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:
 231          *
 232          *      Sequential writes, or
 233          *      Random writes that lay out the file as it goes along
 234          *
 235          * For both these cases, just reading back data in the order we
 236          * wrote it out is the fastest.
 237          *
 238          * One exception is if we don't have a random map AND we are doing
 239          * verifies, in that case we need to check for duplicate blocks and
 240          * drop the old one, which we rely on the rb insert/lookup for
 241          * handling.
 242          */
 243         if (((!td->o.verifysort) || !td_random(td) || !td->o.overwrite) &&
 244               (file_randommap(td, ipo->file) || td->o.verify == VERIFY_NONE)) {
 245                 INIT_FLIST_HEAD(&ipo->list);
 246                 flist_add_tail(&ipo->list, &td->io_hist_list);
 247                 ipo->flags |= IP_F_ONLIST;
 248                 td->io_hist_len++;
 249                 return;
 250         }
 251
 252         RB_CLEAR_NODE(&ipo->rb_node);
 253
 254         /*
 255          * Sort the entry into the verification list
 256          */
 257 restart:
 258         p = &td->io_hist_tree.rb_node;
 259         parent = NULL;
 260         while (*p) {
 261                 int overlap = 0;
 262                 parent = *p;
 263
 264                 __ipo = rb_entry(parent, struct io_piece, rb_node);
 265                 if (ipo->file < __ipo->file)
 266                         p = &(*p)->rb_left;
 267                 else if (ipo->file > __ipo->file)
 268                         p = &(*p)->rb_right;
 269                 else if (ipo->offset < __ipo->offset) {
 270                         p = &(*p)->rb_left;
 271                         overlap = ipo->offset + ipo->len > __ipo->offset;
 272                 }
 273                 else if (ipo->offset > __ipo->offset) {
 274                         p = &(*p)->rb_right;
 275                         overlap = __ipo->offset + __ipo->len > ipo->offset;
 276                 }
 277                 else
 278                         overlap = 1;
 279
 280                 if (overlap) {
 281                         dprint(FD_IO, "iolog: overlap %llu/%lu, %llu/%lu\n",
 282                                 __ipo->offset, __ipo->len,
 283                                 ipo->offset, ipo->len);
 284                         td->io_hist_len--;
 285                         rb_erase(parent, &td->io_hist_tree);
 286                         remove_trim_entry(td, __ipo);
 287                         if (!(__ipo->flags & IP_F_IN_FLIGHT))
 288                                 free(__ipo);
 289                         goto restart;
 290                 }
 291         }
 292
 293         rb_link_node(&ipo->rb_node, parent, p);
 294         rb_insert_color(&ipo->rb_node, &td->io_hist_tree);
 295         ipo->flags |= IP_F_ONRB;
 296         td->io_hist_len++;
 297 }
 298
 299 void unlog_io_piece(struct thread_data *td, struct io_u *io_u)
 300 {
 301         struct io_piece *ipo = io_u->ipo;
 302
 303         if (td->ts.nr_block_infos) {
 304                 uint32_t *info = io_u_block_info(td, io_u);
 305                 if (BLOCK_INFO_STATE(*info) < BLOCK_STATE_TRIM_FAILURE) {
 306                         if (io_u->ddir == DDIR_TRIM)
 307                                 *info = BLOCK_INFO_SET_STATE(*info,
 308                                                 BLOCK_STATE_TRIM_FAILURE);
 309                         else if (io_u->ddir == DDIR_WRITE)
 310                                 *info = BLOCK_INFO_SET_STATE(*info,
 311                                                 BLOCK_STATE_WRITE_FAILURE);
 312                 }
 313         }
 314
 315         if (!ipo)
 316                 return;
 317
 318         if (ipo->flags & IP_F_ONRB)
 319                 rb_erase(&ipo->rb_node, &td->io_hist_tree);
 320         else if (ipo->flags & IP_F_ONLIST)
 321                 flist_del(&ipo->list);
 322
 323         free(ipo);
 324         io_u->ipo = NULL;
 325         td->io_hist_len--;
 326 }
 327
 328 void trim_io_piece(struct thread_data *td, const struct io_u *io_u)
 329 {
 330         struct io_piece *ipo = io_u->ipo;
 331
 332         if (!ipo)
 333                 return;
 334
 335         ipo->len = io_u->xfer_buflen - io_u->resid;
 336 }
 337
 338 void write_iolog_close(struct thread_data *td)
 339 {
 340         fflush(td->iolog_f);
 341         fclose(td->iolog_f);
 342         free(td->iolog_buf);
 343         td->iolog_f = NULL;
 344         td->iolog_buf = NULL;
 345 }
 346
 347 /*
 348  * Read version 2 iolog data. It is enhanced to include per-file logging,
 349  * syncs, etc.
 350  */
 351 static int read_iolog2(struct thread_data *td, FILE *f)
 352 {
 353         unsigned long long offset;
 354         unsigned int bytes;
 355         int reads, writes, waits, fileno = 0, file_action = 0; /* stupid gcc */
 356         char *rfname, *fname, *act;
 357         char *str, *p;
 358         enum fio_ddir rw;
 359
 360         free_release_files(td);
 361
 362         /*
 363          * Read in the read iolog and store it, reuse the infrastructure
 364          * for doing verifications.
 365          */
 366         str = malloc(4096);
 367         rfname = fname = malloc(256+16);
 368         act = malloc(256+16);
 369
 370         reads = writes = waits = 0;
 371         while ((p = fgets(str, 4096, f)) != NULL) {
 372                 struct io_piece *ipo;
 373                 int r;
 374
 375                 r = sscanf(p, "%256s %256s %llu %u", rfname, act, &offset,
 376                                                                         &bytes);
 377
 378                 if (td->o.replay_redirect)
 379                         fname = td->o.replay_redirect;
 380
 381                 if (r == 4) {
 382                         /*
 383                          * Check action first
 384                          */
 385                         if (!strcmp(act, "wait"))
 386                                 rw = DDIR_WAIT;
 387                         else if (!strcmp(act, "read"))
 388                                 rw = DDIR_READ;
 389                         else if (!strcmp(act, "write"))
 390                                 rw = DDIR_WRITE;
 391                         else if (!strcmp(act, "sync"))
 392                                 rw = DDIR_SYNC;
 393                         else if (!strcmp(act, "datasync"))
 394                                 rw = DDIR_DATASYNC;
 395                         else if (!strcmp(act, "trim"))
 396                                 rw = DDIR_TRIM;
 397                         else {
 398                                 log_err("fio: bad iolog file action: %s\n",
 399                                                                         act);
 400                                 continue;
 401                         }
 402                         fileno = get_fileno(td, fname);
 403                 } else if (r == 2) {
 404                         rw = DDIR_INVAL;
 405                         if (!strcmp(act, "add")) {
 406                                 if (td->o.replay_redirect &&
 407                                     get_fileno(td, fname) != -1) {
 408                                         dprint(FD_FILE, "iolog: ignoring"
 409                                                 " re-add of file %s\n", fname);
 410                                 } else {
 411                                         fileno = add_file(td, fname, 0, 1);
 412                                         file_action = FIO_LOG_ADD_FILE;
 413                                 }
 414                                 continue;
 415                         } else if (!strcmp(act, "open")) {
 416                                 fileno = get_fileno(td, fname);
 417                                 file_action = FIO_LOG_OPEN_FILE;
 418                         } else if (!strcmp(act, "close")) {
 419                                 fileno = get_fileno(td, fname);
 420                                 file_action = FIO_LOG_CLOSE_FILE;
 421                         } else {
 422                                 log_err("fio: bad iolog file action: %s\n",
 423                                                                         act);
 424                                 continue;
 425                         }
 426                 } else {
 427                         log_err("bad iolog2: %s\n", p);
 428                         continue;
 429                 }
 430
 431                 if (rw == DDIR_READ)
 432                         reads++;
 433                 else if (rw == DDIR_WRITE) {
 434                         /*
 435                          * Don't add a write for ro mode
 436                          */
 437                         if (read_only)
 438                                 continue;
 439                         writes++;
 440                 } else if (rw == DDIR_WAIT) {
 441                         if (td->o.no_stall)
 442                                 continue;
 443                         waits++;
 444                 } else if (rw == DDIR_INVAL) {
 445                 } else if (!ddir_sync(rw)) {
 446                         log_err("bad ddir: %d\n", rw);
 447                         continue;
 448                 }
 449
 450                 /*
 451                  * Make note of file
 452                  */
 453                 ipo = malloc(sizeof(*ipo));
 454                 init_ipo(ipo);
 455                 ipo->ddir = rw;
 456                 if (rw == DDIR_WAIT) {
 457                         ipo->delay = offset;
 458                 } else {
 459                         if (td->o.replay_scale)
 460                                 ipo->offset = offset / td->o.replay_scale;
 461                         else
 462                                 ipo->offset = offset;
 463                         ipo_bytes_align(td->o.replay_align, ipo);
 464
 465                         ipo->len = bytes;
 466                         if (rw != DDIR_INVAL && bytes > td->o.max_bs[rw])
 467                                 td->o.max_bs[rw] = bytes;
 468                         ipo->fileno = fileno;
 469                         ipo->file_action = file_action;
 470                         td->o.size += bytes;
 471                 }
 472
 473                 queue_io_piece(td, ipo);
 474         }
 475
 476         free(str);
 477         free(act);
 478         free(rfname);
 479
 480         if (writes && read_only) {
 481                 log_err("fio: <%s> skips replay of %d writes due to"
 482                         " read-only\n", td->o.name, writes);
 483                 writes = 0;
 484         }
 485
 486         if (!reads && !writes && !waits)
 487                 return 1;
 488         else if (reads && !writes)
 489                 td->o.td_ddir = TD_DDIR_READ;
 490         else if (!reads && writes)
 491                 td->o.td_ddir = TD_DDIR_WRITE;
 492         else
 493                 td->o.td_ddir = TD_DDIR_RW;
 494
 495         return 0;
 496 }
 497
 498 /*
 499  * open iolog, check version, and call appropriate parser
 500  */
 501 static int init_iolog_read(struct thread_data *td)
 502 {
 503         char buffer[256], *p;
 504         FILE *f;
 505         int ret;
 506
 507         f = fopen(td->o.read_iolog_file, "r");
 508         if (!f) {
 509                 perror("fopen read iolog");
 510                 return 1;
 511         }
 512
 513         p = fgets(buffer, sizeof(buffer), f);
 514         if (!p) {
 515                 td_verror(td, errno, "iolog read");
 516                 log_err("fio: unable to read iolog\n");
 517                 fclose(f);
 518                 return 1;
 519         }
 520
 521         /*
 522          * version 2 of the iolog stores a specific string as the
 523          * first line, check for that
 524          */
 525         if (!strncmp(iolog_ver2, buffer, strlen(iolog_ver2)))
 526                 ret = read_iolog2(td, f);
 527         else {
 528                 log_err("fio: iolog version 1 is no longer supported\n");
 529                 ret = 1;
 530         }
 531
 532         fclose(f);
 533         return ret;
 534 }
 535
 536 /*
 537  * Set up a log for storing io patterns.
 538  */
 539 static int init_iolog_write(struct thread_data *td)
 540 {
 541         struct fio_file *ff;
 542         FILE *f;
 543         unsigned int i;
 544
 545         f = fopen(td->o.write_iolog_file, "a");
 546         if (!f) {
 547                 perror("fopen write iolog");
 548                 return 1;
 549         }
 550
 551         /*
 552          * That's it for writing, setup a log buffer and we're done.
 553           */
 554         td->iolog_f = f;
 555         td->iolog_buf = malloc(8192);
 556         setvbuf(f, td->iolog_buf, _IOFBF, 8192);
 557
 558         /*
 559          * write our version line
 560          */
 561         if (fprintf(f, "%s\n", iolog_ver2) < 0) {
 562                 perror("iolog init\n");
 563                 return 1;
 564         }
 565
 566         /*
 567          * add all known files
 568          */
 569         for_each_file(td, ff, i)
 570                 log_file(td, ff, FIO_LOG_ADD_FILE);
 571
 572         return 0;
 573 }
 574
 575 int init_iolog(struct thread_data *td)
 576 {
 577         int ret = 0;
 578
 579         if (td->o.read_iolog_file) {
 580                 int need_swap;
 581
 582                 /*
 583                  * Check if it's a blktrace file and load that if possible.
 584                  * Otherwise assume it's a normal log file and load that.
 585                  */
 586                 if (is_blktrace(td->o.read_iolog_file, &need_swap))
 587                         ret = load_blktrace(td, td->o.read_iolog_file, need_swap);
 588                 else
 589                         ret = init_iolog_read(td);
 590         } else if (td->o.write_iolog_file)
 591                 ret = init_iolog_write(td);
 592
 593         if (ret)
 594                 td_verror(td, EINVAL, "failed initializing iolog");
 595
 596         return ret;
 597 }
 598
 599 void setup_log(struct io_log **log, struct log_params *p,
 600                const char *filename)
 601 {
 602         struct io_log *l;
 603         int i;
 604         struct io_u_plat_entry *entry;
 605         struct flist_head *list;
 606
 607         l = scalloc(1, sizeof(*l));
 608         INIT_FLIST_HEAD(&l->io_logs);
 609         l->log_type = p->log_type;
 610         l->log_offset = p->log_offset;
 611         l->log_gz = p->log_gz;
 612         l->log_gz_store = p->log_gz_store;
 613         l->avg_msec = p->avg_msec;
 614         l->hist_msec = p->hist_msec;
 615         l->hist_coarseness = p->hist_coarseness;
 616         l->filename = strdup(filename);
 617         l->td = p->td;
 618
 619         /* Initialize histogram lists for each r/w direction,
 620          * with initial io_u_plat of all zeros:
 621          */
 622         for (i = 0; i < DDIR_RWDIR_CNT; i++) {
 623                 list = &l->hist_window[i].list;
 624                 INIT_FLIST_HEAD(list);
 625                 entry = calloc(1, sizeof(struct io_u_plat_entry));
 626                 flist_add(&entry->list, list);
 627         }
 628
 629         if (l->td && l->td->o.io_submit_mode != IO_MODE_OFFLOAD) {
 630                 struct io_logs *p;
 631
 632                 p = calloc(1, sizeof(*l->pending));
 633                 p->max_samples = DEF_LOG_ENTRIES;
 634                 p->log = calloc(p->max_samples, log_entry_sz(l));
 635                 l->pending = p;
 636         }
 637
 638         if (l->log_offset)
 639                 l->log_ddir_mask = LOG_OFFSET_SAMPLE_BIT;
 640
 641         INIT_FLIST_HEAD(&l->chunk_list);
 642
 643         if (l->log_gz && !p->td)
 644                 l->log_gz = 0;
 645         else if (l->log_gz || l->log_gz_store) {
 646                 mutex_init_pshared(&l->chunk_lock);
 647                 mutex_init_pshared(&l->deferred_free_lock);
 648                 p->td->flags |= TD_F_COMPRESS_LOG;
 649         }
 650
 651         *log = l;
 652 }
 653
 654 #ifdef CONFIG_SETVBUF
 655 static void *set_file_buffer(FILE *f)
 656 {
 657         size_t size = 1048576;
 658         void *buf;
 659
 660         buf = malloc(size);
 661         setvbuf(f, buf, _IOFBF, size);
 662         return buf;
 663 }
 664
 665 static void clear_file_buffer(void *buf)
 666 {
 667         free(buf);
 668 }
 669 #else
 670 static void *set_file_buffer(FILE *f)
 671 {
 672         return NULL;
 673 }
 674
 675 static void clear_file_buffer(void *buf)
 676 {
 677 }
 678 #endif
 679
 680 void free_log(struct io_log *log)
 681 {
 682         while (!flist_empty(&log->io_logs)) {
 683                 struct io_logs *cur_log;
 684
 685                 cur_log = flist_first_entry(&log->io_logs, struct io_logs, list);
 686                 flist_del_init(&cur_log->list);
 687                 free(cur_log->log);
 688                 sfree(cur_log);
 689         }
 690
 691         if (log->pending) {
 692                 free(log->pending->log);
 693                 free(log->pending);
 694                 log->pending = NULL;
 695         }
 696
 697         free(log->pending);
 698         free(log->filename);
 699         sfree(log);
 700 }
 701
 702 unsigned long hist_sum(int j, int stride, unsigned int *io_u_plat,
 703                 unsigned int *io_u_plat_last)
 704 {
 705         unsigned long sum;
 706         int k;
 707
 708         if (io_u_plat_last) {
 709                 for (k = sum = 0; k < stride; k++)
 710                         sum += io_u_plat[j + k] - io_u_plat_last[j + k];
 711         } else {
 712                 for (k = sum = 0; k < stride; k++)
 713                         sum += io_u_plat[j + k];
 714         }
 715
 716         return sum;
 717 }
 718
 719 static void flush_hist_samples(FILE *f, int hist_coarseness, void *samples,
 720                                uint64_t sample_size)
 721 {
 722         struct io_sample *s;
 723         int log_offset;
 724         uint64_t i, j, nr_samples;
 725         struct io_u_plat_entry *entry, *entry_before;
 726         unsigned int *io_u_plat;
 727         unsigned int *io_u_plat_before;
 728
 729         int stride = 1 << hist_coarseness;
 730
 731         if (!sample_size)
 732                 return;
 733
 734         s = __get_sample(samples, 0, 0);
 735         log_offset = (s->__ddir & LOG_OFFSET_SAMPLE_BIT) != 0;
 736
 737         nr_samples = sample_size / __log_entry_sz(log_offset);
 738
 739         for (i = 0; i < nr_samples; i++) {
 740                 s = __get_sample(samples, log_offset, i);
 741
 742                 entry = s->data.plat_entry;
 743                 io_u_plat = entry->io_u_plat;
 744
 745                 entry_before = flist_first_entry(&entry->list, struct io_u_plat_entry, list);
 746                 io_u_plat_before = entry_before->io_u_plat;
 747
 748                 fprintf(f, "%lu, %u, %u, ", (unsigned long) s->time,
 749                                                 io_sample_ddir(s), s->bs);
 750                 for (j = 0; j < FIO_IO_U_PLAT_NR - stride; j += stride) {
 751                         fprintf(f, "%lu, ", hist_sum(j, stride, io_u_plat,
 752                                                 io_u_plat_before));
 753                 }
 754                 fprintf(f, "%lu\n", (unsigned long)
 755                         hist_sum(FIO_IO_U_PLAT_NR - stride, stride, io_u_plat,
 756                                         io_u_plat_before));
 757
 758                 flist_del(&entry_before->list);
 759                 free(entry_before);
 760         }
 761 }
 762
 763 void flush_samples(FILE *f, void *samples, uint64_t sample_size)
 764 {
 765         struct io_sample *s;
 766         int log_offset;
 767         uint64_t i, nr_samples;
 768
 769         if (!sample_size)
 770                 return;
 771
 772         s = __get_sample(samples, 0, 0);
 773         log_offset = (s->__ddir & LOG_OFFSET_SAMPLE_BIT) != 0;
 774
 775         nr_samples = sample_size / __log_entry_sz(log_offset);
 776
 777         for (i = 0; i < nr_samples; i++) {
 778                 s = __get_sample(samples, log_offset, i);
 779
 780                 if (!log_offset) {
 781                         fprintf(f, "%lu, %" PRId64 ", %u, %u\n",
 782                                         (unsigned long) s->time,
 783                                         s->data.val,
 784                                         io_sample_ddir(s), s->bs);
 785                 } else {
 786                         struct io_sample_offset *so = (void *) s;
 787
 788                         fprintf(f, "%lu, %" PRId64 ", %u, %u, %llu\n",
 789                                         (unsigned long) s->time,
 790                                         s->data.val,
 791                                         io_sample_ddir(s), s->bs,
 792                                         (unsigned long long) so->offset);
 793                 }
 794         }
 795 }
 796
 797 #ifdef CONFIG_ZLIB
 798
 799 struct iolog_flush_data {
 800         struct workqueue_work work;
 801         struct io_log *log;
 802         void *samples;
 803         uint32_t nr_samples;
 804         bool free;
 805 };
 806
 807 #define GZ_CHUNK        131072
 808
 809 static struct iolog_compress *get_new_chunk(unsigned int seq)
 810 {
 811         struct iolog_compress *c;
 812
 813         c = malloc(sizeof(*c));
 814         INIT_FLIST_HEAD(&c->list);
 815         c->buf = malloc(GZ_CHUNK);
 816         c->len = 0;
 817         c->seq = seq;
 818         return c;
 819 }
 820
 821 static void free_chunk(struct iolog_compress *ic)
 822 {
 823         free(ic->buf);
 824         free(ic);
 825 }
 826
 827 static int z_stream_init(z_stream *stream, int gz_hdr)
 828 {
 829         int wbits = 15;
 830
 831         memset(stream, 0, sizeof(*stream));
 832         stream->zalloc = Z_NULL;
 833         stream->zfree = Z_NULL;
 834         stream->opaque = Z_NULL;
 835         stream->next_in = Z_NULL;
 836
 837         /*
 838          * zlib magic - add 32 for auto-detection of gz header or not,
 839          * if we decide to store files in a gzip friendly format.
 840          */
 841         if (gz_hdr)
 842                 wbits += 32;
 843
 844         if (inflateInit2(stream, wbits) != Z_OK)
 845                 return 1;
 846
 847         return 0;
 848 }
 849
 850 struct inflate_chunk_iter {
 851         unsigned int seq;
 852         int err;
 853         void *buf;
 854         size_t buf_size;
 855         size_t buf_used;
 856         size_t chunk_sz;
 857 };
 858
 859 static void finish_chunk(z_stream *stream, FILE *f,
 860                          struct inflate_chunk_iter *iter)
 861 {
 862         int ret;
 863
 864         ret = inflateEnd(stream);
 865         if (ret != Z_OK)
 866                 log_err("fio: failed to end log inflation seq %d (%d)\n",
 867                                 iter->seq, ret);
 868
 869         flush_samples(f, iter->buf, iter->buf_used);
 870         free(iter->buf);
 871         iter->buf = NULL;
 872         iter->buf_size = iter->buf_used = 0;
 873 }
 874
 875 /*
 876  * Iterative chunk inflation. Handles cases where we cross into a new
 877  * sequence, doing flush finish of previous chunk if needed.
 878  */
 879 static size_t inflate_chunk(struct iolog_compress *ic, int gz_hdr, FILE *f,
 880                             z_stream *stream, struct inflate_chunk_iter *iter)
 881 {
 882         size_t ret;
 883
 884         dprint(FD_COMPRESS, "inflate chunk size=%lu, seq=%u\n",
 885                                 (unsigned long) ic->len, ic->seq);
 886
 887         if (ic->seq != iter->seq) {
 888                 if (iter->seq)
 889                         finish_chunk(stream, f, iter);
 890
 891                 z_stream_init(stream, gz_hdr);
 892                 iter->seq = ic->seq;
 893         }
 894
 895         stream->avail_in = ic->len;
 896         stream->next_in = ic->buf;
 897
 898         if (!iter->buf_size) {
 899                 iter->buf_size = iter->chunk_sz;
 900                 iter->buf = malloc(iter->buf_size);
 901         }
 902
 903         while (stream->avail_in) {
 904                 size_t this_out = iter->buf_size - iter->buf_used;
 905                 int err;
 906
 907                 stream->avail_out = this_out;
 908                 stream->next_out = iter->buf + iter->buf_used;
 909
 910                 err = inflate(stream, Z_NO_FLUSH);
 911                 if (err < 0) {
 912                         log_err("fio: failed inflating log: %d\n", err);
 913                         iter->err = err;
 914                         break;
 915                 }
 916
 917                 iter->buf_used += this_out - stream->avail_out;
 918
 919                 if (!stream->avail_out) {
 920                         iter->buf_size += iter->chunk_sz;
 921                         iter->buf = realloc(iter->buf, iter->buf_size);
 922                         continue;
 923                 }
 924
 925                 if (err == Z_STREAM_END)
 926                         break;
 927         }
 928
 929         ret = (void *) stream->next_in - ic->buf;
 930
 931         dprint(FD_COMPRESS, "inflated to size=%lu\n", (unsigned long) iter->buf_size);
 932
 933         return ret;
 934 }
 935
 936 /*
 937  * Inflate stored compressed chunks, or write them directly to the log
 938  * file if so instructed.
 939  */
 940 static int inflate_gz_chunks(struct io_log *log, FILE *f)
 941 {
 942         struct inflate_chunk_iter iter = { .chunk_sz = log->log_gz, };
 943         z_stream stream;
 944
 945         while (!flist_empty(&log->chunk_list)) {
 946                 struct iolog_compress *ic;
 947
 948                 ic = flist_first_entry(&log->chunk_list, struct iolog_compress, list);
 949                 flist_del(&ic->list);
 950
 951                 if (log->log_gz_store) {
 952                         size_t ret;
 953
 954                         dprint(FD_COMPRESS, "log write chunk size=%lu, "
 955                                 "seq=%u\n", (unsigned long) ic->len, ic->seq);
 956
 957                         ret = fwrite(ic->buf, ic->len, 1, f);
 958                         if (ret != 1 || ferror(f)) {
 959                                 iter.err = errno;
 960                                 log_err("fio: error writing compressed log\n");
 961                         }
 962                 } else
 963                         inflate_chunk(ic, log->log_gz_store, f, &stream, &iter);
 964
 965                 free_chunk(ic);
 966         }
 967
 968         if (iter.seq) {
 969                 finish_chunk(&stream, f, &iter);
 970                 free(iter.buf);
 971         }
 972
 973         return iter.err;
 974 }
 975
 976 /*
 977  * Open compressed log file and decompress the stored chunks and
 978  * write them to stdout. The chunks are stored sequentially in the
 979  * file, so we iterate over them and do them one-by-one.
 980  */
 981 int iolog_file_inflate(const char *file)
 982 {
 983         struct inflate_chunk_iter iter = { .chunk_sz = 64 * 1024 * 1024, };
 984         struct iolog_compress ic;
 985         z_stream stream;
 986         struct stat sb;
 987         ssize_t ret;
 988         size_t total;
 989         void *buf;
 990         FILE *f;
 991
 992         f = fopen(file, "r");
 993         if (!f) {
 994                 perror("fopen");
 995                 return 1;
 996         }
 997
 998         if (stat(file, &sb) < 0) {
 999                 fclose(f);
1000                 perror("stat");
1001                 return 1;
1002         }
1003
1004         ic.buf = buf = malloc(sb.st_size);
1005         ic.len = sb.st_size;
1006         ic.seq = 1;
1007
1008         ret = fread(ic.buf, ic.len, 1, f);
1009         if (ret < 0) {
1010                 perror("fread");
1011                 fclose(f);
1012                 free(buf);
1013                 return 1;
1014         } else if (ret != 1) {
1015                 log_err("fio: short read on reading log\n");
1016                 fclose(f);
1017                 free(buf);
1018                 return 1;
1019         }
1020
1021         fclose(f);
1022
1023         /*
1024          * Each chunk will return Z_STREAM_END. We don't know how many
1025          * chunks are in the file, so we just keep looping and incrementing
1026          * the sequence number until we have consumed the whole compressed
1027          * file.
1028          */
1029         total = ic.len;
1030         do {
1031                 size_t iret;
1032
1033                 iret = inflate_chunk(&ic,  1, stdout, &stream, &iter);
1034                 total -= iret;
1035                 if (!total)
1036                         break;
1037                 if (iter.err)
1038                         break;
1039
1040                 ic.seq++;
1041                 ic.len -= iret;
1042                 ic.buf += iret;
1043         } while (1);
1044
1045         if (iter.seq) {
1046                 finish_chunk(&stream, stdout, &iter);
1047                 free(iter.buf);
1048         }
1049
1050         free(buf);
1051         return iter.err;
1052 }
1053
1054 #else
1055
1056 static int inflate_gz_chunks(struct io_log *log, FILE *f)
1057 {
1058         return 0;
1059 }
1060
1061 int iolog_file_inflate(const char *file)
1062 {
1063         log_err("fio: log inflation not possible without zlib\n");
1064         return 1;
1065 }
1066
1067 #endif
1068
1069 void flush_log(struct io_log *log, bool do_append)
1070 {
1071         void *buf;
1072         FILE *f;
1073
1074         if (!do_append)
1075                 f = fopen(log->filename, "w");
1076         else
1077                 f = fopen(log->filename, "a");
1078         if (!f) {
1079                 perror("fopen log");
1080                 return;
1081         }
1082
1083         buf = set_file_buffer(f);
1084
1085         inflate_gz_chunks(log, f);
1086
1087         while (!flist_empty(&log->io_logs)) {
1088                 struct io_logs *cur_log;
1089
1090                 cur_log = flist_first_entry(&log->io_logs, struct io_logs, list);
1091                 flist_del_init(&cur_log->list);
1092
1093                 if (log->td && log == log->td->clat_hist_log)
1094                         flush_hist_samples(f, log->hist_coarseness, cur_log->log,
1095                                            log_sample_sz(log, cur_log));
1096                 else
1097                         flush_samples(f, cur_log->log, log_sample_sz(log, cur_log));
1098
1099                 sfree(cur_log);
1100         }
1101
1102         fclose(f);
1103         clear_file_buffer(buf);
1104 }
1105
1106 static int finish_log(struct thread_data *td, struct io_log *log, int trylock)
1107 {
1108         if (td->flags & TD_F_COMPRESS_LOG)
1109                 iolog_flush(log);
1110
1111         if (trylock) {
1112                 if (fio_trylock_file(log->filename))
1113                         return 1;
1114         } else
1115                 fio_lock_file(log->filename);
1116
1117         if (td->client_type == FIO_CLIENT_TYPE_GUI || is_backend)
1118                 fio_send_iolog(td, log, log->filename);
1119         else
1120                 flush_log(log, !td->o.per_job_logs);
1121
1122         fio_unlock_file(log->filename);
1123         free_log(log);
1124         return 0;
1125 }
1126
1127 size_t log_chunk_sizes(struct io_log *log)
1128 {
1129         struct flist_head *entry;
1130         size_t ret;
1131
1132         if (flist_empty(&log->chunk_list))
1133                 return 0;
1134
1135         ret = 0;
1136         pthread_mutex_lock(&log->chunk_lock);
1137         flist_for_each(entry, &log->chunk_list) {
1138                 struct iolog_compress *c;
1139
1140                 c = flist_entry(entry, struct iolog_compress, list);
1141                 ret += c->len;
1142         }
1143         pthread_mutex_unlock(&log->chunk_lock);
1144         return ret;
1145 }
1146
1147 #ifdef CONFIG_ZLIB
1148
1149 static bool warned_on_drop;
1150
1151 static void iolog_put_deferred(struct io_log *log, void *ptr)
1152 {
1153         if (!ptr)
1154                 return;
1155
1156         pthread_mutex_lock(&log->deferred_free_lock);
1157         if (log->deferred < IOLOG_MAX_DEFER) {
1158                 log->deferred_items[log->deferred] = ptr;
1159                 log->deferred++;
1160         } else if (!warned_on_drop) {
1161                 log_err("fio: had to drop log entry free\n");
1162                 warned_on_drop = true;
1163         }
1164         pthread_mutex_unlock(&log->deferred_free_lock);
1165 }
1166
1167 static void iolog_free_deferred(struct io_log *log)
1168 {
1169         int i;
1170
1171         if (!log->deferred)
1172                 return;
1173
1174         pthread_mutex_lock(&log->deferred_free_lock);
1175
1176         for (i = 0; i < log->deferred; i++) {
1177                 free(log->deferred_items[i]);
1178                 log->deferred_items[i] = NULL;
1179         }
1180
1181         log->deferred = 0;
1182         pthread_mutex_unlock(&log->deferred_free_lock);
1183 }
1184
1185 static int gz_work(struct iolog_flush_data *data)
1186 {
1187         struct iolog_compress *c = NULL;
1188         struct flist_head list;
1189         unsigned int seq;
1190         z_stream stream;
1191         size_t total = 0;
1192         int ret;
1193
1194         INIT_FLIST_HEAD(&list);
1195
1196         memset(&stream, 0, sizeof(stream));
1197         stream.zalloc = Z_NULL;
1198         stream.zfree = Z_NULL;
1199         stream.opaque = Z_NULL;
1200
1201         ret = deflateInit(&stream, Z_DEFAULT_COMPRESSION);
1202         if (ret != Z_OK) {
1203                 log_err("fio: failed to init gz stream\n");
1204                 goto err;
1205         }
1206
1207         seq = ++data->log->chunk_seq;
1208
1209         stream.next_in = (void *) data->samples;
1210         stream.avail_in = data->nr_samples * log_entry_sz(data->log);
1211
1212         dprint(FD_COMPRESS, "deflate input size=%lu, seq=%u, log=%s\n",
1213                                 (unsigned long) stream.avail_in, seq,
1214                                 data->log->filename);
1215         do {
1216                 if (c)
1217                         dprint(FD_COMPRESS, "seq=%d, chunk=%lu\n", seq,
1218                                 (unsigned long) c->len);
1219                 c = get_new_chunk(seq);
1220                 stream.avail_out = GZ_CHUNK;
1221                 stream.next_out = c->buf;
1222                 ret = deflate(&stream, Z_NO_FLUSH);
1223                 if (ret < 0) {
1224                         log_err("fio: deflate log (%d)\n", ret);
1225                         free_chunk(c);
1226                         goto err;
1227                 }
1228
1229                 c->len = GZ_CHUNK - stream.avail_out;
1230                 flist_add_tail(&c->list, &list);
1231                 total += c->len;
1232         } while (stream.avail_in);
1233
1234         stream.next_out = c->buf + c->len;
1235         stream.avail_out = GZ_CHUNK - c->len;
1236
1237         ret = deflate(&stream, Z_FINISH);
1238         if (ret < 0) {
1239                 /*
1240                  * Z_BUF_ERROR is special, it just means we need more
1241                  * output space. We'll handle that below. Treat any other
1242                  * error as fatal.
1243                  */
1244                 if (ret != Z_BUF_ERROR) {
1245                         log_err("fio: deflate log (%d)\n", ret);
1246                         flist_del(&c->list);
1247                         free_chunk(c);
1248                         goto err;
1249                 }
1250         }
1251
1252         total -= c->len;
1253         c->len = GZ_CHUNK - stream.avail_out;
1254         total += c->len;
1255         dprint(FD_COMPRESS, "seq=%d, chunk=%lu\n", seq, (unsigned long) c->len);
1256
1257         if (ret != Z_STREAM_END) {
1258                 do {
1259                         c = get_new_chunk(seq);
1260                         stream.avail_out = GZ_CHUNK;
1261                         stream.next_out = c->buf;
1262                         ret = deflate(&stream, Z_FINISH);
1263                         c->len = GZ_CHUNK - stream.avail_out;
1264                         total += c->len;
1265                         flist_add_tail(&c->list, &list);
1266                         dprint(FD_COMPRESS, "seq=%d, chunk=%lu\n", seq,
1267                                 (unsigned long) c->len);
1268                 } while (ret != Z_STREAM_END);
1269         }
1270
1271         dprint(FD_COMPRESS, "deflated to size=%lu\n", (unsigned long) total);
1272
1273         ret = deflateEnd(&stream);
1274         if (ret != Z_OK)
1275                 log_err("fio: deflateEnd %d\n", ret);
1276
1277         iolog_put_deferred(data->log, data->samples);
1278
1279         if (!flist_empty(&list)) {
1280                 pthread_mutex_lock(&data->log->chunk_lock);
1281                 flist_splice_tail(&list, &data->log->chunk_list);
1282                 pthread_mutex_unlock(&data->log->chunk_lock);
1283         }
1284
1285         ret = 0;
1286 done:
1287         if (data->free)
1288                 sfree(data);
1289         return ret;
1290 err:
1291         while (!flist_empty(&list)) {
1292                 c = flist_first_entry(list.next, struct iolog_compress, list);
1293                 flist_del(&c->list);
1294                 free_chunk(c);
1295         }
1296         ret = 1;
1297         goto done;
1298 }
1299
1300 /*
1301  * Invoked from our compress helper thread, when logging would have exceeded
1302  * the specified memory limitation. Compresses the previously stored
1303  * entries.
1304  */
1305 static int gz_work_async(struct submit_worker *sw, struct workqueue_work *work)
1306 {
1307         return gz_work(container_of(work, struct iolog_flush_data, work));
1308 }
1309
1310 static int gz_init_worker(struct submit_worker *sw)
1311 {
1312         struct thread_data *td = sw->wq->td;
1313
1314         if (!fio_option_is_set(&td->o, log_gz_cpumask))
1315                 return 0;
1316
1317         if (fio_setaffinity(gettid(), td->o.log_gz_cpumask) == -1) {
1318                 log_err("gz: failed to set CPU affinity\n");
1319                 return 1;
1320         }
1321
1322         return 0;
1323 }
1324
1325 static struct workqueue_ops log_compress_wq_ops = {
1326         .fn             = gz_work_async,
1327         .init_worker_fn = gz_init_worker,
1328         .nice           = 1,
1329 };
1330
1331 int iolog_compress_init(struct thread_data *td, struct sk_out *sk_out)
1332 {
1333         if (!(td->flags & TD_F_COMPRESS_LOG))
1334                 return 0;
1335
1336         workqueue_init(td, &td->log_compress_wq, &log_compress_wq_ops, 1, sk_out);
1337         return 0;
1338 }
1339
1340 void iolog_compress_exit(struct thread_data *td)
1341 {
1342         if (!(td->flags & TD_F_COMPRESS_LOG))
1343                 return;
1344
1345         workqueue_exit(&td->log_compress_wq);
1346 }
1347
1348 /*
1349  * Queue work item to compress the existing log entries. We reset the
1350  * current log to a small size, and reference the existing log in the
1351  * data that we queue for compression. Once compression has been done,
1352  * this old log is freed. If called with finish == true, will not return
1353  * until the log compression has completed, and will flush all previous
1354  * logs too
1355  */
1356 static int iolog_flush(struct io_log *log)
1357 {
1358         struct iolog_flush_data *data;
1359
1360         data = malloc(sizeof(*data));
1361         if (!data)
1362                 return 1;
1363
1364         data->log = log;
1365         data->free = false;
1366
1367         while (!flist_empty(&log->io_logs)) {
1368                 struct io_logs *cur_log;
1369
1370                 cur_log = flist_first_entry(&log->io_logs, struct io_logs, list);
1371                 flist_del_init(&cur_log->list);
1372
1373                 data->samples = cur_log->log;
1374                 data->nr_samples = cur_log->nr_samples;
1375
1376                 sfree(cur_log);
1377
1378                 gz_work(data);
1379         }
1380
1381         free(data);
1382         return 0;
1383 }
1384
1385 int iolog_cur_flush(struct io_log *log, struct io_logs *cur_log)
1386 {
1387         struct iolog_flush_data *data;
1388
1389         data = smalloc(sizeof(*data));
1390         if (!data)
1391                 return 1;
1392
1393         data->log = log;
1394
1395         data->samples = cur_log->log;
1396         data->nr_samples = cur_log->nr_samples;
1397         data->free = true;
1398
1399         cur_log->nr_samples = cur_log->max_samples = 0;
1400         cur_log->log = NULL;
1401
1402         workqueue_enqueue(&log->td->log_compress_wq, &data->work);
1403
1404         iolog_free_deferred(log);
1405
1406         return 0;
1407 }
1408 #else
1409
1410 static int iolog_flush(struct io_log *log)
1411 {
1412         return 1;
1413 }
1414
1415 int iolog_cur_flush(struct io_log *log, struct io_logs *cur_log)
1416 {
1417         return 1;
1418 }
1419
1420 int iolog_compress_init(struct thread_data *td, struct sk_out *sk_out)
1421 {
1422         return 0;
1423 }
1424
1425 void iolog_compress_exit(struct thread_data *td)
1426 {
1427 }
1428
1429 #endif
1430
1431 struct io_logs *iolog_cur_log(struct io_log *log)
1432 {
1433         if (flist_empty(&log->io_logs))
1434                 return NULL;
1435
1436         return flist_last_entry(&log->io_logs, struct io_logs, list);
1437 }
1438
1439 uint64_t iolog_nr_samples(struct io_log *iolog)
1440 {
1441         struct flist_head *entry;
1442         uint64_t ret = 0;
1443
1444         flist_for_each(entry, &iolog->io_logs) {
1445                 struct io_logs *cur_log;
1446
1447                 cur_log = flist_entry(entry, struct io_logs, list);
1448                 ret += cur_log->nr_samples;
1449         }
1450
1451         return ret;
1452 }
1453
1454 static int __write_log(struct thread_data *td, struct io_log *log, int try)
1455 {
1456         if (log)
1457                 return finish_log(td, log, try);
1458
1459         return 0;
1460 }
1461
1462 static int write_iops_log(struct thread_data *td, int try, bool unit_log)
1463 {
1464         int ret;
1465
1466         if (per_unit_log(td->iops_log) != unit_log)
1467                 return 0;
1468
1469         ret = __write_log(td, td->iops_log, try);
1470         if (!ret)
1471                 td->iops_log = NULL;
1472
1473         return ret;
1474 }
1475
1476 static int write_slat_log(struct thread_data *td, int try, bool unit_log)
1477 {
1478         int ret;
1479
1480         if (!unit_log)
1481                 return 0;
1482
1483         ret = __write_log(td, td->slat_log, try);
1484         if (!ret)
1485                 td->slat_log = NULL;
1486
1487         return ret;
1488 }
1489
1490 static int write_clat_log(struct thread_data *td, int try, bool unit_log)
1491 {
1492         int ret;
1493
1494         if (!unit_log)
1495                 return 0;
1496
1497         ret = __write_log(td, td->clat_log, try);
1498         if (!ret)
1499                 td->clat_log = NULL;
1500
1501         return ret;
1502 }
1503
1504 static int write_clat_hist_log(struct thread_data *td, int try, bool unit_log)
1505 {
1506         int ret;
1507
1508         if (!unit_log)
1509                 return 0;
1510
1511         ret = __write_log(td, td->clat_hist_log, try);
1512         if (!ret)
1513                 td->clat_hist_log = NULL;
1514
1515         return ret;
1516 }
1517
1518 static int write_lat_log(struct thread_data *td, int try, bool unit_log)
1519 {
1520         int ret;
1521
1522         if (!unit_log)
1523                 return 0;
1524
1525         ret = __write_log(td, td->lat_log, try);
1526         if (!ret)
1527                 td->lat_log = NULL;
1528
1529         return ret;
1530 }
1531
1532 static int write_bandw_log(struct thread_data *td, int try, bool unit_log)
1533 {
1534         int ret;
1535
1536         if (per_unit_log(td->bw_log) != unit_log)
1537                 return 0;
1538
1539         ret = __write_log(td, td->bw_log, try);
1540         if (!ret)
1541                 td->bw_log = NULL;
1542
1543         return ret;
1544 }
1545
1546 enum {
1547         BW_LOG_MASK     = 1,
1548         LAT_LOG_MASK    = 2,
1549         SLAT_LOG_MASK   = 4,
1550         CLAT_LOG_MASK   = 8,
1551         IOPS_LOG_MASK   = 16,
1552         CLAT_HIST_LOG_MASK = 32,
1553
1554         ALL_LOG_NR      = 6,
1555 };
1556
1557 struct log_type {
1558         unsigned int mask;
1559         int (*fn)(struct thread_data *, int, bool);
1560 };
1561
1562 static struct log_type log_types[] = {
1563         {
1564                 .mask   = BW_LOG_MASK,
1565                 .fn     = write_bandw_log,
1566         },
1567         {
1568                 .mask   = LAT_LOG_MASK,
1569                 .fn     = write_lat_log,
1570         },
1571         {
1572                 .mask   = SLAT_LOG_MASK,
1573                 .fn     = write_slat_log,
1574         },
1575         {
1576                 .mask   = CLAT_LOG_MASK,
1577                 .fn     = write_clat_log,
1578         },
1579         {
1580                 .mask   = IOPS_LOG_MASK,
1581                 .fn     = write_iops_log,
1582         },
1583         {
1584                 .mask   = CLAT_HIST_LOG_MASK,
1585                 .fn     = write_clat_hist_log,
1586         }
1587 };
1588
1589 void td_writeout_logs(struct thread_data *td, bool unit_logs)
1590 {
1591         unsigned int log_mask = 0;
1592         unsigned int log_left = ALL_LOG_NR;
1593         int old_state, i;
1594
1595         old_state = td_bump_runstate(td, TD_FINISHING);
1596
1597         finalize_logs(td, unit_logs);
1598
1599         while (log_left) {
1600                 int prev_log_left = log_left;
1601
1602                 for (i = 0; i < ALL_LOG_NR && log_left; i++) {
1603                         struct log_type *lt = &log_types[i];
1604                         int ret;
1605
1606                         if (!(log_mask & lt->mask)) {
1607                                 ret = lt->fn(td, log_left != 1, unit_logs);
1608                                 if (!ret) {
1609                                         log_left--;
1610                                         log_mask |= lt->mask;
1611                                 }
1612                         }
1613                 }
1614
1615                 if (prev_log_left == log_left)
1616                         usleep(5000);
1617         }
1618
1619         td_restore_runstate(td, old_state);
1620 }
1621
1622 void fio_writeout_logs(bool unit_logs)
1623 {
1624         struct thread_data *td;
1625         int i;
1626
1627         for_each_td(td, i)
1628                 td_writeout_logs(td, unit_logs);
1629 }