2 * Copyright (c) 2008, 2009, 2010 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"
32 #define THIS_MODULE VLM_timeval
37 /* Does this system have monotonic timers? */
38 static bool monotonic_inited;
39 static clockid_t monotonic_clock;
41 /* Has a timer tick occurred? */
42 static volatile sig_atomic_t wall_tick;
43 static volatile sig_atomic_t monotonic_tick;
45 /* The current time, as of the last refresh. */
46 static struct timespec wall_time;
47 static struct timespec monotonic_time;
49 /* Time at which to die with SIGALRM (if not TIME_MIN). */
50 static time_t deadline = TIME_MIN;
52 static void set_up_timer(void);
53 static void set_up_signal(int flags);
54 static void sigalrm_handler(int);
55 static void refresh_wall_if_ticked(void);
56 static void refresh_monotonic_if_ticked(void);
57 static time_t time_add(time_t, time_t);
58 static void block_sigalrm(sigset_t *);
59 static void unblock_sigalrm(const sigset_t *);
60 static void log_poll_interval(long long int last_wakeup,
61 const struct rusage *last_rusage);
63 /* Initializes the timetracking module. */
76 set_up_signal(SA_RESTART);
81 set_up_monotonic(void)
85 if (monotonic_inited) {
89 err = clock_gettime(CLOCK_MONOTONIC, &monotonic_time);
91 monotonic_clock = CLOCK_MONOTONIC;
93 monotonic_clock = CLOCK_REALTIME;
94 VLOG_DBG("monotonic timer not available");
97 monotonic_inited = true;
101 set_up_signal(int flags)
105 memset(&sa, 0, sizeof sa);
106 sa.sa_handler = sigalrm_handler;
107 sigemptyset(&sa.sa_mask);
109 if (sigaction(SIGALRM, &sa, NULL)) {
110 ovs_fatal(errno, "sigaction(SIGALRM) failed");
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)
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)
138 set_up_signal(SA_RESTART);
145 struct itimerspec itimer;
149 if (timer_create(monotonic_clock, NULL, &timer_id)) {
150 ovs_fatal(errno, "timer_create failed");
153 itimer.it_interval.tv_sec = 0;
154 itimer.it_interval.tv_nsec = TIME_UPDATE_INTERVAL * 1000 * 1000;
155 itimer.it_value = itimer.it_interval;
157 if (timer_settime(timer_id, 0, &itimer, NULL)) {
158 ovs_fatal(errno, "timer_settime failed");
162 /* Set up the interval timer, to ensure that time advances even without calling
165 * A child created with fork() does not inherit the parent's interval timer, so
166 * this function needs to be called from the child after fork(). */
176 clock_gettime(CLOCK_REALTIME, &wall_time);
181 refresh_monotonic(void)
185 if (monotonic_clock == CLOCK_MONOTONIC) {
186 clock_gettime(monotonic_clock, &monotonic_time);
188 refresh_wall_if_ticked();
189 monotonic_time = wall_time;
192 monotonic_tick = false;
195 /* Forces a refresh of the current time from the kernel. It is not usually
196 * necessary to call this function, since the time will be refreshed
197 * automatically at least every TIME_UPDATE_INTERVAL milliseconds. */
201 wall_tick = monotonic_tick = true;
204 /* Returns a monotonic timer, in seconds. */
208 refresh_monotonic_if_ticked();
209 return monotonic_time.tv_sec;
212 /* Same as time_now() except does not write to static variables, for use in
213 * signal handlers. set_up_monotonic() must have already been called. */
217 struct timespec cur_time;
219 clock_gettime(monotonic_clock, &cur_time);
220 return cur_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)
273 block_sigalrm(&oldsigs);
274 deadline = secs ? time_add(time_now(), secs) : TIME_MIN;
275 unblock_sigalrm(&oldsigs);
278 /* Like poll(), except:
280 * - On error, returns a negative error code (instead of setting errno).
282 * - If interrupted by a signal, retries automatically until the original
283 * 'timeout' expires. (Because of this property, this function will
284 * never return -EINTR.)
286 * - As a side effect, refreshes the current time (like time_refresh()).
289 time_poll(struct pollfd *pollfds, int n_pollfds, int timeout)
291 static long long int last_wakeup;
292 static struct rusage last_rusage;
299 log_poll_interval(last_wakeup, &last_rusage);
306 long long int elapsed = time_msec() - start;
307 time_left = timeout >= elapsed ? timeout - elapsed : 0;
312 retval = poll(pollfds, n_pollfds, time_left);
317 if (retval != -EINTR) {
321 if (!blocked && deadline == TIME_MIN) {
322 block_sigalrm(&oldsigs);
327 unblock_sigalrm(&oldsigs);
329 last_wakeup = time_msec();
330 getrusage(RUSAGE_SELF, &last_rusage);
334 /* Returns the sum of 'a' and 'b', with saturation on overflow or underflow. */
336 time_add(time_t a, time_t b)
339 ? (b > TIME_MAX - a ? TIME_MAX : a + b)
340 : (b < TIME_MIN - a ? TIME_MIN : a + b));
344 sigalrm_handler(int sig_nr)
347 monotonic_tick = true;
348 if (deadline != TIME_MIN && time_now_sig() > deadline) {
349 fatal_signal_handler(sig_nr);
354 refresh_wall_if_ticked(void)
363 refresh_monotonic_if_ticked(void)
366 if (monotonic_tick) {
372 block_sigalrm(sigset_t *oldsigs)
375 sigemptyset(&sigalrm);
376 sigaddset(&sigalrm, SIGALRM);
377 if (sigprocmask(SIG_BLOCK, &sigalrm, oldsigs)) {
378 ovs_fatal(errno, "sigprocmask");
383 unblock_sigalrm(const sigset_t *oldsigs)
385 if (sigprocmask(SIG_SETMASK, oldsigs, NULL)) {
386 ovs_fatal(errno, "sigprocmask");
391 timespec_to_msec(const struct timespec *ts)
393 return (long long int) ts->tv_sec * 1000 + ts->tv_nsec / (1000 * 1000);
397 timeval_to_msec(const struct timeval *tv)
399 return (long long int) tv->tv_sec * 1000 + tv->tv_usec / 1000;
403 timeval_diff_msec(const struct timeval *a, const struct timeval *b)
405 return timeval_to_msec(a) - timeval_to_msec(b);
409 log_poll_interval(long long int last_wakeup, const struct rusage *last_rusage)
411 static unsigned int mean_interval; /* In 16ths of a millisecond. */
412 static unsigned int n_samples;
415 unsigned int interval; /* In 16ths of a millisecond. */
417 /* Compute interval from last wakeup to now in 16ths of a millisecond,
418 * capped at 10 seconds (16000 in this unit). */
420 interval = MIN(10000, now - last_wakeup) << 4;
422 /* Warn if we took too much time between polls. */
423 if (n_samples > 10 && interval > mean_interval * 8) {
424 struct rusage rusage;
426 getrusage(RUSAGE_SELF, &rusage);
427 VLOG_WARN("%lld ms poll interval (%lld ms user, %lld ms system) "
428 "is over %u times the weighted mean interval %u ms "
431 timeval_diff_msec(&rusage.ru_utime, &last_rusage->ru_utime),
432 timeval_diff_msec(&rusage.ru_stime, &last_rusage->ru_stime),
433 interval / mean_interval,
434 (mean_interval + 8) / 16, n_samples);
435 if (rusage.ru_minflt > last_rusage->ru_minflt
436 || rusage.ru_majflt > last_rusage->ru_majflt) {
437 VLOG_WARN("faults: %ld minor, %ld major",
438 rusage.ru_minflt - last_rusage->ru_minflt,
439 rusage.ru_majflt - last_rusage->ru_majflt);
441 if (rusage.ru_inblock > last_rusage->ru_inblock
442 || rusage.ru_oublock > last_rusage->ru_oublock) {
443 VLOG_WARN("disk: %ld reads, %ld writes",
444 rusage.ru_inblock - last_rusage->ru_inblock,
445 rusage.ru_oublock - last_rusage->ru_oublock);
447 if (rusage.ru_nvcsw > last_rusage->ru_nvcsw
448 || rusage.ru_nivcsw > last_rusage->ru_nivcsw) {
449 VLOG_WARN("context switches: %ld voluntary, %ld involuntary",
450 rusage.ru_nvcsw - last_rusage->ru_nvcsw,
451 rusage.ru_nivcsw - last_rusage->ru_nivcsw);
454 /* Care should be taken in the value chosen for logging. Depending
455 * on the configuration, syslog can write changes synchronously,
456 * which can cause the coverage messages to take longer to log
457 * than the processing delay that triggered it. */
458 coverage_log(VLL_INFO, true);
461 /* Update exponentially weighted moving average. With these parameters, a
462 * given value decays to 1% of its value in about 100 time steps. */
464 mean_interval = (mean_interval * 122 + interval * 6 + 64) / 128;
466 mean_interval = interval;