lib/timeval.c

   1 /*
   2  * Copyright (c) 2008, 2009, 2010, 2011, 2012 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 <assert.h>
  20 #include <errno.h>
  21 #if HAVE_EXECINFO_H
  22 #include <execinfo.h>
  23 #endif
  24 #include <poll.h>
  25 #include <signal.h>
  26 #include <stdlib.h>
  27 #include <string.h>
  28 #include <sys/time.h>
  29 #include <sys/resource.h>
  30 #include <unistd.h>
  31 #include "coverage.h"
  32 #include "dummy.h"
  33 #include "dynamic-string.h"
  34 #include "fatal-signal.h"
  35 #include "signals.h"
  36 #include "unixctl.h"
  37 #include "util.h"
  38 #include "vlog.h"
  39
  40 #ifndef HAVE_EXECINFO_H
  41 #define HAVE_EXECINFO_H 0
  42 #endif
  43
  44 VLOG_DEFINE_THIS_MODULE(timeval);
  45
  46 /* The clock to use for measuring time intervals.  This is CLOCK_MONOTONIC by
  47  * preference, but on systems that don't have a monotonic clock we fall back
  48  * to CLOCK_REALTIME. */
  49 static clockid_t monotonic_clock;
  50
  51 /* Has a timer tick occurred? Only relevant if CACHE_TIME is true.
  52  *
  53  * We initialize these to true to force time_init() to get called on the first
  54  * call to time_msec() or another function that queries the current time. */
  55 static volatile sig_atomic_t wall_tick = true;
  56 static volatile sig_atomic_t monotonic_tick = true;
  57
  58 /* The current time, as of the last refresh. */
  59 static struct timespec wall_time;
  60 static struct timespec monotonic_time;
  61
  62 /* The monotonic time at which the time module was initialized. */
  63 static long long int boot_time;
  64
  65 /* features for use by unit tests. */
  66 static struct timespec warp_offset; /* Offset added to monotonic_time. */
  67 static bool time_stopped;           /* Disables real-time updates, if true. */
  68
  69 /* Time in milliseconds at which to die with SIGALRM (if not LLONG_MAX). */
  70 static long long int deadline = LLONG_MAX;
  71
  72 struct trace {
  73     void *backtrace[32]; /* Populated by backtrace(). */
  74     size_t n_frames;     /* Number of frames in 'backtrace'. */
  75 };
  76
  77 #define MAX_TRACES 50
  78 static struct trace traces[MAX_TRACES];
  79 static size_t trace_head = 0;
  80
  81 static void set_up_timer(void);
  82 static void set_up_signal(int flags);
  83 static void sigalrm_handler(int);
  84 static void refresh_wall_if_ticked(void);
  85 static void refresh_monotonic_if_ticked(void);
  86 static void block_sigalrm(sigset_t *);
  87 static void unblock_sigalrm(const sigset_t *);
  88 static void log_poll_interval(long long int last_wakeup);
  89 static struct rusage *get_recent_rusage(void);
  90 static void refresh_rusage(void);
  91 static void timespec_add(struct timespec *sum,
  92                          const struct timespec *a, const struct timespec *b);
  93 static unixctl_cb_func backtrace_cb;
  94
  95 /* Initializes the timetracking module, if not already initialized. */
  96 static void
  97 time_init(void)
  98 {
  99     static bool inited;
 100
 101     if (inited) {
 102         return;
 103     }
 104     inited = true;
 105
 106     memset(traces, 0, sizeof traces);
 107
 108     if (HAVE_EXECINFO_H && CACHE_TIME) {
 109         unixctl_command_register("backtrace", "", 0, 0, backtrace_cb, NULL);
 110     }
 111
 112     coverage_init();
 113
 114     if (!clock_gettime(CLOCK_MONOTONIC, &monotonic_time)) {
 115         monotonic_clock = CLOCK_MONOTONIC;
 116     } else {
 117         monotonic_clock = CLOCK_REALTIME;
 118         VLOG_DBG("monotonic timer not available");
 119     }
 120
 121     set_up_signal(SA_RESTART);
 122     set_up_timer();
 123
 124     boot_time = time_msec();
 125 }
 126
 127 static void
 128 set_up_signal(int flags)
 129 {
 130     struct sigaction sa;
 131
 132     memset(&sa, 0, sizeof sa);
 133     sa.sa_handler = sigalrm_handler;
 134     sigemptyset(&sa.sa_mask);
 135     sa.sa_flags = flags;
 136     xsigaction(SIGALRM, &sa, NULL);
 137 }
 138
 139 /* Remove SA_RESTART from the flags for SIGALRM, so that any system call that
 140  * is interrupted by the periodic timer interrupt will return EINTR instead of
 141  * continuing after the signal handler returns.
 142  *
 143  * time_disable_restart() and time_enable_restart() may be usefully wrapped
 144  * around function calls that might otherwise block forever unless interrupted
 145  * by a signal, e.g.:
 146  *
 147  *   time_disable_restart();
 148  *   fcntl(fd, F_SETLKW, &lock);
 149  *   time_enable_restart();
 150  */
 151 void
 152 time_disable_restart(void)
 153 {
 154     time_init();
 155     set_up_signal(0);
 156 }
 157
 158 /* Add SA_RESTART to the flags for SIGALRM, so that any system call that
 159  * is interrupted by the periodic timer interrupt will continue after the
 160  * signal handler returns instead of returning EINTR. */
 161 void
 162 time_enable_restart(void)
 163 {
 164     time_init();
 165     set_up_signal(SA_RESTART);
 166 }
 167
 168 static void
 169 set_up_timer(void)
 170 {
 171     static timer_t timer_id;    /* "static" to avoid apparent memory leak. */
 172     struct itimerspec itimer;
 173
 174     if (!CACHE_TIME) {
 175         return;
 176     }
 177
 178     if (timer_create(monotonic_clock, NULL, &timer_id)) {
 179         VLOG_FATAL("timer_create failed (%s)", strerror(errno));
 180     }
 181
 182     itimer.it_interval.tv_sec = 0;
 183     itimer.it_interval.tv_nsec = TIME_UPDATE_INTERVAL * 1000 * 1000;
 184     itimer.it_value = itimer.it_interval;
 185
 186     if (timer_settime(timer_id, 0, &itimer, NULL)) {
 187         VLOG_FATAL("timer_settime failed (%s)", strerror(errno));
 188     }
 189 }
 190
 191 /* Set up the interval timer, to ensure that time advances even without calling
 192  * time_refresh().
 193  *
 194  * A child created with fork() does not inherit the parent's interval timer, so
 195  * this function needs to be called from the child after fork(). */
 196 void
 197 time_postfork(void)
 198 {
 199     time_init();
 200     set_up_timer();
 201 }
 202
 203 static void
 204 refresh_wall(void)
 205 {
 206     time_init();
 207     clock_gettime(CLOCK_REALTIME, &wall_time);
 208     wall_tick = false;
 209 }
 210
 211 static void
 212 refresh_monotonic(void)
 213 {
 214     time_init();
 215
 216     if (!time_stopped) {
 217         if (monotonic_clock == CLOCK_MONOTONIC) {
 218             clock_gettime(monotonic_clock, &monotonic_time);
 219         } else {
 220             refresh_wall_if_ticked();
 221             monotonic_time = wall_time;
 222         }
 223         timespec_add(&monotonic_time, &monotonic_time, &warp_offset);
 224
 225         monotonic_tick = false;
 226     }
 227 }
 228
 229 /* Forces a refresh of the current time from the kernel.  It is not usually
 230  * necessary to call this function, since the time will be refreshed
 231  * automatically at least every TIME_UPDATE_INTERVAL milliseconds.  If
 232  * CACHE_TIME is false, we will always refresh the current time so this
 233  * function has no effect. */
 234 void
 235 time_refresh(void)
 236 {
 237     wall_tick = monotonic_tick = true;
 238 }
 239
 240 /* Returns a monotonic timer, in seconds. */
 241 time_t
 242 time_now(void)
 243 {
 244     refresh_monotonic_if_ticked();
 245     return monotonic_time.tv_sec;
 246 }
 247
 248 /* Returns the current time, in seconds. */
 249 time_t
 250 time_wall(void)
 251 {
 252     refresh_wall_if_ticked();
 253     return wall_time.tv_sec;
 254 }
 255
 256 /* Returns a monotonic timer, in ms (within TIME_UPDATE_INTERVAL ms). */
 257 long long int
 258 time_msec(void)
 259 {
 260     refresh_monotonic_if_ticked();
 261     return timespec_to_msec(&monotonic_time);
 262 }
 263
 264 /* Returns the current time, in ms (within TIME_UPDATE_INTERVAL ms). */
 265 long long int
 266 time_wall_msec(void)
 267 {
 268     refresh_wall_if_ticked();
 269     return timespec_to_msec(&wall_time);
 270 }
 271
 272 /* Stores a monotonic timer, accurate within TIME_UPDATE_INTERVAL ms, into
 273  * '*ts'. */
 274 void
 275 time_timespec(struct timespec *ts)
 276 {
 277     refresh_monotonic_if_ticked();
 278     *ts = monotonic_time;
 279 }
 280
 281 /* Stores the current time, accurate within TIME_UPDATE_INTERVAL ms, into
 282  * '*ts'. */
 283 void
 284 time_wall_timespec(struct timespec *ts)
 285 {
 286     refresh_wall_if_ticked();
 287     *ts = wall_time;
 288 }
 289
 290 /* Configures the program to die with SIGALRM 'secs' seconds from now, if
 291  * 'secs' is nonzero, or disables the feature if 'secs' is zero. */
 292 void
 293 time_alarm(unsigned int secs)
 294 {
 295     long long int now;
 296     long long int msecs;
 297
 298     sigset_t oldsigs;
 299
 300     time_init();
 301     time_refresh();
 302
 303     now = time_msec();
 304     msecs = secs * 1000;
 305
 306     block_sigalrm(&oldsigs);
 307     deadline = now < LLONG_MAX - msecs ? now + msecs : LLONG_MAX;
 308     unblock_sigalrm(&oldsigs);
 309 }
 310
 311 /* Like poll(), except:
 312  *
 313  *      - The timeout is specified as an absolute time, as defined by
 314  *        time_msec(), instead of a duration.
 315  *
 316  *      - On error, returns a negative error code (instead of setting errno).
 317  *
 318  *      - If interrupted by a signal, retries automatically until the original
 319  *        timeout is reached.  (Because of this property, this function will
 320  *        never return -EINTR.)
 321  *
 322  *      - As a side effect, refreshes the current time (like time_refresh()).
 323  *
 324  * Stores the number of milliseconds elapsed during poll in '*elapsed'. */
 325 int
 326 time_poll(struct pollfd *pollfds, int n_pollfds, long long int timeout_when,
 327           int *elapsed)
 328 {
 329     static long long int last_wakeup = 0;
 330     long long int start;
 331     sigset_t oldsigs;
 332     bool blocked;
 333     int retval;
 334
 335     time_refresh();
 336     if (last_wakeup) {
 337         log_poll_interval(last_wakeup);
 338     }
 339     coverage_clear();
 340     start = time_msec();
 341     blocked = false;
 342
 343     timeout_when = MIN(timeout_when, deadline);
 344
 345     for (;;) {
 346         long long int now = time_msec();
 347         int time_left;
 348
 349         if (now >= timeout_when) {
 350             time_left = 0;
 351         } else if ((unsigned long long int) timeout_when - now > INT_MAX) {
 352             time_left = INT_MAX;
 353         } else {
 354             time_left = timeout_when - now;
 355         }
 356
 357         retval = poll(pollfds, n_pollfds, time_left);
 358         if (retval < 0) {
 359             retval = -errno;
 360         }
 361
 362         time_refresh();
 363         if (deadline <= time_msec()) {
 364             fatal_signal_handler(SIGALRM);
 365             if (retval < 0) {
 366                 retval = 0;
 367             }
 368             break;
 369         }
 370
 371         if (retval != -EINTR) {
 372             break;
 373         }
 374
 375         if (!blocked && CACHE_TIME) {
 376             block_sigalrm(&oldsigs);
 377             blocked = true;
 378         }
 379     }
 380     if (blocked) {
 381         unblock_sigalrm(&oldsigs);
 382     }
 383     last_wakeup = time_msec();
 384     refresh_rusage();
 385     *elapsed = last_wakeup - start;
 386     return retval;
 387 }
 388
 389 static void
 390 sigalrm_handler(int sig_nr OVS_UNUSED)
 391 {
 392     wall_tick = true;
 393     monotonic_tick = true;
 394
 395 #if HAVE_EXECINFO_H
 396     if (CACHE_TIME) {
 397         struct trace *trace = &traces[trace_head];
 398
 399         trace->n_frames = backtrace(trace->backtrace,
 400                                     ARRAY_SIZE(trace->backtrace));
 401         trace_head = (trace_head + 1) % MAX_TRACES;
 402     }
 403 #endif
 404 }
 405
 406 static void
 407 refresh_wall_if_ticked(void)
 408 {
 409     if (!CACHE_TIME || wall_tick) {
 410         refresh_wall();
 411     }
 412 }
 413
 414 static void
 415 refresh_monotonic_if_ticked(void)
 416 {
 417     if (!CACHE_TIME || monotonic_tick) {
 418         refresh_monotonic();
 419     }
 420 }
 421
 422 static void
 423 block_sigalrm(sigset_t *oldsigs)
 424 {
 425     sigset_t sigalrm;
 426     sigemptyset(&sigalrm);
 427     sigaddset(&sigalrm, SIGALRM);
 428     xsigprocmask(SIG_BLOCK, &sigalrm, oldsigs);
 429 }
 430
 431 static void
 432 unblock_sigalrm(const sigset_t *oldsigs)
 433 {
 434     xsigprocmask(SIG_SETMASK, oldsigs, NULL);
 435 }
 436
 437 long long int
 438 timespec_to_msec(const struct timespec *ts)
 439 {
 440     return (long long int) ts->tv_sec * 1000 + ts->tv_nsec / (1000 * 1000);
 441 }
 442
 443 long long int
 444 timeval_to_msec(const struct timeval *tv)
 445 {
 446     return (long long int) tv->tv_sec * 1000 + tv->tv_usec / 1000;
 447 }
 448
 449 /* Returns the monotonic time at which the "time" module was initialized, in
 450  * milliseconds(). */
 451 long long int
 452 time_boot_msec(void)
 453 {
 454     time_init();
 455     return boot_time;
 456 }
 457
 458 void
 459 xgettimeofday(struct timeval *tv)
 460 {
 461     if (gettimeofday(tv, NULL) == -1) {
 462         VLOG_FATAL("gettimeofday failed (%s)", strerror(errno));
 463     }
 464 }
 465
 466 static long long int
 467 timeval_diff_msec(const struct timeval *a, const struct timeval *b)
 468 {
 469     return timeval_to_msec(a) - timeval_to_msec(b);
 470 }
 471
 472 static void
 473 timespec_add(struct timespec *sum,
 474              const struct timespec *a,
 475              const struct timespec *b)
 476 {
 477     struct timespec tmp;
 478
 479     tmp.tv_sec = a->tv_sec + b->tv_sec;
 480     tmp.tv_nsec = a->tv_nsec + b->tv_nsec;
 481     if (tmp.tv_nsec >= 1000 * 1000 * 1000) {
 482         tmp.tv_nsec -= 1000 * 1000 * 1000;
 483         tmp.tv_sec++;
 484     }
 485
 486     *sum = tmp;
 487 }
 488
 489 static void
 490 log_poll_interval(long long int last_wakeup)
 491 {
 492     long long int interval = time_msec() - last_wakeup;
 493
 494     if (interval >= 1000) {
 495         const struct rusage *last_rusage = get_recent_rusage();
 496         struct rusage rusage;
 497
 498         getrusage(RUSAGE_SELF, &rusage);
 499         VLOG_WARN("Unreasonably long %lldms poll interval"
 500                   " (%lldms user, %lldms system)",
 501                   interval,
 502                   timeval_diff_msec(&rusage.ru_utime,
 503                                     &last_rusage->ru_utime),
 504                   timeval_diff_msec(&rusage.ru_stime,
 505                                     &last_rusage->ru_stime));
 506         if (rusage.ru_minflt > last_rusage->ru_minflt
 507             || rusage.ru_majflt > last_rusage->ru_majflt) {
 508             VLOG_WARN("faults: %ld minor, %ld major",
 509                       rusage.ru_minflt - last_rusage->ru_minflt,
 510                       rusage.ru_majflt - last_rusage->ru_majflt);
 511         }
 512         if (rusage.ru_inblock > last_rusage->ru_inblock
 513             || rusage.ru_oublock > last_rusage->ru_oublock) {
 514             VLOG_WARN("disk: %ld reads, %ld writes",
 515                       rusage.ru_inblock - last_rusage->ru_inblock,
 516                       rusage.ru_oublock - last_rusage->ru_oublock);
 517         }
 518         if (rusage.ru_nvcsw > last_rusage->ru_nvcsw
 519             || rusage.ru_nivcsw > last_rusage->ru_nivcsw) {
 520             VLOG_WARN("context switches: %ld voluntary, %ld involuntary",
 521                       rusage.ru_nvcsw - last_rusage->ru_nvcsw,
 522                       rusage.ru_nivcsw - last_rusage->ru_nivcsw);
 523         }
 524         coverage_log();
 525     }
 526 }
 527 \f
 528 /* CPU usage tracking. */
 529
 530 struct cpu_usage {
 531     long long int when;         /* Time that this sample was taken. */
 532     unsigned long long int cpu; /* Total user+system CPU usage when sampled. */
 533 };
 534
 535 static struct rusage recent_rusage;
 536 static struct cpu_usage older = { LLONG_MIN, 0 };
 537 static struct cpu_usage newer = { LLONG_MIN, 0 };
 538 static int cpu_usage = -1;
 539
 540 static struct rusage *
 541 get_recent_rusage(void)
 542 {
 543     return &recent_rusage;
 544 }
 545
 546 static void
 547 refresh_rusage(void)
 548 {
 549     long long int now;
 550
 551     now = time_msec();
 552     getrusage(RUSAGE_SELF, &recent_rusage);
 553
 554     if (now >= newer.when + 3 * 1000) {
 555         older = newer;
 556         newer.when = now;
 557         newer.cpu = (timeval_to_msec(&recent_rusage.ru_utime) +
 558                      timeval_to_msec(&recent_rusage.ru_stime));
 559
 560         if (older.when != LLONG_MIN && newer.cpu > older.cpu) {
 561             unsigned int dividend = newer.cpu - older.cpu;
 562             unsigned int divisor = (newer.when - older.when) / 100;
 563             cpu_usage = divisor > 0 ? dividend / divisor : -1;
 564         } else {
 565             cpu_usage = -1;
 566         }
 567     }
 568 }
 569
 570 /* Returns an estimate of this process's CPU usage, as a percentage, over the
 571  * past few seconds of wall-clock time.  Returns -1 if no estimate is available
 572  * (which will happen if the process has not been running long enough to have
 573  * an estimate, and can happen for other reasons as well). */
 574 int
 575 get_cpu_usage(void)
 576 {
 577     return cpu_usage;
 578 }
 579
 580 static void
 581 format_backtraces(struct ds *ds)
 582 {
 583     time_init();
 584
 585 #if HAVE_EXECINFO_H
 586     if (CACHE_TIME) {
 587         sigset_t oldsigs;
 588         size_t i;
 589
 590         block_sigalrm(&oldsigs);
 591
 592         for (i = 0; i < MAX_TRACES; i++) {
 593             struct trace *trace = &traces[i];
 594             char **frame_strs;
 595             size_t j;
 596
 597             if (!trace->n_frames) {
 598                 continue;
 599             }
 600
 601             frame_strs = backtrace_symbols(trace->backtrace, trace->n_frames);
 602
 603             ds_put_format(ds, "Backtrace %zu\n", i + 1);
 604             for (j = 0; j < trace->n_frames; j++) {
 605                 ds_put_format(ds, "%s\n", frame_strs[j]);
 606             }
 607             ds_put_cstr(ds, "\n");
 608
 609             free(frame_strs);
 610         }
 611         ds_chomp(ds, '\n');
 612         unblock_sigalrm(&oldsigs);
 613     }
 614 #endif
 615 }
 616 \f
 617 /* Unixctl interface. */
 618
 619 /* "time/stop" stops the monotonic time returned by e.g. time_msec() from
 620  * advancing, except due to later calls to "time/warp". */
 621 static void
 622 timeval_stop_cb(struct unixctl_conn *conn,
 623                  int argc OVS_UNUSED, const char *argv[] OVS_UNUSED,
 624                  void *aux OVS_UNUSED)
 625 {
 626     time_stopped = true;
 627     unixctl_command_reply(conn, NULL);
 628 }
 629
 630 /* "time/warp MSECS" advances the current monotonic time by the specified
 631  * number of milliseconds.  Unless "time/stop" has also been executed, the
 632  * monotonic clock continues to tick forward at the normal rate afterward.
 633  *
 634  * Does not affect wall clock readings. */
 635 static void
 636 timeval_warp_cb(struct unixctl_conn *conn,
 637                 int argc OVS_UNUSED, const char *argv[], void *aux OVS_UNUSED)
 638 {
 639     struct timespec ts;
 640     int msecs;
 641
 642     msecs = atoi(argv[1]);
 643     if (msecs <= 0) {
 644         unixctl_command_reply_error(conn, "invalid MSECS");
 645         return;
 646     }
 647
 648     ts.tv_sec = msecs / 1000;
 649     ts.tv_nsec = (msecs % 1000) * 1000 * 1000;
 650     timespec_add(&warp_offset, &warp_offset, &ts);
 651     timespec_add(&monotonic_time, &monotonic_time, &ts);
 652     unixctl_command_reply(conn, "warped");
 653 }
 654
 655 static void
 656 backtrace_cb(struct unixctl_conn *conn,
 657              int argc OVS_UNUSED, const char *argv[] OVS_UNUSED,
 658              void *aux OVS_UNUSED)
 659 {
 660     struct ds ds = DS_EMPTY_INITIALIZER;
 661
 662     assert(HAVE_EXECINFO_H && CACHE_TIME);
 663     format_backtraces(&ds);
 664     unixctl_command_reply(conn, ds_cstr(&ds));
 665     ds_destroy(&ds);
 666 }
 667
 668 void
 669 timeval_dummy_register(void)
 670 {
 671     unixctl_command_register("time/stop", "", 0, 0, timeval_stop_cb, NULL);
 672     unixctl_command_register("time/warp", "MSECS", 1, 1,
 673                              timeval_warp_cb, NULL);
 674 }