| 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 | #ifdef CONFIG_ZLIB |
| 10 | #include <zlib.h> |
| 11 | #endif |
| 12 | |
| 13 | #include "flist.h" |
| 14 | #include "fio.h" |
| 15 | #include "verify.h" |
| 16 | #include "trim.h" |
| 17 | #include "filelock.h" |
| 18 | #include "tp.h" |
| 19 | |
| 20 | static const char iolog_ver2[] = "fio version 2 iolog"; |
| 21 | |
| 22 | #ifdef CONFIG_ZLIB |
| 23 | |
| 24 | struct iolog_compress { |
| 25 | struct flist_head list; |
| 26 | void *buf; |
| 27 | size_t len; |
| 28 | unsigned int seq; |
| 29 | }; |
| 30 | |
| 31 | #define GZ_CHUNK 131072 |
| 32 | |
| 33 | static struct iolog_compress *get_new_chunk(unsigned int seq) |
| 34 | { |
| 35 | struct iolog_compress *c; |
| 36 | |
| 37 | c = malloc(sizeof(*c)); |
| 38 | INIT_FLIST_HEAD(&c->list); |
| 39 | c->buf = malloc(GZ_CHUNK); |
| 40 | c->len = 0; |
| 41 | c->seq = seq; |
| 42 | return c; |
| 43 | } |
| 44 | |
| 45 | static void free_chunk(struct iolog_compress *ic) |
| 46 | { |
| 47 | free(ic->buf); |
| 48 | free(ic); |
| 49 | } |
| 50 | |
| 51 | #endif |
| 52 | |
| 53 | void queue_io_piece(struct thread_data *td, struct io_piece *ipo) |
| 54 | { |
| 55 | flist_add_tail(&ipo->list, &td->io_log_list); |
| 56 | td->total_io_size += ipo->len; |
| 57 | } |
| 58 | |
| 59 | void log_io_u(struct thread_data *td, struct io_u *io_u) |
| 60 | { |
| 61 | const char *act[] = { "read", "write", "sync", "datasync", |
| 62 | "sync_file_range", "wait", "trim" }; |
| 63 | |
| 64 | assert(io_u->ddir <= 6); |
| 65 | |
| 66 | if (!td->o.write_iolog_file) |
| 67 | return; |
| 68 | |
| 69 | fprintf(td->iolog_f, "%s %s %llu %lu\n", io_u->file->file_name, |
| 70 | act[io_u->ddir], io_u->offset, |
| 71 | io_u->buflen); |
| 72 | } |
| 73 | |
| 74 | void log_file(struct thread_data *td, struct fio_file *f, |
| 75 | enum file_log_act what) |
| 76 | { |
| 77 | const char *act[] = { "add", "open", "close" }; |
| 78 | |
| 79 | assert(what < 3); |
| 80 | |
| 81 | if (!td->o.write_iolog_file) |
| 82 | return; |
| 83 | |
| 84 | |
| 85 | /* |
| 86 | * this happens on the pre-open/close done before the job starts |
| 87 | */ |
| 88 | if (!td->iolog_f) |
| 89 | return; |
| 90 | |
| 91 | fprintf(td->iolog_f, "%s %s\n", f->file_name, act[what]); |
| 92 | } |
| 93 | |
| 94 | static void iolog_delay(struct thread_data *td, unsigned long delay) |
| 95 | { |
| 96 | unsigned long usec = utime_since_now(&td->last_issue); |
| 97 | unsigned long this_delay; |
| 98 | |
| 99 | if (delay < usec) |
| 100 | return; |
| 101 | |
| 102 | delay -= usec; |
| 103 | |
| 104 | /* |
| 105 | * less than 100 usec delay, just regard it as noise |
| 106 | */ |
| 107 | if (delay < 100) |
| 108 | return; |
| 109 | |
| 110 | while (delay && !td->terminate) { |
| 111 | this_delay = delay; |
| 112 | if (this_delay > 500000) |
| 113 | this_delay = 500000; |
| 114 | |
| 115 | usec_sleep(td, this_delay); |
| 116 | delay -= this_delay; |
| 117 | } |
| 118 | } |
| 119 | |
| 120 | static int ipo_special(struct thread_data *td, struct io_piece *ipo) |
| 121 | { |
| 122 | struct fio_file *f; |
| 123 | int ret; |
| 124 | |
| 125 | /* |
| 126 | * Not a special ipo |
| 127 | */ |
| 128 | if (ipo->ddir != DDIR_INVAL) |
| 129 | return 0; |
| 130 | |
| 131 | f = td->files[ipo->fileno]; |
| 132 | |
| 133 | switch (ipo->file_action) { |
| 134 | case FIO_LOG_OPEN_FILE: |
| 135 | ret = td_io_open_file(td, f); |
| 136 | if (!ret) |
| 137 | break; |
| 138 | td_verror(td, ret, "iolog open file"); |
| 139 | return -1; |
| 140 | case FIO_LOG_CLOSE_FILE: |
| 141 | td_io_close_file(td, f); |
| 142 | break; |
| 143 | case FIO_LOG_UNLINK_FILE: |
| 144 | unlink(f->file_name); |
| 145 | break; |
| 146 | default: |
| 147 | log_err("fio: bad file action %d\n", ipo->file_action); |
| 148 | break; |
| 149 | } |
| 150 | |
| 151 | return 1; |
| 152 | } |
| 153 | |
| 154 | int read_iolog_get(struct thread_data *td, struct io_u *io_u) |
| 155 | { |
| 156 | struct io_piece *ipo; |
| 157 | unsigned long elapsed; |
| 158 | |
| 159 | while (!flist_empty(&td->io_log_list)) { |
| 160 | int ret; |
| 161 | |
| 162 | ipo = flist_entry(td->io_log_list.next, struct io_piece, list); |
| 163 | flist_del(&ipo->list); |
| 164 | remove_trim_entry(td, ipo); |
| 165 | |
| 166 | ret = ipo_special(td, ipo); |
| 167 | if (ret < 0) { |
| 168 | free(ipo); |
| 169 | break; |
| 170 | } else if (ret > 0) { |
| 171 | free(ipo); |
| 172 | continue; |
| 173 | } |
| 174 | |
| 175 | io_u->ddir = ipo->ddir; |
| 176 | if (ipo->ddir != DDIR_WAIT) { |
| 177 | io_u->offset = ipo->offset; |
| 178 | io_u->buflen = ipo->len; |
| 179 | io_u->file = td->files[ipo->fileno]; |
| 180 | get_file(io_u->file); |
| 181 | dprint(FD_IO, "iolog: get %llu/%lu/%s\n", io_u->offset, |
| 182 | io_u->buflen, io_u->file->file_name); |
| 183 | if (ipo->delay) |
| 184 | iolog_delay(td, ipo->delay); |
| 185 | } else { |
| 186 | elapsed = mtime_since_genesis(); |
| 187 | if (ipo->delay > elapsed) |
| 188 | usec_sleep(td, (ipo->delay - elapsed) * 1000); |
| 189 | } |
| 190 | |
| 191 | free(ipo); |
| 192 | |
| 193 | if (io_u->ddir != DDIR_WAIT) |
| 194 | return 0; |
| 195 | } |
| 196 | |
| 197 | td->done = 1; |
| 198 | return 1; |
| 199 | } |
| 200 | |
| 201 | void prune_io_piece_log(struct thread_data *td) |
| 202 | { |
| 203 | struct io_piece *ipo; |
| 204 | struct rb_node *n; |
| 205 | |
| 206 | while ((n = rb_first(&td->io_hist_tree)) != NULL) { |
| 207 | ipo = rb_entry(n, struct io_piece, rb_node); |
| 208 | rb_erase(n, &td->io_hist_tree); |
| 209 | remove_trim_entry(td, ipo); |
| 210 | td->io_hist_len--; |
| 211 | free(ipo); |
| 212 | } |
| 213 | |
| 214 | while (!flist_empty(&td->io_hist_list)) { |
| 215 | ipo = flist_entry(td->io_hist_list.next, struct io_piece, list); |
| 216 | flist_del(&ipo->list); |
| 217 | remove_trim_entry(td, ipo); |
| 218 | td->io_hist_len--; |
| 219 | free(ipo); |
| 220 | } |
| 221 | } |
| 222 | |
| 223 | /* |
| 224 | * log a successful write, so we can unwind the log for verify |
| 225 | */ |
| 226 | void log_io_piece(struct thread_data *td, struct io_u *io_u) |
| 227 | { |
| 228 | struct rb_node **p, *parent; |
| 229 | struct io_piece *ipo, *__ipo; |
| 230 | |
| 231 | ipo = malloc(sizeof(struct io_piece)); |
| 232 | init_ipo(ipo); |
| 233 | ipo->file = io_u->file; |
| 234 | ipo->offset = io_u->offset; |
| 235 | ipo->len = io_u->buflen; |
| 236 | ipo->numberio = io_u->numberio; |
| 237 | ipo->flags = IP_F_IN_FLIGHT; |
| 238 | |
| 239 | io_u->ipo = ipo; |
| 240 | |
| 241 | if (io_u_should_trim(td, io_u)) { |
| 242 | flist_add_tail(&ipo->trim_list, &td->trim_list); |
| 243 | td->trim_entries++; |
| 244 | } |
| 245 | |
| 246 | /* |
| 247 | * We don't need to sort the entries, if: |
| 248 | * |
| 249 | * Sequential writes, or |
| 250 | * Random writes that lay out the file as it goes along |
| 251 | * |
| 252 | * For both these cases, just reading back data in the order we |
| 253 | * wrote it out is the fastest. |
| 254 | * |
| 255 | * One exception is if we don't have a random map AND we are doing |
| 256 | * verifies, in that case we need to check for duplicate blocks and |
| 257 | * drop the old one, which we rely on the rb insert/lookup for |
| 258 | * handling. |
| 259 | */ |
| 260 | if (((!td->o.verifysort) || !td_random(td) || !td->o.overwrite) && |
| 261 | (file_randommap(td, ipo->file) || td->o.verify == VERIFY_NONE)) { |
| 262 | INIT_FLIST_HEAD(&ipo->list); |
| 263 | flist_add_tail(&ipo->list, &td->io_hist_list); |
| 264 | ipo->flags |= IP_F_ONLIST; |
| 265 | td->io_hist_len++; |
| 266 | return; |
| 267 | } |
| 268 | |
| 269 | RB_CLEAR_NODE(&ipo->rb_node); |
| 270 | |
| 271 | /* |
| 272 | * Sort the entry into the verification list |
| 273 | */ |
| 274 | restart: |
| 275 | p = &td->io_hist_tree.rb_node; |
| 276 | parent = NULL; |
| 277 | while (*p) { |
| 278 | parent = *p; |
| 279 | |
| 280 | __ipo = rb_entry(parent, struct io_piece, rb_node); |
| 281 | if (ipo->file < __ipo->file) |
| 282 | p = &(*p)->rb_left; |
| 283 | else if (ipo->file > __ipo->file) |
| 284 | p = &(*p)->rb_right; |
| 285 | else if (ipo->offset < __ipo->offset) |
| 286 | p = &(*p)->rb_left; |
| 287 | else if (ipo->offset > __ipo->offset) |
| 288 | p = &(*p)->rb_right; |
| 289 | else { |
| 290 | dprint(FD_IO, "iolog: overlap %llu/%lu, %llu/%lu", |
| 291 | __ipo->offset, __ipo->len, |
| 292 | ipo->offset, ipo->len); |
| 293 | td->io_hist_len--; |
| 294 | rb_erase(parent, &td->io_hist_tree); |
| 295 | remove_trim_entry(td, __ipo); |
| 296 | free(__ipo); |
| 297 | goto restart; |
| 298 | } |
| 299 | } |
| 300 | |
| 301 | rb_link_node(&ipo->rb_node, parent, p); |
| 302 | rb_insert_color(&ipo->rb_node, &td->io_hist_tree); |
| 303 | ipo->flags |= IP_F_ONRB; |
| 304 | td->io_hist_len++; |
| 305 | } |
| 306 | |
| 307 | void unlog_io_piece(struct thread_data *td, struct io_u *io_u) |
| 308 | { |
| 309 | struct io_piece *ipo = io_u->ipo; |
| 310 | |
| 311 | if (!ipo) |
| 312 | return; |
| 313 | |
| 314 | if (ipo->flags & IP_F_ONRB) |
| 315 | rb_erase(&ipo->rb_node, &td->io_hist_tree); |
| 316 | else if (ipo->flags & IP_F_ONLIST) |
| 317 | flist_del(&ipo->list); |
| 318 | |
| 319 | free(ipo); |
| 320 | io_u->ipo = NULL; |
| 321 | td->io_hist_len--; |
| 322 | } |
| 323 | |
| 324 | void trim_io_piece(struct thread_data *td, struct io_u *io_u) |
| 325 | { |
| 326 | struct io_piece *ipo = io_u->ipo; |
| 327 | |
| 328 | if (!ipo) |
| 329 | return; |
| 330 | |
| 331 | ipo->len = io_u->xfer_buflen - io_u->resid; |
| 332 | } |
| 333 | |
| 334 | void write_iolog_close(struct thread_data *td) |
| 335 | { |
| 336 | fflush(td->iolog_f); |
| 337 | fclose(td->iolog_f); |
| 338 | free(td->iolog_buf); |
| 339 | td->iolog_f = NULL; |
| 340 | td->iolog_buf = NULL; |
| 341 | } |
| 342 | |
| 343 | /* |
| 344 | * Read version 2 iolog data. It is enhanced to include per-file logging, |
| 345 | * syncs, etc. |
| 346 | */ |
| 347 | static int read_iolog2(struct thread_data *td, FILE *f) |
| 348 | { |
| 349 | unsigned long long offset; |
| 350 | unsigned int bytes; |
| 351 | int reads, writes, waits, fileno = 0, file_action = 0; /* stupid gcc */ |
| 352 | char *fname, *act; |
| 353 | char *str, *p; |
| 354 | enum fio_ddir rw; |
| 355 | |
| 356 | free_release_files(td); |
| 357 | |
| 358 | /* |
| 359 | * Read in the read iolog and store it, reuse the infrastructure |
| 360 | * for doing verifications. |
| 361 | */ |
| 362 | str = malloc(4096); |
| 363 | fname = malloc(256+16); |
| 364 | act = malloc(256+16); |
| 365 | |
| 366 | reads = writes = waits = 0; |
| 367 | while ((p = fgets(str, 4096, f)) != NULL) { |
| 368 | struct io_piece *ipo; |
| 369 | int r; |
| 370 | |
| 371 | r = sscanf(p, "%256s %256s %llu %u", fname, act, &offset, |
| 372 | &bytes); |
| 373 | if (r == 4) { |
| 374 | /* |
| 375 | * Check action first |
| 376 | */ |
| 377 | if (!strcmp(act, "wait")) |
| 378 | rw = DDIR_WAIT; |
| 379 | else if (!strcmp(act, "read")) |
| 380 | rw = DDIR_READ; |
| 381 | else if (!strcmp(act, "write")) |
| 382 | rw = DDIR_WRITE; |
| 383 | else if (!strcmp(act, "sync")) |
| 384 | rw = DDIR_SYNC; |
| 385 | else if (!strcmp(act, "datasync")) |
| 386 | rw = DDIR_DATASYNC; |
| 387 | else if (!strcmp(act, "trim")) |
| 388 | rw = DDIR_TRIM; |
| 389 | else { |
| 390 | log_err("fio: bad iolog file action: %s\n", |
| 391 | act); |
| 392 | continue; |
| 393 | } |
| 394 | fileno = get_fileno(td, fname); |
| 395 | } else if (r == 2) { |
| 396 | rw = DDIR_INVAL; |
| 397 | if (!strcmp(act, "add")) { |
| 398 | fileno = add_file(td, fname, 0, 1); |
| 399 | file_action = FIO_LOG_ADD_FILE; |
| 400 | continue; |
| 401 | } else if (!strcmp(act, "open")) { |
| 402 | fileno = get_fileno(td, fname); |
| 403 | file_action = FIO_LOG_OPEN_FILE; |
| 404 | } else if (!strcmp(act, "close")) { |
| 405 | fileno = get_fileno(td, fname); |
| 406 | file_action = FIO_LOG_CLOSE_FILE; |
| 407 | } else { |
| 408 | log_err("fio: bad iolog file action: %s\n", |
| 409 | act); |
| 410 | continue; |
| 411 | } |
| 412 | } else { |
| 413 | log_err("bad iolog2: %s", p); |
| 414 | continue; |
| 415 | } |
| 416 | |
| 417 | if (rw == DDIR_READ) |
| 418 | reads++; |
| 419 | else if (rw == DDIR_WRITE) { |
| 420 | /* |
| 421 | * Don't add a write for ro mode |
| 422 | */ |
| 423 | if (read_only) |
| 424 | continue; |
| 425 | writes++; |
| 426 | } else if (rw == DDIR_WAIT) { |
| 427 | waits++; |
| 428 | } else if (rw == DDIR_INVAL) { |
| 429 | } else if (!ddir_sync(rw)) { |
| 430 | log_err("bad ddir: %d\n", rw); |
| 431 | continue; |
| 432 | } |
| 433 | |
| 434 | /* |
| 435 | * Make note of file |
| 436 | */ |
| 437 | ipo = malloc(sizeof(*ipo)); |
| 438 | init_ipo(ipo); |
| 439 | ipo->ddir = rw; |
| 440 | if (rw == DDIR_WAIT) { |
| 441 | ipo->delay = offset; |
| 442 | } else { |
| 443 | ipo->offset = offset; |
| 444 | ipo->len = bytes; |
| 445 | if (rw != DDIR_INVAL && bytes > td->o.max_bs[rw]) |
| 446 | td->o.max_bs[rw] = bytes; |
| 447 | ipo->fileno = fileno; |
| 448 | ipo->file_action = file_action; |
| 449 | td->o.size += bytes; |
| 450 | } |
| 451 | |
| 452 | queue_io_piece(td, ipo); |
| 453 | } |
| 454 | |
| 455 | free(str); |
| 456 | free(act); |
| 457 | free(fname); |
| 458 | |
| 459 | if (writes && read_only) { |
| 460 | log_err("fio: <%s> skips replay of %d writes due to" |
| 461 | " read-only\n", td->o.name, writes); |
| 462 | writes = 0; |
| 463 | } |
| 464 | |
| 465 | if (!reads && !writes && !waits) |
| 466 | return 1; |
| 467 | else if (reads && !writes) |
| 468 | td->o.td_ddir = TD_DDIR_READ; |
| 469 | else if (!reads && writes) |
| 470 | td->o.td_ddir = TD_DDIR_WRITE; |
| 471 | else |
| 472 | td->o.td_ddir = TD_DDIR_RW; |
| 473 | |
| 474 | return 0; |
| 475 | } |
| 476 | |
| 477 | /* |
| 478 | * open iolog, check version, and call appropriate parser |
| 479 | */ |
| 480 | static int init_iolog_read(struct thread_data *td) |
| 481 | { |
| 482 | char buffer[256], *p; |
| 483 | FILE *f; |
| 484 | int ret; |
| 485 | |
| 486 | f = fopen(td->o.read_iolog_file, "r"); |
| 487 | if (!f) { |
| 488 | perror("fopen read iolog"); |
| 489 | return 1; |
| 490 | } |
| 491 | |
| 492 | p = fgets(buffer, sizeof(buffer), f); |
| 493 | if (!p) { |
| 494 | td_verror(td, errno, "iolog read"); |
| 495 | log_err("fio: unable to read iolog\n"); |
| 496 | fclose(f); |
| 497 | return 1; |
| 498 | } |
| 499 | |
| 500 | /* |
| 501 | * version 2 of the iolog stores a specific string as the |
| 502 | * first line, check for that |
| 503 | */ |
| 504 | if (!strncmp(iolog_ver2, buffer, strlen(iolog_ver2))) |
| 505 | ret = read_iolog2(td, f); |
| 506 | else { |
| 507 | log_err("fio: iolog version 1 is no longer supported\n"); |
| 508 | ret = 1; |
| 509 | } |
| 510 | |
| 511 | fclose(f); |
| 512 | return ret; |
| 513 | } |
| 514 | |
| 515 | /* |
| 516 | * Set up a log for storing io patterns. |
| 517 | */ |
| 518 | static int init_iolog_write(struct thread_data *td) |
| 519 | { |
| 520 | struct fio_file *ff; |
| 521 | FILE *f; |
| 522 | unsigned int i; |
| 523 | |
| 524 | f = fopen(td->o.write_iolog_file, "a"); |
| 525 | if (!f) { |
| 526 | perror("fopen write iolog"); |
| 527 | return 1; |
| 528 | } |
| 529 | |
| 530 | /* |
| 531 | * That's it for writing, setup a log buffer and we're done. |
| 532 | */ |
| 533 | td->iolog_f = f; |
| 534 | td->iolog_buf = malloc(8192); |
| 535 | setvbuf(f, td->iolog_buf, _IOFBF, 8192); |
| 536 | |
| 537 | /* |
| 538 | * write our version line |
| 539 | */ |
| 540 | if (fprintf(f, "%s\n", iolog_ver2) < 0) { |
| 541 | perror("iolog init\n"); |
| 542 | return 1; |
| 543 | } |
| 544 | |
| 545 | /* |
| 546 | * add all known files |
| 547 | */ |
| 548 | for_each_file(td, ff, i) |
| 549 | log_file(td, ff, FIO_LOG_ADD_FILE); |
| 550 | |
| 551 | return 0; |
| 552 | } |
| 553 | |
| 554 | int init_iolog(struct thread_data *td) |
| 555 | { |
| 556 | int ret = 0; |
| 557 | |
| 558 | if (td->o.read_iolog_file) { |
| 559 | int need_swap; |
| 560 | |
| 561 | /* |
| 562 | * Check if it's a blktrace file and load that if possible. |
| 563 | * Otherwise assume it's a normal log file and load that. |
| 564 | */ |
| 565 | if (is_blktrace(td->o.read_iolog_file, &need_swap)) |
| 566 | ret = load_blktrace(td, td->o.read_iolog_file, need_swap); |
| 567 | else |
| 568 | ret = init_iolog_read(td); |
| 569 | } else if (td->o.write_iolog_file) |
| 570 | ret = init_iolog_write(td); |
| 571 | |
| 572 | if (ret) |
| 573 | td_verror(td, EINVAL, "failed initializing iolog"); |
| 574 | |
| 575 | return ret; |
| 576 | } |
| 577 | |
| 578 | void setup_log(struct io_log **log, struct log_params *p, |
| 579 | const char *filename) |
| 580 | { |
| 581 | struct io_log *l = malloc(sizeof(*l)); |
| 582 | |
| 583 | memset(l, 0, sizeof(*l)); |
| 584 | l->nr_samples = 0; |
| 585 | l->max_samples = 1024; |
| 586 | l->log_type = p->log_type; |
| 587 | l->log_offset = p->log_offset; |
| 588 | l->log_gz = p->log_gz; |
| 589 | l->log = malloc(l->max_samples * log_entry_sz(l)); |
| 590 | l->avg_msec = p->avg_msec; |
| 591 | l->filename = strdup(filename); |
| 592 | l->td = p->td; |
| 593 | |
| 594 | INIT_FLIST_HEAD(&l->chunk_list); |
| 595 | |
| 596 | if (l->log_gz && !p->td) |
| 597 | l->log_gz = 0; |
| 598 | else if (l->log_gz) { |
| 599 | pthread_mutex_init(&l->chunk_lock, NULL); |
| 600 | p->td->flags |= TD_F_COMPRESS_LOG; |
| 601 | } |
| 602 | |
| 603 | *log = l; |
| 604 | } |
| 605 | |
| 606 | #ifdef CONFIG_SETVBUF |
| 607 | static void *set_file_buffer(FILE *f) |
| 608 | { |
| 609 | size_t size = 1048576; |
| 610 | void *buf; |
| 611 | |
| 612 | buf = malloc(size); |
| 613 | setvbuf(f, buf, _IOFBF, size); |
| 614 | return buf; |
| 615 | } |
| 616 | |
| 617 | static void clear_file_buffer(void *buf) |
| 618 | { |
| 619 | free(buf); |
| 620 | } |
| 621 | #else |
| 622 | static void *set_file_buffer(FILE *f) |
| 623 | { |
| 624 | return NULL; |
| 625 | } |
| 626 | |
| 627 | static void clear_file_buffer(void *buf) |
| 628 | { |
| 629 | } |
| 630 | #endif |
| 631 | |
| 632 | void free_log(struct io_log *log) |
| 633 | { |
| 634 | free(log->log); |
| 635 | free(log->filename); |
| 636 | free(log); |
| 637 | } |
| 638 | |
| 639 | static void flush_samples(FILE *f, void *samples, uint64_t nr_samples, |
| 640 | int log_offset) |
| 641 | { |
| 642 | uint64_t i; |
| 643 | |
| 644 | for (i = 0; i < nr_samples; i++) { |
| 645 | struct io_sample *s = __get_sample(samples, log_offset, i); |
| 646 | |
| 647 | if (!log_offset) { |
| 648 | fprintf(f, "%lu, %lu, %u, %u\n", |
| 649 | (unsigned long) s->time, |
| 650 | (unsigned long) s->val, |
| 651 | s->ddir, s->bs); |
| 652 | } else { |
| 653 | struct io_sample_offset *so = (void *) s; |
| 654 | |
| 655 | fprintf(f, "%lu, %lu, %u, %u, %llu\n", |
| 656 | (unsigned long) s->time, |
| 657 | (unsigned long) s->val, |
| 658 | s->ddir, s->bs, |
| 659 | (unsigned long long) so->offset); |
| 660 | } |
| 661 | } |
| 662 | } |
| 663 | |
| 664 | #ifdef CONFIG_ZLIB |
| 665 | static int z_stream_init(z_stream *stream) |
| 666 | { |
| 667 | stream->zalloc = Z_NULL; |
| 668 | stream->zfree = Z_NULL; |
| 669 | stream->opaque = Z_NULL; |
| 670 | stream->next_in = Z_NULL; |
| 671 | |
| 672 | if (inflateInit(stream) != Z_OK) |
| 673 | return 1; |
| 674 | |
| 675 | return 0; |
| 676 | } |
| 677 | |
| 678 | struct flush_chunk_iter { |
| 679 | unsigned int seq; |
| 680 | void *buf; |
| 681 | size_t buf_size; |
| 682 | size_t buf_used; |
| 683 | size_t chunk_sz; |
| 684 | }; |
| 685 | |
| 686 | static void finish_chunk(z_stream *stream, int log_offset, FILE *f, |
| 687 | struct flush_chunk_iter *iter) |
| 688 | { |
| 689 | uint64_t nr_samples; |
| 690 | int ret; |
| 691 | |
| 692 | ret = inflateEnd(stream); |
| 693 | if (ret != Z_OK) |
| 694 | log_err("fio: failed to end log inflation (%d)\n", ret); |
| 695 | |
| 696 | nr_samples = iter->buf_used / __log_entry_sz(log_offset); |
| 697 | flush_samples(f, iter->buf, nr_samples, log_offset); |
| 698 | free(iter->buf); |
| 699 | iter->buf = NULL; |
| 700 | iter->buf_size = iter->buf_used = 0; |
| 701 | } |
| 702 | |
| 703 | static int flush_chunk(struct iolog_compress *ic, int log_offset, FILE *f, |
| 704 | z_stream *stream, struct flush_chunk_iter *iter) |
| 705 | { |
| 706 | if (ic->seq != iter->seq) { |
| 707 | if (iter->seq) |
| 708 | finish_chunk(stream, log_offset, f, iter); |
| 709 | |
| 710 | z_stream_init(stream); |
| 711 | iter->seq = ic->seq; |
| 712 | } |
| 713 | |
| 714 | stream->avail_in = ic->len; |
| 715 | stream->next_in = ic->buf; |
| 716 | |
| 717 | if (!iter->buf_size) { |
| 718 | iter->buf_size = iter->chunk_sz; |
| 719 | iter->buf = malloc(iter->buf_size); |
| 720 | } |
| 721 | |
| 722 | while (stream->avail_in) { |
| 723 | int err; |
| 724 | |
| 725 | stream->avail_out = iter->buf_size - iter->buf_used; |
| 726 | stream->next_out = iter->buf + iter->buf_used; |
| 727 | |
| 728 | err = inflate(stream, Z_NO_FLUSH); |
| 729 | if (err < 0) { |
| 730 | log_err("fio: failed inflating log: %d\n", err); |
| 731 | break; |
| 732 | } |
| 733 | |
| 734 | iter->buf_used += iter->buf_size - iter->buf_used - stream->avail_out; |
| 735 | } |
| 736 | |
| 737 | free_chunk(ic); |
| 738 | return 0; |
| 739 | } |
| 740 | |
| 741 | static void flush_gz_chunks(struct io_log *log, FILE *f) |
| 742 | { |
| 743 | struct flush_chunk_iter iter = { .chunk_sz = log->log_gz, }; |
| 744 | struct flist_head *node; |
| 745 | z_stream stream; |
| 746 | |
| 747 | while (!flist_empty(&log->chunk_list)) { |
| 748 | struct iolog_compress *ic; |
| 749 | |
| 750 | node = log->chunk_list.next; |
| 751 | ic = flist_entry(node, struct iolog_compress, list); |
| 752 | flist_del(&ic->list); |
| 753 | flush_chunk(ic, log->log_offset, f, &stream, &iter); |
| 754 | } |
| 755 | |
| 756 | if (iter.seq) { |
| 757 | finish_chunk(&stream, log->log_offset, f, &iter); |
| 758 | free(iter.buf); |
| 759 | } |
| 760 | } |
| 761 | |
| 762 | #else |
| 763 | |
| 764 | static void flush_gz_chunks(struct io_log *log, FILE *f) |
| 765 | { |
| 766 | } |
| 767 | |
| 768 | #endif |
| 769 | |
| 770 | void flush_log(struct io_log *log) |
| 771 | { |
| 772 | void *buf; |
| 773 | FILE *f; |
| 774 | |
| 775 | f = fopen(log->filename, "w"); |
| 776 | if (!f) { |
| 777 | perror("fopen log"); |
| 778 | return; |
| 779 | } |
| 780 | |
| 781 | buf = set_file_buffer(f); |
| 782 | |
| 783 | flush_gz_chunks(log, f); |
| 784 | |
| 785 | flush_samples(f, log->log, log->nr_samples, log->log_offset); |
| 786 | |
| 787 | fclose(f); |
| 788 | clear_file_buffer(buf); |
| 789 | } |
| 790 | |
| 791 | static int finish_log(struct thread_data *td, struct io_log *log, int trylock) |
| 792 | { |
| 793 | if (td->tp_data) |
| 794 | iolog_flush(log, 1); |
| 795 | |
| 796 | if (trylock) { |
| 797 | if (fio_trylock_file(log->filename)) |
| 798 | return 1; |
| 799 | } else |
| 800 | fio_lock_file(log->filename); |
| 801 | |
| 802 | if (td->client_type == FIO_CLIENT_TYPE_GUI) |
| 803 | fio_send_iolog(td, log, log->filename); |
| 804 | else |
| 805 | flush_log(log); |
| 806 | |
| 807 | fio_unlock_file(log->filename); |
| 808 | free_log(log); |
| 809 | return 0; |
| 810 | } |
| 811 | |
| 812 | #ifdef CONFIG_ZLIB |
| 813 | |
| 814 | struct iolog_flush_data { |
| 815 | struct tp_work work; |
| 816 | struct io_log *log; |
| 817 | void *samples; |
| 818 | uint64_t nr_samples; |
| 819 | }; |
| 820 | |
| 821 | static int gz_work(struct tp_work *work) |
| 822 | { |
| 823 | struct iolog_flush_data *data; |
| 824 | struct iolog_compress *c; |
| 825 | struct flist_head list; |
| 826 | unsigned int seq; |
| 827 | z_stream stream; |
| 828 | size_t total = 0; |
| 829 | int ret; |
| 830 | |
| 831 | INIT_FLIST_HEAD(&list); |
| 832 | |
| 833 | data = container_of(work, struct iolog_flush_data, work); |
| 834 | |
| 835 | stream.zalloc = Z_NULL; |
| 836 | stream.zfree = Z_NULL; |
| 837 | stream.opaque = Z_NULL; |
| 838 | |
| 839 | if (deflateInit(&stream, Z_DEFAULT_COMPRESSION) != Z_OK) { |
| 840 | log_err("fio: failed to init gz stream\n"); |
| 841 | return 0; |
| 842 | } |
| 843 | |
| 844 | seq = ++data->log->chunk_seq; |
| 845 | stream.next_in = (void *) data->samples; |
| 846 | stream.avail_in = data->nr_samples * log_entry_sz(data->log); |
| 847 | |
| 848 | do { |
| 849 | c = get_new_chunk(seq); |
| 850 | stream.avail_out = GZ_CHUNK; |
| 851 | stream.next_out = c->buf; |
| 852 | ret = deflate(&stream, Z_NO_FLUSH); |
| 853 | if (ret < 0) { |
| 854 | log_err("fio: deflate log (%d)\n", ret); |
| 855 | break; |
| 856 | } |
| 857 | |
| 858 | c->len = GZ_CHUNK - stream.avail_out; |
| 859 | flist_add_tail(&c->list, &list); |
| 860 | total += c->len; |
| 861 | } while (stream.avail_in); |
| 862 | |
| 863 | stream.next_out = c->buf + c->len; |
| 864 | stream.avail_out = GZ_CHUNK - c->len; |
| 865 | |
| 866 | ret = deflate(&stream, Z_FINISH); |
| 867 | if (ret == Z_STREAM_END) |
| 868 | c->len = GZ_CHUNK - stream.avail_out; |
| 869 | else { |
| 870 | do { |
| 871 | c = get_new_chunk(seq); |
| 872 | stream.avail_out = GZ_CHUNK; |
| 873 | stream.next_out = c->buf; |
| 874 | ret = deflate(&stream, Z_FINISH); |
| 875 | c->len = GZ_CHUNK - stream.avail_out; |
| 876 | flist_add_tail(&c->list, &list); |
| 877 | } while (ret != Z_STREAM_END); |
| 878 | } |
| 879 | |
| 880 | ret = deflateEnd(&stream); |
| 881 | if (ret != Z_OK) |
| 882 | log_err("fio: deflateEnd %d\n", ret); |
| 883 | |
| 884 | free(data->samples); |
| 885 | |
| 886 | if (!flist_empty(&list)) { |
| 887 | pthread_mutex_lock(&data->log->chunk_lock); |
| 888 | flist_splice_tail(&list, &data->log->chunk_list); |
| 889 | pthread_mutex_unlock(&data->log->chunk_lock); |
| 890 | } |
| 891 | |
| 892 | if (work->wait) { |
| 893 | work->done = 1; |
| 894 | pthread_cond_signal(&work->cv); |
| 895 | } else |
| 896 | free(data); |
| 897 | |
| 898 | return 0; |
| 899 | } |
| 900 | |
| 901 | int iolog_flush(struct io_log *log, int wait) |
| 902 | { |
| 903 | struct thread_data *td = log->td; |
| 904 | struct iolog_flush_data *data; |
| 905 | size_t sample_size; |
| 906 | |
| 907 | data = malloc(sizeof(*data)); |
| 908 | if (!data) |
| 909 | return 1; |
| 910 | |
| 911 | data->log = log; |
| 912 | |
| 913 | sample_size = log->nr_samples * log_entry_sz(log); |
| 914 | data->samples = malloc(sample_size); |
| 915 | if (!data->samples) { |
| 916 | free(data); |
| 917 | return 1; |
| 918 | } |
| 919 | |
| 920 | memcpy(data->samples, log->log, sample_size); |
| 921 | data->nr_samples = log->nr_samples; |
| 922 | data->work.fn = gz_work; |
| 923 | log->nr_samples = 0; |
| 924 | |
| 925 | if (wait) { |
| 926 | pthread_mutex_init(&data->work.lock, NULL); |
| 927 | pthread_cond_init(&data->work.cv, NULL); |
| 928 | data->work.wait = 1; |
| 929 | } else |
| 930 | data->work.wait = 0; |
| 931 | |
| 932 | tp_queue_work(td->tp_data, &data->work); |
| 933 | |
| 934 | if (wait) { |
| 935 | pthread_mutex_lock(&data->work.lock); |
| 936 | while (!data->work.done) |
| 937 | pthread_cond_wait(&data->work.cv, &data->work.lock); |
| 938 | pthread_mutex_unlock(&data->work.lock); |
| 939 | free(data); |
| 940 | } |
| 941 | |
| 942 | return 0; |
| 943 | } |
| 944 | |
| 945 | #else |
| 946 | |
| 947 | int iolog_flush(struct io_log *log, int wait) |
| 948 | { |
| 949 | return 1; |
| 950 | } |
| 951 | |
| 952 | #endif |
| 953 | |
| 954 | static int write_iops_log(struct thread_data *td, int try) |
| 955 | { |
| 956 | struct io_log *log = td->iops_log; |
| 957 | |
| 958 | if (!log) |
| 959 | return 0; |
| 960 | |
| 961 | return finish_log(td, log, try); |
| 962 | } |
| 963 | |
| 964 | static int write_slat_log(struct thread_data *td, int try) |
| 965 | { |
| 966 | struct io_log *log = td->slat_log; |
| 967 | |
| 968 | if (!log) |
| 969 | return 0; |
| 970 | |
| 971 | return finish_log(td, log, try); |
| 972 | } |
| 973 | |
| 974 | static int write_clat_log(struct thread_data *td, int try) |
| 975 | { |
| 976 | struct io_log *log = td->clat_log; |
| 977 | |
| 978 | if (!log) |
| 979 | return 0; |
| 980 | |
| 981 | return finish_log(td, log, try); |
| 982 | } |
| 983 | |
| 984 | static int write_lat_log(struct thread_data *td, int try) |
| 985 | { |
| 986 | struct io_log *log = td->lat_log; |
| 987 | |
| 988 | if (!log) |
| 989 | return 0; |
| 990 | |
| 991 | return finish_log(td, log, try); |
| 992 | } |
| 993 | |
| 994 | static int write_bandw_log(struct thread_data *td, int try) |
| 995 | { |
| 996 | struct io_log *log = td->bw_log; |
| 997 | |
| 998 | if (!log) |
| 999 | return 0; |
| 1000 | |
| 1001 | return finish_log(td, log, try); |
| 1002 | } |
| 1003 | |
| 1004 | enum { |
| 1005 | BW_LOG_MASK = 1, |
| 1006 | LAT_LOG_MASK = 2, |
| 1007 | SLAT_LOG_MASK = 4, |
| 1008 | CLAT_LOG_MASK = 8, |
| 1009 | IOPS_LOG_MASK = 16, |
| 1010 | |
| 1011 | ALL_LOG_NR = 5, |
| 1012 | }; |
| 1013 | |
| 1014 | struct log_type { |
| 1015 | unsigned int mask; |
| 1016 | int (*fn)(struct thread_data *, int); |
| 1017 | }; |
| 1018 | |
| 1019 | static struct log_type log_types[] = { |
| 1020 | { |
| 1021 | .mask = BW_LOG_MASK, |
| 1022 | .fn = write_bandw_log, |
| 1023 | }, |
| 1024 | { |
| 1025 | .mask = LAT_LOG_MASK, |
| 1026 | .fn = write_lat_log, |
| 1027 | }, |
| 1028 | { |
| 1029 | .mask = SLAT_LOG_MASK, |
| 1030 | .fn = write_slat_log, |
| 1031 | }, |
| 1032 | { |
| 1033 | .mask = CLAT_LOG_MASK, |
| 1034 | .fn = write_clat_log, |
| 1035 | }, |
| 1036 | { |
| 1037 | .mask = IOPS_LOG_MASK, |
| 1038 | .fn = write_iops_log, |
| 1039 | }, |
| 1040 | }; |
| 1041 | |
| 1042 | void fio_writeout_logs(struct thread_data *td) |
| 1043 | { |
| 1044 | unsigned int log_mask = 0; |
| 1045 | unsigned int log_left = ALL_LOG_NR; |
| 1046 | int old_state, i; |
| 1047 | |
| 1048 | old_state = td_bump_runstate(td, TD_FINISHING); |
| 1049 | |
| 1050 | finalize_logs(td); |
| 1051 | |
| 1052 | while (log_left) { |
| 1053 | int prev_log_left = log_left; |
| 1054 | |
| 1055 | for (i = 0; i < ALL_LOG_NR && log_left; i++) { |
| 1056 | struct log_type *lt = &log_types[i]; |
| 1057 | int ret; |
| 1058 | |
| 1059 | if (!(log_mask & lt->mask)) { |
| 1060 | ret = lt->fn(td, log_left != 1); |
| 1061 | if (!ret) { |
| 1062 | log_left--; |
| 1063 | log_mask |= lt->mask; |
| 1064 | } |
| 1065 | } |
| 1066 | } |
| 1067 | |
| 1068 | if (prev_log_left == log_left) |
| 1069 | usleep(5000); |
| 1070 | } |
| 1071 | |
| 1072 | td_restore_runstate(td, old_state); |
| 1073 | } |