2 * Copyright (c) 2008, 2009, 2010, 2011, 2012, 2013, 2014, 2015 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"
35 #include "ovs-thread.h"
40 #include "openvswitch/vlog.h"
42 VLOG_DEFINE_THIS_MODULE(timeval
);
45 typedef unsigned int clockid_t
;
47 #ifndef CLOCK_MONOTONIC
48 #define CLOCK_MONOTONIC 1
51 #ifndef CLOCK_REALTIME
52 #define CLOCK_REALTIME 2
55 /* Number of 100 ns intervals from January 1, 1601 till January 1, 1970. */
56 const static unsigned long long unix_epoch
= 116444736000000000;
59 /* Structure set by unixctl time/warp command. */
61 struct unixctl_conn
*conn
; /* Connection waiting for warp response. */
62 long long int total_warp
; /* Total offset to be added to monotonic time. */
63 long long int warp
; /* 'total_warp' offset done in steps of 'warp'. */
64 unsigned int main_thread_id
; /* Identification for the main thread. */
68 clockid_t id
; /* CLOCK_MONOTONIC or CLOCK_REALTIME. */
70 /* Features for use by unit tests. Protected by 'mutex'. */
71 struct ovs_mutex mutex
;
72 atomic_bool slow_path
; /* True if warped or stopped. */
73 struct timespec warp OVS_GUARDED
; /* Offset added for unit tests. */
74 bool stopped OVS_GUARDED
; /* Disable real-time updates if true. */
75 struct timespec cache OVS_GUARDED
; /* Last time read from kernel. */
76 struct large_warp large_warp OVS_GUARDED
; /* Connection information waiting
81 static struct clock monotonic_clock
; /* CLOCK_MONOTONIC, if available. */
82 static struct clock wall_clock
; /* CLOCK_REALTIME. */
84 /* The monotonic time at which the time module was initialized. */
85 static long long int boot_time
;
87 /* True only when timeval_dummy_register() is called. */
88 static bool timewarp_enabled
;
89 /* Reference to the seq struct. Threads other than main thread can
90 * wait on timewarp_seq and be waken up when time is warped. */
91 static struct seq
*timewarp_seq
;
92 /* Last value of 'timewarp_seq'. */
93 DEFINE_STATIC_PER_THREAD_DATA(uint64_t, last_seq
, 0);
95 /* Monotonic time in milliseconds at which to die with SIGALRM (if not
97 static long long int deadline
= LLONG_MAX
;
99 /* Monotonic time, in milliseconds, at which the last call to time_poll() woke
101 DEFINE_STATIC_PER_THREAD_DATA(long long int, last_wakeup
, 0);
103 static void log_poll_interval(long long int last_wakeup
);
104 static struct rusage
*get_recent_rusage(void);
105 static int getrusage_thread(struct rusage
*);
106 static void refresh_rusage(void);
107 static void timespec_add(struct timespec
*sum
,
108 const struct timespec
*a
, const struct timespec
*b
);
111 init_clock(struct clock
*c
, clockid_t id
)
113 memset(c
, 0, sizeof *c
);
115 ovs_mutex_init(&c
->mutex
);
116 atomic_init(&c
->slow_path
, false);
117 xclock_gettime(c
->id
, &c
->cache
);
127 timewarp_seq
= seq_create();
128 init_clock(&monotonic_clock
, (!clock_gettime(CLOCK_MONOTONIC
, &ts
)
131 init_clock(&wall_clock
, CLOCK_REALTIME
);
132 boot_time
= timespec_to_msec(&monotonic_clock
.cache
);
135 /* Initializes the timetracking module, if not already initialized. */
139 static pthread_once_t once
= PTHREAD_ONCE_INIT
;
140 pthread_once(&once
, do_init_time
);
144 time_timespec__(struct clock
*c
, struct timespec
*ts
)
150 atomic_read_relaxed(&c
->slow_path
, &slow_path
);
152 xclock_gettime(c
->id
, ts
);
154 struct timespec warp
;
155 struct timespec cache
;
158 ovs_mutex_lock(&c
->mutex
);
159 stopped
= c
->stopped
;
162 ovs_mutex_unlock(&c
->mutex
);
165 xclock_gettime(c
->id
, &cache
);
167 timespec_add(ts
, &cache
, &warp
);
171 /* Stores a monotonic timer, accurate within TIME_UPDATE_INTERVAL ms, into
174 time_timespec(struct timespec
*ts
)
176 time_timespec__(&monotonic_clock
, ts
);
179 /* Stores the current time, accurate within TIME_UPDATE_INTERVAL ms, into
182 time_wall_timespec(struct timespec
*ts
)
184 time_timespec__(&wall_clock
, ts
);
188 time_sec__(struct clock
*c
)
192 time_timespec__(c
, &ts
);
196 /* Returns a monotonic timer, in seconds. */
200 return time_sec__(&monotonic_clock
);
203 /* Returns the current time, in seconds. */
207 return time_sec__(&wall_clock
);
211 time_msec__(struct clock
*c
)
215 time_timespec__(c
, &ts
);
216 return timespec_to_msec(&ts
);
219 /* Returns a monotonic timer, in ms (within TIME_UPDATE_INTERVAL ms). */
223 return time_msec__(&monotonic_clock
);
226 /* Returns the current time, in ms (within TIME_UPDATE_INTERVAL ms). */
230 return time_msec__(&wall_clock
);
233 /* Configures the program to die with SIGALRM 'secs' seconds from now, if
234 * 'secs' is nonzero, or disables the feature if 'secs' is zero. */
236 time_alarm(unsigned int secs
)
241 assert_single_threaded();
245 msecs
= secs
* 1000LL;
246 deadline
= now
< LLONG_MAX
- msecs
? now
+ msecs
: LLONG_MAX
;
249 /* Like poll(), except:
251 * - The timeout is specified as an absolute time, as defined by
252 * time_msec(), instead of a duration.
254 * - On error, returns a negative error code (instead of setting errno).
256 * - If interrupted by a signal, retries automatically until the original
257 * timeout is reached. (Because of this property, this function will
258 * never return -EINTR.)
260 * Stores the number of milliseconds elapsed during poll in '*elapsed'. */
262 time_poll(struct pollfd
*pollfds
, int n_pollfds
, HANDLE
*handles OVS_UNUSED
,
263 long long int timeout_when
, int *elapsed
)
265 long long int *last_wakeup
= last_wakeup_get();
273 if (*last_wakeup
&& !thread_is_pmd()) {
274 log_poll_interval(*last_wakeup
);
278 timeout_when
= MIN(timeout_when
, deadline
);
279 quiescent
= ovsrcu_is_quiescent();
282 long long int now
= time_msec();
285 if (now
>= timeout_when
) {
287 } else if ((unsigned long long int) timeout_when
- now
> INT_MAX
) {
290 time_left
= timeout_when
- now
;
297 ovsrcu_quiesce_start();
302 retval
= poll(pollfds
, n_pollfds
, time_left
);
307 if (n_pollfds
> MAXIMUM_WAIT_OBJECTS
) {
308 VLOG_ERR("Cannot handle more than maximum wait objects\n");
309 } else if (n_pollfds
!= 0) {
310 retval
= WaitForMultipleObjects(n_pollfds
, handles
, FALSE
,
314 /* XXX This will be replace by a win error to errno
315 conversion function */
316 retval
= -WSAGetLastError();
321 if (!quiescent
&& time_left
) {
322 ovsrcu_quiesce_end();
325 if (deadline
<= time_msec()) {
327 fatal_signal_handler(SIGALRM
);
329 VLOG_ERR("wake up from WaitForMultipleObjects after deadline");
330 fatal_signal_handler(SIGTERM
);
338 if (retval
!= -EINTR
) {
342 *last_wakeup
= time_msec();
344 *elapsed
= *last_wakeup
- start
;
349 timespec_to_msec(const struct timespec
*ts
)
351 return (long long int) ts
->tv_sec
* 1000 + ts
->tv_nsec
/ (1000 * 1000);
355 timeval_to_msec(const struct timeval
*tv
)
357 return (long long int) tv
->tv_sec
* 1000 + tv
->tv_usec
/ 1000;
360 /* Returns the monotonic time at which the "time" module was initialized, in
370 static ULARGE_INTEGER
373 ULARGE_INTEGER current_time
;
374 FILETIME current_time_ft
;
376 /* Returns current time in UTC as a 64-bit value representing the number
377 * of 100-nanosecond intervals since January 1, 1601 . */
378 GetSystemTimePreciseAsFileTime(¤t_time_ft
);
379 current_time
.LowPart
= current_time_ft
.dwLowDateTime
;
380 current_time
.HighPart
= current_time_ft
.dwHighDateTime
;
386 clock_gettime(clock_t id
, struct timespec
*ts
)
388 if (id
== CLOCK_MONOTONIC
) {
389 static LARGE_INTEGER freq
;
393 if (!freq
.QuadPart
) {
394 /* Number of counts per second. */
395 QueryPerformanceFrequency(&freq
);
397 /* Total number of counts from a starting point. */
398 QueryPerformanceCounter(&count
);
400 /* Total nano seconds from a starting point. */
401 ns
= (double) count
.QuadPart
/ freq
.QuadPart
* 1000000000;
403 ts
->tv_sec
= count
.QuadPart
/ freq
.QuadPart
;
404 ts
->tv_nsec
= ns
% 1000000000;
405 } else if (id
== CLOCK_REALTIME
) {
406 ULARGE_INTEGER current_time
= xgetfiletime();
408 /* Time from Epoch to now. */
409 ts
->tv_sec
= (current_time
.QuadPart
- unix_epoch
) / 10000000;
410 ts
->tv_nsec
= ((current_time
.QuadPart
- unix_epoch
) %
421 xgettimeofday(struct timeval
*tv
)
424 if (gettimeofday(tv
, NULL
) == -1) {
425 VLOG_FATAL("gettimeofday failed (%s)", ovs_strerror(errno
));
428 ULARGE_INTEGER current_time
= xgetfiletime();
430 tv
->tv_sec
= (current_time
.QuadPart
- unix_epoch
) / 10000000;
431 tv
->tv_usec
= ((current_time
.QuadPart
- unix_epoch
) %
437 xclock_gettime(clock_t id
, struct timespec
*ts
)
439 if (clock_gettime(id
, ts
) == -1) {
440 /* It seems like a bad idea to try to use vlog here because it is
441 * likely to try to check the current time. */
442 ovs_abort(errno
, "xclock_gettime() failed");
447 msec_to_timespec(long long int ms
, struct timespec
*ts
)
449 ts
->tv_sec
= ms
/ 1000;
450 ts
->tv_nsec
= (ms
% 1000) * 1000 * 1000;
456 struct clock
*c
= &monotonic_clock
;
457 struct timespec warp
;
459 ovs_mutex_lock(&c
->mutex
);
460 if (!c
->large_warp
.conn
) {
461 ovs_mutex_unlock(&c
->mutex
);
465 if (c
->large_warp
.total_warp
>= c
->large_warp
.warp
) {
466 msec_to_timespec(c
->large_warp
.warp
, &warp
);
467 timespec_add(&c
->warp
, &c
->warp
, &warp
);
468 c
->large_warp
.total_warp
-= c
->large_warp
.warp
;
469 } else if (c
->large_warp
.total_warp
) {
470 msec_to_timespec(c
->large_warp
.total_warp
, &warp
);
471 timespec_add(&c
->warp
, &c
->warp
, &warp
);
472 c
->large_warp
.total_warp
= 0;
474 /* c->large_warp.total_warp is 0. */
475 msec_to_timespec(c
->large_warp
.warp
, &warp
);
476 timespec_add(&c
->warp
, &c
->warp
, &warp
);
479 if (!c
->large_warp
.total_warp
) {
480 unixctl_command_reply(c
->large_warp
.conn
, "warped");
481 c
->large_warp
.conn
= NULL
;
484 ovs_mutex_unlock(&c
->mutex
);
485 seq_change(timewarp_seq
);
487 /* give threads (eg. monitor) some chances to run */
495 /* Perform work needed for "timewarp_seq"'s producer and consumers. */
499 /* The function is a no-op unless timeval_dummy_register() is called. */
500 if (timewarp_enabled
) {
501 unsigned int thread_id
;
502 ovs_mutex_lock(&monotonic_clock
.mutex
);
503 thread_id
= monotonic_clock
.large_warp
.main_thread_id
;
504 ovs_mutex_unlock(&monotonic_clock
.mutex
);
506 if (thread_id
!= ovsthread_id_self()) {
507 /* For threads other than the thread that changes the sequence,
509 uint64_t *last_seq
= last_seq_get();
511 *last_seq
= seq_read(timewarp_seq
);
512 seq_wait(timewarp_seq
, *last_seq
);
514 /* Work on adding the remaining warps. */
521 timeval_diff_msec(const struct timeval
*a
, const struct timeval
*b
)
523 return timeval_to_msec(a
) - timeval_to_msec(b
);
527 timespec_add(struct timespec
*sum
,
528 const struct timespec
*a
,
529 const struct timespec
*b
)
533 tmp
.tv_sec
= a
->tv_sec
+ b
->tv_sec
;
534 tmp
.tv_nsec
= a
->tv_nsec
+ b
->tv_nsec
;
535 if (tmp
.tv_nsec
>= 1000 * 1000 * 1000) {
536 tmp
.tv_nsec
-= 1000 * 1000 * 1000;
544 is_warped(const struct clock
*c
)
548 ovs_mutex_lock(&c
->mutex
);
549 warped
= monotonic_clock
.warp
.tv_sec
|| monotonic_clock
.warp
.tv_nsec
;
550 ovs_mutex_unlock(&c
->mutex
);
556 log_poll_interval(long long int last_wakeup
)
558 long long int interval
= time_msec() - last_wakeup
;
560 if (interval
>= 1000 && !is_warped(&monotonic_clock
)) {
561 const struct rusage
*last_rusage
= get_recent_rusage();
562 struct rusage rusage
;
564 if (!getrusage_thread(&rusage
)) {
565 VLOG_WARN("Unreasonably long %lldms poll interval"
566 " (%lldms user, %lldms system)",
568 timeval_diff_msec(&rusage
.ru_utime
,
569 &last_rusage
->ru_utime
),
570 timeval_diff_msec(&rusage
.ru_stime
,
571 &last_rusage
->ru_stime
));
573 if (rusage
.ru_minflt
> last_rusage
->ru_minflt
574 || rusage
.ru_majflt
> last_rusage
->ru_majflt
) {
575 VLOG_WARN("faults: %ld minor, %ld major",
576 rusage
.ru_minflt
- last_rusage
->ru_minflt
,
577 rusage
.ru_majflt
- last_rusage
->ru_majflt
);
579 if (rusage
.ru_inblock
> last_rusage
->ru_inblock
580 || rusage
.ru_oublock
> last_rusage
->ru_oublock
) {
581 VLOG_WARN("disk: %ld reads, %ld writes",
582 rusage
.ru_inblock
- last_rusage
->ru_inblock
,
583 rusage
.ru_oublock
- last_rusage
->ru_oublock
);
585 if (rusage
.ru_nvcsw
> last_rusage
->ru_nvcsw
586 || rusage
.ru_nivcsw
> last_rusage
->ru_nivcsw
) {
587 VLOG_WARN("context switches: %ld voluntary, %ld involuntary",
588 rusage
.ru_nvcsw
- last_rusage
->ru_nvcsw
,
589 rusage
.ru_nivcsw
- last_rusage
->ru_nivcsw
);
592 VLOG_WARN("Unreasonably long %lldms poll interval", interval
);
598 /* CPU usage tracking. */
601 long long int when
; /* Time that this sample was taken. */
602 unsigned long long int cpu
; /* Total user+system CPU usage when sampled. */
606 struct cpu_usage older
;
607 struct cpu_usage newer
;
610 struct rusage recent_rusage
;
612 DEFINE_PER_THREAD_MALLOCED_DATA(struct cpu_tracker
*, cpu_tracker_var
);
614 static struct cpu_tracker
*
615 get_cpu_tracker(void)
617 struct cpu_tracker
*t
= cpu_tracker_var_get();
619 t
= xzalloc(sizeof *t
);
620 t
->older
.when
= LLONG_MIN
;
621 t
->newer
.when
= LLONG_MIN
;
622 cpu_tracker_var_set_unsafe(t
);
627 static struct rusage
*
628 get_recent_rusage(void)
630 return &get_cpu_tracker()->recent_rusage
;
634 getrusage_thread(struct rusage
*rusage OVS_UNUSED
)
637 return getrusage(RUSAGE_THREAD
, rusage
);
647 struct cpu_tracker
*t
= get_cpu_tracker();
648 struct rusage
*recent_rusage
= &t
->recent_rusage
;
650 if (!getrusage_thread(recent_rusage
)) {
651 long long int now
= time_msec();
652 if (now
>= t
->newer
.when
+ 3 * 1000) {
655 t
->newer
.cpu
= (timeval_to_msec(&recent_rusage
->ru_utime
) +
656 timeval_to_msec(&recent_rusage
->ru_stime
));
658 if (t
->older
.when
!= LLONG_MIN
&& t
->newer
.cpu
> t
->older
.cpu
) {
659 unsigned int dividend
= t
->newer
.cpu
- t
->older
.cpu
;
660 unsigned int divisor
= (t
->newer
.when
- t
->older
.when
) / 100;
661 t
->cpu_usage
= divisor
> 0 ? dividend
/ divisor
: -1;
669 /* Returns an estimate of this process's CPU usage, as a percentage, over the
670 * past few seconds of wall-clock time. Returns -1 if no estimate is available
671 * (which will happen if the process has not been running long enough to have
672 * an estimate, and can happen for other reasons as well). */
676 return get_cpu_tracker()->cpu_usage
;
679 /* Unixctl interface. */
681 /* "time/stop" stops the monotonic time returned by e.g. time_msec() from
682 * advancing, except due to later calls to "time/warp". */
684 timeval_stop_cb(struct unixctl_conn
*conn
,
685 int argc OVS_UNUSED
, const char *argv
[] OVS_UNUSED
,
686 void *aux OVS_UNUSED
)
688 ovs_mutex_lock(&monotonic_clock
.mutex
);
689 atomic_store_relaxed(&monotonic_clock
.slow_path
, true);
690 monotonic_clock
.stopped
= true;
691 xclock_gettime(monotonic_clock
.id
, &monotonic_clock
.cache
);
692 ovs_mutex_unlock(&monotonic_clock
.mutex
);
694 unixctl_command_reply(conn
, NULL
);
697 /* "time/warp MSECS" advances the current monotonic time by the specified
698 * number of milliseconds. Unless "time/stop" has also been executed, the
699 * monotonic clock continues to tick forward at the normal rate afterward.
701 * "time/warp LARGE_MSECS MSECS" is a variation of the above command. It
702 * advances the current monotonic time by LARGE_MSECS. This is done MSECS
703 * at a time in each run of the main thread. This gives other threads
704 * time to run after the clock has been advanced by MSECS.
706 * Does not affect wall clock readings. */
708 timeval_warp_cb(struct unixctl_conn
*conn
,
709 int argc OVS_UNUSED
, const char *argv
[], void *aux OVS_UNUSED
)
711 long long int total_warp
= argc
> 2 ? atoll(argv
[1]) : 0;
712 long long int msecs
= argc
> 2 ? atoll(argv
[2]) : atoll(argv
[1]);
713 if (msecs
<= 0 || total_warp
< 0) {
714 unixctl_command_reply_error(conn
, "invalid MSECS");
718 ovs_mutex_lock(&monotonic_clock
.mutex
);
719 if (monotonic_clock
.large_warp
.conn
) {
720 ovs_mutex_unlock(&monotonic_clock
.mutex
);
721 unixctl_command_reply_error(conn
, "A previous warp in progress");
724 atomic_store_relaxed(&monotonic_clock
.slow_path
, true);
725 monotonic_clock
.large_warp
.conn
= conn
;
726 monotonic_clock
.large_warp
.total_warp
= total_warp
;
727 monotonic_clock
.large_warp
.warp
= msecs
;
728 monotonic_clock
.large_warp
.main_thread_id
= ovsthread_id_self();
729 ovs_mutex_unlock(&monotonic_clock
.mutex
);
735 timeval_dummy_register(void)
737 timewarp_enabled
= true;
738 unixctl_command_register("time/stop", "", 0, 0, timeval_stop_cb
, NULL
);
739 unixctl_command_register("time/warp", "[large_msecs] msecs", 1, 2,
740 timeval_warp_cb
, NULL
);
745 /* strftime() with an extension for high-resolution timestamps. Any '#'s in
746 * 'format' will be replaced by subseconds, e.g. use "%S.###" to obtain results
749 strftime_msec(char *s
, size_t max
, const char *format
,
750 const struct tm_msec
*tm
)
754 /* Visual Studio 2013's behavior is to crash when 0 is passed as second
755 * argument to strftime. */
756 n
= max
? strftime(s
, max
, format
, &tm
->tm
) : 0;
761 sprintf(decimals
, "%03d", tm
->msec
);
762 for (p
= strchr(s
, '#'); p
; p
= strchr(p
, '#')) {
765 *p
++ = *d
? *d
++ : '0';
774 localtime_msec(long long int now
, struct tm_msec
*result
)
776 time_t now_sec
= now
/ 1000;
777 localtime_r(&now_sec
, &result
->tm
);
778 result
->msec
= now
% 1000;
783 gmtime_msec(long long int now
, struct tm_msec
*result
)
785 time_t now_sec
= now
/ 1000;
786 gmtime_r(&now_sec
, &result
->tm
);
787 result
->msec
= now
% 1000;