X-Git-Url: https://pintos-os.org/cgi-bin/gitweb.cgi?a=blobdiff_plain;f=lib%2Ftimeval.c;h=d29b66187c79789441e8879e2565b23c74014c4d;hb=e615b0a347057f9af7e93922acd4ae794ac87015;hp=3cca338fbac72d813a85dc83fe0ac705bd05fb75;hpb=6bc995e46d64b90ef0b3186abec92f63eab2c2ad;p=openvswitch diff --git a/lib/timeval.c b/lib/timeval.c index 3cca338f..d29b6618 100644 --- a/lib/timeval.c +++ b/lib/timeval.c @@ -1,5 +1,5 @@ /* - * Copyright (c) 2008, 2009 Nicira Networks. + * Copyright (c) 2008, 2009, 2010, 2011, 2012 Nicira, Inc. * * Licensed under the Apache License, Version 2.0 (the "License"); * you may not use this file except in compliance with the License. @@ -20,70 +20,180 @@ #include #include #include +#include #include #include #include #include #include "coverage.h" +#include "dummy.h" #include "fatal-signal.h" +#include "signals.h" +#include "unixctl.h" #include "util.h" - #include "vlog.h" -#define THIS_MODULE VLM_timeval -/* Initialized? */ -static bool inited; +VLOG_DEFINE_THIS_MODULE(timeval); + +/* The clock to use for measuring time intervals. This is CLOCK_MONOTONIC by + * preference, but on systems that don't have a monotonic clock we fall back + * to CLOCK_REALTIME. */ +static clockid_t monotonic_clock; -/* Has a timer tick occurred? */ -static volatile sig_atomic_t tick; +/* Has a timer tick occurred? + * + * We initialize these to true to force time_init() to get called on the first + * call to time_msec() or another function that queries the current time. */ +static volatile sig_atomic_t wall_tick = true; +static volatile sig_atomic_t monotonic_tick = true; /* The current time, as of the last refresh. */ -static struct timeval now; +static struct timespec wall_time; +static struct timespec monotonic_time; + +/* The monotonic time at which the time module was initialized. */ +static long long int boot_time; + +/* features for use by unit tests. */ +static struct timespec warp_offset; /* Offset added to monotonic_time. */ +static bool time_stopped; /* Disables real-time updates, if true. */ /* Time at which to die with SIGALRM (if not TIME_MIN). */ static time_t deadline = TIME_MIN; +static void set_up_timer(void); +static void set_up_signal(int flags); static void sigalrm_handler(int); -static void refresh_if_ticked(void); +static void refresh_wall_if_ticked(void); +static void refresh_monotonic_if_ticked(void); static time_t time_add(time_t, time_t); static void block_sigalrm(sigset_t *); static void unblock_sigalrm(const sigset_t *); -static void log_poll_interval(long long int last_wakeup, - const struct rusage *last_rusage); -static long long int timeval_to_msec(const struct timeval *); +static void log_poll_interval(long long int last_wakeup); +static struct rusage *get_recent_rusage(void); +static void refresh_rusage(void); +static void timespec_add(struct timespec *sum, + const struct timespec *a, const struct timespec *b); -/* Initializes the timetracking module. */ -void +/* Initializes the timetracking module, if not already initialized. */ +static void time_init(void) { - struct sigaction sa; - struct itimerval itimer; - + static bool inited; if (inited) { return; } + inited = true; coverage_init(); - inited = true; - gettimeofday(&now, NULL); - tick = false; + if (!clock_gettime(CLOCK_MONOTONIC, &monotonic_time)) { + monotonic_clock = CLOCK_MONOTONIC; + } else { + monotonic_clock = CLOCK_REALTIME; + VLOG_DBG("monotonic timer not available"); + } + + set_up_signal(SA_RESTART); + set_up_timer(); + boot_time = time_msec(); +} + +static void +set_up_signal(int flags) +{ + struct sigaction sa; - /* Set up signal handler. */ memset(&sa, 0, sizeof sa); sa.sa_handler = sigalrm_handler; sigemptyset(&sa.sa_mask); - sa.sa_flags = SA_RESTART; - if (sigaction(SIGALRM, &sa, NULL)) { - ovs_fatal(errno, "sigaction(SIGALRM) failed"); + sa.sa_flags = flags; + xsigaction(SIGALRM, &sa, NULL); +} + +/* Remove SA_RESTART from the flags for SIGALRM, so that any system call that + * is interrupted by the periodic timer interrupt will return EINTR instead of + * continuing after the signal handler returns. + * + * time_disable_restart() and time_enable_restart() may be usefully wrapped + * around function calls that might otherwise block forever unless interrupted + * by a signal, e.g.: + * + * time_disable_restart(); + * fcntl(fd, F_SETLKW, &lock); + * time_enable_restart(); + */ +void +time_disable_restart(void) +{ + time_init(); + set_up_signal(0); +} + +/* Add SA_RESTART to the flags for SIGALRM, so that any system call that + * is interrupted by the periodic timer interrupt will continue after the + * signal handler returns instead of returning EINTR. */ +void +time_enable_restart(void) +{ + time_init(); + set_up_signal(SA_RESTART); +} + +static void +set_up_timer(void) +{ + static timer_t timer_id; /* "static" to avoid apparent memory leak. */ + struct itimerspec itimer; + + if (timer_create(monotonic_clock, NULL, &timer_id)) { + VLOG_FATAL("timer_create failed (%s)", strerror(errno)); } - /* Set up periodic timer. */ itimer.it_interval.tv_sec = 0; - itimer.it_interval.tv_usec = TIME_UPDATE_INTERVAL * 1000; + itimer.it_interval.tv_nsec = TIME_UPDATE_INTERVAL * 1000 * 1000; itimer.it_value = itimer.it_interval; - if (setitimer(ITIMER_REAL, &itimer, NULL)) { - ovs_fatal(errno, "setitimer failed"); + + if (timer_settime(timer_id, 0, &itimer, NULL)) { + VLOG_FATAL("timer_settime failed (%s)", 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) +{ + time_init(); + set_up_timer(); +} + +static void +refresh_wall(void) +{ + time_init(); + clock_gettime(CLOCK_REALTIME, &wall_time); + wall_tick = false; +} + +static void +refresh_monotonic(void) +{ + time_init(); + + if (!time_stopped) { + if (monotonic_clock == CLOCK_MONOTONIC) { + clock_gettime(monotonic_clock, &monotonic_time); + } else { + refresh_wall_if_ticked(); + monotonic_time = wall_time; + } + timespec_add(&monotonic_time, &monotonic_time, &warp_offset); + + monotonic_tick = false; } } @@ -93,33 +203,68 @@ time_init(void) void time_refresh(void) { - gettimeofday(&now, NULL); - tick = false; + wall_tick = monotonic_tick = true; } -/* Returns the current time, in seconds. */ +/* Returns a monotonic timer, in seconds. */ time_t time_now(void) { - refresh_if_ticked(); - return now.tv_sec; + refresh_monotonic_if_ticked(); + return monotonic_time.tv_sec; } -/* Returns the current time, in ms (within TIME_UPDATE_INTERVAL ms). */ +/* Same as time_now() except does not write to static variables, for use in + * signal handlers. */ +static time_t +time_now_sig(void) +{ + struct timespec cur_time; + + clock_gettime(monotonic_clock, &cur_time); + return cur_time.tv_sec; +} + +/* Returns the current time, in seconds. */ +time_t +time_wall(void) +{ + refresh_wall_if_ticked(); + return wall_time.tv_sec; +} + +/* Returns a monotonic timer, in ms (within TIME_UPDATE_INTERVAL ms). */ long long int time_msec(void) { - refresh_if_ticked(); - return timeval_to_msec(&now); + refresh_monotonic_if_ticked(); + return timespec_to_msec(&monotonic_time); +} + +/* Returns the current time, in ms (within TIME_UPDATE_INTERVAL ms). */ +long long int +time_wall_msec(void) +{ + refresh_wall_if_ticked(); + return timespec_to_msec(&wall_time); +} + +/* Stores a monotonic timer, accurate within TIME_UPDATE_INTERVAL ms, into + * '*ts'. */ +void +time_timespec(struct timespec *ts) +{ + refresh_monotonic_if_ticked(); + *ts = monotonic_time; } /* Stores the current time, accurate within TIME_UPDATE_INTERVAL ms, into - * '*tv'. */ + * '*ts'. */ void -time_timeval(struct timeval *tv) +time_wall_timespec(struct timespec *ts) { - refresh_if_ticked(); - *tv = now; + refresh_wall_if_ticked(); + *ts = wall_time; } /* Configures the program to die with SIGALRM 'secs' seconds from now, if @@ -136,37 +281,44 @@ time_alarm(unsigned int secs) } /* Like poll(), except: + * + * - The timeout is specified as an absolute time, as defined by + * time_msec(), instead of a duration. * * - On error, returns a negative error code (instead of setting errno). * * - If interrupted by a signal, retries automatically until the original - * 'timeout' expires. (Because of this property, this function will + * 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, int timeout) +time_poll(struct pollfd *pollfds, int n_pollfds, long long int timeout_when, + int *elapsed) { static long long int last_wakeup; - static struct rusage last_rusage; long long int start; sigset_t oldsigs; bool blocked; int retval; time_refresh(); - log_poll_interval(last_wakeup, &last_rusage); + log_poll_interval(last_wakeup); coverage_clear(); start = time_msec(); blocked = false; for (;;) { + long long int now = time_msec(); int time_left; - if (timeout > 0) { - long long int elapsed = time_msec() - start; - time_left = timeout >= elapsed ? timeout - elapsed : 0; + + if (now >= timeout_when) { + time_left = 0; + } else if ((unsigned long long int) timeout_when - now > INT_MAX) { + time_left = INT_MAX; } else { - time_left = timeout; + time_left = timeout_when - now; } retval = poll(pollfds, n_pollfds, time_left); @@ -187,7 +339,8 @@ time_poll(struct pollfd *pollfds, int n_pollfds, int timeout) unblock_sigalrm(&oldsigs); } last_wakeup = time_msec(); - getrusage(RUSAGE_SELF, &last_rusage); + refresh_rusage(); + *elapsed = last_wakeup - start; return retval; } @@ -203,18 +356,26 @@ time_add(time_t a, time_t b) static void sigalrm_handler(int sig_nr) { - tick = true; - if (deadline != TIME_MIN && time(0) > deadline) { + wall_tick = true; + monotonic_tick = true; + if (deadline != TIME_MIN && time_now_sig() > deadline) { fatal_signal_handler(sig_nr); } } static void -refresh_if_ticked(void) +refresh_wall_if_ticked(void) { - assert(inited); - if (tick) { - time_refresh(); + if (wall_tick) { + refresh_wall(); + } +} + +static void +refresh_monotonic_if_ticked(void) +{ + if (monotonic_tick) { + refresh_monotonic(); } } @@ -224,25 +385,44 @@ block_sigalrm(sigset_t *oldsigs) sigset_t sigalrm; sigemptyset(&sigalrm); sigaddset(&sigalrm, SIGALRM); - if (sigprocmask(SIG_BLOCK, &sigalrm, oldsigs)) { - ovs_fatal(errno, "sigprocmask"); - } + xsigprocmask(SIG_BLOCK, &sigalrm, oldsigs); } static void unblock_sigalrm(const sigset_t *oldsigs) { - if (sigprocmask(SIG_SETMASK, oldsigs, NULL)) { - ovs_fatal(errno, "sigprocmask"); - } + xsigprocmask(SIG_SETMASK, oldsigs, NULL); } -static long long int +long long int +timespec_to_msec(const struct timespec *ts) +{ + return (long long int) ts->tv_sec * 1000 + ts->tv_nsec / (1000 * 1000); +} + +long long int timeval_to_msec(const struct timeval *tv) { return (long long int) tv->tv_sec * 1000 + tv->tv_usec / 1000; } +/* Returns the monotonic time at which the "time" module was initialized, in + * milliseconds(). */ +long long int +time_boot_msec(void) +{ + time_init(); + return boot_time; +} + +void +xgettimeofday(struct timeval *tv) +{ + if (gettimeofday(tv, NULL) == -1) { + VLOG_FATAL("gettimeofday failed (%s)", strerror(errno)); + } +} + static long long int timeval_diff_msec(const struct timeval *a, const struct timeval *b) { @@ -250,7 +430,24 @@ timeval_diff_msec(const struct timeval *a, const struct timeval *b) } static void -log_poll_interval(long long int last_wakeup, const struct rusage *last_rusage) +timespec_add(struct timespec *sum, + const struct timespec *a, + const struct timespec *b) +{ + struct timespec tmp; + + tmp.tv_sec = a->tv_sec + b->tv_sec; + tmp.tv_nsec = a->tv_nsec + b->tv_nsec; + if (tmp.tv_nsec >= 1000 * 1000 * 1000) { + tmp.tv_nsec -= 1000 * 1000 * 1000; + tmp.tv_sec++; + } + + *sum = tmp; +} + +static void +log_poll_interval(long long int last_wakeup) { static unsigned int mean_interval; /* In 16ths of a millisecond. */ static unsigned int n_samples; @@ -263,43 +460,46 @@ log_poll_interval(long long int last_wakeup, const struct rusage *last_rusage) now = time_msec(); interval = MIN(10000, now - last_wakeup) << 4; - /* Warn if we took too much time between polls. */ - if (n_samples > 10 && interval > mean_interval * 8) { - struct rusage rusage; - - getrusage(RUSAGE_SELF, &rusage); - VLOG_WARN("%u ms poll interval (%lld ms user, %lld ms system) " - "is over %u times the weighted mean interval %u ms " - "(%u samples)", - (interval + 8) / 16, - timeval_diff_msec(&rusage.ru_utime, &last_rusage->ru_utime), - timeval_diff_msec(&rusage.ru_stime, &last_rusage->ru_stime), - interval / mean_interval, - (mean_interval + 8) / 16, n_samples); - 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); + /* Warn if we took too much time between polls: at least 50 ms and at least + * 8X the mean interval. */ + if (n_samples > 10 && interval > mean_interval * 8 && interval > 50 * 16) { + static struct vlog_rate_limit rl = VLOG_RATE_LIMIT_INIT(1, 3); + + if (!VLOG_DROP_WARN(&rl)) { + const struct rusage *last_rusage = get_recent_rusage(); + struct rusage rusage; + + getrusage(RUSAGE_SELF, &rusage); + VLOG_WARN("%lld ms poll interval (%lld ms user, %lld ms system) " + "is over %u times the weighted mean interval %u ms " + "(%u samples)", + now - last_wakeup, + timeval_diff_msec(&rusage.ru_utime, + &last_rusage->ru_utime), + timeval_diff_msec(&rusage.ru_stime, + &last_rusage->ru_stime), + interval / mean_interval, + (mean_interval + 8) / 16, n_samples); + 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 (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); - } - - /* Care should be taken in the value chosen for logging. Depending - * on the configuration, syslog can write changes synchronously, - * which can cause the coverage messages to take longer to log - * than the processing delay that triggered it. */ - coverage_log(VLL_INFO, true); + coverage_log(); } /* Update exponentially weighted moving average. With these parameters, a @@ -310,3 +510,101 @@ log_poll_interval(long long int last_wakeup, const struct rusage *last_rusage) mean_interval = interval; } } + +/* CPU usage tracking. */ + +struct cpu_usage { + long long int when; /* Time that this sample was taken. */ + unsigned long long int cpu; /* Total user+system CPU usage when sampled. */ +}; + +static struct rusage recent_rusage; +static struct cpu_usage older = { LLONG_MIN, 0 }; +static struct cpu_usage newer = { LLONG_MIN, 0 }; +static int cpu_usage = -1; + +static struct rusage * +get_recent_rusage(void) +{ + return &recent_rusage; +} + +static void +refresh_rusage(void) +{ + long long int now; + + now = time_msec(); + getrusage(RUSAGE_SELF, &recent_rusage); + + if (now >= newer.when + 3 * 1000) { + older = newer; + newer.when = now; + newer.cpu = (timeval_to_msec(&recent_rusage.ru_utime) + + timeval_to_msec(&recent_rusage.ru_stime)); + + if (older.when != LLONG_MIN && newer.cpu > older.cpu) { + unsigned int dividend = newer.cpu - older.cpu; + unsigned int divisor = (newer.when - older.when) / 100; + cpu_usage = divisor > 0 ? dividend / divisor : -1; + } else { + cpu_usage = -1; + } + } +} + +/* Returns an estimate of this process's CPU usage, as a percentage, over the + * past few seconds of wall-clock time. Returns -1 if no estimate is available + * (which will happen if the process has not been running long enough to have + * an estimate, and can happen for other reasons as well). */ +int +get_cpu_usage(void) +{ + return cpu_usage; +} + +/* Unixctl interface. */ + +/* "time/stop" stops the monotonic time returned by e.g. time_msec() from + * advancing, except due to later calls to "time/warp". */ +static void +timeval_stop_cb(struct unixctl_conn *conn, + int argc OVS_UNUSED, const char *argv[] OVS_UNUSED, + void *aux OVS_UNUSED) +{ + time_stopped = true; + unixctl_command_reply(conn, NULL); +} + +/* "time/warp MSECS" advances the current monotonic time by the specified + * number of milliseconds. Unless "time/stop" has also been executed, the + * monotonic clock continues to tick forward at the normal rate afterward. + * + * 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) { + unixctl_command_reply_error(conn, "invalid MSECS"); + return; + } + + ts.tv_sec = msecs / 1000; + ts.tv_nsec = (msecs % 1000) * 1000 * 1000; + timespec_add(&warp_offset, &warp_offset, &ts); + timespec_add(&monotonic_time, &monotonic_time, &ts); + unixctl_command_reply(conn, "warped"); +} + +void +timeval_dummy_register(void) +{ + unixctl_command_register("time/stop", "", 0, 0, timeval_stop_cb, NULL); + unixctl_command_register("time/warp", "MSECS", 1, 1, + timeval_warp_cb, NULL); +}