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 <sys/types.h>
  10 #include <sys/stat.h>
  11 #include <unistd.h>
  12 #ifdef CONFIG_ZLIB
  13 #include <zlib.h>
  14 #endif
  15
  16 #include "flist.h"
  17 #include "fio.h"
  18 #include "verify.h"
  19 #include "trim.h"
  20 #include "filelock.h"
  21 #include "smalloc.h"
  22
  23 static int iolog_flush(struct io_log *log);
  24
  25 static const char iolog_ver2[] = "fio version 2 iolog";
  26
  27 void queue_io_piece(struct thread_data *td, struct io_piece *ipo)
  28 {
  29         flist_add_tail(&ipo->list, &td->io_log_list);
  30         td->total_io_size += ipo->len;
  31 }
  32
  33 void log_io_u(const struct thread_data *td, const struct io_u *io_u)
  34 {
  35         if (!td->o.write_iolog_file)
  36                 return;
  37
  38         fprintf(td->iolog_f, "%s %s %llu %lu\n", io_u->file->file_name,
  39                                                 io_ddir_name(io_u->ddir),
  40                                                 io_u->offset, io_u->buflen);
  41 }
  42
  43 void log_file(struct thread_data *td, struct fio_file *f,
  44               enum file_log_act what)
  45 {
  46         const char *act[] = { "add", "open", "close" };
  47
  48         assert(what < 3);
  49
  50         if (!td->o.write_iolog_file)
  51                 return;
  52
  53
  54         /*
  55          * this happens on the pre-open/close done before the job starts
  56          */
  57         if (!td->iolog_f)
  58                 return;
  59
  60         fprintf(td->iolog_f, "%s %s\n", f->file_name, act[what]);
  61 }
  62
  63 static void iolog_delay(struct thread_data *td, unsigned long delay)
  64 {
  65         uint64_t usec = utime_since_now(&td->last_issue);
  66         uint64_t this_delay;
  67         struct timeval tv;
  68
  69         if (delay < td->time_offset) {
  70                 td->time_offset = 0;
  71                 return;
  72         }
  73
  74         delay -= td->time_offset;
  75         if (delay < usec)
  76                 return;
  77
  78         delay -= usec;
  79
  80         fio_gettime(&tv, NULL);
  81         while (delay && !td->terminate) {
  82                 this_delay = delay;
  83                 if (this_delay > 500000)
  84                         this_delay = 500000;
  85
  86                 usec_sleep(td, this_delay);
  87                 delay -= this_delay;
  88         }
  89
  90         usec = utime_since_now(&tv);
  91         if (usec > delay)
  92                 td->time_offset = usec - delay;
  93         else
  94                 td->time_offset = 0;
  95 }
  96
  97 static int ipo_special(struct thread_data *td, struct io_piece *ipo)
  98 {
  99         struct fio_file *f;
 100         int ret;
 101
 102         /*
 103          * Not a special ipo
 104          */
 105         if (ipo->ddir != DDIR_INVAL)
 106                 return 0;
 107
 108         f = td->files[ipo->fileno];
 109
 110         switch (ipo->file_action) {
 111         case FIO_LOG_OPEN_FILE:
 112                 if (td->o.replay_redirect && fio_file_open(f)) {
 113                         dprint(FD_FILE, "iolog: ignoring re-open of file %s\n",
 114                                         f->file_name);
 115                         break;
 116                 }
 117                 ret = td_io_open_file(td, f);
 118                 if (!ret)
 119                         break;
 120                 td_verror(td, ret, "iolog open file");
 121                 return -1;
 122         case FIO_LOG_CLOSE_FILE:
 123                 td_io_close_file(td, f);
 124                 break;
 125         case FIO_LOG_UNLINK_FILE:
 126                 td_io_unlink_file(td, f);
 127                 break;
 128         default:
 129                 log_err("fio: bad file action %d\n", ipo->file_action);
 130                 break;
 131         }
 132
 133         return 1;
 134 }
 135
 136 int read_iolog_get(struct thread_data *td, struct io_u *io_u)
 137 {
 138         struct io_piece *ipo;
 139         unsigned long elapsed;
 140
 141         while (!flist_empty(&td->io_log_list)) {
 142                 int ret;
 143
 144                 ipo = flist_first_entry(&td->io_log_list, struct io_piece, list);
 145                 flist_del(&ipo->list);
 146                 remove_trim_entry(td, ipo);
 147
 148                 ret = ipo_special(td, ipo);
 149                 if (ret < 0) {
 150                         free(ipo);
 151                         break;
 152                 } else if (ret > 0) {
 153                         free(ipo);
 154                         continue;
 155                 }
 156
 157                 io_u->ddir = ipo->ddir;
 158                 if (ipo->ddir != DDIR_WAIT) {
 159                         io_u->offset = ipo->offset;
 160                         io_u->buflen = ipo->len;
 161                         io_u->file = td->files[ipo->fileno];
 162                         get_file(io_u->file);
 163                         dprint(FD_IO, "iolog: get %llu/%lu/%s\n", io_u->offset,
 164                                                 io_u->buflen, io_u->file->file_name);
 165                         if (ipo->delay)
 166                                 iolog_delay(td, ipo->delay);
 167                 } else {
 168                         elapsed = mtime_since_genesis();
 169                         if (ipo->delay > elapsed)
 170                                 usec_sleep(td, (ipo->delay - elapsed) * 1000);
 171                 }
 172
 173                 free(ipo);
 174
 175                 if (io_u->ddir != DDIR_WAIT)
 176                         return 0;
 177         }
 178
 179         td->done = 1;
 180         return 1;
 181 }
 182
 183 void prune_io_piece_log(struct thread_data *td)
 184 {
 185         struct io_piece *ipo;
 186         struct rb_node *n;
 187
 188         while ((n = rb_first(&td->io_hist_tree)) != NULL) {
 189                 ipo = rb_entry(n, struct io_piece, rb_node);
 190                 rb_erase(n, &td->io_hist_tree);
 191                 remove_trim_entry(td, ipo);
 192                 td->io_hist_len--;
 193                 free(ipo);
 194         }
 195
 196         while (!flist_empty(&td->io_hist_list)) {
 197                 ipo = flist_first_entry(&td->io_hist_list, struct io_piece, list);
 198                 flist_del(&ipo->list);
 199                 remove_trim_entry(td, ipo);
 200                 td->io_hist_len--;
 201                 free(ipo);
 202         }
 203 }
 204
 205 /*
 206  * log a successful write, so we can unwind the log for verify
 207  */
 208 void log_io_piece(struct thread_data *td, struct io_u *io_u)
 209 {
 210         struct rb_node **p, *parent;
 211         struct io_piece *ipo, *__ipo;
 212
 213         ipo = malloc(sizeof(struct io_piece));
 214         init_ipo(ipo);
 215         ipo->file = io_u->file;
 216         ipo->offset = io_u->offset;
 217         ipo->len = io_u->buflen;
 218         ipo->numberio = io_u->numberio;
 219         ipo->flags = IP_F_IN_FLIGHT;
 220
 221         io_u->ipo = ipo;
 222
 223         if (io_u_should_trim(td, io_u)) {
 224                 flist_add_tail(&ipo->trim_list, &td->trim_list);
 225                 td->trim_entries++;
 226         }
 227
 228         /*
 229          * We don't need to sort the entries, if:
 230          *
 231          *      Sequential writes, or
 232          *      Random writes that lay out the file as it goes along
 233          *
 234          * For both these cases, just reading back data in the order we
 235          * wrote it out is the fastest.
 236          *
 237          * One exception is if we don't have a random map AND we are doing
 238          * verifies, in that case we need to check for duplicate blocks and
 239          * drop the old one, which we rely on the rb insert/lookup for
 240          * handling.
 241          */
 242         if (((!td->o.verifysort) || !td_random(td) || !td->o.overwrite) &&
 243               (file_randommap(td, ipo->file) || td->o.verify == VERIFY_NONE)) {
 244                 INIT_FLIST_HEAD(&ipo->list);
 245                 flist_add_tail(&ipo->list, &td->io_hist_list);
 246                 ipo->flags |= IP_F_ONLIST;
 247                 td->io_hist_len++;
 248                 return;
 249         }
 250
 251         RB_CLEAR_NODE(&ipo->rb_node);
 252
 253         /*
 254          * Sort the entry into the verification list
 255          */
 256 restart:
 257         p = &td->io_hist_tree.rb_node;
 258         parent = NULL;
 259         while (*p) {
 260                 int overlap = 0;
 261                 parent = *p;
 262
 263                 __ipo = rb_entry(parent, struct io_piece, rb_node);
 264                 if (ipo->file < __ipo->file)
 265                         p = &(*p)->rb_left;
 266                 else if (ipo->file > __ipo->file)
 267                         p = &(*p)->rb_right;
 268                 else if (ipo->offset < __ipo->offset) {
 269                         p = &(*p)->rb_left;
 270                         overlap = ipo->offset + ipo->len > __ipo->offset;
 271                 }
 272                 else if (ipo->offset > __ipo->offset) {
 273                         p = &(*p)->rb_right;
 274                         overlap = __ipo->offset + __ipo->len > ipo->offset;
 275                 }
 276                 else
 277                         overlap = 1;
 278
 279                 if (overlap) {
 280                         dprint(FD_IO, "iolog: overlap %llu/%lu, %llu/%lu",
 281                                 __ipo->offset, __ipo->len,
 282                                 ipo->offset, ipo->len);
 283                         td->io_hist_len--;
 284                         rb_erase(parent, &td->io_hist_tree);
 285                         remove_trim_entry(td, __ipo);
 286                         free(__ipo);
 287                         goto restart;
 288                 }
 289         }
 290
 291         rb_link_node(&ipo->rb_node, parent, p);
 292         rb_insert_color(&ipo->rb_node, &td->io_hist_tree);
 293         ipo->flags |= IP_F_ONRB;
 294         td->io_hist_len++;
 295 }
 296
 297 void unlog_io_piece(struct thread_data *td, struct io_u *io_u)
 298 {
 299         struct io_piece *ipo = io_u->ipo;
 300
 301         if (td->ts.nr_block_infos) {
 302                 uint32_t *info = io_u_block_info(td, io_u);
 303                 if (BLOCK_INFO_STATE(*info) < BLOCK_STATE_TRIM_FAILURE) {
 304                         if (io_u->ddir == DDIR_TRIM)
 305                                 *info = BLOCK_INFO_SET_STATE(*info,
 306                                                 BLOCK_STATE_TRIM_FAILURE);
 307                         else if (io_u->ddir == DDIR_WRITE)
 308                                 *info = BLOCK_INFO_SET_STATE(*info,
 309                                                 BLOCK_STATE_WRITE_FAILURE);
 310                 }
 311         }
 312
 313         if (!ipo)
 314                 return;
 315
 316         if (ipo->flags & IP_F_ONRB)
 317                 rb_erase(&ipo->rb_node, &td->io_hist_tree);
 318         else if (ipo->flags & IP_F_ONLIST)
 319                 flist_del(&ipo->list);
 320
 321         free(ipo);
 322         io_u->ipo = NULL;
 323         td->io_hist_len--;
 324 }
 325
 326 void trim_io_piece(struct thread_data *td, const struct io_u *io_u)
 327 {
 328         struct io_piece *ipo = io_u->ipo;
 329
 330         if (!ipo)
 331                 return;
 332
 333         ipo->len = io_u->xfer_buflen - io_u->resid;
 334 }
 335
 336 void write_iolog_close(struct thread_data *td)
 337 {
 338         fflush(td->iolog_f);
 339         fclose(td->iolog_f);
 340         free(td->iolog_buf);
 341         td->iolog_f = NULL;
 342         td->iolog_buf = NULL;
 343 }
 344
 345 /*
 346  * Read version 2 iolog data. It is enhanced to include per-file logging,
 347  * syncs, etc.
 348  */
 349 static int read_iolog2(struct thread_data *td, FILE *f)
 350 {
 351         unsigned long long offset;
 352         unsigned int bytes;
 353         int reads, writes, waits, fileno = 0, file_action = 0; /* stupid gcc */
 354         char *rfname, *fname, *act;
 355         char *str, *p;
 356         enum fio_ddir rw;
 357
 358         free_release_files(td);
 359
 360         /*
 361          * Read in the read iolog and store it, reuse the infrastructure
 362          * for doing verifications.
 363          */
 364         str = malloc(4096);
 365         rfname = fname = malloc(256+16);
 366         act = malloc(256+16);
 367
 368         reads = writes = waits = 0;
 369         while ((p = fgets(str, 4096, f)) != NULL) {
 370                 struct io_piece *ipo;
 371                 int r;
 372
 373                 r = sscanf(p, "%256s %256s %llu %u", rfname, act, &offset,
 374                                                                         &bytes);
 375
 376                 if (td->o.replay_redirect)
 377                         fname = td->o.replay_redirect;
 378
 379                 if (r == 4) {
 380                         /*
 381                          * Check action first
 382                          */
 383                         if (!strcmp(act, "wait"))
 384                                 rw = DDIR_WAIT;
 385                         else if (!strcmp(act, "read"))
 386                                 rw = DDIR_READ;
 387                         else if (!strcmp(act, "write"))
 388                                 rw = DDIR_WRITE;
 389                         else if (!strcmp(act, "sync"))
 390                                 rw = DDIR_SYNC;
 391                         else if (!strcmp(act, "datasync"))
 392                                 rw = DDIR_DATASYNC;
 393                         else if (!strcmp(act, "trim"))
 394                                 rw = DDIR_TRIM;
 395                         else {
 396                                 log_err("fio: bad iolog file action: %s\n",
 397                                                                         act);
 398                                 continue;
 399                         }
 400                         fileno = get_fileno(td, fname);
 401                 } else if (r == 2) {
 402                         rw = DDIR_INVAL;
 403                         if (!strcmp(act, "add")) {
 404                                 if (td->o.replay_redirect &&
 405                                     get_fileno(td, fname) != -1) {
 406                                         dprint(FD_FILE, "iolog: ignoring"
 407                                                 " re-add of file %s\n", fname);
 408                                 } else {
 409                                         fileno = add_file(td, fname, 0, 1);
 410                                         file_action = FIO_LOG_ADD_FILE;
 411                                 }
 412                                 continue;
 413                         } else if (!strcmp(act, "open")) {
 414                                 fileno = get_fileno(td, fname);
 415                                 file_action = FIO_LOG_OPEN_FILE;
 416                         } else if (!strcmp(act, "close")) {
 417                                 fileno = get_fileno(td, fname);
 418                                 file_action = FIO_LOG_CLOSE_FILE;
 419                         } else {
 420                                 log_err("fio: bad iolog file action: %s\n",
 421                                                                         act);
 422                                 continue;
 423                         }
 424                 } else {
 425                         log_err("bad iolog2: %s", p);
 426                         continue;
 427                 }
 428
 429                 if (rw == DDIR_READ)
 430                         reads++;
 431                 else if (rw == DDIR_WRITE) {
 432                         /*
 433                          * Don't add a write for ro mode
 434                          */
 435                         if (read_only)
 436                                 continue;
 437                         writes++;
 438                 } else if (rw == DDIR_WAIT) {
 439                         if (td->o.no_stall)
 440                                 continue;
 441                         waits++;
 442                 } else if (rw == DDIR_INVAL) {
 443                 } else if (!ddir_sync(rw)) {
 444                         log_err("bad ddir: %d\n", rw);
 445                         continue;
 446                 }
 447
 448                 /*
 449                  * Make note of file
 450                  */
 451                 ipo = malloc(sizeof(*ipo));
 452                 init_ipo(ipo);
 453                 ipo->ddir = rw;
 454                 if (rw == DDIR_WAIT) {
 455                         ipo->delay = offset;
 456                 } else {
 457                         ipo->offset = offset;
 458                         ipo->len = bytes;
 459                         if (rw != DDIR_INVAL && bytes > td->o.max_bs[rw])
 460                                 td->o.max_bs[rw] = bytes;
 461                         ipo->fileno = fileno;
 462                         ipo->file_action = file_action;
 463                         td->o.size += bytes;
 464                 }
 465
 466                 queue_io_piece(td, ipo);
 467         }
 468
 469         free(str);
 470         free(act);
 471         free(rfname);
 472
 473         if (writes && read_only) {
 474                 log_err("fio: <%s> skips replay of %d writes due to"
 475                         " read-only\n", td->o.name, writes);
 476                 writes = 0;
 477         }
 478
 479         if (!reads && !writes && !waits)
 480                 return 1;
 481         else if (reads && !writes)
 482                 td->o.td_ddir = TD_DDIR_READ;
 483         else if (!reads && writes)
 484                 td->o.td_ddir = TD_DDIR_WRITE;
 485         else
 486                 td->o.td_ddir = TD_DDIR_RW;
 487
 488         return 0;
 489 }
 490
 491 /*
 492  * open iolog, check version, and call appropriate parser
 493  */
 494 static int init_iolog_read(struct thread_data *td)
 495 {
 496         char buffer[256], *p;
 497         FILE *f;
 498         int ret;
 499
 500         f = fopen(td->o.read_iolog_file, "r");
 501         if (!f) {
 502                 perror("fopen read iolog");
 503                 return 1;
 504         }
 505
 506         p = fgets(buffer, sizeof(buffer), f);
 507         if (!p) {
 508                 td_verror(td, errno, "iolog read");
 509                 log_err("fio: unable to read iolog\n");
 510                 fclose(f);
 511                 return 1;
 512         }
 513
 514         /*
 515          * version 2 of the iolog stores a specific string as the
 516          * first line, check for that
 517          */
 518         if (!strncmp(iolog_ver2, buffer, strlen(iolog_ver2)))
 519                 ret = read_iolog2(td, f);
 520         else {
 521                 log_err("fio: iolog version 1 is no longer supported\n");
 522                 ret = 1;
 523         }
 524
 525         fclose(f);
 526         return ret;
 527 }
 528
 529 /*
 530  * Set up a log for storing io patterns.
 531  */
 532 static int init_iolog_write(struct thread_data *td)
 533 {
 534         struct fio_file *ff;
 535         FILE *f;
 536         unsigned int i;
 537
 538         f = fopen(td->o.write_iolog_file, "a");
 539         if (!f) {
 540                 perror("fopen write iolog");
 541                 return 1;
 542         }
 543
 544         /*
 545          * That's it for writing, setup a log buffer and we're done.
 546           */
 547         td->iolog_f = f;
 548         td->iolog_buf = malloc(8192);
 549         setvbuf(f, td->iolog_buf, _IOFBF, 8192);
 550
 551         /*
 552          * write our version line
 553          */
 554         if (fprintf(f, "%s\n", iolog_ver2) < 0) {
 555                 perror("iolog init\n");
 556                 return 1;
 557         }
 558
 559         /*
 560          * add all known files
 561          */
 562         for_each_file(td, ff, i)
 563                 log_file(td, ff, FIO_LOG_ADD_FILE);
 564
 565         return 0;
 566 }
 567
 568 int init_iolog(struct thread_data *td)
 569 {
 570         int ret = 0;
 571
 572         if (td->o.read_iolog_file) {
 573                 int need_swap;
 574
 575                 /*
 576                  * Check if it's a blktrace file and load that if possible.
 577                  * Otherwise assume it's a normal log file and load that.
 578                  */
 579                 if (is_blktrace(td->o.read_iolog_file, &need_swap))
 580                         ret = load_blktrace(td, td->o.read_iolog_file, need_swap);
 581                 else
 582                         ret = init_iolog_read(td);
 583         } else if (td->o.write_iolog_file)
 584                 ret = init_iolog_write(td);
 585
 586         if (ret)
 587                 td_verror(td, EINVAL, "failed initializing iolog");
 588
 589         return ret;
 590 }
 591
 592 void setup_log(struct io_log **log, struct log_params *p,
 593                const char *filename)
 594 {
 595         struct io_log *l;
 596         int i;
 597         struct io_u_plat_entry *entry;
 598         struct flist_head *list;
 599
 600         l = scalloc(1, sizeof(*l));
 601         INIT_FLIST_HEAD(&l->io_logs);
 602         l->log_type = p->log_type;
 603         l->log_offset = p->log_offset;
 604         l->log_gz = p->log_gz;
 605         l->log_gz_store = p->log_gz_store;
 606         l->avg_msec = p->avg_msec;
 607         l->hist_msec = p->hist_msec;
 608         l->hist_coarseness = p->hist_coarseness;
 609         l->filename = strdup(filename);
 610         l->td = p->td;
 611
 612         /* Initialize histogram lists for each r/w direction,
 613          * with initial io_u_plat of all zeros:
 614          */
 615         for (i = 0; i < DDIR_RWDIR_CNT; i++) {
 616                 list = &l->hist_window[i].list;
 617                 INIT_FLIST_HEAD(list);
 618                 entry = calloc(1, sizeof(struct io_u_plat_entry));
 619                 flist_add(&entry->list, list);
 620         }
 621
 622         if (l->td && l->td->o.io_submit_mode != IO_MODE_OFFLOAD) {
 623                 struct io_logs *p;
 624
 625                 p = calloc(1, sizeof(*l->pending));
 626                 p->max_samples = DEF_LOG_ENTRIES;
 627                 p->log = calloc(p->max_samples, log_entry_sz(l));
 628                 l->pending = p;
 629         }
 630
 631         if (l->log_offset)
 632                 l->log_ddir_mask = LOG_OFFSET_SAMPLE_BIT;
 633
 634         INIT_FLIST_HEAD(&l->chunk_list);
 635
 636         if (l->log_gz && !p->td)
 637                 l->log_gz = 0;
 638         else if (l->log_gz || l->log_gz_store) {
 639                 mutex_init_pshared(&l->chunk_lock);
 640                 p->td->flags |= TD_F_COMPRESS_LOG;
 641         }
 642
 643         *log = l;
 644 }
 645
 646 #ifdef CONFIG_SETVBUF
 647 static void *set_file_buffer(FILE *f)
 648 {
 649         size_t size = 1048576;
 650         void *buf;
 651
 652         buf = malloc(size);
 653         setvbuf(f, buf, _IOFBF, size);
 654         return buf;
 655 }
 656
 657 static void clear_file_buffer(void *buf)
 658 {
 659         free(buf);
 660 }
 661 #else
 662 static void *set_file_buffer(FILE *f)
 663 {
 664         return NULL;
 665 }
 666
 667 static void clear_file_buffer(void *buf)
 668 {
 669 }
 670 #endif
 671
 672 void free_log(struct io_log *log)
 673 {
 674         while (!flist_empty(&log->io_logs)) {
 675                 struct io_logs *cur_log;
 676
 677                 cur_log = flist_first_entry(&log->io_logs, struct io_logs, list);
 678                 flist_del_init(&cur_log->list);
 679                 free(cur_log->log);
 680                 sfree(cur_log);
 681         }
 682
 683         if (log->pending) {
 684                 free(log->pending->log);
 685                 free(log->pending);
 686                 log->pending = NULL;
 687         }
 688
 689         free(log->pending);
 690         free(log->filename);
 691         sfree(log);
 692 }
 693
 694 inline unsigned long hist_sum(int j, int stride, unsigned int *io_u_plat,
 695                 unsigned int *io_u_plat_last)
 696 {
 697         unsigned long sum;
 698         int k;
 699
 700         if (io_u_plat_last) {
 701                 for (k = sum = 0; k < stride; k++)
 702                         sum += io_u_plat[j + k] - io_u_plat_last[j + k];
 703         } else {
 704                 for (k = sum = 0; k < stride; k++)
 705                         sum += io_u_plat[j + k];
 706         }
 707
 708         return sum;
 709 }
 710
 711 static void flush_hist_samples(FILE *f, int hist_coarseness, void *samples,
 712                                uint64_t sample_size)
 713 {
 714         struct io_sample *s;
 715         int log_offset;
 716         uint64_t i, j, nr_samples;
 717         struct io_u_plat_entry *entry, *entry_before;
 718         unsigned int *io_u_plat;
 719         unsigned int *io_u_plat_before;
 720
 721         int stride = 1 << hist_coarseness;
 722
 723         if (!sample_size)
 724                 return;
 725
 726         s = __get_sample(samples, 0, 0);
 727         log_offset = (s->__ddir & LOG_OFFSET_SAMPLE_BIT) != 0;
 728
 729         nr_samples = sample_size / __log_entry_sz(log_offset);
 730
 731         for (i = 0; i < nr_samples; i++) {
 732                 s = __get_sample(samples, log_offset, i);
 733
 734                 entry = s->data.plat_entry;
 735                 io_u_plat = entry->io_u_plat;
 736
 737                 entry_before = flist_first_entry(&entry->list, struct io_u_plat_entry, list);
 738                 io_u_plat_before = entry_before->io_u_plat;
 739
 740                 fprintf(f, "%lu, %u, %u, ", (unsigned long) s->time,
 741                                                 io_sample_ddir(s), s->bs);
 742                 for (j = 0; j < FIO_IO_U_PLAT_NR - stride; j += stride) {
 743                         fprintf(f, "%lu, ", hist_sum(j, stride, io_u_plat,
 744                                                 io_u_plat_before));
 745                 }
 746                 fprintf(f, "%lu\n", (unsigned long)
 747                         hist_sum(FIO_IO_U_PLAT_NR - stride, stride, io_u_plat,
 748                                         io_u_plat_before));
 749
 750                 flist_del(&entry_before->list);
 751                 free(entry_before);
 752         }
 753 }
 754
 755 void flush_samples(FILE *f, void *samples, uint64_t sample_size)
 756 {
 757         struct io_sample *s;
 758         int log_offset;
 759         uint64_t i, nr_samples;
 760
 761         if (!sample_size)
 762                 return;
 763
 764         s = __get_sample(samples, 0, 0);
 765         log_offset = (s->__ddir & LOG_OFFSET_SAMPLE_BIT) != 0;
 766
 767         nr_samples = sample_size / __log_entry_sz(log_offset);
 768
 769         for (i = 0; i < nr_samples; i++) {
 770                 s = __get_sample(samples, log_offset, i);
 771
 772                 if (!log_offset) {
 773                         fprintf(f, "%lu, %" PRId64 ", %u, %u\n",
 774                                         (unsigned long) s->time,
 775                                         s->data.val,
 776                                         io_sample_ddir(s), s->bs);
 777                 } else {
 778                         struct io_sample_offset *so = (void *) s;
 779
 780                         fprintf(f, "%lu, %" PRId64 ", %u, %u, %llu\n",
 781                                         (unsigned long) s->time,
 782                                         s->data.val,
 783                                         io_sample_ddir(s), s->bs,
 784                                         (unsigned long long) so->offset);
 785                 }
 786         }
 787 }
 788
 789 #ifdef CONFIG_ZLIB
 790
 791 struct iolog_flush_data {
 792         struct workqueue_work work;
 793         struct io_log *log;
 794         void *samples;
 795         uint32_t nr_samples;
 796         bool free;
 797 };
 798
 799 #define GZ_CHUNK        131072
 800
 801 static struct iolog_compress *get_new_chunk(unsigned int seq)
 802 {
 803         struct iolog_compress *c;
 804
 805         c = malloc(sizeof(*c));
 806         INIT_FLIST_HEAD(&c->list);
 807         c->buf = malloc(GZ_CHUNK);
 808         c->len = 0;
 809         c->seq = seq;
 810         return c;
 811 }
 812
 813 static void free_chunk(struct iolog_compress *ic)
 814 {
 815         free(ic->buf);
 816         free(ic);
 817 }
 818
 819 static int z_stream_init(z_stream *stream, int gz_hdr)
 820 {
 821         int wbits = 15;
 822
 823         memset(stream, 0, sizeof(*stream));
 824         stream->zalloc = Z_NULL;
 825         stream->zfree = Z_NULL;
 826         stream->opaque = Z_NULL;
 827         stream->next_in = Z_NULL;
 828
 829         /*
 830          * zlib magic - add 32 for auto-detection of gz header or not,
 831          * if we decide to store files in a gzip friendly format.
 832          */
 833         if (gz_hdr)
 834                 wbits += 32;
 835
 836         if (inflateInit2(stream, wbits) != Z_OK)
 837                 return 1;
 838
 839         return 0;
 840 }
 841
 842 struct inflate_chunk_iter {
 843         unsigned int seq;
 844         int err;
 845         void *buf;
 846         size_t buf_size;
 847         size_t buf_used;
 848         size_t chunk_sz;
 849 };
 850
 851 static void finish_chunk(z_stream *stream, FILE *f,
 852                          struct inflate_chunk_iter *iter)
 853 {
 854         int ret;
 855
 856         ret = inflateEnd(stream);
 857         if (ret != Z_OK)
 858                 log_err("fio: failed to end log inflation seq %d (%d)\n",
 859                                 iter->seq, ret);
 860
 861         flush_samples(f, iter->buf, iter->buf_used);
 862         free(iter->buf);
 863         iter->buf = NULL;
 864         iter->buf_size = iter->buf_used = 0;
 865 }
 866
 867 /*
 868  * Iterative chunk inflation. Handles cases where we cross into a new
 869  * sequence, doing flush finish of previous chunk if needed.
 870  */
 871 static size_t inflate_chunk(struct iolog_compress *ic, int gz_hdr, FILE *f,
 872                             z_stream *stream, struct inflate_chunk_iter *iter)
 873 {
 874         size_t ret;
 875
 876         dprint(FD_COMPRESS, "inflate chunk size=%lu, seq=%u\n",
 877                                 (unsigned long) ic->len, ic->seq);
 878
 879         if (ic->seq != iter->seq) {
 880                 if (iter->seq)
 881                         finish_chunk(stream, f, iter);
 882
 883                 z_stream_init(stream, gz_hdr);
 884                 iter->seq = ic->seq;
 885         }
 886
 887         stream->avail_in = ic->len;
 888         stream->next_in = ic->buf;
 889
 890         if (!iter->buf_size) {
 891                 iter->buf_size = iter->chunk_sz;
 892                 iter->buf = malloc(iter->buf_size);
 893         }
 894
 895         while (stream->avail_in) {
 896                 size_t this_out = iter->buf_size - iter->buf_used;
 897                 int err;
 898
 899                 stream->avail_out = this_out;
 900                 stream->next_out = iter->buf + iter->buf_used;
 901
 902                 err = inflate(stream, Z_NO_FLUSH);
 903                 if (err < 0) {
 904                         log_err("fio: failed inflating log: %d\n", err);
 905                         iter->err = err;
 906                         break;
 907                 }
 908
 909                 iter->buf_used += this_out - stream->avail_out;
 910
 911                 if (!stream->avail_out) {
 912                         iter->buf_size += iter->chunk_sz;
 913                         iter->buf = realloc(iter->buf, iter->buf_size);
 914                         continue;
 915                 }
 916
 917                 if (err == Z_STREAM_END)
 918                         break;
 919         }
 920
 921         ret = (void *) stream->next_in - ic->buf;
 922
 923         dprint(FD_COMPRESS, "inflated to size=%lu\n", (unsigned long) iter->buf_size);
 924
 925         return ret;
 926 }
 927
 928 /*
 929  * Inflate stored compressed chunks, or write them directly to the log
 930  * file if so instructed.
 931  */
 932 static int inflate_gz_chunks(struct io_log *log, FILE *f)
 933 {
 934         struct inflate_chunk_iter iter = { .chunk_sz = log->log_gz, };
 935         z_stream stream;
 936
 937         while (!flist_empty(&log->chunk_list)) {
 938                 struct iolog_compress *ic;
 939
 940                 ic = flist_first_entry(&log->chunk_list, struct iolog_compress, list);
 941                 flist_del(&ic->list);
 942
 943                 if (log->log_gz_store) {
 944                         size_t ret;
 945
 946                         dprint(FD_COMPRESS, "log write chunk size=%lu, "
 947                                 "seq=%u\n", (unsigned long) ic->len, ic->seq);
 948
 949                         ret = fwrite(ic->buf, ic->len, 1, f);
 950                         if (ret != 1 || ferror(f)) {
 951                                 iter.err = errno;
 952                                 log_err("fio: error writing compressed log\n");
 953                         }
 954                 } else
 955                         inflate_chunk(ic, log->log_gz_store, f, &stream, &iter);
 956
 957                 free_chunk(ic);
 958         }
 959
 960         if (iter.seq) {
 961                 finish_chunk(&stream, f, &iter);
 962                 free(iter.buf);
 963         }
 964
 965         return iter.err;
 966 }
 967
 968 /*
 969  * Open compressed log file and decompress the stored chunks and
 970  * write them to stdout. The chunks are stored sequentially in the
 971  * file, so we iterate over them and do them one-by-one.
 972  */
 973 int iolog_file_inflate(const char *file)
 974 {
 975         struct inflate_chunk_iter iter = { .chunk_sz = 64 * 1024 * 1024, };
 976         struct iolog_compress ic;
 977         z_stream stream;
 978         struct stat sb;
 979         ssize_t ret;
 980         size_t total;
 981         void *buf;
 982         FILE *f;
 983
 984         f = fopen(file, "r");
 985         if (!f) {
 986                 perror("fopen");
 987                 return 1;
 988         }
 989
 990         if (stat(file, &sb) < 0) {
 991                 fclose(f);
 992                 perror("stat");
 993                 return 1;
 994         }
 995
 996         ic.buf = buf = malloc(sb.st_size);
 997         ic.len = sb.st_size;
 998         ic.seq = 1;
 999
1000         ret = fread(ic.buf, ic.len, 1, f);
1001         if (ret < 0) {
1002                 perror("fread");
1003                 fclose(f);
1004                 free(buf);
1005                 return 1;
1006         } else if (ret != 1) {
1007                 log_err("fio: short read on reading log\n");
1008                 fclose(f);
1009                 free(buf);
1010                 return 1;
1011         }
1012
1013         fclose(f);
1014
1015         /*
1016          * Each chunk will return Z_STREAM_END. We don't know how many
1017          * chunks are in the file, so we just keep looping and incrementing
1018          * the sequence number until we have consumed the whole compressed
1019          * file.
1020          */
1021         total = ic.len;
1022         do {
1023                 size_t iret;
1024
1025                 iret = inflate_chunk(&ic,  1, stdout, &stream, &iter);
1026                 total -= iret;
1027                 if (!total)
1028                         break;
1029                 if (iter.err)
1030                         break;
1031
1032                 ic.seq++;
1033                 ic.len -= iret;
1034                 ic.buf += iret;
1035         } while (1);
1036
1037         if (iter.seq) {
1038                 finish_chunk(&stream, stdout, &iter);
1039                 free(iter.buf);
1040         }
1041
1042         free(buf);
1043         return iter.err;
1044 }
1045
1046 #else
1047
1048 static int inflate_gz_chunks(struct io_log *log, FILE *f)
1049 {
1050         return 0;
1051 }
1052
1053 int iolog_file_inflate(const char *file)
1054 {
1055         log_err("fio: log inflation not possible without zlib\n");
1056         return 1;
1057 }
1058
1059 #endif
1060
1061 void flush_log(struct io_log *log, bool do_append)
1062 {
1063         void *buf;
1064         FILE *f;
1065
1066         if (!do_append)
1067                 f = fopen(log->filename, "w");
1068         else
1069                 f = fopen(log->filename, "a");
1070         if (!f) {
1071                 perror("fopen log");
1072                 return;
1073         }
1074
1075         buf = set_file_buffer(f);
1076
1077         inflate_gz_chunks(log, f);
1078
1079         while (!flist_empty(&log->io_logs)) {
1080                 struct io_logs *cur_log;
1081
1082                 cur_log = flist_first_entry(&log->io_logs, struct io_logs, list);
1083                 flist_del_init(&cur_log->list);
1084
1085                 if (log->td && log == log->td->clat_hist_log)
1086                         flush_hist_samples(f, log->hist_coarseness, cur_log->log,
1087                                            log_sample_sz(log, cur_log));
1088                 else
1089                         flush_samples(f, cur_log->log, log_sample_sz(log, cur_log));
1090
1091                 sfree(cur_log);
1092         }
1093
1094         fclose(f);
1095         clear_file_buffer(buf);
1096 }
1097
1098 static int finish_log(struct thread_data *td, struct io_log *log, int trylock)
1099 {
1100         if (td->flags & TD_F_COMPRESS_LOG)
1101                 iolog_flush(log);
1102
1103         if (trylock) {
1104                 if (fio_trylock_file(log->filename))
1105                         return 1;
1106         } else
1107                 fio_lock_file(log->filename);
1108
1109         if (td->client_type == FIO_CLIENT_TYPE_GUI || is_backend)
1110                 fio_send_iolog(td, log, log->filename);
1111         else
1112                 flush_log(log, !td->o.per_job_logs);
1113
1114         fio_unlock_file(log->filename);
1115         free_log(log);
1116         return 0;
1117 }
1118
1119 size_t log_chunk_sizes(struct io_log *log)
1120 {
1121         struct flist_head *entry;
1122         size_t ret;
1123
1124         if (flist_empty(&log->chunk_list))
1125                 return 0;
1126
1127         ret = 0;
1128         pthread_mutex_lock(&log->chunk_lock);
1129         flist_for_each(entry, &log->chunk_list) {
1130                 struct iolog_compress *c;
1131
1132                 c = flist_entry(entry, struct iolog_compress, list);
1133                 ret += c->len;
1134         }
1135         pthread_mutex_unlock(&log->chunk_lock);
1136         return ret;
1137 }
1138
1139 #ifdef CONFIG_ZLIB
1140
1141 static int gz_work(struct iolog_flush_data *data)
1142 {
1143         struct iolog_compress *c = NULL;
1144         struct flist_head list;
1145         unsigned int seq;
1146         z_stream stream;
1147         size_t total = 0;
1148         int ret;
1149
1150         INIT_FLIST_HEAD(&list);
1151
1152         memset(&stream, 0, sizeof(stream));
1153         stream.zalloc = Z_NULL;
1154         stream.zfree = Z_NULL;
1155         stream.opaque = Z_NULL;
1156
1157         ret = deflateInit(&stream, Z_DEFAULT_COMPRESSION);
1158         if (ret != Z_OK) {
1159                 log_err("fio: failed to init gz stream\n");
1160                 goto err;
1161         }
1162
1163         seq = ++data->log->chunk_seq;
1164
1165         stream.next_in = (void *) data->samples;
1166         stream.avail_in = data->nr_samples * log_entry_sz(data->log);
1167
1168         dprint(FD_COMPRESS, "deflate input size=%lu, seq=%u, log=%s\n",
1169                                 (unsigned long) stream.avail_in, seq,
1170                                 data->log->filename);
1171         do {
1172                 if (c)
1173                         dprint(FD_COMPRESS, "seq=%d, chunk=%lu\n", seq,
1174                                 (unsigned long) c->len);
1175                 c = get_new_chunk(seq);
1176                 stream.avail_out = GZ_CHUNK;
1177                 stream.next_out = c->buf;
1178                 ret = deflate(&stream, Z_NO_FLUSH);
1179                 if (ret < 0) {
1180                         log_err("fio: deflate log (%d)\n", ret);
1181                         free_chunk(c);
1182                         goto err;
1183                 }
1184
1185                 c->len = GZ_CHUNK - stream.avail_out;
1186                 flist_add_tail(&c->list, &list);
1187                 total += c->len;
1188         } while (stream.avail_in);
1189
1190         stream.next_out = c->buf + c->len;
1191         stream.avail_out = GZ_CHUNK - c->len;
1192
1193         ret = deflate(&stream, Z_FINISH);
1194         if (ret < 0) {
1195                 /*
1196                  * Z_BUF_ERROR is special, it just means we need more
1197                  * output space. We'll handle that below. Treat any other
1198                  * error as fatal.
1199                  */
1200                 if (ret != Z_BUF_ERROR) {
1201                         log_err("fio: deflate log (%d)\n", ret);
1202                         flist_del(&c->list);
1203                         free_chunk(c);
1204                         goto err;
1205                 }
1206         }
1207
1208         total -= c->len;
1209         c->len = GZ_CHUNK - stream.avail_out;
1210         total += c->len;
1211         dprint(FD_COMPRESS, "seq=%d, chunk=%lu\n", seq, (unsigned long) c->len);
1212
1213         if (ret != Z_STREAM_END) {
1214                 do {
1215                         c = get_new_chunk(seq);
1216                         stream.avail_out = GZ_CHUNK;
1217                         stream.next_out = c->buf;
1218                         ret = deflate(&stream, Z_FINISH);
1219                         c->len = GZ_CHUNK - stream.avail_out;
1220                         total += c->len;
1221                         flist_add_tail(&c->list, &list);
1222                         dprint(FD_COMPRESS, "seq=%d, chunk=%lu\n", seq,
1223                                 (unsigned long) c->len);
1224                 } while (ret != Z_STREAM_END);
1225         }
1226
1227         dprint(FD_COMPRESS, "deflated to size=%lu\n", (unsigned long) total);
1228
1229         ret = deflateEnd(&stream);
1230         if (ret != Z_OK)
1231                 log_err("fio: deflateEnd %d\n", ret);
1232
1233         free(data->samples);
1234
1235         if (!flist_empty(&list)) {
1236                 pthread_mutex_lock(&data->log->chunk_lock);
1237                 flist_splice_tail(&list, &data->log->chunk_list);
1238                 pthread_mutex_unlock(&data->log->chunk_lock);
1239         }
1240
1241         ret = 0;
1242 done:
1243         if (data->free)
1244                 free(data);
1245         return ret;
1246 err:
1247         while (!flist_empty(&list)) {
1248                 c = flist_first_entry(list.next, struct iolog_compress, list);
1249                 flist_del(&c->list);
1250                 free_chunk(c);
1251         }
1252         ret = 1;
1253         goto done;
1254 }
1255
1256 /*
1257  * Invoked from our compress helper thread, when logging would have exceeded
1258  * the specified memory limitation. Compresses the previously stored
1259  * entries.
1260  */
1261 static int gz_work_async(struct submit_worker *sw, struct workqueue_work *work)
1262 {
1263         return gz_work(container_of(work, struct iolog_flush_data, work));
1264 }
1265
1266 static int gz_init_worker(struct submit_worker *sw)
1267 {
1268         struct thread_data *td = sw->wq->td;
1269
1270         if (!fio_option_is_set(&td->o, log_gz_cpumask))
1271                 return 0;
1272
1273         if (fio_setaffinity(gettid(), td->o.log_gz_cpumask) == -1) {
1274                 log_err("gz: failed to set CPU affinity\n");
1275                 return 1;
1276         }
1277
1278         return 0;
1279 }
1280
1281 static struct workqueue_ops log_compress_wq_ops = {
1282         .fn             = gz_work_async,
1283         .init_worker_fn = gz_init_worker,
1284         .nice           = 1,
1285 };
1286
1287 int iolog_compress_init(struct thread_data *td, struct sk_out *sk_out)
1288 {
1289         if (!(td->flags & TD_F_COMPRESS_LOG))
1290                 return 0;
1291
1292         workqueue_init(td, &td->log_compress_wq, &log_compress_wq_ops, 1, sk_out);
1293         return 0;
1294 }
1295
1296 void iolog_compress_exit(struct thread_data *td)
1297 {
1298         if (!(td->flags & TD_F_COMPRESS_LOG))
1299                 return;
1300
1301         workqueue_exit(&td->log_compress_wq);
1302 }
1303
1304 /*
1305  * Queue work item to compress the existing log entries. We reset the
1306  * current log to a small size, and reference the existing log in the
1307  * data that we queue for compression. Once compression has been done,
1308  * this old log is freed. If called with finish == true, will not return
1309  * until the log compression has completed, and will flush all previous
1310  * logs too
1311  */
1312 static int iolog_flush(struct io_log *log)
1313 {
1314         struct iolog_flush_data *data;
1315
1316         data = malloc(sizeof(*data));
1317         if (!data)
1318                 return 1;
1319
1320         data->log = log;
1321         data->free = false;
1322
1323         while (!flist_empty(&log->io_logs)) {
1324                 struct io_logs *cur_log;
1325
1326                 cur_log = flist_first_entry(&log->io_logs, struct io_logs, list);
1327                 flist_del_init(&cur_log->list);
1328
1329                 data->samples = cur_log->log;
1330                 data->nr_samples = cur_log->nr_samples;
1331
1332                 sfree(cur_log);
1333
1334                 gz_work(data);
1335         }
1336
1337         free(data);
1338         return 0;
1339 }
1340
1341 int iolog_cur_flush(struct io_log *log, struct io_logs *cur_log)
1342 {
1343         struct iolog_flush_data *data;
1344
1345         data = malloc(sizeof(*data));
1346         if (!data)
1347                 return 1;
1348
1349         data->log = log;
1350
1351         data->samples = cur_log->log;
1352         data->nr_samples = cur_log->nr_samples;
1353         data->free = true;
1354
1355         cur_log->nr_samples = cur_log->max_samples = 0;
1356         cur_log->log = NULL;
1357
1358         workqueue_enqueue(&log->td->log_compress_wq, &data->work);
1359         return 0;
1360 }
1361 #else
1362
1363 static int iolog_flush(struct io_log *log)
1364 {
1365         return 1;
1366 }
1367
1368 int iolog_cur_flush(struct io_log *log, struct io_logs *cur_log)
1369 {
1370         return 1;
1371 }
1372
1373 int iolog_compress_init(struct thread_data *td, struct sk_out *sk_out)
1374 {
1375         return 0;
1376 }
1377
1378 void iolog_compress_exit(struct thread_data *td)
1379 {
1380 }
1381
1382 #endif
1383
1384 struct io_logs *iolog_cur_log(struct io_log *log)
1385 {
1386         if (flist_empty(&log->io_logs))
1387                 return NULL;
1388
1389         return flist_last_entry(&log->io_logs, struct io_logs, list);
1390 }
1391
1392 uint64_t iolog_nr_samples(struct io_log *iolog)
1393 {
1394         struct flist_head *entry;
1395         uint64_t ret = 0;
1396
1397         flist_for_each(entry, &iolog->io_logs) {
1398                 struct io_logs *cur_log;
1399
1400                 cur_log = flist_entry(entry, struct io_logs, list);
1401                 ret += cur_log->nr_samples;
1402         }
1403
1404         return ret;
1405 }
1406
1407 static int __write_log(struct thread_data *td, struct io_log *log, int try)
1408 {
1409         if (log)
1410                 return finish_log(td, log, try);
1411
1412         return 0;
1413 }
1414
1415 static int write_iops_log(struct thread_data *td, int try, bool unit_log)
1416 {
1417         int ret;
1418
1419         if (per_unit_log(td->iops_log) != unit_log)
1420                 return 0;
1421
1422         ret = __write_log(td, td->iops_log, try);
1423         if (!ret)
1424                 td->iops_log = NULL;
1425
1426         return ret;
1427 }
1428
1429 static int write_slat_log(struct thread_data *td, int try, bool unit_log)
1430 {
1431         int ret;
1432
1433         if (!unit_log)
1434                 return 0;
1435
1436         ret = __write_log(td, td->slat_log, try);
1437         if (!ret)
1438                 td->slat_log = NULL;
1439
1440         return ret;
1441 }
1442
1443 static int write_clat_log(struct thread_data *td, int try, bool unit_log)
1444 {
1445         int ret;
1446
1447         if (!unit_log)
1448                 return 0;
1449
1450         ret = __write_log(td, td->clat_log, try);
1451         if (!ret)
1452                 td->clat_log = NULL;
1453
1454         return ret;
1455 }
1456
1457 static int write_clat_hist_log(struct thread_data *td, int try, bool unit_log)
1458 {
1459         int ret;
1460
1461         if (!unit_log)
1462                 return 0;
1463
1464         ret = __write_log(td, td->clat_hist_log, try);
1465         if (!ret)
1466                 td->clat_hist_log = NULL;
1467
1468         return ret;
1469 }
1470
1471 static int write_lat_log(struct thread_data *td, int try, bool unit_log)
1472 {
1473         int ret;
1474
1475         if (!unit_log)
1476                 return 0;
1477
1478         ret = __write_log(td, td->lat_log, try);
1479         if (!ret)
1480                 td->lat_log = NULL;
1481
1482         return ret;
1483 }
1484
1485 static int write_bandw_log(struct thread_data *td, int try, bool unit_log)
1486 {
1487         int ret;
1488
1489         if (per_unit_log(td->bw_log) != unit_log)
1490                 return 0;
1491
1492         ret = __write_log(td, td->bw_log, try);
1493         if (!ret)
1494                 td->bw_log = NULL;
1495
1496         return ret;
1497 }
1498
1499 enum {
1500         BW_LOG_MASK     = 1,
1501         LAT_LOG_MASK    = 2,
1502         SLAT_LOG_MASK   = 4,
1503         CLAT_LOG_MASK   = 8,
1504         IOPS_LOG_MASK   = 16,
1505         CLAT_HIST_LOG_MASK = 32,
1506
1507         ALL_LOG_NR      = 6,
1508 };
1509
1510 struct log_type {
1511         unsigned int mask;
1512         int (*fn)(struct thread_data *, int, bool);
1513 };
1514
1515 static struct log_type log_types[] = {
1516         {
1517                 .mask   = BW_LOG_MASK,
1518                 .fn     = write_bandw_log,
1519         },
1520         {
1521                 .mask   = LAT_LOG_MASK,
1522                 .fn     = write_lat_log,
1523         },
1524         {
1525                 .mask   = SLAT_LOG_MASK,
1526                 .fn     = write_slat_log,
1527         },
1528         {
1529                 .mask   = CLAT_LOG_MASK,
1530                 .fn     = write_clat_log,
1531         },
1532         {
1533                 .mask   = IOPS_LOG_MASK,
1534                 .fn     = write_iops_log,
1535         },
1536         {
1537                 .mask   = CLAT_HIST_LOG_MASK,
1538                 .fn     = write_clat_hist_log,
1539         }
1540 };
1541
1542 void td_writeout_logs(struct thread_data *td, bool unit_logs)
1543 {
1544         unsigned int log_mask = 0;
1545         unsigned int log_left = ALL_LOG_NR;
1546         int old_state, i;
1547
1548         old_state = td_bump_runstate(td, TD_FINISHING);
1549
1550         finalize_logs(td, unit_logs);
1551
1552         while (log_left) {
1553                 int prev_log_left = log_left;
1554
1555                 for (i = 0; i < ALL_LOG_NR && log_left; i++) {
1556                         struct log_type *lt = &log_types[i];
1557                         int ret;
1558
1559                         if (!(log_mask & lt->mask)) {
1560                                 ret = lt->fn(td, log_left != 1, unit_logs);
1561                                 if (!ret) {
1562                                         log_left--;
1563                                         log_mask |= lt->mask;
1564                                 }
1565                         }
1566                 }
1567
1568                 if (prev_log_left == log_left)
1569                         usleep(5000);
1570         }
1571
1572         td_restore_runstate(td, old_state);
1573 }
1574
1575 void fio_writeout_logs(bool unit_logs)
1576 {
1577         struct thread_data *td;
1578         int i;
1579
1580         for_each_td(td, i)
1581                 td_writeout_logs(td, unit_logs);
1582 }