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 struct file_cache {
  17         unsigned int maj;
  18         unsigned int min;
  19         unsigned int fileno;
  20 };
  21
  22 /*
  23  * Just discard the pdu by seeking past it.
  24  */
  25 static int discard_pdu(FILE* f, struct blk_io_trace *t)
  26 {
  27         if (t->pdu_len == 0)
  28                 return 0;
  29
  30         dprint(FD_BLKTRACE, "discard pdu len %u\n", t->pdu_len);
  31         if (fseek(f, t->pdu_len, SEEK_CUR) < 0)
  32                 return -errno;
  33
  34         return t->pdu_len;
  35 }
  36
  37 /*
  38  * Check if this is a blktrace binary data file. We read a single trace
  39  * into memory and check for the magic signature.
  40  */
  41 bool is_blktrace(const char *filename, int *need_swap)
  42 {
  43         struct blk_io_trace t;
  44         int fd, ret;
  45
  46         fd = open(filename, O_RDONLY);
  47         if (fd < 0)
  48                 return false;
  49
  50         ret = read(fd, &t, sizeof(t));
  51         close(fd);
  52
  53         if (ret < 0) {
  54                 perror("read blktrace");
  55                 return false;
  56         } else if (ret != sizeof(t)) {
  57                 log_err("fio: short read on blktrace file\n");
  58                 return false;
  59         }
  60
  61         if ((t.magic & 0xffffff00) == BLK_IO_TRACE_MAGIC) {
  62                 *need_swap = 0;
  63                 return true;
  64         }
  65
  66         /*
  67          * Maybe it needs to be endian swapped...
  68          */
  69         t.magic = fio_swap32(t.magic);
  70         if ((t.magic & 0xffffff00) == BLK_IO_TRACE_MAGIC) {
  71                 *need_swap = 1;
  72                 return true;
  73         }
  74
  75         return false;
  76 }
  77
  78 #define FMINORBITS      20
  79 #define FMINORMASK      ((1U << FMINORBITS) - 1)
  80 #define FMAJOR(dev)     ((unsigned int) ((dev) >> FMINORBITS))
  81 #define FMINOR(dev)     ((unsigned int) ((dev) & FMINORMASK))
  82
  83 static void trace_add_open_close_event(struct thread_data *td, int fileno, enum file_log_act action)
  84 {
  85         struct io_piece *ipo;
  86
  87         ipo = calloc(1, sizeof(*ipo));
  88         init_ipo(ipo);
  89
  90         ipo->ddir = DDIR_INVAL;
  91         ipo->fileno = fileno;
  92         ipo->file_action = action;
  93         flist_add_tail(&ipo->list, &td->io_log_list);
  94 }
  95
  96 static int trace_add_file(struct thread_data *td, __u32 device,
  97                           struct file_cache *cache)
  98 {
  99         unsigned int maj = FMAJOR(device);
 100         unsigned int min = FMINOR(device);
 101         struct fio_file *f;
 102         char dev[256];
 103         unsigned int i;
 104
 105         if (cache->maj == maj && cache->min == min)
 106                 return cache->fileno;
 107
 108         cache->maj = maj;
 109         cache->min = min;
 110
 111         /*
 112          * check for this file in our list
 113          */
 114         for_each_file(td, f, i)
 115                 if (f->major == maj && f->minor == min) {
 116                         cache->fileno = f->fileno;
 117                         return cache->fileno;
 118                 }
 119
 120         strcpy(dev, "/dev");
 121         if (blktrace_lookup_device(td->o.replay_redirect, dev, maj, min)) {
 122                 int fileno;
 123
 124                 if (td->o.replay_redirect)
 125                         dprint(FD_BLKTRACE, "device lookup: %d/%d\n overridden"
 126                                         " with: %s\n", maj, min,
 127                                         td->o.replay_redirect);
 128                 else
 129                         dprint(FD_BLKTRACE, "device lookup: %d/%d\n", maj, min);
 130
 131                 dprint(FD_BLKTRACE, "add devices %s\n", dev);
 132                 fileno = add_file_exclusive(td, dev);
 133                 td->o.open_files++;
 134                 td->files[fileno]->major = maj;
 135                 td->files[fileno]->minor = min;
 136                 trace_add_open_close_event(td, fileno, FIO_LOG_OPEN_FILE);
 137                 cache->fileno = fileno;
 138         }
 139
 140         return cache->fileno;
 141 }
 142
 143 static void t_bytes_align(struct thread_options *o, struct blk_io_trace *t)
 144 {
 145         if (!o->replay_align)
 146                 return;
 147
 148         t->bytes = (t->bytes + o->replay_align - 1) & ~(o->replay_align - 1);
 149 }
 150
 151 /*
 152  * Store blk_io_trace data in an ipo for later retrieval.
 153  */
 154 static void store_ipo(struct thread_data *td, unsigned long long offset,
 155                       unsigned int bytes, int rw, unsigned long long ttime,
 156                       int fileno)
 157 {
 158         struct io_piece *ipo;
 159
 160         ipo = calloc(1, sizeof(*ipo));
 161         init_ipo(ipo);
 162
 163         ipo->offset = offset * 512;
 164         if (td->o.replay_scale)
 165                 ipo->offset = ipo->offset / td->o.replay_scale;
 166         ipo_bytes_align(td->o.replay_align, ipo);
 167         ipo->len = bytes;
 168         ipo->delay = ttime / 1000;
 169         if (rw)
 170                 ipo->ddir = DDIR_WRITE;
 171         else
 172                 ipo->ddir = DDIR_READ;
 173         ipo->fileno = fileno;
 174
 175         dprint(FD_BLKTRACE, "store ddir=%d, off=%llu, len=%lu, delay=%lu\n",
 176                                                         ipo->ddir, ipo->offset,
 177                                                         ipo->len, ipo->delay);
 178         queue_io_piece(td, ipo);
 179 }
 180
 181 static bool handle_trace_notify(struct blk_io_trace *t)
 182 {
 183         switch (t->action) {
 184         case BLK_TN_PROCESS:
 185                 dprint(FD_BLKTRACE, "got process notify: %x, %d\n",
 186                                 t->action, t->pid);
 187                 break;
 188         case BLK_TN_TIMESTAMP:
 189                 dprint(FD_BLKTRACE, "got timestamp notify: %x, %d\n",
 190                                 t->action, t->pid);
 191                 break;
 192         case BLK_TN_MESSAGE:
 193                 break;
 194         default:
 195                 dprint(FD_BLKTRACE, "unknown trace act %x\n", t->action);
 196                 break;
 197         }
 198         return false;
 199 }
 200
 201 static bool handle_trace_discard(struct thread_data *td,
 202                                  struct blk_io_trace *t,
 203                                  unsigned long long ttime,
 204                                  unsigned long *ios, unsigned long long *bs,
 205                                  struct file_cache *cache)
 206 {
 207         struct io_piece *ipo;
 208         int fileno;
 209
 210         if (td->o.replay_skip & (1u << DDIR_TRIM))
 211                 return false;
 212
 213         ipo = calloc(1, sizeof(*ipo));
 214         init_ipo(ipo);
 215         fileno = trace_add_file(td, t->device, cache);
 216
 217         ios[DDIR_TRIM]++;
 218         if (t->bytes > bs[DDIR_TRIM])
 219                 bs[DDIR_TRIM] = t->bytes;
 220
 221         td->o.size += t->bytes;
 222
 223         INIT_FLIST_HEAD(&ipo->list);
 224
 225         ipo->offset = t->sector * 512;
 226         if (td->o.replay_scale)
 227                 ipo->offset = ipo->offset / td->o.replay_scale;
 228         ipo_bytes_align(td->o.replay_align, ipo);
 229         ipo->len = t->bytes;
 230         ipo->delay = ttime / 1000;
 231         ipo->ddir = DDIR_TRIM;
 232         ipo->fileno = fileno;
 233
 234         dprint(FD_BLKTRACE, "store discard, off=%llu, len=%lu, delay=%lu\n",
 235                                                         ipo->offset, ipo->len,
 236                                                         ipo->delay);
 237         queue_io_piece(td, ipo);
 238         return true;
 239 }
 240
 241 static void dump_trace(struct blk_io_trace *t)
 242 {
 243         log_err("blktrace: ignoring zero byte trace: action=%x\n", t->action);
 244 }
 245
 246 static bool handle_trace_fs(struct thread_data *td, struct blk_io_trace *t,
 247                             unsigned long long ttime, unsigned long *ios,
 248                             unsigned long long *bs, struct file_cache *cache)
 249 {
 250         int rw;
 251         int fileno;
 252
 253         fileno = trace_add_file(td, t->device, cache);
 254
 255         rw = (t->action & BLK_TC_ACT(BLK_TC_WRITE)) != 0;
 256
 257         if (rw) {
 258                 if (td->o.replay_skip & (1u << DDIR_WRITE))
 259                         return false;
 260         } else {
 261                 if (td->o.replay_skip & (1u << DDIR_READ))
 262                         return false;
 263         }
 264
 265         if (!t->bytes) {
 266                 if (!fio_did_warn(FIO_WARN_BTRACE_ZERO))
 267                         dump_trace(t);
 268                 return false;
 269         }
 270
 271         if (t->bytes > bs[rw])
 272                 bs[rw] = t->bytes;
 273
 274         ios[rw]++;
 275         td->o.size += t->bytes;
 276         store_ipo(td, t->sector, t->bytes, rw, ttime, fileno);
 277         return true;
 278 }
 279
 280 static bool handle_trace_flush(struct thread_data *td, struct blk_io_trace *t,
 281                                unsigned long long ttime, unsigned long *ios,
 282                                struct file_cache *cache)
 283 {
 284         struct io_piece *ipo;
 285         int fileno;
 286
 287         if (td->o.replay_skip & (1u << DDIR_SYNC))
 288                 return false;
 289
 290         ipo = calloc(1, sizeof(*ipo));
 291         init_ipo(ipo);
 292         fileno = trace_add_file(td, t->device, cache);
 293
 294         ipo->delay = ttime / 1000;
 295         ipo->ddir = DDIR_SYNC;
 296         ipo->fileno = fileno;
 297
 298         ios[DDIR_SYNC]++;
 299         dprint(FD_BLKTRACE, "store flush delay=%lu\n", ipo->delay);
 300         queue_io_piece(td, ipo);
 301         return true;
 302 }
 303
 304 /*
 305  * We only care for queue traces, most of the others are side effects
 306  * due to internal workings of the block layer.
 307  */
 308 static bool queue_trace(struct thread_data *td, struct blk_io_trace *t,
 309                          unsigned long *ios, unsigned long long *bs,
 310                          struct file_cache *cache)
 311 {
 312         unsigned long long *last_ttime = &td->io_log_blktrace_last_ttime;
 313         unsigned long long delay = 0;
 314
 315         if ((t->action & 0xffff) != __BLK_TA_QUEUE)
 316                 return false;
 317
 318         if (!(t->action & BLK_TC_ACT(BLK_TC_NOTIFY))) {
 319                 if (!*last_ttime || td->o.no_stall || t->time < *last_ttime)
 320                         delay = 0;
 321                 else if (td->o.replay_time_scale == 100)
 322                         delay = t->time - *last_ttime;
 323                 else {
 324                         double tmp = t->time - *last_ttime;
 325                         double scale;
 326
 327                         scale = (double) 100.0 / (double) td->o.replay_time_scale;
 328                         tmp *= scale;
 329                         delay = tmp;
 330                 }
 331                 *last_ttime = t->time;
 332         }
 333
 334         t_bytes_align(&td->o, t);
 335
 336         if (t->action & BLK_TC_ACT(BLK_TC_NOTIFY))
 337                 return handle_trace_notify(t);
 338         else if (t->action & BLK_TC_ACT(BLK_TC_DISCARD))
 339                 return handle_trace_discard(td, t, delay, ios, bs, cache);
 340         else if (t->action & BLK_TC_ACT(BLK_TC_FLUSH))
 341                 return handle_trace_flush(td, t, delay, ios, cache);
 342         else
 343                 return handle_trace_fs(td, t, delay, ios, bs, cache);
 344 }
 345
 346 static void byteswap_trace(struct blk_io_trace *t)
 347 {
 348         t->magic = fio_swap32(t->magic);
 349         t->sequence = fio_swap32(t->sequence);
 350         t->time = fio_swap64(t->time);
 351         t->sector = fio_swap64(t->sector);
 352         t->bytes = fio_swap32(t->bytes);
 353         t->action = fio_swap32(t->action);
 354         t->pid = fio_swap32(t->pid);
 355         t->device = fio_swap32(t->device);
 356         t->cpu = fio_swap32(t->cpu);
 357         t->error = fio_swap16(t->error);
 358         t->pdu_len = fio_swap16(t->pdu_len);
 359 }
 360
 361 static bool t_is_write(struct blk_io_trace *t)
 362 {
 363         return (t->action & BLK_TC_ACT(BLK_TC_WRITE | BLK_TC_DISCARD)) != 0;
 364 }
 365
 366 static enum fio_ddir t_get_ddir(struct blk_io_trace *t)
 367 {
 368         if (t->action & BLK_TC_ACT(BLK_TC_READ))
 369                 return DDIR_READ;
 370         else if (t->action & BLK_TC_ACT(BLK_TC_WRITE))
 371                 return DDIR_WRITE;
 372         else if (t->action & BLK_TC_ACT(BLK_TC_DISCARD))
 373                 return DDIR_TRIM;
 374
 375         return DDIR_INVAL;
 376 }
 377
 378 static void depth_inc(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_dec(struct blk_io_trace *t, int *depth)
 388 {
 389         enum fio_ddir ddir;
 390
 391         ddir = t_get_ddir(t);
 392         if (ddir != DDIR_INVAL)
 393                 depth[ddir]--;
 394 }
 395
 396 static void depth_end(struct blk_io_trace *t, int *this_depth, int *depth)
 397 {
 398         enum fio_ddir ddir = DDIR_INVAL;
 399
 400         ddir = t_get_ddir(t);
 401         if (ddir != DDIR_INVAL) {
 402                 depth[ddir] = max(depth[ddir], this_depth[ddir]);
 403                 this_depth[ddir] = 0;
 404         }
 405 }
 406
 407 /*
 408  * Load a blktrace file by reading all the blk_io_trace entries, and storing
 409  * them as io_pieces like the fio text version would do.
 410  */
 411 bool init_blktrace_read(struct thread_data *td, const char *filename, int need_swap)
 412 {
 413         int old_state;
 414
 415         td->io_log_rfile = fopen(filename, "rb");
 416         if (!td->io_log_rfile) {
 417                 td_verror(td, errno, "open blktrace file");
 418                 goto err;
 419         }
 420         td->io_log_blktrace_swap = need_swap;
 421         td->io_log_blktrace_last_ttime = 0;
 422         td->o.size = 0;
 423
 424         free_release_files(td);
 425
 426         old_state = td_bump_runstate(td, TD_SETTING_UP);
 427
 428         if (!read_blktrace(td)) {
 429                 goto err;
 430         }
 431
 432         td_restore_runstate(td, old_state);
 433
 434         if (!td->files_index) {
 435                 log_err("fio: did not find replay device(s)\n");
 436                 return false;
 437         }
 438
 439         return true;
 440
 441 err:
 442         if (td->io_log_rfile) {
 443                 fclose(td->io_log_rfile);
 444                 td->io_log_rfile = NULL;
 445         }
 446         return false;
 447 }
 448
 449 bool read_blktrace(struct thread_data* td)
 450 {
 451         struct blk_io_trace t;
 452         struct file_cache cache = { };
 453         unsigned long ios[DDIR_RWDIR_SYNC_CNT] = { };
 454         unsigned long long rw_bs[DDIR_RWDIR_CNT] = { };
 455         unsigned long skipped_writes;
 456         FILE *f = td->io_log_rfile;
 457         int i, max_depth;
 458         struct fio_file *fiof;
 459         int this_depth[DDIR_RWDIR_CNT] = { };
 460         int depth[DDIR_RWDIR_CNT] = { };
 461         int64_t items_to_fetch = 0;
 462
 463         if (td->o.read_iolog_chunked) {
 464                 items_to_fetch = iolog_items_to_fetch(td);
 465                 if (!items_to_fetch)
 466                         return true;
 467         }
 468
 469         skipped_writes = 0;
 470         do {
 471                 int ret = fread(&t, 1, sizeof(t), f);
 472
 473                 if (ferror(f)) {
 474                         td_verror(td, errno, "read blktrace file");
 475                         goto err;
 476                 } else if (feof(f)) {
 477                         break;
 478                 } else if (ret < (int) sizeof(t)) {
 479                         log_err("fio: iolog short read\n");
 480                         break;
 481                 }
 482
 483                 if (td->io_log_blktrace_swap)
 484                         byteswap_trace(&t);
 485
 486                 if ((t.magic & 0xffffff00) != BLK_IO_TRACE_MAGIC) {
 487                         log_err("fio: bad magic in blktrace data: %x\n",
 488                                                                 t.magic);
 489                         goto err;
 490                 }
 491                 if ((t.magic & 0xff) != BLK_IO_TRACE_VERSION) {
 492                         log_err("fio: bad blktrace version %d\n",
 493                                                                 t.magic & 0xff);
 494                         goto err;
 495                 }
 496                 ret = discard_pdu(f, &t);
 497                 if (ret < 0) {
 498                         td_verror(td, -ret, "blktrace lseek");
 499                         goto err;
 500                 }
 501                 if ((t.action & BLK_TC_ACT(BLK_TC_NOTIFY)) == 0) {
 502                         if ((t.action & 0xffff) == __BLK_TA_QUEUE)
 503                                 depth_inc(&t, this_depth);
 504                         else if (((t.action & 0xffff) == __BLK_TA_BACKMERGE) ||
 505                                 ((t.action & 0xffff) == __BLK_TA_FRONTMERGE))
 506                                 depth_dec(&t, this_depth);
 507                         else if ((t.action & 0xffff) == __BLK_TA_COMPLETE)
 508                                 depth_end(&t, this_depth, depth);
 509
 510                         if (t_is_write(&t) && read_only) {
 511                                 skipped_writes++;
 512                                 continue;
 513                         }
 514                 }
 515
 516                 if (!queue_trace(td, &t, ios, rw_bs, &cache))
 517                         continue;
 518
 519                 if (td->o.read_iolog_chunked) {
 520                         td->io_log_current++;
 521                         items_to_fetch--;
 522                         if (items_to_fetch == 0)
 523                                 break;
 524                 }
 525         } while (1);
 526
 527         if (td->o.read_iolog_chunked) {
 528                 td->io_log_highmark = td->io_log_current;
 529                 td->io_log_checkmark = (td->io_log_highmark + 1) / 2;
 530                 fio_gettime(&td->io_log_highmark_time, NULL);
 531         }
 532
 533         if (skipped_writes)
 534                 log_err("fio: %s skips replay of %lu writes due to read-only\n",
 535                                                 td->o.name, skipped_writes);
 536
 537         if (td->o.read_iolog_chunked) {
 538                 if (td->io_log_current == 0) {
 539                         return false;
 540                 }
 541                 td->o.td_ddir = TD_DDIR_RW;
 542                 if ((rw_bs[DDIR_READ] > td->o.max_bs[DDIR_READ] ||
 543                      rw_bs[DDIR_WRITE] > td->o.max_bs[DDIR_WRITE] ||
 544                      rw_bs[DDIR_TRIM] > td->o.max_bs[DDIR_TRIM]) &&
 545                     td->orig_buffer)
 546                 {
 547                         td->o.max_bs[DDIR_READ] = max(td->o.max_bs[DDIR_READ], rw_bs[DDIR_READ]);
 548                         td->o.max_bs[DDIR_WRITE] = max(td->o.max_bs[DDIR_WRITE], rw_bs[DDIR_WRITE]);
 549                         td->o.max_bs[DDIR_TRIM] = max(td->o.max_bs[DDIR_TRIM], rw_bs[DDIR_TRIM]);
 550                         io_u_quiesce(td);
 551                         free_io_mem(td);
 552                         init_io_u_buffers(td);
 553                 }
 554                 return true;
 555         }
 556
 557         for_each_file(td, fiof, i)
 558                 trace_add_open_close_event(td, fiof->fileno, FIO_LOG_CLOSE_FILE);
 559
 560         fclose(td->io_log_rfile);
 561         td->io_log_rfile = NULL;
 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 (!ios[DDIR_READ] && !ios[DDIR_WRITE] && !ios[DDIR_TRIM] &&
 577             !ios[DDIR_SYNC]) {
 578                 log_err("fio: found no ios in blktrace data\n");
 579                 return false;
 580         }
 581
 582         td->o.td_ddir = 0;
 583         if (ios[DDIR_READ]) {
 584                 td->o.td_ddir |= TD_DDIR_READ;
 585                 td->o.max_bs[DDIR_READ] = rw_bs[DDIR_READ];
 586         }
 587         if (ios[DDIR_WRITE]) {
 588                 td->o.td_ddir |= TD_DDIR_WRITE;
 589                 td->o.max_bs[DDIR_WRITE] = rw_bs[DDIR_WRITE];
 590         }
 591         if (ios[DDIR_TRIM]) {
 592                 td->o.td_ddir |= TD_DDIR_TRIM;
 593                 td->o.max_bs[DDIR_TRIM] = rw_bs[DDIR_TRIM];
 594         }
 595
 596         /*
 597          * If depth wasn't manually set, use probed depth
 598          */
 599         if (!fio_option_is_set(&td->o, iodepth))
 600                 td->o.iodepth = td->o.iodepth_low = max_depth;
 601
 602         return true;
 603 err:
 604         fclose(f);
 605         return false;
 606 }
 607
 608 static int init_merge_param_list(fio_fp64_t *vals, struct blktrace_cursor *bcs,
 609                                  int nr_logs, int def, size_t off)
 610 {
 611         int i = 0, len = 0;
 612
 613         while (len < FIO_IO_U_LIST_MAX_LEN && vals[len].u.f != 0.0)
 614                 len++;
 615
 616         if (len && len != nr_logs)
 617                 return len;
 618
 619         for (i = 0; i < nr_logs; i++) {
 620                 int *val = (int *)((char *)&bcs[i] + off);
 621                 *val = def;
 622                 if (len)
 623                         *val = (int)vals[i].u.f;
 624         }
 625
 626         return 0;
 627
 628 }
 629
 630 static int find_earliest_io(struct blktrace_cursor *bcs, int nr_logs)
 631 {
 632         __u64 time = ~(__u64)0;
 633         int idx = 0, i;
 634
 635         for (i = 0; i < nr_logs; i++) {
 636                 if (bcs[i].t.time < time) {
 637                         time = bcs[i].t.time;
 638                         idx = i;
 639                 }
 640         }
 641
 642         return idx;
 643 }
 644
 645 static void merge_finish_file(struct blktrace_cursor *bcs, int i, int *nr_logs)
 646 {
 647         bcs[i].iter++;
 648         if (bcs[i].iter < bcs[i].nr_iter) {
 649                 fseek(bcs[i].f, 0, SEEK_SET);
 650                 return;
 651         }
 652
 653         *nr_logs -= 1;
 654
 655         /* close file */
 656         fclose(bcs[i].f);
 657
 658         /* keep active files contiguous */
 659         memmove(&bcs[i], &bcs[*nr_logs], sizeof(bcs[i]));
 660 }
 661
 662 static int read_trace(struct thread_data *td, struct blktrace_cursor *bc)
 663 {
 664         int ret = 0;
 665         struct blk_io_trace *t = &bc->t;
 666
 667 read_skip:
 668         /* read an io trace */
 669         ret = fread(&t, 1, sizeof(t), bc->f);
 670         if (ferror(bc->f)) {
 671                 td_verror(td, errno, "read blktrace file");
 672                 return ret;
 673         } else if (feof(bc->f)) {
 674                 if (!bc->length)
 675                         bc->length = bc->t.time;
 676                 return ret;
 677         } else if (ret < (int) sizeof(*t)) {
 678                 log_err("fio: iolog short read\n");
 679                 return -1;
 680         }
 681
 682         if (bc->swap)
 683                 byteswap_trace(t);
 684
 685         /* skip over actions that fio does not care about */
 686         if ((t->action & 0xffff) != __BLK_TA_QUEUE ||
 687             t_get_ddir(t) == DDIR_INVAL) {
 688                 ret = discard_pdu(bc->f, t);
 689                 if (ret < 0) {
 690                         td_verror(td, -ret, "blktrace lseek");
 691                         return ret;
 692                 }
 693                 goto read_skip;
 694         }
 695
 696         t->time = (t->time + bc->iter * bc->length) * bc->scalar / 100;
 697
 698         return ret;
 699 }
 700
 701 static int write_trace(FILE *fp, struct blk_io_trace *t)
 702 {
 703         /* pdu is not used so just write out only the io trace */
 704         t->pdu_len = 0;
 705         return fwrite((void *)t, sizeof(*t), 1, fp);
 706 }
 707
 708 int merge_blktrace_iologs(struct thread_data *td)
 709 {
 710         int nr_logs = get_max_str_idx(td->o.read_iolog_file);
 711         struct blktrace_cursor *bcs = malloc(sizeof(struct blktrace_cursor) *
 712                                              nr_logs);
 713         struct blktrace_cursor *bc;
 714         FILE *merge_fp;
 715         char *str, *ptr, *name, *merge_buf;
 716         int i, ret;
 717
 718         ret = init_merge_param_list(td->o.merge_blktrace_scalars, bcs, nr_logs,
 719                                     100, offsetof(struct blktrace_cursor,
 720                                                   scalar));
 721         if (ret) {
 722                 log_err("fio: merge_blktrace_scalars(%d) != nr_logs(%d)\n",
 723                         ret, nr_logs);
 724                 goto err_param;
 725         }
 726
 727         ret = init_merge_param_list(td->o.merge_blktrace_iters, bcs, nr_logs,
 728                                     1, offsetof(struct blktrace_cursor,
 729                                                 nr_iter));
 730         if (ret) {
 731                 log_err("fio: merge_blktrace_iters(%d) != nr_logs(%d)\n",
 732                         ret, nr_logs);
 733                 goto err_param;
 734         }
 735
 736         /* setup output file */
 737         merge_fp = fopen(td->o.merge_blktrace_file, "w");
 738         merge_buf = malloc(128 * 1024);
 739         if (!merge_buf)
 740                 goto err_out_file;
 741         ret = setvbuf(merge_fp, merge_buf, _IOFBF, 128 * 1024);
 742         if (ret)
 743                 goto err_merge_buf;
 744
 745         /* setup input files */
 746         str = ptr = strdup(td->o.read_iolog_file);
 747         nr_logs = 0;
 748         for (i = 0; (name = get_next_str(&ptr)) != NULL; i++) {
 749                 bcs[i].f = fopen(name, "rb");
 750                 if (!bcs[i].f) {
 751                         log_err("fio: could not open file: %s\n", name);
 752                         ret = -errno;
 753                         free(str);
 754                         goto err_file;
 755                 }
 756                 nr_logs++;
 757
 758                 if (!is_blktrace(name, &bcs[i].swap)) {
 759                         log_err("fio: file is not a blktrace: %s\n", name);
 760                         free(str);
 761                         goto err_file;
 762                 }
 763
 764                 ret = read_trace(td, &bcs[i]);
 765                 if (ret < 0) {
 766                         free(str);
 767                         goto err_file;
 768                 } else if (!ret) {
 769                         merge_finish_file(bcs, i, &nr_logs);
 770                         i--;
 771                 }
 772         }
 773         free(str);
 774
 775         /* merge files */
 776         while (nr_logs) {
 777                 i = find_earliest_io(bcs, nr_logs);
 778                 bc = &bcs[i];
 779                 /* skip over the pdu */
 780                 ret = discard_pdu(bc->f, &bc->t);
 781                 if (ret < 0) {
 782                         td_verror(td, -ret, "blktrace lseek");
 783                         goto err_file;
 784                 }
 785
 786                 ret = write_trace(merge_fp, &bc->t);
 787                 ret = read_trace(td, bc);
 788                 if (ret < 0)
 789                         goto err_file;
 790                 else if (!ret)
 791                         merge_finish_file(bcs, i, &nr_logs);
 792         }
 793
 794         /* set iolog file to read from the newly merged file */
 795         td->o.read_iolog_file = td->o.merge_blktrace_file;
 796         ret = 0;
 797
 798 err_file:
 799         /* cleanup */
 800         for (i = 0; i < nr_logs; i++) {
 801                 fclose(bcs[i].f);
 802         }
 803 err_merge_buf:
 804         free(merge_buf);
 805 err_out_file:
 806         fflush(merge_fp);
 807         fclose(merge_fp);
 808 err_param:
 809         free(bcs);
 810
 811         return ret;
 812 }