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.
28 #include <sys/resource.h>
32 #include "dynamic-string.h"
33 #include "fatal-signal.h"
41 VLOG_DEFINE_THIS_MODULE(timeval);
43 /* The clock to use for measuring time intervals. This is CLOCK_MONOTONIC by
44 * preference, but on systems that don't have a monotonic clock we fall back
45 * to CLOCK_REALTIME. */
46 static clockid_t monotonic_clock;
48 /* Has a timer tick occurred? Only relevant if CACHE_TIME is true.
50 * We initialize these to true to force time_init() to get called on the first
51 * call to time_msec() or another function that queries the current time. */
52 static volatile sig_atomic_t wall_tick = true;
53 static volatile sig_atomic_t monotonic_tick = true;
55 /* The current time, as of the last refresh. */
56 static struct timespec wall_time;
57 static struct timespec monotonic_time;
59 /* The monotonic time at which the time module was initialized. */
60 static long long int boot_time;
62 /* features for use by unit tests. */
63 static struct timespec warp_offset; /* Offset added to monotonic_time. */
64 static bool time_stopped; /* Disables real-time updates, if true. */
66 /* Time in milliseconds at which to die with SIGALRM (if not LLONG_MAX). */
67 static long long int deadline = LLONG_MAX;
70 void *backtrace[32]; /* Populated by backtrace(). */
71 size_t n_frames; /* Number of frames in 'backtrace'. */
73 /* format_backtraces() helper data. */
74 struct hmap_node node;
79 static struct trace traces[MAX_TRACES];
80 static size_t trace_head = 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 unixctl_cb_func backtrace_cb;
96 #ifndef HAVE_EXECINFO_H
97 #define HAVE_EXECINFO_H 0
100 backtrace(void **buffer OVS_UNUSED, int size OVS_UNUSED)
106 backtrace_symbols(void *const *buffer OVS_UNUSED, int size OVS_UNUSED)
112 /* Initializes the timetracking module, if not already initialized. */
123 /* The implementation of backtrace() in glibc does some one time
124 * initialization which is not signal safe. This can cause deadlocks if
125 * run from the signal handler. As a workaround, force the initialization
127 if (HAVE_EXECINFO_H) {
130 backtrace(bt, ARRAY_SIZE(bt));
133 memset(traces, 0, sizeof traces);
135 if (HAVE_EXECINFO_H && CACHE_TIME) {
136 unixctl_command_register("backtrace", "", 0, 0, backtrace_cb, NULL);
141 if (!clock_gettime(CLOCK_MONOTONIC, &monotonic_time)) {
142 monotonic_clock = CLOCK_MONOTONIC;
144 monotonic_clock = CLOCK_REALTIME;
145 VLOG_DBG("monotonic timer not available");
148 set_up_signal(SA_RESTART);
151 boot_time = time_msec();
155 set_up_signal(int flags)
159 memset(&sa, 0, sizeof sa);
160 sa.sa_handler = sigalrm_handler;
161 sigemptyset(&sa.sa_mask);
163 xsigaction(SIGALRM, &sa, NULL);
166 /* Remove SA_RESTART from the flags for SIGALRM, so that any system call that
167 * is interrupted by the periodic timer interrupt will return EINTR instead of
168 * continuing after the signal handler returns.
170 * time_disable_restart() and time_enable_restart() may be usefully wrapped
171 * around function calls that might otherwise block forever unless interrupted
174 * time_disable_restart();
175 * fcntl(fd, F_SETLKW, &lock);
176 * time_enable_restart();
179 time_disable_restart(void)
185 /* Add SA_RESTART to the flags for SIGALRM, so that any system call that
186 * is interrupted by the periodic timer interrupt will continue after the
187 * signal handler returns instead of returning EINTR. */
189 time_enable_restart(void)
192 set_up_signal(SA_RESTART);
198 static timer_t timer_id; /* "static" to avoid apparent memory leak. */
199 struct itimerspec itimer;
205 if (timer_create(monotonic_clock, NULL, &timer_id)) {
206 VLOG_FATAL("timer_create failed (%s)", strerror(errno));
209 itimer.it_interval.tv_sec = 0;
210 itimer.it_interval.tv_nsec = TIME_UPDATE_INTERVAL * 1000 * 1000;
211 itimer.it_value = itimer.it_interval;
213 if (timer_settime(timer_id, 0, &itimer, NULL)) {
214 VLOG_FATAL("timer_settime failed (%s)", strerror(errno));
218 /* Set up the interval timer, to ensure that time advances even without calling
221 * A child created with fork() does not inherit the parent's interval timer, so
222 * this function needs to be called from the child after fork(). */
234 clock_gettime(CLOCK_REALTIME, &wall_time);
239 refresh_monotonic(void)
244 if (monotonic_clock == CLOCK_MONOTONIC) {
245 clock_gettime(monotonic_clock, &monotonic_time);
247 refresh_wall_if_ticked();
248 monotonic_time = wall_time;
250 timespec_add(&monotonic_time, &monotonic_time, &warp_offset);
252 monotonic_tick = false;
256 /* Forces a refresh of the current time from the kernel. It is not usually
257 * necessary to call this function, since the time will be refreshed
258 * automatically at least every TIME_UPDATE_INTERVAL milliseconds. If
259 * CACHE_TIME is false, we will always refresh the current time so this
260 * function has no effect. */
264 wall_tick = monotonic_tick = true;
267 /* Returns a monotonic timer, in seconds. */
271 refresh_monotonic_if_ticked();
272 return monotonic_time.tv_sec;
275 /* Returns the current time, in seconds. */
279 refresh_wall_if_ticked();
280 return wall_time.tv_sec;
283 /* Returns a monotonic timer, in ms (within TIME_UPDATE_INTERVAL ms). */
287 refresh_monotonic_if_ticked();
288 return timespec_to_msec(&monotonic_time);
291 /* Returns the current time, in ms (within TIME_UPDATE_INTERVAL ms). */
295 refresh_wall_if_ticked();
296 return timespec_to_msec(&wall_time);
299 /* Stores a monotonic timer, accurate within TIME_UPDATE_INTERVAL ms, into
302 time_timespec(struct timespec *ts)
304 refresh_monotonic_if_ticked();
305 *ts = monotonic_time;
308 /* Stores the current time, accurate within TIME_UPDATE_INTERVAL ms, into
311 time_wall_timespec(struct timespec *ts)
313 refresh_wall_if_ticked();
317 /* Configures the program to die with SIGALRM 'secs' seconds from now, if
318 * 'secs' is nonzero, or disables the feature if 'secs' is zero. */
320 time_alarm(unsigned int secs)
333 block_sigalrm(&oldsigs);
334 deadline = now < LLONG_MAX - msecs ? now + msecs : LLONG_MAX;
335 unblock_sigalrm(&oldsigs);
338 /* Like poll(), except:
340 * - The timeout is specified as an absolute time, as defined by
341 * time_msec(), instead of a duration.
343 * - On error, returns a negative error code (instead of setting errno).
345 * - If interrupted by a signal, retries automatically until the original
346 * timeout is reached. (Because of this property, this function will
347 * never return -EINTR.)
349 * - As a side effect, refreshes the current time (like time_refresh()).
351 * Stores the number of milliseconds elapsed during poll in '*elapsed'. */
353 time_poll(struct pollfd *pollfds, int n_pollfds, long long int timeout_when,
356 static long long int last_wakeup = 0;
364 log_poll_interval(last_wakeup);
370 timeout_when = MIN(timeout_when, deadline);
373 long long int now = time_msec();
376 if (now >= timeout_when) {
378 } else if ((unsigned long long int) timeout_when - now > INT_MAX) {
381 time_left = timeout_when - now;
384 retval = poll(pollfds, n_pollfds, time_left);
390 if (deadline <= time_msec()) {
391 fatal_signal_handler(SIGALRM);
398 if (retval != -EINTR) {
402 if (!blocked && CACHE_TIME) {
403 block_sigalrm(&oldsigs);
408 unblock_sigalrm(&oldsigs);
410 last_wakeup = time_msec();
412 *elapsed = last_wakeup - start;
417 sigalrm_handler(int sig_nr OVS_UNUSED)
420 monotonic_tick = true;
422 if (HAVE_EXECINFO_H && CACHE_TIME) {
423 struct trace *trace = &traces[trace_head];
425 trace->n_frames = backtrace(trace->backtrace,
426 ARRAY_SIZE(trace->backtrace));
427 trace_head = (trace_head + 1) % MAX_TRACES;
432 refresh_wall_if_ticked(void)
434 if (!CACHE_TIME || wall_tick) {
440 refresh_monotonic_if_ticked(void)
442 if (!CACHE_TIME || monotonic_tick) {
448 block_sigalrm(sigset_t *oldsigs)
451 sigemptyset(&sigalrm);
452 sigaddset(&sigalrm, SIGALRM);
453 xsigprocmask(SIG_BLOCK, &sigalrm, oldsigs);
457 unblock_sigalrm(const sigset_t *oldsigs)
459 xsigprocmask(SIG_SETMASK, oldsigs, NULL);
463 timespec_to_msec(const struct timespec *ts)
465 return (long long int) ts->tv_sec * 1000 + ts->tv_nsec / (1000 * 1000);
469 timeval_to_msec(const struct timeval *tv)
471 return (long long int) tv->tv_sec * 1000 + tv->tv_usec / 1000;
474 /* Returns the monotonic time at which the "time" module was initialized, in
484 xgettimeofday(struct timeval *tv)
486 if (gettimeofday(tv, NULL) == -1) {
487 VLOG_FATAL("gettimeofday failed (%s)", strerror(errno));
492 timeval_diff_msec(const struct timeval *a, const struct timeval *b)
494 return timeval_to_msec(a) - timeval_to_msec(b);
498 timespec_add(struct timespec *sum,
499 const struct timespec *a,
500 const struct timespec *b)
504 tmp.tv_sec = a->tv_sec + b->tv_sec;
505 tmp.tv_nsec = a->tv_nsec + b->tv_nsec;
506 if (tmp.tv_nsec >= 1000 * 1000 * 1000) {
507 tmp.tv_nsec -= 1000 * 1000 * 1000;
515 log_poll_interval(long long int last_wakeup)
517 long long int interval = time_msec() - last_wakeup;
519 if (interval >= 1000) {
520 const struct rusage *last_rusage = get_recent_rusage();
521 struct rusage rusage;
523 getrusage(RUSAGE_SELF, &rusage);
524 VLOG_WARN("Unreasonably long %lldms poll interval"
525 " (%lldms user, %lldms system)",
527 timeval_diff_msec(&rusage.ru_utime,
528 &last_rusage->ru_utime),
529 timeval_diff_msec(&rusage.ru_stime,
530 &last_rusage->ru_stime));
531 if (rusage.ru_minflt > last_rusage->ru_minflt
532 || rusage.ru_majflt > last_rusage->ru_majflt) {
533 VLOG_WARN("faults: %ld minor, %ld major",
534 rusage.ru_minflt - last_rusage->ru_minflt,
535 rusage.ru_majflt - last_rusage->ru_majflt);
537 if (rusage.ru_inblock > last_rusage->ru_inblock
538 || rusage.ru_oublock > last_rusage->ru_oublock) {
539 VLOG_WARN("disk: %ld reads, %ld writes",
540 rusage.ru_inblock - last_rusage->ru_inblock,
541 rusage.ru_oublock - last_rusage->ru_oublock);
543 if (rusage.ru_nvcsw > last_rusage->ru_nvcsw
544 || rusage.ru_nivcsw > last_rusage->ru_nivcsw) {
545 VLOG_WARN("context switches: %ld voluntary, %ld involuntary",
546 rusage.ru_nvcsw - last_rusage->ru_nvcsw,
547 rusage.ru_nivcsw - last_rusage->ru_nivcsw);
553 /* CPU usage tracking. */
556 long long int when; /* Time that this sample was taken. */
557 unsigned long long int cpu; /* Total user+system CPU usage when sampled. */
560 static struct rusage recent_rusage;
561 static struct cpu_usage older = { LLONG_MIN, 0 };
562 static struct cpu_usage newer = { LLONG_MIN, 0 };
563 static int cpu_usage = -1;
565 static struct rusage *
566 get_recent_rusage(void)
568 return &recent_rusage;
577 getrusage(RUSAGE_SELF, &recent_rusage);
579 if (now >= newer.when + 3 * 1000) {
582 newer.cpu = (timeval_to_msec(&recent_rusage.ru_utime) +
583 timeval_to_msec(&recent_rusage.ru_stime));
585 if (older.when != LLONG_MIN && newer.cpu > older.cpu) {
586 unsigned int dividend = newer.cpu - older.cpu;
587 unsigned int divisor = (newer.when - older.when) / 100;
588 cpu_usage = divisor > 0 ? dividend / divisor : -1;
595 /* Returns an estimate of this process's CPU usage, as a percentage, over the
596 * past few seconds of wall-clock time. Returns -1 if no estimate is available
597 * (which will happen if the process has not been running long enough to have
598 * an estimate, and can happen for other reasons as well). */
606 hash_trace(struct trace *trace)
608 return hash_bytes(trace->backtrace,
609 trace->n_frames * sizeof *trace->backtrace, 0);
612 static struct trace *
613 trace_map_lookup(struct hmap *trace_map, struct trace *key)
617 HMAP_FOR_EACH_WITH_HASH (value, node, hash_trace(key), trace_map) {
618 if (key->n_frames == value->n_frames
619 && !memcmp(key->backtrace, value->backtrace,
620 key->n_frames * sizeof *key->backtrace)) {
627 /* Appends a string to 'ds' representing backtraces recorded at regular
628 * intervals in the recent past. This information can be used to get a sense
629 * of what the process has been spending the majority of time doing. Will
630 * ommit any backtraces which have not occurred at least 'min_count' times. */
632 format_backtraces(struct ds *ds, size_t min_count)
636 if (HAVE_EXECINFO_H && CACHE_TIME) {
637 struct hmap trace_map = HMAP_INITIALIZER(&trace_map);
638 struct trace *trace, *next;
642 block_sigalrm(&oldsigs);
644 for (i = 0; i < MAX_TRACES; i++) {
645 struct trace *trace = &traces[i];
646 struct trace *map_trace;
648 if (!trace->n_frames) {
652 map_trace = trace_map_lookup(&trace_map, trace);
656 hmap_insert(&trace_map, &trace->node, hash_trace(trace));
661 HMAP_FOR_EACH_SAFE (trace, next, node, &trace_map) {
665 hmap_remove(&trace_map, &trace->node);
667 if (trace->count < min_count) {
671 frame_strs = backtrace_symbols(trace->backtrace, trace->n_frames);
673 ds_put_format(ds, "Count %zu\n", trace->count);
674 for (j = 0; j < trace->n_frames; j++) {
675 ds_put_format(ds, "%s\n", frame_strs[j]);
677 ds_put_cstr(ds, "\n");
681 hmap_destroy(&trace_map);
684 unblock_sigalrm(&oldsigs);
688 /* Unixctl interface. */
690 /* "time/stop" stops the monotonic time returned by e.g. time_msec() from
691 * advancing, except due to later calls to "time/warp". */
693 timeval_stop_cb(struct unixctl_conn *conn,
694 int argc OVS_UNUSED, const char *argv[] OVS_UNUSED,
695 void *aux OVS_UNUSED)
698 unixctl_command_reply(conn, NULL);
701 /* "time/warp MSECS" advances the current monotonic time by the specified
702 * number of milliseconds. Unless "time/stop" has also been executed, the
703 * monotonic clock continues to tick forward at the normal rate afterward.
705 * Does not affect wall clock readings. */
707 timeval_warp_cb(struct unixctl_conn *conn,
708 int argc OVS_UNUSED, const char *argv[], void *aux OVS_UNUSED)
713 msecs = atoi(argv[1]);
715 unixctl_command_reply_error(conn, "invalid MSECS");
719 ts.tv_sec = msecs / 1000;
720 ts.tv_nsec = (msecs % 1000) * 1000 * 1000;
721 timespec_add(&warp_offset, &warp_offset, &ts);
722 timespec_add(&monotonic_time, &monotonic_time, &ts);
723 unixctl_command_reply(conn, "warped");
727 backtrace_cb(struct unixctl_conn *conn,
728 int argc OVS_UNUSED, const char *argv[] OVS_UNUSED,
729 void *aux OVS_UNUSED)
731 struct ds ds = DS_EMPTY_INITIALIZER;
733 ovs_assert(HAVE_EXECINFO_H && CACHE_TIME);
734 format_backtraces(&ds, 0);
735 unixctl_command_reply(conn, ds_cstr(&ds));
740 timeval_dummy_register(void)
742 unixctl_command_register("time/stop", "", 0, 0, timeval_stop_cb, NULL);
743 unixctl_command_register("time/warp", "MSECS", 1, 1,
744 timeval_warp_cb, NULL);