * The most positive estimate is our best bet. Go through * the list again computing the dispersion. */ bestoff = LFPTOFP(&up->offsets[imax]); dispersion = 0; for (i = 0; i < NCHUCODES; i++) { if (up->haveoffset & (1<offsets[i]); dispersion += (bestoff - tmpoff); } else { dispersion += CHUDELAYPENALTY; } } pp->lasttime = current_time; up->pollcnt = 2; record_clock_stats(&peer->srcadr, pp->a_lastcode); refclock_receive(peer, &up->offsets[imax], 0, dispersion, &up->rectimes[imax], &up->rectimes[imax], pp->leap); /* * Zero out unit for next code series */ up->haveoffset = 0; up->expect = 0; refclock_report(peer, CEVNT_NOMINAL); } /* * chu_poll - called by the transmit procedure */ static void chu_poll(unit, peer) int unit; struct peer *peer; { register struct chuunit *up; struct refclockproc *pp; pp = peer->procptr; up = (struct chuunit *)pp->unitptr; if (up->pollcnt == 0) refclock_report(peer, CEVNT_TIMEOUT); else up->pollcnt--; } #else /* not (REFCLOCK && CHUCLK) */ int refclock_chu_bs; #endif /* not (REFCLOCK && CHUCLK) */