Jens Axboe | ac9b910 | 2011-10-01 15:01:08 -0400 | [diff] [blame] | 1 | /* |
| 2 | * Code related to writing an iolog of what a thread is doing, and to |
| 3 | * later read that back and replay |
| 4 | */ |
| 5 | #include <stdio.h> |
| 6 | #include <stdlib.h> |
| 7 | #include <libgen.h> |
| 8 | #include <assert.h> |
| 9 | #include "flist.h" |
| 10 | #include "fio.h" |
| 11 | #include "verify.h" |
| 12 | #include "trim.h" |
| 13 | |
| 14 | static const char iolog_ver2[] = "fio version 2 iolog"; |
| 15 | |
| 16 | void queue_io_piece(struct thread_data *td, struct io_piece *ipo) |
| 17 | { |
| 18 | flist_add_tail(&ipo->list, &td->io_log_list); |
| 19 | td->total_io_size += ipo->len; |
| 20 | } |
| 21 | |
| 22 | void log_io_u(struct thread_data *td, struct io_u *io_u) |
| 23 | { |
| 24 | const char *act[] = { "read", "write", "sync", "datasync", |
| 25 | "sync_file_range", "wait", "trim" }; |
| 26 | |
| 27 | assert(io_u->ddir <= 6); |
| 28 | |
| 29 | if (!td->o.write_iolog_file) |
| 30 | return; |
| 31 | |
| 32 | fprintf(td->iolog_f, "%s %s %llu %lu\n", io_u->file->file_name, |
| 33 | act[io_u->ddir], io_u->offset, |
| 34 | io_u->buflen); |
| 35 | } |
| 36 | |
| 37 | void log_file(struct thread_data *td, struct fio_file *f, |
| 38 | enum file_log_act what) |
| 39 | { |
| 40 | const char *act[] = { "add", "open", "close" }; |
| 41 | |
| 42 | assert(what < 3); |
| 43 | |
| 44 | if (!td->o.write_iolog_file) |
| 45 | return; |
| 46 | |
| 47 | |
| 48 | /* |
| 49 | * this happens on the pre-open/close done before the job starts |
| 50 | */ |
| 51 | if (!td->iolog_f) |
| 52 | return; |
| 53 | |
| 54 | fprintf(td->iolog_f, "%s %s\n", f->file_name, act[what]); |
| 55 | } |
| 56 | |
| 57 | static void iolog_delay(struct thread_data *td, unsigned long delay) |
| 58 | { |
| 59 | unsigned long usec = utime_since_now(&td->last_issue); |
| 60 | |
| 61 | if (delay < usec) |
| 62 | return; |
| 63 | |
| 64 | delay -= usec; |
| 65 | |
| 66 | /* |
| 67 | * less than 100 usec delay, just regard it as noise |
| 68 | */ |
| 69 | if (delay < 100) |
| 70 | return; |
| 71 | |
| 72 | usec_sleep(td, delay); |
| 73 | } |
| 74 | |
| 75 | static int ipo_special(struct thread_data *td, struct io_piece *ipo) |
| 76 | { |
| 77 | struct fio_file *f; |
| 78 | int ret; |
| 79 | |
| 80 | /* |
| 81 | * Not a special ipo |
| 82 | */ |
| 83 | if (ipo->ddir != DDIR_INVAL) |
| 84 | return 0; |
| 85 | |
| 86 | f = td->files[ipo->fileno]; |
| 87 | |
| 88 | switch (ipo->file_action) { |
| 89 | case FIO_LOG_OPEN_FILE: |
| 90 | ret = td_io_open_file(td, f); |
| 91 | if (!ret) |
| 92 | break; |
| 93 | td_verror(td, ret, "iolog open file"); |
| 94 | return -1; |
| 95 | case FIO_LOG_CLOSE_FILE: |
| 96 | td_io_close_file(td, f); |
| 97 | break; |
| 98 | case FIO_LOG_UNLINK_FILE: |
| 99 | unlink(f->file_name); |
| 100 | break; |
| 101 | default: |
| 102 | log_err("fio: bad file action %d\n", ipo->file_action); |
| 103 | break; |
| 104 | } |
| 105 | |
| 106 | return 1; |
| 107 | } |
| 108 | |
| 109 | int read_iolog_get(struct thread_data *td, struct io_u *io_u) |
| 110 | { |
| 111 | struct io_piece *ipo; |
| 112 | unsigned long elapsed; |
| 113 | |
| 114 | while (!flist_empty(&td->io_log_list)) { |
| 115 | int ret; |
| 116 | |
| 117 | ipo = flist_entry(td->io_log_list.next, struct io_piece, list); |
| 118 | flist_del(&ipo->list); |
| 119 | remove_trim_entry(td, ipo); |
| 120 | |
| 121 | ret = ipo_special(td, ipo); |
| 122 | if (ret < 0) { |
| 123 | free(ipo); |
| 124 | break; |
| 125 | } else if (ret > 0) { |
| 126 | free(ipo); |
| 127 | continue; |
| 128 | } |
| 129 | |
| 130 | io_u->ddir = ipo->ddir; |
| 131 | if (ipo->ddir != DDIR_WAIT) { |
| 132 | io_u->offset = ipo->offset; |
| 133 | io_u->buflen = ipo->len; |
| 134 | io_u->file = td->files[ipo->fileno]; |
| 135 | get_file(io_u->file); |
| 136 | dprint(FD_IO, "iolog: get %llu/%lu/%s\n", io_u->offset, |
| 137 | io_u->buflen, io_u->file->file_name); |
| 138 | if (ipo->delay) |
| 139 | iolog_delay(td, ipo->delay); |
| 140 | } else { |
| 141 | elapsed = mtime_since_genesis(); |
| 142 | if (ipo->delay > elapsed) |
| 143 | usec_sleep(td, (ipo->delay - elapsed) * 1000); |
| 144 | |
| 145 | } |
| 146 | |
| 147 | free(ipo); |
| 148 | |
| 149 | if (io_u->ddir != DDIR_WAIT) |
| 150 | return 0; |
| 151 | } |
| 152 | |
| 153 | td->done = 1; |
| 154 | return 1; |
| 155 | } |
| 156 | |
| 157 | void prune_io_piece_log(struct thread_data *td) |
| 158 | { |
| 159 | struct io_piece *ipo; |
| 160 | struct rb_node *n; |
| 161 | |
| 162 | while ((n = rb_first(&td->io_hist_tree)) != NULL) { |
| 163 | ipo = rb_entry(n, struct io_piece, rb_node); |
| 164 | rb_erase(n, &td->io_hist_tree); |
| 165 | remove_trim_entry(td, ipo); |
| 166 | td->io_hist_len--; |
| 167 | free(ipo); |
| 168 | } |
| 169 | |
| 170 | while (!flist_empty(&td->io_hist_list)) { |
| 171 | ipo = flist_entry(td->io_hist_list.next, struct io_piece, list); |
| 172 | flist_del(&ipo->list); |
| 173 | remove_trim_entry(td, ipo); |
| 174 | td->io_hist_len--; |
| 175 | free(ipo); |
| 176 | } |
| 177 | } |
| 178 | |
Jens Axboe | 86f4028 | 2012-02-10 21:58:46 +0100 | [diff] [blame] | 179 | static void dump_ipo(struct io_piece *ipo, const char *msg) |
| 180 | { |
| 181 | printf("%s\n", msg); |
| 182 | printf("\toffset %llu, len %lu, fl %x, ddir %d\n", ipo->offset, ipo->len, ipo->flags, ipo->ddir); |
| 183 | } |
| 184 | |
Jens Axboe | ac9b910 | 2011-10-01 15:01:08 -0400 | [diff] [blame] | 185 | /* |
| 186 | * log a successful write, so we can unwind the log for verify |
| 187 | */ |
| 188 | void log_io_piece(struct thread_data *td, struct io_u *io_u) |
| 189 | { |
| 190 | struct rb_node **p, *parent; |
| 191 | struct io_piece *ipo, *__ipo; |
| 192 | |
| 193 | ipo = malloc(sizeof(struct io_piece)); |
| 194 | init_ipo(ipo); |
| 195 | ipo->file = io_u->file; |
| 196 | ipo->offset = io_u->offset; |
| 197 | ipo->len = io_u->buflen; |
| 198 | |
| 199 | if (io_u_should_trim(td, io_u)) { |
| 200 | flist_add_tail(&ipo->trim_list, &td->trim_list); |
| 201 | td->trim_entries++; |
| 202 | } |
| 203 | |
| 204 | /* |
| 205 | * We don't need to sort the entries, if: |
| 206 | * |
| 207 | * Sequential writes, or |
| 208 | * Random writes that lay out the file as it goes along |
| 209 | * |
| 210 | * For both these cases, just reading back data in the order we |
| 211 | * wrote it out is the fastest. |
| 212 | * |
| 213 | * One exception is if we don't have a random map AND we are doing |
| 214 | * verifies, in that case we need to check for duplicate blocks and |
| 215 | * drop the old one, which we rely on the rb insert/lookup for |
| 216 | * handling. |
| 217 | */ |
| 218 | if ((!td_random(td) || !td->o.overwrite) && |
| 219 | (file_randommap(td, ipo->file) || td->o.verify == VERIFY_NONE)) { |
| 220 | INIT_FLIST_HEAD(&ipo->list); |
| 221 | flist_add_tail(&ipo->list, &td->io_hist_list); |
| 222 | ipo->flags |= IP_F_ONLIST; |
| 223 | td->io_hist_len++; |
| 224 | return; |
| 225 | } |
| 226 | |
| 227 | RB_CLEAR_NODE(&ipo->rb_node); |
| 228 | |
| 229 | /* |
| 230 | * Sort the entry into the verification list |
| 231 | */ |
| 232 | restart: |
| 233 | p = &td->io_hist_tree.rb_node; |
| 234 | parent = NULL; |
| 235 | while (*p) { |
| 236 | parent = *p; |
| 237 | |
| 238 | __ipo = rb_entry(parent, struct io_piece, rb_node); |
| 239 | if (ipo->file < __ipo->file) |
| 240 | p = &(*p)->rb_left; |
| 241 | else if (ipo->file > __ipo->file) |
| 242 | p = &(*p)->rb_right; |
| 243 | else if (ipo->offset < __ipo->offset) |
| 244 | p = &(*p)->rb_left; |
| 245 | else if (ipo->offset > __ipo->offset) |
| 246 | p = &(*p)->rb_right; |
| 247 | else { |
Jens Axboe | 86f4028 | 2012-02-10 21:58:46 +0100 | [diff] [blame] | 248 | if (ipo->len != __ipo->len) { |
| 249 | log_err("fio: unexpected ipo overlap!\n"); |
| 250 | log_err("fio: please report this issue.\n"); |
| 251 | dump_ipo(ipo, "ipo"); |
| 252 | dump_ipo(__ipo, "__ipo"); |
| 253 | } |
Jens Axboe | ac9b910 | 2011-10-01 15:01:08 -0400 | [diff] [blame] | 254 | td->io_hist_len--; |
| 255 | rb_erase(parent, &td->io_hist_tree); |
| 256 | remove_trim_entry(td, __ipo); |
| 257 | free(__ipo); |
| 258 | goto restart; |
| 259 | } |
| 260 | } |
| 261 | |
| 262 | rb_link_node(&ipo->rb_node, parent, p); |
| 263 | rb_insert_color(&ipo->rb_node, &td->io_hist_tree); |
| 264 | ipo->flags |= IP_F_ONRB; |
| 265 | td->io_hist_len++; |
| 266 | } |
| 267 | |
| 268 | void write_iolog_close(struct thread_data *td) |
| 269 | { |
| 270 | fflush(td->iolog_f); |
| 271 | fclose(td->iolog_f); |
| 272 | free(td->iolog_buf); |
| 273 | td->iolog_f = NULL; |
| 274 | td->iolog_buf = NULL; |
| 275 | } |
| 276 | |
| 277 | /* |
| 278 | * Read version 2 iolog data. It is enhanced to include per-file logging, |
| 279 | * syncs, etc. |
| 280 | */ |
| 281 | static int read_iolog2(struct thread_data *td, FILE *f) |
| 282 | { |
| 283 | unsigned long long offset; |
| 284 | unsigned int bytes; |
| 285 | int reads, writes, waits, fileno = 0, file_action = 0; /* stupid gcc */ |
| 286 | char *fname, *act; |
| 287 | char *str, *p; |
| 288 | enum fio_ddir rw; |
| 289 | |
| 290 | free_release_files(td); |
| 291 | |
| 292 | /* |
| 293 | * Read in the read iolog and store it, reuse the infrastructure |
| 294 | * for doing verifications. |
| 295 | */ |
| 296 | str = malloc(4096); |
| 297 | fname = malloc(256+16); |
| 298 | act = malloc(256+16); |
| 299 | |
| 300 | reads = writes = waits = 0; |
| 301 | while ((p = fgets(str, 4096, f)) != NULL) { |
| 302 | struct io_piece *ipo; |
| 303 | int r; |
| 304 | |
| 305 | r = sscanf(p, "%256s %256s %llu %u", fname, act, &offset, |
| 306 | &bytes); |
| 307 | if (r == 4) { |
| 308 | /* |
| 309 | * Check action first |
| 310 | */ |
| 311 | if (!strcmp(act, "wait")) |
| 312 | rw = DDIR_WAIT; |
| 313 | else if (!strcmp(act, "read")) |
| 314 | rw = DDIR_READ; |
| 315 | else if (!strcmp(act, "write")) |
| 316 | rw = DDIR_WRITE; |
| 317 | else if (!strcmp(act, "sync")) |
| 318 | rw = DDIR_SYNC; |
| 319 | else if (!strcmp(act, "datasync")) |
| 320 | rw = DDIR_DATASYNC; |
| 321 | else if (!strcmp(act, "trim")) |
| 322 | rw = DDIR_TRIM; |
| 323 | else { |
| 324 | log_err("fio: bad iolog file action: %s\n", |
| 325 | act); |
| 326 | continue; |
| 327 | } |
| 328 | } else if (r == 2) { |
| 329 | rw = DDIR_INVAL; |
| 330 | if (!strcmp(act, "add")) { |
| 331 | td->o.nr_files++; |
| 332 | fileno = add_file(td, fname); |
| 333 | file_action = FIO_LOG_ADD_FILE; |
| 334 | continue; |
| 335 | } else if (!strcmp(act, "open")) { |
| 336 | fileno = get_fileno(td, fname); |
| 337 | file_action = FIO_LOG_OPEN_FILE; |
| 338 | } else if (!strcmp(act, "close")) { |
| 339 | fileno = get_fileno(td, fname); |
| 340 | file_action = FIO_LOG_CLOSE_FILE; |
| 341 | } else { |
| 342 | log_err("fio: bad iolog file action: %s\n", |
| 343 | act); |
| 344 | continue; |
| 345 | } |
| 346 | } else { |
| 347 | log_err("bad iolog2: %s", p); |
| 348 | continue; |
| 349 | } |
| 350 | |
| 351 | if (rw == DDIR_READ) |
| 352 | reads++; |
| 353 | else if (rw == DDIR_WRITE) { |
| 354 | /* |
| 355 | * Don't add a write for ro mode |
| 356 | */ |
| 357 | if (read_only) |
| 358 | continue; |
| 359 | writes++; |
| 360 | } else if (rw == DDIR_WAIT) { |
| 361 | waits++; |
| 362 | } else if (rw == DDIR_INVAL) { |
| 363 | } else if (!ddir_sync(rw)) { |
| 364 | log_err("bad ddir: %d\n", rw); |
| 365 | continue; |
| 366 | } |
| 367 | |
| 368 | /* |
| 369 | * Make note of file |
| 370 | */ |
| 371 | ipo = malloc(sizeof(*ipo)); |
| 372 | init_ipo(ipo); |
| 373 | ipo->ddir = rw; |
| 374 | if (rw == DDIR_WAIT) { |
| 375 | ipo->delay = offset; |
| 376 | } else { |
| 377 | ipo->offset = offset; |
| 378 | ipo->len = bytes; |
| 379 | if (bytes > td->o.max_bs[rw]) |
| 380 | td->o.max_bs[rw] = bytes; |
| 381 | ipo->fileno = fileno; |
| 382 | ipo->file_action = file_action; |
| 383 | } |
| 384 | |
| 385 | queue_io_piece(td, ipo); |
| 386 | } |
| 387 | |
| 388 | free(str); |
| 389 | free(act); |
| 390 | free(fname); |
| 391 | |
| 392 | if (writes && read_only) { |
| 393 | log_err("fio: <%s> skips replay of %d writes due to" |
| 394 | " read-only\n", td->o.name, writes); |
| 395 | writes = 0; |
| 396 | } |
| 397 | |
| 398 | if (!reads && !writes && !waits) |
| 399 | return 1; |
| 400 | else if (reads && !writes) |
| 401 | td->o.td_ddir = TD_DDIR_READ; |
| 402 | else if (!reads && writes) |
| 403 | td->o.td_ddir = TD_DDIR_WRITE; |
| 404 | else |
| 405 | td->o.td_ddir = TD_DDIR_RW; |
| 406 | |
| 407 | return 0; |
| 408 | } |
| 409 | |
| 410 | /* |
| 411 | * open iolog, check version, and call appropriate parser |
| 412 | */ |
| 413 | static int init_iolog_read(struct thread_data *td) |
| 414 | { |
| 415 | char buffer[256], *p; |
| 416 | FILE *f; |
| 417 | int ret; |
| 418 | |
| 419 | f = fopen(td->o.read_iolog_file, "r"); |
| 420 | if (!f) { |
| 421 | perror("fopen read iolog"); |
| 422 | return 1; |
| 423 | } |
| 424 | |
| 425 | p = fgets(buffer, sizeof(buffer), f); |
| 426 | if (!p) { |
| 427 | td_verror(td, errno, "iolog read"); |
| 428 | log_err("fio: unable to read iolog\n"); |
| 429 | fclose(f); |
| 430 | return 1; |
| 431 | } |
| 432 | |
| 433 | /* |
| 434 | * version 2 of the iolog stores a specific string as the |
| 435 | * first line, check for that |
| 436 | */ |
| 437 | if (!strncmp(iolog_ver2, buffer, strlen(iolog_ver2))) |
| 438 | ret = read_iolog2(td, f); |
| 439 | else { |
| 440 | log_err("fio: iolog version 1 is no longer supported\n"); |
| 441 | ret = 1; |
| 442 | } |
| 443 | |
| 444 | fclose(f); |
| 445 | return ret; |
| 446 | } |
| 447 | |
| 448 | /* |
| 449 | * Set up a log for storing io patterns. |
| 450 | */ |
| 451 | static int init_iolog_write(struct thread_data *td) |
| 452 | { |
| 453 | struct fio_file *ff; |
| 454 | FILE *f; |
| 455 | unsigned int i; |
| 456 | |
| 457 | f = fopen(td->o.write_iolog_file, "a"); |
| 458 | if (!f) { |
| 459 | perror("fopen write iolog"); |
| 460 | return 1; |
| 461 | } |
| 462 | |
| 463 | /* |
| 464 | * That's it for writing, setup a log buffer and we're done. |
| 465 | */ |
| 466 | td->iolog_f = f; |
| 467 | td->iolog_buf = malloc(8192); |
| 468 | setvbuf(f, td->iolog_buf, _IOFBF, 8192); |
| 469 | |
| 470 | /* |
| 471 | * write our version line |
| 472 | */ |
| 473 | if (fprintf(f, "%s\n", iolog_ver2) < 0) { |
| 474 | perror("iolog init\n"); |
| 475 | return 1; |
| 476 | } |
| 477 | |
| 478 | /* |
| 479 | * add all known files |
| 480 | */ |
| 481 | for_each_file(td, ff, i) |
| 482 | log_file(td, ff, FIO_LOG_ADD_FILE); |
| 483 | |
| 484 | return 0; |
| 485 | } |
| 486 | |
| 487 | int init_iolog(struct thread_data *td) |
| 488 | { |
| 489 | int ret = 0; |
| 490 | |
| 491 | if (td->o.read_iolog_file) { |
| 492 | /* |
| 493 | * Check if it's a blktrace file and load that if possible. |
| 494 | * Otherwise assume it's a normal log file and load that. |
| 495 | */ |
| 496 | if (is_blktrace(td->o.read_iolog_file)) |
| 497 | ret = load_blktrace(td, td->o.read_iolog_file); |
| 498 | else |
| 499 | ret = init_iolog_read(td); |
| 500 | } else if (td->o.write_iolog_file) |
| 501 | ret = init_iolog_write(td); |
| 502 | |
| 503 | return ret; |
| 504 | } |
| 505 | |
Jens Axboe | b8bc8cb | 2011-12-01 09:04:31 +0100 | [diff] [blame] | 506 | void setup_log(struct io_log **log, unsigned long avg_msec) |
Jens Axboe | ac9b910 | 2011-10-01 15:01:08 -0400 | [diff] [blame] | 507 | { |
| 508 | struct io_log *l = malloc(sizeof(*l)); |
| 509 | |
Jens Axboe | b8bc8cb | 2011-12-01 09:04:31 +0100 | [diff] [blame] | 510 | memset(l, 0, sizeof(*l)); |
Jens Axboe | ac9b910 | 2011-10-01 15:01:08 -0400 | [diff] [blame] | 511 | l->nr_samples = 0; |
| 512 | l->max_samples = 1024; |
| 513 | l->log = malloc(l->max_samples * sizeof(struct io_sample)); |
Jens Axboe | b8bc8cb | 2011-12-01 09:04:31 +0100 | [diff] [blame] | 514 | l->avg_msec = avg_msec; |
Jens Axboe | ac9b910 | 2011-10-01 15:01:08 -0400 | [diff] [blame] | 515 | *log = l; |
| 516 | } |
| 517 | |
| 518 | void __finish_log(struct io_log *log, const char *name) |
| 519 | { |
| 520 | unsigned int i; |
| 521 | FILE *f; |
| 522 | |
| 523 | f = fopen(name, "a"); |
| 524 | if (!f) { |
| 525 | perror("fopen log"); |
| 526 | return; |
| 527 | } |
| 528 | |
| 529 | for (i = 0; i < log->nr_samples; i++) { |
| 530 | fprintf(f, "%lu, %lu, %u, %u\n", log->log[i].time, |
| 531 | log->log[i].val, |
| 532 | log->log[i].ddir, |
| 533 | log->log[i].bs); |
| 534 | } |
| 535 | |
| 536 | fclose(f); |
| 537 | free(log->log); |
| 538 | free(log); |
| 539 | } |
| 540 | |
| 541 | void finish_log_named(struct thread_data *td, struct io_log *log, |
| 542 | const char *prefix, const char *postfix) |
| 543 | { |
| 544 | char file_name[256], *p; |
| 545 | |
| 546 | snprintf(file_name, 200, "%s_%s.log", prefix, postfix); |
| 547 | p = basename(file_name); |
| 548 | __finish_log(log, p); |
| 549 | } |
| 550 | |
| 551 | void finish_log(struct thread_data *td, struct io_log *log, const char *name) |
| 552 | { |
| 553 | finish_log_named(td, log, td->o.name, name); |
| 554 | } |