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         INIT_FLIST_HEAD(&ipo->list);
 285
 286         ipo->offset = t->sector * bs;
 287         if (td->o.replay_scale)
 288                 ipo->offset = ipo->offset / td->o.replay_scale;
 289         ipo_bytes_align(td->o.replay_align, ipo);
 290         ipo->len = t->bytes;
 291         ipo->delay = ttime / 1000;
 292         ipo->ddir = DDIR_TRIM;
 293         ipo->fileno = fileno;
 294
 295         dprint(FD_BLKTRACE, "store discard, off=%llu, len=%lu, delay=%lu\n",
 296                                                         ipo->offset, ipo->len,
 297                                                         ipo->delay);
 298         queue_io_piece(td, ipo);
 299 }
 300
 301 static void handle_trace_fs(struct thread_data *td, struct blk_io_trace *t,
 302                             unsigned long long ttime, unsigned long *ios,
 303                             unsigned int *rw_bs)
 304 {
 305         unsigned int bs;
 306         int rw;
 307         int fileno;
 308
 309         fileno = trace_add_file(td, t->device, &bs);
 310
 311         rw = (t->action & BLK_TC_ACT(BLK_TC_WRITE)) != 0;
 312
 313         if (t->bytes > rw_bs[rw])
 314                 rw_bs[rw] = t->bytes;
 315
 316         ios[rw]++;
 317         td->o.size += t->bytes;
 318         store_ipo(td, t->sector, t->bytes, rw, ttime, fileno, bs);
 319 }
 320
 321 /*
 322  * We only care for queue traces, most of the others are side effects
 323  * due to internal workings of the block layer.
 324  */
 325 static void handle_trace(struct thread_data *td, struct blk_io_trace *t,
 326                          unsigned long *ios, unsigned int *bs)
 327 {
 328         static unsigned long long last_ttime;
 329         unsigned long long delay = 0;
 330
 331         if ((t->action & 0xffff) != __BLK_TA_QUEUE)
 332                 return;
 333
 334         if (!(t->action & BLK_TC_ACT(BLK_TC_NOTIFY))) {
 335                 if (!last_ttime || td->o.no_stall)
 336                         delay = 0;
 337                 else if (td->o.replay_time_scale == 100)
 338                         delay = t->time - last_ttime;
 339                 else {
 340                         double tmp = t->time - last_ttime;
 341                         double scale;
 342
 343                         scale = (double) 100.0 / (double) td->o.replay_time_scale;
 344                         tmp *= scale;
 345                         delay = tmp;
 346                 }
 347                 last_ttime = t->time;
 348         }
 349
 350         t_bytes_align(&td->o, t);
 351
 352         if (t->action & BLK_TC_ACT(BLK_TC_NOTIFY))
 353                 handle_trace_notify(t);
 354         else if (t->action & BLK_TC_ACT(BLK_TC_DISCARD))
 355                 handle_trace_discard(td, t, delay, ios, bs);
 356         else
 357                 handle_trace_fs(td, t, delay, ios, bs);
 358 }
 359
 360 static void byteswap_trace(struct blk_io_trace *t)
 361 {
 362         t->magic = fio_swap32(t->magic);
 363         t->sequence = fio_swap32(t->sequence);
 364         t->time = fio_swap64(t->time);
 365         t->sector = fio_swap64(t->sector);
 366         t->bytes = fio_swap32(t->bytes);
 367         t->action = fio_swap32(t->action);
 368         t->pid = fio_swap32(t->pid);
 369         t->device = fio_swap32(t->device);
 370         t->cpu = fio_swap32(t->cpu);
 371         t->error = fio_swap16(t->error);
 372         t->pdu_len = fio_swap16(t->pdu_len);
 373 }
 374
 375 static int t_is_write(struct blk_io_trace *t)
 376 {
 377         return (t->action & BLK_TC_ACT(BLK_TC_WRITE | BLK_TC_DISCARD)) != 0;
 378 }
 379
 380 static enum fio_ddir t_get_ddir(struct blk_io_trace *t)
 381 {
 382         if (t->action & BLK_TC_ACT(BLK_TC_READ))
 383                 return DDIR_READ;
 384         else if (t->action & BLK_TC_ACT(BLK_TC_WRITE))
 385                 return DDIR_WRITE;
 386         else if (t->action & BLK_TC_ACT(BLK_TC_DISCARD))
 387                 return DDIR_TRIM;
 388
 389         return DDIR_INVAL;
 390 }
 391
 392 static void depth_inc(struct blk_io_trace *t, int *depth)
 393 {
 394         enum fio_ddir ddir;
 395
 396         ddir = t_get_ddir(t);
 397         if (ddir != DDIR_INVAL)
 398                 depth[ddir]++;
 399 }
 400
 401 static void depth_dec(struct blk_io_trace *t, int *depth)
 402 {
 403         enum fio_ddir ddir;
 404
 405         ddir = t_get_ddir(t);
 406         if (ddir != DDIR_INVAL)
 407                 depth[ddir]--;
 408 }
 409
 410 static void depth_end(struct blk_io_trace *t, int *this_depth, int *depth)
 411 {
 412         enum fio_ddir ddir = DDIR_INVAL;
 413
 414         ddir = t_get_ddir(t);
 415         if (ddir != DDIR_INVAL) {
 416                 depth[ddir] = max(depth[ddir], this_depth[ddir]);
 417                 this_depth[ddir] = 0;
 418         }
 419 }
 420
 421 /*
 422  * Load a blktrace file by reading all the blk_io_trace entries, and storing
 423  * them as io_pieces like the fio text version would do.
 424  */
 425 int load_blktrace(struct thread_data *td, const char *filename, int need_swap)
 426 {
 427         struct blk_io_trace t;
 428         unsigned long ios[DDIR_RWDIR_CNT], skipped_writes;
 429         unsigned int rw_bs[DDIR_RWDIR_CNT];
 430         struct fifo *fifo;
 431         int fd, i, old_state;
 432         struct fio_file *f;
 433         int this_depth[DDIR_RWDIR_CNT], depth[DDIR_RWDIR_CNT], max_depth;
 434
 435         fd = open(filename, O_RDONLY);
 436         if (fd < 0) {
 437                 td_verror(td, errno, "open blktrace file");
 438                 return 1;
 439         }
 440
 441         fifo = fifo_alloc(TRACE_FIFO_SIZE);
 442
 443         old_state = td_bump_runstate(td, TD_SETTING_UP);
 444
 445         td->o.size = 0;
 446
 447         for (i = 0; i < DDIR_RWDIR_CNT; i++) {
 448                 ios[i] = 0;
 449                 rw_bs[i] = 0;
 450                 this_depth[i] = 0;
 451                 depth[i] = 0;
 452         }
 453
 454         skipped_writes = 0;
 455         do {
 456                 int ret = trace_fifo_get(td, fifo, fd, &t, sizeof(t));
 457
 458                 if (ret < 0)
 459                         goto err;
 460                 else if (!ret)
 461                         break;
 462                 else if (ret < (int) sizeof(t)) {
 463                         log_err("fio: short fifo get\n");
 464                         break;
 465                 }
 466
 467                 if (need_swap)
 468                         byteswap_trace(&t);
 469
 470                 if ((t.magic & 0xffffff00) != BLK_IO_TRACE_MAGIC) {
 471                         log_err("fio: bad magic in blktrace data: %x\n",
 472                                                                 t.magic);
 473                         goto err;
 474                 }
 475                 if ((t.magic & 0xff) != BLK_IO_TRACE_VERSION) {
 476                         log_err("fio: bad blktrace version %d\n",
 477                                                                 t.magic & 0xff);
 478                         goto err;
 479                 }
 480                 ret = discard_pdu(td, fifo, fd, &t);
 481                 if (ret < 0) {
 482                         td_verror(td, ret, "blktrace lseek");
 483                         goto err;
 484                 } else if (t.pdu_len != ret) {
 485                         log_err("fio: discarded %d of %d\n", ret, t.pdu_len);
 486                         goto err;
 487                 }
 488                 if ((t.action & BLK_TC_ACT(BLK_TC_NOTIFY)) == 0) {
 489                         if ((t.action & 0xffff) == __BLK_TA_QUEUE)
 490                                 depth_inc(&t, this_depth);
 491                         else if (((t.action & 0xffff) == __BLK_TA_BACKMERGE) ||
 492                                 ((t.action & 0xffff) == __BLK_TA_FRONTMERGE))
 493                                 depth_dec(&t, this_depth);
 494                         else if ((t.action & 0xffff) == __BLK_TA_COMPLETE)
 495                                 depth_end(&t, this_depth, depth);
 496
 497                         if (t_is_write(&t) && read_only) {
 498                                 skipped_writes++;
 499                                 continue;
 500                         }
 501                 }
 502
 503                 handle_trace(td, &t, ios, rw_bs);
 504         } while (1);
 505
 506         for_each_file(td, f, i)
 507                 trace_add_open_close_event(td, f->fileno, FIO_LOG_CLOSE_FILE);
 508
 509         fifo_free(fifo);
 510         close(fd);
 511
 512         td_restore_runstate(td, old_state);
 513
 514         if (!td->files_index) {
 515                 log_err("fio: did not find replay device(s)\n");
 516                 return 1;
 517         }
 518
 519         /*
 520          * For stacked devices, we don't always get a COMPLETE event so
 521          * the depth grows to insane values. Limit it to something sane(r).
 522          */
 523         max_depth = 0;
 524         for (i = 0; i < DDIR_RWDIR_CNT; i++) {
 525                 if (depth[i] > 1024)
 526                         depth[i] = 1024;
 527                 else if (!depth[i] && ios[i])
 528                         depth[i] = 1;
 529                 max_depth = max(depth[i], max_depth);
 530         }
 531
 532         if (skipped_writes)
 533                 log_err("fio: %s skips replay of %lu writes due to read-only\n",
 534                                                 td->o.name, skipped_writes);
 535
 536         if (!ios[DDIR_READ] && !ios[DDIR_WRITE]) {
 537                 log_err("fio: found no ios in blktrace data\n");
 538                 return 1;
 539         } else if (ios[DDIR_READ] && !ios[DDIR_WRITE]) {
 540                 td->o.td_ddir = TD_DDIR_READ;
 541                 td->o.max_bs[DDIR_READ] = rw_bs[DDIR_READ];
 542         } else if (!ios[DDIR_READ] && ios[DDIR_WRITE]) {
 543                 td->o.td_ddir = TD_DDIR_WRITE;
 544                 td->o.max_bs[DDIR_WRITE] = rw_bs[DDIR_WRITE];
 545         } else {
 546                 td->o.td_ddir = TD_DDIR_RW;
 547                 td->o.max_bs[DDIR_READ] = rw_bs[DDIR_READ];
 548                 td->o.max_bs[DDIR_WRITE] = rw_bs[DDIR_WRITE];
 549                 td->o.max_bs[DDIR_TRIM] = rw_bs[DDIR_TRIM];
 550         }
 551
 552         /*
 553          * We need to do direct/raw ios to the device, to avoid getting
 554          * read-ahead in our way. But only do so if the minimum block size
 555          * is a multiple of 4k, otherwise we don't know if it's safe to do so.
 556          */
 557         if (!fio_option_is_set(&td->o, odirect) && !(td_min_bs(td) & 4095))
 558                 td->o.odirect = 1;
 559
 560         /*
 561          * If depth wasn't manually set, use probed depth
 562          */
 563         if (!fio_option_is_set(&td->o, iodepth))
 564                 td->o.iodepth = td->o.iodepth_low = max_depth;
 565
 566         return 0;
 567 err:
 568         close(fd);
 569         fifo_free(fifo);
 570         return 1;
 571 }