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.
29 #include <sys/resource.h>
33 #include "dynamic-string.h"
34 #include "fatal-signal.h"
40 #ifndef HAVE_EXECINFO_H
41 #define HAVE_EXECINFO_H 0
44 VLOG_DEFINE_THIS_MODULE(timeval);
46 /* The clock to use for measuring time intervals. This is CLOCK_MONOTONIC by
47 * preference, but on systems that don't have a monotonic clock we fall back
48 * to CLOCK_REALTIME. */
49 static clockid_t monotonic_clock;
51 /* Has a timer tick occurred? Only relevant if CACHE_TIME is true.
53 * We initialize these to true to force time_init() to get called on the first
54 * call to time_msec() or another function that queries the current time. */
55 static volatile sig_atomic_t wall_tick = true;
56 static volatile sig_atomic_t monotonic_tick = true;
58 /* The current time, as of the last refresh. */
59 static struct timespec wall_time;
60 static struct timespec monotonic_time;
62 /* The monotonic time at which the time module was initialized. */
63 static long long int boot_time;
65 /* features for use by unit tests. */
66 static struct timespec warp_offset; /* Offset added to monotonic_time. */
67 static bool time_stopped; /* Disables real-time updates, if true. */
69 /* Time in milliseconds at which to die with SIGALRM (if not LLONG_MAX). */
70 static long long int deadline = LLONG_MAX;
73 void *backtrace[32]; /* Populated by backtrace(). */
74 size_t n_frames; /* Number of frames in 'backtrace'. */
78 static struct unixctl_conn *backtrace_conn = NULL;
79 static struct trace *traces = NULL;
80 static size_t n_traces = 0;
82 static void set_up_timer(void);
83 static void set_up_signal(int flags);
84 static void sigalrm_handler(int);
85 static void refresh_wall_if_ticked(void);
86 static void refresh_monotonic_if_ticked(void);
87 static void block_sigalrm(sigset_t *);
88 static void unblock_sigalrm(const sigset_t *);
89 static void log_poll_interval(long long int last_wakeup);
90 static struct rusage *get_recent_rusage(void);
91 static void refresh_rusage(void);
92 static void timespec_add(struct timespec *sum,
93 const struct timespec *a, const struct timespec *b);
94 static void trace_run(void);
95 static unixctl_cb_func backtrace_cb;
97 /* Initializes the timetracking module, if not already initialized. */
103 /* The best place to do this is probably a timeval_run() function.
104 * However, none exists and this function is usually so fast that doing it
105 * here seems fine for now. */
113 if (HAVE_EXECINFO_H && CACHE_TIME) {
114 unixctl_command_register("backtrace", "", 0, 0, backtrace_cb, NULL);
119 if (!clock_gettime(CLOCK_MONOTONIC, &monotonic_time)) {
120 monotonic_clock = CLOCK_MONOTONIC;
122 monotonic_clock = CLOCK_REALTIME;
123 VLOG_DBG("monotonic timer not available");
126 set_up_signal(SA_RESTART);
129 boot_time = time_msec();
133 set_up_signal(int flags)
137 memset(&sa, 0, sizeof sa);
138 sa.sa_handler = sigalrm_handler;
139 sigemptyset(&sa.sa_mask);
141 xsigaction(SIGALRM, &sa, NULL);
144 /* Remove SA_RESTART from the flags for SIGALRM, so that any system call that
145 * is interrupted by the periodic timer interrupt will return EINTR instead of
146 * continuing after the signal handler returns.
148 * time_disable_restart() and time_enable_restart() may be usefully wrapped
149 * around function calls that might otherwise block forever unless interrupted
152 * time_disable_restart();
153 * fcntl(fd, F_SETLKW, &lock);
154 * time_enable_restart();
157 time_disable_restart(void)
163 /* Add SA_RESTART to the flags for SIGALRM, so that any system call that
164 * is interrupted by the periodic timer interrupt will continue after the
165 * signal handler returns instead of returning EINTR. */
167 time_enable_restart(void)
170 set_up_signal(SA_RESTART);
176 static timer_t timer_id; /* "static" to avoid apparent memory leak. */
177 struct itimerspec itimer;
183 if (timer_create(monotonic_clock, NULL, &timer_id)) {
184 VLOG_FATAL("timer_create failed (%s)", strerror(errno));
187 itimer.it_interval.tv_sec = 0;
188 itimer.it_interval.tv_nsec = TIME_UPDATE_INTERVAL * 1000 * 1000;
189 itimer.it_value = itimer.it_interval;
191 if (timer_settime(timer_id, 0, &itimer, NULL)) {
192 VLOG_FATAL("timer_settime failed (%s)", strerror(errno));
196 /* Set up the interval timer, to ensure that time advances even without calling
199 * A child created with fork() does not inherit the parent's interval timer, so
200 * this function needs to be called from the child after fork(). */
212 clock_gettime(CLOCK_REALTIME, &wall_time);
217 refresh_monotonic(void)
222 if (monotonic_clock == CLOCK_MONOTONIC) {
223 clock_gettime(monotonic_clock, &monotonic_time);
225 refresh_wall_if_ticked();
226 monotonic_time = wall_time;
228 timespec_add(&monotonic_time, &monotonic_time, &warp_offset);
230 monotonic_tick = false;
234 /* Forces a refresh of the current time from the kernel. It is not usually
235 * necessary to call this function, since the time will be refreshed
236 * automatically at least every TIME_UPDATE_INTERVAL milliseconds. If
237 * CACHE_TIME is false, we will always refresh the current time so this
238 * function has no effect. */
242 wall_tick = monotonic_tick = true;
245 /* Returns a monotonic timer, in seconds. */
249 refresh_monotonic_if_ticked();
250 return monotonic_time.tv_sec;
253 /* Returns the current time, in seconds. */
257 refresh_wall_if_ticked();
258 return wall_time.tv_sec;
261 /* Returns a monotonic timer, in ms (within TIME_UPDATE_INTERVAL ms). */
265 refresh_monotonic_if_ticked();
266 return timespec_to_msec(&monotonic_time);
269 /* Returns the current time, in ms (within TIME_UPDATE_INTERVAL ms). */
273 refresh_wall_if_ticked();
274 return timespec_to_msec(&wall_time);
277 /* Stores a monotonic timer, accurate within TIME_UPDATE_INTERVAL ms, into
280 time_timespec(struct timespec *ts)
282 refresh_monotonic_if_ticked();
283 *ts = monotonic_time;
286 /* Stores the current time, accurate within TIME_UPDATE_INTERVAL ms, into
289 time_wall_timespec(struct timespec *ts)
291 refresh_wall_if_ticked();
295 /* Configures the program to die with SIGALRM 'secs' seconds from now, if
296 * 'secs' is nonzero, or disables the feature if 'secs' is zero. */
298 time_alarm(unsigned int secs)
311 block_sigalrm(&oldsigs);
312 deadline = now < LLONG_MAX - msecs ? now + msecs : LLONG_MAX;
313 unblock_sigalrm(&oldsigs);
316 /* Like poll(), except:
318 * - The timeout is specified as an absolute time, as defined by
319 * time_msec(), instead of a duration.
321 * - On error, returns a negative error code (instead of setting errno).
323 * - If interrupted by a signal, retries automatically until the original
324 * timeout is reached. (Because of this property, this function will
325 * never return -EINTR.)
327 * - As a side effect, refreshes the current time (like time_refresh()).
329 * Stores the number of milliseconds elapsed during poll in '*elapsed'. */
331 time_poll(struct pollfd *pollfds, int n_pollfds, long long int timeout_when,
334 static long long int last_wakeup;
341 log_poll_interval(last_wakeup);
346 timeout_when = MIN(timeout_when, deadline);
349 long long int now = time_msec();
352 if (now >= timeout_when) {
354 } else if ((unsigned long long int) timeout_when - now > INT_MAX) {
357 time_left = timeout_when - now;
360 retval = poll(pollfds, n_pollfds, time_left);
366 if (deadline <= time_msec()) {
367 fatal_signal_handler(SIGALRM);
374 if (retval != -EINTR) {
378 if (!blocked && CACHE_TIME && !backtrace_conn) {
379 block_sigalrm(&oldsigs);
384 unblock_sigalrm(&oldsigs);
386 last_wakeup = time_msec();
388 *elapsed = last_wakeup - start;
393 sigalrm_handler(int sig_nr OVS_UNUSED)
396 monotonic_tick = true;
399 if (backtrace_conn && n_traces < MAX_TRACES) {
400 struct trace *trace = &traces[n_traces++];
401 trace->n_frames = backtrace(trace->backtrace,
402 ARRAY_SIZE(trace->backtrace));
408 refresh_wall_if_ticked(void)
410 if (!CACHE_TIME || wall_tick) {
416 refresh_monotonic_if_ticked(void)
418 if (!CACHE_TIME || monotonic_tick) {
424 block_sigalrm(sigset_t *oldsigs)
427 sigemptyset(&sigalrm);
428 sigaddset(&sigalrm, SIGALRM);
429 xsigprocmask(SIG_BLOCK, &sigalrm, oldsigs);
433 unblock_sigalrm(const sigset_t *oldsigs)
435 xsigprocmask(SIG_SETMASK, oldsigs, NULL);
439 timespec_to_msec(const struct timespec *ts)
441 return (long long int) ts->tv_sec * 1000 + ts->tv_nsec / (1000 * 1000);
445 timeval_to_msec(const struct timeval *tv)
447 return (long long int) tv->tv_sec * 1000 + tv->tv_usec / 1000;
450 /* Returns the monotonic time at which the "time" module was initialized, in
460 xgettimeofday(struct timeval *tv)
462 if (gettimeofday(tv, NULL) == -1) {
463 VLOG_FATAL("gettimeofday failed (%s)", strerror(errno));
468 timeval_diff_msec(const struct timeval *a, const struct timeval *b)
470 return timeval_to_msec(a) - timeval_to_msec(b);
474 timespec_add(struct timespec *sum,
475 const struct timespec *a,
476 const struct timespec *b)
480 tmp.tv_sec = a->tv_sec + b->tv_sec;
481 tmp.tv_nsec = a->tv_nsec + b->tv_nsec;
482 if (tmp.tv_nsec >= 1000 * 1000 * 1000) {
483 tmp.tv_nsec -= 1000 * 1000 * 1000;
491 log_poll_interval(long long int last_wakeup)
493 static unsigned int mean_interval; /* In 16ths of a millisecond. */
494 static unsigned int n_samples;
497 unsigned int interval; /* In 16ths of a millisecond. */
499 /* Compute interval from last wakeup to now in 16ths of a millisecond,
500 * capped at 10 seconds (16000 in this unit). */
502 interval = MIN(10000, now - last_wakeup) << 4;
504 /* Warn if we took too much time between polls: at least 50 ms and at least
505 * 8X the mean interval. */
506 if (n_samples > 10 && interval > mean_interval * 8 && interval > 50 * 16) {
507 static struct vlog_rate_limit rl = VLOG_RATE_LIMIT_INIT(1, 3);
509 if (!VLOG_DROP_WARN(&rl)) {
510 const struct rusage *last_rusage = get_recent_rusage();
511 struct rusage rusage;
513 getrusage(RUSAGE_SELF, &rusage);
514 VLOG_WARN("%lld ms poll interval (%lld ms user, %lld ms system) "
515 "is over %u times the weighted mean interval %u ms "
518 timeval_diff_msec(&rusage.ru_utime,
519 &last_rusage->ru_utime),
520 timeval_diff_msec(&rusage.ru_stime,
521 &last_rusage->ru_stime),
522 interval / mean_interval,
523 (mean_interval + 8) / 16, n_samples);
524 if (rusage.ru_minflt > last_rusage->ru_minflt
525 || rusage.ru_majflt > last_rusage->ru_majflt) {
526 VLOG_WARN("faults: %ld minor, %ld major",
527 rusage.ru_minflt - last_rusage->ru_minflt,
528 rusage.ru_majflt - last_rusage->ru_majflt);
530 if (rusage.ru_inblock > last_rusage->ru_inblock
531 || rusage.ru_oublock > last_rusage->ru_oublock) {
532 VLOG_WARN("disk: %ld reads, %ld writes",
533 rusage.ru_inblock - last_rusage->ru_inblock,
534 rusage.ru_oublock - last_rusage->ru_oublock);
536 if (rusage.ru_nvcsw > last_rusage->ru_nvcsw
537 || rusage.ru_nivcsw > last_rusage->ru_nivcsw) {
538 VLOG_WARN("context switches: %ld voluntary, %ld involuntary",
539 rusage.ru_nvcsw - last_rusage->ru_nvcsw,
540 rusage.ru_nivcsw - last_rusage->ru_nivcsw);
546 /* Update exponentially weighted moving average. With these parameters, a
547 * given value decays to 1% of its value in about 100 time steps. */
549 mean_interval = (mean_interval * 122 + interval * 6 + 64) / 128;
551 mean_interval = interval;
555 /* CPU usage tracking. */
558 long long int when; /* Time that this sample was taken. */
559 unsigned long long int cpu; /* Total user+system CPU usage when sampled. */
562 static struct rusage recent_rusage;
563 static struct cpu_usage older = { LLONG_MIN, 0 };
564 static struct cpu_usage newer = { LLONG_MIN, 0 };
565 static int cpu_usage = -1;
567 static struct rusage *
568 get_recent_rusage(void)
570 return &recent_rusage;
579 getrusage(RUSAGE_SELF, &recent_rusage);
581 if (now >= newer.when + 3 * 1000) {
584 newer.cpu = (timeval_to_msec(&recent_rusage.ru_utime) +
585 timeval_to_msec(&recent_rusage.ru_stime));
587 if (older.when != LLONG_MIN && newer.cpu > older.cpu) {
588 unsigned int dividend = newer.cpu - older.cpu;
589 unsigned int divisor = (newer.when - older.when) / 100;
590 cpu_usage = divisor > 0 ? dividend / divisor : -1;
597 /* Returns an estimate of this process's CPU usage, as a percentage, over the
598 * past few seconds of wall-clock time. Returns -1 if no estimate is available
599 * (which will happen if the process has not been running long enough to have
600 * an estimate, and can happen for other reasons as well). */
611 if (backtrace_conn && n_traces >= MAX_TRACES) {
612 struct unixctl_conn *reply_conn = backtrace_conn;
613 struct ds ds = DS_EMPTY_INITIALIZER;
617 block_sigalrm(&oldsigs);
619 for (i = 0; i < n_traces; i++) {
620 struct trace *trace = &traces[i];
624 frame_strs = backtrace_symbols(trace->backtrace, trace->n_frames);
626 ds_put_format(&ds, "Backtrace %zu\n", i + 1);
627 for (j = 0; j < trace->n_frames; j++) {
628 ds_put_format(&ds, "%s\n", frame_strs[j]);
630 ds_put_cstr(&ds, "\n");
638 backtrace_conn = NULL;
640 unblock_sigalrm(&oldsigs);
642 unixctl_command_reply(reply_conn, ds_cstr(&ds));
648 /* Unixctl interface. */
650 /* "time/stop" stops the monotonic time returned by e.g. time_msec() from
651 * advancing, except due to later calls to "time/warp". */
653 timeval_stop_cb(struct unixctl_conn *conn,
654 int argc OVS_UNUSED, const char *argv[] OVS_UNUSED,
655 void *aux OVS_UNUSED)
658 unixctl_command_reply(conn, NULL);
661 /* "time/warp MSECS" advances the current monotonic time by the specified
662 * number of milliseconds. Unless "time/stop" has also been executed, the
663 * monotonic clock continues to tick forward at the normal rate afterward.
665 * Does not affect wall clock readings. */
667 timeval_warp_cb(struct unixctl_conn *conn,
668 int argc OVS_UNUSED, const char *argv[], void *aux OVS_UNUSED)
673 msecs = atoi(argv[1]);
675 unixctl_command_reply_error(conn, "invalid MSECS");
679 ts.tv_sec = msecs / 1000;
680 ts.tv_nsec = (msecs % 1000) * 1000 * 1000;
681 timespec_add(&warp_offset, &warp_offset, &ts);
682 timespec_add(&monotonic_time, &monotonic_time, &ts);
683 unixctl_command_reply(conn, "warped");
687 backtrace_cb(struct unixctl_conn *conn,
688 int argc OVS_UNUSED, const char *argv[] OVS_UNUSED,
689 void *aux OVS_UNUSED)
693 assert(HAVE_EXECINFO_H && CACHE_TIME);
695 if (backtrace_conn) {
696 unixctl_command_reply_error(conn, "In Use");
701 block_sigalrm(&oldsigs);
702 backtrace_conn = conn;
703 traces = xmalloc(MAX_TRACES * sizeof *traces);
705 unblock_sigalrm(&oldsigs);
709 timeval_dummy_register(void)
711 unixctl_command_register("time/stop", "", 0, 0, timeval_stop_cb, NULL);
712 unixctl_command_register("time/warp", "MSECS", 1, 1,
713 timeval_warp_cb, NULL);