blktrace.c

   1 /*
   2  * blktrace support code for fio
   3  */
   4 #include <stdio.h>
   5 #include <stdlib.h>
   6 #include <unistd.h>
   7 #include <errno.h>
   8 #include <sys/sysmacros.h>
   9
  10 #include "flist.h"
  11 #include "fio.h"
  12 #include "iolog.h"
  13 #include "blktrace.h"
  14 #include "blktrace_api.h"
  15 #include "oslib/linux-dev-lookup.h"
  16
  17 struct file_cache {
  18         unsigned int maj;
  19         unsigned int min;
  20         unsigned int fileno;
  21 };
  22
  23 /*
  24  * Just discard the pdu by seeking past it.
  25  */
  26 static int discard_pdu(FILE* f, struct blk_io_trace *t)
  27 {
  28         if (t->pdu_len == 0)
  29                 return 0;
  30
  31         dprint(FD_BLKTRACE, "discard pdu len %u\n", t->pdu_len);
  32         if (fseek(f, t->pdu_len, SEEK_CUR) < 0)
  33                 return -errno;
  34
  35         return t->pdu_len;
  36 }
  37
  38 /*
  39  * Check if this is a blktrace binary data file. We read a single trace
  40  * into memory and check for the magic signature.
  41  */
  42 bool is_blktrace(const char *filename, int *need_swap)
  43 {
  44         struct blk_io_trace t;
  45         int fd, ret;
  46
  47         fd = open(filename, O_RDONLY);
  48         if (fd < 0)
  49                 return false;
  50
  51         ret = read(fd, &t, sizeof(t));
  52         close(fd);
  53
  54         if (ret < 0) {
  55                 perror("read blktrace");
  56                 return false;
  57         } else if (ret != sizeof(t)) {
  58                 log_err("fio: short read on blktrace file\n");
  59                 return false;
  60         }
  61
  62         if ((t.magic & 0xffffff00) == BLK_IO_TRACE_MAGIC) {
  63                 *need_swap = 0;
  64                 return true;
  65         }
  66
  67         /*
  68          * Maybe it needs to be endian swapped...
  69          */
  70         t.magic = fio_swap32(t.magic);
  71         if ((t.magic & 0xffffff00) == BLK_IO_TRACE_MAGIC) {
  72                 *need_swap = 1;
  73                 return true;
  74         }
  75
  76         return false;
  77 }
  78
  79 #define FMINORBITS      20
  80 #define FMINORMASK      ((1U << FMINORBITS) - 1)
  81 #define FMAJOR(dev)     ((unsigned int) ((dev) >> FMINORBITS))
  82 #define FMINOR(dev)     ((unsigned int) ((dev) & FMINORMASK))
  83
  84 static void trace_add_open_close_event(struct thread_data *td, int fileno, enum file_log_act action)
  85 {
  86         struct io_piece *ipo;
  87
  88         ipo = calloc(1, sizeof(*ipo));
  89         init_ipo(ipo);
  90
  91         ipo->ddir = DDIR_INVAL;
  92         ipo->fileno = fileno;
  93         ipo->file_action = action;
  94         flist_add_tail(&ipo->list, &td->io_log_list);
  95 }
  96
  97 static int trace_add_file(struct thread_data *td, __u32 device,
  98                           struct file_cache *cache)
  99 {
 100         unsigned int maj = FMAJOR(device);
 101         unsigned int min = FMINOR(device);
 102         struct fio_file *f;
 103         char dev[256];
 104         unsigned int i;
 105
 106         if (cache->maj == maj && cache->min == min)
 107                 return cache->fileno;
 108
 109         cache->maj = maj;
 110         cache->min = min;
 111
 112         /*
 113          * check for this file in our list
 114          */
 115         for_each_file(td, f, i)
 116                 if (f->major == maj && f->minor == min) {
 117                         cache->fileno = f->fileno;
 118                         return cache->fileno;
 119                 }
 120
 121         strcpy(dev, "/dev");
 122         if (blktrace_lookup_device(td->o.replay_redirect, dev, maj, min)) {
 123                 int fileno;
 124
 125                 if (td->o.replay_redirect)
 126                         dprint(FD_BLKTRACE, "device lookup: %d/%d\n overridden"
 127                                         " with: %s\n", maj, min,
 128                                         td->o.replay_redirect);
 129                 else
 130                         dprint(FD_BLKTRACE, "device lookup: %d/%d\n", maj, min);
 131
 132                 dprint(FD_BLKTRACE, "add devices %s\n", dev);
 133                 fileno = add_file_exclusive(td, dev);
 134                 td->o.open_files++;
 135                 td->files[fileno]->major = maj;
 136                 td->files[fileno]->minor = min;
 137                 trace_add_open_close_event(td, fileno, FIO_LOG_OPEN_FILE);
 138                 cache->fileno = fileno;
 139         }
 140
 141         return cache->fileno;
 142 }
 143
 144 static void t_bytes_align(struct thread_options *o, struct blk_io_trace *t)
 145 {
 146         if (!o->replay_align)
 147                 return;
 148
 149         t->bytes = (t->bytes + o->replay_align - 1) & ~(o->replay_align - 1);
 150 }
 151
 152 /*
 153  * Store blk_io_trace data in an ipo for later retrieval.
 154  */
 155 static void store_ipo(struct thread_data *td, unsigned long long offset,
 156                       unsigned int bytes, int rw, unsigned long long ttime,
 157                       int fileno)
 158 {
 159         struct io_piece *ipo;
 160
 161         ipo = calloc(1, sizeof(*ipo));
 162         init_ipo(ipo);
 163
 164         ipo->offset = offset * 512;
 165         if (td->o.replay_scale)
 166                 ipo->offset = ipo->offset / td->o.replay_scale;
 167         ipo_bytes_align(td->o.replay_align, ipo);
 168         ipo->len = bytes;
 169         ipo->delay = ttime / 1000;
 170         if (rw)
 171                 ipo->ddir = DDIR_WRITE;
 172         else
 173                 ipo->ddir = DDIR_READ;
 174         ipo->fileno = fileno;
 175
 176         dprint(FD_BLKTRACE, "store ddir=%d, off=%llu, len=%lu, delay=%lu\n",
 177                                                         ipo->ddir, ipo->offset,
 178                                                         ipo->len, ipo->delay);
 179         queue_io_piece(td, ipo);
 180 }
 181
 182 static bool handle_trace_notify(struct blk_io_trace *t)
 183 {
 184         switch (t->action) {
 185         case BLK_TN_PROCESS:
 186                 dprint(FD_BLKTRACE, "got process notify: %x, %d\n",
 187                                 t->action, t->pid);
 188                 break;
 189         case BLK_TN_TIMESTAMP:
 190                 dprint(FD_BLKTRACE, "got timestamp notify: %x, %d\n",
 191                                 t->action, t->pid);
 192                 break;
 193         case BLK_TN_MESSAGE:
 194                 break;
 195         default:
 196                 dprint(FD_BLKTRACE, "unknown trace act %x\n", t->action);
 197                 break;
 198         }
 199         return false;
 200 }
 201
 202 static bool handle_trace_discard(struct thread_data *td,
 203                                  struct blk_io_trace *t,
 204                                  unsigned long long ttime,
 205                                  unsigned long *ios, unsigned long long *bs,
 206                                  struct file_cache *cache)
 207 {
 208         struct io_piece *ipo;
 209         int fileno;
 210
 211         if (td->o.replay_skip & (1u << DDIR_TRIM))
 212                 return false;
 213
 214         ipo = calloc(1, sizeof(*ipo));
 215         init_ipo(ipo);
 216         fileno = trace_add_file(td, t->device, cache);
 217
 218         ios[DDIR_TRIM]++;
 219         if (t->bytes > bs[DDIR_TRIM])
 220                 bs[DDIR_TRIM] = t->bytes;
 221
 222         td->o.size += t->bytes;
 223
 224         INIT_FLIST_HEAD(&ipo->list);
 225
 226         ipo->offset = t->sector * 512;
 227         if (td->o.replay_scale)
 228                 ipo->offset = ipo->offset / td->o.replay_scale;
 229         ipo_bytes_align(td->o.replay_align, ipo);
 230         ipo->len = t->bytes;
 231         ipo->delay = ttime / 1000;
 232         ipo->ddir = DDIR_TRIM;
 233         ipo->fileno = fileno;
 234
 235         dprint(FD_BLKTRACE, "store discard, off=%llu, len=%lu, delay=%lu\n",
 236                                                         ipo->offset, ipo->len,
 237                                                         ipo->delay);
 238         queue_io_piece(td, ipo);
 239         return true;
 240 }
 241
 242 static void dump_trace(struct blk_io_trace *t)
 243 {
 244         log_err("blktrace: ignoring zero byte trace: action=%x\n", t->action);
 245 }
 246
 247 static bool handle_trace_fs(struct thread_data *td, struct blk_io_trace *t,
 248                             unsigned long long ttime, unsigned long *ios,
 249                             unsigned long long *bs, struct file_cache *cache)
 250 {
 251         int rw;
 252         int fileno;
 253
 254         fileno = trace_add_file(td, t->device, cache);
 255
 256         rw = (t->action & BLK_TC_ACT(BLK_TC_WRITE)) != 0;
 257
 258         if (rw) {
 259                 if (td->o.replay_skip & (1u << DDIR_WRITE))
 260                         return false;
 261         } else {
 262                 if (td->o.replay_skip & (1u << DDIR_READ))
 263                         return false;
 264         }
 265
 266         if (!t->bytes) {
 267                 if (!fio_did_warn(FIO_WARN_BTRACE_ZERO))
 268                         dump_trace(t);
 269                 return false;
 270         }
 271
 272         if (t->bytes > bs[rw])
 273                 bs[rw] = t->bytes;
 274
 275         ios[rw]++;
 276         td->o.size += t->bytes;
 277         store_ipo(td, t->sector, t->bytes, rw, ttime, fileno);
 278         return true;
 279 }
 280
 281 static bool handle_trace_flush(struct thread_data *td, struct blk_io_trace *t,
 282                                unsigned long long ttime, unsigned long *ios,
 283                                struct file_cache *cache)
 284 {
 285         struct io_piece *ipo;
 286         int fileno;
 287
 288         if (td->o.replay_skip & (1u << DDIR_SYNC))
 289                 return false;
 290
 291         ipo = calloc(1, sizeof(*ipo));
 292         init_ipo(ipo);
 293         fileno = trace_add_file(td, t->device, cache);
 294
 295         ipo->delay = ttime / 1000;
 296         ipo->ddir = DDIR_SYNC;
 297         ipo->fileno = fileno;
 298
 299         ios[DDIR_SYNC]++;
 300         dprint(FD_BLKTRACE, "store flush delay=%lu\n", ipo->delay);
 301
 302         if (!(td->flags & TD_F_SYNCS))
 303                 td->flags |= TD_F_SYNCS;
 304
 305         queue_io_piece(td, ipo);
 306         return true;
 307 }
 308
 309 /*
 310  * We only care for queue traces, most of the others are side effects
 311  * due to internal workings of the block layer.
 312  */
 313 static bool queue_trace(struct thread_data *td, struct blk_io_trace *t,
 314                          unsigned long *ios, unsigned long long *bs,
 315                          struct file_cache *cache)
 316 {
 317         unsigned long long *last_ttime = &td->io_log_last_ttime;
 318         unsigned long long delay = 0;
 319
 320         if ((t->action & 0xffff) != __BLK_TA_QUEUE)
 321                 return false;
 322
 323         if (!(t->action & BLK_TC_ACT(BLK_TC_NOTIFY))) {
 324                 delay = delay_since_ttime(td, t->time);
 325                 *last_ttime = t->time;
 326         }
 327
 328         t_bytes_align(&td->o, t);
 329
 330         if (t->action & BLK_TC_ACT(BLK_TC_NOTIFY))
 331                 return handle_trace_notify(t);
 332         else if (t->action & BLK_TC_ACT(BLK_TC_DISCARD))
 333                 return handle_trace_discard(td, t, delay, ios, bs, cache);
 334         else if (t->action & BLK_TC_ACT(BLK_TC_FLUSH))
 335                 return handle_trace_flush(td, t, delay, ios, cache);
 336         else
 337                 return handle_trace_fs(td, t, delay, ios, bs, cache);
 338 }
 339
 340 static void byteswap_trace(struct blk_io_trace *t)
 341 {
 342         t->magic = fio_swap32(t->magic);
 343         t->sequence = fio_swap32(t->sequence);
 344         t->time = fio_swap64(t->time);
 345         t->sector = fio_swap64(t->sector);
 346         t->bytes = fio_swap32(t->bytes);
 347         t->action = fio_swap32(t->action);
 348         t->pid = fio_swap32(t->pid);
 349         t->device = fio_swap32(t->device);
 350         t->cpu = fio_swap32(t->cpu);
 351         t->error = fio_swap16(t->error);
 352         t->pdu_len = fio_swap16(t->pdu_len);
 353 }
 354
 355 static bool t_is_write(struct blk_io_trace *t)
 356 {
 357         return (t->action & BLK_TC_ACT(BLK_TC_WRITE | BLK_TC_DISCARD)) != 0;
 358 }
 359
 360 static enum fio_ddir t_get_ddir(struct blk_io_trace *t)
 361 {
 362         if (t->action & BLK_TC_ACT(BLK_TC_READ))
 363                 return DDIR_READ;
 364         else if (t->action & BLK_TC_ACT(BLK_TC_WRITE))
 365                 return DDIR_WRITE;
 366         else if (t->action & BLK_TC_ACT(BLK_TC_DISCARD))
 367                 return DDIR_TRIM;
 368
 369         return DDIR_INVAL;
 370 }
 371
 372 static void depth_inc(struct blk_io_trace *t, int *depth)
 373 {
 374         enum fio_ddir ddir;
 375
 376         ddir = t_get_ddir(t);
 377         if (ddir != DDIR_INVAL)
 378                 depth[ddir]++;
 379 }
 380
 381 static void depth_dec(struct blk_io_trace *t, int *depth)
 382 {
 383         enum fio_ddir ddir;
 384
 385         ddir = t_get_ddir(t);
 386         if (ddir != DDIR_INVAL)
 387                 depth[ddir]--;
 388 }
 389
 390 static void depth_end(struct blk_io_trace *t, int *this_depth, int *depth)
 391 {
 392         enum fio_ddir ddir = DDIR_INVAL;
 393
 394         ddir = t_get_ddir(t);
 395         if (ddir != DDIR_INVAL) {
 396                 depth[ddir] = max(depth[ddir], this_depth[ddir]);
 397                 this_depth[ddir] = 0;
 398         }
 399 }
 400
 401 /*
 402  * Load a blktrace file by reading all the blk_io_trace entries, and storing
 403  * them as io_pieces like the fio text version would do.
 404  */
 405 bool init_blktrace_read(struct thread_data *td, const char *filename, int need_swap)
 406 {
 407         int old_state;
 408
 409         td->io_log_rfile = fopen(filename, "rb");
 410         if (!td->io_log_rfile) {
 411                 td_verror(td, errno, "open blktrace file");
 412                 goto err;
 413         }
 414         td->io_log_blktrace_swap = need_swap;
 415         td->io_log_last_ttime = 0;
 416         td->o.size = 0;
 417
 418         free_release_files(td);
 419
 420         old_state = td_bump_runstate(td, TD_SETTING_UP);
 421
 422         if (!read_blktrace(td)) {
 423                 goto err;
 424         }
 425
 426         td_restore_runstate(td, old_state);
 427
 428         if (!td->files_index) {
 429                 log_err("fio: did not find replay device(s)\n");
 430                 return false;
 431         }
 432
 433         return true;
 434
 435 err:
 436         if (td->io_log_rfile) {
 437                 fclose(td->io_log_rfile);
 438                 td->io_log_rfile = NULL;
 439         }
 440         return false;
 441 }
 442
 443 bool read_blktrace(struct thread_data* td)
 444 {
 445         struct blk_io_trace t;
 446         struct file_cache cache = {
 447                 .maj = ~0U,
 448                 .min = ~0U,
 449         };
 450         unsigned long ios[DDIR_RWDIR_SYNC_CNT] = { };
 451         unsigned long long rw_bs[DDIR_RWDIR_CNT] = { };
 452         unsigned long skipped_writes;
 453         FILE *f = td->io_log_rfile;
 454         int i, max_depth;
 455         struct fio_file *fiof;
 456         int this_depth[DDIR_RWDIR_CNT] = { };
 457         int depth[DDIR_RWDIR_CNT] = { };
 458         int64_t items_to_fetch = 0;
 459
 460         if (td->o.read_iolog_chunked) {
 461                 items_to_fetch = iolog_items_to_fetch(td);
 462                 if (!items_to_fetch)
 463                         return true;
 464         }
 465
 466         skipped_writes = 0;
 467         do {
 468                 int ret = fread(&t, 1, sizeof(t), f);
 469
 470                 if (ferror(f)) {
 471                         td_verror(td, errno, "read blktrace file");
 472                         goto err;
 473                 } else if (feof(f)) {
 474                         break;
 475                 } else if (ret < (int) sizeof(t)) {
 476                         log_err("fio: iolog short read\n");
 477                         break;
 478                 }
 479
 480                 if (td->io_log_blktrace_swap)
 481                         byteswap_trace(&t);
 482
 483                 if ((t.magic & 0xffffff00) != BLK_IO_TRACE_MAGIC) {
 484                         log_err("fio: bad magic in blktrace data: %x\n",
 485                                                                 t.magic);
 486                         goto err;
 487                 }
 488                 if ((t.magic & 0xff) != BLK_IO_TRACE_VERSION) {
 489                         log_err("fio: bad blktrace version %d\n",
 490                                                                 t.magic & 0xff);
 491                         goto err;
 492                 }
 493                 ret = discard_pdu(f, &t);
 494                 if (ret < 0) {
 495                         td_verror(td, -ret, "blktrace lseek");
 496                         goto err;
 497                 }
 498                 if ((t.action & BLK_TC_ACT(BLK_TC_NOTIFY)) == 0) {
 499                         if ((t.action & 0xffff) == __BLK_TA_QUEUE)
 500                                 depth_inc(&t, this_depth);
 501                         else if (((t.action & 0xffff) == __BLK_TA_BACKMERGE) ||
 502                                 ((t.action & 0xffff) == __BLK_TA_FRONTMERGE))
 503                                 depth_dec(&t, this_depth);
 504                         else if ((t.action & 0xffff) == __BLK_TA_COMPLETE)
 505                                 depth_end(&t, this_depth, depth);
 506
 507                         if (t_is_write(&t) && read_only) {
 508                                 skipped_writes++;
 509                                 continue;
 510                         }
 511                 }
 512
 513                 if (!queue_trace(td, &t, ios, rw_bs, &cache))
 514                         continue;
 515
 516                 if (td->o.read_iolog_chunked) {
 517                         td->io_log_current++;
 518                         items_to_fetch--;
 519                         if (items_to_fetch == 0)
 520                                 break;
 521                 }
 522         } while (1);
 523
 524         if (td->o.read_iolog_chunked) {
 525                 td->io_log_highmark = td->io_log_current;
 526                 td->io_log_checkmark = (td->io_log_highmark + 1) / 2;
 527                 fio_gettime(&td->io_log_highmark_time, NULL);
 528         }
 529
 530         if (skipped_writes)
 531                 log_err("fio: %s skips replay of %lu writes due to read-only\n",
 532                                                 td->o.name, skipped_writes);
 533
 534         if (td->o.read_iolog_chunked) {
 535                 if (td->io_log_current == 0) {
 536                         return false;
 537                 }
 538                 td->o.td_ddir = TD_DDIR_RW;
 539                 if ((rw_bs[DDIR_READ] > td->o.max_bs[DDIR_READ] ||
 540                      rw_bs[DDIR_WRITE] > td->o.max_bs[DDIR_WRITE] ||
 541                      rw_bs[DDIR_TRIM] > td->o.max_bs[DDIR_TRIM]) &&
 542                     td->orig_buffer)
 543                 {
 544                         td->o.max_bs[DDIR_READ] = max(td->o.max_bs[DDIR_READ], rw_bs[DDIR_READ]);
 545                         td->o.max_bs[DDIR_WRITE] = max(td->o.max_bs[DDIR_WRITE], rw_bs[DDIR_WRITE]);
 546                         td->o.max_bs[DDIR_TRIM] = max(td->o.max_bs[DDIR_TRIM], rw_bs[DDIR_TRIM]);
 547                         io_u_quiesce(td);
 548                         free_io_mem(td);
 549                         if (init_io_u_buffers(td))
 550                                 return false;
 551                 }
 552                 return true;
 553         }
 554
 555         for_each_file(td, fiof, i)
 556                 trace_add_open_close_event(td, fiof->fileno, FIO_LOG_CLOSE_FILE);
 557
 558         fclose(td->io_log_rfile);
 559         td->io_log_rfile = NULL;
 560
 561         /*
 562          * For stacked devices, we don't always get a COMPLETE event so
 563          * the depth grows to insane values. Limit it to something sane(r).
 564          */
 565         max_depth = 0;
 566         for (i = 0; i < DDIR_RWDIR_CNT; i++) {
 567                 if (depth[i] > 1024)
 568                         depth[i] = 1024;
 569                 else if (!depth[i] && ios[i])
 570                         depth[i] = 1;
 571                 max_depth = max(depth[i], max_depth);
 572         }
 573
 574         if (!ios[DDIR_READ] && !ios[DDIR_WRITE] && !ios[DDIR_TRIM] &&
 575             !ios[DDIR_SYNC]) {
 576                 log_err("fio: found no ios in blktrace data\n");
 577                 return false;
 578         }
 579
 580         td->o.td_ddir = 0;
 581         if (ios[DDIR_READ]) {
 582                 td->o.td_ddir |= TD_DDIR_READ;
 583                 td->o.max_bs[DDIR_READ] = rw_bs[DDIR_READ];
 584         }
 585         if (ios[DDIR_WRITE]) {
 586                 td->o.td_ddir |= TD_DDIR_WRITE;
 587                 td->o.max_bs[DDIR_WRITE] = rw_bs[DDIR_WRITE];
 588         }
 589         if (ios[DDIR_TRIM]) {
 590                 td->o.td_ddir |= TD_DDIR_TRIM;
 591                 td->o.max_bs[DDIR_TRIM] = rw_bs[DDIR_TRIM];
 592         }
 593
 594         /*
 595          * If depth wasn't manually set, use probed depth
 596          */
 597         if (!fio_option_is_set(&td->o, iodepth))
 598                 td->o.iodepth = td->o.iodepth_low = max_depth;
 599
 600         return true;
 601 err:
 602         fclose(f);
 603         return false;
 604 }
 605
 606 static int init_merge_param_list(fio_fp64_t *vals, struct blktrace_cursor *bcs,
 607                                  int nr_logs, int def, size_t off)
 608 {
 609         int i = 0, len = 0;
 610
 611         while (len < FIO_IO_U_LIST_MAX_LEN && vals[len].u.f != 0.0)
 612                 len++;
 613
 614         if (len && len != nr_logs)
 615                 return len;
 616
 617         for (i = 0; i < nr_logs; i++) {
 618                 int *val = (int *)((char *)&bcs[i] + off);
 619                 *val = def;
 620                 if (len)
 621                         *val = (int)vals[i].u.f;
 622         }
 623
 624         return 0;
 625
 626 }
 627
 628 static int find_earliest_io(struct blktrace_cursor *bcs, int nr_logs)
 629 {
 630         __u64 time = ~(__u64)0;
 631         int idx = 0, i;
 632
 633         for (i = 0; i < nr_logs; i++) {
 634                 if (bcs[i].t.time < time) {
 635                         time = bcs[i].t.time;
 636                         idx = i;
 637                 }
 638         }
 639
 640         return idx;
 641 }
 642
 643 static void merge_finish_file(struct blktrace_cursor *bcs, int i, int *nr_logs)
 644 {
 645         bcs[i].iter++;
 646         if (bcs[i].iter < bcs[i].nr_iter) {
 647                 fseek(bcs[i].f, 0, SEEK_SET);
 648                 return;
 649         }
 650
 651         *nr_logs -= 1;
 652
 653         /* close file */
 654         fclose(bcs[i].f);
 655
 656         /* keep active files contiguous */
 657         memmove(&bcs[i], &bcs[*nr_logs], sizeof(bcs[i]));
 658 }
 659
 660 static int read_trace(struct thread_data *td, struct blktrace_cursor *bc)
 661 {
 662         int ret = 0;
 663         struct blk_io_trace *t = &bc->t;
 664
 665 read_skip:
 666         /* read an io trace */
 667         ret = fread(&t, 1, sizeof(t), bc->f);
 668         if (ferror(bc->f)) {
 669                 td_verror(td, errno, "read blktrace file");
 670                 return ret;
 671         } else if (feof(bc->f)) {
 672                 if (!bc->length)
 673                         bc->length = bc->t.time;
 674                 return ret;
 675         } else if (ret < (int) sizeof(*t)) {
 676                 log_err("fio: iolog short read\n");
 677                 return -1;
 678         }
 679
 680         if (bc->swap)
 681                 byteswap_trace(t);
 682
 683         /* skip over actions that fio does not care about */
 684         if ((t->action & 0xffff) != __BLK_TA_QUEUE ||
 685             t_get_ddir(t) == DDIR_INVAL) {
 686                 ret = discard_pdu(bc->f, t);
 687                 if (ret < 0) {
 688                         td_verror(td, -ret, "blktrace lseek");
 689                         return ret;
 690                 }
 691                 goto read_skip;
 692         }
 693
 694         t->time = (t->time + bc->iter * bc->length) * bc->scalar / 100;
 695
 696         return ret;
 697 }
 698
 699 static int write_trace(FILE *fp, struct blk_io_trace *t)
 700 {
 701         /* pdu is not used so just write out only the io trace */
 702         t->pdu_len = 0;
 703         return fwrite((void *)t, sizeof(*t), 1, fp);
 704 }
 705
 706 int merge_blktrace_iologs(struct thread_data *td)
 707 {
 708         int nr_logs = get_max_str_idx(td->o.read_iolog_file);
 709         struct blktrace_cursor *bcs = malloc(sizeof(struct blktrace_cursor) *
 710                                              nr_logs);
 711         struct blktrace_cursor *bc;
 712         FILE *merge_fp;
 713         char *str, *ptr, *name, *merge_buf;
 714         int i, ret;
 715
 716         ret = init_merge_param_list(td->o.merge_blktrace_scalars, bcs, nr_logs,
 717                                     100, offsetof(struct blktrace_cursor,
 718                                                   scalar));
 719         if (ret) {
 720                 log_err("fio: merge_blktrace_scalars(%d) != nr_logs(%d)\n",
 721                         ret, nr_logs);
 722                 goto err_param;
 723         }
 724
 725         ret = init_merge_param_list(td->o.merge_blktrace_iters, bcs, nr_logs,
 726                                     1, offsetof(struct blktrace_cursor,
 727                                                 nr_iter));
 728         if (ret) {
 729                 log_err("fio: merge_blktrace_iters(%d) != nr_logs(%d)\n",
 730                         ret, nr_logs);
 731                 goto err_param;
 732         }
 733
 734         /* setup output file */
 735         merge_fp = fopen(td->o.merge_blktrace_file, "w");
 736         merge_buf = malloc(128 * 1024);
 737         if (!merge_buf)
 738                 goto err_out_file;
 739         ret = setvbuf(merge_fp, merge_buf, _IOFBF, 128 * 1024);
 740         if (ret)
 741                 goto err_merge_buf;
 742
 743         /* setup input files */
 744         str = ptr = strdup(td->o.read_iolog_file);
 745         nr_logs = 0;
 746         for (i = 0; (name = get_next_str(&ptr)) != NULL; i++) {
 747                 bcs[i].f = fopen(name, "rb");
 748                 if (!bcs[i].f) {
 749                         log_err("fio: could not open file: %s\n", name);
 750                         ret = -errno;
 751                         free(str);
 752                         goto err_file;
 753                 }
 754                 nr_logs++;
 755
 756                 if (!is_blktrace(name, &bcs[i].swap)) {
 757                         log_err("fio: file is not a blktrace: %s\n", name);
 758                         free(str);
 759                         goto err_file;
 760                 }
 761
 762                 ret = read_trace(td, &bcs[i]);
 763                 if (ret < 0) {
 764                         free(str);
 765                         goto err_file;
 766                 } else if (!ret) {
 767                         merge_finish_file(bcs, i, &nr_logs);
 768                         i--;
 769                 }
 770         }
 771         free(str);
 772
 773         /* merge files */
 774         while (nr_logs) {
 775                 i = find_earliest_io(bcs, nr_logs);
 776                 bc = &bcs[i];
 777                 /* skip over the pdu */
 778                 ret = discard_pdu(bc->f, &bc->t);
 779                 if (ret < 0) {
 780                         td_verror(td, -ret, "blktrace lseek");
 781                         goto err_file;
 782                 }
 783
 784                 ret = write_trace(merge_fp, &bc->t);
 785                 ret = read_trace(td, bc);
 786                 if (ret < 0)
 787                         goto err_file;
 788                 else if (!ret)
 789                         merge_finish_file(bcs, i, &nr_logs);
 790         }
 791
 792         /* set iolog file to read from the newly merged file */
 793         td->o.read_iolog_file = td->o.merge_blktrace_file;
 794         ret = 0;
 795
 796 err_file:
 797         /* cleanup */
 798         for (i = 0; i < nr_logs; i++) {
 799                 fclose(bcs[i].f);
 800         }
 801 err_merge_buf:
 802         free(merge_buf);
 803 err_out_file:
 804         fflush(merge_fp);
 805         fclose(merge_fp);
 806 err_param:
 807         free(bcs);
 808
 809         return ret;
 810 }