iolog.c

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