iolog.c

   1 /*
   2  * Code related to writing an iolog of what a thread is doing, and to
   3  * later read that back and replay
   4  */
   5 #include <stdio.h>
   6 #include <stdlib.h>
   7 #include <libgen.h>
   8 #include <assert.h>
   9 #include "flist.h"
  10 #include "fio.h"
  11 #include "verify.h"
  12 #include "trim.h"
  13 #include "filelock.h"
  14
  15 static const char iolog_ver2[] = "fio version 2 iolog";
  16
  17 void queue_io_piece(struct thread_data *td, struct io_piece *ipo)
  18 {
  19         flist_add_tail(&ipo->list, &td->io_log_list);
  20         td->total_io_size += ipo->len;
  21 }
  22
  23 void log_io_u(struct thread_data *td, struct io_u *io_u)
  24 {
  25         const char *act[] = { "read", "write", "sync", "datasync",
  26                                 "sync_file_range", "wait", "trim" };
  27
  28         assert(io_u->ddir <= 6);
  29
  30         if (!td->o.write_iolog_file)
  31                 return;
  32
  33         fprintf(td->iolog_f, "%s %s %llu %lu\n", io_u->file->file_name,
  34                                                 act[io_u->ddir], io_u->offset,
  35                                                 io_u->buflen);
  36 }
  37
  38 void log_file(struct thread_data *td, struct fio_file *f,
  39               enum file_log_act what)
  40 {
  41         const char *act[] = { "add", "open", "close" };
  42
  43         assert(what < 3);
  44
  45         if (!td->o.write_iolog_file)
  46                 return;
  47
  48
  49         /*
  50          * this happens on the pre-open/close done before the job starts
  51          */
  52         if (!td->iolog_f)
  53                 return;
  54
  55         fprintf(td->iolog_f, "%s %s\n", f->file_name, act[what]);
  56 }
  57
  58 static void iolog_delay(struct thread_data *td, unsigned long delay)
  59 {
  60         unsigned long usec = utime_since_now(&td->last_issue);
  61         unsigned long this_delay;
  62
  63         if (delay < usec)
  64                 return;
  65
  66         delay -= usec;
  67
  68         /*
  69          * less than 100 usec delay, just regard it as noise
  70          */
  71         if (delay < 100)
  72                 return;
  73
  74         while (delay && !td->terminate) {
  75                 this_delay = delay;
  76                 if (this_delay > 500000)
  77                         this_delay = 500000;
  78
  79                 usec_sleep(td, this_delay);
  80                 delay -= this_delay;
  81         }
  82 }
  83
  84 static int ipo_special(struct thread_data *td, struct io_piece *ipo)
  85 {
  86         struct fio_file *f;
  87         int ret;
  88
  89         /*
  90          * Not a special ipo
  91          */
  92         if (ipo->ddir != DDIR_INVAL)
  93                 return 0;
  94
  95         f = td->files[ipo->fileno];
  96
  97         switch (ipo->file_action) {
  98         case FIO_LOG_OPEN_FILE:
  99                 ret = td_io_open_file(td, f);
 100                 if (!ret)
 101                         break;
 102                 td_verror(td, ret, "iolog open file");
 103                 return -1;
 104         case FIO_LOG_CLOSE_FILE:
 105                 td_io_close_file(td, f);
 106                 break;
 107         case FIO_LOG_UNLINK_FILE:
 108                 unlink(f->file_name);
 109                 break;
 110         default:
 111                 log_err("fio: bad file action %d\n", ipo->file_action);
 112                 break;
 113         }
 114
 115         return 1;
 116 }
 117
 118 int read_iolog_get(struct thread_data *td, struct io_u *io_u)
 119 {
 120         struct io_piece *ipo;
 121         unsigned long elapsed;
 122
 123         while (!flist_empty(&td->io_log_list)) {
 124                 int ret;
 125
 126                 ipo = flist_entry(td->io_log_list.next, struct io_piece, list);
 127                 flist_del(&ipo->list);
 128                 remove_trim_entry(td, ipo);
 129
 130                 ret = ipo_special(td, ipo);
 131                 if (ret < 0) {
 132                         free(ipo);
 133                         break;
 134                 } else if (ret > 0) {
 135                         free(ipo);
 136                         continue;
 137                 }
 138
 139                 io_u->ddir = ipo->ddir;
 140                 if (ipo->ddir != DDIR_WAIT) {
 141                         io_u->offset = ipo->offset;
 142                         io_u->buflen = ipo->len;
 143                         io_u->file = td->files[ipo->fileno];
 144                         get_file(io_u->file);
 145                         dprint(FD_IO, "iolog: get %llu/%lu/%s\n", io_u->offset,
 146                                                 io_u->buflen, io_u->file->file_name);
 147                         if (ipo->delay)
 148                                 iolog_delay(td, ipo->delay);
 149                 } else {
 150                         elapsed = mtime_since_genesis();
 151                         if (ipo->delay > elapsed)
 152                                 usec_sleep(td, (ipo->delay - elapsed) * 1000);
 153                 }
 154
 155                 free(ipo);
 156
 157                 if (io_u->ddir != DDIR_WAIT)
 158                         return 0;
 159         }
 160
 161         td->done = 1;
 162         return 1;
 163 }
 164
 165 void prune_io_piece_log(struct thread_data *td)
 166 {
 167         struct io_piece *ipo;
 168         struct rb_node *n;
 169
 170         while ((n = rb_first(&td->io_hist_tree)) != NULL) {
 171                 ipo = rb_entry(n, struct io_piece, rb_node);
 172                 rb_erase(n, &td->io_hist_tree);
 173                 remove_trim_entry(td, ipo);
 174                 td->io_hist_len--;
 175                 free(ipo);
 176         }
 177
 178         while (!flist_empty(&td->io_hist_list)) {
 179                 ipo = flist_entry(td->io_hist_list.next, struct io_piece, list);
 180                 flist_del(&ipo->list);
 181                 remove_trim_entry(td, ipo);
 182                 td->io_hist_len--;
 183                 free(ipo);
 184         }
 185 }
 186
 187 /*
 188  * log a successful write, so we can unwind the log for verify
 189  */
 190 void log_io_piece(struct thread_data *td, struct io_u *io_u)
 191 {
 192         struct rb_node **p, *parent;
 193         struct io_piece *ipo, *__ipo;
 194
 195         ipo = malloc(sizeof(struct io_piece));
 196         init_ipo(ipo);
 197         ipo->file = io_u->file;
 198         ipo->offset = io_u->offset;
 199         ipo->len = io_u->buflen;
 200         ipo->numberio = io_u->numberio;
 201         ipo->flags = IP_F_IN_FLIGHT;
 202
 203         io_u->ipo = ipo;
 204
 205         if (io_u_should_trim(td, io_u)) {
 206                 flist_add_tail(&ipo->trim_list, &td->trim_list);
 207                 td->trim_entries++;
 208         }
 209
 210         /*
 211          * We don't need to sort the entries, if:
 212          *
 213          *      Sequential writes, or
 214          *      Random writes that lay out the file as it goes along
 215          *
 216          * For both these cases, just reading back data in the order we
 217          * wrote it out is the fastest.
 218          *
 219          * One exception is if we don't have a random map AND we are doing
 220          * verifies, in that case we need to check for duplicate blocks and
 221          * drop the old one, which we rely on the rb insert/lookup for
 222          * handling.
 223          */
 224         if (((!td->o.verifysort) || !td_random(td) || !td->o.overwrite) &&
 225               (file_randommap(td, ipo->file) || td->o.verify == VERIFY_NONE)) {
 226                 INIT_FLIST_HEAD(&ipo->list);
 227                 flist_add_tail(&ipo->list, &td->io_hist_list);
 228                 ipo->flags |= IP_F_ONLIST;
 229                 td->io_hist_len++;
 230                 return;
 231         }
 232
 233         RB_CLEAR_NODE(&ipo->rb_node);
 234
 235         /*
 236          * Sort the entry into the verification list
 237          */
 238 restart:
 239         p = &td->io_hist_tree.rb_node;
 240         parent = NULL;
 241         while (*p) {
 242                 parent = *p;
 243
 244                 __ipo = rb_entry(parent, struct io_piece, rb_node);
 245                 if (ipo->file < __ipo->file)
 246                         p = &(*p)->rb_left;
 247                 else if (ipo->file > __ipo->file)
 248                         p = &(*p)->rb_right;
 249                 else if (ipo->offset < __ipo->offset)
 250                         p = &(*p)->rb_left;
 251                 else if (ipo->offset > __ipo->offset)
 252                         p = &(*p)->rb_right;
 253                 else {
 254                         dprint(FD_IO, "iolog: overlap %llu/%lu, %llu/%lu",
 255                                 __ipo->offset, __ipo->len,
 256                                 ipo->offset, ipo->len);
 257                         td->io_hist_len--;
 258                         rb_erase(parent, &td->io_hist_tree);
 259                         remove_trim_entry(td, __ipo);
 260                         free(__ipo);
 261                         goto restart;
 262                 }
 263         }
 264
 265         rb_link_node(&ipo->rb_node, parent, p);
 266         rb_insert_color(&ipo->rb_node, &td->io_hist_tree);
 267         ipo->flags |= IP_F_ONRB;
 268         td->io_hist_len++;
 269 }
 270
 271 void write_iolog_close(struct thread_data *td)
 272 {
 273         fflush(td->iolog_f);
 274         fclose(td->iolog_f);
 275         free(td->iolog_buf);
 276         td->iolog_f = NULL;
 277         td->iolog_buf = NULL;
 278 }
 279
 280 /*
 281  * Read version 2 iolog data. It is enhanced to include per-file logging,
 282  * syncs, etc.
 283  */
 284 static int read_iolog2(struct thread_data *td, FILE *f)
 285 {
 286         unsigned long long offset;
 287         unsigned int bytes;
 288         int reads, writes, waits, fileno = 0, file_action = 0; /* stupid gcc */
 289         char *fname, *act;
 290         char *str, *p;
 291         enum fio_ddir rw;
 292
 293         free_release_files(td);
 294
 295         /*
 296          * Read in the read iolog and store it, reuse the infrastructure
 297          * for doing verifications.
 298          */
 299         str = malloc(4096);
 300         fname = malloc(256+16);
 301         act = malloc(256+16);
 302
 303         reads = writes = waits = 0;
 304         while ((p = fgets(str, 4096, f)) != NULL) {
 305                 struct io_piece *ipo;
 306                 int r;
 307
 308                 r = sscanf(p, "%256s %256s %llu %u", fname, act, &offset,
 309                                                                         &bytes);
 310                 if (r == 4) {
 311                         /*
 312                          * Check action first
 313                          */
 314                         if (!strcmp(act, "wait"))
 315                                 rw = DDIR_WAIT;
 316                         else if (!strcmp(act, "read"))
 317                                 rw = DDIR_READ;
 318                         else if (!strcmp(act, "write"))
 319                                 rw = DDIR_WRITE;
 320                         else if (!strcmp(act, "sync"))
 321                                 rw = DDIR_SYNC;
 322                         else if (!strcmp(act, "datasync"))
 323                                 rw = DDIR_DATASYNC;
 324                         else if (!strcmp(act, "trim"))
 325                                 rw = DDIR_TRIM;
 326                         else {
 327                                 log_err("fio: bad iolog file action: %s\n",
 328                                                                         act);
 329                                 continue;
 330                         }
 331                         fileno = get_fileno(td, fname);
 332                 } else if (r == 2) {
 333                         rw = DDIR_INVAL;
 334                         if (!strcmp(act, "add")) {
 335                                 fileno = add_file(td, fname, 0, 1);
 336                                 file_action = FIO_LOG_ADD_FILE;
 337                                 continue;
 338                         } else if (!strcmp(act, "open")) {
 339                                 fileno = get_fileno(td, fname);
 340                                 file_action = FIO_LOG_OPEN_FILE;
 341                         } else if (!strcmp(act, "close")) {
 342                                 fileno = get_fileno(td, fname);
 343                                 file_action = FIO_LOG_CLOSE_FILE;
 344                         } else {
 345                                 log_err("fio: bad iolog file action: %s\n",
 346                                                                         act);
 347                                 continue;
 348                         }
 349                 } else {
 350                         log_err("bad iolog2: %s", p);
 351                         continue;
 352                 }
 353
 354                 if (rw == DDIR_READ)
 355                         reads++;
 356                 else if (rw == DDIR_WRITE) {
 357                         /*
 358                          * Don't add a write for ro mode
 359                          */
 360                         if (read_only)
 361                                 continue;
 362                         writes++;
 363                 } else if (rw == DDIR_WAIT) {
 364                         waits++;
 365                 } else if (rw == DDIR_INVAL) {
 366                 } else if (!ddir_sync(rw)) {
 367                         log_err("bad ddir: %d\n", rw);
 368                         continue;
 369                 }
 370
 371                 /*
 372                  * Make note of file
 373                  */
 374                 ipo = malloc(sizeof(*ipo));
 375                 init_ipo(ipo);
 376                 ipo->ddir = rw;
 377                 if (rw == DDIR_WAIT) {
 378                         ipo->delay = offset;
 379                 } else {
 380                         ipo->offset = offset;
 381                         ipo->len = bytes;
 382                         if (rw != DDIR_INVAL && bytes > td->o.max_bs[rw])
 383                                 td->o.max_bs[rw] = bytes;
 384                         ipo->fileno = fileno;
 385                         ipo->file_action = file_action;
 386                 }
 387
 388                 queue_io_piece(td, ipo);
 389         }
 390
 391         free(str);
 392         free(act);
 393         free(fname);
 394
 395         if (writes && read_only) {
 396                 log_err("fio: <%s> skips replay of %d writes due to"
 397                         " read-only\n", td->o.name, writes);
 398                 writes = 0;
 399         }
 400
 401         if (!reads && !writes && !waits)
 402                 return 1;
 403         else if (reads && !writes)
 404                 td->o.td_ddir = TD_DDIR_READ;
 405         else if (!reads && writes)
 406                 td->o.td_ddir = TD_DDIR_WRITE;
 407         else
 408                 td->o.td_ddir = TD_DDIR_RW;
 409
 410         return 0;
 411 }
 412
 413 /*
 414  * open iolog, check version, and call appropriate parser
 415  */
 416 static int init_iolog_read(struct thread_data *td)
 417 {
 418         char buffer[256], *p;
 419         FILE *f;
 420         int ret;
 421
 422         f = fopen(td->o.read_iolog_file, "r");
 423         if (!f) {
 424                 perror("fopen read iolog");
 425                 return 1;
 426         }
 427
 428         p = fgets(buffer, sizeof(buffer), f);
 429         if (!p) {
 430                 td_verror(td, errno, "iolog read");
 431                 log_err("fio: unable to read iolog\n");
 432                 fclose(f);
 433                 return 1;
 434         }
 435
 436         /*
 437          * version 2 of the iolog stores a specific string as the
 438          * first line, check for that
 439          */
 440         if (!strncmp(iolog_ver2, buffer, strlen(iolog_ver2)))
 441                 ret = read_iolog2(td, f);
 442         else {
 443                 log_err("fio: iolog version 1 is no longer supported\n");
 444                 ret = 1;
 445         }
 446
 447         fclose(f);
 448         return ret;
 449 }
 450
 451 /*
 452  * Set up a log for storing io patterns.
 453  */
 454 static int init_iolog_write(struct thread_data *td)
 455 {
 456         struct fio_file *ff;
 457         FILE *f;
 458         unsigned int i;
 459
 460         f = fopen(td->o.write_iolog_file, "a");
 461         if (!f) {
 462                 perror("fopen write iolog");
 463                 return 1;
 464         }
 465
 466         /*
 467          * That's it for writing, setup a log buffer and we're done.
 468           */
 469         td->iolog_f = f;
 470         td->iolog_buf = malloc(8192);
 471         setvbuf(f, td->iolog_buf, _IOFBF, 8192);
 472
 473         /*
 474          * write our version line
 475          */
 476         if (fprintf(f, "%s\n", iolog_ver2) < 0) {
 477                 perror("iolog init\n");
 478                 return 1;
 479         }
 480
 481         /*
 482          * add all known files
 483          */
 484         for_each_file(td, ff, i)
 485                 log_file(td, ff, FIO_LOG_ADD_FILE);
 486
 487         return 0;
 488 }
 489
 490 int init_iolog(struct thread_data *td)
 491 {
 492         int ret = 0;
 493
 494         if (td->o.read_iolog_file) {
 495                 int need_swap;
 496
 497                 /*
 498                  * Check if it's a blktrace file and load that if possible.
 499                  * Otherwise assume it's a normal log file and load that.
 500                  */
 501                 if (is_blktrace(td->o.read_iolog_file, &need_swap))
 502                         ret = load_blktrace(td, td->o.read_iolog_file, need_swap);
 503                 else
 504                         ret = init_iolog_read(td);
 505         } else if (td->o.write_iolog_file)
 506                 ret = init_iolog_write(td);
 507
 508         if (ret)
 509                 td_verror(td, EINVAL, "failed initializing iolog");
 510
 511         return ret;
 512 }
 513
 514 void setup_log(struct io_log **log, unsigned long avg_msec, int log_type)
 515 {
 516         struct io_log *l = malloc(sizeof(*l));
 517
 518         memset(l, 0, sizeof(*l));
 519         l->nr_samples = 0;
 520         l->max_samples = 1024;
 521         l->log_type = log_type;
 522         l->log = malloc(l->max_samples * sizeof(struct io_sample));
 523         l->avg_msec = avg_msec;
 524         *log = l;
 525 }
 526
 527 #ifdef CONFIG_SETVBUF
 528 static void *set_file_buffer(FILE *f)
 529 {
 530         size_t size = 1048576;
 531         void *buf;
 532
 533         buf = malloc(size);
 534         setvbuf(f, buf, _IOFBF, size);
 535         return buf;
 536 }
 537
 538 static void clear_file_buffer(void *buf)
 539 {
 540         free(buf);
 541 }
 542 #else
 543 static void *set_file_buffer(FILE *f)
 544 {
 545         return NULL;
 546 }
 547
 548 static void clear_file_buffer(void *buf)
 549 {
 550 }
 551 #endif
 552
 553 void __finish_log(struct io_log *log, const char *name)
 554 {
 555         unsigned int i;
 556         void *buf;
 557         FILE *f;
 558
 559         f = fopen(name, "a");
 560         if (!f) {
 561                 perror("fopen log");
 562                 return;
 563         }
 564
 565         buf = set_file_buffer(f);
 566
 567         for (i = 0; i < log->nr_samples; i++) {
 568                 fprintf(f, "%lu, %lu, %u, %u\n",
 569                                 (unsigned long) log->log[i].time,
 570                                 (unsigned long) log->log[i].val,
 571                                 log->log[i].ddir, log->log[i].bs);
 572         }
 573
 574         fclose(f);
 575         clear_file_buffer(buf);
 576         free(log->log);
 577         free(log);
 578 }
 579
 580 static int finish_log_named(struct thread_data *td, struct io_log *log,
 581                             const char *prefix, const char *postfix,
 582                             int trylock)
 583 {
 584         char file_name[256];
 585
 586         snprintf(file_name, sizeof(file_name), "%s_%s.log", prefix, postfix);
 587
 588         if (trylock) {
 589                 if (fio_trylock_file(file_name))
 590                         return 1;
 591         } else
 592                 fio_lock_file(file_name);
 593
 594         if (td->client_type == FIO_CLIENT_TYPE_GUI) {
 595                 fio_send_iolog(td, log, file_name);
 596                 free(log->log);
 597                 free(log);
 598         } else
 599                 __finish_log(log, file_name);
 600
 601         fio_unlock_file(file_name);
 602         return 0;
 603 }
 604
 605 static int finish_log(struct thread_data *td, struct io_log *log,
 606                       const char *name, int trylock)
 607 {
 608         return finish_log_named(td, log, td->o.name, name, trylock);
 609 }
 610
 611 static int write_this_log(struct thread_data *td, struct io_log *log,
 612                           const char *log_file, const char *name, int try)
 613 {
 614         int ret;
 615
 616         if (!log)
 617                 return 0;
 618
 619         if (log_file)
 620                 ret = finish_log_named(td, log, log_file, name, try);
 621         else
 622                 ret = finish_log(td, log, name, try);
 623
 624         return ret;
 625 }
 626
 627 static int write_iops_log(struct thread_data *td, int try)
 628 {
 629         struct thread_options *o = &td->o;
 630
 631         return write_this_log(td, td->iops_log, o->iops_log_file, "iops", try);
 632 }
 633
 634 static int write_slat_log(struct thread_data *td, int try)
 635 {
 636         struct thread_options *o = &td->o;
 637
 638         return write_this_log(td, td->slat_log, o->lat_log_file, "slat", try);
 639 }
 640
 641 static int write_clat_log(struct thread_data *td, int try)
 642 {
 643         struct thread_options *o = &td->o;
 644
 645         return write_this_log(td, td->clat_log, o->lat_log_file, "clat" , try);
 646 }
 647
 648 static int write_lat_log(struct thread_data *td, int try)
 649 {
 650         struct thread_options *o = &td->o;
 651
 652         return write_this_log(td, td->lat_log, o->lat_log_file, "lat", try);
 653 }
 654
 655 static int write_bandw_log(struct thread_data *td, int try)
 656 {
 657         struct thread_options *o = &td->o;
 658
 659         return write_this_log(td, td->bw_log, o->bw_log_file, "bw", try);
 660 }
 661
 662 enum {
 663         BW_LOG_MASK     = 1,
 664         LAT_LOG_MASK    = 2,
 665         SLAT_LOG_MASK   = 4,
 666         CLAT_LOG_MASK   = 8,
 667         IOPS_LOG_MASK   = 16,
 668
 669         ALL_LOG_NR      = 5,
 670 };
 671
 672 struct log_type {
 673         unsigned int mask;
 674         int (*fn)(struct thread_data *, int);
 675 };
 676
 677 static struct log_type log_types[] = {
 678         {
 679                 .mask   = BW_LOG_MASK,
 680                 .fn     = write_bandw_log,
 681         },
 682         {
 683                 .mask   = LAT_LOG_MASK,
 684                 .fn     = write_lat_log,
 685         },
 686         {
 687                 .mask   = SLAT_LOG_MASK,
 688                 .fn     = write_slat_log,
 689         },
 690         {
 691                 .mask   = CLAT_LOG_MASK,
 692                 .fn     = write_clat_log,
 693         },
 694         {
 695                 .mask   = IOPS_LOG_MASK,
 696                 .fn     = write_iops_log,
 697         },
 698 };
 699
 700 void fio_writeout_logs(struct thread_data *td)
 701 {
 702         unsigned int log_mask = 0;
 703         unsigned int log_left = ALL_LOG_NR;
 704         int old_state, i;
 705
 706         old_state = td_bump_runstate(td, TD_FINISHING);
 707
 708         finalize_logs(td);
 709
 710         while (log_left) {
 711                 int prev_log_left = log_left;
 712
 713                 for (i = 0; i < ALL_LOG_NR && log_left; i++) {
 714                         struct log_type *lt = &log_types[i];
 715                         int ret;
 716
 717                         if (!(log_mask & lt->mask)) {
 718                                 ret = lt->fn(td, log_left != 1);
 719                                 if (!ret) {
 720                                         log_left--;
 721                                         log_mask |= lt->mask;
 722                                 }
 723                         }
 724                 }
 725
 726                 if (prev_log_left == log_left)
 727                         usleep(5000);
 728         }
 729
 730         td_restore_runstate(td, old_state);
 731 }