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 = malloc(sizeof(*ipo));
 226
 227         init_ipo(ipo);
 228
 229         ipo->offset = offset * bs;
 230         if (td->o.replay_scale)
 231                 ipo->offset = ipo->offset / td->o.replay_scale;
 232         ipo_bytes_align(td->o.replay_align, ipo);
 233         ipo->len = bytes;
 234         ipo->delay = ttime / 1000;
 235         if (rw)
 236                 ipo->ddir = DDIR_WRITE;
 237         else
 238                 ipo->ddir = DDIR_READ;
 239         ipo->fileno = fileno;
 240
 241         dprint(FD_BLKTRACE, "store ddir=%d, off=%llu, len=%lu, delay=%lu\n",
 242                                                         ipo->ddir, ipo->offset,
 243                                                         ipo->len, ipo->delay);
 244         queue_io_piece(td, ipo);
 245 }
 246
 247 static void handle_trace_notify(struct blk_io_trace *t)
 248 {
 249         switch (t->action) {
 250         case BLK_TN_PROCESS:
 251                 dprint(FD_BLKTRACE, "got process notify: %x, %d\n",
 252                                 t->action, t->pid);
 253                 break;
 254         case BLK_TN_TIMESTAMP:
 255                 dprint(FD_BLKTRACE, "got timestamp notify: %x, %d\n",
 256                                 t->action, t->pid);
 257                 break;
 258         case BLK_TN_MESSAGE:
 259                 break;
 260         default:
 261                 dprint(FD_BLKTRACE, "unknown trace act %x\n", t->action);
 262                 break;
 263         }
 264 }
 265
 266 static void handle_trace_discard(struct thread_data *td,
 267                                  struct blk_io_trace *t,
 268                                  unsigned long long ttime,
 269                                  unsigned long *ios, unsigned int *rw_bs)
 270 {
 271         struct io_piece *ipo = malloc(sizeof(*ipo));
 272         unsigned int bs;
 273         int fileno;
 274
 275         init_ipo(ipo);
 276         fileno = trace_add_file(td, t->device, &bs);
 277
 278         ios[DDIR_TRIM]++;
 279         if (t->bytes > rw_bs[DDIR_TRIM])
 280                 rw_bs[DDIR_TRIM] = t->bytes;
 281
 282         td->o.size += t->bytes;
 283
 284         memset(ipo, 0, sizeof(*ipo));
 285         INIT_FLIST_HEAD(&ipo->list);
 286
 287         ipo->offset = t->sector * bs;
 288         if (td->o.replay_scale)
 289                 ipo->offset = ipo->offset / td->o.replay_scale;
 290         ipo_bytes_align(td->o.replay_align, ipo);
 291         ipo->len = t->bytes;
 292         ipo->delay = ttime / 1000;
 293         ipo->ddir = DDIR_TRIM;
 294         ipo->fileno = fileno;
 295
 296         dprint(FD_BLKTRACE, "store discard, off=%llu, len=%lu, delay=%lu\n",
 297                                                         ipo->offset, ipo->len,
 298                                                         ipo->delay);
 299         queue_io_piece(td, ipo);
 300 }
 301
 302 static void handle_trace_fs(struct thread_data *td, struct blk_io_trace *t,
 303                             unsigned long long ttime, unsigned long *ios,
 304                             unsigned int *rw_bs)
 305 {
 306         unsigned int bs;
 307         int rw;
 308         int fileno;
 309
 310         fileno = trace_add_file(td, t->device, &bs);
 311
 312         rw = (t->action & BLK_TC_ACT(BLK_TC_WRITE)) != 0;
 313
 314         if (t->bytes > rw_bs[rw])
 315                 rw_bs[rw] = t->bytes;
 316
 317         ios[rw]++;
 318         td->o.size += t->bytes;
 319         store_ipo(td, t->sector, t->bytes, rw, ttime, fileno, bs);
 320 }
 321
 322 /*
 323  * We only care for queue traces, most of the others are side effects
 324  * due to internal workings of the block layer.
 325  */
 326 static void handle_trace(struct thread_data *td, struct blk_io_trace *t,
 327                          unsigned long *ios, unsigned int *bs)
 328 {
 329         static unsigned long long last_ttime;
 330         unsigned long long delay = 0;
 331
 332         if ((t->action & 0xffff) != __BLK_TA_QUEUE)
 333                 return;
 334
 335         if (!(t->action & BLK_TC_ACT(BLK_TC_NOTIFY))) {
 336                 if (!last_ttime || td->o.no_stall) {
 337                         last_ttime = t->time;
 338                         delay = 0;
 339                 } else {
 340                         delay = t->time - last_ttime;
 341                         last_ttime = t->time;
 342                 }
 343         }
 344
 345         t_bytes_align(&td->o, t);
 346
 347         if (t->action & BLK_TC_ACT(BLK_TC_NOTIFY))
 348                 handle_trace_notify(t);
 349         else if (t->action & BLK_TC_ACT(BLK_TC_DISCARD))
 350                 handle_trace_discard(td, t, delay, ios, bs);
 351         else
 352                 handle_trace_fs(td, t, delay, ios, bs);
 353 }
 354
 355 static void byteswap_trace(struct blk_io_trace *t)
 356 {
 357         t->magic = fio_swap32(t->magic);
 358         t->sequence = fio_swap32(t->sequence);
 359         t->time = fio_swap64(t->time);
 360         t->sector = fio_swap64(t->sector);
 361         t->bytes = fio_swap32(t->bytes);
 362         t->action = fio_swap32(t->action);
 363         t->pid = fio_swap32(t->pid);
 364         t->device = fio_swap32(t->device);
 365         t->cpu = fio_swap32(t->cpu);
 366         t->error = fio_swap16(t->error);
 367         t->pdu_len = fio_swap16(t->pdu_len);
 368 }
 369
 370 static int t_is_write(struct blk_io_trace *t)
 371 {
 372         return (t->action & BLK_TC_ACT(BLK_TC_WRITE | BLK_TC_DISCARD)) != 0;
 373 }
 374
 375 static enum fio_ddir t_get_ddir(struct blk_io_trace *t)
 376 {
 377         if (t->action & BLK_TC_ACT(BLK_TC_READ))
 378                 return DDIR_READ;
 379         else if (t->action & BLK_TC_ACT(BLK_TC_WRITE))
 380                 return DDIR_WRITE;
 381         else if (t->action & BLK_TC_ACT(BLK_TC_DISCARD))
 382                 return DDIR_TRIM;
 383
 384         return DDIR_INVAL;
 385 }
 386
 387 static void depth_inc(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_dec(struct blk_io_trace *t, int *depth)
 397 {
 398         enum fio_ddir ddir;
 399
 400         ddir = t_get_ddir(t);
 401         if (ddir != DDIR_INVAL)
 402                 depth[ddir]--;
 403 }
 404
 405 static void depth_end(struct blk_io_trace *t, int *this_depth, int *depth)
 406 {
 407         enum fio_ddir ddir = DDIR_INVAL;
 408
 409         ddir = t_get_ddir(t);
 410         if (ddir != DDIR_INVAL) {
 411                 depth[ddir] = max(depth[ddir], this_depth[ddir]);
 412                 this_depth[ddir] = 0;
 413         }
 414 }
 415
 416 /*
 417  * Load a blktrace file by reading all the blk_io_trace entries, and storing
 418  * them as io_pieces like the fio text version would do.
 419  */
 420 int load_blktrace(struct thread_data *td, const char *filename, int need_swap)
 421 {
 422         struct blk_io_trace t;
 423         unsigned long ios[DDIR_RWDIR_CNT], skipped_writes;
 424         unsigned int rw_bs[DDIR_RWDIR_CNT];
 425         struct fifo *fifo;
 426         int fd, i, old_state;
 427         struct fio_file *f;
 428         int this_depth[DDIR_RWDIR_CNT], depth[DDIR_RWDIR_CNT], max_depth;
 429
 430         fd = open(filename, O_RDONLY);
 431         if (fd < 0) {
 432                 td_verror(td, errno, "open blktrace file");
 433                 return 1;
 434         }
 435
 436         fifo = fifo_alloc(TRACE_FIFO_SIZE);
 437
 438         old_state = td_bump_runstate(td, TD_SETTING_UP);
 439
 440         td->o.size = 0;
 441
 442         for (i = 0; i < DDIR_RWDIR_CNT; i++) {
 443                 ios[i] = 0;
 444                 rw_bs[i] = 0;
 445                 this_depth[i] = 0;
 446                 depth[i] = 0;
 447         }
 448
 449         skipped_writes = 0;
 450         do {
 451                 int ret = trace_fifo_get(td, fifo, fd, &t, sizeof(t));
 452
 453                 if (ret < 0)
 454                         goto err;
 455                 else if (!ret)
 456                         break;
 457                 else if (ret < (int) sizeof(t)) {
 458                         log_err("fio: short fifo get\n");
 459                         break;
 460                 }
 461
 462                 if (need_swap)
 463                         byteswap_trace(&t);
 464
 465                 if ((t.magic & 0xffffff00) != BLK_IO_TRACE_MAGIC) {
 466                         log_err("fio: bad magic in blktrace data: %x\n",
 467                                                                 t.magic);
 468                         goto err;
 469                 }
 470                 if ((t.magic & 0xff) != BLK_IO_TRACE_VERSION) {
 471                         log_err("fio: bad blktrace version %d\n",
 472                                                                 t.magic & 0xff);
 473                         goto err;
 474                 }
 475                 ret = discard_pdu(td, fifo, fd, &t);
 476                 if (ret < 0) {
 477                         td_verror(td, ret, "blktrace lseek");
 478                         goto err;
 479                 } else if (t.pdu_len != ret) {
 480                         log_err("fio: discarded %d of %d\n", ret, t.pdu_len);
 481                         goto err;
 482                 }
 483                 if ((t.action & BLK_TC_ACT(BLK_TC_NOTIFY)) == 0) {
 484                         if ((t.action & 0xffff) == __BLK_TA_QUEUE)
 485                                 depth_inc(&t, this_depth);
 486                         else if (((t.action & 0xffff) == __BLK_TA_BACKMERGE) ||
 487                                 ((t.action & 0xffff) == __BLK_TA_FRONTMERGE))
 488                                 depth_dec(&t, this_depth);
 489                         else if ((t.action & 0xffff) == __BLK_TA_COMPLETE)
 490                                 depth_end(&t, this_depth, depth);
 491
 492                         if (t_is_write(&t) && read_only) {
 493                                 skipped_writes++;
 494                                 continue;
 495                         }
 496                 }
 497
 498                 handle_trace(td, &t, ios, rw_bs);
 499         } while (1);
 500
 501         for_each_file(td, f, i)
 502                 trace_add_open_close_event(td, f->fileno, FIO_LOG_CLOSE_FILE);
 503
 504         fifo_free(fifo);
 505         close(fd);
 506
 507         td_restore_runstate(td, old_state);
 508
 509         if (!td->files_index) {
 510                 log_err("fio: did not find replay device(s)\n");
 511                 return 1;
 512         }
 513
 514         /*
 515          * For stacked devices, we don't always get a COMPLETE event so
 516          * the depth grows to insane values. Limit it to something sane(r).
 517          */
 518         max_depth = 0;
 519         for (i = 0; i < DDIR_RWDIR_CNT; i++) {
 520                 if (depth[i] > 1024)
 521                         depth[i] = 1024;
 522                 else if (!depth[i] && ios[i])
 523                         depth[i] = 1;
 524                 max_depth = max(depth[i], max_depth);
 525         }
 526
 527         if (skipped_writes)
 528                 log_err("fio: %s skips replay of %lu writes due to read-only\n",
 529                                                 td->o.name, skipped_writes);
 530
 531         if (!ios[DDIR_READ] && !ios[DDIR_WRITE]) {
 532                 log_err("fio: found no ios in blktrace data\n");
 533                 return 1;
 534         } else if (ios[DDIR_READ] && !ios[DDIR_WRITE]) {
 535                 td->o.td_ddir = TD_DDIR_READ;
 536                 td->o.max_bs[DDIR_READ] = rw_bs[DDIR_READ];
 537         } else if (!ios[DDIR_READ] && ios[DDIR_WRITE]) {
 538                 td->o.td_ddir = TD_DDIR_WRITE;
 539                 td->o.max_bs[DDIR_WRITE] = rw_bs[DDIR_WRITE];
 540         } else {
 541                 td->o.td_ddir = TD_DDIR_RW;
 542                 td->o.max_bs[DDIR_READ] = rw_bs[DDIR_READ];
 543                 td->o.max_bs[DDIR_WRITE] = rw_bs[DDIR_WRITE];
 544                 td->o.max_bs[DDIR_TRIM] = rw_bs[DDIR_TRIM];
 545         }
 546
 547         /*
 548          * We need to do direct/raw ios to the device, to avoid getting
 549          * read-ahead in our way. But only do so if the minimum block size
 550          * is a multiple of 4k, otherwise we don't know if it's safe to do so.
 551          */
 552         if (!fio_option_is_set(&td->o, odirect) && !(td_min_bs(td) & 4095))
 553                 td->o.odirect = 1;
 554
 555         /*
 556          * If depth wasn't manually set, use probed depth
 557          */
 558         if (!fio_option_is_set(&td->o, iodepth))
 559                 td->o.iodepth = td->o.iodepth_low = max_depth;
 560
 561         return 0;
 562 err:
 563         close(fd);
 564         fifo_free(fifo);
 565         return 1;
 566 }