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 <assert.h>
   8 #include <sys/types.h>
   9 #include <sys/stat.h>
  10 #include <unistd.h>
  11 #ifdef CONFIG_ZLIB
  12 #include <zlib.h>
  13 #endif
  14
  15 #include "flist.h"
  16 #include "fio.h"
  17 #include "trim.h"
  18 #include "filelock.h"
  19 #include "smalloc.h"
  20 #include "blktrace.h"
  21 #include "pshared.h"
  22
  23 #include <netinet/in.h>
  24 #include <netinet/tcp.h>
  25 #include <arpa/inet.h>
  26 #include <sys/stat.h>
  27 #include <sys/socket.h>
  28 #include <sys/un.h>
  29
  30 static int iolog_flush(struct io_log *log);
  31
  32 static const char iolog_ver2[] = "fio version 2 iolog";
  33
  34 void queue_io_piece(struct thread_data *td, struct io_piece *ipo)
  35 {
  36         flist_add_tail(&ipo->list, &td->io_log_list);
  37         td->total_io_size += ipo->len;
  38 }
  39
  40 void log_io_u(const struct thread_data *td, const struct io_u *io_u)
  41 {
  42         if (!td->o.write_iolog_file)
  43                 return;
  44
  45         fprintf(td->iolog_f, "%s %s %llu %llu\n", io_u->file->file_name,
  46                                                 io_ddir_name(io_u->ddir),
  47                                                 io_u->offset, io_u->buflen);
  48 }
  49
  50 void log_file(struct thread_data *td, struct fio_file *f,
  51               enum file_log_act what)
  52 {
  53         const char *act[] = { "add", "open", "close" };
  54
  55         assert(what < 3);
  56
  57         if (!td->o.write_iolog_file)
  58                 return;
  59
  60
  61         /*
  62          * this happens on the pre-open/close done before the job starts
  63          */
  64         if (!td->iolog_f)
  65                 return;
  66
  67         fprintf(td->iolog_f, "%s %s\n", f->file_name, act[what]);
  68 }
  69
  70 static void iolog_delay(struct thread_data *td, unsigned long delay)
  71 {
  72         uint64_t usec = utime_since_now(&td->last_issue);
  73         unsigned long orig_delay = delay;
  74         uint64_t this_delay;
  75         struct timespec ts;
  76
  77         if (delay < td->time_offset) {
  78                 td->time_offset = 0;
  79                 return;
  80         }
  81
  82         delay -= td->time_offset;
  83         if (delay < usec)
  84                 return;
  85
  86         delay -= usec;
  87
  88         fio_gettime(&ts, NULL);
  89         while (delay && !td->terminate) {
  90                 this_delay = delay;
  91                 if (this_delay > 500000)
  92                         this_delay = 500000;
  93
  94                 usec_sleep(td, this_delay);
  95                 delay -= this_delay;
  96         }
  97
  98         usec = utime_since_now(&ts);
  99         if (usec > orig_delay)
 100                 td->time_offset = usec - orig_delay;
 101         else
 102                 td->time_offset = 0;
 103 }
 104
 105 static int ipo_special(struct thread_data *td, struct io_piece *ipo)
 106 {
 107         struct fio_file *f;
 108         int ret;
 109
 110         /*
 111          * Not a special ipo
 112          */
 113         if (ipo->ddir != DDIR_INVAL)
 114                 return 0;
 115
 116         f = td->files[ipo->fileno];
 117
 118         switch (ipo->file_action) {
 119         case FIO_LOG_OPEN_FILE:
 120                 if (td->o.replay_redirect && fio_file_open(f)) {
 121                         dprint(FD_FILE, "iolog: ignoring re-open of file %s\n",
 122                                         f->file_name);
 123                         break;
 124                 }
 125                 ret = td_io_open_file(td, f);
 126                 if (!ret)
 127                         break;
 128                 td_verror(td, ret, "iolog open file");
 129                 return -1;
 130         case FIO_LOG_CLOSE_FILE:
 131                 td_io_close_file(td, f);
 132                 break;
 133         case FIO_LOG_UNLINK_FILE:
 134                 td_io_unlink_file(td, f);
 135                 break;
 136         default:
 137                 log_err("fio: bad file action %d\n", ipo->file_action);
 138                 break;
 139         }
 140
 141         return 1;
 142 }
 143
 144 static bool read_iolog2(struct thread_data *td);
 145
 146 int read_iolog_get(struct thread_data *td, struct io_u *io_u)
 147 {
 148         struct io_piece *ipo;
 149         unsigned long elapsed;
 150
 151         while (!flist_empty(&td->io_log_list)) {
 152                 int ret;
 153                 if (td->o.read_iolog_chunked) {
 154                         if (td->io_log_checkmark == td->io_log_current) {
 155                                 if (!read_iolog2(td))
 156                                         return 1;
 157                         }
 158                         td->io_log_current--;
 159                 }
 160                 ipo = flist_first_entry(&td->io_log_list, struct io_piece, list);
 161                 flist_del(&ipo->list);
 162                 remove_trim_entry(td, ipo);
 163
 164                 ret = ipo_special(td, ipo);
 165                 if (ret < 0) {
 166                         free(ipo);
 167                         break;
 168                 } else if (ret > 0) {
 169                         free(ipo);
 170                         continue;
 171                 }
 172
 173                 io_u->ddir = ipo->ddir;
 174                 if (ipo->ddir != DDIR_WAIT) {
 175                         io_u->offset = ipo->offset;
 176                         io_u->buflen = ipo->len;
 177                         io_u->file = td->files[ipo->fileno];
 178                         get_file(io_u->file);
 179                         dprint(FD_IO, "iolog: get %llu/%llu/%s\n", io_u->offset,
 180                                                 io_u->buflen, io_u->file->file_name);
 181                         if (ipo->delay)
 182                                 iolog_delay(td, ipo->delay);
 183                 } else {
 184                         elapsed = mtime_since_genesis();
 185                         if (ipo->delay > elapsed)
 186                                 usec_sleep(td, (ipo->delay - elapsed) * 1000);
 187                 }
 188
 189                 free(ipo);
 190
 191                 if (io_u->ddir != DDIR_WAIT)
 192                         return 0;
 193         }
 194
 195         td->done = 1;
 196         return 1;
 197 }
 198
 199 void prune_io_piece_log(struct thread_data *td)
 200 {
 201         struct io_piece *ipo;
 202         struct fio_rb_node *n;
 203
 204         while ((n = rb_first(&td->io_hist_tree)) != NULL) {
 205                 ipo = rb_entry(n, struct io_piece, rb_node);
 206                 rb_erase(n, &td->io_hist_tree);
 207                 remove_trim_entry(td, ipo);
 208                 td->io_hist_len--;
 209                 free(ipo);
 210         }
 211
 212         while (!flist_empty(&td->io_hist_list)) {
 213                 ipo = flist_first_entry(&td->io_hist_list, struct io_piece, list);
 214                 flist_del(&ipo->list);
 215                 remove_trim_entry(td, ipo);
 216                 td->io_hist_len--;
 217                 free(ipo);
 218         }
 219 }
 220
 221 /*
 222  * log a successful write, so we can unwind the log for verify
 223  */
 224 void log_io_piece(struct thread_data *td, struct io_u *io_u)
 225 {
 226         struct fio_rb_node **p, *parent;
 227         struct io_piece *ipo, *__ipo;
 228
 229         ipo = calloc(1, sizeof(struct io_piece));
 230         init_ipo(ipo);
 231         ipo->file = io_u->file;
 232         ipo->offset = io_u->offset;
 233         ipo->len = io_u->buflen;
 234         ipo->numberio = io_u->numberio;
 235         ipo->flags = IP_F_IN_FLIGHT;
 236
 237         io_u->ipo = ipo;
 238
 239         if (io_u_should_trim(td, io_u)) {
 240                 flist_add_tail(&ipo->trim_list, &td->trim_list);
 241                 td->trim_entries++;
 242         }
 243
 244         /*
 245          * Only sort writes if we don't have a random map in which case we need
 246          * to check for duplicate blocks and drop the old one, which we rely on
 247          * the rb insert/lookup for handling.
 248          */
 249         if (file_randommap(td, ipo->file)) {
 250                 INIT_FLIST_HEAD(&ipo->list);
 251                 flist_add_tail(&ipo->list, &td->io_hist_list);
 252                 ipo->flags |= IP_F_ONLIST;
 253                 td->io_hist_len++;
 254                 return;
 255         }
 256
 257         RB_CLEAR_NODE(&ipo->rb_node);
 258
 259         /*
 260          * Sort the entry into the verification list
 261          */
 262 restart:
 263         p = &td->io_hist_tree.rb_node;
 264         parent = NULL;
 265         while (*p) {
 266                 int overlap = 0;
 267                 parent = *p;
 268
 269                 __ipo = rb_entry(parent, struct io_piece, rb_node);
 270                 if (ipo->file < __ipo->file)
 271                         p = &(*p)->rb_left;
 272                 else if (ipo->file > __ipo->file)
 273                         p = &(*p)->rb_right;
 274                 else if (ipo->offset < __ipo->offset) {
 275                         p = &(*p)->rb_left;
 276                         overlap = ipo->offset + ipo->len > __ipo->offset;
 277                 }
 278                 else if (ipo->offset > __ipo->offset) {
 279                         p = &(*p)->rb_right;
 280                         overlap = __ipo->offset + __ipo->len > ipo->offset;
 281                 }
 282                 else
 283                         overlap = 1;
 284
 285                 if (overlap) {
 286                         dprint(FD_IO, "iolog: overlap %llu/%lu, %llu/%lu\n",
 287                                 __ipo->offset, __ipo->len,
 288                                 ipo->offset, ipo->len);
 289                         td->io_hist_len--;
 290                         rb_erase(parent, &td->io_hist_tree);
 291                         remove_trim_entry(td, __ipo);
 292                         if (!(__ipo->flags & IP_F_IN_FLIGHT))
 293                                 free(__ipo);
 294                         goto restart;
 295                 }
 296         }
 297
 298         rb_link_node(&ipo->rb_node, parent, p);
 299         rb_insert_color(&ipo->rb_node, &td->io_hist_tree);
 300         ipo->flags |= IP_F_ONRB;
 301         td->io_hist_len++;
 302 }
 303
 304 void unlog_io_piece(struct thread_data *td, struct io_u *io_u)
 305 {
 306         struct io_piece *ipo = io_u->ipo;
 307
 308         if (td->ts.nr_block_infos) {
 309                 uint32_t *info = io_u_block_info(td, io_u);
 310                 if (BLOCK_INFO_STATE(*info) < BLOCK_STATE_TRIM_FAILURE) {
 311                         if (io_u->ddir == DDIR_TRIM)
 312                                 *info = BLOCK_INFO_SET_STATE(*info,
 313                                                 BLOCK_STATE_TRIM_FAILURE);
 314                         else if (io_u->ddir == DDIR_WRITE)
 315                                 *info = BLOCK_INFO_SET_STATE(*info,
 316                                                 BLOCK_STATE_WRITE_FAILURE);
 317                 }
 318         }
 319
 320         if (!ipo)
 321                 return;
 322
 323         if (ipo->flags & IP_F_ONRB)
 324                 rb_erase(&ipo->rb_node, &td->io_hist_tree);
 325         else if (ipo->flags & IP_F_ONLIST)
 326                 flist_del(&ipo->list);
 327
 328         free(ipo);
 329         io_u->ipo = NULL;
 330         td->io_hist_len--;
 331 }
 332
 333 void trim_io_piece(const struct io_u *io_u)
 334 {
 335         struct io_piece *ipo = io_u->ipo;
 336
 337         if (!ipo)
 338                 return;
 339
 340         ipo->len = io_u->xfer_buflen - io_u->resid;
 341 }
 342
 343 void write_iolog_close(struct thread_data *td)
 344 {
 345         fflush(td->iolog_f);
 346         fclose(td->iolog_f);
 347         free(td->iolog_buf);
 348         td->iolog_f = NULL;
 349         td->iolog_buf = NULL;
 350 }
 351
 352 static int64_t iolog_items_to_fetch(struct thread_data *td)
 353 {
 354         struct timespec now;
 355         uint64_t elapsed;
 356         uint64_t for_1s;
 357         int64_t items_to_fetch;
 358
 359         if (!td->io_log_highmark)
 360                 return 10;
 361
 362
 363         fio_gettime(&now, NULL);
 364         elapsed = ntime_since(&td->io_log_highmark_time, &now);
 365         if (elapsed) {
 366                 for_1s = (td->io_log_highmark - td->io_log_current) * 1000000000 / elapsed;
 367                 items_to_fetch = for_1s - td->io_log_current;
 368                 if (items_to_fetch < 0)
 369                         items_to_fetch = 0;
 370         } else
 371                 items_to_fetch = 0;
 372
 373         td->io_log_highmark = td->io_log_current + items_to_fetch;
 374         td->io_log_checkmark = (td->io_log_highmark + 1) / 2;
 375         fio_gettime(&td->io_log_highmark_time, NULL);
 376
 377         return items_to_fetch;
 378 }
 379
 380 /*
 381  * Read version 2 iolog data. It is enhanced to include per-file logging,
 382  * syncs, etc.
 383  */
 384 static bool read_iolog2(struct thread_data *td)
 385 {
 386         unsigned long long offset;
 387         unsigned int bytes;
 388         int reads, writes, waits, fileno = 0, file_action = 0; /* stupid gcc */
 389         char *rfname, *fname, *act;
 390         char *str, *p;
 391         enum fio_ddir rw;
 392         bool realloc = false;
 393         int64_t items_to_fetch = 0;
 394
 395         if (td->o.read_iolog_chunked) {
 396                 items_to_fetch = iolog_items_to_fetch(td);
 397                 if (!items_to_fetch)
 398                         return true;
 399         }
 400
 401         /*
 402          * Read in the read iolog and store it, reuse the infrastructure
 403          * for doing verifications.
 404          */
 405         str = malloc(4096);
 406         rfname = fname = malloc(256+16);
 407         act = malloc(256+16);
 408
 409         reads = writes = waits = 0;
 410         while ((p = fgets(str, 4096, td->io_log_rfile)) != NULL) {
 411                 struct io_piece *ipo;
 412                 int r;
 413
 414                 r = sscanf(p, "%256s %256s %llu %u", rfname, act, &offset,
 415                                                                         &bytes);
 416
 417                 if (td->o.replay_redirect)
 418                         fname = td->o.replay_redirect;
 419
 420                 if (r == 4) {
 421                         /*
 422                          * Check action first
 423                          */
 424                         if (!strcmp(act, "wait"))
 425                                 rw = DDIR_WAIT;
 426                         else if (!strcmp(act, "read"))
 427                                 rw = DDIR_READ;
 428                         else if (!strcmp(act, "write"))
 429                                 rw = DDIR_WRITE;
 430                         else if (!strcmp(act, "sync"))
 431                                 rw = DDIR_SYNC;
 432                         else if (!strcmp(act, "datasync"))
 433                                 rw = DDIR_DATASYNC;
 434                         else if (!strcmp(act, "trim"))
 435                                 rw = DDIR_TRIM;
 436                         else {
 437                                 log_err("fio: bad iolog file action: %s\n",
 438                                                                         act);
 439                                 continue;
 440                         }
 441                         fileno = get_fileno(td, fname);
 442                 } else if (r == 2) {
 443                         rw = DDIR_INVAL;
 444                         if (!strcmp(act, "add")) {
 445                                 if (td->o.replay_redirect &&
 446                                     get_fileno(td, fname) != -1) {
 447                                         dprint(FD_FILE, "iolog: ignoring"
 448                                                 " re-add of file %s\n", fname);
 449                                 } else {
 450                                         fileno = add_file(td, fname, td->subjob_number, 1);
 451                                         file_action = FIO_LOG_ADD_FILE;
 452                                 }
 453                                 continue;
 454                         } else if (!strcmp(act, "open")) {
 455                                 fileno = get_fileno(td, fname);
 456                                 file_action = FIO_LOG_OPEN_FILE;
 457                         } else if (!strcmp(act, "close")) {
 458                                 fileno = get_fileno(td, fname);
 459                                 file_action = FIO_LOG_CLOSE_FILE;
 460                         } else {
 461                                 log_err("fio: bad iolog file action: %s\n",
 462                                                                         act);
 463                                 continue;
 464                         }
 465                 } else {
 466                         log_err("bad iolog2: %s\n", p);
 467                         continue;
 468                 }
 469
 470                 if (rw == DDIR_READ)
 471                         reads++;
 472                 else if (rw == DDIR_WRITE) {
 473                         /*
 474                          * Don't add a write for ro mode
 475                          */
 476                         if (read_only)
 477                                 continue;
 478                         writes++;
 479                 } else if (rw == DDIR_WAIT) {
 480                         if (td->o.no_stall)
 481                                 continue;
 482                         waits++;
 483                 } else if (rw == DDIR_INVAL) {
 484                 } else if (!ddir_sync(rw)) {
 485                         log_err("bad ddir: %d\n", rw);
 486                         continue;
 487                 }
 488
 489                 /*
 490                  * Make note of file
 491                  */
 492                 ipo = calloc(1, sizeof(*ipo));
 493                 init_ipo(ipo);
 494                 ipo->ddir = rw;
 495                 if (rw == DDIR_WAIT) {
 496                         ipo->delay = offset;
 497                 } else {
 498                         if (td->o.replay_scale)
 499                                 ipo->offset = offset / td->o.replay_scale;
 500                         else
 501                                 ipo->offset = offset;
 502                         ipo_bytes_align(td->o.replay_align, ipo);
 503
 504                         ipo->len = bytes;
 505                         if (rw != DDIR_INVAL && bytes > td->o.max_bs[rw]) {
 506                                 realloc = true;
 507                                 td->o.max_bs[rw] = bytes;
 508                         }
 509                         ipo->fileno = fileno;
 510                         ipo->file_action = file_action;
 511                         td->o.size += bytes;
 512                 }
 513
 514                 queue_io_piece(td, ipo);
 515
 516                 if (td->o.read_iolog_chunked) {
 517                         td->io_log_current++;
 518                         items_to_fetch--;
 519                         if (items_to_fetch == 0)
 520                                 break;
 521                 }
 522         }
 523
 524         free(str);
 525         free(act);
 526         free(rfname);
 527
 528         if (td->o.read_iolog_chunked) {
 529                 td->io_log_highmark = td->io_log_current;
 530                 td->io_log_checkmark = (td->io_log_highmark + 1) / 2;
 531                 fio_gettime(&td->io_log_highmark_time, NULL);
 532         }
 533
 534         if (writes && read_only) {
 535                 log_err("fio: <%s> skips replay of %d writes due to"
 536                         " read-only\n", td->o.name, writes);
 537                 writes = 0;
 538         }
 539
 540         if (td->o.read_iolog_chunked) {
 541                 if (td->io_log_current == 0) {
 542                         return false;
 543                 }
 544                 td->o.td_ddir = TD_DDIR_RW;
 545                 if (realloc && td->orig_buffer)
 546                 {
 547                         io_u_quiesce(td);
 548                         free_io_mem(td);
 549                         init_io_u_buffers(td);
 550                 }
 551                 return true;
 552         }
 553
 554         if (!reads && !writes && !waits)
 555                 return false;
 556         else if (reads && !writes)
 557                 td->o.td_ddir = TD_DDIR_READ;
 558         else if (!reads && writes)
 559                 td->o.td_ddir = TD_DDIR_WRITE;
 560         else
 561                 td->o.td_ddir = TD_DDIR_RW;
 562
 563         return true;
 564 }
 565
 566 static bool is_socket(const char *path)
 567 {
 568         struct stat buf;
 569         int r = stat(path, &buf);
 570         if (r == -1)
 571                 return false;
 572
 573         return S_ISSOCK(buf.st_mode);
 574 }
 575
 576 static int open_socket(const char *path)
 577 {
 578         int fd = socket(AF_UNIX, SOCK_STREAM, 0);
 579         struct sockaddr_un addr;
 580         if (fd < 0)
 581                 return fd;
 582         addr.sun_family = AF_UNIX;
 583         if (snprintf(addr.sun_path, sizeof(addr.sun_path), "%s", path) >=
 584             sizeof(addr.sun_path))
 585                 log_err("%s: path name %s is too long for a Unix socket\n",
 586                         __func__, path);
 587         if (connect(fd, (const struct sockaddr *)&addr, strlen(path) + sizeof(addr.sun_family)) == 0)
 588                 return fd;
 589         else
 590                 close(fd);
 591         return -1;
 592 }
 593
 594 /*
 595  * open iolog, check version, and call appropriate parser
 596  */
 597 static bool init_iolog_read(struct thread_data *td)
 598 {
 599         char buffer[256], *p;
 600         FILE *f = NULL;
 601         bool ret;
 602         char* fname = get_name_by_idx(td->o.read_iolog_file, td->subjob_number);
 603         dprint(FD_IO, "iolog: name=%s\n", fname);
 604
 605         if (is_socket(fname)) {
 606                 int fd = open_socket(fname);
 607                 if (fd >= 0) {
 608                         f = fdopen(fd, "r");
 609                 }
 610         } else
 611                 f = fopen(fname, "r");
 612         free(fname);
 613         if (!f) {
 614                 perror("fopen read iolog");
 615                 return false;
 616         }
 617
 618         p = fgets(buffer, sizeof(buffer), f);
 619         if (!p) {
 620                 td_verror(td, errno, "iolog read");
 621                 log_err("fio: unable to read iolog\n");
 622                 fclose(f);
 623                 return false;
 624         }
 625         td->io_log_rfile = f;
 626         /*
 627          * version 2 of the iolog stores a specific string as the
 628          * first line, check for that
 629          */
 630         if (!strncmp(iolog_ver2, buffer, strlen(iolog_ver2))) {
 631                 free_release_files(td);
 632                 ret = read_iolog2(td);
 633         }
 634         else {
 635                 log_err("fio: iolog version 1 is no longer supported\n");
 636                 ret = false;
 637         }
 638
 639         return ret;
 640 }
 641
 642 /*
 643  * Set up a log for storing io patterns.
 644  */
 645 static bool init_iolog_write(struct thread_data *td)
 646 {
 647         struct fio_file *ff;
 648         FILE *f;
 649         unsigned int i;
 650
 651         f = fopen(td->o.write_iolog_file, "a");
 652         if (!f) {
 653                 perror("fopen write iolog");
 654                 return false;
 655         }
 656
 657         /*
 658          * That's it for writing, setup a log buffer and we're done.
 659           */
 660         td->iolog_f = f;
 661         td->iolog_buf = malloc(8192);
 662         setvbuf(f, td->iolog_buf, _IOFBF, 8192);
 663
 664         /*
 665          * write our version line
 666          */
 667         if (fprintf(f, "%s\n", iolog_ver2) < 0) {
 668                 perror("iolog init\n");
 669                 return false;
 670         }
 671
 672         /*
 673          * add all known files
 674          */
 675         for_each_file(td, ff, i)
 676                 log_file(td, ff, FIO_LOG_ADD_FILE);
 677
 678         return true;
 679 }
 680
 681 bool init_iolog(struct thread_data *td)
 682 {
 683         bool ret;
 684
 685         if (td->o.read_iolog_file) {
 686                 int need_swap;
 687
 688                 /*
 689                  * Check if it's a blktrace file and load that if possible.
 690                  * Otherwise assume it's a normal log file and load that.
 691                  */
 692                 if (is_blktrace(td->o.read_iolog_file, &need_swap))
 693                         ret = load_blktrace(td, td->o.read_iolog_file, need_swap);
 694                 else
 695                         ret = init_iolog_read(td);
 696         } else if (td->o.write_iolog_file)
 697                 ret = init_iolog_write(td);
 698         else
 699                 ret = true;
 700
 701         if (!ret)
 702                 td_verror(td, EINVAL, "failed initializing iolog");
 703
 704         return ret;
 705 }
 706
 707 void setup_log(struct io_log **log, struct log_params *p,
 708                const char *filename)
 709 {
 710         struct io_log *l;
 711         int i;
 712         struct io_u_plat_entry *entry;
 713         struct flist_head *list;
 714
 715         l = scalloc(1, sizeof(*l));
 716         INIT_FLIST_HEAD(&l->io_logs);
 717         l->log_type = p->log_type;
 718         l->log_offset = p->log_offset;
 719         l->log_gz = p->log_gz;
 720         l->log_gz_store = p->log_gz_store;
 721         l->avg_msec = p->avg_msec;
 722         l->hist_msec = p->hist_msec;
 723         l->hist_coarseness = p->hist_coarseness;
 724         l->filename = strdup(filename);
 725         l->td = p->td;
 726
 727         /* Initialize histogram lists for each r/w direction,
 728          * with initial io_u_plat of all zeros:
 729          */
 730         for (i = 0; i < DDIR_RWDIR_CNT; i++) {
 731                 list = &l->hist_window[i].list;
 732                 INIT_FLIST_HEAD(list);
 733                 entry = calloc(1, sizeof(struct io_u_plat_entry));
 734                 flist_add(&entry->list, list);
 735         }
 736
 737         if (l->td && l->td->o.io_submit_mode != IO_MODE_OFFLOAD) {
 738                 struct io_logs *__p;
 739
 740                 __p = calloc(1, sizeof(*l->pending));
 741                 __p->max_samples = DEF_LOG_ENTRIES;
 742                 __p->log = calloc(__p->max_samples, log_entry_sz(l));
 743                 l->pending = __p;
 744         }
 745
 746         if (l->log_offset)
 747                 l->log_ddir_mask = LOG_OFFSET_SAMPLE_BIT;
 748
 749         INIT_FLIST_HEAD(&l->chunk_list);
 750
 751         if (l->log_gz && !p->td)
 752                 l->log_gz = 0;
 753         else if (l->log_gz || l->log_gz_store) {
 754                 mutex_init_pshared(&l->chunk_lock);
 755                 mutex_init_pshared(&l->deferred_free_lock);
 756                 p->td->flags |= TD_F_COMPRESS_LOG;
 757         }
 758
 759         *log = l;
 760 }
 761
 762 #ifdef CONFIG_SETVBUF
 763 static void *set_file_buffer(FILE *f)
 764 {
 765         size_t size = 1048576;
 766         void *buf;
 767
 768         buf = malloc(size);
 769         setvbuf(f, buf, _IOFBF, size);
 770         return buf;
 771 }
 772
 773 static void clear_file_buffer(void *buf)
 774 {
 775         free(buf);
 776 }
 777 #else
 778 static void *set_file_buffer(FILE *f)
 779 {
 780         return NULL;
 781 }
 782
 783 static void clear_file_buffer(void *buf)
 784 {
 785 }
 786 #endif
 787
 788 void free_log(struct io_log *log)
 789 {
 790         while (!flist_empty(&log->io_logs)) {
 791                 struct io_logs *cur_log;
 792
 793                 cur_log = flist_first_entry(&log->io_logs, struct io_logs, list);
 794                 flist_del_init(&cur_log->list);
 795                 free(cur_log->log);
 796                 sfree(cur_log);
 797         }
 798
 799         if (log->pending) {
 800                 free(log->pending->log);
 801                 free(log->pending);
 802                 log->pending = NULL;
 803         }
 804
 805         free(log->pending);
 806         free(log->filename);
 807         sfree(log);
 808 }
 809
 810 uint64_t hist_sum(int j, int stride, uint64_t *io_u_plat,
 811                 uint64_t *io_u_plat_last)
 812 {
 813         uint64_t sum;
 814         int k;
 815
 816         if (io_u_plat_last) {
 817                 for (k = sum = 0; k < stride; k++)
 818                         sum += io_u_plat[j + k] - io_u_plat_last[j + k];
 819         } else {
 820                 for (k = sum = 0; k < stride; k++)
 821                         sum += io_u_plat[j + k];
 822         }
 823
 824         return sum;
 825 }
 826
 827 static void flush_hist_samples(FILE *f, int hist_coarseness, void *samples,
 828                                uint64_t sample_size)
 829 {
 830         struct io_sample *s;
 831         int log_offset;
 832         uint64_t i, j, nr_samples;
 833         struct io_u_plat_entry *entry, *entry_before;
 834         uint64_t *io_u_plat;
 835         uint64_t *io_u_plat_before;
 836
 837         int stride = 1 << hist_coarseness;
 838
 839         if (!sample_size)
 840                 return;
 841
 842         s = __get_sample(samples, 0, 0);
 843         log_offset = (s->__ddir & LOG_OFFSET_SAMPLE_BIT) != 0;
 844
 845         nr_samples = sample_size / __log_entry_sz(log_offset);
 846
 847         for (i = 0; i < nr_samples; i++) {
 848                 s = __get_sample(samples, log_offset, i);
 849
 850                 entry = s->data.plat_entry;
 851                 io_u_plat = entry->io_u_plat;
 852
 853                 entry_before = flist_first_entry(&entry->list, struct io_u_plat_entry, list);
 854                 io_u_plat_before = entry_before->io_u_plat;
 855
 856                 fprintf(f, "%lu, %u, %llu, ", (unsigned long) s->time,
 857                                                 io_sample_ddir(s), (unsigned long long) s->bs);
 858                 for (j = 0; j < FIO_IO_U_PLAT_NR - stride; j += stride) {
 859                         fprintf(f, "%llu, ", (unsigned long long)
 860                                 hist_sum(j, stride, io_u_plat, io_u_plat_before));
 861                 }
 862                 fprintf(f, "%llu\n", (unsigned long long)
 863                         hist_sum(FIO_IO_U_PLAT_NR - stride, stride, io_u_plat,
 864                                         io_u_plat_before));
 865
 866                 flist_del(&entry_before->list);
 867                 free(entry_before);
 868         }
 869 }
 870
 871 void flush_samples(FILE *f, void *samples, uint64_t sample_size)
 872 {
 873         struct io_sample *s;
 874         int log_offset;
 875         uint64_t i, nr_samples;
 876
 877         if (!sample_size)
 878                 return;
 879
 880         s = __get_sample(samples, 0, 0);
 881         log_offset = (s->__ddir & LOG_OFFSET_SAMPLE_BIT) != 0;
 882
 883         nr_samples = sample_size / __log_entry_sz(log_offset);
 884
 885         for (i = 0; i < nr_samples; i++) {
 886                 s = __get_sample(samples, log_offset, i);
 887
 888                 if (!log_offset) {
 889                         fprintf(f, "%lu, %" PRId64 ", %u, %llu\n",
 890                                         (unsigned long) s->time,
 891                                         s->data.val,
 892                                         io_sample_ddir(s), (unsigned long long) s->bs);
 893                 } else {
 894                         struct io_sample_offset *so = (void *) s;
 895
 896                         fprintf(f, "%lu, %" PRId64 ", %u, %llu, %llu\n",
 897                                         (unsigned long) s->time,
 898                                         s->data.val,
 899                                         io_sample_ddir(s), (unsigned long long) s->bs,
 900                                         (unsigned long long) so->offset);
 901                 }
 902         }
 903 }
 904
 905 #ifdef CONFIG_ZLIB
 906
 907 struct iolog_flush_data {
 908         struct workqueue_work work;
 909         struct io_log *log;
 910         void *samples;
 911         uint32_t nr_samples;
 912         bool free;
 913 };
 914
 915 #define GZ_CHUNK        131072
 916
 917 static struct iolog_compress *get_new_chunk(unsigned int seq)
 918 {
 919         struct iolog_compress *c;
 920
 921         c = malloc(sizeof(*c));
 922         INIT_FLIST_HEAD(&c->list);
 923         c->buf = malloc(GZ_CHUNK);
 924         c->len = 0;
 925         c->seq = seq;
 926         return c;
 927 }
 928
 929 static void free_chunk(struct iolog_compress *ic)
 930 {
 931         free(ic->buf);
 932         free(ic);
 933 }
 934
 935 static int z_stream_init(z_stream *stream, int gz_hdr)
 936 {
 937         int wbits = 15;
 938
 939         memset(stream, 0, sizeof(*stream));
 940         stream->zalloc = Z_NULL;
 941         stream->zfree = Z_NULL;
 942         stream->opaque = Z_NULL;
 943         stream->next_in = Z_NULL;
 944
 945         /*
 946          * zlib magic - add 32 for auto-detection of gz header or not,
 947          * if we decide to store files in a gzip friendly format.
 948          */
 949         if (gz_hdr)
 950                 wbits += 32;
 951
 952         if (inflateInit2(stream, wbits) != Z_OK)
 953                 return 1;
 954
 955         return 0;
 956 }
 957
 958 struct inflate_chunk_iter {
 959         unsigned int seq;
 960         int err;
 961         void *buf;
 962         size_t buf_size;
 963         size_t buf_used;
 964         size_t chunk_sz;
 965 };
 966
 967 static void finish_chunk(z_stream *stream, FILE *f,
 968                          struct inflate_chunk_iter *iter)
 969 {
 970         int ret;
 971
 972         ret = inflateEnd(stream);
 973         if (ret != Z_OK)
 974                 log_err("fio: failed to end log inflation seq %d (%d)\n",
 975                                 iter->seq, ret);
 976
 977         flush_samples(f, iter->buf, iter->buf_used);
 978         free(iter->buf);
 979         iter->buf = NULL;
 980         iter->buf_size = iter->buf_used = 0;
 981 }
 982
 983 /*
 984  * Iterative chunk inflation. Handles cases where we cross into a new
 985  * sequence, doing flush finish of previous chunk if needed.
 986  */
 987 static size_t inflate_chunk(struct iolog_compress *ic, int gz_hdr, FILE *f,
 988                             z_stream *stream, struct inflate_chunk_iter *iter)
 989 {
 990         size_t ret;
 991
 992         dprint(FD_COMPRESS, "inflate chunk size=%lu, seq=%u\n",
 993                                 (unsigned long) ic->len, ic->seq);
 994
 995         if (ic->seq != iter->seq) {
 996                 if (iter->seq)
 997                         finish_chunk(stream, f, iter);
 998
 999                 z_stream_init(stream, gz_hdr);
1000                 iter->seq = ic->seq;
1001         }
1002
1003         stream->avail_in = ic->len;
1004         stream->next_in = ic->buf;
1005
1006         if (!iter->buf_size) {
1007                 iter->buf_size = iter->chunk_sz;
1008                 iter->buf = malloc(iter->buf_size);
1009         }
1010
1011         while (stream->avail_in) {
1012                 size_t this_out = iter->buf_size - iter->buf_used;
1013                 int err;
1014
1015                 stream->avail_out = this_out;
1016                 stream->next_out = iter->buf + iter->buf_used;
1017
1018                 err = inflate(stream, Z_NO_FLUSH);
1019                 if (err < 0) {
1020                         log_err("fio: failed inflating log: %d\n", err);
1021                         iter->err = err;
1022                         break;
1023                 }
1024
1025                 iter->buf_used += this_out - stream->avail_out;
1026
1027                 if (!stream->avail_out) {
1028                         iter->buf_size += iter->chunk_sz;
1029                         iter->buf = realloc(iter->buf, iter->buf_size);
1030                         continue;
1031                 }
1032
1033                 if (err == Z_STREAM_END)
1034                         break;
1035         }
1036
1037         ret = (void *) stream->next_in - ic->buf;
1038
1039         dprint(FD_COMPRESS, "inflated to size=%lu\n", (unsigned long) iter->buf_size);
1040
1041         return ret;
1042 }
1043
1044 /*
1045  * Inflate stored compressed chunks, or write them directly to the log
1046  * file if so instructed.
1047  */
1048 static int inflate_gz_chunks(struct io_log *log, FILE *f)
1049 {
1050         struct inflate_chunk_iter iter = { .chunk_sz = log->log_gz, };
1051         z_stream stream;
1052
1053         while (!flist_empty(&log->chunk_list)) {
1054                 struct iolog_compress *ic;
1055
1056                 ic = flist_first_entry(&log->chunk_list, struct iolog_compress, list);
1057                 flist_del(&ic->list);
1058
1059                 if (log->log_gz_store) {
1060                         size_t ret;
1061
1062                         dprint(FD_COMPRESS, "log write chunk size=%lu, "
1063                                 "seq=%u\n", (unsigned long) ic->len, ic->seq);
1064
1065                         ret = fwrite(ic->buf, ic->len, 1, f);
1066                         if (ret != 1 || ferror(f)) {
1067                                 iter.err = errno;
1068                                 log_err("fio: error writing compressed log\n");
1069                         }
1070                 } else
1071                         inflate_chunk(ic, log->log_gz_store, f, &stream, &iter);
1072
1073                 free_chunk(ic);
1074         }
1075
1076         if (iter.seq) {
1077                 finish_chunk(&stream, f, &iter);
1078                 free(iter.buf);
1079         }
1080
1081         return iter.err;
1082 }
1083
1084 /*
1085  * Open compressed log file and decompress the stored chunks and
1086  * write them to stdout. The chunks are stored sequentially in the
1087  * file, so we iterate over them and do them one-by-one.
1088  */
1089 int iolog_file_inflate(const char *file)
1090 {
1091         struct inflate_chunk_iter iter = { .chunk_sz = 64 * 1024 * 1024, };
1092         struct iolog_compress ic;
1093         z_stream stream;
1094         struct stat sb;
1095         size_t ret;
1096         size_t total;
1097         void *buf;
1098         FILE *f;
1099
1100         f = fopen(file, "r");
1101         if (!f) {
1102                 perror("fopen");
1103                 return 1;
1104         }
1105
1106         if (stat(file, &sb) < 0) {
1107                 fclose(f);
1108                 perror("stat");
1109                 return 1;
1110         }
1111
1112         ic.buf = buf = malloc(sb.st_size);
1113         ic.len = sb.st_size;
1114         ic.seq = 1;
1115
1116         ret = fread(ic.buf, ic.len, 1, f);
1117         if (ret == 0 && ferror(f)) {
1118                 perror("fread");
1119                 fclose(f);
1120                 free(buf);
1121                 return 1;
1122         } else if (ferror(f) || (!feof(f) && ret != 1)) {
1123                 log_err("fio: short read on reading log\n");
1124                 fclose(f);
1125                 free(buf);
1126                 return 1;
1127         }
1128
1129         fclose(f);
1130
1131         /*
1132          * Each chunk will return Z_STREAM_END. We don't know how many
1133          * chunks are in the file, so we just keep looping and incrementing
1134          * the sequence number until we have consumed the whole compressed
1135          * file.
1136          */
1137         total = ic.len;
1138         do {
1139                 size_t iret;
1140
1141                 iret = inflate_chunk(&ic,  1, stdout, &stream, &iter);
1142                 total -= iret;
1143                 if (!total)
1144                         break;
1145                 if (iter.err)
1146                         break;
1147
1148                 ic.seq++;
1149                 ic.len -= iret;
1150                 ic.buf += iret;
1151         } while (1);
1152
1153         if (iter.seq) {
1154                 finish_chunk(&stream, stdout, &iter);
1155                 free(iter.buf);
1156         }
1157
1158         free(buf);
1159         return iter.err;
1160 }
1161
1162 #else
1163
1164 static int inflate_gz_chunks(struct io_log *log, FILE *f)
1165 {
1166         return 0;
1167 }
1168
1169 int iolog_file_inflate(const char *file)
1170 {
1171         log_err("fio: log inflation not possible without zlib\n");
1172         return 1;
1173 }
1174
1175 #endif
1176
1177 void flush_log(struct io_log *log, bool do_append)
1178 {
1179         void *buf;
1180         FILE *f;
1181
1182         if (!do_append)
1183                 f = fopen(log->filename, "w");
1184         else
1185                 f = fopen(log->filename, "a");
1186         if (!f) {
1187                 perror("fopen log");
1188                 return;
1189         }
1190
1191         buf = set_file_buffer(f);
1192
1193         inflate_gz_chunks(log, f);
1194
1195         while (!flist_empty(&log->io_logs)) {
1196                 struct io_logs *cur_log;
1197
1198                 cur_log = flist_first_entry(&log->io_logs, struct io_logs, list);
1199                 flist_del_init(&cur_log->list);
1200
1201                 if (log->td && log == log->td->clat_hist_log)
1202                         flush_hist_samples(f, log->hist_coarseness, cur_log->log,
1203                                            log_sample_sz(log, cur_log));
1204                 else
1205                         flush_samples(f, cur_log->log, log_sample_sz(log, cur_log));
1206
1207                 sfree(cur_log);
1208         }
1209
1210         fclose(f);
1211         clear_file_buffer(buf);
1212 }
1213
1214 static int finish_log(struct thread_data *td, struct io_log *log, int trylock)
1215 {
1216         if (td->flags & TD_F_COMPRESS_LOG)
1217                 iolog_flush(log);
1218
1219         if (trylock) {
1220                 if (fio_trylock_file(log->filename))
1221                         return 1;
1222         } else
1223                 fio_lock_file(log->filename);
1224
1225         if (td->client_type == FIO_CLIENT_TYPE_GUI || is_backend)
1226                 fio_send_iolog(td, log, log->filename);
1227         else
1228                 flush_log(log, !td->o.per_job_logs);
1229
1230         fio_unlock_file(log->filename);
1231         free_log(log);
1232         return 0;
1233 }
1234
1235 size_t log_chunk_sizes(struct io_log *log)
1236 {
1237         struct flist_head *entry;
1238         size_t ret;
1239
1240         if (flist_empty(&log->chunk_list))
1241                 return 0;
1242
1243         ret = 0;
1244         pthread_mutex_lock(&log->chunk_lock);
1245         flist_for_each(entry, &log->chunk_list) {
1246                 struct iolog_compress *c;
1247
1248                 c = flist_entry(entry, struct iolog_compress, list);
1249                 ret += c->len;
1250         }
1251         pthread_mutex_unlock(&log->chunk_lock);
1252         return ret;
1253 }
1254
1255 #ifdef CONFIG_ZLIB
1256
1257 static void iolog_put_deferred(struct io_log *log, void *ptr)
1258 {
1259         if (!ptr)
1260                 return;
1261
1262         pthread_mutex_lock(&log->deferred_free_lock);
1263         if (log->deferred < IOLOG_MAX_DEFER) {
1264                 log->deferred_items[log->deferred] = ptr;
1265                 log->deferred++;
1266         } else if (!fio_did_warn(FIO_WARN_IOLOG_DROP))
1267                 log_err("fio: had to drop log entry free\n");
1268         pthread_mutex_unlock(&log->deferred_free_lock);
1269 }
1270
1271 static void iolog_free_deferred(struct io_log *log)
1272 {
1273         int i;
1274
1275         if (!log->deferred)
1276                 return;
1277
1278         pthread_mutex_lock(&log->deferred_free_lock);
1279
1280         for (i = 0; i < log->deferred; i++) {
1281                 free(log->deferred_items[i]);
1282                 log->deferred_items[i] = NULL;
1283         }
1284
1285         log->deferred = 0;
1286         pthread_mutex_unlock(&log->deferred_free_lock);
1287 }
1288
1289 static int gz_work(struct iolog_flush_data *data)
1290 {
1291         struct iolog_compress *c = NULL;
1292         struct flist_head list;
1293         unsigned int seq;
1294         z_stream stream;
1295         size_t total = 0;
1296         int ret;
1297
1298         INIT_FLIST_HEAD(&list);
1299
1300         memset(&stream, 0, sizeof(stream));
1301         stream.zalloc = Z_NULL;
1302         stream.zfree = Z_NULL;
1303         stream.opaque = Z_NULL;
1304
1305         ret = deflateInit(&stream, Z_DEFAULT_COMPRESSION);
1306         if (ret != Z_OK) {
1307                 log_err("fio: failed to init gz stream\n");
1308                 goto err;
1309         }
1310
1311         seq = ++data->log->chunk_seq;
1312
1313         stream.next_in = (void *) data->samples;
1314         stream.avail_in = data->nr_samples * log_entry_sz(data->log);
1315
1316         dprint(FD_COMPRESS, "deflate input size=%lu, seq=%u, log=%s\n",
1317                                 (unsigned long) stream.avail_in, seq,
1318                                 data->log->filename);
1319         do {
1320                 if (c)
1321                         dprint(FD_COMPRESS, "seq=%d, chunk=%lu\n", seq,
1322                                 (unsigned long) c->len);
1323                 c = get_new_chunk(seq);
1324                 stream.avail_out = GZ_CHUNK;
1325                 stream.next_out = c->buf;
1326                 ret = deflate(&stream, Z_NO_FLUSH);
1327                 if (ret < 0) {
1328                         log_err("fio: deflate log (%d)\n", ret);
1329                         free_chunk(c);
1330                         goto err;
1331                 }
1332
1333                 c->len = GZ_CHUNK - stream.avail_out;
1334                 flist_add_tail(&c->list, &list);
1335                 total += c->len;
1336         } while (stream.avail_in);
1337
1338         stream.next_out = c->buf + c->len;
1339         stream.avail_out = GZ_CHUNK - c->len;
1340
1341         ret = deflate(&stream, Z_FINISH);
1342         if (ret < 0) {
1343                 /*
1344                  * Z_BUF_ERROR is special, it just means we need more
1345                  * output space. We'll handle that below. Treat any other
1346                  * error as fatal.
1347                  */
1348                 if (ret != Z_BUF_ERROR) {
1349                         log_err("fio: deflate log (%d)\n", ret);
1350                         flist_del(&c->list);
1351                         free_chunk(c);
1352                         goto err;
1353                 }
1354         }
1355
1356         total -= c->len;
1357         c->len = GZ_CHUNK - stream.avail_out;
1358         total += c->len;
1359         dprint(FD_COMPRESS, "seq=%d, chunk=%lu\n", seq, (unsigned long) c->len);
1360
1361         if (ret != Z_STREAM_END) {
1362                 do {
1363                         c = get_new_chunk(seq);
1364                         stream.avail_out = GZ_CHUNK;
1365                         stream.next_out = c->buf;
1366                         ret = deflate(&stream, Z_FINISH);
1367                         c->len = GZ_CHUNK - stream.avail_out;
1368                         total += c->len;
1369                         flist_add_tail(&c->list, &list);
1370                         dprint(FD_COMPRESS, "seq=%d, chunk=%lu\n", seq,
1371                                 (unsigned long) c->len);
1372                 } while (ret != Z_STREAM_END);
1373         }
1374
1375         dprint(FD_COMPRESS, "deflated to size=%lu\n", (unsigned long) total);
1376
1377         ret = deflateEnd(&stream);
1378         if (ret != Z_OK)
1379                 log_err("fio: deflateEnd %d\n", ret);
1380
1381         iolog_put_deferred(data->log, data->samples);
1382
1383         if (!flist_empty(&list)) {
1384                 pthread_mutex_lock(&data->log->chunk_lock);
1385                 flist_splice_tail(&list, &data->log->chunk_list);
1386                 pthread_mutex_unlock(&data->log->chunk_lock);
1387         }
1388
1389         ret = 0;
1390 done:
1391         if (data->free)
1392                 sfree(data);
1393         return ret;
1394 err:
1395         while (!flist_empty(&list)) {
1396                 c = flist_first_entry(list.next, struct iolog_compress, list);
1397                 flist_del(&c->list);
1398                 free_chunk(c);
1399         }
1400         ret = 1;
1401         goto done;
1402 }
1403
1404 /*
1405  * Invoked from our compress helper thread, when logging would have exceeded
1406  * the specified memory limitation. Compresses the previously stored
1407  * entries.
1408  */
1409 static int gz_work_async(struct submit_worker *sw, struct workqueue_work *work)
1410 {
1411         return gz_work(container_of(work, struct iolog_flush_data, work));
1412 }
1413
1414 static int gz_init_worker(struct submit_worker *sw)
1415 {
1416         struct thread_data *td = sw->wq->td;
1417
1418         if (!fio_option_is_set(&td->o, log_gz_cpumask))
1419                 return 0;
1420
1421         if (fio_setaffinity(gettid(), td->o.log_gz_cpumask) == -1) {
1422                 log_err("gz: failed to set CPU affinity\n");
1423                 return 1;
1424         }
1425
1426         return 0;
1427 }
1428
1429 static struct workqueue_ops log_compress_wq_ops = {
1430         .fn             = gz_work_async,
1431         .init_worker_fn = gz_init_worker,
1432         .nice           = 1,
1433 };
1434
1435 int iolog_compress_init(struct thread_data *td, struct sk_out *sk_out)
1436 {
1437         if (!(td->flags & TD_F_COMPRESS_LOG))
1438                 return 0;
1439
1440         workqueue_init(td, &td->log_compress_wq, &log_compress_wq_ops, 1, sk_out);
1441         return 0;
1442 }
1443
1444 void iolog_compress_exit(struct thread_data *td)
1445 {
1446         if (!(td->flags & TD_F_COMPRESS_LOG))
1447                 return;
1448
1449         workqueue_exit(&td->log_compress_wq);
1450 }
1451
1452 /*
1453  * Queue work item to compress the existing log entries. We reset the
1454  * current log to a small size, and reference the existing log in the
1455  * data that we queue for compression. Once compression has been done,
1456  * this old log is freed. If called with finish == true, will not return
1457  * until the log compression has completed, and will flush all previous
1458  * logs too
1459  */
1460 static int iolog_flush(struct io_log *log)
1461 {
1462         struct iolog_flush_data *data;
1463
1464         data = malloc(sizeof(*data));
1465         if (!data)
1466                 return 1;
1467
1468         data->log = log;
1469         data->free = false;
1470
1471         while (!flist_empty(&log->io_logs)) {
1472                 struct io_logs *cur_log;
1473
1474                 cur_log = flist_first_entry(&log->io_logs, struct io_logs, list);
1475                 flist_del_init(&cur_log->list);
1476
1477                 data->samples = cur_log->log;
1478                 data->nr_samples = cur_log->nr_samples;
1479
1480                 sfree(cur_log);
1481
1482                 gz_work(data);
1483         }
1484
1485         free(data);
1486         return 0;
1487 }
1488
1489 int iolog_cur_flush(struct io_log *log, struct io_logs *cur_log)
1490 {
1491         struct iolog_flush_data *data;
1492
1493         data = smalloc(sizeof(*data));
1494         if (!data)
1495                 return 1;
1496
1497         data->log = log;
1498
1499         data->samples = cur_log->log;
1500         data->nr_samples = cur_log->nr_samples;
1501         data->free = true;
1502
1503         cur_log->nr_samples = cur_log->max_samples = 0;
1504         cur_log->log = NULL;
1505
1506         workqueue_enqueue(&log->td->log_compress_wq, &data->work);
1507
1508         iolog_free_deferred(log);
1509
1510         return 0;
1511 }
1512 #else
1513
1514 static int iolog_flush(struct io_log *log)
1515 {
1516         return 1;
1517 }
1518
1519 int iolog_cur_flush(struct io_log *log, struct io_logs *cur_log)
1520 {
1521         return 1;
1522 }
1523
1524 int iolog_compress_init(struct thread_data *td, struct sk_out *sk_out)
1525 {
1526         return 0;
1527 }
1528
1529 void iolog_compress_exit(struct thread_data *td)
1530 {
1531 }
1532
1533 #endif
1534
1535 struct io_logs *iolog_cur_log(struct io_log *log)
1536 {
1537         if (flist_empty(&log->io_logs))
1538                 return NULL;
1539
1540         return flist_last_entry(&log->io_logs, struct io_logs, list);
1541 }
1542
1543 uint64_t iolog_nr_samples(struct io_log *iolog)
1544 {
1545         struct flist_head *entry;
1546         uint64_t ret = 0;
1547
1548         flist_for_each(entry, &iolog->io_logs) {
1549                 struct io_logs *cur_log;
1550
1551                 cur_log = flist_entry(entry, struct io_logs, list);
1552                 ret += cur_log->nr_samples;
1553         }
1554
1555         return ret;
1556 }
1557
1558 static int __write_log(struct thread_data *td, struct io_log *log, int try)
1559 {
1560         if (log)
1561                 return finish_log(td, log, try);
1562
1563         return 0;
1564 }
1565
1566 static int write_iops_log(struct thread_data *td, int try, bool unit_log)
1567 {
1568         int ret;
1569
1570         if (per_unit_log(td->iops_log) != unit_log)
1571                 return 0;
1572
1573         ret = __write_log(td, td->iops_log, try);
1574         if (!ret)
1575                 td->iops_log = NULL;
1576
1577         return ret;
1578 }
1579
1580 static int write_slat_log(struct thread_data *td, int try, bool unit_log)
1581 {
1582         int ret;
1583
1584         if (!unit_log)
1585                 return 0;
1586
1587         ret = __write_log(td, td->slat_log, try);
1588         if (!ret)
1589                 td->slat_log = NULL;
1590
1591         return ret;
1592 }
1593
1594 static int write_clat_log(struct thread_data *td, int try, bool unit_log)
1595 {
1596         int ret;
1597
1598         if (!unit_log)
1599                 return 0;
1600
1601         ret = __write_log(td, td->clat_log, try);
1602         if (!ret)
1603                 td->clat_log = NULL;
1604
1605         return ret;
1606 }
1607
1608 static int write_clat_hist_log(struct thread_data *td, int try, bool unit_log)
1609 {
1610         int ret;
1611
1612         if (!unit_log)
1613                 return 0;
1614
1615         ret = __write_log(td, td->clat_hist_log, try);
1616         if (!ret)
1617                 td->clat_hist_log = NULL;
1618
1619         return ret;
1620 }
1621
1622 static int write_lat_log(struct thread_data *td, int try, bool unit_log)
1623 {
1624         int ret;
1625
1626         if (!unit_log)
1627                 return 0;
1628
1629         ret = __write_log(td, td->lat_log, try);
1630         if (!ret)
1631                 td->lat_log = NULL;
1632
1633         return ret;
1634 }
1635
1636 static int write_bandw_log(struct thread_data *td, int try, bool unit_log)
1637 {
1638         int ret;
1639
1640         if (per_unit_log(td->bw_log) != unit_log)
1641                 return 0;
1642
1643         ret = __write_log(td, td->bw_log, try);
1644         if (!ret)
1645                 td->bw_log = NULL;
1646
1647         return ret;
1648 }
1649
1650 enum {
1651         BW_LOG_MASK     = 1,
1652         LAT_LOG_MASK    = 2,
1653         SLAT_LOG_MASK   = 4,
1654         CLAT_LOG_MASK   = 8,
1655         IOPS_LOG_MASK   = 16,
1656         CLAT_HIST_LOG_MASK = 32,
1657
1658         ALL_LOG_NR      = 6,
1659 };
1660
1661 struct log_type {
1662         unsigned int mask;
1663         int (*fn)(struct thread_data *, int, bool);
1664 };
1665
1666 static struct log_type log_types[] = {
1667         {
1668                 .mask   = BW_LOG_MASK,
1669                 .fn     = write_bandw_log,
1670         },
1671         {
1672                 .mask   = LAT_LOG_MASK,
1673                 .fn     = write_lat_log,
1674         },
1675         {
1676                 .mask   = SLAT_LOG_MASK,
1677                 .fn     = write_slat_log,
1678         },
1679         {
1680                 .mask   = CLAT_LOG_MASK,
1681                 .fn     = write_clat_log,
1682         },
1683         {
1684                 .mask   = IOPS_LOG_MASK,
1685                 .fn     = write_iops_log,
1686         },
1687         {
1688                 .mask   = CLAT_HIST_LOG_MASK,
1689                 .fn     = write_clat_hist_log,
1690         }
1691 };
1692
1693 void td_writeout_logs(struct thread_data *td, bool unit_logs)
1694 {
1695         unsigned int log_mask = 0;
1696         unsigned int log_left = ALL_LOG_NR;
1697         int old_state, i;
1698
1699         old_state = td_bump_runstate(td, TD_FINISHING);
1700
1701         finalize_logs(td, unit_logs);
1702
1703         while (log_left) {
1704                 int prev_log_left = log_left;
1705
1706                 for (i = 0; i < ALL_LOG_NR && log_left; i++) {
1707                         struct log_type *lt = &log_types[i];
1708                         int ret;
1709
1710                         if (!(log_mask & lt->mask)) {
1711                                 ret = lt->fn(td, log_left != 1, unit_logs);
1712                                 if (!ret) {
1713                                         log_left--;
1714                                         log_mask |= lt->mask;
1715                                 }
1716                         }
1717                 }
1718
1719                 if (prev_log_left == log_left)
1720                         usleep(5000);
1721         }
1722
1723         td_restore_runstate(td, old_state);
1724 }
1725
1726 void fio_writeout_logs(bool unit_logs)
1727 {
1728         struct thread_data *td;
1729         int i;
1730
1731         for_each_td(td, i)
1732                 td_writeout_logs(td, unit_logs);
1733 }