lib/timeval.c

   1 /*
   2  * Copyright (c) 2008, 2009, 2010, 2011, 2012, 2013 Nicira, Inc.
   3  *
   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:
   7  *
   8  *     http://www.apache.org/licenses/LICENSE-2.0
   9  *
  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.
  15  */
  16
  17 #include <config.h>
  18 #include "timeval.h"
  19 #include <errno.h>
  20 #include <poll.h>
  21 #include <pthread.h>
  22 #include <signal.h>
  23 #include <stdlib.h>
  24 #include <string.h>
  25 #include <sys/time.h>
  26 #include <sys/resource.h>
  27 #include <unistd.h>
  28 #include "coverage.h"
  29 #include "dummy.h"
  30 #include "dynamic-string.h"
  31 #include "fatal-signal.h"
  32 #include "hash.h"
  33 #include "hmap.h"
  34 #include "ovs-thread.h"
  35 #include "signals.h"
  36 #include "unixctl.h"
  37 #include "util.h"
  38 #include "vlog.h"
  39
  40 VLOG_DEFINE_THIS_MODULE(timeval);
  41
  42 struct clock {
  43     clockid_t id;               /* CLOCK_MONOTONIC or CLOCK_REALTIME. */
  44     pthread_rwlock_t rwlock;    /* Mutual exclusion for 'cache'. */
  45
  46     /* Features for use by unit tests.  Protected by 'rwlock'. */
  47     struct timespec warp;       /* Offset added for unit tests. */
  48     bool stopped;               /* Disables real-time updates if true.  */
  49
  50     /* Relevant only if CACHE_TIME is true. */
  51     volatile sig_atomic_t tick; /* Has the timer ticked?  Set by signal. */
  52     struct timespec cache;      /* Last time read from kernel. */
  53 };
  54
  55 /* Our clocks. */
  56 static struct clock monotonic_clock; /* CLOCK_MONOTONIC, if available. */
  57 static struct clock wall_clock;      /* CLOCK_REALTIME. */
  58
  59 /* The monotonic time at which the time module was initialized. */
  60 static long long int boot_time;
  61
  62 /* Monotonic time in milliseconds at which to die with SIGALRM (if not
  63  * LLONG_MAX). */
  64 static long long int deadline = LLONG_MAX;
  65
  66 /* Monotonic time, in milliseconds, at which the last call to time_poll() woke
  67  * up. */
  68 DEFINE_PER_THREAD_DATA(long long int, last_wakeup, 0);
  69
  70 static void set_up_timer(void);
  71 static void set_up_signal(int flags);
  72 static void sigalrm_handler(int);
  73 static void block_sigalrm(sigset_t *);
  74 static void unblock_sigalrm(const sigset_t *);
  75 static void log_poll_interval(long long int last_wakeup);
  76 static struct rusage *get_recent_rusage(void);
  77 static void refresh_rusage(void);
  78 static void timespec_add(struct timespec *sum,
  79                          const struct timespec *a, const struct timespec *b);
  80
  81 static void
  82 init_clock(struct clock *c, clockid_t id)
  83 {
  84     memset(c, 0, sizeof *c);
  85     c->id = id;
  86     xpthread_rwlock_init(&c->rwlock, NULL);
  87     xclock_gettime(c->id, &c->cache);
  88 }
  89
  90 static void
  91 do_init_time(void)
  92 {
  93     struct timespec ts;
  94
  95     coverage_init();
  96
  97     init_clock(&monotonic_clock, (!clock_gettime(CLOCK_MONOTONIC, &ts)
  98                                   ? CLOCK_MONOTONIC
  99                                   : CLOCK_REALTIME));
 100     init_clock(&wall_clock, CLOCK_REALTIME);
 101     boot_time = timespec_to_msec(&monotonic_clock.cache);
 102
 103     set_up_signal(SA_RESTART);
 104     set_up_timer();
 105 }
 106
 107 /* Initializes the timetracking module, if not already initialized. */
 108 static void
 109 time_init(void)
 110 {
 111     static pthread_once_t once = PTHREAD_ONCE_INIT;
 112     pthread_once(&once, do_init_time);
 113 }
 114
 115 static void
 116 set_up_signal(int flags)
 117 {
 118     struct sigaction sa;
 119
 120     memset(&sa, 0, sizeof sa);
 121     sa.sa_handler = sigalrm_handler;
 122     sigemptyset(&sa.sa_mask);
 123     sa.sa_flags = flags;
 124     xsigaction(SIGALRM, &sa, NULL);
 125 }
 126
 127 static void
 128 set_up_timer(void)
 129 {
 130     static timer_t timer_id;    /* "static" to avoid apparent memory leak. */
 131     struct itimerspec itimer;
 132
 133     if (!CACHE_TIME) {
 134         return;
 135     }
 136
 137     if (timer_create(monotonic_clock.id, NULL, &timer_id)) {
 138         VLOG_FATAL("timer_create failed (%s)", ovs_strerror(errno));
 139     }
 140
 141     itimer.it_interval.tv_sec = 0;
 142     itimer.it_interval.tv_nsec = TIME_UPDATE_INTERVAL * 1000 * 1000;
 143     itimer.it_value = itimer.it_interval;
 144
 145     if (timer_settime(timer_id, 0, &itimer, NULL)) {
 146         VLOG_FATAL("timer_settime failed (%s)", ovs_strerror(errno));
 147     }
 148 }
 149
 150 /* Set up the interval timer, to ensure that time advances even without calling
 151  * time_refresh().
 152  *
 153  * A child created with fork() does not inherit the parent's interval timer, so
 154  * this function needs to be called from the child after fork(). */
 155 void
 156 time_postfork(void)
 157 {
 158     assert_single_threaded();
 159     time_init();
 160     set_up_timer();
 161 }
 162
 163 /* Forces a refresh of the current time from the kernel.  It is not usually
 164  * necessary to call this function, since the time will be refreshed
 165  * automatically at least every TIME_UPDATE_INTERVAL milliseconds.  If
 166  * CACHE_TIME is false, we will always refresh the current time so this
 167  * function has no effect. */
 168 void
 169 time_refresh(void)
 170 {
 171     monotonic_clock.tick = wall_clock.tick = true;
 172 }
 173
 174 static void
 175 time_timespec__(struct clock *c, struct timespec *ts)
 176 {
 177     time_init();
 178     for (;;) {
 179         /* Use the cached time by preference, but fall through if there's been
 180          * a clock tick.  */
 181         xpthread_rwlock_rdlock(&c->rwlock);
 182         if (c->stopped || !c->tick) {
 183             timespec_add(ts, &c->cache, &c->warp);
 184             xpthread_rwlock_unlock(&c->rwlock);
 185             return;
 186         }
 187         xpthread_rwlock_unlock(&c->rwlock);
 188
 189         /* Refresh the cache. */
 190         xpthread_rwlock_wrlock(&c->rwlock);
 191         if (c->tick) {
 192             c->tick = false;
 193             xclock_gettime(c->id, &c->cache);
 194         }
 195         xpthread_rwlock_unlock(&c->rwlock);
 196     }
 197 }
 198
 199 /* Stores a monotonic timer, accurate within TIME_UPDATE_INTERVAL ms, into
 200  * '*ts'. */
 201 void
 202 time_timespec(struct timespec *ts)
 203 {
 204     time_timespec__(&monotonic_clock, ts);
 205 }
 206
 207 /* Stores the current time, accurate within TIME_UPDATE_INTERVAL ms, into
 208  * '*ts'. */
 209 void
 210 time_wall_timespec(struct timespec *ts)
 211 {
 212     time_timespec__(&wall_clock, ts);
 213 }
 214
 215 static time_t
 216 time_sec__(struct clock *c)
 217 {
 218     struct timespec ts;
 219
 220     time_timespec__(c, &ts);
 221     return ts.tv_sec;
 222 }
 223
 224 /* Returns a monotonic timer, in seconds. */
 225 time_t
 226 time_now(void)
 227 {
 228     return time_sec__(&monotonic_clock);
 229 }
 230
 231 /* Returns the current time, in seconds. */
 232 time_t
 233 time_wall(void)
 234 {
 235     return time_sec__(&wall_clock);
 236 }
 237
 238 static long long int
 239 time_msec__(struct clock *c)
 240 {
 241     struct timespec ts;
 242
 243     time_timespec__(c, &ts);
 244     return timespec_to_msec(&ts);
 245 }
 246
 247 /* Returns a monotonic timer, in ms (within TIME_UPDATE_INTERVAL ms). */
 248 long long int
 249 time_msec(void)
 250 {
 251     return time_msec__(&monotonic_clock);
 252 }
 253
 254 /* Returns the current time, in ms (within TIME_UPDATE_INTERVAL ms). */
 255 long long int
 256 time_wall_msec(void)
 257 {
 258     return time_msec__(&wall_clock);
 259 }
 260
 261 /* Configures the program to die with SIGALRM 'secs' seconds from now, if
 262  * 'secs' is nonzero, or disables the feature if 'secs' is zero. */
 263 void
 264 time_alarm(unsigned int secs)
 265 {
 266     long long int now;
 267     long long int msecs;
 268
 269     assert_single_threaded();
 270     time_init();
 271     time_refresh();
 272
 273     now = time_msec();
 274     msecs = secs * 1000LL;
 275     deadline = now < LLONG_MAX - msecs ? now + msecs : LLONG_MAX;
 276 }
 277
 278 /* Like poll(), except:
 279  *
 280  *      - The timeout is specified as an absolute time, as defined by
 281  *        time_msec(), instead of a duration.
 282  *
 283  *      - On error, returns a negative error code (instead of setting errno).
 284  *
 285  *      - If interrupted by a signal, retries automatically until the original
 286  *        timeout is reached.  (Because of this property, this function will
 287  *        never return -EINTR.)
 288  *
 289  *      - As a side effect, refreshes the current time (like time_refresh()).
 290  *
 291  * Stores the number of milliseconds elapsed during poll in '*elapsed'. */
 292 int
 293 time_poll(struct pollfd *pollfds, int n_pollfds, long long int timeout_when,
 294           int *elapsed)
 295 {
 296     long long int *last_wakeup = last_wakeup_get();
 297     long long int start;
 298     sigset_t oldsigs;
 299     bool blocked;
 300     int retval;
 301
 302     time_init();
 303     time_refresh();
 304     if (*last_wakeup) {
 305         log_poll_interval(*last_wakeup);
 306     }
 307     coverage_clear();
 308     start = time_msec();
 309     blocked = false;
 310
 311     timeout_when = MIN(timeout_when, deadline);
 312
 313     for (;;) {
 314         long long int now = time_msec();
 315         int time_left;
 316
 317         if (now >= timeout_when) {
 318             time_left = 0;
 319         } else if ((unsigned long long int) timeout_when - now > INT_MAX) {
 320             time_left = INT_MAX;
 321         } else {
 322             time_left = timeout_when - now;
 323         }
 324
 325         retval = poll(pollfds, n_pollfds, time_left);
 326         if (retval < 0) {
 327             retval = -errno;
 328         }
 329
 330         time_refresh();
 331         if (deadline <= time_msec()) {
 332             fatal_signal_handler(SIGALRM);
 333             if (retval < 0) {
 334                 retval = 0;
 335             }
 336             break;
 337         }
 338
 339         if (retval != -EINTR) {
 340             break;
 341         }
 342
 343         if (!blocked && CACHE_TIME) {
 344             block_sigalrm(&oldsigs);
 345             blocked = true;
 346         }
 347     }
 348     if (blocked) {
 349         unblock_sigalrm(&oldsigs);
 350     }
 351     *last_wakeup = time_msec();
 352     refresh_rusage();
 353     *elapsed = *last_wakeup - start;
 354     return retval;
 355 }
 356
 357 static void
 358 sigalrm_handler(int sig_nr OVS_UNUSED)
 359 {
 360     monotonic_clock.tick = wall_clock.tick = true;
 361 }
 362
 363 static void
 364 block_sigalrm(sigset_t *oldsigs)
 365 {
 366     sigset_t sigalrm;
 367     sigemptyset(&sigalrm);
 368     sigaddset(&sigalrm, SIGALRM);
 369     xpthread_sigmask(SIG_BLOCK, &sigalrm, oldsigs);
 370 }
 371
 372 static void
 373 unblock_sigalrm(const sigset_t *oldsigs)
 374 {
 375     xpthread_sigmask(SIG_SETMASK, oldsigs, NULL);
 376 }
 377
 378 long long int
 379 timespec_to_msec(const struct timespec *ts)
 380 {
 381     return (long long int) ts->tv_sec * 1000 + ts->tv_nsec / (1000 * 1000);
 382 }
 383
 384 long long int
 385 timeval_to_msec(const struct timeval *tv)
 386 {
 387     return (long long int) tv->tv_sec * 1000 + tv->tv_usec / 1000;
 388 }
 389
 390 /* Returns the monotonic time at which the "time" module was initialized, in
 391  * milliseconds. */
 392 long long int
 393 time_boot_msec(void)
 394 {
 395     time_init();
 396     return boot_time;
 397 }
 398
 399 void
 400 xgettimeofday(struct timeval *tv)
 401 {
 402     if (gettimeofday(tv, NULL) == -1) {
 403         VLOG_FATAL("gettimeofday failed (%s)", ovs_strerror(errno));
 404     }
 405 }
 406
 407 void
 408 xclock_gettime(clock_t id, struct timespec *ts)
 409 {
 410     if (clock_gettime(id, ts) == -1) {
 411         /* It seems like a bad idea to try to use vlog here because it is
 412          * likely to try to check the current time. */
 413         ovs_abort(errno, "xclock_gettime() failed");
 414     }
 415 }
 416
 417 static long long int
 418 timeval_diff_msec(const struct timeval *a, const struct timeval *b)
 419 {
 420     return timeval_to_msec(a) - timeval_to_msec(b);
 421 }
 422
 423 static void
 424 timespec_add(struct timespec *sum,
 425              const struct timespec *a,
 426              const struct timespec *b)
 427 {
 428     struct timespec tmp;
 429
 430     tmp.tv_sec = a->tv_sec + b->tv_sec;
 431     tmp.tv_nsec = a->tv_nsec + b->tv_nsec;
 432     if (tmp.tv_nsec >= 1000 * 1000 * 1000) {
 433         tmp.tv_nsec -= 1000 * 1000 * 1000;
 434         tmp.tv_sec++;
 435     }
 436
 437     *sum = tmp;
 438 }
 439
 440 static void
 441 log_poll_interval(long long int last_wakeup)
 442 {
 443     long long int interval = time_msec() - last_wakeup;
 444
 445     if (interval >= 1000
 446         && !monotonic_clock.warp.tv_sec
 447         && !monotonic_clock.warp.tv_nsec) {
 448         const struct rusage *last_rusage = get_recent_rusage();
 449         struct rusage rusage;
 450
 451         getrusage(RUSAGE_SELF, &rusage);
 452         VLOG_WARN("Unreasonably long %lldms poll interval"
 453                   " (%lldms user, %lldms system)",
 454                   interval,
 455                   timeval_diff_msec(&rusage.ru_utime,
 456                                     &last_rusage->ru_utime),
 457                   timeval_diff_msec(&rusage.ru_stime,
 458                                     &last_rusage->ru_stime));
 459         if (rusage.ru_minflt > last_rusage->ru_minflt
 460             || rusage.ru_majflt > last_rusage->ru_majflt) {
 461             VLOG_WARN("faults: %ld minor, %ld major",
 462                       rusage.ru_minflt - last_rusage->ru_minflt,
 463                       rusage.ru_majflt - last_rusage->ru_majflt);
 464         }
 465         if (rusage.ru_inblock > last_rusage->ru_inblock
 466             || rusage.ru_oublock > last_rusage->ru_oublock) {
 467             VLOG_WARN("disk: %ld reads, %ld writes",
 468                       rusage.ru_inblock - last_rusage->ru_inblock,
 469                       rusage.ru_oublock - last_rusage->ru_oublock);
 470         }
 471         if (rusage.ru_nvcsw > last_rusage->ru_nvcsw
 472             || rusage.ru_nivcsw > last_rusage->ru_nivcsw) {
 473             VLOG_WARN("context switches: %ld voluntary, %ld involuntary",
 474                       rusage.ru_nvcsw - last_rusage->ru_nvcsw,
 475                       rusage.ru_nivcsw - last_rusage->ru_nivcsw);
 476         }
 477         coverage_log();
 478     }
 479 }
 480 \f
 481 /* CPU usage tracking. */
 482
 483 struct cpu_usage {
 484     long long int when;         /* Time that this sample was taken. */
 485     unsigned long long int cpu; /* Total user+system CPU usage when sampled. */
 486 };
 487
 488 struct cpu_tracker {
 489     struct cpu_usage older;
 490     struct cpu_usage newer;
 491     int cpu_usage;
 492
 493     struct rusage recent_rusage;
 494 };
 495 DEFINE_PER_THREAD_MALLOCED_DATA(struct cpu_tracker *, cpu_tracker_var);
 496
 497 static struct cpu_tracker *
 498 get_cpu_tracker(void)
 499 {
 500     struct cpu_tracker *t = cpu_tracker_var_get();
 501     if (!t) {
 502         t = xzalloc(sizeof *t);
 503         t->older.when = LLONG_MIN;
 504         t->newer.when = LLONG_MIN;
 505         cpu_tracker_var_set_unsafe(t);
 506     }
 507     return t;
 508 }
 509
 510 static struct rusage *
 511 get_recent_rusage(void)
 512 {
 513     return &get_cpu_tracker()->recent_rusage;
 514 }
 515
 516 static int
 517 getrusage_thread(struct rusage *rusage OVS_UNUSED)
 518 {
 519 #ifdef RUSAGE_THREAD
 520     return getrusage(RUSAGE_THREAD, rusage);
 521 #else
 522     errno = EINVAL;
 523     return -1;
 524 #endif
 525 }
 526
 527 static void
 528 refresh_rusage(void)
 529 {
 530     struct cpu_tracker *t = get_cpu_tracker();
 531     struct rusage *recent_rusage = &t->recent_rusage;
 532
 533     if (!getrusage_thread(recent_rusage)) {
 534         long long int now = time_msec();
 535         if (now >= t->newer.when + 3 * 1000) {
 536             t->older = t->newer;
 537             t->newer.when = now;
 538             t->newer.cpu = (timeval_to_msec(&recent_rusage->ru_utime) +
 539                             timeval_to_msec(&recent_rusage->ru_stime));
 540
 541             if (t->older.when != LLONG_MIN && t->newer.cpu > t->older.cpu) {
 542                 unsigned int dividend = t->newer.cpu - t->older.cpu;
 543                 unsigned int divisor = (t->newer.when - t->older.when) / 100;
 544                 t->cpu_usage = divisor > 0 ? dividend / divisor : -1;
 545             } else {
 546                 t->cpu_usage = -1;
 547             }
 548         }
 549     }
 550 }
 551
 552 /* Returns an estimate of this process's CPU usage, as a percentage, over the
 553  * past few seconds of wall-clock time.  Returns -1 if no estimate is available
 554  * (which will happen if the process has not been running long enough to have
 555  * an estimate, and can happen for other reasons as well). */
 556 int
 557 get_cpu_usage(void)
 558 {
 559     return get_cpu_tracker()->cpu_usage;
 560 }
 561 \f
 562 /* Unixctl interface. */
 563
 564 /* "time/stop" stops the monotonic time returned by e.g. time_msec() from
 565  * advancing, except due to later calls to "time/warp". */
 566 static void
 567 timeval_stop_cb(struct unixctl_conn *conn,
 568                  int argc OVS_UNUSED, const char *argv[] OVS_UNUSED,
 569                  void *aux OVS_UNUSED)
 570 {
 571     xpthread_rwlock_wrlock(&monotonic_clock.rwlock);
 572     monotonic_clock.stopped = true;
 573     xpthread_rwlock_unlock(&monotonic_clock.rwlock);
 574
 575     unixctl_command_reply(conn, NULL);
 576 }
 577
 578 /* "time/warp MSECS" advances the current monotonic time by the specified
 579  * number of milliseconds.  Unless "time/stop" has also been executed, the
 580  * monotonic clock continues to tick forward at the normal rate afterward.
 581  *
 582  * Does not affect wall clock readings. */
 583 static void
 584 timeval_warp_cb(struct unixctl_conn *conn,
 585                 int argc OVS_UNUSED, const char *argv[], void *aux OVS_UNUSED)
 586 {
 587     struct timespec ts;
 588     int msecs;
 589
 590     msecs = atoi(argv[1]);
 591     if (msecs <= 0) {
 592         unixctl_command_reply_error(conn, "invalid MSECS");
 593         return;
 594     }
 595
 596     ts.tv_sec = msecs / 1000;
 597     ts.tv_nsec = (msecs % 1000) * 1000 * 1000;
 598
 599     xpthread_rwlock_wrlock(&monotonic_clock.rwlock);
 600     timespec_add(&monotonic_clock.warp, &monotonic_clock.warp, &ts);
 601     xpthread_rwlock_unlock(&monotonic_clock.rwlock);
 602
 603     unixctl_command_reply(conn, "warped");
 604 }
 605
 606 void
 607 timeval_dummy_register(void)
 608 {
 609     unixctl_command_register("time/stop", "", 0, 0, timeval_stop_cb, NULL);
 610     unixctl_command_register("time/warp", "MSECS", 1, 1,
 611                              timeval_warp_cb, NULL);
 612 }