eta.c

   1 /*
   2  * Status and ETA code
   3  */
   4 #include <unistd.h>
   5 #include <fcntl.h>
   6 #include <string.h>
   7
   8 #include "fio.h"
   9 #include "lib/pow2.h"
  10
  11 static char __run_str[REAL_MAX_JOBS + 1];
  12 static char run_str[__THREAD_RUNSTR_SZ(REAL_MAX_JOBS)];
  13
  14 static void update_condensed_str(char *rstr, char *run_str_condensed)
  15 {
  16         if (*rstr) {
  17                 while (*rstr) {
  18                         int nr = 1;
  19
  20                         *run_str_condensed++ = *rstr++;
  21                         while (*(rstr - 1) == *rstr) {
  22                                 rstr++;
  23                                 nr++;
  24                         }
  25                         run_str_condensed += sprintf(run_str_condensed, "(%u),", nr);
  26                 }
  27                 run_str_condensed--;
  28         }
  29         *run_str_condensed = '\0';
  30 }
  31
  32 /*
  33  * Sets the status of the 'td' in the printed status map.
  34  */
  35 static void check_str_update(struct thread_data *td)
  36 {
  37         char c = __run_str[td->thread_number - 1];
  38
  39         switch (td->runstate) {
  40         case TD_REAPED:
  41                 if (td->error)
  42                         c = 'X';
  43                 else if (td->sig)
  44                         c = 'K';
  45                 else
  46                         c = '_';
  47                 break;
  48         case TD_EXITED:
  49                 c = 'E';
  50                 break;
  51         case TD_RAMP:
  52                 c = '/';
  53                 break;
  54         case TD_RUNNING:
  55                 if (td_rw(td)) {
  56                         if (td_random(td)) {
  57                                 if (td->o.rwmix[DDIR_READ] == 100)
  58                                         c = 'r';
  59                                 else if (td->o.rwmix[DDIR_WRITE] == 100)
  60                                         c = 'w';
  61                                 else
  62                                         c = 'm';
  63                         } else {
  64                                 if (td->o.rwmix[DDIR_READ] == 100)
  65                                         c = 'R';
  66                                 else if (td->o.rwmix[DDIR_WRITE] == 100)
  67                                         c = 'W';
  68                                 else
  69                                         c = 'M';
  70                         }
  71                 } else if (td_read(td)) {
  72                         if (td_random(td))
  73                                 c = 'r';
  74                         else
  75                                 c = 'R';
  76                 } else if (td_write(td)) {
  77                         if (td_random(td))
  78                                 c = 'w';
  79                         else
  80                                 c = 'W';
  81                 } else {
  82                         if (td_random(td))
  83                                 c = 'd';
  84                         else
  85                                 c = 'D';
  86                 }
  87                 break;
  88         case TD_PRE_READING:
  89                 c = 'p';
  90                 break;
  91         case TD_VERIFYING:
  92                 c = 'V';
  93                 break;
  94         case TD_FSYNCING:
  95                 c = 'F';
  96                 break;
  97         case TD_FINISHING:
  98                 c = 'f';
  99                 break;
 100         case TD_CREATED:
 101                 c = 'C';
 102                 break;
 103         case TD_INITIALIZED:
 104         case TD_SETTING_UP:
 105                 c = 'I';
 106                 break;
 107         case TD_NOT_CREATED:
 108                 c = 'P';
 109                 break;
 110         default:
 111                 log_err("state %d\n", td->runstate);
 112         }
 113
 114         __run_str[td->thread_number - 1] = c;
 115         update_condensed_str(__run_str, run_str);
 116 }
 117
 118 /*
 119  * Convert seconds to a printable string.
 120  */
 121 void eta_to_str(char *str, unsigned long eta_sec)
 122 {
 123         unsigned int d, h, m, s;
 124         int disp_hour = 0;
 125
 126         if (eta_sec == -1) {
 127                 sprintf(str, "--");
 128                 return;
 129         }
 130
 131         s = eta_sec % 60;
 132         eta_sec /= 60;
 133         m = eta_sec % 60;
 134         eta_sec /= 60;
 135         h = eta_sec % 24;
 136         eta_sec /= 24;
 137         d = eta_sec;
 138
 139         if (d) {
 140                 disp_hour = 1;
 141                 str += sprintf(str, "%02ud:", d);
 142         }
 143
 144         if (h || disp_hour)
 145                 str += sprintf(str, "%02uh:", h);
 146
 147         str += sprintf(str, "%02um:", m);
 148         str += sprintf(str, "%02us", s);
 149 }
 150
 151 /*
 152  * Best effort calculation of the estimated pending runtime of a job.
 153  */
 154 static unsigned long thread_eta(struct thread_data *td)
 155 {
 156         unsigned long long bytes_total, bytes_done;
 157         unsigned long eta_sec = 0;
 158         unsigned long elapsed;
 159         uint64_t timeout;
 160
 161         elapsed = (mtime_since_now(&td->epoch) + 999) / 1000;
 162         timeout = td->o.timeout / 1000000UL;
 163
 164         bytes_total = td->total_io_size;
 165
 166         if (td->flags & TD_F_NO_PROGRESS)
 167                 return -1;
 168
 169         if (td->o.fill_device && td->o.size  == -1ULL) {
 170                 if (!td->fill_device_size || td->fill_device_size == -1ULL)
 171                         return 0;
 172
 173                 bytes_total = td->fill_device_size;
 174         }
 175
 176         if (td->o.zone_size && td->o.zone_skip && bytes_total) {
 177                 unsigned int nr_zones;
 178                 uint64_t zone_bytes;
 179
 180                 zone_bytes = bytes_total + td->o.zone_size + td->o.zone_skip;
 181                 nr_zones = (zone_bytes - 1) / (td->o.zone_size + td->o.zone_skip);
 182                 bytes_total -= nr_zones * td->o.zone_skip;
 183         }
 184
 185         /*
 186          * if writing and verifying afterwards, bytes_total will be twice the
 187          * size. In a mixed workload, verify phase will be the size of the
 188          * first stage writes.
 189          */
 190         if (td->o.do_verify && td->o.verify && td_write(td)) {
 191                 if (td_rw(td)) {
 192                         unsigned int perc = 50;
 193
 194                         if (td->o.rwmix[DDIR_WRITE])
 195                                 perc = td->o.rwmix[DDIR_WRITE];
 196
 197                         bytes_total += (bytes_total * perc) / 100;
 198                 } else
 199                         bytes_total <<= 1;
 200         }
 201
 202         if (td->runstate == TD_RUNNING || td->runstate == TD_VERIFYING) {
 203                 double perc, perc_t;
 204
 205                 bytes_done = ddir_rw_sum(td->io_bytes);
 206
 207                 if (bytes_total) {
 208                         perc = (double) bytes_done / (double) bytes_total;
 209                         if (perc > 1.0)
 210                                 perc = 1.0;
 211                 } else
 212                         perc = 0.0;
 213
 214                 if (td->o.time_based) {
 215                         if (timeout) {
 216                                 perc_t = (double) elapsed / (double) timeout;
 217                                 if (perc_t < perc)
 218                                         perc = perc_t;
 219                         } else {
 220                                 /*
 221                                  * Will never hit, we can't have time_based
 222                                  * without a timeout set.
 223                                  */
 224                                 perc = 0.0;
 225                         }
 226                 }
 227
 228                 eta_sec = (unsigned long) (elapsed * (1.0 / perc)) - elapsed;
 229
 230                 if (td->o.timeout &&
 231                     eta_sec > (timeout + done_secs - elapsed))
 232                         eta_sec = timeout + done_secs - elapsed;
 233         } else if (td->runstate == TD_NOT_CREATED || td->runstate == TD_CREATED
 234                         || td->runstate == TD_INITIALIZED
 235                         || td->runstate == TD_SETTING_UP
 236                         || td->runstate == TD_RAMP
 237                         || td->runstate == TD_PRE_READING) {
 238                 int64_t t_eta = 0, r_eta = 0;
 239                 unsigned long long rate_bytes;
 240
 241                 /*
 242                  * We can only guess - assume it'll run the full timeout
 243                  * if given, otherwise assume it'll run at the specified rate.
 244                  */
 245                 if (td->o.timeout) {
 246                         uint64_t __timeout = td->o.timeout;
 247                         uint64_t start_delay = td->o.start_delay;
 248                         uint64_t ramp_time = td->o.ramp_time;
 249
 250                         t_eta = __timeout + start_delay + ramp_time;
 251                         t_eta /= 1000000ULL;
 252
 253                         if ((td->runstate == TD_RAMP) && in_ramp_time(td)) {
 254                                 unsigned long ramp_left;
 255
 256                                 ramp_left = mtime_since_now(&td->epoch);
 257                                 ramp_left = (ramp_left + 999) / 1000;
 258                                 if (ramp_left <= t_eta)
 259                                         t_eta -= ramp_left;
 260                         }
 261                 }
 262                 rate_bytes = 0;
 263                 if (td_read(td))
 264                         rate_bytes  = td->o.rate[DDIR_READ];
 265                 if (td_write(td))
 266                         rate_bytes += td->o.rate[DDIR_WRITE];
 267                 if (td_trim(td))
 268                         rate_bytes += td->o.rate[DDIR_TRIM];
 269
 270                 if (rate_bytes) {
 271                         r_eta = bytes_total / rate_bytes;
 272                         r_eta += (td->o.start_delay / 1000000ULL);
 273                 }
 274
 275                 if (r_eta && t_eta)
 276                         eta_sec = min(r_eta, t_eta);
 277                 else if (r_eta)
 278                         eta_sec = r_eta;
 279                 else if (t_eta)
 280                         eta_sec = t_eta;
 281                 else
 282                         eta_sec = 0;
 283         } else {
 284                 /*
 285                  * thread is already done or waiting for fsync
 286                  */
 287                 eta_sec = 0;
 288         }
 289
 290         return eta_sec;
 291 }
 292
 293 static void calc_rate(int unified_rw_rep, unsigned long mtime,
 294                       unsigned long long *io_bytes,
 295                       unsigned long long *prev_io_bytes, uint64_t *rate)
 296 {
 297         int i;
 298
 299         for (i = 0; i < DDIR_RWDIR_CNT; i++) {
 300                 unsigned long long diff, this_rate;
 301
 302                 diff = io_bytes[i] - prev_io_bytes[i];
 303                 if (mtime)
 304                         this_rate = ((1000 * diff) / mtime) / 1024;
 305                 else
 306                         this_rate = 0;
 307
 308                 if (unified_rw_rep) {
 309                         rate[i] = 0;
 310                         rate[0] += this_rate;
 311                 } else
 312                         rate[i] = this_rate;
 313
 314                 prev_io_bytes[i] = io_bytes[i];
 315         }
 316 }
 317
 318 static void calc_iops(int unified_rw_rep, unsigned long mtime,
 319                       unsigned long long *io_iops,
 320                       unsigned long long *prev_io_iops, unsigned int *iops)
 321 {
 322         int i;
 323
 324         for (i = 0; i < DDIR_RWDIR_CNT; i++) {
 325                 unsigned long long diff, this_iops;
 326
 327                 diff = io_iops[i] - prev_io_iops[i];
 328                 if (mtime)
 329                         this_iops = (diff * 1000) / mtime;
 330                 else
 331                         this_iops = 0;
 332
 333                 if (unified_rw_rep) {
 334                         iops[i] = 0;
 335                         iops[0] += this_iops;
 336                 } else
 337                         iops[i] = this_iops;
 338
 339                 prev_io_iops[i] = io_iops[i];
 340         }
 341 }
 342
 343 /*
 344  * Print status of the jobs we know about. This includes rate estimates,
 345  * ETA, thread state, etc.
 346  */
 347 bool calc_thread_status(struct jobs_eta *je, int force)
 348 {
 349         struct thread_data *td;
 350         int i, unified_rw_rep;
 351         uint64_t rate_time, disp_time, bw_avg_time, *eta_secs;
 352         unsigned long long io_bytes[DDIR_RWDIR_CNT];
 353         unsigned long long io_iops[DDIR_RWDIR_CNT];
 354         struct timeval now;
 355
 356         static unsigned long long rate_io_bytes[DDIR_RWDIR_CNT];
 357         static unsigned long long disp_io_bytes[DDIR_RWDIR_CNT];
 358         static unsigned long long disp_io_iops[DDIR_RWDIR_CNT];
 359         static struct timeval rate_prev_time, disp_prev_time;
 360
 361         if (!force) {
 362                 if (!(output_format & FIO_OUTPUT_NORMAL) &&
 363                     f_out == stdout)
 364                         return false;
 365                 if (temp_stall_ts || eta_print == FIO_ETA_NEVER)
 366                         return false;
 367
 368                 if (!isatty(STDOUT_FILENO) && (eta_print != FIO_ETA_ALWAYS))
 369                         return false;
 370         }
 371
 372         if (!ddir_rw_sum(rate_io_bytes))
 373                 fill_start_time(&rate_prev_time);
 374         if (!ddir_rw_sum(disp_io_bytes))
 375                 fill_start_time(&disp_prev_time);
 376
 377         eta_secs = malloc(thread_number * sizeof(uint64_t));
 378         memset(eta_secs, 0, thread_number * sizeof(uint64_t));
 379
 380         je->elapsed_sec = (mtime_since_genesis() + 999) / 1000;
 381
 382         io_bytes[DDIR_READ] = io_bytes[DDIR_WRITE] = io_bytes[DDIR_TRIM] = 0;
 383         io_iops[DDIR_READ] = io_iops[DDIR_WRITE] = io_iops[DDIR_TRIM] = 0;
 384         bw_avg_time = ULONG_MAX;
 385         unified_rw_rep = 0;
 386         for_each_td(td, i) {
 387                 unified_rw_rep += td->o.unified_rw_rep;
 388                 if (is_power_of_2(td->o.kb_base))
 389                         je->is_pow2 = 1;
 390                 je->unit_base = td->o.unit_base;
 391                 if (td->o.bw_avg_time < bw_avg_time)
 392                         bw_avg_time = td->o.bw_avg_time;
 393                 if (td->runstate == TD_RUNNING || td->runstate == TD_VERIFYING
 394                     || td->runstate == TD_FSYNCING
 395                     || td->runstate == TD_PRE_READING
 396                     || td->runstate == TD_FINISHING) {
 397                         je->nr_running++;
 398                         if (td_read(td)) {
 399                                 je->t_rate[0] += td->o.rate[DDIR_READ];
 400                                 je->t_iops[0] += td->o.rate_iops[DDIR_READ];
 401                                 je->m_rate[0] += td->o.ratemin[DDIR_READ];
 402                                 je->m_iops[0] += td->o.rate_iops_min[DDIR_READ];
 403                         }
 404                         if (td_write(td)) {
 405                                 je->t_rate[1] += td->o.rate[DDIR_WRITE];
 406                                 je->t_iops[1] += td->o.rate_iops[DDIR_WRITE];
 407                                 je->m_rate[1] += td->o.ratemin[DDIR_WRITE];
 408                                 je->m_iops[1] += td->o.rate_iops_min[DDIR_WRITE];
 409                         }
 410                         if (td_trim(td)) {
 411                                 je->t_rate[2] += td->o.rate[DDIR_TRIM];
 412                                 je->t_iops[2] += td->o.rate_iops[DDIR_TRIM];
 413                                 je->m_rate[2] += td->o.ratemin[DDIR_TRIM];
 414                                 je->m_iops[2] += td->o.rate_iops_min[DDIR_TRIM];
 415                         }
 416
 417                         je->files_open += td->nr_open_files;
 418                 } else if (td->runstate == TD_RAMP) {
 419                         je->nr_running++;
 420                         je->nr_ramp++;
 421                 } else if (td->runstate == TD_SETTING_UP)
 422                         je->nr_setting_up++;
 423                 else if (td->runstate < TD_RUNNING)
 424                         je->nr_pending++;
 425
 426                 if (je->elapsed_sec >= 3)
 427                         eta_secs[i] = thread_eta(td);
 428                 else
 429                         eta_secs[i] = INT_MAX;
 430
 431                 check_str_update(td);
 432
 433                 if (td->runstate > TD_SETTING_UP) {
 434                         int ddir;
 435
 436                         for (ddir = DDIR_READ; ddir < DDIR_RWDIR_CNT; ddir++) {
 437                                 if (unified_rw_rep) {
 438                                         io_bytes[0] += td->io_bytes[ddir];
 439                                         io_iops[0] += td->io_blocks[ddir];
 440                                 } else {
 441                                         io_bytes[ddir] += td->io_bytes[ddir];
 442                                         io_iops[ddir] += td->io_blocks[ddir];
 443                                 }
 444                         }
 445                 }
 446         }
 447
 448         if (exitall_on_terminate) {
 449                 je->eta_sec = INT_MAX;
 450                 for_each_td(td, i) {
 451                         if (eta_secs[i] < je->eta_sec)
 452                                 je->eta_sec = eta_secs[i];
 453                 }
 454         } else {
 455                 unsigned long eta_stone = 0;
 456
 457                 je->eta_sec = 0;
 458                 for_each_td(td, i) {
 459                         if ((td->runstate == TD_NOT_CREATED) && td->o.stonewall)
 460                                 eta_stone += eta_secs[i];
 461                         else {
 462                                 if (eta_secs[i] > je->eta_sec)
 463                                         je->eta_sec = eta_secs[i];
 464                         }
 465                 }
 466                 je->eta_sec += eta_stone;
 467         }
 468
 469         free(eta_secs);
 470
 471         fio_gettime(&now, NULL);
 472         rate_time = mtime_since(&rate_prev_time, &now);
 473
 474         if (write_bw_log && rate_time > bw_avg_time && !in_ramp_time(td)) {
 475                 calc_rate(unified_rw_rep, rate_time, io_bytes, rate_io_bytes,
 476                                 je->rate);
 477                 memcpy(&rate_prev_time, &now, sizeof(now));
 478                 add_agg_sample(je->rate[DDIR_READ], DDIR_READ, 0);
 479                 add_agg_sample(je->rate[DDIR_WRITE], DDIR_WRITE, 0);
 480                 add_agg_sample(je->rate[DDIR_TRIM], DDIR_TRIM, 0);
 481         }
 482
 483         disp_time = mtime_since(&disp_prev_time, &now);
 484
 485         /*
 486          * Allow a little slack, the target is to print it every 1000 msecs
 487          */
 488         if (!force && disp_time < 900)
 489                 return false;
 490
 491         calc_rate(unified_rw_rep, disp_time, io_bytes, disp_io_bytes, je->rate);
 492         calc_iops(unified_rw_rep, disp_time, io_iops, disp_io_iops, je->iops);
 493
 494         memcpy(&disp_prev_time, &now, sizeof(now));
 495
 496         if (!force && !je->nr_running && !je->nr_pending)
 497                 return false;
 498
 499         je->nr_threads = thread_number;
 500         update_condensed_str(__run_str, run_str);
 501         memcpy(je->run_str, run_str, strlen(run_str));
 502         return true;
 503 }
 504
 505 void display_thread_status(struct jobs_eta *je)
 506 {
 507         static struct timeval disp_eta_new_line;
 508         static int eta_new_line_init, eta_new_line_pending;
 509         static int linelen_last;
 510         static int eta_good;
 511         char output[REAL_MAX_JOBS + 512], *p = output;
 512         char eta_str[128];
 513         double perc = 0.0;
 514
 515         if (je->eta_sec != INT_MAX && je->elapsed_sec) {
 516                 perc = (double) je->elapsed_sec / (double) (je->elapsed_sec + je->eta_sec);
 517                 eta_to_str(eta_str, je->eta_sec);
 518         }
 519
 520         if (eta_new_line_pending) {
 521                 eta_new_line_pending = 0;
 522                 p += sprintf(p, "\n");
 523         }
 524
 525         p += sprintf(p, "Jobs: %d (f=%d)", je->nr_running, je->files_open);
 526         if (je->m_rate[0] || je->m_rate[1] || je->t_rate[0] || je->t_rate[1]) {
 527                 char *tr, *mr;
 528
 529                 mr = num2str(je->m_rate[0] + je->m_rate[1], 4, 0, je->is_pow2, 8);
 530                 tr = num2str(je->t_rate[0] + je->t_rate[1], 4, 0, je->is_pow2, 8);
 531                 p += sprintf(p, ", CR=%s/%s KB/s", tr, mr);
 532                 free(tr);
 533                 free(mr);
 534         } else if (je->m_iops[0] || je->m_iops[1] || je->t_iops[0] || je->t_iops[1]) {
 535                 p += sprintf(p, ", CR=%d/%d IOPS",
 536                                         je->t_iops[0] + je->t_iops[1],
 537                                         je->m_iops[0] + je->m_iops[1]);
 538         }
 539         if (je->eta_sec != INT_MAX && je->nr_running) {
 540                 char perc_str[32];
 541                 char *iops_str[DDIR_RWDIR_CNT];
 542                 char *rate_str[DDIR_RWDIR_CNT];
 543                 size_t left;
 544                 int l;
 545                 int ddir;
 546
 547                 if ((!je->eta_sec && !eta_good) || je->nr_ramp == je->nr_running ||
 548                     je->eta_sec == -1)
 549                         strcpy(perc_str, "-.-% done");
 550                 else {
 551                         double mult = 100.0;
 552
 553                         if (je->nr_setting_up && je->nr_running)
 554                                 mult *= (1.0 - (double) je->nr_setting_up / (double) je->nr_running);
 555
 556                         eta_good = 1;
 557                         perc *= mult;
 558                         sprintf(perc_str, "%3.1f%% done", perc);
 559                 }
 560
 561                 for (ddir = DDIR_READ; ddir < DDIR_RWDIR_CNT; ddir++) {
 562                         rate_str[ddir] = num2str(je->rate[ddir], 5,
 563                                                 1024, je->is_pow2, je->unit_base);
 564                         iops_str[ddir] = num2str(je->iops[ddir], 4, 1, 0, 0);
 565                 }
 566
 567                 left = sizeof(output) - (p - output) - 1;
 568
 569                 l = snprintf(p, left, ": [%s] [%s] [%s/%s/%s /s] [%s/%s/%s iops] [eta %s]",
 570                                 je->run_str, perc_str, rate_str[DDIR_READ],
 571                                 rate_str[DDIR_WRITE], rate_str[DDIR_TRIM],
 572                                 iops_str[DDIR_READ], iops_str[DDIR_WRITE],
 573                                 iops_str[DDIR_TRIM], eta_str);
 574                 p += l;
 575                 if (l >= 0 && l < linelen_last)
 576                         p += sprintf(p, "%*s", linelen_last - l, "");
 577                 linelen_last = l;
 578
 579                 for (ddir = DDIR_READ; ddir < DDIR_RWDIR_CNT; ddir++) {
 580                         free(rate_str[ddir]);
 581                         free(iops_str[ddir]);
 582                 }
 583         }
 584         p += sprintf(p, "\r");
 585
 586         printf("%s", output);
 587
 588         if (!eta_new_line_init) {
 589                 fio_gettime(&disp_eta_new_line, NULL);
 590                 eta_new_line_init = 1;
 591         } else if (eta_new_line && mtime_since_now(&disp_eta_new_line) > eta_new_line) {
 592                 fio_gettime(&disp_eta_new_line, NULL);
 593                 eta_new_line_pending = 1;
 594         }
 595
 596         fflush(stdout);
 597 }
 598
 599 struct jobs_eta *get_jobs_eta(bool force, size_t *size)
 600 {
 601         struct jobs_eta *je;
 602
 603         if (!thread_number)
 604                 return NULL;
 605
 606         *size = sizeof(*je) + THREAD_RUNSTR_SZ + 8;
 607         je = malloc(*size);
 608         if (!je)
 609                 return NULL;
 610         memset(je, 0, *size);
 611
 612         if (!calc_thread_status(je, force)) {
 613                 free(je);
 614                 return NULL;
 615         }
 616
 617         *size = sizeof(*je) + strlen((char *) je->run_str) + 1;
 618         return je;
 619 }
 620
 621 void print_thread_status(void)
 622 {
 623         struct jobs_eta *je;
 624         size_t size;
 625
 626         je = get_jobs_eta(false, &size);
 627         if (je)
 628                 display_thread_status(je);
 629
 630         free(je);
 631 }
 632
 633 void print_status_init(int thr_number)
 634 {
 635         __run_str[thr_number] = 'P';
 636         update_condensed_str(__run_str, run_str);
 637 }