2 * Copyright (c) 2008, 2009, 2010, 2011, 2012 Nicira Networks.
4 * Licensed under the Apache License, Version 2.0 (the "License");
5 * you may not use this file except in compliance with the License.
6 * You may obtain a copy of the License at:
8 * http://www.apache.org/licenses/LICENSE-2.0
10 * Unless required by applicable law or agreed to in writing, software
11 * distributed under the License is distributed on an "AS IS" BASIS,
12 * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
13 * See the License for the specific language governing permissions and
14 * limitations under the License.
26 #include <sys/resource.h>
30 #include "fatal-signal.h"
36 VLOG_DEFINE_THIS_MODULE(timeval);
38 /* The clock to use for measuring time intervals. This is CLOCK_MONOTONIC by
39 * preference, but on systems that don't have a monotonic clock we fall back
40 * to CLOCK_REALTIME. */
41 static clockid_t monotonic_clock;
43 /* Has a timer tick occurred?
45 * We initialize these to true to force time_init() to get called on the first
46 * call to time_msec() or another function that queries the current time. */
47 static volatile sig_atomic_t wall_tick = true;
48 static volatile sig_atomic_t monotonic_tick = true;
50 /* The current time, as of the last refresh. */
51 static struct timespec wall_time;
52 static struct timespec monotonic_time;
54 /* The monotonic time at which the time module was initialized. */
55 static long long int boot_time;
57 /* Fixed monotonic time offset, for use by unit tests. */
58 static struct timespec warp_offset;
60 /* Time at which to die with SIGALRM (if not TIME_MIN). */
61 static time_t deadline = TIME_MIN;
63 static void set_up_timer(void);
64 static void set_up_signal(int flags);
65 static void sigalrm_handler(int);
66 static void refresh_wall_if_ticked(void);
67 static void refresh_monotonic_if_ticked(void);
68 static time_t time_add(time_t, time_t);
69 static void block_sigalrm(sigset_t *);
70 static void unblock_sigalrm(const sigset_t *);
71 static void log_poll_interval(long long int last_wakeup);
72 static struct rusage *get_recent_rusage(void);
73 static void refresh_rusage(void);
74 static void timespec_add(struct timespec *sum,
75 const struct timespec *a, const struct timespec *b);
77 /* Initializes the timetracking module, if not already initialized. */
89 if (!clock_gettime(CLOCK_MONOTONIC, &monotonic_time)) {
90 monotonic_clock = CLOCK_MONOTONIC;
92 monotonic_clock = CLOCK_REALTIME;
93 VLOG_DBG("monotonic timer not available");
96 set_up_signal(SA_RESTART);
98 boot_time = time_msec();
102 set_up_signal(int flags)
106 memset(&sa, 0, sizeof sa);
107 sa.sa_handler = sigalrm_handler;
108 sigemptyset(&sa.sa_mask);
110 xsigaction(SIGALRM, &sa, NULL);
113 /* Remove SA_RESTART from the flags for SIGALRM, so that any system call that
114 * is interrupted by the periodic timer interrupt will return EINTR instead of
115 * continuing after the signal handler returns.
117 * time_disable_restart() and time_enable_restart() may be usefully wrapped
118 * around function calls that might otherwise block forever unless interrupted
121 * time_disable_restart();
122 * fcntl(fd, F_SETLKW, &lock);
123 * time_enable_restart();
126 time_disable_restart(void)
132 /* Add SA_RESTART to the flags for SIGALRM, so that any system call that
133 * is interrupted by the periodic timer interrupt will continue after the
134 * signal handler returns instead of returning EINTR. */
136 time_enable_restart(void)
139 set_up_signal(SA_RESTART);
145 static timer_t timer_id; /* "static" to avoid apparent memory leak. */
146 struct itimerspec itimer;
148 if (timer_create(monotonic_clock, NULL, &timer_id)) {
149 VLOG_FATAL("timer_create failed (%s)", strerror(errno));
152 itimer.it_interval.tv_sec = 0;
153 itimer.it_interval.tv_nsec = TIME_UPDATE_INTERVAL * 1000 * 1000;
154 itimer.it_value = itimer.it_interval;
156 if (timer_settime(timer_id, 0, &itimer, NULL)) {
157 VLOG_FATAL("timer_settime failed (%s)", strerror(errno));
161 /* Set up the interval timer, to ensure that time advances even without calling
164 * A child created with fork() does not inherit the parent's interval timer, so
165 * this function needs to be called from the child after fork(). */
177 clock_gettime(CLOCK_REALTIME, &wall_time);
182 refresh_monotonic(void)
186 if (monotonic_clock == CLOCK_MONOTONIC) {
187 clock_gettime(monotonic_clock, &monotonic_time);
189 refresh_wall_if_ticked();
190 monotonic_time = wall_time;
192 timespec_add(&monotonic_time, &monotonic_time, &warp_offset);
194 monotonic_tick = false;
197 /* Forces a refresh of the current time from the kernel. It is not usually
198 * necessary to call this function, since the time will be refreshed
199 * automatically at least every TIME_UPDATE_INTERVAL milliseconds. */
203 wall_tick = monotonic_tick = true;
206 /* Returns a monotonic timer, in seconds. */
210 refresh_monotonic_if_ticked();
211 return monotonic_time.tv_sec;
214 /* Same as time_now() except does not write to static variables, for use in
215 * signal handlers. */
219 struct timespec cur_time;
221 clock_gettime(monotonic_clock, &cur_time);
222 return cur_time.tv_sec;
225 /* Returns the current time, in seconds. */
229 refresh_wall_if_ticked();
230 return wall_time.tv_sec;
233 /* Returns a monotonic timer, in ms (within TIME_UPDATE_INTERVAL ms). */
237 refresh_monotonic_if_ticked();
238 return timespec_to_msec(&monotonic_time);
241 /* Returns the current time, in ms (within TIME_UPDATE_INTERVAL ms). */
245 refresh_wall_if_ticked();
246 return timespec_to_msec(&wall_time);
249 /* Stores a monotonic timer, accurate within TIME_UPDATE_INTERVAL ms, into
252 time_timespec(struct timespec *ts)
254 refresh_monotonic_if_ticked();
255 *ts = monotonic_time;
258 /* Stores the current time, accurate within TIME_UPDATE_INTERVAL ms, into
261 time_wall_timespec(struct timespec *ts)
263 refresh_wall_if_ticked();
267 /* Configures the program to die with SIGALRM 'secs' seconds from now, if
268 * 'secs' is nonzero, or disables the feature if 'secs' is zero. */
270 time_alarm(unsigned int secs)
275 block_sigalrm(&oldsigs);
276 deadline = secs ? time_add(time_now(), secs) : TIME_MIN;
277 unblock_sigalrm(&oldsigs);
280 /* Like poll(), except:
282 * - The timeout is specified as an absolute time, as defined by
283 * time_msec(), instead of a duration.
285 * - On error, returns a negative error code (instead of setting errno).
287 * - If interrupted by a signal, retries automatically until the original
288 * timeout is reached. (Because of this property, this function will
289 * never return -EINTR.)
291 * - As a side effect, refreshes the current time (like time_refresh()).
293 * Stores the number of milliseconds elapsed during poll in '*elapsed'. */
295 time_poll(struct pollfd *pollfds, int n_pollfds, long long int timeout_when,
298 static long long int last_wakeup;
305 log_poll_interval(last_wakeup);
310 long long int now = time_msec();
313 if (now >= timeout_when) {
315 } else if ((unsigned long long int) timeout_when - now > INT_MAX) {
318 time_left = timeout_when - now;
321 retval = poll(pollfds, n_pollfds, time_left);
326 if (retval != -EINTR) {
330 if (!blocked && deadline == TIME_MIN) {
331 block_sigalrm(&oldsigs);
336 unblock_sigalrm(&oldsigs);
338 last_wakeup = time_msec();
340 *elapsed = last_wakeup - start;
344 /* Returns the sum of 'a' and 'b', with saturation on overflow or underflow. */
346 time_add(time_t a, time_t b)
349 ? (b > TIME_MAX - a ? TIME_MAX : a + b)
350 : (b < TIME_MIN - a ? TIME_MIN : a + b));
354 sigalrm_handler(int sig_nr)
357 monotonic_tick = true;
358 if (deadline != TIME_MIN && time_now_sig() > deadline) {
359 fatal_signal_handler(sig_nr);
364 refresh_wall_if_ticked(void)
372 refresh_monotonic_if_ticked(void)
374 if (monotonic_tick) {
380 block_sigalrm(sigset_t *oldsigs)
383 sigemptyset(&sigalrm);
384 sigaddset(&sigalrm, SIGALRM);
385 xsigprocmask(SIG_BLOCK, &sigalrm, oldsigs);
389 unblock_sigalrm(const sigset_t *oldsigs)
391 xsigprocmask(SIG_SETMASK, oldsigs, NULL);
395 timespec_to_msec(const struct timespec *ts)
397 return (long long int) ts->tv_sec * 1000 + ts->tv_nsec / (1000 * 1000);
401 timeval_to_msec(const struct timeval *tv)
403 return (long long int) tv->tv_sec * 1000 + tv->tv_usec / 1000;
406 /* Returns the monotonic time at which the "time" module was initialized, in
416 xgettimeofday(struct timeval *tv)
418 if (gettimeofday(tv, NULL) == -1) {
419 VLOG_FATAL("gettimeofday failed (%s)", strerror(errno));
424 timeval_diff_msec(const struct timeval *a, const struct timeval *b)
426 return timeval_to_msec(a) - timeval_to_msec(b);
430 timespec_add(struct timespec *sum,
431 const struct timespec *a,
432 const struct timespec *b)
436 tmp.tv_sec = a->tv_sec + b->tv_sec;
437 tmp.tv_nsec = a->tv_nsec + b->tv_nsec;
438 if (tmp.tv_nsec >= 1000 * 1000 * 1000) {
439 tmp.tv_nsec -= 1000 * 1000 * 1000;
447 log_poll_interval(long long int last_wakeup)
449 static unsigned int mean_interval; /* In 16ths of a millisecond. */
450 static unsigned int n_samples;
453 unsigned int interval; /* In 16ths of a millisecond. */
455 /* Compute interval from last wakeup to now in 16ths of a millisecond,
456 * capped at 10 seconds (16000 in this unit). */
458 interval = MIN(10000, now - last_wakeup) << 4;
460 /* Warn if we took too much time between polls: at least 50 ms and at least
461 * 8X the mean interval. */
462 if (n_samples > 10 && interval > mean_interval * 8 && interval > 50 * 16) {
463 const struct rusage *last_rusage = get_recent_rusage();
464 struct rusage rusage;
466 getrusage(RUSAGE_SELF, &rusage);
467 VLOG_WARN("%lld ms poll interval (%lld ms user, %lld ms system) "
468 "is over %u times the weighted mean interval %u ms "
471 timeval_diff_msec(&rusage.ru_utime, &last_rusage->ru_utime),
472 timeval_diff_msec(&rusage.ru_stime, &last_rusage->ru_stime),
473 interval / mean_interval,
474 (mean_interval + 8) / 16, n_samples);
475 if (rusage.ru_minflt > last_rusage->ru_minflt
476 || rusage.ru_majflt > last_rusage->ru_majflt) {
477 VLOG_WARN("faults: %ld minor, %ld major",
478 rusage.ru_minflt - last_rusage->ru_minflt,
479 rusage.ru_majflt - last_rusage->ru_majflt);
481 if (rusage.ru_inblock > last_rusage->ru_inblock
482 || rusage.ru_oublock > last_rusage->ru_oublock) {
483 VLOG_WARN("disk: %ld reads, %ld writes",
484 rusage.ru_inblock - last_rusage->ru_inblock,
485 rusage.ru_oublock - last_rusage->ru_oublock);
487 if (rusage.ru_nvcsw > last_rusage->ru_nvcsw
488 || rusage.ru_nivcsw > last_rusage->ru_nivcsw) {
489 VLOG_WARN("context switches: %ld voluntary, %ld involuntary",
490 rusage.ru_nvcsw - last_rusage->ru_nvcsw,
491 rusage.ru_nivcsw - last_rusage->ru_nivcsw);
494 /* Care should be taken in the value chosen for logging. Depending
495 * on the configuration, syslog can write changes synchronously,
496 * which can cause the coverage messages to take longer to log
497 * than the processing delay that triggered it. */
498 coverage_log(VLL_INFO, true);
501 /* Update exponentially weighted moving average. With these parameters, a
502 * given value decays to 1% of its value in about 100 time steps. */
504 mean_interval = (mean_interval * 122 + interval * 6 + 64) / 128;
506 mean_interval = interval;
510 /* CPU usage tracking. */
513 long long int when; /* Time that this sample was taken. */
514 unsigned long long int cpu; /* Total user+system CPU usage when sampled. */
517 static struct rusage recent_rusage;
518 static struct cpu_usage older = { LLONG_MIN, 0 };
519 static struct cpu_usage newer = { LLONG_MIN, 0 };
520 static int cpu_usage = -1;
522 static struct rusage *
523 get_recent_rusage(void)
525 return &recent_rusage;
534 getrusage(RUSAGE_SELF, &recent_rusage);
536 if (now >= newer.when + 3 * 1000) {
539 newer.cpu = (timeval_to_msec(&recent_rusage.ru_utime) +
540 timeval_to_msec(&recent_rusage.ru_stime));
542 if (older.when != LLONG_MIN && newer.cpu > older.cpu) {
543 unsigned int dividend = newer.cpu - older.cpu;
544 unsigned int divisor = (newer.when - older.when) / 100;
545 cpu_usage = divisor > 0 ? dividend / divisor : -1;
552 /* Returns an estimate of this process's CPU usage, as a percentage, over the
553 * past few seconds of wall-clock time. Returns -1 if no estimate is available
554 * (which will happen if the process has not been running long enough to have
555 * an estimate, and can happen for other reasons as well). */
562 /* Unixctl interface. */
565 timeval_warp_cb(struct unixctl_conn *conn,
566 int argc OVS_UNUSED, const char *argv[], void *aux OVS_UNUSED)
571 msecs = atoi(argv[1]);
573 unixctl_command_reply_error(conn, "invalid MSECS");
577 ts.tv_sec = msecs / 1000;
578 ts.tv_nsec = (msecs % 1000) * 1000 * 1000;
579 timespec_add(&warp_offset, &warp_offset, &ts);
580 unixctl_command_reply(conn, "warped");
584 timeval_dummy_register(void)
586 unixctl_command_register("time/warp", "MSECS", 1, 1,
587 timeval_warp_cb, NULL);