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
 623                 f = fopen(fname, "r");
 624
 625         free(fname);
 626
 627         if (!f) {
 628                 perror("fopen read iolog");
 629                 return false;
 630         }
 631
 632         p = fgets(buffer, sizeof(buffer), f);
 633         if (!p) {
 634                 td_verror(td, errno, "iolog read");
 635                 log_err("fio: unable to read iolog\n");
 636                 fclose(f);
 637                 return false;
 638         }
 639
 640         /*
 641          * version 2 of the iolog stores a specific string as the
 642          * first line, check for that
 643          */
 644         if (!strncmp(iolog_ver2, buffer, strlen(iolog_ver2))) {
 645                 free_release_files(td);
 646                 td->io_log_rfile = f;
 647                 return read_iolog2(td);
 648         }
 649
 650         log_err("fio: iolog version 1 is no longer supported\n");
 651         fclose(f);
 652         return false;
 653 }
 654
 655 /*
 656  * Set up a log for storing io patterns.
 657  */
 658 static bool init_iolog_write(struct thread_data *td)
 659 {
 660         struct fio_file *ff;
 661         FILE *f;
 662         unsigned int i;
 663
 664         f = fopen(td->o.write_iolog_file, "a");
 665         if (!f) {
 666                 perror("fopen write iolog");
 667                 return false;
 668         }
 669
 670         /*
 671          * That's it for writing, setup a log buffer and we're done.
 672           */
 673         td->iolog_f = f;
 674         td->iolog_buf = malloc(8192);
 675         setvbuf(f, td->iolog_buf, _IOFBF, 8192);
 676
 677         /*
 678          * write our version line
 679          */
 680         if (fprintf(f, "%s\n", iolog_ver2) < 0) {
 681                 perror("iolog init\n");
 682                 return false;
 683         }
 684
 685         /*
 686          * add all known files
 687          */
 688         for_each_file(td, ff, i)
 689                 log_file(td, ff, FIO_LOG_ADD_FILE);
 690
 691         return true;
 692 }
 693
 694 bool init_iolog(struct thread_data *td)
 695 {
 696         bool ret;
 697
 698         if (td->o.read_iolog_file) {
 699                 int need_swap;
 700
 701                 /*
 702                  * Check if it's a blktrace file and load that if possible.
 703                  * Otherwise assume it's a normal log file and load that.
 704                  */
 705                 if (is_blktrace(td->o.read_iolog_file, &need_swap))
 706                         ret = load_blktrace(td, td->o.read_iolog_file, need_swap);
 707                 else
 708                         ret = init_iolog_read(td);
 709         } else if (td->o.write_iolog_file)
 710                 ret = init_iolog_write(td);
 711         else
 712                 ret = true;
 713
 714         if (!ret)
 715                 td_verror(td, EINVAL, "failed initializing iolog");
 716
 717         return ret;
 718 }
 719
 720 void setup_log(struct io_log **log, struct log_params *p,
 721                const char *filename)
 722 {
 723         struct io_log *l;
 724         int i;
 725         struct io_u_plat_entry *entry;
 726         struct flist_head *list;
 727
 728         l = scalloc(1, sizeof(*l));
 729         INIT_FLIST_HEAD(&l->io_logs);
 730         l->log_type = p->log_type;
 731         l->log_offset = p->log_offset;
 732         l->log_gz = p->log_gz;
 733         l->log_gz_store = p->log_gz_store;
 734         l->avg_msec = p->avg_msec;
 735         l->hist_msec = p->hist_msec;
 736         l->hist_coarseness = p->hist_coarseness;
 737         l->filename = strdup(filename);
 738         l->td = p->td;
 739
 740         /* Initialize histogram lists for each r/w direction,
 741          * with initial io_u_plat of all zeros:
 742          */
 743         for (i = 0; i < DDIR_RWDIR_CNT; i++) {
 744                 list = &l->hist_window[i].list;
 745                 INIT_FLIST_HEAD(list);
 746                 entry = calloc(1, sizeof(struct io_u_plat_entry));
 747                 flist_add(&entry->list, list);
 748         }
 749
 750         if (l->td && l->td->o.io_submit_mode != IO_MODE_OFFLOAD) {
 751                 struct io_logs *__p;
 752
 753                 __p = calloc(1, sizeof(*l->pending));
 754                 __p->max_samples = DEF_LOG_ENTRIES;
 755                 __p->log = calloc(__p->max_samples, log_entry_sz(l));
 756                 l->pending = __p;
 757         }
 758
 759         if (l->log_offset)
 760                 l->log_ddir_mask = LOG_OFFSET_SAMPLE_BIT;
 761
 762         INIT_FLIST_HEAD(&l->chunk_list);
 763
 764         if (l->log_gz && !p->td)
 765                 l->log_gz = 0;
 766         else if (l->log_gz || l->log_gz_store) {
 767                 mutex_init_pshared(&l->chunk_lock);
 768                 mutex_init_pshared(&l->deferred_free_lock);
 769                 p->td->flags |= TD_F_COMPRESS_LOG;
 770         }
 771
 772         *log = l;
 773 }
 774
 775 #ifdef CONFIG_SETVBUF
 776 static void *set_file_buffer(FILE *f)
 777 {
 778         size_t size = 1048576;
 779         void *buf;
 780
 781         buf = malloc(size);
 782         setvbuf(f, buf, _IOFBF, size);
 783         return buf;
 784 }
 785
 786 static void clear_file_buffer(void *buf)
 787 {
 788         free(buf);
 789 }
 790 #else
 791 static void *set_file_buffer(FILE *f)
 792 {
 793         return NULL;
 794 }
 795
 796 static void clear_file_buffer(void *buf)
 797 {
 798 }
 799 #endif
 800
 801 void free_log(struct io_log *log)
 802 {
 803         while (!flist_empty(&log->io_logs)) {
 804                 struct io_logs *cur_log;
 805
 806                 cur_log = flist_first_entry(&log->io_logs, struct io_logs, list);
 807                 flist_del_init(&cur_log->list);
 808                 free(cur_log->log);
 809                 sfree(cur_log);
 810         }
 811
 812         if (log->pending) {
 813                 free(log->pending->log);
 814                 free(log->pending);
 815                 log->pending = NULL;
 816         }
 817
 818         free(log->pending);
 819         free(log->filename);
 820         sfree(log);
 821 }
 822
 823 uint64_t hist_sum(int j, int stride, uint64_t *io_u_plat,
 824                 uint64_t *io_u_plat_last)
 825 {
 826         uint64_t sum;
 827         int k;
 828
 829         if (io_u_plat_last) {
 830                 for (k = sum = 0; k < stride; k++)
 831                         sum += io_u_plat[j + k] - io_u_plat_last[j + k];
 832         } else {
 833                 for (k = sum = 0; k < stride; k++)
 834                         sum += io_u_plat[j + k];
 835         }
 836
 837         return sum;
 838 }
 839
 840 static void flush_hist_samples(FILE *f, int hist_coarseness, void *samples,
 841                                uint64_t sample_size)
 842 {
 843         struct io_sample *s;
 844         int log_offset;
 845         uint64_t i, j, nr_samples;
 846         struct io_u_plat_entry *entry, *entry_before;
 847         uint64_t *io_u_plat;
 848         uint64_t *io_u_plat_before;
 849
 850         int stride = 1 << hist_coarseness;
 851
 852         if (!sample_size)
 853                 return;
 854
 855         s = __get_sample(samples, 0, 0);
 856         log_offset = (s->__ddir & LOG_OFFSET_SAMPLE_BIT) != 0;
 857
 858         nr_samples = sample_size / __log_entry_sz(log_offset);
 859
 860         for (i = 0; i < nr_samples; i++) {
 861                 s = __get_sample(samples, log_offset, i);
 862
 863                 entry = s->data.plat_entry;
 864                 io_u_plat = entry->io_u_plat;
 865
 866                 entry_before = flist_first_entry(&entry->list, struct io_u_plat_entry, list);
 867                 io_u_plat_before = entry_before->io_u_plat;
 868
 869                 fprintf(f, "%lu, %u, %llu, ", (unsigned long) s->time,
 870                                                 io_sample_ddir(s), (unsigned long long) s->bs);
 871                 for (j = 0; j < FIO_IO_U_PLAT_NR - stride; j += stride) {
 872                         fprintf(f, "%llu, ", (unsigned long long)
 873                                 hist_sum(j, stride, io_u_plat, io_u_plat_before));
 874                 }
 875                 fprintf(f, "%llu\n", (unsigned long long)
 876                         hist_sum(FIO_IO_U_PLAT_NR - stride, stride, io_u_plat,
 877                                         io_u_plat_before));
 878
 879                 flist_del(&entry_before->list);
 880                 free(entry_before);
 881         }
 882 }
 883
 884 void flush_samples(FILE *f, void *samples, uint64_t sample_size)
 885 {
 886         struct io_sample *s;
 887         int log_offset;
 888         uint64_t i, nr_samples;
 889
 890         if (!sample_size)
 891                 return;
 892
 893         s = __get_sample(samples, 0, 0);
 894         log_offset = (s->__ddir & LOG_OFFSET_SAMPLE_BIT) != 0;
 895
 896         nr_samples = sample_size / __log_entry_sz(log_offset);
 897
 898         for (i = 0; i < nr_samples; i++) {
 899                 s = __get_sample(samples, log_offset, i);
 900
 901                 if (!log_offset) {
 902                         fprintf(f, "%lu, %" PRId64 ", %u, %llu, %u\n",
 903                                         (unsigned long) s->time,
 904                                         s->data.val,
 905                                         io_sample_ddir(s), (unsigned long long) s->bs, s->priority_bit);
 906                 } else {
 907                         struct io_sample_offset *so = (void *) s;
 908
 909                         fprintf(f, "%lu, %" PRId64 ", %u, %llu, %llu, %u\n",
 910                                         (unsigned long) s->time,
 911                                         s->data.val,
 912                                         io_sample_ddir(s), (unsigned long long) s->bs,
 913                                         (unsigned long long) so->offset, s->priority_bit);
 914                 }
 915         }
 916 }
 917
 918 #ifdef CONFIG_ZLIB
 919
 920 struct iolog_flush_data {
 921         struct workqueue_work work;
 922         struct io_log *log;
 923         void *samples;
 924         uint32_t nr_samples;
 925         bool free;
 926 };
 927
 928 #define GZ_CHUNK        131072
 929
 930 static struct iolog_compress *get_new_chunk(unsigned int seq)
 931 {
 932         struct iolog_compress *c;
 933
 934         c = malloc(sizeof(*c));
 935         INIT_FLIST_HEAD(&c->list);
 936         c->buf = malloc(GZ_CHUNK);
 937         c->len = 0;
 938         c->seq = seq;
 939         return c;
 940 }
 941
 942 static void free_chunk(struct iolog_compress *ic)
 943 {
 944         free(ic->buf);
 945         free(ic);
 946 }
 947
 948 static int z_stream_init(z_stream *stream, int gz_hdr)
 949 {
 950         int wbits = 15;
 951
 952         memset(stream, 0, sizeof(*stream));
 953         stream->zalloc = Z_NULL;
 954         stream->zfree = Z_NULL;
 955         stream->opaque = Z_NULL;
 956         stream->next_in = Z_NULL;
 957
 958         /*
 959          * zlib magic - add 32 for auto-detection of gz header or not,
 960          * if we decide to store files in a gzip friendly format.
 961          */
 962         if (gz_hdr)
 963                 wbits += 32;
 964
 965         if (inflateInit2(stream, wbits) != Z_OK)
 966                 return 1;
 967
 968         return 0;
 969 }
 970
 971 struct inflate_chunk_iter {
 972         unsigned int seq;
 973         int err;
 974         void *buf;
 975         size_t buf_size;
 976         size_t buf_used;
 977         size_t chunk_sz;
 978 };
 979
 980 static void finish_chunk(z_stream *stream, FILE *f,
 981                          struct inflate_chunk_iter *iter)
 982 {
 983         int ret;
 984
 985         ret = inflateEnd(stream);
 986         if (ret != Z_OK)
 987                 log_err("fio: failed to end log inflation seq %d (%d)\n",
 988                                 iter->seq, ret);
 989
 990         flush_samples(f, iter->buf, iter->buf_used);
 991         free(iter->buf);
 992         iter->buf = NULL;
 993         iter->buf_size = iter->buf_used = 0;
 994 }
 995
 996 /*
 997  * Iterative chunk inflation. Handles cases where we cross into a new
 998  * sequence, doing flush finish of previous chunk if needed.
 999  */
1000 static size_t inflate_chunk(struct iolog_compress *ic, int gz_hdr, FILE *f,
1001                             z_stream *stream, struct inflate_chunk_iter *iter)
1002 {
1003         size_t ret;
1004
1005         dprint(FD_COMPRESS, "inflate chunk size=%lu, seq=%u\n",
1006                                 (unsigned long) ic->len, ic->seq);
1007
1008         if (ic->seq != iter->seq) {
1009                 if (iter->seq)
1010                         finish_chunk(stream, f, iter);
1011
1012                 z_stream_init(stream, gz_hdr);
1013                 iter->seq = ic->seq;
1014         }
1015
1016         stream->avail_in = ic->len;
1017         stream->next_in = ic->buf;
1018
1019         if (!iter->buf_size) {
1020                 iter->buf_size = iter->chunk_sz;
1021                 iter->buf = malloc(iter->buf_size);
1022         }
1023
1024         while (stream->avail_in) {
1025                 size_t this_out = iter->buf_size - iter->buf_used;
1026                 int err;
1027
1028                 stream->avail_out = this_out;
1029                 stream->next_out = iter->buf + iter->buf_used;
1030
1031                 err = inflate(stream, Z_NO_FLUSH);
1032                 if (err < 0) {
1033                         log_err("fio: failed inflating log: %d\n", err);
1034                         iter->err = err;
1035                         break;
1036                 }
1037
1038                 iter->buf_used += this_out - stream->avail_out;
1039
1040                 if (!stream->avail_out) {
1041                         iter->buf_size += iter->chunk_sz;
1042                         iter->buf = realloc(iter->buf, iter->buf_size);
1043                         continue;
1044                 }
1045
1046                 if (err == Z_STREAM_END)
1047                         break;
1048         }
1049
1050         ret = (void *) stream->next_in - ic->buf;
1051
1052         dprint(FD_COMPRESS, "inflated to size=%lu\n", (unsigned long) iter->buf_size);
1053
1054         return ret;
1055 }
1056
1057 /*
1058  * Inflate stored compressed chunks, or write them directly to the log
1059  * file if so instructed.
1060  */
1061 static int inflate_gz_chunks(struct io_log *log, FILE *f)
1062 {
1063         struct inflate_chunk_iter iter = { .chunk_sz = log->log_gz, };
1064         z_stream stream;
1065
1066         while (!flist_empty(&log->chunk_list)) {
1067                 struct iolog_compress *ic;
1068
1069                 ic = flist_first_entry(&log->chunk_list, struct iolog_compress, list);
1070                 flist_del(&ic->list);
1071
1072                 if (log->log_gz_store) {
1073                         size_t ret;
1074
1075                         dprint(FD_COMPRESS, "log write chunk size=%lu, "
1076                                 "seq=%u\n", (unsigned long) ic->len, ic->seq);
1077
1078                         ret = fwrite(ic->buf, ic->len, 1, f);
1079                         if (ret != 1 || ferror(f)) {
1080                                 iter.err = errno;
1081                                 log_err("fio: error writing compressed log\n");
1082                         }
1083                 } else
1084                         inflate_chunk(ic, log->log_gz_store, f, &stream, &iter);
1085
1086                 free_chunk(ic);
1087         }
1088
1089         if (iter.seq) {
1090                 finish_chunk(&stream, f, &iter);
1091                 free(iter.buf);
1092         }
1093
1094         return iter.err;
1095 }
1096
1097 /*
1098  * Open compressed log file and decompress the stored chunks and
1099  * write them to stdout. The chunks are stored sequentially in the
1100  * file, so we iterate over them and do them one-by-one.
1101  */
1102 int iolog_file_inflate(const char *file)
1103 {
1104         struct inflate_chunk_iter iter = { .chunk_sz = 64 * 1024 * 1024, };
1105         struct iolog_compress ic;
1106         z_stream stream;
1107         struct stat sb;
1108         size_t ret;
1109         size_t total;
1110         void *buf;
1111         FILE *f;
1112
1113         f = fopen(file, "r");
1114         if (!f) {
1115                 perror("fopen");
1116                 return 1;
1117         }
1118
1119         if (stat(file, &sb) < 0) {
1120                 fclose(f);
1121                 perror("stat");
1122                 return 1;
1123         }
1124
1125         ic.buf = buf = malloc(sb.st_size);
1126         ic.len = sb.st_size;
1127         ic.seq = 1;
1128
1129         ret = fread(ic.buf, ic.len, 1, f);
1130         if (ret == 0 && ferror(f)) {
1131                 perror("fread");
1132                 fclose(f);
1133                 free(buf);
1134                 return 1;
1135         } else if (ferror(f) || (!feof(f) && ret != 1)) {
1136                 log_err("fio: short read on reading log\n");
1137                 fclose(f);
1138                 free(buf);
1139                 return 1;
1140         }
1141
1142         fclose(f);
1143
1144         /*
1145          * Each chunk will return Z_STREAM_END. We don't know how many
1146          * chunks are in the file, so we just keep looping and incrementing
1147          * the sequence number until we have consumed the whole compressed
1148          * file.
1149          */
1150         total = ic.len;
1151         do {
1152                 size_t iret;
1153
1154                 iret = inflate_chunk(&ic,  1, stdout, &stream, &iter);
1155                 total -= iret;
1156                 if (!total)
1157                         break;
1158                 if (iter.err)
1159                         break;
1160
1161                 ic.seq++;
1162                 ic.len -= iret;
1163                 ic.buf += iret;
1164         } while (1);
1165
1166         if (iter.seq) {
1167                 finish_chunk(&stream, stdout, &iter);
1168                 free(iter.buf);
1169         }
1170
1171         free(buf);
1172         return iter.err;
1173 }
1174
1175 #else
1176
1177 static int inflate_gz_chunks(struct io_log *log, FILE *f)
1178 {
1179         return 0;
1180 }
1181
1182 int iolog_file_inflate(const char *file)
1183 {
1184         log_err("fio: log inflation not possible without zlib\n");
1185         return 1;
1186 }
1187
1188 #endif
1189
1190 void flush_log(struct io_log *log, bool do_append)
1191 {
1192         void *buf;
1193         FILE *f;
1194
1195         if (!do_append)
1196                 f = fopen(log->filename, "w");
1197         else
1198                 f = fopen(log->filename, "a");
1199         if (!f) {
1200                 perror("fopen log");
1201                 return;
1202         }
1203
1204         buf = set_file_buffer(f);
1205
1206         inflate_gz_chunks(log, f);
1207
1208         while (!flist_empty(&log->io_logs)) {
1209                 struct io_logs *cur_log;
1210
1211                 cur_log = flist_first_entry(&log->io_logs, struct io_logs, list);
1212                 flist_del_init(&cur_log->list);
1213
1214                 if (log->td && log == log->td->clat_hist_log)
1215                         flush_hist_samples(f, log->hist_coarseness, cur_log->log,
1216                                            log_sample_sz(log, cur_log));
1217                 else
1218                         flush_samples(f, cur_log->log, log_sample_sz(log, cur_log));
1219
1220                 sfree(cur_log);
1221         }
1222
1223         fclose(f);
1224         clear_file_buffer(buf);
1225 }
1226
1227 static int finish_log(struct thread_data *td, struct io_log *log, int trylock)
1228 {
1229         if (td->flags & TD_F_COMPRESS_LOG)
1230                 iolog_flush(log);
1231
1232         if (trylock) {
1233                 if (fio_trylock_file(log->filename))
1234                         return 1;
1235         } else
1236                 fio_lock_file(log->filename);
1237
1238         if (td->client_type == FIO_CLIENT_TYPE_GUI || is_backend)
1239                 fio_send_iolog(td, log, log->filename);
1240         else
1241                 flush_log(log, !td->o.per_job_logs);
1242
1243         fio_unlock_file(log->filename);
1244         free_log(log);
1245         return 0;
1246 }
1247
1248 size_t log_chunk_sizes(struct io_log *log)
1249 {
1250         struct flist_head *entry;
1251         size_t ret;
1252
1253         if (flist_empty(&log->chunk_list))
1254                 return 0;
1255
1256         ret = 0;
1257         pthread_mutex_lock(&log->chunk_lock);
1258         flist_for_each(entry, &log->chunk_list) {
1259                 struct iolog_compress *c;
1260
1261                 c = flist_entry(entry, struct iolog_compress, list);
1262                 ret += c->len;
1263         }
1264         pthread_mutex_unlock(&log->chunk_lock);
1265         return ret;
1266 }
1267
1268 #ifdef CONFIG_ZLIB
1269
1270 static void iolog_put_deferred(struct io_log *log, void *ptr)
1271 {
1272         if (!ptr)
1273                 return;
1274
1275         pthread_mutex_lock(&log->deferred_free_lock);
1276         if (log->deferred < IOLOG_MAX_DEFER) {
1277                 log->deferred_items[log->deferred] = ptr;
1278                 log->deferred++;
1279         } else if (!fio_did_warn(FIO_WARN_IOLOG_DROP))
1280                 log_err("fio: had to drop log entry free\n");
1281         pthread_mutex_unlock(&log->deferred_free_lock);
1282 }
1283
1284 static void iolog_free_deferred(struct io_log *log)
1285 {
1286         int i;
1287
1288         if (!log->deferred)
1289                 return;
1290
1291         pthread_mutex_lock(&log->deferred_free_lock);
1292
1293         for (i = 0; i < log->deferred; i++) {
1294                 free(log->deferred_items[i]);
1295                 log->deferred_items[i] = NULL;
1296         }
1297
1298         log->deferred = 0;
1299         pthread_mutex_unlock(&log->deferred_free_lock);
1300 }
1301
1302 static int gz_work(struct iolog_flush_data *data)
1303 {
1304         struct iolog_compress *c = NULL;
1305         struct flist_head list;
1306         unsigned int seq;
1307         z_stream stream;
1308         size_t total = 0;
1309         int ret;
1310
1311         INIT_FLIST_HEAD(&list);
1312
1313         memset(&stream, 0, sizeof(stream));
1314         stream.zalloc = Z_NULL;
1315         stream.zfree = Z_NULL;
1316         stream.opaque = Z_NULL;
1317
1318         ret = deflateInit(&stream, Z_DEFAULT_COMPRESSION);
1319         if (ret != Z_OK) {
1320                 log_err("fio: failed to init gz stream\n");
1321                 goto err;
1322         }
1323
1324         seq = ++data->log->chunk_seq;
1325
1326         stream.next_in = (void *) data->samples;
1327         stream.avail_in = data->nr_samples * log_entry_sz(data->log);
1328
1329         dprint(FD_COMPRESS, "deflate input size=%lu, seq=%u, log=%s\n",
1330                                 (unsigned long) stream.avail_in, seq,
1331                                 data->log->filename);
1332         do {
1333                 if (c)
1334                         dprint(FD_COMPRESS, "seq=%d, chunk=%lu\n", seq,
1335                                 (unsigned long) c->len);
1336                 c = get_new_chunk(seq);
1337                 stream.avail_out = GZ_CHUNK;
1338                 stream.next_out = c->buf;
1339                 ret = deflate(&stream, Z_NO_FLUSH);
1340                 if (ret < 0) {
1341                         log_err("fio: deflate log (%d)\n", ret);
1342                         free_chunk(c);
1343                         goto err;
1344                 }
1345
1346                 c->len = GZ_CHUNK - stream.avail_out;
1347                 flist_add_tail(&c->list, &list);
1348                 total += c->len;
1349         } while (stream.avail_in);
1350
1351         stream.next_out = c->buf + c->len;
1352         stream.avail_out = GZ_CHUNK - c->len;
1353
1354         ret = deflate(&stream, Z_FINISH);
1355         if (ret < 0) {
1356                 /*
1357                  * Z_BUF_ERROR is special, it just means we need more
1358                  * output space. We'll handle that below. Treat any other
1359                  * error as fatal.
1360                  */
1361                 if (ret != Z_BUF_ERROR) {
1362                         log_err("fio: deflate log (%d)\n", ret);
1363                         flist_del(&c->list);
1364                         free_chunk(c);
1365                         goto err;
1366                 }
1367         }
1368
1369         total -= c->len;
1370         c->len = GZ_CHUNK - stream.avail_out;
1371         total += c->len;
1372         dprint(FD_COMPRESS, "seq=%d, chunk=%lu\n", seq, (unsigned long) c->len);
1373
1374         if (ret != Z_STREAM_END) {
1375                 do {
1376                         c = get_new_chunk(seq);
1377                         stream.avail_out = GZ_CHUNK;
1378                         stream.next_out = c->buf;
1379                         ret = deflate(&stream, Z_FINISH);
1380                         c->len = GZ_CHUNK - stream.avail_out;
1381                         total += c->len;
1382                         flist_add_tail(&c->list, &list);
1383                         dprint(FD_COMPRESS, "seq=%d, chunk=%lu\n", seq,
1384                                 (unsigned long) c->len);
1385                 } while (ret != Z_STREAM_END);
1386         }
1387
1388         dprint(FD_COMPRESS, "deflated to size=%lu\n", (unsigned long) total);
1389
1390         ret = deflateEnd(&stream);
1391         if (ret != Z_OK)
1392                 log_err("fio: deflateEnd %d\n", ret);
1393
1394         iolog_put_deferred(data->log, data->samples);
1395
1396         if (!flist_empty(&list)) {
1397                 pthread_mutex_lock(&data->log->chunk_lock);
1398                 flist_splice_tail(&list, &data->log->chunk_list);
1399                 pthread_mutex_unlock(&data->log->chunk_lock);
1400         }
1401
1402         ret = 0;
1403 done:
1404         if (data->free)
1405                 sfree(data);
1406         return ret;
1407 err:
1408         while (!flist_empty(&list)) {
1409                 c = flist_first_entry(list.next, struct iolog_compress, list);
1410                 flist_del(&c->list);
1411                 free_chunk(c);
1412         }
1413         ret = 1;
1414         goto done;
1415 }
1416
1417 /*
1418  * Invoked from our compress helper thread, when logging would have exceeded
1419  * the specified memory limitation. Compresses the previously stored
1420  * entries.
1421  */
1422 static int gz_work_async(struct submit_worker *sw, struct workqueue_work *work)
1423 {
1424         return gz_work(container_of(work, struct iolog_flush_data, work));
1425 }
1426
1427 static int gz_init_worker(struct submit_worker *sw)
1428 {
1429         struct thread_data *td = sw->wq->td;
1430
1431         if (!fio_option_is_set(&td->o, log_gz_cpumask))
1432                 return 0;
1433
1434         if (fio_setaffinity(gettid(), td->o.log_gz_cpumask) == -1) {
1435                 log_err("gz: failed to set CPU affinity\n");
1436                 return 1;
1437         }
1438
1439         return 0;
1440 }
1441
1442 static struct workqueue_ops log_compress_wq_ops = {
1443         .fn             = gz_work_async,
1444         .init_worker_fn = gz_init_worker,
1445         .nice           = 1,
1446 };
1447
1448 int iolog_compress_init(struct thread_data *td, struct sk_out *sk_out)
1449 {
1450         if (!(td->flags & TD_F_COMPRESS_LOG))
1451                 return 0;
1452
1453         workqueue_init(td, &td->log_compress_wq, &log_compress_wq_ops, 1, sk_out);
1454         return 0;
1455 }
1456
1457 void iolog_compress_exit(struct thread_data *td)
1458 {
1459         if (!(td->flags & TD_F_COMPRESS_LOG))
1460                 return;
1461
1462         workqueue_exit(&td->log_compress_wq);
1463 }
1464
1465 /*
1466  * Queue work item to compress the existing log entries. We reset the
1467  * current log to a small size, and reference the existing log in the
1468  * data that we queue for compression. Once compression has been done,
1469  * this old log is freed. If called with finish == true, will not return
1470  * until the log compression has completed, and will flush all previous
1471  * logs too
1472  */
1473 static int iolog_flush(struct io_log *log)
1474 {
1475         struct iolog_flush_data *data;
1476
1477         data = malloc(sizeof(*data));
1478         if (!data)
1479                 return 1;
1480
1481         data->log = log;
1482         data->free = false;
1483
1484         while (!flist_empty(&log->io_logs)) {
1485                 struct io_logs *cur_log;
1486
1487                 cur_log = flist_first_entry(&log->io_logs, struct io_logs, list);
1488                 flist_del_init(&cur_log->list);
1489
1490                 data->samples = cur_log->log;
1491                 data->nr_samples = cur_log->nr_samples;
1492
1493                 sfree(cur_log);
1494
1495                 gz_work(data);
1496         }
1497
1498         free(data);
1499         return 0;
1500 }
1501
1502 int iolog_cur_flush(struct io_log *log, struct io_logs *cur_log)
1503 {
1504         struct iolog_flush_data *data;
1505
1506         data = smalloc(sizeof(*data));
1507         if (!data)
1508                 return 1;
1509
1510         data->log = log;
1511
1512         data->samples = cur_log->log;
1513         data->nr_samples = cur_log->nr_samples;
1514         data->free = true;
1515
1516         cur_log->nr_samples = cur_log->max_samples = 0;
1517         cur_log->log = NULL;
1518
1519         workqueue_enqueue(&log->td->log_compress_wq, &data->work);
1520
1521         iolog_free_deferred(log);
1522
1523         return 0;
1524 }
1525 #else
1526
1527 static int iolog_flush(struct io_log *log)
1528 {
1529         return 1;
1530 }
1531
1532 int iolog_cur_flush(struct io_log *log, struct io_logs *cur_log)
1533 {
1534         return 1;
1535 }
1536
1537 int iolog_compress_init(struct thread_data *td, struct sk_out *sk_out)
1538 {
1539         return 0;
1540 }
1541
1542 void iolog_compress_exit(struct thread_data *td)
1543 {
1544 }
1545
1546 #endif
1547
1548 struct io_logs *iolog_cur_log(struct io_log *log)
1549 {
1550         if (flist_empty(&log->io_logs))
1551                 return NULL;
1552
1553         return flist_last_entry(&log->io_logs, struct io_logs, list);
1554 }
1555
1556 uint64_t iolog_nr_samples(struct io_log *iolog)
1557 {
1558         struct flist_head *entry;
1559         uint64_t ret = 0;
1560
1561         flist_for_each(entry, &iolog->io_logs) {
1562                 struct io_logs *cur_log;
1563
1564                 cur_log = flist_entry(entry, struct io_logs, list);
1565                 ret += cur_log->nr_samples;
1566         }
1567
1568         return ret;
1569 }
1570
1571 static int __write_log(struct thread_data *td, struct io_log *log, int try)
1572 {
1573         if (log)
1574                 return finish_log(td, log, try);
1575
1576         return 0;
1577 }
1578
1579 static int write_iops_log(struct thread_data *td, int try, bool unit_log)
1580 {
1581         int ret;
1582
1583         if (per_unit_log(td->iops_log) != unit_log)
1584                 return 0;
1585
1586         ret = __write_log(td, td->iops_log, try);
1587         if (!ret)
1588                 td->iops_log = NULL;
1589
1590         return ret;
1591 }
1592
1593 static int write_slat_log(struct thread_data *td, int try, bool unit_log)
1594 {
1595         int ret;
1596
1597         if (!unit_log)
1598                 return 0;
1599
1600         ret = __write_log(td, td->slat_log, try);
1601         if (!ret)
1602                 td->slat_log = NULL;
1603
1604         return ret;
1605 }
1606
1607 static int write_clat_log(struct thread_data *td, int try, bool unit_log)
1608 {
1609         int ret;
1610
1611         if (!unit_log)
1612                 return 0;
1613
1614         ret = __write_log(td, td->clat_log, try);
1615         if (!ret)
1616                 td->clat_log = NULL;
1617
1618         return ret;
1619 }
1620
1621 static int write_clat_hist_log(struct thread_data *td, int try, bool unit_log)
1622 {
1623         int ret;
1624
1625         if (!unit_log)
1626                 return 0;
1627
1628         ret = __write_log(td, td->clat_hist_log, try);
1629         if (!ret)
1630                 td->clat_hist_log = NULL;
1631
1632         return ret;
1633 }
1634
1635 static int write_lat_log(struct thread_data *td, int try, bool unit_log)
1636 {
1637         int ret;
1638
1639         if (!unit_log)
1640                 return 0;
1641
1642         ret = __write_log(td, td->lat_log, try);
1643         if (!ret)
1644                 td->lat_log = NULL;
1645
1646         return ret;
1647 }
1648
1649 static int write_bandw_log(struct thread_data *td, int try, bool unit_log)
1650 {
1651         int ret;
1652
1653         if (per_unit_log(td->bw_log) != unit_log)
1654                 return 0;
1655
1656         ret = __write_log(td, td->bw_log, try);
1657         if (!ret)
1658                 td->bw_log = NULL;
1659
1660         return ret;
1661 }
1662
1663 enum {
1664         BW_LOG_MASK     = 1,
1665         LAT_LOG_MASK    = 2,
1666         SLAT_LOG_MASK   = 4,
1667         CLAT_LOG_MASK   = 8,
1668         IOPS_LOG_MASK   = 16,
1669         CLAT_HIST_LOG_MASK = 32,
1670
1671         ALL_LOG_NR      = 6,
1672 };
1673
1674 struct log_type {
1675         unsigned int mask;
1676         int (*fn)(struct thread_data *, int, bool);
1677 };
1678
1679 static struct log_type log_types[] = {
1680         {
1681                 .mask   = BW_LOG_MASK,
1682                 .fn     = write_bandw_log,
1683         },
1684         {
1685                 .mask   = LAT_LOG_MASK,
1686                 .fn     = write_lat_log,
1687         },
1688         {
1689                 .mask   = SLAT_LOG_MASK,
1690                 .fn     = write_slat_log,
1691         },
1692         {
1693                 .mask   = CLAT_LOG_MASK,
1694                 .fn     = write_clat_log,
1695         },
1696         {
1697                 .mask   = IOPS_LOG_MASK,
1698                 .fn     = write_iops_log,
1699         },
1700         {
1701                 .mask   = CLAT_HIST_LOG_MASK,
1702                 .fn     = write_clat_hist_log,
1703         }
1704 };
1705
1706 void td_writeout_logs(struct thread_data *td, bool unit_logs)
1707 {
1708         unsigned int log_mask = 0;
1709         unsigned int log_left = ALL_LOG_NR;
1710         int old_state, i;
1711
1712         old_state = td_bump_runstate(td, TD_FINISHING);
1713
1714         finalize_logs(td, unit_logs);
1715
1716         while (log_left) {
1717                 int prev_log_left = log_left;
1718
1719                 for (i = 0; i < ALL_LOG_NR && log_left; i++) {
1720                         struct log_type *lt = &log_types[i];
1721                         int ret;
1722
1723                         if (!(log_mask & lt->mask)) {
1724                                 ret = lt->fn(td, log_left != 1, unit_logs);
1725                                 if (!ret) {
1726                                         log_left--;
1727                                         log_mask |= lt->mask;
1728                                 }
1729                         }
1730                 }
1731
1732                 if (prev_log_left == log_left)
1733                         usleep(5000);
1734         }
1735
1736         td_restore_runstate(td, old_state);
1737 }
1738
1739 void fio_writeout_logs(bool unit_logs)
1740 {
1741         struct thread_data *td;
1742         int i;
1743
1744         for_each_td(td, i)
1745                 td_writeout_logs(td, unit_logs);
1746 }