iolog.c

   1 /*
   2  * Code related to writing an iolog of what a thread is doing, and to
   3  * later read that back and replay
   4  */
   5 #include <stdio.h>
   6 #include <stdlib.h>
   7 #include <assert.h>
   8 #include <sys/types.h>
   9 #include <sys/stat.h>
  10 #include <unistd.h>
  11 #ifdef CONFIG_ZLIB
  12 #include <zlib.h>
  13 #endif
  14
  15 #include "flist.h"
  16 #include "fio.h"
  17 #include "trim.h"
  18 #include "filelock.h"
  19 #include "smalloc.h"
  20 #include "blktrace.h"
  21 #include "pshared.h"
  22
  23 #include <netinet/in.h>
  24 #include <netinet/tcp.h>
  25 #include <arpa/inet.h>
  26 #include <sys/stat.h>
  27 #include <sys/socket.h>
  28 #include <sys/un.h>
  29
  30 static int iolog_flush(struct io_log *log);
  31
  32 static const char iolog_ver2[] = "fio version 2 iolog";
  33
  34 void queue_io_piece(struct thread_data *td, struct io_piece *ipo)
  35 {
  36         flist_add_tail(&ipo->list, &td->io_log_list);
  37         td->total_io_size += ipo->len;
  38 }
  39
  40 void log_io_u(const struct thread_data *td, const struct io_u *io_u)
  41 {
  42         if (!td->o.write_iolog_file)
  43                 return;
  44
  45         fprintf(td->iolog_f, "%s %s %llu %llu\n", io_u->file->file_name,
  46                                                 io_ddir_name(io_u->ddir),
  47                                                 io_u->offset, io_u->buflen);
  48 }
  49
  50 void log_file(struct thread_data *td, struct fio_file *f,
  51               enum file_log_act what)
  52 {
  53         const char *act[] = { "add", "open", "close" };
  54
  55         assert(what < 3);
  56
  57         if (!td->o.write_iolog_file)
  58                 return;
  59
  60
  61         /*
  62          * this happens on the pre-open/close done before the job starts
  63          */
  64         if (!td->iolog_f)
  65                 return;
  66
  67         fprintf(td->iolog_f, "%s %s\n", f->file_name, act[what]);
  68 }
  69
  70 static void iolog_delay(struct thread_data *td, unsigned long delay)
  71 {
  72         uint64_t usec = utime_since_now(&td->last_issue);
  73         unsigned long orig_delay = delay;
  74         uint64_t this_delay;
  75         struct timespec ts;
  76
  77         if (delay < td->time_offset) {
  78                 td->time_offset = 0;
  79                 return;
  80         }
  81
  82         delay -= td->time_offset;
  83         if (delay < usec)
  84                 return;
  85
  86         delay -= usec;
  87
  88         fio_gettime(&ts, NULL);
  89         while (delay && !td->terminate) {
  90                 this_delay = delay;
  91                 if (this_delay > 500000)
  92                         this_delay = 500000;
  93
  94                 usec_sleep(td, this_delay);
  95                 delay -= this_delay;
  96         }
  97
  98         usec = utime_since_now(&ts);
  99         if (usec > orig_delay)
 100                 td->time_offset = usec - orig_delay;
 101         else
 102                 td->time_offset = 0;
 103 }
 104
 105 static int ipo_special(struct thread_data *td, struct io_piece *ipo)
 106 {
 107         struct fio_file *f;
 108         int ret;
 109
 110         /*
 111          * Not a special ipo
 112          */
 113         if (ipo->ddir != DDIR_INVAL)
 114                 return 0;
 115
 116         f = td->files[ipo->fileno];
 117
 118         switch (ipo->file_action) {
 119         case FIO_LOG_OPEN_FILE:
 120                 if (td->o.replay_redirect && fio_file_open(f)) {
 121                         dprint(FD_FILE, "iolog: ignoring re-open of file %s\n",
 122                                         f->file_name);
 123                         break;
 124                 }
 125                 ret = td_io_open_file(td, f);
 126                 if (!ret)
 127                         break;
 128                 td_verror(td, ret, "iolog open file");
 129                 return -1;
 130         case FIO_LOG_CLOSE_FILE:
 131                 td_io_close_file(td, f);
 132                 break;
 133         case FIO_LOG_UNLINK_FILE:
 134                 td_io_unlink_file(td, f);
 135                 break;
 136         default:
 137                 log_err("fio: bad file action %d\n", ipo->file_action);
 138                 break;
 139         }
 140
 141         return 1;
 142 }
 143
 144 static bool read_iolog2(struct thread_data *td);
 145
 146 int read_iolog_get(struct thread_data *td, struct io_u *io_u)
 147 {
 148         struct io_piece *ipo;
 149         unsigned long elapsed;
 150
 151         while (!flist_empty(&td->io_log_list)) {
 152                 int ret;
 153                 if (td->o.read_iolog_chunked) {
 154                         if (td->io_log_checkmark == td->io_log_current) {
 155                                 if (!read_iolog2(td))
 156                                         return 1;
 157                         }
 158                         td->io_log_current--;
 159                 }
 160                 ipo = flist_first_entry(&td->io_log_list, struct io_piece, list);
 161                 flist_del(&ipo->list);
 162                 remove_trim_entry(td, ipo);
 163
 164                 ret = ipo_special(td, ipo);
 165                 if (ret < 0) {
 166                         free(ipo);
 167                         break;
 168                 } else if (ret > 0) {
 169                         free(ipo);
 170                         continue;
 171                 }
 172
 173                 io_u->ddir = ipo->ddir;
 174                 if (ipo->ddir != DDIR_WAIT) {
 175                         io_u->offset = ipo->offset;
 176                         io_u->buflen = ipo->len;
 177                         io_u->file = td->files[ipo->fileno];
 178                         get_file(io_u->file);
 179                         dprint(FD_IO, "iolog: get %llu/%llu/%s\n", io_u->offset,
 180                                                 io_u->buflen, io_u->file->file_name);
 181                         if (ipo->delay)
 182                                 iolog_delay(td, ipo->delay);
 183                 } else {
 184                         elapsed = mtime_since_genesis();
 185                         if (ipo->delay > elapsed)
 186                                 usec_sleep(td, (ipo->delay - elapsed) * 1000);
 187                 }
 188
 189                 free(ipo);
 190
 191                 if (io_u->ddir != DDIR_WAIT)
 192                         return 0;
 193         }
 194
 195         td->done = 1;
 196         return 1;
 197 }
 198
 199 void prune_io_piece_log(struct thread_data *td)
 200 {
 201         struct io_piece *ipo;
 202         struct fio_rb_node *n;
 203
 204         while ((n = rb_first(&td->io_hist_tree)) != NULL) {
 205                 ipo = rb_entry(n, struct io_piece, rb_node);
 206                 rb_erase(n, &td->io_hist_tree);
 207                 remove_trim_entry(td, ipo);
 208                 td->io_hist_len--;
 209                 free(ipo);
 210         }
 211
 212         while (!flist_empty(&td->io_hist_list)) {
 213                 ipo = flist_first_entry(&td->io_hist_list, struct io_piece, list);
 214                 flist_del(&ipo->list);
 215                 remove_trim_entry(td, ipo);
 216                 td->io_hist_len--;
 217                 free(ipo);
 218         }
 219 }
 220
 221 /*
 222  * log a successful write, so we can unwind the log for verify
 223  */
 224 void log_io_piece(struct thread_data *td, struct io_u *io_u)
 225 {
 226         struct fio_rb_node **p, *parent;
 227         struct io_piece *ipo, *__ipo;
 228
 229         ipo = calloc(1, sizeof(struct io_piece));
 230         init_ipo(ipo);
 231         ipo->file = io_u->file;
 232         ipo->offset = io_u->offset;
 233         ipo->len = io_u->buflen;
 234         ipo->numberio = io_u->numberio;
 235         ipo->flags = IP_F_IN_FLIGHT;
 236
 237         io_u->ipo = ipo;
 238
 239         if (io_u_should_trim(td, io_u)) {
 240                 flist_add_tail(&ipo->trim_list, &td->trim_list);
 241                 td->trim_entries++;
 242         }
 243
 244         /*
 245          * Only sort writes if we don't have a random map in which case we need
 246          * to check for duplicate blocks and drop the old one, which we rely on
 247          * the rb insert/lookup for handling.
 248          */
 249         if (file_randommap(td, ipo->file)) {
 250                 INIT_FLIST_HEAD(&ipo->list);
 251                 flist_add_tail(&ipo->list, &td->io_hist_list);
 252                 ipo->flags |= IP_F_ONLIST;
 253                 td->io_hist_len++;
 254                 return;
 255         }
 256
 257         RB_CLEAR_NODE(&ipo->rb_node);
 258
 259         /*
 260          * Sort the entry into the verification list
 261          */
 262 restart:
 263         p = &td->io_hist_tree.rb_node;
 264         parent = NULL;
 265         while (*p) {
 266                 int overlap = 0;
 267                 parent = *p;
 268
 269                 __ipo = rb_entry(parent, struct io_piece, rb_node);
 270                 if (ipo->file < __ipo->file)
 271                         p = &(*p)->rb_left;
 272                 else if (ipo->file > __ipo->file)
 273                         p = &(*p)->rb_right;
 274                 else if (ipo->offset < __ipo->offset) {
 275                         p = &(*p)->rb_left;
 276                         overlap = ipo->offset + ipo->len > __ipo->offset;
 277                 }
 278                 else if (ipo->offset > __ipo->offset) {
 279                         p = &(*p)->rb_right;
 280                         overlap = __ipo->offset + __ipo->len > ipo->offset;
 281                 }
 282                 else
 283                         overlap = 1;
 284
 285                 if (overlap) {
 286                         dprint(FD_IO, "iolog: overlap %llu/%lu, %llu/%lu\n",
 287                                 __ipo->offset, __ipo->len,
 288                                 ipo->offset, ipo->len);
 289                         td->io_hist_len--;
 290                         rb_erase(parent, &td->io_hist_tree);
 291                         remove_trim_entry(td, __ipo);
 292                         if (!(__ipo->flags & IP_F_IN_FLIGHT))
 293                                 free(__ipo);
 294                         goto restart;
 295                 }
 296         }
 297
 298         rb_link_node(&ipo->rb_node, parent, p);
 299         rb_insert_color(&ipo->rb_node, &td->io_hist_tree);
 300         ipo->flags |= IP_F_ONRB;
 301         td->io_hist_len++;
 302 }
 303
 304 void unlog_io_piece(struct thread_data *td, struct io_u *io_u)
 305 {
 306         struct io_piece *ipo = io_u->ipo;
 307
 308         if (td->ts.nr_block_infos) {
 309                 uint32_t *info = io_u_block_info(td, io_u);
 310                 if (BLOCK_INFO_STATE(*info) < BLOCK_STATE_TRIM_FAILURE) {
 311                         if (io_u->ddir == DDIR_TRIM)
 312                                 *info = BLOCK_INFO_SET_STATE(*info,
 313                                                 BLOCK_STATE_TRIM_FAILURE);
 314                         else if (io_u->ddir == DDIR_WRITE)
 315                                 *info = BLOCK_INFO_SET_STATE(*info,
 316                                                 BLOCK_STATE_WRITE_FAILURE);
 317                 }
 318         }
 319
 320         if (!ipo)
 321                 return;
 322
 323         if (ipo->flags & IP_F_ONRB)
 324                 rb_erase(&ipo->rb_node, &td->io_hist_tree);
 325         else if (ipo->flags & IP_F_ONLIST)
 326                 flist_del(&ipo->list);
 327
 328         free(ipo);
 329         io_u->ipo = NULL;
 330         td->io_hist_len--;
 331 }
 332
 333 void trim_io_piece(const struct io_u *io_u)
 334 {
 335         struct io_piece *ipo = io_u->ipo;
 336
 337         if (!ipo)
 338                 return;
 339
 340         ipo->len = io_u->xfer_buflen - io_u->resid;
 341 }
 342
 343 void write_iolog_close(struct thread_data *td)
 344 {
 345         if (!td->iolog_f)
 346                 return;
 347
 348         fflush(td->iolog_f);
 349         fclose(td->iolog_f);
 350         free(td->iolog_buf);
 351         td->iolog_f = NULL;
 352         td->iolog_buf = NULL;
 353 }
 354
 355 static int64_t iolog_items_to_fetch(struct thread_data *td)
 356 {
 357         struct timespec now;
 358         uint64_t elapsed;
 359         uint64_t for_1s;
 360         int64_t items_to_fetch;
 361
 362         if (!td->io_log_highmark)
 363                 return 10;
 364
 365
 366         fio_gettime(&now, NULL);
 367         elapsed = ntime_since(&td->io_log_highmark_time, &now);
 368         if (elapsed) {
 369                 for_1s = (td->io_log_highmark - td->io_log_current) * 1000000000 / elapsed;
 370                 items_to_fetch = for_1s - td->io_log_current;
 371                 if (items_to_fetch < 0)
 372                         items_to_fetch = 0;
 373         } else
 374                 items_to_fetch = 0;
 375
 376         td->io_log_highmark = td->io_log_current + items_to_fetch;
 377         td->io_log_checkmark = (td->io_log_highmark + 1) / 2;
 378         fio_gettime(&td->io_log_highmark_time, NULL);
 379
 380         return items_to_fetch;
 381 }
 382
 383 /*
 384  * Read version 2 iolog data. It is enhanced to include per-file logging,
 385  * syncs, etc.
 386  */
 387 static bool read_iolog2(struct thread_data *td)
 388 {
 389         unsigned long long offset;
 390         unsigned int bytes;
 391         int reads, writes, waits, fileno = 0, file_action = 0; /* stupid gcc */
 392         char *rfname, *fname, *act;
 393         char *str, *p;
 394         enum fio_ddir rw;
 395         bool realloc = false;
 396         int64_t items_to_fetch = 0;
 397
 398         if (td->o.read_iolog_chunked) {
 399                 items_to_fetch = iolog_items_to_fetch(td);
 400                 if (!items_to_fetch)
 401                         return true;
 402         }
 403
 404         /*
 405          * Read in the read iolog and store it, reuse the infrastructure
 406          * for doing verifications.
 407          */
 408         str = malloc(4096);
 409         rfname = fname = malloc(256+16);
 410         act = malloc(256+16);
 411
 412         reads = writes = waits = 0;
 413         while ((p = fgets(str, 4096, td->io_log_rfile)) != NULL) {
 414                 struct io_piece *ipo;
 415                 int r;
 416
 417                 r = sscanf(p, "%256s %256s %llu %u", rfname, act, &offset,
 418                                                                         &bytes);
 419
 420                 if (td->o.replay_redirect)
 421                         fname = td->o.replay_redirect;
 422
 423                 if (r == 4) {
 424                         /*
 425                          * Check action first
 426                          */
 427                         if (!strcmp(act, "wait"))
 428                                 rw = DDIR_WAIT;
 429                         else if (!strcmp(act, "read"))
 430                                 rw = DDIR_READ;
 431                         else if (!strcmp(act, "write"))
 432                                 rw = DDIR_WRITE;
 433                         else if (!strcmp(act, "sync"))
 434                                 rw = DDIR_SYNC;
 435                         else if (!strcmp(act, "datasync"))
 436                                 rw = DDIR_DATASYNC;
 437                         else if (!strcmp(act, "trim"))
 438                                 rw = DDIR_TRIM;
 439                         else {
 440                                 log_err("fio: bad iolog file action: %s\n",
 441                                                                         act);
 442                                 continue;
 443                         }
 444                         fileno = get_fileno(td, fname);
 445                 } else if (r == 2) {
 446                         rw = DDIR_INVAL;
 447                         if (!strcmp(act, "add")) {
 448                                 if (td->o.replay_redirect &&
 449                                     get_fileno(td, fname) != -1) {
 450                                         dprint(FD_FILE, "iolog: ignoring"
 451                                                 " re-add of file %s\n", fname);
 452                                 } else {
 453                                         fileno = add_file(td, fname, td->subjob_number, 1);
 454                                         file_action = FIO_LOG_ADD_FILE;
 455                                 }
 456                                 continue;
 457                         } else if (!strcmp(act, "open")) {
 458                                 fileno = get_fileno(td, fname);
 459                                 file_action = FIO_LOG_OPEN_FILE;
 460                         } else if (!strcmp(act, "close")) {
 461                                 fileno = get_fileno(td, fname);
 462                                 file_action = FIO_LOG_CLOSE_FILE;
 463                         } else {
 464                                 log_err("fio: bad iolog file action: %s\n",
 465                                                                         act);
 466                                 continue;
 467                         }
 468                 } else {
 469                         log_err("bad iolog2: %s\n", p);
 470                         continue;
 471                 }
 472
 473                 if (rw == DDIR_READ)
 474                         reads++;
 475                 else if (rw == DDIR_WRITE) {
 476                         /*
 477                          * Don't add a write for ro mode
 478                          */
 479                         if (read_only)
 480                                 continue;
 481                         writes++;
 482                 } else if (rw == DDIR_WAIT) {
 483                         if (td->o.no_stall)
 484                                 continue;
 485                         waits++;
 486                 } else if (rw == DDIR_INVAL) {
 487                 } else if (!ddir_sync(rw)) {
 488                         log_err("bad ddir: %d\n", rw);
 489                         continue;
 490                 }
 491
 492                 /*
 493                  * Make note of file
 494                  */
 495                 ipo = calloc(1, sizeof(*ipo));
 496                 init_ipo(ipo);
 497                 ipo->ddir = rw;
 498                 if (rw == DDIR_WAIT) {
 499                         ipo->delay = offset;
 500                 } else {
 501                         if (td->o.replay_scale)
 502                                 ipo->offset = offset / td->o.replay_scale;
 503                         else
 504                                 ipo->offset = offset;
 505                         ipo_bytes_align(td->o.replay_align, ipo);
 506
 507                         ipo->len = bytes;
 508                         if (rw != DDIR_INVAL && bytes > td->o.max_bs[rw]) {
 509                                 realloc = true;
 510                                 td->o.max_bs[rw] = bytes;
 511                         }
 512                         ipo->fileno = fileno;
 513                         ipo->file_action = file_action;
 514                         td->o.size += bytes;
 515                 }
 516
 517                 queue_io_piece(td, ipo);
 518
 519                 if (td->o.read_iolog_chunked) {
 520                         td->io_log_current++;
 521                         items_to_fetch--;
 522                         if (items_to_fetch == 0)
 523                                 break;
 524                 }
 525         }
 526
 527         free(str);
 528         free(act);
 529         free(rfname);
 530
 531         if (td->o.read_iolog_chunked) {
 532                 td->io_log_highmark = td->io_log_current;
 533                 td->io_log_checkmark = (td->io_log_highmark + 1) / 2;
 534                 fio_gettime(&td->io_log_highmark_time, NULL);
 535         }
 536
 537         if (writes && read_only) {
 538                 log_err("fio: <%s> skips replay of %d writes due to"
 539                         " read-only\n", td->o.name, writes);
 540                 writes = 0;
 541         }
 542
 543         if (td->o.read_iolog_chunked) {
 544                 if (td->io_log_current == 0) {
 545                         return false;
 546                 }
 547                 td->o.td_ddir = TD_DDIR_RW;
 548                 if (realloc && td->orig_buffer)
 549                 {
 550                         io_u_quiesce(td);
 551                         free_io_mem(td);
 552                         init_io_u_buffers(td);
 553                 }
 554                 return true;
 555         }
 556
 557         if (!reads && !writes && !waits)
 558                 return false;
 559         else if (reads && !writes)
 560                 td->o.td_ddir = TD_DDIR_READ;
 561         else if (!reads && writes)
 562                 td->o.td_ddir = TD_DDIR_WRITE;
 563         else
 564                 td->o.td_ddir = TD_DDIR_RW;
 565
 566         return true;
 567 }
 568
 569 static bool is_socket(const char *path)
 570 {
 571         struct stat buf;
 572         int r;
 573
 574         r = stat(path, &buf);
 575         if (r == -1)
 576                 return false;
 577
 578         return S_ISSOCK(buf.st_mode);
 579 }
 580
 581 static int open_socket(const char *path)
 582 {
 583         struct sockaddr_un addr;
 584         int ret, fd;
 585
 586         fd = socket(AF_UNIX, SOCK_STREAM, 0);
 587         if (fd < 0)
 588                 return fd;
 589
 590         addr.sun_family = AF_UNIX;
 591         if (snprintf(addr.sun_path, sizeof(addr.sun_path), "%s", path) >=
 592             sizeof(addr.sun_path)) {
 593                 log_err("%s: path name %s is too long for a Unix socket\n",
 594                         __func__, path);
 595         }
 596
 597         ret = connect(fd, (const struct sockaddr *)&addr, strlen(path) + sizeof(addr.sun_family));
 598         if (!ret)
 599                 return fd;
 600
 601         close(fd);
 602         return -1;
 603 }
 604
 605 /*
 606  * open iolog, check version, and call appropriate parser
 607  */
 608 static bool init_iolog_read(struct thread_data *td)
 609 {
 610         char buffer[256], *p, *fname;
 611         FILE *f = NULL;
 612
 613         fname = get_name_by_idx(td->o.read_iolog_file, td->subjob_number);
 614         dprint(FD_IO, "iolog: name=%s\n", fname);
 615
 616         if (is_socket(fname)) {
 617                 int fd;
 618
 619                 fd = open_socket(fname);
 620                 if (fd >= 0)
 621                         f = fdopen(fd, "r");
 622         } else if (!strcmp(fname, "-")) {
 623                 f = stdin;
 624         } else
 625                 f = fopen(fname, "r");
 626
 627         free(fname);
 628
 629         if (!f) {
 630                 perror("fopen read iolog");
 631                 return false;
 632         }
 633
 634         p = fgets(buffer, sizeof(buffer), f);
 635         if (!p) {
 636                 td_verror(td, errno, "iolog read");
 637                 log_err("fio: unable to read iolog\n");
 638                 fclose(f);
 639                 return false;
 640         }
 641
 642         /*
 643          * version 2 of the iolog stores a specific string as the
 644          * first line, check for that
 645          */
 646         if (!strncmp(iolog_ver2, buffer, strlen(iolog_ver2))) {
 647                 free_release_files(td);
 648                 td->io_log_rfile = f;
 649                 return read_iolog2(td);
 650         }
 651
 652         log_err("fio: iolog version 1 is no longer supported\n");
 653         fclose(f);
 654         return false;
 655 }
 656
 657 /*
 658  * Set up a log for storing io patterns.
 659  */
 660 static bool init_iolog_write(struct thread_data *td)
 661 {
 662         struct fio_file *ff;
 663         FILE *f;
 664         unsigned int i;
 665
 666         f = fopen(td->o.write_iolog_file, "a");
 667         if (!f) {
 668                 perror("fopen write iolog");
 669                 return false;
 670         }
 671
 672         /*
 673          * That's it for writing, setup a log buffer and we're done.
 674           */
 675         td->iolog_f = f;
 676         td->iolog_buf = malloc(8192);
 677         setvbuf(f, td->iolog_buf, _IOFBF, 8192);
 678
 679         /*
 680          * write our version line
 681          */
 682         if (fprintf(f, "%s\n", iolog_ver2) < 0) {
 683                 perror("iolog init\n");
 684                 return false;
 685         }
 686
 687         /*
 688          * add all known files
 689          */
 690         for_each_file(td, ff, i)
 691                 log_file(td, ff, FIO_LOG_ADD_FILE);
 692
 693         return true;
 694 }
 695
 696 bool init_iolog(struct thread_data *td)
 697 {
 698         bool ret;
 699
 700         if (td->o.read_iolog_file) {
 701                 int need_swap;
 702
 703                 /*
 704                  * Check if it's a blktrace file and load that if possible.
 705                  * Otherwise assume it's a normal log file and load that.
 706                  */
 707                 if (is_blktrace(td->o.read_iolog_file, &need_swap))
 708                         ret = load_blktrace(td, td->o.read_iolog_file, need_swap);
 709                 else
 710                         ret = init_iolog_read(td);
 711         } else if (td->o.write_iolog_file)
 712                 ret = init_iolog_write(td);
 713         else
 714                 ret = true;
 715
 716         if (!ret)
 717                 td_verror(td, EINVAL, "failed initializing iolog");
 718
 719         return ret;
 720 }
 721
 722 void setup_log(struct io_log **log, struct log_params *p,
 723                const char *filename)
 724 {
 725         struct io_log *l;
 726         int i;
 727         struct io_u_plat_entry *entry;
 728         struct flist_head *list;
 729
 730         l = scalloc(1, sizeof(*l));
 731         INIT_FLIST_HEAD(&l->io_logs);
 732         l->log_type = p->log_type;
 733         l->log_offset = p->log_offset;
 734         l->log_gz = p->log_gz;
 735         l->log_gz_store = p->log_gz_store;
 736         l->avg_msec = p->avg_msec;
 737         l->hist_msec = p->hist_msec;
 738         l->hist_coarseness = p->hist_coarseness;
 739         l->filename = strdup(filename);
 740         l->td = p->td;
 741
 742         /* Initialize histogram lists for each r/w direction,
 743          * with initial io_u_plat of all zeros:
 744          */
 745         for (i = 0; i < DDIR_RWDIR_CNT; i++) {
 746                 list = &l->hist_window[i].list;
 747                 INIT_FLIST_HEAD(list);
 748                 entry = calloc(1, sizeof(struct io_u_plat_entry));
 749                 flist_add(&entry->list, list);
 750         }
 751
 752         if (l->td && l->td->o.io_submit_mode != IO_MODE_OFFLOAD) {
 753                 struct io_logs *__p;
 754
 755                 __p = calloc(1, sizeof(*l->pending));
 756                 __p->max_samples = DEF_LOG_ENTRIES;
 757                 __p->log = calloc(__p->max_samples, log_entry_sz(l));
 758                 l->pending = __p;
 759         }
 760
 761         if (l->log_offset)
 762                 l->log_ddir_mask = LOG_OFFSET_SAMPLE_BIT;
 763
 764         INIT_FLIST_HEAD(&l->chunk_list);
 765
 766         if (l->log_gz && !p->td)
 767                 l->log_gz = 0;
 768         else if (l->log_gz || l->log_gz_store) {
 769                 mutex_init_pshared(&l->chunk_lock);
 770                 mutex_init_pshared(&l->deferred_free_lock);
 771                 p->td->flags |= TD_F_COMPRESS_LOG;
 772         }
 773
 774         *log = l;
 775 }
 776
 777 #ifdef CONFIG_SETVBUF
 778 static void *set_file_buffer(FILE *f)
 779 {
 780         size_t size = 1048576;
 781         void *buf;
 782
 783         buf = malloc(size);
 784         setvbuf(f, buf, _IOFBF, size);
 785         return buf;
 786 }
 787
 788 static void clear_file_buffer(void *buf)
 789 {
 790         free(buf);
 791 }
 792 #else
 793 static void *set_file_buffer(FILE *f)
 794 {
 795         return NULL;
 796 }
 797
 798 static void clear_file_buffer(void *buf)
 799 {
 800 }
 801 #endif
 802
 803 void free_log(struct io_log *log)
 804 {
 805         while (!flist_empty(&log->io_logs)) {
 806                 struct io_logs *cur_log;
 807
 808                 cur_log = flist_first_entry(&log->io_logs, struct io_logs, list);
 809                 flist_del_init(&cur_log->list);
 810                 free(cur_log->log);
 811                 sfree(cur_log);
 812         }
 813
 814         if (log->pending) {
 815                 free(log->pending->log);
 816                 free(log->pending);
 817                 log->pending = NULL;
 818         }
 819
 820         free(log->pending);
 821         free(log->filename);
 822         sfree(log);
 823 }
 824
 825 uint64_t hist_sum(int j, int stride, uint64_t *io_u_plat,
 826                 uint64_t *io_u_plat_last)
 827 {
 828         uint64_t sum;
 829         int k;
 830
 831         if (io_u_plat_last) {
 832                 for (k = sum = 0; k < stride; k++)
 833                         sum += io_u_plat[j + k] - io_u_plat_last[j + k];
 834         } else {
 835                 for (k = sum = 0; k < stride; k++)
 836                         sum += io_u_plat[j + k];
 837         }
 838
 839         return sum;
 840 }
 841
 842 static void flush_hist_samples(FILE *f, int hist_coarseness, void *samples,
 843                                uint64_t sample_size)
 844 {
 845         struct io_sample *s;
 846         int log_offset;
 847         uint64_t i, j, nr_samples;
 848         struct io_u_plat_entry *entry, *entry_before;
 849         uint64_t *io_u_plat;
 850         uint64_t *io_u_plat_before;
 851
 852         int stride = 1 << hist_coarseness;
 853
 854         if (!sample_size)
 855                 return;
 856
 857         s = __get_sample(samples, 0, 0);
 858         log_offset = (s->__ddir & LOG_OFFSET_SAMPLE_BIT) != 0;
 859
 860         nr_samples = sample_size / __log_entry_sz(log_offset);
 861
 862         for (i = 0; i < nr_samples; i++) {
 863                 s = __get_sample(samples, log_offset, i);
 864
 865                 entry = s->data.plat_entry;
 866                 io_u_plat = entry->io_u_plat;
 867
 868                 entry_before = flist_first_entry(&entry->list, struct io_u_plat_entry, list);
 869                 io_u_plat_before = entry_before->io_u_plat;
 870
 871                 fprintf(f, "%lu, %u, %llu, ", (unsigned long) s->time,
 872                                                 io_sample_ddir(s), (unsigned long long) s->bs);
 873                 for (j = 0; j < FIO_IO_U_PLAT_NR - stride; j += stride) {
 874                         fprintf(f, "%llu, ", (unsigned long long)
 875                                 hist_sum(j, stride, io_u_plat, io_u_plat_before));
 876                 }
 877                 fprintf(f, "%llu\n", (unsigned long long)
 878                         hist_sum(FIO_IO_U_PLAT_NR - stride, stride, io_u_plat,
 879                                         io_u_plat_before));
 880
 881                 flist_del(&entry_before->list);
 882                 free(entry_before);
 883         }
 884 }
 885
 886 void flush_samples(FILE *f, void *samples, uint64_t sample_size)
 887 {
 888         struct io_sample *s;
 889         int log_offset;
 890         uint64_t i, nr_samples;
 891
 892         if (!sample_size)
 893                 return;
 894
 895         s = __get_sample(samples, 0, 0);
 896         log_offset = (s->__ddir & LOG_OFFSET_SAMPLE_BIT) != 0;
 897
 898         nr_samples = sample_size / __log_entry_sz(log_offset);
 899
 900         for (i = 0; i < nr_samples; i++) {
 901                 s = __get_sample(samples, log_offset, i);
 902
 903                 if (!log_offset) {
 904                         fprintf(f, "%lu, %" PRId64 ", %u, %llu, %u\n",
 905                                         (unsigned long) s->time,
 906                                         s->data.val,
 907                                         io_sample_ddir(s), (unsigned long long) s->bs, s->priority_bit);
 908                 } else {
 909                         struct io_sample_offset *so = (void *) s;
 910
 911                         fprintf(f, "%lu, %" PRId64 ", %u, %llu, %llu, %u\n",
 912                                         (unsigned long) s->time,
 913                                         s->data.val,
 914                                         io_sample_ddir(s), (unsigned long long) s->bs,
 915                                         (unsigned long long) so->offset, s->priority_bit);
 916                 }
 917         }
 918 }
 919
 920 #ifdef CONFIG_ZLIB
 921
 922 struct iolog_flush_data {
 923         struct workqueue_work work;
 924         struct io_log *log;
 925         void *samples;
 926         uint32_t nr_samples;
 927         bool free;
 928 };
 929
 930 #define GZ_CHUNK        131072
 931
 932 static struct iolog_compress *get_new_chunk(unsigned int seq)
 933 {
 934         struct iolog_compress *c;
 935
 936         c = malloc(sizeof(*c));
 937         INIT_FLIST_HEAD(&c->list);
 938         c->buf = malloc(GZ_CHUNK);
 939         c->len = 0;
 940         c->seq = seq;
 941         return c;
 942 }
 943
 944 static void free_chunk(struct iolog_compress *ic)
 945 {
 946         free(ic->buf);
 947         free(ic);
 948 }
 949
 950 static int z_stream_init(z_stream *stream, int gz_hdr)
 951 {
 952         int wbits = 15;
 953
 954         memset(stream, 0, sizeof(*stream));
 955         stream->zalloc = Z_NULL;
 956         stream->zfree = Z_NULL;
 957         stream->opaque = Z_NULL;
 958         stream->next_in = Z_NULL;
 959
 960         /*
 961          * zlib magic - add 32 for auto-detection of gz header or not,
 962          * if we decide to store files in a gzip friendly format.
 963          */
 964         if (gz_hdr)
 965                 wbits += 32;
 966
 967         if (inflateInit2(stream, wbits) != Z_OK)
 968                 return 1;
 969
 970         return 0;
 971 }
 972
 973 struct inflate_chunk_iter {
 974         unsigned int seq;
 975         int err;
 976         void *buf;
 977         size_t buf_size;
 978         size_t buf_used;
 979         size_t chunk_sz;
 980 };
 981
 982 static void finish_chunk(z_stream *stream, FILE *f,
 983                          struct inflate_chunk_iter *iter)
 984 {
 985         int ret;
 986
 987         ret = inflateEnd(stream);
 988         if (ret != Z_OK)
 989                 log_err("fio: failed to end log inflation seq %d (%d)\n",
 990                                 iter->seq, ret);
 991
 992         flush_samples(f, iter->buf, iter->buf_used);
 993         free(iter->buf);
 994         iter->buf = NULL;
 995         iter->buf_size = iter->buf_used = 0;
 996 }
 997
 998 /*
 999  * Iterative chunk inflation. Handles cases where we cross into a new
1000  * sequence, doing flush finish of previous chunk if needed.
1001  */
1002 static size_t inflate_chunk(struct iolog_compress *ic, int gz_hdr, FILE *f,
1003                             z_stream *stream, struct inflate_chunk_iter *iter)
1004 {
1005         size_t ret;
1006
1007         dprint(FD_COMPRESS, "inflate chunk size=%lu, seq=%u\n",
1008                                 (unsigned long) ic->len, ic->seq);
1009
1010         if (ic->seq != iter->seq) {
1011                 if (iter->seq)
1012                         finish_chunk(stream, f, iter);
1013
1014                 z_stream_init(stream, gz_hdr);
1015                 iter->seq = ic->seq;
1016         }
1017
1018         stream->avail_in = ic->len;
1019         stream->next_in = ic->buf;
1020
1021         if (!iter->buf_size) {
1022                 iter->buf_size = iter->chunk_sz;
1023                 iter->buf = malloc(iter->buf_size);
1024         }
1025
1026         while (stream->avail_in) {
1027                 size_t this_out = iter->buf_size - iter->buf_used;
1028                 int err;
1029
1030                 stream->avail_out = this_out;
1031                 stream->next_out = iter->buf + iter->buf_used;
1032
1033                 err = inflate(stream, Z_NO_FLUSH);
1034                 if (err < 0) {
1035                         log_err("fio: failed inflating log: %d\n", err);
1036                         iter->err = err;
1037                         break;
1038                 }
1039
1040                 iter->buf_used += this_out - stream->avail_out;
1041
1042                 if (!stream->avail_out) {
1043                         iter->buf_size += iter->chunk_sz;
1044                         iter->buf = realloc(iter->buf, iter->buf_size);
1045                         continue;
1046                 }
1047
1048                 if (err == Z_STREAM_END)
1049                         break;
1050         }
1051
1052         ret = (void *) stream->next_in - ic->buf;
1053
1054         dprint(FD_COMPRESS, "inflated to size=%lu\n", (unsigned long) iter->buf_size);
1055
1056         return ret;
1057 }
1058
1059 /*
1060  * Inflate stored compressed chunks, or write them directly to the log
1061  * file if so instructed.
1062  */
1063 static int inflate_gz_chunks(struct io_log *log, FILE *f)
1064 {
1065         struct inflate_chunk_iter iter = { .chunk_sz = log->log_gz, };
1066         z_stream stream;
1067
1068         while (!flist_empty(&log->chunk_list)) {
1069                 struct iolog_compress *ic;
1070
1071                 ic = flist_first_entry(&log->chunk_list, struct iolog_compress, list);
1072                 flist_del(&ic->list);
1073
1074                 if (log->log_gz_store) {
1075                         size_t ret;
1076
1077                         dprint(FD_COMPRESS, "log write chunk size=%lu, "
1078                                 "seq=%u\n", (unsigned long) ic->len, ic->seq);
1079
1080                         ret = fwrite(ic->buf, ic->len, 1, f);
1081                         if (ret != 1 || ferror(f)) {
1082                                 iter.err = errno;
1083                                 log_err("fio: error writing compressed log\n");
1084                         }
1085                 } else
1086                         inflate_chunk(ic, log->log_gz_store, f, &stream, &iter);
1087
1088                 free_chunk(ic);
1089         }
1090
1091         if (iter.seq) {
1092                 finish_chunk(&stream, f, &iter);
1093                 free(iter.buf);
1094         }
1095
1096         return iter.err;
1097 }
1098
1099 /*
1100  * Open compressed log file and decompress the stored chunks and
1101  * write them to stdout. The chunks are stored sequentially in the
1102  * file, so we iterate over them and do them one-by-one.
1103  */
1104 int iolog_file_inflate(const char *file)
1105 {
1106         struct inflate_chunk_iter iter = { .chunk_sz = 64 * 1024 * 1024, };
1107         struct iolog_compress ic;
1108         z_stream stream;
1109         struct stat sb;
1110         size_t ret;
1111         size_t total;
1112         void *buf;
1113         FILE *f;
1114
1115         f = fopen(file, "r");
1116         if (!f) {
1117                 perror("fopen");
1118                 return 1;
1119         }
1120
1121         if (stat(file, &sb) < 0) {
1122                 fclose(f);
1123                 perror("stat");
1124                 return 1;
1125         }
1126
1127         ic.buf = buf = malloc(sb.st_size);
1128         ic.len = sb.st_size;
1129         ic.seq = 1;
1130
1131         ret = fread(ic.buf, ic.len, 1, f);
1132         if (ret == 0 && ferror(f)) {
1133                 perror("fread");
1134                 fclose(f);
1135                 free(buf);
1136                 return 1;
1137         } else if (ferror(f) || (!feof(f) && ret != 1)) {
1138                 log_err("fio: short read on reading log\n");
1139                 fclose(f);
1140                 free(buf);
1141                 return 1;
1142         }
1143
1144         fclose(f);
1145
1146         /*
1147          * Each chunk will return Z_STREAM_END. We don't know how many
1148          * chunks are in the file, so we just keep looping and incrementing
1149          * the sequence number until we have consumed the whole compressed
1150          * file.
1151          */
1152         total = ic.len;
1153         do {
1154                 size_t iret;
1155
1156                 iret = inflate_chunk(&ic,  1, stdout, &stream, &iter);
1157                 total -= iret;
1158                 if (!total)
1159                         break;
1160                 if (iter.err)
1161                         break;
1162
1163                 ic.seq++;
1164                 ic.len -= iret;
1165                 ic.buf += iret;
1166         } while (1);
1167
1168         if (iter.seq) {
1169                 finish_chunk(&stream, stdout, &iter);
1170                 free(iter.buf);
1171         }
1172
1173         free(buf);
1174         return iter.err;
1175 }
1176
1177 #else
1178
1179 static int inflate_gz_chunks(struct io_log *log, FILE *f)
1180 {
1181         return 0;
1182 }
1183
1184 int iolog_file_inflate(const char *file)
1185 {
1186         log_err("fio: log inflation not possible without zlib\n");
1187         return 1;
1188 }
1189
1190 #endif
1191
1192 void flush_log(struct io_log *log, bool do_append)
1193 {
1194         void *buf;
1195         FILE *f;
1196
1197         if (!do_append)
1198                 f = fopen(log->filename, "w");
1199         else
1200                 f = fopen(log->filename, "a");
1201         if (!f) {
1202                 perror("fopen log");
1203                 return;
1204         }
1205
1206         buf = set_file_buffer(f);
1207
1208         inflate_gz_chunks(log, f);
1209
1210         while (!flist_empty(&log->io_logs)) {
1211                 struct io_logs *cur_log;
1212
1213                 cur_log = flist_first_entry(&log->io_logs, struct io_logs, list);
1214                 flist_del_init(&cur_log->list);
1215
1216                 if (log->td && log == log->td->clat_hist_log)
1217                         flush_hist_samples(f, log->hist_coarseness, cur_log->log,
1218                                            log_sample_sz(log, cur_log));
1219                 else
1220                         flush_samples(f, cur_log->log, log_sample_sz(log, cur_log));
1221
1222                 sfree(cur_log);
1223         }
1224
1225         fclose(f);
1226         clear_file_buffer(buf);
1227 }
1228
1229 static int finish_log(struct thread_data *td, struct io_log *log, int trylock)
1230 {
1231         if (td->flags & TD_F_COMPRESS_LOG)
1232                 iolog_flush(log);
1233
1234         if (trylock) {
1235                 if (fio_trylock_file(log->filename))
1236                         return 1;
1237         } else
1238                 fio_lock_file(log->filename);
1239
1240         if (td->client_type == FIO_CLIENT_TYPE_GUI || is_backend)
1241                 fio_send_iolog(td, log, log->filename);
1242         else
1243                 flush_log(log, !td->o.per_job_logs);
1244
1245         fio_unlock_file(log->filename);
1246         free_log(log);
1247         return 0;
1248 }
1249
1250 size_t log_chunk_sizes(struct io_log *log)
1251 {
1252         struct flist_head *entry;
1253         size_t ret;
1254
1255         if (flist_empty(&log->chunk_list))
1256                 return 0;
1257
1258         ret = 0;
1259         pthread_mutex_lock(&log->chunk_lock);
1260         flist_for_each(entry, &log->chunk_list) {
1261                 struct iolog_compress *c;
1262
1263                 c = flist_entry(entry, struct iolog_compress, list);
1264                 ret += c->len;
1265         }
1266         pthread_mutex_unlock(&log->chunk_lock);
1267         return ret;
1268 }
1269
1270 #ifdef CONFIG_ZLIB
1271
1272 static void iolog_put_deferred(struct io_log *log, void *ptr)
1273 {
1274         if (!ptr)
1275                 return;
1276
1277         pthread_mutex_lock(&log->deferred_free_lock);
1278         if (log->deferred < IOLOG_MAX_DEFER) {
1279                 log->deferred_items[log->deferred] = ptr;
1280                 log->deferred++;
1281         } else if (!fio_did_warn(FIO_WARN_IOLOG_DROP))
1282                 log_err("fio: had to drop log entry free\n");
1283         pthread_mutex_unlock(&log->deferred_free_lock);
1284 }
1285
1286 static void iolog_free_deferred(struct io_log *log)
1287 {
1288         int i;
1289
1290         if (!log->deferred)
1291                 return;
1292
1293         pthread_mutex_lock(&log->deferred_free_lock);
1294
1295         for (i = 0; i < log->deferred; i++) {
1296                 free(log->deferred_items[i]);
1297                 log->deferred_items[i] = NULL;
1298         }
1299
1300         log->deferred = 0;
1301         pthread_mutex_unlock(&log->deferred_free_lock);
1302 }
1303
1304 static int gz_work(struct iolog_flush_data *data)
1305 {
1306         struct iolog_compress *c = NULL;
1307         struct flist_head list;
1308         unsigned int seq;
1309         z_stream stream;
1310         size_t total = 0;
1311         int ret;
1312
1313         INIT_FLIST_HEAD(&list);
1314
1315         memset(&stream, 0, sizeof(stream));
1316         stream.zalloc = Z_NULL;
1317         stream.zfree = Z_NULL;
1318         stream.opaque = Z_NULL;
1319
1320         ret = deflateInit(&stream, Z_DEFAULT_COMPRESSION);
1321         if (ret != Z_OK) {
1322                 log_err("fio: failed to init gz stream\n");
1323                 goto err;
1324         }
1325
1326         seq = ++data->log->chunk_seq;
1327
1328         stream.next_in = (void *) data->samples;
1329         stream.avail_in = data->nr_samples * log_entry_sz(data->log);
1330
1331         dprint(FD_COMPRESS, "deflate input size=%lu, seq=%u, log=%s\n",
1332                                 (unsigned long) stream.avail_in, seq,
1333                                 data->log->filename);
1334         do {
1335                 if (c)
1336                         dprint(FD_COMPRESS, "seq=%d, chunk=%lu\n", seq,
1337                                 (unsigned long) c->len);
1338                 c = get_new_chunk(seq);
1339                 stream.avail_out = GZ_CHUNK;
1340                 stream.next_out = c->buf;
1341                 ret = deflate(&stream, Z_NO_FLUSH);
1342                 if (ret < 0) {
1343                         log_err("fio: deflate log (%d)\n", ret);
1344                         free_chunk(c);
1345                         goto err;
1346                 }
1347
1348                 c->len = GZ_CHUNK - stream.avail_out;
1349                 flist_add_tail(&c->list, &list);
1350                 total += c->len;
1351         } while (stream.avail_in);
1352
1353         stream.next_out = c->buf + c->len;
1354         stream.avail_out = GZ_CHUNK - c->len;
1355
1356         ret = deflate(&stream, Z_FINISH);
1357         if (ret < 0) {
1358                 /*
1359                  * Z_BUF_ERROR is special, it just means we need more
1360                  * output space. We'll handle that below. Treat any other
1361                  * error as fatal.
1362                  */
1363                 if (ret != Z_BUF_ERROR) {
1364                         log_err("fio: deflate log (%d)\n", ret);
1365                         flist_del(&c->list);
1366                         free_chunk(c);
1367                         goto err;
1368                 }
1369         }
1370
1371         total -= c->len;
1372         c->len = GZ_CHUNK - stream.avail_out;
1373         total += c->len;
1374         dprint(FD_COMPRESS, "seq=%d, chunk=%lu\n", seq, (unsigned long) c->len);
1375
1376         if (ret != Z_STREAM_END) {
1377                 do {
1378                         c = get_new_chunk(seq);
1379                         stream.avail_out = GZ_CHUNK;
1380                         stream.next_out = c->buf;
1381                         ret = deflate(&stream, Z_FINISH);
1382                         c->len = GZ_CHUNK - stream.avail_out;
1383                         total += c->len;
1384                         flist_add_tail(&c->list, &list);
1385                         dprint(FD_COMPRESS, "seq=%d, chunk=%lu\n", seq,
1386                                 (unsigned long) c->len);
1387                 } while (ret != Z_STREAM_END);
1388         }
1389
1390         dprint(FD_COMPRESS, "deflated to size=%lu\n", (unsigned long) total);
1391
1392         ret = deflateEnd(&stream);
1393         if (ret != Z_OK)
1394                 log_err("fio: deflateEnd %d\n", ret);
1395
1396         iolog_put_deferred(data->log, data->samples);
1397
1398         if (!flist_empty(&list)) {
1399                 pthread_mutex_lock(&data->log->chunk_lock);
1400                 flist_splice_tail(&list, &data->log->chunk_list);
1401                 pthread_mutex_unlock(&data->log->chunk_lock);
1402         }
1403
1404         ret = 0;
1405 done:
1406         if (data->free)
1407                 sfree(data);
1408         return ret;
1409 err:
1410         while (!flist_empty(&list)) {
1411                 c = flist_first_entry(list.next, struct iolog_compress, list);
1412                 flist_del(&c->list);
1413                 free_chunk(c);
1414         }
1415         ret = 1;
1416         goto done;
1417 }
1418
1419 /*
1420  * Invoked from our compress helper thread, when logging would have exceeded
1421  * the specified memory limitation. Compresses the previously stored
1422  * entries.
1423  */
1424 static int gz_work_async(struct submit_worker *sw, struct workqueue_work *work)
1425 {
1426         return gz_work(container_of(work, struct iolog_flush_data, work));
1427 }
1428
1429 static int gz_init_worker(struct submit_worker *sw)
1430 {
1431         struct thread_data *td = sw->wq->td;
1432
1433         if (!fio_option_is_set(&td->o, log_gz_cpumask))
1434                 return 0;
1435
1436         if (fio_setaffinity(gettid(), td->o.log_gz_cpumask) == -1) {
1437                 log_err("gz: failed to set CPU affinity\n");
1438                 return 1;
1439         }
1440
1441         return 0;
1442 }
1443
1444 static struct workqueue_ops log_compress_wq_ops = {
1445         .fn             = gz_work_async,
1446         .init_worker_fn = gz_init_worker,
1447         .nice           = 1,
1448 };
1449
1450 int iolog_compress_init(struct thread_data *td, struct sk_out *sk_out)
1451 {
1452         if (!(td->flags & TD_F_COMPRESS_LOG))
1453                 return 0;
1454
1455         workqueue_init(td, &td->log_compress_wq, &log_compress_wq_ops, 1, sk_out);
1456         return 0;
1457 }
1458
1459 void iolog_compress_exit(struct thread_data *td)
1460 {
1461         if (!(td->flags & TD_F_COMPRESS_LOG))
1462                 return;
1463
1464         workqueue_exit(&td->log_compress_wq);
1465 }
1466
1467 /*
1468  * Queue work item to compress the existing log entries. We reset the
1469  * current log to a small size, and reference the existing log in the
1470  * data that we queue for compression. Once compression has been done,
1471  * this old log is freed. If called with finish == true, will not return
1472  * until the log compression has completed, and will flush all previous
1473  * logs too
1474  */
1475 static int iolog_flush(struct io_log *log)
1476 {
1477         struct iolog_flush_data *data;
1478
1479         data = malloc(sizeof(*data));
1480         if (!data)
1481                 return 1;
1482
1483         data->log = log;
1484         data->free = false;
1485
1486         while (!flist_empty(&log->io_logs)) {
1487                 struct io_logs *cur_log;
1488
1489                 cur_log = flist_first_entry(&log->io_logs, struct io_logs, list);
1490                 flist_del_init(&cur_log->list);
1491
1492                 data->samples = cur_log->log;
1493                 data->nr_samples = cur_log->nr_samples;
1494
1495                 sfree(cur_log);
1496
1497                 gz_work(data);
1498         }
1499
1500         free(data);
1501         return 0;
1502 }
1503
1504 int iolog_cur_flush(struct io_log *log, struct io_logs *cur_log)
1505 {
1506         struct iolog_flush_data *data;
1507
1508         data = smalloc(sizeof(*data));
1509         if (!data)
1510                 return 1;
1511
1512         data->log = log;
1513
1514         data->samples = cur_log->log;
1515         data->nr_samples = cur_log->nr_samples;
1516         data->free = true;
1517
1518         cur_log->nr_samples = cur_log->max_samples = 0;
1519         cur_log->log = NULL;
1520
1521         workqueue_enqueue(&log->td->log_compress_wq, &data->work);
1522
1523         iolog_free_deferred(log);
1524
1525         return 0;
1526 }
1527 #else
1528
1529 static int iolog_flush(struct io_log *log)
1530 {
1531         return 1;
1532 }
1533
1534 int iolog_cur_flush(struct io_log *log, struct io_logs *cur_log)
1535 {
1536         return 1;
1537 }
1538
1539 int iolog_compress_init(struct thread_data *td, struct sk_out *sk_out)
1540 {
1541         return 0;
1542 }
1543
1544 void iolog_compress_exit(struct thread_data *td)
1545 {
1546 }
1547
1548 #endif
1549
1550 struct io_logs *iolog_cur_log(struct io_log *log)
1551 {
1552         if (flist_empty(&log->io_logs))
1553                 return NULL;
1554
1555         return flist_last_entry(&log->io_logs, struct io_logs, list);
1556 }
1557
1558 uint64_t iolog_nr_samples(struct io_log *iolog)
1559 {
1560         struct flist_head *entry;
1561         uint64_t ret = 0;
1562
1563         flist_for_each(entry, &iolog->io_logs) {
1564                 struct io_logs *cur_log;
1565
1566                 cur_log = flist_entry(entry, struct io_logs, list);
1567                 ret += cur_log->nr_samples;
1568         }
1569
1570         return ret;
1571 }
1572
1573 static int __write_log(struct thread_data *td, struct io_log *log, int try)
1574 {
1575         if (log)
1576                 return finish_log(td, log, try);
1577
1578         return 0;
1579 }
1580
1581 static int write_iops_log(struct thread_data *td, int try, bool unit_log)
1582 {
1583         int ret;
1584
1585         if (per_unit_log(td->iops_log) != unit_log)
1586                 return 0;
1587
1588         ret = __write_log(td, td->iops_log, try);
1589         if (!ret)
1590                 td->iops_log = NULL;
1591
1592         return ret;
1593 }
1594
1595 static int write_slat_log(struct thread_data *td, int try, bool unit_log)
1596 {
1597         int ret;
1598
1599         if (!unit_log)
1600                 return 0;
1601
1602         ret = __write_log(td, td->slat_log, try);
1603         if (!ret)
1604                 td->slat_log = NULL;
1605
1606         return ret;
1607 }
1608
1609 static int write_clat_log(struct thread_data *td, int try, bool unit_log)
1610 {
1611         int ret;
1612
1613         if (!unit_log)
1614                 return 0;
1615
1616         ret = __write_log(td, td->clat_log, try);
1617         if (!ret)
1618                 td->clat_log = NULL;
1619
1620         return ret;
1621 }
1622
1623 static int write_clat_hist_log(struct thread_data *td, int try, bool unit_log)
1624 {
1625         int ret;
1626
1627         if (!unit_log)
1628                 return 0;
1629
1630         ret = __write_log(td, td->clat_hist_log, try);
1631         if (!ret)
1632                 td->clat_hist_log = NULL;
1633
1634         return ret;
1635 }
1636
1637 static int write_lat_log(struct thread_data *td, int try, bool unit_log)
1638 {
1639         int ret;
1640
1641         if (!unit_log)
1642                 return 0;
1643
1644         ret = __write_log(td, td->lat_log, try);
1645         if (!ret)
1646                 td->lat_log = NULL;
1647
1648         return ret;
1649 }
1650
1651 static int write_bandw_log(struct thread_data *td, int try, bool unit_log)
1652 {
1653         int ret;
1654
1655         if (per_unit_log(td->bw_log) != unit_log)
1656                 return 0;
1657
1658         ret = __write_log(td, td->bw_log, try);
1659         if (!ret)
1660                 td->bw_log = NULL;
1661
1662         return ret;
1663 }
1664
1665 enum {
1666         BW_LOG_MASK     = 1,
1667         LAT_LOG_MASK    = 2,
1668         SLAT_LOG_MASK   = 4,
1669         CLAT_LOG_MASK   = 8,
1670         IOPS_LOG_MASK   = 16,
1671         CLAT_HIST_LOG_MASK = 32,
1672
1673         ALL_LOG_NR      = 6,
1674 };
1675
1676 struct log_type {
1677         unsigned int mask;
1678         int (*fn)(struct thread_data *, int, bool);
1679 };
1680
1681 static struct log_type log_types[] = {
1682         {
1683                 .mask   = BW_LOG_MASK,
1684                 .fn     = write_bandw_log,
1685         },
1686         {
1687                 .mask   = LAT_LOG_MASK,
1688                 .fn     = write_lat_log,
1689         },
1690         {
1691                 .mask   = SLAT_LOG_MASK,
1692                 .fn     = write_slat_log,
1693         },
1694         {
1695                 .mask   = CLAT_LOG_MASK,
1696                 .fn     = write_clat_log,
1697         },
1698         {
1699                 .mask   = IOPS_LOG_MASK,
1700                 .fn     = write_iops_log,
1701         },
1702         {
1703                 .mask   = CLAT_HIST_LOG_MASK,
1704                 .fn     = write_clat_hist_log,
1705         }
1706 };
1707
1708 void td_writeout_logs(struct thread_data *td, bool unit_logs)
1709 {
1710         unsigned int log_mask = 0;
1711         unsigned int log_left = ALL_LOG_NR;
1712         int old_state, i;
1713
1714         old_state = td_bump_runstate(td, TD_FINISHING);
1715
1716         finalize_logs(td, unit_logs);
1717
1718         while (log_left) {
1719                 int prev_log_left = log_left;
1720
1721                 for (i = 0; i < ALL_LOG_NR && log_left; i++) {
1722                         struct log_type *lt = &log_types[i];
1723                         int ret;
1724
1725                         if (!(log_mask & lt->mask)) {
1726                                 ret = lt->fn(td, log_left != 1, unit_logs);
1727                                 if (!ret) {
1728                                         log_left--;
1729                                         log_mask |= lt->mask;
1730                                 }
1731                         }
1732                 }
1733
1734                 if (prev_log_left == log_left)
1735                         usleep(5000);
1736         }
1737
1738         td_restore_runstate(td, old_state);
1739 }
1740
1741 void fio_writeout_logs(bool unit_logs)
1742 {
1743         struct thread_data *td;
1744         int i;
1745
1746         for_each_td(td, i)
1747                 td_writeout_logs(td, unit_logs);
1748 }