blktrace.c

   1 /*
   2  * blktrace support code for fio
   3  */
   4 #include <stdio.h>
   5 #include <stdlib.h>
   6 #include <sys/ioctl.h>
   7 #include <unistd.h>
   8 #include <linux/fs.h>
   9
  10 #include "flist.h"
  11 #include "fio.h"
  12 #include "blktrace.h"
  13 #include "blktrace_api.h"
  14 #include "oslib/linux-dev-lookup.h"
  15
  16 #define TRACE_FIFO_SIZE 8192
  17
  18 /*
  19  * fifo refill frontend, to avoid reading data in trace sized bites
  20  */
  21 static int refill_fifo(struct thread_data *td, struct fifo *fifo, int fd)
  22 {
  23         char buf[TRACE_FIFO_SIZE];
  24         unsigned int total;
  25         int ret;
  26
  27         total = sizeof(buf);
  28         if (total > fifo_room(fifo))
  29                 total = fifo_room(fifo);
  30
  31         ret = read(fd, buf, total);
  32         if (ret < 0) {
  33                 td_verror(td, errno, "read blktrace file");
  34                 return -1;
  35         }
  36
  37         if (ret > 0)
  38                 ret = fifo_put(fifo, buf, ret);
  39
  40         dprint(FD_BLKTRACE, "refill: filled %d bytes\n", ret);
  41         return ret;
  42 }
  43
  44 /*
  45  * Retrieve 'len' bytes from the fifo, refilling if necessary.
  46  */
  47 static int trace_fifo_get(struct thread_data *td, struct fifo *fifo, int fd,
  48                           void *buf, unsigned int len)
  49 {
  50         if (fifo_len(fifo) < len) {
  51                 int ret = refill_fifo(td, fifo, fd);
  52
  53                 if (ret < 0)
  54                         return ret;
  55         }
  56
  57         return fifo_get(fifo, buf, len);
  58 }
  59
  60 /*
  61  * Just discard the pdu by seeking past it.
  62  */
  63 static int discard_pdu(struct thread_data *td, struct fifo *fifo, int fd,
  64                        struct blk_io_trace *t)
  65 {
  66         if (t->pdu_len == 0)
  67                 return 0;
  68
  69         dprint(FD_BLKTRACE, "discard pdu len %u\n", t->pdu_len);
  70         return trace_fifo_get(td, fifo, fd, NULL, t->pdu_len);
  71 }
  72
  73 /*
  74  * Check if this is a blktrace binary data file. We read a single trace
  75  * into memory and check for the magic signature.
  76  */
  77 bool is_blktrace(const char *filename, int *need_swap)
  78 {
  79         struct blk_io_trace t;
  80         int fd, ret;
  81
  82         fd = open(filename, O_RDONLY);
  83         if (fd < 0)
  84                 return false;
  85
  86         ret = read(fd, &t, sizeof(t));
  87         close(fd);
  88
  89         if (ret < 0) {
  90                 perror("read blktrace");
  91                 return false;
  92         } else if (ret != sizeof(t)) {
  93                 log_err("fio: short read on blktrace file\n");
  94                 return false;
  95         }
  96
  97         if ((t.magic & 0xffffff00) == BLK_IO_TRACE_MAGIC) {
  98                 *need_swap = 0;
  99                 return true;
 100         }
 101
 102         /*
 103          * Maybe it needs to be endian swapped...
 104          */
 105         t.magic = fio_swap32(t.magic);
 106         if ((t.magic & 0xffffff00) == BLK_IO_TRACE_MAGIC) {
 107                 *need_swap = 1;
 108                 return true;
 109         }
 110
 111         return false;
 112 }
 113
 114 #define FMINORBITS      20
 115 #define FMINORMASK      ((1U << FMINORBITS) - 1)
 116 #define FMAJOR(dev)     ((unsigned int) ((dev) >> FMINORBITS))
 117 #define FMINOR(dev)     ((unsigned int) ((dev) & FMINORMASK))
 118
 119 static void trace_add_open_close_event(struct thread_data *td, int fileno, enum file_log_act action)
 120 {
 121         struct io_piece *ipo;
 122
 123         ipo = calloc(1, sizeof(*ipo));
 124         init_ipo(ipo);
 125
 126         ipo->ddir = DDIR_INVAL;
 127         ipo->fileno = fileno;
 128         ipo->file_action = action;
 129         flist_add_tail(&ipo->list, &td->io_log_list);
 130 }
 131
 132 static int get_dev_blocksize(const char *dev, unsigned int *bs)
 133 {
 134         int fd;
 135
 136         fd = open(dev, O_RDONLY);
 137         if (fd < 0)
 138                 return 1;
 139
 140         if (ioctl(fd, BLKSSZGET, bs) < 0) {
 141                 close(fd);
 142                 return 1;
 143         }
 144
 145         close(fd);
 146         return 0;
 147 }
 148
 149 static int trace_add_file(struct thread_data *td, __u32 device,
 150                           unsigned int *bs)
 151 {
 152         static unsigned int last_maj, last_min, last_fileno, last_bs;
 153         unsigned int maj = FMAJOR(device);
 154         unsigned int min = FMINOR(device);
 155         struct fio_file *f;
 156         unsigned int i;
 157         char dev[256];
 158
 159         if (last_maj == maj && last_min == min) {
 160                 *bs = last_bs;
 161                 return last_fileno;
 162         }
 163
 164         last_maj = maj;
 165         last_min = min;
 166
 167         /*
 168          * check for this file in our list
 169          */
 170         for_each_file(td, f, i) {
 171                 if (f->major == maj && f->minor == min) {
 172                         last_fileno = f->fileno;
 173                         last_bs = f->bs;
 174                         goto out;
 175                 }
 176         }
 177
 178         strcpy(dev, "/dev");
 179         if (blktrace_lookup_device(td->o.replay_redirect, dev, maj, min)) {
 180                 unsigned int this_bs;
 181                 int fileno;
 182
 183                 if (td->o.replay_redirect)
 184                         dprint(FD_BLKTRACE, "device lookup: %d/%d\n overridden"
 185                                         " with: %s\n", maj, min,
 186                                         td->o.replay_redirect);
 187                 else
 188                         dprint(FD_BLKTRACE, "device lookup: %d/%d\n", maj, min);
 189
 190                 dprint(FD_BLKTRACE, "add devices %s\n", dev);
 191                 fileno = add_file_exclusive(td, dev);
 192
 193                 if (get_dev_blocksize(dev, &this_bs))
 194                         this_bs = 512;
 195
 196                 td->o.open_files++;
 197                 td->files[fileno]->major = maj;
 198                 td->files[fileno]->minor = min;
 199                 td->files[fileno]->bs = this_bs;
 200                 trace_add_open_close_event(td, fileno, FIO_LOG_OPEN_FILE);
 201
 202                 last_fileno = fileno;
 203                 last_bs = this_bs;
 204         }
 205
 206 out:
 207         *bs = last_bs;
 208         return last_fileno;
 209 }
 210
 211 static void t_bytes_align(struct thread_options *o, struct blk_io_trace *t)
 212 {
 213         if (!o->replay_align)
 214                 return;
 215
 216         t->bytes = (t->bytes + o->replay_align - 1) & ~(o->replay_align - 1);
 217 }
 218
 219 /*
 220  * Store blk_io_trace data in an ipo for later retrieval.
 221  */
 222 static void store_ipo(struct thread_data *td, unsigned long long offset,
 223                       unsigned int bytes, int rw, unsigned long long ttime,
 224                       int fileno, unsigned int bs)
 225 {
 226         struct io_piece *ipo;
 227
 228         ipo = calloc(1, sizeof(*ipo));
 229         init_ipo(ipo);
 230
 231         ipo->offset = offset * bs;
 232         if (td->o.replay_scale)
 233                 ipo->offset = ipo->offset / td->o.replay_scale;
 234         ipo_bytes_align(td->o.replay_align, ipo);
 235         ipo->len = bytes;
 236         ipo->delay = ttime / 1000;
 237         if (rw)
 238                 ipo->ddir = DDIR_WRITE;
 239         else
 240                 ipo->ddir = DDIR_READ;
 241         ipo->fileno = fileno;
 242
 243         dprint(FD_BLKTRACE, "store ddir=%d, off=%llu, len=%lu, delay=%lu\n",
 244                                                         ipo->ddir, ipo->offset,
 245                                                         ipo->len, ipo->delay);
 246         queue_io_piece(td, ipo);
 247 }
 248
 249 static void handle_trace_notify(struct blk_io_trace *t)
 250 {
 251         switch (t->action) {
 252         case BLK_TN_PROCESS:
 253                 dprint(FD_BLKTRACE, "got process notify: %x, %d\n",
 254                                 t->action, t->pid);
 255                 break;
 256         case BLK_TN_TIMESTAMP:
 257                 dprint(FD_BLKTRACE, "got timestamp notify: %x, %d\n",
 258                                 t->action, t->pid);
 259                 break;
 260         case BLK_TN_MESSAGE:
 261                 break;
 262         default:
 263                 dprint(FD_BLKTRACE, "unknown trace act %x\n", t->action);
 264                 break;
 265         }
 266 }
 267
 268 static void handle_trace_discard(struct thread_data *td,
 269                                  struct blk_io_trace *t,
 270                                  unsigned long long ttime,
 271                                  unsigned long *ios, unsigned int *rw_bs)
 272 {
 273         struct io_piece *ipo;
 274         unsigned int bs;
 275         int fileno;
 276
 277         if (td->o.replay_skip & (1u << DDIR_TRIM))
 278                 return;
 279
 280         ipo = calloc(1, sizeof(*ipo));
 281         init_ipo(ipo);
 282         fileno = trace_add_file(td, t->device, &bs);
 283
 284         ios[DDIR_TRIM]++;
 285         if (t->bytes > rw_bs[DDIR_TRIM])
 286                 rw_bs[DDIR_TRIM] = t->bytes;
 287
 288         td->o.size += t->bytes;
 289
 290         INIT_FLIST_HEAD(&ipo->list);
 291
 292         ipo->offset = t->sector * bs;
 293         if (td->o.replay_scale)
 294                 ipo->offset = ipo->offset / td->o.replay_scale;
 295         ipo_bytes_align(td->o.replay_align, ipo);
 296         ipo->len = t->bytes;
 297         ipo->delay = ttime / 1000;
 298         ipo->ddir = DDIR_TRIM;
 299         ipo->fileno = fileno;
 300
 301         dprint(FD_BLKTRACE, "store discard, off=%llu, len=%lu, delay=%lu\n",
 302                                                         ipo->offset, ipo->len,
 303                                                         ipo->delay);
 304         queue_io_piece(td, ipo);
 305 }
 306
 307 static void dump_trace(struct blk_io_trace *t)
 308 {
 309         log_err("blktrace: ignoring zero byte trace: action=%x\n", t->action);
 310 }
 311
 312 static void handle_trace_fs(struct thread_data *td, struct blk_io_trace *t,
 313                             unsigned long long ttime, unsigned long *ios,
 314                             unsigned int *rw_bs)
 315 {
 316         unsigned int bs;
 317         int rw;
 318         int fileno;
 319
 320         fileno = trace_add_file(td, t->device, &bs);
 321
 322         rw = (t->action & BLK_TC_ACT(BLK_TC_WRITE)) != 0;
 323
 324         if (rw) {
 325                 if (td->o.replay_skip & (1u << DDIR_WRITE))
 326                         return;
 327         } else {
 328                 if (td->o.replay_skip & (1u << DDIR_READ))
 329                         return;
 330         }
 331
 332         if (!t->bytes) {
 333                 if (!fio_did_warn(FIO_WARN_BTRACE_ZERO))
 334                         dump_trace(t);
 335                 return;
 336         }
 337
 338         if (t->bytes > rw_bs[rw])
 339                 rw_bs[rw] = t->bytes;
 340
 341         ios[rw]++;
 342         td->o.size += t->bytes;
 343         store_ipo(td, t->sector, t->bytes, rw, ttime, fileno, bs);
 344 }
 345
 346 static void handle_trace_flush(struct thread_data *td, struct blk_io_trace *t,
 347                                unsigned long long ttime, unsigned long *ios)
 348 {
 349         struct io_piece *ipo;
 350         unsigned int bs;
 351         int fileno;
 352
 353         if (td->o.replay_skip & (1u << DDIR_SYNC))
 354                 return;
 355
 356         ipo = calloc(1, sizeof(*ipo));
 357         init_ipo(ipo);
 358         fileno = trace_add_file(td, t->device, &bs);
 359
 360         ipo->delay = ttime / 1000;
 361         ipo->ddir = DDIR_SYNC;
 362         ipo->fileno = fileno;
 363
 364         ios[DDIR_SYNC]++;
 365         dprint(FD_BLKTRACE, "store flush delay=%lu\n", ipo->delay);
 366         queue_io_piece(td, ipo);
 367 }
 368
 369 /*
 370  * We only care for queue traces, most of the others are side effects
 371  * due to internal workings of the block layer.
 372  */
 373 static void handle_trace(struct thread_data *td, struct blk_io_trace *t,
 374                          unsigned long *ios, unsigned int *bs)
 375 {
 376         static unsigned long long last_ttime;
 377         unsigned long long delay = 0;
 378
 379         if ((t->action & 0xffff) != __BLK_TA_QUEUE)
 380                 return;
 381
 382         if (!(t->action & BLK_TC_ACT(BLK_TC_NOTIFY))) {
 383                 if (!last_ttime || td->o.no_stall)
 384                         delay = 0;
 385                 else if (td->o.replay_time_scale == 100)
 386                         delay = t->time - last_ttime;
 387                 else {
 388                         double tmp = t->time - last_ttime;
 389                         double scale;
 390
 391                         scale = (double) 100.0 / (double) td->o.replay_time_scale;
 392                         tmp *= scale;
 393                         delay = tmp;
 394                 }
 395                 last_ttime = t->time;
 396         }
 397
 398         t_bytes_align(&td->o, t);
 399
 400         if (t->action & BLK_TC_ACT(BLK_TC_NOTIFY))
 401                 handle_trace_notify(t);
 402         else if (t->action & BLK_TC_ACT(BLK_TC_DISCARD))
 403                 handle_trace_discard(td, t, delay, ios, bs);
 404         else if (t->action & BLK_TC_ACT(BLK_TC_FLUSH))
 405                 handle_trace_flush(td, t, delay, ios);
 406         else
 407                 handle_trace_fs(td, t, delay, ios, bs);
 408 }
 409
 410 static void byteswap_trace(struct blk_io_trace *t)
 411 {
 412         t->magic = fio_swap32(t->magic);
 413         t->sequence = fio_swap32(t->sequence);
 414         t->time = fio_swap64(t->time);
 415         t->sector = fio_swap64(t->sector);
 416         t->bytes = fio_swap32(t->bytes);
 417         t->action = fio_swap32(t->action);
 418         t->pid = fio_swap32(t->pid);
 419         t->device = fio_swap32(t->device);
 420         t->cpu = fio_swap32(t->cpu);
 421         t->error = fio_swap16(t->error);
 422         t->pdu_len = fio_swap16(t->pdu_len);
 423 }
 424
 425 static bool t_is_write(struct blk_io_trace *t)
 426 {
 427         return (t->action & BLK_TC_ACT(BLK_TC_WRITE | BLK_TC_DISCARD)) != 0;
 428 }
 429
 430 static enum fio_ddir t_get_ddir(struct blk_io_trace *t)
 431 {
 432         if (t->action & BLK_TC_ACT(BLK_TC_READ))
 433                 return DDIR_READ;
 434         else if (t->action & BLK_TC_ACT(BLK_TC_WRITE))
 435                 return DDIR_WRITE;
 436         else if (t->action & BLK_TC_ACT(BLK_TC_DISCARD))
 437                 return DDIR_TRIM;
 438
 439         return DDIR_INVAL;
 440 }
 441
 442 static void depth_inc(struct blk_io_trace *t, int *depth)
 443 {
 444         enum fio_ddir ddir;
 445
 446         ddir = t_get_ddir(t);
 447         if (ddir != DDIR_INVAL)
 448                 depth[ddir]++;
 449 }
 450
 451 static void depth_dec(struct blk_io_trace *t, int *depth)
 452 {
 453         enum fio_ddir ddir;
 454
 455         ddir = t_get_ddir(t);
 456         if (ddir != DDIR_INVAL)
 457                 depth[ddir]--;
 458 }
 459
 460 static void depth_end(struct blk_io_trace *t, int *this_depth, int *depth)
 461 {
 462         enum fio_ddir ddir = DDIR_INVAL;
 463
 464         ddir = t_get_ddir(t);
 465         if (ddir != DDIR_INVAL) {
 466                 depth[ddir] = max(depth[ddir], this_depth[ddir]);
 467                 this_depth[ddir] = 0;
 468         }
 469 }
 470
 471 /*
 472  * Load a blktrace file by reading all the blk_io_trace entries, and storing
 473  * them as io_pieces like the fio text version would do.
 474  */
 475 bool load_blktrace(struct thread_data *td, const char *filename, int need_swap)
 476 {
 477         struct blk_io_trace t;
 478         unsigned long ios[DDIR_RWDIR_SYNC_CNT] = { };
 479         unsigned int rw_bs[DDIR_RWDIR_CNT] = { };
 480         unsigned long skipped_writes;
 481         struct fifo *fifo;
 482         int fd, i, old_state, max_depth;
 483         struct fio_file *f;
 484         int this_depth[DDIR_RWDIR_CNT] = { };
 485         int depth[DDIR_RWDIR_CNT] = { };
 486
 487         fd = open(filename, O_RDONLY);
 488         if (fd < 0) {
 489                 td_verror(td, errno, "open blktrace file");
 490                 return false;
 491         }
 492
 493         fifo = fifo_alloc(TRACE_FIFO_SIZE);
 494
 495         old_state = td_bump_runstate(td, TD_SETTING_UP);
 496
 497         td->o.size = 0;
 498         skipped_writes = 0;
 499         do {
 500                 int ret = trace_fifo_get(td, fifo, fd, &t, sizeof(t));
 501
 502                 if (ret < 0)
 503                         goto err;
 504                 else if (!ret)
 505                         break;
 506                 else if (ret < (int) sizeof(t)) {
 507                         log_err("fio: short fifo get\n");
 508                         break;
 509                 }
 510
 511                 if (need_swap)
 512                         byteswap_trace(&t);
 513
 514                 if ((t.magic & 0xffffff00) != BLK_IO_TRACE_MAGIC) {
 515                         log_err("fio: bad magic in blktrace data: %x\n",
 516                                                                 t.magic);
 517                         goto err;
 518                 }
 519                 if ((t.magic & 0xff) != BLK_IO_TRACE_VERSION) {
 520                         log_err("fio: bad blktrace version %d\n",
 521                                                                 t.magic & 0xff);
 522                         goto err;
 523                 }
 524                 ret = discard_pdu(td, fifo, fd, &t);
 525                 if (ret < 0) {
 526                         td_verror(td, ret, "blktrace lseek");
 527                         goto err;
 528                 } else if (t.pdu_len != ret) {
 529                         log_err("fio: discarded %d of %d\n", ret, t.pdu_len);
 530                         goto err;
 531                 }
 532                 if ((t.action & BLK_TC_ACT(BLK_TC_NOTIFY)) == 0) {
 533                         if ((t.action & 0xffff) == __BLK_TA_QUEUE)
 534                                 depth_inc(&t, this_depth);
 535                         else if (((t.action & 0xffff) == __BLK_TA_BACKMERGE) ||
 536                                 ((t.action & 0xffff) == __BLK_TA_FRONTMERGE))
 537                                 depth_dec(&t, this_depth);
 538                         else if ((t.action & 0xffff) == __BLK_TA_COMPLETE)
 539                                 depth_end(&t, this_depth, depth);
 540
 541                         if (t_is_write(&t) && read_only) {
 542                                 skipped_writes++;
 543                                 continue;
 544                         }
 545                 }
 546
 547                 handle_trace(td, &t, ios, rw_bs);
 548         } while (1);
 549
 550         for_each_file(td, f, i)
 551                 trace_add_open_close_event(td, f->fileno, FIO_LOG_CLOSE_FILE);
 552
 553         fifo_free(fifo);
 554         close(fd);
 555
 556         td_restore_runstate(td, old_state);
 557
 558         if (!td->files_index) {
 559                 log_err("fio: did not find replay device(s)\n");
 560                 return false;
 561         }
 562
 563         /*
 564          * For stacked devices, we don't always get a COMPLETE event so
 565          * the depth grows to insane values. Limit it to something sane(r).
 566          */
 567         max_depth = 0;
 568         for (i = 0; i < DDIR_RWDIR_CNT; i++) {
 569                 if (depth[i] > 1024)
 570                         depth[i] = 1024;
 571                 else if (!depth[i] && ios[i])
 572                         depth[i] = 1;
 573                 max_depth = max(depth[i], max_depth);
 574         }
 575
 576         if (skipped_writes)
 577                 log_err("fio: %s skips replay of %lu writes due to read-only\n",
 578                                                 td->o.name, skipped_writes);
 579
 580         if (!ios[DDIR_READ] && !ios[DDIR_WRITE] && !ios[DDIR_TRIM] &&
 581             !ios[DDIR_SYNC]) {
 582                 log_err("fio: found no ios in blktrace data\n");
 583                 return false;
 584         } else if (ios[DDIR_READ] && !ios[DDIR_WRITE]) {
 585                 td->o.td_ddir = TD_DDIR_READ;
 586                 td->o.max_bs[DDIR_READ] = rw_bs[DDIR_READ];
 587         } else if (!ios[DDIR_READ] && ios[DDIR_WRITE]) {
 588                 td->o.td_ddir = TD_DDIR_WRITE;
 589                 td->o.max_bs[DDIR_WRITE] = rw_bs[DDIR_WRITE];
 590         } else {
 591                 td->o.td_ddir = TD_DDIR_RW;
 592                 td->o.max_bs[DDIR_READ] = rw_bs[DDIR_READ];
 593                 td->o.max_bs[DDIR_WRITE] = rw_bs[DDIR_WRITE];
 594                 td->o.max_bs[DDIR_TRIM] = rw_bs[DDIR_TRIM];
 595         }
 596
 597         /*
 598          * We need to do direct/raw ios to the device, to avoid getting
 599          * read-ahead in our way. But only do so if the minimum block size
 600          * is a multiple of 4k, otherwise we don't know if it's safe to do so.
 601          */
 602         if (!fio_option_is_set(&td->o, odirect) && !(td_min_bs(td) & 4095))
 603                 td->o.odirect = 1;
 604
 605         /*
 606          * If depth wasn't manually set, use probed depth
 607          */
 608         if (!fio_option_is_set(&td->o, iodepth))
 609                 td->o.iodepth = td->o.iodepth_low = max_depth;
 610
 611         return true;
 612 err:
 613         close(fd);
 614         fifo_free(fifo);
 615         return false;
 616 }
 617
 618 static int init_merge_param_list(fio_fp64_t *vals, struct blktrace_cursor *bcs,
 619                                  int nr_logs, int def, size_t off)
 620 {
 621         int i = 0, len = 0;
 622
 623         while (len < FIO_IO_U_LIST_MAX_LEN && vals[len].u.f != 0.0)
 624                 len++;
 625
 626         if (len && len != nr_logs)
 627                 return len;
 628
 629         for (i = 0; i < nr_logs; i++) {
 630                 int *val = (int *)((char *)&bcs[i] + off);
 631                 *val = def;
 632                 if (len)
 633                         *val = (int)vals[i].u.f;
 634         }
 635
 636         return 0;
 637
 638 }
 639
 640 static int find_earliest_io(struct blktrace_cursor *bcs, int nr_logs)
 641 {
 642         __u64 time = ~(__u64)0;
 643         int idx = 0, i;
 644
 645         for (i = 0; i < nr_logs; i++) {
 646                 if (bcs[i].t.time < time) {
 647                         time = bcs[i].t.time;
 648                         idx = i;
 649                 }
 650         }
 651
 652         return idx;
 653 }
 654
 655 static void merge_finish_file(struct blktrace_cursor *bcs, int i, int *nr_logs)
 656 {
 657         bcs[i].iter++;
 658         if (bcs[i].iter < bcs[i].nr_iter) {
 659                 lseek(bcs[i].fd, 0, SEEK_SET);
 660                 return;
 661         }
 662
 663         *nr_logs -= 1;
 664
 665         /* close file */
 666         fifo_free(bcs[i].fifo);
 667         close(bcs[i].fd);
 668
 669         /* keep active files contiguous */
 670         memmove(&bcs[i], &bcs[*nr_logs], sizeof(bcs[i]));
 671 }
 672
 673 static int read_trace(struct thread_data *td, struct blktrace_cursor *bc)
 674 {
 675         int ret = 0;
 676         struct blk_io_trace *t = &bc->t;
 677
 678 read_skip:
 679         /* read an io trace */
 680         ret = trace_fifo_get(td, bc->fifo, bc->fd, t, sizeof(*t));
 681         if (ret < 0) {
 682                 return ret;
 683         } else if (!ret) {
 684                 if (!bc->length)
 685                         bc->length = bc->t.time;
 686                 return ret;
 687         } else if (ret < (int) sizeof(*t)) {
 688                 log_err("fio: short fifo get\n");
 689                 return -1;
 690         }
 691
 692         if (bc->swap)
 693                 byteswap_trace(t);
 694
 695         /* skip over actions that fio does not care about */
 696         if ((t->action & 0xffff) != __BLK_TA_QUEUE ||
 697             t_get_ddir(t) == DDIR_INVAL) {
 698                 ret = discard_pdu(td, bc->fifo, bc->fd, t);
 699                 if (ret < 0) {
 700                         td_verror(td, ret, "blktrace lseek");
 701                         return ret;
 702                 } else if (t->pdu_len != ret) {
 703                         log_err("fio: discarded %d of %d\n", ret,
 704                                 t->pdu_len);
 705                         return -1;
 706                 }
 707                 goto read_skip;
 708         }
 709
 710         t->time = (t->time + bc->iter * bc->length) * bc->scalar / 100;
 711
 712         return ret;
 713 }
 714
 715 static int write_trace(FILE *fp, struct blk_io_trace *t)
 716 {
 717         /* pdu is not used so just write out only the io trace */
 718         t->pdu_len = 0;
 719         return fwrite((void *)t, sizeof(*t), 1, fp);
 720 }
 721
 722 int merge_blktrace_iologs(struct thread_data *td)
 723 {
 724         int nr_logs = get_max_str_idx(td->o.read_iolog_file);
 725         struct blktrace_cursor *bcs = malloc(sizeof(struct blktrace_cursor) *
 726                                              nr_logs);
 727         struct blktrace_cursor *bc;
 728         FILE *merge_fp;
 729         char *str, *ptr, *name, *merge_buf;
 730         int i, ret;
 731
 732         ret = init_merge_param_list(td->o.merge_blktrace_scalars, bcs, nr_logs,
 733                                     100, offsetof(struct blktrace_cursor,
 734                                                   scalar));
 735         if (ret) {
 736                 log_err("fio: merge_blktrace_scalars(%d) != nr_logs(%d)\n",
 737                         ret, nr_logs);
 738                 goto err_param;
 739         }
 740
 741         ret = init_merge_param_list(td->o.merge_blktrace_iters, bcs, nr_logs,
 742                                     1, offsetof(struct blktrace_cursor,
 743                                                 nr_iter));
 744         if (ret) {
 745                 log_err("fio: merge_blktrace_iters(%d) != nr_logs(%d)\n",
 746                         ret, nr_logs);
 747                 goto err_param;
 748         }
 749
 750         /* setup output file */
 751         merge_fp = fopen(td->o.merge_blktrace_file, "w");
 752         merge_buf = malloc(128 * 1024);
 753         ret = setvbuf(merge_fp, merge_buf, _IOFBF, 128 * 1024);
 754         if (ret)
 755                 goto err_out_file;
 756
 757         /* setup input files */
 758         str = ptr = strdup(td->o.read_iolog_file);
 759         nr_logs = 0;
 760         for (i = 0; (name = get_next_str(&ptr)) != NULL; i++) {
 761                 bcs[i].fd = open(name, O_RDONLY);
 762                 if (bcs[i].fd < 0) {
 763                         log_err("fio: could not open file: %s\n", name);
 764                         ret = bcs[i].fd;
 765                         goto err_file;
 766                 }
 767                 bcs[i].fifo = fifo_alloc(TRACE_FIFO_SIZE);
 768                 nr_logs++;
 769
 770                 if (!is_blktrace(name, &bcs[i].swap)) {
 771                         log_err("fio: file is not a blktrace: %s\n", name);
 772                         goto err_file;
 773                 }
 774
 775                 ret = read_trace(td, &bcs[i]);
 776                 if (ret < 0) {
 777                         goto err_file;
 778                 } else if (!ret) {
 779                         merge_finish_file(bcs, i, &nr_logs);
 780                         i--;
 781                 }
 782         }
 783         free(str);
 784
 785         /* merge files */
 786         while (nr_logs) {
 787                 i = find_earliest_io(bcs, nr_logs);
 788                 bc = &bcs[i];
 789                 /* skip over the pdu */
 790                 ret = discard_pdu(td, bc->fifo, bc->fd, &bc->t);
 791                 if (ret < 0) {
 792                         td_verror(td, ret, "blktrace lseek");
 793                         goto err_file;
 794                 } else if (bc->t.pdu_len != ret) {
 795                         log_err("fio: discarded %d of %d\n", ret,
 796                                 bc->t.pdu_len);
 797                         goto err_file;
 798                 }
 799
 800                 ret = write_trace(merge_fp, &bc->t);
 801                 ret = read_trace(td, bc);
 802                 if (ret < 0)
 803                         goto err_file;
 804                 else if (!ret)
 805                         merge_finish_file(bcs, i, &nr_logs);
 806         }
 807
 808         /* set iolog file to read from the newly merged file */
 809         td->o.read_iolog_file = td->o.merge_blktrace_file;
 810         ret = 0;
 811
 812 err_file:
 813         /* cleanup */
 814         for (i = 0; i < nr_logs; i++) {
 815                 fifo_free(bcs[i].fifo);
 816                 close(bcs[i].fd);
 817         }
 818 err_out_file:
 819         fflush(merge_fp);
 820         fclose(merge_fp);
 821         free(merge_buf);
 822 err_param:
 823         free(bcs);
 824
 825         return ret;
 826 }