iolog.c

   1 /*
   2  * Code related to writing an iolog of what a thread is doing, and to
   3  * later read that back and replay
   4  */
   5 #include <stdio.h>
   6 #include <stdlib.h>
   7 #include <assert.h>
   8 #include <sys/types.h>
   9 #include <sys/stat.h>
  10 #include <unistd.h>
  11 #ifdef CONFIG_ZLIB
  12 #include <zlib.h>
  13 #endif
  14
  15 #include "flist.h"
  16 #include "fio.h"
  17 #include "trim.h"
  18 #include "filelock.h"
  19 #include "smalloc.h"
  20 #include "blktrace.h"
  21 #include "pshared.h"
  22 #include "lib/roundup.h"
  23
  24 #include <netinet/in.h>
  25 #include <netinet/tcp.h>
  26 #include <arpa/inet.h>
  27 #include <sys/stat.h>
  28 #include <sys/socket.h>
  29 #include <sys/un.h>
  30
  31 static int iolog_flush(struct io_log *log);
  32
  33 static const char iolog_ver2[] = "fio version 2 iolog";
  34 static const char iolog_ver3[] = "fio version 3 iolog";
  35
  36 void queue_io_piece(struct thread_data *td, struct io_piece *ipo)
  37 {
  38         flist_add_tail(&ipo->list, &td->io_log_list);
  39         td->total_io_size += ipo->len;
  40 }
  41
  42 void log_io_u(const struct thread_data *td, const struct io_u *io_u)
  43 {
  44         struct timespec now;
  45
  46         if (!td->o.write_iolog_file)
  47                 return;
  48
  49         fio_gettime(&now, NULL);
  50         fprintf(td->iolog_f, "%llu %s %s %llu %llu\n",
  51                 (unsigned long long) utime_since_now(&td->io_log_start_time),
  52                 io_u->file->file_name, io_ddir_name(io_u->ddir), io_u->offset,
  53                 io_u->buflen);
  54
  55 }
  56
  57 void log_file(struct thread_data *td, struct fio_file *f,
  58               enum file_log_act what)
  59 {
  60         const char *act[] = { "add", "open", "close" };
  61         struct timespec now;
  62
  63         assert(what < 3);
  64
  65         if (!td->o.write_iolog_file)
  66                 return;
  67
  68
  69         /*
  70          * this happens on the pre-open/close done before the job starts
  71          */
  72         if (!td->iolog_f)
  73                 return;
  74
  75         fio_gettime(&now, NULL);
  76         fprintf(td->iolog_f, "%llu %s %s\n",
  77                 (unsigned long long) utime_since_now(&td->io_log_start_time),
  78                 f->file_name, act[what]);
  79 }
  80
  81 static void iolog_delay(struct thread_data *td, unsigned long delay)
  82 {
  83         uint64_t usec = utime_since_now(&td->last_issue);
  84         unsigned long orig_delay = delay;
  85         struct timespec ts;
  86         int ret = 0;
  87
  88         if (delay < td->time_offset) {
  89                 td->time_offset = 0;
  90                 return;
  91         }
  92
  93         delay -= td->time_offset;
  94         if (delay < usec)
  95                 return;
  96
  97         delay -= usec;
  98
  99         fio_gettime(&ts, NULL);
 100
 101         while (delay && !td->terminate) {
 102                 ret = io_u_queued_complete(td, 0);
 103                 if (ret < 0)
 104                         td_verror(td, -ret, "io_u_queued_complete");
 105                 if (td->flags & TD_F_REGROW_LOGS)
 106                         regrow_logs(td);
 107                 if (utime_since_now(&ts) > delay)
 108                         break;
 109         }
 110
 111         usec = utime_since_now(&ts);
 112         if (usec > orig_delay)
 113                 td->time_offset = usec - orig_delay;
 114         else
 115                 td->time_offset = 0;
 116 }
 117
 118 static int ipo_special(struct thread_data *td, struct io_piece *ipo)
 119 {
 120         struct fio_file *f;
 121         int ret;
 122
 123         /*
 124          * Not a special ipo
 125          */
 126         if (ipo->ddir != DDIR_INVAL)
 127                 return 0;
 128
 129         f = td->files[ipo->fileno];
 130
 131         if (ipo->delay)
 132                 iolog_delay(td, ipo->delay);
 133         if (fio_fill_issue_time(td))
 134                 fio_gettime(&td->last_issue, NULL);
 135         switch (ipo->file_action) {
 136         case FIO_LOG_OPEN_FILE:
 137                 if (td->o.replay_redirect && fio_file_open(f)) {
 138                         dprint(FD_FILE, "iolog: ignoring re-open of file %s\n",
 139                                         f->file_name);
 140                         break;
 141                 }
 142                 ret = td_io_open_file(td, f);
 143                 if (!ret)
 144                         break;
 145                 td_verror(td, ret, "iolog open file");
 146                 return -1;
 147         case FIO_LOG_CLOSE_FILE:
 148                 td_io_close_file(td, f);
 149                 break;
 150         case FIO_LOG_UNLINK_FILE:
 151                 td_io_unlink_file(td, f);
 152                 break;
 153         case FIO_LOG_ADD_FILE:
 154                 /*
 155                  * Nothing to do
 156                  */
 157                 break;
 158         default:
 159                 log_err("fio: bad file action %d\n", ipo->file_action);
 160                 break;
 161         }
 162
 163         return 1;
 164 }
 165
 166 static bool read_iolog(struct thread_data *td);
 167
 168 unsigned long long delay_since_ttime(const struct thread_data *td,
 169                unsigned long long time)
 170 {
 171         double tmp;
 172         double scale;
 173         const unsigned long long *last_ttime = &td->io_log_last_ttime;
 174
 175         if (!*last_ttime || td->o.no_stall || time < *last_ttime)
 176                 return 0;
 177         else if (td->o.replay_time_scale == 100)
 178                 return time - *last_ttime;
 179
 180
 181         scale = (double) 100.0 / (double) td->o.replay_time_scale;
 182         tmp = time - *last_ttime;
 183         return tmp * scale;
 184 }
 185
 186 int read_iolog_get(struct thread_data *td, struct io_u *io_u)
 187 {
 188         struct io_piece *ipo;
 189         unsigned long elapsed;
 190
 191         while (!flist_empty(&td->io_log_list)) {
 192                 int ret;
 193
 194                 if (td->o.read_iolog_chunked) {
 195                         if (td->io_log_checkmark == td->io_log_current) {
 196                                 if (td->io_log_blktrace) {
 197                                         if (!read_blktrace(td))
 198                                                 return 1;
 199                                 } else {
 200                                         if (!read_iolog(td))
 201                                                 return 1;
 202                                 }
 203                         }
 204                         td->io_log_current--;
 205                 }
 206                 ipo = flist_first_entry(&td->io_log_list, struct io_piece, list);
 207                 flist_del(&ipo->list);
 208                 remove_trim_entry(td, ipo);
 209
 210                 ret = ipo_special(td, ipo);
 211                 if (ret < 0) {
 212                         free(ipo);
 213                         break;
 214                 } else if (ret > 0) {
 215                         free(ipo);
 216                         continue;
 217                 }
 218
 219                 io_u->ddir = ipo->ddir;
 220                 if (ipo->ddir != DDIR_WAIT) {
 221                         io_u->offset = ipo->offset;
 222                         io_u->verify_offset = ipo->offset;
 223                         io_u->buflen = ipo->len;
 224                         io_u->file = td->files[ipo->fileno];
 225                         get_file(io_u->file);
 226                         dprint(FD_IO, "iolog: get %llu/%llu/%s\n", io_u->offset,
 227                                                 io_u->buflen, io_u->file->file_name);
 228                         if (ipo->delay)
 229                                 iolog_delay(td, ipo->delay);
 230                 } else {
 231                         elapsed = mtime_since_genesis();
 232                         if (ipo->delay > elapsed)
 233                                 usec_sleep(td, (ipo->delay - elapsed) * 1000);
 234                 }
 235
 236                 free(ipo);
 237
 238                 if (io_u->ddir != DDIR_WAIT)
 239                         return 0;
 240         }
 241
 242         td->done = 1;
 243         return 1;
 244 }
 245
 246 void prune_io_piece_log(struct thread_data *td)
 247 {
 248         struct io_piece *ipo;
 249         struct fio_rb_node *n;
 250
 251         while ((n = rb_first(&td->io_hist_tree)) != NULL) {
 252                 ipo = rb_entry(n, struct io_piece, rb_node);
 253                 rb_erase(n, &td->io_hist_tree);
 254                 remove_trim_entry(td, ipo);
 255                 td->io_hist_len--;
 256                 free(ipo);
 257         }
 258
 259         while (!flist_empty(&td->io_hist_list)) {
 260                 ipo = flist_first_entry(&td->io_hist_list, struct io_piece, list);
 261                 flist_del(&ipo->list);
 262                 remove_trim_entry(td, ipo);
 263                 td->io_hist_len--;
 264                 free(ipo);
 265         }
 266 }
 267
 268 /*
 269  * log a successful write, so we can unwind the log for verify
 270  */
 271 void log_io_piece(struct thread_data *td, struct io_u *io_u)
 272 {
 273         struct fio_rb_node **p, *parent;
 274         struct io_piece *ipo, *__ipo;
 275
 276         ipo = calloc(1, sizeof(struct io_piece));
 277         init_ipo(ipo);
 278         ipo->file = io_u->file;
 279         ipo->offset = io_u->offset;
 280         ipo->len = io_u->buflen;
 281         ipo->numberio = io_u->numberio;
 282         ipo->flags = IP_F_IN_FLIGHT;
 283
 284         io_u->ipo = ipo;
 285
 286         if (io_u_should_trim(td, io_u)) {
 287                 flist_add_tail(&ipo->trim_list, &td->trim_list);
 288                 td->trim_entries++;
 289         }
 290
 291         /*
 292          * Only sort writes if we don't have a random map in which case we need
 293          * to check for duplicate blocks and drop the old one, which we rely on
 294          * the rb insert/lookup for handling.
 295          */
 296         if (file_randommap(td, ipo->file)) {
 297                 INIT_FLIST_HEAD(&ipo->list);
 298                 flist_add_tail(&ipo->list, &td->io_hist_list);
 299                 ipo->flags |= IP_F_ONLIST;
 300                 td->io_hist_len++;
 301                 return;
 302         }
 303
 304         RB_CLEAR_NODE(&ipo->rb_node);
 305
 306         /*
 307          * Sort the entry into the verification list
 308          */
 309 restart:
 310         p = &td->io_hist_tree.rb_node;
 311         parent = NULL;
 312         while (*p) {
 313                 int overlap = 0;
 314                 parent = *p;
 315
 316                 __ipo = rb_entry(parent, struct io_piece, rb_node);
 317                 if (ipo->file < __ipo->file)
 318                         p = &(*p)->rb_left;
 319                 else if (ipo->file > __ipo->file)
 320                         p = &(*p)->rb_right;
 321                 else if (ipo->offset < __ipo->offset) {
 322                         p = &(*p)->rb_left;
 323                         overlap = ipo->offset + ipo->len > __ipo->offset;
 324                 }
 325                 else if (ipo->offset > __ipo->offset) {
 326                         p = &(*p)->rb_right;
 327                         overlap = __ipo->offset + __ipo->len > ipo->offset;
 328                 }
 329                 else
 330                         overlap = 1;
 331
 332                 if (overlap) {
 333                         dprint(FD_IO, "iolog: overlap %llu/%lu, %llu/%lu\n",
 334                                 __ipo->offset, __ipo->len,
 335                                 ipo->offset, ipo->len);
 336                         td->io_hist_len--;
 337                         rb_erase(parent, &td->io_hist_tree);
 338                         remove_trim_entry(td, __ipo);
 339                         if (!(__ipo->flags & IP_F_IN_FLIGHT))
 340                                 free(__ipo);
 341                         goto restart;
 342                 }
 343         }
 344
 345         rb_link_node(&ipo->rb_node, parent, p);
 346         rb_insert_color(&ipo->rb_node, &td->io_hist_tree);
 347         ipo->flags |= IP_F_ONRB;
 348         td->io_hist_len++;
 349 }
 350
 351 void unlog_io_piece(struct thread_data *td, struct io_u *io_u)
 352 {
 353         struct io_piece *ipo = io_u->ipo;
 354
 355         if (td->ts.nr_block_infos) {
 356                 uint32_t *info = io_u_block_info(td, io_u);
 357                 if (BLOCK_INFO_STATE(*info) < BLOCK_STATE_TRIM_FAILURE) {
 358                         if (io_u->ddir == DDIR_TRIM)
 359                                 *info = BLOCK_INFO_SET_STATE(*info,
 360                                                 BLOCK_STATE_TRIM_FAILURE);
 361                         else if (io_u->ddir == DDIR_WRITE)
 362                                 *info = BLOCK_INFO_SET_STATE(*info,
 363                                                 BLOCK_STATE_WRITE_FAILURE);
 364                 }
 365         }
 366
 367         if (!ipo)
 368                 return;
 369
 370         if (ipo->flags & IP_F_ONRB)
 371                 rb_erase(&ipo->rb_node, &td->io_hist_tree);
 372         else if (ipo->flags & IP_F_ONLIST)
 373                 flist_del(&ipo->list);
 374
 375         free(ipo);
 376         io_u->ipo = NULL;
 377         td->io_hist_len--;
 378 }
 379
 380 void trim_io_piece(const struct io_u *io_u)
 381 {
 382         struct io_piece *ipo = io_u->ipo;
 383
 384         if (!ipo)
 385                 return;
 386
 387         ipo->len = io_u->xfer_buflen - io_u->resid;
 388 }
 389
 390 void write_iolog_close(struct thread_data *td)
 391 {
 392         if (!td->iolog_f)
 393                 return;
 394
 395         fflush(td->iolog_f);
 396         fclose(td->iolog_f);
 397         free(td->iolog_buf);
 398         td->iolog_f = NULL;
 399         td->iolog_buf = NULL;
 400 }
 401
 402 int64_t iolog_items_to_fetch(struct thread_data *td)
 403 {
 404         struct timespec now;
 405         uint64_t elapsed;
 406         uint64_t for_1s;
 407         int64_t items_to_fetch;
 408
 409         if (!td->io_log_highmark)
 410                 return 10;
 411
 412
 413         fio_gettime(&now, NULL);
 414         elapsed = ntime_since(&td->io_log_highmark_time, &now);
 415         if (elapsed) {
 416                 for_1s = (td->io_log_highmark - td->io_log_current) * 1000000000 / elapsed;
 417                 items_to_fetch = for_1s - td->io_log_current;
 418                 if (items_to_fetch < 0)
 419                         items_to_fetch = 0;
 420         } else
 421                 items_to_fetch = 0;
 422
 423         td->io_log_highmark = td->io_log_current + items_to_fetch;
 424         td->io_log_checkmark = (td->io_log_highmark + 1) / 2;
 425         fio_gettime(&td->io_log_highmark_time, NULL);
 426
 427         return items_to_fetch;
 428 }
 429
 430 #define io_act(_td, _r) (((_td)->io_log_version == 3 && (r) == 5) || \
 431                                         ((_td)->io_log_version == 2 && (r) == 4))
 432 #define file_act(_td, _r) (((_td)->io_log_version == 3 && (r) == 3) || \
 433                                         ((_td)->io_log_version == 2 && (r) == 2))
 434
 435 /*
 436  * Read version 2 and 3 iolog data. It is enhanced to include per-file logging,
 437  * syncs, etc.
 438  */
 439 static bool read_iolog(struct thread_data *td)
 440 {
 441         unsigned long long offset;
 442         unsigned int bytes;
 443         unsigned long long delay = 0;
 444         int reads, writes, trims, waits, fileno = 0, file_action = 0; /* stupid gcc */
 445         char *rfname, *fname, *act;
 446         char *str, *p;
 447         enum fio_ddir rw;
 448         bool realloc = false;
 449         int64_t items_to_fetch = 0;
 450         int syncs;
 451
 452         if (td->o.read_iolog_chunked) {
 453                 items_to_fetch = iolog_items_to_fetch(td);
 454                 if (!items_to_fetch)
 455                         return true;
 456         }
 457
 458         /*
 459          * Read in the read iolog and store it, reuse the infrastructure
 460          * for doing verifications.
 461          */
 462         str = malloc(4096);
 463         rfname = fname = malloc(256+16);
 464         act = malloc(256+16);
 465
 466         syncs = reads = writes = trims = waits = 0;
 467         while ((p = fgets(str, 4096, td->io_log_rfile)) != NULL) {
 468                 struct io_piece *ipo;
 469                 int r;
 470                 unsigned long long ttime;
 471
 472                 if (td->io_log_version == 3) {
 473                         r = sscanf(p, "%llu %256s %256s %llu %u", &ttime, rfname, act,
 474                                                         &offset, &bytes);
 475                         delay = delay_since_ttime(td, ttime);
 476                         td->io_log_last_ttime = ttime;
 477                         /*
 478                          * "wait" is not allowed with version 3
 479                          */
 480                         if (!strcmp(act, "wait")) {
 481                                 log_err("iolog: ignoring wait command with"
 482                                         " version 3 for file %s\n", fname);
 483                                 continue;
 484                         }
 485                 } else /* version 2 */
 486                         r = sscanf(p, "%256s %256s %llu %u", rfname, act, &offset, &bytes);
 487
 488                 if (td->o.replay_redirect)
 489                         fname = td->o.replay_redirect;
 490
 491                 if (io_act(td, r)) {
 492                         /*
 493                          * Check action first
 494                          */
 495                         if (!strcmp(act, "wait"))
 496                                 rw = DDIR_WAIT;
 497                         else if (!strcmp(act, "read")) {
 498                                 if (td->o.replay_skip & (1u << DDIR_READ))
 499                                         continue;
 500                                 rw = DDIR_READ;
 501                         } else if (!strcmp(act, "write")) {
 502                                 if (td->o.replay_skip & (1u << DDIR_WRITE))
 503                                         continue;
 504                                 rw = DDIR_WRITE;
 505                         } else if (!strcmp(act, "sync")) {
 506                                 if (td->o.replay_skip & (1u << DDIR_SYNC))
 507                                         continue;
 508                                 rw = DDIR_SYNC;
 509                         } else if (!strcmp(act, "datasync"))
 510                                 rw = DDIR_DATASYNC;
 511                         else if (!strcmp(act, "trim")) {
 512                                 if (td->o.replay_skip & (1u << DDIR_TRIM))
 513                                         continue;
 514                                 rw = DDIR_TRIM;
 515                         } else {
 516                                 log_err("fio: bad iolog file action: %s\n",
 517                                                                         act);
 518                                 continue;
 519                         }
 520                         fileno = get_fileno(td, fname);
 521                 } else if (file_act(td, r)) {
 522                         rw = DDIR_INVAL;
 523                         if (!strcmp(act, "add")) {
 524                                 if (td->o.replay_redirect &&
 525                                     get_fileno(td, fname) != -1) {
 526                                         dprint(FD_FILE, "iolog: ignoring"
 527                                                 " re-add of file %s\n", fname);
 528                                 } else {
 529                                         fileno = add_file(td, fname, td->subjob_number, 1);
 530                                         file_action = FIO_LOG_ADD_FILE;
 531                                 }
 532                         } else if (!strcmp(act, "open")) {
 533                                 fileno = get_fileno(td, fname);
 534                                 file_action = FIO_LOG_OPEN_FILE;
 535                         } else if (!strcmp(act, "close")) {
 536                                 fileno = get_fileno(td, fname);
 537                                 file_action = FIO_LOG_CLOSE_FILE;
 538                         } else {
 539                                 log_err("fio: bad iolog file action: %s\n",
 540                                                                         act);
 541                                 continue;
 542                         }
 543                 } else {
 544                         log_err("bad iolog%d: %s\n", td->io_log_version, p);
 545                         continue;
 546                 }
 547
 548                 if (rw == DDIR_READ)
 549                         reads++;
 550                 else if (rw == DDIR_WRITE) {
 551                         /*
 552                          * Don't add a write for ro mode
 553                          */
 554                         if (read_only)
 555                                 continue;
 556                         writes++;
 557                 } else if (rw == DDIR_TRIM) {
 558                         /*
 559                          * Don't add a trim for ro mode
 560                          */
 561                         if (read_only)
 562                                 continue;
 563                         trims++;
 564                 } else if (rw == DDIR_WAIT) {
 565                         if (td->o.no_stall)
 566                                 continue;
 567                         waits++;
 568                 } else if (rw == DDIR_INVAL) {
 569                 } else if (ddir_sync(rw)) {
 570                         syncs++;
 571                 } else {
 572                         log_err("bad ddir: %d\n", rw);
 573                         continue;
 574                 }
 575
 576                 /*
 577                  * Make note of file
 578                  */
 579                 ipo = calloc(1, sizeof(*ipo));
 580                 init_ipo(ipo);
 581                 ipo->ddir = rw;
 582                 if (td->io_log_version == 3)
 583                         ipo->delay = delay;
 584                 if (rw == DDIR_WAIT) {
 585                         ipo->delay = offset;
 586                 } else {
 587                         if (td->o.replay_scale)
 588                                 ipo->offset = offset / td->o.replay_scale;
 589                         else
 590                                 ipo->offset = offset;
 591                         ipo_bytes_align(td->o.replay_align, ipo);
 592
 593                         ipo->len = bytes;
 594                         if (rw != DDIR_INVAL && bytes > td->o.max_bs[rw]) {
 595                                 realloc = true;
 596                                 td->o.max_bs[rw] = bytes;
 597                         }
 598                         ipo->fileno = fileno;
 599                         ipo->file_action = file_action;
 600                         td->o.size += bytes;
 601                 }
 602
 603                 queue_io_piece(td, ipo);
 604
 605                 if (td->o.read_iolog_chunked) {
 606                         td->io_log_current++;
 607                         items_to_fetch--;
 608                         if (items_to_fetch == 0)
 609                                 break;
 610                 }
 611         }
 612
 613         free(str);
 614         free(act);
 615         free(rfname);
 616
 617         if (td->o.read_iolog_chunked) {
 618                 td->io_log_highmark = td->io_log_current;
 619                 td->io_log_checkmark = (td->io_log_highmark + 1) / 2;
 620                 fio_gettime(&td->io_log_highmark_time, NULL);
 621         }
 622
 623         if (writes && read_only) {
 624                 log_err("fio: <%s> skips replay of %d writes due to"
 625                         " read-only\n", td->o.name, writes);
 626                 writes = 0;
 627         }
 628         if (syncs)
 629                 td->flags |= TD_F_SYNCS;
 630
 631         if (td->o.read_iolog_chunked) {
 632                 if (td->io_log_current == 0) {
 633                         return false;
 634                 }
 635                 td->o.td_ddir = TD_DDIR_RW;
 636                 if (realloc && td->orig_buffer)
 637                 {
 638                         io_u_quiesce(td);
 639                         free_io_mem(td);
 640                         if (init_io_u_buffers(td))
 641                                 return false;
 642                 }
 643                 return true;
 644         }
 645
 646         if (!reads && !writes && !waits && !trims)
 647                 return false;
 648
 649         td->o.td_ddir = 0;
 650         if (reads)
 651                 td->o.td_ddir |= TD_DDIR_READ;
 652         if (writes)
 653                 td->o.td_ddir |= TD_DDIR_WRITE;
 654         if (trims)
 655                 td->o.td_ddir |= TD_DDIR_TRIM;
 656
 657         return true;
 658 }
 659
 660 static bool is_socket(const char *path)
 661 {
 662         struct stat buf;
 663         int r;
 664
 665         r = stat(path, &buf);
 666         if (r == -1)
 667                 return false;
 668
 669         return S_ISSOCK(buf.st_mode);
 670 }
 671
 672 static int open_socket(const char *path)
 673 {
 674         struct sockaddr_un addr;
 675         int ret, fd;
 676
 677         fd = socket(AF_UNIX, SOCK_STREAM, 0);
 678         if (fd < 0)
 679                 return fd;
 680
 681         addr.sun_family = AF_UNIX;
 682         if (snprintf(addr.sun_path, sizeof(addr.sun_path), "%s", path) >=
 683             sizeof(addr.sun_path)) {
 684                 log_err("%s: path name %s is too long for a Unix socket\n",
 685                         __func__, path);
 686         }
 687
 688         ret = connect(fd, (const struct sockaddr *)&addr, strlen(path) + sizeof(addr.sun_family));
 689         if (!ret)
 690                 return fd;
 691
 692         close(fd);
 693         return -1;
 694 }
 695
 696 /*
 697  * open iolog, check version, and call appropriate parser
 698  */
 699 static bool init_iolog_read(struct thread_data *td, char *fname)
 700 {
 701         char buffer[256], *p;
 702         FILE *f = NULL;
 703
 704         dprint(FD_IO, "iolog: name=%s\n", fname);
 705
 706         if (is_socket(fname)) {
 707                 int fd;
 708
 709                 fd = open_socket(fname);
 710                 if (fd >= 0)
 711                         f = fdopen(fd, "r");
 712         } else if (!strcmp(fname, "-")) {
 713                 f = stdin;
 714         } else
 715                 f = fopen(fname, "r");
 716
 717         if (!f) {
 718                 perror("fopen read iolog");
 719                 return false;
 720         }
 721
 722         p = fgets(buffer, sizeof(buffer), f);
 723         if (!p) {
 724                 td_verror(td, errno, "iolog read");
 725                 log_err("fio: unable to read iolog\n");
 726                 fclose(f);
 727                 return false;
 728         }
 729
 730         /*
 731          * versions 2 and 3 of the iolog store a specific string as the
 732          * first line, check for that
 733          */
 734         if (!strncmp(iolog_ver2, buffer, strlen(iolog_ver2)))
 735                 td->io_log_version = 2;
 736         else if (!strncmp(iolog_ver3, buffer, strlen(iolog_ver3)))
 737                 td->io_log_version = 3;
 738         else {
 739                 log_err("fio: iolog version 1 is no longer supported\n");
 740                 fclose(f);
 741                 return false;
 742         }
 743
 744         free_release_files(td);
 745         td->io_log_rfile = f;
 746         return read_iolog(td);
 747 }
 748
 749 /*
 750  * Set up a log for storing io patterns.
 751  */
 752 static bool init_iolog_write(struct thread_data *td)
 753 {
 754         struct fio_file *ff;
 755         FILE *f;
 756         unsigned int i;
 757
 758         f = fopen(td->o.write_iolog_file, "a");
 759         if (!f) {
 760                 perror("fopen write iolog");
 761                 return false;
 762         }
 763
 764         /*
 765          * That's it for writing, setup a log buffer and we're done.
 766           */
 767         td->iolog_f = f;
 768         td->iolog_buf = malloc(8192);
 769         setvbuf(f, td->iolog_buf, _IOFBF, 8192);
 770         fio_gettime(&td->io_log_start_time, NULL);
 771
 772         /*
 773          * write our version line
 774          */
 775         if (fprintf(f, "%s\n", iolog_ver3) < 0) {
 776                 perror("iolog init\n");
 777                 return false;
 778         }
 779
 780         /*
 781          * add all known files
 782          */
 783         for_each_file(td, ff, i)
 784                 log_file(td, ff, FIO_LOG_ADD_FILE);
 785
 786         return true;
 787 }
 788
 789 bool init_iolog(struct thread_data *td)
 790 {
 791         bool ret;
 792
 793         if (td->o.read_iolog_file) {
 794                 int need_swap;
 795                 char * fname = get_name_by_idx(td->o.read_iolog_file, td->subjob_number);
 796
 797                 /*
 798                  * Check if it's a blktrace file and load that if possible.
 799                  * Otherwise assume it's a normal log file and load that.
 800                  */
 801                 if (is_blktrace(fname, &need_swap)) {
 802                         td->io_log_blktrace = 1;
 803                         ret = init_blktrace_read(td, fname, need_swap);
 804                 } else {
 805                         td->io_log_blktrace = 0;
 806                         ret = init_iolog_read(td, fname);
 807                 }
 808                 free(fname);
 809         } else if (td->o.write_iolog_file)
 810                 ret = init_iolog_write(td);
 811         else
 812                 ret = true;
 813
 814         if (!ret)
 815                 td_verror(td, EINVAL, "failed initializing iolog");
 816
 817         init_disk_util(td);
 818
 819         return ret;
 820 }
 821
 822 void setup_log(struct io_log **log, struct log_params *p,
 823                const char *filename)
 824 {
 825         struct io_log *l;
 826         int i;
 827         struct io_u_plat_entry *entry;
 828         struct flist_head *list;
 829
 830         l = scalloc(1, sizeof(*l));
 831         INIT_FLIST_HEAD(&l->io_logs);
 832         l->log_type = p->log_type;
 833         l->log_offset = p->log_offset;
 834         l->log_prio = p->log_prio;
 835         l->log_gz = p->log_gz;
 836         l->log_gz_store = p->log_gz_store;
 837         l->avg_msec = p->avg_msec;
 838         l->hist_msec = p->hist_msec;
 839         l->hist_coarseness = p->hist_coarseness;
 840         l->filename = strdup(filename);
 841         l->td = p->td;
 842
 843         /* Initialize histogram lists for each r/w direction,
 844          * with initial io_u_plat of all zeros:
 845          */
 846         for (i = 0; i < DDIR_RWDIR_CNT; i++) {
 847                 list = &l->hist_window[i].list;
 848                 INIT_FLIST_HEAD(list);
 849                 entry = calloc(1, sizeof(struct io_u_plat_entry));
 850                 flist_add(&entry->list, list);
 851         }
 852
 853         if (l->td && l->td->o.io_submit_mode != IO_MODE_OFFLOAD) {
 854                 unsigned int def_samples = DEF_LOG_ENTRIES;
 855                 struct io_logs *__p;
 856
 857                 __p = calloc(1, sizeof(*l->pending));
 858                 if (l->td->o.iodepth > DEF_LOG_ENTRIES)
 859                         def_samples = roundup_pow2(l->td->o.iodepth);
 860                 __p->max_samples = def_samples;
 861                 __p->log = calloc(__p->max_samples, log_entry_sz(l));
 862                 l->pending = __p;
 863         }
 864
 865         if (l->log_offset)
 866                 l->log_ddir_mask = LOG_OFFSET_SAMPLE_BIT;
 867         if (l->log_prio)
 868                 l->log_ddir_mask |= LOG_PRIO_SAMPLE_BIT;
 869         /*
 870          * The bandwidth-log option generates agg-read_bw.log,
 871          * agg-write_bw.log and agg-trim_bw.log for which l->td is NULL.
 872          * Check if l->td is valid before dereferencing it.
 873          */
 874         if (l->td && l->td->o.log_max == IO_LOG_SAMPLE_BOTH)
 875                 l->log_ddir_mask |= LOG_AVG_MAX_SAMPLE_BIT;
 876
 877         INIT_FLIST_HEAD(&l->chunk_list);
 878
 879         if (l->log_gz && !p->td)
 880                 l->log_gz = 0;
 881         else if (l->log_gz || l->log_gz_store) {
 882                 mutex_init_pshared(&l->chunk_lock);
 883                 mutex_init_pshared(&l->deferred_free_lock);
 884                 p->td->flags |= TD_F_COMPRESS_LOG;
 885         }
 886
 887         *log = l;
 888 }
 889
 890 #ifdef CONFIG_SETVBUF
 891 static void *set_file_buffer(FILE *f)
 892 {
 893         size_t size = 1048576;
 894         void *buf;
 895
 896         buf = malloc(size);
 897         setvbuf(f, buf, _IOFBF, size);
 898         return buf;
 899 }
 900
 901 static void clear_file_buffer(void *buf)
 902 {
 903         free(buf);
 904 }
 905 #else
 906 static void *set_file_buffer(FILE *f)
 907 {
 908         return NULL;
 909 }
 910
 911 static void clear_file_buffer(void *buf)
 912 {
 913 }
 914 #endif
 915
 916 void free_log(struct io_log *log)
 917 {
 918         while (!flist_empty(&log->io_logs)) {
 919                 struct io_logs *cur_log;
 920
 921                 cur_log = flist_first_entry(&log->io_logs, struct io_logs, list);
 922                 flist_del_init(&cur_log->list);
 923                 free(cur_log->log);
 924                 sfree(cur_log);
 925         }
 926
 927         if (log->pending) {
 928                 free(log->pending->log);
 929                 free(log->pending);
 930                 log->pending = NULL;
 931         }
 932
 933         free(log->pending);
 934         free(log->filename);
 935         sfree(log);
 936 }
 937
 938 uint64_t hist_sum(int j, int stride, uint64_t *io_u_plat,
 939                 uint64_t *io_u_plat_last)
 940 {
 941         uint64_t sum;
 942         int k;
 943
 944         if (io_u_plat_last) {
 945                 for (k = sum = 0; k < stride; k++)
 946                         sum += io_u_plat[j + k] - io_u_plat_last[j + k];
 947         } else {
 948                 for (k = sum = 0; k < stride; k++)
 949                         sum += io_u_plat[j + k];
 950         }
 951
 952         return sum;
 953 }
 954
 955 static void flush_hist_samples(FILE *f, int hist_coarseness, void *samples,
 956                                uint64_t sample_size)
 957 {
 958         struct io_sample *s;
 959         int log_offset;
 960         uint64_t i, j, nr_samples;
 961         struct io_u_plat_entry *entry, *entry_before;
 962         uint64_t *io_u_plat;
 963         uint64_t *io_u_plat_before;
 964
 965         int stride = 1 << hist_coarseness;
 966
 967         if (!sample_size)
 968                 return;
 969
 970         s = __get_sample(samples, 0, 0);
 971         log_offset = (s->__ddir & LOG_OFFSET_SAMPLE_BIT) != 0;
 972
 973         nr_samples = sample_size / __log_entry_sz(log_offset);
 974
 975         for (i = 0; i < nr_samples; i++) {
 976                 s = __get_sample(samples, log_offset, i);
 977
 978                 entry = s->data.plat_entry;
 979                 io_u_plat = entry->io_u_plat;
 980
 981                 entry_before = flist_first_entry(&entry->list, struct io_u_plat_entry, list);
 982                 io_u_plat_before = entry_before->io_u_plat;
 983
 984                 fprintf(f, "%lu, %u, %llu, ", (unsigned long) s->time,
 985                                                 io_sample_ddir(s), (unsigned long long) s->bs);
 986                 for (j = 0; j < FIO_IO_U_PLAT_NR - stride; j += stride) {
 987                         fprintf(f, "%llu, ", (unsigned long long)
 988                                 hist_sum(j, stride, io_u_plat, io_u_plat_before));
 989                 }
 990                 fprintf(f, "%llu\n", (unsigned long long)
 991                         hist_sum(FIO_IO_U_PLAT_NR - stride, stride, io_u_plat,
 992                                         io_u_plat_before));
 993
 994                 flist_del(&entry_before->list);
 995                 free(entry_before);
 996         }
 997 }
 998
 999 void flush_samples(FILE *f, void *samples, uint64_t sample_size)
1000 {
1001         struct io_sample *s;
1002         int log_offset, log_prio, log_avg_max;
1003         uint64_t i, nr_samples;
1004         unsigned int prio_val;
1005         const char *fmt;
1006
1007         if (!sample_size)
1008                 return;
1009
1010         s = __get_sample(samples, 0, 0);
1011         log_offset = (s->__ddir & LOG_OFFSET_SAMPLE_BIT) != 0;
1012         log_prio = (s->__ddir & LOG_PRIO_SAMPLE_BIT) != 0;
1013         log_avg_max = (s->__ddir & LOG_AVG_MAX_SAMPLE_BIT) != 0;
1014
1015         if (log_offset) {
1016                 if (log_prio) {
1017                         if (log_avg_max)
1018                                 fmt = "%" PRIu64 ", %" PRId64 ", %" PRId64 ", %u, %llu, %llu, 0x%04x\n";
1019                         else
1020                                 fmt = "%" PRIu64 ", %" PRId64 ", %u, %llu, %llu, 0x%04x\n";
1021                 } else {
1022                         if (log_avg_max)
1023                                 fmt = "%" PRIu64 ", %" PRId64 ", %" PRId64 ", %u, %llu, %llu, %u\n";
1024                         else
1025                                 fmt = "%" PRIu64 ", %" PRId64 ", %u, %llu, %llu, %u\n";
1026                 }
1027         } else {
1028                 if (log_prio) {
1029                         if (log_avg_max)
1030                                 fmt = "%" PRIu64 ", %" PRId64 ", %" PRId64 ", %u, %llu, 0x%04x\n";
1031                         else
1032                                 fmt = "%" PRIu64 ", %" PRId64 ", %u, %llu, 0x%04x\n";
1033                 } else {
1034                         if (log_avg_max)
1035                                 fmt = "%" PRIu64 ", %" PRId64 ", %" PRId64 ", %u, %llu, %u\n";
1036                         else
1037                                 fmt = "%" PRIu64 ", %" PRId64 ", %u, %llu, %u\n";
1038                 }
1039         }
1040
1041         nr_samples = sample_size / __log_entry_sz(log_offset);
1042
1043         for (i = 0; i < nr_samples; i++) {
1044                 s = __get_sample(samples, log_offset, i);
1045
1046                 if (log_prio)
1047                         prio_val = s->priority;
1048                 else
1049                         prio_val = ioprio_value_is_class_rt(s->priority);
1050
1051                 if (!log_offset) {
1052                         if (log_avg_max)
1053                                 fprintf(f, fmt,
1054                                         s->time,
1055                                         s->data.val.val0,
1056                                         s->data.val.val1,
1057                                         io_sample_ddir(s), (unsigned long long) s->bs,
1058                                         prio_val);
1059                         else
1060                                 fprintf(f, fmt,
1061                                         s->time,
1062                                         s->data.val.val0,
1063                                         io_sample_ddir(s), (unsigned long long) s->bs,
1064                                         prio_val);
1065                 } else {
1066                         struct io_sample_offset *so = (void *) s;
1067
1068                         if (log_avg_max)
1069                                 fprintf(f, fmt,
1070                                         s->time,
1071                                         s->data.val.val0,
1072                                         s->data.val.val1,
1073                                         io_sample_ddir(s), (unsigned long long) s->bs,
1074                                         (unsigned long long) so->offset,
1075                                         prio_val);
1076                         else
1077                                 fprintf(f, fmt,
1078                                         s->time,
1079                                         s->data.val.val0,
1080                                         io_sample_ddir(s), (unsigned long long) s->bs,
1081                                         (unsigned long long) so->offset,
1082                                         prio_val);
1083                 }
1084         }
1085 }
1086
1087 #ifdef CONFIG_ZLIB
1088
1089 struct iolog_flush_data {
1090         struct workqueue_work work;
1091         struct io_log *log;
1092         void *samples;
1093         uint32_t nr_samples;
1094         bool free;
1095 };
1096
1097 #define GZ_CHUNK        131072
1098
1099 static struct iolog_compress *get_new_chunk(unsigned int seq)
1100 {
1101         struct iolog_compress *c;
1102
1103         c = malloc(sizeof(*c));
1104         INIT_FLIST_HEAD(&c->list);
1105         c->buf = malloc(GZ_CHUNK);
1106         c->len = 0;
1107         c->seq = seq;
1108         return c;
1109 }
1110
1111 static void free_chunk(struct iolog_compress *ic)
1112 {
1113         free(ic->buf);
1114         free(ic);
1115 }
1116
1117 static int z_stream_init(z_stream *stream, int gz_hdr)
1118 {
1119         int wbits = 15;
1120
1121         memset(stream, 0, sizeof(*stream));
1122         stream->zalloc = Z_NULL;
1123         stream->zfree = Z_NULL;
1124         stream->opaque = Z_NULL;
1125         stream->next_in = Z_NULL;
1126
1127         /*
1128          * zlib magic - add 32 for auto-detection of gz header or not,
1129          * if we decide to store files in a gzip friendly format.
1130          */
1131         if (gz_hdr)
1132                 wbits += 32;
1133
1134         if (inflateInit2(stream, wbits) != Z_OK)
1135                 return 1;
1136
1137         return 0;
1138 }
1139
1140 struct inflate_chunk_iter {
1141         unsigned int seq;
1142         int err;
1143         void *buf;
1144         size_t buf_size;
1145         size_t buf_used;
1146         size_t chunk_sz;
1147 };
1148
1149 static void finish_chunk(z_stream *stream, FILE *f,
1150                          struct inflate_chunk_iter *iter)
1151 {
1152         int ret;
1153
1154         ret = inflateEnd(stream);
1155         if (ret != Z_OK)
1156                 log_err("fio: failed to end log inflation seq %d (%d)\n",
1157                                 iter->seq, ret);
1158
1159         flush_samples(f, iter->buf, iter->buf_used);
1160         free(iter->buf);
1161         iter->buf = NULL;
1162         iter->buf_size = iter->buf_used = 0;
1163 }
1164
1165 /*
1166  * Iterative chunk inflation. Handles cases where we cross into a new
1167  * sequence, doing flush finish of previous chunk if needed.
1168  */
1169 static size_t inflate_chunk(struct iolog_compress *ic, int gz_hdr, FILE *f,
1170                             z_stream *stream, struct inflate_chunk_iter *iter)
1171 {
1172         size_t ret;
1173
1174         dprint(FD_COMPRESS, "inflate chunk size=%lu, seq=%u\n",
1175                                 (unsigned long) ic->len, ic->seq);
1176
1177         if (ic->seq != iter->seq) {
1178                 if (iter->seq)
1179                         finish_chunk(stream, f, iter);
1180
1181                 z_stream_init(stream, gz_hdr);
1182                 iter->seq = ic->seq;
1183         }
1184
1185         stream->avail_in = ic->len;
1186         stream->next_in = ic->buf;
1187
1188         if (!iter->buf_size) {
1189                 iter->buf_size = iter->chunk_sz;
1190                 iter->buf = malloc(iter->buf_size);
1191         }
1192
1193         while (stream->avail_in) {
1194                 size_t this_out = iter->buf_size - iter->buf_used;
1195                 int err;
1196
1197                 stream->avail_out = this_out;
1198                 stream->next_out = iter->buf + iter->buf_used;
1199
1200                 err = inflate(stream, Z_NO_FLUSH);
1201                 if (err < 0) {
1202                         log_err("fio: failed inflating log: %d\n", err);
1203                         iter->err = err;
1204                         break;
1205                 }
1206
1207                 iter->buf_used += this_out - stream->avail_out;
1208
1209                 if (!stream->avail_out) {
1210                         iter->buf_size += iter->chunk_sz;
1211                         iter->buf = realloc(iter->buf, iter->buf_size);
1212                         continue;
1213                 }
1214
1215                 if (err == Z_STREAM_END)
1216                         break;
1217         }
1218
1219         ret = (void *) stream->next_in - ic->buf;
1220
1221         dprint(FD_COMPRESS, "inflated to size=%lu\n", (unsigned long) iter->buf_size);
1222
1223         return ret;
1224 }
1225
1226 /*
1227  * Inflate stored compressed chunks, or write them directly to the log
1228  * file if so instructed.
1229  */
1230 static int inflate_gz_chunks(struct io_log *log, FILE *f)
1231 {
1232         struct inflate_chunk_iter iter = { .chunk_sz = log->log_gz, };
1233         z_stream stream;
1234
1235         while (!flist_empty(&log->chunk_list)) {
1236                 struct iolog_compress *ic;
1237
1238                 ic = flist_first_entry(&log->chunk_list, struct iolog_compress, list);
1239                 flist_del(&ic->list);
1240
1241                 if (log->log_gz_store) {
1242                         size_t ret;
1243
1244                         dprint(FD_COMPRESS, "log write chunk size=%lu, "
1245                                 "seq=%u\n", (unsigned long) ic->len, ic->seq);
1246
1247                         ret = fwrite(ic->buf, ic->len, 1, f);
1248                         if (ret != 1 || ferror(f)) {
1249                                 iter.err = errno;
1250                                 log_err("fio: error writing compressed log\n");
1251                         }
1252                 } else
1253                         inflate_chunk(ic, log->log_gz_store, f, &stream, &iter);
1254
1255                 free_chunk(ic);
1256         }
1257
1258         if (iter.seq) {
1259                 finish_chunk(&stream, f, &iter);
1260                 free(iter.buf);
1261         }
1262
1263         return iter.err;
1264 }
1265
1266 /*
1267  * Open compressed log file and decompress the stored chunks and
1268  * write them to stdout. The chunks are stored sequentially in the
1269  * file, so we iterate over them and do them one-by-one.
1270  */
1271 int iolog_file_inflate(const char *file)
1272 {
1273         struct inflate_chunk_iter iter = { .chunk_sz = 64 * 1024 * 1024, };
1274         struct iolog_compress ic;
1275         z_stream stream;
1276         struct stat sb;
1277         size_t ret;
1278         size_t total;
1279         void *buf;
1280         FILE *f;
1281
1282         f = fopen(file, "rb");
1283         if (!f) {
1284                 perror("fopen");
1285                 return 1;
1286         }
1287
1288         if (stat(file, &sb) < 0) {
1289                 fclose(f);
1290                 perror("stat");
1291                 return 1;
1292         }
1293
1294         ic.buf = buf = malloc(sb.st_size);
1295         ic.len = sb.st_size;
1296         ic.seq = 1;
1297
1298         ret = fread(ic.buf, ic.len, 1, f);
1299         if (ret == 0 && ferror(f)) {
1300                 perror("fread");
1301                 fclose(f);
1302                 free(buf);
1303                 return 1;
1304         } else if (ferror(f) || (!feof(f) && ret != 1)) {
1305                 log_err("fio: short read on reading log\n");
1306                 fclose(f);
1307                 free(buf);
1308                 return 1;
1309         }
1310
1311         fclose(f);
1312
1313         /*
1314          * Each chunk will return Z_STREAM_END. We don't know how many
1315          * chunks are in the file, so we just keep looping and incrementing
1316          * the sequence number until we have consumed the whole compressed
1317          * file.
1318          */
1319         total = ic.len;
1320         do {
1321                 size_t iret;
1322
1323                 iret = inflate_chunk(&ic,  1, stdout, &stream, &iter);
1324                 total -= iret;
1325                 if (!total)
1326                         break;
1327                 if (iter.err)
1328                         break;
1329
1330                 ic.seq++;
1331                 ic.len -= iret;
1332                 ic.buf += iret;
1333         } while (1);
1334
1335         if (iter.seq) {
1336                 finish_chunk(&stream, stdout, &iter);
1337                 free(iter.buf);
1338         }
1339
1340         free(buf);
1341         return iter.err;
1342 }
1343
1344 #else
1345
1346 static int inflate_gz_chunks(struct io_log *log, FILE *f)
1347 {
1348         return 0;
1349 }
1350
1351 int iolog_file_inflate(const char *file)
1352 {
1353         log_err("fio: log inflation not possible without zlib\n");
1354         return 1;
1355 }
1356
1357 #endif
1358
1359 void flush_log(struct io_log *log, bool do_append)
1360 {
1361         void *buf;
1362         FILE *f;
1363
1364         /*
1365          * If log_gz_store is true, we are writing a binary file.
1366          * Set the mode appropriately (on all platforms) to avoid issues
1367          * on windows (line-ending conversions, etc.)
1368          */
1369         if (!do_append)
1370                 if (log->log_gz_store)
1371                         f = fopen(log->filename, "wb");
1372                 else
1373                         f = fopen(log->filename, "w");
1374         else
1375                 if (log->log_gz_store)
1376                         f = fopen(log->filename, "ab");
1377                 else
1378                         f = fopen(log->filename, "a");
1379         if (!f) {
1380                 perror("fopen log");
1381                 return;
1382         }
1383
1384         buf = set_file_buffer(f);
1385
1386         inflate_gz_chunks(log, f);
1387
1388         while (!flist_empty(&log->io_logs)) {
1389                 struct io_logs *cur_log;
1390
1391                 cur_log = flist_first_entry(&log->io_logs, struct io_logs, list);
1392                 flist_del_init(&cur_log->list);
1393
1394                 if (log->td && log == log->td->clat_hist_log)
1395                         flush_hist_samples(f, log->hist_coarseness, cur_log->log,
1396                                            log_sample_sz(log, cur_log));
1397                 else
1398                         flush_samples(f, cur_log->log, log_sample_sz(log, cur_log));
1399
1400                 sfree(cur_log);
1401         }
1402
1403         fclose(f);
1404         clear_file_buffer(buf);
1405 }
1406
1407 static int finish_log(struct thread_data *td, struct io_log *log, int trylock)
1408 {
1409         if (td->flags & TD_F_COMPRESS_LOG)
1410                 iolog_flush(log);
1411
1412         if (trylock) {
1413                 if (fio_trylock_file(log->filename))
1414                         return 1;
1415         } else
1416                 fio_lock_file(log->filename);
1417
1418         if (td->client_type == FIO_CLIENT_TYPE_GUI || is_backend)
1419                 fio_send_iolog(td, log, log->filename);
1420         else
1421                 flush_log(log, !td->o.per_job_logs);
1422
1423         fio_unlock_file(log->filename);
1424         free_log(log);
1425         return 0;
1426 }
1427
1428 size_t log_chunk_sizes(struct io_log *log)
1429 {
1430         struct flist_head *entry;
1431         size_t ret;
1432
1433         if (flist_empty(&log->chunk_list))
1434                 return 0;
1435
1436         ret = 0;
1437         pthread_mutex_lock(&log->chunk_lock);
1438         flist_for_each(entry, &log->chunk_list) {
1439                 struct iolog_compress *c;
1440
1441                 c = flist_entry(entry, struct iolog_compress, list);
1442                 ret += c->len;
1443         }
1444         pthread_mutex_unlock(&log->chunk_lock);
1445         return ret;
1446 }
1447
1448 #ifdef CONFIG_ZLIB
1449
1450 static void iolog_put_deferred(struct io_log *log, void *ptr)
1451 {
1452         if (!ptr)
1453                 return;
1454
1455         pthread_mutex_lock(&log->deferred_free_lock);
1456         if (log->deferred < IOLOG_MAX_DEFER) {
1457                 log->deferred_items[log->deferred] = ptr;
1458                 log->deferred++;
1459         } else if (!fio_did_warn(FIO_WARN_IOLOG_DROP))
1460                 log_err("fio: had to drop log entry free\n");
1461         pthread_mutex_unlock(&log->deferred_free_lock);
1462 }
1463
1464 static void iolog_free_deferred(struct io_log *log)
1465 {
1466         int i;
1467
1468         if (!log->deferred)
1469                 return;
1470
1471         pthread_mutex_lock(&log->deferred_free_lock);
1472
1473         for (i = 0; i < log->deferred; i++) {
1474                 free(log->deferred_items[i]);
1475                 log->deferred_items[i] = NULL;
1476         }
1477
1478         log->deferred = 0;
1479         pthread_mutex_unlock(&log->deferred_free_lock);
1480 }
1481
1482 static int gz_work(struct iolog_flush_data *data)
1483 {
1484         struct iolog_compress *c = NULL;
1485         struct flist_head list;
1486         unsigned int seq;
1487         z_stream stream;
1488         size_t total = 0;
1489         int ret;
1490
1491         INIT_FLIST_HEAD(&list);
1492
1493         memset(&stream, 0, sizeof(stream));
1494         stream.zalloc = Z_NULL;
1495         stream.zfree = Z_NULL;
1496         stream.opaque = Z_NULL;
1497
1498         ret = deflateInit(&stream, Z_DEFAULT_COMPRESSION);
1499         if (ret != Z_OK) {
1500                 log_err("fio: failed to init gz stream\n");
1501                 goto err;
1502         }
1503
1504         seq = ++data->log->chunk_seq;
1505
1506         stream.next_in = (void *) data->samples;
1507         stream.avail_in = data->nr_samples * log_entry_sz(data->log);
1508
1509         dprint(FD_COMPRESS, "deflate input size=%lu, seq=%u, log=%s\n",
1510                                 (unsigned long) stream.avail_in, seq,
1511                                 data->log->filename);
1512         do {
1513                 if (c)
1514                         dprint(FD_COMPRESS, "seq=%d, chunk=%lu\n", seq,
1515                                 (unsigned long) c->len);
1516                 c = get_new_chunk(seq);
1517                 stream.avail_out = GZ_CHUNK;
1518                 stream.next_out = c->buf;
1519                 ret = deflate(&stream, Z_NO_FLUSH);
1520                 if (ret < 0) {
1521                         log_err("fio: deflate log (%d)\n", ret);
1522                         free_chunk(c);
1523                         goto err;
1524                 }
1525
1526                 c->len = GZ_CHUNK - stream.avail_out;
1527                 flist_add_tail(&c->list, &list);
1528                 total += c->len;
1529         } while (stream.avail_in);
1530
1531         stream.next_out = c->buf + c->len;
1532         stream.avail_out = GZ_CHUNK - c->len;
1533
1534         ret = deflate(&stream, Z_FINISH);
1535         if (ret < 0) {
1536                 /*
1537                  * Z_BUF_ERROR is special, it just means we need more
1538                  * output space. We'll handle that below. Treat any other
1539                  * error as fatal.
1540                  */
1541                 if (ret != Z_BUF_ERROR) {
1542                         log_err("fio: deflate log (%d)\n", ret);
1543                         flist_del(&c->list);
1544                         free_chunk(c);
1545                         goto err;
1546                 }
1547         }
1548
1549         total -= c->len;
1550         c->len = GZ_CHUNK - stream.avail_out;
1551         total += c->len;
1552         dprint(FD_COMPRESS, "seq=%d, chunk=%lu\n", seq, (unsigned long) c->len);
1553
1554         if (ret != Z_STREAM_END) {
1555                 do {
1556                         c = get_new_chunk(seq);
1557                         stream.avail_out = GZ_CHUNK;
1558                         stream.next_out = c->buf;
1559                         ret = deflate(&stream, Z_FINISH);
1560                         c->len = GZ_CHUNK - stream.avail_out;
1561                         total += c->len;
1562                         flist_add_tail(&c->list, &list);
1563                         dprint(FD_COMPRESS, "seq=%d, chunk=%lu\n", seq,
1564                                 (unsigned long) c->len);
1565                 } while (ret != Z_STREAM_END);
1566         }
1567
1568         dprint(FD_COMPRESS, "deflated to size=%lu\n", (unsigned long) total);
1569
1570         ret = deflateEnd(&stream);
1571         if (ret != Z_OK)
1572                 log_err("fio: deflateEnd %d\n", ret);
1573
1574         iolog_put_deferred(data->log, data->samples);
1575
1576         if (!flist_empty(&list)) {
1577                 pthread_mutex_lock(&data->log->chunk_lock);
1578                 flist_splice_tail(&list, &data->log->chunk_list);
1579                 pthread_mutex_unlock(&data->log->chunk_lock);
1580         }
1581
1582         ret = 0;
1583 done:
1584         if (data->free)
1585                 sfree(data);
1586         return ret;
1587 err:
1588         while (!flist_empty(&list)) {
1589                 c = flist_first_entry(list.next, struct iolog_compress, list);
1590                 flist_del(&c->list);
1591                 free_chunk(c);
1592         }
1593         ret = 1;
1594         goto done;
1595 }
1596
1597 /*
1598  * Invoked from our compress helper thread, when logging would have exceeded
1599  * the specified memory limitation. Compresses the previously stored
1600  * entries.
1601  */
1602 static int gz_work_async(struct submit_worker *sw, struct workqueue_work *work)
1603 {
1604         return gz_work(container_of(work, struct iolog_flush_data, work));
1605 }
1606
1607 static int gz_init_worker(struct submit_worker *sw)
1608 {
1609         struct thread_data *td = sw->wq->td;
1610
1611         if (!fio_option_is_set(&td->o, log_gz_cpumask))
1612                 return 0;
1613
1614         if (fio_setaffinity(gettid(), td->o.log_gz_cpumask) == -1) {
1615                 log_err("gz: failed to set CPU affinity\n");
1616                 return 1;
1617         }
1618
1619         return 0;
1620 }
1621
1622 static struct workqueue_ops log_compress_wq_ops = {
1623         .fn             = gz_work_async,
1624         .init_worker_fn = gz_init_worker,
1625         .nice           = 1,
1626 };
1627
1628 int iolog_compress_init(struct thread_data *td, struct sk_out *sk_out)
1629 {
1630         if (!(td->flags & TD_F_COMPRESS_LOG))
1631                 return 0;
1632
1633         workqueue_init(td, &td->log_compress_wq, &log_compress_wq_ops, 1, sk_out);
1634         return 0;
1635 }
1636
1637 void iolog_compress_exit(struct thread_data *td)
1638 {
1639         if (!(td->flags & TD_F_COMPRESS_LOG))
1640                 return;
1641
1642         workqueue_exit(&td->log_compress_wq);
1643 }
1644
1645 /*
1646  * Queue work item to compress the existing log entries. We reset the
1647  * current log to a small size, and reference the existing log in the
1648  * data that we queue for compression. Once compression has been done,
1649  * this old log is freed. Will not return until the log compression
1650  * has completed, and will flush all previous logs too
1651  */
1652 static int iolog_flush(struct io_log *log)
1653 {
1654         struct iolog_flush_data *data;
1655
1656         workqueue_flush(&log->td->log_compress_wq);
1657         data = malloc(sizeof(*data));
1658         if (!data)
1659                 return 1;
1660
1661         data->log = log;
1662         data->free = false;
1663
1664         while (!flist_empty(&log->io_logs)) {
1665                 struct io_logs *cur_log;
1666
1667                 cur_log = flist_first_entry(&log->io_logs, struct io_logs, list);
1668                 flist_del_init(&cur_log->list);
1669
1670                 data->samples = cur_log->log;
1671                 data->nr_samples = cur_log->nr_samples;
1672
1673                 sfree(cur_log);
1674
1675                 gz_work(data);
1676         }
1677
1678         free(data);
1679         return 0;
1680 }
1681
1682 int iolog_cur_flush(struct io_log *log, struct io_logs *cur_log)
1683 {
1684         struct iolog_flush_data *data;
1685
1686         data = smalloc(sizeof(*data));
1687         if (!data)
1688                 return 1;
1689
1690         data->log = log;
1691
1692         data->samples = cur_log->log;
1693         data->nr_samples = cur_log->nr_samples;
1694         data->free = true;
1695
1696         cur_log->nr_samples = cur_log->max_samples = 0;
1697         cur_log->log = NULL;
1698
1699         workqueue_enqueue(&log->td->log_compress_wq, &data->work);
1700
1701         iolog_free_deferred(log);
1702
1703         return 0;
1704 }
1705 #else
1706
1707 static int iolog_flush(struct io_log *log)
1708 {
1709         return 1;
1710 }
1711
1712 int iolog_cur_flush(struct io_log *log, struct io_logs *cur_log)
1713 {
1714         return 1;
1715 }
1716
1717 int iolog_compress_init(struct thread_data *td, struct sk_out *sk_out)
1718 {
1719         return 0;
1720 }
1721
1722 void iolog_compress_exit(struct thread_data *td)
1723 {
1724 }
1725
1726 #endif
1727
1728 struct io_logs *iolog_cur_log(struct io_log *log)
1729 {
1730         if (flist_empty(&log->io_logs))
1731                 return NULL;
1732
1733         return flist_last_entry(&log->io_logs, struct io_logs, list);
1734 }
1735
1736 uint64_t iolog_nr_samples(struct io_log *iolog)
1737 {
1738         struct flist_head *entry;
1739         uint64_t ret = 0;
1740
1741         flist_for_each(entry, &iolog->io_logs) {
1742                 struct io_logs *cur_log;
1743
1744                 cur_log = flist_entry(entry, struct io_logs, list);
1745                 ret += cur_log->nr_samples;
1746         }
1747
1748         return ret;
1749 }
1750
1751 static int __write_log(struct thread_data *td, struct io_log *log, int try)
1752 {
1753         if (log)
1754                 return finish_log(td, log, try);
1755
1756         return 0;
1757 }
1758
1759 static int write_iops_log(struct thread_data *td, int try, bool unit_log)
1760 {
1761         int ret;
1762
1763         if (per_unit_log(td->iops_log) != unit_log)
1764                 return 0;
1765
1766         ret = __write_log(td, td->iops_log, try);
1767         if (!ret)
1768                 td->iops_log = NULL;
1769
1770         return ret;
1771 }
1772
1773 static int write_slat_log(struct thread_data *td, int try, bool unit_log)
1774 {
1775         int ret;
1776
1777         if (!unit_log)
1778                 return 0;
1779
1780         ret = __write_log(td, td->slat_log, try);
1781         if (!ret)
1782                 td->slat_log = NULL;
1783
1784         return ret;
1785 }
1786
1787 static int write_clat_log(struct thread_data *td, int try, bool unit_log)
1788 {
1789         int ret;
1790
1791         if (!unit_log)
1792                 return 0;
1793
1794         ret = __write_log(td, td->clat_log, try);
1795         if (!ret)
1796                 td->clat_log = NULL;
1797
1798         return ret;
1799 }
1800
1801 static int write_clat_hist_log(struct thread_data *td, int try, bool unit_log)
1802 {
1803         int ret;
1804
1805         if (!unit_log)
1806                 return 0;
1807
1808         ret = __write_log(td, td->clat_hist_log, try);
1809         if (!ret)
1810                 td->clat_hist_log = NULL;
1811
1812         return ret;
1813 }
1814
1815 static int write_lat_log(struct thread_data *td, int try, bool unit_log)
1816 {
1817         int ret;
1818
1819         if (!unit_log)
1820                 return 0;
1821
1822         ret = __write_log(td, td->lat_log, try);
1823         if (!ret)
1824                 td->lat_log = NULL;
1825
1826         return ret;
1827 }
1828
1829 static int write_bandw_log(struct thread_data *td, int try, bool unit_log)
1830 {
1831         int ret;
1832
1833         if (per_unit_log(td->bw_log) != unit_log)
1834                 return 0;
1835
1836         ret = __write_log(td, td->bw_log, try);
1837         if (!ret)
1838                 td->bw_log = NULL;
1839
1840         return ret;
1841 }
1842
1843 enum {
1844         BW_LOG_MASK     = 1,
1845         LAT_LOG_MASK    = 2,
1846         SLAT_LOG_MASK   = 4,
1847         CLAT_LOG_MASK   = 8,
1848         IOPS_LOG_MASK   = 16,
1849         CLAT_HIST_LOG_MASK = 32,
1850
1851         ALL_LOG_NR      = 6,
1852 };
1853
1854 struct log_type {
1855         unsigned int mask;
1856         int (*fn)(struct thread_data *, int, bool);
1857 };
1858
1859 static struct log_type log_types[] = {
1860         {
1861                 .mask   = BW_LOG_MASK,
1862                 .fn     = write_bandw_log,
1863         },
1864         {
1865                 .mask   = LAT_LOG_MASK,
1866                 .fn     = write_lat_log,
1867         },
1868         {
1869                 .mask   = SLAT_LOG_MASK,
1870                 .fn     = write_slat_log,
1871         },
1872         {
1873                 .mask   = CLAT_LOG_MASK,
1874                 .fn     = write_clat_log,
1875         },
1876         {
1877                 .mask   = IOPS_LOG_MASK,
1878                 .fn     = write_iops_log,
1879         },
1880         {
1881                 .mask   = CLAT_HIST_LOG_MASK,
1882                 .fn     = write_clat_hist_log,
1883         }
1884 };
1885
1886 void td_writeout_logs(struct thread_data *td, bool unit_logs)
1887 {
1888         unsigned int log_mask = 0;
1889         unsigned int log_left = ALL_LOG_NR;
1890         int old_state, i;
1891
1892         old_state = td_bump_runstate(td, TD_FINISHING);
1893
1894         finalize_logs(td, unit_logs);
1895
1896         while (log_left) {
1897                 int prev_log_left = log_left;
1898
1899                 for (i = 0; i < ALL_LOG_NR && log_left; i++) {
1900                         struct log_type *lt = &log_types[i];
1901                         int ret;
1902
1903                         if (!(log_mask & lt->mask)) {
1904                                 ret = lt->fn(td, log_left != 1, unit_logs);
1905                                 if (!ret) {
1906                                         log_left--;
1907                                         log_mask |= lt->mask;
1908                                 }
1909                         }
1910                 }
1911
1912                 if (prev_log_left == log_left)
1913                         usleep(5000);
1914         }
1915
1916         td_restore_runstate(td, old_state);
1917 }
1918
1919 void fio_writeout_logs(bool unit_logs)
1920 {
1921         for_each_td(td) {
1922                 td_writeout_logs(td, unit_logs);
1923         } end_for_each();
1924 }