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         int64_t items_to_fetch = 0;
 393
 394         if (td->o.read_iolog_chunked) {
 395                 items_to_fetch = iolog_items_to_fetch(td);
 396                 if (!items_to_fetch)
 397                         return true;
 398         }
 399
 400         /*
 401          * Read in the read iolog and store it, reuse the infrastructure
 402          * for doing verifications.
 403          */
 404         str = malloc(4096);
 405         rfname = fname = malloc(256+16);
 406         act = malloc(256+16);
 407
 408         reads = writes = waits = 0;
 409         while ((p = fgets(str, 4096, td->io_log_rfile)) != NULL) {
 410                 struct io_piece *ipo;
 411                 int r;
 412
 413                 r = sscanf(p, "%256s %256s %llu %u", rfname, act, &offset,
 414                                                                         &bytes);
 415
 416                 if (td->o.replay_redirect)
 417                         fname = td->o.replay_redirect;
 418
 419                 if (r == 4) {
 420                         /*
 421                          * Check action first
 422                          */
 423                         if (!strcmp(act, "wait"))
 424                                 rw = DDIR_WAIT;
 425                         else if (!strcmp(act, "read"))
 426                                 rw = DDIR_READ;
 427                         else if (!strcmp(act, "write"))
 428                                 rw = DDIR_WRITE;
 429                         else if (!strcmp(act, "sync"))
 430                                 rw = DDIR_SYNC;
 431                         else if (!strcmp(act, "datasync"))
 432                                 rw = DDIR_DATASYNC;
 433                         else if (!strcmp(act, "trim"))
 434                                 rw = DDIR_TRIM;
 435                         else {
 436                                 log_err("fio: bad iolog file action: %s\n",
 437                                                                         act);
 438                                 continue;
 439                         }
 440                         fileno = get_fileno(td, fname);
 441                 } else if (r == 2) {
 442                         rw = DDIR_INVAL;
 443                         if (!strcmp(act, "add")) {
 444                                 if (td->o.replay_redirect &&
 445                                     get_fileno(td, fname) != -1) {
 446                                         dprint(FD_FILE, "iolog: ignoring"
 447                                                 " re-add of file %s\n", fname);
 448                                 } else {
 449                                         fileno = add_file(td, fname, 0, 1);
 450                                         file_action = FIO_LOG_ADD_FILE;
 451                                 }
 452                                 continue;
 453                         } else if (!strcmp(act, "open")) {
 454                                 fileno = get_fileno(td, fname);
 455                                 file_action = FIO_LOG_OPEN_FILE;
 456                         } else if (!strcmp(act, "close")) {
 457                                 fileno = get_fileno(td, fname);
 458                                 file_action = FIO_LOG_CLOSE_FILE;
 459                         } else {
 460                                 log_err("fio: bad iolog file action: %s\n",
 461                                                                         act);
 462                                 continue;
 463                         }
 464                 } else {
 465                         log_err("bad iolog2: %s\n", p);
 466                         continue;
 467                 }
 468
 469                 if (rw == DDIR_READ)
 470                         reads++;
 471                 else if (rw == DDIR_WRITE) {
 472                         /*
 473                          * Don't add a write for ro mode
 474                          */
 475                         if (read_only)
 476                                 continue;
 477                         writes++;
 478                 } else if (rw == DDIR_WAIT) {
 479                         if (td->o.no_stall)
 480                                 continue;
 481                         waits++;
 482                 } else if (rw == DDIR_INVAL) {
 483                 } else if (!ddir_sync(rw)) {
 484                         log_err("bad ddir: %d\n", rw);
 485                         continue;
 486                 }
 487
 488                 /*
 489                  * Make note of file
 490                  */
 491                 ipo = calloc(1, sizeof(*ipo));
 492                 init_ipo(ipo);
 493                 ipo->ddir = rw;
 494                 if (rw == DDIR_WAIT) {
 495                         ipo->delay = offset;
 496                 } else {
 497                         if (td->o.replay_scale)
 498                                 ipo->offset = offset / td->o.replay_scale;
 499                         else
 500                                 ipo->offset = offset;
 501                         ipo_bytes_align(td->o.replay_align, ipo);
 502
 503                         ipo->len = bytes;
 504                         if (rw != DDIR_INVAL && bytes > td->o.max_bs[rw])
 505                                 td->o.max_bs[rw] = bytes;
 506                         ipo->fileno = fileno;
 507                         ipo->file_action = file_action;
 508                         td->o.size += bytes;
 509                 }
 510
 511                 queue_io_piece(td, ipo);
 512
 513                 if (td->o.read_iolog_chunked) {
 514                         td->io_log_current++;
 515                         items_to_fetch--;
 516                         if (items_to_fetch == 0)
 517                                 break;
 518                 }
 519         }
 520
 521         free(str);
 522         free(act);
 523         free(rfname);
 524
 525         if (td->o.read_iolog_chunked) {
 526                 td->io_log_highmark = td->io_log_current;
 527                 td->io_log_checkmark = (td->io_log_highmark + 1) / 2;
 528                 fio_gettime(&td->io_log_highmark_time, NULL);
 529         }
 530
 531         if (writes && read_only) {
 532                 log_err("fio: <%s> skips replay of %d writes due to"
 533                         " read-only\n", td->o.name, writes);
 534                 writes = 0;
 535         }
 536
 537         if (td->o.read_iolog_chunked) {
 538                 if (td->io_log_current == 0) {
 539                         return false;
 540                 }
 541                 td->o.td_ddir = TD_DDIR_RW;
 542                 return true;
 543         }
 544
 545         if (!reads && !writes && !waits)
 546                 return false;
 547         else if (reads && !writes)
 548                 td->o.td_ddir = TD_DDIR_READ;
 549         else if (!reads && writes)
 550                 td->o.td_ddir = TD_DDIR_WRITE;
 551         else
 552                 td->o.td_ddir = TD_DDIR_RW;
 553
 554         return true;
 555 }
 556
 557 static bool is_socket(const char *path)
 558 {
 559         struct stat buf;
 560         int r = stat(path, &buf);
 561         if (r == -1)
 562                 return false;
 563
 564         return S_ISSOCK(buf.st_mode);
 565 }
 566
 567 static int open_socket(const char *path)
 568 {
 569         int fd = socket(AF_UNIX, SOCK_STREAM, 0);
 570         struct sockaddr_un addr;
 571         if (fd < 0)
 572                 return fd;
 573         addr.sun_family = AF_UNIX;
 574         strncpy(addr.sun_path, path, sizeof(addr.sun_path));
 575         if (connect(fd, (const struct sockaddr *)&addr, strlen(path) + sizeof(addr.sun_family)) == 0)
 576                 return fd;
 577         else
 578                 close(fd);
 579         return -1;
 580 }
 581
 582 /*
 583  * open iolog, check version, and call appropriate parser
 584  */
 585 static bool init_iolog_read(struct thread_data *td)
 586 {
 587         char buffer[256], *p;
 588         FILE *f = NULL;
 589         bool ret;
 590         if (is_socket(td->o.read_iolog_file)) {
 591                 int fd = open_socket(td->o.read_iolog_file);
 592                 if (fd >= 0) {
 593                         f = fdopen(fd, "r");
 594                 }
 595         } else
 596                 f = fopen(td->o.read_iolog_file, "r");
 597         if (!f) {
 598                 perror("fopen read iolog");
 599                 return false;
 600         }
 601
 602         p = fgets(buffer, sizeof(buffer), f);
 603         if (!p) {
 604                 td_verror(td, errno, "iolog read");
 605                 log_err("fio: unable to read iolog\n");
 606                 fclose(f);
 607                 return false;
 608         }
 609         td->io_log_rfile = f;
 610         /*
 611          * version 2 of the iolog stores a specific string as the
 612          * first line, check for that
 613          */
 614         if (!strncmp(iolog_ver2, buffer, strlen(iolog_ver2))) {
 615                 free_release_files(td);
 616                 ret = read_iolog2(td);
 617         }
 618         else {
 619                 log_err("fio: iolog version 1 is no longer supported\n");
 620                 ret = false;
 621         }
 622
 623         return ret;
 624 }
 625
 626 /*
 627  * Set up a log for storing io patterns.
 628  */
 629 static bool init_iolog_write(struct thread_data *td)
 630 {
 631         struct fio_file *ff;
 632         FILE *f;
 633         unsigned int i;
 634
 635         f = fopen(td->o.write_iolog_file, "a");
 636         if (!f) {
 637                 perror("fopen write iolog");
 638                 return false;
 639         }
 640
 641         /*
 642          * That's it for writing, setup a log buffer and we're done.
 643           */
 644         td->iolog_f = f;
 645         td->iolog_buf = malloc(8192);
 646         setvbuf(f, td->iolog_buf, _IOFBF, 8192);
 647
 648         /*
 649          * write our version line
 650          */
 651         if (fprintf(f, "%s\n", iolog_ver2) < 0) {
 652                 perror("iolog init\n");
 653                 return false;
 654         }
 655
 656         /*
 657          * add all known files
 658          */
 659         for_each_file(td, ff, i)
 660                 log_file(td, ff, FIO_LOG_ADD_FILE);
 661
 662         return true;
 663 }
 664
 665 bool init_iolog(struct thread_data *td)
 666 {
 667         bool ret;
 668
 669         if (td->o.read_iolog_file) {
 670                 int need_swap;
 671
 672                 /*
 673                  * Check if it's a blktrace file and load that if possible.
 674                  * Otherwise assume it's a normal log file and load that.
 675                  */
 676                 if (is_blktrace(td->o.read_iolog_file, &need_swap))
 677                         ret = load_blktrace(td, td->o.read_iolog_file, need_swap);
 678                 else
 679                         ret = init_iolog_read(td);
 680         } else if (td->o.write_iolog_file)
 681                 ret = init_iolog_write(td);
 682         else
 683                 ret = true;
 684
 685         if (!ret)
 686                 td_verror(td, EINVAL, "failed initializing iolog");
 687
 688         return ret;
 689 }
 690
 691 void setup_log(struct io_log **log, struct log_params *p,
 692                const char *filename)
 693 {
 694         struct io_log *l;
 695         int i;
 696         struct io_u_plat_entry *entry;
 697         struct flist_head *list;
 698
 699         l = scalloc(1, sizeof(*l));
 700         INIT_FLIST_HEAD(&l->io_logs);
 701         l->log_type = p->log_type;
 702         l->log_offset = p->log_offset;
 703         l->log_gz = p->log_gz;
 704         l->log_gz_store = p->log_gz_store;
 705         l->avg_msec = p->avg_msec;
 706         l->hist_msec = p->hist_msec;
 707         l->hist_coarseness = p->hist_coarseness;
 708         l->filename = strdup(filename);
 709         l->td = p->td;
 710
 711         /* Initialize histogram lists for each r/w direction,
 712          * with initial io_u_plat of all zeros:
 713          */
 714         for (i = 0; i < DDIR_RWDIR_CNT; i++) {
 715                 list = &l->hist_window[i].list;
 716                 INIT_FLIST_HEAD(list);
 717                 entry = calloc(1, sizeof(struct io_u_plat_entry));
 718                 flist_add(&entry->list, list);
 719         }
 720
 721         if (l->td && l->td->o.io_submit_mode != IO_MODE_OFFLOAD) {
 722                 struct io_logs *__p;
 723
 724                 __p = calloc(1, sizeof(*l->pending));
 725                 __p->max_samples = DEF_LOG_ENTRIES;
 726                 __p->log = calloc(__p->max_samples, log_entry_sz(l));
 727                 l->pending = __p;
 728         }
 729
 730         if (l->log_offset)
 731                 l->log_ddir_mask = LOG_OFFSET_SAMPLE_BIT;
 732
 733         INIT_FLIST_HEAD(&l->chunk_list);
 734
 735         if (l->log_gz && !p->td)
 736                 l->log_gz = 0;
 737         else if (l->log_gz || l->log_gz_store) {
 738                 mutex_init_pshared(&l->chunk_lock);
 739                 mutex_init_pshared(&l->deferred_free_lock);
 740                 p->td->flags |= TD_F_COMPRESS_LOG;
 741         }
 742
 743         *log = l;
 744 }
 745
 746 #ifdef CONFIG_SETVBUF
 747 static void *set_file_buffer(FILE *f)
 748 {
 749         size_t size = 1048576;
 750         void *buf;
 751
 752         buf = malloc(size);
 753         setvbuf(f, buf, _IOFBF, size);
 754         return buf;
 755 }
 756
 757 static void clear_file_buffer(void *buf)
 758 {
 759         free(buf);
 760 }
 761 #else
 762 static void *set_file_buffer(FILE *f)
 763 {
 764         return NULL;
 765 }
 766
 767 static void clear_file_buffer(void *buf)
 768 {
 769 }
 770 #endif
 771
 772 void free_log(struct io_log *log)
 773 {
 774         while (!flist_empty(&log->io_logs)) {
 775                 struct io_logs *cur_log;
 776
 777                 cur_log = flist_first_entry(&log->io_logs, struct io_logs, list);
 778                 flist_del_init(&cur_log->list);
 779                 free(cur_log->log);
 780                 sfree(cur_log);
 781         }
 782
 783         if (log->pending) {
 784                 free(log->pending->log);
 785                 free(log->pending);
 786                 log->pending = NULL;
 787         }
 788
 789         free(log->pending);
 790         free(log->filename);
 791         sfree(log);
 792 }
 793
 794 uint64_t hist_sum(int j, int stride, uint64_t *io_u_plat,
 795                 uint64_t *io_u_plat_last)
 796 {
 797         uint64_t sum;
 798         int k;
 799
 800         if (io_u_plat_last) {
 801                 for (k = sum = 0; k < stride; k++)
 802                         sum += io_u_plat[j + k] - io_u_plat_last[j + k];
 803         } else {
 804                 for (k = sum = 0; k < stride; k++)
 805                         sum += io_u_plat[j + k];
 806         }
 807
 808         return sum;
 809 }
 810
 811 static void flush_hist_samples(FILE *f, int hist_coarseness, void *samples,
 812                                uint64_t sample_size)
 813 {
 814         struct io_sample *s;
 815         int log_offset;
 816         uint64_t i, j, nr_samples;
 817         struct io_u_plat_entry *entry, *entry_before;
 818         uint64_t *io_u_plat;
 819         uint64_t *io_u_plat_before;
 820
 821         int stride = 1 << hist_coarseness;
 822
 823         if (!sample_size)
 824                 return;
 825
 826         s = __get_sample(samples, 0, 0);
 827         log_offset = (s->__ddir & LOG_OFFSET_SAMPLE_BIT) != 0;
 828
 829         nr_samples = sample_size / __log_entry_sz(log_offset);
 830
 831         for (i = 0; i < nr_samples; i++) {
 832                 s = __get_sample(samples, log_offset, i);
 833
 834                 entry = s->data.plat_entry;
 835                 io_u_plat = entry->io_u_plat;
 836
 837                 entry_before = flist_first_entry(&entry->list, struct io_u_plat_entry, list);
 838                 io_u_plat_before = entry_before->io_u_plat;
 839
 840                 fprintf(f, "%lu, %u, %llu, ", (unsigned long) s->time,
 841                                                 io_sample_ddir(s), (unsigned long long) s->bs);
 842                 for (j = 0; j < FIO_IO_U_PLAT_NR - stride; j += stride) {
 843                         fprintf(f, "%llu, ", (unsigned long long)
 844                                 hist_sum(j, stride, io_u_plat, io_u_plat_before));
 845                 }
 846                 fprintf(f, "%llu\n", (unsigned long long)
 847                         hist_sum(FIO_IO_U_PLAT_NR - stride, stride, io_u_plat,
 848                                         io_u_plat_before));
 849
 850                 flist_del(&entry_before->list);
 851                 free(entry_before);
 852         }
 853 }
 854
 855 void flush_samples(FILE *f, void *samples, uint64_t sample_size)
 856 {
 857         struct io_sample *s;
 858         int log_offset;
 859         uint64_t i, nr_samples;
 860
 861         if (!sample_size)
 862                 return;
 863
 864         s = __get_sample(samples, 0, 0);
 865         log_offset = (s->__ddir & LOG_OFFSET_SAMPLE_BIT) != 0;
 866
 867         nr_samples = sample_size / __log_entry_sz(log_offset);
 868
 869         for (i = 0; i < nr_samples; i++) {
 870                 s = __get_sample(samples, log_offset, i);
 871
 872                 if (!log_offset) {
 873                         fprintf(f, "%lu, %" PRId64 ", %u, %llu\n",
 874                                         (unsigned long) s->time,
 875                                         s->data.val,
 876                                         io_sample_ddir(s), (unsigned long long) s->bs);
 877                 } else {
 878                         struct io_sample_offset *so = (void *) s;
 879
 880                         fprintf(f, "%lu, %" PRId64 ", %u, %llu, %llu\n",
 881                                         (unsigned long) s->time,
 882                                         s->data.val,
 883                                         io_sample_ddir(s), (unsigned long long) s->bs,
 884                                         (unsigned long long) so->offset);
 885                 }
 886         }
 887 }
 888
 889 #ifdef CONFIG_ZLIB
 890
 891 struct iolog_flush_data {
 892         struct workqueue_work work;
 893         struct io_log *log;
 894         void *samples;
 895         uint32_t nr_samples;
 896         bool free;
 897 };
 898
 899 #define GZ_CHUNK        131072
 900
 901 static struct iolog_compress *get_new_chunk(unsigned int seq)
 902 {
 903         struct iolog_compress *c;
 904
 905         c = malloc(sizeof(*c));
 906         INIT_FLIST_HEAD(&c->list);
 907         c->buf = malloc(GZ_CHUNK);
 908         c->len = 0;
 909         c->seq = seq;
 910         return c;
 911 }
 912
 913 static void free_chunk(struct iolog_compress *ic)
 914 {
 915         free(ic->buf);
 916         free(ic);
 917 }
 918
 919 static int z_stream_init(z_stream *stream, int gz_hdr)
 920 {
 921         int wbits = 15;
 922
 923         memset(stream, 0, sizeof(*stream));
 924         stream->zalloc = Z_NULL;
 925         stream->zfree = Z_NULL;
 926         stream->opaque = Z_NULL;
 927         stream->next_in = Z_NULL;
 928
 929         /*
 930          * zlib magic - add 32 for auto-detection of gz header or not,
 931          * if we decide to store files in a gzip friendly format.
 932          */
 933         if (gz_hdr)
 934                 wbits += 32;
 935
 936         if (inflateInit2(stream, wbits) != Z_OK)
 937                 return 1;
 938
 939         return 0;
 940 }
 941
 942 struct inflate_chunk_iter {
 943         unsigned int seq;
 944         int err;
 945         void *buf;
 946         size_t buf_size;
 947         size_t buf_used;
 948         size_t chunk_sz;
 949 };
 950
 951 static void finish_chunk(z_stream *stream, FILE *f,
 952                          struct inflate_chunk_iter *iter)
 953 {
 954         int ret;
 955
 956         ret = inflateEnd(stream);
 957         if (ret != Z_OK)
 958                 log_err("fio: failed to end log inflation seq %d (%d)\n",
 959                                 iter->seq, ret);
 960
 961         flush_samples(f, iter->buf, iter->buf_used);
 962         free(iter->buf);
 963         iter->buf = NULL;
 964         iter->buf_size = iter->buf_used = 0;
 965 }
 966
 967 /*
 968  * Iterative chunk inflation. Handles cases where we cross into a new
 969  * sequence, doing flush finish of previous chunk if needed.
 970  */
 971 static size_t inflate_chunk(struct iolog_compress *ic, int gz_hdr, FILE *f,
 972                             z_stream *stream, struct inflate_chunk_iter *iter)
 973 {
 974         size_t ret;
 975
 976         dprint(FD_COMPRESS, "inflate chunk size=%lu, seq=%u\n",
 977                                 (unsigned long) ic->len, ic->seq);
 978
 979         if (ic->seq != iter->seq) {
 980                 if (iter->seq)
 981                         finish_chunk(stream, f, iter);
 982
 983                 z_stream_init(stream, gz_hdr);
 984                 iter->seq = ic->seq;
 985         }
 986
 987         stream->avail_in = ic->len;
 988         stream->next_in = ic->buf;
 989
 990         if (!iter->buf_size) {
 991                 iter->buf_size = iter->chunk_sz;
 992                 iter->buf = malloc(iter->buf_size);
 993         }
 994
 995         while (stream->avail_in) {
 996                 size_t this_out = iter->buf_size - iter->buf_used;
 997                 int err;
 998
 999                 stream->avail_out = this_out;
1000                 stream->next_out = iter->buf + iter->buf_used;
1001
1002                 err = inflate(stream, Z_NO_FLUSH);
1003                 if (err < 0) {
1004                         log_err("fio: failed inflating log: %d\n", err);
1005                         iter->err = err;
1006                         break;
1007                 }
1008
1009                 iter->buf_used += this_out - stream->avail_out;
1010
1011                 if (!stream->avail_out) {
1012                         iter->buf_size += iter->chunk_sz;
1013                         iter->buf = realloc(iter->buf, iter->buf_size);
1014                         continue;
1015                 }
1016
1017                 if (err == Z_STREAM_END)
1018                         break;
1019         }
1020
1021         ret = (void *) stream->next_in - ic->buf;
1022
1023         dprint(FD_COMPRESS, "inflated to size=%lu\n", (unsigned long) iter->buf_size);
1024
1025         return ret;
1026 }
1027
1028 /*
1029  * Inflate stored compressed chunks, or write them directly to the log
1030  * file if so instructed.
1031  */
1032 static int inflate_gz_chunks(struct io_log *log, FILE *f)
1033 {
1034         struct inflate_chunk_iter iter = { .chunk_sz = log->log_gz, };
1035         z_stream stream;
1036
1037         while (!flist_empty(&log->chunk_list)) {
1038                 struct iolog_compress *ic;
1039
1040                 ic = flist_first_entry(&log->chunk_list, struct iolog_compress, list);
1041                 flist_del(&ic->list);
1042
1043                 if (log->log_gz_store) {
1044                         size_t ret;
1045
1046                         dprint(FD_COMPRESS, "log write chunk size=%lu, "
1047                                 "seq=%u\n", (unsigned long) ic->len, ic->seq);
1048
1049                         ret = fwrite(ic->buf, ic->len, 1, f);
1050                         if (ret != 1 || ferror(f)) {
1051                                 iter.err = errno;
1052                                 log_err("fio: error writing compressed log\n");
1053                         }
1054                 } else
1055                         inflate_chunk(ic, log->log_gz_store, f, &stream, &iter);
1056
1057                 free_chunk(ic);
1058         }
1059
1060         if (iter.seq) {
1061                 finish_chunk(&stream, f, &iter);
1062                 free(iter.buf);
1063         }
1064
1065         return iter.err;
1066 }
1067
1068 /*
1069  * Open compressed log file and decompress the stored chunks and
1070  * write them to stdout. The chunks are stored sequentially in the
1071  * file, so we iterate over them and do them one-by-one.
1072  */
1073 int iolog_file_inflate(const char *file)
1074 {
1075         struct inflate_chunk_iter iter = { .chunk_sz = 64 * 1024 * 1024, };
1076         struct iolog_compress ic;
1077         z_stream stream;
1078         struct stat sb;
1079         size_t ret;
1080         size_t total;
1081         void *buf;
1082         FILE *f;
1083
1084         f = fopen(file, "r");
1085         if (!f) {
1086                 perror("fopen");
1087                 return 1;
1088         }
1089
1090         if (stat(file, &sb) < 0) {
1091                 fclose(f);
1092                 perror("stat");
1093                 return 1;
1094         }
1095
1096         ic.buf = buf = malloc(sb.st_size);
1097         ic.len = sb.st_size;
1098         ic.seq = 1;
1099
1100         ret = fread(ic.buf, ic.len, 1, f);
1101         if (ret == 0 && ferror(f)) {
1102                 perror("fread");
1103                 fclose(f);
1104                 free(buf);
1105                 return 1;
1106         } else if (ferror(f) || (!feof(f) && ret != 1)) {
1107                 log_err("fio: short read on reading log\n");
1108                 fclose(f);
1109                 free(buf);
1110                 return 1;
1111         }
1112
1113         fclose(f);
1114
1115         /*
1116          * Each chunk will return Z_STREAM_END. We don't know how many
1117          * chunks are in the file, so we just keep looping and incrementing
1118          * the sequence number until we have consumed the whole compressed
1119          * file.
1120          */
1121         total = ic.len;
1122         do {
1123                 size_t iret;
1124
1125                 iret = inflate_chunk(&ic,  1, stdout, &stream, &iter);
1126                 total -= iret;
1127                 if (!total)
1128                         break;
1129                 if (iter.err)
1130                         break;
1131
1132                 ic.seq++;
1133                 ic.len -= iret;
1134                 ic.buf += iret;
1135         } while (1);
1136
1137         if (iter.seq) {
1138                 finish_chunk(&stream, stdout, &iter);
1139                 free(iter.buf);
1140         }
1141
1142         free(buf);
1143         return iter.err;
1144 }
1145
1146 #else
1147
1148 static int inflate_gz_chunks(struct io_log *log, FILE *f)
1149 {
1150         return 0;
1151 }
1152
1153 int iolog_file_inflate(const char *file)
1154 {
1155         log_err("fio: log inflation not possible without zlib\n");
1156         return 1;
1157 }
1158
1159 #endif
1160
1161 void flush_log(struct io_log *log, bool do_append)
1162 {
1163         void *buf;
1164         FILE *f;
1165
1166         if (!do_append)
1167                 f = fopen(log->filename, "w");
1168         else
1169                 f = fopen(log->filename, "a");
1170         if (!f) {
1171                 perror("fopen log");
1172                 return;
1173         }
1174
1175         buf = set_file_buffer(f);
1176
1177         inflate_gz_chunks(log, f);
1178
1179         while (!flist_empty(&log->io_logs)) {
1180                 struct io_logs *cur_log;
1181
1182                 cur_log = flist_first_entry(&log->io_logs, struct io_logs, list);
1183                 flist_del_init(&cur_log->list);
1184
1185                 if (log->td && log == log->td->clat_hist_log)
1186                         flush_hist_samples(f, log->hist_coarseness, cur_log->log,
1187                                            log_sample_sz(log, cur_log));
1188                 else
1189                         flush_samples(f, cur_log->log, log_sample_sz(log, cur_log));
1190
1191                 sfree(cur_log);
1192         }
1193
1194         fclose(f);
1195         clear_file_buffer(buf);
1196 }
1197
1198 static int finish_log(struct thread_data *td, struct io_log *log, int trylock)
1199 {
1200         if (td->flags & TD_F_COMPRESS_LOG)
1201                 iolog_flush(log);
1202
1203         if (trylock) {
1204                 if (fio_trylock_file(log->filename))
1205                         return 1;
1206         } else
1207                 fio_lock_file(log->filename);
1208
1209         if (td->client_type == FIO_CLIENT_TYPE_GUI || is_backend)
1210                 fio_send_iolog(td, log, log->filename);
1211         else
1212                 flush_log(log, !td->o.per_job_logs);
1213
1214         fio_unlock_file(log->filename);
1215         free_log(log);
1216         return 0;
1217 }
1218
1219 size_t log_chunk_sizes(struct io_log *log)
1220 {
1221         struct flist_head *entry;
1222         size_t ret;
1223
1224         if (flist_empty(&log->chunk_list))
1225                 return 0;
1226
1227         ret = 0;
1228         pthread_mutex_lock(&log->chunk_lock);
1229         flist_for_each(entry, &log->chunk_list) {
1230                 struct iolog_compress *c;
1231
1232                 c = flist_entry(entry, struct iolog_compress, list);
1233                 ret += c->len;
1234         }
1235         pthread_mutex_unlock(&log->chunk_lock);
1236         return ret;
1237 }
1238
1239 #ifdef CONFIG_ZLIB
1240
1241 static void iolog_put_deferred(struct io_log *log, void *ptr)
1242 {
1243         if (!ptr)
1244                 return;
1245
1246         pthread_mutex_lock(&log->deferred_free_lock);
1247         if (log->deferred < IOLOG_MAX_DEFER) {
1248                 log->deferred_items[log->deferred] = ptr;
1249                 log->deferred++;
1250         } else if (!fio_did_warn(FIO_WARN_IOLOG_DROP))
1251                 log_err("fio: had to drop log entry free\n");
1252         pthread_mutex_unlock(&log->deferred_free_lock);
1253 }
1254
1255 static void iolog_free_deferred(struct io_log *log)
1256 {
1257         int i;
1258
1259         if (!log->deferred)
1260                 return;
1261
1262         pthread_mutex_lock(&log->deferred_free_lock);
1263
1264         for (i = 0; i < log->deferred; i++) {
1265                 free(log->deferred_items[i]);
1266                 log->deferred_items[i] = NULL;
1267         }
1268
1269         log->deferred = 0;
1270         pthread_mutex_unlock(&log->deferred_free_lock);
1271 }
1272
1273 static int gz_work(struct iolog_flush_data *data)
1274 {
1275         struct iolog_compress *c = NULL;
1276         struct flist_head list;
1277         unsigned int seq;
1278         z_stream stream;
1279         size_t total = 0;
1280         int ret;
1281
1282         INIT_FLIST_HEAD(&list);
1283
1284         memset(&stream, 0, sizeof(stream));
1285         stream.zalloc = Z_NULL;
1286         stream.zfree = Z_NULL;
1287         stream.opaque = Z_NULL;
1288
1289         ret = deflateInit(&stream, Z_DEFAULT_COMPRESSION);
1290         if (ret != Z_OK) {
1291                 log_err("fio: failed to init gz stream\n");
1292                 goto err;
1293         }
1294
1295         seq = ++data->log->chunk_seq;
1296
1297         stream.next_in = (void *) data->samples;
1298         stream.avail_in = data->nr_samples * log_entry_sz(data->log);
1299
1300         dprint(FD_COMPRESS, "deflate input size=%lu, seq=%u, log=%s\n",
1301                                 (unsigned long) stream.avail_in, seq,
1302                                 data->log->filename);
1303         do {
1304                 if (c)
1305                         dprint(FD_COMPRESS, "seq=%d, chunk=%lu\n", seq,
1306                                 (unsigned long) c->len);
1307                 c = get_new_chunk(seq);
1308                 stream.avail_out = GZ_CHUNK;
1309                 stream.next_out = c->buf;
1310                 ret = deflate(&stream, Z_NO_FLUSH);
1311                 if (ret < 0) {
1312                         log_err("fio: deflate log (%d)\n", ret);
1313                         free_chunk(c);
1314                         goto err;
1315                 }
1316
1317                 c->len = GZ_CHUNK - stream.avail_out;
1318                 flist_add_tail(&c->list, &list);
1319                 total += c->len;
1320         } while (stream.avail_in);
1321
1322         stream.next_out = c->buf + c->len;
1323         stream.avail_out = GZ_CHUNK - c->len;
1324
1325         ret = deflate(&stream, Z_FINISH);
1326         if (ret < 0) {
1327                 /*
1328                  * Z_BUF_ERROR is special, it just means we need more
1329                  * output space. We'll handle that below. Treat any other
1330                  * error as fatal.
1331                  */
1332                 if (ret != Z_BUF_ERROR) {
1333                         log_err("fio: deflate log (%d)\n", ret);
1334                         flist_del(&c->list);
1335                         free_chunk(c);
1336                         goto err;
1337                 }
1338         }
1339
1340         total -= c->len;
1341         c->len = GZ_CHUNK - stream.avail_out;
1342         total += c->len;
1343         dprint(FD_COMPRESS, "seq=%d, chunk=%lu\n", seq, (unsigned long) c->len);
1344
1345         if (ret != Z_STREAM_END) {
1346                 do {
1347                         c = get_new_chunk(seq);
1348                         stream.avail_out = GZ_CHUNK;
1349                         stream.next_out = c->buf;
1350                         ret = deflate(&stream, Z_FINISH);
1351                         c->len = GZ_CHUNK - stream.avail_out;
1352                         total += c->len;
1353                         flist_add_tail(&c->list, &list);
1354                         dprint(FD_COMPRESS, "seq=%d, chunk=%lu\n", seq,
1355                                 (unsigned long) c->len);
1356                 } while (ret != Z_STREAM_END);
1357         }
1358
1359         dprint(FD_COMPRESS, "deflated to size=%lu\n", (unsigned long) total);
1360
1361         ret = deflateEnd(&stream);
1362         if (ret != Z_OK)
1363                 log_err("fio: deflateEnd %d\n", ret);
1364
1365         iolog_put_deferred(data->log, data->samples);
1366
1367         if (!flist_empty(&list)) {
1368                 pthread_mutex_lock(&data->log->chunk_lock);
1369                 flist_splice_tail(&list, &data->log->chunk_list);
1370                 pthread_mutex_unlock(&data->log->chunk_lock);
1371         }
1372
1373         ret = 0;
1374 done:
1375         if (data->free)
1376                 sfree(data);
1377         return ret;
1378 err:
1379         while (!flist_empty(&list)) {
1380                 c = flist_first_entry(list.next, struct iolog_compress, list);
1381                 flist_del(&c->list);
1382                 free_chunk(c);
1383         }
1384         ret = 1;
1385         goto done;
1386 }
1387
1388 /*
1389  * Invoked from our compress helper thread, when logging would have exceeded
1390  * the specified memory limitation. Compresses the previously stored
1391  * entries.
1392  */
1393 static int gz_work_async(struct submit_worker *sw, struct workqueue_work *work)
1394 {
1395         return gz_work(container_of(work, struct iolog_flush_data, work));
1396 }
1397
1398 static int gz_init_worker(struct submit_worker *sw)
1399 {
1400         struct thread_data *td = sw->wq->td;
1401
1402         if (!fio_option_is_set(&td->o, log_gz_cpumask))
1403                 return 0;
1404
1405         if (fio_setaffinity(gettid(), td->o.log_gz_cpumask) == -1) {
1406                 log_err("gz: failed to set CPU affinity\n");
1407                 return 1;
1408         }
1409
1410         return 0;
1411 }
1412
1413 static struct workqueue_ops log_compress_wq_ops = {
1414         .fn             = gz_work_async,
1415         .init_worker_fn = gz_init_worker,
1416         .nice           = 1,
1417 };
1418
1419 int iolog_compress_init(struct thread_data *td, struct sk_out *sk_out)
1420 {
1421         if (!(td->flags & TD_F_COMPRESS_LOG))
1422                 return 0;
1423
1424         workqueue_init(td, &td->log_compress_wq, &log_compress_wq_ops, 1, sk_out);
1425         return 0;
1426 }
1427
1428 void iolog_compress_exit(struct thread_data *td)
1429 {
1430         if (!(td->flags & TD_F_COMPRESS_LOG))
1431                 return;
1432
1433         workqueue_exit(&td->log_compress_wq);
1434 }
1435
1436 /*
1437  * Queue work item to compress the existing log entries. We reset the
1438  * current log to a small size, and reference the existing log in the
1439  * data that we queue for compression. Once compression has been done,
1440  * this old log is freed. If called with finish == true, will not return
1441  * until the log compression has completed, and will flush all previous
1442  * logs too
1443  */
1444 static int iolog_flush(struct io_log *log)
1445 {
1446         struct iolog_flush_data *data;
1447
1448         data = malloc(sizeof(*data));
1449         if (!data)
1450                 return 1;
1451
1452         data->log = log;
1453         data->free = false;
1454
1455         while (!flist_empty(&log->io_logs)) {
1456                 struct io_logs *cur_log;
1457
1458                 cur_log = flist_first_entry(&log->io_logs, struct io_logs, list);
1459                 flist_del_init(&cur_log->list);
1460
1461                 data->samples = cur_log->log;
1462                 data->nr_samples = cur_log->nr_samples;
1463
1464                 sfree(cur_log);
1465
1466                 gz_work(data);
1467         }
1468
1469         free(data);
1470         return 0;
1471 }
1472
1473 int iolog_cur_flush(struct io_log *log, struct io_logs *cur_log)
1474 {
1475         struct iolog_flush_data *data;
1476
1477         data = smalloc(sizeof(*data));
1478         if (!data)
1479                 return 1;
1480
1481         data->log = log;
1482
1483         data->samples = cur_log->log;
1484         data->nr_samples = cur_log->nr_samples;
1485         data->free = true;
1486
1487         cur_log->nr_samples = cur_log->max_samples = 0;
1488         cur_log->log = NULL;
1489
1490         workqueue_enqueue(&log->td->log_compress_wq, &data->work);
1491
1492         iolog_free_deferred(log);
1493
1494         return 0;
1495 }
1496 #else
1497
1498 static int iolog_flush(struct io_log *log)
1499 {
1500         return 1;
1501 }
1502
1503 int iolog_cur_flush(struct io_log *log, struct io_logs *cur_log)
1504 {
1505         return 1;
1506 }
1507
1508 int iolog_compress_init(struct thread_data *td, struct sk_out *sk_out)
1509 {
1510         return 0;
1511 }
1512
1513 void iolog_compress_exit(struct thread_data *td)
1514 {
1515 }
1516
1517 #endif
1518
1519 struct io_logs *iolog_cur_log(struct io_log *log)
1520 {
1521         if (flist_empty(&log->io_logs))
1522                 return NULL;
1523
1524         return flist_last_entry(&log->io_logs, struct io_logs, list);
1525 }
1526
1527 uint64_t iolog_nr_samples(struct io_log *iolog)
1528 {
1529         struct flist_head *entry;
1530         uint64_t ret = 0;
1531
1532         flist_for_each(entry, &iolog->io_logs) {
1533                 struct io_logs *cur_log;
1534
1535                 cur_log = flist_entry(entry, struct io_logs, list);
1536                 ret += cur_log->nr_samples;
1537         }
1538
1539         return ret;
1540 }
1541
1542 static int __write_log(struct thread_data *td, struct io_log *log, int try)
1543 {
1544         if (log)
1545                 return finish_log(td, log, try);
1546
1547         return 0;
1548 }
1549
1550 static int write_iops_log(struct thread_data *td, int try, bool unit_log)
1551 {
1552         int ret;
1553
1554         if (per_unit_log(td->iops_log) != unit_log)
1555                 return 0;
1556
1557         ret = __write_log(td, td->iops_log, try);
1558         if (!ret)
1559                 td->iops_log = NULL;
1560
1561         return ret;
1562 }
1563
1564 static int write_slat_log(struct thread_data *td, int try, bool unit_log)
1565 {
1566         int ret;
1567
1568         if (!unit_log)
1569                 return 0;
1570
1571         ret = __write_log(td, td->slat_log, try);
1572         if (!ret)
1573                 td->slat_log = NULL;
1574
1575         return ret;
1576 }
1577
1578 static int write_clat_log(struct thread_data *td, int try, bool unit_log)
1579 {
1580         int ret;
1581
1582         if (!unit_log)
1583                 return 0;
1584
1585         ret = __write_log(td, td->clat_log, try);
1586         if (!ret)
1587                 td->clat_log = NULL;
1588
1589         return ret;
1590 }
1591
1592 static int write_clat_hist_log(struct thread_data *td, int try, bool unit_log)
1593 {
1594         int ret;
1595
1596         if (!unit_log)
1597                 return 0;
1598
1599         ret = __write_log(td, td->clat_hist_log, try);
1600         if (!ret)
1601                 td->clat_hist_log = NULL;
1602
1603         return ret;
1604 }
1605
1606 static int write_lat_log(struct thread_data *td, int try, bool unit_log)
1607 {
1608         int ret;
1609
1610         if (!unit_log)
1611                 return 0;
1612
1613         ret = __write_log(td, td->lat_log, try);
1614         if (!ret)
1615                 td->lat_log = NULL;
1616
1617         return ret;
1618 }
1619
1620 static int write_bandw_log(struct thread_data *td, int try, bool unit_log)
1621 {
1622         int ret;
1623
1624         if (per_unit_log(td->bw_log) != unit_log)
1625                 return 0;
1626
1627         ret = __write_log(td, td->bw_log, try);
1628         if (!ret)
1629                 td->bw_log = NULL;
1630
1631         return ret;
1632 }
1633
1634 enum {
1635         BW_LOG_MASK     = 1,
1636         LAT_LOG_MASK    = 2,
1637         SLAT_LOG_MASK   = 4,
1638         CLAT_LOG_MASK   = 8,
1639         IOPS_LOG_MASK   = 16,
1640         CLAT_HIST_LOG_MASK = 32,
1641
1642         ALL_LOG_NR      = 6,
1643 };
1644
1645 struct log_type {
1646         unsigned int mask;
1647         int (*fn)(struct thread_data *, int, bool);
1648 };
1649
1650 static struct log_type log_types[] = {
1651         {
1652                 .mask   = BW_LOG_MASK,
1653                 .fn     = write_bandw_log,
1654         },
1655         {
1656                 .mask   = LAT_LOG_MASK,
1657                 .fn     = write_lat_log,
1658         },
1659         {
1660                 .mask   = SLAT_LOG_MASK,
1661                 .fn     = write_slat_log,
1662         },
1663         {
1664                 .mask   = CLAT_LOG_MASK,
1665                 .fn     = write_clat_log,
1666         },
1667         {
1668                 .mask   = IOPS_LOG_MASK,
1669                 .fn     = write_iops_log,
1670         },
1671         {
1672                 .mask   = CLAT_HIST_LOG_MASK,
1673                 .fn     = write_clat_hist_log,
1674         }
1675 };
1676
1677 void td_writeout_logs(struct thread_data *td, bool unit_logs)
1678 {
1679         unsigned int log_mask = 0;
1680         unsigned int log_left = ALL_LOG_NR;
1681         int old_state, i;
1682
1683         old_state = td_bump_runstate(td, TD_FINISHING);
1684
1685         finalize_logs(td, unit_logs);
1686
1687         while (log_left) {
1688                 int prev_log_left = log_left;
1689
1690                 for (i = 0; i < ALL_LOG_NR && log_left; i++) {
1691                         struct log_type *lt = &log_types[i];
1692                         int ret;
1693
1694                         if (!(log_mask & lt->mask)) {
1695                                 ret = lt->fn(td, log_left != 1, unit_logs);
1696                                 if (!ret) {
1697                                         log_left--;
1698                                         log_mask |= lt->mask;
1699                                 }
1700                         }
1701                 }
1702
1703                 if (prev_log_left == log_left)
1704                         usleep(5000);
1705         }
1706
1707         td_restore_runstate(td, old_state);
1708 }
1709
1710 void fio_writeout_logs(bool unit_logs)
1711 {
1712         struct thread_data *td;
1713         int i;
1714
1715         for_each_td(td, i)
1716                 td_writeout_logs(td, unit_logs);
1717 }