xref: /freebsd/contrib/ntp/ntpd/refclock_shm.c (revision 63a938566d524836885917d95bd491aa4400b181)
1 /*
2  * refclock_shm - clock driver for utc via shared memory
3  * - under construction -
4  * To add new modes: Extend or union the shmTime-struct. Do not
5  * extend/shrink size, because otherwise existing implementations
6  * will specify wrong size of shared memory-segment
7  * PB 18.3.97
8  */
9 
10 #ifdef HAVE_CONFIG_H
11 # include <config.h>
12 #endif
13 
14 #include "ntp_types.h"
15 
16 #if defined(REFCLOCK) && defined(CLOCK_SHM)
17 
18 #include "ntpd.h"
19 #undef fileno
20 #include "ntp_io.h"
21 #undef fileno
22 #include "ntp_refclock.h"
23 #undef fileno
24 #include "timespecops.h"
25 #undef fileno
26 #include "ntp_stdlib.h"
27 #include "ntp_assert.h"
28 
29 #undef fileno
30 #include <ctype.h>
31 #undef fileno
32 
33 #ifndef SYS_WINNT
34 # include <sys/ipc.h>
35 # include <sys/shm.h>
36 # include <assert.h>
37 # include <unistd.h>
38 # include <stdio.h>
39 #endif
40 
41 #ifdef HAVE_STDATOMIC_H
42 # include <stdatomic.h>
43 #endif /* HAVE_STDATOMIC_H */
44 
45 /*
46  * This driver supports a reference clock attached thru shared memory
47  */
48 
49 /*
50  * SHM interface definitions
51  */
52 #define PRECISION       (-1)    /* precision assumed (0.5 s) */
53 #define REFID           "SHM"   /* reference ID */
54 #define DESCRIPTION     "SHM/Shared memory interface"
55 
56 #define NSAMPLES        3       /* stages of median filter */
57 
58 /*
59  * Mode flags
60  */
61 #define SHM_MODE_PRIVATE 0x0001
62 
63 /*
64  * Function prototypes
65  */
66 static  int     shm_start       (int unit, struct peer *peer);
67 static  void    shm_shutdown    (int unit, struct peer *peer);
68 static  void    shm_poll        (int unit, struct peer *peer);
69 static  void    shm_timer       (int unit, struct peer *peer);
70 static	void	shm_clockstats  (int unit, struct peer *peer);
71 static	void	shm_control	(int unit, const struct refclockstat * in_st,
72 				 struct refclockstat * out_st, struct peer *peer);
73 
74 /*
75  * Transfer vector
76  */
77 struct  refclock refclock_shm = {
78 	shm_start,              /* start up driver */
79 	shm_shutdown,           /* shut down driver */
80 	shm_poll,		/* transmit poll message */
81 	shm_control,		/* control settings */
82 	noentry,		/* not used: init */
83 	noentry,		/* not used: buginfo */
84 	shm_timer,              /* once per second */
85 };
86 
87 struct shmTime {
88 	int    mode; /* 0 - if valid is set:
89 		      *       use values,
90 		      *       clear valid
91 		      * 1 - if valid is set:
92 		      *       if count before and after read of values is equal,
93 		      *         use values
94 		      *       clear valid
95 		      */
96 	volatile int    count;
97 	time_t		clockTimeStampSec;
98 	int		clockTimeStampUSec;
99 	time_t		receiveTimeStampSec;
100 	int		receiveTimeStampUSec;
101 	int		leap;
102 	int		precision;
103 	int		nsamples;
104 	volatile int    valid;
105 	unsigned	clockTimeStampNSec;	/* Unsigned ns timestamps */
106 	unsigned	receiveTimeStampNSec;	/* Unsigned ns timestamps */
107 	int		dummy[8];
108 };
109 
110 struct shmunit {
111 	struct shmTime *shm;	/* pointer to shared memory segment */
112 	int forall;		/* access for all UIDs?	*/
113 
114 	/* debugging/monitoring counters - reset when printed */
115 	int ticks;		/* number of attempts to read data*/
116 	int good;		/* number of valid samples */
117 	int notready;		/* number of peeks without data ready */
118 	int bad;		/* number of invalid samples */
119 	int clash;		/* number of access clashes while reading */
120 
121 	time_t max_delta;	/* difference limit */
122 	time_t max_delay;	/* age/stale limit */
123 };
124 
125 
126 static struct shmTime*
127 getShmTime(
128 	int unit,
129 	int/*BOOL*/ forall
130 	)
131 {
132 	struct shmTime *p = NULL;
133 
134 #ifndef SYS_WINNT
135 
136 	int shmid;
137 
138 	/* 0x4e545030 is NTP0.
139 	 * Big units will give non-ascii but that's OK
140 	 * as long as everybody does it the same way.
141 	 */
142 	shmid=shmget(0x4e545030 + unit, sizeof (struct shmTime),
143 		      IPC_CREAT | (forall ? 0666 : 0600));
144 	if (shmid == -1) { /* error */
145 		msyslog(LOG_ERR, "SHM shmget (unit %d): %m", unit);
146 		return NULL;
147 	}
148 	p = (struct shmTime *)shmat (shmid, 0, 0);
149 	if (p == (struct shmTime *)-1) { /* error */
150 		msyslog(LOG_ERR, "SHM shmat (unit %d): %m", unit);
151 		return NULL;
152 	}
153 
154 	return p;
155 #else
156 
157 	static const char * nspref[2] = { "Local", "Global" };
158 	char buf[20];
159 	LPSECURITY_ATTRIBUTES psec = 0;
160 	HANDLE shmid = 0;
161 	SECURITY_DESCRIPTOR sd;
162 	SECURITY_ATTRIBUTES sa;
163 	unsigned int numch;
164 
165 	numch = snprintf(buf, sizeof(buf), "%s\\NTP%d",
166 			 nspref[forall != 0], (unit & 0xFF));
167 	if (numch >= sizeof(buf)) {
168 		msyslog(LOG_ERR, "SHM name too long (unit %d)", unit);
169 		return NULL;
170 	}
171 	if (forall) { /* world access */
172 		if (!InitializeSecurityDescriptor(&sd, SECURITY_DESCRIPTOR_REVISION)) {
173 			msyslog(LOG_ERR,"SHM InitializeSecurityDescriptor (unit %d): %m", unit);
174 			return NULL;
175 		}
176 		if (!SetSecurityDescriptorDacl(&sd, TRUE, NULL, FALSE)) {
177 			msyslog(LOG_ERR, "SHM SetSecurityDescriptorDacl (unit %d): %m", unit);
178 			return NULL;
179 		}
180 		sa.nLength = sizeof(SECURITY_ATTRIBUTES);
181 		sa.lpSecurityDescriptor = &sd;
182 		sa.bInheritHandle = FALSE;
183 		psec = &sa;
184 	}
185 	shmid = CreateFileMapping ((HANDLE)0xffffffff, psec, PAGE_READWRITE,
186 				   0, sizeof (struct shmTime), buf);
187 	if (shmid == NULL) { /*error*/
188 		char buf[1000];
189 		FormatMessage (FORMAT_MESSAGE_FROM_SYSTEM,
190 			       0, GetLastError (), 0, buf, sizeof (buf), 0);
191 		msyslog(LOG_ERR, "SHM CreateFileMapping (unit %d): %s", unit, buf);
192 		return NULL;
193 	}
194 	p = (struct shmTime *)MapViewOfFile(shmid, FILE_MAP_WRITE, 0, 0,
195 					    sizeof (struct shmTime));
196 	if (p == NULL) { /*error*/
197 		char buf[1000];
198 		FormatMessage (FORMAT_MESSAGE_FROM_SYSTEM,
199 			       0, GetLastError (), 0, buf, sizeof (buf), 0);
200 		msyslog(LOG_ERR,"SHM MapViewOfFile (unit %d): %s", unit, buf);
201 		return NULL;
202 	}
203 
204 	return p;
205 #endif
206 
207 	/* NOTREACHED */
208 	ENSURE(!"getShmTime(): Not reached.");
209 }
210 
211 
212 /*
213  * shm_start - attach to shared memory
214  */
215 static int
216 shm_start(
217 	int unit,
218 	struct peer *peer
219 	)
220 {
221 	struct refclockproc * const pp = peer->procptr;
222 	struct shmunit *      const up = emalloc_zero(sizeof(*up));
223 
224 	pp->io.clock_recv = noentry;
225 	pp->io.srcclock = peer;
226 	pp->io.datalen = 0;
227 	pp->io.fd = -1;
228 
229 	up->forall = (unit >= 2) && !(peer->ttl & SHM_MODE_PRIVATE);
230 
231 	up->shm = getShmTime(unit, up->forall);
232 
233 	/*
234 	 * Initialize miscellaneous peer variables
235 	 */
236 	memcpy((char *)&pp->refid, REFID, 4);
237 	if (up->shm != 0) {
238 		pp->unitptr = up;
239 		up->shm->precision = PRECISION;
240 		peer->precision = up->shm->precision;
241 		up->shm->valid = 0;
242 		up->shm->nsamples = NSAMPLES;
243 		pp->clockdesc = DESCRIPTION;
244 		/* items to be changed later in 'shm_control()': */
245 		up->max_delay = 5;
246 		up->max_delta = 4*3600;
247 		return 1;
248 	} else {
249 		free(up);
250 		pp->unitptr = NULL;
251 		return 0;
252 	}
253 }
254 
255 
256 /*
257  * shm_control - configure flag1/time2 params
258  *
259  * These are not yet available during 'shm_start', so we have to do any
260  * pre-computations we want to avoid during regular poll/timer callbacks
261  * in this callback.
262  */
263 static void
264 shm_control(
265 	int                         unit,
266 	const struct refclockstat * in_st,
267 	struct refclockstat       * out_st,
268 	struct peer               * peer
269 	)
270 {
271 	struct refclockproc * const pp = peer->procptr;
272 	struct shmunit *      const up = pp->unitptr;
273 
274 	UNUSED_ARG(unit);
275 	UNUSED_ARG(in_st);
276 	UNUSED_ARG(out_st);
277 	if (NULL == up)
278 		return;
279 	if (pp->sloppyclockflag & CLK_FLAG1)
280 		up->max_delta = 0;
281 	else if (pp->fudgetime2 < 1. || pp->fudgetime2 > 86400.)
282 		up->max_delta = 4*3600;
283 	else
284 		up->max_delta = (time_t)floor(pp->fudgetime2 + 0.5);
285 }
286 
287 
288 /*
289  * shm_shutdown - shut down the clock
290  */
291 static void
292 shm_shutdown(
293 	int unit,
294 	struct peer *peer
295 	)
296 {
297 	struct refclockproc * const pp = peer->procptr;
298 	struct shmunit *      const up = pp->unitptr;
299 
300 	UNUSED_ARG(unit);
301 	if (NULL == up)
302 		return;
303 #ifndef SYS_WINNT
304 
305 	/* HMS: shmdt() wants char* or const void * */
306 	(void)shmdt((char *)up->shm);
307 
308 #else
309 
310 	UnmapViewOfFile(up->shm);
311 
312 #endif
313 	free(up);
314 }
315 
316 
317 /*
318  * shm_poll - called by the transmit procedure
319  */
320 static void
321 shm_poll(
322 	int unit,
323 	struct peer *peer
324 	)
325 {
326 	struct refclockproc * const pp = peer->procptr;
327 	struct shmunit *      const up = pp->unitptr;
328 	int major_error;
329 
330 	pp->polls++;
331 
332 	/* get dominant reason if we have no samples at all */
333 	major_error = max(up->notready, up->bad);
334 	major_error = max(major_error, up->clash);
335 
336         /*
337          * Process median filter samples. If none received, see what
338          * happened, tell the core and keep going.
339          */
340         if (pp->coderecv != pp->codeproc) {
341 		/* have some samples, everything OK */
342 		pp->lastref = pp->lastrec;
343 		refclock_receive(peer);
344 	} else if (NULL == up->shm) { /* is this possible at all? */
345 		/* we're out of business without SHM access */
346 		refclock_report(peer, CEVNT_FAULT);
347 	} else if (major_error == up->clash) {
348 		/* too many collisions is like a bad signal */
349                 refclock_report(peer, CEVNT_PROP);
350 	} else if (major_error == up->bad) {
351 		/* too much stale/bad/garbled data */
352                 refclock_report(peer, CEVNT_BADREPLY);
353 	} else {
354 		/* in any other case assume it's just a timeout */
355                 refclock_report(peer, CEVNT_TIMEOUT);
356         }
357 	/* shm_clockstats() clears the tallies, so it must be last... */
358 	shm_clockstats(unit, peer);
359 }
360 
361 
362 enum segstat_t {
363     OK, NO_SEGMENT, NOT_READY, BAD_MODE, CLASH
364 };
365 
366 struct shm_stat_t {
367     int status;
368     int mode;
369     struct timespec tvc, tvr, tvt;
370     int precision;
371     int leap;
372 };
373 
374 static inline void memory_barrier(void)
375 {
376 #ifdef HAVE_ATOMIC_THREAD_FENCE
377     atomic_thread_fence(memory_order_seq_cst);
378 #endif /* HAVE_ATOMIC_THREAD_FENCE */
379 }
380 
381 static enum segstat_t shm_query(volatile struct shmTime *shm_in, struct shm_stat_t *shm_stat)
382 /* try to grab a sample from the specified SHM segment */
383 {
384     struct shmTime shmcopy;
385     volatile struct shmTime *shm = shm_in;
386     volatile int cnt;
387 
388     unsigned int cns_new, rns_new;
389 
390     /*
391      * This is the main routine. It snatches the time from the shm
392      * board and tacks on a local timestamp.
393      */
394     if (shm == NULL) {
395 	shm_stat->status = NO_SEGMENT;
396 	return NO_SEGMENT;
397     }
398 
399     /*@-type@*//* splint is confused about struct timespec */
400     shm_stat->tvc.tv_sec = shm_stat->tvc.tv_nsec = 0;
401     {
402 	time_t now;
403 
404 	time(&now);
405 	shm_stat->tvc.tv_sec = now;
406     }
407 
408     /* relying on word access to be atomic here */
409     if (shm->valid == 0) {
410 	shm_stat->status = NOT_READY;
411 	return NOT_READY;
412     }
413 
414     cnt = shm->count;
415 
416     /*
417      * This is proof against concurrency issues if either
418      * (a) the memory_barrier() call works on this host, or
419      * (b) memset compiles to an uninterruptible single-instruction bitblt.
420      */
421     memory_barrier();
422     memcpy(&shmcopy, (void*)(uintptr_t)shm, sizeof(struct shmTime));
423     shm->valid = 0;
424     memory_barrier();
425 
426     /*
427      * Clash detection in case neither (a) nor (b) was true.
428      * Not supported in mode 0, and word access to the count field
429      * must be atomic for this to work.
430      */
431     if (shmcopy.mode > 0 && cnt != shm->count) {
432 	shm_stat->status = CLASH;
433 	return shm_stat->status;
434     }
435 
436     shm_stat->status = OK;
437     shm_stat->mode = shmcopy.mode;
438 
439     switch (shmcopy.mode) {
440     case 0:
441 	shm_stat->tvr.tv_sec	= shmcopy.receiveTimeStampSec;
442 	shm_stat->tvr.tv_nsec	= shmcopy.receiveTimeStampUSec * 1000;
443 	rns_new		= shmcopy.receiveTimeStampNSec;
444 	shm_stat->tvt.tv_sec	= shmcopy.clockTimeStampSec;
445 	shm_stat->tvt.tv_nsec	= shmcopy.clockTimeStampUSec * 1000;
446 	cns_new		= shmcopy.clockTimeStampNSec;
447 
448 	/* Since the following comparisons are between unsigned
449 	** variables they are always well defined, and any
450 	** (signed) underflow will turn into very large unsigned
451 	** values, well above the 1000 cutoff.
452 	**
453 	** Note: The usecs *must* be a *truncated*
454 	** representation of the nsecs. This code will fail for
455 	** *rounded* usecs, and the logic to deal with
456 	** wrap-arounds in the presence of rounded values is
457 	** much more convoluted.
458 	*/
459 	if (   ((cns_new - (unsigned)shm_stat->tvt.tv_nsec) < 1000)
460 	       && ((rns_new - (unsigned)shm_stat->tvr.tv_nsec) < 1000)) {
461 	    shm_stat->tvt.tv_nsec = cns_new;
462 	    shm_stat->tvr.tv_nsec = rns_new;
463 	}
464 	/* At this point shm_stat->tvr and shm_stat->tvt contain valid ns-level
465 	** timestamps, possibly generated by extending the old
466 	** us-level timestamps
467 	*/
468 	break;
469 
470     case 1:
471 
472 	shm_stat->tvr.tv_sec	= shmcopy.receiveTimeStampSec;
473 	shm_stat->tvr.tv_nsec	= shmcopy.receiveTimeStampUSec * 1000;
474 	rns_new		= shmcopy.receiveTimeStampNSec;
475 	shm_stat->tvt.tv_sec	= shmcopy.clockTimeStampSec;
476 	shm_stat->tvt.tv_nsec	= shmcopy.clockTimeStampUSec * 1000;
477 	cns_new		= shmcopy.clockTimeStampNSec;
478 
479 	/* See the case above for an explanation of the
480 	** following test.
481 	*/
482 	if (   ((cns_new - (unsigned)shm_stat->tvt.tv_nsec) < 1000)
483 	       && ((rns_new - (unsigned)shm_stat->tvr.tv_nsec) < 1000)) {
484 	    shm_stat->tvt.tv_nsec = cns_new;
485 	    shm_stat->tvr.tv_nsec = rns_new;
486 	}
487 	/* At this point shm_stat->tvr and shm_stat->tvt contains valid ns-level
488 	** timestamps, possibly generated by extending the old
489 	** us-level timestamps
490 	*/
491 	break;
492 
493     default:
494 	shm_stat->status = BAD_MODE;
495 	break;
496     }
497     /*@-type@*/
498 
499     /*
500      * leap field is not a leap offset but a leap notification code.
501      * The values are magic numbers used by NTP and set by GPSD, if at all, in
502      * the subframe code.
503      */
504     shm_stat->leap = shmcopy.leap;
505     shm_stat->precision = shmcopy.precision;
506 
507     return shm_stat->status;
508 }
509 
510 /*
511  * shm_timer - called once every second.
512  *
513  * This tries to grab a sample from the SHM segment, filtering bad ones
514  */
515 static void
516 shm_timer(
517 	int unit,
518 	struct peer *peer
519 	)
520 {
521 	struct refclockproc * const pp = peer->procptr;
522 	struct shmunit *      const up = pp->unitptr;
523 
524 	volatile struct shmTime *shm;
525 
526 	l_fp tsrcv;
527 	l_fp tsref;
528 	int c;
529 
530 	/* for formatting 'a_lastcode': */
531 	struct calendar cd;
532 	time_t tt;
533 	vint64 ts;
534 
535 	enum segstat_t status;
536 	struct shm_stat_t shm_stat;
537 
538 	up->ticks++;
539 	if ((shm = up->shm) == NULL) {
540 		/* try to map again - this may succeed if meanwhile some-
541 		body has ipcrm'ed the old (unaccessible) shared mem segment */
542 		shm = up->shm = getShmTime(unit, up->forall);
543 		if (shm == NULL) {
544 			DPRINTF(1, ("%s: no SHM segment\n",
545 				    refnumtoa(&peer->srcadr)));
546 			return;
547 		}
548 	}
549 
550 	/* query the segment, atomically */
551 	status = shm_query(shm, &shm_stat);
552 
553 	switch (status) {
554 	case OK:
555 	    DPRINTF(2, ("%s: SHM type %d sample\n",
556 			refnumtoa(&peer->srcadr), shm_stat.mode));
557 	    break;
558 	case NO_SEGMENT:
559 	    /* should never happen, but is harmless */
560 	    return;
561 	case NOT_READY:
562 	    DPRINTF(1, ("%s: SHM not ready\n",refnumtoa(&peer->srcadr)));
563 	    up->notready++;
564 	    return;
565 	case BAD_MODE:
566 	    DPRINTF(1, ("%s: SHM type blooper, mode=%d\n",
567 			refnumtoa(&peer->srcadr), shm->mode));
568 	    up->bad++;
569 	    msyslog (LOG_ERR, "SHM: bad mode found in shared memory: %d",
570 		     shm->mode);
571 	    return;
572 	case CLASH:
573 	    DPRINTF(1, ("%s: type 1 access clash\n",
574 			refnumtoa(&peer->srcadr)));
575 	    msyslog (LOG_NOTICE, "SHM: access clash in shared memory");
576 	    up->clash++;
577 	    return;
578 	default:
579 	    DPRINTF(1, ("%s: internal error, unknown SHM fetch status\n",
580 			refnumtoa(&peer->srcadr)));
581 	    msyslog (LOG_NOTICE, "internal error, unknown SHM fetch status");
582 	    up->bad++;
583 	    return;
584 	}
585 
586 
587 	/* format the last time code in human-readable form into
588 	 * 'pp->a_lastcode'. Someone claimed: "NetBSD has incompatible
589 	 * tv_sec". I can't find a base for this claim, but we can work
590 	 * around that potential problem. BTW, simply casting a pointer
591 	 * is a receipe for disaster on some architectures.
592 	 */
593 	tt = (time_t)shm_stat.tvt.tv_sec;
594 	ts = time_to_vint64(&tt);
595 	ntpcal_time_to_date(&cd, &ts);
596 
597 	/* add ntpq -c cv timecode in ISO 8601 format */
598 	c = snprintf(pp->a_lastcode, sizeof(pp->a_lastcode),
599 		     "%04u-%02u-%02uT%02u:%02u:%02u.%09ldZ",
600 		     cd.year, cd.month, cd.monthday,
601 		     cd.hour, cd.minute, cd.second,
602 		     (long)shm_stat.tvt.tv_nsec);
603 	pp->lencode = (c > 0 && (size_t)c < sizeof(pp->a_lastcode)) ? c : 0;
604 
605 	/* check 1: age control of local time stamp */
606 	tt = shm_stat.tvc.tv_sec - shm_stat.tvr.tv_sec;
607 	if (tt < 0 || tt > up->max_delay) {
608 		DPRINTF(1, ("%s:SHM stale/bad receive time, delay=%llds\n",
609 			    refnumtoa(&peer->srcadr), (long long)tt));
610 		up->bad++;
611 		msyslog (LOG_ERR, "SHM: stale/bad receive time, delay=%llds",
612 			 (long long)tt);
613 		return;
614 	}
615 
616 	/* check 2: delta check */
617 	tt = shm_stat.tvr.tv_sec - shm_stat.tvt.tv_sec - (shm_stat.tvr.tv_nsec < shm_stat.tvt.tv_nsec);
618 	if (tt < 0)
619 		tt = -tt;
620 	if (up->max_delta > 0 && tt > up->max_delta) {
621 		DPRINTF(1, ("%s: SHM diff limit exceeded, delta=%llds\n",
622 			    refnumtoa(&peer->srcadr), (long long)tt));
623 		up->bad++;
624 		msyslog (LOG_ERR, "SHM: difference limit exceeded, delta=%llds\n",
625 			 (long long)tt);
626 		return;
627 	}
628 
629 	/* if we really made it to this point... we're winners! */
630 	DPRINTF(2, ("%s: SHM feeding data\n",
631 		    refnumtoa(&peer->srcadr)));
632 	tsrcv = tspec_stamp_to_lfp(shm_stat.tvr);
633 	tsref = tspec_stamp_to_lfp(shm_stat.tvt);
634 	pp->leap = shm_stat.leap;
635 	peer->precision = shm_stat.precision;
636 	refclock_process_offset(pp, tsref, tsrcv, pp->fudgetime1);
637 	up->good++;
638 }
639 
640 /*
641  * shm_clockstats - dump and reset counters
642  */
643 static void shm_clockstats(
644 	int unit,
645 	struct peer *peer
646 	)
647 {
648 	struct refclockproc * const pp = peer->procptr;
649 	struct shmunit *      const up = pp->unitptr;
650 
651 	UNUSED_ARG(unit);
652 	if (pp->sloppyclockflag & CLK_FLAG4) {
653 		mprintf_clock_stats(
654 			&peer->srcadr, "%3d %3d %3d %3d %3d",
655 			up->ticks, up->good, up->notready,
656 			up->bad, up->clash);
657 	}
658 	up->ticks = up->good = up->notready = up->bad = up->clash = 0;
659 }
660 
661 #else
662 NONEMPTY_TRANSLATION_UNIT
663 #endif /* REFCLOCK */
664