iolog.c

   1 /*
   2  * Code related to writing an iolog of what a thread is doing, and to
   3  * later read that back and replay
   4  */
   5 #include <stdio.h>
   6 #include <stdlib.h>
   7 #include <assert.h>
   8 #include <sys/types.h>
   9 #include <sys/stat.h>
  10 #include <unistd.h>
  11 #ifdef CONFIG_ZLIB
  12 #include <zlib.h>
  13 #endif
  14
  15 #include "flist.h"
  16 #include "fio.h"
  17 #include "trim.h"
  18 #include "filelock.h"
  19 #include "smalloc.h"
  20 #include "blktrace.h"
  21 #include "pshared.h"
  22 #include "lib/roundup.h"
  23
  24 #include <netinet/in.h>
  25 #include <netinet/tcp.h>
  26 #include <arpa/inet.h>
  27 #include <sys/stat.h>
  28 #include <sys/socket.h>
  29 #include <sys/un.h>
  30
  31 static int iolog_flush(struct io_log *log);
  32
  33 static const char iolog_ver2[] = "fio version 2 iolog";
  34
  35 void queue_io_piece(struct thread_data *td, struct io_piece *ipo)
  36 {
  37         flist_add_tail(&ipo->list, &td->io_log_list);
  38         td->total_io_size += ipo->len;
  39 }
  40
  41 void log_io_u(const struct thread_data *td, const struct io_u *io_u)
  42 {
  43         if (!td->o.write_iolog_file)
  44                 return;
  45
  46         fprintf(td->iolog_f, "%s %s %llu %llu\n", io_u->file->file_name,
  47                                                 io_ddir_name(io_u->ddir),
  48                                                 io_u->offset, io_u->buflen);
  49 }
  50
  51 void log_file(struct thread_data *td, struct fio_file *f,
  52               enum file_log_act what)
  53 {
  54         const char *act[] = { "add", "open", "close" };
  55
  56         assert(what < 3);
  57
  58         if (!td->o.write_iolog_file)
  59                 return;
  60
  61
  62         /*
  63          * this happens on the pre-open/close done before the job starts
  64          */
  65         if (!td->iolog_f)
  66                 return;
  67
  68         fprintf(td->iolog_f, "%s %s\n", f->file_name, act[what]);
  69 }
  70
  71 static void iolog_delay(struct thread_data *td, unsigned long delay)
  72 {
  73         uint64_t usec = utime_since_now(&td->last_issue);
  74         unsigned long orig_delay = delay;
  75         uint64_t this_delay;
  76         struct timespec ts;
  77
  78         if (delay < td->time_offset) {
  79                 td->time_offset = 0;
  80                 return;
  81         }
  82
  83         delay -= td->time_offset;
  84         if (delay < usec)
  85                 return;
  86
  87         delay -= usec;
  88
  89         fio_gettime(&ts, NULL);
  90         while (delay && !td->terminate) {
  91                 this_delay = delay;
  92                 if (this_delay > 500000)
  93                         this_delay = 500000;
  94
  95                 usec_sleep(td, this_delay);
  96                 delay -= this_delay;
  97         }
  98
  99         usec = utime_since_now(&ts);
 100         if (usec > orig_delay)
 101                 td->time_offset = usec - orig_delay;
 102         else
 103                 td->time_offset = 0;
 104 }
 105
 106 static int ipo_special(struct thread_data *td, struct io_piece *ipo)
 107 {
 108         struct fio_file *f;
 109         int ret;
 110
 111         /*
 112          * Not a special ipo
 113          */
 114         if (ipo->ddir != DDIR_INVAL)
 115                 return 0;
 116
 117         f = td->files[ipo->fileno];
 118
 119         switch (ipo->file_action) {
 120         case FIO_LOG_OPEN_FILE:
 121                 if (td->o.replay_redirect && fio_file_open(f)) {
 122                         dprint(FD_FILE, "iolog: ignoring re-open of file %s\n",
 123                                         f->file_name);
 124                         break;
 125                 }
 126                 ret = td_io_open_file(td, f);
 127                 if (!ret)
 128                         break;
 129                 td_verror(td, ret, "iolog open file");
 130                 return -1;
 131         case FIO_LOG_CLOSE_FILE:
 132                 td_io_close_file(td, f);
 133                 break;
 134         case FIO_LOG_UNLINK_FILE:
 135                 td_io_unlink_file(td, f);
 136                 break;
 137         default:
 138                 log_err("fio: bad file action %d\n", ipo->file_action);
 139                 break;
 140         }
 141
 142         return 1;
 143 }
 144
 145 static bool read_iolog2(struct thread_data *td);
 146
 147 int read_iolog_get(struct thread_data *td, struct io_u *io_u)
 148 {
 149         struct io_piece *ipo;
 150         unsigned long elapsed;
 151
 152         while (!flist_empty(&td->io_log_list)) {
 153                 int ret;
 154
 155                 if (!td->io_log_blktrace && td->o.read_iolog_chunked) {
 156                         if (td->io_log_checkmark == td->io_log_current) {
 157                                 if (!read_iolog2(td))
 158                                         return 1;
 159                         }
 160                         td->io_log_current--;
 161                 }
 162                 ipo = flist_first_entry(&td->io_log_list, struct io_piece, list);
 163                 flist_del(&ipo->list);
 164                 remove_trim_entry(td, ipo);
 165
 166                 ret = ipo_special(td, ipo);
 167                 if (ret < 0) {
 168                         free(ipo);
 169                         break;
 170                 } else if (ret > 0) {
 171                         free(ipo);
 172                         continue;
 173                 }
 174
 175                 io_u->ddir = ipo->ddir;
 176                 if (ipo->ddir != DDIR_WAIT) {
 177                         io_u->offset = ipo->offset;
 178                         io_u->verify_offset = ipo->offset;
 179                         io_u->buflen = ipo->len;
 180                         io_u->file = td->files[ipo->fileno];
 181                         get_file(io_u->file);
 182                         dprint(FD_IO, "iolog: get %llu/%llu/%s\n", io_u->offset,
 183                                                 io_u->buflen, io_u->file->file_name);
 184                         if (ipo->delay)
 185                                 iolog_delay(td, ipo->delay);
 186                 } else {
 187                         elapsed = mtime_since_genesis();
 188                         if (ipo->delay > elapsed)
 189                                 usec_sleep(td, (ipo->delay - elapsed) * 1000);
 190                 }
 191
 192                 free(ipo);
 193
 194                 if (io_u->ddir != DDIR_WAIT)
 195                         return 0;
 196         }
 197
 198         td->done = 1;
 199         return 1;
 200 }
 201
 202 void prune_io_piece_log(struct thread_data *td)
 203 {
 204         struct io_piece *ipo;
 205         struct fio_rb_node *n;
 206
 207         while ((n = rb_first(&td->io_hist_tree)) != NULL) {
 208                 ipo = rb_entry(n, struct io_piece, rb_node);
 209                 rb_erase(n, &td->io_hist_tree);
 210                 remove_trim_entry(td, ipo);
 211                 td->io_hist_len--;
 212                 free(ipo);
 213         }
 214
 215         while (!flist_empty(&td->io_hist_list)) {
 216                 ipo = flist_first_entry(&td->io_hist_list, struct io_piece, list);
 217                 flist_del(&ipo->list);
 218                 remove_trim_entry(td, ipo);
 219                 td->io_hist_len--;
 220                 free(ipo);
 221         }
 222 }
 223
 224 /*
 225  * log a successful write, so we can unwind the log for verify
 226  */
 227 void log_io_piece(struct thread_data *td, struct io_u *io_u)
 228 {
 229         struct fio_rb_node **p, *parent;
 230         struct io_piece *ipo, *__ipo;
 231
 232         ipo = calloc(1, sizeof(struct io_piece));
 233         init_ipo(ipo);
 234         ipo->file = io_u->file;
 235         ipo->offset = io_u->offset;
 236         ipo->len = io_u->buflen;
 237         ipo->numberio = io_u->numberio;
 238         ipo->flags = IP_F_IN_FLIGHT;
 239
 240         io_u->ipo = ipo;
 241
 242         if (io_u_should_trim(td, io_u)) {
 243                 flist_add_tail(&ipo->trim_list, &td->trim_list);
 244                 td->trim_entries++;
 245         }
 246
 247         /*
 248          * Only sort writes if we don't have a random map in which case we need
 249          * to check for duplicate blocks and drop the old one, which we rely on
 250          * the rb insert/lookup for handling.
 251          */
 252         if (file_randommap(td, ipo->file)) {
 253                 INIT_FLIST_HEAD(&ipo->list);
 254                 flist_add_tail(&ipo->list, &td->io_hist_list);
 255                 ipo->flags |= IP_F_ONLIST;
 256                 td->io_hist_len++;
 257                 return;
 258         }
 259
 260         RB_CLEAR_NODE(&ipo->rb_node);
 261
 262         /*
 263          * Sort the entry into the verification list
 264          */
 265 restart:
 266         p = &td->io_hist_tree.rb_node;
 267         parent = NULL;
 268         while (*p) {
 269                 int overlap = 0;
 270                 parent = *p;
 271
 272                 __ipo = rb_entry(parent, struct io_piece, rb_node);
 273                 if (ipo->file < __ipo->file)
 274                         p = &(*p)->rb_left;
 275                 else if (ipo->file > __ipo->file)
 276                         p = &(*p)->rb_right;
 277                 else if (ipo->offset < __ipo->offset) {
 278                         p = &(*p)->rb_left;
 279                         overlap = ipo->offset + ipo->len > __ipo->offset;
 280                 }
 281                 else if (ipo->offset > __ipo->offset) {
 282                         p = &(*p)->rb_right;
 283                         overlap = __ipo->offset + __ipo->len > ipo->offset;
 284                 }
 285                 else
 286                         overlap = 1;
 287
 288                 if (overlap) {
 289                         dprint(FD_IO, "iolog: overlap %llu/%lu, %llu/%lu\n",
 290                                 __ipo->offset, __ipo->len,
 291                                 ipo->offset, ipo->len);
 292                         td->io_hist_len--;
 293                         rb_erase(parent, &td->io_hist_tree);
 294                         remove_trim_entry(td, __ipo);
 295                         if (!(__ipo->flags & IP_F_IN_FLIGHT))
 296                                 free(__ipo);
 297                         goto restart;
 298                 }
 299         }
 300
 301         rb_link_node(&ipo->rb_node, parent, p);
 302         rb_insert_color(&ipo->rb_node, &td->io_hist_tree);
 303         ipo->flags |= IP_F_ONRB;
 304         td->io_hist_len++;
 305 }
 306
 307 void unlog_io_piece(struct thread_data *td, struct io_u *io_u)
 308 {
 309         struct io_piece *ipo = io_u->ipo;
 310
 311         if (td->ts.nr_block_infos) {
 312                 uint32_t *info = io_u_block_info(td, io_u);
 313                 if (BLOCK_INFO_STATE(*info) < BLOCK_STATE_TRIM_FAILURE) {
 314                         if (io_u->ddir == DDIR_TRIM)
 315                                 *info = BLOCK_INFO_SET_STATE(*info,
 316                                                 BLOCK_STATE_TRIM_FAILURE);
 317                         else if (io_u->ddir == DDIR_WRITE)
 318                                 *info = BLOCK_INFO_SET_STATE(*info,
 319                                                 BLOCK_STATE_WRITE_FAILURE);
 320                 }
 321         }
 322
 323         if (!ipo)
 324                 return;
 325
 326         if (ipo->flags & IP_F_ONRB)
 327                 rb_erase(&ipo->rb_node, &td->io_hist_tree);
 328         else if (ipo->flags & IP_F_ONLIST)
 329                 flist_del(&ipo->list);
 330
 331         free(ipo);
 332         io_u->ipo = NULL;
 333         td->io_hist_len--;
 334 }
 335
 336 void trim_io_piece(const struct io_u *io_u)
 337 {
 338         struct io_piece *ipo = io_u->ipo;
 339
 340         if (!ipo)
 341                 return;
 342
 343         ipo->len = io_u->xfer_buflen - io_u->resid;
 344 }
 345
 346 void write_iolog_close(struct thread_data *td)
 347 {
 348         if (!td->iolog_f)
 349                 return;
 350
 351         fflush(td->iolog_f);
 352         fclose(td->iolog_f);
 353         free(td->iolog_buf);
 354         td->iolog_f = NULL;
 355         td->iolog_buf = NULL;
 356 }
 357
 358 static int64_t iolog_items_to_fetch(struct thread_data *td)
 359 {
 360         struct timespec now;
 361         uint64_t elapsed;
 362         uint64_t for_1s;
 363         int64_t items_to_fetch;
 364
 365         if (!td->io_log_highmark)
 366                 return 10;
 367
 368
 369         fio_gettime(&now, NULL);
 370         elapsed = ntime_since(&td->io_log_highmark_time, &now);
 371         if (elapsed) {
 372                 for_1s = (td->io_log_highmark - td->io_log_current) * 1000000000 / elapsed;
 373                 items_to_fetch = for_1s - td->io_log_current;
 374                 if (items_to_fetch < 0)
 375                         items_to_fetch = 0;
 376         } else
 377                 items_to_fetch = 0;
 378
 379         td->io_log_highmark = td->io_log_current + items_to_fetch;
 380         td->io_log_checkmark = (td->io_log_highmark + 1) / 2;
 381         fio_gettime(&td->io_log_highmark_time, NULL);
 382
 383         return items_to_fetch;
 384 }
 385
 386 /*
 387  * Read version 2 iolog data. It is enhanced to include per-file logging,
 388  * syncs, etc.
 389  */
 390 static bool read_iolog2(struct thread_data *td)
 391 {
 392         unsigned long long offset;
 393         unsigned int bytes;
 394         int reads, writes, waits, fileno = 0, file_action = 0; /* stupid gcc */
 395         char *rfname, *fname, *act;
 396         char *str, *p;
 397         enum fio_ddir rw;
 398         bool realloc = false;
 399         int64_t items_to_fetch = 0;
 400
 401         if (td->o.read_iolog_chunked) {
 402                 items_to_fetch = iolog_items_to_fetch(td);
 403                 if (!items_to_fetch)
 404                         return true;
 405         }
 406
 407         /*
 408          * Read in the read iolog and store it, reuse the infrastructure
 409          * for doing verifications.
 410          */
 411         str = malloc(4096);
 412         rfname = fname = malloc(256+16);
 413         act = malloc(256+16);
 414
 415         reads = writes = waits = 0;
 416         while ((p = fgets(str, 4096, td->io_log_rfile)) != NULL) {
 417                 struct io_piece *ipo;
 418                 int r;
 419
 420                 r = sscanf(p, "%256s %256s %llu %u", rfname, act, &offset,
 421                                                                         &bytes);
 422
 423                 if (td->o.replay_redirect)
 424                         fname = td->o.replay_redirect;
 425
 426                 if (r == 4) {
 427                         /*
 428                          * Check action first
 429                          */
 430                         if (!strcmp(act, "wait"))
 431                                 rw = DDIR_WAIT;
 432                         else if (!strcmp(act, "read"))
 433                                 rw = DDIR_READ;
 434                         else if (!strcmp(act, "write"))
 435                                 rw = DDIR_WRITE;
 436                         else if (!strcmp(act, "sync"))
 437                                 rw = DDIR_SYNC;
 438                         else if (!strcmp(act, "datasync"))
 439                                 rw = DDIR_DATASYNC;
 440                         else if (!strcmp(act, "trim"))
 441                                 rw = DDIR_TRIM;
 442                         else {
 443                                 log_err("fio: bad iolog file action: %s\n",
 444                                                                         act);
 445                                 continue;
 446                         }
 447                         fileno = get_fileno(td, fname);
 448                 } else if (r == 2) {
 449                         rw = DDIR_INVAL;
 450                         if (!strcmp(act, "add")) {
 451                                 if (td->o.replay_redirect &&
 452                                     get_fileno(td, fname) != -1) {
 453                                         dprint(FD_FILE, "iolog: ignoring"
 454                                                 " re-add of file %s\n", fname);
 455                                 } else {
 456                                         fileno = add_file(td, fname, td->subjob_number, 1);
 457                                         file_action = FIO_LOG_ADD_FILE;
 458                                 }
 459                                 continue;
 460                         } else if (!strcmp(act, "open")) {
 461                                 fileno = get_fileno(td, fname);
 462                                 file_action = FIO_LOG_OPEN_FILE;
 463                         } else if (!strcmp(act, "close")) {
 464                                 fileno = get_fileno(td, fname);
 465                                 file_action = FIO_LOG_CLOSE_FILE;
 466                         } else {
 467                                 log_err("fio: bad iolog file action: %s\n",
 468                                                                         act);
 469                                 continue;
 470                         }
 471                 } else {
 472                         log_err("bad iolog2: %s\n", p);
 473                         continue;
 474                 }
 475
 476                 if (rw == DDIR_READ)
 477                         reads++;
 478                 else if (rw == DDIR_WRITE) {
 479                         /*
 480                          * Don't add a write for ro mode
 481                          */
 482                         if (read_only)
 483                                 continue;
 484                         writes++;
 485                 } else if (rw == DDIR_WAIT) {
 486                         if (td->o.no_stall)
 487                                 continue;
 488                         waits++;
 489                 } else if (rw == DDIR_INVAL) {
 490                 } else if (!ddir_sync(rw)) {
 491                         log_err("bad ddir: %d\n", rw);
 492                         continue;
 493                 }
 494
 495                 /*
 496                  * Make note of file
 497                  */
 498                 ipo = calloc(1, sizeof(*ipo));
 499                 init_ipo(ipo);
 500                 ipo->ddir = rw;
 501                 if (rw == DDIR_WAIT) {
 502                         ipo->delay = offset;
 503                 } else {
 504                         if (td->o.replay_scale)
 505                                 ipo->offset = offset / td->o.replay_scale;
 506                         else
 507                                 ipo->offset = offset;
 508                         ipo_bytes_align(td->o.replay_align, ipo);
 509
 510                         ipo->len = bytes;
 511                         if (rw != DDIR_INVAL && bytes > td->o.max_bs[rw]) {
 512                                 realloc = true;
 513                                 td->o.max_bs[rw] = bytes;
 514                         }
 515                         ipo->fileno = fileno;
 516                         ipo->file_action = file_action;
 517                         td->o.size += bytes;
 518                 }
 519
 520                 queue_io_piece(td, ipo);
 521
 522                 if (td->o.read_iolog_chunked) {
 523                         td->io_log_current++;
 524                         items_to_fetch--;
 525                         if (items_to_fetch == 0)
 526                                 break;
 527                 }
 528         }
 529
 530         free(str);
 531         free(act);
 532         free(rfname);
 533
 534         if (td->o.read_iolog_chunked) {
 535                 td->io_log_highmark = td->io_log_current;
 536                 td->io_log_checkmark = (td->io_log_highmark + 1) / 2;
 537                 fio_gettime(&td->io_log_highmark_time, NULL);
 538         }
 539
 540         if (writes && read_only) {
 541                 log_err("fio: <%s> skips replay of %d writes due to"
 542                         " read-only\n", td->o.name, writes);
 543                 writes = 0;
 544         }
 545
 546         if (td->o.read_iolog_chunked) {
 547                 if (td->io_log_current == 0) {
 548                         return false;
 549                 }
 550                 td->o.td_ddir = TD_DDIR_RW;
 551                 if (realloc && td->orig_buffer)
 552                 {
 553                         io_u_quiesce(td);
 554                         free_io_mem(td);
 555                         init_io_u_buffers(td);
 556                 }
 557                 return true;
 558         }
 559
 560         if (!reads && !writes && !waits)
 561                 return false;
 562         else if (reads && !writes)
 563                 td->o.td_ddir = TD_DDIR_READ;
 564         else if (!reads && writes)
 565                 td->o.td_ddir = TD_DDIR_WRITE;
 566         else
 567                 td->o.td_ddir = TD_DDIR_RW;
 568
 569         return true;
 570 }
 571
 572 static bool is_socket(const char *path)
 573 {
 574         struct stat buf;
 575         int r;
 576
 577         r = stat(path, &buf);
 578         if (r == -1)
 579                 return false;
 580
 581         return S_ISSOCK(buf.st_mode);
 582 }
 583
 584 static int open_socket(const char *path)
 585 {
 586         struct sockaddr_un addr;
 587         int ret, fd;
 588
 589         fd = socket(AF_UNIX, SOCK_STREAM, 0);
 590         if (fd < 0)
 591                 return fd;
 592
 593         addr.sun_family = AF_UNIX;
 594         if (snprintf(addr.sun_path, sizeof(addr.sun_path), "%s", path) >=
 595             sizeof(addr.sun_path)) {
 596                 log_err("%s: path name %s is too long for a Unix socket\n",
 597                         __func__, path);
 598         }
 599
 600         ret = connect(fd, (const struct sockaddr *)&addr, strlen(path) + sizeof(addr.sun_family));
 601         if (!ret)
 602                 return fd;
 603
 604         close(fd);
 605         return -1;
 606 }
 607
 608 /*
 609  * open iolog, check version, and call appropriate parser
 610  */
 611 static bool init_iolog_read(struct thread_data *td, char *fname)
 612 {
 613         char buffer[256], *p;
 614         FILE *f = NULL;
 615
 616         dprint(FD_IO, "iolog: name=%s\n", fname);
 617
 618         if (is_socket(fname)) {
 619                 int fd;
 620
 621                 fd = open_socket(fname);
 622                 if (fd >= 0)
 623                         f = fdopen(fd, "r");
 624         } else if (!strcmp(fname, "-")) {
 625                 f = stdin;
 626         } else
 627                 f = fopen(fname, "r");
 628
 629         free(fname);
 630
 631         if (!f) {
 632                 perror("fopen read iolog");
 633                 return false;
 634         }
 635
 636         p = fgets(buffer, sizeof(buffer), f);
 637         if (!p) {
 638                 td_verror(td, errno, "iolog read");
 639                 log_err("fio: unable to read iolog\n");
 640                 fclose(f);
 641                 return false;
 642         }
 643
 644         /*
 645          * version 2 of the iolog stores a specific string as the
 646          * first line, check for that
 647          */
 648         if (!strncmp(iolog_ver2, buffer, strlen(iolog_ver2))) {
 649                 free_release_files(td);
 650                 td->io_log_rfile = f;
 651                 return read_iolog2(td);
 652         }
 653
 654         log_err("fio: iolog version 1 is no longer supported\n");
 655         fclose(f);
 656         return false;
 657 }
 658
 659 /*
 660  * Set up a log for storing io patterns.
 661  */
 662 static bool init_iolog_write(struct thread_data *td)
 663 {
 664         struct fio_file *ff;
 665         FILE *f;
 666         unsigned int i;
 667
 668         f = fopen(td->o.write_iolog_file, "a");
 669         if (!f) {
 670                 perror("fopen write iolog");
 671                 return false;
 672         }
 673
 674         /*
 675          * That's it for writing, setup a log buffer and we're done.
 676           */
 677         td->iolog_f = f;
 678         td->iolog_buf = malloc(8192);
 679         setvbuf(f, td->iolog_buf, _IOFBF, 8192);
 680
 681         /*
 682          * write our version line
 683          */
 684         if (fprintf(f, "%s\n", iolog_ver2) < 0) {
 685                 perror("iolog init\n");
 686                 return false;
 687         }
 688
 689         /*
 690          * add all known files
 691          */
 692         for_each_file(td, ff, i)
 693                 log_file(td, ff, FIO_LOG_ADD_FILE);
 694
 695         return true;
 696 }
 697
 698 bool init_iolog(struct thread_data *td)
 699 {
 700         bool ret;
 701
 702         if (td->o.read_iolog_file) {
 703                 int need_swap;
 704                 char * fname = get_name_by_idx(td->o.read_iolog_file, td->subjob_number);
 705
 706                 /*
 707                  * Check if it's a blktrace file and load that if possible.
 708                  * Otherwise assume it's a normal log file and load that.
 709                  */
 710                 if (is_blktrace(fname, &need_swap)) {
 711                         td->io_log_blktrace = 1;
 712                         ret = load_blktrace(td, fname, need_swap);
 713                 } else {
 714                         td->io_log_blktrace = 0;
 715                         ret = init_iolog_read(td, fname);
 716                 }
 717         } else if (td->o.write_iolog_file)
 718                 ret = init_iolog_write(td);
 719         else
 720                 ret = true;
 721
 722         if (!ret)
 723                 td_verror(td, EINVAL, "failed initializing iolog");
 724
 725         return ret;
 726 }
 727
 728 void setup_log(struct io_log **log, struct log_params *p,
 729                const char *filename)
 730 {
 731         struct io_log *l;
 732         int i;
 733         struct io_u_plat_entry *entry;
 734         struct flist_head *list;
 735
 736         l = scalloc(1, sizeof(*l));
 737         INIT_FLIST_HEAD(&l->io_logs);
 738         l->log_type = p->log_type;
 739         l->log_offset = p->log_offset;
 740         l->log_gz = p->log_gz;
 741         l->log_gz_store = p->log_gz_store;
 742         l->avg_msec = p->avg_msec;
 743         l->hist_msec = p->hist_msec;
 744         l->hist_coarseness = p->hist_coarseness;
 745         l->filename = strdup(filename);
 746         l->td = p->td;
 747
 748         /* Initialize histogram lists for each r/w direction,
 749          * with initial io_u_plat of all zeros:
 750          */
 751         for (i = 0; i < DDIR_RWDIR_CNT; i++) {
 752                 list = &l->hist_window[i].list;
 753                 INIT_FLIST_HEAD(list);
 754                 entry = calloc(1, sizeof(struct io_u_plat_entry));
 755                 flist_add(&entry->list, list);
 756         }
 757
 758         if (l->td && l->td->o.io_submit_mode != IO_MODE_OFFLOAD) {
 759                 unsigned int def_samples = DEF_LOG_ENTRIES;
 760                 struct io_logs *__p;
 761
 762                 __p = calloc(1, sizeof(*l->pending));
 763                 if (l->td->o.iodepth > DEF_LOG_ENTRIES)
 764                         def_samples = roundup_pow2(l->td->o.iodepth);
 765                 __p->max_samples = def_samples;
 766                 __p->log = calloc(__p->max_samples, log_entry_sz(l));
 767                 l->pending = __p;
 768         }
 769
 770         if (l->log_offset)
 771                 l->log_ddir_mask = LOG_OFFSET_SAMPLE_BIT;
 772
 773         INIT_FLIST_HEAD(&l->chunk_list);
 774
 775         if (l->log_gz && !p->td)
 776                 l->log_gz = 0;
 777         else if (l->log_gz || l->log_gz_store) {
 778                 mutex_init_pshared(&l->chunk_lock);
 779                 mutex_init_pshared(&l->deferred_free_lock);
 780                 p->td->flags |= TD_F_COMPRESS_LOG;
 781         }
 782
 783         *log = l;
 784 }
 785
 786 #ifdef CONFIG_SETVBUF
 787 static void *set_file_buffer(FILE *f)
 788 {
 789         size_t size = 1048576;
 790         void *buf;
 791
 792         buf = malloc(size);
 793         setvbuf(f, buf, _IOFBF, size);
 794         return buf;
 795 }
 796
 797 static void clear_file_buffer(void *buf)
 798 {
 799         free(buf);
 800 }
 801 #else
 802 static void *set_file_buffer(FILE *f)
 803 {
 804         return NULL;
 805 }
 806
 807 static void clear_file_buffer(void *buf)
 808 {
 809 }
 810 #endif
 811
 812 void free_log(struct io_log *log)
 813 {
 814         while (!flist_empty(&log->io_logs)) {
 815                 struct io_logs *cur_log;
 816
 817                 cur_log = flist_first_entry(&log->io_logs, struct io_logs, list);
 818                 flist_del_init(&cur_log->list);
 819                 free(cur_log->log);
 820                 sfree(cur_log);
 821         }
 822
 823         if (log->pending) {
 824                 free(log->pending->log);
 825                 free(log->pending);
 826                 log->pending = NULL;
 827         }
 828
 829         free(log->pending);
 830         free(log->filename);
 831         sfree(log);
 832 }
 833
 834 uint64_t hist_sum(int j, int stride, uint64_t *io_u_plat,
 835                 uint64_t *io_u_plat_last)
 836 {
 837         uint64_t sum;
 838         int k;
 839
 840         if (io_u_plat_last) {
 841                 for (k = sum = 0; k < stride; k++)
 842                         sum += io_u_plat[j + k] - io_u_plat_last[j + k];
 843         } else {
 844                 for (k = sum = 0; k < stride; k++)
 845                         sum += io_u_plat[j + k];
 846         }
 847
 848         return sum;
 849 }
 850
 851 static void flush_hist_samples(FILE *f, int hist_coarseness, void *samples,
 852                                uint64_t sample_size)
 853 {
 854         struct io_sample *s;
 855         int log_offset;
 856         uint64_t i, j, nr_samples;
 857         struct io_u_plat_entry *entry, *entry_before;
 858         uint64_t *io_u_plat;
 859         uint64_t *io_u_plat_before;
 860
 861         int stride = 1 << hist_coarseness;
 862
 863         if (!sample_size)
 864                 return;
 865
 866         s = __get_sample(samples, 0, 0);
 867         log_offset = (s->__ddir & LOG_OFFSET_SAMPLE_BIT) != 0;
 868
 869         nr_samples = sample_size / __log_entry_sz(log_offset);
 870
 871         for (i = 0; i < nr_samples; i++) {
 872                 s = __get_sample(samples, log_offset, i);
 873
 874                 entry = s->data.plat_entry;
 875                 io_u_plat = entry->io_u_plat;
 876
 877                 entry_before = flist_first_entry(&entry->list, struct io_u_plat_entry, list);
 878                 io_u_plat_before = entry_before->io_u_plat;
 879
 880                 fprintf(f, "%lu, %u, %llu, ", (unsigned long) s->time,
 881                                                 io_sample_ddir(s), (unsigned long long) s->bs);
 882                 for (j = 0; j < FIO_IO_U_PLAT_NR - stride; j += stride) {
 883                         fprintf(f, "%llu, ", (unsigned long long)
 884                                 hist_sum(j, stride, io_u_plat, io_u_plat_before));
 885                 }
 886                 fprintf(f, "%llu\n", (unsigned long long)
 887                         hist_sum(FIO_IO_U_PLAT_NR - stride, stride, io_u_plat,
 888                                         io_u_plat_before));
 889
 890                 flist_del(&entry_before->list);
 891                 free(entry_before);
 892         }
 893 }
 894
 895 void flush_samples(FILE *f, void *samples, uint64_t sample_size)
 896 {
 897         struct io_sample *s;
 898         int log_offset;
 899         uint64_t i, nr_samples;
 900
 901         if (!sample_size)
 902                 return;
 903
 904         s = __get_sample(samples, 0, 0);
 905         log_offset = (s->__ddir & LOG_OFFSET_SAMPLE_BIT) != 0;
 906
 907         nr_samples = sample_size / __log_entry_sz(log_offset);
 908
 909         for (i = 0; i < nr_samples; i++) {
 910                 s = __get_sample(samples, log_offset, i);
 911
 912                 if (!log_offset) {
 913                         fprintf(f, "%lu, %" PRId64 ", %u, %llu, %u\n",
 914                                         (unsigned long) s->time,
 915                                         s->data.val,
 916                                         io_sample_ddir(s), (unsigned long long) s->bs, s->priority_bit);
 917                 } else {
 918                         struct io_sample_offset *so = (void *) s;
 919
 920                         fprintf(f, "%lu, %" PRId64 ", %u, %llu, %llu, %u\n",
 921                                         (unsigned long) s->time,
 922                                         s->data.val,
 923                                         io_sample_ddir(s), (unsigned long long) s->bs,
 924                                         (unsigned long long) so->offset, s->priority_bit);
 925                 }
 926         }
 927 }
 928
 929 #ifdef CONFIG_ZLIB
 930
 931 struct iolog_flush_data {
 932         struct workqueue_work work;
 933         struct io_log *log;
 934         void *samples;
 935         uint32_t nr_samples;
 936         bool free;
 937 };
 938
 939 #define GZ_CHUNK        131072
 940
 941 static struct iolog_compress *get_new_chunk(unsigned int seq)
 942 {
 943         struct iolog_compress *c;
 944
 945         c = malloc(sizeof(*c));
 946         INIT_FLIST_HEAD(&c->list);
 947         c->buf = malloc(GZ_CHUNK);
 948         c->len = 0;
 949         c->seq = seq;
 950         return c;
 951 }
 952
 953 static void free_chunk(struct iolog_compress *ic)
 954 {
 955         free(ic->buf);
 956         free(ic);
 957 }
 958
 959 static int z_stream_init(z_stream *stream, int gz_hdr)
 960 {
 961         int wbits = 15;
 962
 963         memset(stream, 0, sizeof(*stream));
 964         stream->zalloc = Z_NULL;
 965         stream->zfree = Z_NULL;
 966         stream->opaque = Z_NULL;
 967         stream->next_in = Z_NULL;
 968
 969         /*
 970          * zlib magic - add 32 for auto-detection of gz header or not,
 971          * if we decide to store files in a gzip friendly format.
 972          */
 973         if (gz_hdr)
 974                 wbits += 32;
 975
 976         if (inflateInit2(stream, wbits) != Z_OK)
 977                 return 1;
 978
 979         return 0;
 980 }
 981
 982 struct inflate_chunk_iter {
 983         unsigned int seq;
 984         int err;
 985         void *buf;
 986         size_t buf_size;
 987         size_t buf_used;
 988         size_t chunk_sz;
 989 };
 990
 991 static void finish_chunk(z_stream *stream, FILE *f,
 992                          struct inflate_chunk_iter *iter)
 993 {
 994         int ret;
 995
 996         ret = inflateEnd(stream);
 997         if (ret != Z_OK)
 998                 log_err("fio: failed to end log inflation seq %d (%d)\n",
 999                                 iter->seq, ret);
1000
1001         flush_samples(f, iter->buf, iter->buf_used);
1002         free(iter->buf);
1003         iter->buf = NULL;
1004         iter->buf_size = iter->buf_used = 0;
1005 }
1006
1007 /*
1008  * Iterative chunk inflation. Handles cases where we cross into a new
1009  * sequence, doing flush finish of previous chunk if needed.
1010  */
1011 static size_t inflate_chunk(struct iolog_compress *ic, int gz_hdr, FILE *f,
1012                             z_stream *stream, struct inflate_chunk_iter *iter)
1013 {
1014         size_t ret;
1015
1016         dprint(FD_COMPRESS, "inflate chunk size=%lu, seq=%u\n",
1017                                 (unsigned long) ic->len, ic->seq);
1018
1019         if (ic->seq != iter->seq) {
1020                 if (iter->seq)
1021                         finish_chunk(stream, f, iter);
1022
1023                 z_stream_init(stream, gz_hdr);
1024                 iter->seq = ic->seq;
1025         }
1026
1027         stream->avail_in = ic->len;
1028         stream->next_in = ic->buf;
1029
1030         if (!iter->buf_size) {
1031                 iter->buf_size = iter->chunk_sz;
1032                 iter->buf = malloc(iter->buf_size);
1033         }
1034
1035         while (stream->avail_in) {
1036                 size_t this_out = iter->buf_size - iter->buf_used;
1037                 int err;
1038
1039                 stream->avail_out = this_out;
1040                 stream->next_out = iter->buf + iter->buf_used;
1041
1042                 err = inflate(stream, Z_NO_FLUSH);
1043                 if (err < 0) {
1044                         log_err("fio: failed inflating log: %d\n", err);
1045                         iter->err = err;
1046                         break;
1047                 }
1048
1049                 iter->buf_used += this_out - stream->avail_out;
1050
1051                 if (!stream->avail_out) {
1052                         iter->buf_size += iter->chunk_sz;
1053                         iter->buf = realloc(iter->buf, iter->buf_size);
1054                         continue;
1055                 }
1056
1057                 if (err == Z_STREAM_END)
1058                         break;
1059         }
1060
1061         ret = (void *) stream->next_in - ic->buf;
1062
1063         dprint(FD_COMPRESS, "inflated to size=%lu\n", (unsigned long) iter->buf_size);
1064
1065         return ret;
1066 }
1067
1068 /*
1069  * Inflate stored compressed chunks, or write them directly to the log
1070  * file if so instructed.
1071  */
1072 static int inflate_gz_chunks(struct io_log *log, FILE *f)
1073 {
1074         struct inflate_chunk_iter iter = { .chunk_sz = log->log_gz, };
1075         z_stream stream;
1076
1077         while (!flist_empty(&log->chunk_list)) {
1078                 struct iolog_compress *ic;
1079
1080                 ic = flist_first_entry(&log->chunk_list, struct iolog_compress, list);
1081                 flist_del(&ic->list);
1082
1083                 if (log->log_gz_store) {
1084                         size_t ret;
1085
1086                         dprint(FD_COMPRESS, "log write chunk size=%lu, "
1087                                 "seq=%u\n", (unsigned long) ic->len, ic->seq);
1088
1089                         ret = fwrite(ic->buf, ic->len, 1, f);
1090                         if (ret != 1 || ferror(f)) {
1091                                 iter.err = errno;
1092                                 log_err("fio: error writing compressed log\n");
1093                         }
1094                 } else
1095                         inflate_chunk(ic, log->log_gz_store, f, &stream, &iter);
1096
1097                 free_chunk(ic);
1098         }
1099
1100         if (iter.seq) {
1101                 finish_chunk(&stream, f, &iter);
1102                 free(iter.buf);
1103         }
1104
1105         return iter.err;
1106 }
1107
1108 /*
1109  * Open compressed log file and decompress the stored chunks and
1110  * write them to stdout. The chunks are stored sequentially in the
1111  * file, so we iterate over them and do them one-by-one.
1112  */
1113 int iolog_file_inflate(const char *file)
1114 {
1115         struct inflate_chunk_iter iter = { .chunk_sz = 64 * 1024 * 1024, };
1116         struct iolog_compress ic;
1117         z_stream stream;
1118         struct stat sb;
1119         size_t ret;
1120         size_t total;
1121         void *buf;
1122         FILE *f;
1123
1124         f = fopen(file, "r");
1125         if (!f) {
1126                 perror("fopen");
1127                 return 1;
1128         }
1129
1130         if (stat(file, &sb) < 0) {
1131                 fclose(f);
1132                 perror("stat");
1133                 return 1;
1134         }
1135
1136         ic.buf = buf = malloc(sb.st_size);
1137         ic.len = sb.st_size;
1138         ic.seq = 1;
1139
1140         ret = fread(ic.buf, ic.len, 1, f);
1141         if (ret == 0 && ferror(f)) {
1142                 perror("fread");
1143                 fclose(f);
1144                 free(buf);
1145                 return 1;
1146         } else if (ferror(f) || (!feof(f) && ret != 1)) {
1147                 log_err("fio: short read on reading log\n");
1148                 fclose(f);
1149                 free(buf);
1150                 return 1;
1151         }
1152
1153         fclose(f);
1154
1155         /*
1156          * Each chunk will return Z_STREAM_END. We don't know how many
1157          * chunks are in the file, so we just keep looping and incrementing
1158          * the sequence number until we have consumed the whole compressed
1159          * file.
1160          */
1161         total = ic.len;
1162         do {
1163                 size_t iret;
1164
1165                 iret = inflate_chunk(&ic,  1, stdout, &stream, &iter);
1166                 total -= iret;
1167                 if (!total)
1168                         break;
1169                 if (iter.err)
1170                         break;
1171
1172                 ic.seq++;
1173                 ic.len -= iret;
1174                 ic.buf += iret;
1175         } while (1);
1176
1177         if (iter.seq) {
1178                 finish_chunk(&stream, stdout, &iter);
1179                 free(iter.buf);
1180         }
1181
1182         free(buf);
1183         return iter.err;
1184 }
1185
1186 #else
1187
1188 static int inflate_gz_chunks(struct io_log *log, FILE *f)
1189 {
1190         return 0;
1191 }
1192
1193 int iolog_file_inflate(const char *file)
1194 {
1195         log_err("fio: log inflation not possible without zlib\n");
1196         return 1;
1197 }
1198
1199 #endif
1200
1201 void flush_log(struct io_log *log, bool do_append)
1202 {
1203         void *buf;
1204         FILE *f;
1205
1206         if (!do_append)
1207                 f = fopen(log->filename, "w");
1208         else
1209                 f = fopen(log->filename, "a");
1210         if (!f) {
1211                 perror("fopen log");
1212                 return;
1213         }
1214
1215         buf = set_file_buffer(f);
1216
1217         inflate_gz_chunks(log, f);
1218
1219         while (!flist_empty(&log->io_logs)) {
1220                 struct io_logs *cur_log;
1221
1222                 cur_log = flist_first_entry(&log->io_logs, struct io_logs, list);
1223                 flist_del_init(&cur_log->list);
1224
1225                 if (log->td && log == log->td->clat_hist_log)
1226                         flush_hist_samples(f, log->hist_coarseness, cur_log->log,
1227                                            log_sample_sz(log, cur_log));
1228                 else
1229                         flush_samples(f, cur_log->log, log_sample_sz(log, cur_log));
1230
1231                 sfree(cur_log);
1232         }
1233
1234         fclose(f);
1235         clear_file_buffer(buf);
1236 }
1237
1238 static int finish_log(struct thread_data *td, struct io_log *log, int trylock)
1239 {
1240         if (td->flags & TD_F_COMPRESS_LOG)
1241                 iolog_flush(log);
1242
1243         if (trylock) {
1244                 if (fio_trylock_file(log->filename))
1245                         return 1;
1246         } else
1247                 fio_lock_file(log->filename);
1248
1249         if (td->client_type == FIO_CLIENT_TYPE_GUI || is_backend)
1250                 fio_send_iolog(td, log, log->filename);
1251         else
1252                 flush_log(log, !td->o.per_job_logs);
1253
1254         fio_unlock_file(log->filename);
1255         free_log(log);
1256         return 0;
1257 }
1258
1259 size_t log_chunk_sizes(struct io_log *log)
1260 {
1261         struct flist_head *entry;
1262         size_t ret;
1263
1264         if (flist_empty(&log->chunk_list))
1265                 return 0;
1266
1267         ret = 0;
1268         pthread_mutex_lock(&log->chunk_lock);
1269         flist_for_each(entry, &log->chunk_list) {
1270                 struct iolog_compress *c;
1271
1272                 c = flist_entry(entry, struct iolog_compress, list);
1273                 ret += c->len;
1274         }
1275         pthread_mutex_unlock(&log->chunk_lock);
1276         return ret;
1277 }
1278
1279 #ifdef CONFIG_ZLIB
1280
1281 static void iolog_put_deferred(struct io_log *log, void *ptr)
1282 {
1283         if (!ptr)
1284                 return;
1285
1286         pthread_mutex_lock(&log->deferred_free_lock);
1287         if (log->deferred < IOLOG_MAX_DEFER) {
1288                 log->deferred_items[log->deferred] = ptr;
1289                 log->deferred++;
1290         } else if (!fio_did_warn(FIO_WARN_IOLOG_DROP))
1291                 log_err("fio: had to drop log entry free\n");
1292         pthread_mutex_unlock(&log->deferred_free_lock);
1293 }
1294
1295 static void iolog_free_deferred(struct io_log *log)
1296 {
1297         int i;
1298
1299         if (!log->deferred)
1300                 return;
1301
1302         pthread_mutex_lock(&log->deferred_free_lock);
1303
1304         for (i = 0; i < log->deferred; i++) {
1305                 free(log->deferred_items[i]);
1306                 log->deferred_items[i] = NULL;
1307         }
1308
1309         log->deferred = 0;
1310         pthread_mutex_unlock(&log->deferred_free_lock);
1311 }
1312
1313 static int gz_work(struct iolog_flush_data *data)
1314 {
1315         struct iolog_compress *c = NULL;
1316         struct flist_head list;
1317         unsigned int seq;
1318         z_stream stream;
1319         size_t total = 0;
1320         int ret;
1321
1322         INIT_FLIST_HEAD(&list);
1323
1324         memset(&stream, 0, sizeof(stream));
1325         stream.zalloc = Z_NULL;
1326         stream.zfree = Z_NULL;
1327         stream.opaque = Z_NULL;
1328
1329         ret = deflateInit(&stream, Z_DEFAULT_COMPRESSION);
1330         if (ret != Z_OK) {
1331                 log_err("fio: failed to init gz stream\n");
1332                 goto err;
1333         }
1334
1335         seq = ++data->log->chunk_seq;
1336
1337         stream.next_in = (void *) data->samples;
1338         stream.avail_in = data->nr_samples * log_entry_sz(data->log);
1339
1340         dprint(FD_COMPRESS, "deflate input size=%lu, seq=%u, log=%s\n",
1341                                 (unsigned long) stream.avail_in, seq,
1342                                 data->log->filename);
1343         do {
1344                 if (c)
1345                         dprint(FD_COMPRESS, "seq=%d, chunk=%lu\n", seq,
1346                                 (unsigned long) c->len);
1347                 c = get_new_chunk(seq);
1348                 stream.avail_out = GZ_CHUNK;
1349                 stream.next_out = c->buf;
1350                 ret = deflate(&stream, Z_NO_FLUSH);
1351                 if (ret < 0) {
1352                         log_err("fio: deflate log (%d)\n", ret);
1353                         free_chunk(c);
1354                         goto err;
1355                 }
1356
1357                 c->len = GZ_CHUNK - stream.avail_out;
1358                 flist_add_tail(&c->list, &list);
1359                 total += c->len;
1360         } while (stream.avail_in);
1361
1362         stream.next_out = c->buf + c->len;
1363         stream.avail_out = GZ_CHUNK - c->len;
1364
1365         ret = deflate(&stream, Z_FINISH);
1366         if (ret < 0) {
1367                 /*
1368                  * Z_BUF_ERROR is special, it just means we need more
1369                  * output space. We'll handle that below. Treat any other
1370                  * error as fatal.
1371                  */
1372                 if (ret != Z_BUF_ERROR) {
1373                         log_err("fio: deflate log (%d)\n", ret);
1374                         flist_del(&c->list);
1375                         free_chunk(c);
1376                         goto err;
1377                 }
1378         }
1379
1380         total -= c->len;
1381         c->len = GZ_CHUNK - stream.avail_out;
1382         total += c->len;
1383         dprint(FD_COMPRESS, "seq=%d, chunk=%lu\n", seq, (unsigned long) c->len);
1384
1385         if (ret != Z_STREAM_END) {
1386                 do {
1387                         c = get_new_chunk(seq);
1388                         stream.avail_out = GZ_CHUNK;
1389                         stream.next_out = c->buf;
1390                         ret = deflate(&stream, Z_FINISH);
1391                         c->len = GZ_CHUNK - stream.avail_out;
1392                         total += c->len;
1393                         flist_add_tail(&c->list, &list);
1394                         dprint(FD_COMPRESS, "seq=%d, chunk=%lu\n", seq,
1395                                 (unsigned long) c->len);
1396                 } while (ret != Z_STREAM_END);
1397         }
1398
1399         dprint(FD_COMPRESS, "deflated to size=%lu\n", (unsigned long) total);
1400
1401         ret = deflateEnd(&stream);
1402         if (ret != Z_OK)
1403                 log_err("fio: deflateEnd %d\n", ret);
1404
1405         iolog_put_deferred(data->log, data->samples);
1406
1407         if (!flist_empty(&list)) {
1408                 pthread_mutex_lock(&data->log->chunk_lock);
1409                 flist_splice_tail(&list, &data->log->chunk_list);
1410                 pthread_mutex_unlock(&data->log->chunk_lock);
1411         }
1412
1413         ret = 0;
1414 done:
1415         if (data->free)
1416                 sfree(data);
1417         return ret;
1418 err:
1419         while (!flist_empty(&list)) {
1420                 c = flist_first_entry(list.next, struct iolog_compress, list);
1421                 flist_del(&c->list);
1422                 free_chunk(c);
1423         }
1424         ret = 1;
1425         goto done;
1426 }
1427
1428 /*
1429  * Invoked from our compress helper thread, when logging would have exceeded
1430  * the specified memory limitation. Compresses the previously stored
1431  * entries.
1432  */
1433 static int gz_work_async(struct submit_worker *sw, struct workqueue_work *work)
1434 {
1435         return gz_work(container_of(work, struct iolog_flush_data, work));
1436 }
1437
1438 static int gz_init_worker(struct submit_worker *sw)
1439 {
1440         struct thread_data *td = sw->wq->td;
1441
1442         if (!fio_option_is_set(&td->o, log_gz_cpumask))
1443                 return 0;
1444
1445         if (fio_setaffinity(gettid(), td->o.log_gz_cpumask) == -1) {
1446                 log_err("gz: failed to set CPU affinity\n");
1447                 return 1;
1448         }
1449
1450         return 0;
1451 }
1452
1453 static struct workqueue_ops log_compress_wq_ops = {
1454         .fn             = gz_work_async,
1455         .init_worker_fn = gz_init_worker,
1456         .nice           = 1,
1457 };
1458
1459 int iolog_compress_init(struct thread_data *td, struct sk_out *sk_out)
1460 {
1461         if (!(td->flags & TD_F_COMPRESS_LOG))
1462                 return 0;
1463
1464         workqueue_init(td, &td->log_compress_wq, &log_compress_wq_ops, 1, sk_out);
1465         return 0;
1466 }
1467
1468 void iolog_compress_exit(struct thread_data *td)
1469 {
1470         if (!(td->flags & TD_F_COMPRESS_LOG))
1471                 return;
1472
1473         workqueue_exit(&td->log_compress_wq);
1474 }
1475
1476 /*
1477  * Queue work item to compress the existing log entries. We reset the
1478  * current log to a small size, and reference the existing log in the
1479  * data that we queue for compression. Once compression has been done,
1480  * this old log is freed. If called with finish == true, will not return
1481  * until the log compression has completed, and will flush all previous
1482  * logs too
1483  */
1484 static int iolog_flush(struct io_log *log)
1485 {
1486         struct iolog_flush_data *data;
1487
1488         data = malloc(sizeof(*data));
1489         if (!data)
1490                 return 1;
1491
1492         data->log = log;
1493         data->free = false;
1494
1495         while (!flist_empty(&log->io_logs)) {
1496                 struct io_logs *cur_log;
1497
1498                 cur_log = flist_first_entry(&log->io_logs, struct io_logs, list);
1499                 flist_del_init(&cur_log->list);
1500
1501                 data->samples = cur_log->log;
1502                 data->nr_samples = cur_log->nr_samples;
1503
1504                 sfree(cur_log);
1505
1506                 gz_work(data);
1507         }
1508
1509         free(data);
1510         return 0;
1511 }
1512
1513 int iolog_cur_flush(struct io_log *log, struct io_logs *cur_log)
1514 {
1515         struct iolog_flush_data *data;
1516
1517         data = smalloc(sizeof(*data));
1518         if (!data)
1519                 return 1;
1520
1521         data->log = log;
1522
1523         data->samples = cur_log->log;
1524         data->nr_samples = cur_log->nr_samples;
1525         data->free = true;
1526
1527         cur_log->nr_samples = cur_log->max_samples = 0;
1528         cur_log->log = NULL;
1529
1530         workqueue_enqueue(&log->td->log_compress_wq, &data->work);
1531
1532         iolog_free_deferred(log);
1533
1534         return 0;
1535 }
1536 #else
1537
1538 static int iolog_flush(struct io_log *log)
1539 {
1540         return 1;
1541 }
1542
1543 int iolog_cur_flush(struct io_log *log, struct io_logs *cur_log)
1544 {
1545         return 1;
1546 }
1547
1548 int iolog_compress_init(struct thread_data *td, struct sk_out *sk_out)
1549 {
1550         return 0;
1551 }
1552
1553 void iolog_compress_exit(struct thread_data *td)
1554 {
1555 }
1556
1557 #endif
1558
1559 struct io_logs *iolog_cur_log(struct io_log *log)
1560 {
1561         if (flist_empty(&log->io_logs))
1562                 return NULL;
1563
1564         return flist_last_entry(&log->io_logs, struct io_logs, list);
1565 }
1566
1567 uint64_t iolog_nr_samples(struct io_log *iolog)
1568 {
1569         struct flist_head *entry;
1570         uint64_t ret = 0;
1571
1572         flist_for_each(entry, &iolog->io_logs) {
1573                 struct io_logs *cur_log;
1574
1575                 cur_log = flist_entry(entry, struct io_logs, list);
1576                 ret += cur_log->nr_samples;
1577         }
1578
1579         return ret;
1580 }
1581
1582 static int __write_log(struct thread_data *td, struct io_log *log, int try)
1583 {
1584         if (log)
1585                 return finish_log(td, log, try);
1586
1587         return 0;
1588 }
1589
1590 static int write_iops_log(struct thread_data *td, int try, bool unit_log)
1591 {
1592         int ret;
1593
1594         if (per_unit_log(td->iops_log) != unit_log)
1595                 return 0;
1596
1597         ret = __write_log(td, td->iops_log, try);
1598         if (!ret)
1599                 td->iops_log = NULL;
1600
1601         return ret;
1602 }
1603
1604 static int write_slat_log(struct thread_data *td, int try, bool unit_log)
1605 {
1606         int ret;
1607
1608         if (!unit_log)
1609                 return 0;
1610
1611         ret = __write_log(td, td->slat_log, try);
1612         if (!ret)
1613                 td->slat_log = NULL;
1614
1615         return ret;
1616 }
1617
1618 static int write_clat_log(struct thread_data *td, int try, bool unit_log)
1619 {
1620         int ret;
1621
1622         if (!unit_log)
1623                 return 0;
1624
1625         ret = __write_log(td, td->clat_log, try);
1626         if (!ret)
1627                 td->clat_log = NULL;
1628
1629         return ret;
1630 }
1631
1632 static int write_clat_hist_log(struct thread_data *td, int try, bool unit_log)
1633 {
1634         int ret;
1635
1636         if (!unit_log)
1637                 return 0;
1638
1639         ret = __write_log(td, td->clat_hist_log, try);
1640         if (!ret)
1641                 td->clat_hist_log = NULL;
1642
1643         return ret;
1644 }
1645
1646 static int write_lat_log(struct thread_data *td, int try, bool unit_log)
1647 {
1648         int ret;
1649
1650         if (!unit_log)
1651                 return 0;
1652
1653         ret = __write_log(td, td->lat_log, try);
1654         if (!ret)
1655                 td->lat_log = NULL;
1656
1657         return ret;
1658 }
1659
1660 static int write_bandw_log(struct thread_data *td, int try, bool unit_log)
1661 {
1662         int ret;
1663
1664         if (per_unit_log(td->bw_log) != unit_log)
1665                 return 0;
1666
1667         ret = __write_log(td, td->bw_log, try);
1668         if (!ret)
1669                 td->bw_log = NULL;
1670
1671         return ret;
1672 }
1673
1674 enum {
1675         BW_LOG_MASK     = 1,
1676         LAT_LOG_MASK    = 2,
1677         SLAT_LOG_MASK   = 4,
1678         CLAT_LOG_MASK   = 8,
1679         IOPS_LOG_MASK   = 16,
1680         CLAT_HIST_LOG_MASK = 32,
1681
1682         ALL_LOG_NR      = 6,
1683 };
1684
1685 struct log_type {
1686         unsigned int mask;
1687         int (*fn)(struct thread_data *, int, bool);
1688 };
1689
1690 static struct log_type log_types[] = {
1691         {
1692                 .mask   = BW_LOG_MASK,
1693                 .fn     = write_bandw_log,
1694         },
1695         {
1696                 .mask   = LAT_LOG_MASK,
1697                 .fn     = write_lat_log,
1698         },
1699         {
1700                 .mask   = SLAT_LOG_MASK,
1701                 .fn     = write_slat_log,
1702         },
1703         {
1704                 .mask   = CLAT_LOG_MASK,
1705                 .fn     = write_clat_log,
1706         },
1707         {
1708                 .mask   = IOPS_LOG_MASK,
1709                 .fn     = write_iops_log,
1710         },
1711         {
1712                 .mask   = CLAT_HIST_LOG_MASK,
1713                 .fn     = write_clat_hist_log,
1714         }
1715 };
1716
1717 void td_writeout_logs(struct thread_data *td, bool unit_logs)
1718 {
1719         unsigned int log_mask = 0;
1720         unsigned int log_left = ALL_LOG_NR;
1721         int old_state, i;
1722
1723         old_state = td_bump_runstate(td, TD_FINISHING);
1724
1725         finalize_logs(td, unit_logs);
1726
1727         while (log_left) {
1728                 int prev_log_left = log_left;
1729
1730                 for (i = 0; i < ALL_LOG_NR && log_left; i++) {
1731                         struct log_type *lt = &log_types[i];
1732                         int ret;
1733
1734                         if (!(log_mask & lt->mask)) {
1735                                 ret = lt->fn(td, log_left != 1, unit_logs);
1736                                 if (!ret) {
1737                                         log_left--;
1738                                         log_mask |= lt->mask;
1739                                 }
1740                         }
1741                 }
1742
1743                 if (prev_log_left == log_left)
1744                         usleep(5000);
1745         }
1746
1747         td_restore_runstate(td, old_state);
1748 }
1749
1750 void fio_writeout_logs(bool unit_logs)
1751 {
1752         struct thread_data *td;
1753         int i;
1754
1755         for_each_td(td, i)
1756                 td_writeout_logs(td, unit_logs);
1757 }