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 /* features for use by unit tests. */
55 static struct timespec warp_offset; /* Offset added to monotonic_time. */
56 static bool time_stopped; /* Disables real-time updates, if true. */
58 /* Time at which to die with SIGALRM (if not TIME_MIN). */
59 static time_t deadline = TIME_MIN;
61 static void set_up_timer(void);
62 static void set_up_signal(int flags);
63 static void sigalrm_handler(int);
64 static void refresh_wall_if_ticked(void);
65 static void refresh_monotonic_if_ticked(void);
66 static time_t time_add(time_t, time_t);
67 static void block_sigalrm(sigset_t *);
68 static void unblock_sigalrm(const sigset_t *);
69 static void log_poll_interval(long long int last_wakeup);
70 static struct rusage *get_recent_rusage(void);
71 static void refresh_rusage(void);
72 static void timespec_add(struct timespec *sum,
73 const struct timespec *a, const struct timespec *b);
75 /* Initializes the timetracking module.
77 * It is not necessary to call this function directly, because other time
78 * functions will call it automatically, but it doesn't hurt. */
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");
97 set_up_signal(SA_RESTART);
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)
187 if (monotonic_clock == CLOCK_MONOTONIC) {
188 clock_gettime(monotonic_clock, &monotonic_time);
190 refresh_wall_if_ticked();
191 monotonic_time = wall_time;
193 timespec_add(&monotonic_time, &monotonic_time, &warp_offset);
195 monotonic_tick = false;
199 /* Forces a refresh of the current time from the kernel. It is not usually
200 * necessary to call this function, since the time will be refreshed
201 * automatically at least every TIME_UPDATE_INTERVAL milliseconds. */
205 wall_tick = monotonic_tick = true;
208 /* Returns a monotonic timer, in seconds. */
212 refresh_monotonic_if_ticked();
213 return monotonic_time.tv_sec;
216 /* Same as time_now() except does not write to static variables, for use in
217 * signal handlers. */
221 struct timespec cur_time;
223 clock_gettime(monotonic_clock, &cur_time);
224 return cur_time.tv_sec;
227 /* Returns the current time, in seconds. */
231 refresh_wall_if_ticked();
232 return wall_time.tv_sec;
235 /* Returns a monotonic timer, in ms (within TIME_UPDATE_INTERVAL ms). */
239 refresh_monotonic_if_ticked();
240 return timespec_to_msec(&monotonic_time);
243 /* Returns the current time, in ms (within TIME_UPDATE_INTERVAL ms). */
247 refresh_wall_if_ticked();
248 return timespec_to_msec(&wall_time);
251 /* Stores a monotonic timer, accurate within TIME_UPDATE_INTERVAL ms, into
254 time_timespec(struct timespec *ts)
256 refresh_monotonic_if_ticked();
257 *ts = monotonic_time;
260 /* Stores the current time, accurate within TIME_UPDATE_INTERVAL ms, into
263 time_wall_timespec(struct timespec *ts)
265 refresh_wall_if_ticked();
269 /* Configures the program to die with SIGALRM 'secs' seconds from now, if
270 * 'secs' is nonzero, or disables the feature if 'secs' is zero. */
272 time_alarm(unsigned int secs)
277 block_sigalrm(&oldsigs);
278 deadline = secs ? time_add(time_now(), secs) : TIME_MIN;
279 unblock_sigalrm(&oldsigs);
282 /* Like poll(), except:
284 * - The timeout is specified as an absolute time, as defined by
285 * time_msec(), instead of a duration.
287 * - On error, returns a negative error code (instead of setting errno).
289 * - If interrupted by a signal, retries automatically until the original
290 * timeout is reached. (Because of this property, this function will
291 * never return -EINTR.)
293 * - As a side effect, refreshes the current time (like time_refresh()).
295 * Stores the number of milliseconds elapsed during poll in '*elapsed'. */
297 time_poll(struct pollfd *pollfds, int n_pollfds, long long int timeout_when,
300 static long long int last_wakeup;
307 log_poll_interval(last_wakeup);
312 long long int now = time_msec();
315 if (now >= timeout_when) {
317 } else if ((unsigned long long int) timeout_when - now > INT_MAX) {
320 time_left = timeout_when - now;
323 retval = poll(pollfds, n_pollfds, time_left);
328 if (retval != -EINTR) {
332 if (!blocked && deadline == TIME_MIN) {
333 block_sigalrm(&oldsigs);
338 unblock_sigalrm(&oldsigs);
340 last_wakeup = time_msec();
342 *elapsed = last_wakeup - start;
346 /* Returns the sum of 'a' and 'b', with saturation on overflow or underflow. */
348 time_add(time_t a, time_t b)
351 ? (b > TIME_MAX - a ? TIME_MAX : a + b)
352 : (b < TIME_MIN - a ? TIME_MIN : a + b));
356 sigalrm_handler(int sig_nr)
359 monotonic_tick = true;
360 if (deadline != TIME_MIN && time_now_sig() > deadline) {
361 fatal_signal_handler(sig_nr);
366 refresh_wall_if_ticked(void)
374 refresh_monotonic_if_ticked(void)
376 if (monotonic_tick) {
382 block_sigalrm(sigset_t *oldsigs)
385 sigemptyset(&sigalrm);
386 sigaddset(&sigalrm, SIGALRM);
387 xsigprocmask(SIG_BLOCK, &sigalrm, oldsigs);
391 unblock_sigalrm(const sigset_t *oldsigs)
393 xsigprocmask(SIG_SETMASK, oldsigs, NULL);
397 timespec_to_msec(const struct timespec *ts)
399 return (long long int) ts->tv_sec * 1000 + ts->tv_nsec / (1000 * 1000);
403 timeval_to_msec(const struct timeval *tv)
405 return (long long int) tv->tv_sec * 1000 + tv->tv_usec / 1000;
409 xgettimeofday(struct timeval *tv)
411 if (gettimeofday(tv, NULL) == -1) {
412 VLOG_FATAL("gettimeofday failed (%s)", strerror(errno));
417 timeval_diff_msec(const struct timeval *a, const struct timeval *b)
419 return timeval_to_msec(a) - timeval_to_msec(b);
423 timespec_add(struct timespec *sum,
424 const struct timespec *a,
425 const struct timespec *b)
429 tmp.tv_sec = a->tv_sec + b->tv_sec;
430 tmp.tv_nsec = a->tv_nsec + b->tv_nsec;
431 if (tmp.tv_nsec >= 1000 * 1000 * 1000) {
432 tmp.tv_nsec -= 1000 * 1000 * 1000;
440 log_poll_interval(long long int last_wakeup)
442 static unsigned int mean_interval; /* In 16ths of a millisecond. */
443 static unsigned int n_samples;
446 unsigned int interval; /* In 16ths of a millisecond. */
448 /* Compute interval from last wakeup to now in 16ths of a millisecond,
449 * capped at 10 seconds (16000 in this unit). */
451 interval = MIN(10000, now - last_wakeup) << 4;
453 /* Warn if we took too much time between polls: at least 50 ms and at least
454 * 8X the mean interval. */
455 if (n_samples > 10 && interval > mean_interval * 8 && interval > 50 * 16) {
456 const struct rusage *last_rusage = get_recent_rusage();
457 struct rusage rusage;
459 getrusage(RUSAGE_SELF, &rusage);
460 VLOG_WARN("%lld ms poll interval (%lld ms user, %lld ms system) "
461 "is over %u times the weighted mean interval %u ms "
464 timeval_diff_msec(&rusage.ru_utime, &last_rusage->ru_utime),
465 timeval_diff_msec(&rusage.ru_stime, &last_rusage->ru_stime),
466 interval / mean_interval,
467 (mean_interval + 8) / 16, n_samples);
468 if (rusage.ru_minflt > last_rusage->ru_minflt
469 || rusage.ru_majflt > last_rusage->ru_majflt) {
470 VLOG_WARN("faults: %ld minor, %ld major",
471 rusage.ru_minflt - last_rusage->ru_minflt,
472 rusage.ru_majflt - last_rusage->ru_majflt);
474 if (rusage.ru_inblock > last_rusage->ru_inblock
475 || rusage.ru_oublock > last_rusage->ru_oublock) {
476 VLOG_WARN("disk: %ld reads, %ld writes",
477 rusage.ru_inblock - last_rusage->ru_inblock,
478 rusage.ru_oublock - last_rusage->ru_oublock);
480 if (rusage.ru_nvcsw > last_rusage->ru_nvcsw
481 || rusage.ru_nivcsw > last_rusage->ru_nivcsw) {
482 VLOG_WARN("context switches: %ld voluntary, %ld involuntary",
483 rusage.ru_nvcsw - last_rusage->ru_nvcsw,
484 rusage.ru_nivcsw - last_rusage->ru_nivcsw);
487 /* Care should be taken in the value chosen for logging. Depending
488 * on the configuration, syslog can write changes synchronously,
489 * which can cause the coverage messages to take longer to log
490 * than the processing delay that triggered it. */
491 coverage_log(VLL_INFO, true);
494 /* Update exponentially weighted moving average. With these parameters, a
495 * given value decays to 1% of its value in about 100 time steps. */
497 mean_interval = (mean_interval * 122 + interval * 6 + 64) / 128;
499 mean_interval = interval;
503 /* CPU usage tracking. */
506 long long int when; /* Time that this sample was taken. */
507 unsigned long long int cpu; /* Total user+system CPU usage when sampled. */
510 static struct rusage recent_rusage;
511 static struct cpu_usage older = { LLONG_MIN, 0 };
512 static struct cpu_usage newer = { LLONG_MIN, 0 };
513 static int cpu_usage = -1;
515 static struct rusage *
516 get_recent_rusage(void)
518 return &recent_rusage;
527 getrusage(RUSAGE_SELF, &recent_rusage);
529 if (now >= newer.when + 3 * 1000) {
532 newer.cpu = (timeval_to_msec(&recent_rusage.ru_utime) +
533 timeval_to_msec(&recent_rusage.ru_stime));
535 if (older.when != LLONG_MIN && newer.cpu > older.cpu) {
536 unsigned int dividend = newer.cpu - older.cpu;
537 unsigned int divisor = (newer.when - older.when) / 100;
538 cpu_usage = divisor > 0 ? dividend / divisor : -1;
545 /* Returns an estimate of this process's CPU usage, as a percentage, over the
546 * past few seconds of wall-clock time. Returns -1 if no estimate is available
547 * (which will happen if the process has not been running long enough to have
548 * an estimate, and can happen for other reasons as well). */
555 /* Unixctl interface. */
557 /* "time/stop" stops the monotonic time returned by e.g. time_msec() from
558 * advancing, except due to later calls to "time/warp". */
560 timeval_stop_cb(struct unixctl_conn *conn,
561 int argc OVS_UNUSED, const char *argv[] OVS_UNUSED,
562 void *aux OVS_UNUSED)
565 unixctl_command_reply(conn, 200, NULL);
568 /* "time/warp MSECS" advances the current monotonic time by the specified
569 * number of milliseconds. Unless "time/stop" has also been executed, the
570 * monotonic clock continues to tick forward at the normal rate afterward.
572 * Does not affect wall clock readings. */
574 timeval_warp_cb(struct unixctl_conn *conn,
575 int argc OVS_UNUSED, const char *argv[], void *aux OVS_UNUSED)
580 msecs = atoi(argv[1]);
582 unixctl_command_reply(conn, 501, "invalid MSECS");
586 ts.tv_sec = msecs / 1000;
587 ts.tv_nsec = (msecs % 1000) * 1000 * 1000;
588 timespec_add(&warp_offset, &warp_offset, &ts);
589 timespec_add(&monotonic_time, &monotonic_time, &ts);
590 unixctl_command_reply(conn, 200, "warped");
594 timeval_dummy_register(void)
596 unixctl_command_register("time/stop", "", 0, 0, timeval_stop_cb, NULL);
597 unixctl_command_register("time/warp", "MSECS", 1, 1,
598 timeval_warp_cb, NULL);