systime.c revision 293896
1/*
2 * systime -- routines to fiddle a UNIX clock.
3 *
4 * ATTENTION: Get approval from Dave Mills on all changes to this file!
5 *
6 */
7#include <config.h>
8
9#include "ntp.h"
10#include "ntp_syslog.h"
11#include "ntp_stdlib.h"
12#include "ntp_random.h"
13#include "iosignal.h"
14#include "timevalops.h"
15#include "timespecops.h"
16#include "ntp_calendar.h"
17
18#ifdef HAVE_SYS_PARAM_H
19# include <sys/param.h>
20#endif
21#ifdef HAVE_UTMP_H
22# include <utmp.h>
23#endif /* HAVE_UTMP_H */
24#ifdef HAVE_UTMPX_H
25# include <utmpx.h>
26#endif /* HAVE_UTMPX_H */
27
28int	allow_panic = FALSE;		/* allow panic correction (-g) */
29int	enable_panic_check = TRUE;	/* Can we check allow_panic's state? */
30
31#ifndef USE_COMPILETIME_PIVOT
32# define USE_COMPILETIME_PIVOT 1
33#endif
34
35/*
36 * These routines (get_systime, step_systime, adj_systime) implement an
37 * interface between the system independent NTP clock and the Unix
38 * system clock in various architectures and operating systems. Time is
39 * a precious quantity in these routines and every effort is made to
40 * minimize errors by unbiased rounding and amortizing adjustment
41 * residues.
42 *
43 * In order to improve the apparent resolution, provide unbiased
44 * rounding and most importantly ensure that the readings cannot be
45 * predicted, the low-order unused portion of the time below the minimum
46 * time to read the clock is filled with an unbiased random fuzz.
47 *
48 * The sys_tick variable specifies the system clock tick interval in
49 * seconds, for stepping clocks, defined as those which return times
50 * less than MINSTEP greater than the previous reading. For systems that
51 * use a high-resolution counter such that each clock reading is always
52 * at least MINSTEP greater than the prior, sys_tick is the time to read
53 * the system clock.
54 *
55 * The sys_fuzz variable measures the minimum time to read the system
56 * clock, regardless of its precision.  When reading the system clock
57 * using get_systime() after sys_tick and sys_fuzz have been determined,
58 * ntpd ensures each unprocessed clock reading is no less than sys_fuzz
59 * later than the prior unprocessed reading, and then fuzzes the bits
60 * below sys_fuzz in the timestamp returned, ensuring each of its
61 * resulting readings is strictly later than the previous.
62 *
63 * When slewing the system clock using adj_systime() (with the kernel
64 * loop discipline unavailable or disabled), adjtime() offsets are
65 * quantized to sys_tick, if sys_tick is greater than sys_fuzz, which
66 * is to say if the OS presents a stepping clock.  Otherwise, offsets
67 * are quantized to the microsecond resolution of adjtime()'s timeval
68 * input.  The remaining correction sys_residual is carried into the
69 * next adjtime() and meanwhile is also factored into get_systime()
70 * readings.
71 */
72double	sys_tick = 0;		/* tick size or time to read (s) */
73double	sys_fuzz = 0;		/* min. time to read the clock (s) */
74long	sys_fuzz_nsec = 0;	/* min. time to read the clock (ns) */
75double	measured_tick;		/* non-overridable sys_tick (s) */
76double	sys_residual = 0;	/* adjustment residue (s) */
77int	trunc_os_clock;		/* sys_tick > measured_tick */
78time_stepped_callback	step_callback;
79
80#ifndef SIM
81/* perlinger@ntp.org: As 'get_sysime()' does it's own check for clock
82 * backstepping, this could probably become a local variable in
83 * 'get_systime()' and the cruft associated with communicating via a
84 * static value could be removed after the v4.2.8 release.
85 */
86static int lamport_violated;	/* clock was stepped back */
87#endif	/* !SIM */
88
89#ifdef DEBUG
90static int systime_init_done;
91# define DONE_SYSTIME_INIT()	systime_init_done = TRUE
92#else
93# define DONE_SYSTIME_INIT()	do {} while (FALSE)
94#endif
95
96#ifdef HAVE_SIGNALED_IO
97int using_sigio;
98#endif
99
100#ifdef SYS_WINNT
101CRITICAL_SECTION get_systime_cs;
102#endif
103
104
105void
106set_sys_fuzz(
107	double	fuzz_val
108	)
109{
110	sys_fuzz = fuzz_val;
111	INSIST(sys_fuzz >= 0);
112	INSIST(sys_fuzz <= 1.0);
113	sys_fuzz_nsec = (long)(sys_fuzz * 1e9 + 0.5);
114}
115
116
117void
118init_systime(void)
119{
120	INIT_GET_SYSTIME_CRITSEC();
121	INIT_WIN_PRECISE_TIME();
122	DONE_SYSTIME_INIT();
123}
124
125
126#ifndef SIM	/* ntpsim.c has get_systime() and friends for sim */
127
128static inline void
129get_ostime(
130	struct timespec *	tsp
131	)
132{
133	int	rc;
134	long	ticks;
135
136#if defined(HAVE_CLOCK_GETTIME)
137	rc = clock_gettime(CLOCK_REALTIME, tsp);
138#elif defined(HAVE_GETCLOCK)
139	rc = getclock(TIMEOFDAY, tsp);
140#else
141	struct timeval		tv;
142
143	rc = GETTIMEOFDAY(&tv, NULL);
144	tsp->tv_sec = tv.tv_sec;
145	tsp->tv_nsec = tv.tv_usec * 1000;
146#endif
147	if (rc < 0) {
148		msyslog(LOG_ERR, "read system clock failed: %m (%d)",
149			errno);
150		exit(1);
151	}
152
153	if (trunc_os_clock) {
154		ticks = (long)((tsp->tv_nsec * 1e-9) / sys_tick);
155		tsp->tv_nsec = (long)(ticks * 1e9 * sys_tick);
156	}
157}
158
159
160/*
161 * get_systime - return system time in NTP timestamp format.
162 */
163void
164get_systime(
165	l_fp *now		/* system time */
166	)
167{
168        static struct timespec  ts_last;        /* last sampled os time */
169	static struct timespec	ts_prev;	/* prior os time */
170	static l_fp		lfp_prev;	/* prior result */
171	static double		dfuzz_prev;	/* prior fuzz */
172	struct timespec ts;	/* seconds and nanoseconds */
173	struct timespec ts_min;	/* earliest permissible */
174	struct timespec ts_lam;	/* lamport fictional increment */
175	struct timespec ts_prev_log;	/* for msyslog only */
176	double	dfuzz;
177	double	ddelta;
178	l_fp	result;
179	l_fp	lfpfuzz;
180	l_fp	lfpdelta;
181
182	get_ostime(&ts);
183	DEBUG_REQUIRE(systime_init_done);
184	ENTER_GET_SYSTIME_CRITSEC();
185
186        /* First check if here was a Lamport violation, that is, two
187         * successive calls to 'get_ostime()' resulted in negative
188         * time difference. Use a few milliseconds of permissible
189         * tolerance -- being too sharp can hurt here. (This is intented
190         * for the Win32 target, where the HPC interpolation might
191         * introduce small steps backward. It should not be an issue on
192         * systems where get_ostime() results in a true syscall.)
193         */
194        if (cmp_tspec(add_tspec_ns(ts, 50000000), ts_last) < 0)
195                lamport_violated = 1;
196        ts_last = ts;
197
198	/*
199	 * After default_get_precision() has set a nonzero sys_fuzz,
200	 * ensure every reading of the OS clock advances by at least
201	 * sys_fuzz over the prior reading, thereby assuring each
202	 * fuzzed result is strictly later than the prior.  Limit the
203	 * necessary fiction to 1 second.
204	 */
205	if (!USING_SIGIO()) {
206		ts_min = add_tspec_ns(ts_prev, sys_fuzz_nsec);
207		if (cmp_tspec(ts, ts_min) < 0) {
208			ts_lam = sub_tspec(ts_min, ts);
209			if (ts_lam.tv_sec > 0 && !lamport_violated) {
210				msyslog(LOG_ERR,
211					"get_systime Lamport advance exceeds one second (%.9f)",
212					ts_lam.tv_sec +
213					    1e-9 * ts_lam.tv_nsec);
214				exit(1);
215			}
216			if (!lamport_violated)
217				ts = ts_min;
218		}
219		ts_prev_log = ts_prev;
220		ts_prev = ts;
221	} else {
222		/*
223		 * Quiet "ts_prev_log.tv_sec may be used uninitialized"
224		 * warning from x86 gcc 4.5.2.
225		 */
226		ZERO(ts_prev_log);
227	}
228
229	/* convert from timespec to l_fp fixed-point */
230	result = tspec_stamp_to_lfp(ts);
231
232	/*
233	 * Add in the fuzz.
234	 */
235	dfuzz = ntp_random() * 2. / FRAC * sys_fuzz;
236	DTOLFP(dfuzz, &lfpfuzz);
237	L_ADD(&result, &lfpfuzz);
238
239	/*
240	 * Ensure result is strictly greater than prior result (ignoring
241	 * sys_residual's effect for now) once sys_fuzz has been
242	 * determined.
243	 */
244	if (!USING_SIGIO()) {
245		if (!L_ISZERO(&lfp_prev) && !lamport_violated) {
246			if (!L_ISGTU(&result, &lfp_prev) &&
247			    sys_fuzz > 0.) {
248				msyslog(LOG_ERR, "ts_prev %s ts_min %s",
249					tspectoa(ts_prev_log),
250					tspectoa(ts_min));
251				msyslog(LOG_ERR, "ts %s", tspectoa(ts));
252				msyslog(LOG_ERR, "sys_fuzz %ld nsec, prior fuzz %.9f",
253					sys_fuzz_nsec, dfuzz_prev);
254				msyslog(LOG_ERR, "this fuzz %.9f",
255					dfuzz);
256				lfpdelta = lfp_prev;
257				L_SUB(&lfpdelta, &result);
258				LFPTOD(&lfpdelta, ddelta);
259				msyslog(LOG_ERR,
260					"prev get_systime 0x%x.%08x is %.9f later than 0x%x.%08x",
261					lfp_prev.l_ui, lfp_prev.l_uf,
262					ddelta, result.l_ui, result.l_uf);
263			}
264		}
265		lfp_prev = result;
266		dfuzz_prev = dfuzz;
267		if (lamport_violated)
268			lamport_violated = FALSE;
269	}
270	LEAVE_GET_SYSTIME_CRITSEC();
271	*now = result;
272}
273
274
275/*
276 * adj_systime - adjust system time by the argument.
277 */
278#if !defined SYS_WINNT
279int				/* 0 okay, 1 error */
280adj_systime(
281	double now		/* adjustment (s) */
282	)
283{
284	struct timeval adjtv;	/* new adjustment */
285	struct timeval oadjtv;	/* residual adjustment */
286	double	quant;		/* quantize to multiples of */
287	double	dtemp;
288	long	ticks;
289	int	isneg = 0;
290
291	/*
292	 * The Windows port adj_systime() depends on being called each
293	 * second even when there's no additional correction, to allow
294	 * emulation of adjtime() behavior on top of an API that simply
295	 * sets the current rate.  This POSIX implementation needs to
296	 * ignore invocations with zero correction, otherwise ongoing
297	 * EVNT_NSET adjtime() can be aborted by a tiny adjtime()
298	 * triggered by sys_residual.
299	 */
300	if (0. == now) {
301		if (enable_panic_check && allow_panic) {
302			msyslog(LOG_ERR, "adj_systime: allow_panic is TRUE!");
303			INSIST(!allow_panic);
304		}
305		return TRUE;
306	}
307
308	/*
309	 * Most Unix adjtime() implementations adjust the system clock
310	 * in microsecond quanta, but some adjust in 10-ms quanta. We
311	 * carefully round the adjustment to the nearest quantum, then
312	 * adjust in quanta and keep the residue for later.
313	 */
314	dtemp = now + sys_residual;
315	if (dtemp < 0) {
316		isneg = 1;
317		dtemp = -dtemp;
318	}
319	adjtv.tv_sec = (long)dtemp;
320	dtemp -= adjtv.tv_sec;
321	if (sys_tick > sys_fuzz)
322		quant = sys_tick;
323	else
324		quant = 1e-6;
325	ticks = (long)(dtemp / quant + .5);
326	adjtv.tv_usec = (long)(ticks * quant * 1e6);
327	dtemp -= adjtv.tv_usec / 1e6;
328	sys_residual = dtemp;
329
330	/*
331	 * Convert to signed seconds and microseconds for the Unix
332	 * adjtime() system call. Note we purposely lose the adjtime()
333	 * leftover.
334	 */
335	if (isneg) {
336		adjtv.tv_sec = -adjtv.tv_sec;
337		adjtv.tv_usec = -adjtv.tv_usec;
338		sys_residual = -sys_residual;
339	}
340	if (adjtv.tv_sec != 0 || adjtv.tv_usec != 0) {
341		if (adjtime(&adjtv, &oadjtv) < 0) {
342			msyslog(LOG_ERR, "adj_systime: %m");
343			if (enable_panic_check && allow_panic) {
344				msyslog(LOG_ERR, "adj_systime: allow_panic is TRUE!");
345			}
346			return FALSE;
347		}
348	}
349	if (enable_panic_check && allow_panic) {
350		msyslog(LOG_ERR, "adj_systime: allow_panic is TRUE!");
351	}
352	return TRUE;
353}
354#endif
355
356
357/*
358 * step_systime - step the system clock.
359 */
360
361int
362step_systime(
363	double step
364	)
365{
366	time_t pivot; /* for ntp era unfolding */
367	struct timeval timetv, tvlast, tvdiff;
368	struct timespec timets;
369	struct calendar jd;
370	l_fp fp_ofs, fp_sys; /* offset and target system time in FP */
371
372	/*
373	 * Get pivot time for NTP era unfolding. Since we don't step
374	 * very often, we can afford to do the whole calculation from
375	 * scratch. And we're not in the time-critical path yet.
376	 */
377#if SIZEOF_TIME_T > 4
378	/*
379	 * This code makes sure the resulting time stamp for the new
380	 * system time is in the 2^32 seconds starting at 1970-01-01,
381	 * 00:00:00 UTC.
382	 */
383	pivot = 0x80000000;
384#if USE_COMPILETIME_PIVOT
385	/*
386	 * Add the compile time minus 10 years to get a possible target
387	 * area of (compile time - 10 years) to (compile time + 126
388	 * years).  This should be sufficient for a given binary of
389	 * NTPD.
390	 */
391	if (ntpcal_get_build_date(&jd)) {
392		jd.year -= 10;
393		pivot += ntpcal_date_to_time(&jd);
394	} else {
395		msyslog(LOG_ERR,
396			"step-systime: assume 1970-01-01 as build date");
397	}
398#else
399	UNUSED_LOCAL(jd);
400#endif /* USE_COMPILETIME_PIVOT */
401#else
402	UNUSED_LOCAL(jd);
403	/* This makes sure the resulting time stamp is on or after
404	 * 1969-12-31/23:59:59 UTC and gives us additional two years,
405	 * from the change of NTP era in 2036 to the UNIX rollover in
406	 * 2038. (Minus one second, but that won't hurt.) We *really*
407	 * need a longer 'time_t' after that!  Or a different baseline,
408	 * but that would cause other serious trouble, too.
409	 */
410	pivot = 0x7FFFFFFF;
411#endif
412
413	/* get the complete jump distance as l_fp */
414	DTOLFP(sys_residual, &fp_sys);
415	DTOLFP(step,         &fp_ofs);
416	L_ADD(&fp_ofs, &fp_sys);
417
418	/* ---> time-critical path starts ---> */
419
420	/* get the current time as l_fp (without fuzz) and as struct timeval */
421	get_ostime(&timets);
422	fp_sys = tspec_stamp_to_lfp(timets);
423	tvlast.tv_sec = timets.tv_sec;
424	tvlast.tv_usec = (timets.tv_nsec + 500) / 1000;
425
426	/* get the target time as l_fp */
427	L_ADD(&fp_sys, &fp_ofs);
428
429	/* unfold the new system time */
430	timetv = lfp_stamp_to_tval(fp_sys, &pivot);
431
432	/* now set new system time */
433	if (ntp_set_tod(&timetv, NULL) != 0) {
434		msyslog(LOG_ERR, "step-systime: %m");
435		if (enable_panic_check && allow_panic) {
436			msyslog(LOG_ERR, "step_systime: allow_panic is TRUE!");
437		}
438		return FALSE;
439	}
440
441	/* <--- time-critical path ended with 'ntp_set_tod()' <--- */
442
443	sys_residual = 0;
444	lamport_violated = (step < 0);
445	if (step_callback)
446		(*step_callback)();
447
448#ifdef NEED_HPUX_ADJTIME
449	/*
450	 * CHECKME: is this correct when called by ntpdate?????
451	 */
452	_clear_adjtime();
453#endif
454
455	/*
456	 * FreeBSD, for example, has:
457	 * struct utmp {
458	 *	   char    ut_line[UT_LINESIZE];
459	 *	   char    ut_name[UT_NAMESIZE];
460	 *	   char    ut_host[UT_HOSTSIZE];
461	 *	   long    ut_time;
462	 * };
463	 * and appends line="|", name="date", host="", time for the OLD
464	 * and appends line="{", name="date", host="", time for the NEW // }
465	 * to _PATH_WTMP .
466	 *
467	 * Some OSes have utmp, some have utmpx.
468	 */
469
470	/*
471	 * Write old and new time entries in utmp and wtmp if step
472	 * adjustment is greater than one second.
473	 *
474	 * This might become even Uglier...
475	 */
476	tvdiff = abs_tval(sub_tval(timetv, tvlast));
477	if (tvdiff.tv_sec > 0) {
478#ifdef HAVE_UTMP_H
479		struct utmp ut;
480#endif
481#ifdef HAVE_UTMPX_H
482		struct utmpx utx;
483#endif
484
485#ifdef HAVE_UTMP_H
486		ZERO(ut);
487#endif
488#ifdef HAVE_UTMPX_H
489		ZERO(utx);
490#endif
491
492		/* UTMP */
493
494#ifdef UPDATE_UTMP
495# ifdef HAVE_PUTUTLINE
496#  ifndef _PATH_UTMP
497#   define _PATH_UTMP UTMP_FILE
498#  endif
499		utmpname(_PATH_UTMP);
500		ut.ut_type = OLD_TIME;
501		strlcpy(ut.ut_line, OTIME_MSG, sizeof(ut.ut_line));
502		ut.ut_time = tvlast.tv_sec;
503		setutent();
504		pututline(&ut);
505		ut.ut_type = NEW_TIME;
506		strlcpy(ut.ut_line, NTIME_MSG, sizeof(ut.ut_line));
507		ut.ut_time = timetv.tv_sec;
508		setutent();
509		pututline(&ut);
510		endutent();
511# else /* not HAVE_PUTUTLINE */
512# endif /* not HAVE_PUTUTLINE */
513#endif /* UPDATE_UTMP */
514
515		/* UTMPX */
516
517#ifdef UPDATE_UTMPX
518# ifdef HAVE_PUTUTXLINE
519		utx.ut_type = OLD_TIME;
520		strlcpy(utx.ut_line, OTIME_MSG, sizeof(utx.ut_line));
521		utx.ut_tv = tvlast;
522		setutxent();
523		pututxline(&utx);
524		utx.ut_type = NEW_TIME;
525		strlcpy(utx.ut_line, NTIME_MSG, sizeof(utx.ut_line));
526		utx.ut_tv = timetv;
527		setutxent();
528		pututxline(&utx);
529		endutxent();
530# else /* not HAVE_PUTUTXLINE */
531# endif /* not HAVE_PUTUTXLINE */
532#endif /* UPDATE_UTMPX */
533
534		/* WTMP */
535
536#ifdef UPDATE_WTMP
537# ifdef HAVE_PUTUTLINE
538#  ifndef _PATH_WTMP
539#   define _PATH_WTMP WTMP_FILE
540#  endif
541		utmpname(_PATH_WTMP);
542		ut.ut_type = OLD_TIME;
543		strlcpy(ut.ut_line, OTIME_MSG, sizeof(ut.ut_line));
544		ut.ut_time = tvlast.tv_sec;
545		setutent();
546		pututline(&ut);
547		ut.ut_type = NEW_TIME;
548		strlcpy(ut.ut_line, NTIME_MSG, sizeof(ut.ut_line));
549		ut.ut_time = timetv.tv_sec;
550		setutent();
551		pututline(&ut);
552		endutent();
553# else /* not HAVE_PUTUTLINE */
554# endif /* not HAVE_PUTUTLINE */
555#endif /* UPDATE_WTMP */
556
557		/* WTMPX */
558
559#ifdef UPDATE_WTMPX
560# ifdef HAVE_PUTUTXLINE
561		utx.ut_type = OLD_TIME;
562		utx.ut_tv = tvlast;
563		strlcpy(utx.ut_line, OTIME_MSG, sizeof(utx.ut_line));
564#  ifdef HAVE_UPDWTMPX
565		updwtmpx(WTMPX_FILE, &utx);
566#  else /* not HAVE_UPDWTMPX */
567#  endif /* not HAVE_UPDWTMPX */
568# else /* not HAVE_PUTUTXLINE */
569# endif /* not HAVE_PUTUTXLINE */
570# ifdef HAVE_PUTUTXLINE
571		utx.ut_type = NEW_TIME;
572		utx.ut_tv = timetv;
573		strlcpy(utx.ut_line, NTIME_MSG, sizeof(utx.ut_line));
574#  ifdef HAVE_UPDWTMPX
575		updwtmpx(WTMPX_FILE, &utx);
576#  else /* not HAVE_UPDWTMPX */
577#  endif /* not HAVE_UPDWTMPX */
578# else /* not HAVE_PUTUTXLINE */
579# endif /* not HAVE_PUTUTXLINE */
580#endif /* UPDATE_WTMPX */
581
582	}
583	if (enable_panic_check && allow_panic) {
584		msyslog(LOG_ERR, "step_systime: allow_panic is TRUE!");
585		INSIST(!allow_panic);
586	}
587	return TRUE;
588}
589
590#endif	/* !SIM */
591