2 * Copyright (c) 2008, 2009, 2010, 2011, 2012, 2013, 2014 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 "dynamic-string.h"
31 #include "fatal-signal.h"
34 #include "ovs-thread.h"
41 VLOG_DEFINE_THIS_MODULE(timeval
);
44 typedef unsigned int clockid_t
;
46 #ifndef CLOCK_MONOTONIC
47 #define CLOCK_MONOTONIC 1
50 #ifndef CLOCK_REALTIME
51 #define CLOCK_REALTIME 2
54 /* Number of 100 ns intervals from January 1, 1601 till January 1, 1970. */
55 static ULARGE_INTEGER unix_epoch
;
59 clockid_t id
; /* CLOCK_MONOTONIC or CLOCK_REALTIME. */
61 /* Features for use by unit tests. Protected by 'mutex'. */
62 struct ovs_mutex mutex
;
63 atomic_bool slow_path
; /* True if warped or stopped. */
64 struct timespec warp OVS_GUARDED
; /* Offset added for unit tests. */
65 bool stopped OVS_GUARDED
; /* Disable real-time updates if true. */
66 struct timespec cache OVS_GUARDED
; /* Last time read from kernel. */
70 static struct clock monotonic_clock
; /* CLOCK_MONOTONIC, if available. */
71 static struct clock wall_clock
; /* CLOCK_REALTIME. */
73 /* The monotonic time at which the time module was initialized. */
74 static long long int boot_time
;
76 /* True only when timeval_dummy_register() is called. */
77 static bool timewarp_enabled
;
78 /* Reference to the seq struct. Threads other than main thread can
79 * wait on timewarp_seq and be waken up when time is warped. */
80 static struct seq
*timewarp_seq
;
81 /* Last value of 'timewarp_seq'. */
82 DEFINE_STATIC_PER_THREAD_DATA(uint64_t, last_seq
, 0);
84 /* Monotonic time in milliseconds at which to die with SIGALRM (if not
86 static long long int deadline
= LLONG_MAX
;
88 /* Monotonic time, in milliseconds, at which the last call to time_poll() woke
90 DEFINE_STATIC_PER_THREAD_DATA(long long int, last_wakeup
, 0);
92 static void log_poll_interval(long long int last_wakeup
);
93 static struct rusage
*get_recent_rusage(void);
94 static void refresh_rusage(void);
95 static void timespec_add(struct timespec
*sum
,
96 const struct timespec
*a
, const struct timespec
*b
);
99 init_clock(struct clock
*c
, clockid_t id
)
101 memset(c
, 0, sizeof *c
);
103 ovs_mutex_init(&c
->mutex
);
104 atomic_init(&c
->slow_path
, false);
105 xclock_gettime(c
->id
, &c
->cache
);
106 timewarp_seq
= seq_create();
115 /* Calculate number of 100-nanosecond intervals till 01/01/1970. */
116 SYSTEMTIME unix_epoch_st
= { 1970, 1, 0, 1, 0, 0, 0, 0};
117 FILETIME unix_epoch_ft
;
119 SystemTimeToFileTime(&unix_epoch_st
, &unix_epoch_ft
);
120 unix_epoch
.LowPart
= unix_epoch_ft
.dwLowDateTime
;
121 unix_epoch
.HighPart
= unix_epoch_ft
.dwHighDateTime
;
126 init_clock(&monotonic_clock
, (!clock_gettime(CLOCK_MONOTONIC
, &ts
)
129 init_clock(&wall_clock
, CLOCK_REALTIME
);
130 boot_time
= timespec_to_msec(&monotonic_clock
.cache
);
133 /* Initializes the timetracking module, if not already initialized. */
137 static pthread_once_t once
= PTHREAD_ONCE_INIT
;
138 pthread_once(&once
, do_init_time
);
142 time_timespec__(struct clock
*c
, struct timespec
*ts
)
148 atomic_read_explicit(&c
->slow_path
, &slow_path
, memory_order_relaxed
);
150 xclock_gettime(c
->id
, ts
);
152 struct timespec warp
;
153 struct timespec cache
;
156 ovs_mutex_lock(&c
->mutex
);
157 stopped
= c
->stopped
;
160 ovs_mutex_unlock(&c
->mutex
);
163 xclock_gettime(c
->id
, &cache
);
165 timespec_add(ts
, &cache
, &warp
);
169 /* Stores a monotonic timer, accurate within TIME_UPDATE_INTERVAL ms, into
172 time_timespec(struct timespec
*ts
)
174 time_timespec__(&monotonic_clock
, ts
);
177 /* Stores the current time, accurate within TIME_UPDATE_INTERVAL ms, into
180 time_wall_timespec(struct timespec
*ts
)
182 time_timespec__(&wall_clock
, ts
);
186 time_sec__(struct clock
*c
)
190 time_timespec__(c
, &ts
);
194 /* Returns a monotonic timer, in seconds. */
198 return time_sec__(&monotonic_clock
);
201 /* Returns the current time, in seconds. */
205 return time_sec__(&wall_clock
);
209 time_msec__(struct clock
*c
)
213 time_timespec__(c
, &ts
);
214 return timespec_to_msec(&ts
);
217 /* Returns a monotonic timer, in ms (within TIME_UPDATE_INTERVAL ms). */
221 return time_msec__(&monotonic_clock
);
224 /* Returns the current time, in ms (within TIME_UPDATE_INTERVAL ms). */
228 return time_msec__(&wall_clock
);
231 /* Configures the program to die with SIGALRM 'secs' seconds from now, if
232 * 'secs' is nonzero, or disables the feature if 'secs' is zero. */
234 time_alarm(unsigned int secs
)
239 assert_single_threaded();
243 msecs
= secs
* 1000LL;
244 deadline
= now
< LLONG_MAX
- msecs
? now
+ msecs
: LLONG_MAX
;
247 /* Like poll(), except:
249 * - The timeout is specified as an absolute time, as defined by
250 * time_msec(), instead of a duration.
252 * - On error, returns a negative error code (instead of setting errno).
254 * - If interrupted by a signal, retries automatically until the original
255 * timeout is reached. (Because of this property, this function will
256 * never return -EINTR.)
258 * Stores the number of milliseconds elapsed during poll in '*elapsed'. */
260 time_poll(struct pollfd
*pollfds
, int n_pollfds
, HANDLE
*handles OVS_UNUSED
,
261 long long int timeout_when
, int *elapsed
)
263 long long int *last_wakeup
= last_wakeup_get();
271 log_poll_interval(*last_wakeup
);
275 timeout_when
= MIN(timeout_when
, deadline
);
278 long long int now
= time_msec();
281 if (now
>= timeout_when
) {
283 } else if ((unsigned long long int) timeout_when
- now
> INT_MAX
) {
286 time_left
= timeout_when
- now
;
290 retval
= poll(pollfds
, n_pollfds
, time_left
);
295 if (n_pollfds
> MAXIMUM_WAIT_OBJECTS
) {
296 VLOG_ERR("Cannot handle more than maximum wait objects\n");
297 } else if (n_pollfds
!= 0) {
298 retval
= WaitForMultipleObjects(n_pollfds
, handles
, FALSE
,
302 /* XXX This will be replace by a win error to errno
303 conversion function */
304 retval
= -WSAGetLastError();
309 if (deadline
<= time_msec()) {
311 fatal_signal_handler(SIGALRM
);
313 VLOG_ERR("wake up from WaitForMultipleObjects after deadline");
314 fatal_signal_handler(SIGTERM
);
322 if (retval
!= -EINTR
) {
326 *last_wakeup
= time_msec();
328 *elapsed
= *last_wakeup
- start
;
333 timespec_to_msec(const struct timespec
*ts
)
335 return (long long int) ts
->tv_sec
* 1000 + ts
->tv_nsec
/ (1000 * 1000);
339 timeval_to_msec(const struct timeval
*tv
)
341 return (long long int) tv
->tv_sec
* 1000 + tv
->tv_usec
/ 1000;
344 /* Returns the monotonic time at which the "time" module was initialized, in
354 static ULARGE_INTEGER
357 ULARGE_INTEGER current_time
;
358 FILETIME current_time_ft
;
360 /* Returns current time in UTC as a 64-bit value representing the number
361 * of 100-nanosecond intervals since January 1, 1601 . */
362 GetSystemTimePreciseAsFileTime(¤t_time_ft
);
363 current_time
.LowPart
= current_time_ft
.dwLowDateTime
;
364 current_time
.HighPart
= current_time_ft
.dwHighDateTime
;
370 clock_gettime(clock_t id
, struct timespec
*ts
)
372 if (id
== CLOCK_MONOTONIC
) {
373 static LARGE_INTEGER freq
;
377 if (!freq
.QuadPart
) {
378 /* Number of counts per second. */
379 QueryPerformanceFrequency(&freq
);
381 /* Total number of counts from a starting point. */
382 QueryPerformanceCounter(&count
);
384 /* Total nano seconds from a starting point. */
385 ns
= (double) count
.QuadPart
/ freq
.QuadPart
* 1000000000;
387 ts
->tv_sec
= count
.QuadPart
/ freq
.QuadPart
;
388 ts
->tv_nsec
= ns
% 1000000000;
389 } else if (id
== CLOCK_REALTIME
) {
390 ULARGE_INTEGER current_time
= xgetfiletime();
392 /* Time from Epoch to now. */
393 ts
->tv_sec
= (current_time
.QuadPart
- unix_epoch
.QuadPart
) / 10000000;
394 ts
->tv_nsec
= ((current_time
.QuadPart
- unix_epoch
.QuadPart
) %
403 xgettimeofday(struct timeval
*tv
)
406 if (gettimeofday(tv
, NULL
) == -1) {
407 VLOG_FATAL("gettimeofday failed (%s)", ovs_strerror(errno
));
410 ULARGE_INTEGER current_time
= xgetfiletime();
412 tv
->tv_sec
= (current_time
.QuadPart
- unix_epoch
.QuadPart
) / 10000000;
413 tv
->tv_usec
= ((current_time
.QuadPart
- unix_epoch
.QuadPart
) %
419 xclock_gettime(clock_t id
, struct timespec
*ts
)
421 if (clock_gettime(id
, ts
) == -1) {
422 /* It seems like a bad idea to try to use vlog here because it is
423 * likely to try to check the current time. */
424 ovs_abort(errno
, "xclock_gettime() failed");
428 /* Makes threads wait on timewarp_seq and be waken up when time is warped.
429 * This function will be no-op unless timeval_dummy_register() is called. */
433 if (timewarp_enabled
) {
434 uint64_t *last_seq
= last_seq_get();
436 *last_seq
= seq_read(timewarp_seq
);
437 seq_wait(timewarp_seq
, *last_seq
);
442 timeval_diff_msec(const struct timeval
*a
, const struct timeval
*b
)
444 return timeval_to_msec(a
) - timeval_to_msec(b
);
448 timespec_add(struct timespec
*sum
,
449 const struct timespec
*a
,
450 const struct timespec
*b
)
454 tmp
.tv_sec
= a
->tv_sec
+ b
->tv_sec
;
455 tmp
.tv_nsec
= a
->tv_nsec
+ b
->tv_nsec
;
456 if (tmp
.tv_nsec
>= 1000 * 1000 * 1000) {
457 tmp
.tv_nsec
-= 1000 * 1000 * 1000;
465 is_warped(const struct clock
*c
)
469 ovs_mutex_lock(&c
->mutex
);
470 warped
= monotonic_clock
.warp
.tv_sec
|| monotonic_clock
.warp
.tv_nsec
;
471 ovs_mutex_unlock(&c
->mutex
);
477 log_poll_interval(long long int last_wakeup
)
479 long long int interval
= time_msec() - last_wakeup
;
481 if (interval
>= 1000 && !is_warped(&monotonic_clock
)) {
482 const struct rusage
*last_rusage
= get_recent_rusage();
483 struct rusage rusage
;
485 getrusage(RUSAGE_SELF
, &rusage
);
486 VLOG_WARN("Unreasonably long %lldms poll interval"
487 " (%lldms user, %lldms system)",
489 timeval_diff_msec(&rusage
.ru_utime
,
490 &last_rusage
->ru_utime
),
491 timeval_diff_msec(&rusage
.ru_stime
,
492 &last_rusage
->ru_stime
));
493 if (rusage
.ru_minflt
> last_rusage
->ru_minflt
494 || rusage
.ru_majflt
> last_rusage
->ru_majflt
) {
495 VLOG_WARN("faults: %ld minor, %ld major",
496 rusage
.ru_minflt
- last_rusage
->ru_minflt
,
497 rusage
.ru_majflt
- last_rusage
->ru_majflt
);
499 if (rusage
.ru_inblock
> last_rusage
->ru_inblock
500 || rusage
.ru_oublock
> last_rusage
->ru_oublock
) {
501 VLOG_WARN("disk: %ld reads, %ld writes",
502 rusage
.ru_inblock
- last_rusage
->ru_inblock
,
503 rusage
.ru_oublock
- last_rusage
->ru_oublock
);
505 if (rusage
.ru_nvcsw
> last_rusage
->ru_nvcsw
506 || rusage
.ru_nivcsw
> last_rusage
->ru_nivcsw
) {
507 VLOG_WARN("context switches: %ld voluntary, %ld involuntary",
508 rusage
.ru_nvcsw
- last_rusage
->ru_nvcsw
,
509 rusage
.ru_nivcsw
- last_rusage
->ru_nivcsw
);
515 /* CPU usage tracking. */
518 long long int when
; /* Time that this sample was taken. */
519 unsigned long long int cpu
; /* Total user+system CPU usage when sampled. */
523 struct cpu_usage older
;
524 struct cpu_usage newer
;
527 struct rusage recent_rusage
;
529 DEFINE_PER_THREAD_MALLOCED_DATA(struct cpu_tracker
*, cpu_tracker_var
);
531 static struct cpu_tracker
*
532 get_cpu_tracker(void)
534 struct cpu_tracker
*t
= cpu_tracker_var_get();
536 t
= xzalloc(sizeof *t
);
537 t
->older
.when
= LLONG_MIN
;
538 t
->newer
.when
= LLONG_MIN
;
539 cpu_tracker_var_set_unsafe(t
);
544 static struct rusage
*
545 get_recent_rusage(void)
547 return &get_cpu_tracker()->recent_rusage
;
551 getrusage_thread(struct rusage
*rusage OVS_UNUSED
)
554 return getrusage(RUSAGE_THREAD
, rusage
);
564 struct cpu_tracker
*t
= get_cpu_tracker();
565 struct rusage
*recent_rusage
= &t
->recent_rusage
;
567 if (!getrusage_thread(recent_rusage
)) {
568 long long int now
= time_msec();
569 if (now
>= t
->newer
.when
+ 3 * 1000) {
572 t
->newer
.cpu
= (timeval_to_msec(&recent_rusage
->ru_utime
) +
573 timeval_to_msec(&recent_rusage
->ru_stime
));
575 if (t
->older
.when
!= LLONG_MIN
&& t
->newer
.cpu
> t
->older
.cpu
) {
576 unsigned int dividend
= t
->newer
.cpu
- t
->older
.cpu
;
577 unsigned int divisor
= (t
->newer
.when
- t
->older
.when
) / 100;
578 t
->cpu_usage
= divisor
> 0 ? dividend
/ divisor
: -1;
586 /* Returns an estimate of this process's CPU usage, as a percentage, over the
587 * past few seconds of wall-clock time. Returns -1 if no estimate is available
588 * (which will happen if the process has not been running long enough to have
589 * an estimate, and can happen for other reasons as well). */
593 return get_cpu_tracker()->cpu_usage
;
596 /* Unixctl interface. */
598 /* "time/stop" stops the monotonic time returned by e.g. time_msec() from
599 * advancing, except due to later calls to "time/warp". */
601 timeval_stop_cb(struct unixctl_conn
*conn
,
602 int argc OVS_UNUSED
, const char *argv
[] OVS_UNUSED
,
603 void *aux OVS_UNUSED
)
605 ovs_mutex_lock(&monotonic_clock
.mutex
);
606 atomic_store(&monotonic_clock
.slow_path
, true);
607 monotonic_clock
.stopped
= true;
608 xclock_gettime(monotonic_clock
.id
, &monotonic_clock
.cache
);
609 ovs_mutex_unlock(&monotonic_clock
.mutex
);
611 unixctl_command_reply(conn
, NULL
);
614 /* "time/warp MSECS" advances the current monotonic time by the specified
615 * number of milliseconds. Unless "time/stop" has also been executed, the
616 * monotonic clock continues to tick forward at the normal rate afterward.
618 * Does not affect wall clock readings. */
620 timeval_warp_cb(struct unixctl_conn
*conn
,
621 int argc OVS_UNUSED
, const char *argv
[], void *aux OVS_UNUSED
)
626 msecs
= atoi(argv
[1]);
628 unixctl_command_reply_error(conn
, "invalid MSECS");
632 ts
.tv_sec
= msecs
/ 1000;
633 ts
.tv_nsec
= (msecs
% 1000) * 1000 * 1000;
635 ovs_mutex_lock(&monotonic_clock
.mutex
);
636 atomic_store(&monotonic_clock
.slow_path
, true);
637 timespec_add(&monotonic_clock
.warp
, &monotonic_clock
.warp
, &ts
);
638 ovs_mutex_unlock(&monotonic_clock
.mutex
);
639 seq_change(timewarp_seq
);
640 /* give threads (eg. monitor) some chances to run */
646 unixctl_command_reply(conn
, "warped");
650 timeval_dummy_register(void)
652 timewarp_enabled
= true;
653 unixctl_command_register("time/stop", "", 0, 0, timeval_stop_cb
, NULL
);
654 unixctl_command_register("time/warp", "MSECS", 1, 1,
655 timeval_warp_cb
, NULL
);
660 /* strftime() with an extension for high-resolution timestamps. Any '#'s in
661 * 'format' will be replaced by subseconds, e.g. use "%S.###" to obtain results
664 strftime_msec(char *s
, size_t max
, const char *format
,
665 const struct tm_msec
*tm
)
669 n
= strftime(s
, max
, format
, &tm
->tm
);
674 sprintf(decimals
, "%03d", tm
->msec
);
675 for (p
= strchr(s
, '#'); p
; p
= strchr(p
, '#')) {
678 *p
++ = *d
? *d
++ : '0';
687 localtime_msec(long long int now
, struct tm_msec
*result
)
689 time_t now_sec
= now
/ 1000;
690 localtime_r(&now_sec
, &result
->tm
);
691 result
->msec
= now
% 1000;
696 gmtime_msec(long long int now
, struct tm_msec
*result
)
698 time_t now_sec
= now
/ 1000;
699 gmtime_r(&now_sec
, &result
->tm
);
700 result
->msec
= now
% 1000;