refclock_shm.c revision 1.1.1.6 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 volatile struct shmTime shmcopy, *shm = shm_in;
385 volatile int cnt;
386
387 unsigned int cns_new, rns_new;
388
389 /*
390 * This is the main routine. It snatches the time from the shm
391 * board and tacks on a local timestamp.
392 */
393 if (shm == NULL) {
394 shm_stat->status = NO_SEGMENT;
395 return NO_SEGMENT;
396 }
397
398 /*@-type@*//* splint is confused about struct timespec */
399 shm_stat->tvc.tv_sec = shm_stat->tvc.tv_nsec = 0;
400 {
401 time_t now;
402
403 time(&now);
404 shm_stat->tvc.tv_sec = now;
405 }
406
407 /* relying on word access to be atomic here */
408 if (shm->valid == 0) {
409 shm_stat->status = NOT_READY;
410 return NOT_READY;
411 }
412
413 cnt = shm->count;
414
415 /*
416 * This is proof against concurrency issues if either
417 * (a) the memory_barrier() call works on this host, or
418 * (b) memset compiles to an uninterruptible single-instruction bitblt.
419 */
420 memory_barrier();
421 memcpy((void *)&shmcopy, (void *)shm, sizeof(struct shmTime));
422 shm->valid = 0;
423 memory_barrier();
424
425 /*
426 * Clash detection in case neither (a) nor (b) was true.
427 * Not supported in mode 0, and word access to the count field
428 * must be atomic for this to work.
429 */
430 if (shmcopy.mode > 0 && cnt != shm->count) {
431 shm_stat->status = CLASH;
432 return shm_stat->status;
433 }
434
435 shm_stat->status = OK;
436 shm_stat->mode = shmcopy.mode;
437
438 switch (shmcopy.mode) {
439 case 0:
440 shm_stat->tvr.tv_sec = shmcopy.receiveTimeStampSec;
441 shm_stat->tvr.tv_nsec = shmcopy.receiveTimeStampUSec * 1000;
442 rns_new = shmcopy.receiveTimeStampNSec;
443 shm_stat->tvt.tv_sec = shmcopy.clockTimeStampSec;
444 shm_stat->tvt.tv_nsec = shmcopy.clockTimeStampUSec * 1000;
445 cns_new = shmcopy.clockTimeStampNSec;
446
447 /* Since the following comparisons are between unsigned
448 ** variables they are always well defined, and any
449 ** (signed) underflow will turn into very large unsigned
450 ** values, well above the 1000 cutoff.
451 **
452 ** Note: The usecs *must* be a *truncated*
453 ** representation of the nsecs. This code will fail for
454 ** *rounded* usecs, and the logic to deal with
455 ** wrap-arounds in the presence of rounded values is
456 ** much more convoluted.
457 */
458 if ( ((cns_new - (unsigned)shm_stat->tvt.tv_nsec) < 1000)
459 && ((rns_new - (unsigned)shm_stat->tvr.tv_nsec) < 1000)) {
460 shm_stat->tvt.tv_nsec = cns_new;
461 shm_stat->tvr.tv_nsec = rns_new;
462 }
463 /* At this point shm_stat->tvr and shm_stat->tvt contain valid ns-level
464 ** timestamps, possibly generated by extending the old
465 ** us-level timestamps
466 */
467 break;
468
469 case 1:
470
471 shm_stat->tvr.tv_sec = shmcopy.receiveTimeStampSec;
472 shm_stat->tvr.tv_nsec = shmcopy.receiveTimeStampUSec * 1000;
473 rns_new = shmcopy.receiveTimeStampNSec;
474 shm_stat->tvt.tv_sec = shmcopy.clockTimeStampSec;
475 shm_stat->tvt.tv_nsec = shmcopy.clockTimeStampUSec * 1000;
476 cns_new = shmcopy.clockTimeStampNSec;
477
478 /* See the case above for an explanation of the
479 ** following test.
480 */
481 if ( ((cns_new - (unsigned)shm_stat->tvt.tv_nsec) < 1000)
482 && ((rns_new - (unsigned)shm_stat->tvr.tv_nsec) < 1000)) {
483 shm_stat->tvt.tv_nsec = cns_new;
484 shm_stat->tvr.tv_nsec = rns_new;
485 }
486 /* At this point shm_stat->tvr and shm_stat->tvt contains valid ns-level
487 ** timestamps, possibly generated by extending the old
488 ** us-level timestamps
489 */
490 break;
491
492 default:
493 shm_stat->status = BAD_MODE;
494 break;
495 }
496 /*@-type@*/
497
498 /*
499 * leap field is not a leap offset but a leap notification code.
500 * The values are magic numbers used by NTP and set by GPSD, if at all, in
501 * the subframe code.
502 */
503 shm_stat->leap = shmcopy.leap;
504 shm_stat->precision = shmcopy.precision;
505
506 return shm_stat->status;
507 }
508
509 /*
510 * shm_timer - called once every second.
511 *
512 * This tries to grab a sample from the SHM segment, filtering bad ones
513 */
514 static void
515 shm_timer(
516 int unit,
517 struct peer *peer
518 )
519 {
520 struct refclockproc * const pp = peer->procptr;
521 struct shmunit * const up = pp->unitptr;
522
523 volatile struct shmTime *shm;
524
525 l_fp tsrcv;
526 l_fp tsref;
527 int c;
528
529 /* for formatting 'a_lastcode': */
530 struct calendar cd;
531 time_t tt;
532 vint64 ts;
533
534 enum segstat_t status;
535 struct shm_stat_t shm_stat;
536
537 up->ticks++;
538 if ((shm = up->shm) == NULL) {
539 /* try to map again - this may succeed if meanwhile some-
540 body has ipcrm'ed the old (unaccessible) shared mem segment */
541 shm = up->shm = getShmTime(unit, up->forall);
542 if (shm == NULL) {
543 DPRINTF(1, ("%s: no SHM segment\n",
544 refnumtoa(&peer->srcadr)));
545 return;
546 }
547 }
548
549 /* query the segment, atomically */
550 status = shm_query(shm, &shm_stat);
551
552 switch (status) {
553 case OK:
554 DPRINTF(2, ("%s: SHM type %d sample\n",
555 refnumtoa(&peer->srcadr), shm_stat.mode));
556 break;
557 case NO_SEGMENT:
558 /* should never happen, but is harmless */
559 return;
560 case NOT_READY:
561 DPRINTF(1, ("%s: SHM not ready\n",refnumtoa(&peer->srcadr)));
562 up->notready++;
563 return;
564 case BAD_MODE:
565 DPRINTF(1, ("%s: SHM type blooper, mode=%d\n",
566 refnumtoa(&peer->srcadr), shm->mode));
567 up->bad++;
568 msyslog (LOG_ERR, "SHM: bad mode found in shared memory: %d",
569 shm->mode);
570 return;
571 case CLASH:
572 DPRINTF(1, ("%s: type 1 access clash\n",
573 refnumtoa(&peer->srcadr)));
574 msyslog (LOG_NOTICE, "SHM: access clash in shared memory");
575 up->clash++;
576 return;
577 default:
578 DPRINTF(1, ("%s: internal error, unknown SHM fetch status\n",
579 refnumtoa(&peer->srcadr)));
580 msyslog (LOG_NOTICE, "internal error, unknown SHM fetch status");
581 up->bad++;
582 return;
583 }
584
585
586 /* format the last time code in human-readable form into
587 * 'pp->a_lastcode'. Someone claimed: "NetBSD has incompatible
588 * tv_sec". I can't find a base for this claim, but we can work
589 * around that potential problem. BTW, simply casting a pointer
590 * is a receipe for disaster on some architectures.
591 */
592 tt = (time_t)shm_stat.tvt.tv_sec;
593 ts = time_to_vint64(&tt);
594 ntpcal_time_to_date(&cd, &ts);
595
596 /* add ntpq -c cv timecode in ISO 8601 format */
597 c = snprintf(pp->a_lastcode, sizeof(pp->a_lastcode),
598 "%04u-%02u-%02uT%02u:%02u:%02u.%09ldZ",
599 cd.year, cd.month, cd.monthday,
600 cd.hour, cd.minute, cd.second,
601 (long)shm_stat.tvt.tv_nsec);
602 pp->lencode = (c < sizeof(pp->a_lastcode)) ? c : 0;
603
604 /* check 1: age control of local time stamp */
605 tt = shm_stat.tvc.tv_sec - shm_stat.tvr.tv_sec;
606 if (tt < 0 || tt > up->max_delay) {
607 DPRINTF(1, ("%s:SHM stale/bad receive time, delay=%llds\n",
608 refnumtoa(&peer->srcadr), (long long)tt));
609 up->bad++;
610 msyslog (LOG_ERR, "SHM: stale/bad receive time, delay=%llds",
611 (long long)tt);
612 return;
613 }
614
615 /* check 2: delta check */
616 tt = shm_stat.tvr.tv_sec - shm_stat.tvt.tv_sec - (shm_stat.tvr.tv_nsec < shm_stat.tvt.tv_nsec);
617 if (tt < 0)
618 tt = -tt;
619 if (up->max_delta > 0 && tt > up->max_delta) {
620 DPRINTF(1, ("%s: SHM diff limit exceeded, delta=%llds\n",
621 refnumtoa(&peer->srcadr), (long long)tt));
622 up->bad++;
623 msyslog (LOG_ERR, "SHM: difference limit exceeded, delta=%llds\n",
624 (long long)tt);
625 return;
626 }
627
628 /* if we really made it to this point... we're winners! */
629 DPRINTF(2, ("%s: SHM feeding data\n",
630 refnumtoa(&peer->srcadr)));
631 tsrcv = tspec_stamp_to_lfp(shm_stat.tvr);
632 tsref = tspec_stamp_to_lfp(shm_stat.tvt);
633 pp->leap = shm_stat.leap;
634 peer->precision = shm_stat.precision;
635 refclock_process_offset(pp, tsref, tsrcv, pp->fudgetime1);
636 up->good++;
637 }
638
639 /*
640 * shm_clockstats - dump and reset counters
641 */
642 static void shm_clockstats(
643 int unit,
644 struct peer *peer
645 )
646 {
647 struct refclockproc * const pp = peer->procptr;
648 struct shmunit * const up = pp->unitptr;
649
650 UNUSED_ARG(unit);
651 if (pp->sloppyclockflag & CLK_FLAG4) {
652 mprintf_clock_stats(
653 &peer->srcadr, "%3d %3d %3d %3d %3d",
654 up->ticks, up->good, up->notready,
655 up->bad, up->clash);
656 }
657 up->ticks = up->good = up->notready = up->bad = up->clash = 0;
658 }
659
660 #else
661 NONEMPTY_TRANSLATION_UNIT
662 #endif /* REFCLOCK */
663