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 #include "blktrace.h"
  23
  24 static int iolog_flush(struct io_log *log);
  25
  26 static const char iolog_ver2[] = "fio version 2 iolog";
  27
  28 void queue_io_piece(struct thread_data *td, struct io_piece *ipo)
  29 {
  30         flist_add_tail(&ipo->list, &td->io_log_list);
  31         td->total_io_size += ipo->len;
  32 }
  33
  34 void log_io_u(const struct thread_data *td, const struct io_u *io_u)
  35 {
  36         if (!td->o.write_iolog_file)
  37                 return;
  38
  39         fprintf(td->iolog_f, "%s %s %llu %lu\n", io_u->file->file_name,
  40                                                 io_ddir_name(io_u->ddir),
  41                                                 io_u->offset, io_u->buflen);
  42 }
  43
  44 void log_file(struct thread_data *td, struct fio_file *f,
  45               enum file_log_act what)
  46 {
  47         const char *act[] = { "add", "open", "close" };
  48
  49         assert(what < 3);
  50
  51         if (!td->o.write_iolog_file)
  52                 return;
  53
  54
  55         /*
  56          * this happens on the pre-open/close done before the job starts
  57          */
  58         if (!td->iolog_f)
  59                 return;
  60
  61         fprintf(td->iolog_f, "%s %s\n", f->file_name, act[what]);
  62 }
  63
  64 static void iolog_delay(struct thread_data *td, unsigned long delay)
  65 {
  66         uint64_t usec = utime_since_now(&td->last_issue);
  67         uint64_t this_delay;
  68         struct timespec ts;
  69
  70         if (delay < td->time_offset) {
  71                 td->time_offset = 0;
  72                 return;
  73         }
  74
  75         delay -= td->time_offset;
  76         if (delay < usec)
  77                 return;
  78
  79         delay -= usec;
  80
  81         fio_gettime(&ts, NULL);
  82         while (delay && !td->terminate) {
  83                 this_delay = delay;
  84                 if (this_delay > 500000)
  85                         this_delay = 500000;
  86
  87                 usec_sleep(td, this_delay);
  88                 delay -= this_delay;
  89         }
  90
  91         usec = utime_since_now(&ts);
  92         if (usec > delay)
  93                 td->time_offset = usec - delay;
  94         else
  95                 td->time_offset = 0;
  96 }
  97
  98 static int ipo_special(struct thread_data *td, struct io_piece *ipo)
  99 {
 100         struct fio_file *f;
 101         int ret;
 102
 103         /*
 104          * Not a special ipo
 105          */
 106         if (ipo->ddir != DDIR_INVAL)
 107                 return 0;
 108
 109         f = td->files[ipo->fileno];
 110
 111         switch (ipo->file_action) {
 112         case FIO_LOG_OPEN_FILE:
 113                 if (td->o.replay_redirect && fio_file_open(f)) {
 114                         dprint(FD_FILE, "iolog: ignoring re-open of file %s\n",
 115                                         f->file_name);
 116                         break;
 117                 }
 118                 ret = td_io_open_file(td, f);
 119                 if (!ret)
 120                         break;
 121                 td_verror(td, ret, "iolog open file");
 122                 return -1;
 123         case FIO_LOG_CLOSE_FILE:
 124                 td_io_close_file(td, f);
 125                 break;
 126         case FIO_LOG_UNLINK_FILE:
 127                 td_io_unlink_file(td, f);
 128                 break;
 129         default:
 130                 log_err("fio: bad file action %d\n", ipo->file_action);
 131                 break;
 132         }
 133
 134         return 1;
 135 }
 136
 137 int read_iolog_get(struct thread_data *td, struct io_u *io_u)
 138 {
 139         struct io_piece *ipo;
 140         unsigned long elapsed;
 141
 142         while (!flist_empty(&td->io_log_list)) {
 143                 int ret;
 144
 145                 ipo = flist_first_entry(&td->io_log_list, struct io_piece, list);
 146                 flist_del(&ipo->list);
 147                 remove_trim_entry(td, ipo);
 148
 149                 ret = ipo_special(td, ipo);
 150                 if (ret < 0) {
 151                         free(ipo);
 152                         break;
 153                 } else if (ret > 0) {
 154                         free(ipo);
 155                         continue;
 156                 }
 157
 158                 io_u->ddir = ipo->ddir;
 159                 if (ipo->ddir != DDIR_WAIT) {
 160                         io_u->offset = ipo->offset;
 161                         io_u->buflen = ipo->len;
 162                         io_u->file = td->files[ipo->fileno];
 163                         get_file(io_u->file);
 164                         dprint(FD_IO, "iolog: get %llu/%lu/%s\n", io_u->offset,
 165                                                 io_u->buflen, io_u->file->file_name);
 166                         if (ipo->delay)
 167                                 iolog_delay(td, ipo->delay);
 168                 } else {
 169                         elapsed = mtime_since_genesis();
 170                         if (ipo->delay > elapsed)
 171                                 usec_sleep(td, (ipo->delay - elapsed) * 1000);
 172                 }
 173
 174                 free(ipo);
 175
 176                 if (io_u->ddir != DDIR_WAIT)
 177                         return 0;
 178         }
 179
 180         td->done = 1;
 181         return 1;
 182 }
 183
 184 void prune_io_piece_log(struct thread_data *td)
 185 {
 186         struct io_piece *ipo;
 187         struct rb_node *n;
 188
 189         while ((n = rb_first(&td->io_hist_tree)) != NULL) {
 190                 ipo = rb_entry(n, struct io_piece, rb_node);
 191                 rb_erase(n, &td->io_hist_tree);
 192                 remove_trim_entry(td, ipo);
 193                 td->io_hist_len--;
 194                 free(ipo);
 195         }
 196
 197         while (!flist_empty(&td->io_hist_list)) {
 198                 ipo = flist_first_entry(&td->io_hist_list, struct io_piece, list);
 199                 flist_del(&ipo->list);
 200                 remove_trim_entry(td, ipo);
 201                 td->io_hist_len--;
 202                 free(ipo);
 203         }
 204 }
 205
 206 /*
 207  * log a successful write, so we can unwind the log for verify
 208  */
 209 void log_io_piece(struct thread_data *td, struct io_u *io_u)
 210 {
 211         struct rb_node **p, *parent;
 212         struct io_piece *ipo, *__ipo;
 213
 214         ipo = malloc(sizeof(struct io_piece));
 215         init_ipo(ipo);
 216         ipo->file = io_u->file;
 217         ipo->offset = io_u->offset;
 218         ipo->len = io_u->buflen;
 219         ipo->numberio = io_u->numberio;
 220         ipo->flags = IP_F_IN_FLIGHT;
 221
 222         io_u->ipo = ipo;
 223
 224         if (io_u_should_trim(td, io_u)) {
 225                 flist_add_tail(&ipo->trim_list, &td->trim_list);
 226                 td->trim_entries++;
 227         }
 228
 229         /*
 230          * We don't need to sort the entries if we only performed sequential
 231          * writes. In this case, just reading back data in the order we wrote
 232          * it out is the faster but still safe.
 233          *
 234          * One exception is if we don't have a random map in which case we need
 235          * to check for duplicate blocks and drop the old one, which we rely on
 236          * the rb insert/lookup for handling.
 237          */
 238         if (((!td->o.verifysort) || !td_random(td)) &&
 239               file_randommap(td, ipo->file)) {
 240                 INIT_FLIST_HEAD(&ipo->list);
 241                 flist_add_tail(&ipo->list, &td->io_hist_list);
 242                 ipo->flags |= IP_F_ONLIST;
 243                 td->io_hist_len++;
 244                 return;
 245         }
 246
 247         RB_CLEAR_NODE(&ipo->rb_node);
 248
 249         /*
 250          * Sort the entry into the verification list
 251          */
 252 restart:
 253         p = &td->io_hist_tree.rb_node;
 254         parent = NULL;
 255         while (*p) {
 256                 int overlap = 0;
 257                 parent = *p;
 258
 259                 __ipo = rb_entry(parent, struct io_piece, rb_node);
 260                 if (ipo->file < __ipo->file)
 261                         p = &(*p)->rb_left;
 262                 else if (ipo->file > __ipo->file)
 263                         p = &(*p)->rb_right;
 264                 else if (ipo->offset < __ipo->offset) {
 265                         p = &(*p)->rb_left;
 266                         overlap = ipo->offset + ipo->len > __ipo->offset;
 267                 }
 268                 else if (ipo->offset > __ipo->offset) {
 269                         p = &(*p)->rb_right;
 270                         overlap = __ipo->offset + __ipo->len > ipo->offset;
 271                 }
 272                 else
 273                         overlap = 1;
 274
 275                 if (overlap) {
 276                         dprint(FD_IO, "iolog: overlap %llu/%lu, %llu/%lu\n",
 277                                 __ipo->offset, __ipo->len,
 278                                 ipo->offset, ipo->len);
 279                         td->io_hist_len--;
 280                         rb_erase(parent, &td->io_hist_tree);
 281                         remove_trim_entry(td, __ipo);
 282                         if (!(__ipo->flags & IP_F_IN_FLIGHT))
 283                                 free(__ipo);
 284                         goto restart;
 285                 }
 286         }
 287
 288         rb_link_node(&ipo->rb_node, parent, p);
 289         rb_insert_color(&ipo->rb_node, &td->io_hist_tree);
 290         ipo->flags |= IP_F_ONRB;
 291         td->io_hist_len++;
 292 }
 293
 294 void unlog_io_piece(struct thread_data *td, struct io_u *io_u)
 295 {
 296         struct io_piece *ipo = io_u->ipo;
 297
 298         if (td->ts.nr_block_infos) {
 299                 uint32_t *info = io_u_block_info(td, io_u);
 300                 if (BLOCK_INFO_STATE(*info) < BLOCK_STATE_TRIM_FAILURE) {
 301                         if (io_u->ddir == DDIR_TRIM)
 302                                 *info = BLOCK_INFO_SET_STATE(*info,
 303                                                 BLOCK_STATE_TRIM_FAILURE);
 304                         else if (io_u->ddir == DDIR_WRITE)
 305                                 *info = BLOCK_INFO_SET_STATE(*info,
 306                                                 BLOCK_STATE_WRITE_FAILURE);
 307                 }
 308         }
 309
 310         if (!ipo)
 311                 return;
 312
 313         if (ipo->flags & IP_F_ONRB)
 314                 rb_erase(&ipo->rb_node, &td->io_hist_tree);
 315         else if (ipo->flags & IP_F_ONLIST)
 316                 flist_del(&ipo->list);
 317
 318         free(ipo);
 319         io_u->ipo = NULL;
 320         td->io_hist_len--;
 321 }
 322
 323 void trim_io_piece(struct thread_data *td, const struct io_u *io_u)
 324 {
 325         struct io_piece *ipo = io_u->ipo;
 326
 327         if (!ipo)
 328                 return;
 329
 330         ipo->len = io_u->xfer_buflen - io_u->resid;
 331 }
 332
 333 void write_iolog_close(struct thread_data *td)
 334 {
 335         fflush(td->iolog_f);
 336         fclose(td->iolog_f);
 337         free(td->iolog_buf);
 338         td->iolog_f = NULL;
 339         td->iolog_buf = NULL;
 340 }
 341
 342 /*
 343  * Read version 2 iolog data. It is enhanced to include per-file logging,
 344  * syncs, etc.
 345  */
 346 static int read_iolog2(struct thread_data *td, FILE *f)
 347 {
 348         unsigned long long offset;
 349         unsigned int bytes;
 350         int reads, writes, waits, fileno = 0, file_action = 0; /* stupid gcc */
 351         char *rfname, *fname, *act;
 352         char *str, *p;
 353         enum fio_ddir rw;
 354
 355         free_release_files(td);
 356
 357         /*
 358          * Read in the read iolog and store it, reuse the infrastructure
 359          * for doing verifications.
 360          */
 361         str = malloc(4096);
 362         rfname = fname = malloc(256+16);
 363         act = malloc(256+16);
 364
 365         reads = writes = waits = 0;
 366         while ((p = fgets(str, 4096, f)) != NULL) {
 367                 struct io_piece *ipo;
 368                 int r;
 369
 370                 r = sscanf(p, "%256s %256s %llu %u", rfname, act, &offset,
 371                                                                         &bytes);
 372
 373                 if (td->o.replay_redirect)
 374                         fname = td->o.replay_redirect;
 375
 376                 if (r == 4) {
 377                         /*
 378                          * Check action first
 379                          */
 380                         if (!strcmp(act, "wait"))
 381                                 rw = DDIR_WAIT;
 382                         else if (!strcmp(act, "read"))
 383                                 rw = DDIR_READ;
 384                         else if (!strcmp(act, "write"))
 385                                 rw = DDIR_WRITE;
 386                         else if (!strcmp(act, "sync"))
 387                                 rw = DDIR_SYNC;
 388                         else if (!strcmp(act, "datasync"))
 389                                 rw = DDIR_DATASYNC;
 390                         else if (!strcmp(act, "trim"))
 391                                 rw = DDIR_TRIM;
 392                         else {
 393                                 log_err("fio: bad iolog file action: %s\n",
 394                                                                         act);
 395                                 continue;
 396                         }
 397                         fileno = get_fileno(td, fname);
 398                 } else if (r == 2) {
 399                         rw = DDIR_INVAL;
 400                         if (!strcmp(act, "add")) {
 401                                 if (td->o.replay_redirect &&
 402                                     get_fileno(td, fname) != -1) {
 403                                         dprint(FD_FILE, "iolog: ignoring"
 404                                                 " re-add of file %s\n", fname);
 405                                 } else {
 406                                         fileno = add_file(td, fname, 0, 1);
 407                                         file_action = FIO_LOG_ADD_FILE;
 408                                 }
 409                                 continue;
 410                         } else if (!strcmp(act, "open")) {
 411                                 fileno = get_fileno(td, fname);
 412                                 file_action = FIO_LOG_OPEN_FILE;
 413                         } else if (!strcmp(act, "close")) {
 414                                 fileno = get_fileno(td, fname);
 415                                 file_action = FIO_LOG_CLOSE_FILE;
 416                         } else {
 417                                 log_err("fio: bad iolog file action: %s\n",
 418                                                                         act);
 419                                 continue;
 420                         }
 421                 } else {
 422                         log_err("bad iolog2: %s\n", p);
 423                         continue;
 424                 }
 425
 426                 if (rw == DDIR_READ)
 427                         reads++;
 428                 else if (rw == DDIR_WRITE) {
 429                         /*
 430                          * Don't add a write for ro mode
 431                          */
 432                         if (read_only)
 433                                 continue;
 434                         writes++;
 435                 } else if (rw == DDIR_WAIT) {
 436                         if (td->o.no_stall)
 437                                 continue;
 438                         waits++;
 439                 } else if (rw == DDIR_INVAL) {
 440                 } else if (!ddir_sync(rw)) {
 441                         log_err("bad ddir: %d\n", rw);
 442                         continue;
 443                 }
 444
 445                 /*
 446                  * Make note of file
 447                  */
 448                 ipo = malloc(sizeof(*ipo));
 449                 init_ipo(ipo);
 450                 ipo->ddir = rw;
 451                 if (rw == DDIR_WAIT) {
 452                         ipo->delay = offset;
 453                 } else {
 454                         if (td->o.replay_scale)
 455                                 ipo->offset = offset / td->o.replay_scale;
 456                         else
 457                                 ipo->offset = offset;
 458                         ipo_bytes_align(td->o.replay_align, ipo);
 459
 460                         ipo->len = bytes;
 461                         if (rw != DDIR_INVAL && bytes > td->o.max_bs[rw])
 462                                 td->o.max_bs[rw] = bytes;
 463                         ipo->fileno = fileno;
 464                         ipo->file_action = file_action;
 465                         td->o.size += bytes;
 466                 }
 467
 468                 queue_io_piece(td, ipo);
 469         }
 470
 471         free(str);
 472         free(act);
 473         free(rfname);
 474
 475         if (writes && read_only) {
 476                 log_err("fio: <%s> skips replay of %d writes due to"
 477                         " read-only\n", td->o.name, writes);
 478                 writes = 0;
 479         }
 480
 481         if (!reads && !writes && !waits)
 482                 return 1;
 483         else if (reads && !writes)
 484                 td->o.td_ddir = TD_DDIR_READ;
 485         else if (!reads && writes)
 486                 td->o.td_ddir = TD_DDIR_WRITE;
 487         else
 488                 td->o.td_ddir = TD_DDIR_RW;
 489
 490         return 0;
 491 }
 492
 493 /*
 494  * open iolog, check version, and call appropriate parser
 495  */
 496 static int init_iolog_read(struct thread_data *td)
 497 {
 498         char buffer[256], *p;
 499         FILE *f;
 500         int ret;
 501
 502         f = fopen(td->o.read_iolog_file, "r");
 503         if (!f) {
 504                 perror("fopen read iolog");
 505                 return 1;
 506         }
 507
 508         p = fgets(buffer, sizeof(buffer), f);
 509         if (!p) {
 510                 td_verror(td, errno, "iolog read");
 511                 log_err("fio: unable to read iolog\n");
 512                 fclose(f);
 513                 return 1;
 514         }
 515
 516         /*
 517          * version 2 of the iolog stores a specific string as the
 518          * first line, check for that
 519          */
 520         if (!strncmp(iolog_ver2, buffer, strlen(iolog_ver2)))
 521                 ret = read_iolog2(td, f);
 522         else {
 523                 log_err("fio: iolog version 1 is no longer supported\n");
 524                 ret = 1;
 525         }
 526
 527         fclose(f);
 528         return ret;
 529 }
 530
 531 /*
 532  * Set up a log for storing io patterns.
 533  */
 534 static int init_iolog_write(struct thread_data *td)
 535 {
 536         struct fio_file *ff;
 537         FILE *f;
 538         unsigned int i;
 539
 540         f = fopen(td->o.write_iolog_file, "a");
 541         if (!f) {
 542                 perror("fopen write iolog");
 543                 return 1;
 544         }
 545
 546         /*
 547          * That's it for writing, setup a log buffer and we're done.
 548           */
 549         td->iolog_f = f;
 550         td->iolog_buf = malloc(8192);
 551         setvbuf(f, td->iolog_buf, _IOFBF, 8192);
 552
 553         /*
 554          * write our version line
 555          */
 556         if (fprintf(f, "%s\n", iolog_ver2) < 0) {
 557                 perror("iolog init\n");
 558                 return 1;
 559         }
 560
 561         /*
 562          * add all known files
 563          */
 564         for_each_file(td, ff, i)
 565                 log_file(td, ff, FIO_LOG_ADD_FILE);
 566
 567         return 0;
 568 }
 569
 570 int init_iolog(struct thread_data *td)
 571 {
 572         int ret = 0;
 573
 574         if (td->o.read_iolog_file) {
 575                 int need_swap;
 576
 577                 /*
 578                  * Check if it's a blktrace file and load that if possible.
 579                  * Otherwise assume it's a normal log file and load that.
 580                  */
 581                 if (is_blktrace(td->o.read_iolog_file, &need_swap))
 582                         ret = load_blktrace(td, td->o.read_iolog_file, need_swap);
 583                 else
 584                         ret = init_iolog_read(td);
 585         } else if (td->o.write_iolog_file)
 586                 ret = init_iolog_write(td);
 587
 588         if (ret)
 589                 td_verror(td, EINVAL, "failed initializing iolog");
 590
 591         init_disk_util(td);
 592
 593         return ret;
 594 }
 595
 596 void setup_log(struct io_log **log, struct log_params *p,
 597                const char *filename)
 598 {
 599         struct io_log *l;
 600         int i;
 601         struct io_u_plat_entry *entry;
 602         struct flist_head *list;
 603
 604         l = scalloc(1, sizeof(*l));
 605         INIT_FLIST_HEAD(&l->io_logs);
 606         l->log_type = p->log_type;
 607         l->log_offset = p->log_offset;
 608         l->log_gz = p->log_gz;
 609         l->log_gz_store = p->log_gz_store;
 610         l->avg_msec = p->avg_msec;
 611         l->hist_msec = p->hist_msec;
 612         l->hist_coarseness = p->hist_coarseness;
 613         l->filename = strdup(filename);
 614         l->td = p->td;
 615
 616         /* Initialize histogram lists for each r/w direction,
 617          * with initial io_u_plat of all zeros:
 618          */
 619         for (i = 0; i < DDIR_RWDIR_CNT; i++) {
 620                 list = &l->hist_window[i].list;
 621                 INIT_FLIST_HEAD(list);
 622                 entry = calloc(1, sizeof(struct io_u_plat_entry));
 623                 flist_add(&entry->list, list);
 624         }
 625
 626         if (l->td && l->td->o.io_submit_mode != IO_MODE_OFFLOAD) {
 627                 struct io_logs *p;
 628
 629                 p = calloc(1, sizeof(*l->pending));
 630                 p->max_samples = DEF_LOG_ENTRIES;
 631                 p->log = calloc(p->max_samples, log_entry_sz(l));
 632                 l->pending = p;
 633         }
 634
 635         if (l->log_offset)
 636                 l->log_ddir_mask = LOG_OFFSET_SAMPLE_BIT;
 637
 638         INIT_FLIST_HEAD(&l->chunk_list);
 639
 640         if (l->log_gz && !p->td)
 641                 l->log_gz = 0;
 642         else if (l->log_gz || l->log_gz_store) {
 643                 mutex_init_pshared(&l->chunk_lock);
 644                 mutex_init_pshared(&l->deferred_free_lock);
 645                 p->td->flags |= TD_F_COMPRESS_LOG;
 646         }
 647
 648         *log = l;
 649 }
 650
 651 #ifdef CONFIG_SETVBUF
 652 static void *set_file_buffer(FILE *f)
 653 {
 654         size_t size = 1048576;
 655         void *buf;
 656
 657         buf = malloc(size);
 658         setvbuf(f, buf, _IOFBF, size);
 659         return buf;
 660 }
 661
 662 static void clear_file_buffer(void *buf)
 663 {
 664         free(buf);
 665 }
 666 #else
 667 static void *set_file_buffer(FILE *f)
 668 {
 669         return NULL;
 670 }
 671
 672 static void clear_file_buffer(void *buf)
 673 {
 674 }
 675 #endif
 676
 677 void free_log(struct io_log *log)
 678 {
 679         while (!flist_empty(&log->io_logs)) {
 680                 struct io_logs *cur_log;
 681
 682                 cur_log = flist_first_entry(&log->io_logs, struct io_logs, list);
 683                 flist_del_init(&cur_log->list);
 684                 free(cur_log->log);
 685                 sfree(cur_log);
 686         }
 687
 688         if (log->pending) {
 689                 free(log->pending->log);
 690                 free(log->pending);
 691                 log->pending = NULL;
 692         }
 693
 694         free(log->pending);
 695         free(log->filename);
 696         sfree(log);
 697 }
 698
 699 unsigned long hist_sum(int j, int stride, unsigned int *io_u_plat,
 700                 unsigned int *io_u_plat_last)
 701 {
 702         unsigned long sum;
 703         int k;
 704
 705         if (io_u_plat_last) {
 706                 for (k = sum = 0; k < stride; k++)
 707                         sum += io_u_plat[j + k] - io_u_plat_last[j + k];
 708         } else {
 709                 for (k = sum = 0; k < stride; k++)
 710                         sum += io_u_plat[j + k];
 711         }
 712
 713         return sum;
 714 }
 715
 716 static void flush_hist_samples(FILE *f, int hist_coarseness, void *samples,
 717                                uint64_t sample_size)
 718 {
 719         struct io_sample *s;
 720         int log_offset;
 721         uint64_t i, j, nr_samples;
 722         struct io_u_plat_entry *entry, *entry_before;
 723         unsigned int *io_u_plat;
 724         unsigned int *io_u_plat_before;
 725
 726         int stride = 1 << hist_coarseness;
 727
 728         if (!sample_size)
 729                 return;
 730
 731         s = __get_sample(samples, 0, 0);
 732         log_offset = (s->__ddir & LOG_OFFSET_SAMPLE_BIT) != 0;
 733
 734         nr_samples = sample_size / __log_entry_sz(log_offset);
 735
 736         for (i = 0; i < nr_samples; i++) {
 737                 s = __get_sample(samples, log_offset, i);
 738
 739                 entry = s->data.plat_entry;
 740                 io_u_plat = entry->io_u_plat;
 741
 742                 entry_before = flist_first_entry(&entry->list, struct io_u_plat_entry, list);
 743                 io_u_plat_before = entry_before->io_u_plat;
 744
 745                 fprintf(f, "%lu, %u, %u, ", (unsigned long) s->time,
 746                                                 io_sample_ddir(s), s->bs);
 747                 for (j = 0; j < FIO_IO_U_PLAT_NR - stride; j += stride) {
 748                         fprintf(f, "%lu, ", hist_sum(j, stride, io_u_plat,
 749                                                 io_u_plat_before));
 750                 }
 751                 fprintf(f, "%lu\n", (unsigned long)
 752                         hist_sum(FIO_IO_U_PLAT_NR - stride, stride, io_u_plat,
 753                                         io_u_plat_before));
 754
 755                 flist_del(&entry_before->list);
 756                 free(entry_before);
 757         }
 758 }
 759
 760 void flush_samples(FILE *f, void *samples, uint64_t sample_size)
 761 {
 762         struct io_sample *s;
 763         int log_offset;
 764         uint64_t i, nr_samples;
 765
 766         if (!sample_size)
 767                 return;
 768
 769         s = __get_sample(samples, 0, 0);
 770         log_offset = (s->__ddir & LOG_OFFSET_SAMPLE_BIT) != 0;
 771
 772         nr_samples = sample_size / __log_entry_sz(log_offset);
 773
 774         for (i = 0; i < nr_samples; i++) {
 775                 s = __get_sample(samples, log_offset, i);
 776
 777                 if (!log_offset) {
 778                         fprintf(f, "%lu, %" PRId64 ", %u, %u\n",
 779                                         (unsigned long) s->time,
 780                                         s->data.val,
 781                                         io_sample_ddir(s), s->bs);
 782                 } else {
 783                         struct io_sample_offset *so = (void *) s;
 784
 785                         fprintf(f, "%lu, %" PRId64 ", %u, %u, %llu\n",
 786                                         (unsigned long) s->time,
 787                                         s->data.val,
 788                                         io_sample_ddir(s), s->bs,
 789                                         (unsigned long long) so->offset);
 790                 }
 791         }
 792 }
 793
 794 #ifdef CONFIG_ZLIB
 795
 796 struct iolog_flush_data {
 797         struct workqueue_work work;
 798         struct io_log *log;
 799         void *samples;
 800         uint32_t nr_samples;
 801         bool free;
 802 };
 803
 804 #define GZ_CHUNK        131072
 805
 806 static struct iolog_compress *get_new_chunk(unsigned int seq)
 807 {
 808         struct iolog_compress *c;
 809
 810         c = malloc(sizeof(*c));
 811         INIT_FLIST_HEAD(&c->list);
 812         c->buf = malloc(GZ_CHUNK);
 813         c->len = 0;
 814         c->seq = seq;
 815         return c;
 816 }
 817
 818 static void free_chunk(struct iolog_compress *ic)
 819 {
 820         free(ic->buf);
 821         free(ic);
 822 }
 823
 824 static int z_stream_init(z_stream *stream, int gz_hdr)
 825 {
 826         int wbits = 15;
 827
 828         memset(stream, 0, sizeof(*stream));
 829         stream->zalloc = Z_NULL;
 830         stream->zfree = Z_NULL;
 831         stream->opaque = Z_NULL;
 832         stream->next_in = Z_NULL;
 833
 834         /*
 835          * zlib magic - add 32 for auto-detection of gz header or not,
 836          * if we decide to store files in a gzip friendly format.
 837          */
 838         if (gz_hdr)
 839                 wbits += 32;
 840
 841         if (inflateInit2(stream, wbits) != Z_OK)
 842                 return 1;
 843
 844         return 0;
 845 }
 846
 847 struct inflate_chunk_iter {
 848         unsigned int seq;
 849         int err;
 850         void *buf;
 851         size_t buf_size;
 852         size_t buf_used;
 853         size_t chunk_sz;
 854 };
 855
 856 static void finish_chunk(z_stream *stream, FILE *f,
 857                          struct inflate_chunk_iter *iter)
 858 {
 859         int ret;
 860
 861         ret = inflateEnd(stream);
 862         if (ret != Z_OK)
 863                 log_err("fio: failed to end log inflation seq %d (%d)\n",
 864                                 iter->seq, ret);
 865
 866         flush_samples(f, iter->buf, iter->buf_used);
 867         free(iter->buf);
 868         iter->buf = NULL;
 869         iter->buf_size = iter->buf_used = 0;
 870 }
 871
 872 /*
 873  * Iterative chunk inflation. Handles cases where we cross into a new
 874  * sequence, doing flush finish of previous chunk if needed.
 875  */
 876 static size_t inflate_chunk(struct iolog_compress *ic, int gz_hdr, FILE *f,
 877                             z_stream *stream, struct inflate_chunk_iter *iter)
 878 {
 879         size_t ret;
 880
 881         dprint(FD_COMPRESS, "inflate chunk size=%lu, seq=%u\n",
 882                                 (unsigned long) ic->len, ic->seq);
 883
 884         if (ic->seq != iter->seq) {
 885                 if (iter->seq)
 886                         finish_chunk(stream, f, iter);
 887
 888                 z_stream_init(stream, gz_hdr);
 889                 iter->seq = ic->seq;
 890         }
 891
 892         stream->avail_in = ic->len;
 893         stream->next_in = ic->buf;
 894
 895         if (!iter->buf_size) {
 896                 iter->buf_size = iter->chunk_sz;
 897                 iter->buf = malloc(iter->buf_size);
 898         }
 899
 900         while (stream->avail_in) {
 901                 size_t this_out = iter->buf_size - iter->buf_used;
 902                 int err;
 903
 904                 stream->avail_out = this_out;
 905                 stream->next_out = iter->buf + iter->buf_used;
 906
 907                 err = inflate(stream, Z_NO_FLUSH);
 908                 if (err < 0) {
 909                         log_err("fio: failed inflating log: %d\n", err);
 910                         iter->err = err;
 911                         break;
 912                 }
 913
 914                 iter->buf_used += this_out - stream->avail_out;
 915
 916                 if (!stream->avail_out) {
 917                         iter->buf_size += iter->chunk_sz;
 918                         iter->buf = realloc(iter->buf, iter->buf_size);
 919                         continue;
 920                 }
 921
 922                 if (err == Z_STREAM_END)
 923                         break;
 924         }
 925
 926         ret = (void *) stream->next_in - ic->buf;
 927
 928         dprint(FD_COMPRESS, "inflated to size=%lu\n", (unsigned long) iter->buf_size);
 929
 930         return ret;
 931 }
 932
 933 /*
 934  * Inflate stored compressed chunks, or write them directly to the log
 935  * file if so instructed.
 936  */
 937 static int inflate_gz_chunks(struct io_log *log, FILE *f)
 938 {
 939         struct inflate_chunk_iter iter = { .chunk_sz = log->log_gz, };
 940         z_stream stream;
 941
 942         while (!flist_empty(&log->chunk_list)) {
 943                 struct iolog_compress *ic;
 944
 945                 ic = flist_first_entry(&log->chunk_list, struct iolog_compress, list);
 946                 flist_del(&ic->list);
 947
 948                 if (log->log_gz_store) {
 949                         size_t ret;
 950
 951                         dprint(FD_COMPRESS, "log write chunk size=%lu, "
 952                                 "seq=%u\n", (unsigned long) ic->len, ic->seq);
 953
 954                         ret = fwrite(ic->buf, ic->len, 1, f);
 955                         if (ret != 1 || ferror(f)) {
 956                                 iter.err = errno;
 957                                 log_err("fio: error writing compressed log\n");
 958                         }
 959                 } else
 960                         inflate_chunk(ic, log->log_gz_store, f, &stream, &iter);
 961
 962                 free_chunk(ic);
 963         }
 964
 965         if (iter.seq) {
 966                 finish_chunk(&stream, f, &iter);
 967                 free(iter.buf);
 968         }
 969
 970         return iter.err;
 971 }
 972
 973 /*
 974  * Open compressed log file and decompress the stored chunks and
 975  * write them to stdout. The chunks are stored sequentially in the
 976  * file, so we iterate over them and do them one-by-one.
 977  */
 978 int iolog_file_inflate(const char *file)
 979 {
 980         struct inflate_chunk_iter iter = { .chunk_sz = 64 * 1024 * 1024, };
 981         struct iolog_compress ic;
 982         z_stream stream;
 983         struct stat sb;
 984         ssize_t ret;
 985         size_t total;
 986         void *buf;
 987         FILE *f;
 988
 989         f = fopen(file, "r");
 990         if (!f) {
 991                 perror("fopen");
 992                 return 1;
 993         }
 994
 995         if (stat(file, &sb) < 0) {
 996                 fclose(f);
 997                 perror("stat");
 998                 return 1;
 999         }
1000
1001         ic.buf = buf = malloc(sb.st_size);
1002         ic.len = sb.st_size;
1003         ic.seq = 1;
1004
1005         ret = fread(ic.buf, ic.len, 1, f);
1006         if (ret < 0) {
1007                 perror("fread");
1008                 fclose(f);
1009                 free(buf);
1010                 return 1;
1011         } else if (ret != 1) {
1012                 log_err("fio: short read on reading log\n");
1013                 fclose(f);
1014                 free(buf);
1015                 return 1;
1016         }
1017
1018         fclose(f);
1019
1020         /*
1021          * Each chunk will return Z_STREAM_END. We don't know how many
1022          * chunks are in the file, so we just keep looping and incrementing
1023          * the sequence number until we have consumed the whole compressed
1024          * file.
1025          */
1026         total = ic.len;
1027         do {
1028                 size_t iret;
1029
1030                 iret = inflate_chunk(&ic,  1, stdout, &stream, &iter);
1031                 total -= iret;
1032                 if (!total)
1033                         break;
1034                 if (iter.err)
1035                         break;
1036
1037                 ic.seq++;
1038                 ic.len -= iret;
1039                 ic.buf += iret;
1040         } while (1);
1041
1042         if (iter.seq) {
1043                 finish_chunk(&stream, stdout, &iter);
1044                 free(iter.buf);
1045         }
1046
1047         free(buf);
1048         return iter.err;
1049 }
1050
1051 #else
1052
1053 static int inflate_gz_chunks(struct io_log *log, FILE *f)
1054 {
1055         return 0;
1056 }
1057
1058 int iolog_file_inflate(const char *file)
1059 {
1060         log_err("fio: log inflation not possible without zlib\n");
1061         return 1;
1062 }
1063
1064 #endif
1065
1066 void flush_log(struct io_log *log, bool do_append)
1067 {
1068         void *buf;
1069         FILE *f;
1070
1071         if (!do_append)
1072                 f = fopen(log->filename, "w");
1073         else
1074                 f = fopen(log->filename, "a");
1075         if (!f) {
1076                 perror("fopen log");
1077                 return;
1078         }
1079
1080         buf = set_file_buffer(f);
1081
1082         inflate_gz_chunks(log, f);
1083
1084         while (!flist_empty(&log->io_logs)) {
1085                 struct io_logs *cur_log;
1086
1087                 cur_log = flist_first_entry(&log->io_logs, struct io_logs, list);
1088                 flist_del_init(&cur_log->list);
1089
1090                 if (log->td && log == log->td->clat_hist_log)
1091                         flush_hist_samples(f, log->hist_coarseness, cur_log->log,
1092                                            log_sample_sz(log, cur_log));
1093                 else
1094                         flush_samples(f, cur_log->log, log_sample_sz(log, cur_log));
1095
1096                 sfree(cur_log);
1097         }
1098
1099         fclose(f);
1100         clear_file_buffer(buf);
1101 }
1102
1103 static int finish_log(struct thread_data *td, struct io_log *log, int trylock)
1104 {
1105         if (td->flags & TD_F_COMPRESS_LOG)
1106                 iolog_flush(log);
1107
1108         if (trylock) {
1109                 if (fio_trylock_file(log->filename))
1110                         return 1;
1111         } else
1112                 fio_lock_file(log->filename);
1113
1114         if (td->client_type == FIO_CLIENT_TYPE_GUI || is_backend)
1115                 fio_send_iolog(td, log, log->filename);
1116         else
1117                 flush_log(log, !td->o.per_job_logs);
1118
1119         fio_unlock_file(log->filename);
1120         free_log(log);
1121         return 0;
1122 }
1123
1124 size_t log_chunk_sizes(struct io_log *log)
1125 {
1126         struct flist_head *entry;
1127         size_t ret;
1128
1129         if (flist_empty(&log->chunk_list))
1130                 return 0;
1131
1132         ret = 0;
1133         pthread_mutex_lock(&log->chunk_lock);
1134         flist_for_each(entry, &log->chunk_list) {
1135                 struct iolog_compress *c;
1136
1137                 c = flist_entry(entry, struct iolog_compress, list);
1138                 ret += c->len;
1139         }
1140         pthread_mutex_unlock(&log->chunk_lock);
1141         return ret;
1142 }
1143
1144 #ifdef CONFIG_ZLIB
1145
1146 static bool warned_on_drop;
1147
1148 static void iolog_put_deferred(struct io_log *log, void *ptr)
1149 {
1150         if (!ptr)
1151                 return;
1152
1153         pthread_mutex_lock(&log->deferred_free_lock);
1154         if (log->deferred < IOLOG_MAX_DEFER) {
1155                 log->deferred_items[log->deferred] = ptr;
1156                 log->deferred++;
1157         } else if (!warned_on_drop) {
1158                 log_err("fio: had to drop log entry free\n");
1159                 warned_on_drop = true;
1160         }
1161         pthread_mutex_unlock(&log->deferred_free_lock);
1162 }
1163
1164 static void iolog_free_deferred(struct io_log *log)
1165 {
1166         int i;
1167
1168         if (!log->deferred)
1169                 return;
1170
1171         pthread_mutex_lock(&log->deferred_free_lock);
1172
1173         for (i = 0; i < log->deferred; i++) {
1174                 free(log->deferred_items[i]);
1175                 log->deferred_items[i] = NULL;
1176         }
1177
1178         log->deferred = 0;
1179         pthread_mutex_unlock(&log->deferred_free_lock);
1180 }
1181
1182 static int gz_work(struct iolog_flush_data *data)
1183 {
1184         struct iolog_compress *c = NULL;
1185         struct flist_head list;
1186         unsigned int seq;
1187         z_stream stream;
1188         size_t total = 0;
1189         int ret;
1190
1191         INIT_FLIST_HEAD(&list);
1192
1193         memset(&stream, 0, sizeof(stream));
1194         stream.zalloc = Z_NULL;
1195         stream.zfree = Z_NULL;
1196         stream.opaque = Z_NULL;
1197
1198         ret = deflateInit(&stream, Z_DEFAULT_COMPRESSION);
1199         if (ret != Z_OK) {
1200                 log_err("fio: failed to init gz stream\n");
1201                 goto err;
1202         }
1203
1204         seq = ++data->log->chunk_seq;
1205
1206         stream.next_in = (void *) data->samples;
1207         stream.avail_in = data->nr_samples * log_entry_sz(data->log);
1208
1209         dprint(FD_COMPRESS, "deflate input size=%lu, seq=%u, log=%s\n",
1210                                 (unsigned long) stream.avail_in, seq,
1211                                 data->log->filename);
1212         do {
1213                 if (c)
1214                         dprint(FD_COMPRESS, "seq=%d, chunk=%lu\n", seq,
1215                                 (unsigned long) c->len);
1216                 c = get_new_chunk(seq);
1217                 stream.avail_out = GZ_CHUNK;
1218                 stream.next_out = c->buf;
1219                 ret = deflate(&stream, Z_NO_FLUSH);
1220                 if (ret < 0) {
1221                         log_err("fio: deflate log (%d)\n", ret);
1222                         free_chunk(c);
1223                         goto err;
1224                 }
1225
1226                 c->len = GZ_CHUNK - stream.avail_out;
1227                 flist_add_tail(&c->list, &list);
1228                 total += c->len;
1229         } while (stream.avail_in);
1230
1231         stream.next_out = c->buf + c->len;
1232         stream.avail_out = GZ_CHUNK - c->len;
1233
1234         ret = deflate(&stream, Z_FINISH);
1235         if (ret < 0) {
1236                 /*
1237                  * Z_BUF_ERROR is special, it just means we need more
1238                  * output space. We'll handle that below. Treat any other
1239                  * error as fatal.
1240                  */
1241                 if (ret != Z_BUF_ERROR) {
1242                         log_err("fio: deflate log (%d)\n", ret);
1243                         flist_del(&c->list);
1244                         free_chunk(c);
1245                         goto err;
1246                 }
1247         }
1248
1249         total -= c->len;
1250         c->len = GZ_CHUNK - stream.avail_out;
1251         total += c->len;
1252         dprint(FD_COMPRESS, "seq=%d, chunk=%lu\n", seq, (unsigned long) c->len);
1253
1254         if (ret != Z_STREAM_END) {
1255                 do {
1256                         c = get_new_chunk(seq);
1257                         stream.avail_out = GZ_CHUNK;
1258                         stream.next_out = c->buf;
1259                         ret = deflate(&stream, Z_FINISH);
1260                         c->len = GZ_CHUNK - stream.avail_out;
1261                         total += c->len;
1262                         flist_add_tail(&c->list, &list);
1263                         dprint(FD_COMPRESS, "seq=%d, chunk=%lu\n", seq,
1264                                 (unsigned long) c->len);
1265                 } while (ret != Z_STREAM_END);
1266         }
1267
1268         dprint(FD_COMPRESS, "deflated to size=%lu\n", (unsigned long) total);
1269
1270         ret = deflateEnd(&stream);
1271         if (ret != Z_OK)
1272                 log_err("fio: deflateEnd %d\n", ret);
1273
1274         iolog_put_deferred(data->log, data->samples);
1275
1276         if (!flist_empty(&list)) {
1277                 pthread_mutex_lock(&data->log->chunk_lock);
1278                 flist_splice_tail(&list, &data->log->chunk_list);
1279                 pthread_mutex_unlock(&data->log->chunk_lock);
1280         }
1281
1282         ret = 0;
1283 done:
1284         if (data->free)
1285                 sfree(data);
1286         return ret;
1287 err:
1288         while (!flist_empty(&list)) {
1289                 c = flist_first_entry(list.next, struct iolog_compress, list);
1290                 flist_del(&c->list);
1291                 free_chunk(c);
1292         }
1293         ret = 1;
1294         goto done;
1295 }
1296
1297 /*
1298  * Invoked from our compress helper thread, when logging would have exceeded
1299  * the specified memory limitation. Compresses the previously stored
1300  * entries.
1301  */
1302 static int gz_work_async(struct submit_worker *sw, struct workqueue_work *work)
1303 {
1304         return gz_work(container_of(work, struct iolog_flush_data, work));
1305 }
1306
1307 static int gz_init_worker(struct submit_worker *sw)
1308 {
1309         struct thread_data *td = sw->wq->td;
1310
1311         if (!fio_option_is_set(&td->o, log_gz_cpumask))
1312                 return 0;
1313
1314         if (fio_setaffinity(gettid(), td->o.log_gz_cpumask) == -1) {
1315                 log_err("gz: failed to set CPU affinity\n");
1316                 return 1;
1317         }
1318
1319         return 0;
1320 }
1321
1322 static struct workqueue_ops log_compress_wq_ops = {
1323         .fn             = gz_work_async,
1324         .init_worker_fn = gz_init_worker,
1325         .nice           = 1,
1326 };
1327
1328 int iolog_compress_init(struct thread_data *td, struct sk_out *sk_out)
1329 {
1330         if (!(td->flags & TD_F_COMPRESS_LOG))
1331                 return 0;
1332
1333         workqueue_init(td, &td->log_compress_wq, &log_compress_wq_ops, 1, sk_out);
1334         return 0;
1335 }
1336
1337 void iolog_compress_exit(struct thread_data *td)
1338 {
1339         if (!(td->flags & TD_F_COMPRESS_LOG))
1340                 return;
1341
1342         workqueue_exit(&td->log_compress_wq);
1343 }
1344
1345 /*
1346  * Queue work item to compress the existing log entries. We reset the
1347  * current log to a small size, and reference the existing log in the
1348  * data that we queue for compression. Once compression has been done,
1349  * this old log is freed. If called with finish == true, will not return
1350  * until the log compression has completed, and will flush all previous
1351  * logs too
1352  */
1353 static int iolog_flush(struct io_log *log)
1354 {
1355         struct iolog_flush_data *data;
1356
1357         data = malloc(sizeof(*data));
1358         if (!data)
1359                 return 1;
1360
1361         data->log = log;
1362         data->free = false;
1363
1364         while (!flist_empty(&log->io_logs)) {
1365                 struct io_logs *cur_log;
1366
1367                 cur_log = flist_first_entry(&log->io_logs, struct io_logs, list);
1368                 flist_del_init(&cur_log->list);
1369
1370                 data->samples = cur_log->log;
1371                 data->nr_samples = cur_log->nr_samples;
1372
1373                 sfree(cur_log);
1374
1375                 gz_work(data);
1376         }
1377
1378         free(data);
1379         return 0;
1380 }
1381
1382 int iolog_cur_flush(struct io_log *log, struct io_logs *cur_log)
1383 {
1384         struct iolog_flush_data *data;
1385
1386         data = smalloc(sizeof(*data));
1387         if (!data)
1388                 return 1;
1389
1390         data->log = log;
1391
1392         data->samples = cur_log->log;
1393         data->nr_samples = cur_log->nr_samples;
1394         data->free = true;
1395
1396         cur_log->nr_samples = cur_log->max_samples = 0;
1397         cur_log->log = NULL;
1398
1399         workqueue_enqueue(&log->td->log_compress_wq, &data->work);
1400
1401         iolog_free_deferred(log);
1402
1403         return 0;
1404 }
1405 #else
1406
1407 static int iolog_flush(struct io_log *log)
1408 {
1409         return 1;
1410 }
1411
1412 int iolog_cur_flush(struct io_log *log, struct io_logs *cur_log)
1413 {
1414         return 1;
1415 }
1416
1417 int iolog_compress_init(struct thread_data *td, struct sk_out *sk_out)
1418 {
1419         return 0;
1420 }
1421
1422 void iolog_compress_exit(struct thread_data *td)
1423 {
1424 }
1425
1426 #endif
1427
1428 struct io_logs *iolog_cur_log(struct io_log *log)
1429 {
1430         if (flist_empty(&log->io_logs))
1431                 return NULL;
1432
1433         return flist_last_entry(&log->io_logs, struct io_logs, list);
1434 }
1435
1436 uint64_t iolog_nr_samples(struct io_log *iolog)
1437 {
1438         struct flist_head *entry;
1439         uint64_t ret = 0;
1440
1441         flist_for_each(entry, &iolog->io_logs) {
1442                 struct io_logs *cur_log;
1443
1444                 cur_log = flist_entry(entry, struct io_logs, list);
1445                 ret += cur_log->nr_samples;
1446         }
1447
1448         return ret;
1449 }
1450
1451 static int __write_log(struct thread_data *td, struct io_log *log, int try)
1452 {
1453         if (log)
1454                 return finish_log(td, log, try);
1455
1456         return 0;
1457 }
1458
1459 static int write_iops_log(struct thread_data *td, int try, bool unit_log)
1460 {
1461         int ret;
1462
1463         if (per_unit_log(td->iops_log) != unit_log)
1464                 return 0;
1465
1466         ret = __write_log(td, td->iops_log, try);
1467         if (!ret)
1468                 td->iops_log = NULL;
1469
1470         return ret;
1471 }
1472
1473 static int write_slat_log(struct thread_data *td, int try, bool unit_log)
1474 {
1475         int ret;
1476
1477         if (!unit_log)
1478                 return 0;
1479
1480         ret = __write_log(td, td->slat_log, try);
1481         if (!ret)
1482                 td->slat_log = NULL;
1483
1484         return ret;
1485 }
1486
1487 static int write_clat_log(struct thread_data *td, int try, bool unit_log)
1488 {
1489         int ret;
1490
1491         if (!unit_log)
1492                 return 0;
1493
1494         ret = __write_log(td, td->clat_log, try);
1495         if (!ret)
1496                 td->clat_log = NULL;
1497
1498         return ret;
1499 }
1500
1501 static int write_clat_hist_log(struct thread_data *td, int try, bool unit_log)
1502 {
1503         int ret;
1504
1505         if (!unit_log)
1506                 return 0;
1507
1508         ret = __write_log(td, td->clat_hist_log, try);
1509         if (!ret)
1510                 td->clat_hist_log = NULL;
1511
1512         return ret;
1513 }
1514
1515 static int write_lat_log(struct thread_data *td, int try, bool unit_log)
1516 {
1517         int ret;
1518
1519         if (!unit_log)
1520                 return 0;
1521
1522         ret = __write_log(td, td->lat_log, try);
1523         if (!ret)
1524                 td->lat_log = NULL;
1525
1526         return ret;
1527 }
1528
1529 static int write_bandw_log(struct thread_data *td, int try, bool unit_log)
1530 {
1531         int ret;
1532
1533         if (per_unit_log(td->bw_log) != unit_log)
1534                 return 0;
1535
1536         ret = __write_log(td, td->bw_log, try);
1537         if (!ret)
1538                 td->bw_log = NULL;
1539
1540         return ret;
1541 }
1542
1543 enum {
1544         BW_LOG_MASK     = 1,
1545         LAT_LOG_MASK    = 2,
1546         SLAT_LOG_MASK   = 4,
1547         CLAT_LOG_MASK   = 8,
1548         IOPS_LOG_MASK   = 16,
1549         CLAT_HIST_LOG_MASK = 32,
1550
1551         ALL_LOG_NR      = 6,
1552 };
1553
1554 struct log_type {
1555         unsigned int mask;
1556         int (*fn)(struct thread_data *, int, bool);
1557 };
1558
1559 static struct log_type log_types[] = {
1560         {
1561                 .mask   = BW_LOG_MASK,
1562                 .fn     = write_bandw_log,
1563         },
1564         {
1565                 .mask   = LAT_LOG_MASK,
1566                 .fn     = write_lat_log,
1567         },
1568         {
1569                 .mask   = SLAT_LOG_MASK,
1570                 .fn     = write_slat_log,
1571         },
1572         {
1573                 .mask   = CLAT_LOG_MASK,
1574                 .fn     = write_clat_log,
1575         },
1576         {
1577                 .mask   = IOPS_LOG_MASK,
1578                 .fn     = write_iops_log,
1579         },
1580         {
1581                 .mask   = CLAT_HIST_LOG_MASK,
1582                 .fn     = write_clat_hist_log,
1583         }
1584 };
1585
1586 void td_writeout_logs(struct thread_data *td, bool unit_logs)
1587 {
1588         unsigned int log_mask = 0;
1589         unsigned int log_left = ALL_LOG_NR;
1590         int old_state, i;
1591
1592         old_state = td_bump_runstate(td, TD_FINISHING);
1593
1594         finalize_logs(td, unit_logs);
1595
1596         while (log_left) {
1597                 int prev_log_left = log_left;
1598
1599                 for (i = 0; i < ALL_LOG_NR && log_left; i++) {
1600                         struct log_type *lt = &log_types[i];
1601                         int ret;
1602
1603                         if (!(log_mask & lt->mask)) {
1604                                 ret = lt->fn(td, log_left != 1, unit_logs);
1605                                 if (!ret) {
1606                                         log_left--;
1607                                         log_mask |= lt->mask;
1608                                 }
1609                         }
1610                 }
1611
1612                 if (prev_log_left == log_left)
1613                         usleep(5000);
1614         }
1615
1616         td_restore_runstate(td, old_state);
1617 }
1618
1619 void fio_writeout_logs(bool unit_logs)
1620 {
1621         struct thread_data *td;
1622         int i;
1623
1624         for_each_td(td, i)
1625                 td_writeout_logs(td, unit_logs);
1626 }