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