X-Git-Url: http://git.onelab.eu/?a=blobdiff_plain;f=lib%2Ftimeval.c;h=2f003552825919442ff0958099c9d4820764c627;hb=67680b012be30d9c26eab999b83b08c6eb32dbd2;hp=3cca338fbac72d813a85dc83fe0ac705bd05fb75;hpb=6bc995e46d64b90ef0b3186abec92f63eab2c2ad;p=sliver-openvswitch.git diff --git a/lib/timeval.c b/lib/timeval.c index 3cca338fb..2f0035528 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, 2013 Nicira, Inc. * * Licensed under the Apache License, Version 2.0 (the "License"); * you may not use this file except in compliance with the License. @@ -16,110 +16,302 @@ #include #include "timeval.h" -#include #include +#if HAVE_EXECINFO_H +#include +#endif #include #include +#include #include #include #include #include #include "coverage.h" +#include "dummy.h" +#include "dynamic-string.h" #include "fatal-signal.h" +#include "hash.h" +#include "hmap.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? Only relevant if CACHE_TIME is true. + * + * 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 in milliseconds at which to die with SIGALRM (if not LLONG_MAX). */ +static long long int deadline = LLONG_MAX; + +struct trace { + void *backtrace[32]; /* Populated by backtrace(). */ + size_t n_frames; /* Number of frames in 'backtrace'. */ -/* Time at which to die with SIGALRM (if not TIME_MIN). */ -static time_t deadline = TIME_MIN; + /* format_backtraces() helper data. */ + struct hmap_node node; + size_t count; +}; +#define MAX_TRACES 50 +static struct trace traces[MAX_TRACES]; +static size_t trace_head = 0; + +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 time_t time_add(time_t, time_t); +static void refresh_wall_if_ticked(void); +static void refresh_monotonic_if_ticked(void); 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); +static unixctl_cb_func backtrace_cb; + +#ifndef HAVE_EXECINFO_H +#define HAVE_EXECINFO_H 0 + +static int +backtrace(void **buffer OVS_UNUSED, int size OVS_UNUSED) +{ + NOT_REACHED(); +} -/* Initializes the timetracking module. */ -void +static char ** +backtrace_symbols(void *const *buffer OVS_UNUSED, int size OVS_UNUSED) +{ + NOT_REACHED(); +} +#endif + +/* 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; + + /* The implementation of backtrace() in glibc does some one time + * initialization which is not signal safe. This can cause deadlocks if + * run from the signal handler. As a workaround, force the initialization + * to happen here. */ + if (HAVE_EXECINFO_H) { + void *bt[1]; + + backtrace(bt, ARRAY_SIZE(bt)); + } + + memset(traces, 0, sizeof traces); + + if (HAVE_EXECINFO_H && CACHE_TIME) { + unixctl_command_register("backtrace", "", 0, 0, backtrace_cb, NULL); + } 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 (!CACHE_TIME) { + return; + } + + 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; } } /* 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. */ + * 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) { - 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). */ +/* 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 @@ -127,58 +319,87 @@ time_timeval(struct timeval *tv) void time_alarm(unsigned int secs) { + long long int now; + long long int msecs; + sigset_t oldsigs; time_init(); + time_refresh(); + + now = time_msec(); + msecs = secs * 1000LL; + block_sigalrm(&oldsigs); - deadline = secs ? time_add(time_now(), secs) : TIME_MIN; + deadline = now < LLONG_MAX - msecs ? now + msecs : LLONG_MAX; unblock_sigalrm(&oldsigs); } /* 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; + static long long int last_wakeup = 0; long long int start; sigset_t oldsigs; bool blocked; int retval; time_refresh(); - log_poll_interval(last_wakeup, &last_rusage); + if (last_wakeup) { + log_poll_interval(last_wakeup); + } coverage_clear(); start = time_msec(); blocked = false; + + timeout_when = MIN(timeout_when, deadline); + 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); if (retval < 0) { retval = -errno; } + time_refresh(); + if (deadline <= time_msec()) { + fatal_signal_handler(SIGALRM); + if (retval < 0) { + retval = 0; + } + break; + } + if (retval != -EINTR) { break; } - if (!blocked && deadline == TIME_MIN) { + if (!blocked && CACHE_TIME) { block_sigalrm(&oldsigs); blocked = true; } @@ -187,34 +408,39 @@ 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; } -/* Returns the sum of 'a' and 'b', with saturation on overflow or underflow. */ -static time_t -time_add(time_t a, time_t b) +static void +sigalrm_handler(int sig_nr OVS_UNUSED) { - return (a >= 0 - ? (b > TIME_MAX - a ? TIME_MAX : a + b) - : (b < TIME_MIN - a ? TIME_MIN : a + b)); + wall_tick = true; + monotonic_tick = true; + + if (HAVE_EXECINFO_H && CACHE_TIME) { + struct trace *trace = &traces[trace_head]; + + trace->n_frames = backtrace(trace->backtrace, + ARRAY_SIZE(trace->backtrace)); + trace_head = (trace_head + 1) % MAX_TRACES; + } } static void -sigalrm_handler(int sig_nr) +refresh_wall_if_ticked(void) { - tick = true; - if (deadline != TIME_MIN && time(0) > deadline) { - fatal_signal_handler(sig_nr); + if (!CACHE_TIME || wall_tick) { + refresh_wall(); } } static void -refresh_if_ticked(void) +refresh_monotonic_if_ticked(void) { - assert(inited); - if (tick) { - time_refresh(); + if (!CACHE_TIME || monotonic_tick) { + refresh_monotonic(); } } @@ -224,25 +450,44 @@ 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) { @@ -250,32 +495,39 @@ 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) { - static unsigned int mean_interval; /* In 16ths of a millisecond. */ - static unsigned int n_samples; + struct timespec tmp; - long long int now; - unsigned int interval; /* In 16ths of a millisecond. */ + 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++; + } - /* Compute interval from last wakeup to now in 16ths of a millisecond, - * capped at 10 seconds (16000 in this unit). */ - now = time_msec(); - interval = MIN(10000, now - last_wakeup) << 4; + *sum = tmp; +} + +static void +log_poll_interval(long long int last_wakeup) +{ + long long int interval = time_msec() - last_wakeup; - /* Warn if we took too much time between polls. */ - if (n_samples > 10 && interval > mean_interval * 8) { + if (interval >= 1000 && !warp_offset.tv_sec && !warp_offset.tv_nsec) { + const struct rusage *last_rusage = get_recent_rusage(); 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); + 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", @@ -294,19 +546,200 @@ log_poll_interval(long long int last_wakeup, const struct rusage *last_rusage) rusage.ru_nvcsw - last_rusage->ru_nvcsw, rusage.ru_nivcsw - last_rusage->ru_nivcsw); } + coverage_log(); + } +} + +/* 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; +} - /* 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); +static uint32_t +hash_trace(struct trace *trace) +{ + return hash_bytes(trace->backtrace, + trace->n_frames * sizeof *trace->backtrace, 0); +} + +static struct trace * +trace_map_lookup(struct hmap *trace_map, struct trace *key) +{ + struct trace *value; + + HMAP_FOR_EACH_WITH_HASH (value, node, hash_trace(key), trace_map) { + if (key->n_frames == value->n_frames + && !memcmp(key->backtrace, value->backtrace, + key->n_frames * sizeof *key->backtrace)) { + return value; + } } + return NULL; +} - /* Update exponentially weighted moving average. With these parameters, a - * given value decays to 1% of its value in about 100 time steps. */ - if (n_samples++) { - mean_interval = (mean_interval * 122 + interval * 6 + 64) / 128; - } else { - mean_interval = interval; +/* Appends a string to 'ds' representing backtraces recorded at regular + * intervals in the recent past. This information can be used to get a sense + * of what the process has been spending the majority of time doing. Will + * ommit any backtraces which have not occurred at least 'min_count' times. */ +void +format_backtraces(struct ds *ds, size_t min_count) +{ + time_init(); + + if (HAVE_EXECINFO_H && CACHE_TIME) { + struct hmap trace_map = HMAP_INITIALIZER(&trace_map); + struct trace *trace, *next; + sigset_t oldsigs; + size_t i; + + block_sigalrm(&oldsigs); + + for (i = 0; i < MAX_TRACES; i++) { + struct trace *trace = &traces[i]; + struct trace *map_trace; + + if (!trace->n_frames) { + continue; + } + + map_trace = trace_map_lookup(&trace_map, trace); + if (map_trace) { + map_trace->count++; + } else { + hmap_insert(&trace_map, &trace->node, hash_trace(trace)); + trace->count = 1; + } + } + + HMAP_FOR_EACH_SAFE (trace, next, node, &trace_map) { + char **frame_strs; + size_t j; + + hmap_remove(&trace_map, &trace->node); + + if (trace->count < min_count) { + continue; + } + + frame_strs = backtrace_symbols(trace->backtrace, trace->n_frames); + + ds_put_format(ds, "Count %zu\n", trace->count); + for (j = 0; j < trace->n_frames; j++) { + ds_put_format(ds, "%s\n", frame_strs[j]); + } + ds_put_cstr(ds, "\n"); + + free(frame_strs); + } + hmap_destroy(&trace_map); + + ds_chomp(ds, '\n'); + unblock_sigalrm(&oldsigs); } } + +/* 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"); +} + +static void +backtrace_cb(struct unixctl_conn *conn, + int argc OVS_UNUSED, const char *argv[] OVS_UNUSED, + void *aux OVS_UNUSED) +{ + struct ds ds = DS_EMPTY_INITIALIZER; + + ovs_assert(HAVE_EXECINFO_H && CACHE_TIME); + format_backtraces(&ds, 0); + unixctl_command_reply(conn, ds_cstr(&ds)); + ds_destroy(&ds); +} + +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); +}