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