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 1.
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");
97 set_up_signal(SA_RESTART);
101 boot_time = time_msec();
105 set_up_signal(int flags)
109 memset(&sa, 0, sizeof sa);
110 sa.sa_handler = sigalrm_handler;
111 sigemptyset(&sa.sa_mask);
113 xsigaction(SIGALRM, &sa, NULL);
116 /* Remove SA_RESTART from the flags for SIGALRM, so that any system call that
117 * is interrupted by the periodic timer interrupt will return EINTR instead of
118 * continuing after the signal handler returns.
120 * time_disable_restart() and time_enable_restart() may be usefully wrapped
121 * around function calls that might otherwise block forever unless interrupted
124 * time_disable_restart();
125 * fcntl(fd, F_SETLKW, &lock);
126 * time_enable_restart();
129 time_disable_restart(void)
135 /* Add SA_RESTART to the flags for SIGALRM, so that any system call that
136 * is interrupted by the periodic timer interrupt will continue after the
137 * signal handler returns instead of returning EINTR. */
139 time_enable_restart(void)
142 set_up_signal(SA_RESTART);
148 static timer_t timer_id; /* "static" to avoid apparent memory leak. */
149 struct itimerspec itimer;
151 if (timer_create(monotonic_clock, NULL, &timer_id)) {
152 VLOG_FATAL("timer_create failed (%s)", strerror(errno));
155 itimer.it_interval.tv_sec = 0;
156 itimer.it_interval.tv_nsec = TIME_UPDATE_INTERVAL * 1000 * 1000;
157 itimer.it_value = itimer.it_interval;
159 if (timer_settime(timer_id, 0, &itimer, NULL)) {
160 VLOG_FATAL("timer_settime failed (%s)", strerror(errno));
164 /* Set up the interval timer, to ensure that time advances even without calling
167 * A child created with fork() does not inherit the parent's interval timer, so
168 * this function needs to be called from the child after fork(). */
183 clock_gettime(CLOCK_REALTIME, &wall_time);
188 refresh_monotonic(void)
193 if (monotonic_clock == CLOCK_MONOTONIC) {
194 clock_gettime(monotonic_clock, &monotonic_time);
196 refresh_wall_if_ticked();
197 monotonic_time = wall_time;
199 timespec_add(&monotonic_time, &monotonic_time, &warp_offset);
201 monotonic_tick = false;
205 /* Forces a refresh of the current time from the kernel. It is not usually
206 * necessary to call this function, since the time will be refreshed
207 * automatically at least every TIME_UPDATE_INTERVAL milliseconds. If
208 * CACHE_TIME is 0, we will always refresh the current time so this
209 * function has no effect. */
213 wall_tick = monotonic_tick = true;
216 /* Returns a monotonic timer, in seconds. */
220 refresh_monotonic_if_ticked();
221 return monotonic_time.tv_sec;
224 /* Returns the current time, in seconds. */
228 refresh_wall_if_ticked();
229 return wall_time.tv_sec;
232 /* Returns a monotonic timer, in ms (within TIME_UPDATE_INTERVAL ms). */
236 refresh_monotonic_if_ticked();
237 return timespec_to_msec(&monotonic_time);
240 /* Returns the current time, in ms (within TIME_UPDATE_INTERVAL ms). */
244 refresh_wall_if_ticked();
245 return timespec_to_msec(&wall_time);
248 /* Stores a monotonic timer, accurate within TIME_UPDATE_INTERVAL ms, into
251 time_timespec(struct timespec *ts)
253 refresh_monotonic_if_ticked();
254 *ts = monotonic_time;
257 /* Stores the current time, accurate within TIME_UPDATE_INTERVAL ms, into
260 time_wall_timespec(struct timespec *ts)
262 refresh_wall_if_ticked();
266 /* Configures the program to die with SIGALRM 'secs' seconds from now, if
267 * 'secs' is nonzero, or disables the feature if 'secs' is zero. */
269 time_alarm(unsigned int secs)
282 block_sigalrm(&oldsigs);
283 deadline = now < LLONG_MAX - msecs ? now + msecs : LLONG_MAX;
284 unblock_sigalrm(&oldsigs);
287 /* Like poll(), except:
289 * - The timeout is specified as an absolute time, as defined by
290 * time_msec(), instead of a duration.
292 * - On error, returns a negative error code (instead of setting errno).
294 * - If interrupted by a signal, retries automatically until the original
295 * timeout is reached. (Because of this property, this function will
296 * never return -EINTR.)
298 * - As a side effect, refreshes the current time (like time_refresh()).
300 * Stores the number of milliseconds elapsed during poll in '*elapsed'. */
302 time_poll(struct pollfd *pollfds, int n_pollfds, long long int timeout_when,
305 static long long int last_wakeup;
312 log_poll_interval(last_wakeup);
317 timeout_when = MIN(timeout_when, deadline);
320 long long int now = time_msec();
323 if (now >= timeout_when) {
325 } else if ((unsigned long long int) timeout_when - now > INT_MAX) {
328 time_left = timeout_when - now;
331 retval = poll(pollfds, n_pollfds, time_left);
337 if (deadline <= time_msec()) {
338 fatal_signal_handler(SIGALRM);
345 if (retval != -EINTR) {
349 if (!blocked && !CACHE_TIME) {
350 block_sigalrm(&oldsigs);
355 unblock_sigalrm(&oldsigs);
357 last_wakeup = time_msec();
359 *elapsed = last_wakeup - start;
364 sigalrm_handler(int sig_nr OVS_UNUSED)
367 monotonic_tick = true;
371 refresh_wall_if_ticked(void)
373 if (!CACHE_TIME || wall_tick) {
379 refresh_monotonic_if_ticked(void)
381 if (!CACHE_TIME || monotonic_tick) {
387 block_sigalrm(sigset_t *oldsigs)
390 sigemptyset(&sigalrm);
391 sigaddset(&sigalrm, SIGALRM);
392 xsigprocmask(SIG_BLOCK, &sigalrm, oldsigs);
396 unblock_sigalrm(const sigset_t *oldsigs)
398 xsigprocmask(SIG_SETMASK, oldsigs, NULL);
402 timespec_to_msec(const struct timespec *ts)
404 return (long long int) ts->tv_sec * 1000 + ts->tv_nsec / (1000 * 1000);
408 timeval_to_msec(const struct timeval *tv)
410 return (long long int) tv->tv_sec * 1000 + tv->tv_usec / 1000;
413 /* Returns the monotonic time at which the "time" module was initialized, in
423 xgettimeofday(struct timeval *tv)
425 if (gettimeofday(tv, NULL) == -1) {
426 VLOG_FATAL("gettimeofday failed (%s)", strerror(errno));
431 timeval_diff_msec(const struct timeval *a, const struct timeval *b)
433 return timeval_to_msec(a) - timeval_to_msec(b);
437 timespec_add(struct timespec *sum,
438 const struct timespec *a,
439 const struct timespec *b)
443 tmp.tv_sec = a->tv_sec + b->tv_sec;
444 tmp.tv_nsec = a->tv_nsec + b->tv_nsec;
445 if (tmp.tv_nsec >= 1000 * 1000 * 1000) {
446 tmp.tv_nsec -= 1000 * 1000 * 1000;
454 log_poll_interval(long long int last_wakeup)
456 static unsigned int mean_interval; /* In 16ths of a millisecond. */
457 static unsigned int n_samples;
460 unsigned int interval; /* In 16ths of a millisecond. */
462 /* Compute interval from last wakeup to now in 16ths of a millisecond,
463 * capped at 10 seconds (16000 in this unit). */
465 interval = MIN(10000, now - last_wakeup) << 4;
467 /* Warn if we took too much time between polls: at least 50 ms and at least
468 * 8X the mean interval. */
469 if (n_samples > 10 && interval > mean_interval * 8 && interval > 50 * 16) {
470 static struct vlog_rate_limit rl = VLOG_RATE_LIMIT_INIT(1, 3);
472 if (!VLOG_DROP_WARN(&rl)) {
473 const struct rusage *last_rusage = get_recent_rusage();
474 struct rusage rusage;
476 getrusage(RUSAGE_SELF, &rusage);
477 VLOG_WARN("%lld ms poll interval (%lld ms user, %lld ms system) "
478 "is over %u times the weighted mean interval %u ms "
481 timeval_diff_msec(&rusage.ru_utime,
482 &last_rusage->ru_utime),
483 timeval_diff_msec(&rusage.ru_stime,
484 &last_rusage->ru_stime),
485 interval / mean_interval,
486 (mean_interval + 8) / 16, n_samples);
487 if (rusage.ru_minflt > last_rusage->ru_minflt
488 || rusage.ru_majflt > last_rusage->ru_majflt) {
489 VLOG_WARN("faults: %ld minor, %ld major",
490 rusage.ru_minflt - last_rusage->ru_minflt,
491 rusage.ru_majflt - last_rusage->ru_majflt);
493 if (rusage.ru_inblock > last_rusage->ru_inblock
494 || rusage.ru_oublock > last_rusage->ru_oublock) {
495 VLOG_WARN("disk: %ld reads, %ld writes",
496 rusage.ru_inblock - last_rusage->ru_inblock,
497 rusage.ru_oublock - last_rusage->ru_oublock);
499 if (rusage.ru_nvcsw > last_rusage->ru_nvcsw
500 || rusage.ru_nivcsw > last_rusage->ru_nivcsw) {
501 VLOG_WARN("context switches: %ld voluntary, %ld involuntary",
502 rusage.ru_nvcsw - last_rusage->ru_nvcsw,
503 rusage.ru_nivcsw - last_rusage->ru_nivcsw);
509 /* Update exponentially weighted moving average. With these parameters, a
510 * given value decays to 1% of its value in about 100 time steps. */
512 mean_interval = (mean_interval * 122 + interval * 6 + 64) / 128;
514 mean_interval = interval;
518 /* CPU usage tracking. */
521 long long int when; /* Time that this sample was taken. */
522 unsigned long long int cpu; /* Total user+system CPU usage when sampled. */
525 static struct rusage recent_rusage;
526 static struct cpu_usage older = { LLONG_MIN, 0 };
527 static struct cpu_usage newer = { LLONG_MIN, 0 };
528 static int cpu_usage = -1;
530 static struct rusage *
531 get_recent_rusage(void)
533 return &recent_rusage;
542 getrusage(RUSAGE_SELF, &recent_rusage);
544 if (now >= newer.when + 3 * 1000) {
547 newer.cpu = (timeval_to_msec(&recent_rusage.ru_utime) +
548 timeval_to_msec(&recent_rusage.ru_stime));
550 if (older.when != LLONG_MIN && newer.cpu > older.cpu) {
551 unsigned int dividend = newer.cpu - older.cpu;
552 unsigned int divisor = (newer.when - older.when) / 100;
553 cpu_usage = divisor > 0 ? dividend / divisor : -1;
560 /* Returns an estimate of this process's CPU usage, as a percentage, over the
561 * past few seconds of wall-clock time. Returns -1 if no estimate is available
562 * (which will happen if the process has not been running long enough to have
563 * an estimate, and can happen for other reasons as well). */
570 /* Unixctl interface. */
572 /* "time/stop" stops the monotonic time returned by e.g. time_msec() from
573 * advancing, except due to later calls to "time/warp". */
575 timeval_stop_cb(struct unixctl_conn *conn,
576 int argc OVS_UNUSED, const char *argv[] OVS_UNUSED,
577 void *aux OVS_UNUSED)
580 unixctl_command_reply(conn, NULL);
583 /* "time/warp MSECS" advances the current monotonic time by the specified
584 * number of milliseconds. Unless "time/stop" has also been executed, the
585 * monotonic clock continues to tick forward at the normal rate afterward.
587 * Does not affect wall clock readings. */
589 timeval_warp_cb(struct unixctl_conn *conn,
590 int argc OVS_UNUSED, const char *argv[], void *aux OVS_UNUSED)
595 msecs = atoi(argv[1]);
597 unixctl_command_reply_error(conn, "invalid MSECS");
601 ts.tv_sec = msecs / 1000;
602 ts.tv_nsec = (msecs % 1000) * 1000 * 1000;
603 timespec_add(&warp_offset, &warp_offset, &ts);
604 timespec_add(&monotonic_time, &monotonic_time, &ts);
605 unixctl_command_reply(conn, "warped");
609 timeval_dummy_register(void)
611 unixctl_command_register("time/stop", "", 0, 0, timeval_stop_cb, NULL);
612 unixctl_command_register("time/warp", "MSECS", 1, 1,
613 timeval_warp_cb, NULL);