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
   9 #include "flist.h"
  10 #include "fio.h"
  11 #include "iolog.h"
  12 #include "blktrace.h"
  13 #include "blktrace_api.h"
  14 #include "oslib/linux-dev-lookup.h"
  15
  16 /*
  17  * Just discard the pdu by seeking past it.
  18  */
  19 static int discard_pdu(FILE* f, struct blk_io_trace *t)
  20 {
  21         if (t->pdu_len == 0)
  22                 return 0;
  23
  24         dprint(FD_BLKTRACE, "discard pdu len %u\n", t->pdu_len);
  25         if (fseek(f, t->pdu_len, SEEK_CUR) < 0)
  26                 return -errno;
  27
  28         return t->pdu_len;
  29 }
  30
  31 /*
  32  * Check if this is a blktrace binary data file. We read a single trace
  33  * into memory and check for the magic signature.
  34  */
  35 bool is_blktrace(const char *filename, int *need_swap)
  36 {
  37         struct blk_io_trace t;
  38         int fd, ret;
  39
  40         fd = open(filename, O_RDONLY);
  41         if (fd < 0)
  42                 return false;
  43
  44         ret = read(fd, &t, sizeof(t));
  45         close(fd);
  46
  47         if (ret < 0) {
  48                 perror("read blktrace");
  49                 return false;
  50         } else if (ret != sizeof(t)) {
  51                 log_err("fio: short read on blktrace file\n");
  52                 return false;
  53         }
  54
  55         if ((t.magic & 0xffffff00) == BLK_IO_TRACE_MAGIC) {
  56                 *need_swap = 0;
  57                 return true;
  58         }
  59
  60         /*
  61          * Maybe it needs to be endian swapped...
  62          */
  63         t.magic = fio_swap32(t.magic);
  64         if ((t.magic & 0xffffff00) == BLK_IO_TRACE_MAGIC) {
  65                 *need_swap = 1;
  66                 return true;
  67         }
  68
  69         return false;
  70 }
  71
  72 #define FMINORBITS      20
  73 #define FMINORMASK      ((1U << FMINORBITS) - 1)
  74 #define FMAJOR(dev)     ((unsigned int) ((dev) >> FMINORBITS))
  75 #define FMINOR(dev)     ((unsigned int) ((dev) & FMINORMASK))
  76
  77 static void trace_add_open_close_event(struct thread_data *td, int fileno, enum file_log_act action)
  78 {
  79         struct io_piece *ipo;
  80
  81         ipo = calloc(1, sizeof(*ipo));
  82         init_ipo(ipo);
  83
  84         ipo->ddir = DDIR_INVAL;
  85         ipo->fileno = fileno;
  86         ipo->file_action = action;
  87         flist_add_tail(&ipo->list, &td->io_log_list);
  88 }
  89
  90 static int trace_add_file(struct thread_data *td, __u32 device)
  91 {
  92         static unsigned int last_maj, last_min, last_fileno;
  93         unsigned int maj = FMAJOR(device);
  94         unsigned int min = FMINOR(device);
  95         struct fio_file *f;
  96         char dev[256];
  97         unsigned int i;
  98
  99         if (last_maj == maj && last_min == min)
 100                 return last_fileno;
 101
 102         last_maj = maj;
 103         last_min = min;
 104
 105         /*
 106          * check for this file in our list
 107          */
 108         for_each_file(td, f, i)
 109                 if (f->major == maj && f->minor == min) {
 110                         last_fileno = f->fileno;
 111                         return last_fileno;
 112                 }
 113
 114         strcpy(dev, "/dev");
 115         if (blktrace_lookup_device(td->o.replay_redirect, dev, maj, min)) {
 116                 int fileno;
 117
 118                 if (td->o.replay_redirect)
 119                         dprint(FD_BLKTRACE, "device lookup: %d/%d\n overridden"
 120                                         " with: %s\n", maj, min,
 121                                         td->o.replay_redirect);
 122                 else
 123                         dprint(FD_BLKTRACE, "device lookup: %d/%d\n", maj, min);
 124
 125                 dprint(FD_BLKTRACE, "add devices %s\n", dev);
 126                 fileno = add_file_exclusive(td, dev);
 127                 td->o.open_files++;
 128                 td->files[fileno]->major = maj;
 129                 td->files[fileno]->minor = min;
 130                 trace_add_open_close_event(td, fileno, FIO_LOG_OPEN_FILE);
 131                 last_fileno = fileno;
 132         }
 133
 134         return last_fileno;
 135 }
 136
 137 static void t_bytes_align(struct thread_options *o, struct blk_io_trace *t)
 138 {
 139         if (!o->replay_align)
 140                 return;
 141
 142         t->bytes = (t->bytes + o->replay_align - 1) & ~(o->replay_align - 1);
 143 }
 144
 145 /*
 146  * Store blk_io_trace data in an ipo for later retrieval.
 147  */
 148 static void store_ipo(struct thread_data *td, unsigned long long offset,
 149                       unsigned int bytes, int rw, unsigned long long ttime,
 150                       int fileno)
 151 {
 152         struct io_piece *ipo;
 153
 154         ipo = calloc(1, sizeof(*ipo));
 155         init_ipo(ipo);
 156
 157         ipo->offset = offset * 512;
 158         if (td->o.replay_scale)
 159                 ipo->offset = ipo->offset / td->o.replay_scale;
 160         ipo_bytes_align(td->o.replay_align, ipo);
 161         ipo->len = bytes;
 162         ipo->delay = ttime / 1000;
 163         if (rw)
 164                 ipo->ddir = DDIR_WRITE;
 165         else
 166                 ipo->ddir = DDIR_READ;
 167         ipo->fileno = fileno;
 168
 169         dprint(FD_BLKTRACE, "store ddir=%d, off=%llu, len=%lu, delay=%lu\n",
 170                                                         ipo->ddir, ipo->offset,
 171                                                         ipo->len, ipo->delay);
 172         queue_io_piece(td, ipo);
 173 }
 174
 175 static bool handle_trace_notify(struct blk_io_trace *t)
 176 {
 177         switch (t->action) {
 178         case BLK_TN_PROCESS:
 179                 dprint(FD_BLKTRACE, "got process notify: %x, %d\n",
 180                                 t->action, t->pid);
 181                 break;
 182         case BLK_TN_TIMESTAMP:
 183                 dprint(FD_BLKTRACE, "got timestamp notify: %x, %d\n",
 184                                 t->action, t->pid);
 185                 break;
 186         case BLK_TN_MESSAGE:
 187                 break;
 188         default:
 189                 dprint(FD_BLKTRACE, "unknown trace act %x\n", t->action);
 190                 break;
 191         }
 192         return false;
 193 }
 194
 195 static bool handle_trace_discard(struct thread_data *td,
 196                                  struct blk_io_trace *t,
 197                                  unsigned long long ttime,
 198                                  unsigned long *ios, unsigned long long *bs)
 199 {
 200         struct io_piece *ipo;
 201         int fileno;
 202
 203         if (td->o.replay_skip & (1u << DDIR_TRIM))
 204                 return false;
 205
 206         ipo = calloc(1, sizeof(*ipo));
 207         init_ipo(ipo);
 208         fileno = trace_add_file(td, t->device);
 209
 210         ios[DDIR_TRIM]++;
 211         if (t->bytes > bs[DDIR_TRIM])
 212                 bs[DDIR_TRIM] = t->bytes;
 213
 214         td->o.size += t->bytes;
 215
 216         INIT_FLIST_HEAD(&ipo->list);
 217
 218         ipo->offset = t->sector * 512;
 219         if (td->o.replay_scale)
 220                 ipo->offset = ipo->offset / td->o.replay_scale;
 221         ipo_bytes_align(td->o.replay_align, ipo);
 222         ipo->len = t->bytes;
 223         ipo->delay = ttime / 1000;
 224         ipo->ddir = DDIR_TRIM;
 225         ipo->fileno = fileno;
 226
 227         dprint(FD_BLKTRACE, "store discard, off=%llu, len=%lu, delay=%lu\n",
 228                                                         ipo->offset, ipo->len,
 229                                                         ipo->delay);
 230         queue_io_piece(td, ipo);
 231         return true;
 232 }
 233
 234 static void dump_trace(struct blk_io_trace *t)
 235 {
 236         log_err("blktrace: ignoring zero byte trace: action=%x\n", t->action);
 237 }
 238
 239 static bool handle_trace_fs(struct thread_data *td, struct blk_io_trace *t,
 240                             unsigned long long ttime, unsigned long *ios,
 241                             unsigned long long *bs)
 242 {
 243         int rw;
 244         int fileno;
 245
 246         fileno = trace_add_file(td, t->device);
 247
 248         rw = (t->action & BLK_TC_ACT(BLK_TC_WRITE)) != 0;
 249
 250         if (rw) {
 251                 if (td->o.replay_skip & (1u << DDIR_WRITE))
 252                         return false;
 253         } else {
 254                 if (td->o.replay_skip & (1u << DDIR_READ))
 255                         return false;
 256         }
 257
 258         if (!t->bytes) {
 259                 if (!fio_did_warn(FIO_WARN_BTRACE_ZERO))
 260                         dump_trace(t);
 261                 return false;
 262         }
 263
 264         if (t->bytes > bs[rw])
 265                 bs[rw] = t->bytes;
 266
 267         ios[rw]++;
 268         td->o.size += t->bytes;
 269         store_ipo(td, t->sector, t->bytes, rw, ttime, fileno);
 270         return true;
 271 }
 272
 273 static bool handle_trace_flush(struct thread_data *td, struct blk_io_trace *t,
 274                                unsigned long long ttime, unsigned long *ios)
 275 {
 276         struct io_piece *ipo;
 277         int fileno;
 278
 279         if (td->o.replay_skip & (1u << DDIR_SYNC))
 280                 return false;
 281
 282         ipo = calloc(1, sizeof(*ipo));
 283         init_ipo(ipo);
 284         fileno = trace_add_file(td, t->device);
 285
 286         ipo->delay = ttime / 1000;
 287         ipo->ddir = DDIR_SYNC;
 288         ipo->fileno = fileno;
 289
 290         ios[DDIR_SYNC]++;
 291         dprint(FD_BLKTRACE, "store flush delay=%lu\n", ipo->delay);
 292         queue_io_piece(td, ipo);
 293         return true;
 294 }
 295
 296 /*
 297  * We only care for queue traces, most of the others are side effects
 298  * due to internal workings of the block layer.
 299  */
 300 static bool queue_trace(struct thread_data *td, struct blk_io_trace *t,
 301                          unsigned long *ios, unsigned long long *bs)
 302 {
 303         static unsigned long long last_ttime;
 304         unsigned long long delay = 0;
 305
 306         if ((t->action & 0xffff) != __BLK_TA_QUEUE)
 307                 return false;
 308
 309         if (!(t->action & BLK_TC_ACT(BLK_TC_NOTIFY))) {
 310                 if (!last_ttime || td->o.no_stall)
 311                         delay = 0;
 312                 else if (td->o.replay_time_scale == 100)
 313                         delay = t->time - last_ttime;
 314                 else {
 315                         double tmp = t->time - last_ttime;
 316                         double scale;
 317
 318                         scale = (double) 100.0 / (double) td->o.replay_time_scale;
 319                         tmp *= scale;
 320                         delay = tmp;
 321                 }
 322                 last_ttime = t->time;
 323         }
 324
 325         t_bytes_align(&td->o, t);
 326
 327         if (t->action & BLK_TC_ACT(BLK_TC_NOTIFY))
 328                 return handle_trace_notify(t);
 329         else if (t->action & BLK_TC_ACT(BLK_TC_DISCARD))
 330                 return handle_trace_discard(td, t, delay, ios, bs);
 331         else if (t->action & BLK_TC_ACT(BLK_TC_FLUSH))
 332                 return handle_trace_flush(td, t, delay, ios);
 333         else
 334                 return handle_trace_fs(td, t, delay, ios, bs);
 335 }
 336
 337 static void byteswap_trace(struct blk_io_trace *t)
 338 {
 339         t->magic = fio_swap32(t->magic);
 340         t->sequence = fio_swap32(t->sequence);
 341         t->time = fio_swap64(t->time);
 342         t->sector = fio_swap64(t->sector);
 343         t->bytes = fio_swap32(t->bytes);
 344         t->action = fio_swap32(t->action);
 345         t->pid = fio_swap32(t->pid);
 346         t->device = fio_swap32(t->device);
 347         t->cpu = fio_swap32(t->cpu);
 348         t->error = fio_swap16(t->error);
 349         t->pdu_len = fio_swap16(t->pdu_len);
 350 }
 351
 352 static bool t_is_write(struct blk_io_trace *t)
 353 {
 354         return (t->action & BLK_TC_ACT(BLK_TC_WRITE | BLK_TC_DISCARD)) != 0;
 355 }
 356
 357 static enum fio_ddir t_get_ddir(struct blk_io_trace *t)
 358 {
 359         if (t->action & BLK_TC_ACT(BLK_TC_READ))
 360                 return DDIR_READ;
 361         else if (t->action & BLK_TC_ACT(BLK_TC_WRITE))
 362                 return DDIR_WRITE;
 363         else if (t->action & BLK_TC_ACT(BLK_TC_DISCARD))
 364                 return DDIR_TRIM;
 365
 366         return DDIR_INVAL;
 367 }
 368
 369 static void depth_inc(struct blk_io_trace *t, int *depth)
 370 {
 371         enum fio_ddir ddir;
 372
 373         ddir = t_get_ddir(t);
 374         if (ddir != DDIR_INVAL)
 375                 depth[ddir]++;
 376 }
 377
 378 static void depth_dec(struct blk_io_trace *t, int *depth)
 379 {
 380         enum fio_ddir ddir;
 381
 382         ddir = t_get_ddir(t);
 383         if (ddir != DDIR_INVAL)
 384                 depth[ddir]--;
 385 }
 386
 387 static void depth_end(struct blk_io_trace *t, int *this_depth, int *depth)
 388 {
 389         enum fio_ddir ddir = DDIR_INVAL;
 390
 391         ddir = t_get_ddir(t);
 392         if (ddir != DDIR_INVAL) {
 393                 depth[ddir] = max(depth[ddir], this_depth[ddir]);
 394                 this_depth[ddir] = 0;
 395         }
 396 }
 397
 398 /*
 399  * Load a blktrace file by reading all the blk_io_trace entries, and storing
 400  * them as io_pieces like the fio text version would do.
 401  */
 402 bool init_blktrace_read(struct thread_data *td, const char *filename, int need_swap)
 403 {
 404         int old_state;
 405
 406         td->io_log_rfile = fopen(filename, "rb");
 407         if (!td->io_log_rfile) {
 408                 td_verror(td, errno, "open blktrace file");
 409                 goto err;
 410         }
 411         td->io_log_blktrace_swap = need_swap;
 412         td->o.size = 0;
 413
 414         free_release_files(td);
 415
 416         old_state = td_bump_runstate(td, TD_SETTING_UP);
 417
 418         if (!read_blktrace(td)) {
 419                 goto err;
 420         }
 421
 422         td_restore_runstate(td, old_state);
 423
 424         if (!td->files_index) {
 425                 log_err("fio: did not find replay device(s)\n");
 426                 return false;
 427         }
 428
 429         return true;
 430
 431 err:
 432         if (td->io_log_rfile) {
 433                 fclose(td->io_log_rfile);
 434                 td->io_log_rfile = NULL;
 435         }
 436         return false;
 437 }
 438
 439 bool read_blktrace(struct thread_data* td)
 440 {
 441         struct blk_io_trace t;
 442         unsigned long ios[DDIR_RWDIR_SYNC_CNT] = { };
 443         unsigned long long rw_bs[DDIR_RWDIR_CNT] = { };
 444         unsigned long skipped_writes;
 445         FILE *f = td->io_log_rfile;
 446         int i, max_depth;
 447         struct fio_file *fiof;
 448         int this_depth[DDIR_RWDIR_CNT] = { };
 449         int depth[DDIR_RWDIR_CNT] = { };
 450         int64_t items_to_fetch = 0;
 451
 452         if (td->o.read_iolog_chunked) {
 453                 items_to_fetch = iolog_items_to_fetch(td);
 454                 if (!items_to_fetch)
 455                         return true;
 456         }
 457
 458         skipped_writes = 0;
 459         do {
 460                 int ret = fread(&t, 1, sizeof(t), f);
 461
 462                 if (ferror(f)) {
 463                         td_verror(td, errno, "read blktrace file");
 464                         goto err;
 465                 } else if (feof(f)) {
 466                         break;
 467                 } else if (ret < (int) sizeof(t)) {
 468                         log_err("fio: iolog short read\n");
 469                         break;
 470                 }
 471
 472                 if (td->io_log_blktrace_swap)
 473                         byteswap_trace(&t);
 474
 475                 if ((t.magic & 0xffffff00) != BLK_IO_TRACE_MAGIC) {
 476                         log_err("fio: bad magic in blktrace data: %x\n",
 477                                                                 t.magic);
 478                         goto err;
 479                 }
 480                 if ((t.magic & 0xff) != BLK_IO_TRACE_VERSION) {
 481                         log_err("fio: bad blktrace version %d\n",
 482                                                                 t.magic & 0xff);
 483                         goto err;
 484                 }
 485                 ret = discard_pdu(f, &t);
 486                 if (ret < 0) {
 487                         td_verror(td, -ret, "blktrace lseek");
 488                         goto err;
 489                 }
 490                 if ((t.action & BLK_TC_ACT(BLK_TC_NOTIFY)) == 0) {
 491                         if ((t.action & 0xffff) == __BLK_TA_QUEUE)
 492                                 depth_inc(&t, this_depth);
 493                         else if (((t.action & 0xffff) == __BLK_TA_BACKMERGE) ||
 494                                 ((t.action & 0xffff) == __BLK_TA_FRONTMERGE))
 495                                 depth_dec(&t, this_depth);
 496                         else if ((t.action & 0xffff) == __BLK_TA_COMPLETE)
 497                                 depth_end(&t, this_depth, depth);
 498
 499                         if (t_is_write(&t) && read_only) {
 500                                 skipped_writes++;
 501                                 continue;
 502                         }
 503                 }
 504
 505                 if (!queue_trace(td, &t, ios, rw_bs))
 506                         continue;
 507
 508                 if (td->o.read_iolog_chunked) {
 509                         td->io_log_current++;
 510                         items_to_fetch--;
 511                         if (items_to_fetch == 0)
 512                                 break;
 513                 }
 514         } while (1);
 515
 516         if (td->o.read_iolog_chunked) {
 517                 td->io_log_highmark = td->io_log_current;
 518                 td->io_log_checkmark = (td->io_log_highmark + 1) / 2;
 519                 fio_gettime(&td->io_log_highmark_time, NULL);
 520         }
 521
 522         if (skipped_writes)
 523                 log_err("fio: %s skips replay of %lu writes due to read-only\n",
 524                                                 td->o.name, skipped_writes);
 525
 526         if (td->o.read_iolog_chunked) {
 527                 if (td->io_log_current == 0) {
 528                         return false;
 529                 }
 530                 td->o.td_ddir = TD_DDIR_RW;
 531                 if ((rw_bs[DDIR_READ] > td->o.max_bs[DDIR_READ] ||
 532                      rw_bs[DDIR_WRITE] > td->o.max_bs[DDIR_WRITE] ||
 533                      rw_bs[DDIR_TRIM] > td->o.max_bs[DDIR_TRIM]) &&
 534                     td->orig_buffer)
 535                 {
 536                         td->o.max_bs[DDIR_READ] = max(td->o.max_bs[DDIR_READ], rw_bs[DDIR_READ]);
 537                         td->o.max_bs[DDIR_WRITE] = max(td->o.max_bs[DDIR_WRITE], rw_bs[DDIR_WRITE]);
 538                         td->o.max_bs[DDIR_TRIM] = max(td->o.max_bs[DDIR_TRIM], rw_bs[DDIR_TRIM]);
 539                         io_u_quiesce(td);
 540                         free_io_mem(td);
 541                         init_io_u_buffers(td);
 542                 }
 543                 return true;
 544         }
 545
 546         for_each_file(td, fiof, i)
 547                 trace_add_open_close_event(td, fiof->fileno, FIO_LOG_CLOSE_FILE);
 548
 549         fclose(td->io_log_rfile);
 550         td->io_log_rfile = NULL;
 551
 552         /*
 553          * For stacked devices, we don't always get a COMPLETE event so
 554          * the depth grows to insane values. Limit it to something sane(r).
 555          */
 556         max_depth = 0;
 557         for (i = 0; i < DDIR_RWDIR_CNT; i++) {
 558                 if (depth[i] > 1024)
 559                         depth[i] = 1024;
 560                 else if (!depth[i] && ios[i])
 561                         depth[i] = 1;
 562                 max_depth = max(depth[i], max_depth);
 563         }
 564
 565         if (!ios[DDIR_READ] && !ios[DDIR_WRITE] && !ios[DDIR_TRIM] &&
 566             !ios[DDIR_SYNC]) {
 567                 log_err("fio: found no ios in blktrace data\n");
 568                 return false;
 569         }
 570
 571         td->o.td_ddir = 0;
 572         if (ios[DDIR_READ]) {
 573                 td->o.td_ddir |= TD_DDIR_READ;
 574                 td->o.max_bs[DDIR_READ] = rw_bs[DDIR_READ];
 575         }
 576         if (ios[DDIR_WRITE]) {
 577                 td->o.td_ddir |= TD_DDIR_WRITE;
 578                 td->o.max_bs[DDIR_WRITE] = rw_bs[DDIR_WRITE];
 579         }
 580         if (ios[DDIR_TRIM]) {
 581                 td->o.td_ddir |= TD_DDIR_TRIM;
 582                 td->o.max_bs[DDIR_TRIM] = rw_bs[DDIR_TRIM];
 583         }
 584
 585         /*
 586          * We need to do direct/raw ios to the device, to avoid getting
 587          * read-ahead in our way. But only do so if the minimum block size
 588          * is a multiple of 4k, otherwise we don't know if it's safe to do so.
 589          */
 590         if (!fio_option_is_set(&td->o, odirect) && !(td_min_bs(td) & 4095))
 591                 td->o.odirect = 1;
 592
 593         /*
 594          * If depth wasn't manually set, use probed depth
 595          */
 596         if (!fio_option_is_set(&td->o, iodepth))
 597                 td->o.iodepth = td->o.iodepth_low = max_depth;
 598
 599         return true;
 600 err:
 601         fclose(f);
 602         return false;
 603 }
 604
 605 static int init_merge_param_list(fio_fp64_t *vals, struct blktrace_cursor *bcs,
 606                                  int nr_logs, int def, size_t off)
 607 {
 608         int i = 0, len = 0;
 609
 610         while (len < FIO_IO_U_LIST_MAX_LEN && vals[len].u.f != 0.0)
 611                 len++;
 612
 613         if (len && len != nr_logs)
 614                 return len;
 615
 616         for (i = 0; i < nr_logs; i++) {
 617                 int *val = (int *)((char *)&bcs[i] + off);
 618                 *val = def;
 619                 if (len)
 620                         *val = (int)vals[i].u.f;
 621         }
 622
 623         return 0;
 624
 625 }
 626
 627 static int find_earliest_io(struct blktrace_cursor *bcs, int nr_logs)
 628 {
 629         __u64 time = ~(__u64)0;
 630         int idx = 0, i;
 631
 632         for (i = 0; i < nr_logs; i++) {
 633                 if (bcs[i].t.time < time) {
 634                         time = bcs[i].t.time;
 635                         idx = i;
 636                 }
 637         }
 638
 639         return idx;
 640 }
 641
 642 static void merge_finish_file(struct blktrace_cursor *bcs, int i, int *nr_logs)
 643 {
 644         bcs[i].iter++;
 645         if (bcs[i].iter < bcs[i].nr_iter) {
 646                 fseek(bcs[i].f, 0, SEEK_SET);
 647                 return;
 648         }
 649
 650         *nr_logs -= 1;
 651
 652         /* close file */
 653         fclose(bcs[i].f);
 654
 655         /* keep active files contiguous */
 656         memmove(&bcs[i], &bcs[*nr_logs], sizeof(bcs[i]));
 657 }
 658
 659 static int read_trace(struct thread_data *td, struct blktrace_cursor *bc)
 660 {
 661         int ret = 0;
 662         struct blk_io_trace *t = &bc->t;
 663
 664 read_skip:
 665         /* read an io trace */
 666         ret = fread(&t, 1, sizeof(t), bc->f);
 667         if (ferror(bc->f)) {
 668                 td_verror(td, errno, "read blktrace file");
 669                 return ret;
 670         } else if (feof(bc->f)) {
 671                 if (!bc->length)
 672                         bc->length = bc->t.time;
 673                 return ret;
 674         } else if (ret < (int) sizeof(*t)) {
 675                 log_err("fio: iolog short read\n");
 676                 return -1;
 677         }
 678
 679         if (bc->swap)
 680                 byteswap_trace(t);
 681
 682         /* skip over actions that fio does not care about */
 683         if ((t->action & 0xffff) != __BLK_TA_QUEUE ||
 684             t_get_ddir(t) == DDIR_INVAL) {
 685                 ret = discard_pdu(bc->f, t);
 686                 if (ret < 0) {
 687                         td_verror(td, -ret, "blktrace lseek");
 688                         return ret;
 689                 }
 690                 goto read_skip;
 691         }
 692
 693         t->time = (t->time + bc->iter * bc->length) * bc->scalar / 100;
 694
 695         return ret;
 696 }
 697
 698 static int write_trace(FILE *fp, struct blk_io_trace *t)
 699 {
 700         /* pdu is not used so just write out only the io trace */
 701         t->pdu_len = 0;
 702         return fwrite((void *)t, sizeof(*t), 1, fp);
 703 }
 704
 705 int merge_blktrace_iologs(struct thread_data *td)
 706 {
 707         int nr_logs = get_max_str_idx(td->o.read_iolog_file);
 708         struct blktrace_cursor *bcs = malloc(sizeof(struct blktrace_cursor) *
 709                                              nr_logs);
 710         struct blktrace_cursor *bc;
 711         FILE *merge_fp;
 712         char *str, *ptr, *name, *merge_buf;
 713         int i, ret;
 714
 715         ret = init_merge_param_list(td->o.merge_blktrace_scalars, bcs, nr_logs,
 716                                     100, offsetof(struct blktrace_cursor,
 717                                                   scalar));
 718         if (ret) {
 719                 log_err("fio: merge_blktrace_scalars(%d) != nr_logs(%d)\n",
 720                         ret, nr_logs);
 721                 goto err_param;
 722         }
 723
 724         ret = init_merge_param_list(td->o.merge_blktrace_iters, bcs, nr_logs,
 725                                     1, offsetof(struct blktrace_cursor,
 726                                                 nr_iter));
 727         if (ret) {
 728                 log_err("fio: merge_blktrace_iters(%d) != nr_logs(%d)\n",
 729                         ret, nr_logs);
 730                 goto err_param;
 731         }
 732
 733         /* setup output file */
 734         merge_fp = fopen(td->o.merge_blktrace_file, "w");
 735         merge_buf = malloc(128 * 1024);
 736         if (!merge_buf)
 737                 goto err_out_file;
 738         ret = setvbuf(merge_fp, merge_buf, _IOFBF, 128 * 1024);
 739         if (ret)
 740                 goto err_merge_buf;
 741
 742         /* setup input files */
 743         str = ptr = strdup(td->o.read_iolog_file);
 744         nr_logs = 0;
 745         for (i = 0; (name = get_next_str(&ptr)) != NULL; i++) {
 746                 bcs[i].f = fopen(name, "rb");
 747                 if (!bcs[i].f) {
 748                         log_err("fio: could not open file: %s\n", name);
 749                         ret = -errno;
 750                         free(str);
 751                         goto err_file;
 752                 }
 753                 nr_logs++;
 754
 755                 if (!is_blktrace(name, &bcs[i].swap)) {
 756                         log_err("fio: file is not a blktrace: %s\n", name);
 757                         free(str);
 758                         goto err_file;
 759                 }
 760
 761                 ret = read_trace(td, &bcs[i]);
 762                 if (ret < 0) {
 763                         free(str);
 764                         goto err_file;
 765                 } else if (!ret) {
 766                         merge_finish_file(bcs, i, &nr_logs);
 767                         i--;
 768                 }
 769         }
 770         free(str);
 771
 772         /* merge files */
 773         while (nr_logs) {
 774                 i = find_earliest_io(bcs, nr_logs);
 775                 bc = &bcs[i];
 776                 /* skip over the pdu */
 777                 ret = discard_pdu(bc->f, &bc->t);
 778                 if (ret < 0) {
 779                         td_verror(td, -ret, "blktrace lseek");
 780                         goto err_file;
 781                 }
 782
 783                 ret = write_trace(merge_fp, &bc->t);
 784                 ret = read_trace(td, bc);
 785                 if (ret < 0)
 786                         goto err_file;
 787                 else if (!ret)
 788                         merge_finish_file(bcs, i, &nr_logs);
 789         }
 790
 791         /* set iolog file to read from the newly merged file */
 792         td->o.read_iolog_file = td->o.merge_blktrace_file;
 793         ret = 0;
 794
 795 err_file:
 796         /* cleanup */
 797         for (i = 0; i < nr_logs; i++) {
 798                 fclose(bcs[i].f);
 799         }
 800 err_merge_buf:
 801         free(merge_buf);
 802 err_out_file:
 803         fflush(merge_fp);
 804         fclose(merge_fp);
 805 err_param:
 806         free(bcs);
 807
 808         return ret;
 809 }