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