| 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 <assert.h> |
| 8 | #include <sys/types.h> |
| 9 | #include <sys/stat.h> |
| 10 | #include <unistd.h> |
| 11 | #ifdef CONFIG_ZLIB |
| 12 | #include <zlib.h> |
| 13 | #endif |
| 14 | |
| 15 | #include "flist.h" |
| 16 | #include "fio.h" |
| 17 | #include "trim.h" |
| 18 | #include "filelock.h" |
| 19 | #include "smalloc.h" |
| 20 | #include "blktrace.h" |
| 21 | #include "pshared.h" |
| 22 | #include "lib/roundup.h" |
| 23 | |
| 24 | #include <netinet/in.h> |
| 25 | #include <netinet/tcp.h> |
| 26 | #include <arpa/inet.h> |
| 27 | #include <sys/stat.h> |
| 28 | #include <sys/socket.h> |
| 29 | #include <sys/un.h> |
| 30 | |
| 31 | static int iolog_flush(struct io_log *log); |
| 32 | |
| 33 | static const char iolog_ver2[] = "fio version 2 iolog"; |
| 34 | static const char iolog_ver3[] = "fio version 3 iolog"; |
| 35 | |
| 36 | void queue_io_piece(struct thread_data *td, struct io_piece *ipo) |
| 37 | { |
| 38 | flist_add_tail(&ipo->list, &td->io_log_list); |
| 39 | td->total_io_size += ipo->len; |
| 40 | } |
| 41 | |
| 42 | void log_io_u(const struct thread_data *td, const struct io_u *io_u) |
| 43 | { |
| 44 | struct timespec now; |
| 45 | |
| 46 | if (!td->o.write_iolog_file) |
| 47 | return; |
| 48 | |
| 49 | fio_gettime(&now, NULL); |
| 50 | fprintf(td->iolog_f, "%llu %s %s %llu %llu\n", |
| 51 | (unsigned long long) utime_since_now(&td->io_log_start_time), |
| 52 | io_u->file->file_name, io_ddir_name(io_u->ddir), io_u->offset, |
| 53 | io_u->buflen); |
| 54 | |
| 55 | } |
| 56 | |
| 57 | void log_file(struct thread_data *td, struct fio_file *f, |
| 58 | enum file_log_act what) |
| 59 | { |
| 60 | const char *act[] = { "add", "open", "close" }; |
| 61 | struct timespec now; |
| 62 | |
| 63 | assert(what < 3); |
| 64 | |
| 65 | if (!td->o.write_iolog_file) |
| 66 | return; |
| 67 | |
| 68 | |
| 69 | /* |
| 70 | * this happens on the pre-open/close done before the job starts |
| 71 | */ |
| 72 | if (!td->iolog_f) |
| 73 | return; |
| 74 | |
| 75 | fio_gettime(&now, NULL); |
| 76 | fprintf(td->iolog_f, "%llu %s %s\n", |
| 77 | (unsigned long long) utime_since_now(&td->io_log_start_time), |
| 78 | f->file_name, act[what]); |
| 79 | } |
| 80 | |
| 81 | static void iolog_delay(struct thread_data *td, unsigned long delay) |
| 82 | { |
| 83 | uint64_t usec = utime_since_now(&td->last_issue); |
| 84 | unsigned long orig_delay = delay; |
| 85 | uint64_t this_delay; |
| 86 | struct timespec ts; |
| 87 | |
| 88 | if (delay < td->time_offset) { |
| 89 | td->time_offset = 0; |
| 90 | return; |
| 91 | } |
| 92 | |
| 93 | delay -= td->time_offset; |
| 94 | if (delay < usec) |
| 95 | return; |
| 96 | |
| 97 | delay -= usec; |
| 98 | |
| 99 | fio_gettime(&ts, NULL); |
| 100 | while (delay && !td->terminate) { |
| 101 | this_delay = delay; |
| 102 | if (this_delay > 500000) |
| 103 | this_delay = 500000; |
| 104 | |
| 105 | usec_sleep(td, this_delay); |
| 106 | delay -= this_delay; |
| 107 | } |
| 108 | |
| 109 | usec = utime_since_now(&ts); |
| 110 | if (usec > orig_delay) |
| 111 | td->time_offset = usec - orig_delay; |
| 112 | else |
| 113 | td->time_offset = 0; |
| 114 | } |
| 115 | |
| 116 | static int ipo_special(struct thread_data *td, struct io_piece *ipo) |
| 117 | { |
| 118 | struct fio_file *f; |
| 119 | int ret; |
| 120 | |
| 121 | /* |
| 122 | * Not a special ipo |
| 123 | */ |
| 124 | if (ipo->ddir != DDIR_INVAL) |
| 125 | return 0; |
| 126 | |
| 127 | f = td->files[ipo->fileno]; |
| 128 | |
| 129 | if (ipo->delay) |
| 130 | iolog_delay(td, ipo->delay); |
| 131 | if (fio_fill_issue_time(td)) |
| 132 | fio_gettime(&td->last_issue, NULL); |
| 133 | switch (ipo->file_action) { |
| 134 | case FIO_LOG_OPEN_FILE: |
| 135 | if (td->o.replay_redirect && fio_file_open(f)) { |
| 136 | dprint(FD_FILE, "iolog: ignoring re-open of file %s\n", |
| 137 | f->file_name); |
| 138 | break; |
| 139 | } |
| 140 | ret = td_io_open_file(td, f); |
| 141 | if (!ret) |
| 142 | break; |
| 143 | td_verror(td, ret, "iolog open file"); |
| 144 | return -1; |
| 145 | case FIO_LOG_CLOSE_FILE: |
| 146 | td_io_close_file(td, f); |
| 147 | break; |
| 148 | case FIO_LOG_UNLINK_FILE: |
| 149 | td_io_unlink_file(td, f); |
| 150 | break; |
| 151 | case FIO_LOG_ADD_FILE: |
| 152 | /* |
| 153 | * Nothing to do |
| 154 | */ |
| 155 | break; |
| 156 | default: |
| 157 | log_err("fio: bad file action %d\n", ipo->file_action); |
| 158 | break; |
| 159 | } |
| 160 | |
| 161 | return 1; |
| 162 | } |
| 163 | |
| 164 | static bool read_iolog(struct thread_data *td); |
| 165 | |
| 166 | unsigned long long delay_since_ttime(const struct thread_data *td, |
| 167 | unsigned long long time) |
| 168 | { |
| 169 | double tmp; |
| 170 | double scale; |
| 171 | const unsigned long long *last_ttime = &td->io_log_last_ttime; |
| 172 | |
| 173 | if (!*last_ttime || td->o.no_stall || time < *last_ttime) |
| 174 | return 0; |
| 175 | else if (td->o.replay_time_scale == 100) |
| 176 | return time - *last_ttime; |
| 177 | |
| 178 | |
| 179 | scale = (double) 100.0 / (double) td->o.replay_time_scale; |
| 180 | tmp = time - *last_ttime; |
| 181 | return tmp * scale; |
| 182 | } |
| 183 | |
| 184 | int read_iolog_get(struct thread_data *td, struct io_u *io_u) |
| 185 | { |
| 186 | struct io_piece *ipo; |
| 187 | unsigned long elapsed; |
| 188 | |
| 189 | while (!flist_empty(&td->io_log_list)) { |
| 190 | int ret; |
| 191 | |
| 192 | if (td->o.read_iolog_chunked) { |
| 193 | if (td->io_log_checkmark == td->io_log_current) { |
| 194 | if (td->io_log_blktrace) { |
| 195 | if (!read_blktrace(td)) |
| 196 | return 1; |
| 197 | } else { |
| 198 | if (!read_iolog(td)) |
| 199 | return 1; |
| 200 | } |
| 201 | } |
| 202 | td->io_log_current--; |
| 203 | } |
| 204 | ipo = flist_first_entry(&td->io_log_list, struct io_piece, list); |
| 205 | flist_del(&ipo->list); |
| 206 | remove_trim_entry(td, ipo); |
| 207 | |
| 208 | ret = ipo_special(td, ipo); |
| 209 | if (ret < 0) { |
| 210 | free(ipo); |
| 211 | break; |
| 212 | } else if (ret > 0) { |
| 213 | free(ipo); |
| 214 | continue; |
| 215 | } |
| 216 | |
| 217 | io_u->ddir = ipo->ddir; |
| 218 | if (ipo->ddir != DDIR_WAIT) { |
| 219 | io_u->offset = ipo->offset; |
| 220 | io_u->verify_offset = ipo->offset; |
| 221 | io_u->buflen = ipo->len; |
| 222 | io_u->file = td->files[ipo->fileno]; |
| 223 | get_file(io_u->file); |
| 224 | dprint(FD_IO, "iolog: get %llu/%llu/%s\n", io_u->offset, |
| 225 | io_u->buflen, io_u->file->file_name); |
| 226 | if (ipo->delay) |
| 227 | iolog_delay(td, ipo->delay); |
| 228 | } else { |
| 229 | elapsed = mtime_since_genesis(); |
| 230 | if (ipo->delay > elapsed) |
| 231 | usec_sleep(td, (ipo->delay - elapsed) * 1000); |
| 232 | } |
| 233 | |
| 234 | free(ipo); |
| 235 | |
| 236 | if (io_u->ddir != DDIR_WAIT) |
| 237 | return 0; |
| 238 | } |
| 239 | |
| 240 | td->done = 1; |
| 241 | return 1; |
| 242 | } |
| 243 | |
| 244 | void prune_io_piece_log(struct thread_data *td) |
| 245 | { |
| 246 | struct io_piece *ipo; |
| 247 | struct fio_rb_node *n; |
| 248 | |
| 249 | while ((n = rb_first(&td->io_hist_tree)) != NULL) { |
| 250 | ipo = rb_entry(n, struct io_piece, rb_node); |
| 251 | rb_erase(n, &td->io_hist_tree); |
| 252 | remove_trim_entry(td, ipo); |
| 253 | td->io_hist_len--; |
| 254 | free(ipo); |
| 255 | } |
| 256 | |
| 257 | while (!flist_empty(&td->io_hist_list)) { |
| 258 | ipo = flist_first_entry(&td->io_hist_list, struct io_piece, list); |
| 259 | flist_del(&ipo->list); |
| 260 | remove_trim_entry(td, ipo); |
| 261 | td->io_hist_len--; |
| 262 | free(ipo); |
| 263 | } |
| 264 | } |
| 265 | |
| 266 | /* |
| 267 | * log a successful write, so we can unwind the log for verify |
| 268 | */ |
| 269 | void log_io_piece(struct thread_data *td, struct io_u *io_u) |
| 270 | { |
| 271 | struct fio_rb_node **p, *parent; |
| 272 | struct io_piece *ipo, *__ipo; |
| 273 | |
| 274 | ipo = calloc(1, sizeof(struct io_piece)); |
| 275 | init_ipo(ipo); |
| 276 | ipo->file = io_u->file; |
| 277 | ipo->offset = io_u->offset; |
| 278 | ipo->len = io_u->buflen; |
| 279 | ipo->numberio = io_u->numberio; |
| 280 | ipo->flags = IP_F_IN_FLIGHT; |
| 281 | |
| 282 | io_u->ipo = ipo; |
| 283 | |
| 284 | if (io_u_should_trim(td, io_u)) { |
| 285 | flist_add_tail(&ipo->trim_list, &td->trim_list); |
| 286 | td->trim_entries++; |
| 287 | } |
| 288 | |
| 289 | /* |
| 290 | * Only sort writes if we don't have a random map in which case we need |
| 291 | * to check for duplicate blocks and drop the old one, which we rely on |
| 292 | * the rb insert/lookup for handling. |
| 293 | */ |
| 294 | if (file_randommap(td, ipo->file)) { |
| 295 | INIT_FLIST_HEAD(&ipo->list); |
| 296 | flist_add_tail(&ipo->list, &td->io_hist_list); |
| 297 | ipo->flags |= IP_F_ONLIST; |
| 298 | td->io_hist_len++; |
| 299 | return; |
| 300 | } |
| 301 | |
| 302 | RB_CLEAR_NODE(&ipo->rb_node); |
| 303 | |
| 304 | /* |
| 305 | * Sort the entry into the verification list |
| 306 | */ |
| 307 | restart: |
| 308 | p = &td->io_hist_tree.rb_node; |
| 309 | parent = NULL; |
| 310 | while (*p) { |
| 311 | int overlap = 0; |
| 312 | parent = *p; |
| 313 | |
| 314 | __ipo = rb_entry(parent, struct io_piece, rb_node); |
| 315 | if (ipo->file < __ipo->file) |
| 316 | p = &(*p)->rb_left; |
| 317 | else if (ipo->file > __ipo->file) |
| 318 | p = &(*p)->rb_right; |
| 319 | else if (ipo->offset < __ipo->offset) { |
| 320 | p = &(*p)->rb_left; |
| 321 | overlap = ipo->offset + ipo->len > __ipo->offset; |
| 322 | } |
| 323 | else if (ipo->offset > __ipo->offset) { |
| 324 | p = &(*p)->rb_right; |
| 325 | overlap = __ipo->offset + __ipo->len > ipo->offset; |
| 326 | } |
| 327 | else |
| 328 | overlap = 1; |
| 329 | |
| 330 | if (overlap) { |
| 331 | dprint(FD_IO, "iolog: overlap %llu/%lu, %llu/%lu\n", |
| 332 | __ipo->offset, __ipo->len, |
| 333 | ipo->offset, ipo->len); |
| 334 | td->io_hist_len--; |
| 335 | rb_erase(parent, &td->io_hist_tree); |
| 336 | remove_trim_entry(td, __ipo); |
| 337 | if (!(__ipo->flags & IP_F_IN_FLIGHT)) |
| 338 | free(__ipo); |
| 339 | goto restart; |
| 340 | } |
| 341 | } |
| 342 | |
| 343 | rb_link_node(&ipo->rb_node, parent, p); |
| 344 | rb_insert_color(&ipo->rb_node, &td->io_hist_tree); |
| 345 | ipo->flags |= IP_F_ONRB; |
| 346 | td->io_hist_len++; |
| 347 | } |
| 348 | |
| 349 | void unlog_io_piece(struct thread_data *td, struct io_u *io_u) |
| 350 | { |
| 351 | struct io_piece *ipo = io_u->ipo; |
| 352 | |
| 353 | if (td->ts.nr_block_infos) { |
| 354 | uint32_t *info = io_u_block_info(td, io_u); |
| 355 | if (BLOCK_INFO_STATE(*info) < BLOCK_STATE_TRIM_FAILURE) { |
| 356 | if (io_u->ddir == DDIR_TRIM) |
| 357 | *info = BLOCK_INFO_SET_STATE(*info, |
| 358 | BLOCK_STATE_TRIM_FAILURE); |
| 359 | else if (io_u->ddir == DDIR_WRITE) |
| 360 | *info = BLOCK_INFO_SET_STATE(*info, |
| 361 | BLOCK_STATE_WRITE_FAILURE); |
| 362 | } |
| 363 | } |
| 364 | |
| 365 | if (!ipo) |
| 366 | return; |
| 367 | |
| 368 | if (ipo->flags & IP_F_ONRB) |
| 369 | rb_erase(&ipo->rb_node, &td->io_hist_tree); |
| 370 | else if (ipo->flags & IP_F_ONLIST) |
| 371 | flist_del(&ipo->list); |
| 372 | |
| 373 | free(ipo); |
| 374 | io_u->ipo = NULL; |
| 375 | td->io_hist_len--; |
| 376 | } |
| 377 | |
| 378 | void trim_io_piece(const struct io_u *io_u) |
| 379 | { |
| 380 | struct io_piece *ipo = io_u->ipo; |
| 381 | |
| 382 | if (!ipo) |
| 383 | return; |
| 384 | |
| 385 | ipo->len = io_u->xfer_buflen - io_u->resid; |
| 386 | } |
| 387 | |
| 388 | void write_iolog_close(struct thread_data *td) |
| 389 | { |
| 390 | if (!td->iolog_f) |
| 391 | return; |
| 392 | |
| 393 | fflush(td->iolog_f); |
| 394 | fclose(td->iolog_f); |
| 395 | free(td->iolog_buf); |
| 396 | td->iolog_f = NULL; |
| 397 | td->iolog_buf = NULL; |
| 398 | } |
| 399 | |
| 400 | int64_t iolog_items_to_fetch(struct thread_data *td) |
| 401 | { |
| 402 | struct timespec now; |
| 403 | uint64_t elapsed; |
| 404 | uint64_t for_1s; |
| 405 | int64_t items_to_fetch; |
| 406 | |
| 407 | if (!td->io_log_highmark) |
| 408 | return 10; |
| 409 | |
| 410 | |
| 411 | fio_gettime(&now, NULL); |
| 412 | elapsed = ntime_since(&td->io_log_highmark_time, &now); |
| 413 | if (elapsed) { |
| 414 | for_1s = (td->io_log_highmark - td->io_log_current) * 1000000000 / elapsed; |
| 415 | items_to_fetch = for_1s - td->io_log_current; |
| 416 | if (items_to_fetch < 0) |
| 417 | items_to_fetch = 0; |
| 418 | } else |
| 419 | items_to_fetch = 0; |
| 420 | |
| 421 | td->io_log_highmark = td->io_log_current + items_to_fetch; |
| 422 | td->io_log_checkmark = (td->io_log_highmark + 1) / 2; |
| 423 | fio_gettime(&td->io_log_highmark_time, NULL); |
| 424 | |
| 425 | return items_to_fetch; |
| 426 | } |
| 427 | |
| 428 | #define io_act(_td, _r) (((_td)->io_log_version == 3 && (r) == 5) || \ |
| 429 | ((_td)->io_log_version == 2 && (r) == 4)) |
| 430 | #define file_act(_td, _r) (((_td)->io_log_version == 3 && (r) == 3) || \ |
| 431 | ((_td)->io_log_version == 2 && (r) == 2)) |
| 432 | |
| 433 | /* |
| 434 | * Read version 2 and 3 iolog data. It is enhanced to include per-file logging, |
| 435 | * syncs, etc. |
| 436 | */ |
| 437 | static bool read_iolog(struct thread_data *td) |
| 438 | { |
| 439 | unsigned long long offset; |
| 440 | unsigned int bytes; |
| 441 | unsigned long long delay = 0; |
| 442 | int reads, writes, trims, waits, fileno = 0, file_action = 0; /* stupid gcc */ |
| 443 | char *rfname, *fname, *act; |
| 444 | char *str, *p; |
| 445 | enum fio_ddir rw; |
| 446 | bool realloc = false; |
| 447 | int64_t items_to_fetch = 0; |
| 448 | int syncs; |
| 449 | |
| 450 | if (td->o.read_iolog_chunked) { |
| 451 | items_to_fetch = iolog_items_to_fetch(td); |
| 452 | if (!items_to_fetch) |
| 453 | return true; |
| 454 | } |
| 455 | |
| 456 | /* |
| 457 | * Read in the read iolog and store it, reuse the infrastructure |
| 458 | * for doing verifications. |
| 459 | */ |
| 460 | str = malloc(4096); |
| 461 | rfname = fname = malloc(256+16); |
| 462 | act = malloc(256+16); |
| 463 | |
| 464 | syncs = reads = writes = trims = waits = 0; |
| 465 | while ((p = fgets(str, 4096, td->io_log_rfile)) != NULL) { |
| 466 | struct io_piece *ipo; |
| 467 | int r; |
| 468 | unsigned long long ttime; |
| 469 | |
| 470 | if (td->io_log_version == 3) { |
| 471 | r = sscanf(p, "%llu %256s %256s %llu %u", &ttime, rfname, act, |
| 472 | &offset, &bytes); |
| 473 | delay = delay_since_ttime(td, ttime); |
| 474 | td->io_log_last_ttime = ttime; |
| 475 | /* |
| 476 | * "wait" is not allowed with version 3 |
| 477 | */ |
| 478 | if (!strcmp(act, "wait")) { |
| 479 | log_err("iolog: ignoring wait command with" |
| 480 | " version 3 for file %s\n", fname); |
| 481 | continue; |
| 482 | } |
| 483 | } else /* version 2 */ |
| 484 | r = sscanf(p, "%256s %256s %llu %u", rfname, act, &offset, &bytes); |
| 485 | |
| 486 | if (td->o.replay_redirect) |
| 487 | fname = td->o.replay_redirect; |
| 488 | |
| 489 | if (io_act(td, r)) { |
| 490 | /* |
| 491 | * Check action first |
| 492 | */ |
| 493 | if (!strcmp(act, "wait")) |
| 494 | rw = DDIR_WAIT; |
| 495 | else if (!strcmp(act, "read")) { |
| 496 | if (td->o.replay_skip & (1u << DDIR_READ)) |
| 497 | continue; |
| 498 | rw = DDIR_READ; |
| 499 | } else if (!strcmp(act, "write")) { |
| 500 | if (td->o.replay_skip & (1u << DDIR_WRITE)) |
| 501 | continue; |
| 502 | rw = DDIR_WRITE; |
| 503 | } else if (!strcmp(act, "sync")) { |
| 504 | if (td->o.replay_skip & (1u << DDIR_SYNC)) |
| 505 | continue; |
| 506 | rw = DDIR_SYNC; |
| 507 | } else if (!strcmp(act, "datasync")) |
| 508 | rw = DDIR_DATASYNC; |
| 509 | else if (!strcmp(act, "trim")) { |
| 510 | if (td->o.replay_skip & (1u << DDIR_TRIM)) |
| 511 | continue; |
| 512 | rw = DDIR_TRIM; |
| 513 | } else { |
| 514 | log_err("fio: bad iolog file action: %s\n", |
| 515 | act); |
| 516 | continue; |
| 517 | } |
| 518 | fileno = get_fileno(td, fname); |
| 519 | } else if (file_act(td, r)) { |
| 520 | rw = DDIR_INVAL; |
| 521 | if (!strcmp(act, "add")) { |
| 522 | if (td->o.replay_redirect && |
| 523 | get_fileno(td, fname) != -1) { |
| 524 | dprint(FD_FILE, "iolog: ignoring" |
| 525 | " re-add of file %s\n", fname); |
| 526 | } else { |
| 527 | fileno = add_file(td, fname, td->subjob_number, 1); |
| 528 | file_action = FIO_LOG_ADD_FILE; |
| 529 | } |
| 530 | } else if (!strcmp(act, "open")) { |
| 531 | fileno = get_fileno(td, fname); |
| 532 | file_action = FIO_LOG_OPEN_FILE; |
| 533 | } else if (!strcmp(act, "close")) { |
| 534 | fileno = get_fileno(td, fname); |
| 535 | file_action = FIO_LOG_CLOSE_FILE; |
| 536 | } else { |
| 537 | log_err("fio: bad iolog file action: %s\n", |
| 538 | act); |
| 539 | continue; |
| 540 | } |
| 541 | } else { |
| 542 | log_err("bad iolog%d: %s\n", td->io_log_version, p); |
| 543 | continue; |
| 544 | } |
| 545 | |
| 546 | if (rw == DDIR_READ) |
| 547 | reads++; |
| 548 | else if (rw == DDIR_WRITE) { |
| 549 | /* |
| 550 | * Don't add a write for ro mode |
| 551 | */ |
| 552 | if (read_only) |
| 553 | continue; |
| 554 | writes++; |
| 555 | } else if (rw == DDIR_TRIM) { |
| 556 | /* |
| 557 | * Don't add a trim for ro mode |
| 558 | */ |
| 559 | if (read_only) |
| 560 | continue; |
| 561 | trims++; |
| 562 | } else if (rw == DDIR_WAIT) { |
| 563 | if (td->o.no_stall) |
| 564 | continue; |
| 565 | waits++; |
| 566 | } else if (rw == DDIR_INVAL) { |
| 567 | } else if (ddir_sync(rw)) { |
| 568 | syncs++; |
| 569 | } else { |
| 570 | log_err("bad ddir: %d\n", rw); |
| 571 | continue; |
| 572 | } |
| 573 | |
| 574 | /* |
| 575 | * Make note of file |
| 576 | */ |
| 577 | ipo = calloc(1, sizeof(*ipo)); |
| 578 | init_ipo(ipo); |
| 579 | ipo->ddir = rw; |
| 580 | if (td->io_log_version == 3) |
| 581 | ipo->delay = delay; |
| 582 | if (rw == DDIR_WAIT) { |
| 583 | ipo->delay = offset; |
| 584 | } else { |
| 585 | if (td->o.replay_scale) |
| 586 | ipo->offset = offset / td->o.replay_scale; |
| 587 | else |
| 588 | ipo->offset = offset; |
| 589 | ipo_bytes_align(td->o.replay_align, ipo); |
| 590 | |
| 591 | ipo->len = bytes; |
| 592 | if (rw != DDIR_INVAL && bytes > td->o.max_bs[rw]) { |
| 593 | realloc = true; |
| 594 | td->o.max_bs[rw] = bytes; |
| 595 | } |
| 596 | ipo->fileno = fileno; |
| 597 | ipo->file_action = file_action; |
| 598 | td->o.size += bytes; |
| 599 | } |
| 600 | |
| 601 | queue_io_piece(td, ipo); |
| 602 | |
| 603 | if (td->o.read_iolog_chunked) { |
| 604 | td->io_log_current++; |
| 605 | items_to_fetch--; |
| 606 | if (items_to_fetch == 0) |
| 607 | break; |
| 608 | } |
| 609 | } |
| 610 | |
| 611 | free(str); |
| 612 | free(act); |
| 613 | free(rfname); |
| 614 | |
| 615 | if (td->o.read_iolog_chunked) { |
| 616 | td->io_log_highmark = td->io_log_current; |
| 617 | td->io_log_checkmark = (td->io_log_highmark + 1) / 2; |
| 618 | fio_gettime(&td->io_log_highmark_time, NULL); |
| 619 | } |
| 620 | |
| 621 | if (writes && read_only) { |
| 622 | log_err("fio: <%s> skips replay of %d writes due to" |
| 623 | " read-only\n", td->o.name, writes); |
| 624 | writes = 0; |
| 625 | } |
| 626 | if (syncs) |
| 627 | td->flags |= TD_F_SYNCS; |
| 628 | |
| 629 | if (td->o.read_iolog_chunked) { |
| 630 | if (td->io_log_current == 0) { |
| 631 | return false; |
| 632 | } |
| 633 | td->o.td_ddir = TD_DDIR_RW; |
| 634 | if (realloc && td->orig_buffer) |
| 635 | { |
| 636 | io_u_quiesce(td); |
| 637 | free_io_mem(td); |
| 638 | if (init_io_u_buffers(td)) |
| 639 | return false; |
| 640 | } |
| 641 | return true; |
| 642 | } |
| 643 | |
| 644 | if (!reads && !writes && !waits && !trims) |
| 645 | return false; |
| 646 | |
| 647 | td->o.td_ddir = 0; |
| 648 | if (reads) |
| 649 | td->o.td_ddir |= TD_DDIR_READ; |
| 650 | if (writes) |
| 651 | td->o.td_ddir |= TD_DDIR_WRITE; |
| 652 | if (trims) |
| 653 | td->o.td_ddir |= TD_DDIR_TRIM; |
| 654 | |
| 655 | return true; |
| 656 | } |
| 657 | |
| 658 | static bool is_socket(const char *path) |
| 659 | { |
| 660 | struct stat buf; |
| 661 | int r; |
| 662 | |
| 663 | r = stat(path, &buf); |
| 664 | if (r == -1) |
| 665 | return false; |
| 666 | |
| 667 | return S_ISSOCK(buf.st_mode); |
| 668 | } |
| 669 | |
| 670 | static int open_socket(const char *path) |
| 671 | { |
| 672 | struct sockaddr_un addr; |
| 673 | int ret, fd; |
| 674 | |
| 675 | fd = socket(AF_UNIX, SOCK_STREAM, 0); |
| 676 | if (fd < 0) |
| 677 | return fd; |
| 678 | |
| 679 | addr.sun_family = AF_UNIX; |
| 680 | if (snprintf(addr.sun_path, sizeof(addr.sun_path), "%s", path) >= |
| 681 | sizeof(addr.sun_path)) { |
| 682 | log_err("%s: path name %s is too long for a Unix socket\n", |
| 683 | __func__, path); |
| 684 | } |
| 685 | |
| 686 | ret = connect(fd, (const struct sockaddr *)&addr, strlen(path) + sizeof(addr.sun_family)); |
| 687 | if (!ret) |
| 688 | return fd; |
| 689 | |
| 690 | close(fd); |
| 691 | return -1; |
| 692 | } |
| 693 | |
| 694 | /* |
| 695 | * open iolog, check version, and call appropriate parser |
| 696 | */ |
| 697 | static bool init_iolog_read(struct thread_data *td, char *fname) |
| 698 | { |
| 699 | char buffer[256], *p; |
| 700 | FILE *f = NULL; |
| 701 | |
| 702 | dprint(FD_IO, "iolog: name=%s\n", fname); |
| 703 | |
| 704 | if (is_socket(fname)) { |
| 705 | int fd; |
| 706 | |
| 707 | fd = open_socket(fname); |
| 708 | if (fd >= 0) |
| 709 | f = fdopen(fd, "r"); |
| 710 | } else if (!strcmp(fname, "-")) { |
| 711 | f = stdin; |
| 712 | } else |
| 713 | f = fopen(fname, "r"); |
| 714 | |
| 715 | if (!f) { |
| 716 | perror("fopen read iolog"); |
| 717 | return false; |
| 718 | } |
| 719 | |
| 720 | p = fgets(buffer, sizeof(buffer), f); |
| 721 | if (!p) { |
| 722 | td_verror(td, errno, "iolog read"); |
| 723 | log_err("fio: unable to read iolog\n"); |
| 724 | fclose(f); |
| 725 | return false; |
| 726 | } |
| 727 | |
| 728 | /* |
| 729 | * versions 2 and 3 of the iolog store a specific string as the |
| 730 | * first line, check for that |
| 731 | */ |
| 732 | if (!strncmp(iolog_ver2, buffer, strlen(iolog_ver2))) |
| 733 | td->io_log_version = 2; |
| 734 | else if (!strncmp(iolog_ver3, buffer, strlen(iolog_ver3))) |
| 735 | td->io_log_version = 3; |
| 736 | else { |
| 737 | log_err("fio: iolog version 1 is no longer supported\n"); |
| 738 | fclose(f); |
| 739 | return false; |
| 740 | } |
| 741 | |
| 742 | free_release_files(td); |
| 743 | td->io_log_rfile = f; |
| 744 | return read_iolog(td); |
| 745 | } |
| 746 | |
| 747 | /* |
| 748 | * Set up a log for storing io patterns. |
| 749 | */ |
| 750 | static bool init_iolog_write(struct thread_data *td) |
| 751 | { |
| 752 | struct fio_file *ff; |
| 753 | FILE *f; |
| 754 | unsigned int i; |
| 755 | |
| 756 | f = fopen(td->o.write_iolog_file, "a"); |
| 757 | if (!f) { |
| 758 | perror("fopen write iolog"); |
| 759 | return false; |
| 760 | } |
| 761 | |
| 762 | /* |
| 763 | * That's it for writing, setup a log buffer and we're done. |
| 764 | */ |
| 765 | td->iolog_f = f; |
| 766 | td->iolog_buf = malloc(8192); |
| 767 | setvbuf(f, td->iolog_buf, _IOFBF, 8192); |
| 768 | fio_gettime(&td->io_log_start_time, NULL); |
| 769 | |
| 770 | /* |
| 771 | * write our version line |
| 772 | */ |
| 773 | if (fprintf(f, "%s\n", iolog_ver3) < 0) { |
| 774 | perror("iolog init\n"); |
| 775 | return false; |
| 776 | } |
| 777 | |
| 778 | /* |
| 779 | * add all known files |
| 780 | */ |
| 781 | for_each_file(td, ff, i) |
| 782 | log_file(td, ff, FIO_LOG_ADD_FILE); |
| 783 | |
| 784 | return true; |
| 785 | } |
| 786 | |
| 787 | bool init_iolog(struct thread_data *td) |
| 788 | { |
| 789 | bool ret; |
| 790 | |
| 791 | if (td->o.read_iolog_file) { |
| 792 | int need_swap; |
| 793 | char * fname = get_name_by_idx(td->o.read_iolog_file, td->subjob_number); |
| 794 | |
| 795 | /* |
| 796 | * Check if it's a blktrace file and load that if possible. |
| 797 | * Otherwise assume it's a normal log file and load that. |
| 798 | */ |
| 799 | if (is_blktrace(fname, &need_swap)) { |
| 800 | td->io_log_blktrace = 1; |
| 801 | ret = init_blktrace_read(td, fname, need_swap); |
| 802 | } else { |
| 803 | td->io_log_blktrace = 0; |
| 804 | ret = init_iolog_read(td, fname); |
| 805 | } |
| 806 | free(fname); |
| 807 | } else if (td->o.write_iolog_file) |
| 808 | ret = init_iolog_write(td); |
| 809 | else |
| 810 | ret = true; |
| 811 | |
| 812 | if (!ret) |
| 813 | td_verror(td, EINVAL, "failed initializing iolog"); |
| 814 | |
| 815 | return ret; |
| 816 | } |
| 817 | |
| 818 | void setup_log(struct io_log **log, struct log_params *p, |
| 819 | const char *filename) |
| 820 | { |
| 821 | struct io_log *l; |
| 822 | int i; |
| 823 | struct io_u_plat_entry *entry; |
| 824 | struct flist_head *list; |
| 825 | |
| 826 | l = scalloc(1, sizeof(*l)); |
| 827 | INIT_FLIST_HEAD(&l->io_logs); |
| 828 | l->log_type = p->log_type; |
| 829 | l->log_offset = p->log_offset; |
| 830 | l->log_prio = p->log_prio; |
| 831 | l->log_gz = p->log_gz; |
| 832 | l->log_gz_store = p->log_gz_store; |
| 833 | l->avg_msec = p->avg_msec; |
| 834 | l->hist_msec = p->hist_msec; |
| 835 | l->hist_coarseness = p->hist_coarseness; |
| 836 | l->filename = strdup(filename); |
| 837 | l->td = p->td; |
| 838 | |
| 839 | /* Initialize histogram lists for each r/w direction, |
| 840 | * with initial io_u_plat of all zeros: |
| 841 | */ |
| 842 | for (i = 0; i < DDIR_RWDIR_CNT; i++) { |
| 843 | list = &l->hist_window[i].list; |
| 844 | INIT_FLIST_HEAD(list); |
| 845 | entry = calloc(1, sizeof(struct io_u_plat_entry)); |
| 846 | flist_add(&entry->list, list); |
| 847 | } |
| 848 | |
| 849 | if (l->td && l->td->o.io_submit_mode != IO_MODE_OFFLOAD) { |
| 850 | unsigned int def_samples = DEF_LOG_ENTRIES; |
| 851 | struct io_logs *__p; |
| 852 | |
| 853 | __p = calloc(1, sizeof(*l->pending)); |
| 854 | if (l->td->o.iodepth > DEF_LOG_ENTRIES) |
| 855 | def_samples = roundup_pow2(l->td->o.iodepth); |
| 856 | __p->max_samples = def_samples; |
| 857 | __p->log = calloc(__p->max_samples, log_entry_sz(l)); |
| 858 | l->pending = __p; |
| 859 | } |
| 860 | |
| 861 | if (l->log_offset) |
| 862 | l->log_ddir_mask = LOG_OFFSET_SAMPLE_BIT; |
| 863 | if (l->log_prio) |
| 864 | l->log_ddir_mask |= LOG_PRIO_SAMPLE_BIT; |
| 865 | |
| 866 | INIT_FLIST_HEAD(&l->chunk_list); |
| 867 | |
| 868 | if (l->log_gz && !p->td) |
| 869 | l->log_gz = 0; |
| 870 | else if (l->log_gz || l->log_gz_store) { |
| 871 | mutex_init_pshared(&l->chunk_lock); |
| 872 | mutex_init_pshared(&l->deferred_free_lock); |
| 873 | p->td->flags |= TD_F_COMPRESS_LOG; |
| 874 | } |
| 875 | |
| 876 | *log = l; |
| 877 | } |
| 878 | |
| 879 | #ifdef CONFIG_SETVBUF |
| 880 | static void *set_file_buffer(FILE *f) |
| 881 | { |
| 882 | size_t size = 1048576; |
| 883 | void *buf; |
| 884 | |
| 885 | buf = malloc(size); |
| 886 | setvbuf(f, buf, _IOFBF, size); |
| 887 | return buf; |
| 888 | } |
| 889 | |
| 890 | static void clear_file_buffer(void *buf) |
| 891 | { |
| 892 | free(buf); |
| 893 | } |
| 894 | #else |
| 895 | static void *set_file_buffer(FILE *f) |
| 896 | { |
| 897 | return NULL; |
| 898 | } |
| 899 | |
| 900 | static void clear_file_buffer(void *buf) |
| 901 | { |
| 902 | } |
| 903 | #endif |
| 904 | |
| 905 | void free_log(struct io_log *log) |
| 906 | { |
| 907 | while (!flist_empty(&log->io_logs)) { |
| 908 | struct io_logs *cur_log; |
| 909 | |
| 910 | cur_log = flist_first_entry(&log->io_logs, struct io_logs, list); |
| 911 | flist_del_init(&cur_log->list); |
| 912 | free(cur_log->log); |
| 913 | sfree(cur_log); |
| 914 | } |
| 915 | |
| 916 | if (log->pending) { |
| 917 | free(log->pending->log); |
| 918 | free(log->pending); |
| 919 | log->pending = NULL; |
| 920 | } |
| 921 | |
| 922 | free(log->pending); |
| 923 | free(log->filename); |
| 924 | sfree(log); |
| 925 | } |
| 926 | |
| 927 | uint64_t hist_sum(int j, int stride, uint64_t *io_u_plat, |
| 928 | uint64_t *io_u_plat_last) |
| 929 | { |
| 930 | uint64_t sum; |
| 931 | int k; |
| 932 | |
| 933 | if (io_u_plat_last) { |
| 934 | for (k = sum = 0; k < stride; k++) |
| 935 | sum += io_u_plat[j + k] - io_u_plat_last[j + k]; |
| 936 | } else { |
| 937 | for (k = sum = 0; k < stride; k++) |
| 938 | sum += io_u_plat[j + k]; |
| 939 | } |
| 940 | |
| 941 | return sum; |
| 942 | } |
| 943 | |
| 944 | static void flush_hist_samples(FILE *f, int hist_coarseness, void *samples, |
| 945 | uint64_t sample_size) |
| 946 | { |
| 947 | struct io_sample *s; |
| 948 | int log_offset; |
| 949 | uint64_t i, j, nr_samples; |
| 950 | struct io_u_plat_entry *entry, *entry_before; |
| 951 | uint64_t *io_u_plat; |
| 952 | uint64_t *io_u_plat_before; |
| 953 | |
| 954 | int stride = 1 << hist_coarseness; |
| 955 | |
| 956 | if (!sample_size) |
| 957 | return; |
| 958 | |
| 959 | s = __get_sample(samples, 0, 0); |
| 960 | log_offset = (s->__ddir & LOG_OFFSET_SAMPLE_BIT) != 0; |
| 961 | |
| 962 | nr_samples = sample_size / __log_entry_sz(log_offset); |
| 963 | |
| 964 | for (i = 0; i < nr_samples; i++) { |
| 965 | s = __get_sample(samples, log_offset, i); |
| 966 | |
| 967 | entry = s->data.plat_entry; |
| 968 | io_u_plat = entry->io_u_plat; |
| 969 | |
| 970 | entry_before = flist_first_entry(&entry->list, struct io_u_plat_entry, list); |
| 971 | io_u_plat_before = entry_before->io_u_plat; |
| 972 | |
| 973 | fprintf(f, "%lu, %u, %llu, ", (unsigned long) s->time, |
| 974 | io_sample_ddir(s), (unsigned long long) s->bs); |
| 975 | for (j = 0; j < FIO_IO_U_PLAT_NR - stride; j += stride) { |
| 976 | fprintf(f, "%llu, ", (unsigned long long) |
| 977 | hist_sum(j, stride, io_u_plat, io_u_plat_before)); |
| 978 | } |
| 979 | fprintf(f, "%llu\n", (unsigned long long) |
| 980 | hist_sum(FIO_IO_U_PLAT_NR - stride, stride, io_u_plat, |
| 981 | io_u_plat_before)); |
| 982 | |
| 983 | flist_del(&entry_before->list); |
| 984 | free(entry_before); |
| 985 | } |
| 986 | } |
| 987 | |
| 988 | void flush_samples(FILE *f, void *samples, uint64_t sample_size) |
| 989 | { |
| 990 | struct io_sample *s; |
| 991 | int log_offset, log_prio; |
| 992 | uint64_t i, nr_samples; |
| 993 | unsigned int prio_val; |
| 994 | const char *fmt; |
| 995 | |
| 996 | if (!sample_size) |
| 997 | return; |
| 998 | |
| 999 | s = __get_sample(samples, 0, 0); |
| 1000 | log_offset = (s->__ddir & LOG_OFFSET_SAMPLE_BIT) != 0; |
| 1001 | log_prio = (s->__ddir & LOG_PRIO_SAMPLE_BIT) != 0; |
| 1002 | |
| 1003 | if (log_offset) { |
| 1004 | if (log_prio) |
| 1005 | fmt = "%lu, %" PRId64 ", %u, %llu, %llu, 0x%04x\n"; |
| 1006 | else |
| 1007 | fmt = "%lu, %" PRId64 ", %u, %llu, %llu, %u\n"; |
| 1008 | } else { |
| 1009 | if (log_prio) |
| 1010 | fmt = "%lu, %" PRId64 ", %u, %llu, 0x%04x\n"; |
| 1011 | else |
| 1012 | fmt = "%lu, %" PRId64 ", %u, %llu, %u\n"; |
| 1013 | } |
| 1014 | |
| 1015 | nr_samples = sample_size / __log_entry_sz(log_offset); |
| 1016 | |
| 1017 | for (i = 0; i < nr_samples; i++) { |
| 1018 | s = __get_sample(samples, log_offset, i); |
| 1019 | |
| 1020 | if (log_prio) |
| 1021 | prio_val = s->priority; |
| 1022 | else |
| 1023 | prio_val = ioprio_value_is_class_rt(s->priority); |
| 1024 | |
| 1025 | if (!log_offset) { |
| 1026 | fprintf(f, fmt, |
| 1027 | (unsigned long) s->time, |
| 1028 | s->data.val, |
| 1029 | io_sample_ddir(s), (unsigned long long) s->bs, |
| 1030 | prio_val); |
| 1031 | } else { |
| 1032 | struct io_sample_offset *so = (void *) s; |
| 1033 | |
| 1034 | fprintf(f, fmt, |
| 1035 | (unsigned long) s->time, |
| 1036 | s->data.val, |
| 1037 | io_sample_ddir(s), (unsigned long long) s->bs, |
| 1038 | (unsigned long long) so->offset, |
| 1039 | prio_val); |
| 1040 | } |
| 1041 | } |
| 1042 | } |
| 1043 | |
| 1044 | #ifdef CONFIG_ZLIB |
| 1045 | |
| 1046 | struct iolog_flush_data { |
| 1047 | struct workqueue_work work; |
| 1048 | struct io_log *log; |
| 1049 | void *samples; |
| 1050 | uint32_t nr_samples; |
| 1051 | bool free; |
| 1052 | }; |
| 1053 | |
| 1054 | #define GZ_CHUNK 131072 |
| 1055 | |
| 1056 | static struct iolog_compress *get_new_chunk(unsigned int seq) |
| 1057 | { |
| 1058 | struct iolog_compress *c; |
| 1059 | |
| 1060 | c = malloc(sizeof(*c)); |
| 1061 | INIT_FLIST_HEAD(&c->list); |
| 1062 | c->buf = malloc(GZ_CHUNK); |
| 1063 | c->len = 0; |
| 1064 | c->seq = seq; |
| 1065 | return c; |
| 1066 | } |
| 1067 | |
| 1068 | static void free_chunk(struct iolog_compress *ic) |
| 1069 | { |
| 1070 | free(ic->buf); |
| 1071 | free(ic); |
| 1072 | } |
| 1073 | |
| 1074 | static int z_stream_init(z_stream *stream, int gz_hdr) |
| 1075 | { |
| 1076 | int wbits = 15; |
| 1077 | |
| 1078 | memset(stream, 0, sizeof(*stream)); |
| 1079 | stream->zalloc = Z_NULL; |
| 1080 | stream->zfree = Z_NULL; |
| 1081 | stream->opaque = Z_NULL; |
| 1082 | stream->next_in = Z_NULL; |
| 1083 | |
| 1084 | /* |
| 1085 | * zlib magic - add 32 for auto-detection of gz header or not, |
| 1086 | * if we decide to store files in a gzip friendly format. |
| 1087 | */ |
| 1088 | if (gz_hdr) |
| 1089 | wbits += 32; |
| 1090 | |
| 1091 | if (inflateInit2(stream, wbits) != Z_OK) |
| 1092 | return 1; |
| 1093 | |
| 1094 | return 0; |
| 1095 | } |
| 1096 | |
| 1097 | struct inflate_chunk_iter { |
| 1098 | unsigned int seq; |
| 1099 | int err; |
| 1100 | void *buf; |
| 1101 | size_t buf_size; |
| 1102 | size_t buf_used; |
| 1103 | size_t chunk_sz; |
| 1104 | }; |
| 1105 | |
| 1106 | static void finish_chunk(z_stream *stream, FILE *f, |
| 1107 | struct inflate_chunk_iter *iter) |
| 1108 | { |
| 1109 | int ret; |
| 1110 | |
| 1111 | ret = inflateEnd(stream); |
| 1112 | if (ret != Z_OK) |
| 1113 | log_err("fio: failed to end log inflation seq %d (%d)\n", |
| 1114 | iter->seq, ret); |
| 1115 | |
| 1116 | flush_samples(f, iter->buf, iter->buf_used); |
| 1117 | free(iter->buf); |
| 1118 | iter->buf = NULL; |
| 1119 | iter->buf_size = iter->buf_used = 0; |
| 1120 | } |
| 1121 | |
| 1122 | /* |
| 1123 | * Iterative chunk inflation. Handles cases where we cross into a new |
| 1124 | * sequence, doing flush finish of previous chunk if needed. |
| 1125 | */ |
| 1126 | static size_t inflate_chunk(struct iolog_compress *ic, int gz_hdr, FILE *f, |
| 1127 | z_stream *stream, struct inflate_chunk_iter *iter) |
| 1128 | { |
| 1129 | size_t ret; |
| 1130 | |
| 1131 | dprint(FD_COMPRESS, "inflate chunk size=%lu, seq=%u\n", |
| 1132 | (unsigned long) ic->len, ic->seq); |
| 1133 | |
| 1134 | if (ic->seq != iter->seq) { |
| 1135 | if (iter->seq) |
| 1136 | finish_chunk(stream, f, iter); |
| 1137 | |
| 1138 | z_stream_init(stream, gz_hdr); |
| 1139 | iter->seq = ic->seq; |
| 1140 | } |
| 1141 | |
| 1142 | stream->avail_in = ic->len; |
| 1143 | stream->next_in = ic->buf; |
| 1144 | |
| 1145 | if (!iter->buf_size) { |
| 1146 | iter->buf_size = iter->chunk_sz; |
| 1147 | iter->buf = malloc(iter->buf_size); |
| 1148 | } |
| 1149 | |
| 1150 | while (stream->avail_in) { |
| 1151 | size_t this_out = iter->buf_size - iter->buf_used; |
| 1152 | int err; |
| 1153 | |
| 1154 | stream->avail_out = this_out; |
| 1155 | stream->next_out = iter->buf + iter->buf_used; |
| 1156 | |
| 1157 | err = inflate(stream, Z_NO_FLUSH); |
| 1158 | if (err < 0) { |
| 1159 | log_err("fio: failed inflating log: %d\n", err); |
| 1160 | iter->err = err; |
| 1161 | break; |
| 1162 | } |
| 1163 | |
| 1164 | iter->buf_used += this_out - stream->avail_out; |
| 1165 | |
| 1166 | if (!stream->avail_out) { |
| 1167 | iter->buf_size += iter->chunk_sz; |
| 1168 | iter->buf = realloc(iter->buf, iter->buf_size); |
| 1169 | continue; |
| 1170 | } |
| 1171 | |
| 1172 | if (err == Z_STREAM_END) |
| 1173 | break; |
| 1174 | } |
| 1175 | |
| 1176 | ret = (void *) stream->next_in - ic->buf; |
| 1177 | |
| 1178 | dprint(FD_COMPRESS, "inflated to size=%lu\n", (unsigned long) iter->buf_size); |
| 1179 | |
| 1180 | return ret; |
| 1181 | } |
| 1182 | |
| 1183 | /* |
| 1184 | * Inflate stored compressed chunks, or write them directly to the log |
| 1185 | * file if so instructed. |
| 1186 | */ |
| 1187 | static int inflate_gz_chunks(struct io_log *log, FILE *f) |
| 1188 | { |
| 1189 | struct inflate_chunk_iter iter = { .chunk_sz = log->log_gz, }; |
| 1190 | z_stream stream; |
| 1191 | |
| 1192 | while (!flist_empty(&log->chunk_list)) { |
| 1193 | struct iolog_compress *ic; |
| 1194 | |
| 1195 | ic = flist_first_entry(&log->chunk_list, struct iolog_compress, list); |
| 1196 | flist_del(&ic->list); |
| 1197 | |
| 1198 | if (log->log_gz_store) { |
| 1199 | size_t ret; |
| 1200 | |
| 1201 | dprint(FD_COMPRESS, "log write chunk size=%lu, " |
| 1202 | "seq=%u\n", (unsigned long) ic->len, ic->seq); |
| 1203 | |
| 1204 | ret = fwrite(ic->buf, ic->len, 1, f); |
| 1205 | if (ret != 1 || ferror(f)) { |
| 1206 | iter.err = errno; |
| 1207 | log_err("fio: error writing compressed log\n"); |
| 1208 | } |
| 1209 | } else |
| 1210 | inflate_chunk(ic, log->log_gz_store, f, &stream, &iter); |
| 1211 | |
| 1212 | free_chunk(ic); |
| 1213 | } |
| 1214 | |
| 1215 | if (iter.seq) { |
| 1216 | finish_chunk(&stream, f, &iter); |
| 1217 | free(iter.buf); |
| 1218 | } |
| 1219 | |
| 1220 | return iter.err; |
| 1221 | } |
| 1222 | |
| 1223 | /* |
| 1224 | * Open compressed log file and decompress the stored chunks and |
| 1225 | * write them to stdout. The chunks are stored sequentially in the |
| 1226 | * file, so we iterate over them and do them one-by-one. |
| 1227 | */ |
| 1228 | int iolog_file_inflate(const char *file) |
| 1229 | { |
| 1230 | struct inflate_chunk_iter iter = { .chunk_sz = 64 * 1024 * 1024, }; |
| 1231 | struct iolog_compress ic; |
| 1232 | z_stream stream; |
| 1233 | struct stat sb; |
| 1234 | size_t ret; |
| 1235 | size_t total; |
| 1236 | void *buf; |
| 1237 | FILE *f; |
| 1238 | |
| 1239 | f = fopen(file, "rb"); |
| 1240 | if (!f) { |
| 1241 | perror("fopen"); |
| 1242 | return 1; |
| 1243 | } |
| 1244 | |
| 1245 | if (stat(file, &sb) < 0) { |
| 1246 | fclose(f); |
| 1247 | perror("stat"); |
| 1248 | return 1; |
| 1249 | } |
| 1250 | |
| 1251 | ic.buf = buf = malloc(sb.st_size); |
| 1252 | ic.len = sb.st_size; |
| 1253 | ic.seq = 1; |
| 1254 | |
| 1255 | ret = fread(ic.buf, ic.len, 1, f); |
| 1256 | if (ret == 0 && ferror(f)) { |
| 1257 | perror("fread"); |
| 1258 | fclose(f); |
| 1259 | free(buf); |
| 1260 | return 1; |
| 1261 | } else if (ferror(f) || (!feof(f) && ret != 1)) { |
| 1262 | log_err("fio: short read on reading log\n"); |
| 1263 | fclose(f); |
| 1264 | free(buf); |
| 1265 | return 1; |
| 1266 | } |
| 1267 | |
| 1268 | fclose(f); |
| 1269 | |
| 1270 | /* |
| 1271 | * Each chunk will return Z_STREAM_END. We don't know how many |
| 1272 | * chunks are in the file, so we just keep looping and incrementing |
| 1273 | * the sequence number until we have consumed the whole compressed |
| 1274 | * file. |
| 1275 | */ |
| 1276 | total = ic.len; |
| 1277 | do { |
| 1278 | size_t iret; |
| 1279 | |
| 1280 | iret = inflate_chunk(&ic, 1, stdout, &stream, &iter); |
| 1281 | total -= iret; |
| 1282 | if (!total) |
| 1283 | break; |
| 1284 | if (iter.err) |
| 1285 | break; |
| 1286 | |
| 1287 | ic.seq++; |
| 1288 | ic.len -= iret; |
| 1289 | ic.buf += iret; |
| 1290 | } while (1); |
| 1291 | |
| 1292 | if (iter.seq) { |
| 1293 | finish_chunk(&stream, stdout, &iter); |
| 1294 | free(iter.buf); |
| 1295 | } |
| 1296 | |
| 1297 | free(buf); |
| 1298 | return iter.err; |
| 1299 | } |
| 1300 | |
| 1301 | #else |
| 1302 | |
| 1303 | static int inflate_gz_chunks(struct io_log *log, FILE *f) |
| 1304 | { |
| 1305 | return 0; |
| 1306 | } |
| 1307 | |
| 1308 | int iolog_file_inflate(const char *file) |
| 1309 | { |
| 1310 | log_err("fio: log inflation not possible without zlib\n"); |
| 1311 | return 1; |
| 1312 | } |
| 1313 | |
| 1314 | #endif |
| 1315 | |
| 1316 | void flush_log(struct io_log *log, bool do_append) |
| 1317 | { |
| 1318 | void *buf; |
| 1319 | FILE *f; |
| 1320 | |
| 1321 | /* |
| 1322 | * If log_gz_store is true, we are writing a binary file. |
| 1323 | * Set the mode appropriately (on all platforms) to avoid issues |
| 1324 | * on windows (line-ending conversions, etc.) |
| 1325 | */ |
| 1326 | if (!do_append) |
| 1327 | if (log->log_gz_store) |
| 1328 | f = fopen(log->filename, "wb"); |
| 1329 | else |
| 1330 | f = fopen(log->filename, "w"); |
| 1331 | else |
| 1332 | if (log->log_gz_store) |
| 1333 | f = fopen(log->filename, "ab"); |
| 1334 | else |
| 1335 | f = fopen(log->filename, "a"); |
| 1336 | if (!f) { |
| 1337 | perror("fopen log"); |
| 1338 | return; |
| 1339 | } |
| 1340 | |
| 1341 | buf = set_file_buffer(f); |
| 1342 | |
| 1343 | inflate_gz_chunks(log, f); |
| 1344 | |
| 1345 | while (!flist_empty(&log->io_logs)) { |
| 1346 | struct io_logs *cur_log; |
| 1347 | |
| 1348 | cur_log = flist_first_entry(&log->io_logs, struct io_logs, list); |
| 1349 | flist_del_init(&cur_log->list); |
| 1350 | |
| 1351 | if (log->td && log == log->td->clat_hist_log) |
| 1352 | flush_hist_samples(f, log->hist_coarseness, cur_log->log, |
| 1353 | log_sample_sz(log, cur_log)); |
| 1354 | else |
| 1355 | flush_samples(f, cur_log->log, log_sample_sz(log, cur_log)); |
| 1356 | |
| 1357 | sfree(cur_log); |
| 1358 | } |
| 1359 | |
| 1360 | fclose(f); |
| 1361 | clear_file_buffer(buf); |
| 1362 | } |
| 1363 | |
| 1364 | static int finish_log(struct thread_data *td, struct io_log *log, int trylock) |
| 1365 | { |
| 1366 | if (td->flags & TD_F_COMPRESS_LOG) |
| 1367 | iolog_flush(log); |
| 1368 | |
| 1369 | if (trylock) { |
| 1370 | if (fio_trylock_file(log->filename)) |
| 1371 | return 1; |
| 1372 | } else |
| 1373 | fio_lock_file(log->filename); |
| 1374 | |
| 1375 | if (td->client_type == FIO_CLIENT_TYPE_GUI || is_backend) |
| 1376 | fio_send_iolog(td, log, log->filename); |
| 1377 | else |
| 1378 | flush_log(log, !td->o.per_job_logs); |
| 1379 | |
| 1380 | fio_unlock_file(log->filename); |
| 1381 | free_log(log); |
| 1382 | return 0; |
| 1383 | } |
| 1384 | |
| 1385 | size_t log_chunk_sizes(struct io_log *log) |
| 1386 | { |
| 1387 | struct flist_head *entry; |
| 1388 | size_t ret; |
| 1389 | |
| 1390 | if (flist_empty(&log->chunk_list)) |
| 1391 | return 0; |
| 1392 | |
| 1393 | ret = 0; |
| 1394 | pthread_mutex_lock(&log->chunk_lock); |
| 1395 | flist_for_each(entry, &log->chunk_list) { |
| 1396 | struct iolog_compress *c; |
| 1397 | |
| 1398 | c = flist_entry(entry, struct iolog_compress, list); |
| 1399 | ret += c->len; |
| 1400 | } |
| 1401 | pthread_mutex_unlock(&log->chunk_lock); |
| 1402 | return ret; |
| 1403 | } |
| 1404 | |
| 1405 | #ifdef CONFIG_ZLIB |
| 1406 | |
| 1407 | static void iolog_put_deferred(struct io_log *log, void *ptr) |
| 1408 | { |
| 1409 | if (!ptr) |
| 1410 | return; |
| 1411 | |
| 1412 | pthread_mutex_lock(&log->deferred_free_lock); |
| 1413 | if (log->deferred < IOLOG_MAX_DEFER) { |
| 1414 | log->deferred_items[log->deferred] = ptr; |
| 1415 | log->deferred++; |
| 1416 | } else if (!fio_did_warn(FIO_WARN_IOLOG_DROP)) |
| 1417 | log_err("fio: had to drop log entry free\n"); |
| 1418 | pthread_mutex_unlock(&log->deferred_free_lock); |
| 1419 | } |
| 1420 | |
| 1421 | static void iolog_free_deferred(struct io_log *log) |
| 1422 | { |
| 1423 | int i; |
| 1424 | |
| 1425 | if (!log->deferred) |
| 1426 | return; |
| 1427 | |
| 1428 | pthread_mutex_lock(&log->deferred_free_lock); |
| 1429 | |
| 1430 | for (i = 0; i < log->deferred; i++) { |
| 1431 | free(log->deferred_items[i]); |
| 1432 | log->deferred_items[i] = NULL; |
| 1433 | } |
| 1434 | |
| 1435 | log->deferred = 0; |
| 1436 | pthread_mutex_unlock(&log->deferred_free_lock); |
| 1437 | } |
| 1438 | |
| 1439 | static int gz_work(struct iolog_flush_data *data) |
| 1440 | { |
| 1441 | struct iolog_compress *c = NULL; |
| 1442 | struct flist_head list; |
| 1443 | unsigned int seq; |
| 1444 | z_stream stream; |
| 1445 | size_t total = 0; |
| 1446 | int ret; |
| 1447 | |
| 1448 | INIT_FLIST_HEAD(&list); |
| 1449 | |
| 1450 | memset(&stream, 0, sizeof(stream)); |
| 1451 | stream.zalloc = Z_NULL; |
| 1452 | stream.zfree = Z_NULL; |
| 1453 | stream.opaque = Z_NULL; |
| 1454 | |
| 1455 | ret = deflateInit(&stream, Z_DEFAULT_COMPRESSION); |
| 1456 | if (ret != Z_OK) { |
| 1457 | log_err("fio: failed to init gz stream\n"); |
| 1458 | goto err; |
| 1459 | } |
| 1460 | |
| 1461 | seq = ++data->log->chunk_seq; |
| 1462 | |
| 1463 | stream.next_in = (void *) data->samples; |
| 1464 | stream.avail_in = data->nr_samples * log_entry_sz(data->log); |
| 1465 | |
| 1466 | dprint(FD_COMPRESS, "deflate input size=%lu, seq=%u, log=%s\n", |
| 1467 | (unsigned long) stream.avail_in, seq, |
| 1468 | data->log->filename); |
| 1469 | do { |
| 1470 | if (c) |
| 1471 | dprint(FD_COMPRESS, "seq=%d, chunk=%lu\n", seq, |
| 1472 | (unsigned long) c->len); |
| 1473 | c = get_new_chunk(seq); |
| 1474 | stream.avail_out = GZ_CHUNK; |
| 1475 | stream.next_out = c->buf; |
| 1476 | ret = deflate(&stream, Z_NO_FLUSH); |
| 1477 | if (ret < 0) { |
| 1478 | log_err("fio: deflate log (%d)\n", ret); |
| 1479 | free_chunk(c); |
| 1480 | goto err; |
| 1481 | } |
| 1482 | |
| 1483 | c->len = GZ_CHUNK - stream.avail_out; |
| 1484 | flist_add_tail(&c->list, &list); |
| 1485 | total += c->len; |
| 1486 | } while (stream.avail_in); |
| 1487 | |
| 1488 | stream.next_out = c->buf + c->len; |
| 1489 | stream.avail_out = GZ_CHUNK - c->len; |
| 1490 | |
| 1491 | ret = deflate(&stream, Z_FINISH); |
| 1492 | if (ret < 0) { |
| 1493 | /* |
| 1494 | * Z_BUF_ERROR is special, it just means we need more |
| 1495 | * output space. We'll handle that below. Treat any other |
| 1496 | * error as fatal. |
| 1497 | */ |
| 1498 | if (ret != Z_BUF_ERROR) { |
| 1499 | log_err("fio: deflate log (%d)\n", ret); |
| 1500 | flist_del(&c->list); |
| 1501 | free_chunk(c); |
| 1502 | goto err; |
| 1503 | } |
| 1504 | } |
| 1505 | |
| 1506 | total -= c->len; |
| 1507 | c->len = GZ_CHUNK - stream.avail_out; |
| 1508 | total += c->len; |
| 1509 | dprint(FD_COMPRESS, "seq=%d, chunk=%lu\n", seq, (unsigned long) c->len); |
| 1510 | |
| 1511 | if (ret != Z_STREAM_END) { |
| 1512 | do { |
| 1513 | c = get_new_chunk(seq); |
| 1514 | stream.avail_out = GZ_CHUNK; |
| 1515 | stream.next_out = c->buf; |
| 1516 | ret = deflate(&stream, Z_FINISH); |
| 1517 | c->len = GZ_CHUNK - stream.avail_out; |
| 1518 | total += c->len; |
| 1519 | flist_add_tail(&c->list, &list); |
| 1520 | dprint(FD_COMPRESS, "seq=%d, chunk=%lu\n", seq, |
| 1521 | (unsigned long) c->len); |
| 1522 | } while (ret != Z_STREAM_END); |
| 1523 | } |
| 1524 | |
| 1525 | dprint(FD_COMPRESS, "deflated to size=%lu\n", (unsigned long) total); |
| 1526 | |
| 1527 | ret = deflateEnd(&stream); |
| 1528 | if (ret != Z_OK) |
| 1529 | log_err("fio: deflateEnd %d\n", ret); |
| 1530 | |
| 1531 | iolog_put_deferred(data->log, data->samples); |
| 1532 | |
| 1533 | if (!flist_empty(&list)) { |
| 1534 | pthread_mutex_lock(&data->log->chunk_lock); |
| 1535 | flist_splice_tail(&list, &data->log->chunk_list); |
| 1536 | pthread_mutex_unlock(&data->log->chunk_lock); |
| 1537 | } |
| 1538 | |
| 1539 | ret = 0; |
| 1540 | done: |
| 1541 | if (data->free) |
| 1542 | sfree(data); |
| 1543 | return ret; |
| 1544 | err: |
| 1545 | while (!flist_empty(&list)) { |
| 1546 | c = flist_first_entry(list.next, struct iolog_compress, list); |
| 1547 | flist_del(&c->list); |
| 1548 | free_chunk(c); |
| 1549 | } |
| 1550 | ret = 1; |
| 1551 | goto done; |
| 1552 | } |
| 1553 | |
| 1554 | /* |
| 1555 | * Invoked from our compress helper thread, when logging would have exceeded |
| 1556 | * the specified memory limitation. Compresses the previously stored |
| 1557 | * entries. |
| 1558 | */ |
| 1559 | static int gz_work_async(struct submit_worker *sw, struct workqueue_work *work) |
| 1560 | { |
| 1561 | return gz_work(container_of(work, struct iolog_flush_data, work)); |
| 1562 | } |
| 1563 | |
| 1564 | static int gz_init_worker(struct submit_worker *sw) |
| 1565 | { |
| 1566 | struct thread_data *td = sw->wq->td; |
| 1567 | |
| 1568 | if (!fio_option_is_set(&td->o, log_gz_cpumask)) |
| 1569 | return 0; |
| 1570 | |
| 1571 | if (fio_setaffinity(gettid(), td->o.log_gz_cpumask) == -1) { |
| 1572 | log_err("gz: failed to set CPU affinity\n"); |
| 1573 | return 1; |
| 1574 | } |
| 1575 | |
| 1576 | return 0; |
| 1577 | } |
| 1578 | |
| 1579 | static struct workqueue_ops log_compress_wq_ops = { |
| 1580 | .fn = gz_work_async, |
| 1581 | .init_worker_fn = gz_init_worker, |
| 1582 | .nice = 1, |
| 1583 | }; |
| 1584 | |
| 1585 | int iolog_compress_init(struct thread_data *td, struct sk_out *sk_out) |
| 1586 | { |
| 1587 | if (!(td->flags & TD_F_COMPRESS_LOG)) |
| 1588 | return 0; |
| 1589 | |
| 1590 | workqueue_init(td, &td->log_compress_wq, &log_compress_wq_ops, 1, sk_out); |
| 1591 | return 0; |
| 1592 | } |
| 1593 | |
| 1594 | void iolog_compress_exit(struct thread_data *td) |
| 1595 | { |
| 1596 | if (!(td->flags & TD_F_COMPRESS_LOG)) |
| 1597 | return; |
| 1598 | |
| 1599 | workqueue_exit(&td->log_compress_wq); |
| 1600 | } |
| 1601 | |
| 1602 | /* |
| 1603 | * Queue work item to compress the existing log entries. We reset the |
| 1604 | * current log to a small size, and reference the existing log in the |
| 1605 | * data that we queue for compression. Once compression has been done, |
| 1606 | * this old log is freed. Will not return until the log compression |
| 1607 | * has completed, and will flush all previous logs too |
| 1608 | */ |
| 1609 | static int iolog_flush(struct io_log *log) |
| 1610 | { |
| 1611 | struct iolog_flush_data *data; |
| 1612 | |
| 1613 | workqueue_flush(&log->td->log_compress_wq); |
| 1614 | data = malloc(sizeof(*data)); |
| 1615 | if (!data) |
| 1616 | return 1; |
| 1617 | |
| 1618 | data->log = log; |
| 1619 | data->free = false; |
| 1620 | |
| 1621 | while (!flist_empty(&log->io_logs)) { |
| 1622 | struct io_logs *cur_log; |
| 1623 | |
| 1624 | cur_log = flist_first_entry(&log->io_logs, struct io_logs, list); |
| 1625 | flist_del_init(&cur_log->list); |
| 1626 | |
| 1627 | data->samples = cur_log->log; |
| 1628 | data->nr_samples = cur_log->nr_samples; |
| 1629 | |
| 1630 | sfree(cur_log); |
| 1631 | |
| 1632 | gz_work(data); |
| 1633 | } |
| 1634 | |
| 1635 | free(data); |
| 1636 | return 0; |
| 1637 | } |
| 1638 | |
| 1639 | int iolog_cur_flush(struct io_log *log, struct io_logs *cur_log) |
| 1640 | { |
| 1641 | struct iolog_flush_data *data; |
| 1642 | |
| 1643 | data = smalloc(sizeof(*data)); |
| 1644 | if (!data) |
| 1645 | return 1; |
| 1646 | |
| 1647 | data->log = log; |
| 1648 | |
| 1649 | data->samples = cur_log->log; |
| 1650 | data->nr_samples = cur_log->nr_samples; |
| 1651 | data->free = true; |
| 1652 | |
| 1653 | cur_log->nr_samples = cur_log->max_samples = 0; |
| 1654 | cur_log->log = NULL; |
| 1655 | |
| 1656 | workqueue_enqueue(&log->td->log_compress_wq, &data->work); |
| 1657 | |
| 1658 | iolog_free_deferred(log); |
| 1659 | |
| 1660 | return 0; |
| 1661 | } |
| 1662 | #else |
| 1663 | |
| 1664 | static int iolog_flush(struct io_log *log) |
| 1665 | { |
| 1666 | return 1; |
| 1667 | } |
| 1668 | |
| 1669 | int iolog_cur_flush(struct io_log *log, struct io_logs *cur_log) |
| 1670 | { |
| 1671 | return 1; |
| 1672 | } |
| 1673 | |
| 1674 | int iolog_compress_init(struct thread_data *td, struct sk_out *sk_out) |
| 1675 | { |
| 1676 | return 0; |
| 1677 | } |
| 1678 | |
| 1679 | void iolog_compress_exit(struct thread_data *td) |
| 1680 | { |
| 1681 | } |
| 1682 | |
| 1683 | #endif |
| 1684 | |
| 1685 | struct io_logs *iolog_cur_log(struct io_log *log) |
| 1686 | { |
| 1687 | if (flist_empty(&log->io_logs)) |
| 1688 | return NULL; |
| 1689 | |
| 1690 | return flist_last_entry(&log->io_logs, struct io_logs, list); |
| 1691 | } |
| 1692 | |
| 1693 | uint64_t iolog_nr_samples(struct io_log *iolog) |
| 1694 | { |
| 1695 | struct flist_head *entry; |
| 1696 | uint64_t ret = 0; |
| 1697 | |
| 1698 | flist_for_each(entry, &iolog->io_logs) { |
| 1699 | struct io_logs *cur_log; |
| 1700 | |
| 1701 | cur_log = flist_entry(entry, struct io_logs, list); |
| 1702 | ret += cur_log->nr_samples; |
| 1703 | } |
| 1704 | |
| 1705 | return ret; |
| 1706 | } |
| 1707 | |
| 1708 | static int __write_log(struct thread_data *td, struct io_log *log, int try) |
| 1709 | { |
| 1710 | if (log) |
| 1711 | return finish_log(td, log, try); |
| 1712 | |
| 1713 | return 0; |
| 1714 | } |
| 1715 | |
| 1716 | static int write_iops_log(struct thread_data *td, int try, bool unit_log) |
| 1717 | { |
| 1718 | int ret; |
| 1719 | |
| 1720 | if (per_unit_log(td->iops_log) != unit_log) |
| 1721 | return 0; |
| 1722 | |
| 1723 | ret = __write_log(td, td->iops_log, try); |
| 1724 | if (!ret) |
| 1725 | td->iops_log = NULL; |
| 1726 | |
| 1727 | return ret; |
| 1728 | } |
| 1729 | |
| 1730 | static int write_slat_log(struct thread_data *td, int try, bool unit_log) |
| 1731 | { |
| 1732 | int ret; |
| 1733 | |
| 1734 | if (!unit_log) |
| 1735 | return 0; |
| 1736 | |
| 1737 | ret = __write_log(td, td->slat_log, try); |
| 1738 | if (!ret) |
| 1739 | td->slat_log = NULL; |
| 1740 | |
| 1741 | return ret; |
| 1742 | } |
| 1743 | |
| 1744 | static int write_clat_log(struct thread_data *td, int try, bool unit_log) |
| 1745 | { |
| 1746 | int ret; |
| 1747 | |
| 1748 | if (!unit_log) |
| 1749 | return 0; |
| 1750 | |
| 1751 | ret = __write_log(td, td->clat_log, try); |
| 1752 | if (!ret) |
| 1753 | td->clat_log = NULL; |
| 1754 | |
| 1755 | return ret; |
| 1756 | } |
| 1757 | |
| 1758 | static int write_clat_hist_log(struct thread_data *td, int try, bool unit_log) |
| 1759 | { |
| 1760 | int ret; |
| 1761 | |
| 1762 | if (!unit_log) |
| 1763 | return 0; |
| 1764 | |
| 1765 | ret = __write_log(td, td->clat_hist_log, try); |
| 1766 | if (!ret) |
| 1767 | td->clat_hist_log = NULL; |
| 1768 | |
| 1769 | return ret; |
| 1770 | } |
| 1771 | |
| 1772 | static int write_lat_log(struct thread_data *td, int try, bool unit_log) |
| 1773 | { |
| 1774 | int ret; |
| 1775 | |
| 1776 | if (!unit_log) |
| 1777 | return 0; |
| 1778 | |
| 1779 | ret = __write_log(td, td->lat_log, try); |
| 1780 | if (!ret) |
| 1781 | td->lat_log = NULL; |
| 1782 | |
| 1783 | return ret; |
| 1784 | } |
| 1785 | |
| 1786 | static int write_bandw_log(struct thread_data *td, int try, bool unit_log) |
| 1787 | { |
| 1788 | int ret; |
| 1789 | |
| 1790 | if (per_unit_log(td->bw_log) != unit_log) |
| 1791 | return 0; |
| 1792 | |
| 1793 | ret = __write_log(td, td->bw_log, try); |
| 1794 | if (!ret) |
| 1795 | td->bw_log = NULL; |
| 1796 | |
| 1797 | return ret; |
| 1798 | } |
| 1799 | |
| 1800 | enum { |
| 1801 | BW_LOG_MASK = 1, |
| 1802 | LAT_LOG_MASK = 2, |
| 1803 | SLAT_LOG_MASK = 4, |
| 1804 | CLAT_LOG_MASK = 8, |
| 1805 | IOPS_LOG_MASK = 16, |
| 1806 | CLAT_HIST_LOG_MASK = 32, |
| 1807 | |
| 1808 | ALL_LOG_NR = 6, |
| 1809 | }; |
| 1810 | |
| 1811 | struct log_type { |
| 1812 | unsigned int mask; |
| 1813 | int (*fn)(struct thread_data *, int, bool); |
| 1814 | }; |
| 1815 | |
| 1816 | static struct log_type log_types[] = { |
| 1817 | { |
| 1818 | .mask = BW_LOG_MASK, |
| 1819 | .fn = write_bandw_log, |
| 1820 | }, |
| 1821 | { |
| 1822 | .mask = LAT_LOG_MASK, |
| 1823 | .fn = write_lat_log, |
| 1824 | }, |
| 1825 | { |
| 1826 | .mask = SLAT_LOG_MASK, |
| 1827 | .fn = write_slat_log, |
| 1828 | }, |
| 1829 | { |
| 1830 | .mask = CLAT_LOG_MASK, |
| 1831 | .fn = write_clat_log, |
| 1832 | }, |
| 1833 | { |
| 1834 | .mask = IOPS_LOG_MASK, |
| 1835 | .fn = write_iops_log, |
| 1836 | }, |
| 1837 | { |
| 1838 | .mask = CLAT_HIST_LOG_MASK, |
| 1839 | .fn = write_clat_hist_log, |
| 1840 | } |
| 1841 | }; |
| 1842 | |
| 1843 | void td_writeout_logs(struct thread_data *td, bool unit_logs) |
| 1844 | { |
| 1845 | unsigned int log_mask = 0; |
| 1846 | unsigned int log_left = ALL_LOG_NR; |
| 1847 | int old_state, i; |
| 1848 | |
| 1849 | old_state = td_bump_runstate(td, TD_FINISHING); |
| 1850 | |
| 1851 | finalize_logs(td, unit_logs); |
| 1852 | |
| 1853 | while (log_left) { |
| 1854 | int prev_log_left = log_left; |
| 1855 | |
| 1856 | for (i = 0; i < ALL_LOG_NR && log_left; i++) { |
| 1857 | struct log_type *lt = &log_types[i]; |
| 1858 | int ret; |
| 1859 | |
| 1860 | if (!(log_mask & lt->mask)) { |
| 1861 | ret = lt->fn(td, log_left != 1, unit_logs); |
| 1862 | if (!ret) { |
| 1863 | log_left--; |
| 1864 | log_mask |= lt->mask; |
| 1865 | } |
| 1866 | } |
| 1867 | } |
| 1868 | |
| 1869 | if (prev_log_left == log_left) |
| 1870 | usleep(5000); |
| 1871 | } |
| 1872 | |
| 1873 | td_restore_runstate(td, old_state); |
| 1874 | } |
| 1875 | |
| 1876 | void fio_writeout_logs(bool unit_logs) |
| 1877 | { |
| 1878 | for_each_td(td) { |
| 1879 | td_writeout_logs(td, unit_logs); |
| 1880 | } end_for_each(); |
| 1881 | } |