2 * Copyright (c) 2008, 2009, 2010, 2011, 2012 Nicira, Inc.
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? Only relevant if CACHE_TIME is 1.
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 /* features for use by unit tests. */
58 static struct timespec warp_offset; /* Offset added to monotonic_time. */
59 static bool time_stopped; /* Disables real-time updates, if true. */
61 /* Time at which to die with SIGALRM (if not TIME_MIN). */
62 static time_t deadline = TIME_MIN;
64 static void set_up_timer(void);
65 static void set_up_signal(int flags);
66 static void sigalrm_handler(int);
67 static void refresh_wall_if_ticked(void);
68 static void refresh_monotonic_if_ticked(void);
69 static time_t time_add(time_t, time_t);
70 static void block_sigalrm(sigset_t *);
71 static void unblock_sigalrm(const sigset_t *);
72 static void log_poll_interval(long long int last_wakeup);
73 static struct rusage *get_recent_rusage(void);
74 static void refresh_rusage(void);
75 static void timespec_add(struct timespec *sum,
76 const struct timespec *a, const struct timespec *b);
78 /* Initializes the timetracking module, if not already initialized. */
90 if (!clock_gettime(CLOCK_MONOTONIC, &monotonic_time)) {
91 monotonic_clock = CLOCK_MONOTONIC;
93 monotonic_clock = CLOCK_REALTIME;
94 VLOG_DBG("monotonic timer not available");
98 set_up_signal(SA_RESTART);
102 boot_time = time_msec();
106 set_up_signal(int flags)
110 memset(&sa, 0, sizeof sa);
111 sa.sa_handler = sigalrm_handler;
112 sigemptyset(&sa.sa_mask);
114 xsigaction(SIGALRM, &sa, NULL);
117 /* Remove SA_RESTART from the flags for SIGALRM, so that any system call that
118 * is interrupted by the periodic timer interrupt will return EINTR instead of
119 * continuing after the signal handler returns.
121 * time_disable_restart() and time_enable_restart() may be usefully wrapped
122 * around function calls that might otherwise block forever unless interrupted
125 * time_disable_restart();
126 * fcntl(fd, F_SETLKW, &lock);
127 * time_enable_restart();
130 time_disable_restart(void)
136 /* Add SA_RESTART to the flags for SIGALRM, so that any system call that
137 * is interrupted by the periodic timer interrupt will continue after the
138 * signal handler returns instead of returning EINTR. */
140 time_enable_restart(void)
143 set_up_signal(SA_RESTART);
149 static timer_t timer_id; /* "static" to avoid apparent memory leak. */
150 struct itimerspec itimer;
152 if (timer_create(monotonic_clock, NULL, &timer_id)) {
153 VLOG_FATAL("timer_create failed (%s)", strerror(errno));
156 itimer.it_interval.tv_sec = 0;
157 itimer.it_interval.tv_nsec = TIME_UPDATE_INTERVAL * 1000 * 1000;
158 itimer.it_value = itimer.it_interval;
160 if (timer_settime(timer_id, 0, &itimer, NULL)) {
161 VLOG_FATAL("timer_settime failed (%s)", strerror(errno));
165 /* Set up the interval timer, to ensure that time advances even without calling
168 * A child created with fork() does not inherit the parent's interval timer, so
169 * this function needs to be called from the child after fork(). */
178 /* If we are not caching kernel time, the only reason the timer should
179 * exist is if time_alarm() was called and deadline is set */
180 if (deadline != TIME_MIN) {
190 clock_gettime(CLOCK_REALTIME, &wall_time);
195 refresh_monotonic(void)
200 if (monotonic_clock == CLOCK_MONOTONIC) {
201 clock_gettime(monotonic_clock, &monotonic_time);
203 refresh_wall_if_ticked();
204 monotonic_time = wall_time;
206 timespec_add(&monotonic_time, &monotonic_time, &warp_offset);
208 monotonic_tick = false;
212 /* Forces a refresh of the current time from the kernel. It is not usually
213 * necessary to call this function, since the time will be refreshed
214 * automatically at least every TIME_UPDATE_INTERVAL milliseconds. If
215 * CACHE_TIME is 0, we will always refresh the current time so this
216 * function has no effect. */
220 wall_tick = monotonic_tick = true;
223 /* Returns a monotonic timer, in seconds. */
227 refresh_monotonic_if_ticked();
228 return monotonic_time.tv_sec;
231 /* Same as time_now() except does not write to static variables, for use in
232 * signal handlers. */
236 struct timespec cur_time;
238 clock_gettime(monotonic_clock, &cur_time);
239 return cur_time.tv_sec;
242 /* Returns the current time, in seconds. */
246 refresh_wall_if_ticked();
247 return wall_time.tv_sec;
250 /* Returns a monotonic timer, in ms (within TIME_UPDATE_INTERVAL ms). */
254 refresh_monotonic_if_ticked();
255 return timespec_to_msec(&monotonic_time);
258 /* Returns the current time, in ms (within TIME_UPDATE_INTERVAL ms). */
262 refresh_wall_if_ticked();
263 return timespec_to_msec(&wall_time);
266 /* Stores a monotonic timer, accurate within TIME_UPDATE_INTERVAL ms, into
269 time_timespec(struct timespec *ts)
271 refresh_monotonic_if_ticked();
272 *ts = monotonic_time;
275 /* Stores the current time, accurate within TIME_UPDATE_INTERVAL ms, into
278 time_wall_timespec(struct timespec *ts)
280 refresh_wall_if_ticked();
284 /* Configures the program to die with SIGALRM 'secs' seconds from now, if
285 * 'secs' is nonzero, or disables the feature if 'secs' is zero. */
287 time_alarm(unsigned int secs)
293 block_sigalrm(&oldsigs);
294 deadline = secs ? time_add(time_now(), secs) : TIME_MIN;
295 unblock_sigalrm(&oldsigs);
298 /* If we aren't timing the gaps between kernel time refreshes we need to
299 * to start the timer up now */
300 set_up_signal(SA_RESTART);
305 /* Like poll(), except:
307 * - The timeout is specified as an absolute time, as defined by
308 * time_msec(), instead of a duration.
310 * - On error, returns a negative error code (instead of setting errno).
312 * - If interrupted by a signal, retries automatically until the original
313 * timeout is reached. (Because of this property, this function will
314 * never return -EINTR.)
316 * - As a side effect, refreshes the current time (like time_refresh()).
318 * Stores the number of milliseconds elapsed during poll in '*elapsed'. */
320 time_poll(struct pollfd *pollfds, int n_pollfds, long long int timeout_when,
323 static long long int last_wakeup;
330 log_poll_interval(last_wakeup);
335 long long int now = time_msec();
338 if (now >= timeout_when) {
340 } else if ((unsigned long long int) timeout_when - now > INT_MAX) {
343 time_left = timeout_when - now;
346 retval = poll(pollfds, n_pollfds, time_left);
351 if (retval != -EINTR) {
355 if (!blocked && deadline == TIME_MIN) {
356 block_sigalrm(&oldsigs);
361 unblock_sigalrm(&oldsigs);
363 last_wakeup = time_msec();
365 *elapsed = last_wakeup - start;
369 /* Returns the sum of 'a' and 'b', with saturation on overflow or underflow. */
371 time_add(time_t a, time_t b)
374 ? (b > TIME_MAX - a ? TIME_MAX : a + b)
375 : (b < TIME_MIN - a ? TIME_MIN : a + b));
379 sigalrm_handler(int sig_nr)
382 monotonic_tick = true;
383 if (deadline != TIME_MIN && time_now_sig() > deadline) {
384 fatal_signal_handler(sig_nr);
389 refresh_wall_if_ticked(void)
391 if (!CACHE_TIME || wall_tick) {
397 refresh_monotonic_if_ticked(void)
399 if (!CACHE_TIME || monotonic_tick) {
405 block_sigalrm(sigset_t *oldsigs)
408 sigemptyset(&sigalrm);
409 sigaddset(&sigalrm, SIGALRM);
410 xsigprocmask(SIG_BLOCK, &sigalrm, oldsigs);
414 unblock_sigalrm(const sigset_t *oldsigs)
416 xsigprocmask(SIG_SETMASK, oldsigs, NULL);
420 timespec_to_msec(const struct timespec *ts)
422 return (long long int) ts->tv_sec * 1000 + ts->tv_nsec / (1000 * 1000);
426 timeval_to_msec(const struct timeval *tv)
428 return (long long int) tv->tv_sec * 1000 + tv->tv_usec / 1000;
431 /* Returns the monotonic time at which the "time" module was initialized, in
441 xgettimeofday(struct timeval *tv)
443 if (gettimeofday(tv, NULL) == -1) {
444 VLOG_FATAL("gettimeofday failed (%s)", strerror(errno));
449 timeval_diff_msec(const struct timeval *a, const struct timeval *b)
451 return timeval_to_msec(a) - timeval_to_msec(b);
455 timespec_add(struct timespec *sum,
456 const struct timespec *a,
457 const struct timespec *b)
461 tmp.tv_sec = a->tv_sec + b->tv_sec;
462 tmp.tv_nsec = a->tv_nsec + b->tv_nsec;
463 if (tmp.tv_nsec >= 1000 * 1000 * 1000) {
464 tmp.tv_nsec -= 1000 * 1000 * 1000;
472 log_poll_interval(long long int last_wakeup)
474 static unsigned int mean_interval; /* In 16ths of a millisecond. */
475 static unsigned int n_samples;
478 unsigned int interval; /* In 16ths of a millisecond. */
480 /* Compute interval from last wakeup to now in 16ths of a millisecond,
481 * capped at 10 seconds (16000 in this unit). */
483 interval = MIN(10000, now - last_wakeup) << 4;
485 /* Warn if we took too much time between polls: at least 50 ms and at least
486 * 8X the mean interval. */
487 if (n_samples > 10 && interval > mean_interval * 8 && interval > 50 * 16) {
488 static struct vlog_rate_limit rl = VLOG_RATE_LIMIT_INIT(1, 3);
490 if (!VLOG_DROP_WARN(&rl)) {
491 const struct rusage *last_rusage = get_recent_rusage();
492 struct rusage rusage;
494 getrusage(RUSAGE_SELF, &rusage);
495 VLOG_WARN("%lld ms poll interval (%lld ms user, %lld ms system) "
496 "is over %u times the weighted mean interval %u ms "
499 timeval_diff_msec(&rusage.ru_utime,
500 &last_rusage->ru_utime),
501 timeval_diff_msec(&rusage.ru_stime,
502 &last_rusage->ru_stime),
503 interval / mean_interval,
504 (mean_interval + 8) / 16, n_samples);
505 if (rusage.ru_minflt > last_rusage->ru_minflt
506 || rusage.ru_majflt > last_rusage->ru_majflt) {
507 VLOG_WARN("faults: %ld minor, %ld major",
508 rusage.ru_minflt - last_rusage->ru_minflt,
509 rusage.ru_majflt - last_rusage->ru_majflt);
511 if (rusage.ru_inblock > last_rusage->ru_inblock
512 || rusage.ru_oublock > last_rusage->ru_oublock) {
513 VLOG_WARN("disk: %ld reads, %ld writes",
514 rusage.ru_inblock - last_rusage->ru_inblock,
515 rusage.ru_oublock - last_rusage->ru_oublock);
517 if (rusage.ru_nvcsw > last_rusage->ru_nvcsw
518 || rusage.ru_nivcsw > last_rusage->ru_nivcsw) {
519 VLOG_WARN("context switches: %ld voluntary, %ld involuntary",
520 rusage.ru_nvcsw - last_rusage->ru_nvcsw,
521 rusage.ru_nivcsw - last_rusage->ru_nivcsw);
527 /* Update exponentially weighted moving average. With these parameters, a
528 * given value decays to 1% of its value in about 100 time steps. */
530 mean_interval = (mean_interval * 122 + interval * 6 + 64) / 128;
532 mean_interval = interval;
536 /* CPU usage tracking. */
539 long long int when; /* Time that this sample was taken. */
540 unsigned long long int cpu; /* Total user+system CPU usage when sampled. */
543 static struct rusage recent_rusage;
544 static struct cpu_usage older = { LLONG_MIN, 0 };
545 static struct cpu_usage newer = { LLONG_MIN, 0 };
546 static int cpu_usage = -1;
548 static struct rusage *
549 get_recent_rusage(void)
551 return &recent_rusage;
560 getrusage(RUSAGE_SELF, &recent_rusage);
562 if (now >= newer.when + 3 * 1000) {
565 newer.cpu = (timeval_to_msec(&recent_rusage.ru_utime) +
566 timeval_to_msec(&recent_rusage.ru_stime));
568 if (older.when != LLONG_MIN && newer.cpu > older.cpu) {
569 unsigned int dividend = newer.cpu - older.cpu;
570 unsigned int divisor = (newer.when - older.when) / 100;
571 cpu_usage = divisor > 0 ? dividend / divisor : -1;
578 /* Returns an estimate of this process's CPU usage, as a percentage, over the
579 * past few seconds of wall-clock time. Returns -1 if no estimate is available
580 * (which will happen if the process has not been running long enough to have
581 * an estimate, and can happen for other reasons as well). */
588 /* Unixctl interface. */
590 /* "time/stop" stops the monotonic time returned by e.g. time_msec() from
591 * advancing, except due to later calls to "time/warp". */
593 timeval_stop_cb(struct unixctl_conn *conn,
594 int argc OVS_UNUSED, const char *argv[] OVS_UNUSED,
595 void *aux OVS_UNUSED)
598 unixctl_command_reply(conn, NULL);
601 /* "time/warp MSECS" advances the current monotonic time by the specified
602 * number of milliseconds. Unless "time/stop" has also been executed, the
603 * monotonic clock continues to tick forward at the normal rate afterward.
605 * Does not affect wall clock readings. */
607 timeval_warp_cb(struct unixctl_conn *conn,
608 int argc OVS_UNUSED, const char *argv[], void *aux OVS_UNUSED)
613 msecs = atoi(argv[1]);
615 unixctl_command_reply_error(conn, "invalid MSECS");
619 ts.tv_sec = msecs / 1000;
620 ts.tv_nsec = (msecs % 1000) * 1000 * 1000;
621 timespec_add(&warp_offset, &warp_offset, &ts);
622 timespec_add(&monotonic_time, &monotonic_time, &ts);
623 unixctl_command_reply(conn, "warped");
627 timeval_dummy_register(void)
629 unixctl_command_register("time/stop", "", 0, 0, timeval_stop_cb, NULL);
630 unixctl_command_register("time/warp", "MSECS", 1, 1,
631 timeval_warp_cb, NULL);