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
  35 void queue_io_piece(struct thread_data *td, struct io_piece *ipo)
  36 {
  37         flist_add_tail(&ipo->list, &td->io_log_list);
  38         td->total_io_size += ipo->len;
  39 }
  40
  41 void log_io_u(const struct thread_data *td, const struct io_u *io_u)
  42 {
  43         if (!td->o.write_iolog_file)
  44                 return;
  45
  46         fprintf(td->iolog_f, "%s %s %llu %llu\n", io_u->file->file_name,
  47                                                 io_ddir_name(io_u->ddir),
  48                                                 io_u->offset, io_u->buflen);
  49 }
  50
  51 void log_file(struct thread_data *td, struct fio_file *f,
  52               enum file_log_act what)
  53 {
  54         const char *act[] = { "add", "open", "close" };
  55
  56         assert(what < 3);
  57
  58         if (!td->o.write_iolog_file)
  59                 return;
  60
  61
  62         /*
  63          * this happens on the pre-open/close done before the job starts
  64          */
  65         if (!td->iolog_f)
  66                 return;
  67
  68         fprintf(td->iolog_f, "%s %s\n", f->file_name, act[what]);
  69 }
  70
  71 static void iolog_delay(struct thread_data *td, unsigned long delay)
  72 {
  73         uint64_t usec = utime_since_now(&td->last_issue);
  74         unsigned long orig_delay = delay;
  75         uint64_t this_delay;
  76         struct timespec ts;
  77
  78         if (delay < td->time_offset) {
  79                 td->time_offset = 0;
  80                 return;
  81         }
  82
  83         delay -= td->time_offset;
  84         if (delay < usec)
  85                 return;
  86
  87         delay -= usec;
  88
  89         fio_gettime(&ts, NULL);
  90         while (delay && !td->terminate) {
  91                 this_delay = delay;
  92                 if (this_delay > 500000)
  93                         this_delay = 500000;
  94
  95                 usec_sleep(td, this_delay);
  96                 delay -= this_delay;
  97         }
  98
  99         usec = utime_since_now(&ts);
 100         if (usec > orig_delay)
 101                 td->time_offset = usec - orig_delay;
 102         else
 103                 td->time_offset = 0;
 104 }
 105
 106 static int ipo_special(struct thread_data *td, struct io_piece *ipo)
 107 {
 108         struct fio_file *f;
 109         int ret;
 110
 111         /*
 112          * Not a special ipo
 113          */
 114         if (ipo->ddir != DDIR_INVAL)
 115                 return 0;
 116
 117         f = td->files[ipo->fileno];
 118
 119         switch (ipo->file_action) {
 120         case FIO_LOG_OPEN_FILE:
 121                 if (td->o.replay_redirect && fio_file_open(f)) {
 122                         dprint(FD_FILE, "iolog: ignoring re-open of file %s\n",
 123                                         f->file_name);
 124                         break;
 125                 }
 126                 ret = td_io_open_file(td, f);
 127                 if (!ret)
 128                         break;
 129                 td_verror(td, ret, "iolog open file");
 130                 return -1;
 131         case FIO_LOG_CLOSE_FILE:
 132                 td_io_close_file(td, f);
 133                 break;
 134         case FIO_LOG_UNLINK_FILE:
 135                 td_io_unlink_file(td, f);
 136                 break;
 137         default:
 138                 log_err("fio: bad file action %d\n", ipo->file_action);
 139                 break;
 140         }
 141
 142         return 1;
 143 }
 144
 145 static bool read_iolog2(struct thread_data *td);
 146
 147 int read_iolog_get(struct thread_data *td, struct io_u *io_u)
 148 {
 149         struct io_piece *ipo;
 150         unsigned long elapsed;
 151
 152         while (!flist_empty(&td->io_log_list)) {
 153                 int ret;
 154
 155                 if (!td->io_log_blktrace && td->o.read_iolog_chunked) {
 156                         if (td->io_log_checkmark == td->io_log_current) {
 157                                 if (!read_iolog2(td))
 158                                         return 1;
 159                         }
 160                         td->io_log_current--;
 161                 }
 162                 ipo = flist_first_entry(&td->io_log_list, struct io_piece, list);
 163                 flist_del(&ipo->list);
 164                 remove_trim_entry(td, ipo);
 165
 166                 ret = ipo_special(td, ipo);
 167                 if (ret < 0) {
 168                         free(ipo);
 169                         break;
 170                 } else if (ret > 0) {
 171                         free(ipo);
 172                         continue;
 173                 }
 174
 175                 io_u->ddir = ipo->ddir;
 176                 if (ipo->ddir != DDIR_WAIT) {
 177                         io_u->offset = ipo->offset;
 178                         io_u->verify_offset = ipo->offset;
 179                         io_u->buflen = ipo->len;
 180                         io_u->file = td->files[ipo->fileno];
 181                         get_file(io_u->file);
 182                         dprint(FD_IO, "iolog: get %llu/%llu/%s\n", io_u->offset,
 183                                                 io_u->buflen, io_u->file->file_name);
 184                         if (ipo->delay)
 185                                 iolog_delay(td, ipo->delay);
 186                 } else {
 187                         elapsed = mtime_since_genesis();
 188                         if (ipo->delay > elapsed)
 189                                 usec_sleep(td, (ipo->delay - elapsed) * 1000);
 190                 }
 191
 192                 free(ipo);
 193
 194                 if (io_u->ddir != DDIR_WAIT)
 195                         return 0;
 196         }
 197
 198         td->done = 1;
 199         return 1;
 200 }
 201
 202 void prune_io_piece_log(struct thread_data *td)
 203 {
 204         struct io_piece *ipo;
 205         struct fio_rb_node *n;
 206
 207         while ((n = rb_first(&td->io_hist_tree)) != NULL) {
 208                 ipo = rb_entry(n, struct io_piece, rb_node);
 209                 rb_erase(n, &td->io_hist_tree);
 210                 remove_trim_entry(td, ipo);
 211                 td->io_hist_len--;
 212                 free(ipo);
 213         }
 214
 215         while (!flist_empty(&td->io_hist_list)) {
 216                 ipo = flist_first_entry(&td->io_hist_list, struct io_piece, list);
 217                 flist_del(&ipo->list);
 218                 remove_trim_entry(td, ipo);
 219                 td->io_hist_len--;
 220                 free(ipo);
 221         }
 222 }
 223
 224 /*
 225  * log a successful write, so we can unwind the log for verify
 226  */
 227 void log_io_piece(struct thread_data *td, struct io_u *io_u)
 228 {
 229         struct fio_rb_node **p, *parent;
 230         struct io_piece *ipo, *__ipo;
 231
 232         ipo = calloc(1, sizeof(struct io_piece));
 233         init_ipo(ipo);
 234         ipo->file = io_u->file;
 235         ipo->offset = io_u->offset;
 236         ipo->len = io_u->buflen;
 237         ipo->numberio = io_u->numberio;
 238         ipo->flags = IP_F_IN_FLIGHT;
 239
 240         io_u->ipo = ipo;
 241
 242         if (io_u_should_trim(td, io_u)) {
 243                 flist_add_tail(&ipo->trim_list, &td->trim_list);
 244                 td->trim_entries++;
 245         }
 246
 247         /*
 248          * Only sort writes if we don't have a random map in which case we need
 249          * to check for duplicate blocks and drop the old one, which we rely on
 250          * the rb insert/lookup for handling.
 251          */
 252         if (file_randommap(td, ipo->file)) {
 253                 INIT_FLIST_HEAD(&ipo->list);
 254                 flist_add_tail(&ipo->list, &td->io_hist_list);
 255                 ipo->flags |= IP_F_ONLIST;
 256                 td->io_hist_len++;
 257                 return;
 258         }
 259
 260         RB_CLEAR_NODE(&ipo->rb_node);
 261
 262         /*
 263          * Sort the entry into the verification list
 264          */
 265 restart:
 266         p = &td->io_hist_tree.rb_node;
 267         parent = NULL;
 268         while (*p) {
 269                 int overlap = 0;
 270                 parent = *p;
 271
 272                 __ipo = rb_entry(parent, struct io_piece, rb_node);
 273                 if (ipo->file < __ipo->file)
 274                         p = &(*p)->rb_left;
 275                 else if (ipo->file > __ipo->file)
 276                         p = &(*p)->rb_right;
 277                 else if (ipo->offset < __ipo->offset) {
 278                         p = &(*p)->rb_left;
 279                         overlap = ipo->offset + ipo->len > __ipo->offset;
 280                 }
 281                 else if (ipo->offset > __ipo->offset) {
 282                         p = &(*p)->rb_right;
 283                         overlap = __ipo->offset + __ipo->len > ipo->offset;
 284                 }
 285                 else
 286                         overlap = 1;
 287
 288                 if (overlap) {
 289                         dprint(FD_IO, "iolog: overlap %llu/%lu, %llu/%lu\n",
 290                                 __ipo->offset, __ipo->len,
 291                                 ipo->offset, ipo->len);
 292                         td->io_hist_len--;
 293                         rb_erase(parent, &td->io_hist_tree);
 294                         remove_trim_entry(td, __ipo);
 295                         if (!(__ipo->flags & IP_F_IN_FLIGHT))
 296                                 free(__ipo);
 297                         goto restart;
 298                 }
 299         }
 300
 301         rb_link_node(&ipo->rb_node, parent, p);
 302         rb_insert_color(&ipo->rb_node, &td->io_hist_tree);
 303         ipo->flags |= IP_F_ONRB;
 304         td->io_hist_len++;
 305 }
 306
 307 void unlog_io_piece(struct thread_data *td, struct io_u *io_u)
 308 {
 309         struct io_piece *ipo = io_u->ipo;
 310
 311         if (td->ts.nr_block_infos) {
 312                 uint32_t *info = io_u_block_info(td, io_u);
 313                 if (BLOCK_INFO_STATE(*info) < BLOCK_STATE_TRIM_FAILURE) {
 314                         if (io_u->ddir == DDIR_TRIM)
 315                                 *info = BLOCK_INFO_SET_STATE(*info,
 316                                                 BLOCK_STATE_TRIM_FAILURE);
 317                         else if (io_u->ddir == DDIR_WRITE)
 318                                 *info = BLOCK_INFO_SET_STATE(*info,
 319                                                 BLOCK_STATE_WRITE_FAILURE);
 320                 }
 321         }
 322
 323         if (!ipo)
 324                 return;
 325
 326         if (ipo->flags & IP_F_ONRB)
 327                 rb_erase(&ipo->rb_node, &td->io_hist_tree);
 328         else if (ipo->flags & IP_F_ONLIST)
 329                 flist_del(&ipo->list);
 330
 331         free(ipo);
 332         io_u->ipo = NULL;
 333         td->io_hist_len--;
 334 }
 335
 336 void trim_io_piece(const struct io_u *io_u)
 337 {
 338         struct io_piece *ipo = io_u->ipo;
 339
 340         if (!ipo)
 341                 return;
 342
 343         ipo->len = io_u->xfer_buflen - io_u->resid;
 344 }
 345
 346 void write_iolog_close(struct thread_data *td)
 347 {
 348         if (!td->iolog_f)
 349                 return;
 350
 351         fflush(td->iolog_f);
 352         fclose(td->iolog_f);
 353         free(td->iolog_buf);
 354         td->iolog_f = NULL;
 355         td->iolog_buf = NULL;
 356 }
 357
 358 static int64_t iolog_items_to_fetch(struct thread_data *td)
 359 {
 360         struct timespec now;
 361         uint64_t elapsed;
 362         uint64_t for_1s;
 363         int64_t items_to_fetch;
 364
 365         if (!td->io_log_highmark)
 366                 return 10;
 367
 368
 369         fio_gettime(&now, NULL);
 370         elapsed = ntime_since(&td->io_log_highmark_time, &now);
 371         if (elapsed) {
 372                 for_1s = (td->io_log_highmark - td->io_log_current) * 1000000000 / elapsed;
 373                 items_to_fetch = for_1s - td->io_log_current;
 374                 if (items_to_fetch < 0)
 375                         items_to_fetch = 0;
 376         } else
 377                 items_to_fetch = 0;
 378
 379         td->io_log_highmark = td->io_log_current + items_to_fetch;
 380         td->io_log_checkmark = (td->io_log_highmark + 1) / 2;
 381         fio_gettime(&td->io_log_highmark_time, NULL);
 382
 383         return items_to_fetch;
 384 }
 385
 386 /*
 387  * Read version 2 iolog data. It is enhanced to include per-file logging,
 388  * syncs, etc.
 389  */
 390 static bool read_iolog2(struct thread_data *td)
 391 {
 392         unsigned long long offset;
 393         unsigned int bytes;
 394         int reads, writes, waits, fileno = 0, file_action = 0; /* stupid gcc */
 395         char *rfname, *fname, *act;
 396         char *str, *p;
 397         enum fio_ddir rw;
 398         bool realloc = false;
 399         int64_t items_to_fetch = 0;
 400
 401         if (td->o.read_iolog_chunked) {
 402                 items_to_fetch = iolog_items_to_fetch(td);
 403                 if (!items_to_fetch)
 404                         return true;
 405         }
 406
 407         /*
 408          * Read in the read iolog and store it, reuse the infrastructure
 409          * for doing verifications.
 410          */
 411         str = malloc(4096);
 412         rfname = fname = malloc(256+16);
 413         act = malloc(256+16);
 414
 415         reads = writes = waits = 0;
 416         while ((p = fgets(str, 4096, td->io_log_rfile)) != NULL) {
 417                 struct io_piece *ipo;
 418                 int r;
 419
 420                 r = sscanf(p, "%256s %256s %llu %u", rfname, act, &offset,
 421                                                                         &bytes);
 422
 423                 if (td->o.replay_redirect)
 424                         fname = td->o.replay_redirect;
 425
 426                 if (r == 4) {
 427                         /*
 428                          * Check action first
 429                          */
 430                         if (!strcmp(act, "wait"))
 431                                 rw = DDIR_WAIT;
 432                         else if (!strcmp(act, "read"))
 433                                 rw = DDIR_READ;
 434                         else if (!strcmp(act, "write"))
 435                                 rw = DDIR_WRITE;
 436                         else if (!strcmp(act, "sync"))
 437                                 rw = DDIR_SYNC;
 438                         else if (!strcmp(act, "datasync"))
 439                                 rw = DDIR_DATASYNC;
 440                         else if (!strcmp(act, "trim"))
 441                                 rw = DDIR_TRIM;
 442                         else {
 443                                 log_err("fio: bad iolog file action: %s\n",
 444                                                                         act);
 445                                 continue;
 446                         }
 447                         fileno = get_fileno(td, fname);
 448                 } else if (r == 2) {
 449                         rw = DDIR_INVAL;
 450                         if (!strcmp(act, "add")) {
 451                                 if (td->o.replay_redirect &&
 452                                     get_fileno(td, fname) != -1) {
 453                                         dprint(FD_FILE, "iolog: ignoring"
 454                                                 " re-add of file %s\n", fname);
 455                                 } else {
 456                                         fileno = add_file(td, fname, td->subjob_number, 1);
 457                                         file_action = FIO_LOG_ADD_FILE;
 458                                 }
 459                                 continue;
 460                         } else if (!strcmp(act, "open")) {
 461                                 fileno = get_fileno(td, fname);
 462                                 file_action = FIO_LOG_OPEN_FILE;
 463                         } else if (!strcmp(act, "close")) {
 464                                 fileno = get_fileno(td, fname);
 465                                 file_action = FIO_LOG_CLOSE_FILE;
 466                         } else {
 467                                 log_err("fio: bad iolog file action: %s\n",
 468                                                                         act);
 469                                 continue;
 470                         }
 471                 } else {
 472                         log_err("bad iolog2: %s\n", p);
 473                         continue;
 474                 }
 475
 476                 if (rw == DDIR_READ)
 477                         reads++;
 478                 else if (rw == DDIR_WRITE) {
 479                         /*
 480                          * Don't add a write for ro mode
 481                          */
 482                         if (read_only)
 483                                 continue;
 484                         writes++;
 485                 } else if (rw == DDIR_WAIT) {
 486                         if (td->o.no_stall)
 487                                 continue;
 488                         waits++;
 489                 } else if (rw == DDIR_INVAL) {
 490                 } else if (!ddir_sync(rw)) {
 491                         log_err("bad ddir: %d\n", rw);
 492                         continue;
 493                 }
 494
 495                 /*
 496                  * Make note of file
 497                  */
 498                 ipo = calloc(1, sizeof(*ipo));
 499                 init_ipo(ipo);
 500                 ipo->ddir = rw;
 501                 if (rw == DDIR_WAIT) {
 502                         ipo->delay = offset;
 503                 } else {
 504                         if (td->o.replay_scale)
 505                                 ipo->offset = offset / td->o.replay_scale;
 506                         else
 507                                 ipo->offset = offset;
 508                         ipo_bytes_align(td->o.replay_align, ipo);
 509
 510                         ipo->len = bytes;
 511                         if (rw != DDIR_INVAL && bytes > td->o.max_bs[rw]) {
 512                                 realloc = true;
 513                                 td->o.max_bs[rw] = bytes;
 514                         }
 515                         ipo->fileno = fileno;
 516                         ipo->file_action = file_action;
 517                         td->o.size += bytes;
 518                 }
 519
 520                 queue_io_piece(td, ipo);
 521
 522                 if (td->o.read_iolog_chunked) {
 523                         td->io_log_current++;
 524                         items_to_fetch--;
 525                         if (items_to_fetch == 0)
 526                                 break;
 527                 }
 528         }
 529
 530         free(str);
 531         free(act);
 532         free(rfname);
 533
 534         if (td->o.read_iolog_chunked) {
 535                 td->io_log_highmark = td->io_log_current;
 536                 td->io_log_checkmark = (td->io_log_highmark + 1) / 2;
 537                 fio_gettime(&td->io_log_highmark_time, NULL);
 538         }
 539
 540         if (writes && read_only) {
 541                 log_err("fio: <%s> skips replay of %d writes due to"
 542                         " read-only\n", td->o.name, writes);
 543                 writes = 0;
 544         }
 545
 546         if (td->o.read_iolog_chunked) {
 547                 if (td->io_log_current == 0) {
 548                         return false;
 549                 }
 550                 td->o.td_ddir = TD_DDIR_RW;
 551                 if (realloc && td->orig_buffer)
 552                 {
 553                         io_u_quiesce(td);
 554                         free_io_mem(td);
 555                         init_io_u_buffers(td);
 556                 }
 557                 return true;
 558         }
 559
 560         if (!reads && !writes && !waits)
 561                 return false;
 562         else if (reads && !writes)
 563                 td->o.td_ddir = TD_DDIR_READ;
 564         else if (!reads && writes)
 565                 td->o.td_ddir = TD_DDIR_WRITE;
 566         else
 567                 td->o.td_ddir = TD_DDIR_RW;
 568
 569         return true;
 570 }
 571
 572 static bool is_socket(const char *path)
 573 {
 574         struct stat buf;
 575         int r;
 576
 577         r = stat(path, &buf);
 578         if (r == -1)
 579                 return false;
 580
 581         return S_ISSOCK(buf.st_mode);
 582 }
 583
 584 static int open_socket(const char *path)
 585 {
 586         struct sockaddr_un addr;
 587         int ret, fd;
 588
 589         fd = socket(AF_UNIX, SOCK_STREAM, 0);
 590         if (fd < 0)
 591                 return fd;
 592
 593         addr.sun_family = AF_UNIX;
 594         if (snprintf(addr.sun_path, sizeof(addr.sun_path), "%s", path) >=
 595             sizeof(addr.sun_path)) {
 596                 log_err("%s: path name %s is too long for a Unix socket\n",
 597                         __func__, path);
 598         }
 599
 600         ret = connect(fd, (const struct sockaddr *)&addr, strlen(path) + sizeof(addr.sun_family));
 601         if (!ret)
 602                 return fd;
 603
 604         close(fd);
 605         return -1;
 606 }
 607
 608 /*
 609  * open iolog, check version, and call appropriate parser
 610  */
 611 static bool init_iolog_read(struct thread_data *td, char *fname)
 612 {
 613         char buffer[256], *p;
 614         FILE *f = NULL;
 615
 616         dprint(FD_IO, "iolog: name=%s\n", fname);
 617
 618         if (is_socket(fname)) {
 619                 int fd;
 620
 621                 fd = open_socket(fname);
 622                 if (fd >= 0)
 623                         f = fdopen(fd, "r");
 624         } else if (!strcmp(fname, "-")) {
 625                 f = stdin;
 626         } else
 627                 f = fopen(fname, "r");
 628
 629         free(fname);
 630
 631         if (!f) {
 632                 perror("fopen read iolog");
 633                 return false;
 634         }
 635
 636         p = fgets(buffer, sizeof(buffer), f);
 637         if (!p) {
 638                 td_verror(td, errno, "iolog read");
 639                 log_err("fio: unable to read iolog\n");
 640                 fclose(f);
 641                 return false;
 642         }
 643
 644         /*
 645          * version 2 of the iolog stores a specific string as the
 646          * first line, check for that
 647          */
 648         if (!strncmp(iolog_ver2, buffer, strlen(iolog_ver2))) {
 649                 free_release_files(td);
 650                 td->io_log_rfile = f;
 651                 return read_iolog2(td);
 652         }
 653
 654         log_err("fio: iolog version 1 is no longer supported\n");
 655         fclose(f);
 656         return false;
 657 }
 658
 659 /*
 660  * Set up a log for storing io patterns.
 661  */
 662 static bool init_iolog_write(struct thread_data *td)
 663 {
 664         struct fio_file *ff;
 665         FILE *f;
 666         unsigned int i;
 667
 668         f = fopen(td->o.write_iolog_file, "a");
 669         if (!f) {
 670                 perror("fopen write iolog");
 671                 return false;
 672         }
 673
 674         /*
 675          * That's it for writing, setup a log buffer and we're done.
 676           */
 677         td->iolog_f = f;
 678         td->iolog_buf = malloc(8192);
 679         setvbuf(f, td->iolog_buf, _IOFBF, 8192);
 680
 681         /*
 682          * write our version line
 683          */
 684         if (fprintf(f, "%s\n", iolog_ver2) < 0) {
 685                 perror("iolog init\n");
 686                 return false;
 687         }
 688
 689         /*
 690          * add all known files
 691          */
 692         for_each_file(td, ff, i)
 693                 log_file(td, ff, FIO_LOG_ADD_FILE);
 694
 695         return true;
 696 }
 697
 698 bool init_iolog(struct thread_data *td)
 699 {
 700         bool ret;
 701
 702         if (td->o.read_iolog_file) {
 703                 int need_swap;
 704                 char * fname = get_name_by_idx(td->o.read_iolog_file, td->subjob_number);
 705
 706                 /*
 707                  * Check if it's a blktrace file and load that if possible.
 708                  * Otherwise assume it's a normal log file and load that.
 709                  */
 710                 if (is_blktrace(fname, &need_swap)) {
 711                         td->io_log_blktrace = 1;
 712                         ret = load_blktrace(td, fname, need_swap);
 713                 } else {
 714                         td->io_log_blktrace = 0;
 715                         ret = init_iolog_read(td, fname);
 716                 }
 717         } else if (td->o.write_iolog_file)
 718                 ret = init_iolog_write(td);
 719         else
 720                 ret = true;
 721
 722         if (!ret)
 723                 td_verror(td, EINVAL, "failed initializing iolog");
 724
 725         return ret;
 726 }
 727
 728 void setup_log(struct io_log **log, struct log_params *p,
 729                const char *filename)
 730 {
 731         struct io_log *l;
 732         int i;
 733         struct io_u_plat_entry *entry;
 734         struct flist_head *list;
 735
 736         l = scalloc(1, sizeof(*l));
 737         INIT_FLIST_HEAD(&l->io_logs);
 738         l->log_type = p->log_type;
 739         l->log_offset = p->log_offset;
 740         l->log_prio = p->log_prio;
 741         l->log_gz = p->log_gz;
 742         l->log_gz_store = p->log_gz_store;
 743         l->avg_msec = p->avg_msec;
 744         l->hist_msec = p->hist_msec;
 745         l->hist_coarseness = p->hist_coarseness;
 746         l->filename = strdup(filename);
 747         l->td = p->td;
 748
 749         /* Initialize histogram lists for each r/w direction,
 750          * with initial io_u_plat of all zeros:
 751          */
 752         for (i = 0; i < DDIR_RWDIR_CNT; i++) {
 753                 list = &l->hist_window[i].list;
 754                 INIT_FLIST_HEAD(list);
 755                 entry = calloc(1, sizeof(struct io_u_plat_entry));
 756                 flist_add(&entry->list, list);
 757         }
 758
 759         if (l->td && l->td->o.io_submit_mode != IO_MODE_OFFLOAD) {
 760                 unsigned int def_samples = DEF_LOG_ENTRIES;
 761                 struct io_logs *__p;
 762
 763                 __p = calloc(1, sizeof(*l->pending));
 764                 if (l->td->o.iodepth > DEF_LOG_ENTRIES)
 765                         def_samples = roundup_pow2(l->td->o.iodepth);
 766                 __p->max_samples = def_samples;
 767                 __p->log = calloc(__p->max_samples, log_entry_sz(l));
 768                 l->pending = __p;
 769         }
 770
 771         if (l->log_offset)
 772                 l->log_ddir_mask = LOG_OFFSET_SAMPLE_BIT;
 773         if (l->log_prio)
 774                 l->log_ddir_mask |= LOG_PRIO_SAMPLE_BIT;
 775
 776         INIT_FLIST_HEAD(&l->chunk_list);
 777
 778         if (l->log_gz && !p->td)
 779                 l->log_gz = 0;
 780         else if (l->log_gz || l->log_gz_store) {
 781                 mutex_init_pshared(&l->chunk_lock);
 782                 mutex_init_pshared(&l->deferred_free_lock);
 783                 p->td->flags |= TD_F_COMPRESS_LOG;
 784         }
 785
 786         *log = l;
 787 }
 788
 789 #ifdef CONFIG_SETVBUF
 790 static void *set_file_buffer(FILE *f)
 791 {
 792         size_t size = 1048576;
 793         void *buf;
 794
 795         buf = malloc(size);
 796         setvbuf(f, buf, _IOFBF, size);
 797         return buf;
 798 }
 799
 800 static void clear_file_buffer(void *buf)
 801 {
 802         free(buf);
 803 }
 804 #else
 805 static void *set_file_buffer(FILE *f)
 806 {
 807         return NULL;
 808 }
 809
 810 static void clear_file_buffer(void *buf)
 811 {
 812 }
 813 #endif
 814
 815 void free_log(struct io_log *log)
 816 {
 817         while (!flist_empty(&log->io_logs)) {
 818                 struct io_logs *cur_log;
 819
 820                 cur_log = flist_first_entry(&log->io_logs, struct io_logs, list);
 821                 flist_del_init(&cur_log->list);
 822                 free(cur_log->log);
 823                 sfree(cur_log);
 824         }
 825
 826         if (log->pending) {
 827                 free(log->pending->log);
 828                 free(log->pending);
 829                 log->pending = NULL;
 830         }
 831
 832         free(log->pending);
 833         free(log->filename);
 834         sfree(log);
 835 }
 836
 837 uint64_t hist_sum(int j, int stride, uint64_t *io_u_plat,
 838                 uint64_t *io_u_plat_last)
 839 {
 840         uint64_t sum;
 841         int k;
 842
 843         if (io_u_plat_last) {
 844                 for (k = sum = 0; k < stride; k++)
 845                         sum += io_u_plat[j + k] - io_u_plat_last[j + k];
 846         } else {
 847                 for (k = sum = 0; k < stride; k++)
 848                         sum += io_u_plat[j + k];
 849         }
 850
 851         return sum;
 852 }
 853
 854 static void flush_hist_samples(FILE *f, int hist_coarseness, void *samples,
 855                                uint64_t sample_size)
 856 {
 857         struct io_sample *s;
 858         int log_offset;
 859         uint64_t i, j, nr_samples;
 860         struct io_u_plat_entry *entry, *entry_before;
 861         uint64_t *io_u_plat;
 862         uint64_t *io_u_plat_before;
 863
 864         int stride = 1 << hist_coarseness;
 865
 866         if (!sample_size)
 867                 return;
 868
 869         s = __get_sample(samples, 0, 0);
 870         log_offset = (s->__ddir & LOG_OFFSET_SAMPLE_BIT) != 0;
 871
 872         nr_samples = sample_size / __log_entry_sz(log_offset);
 873
 874         for (i = 0; i < nr_samples; i++) {
 875                 s = __get_sample(samples, log_offset, i);
 876
 877                 entry = s->data.plat_entry;
 878                 io_u_plat = entry->io_u_plat;
 879
 880                 entry_before = flist_first_entry(&entry->list, struct io_u_plat_entry, list);
 881                 io_u_plat_before = entry_before->io_u_plat;
 882
 883                 fprintf(f, "%lu, %u, %llu, ", (unsigned long) s->time,
 884                                                 io_sample_ddir(s), (unsigned long long) s->bs);
 885                 for (j = 0; j < FIO_IO_U_PLAT_NR - stride; j += stride) {
 886                         fprintf(f, "%llu, ", (unsigned long long)
 887                                 hist_sum(j, stride, io_u_plat, io_u_plat_before));
 888                 }
 889                 fprintf(f, "%llu\n", (unsigned long long)
 890                         hist_sum(FIO_IO_U_PLAT_NR - stride, stride, io_u_plat,
 891                                         io_u_plat_before));
 892
 893                 flist_del(&entry_before->list);
 894                 free(entry_before);
 895         }
 896 }
 897
 898 void flush_samples(FILE *f, void *samples, uint64_t sample_size)
 899 {
 900         struct io_sample *s;
 901         int log_offset, log_prio;
 902         uint64_t i, nr_samples;
 903         unsigned int prio_val;
 904         const char *fmt;
 905
 906         if (!sample_size)
 907                 return;
 908
 909         s = __get_sample(samples, 0, 0);
 910         log_offset = (s->__ddir & LOG_OFFSET_SAMPLE_BIT) != 0;
 911         log_prio = (s->__ddir & LOG_PRIO_SAMPLE_BIT) != 0;
 912
 913         if (log_offset) {
 914                 if (log_prio)
 915                         fmt = "%lu, %" PRId64 ", %u, %llu, %llu, 0x%04x\n";
 916                 else
 917                         fmt = "%lu, %" PRId64 ", %u, %llu, %llu, %u\n";
 918         } else {
 919                 if (log_prio)
 920                         fmt = "%lu, %" PRId64 ", %u, %llu, 0x%04x\n";
 921                 else
 922                         fmt = "%lu, %" PRId64 ", %u, %llu, %u\n";
 923         }
 924
 925         nr_samples = sample_size / __log_entry_sz(log_offset);
 926
 927         for (i = 0; i < nr_samples; i++) {
 928                 s = __get_sample(samples, log_offset, i);
 929
 930                 if (log_prio)
 931                         prio_val = s->priority;
 932                 else
 933                         prio_val = ioprio_value_is_class_rt(s->priority);
 934
 935                 if (!log_offset) {
 936                         fprintf(f, fmt,
 937                                 (unsigned long) s->time,
 938                                 s->data.val,
 939                                 io_sample_ddir(s), (unsigned long long) s->bs,
 940                                 prio_val);
 941                 } else {
 942                         struct io_sample_offset *so = (void *) s;
 943
 944                         fprintf(f, fmt,
 945                                 (unsigned long) s->time,
 946                                 s->data.val,
 947                                 io_sample_ddir(s), (unsigned long long) s->bs,
 948                                 (unsigned long long) so->offset,
 949                                 prio_val);
 950                 }
 951         }
 952 }
 953
 954 #ifdef CONFIG_ZLIB
 955
 956 struct iolog_flush_data {
 957         struct workqueue_work work;
 958         struct io_log *log;
 959         void *samples;
 960         uint32_t nr_samples;
 961         bool free;
 962 };
 963
 964 #define GZ_CHUNK        131072
 965
 966 static struct iolog_compress *get_new_chunk(unsigned int seq)
 967 {
 968         struct iolog_compress *c;
 969
 970         c = malloc(sizeof(*c));
 971         INIT_FLIST_HEAD(&c->list);
 972         c->buf = malloc(GZ_CHUNK);
 973         c->len = 0;
 974         c->seq = seq;
 975         return c;
 976 }
 977
 978 static void free_chunk(struct iolog_compress *ic)
 979 {
 980         free(ic->buf);
 981         free(ic);
 982 }
 983
 984 static int z_stream_init(z_stream *stream, int gz_hdr)
 985 {
 986         int wbits = 15;
 987
 988         memset(stream, 0, sizeof(*stream));
 989         stream->zalloc = Z_NULL;
 990         stream->zfree = Z_NULL;
 991         stream->opaque = Z_NULL;
 992         stream->next_in = Z_NULL;
 993
 994         /*
 995          * zlib magic - add 32 for auto-detection of gz header or not,
 996          * if we decide to store files in a gzip friendly format.
 997          */
 998         if (gz_hdr)
 999                 wbits += 32;
1000
1001         if (inflateInit2(stream, wbits) != Z_OK)
1002                 return 1;
1003
1004         return 0;
1005 }
1006
1007 struct inflate_chunk_iter {
1008         unsigned int seq;
1009         int err;
1010         void *buf;
1011         size_t buf_size;
1012         size_t buf_used;
1013         size_t chunk_sz;
1014 };
1015
1016 static void finish_chunk(z_stream *stream, FILE *f,
1017                          struct inflate_chunk_iter *iter)
1018 {
1019         int ret;
1020
1021         ret = inflateEnd(stream);
1022         if (ret != Z_OK)
1023                 log_err("fio: failed to end log inflation seq %d (%d)\n",
1024                                 iter->seq, ret);
1025
1026         flush_samples(f, iter->buf, iter->buf_used);
1027         free(iter->buf);
1028         iter->buf = NULL;
1029         iter->buf_size = iter->buf_used = 0;
1030 }
1031
1032 /*
1033  * Iterative chunk inflation. Handles cases where we cross into a new
1034  * sequence, doing flush finish of previous chunk if needed.
1035  */
1036 static size_t inflate_chunk(struct iolog_compress *ic, int gz_hdr, FILE *f,
1037                             z_stream *stream, struct inflate_chunk_iter *iter)
1038 {
1039         size_t ret;
1040
1041         dprint(FD_COMPRESS, "inflate chunk size=%lu, seq=%u\n",
1042                                 (unsigned long) ic->len, ic->seq);
1043
1044         if (ic->seq != iter->seq) {
1045                 if (iter->seq)
1046                         finish_chunk(stream, f, iter);
1047
1048                 z_stream_init(stream, gz_hdr);
1049                 iter->seq = ic->seq;
1050         }
1051
1052         stream->avail_in = ic->len;
1053         stream->next_in = ic->buf;
1054
1055         if (!iter->buf_size) {
1056                 iter->buf_size = iter->chunk_sz;
1057                 iter->buf = malloc(iter->buf_size);
1058         }
1059
1060         while (stream->avail_in) {
1061                 size_t this_out = iter->buf_size - iter->buf_used;
1062                 int err;
1063
1064                 stream->avail_out = this_out;
1065                 stream->next_out = iter->buf + iter->buf_used;
1066
1067                 err = inflate(stream, Z_NO_FLUSH);
1068                 if (err < 0) {
1069                         log_err("fio: failed inflating log: %d\n", err);
1070                         iter->err = err;
1071                         break;
1072                 }
1073
1074                 iter->buf_used += this_out - stream->avail_out;
1075
1076                 if (!stream->avail_out) {
1077                         iter->buf_size += iter->chunk_sz;
1078                         iter->buf = realloc(iter->buf, iter->buf_size);
1079                         continue;
1080                 }
1081
1082                 if (err == Z_STREAM_END)
1083                         break;
1084         }
1085
1086         ret = (void *) stream->next_in - ic->buf;
1087
1088         dprint(FD_COMPRESS, "inflated to size=%lu\n", (unsigned long) iter->buf_size);
1089
1090         return ret;
1091 }
1092
1093 /*
1094  * Inflate stored compressed chunks, or write them directly to the log
1095  * file if so instructed.
1096  */
1097 static int inflate_gz_chunks(struct io_log *log, FILE *f)
1098 {
1099         struct inflate_chunk_iter iter = { .chunk_sz = log->log_gz, };
1100         z_stream stream;
1101
1102         while (!flist_empty(&log->chunk_list)) {
1103                 struct iolog_compress *ic;
1104
1105                 ic = flist_first_entry(&log->chunk_list, struct iolog_compress, list);
1106                 flist_del(&ic->list);
1107
1108                 if (log->log_gz_store) {
1109                         size_t ret;
1110
1111                         dprint(FD_COMPRESS, "log write chunk size=%lu, "
1112                                 "seq=%u\n", (unsigned long) ic->len, ic->seq);
1113
1114                         ret = fwrite(ic->buf, ic->len, 1, f);
1115                         if (ret != 1 || ferror(f)) {
1116                                 iter.err = errno;
1117                                 log_err("fio: error writing compressed log\n");
1118                         }
1119                 } else
1120                         inflate_chunk(ic, log->log_gz_store, f, &stream, &iter);
1121
1122                 free_chunk(ic);
1123         }
1124
1125         if (iter.seq) {
1126                 finish_chunk(&stream, f, &iter);
1127                 free(iter.buf);
1128         }
1129
1130         return iter.err;
1131 }
1132
1133 /*
1134  * Open compressed log file and decompress the stored chunks and
1135  * write them to stdout. The chunks are stored sequentially in the
1136  * file, so we iterate over them and do them one-by-one.
1137  */
1138 int iolog_file_inflate(const char *file)
1139 {
1140         struct inflate_chunk_iter iter = { .chunk_sz = 64 * 1024 * 1024, };
1141         struct iolog_compress ic;
1142         z_stream stream;
1143         struct stat sb;
1144         size_t ret;
1145         size_t total;
1146         void *buf;
1147         FILE *f;
1148
1149         f = fopen(file, "r");
1150         if (!f) {
1151                 perror("fopen");
1152                 return 1;
1153         }
1154
1155         if (stat(file, &sb) < 0) {
1156                 fclose(f);
1157                 perror("stat");
1158                 return 1;
1159         }
1160
1161         ic.buf = buf = malloc(sb.st_size);
1162         ic.len = sb.st_size;
1163         ic.seq = 1;
1164
1165         ret = fread(ic.buf, ic.len, 1, f);
1166         if (ret == 0 && ferror(f)) {
1167                 perror("fread");
1168                 fclose(f);
1169                 free(buf);
1170                 return 1;
1171         } else if (ferror(f) || (!feof(f) && ret != 1)) {
1172                 log_err("fio: short read on reading log\n");
1173                 fclose(f);
1174                 free(buf);
1175                 return 1;
1176         }
1177
1178         fclose(f);
1179
1180         /*
1181          * Each chunk will return Z_STREAM_END. We don't know how many
1182          * chunks are in the file, so we just keep looping and incrementing
1183          * the sequence number until we have consumed the whole compressed
1184          * file.
1185          */
1186         total = ic.len;
1187         do {
1188                 size_t iret;
1189
1190                 iret = inflate_chunk(&ic,  1, stdout, &stream, &iter);
1191                 total -= iret;
1192                 if (!total)
1193                         break;
1194                 if (iter.err)
1195                         break;
1196
1197                 ic.seq++;
1198                 ic.len -= iret;
1199                 ic.buf += iret;
1200         } while (1);
1201
1202         if (iter.seq) {
1203                 finish_chunk(&stream, stdout, &iter);
1204                 free(iter.buf);
1205         }
1206
1207         free(buf);
1208         return iter.err;
1209 }
1210
1211 #else
1212
1213 static int inflate_gz_chunks(struct io_log *log, FILE *f)
1214 {
1215         return 0;
1216 }
1217
1218 int iolog_file_inflate(const char *file)
1219 {
1220         log_err("fio: log inflation not possible without zlib\n");
1221         return 1;
1222 }
1223
1224 #endif
1225
1226 void flush_log(struct io_log *log, bool do_append)
1227 {
1228         void *buf;
1229         FILE *f;
1230
1231         if (!do_append)
1232                 f = fopen(log->filename, "w");
1233         else
1234                 f = fopen(log->filename, "a");
1235         if (!f) {
1236                 perror("fopen log");
1237                 return;
1238         }
1239
1240         buf = set_file_buffer(f);
1241
1242         inflate_gz_chunks(log, f);
1243
1244         while (!flist_empty(&log->io_logs)) {
1245                 struct io_logs *cur_log;
1246
1247                 cur_log = flist_first_entry(&log->io_logs, struct io_logs, list);
1248                 flist_del_init(&cur_log->list);
1249
1250                 if (log->td && log == log->td->clat_hist_log)
1251                         flush_hist_samples(f, log->hist_coarseness, cur_log->log,
1252                                            log_sample_sz(log, cur_log));
1253                 else
1254                         flush_samples(f, cur_log->log, log_sample_sz(log, cur_log));
1255
1256                 sfree(cur_log);
1257         }
1258
1259         fclose(f);
1260         clear_file_buffer(buf);
1261 }
1262
1263 static int finish_log(struct thread_data *td, struct io_log *log, int trylock)
1264 {
1265         if (td->flags & TD_F_COMPRESS_LOG)
1266                 iolog_flush(log);
1267
1268         if (trylock) {
1269                 if (fio_trylock_file(log->filename))
1270                         return 1;
1271         } else
1272                 fio_lock_file(log->filename);
1273
1274         if (td->client_type == FIO_CLIENT_TYPE_GUI || is_backend)
1275                 fio_send_iolog(td, log, log->filename);
1276         else
1277                 flush_log(log, !td->o.per_job_logs);
1278
1279         fio_unlock_file(log->filename);
1280         free_log(log);
1281         return 0;
1282 }
1283
1284 size_t log_chunk_sizes(struct io_log *log)
1285 {
1286         struct flist_head *entry;
1287         size_t ret;
1288
1289         if (flist_empty(&log->chunk_list))
1290                 return 0;
1291
1292         ret = 0;
1293         pthread_mutex_lock(&log->chunk_lock);
1294         flist_for_each(entry, &log->chunk_list) {
1295                 struct iolog_compress *c;
1296
1297                 c = flist_entry(entry, struct iolog_compress, list);
1298                 ret += c->len;
1299         }
1300         pthread_mutex_unlock(&log->chunk_lock);
1301         return ret;
1302 }
1303
1304 #ifdef CONFIG_ZLIB
1305
1306 static void iolog_put_deferred(struct io_log *log, void *ptr)
1307 {
1308         if (!ptr)
1309                 return;
1310
1311         pthread_mutex_lock(&log->deferred_free_lock);
1312         if (log->deferred < IOLOG_MAX_DEFER) {
1313                 log->deferred_items[log->deferred] = ptr;
1314                 log->deferred++;
1315         } else if (!fio_did_warn(FIO_WARN_IOLOG_DROP))
1316                 log_err("fio: had to drop log entry free\n");
1317         pthread_mutex_unlock(&log->deferred_free_lock);
1318 }
1319
1320 static void iolog_free_deferred(struct io_log *log)
1321 {
1322         int i;
1323
1324         if (!log->deferred)
1325                 return;
1326
1327         pthread_mutex_lock(&log->deferred_free_lock);
1328
1329         for (i = 0; i < log->deferred; i++) {
1330                 free(log->deferred_items[i]);
1331                 log->deferred_items[i] = NULL;
1332         }
1333
1334         log->deferred = 0;
1335         pthread_mutex_unlock(&log->deferred_free_lock);
1336 }
1337
1338 static int gz_work(struct iolog_flush_data *data)
1339 {
1340         struct iolog_compress *c = NULL;
1341         struct flist_head list;
1342         unsigned int seq;
1343         z_stream stream;
1344         size_t total = 0;
1345         int ret;
1346
1347         INIT_FLIST_HEAD(&list);
1348
1349         memset(&stream, 0, sizeof(stream));
1350         stream.zalloc = Z_NULL;
1351         stream.zfree = Z_NULL;
1352         stream.opaque = Z_NULL;
1353
1354         ret = deflateInit(&stream, Z_DEFAULT_COMPRESSION);
1355         if (ret != Z_OK) {
1356                 log_err("fio: failed to init gz stream\n");
1357                 goto err;
1358         }
1359
1360         seq = ++data->log->chunk_seq;
1361
1362         stream.next_in = (void *) data->samples;
1363         stream.avail_in = data->nr_samples * log_entry_sz(data->log);
1364
1365         dprint(FD_COMPRESS, "deflate input size=%lu, seq=%u, log=%s\n",
1366                                 (unsigned long) stream.avail_in, seq,
1367                                 data->log->filename);
1368         do {
1369                 if (c)
1370                         dprint(FD_COMPRESS, "seq=%d, chunk=%lu\n", seq,
1371                                 (unsigned long) c->len);
1372                 c = get_new_chunk(seq);
1373                 stream.avail_out = GZ_CHUNK;
1374                 stream.next_out = c->buf;
1375                 ret = deflate(&stream, Z_NO_FLUSH);
1376                 if (ret < 0) {
1377                         log_err("fio: deflate log (%d)\n", ret);
1378                         free_chunk(c);
1379                         goto err;
1380                 }
1381
1382                 c->len = GZ_CHUNK - stream.avail_out;
1383                 flist_add_tail(&c->list, &list);
1384                 total += c->len;
1385         } while (stream.avail_in);
1386
1387         stream.next_out = c->buf + c->len;
1388         stream.avail_out = GZ_CHUNK - c->len;
1389
1390         ret = deflate(&stream, Z_FINISH);
1391         if (ret < 0) {
1392                 /*
1393                  * Z_BUF_ERROR is special, it just means we need more
1394                  * output space. We'll handle that below. Treat any other
1395                  * error as fatal.
1396                  */
1397                 if (ret != Z_BUF_ERROR) {
1398                         log_err("fio: deflate log (%d)\n", ret);
1399                         flist_del(&c->list);
1400                         free_chunk(c);
1401                         goto err;
1402                 }
1403         }
1404
1405         total -= c->len;
1406         c->len = GZ_CHUNK - stream.avail_out;
1407         total += c->len;
1408         dprint(FD_COMPRESS, "seq=%d, chunk=%lu\n", seq, (unsigned long) c->len);
1409
1410         if (ret != Z_STREAM_END) {
1411                 do {
1412                         c = get_new_chunk(seq);
1413                         stream.avail_out = GZ_CHUNK;
1414                         stream.next_out = c->buf;
1415                         ret = deflate(&stream, Z_FINISH);
1416                         c->len = GZ_CHUNK - stream.avail_out;
1417                         total += c->len;
1418                         flist_add_tail(&c->list, &list);
1419                         dprint(FD_COMPRESS, "seq=%d, chunk=%lu\n", seq,
1420                                 (unsigned long) c->len);
1421                 } while (ret != Z_STREAM_END);
1422         }
1423
1424         dprint(FD_COMPRESS, "deflated to size=%lu\n", (unsigned long) total);
1425
1426         ret = deflateEnd(&stream);
1427         if (ret != Z_OK)
1428                 log_err("fio: deflateEnd %d\n", ret);
1429
1430         iolog_put_deferred(data->log, data->samples);
1431
1432         if (!flist_empty(&list)) {
1433                 pthread_mutex_lock(&data->log->chunk_lock);
1434                 flist_splice_tail(&list, &data->log->chunk_list);
1435                 pthread_mutex_unlock(&data->log->chunk_lock);
1436         }
1437
1438         ret = 0;
1439 done:
1440         if (data->free)
1441                 sfree(data);
1442         return ret;
1443 err:
1444         while (!flist_empty(&list)) {
1445                 c = flist_first_entry(list.next, struct iolog_compress, list);
1446                 flist_del(&c->list);
1447                 free_chunk(c);
1448         }
1449         ret = 1;
1450         goto done;
1451 }
1452
1453 /*
1454  * Invoked from our compress helper thread, when logging would have exceeded
1455  * the specified memory limitation. Compresses the previously stored
1456  * entries.
1457  */
1458 static int gz_work_async(struct submit_worker *sw, struct workqueue_work *work)
1459 {
1460         return gz_work(container_of(work, struct iolog_flush_data, work));
1461 }
1462
1463 static int gz_init_worker(struct submit_worker *sw)
1464 {
1465         struct thread_data *td = sw->wq->td;
1466
1467         if (!fio_option_is_set(&td->o, log_gz_cpumask))
1468                 return 0;
1469
1470         if (fio_setaffinity(gettid(), td->o.log_gz_cpumask) == -1) {
1471                 log_err("gz: failed to set CPU affinity\n");
1472                 return 1;
1473         }
1474
1475         return 0;
1476 }
1477
1478 static struct workqueue_ops log_compress_wq_ops = {
1479         .fn             = gz_work_async,
1480         .init_worker_fn = gz_init_worker,
1481         .nice           = 1,
1482 };
1483
1484 int iolog_compress_init(struct thread_data *td, struct sk_out *sk_out)
1485 {
1486         if (!(td->flags & TD_F_COMPRESS_LOG))
1487                 return 0;
1488
1489         workqueue_init(td, &td->log_compress_wq, &log_compress_wq_ops, 1, sk_out);
1490         return 0;
1491 }
1492
1493 void iolog_compress_exit(struct thread_data *td)
1494 {
1495         if (!(td->flags & TD_F_COMPRESS_LOG))
1496                 return;
1497
1498         workqueue_exit(&td->log_compress_wq);
1499 }
1500
1501 /*
1502  * Queue work item to compress the existing log entries. We reset the
1503  * current log to a small size, and reference the existing log in the
1504  * data that we queue for compression. Once compression has been done,
1505  * this old log is freed. If called with finish == true, will not return
1506  * until the log compression has completed, and will flush all previous
1507  * logs too
1508  */
1509 static int iolog_flush(struct io_log *log)
1510 {
1511         struct iolog_flush_data *data;
1512
1513         data = malloc(sizeof(*data));
1514         if (!data)
1515                 return 1;
1516
1517         data->log = log;
1518         data->free = false;
1519
1520         while (!flist_empty(&log->io_logs)) {
1521                 struct io_logs *cur_log;
1522
1523                 cur_log = flist_first_entry(&log->io_logs, struct io_logs, list);
1524                 flist_del_init(&cur_log->list);
1525
1526                 data->samples = cur_log->log;
1527                 data->nr_samples = cur_log->nr_samples;
1528
1529                 sfree(cur_log);
1530
1531                 gz_work(data);
1532         }
1533
1534         free(data);
1535         return 0;
1536 }
1537
1538 int iolog_cur_flush(struct io_log *log, struct io_logs *cur_log)
1539 {
1540         struct iolog_flush_data *data;
1541
1542         data = smalloc(sizeof(*data));
1543         if (!data)
1544                 return 1;
1545
1546         data->log = log;
1547
1548         data->samples = cur_log->log;
1549         data->nr_samples = cur_log->nr_samples;
1550         data->free = true;
1551
1552         cur_log->nr_samples = cur_log->max_samples = 0;
1553         cur_log->log = NULL;
1554
1555         workqueue_enqueue(&log->td->log_compress_wq, &data->work);
1556
1557         iolog_free_deferred(log);
1558
1559         return 0;
1560 }
1561 #else
1562
1563 static int iolog_flush(struct io_log *log)
1564 {
1565         return 1;
1566 }
1567
1568 int iolog_cur_flush(struct io_log *log, struct io_logs *cur_log)
1569 {
1570         return 1;
1571 }
1572
1573 int iolog_compress_init(struct thread_data *td, struct sk_out *sk_out)
1574 {
1575         return 0;
1576 }
1577
1578 void iolog_compress_exit(struct thread_data *td)
1579 {
1580 }
1581
1582 #endif
1583
1584 struct io_logs *iolog_cur_log(struct io_log *log)
1585 {
1586         if (flist_empty(&log->io_logs))
1587                 return NULL;
1588
1589         return flist_last_entry(&log->io_logs, struct io_logs, list);
1590 }
1591
1592 uint64_t iolog_nr_samples(struct io_log *iolog)
1593 {
1594         struct flist_head *entry;
1595         uint64_t ret = 0;
1596
1597         flist_for_each(entry, &iolog->io_logs) {
1598                 struct io_logs *cur_log;
1599
1600                 cur_log = flist_entry(entry, struct io_logs, list);
1601                 ret += cur_log->nr_samples;
1602         }
1603
1604         return ret;
1605 }
1606
1607 static int __write_log(struct thread_data *td, struct io_log *log, int try)
1608 {
1609         if (log)
1610                 return finish_log(td, log, try);
1611
1612         return 0;
1613 }
1614
1615 static int write_iops_log(struct thread_data *td, int try, bool unit_log)
1616 {
1617         int ret;
1618
1619         if (per_unit_log(td->iops_log) != unit_log)
1620                 return 0;
1621
1622         ret = __write_log(td, td->iops_log, try);
1623         if (!ret)
1624                 td->iops_log = NULL;
1625
1626         return ret;
1627 }
1628
1629 static int write_slat_log(struct thread_data *td, int try, bool unit_log)
1630 {
1631         int ret;
1632
1633         if (!unit_log)
1634                 return 0;
1635
1636         ret = __write_log(td, td->slat_log, try);
1637         if (!ret)
1638                 td->slat_log = NULL;
1639
1640         return ret;
1641 }
1642
1643 static int write_clat_log(struct thread_data *td, int try, bool unit_log)
1644 {
1645         int ret;
1646
1647         if (!unit_log)
1648                 return 0;
1649
1650         ret = __write_log(td, td->clat_log, try);
1651         if (!ret)
1652                 td->clat_log = NULL;
1653
1654         return ret;
1655 }
1656
1657 static int write_clat_hist_log(struct thread_data *td, int try, bool unit_log)
1658 {
1659         int ret;
1660
1661         if (!unit_log)
1662                 return 0;
1663
1664         ret = __write_log(td, td->clat_hist_log, try);
1665         if (!ret)
1666                 td->clat_hist_log = NULL;
1667
1668         return ret;
1669 }
1670
1671 static int write_lat_log(struct thread_data *td, int try, bool unit_log)
1672 {
1673         int ret;
1674
1675         if (!unit_log)
1676                 return 0;
1677
1678         ret = __write_log(td, td->lat_log, try);
1679         if (!ret)
1680                 td->lat_log = NULL;
1681
1682         return ret;
1683 }
1684
1685 static int write_bandw_log(struct thread_data *td, int try, bool unit_log)
1686 {
1687         int ret;
1688
1689         if (per_unit_log(td->bw_log) != unit_log)
1690                 return 0;
1691
1692         ret = __write_log(td, td->bw_log, try);
1693         if (!ret)
1694                 td->bw_log = NULL;
1695
1696         return ret;
1697 }
1698
1699 enum {
1700         BW_LOG_MASK     = 1,
1701         LAT_LOG_MASK    = 2,
1702         SLAT_LOG_MASK   = 4,
1703         CLAT_LOG_MASK   = 8,
1704         IOPS_LOG_MASK   = 16,
1705         CLAT_HIST_LOG_MASK = 32,
1706
1707         ALL_LOG_NR      = 6,
1708 };
1709
1710 struct log_type {
1711         unsigned int mask;
1712         int (*fn)(struct thread_data *, int, bool);
1713 };
1714
1715 static struct log_type log_types[] = {
1716         {
1717                 .mask   = BW_LOG_MASK,
1718                 .fn     = write_bandw_log,
1719         },
1720         {
1721                 .mask   = LAT_LOG_MASK,
1722                 .fn     = write_lat_log,
1723         },
1724         {
1725                 .mask   = SLAT_LOG_MASK,
1726                 .fn     = write_slat_log,
1727         },
1728         {
1729                 .mask   = CLAT_LOG_MASK,
1730                 .fn     = write_clat_log,
1731         },
1732         {
1733                 .mask   = IOPS_LOG_MASK,
1734                 .fn     = write_iops_log,
1735         },
1736         {
1737                 .mask   = CLAT_HIST_LOG_MASK,
1738                 .fn     = write_clat_hist_log,
1739         }
1740 };
1741
1742 void td_writeout_logs(struct thread_data *td, bool unit_logs)
1743 {
1744         unsigned int log_mask = 0;
1745         unsigned int log_left = ALL_LOG_NR;
1746         int old_state, i;
1747
1748         old_state = td_bump_runstate(td, TD_FINISHING);
1749
1750         finalize_logs(td, unit_logs);
1751
1752         while (log_left) {
1753                 int prev_log_left = log_left;
1754
1755                 for (i = 0; i < ALL_LOG_NR && log_left; i++) {
1756                         struct log_type *lt = &log_types[i];
1757                         int ret;
1758
1759                         if (!(log_mask & lt->mask)) {
1760                                 ret = lt->fn(td, log_left != 1, unit_logs);
1761                                 if (!ret) {
1762                                         log_left--;
1763                                         log_mask |= lt->mask;
1764                                 }
1765                         }
1766                 }
1767
1768                 if (prev_log_left == log_left)
1769                         usleep(5000);
1770         }
1771
1772         td_restore_runstate(td, old_state);
1773 }
1774
1775 void fio_writeout_logs(bool unit_logs)
1776 {
1777         struct thread_data *td;
1778         int i;
1779
1780         for_each_td(td, i)
1781                 td_writeout_logs(td, unit_logs);
1782 }