/*
- * Copyright (c) 2008, 2009, 2010, 2011, 2012, 2013 Nicira, Inc.
+ * Copyright (c) 2008, 2009, 2010, 2011, 2012, 2013, 2014, 2015 Nicira, Inc.
*
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
#include "fatal-signal.h"
#include "hash.h"
#include "hmap.h"
+#include "ovs-rcu.h"
#include "ovs-thread.h"
#include "signals.h"
+#include "seq.h"
#include "unixctl.h"
#include "util.h"
-#include "vlog.h"
+#include "openvswitch/vlog.h"
VLOG_DEFINE_THIS_MODULE(timeval);
+#ifdef _WIN32
+typedef unsigned int clockid_t;
+
+#ifndef CLOCK_MONOTONIC
+#define CLOCK_MONOTONIC 1
+#endif
+
+#ifndef CLOCK_REALTIME
+#define CLOCK_REALTIME 2
+#endif
+
+/* Number of 100 ns intervals from January 1, 1601 till January 1, 1970. */
+const static unsigned long long unix_epoch = 116444736000000000;
+#endif /* _WIN32 */
+
+/* Structure set by unixctl time/warp command. */
+struct large_warp {
+ struct unixctl_conn *conn; /* Connection waiting for warp response. */
+ long long int total_warp; /* Total offset to be added to monotonic time. */
+ long long int warp; /* 'total_warp' offset done in steps of 'warp'. */
+ unsigned int main_thread_id; /* Identification for the main thread. */
+};
+
struct clock {
clockid_t id; /* CLOCK_MONOTONIC or CLOCK_REALTIME. */
- struct ovs_rwlock rwlock; /* Mutual exclusion for 'cache'. */
- /* Features for use by unit tests. Protected by 'rwlock'. */
- struct timespec warp; /* Offset added for unit tests. */
- bool stopped; /* Disables real-time updates if true. */
-
- /* Relevant only if CACHE_TIME is true. */
- volatile sig_atomic_t tick; /* Has the timer ticked? Set by signal. */
- struct timespec cache; /* Last time read from kernel. */
+ /* Features for use by unit tests. Protected by 'mutex'. */
+ struct ovs_mutex mutex;
+ atomic_bool slow_path; /* True if warped or stopped. */
+ struct timespec warp OVS_GUARDED; /* Offset added for unit tests. */
+ bool stopped OVS_GUARDED; /* Disable real-time updates if true. */
+ struct timespec cache OVS_GUARDED; /* Last time read from kernel. */
+ struct large_warp large_warp OVS_GUARDED; /* Connection information waiting
+ for warp response. */
};
/* Our clocks. */
/* The monotonic time at which the time module was initialized. */
static long long int boot_time;
+/* True only when timeval_dummy_register() is called. */
+static bool timewarp_enabled;
+/* Reference to the seq struct. Threads other than main thread can
+ * wait on timewarp_seq and be waken up when time is warped. */
+static struct seq *timewarp_seq;
+/* Last value of 'timewarp_seq'. */
+DEFINE_STATIC_PER_THREAD_DATA(uint64_t, last_seq, 0);
+
/* Monotonic time in milliseconds at which to die with SIGALRM (if not
* LLONG_MAX). */
static long long int deadline = LLONG_MAX;
/* Monotonic time, in milliseconds, at which the last call to time_poll() woke
* up. */
-DEFINE_PER_THREAD_DATA(long long int, last_wakeup, 0);
+DEFINE_STATIC_PER_THREAD_DATA(long long int, last_wakeup, 0);
-static void set_up_timer(void);
-static void set_up_signal(int flags);
-static void sigalrm_handler(int);
-static void block_sigalrm(sigset_t *);
-static void unblock_sigalrm(const sigset_t *);
static void log_poll_interval(long long int last_wakeup);
static struct rusage *get_recent_rusage(void);
+static int getrusage_thread(struct rusage *);
static void refresh_rusage(void);
static void timespec_add(struct timespec *sum,
const struct timespec *a, const struct timespec *b);
{
memset(c, 0, sizeof *c);
c->id = id;
- ovs_rwlock_init(&c->rwlock);
+ ovs_mutex_init(&c->mutex);
+ atomic_init(&c->slow_path, false);
xclock_gettime(c->id, &c->cache);
}
coverage_init();
+ timewarp_seq = seq_create();
init_clock(&monotonic_clock, (!clock_gettime(CLOCK_MONOTONIC, &ts)
? CLOCK_MONOTONIC
: CLOCK_REALTIME));
init_clock(&wall_clock, CLOCK_REALTIME);
boot_time = timespec_to_msec(&monotonic_clock.cache);
-
- set_up_signal(SA_RESTART);
- set_up_timer();
}
/* Initializes the timetracking module, if not already initialized. */
pthread_once(&once, do_init_time);
}
-static void
-set_up_signal(int flags)
-{
- struct sigaction sa;
-
- memset(&sa, 0, sizeof sa);
- sa.sa_handler = sigalrm_handler;
- sigemptyset(&sa.sa_mask);
- sa.sa_flags = flags;
- xsigaction(SIGALRM, &sa, NULL);
-}
-
-static void
-set_up_timer(void)
-{
- static timer_t timer_id; /* "static" to avoid apparent memory leak. */
- struct itimerspec itimer;
-
- if (!CACHE_TIME) {
- return;
- }
-
- if (timer_create(monotonic_clock.id, NULL, &timer_id)) {
- VLOG_FATAL("timer_create failed (%s)", ovs_strerror(errno));
- }
-
- itimer.it_interval.tv_sec = 0;
- itimer.it_interval.tv_nsec = TIME_UPDATE_INTERVAL * 1000 * 1000;
- itimer.it_value = itimer.it_interval;
-
- if (timer_settime(timer_id, 0, &itimer, NULL)) {
- VLOG_FATAL("timer_settime failed (%s)", ovs_strerror(errno));
- }
-}
-
-/* Set up the interval timer, to ensure that time advances even without calling
- * time_refresh().
- *
- * A child created with fork() does not inherit the parent's interval timer, so
- * this function needs to be called from the child after fork(). */
-void
-time_postfork(void)
-{
- assert_single_threaded();
- time_init();
- set_up_timer();
-}
-
-/* Forces a refresh of the current time from the kernel. It is not usually
- * necessary to call this function, since the time will be refreshed
- * automatically at least every TIME_UPDATE_INTERVAL milliseconds. If
- * CACHE_TIME is false, we will always refresh the current time so this
- * function has no effect. */
-void
-time_refresh(void)
-{
- monotonic_clock.tick = wall_clock.tick = true;
-}
-
static void
time_timespec__(struct clock *c, struct timespec *ts)
{
+ bool slow_path;
+
time_init();
- for (;;) {
- /* Use the cached time by preference, but fall through if there's been
- * a clock tick. */
- ovs_rwlock_rdlock(&c->rwlock);
- if (c->stopped || !c->tick) {
- timespec_add(ts, &c->cache, &c->warp);
- ovs_rwlock_unlock(&c->rwlock);
- return;
- }
- ovs_rwlock_unlock(&c->rwlock);
- /* Refresh the cache. */
- ovs_rwlock_wrlock(&c->rwlock);
- if (c->tick) {
- c->tick = false;
- xclock_gettime(c->id, &c->cache);
+ atomic_read_relaxed(&c->slow_path, &slow_path);
+ if (!slow_path) {
+ xclock_gettime(c->id, ts);
+ } else {
+ struct timespec warp;
+ struct timespec cache;
+ bool stopped;
+
+ ovs_mutex_lock(&c->mutex);
+ stopped = c->stopped;
+ warp = c->warp;
+ cache = c->cache;
+ ovs_mutex_unlock(&c->mutex);
+
+ if (!stopped) {
+ xclock_gettime(c->id, &cache);
}
- ovs_rwlock_unlock(&c->rwlock);
+ timespec_add(ts, &cache, &warp);
}
}
assert_single_threaded();
time_init();
- time_refresh();
now = time_msec();
msecs = secs * 1000LL;
* timeout is reached. (Because of this property, this function will
* never return -EINTR.)
*
- * - As a side effect, refreshes the current time (like time_refresh()).
- *
* Stores the number of milliseconds elapsed during poll in '*elapsed'. */
int
-time_poll(struct pollfd *pollfds, int n_pollfds, long long int timeout_when,
- int *elapsed)
+time_poll(struct pollfd *pollfds, int n_pollfds, HANDLE *handles OVS_UNUSED,
+ long long int timeout_when, int *elapsed)
{
long long int *last_wakeup = last_wakeup_get();
long long int start;
- sigset_t oldsigs;
- bool blocked;
- int retval;
+ bool quiescent;
+ int retval = 0;
time_init();
- time_refresh();
- if (*last_wakeup) {
+ coverage_clear();
+ coverage_run();
+ if (*last_wakeup && !thread_is_pmd()) {
log_poll_interval(*last_wakeup);
}
- coverage_clear();
start = time_msec();
- blocked = false;
timeout_when = MIN(timeout_when, deadline);
+ quiescent = ovsrcu_is_quiescent();
for (;;) {
long long int now = time_msec();
time_left = timeout_when - now;
}
+ if (!quiescent) {
+ if (!time_left) {
+ ovsrcu_quiesce();
+ } else {
+ ovsrcu_quiesce_start();
+ }
+ }
+
+#ifndef _WIN32
retval = poll(pollfds, n_pollfds, time_left);
if (retval < 0) {
retval = -errno;
}
+#else
+ if (n_pollfds > MAXIMUM_WAIT_OBJECTS) {
+ VLOG_ERR("Cannot handle more than maximum wait objects\n");
+ } else if (n_pollfds != 0) {
+ retval = WaitForMultipleObjects(n_pollfds, handles, FALSE,
+ time_left);
+ }
+ if (retval < 0) {
+ /* XXX This will be replace by a win error to errno
+ conversion function */
+ retval = -WSAGetLastError();
+ retval = -EINVAL;
+ }
+#endif
+
+ if (!quiescent && time_left) {
+ ovsrcu_quiesce_end();
+ }
- time_refresh();
if (deadline <= time_msec()) {
+#ifndef _WIN32
fatal_signal_handler(SIGALRM);
+#else
+ VLOG_ERR("wake up from WaitForMultipleObjects after deadline");
+ fatal_signal_handler(SIGTERM);
+#endif
if (retval < 0) {
retval = 0;
}
if (retval != -EINTR) {
break;
}
-
- if (!blocked && CACHE_TIME) {
- block_sigalrm(&oldsigs);
- blocked = true;
- }
- }
- if (blocked) {
- unblock_sigalrm(&oldsigs);
}
*last_wakeup = time_msec();
refresh_rusage();
return retval;
}
-static void
-sigalrm_handler(int sig_nr OVS_UNUSED)
-{
- monotonic_clock.tick = wall_clock.tick = true;
-}
-
-static void
-block_sigalrm(sigset_t *oldsigs)
-{
- sigset_t sigalrm;
- sigemptyset(&sigalrm);
- sigaddset(&sigalrm, SIGALRM);
- xpthread_sigmask(SIG_BLOCK, &sigalrm, oldsigs);
-}
-
-static void
-unblock_sigalrm(const sigset_t *oldsigs)
-{
- xpthread_sigmask(SIG_SETMASK, oldsigs, NULL);
-}
-
long long int
timespec_to_msec(const struct timespec *ts)
{
return boot_time;
}
+#ifdef _WIN32
+static ULARGE_INTEGER
+xgetfiletime(void)
+{
+ ULARGE_INTEGER current_time;
+ FILETIME current_time_ft;
+
+ /* Returns current time in UTC as a 64-bit value representing the number
+ * of 100-nanosecond intervals since January 1, 1601 . */
+ GetSystemTimePreciseAsFileTime(¤t_time_ft);
+ current_time.LowPart = current_time_ft.dwLowDateTime;
+ current_time.HighPart = current_time_ft.dwHighDateTime;
+
+ return current_time;
+}
+
+static int
+clock_gettime(clock_t id, struct timespec *ts)
+{
+ if (id == CLOCK_MONOTONIC) {
+ static LARGE_INTEGER freq;
+ LARGE_INTEGER count;
+ long long int ns;
+
+ if (!freq.QuadPart) {
+ /* Number of counts per second. */
+ QueryPerformanceFrequency(&freq);
+ }
+ /* Total number of counts from a starting point. */
+ QueryPerformanceCounter(&count);
+
+ /* Total nano seconds from a starting point. */
+ ns = (double) count.QuadPart / freq.QuadPart * 1000000000;
+
+ ts->tv_sec = count.QuadPart / freq.QuadPart;
+ ts->tv_nsec = ns % 1000000000;
+ } else if (id == CLOCK_REALTIME) {
+ ULARGE_INTEGER current_time = xgetfiletime();
+
+ /* Time from Epoch to now. */
+ ts->tv_sec = (current_time.QuadPart - unix_epoch) / 10000000;
+ ts->tv_nsec = ((current_time.QuadPart - unix_epoch) %
+ 10000000) * 100;
+ } else {
+ return -1;
+ }
+
+ return 0;
+}
+#endif /* _WIN32 */
+
void
xgettimeofday(struct timeval *tv)
{
+#ifndef _WIN32
if (gettimeofday(tv, NULL) == -1) {
VLOG_FATAL("gettimeofday failed (%s)", ovs_strerror(errno));
}
+#else
+ ULARGE_INTEGER current_time = xgetfiletime();
+
+ tv->tv_sec = (current_time.QuadPart - unix_epoch) / 10000000;
+ tv->tv_usec = ((current_time.QuadPart - unix_epoch) %
+ 10000000) / 10;
+#endif
}
void
}
}
+static void
+msec_to_timespec(long long int ms, struct timespec *ts)
+{
+ ts->tv_sec = ms / 1000;
+ ts->tv_nsec = (ms % 1000) * 1000 * 1000;
+}
+
+static void
+timewarp_work(void)
+{
+ struct clock *c = &monotonic_clock;
+ struct timespec warp;
+
+ ovs_mutex_lock(&c->mutex);
+ if (!c->large_warp.conn) {
+ ovs_mutex_unlock(&c->mutex);
+ return;
+ }
+
+ if (c->large_warp.total_warp >= c->large_warp.warp) {
+ msec_to_timespec(c->large_warp.warp, &warp);
+ timespec_add(&c->warp, &c->warp, &warp);
+ c->large_warp.total_warp -= c->large_warp.warp;
+ } else if (c->large_warp.total_warp) {
+ msec_to_timespec(c->large_warp.total_warp, &warp);
+ timespec_add(&c->warp, &c->warp, &warp);
+ c->large_warp.total_warp = 0;
+ } else {
+ /* c->large_warp.total_warp is 0. */
+ msec_to_timespec(c->large_warp.warp, &warp);
+ timespec_add(&c->warp, &c->warp, &warp);
+ }
+
+ if (!c->large_warp.total_warp) {
+ unixctl_command_reply(c->large_warp.conn, "warped");
+ c->large_warp.conn = NULL;
+ }
+
+ ovs_mutex_unlock(&c->mutex);
+ seq_change(timewarp_seq);
+
+ /* give threads (eg. monitor) some chances to run */
+#ifndef _WIN32
+ poll(NULL, 0, 10);
+#else
+ Sleep(10);
+#endif
+}
+
+/* Perform work needed for "timewarp_seq"'s producer and consumers. */
+void
+timewarp_run(void)
+{
+ /* The function is a no-op unless timeval_dummy_register() is called. */
+ if (timewarp_enabled) {
+ unsigned int thread_id;
+ ovs_mutex_lock(&monotonic_clock.mutex);
+ thread_id = monotonic_clock.large_warp.main_thread_id;
+ ovs_mutex_unlock(&monotonic_clock.mutex);
+
+ if (thread_id != ovsthread_id_self()) {
+ /* For threads other than the thread that changes the sequence,
+ * wait on it. */
+ uint64_t *last_seq = last_seq_get();
+
+ *last_seq = seq_read(timewarp_seq);
+ seq_wait(timewarp_seq, *last_seq);
+ } else {
+ /* Work on adding the remaining warps. */
+ timewarp_work();
+ }
+ }
+}
+
static long long int
timeval_diff_msec(const struct timeval *a, const struct timeval *b)
{
*sum = tmp;
}
+static bool
+is_warped(const struct clock *c)
+{
+ bool warped;
+
+ ovs_mutex_lock(&c->mutex);
+ warped = monotonic_clock.warp.tv_sec || monotonic_clock.warp.tv_nsec;
+ ovs_mutex_unlock(&c->mutex);
+
+ return warped;
+}
+
static void
log_poll_interval(long long int last_wakeup)
{
long long int interval = time_msec() - last_wakeup;
- if (interval >= 1000
- && !monotonic_clock.warp.tv_sec
- && !monotonic_clock.warp.tv_nsec) {
+ if (interval >= 1000 && !is_warped(&monotonic_clock)) {
const struct rusage *last_rusage = get_recent_rusage();
struct rusage rusage;
- getrusage(RUSAGE_SELF, &rusage);
- VLOG_WARN("Unreasonably long %lldms poll interval"
- " (%lldms user, %lldms system)",
- interval,
- timeval_diff_msec(&rusage.ru_utime,
- &last_rusage->ru_utime),
- timeval_diff_msec(&rusage.ru_stime,
- &last_rusage->ru_stime));
- if (rusage.ru_minflt > last_rusage->ru_minflt
- || rusage.ru_majflt > last_rusage->ru_majflt) {
- VLOG_WARN("faults: %ld minor, %ld major",
- rusage.ru_minflt - last_rusage->ru_minflt,
- rusage.ru_majflt - last_rusage->ru_majflt);
- }
- if (rusage.ru_inblock > last_rusage->ru_inblock
- || rusage.ru_oublock > last_rusage->ru_oublock) {
- VLOG_WARN("disk: %ld reads, %ld writes",
- rusage.ru_inblock - last_rusage->ru_inblock,
- rusage.ru_oublock - last_rusage->ru_oublock);
- }
- if (rusage.ru_nvcsw > last_rusage->ru_nvcsw
- || rusage.ru_nivcsw > last_rusage->ru_nivcsw) {
- VLOG_WARN("context switches: %ld voluntary, %ld involuntary",
- rusage.ru_nvcsw - last_rusage->ru_nvcsw,
- rusage.ru_nivcsw - last_rusage->ru_nivcsw);
+ if (!getrusage_thread(&rusage)) {
+ VLOG_WARN("Unreasonably long %lldms poll interval"
+ " (%lldms user, %lldms system)",
+ interval,
+ timeval_diff_msec(&rusage.ru_utime,
+ &last_rusage->ru_utime),
+ timeval_diff_msec(&rusage.ru_stime,
+ &last_rusage->ru_stime));
+
+ if (rusage.ru_minflt > last_rusage->ru_minflt
+ || rusage.ru_majflt > last_rusage->ru_majflt) {
+ VLOG_WARN("faults: %ld minor, %ld major",
+ rusage.ru_minflt - last_rusage->ru_minflt,
+ rusage.ru_majflt - last_rusage->ru_majflt);
+ }
+ if (rusage.ru_inblock > last_rusage->ru_inblock
+ || rusage.ru_oublock > last_rusage->ru_oublock) {
+ VLOG_WARN("disk: %ld reads, %ld writes",
+ rusage.ru_inblock - last_rusage->ru_inblock,
+ rusage.ru_oublock - last_rusage->ru_oublock);
+ }
+ if (rusage.ru_nvcsw > last_rusage->ru_nvcsw
+ || rusage.ru_nivcsw > last_rusage->ru_nivcsw) {
+ VLOG_WARN("context switches: %ld voluntary, %ld involuntary",
+ rusage.ru_nvcsw - last_rusage->ru_nvcsw,
+ rusage.ru_nivcsw - last_rusage->ru_nivcsw);
+ }
+ } else {
+ VLOG_WARN("Unreasonably long %lldms poll interval", interval);
}
coverage_log();
}
int argc OVS_UNUSED, const char *argv[] OVS_UNUSED,
void *aux OVS_UNUSED)
{
- ovs_rwlock_wrlock(&monotonic_clock.rwlock);
+ ovs_mutex_lock(&monotonic_clock.mutex);
+ atomic_store_relaxed(&monotonic_clock.slow_path, true);
monotonic_clock.stopped = true;
- ovs_rwlock_unlock(&monotonic_clock.rwlock);
+ xclock_gettime(monotonic_clock.id, &monotonic_clock.cache);
+ ovs_mutex_unlock(&monotonic_clock.mutex);
unixctl_command_reply(conn, NULL);
}
* number of milliseconds. Unless "time/stop" has also been executed, the
* monotonic clock continues to tick forward at the normal rate afterward.
*
+ * "time/warp LARGE_MSECS MSECS" is a variation of the above command. It
+ * advances the current monotonic time by LARGE_MSECS. This is done MSECS
+ * at a time in each run of the main thread. This gives other threads
+ * time to run after the clock has been advanced by MSECS.
+ *
* Does not affect wall clock readings. */
static void
timeval_warp_cb(struct unixctl_conn *conn,
int argc OVS_UNUSED, const char *argv[], void *aux OVS_UNUSED)
{
- struct timespec ts;
- int msecs;
-
- msecs = atoi(argv[1]);
- if (msecs <= 0) {
+ long long int total_warp = argc > 2 ? atoll(argv[1]) : 0;
+ long long int msecs = argc > 2 ? atoll(argv[2]) : atoll(argv[1]);
+ if (msecs <= 0 || total_warp < 0) {
unixctl_command_reply_error(conn, "invalid MSECS");
return;
}
- ts.tv_sec = msecs / 1000;
- ts.tv_nsec = (msecs % 1000) * 1000 * 1000;
-
- ovs_rwlock_wrlock(&monotonic_clock.rwlock);
- timespec_add(&monotonic_clock.warp, &monotonic_clock.warp, &ts);
- ovs_rwlock_unlock(&monotonic_clock.rwlock);
+ ovs_mutex_lock(&monotonic_clock.mutex);
+ if (monotonic_clock.large_warp.conn) {
+ ovs_mutex_unlock(&monotonic_clock.mutex);
+ unixctl_command_reply_error(conn, "A previous warp in progress");
+ return;
+ }
+ atomic_store_relaxed(&monotonic_clock.slow_path, true);
+ monotonic_clock.large_warp.conn = conn;
+ monotonic_clock.large_warp.total_warp = total_warp;
+ monotonic_clock.large_warp.warp = msecs;
+ monotonic_clock.large_warp.main_thread_id = ovsthread_id_self();
+ ovs_mutex_unlock(&monotonic_clock.mutex);
- unixctl_command_reply(conn, "warped");
+ timewarp_work();
}
void
timeval_dummy_register(void)
{
+ timewarp_enabled = true;
unixctl_command_register("time/stop", "", 0, 0, timeval_stop_cb, NULL);
- unixctl_command_register("time/warp", "MSECS", 1, 1,
+ unixctl_command_register("time/warp", "[large_msecs] msecs", 1, 2,
timeval_warp_cb, NULL);
}
+
+
+
+/* strftime() with an extension for high-resolution timestamps. Any '#'s in
+ * 'format' will be replaced by subseconds, e.g. use "%S.###" to obtain results
+ * like "01.123". */
+size_t
+strftime_msec(char *s, size_t max, const char *format,
+ const struct tm_msec *tm)
+{
+ size_t n;
+
+ /* Visual Studio 2013's behavior is to crash when 0 is passed as second
+ * argument to strftime. */
+ n = max ? strftime(s, max, format, &tm->tm) : 0;
+ if (n) {
+ char decimals[4];
+ char *p;
+
+ sprintf(decimals, "%03d", tm->msec);
+ for (p = strchr(s, '#'); p; p = strchr(p, '#')) {
+ char *d = decimals;
+ while (*p == '#') {
+ *p++ = *d ? *d++ : '0';
+ }
+ }
+ }
+
+ return n;
+}
+
+struct tm_msec *
+localtime_msec(long long int now, struct tm_msec *result)
+{
+ time_t now_sec = now / 1000;
+ localtime_r(&now_sec, &result->tm);
+ result->msec = now % 1000;
+ return result;
+}
+
+struct tm_msec *
+gmtime_msec(long long int now, struct tm_msec *result)
+{
+ time_t now_sec = now / 1000;
+ gmtime_r(&now_sec, &result->tm);
+ result->msec = now % 1000;
+ return result;
+}
projects / cascardo / ovs.git / blobdiff