2 * Copyright (c) 2008, 2009, 2010, 2011 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.
25 #include <sys/resource.h>
28 #include "fatal-signal.h"
33 VLOG_DEFINE_THIS_MODULE(timeval);
35 /* The clock to use for measuring time intervals. This is CLOCK_MONOTONIC by
36 * preference, but on systems that don't have a monotonic clock we fall back
37 * to CLOCK_REALTIME. */
38 static clockid_t monotonic_clock;
40 /* Has a timer tick occurred?
42 * We initialize these to true to force time_init() to get called on the first
43 * call to time_msec() or another function that queries the current time. */
44 static volatile sig_atomic_t wall_tick = true;
45 static volatile sig_atomic_t monotonic_tick = true;
47 /* The current time, as of the last refresh. */
48 static struct timespec wall_time;
49 static struct timespec monotonic_time;
51 /* Time at which to die with SIGALRM (if not TIME_MIN). */
52 static time_t deadline = TIME_MIN;
54 static void set_up_timer(void);
55 static void set_up_signal(int flags);
56 static void sigalrm_handler(int);
57 static void refresh_wall_if_ticked(void);
58 static void refresh_monotonic_if_ticked(void);
59 static time_t time_add(time_t, time_t);
60 static void block_sigalrm(sigset_t *);
61 static void unblock_sigalrm(const sigset_t *);
62 static void log_poll_interval(long long int last_wakeup,
63 const struct rusage *last_rusage);
65 /* Initializes the timetracking module.
67 * It is not necessary to call this function directly, because other time
68 * functions will call it automatically, but it doesn't hurt. */
80 if (!clock_gettime(CLOCK_MONOTONIC, &monotonic_time)) {
81 monotonic_clock = CLOCK_MONOTONIC;
83 monotonic_clock = CLOCK_REALTIME;
84 VLOG_DBG("monotonic timer not available");
87 set_up_signal(SA_RESTART);
92 set_up_signal(int flags)
96 memset(&sa, 0, sizeof sa);
97 sa.sa_handler = sigalrm_handler;
98 sigemptyset(&sa.sa_mask);
100 xsigaction(SIGALRM, &sa, NULL);
103 /* Remove SA_RESTART from the flags for SIGALRM, so that any system call that
104 * is interrupted by the periodic timer interrupt will return EINTR instead of
105 * continuing after the signal handler returns.
107 * time_disable_restart() and time_enable_restart() may be usefully wrapped
108 * around function calls that might otherwise block forever unless interrupted
111 * time_disable_restart();
112 * fcntl(fd, F_SETLKW, &lock);
113 * time_enable_restart();
116 time_disable_restart(void)
122 /* Add SA_RESTART to the flags for SIGALRM, so that any system call that
123 * is interrupted by the periodic timer interrupt will continue after the
124 * signal handler returns instead of returning EINTR. */
126 time_enable_restart(void)
129 set_up_signal(SA_RESTART);
135 static timer_t timer_id; /* "static" to avoid apparent memory leak. */
136 struct itimerspec itimer;
138 if (timer_create(monotonic_clock, NULL, &timer_id)) {
139 VLOG_FATAL("timer_create failed (%s)", strerror(errno));
142 itimer.it_interval.tv_sec = 0;
143 itimer.it_interval.tv_nsec = TIME_UPDATE_INTERVAL * 1000 * 1000;
144 itimer.it_value = itimer.it_interval;
146 if (timer_settime(timer_id, 0, &itimer, NULL)) {
147 VLOG_FATAL("timer_settime failed (%s)", strerror(errno));
151 /* Set up the interval timer, to ensure that time advances even without calling
154 * A child created with fork() does not inherit the parent's interval timer, so
155 * this function needs to be called from the child after fork(). */
167 clock_gettime(CLOCK_REALTIME, &wall_time);
172 refresh_monotonic(void)
176 if (monotonic_clock == CLOCK_MONOTONIC) {
177 clock_gettime(monotonic_clock, &monotonic_time);
179 refresh_wall_if_ticked();
180 monotonic_time = wall_time;
183 monotonic_tick = false;
186 /* Forces a refresh of the current time from the kernel. It is not usually
187 * necessary to call this function, since the time will be refreshed
188 * automatically at least every TIME_UPDATE_INTERVAL milliseconds. */
192 wall_tick = monotonic_tick = true;
195 /* Returns a monotonic timer, in seconds. */
199 refresh_monotonic_if_ticked();
200 return monotonic_time.tv_sec;
203 /* Same as time_now() except does not write to static variables, for use in
204 * signal handlers. */
208 struct timespec cur_time;
210 clock_gettime(monotonic_clock, &cur_time);
211 return cur_time.tv_sec;
214 /* Returns the current time, in seconds. */
218 refresh_wall_if_ticked();
219 return wall_time.tv_sec;
222 /* Returns a monotonic timer, in ms (within TIME_UPDATE_INTERVAL ms). */
226 refresh_monotonic_if_ticked();
227 return timespec_to_msec(&monotonic_time);
230 /* Returns the current time, in ms (within TIME_UPDATE_INTERVAL ms). */
234 refresh_wall_if_ticked();
235 return timespec_to_msec(&wall_time);
238 /* Stores a monotonic timer, accurate within TIME_UPDATE_INTERVAL ms, into
241 time_timespec(struct timespec *ts)
243 refresh_monotonic_if_ticked();
244 *ts = monotonic_time;
247 /* Stores the current time, accurate within TIME_UPDATE_INTERVAL ms, into
250 time_wall_timespec(struct timespec *ts)
252 refresh_wall_if_ticked();
256 /* Configures the program to die with SIGALRM 'secs' seconds from now, if
257 * 'secs' is nonzero, or disables the feature if 'secs' is zero. */
259 time_alarm(unsigned int secs)
264 block_sigalrm(&oldsigs);
265 deadline = secs ? time_add(time_now(), secs) : TIME_MIN;
266 unblock_sigalrm(&oldsigs);
269 /* Like poll(), except:
271 * - On error, returns a negative error code (instead of setting errno).
273 * - If interrupted by a signal, retries automatically until the original
274 * 'timeout' expires. (Because of this property, this function will
275 * never return -EINTR.)
277 * - As a side effect, refreshes the current time (like time_refresh()).
280 time_poll(struct pollfd *pollfds, int n_pollfds, int timeout)
282 static long long int last_wakeup;
283 static struct rusage last_rusage;
290 log_poll_interval(last_wakeup, &last_rusage);
297 long long int elapsed = time_msec() - start;
298 time_left = timeout >= elapsed ? timeout - elapsed : 0;
303 retval = poll(pollfds, n_pollfds, time_left);
308 if (retval != -EINTR) {
312 if (!blocked && deadline == TIME_MIN) {
313 block_sigalrm(&oldsigs);
318 unblock_sigalrm(&oldsigs);
320 last_wakeup = time_msec();
321 getrusage(RUSAGE_SELF, &last_rusage);
325 /* Returns the sum of 'a' and 'b', with saturation on overflow or underflow. */
327 time_add(time_t a, time_t b)
330 ? (b > TIME_MAX - a ? TIME_MAX : a + b)
331 : (b < TIME_MIN - a ? TIME_MIN : a + b));
335 sigalrm_handler(int sig_nr)
338 monotonic_tick = true;
339 if (deadline != TIME_MIN && time_now_sig() > deadline) {
340 fatal_signal_handler(sig_nr);
345 refresh_wall_if_ticked(void)
353 refresh_monotonic_if_ticked(void)
355 if (monotonic_tick) {
361 block_sigalrm(sigset_t *oldsigs)
364 sigemptyset(&sigalrm);
365 sigaddset(&sigalrm, SIGALRM);
366 xsigprocmask(SIG_BLOCK, &sigalrm, oldsigs);
370 unblock_sigalrm(const sigset_t *oldsigs)
372 xsigprocmask(SIG_SETMASK, oldsigs, NULL);
376 timespec_to_msec(const struct timespec *ts)
378 return (long long int) ts->tv_sec * 1000 + ts->tv_nsec / (1000 * 1000);
382 timeval_to_msec(const struct timeval *tv)
384 return (long long int) tv->tv_sec * 1000 + tv->tv_usec / 1000;
388 xgettimeofday(struct timeval *tv)
390 if (gettimeofday(tv, NULL) == -1) {
391 VLOG_FATAL("gettimeofday failed (%s)", strerror(errno));
396 timeval_diff_msec(const struct timeval *a, const struct timeval *b)
398 return timeval_to_msec(a) - timeval_to_msec(b);
402 log_poll_interval(long long int last_wakeup, const struct rusage *last_rusage)
404 static unsigned int mean_interval; /* In 16ths of a millisecond. */
405 static unsigned int n_samples;
408 unsigned int interval; /* In 16ths of a millisecond. */
410 /* Compute interval from last wakeup to now in 16ths of a millisecond,
411 * capped at 10 seconds (16000 in this unit). */
413 interval = MIN(10000, now - last_wakeup) << 4;
415 /* Warn if we took too much time between polls: at least 50 ms and at least
416 * 8X the mean interval. */
417 if (n_samples > 10 && interval > mean_interval * 8 && interval > 50 * 16) {
418 struct rusage rusage;
420 getrusage(RUSAGE_SELF, &rusage);
421 VLOG_WARN("%lld ms poll interval (%lld ms user, %lld ms system) "
422 "is over %u times the weighted mean interval %u ms "
425 timeval_diff_msec(&rusage.ru_utime, &last_rusage->ru_utime),
426 timeval_diff_msec(&rusage.ru_stime, &last_rusage->ru_stime),
427 interval / mean_interval,
428 (mean_interval + 8) / 16, n_samples);
429 if (rusage.ru_minflt > last_rusage->ru_minflt
430 || rusage.ru_majflt > last_rusage->ru_majflt) {
431 VLOG_WARN("faults: %ld minor, %ld major",
432 rusage.ru_minflt - last_rusage->ru_minflt,
433 rusage.ru_majflt - last_rusage->ru_majflt);
435 if (rusage.ru_inblock > last_rusage->ru_inblock
436 || rusage.ru_oublock > last_rusage->ru_oublock) {
437 VLOG_WARN("disk: %ld reads, %ld writes",
438 rusage.ru_inblock - last_rusage->ru_inblock,
439 rusage.ru_oublock - last_rusage->ru_oublock);
441 if (rusage.ru_nvcsw > last_rusage->ru_nvcsw
442 || rusage.ru_nivcsw > last_rusage->ru_nivcsw) {
443 VLOG_WARN("context switches: %ld voluntary, %ld involuntary",
444 rusage.ru_nvcsw - last_rusage->ru_nvcsw,
445 rusage.ru_nivcsw - last_rusage->ru_nivcsw);
448 /* Care should be taken in the value chosen for logging. Depending
449 * on the configuration, syslog can write changes synchronously,
450 * which can cause the coverage messages to take longer to log
451 * than the processing delay that triggered it. */
452 coverage_log(VLL_INFO, true);
455 /* Update exponentially weighted moving average. With these parameters, a
456 * given value decays to 1% of its value in about 100 time steps. */
458 mean_interval = (mean_interval * 122 + interval * 6 + 64) / 128;
460 mean_interval = interval;