2 * Copyright (c) 2008, 2009, 2010, 2011, 2012, 2013 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.
25 #include <sys/resource.h>
29 #include "dynamic-string.h"
30 #include "fatal-signal.h"
38 /* backtrace() from <execinfo.h> is really useful, but it is not signal safe
39 * everywhere, such as on x86-64. */
40 #if HAVE_BACKTRACE && !defined __x86_64__
41 # define USE_BACKTRACE 1
42 # include <execinfo.h>
44 # define USE_BACKTRACE 0
47 VLOG_DEFINE_THIS_MODULE(timeval);
49 /* The clock to use for measuring time intervals. This is CLOCK_MONOTONIC by
50 * preference, but on systems that don't have a monotonic clock we fall back
51 * to CLOCK_REALTIME. */
52 static clockid_t monotonic_clock;
54 /* Has a timer tick occurred? Only relevant if CACHE_TIME is true.
56 * We initialize these to true to force time_init() to get called on the first
57 * call to time_msec() or another function that queries the current time. */
58 static volatile sig_atomic_t wall_tick = true;
59 static volatile sig_atomic_t monotonic_tick = true;
61 /* The current time, as of the last refresh. */
62 static struct timespec wall_time;
63 static struct timespec monotonic_time;
65 /* The monotonic time at which the time module was initialized. */
66 static long long int boot_time;
68 /* features for use by unit tests. */
69 static struct timespec warp_offset; /* Offset added to monotonic_time. */
70 static bool time_stopped; /* Disables real-time updates, if true. */
72 /* Time in milliseconds at which to die with SIGALRM (if not LLONG_MAX). */
73 static long long int deadline = LLONG_MAX;
76 void *backtrace[32]; /* Populated by backtrace(). */
77 size_t n_frames; /* Number of frames in 'backtrace'. */
79 /* format_backtraces() helper data. */
80 struct hmap_node node;
85 static struct trace traces[MAX_TRACES];
86 static size_t trace_head = 0;
88 static void set_up_timer(void);
89 static void set_up_signal(int flags);
90 static void sigalrm_handler(int);
91 static void refresh_wall_if_ticked(void);
92 static void refresh_monotonic_if_ticked(void);
93 static void block_sigalrm(sigset_t *);
94 static void unblock_sigalrm(const sigset_t *);
95 static void log_poll_interval(long long int last_wakeup);
96 static struct rusage *get_recent_rusage(void);
97 static void refresh_rusage(void);
98 static void timespec_add(struct timespec *sum,
99 const struct timespec *a, const struct timespec *b);
100 static unixctl_cb_func backtrace_cb;
104 backtrace(void **buffer OVS_UNUSED, int size OVS_UNUSED)
110 backtrace_symbols(void *const *buffer OVS_UNUSED, int size OVS_UNUSED)
114 #endif /* !USE_BACKTRACE */
116 /* Initializes the timetracking module, if not already initialized. */
127 /* The implementation of backtrace() in glibc does some one time
128 * initialization which is not signal safe. This can cause deadlocks if
129 * run from the signal handler. As a workaround, force the initialization
134 backtrace(bt, ARRAY_SIZE(bt));
137 memset(traces, 0, sizeof traces);
139 if (USE_BACKTRACE && CACHE_TIME) {
140 unixctl_command_register("backtrace", "", 0, 0, backtrace_cb, NULL);
145 if (!clock_gettime(CLOCK_MONOTONIC, &monotonic_time)) {
146 monotonic_clock = CLOCK_MONOTONIC;
148 monotonic_clock = CLOCK_REALTIME;
149 VLOG_DBG("monotonic timer not available");
152 set_up_signal(SA_RESTART);
155 boot_time = time_msec();
159 set_up_signal(int flags)
163 memset(&sa, 0, sizeof sa);
164 sa.sa_handler = sigalrm_handler;
165 sigemptyset(&sa.sa_mask);
167 xsigaction(SIGALRM, &sa, NULL);
173 static timer_t timer_id; /* "static" to avoid apparent memory leak. */
174 struct itimerspec itimer;
180 if (timer_create(monotonic_clock, NULL, &timer_id)) {
181 VLOG_FATAL("timer_create failed (%s)", strerror(errno));
184 itimer.it_interval.tv_sec = 0;
185 itimer.it_interval.tv_nsec = TIME_UPDATE_INTERVAL * 1000 * 1000;
186 itimer.it_value = itimer.it_interval;
188 if (timer_settime(timer_id, 0, &itimer, NULL)) {
189 VLOG_FATAL("timer_settime failed (%s)", strerror(errno));
193 /* Set up the interval timer, to ensure that time advances even without calling
196 * A child created with fork() does not inherit the parent's interval timer, so
197 * this function needs to be called from the child after fork(). */
209 clock_gettime(CLOCK_REALTIME, &wall_time);
214 refresh_monotonic(void)
219 if (monotonic_clock == CLOCK_MONOTONIC) {
220 clock_gettime(monotonic_clock, &monotonic_time);
222 refresh_wall_if_ticked();
223 monotonic_time = wall_time;
225 timespec_add(&monotonic_time, &monotonic_time, &warp_offset);
227 monotonic_tick = false;
231 /* Forces a refresh of the current time from the kernel. It is not usually
232 * necessary to call this function, since the time will be refreshed
233 * automatically at least every TIME_UPDATE_INTERVAL milliseconds. If
234 * CACHE_TIME is false, we will always refresh the current time so this
235 * function has no effect. */
239 wall_tick = monotonic_tick = true;
242 /* Returns a monotonic timer, in seconds. */
246 refresh_monotonic_if_ticked();
247 return monotonic_time.tv_sec;
250 /* Returns the current time, in seconds. */
254 refresh_wall_if_ticked();
255 return wall_time.tv_sec;
258 /* Returns a monotonic timer, in ms (within TIME_UPDATE_INTERVAL ms). */
262 refresh_monotonic_if_ticked();
263 return timespec_to_msec(&monotonic_time);
266 /* Returns the current time, in ms (within TIME_UPDATE_INTERVAL ms). */
270 refresh_wall_if_ticked();
271 return timespec_to_msec(&wall_time);
274 /* Stores a monotonic timer, accurate within TIME_UPDATE_INTERVAL ms, into
277 time_timespec(struct timespec *ts)
279 refresh_monotonic_if_ticked();
280 *ts = monotonic_time;
283 /* Stores the current time, accurate within TIME_UPDATE_INTERVAL ms, into
286 time_wall_timespec(struct timespec *ts)
288 refresh_wall_if_ticked();
292 /* Configures the program to die with SIGALRM 'secs' seconds from now, if
293 * 'secs' is nonzero, or disables the feature if 'secs' is zero. */
295 time_alarm(unsigned int secs)
304 msecs = secs * 1000LL;
305 deadline = now < LLONG_MAX - msecs ? now + msecs : LLONG_MAX;
308 /* Like poll(), except:
310 * - The timeout is specified as an absolute time, as defined by
311 * time_msec(), instead of a duration.
313 * - On error, returns a negative error code (instead of setting errno).
315 * - If interrupted by a signal, retries automatically until the original
316 * timeout is reached. (Because of this property, this function will
317 * never return -EINTR.)
319 * - As a side effect, refreshes the current time (like time_refresh()).
321 * Stores the number of milliseconds elapsed during poll in '*elapsed'. */
323 time_poll(struct pollfd *pollfds, int n_pollfds, long long int timeout_when,
326 static long long int last_wakeup = 0;
334 log_poll_interval(last_wakeup);
340 timeout_when = MIN(timeout_when, deadline);
343 long long int now = time_msec();
346 if (now >= timeout_when) {
348 } else if ((unsigned long long int) timeout_when - now > INT_MAX) {
351 time_left = timeout_when - now;
354 retval = poll(pollfds, n_pollfds, time_left);
360 if (deadline <= time_msec()) {
361 fatal_signal_handler(SIGALRM);
368 if (retval != -EINTR) {
372 if (!blocked && CACHE_TIME) {
373 block_sigalrm(&oldsigs);
378 unblock_sigalrm(&oldsigs);
380 last_wakeup = time_msec();
382 *elapsed = last_wakeup - start;
387 sigalrm_handler(int sig_nr OVS_UNUSED)
390 monotonic_tick = true;
392 if (USE_BACKTRACE && CACHE_TIME) {
393 struct trace *trace = &traces[trace_head];
395 trace->n_frames = backtrace(trace->backtrace,
396 ARRAY_SIZE(trace->backtrace));
397 trace_head = (trace_head + 1) % MAX_TRACES;
402 refresh_wall_if_ticked(void)
404 if (!CACHE_TIME || wall_tick) {
410 refresh_monotonic_if_ticked(void)
412 if (!CACHE_TIME || monotonic_tick) {
418 block_sigalrm(sigset_t *oldsigs)
421 sigemptyset(&sigalrm);
422 sigaddset(&sigalrm, SIGALRM);
423 xpthread_sigmask(SIG_BLOCK, &sigalrm, oldsigs);
427 unblock_sigalrm(const sigset_t *oldsigs)
429 xpthread_sigmask(SIG_SETMASK, oldsigs, NULL);
433 timespec_to_msec(const struct timespec *ts)
435 return (long long int) ts->tv_sec * 1000 + ts->tv_nsec / (1000 * 1000);
439 timeval_to_msec(const struct timeval *tv)
441 return (long long int) tv->tv_sec * 1000 + tv->tv_usec / 1000;
444 /* Returns the monotonic time at which the "time" module was initialized, in
454 xgettimeofday(struct timeval *tv)
456 if (gettimeofday(tv, NULL) == -1) {
457 VLOG_FATAL("gettimeofday failed (%s)", strerror(errno));
462 timeval_diff_msec(const struct timeval *a, const struct timeval *b)
464 return timeval_to_msec(a) - timeval_to_msec(b);
468 timespec_add(struct timespec *sum,
469 const struct timespec *a,
470 const struct timespec *b)
474 tmp.tv_sec = a->tv_sec + b->tv_sec;
475 tmp.tv_nsec = a->tv_nsec + b->tv_nsec;
476 if (tmp.tv_nsec >= 1000 * 1000 * 1000) {
477 tmp.tv_nsec -= 1000 * 1000 * 1000;
485 log_poll_interval(long long int last_wakeup)
487 long long int interval = time_msec() - last_wakeup;
489 if (interval >= 1000 && !warp_offset.tv_sec && !warp_offset.tv_nsec) {
490 const struct rusage *last_rusage = get_recent_rusage();
491 struct rusage rusage;
493 getrusage(RUSAGE_SELF, &rusage);
494 VLOG_WARN("Unreasonably long %lldms poll interval"
495 " (%lldms user, %lldms system)",
497 timeval_diff_msec(&rusage.ru_utime,
498 &last_rusage->ru_utime),
499 timeval_diff_msec(&rusage.ru_stime,
500 &last_rusage->ru_stime));
501 if (rusage.ru_minflt > last_rusage->ru_minflt
502 || rusage.ru_majflt > last_rusage->ru_majflt) {
503 VLOG_WARN("faults: %ld minor, %ld major",
504 rusage.ru_minflt - last_rusage->ru_minflt,
505 rusage.ru_majflt - last_rusage->ru_majflt);
507 if (rusage.ru_inblock > last_rusage->ru_inblock
508 || rusage.ru_oublock > last_rusage->ru_oublock) {
509 VLOG_WARN("disk: %ld reads, %ld writes",
510 rusage.ru_inblock - last_rusage->ru_inblock,
511 rusage.ru_oublock - last_rusage->ru_oublock);
513 if (rusage.ru_nvcsw > last_rusage->ru_nvcsw
514 || rusage.ru_nivcsw > last_rusage->ru_nivcsw) {
515 VLOG_WARN("context switches: %ld voluntary, %ld involuntary",
516 rusage.ru_nvcsw - last_rusage->ru_nvcsw,
517 rusage.ru_nivcsw - last_rusage->ru_nivcsw);
523 /* CPU usage tracking. */
526 long long int when; /* Time that this sample was taken. */
527 unsigned long long int cpu; /* Total user+system CPU usage when sampled. */
530 static struct rusage recent_rusage;
531 static struct cpu_usage older = { LLONG_MIN, 0 };
532 static struct cpu_usage newer = { LLONG_MIN, 0 };
533 static int cpu_usage = -1;
535 static struct rusage *
536 get_recent_rusage(void)
538 return &recent_rusage;
547 getrusage(RUSAGE_SELF, &recent_rusage);
549 if (now >= newer.when + 3 * 1000) {
552 newer.cpu = (timeval_to_msec(&recent_rusage.ru_utime) +
553 timeval_to_msec(&recent_rusage.ru_stime));
555 if (older.when != LLONG_MIN && newer.cpu > older.cpu) {
556 unsigned int dividend = newer.cpu - older.cpu;
557 unsigned int divisor = (newer.when - older.when) / 100;
558 cpu_usage = divisor > 0 ? dividend / divisor : -1;
565 /* Returns an estimate of this process's CPU usage, as a percentage, over the
566 * past few seconds of wall-clock time. Returns -1 if no estimate is available
567 * (which will happen if the process has not been running long enough to have
568 * an estimate, and can happen for other reasons as well). */
576 hash_trace(struct trace *trace)
578 return hash_bytes(trace->backtrace,
579 trace->n_frames * sizeof *trace->backtrace, 0);
582 static struct trace *
583 trace_map_lookup(struct hmap *trace_map, struct trace *key)
587 HMAP_FOR_EACH_WITH_HASH (value, node, hash_trace(key), trace_map) {
588 if (key->n_frames == value->n_frames
589 && !memcmp(key->backtrace, value->backtrace,
590 key->n_frames * sizeof *key->backtrace)) {
597 /* Appends a string to 'ds' representing backtraces recorded at regular
598 * intervals in the recent past. This information can be used to get a sense
599 * of what the process has been spending the majority of time doing. Will
600 * ommit any backtraces which have not occurred at least 'min_count' times. */
602 format_backtraces(struct ds *ds, size_t min_count)
606 if (USE_BACKTRACE && CACHE_TIME) {
607 struct hmap trace_map = HMAP_INITIALIZER(&trace_map);
608 struct trace *trace, *next;
612 block_sigalrm(&oldsigs);
614 for (i = 0; i < MAX_TRACES; i++) {
615 struct trace *trace = &traces[i];
616 struct trace *map_trace;
618 if (!trace->n_frames) {
622 map_trace = trace_map_lookup(&trace_map, trace);
626 hmap_insert(&trace_map, &trace->node, hash_trace(trace));
631 HMAP_FOR_EACH_SAFE (trace, next, node, &trace_map) {
635 hmap_remove(&trace_map, &trace->node);
637 if (trace->count < min_count) {
641 frame_strs = backtrace_symbols(trace->backtrace, trace->n_frames);
643 ds_put_format(ds, "Count %zu\n", trace->count);
644 for (j = 0; j < trace->n_frames; j++) {
645 ds_put_format(ds, "%s\n", frame_strs[j]);
647 ds_put_cstr(ds, "\n");
651 hmap_destroy(&trace_map);
654 unblock_sigalrm(&oldsigs);
658 /* Unixctl interface. */
660 /* "time/stop" stops the monotonic time returned by e.g. time_msec() from
661 * advancing, except due to later calls to "time/warp". */
663 timeval_stop_cb(struct unixctl_conn *conn,
664 int argc OVS_UNUSED, const char *argv[] OVS_UNUSED,
665 void *aux OVS_UNUSED)
668 unixctl_command_reply(conn, NULL);
671 /* "time/warp MSECS" advances the current monotonic time by the specified
672 * number of milliseconds. Unless "time/stop" has also been executed, the
673 * monotonic clock continues to tick forward at the normal rate afterward.
675 * Does not affect wall clock readings. */
677 timeval_warp_cb(struct unixctl_conn *conn,
678 int argc OVS_UNUSED, const char *argv[], void *aux OVS_UNUSED)
683 msecs = atoi(argv[1]);
685 unixctl_command_reply_error(conn, "invalid MSECS");
689 ts.tv_sec = msecs / 1000;
690 ts.tv_nsec = (msecs % 1000) * 1000 * 1000;
691 timespec_add(&warp_offset, &warp_offset, &ts);
692 timespec_add(&monotonic_time, &monotonic_time, &ts);
693 unixctl_command_reply(conn, "warped");
697 backtrace_cb(struct unixctl_conn *conn,
698 int argc OVS_UNUSED, const char *argv[] OVS_UNUSED,
699 void *aux OVS_UNUSED)
701 struct ds ds = DS_EMPTY_INITIALIZER;
703 ovs_assert(USE_BACKTRACE && CACHE_TIME);
704 format_backtraces(&ds, 0);
705 unixctl_command_reply(conn, ds_cstr(&ds));
710 timeval_dummy_register(void)
712 unixctl_command_register("time/stop", "", 0, 0, timeval_stop_cb, NULL);
713 unixctl_command_register("time/warp", "MSECS", 1, 1,
714 timeval_warp_cb, NULL);