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 <linux/fs.h>
   8
   9 #include "flist.h"
  10 #include "fio.h"
  11 #include "blktrace.h"
  12 #include "blktrace_api.h"
  13 #include "oslib/linux-dev-lookup.h"
  14
  15 #define TRACE_FIFO_SIZE 8192
  16
  17 /*
  18  * fifo refill frontend, to avoid reading data in trace sized bites
  19  */
  20 static int refill_fifo(struct thread_data *td, struct fifo *fifo, int fd)
  21 {
  22         char buf[TRACE_FIFO_SIZE];
  23         unsigned int total;
  24         int ret;
  25
  26         total = sizeof(buf);
  27         if (total > fifo_room(fifo))
  28                 total = fifo_room(fifo);
  29
  30         ret = read(fd, buf, total);
  31         if (ret < 0) {
  32                 td_verror(td, errno, "read blktrace file");
  33                 return -1;
  34         }
  35
  36         if (ret > 0)
  37                 ret = fifo_put(fifo, buf, ret);
  38
  39         dprint(FD_BLKTRACE, "refill: filled %d bytes\n", ret);
  40         return ret;
  41 }
  42
  43 /*
  44  * Retrieve 'len' bytes from the fifo, refilling if necessary.
  45  */
  46 static int trace_fifo_get(struct thread_data *td, struct fifo *fifo, int fd,
  47                           void *buf, unsigned int len)
  48 {
  49         if (fifo_len(fifo) < len) {
  50                 int ret = refill_fifo(td, fifo, fd);
  51
  52                 if (ret < 0)
  53                         return ret;
  54         }
  55
  56         return fifo_get(fifo, buf, len);
  57 }
  58
  59 /*
  60  * Just discard the pdu by seeking past it.
  61  */
  62 static int discard_pdu(struct thread_data *td, struct fifo *fifo, int fd,
  63                        struct blk_io_trace *t)
  64 {
  65         if (t->pdu_len == 0)
  66                 return 0;
  67
  68         dprint(FD_BLKTRACE, "discard pdu len %u\n", t->pdu_len);
  69         return trace_fifo_get(td, fifo, fd, NULL, t->pdu_len);
  70 }
  71
  72 /*
  73  * Check if this is a blktrace binary data file. We read a single trace
  74  * into memory and check for the magic signature.
  75  */
  76 int is_blktrace(const char *filename, int *need_swap)
  77 {
  78         struct blk_io_trace t;
  79         int fd, ret;
  80
  81         fd = open(filename, O_RDONLY);
  82         if (fd < 0)
  83                 return 0;
  84
  85         ret = read(fd, &t, sizeof(t));
  86         close(fd);
  87
  88         if (ret < 0) {
  89                 perror("read blktrace");
  90                 return 0;
  91         } else if (ret != sizeof(t)) {
  92                 log_err("fio: short read on blktrace file\n");
  93                 return 0;
  94         }
  95
  96         if ((t.magic & 0xffffff00) == BLK_IO_TRACE_MAGIC) {
  97                 *need_swap = 0;
  98                 return 1;
  99         }
 100
 101         /*
 102          * Maybe it needs to be endian swapped...
 103          */
 104         t.magic = fio_swap32(t.magic);
 105         if ((t.magic & 0xffffff00) == BLK_IO_TRACE_MAGIC) {
 106                 *need_swap = 1;
 107                 return 1;
 108         }
 109
 110         return 0;
 111 }
 112
 113 #define FMINORBITS      20
 114 #define FMINORMASK      ((1U << FMINORBITS) - 1)
 115 #define FMAJOR(dev)     ((unsigned int) ((dev) >> FMINORBITS))
 116 #define FMINOR(dev)     ((unsigned int) ((dev) & FMINORMASK))
 117
 118 static void trace_add_open_close_event(struct thread_data *td, int fileno, enum file_log_act action)
 119 {
 120         struct io_piece *ipo;
 121
 122         ipo = calloc(1, sizeof(*ipo));
 123         init_ipo(ipo);
 124
 125         ipo->ddir = DDIR_INVAL;
 126         ipo->fileno = fileno;
 127         ipo->file_action = action;
 128         flist_add_tail(&ipo->list, &td->io_log_list);
 129 }
 130
 131 static int get_dev_blocksize(const char *dev, unsigned int *bs)
 132 {
 133         int fd;
 134
 135         fd = open(dev, O_RDONLY);
 136         if (fd < 0)
 137                 return 1;
 138
 139         if (ioctl(fd, BLKSSZGET, bs) < 0) {
 140                 close(fd);
 141                 return 1;
 142         }
 143
 144         close(fd);
 145         return 0;
 146 }
 147
 148 static int trace_add_file(struct thread_data *td, __u32 device,
 149                           unsigned int *bs)
 150 {
 151         static unsigned int last_maj, last_min, last_fileno, last_bs;
 152         unsigned int maj = FMAJOR(device);
 153         unsigned int min = FMINOR(device);
 154         struct fio_file *f;
 155         unsigned int i;
 156         char dev[256];
 157
 158         if (last_maj == maj && last_min == min) {
 159                 *bs = last_bs;
 160                 return last_fileno;
 161         }
 162
 163         last_maj = maj;
 164         last_min = min;
 165
 166         /*
 167          * check for this file in our list
 168          */
 169         for_each_file(td, f, i) {
 170                 if (f->major == maj && f->minor == min) {
 171                         last_fileno = f->fileno;
 172                         last_bs = f->bs;
 173                         goto out;
 174                 }
 175         }
 176
 177         strcpy(dev, "/dev");
 178         if (blktrace_lookup_device(td->o.replay_redirect, dev, maj, min)) {
 179                 unsigned int this_bs;
 180                 int fileno;
 181
 182                 if (td->o.replay_redirect)
 183                         dprint(FD_BLKTRACE, "device lookup: %d/%d\n overridden"
 184                                         " with: %s\n", maj, min,
 185                                         td->o.replay_redirect);
 186                 else
 187                         dprint(FD_BLKTRACE, "device lookup: %d/%d\n", maj, min);
 188
 189                 dprint(FD_BLKTRACE, "add devices %s\n", dev);
 190                 fileno = add_file_exclusive(td, dev);
 191
 192                 if (get_dev_blocksize(dev, &this_bs))
 193                         this_bs = 512;
 194
 195                 td->o.open_files++;
 196                 td->files[fileno]->major = maj;
 197                 td->files[fileno]->minor = min;
 198                 td->files[fileno]->bs = this_bs;
 199                 trace_add_open_close_event(td, fileno, FIO_LOG_OPEN_FILE);
 200
 201                 last_fileno = fileno;
 202                 last_bs = this_bs;
 203         }
 204
 205 out:
 206         *bs = last_bs;
 207         return last_fileno;
 208 }
 209
 210 static void t_bytes_align(struct thread_options *o, struct blk_io_trace *t)
 211 {
 212         if (!o->replay_align)
 213                 return;
 214
 215         t->bytes = (t->bytes + o->replay_align - 1) & ~(o->replay_align - 1);
 216 }
 217
 218 /*
 219  * Store blk_io_trace data in an ipo for later retrieval.
 220  */
 221 static void store_ipo(struct thread_data *td, unsigned long long offset,
 222                       unsigned int bytes, int rw, unsigned long long ttime,
 223                       int fileno, unsigned int bs)
 224 {
 225         struct io_piece *ipo;
 226
 227         ipo = calloc(1, sizeof(*ipo));
 228         init_ipo(ipo);
 229
 230         ipo->offset = offset * bs;
 231         if (td->o.replay_scale)
 232                 ipo->offset = ipo->offset / td->o.replay_scale;
 233         ipo_bytes_align(td->o.replay_align, ipo);
 234         ipo->len = bytes;
 235         ipo->delay = ttime / 1000;
 236         if (rw)
 237                 ipo->ddir = DDIR_WRITE;
 238         else
 239                 ipo->ddir = DDIR_READ;
 240         ipo->fileno = fileno;
 241
 242         dprint(FD_BLKTRACE, "store ddir=%d, off=%llu, len=%lu, delay=%lu\n",
 243                                                         ipo->ddir, ipo->offset,
 244                                                         ipo->len, ipo->delay);
 245         queue_io_piece(td, ipo);
 246 }
 247
 248 static void handle_trace_notify(struct blk_io_trace *t)
 249 {
 250         switch (t->action) {
 251         case BLK_TN_PROCESS:
 252                 dprint(FD_BLKTRACE, "got process notify: %x, %d\n",
 253                                 t->action, t->pid);
 254                 break;
 255         case BLK_TN_TIMESTAMP:
 256                 dprint(FD_BLKTRACE, "got timestamp notify: %x, %d\n",
 257                                 t->action, t->pid);
 258                 break;
 259         case BLK_TN_MESSAGE:
 260                 break;
 261         default:
 262                 dprint(FD_BLKTRACE, "unknown trace act %x\n", t->action);
 263                 break;
 264         }
 265 }
 266
 267 static void handle_trace_discard(struct thread_data *td,
 268                                  struct blk_io_trace *t,
 269                                  unsigned long long ttime,
 270                                  unsigned long *ios, unsigned int *rw_bs)
 271 {
 272         struct io_piece *ipo;
 273         unsigned int bs;
 274         int fileno;
 275
 276         ipo = calloc(1, sizeof(*ipo));
 277         init_ipo(ipo);
 278         fileno = trace_add_file(td, t->device, &bs);
 279
 280         ios[DDIR_TRIM]++;
 281         if (t->bytes > rw_bs[DDIR_TRIM])
 282                 rw_bs[DDIR_TRIM] = t->bytes;
 283
 284         td->o.size += t->bytes;
 285
 286         INIT_FLIST_HEAD(&ipo->list);
 287
 288         ipo->offset = t->sector * bs;
 289         if (td->o.replay_scale)
 290                 ipo->offset = ipo->offset / td->o.replay_scale;
 291         ipo_bytes_align(td->o.replay_align, ipo);
 292         ipo->len = t->bytes;
 293         ipo->delay = ttime / 1000;
 294         ipo->ddir = DDIR_TRIM;
 295         ipo->fileno = fileno;
 296
 297         dprint(FD_BLKTRACE, "store discard, off=%llu, len=%lu, delay=%lu\n",
 298                                                         ipo->offset, ipo->len,
 299                                                         ipo->delay);
 300         queue_io_piece(td, ipo);
 301 }
 302
 303 static void handle_trace_fs(struct thread_data *td, struct blk_io_trace *t,
 304                             unsigned long long ttime, unsigned long *ios,
 305                             unsigned int *rw_bs)
 306 {
 307         unsigned int bs;
 308         int rw;
 309         int fileno;
 310
 311         fileno = trace_add_file(td, t->device, &bs);
 312
 313         rw = (t->action & BLK_TC_ACT(BLK_TC_WRITE)) != 0;
 314
 315         assert(t->bytes);
 316
 317         if (t->bytes > rw_bs[rw])
 318                 rw_bs[rw] = t->bytes;
 319
 320         ios[rw]++;
 321         td->o.size += t->bytes;
 322         store_ipo(td, t->sector, t->bytes, rw, ttime, fileno, bs);
 323 }
 324
 325 static void handle_trace_flush(struct thread_data *td, struct blk_io_trace *t,
 326                                unsigned long long ttime)
 327 {
 328         struct io_piece *ipo;
 329         unsigned int bs;
 330         int fileno;
 331
 332         ipo = calloc(1, sizeof(*ipo));
 333         init_ipo(ipo);
 334         fileno = trace_add_file(td, t->device, &bs);
 335
 336         ipo->delay = ttime / 1000;
 337         ipo->ddir = DDIR_SYNC;
 338         ipo->fileno = fileno;
 339
 340         dprint(FD_BLKTRACE, "store flush delay=%lu\n", ipo->delay);
 341         queue_io_piece(td, ipo);
 342 }
 343
 344 /*
 345  * We only care for queue traces, most of the others are side effects
 346  * due to internal workings of the block layer.
 347  */
 348 static void handle_trace(struct thread_data *td, struct blk_io_trace *t,
 349                          unsigned long *ios, unsigned int *bs)
 350 {
 351         static unsigned long long last_ttime;
 352         unsigned long long delay = 0;
 353
 354         if ((t->action & 0xffff) != __BLK_TA_QUEUE)
 355                 return;
 356
 357         if (!(t->action & BLK_TC_ACT(BLK_TC_NOTIFY))) {
 358                 if (!last_ttime || td->o.no_stall)
 359                         delay = 0;
 360                 else if (td->o.replay_time_scale == 100)
 361                         delay = t->time - last_ttime;
 362                 else {
 363                         double tmp = t->time - last_ttime;
 364                         double scale;
 365
 366                         scale = (double) 100.0 / (double) td->o.replay_time_scale;
 367                         tmp *= scale;
 368                         delay = tmp;
 369                 }
 370                 last_ttime = t->time;
 371         }
 372
 373         t_bytes_align(&td->o, t);
 374
 375         if (t->action & BLK_TC_ACT(BLK_TC_NOTIFY))
 376                 handle_trace_notify(t);
 377         else if (t->action & BLK_TC_ACT(BLK_TC_DISCARD))
 378                 handle_trace_discard(td, t, delay, ios, bs);
 379         else if (t->action & BLK_TC_ACT(BLK_TC_FLUSH))
 380                 handle_trace_flush(td, t, delay);
 381         else
 382                 handle_trace_fs(td, t, delay, ios, bs);
 383 }
 384
 385 static void byteswap_trace(struct blk_io_trace *t)
 386 {
 387         t->magic = fio_swap32(t->magic);
 388         t->sequence = fio_swap32(t->sequence);
 389         t->time = fio_swap64(t->time);
 390         t->sector = fio_swap64(t->sector);
 391         t->bytes = fio_swap32(t->bytes);
 392         t->action = fio_swap32(t->action);
 393         t->pid = fio_swap32(t->pid);
 394         t->device = fio_swap32(t->device);
 395         t->cpu = fio_swap32(t->cpu);
 396         t->error = fio_swap16(t->error);
 397         t->pdu_len = fio_swap16(t->pdu_len);
 398 }
 399
 400 static int t_is_write(struct blk_io_trace *t)
 401 {
 402         return (t->action & BLK_TC_ACT(BLK_TC_WRITE | BLK_TC_DISCARD)) != 0;
 403 }
 404
 405 static enum fio_ddir t_get_ddir(struct blk_io_trace *t)
 406 {
 407         if (t->action & BLK_TC_ACT(BLK_TC_READ))
 408                 return DDIR_READ;
 409         else if (t->action & BLK_TC_ACT(BLK_TC_WRITE))
 410                 return DDIR_WRITE;
 411         else if (t->action & BLK_TC_ACT(BLK_TC_DISCARD))
 412                 return DDIR_TRIM;
 413
 414         return DDIR_INVAL;
 415 }
 416
 417 static void depth_inc(struct blk_io_trace *t, int *depth)
 418 {
 419         enum fio_ddir ddir;
 420
 421         ddir = t_get_ddir(t);
 422         if (ddir != DDIR_INVAL)
 423                 depth[ddir]++;
 424 }
 425
 426 static void depth_dec(struct blk_io_trace *t, int *depth)
 427 {
 428         enum fio_ddir ddir;
 429
 430         ddir = t_get_ddir(t);
 431         if (ddir != DDIR_INVAL)
 432                 depth[ddir]--;
 433 }
 434
 435 static void depth_end(struct blk_io_trace *t, int *this_depth, int *depth)
 436 {
 437         enum fio_ddir ddir = DDIR_INVAL;
 438
 439         ddir = t_get_ddir(t);
 440         if (ddir != DDIR_INVAL) {
 441                 depth[ddir] = max(depth[ddir], this_depth[ddir]);
 442                 this_depth[ddir] = 0;
 443         }
 444 }
 445
 446 /*
 447  * Load a blktrace file by reading all the blk_io_trace entries, and storing
 448  * them as io_pieces like the fio text version would do.
 449  */
 450 int load_blktrace(struct thread_data *td, const char *filename, int need_swap)
 451 {
 452         struct blk_io_trace t;
 453         unsigned long ios[DDIR_RWDIR_CNT], skipped_writes;
 454         unsigned int rw_bs[DDIR_RWDIR_CNT];
 455         struct fifo *fifo;
 456         int fd, i, old_state;
 457         struct fio_file *f;
 458         int this_depth[DDIR_RWDIR_CNT], depth[DDIR_RWDIR_CNT], max_depth;
 459
 460         fd = open(filename, O_RDONLY);
 461         if (fd < 0) {
 462                 td_verror(td, errno, "open blktrace file");
 463                 return 1;
 464         }
 465
 466         fifo = fifo_alloc(TRACE_FIFO_SIZE);
 467
 468         old_state = td_bump_runstate(td, TD_SETTING_UP);
 469
 470         td->o.size = 0;
 471
 472         for (i = 0; i < DDIR_RWDIR_CNT; i++) {
 473                 ios[i] = 0;
 474                 rw_bs[i] = 0;
 475                 this_depth[i] = 0;
 476                 depth[i] = 0;
 477         }
 478
 479         skipped_writes = 0;
 480         do {
 481                 int ret = trace_fifo_get(td, fifo, fd, &t, sizeof(t));
 482
 483                 if (ret < 0)
 484                         goto err;
 485                 else if (!ret)
 486                         break;
 487                 else if (ret < (int) sizeof(t)) {
 488                         log_err("fio: short fifo get\n");
 489                         break;
 490                 }
 491
 492                 if (need_swap)
 493                         byteswap_trace(&t);
 494
 495                 if ((t.magic & 0xffffff00) != BLK_IO_TRACE_MAGIC) {
 496                         log_err("fio: bad magic in blktrace data: %x\n",
 497                                                                 t.magic);
 498                         goto err;
 499                 }
 500                 if ((t.magic & 0xff) != BLK_IO_TRACE_VERSION) {
 501                         log_err("fio: bad blktrace version %d\n",
 502                                                                 t.magic & 0xff);
 503                         goto err;
 504                 }
 505                 ret = discard_pdu(td, fifo, fd, &t);
 506                 if (ret < 0) {
 507                         td_verror(td, ret, "blktrace lseek");
 508                         goto err;
 509                 } else if (t.pdu_len != ret) {
 510                         log_err("fio: discarded %d of %d\n", ret, t.pdu_len);
 511                         goto err;
 512                 }
 513                 if ((t.action & BLK_TC_ACT(BLK_TC_NOTIFY)) == 0) {
 514                         if ((t.action & 0xffff) == __BLK_TA_QUEUE)
 515                                 depth_inc(&t, this_depth);
 516                         else if (((t.action & 0xffff) == __BLK_TA_BACKMERGE) ||
 517                                 ((t.action & 0xffff) == __BLK_TA_FRONTMERGE))
 518                                 depth_dec(&t, this_depth);
 519                         else if ((t.action & 0xffff) == __BLK_TA_COMPLETE)
 520                                 depth_end(&t, this_depth, depth);
 521
 522                         if (t_is_write(&t) && read_only) {
 523                                 skipped_writes++;
 524                                 continue;
 525                         }
 526                 }
 527
 528                 handle_trace(td, &t, ios, rw_bs);
 529         } while (1);
 530
 531         for_each_file(td, f, i)
 532                 trace_add_open_close_event(td, f->fileno, FIO_LOG_CLOSE_FILE);
 533
 534         fifo_free(fifo);
 535         close(fd);
 536
 537         td_restore_runstate(td, old_state);
 538
 539         if (!td->files_index) {
 540                 log_err("fio: did not find replay device(s)\n");
 541                 return 1;
 542         }
 543
 544         /*
 545          * For stacked devices, we don't always get a COMPLETE event so
 546          * the depth grows to insane values. Limit it to something sane(r).
 547          */
 548         max_depth = 0;
 549         for (i = 0; i < DDIR_RWDIR_CNT; i++) {
 550                 if (depth[i] > 1024)
 551                         depth[i] = 1024;
 552                 else if (!depth[i] && ios[i])
 553                         depth[i] = 1;
 554                 max_depth = max(depth[i], max_depth);
 555         }
 556
 557         if (skipped_writes)
 558                 log_err("fio: %s skips replay of %lu writes due to read-only\n",
 559                                                 td->o.name, skipped_writes);
 560
 561         if (!ios[DDIR_READ] && !ios[DDIR_WRITE]) {
 562                 log_err("fio: found no ios in blktrace data\n");
 563                 return 1;
 564         } else if (ios[DDIR_READ] && !ios[DDIR_WRITE]) {
 565                 td->o.td_ddir = TD_DDIR_READ;
 566                 td->o.max_bs[DDIR_READ] = rw_bs[DDIR_READ];
 567         } else if (!ios[DDIR_READ] && ios[DDIR_WRITE]) {
 568                 td->o.td_ddir = TD_DDIR_WRITE;
 569                 td->o.max_bs[DDIR_WRITE] = rw_bs[DDIR_WRITE];
 570         } else {
 571                 td->o.td_ddir = TD_DDIR_RW;
 572                 td->o.max_bs[DDIR_READ] = rw_bs[DDIR_READ];
 573                 td->o.max_bs[DDIR_WRITE] = rw_bs[DDIR_WRITE];
 574                 td->o.max_bs[DDIR_TRIM] = rw_bs[DDIR_TRIM];
 575         }
 576
 577         /*
 578          * We need to do direct/raw ios to the device, to avoid getting
 579          * read-ahead in our way. But only do so if the minimum block size
 580          * is a multiple of 4k, otherwise we don't know if it's safe to do so.
 581          */
 582         if (!fio_option_is_set(&td->o, odirect) && !(td_min_bs(td) & 4095))
 583                 td->o.odirect = 1;
 584
 585         /*
 586          * If depth wasn't manually set, use probed depth
 587          */
 588         if (!fio_option_is_set(&td->o, iodepth))
 589                 td->o.iodepth = td->o.iodepth_low = max_depth;
 590
 591         return 0;
 592 err:
 593         close(fd);
 594         fifo_free(fifo);
 595         return 1;
 596 }