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         return ret;
 818 }
 819
 820 void setup_log(struct io_log **log, struct log_params *p,
 821                const char *filename)
 822 {
 823         struct io_log *l;
 824         int i;
 825         struct io_u_plat_entry *entry;
 826         struct flist_head *list;
 827
 828         l = scalloc(1, sizeof(*l));
 829         INIT_FLIST_HEAD(&l->io_logs);
 830         l->log_type = p->log_type;
 831         l->log_offset = p->log_offset;
 832         l->log_prio = p->log_prio;
 833         l->log_gz = p->log_gz;
 834         l->log_gz_store = p->log_gz_store;
 835         l->avg_msec = p->avg_msec;
 836         l->hist_msec = p->hist_msec;
 837         l->hist_coarseness = p->hist_coarseness;
 838         l->filename = strdup(filename);
 839         l->td = p->td;
 840
 841         /* Initialize histogram lists for each r/w direction,
 842          * with initial io_u_plat of all zeros:
 843          */
 844         for (i = 0; i < DDIR_RWDIR_CNT; i++) {
 845                 list = &l->hist_window[i].list;
 846                 INIT_FLIST_HEAD(list);
 847                 entry = calloc(1, sizeof(struct io_u_plat_entry));
 848                 flist_add(&entry->list, list);
 849         }
 850
 851         if (l->td && l->td->o.io_submit_mode != IO_MODE_OFFLOAD) {
 852                 unsigned int def_samples = DEF_LOG_ENTRIES;
 853                 struct io_logs *__p;
 854
 855                 __p = calloc(1, sizeof(*l->pending));
 856                 if (l->td->o.iodepth > DEF_LOG_ENTRIES)
 857                         def_samples = roundup_pow2(l->td->o.iodepth);
 858                 __p->max_samples = def_samples;
 859                 __p->log = calloc(__p->max_samples, log_entry_sz(l));
 860                 l->pending = __p;
 861         }
 862
 863         if (l->log_offset)
 864                 l->log_ddir_mask = LOG_OFFSET_SAMPLE_BIT;
 865         if (l->log_prio)
 866                 l->log_ddir_mask |= LOG_PRIO_SAMPLE_BIT;
 867         /*
 868          * The bandwidth-log option generates agg-read_bw.log,
 869          * agg-write_bw.log and agg-trim_bw.log for which l->td is NULL.
 870          * Check if l->td is valid before dereferencing it.
 871          */
 872         if (l->td && l->td->o.log_max == IO_LOG_SAMPLE_BOTH)
 873                 l->log_ddir_mask |= LOG_AVG_MAX_SAMPLE_BIT;
 874
 875         INIT_FLIST_HEAD(&l->chunk_list);
 876
 877         if (l->log_gz && !p->td)
 878                 l->log_gz = 0;
 879         else if (l->log_gz || l->log_gz_store) {
 880                 mutex_init_pshared(&l->chunk_lock);
 881                 mutex_init_pshared(&l->deferred_free_lock);
 882                 p->td->flags |= TD_F_COMPRESS_LOG;
 883         }
 884
 885         *log = l;
 886 }
 887
 888 #ifdef CONFIG_SETVBUF
 889 static void *set_file_buffer(FILE *f)
 890 {
 891         size_t size = 1048576;
 892         void *buf;
 893
 894         buf = malloc(size);
 895         setvbuf(f, buf, _IOFBF, size);
 896         return buf;
 897 }
 898
 899 static void clear_file_buffer(void *buf)
 900 {
 901         free(buf);
 902 }
 903 #else
 904 static void *set_file_buffer(FILE *f)
 905 {
 906         return NULL;
 907 }
 908
 909 static void clear_file_buffer(void *buf)
 910 {
 911 }
 912 #endif
 913
 914 void free_log(struct io_log *log)
 915 {
 916         while (!flist_empty(&log->io_logs)) {
 917                 struct io_logs *cur_log;
 918
 919                 cur_log = flist_first_entry(&log->io_logs, struct io_logs, list);
 920                 flist_del_init(&cur_log->list);
 921                 free(cur_log->log);
 922                 sfree(cur_log);
 923         }
 924
 925         if (log->pending) {
 926                 free(log->pending->log);
 927                 free(log->pending);
 928                 log->pending = NULL;
 929         }
 930
 931         free(log->pending);
 932         free(log->filename);
 933         sfree(log);
 934 }
 935
 936 uint64_t hist_sum(int j, int stride, uint64_t *io_u_plat,
 937                 uint64_t *io_u_plat_last)
 938 {
 939         uint64_t sum;
 940         int k;
 941
 942         if (io_u_plat_last) {
 943                 for (k = sum = 0; k < stride; k++)
 944                         sum += io_u_plat[j + k] - io_u_plat_last[j + k];
 945         } else {
 946                 for (k = sum = 0; k < stride; k++)
 947                         sum += io_u_plat[j + k];
 948         }
 949
 950         return sum;
 951 }
 952
 953 static void flush_hist_samples(FILE *f, int hist_coarseness, void *samples,
 954                                uint64_t sample_size)
 955 {
 956         struct io_sample *s;
 957         int log_offset;
 958         uint64_t i, j, nr_samples;
 959         struct io_u_plat_entry *entry, *entry_before;
 960         uint64_t *io_u_plat;
 961         uint64_t *io_u_plat_before;
 962
 963         int stride = 1 << hist_coarseness;
 964
 965         if (!sample_size)
 966                 return;
 967
 968         s = __get_sample(samples, 0, 0);
 969         log_offset = (s->__ddir & LOG_OFFSET_SAMPLE_BIT) != 0;
 970
 971         nr_samples = sample_size / __log_entry_sz(log_offset);
 972
 973         for (i = 0; i < nr_samples; i++) {
 974                 s = __get_sample(samples, log_offset, i);
 975
 976                 entry = s->data.plat_entry;
 977                 io_u_plat = entry->io_u_plat;
 978
 979                 entry_before = flist_first_entry(&entry->list, struct io_u_plat_entry, list);
 980                 io_u_plat_before = entry_before->io_u_plat;
 981
 982                 fprintf(f, "%lu, %u, %llu, ", (unsigned long) s->time,
 983                                                 io_sample_ddir(s), (unsigned long long) s->bs);
 984                 for (j = 0; j < FIO_IO_U_PLAT_NR - stride; j += stride) {
 985                         fprintf(f, "%llu, ", (unsigned long long)
 986                                 hist_sum(j, stride, io_u_plat, io_u_plat_before));
 987                 }
 988                 fprintf(f, "%llu\n", (unsigned long long)
 989                         hist_sum(FIO_IO_U_PLAT_NR - stride, stride, io_u_plat,
 990                                         io_u_plat_before));
 991
 992                 flist_del(&entry_before->list);
 993                 free(entry_before);
 994         }
 995 }
 996
 997 void flush_samples(FILE *f, void *samples, uint64_t sample_size)
 998 {
 999         struct io_sample *s;
1000         int log_offset, log_prio, log_avg_max;
1001         uint64_t i, nr_samples;
1002         unsigned int prio_val;
1003         const char *fmt;
1004
1005         if (!sample_size)
1006                 return;
1007
1008         s = __get_sample(samples, 0, 0);
1009         log_offset = (s->__ddir & LOG_OFFSET_SAMPLE_BIT) != 0;
1010         log_prio = (s->__ddir & LOG_PRIO_SAMPLE_BIT) != 0;
1011         log_avg_max = (s->__ddir & LOG_AVG_MAX_SAMPLE_BIT) != 0;
1012
1013         if (log_offset) {
1014                 if (log_prio) {
1015                         if (log_avg_max)
1016                                 fmt = "%" PRIu64 ", %" PRId64 ", %" PRId64 ", %u, %llu, %llu, 0x%04x\n";
1017                         else
1018                                 fmt = "%" PRIu64 ", %" PRId64 ", %u, %llu, %llu, 0x%04x\n";
1019                 } else {
1020                         if (log_avg_max)
1021                                 fmt = "%" PRIu64 ", %" PRId64 ", %" PRId64 ", %u, %llu, %llu, %u\n";
1022                         else
1023                                 fmt = "%" PRIu64 ", %" PRId64 ", %u, %llu, %llu, %u\n";
1024                 }
1025         } else {
1026                 if (log_prio) {
1027                         if (log_avg_max)
1028                                 fmt = "%" PRIu64 ", %" PRId64 ", %" PRId64 ", %u, %llu, 0x%04x\n";
1029                         else
1030                                 fmt = "%" PRIu64 ", %" PRId64 ", %u, %llu, 0x%04x\n";
1031                 } else {
1032                         if (log_avg_max)
1033                                 fmt = "%" PRIu64 ", %" PRId64 ", %" PRId64 ", %u, %llu, %u\n";
1034                         else
1035                                 fmt = "%" PRIu64 ", %" PRId64 ", %u, %llu, %u\n";
1036                 }
1037         }
1038
1039         nr_samples = sample_size / __log_entry_sz(log_offset);
1040
1041         for (i = 0; i < nr_samples; i++) {
1042                 s = __get_sample(samples, log_offset, i);
1043
1044                 if (log_prio)
1045                         prio_val = s->priority;
1046                 else
1047                         prio_val = ioprio_value_is_class_rt(s->priority);
1048
1049                 if (!log_offset) {
1050                         if (log_avg_max)
1051                                 fprintf(f, fmt,
1052                                         s->time,
1053                                         s->data.val.val0,
1054                                         s->data.val.val1,
1055                                         io_sample_ddir(s), (unsigned long long) s->bs,
1056                                         prio_val);
1057                         else
1058                                 fprintf(f, fmt,
1059                                         s->time,
1060                                         s->data.val.val0,
1061                                         io_sample_ddir(s), (unsigned long long) s->bs,
1062                                         prio_val);
1063                 } else {
1064                         struct io_sample_offset *so = (void *) s;
1065
1066                         if (log_avg_max)
1067                                 fprintf(f, fmt,
1068                                         s->time,
1069                                         s->data.val.val0,
1070                                         s->data.val.val1,
1071                                         io_sample_ddir(s), (unsigned long long) s->bs,
1072                                         (unsigned long long) so->offset,
1073                                         prio_val);
1074                         else
1075                                 fprintf(f, fmt,
1076                                         s->time,
1077                                         s->data.val.val0,
1078                                         io_sample_ddir(s), (unsigned long long) s->bs,
1079                                         (unsigned long long) so->offset,
1080                                         prio_val);
1081                 }
1082         }
1083 }
1084
1085 #ifdef CONFIG_ZLIB
1086
1087 struct iolog_flush_data {
1088         struct workqueue_work work;
1089         struct io_log *log;
1090         void *samples;
1091         uint32_t nr_samples;
1092         bool free;
1093 };
1094
1095 #define GZ_CHUNK        131072
1096
1097 static struct iolog_compress *get_new_chunk(unsigned int seq)
1098 {
1099         struct iolog_compress *c;
1100
1101         c = malloc(sizeof(*c));
1102         INIT_FLIST_HEAD(&c->list);
1103         c->buf = malloc(GZ_CHUNK);
1104         c->len = 0;
1105         c->seq = seq;
1106         return c;
1107 }
1108
1109 static void free_chunk(struct iolog_compress *ic)
1110 {
1111         free(ic->buf);
1112         free(ic);
1113 }
1114
1115 static int z_stream_init(z_stream *stream, int gz_hdr)
1116 {
1117         int wbits = 15;
1118
1119         memset(stream, 0, sizeof(*stream));
1120         stream->zalloc = Z_NULL;
1121         stream->zfree = Z_NULL;
1122         stream->opaque = Z_NULL;
1123         stream->next_in = Z_NULL;
1124
1125         /*
1126          * zlib magic - add 32 for auto-detection of gz header or not,
1127          * if we decide to store files in a gzip friendly format.
1128          */
1129         if (gz_hdr)
1130                 wbits += 32;
1131
1132         if (inflateInit2(stream, wbits) != Z_OK)
1133                 return 1;
1134
1135         return 0;
1136 }
1137
1138 struct inflate_chunk_iter {
1139         unsigned int seq;
1140         int err;
1141         void *buf;
1142         size_t buf_size;
1143         size_t buf_used;
1144         size_t chunk_sz;
1145 };
1146
1147 static void finish_chunk(z_stream *stream, FILE *f,
1148                          struct inflate_chunk_iter *iter)
1149 {
1150         int ret;
1151
1152         ret = inflateEnd(stream);
1153         if (ret != Z_OK)
1154                 log_err("fio: failed to end log inflation seq %d (%d)\n",
1155                                 iter->seq, ret);
1156
1157         flush_samples(f, iter->buf, iter->buf_used);
1158         free(iter->buf);
1159         iter->buf = NULL;
1160         iter->buf_size = iter->buf_used = 0;
1161 }
1162
1163 /*
1164  * Iterative chunk inflation. Handles cases where we cross into a new
1165  * sequence, doing flush finish of previous chunk if needed.
1166  */
1167 static size_t inflate_chunk(struct iolog_compress *ic, int gz_hdr, FILE *f,
1168                             z_stream *stream, struct inflate_chunk_iter *iter)
1169 {
1170         size_t ret;
1171
1172         dprint(FD_COMPRESS, "inflate chunk size=%lu, seq=%u\n",
1173                                 (unsigned long) ic->len, ic->seq);
1174
1175         if (ic->seq != iter->seq) {
1176                 if (iter->seq)
1177                         finish_chunk(stream, f, iter);
1178
1179                 z_stream_init(stream, gz_hdr);
1180                 iter->seq = ic->seq;
1181         }
1182
1183         stream->avail_in = ic->len;
1184         stream->next_in = ic->buf;
1185
1186         if (!iter->buf_size) {
1187                 iter->buf_size = iter->chunk_sz;
1188                 iter->buf = malloc(iter->buf_size);
1189         }
1190
1191         while (stream->avail_in) {
1192                 size_t this_out = iter->buf_size - iter->buf_used;
1193                 int err;
1194
1195                 stream->avail_out = this_out;
1196                 stream->next_out = iter->buf + iter->buf_used;
1197
1198                 err = inflate(stream, Z_NO_FLUSH);
1199                 if (err < 0) {
1200                         log_err("fio: failed inflating log: %d\n", err);
1201                         iter->err = err;
1202                         break;
1203                 }
1204
1205                 iter->buf_used += this_out - stream->avail_out;
1206
1207                 if (!stream->avail_out) {
1208                         iter->buf_size += iter->chunk_sz;
1209                         iter->buf = realloc(iter->buf, iter->buf_size);
1210                         continue;
1211                 }
1212
1213                 if (err == Z_STREAM_END)
1214                         break;
1215         }
1216
1217         ret = (void *) stream->next_in - ic->buf;
1218
1219         dprint(FD_COMPRESS, "inflated to size=%lu\n", (unsigned long) iter->buf_size);
1220
1221         return ret;
1222 }
1223
1224 /*
1225  * Inflate stored compressed chunks, or write them directly to the log
1226  * file if so instructed.
1227  */
1228 static int inflate_gz_chunks(struct io_log *log, FILE *f)
1229 {
1230         struct inflate_chunk_iter iter = { .chunk_sz = log->log_gz, };
1231         z_stream stream;
1232
1233         while (!flist_empty(&log->chunk_list)) {
1234                 struct iolog_compress *ic;
1235
1236                 ic = flist_first_entry(&log->chunk_list, struct iolog_compress, list);
1237                 flist_del(&ic->list);
1238
1239                 if (log->log_gz_store) {
1240                         size_t ret;
1241
1242                         dprint(FD_COMPRESS, "log write chunk size=%lu, "
1243                                 "seq=%u\n", (unsigned long) ic->len, ic->seq);
1244
1245                         ret = fwrite(ic->buf, ic->len, 1, f);
1246                         if (ret != 1 || ferror(f)) {
1247                                 iter.err = errno;
1248                                 log_err("fio: error writing compressed log\n");
1249                         }
1250                 } else
1251                         inflate_chunk(ic, log->log_gz_store, f, &stream, &iter);
1252
1253                 free_chunk(ic);
1254         }
1255
1256         if (iter.seq) {
1257                 finish_chunk(&stream, f, &iter);
1258                 free(iter.buf);
1259         }
1260
1261         return iter.err;
1262 }
1263
1264 /*
1265  * Open compressed log file and decompress the stored chunks and
1266  * write them to stdout. The chunks are stored sequentially in the
1267  * file, so we iterate over them and do them one-by-one.
1268  */
1269 int iolog_file_inflate(const char *file)
1270 {
1271         struct inflate_chunk_iter iter = { .chunk_sz = 64 * 1024 * 1024, };
1272         struct iolog_compress ic;
1273         z_stream stream;
1274         struct stat sb;
1275         size_t ret;
1276         size_t total;
1277         void *buf;
1278         FILE *f;
1279
1280         f = fopen(file, "rb");
1281         if (!f) {
1282                 perror("fopen");
1283                 return 1;
1284         }
1285
1286         if (stat(file, &sb) < 0) {
1287                 fclose(f);
1288                 perror("stat");
1289                 return 1;
1290         }
1291
1292         ic.buf = buf = malloc(sb.st_size);
1293         ic.len = sb.st_size;
1294         ic.seq = 1;
1295
1296         ret = fread(ic.buf, ic.len, 1, f);
1297         if (ret == 0 && ferror(f)) {
1298                 perror("fread");
1299                 fclose(f);
1300                 free(buf);
1301                 return 1;
1302         } else if (ferror(f) || (!feof(f) && ret != 1)) {
1303                 log_err("fio: short read on reading log\n");
1304                 fclose(f);
1305                 free(buf);
1306                 return 1;
1307         }
1308
1309         fclose(f);
1310
1311         /*
1312          * Each chunk will return Z_STREAM_END. We don't know how many
1313          * chunks are in the file, so we just keep looping and incrementing
1314          * the sequence number until we have consumed the whole compressed
1315          * file.
1316          */
1317         total = ic.len;
1318         do {
1319                 size_t iret;
1320
1321                 iret = inflate_chunk(&ic,  1, stdout, &stream, &iter);
1322                 total -= iret;
1323                 if (!total)
1324                         break;
1325                 if (iter.err)
1326                         break;
1327
1328                 ic.seq++;
1329                 ic.len -= iret;
1330                 ic.buf += iret;
1331         } while (1);
1332
1333         if (iter.seq) {
1334                 finish_chunk(&stream, stdout, &iter);
1335                 free(iter.buf);
1336         }
1337
1338         free(buf);
1339         return iter.err;
1340 }
1341
1342 #else
1343
1344 static int inflate_gz_chunks(struct io_log *log, FILE *f)
1345 {
1346         return 0;
1347 }
1348
1349 int iolog_file_inflate(const char *file)
1350 {
1351         log_err("fio: log inflation not possible without zlib\n");
1352         return 1;
1353 }
1354
1355 #endif
1356
1357 void flush_log(struct io_log *log, bool do_append)
1358 {
1359         void *buf;
1360         FILE *f;
1361
1362         /*
1363          * If log_gz_store is true, we are writing a binary file.
1364          * Set the mode appropriately (on all platforms) to avoid issues
1365          * on windows (line-ending conversions, etc.)
1366          */
1367         if (!do_append)
1368                 if (log->log_gz_store)
1369                         f = fopen(log->filename, "wb");
1370                 else
1371                         f = fopen(log->filename, "w");
1372         else
1373                 if (log->log_gz_store)
1374                         f = fopen(log->filename, "ab");
1375                 else
1376                         f = fopen(log->filename, "a");
1377         if (!f) {
1378                 perror("fopen log");
1379                 return;
1380         }
1381
1382         buf = set_file_buffer(f);
1383
1384         inflate_gz_chunks(log, f);
1385
1386         while (!flist_empty(&log->io_logs)) {
1387                 struct io_logs *cur_log;
1388
1389                 cur_log = flist_first_entry(&log->io_logs, struct io_logs, list);
1390                 flist_del_init(&cur_log->list);
1391
1392                 if (log->td && log == log->td->clat_hist_log)
1393                         flush_hist_samples(f, log->hist_coarseness, cur_log->log,
1394                                            log_sample_sz(log, cur_log));
1395                 else
1396                         flush_samples(f, cur_log->log, log_sample_sz(log, cur_log));
1397
1398                 sfree(cur_log);
1399         }
1400
1401         fclose(f);
1402         clear_file_buffer(buf);
1403 }
1404
1405 static int finish_log(struct thread_data *td, struct io_log *log, int trylock)
1406 {
1407         if (td->flags & TD_F_COMPRESS_LOG)
1408                 iolog_flush(log);
1409
1410         if (trylock) {
1411                 if (fio_trylock_file(log->filename))
1412                         return 1;
1413         } else
1414                 fio_lock_file(log->filename);
1415
1416         if (td->client_type == FIO_CLIENT_TYPE_GUI || is_backend)
1417                 fio_send_iolog(td, log, log->filename);
1418         else
1419                 flush_log(log, !td->o.per_job_logs);
1420
1421         fio_unlock_file(log->filename);
1422         free_log(log);
1423         return 0;
1424 }
1425
1426 size_t log_chunk_sizes(struct io_log *log)
1427 {
1428         struct flist_head *entry;
1429         size_t ret;
1430
1431         if (flist_empty(&log->chunk_list))
1432                 return 0;
1433
1434         ret = 0;
1435         pthread_mutex_lock(&log->chunk_lock);
1436         flist_for_each(entry, &log->chunk_list) {
1437                 struct iolog_compress *c;
1438
1439                 c = flist_entry(entry, struct iolog_compress, list);
1440                 ret += c->len;
1441         }
1442         pthread_mutex_unlock(&log->chunk_lock);
1443         return ret;
1444 }
1445
1446 #ifdef CONFIG_ZLIB
1447
1448 static void iolog_put_deferred(struct io_log *log, void *ptr)
1449 {
1450         if (!ptr)
1451                 return;
1452
1453         pthread_mutex_lock(&log->deferred_free_lock);
1454         if (log->deferred < IOLOG_MAX_DEFER) {
1455                 log->deferred_items[log->deferred] = ptr;
1456                 log->deferred++;
1457         } else if (!fio_did_warn(FIO_WARN_IOLOG_DROP))
1458                 log_err("fio: had to drop log entry free\n");
1459         pthread_mutex_unlock(&log->deferred_free_lock);
1460 }
1461
1462 static void iolog_free_deferred(struct io_log *log)
1463 {
1464         int i;
1465
1466         if (!log->deferred)
1467                 return;
1468
1469         pthread_mutex_lock(&log->deferred_free_lock);
1470
1471         for (i = 0; i < log->deferred; i++) {
1472                 free(log->deferred_items[i]);
1473                 log->deferred_items[i] = NULL;
1474         }
1475
1476         log->deferred = 0;
1477         pthread_mutex_unlock(&log->deferred_free_lock);
1478 }
1479
1480 static int gz_work(struct iolog_flush_data *data)
1481 {
1482         struct iolog_compress *c = NULL;
1483         struct flist_head list;
1484         unsigned int seq;
1485         z_stream stream;
1486         size_t total = 0;
1487         int ret;
1488
1489         INIT_FLIST_HEAD(&list);
1490
1491         memset(&stream, 0, sizeof(stream));
1492         stream.zalloc = Z_NULL;
1493         stream.zfree = Z_NULL;
1494         stream.opaque = Z_NULL;
1495
1496         ret = deflateInit(&stream, Z_DEFAULT_COMPRESSION);
1497         if (ret != Z_OK) {
1498                 log_err("fio: failed to init gz stream\n");
1499                 goto err;
1500         }
1501
1502         seq = ++data->log->chunk_seq;
1503
1504         stream.next_in = (void *) data->samples;
1505         stream.avail_in = data->nr_samples * log_entry_sz(data->log);
1506
1507         dprint(FD_COMPRESS, "deflate input size=%lu, seq=%u, log=%s\n",
1508                                 (unsigned long) stream.avail_in, seq,
1509                                 data->log->filename);
1510         do {
1511                 if (c)
1512                         dprint(FD_COMPRESS, "seq=%d, chunk=%lu\n", seq,
1513                                 (unsigned long) c->len);
1514                 c = get_new_chunk(seq);
1515                 stream.avail_out = GZ_CHUNK;
1516                 stream.next_out = c->buf;
1517                 ret = deflate(&stream, Z_NO_FLUSH);
1518                 if (ret < 0) {
1519                         log_err("fio: deflate log (%d)\n", ret);
1520                         free_chunk(c);
1521                         goto err;
1522                 }
1523
1524                 c->len = GZ_CHUNK - stream.avail_out;
1525                 flist_add_tail(&c->list, &list);
1526                 total += c->len;
1527         } while (stream.avail_in);
1528
1529         stream.next_out = c->buf + c->len;
1530         stream.avail_out = GZ_CHUNK - c->len;
1531
1532         ret = deflate(&stream, Z_FINISH);
1533         if (ret < 0) {
1534                 /*
1535                  * Z_BUF_ERROR is special, it just means we need more
1536                  * output space. We'll handle that below. Treat any other
1537                  * error as fatal.
1538                  */
1539                 if (ret != Z_BUF_ERROR) {
1540                         log_err("fio: deflate log (%d)\n", ret);
1541                         flist_del(&c->list);
1542                         free_chunk(c);
1543                         goto err;
1544                 }
1545         }
1546
1547         total -= c->len;
1548         c->len = GZ_CHUNK - stream.avail_out;
1549         total += c->len;
1550         dprint(FD_COMPRESS, "seq=%d, chunk=%lu\n", seq, (unsigned long) c->len);
1551
1552         if (ret != Z_STREAM_END) {
1553                 do {
1554                         c = get_new_chunk(seq);
1555                         stream.avail_out = GZ_CHUNK;
1556                         stream.next_out = c->buf;
1557                         ret = deflate(&stream, Z_FINISH);
1558                         c->len = GZ_CHUNK - stream.avail_out;
1559                         total += c->len;
1560                         flist_add_tail(&c->list, &list);
1561                         dprint(FD_COMPRESS, "seq=%d, chunk=%lu\n", seq,
1562                                 (unsigned long) c->len);
1563                 } while (ret != Z_STREAM_END);
1564         }
1565
1566         dprint(FD_COMPRESS, "deflated to size=%lu\n", (unsigned long) total);
1567
1568         ret = deflateEnd(&stream);
1569         if (ret != Z_OK)
1570                 log_err("fio: deflateEnd %d\n", ret);
1571
1572         iolog_put_deferred(data->log, data->samples);
1573
1574         if (!flist_empty(&list)) {
1575                 pthread_mutex_lock(&data->log->chunk_lock);
1576                 flist_splice_tail(&list, &data->log->chunk_list);
1577                 pthread_mutex_unlock(&data->log->chunk_lock);
1578         }
1579
1580         ret = 0;
1581 done:
1582         if (data->free)
1583                 sfree(data);
1584         return ret;
1585 err:
1586         while (!flist_empty(&list)) {
1587                 c = flist_first_entry(list.next, struct iolog_compress, list);
1588                 flist_del(&c->list);
1589                 free_chunk(c);
1590         }
1591         ret = 1;
1592         goto done;
1593 }
1594
1595 /*
1596  * Invoked from our compress helper thread, when logging would have exceeded
1597  * the specified memory limitation. Compresses the previously stored
1598  * entries.
1599  */
1600 static int gz_work_async(struct submit_worker *sw, struct workqueue_work *work)
1601 {
1602         return gz_work(container_of(work, struct iolog_flush_data, work));
1603 }
1604
1605 static int gz_init_worker(struct submit_worker *sw)
1606 {
1607         struct thread_data *td = sw->wq->td;
1608
1609         if (!fio_option_is_set(&td->o, log_gz_cpumask))
1610                 return 0;
1611
1612         if (fio_setaffinity(gettid(), td->o.log_gz_cpumask) == -1) {
1613                 log_err("gz: failed to set CPU affinity\n");
1614                 return 1;
1615         }
1616
1617         return 0;
1618 }
1619
1620 static struct workqueue_ops log_compress_wq_ops = {
1621         .fn             = gz_work_async,
1622         .init_worker_fn = gz_init_worker,
1623         .nice           = 1,
1624 };
1625
1626 int iolog_compress_init(struct thread_data *td, struct sk_out *sk_out)
1627 {
1628         if (!(td->flags & TD_F_COMPRESS_LOG))
1629                 return 0;
1630
1631         workqueue_init(td, &td->log_compress_wq, &log_compress_wq_ops, 1, sk_out);
1632         return 0;
1633 }
1634
1635 void iolog_compress_exit(struct thread_data *td)
1636 {
1637         if (!(td->flags & TD_F_COMPRESS_LOG))
1638                 return;
1639
1640         workqueue_exit(&td->log_compress_wq);
1641 }
1642
1643 /*
1644  * Queue work item to compress the existing log entries. We reset the
1645  * current log to a small size, and reference the existing log in the
1646  * data that we queue for compression. Once compression has been done,
1647  * this old log is freed. Will not return until the log compression
1648  * has completed, and will flush all previous logs too
1649  */
1650 static int iolog_flush(struct io_log *log)
1651 {
1652         struct iolog_flush_data *data;
1653
1654         workqueue_flush(&log->td->log_compress_wq);
1655         data = malloc(sizeof(*data));
1656         if (!data)
1657                 return 1;
1658
1659         data->log = log;
1660         data->free = false;
1661
1662         while (!flist_empty(&log->io_logs)) {
1663                 struct io_logs *cur_log;
1664
1665                 cur_log = flist_first_entry(&log->io_logs, struct io_logs, list);
1666                 flist_del_init(&cur_log->list);
1667
1668                 data->samples = cur_log->log;
1669                 data->nr_samples = cur_log->nr_samples;
1670
1671                 sfree(cur_log);
1672
1673                 gz_work(data);
1674         }
1675
1676         free(data);
1677         return 0;
1678 }
1679
1680 int iolog_cur_flush(struct io_log *log, struct io_logs *cur_log)
1681 {
1682         struct iolog_flush_data *data;
1683
1684         data = smalloc(sizeof(*data));
1685         if (!data)
1686                 return 1;
1687
1688         data->log = log;
1689
1690         data->samples = cur_log->log;
1691         data->nr_samples = cur_log->nr_samples;
1692         data->free = true;
1693
1694         cur_log->nr_samples = cur_log->max_samples = 0;
1695         cur_log->log = NULL;
1696
1697         workqueue_enqueue(&log->td->log_compress_wq, &data->work);
1698
1699         iolog_free_deferred(log);
1700
1701         return 0;
1702 }
1703 #else
1704
1705 static int iolog_flush(struct io_log *log)
1706 {
1707         return 1;
1708 }
1709
1710 int iolog_cur_flush(struct io_log *log, struct io_logs *cur_log)
1711 {
1712         return 1;
1713 }
1714
1715 int iolog_compress_init(struct thread_data *td, struct sk_out *sk_out)
1716 {
1717         return 0;
1718 }
1719
1720 void iolog_compress_exit(struct thread_data *td)
1721 {
1722 }
1723
1724 #endif
1725
1726 struct io_logs *iolog_cur_log(struct io_log *log)
1727 {
1728         if (flist_empty(&log->io_logs))
1729                 return NULL;
1730
1731         return flist_last_entry(&log->io_logs, struct io_logs, list);
1732 }
1733
1734 uint64_t iolog_nr_samples(struct io_log *iolog)
1735 {
1736         struct flist_head *entry;
1737         uint64_t ret = 0;
1738
1739         flist_for_each(entry, &iolog->io_logs) {
1740                 struct io_logs *cur_log;
1741
1742                 cur_log = flist_entry(entry, struct io_logs, list);
1743                 ret += cur_log->nr_samples;
1744         }
1745
1746         return ret;
1747 }
1748
1749 static int __write_log(struct thread_data *td, struct io_log *log, int try)
1750 {
1751         if (log)
1752                 return finish_log(td, log, try);
1753
1754         return 0;
1755 }
1756
1757 static int write_iops_log(struct thread_data *td, int try, bool unit_log)
1758 {
1759         int ret;
1760
1761         if (per_unit_log(td->iops_log) != unit_log)
1762                 return 0;
1763
1764         ret = __write_log(td, td->iops_log, try);
1765         if (!ret)
1766                 td->iops_log = NULL;
1767
1768         return ret;
1769 }
1770
1771 static int write_slat_log(struct thread_data *td, int try, bool unit_log)
1772 {
1773         int ret;
1774
1775         if (!unit_log)
1776                 return 0;
1777
1778         ret = __write_log(td, td->slat_log, try);
1779         if (!ret)
1780                 td->slat_log = NULL;
1781
1782         return ret;
1783 }
1784
1785 static int write_clat_log(struct thread_data *td, int try, bool unit_log)
1786 {
1787         int ret;
1788
1789         if (!unit_log)
1790                 return 0;
1791
1792         ret = __write_log(td, td->clat_log, try);
1793         if (!ret)
1794                 td->clat_log = NULL;
1795
1796         return ret;
1797 }
1798
1799 static int write_clat_hist_log(struct thread_data *td, int try, bool unit_log)
1800 {
1801         int ret;
1802
1803         if (!unit_log)
1804                 return 0;
1805
1806         ret = __write_log(td, td->clat_hist_log, try);
1807         if (!ret)
1808                 td->clat_hist_log = NULL;
1809
1810         return ret;
1811 }
1812
1813 static int write_lat_log(struct thread_data *td, int try, bool unit_log)
1814 {
1815         int ret;
1816
1817         if (!unit_log)
1818                 return 0;
1819
1820         ret = __write_log(td, td->lat_log, try);
1821         if (!ret)
1822                 td->lat_log = NULL;
1823
1824         return ret;
1825 }
1826
1827 static int write_bandw_log(struct thread_data *td, int try, bool unit_log)
1828 {
1829         int ret;
1830
1831         if (per_unit_log(td->bw_log) != unit_log)
1832                 return 0;
1833
1834         ret = __write_log(td, td->bw_log, try);
1835         if (!ret)
1836                 td->bw_log = NULL;
1837
1838         return ret;
1839 }
1840
1841 enum {
1842         BW_LOG_MASK     = 1,
1843         LAT_LOG_MASK    = 2,
1844         SLAT_LOG_MASK   = 4,
1845         CLAT_LOG_MASK   = 8,
1846         IOPS_LOG_MASK   = 16,
1847         CLAT_HIST_LOG_MASK = 32,
1848
1849         ALL_LOG_NR      = 6,
1850 };
1851
1852 struct log_type {
1853         unsigned int mask;
1854         int (*fn)(struct thread_data *, int, bool);
1855 };
1856
1857 static struct log_type log_types[] = {
1858         {
1859                 .mask   = BW_LOG_MASK,
1860                 .fn     = write_bandw_log,
1861         },
1862         {
1863                 .mask   = LAT_LOG_MASK,
1864                 .fn     = write_lat_log,
1865         },
1866         {
1867                 .mask   = SLAT_LOG_MASK,
1868                 .fn     = write_slat_log,
1869         },
1870         {
1871                 .mask   = CLAT_LOG_MASK,
1872                 .fn     = write_clat_log,
1873         },
1874         {
1875                 .mask   = IOPS_LOG_MASK,
1876                 .fn     = write_iops_log,
1877         },
1878         {
1879                 .mask   = CLAT_HIST_LOG_MASK,
1880                 .fn     = write_clat_hist_log,
1881         }
1882 };
1883
1884 void td_writeout_logs(struct thread_data *td, bool unit_logs)
1885 {
1886         unsigned int log_mask = 0;
1887         unsigned int log_left = ALL_LOG_NR;
1888         int old_state, i;
1889
1890         old_state = td_bump_runstate(td, TD_FINISHING);
1891
1892         finalize_logs(td, unit_logs);
1893
1894         while (log_left) {
1895                 int prev_log_left = log_left;
1896
1897                 for (i = 0; i < ALL_LOG_NR && log_left; i++) {
1898                         struct log_type *lt = &log_types[i];
1899                         int ret;
1900
1901                         if (!(log_mask & lt->mask)) {
1902                                 ret = lt->fn(td, log_left != 1, unit_logs);
1903                                 if (!ret) {
1904                                         log_left--;
1905                                         log_mask |= lt->mask;
1906                                 }
1907                         }
1908                 }
1909
1910                 if (prev_log_left == log_left)
1911                         usleep(5000);
1912         }
1913
1914         td_restore_runstate(td, old_state);
1915 }
1916
1917 void fio_writeout_logs(bool unit_logs)
1918 {
1919         for_each_td(td) {
1920                 td_writeout_logs(td, unit_logs);
1921         } end_for_each();
1922 }