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 /* The clock to use for measuring time intervals. This is CLOCK_MONOTONIC by
38 * preference, but on systems that don't have a monotonic clock we fall back
39 * to CLOCK_REALTIME. */
40 static clockid_t monotonic_clock;
42 /* Has a timer tick occurred? */
43 static volatile sig_atomic_t wall_tick;
44 static volatile sig_atomic_t monotonic_tick;
46 /* The current time, as of the last refresh. */
47 static struct timespec wall_time;
48 static struct timespec monotonic_time;
50 /* Time at which to die with SIGALRM (if not TIME_MIN). */
51 static time_t deadline = TIME_MIN;
53 static void set_up_timer(void);
54 static void set_up_signal(int flags);
55 static void sigalrm_handler(int);
56 static void refresh_wall_if_ticked(void);
57 static void refresh_monotonic_if_ticked(void);
58 static time_t time_add(time_t, time_t);
59 static void block_sigalrm(sigset_t *);
60 static void unblock_sigalrm(const sigset_t *);
61 static void log_poll_interval(long long int last_wakeup,
62 const struct rusage *last_rusage);
64 /* Initializes the timetracking module. */
77 if (!clock_gettime(CLOCK_MONOTONIC, &monotonic_time)) {
78 monotonic_clock = CLOCK_MONOTONIC;
80 monotonic_clock = CLOCK_REALTIME;
81 VLOG_DBG("monotonic timer not available");
84 set_up_signal(SA_RESTART);
89 set_up_signal(int flags)
93 memset(&sa, 0, sizeof sa);
94 sa.sa_handler = sigalrm_handler;
95 sigemptyset(&sa.sa_mask);
97 if (sigaction(SIGALRM, &sa, NULL)) {
98 ovs_fatal(errno, "sigaction(SIGALRM) failed");
102 /* Remove SA_RESTART from the flags for SIGALRM, so that any system call that
103 * is interrupted by the periodic timer interrupt will return EINTR instead of
104 * continuing after the signal handler returns.
106 * time_disable_restart() and time_enable_restart() may be usefully wrapped
107 * around function calls that might otherwise block forever unless interrupted
110 * time_disable_restart();
111 * fcntl(fd, F_SETLKW, &lock);
112 * time_enable_restart();
115 time_disable_restart(void)
120 /* Add SA_RESTART to the flags for SIGALRM, so that any system call that
121 * is interrupted by the periodic timer interrupt will continue after the
122 * signal handler returns instead of returning EINTR. */
124 time_enable_restart(void)
126 set_up_signal(SA_RESTART);
132 static timer_t timer_id; /* "static" to avoid apparent memory leak. */
133 struct itimerspec itimer;
135 if (timer_create(monotonic_clock, NULL, &timer_id)) {
136 ovs_fatal(errno, "timer_create failed");
139 itimer.it_interval.tv_sec = 0;
140 itimer.it_interval.tv_nsec = TIME_UPDATE_INTERVAL * 1000 * 1000;
141 itimer.it_value = itimer.it_interval;
143 if (timer_settime(timer_id, 0, &itimer, NULL)) {
144 ovs_fatal(errno, "timer_settime failed");
148 /* Set up the interval timer, to ensure that time advances even without calling
151 * A child created with fork() does not inherit the parent's interval timer, so
152 * this function needs to be called from the child after fork(). */
163 clock_gettime(CLOCK_REALTIME, &wall_time);
168 refresh_monotonic(void)
172 if (monotonic_clock == CLOCK_MONOTONIC) {
173 clock_gettime(monotonic_clock, &monotonic_time);
175 refresh_wall_if_ticked();
176 monotonic_time = wall_time;
179 monotonic_tick = false;
182 /* Forces a refresh of the current time from the kernel. It is not usually
183 * necessary to call this function, since the time will be refreshed
184 * automatically at least every TIME_UPDATE_INTERVAL milliseconds. */
188 wall_tick = monotonic_tick = true;
191 /* Returns a monotonic timer, in seconds. */
195 refresh_monotonic_if_ticked();
196 return monotonic_time.tv_sec;
199 /* Same as time_now() except does not write to static variables, for use in
200 * signal handlers. */
204 struct timespec cur_time;
206 clock_gettime(monotonic_clock, &cur_time);
207 return cur_time.tv_sec;
210 /* Returns the current time, in seconds. */
214 refresh_wall_if_ticked();
215 return wall_time.tv_sec;
218 /* Returns a monotonic timer, in ms (within TIME_UPDATE_INTERVAL ms). */
222 refresh_monotonic_if_ticked();
223 return timespec_to_msec(&monotonic_time);
226 /* Returns the current time, in ms (within TIME_UPDATE_INTERVAL ms). */
230 refresh_wall_if_ticked();
231 return timespec_to_msec(&wall_time);
234 /* Stores a monotonic timer, accurate within TIME_UPDATE_INTERVAL ms, into
237 time_timespec(struct timespec *ts)
239 refresh_monotonic_if_ticked();
240 *ts = monotonic_time;
243 /* Stores the current time, accurate within TIME_UPDATE_INTERVAL ms, into
246 time_wall_timespec(struct timespec *ts)
248 refresh_wall_if_ticked();
252 /* Configures the program to die with SIGALRM 'secs' seconds from now, if
253 * 'secs' is nonzero, or disables the feature if 'secs' is zero. */
255 time_alarm(unsigned int secs)
260 block_sigalrm(&oldsigs);
261 deadline = secs ? time_add(time_now(), secs) : TIME_MIN;
262 unblock_sigalrm(&oldsigs);
265 /* Like poll(), except:
267 * - On error, returns a negative error code (instead of setting errno).
269 * - If interrupted by a signal, retries automatically until the original
270 * 'timeout' expires. (Because of this property, this function will
271 * never return -EINTR.)
273 * - As a side effect, refreshes the current time (like time_refresh()).
276 time_poll(struct pollfd *pollfds, int n_pollfds, int timeout)
278 static long long int last_wakeup;
279 static struct rusage last_rusage;
286 log_poll_interval(last_wakeup, &last_rusage);
293 long long int elapsed = time_msec() - start;
294 time_left = timeout >= elapsed ? timeout - elapsed : 0;
299 retval = poll(pollfds, n_pollfds, time_left);
304 if (retval != -EINTR) {
308 if (!blocked && deadline == TIME_MIN) {
309 block_sigalrm(&oldsigs);
314 unblock_sigalrm(&oldsigs);
316 last_wakeup = time_msec();
317 getrusage(RUSAGE_SELF, &last_rusage);
321 /* Returns the sum of 'a' and 'b', with saturation on overflow or underflow. */
323 time_add(time_t a, time_t b)
326 ? (b > TIME_MAX - a ? TIME_MAX : a + b)
327 : (b < TIME_MIN - a ? TIME_MIN : a + b));
331 sigalrm_handler(int sig_nr)
334 monotonic_tick = true;
335 if (deadline != TIME_MIN && time_now_sig() > deadline) {
336 fatal_signal_handler(sig_nr);
341 refresh_wall_if_ticked(void)
350 refresh_monotonic_if_ticked(void)
353 if (monotonic_tick) {
359 block_sigalrm(sigset_t *oldsigs)
362 sigemptyset(&sigalrm);
363 sigaddset(&sigalrm, SIGALRM);
364 if (sigprocmask(SIG_BLOCK, &sigalrm, oldsigs)) {
365 ovs_fatal(errno, "sigprocmask");
370 unblock_sigalrm(const sigset_t *oldsigs)
372 if (sigprocmask(SIG_SETMASK, oldsigs, NULL)) {
373 ovs_fatal(errno, "sigprocmask");
378 timespec_to_msec(const struct timespec *ts)
380 return (long long int) ts->tv_sec * 1000 + ts->tv_nsec / (1000 * 1000);
384 timeval_to_msec(const struct timeval *tv)
386 return (long long int) tv->tv_sec * 1000 + tv->tv_usec / 1000;
390 timeval_diff_msec(const struct timeval *a, const struct timeval *b)
392 return timeval_to_msec(a) - timeval_to_msec(b);
396 log_poll_interval(long long int last_wakeup, const struct rusage *last_rusage)
398 static unsigned int mean_interval; /* In 16ths of a millisecond. */
399 static unsigned int n_samples;
402 unsigned int interval; /* In 16ths of a millisecond. */
404 /* Compute interval from last wakeup to now in 16ths of a millisecond,
405 * capped at 10 seconds (16000 in this unit). */
407 interval = MIN(10000, now - last_wakeup) << 4;
409 /* Warn if we took too much time between polls. */
410 if (n_samples > 10 && interval > mean_interval * 8) {
411 struct rusage rusage;
413 getrusage(RUSAGE_SELF, &rusage);
414 VLOG_WARN("%lld ms poll interval (%lld ms user, %lld ms system) "
415 "is over %u times the weighted mean interval %u ms "
418 timeval_diff_msec(&rusage.ru_utime, &last_rusage->ru_utime),
419 timeval_diff_msec(&rusage.ru_stime, &last_rusage->ru_stime),
420 interval / mean_interval,
421 (mean_interval + 8) / 16, n_samples);
422 if (rusage.ru_minflt > last_rusage->ru_minflt
423 || rusage.ru_majflt > last_rusage->ru_majflt) {
424 VLOG_WARN("faults: %ld minor, %ld major",
425 rusage.ru_minflt - last_rusage->ru_minflt,
426 rusage.ru_majflt - last_rusage->ru_majflt);
428 if (rusage.ru_inblock > last_rusage->ru_inblock
429 || rusage.ru_oublock > last_rusage->ru_oublock) {
430 VLOG_WARN("disk: %ld reads, %ld writes",
431 rusage.ru_inblock - last_rusage->ru_inblock,
432 rusage.ru_oublock - last_rusage->ru_oublock);
434 if (rusage.ru_nvcsw > last_rusage->ru_nvcsw
435 || rusage.ru_nivcsw > last_rusage->ru_nivcsw) {
436 VLOG_WARN("context switches: %ld voluntary, %ld involuntary",
437 rusage.ru_nvcsw - last_rusage->ru_nvcsw,
438 rusage.ru_nivcsw - last_rusage->ru_nivcsw);
441 /* Care should be taken in the value chosen for logging. Depending
442 * on the configuration, syslog can write changes synchronously,
443 * which can cause the coverage messages to take longer to log
444 * than the processing delay that triggered it. */
445 coverage_log(VLL_INFO, true);
448 /* Update exponentially weighted moving average. With these parameters, a
449 * given value decays to 1% of its value in about 100 time steps. */
451 mean_interval = (mean_interval * 122 + interval * 6 + 64) / 128;
453 mean_interval = interval;