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 #include "lib/roundup.h"
  23
  24 #include <netinet/in.h>
  25 #include <netinet/tcp.h>
  26 #include <arpa/inet.h>
  27 #include <sys/stat.h>
  28 #include <sys/socket.h>
  29 #include <sys/un.h>
  30
  31 static int iolog_flush(struct io_log *log);
  32
  33 static const char iolog_ver2[] = "fio version 2 iolog";
  34 static const char iolog_ver3[] = "fio version 3 iolog";
  35
  36 void queue_io_piece(struct thread_data *td, struct io_piece *ipo)
  37 {
  38         flist_add_tail(&ipo->list, &td->io_log_list);
  39         td->total_io_size += ipo->len;
  40 }
  41
  42 void log_io_u(const struct thread_data *td, const struct io_u *io_u)
  43 {
  44         struct timespec now;
  45
  46         if (!td->o.write_iolog_file)
  47                 return;
  48
  49         fio_gettime(&now, NULL);
  50         fprintf(td->iolog_f, "%llu %s %s %llu %llu\n",
  51                 (unsigned long long) utime_since_now(&td->io_log_start_time),
  52                 io_u->file->file_name, io_ddir_name(io_u->ddir), io_u->offset,
  53                 io_u->buflen);
  54
  55 }
  56
  57 void log_file(struct thread_data *td, struct fio_file *f,
  58               enum file_log_act what)
  59 {
  60         const char *act[] = { "add", "open", "close" };
  61         struct timespec now;
  62
  63         assert(what < 3);
  64
  65         if (!td->o.write_iolog_file)
  66                 return;
  67
  68
  69         /*
  70          * this happens on the pre-open/close done before the job starts
  71          */
  72         if (!td->iolog_f)
  73                 return;
  74
  75         fio_gettime(&now, NULL);
  76         fprintf(td->iolog_f, "%llu %s %s\n",
  77                 (unsigned long long) utime_since_now(&td->io_log_start_time),
  78                 f->file_name, act[what]);
  79 }
  80
  81 static void iolog_delay(struct thread_data *td, unsigned long delay)
  82 {
  83         uint64_t usec = utime_since_now(&td->last_issue);
  84         unsigned long orig_delay = delay;
  85         uint64_t this_delay;
  86         struct timespec ts;
  87
  88         if (delay < td->time_offset) {
  89                 td->time_offset = 0;
  90                 return;
  91         }
  92
  93         delay -= td->time_offset;
  94         if (delay < usec)
  95                 return;
  96
  97         delay -= usec;
  98
  99         fio_gettime(&ts, NULL);
 100         while (delay && !td->terminate) {
 101                 this_delay = delay;
 102                 if (this_delay > 500000)
 103                         this_delay = 500000;
 104
 105                 usec_sleep(td, this_delay);
 106                 delay -= this_delay;
 107         }
 108
 109         usec = utime_since_now(&ts);
 110         if (usec > orig_delay)
 111                 td->time_offset = usec - orig_delay;
 112         else
 113                 td->time_offset = 0;
 114 }
 115
 116 static int ipo_special(struct thread_data *td, struct io_piece *ipo)
 117 {
 118         struct fio_file *f;
 119         int ret;
 120
 121         /*
 122          * Not a special ipo
 123          */
 124         if (ipo->ddir != DDIR_INVAL)
 125                 return 0;
 126
 127         f = td->files[ipo->fileno];
 128
 129         if (ipo->delay)
 130                 iolog_delay(td, ipo->delay);
 131         if (fio_fill_issue_time(td))
 132                 fio_gettime(&td->last_issue, NULL);
 133         switch (ipo->file_action) {
 134         case FIO_LOG_OPEN_FILE:
 135                 if (td->o.replay_redirect && fio_file_open(f)) {
 136                         dprint(FD_FILE, "iolog: ignoring re-open of file %s\n",
 137                                         f->file_name);
 138                         break;
 139                 }
 140                 ret = td_io_open_file(td, f);
 141                 if (!ret)
 142                         break;
 143                 td_verror(td, ret, "iolog open file");
 144                 return -1;
 145         case FIO_LOG_CLOSE_FILE:
 146                 td_io_close_file(td, f);
 147                 break;
 148         case FIO_LOG_UNLINK_FILE:
 149                 td_io_unlink_file(td, f);
 150                 break;
 151         case FIO_LOG_ADD_FILE:
 152                 /*
 153                  * Nothing to do
 154                  */
 155                 break;
 156         default:
 157                 log_err("fio: bad file action %d\n", ipo->file_action);
 158                 break;
 159         }
 160
 161         return 1;
 162 }
 163
 164 static bool read_iolog(struct thread_data *td);
 165
 166 unsigned long long delay_since_ttime(const struct thread_data *td,
 167                unsigned long long time)
 168 {
 169         double tmp;
 170         double scale;
 171         const unsigned long long *last_ttime = &td->io_log_last_ttime;
 172
 173         if (!*last_ttime || td->o.no_stall || time < *last_ttime)
 174                 return 0;
 175         else if (td->o.replay_time_scale == 100)
 176                 return time - *last_ttime;
 177
 178
 179         scale = (double) 100.0 / (double) td->o.replay_time_scale;
 180         tmp = time - *last_ttime;
 181         return tmp * scale;
 182 }
 183
 184 int read_iolog_get(struct thread_data *td, struct io_u *io_u)
 185 {
 186         struct io_piece *ipo;
 187         unsigned long elapsed;
 188
 189         while (!flist_empty(&td->io_log_list)) {
 190                 int ret;
 191
 192                 if (td->o.read_iolog_chunked) {
 193                         if (td->io_log_checkmark == td->io_log_current) {
 194                                 if (td->io_log_blktrace) {
 195                                         if (!read_blktrace(td))
 196                                                 return 1;
 197                                 } else {
 198                                         if (!read_iolog(td))
 199                                                 return 1;
 200                                 }
 201                         }
 202                         td->io_log_current--;
 203                 }
 204                 ipo = flist_first_entry(&td->io_log_list, struct io_piece, list);
 205                 flist_del(&ipo->list);
 206                 remove_trim_entry(td, ipo);
 207
 208                 ret = ipo_special(td, ipo);
 209                 if (ret < 0) {
 210                         free(ipo);
 211                         break;
 212                 } else if (ret > 0) {
 213                         free(ipo);
 214                         continue;
 215                 }
 216
 217                 io_u->ddir = ipo->ddir;
 218                 if (ipo->ddir != DDIR_WAIT) {
 219                         io_u->offset = ipo->offset;
 220                         io_u->verify_offset = ipo->offset;
 221                         io_u->buflen = ipo->len;
 222                         io_u->file = td->files[ipo->fileno];
 223                         get_file(io_u->file);
 224                         dprint(FD_IO, "iolog: get %llu/%llu/%s\n", io_u->offset,
 225                                                 io_u->buflen, io_u->file->file_name);
 226                         if (ipo->delay)
 227                                 iolog_delay(td, ipo->delay);
 228                 } else {
 229                         elapsed = mtime_since_genesis();
 230                         if (ipo->delay > elapsed)
 231                                 usec_sleep(td, (ipo->delay - elapsed) * 1000);
 232                 }
 233
 234                 free(ipo);
 235
 236                 if (io_u->ddir != DDIR_WAIT)
 237                         return 0;
 238         }
 239
 240         td->done = 1;
 241         return 1;
 242 }
 243
 244 void prune_io_piece_log(struct thread_data *td)
 245 {
 246         struct io_piece *ipo;
 247         struct fio_rb_node *n;
 248
 249         while ((n = rb_first(&td->io_hist_tree)) != NULL) {
 250                 ipo = rb_entry(n, struct io_piece, rb_node);
 251                 rb_erase(n, &td->io_hist_tree);
 252                 remove_trim_entry(td, ipo);
 253                 td->io_hist_len--;
 254                 free(ipo);
 255         }
 256
 257         while (!flist_empty(&td->io_hist_list)) {
 258                 ipo = flist_first_entry(&td->io_hist_list, struct io_piece, list);
 259                 flist_del(&ipo->list);
 260                 remove_trim_entry(td, ipo);
 261                 td->io_hist_len--;
 262                 free(ipo);
 263         }
 264 }
 265
 266 /*
 267  * log a successful write, so we can unwind the log for verify
 268  */
 269 void log_io_piece(struct thread_data *td, struct io_u *io_u)
 270 {
 271         struct fio_rb_node **p, *parent;
 272         struct io_piece *ipo, *__ipo;
 273
 274         ipo = calloc(1, sizeof(struct io_piece));
 275         init_ipo(ipo);
 276         ipo->file = io_u->file;
 277         ipo->offset = io_u->offset;
 278         ipo->len = io_u->buflen;
 279         ipo->numberio = io_u->numberio;
 280         ipo->flags = IP_F_IN_FLIGHT;
 281
 282         io_u->ipo = ipo;
 283
 284         if (io_u_should_trim(td, io_u)) {
 285                 flist_add_tail(&ipo->trim_list, &td->trim_list);
 286                 td->trim_entries++;
 287         }
 288
 289         /*
 290          * Only sort writes if we don't have a random map in which case we need
 291          * to check for duplicate blocks and drop the old one, which we rely on
 292          * the rb insert/lookup for handling.
 293          */
 294         if (file_randommap(td, ipo->file)) {
 295                 INIT_FLIST_HEAD(&ipo->list);
 296                 flist_add_tail(&ipo->list, &td->io_hist_list);
 297                 ipo->flags |= IP_F_ONLIST;
 298                 td->io_hist_len++;
 299                 return;
 300         }
 301
 302         RB_CLEAR_NODE(&ipo->rb_node);
 303
 304         /*
 305          * Sort the entry into the verification list
 306          */
 307 restart:
 308         p = &td->io_hist_tree.rb_node;
 309         parent = NULL;
 310         while (*p) {
 311                 int overlap = 0;
 312                 parent = *p;
 313
 314                 __ipo = rb_entry(parent, struct io_piece, rb_node);
 315                 if (ipo->file < __ipo->file)
 316                         p = &(*p)->rb_left;
 317                 else if (ipo->file > __ipo->file)
 318                         p = &(*p)->rb_right;
 319                 else if (ipo->offset < __ipo->offset) {
 320                         p = &(*p)->rb_left;
 321                         overlap = ipo->offset + ipo->len > __ipo->offset;
 322                 }
 323                 else if (ipo->offset > __ipo->offset) {
 324                         p = &(*p)->rb_right;
 325                         overlap = __ipo->offset + __ipo->len > ipo->offset;
 326                 }
 327                 else
 328                         overlap = 1;
 329
 330                 if (overlap) {
 331                         dprint(FD_IO, "iolog: overlap %llu/%lu, %llu/%lu\n",
 332                                 __ipo->offset, __ipo->len,
 333                                 ipo->offset, ipo->len);
 334                         td->io_hist_len--;
 335                         rb_erase(parent, &td->io_hist_tree);
 336                         remove_trim_entry(td, __ipo);
 337                         if (!(__ipo->flags & IP_F_IN_FLIGHT))
 338                                 free(__ipo);
 339                         goto restart;
 340                 }
 341         }
 342
 343         rb_link_node(&ipo->rb_node, parent, p);
 344         rb_insert_color(&ipo->rb_node, &td->io_hist_tree);
 345         ipo->flags |= IP_F_ONRB;
 346         td->io_hist_len++;
 347 }
 348
 349 void unlog_io_piece(struct thread_data *td, struct io_u *io_u)
 350 {
 351         struct io_piece *ipo = io_u->ipo;
 352
 353         if (td->ts.nr_block_infos) {
 354                 uint32_t *info = io_u_block_info(td, io_u);
 355                 if (BLOCK_INFO_STATE(*info) < BLOCK_STATE_TRIM_FAILURE) {
 356                         if (io_u->ddir == DDIR_TRIM)
 357                                 *info = BLOCK_INFO_SET_STATE(*info,
 358                                                 BLOCK_STATE_TRIM_FAILURE);
 359                         else if (io_u->ddir == DDIR_WRITE)
 360                                 *info = BLOCK_INFO_SET_STATE(*info,
 361                                                 BLOCK_STATE_WRITE_FAILURE);
 362                 }
 363         }
 364
 365         if (!ipo)
 366                 return;
 367
 368         if (ipo->flags & IP_F_ONRB)
 369                 rb_erase(&ipo->rb_node, &td->io_hist_tree);
 370         else if (ipo->flags & IP_F_ONLIST)
 371                 flist_del(&ipo->list);
 372
 373         free(ipo);
 374         io_u->ipo = NULL;
 375         td->io_hist_len--;
 376 }
 377
 378 void trim_io_piece(const struct io_u *io_u)
 379 {
 380         struct io_piece *ipo = io_u->ipo;
 381
 382         if (!ipo)
 383                 return;
 384
 385         ipo->len = io_u->xfer_buflen - io_u->resid;
 386 }
 387
 388 void write_iolog_close(struct thread_data *td)
 389 {
 390         if (!td->iolog_f)
 391                 return;
 392
 393         fflush(td->iolog_f);
 394         fclose(td->iolog_f);
 395         free(td->iolog_buf);
 396         td->iolog_f = NULL;
 397         td->iolog_buf = NULL;
 398 }
 399
 400 int64_t iolog_items_to_fetch(struct thread_data *td)
 401 {
 402         struct timespec now;
 403         uint64_t elapsed;
 404         uint64_t for_1s;
 405         int64_t items_to_fetch;
 406
 407         if (!td->io_log_highmark)
 408                 return 10;
 409
 410
 411         fio_gettime(&now, NULL);
 412         elapsed = ntime_since(&td->io_log_highmark_time, &now);
 413         if (elapsed) {
 414                 for_1s = (td->io_log_highmark - td->io_log_current) * 1000000000 / elapsed;
 415                 items_to_fetch = for_1s - td->io_log_current;
 416                 if (items_to_fetch < 0)
 417                         items_to_fetch = 0;
 418         } else
 419                 items_to_fetch = 0;
 420
 421         td->io_log_highmark = td->io_log_current + items_to_fetch;
 422         td->io_log_checkmark = (td->io_log_highmark + 1) / 2;
 423         fio_gettime(&td->io_log_highmark_time, NULL);
 424
 425         return items_to_fetch;
 426 }
 427
 428 #define io_act(_td, _r) (((_td)->io_log_version == 3 && (r) == 5) || \
 429                                         ((_td)->io_log_version == 2 && (r) == 4))
 430 #define file_act(_td, _r) (((_td)->io_log_version == 3 && (r) == 3) || \
 431                                         ((_td)->io_log_version == 2 && (r) == 2))
 432
 433 /*
 434  * Read version 2 and 3 iolog data. It is enhanced to include per-file logging,
 435  * syncs, etc.
 436  */
 437 static bool read_iolog(struct thread_data *td)
 438 {
 439         unsigned long long offset;
 440         unsigned int bytes;
 441         unsigned long long delay = 0;
 442         int reads, writes, waits, fileno = 0, file_action = 0; /* stupid gcc */
 443         char *rfname, *fname, *act;
 444         char *str, *p;
 445         enum fio_ddir rw;
 446         bool realloc = false;
 447         int64_t items_to_fetch = 0;
 448         int syncs;
 449
 450         if (td->o.read_iolog_chunked) {
 451                 items_to_fetch = iolog_items_to_fetch(td);
 452                 if (!items_to_fetch)
 453                         return true;
 454         }
 455
 456         /*
 457          * Read in the read iolog and store it, reuse the infrastructure
 458          * for doing verifications.
 459          */
 460         str = malloc(4096);
 461         rfname = fname = malloc(256+16);
 462         act = malloc(256+16);
 463
 464         syncs = reads = writes = waits = 0;
 465         while ((p = fgets(str, 4096, td->io_log_rfile)) != NULL) {
 466                 struct io_piece *ipo;
 467                 int r;
 468                 unsigned long long ttime;
 469
 470                 if (td->io_log_version == 3) {
 471                         r = sscanf(p, "%llu %256s %256s %llu %u", &ttime, rfname, act,
 472                                                         &offset, &bytes);
 473                         delay = delay_since_ttime(td, ttime);
 474                         td->io_log_last_ttime = ttime;
 475                         /*
 476                          * "wait" is not allowed with version 3
 477                          */
 478                         if (!strcmp(act, "wait")) {
 479                                 log_err("iolog: ignoring wait command with"
 480                                         " version 3 for file %s\n", fname);
 481                                 continue;
 482                         }
 483                 } else /* version 2 */
 484                         r = sscanf(p, "%256s %256s %llu %u", rfname, act, &offset, &bytes);
 485
 486                 if (td->o.replay_redirect)
 487                         fname = td->o.replay_redirect;
 488
 489                 if (io_act(td, r)) {
 490                         /*
 491                          * Check action first
 492                          */
 493                         if (!strcmp(act, "wait"))
 494                                 rw = DDIR_WAIT;
 495                         else if (!strcmp(act, "read"))
 496                                 rw = DDIR_READ;
 497                         else if (!strcmp(act, "write"))
 498                                 rw = DDIR_WRITE;
 499                         else if (!strcmp(act, "sync"))
 500                                 rw = DDIR_SYNC;
 501                         else if (!strcmp(act, "datasync"))
 502                                 rw = DDIR_DATASYNC;
 503                         else if (!strcmp(act, "trim"))
 504                                 rw = DDIR_TRIM;
 505                         else {
 506                                 log_err("fio: bad iolog file action: %s\n",
 507                                                                         act);
 508                                 continue;
 509                         }
 510                         fileno = get_fileno(td, fname);
 511                 } else if (file_act(td, r)) {
 512                         rw = DDIR_INVAL;
 513                         if (!strcmp(act, "add")) {
 514                                 if (td->o.replay_redirect &&
 515                                     get_fileno(td, fname) != -1) {
 516                                         dprint(FD_FILE, "iolog: ignoring"
 517                                                 " re-add of file %s\n", fname);
 518                                 } else {
 519                                         fileno = add_file(td, fname, td->subjob_number, 1);
 520                                         file_action = FIO_LOG_ADD_FILE;
 521                                 }
 522                         } else if (!strcmp(act, "open")) {
 523                                 fileno = get_fileno(td, fname);
 524                                 file_action = FIO_LOG_OPEN_FILE;
 525                         } else if (!strcmp(act, "close")) {
 526                                 fileno = get_fileno(td, fname);
 527                                 file_action = FIO_LOG_CLOSE_FILE;
 528                         } else {
 529                                 log_err("fio: bad iolog file action: %s\n",
 530                                                                         act);
 531                                 continue;
 532                         }
 533                 } else {
 534                         log_err("bad iolog%d: %s\n", td->io_log_version, p);
 535                         continue;
 536                 }
 537
 538                 if (rw == DDIR_READ)
 539                         reads++;
 540                 else if (rw == DDIR_WRITE) {
 541                         /*
 542                          * Don't add a write for ro mode
 543                          */
 544                         if (read_only)
 545                                 continue;
 546                         writes++;
 547                 } else if (rw == DDIR_WAIT) {
 548                         if (td->o.no_stall)
 549                                 continue;
 550                         waits++;
 551                 } else if (rw == DDIR_INVAL) {
 552                 } else if (ddir_sync(rw)) {
 553                         syncs++;
 554                 } else {
 555                         log_err("bad ddir: %d\n", rw);
 556                         continue;
 557                 }
 558
 559                 /*
 560                  * Make note of file
 561                  */
 562                 ipo = calloc(1, sizeof(*ipo));
 563                 init_ipo(ipo);
 564                 ipo->ddir = rw;
 565                 if (td->io_log_version == 3)
 566                         ipo->delay = delay;
 567                 if (rw == DDIR_WAIT) {
 568                         ipo->delay = offset;
 569                 } else {
 570                         if (td->o.replay_scale)
 571                                 ipo->offset = offset / td->o.replay_scale;
 572                         else
 573                                 ipo->offset = offset;
 574                         ipo_bytes_align(td->o.replay_align, ipo);
 575
 576                         ipo->len = bytes;
 577                         if (rw != DDIR_INVAL && bytes > td->o.max_bs[rw]) {
 578                                 realloc = true;
 579                                 td->o.max_bs[rw] = bytes;
 580                         }
 581                         ipo->fileno = fileno;
 582                         ipo->file_action = file_action;
 583                         td->o.size += bytes;
 584                 }
 585
 586                 queue_io_piece(td, ipo);
 587
 588                 if (td->o.read_iolog_chunked) {
 589                         td->io_log_current++;
 590                         items_to_fetch--;
 591                         if (items_to_fetch == 0)
 592                                 break;
 593                 }
 594         }
 595
 596         free(str);
 597         free(act);
 598         free(rfname);
 599
 600         if (td->o.read_iolog_chunked) {
 601                 td->io_log_highmark = td->io_log_current;
 602                 td->io_log_checkmark = (td->io_log_highmark + 1) / 2;
 603                 fio_gettime(&td->io_log_highmark_time, NULL);
 604         }
 605
 606         if (writes && read_only) {
 607                 log_err("fio: <%s> skips replay of %d writes due to"
 608                         " read-only\n", td->o.name, writes);
 609                 writes = 0;
 610         }
 611         if (syncs)
 612                 td->flags |= TD_F_SYNCS;
 613
 614         if (td->o.read_iolog_chunked) {
 615                 if (td->io_log_current == 0) {
 616                         return false;
 617                 }
 618                 td->o.td_ddir = TD_DDIR_RW;
 619                 if (realloc && td->orig_buffer)
 620                 {
 621                         io_u_quiesce(td);
 622                         free_io_mem(td);
 623                         init_io_u_buffers(td);
 624                 }
 625                 return true;
 626         }
 627
 628         if (!reads && !writes && !waits)
 629                 return false;
 630         else if (reads && !writes)
 631                 td->o.td_ddir = TD_DDIR_READ;
 632         else if (!reads && writes)
 633                 td->o.td_ddir = TD_DDIR_WRITE;
 634         else
 635                 td->o.td_ddir = TD_DDIR_RW;
 636
 637         return true;
 638 }
 639
 640 static bool is_socket(const char *path)
 641 {
 642         struct stat buf;
 643         int r;
 644
 645         r = stat(path, &buf);
 646         if (r == -1)
 647                 return false;
 648
 649         return S_ISSOCK(buf.st_mode);
 650 }
 651
 652 static int open_socket(const char *path)
 653 {
 654         struct sockaddr_un addr;
 655         int ret, fd;
 656
 657         fd = socket(AF_UNIX, SOCK_STREAM, 0);
 658         if (fd < 0)
 659                 return fd;
 660
 661         addr.sun_family = AF_UNIX;
 662         if (snprintf(addr.sun_path, sizeof(addr.sun_path), "%s", path) >=
 663             sizeof(addr.sun_path)) {
 664                 log_err("%s: path name %s is too long for a Unix socket\n",
 665                         __func__, path);
 666         }
 667
 668         ret = connect(fd, (const struct sockaddr *)&addr, strlen(path) + sizeof(addr.sun_family));
 669         if (!ret)
 670                 return fd;
 671
 672         close(fd);
 673         return -1;
 674 }
 675
 676 /*
 677  * open iolog, check version, and call appropriate parser
 678  */
 679 static bool init_iolog_read(struct thread_data *td, char *fname)
 680 {
 681         char buffer[256], *p;
 682         FILE *f = NULL;
 683
 684         dprint(FD_IO, "iolog: name=%s\n", fname);
 685
 686         if (is_socket(fname)) {
 687                 int fd;
 688
 689                 fd = open_socket(fname);
 690                 if (fd >= 0)
 691                         f = fdopen(fd, "r");
 692         } else if (!strcmp(fname, "-")) {
 693                 f = stdin;
 694         } else
 695                 f = fopen(fname, "r");
 696
 697         if (!f) {
 698                 perror("fopen read iolog");
 699                 return false;
 700         }
 701
 702         p = fgets(buffer, sizeof(buffer), f);
 703         if (!p) {
 704                 td_verror(td, errno, "iolog read");
 705                 log_err("fio: unable to read iolog\n");
 706                 fclose(f);
 707                 return false;
 708         }
 709
 710         /*
 711          * versions 2 and 3 of the iolog store a specific string as the
 712          * first line, check for that
 713          */
 714         if (!strncmp(iolog_ver2, buffer, strlen(iolog_ver2)))
 715                 td->io_log_version = 2;
 716         else if (!strncmp(iolog_ver3, buffer, strlen(iolog_ver3)))
 717                 td->io_log_version = 3;
 718         else {
 719                 log_err("fio: iolog version 1 is no longer supported\n");
 720                 fclose(f);
 721                 return false;
 722         }
 723
 724         free_release_files(td);
 725         td->io_log_rfile = f;
 726         return read_iolog(td);
 727 }
 728
 729 /*
 730  * Set up a log for storing io patterns.
 731  */
 732 static bool init_iolog_write(struct thread_data *td)
 733 {
 734         struct fio_file *ff;
 735         FILE *f;
 736         unsigned int i;
 737
 738         f = fopen(td->o.write_iolog_file, "a");
 739         if (!f) {
 740                 perror("fopen write iolog");
 741                 return false;
 742         }
 743
 744         /*
 745          * That's it for writing, setup a log buffer and we're done.
 746           */
 747         td->iolog_f = f;
 748         td->iolog_buf = malloc(8192);
 749         setvbuf(f, td->iolog_buf, _IOFBF, 8192);
 750         fio_gettime(&td->io_log_start_time, NULL);
 751
 752         /*
 753          * write our version line
 754          */
 755         if (fprintf(f, "%s\n", iolog_ver3) < 0) {
 756                 perror("iolog init\n");
 757                 return false;
 758         }
 759
 760         /*
 761          * add all known files
 762          */
 763         for_each_file(td, ff, i)
 764                 log_file(td, ff, FIO_LOG_ADD_FILE);
 765
 766         return true;
 767 }
 768
 769 bool init_iolog(struct thread_data *td)
 770 {
 771         bool ret;
 772
 773         if (td->o.read_iolog_file) {
 774                 int need_swap;
 775                 char * fname = get_name_by_idx(td->o.read_iolog_file, td->subjob_number);
 776
 777                 /*
 778                  * Check if it's a blktrace file and load that if possible.
 779                  * Otherwise assume it's a normal log file and load that.
 780                  */
 781                 if (is_blktrace(fname, &need_swap)) {
 782                         td->io_log_blktrace = 1;
 783                         ret = init_blktrace_read(td, fname, need_swap);
 784                 } else {
 785                         td->io_log_blktrace = 0;
 786                         ret = init_iolog_read(td, fname);
 787                 }
 788                 free(fname);
 789         } else if (td->o.write_iolog_file)
 790                 ret = init_iolog_write(td);
 791         else
 792                 ret = true;
 793
 794         if (!ret)
 795                 td_verror(td, EINVAL, "failed initializing iolog");
 796
 797         return ret;
 798 }
 799
 800 void setup_log(struct io_log **log, struct log_params *p,
 801                const char *filename)
 802 {
 803         struct io_log *l;
 804         int i;
 805         struct io_u_plat_entry *entry;
 806         struct flist_head *list;
 807
 808         l = scalloc(1, sizeof(*l));
 809         INIT_FLIST_HEAD(&l->io_logs);
 810         l->log_type = p->log_type;
 811         l->log_offset = p->log_offset;
 812         l->log_prio = p->log_prio;
 813         l->log_gz = p->log_gz;
 814         l->log_gz_store = p->log_gz_store;
 815         l->avg_msec = p->avg_msec;
 816         l->hist_msec = p->hist_msec;
 817         l->hist_coarseness = p->hist_coarseness;
 818         l->filename = strdup(filename);
 819         l->td = p->td;
 820
 821         /* Initialize histogram lists for each r/w direction,
 822          * with initial io_u_plat of all zeros:
 823          */
 824         for (i = 0; i < DDIR_RWDIR_CNT; i++) {
 825                 list = &l->hist_window[i].list;
 826                 INIT_FLIST_HEAD(list);
 827                 entry = calloc(1, sizeof(struct io_u_plat_entry));
 828                 flist_add(&entry->list, list);
 829         }
 830
 831         if (l->td && l->td->o.io_submit_mode != IO_MODE_OFFLOAD) {
 832                 unsigned int def_samples = DEF_LOG_ENTRIES;
 833                 struct io_logs *__p;
 834
 835                 __p = calloc(1, sizeof(*l->pending));
 836                 if (l->td->o.iodepth > DEF_LOG_ENTRIES)
 837                         def_samples = roundup_pow2(l->td->o.iodepth);
 838                 __p->max_samples = def_samples;
 839                 __p->log = calloc(__p->max_samples, log_entry_sz(l));
 840                 l->pending = __p;
 841         }
 842
 843         if (l->log_offset)
 844                 l->log_ddir_mask = LOG_OFFSET_SAMPLE_BIT;
 845         if (l->log_prio)
 846                 l->log_ddir_mask |= LOG_PRIO_SAMPLE_BIT;
 847
 848         INIT_FLIST_HEAD(&l->chunk_list);
 849
 850         if (l->log_gz && !p->td)
 851                 l->log_gz = 0;
 852         else if (l->log_gz || l->log_gz_store) {
 853                 mutex_init_pshared(&l->chunk_lock);
 854                 mutex_init_pshared(&l->deferred_free_lock);
 855                 p->td->flags |= TD_F_COMPRESS_LOG;
 856         }
 857
 858         *log = l;
 859 }
 860
 861 #ifdef CONFIG_SETVBUF
 862 static void *set_file_buffer(FILE *f)
 863 {
 864         size_t size = 1048576;
 865         void *buf;
 866
 867         buf = malloc(size);
 868         setvbuf(f, buf, _IOFBF, size);
 869         return buf;
 870 }
 871
 872 static void clear_file_buffer(void *buf)
 873 {
 874         free(buf);
 875 }
 876 #else
 877 static void *set_file_buffer(FILE *f)
 878 {
 879         return NULL;
 880 }
 881
 882 static void clear_file_buffer(void *buf)
 883 {
 884 }
 885 #endif
 886
 887 void free_log(struct io_log *log)
 888 {
 889         while (!flist_empty(&log->io_logs)) {
 890                 struct io_logs *cur_log;
 891
 892                 cur_log = flist_first_entry(&log->io_logs, struct io_logs, list);
 893                 flist_del_init(&cur_log->list);
 894                 free(cur_log->log);
 895                 sfree(cur_log);
 896         }
 897
 898         if (log->pending) {
 899                 free(log->pending->log);
 900                 free(log->pending);
 901                 log->pending = NULL;
 902         }
 903
 904         free(log->pending);
 905         free(log->filename);
 906         sfree(log);
 907 }
 908
 909 uint64_t hist_sum(int j, int stride, uint64_t *io_u_plat,
 910                 uint64_t *io_u_plat_last)
 911 {
 912         uint64_t sum;
 913         int k;
 914
 915         if (io_u_plat_last) {
 916                 for (k = sum = 0; k < stride; k++)
 917                         sum += io_u_plat[j + k] - io_u_plat_last[j + k];
 918         } else {
 919                 for (k = sum = 0; k < stride; k++)
 920                         sum += io_u_plat[j + k];
 921         }
 922
 923         return sum;
 924 }
 925
 926 static void flush_hist_samples(FILE *f, int hist_coarseness, void *samples,
 927                                uint64_t sample_size)
 928 {
 929         struct io_sample *s;
 930         int log_offset;
 931         uint64_t i, j, nr_samples;
 932         struct io_u_plat_entry *entry, *entry_before;
 933         uint64_t *io_u_plat;
 934         uint64_t *io_u_plat_before;
 935
 936         int stride = 1 << hist_coarseness;
 937
 938         if (!sample_size)
 939                 return;
 940
 941         s = __get_sample(samples, 0, 0);
 942         log_offset = (s->__ddir & LOG_OFFSET_SAMPLE_BIT) != 0;
 943
 944         nr_samples = sample_size / __log_entry_sz(log_offset);
 945
 946         for (i = 0; i < nr_samples; i++) {
 947                 s = __get_sample(samples, log_offset, i);
 948
 949                 entry = s->data.plat_entry;
 950                 io_u_plat = entry->io_u_plat;
 951
 952                 entry_before = flist_first_entry(&entry->list, struct io_u_plat_entry, list);
 953                 io_u_plat_before = entry_before->io_u_plat;
 954
 955                 fprintf(f, "%lu, %u, %llu, ", (unsigned long) s->time,
 956                                                 io_sample_ddir(s), (unsigned long long) s->bs);
 957                 for (j = 0; j < FIO_IO_U_PLAT_NR - stride; j += stride) {
 958                         fprintf(f, "%llu, ", (unsigned long long)
 959                                 hist_sum(j, stride, io_u_plat, io_u_plat_before));
 960                 }
 961                 fprintf(f, "%llu\n", (unsigned long long)
 962                         hist_sum(FIO_IO_U_PLAT_NR - stride, stride, io_u_plat,
 963                                         io_u_plat_before));
 964
 965                 flist_del(&entry_before->list);
 966                 free(entry_before);
 967         }
 968 }
 969
 970 void flush_samples(FILE *f, void *samples, uint64_t sample_size)
 971 {
 972         struct io_sample *s;
 973         int log_offset, log_prio;
 974         uint64_t i, nr_samples;
 975         unsigned int prio_val;
 976         const char *fmt;
 977
 978         if (!sample_size)
 979                 return;
 980
 981         s = __get_sample(samples, 0, 0);
 982         log_offset = (s->__ddir & LOG_OFFSET_SAMPLE_BIT) != 0;
 983         log_prio = (s->__ddir & LOG_PRIO_SAMPLE_BIT) != 0;
 984
 985         if (log_offset) {
 986                 if (log_prio)
 987                         fmt = "%lu, %" PRId64 ", %u, %llu, %llu, 0x%04x\n";
 988                 else
 989                         fmt = "%lu, %" PRId64 ", %u, %llu, %llu, %u\n";
 990         } else {
 991                 if (log_prio)
 992                         fmt = "%lu, %" PRId64 ", %u, %llu, 0x%04x\n";
 993                 else
 994                         fmt = "%lu, %" PRId64 ", %u, %llu, %u\n";
 995         }
 996
 997         nr_samples = sample_size / __log_entry_sz(log_offset);
 998
 999         for (i = 0; i < nr_samples; i++) {
1000                 s = __get_sample(samples, log_offset, i);
1001
1002                 if (log_prio)
1003                         prio_val = s->priority;
1004                 else
1005                         prio_val = ioprio_value_is_class_rt(s->priority);
1006
1007                 if (!log_offset) {
1008                         fprintf(f, fmt,
1009                                 (unsigned long) s->time,
1010                                 s->data.val,
1011                                 io_sample_ddir(s), (unsigned long long) s->bs,
1012                                 prio_val);
1013                 } else {
1014                         struct io_sample_offset *so = (void *) s;
1015
1016                         fprintf(f, fmt,
1017                                 (unsigned long) s->time,
1018                                 s->data.val,
1019                                 io_sample_ddir(s), (unsigned long long) s->bs,
1020                                 (unsigned long long) so->offset,
1021                                 prio_val);
1022                 }
1023         }
1024 }
1025
1026 #ifdef CONFIG_ZLIB
1027
1028 struct iolog_flush_data {
1029         struct workqueue_work work;
1030         struct io_log *log;
1031         void *samples;
1032         uint32_t nr_samples;
1033         bool free;
1034 };
1035
1036 #define GZ_CHUNK        131072
1037
1038 static struct iolog_compress *get_new_chunk(unsigned int seq)
1039 {
1040         struct iolog_compress *c;
1041
1042         c = malloc(sizeof(*c));
1043         INIT_FLIST_HEAD(&c->list);
1044         c->buf = malloc(GZ_CHUNK);
1045         c->len = 0;
1046         c->seq = seq;
1047         return c;
1048 }
1049
1050 static void free_chunk(struct iolog_compress *ic)
1051 {
1052         free(ic->buf);
1053         free(ic);
1054 }
1055
1056 static int z_stream_init(z_stream *stream, int gz_hdr)
1057 {
1058         int wbits = 15;
1059
1060         memset(stream, 0, sizeof(*stream));
1061         stream->zalloc = Z_NULL;
1062         stream->zfree = Z_NULL;
1063         stream->opaque = Z_NULL;
1064         stream->next_in = Z_NULL;
1065
1066         /*
1067          * zlib magic - add 32 for auto-detection of gz header or not,
1068          * if we decide to store files in a gzip friendly format.
1069          */
1070         if (gz_hdr)
1071                 wbits += 32;
1072
1073         if (inflateInit2(stream, wbits) != Z_OK)
1074                 return 1;
1075
1076         return 0;
1077 }
1078
1079 struct inflate_chunk_iter {
1080         unsigned int seq;
1081         int err;
1082         void *buf;
1083         size_t buf_size;
1084         size_t buf_used;
1085         size_t chunk_sz;
1086 };
1087
1088 static void finish_chunk(z_stream *stream, FILE *f,
1089                          struct inflate_chunk_iter *iter)
1090 {
1091         int ret;
1092
1093         ret = inflateEnd(stream);
1094         if (ret != Z_OK)
1095                 log_err("fio: failed to end log inflation seq %d (%d)\n",
1096                                 iter->seq, ret);
1097
1098         flush_samples(f, iter->buf, iter->buf_used);
1099         free(iter->buf);
1100         iter->buf = NULL;
1101         iter->buf_size = iter->buf_used = 0;
1102 }
1103
1104 /*
1105  * Iterative chunk inflation. Handles cases where we cross into a new
1106  * sequence, doing flush finish of previous chunk if needed.
1107  */
1108 static size_t inflate_chunk(struct iolog_compress *ic, int gz_hdr, FILE *f,
1109                             z_stream *stream, struct inflate_chunk_iter *iter)
1110 {
1111         size_t ret;
1112
1113         dprint(FD_COMPRESS, "inflate chunk size=%lu, seq=%u\n",
1114                                 (unsigned long) ic->len, ic->seq);
1115
1116         if (ic->seq != iter->seq) {
1117                 if (iter->seq)
1118                         finish_chunk(stream, f, iter);
1119
1120                 z_stream_init(stream, gz_hdr);
1121                 iter->seq = ic->seq;
1122         }
1123
1124         stream->avail_in = ic->len;
1125         stream->next_in = ic->buf;
1126
1127         if (!iter->buf_size) {
1128                 iter->buf_size = iter->chunk_sz;
1129                 iter->buf = malloc(iter->buf_size);
1130         }
1131
1132         while (stream->avail_in) {
1133                 size_t this_out = iter->buf_size - iter->buf_used;
1134                 int err;
1135
1136                 stream->avail_out = this_out;
1137                 stream->next_out = iter->buf + iter->buf_used;
1138
1139                 err = inflate(stream, Z_NO_FLUSH);
1140                 if (err < 0) {
1141                         log_err("fio: failed inflating log: %d\n", err);
1142                         iter->err = err;
1143                         break;
1144                 }
1145
1146                 iter->buf_used += this_out - stream->avail_out;
1147
1148                 if (!stream->avail_out) {
1149                         iter->buf_size += iter->chunk_sz;
1150                         iter->buf = realloc(iter->buf, iter->buf_size);
1151                         continue;
1152                 }
1153
1154                 if (err == Z_STREAM_END)
1155                         break;
1156         }
1157
1158         ret = (void *) stream->next_in - ic->buf;
1159
1160         dprint(FD_COMPRESS, "inflated to size=%lu\n", (unsigned long) iter->buf_size);
1161
1162         return ret;
1163 }
1164
1165 /*
1166  * Inflate stored compressed chunks, or write them directly to the log
1167  * file if so instructed.
1168  */
1169 static int inflate_gz_chunks(struct io_log *log, FILE *f)
1170 {
1171         struct inflate_chunk_iter iter = { .chunk_sz = log->log_gz, };
1172         z_stream stream;
1173
1174         while (!flist_empty(&log->chunk_list)) {
1175                 struct iolog_compress *ic;
1176
1177                 ic = flist_first_entry(&log->chunk_list, struct iolog_compress, list);
1178                 flist_del(&ic->list);
1179
1180                 if (log->log_gz_store) {
1181                         size_t ret;
1182
1183                         dprint(FD_COMPRESS, "log write chunk size=%lu, "
1184                                 "seq=%u\n", (unsigned long) ic->len, ic->seq);
1185
1186                         ret = fwrite(ic->buf, ic->len, 1, f);
1187                         if (ret != 1 || ferror(f)) {
1188                                 iter.err = errno;
1189                                 log_err("fio: error writing compressed log\n");
1190                         }
1191                 } else
1192                         inflate_chunk(ic, log->log_gz_store, f, &stream, &iter);
1193
1194                 free_chunk(ic);
1195         }
1196
1197         if (iter.seq) {
1198                 finish_chunk(&stream, f, &iter);
1199                 free(iter.buf);
1200         }
1201
1202         return iter.err;
1203 }
1204
1205 /*
1206  * Open compressed log file and decompress the stored chunks and
1207  * write them to stdout. The chunks are stored sequentially in the
1208  * file, so we iterate over them and do them one-by-one.
1209  */
1210 int iolog_file_inflate(const char *file)
1211 {
1212         struct inflate_chunk_iter iter = { .chunk_sz = 64 * 1024 * 1024, };
1213         struct iolog_compress ic;
1214         z_stream stream;
1215         struct stat sb;
1216         size_t ret;
1217         size_t total;
1218         void *buf;
1219         FILE *f;
1220
1221         f = fopen(file, "rb");
1222         if (!f) {
1223                 perror("fopen");
1224                 return 1;
1225         }
1226
1227         if (stat(file, &sb) < 0) {
1228                 fclose(f);
1229                 perror("stat");
1230                 return 1;
1231         }
1232
1233         ic.buf = buf = malloc(sb.st_size);
1234         ic.len = sb.st_size;
1235         ic.seq = 1;
1236
1237         ret = fread(ic.buf, ic.len, 1, f);
1238         if (ret == 0 && ferror(f)) {
1239                 perror("fread");
1240                 fclose(f);
1241                 free(buf);
1242                 return 1;
1243         } else if (ferror(f) || (!feof(f) && ret != 1)) {
1244                 log_err("fio: short read on reading log\n");
1245                 fclose(f);
1246                 free(buf);
1247                 return 1;
1248         }
1249
1250         fclose(f);
1251
1252         /*
1253          * Each chunk will return Z_STREAM_END. We don't know how many
1254          * chunks are in the file, so we just keep looping and incrementing
1255          * the sequence number until we have consumed the whole compressed
1256          * file.
1257          */
1258         total = ic.len;
1259         do {
1260                 size_t iret;
1261
1262                 iret = inflate_chunk(&ic,  1, stdout, &stream, &iter);
1263                 total -= iret;
1264                 if (!total)
1265                         break;
1266                 if (iter.err)
1267                         break;
1268
1269                 ic.seq++;
1270                 ic.len -= iret;
1271                 ic.buf += iret;
1272         } while (1);
1273
1274         if (iter.seq) {
1275                 finish_chunk(&stream, stdout, &iter);
1276                 free(iter.buf);
1277         }
1278
1279         free(buf);
1280         return iter.err;
1281 }
1282
1283 #else
1284
1285 static int inflate_gz_chunks(struct io_log *log, FILE *f)
1286 {
1287         return 0;
1288 }
1289
1290 int iolog_file_inflate(const char *file)
1291 {
1292         log_err("fio: log inflation not possible without zlib\n");
1293         return 1;
1294 }
1295
1296 #endif
1297
1298 void flush_log(struct io_log *log, bool do_append)
1299 {
1300         void *buf;
1301         FILE *f;
1302
1303         /*
1304          * If log_gz_store is true, we are writing a binary file.
1305          * Set the mode appropriately (on all platforms) to avoid issues
1306          * on windows (line-ending conversions, etc.)
1307          */
1308         if (!do_append)
1309                 if (log->log_gz_store)
1310                         f = fopen(log->filename, "wb");
1311                 else
1312                         f = fopen(log->filename, "w");
1313         else
1314                 if (log->log_gz_store)
1315                         f = fopen(log->filename, "ab");
1316                 else
1317                         f = fopen(log->filename, "a");
1318         if (!f) {
1319                 perror("fopen log");
1320                 return;
1321         }
1322
1323         buf = set_file_buffer(f);
1324
1325         inflate_gz_chunks(log, f);
1326
1327         while (!flist_empty(&log->io_logs)) {
1328                 struct io_logs *cur_log;
1329
1330                 cur_log = flist_first_entry(&log->io_logs, struct io_logs, list);
1331                 flist_del_init(&cur_log->list);
1332
1333                 if (log->td && log == log->td->clat_hist_log)
1334                         flush_hist_samples(f, log->hist_coarseness, cur_log->log,
1335                                            log_sample_sz(log, cur_log));
1336                 else
1337                         flush_samples(f, cur_log->log, log_sample_sz(log, cur_log));
1338
1339                 sfree(cur_log);
1340         }
1341
1342         fclose(f);
1343         clear_file_buffer(buf);
1344 }
1345
1346 static int finish_log(struct thread_data *td, struct io_log *log, int trylock)
1347 {
1348         if (td->flags & TD_F_COMPRESS_LOG)
1349                 iolog_flush(log);
1350
1351         if (trylock) {
1352                 if (fio_trylock_file(log->filename))
1353                         return 1;
1354         } else
1355                 fio_lock_file(log->filename);
1356
1357         if (td->client_type == FIO_CLIENT_TYPE_GUI || is_backend)
1358                 fio_send_iolog(td, log, log->filename);
1359         else
1360                 flush_log(log, !td->o.per_job_logs);
1361
1362         fio_unlock_file(log->filename);
1363         free_log(log);
1364         return 0;
1365 }
1366
1367 size_t log_chunk_sizes(struct io_log *log)
1368 {
1369         struct flist_head *entry;
1370         size_t ret;
1371
1372         if (flist_empty(&log->chunk_list))
1373                 return 0;
1374
1375         ret = 0;
1376         pthread_mutex_lock(&log->chunk_lock);
1377         flist_for_each(entry, &log->chunk_list) {
1378                 struct iolog_compress *c;
1379
1380                 c = flist_entry(entry, struct iolog_compress, list);
1381                 ret += c->len;
1382         }
1383         pthread_mutex_unlock(&log->chunk_lock);
1384         return ret;
1385 }
1386
1387 #ifdef CONFIG_ZLIB
1388
1389 static void iolog_put_deferred(struct io_log *log, void *ptr)
1390 {
1391         if (!ptr)
1392                 return;
1393
1394         pthread_mutex_lock(&log->deferred_free_lock);
1395         if (log->deferred < IOLOG_MAX_DEFER) {
1396                 log->deferred_items[log->deferred] = ptr;
1397                 log->deferred++;
1398         } else if (!fio_did_warn(FIO_WARN_IOLOG_DROP))
1399                 log_err("fio: had to drop log entry free\n");
1400         pthread_mutex_unlock(&log->deferred_free_lock);
1401 }
1402
1403 static void iolog_free_deferred(struct io_log *log)
1404 {
1405         int i;
1406
1407         if (!log->deferred)
1408                 return;
1409
1410         pthread_mutex_lock(&log->deferred_free_lock);
1411
1412         for (i = 0; i < log->deferred; i++) {
1413                 free(log->deferred_items[i]);
1414                 log->deferred_items[i] = NULL;
1415         }
1416
1417         log->deferred = 0;
1418         pthread_mutex_unlock(&log->deferred_free_lock);
1419 }
1420
1421 static int gz_work(struct iolog_flush_data *data)
1422 {
1423         struct iolog_compress *c = NULL;
1424         struct flist_head list;
1425         unsigned int seq;
1426         z_stream stream;
1427         size_t total = 0;
1428         int ret;
1429
1430         INIT_FLIST_HEAD(&list);
1431
1432         memset(&stream, 0, sizeof(stream));
1433         stream.zalloc = Z_NULL;
1434         stream.zfree = Z_NULL;
1435         stream.opaque = Z_NULL;
1436
1437         ret = deflateInit(&stream, Z_DEFAULT_COMPRESSION);
1438         if (ret != Z_OK) {
1439                 log_err("fio: failed to init gz stream\n");
1440                 goto err;
1441         }
1442
1443         seq = ++data->log->chunk_seq;
1444
1445         stream.next_in = (void *) data->samples;
1446         stream.avail_in = data->nr_samples * log_entry_sz(data->log);
1447
1448         dprint(FD_COMPRESS, "deflate input size=%lu, seq=%u, log=%s\n",
1449                                 (unsigned long) stream.avail_in, seq,
1450                                 data->log->filename);
1451         do {
1452                 if (c)
1453                         dprint(FD_COMPRESS, "seq=%d, chunk=%lu\n", seq,
1454                                 (unsigned long) c->len);
1455                 c = get_new_chunk(seq);
1456                 stream.avail_out = GZ_CHUNK;
1457                 stream.next_out = c->buf;
1458                 ret = deflate(&stream, Z_NO_FLUSH);
1459                 if (ret < 0) {
1460                         log_err("fio: deflate log (%d)\n", ret);
1461                         free_chunk(c);
1462                         goto err;
1463                 }
1464
1465                 c->len = GZ_CHUNK - stream.avail_out;
1466                 flist_add_tail(&c->list, &list);
1467                 total += c->len;
1468         } while (stream.avail_in);
1469
1470         stream.next_out = c->buf + c->len;
1471         stream.avail_out = GZ_CHUNK - c->len;
1472
1473         ret = deflate(&stream, Z_FINISH);
1474         if (ret < 0) {
1475                 /*
1476                  * Z_BUF_ERROR is special, it just means we need more
1477                  * output space. We'll handle that below. Treat any other
1478                  * error as fatal.
1479                  */
1480                 if (ret != Z_BUF_ERROR) {
1481                         log_err("fio: deflate log (%d)\n", ret);
1482                         flist_del(&c->list);
1483                         free_chunk(c);
1484                         goto err;
1485                 }
1486         }
1487
1488         total -= c->len;
1489         c->len = GZ_CHUNK - stream.avail_out;
1490         total += c->len;
1491         dprint(FD_COMPRESS, "seq=%d, chunk=%lu\n", seq, (unsigned long) c->len);
1492
1493         if (ret != Z_STREAM_END) {
1494                 do {
1495                         c = get_new_chunk(seq);
1496                         stream.avail_out = GZ_CHUNK;
1497                         stream.next_out = c->buf;
1498                         ret = deflate(&stream, Z_FINISH);
1499                         c->len = GZ_CHUNK - stream.avail_out;
1500                         total += c->len;
1501                         flist_add_tail(&c->list, &list);
1502                         dprint(FD_COMPRESS, "seq=%d, chunk=%lu\n", seq,
1503                                 (unsigned long) c->len);
1504                 } while (ret != Z_STREAM_END);
1505         }
1506
1507         dprint(FD_COMPRESS, "deflated to size=%lu\n", (unsigned long) total);
1508
1509         ret = deflateEnd(&stream);
1510         if (ret != Z_OK)
1511                 log_err("fio: deflateEnd %d\n", ret);
1512
1513         iolog_put_deferred(data->log, data->samples);
1514
1515         if (!flist_empty(&list)) {
1516                 pthread_mutex_lock(&data->log->chunk_lock);
1517                 flist_splice_tail(&list, &data->log->chunk_list);
1518                 pthread_mutex_unlock(&data->log->chunk_lock);
1519         }
1520
1521         ret = 0;
1522 done:
1523         if (data->free)
1524                 sfree(data);
1525         return ret;
1526 err:
1527         while (!flist_empty(&list)) {
1528                 c = flist_first_entry(list.next, struct iolog_compress, list);
1529                 flist_del(&c->list);
1530                 free_chunk(c);
1531         }
1532         ret = 1;
1533         goto done;
1534 }
1535
1536 /*
1537  * Invoked from our compress helper thread, when logging would have exceeded
1538  * the specified memory limitation. Compresses the previously stored
1539  * entries.
1540  */
1541 static int gz_work_async(struct submit_worker *sw, struct workqueue_work *work)
1542 {
1543         return gz_work(container_of(work, struct iolog_flush_data, work));
1544 }
1545
1546 static int gz_init_worker(struct submit_worker *sw)
1547 {
1548         struct thread_data *td = sw->wq->td;
1549
1550         if (!fio_option_is_set(&td->o, log_gz_cpumask))
1551                 return 0;
1552
1553         if (fio_setaffinity(gettid(), td->o.log_gz_cpumask) == -1) {
1554                 log_err("gz: failed to set CPU affinity\n");
1555                 return 1;
1556         }
1557
1558         return 0;
1559 }
1560
1561 static struct workqueue_ops log_compress_wq_ops = {
1562         .fn             = gz_work_async,
1563         .init_worker_fn = gz_init_worker,
1564         .nice           = 1,
1565 };
1566
1567 int iolog_compress_init(struct thread_data *td, struct sk_out *sk_out)
1568 {
1569         if (!(td->flags & TD_F_COMPRESS_LOG))
1570                 return 0;
1571
1572         workqueue_init(td, &td->log_compress_wq, &log_compress_wq_ops, 1, sk_out);
1573         return 0;
1574 }
1575
1576 void iolog_compress_exit(struct thread_data *td)
1577 {
1578         if (!(td->flags & TD_F_COMPRESS_LOG))
1579                 return;
1580
1581         workqueue_exit(&td->log_compress_wq);
1582 }
1583
1584 /*
1585  * Queue work item to compress the existing log entries. We reset the
1586  * current log to a small size, and reference the existing log in the
1587  * data that we queue for compression. Once compression has been done,
1588  * this old log is freed. Will not return until the log compression
1589  * has completed, and will flush all previous logs too
1590  */
1591 static int iolog_flush(struct io_log *log)
1592 {
1593         struct iolog_flush_data *data;
1594
1595         workqueue_flush(&log->td->log_compress_wq);
1596         data = malloc(sizeof(*data));
1597         if (!data)
1598                 return 1;
1599
1600         data->log = log;
1601         data->free = false;
1602
1603         while (!flist_empty(&log->io_logs)) {
1604                 struct io_logs *cur_log;
1605
1606                 cur_log = flist_first_entry(&log->io_logs, struct io_logs, list);
1607                 flist_del_init(&cur_log->list);
1608
1609                 data->samples = cur_log->log;
1610                 data->nr_samples = cur_log->nr_samples;
1611
1612                 sfree(cur_log);
1613
1614                 gz_work(data);
1615         }
1616
1617         free(data);
1618         return 0;
1619 }
1620
1621 int iolog_cur_flush(struct io_log *log, struct io_logs *cur_log)
1622 {
1623         struct iolog_flush_data *data;
1624
1625         data = smalloc(sizeof(*data));
1626         if (!data)
1627                 return 1;
1628
1629         data->log = log;
1630
1631         data->samples = cur_log->log;
1632         data->nr_samples = cur_log->nr_samples;
1633         data->free = true;
1634
1635         cur_log->nr_samples = cur_log->max_samples = 0;
1636         cur_log->log = NULL;
1637
1638         workqueue_enqueue(&log->td->log_compress_wq, &data->work);
1639
1640         iolog_free_deferred(log);
1641
1642         return 0;
1643 }
1644 #else
1645
1646 static int iolog_flush(struct io_log *log)
1647 {
1648         return 1;
1649 }
1650
1651 int iolog_cur_flush(struct io_log *log, struct io_logs *cur_log)
1652 {
1653         return 1;
1654 }
1655
1656 int iolog_compress_init(struct thread_data *td, struct sk_out *sk_out)
1657 {
1658         return 0;
1659 }
1660
1661 void iolog_compress_exit(struct thread_data *td)
1662 {
1663 }
1664
1665 #endif
1666
1667 struct io_logs *iolog_cur_log(struct io_log *log)
1668 {
1669         if (flist_empty(&log->io_logs))
1670                 return NULL;
1671
1672         return flist_last_entry(&log->io_logs, struct io_logs, list);
1673 }
1674
1675 uint64_t iolog_nr_samples(struct io_log *iolog)
1676 {
1677         struct flist_head *entry;
1678         uint64_t ret = 0;
1679
1680         flist_for_each(entry, &iolog->io_logs) {
1681                 struct io_logs *cur_log;
1682
1683                 cur_log = flist_entry(entry, struct io_logs, list);
1684                 ret += cur_log->nr_samples;
1685         }
1686
1687         return ret;
1688 }
1689
1690 static int __write_log(struct thread_data *td, struct io_log *log, int try)
1691 {
1692         if (log)
1693                 return finish_log(td, log, try);
1694
1695         return 0;
1696 }
1697
1698 static int write_iops_log(struct thread_data *td, int try, bool unit_log)
1699 {
1700         int ret;
1701
1702         if (per_unit_log(td->iops_log) != unit_log)
1703                 return 0;
1704
1705         ret = __write_log(td, td->iops_log, try);
1706         if (!ret)
1707                 td->iops_log = NULL;
1708
1709         return ret;
1710 }
1711
1712 static int write_slat_log(struct thread_data *td, int try, bool unit_log)
1713 {
1714         int ret;
1715
1716         if (!unit_log)
1717                 return 0;
1718
1719         ret = __write_log(td, td->slat_log, try);
1720         if (!ret)
1721                 td->slat_log = NULL;
1722
1723         return ret;
1724 }
1725
1726 static int write_clat_log(struct thread_data *td, int try, bool unit_log)
1727 {
1728         int ret;
1729
1730         if (!unit_log)
1731                 return 0;
1732
1733         ret = __write_log(td, td->clat_log, try);
1734         if (!ret)
1735                 td->clat_log = NULL;
1736
1737         return ret;
1738 }
1739
1740 static int write_clat_hist_log(struct thread_data *td, int try, bool unit_log)
1741 {
1742         int ret;
1743
1744         if (!unit_log)
1745                 return 0;
1746
1747         ret = __write_log(td, td->clat_hist_log, try);
1748         if (!ret)
1749                 td->clat_hist_log = NULL;
1750
1751         return ret;
1752 }
1753
1754 static int write_lat_log(struct thread_data *td, int try, bool unit_log)
1755 {
1756         int ret;
1757
1758         if (!unit_log)
1759                 return 0;
1760
1761         ret = __write_log(td, td->lat_log, try);
1762         if (!ret)
1763                 td->lat_log = NULL;
1764
1765         return ret;
1766 }
1767
1768 static int write_bandw_log(struct thread_data *td, int try, bool unit_log)
1769 {
1770         int ret;
1771
1772         if (per_unit_log(td->bw_log) != unit_log)
1773                 return 0;
1774
1775         ret = __write_log(td, td->bw_log, try);
1776         if (!ret)
1777                 td->bw_log = NULL;
1778
1779         return ret;
1780 }
1781
1782 enum {
1783         BW_LOG_MASK     = 1,
1784         LAT_LOG_MASK    = 2,
1785         SLAT_LOG_MASK   = 4,
1786         CLAT_LOG_MASK   = 8,
1787         IOPS_LOG_MASK   = 16,
1788         CLAT_HIST_LOG_MASK = 32,
1789
1790         ALL_LOG_NR      = 6,
1791 };
1792
1793 struct log_type {
1794         unsigned int mask;
1795         int (*fn)(struct thread_data *, int, bool);
1796 };
1797
1798 static struct log_type log_types[] = {
1799         {
1800                 .mask   = BW_LOG_MASK,
1801                 .fn     = write_bandw_log,
1802         },
1803         {
1804                 .mask   = LAT_LOG_MASK,
1805                 .fn     = write_lat_log,
1806         },
1807         {
1808                 .mask   = SLAT_LOG_MASK,
1809                 .fn     = write_slat_log,
1810         },
1811         {
1812                 .mask   = CLAT_LOG_MASK,
1813                 .fn     = write_clat_log,
1814         },
1815         {
1816                 .mask   = IOPS_LOG_MASK,
1817                 .fn     = write_iops_log,
1818         },
1819         {
1820                 .mask   = CLAT_HIST_LOG_MASK,
1821                 .fn     = write_clat_hist_log,
1822         }
1823 };
1824
1825 void td_writeout_logs(struct thread_data *td, bool unit_logs)
1826 {
1827         unsigned int log_mask = 0;
1828         unsigned int log_left = ALL_LOG_NR;
1829         int old_state, i;
1830
1831         old_state = td_bump_runstate(td, TD_FINISHING);
1832
1833         finalize_logs(td, unit_logs);
1834
1835         while (log_left) {
1836                 int prev_log_left = log_left;
1837
1838                 for (i = 0; i < ALL_LOG_NR && log_left; i++) {
1839                         struct log_type *lt = &log_types[i];
1840                         int ret;
1841
1842                         if (!(log_mask & lt->mask)) {
1843                                 ret = lt->fn(td, log_left != 1, unit_logs);
1844                                 if (!ret) {
1845                                         log_left--;
1846                                         log_mask |= lt->mask;
1847                                 }
1848                         }
1849                 }
1850
1851                 if (prev_log_left == log_left)
1852                         usleep(5000);
1853         }
1854
1855         td_restore_runstate(td, old_state);
1856 }
1857
1858 void fio_writeout_logs(bool unit_logs)
1859 {
1860         struct thread_data *td;
1861         int i;
1862
1863         for_each_td(td, i)
1864                 td_writeout_logs(td, unit_logs);
1865 }