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"
33 #include "ovs-thread.h"
39 /* backtrace() from <execinfo.h> is really useful, but it is not signal safe
40 * everywhere, such as on x86-64. */
41 #if HAVE_BACKTRACE && !defined __x86_64__
42 # define USE_BACKTRACE 1
43 # include <execinfo.h>
45 # define USE_BACKTRACE 0
48 VLOG_DEFINE_THIS_MODULE(timeval);
50 /* The clock to use for measuring time intervals. This is CLOCK_MONOTONIC by
51 * preference, but on systems that don't have a monotonic clock we fall back
52 * to CLOCK_REALTIME. */
53 static clockid_t monotonic_clock;
55 /* Has a timer tick occurred? Only relevant if CACHE_TIME is true.
57 * We initialize these to true to force time_init() to get called on the first
58 * call to time_msec() or another function that queries the current time. */
59 static volatile sig_atomic_t wall_tick = true;
60 static volatile sig_atomic_t monotonic_tick = true;
62 /* The current time, as of the last refresh. */
63 static struct timespec wall_time;
64 static struct timespec monotonic_time;
66 /* The monotonic time at which the time module was initialized. */
67 static long long int boot_time;
69 /* features for use by unit tests. */
70 static struct timespec warp_offset; /* Offset added to monotonic_time. */
71 static bool time_stopped; /* Disables real-time updates, if true. */
73 /* Time in milliseconds at which to die with SIGALRM (if not LLONG_MAX). */
74 static long long int deadline = LLONG_MAX;
77 void *backtrace[32]; /* Populated by backtrace(). */
78 size_t n_frames; /* Number of frames in 'backtrace'. */
80 /* format_backtraces() helper data. */
81 struct hmap_node node;
86 static struct trace traces[MAX_TRACES];
87 static size_t trace_head = 0;
89 static void set_up_timer(void);
90 static void set_up_signal(int flags);
91 static void sigalrm_handler(int);
92 static void refresh_wall_if_ticked(void);
93 static void refresh_monotonic_if_ticked(void);
94 static void block_sigalrm(sigset_t *);
95 static void unblock_sigalrm(const sigset_t *);
96 static void log_poll_interval(long long int last_wakeup);
97 static struct rusage *get_recent_rusage(void);
98 static void refresh_rusage(void);
99 static void timespec_add(struct timespec *sum,
100 const struct timespec *a, const struct timespec *b);
101 static unixctl_cb_func backtrace_cb;
105 backtrace(void **buffer OVS_UNUSED, int size OVS_UNUSED)
111 backtrace_symbols(void *const *buffer OVS_UNUSED, int size OVS_UNUSED)
115 #endif /* !USE_BACKTRACE */
117 /* Initializes the timetracking module, if not already initialized. */
128 /* The implementation of backtrace() in glibc does some one time
129 * initialization which is not signal safe. This can cause deadlocks if
130 * run from the signal handler. As a workaround, force the initialization
135 backtrace(bt, ARRAY_SIZE(bt));
138 memset(traces, 0, sizeof traces);
140 if (USE_BACKTRACE && CACHE_TIME) {
141 unixctl_command_register("backtrace", "", 0, 0, backtrace_cb, NULL);
146 if (!clock_gettime(CLOCK_MONOTONIC, &monotonic_time)) {
147 monotonic_clock = CLOCK_MONOTONIC;
149 monotonic_clock = CLOCK_REALTIME;
150 VLOG_DBG("monotonic timer not available");
153 set_up_signal(SA_RESTART);
156 boot_time = time_msec();
160 set_up_signal(int flags)
164 memset(&sa, 0, sizeof sa);
165 sa.sa_handler = sigalrm_handler;
166 sigemptyset(&sa.sa_mask);
168 xsigaction(SIGALRM, &sa, NULL);
174 static timer_t timer_id; /* "static" to avoid apparent memory leak. */
175 struct itimerspec itimer;
181 if (timer_create(monotonic_clock, NULL, &timer_id)) {
182 VLOG_FATAL("timer_create failed (%s)", ovs_strerror(errno));
185 itimer.it_interval.tv_sec = 0;
186 itimer.it_interval.tv_nsec = TIME_UPDATE_INTERVAL * 1000 * 1000;
187 itimer.it_value = itimer.it_interval;
189 if (timer_settime(timer_id, 0, &itimer, NULL)) {
190 VLOG_FATAL("timer_settime failed (%s)", ovs_strerror(errno));
194 /* Set up the interval timer, to ensure that time advances even without calling
197 * A child created with fork() does not inherit the parent's interval timer, so
198 * this function needs to be called from the child after fork(). */
210 clock_gettime(CLOCK_REALTIME, &wall_time);
215 refresh_monotonic(void)
220 if (monotonic_clock == CLOCK_MONOTONIC) {
221 clock_gettime(monotonic_clock, &monotonic_time);
223 refresh_wall_if_ticked();
224 monotonic_time = wall_time;
226 timespec_add(&monotonic_time, &monotonic_time, &warp_offset);
228 monotonic_tick = false;
232 /* Forces a refresh of the current time from the kernel. It is not usually
233 * necessary to call this function, since the time will be refreshed
234 * automatically at least every TIME_UPDATE_INTERVAL milliseconds. If
235 * CACHE_TIME is false, we will always refresh the current time so this
236 * function has no effect. */
240 wall_tick = monotonic_tick = true;
243 /* Returns a monotonic timer, in seconds. */
247 refresh_monotonic_if_ticked();
248 return monotonic_time.tv_sec;
251 /* Returns the current time, in seconds. */
255 refresh_wall_if_ticked();
256 return wall_time.tv_sec;
259 /* Returns a monotonic timer, in ms (within TIME_UPDATE_INTERVAL ms). */
263 refresh_monotonic_if_ticked();
264 return timespec_to_msec(&monotonic_time);
267 /* Returns the current time, in ms (within TIME_UPDATE_INTERVAL ms). */
271 refresh_wall_if_ticked();
272 return timespec_to_msec(&wall_time);
275 /* Stores a monotonic timer, accurate within TIME_UPDATE_INTERVAL ms, into
278 time_timespec(struct timespec *ts)
280 refresh_monotonic_if_ticked();
281 *ts = monotonic_time;
284 /* Stores the current time, accurate within TIME_UPDATE_INTERVAL ms, into
287 time_wall_timespec(struct timespec *ts)
289 refresh_wall_if_ticked();
293 /* Configures the program to die with SIGALRM 'secs' seconds from now, if
294 * 'secs' is nonzero, or disables the feature if 'secs' is zero. */
296 time_alarm(unsigned int secs)
301 assert_single_threaded();
306 msecs = secs * 1000LL;
307 deadline = now < LLONG_MAX - msecs ? now + msecs : LLONG_MAX;
310 /* Like poll(), except:
312 * - The timeout is specified as an absolute time, as defined by
313 * time_msec(), instead of a duration.
315 * - On error, returns a negative error code (instead of setting errno).
317 * - If interrupted by a signal, retries automatically until the original
318 * timeout is reached. (Because of this property, this function will
319 * never return -EINTR.)
321 * - As a side effect, refreshes the current time (like time_refresh()).
323 * Stores the number of milliseconds elapsed during poll in '*elapsed'. */
325 time_poll(struct pollfd *pollfds, int n_pollfds, long long int timeout_when,
328 static long long int last_wakeup = 0;
336 log_poll_interval(last_wakeup);
342 timeout_when = MIN(timeout_when, deadline);
345 long long int now = time_msec();
348 if (now >= timeout_when) {
350 } else if ((unsigned long long int) timeout_when - now > INT_MAX) {
353 time_left = timeout_when - now;
356 retval = poll(pollfds, n_pollfds, time_left);
362 if (deadline <= time_msec()) {
363 fatal_signal_handler(SIGALRM);
370 if (retval != -EINTR) {
374 if (!blocked && CACHE_TIME) {
375 block_sigalrm(&oldsigs);
380 unblock_sigalrm(&oldsigs);
382 last_wakeup = time_msec();
384 *elapsed = last_wakeup - start;
389 sigalrm_handler(int sig_nr OVS_UNUSED)
392 monotonic_tick = true;
394 if (USE_BACKTRACE && CACHE_TIME) {
395 struct trace *trace = &traces[trace_head];
397 trace->n_frames = backtrace(trace->backtrace,
398 ARRAY_SIZE(trace->backtrace));
399 trace_head = (trace_head + 1) % MAX_TRACES;
404 refresh_wall_if_ticked(void)
406 if (!CACHE_TIME || wall_tick) {
412 refresh_monotonic_if_ticked(void)
414 if (!CACHE_TIME || monotonic_tick) {
420 block_sigalrm(sigset_t *oldsigs)
423 sigemptyset(&sigalrm);
424 sigaddset(&sigalrm, SIGALRM);
425 xpthread_sigmask(SIG_BLOCK, &sigalrm, oldsigs);
429 unblock_sigalrm(const sigset_t *oldsigs)
431 xpthread_sigmask(SIG_SETMASK, oldsigs, NULL);
435 timespec_to_msec(const struct timespec *ts)
437 return (long long int) ts->tv_sec * 1000 + ts->tv_nsec / (1000 * 1000);
441 timeval_to_msec(const struct timeval *tv)
443 return (long long int) tv->tv_sec * 1000 + tv->tv_usec / 1000;
446 /* Returns the monotonic time at which the "time" module was initialized, in
456 xgettimeofday(struct timeval *tv)
458 if (gettimeofday(tv, NULL) == -1) {
459 VLOG_FATAL("gettimeofday failed (%s)", ovs_strerror(errno));
464 timeval_diff_msec(const struct timeval *a, const struct timeval *b)
466 return timeval_to_msec(a) - timeval_to_msec(b);
470 timespec_add(struct timespec *sum,
471 const struct timespec *a,
472 const struct timespec *b)
476 tmp.tv_sec = a->tv_sec + b->tv_sec;
477 tmp.tv_nsec = a->tv_nsec + b->tv_nsec;
478 if (tmp.tv_nsec >= 1000 * 1000 * 1000) {
479 tmp.tv_nsec -= 1000 * 1000 * 1000;
487 log_poll_interval(long long int last_wakeup)
489 long long int interval = time_msec() - last_wakeup;
491 if (interval >= 1000 && !warp_offset.tv_sec && !warp_offset.tv_nsec) {
492 const struct rusage *last_rusage = get_recent_rusage();
493 struct rusage rusage;
495 getrusage(RUSAGE_SELF, &rusage);
496 VLOG_WARN("Unreasonably long %lldms poll interval"
497 " (%lldms user, %lldms system)",
499 timeval_diff_msec(&rusage.ru_utime,
500 &last_rusage->ru_utime),
501 timeval_diff_msec(&rusage.ru_stime,
502 &last_rusage->ru_stime));
503 if (rusage.ru_minflt > last_rusage->ru_minflt
504 || rusage.ru_majflt > last_rusage->ru_majflt) {
505 VLOG_WARN("faults: %ld minor, %ld major",
506 rusage.ru_minflt - last_rusage->ru_minflt,
507 rusage.ru_majflt - last_rusage->ru_majflt);
509 if (rusage.ru_inblock > last_rusage->ru_inblock
510 || rusage.ru_oublock > last_rusage->ru_oublock) {
511 VLOG_WARN("disk: %ld reads, %ld writes",
512 rusage.ru_inblock - last_rusage->ru_inblock,
513 rusage.ru_oublock - last_rusage->ru_oublock);
515 if (rusage.ru_nvcsw > last_rusage->ru_nvcsw
516 || rusage.ru_nivcsw > last_rusage->ru_nivcsw) {
517 VLOG_WARN("context switches: %ld voluntary, %ld involuntary",
518 rusage.ru_nvcsw - last_rusage->ru_nvcsw,
519 rusage.ru_nivcsw - last_rusage->ru_nivcsw);
525 /* CPU usage tracking. */
528 long long int when; /* Time that this sample was taken. */
529 unsigned long long int cpu; /* Total user+system CPU usage when sampled. */
532 static struct rusage recent_rusage;
533 static struct cpu_usage older = { LLONG_MIN, 0 };
534 static struct cpu_usage newer = { LLONG_MIN, 0 };
535 static int cpu_usage = -1;
537 static struct rusage *
538 get_recent_rusage(void)
540 return &recent_rusage;
549 getrusage(RUSAGE_SELF, &recent_rusage);
551 if (now >= newer.when + 3 * 1000) {
554 newer.cpu = (timeval_to_msec(&recent_rusage.ru_utime) +
555 timeval_to_msec(&recent_rusage.ru_stime));
557 if (older.when != LLONG_MIN && newer.cpu > older.cpu) {
558 unsigned int dividend = newer.cpu - older.cpu;
559 unsigned int divisor = (newer.when - older.when) / 100;
560 cpu_usage = divisor > 0 ? dividend / divisor : -1;
567 /* Returns an estimate of this process's CPU usage, as a percentage, over the
568 * past few seconds of wall-clock time. Returns -1 if no estimate is available
569 * (which will happen if the process has not been running long enough to have
570 * an estimate, and can happen for other reasons as well). */
578 hash_trace(struct trace *trace)
580 return hash_bytes(trace->backtrace,
581 trace->n_frames * sizeof *trace->backtrace, 0);
584 static struct trace *
585 trace_map_lookup(struct hmap *trace_map, struct trace *key)
589 HMAP_FOR_EACH_WITH_HASH (value, node, hash_trace(key), trace_map) {
590 if (key->n_frames == value->n_frames
591 && !memcmp(key->backtrace, value->backtrace,
592 key->n_frames * sizeof *key->backtrace)) {
599 /* Appends a string to 'ds' representing backtraces recorded at regular
600 * intervals in the recent past. This information can be used to get a sense
601 * of what the process has been spending the majority of time doing. Will
602 * ommit any backtraces which have not occurred at least 'min_count' times. */
604 format_backtraces(struct ds *ds, size_t min_count)
608 if (USE_BACKTRACE && CACHE_TIME) {
609 struct hmap trace_map = HMAP_INITIALIZER(&trace_map);
610 struct trace *trace, *next;
614 block_sigalrm(&oldsigs);
616 for (i = 0; i < MAX_TRACES; i++) {
617 struct trace *trace = &traces[i];
618 struct trace *map_trace;
620 if (!trace->n_frames) {
624 map_trace = trace_map_lookup(&trace_map, trace);
628 hmap_insert(&trace_map, &trace->node, hash_trace(trace));
633 HMAP_FOR_EACH_SAFE (trace, next, node, &trace_map) {
637 hmap_remove(&trace_map, &trace->node);
639 if (trace->count < min_count) {
643 frame_strs = backtrace_symbols(trace->backtrace, trace->n_frames);
645 ds_put_format(ds, "Count %zu\n", trace->count);
646 for (j = 0; j < trace->n_frames; j++) {
647 ds_put_format(ds, "%s\n", frame_strs[j]);
649 ds_put_cstr(ds, "\n");
653 hmap_destroy(&trace_map);
656 unblock_sigalrm(&oldsigs);
660 /* Unixctl interface. */
662 /* "time/stop" stops the monotonic time returned by e.g. time_msec() from
663 * advancing, except due to later calls to "time/warp". */
665 timeval_stop_cb(struct unixctl_conn *conn,
666 int argc OVS_UNUSED, const char *argv[] OVS_UNUSED,
667 void *aux OVS_UNUSED)
670 unixctl_command_reply(conn, NULL);
673 /* "time/warp MSECS" advances the current monotonic time by the specified
674 * number of milliseconds. Unless "time/stop" has also been executed, the
675 * monotonic clock continues to tick forward at the normal rate afterward.
677 * Does not affect wall clock readings. */
679 timeval_warp_cb(struct unixctl_conn *conn,
680 int argc OVS_UNUSED, const char *argv[], void *aux OVS_UNUSED)
685 msecs = atoi(argv[1]);
687 unixctl_command_reply_error(conn, "invalid MSECS");
691 ts.tv_sec = msecs / 1000;
692 ts.tv_nsec = (msecs % 1000) * 1000 * 1000;
693 timespec_add(&warp_offset, &warp_offset, &ts);
694 timespec_add(&monotonic_time, &monotonic_time, &ts);
695 unixctl_command_reply(conn, "warped");
699 backtrace_cb(struct unixctl_conn *conn,
700 int argc OVS_UNUSED, const char *argv[] OVS_UNUSED,
701 void *aux OVS_UNUSED)
703 struct ds ds = DS_EMPTY_INITIALIZER;
705 ovs_assert(USE_BACKTRACE && CACHE_TIME);
706 format_backtraces(&ds, 0);
707 unixctl_command_reply(conn, ds_cstr(&ds));
712 timeval_dummy_register(void)
714 unixctl_command_register("time/stop", "", 0, 0, timeval_stop_cb, NULL);
715 unixctl_command_register("time/warp", "MSECS", 1, 1,
716 timeval_warp_cb, NULL);