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