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 true.
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 in milliseconds at which to die with SIGALRM (if not LLONG_MAX). */
62 static long long int deadline = LLONG_MAX;
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 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);
99 boot_time = time_msec();
103 set_up_signal(int flags)
107 memset(&sa, 0, sizeof sa);
108 sa.sa_handler = sigalrm_handler;
109 sigemptyset(&sa.sa_mask);
111 xsigaction(SIGALRM, &sa, NULL);
114 /* Remove SA_RESTART from the flags for SIGALRM, so that any system call that
115 * is interrupted by the periodic timer interrupt will return EINTR instead of
116 * continuing after the signal handler returns.
118 * time_disable_restart() and time_enable_restart() may be usefully wrapped
119 * around function calls that might otherwise block forever unless interrupted
122 * time_disable_restart();
123 * fcntl(fd, F_SETLKW, &lock);
124 * time_enable_restart();
127 time_disable_restart(void)
133 /* Add SA_RESTART to the flags for SIGALRM, so that any system call that
134 * is interrupted by the periodic timer interrupt will continue after the
135 * signal handler returns instead of returning EINTR. */
137 time_enable_restart(void)
140 set_up_signal(SA_RESTART);
146 static timer_t timer_id; /* "static" to avoid apparent memory leak. */
147 struct itimerspec itimer;
153 if (timer_create(monotonic_clock, NULL, &timer_id)) {
154 VLOG_FATAL("timer_create failed (%s)", strerror(errno));
157 itimer.it_interval.tv_sec = 0;
158 itimer.it_interval.tv_nsec = TIME_UPDATE_INTERVAL * 1000 * 1000;
159 itimer.it_value = itimer.it_interval;
161 if (timer_settime(timer_id, 0, &itimer, NULL)) {
162 VLOG_FATAL("timer_settime failed (%s)", strerror(errno));
166 /* Set up the interval timer, to ensure that time advances even without calling
169 * A child created with fork() does not inherit the parent's interval timer, so
170 * this function needs to be called from the child after fork(). */
182 clock_gettime(CLOCK_REALTIME, &wall_time);
187 refresh_monotonic(void)
192 if (monotonic_clock == CLOCK_MONOTONIC) {
193 clock_gettime(monotonic_clock, &monotonic_time);
195 refresh_wall_if_ticked();
196 monotonic_time = wall_time;
198 timespec_add(&monotonic_time, &monotonic_time, &warp_offset);
200 monotonic_tick = false;
204 /* Forces a refresh of the current time from the kernel. It is not usually
205 * necessary to call this function, since the time will be refreshed
206 * automatically at least every TIME_UPDATE_INTERVAL milliseconds. If
207 * CACHE_TIME is false, we will always refresh the current time so this
208 * function has no effect. */
212 wall_tick = monotonic_tick = true;
215 /* Returns a monotonic timer, in seconds. */
219 refresh_monotonic_if_ticked();
220 return monotonic_time.tv_sec;
223 /* Returns the current time, in seconds. */
227 refresh_wall_if_ticked();
228 return wall_time.tv_sec;
231 /* Returns a monotonic timer, in ms (within TIME_UPDATE_INTERVAL ms). */
235 refresh_monotonic_if_ticked();
236 return timespec_to_msec(&monotonic_time);
239 /* Returns the current time, in ms (within TIME_UPDATE_INTERVAL ms). */
243 refresh_wall_if_ticked();
244 return timespec_to_msec(&wall_time);
247 /* Stores a monotonic timer, accurate within TIME_UPDATE_INTERVAL ms, into
250 time_timespec(struct timespec *ts)
252 refresh_monotonic_if_ticked();
253 *ts = monotonic_time;
256 /* Stores the current time, accurate within TIME_UPDATE_INTERVAL ms, into
259 time_wall_timespec(struct timespec *ts)
261 refresh_wall_if_ticked();
265 /* Configures the program to die with SIGALRM 'secs' seconds from now, if
266 * 'secs' is nonzero, or disables the feature if 'secs' is zero. */
268 time_alarm(unsigned int secs)
281 block_sigalrm(&oldsigs);
282 deadline = now < LLONG_MAX - msecs ? now + msecs : LLONG_MAX;
283 unblock_sigalrm(&oldsigs);
286 /* Like poll(), except:
288 * - The timeout is specified as an absolute time, as defined by
289 * time_msec(), instead of a duration.
291 * - On error, returns a negative error code (instead of setting errno).
293 * - If interrupted by a signal, retries automatically until the original
294 * timeout is reached. (Because of this property, this function will
295 * never return -EINTR.)
297 * - As a side effect, refreshes the current time (like time_refresh()).
299 * Stores the number of milliseconds elapsed during poll in '*elapsed'. */
301 time_poll(struct pollfd *pollfds, int n_pollfds, long long int timeout_when,
304 static long long int last_wakeup;
311 log_poll_interval(last_wakeup);
316 timeout_when = MIN(timeout_when, deadline);
319 long long int now = time_msec();
322 if (now >= timeout_when) {
324 } else if ((unsigned long long int) timeout_when - now > INT_MAX) {
327 time_left = timeout_when - now;
330 retval = poll(pollfds, n_pollfds, time_left);
336 if (deadline <= time_msec()) {
337 fatal_signal_handler(SIGALRM);
344 if (retval != -EINTR) {
348 if (!blocked && !CACHE_TIME) {
349 block_sigalrm(&oldsigs);
354 unblock_sigalrm(&oldsigs);
356 last_wakeup = time_msec();
358 *elapsed = last_wakeup - start;
363 sigalrm_handler(int sig_nr OVS_UNUSED)
366 monotonic_tick = true;
370 refresh_wall_if_ticked(void)
372 if (!CACHE_TIME || wall_tick) {
378 refresh_monotonic_if_ticked(void)
380 if (!CACHE_TIME || monotonic_tick) {
386 block_sigalrm(sigset_t *oldsigs)
389 sigemptyset(&sigalrm);
390 sigaddset(&sigalrm, SIGALRM);
391 xsigprocmask(SIG_BLOCK, &sigalrm, oldsigs);
395 unblock_sigalrm(const sigset_t *oldsigs)
397 xsigprocmask(SIG_SETMASK, oldsigs, NULL);
401 timespec_to_msec(const struct timespec *ts)
403 return (long long int) ts->tv_sec * 1000 + ts->tv_nsec / (1000 * 1000);
407 timeval_to_msec(const struct timeval *tv)
409 return (long long int) tv->tv_sec * 1000 + tv->tv_usec / 1000;
412 /* Returns the monotonic time at which the "time" module was initialized, in
422 xgettimeofday(struct timeval *tv)
424 if (gettimeofday(tv, NULL) == -1) {
425 VLOG_FATAL("gettimeofday failed (%s)", strerror(errno));
430 timeval_diff_msec(const struct timeval *a, const struct timeval *b)
432 return timeval_to_msec(a) - timeval_to_msec(b);
436 timespec_add(struct timespec *sum,
437 const struct timespec *a,
438 const struct timespec *b)
442 tmp.tv_sec = a->tv_sec + b->tv_sec;
443 tmp.tv_nsec = a->tv_nsec + b->tv_nsec;
444 if (tmp.tv_nsec >= 1000 * 1000 * 1000) {
445 tmp.tv_nsec -= 1000 * 1000 * 1000;
453 log_poll_interval(long long int last_wakeup)
455 static unsigned int mean_interval; /* In 16ths of a millisecond. */
456 static unsigned int n_samples;
459 unsigned int interval; /* In 16ths of a millisecond. */
461 /* Compute interval from last wakeup to now in 16ths of a millisecond,
462 * capped at 10 seconds (16000 in this unit). */
464 interval = MIN(10000, now - last_wakeup) << 4;
466 /* Warn if we took too much time between polls: at least 50 ms and at least
467 * 8X the mean interval. */
468 if (n_samples > 10 && interval > mean_interval * 8 && interval > 50 * 16) {
469 static struct vlog_rate_limit rl = VLOG_RATE_LIMIT_INIT(1, 3);
471 if (!VLOG_DROP_WARN(&rl)) {
472 const struct rusage *last_rusage = get_recent_rusage();
473 struct rusage rusage;
475 getrusage(RUSAGE_SELF, &rusage);
476 VLOG_WARN("%lld ms poll interval (%lld ms user, %lld ms system) "
477 "is over %u times the weighted mean interval %u ms "
480 timeval_diff_msec(&rusage.ru_utime,
481 &last_rusage->ru_utime),
482 timeval_diff_msec(&rusage.ru_stime,
483 &last_rusage->ru_stime),
484 interval / mean_interval,
485 (mean_interval + 8) / 16, n_samples);
486 if (rusage.ru_minflt > last_rusage->ru_minflt
487 || rusage.ru_majflt > last_rusage->ru_majflt) {
488 VLOG_WARN("faults: %ld minor, %ld major",
489 rusage.ru_minflt - last_rusage->ru_minflt,
490 rusage.ru_majflt - last_rusage->ru_majflt);
492 if (rusage.ru_inblock > last_rusage->ru_inblock
493 || rusage.ru_oublock > last_rusage->ru_oublock) {
494 VLOG_WARN("disk: %ld reads, %ld writes",
495 rusage.ru_inblock - last_rusage->ru_inblock,
496 rusage.ru_oublock - last_rusage->ru_oublock);
498 if (rusage.ru_nvcsw > last_rusage->ru_nvcsw
499 || rusage.ru_nivcsw > last_rusage->ru_nivcsw) {
500 VLOG_WARN("context switches: %ld voluntary, %ld involuntary",
501 rusage.ru_nvcsw - last_rusage->ru_nvcsw,
502 rusage.ru_nivcsw - last_rusage->ru_nivcsw);
508 /* Update exponentially weighted moving average. With these parameters, a
509 * given value decays to 1% of its value in about 100 time steps. */
511 mean_interval = (mean_interval * 122 + interval * 6 + 64) / 128;
513 mean_interval = interval;
517 /* CPU usage tracking. */
520 long long int when; /* Time that this sample was taken. */
521 unsigned long long int cpu; /* Total user+system CPU usage when sampled. */
524 static struct rusage recent_rusage;
525 static struct cpu_usage older = { LLONG_MIN, 0 };
526 static struct cpu_usage newer = { LLONG_MIN, 0 };
527 static int cpu_usage = -1;
529 static struct rusage *
530 get_recent_rusage(void)
532 return &recent_rusage;
541 getrusage(RUSAGE_SELF, &recent_rusage);
543 if (now >= newer.when + 3 * 1000) {
546 newer.cpu = (timeval_to_msec(&recent_rusage.ru_utime) +
547 timeval_to_msec(&recent_rusage.ru_stime));
549 if (older.when != LLONG_MIN && newer.cpu > older.cpu) {
550 unsigned int dividend = newer.cpu - older.cpu;
551 unsigned int divisor = (newer.when - older.when) / 100;
552 cpu_usage = divisor > 0 ? dividend / divisor : -1;
559 /* Returns an estimate of this process's CPU usage, as a percentage, over the
560 * past few seconds of wall-clock time. Returns -1 if no estimate is available
561 * (which will happen if the process has not been running long enough to have
562 * an estimate, and can happen for other reasons as well). */
569 /* Unixctl interface. */
571 /* "time/stop" stops the monotonic time returned by e.g. time_msec() from
572 * advancing, except due to later calls to "time/warp". */
574 timeval_stop_cb(struct unixctl_conn *conn,
575 int argc OVS_UNUSED, const char *argv[] OVS_UNUSED,
576 void *aux OVS_UNUSED)
579 unixctl_command_reply(conn, NULL);
582 /* "time/warp MSECS" advances the current monotonic time by the specified
583 * number of milliseconds. Unless "time/stop" has also been executed, the
584 * monotonic clock continues to tick forward at the normal rate afterward.
586 * Does not affect wall clock readings. */
588 timeval_warp_cb(struct unixctl_conn *conn,
589 int argc OVS_UNUSED, const char *argv[], void *aux OVS_UNUSED)
594 msecs = atoi(argv[1]);
596 unixctl_command_reply_error(conn, "invalid MSECS");
600 ts.tv_sec = msecs / 1000;
601 ts.tv_nsec = (msecs % 1000) * 1000 * 1000;
602 timespec_add(&warp_offset, &warp_offset, &ts);
603 timespec_add(&monotonic_time, &monotonic_time, &ts);
604 unixctl_command_reply(conn, "warped");
608 timeval_dummy_register(void)
610 unixctl_command_register("time/stop", "", 0, 0, timeval_stop_cb, NULL);
611 unixctl_command_register("time/warp", "MSECS", 1, 1,
612 timeval_warp_cb, NULL);