blktrace.c

   1 /*
   2  * blktrace support code for fio
   3  */
   4 #include <stdio.h>
   5 #include <stdlib.h>
   6 #include <sys/stat.h>
   7 #include <sys/ioctl.h>
   8 #include <linux/fs.h>
   9 #include <dirent.h>
  10
  11 #include "flist.h"
  12 #include "fio.h"
  13 #include "blktrace_api.h"
  14 #include "lib/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 int 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 0;
  85
  86         ret = read(fd, &t, sizeof(t));
  87         close(fd);
  88
  89         if (ret < 0) {
  90                 perror("read blktrace");
  91                 return 0;
  92         } else if (ret != sizeof(t)) {
  93                 log_err("fio: short read on blktrace file\n");
  94                 return 0;
  95         }
  96
  97         if ((t.magic & 0xffffff00) == BLK_IO_TRACE_MAGIC) {
  98                 *need_swap = 0;
  99                 return 1;
 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 1;
 109         }
 110
 111         return 0;
 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 /*
 212  * Store blk_io_trace data in an ipo for later retrieval.
 213  */
 214 static void store_ipo(struct thread_data *td, unsigned long long offset,
 215                       unsigned int bytes, int rw, unsigned long long ttime,
 216                       int fileno, unsigned int bs)
 217 {
 218         struct io_piece *ipo = malloc(sizeof(*ipo));
 219
 220         init_ipo(ipo);
 221
 222         ipo->offset = offset * bs;
 223         ipo->len = bytes;
 224         ipo->delay = ttime / 1000;
 225         if (rw)
 226                 ipo->ddir = DDIR_WRITE;
 227         else
 228                 ipo->ddir = DDIR_READ;
 229         ipo->fileno = fileno;
 230
 231         dprint(FD_BLKTRACE, "store ddir=%d, off=%llu, len=%lu, delay=%lu\n",
 232                                                         ipo->ddir, ipo->offset,
 233                                                         ipo->len, ipo->delay);
 234         queue_io_piece(td, ipo);
 235 }
 236
 237 static void handle_trace_notify(struct blk_io_trace *t)
 238 {
 239         switch (t->action) {
 240         case BLK_TN_PROCESS:
 241                 dprint(FD_BLKTRACE, "got process notify: %x, %d\n",
 242                                 t->action, t->pid);
 243                 break;
 244         case BLK_TN_TIMESTAMP:
 245                 dprint(FD_BLKTRACE, "got timestamp notify: %x, %d\n",
 246                                 t->action, t->pid);
 247                 break;
 248         case BLK_TN_MESSAGE:
 249                 break;
 250         default:
 251                 dprint(FD_BLKTRACE, "unknown trace act %x\n", t->action);
 252                 break;
 253         }
 254 }
 255
 256 static void handle_trace_discard(struct thread_data *td,
 257                                  struct blk_io_trace *t,
 258                                  unsigned long long ttime,
 259                                  unsigned long *ios, unsigned int *rw_bs)
 260 {
 261         struct io_piece *ipo = malloc(sizeof(*ipo));
 262         unsigned int bs;
 263         int fileno;
 264
 265         init_ipo(ipo);
 266         fileno = trace_add_file(td, t->device, &bs);
 267
 268         ios[DDIR_TRIM]++;
 269         if (t->bytes > rw_bs[DDIR_TRIM])
 270                 rw_bs[DDIR_TRIM] = t->bytes;
 271
 272         td->o.size += t->bytes;
 273
 274         memset(ipo, 0, sizeof(*ipo));
 275         INIT_FLIST_HEAD(&ipo->list);
 276
 277         ipo->offset = t->sector * bs;
 278         ipo->len = t->bytes;
 279         ipo->delay = ttime / 1000;
 280         ipo->ddir = DDIR_TRIM;
 281         ipo->fileno = fileno;
 282
 283         dprint(FD_BLKTRACE, "store discard, off=%llu, len=%lu, delay=%lu\n",
 284                                                         ipo->offset, ipo->len,
 285                                                         ipo->delay);
 286         queue_io_piece(td, ipo);
 287 }
 288
 289 static void handle_trace_fs(struct thread_data *td, struct blk_io_trace *t,
 290                             unsigned long long ttime, unsigned long *ios,
 291                             unsigned int *rw_bs)
 292 {
 293         unsigned int bs;
 294         int rw;
 295         int fileno;
 296
 297         fileno = trace_add_file(td, t->device, &bs);
 298
 299         rw = (t->action & BLK_TC_ACT(BLK_TC_WRITE)) != 0;
 300
 301         if (t->bytes > rw_bs[rw])
 302                 rw_bs[rw] = t->bytes;
 303
 304         ios[rw]++;
 305         td->o.size += t->bytes;
 306         store_ipo(td, t->sector, t->bytes, rw, ttime, fileno, bs);
 307 }
 308
 309 /*
 310  * We only care for queue traces, most of the others are side effects
 311  * due to internal workings of the block layer.
 312  */
 313 static void handle_trace(struct thread_data *td, struct blk_io_trace *t,
 314                          unsigned long *ios, unsigned int *bs)
 315 {
 316         static unsigned long long last_ttime;
 317         unsigned long long delay;
 318
 319         if ((t->action & 0xffff) != __BLK_TA_QUEUE)
 320                 return;
 321
 322         if (!(t->action & BLK_TC_ACT(BLK_TC_NOTIFY))) {
 323                 if (!last_ttime || td->o.no_stall) {
 324                         last_ttime = t->time;
 325                         delay = 0;
 326                 } else {
 327                         delay = t->time - last_ttime;
 328                         last_ttime = t->time;
 329                 }
 330         }
 331
 332         if (t->action & BLK_TC_ACT(BLK_TC_NOTIFY))
 333                 handle_trace_notify(t);
 334         else if (t->action & BLK_TC_ACT(BLK_TC_DISCARD))
 335                 handle_trace_discard(td, t, delay, ios, bs);
 336         else
 337                 handle_trace_fs(td, t, delay, ios, bs);
 338 }
 339
 340 static void byteswap_trace(struct blk_io_trace *t)
 341 {
 342         t->magic = fio_swap32(t->magic);
 343         t->sequence = fio_swap32(t->sequence);
 344         t->time = fio_swap64(t->time);
 345         t->sector = fio_swap64(t->sector);
 346         t->bytes = fio_swap32(t->bytes);
 347         t->action = fio_swap32(t->action);
 348         t->pid = fio_swap32(t->pid);
 349         t->device = fio_swap32(t->device);
 350         t->cpu = fio_swap32(t->cpu);
 351         t->error = fio_swap16(t->error);
 352         t->pdu_len = fio_swap16(t->pdu_len);
 353 }
 354
 355 static int t_is_write(struct blk_io_trace *t)
 356 {
 357         return (t->action & BLK_TC_ACT(BLK_TC_WRITE | BLK_TC_DISCARD)) != 0;
 358 }
 359
 360 static enum fio_ddir t_get_ddir(struct blk_io_trace *t)
 361 {
 362         if (t->action & BLK_TC_ACT(BLK_TC_READ))
 363                 return DDIR_READ;
 364         else if (t->action & BLK_TC_ACT(BLK_TC_WRITE))
 365                 return DDIR_WRITE;
 366         else if (t->action & BLK_TC_ACT(BLK_TC_DISCARD))
 367                 return DDIR_TRIM;
 368
 369         return DDIR_INVAL;
 370 }
 371
 372 static void depth_inc(struct blk_io_trace *t, int *depth)
 373 {
 374         enum fio_ddir ddir;
 375
 376         ddir = t_get_ddir(t);
 377         if (ddir != DDIR_INVAL)
 378                 depth[ddir]++;
 379 }
 380
 381 static void depth_dec(struct blk_io_trace *t, int *depth)
 382 {
 383         enum fio_ddir ddir;
 384
 385         ddir = t_get_ddir(t);
 386         if (ddir != DDIR_INVAL)
 387                 depth[ddir]--;
 388 }
 389
 390 static void depth_end(struct blk_io_trace *t, int *this_depth, int *depth)
 391 {
 392         enum fio_ddir ddir = DDIR_INVAL;
 393
 394         ddir = t_get_ddir(t);
 395         if (ddir != DDIR_INVAL) {
 396                 depth[ddir] = max(depth[ddir], this_depth[ddir]);
 397                 this_depth[ddir] = 0;
 398         }
 399 }
 400
 401 /*
 402  * Load a blktrace file by reading all the blk_io_trace entries, and storing
 403  * them as io_pieces like the fio text version would do.
 404  */
 405 int load_blktrace(struct thread_data *td, const char *filename, int need_swap)
 406 {
 407         struct blk_io_trace t;
 408         unsigned long ios[DDIR_RWDIR_CNT], skipped_writes;
 409         unsigned int rw_bs[DDIR_RWDIR_CNT];
 410         struct fifo *fifo;
 411         int fd, i, old_state;
 412         struct fio_file *f;
 413         int this_depth[DDIR_RWDIR_CNT], depth[DDIR_RWDIR_CNT], max_depth;
 414
 415         fd = open(filename, O_RDONLY);
 416         if (fd < 0) {
 417                 td_verror(td, errno, "open blktrace file");
 418                 return 1;
 419         }
 420
 421         fifo = fifo_alloc(TRACE_FIFO_SIZE);
 422
 423         old_state = td_bump_runstate(td, TD_SETTING_UP);
 424
 425         td->o.size = 0;
 426
 427         for (i = 0; i < DDIR_RWDIR_CNT; i++) {
 428                 ios[i] = 0;
 429                 rw_bs[i] = 0;
 430                 this_depth[i] = 0;
 431                 depth[i] = 0;
 432         }
 433
 434         skipped_writes = 0;
 435         do {
 436                 int ret = trace_fifo_get(td, fifo, fd, &t, sizeof(t));
 437
 438                 if (ret < 0)
 439                         goto err;
 440                 else if (!ret)
 441                         break;
 442                 else if (ret < (int) sizeof(t)) {
 443                         log_err("fio: short fifo get\n");
 444                         break;
 445                 }
 446
 447                 if (need_swap)
 448                         byteswap_trace(&t);
 449
 450                 if ((t.magic & 0xffffff00) != BLK_IO_TRACE_MAGIC) {
 451                         log_err("fio: bad magic in blktrace data: %x\n",
 452                                                                 t.magic);
 453                         goto err;
 454                 }
 455                 if ((t.magic & 0xff) != BLK_IO_TRACE_VERSION) {
 456                         log_err("fio: bad blktrace version %d\n",
 457                                                                 t.magic & 0xff);
 458                         goto err;
 459                 }
 460                 ret = discard_pdu(td, fifo, fd, &t);
 461                 if (ret < 0) {
 462                         td_verror(td, ret, "blktrace lseek");
 463                         goto err;
 464                 } else if (t.pdu_len != ret) {
 465                         log_err("fio: discarded %d of %d\n", ret, t.pdu_len);
 466                         goto err;
 467                 }
 468                 if ((t.action & BLK_TC_ACT(BLK_TC_NOTIFY)) == 0) {
 469                         if ((t.action & 0xffff) == __BLK_TA_QUEUE)
 470                                 depth_inc(&t, this_depth);
 471                         else if (((t.action & 0xffff) == __BLK_TA_BACKMERGE) ||
 472                                 ((t.action & 0xffff) == __BLK_TA_FRONTMERGE))
 473                                 depth_dec(&t, this_depth);
 474                         else if ((t.action & 0xffff) == __BLK_TA_COMPLETE)
 475                                 depth_end(&t, this_depth, depth);
 476
 477                         if (t_is_write(&t) && read_only) {
 478                                 skipped_writes++;
 479                                 continue;
 480                         }
 481                 }
 482
 483                 handle_trace(td, &t, ios, rw_bs);
 484         } while (1);
 485
 486         for (i = 0; i < td->files_index; i++) {
 487                 f = td->files[i];
 488                 trace_add_open_close_event(td, f->fileno, FIO_LOG_CLOSE_FILE);
 489         }
 490
 491         fifo_free(fifo);
 492         close(fd);
 493
 494         td_restore_runstate(td, old_state);
 495
 496         if (!td->files_index) {
 497                 log_err("fio: did not find replay device(s)\n");
 498                 return 1;
 499         }
 500
 501         /*
 502          * For stacked devices, we don't always get a COMPLETE event so
 503          * the depth grows to insane values. Limit it to something sane(r).
 504          */
 505         max_depth = 0;
 506         for (i = 0; i < DDIR_RWDIR_CNT; i++) {
 507                 if (depth[i] > 1024)
 508                         depth[i] = 1024;
 509                 else if (!depth[i] && ios[i])
 510                         depth[i] = 1;
 511                 max_depth = max(depth[i], max_depth);
 512         }
 513
 514         if (skipped_writes)
 515                 log_err("fio: %s skips replay of %lu writes due to read-only\n",
 516                                                 td->o.name, skipped_writes);
 517
 518         if (!ios[DDIR_READ] && !ios[DDIR_WRITE]) {
 519                 log_err("fio: found no ios in blktrace data\n");
 520                 return 1;
 521         } else if (ios[DDIR_READ] && !ios[DDIR_WRITE]) {
 522                 td->o.td_ddir = TD_DDIR_READ;
 523                 td->o.max_bs[DDIR_READ] = rw_bs[DDIR_READ];
 524         } else if (!ios[DDIR_READ] && ios[DDIR_WRITE]) {
 525                 td->o.td_ddir = TD_DDIR_WRITE;
 526                 td->o.max_bs[DDIR_WRITE] = rw_bs[DDIR_WRITE];
 527         } else {
 528                 td->o.td_ddir = TD_DDIR_RW;
 529                 td->o.max_bs[DDIR_READ] = rw_bs[DDIR_READ];
 530                 td->o.max_bs[DDIR_WRITE] = rw_bs[DDIR_WRITE];
 531                 td->o.max_bs[DDIR_TRIM] = rw_bs[DDIR_TRIM];
 532         }
 533
 534         /*
 535          * We need to do direct/raw ios to the device, to avoid getting
 536          * read-ahead in our way. But only do so if the minimum block size
 537          * is a multiple of 4k, otherwise we don't know if it's safe to do so.
 538          */
 539         if (!fio_option_is_set(&td->o, odirect) && !(td_min_bs(td) & 4095))
 540                 td->o.odirect = 1;
 541
 542         /*
 543          * If depth wasn't manually set, use probed depth
 544          */
 545         if (!fio_option_is_set(&td->o, iodepth))
 546                 td->o.iodepth = td->o.iodepth_low = max_depth;
 547
 548         return 0;
 549 err:
 550         close(fd);
 551         fifo_free(fifo);
 552         return 1;
 553 }