iolog.c

   1 /*
   2  * Code related to writing an iolog of what a thread is doing, and to
   3  * later read that back and replay
   4  */
   5 #include <stdio.h>
   6 #include <stdlib.h>
   7 #include <assert.h>
   8 #include <sys/types.h>
   9 #include <sys/stat.h>
  10 #include <unistd.h>
  11 #ifdef CONFIG_ZLIB
  12 #include <zlib.h>
  13 #endif
  14
  15 #include "flist.h"
  16 #include "fio.h"
  17 #include "trim.h"
  18 #include "filelock.h"
  19 #include "smalloc.h"
  20 #include "blktrace.h"
  21 #include "pshared.h"
  22 #include "lib/roundup.h"
  23
  24 #include <netinet/in.h>
  25 #include <netinet/tcp.h>
  26 #include <arpa/inet.h>
  27 #include <sys/stat.h>
  28 #include <sys/socket.h>
  29 #include <sys/un.h>
  30
  31 static int iolog_flush(struct io_log *log);
  32
  33 static const char iolog_ver2[] = "fio version 2 iolog";
  34 static const char iolog_ver3[] = "fio version 3 iolog";
  35
  36 void queue_io_piece(struct thread_data *td, struct io_piece *ipo)
  37 {
  38         flist_add_tail(&ipo->list, &td->io_log_list);
  39         td->total_io_size += ipo->len;
  40 }
  41
  42 void log_io_u(const struct thread_data *td, const struct io_u *io_u)
  43 {
  44         if (!td->o.write_iolog_file)
  45                 return;
  46
  47         fprintf(td->iolog_f, "%s %s %llu %llu\n", io_u->file->file_name,
  48                                                 io_ddir_name(io_u->ddir),
  49                                                 io_u->offset, io_u->buflen);
  50 }
  51
  52 void log_file(struct thread_data *td, struct fio_file *f,
  53               enum file_log_act what)
  54 {
  55         const char *act[] = { "add", "open", "close" };
  56
  57         assert(what < 3);
  58
  59         if (!td->o.write_iolog_file)
  60                 return;
  61
  62
  63         /*
  64          * this happens on the pre-open/close done before the job starts
  65          */
  66         if (!td->iolog_f)
  67                 return;
  68
  69         fprintf(td->iolog_f, "%s %s\n", f->file_name, act[what]);
  70 }
  71
  72 static void iolog_delay(struct thread_data *td, unsigned long delay)
  73 {
  74         uint64_t usec = utime_since_now(&td->last_issue);
  75         unsigned long orig_delay = delay;
  76         uint64_t this_delay;
  77         struct timespec ts;
  78
  79         if (delay < td->time_offset) {
  80                 td->time_offset = 0;
  81                 return;
  82         }
  83
  84         delay -= td->time_offset;
  85         if (delay < usec)
  86                 return;
  87
  88         delay -= usec;
  89
  90         fio_gettime(&ts, NULL);
  91         while (delay && !td->terminate) {
  92                 this_delay = delay;
  93                 if (this_delay > 500000)
  94                         this_delay = 500000;
  95
  96                 usec_sleep(td, this_delay);
  97                 delay -= this_delay;
  98         }
  99
 100         usec = utime_since_now(&ts);
 101         if (usec > orig_delay)
 102                 td->time_offset = usec - orig_delay;
 103         else
 104                 td->time_offset = 0;
 105 }
 106
 107 static int ipo_special(struct thread_data *td, struct io_piece *ipo)
 108 {
 109         struct fio_file *f;
 110         int ret;
 111
 112         /*
 113          * Not a special ipo
 114          */
 115         if (ipo->ddir != DDIR_INVAL)
 116                 return 0;
 117
 118         f = td->files[ipo->fileno];
 119
 120         if (ipo->delay)
 121                 iolog_delay(td, ipo->delay);
 122         if (fio_fill_issue_time(td))
 123                 fio_gettime(&td->last_issue, NULL);
 124         switch (ipo->file_action) {
 125         case FIO_LOG_OPEN_FILE:
 126                 if (td->o.replay_redirect && fio_file_open(f)) {
 127                         dprint(FD_FILE, "iolog: ignoring re-open of file %s\n",
 128                                         f->file_name);
 129                         break;
 130                 }
 131                 ret = td_io_open_file(td, f);
 132                 if (!ret)
 133                         break;
 134                 td_verror(td, ret, "iolog open file");
 135                 return -1;
 136         case FIO_LOG_CLOSE_FILE:
 137                 td_io_close_file(td, f);
 138                 break;
 139         case FIO_LOG_UNLINK_FILE:
 140                 td_io_unlink_file(td, f);
 141                 break;
 142         case FIO_LOG_ADD_FILE:
 143                 /*
 144                  * Nothing to do
 145                  */
 146                 break;
 147         default:
 148                 log_err("fio: bad file action %d\n", ipo->file_action);
 149                 break;
 150         }
 151
 152         return 1;
 153 }
 154
 155 static bool read_iolog(struct thread_data *td);
 156
 157 unsigned long long delay_since_ttime(const struct thread_data *td,
 158                unsigned long long time)
 159 {
 160         double tmp;
 161         double scale;
 162         const unsigned long long *last_ttime = &td->io_log_last_ttime;
 163
 164         if (!*last_ttime || td->o.no_stall || time < *last_ttime)
 165                 return 0;
 166         else if (td->o.replay_time_scale == 100)
 167                 return time - *last_ttime;
 168
 169
 170         scale = (double) 100.0 / (double) td->o.replay_time_scale;
 171         tmp = time - *last_ttime;
 172         return tmp * scale;
 173 }
 174
 175 int read_iolog_get(struct thread_data *td, struct io_u *io_u)
 176 {
 177         struct io_piece *ipo;
 178         unsigned long elapsed;
 179
 180         while (!flist_empty(&td->io_log_list)) {
 181                 int ret;
 182
 183                 if (td->o.read_iolog_chunked) {
 184                         if (td->io_log_checkmark == td->io_log_current) {
 185                                 if (td->io_log_blktrace) {
 186                                         if (!read_blktrace(td))
 187                                                 return 1;
 188                                 } else {
 189                                         if (!read_iolog(td))
 190                                                 return 1;
 191                                 }
 192                         }
 193                         td->io_log_current--;
 194                 }
 195                 ipo = flist_first_entry(&td->io_log_list, struct io_piece, list);
 196                 flist_del(&ipo->list);
 197                 remove_trim_entry(td, ipo);
 198
 199                 ret = ipo_special(td, ipo);
 200                 if (ret < 0) {
 201                         free(ipo);
 202                         break;
 203                 } else if (ret > 0) {
 204                         free(ipo);
 205                         continue;
 206                 }
 207
 208                 io_u->ddir = ipo->ddir;
 209                 if (ipo->ddir != DDIR_WAIT) {
 210                         io_u->offset = ipo->offset;
 211                         io_u->verify_offset = ipo->offset;
 212                         io_u->buflen = ipo->len;
 213                         io_u->file = td->files[ipo->fileno];
 214                         get_file(io_u->file);
 215                         dprint(FD_IO, "iolog: get %llu/%llu/%s\n", io_u->offset,
 216                                                 io_u->buflen, io_u->file->file_name);
 217                         if (ipo->delay)
 218                                 iolog_delay(td, ipo->delay);
 219                 } else {
 220                         elapsed = mtime_since_genesis();
 221                         if (ipo->delay > elapsed)
 222                                 usec_sleep(td, (ipo->delay - elapsed) * 1000);
 223                 }
 224
 225                 free(ipo);
 226
 227                 if (io_u->ddir != DDIR_WAIT)
 228                         return 0;
 229         }
 230
 231         td->done = 1;
 232         return 1;
 233 }
 234
 235 void prune_io_piece_log(struct thread_data *td)
 236 {
 237         struct io_piece *ipo;
 238         struct fio_rb_node *n;
 239
 240         while ((n = rb_first(&td->io_hist_tree)) != NULL) {
 241                 ipo = rb_entry(n, struct io_piece, rb_node);
 242                 rb_erase(n, &td->io_hist_tree);
 243                 remove_trim_entry(td, ipo);
 244                 td->io_hist_len--;
 245                 free(ipo);
 246         }
 247
 248         while (!flist_empty(&td->io_hist_list)) {
 249                 ipo = flist_first_entry(&td->io_hist_list, struct io_piece, list);
 250                 flist_del(&ipo->list);
 251                 remove_trim_entry(td, ipo);
 252                 td->io_hist_len--;
 253                 free(ipo);
 254         }
 255 }
 256
 257 /*
 258  * log a successful write, so we can unwind the log for verify
 259  */
 260 void log_io_piece(struct thread_data *td, struct io_u *io_u)
 261 {
 262         struct fio_rb_node **p, *parent;
 263         struct io_piece *ipo, *__ipo;
 264
 265         ipo = calloc(1, sizeof(struct io_piece));
 266         init_ipo(ipo);
 267         ipo->file = io_u->file;
 268         ipo->offset = io_u->offset;
 269         ipo->len = io_u->buflen;
 270         ipo->numberio = io_u->numberio;
 271         ipo->flags = IP_F_IN_FLIGHT;
 272
 273         io_u->ipo = ipo;
 274
 275         if (io_u_should_trim(td, io_u)) {
 276                 flist_add_tail(&ipo->trim_list, &td->trim_list);
 277                 td->trim_entries++;
 278         }
 279
 280         /*
 281          * Only sort writes if we don't have a random map in which case we need
 282          * to check for duplicate blocks and drop the old one, which we rely on
 283          * the rb insert/lookup for handling.
 284          */
 285         if (file_randommap(td, ipo->file)) {
 286                 INIT_FLIST_HEAD(&ipo->list);
 287                 flist_add_tail(&ipo->list, &td->io_hist_list);
 288                 ipo->flags |= IP_F_ONLIST;
 289                 td->io_hist_len++;
 290                 return;
 291         }
 292
 293         RB_CLEAR_NODE(&ipo->rb_node);
 294
 295         /*
 296          * Sort the entry into the verification list
 297          */
 298 restart:
 299         p = &td->io_hist_tree.rb_node;
 300         parent = NULL;
 301         while (*p) {
 302                 int overlap = 0;
 303                 parent = *p;
 304
 305                 __ipo = rb_entry(parent, struct io_piece, rb_node);
 306                 if (ipo->file < __ipo->file)
 307                         p = &(*p)->rb_left;
 308                 else if (ipo->file > __ipo->file)
 309                         p = &(*p)->rb_right;
 310                 else if (ipo->offset < __ipo->offset) {
 311                         p = &(*p)->rb_left;
 312                         overlap = ipo->offset + ipo->len > __ipo->offset;
 313                 }
 314                 else if (ipo->offset > __ipo->offset) {
 315                         p = &(*p)->rb_right;
 316                         overlap = __ipo->offset + __ipo->len > ipo->offset;
 317                 }
 318                 else
 319                         overlap = 1;
 320
 321                 if (overlap) {
 322                         dprint(FD_IO, "iolog: overlap %llu/%lu, %llu/%lu\n",
 323                                 __ipo->offset, __ipo->len,
 324                                 ipo->offset, ipo->len);
 325                         td->io_hist_len--;
 326                         rb_erase(parent, &td->io_hist_tree);
 327                         remove_trim_entry(td, __ipo);
 328                         if (!(__ipo->flags & IP_F_IN_FLIGHT))
 329                                 free(__ipo);
 330                         goto restart;
 331                 }
 332         }
 333
 334         rb_link_node(&ipo->rb_node, parent, p);
 335         rb_insert_color(&ipo->rb_node, &td->io_hist_tree);
 336         ipo->flags |= IP_F_ONRB;
 337         td->io_hist_len++;
 338 }
 339
 340 void unlog_io_piece(struct thread_data *td, struct io_u *io_u)
 341 {
 342         struct io_piece *ipo = io_u->ipo;
 343
 344         if (td->ts.nr_block_infos) {
 345                 uint32_t *info = io_u_block_info(td, io_u);
 346                 if (BLOCK_INFO_STATE(*info) < BLOCK_STATE_TRIM_FAILURE) {
 347                         if (io_u->ddir == DDIR_TRIM)
 348                                 *info = BLOCK_INFO_SET_STATE(*info,
 349                                                 BLOCK_STATE_TRIM_FAILURE);
 350                         else if (io_u->ddir == DDIR_WRITE)
 351                                 *info = BLOCK_INFO_SET_STATE(*info,
 352                                                 BLOCK_STATE_WRITE_FAILURE);
 353                 }
 354         }
 355
 356         if (!ipo)
 357                 return;
 358
 359         if (ipo->flags & IP_F_ONRB)
 360                 rb_erase(&ipo->rb_node, &td->io_hist_tree);
 361         else if (ipo->flags & IP_F_ONLIST)
 362                 flist_del(&ipo->list);
 363
 364         free(ipo);
 365         io_u->ipo = NULL;
 366         td->io_hist_len--;
 367 }
 368
 369 void trim_io_piece(const struct io_u *io_u)
 370 {
 371         struct io_piece *ipo = io_u->ipo;
 372
 373         if (!ipo)
 374                 return;
 375
 376         ipo->len = io_u->xfer_buflen - io_u->resid;
 377 }
 378
 379 void write_iolog_close(struct thread_data *td)
 380 {
 381         if (!td->iolog_f)
 382                 return;
 383
 384         fflush(td->iolog_f);
 385         fclose(td->iolog_f);
 386         free(td->iolog_buf);
 387         td->iolog_f = NULL;
 388         td->iolog_buf = NULL;
 389 }
 390
 391 int64_t iolog_items_to_fetch(struct thread_data *td)
 392 {
 393         struct timespec now;
 394         uint64_t elapsed;
 395         uint64_t for_1s;
 396         int64_t items_to_fetch;
 397
 398         if (!td->io_log_highmark)
 399                 return 10;
 400
 401
 402         fio_gettime(&now, NULL);
 403         elapsed = ntime_since(&td->io_log_highmark_time, &now);
 404         if (elapsed) {
 405                 for_1s = (td->io_log_highmark - td->io_log_current) * 1000000000 / elapsed;
 406                 items_to_fetch = for_1s - td->io_log_current;
 407                 if (items_to_fetch < 0)
 408                         items_to_fetch = 0;
 409         } else
 410                 items_to_fetch = 0;
 411
 412         td->io_log_highmark = td->io_log_current + items_to_fetch;
 413         td->io_log_checkmark = (td->io_log_highmark + 1) / 2;
 414         fio_gettime(&td->io_log_highmark_time, NULL);
 415
 416         return items_to_fetch;
 417 }
 418
 419 #define io_act(_td, _r) (((_td)->io_log_version == 3 && (r) == 5) || \
 420                                         ((_td)->io_log_version == 2 && (r) == 4))
 421 #define file_act(_td, _r) (((_td)->io_log_version == 3 && (r) == 3) || \
 422                                         ((_td)->io_log_version == 2 && (r) == 2))
 423
 424 /*
 425  * Read version 2 and 3 iolog data. It is enhanced to include per-file logging,
 426  * syncs, etc.
 427  */
 428 static bool read_iolog(struct thread_data *td)
 429 {
 430         unsigned long long offset;
 431         unsigned int bytes;
 432         unsigned long long delay = 0;
 433         int reads, writes, waits, fileno = 0, file_action = 0; /* stupid gcc */
 434         char *rfname, *fname, *act;
 435         char *str, *p;
 436         enum fio_ddir rw;
 437         bool realloc = false;
 438         int64_t items_to_fetch = 0;
 439         int syncs;
 440
 441         if (td->o.read_iolog_chunked) {
 442                 items_to_fetch = iolog_items_to_fetch(td);
 443                 if (!items_to_fetch)
 444                         return true;
 445         }
 446
 447         /*
 448          * Read in the read iolog and store it, reuse the infrastructure
 449          * for doing verifications.
 450          */
 451         str = malloc(4096);
 452         rfname = fname = malloc(256+16);
 453         act = malloc(256+16);
 454
 455         syncs = reads = writes = waits = 0;
 456         while ((p = fgets(str, 4096, td->io_log_rfile)) != NULL) {
 457                 struct io_piece *ipo;
 458                 int r;
 459                 unsigned long long ttime;
 460
 461                 if (td->io_log_version == 3) {
 462                         r = sscanf(p, "%llu %256s %256s %llu %u", &ttime, rfname, act,
 463                                                         &offset, &bytes);
 464                         delay = delay_since_ttime(td, ttime);
 465                         td->io_log_last_ttime = ttime;
 466                         /*
 467                          * "wait" is not allowed with version 3
 468                          */
 469                         if (!strcmp(act, "wait")) {
 470                                 log_err("iolog: ignoring wait command with"
 471                                         " version 3 for file %s\n", fname);
 472                                 continue;
 473                         }
 474                 } else /* version 2 */
 475                         r = sscanf(p, "%256s %256s %llu %u", rfname, act, &offset, &bytes);
 476
 477                 if (td->o.replay_redirect)
 478                         fname = td->o.replay_redirect;
 479
 480                 if (io_act(td, r)) {
 481                         /*
 482                          * Check action first
 483                          */
 484                         if (!strcmp(act, "wait"))
 485                                 rw = DDIR_WAIT;
 486                         else if (!strcmp(act, "read"))
 487                                 rw = DDIR_READ;
 488                         else if (!strcmp(act, "write"))
 489                                 rw = DDIR_WRITE;
 490                         else if (!strcmp(act, "sync"))
 491                                 rw = DDIR_SYNC;
 492                         else if (!strcmp(act, "datasync"))
 493                                 rw = DDIR_DATASYNC;
 494                         else if (!strcmp(act, "trim"))
 495                                 rw = DDIR_TRIM;
 496                         else {
 497                                 log_err("fio: bad iolog file action: %s\n",
 498                                                                         act);
 499                                 continue;
 500                         }
 501                         fileno = get_fileno(td, fname);
 502                 } else if (file_act(td, r)) {
 503                         rw = DDIR_INVAL;
 504                         if (!strcmp(act, "add")) {
 505                                 if (td->o.replay_redirect &&
 506                                     get_fileno(td, fname) != -1) {
 507                                         dprint(FD_FILE, "iolog: ignoring"
 508                                                 " re-add of file %s\n", fname);
 509                                 } else {
 510                                         fileno = add_file(td, fname, td->subjob_number, 1);
 511                                         file_action = FIO_LOG_ADD_FILE;
 512                                 }
 513                         } else if (!strcmp(act, "open")) {
 514                                 fileno = get_fileno(td, fname);
 515                                 file_action = FIO_LOG_OPEN_FILE;
 516                         } else if (!strcmp(act, "close")) {
 517                                 fileno = get_fileno(td, fname);
 518                                 file_action = FIO_LOG_CLOSE_FILE;
 519                         } else {
 520                                 log_err("fio: bad iolog file action: %s\n",
 521                                                                         act);
 522                                 continue;
 523                         }
 524                 } else {
 525                         log_err("bad iolog%d: %s\n", td->io_log_version, p);
 526                         continue;
 527                 }
 528
 529                 if (rw == DDIR_READ)
 530                         reads++;
 531                 else if (rw == DDIR_WRITE) {
 532                         /*
 533                          * Don't add a write for ro mode
 534                          */
 535                         if (read_only)
 536                                 continue;
 537                         writes++;
 538                 } else if (rw == DDIR_WAIT) {
 539                         if (td->o.no_stall)
 540                                 continue;
 541                         waits++;
 542                 } else if (rw == DDIR_INVAL) {
 543                 } else if (ddir_sync(rw)) {
 544                         syncs++;
 545                 } else {
 546                         log_err("bad ddir: %d\n", rw);
 547                         continue;
 548                 }
 549
 550                 /*
 551                  * Make note of file
 552                  */
 553                 ipo = calloc(1, sizeof(*ipo));
 554                 init_ipo(ipo);
 555                 ipo->ddir = rw;
 556                 if (td->io_log_version == 3)
 557                         ipo->delay = delay;
 558                 if (rw == DDIR_WAIT) {
 559                         ipo->delay = offset;
 560                 } else {
 561                         if (td->o.replay_scale)
 562                                 ipo->offset = offset / td->o.replay_scale;
 563                         else
 564                                 ipo->offset = offset;
 565                         ipo_bytes_align(td->o.replay_align, ipo);
 566
 567                         ipo->len = bytes;
 568                         if (rw != DDIR_INVAL && bytes > td->o.max_bs[rw]) {
 569                                 realloc = true;
 570                                 td->o.max_bs[rw] = bytes;
 571                         }
 572                         ipo->fileno = fileno;
 573                         ipo->file_action = file_action;
 574                         td->o.size += bytes;
 575                 }
 576
 577                 queue_io_piece(td, ipo);
 578
 579                 if (td->o.read_iolog_chunked) {
 580                         td->io_log_current++;
 581                         items_to_fetch--;
 582                         if (items_to_fetch == 0)
 583                                 break;
 584                 }
 585         }
 586
 587         free(str);
 588         free(act);
 589         free(rfname);
 590
 591         if (td->o.read_iolog_chunked) {
 592                 td->io_log_highmark = td->io_log_current;
 593                 td->io_log_checkmark = (td->io_log_highmark + 1) / 2;
 594                 fio_gettime(&td->io_log_highmark_time, NULL);
 595         }
 596
 597         if (writes && read_only) {
 598                 log_err("fio: <%s> skips replay of %d writes due to"
 599                         " read-only\n", td->o.name, writes);
 600                 writes = 0;
 601         }
 602         if (syncs)
 603                 td->flags |= TD_F_SYNCS;
 604
 605         if (td->o.read_iolog_chunked) {
 606                 if (td->io_log_current == 0) {
 607                         return false;
 608                 }
 609                 td->o.td_ddir = TD_DDIR_RW;
 610                 if (realloc && td->orig_buffer)
 611                 {
 612                         io_u_quiesce(td);
 613                         free_io_mem(td);
 614                         init_io_u_buffers(td);
 615                 }
 616                 return true;
 617         }
 618
 619         if (!reads && !writes && !waits)
 620                 return false;
 621         else if (reads && !writes)
 622                 td->o.td_ddir = TD_DDIR_READ;
 623         else if (!reads && writes)
 624                 td->o.td_ddir = TD_DDIR_WRITE;
 625         else
 626                 td->o.td_ddir = TD_DDIR_RW;
 627
 628         return true;
 629 }
 630
 631 static bool is_socket(const char *path)
 632 {
 633         struct stat buf;
 634         int r;
 635
 636         r = stat(path, &buf);
 637         if (r == -1)
 638                 return false;
 639
 640         return S_ISSOCK(buf.st_mode);
 641 }
 642
 643 static int open_socket(const char *path)
 644 {
 645         struct sockaddr_un addr;
 646         int ret, fd;
 647
 648         fd = socket(AF_UNIX, SOCK_STREAM, 0);
 649         if (fd < 0)
 650                 return fd;
 651
 652         addr.sun_family = AF_UNIX;
 653         if (snprintf(addr.sun_path, sizeof(addr.sun_path), "%s", path) >=
 654             sizeof(addr.sun_path)) {
 655                 log_err("%s: path name %s is too long for a Unix socket\n",
 656                         __func__, path);
 657         }
 658
 659         ret = connect(fd, (const struct sockaddr *)&addr, strlen(path) + sizeof(addr.sun_family));
 660         if (!ret)
 661                 return fd;
 662
 663         close(fd);
 664         return -1;
 665 }
 666
 667 /*
 668  * open iolog, check version, and call appropriate parser
 669  */
 670 static bool init_iolog_read(struct thread_data *td, char *fname)
 671 {
 672         char buffer[256], *p;
 673         FILE *f = NULL;
 674
 675         dprint(FD_IO, "iolog: name=%s\n", fname);
 676
 677         if (is_socket(fname)) {
 678                 int fd;
 679
 680                 fd = open_socket(fname);
 681                 if (fd >= 0)
 682                         f = fdopen(fd, "r");
 683         } else if (!strcmp(fname, "-")) {
 684                 f = stdin;
 685         } else
 686                 f = fopen(fname, "r");
 687
 688         if (!f) {
 689                 perror("fopen read iolog");
 690                 return false;
 691         }
 692
 693         p = fgets(buffer, sizeof(buffer), f);
 694         if (!p) {
 695                 td_verror(td, errno, "iolog read");
 696                 log_err("fio: unable to read iolog\n");
 697                 fclose(f);
 698                 return false;
 699         }
 700
 701         /*
 702          * versions 2 and 3 of the iolog store a specific string as the
 703          * first line, check for that
 704          */
 705         if (!strncmp(iolog_ver2, buffer, strlen(iolog_ver2)))
 706                 td->io_log_version = 2;
 707         else if (!strncmp(iolog_ver3, buffer, strlen(iolog_ver3)))
 708                 td->io_log_version = 3;
 709         else {
 710                 log_err("fio: iolog version 1 is no longer supported\n");
 711                 fclose(f);
 712                 return false;
 713         }
 714
 715         free_release_files(td);
 716         td->io_log_rfile = f;
 717         return read_iolog(td);
 718 }
 719
 720 /*
 721  * Set up a log for storing io patterns.
 722  */
 723 static bool init_iolog_write(struct thread_data *td)
 724 {
 725         struct fio_file *ff;
 726         FILE *f;
 727         unsigned int i;
 728
 729         f = fopen(td->o.write_iolog_file, "a");
 730         if (!f) {
 731                 perror("fopen write iolog");
 732                 return false;
 733         }
 734
 735         /*
 736          * That's it for writing, setup a log buffer and we're done.
 737           */
 738         td->iolog_f = f;
 739         td->iolog_buf = malloc(8192);
 740         setvbuf(f, td->iolog_buf, _IOFBF, 8192);
 741
 742         /*
 743          * write our version line
 744          */
 745         if (fprintf(f, "%s\n", iolog_ver2) < 0) {
 746                 perror("iolog init\n");
 747                 return false;
 748         }
 749
 750         /*
 751          * add all known files
 752          */
 753         for_each_file(td, ff, i)
 754                 log_file(td, ff, FIO_LOG_ADD_FILE);
 755
 756         return true;
 757 }
 758
 759 bool init_iolog(struct thread_data *td)
 760 {
 761         bool ret;
 762
 763         if (td->o.read_iolog_file) {
 764                 int need_swap;
 765                 char * fname = get_name_by_idx(td->o.read_iolog_file, td->subjob_number);
 766
 767                 /*
 768                  * Check if it's a blktrace file and load that if possible.
 769                  * Otherwise assume it's a normal log file and load that.
 770                  */
 771                 if (is_blktrace(fname, &need_swap)) {
 772                         td->io_log_blktrace = 1;
 773                         ret = init_blktrace_read(td, fname, need_swap);
 774                 } else {
 775                         td->io_log_blktrace = 0;
 776                         ret = init_iolog_read(td, fname);
 777                 }
 778                 free(fname);
 779         } else if (td->o.write_iolog_file)
 780                 ret = init_iolog_write(td);
 781         else
 782                 ret = true;
 783
 784         if (!ret)
 785                 td_verror(td, EINVAL, "failed initializing iolog");
 786
 787         return ret;
 788 }
 789
 790 void setup_log(struct io_log **log, struct log_params *p,
 791                const char *filename)
 792 {
 793         struct io_log *l;
 794         int i;
 795         struct io_u_plat_entry *entry;
 796         struct flist_head *list;
 797
 798         l = scalloc(1, sizeof(*l));
 799         INIT_FLIST_HEAD(&l->io_logs);
 800         l->log_type = p->log_type;
 801         l->log_offset = p->log_offset;
 802         l->log_prio = p->log_prio;
 803         l->log_gz = p->log_gz;
 804         l->log_gz_store = p->log_gz_store;
 805         l->avg_msec = p->avg_msec;
 806         l->hist_msec = p->hist_msec;
 807         l->hist_coarseness = p->hist_coarseness;
 808         l->filename = strdup(filename);
 809         l->td = p->td;
 810
 811         /* Initialize histogram lists for each r/w direction,
 812          * with initial io_u_plat of all zeros:
 813          */
 814         for (i = 0; i < DDIR_RWDIR_CNT; i++) {
 815                 list = &l->hist_window[i].list;
 816                 INIT_FLIST_HEAD(list);
 817                 entry = calloc(1, sizeof(struct io_u_plat_entry));
 818                 flist_add(&entry->list, list);
 819         }
 820
 821         if (l->td && l->td->o.io_submit_mode != IO_MODE_OFFLOAD) {
 822                 unsigned int def_samples = DEF_LOG_ENTRIES;
 823                 struct io_logs *__p;
 824
 825                 __p = calloc(1, sizeof(*l->pending));
 826                 if (l->td->o.iodepth > DEF_LOG_ENTRIES)
 827                         def_samples = roundup_pow2(l->td->o.iodepth);
 828                 __p->max_samples = def_samples;
 829                 __p->log = calloc(__p->max_samples, log_entry_sz(l));
 830                 l->pending = __p;
 831         }
 832
 833         if (l->log_offset)
 834                 l->log_ddir_mask = LOG_OFFSET_SAMPLE_BIT;
 835         if (l->log_prio)
 836                 l->log_ddir_mask |= LOG_PRIO_SAMPLE_BIT;
 837
 838         INIT_FLIST_HEAD(&l->chunk_list);
 839
 840         if (l->log_gz && !p->td)
 841                 l->log_gz = 0;
 842         else if (l->log_gz || l->log_gz_store) {
 843                 mutex_init_pshared(&l->chunk_lock);
 844                 mutex_init_pshared(&l->deferred_free_lock);
 845                 p->td->flags |= TD_F_COMPRESS_LOG;
 846         }
 847
 848         *log = l;
 849 }
 850
 851 #ifdef CONFIG_SETVBUF
 852 static void *set_file_buffer(FILE *f)
 853 {
 854         size_t size = 1048576;
 855         void *buf;
 856
 857         buf = malloc(size);
 858         setvbuf(f, buf, _IOFBF, size);
 859         return buf;
 860 }
 861
 862 static void clear_file_buffer(void *buf)
 863 {
 864         free(buf);
 865 }
 866 #else
 867 static void *set_file_buffer(FILE *f)
 868 {
 869         return NULL;
 870 }
 871
 872 static void clear_file_buffer(void *buf)
 873 {
 874 }
 875 #endif
 876
 877 void free_log(struct io_log *log)
 878 {
 879         while (!flist_empty(&log->io_logs)) {
 880                 struct io_logs *cur_log;
 881
 882                 cur_log = flist_first_entry(&log->io_logs, struct io_logs, list);
 883                 flist_del_init(&cur_log->list);
 884                 free(cur_log->log);
 885                 sfree(cur_log);
 886         }
 887
 888         if (log->pending) {
 889                 free(log->pending->log);
 890                 free(log->pending);
 891                 log->pending = NULL;
 892         }
 893
 894         free(log->pending);
 895         free(log->filename);
 896         sfree(log);
 897 }
 898
 899 uint64_t hist_sum(int j, int stride, uint64_t *io_u_plat,
 900                 uint64_t *io_u_plat_last)
 901 {
 902         uint64_t sum;
 903         int k;
 904
 905         if (io_u_plat_last) {
 906                 for (k = sum = 0; k < stride; k++)
 907                         sum += io_u_plat[j + k] - io_u_plat_last[j + k];
 908         } else {
 909                 for (k = sum = 0; k < stride; k++)
 910                         sum += io_u_plat[j + k];
 911         }
 912
 913         return sum;
 914 }
 915
 916 static void flush_hist_samples(FILE *f, int hist_coarseness, void *samples,
 917                                uint64_t sample_size)
 918 {
 919         struct io_sample *s;
 920         int log_offset;
 921         uint64_t i, j, nr_samples;
 922         struct io_u_plat_entry *entry, *entry_before;
 923         uint64_t *io_u_plat;
 924         uint64_t *io_u_plat_before;
 925
 926         int stride = 1 << hist_coarseness;
 927
 928         if (!sample_size)
 929                 return;
 930
 931         s = __get_sample(samples, 0, 0);
 932         log_offset = (s->__ddir & LOG_OFFSET_SAMPLE_BIT) != 0;
 933
 934         nr_samples = sample_size / __log_entry_sz(log_offset);
 935
 936         for (i = 0; i < nr_samples; i++) {
 937                 s = __get_sample(samples, log_offset, i);
 938
 939                 entry = s->data.plat_entry;
 940                 io_u_plat = entry->io_u_plat;
 941
 942                 entry_before = flist_first_entry(&entry->list, struct io_u_plat_entry, list);
 943                 io_u_plat_before = entry_before->io_u_plat;
 944
 945                 fprintf(f, "%lu, %u, %llu, ", (unsigned long) s->time,
 946                                                 io_sample_ddir(s), (unsigned long long) s->bs);
 947                 for (j = 0; j < FIO_IO_U_PLAT_NR - stride; j += stride) {
 948                         fprintf(f, "%llu, ", (unsigned long long)
 949                                 hist_sum(j, stride, io_u_plat, io_u_plat_before));
 950                 }
 951                 fprintf(f, "%llu\n", (unsigned long long)
 952                         hist_sum(FIO_IO_U_PLAT_NR - stride, stride, io_u_plat,
 953                                         io_u_plat_before));
 954
 955                 flist_del(&entry_before->list);
 956                 free(entry_before);
 957         }
 958 }
 959
 960 void flush_samples(FILE *f, void *samples, uint64_t sample_size)
 961 {
 962         struct io_sample *s;
 963         int log_offset, log_prio;
 964         uint64_t i, nr_samples;
 965         unsigned int prio_val;
 966         const char *fmt;
 967
 968         if (!sample_size)
 969                 return;
 970
 971         s = __get_sample(samples, 0, 0);
 972         log_offset = (s->__ddir & LOG_OFFSET_SAMPLE_BIT) != 0;
 973         log_prio = (s->__ddir & LOG_PRIO_SAMPLE_BIT) != 0;
 974
 975         if (log_offset) {
 976                 if (log_prio)
 977                         fmt = "%lu, %" PRId64 ", %u, %llu, %llu, 0x%04x\n";
 978                 else
 979                         fmt = "%lu, %" PRId64 ", %u, %llu, %llu, %u\n";
 980         } else {
 981                 if (log_prio)
 982                         fmt = "%lu, %" PRId64 ", %u, %llu, 0x%04x\n";
 983                 else
 984                         fmt = "%lu, %" PRId64 ", %u, %llu, %u\n";
 985         }
 986
 987         nr_samples = sample_size / __log_entry_sz(log_offset);
 988
 989         for (i = 0; i < nr_samples; i++) {
 990                 s = __get_sample(samples, log_offset, i);
 991
 992                 if (log_prio)
 993                         prio_val = s->priority;
 994                 else
 995                         prio_val = ioprio_value_is_class_rt(s->priority);
 996
 997                 if (!log_offset) {
 998                         fprintf(f, fmt,
 999                                 (unsigned long) s->time,
1000                                 s->data.val,
1001                                 io_sample_ddir(s), (unsigned long long) s->bs,
1002                                 prio_val);
1003                 } else {
1004                         struct io_sample_offset *so = (void *) s;
1005
1006                         fprintf(f, fmt,
1007                                 (unsigned long) s->time,
1008                                 s->data.val,
1009                                 io_sample_ddir(s), (unsigned long long) s->bs,
1010                                 (unsigned long long) so->offset,
1011                                 prio_val);
1012                 }
1013         }
1014 }
1015
1016 #ifdef CONFIG_ZLIB
1017
1018 struct iolog_flush_data {
1019         struct workqueue_work work;
1020         struct io_log *log;
1021         void *samples;
1022         uint32_t nr_samples;
1023         bool free;
1024 };
1025
1026 #define GZ_CHUNK        131072
1027
1028 static struct iolog_compress *get_new_chunk(unsigned int seq)
1029 {
1030         struct iolog_compress *c;
1031
1032         c = malloc(sizeof(*c));
1033         INIT_FLIST_HEAD(&c->list);
1034         c->buf = malloc(GZ_CHUNK);
1035         c->len = 0;
1036         c->seq = seq;
1037         return c;
1038 }
1039
1040 static void free_chunk(struct iolog_compress *ic)
1041 {
1042         free(ic->buf);
1043         free(ic);
1044 }
1045
1046 static int z_stream_init(z_stream *stream, int gz_hdr)
1047 {
1048         int wbits = 15;
1049
1050         memset(stream, 0, sizeof(*stream));
1051         stream->zalloc = Z_NULL;
1052         stream->zfree = Z_NULL;
1053         stream->opaque = Z_NULL;
1054         stream->next_in = Z_NULL;
1055
1056         /*
1057          * zlib magic - add 32 for auto-detection of gz header or not,
1058          * if we decide to store files in a gzip friendly format.
1059          */
1060         if (gz_hdr)
1061                 wbits += 32;
1062
1063         if (inflateInit2(stream, wbits) != Z_OK)
1064                 return 1;
1065
1066         return 0;
1067 }
1068
1069 struct inflate_chunk_iter {
1070         unsigned int seq;
1071         int err;
1072         void *buf;
1073         size_t buf_size;
1074         size_t buf_used;
1075         size_t chunk_sz;
1076 };
1077
1078 static void finish_chunk(z_stream *stream, FILE *f,
1079                          struct inflate_chunk_iter *iter)
1080 {
1081         int ret;
1082
1083         ret = inflateEnd(stream);
1084         if (ret != Z_OK)
1085                 log_err("fio: failed to end log inflation seq %d (%d)\n",
1086                                 iter->seq, ret);
1087
1088         flush_samples(f, iter->buf, iter->buf_used);
1089         free(iter->buf);
1090         iter->buf = NULL;
1091         iter->buf_size = iter->buf_used = 0;
1092 }
1093
1094 /*
1095  * Iterative chunk inflation. Handles cases where we cross into a new
1096  * sequence, doing flush finish of previous chunk if needed.
1097  */
1098 static size_t inflate_chunk(struct iolog_compress *ic, int gz_hdr, FILE *f,
1099                             z_stream *stream, struct inflate_chunk_iter *iter)
1100 {
1101         size_t ret;
1102
1103         dprint(FD_COMPRESS, "inflate chunk size=%lu, seq=%u\n",
1104                                 (unsigned long) ic->len, ic->seq);
1105
1106         if (ic->seq != iter->seq) {
1107                 if (iter->seq)
1108                         finish_chunk(stream, f, iter);
1109
1110                 z_stream_init(stream, gz_hdr);
1111                 iter->seq = ic->seq;
1112         }
1113
1114         stream->avail_in = ic->len;
1115         stream->next_in = ic->buf;
1116
1117         if (!iter->buf_size) {
1118                 iter->buf_size = iter->chunk_sz;
1119                 iter->buf = malloc(iter->buf_size);
1120         }
1121
1122         while (stream->avail_in) {
1123                 size_t this_out = iter->buf_size - iter->buf_used;
1124                 int err;
1125
1126                 stream->avail_out = this_out;
1127                 stream->next_out = iter->buf + iter->buf_used;
1128
1129                 err = inflate(stream, Z_NO_FLUSH);
1130                 if (err < 0) {
1131                         log_err("fio: failed inflating log: %d\n", err);
1132                         iter->err = err;
1133                         break;
1134                 }
1135
1136                 iter->buf_used += this_out - stream->avail_out;
1137
1138                 if (!stream->avail_out) {
1139                         iter->buf_size += iter->chunk_sz;
1140                         iter->buf = realloc(iter->buf, iter->buf_size);
1141                         continue;
1142                 }
1143
1144                 if (err == Z_STREAM_END)
1145                         break;
1146         }
1147
1148         ret = (void *) stream->next_in - ic->buf;
1149
1150         dprint(FD_COMPRESS, "inflated to size=%lu\n", (unsigned long) iter->buf_size);
1151
1152         return ret;
1153 }
1154
1155 /*
1156  * Inflate stored compressed chunks, or write them directly to the log
1157  * file if so instructed.
1158  */
1159 static int inflate_gz_chunks(struct io_log *log, FILE *f)
1160 {
1161         struct inflate_chunk_iter iter = { .chunk_sz = log->log_gz, };
1162         z_stream stream;
1163
1164         while (!flist_empty(&log->chunk_list)) {
1165                 struct iolog_compress *ic;
1166
1167                 ic = flist_first_entry(&log->chunk_list, struct iolog_compress, list);
1168                 flist_del(&ic->list);
1169
1170                 if (log->log_gz_store) {
1171                         size_t ret;
1172
1173                         dprint(FD_COMPRESS, "log write chunk size=%lu, "
1174                                 "seq=%u\n", (unsigned long) ic->len, ic->seq);
1175
1176                         ret = fwrite(ic->buf, ic->len, 1, f);
1177                         if (ret != 1 || ferror(f)) {
1178                                 iter.err = errno;
1179                                 log_err("fio: error writing compressed log\n");
1180                         }
1181                 } else
1182                         inflate_chunk(ic, log->log_gz_store, f, &stream, &iter);
1183
1184                 free_chunk(ic);
1185         }
1186
1187         if (iter.seq) {
1188                 finish_chunk(&stream, f, &iter);
1189                 free(iter.buf);
1190         }
1191
1192         return iter.err;
1193 }
1194
1195 /*
1196  * Open compressed log file and decompress the stored chunks and
1197  * write them to stdout. The chunks are stored sequentially in the
1198  * file, so we iterate over them and do them one-by-one.
1199  */
1200 int iolog_file_inflate(const char *file)
1201 {
1202         struct inflate_chunk_iter iter = { .chunk_sz = 64 * 1024 * 1024, };
1203         struct iolog_compress ic;
1204         z_stream stream;
1205         struct stat sb;
1206         size_t ret;
1207         size_t total;
1208         void *buf;
1209         FILE *f;
1210
1211         f = fopen(file, "r");
1212         if (!f) {
1213                 perror("fopen");
1214                 return 1;
1215         }
1216
1217         if (stat(file, &sb) < 0) {
1218                 fclose(f);
1219                 perror("stat");
1220                 return 1;
1221         }
1222
1223         ic.buf = buf = malloc(sb.st_size);
1224         ic.len = sb.st_size;
1225         ic.seq = 1;
1226
1227         ret = fread(ic.buf, ic.len, 1, f);
1228         if (ret == 0 && ferror(f)) {
1229                 perror("fread");
1230                 fclose(f);
1231                 free(buf);
1232                 return 1;
1233         } else if (ferror(f) || (!feof(f) && ret != 1)) {
1234                 log_err("fio: short read on reading log\n");
1235                 fclose(f);
1236                 free(buf);
1237                 return 1;
1238         }
1239
1240         fclose(f);
1241
1242         /*
1243          * Each chunk will return Z_STREAM_END. We don't know how many
1244          * chunks are in the file, so we just keep looping and incrementing
1245          * the sequence number until we have consumed the whole compressed
1246          * file.
1247          */
1248         total = ic.len;
1249         do {
1250                 size_t iret;
1251
1252                 iret = inflate_chunk(&ic,  1, stdout, &stream, &iter);
1253                 total -= iret;
1254                 if (!total)
1255                         break;
1256                 if (iter.err)
1257                         break;
1258
1259                 ic.seq++;
1260                 ic.len -= iret;
1261                 ic.buf += iret;
1262         } while (1);
1263
1264         if (iter.seq) {
1265                 finish_chunk(&stream, stdout, &iter);
1266                 free(iter.buf);
1267         }
1268
1269         free(buf);
1270         return iter.err;
1271 }
1272
1273 #else
1274
1275 static int inflate_gz_chunks(struct io_log *log, FILE *f)
1276 {
1277         return 0;
1278 }
1279
1280 int iolog_file_inflate(const char *file)
1281 {
1282         log_err("fio: log inflation not possible without zlib\n");
1283         return 1;
1284 }
1285
1286 #endif
1287
1288 void flush_log(struct io_log *log, bool do_append)
1289 {
1290         void *buf;
1291         FILE *f;
1292
1293         if (!do_append)
1294                 f = fopen(log->filename, "w");
1295         else
1296                 f = fopen(log->filename, "a");
1297         if (!f) {
1298                 perror("fopen log");
1299                 return;
1300         }
1301
1302         buf = set_file_buffer(f);
1303
1304         inflate_gz_chunks(log, f);
1305
1306         while (!flist_empty(&log->io_logs)) {
1307                 struct io_logs *cur_log;
1308
1309                 cur_log = flist_first_entry(&log->io_logs, struct io_logs, list);
1310                 flist_del_init(&cur_log->list);
1311
1312                 if (log->td && log == log->td->clat_hist_log)
1313                         flush_hist_samples(f, log->hist_coarseness, cur_log->log,
1314                                            log_sample_sz(log, cur_log));
1315                 else
1316                         flush_samples(f, cur_log->log, log_sample_sz(log, cur_log));
1317
1318                 sfree(cur_log);
1319         }
1320
1321         fclose(f);
1322         clear_file_buffer(buf);
1323 }
1324
1325 static int finish_log(struct thread_data *td, struct io_log *log, int trylock)
1326 {
1327         if (td->flags & TD_F_COMPRESS_LOG)
1328                 iolog_flush(log);
1329
1330         if (trylock) {
1331                 if (fio_trylock_file(log->filename))
1332                         return 1;
1333         } else
1334                 fio_lock_file(log->filename);
1335
1336         if (td->client_type == FIO_CLIENT_TYPE_GUI || is_backend)
1337                 fio_send_iolog(td, log, log->filename);
1338         else
1339                 flush_log(log, !td->o.per_job_logs);
1340
1341         fio_unlock_file(log->filename);
1342         free_log(log);
1343         return 0;
1344 }
1345
1346 size_t log_chunk_sizes(struct io_log *log)
1347 {
1348         struct flist_head *entry;
1349         size_t ret;
1350
1351         if (flist_empty(&log->chunk_list))
1352                 return 0;
1353
1354         ret = 0;
1355         pthread_mutex_lock(&log->chunk_lock);
1356         flist_for_each(entry, &log->chunk_list) {
1357                 struct iolog_compress *c;
1358
1359                 c = flist_entry(entry, struct iolog_compress, list);
1360                 ret += c->len;
1361         }
1362         pthread_mutex_unlock(&log->chunk_lock);
1363         return ret;
1364 }
1365
1366 #ifdef CONFIG_ZLIB
1367
1368 static void iolog_put_deferred(struct io_log *log, void *ptr)
1369 {
1370         if (!ptr)
1371                 return;
1372
1373         pthread_mutex_lock(&log->deferred_free_lock);
1374         if (log->deferred < IOLOG_MAX_DEFER) {
1375                 log->deferred_items[log->deferred] = ptr;
1376                 log->deferred++;
1377         } else if (!fio_did_warn(FIO_WARN_IOLOG_DROP))
1378                 log_err("fio: had to drop log entry free\n");
1379         pthread_mutex_unlock(&log->deferred_free_lock);
1380 }
1381
1382 static void iolog_free_deferred(struct io_log *log)
1383 {
1384         int i;
1385
1386         if (!log->deferred)
1387                 return;
1388
1389         pthread_mutex_lock(&log->deferred_free_lock);
1390
1391         for (i = 0; i < log->deferred; i++) {
1392                 free(log->deferred_items[i]);
1393                 log->deferred_items[i] = NULL;
1394         }
1395
1396         log->deferred = 0;
1397         pthread_mutex_unlock(&log->deferred_free_lock);
1398 }
1399
1400 static int gz_work(struct iolog_flush_data *data)
1401 {
1402         struct iolog_compress *c = NULL;
1403         struct flist_head list;
1404         unsigned int seq;
1405         z_stream stream;
1406         size_t total = 0;
1407         int ret;
1408
1409         INIT_FLIST_HEAD(&list);
1410
1411         memset(&stream, 0, sizeof(stream));
1412         stream.zalloc = Z_NULL;
1413         stream.zfree = Z_NULL;
1414         stream.opaque = Z_NULL;
1415
1416         ret = deflateInit(&stream, Z_DEFAULT_COMPRESSION);
1417         if (ret != Z_OK) {
1418                 log_err("fio: failed to init gz stream\n");
1419                 goto err;
1420         }
1421
1422         seq = ++data->log->chunk_seq;
1423
1424         stream.next_in = (void *) data->samples;
1425         stream.avail_in = data->nr_samples * log_entry_sz(data->log);
1426
1427         dprint(FD_COMPRESS, "deflate input size=%lu, seq=%u, log=%s\n",
1428                                 (unsigned long) stream.avail_in, seq,
1429                                 data->log->filename);
1430         do {
1431                 if (c)
1432                         dprint(FD_COMPRESS, "seq=%d, chunk=%lu\n", seq,
1433                                 (unsigned long) c->len);
1434                 c = get_new_chunk(seq);
1435                 stream.avail_out = GZ_CHUNK;
1436                 stream.next_out = c->buf;
1437                 ret = deflate(&stream, Z_NO_FLUSH);
1438                 if (ret < 0) {
1439                         log_err("fio: deflate log (%d)\n", ret);
1440                         free_chunk(c);
1441                         goto err;
1442                 }
1443
1444                 c->len = GZ_CHUNK - stream.avail_out;
1445                 flist_add_tail(&c->list, &list);
1446                 total += c->len;
1447         } while (stream.avail_in);
1448
1449         stream.next_out = c->buf + c->len;
1450         stream.avail_out = GZ_CHUNK - c->len;
1451
1452         ret = deflate(&stream, Z_FINISH);
1453         if (ret < 0) {
1454                 /*
1455                  * Z_BUF_ERROR is special, it just means we need more
1456                  * output space. We'll handle that below. Treat any other
1457                  * error as fatal.
1458                  */
1459                 if (ret != Z_BUF_ERROR) {
1460                         log_err("fio: deflate log (%d)\n", ret);
1461                         flist_del(&c->list);
1462                         free_chunk(c);
1463                         goto err;
1464                 }
1465         }
1466
1467         total -= c->len;
1468         c->len = GZ_CHUNK - stream.avail_out;
1469         total += c->len;
1470         dprint(FD_COMPRESS, "seq=%d, chunk=%lu\n", seq, (unsigned long) c->len);
1471
1472         if (ret != Z_STREAM_END) {
1473                 do {
1474                         c = get_new_chunk(seq);
1475                         stream.avail_out = GZ_CHUNK;
1476                         stream.next_out = c->buf;
1477                         ret = deflate(&stream, Z_FINISH);
1478                         c->len = GZ_CHUNK - stream.avail_out;
1479                         total += c->len;
1480                         flist_add_tail(&c->list, &list);
1481                         dprint(FD_COMPRESS, "seq=%d, chunk=%lu\n", seq,
1482                                 (unsigned long) c->len);
1483                 } while (ret != Z_STREAM_END);
1484         }
1485
1486         dprint(FD_COMPRESS, "deflated to size=%lu\n", (unsigned long) total);
1487
1488         ret = deflateEnd(&stream);
1489         if (ret != Z_OK)
1490                 log_err("fio: deflateEnd %d\n", ret);
1491
1492         iolog_put_deferred(data->log, data->samples);
1493
1494         if (!flist_empty(&list)) {
1495                 pthread_mutex_lock(&data->log->chunk_lock);
1496                 flist_splice_tail(&list, &data->log->chunk_list);
1497                 pthread_mutex_unlock(&data->log->chunk_lock);
1498         }
1499
1500         ret = 0;
1501 done:
1502         if (data->free)
1503                 sfree(data);
1504         return ret;
1505 err:
1506         while (!flist_empty(&list)) {
1507                 c = flist_first_entry(list.next, struct iolog_compress, list);
1508                 flist_del(&c->list);
1509                 free_chunk(c);
1510         }
1511         ret = 1;
1512         goto done;
1513 }
1514
1515 /*
1516  * Invoked from our compress helper thread, when logging would have exceeded
1517  * the specified memory limitation. Compresses the previously stored
1518  * entries.
1519  */
1520 static int gz_work_async(struct submit_worker *sw, struct workqueue_work *work)
1521 {
1522         return gz_work(container_of(work, struct iolog_flush_data, work));
1523 }
1524
1525 static int gz_init_worker(struct submit_worker *sw)
1526 {
1527         struct thread_data *td = sw->wq->td;
1528
1529         if (!fio_option_is_set(&td->o, log_gz_cpumask))
1530                 return 0;
1531
1532         if (fio_setaffinity(gettid(), td->o.log_gz_cpumask) == -1) {
1533                 log_err("gz: failed to set CPU affinity\n");
1534                 return 1;
1535         }
1536
1537         return 0;
1538 }
1539
1540 static struct workqueue_ops log_compress_wq_ops = {
1541         .fn             = gz_work_async,
1542         .init_worker_fn = gz_init_worker,
1543         .nice           = 1,
1544 };
1545
1546 int iolog_compress_init(struct thread_data *td, struct sk_out *sk_out)
1547 {
1548         if (!(td->flags & TD_F_COMPRESS_LOG))
1549                 return 0;
1550
1551         workqueue_init(td, &td->log_compress_wq, &log_compress_wq_ops, 1, sk_out);
1552         return 0;
1553 }
1554
1555 void iolog_compress_exit(struct thread_data *td)
1556 {
1557         if (!(td->flags & TD_F_COMPRESS_LOG))
1558                 return;
1559
1560         workqueue_exit(&td->log_compress_wq);
1561 }
1562
1563 /*
1564  * Queue work item to compress the existing log entries. We reset the
1565  * current log to a small size, and reference the existing log in the
1566  * data that we queue for compression. Once compression has been done,
1567  * this old log is freed. If called with finish == true, will not return
1568  * until the log compression has completed, and will flush all previous
1569  * logs too
1570  */
1571 static int iolog_flush(struct io_log *log)
1572 {
1573         struct iolog_flush_data *data;
1574
1575         data = malloc(sizeof(*data));
1576         if (!data)
1577                 return 1;
1578
1579         data->log = log;
1580         data->free = false;
1581
1582         while (!flist_empty(&log->io_logs)) {
1583                 struct io_logs *cur_log;
1584
1585                 cur_log = flist_first_entry(&log->io_logs, struct io_logs, list);
1586                 flist_del_init(&cur_log->list);
1587
1588                 data->samples = cur_log->log;
1589                 data->nr_samples = cur_log->nr_samples;
1590
1591                 sfree(cur_log);
1592
1593                 gz_work(data);
1594         }
1595
1596         free(data);
1597         return 0;
1598 }
1599
1600 int iolog_cur_flush(struct io_log *log, struct io_logs *cur_log)
1601 {
1602         struct iolog_flush_data *data;
1603
1604         data = smalloc(sizeof(*data));
1605         if (!data)
1606                 return 1;
1607
1608         data->log = log;
1609
1610         data->samples = cur_log->log;
1611         data->nr_samples = cur_log->nr_samples;
1612         data->free = true;
1613
1614         cur_log->nr_samples = cur_log->max_samples = 0;
1615         cur_log->log = NULL;
1616
1617         workqueue_enqueue(&log->td->log_compress_wq, &data->work);
1618
1619         iolog_free_deferred(log);
1620
1621         return 0;
1622 }
1623 #else
1624
1625 static int iolog_flush(struct io_log *log)
1626 {
1627         return 1;
1628 }
1629
1630 int iolog_cur_flush(struct io_log *log, struct io_logs *cur_log)
1631 {
1632         return 1;
1633 }
1634
1635 int iolog_compress_init(struct thread_data *td, struct sk_out *sk_out)
1636 {
1637         return 0;
1638 }
1639
1640 void iolog_compress_exit(struct thread_data *td)
1641 {
1642 }
1643
1644 #endif
1645
1646 struct io_logs *iolog_cur_log(struct io_log *log)
1647 {
1648         if (flist_empty(&log->io_logs))
1649                 return NULL;
1650
1651         return flist_last_entry(&log->io_logs, struct io_logs, list);
1652 }
1653
1654 uint64_t iolog_nr_samples(struct io_log *iolog)
1655 {
1656         struct flist_head *entry;
1657         uint64_t ret = 0;
1658
1659         flist_for_each(entry, &iolog->io_logs) {
1660                 struct io_logs *cur_log;
1661
1662                 cur_log = flist_entry(entry, struct io_logs, list);
1663                 ret += cur_log->nr_samples;
1664         }
1665
1666         return ret;
1667 }
1668
1669 static int __write_log(struct thread_data *td, struct io_log *log, int try)
1670 {
1671         if (log)
1672                 return finish_log(td, log, try);
1673
1674         return 0;
1675 }
1676
1677 static int write_iops_log(struct thread_data *td, int try, bool unit_log)
1678 {
1679         int ret;
1680
1681         if (per_unit_log(td->iops_log) != unit_log)
1682                 return 0;
1683
1684         ret = __write_log(td, td->iops_log, try);
1685         if (!ret)
1686                 td->iops_log = NULL;
1687
1688         return ret;
1689 }
1690
1691 static int write_slat_log(struct thread_data *td, int try, bool unit_log)
1692 {
1693         int ret;
1694
1695         if (!unit_log)
1696                 return 0;
1697
1698         ret = __write_log(td, td->slat_log, try);
1699         if (!ret)
1700                 td->slat_log = NULL;
1701
1702         return ret;
1703 }
1704
1705 static int write_clat_log(struct thread_data *td, int try, bool unit_log)
1706 {
1707         int ret;
1708
1709         if (!unit_log)
1710                 return 0;
1711
1712         ret = __write_log(td, td->clat_log, try);
1713         if (!ret)
1714                 td->clat_log = NULL;
1715
1716         return ret;
1717 }
1718
1719 static int write_clat_hist_log(struct thread_data *td, int try, bool unit_log)
1720 {
1721         int ret;
1722
1723         if (!unit_log)
1724                 return 0;
1725
1726         ret = __write_log(td, td->clat_hist_log, try);
1727         if (!ret)
1728                 td->clat_hist_log = NULL;
1729
1730         return ret;
1731 }
1732
1733 static int write_lat_log(struct thread_data *td, int try, bool unit_log)
1734 {
1735         int ret;
1736
1737         if (!unit_log)
1738                 return 0;
1739
1740         ret = __write_log(td, td->lat_log, try);
1741         if (!ret)
1742                 td->lat_log = NULL;
1743
1744         return ret;
1745 }
1746
1747 static int write_bandw_log(struct thread_data *td, int try, bool unit_log)
1748 {
1749         int ret;
1750
1751         if (per_unit_log(td->bw_log) != unit_log)
1752                 return 0;
1753
1754         ret = __write_log(td, td->bw_log, try);
1755         if (!ret)
1756                 td->bw_log = NULL;
1757
1758         return ret;
1759 }
1760
1761 enum {
1762         BW_LOG_MASK     = 1,
1763         LAT_LOG_MASK    = 2,
1764         SLAT_LOG_MASK   = 4,
1765         CLAT_LOG_MASK   = 8,
1766         IOPS_LOG_MASK   = 16,
1767         CLAT_HIST_LOG_MASK = 32,
1768
1769         ALL_LOG_NR      = 6,
1770 };
1771
1772 struct log_type {
1773         unsigned int mask;
1774         int (*fn)(struct thread_data *, int, bool);
1775 };
1776
1777 static struct log_type log_types[] = {
1778         {
1779                 .mask   = BW_LOG_MASK,
1780                 .fn     = write_bandw_log,
1781         },
1782         {
1783                 .mask   = LAT_LOG_MASK,
1784                 .fn     = write_lat_log,
1785         },
1786         {
1787                 .mask   = SLAT_LOG_MASK,
1788                 .fn     = write_slat_log,
1789         },
1790         {
1791                 .mask   = CLAT_LOG_MASK,
1792                 .fn     = write_clat_log,
1793         },
1794         {
1795                 .mask   = IOPS_LOG_MASK,
1796                 .fn     = write_iops_log,
1797         },
1798         {
1799                 .mask   = CLAT_HIST_LOG_MASK,
1800                 .fn     = write_clat_hist_log,
1801         }
1802 };
1803
1804 void td_writeout_logs(struct thread_data *td, bool unit_logs)
1805 {
1806         unsigned int log_mask = 0;
1807         unsigned int log_left = ALL_LOG_NR;
1808         int old_state, i;
1809
1810         old_state = td_bump_runstate(td, TD_FINISHING);
1811
1812         finalize_logs(td, unit_logs);
1813
1814         while (log_left) {
1815                 int prev_log_left = log_left;
1816
1817                 for (i = 0; i < ALL_LOG_NR && log_left; i++) {
1818                         struct log_type *lt = &log_types[i];
1819                         int ret;
1820
1821                         if (!(log_mask & lt->mask)) {
1822                                 ret = lt->fn(td, log_left != 1, unit_logs);
1823                                 if (!ret) {
1824                                         log_left--;
1825                                         log_mask |= lt->mask;
1826                                 }
1827                         }
1828                 }
1829
1830                 if (prev_log_left == log_left)
1831                         usleep(5000);
1832         }
1833
1834         td_restore_runstate(td, old_state);
1835 }
1836
1837 void fio_writeout_logs(bool unit_logs)
1838 {
1839         struct thread_data *td;
1840         int i;
1841
1842         for_each_td(td, i)
1843                 td_writeout_logs(td, unit_logs);
1844 }